diff options
Diffstat (limited to 'src/compiler')
76 files changed, 24404 insertions, 0 deletions
diff --git a/src/compiler/Android.gen.mk b/src/compiler/Android.gen.mk new file mode 100644 index 0000000..fcd5f94 --- /dev/null +++ b/src/compiler/Android.gen.mk @@ -0,0 +1,91 @@ +# Mesa 3-D graphics library +# +# Copyright (C) 2010-2011 Chia-I Wu <olvaffe@gmail.com> +# Copyright (C) 2010-2011 LunarG Inc. +# +# Permission is hereby granted, free of charge, to any person obtaining a +# copy of this software and associated documentation files (the "Software"), +# to deal in the Software without restriction, including without limitation +# the rights to use, copy, modify, merge, publish, distribute, sublicense, +# and/or sell copies of the Software, and to permit persons to whom the +# Software is furnished to do so, subject to the following conditions: +# +# The above copyright notice and this permission notice shall be included +# in all copies or substantial portions of the Software. +# +# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL +# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER +# DEALINGS IN THE SOFTWARE. + +# included by glsl Android.mk for source generation + +ifeq ($(LOCAL_MODULE_CLASS),) +LOCAL_MODULE_CLASS := STATIC_LIBRARIES +endif + +intermediates := $(call local-generated-sources-dir) + +LOCAL_SRC_FILES := $(LOCAL_SRC_FILES) + +LOCAL_C_INCLUDES += \ + $(intermediates)/nir \ + $(MESA_TOP)/src/compiler/nir + +LOCAL_EXPORT_C_INCLUDE_DIRS += \ + $(intermediates)/nir \ + $(MESA_TOP)/src/compiler/nir + +LOCAL_GENERATED_SOURCES += $(addprefix $(intermediates)/, \ + $(NIR_GENERATED_FILES)) + + +nir_builder_opcodes_gen := $(LOCAL_PATH)/nir/nir_builder_opcodes_h.py +nir_builder_opcodes_deps := \ + $(LOCAL_PATH)/nir/nir_opcodes.py \ + $(LOCAL_PATH)/nir/nir_builder_opcodes_h.py + +$(intermediates)/nir/nir_builder_opcodes.h: $(nir_builder_opcodes_deps) + @mkdir -p $(dir $@) + $(hide) $(MESA_PYTHON2) $(nir_builder_opcodes_gen) $< > $@ + +nir_constant_expressions_gen := $(LOCAL_PATH)/nir/nir_constant_expressions.py +nir_constant_expressions_deps := \ + $(LOCAL_PATH)/nir/nir_opcodes.py \ + $(LOCAL_PATH)/nir/nir_constant_expressions.py + +$(intermediates)/nir/nir_constant_expressions.c: $(nir_constant_expressions_deps) + @mkdir -p $(dir $@) + $(hide) $(MESA_PYTHON2) $(nir_constant_expressions_gen) $< > $@ + +nir_opcodes_h_gen := $(LOCAL_PATH)/nir/nir_opcodes_h.py +nir_opcodes_h_deps := \ + $(LOCAL_PATH)/nir/nir_opcodes.py \ + $(LOCAL_PATH)/nir/nir_opcodes_h.py + +$(intermediates)/nir/nir_opcodes.h: $(nir_opcodes_h_deps) + @mkdir -p $(dir $@) + $(hide) $(MESA_PYTHON2) $(nir_opcodes_h_gen) $< > $@ + +$(LOCAL_PATH)/nir/nir.h: $(intermediates)/nir/nir_opcodes.h + +nir_opcodes_c_gen := $(LOCAL_PATH)/nir/nir_opcodes_c.py +nir_opcodes_c_deps := \ + $(LOCAL_PATH)/nir/nir_opcodes.py \ + $(LOCAL_PATH)/nir/nir_opcodes_c.py + +$(intermediates)/nir/nir_opcodes.c: $(nir_opcodes_c_deps) + @mkdir -p $(dir $@) + $(hide) $(MESA_PYTHON2) $(nir_opcodes_c_gen) $< > $@ + +nir_opt_algebraic_gen := $(LOCAL_PATH)/nir/nir_opt_algebraic.py +nir_opt_algebraic_deps := \ + $(LOCAL_PATH)/nir/nir_opt_algebraic.py \ + $(LOCAL_PATH)/nir/nir_algebraic.py + +$(intermediates)/nir/nir_opt_algebraic.c: $(nir_opt_algebraic_deps) + @mkdir -p $(dir $@) + $(hide) $(MESA_PYTHON2) $(nir_opt_algebraic_gen) $< > $@ diff --git a/src/compiler/Android.mk b/src/compiler/Android.mk index 8ebd497..888780b 100644 --- a/src/compiler/Android.mk +++ b/src/compiler/Android.mk @@ -42,3 +42,26 @@ LOCAL_MODULE := libmesa_compiler include $(MESA_COMMON_MK) include $(BUILD_STATIC_LIBRARY) + +# --------------------------------------- +# Build libmesa_nir +# --------------------------------------- + +include $(CLEAR_VARS) + +LOCAL_SRC_FILES := \ + $(NIR_FILES) + +LOCAL_C_INCLUDES := \ + $(MESA_TOP)/src/mapi \ + $(MESA_TOP)/src/mesa \ + $(MESA_TOP)/src/gallium/include \ + $(MESA_TOP)/src/gallium/auxiliary + +LOCAL_STATIC_LIBRARIES := libmesa_compiler + +LOCAL_MODULE := libmesa_nir + +include $(LOCAL_PATH)/Android.gen.mk +include $(MESA_COMMON_MK) +include $(BUILD_STATIC_LIBRARY) diff --git a/src/compiler/Makefile.am b/src/compiler/Makefile.am index fd1dd4b..1e3778d 100644 --- a/src/compiler/Makefile.am +++ b/src/compiler/Makefile.am @@ -29,6 +29,7 @@ AM_CPPFLAGS = \ -I$(top_srcdir)/src/mesa/ \ -I$(top_srcdir)/src/gallium/include \ -I$(top_srcdir)/src/gallium/auxiliary \ + -I$(top_srcdir)/src/gtest/include \ $(DEFINES) AM_CFLAGS = \ @@ -43,4 +44,81 @@ noinst_LTLIBRARIES = libcompiler.la libcompiler_la_SOURCES = $(LIBCOMPILER_FILES) +check_PROGRAMS = +TESTS = +BUILT_SOURCES = +CLEANFILES = EXTRA_DIST = SConscript + + +noinst_LTLIBRARIES += nir/libnir.la + +nir_libnir_la_CPPFLAGS = \ + $(AM_CPPFLAGS) \ + -I$(top_builddir)/src/compiler/nir \ + -I$(top_srcdir)/src/compiler/nir + +nir_libnir_la_LIBADD = \ + libcompiler.la + +nir_libnir_la_SOURCES = \ + $(NIR_FILES) \ + $(NIR_GENERATED_FILES) + +PYTHON_GEN = $(AM_V_GEN)$(PYTHON2) $(PYTHON_FLAGS) + +nir/nir_builder_opcodes.h: nir/nir_opcodes.py nir/nir_builder_opcodes_h.py + $(MKDIR_GEN) + $(PYTHON_GEN) $(srcdir)/nir/nir_builder_opcodes_h.py > $@ || ($(RM) $@; false) + +nir/nir_constant_expressions.c: nir/nir_opcodes.py nir/nir_constant_expressions.py + $(MKDIR_GEN) + $(PYTHON_GEN) $(srcdir)/nir/nir_constant_expressions.py > $@ || ($(RM) $@; false) + +nir/nir_opcodes.h: nir/nir_opcodes.py nir/nir_opcodes_h.py + $(MKDIR_GEN) + $(PYTHON_GEN) $(srcdir)/nir/nir_opcodes_h.py > $@ || ($(RM) $@; false) + +nir/nir_opcodes.c: nir/nir_opcodes.py nir/nir_opcodes_c.py + $(MKDIR_GEN) + $(PYTHON_GEN) $(srcdir)/nir/nir_opcodes_c.py > $@ || ($(RM) $@; false) + +nir/nir_opt_algebraic.c: nir/nir_opt_algebraic.py nir/nir_algebraic.py + $(MKDIR_GEN) + $(PYTHON_GEN) $(srcdir)/nir/nir_opt_algebraic.py > $@ || ($(RM) $@; false) + + +check_PROGRAMS += nir/tests/control_flow_tests + +nir_tests_control_flow_tests_CPPFLAGS = \ + $(AM_CPPFLAGS) \ + -I$(top_builddir)/src/compiler/nir \ + -I$(top_srcdir)/src/compiler/nir + +nir_tests_control_flow_tests_SOURCES = \ + nir/tests/control_flow_tests.cpp +nir_tests_control_flow_tests_CFLAGS = \ + $(PTHREAD_CFLAGS) +nir_tests_control_flow_tests_LDADD = \ + $(top_builddir)/src/gtest/libgtest.la \ + $(top_builddir)/src/compiler/nir/libnir.la \ + $(top_builddir)/src/util/libmesautil.la \ + $(PTHREAD_LIBS) + + +TESTS += nir/tests/control_flow_tests + + +BUILT_SOURCES += $(NIR_GENERATED_FILES) +CLEANFILES += $(NIR_GENERATED_FILES) + +EXTRA_DIST += \ + nir/nir_algebraic.py \ + nir/nir_builder_opcodes_h.py \ + nir/nir_constant_expressions.py \ + nir/nir_opcodes.py \ + nir/nir_opcodes_c.py \ + nir/nir_opcodes_h.py \ + nir/nir_opt_algebraic.py \ + nir/tests \ + nir/Makefile.sources diff --git a/src/compiler/Makefile.sources b/src/compiler/Makefile.sources index e1228ca..fe12e41 100644 --- a/src/compiler/Makefile.sources +++ b/src/compiler/Makefile.sources @@ -6,3 +6,74 @@ LIBCOMPILER_FILES = \ nir_types.h \ shader_enums.c \ shader_enums.h + +NIR_GENERATED_FILES = \ + nir/nir_builder_opcodes.h \ + nir/nir_constant_expressions.c \ + nir/nir_opcodes.c \ + nir/nir_opcodes.h \ + nir/nir_opt_algebraic.c + +NIR_FILES = \ + nir/glsl_to_nir.cpp \ + nir/glsl_to_nir.h \ + nir/nir.c \ + nir/nir.h \ + nir/nir_array.h \ + nir/nir_builder.h \ + nir/nir_clone.c \ + nir/nir_constant_expressions.h \ + nir/nir_control_flow.c \ + nir/nir_control_flow.h \ + nir/nir_control_flow_private.h \ + nir/nir_dominance.c \ + nir/nir_from_ssa.c \ + nir/nir_gs_count_vertices.c \ + nir/nir_intrinsics.c \ + nir/nir_intrinsics.h \ + nir/nir_instr_set.c \ + nir/nir_instr_set.h \ + nir/nir_liveness.c \ + nir/nir_lower_alu_to_scalar.c \ + nir/nir_lower_atomics.c \ + nir/nir_lower_clip.c \ + nir/nir_lower_global_vars_to_local.c \ + nir/nir_lower_gs_intrinsics.c \ + nir/nir_lower_load_const_to_scalar.c \ + nir/nir_lower_locals_to_regs.c \ + nir/nir_lower_idiv.c \ + nir/nir_lower_io.c \ + nir/nir_lower_outputs_to_temporaries.c \ + nir/nir_lower_phis_to_scalar.c \ + nir/nir_lower_samplers.c \ + nir/nir_lower_system_values.c \ + nir/nir_lower_tex.c \ + nir/nir_lower_to_source_mods.c \ + nir/nir_lower_two_sided_color.c \ + nir/nir_lower_vars_to_ssa.c \ + nir/nir_lower_var_copies.c \ + nir/nir_lower_vec_to_movs.c \ + nir/nir_metadata.c \ + nir/nir_move_vec_src_uses_to_dest.c \ + nir/nir_normalize_cubemap_coords.c \ + nir/nir_opt_constant_folding.c \ + nir/nir_opt_copy_propagate.c \ + nir/nir_opt_cse.c \ + nir/nir_opt_dce.c \ + nir/nir_opt_dead_cf.c \ + nir/nir_opt_gcm.c \ + nir/nir_opt_global_to_local.c \ + nir/nir_opt_peephole_select.c \ + nir/nir_opt_remove_phis.c \ + nir/nir_opt_undef.c \ + nir/nir_print.c \ + nir/nir_remove_dead_variables.c \ + nir/nir_search.c \ + nir/nir_search.h \ + nir/nir_split_var_copies.c \ + nir/nir_sweep.c \ + nir/nir_to_ssa.c \ + nir/nir_validate.c \ + nir/nir_vla.h \ + nir/nir_worklist.c \ + nir/nir_worklist.h diff --git a/src/compiler/nir/.gitignore b/src/compiler/nir/.gitignore new file mode 100644 index 0000000..64828eb --- /dev/null +++ b/src/compiler/nir/.gitignore @@ -0,0 +1,5 @@ +nir_builder_opcodes.h +nir_opt_algebraic.c +nir_opcodes.c +nir_opcodes.h +nir_constant_expressions.c diff --git a/src/compiler/nir/README b/src/compiler/nir/README new file mode 100644 index 0000000..2c81db9 --- /dev/null +++ b/src/compiler/nir/README @@ -0,0 +1,118 @@ +New IR, or NIR, is an IR for Mesa intended to sit below GLSL IR and Mesa IR. +Its design inherits from the various IR's that Mesa has used in the past, as +well as Direct3D assembly, and it includes a few new ideas as well. It is a +flat (in terms of using instructions instead of expressions), typeless IR, +similar to TGSI and Mesa IR. It also supports SSA (although it doesn't require +it). + +Variables +========= + +NIR includes support for source-level GLSL variables through a structure mostly +copied from GLSL IR. These will be used for linking and conversion from GLSL IR +(and later, from an AST), but for the most part, they will be lowered to +registers (see below) and loads/stores. + +Registers +========= + +Registers are light-weight; they consist of a structure that only contains its +size, its index for liveness analysis, and an optional name for debugging. In +addition, registers can be local to a function or global to the entire shader; +the latter will be used in ARB_shader_subroutine for passing parameters and +getting return values from subroutines. Registers can also be an array, in which +case they can be accessed indirectly. Each ALU instruction (add, subtract, etc.) +works directly with registers or SSA values (see below). + +SSA +======== + +Everywhere a register can be loaded/stored, an SSA value can be used instead. +The only exception is that arrays/indirect addressing are not supported with +SSA; although research has been done on extensions of SSA to arrays before, it's +usually for the purpose of parallelization (which we're not interested in), and +adds some overhead in the form of adding copies or extra arrays (which is much +more expensive than introducing copies between non-array registers). SSA uses +point directly to their corresponding definition, which in turn points to the +instruction it is part of. This creates an implicit use-def chain and avoids the +need for an external structure for each SSA register. + +Functions +========= + +Support for function calls is mostly similar to GLSL IR. Each shader contains a +list of functions, and each function has a list of overloads. Each overload +contains a list of parameters, and may contain an implementation which specifies +the variables that correspond to the parameters and return value. Inlining a +function, assuming it has a single return point, is as simple as copying its +instructions, registers, and local variables into the target function and then +inserting copies to and from the new parameters as appropriate. After functions +are inlined and any non-subroutine functions are deleted, parameters and return +variables will be converted to global variables and then global registers. We +don't do this lowering earlier (i.e. the fortranizer idea) for a few reasons: + +- If we want to do optimizations before link time, we need to have the function +signature available during link-time. + +- If we do any inlining before link time, then we might wind up with the +inlined function and the non-inlined function using the same global +variables/registers which would preclude optimization. + +Intrinsics +========= + +Any operation (other than function calls and textures) which touches a variable +or is not referentially transparent is represented by an intrinsic. Intrinsics +are similar to the idea of a "builtin function," i.e. a function declaration +whose implementation is provided by the backend, except they are more powerful +in the following ways: + +- They can also load and store registers when appropriate, which limits the +number of variables needed in later stages of the IR while obviating the need +for a separate load/store variable instruction. + +- Intrinsics can be marked as side-effect free, which permits them to be +treated like any other instruction when it comes to optimizations. This allows +load intrinsics to be represented as intrinsics while still being optimized +away by dead code elimination, common subexpression elimination, etc. + +Intrinsics are used for: + +- Atomic operations +- Memory barriers +- Subroutine calls +- Geometry shader emitVertex and endPrimitive +- Loading and storing variables (before lowering) +- Loading and storing uniforms, shader inputs and outputs, etc (after lowering) +- Copying variables (cases where in GLSL the destination is a structure or +array) +- The kitchen sink +- ... + +Textures +========= + +Unfortunately, there are far too many texture operations to represent each one +of them with an intrinsic, so there's a special texture instruction similar to +the GLSL IR one. The biggest difference is that, while the texture instruction +has a sampler dereference field used just like in GLSL IR, this gets lowered to +a texture unit index (with a possible indirect offset) while the type +information of the original sampler is kept around for backends. Also, all the +non-constant sources are stored in a single array to make it easier for +optimization passes to iterate over all the sources. + +Control Flow +========= + +Like in GLSL IR, control flow consists of a tree of "control flow nodes", which +include if statements and loops, and jump instructions (break, continue, and +return). Unlike GLSL IR, though, the leaves of the tree aren't statements but +basic blocks. Each basic block also keeps track of its successors and +predecessors, and function implementations keep track of the beginning basic +block (the first basic block of the function) and the ending basic block (a fake +basic block that every return statement points to). Together, these elements +make up the control flow graph, in this case a redundant piece of information on +top of the control flow tree that will be used by almost all the optimizations. +There are helper functions to add and remove control flow nodes that also update +the control flow graph, and so usually it doesn't need to be touched by passes +that modify control flow nodes. diff --git a/src/compiler/nir/glsl_to_nir.cpp b/src/compiler/nir/glsl_to_nir.cpp new file mode 100644 index 0000000..33b1f5c --- /dev/null +++ b/src/compiler/nir/glsl_to_nir.cpp @@ -0,0 +1,2031 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +#include "glsl_to_nir.h" +#include "nir_control_flow.h" +#include "nir_builder.h" +#include "glsl/ir_visitor.h" +#include "glsl/ir_hierarchical_visitor.h" +#include "glsl/ir.h" +#include "main/imports.h" + +/* + * pass to lower GLSL IR to NIR + * + * This will lower variable dereferences to loads/stores of corresponding + * variables in NIR - the variables will be converted to registers in a later + * pass. + */ + +namespace { + +class nir_visitor : public ir_visitor +{ +public: + nir_visitor(nir_shader *shader); + ~nir_visitor(); + + virtual void visit(ir_variable *); + virtual void visit(ir_function *); + virtual void visit(ir_function_signature *); + virtual void visit(ir_loop *); + virtual void visit(ir_if *); + virtual void visit(ir_discard *); + virtual void visit(ir_loop_jump *); + virtual void visit(ir_return *); + virtual void visit(ir_call *); + virtual void visit(ir_assignment *); + virtual void visit(ir_emit_vertex *); + virtual void visit(ir_end_primitive *); + virtual void visit(ir_expression *); + virtual void visit(ir_swizzle *); + virtual void visit(ir_texture *); + virtual void visit(ir_constant *); + virtual void visit(ir_dereference_variable *); + virtual void visit(ir_dereference_record *); + virtual void visit(ir_dereference_array *); + virtual void visit(ir_barrier *); + + void create_function(ir_function_signature *ir); + +private: + void add_instr(nir_instr *instr, unsigned num_components); + nir_ssa_def *evaluate_rvalue(ir_rvalue *ir); + + nir_alu_instr *emit(nir_op op, unsigned dest_size, nir_ssa_def **srcs); + nir_alu_instr *emit(nir_op op, unsigned dest_size, nir_ssa_def *src1); + nir_alu_instr *emit(nir_op op, unsigned dest_size, nir_ssa_def *src1, + nir_ssa_def *src2); + nir_alu_instr *emit(nir_op op, unsigned dest_size, nir_ssa_def *src1, + nir_ssa_def *src2, nir_ssa_def *src3); + + bool supports_ints; + + nir_shader *shader; + nir_function_impl *impl; + nir_builder b; + nir_ssa_def *result; /* result of the expression tree last visited */ + + nir_deref_var *evaluate_deref(nir_instr *mem_ctx, ir_instruction *ir); + + /* the head of the dereference chain we're creating */ + nir_deref_var *deref_head; + /* the tail of the dereference chain we're creating */ + nir_deref *deref_tail; + + nir_variable *var; /* variable created by ir_variable visitor */ + + /* whether the IR we're operating on is per-function or global */ + bool is_global; + + /* map of ir_variable -> nir_variable */ + struct hash_table *var_table; + + /* map of ir_function_signature -> nir_function_overload */ + struct hash_table *overload_table; +}; + +/* + * This visitor runs before the main visitor, calling create_function() for + * each function so that the main visitor can resolve forward references in + * calls. + */ + +class nir_function_visitor : public ir_hierarchical_visitor +{ +public: + nir_function_visitor(nir_visitor *v) : visitor(v) + { + } + virtual ir_visitor_status visit_enter(ir_function *); + +private: + nir_visitor *visitor; +}; + +}; /* end of anonymous namespace */ + +nir_shader * +glsl_to_nir(const struct gl_shader_program *shader_prog, + gl_shader_stage stage, + const nir_shader_compiler_options *options) +{ + struct gl_shader *sh = shader_prog->_LinkedShaders[stage]; + + nir_shader *shader = nir_shader_create(NULL, stage, options); + + nir_visitor v1(shader); + nir_function_visitor v2(&v1); + v2.run(sh->ir); + visit_exec_list(sh->ir, &v1); + + nir_lower_outputs_to_temporaries(shader); + + shader->info.name = ralloc_asprintf(shader, "GLSL%d", shader_prog->Name); + if (shader_prog->Label) + shader->info.label = ralloc_strdup(shader, shader_prog->Label); + shader->info.num_textures = _mesa_fls(sh->Program->SamplersUsed); + shader->info.num_ubos = sh->NumUniformBlocks; + shader->info.num_abos = shader_prog->NumAtomicBuffers; + shader->info.num_ssbos = sh->NumShaderStorageBlocks; + shader->info.num_images = sh->NumImages; + shader->info.inputs_read = sh->Program->InputsRead; + shader->info.outputs_written = sh->Program->OutputsWritten; + shader->info.patch_inputs_read = sh->Program->PatchInputsRead; + shader->info.patch_outputs_written = sh->Program->PatchOutputsWritten; + shader->info.system_values_read = sh->Program->SystemValuesRead; + shader->info.uses_texture_gather = sh->Program->UsesGather; + shader->info.uses_clip_distance_out = + sh->Program->ClipDistanceArraySize != 0; + shader->info.separate_shader = shader_prog->SeparateShader; + shader->info.has_transform_feedback_varyings = + shader_prog->TransformFeedback.NumVarying > 0; + + switch (stage) { + case MESA_SHADER_TESS_CTRL: + shader->info.tcs.vertices_out = shader_prog->TessCtrl.VerticesOut; + break; + + case MESA_SHADER_GEOMETRY: + shader->info.gs.vertices_in = shader_prog->Geom.VerticesIn; + shader->info.gs.output_primitive = sh->Geom.OutputType; + shader->info.gs.vertices_out = sh->Geom.VerticesOut; + shader->info.gs.invocations = sh->Geom.Invocations; + shader->info.gs.uses_end_primitive = shader_prog->Geom.UsesEndPrimitive; + shader->info.gs.uses_streams = shader_prog->Geom.UsesStreams; + break; + + case MESA_SHADER_FRAGMENT: { + struct gl_fragment_program *fp = + (struct gl_fragment_program *)sh->Program; + + shader->info.fs.uses_discard = fp->UsesKill; + shader->info.fs.early_fragment_tests = sh->EarlyFragmentTests; + shader->info.fs.depth_layout = fp->FragDepthLayout; + break; + } + + case MESA_SHADER_COMPUTE: { + struct gl_compute_program *cp = (struct gl_compute_program *)sh->Program; + shader->info.cs.local_size[0] = cp->LocalSize[0]; + shader->info.cs.local_size[1] = cp->LocalSize[1]; + shader->info.cs.local_size[2] = cp->LocalSize[2]; + break; + } + + default: + break; /* No stage-specific info */ + } + + return shader; +} + +nir_visitor::nir_visitor(nir_shader *shader) +{ + this->supports_ints = shader->options->native_integers; + this->shader = shader; + this->is_global = true; + this->var_table = _mesa_hash_table_create(NULL, _mesa_hash_pointer, + _mesa_key_pointer_equal); + this->overload_table = _mesa_hash_table_create(NULL, _mesa_hash_pointer, + _mesa_key_pointer_equal); +} + +nir_visitor::~nir_visitor() +{ + _mesa_hash_table_destroy(this->var_table, NULL); + _mesa_hash_table_destroy(this->overload_table, NULL); +} + +nir_deref_var * +nir_visitor::evaluate_deref(nir_instr *mem_ctx, ir_instruction *ir) +{ + ir->accept(this); + ralloc_steal(mem_ctx, this->deref_head); + return this->deref_head; +} + +static nir_constant * +constant_copy(ir_constant *ir, void *mem_ctx) +{ + if (ir == NULL) + return NULL; + + nir_constant *ret = ralloc(mem_ctx, nir_constant); + + unsigned total_elems = ir->type->components(); + unsigned i; + + ret->num_elements = 0; + switch (ir->type->base_type) { + case GLSL_TYPE_UINT: + for (i = 0; i < total_elems; i++) + ret->value.u[i] = ir->value.u[i]; + break; + + case GLSL_TYPE_INT: + for (i = 0; i < total_elems; i++) + ret->value.i[i] = ir->value.i[i]; + break; + + case GLSL_TYPE_FLOAT: + for (i = 0; i < total_elems; i++) + ret->value.f[i] = ir->value.f[i]; + break; + + case GLSL_TYPE_BOOL: + for (i = 0; i < total_elems; i++) + ret->value.b[i] = ir->value.b[i]; + break; + + case GLSL_TYPE_STRUCT: + ret->elements = ralloc_array(mem_ctx, nir_constant *, + ir->type->length); + ret->num_elements = ir->type->length; + + i = 0; + foreach_in_list(ir_constant, field, &ir->components) { + ret->elements[i] = constant_copy(field, mem_ctx); + i++; + } + break; + + case GLSL_TYPE_ARRAY: + ret->elements = ralloc_array(mem_ctx, nir_constant *, + ir->type->length); + ret->num_elements = ir->type->length; + + for (i = 0; i < ir->type->length; i++) + ret->elements[i] = constant_copy(ir->array_elements[i], mem_ctx); + break; + + default: + unreachable("not reached"); + } + + return ret; +} + +void +nir_visitor::visit(ir_variable *ir) +{ + nir_variable *var = ralloc(shader, nir_variable); + var->type = ir->type; + var->name = ralloc_strdup(var, ir->name); + + var->data.read_only = ir->data.read_only; + var->data.centroid = ir->data.centroid; + var->data.sample = ir->data.sample; + var->data.patch = ir->data.patch; + var->data.invariant = ir->data.invariant; + var->data.location = ir->data.location; + + switch(ir->data.mode) { + case ir_var_auto: + case ir_var_temporary: + if (is_global) + var->data.mode = nir_var_global; + else + var->data.mode = nir_var_local; + break; + + case ir_var_function_in: + case ir_var_function_out: + case ir_var_function_inout: + case ir_var_const_in: + var->data.mode = nir_var_local; + break; + + case ir_var_shader_in: + if (shader->stage == MESA_SHADER_FRAGMENT && + ir->data.location == VARYING_SLOT_FACE) { + /* For whatever reason, GLSL IR makes gl_FrontFacing an input */ + var->data.location = SYSTEM_VALUE_FRONT_FACE; + var->data.mode = nir_var_system_value; + } else if (shader->stage == MESA_SHADER_GEOMETRY && + ir->data.location == VARYING_SLOT_PRIMITIVE_ID) { + /* For whatever reason, GLSL IR makes gl_PrimitiveIDIn an input */ + var->data.location = SYSTEM_VALUE_PRIMITIVE_ID; + var->data.mode = nir_var_system_value; + } else { + var->data.mode = nir_var_shader_in; + } + break; + + case ir_var_shader_out: + var->data.mode = nir_var_shader_out; + break; + + case ir_var_uniform: + var->data.mode = nir_var_uniform; + break; + + case ir_var_shader_storage: + var->data.mode = nir_var_shader_storage; + break; + + case ir_var_system_value: + var->data.mode = nir_var_system_value; + break; + + default: + unreachable("not reached"); + } + + var->data.interpolation = ir->data.interpolation; + var->data.origin_upper_left = ir->data.origin_upper_left; + var->data.pixel_center_integer = ir->data.pixel_center_integer; + var->data.explicit_location = ir->data.explicit_location; + var->data.explicit_index = ir->data.explicit_index; + var->data.explicit_binding = ir->data.explicit_binding; + var->data.has_initializer = ir->data.has_initializer; + var->data.location_frac = ir->data.location_frac; + var->data.from_named_ifc_block_array = ir->data.from_named_ifc_block_array; + var->data.from_named_ifc_block_nonarray = ir->data.from_named_ifc_block_nonarray; + + switch (ir->data.depth_layout) { + case ir_depth_layout_none: + var->data.depth_layout = nir_depth_layout_none; + break; + case ir_depth_layout_any: + var->data.depth_layout = nir_depth_layout_any; + break; + case ir_depth_layout_greater: + var->data.depth_layout = nir_depth_layout_greater; + break; + case ir_depth_layout_less: + var->data.depth_layout = nir_depth_layout_less; + break; + case ir_depth_layout_unchanged: + var->data.depth_layout = nir_depth_layout_unchanged; + break; + default: + unreachable("not reached"); + } + + var->data.index = ir->data.index; + var->data.binding = ir->data.binding; + var->data.offset = ir->data.offset; + var->data.image.read_only = ir->data.image_read_only; + var->data.image.write_only = ir->data.image_write_only; + var->data.image.coherent = ir->data.image_coherent; + var->data.image._volatile = ir->data.image_volatile; + var->data.image.restrict_flag = ir->data.image_restrict; + var->data.image.format = ir->data.image_format; + var->data.max_array_access = ir->data.max_array_access; + + var->num_state_slots = ir->get_num_state_slots(); + if (var->num_state_slots > 0) { + var->state_slots = ralloc_array(var, nir_state_slot, + var->num_state_slots); + + ir_state_slot *state_slots = ir->get_state_slots(); + for (unsigned i = 0; i < var->num_state_slots; i++) { + for (unsigned j = 0; j < 5; j++) + var->state_slots[i].tokens[j] = state_slots[i].tokens[j]; + var->state_slots[i].swizzle = state_slots[i].swizzle; + } + } else { + var->state_slots = NULL; + } + + var->constant_initializer = constant_copy(ir->constant_initializer, var); + + var->interface_type = ir->get_interface_type(); + + if (var->data.mode == nir_var_local) + nir_function_impl_add_variable(impl, var); + else + nir_shader_add_variable(shader, var); + + _mesa_hash_table_insert(var_table, ir, var); + this->var = var; +} + +ir_visitor_status +nir_function_visitor::visit_enter(ir_function *ir) +{ + foreach_in_list(ir_function_signature, sig, &ir->signatures) { + visitor->create_function(sig); + } + return visit_continue_with_parent; +} + +void +nir_visitor::create_function(ir_function_signature *ir) +{ + if (ir->is_intrinsic) + return; + + nir_function *func = nir_function_create(shader, ir->function_name()); + + unsigned num_params = ir->parameters.length(); + func->num_params = num_params; + func->params = ralloc_array(shader, nir_parameter, num_params); + + unsigned i = 0; + foreach_in_list(ir_variable, param, &ir->parameters) { + switch (param->data.mode) { + case ir_var_function_in: + func->params[i].param_type = nir_parameter_in; + break; + + case ir_var_function_out: + func->params[i].param_type = nir_parameter_out; + break; + + case ir_var_function_inout: + func->params[i].param_type = nir_parameter_inout; + break; + + default: + unreachable("not reached"); + } + + func->params[i].type = param->type; + i++; + } + + func->return_type = ir->return_type; + + _mesa_hash_table_insert(this->overload_table, ir, func); +} + +void +nir_visitor::visit(ir_function *ir) +{ + foreach_in_list(ir_function_signature, sig, &ir->signatures) + sig->accept(this); +} + +void +nir_visitor::visit(ir_function_signature *ir) +{ + if (ir->is_intrinsic) + return; + + struct hash_entry *entry = + _mesa_hash_table_search(this->overload_table, ir); + + assert(entry); + nir_function *func = (nir_function *) entry->data; + + if (ir->is_defined) { + nir_function_impl *impl = nir_function_impl_create(func); + this->impl = impl; + + unsigned num_params = func->num_params; + impl->num_params = num_params; + impl->params = ralloc_array(this->shader, nir_variable *, num_params); + unsigned i = 0; + foreach_in_list(ir_variable, param, &ir->parameters) { + param->accept(this); + impl->params[i] = this->var; + i++; + } + + if (func->return_type == glsl_type::void_type) { + impl->return_var = NULL; + } else { + impl->return_var = ralloc(this->shader, nir_variable); + impl->return_var->name = ralloc_strdup(impl->return_var, + "return_var"); + impl->return_var->type = func->return_type; + } + + this->is_global = false; + + nir_builder_init(&b, impl); + b.cursor = nir_after_cf_list(&impl->body); + visit_exec_list(&ir->body, this); + + this->is_global = true; + } else { + func->impl = NULL; + } +} + +void +nir_visitor::visit(ir_loop *ir) +{ + nir_loop *loop = nir_loop_create(this->shader); + nir_builder_cf_insert(&b, &loop->cf_node); + + b.cursor = nir_after_cf_list(&loop->body); + visit_exec_list(&ir->body_instructions, this); + b.cursor = nir_after_cf_node(&loop->cf_node); +} + +void +nir_visitor::visit(ir_if *ir) +{ + nir_src condition = + nir_src_for_ssa(evaluate_rvalue(ir->condition)); + + nir_if *if_stmt = nir_if_create(this->shader); + if_stmt->condition = condition; + nir_builder_cf_insert(&b, &if_stmt->cf_node); + + b.cursor = nir_after_cf_list(&if_stmt->then_list); + visit_exec_list(&ir->then_instructions, this); + + b.cursor = nir_after_cf_list(&if_stmt->else_list); + visit_exec_list(&ir->else_instructions, this); + + b.cursor = nir_after_cf_node(&if_stmt->cf_node); +} + +void +nir_visitor::visit(ir_discard *ir) +{ + /* + * discards aren't treated as control flow, because before we lower them + * they can appear anywhere in the shader and the stuff after them may still + * be executed (yay, crazy GLSL rules!). However, after lowering, all the + * discards will be immediately followed by a return. + */ + + nir_intrinsic_instr *discard; + if (ir->condition) { + discard = nir_intrinsic_instr_create(this->shader, + nir_intrinsic_discard_if); + discard->src[0] = + nir_src_for_ssa(evaluate_rvalue(ir->condition)); + } else { + discard = nir_intrinsic_instr_create(this->shader, nir_intrinsic_discard); + } + + nir_builder_instr_insert(&b, &discard->instr); +} + +void +nir_visitor::visit(ir_emit_vertex *ir) +{ + nir_intrinsic_instr *instr = + nir_intrinsic_instr_create(this->shader, nir_intrinsic_emit_vertex); + instr->const_index[0] = ir->stream_id(); + nir_builder_instr_insert(&b, &instr->instr); +} + +void +nir_visitor::visit(ir_end_primitive *ir) +{ + nir_intrinsic_instr *instr = + nir_intrinsic_instr_create(this->shader, nir_intrinsic_end_primitive); + instr->const_index[0] = ir->stream_id(); + nir_builder_instr_insert(&b, &instr->instr); +} + +void +nir_visitor::visit(ir_loop_jump *ir) +{ + nir_jump_type type; + switch (ir->mode) { + case ir_loop_jump::jump_break: + type = nir_jump_break; + break; + case ir_loop_jump::jump_continue: + type = nir_jump_continue; + break; + default: + unreachable("not reached"); + } + + nir_jump_instr *instr = nir_jump_instr_create(this->shader, type); + nir_builder_instr_insert(&b, &instr->instr); +} + +void +nir_visitor::visit(ir_return *ir) +{ + if (ir->value != NULL) { + nir_intrinsic_instr *copy = + nir_intrinsic_instr_create(this->shader, nir_intrinsic_copy_var); + + copy->variables[0] = nir_deref_var_create(copy, this->impl->return_var); + copy->variables[1] = evaluate_deref(©->instr, ir->value); + } + + nir_jump_instr *instr = nir_jump_instr_create(this->shader, nir_jump_return); + nir_builder_instr_insert(&b, &instr->instr); +} + +void +nir_visitor::visit(ir_call *ir) +{ + if (ir->callee->is_intrinsic) { + nir_intrinsic_op op; + if (strcmp(ir->callee_name(), "__intrinsic_atomic_read") == 0) { + op = nir_intrinsic_atomic_counter_read_var; + } else if (strcmp(ir->callee_name(), "__intrinsic_atomic_increment") == 0) { + op = nir_intrinsic_atomic_counter_inc_var; + } else if (strcmp(ir->callee_name(), "__intrinsic_atomic_predecrement") == 0) { + op = nir_intrinsic_atomic_counter_dec_var; + } else if (strcmp(ir->callee_name(), "__intrinsic_image_load") == 0) { + op = nir_intrinsic_image_load; + } else if (strcmp(ir->callee_name(), "__intrinsic_image_store") == 0) { + op = nir_intrinsic_image_store; + } else if (strcmp(ir->callee_name(), "__intrinsic_image_atomic_add") == 0) { + op = nir_intrinsic_image_atomic_add; + } else if (strcmp(ir->callee_name(), "__intrinsic_image_atomic_min") == 0) { + op = nir_intrinsic_image_atomic_min; + } else if (strcmp(ir->callee_name(), "__intrinsic_image_atomic_max") == 0) { + op = nir_intrinsic_image_atomic_max; + } else if (strcmp(ir->callee_name(), "__intrinsic_image_atomic_and") == 0) { + op = nir_intrinsic_image_atomic_and; + } else if (strcmp(ir->callee_name(), "__intrinsic_image_atomic_or") == 0) { + op = nir_intrinsic_image_atomic_or; + } else if (strcmp(ir->callee_name(), "__intrinsic_image_atomic_xor") == 0) { + op = nir_intrinsic_image_atomic_xor; + } else if (strcmp(ir->callee_name(), "__intrinsic_image_atomic_exchange") == 0) { + op = nir_intrinsic_image_atomic_exchange; + } else if (strcmp(ir->callee_name(), "__intrinsic_image_atomic_comp_swap") == 0) { + op = nir_intrinsic_image_atomic_comp_swap; + } else if (strcmp(ir->callee_name(), "__intrinsic_memory_barrier") == 0) { + op = nir_intrinsic_memory_barrier; + } else if (strcmp(ir->callee_name(), "__intrinsic_image_size") == 0) { + op = nir_intrinsic_image_size; + } else if (strcmp(ir->callee_name(), "__intrinsic_image_samples") == 0) { + op = nir_intrinsic_image_samples; + } else if (strcmp(ir->callee_name(), "__intrinsic_store_ssbo") == 0) { + op = nir_intrinsic_store_ssbo; + } else if (strcmp(ir->callee_name(), "__intrinsic_load_ssbo") == 0) { + op = nir_intrinsic_load_ssbo; + } else if (strcmp(ir->callee_name(), "__intrinsic_atomic_add_ssbo") == 0) { + op = nir_intrinsic_ssbo_atomic_add; + } else if (strcmp(ir->callee_name(), "__intrinsic_atomic_and_ssbo") == 0) { + op = nir_intrinsic_ssbo_atomic_and; + } else if (strcmp(ir->callee_name(), "__intrinsic_atomic_or_ssbo") == 0) { + op = nir_intrinsic_ssbo_atomic_or; + } else if (strcmp(ir->callee_name(), "__intrinsic_atomic_xor_ssbo") == 0) { + op = nir_intrinsic_ssbo_atomic_xor; + } else if (strcmp(ir->callee_name(), "__intrinsic_atomic_min_ssbo") == 0) { + assert(ir->return_deref); + if (ir->return_deref->type == glsl_type::int_type) + op = nir_intrinsic_ssbo_atomic_imin; + else if (ir->return_deref->type == glsl_type::uint_type) + op = nir_intrinsic_ssbo_atomic_umin; + else + unreachable("Invalid type"); + } else if (strcmp(ir->callee_name(), "__intrinsic_atomic_max_ssbo") == 0) { + assert(ir->return_deref); + if (ir->return_deref->type == glsl_type::int_type) + op = nir_intrinsic_ssbo_atomic_imax; + else if (ir->return_deref->type == glsl_type::uint_type) + op = nir_intrinsic_ssbo_atomic_umax; + else + unreachable("Invalid type"); + } else if (strcmp(ir->callee_name(), "__intrinsic_atomic_exchange_ssbo") == 0) { + op = nir_intrinsic_ssbo_atomic_exchange; + } else if (strcmp(ir->callee_name(), "__intrinsic_atomic_comp_swap_ssbo") == 0) { + op = nir_intrinsic_ssbo_atomic_comp_swap; + } else if (strcmp(ir->callee_name(), "__intrinsic_shader_clock") == 0) { + op = nir_intrinsic_shader_clock; + } else if (strcmp(ir->callee_name(), "__intrinsic_group_memory_barrier") == 0) { + op = nir_intrinsic_group_memory_barrier; + } else if (strcmp(ir->callee_name(), "__intrinsic_memory_barrier_atomic_counter") == 0) { + op = nir_intrinsic_memory_barrier_atomic_counter; + } else if (strcmp(ir->callee_name(), "__intrinsic_memory_barrier_buffer") == 0) { + op = nir_intrinsic_memory_barrier_buffer; + } else if (strcmp(ir->callee_name(), "__intrinsic_memory_barrier_image") == 0) { + op = nir_intrinsic_memory_barrier_image; + } else if (strcmp(ir->callee_name(), "__intrinsic_memory_barrier_shared") == 0) { + op = nir_intrinsic_memory_barrier_shared; + } else if (strcmp(ir->callee_name(), "__intrinsic_load_shared") == 0) { + op = nir_intrinsic_load_shared; + } else if (strcmp(ir->callee_name(), "__intrinsic_store_shared") == 0) { + op = nir_intrinsic_store_shared; + } else if (strcmp(ir->callee_name(), "__intrinsic_atomic_add_shared") == 0) { + op = nir_intrinsic_shared_atomic_add; + } else if (strcmp(ir->callee_name(), "__intrinsic_atomic_and_shared") == 0) { + op = nir_intrinsic_shared_atomic_and; + } else if (strcmp(ir->callee_name(), "__intrinsic_atomic_or_shared") == 0) { + op = nir_intrinsic_shared_atomic_or; + } else if (strcmp(ir->callee_name(), "__intrinsic_atomic_xor_shared") == 0) { + op = nir_intrinsic_shared_atomic_xor; + } else if (strcmp(ir->callee_name(), "__intrinsic_atomic_min_shared") == 0) { + assert(ir->return_deref); + if (ir->return_deref->type == glsl_type::int_type) + op = nir_intrinsic_shared_atomic_imin; + else if (ir->return_deref->type == glsl_type::uint_type) + op = nir_intrinsic_shared_atomic_umin; + else + unreachable("Invalid type"); + } else if (strcmp(ir->callee_name(), "__intrinsic_atomic_max_shared") == 0) { + assert(ir->return_deref); + if (ir->return_deref->type == glsl_type::int_type) + op = nir_intrinsic_shared_atomic_imax; + else if (ir->return_deref->type == glsl_type::uint_type) + op = nir_intrinsic_shared_atomic_umax; + else + unreachable("Invalid type"); + } else if (strcmp(ir->callee_name(), "__intrinsic_atomic_exchange_shared") == 0) { + op = nir_intrinsic_shared_atomic_exchange; + } else if (strcmp(ir->callee_name(), "__intrinsic_atomic_comp_swap_shared") == 0) { + op = nir_intrinsic_shared_atomic_comp_swap; + } else { + unreachable("not reached"); + } + + nir_intrinsic_instr *instr = nir_intrinsic_instr_create(shader, op); + nir_dest *dest = &instr->dest; + + switch (op) { + case nir_intrinsic_atomic_counter_read_var: + case nir_intrinsic_atomic_counter_inc_var: + case nir_intrinsic_atomic_counter_dec_var: { + ir_dereference *param = + (ir_dereference *) ir->actual_parameters.get_head(); + instr->variables[0] = evaluate_deref(&instr->instr, param); + nir_ssa_dest_init(&instr->instr, &instr->dest, 1, NULL); + nir_builder_instr_insert(&b, &instr->instr); + break; + } + case nir_intrinsic_image_load: + case nir_intrinsic_image_store: + case nir_intrinsic_image_atomic_add: + case nir_intrinsic_image_atomic_min: + case nir_intrinsic_image_atomic_max: + case nir_intrinsic_image_atomic_and: + case nir_intrinsic_image_atomic_or: + case nir_intrinsic_image_atomic_xor: + case nir_intrinsic_image_atomic_exchange: + case nir_intrinsic_image_atomic_comp_swap: + case nir_intrinsic_image_samples: + case nir_intrinsic_image_size: { + nir_ssa_undef_instr *instr_undef = + nir_ssa_undef_instr_create(shader, 1); + nir_builder_instr_insert(&b, &instr_undef->instr); + + /* Set the image variable dereference. */ + exec_node *param = ir->actual_parameters.get_head(); + ir_dereference *image = (ir_dereference *)param; + const glsl_type *type = + image->variable_referenced()->type->without_array(); + + instr->variables[0] = evaluate_deref(&instr->instr, image); + param = param->get_next(); + + /* Set the intrinsic destination. */ + if (ir->return_deref) { + const nir_intrinsic_info *info = + &nir_intrinsic_infos[instr->intrinsic]; + nir_ssa_dest_init(&instr->instr, &instr->dest, + info->dest_components, NULL); + } + + if (op == nir_intrinsic_image_size || + op == nir_intrinsic_image_samples) { + nir_builder_instr_insert(&b, &instr->instr); + break; + } + + /* Set the address argument, extending the coordinate vector to four + * components. + */ + nir_ssa_def *src_addr = + evaluate_rvalue((ir_dereference *)param); + nir_ssa_def *srcs[4]; + + for (int i = 0; i < 4; i++) { + if (i < type->coordinate_components()) + srcs[i] = nir_channel(&b, src_addr, i); + else + srcs[i] = &instr_undef->def; + } + + instr->src[0] = nir_src_for_ssa(nir_vec(&b, srcs, 4)); + param = param->get_next(); + + /* Set the sample argument, which is undefined for single-sample + * images. + */ + if (type->sampler_dimensionality == GLSL_SAMPLER_DIM_MS) { + instr->src[1] = + nir_src_for_ssa(evaluate_rvalue((ir_dereference *)param)); + param = param->get_next(); + } else { + instr->src[1] = nir_src_for_ssa(&instr_undef->def); + } + + /* Set the intrinsic parameters. */ + if (!param->is_tail_sentinel()) { + instr->src[2] = + nir_src_for_ssa(evaluate_rvalue((ir_dereference *)param)); + param = param->get_next(); + } + + if (!param->is_tail_sentinel()) { + instr->src[3] = + nir_src_for_ssa(evaluate_rvalue((ir_dereference *)param)); + param = param->get_next(); + } + nir_builder_instr_insert(&b, &instr->instr); + break; + } + case nir_intrinsic_memory_barrier: + case nir_intrinsic_group_memory_barrier: + case nir_intrinsic_memory_barrier_atomic_counter: + case nir_intrinsic_memory_barrier_buffer: + case nir_intrinsic_memory_barrier_image: + case nir_intrinsic_memory_barrier_shared: + nir_builder_instr_insert(&b, &instr->instr); + break; + case nir_intrinsic_shader_clock: + nir_ssa_dest_init(&instr->instr, &instr->dest, 1, NULL); + nir_builder_instr_insert(&b, &instr->instr); + break; + case nir_intrinsic_store_ssbo: { + exec_node *param = ir->actual_parameters.get_head(); + ir_rvalue *block = ((ir_instruction *)param)->as_rvalue(); + + param = param->get_next(); + ir_rvalue *offset = ((ir_instruction *)param)->as_rvalue(); + + param = param->get_next(); + ir_rvalue *val = ((ir_instruction *)param)->as_rvalue(); + + param = param->get_next(); + ir_constant *write_mask = ((ir_instruction *)param)->as_constant(); + assert(write_mask); + + instr->src[0] = nir_src_for_ssa(evaluate_rvalue(val)); + instr->src[1] = nir_src_for_ssa(evaluate_rvalue(block)); + instr->src[2] = nir_src_for_ssa(evaluate_rvalue(offset)); + instr->const_index[0] = write_mask->value.u[0]; + instr->num_components = val->type->vector_elements; + + nir_builder_instr_insert(&b, &instr->instr); + break; + } + case nir_intrinsic_load_ssbo: { + exec_node *param = ir->actual_parameters.get_head(); + ir_rvalue *block = ((ir_instruction *)param)->as_rvalue(); + + param = param->get_next(); + ir_rvalue *offset = ((ir_instruction *)param)->as_rvalue(); + + instr->src[0] = nir_src_for_ssa(evaluate_rvalue(block)); + instr->src[1] = nir_src_for_ssa(evaluate_rvalue(offset)); + + const glsl_type *type = ir->return_deref->var->type; + instr->num_components = type->vector_elements; + + /* Setup destination register */ + nir_ssa_dest_init(&instr->instr, &instr->dest, + type->vector_elements, NULL); + + /* Insert the created nir instruction now since in the case of boolean + * result we will need to emit another instruction after it + */ + nir_builder_instr_insert(&b, &instr->instr); + + /* + * In SSBO/UBO's, a true boolean value is any non-zero value, but we + * consider a true boolean to be ~0. Fix this up with a != 0 + * comparison. + */ + if (type->base_type == GLSL_TYPE_BOOL) { + nir_alu_instr *load_ssbo_compare = + nir_alu_instr_create(shader, nir_op_ine); + load_ssbo_compare->src[0].src.is_ssa = true; + load_ssbo_compare->src[0].src.ssa = &instr->dest.ssa; + load_ssbo_compare->src[1].src = + nir_src_for_ssa(nir_imm_int(&b, 0)); + for (unsigned i = 0; i < type->vector_elements; i++) + load_ssbo_compare->src[1].swizzle[i] = 0; + nir_ssa_dest_init(&load_ssbo_compare->instr, + &load_ssbo_compare->dest.dest, + type->vector_elements, NULL); + load_ssbo_compare->dest.write_mask = (1 << type->vector_elements) - 1; + nir_builder_instr_insert(&b, &load_ssbo_compare->instr); + dest = &load_ssbo_compare->dest.dest; + } + break; + } + case nir_intrinsic_ssbo_atomic_add: + case nir_intrinsic_ssbo_atomic_imin: + case nir_intrinsic_ssbo_atomic_umin: + case nir_intrinsic_ssbo_atomic_imax: + case nir_intrinsic_ssbo_atomic_umax: + case nir_intrinsic_ssbo_atomic_and: + case nir_intrinsic_ssbo_atomic_or: + case nir_intrinsic_ssbo_atomic_xor: + case nir_intrinsic_ssbo_atomic_exchange: + case nir_intrinsic_ssbo_atomic_comp_swap: { + int param_count = ir->actual_parameters.length(); + assert(param_count == 3 || param_count == 4); + + /* Block index */ + exec_node *param = ir->actual_parameters.get_head(); + ir_instruction *inst = (ir_instruction *) param; + instr->src[0] = nir_src_for_ssa(evaluate_rvalue(inst->as_rvalue())); + + /* Offset */ + param = param->get_next(); + inst = (ir_instruction *) param; + instr->src[1] = nir_src_for_ssa(evaluate_rvalue(inst->as_rvalue())); + + /* data1 parameter (this is always present) */ + param = param->get_next(); + inst = (ir_instruction *) param; + instr->src[2] = nir_src_for_ssa(evaluate_rvalue(inst->as_rvalue())); + + /* data2 parameter (only with atomic_comp_swap) */ + if (param_count == 4) { + assert(op == nir_intrinsic_ssbo_atomic_comp_swap); + param = param->get_next(); + inst = (ir_instruction *) param; + instr->src[3] = nir_src_for_ssa(evaluate_rvalue(inst->as_rvalue())); + } + + /* Atomic result */ + assert(ir->return_deref); + nir_ssa_dest_init(&instr->instr, &instr->dest, + ir->return_deref->type->vector_elements, NULL); + nir_builder_instr_insert(&b, &instr->instr); + break; + } + case nir_intrinsic_load_shared: { + exec_node *param = ir->actual_parameters.get_head(); + ir_rvalue *offset = ((ir_instruction *)param)->as_rvalue(); + + instr->const_index[0] = 0; + instr->src[0] = nir_src_for_ssa(evaluate_rvalue(offset)); + + const glsl_type *type = ir->return_deref->var->type; + instr->num_components = type->vector_elements; + + /* Setup destination register */ + nir_ssa_dest_init(&instr->instr, &instr->dest, + type->vector_elements, NULL); + + nir_builder_instr_insert(&b, &instr->instr); + break; + } + case nir_intrinsic_store_shared: { + exec_node *param = ir->actual_parameters.get_head(); + ir_rvalue *offset = ((ir_instruction *)param)->as_rvalue(); + + param = param->get_next(); + ir_rvalue *val = ((ir_instruction *)param)->as_rvalue(); + + param = param->get_next(); + ir_constant *write_mask = ((ir_instruction *)param)->as_constant(); + assert(write_mask); + + instr->const_index[0] = 0; + instr->src[1] = nir_src_for_ssa(evaluate_rvalue(offset)); + + instr->const_index[1] = write_mask->value.u[0]; + + instr->src[0] = nir_src_for_ssa(evaluate_rvalue(val)); + instr->num_components = val->type->vector_elements; + + nir_builder_instr_insert(&b, &instr->instr); + break; + } + case nir_intrinsic_shared_atomic_add: + case nir_intrinsic_shared_atomic_imin: + case nir_intrinsic_shared_atomic_umin: + case nir_intrinsic_shared_atomic_imax: + case nir_intrinsic_shared_atomic_umax: + case nir_intrinsic_shared_atomic_and: + case nir_intrinsic_shared_atomic_or: + case nir_intrinsic_shared_atomic_xor: + case nir_intrinsic_shared_atomic_exchange: + case nir_intrinsic_shared_atomic_comp_swap: { + int param_count = ir->actual_parameters.length(); + assert(param_count == 2 || param_count == 3); + + /* Offset */ + exec_node *param = ir->actual_parameters.get_head(); + ir_instruction *inst = (ir_instruction *) param; + instr->src[0] = nir_src_for_ssa(evaluate_rvalue(inst->as_rvalue())); + + /* data1 parameter (this is always present) */ + param = param->get_next(); + inst = (ir_instruction *) param; + instr->src[1] = nir_src_for_ssa(evaluate_rvalue(inst->as_rvalue())); + + /* data2 parameter (only with atomic_comp_swap) */ + if (param_count == 3) { + assert(op == nir_intrinsic_shared_atomic_comp_swap); + param = param->get_next(); + inst = (ir_instruction *) param; + instr->src[2] = + nir_src_for_ssa(evaluate_rvalue(inst->as_rvalue())); + } + + /* Atomic result */ + assert(ir->return_deref); + nir_ssa_dest_init(&instr->instr, &instr->dest, + ir->return_deref->type->vector_elements, NULL); + nir_builder_instr_insert(&b, &instr->instr); + break; + } + default: + unreachable("not reached"); + } + + if (ir->return_deref) { + nir_intrinsic_instr *store_instr = + nir_intrinsic_instr_create(shader, nir_intrinsic_store_var); + store_instr->num_components = ir->return_deref->type->vector_elements; + store_instr->const_index[0] = (1 << store_instr->num_components) - 1; + + store_instr->variables[0] = + evaluate_deref(&store_instr->instr, ir->return_deref); + store_instr->src[0] = nir_src_for_ssa(&dest->ssa); + + nir_builder_instr_insert(&b, &store_instr->instr); + } + + return; + } + + struct hash_entry *entry = + _mesa_hash_table_search(this->overload_table, ir->callee); + assert(entry); + nir_function *callee = (nir_function *) entry->data; + + nir_call_instr *instr = nir_call_instr_create(this->shader, callee); + + unsigned i = 0; + foreach_in_list(ir_dereference, param, &ir->actual_parameters) { + instr->params[i] = evaluate_deref(&instr->instr, param); + i++; + } + + instr->return_deref = evaluate_deref(&instr->instr, ir->return_deref); + nir_builder_instr_insert(&b, &instr->instr); +} + +void +nir_visitor::visit(ir_assignment *ir) +{ + unsigned num_components = ir->lhs->type->vector_elements; + + if ((ir->rhs->as_dereference() || ir->rhs->as_constant()) && + (ir->write_mask == (1 << num_components) - 1 || ir->write_mask == 0)) { + /* We're doing a plain-as-can-be copy, so emit a copy_var */ + nir_intrinsic_instr *copy = + nir_intrinsic_instr_create(this->shader, nir_intrinsic_copy_var); + + copy->variables[0] = evaluate_deref(©->instr, ir->lhs); + copy->variables[1] = evaluate_deref(©->instr, ir->rhs); + + if (ir->condition) { + nir_if *if_stmt = nir_if_create(this->shader); + if_stmt->condition = nir_src_for_ssa(evaluate_rvalue(ir->condition)); + nir_builder_cf_insert(&b, &if_stmt->cf_node); + nir_instr_insert_after_cf_list(&if_stmt->then_list, ©->instr); + b.cursor = nir_after_cf_node(&if_stmt->cf_node); + } else { + nir_builder_instr_insert(&b, ©->instr); + } + return; + } + + assert(ir->rhs->type->is_scalar() || ir->rhs->type->is_vector()); + + ir->lhs->accept(this); + nir_deref_var *lhs_deref = this->deref_head; + nir_ssa_def *src = evaluate_rvalue(ir->rhs); + + if (ir->write_mask != (1 << num_components) - 1 && ir->write_mask != 0) { + /* GLSL IR will give us the input to the write-masked assignment in a + * single packed vector. So, for example, if the writemask is xzw, then + * we have to swizzle x -> x, y -> z, and z -> w and get the y component + * from the load. + */ + unsigned swiz[4]; + unsigned component = 0; + for (unsigned i = 0; i < 4; i++) { + swiz[i] = ir->write_mask & (1 << i) ? component++ : 0; + } + src = nir_swizzle(&b, src, swiz, num_components, !supports_ints); + } + + nir_intrinsic_instr *store = + nir_intrinsic_instr_create(this->shader, nir_intrinsic_store_var); + store->num_components = ir->lhs->type->vector_elements; + store->const_index[0] = ir->write_mask; + nir_deref *store_deref = nir_copy_deref(store, &lhs_deref->deref); + store->variables[0] = nir_deref_as_var(store_deref); + store->src[0] = nir_src_for_ssa(src); + + if (ir->condition) { + nir_if *if_stmt = nir_if_create(this->shader); + if_stmt->condition = nir_src_for_ssa(evaluate_rvalue(ir->condition)); + nir_builder_cf_insert(&b, &if_stmt->cf_node); + nir_instr_insert_after_cf_list(&if_stmt->then_list, &store->instr); + b.cursor = nir_after_cf_node(&if_stmt->cf_node); + } else { + nir_builder_instr_insert(&b, &store->instr); + } +} + +/* + * Given an instruction, returns a pointer to its destination or NULL if there + * is no destination. + * + * Note that this only handles instructions we generate at this level. + */ +static nir_dest * +get_instr_dest(nir_instr *instr) +{ + nir_alu_instr *alu_instr; + nir_intrinsic_instr *intrinsic_instr; + nir_tex_instr *tex_instr; + + switch (instr->type) { + case nir_instr_type_alu: + alu_instr = nir_instr_as_alu(instr); + return &alu_instr->dest.dest; + + case nir_instr_type_intrinsic: + intrinsic_instr = nir_instr_as_intrinsic(instr); + if (nir_intrinsic_infos[intrinsic_instr->intrinsic].has_dest) + return &intrinsic_instr->dest; + else + return NULL; + + case nir_instr_type_tex: + tex_instr = nir_instr_as_tex(instr); + return &tex_instr->dest; + + default: + unreachable("not reached"); + } + + return NULL; +} + +void +nir_visitor::add_instr(nir_instr *instr, unsigned num_components) +{ + nir_dest *dest = get_instr_dest(instr); + + if (dest) + nir_ssa_dest_init(instr, dest, num_components, NULL); + + nir_builder_instr_insert(&b, instr); + + if (dest) { + assert(dest->is_ssa); + this->result = &dest->ssa; + } +} + +nir_ssa_def * +nir_visitor::evaluate_rvalue(ir_rvalue* ir) +{ + ir->accept(this); + if (ir->as_dereference() || ir->as_constant()) { + /* + * A dereference is being used on the right hand side, which means we + * must emit a variable load. + */ + + nir_intrinsic_instr *load_instr = + nir_intrinsic_instr_create(this->shader, nir_intrinsic_load_var); + load_instr->num_components = ir->type->vector_elements; + load_instr->variables[0] = this->deref_head; + ralloc_steal(load_instr, load_instr->variables[0]); + add_instr(&load_instr->instr, ir->type->vector_elements); + } + + return this->result; +} + +void +nir_visitor::visit(ir_expression *ir) +{ + /* Some special cases */ + switch (ir->operation) { + case ir_binop_ubo_load: { + nir_intrinsic_instr *load = + nir_intrinsic_instr_create(this->shader, nir_intrinsic_load_ubo); + load->num_components = ir->type->vector_elements; + load->src[0] = nir_src_for_ssa(evaluate_rvalue(ir->operands[0])); + load->src[1] = nir_src_for_ssa(evaluate_rvalue(ir->operands[1])); + add_instr(&load->instr, ir->type->vector_elements); + + /* + * In UBO's, a true boolean value is any non-zero value, but we consider + * a true boolean to be ~0. Fix this up with a != 0 comparison. + */ + + if (ir->type->base_type == GLSL_TYPE_BOOL) + this->result = nir_ine(&b, &load->dest.ssa, nir_imm_int(&b, 0)); + + return; + } + + case ir_unop_interpolate_at_centroid: + case ir_binop_interpolate_at_offset: + case ir_binop_interpolate_at_sample: { + ir_dereference *deref = ir->operands[0]->as_dereference(); + ir_swizzle *swizzle = NULL; + if (!deref) { + /* the api does not allow a swizzle here, but the varying packing code + * may have pushed one into here. + */ + swizzle = ir->operands[0]->as_swizzle(); + assert(swizzle); + deref = swizzle->val->as_dereference(); + assert(deref); + } + + deref->accept(this); + + nir_intrinsic_op op; + if (this->deref_head->var->data.mode == nir_var_shader_in) { + switch (ir->operation) { + case ir_unop_interpolate_at_centroid: + op = nir_intrinsic_interp_var_at_centroid; + break; + case ir_binop_interpolate_at_offset: + op = nir_intrinsic_interp_var_at_offset; + break; + case ir_binop_interpolate_at_sample: + op = nir_intrinsic_interp_var_at_sample; + break; + default: + unreachable("Invalid interpolation intrinsic"); + } + } else { + /* This case can happen if the vertex shader does not write the + * given varying. In this case, the linker will lower it to a + * global variable. Since interpolating a variable makes no + * sense, we'll just turn it into a load which will probably + * eventually end up as an SSA definition. + */ + assert(this->deref_head->var->data.mode == nir_var_global); + op = nir_intrinsic_load_var; + } + + nir_intrinsic_instr *intrin = nir_intrinsic_instr_create(shader, op); + intrin->num_components = deref->type->vector_elements; + intrin->variables[0] = this->deref_head; + ralloc_steal(intrin, intrin->variables[0]); + + if (intrin->intrinsic == nir_intrinsic_interp_var_at_offset || + intrin->intrinsic == nir_intrinsic_interp_var_at_sample) + intrin->src[0] = nir_src_for_ssa(evaluate_rvalue(ir->operands[1])); + + add_instr(&intrin->instr, deref->type->vector_elements); + + if (swizzle) { + unsigned swiz[4] = { + swizzle->mask.x, swizzle->mask.y, swizzle->mask.z, swizzle->mask.w + }; + + result = nir_swizzle(&b, result, swiz, + swizzle->type->vector_elements, false); + } + + return; + } + + default: + break; + } + + nir_ssa_def *srcs[4]; + for (unsigned i = 0; i < ir->get_num_operands(); i++) + srcs[i] = evaluate_rvalue(ir->operands[i]); + + glsl_base_type types[4]; + for (unsigned i = 0; i < ir->get_num_operands(); i++) + if (supports_ints) + types[i] = ir->operands[i]->type->base_type; + else + types[i] = GLSL_TYPE_FLOAT; + + glsl_base_type out_type; + if (supports_ints) + out_type = ir->type->base_type; + else + out_type = GLSL_TYPE_FLOAT; + + switch (ir->operation) { + case ir_unop_bit_not: result = nir_inot(&b, srcs[0]); break; + case ir_unop_logic_not: + result = supports_ints ? nir_inot(&b, srcs[0]) : nir_fnot(&b, srcs[0]); + break; + case ir_unop_neg: + result = (types[0] == GLSL_TYPE_FLOAT) ? nir_fneg(&b, srcs[0]) + : nir_ineg(&b, srcs[0]); + break; + case ir_unop_abs: + result = (types[0] == GLSL_TYPE_FLOAT) ? nir_fabs(&b, srcs[0]) + : nir_iabs(&b, srcs[0]); + break; + case ir_unop_saturate: + assert(types[0] == GLSL_TYPE_FLOAT); + result = nir_fsat(&b, srcs[0]); + break; + case ir_unop_sign: + result = (types[0] == GLSL_TYPE_FLOAT) ? nir_fsign(&b, srcs[0]) + : nir_isign(&b, srcs[0]); + break; + case ir_unop_rcp: result = nir_frcp(&b, srcs[0]); break; + case ir_unop_rsq: result = nir_frsq(&b, srcs[0]); break; + case ir_unop_sqrt: result = nir_fsqrt(&b, srcs[0]); break; + case ir_unop_exp: unreachable("ir_unop_exp should have been lowered"); + case ir_unop_log: unreachable("ir_unop_log should have been lowered"); + case ir_unop_exp2: result = nir_fexp2(&b, srcs[0]); break; + case ir_unop_log2: result = nir_flog2(&b, srcs[0]); break; + case ir_unop_i2f: + result = supports_ints ? nir_i2f(&b, srcs[0]) : nir_fmov(&b, srcs[0]); + break; + case ir_unop_u2f: + result = supports_ints ? nir_u2f(&b, srcs[0]) : nir_fmov(&b, srcs[0]); + break; + case ir_unop_b2f: + result = supports_ints ? nir_b2f(&b, srcs[0]) : nir_fmov(&b, srcs[0]); + break; + case ir_unop_f2i: result = nir_f2i(&b, srcs[0]); break; + case ir_unop_f2u: result = nir_f2u(&b, srcs[0]); break; + case ir_unop_f2b: result = nir_f2b(&b, srcs[0]); break; + case ir_unop_i2b: result = nir_i2b(&b, srcs[0]); break; + case ir_unop_b2i: result = nir_b2i(&b, srcs[0]); break; + case ir_unop_i2u: + case ir_unop_u2i: + case ir_unop_bitcast_i2f: + case ir_unop_bitcast_f2i: + case ir_unop_bitcast_u2f: + case ir_unop_bitcast_f2u: + case ir_unop_subroutine_to_int: + /* no-op */ + result = nir_imov(&b, srcs[0]); + break; + case ir_unop_trunc: result = nir_ftrunc(&b, srcs[0]); break; + case ir_unop_ceil: result = nir_fceil(&b, srcs[0]); break; + case ir_unop_floor: result = nir_ffloor(&b, srcs[0]); break; + case ir_unop_fract: result = nir_ffract(&b, srcs[0]); break; + case ir_unop_round_even: result = nir_fround_even(&b, srcs[0]); break; + case ir_unop_sin: result = nir_fsin(&b, srcs[0]); break; + case ir_unop_cos: result = nir_fcos(&b, srcs[0]); break; + case ir_unop_dFdx: result = nir_fddx(&b, srcs[0]); break; + case ir_unop_dFdy: result = nir_fddy(&b, srcs[0]); break; + case ir_unop_dFdx_fine: result = nir_fddx_fine(&b, srcs[0]); break; + case ir_unop_dFdy_fine: result = nir_fddy_fine(&b, srcs[0]); break; + case ir_unop_dFdx_coarse: result = nir_fddx_coarse(&b, srcs[0]); break; + case ir_unop_dFdy_coarse: result = nir_fddy_coarse(&b, srcs[0]); break; + case ir_unop_pack_snorm_2x16: + result = nir_pack_snorm_2x16(&b, srcs[0]); + break; + case ir_unop_pack_snorm_4x8: + result = nir_pack_snorm_4x8(&b, srcs[0]); + break; + case ir_unop_pack_unorm_2x16: + result = nir_pack_unorm_2x16(&b, srcs[0]); + break; + case ir_unop_pack_unorm_4x8: + result = nir_pack_unorm_4x8(&b, srcs[0]); + break; + case ir_unop_pack_half_2x16: + result = nir_pack_half_2x16(&b, srcs[0]); + break; + case ir_unop_unpack_snorm_2x16: + result = nir_unpack_snorm_2x16(&b, srcs[0]); + break; + case ir_unop_unpack_snorm_4x8: + result = nir_unpack_snorm_4x8(&b, srcs[0]); + break; + case ir_unop_unpack_unorm_2x16: + result = nir_unpack_unorm_2x16(&b, srcs[0]); + break; + case ir_unop_unpack_unorm_4x8: + result = nir_unpack_unorm_4x8(&b, srcs[0]); + break; + case ir_unop_unpack_half_2x16: + result = nir_unpack_half_2x16(&b, srcs[0]); + break; + case ir_unop_unpack_half_2x16_split_x: + result = nir_unpack_half_2x16_split_x(&b, srcs[0]); + break; + case ir_unop_unpack_half_2x16_split_y: + result = nir_unpack_half_2x16_split_y(&b, srcs[0]); + break; + case ir_unop_bitfield_reverse: + result = nir_bitfield_reverse(&b, srcs[0]); + break; + case ir_unop_bit_count: + result = nir_bit_count(&b, srcs[0]); + break; + case ir_unop_find_msb: + switch (types[0]) { + case GLSL_TYPE_UINT: + result = nir_ufind_msb(&b, srcs[0]); + break; + case GLSL_TYPE_INT: + result = nir_ifind_msb(&b, srcs[0]); + break; + default: + unreachable("Invalid type for findMSB()"); + } + break; + case ir_unop_find_lsb: + result = nir_find_lsb(&b, srcs[0]); + break; + + case ir_unop_noise: + switch (ir->type->vector_elements) { + case 1: + switch (ir->operands[0]->type->vector_elements) { + case 1: result = nir_fnoise1_1(&b, srcs[0]); break; + case 2: result = nir_fnoise1_2(&b, srcs[0]); break; + case 3: result = nir_fnoise1_3(&b, srcs[0]); break; + case 4: result = nir_fnoise1_4(&b, srcs[0]); break; + default: unreachable("not reached"); + } + break; + case 2: + switch (ir->operands[0]->type->vector_elements) { + case 1: result = nir_fnoise2_1(&b, srcs[0]); break; + case 2: result = nir_fnoise2_2(&b, srcs[0]); break; + case 3: result = nir_fnoise2_3(&b, srcs[0]); break; + case 4: result = nir_fnoise2_4(&b, srcs[0]); break; + default: unreachable("not reached"); + } + break; + case 3: + switch (ir->operands[0]->type->vector_elements) { + case 1: result = nir_fnoise3_1(&b, srcs[0]); break; + case 2: result = nir_fnoise3_2(&b, srcs[0]); break; + case 3: result = nir_fnoise3_3(&b, srcs[0]); break; + case 4: result = nir_fnoise3_4(&b, srcs[0]); break; + default: unreachable("not reached"); + } + break; + case 4: + switch (ir->operands[0]->type->vector_elements) { + case 1: result = nir_fnoise4_1(&b, srcs[0]); break; + case 2: result = nir_fnoise4_2(&b, srcs[0]); break; + case 3: result = nir_fnoise4_3(&b, srcs[0]); break; + case 4: result = nir_fnoise4_4(&b, srcs[0]); break; + default: unreachable("not reached"); + } + break; + default: + unreachable("not reached"); + } + break; + case ir_unop_get_buffer_size: { + nir_intrinsic_instr *load = nir_intrinsic_instr_create( + this->shader, + nir_intrinsic_get_buffer_size); + load->num_components = ir->type->vector_elements; + load->src[0] = nir_src_for_ssa(evaluate_rvalue(ir->operands[0])); + add_instr(&load->instr, ir->type->vector_elements); + return; + } + + case ir_binop_add: + result = (out_type == GLSL_TYPE_FLOAT) ? nir_fadd(&b, srcs[0], srcs[1]) + : nir_iadd(&b, srcs[0], srcs[1]); + break; + case ir_binop_sub: + result = (out_type == GLSL_TYPE_FLOAT) ? nir_fsub(&b, srcs[0], srcs[1]) + : nir_isub(&b, srcs[0], srcs[1]); + break; + case ir_binop_mul: + result = (out_type == GLSL_TYPE_FLOAT) ? nir_fmul(&b, srcs[0], srcs[1]) + : nir_imul(&b, srcs[0], srcs[1]); + break; + case ir_binop_div: + if (out_type == GLSL_TYPE_FLOAT) + result = nir_fdiv(&b, srcs[0], srcs[1]); + else if (out_type == GLSL_TYPE_INT) + result = nir_idiv(&b, srcs[0], srcs[1]); + else + result = nir_udiv(&b, srcs[0], srcs[1]); + break; + case ir_binop_mod: + result = (out_type == GLSL_TYPE_FLOAT) ? nir_fmod(&b, srcs[0], srcs[1]) + : nir_umod(&b, srcs[0], srcs[1]); + break; + case ir_binop_min: + if (out_type == GLSL_TYPE_FLOAT) + result = nir_fmin(&b, srcs[0], srcs[1]); + else if (out_type == GLSL_TYPE_INT) + result = nir_imin(&b, srcs[0], srcs[1]); + else + result = nir_umin(&b, srcs[0], srcs[1]); + break; + case ir_binop_max: + if (out_type == GLSL_TYPE_FLOAT) + result = nir_fmax(&b, srcs[0], srcs[1]); + else if (out_type == GLSL_TYPE_INT) + result = nir_imax(&b, srcs[0], srcs[1]); + else + result = nir_umax(&b, srcs[0], srcs[1]); + break; + case ir_binop_pow: result = nir_fpow(&b, srcs[0], srcs[1]); break; + case ir_binop_bit_and: result = nir_iand(&b, srcs[0], srcs[1]); break; + case ir_binop_bit_or: result = nir_ior(&b, srcs[0], srcs[1]); break; + case ir_binop_bit_xor: result = nir_ixor(&b, srcs[0], srcs[1]); break; + case ir_binop_logic_and: + result = supports_ints ? nir_iand(&b, srcs[0], srcs[1]) + : nir_fand(&b, srcs[0], srcs[1]); + break; + case ir_binop_logic_or: + result = supports_ints ? nir_ior(&b, srcs[0], srcs[1]) + : nir_for(&b, srcs[0], srcs[1]); + break; + case ir_binop_logic_xor: + result = supports_ints ? nir_ixor(&b, srcs[0], srcs[1]) + : nir_fxor(&b, srcs[0], srcs[1]); + break; + case ir_binop_lshift: result = nir_ishl(&b, srcs[0], srcs[1]); break; + case ir_binop_rshift: + result = (out_type == GLSL_TYPE_INT) ? nir_ishr(&b, srcs[0], srcs[1]) + : nir_ushr(&b, srcs[0], srcs[1]); + break; + case ir_binop_imul_high: + result = (out_type == GLSL_TYPE_INT) ? nir_imul_high(&b, srcs[0], srcs[1]) + : nir_umul_high(&b, srcs[0], srcs[1]); + break; + case ir_binop_carry: result = nir_uadd_carry(&b, srcs[0], srcs[1]); break; + case ir_binop_borrow: result = nir_usub_borrow(&b, srcs[0], srcs[1]); break; + case ir_binop_less: + if (supports_ints) { + if (types[0] == GLSL_TYPE_FLOAT) + result = nir_flt(&b, srcs[0], srcs[1]); + else if (types[0] == GLSL_TYPE_INT) + result = nir_ilt(&b, srcs[0], srcs[1]); + else + result = nir_ult(&b, srcs[0], srcs[1]); + } else { + result = nir_slt(&b, srcs[0], srcs[1]); + } + break; + case ir_binop_greater: + if (supports_ints) { + if (types[0] == GLSL_TYPE_FLOAT) + result = nir_flt(&b, srcs[1], srcs[0]); + else if (types[0] == GLSL_TYPE_INT) + result = nir_ilt(&b, srcs[1], srcs[0]); + else + result = nir_ult(&b, srcs[1], srcs[0]); + } else { + result = nir_slt(&b, srcs[1], srcs[0]); + } + break; + case ir_binop_lequal: + if (supports_ints) { + if (types[0] == GLSL_TYPE_FLOAT) + result = nir_fge(&b, srcs[1], srcs[0]); + else if (types[0] == GLSL_TYPE_INT) + result = nir_ige(&b, srcs[1], srcs[0]); + else + result = nir_uge(&b, srcs[1], srcs[0]); + } else { + result = nir_slt(&b, srcs[1], srcs[0]); + } + break; + case ir_binop_gequal: + if (supports_ints) { + if (types[0] == GLSL_TYPE_FLOAT) + result = nir_fge(&b, srcs[0], srcs[1]); + else if (types[0] == GLSL_TYPE_INT) + result = nir_ige(&b, srcs[0], srcs[1]); + else + result = nir_uge(&b, srcs[0], srcs[1]); + } else { + result = nir_slt(&b, srcs[0], srcs[1]); + } + break; + case ir_binop_equal: + if (supports_ints) { + if (types[0] == GLSL_TYPE_FLOAT) + result = nir_feq(&b, srcs[0], srcs[1]); + else + result = nir_ieq(&b, srcs[0], srcs[1]); + } else { + result = nir_seq(&b, srcs[0], srcs[1]); + } + break; + case ir_binop_nequal: + if (supports_ints) { + if (types[0] == GLSL_TYPE_FLOAT) + result = nir_fne(&b, srcs[0], srcs[1]); + else + result = nir_ine(&b, srcs[0], srcs[1]); + } else { + result = nir_sne(&b, srcs[0], srcs[1]); + } + break; + case ir_binop_all_equal: + if (supports_ints) { + if (types[0] == GLSL_TYPE_FLOAT) { + switch (ir->operands[0]->type->vector_elements) { + case 1: result = nir_feq(&b, srcs[0], srcs[1]); break; + case 2: result = nir_ball_fequal2(&b, srcs[0], srcs[1]); break; + case 3: result = nir_ball_fequal3(&b, srcs[0], srcs[1]); break; + case 4: result = nir_ball_fequal4(&b, srcs[0], srcs[1]); break; + default: + unreachable("not reached"); + } + } else { + switch (ir->operands[0]->type->vector_elements) { + case 1: result = nir_ieq(&b, srcs[0], srcs[1]); break; + case 2: result = nir_ball_iequal2(&b, srcs[0], srcs[1]); break; + case 3: result = nir_ball_iequal3(&b, srcs[0], srcs[1]); break; + case 4: result = nir_ball_iequal4(&b, srcs[0], srcs[1]); break; + default: + unreachable("not reached"); + } + } + } else { + switch (ir->operands[0]->type->vector_elements) { + case 1: result = nir_seq(&b, srcs[0], srcs[1]); break; + case 2: result = nir_fall_equal2(&b, srcs[0], srcs[1]); break; + case 3: result = nir_fall_equal3(&b, srcs[0], srcs[1]); break; + case 4: result = nir_fall_equal4(&b, srcs[0], srcs[1]); break; + default: + unreachable("not reached"); + } + } + break; + case ir_binop_any_nequal: + if (supports_ints) { + if (types[0] == GLSL_TYPE_FLOAT) { + switch (ir->operands[0]->type->vector_elements) { + case 1: result = nir_fne(&b, srcs[0], srcs[1]); break; + case 2: result = nir_bany_fnequal2(&b, srcs[0], srcs[1]); break; + case 3: result = nir_bany_fnequal3(&b, srcs[0], srcs[1]); break; + case 4: result = nir_bany_fnequal4(&b, srcs[0], srcs[1]); break; + default: + unreachable("not reached"); + } + } else { + switch (ir->operands[0]->type->vector_elements) { + case 1: result = nir_ine(&b, srcs[0], srcs[1]); break; + case 2: result = nir_bany_inequal2(&b, srcs[0], srcs[1]); break; + case 3: result = nir_bany_inequal3(&b, srcs[0], srcs[1]); break; + case 4: result = nir_bany_inequal4(&b, srcs[0], srcs[1]); break; + default: + unreachable("not reached"); + } + } + } else { + switch (ir->operands[0]->type->vector_elements) { + case 1: result = nir_sne(&b, srcs[0], srcs[1]); break; + case 2: result = nir_fany_nequal2(&b, srcs[0], srcs[1]); break; + case 3: result = nir_fany_nequal3(&b, srcs[0], srcs[1]); break; + case 4: result = nir_fany_nequal4(&b, srcs[0], srcs[1]); break; + default: + unreachable("not reached"); + } + } + break; + case ir_binop_dot: + switch (ir->operands[0]->type->vector_elements) { + case 2: result = nir_fdot2(&b, srcs[0], srcs[1]); break; + case 3: result = nir_fdot3(&b, srcs[0], srcs[1]); break; + case 4: result = nir_fdot4(&b, srcs[0], srcs[1]); break; + default: + unreachable("not reached"); + } + break; + + case ir_binop_pack_half_2x16_split: + result = nir_pack_half_2x16_split(&b, srcs[0], srcs[1]); + break; + case ir_binop_ldexp: result = nir_ldexp(&b, srcs[0], srcs[1]); break; + case ir_triop_fma: + result = nir_ffma(&b, srcs[0], srcs[1], srcs[2]); + break; + case ir_triop_lrp: + result = nir_flrp(&b, srcs[0], srcs[1], srcs[2]); + break; + case ir_triop_csel: + if (supports_ints) + result = nir_bcsel(&b, srcs[0], srcs[1], srcs[2]); + else + result = nir_fcsel(&b, srcs[0], srcs[1], srcs[2]); + break; + case ir_triop_bitfield_extract: + result = (out_type == GLSL_TYPE_INT) ? + nir_ibitfield_extract(&b, srcs[0], srcs[1], srcs[2]) : + nir_ubitfield_extract(&b, srcs[0], srcs[1], srcs[2]); + break; + case ir_quadop_bitfield_insert: + result = nir_bitfield_insert(&b, srcs[0], srcs[1], srcs[2], srcs[3]); + break; + case ir_quadop_vector: + result = nir_vec(&b, srcs, ir->type->vector_elements); + break; + + default: + unreachable("not reached"); + } +} + +void +nir_visitor::visit(ir_swizzle *ir) +{ + unsigned swizzle[4] = { ir->mask.x, ir->mask.y, ir->mask.z, ir->mask.w }; + result = nir_swizzle(&b, evaluate_rvalue(ir->val), swizzle, + ir->type->vector_elements, !supports_ints); +} + +void +nir_visitor::visit(ir_texture *ir) +{ + unsigned num_srcs; + nir_texop op; + switch (ir->op) { + case ir_tex: + op = nir_texop_tex; + num_srcs = 1; /* coordinate */ + break; + + case ir_txb: + case ir_txl: + op = (ir->op == ir_txb) ? nir_texop_txb : nir_texop_txl; + num_srcs = 2; /* coordinate, bias/lod */ + break; + + case ir_txd: + op = nir_texop_txd; /* coordinate, dPdx, dPdy */ + num_srcs = 3; + break; + + case ir_txf: + op = nir_texop_txf; + if (ir->lod_info.lod != NULL) + num_srcs = 2; /* coordinate, lod */ + else + num_srcs = 1; /* coordinate */ + break; + + case ir_txf_ms: + op = nir_texop_txf_ms; + num_srcs = 2; /* coordinate, sample_index */ + break; + + case ir_txs: + op = nir_texop_txs; + if (ir->lod_info.lod != NULL) + num_srcs = 1; /* lod */ + else + num_srcs = 0; + break; + + case ir_lod: + op = nir_texop_lod; + num_srcs = 1; /* coordinate */ + break; + + case ir_tg4: + op = nir_texop_tg4; + num_srcs = 1; /* coordinate */ + break; + + case ir_query_levels: + op = nir_texop_query_levels; + num_srcs = 0; + break; + + case ir_texture_samples: + op = nir_texop_texture_samples; + num_srcs = 0; + break; + + case ir_samples_identical: + op = nir_texop_samples_identical; + num_srcs = 1; /* coordinate */ + break; + + default: + unreachable("not reached"); + } + + if (ir->projector != NULL) + num_srcs++; + if (ir->shadow_comparitor != NULL) + num_srcs++; + if (ir->offset != NULL && ir->offset->as_constant() == NULL) + num_srcs++; + + nir_tex_instr *instr = nir_tex_instr_create(this->shader, num_srcs); + + instr->op = op; + instr->sampler_dim = + (glsl_sampler_dim) ir->sampler->type->sampler_dimensionality; + instr->is_array = ir->sampler->type->sampler_array; + instr->is_shadow = ir->sampler->type->sampler_shadow; + if (instr->is_shadow) + instr->is_new_style_shadow = (ir->type->vector_elements == 1); + switch (ir->type->base_type) { + case GLSL_TYPE_FLOAT: + instr->dest_type = nir_type_float; + break; + case GLSL_TYPE_INT: + instr->dest_type = nir_type_int; + break; + case GLSL_TYPE_BOOL: + case GLSL_TYPE_UINT: + instr->dest_type = nir_type_uint; + break; + default: + unreachable("not reached"); + } + + instr->sampler = evaluate_deref(&instr->instr, ir->sampler); + + unsigned src_number = 0; + + if (ir->coordinate != NULL) { + instr->coord_components = ir->coordinate->type->vector_elements; + instr->src[src_number].src = + nir_src_for_ssa(evaluate_rvalue(ir->coordinate)); + instr->src[src_number].src_type = nir_tex_src_coord; + src_number++; + } + + if (ir->projector != NULL) { + instr->src[src_number].src = + nir_src_for_ssa(evaluate_rvalue(ir->projector)); + instr->src[src_number].src_type = nir_tex_src_projector; + src_number++; + } + + if (ir->shadow_comparitor != NULL) { + instr->src[src_number].src = + nir_src_for_ssa(evaluate_rvalue(ir->shadow_comparitor)); + instr->src[src_number].src_type = nir_tex_src_comparitor; + src_number++; + } + + if (ir->offset != NULL) { + /* we don't support multiple offsets yet */ + assert(ir->offset->type->is_vector() || ir->offset->type->is_scalar()); + + ir_constant *const_offset = ir->offset->as_constant(); + if (const_offset != NULL) { + for (unsigned i = 0; i < const_offset->type->vector_elements; i++) + instr->const_offset[i] = const_offset->value.i[i]; + } else { + instr->src[src_number].src = + nir_src_for_ssa(evaluate_rvalue(ir->offset)); + instr->src[src_number].src_type = nir_tex_src_offset; + src_number++; + } + } + + switch (ir->op) { + case ir_txb: + instr->src[src_number].src = + nir_src_for_ssa(evaluate_rvalue(ir->lod_info.bias)); + instr->src[src_number].src_type = nir_tex_src_bias; + src_number++; + break; + + case ir_txl: + case ir_txf: + case ir_txs: + if (ir->lod_info.lod != NULL) { + instr->src[src_number].src = + nir_src_for_ssa(evaluate_rvalue(ir->lod_info.lod)); + instr->src[src_number].src_type = nir_tex_src_lod; + src_number++; + } + break; + + case ir_txd: + instr->src[src_number].src = + nir_src_for_ssa(evaluate_rvalue(ir->lod_info.grad.dPdx)); + instr->src[src_number].src_type = nir_tex_src_ddx; + src_number++; + instr->src[src_number].src = + nir_src_for_ssa(evaluate_rvalue(ir->lod_info.grad.dPdy)); + instr->src[src_number].src_type = nir_tex_src_ddy; + src_number++; + break; + + case ir_txf_ms: + instr->src[src_number].src = + nir_src_for_ssa(evaluate_rvalue(ir->lod_info.sample_index)); + instr->src[src_number].src_type = nir_tex_src_ms_index; + src_number++; + break; + + case ir_tg4: + instr->component = ir->lod_info.component->as_constant()->value.u[0]; + break; + + default: + break; + } + + assert(src_number == num_srcs); + + add_instr(&instr->instr, nir_tex_instr_dest_size(instr)); +} + +void +nir_visitor::visit(ir_constant *ir) +{ + /* + * We don't know if this variable is an an array or struct that gets + * dereferenced, so do the safe thing an make it a variable with a + * constant initializer and return a dereference. + */ + + nir_variable *var = + nir_local_variable_create(this->impl, ir->type, "const_temp"); + var->data.read_only = true; + var->constant_initializer = constant_copy(ir, var); + + this->deref_head = nir_deref_var_create(this->shader, var); + this->deref_tail = &this->deref_head->deref; +} + +void +nir_visitor::visit(ir_dereference_variable *ir) +{ + struct hash_entry *entry = + _mesa_hash_table_search(this->var_table, ir->var); + assert(entry); + nir_variable *var = (nir_variable *) entry->data; + + nir_deref_var *deref = nir_deref_var_create(this->shader, var); + this->deref_head = deref; + this->deref_tail = &deref->deref; +} + +void +nir_visitor::visit(ir_dereference_record *ir) +{ + ir->record->accept(this); + + int field_index = this->deref_tail->type->field_index(ir->field); + assert(field_index >= 0); + + nir_deref_struct *deref = nir_deref_struct_create(this->deref_tail, field_index); + deref->deref.type = ir->type; + this->deref_tail->child = &deref->deref; + this->deref_tail = &deref->deref; +} + +void +nir_visitor::visit(ir_dereference_array *ir) +{ + nir_deref_array *deref = nir_deref_array_create(this->shader); + deref->deref.type = ir->type; + + ir_constant *const_index = ir->array_index->as_constant(); + if (const_index != NULL) { + deref->deref_array_type = nir_deref_array_type_direct; + deref->base_offset = const_index->value.u[0]; + } else { + deref->deref_array_type = nir_deref_array_type_indirect; + deref->indirect = + nir_src_for_ssa(evaluate_rvalue(ir->array_index)); + } + + ir->array->accept(this); + + this->deref_tail->child = &deref->deref; + ralloc_steal(this->deref_tail, deref); + this->deref_tail = &deref->deref; +} + +void +nir_visitor::visit(ir_barrier *ir) +{ + nir_intrinsic_instr *instr = + nir_intrinsic_instr_create(this->shader, nir_intrinsic_barrier); + nir_builder_instr_insert(&b, &instr->instr); +} diff --git a/src/compiler/nir/glsl_to_nir.h b/src/compiler/nir/glsl_to_nir.h new file mode 100644 index 0000000..20d2a38 --- /dev/null +++ b/src/compiler/nir/glsl_to_nir.h @@ -0,0 +1,41 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +#include "nir.h" +#include "glsl/glsl_parser_extras.h" + +#ifdef __cplusplus +extern "C" { +#endif + +nir_shader *glsl_to_nir(const struct gl_shader_program *shader_prog, + gl_shader_stage stage, + const nir_shader_compiler_options *options); + +#ifdef __cplusplus +} +#endif diff --git a/src/compiler/nir/nir.c b/src/compiler/nir/nir.c new file mode 100644 index 0000000..21bf678 --- /dev/null +++ b/src/compiler/nir/nir.c @@ -0,0 +1,1665 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +#include "nir.h" +#include "nir_control_flow_private.h" +#include <assert.h> + +nir_shader * +nir_shader_create(void *mem_ctx, + gl_shader_stage stage, + const nir_shader_compiler_options *options) +{ + nir_shader *shader = ralloc(mem_ctx, nir_shader); + + exec_list_make_empty(&shader->uniforms); + exec_list_make_empty(&shader->inputs); + exec_list_make_empty(&shader->outputs); + + shader->options = options; + memset(&shader->info, 0, sizeof(shader->info)); + + exec_list_make_empty(&shader->functions); + exec_list_make_empty(&shader->registers); + exec_list_make_empty(&shader->globals); + exec_list_make_empty(&shader->system_values); + shader->reg_alloc = 0; + + shader->num_inputs = 0; + shader->num_outputs = 0; + shader->num_uniforms = 0; + + shader->stage = stage; + + return shader; +} + +static nir_register * +reg_create(void *mem_ctx, struct exec_list *list) +{ + nir_register *reg = ralloc(mem_ctx, nir_register); + + list_inithead(®->uses); + list_inithead(®->defs); + list_inithead(®->if_uses); + + reg->num_components = 0; + reg->num_array_elems = 0; + reg->is_packed = false; + reg->name = NULL; + + exec_list_push_tail(list, ®->node); + + return reg; +} + +nir_register * +nir_global_reg_create(nir_shader *shader) +{ + nir_register *reg = reg_create(shader, &shader->registers); + reg->index = shader->reg_alloc++; + reg->is_global = true; + + return reg; +} + +nir_register * +nir_local_reg_create(nir_function_impl *impl) +{ + nir_register *reg = reg_create(ralloc_parent(impl), &impl->registers); + reg->index = impl->reg_alloc++; + reg->is_global = false; + + return reg; +} + +void +nir_reg_remove(nir_register *reg) +{ + exec_node_remove(®->node); +} + +void +nir_shader_add_variable(nir_shader *shader, nir_variable *var) +{ + switch (var->data.mode) { + case nir_var_all: + assert(!"invalid mode"); + break; + + case nir_var_local: + assert(!"nir_shader_add_variable cannot be used for local variables"); + break; + + case nir_var_global: + exec_list_push_tail(&shader->globals, &var->node); + break; + + case nir_var_shader_in: + exec_list_push_tail(&shader->inputs, &var->node); + break; + + case nir_var_shader_out: + exec_list_push_tail(&shader->outputs, &var->node); + break; + + case nir_var_uniform: + case nir_var_shader_storage: + exec_list_push_tail(&shader->uniforms, &var->node); + break; + + case nir_var_system_value: + exec_list_push_tail(&shader->system_values, &var->node); + break; + } +} + +nir_variable * +nir_variable_create(nir_shader *shader, nir_variable_mode mode, + const struct glsl_type *type, const char *name) +{ + nir_variable *var = rzalloc(shader, nir_variable); + var->name = ralloc_strdup(var, name); + var->type = type; + var->data.mode = mode; + + if ((mode == nir_var_shader_in && shader->stage != MESA_SHADER_VERTEX) || + (mode == nir_var_shader_out && shader->stage != MESA_SHADER_FRAGMENT)) + var->data.interpolation = INTERP_QUALIFIER_SMOOTH; + + if (mode == nir_var_shader_in || mode == nir_var_uniform) + var->data.read_only = true; + + nir_shader_add_variable(shader, var); + + return var; +} + +nir_variable * +nir_local_variable_create(nir_function_impl *impl, + const struct glsl_type *type, const char *name) +{ + nir_variable *var = rzalloc(impl->function->shader, nir_variable); + var->name = ralloc_strdup(var, name); + var->type = type; + var->data.mode = nir_var_local; + + nir_function_impl_add_variable(impl, var); + + return var; +} + +nir_function * +nir_function_create(nir_shader *shader, const char *name) +{ + nir_function *func = ralloc(shader, nir_function); + + exec_list_push_tail(&shader->functions, &func->node); + + func->name = ralloc_strdup(func, name); + func->shader = shader; + func->num_params = 0; + func->params = NULL; + func->return_type = glsl_void_type(); + func->impl = NULL; + + return func; +} + +void nir_src_copy(nir_src *dest, const nir_src *src, void *mem_ctx) +{ + dest->is_ssa = src->is_ssa; + if (src->is_ssa) { + dest->ssa = src->ssa; + } else { + dest->reg.base_offset = src->reg.base_offset; + dest->reg.reg = src->reg.reg; + if (src->reg.indirect) { + dest->reg.indirect = ralloc(mem_ctx, nir_src); + nir_src_copy(dest->reg.indirect, src->reg.indirect, mem_ctx); + } else { + dest->reg.indirect = NULL; + } + } +} + +void nir_dest_copy(nir_dest *dest, const nir_dest *src, nir_instr *instr) +{ + /* Copying an SSA definition makes no sense whatsoever. */ + assert(!src->is_ssa); + + dest->is_ssa = false; + + dest->reg.base_offset = src->reg.base_offset; + dest->reg.reg = src->reg.reg; + if (src->reg.indirect) { + dest->reg.indirect = ralloc(instr, nir_src); + nir_src_copy(dest->reg.indirect, src->reg.indirect, instr); + } else { + dest->reg.indirect = NULL; + } +} + +void +nir_alu_src_copy(nir_alu_src *dest, const nir_alu_src *src, + nir_alu_instr *instr) +{ + nir_src_copy(&dest->src, &src->src, &instr->instr); + dest->abs = src->abs; + dest->negate = src->negate; + for (unsigned i = 0; i < 4; i++) + dest->swizzle[i] = src->swizzle[i]; +} + +void +nir_alu_dest_copy(nir_alu_dest *dest, const nir_alu_dest *src, + nir_alu_instr *instr) +{ + nir_dest_copy(&dest->dest, &src->dest, &instr->instr); + dest->write_mask = src->write_mask; + dest->saturate = src->saturate; +} + + +static void +cf_init(nir_cf_node *node, nir_cf_node_type type) +{ + exec_node_init(&node->node); + node->parent = NULL; + node->type = type; +} + +nir_function_impl * +nir_function_impl_create(nir_function *function) +{ + assert(function->impl == NULL); + + void *mem_ctx = ralloc_parent(function); + + nir_function_impl *impl = ralloc(mem_ctx, nir_function_impl); + + function->impl = impl; + impl->function = function; + + cf_init(&impl->cf_node, nir_cf_node_function); + + exec_list_make_empty(&impl->body); + exec_list_make_empty(&impl->registers); + exec_list_make_empty(&impl->locals); + impl->num_params = 0; + impl->params = NULL; + impl->return_var = NULL; + impl->reg_alloc = 0; + impl->ssa_alloc = 0; + impl->valid_metadata = nir_metadata_none; + + /* create start & end blocks */ + nir_block *start_block = nir_block_create(mem_ctx); + nir_block *end_block = nir_block_create(mem_ctx); + start_block->cf_node.parent = &impl->cf_node; + end_block->cf_node.parent = &impl->cf_node; + impl->end_block = end_block; + + exec_list_push_tail(&impl->body, &start_block->cf_node.node); + + start_block->successors[0] = end_block; + _mesa_set_add(end_block->predecessors, start_block); + return impl; +} + +nir_block * +nir_block_create(nir_shader *shader) +{ + nir_block *block = ralloc(shader, nir_block); + + cf_init(&block->cf_node, nir_cf_node_block); + + block->successors[0] = block->successors[1] = NULL; + block->predecessors = _mesa_set_create(block, _mesa_hash_pointer, + _mesa_key_pointer_equal); + block->imm_dom = NULL; + /* XXX maybe it would be worth it to defer allocation? This + * way it doesn't get allocated for shader ref's that never run + * nir_calc_dominance? For example, state-tracker creates an + * initial IR, clones that, runs appropriate lowering pass, passes + * to driver which does common lowering/opt, and then stores ref + * which is later used to do state specific lowering and futher + * opt. Do any of the references not need dominance metadata? + */ + block->dom_frontier = _mesa_set_create(block, _mesa_hash_pointer, + _mesa_key_pointer_equal); + + exec_list_make_empty(&block->instr_list); + + return block; +} + +static inline void +src_init(nir_src *src) +{ + src->is_ssa = false; + src->reg.reg = NULL; + src->reg.indirect = NULL; + src->reg.base_offset = 0; +} + +nir_if * +nir_if_create(nir_shader *shader) +{ + nir_if *if_stmt = ralloc(shader, nir_if); + + cf_init(&if_stmt->cf_node, nir_cf_node_if); + src_init(&if_stmt->condition); + + nir_block *then = nir_block_create(shader); + exec_list_make_empty(&if_stmt->then_list); + exec_list_push_tail(&if_stmt->then_list, &then->cf_node.node); + then->cf_node.parent = &if_stmt->cf_node; + + nir_block *else_stmt = nir_block_create(shader); + exec_list_make_empty(&if_stmt->else_list); + exec_list_push_tail(&if_stmt->else_list, &else_stmt->cf_node.node); + else_stmt->cf_node.parent = &if_stmt->cf_node; + + return if_stmt; +} + +nir_loop * +nir_loop_create(nir_shader *shader) +{ + nir_loop *loop = ralloc(shader, nir_loop); + + cf_init(&loop->cf_node, nir_cf_node_loop); + + nir_block *body = nir_block_create(shader); + exec_list_make_empty(&loop->body); + exec_list_push_tail(&loop->body, &body->cf_node.node); + body->cf_node.parent = &loop->cf_node; + + body->successors[0] = body; + _mesa_set_add(body->predecessors, body); + + return loop; +} + +static void +instr_init(nir_instr *instr, nir_instr_type type) +{ + instr->type = type; + instr->block = NULL; + exec_node_init(&instr->node); +} + +static void +dest_init(nir_dest *dest) +{ + dest->is_ssa = false; + dest->reg.reg = NULL; + dest->reg.indirect = NULL; + dest->reg.base_offset = 0; +} + +static void +alu_dest_init(nir_alu_dest *dest) +{ + dest_init(&dest->dest); + dest->saturate = false; + dest->write_mask = 0xf; +} + +static void +alu_src_init(nir_alu_src *src) +{ + src_init(&src->src); + src->abs = src->negate = false; + src->swizzle[0] = 0; + src->swizzle[1] = 1; + src->swizzle[2] = 2; + src->swizzle[3] = 3; +} + +nir_alu_instr * +nir_alu_instr_create(nir_shader *shader, nir_op op) +{ + unsigned num_srcs = nir_op_infos[op].num_inputs; + nir_alu_instr *instr = + ralloc_size(shader, + sizeof(nir_alu_instr) + num_srcs * sizeof(nir_alu_src)); + + instr_init(&instr->instr, nir_instr_type_alu); + instr->op = op; + alu_dest_init(&instr->dest); + for (unsigned i = 0; i < num_srcs; i++) + alu_src_init(&instr->src[i]); + + return instr; +} + +nir_jump_instr * +nir_jump_instr_create(nir_shader *shader, nir_jump_type type) +{ + nir_jump_instr *instr = ralloc(shader, nir_jump_instr); + instr_init(&instr->instr, nir_instr_type_jump); + instr->type = type; + return instr; +} + +nir_load_const_instr * +nir_load_const_instr_create(nir_shader *shader, unsigned num_components) +{ + nir_load_const_instr *instr = ralloc(shader, nir_load_const_instr); + instr_init(&instr->instr, nir_instr_type_load_const); + + nir_ssa_def_init(&instr->instr, &instr->def, num_components, NULL); + + return instr; +} + +nir_intrinsic_instr * +nir_intrinsic_instr_create(nir_shader *shader, nir_intrinsic_op op) +{ + unsigned num_srcs = nir_intrinsic_infos[op].num_srcs; + nir_intrinsic_instr *instr = + ralloc_size(shader, + sizeof(nir_intrinsic_instr) + num_srcs * sizeof(nir_src)); + + instr_init(&instr->instr, nir_instr_type_intrinsic); + instr->intrinsic = op; + + if (nir_intrinsic_infos[op].has_dest) + dest_init(&instr->dest); + + for (unsigned i = 0; i < num_srcs; i++) + src_init(&instr->src[i]); + + return instr; +} + +nir_call_instr * +nir_call_instr_create(nir_shader *shader, nir_function *callee) +{ + nir_call_instr *instr = ralloc(shader, nir_call_instr); + instr_init(&instr->instr, nir_instr_type_call); + + instr->callee = callee; + instr->num_params = callee->num_params; + instr->params = ralloc_array(instr, nir_deref_var *, instr->num_params); + instr->return_deref = NULL; + + return instr; +} + +nir_tex_instr * +nir_tex_instr_create(nir_shader *shader, unsigned num_srcs) +{ + nir_tex_instr *instr = rzalloc(shader, nir_tex_instr); + instr_init(&instr->instr, nir_instr_type_tex); + + dest_init(&instr->dest); + + instr->num_srcs = num_srcs; + instr->src = ralloc_array(instr, nir_tex_src, num_srcs); + for (unsigned i = 0; i < num_srcs; i++) + src_init(&instr->src[i].src); + + instr->sampler_index = 0; + instr->sampler_array_size = 0; + instr->sampler = NULL; + + return instr; +} + +nir_phi_instr * +nir_phi_instr_create(nir_shader *shader) +{ + nir_phi_instr *instr = ralloc(shader, nir_phi_instr); + instr_init(&instr->instr, nir_instr_type_phi); + + dest_init(&instr->dest); + exec_list_make_empty(&instr->srcs); + return instr; +} + +nir_parallel_copy_instr * +nir_parallel_copy_instr_create(nir_shader *shader) +{ + nir_parallel_copy_instr *instr = ralloc(shader, nir_parallel_copy_instr); + instr_init(&instr->instr, nir_instr_type_parallel_copy); + + exec_list_make_empty(&instr->entries); + + return instr; +} + +nir_ssa_undef_instr * +nir_ssa_undef_instr_create(nir_shader *shader, unsigned num_components) +{ + nir_ssa_undef_instr *instr = ralloc(shader, nir_ssa_undef_instr); + instr_init(&instr->instr, nir_instr_type_ssa_undef); + + nir_ssa_def_init(&instr->instr, &instr->def, num_components, NULL); + + return instr; +} + +nir_deref_var * +nir_deref_var_create(void *mem_ctx, nir_variable *var) +{ + nir_deref_var *deref = ralloc(mem_ctx, nir_deref_var); + deref->deref.deref_type = nir_deref_type_var; + deref->deref.child = NULL; + deref->deref.type = var->type; + deref->var = var; + return deref; +} + +nir_deref_array * +nir_deref_array_create(void *mem_ctx) +{ + nir_deref_array *deref = ralloc(mem_ctx, nir_deref_array); + deref->deref.deref_type = nir_deref_type_array; + deref->deref.child = NULL; + deref->deref_array_type = nir_deref_array_type_direct; + src_init(&deref->indirect); + deref->base_offset = 0; + return deref; +} + +nir_deref_struct * +nir_deref_struct_create(void *mem_ctx, unsigned field_index) +{ + nir_deref_struct *deref = ralloc(mem_ctx, nir_deref_struct); + deref->deref.deref_type = nir_deref_type_struct; + deref->deref.child = NULL; + deref->index = field_index; + return deref; +} + +static nir_deref_var * +copy_deref_var(void *mem_ctx, nir_deref_var *deref) +{ + nir_deref_var *ret = nir_deref_var_create(mem_ctx, deref->var); + ret->deref.type = deref->deref.type; + if (deref->deref.child) + ret->deref.child = nir_copy_deref(ret, deref->deref.child); + return ret; +} + +static nir_deref_array * +copy_deref_array(void *mem_ctx, nir_deref_array *deref) +{ + nir_deref_array *ret = nir_deref_array_create(mem_ctx); + ret->base_offset = deref->base_offset; + ret->deref_array_type = deref->deref_array_type; + if (deref->deref_array_type == nir_deref_array_type_indirect) { + nir_src_copy(&ret->indirect, &deref->indirect, mem_ctx); + } + ret->deref.type = deref->deref.type; + if (deref->deref.child) + ret->deref.child = nir_copy_deref(ret, deref->deref.child); + return ret; +} + +static nir_deref_struct * +copy_deref_struct(void *mem_ctx, nir_deref_struct *deref) +{ + nir_deref_struct *ret = nir_deref_struct_create(mem_ctx, deref->index); + ret->deref.type = deref->deref.type; + if (deref->deref.child) + ret->deref.child = nir_copy_deref(ret, deref->deref.child); + return ret; +} + +nir_deref * +nir_copy_deref(void *mem_ctx, nir_deref *deref) +{ + switch (deref->deref_type) { + case nir_deref_type_var: + return ©_deref_var(mem_ctx, nir_deref_as_var(deref))->deref; + case nir_deref_type_array: + return ©_deref_array(mem_ctx, nir_deref_as_array(deref))->deref; + case nir_deref_type_struct: + return ©_deref_struct(mem_ctx, nir_deref_as_struct(deref))->deref; + default: + unreachable("Invalid dereference type"); + } + + return NULL; +} + +/* Returns a load_const instruction that represents the constant + * initializer for the given deref chain. The caller is responsible for + * ensuring that there actually is a constant initializer. + */ +nir_load_const_instr * +nir_deref_get_const_initializer_load(nir_shader *shader, nir_deref_var *deref) +{ + nir_constant *constant = deref->var->constant_initializer; + assert(constant); + + const nir_deref *tail = &deref->deref; + unsigned matrix_offset = 0; + while (tail->child) { + switch (tail->child->deref_type) { + case nir_deref_type_array: { + nir_deref_array *arr = nir_deref_as_array(tail->child); + assert(arr->deref_array_type == nir_deref_array_type_direct); + if (glsl_type_is_matrix(tail->type)) { + assert(arr->deref.child == NULL); + matrix_offset = arr->base_offset; + } else { + constant = constant->elements[arr->base_offset]; + } + break; + } + + case nir_deref_type_struct: { + constant = constant->elements[nir_deref_as_struct(tail->child)->index]; + break; + } + + default: + unreachable("Invalid deref child type"); + } + + tail = tail->child; + } + + nir_load_const_instr *load = + nir_load_const_instr_create(shader, glsl_get_vector_elements(tail->type)); + + matrix_offset *= load->def.num_components; + for (unsigned i = 0; i < load->def.num_components; i++) { + switch (glsl_get_base_type(tail->type)) { + case GLSL_TYPE_FLOAT: + case GLSL_TYPE_INT: + case GLSL_TYPE_UINT: + load->value.u[i] = constant->value.u[matrix_offset + i]; + break; + case GLSL_TYPE_BOOL: + load->value.u[i] = constant->value.b[matrix_offset + i] ? + NIR_TRUE : NIR_FALSE; + break; + default: + unreachable("Invalid immediate type"); + } + } + + return load; +} + +nir_function_impl * +nir_cf_node_get_function(nir_cf_node *node) +{ + while (node->type != nir_cf_node_function) { + node = node->parent; + } + + return nir_cf_node_as_function(node); +} + +static bool +add_use_cb(nir_src *src, void *state) +{ + nir_instr *instr = state; + + src->parent_instr = instr; + list_addtail(&src->use_link, + src->is_ssa ? &src->ssa->uses : &src->reg.reg->uses); + + return true; +} + +static bool +add_ssa_def_cb(nir_ssa_def *def, void *state) +{ + nir_instr *instr = state; + + if (instr->block && def->index == UINT_MAX) { + nir_function_impl *impl = + nir_cf_node_get_function(&instr->block->cf_node); + + def->index = impl->ssa_alloc++; + } + + return true; +} + +static bool +add_reg_def_cb(nir_dest *dest, void *state) +{ + nir_instr *instr = state; + + if (!dest->is_ssa) { + dest->reg.parent_instr = instr; + list_addtail(&dest->reg.def_link, &dest->reg.reg->defs); + } + + return true; +} + +static void +add_defs_uses(nir_instr *instr) +{ + nir_foreach_src(instr, add_use_cb, instr); + nir_foreach_dest(instr, add_reg_def_cb, instr); + nir_foreach_ssa_def(instr, add_ssa_def_cb, instr); +} + +void +nir_instr_insert(nir_cursor cursor, nir_instr *instr) +{ + switch (cursor.option) { + case nir_cursor_before_block: + /* Only allow inserting jumps into empty blocks. */ + if (instr->type == nir_instr_type_jump) + assert(exec_list_is_empty(&cursor.block->instr_list)); + + instr->block = cursor.block; + add_defs_uses(instr); + exec_list_push_head(&cursor.block->instr_list, &instr->node); + break; + case nir_cursor_after_block: { + /* Inserting instructions after a jump is illegal. */ + nir_instr *last = nir_block_last_instr(cursor.block); + assert(last == NULL || last->type != nir_instr_type_jump); + (void) last; + + instr->block = cursor.block; + add_defs_uses(instr); + exec_list_push_tail(&cursor.block->instr_list, &instr->node); + break; + } + case nir_cursor_before_instr: + assert(instr->type != nir_instr_type_jump); + instr->block = cursor.instr->block; + add_defs_uses(instr); + exec_node_insert_node_before(&cursor.instr->node, &instr->node); + break; + case nir_cursor_after_instr: + /* Inserting instructions after a jump is illegal. */ + assert(cursor.instr->type != nir_instr_type_jump); + + /* Only allow inserting jumps at the end of the block. */ + if (instr->type == nir_instr_type_jump) + assert(cursor.instr == nir_block_last_instr(cursor.instr->block)); + + instr->block = cursor.instr->block; + add_defs_uses(instr); + exec_node_insert_after(&cursor.instr->node, &instr->node); + break; + } + + if (instr->type == nir_instr_type_jump) + nir_handle_add_jump(instr->block); +} + +static bool +src_is_valid(const nir_src *src) +{ + return src->is_ssa ? (src->ssa != NULL) : (src->reg.reg != NULL); +} + +static bool +remove_use_cb(nir_src *src, void *state) +{ + if (src_is_valid(src)) + list_del(&src->use_link); + + return true; +} + +static bool +remove_def_cb(nir_dest *dest, void *state) +{ + if (!dest->is_ssa) + list_del(&dest->reg.def_link); + + return true; +} + +static void +remove_defs_uses(nir_instr *instr) +{ + nir_foreach_dest(instr, remove_def_cb, instr); + nir_foreach_src(instr, remove_use_cb, instr); +} + +void nir_instr_remove(nir_instr *instr) +{ + remove_defs_uses(instr); + exec_node_remove(&instr->node); + + if (instr->type == nir_instr_type_jump) { + nir_jump_instr *jump_instr = nir_instr_as_jump(instr); + nir_handle_remove_jump(instr->block, jump_instr->type); + } +} + +/*@}*/ + +void +nir_index_local_regs(nir_function_impl *impl) +{ + unsigned index = 0; + foreach_list_typed(nir_register, reg, node, &impl->registers) { + reg->index = index++; + } + impl->reg_alloc = index; +} + +void +nir_index_global_regs(nir_shader *shader) +{ + unsigned index = 0; + foreach_list_typed(nir_register, reg, node, &shader->registers) { + reg->index = index++; + } + shader->reg_alloc = index; +} + +static bool +visit_alu_dest(nir_alu_instr *instr, nir_foreach_dest_cb cb, void *state) +{ + return cb(&instr->dest.dest, state); +} + +static bool +visit_intrinsic_dest(nir_intrinsic_instr *instr, nir_foreach_dest_cb cb, + void *state) +{ + if (nir_intrinsic_infos[instr->intrinsic].has_dest) + return cb(&instr->dest, state); + + return true; +} + +static bool +visit_texture_dest(nir_tex_instr *instr, nir_foreach_dest_cb cb, + void *state) +{ + return cb(&instr->dest, state); +} + +static bool +visit_phi_dest(nir_phi_instr *instr, nir_foreach_dest_cb cb, void *state) +{ + return cb(&instr->dest, state); +} + +static bool +visit_parallel_copy_dest(nir_parallel_copy_instr *instr, + nir_foreach_dest_cb cb, void *state) +{ + nir_foreach_parallel_copy_entry(instr, entry) { + if (!cb(&entry->dest, state)) + return false; + } + + return true; +} + +bool +nir_foreach_dest(nir_instr *instr, nir_foreach_dest_cb cb, void *state) +{ + switch (instr->type) { + case nir_instr_type_alu: + return visit_alu_dest(nir_instr_as_alu(instr), cb, state); + case nir_instr_type_intrinsic: + return visit_intrinsic_dest(nir_instr_as_intrinsic(instr), cb, state); + case nir_instr_type_tex: + return visit_texture_dest(nir_instr_as_tex(instr), cb, state); + case nir_instr_type_phi: + return visit_phi_dest(nir_instr_as_phi(instr), cb, state); + case nir_instr_type_parallel_copy: + return visit_parallel_copy_dest(nir_instr_as_parallel_copy(instr), + cb, state); + + case nir_instr_type_load_const: + case nir_instr_type_ssa_undef: + case nir_instr_type_call: + case nir_instr_type_jump: + break; + + default: + unreachable("Invalid instruction type"); + break; + } + + return true; +} + +struct foreach_ssa_def_state { + nir_foreach_ssa_def_cb cb; + void *client_state; +}; + +static inline bool +nir_ssa_def_visitor(nir_dest *dest, void *void_state) +{ + struct foreach_ssa_def_state *state = void_state; + + if (dest->is_ssa) + return state->cb(&dest->ssa, state->client_state); + else + return true; +} + +bool +nir_foreach_ssa_def(nir_instr *instr, nir_foreach_ssa_def_cb cb, void *state) +{ + switch (instr->type) { + case nir_instr_type_alu: + case nir_instr_type_tex: + case nir_instr_type_intrinsic: + case nir_instr_type_phi: + case nir_instr_type_parallel_copy: { + struct foreach_ssa_def_state foreach_state = {cb, state}; + return nir_foreach_dest(instr, nir_ssa_def_visitor, &foreach_state); + } + + case nir_instr_type_load_const: + return cb(&nir_instr_as_load_const(instr)->def, state); + case nir_instr_type_ssa_undef: + return cb(&nir_instr_as_ssa_undef(instr)->def, state); + case nir_instr_type_call: + case nir_instr_type_jump: + return true; + default: + unreachable("Invalid instruction type"); + } +} + +static bool +visit_src(nir_src *src, nir_foreach_src_cb cb, void *state) +{ + if (!cb(src, state)) + return false; + if (!src->is_ssa && src->reg.indirect) + return cb(src->reg.indirect, state); + return true; +} + +static bool +visit_deref_array_src(nir_deref_array *deref, nir_foreach_src_cb cb, + void *state) +{ + if (deref->deref_array_type == nir_deref_array_type_indirect) + return visit_src(&deref->indirect, cb, state); + return true; +} + +static bool +visit_deref_src(nir_deref_var *deref, nir_foreach_src_cb cb, void *state) +{ + nir_deref *cur = &deref->deref; + while (cur != NULL) { + if (cur->deref_type == nir_deref_type_array) + if (!visit_deref_array_src(nir_deref_as_array(cur), cb, state)) + return false; + + cur = cur->child; + } + + return true; +} + +static bool +visit_alu_src(nir_alu_instr *instr, nir_foreach_src_cb cb, void *state) +{ + for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) + if (!visit_src(&instr->src[i].src, cb, state)) + return false; + + return true; +} + +static bool +visit_tex_src(nir_tex_instr *instr, nir_foreach_src_cb cb, void *state) +{ + for (unsigned i = 0; i < instr->num_srcs; i++) + if (!visit_src(&instr->src[i].src, cb, state)) + return false; + + if (instr->sampler != NULL) + if (!visit_deref_src(instr->sampler, cb, state)) + return false; + + return true; +} + +static bool +visit_intrinsic_src(nir_intrinsic_instr *instr, nir_foreach_src_cb cb, + void *state) +{ + unsigned num_srcs = nir_intrinsic_infos[instr->intrinsic].num_srcs; + for (unsigned i = 0; i < num_srcs; i++) + if (!visit_src(&instr->src[i], cb, state)) + return false; + + unsigned num_vars = + nir_intrinsic_infos[instr->intrinsic].num_variables; + for (unsigned i = 0; i < num_vars; i++) + if (!visit_deref_src(instr->variables[i], cb, state)) + return false; + + return true; +} + +static bool +visit_call_src(nir_call_instr *instr, nir_foreach_src_cb cb, void *state) +{ + return true; +} + +static bool +visit_load_const_src(nir_load_const_instr *instr, nir_foreach_src_cb cb, + void *state) +{ + return true; +} + +static bool +visit_phi_src(nir_phi_instr *instr, nir_foreach_src_cb cb, void *state) +{ + nir_foreach_phi_src(instr, src) { + if (!visit_src(&src->src, cb, state)) + return false; + } + + return true; +} + +static bool +visit_parallel_copy_src(nir_parallel_copy_instr *instr, + nir_foreach_src_cb cb, void *state) +{ + nir_foreach_parallel_copy_entry(instr, entry) { + if (!visit_src(&entry->src, cb, state)) + return false; + } + + return true; +} + +typedef struct { + void *state; + nir_foreach_src_cb cb; +} visit_dest_indirect_state; + +static bool +visit_dest_indirect(nir_dest *dest, void *_state) +{ + visit_dest_indirect_state *state = (visit_dest_indirect_state *) _state; + + if (!dest->is_ssa && dest->reg.indirect) + return state->cb(dest->reg.indirect, state->state); + + return true; +} + +bool +nir_foreach_src(nir_instr *instr, nir_foreach_src_cb cb, void *state) +{ + switch (instr->type) { + case nir_instr_type_alu: + if (!visit_alu_src(nir_instr_as_alu(instr), cb, state)) + return false; + break; + case nir_instr_type_intrinsic: + if (!visit_intrinsic_src(nir_instr_as_intrinsic(instr), cb, state)) + return false; + break; + case nir_instr_type_tex: + if (!visit_tex_src(nir_instr_as_tex(instr), cb, state)) + return false; + break; + case nir_instr_type_call: + if (!visit_call_src(nir_instr_as_call(instr), cb, state)) + return false; + break; + case nir_instr_type_load_const: + if (!visit_load_const_src(nir_instr_as_load_const(instr), cb, state)) + return false; + break; + case nir_instr_type_phi: + if (!visit_phi_src(nir_instr_as_phi(instr), cb, state)) + return false; + break; + case nir_instr_type_parallel_copy: + if (!visit_parallel_copy_src(nir_instr_as_parallel_copy(instr), + cb, state)) + return false; + break; + case nir_instr_type_jump: + case nir_instr_type_ssa_undef: + return true; + + default: + unreachable("Invalid instruction type"); + break; + } + + visit_dest_indirect_state dest_state; + dest_state.state = state; + dest_state.cb = cb; + return nir_foreach_dest(instr, visit_dest_indirect, &dest_state); +} + +nir_const_value * +nir_src_as_const_value(nir_src src) +{ + if (!src.is_ssa) + return NULL; + + if (src.ssa->parent_instr->type != nir_instr_type_load_const) + return NULL; + + nir_load_const_instr *load = nir_instr_as_load_const(src.ssa->parent_instr); + + return &load->value; +} + +/** + * Returns true if the source is known to be dynamically uniform. Otherwise it + * returns false which means it may or may not be dynamically uniform but it + * can't be determined. + */ +bool +nir_src_is_dynamically_uniform(nir_src src) +{ + if (!src.is_ssa) + return false; + + /* Constants are trivially dynamically uniform */ + if (src.ssa->parent_instr->type == nir_instr_type_load_const) + return true; + + /* As are uniform variables */ + if (src.ssa->parent_instr->type == nir_instr_type_intrinsic) { + nir_intrinsic_instr *intr = nir_instr_as_intrinsic(src.ssa->parent_instr); + + if (intr->intrinsic == nir_intrinsic_load_uniform) + return true; + } + + /* XXX: this could have many more tests, such as when a sampler function is + * called with dynamically uniform arguments. + */ + return false; +} + +static void +src_remove_all_uses(nir_src *src) +{ + for (; src; src = src->is_ssa ? NULL : src->reg.indirect) { + if (!src_is_valid(src)) + continue; + + list_del(&src->use_link); + } +} + +static void +src_add_all_uses(nir_src *src, nir_instr *parent_instr, nir_if *parent_if) +{ + for (; src; src = src->is_ssa ? NULL : src->reg.indirect) { + if (!src_is_valid(src)) + continue; + + if (parent_instr) { + src->parent_instr = parent_instr; + if (src->is_ssa) + list_addtail(&src->use_link, &src->ssa->uses); + else + list_addtail(&src->use_link, &src->reg.reg->uses); + } else { + assert(parent_if); + src->parent_if = parent_if; + if (src->is_ssa) + list_addtail(&src->use_link, &src->ssa->if_uses); + else + list_addtail(&src->use_link, &src->reg.reg->if_uses); + } + } +} + +void +nir_instr_rewrite_src(nir_instr *instr, nir_src *src, nir_src new_src) +{ + assert(!src_is_valid(src) || src->parent_instr == instr); + + src_remove_all_uses(src); + *src = new_src; + src_add_all_uses(src, instr, NULL); +} + +void +nir_instr_move_src(nir_instr *dest_instr, nir_src *dest, nir_src *src) +{ + assert(!src_is_valid(dest) || dest->parent_instr == dest_instr); + + src_remove_all_uses(dest); + src_remove_all_uses(src); + *dest = *src; + *src = NIR_SRC_INIT; + src_add_all_uses(dest, dest_instr, NULL); +} + +void +nir_if_rewrite_condition(nir_if *if_stmt, nir_src new_src) +{ + nir_src *src = &if_stmt->condition; + assert(!src_is_valid(src) || src->parent_if == if_stmt); + + src_remove_all_uses(src); + *src = new_src; + src_add_all_uses(src, NULL, if_stmt); +} + +void +nir_instr_rewrite_dest(nir_instr *instr, nir_dest *dest, nir_dest new_dest) +{ + if (dest->is_ssa) { + /* We can only overwrite an SSA destination if it has no uses. */ + assert(list_empty(&dest->ssa.uses) && list_empty(&dest->ssa.if_uses)); + } else { + list_del(&dest->reg.def_link); + if (dest->reg.indirect) + src_remove_all_uses(dest->reg.indirect); + } + + /* We can't re-write with an SSA def */ + assert(!new_dest.is_ssa); + + nir_dest_copy(dest, &new_dest, instr); + + dest->reg.parent_instr = instr; + list_addtail(&dest->reg.def_link, &new_dest.reg.reg->defs); + + if (dest->reg.indirect) + src_add_all_uses(dest->reg.indirect, instr, NULL); +} + +void +nir_ssa_def_init(nir_instr *instr, nir_ssa_def *def, + unsigned num_components, const char *name) +{ + def->name = name; + def->parent_instr = instr; + list_inithead(&def->uses); + list_inithead(&def->if_uses); + def->num_components = num_components; + + if (instr->block) { + nir_function_impl *impl = + nir_cf_node_get_function(&instr->block->cf_node); + + def->index = impl->ssa_alloc++; + } else { + def->index = UINT_MAX; + } +} + +void +nir_ssa_dest_init(nir_instr *instr, nir_dest *dest, + unsigned num_components, const char *name) +{ + dest->is_ssa = true; + nir_ssa_def_init(instr, &dest->ssa, num_components, name); +} + +void +nir_ssa_def_rewrite_uses(nir_ssa_def *def, nir_src new_src) +{ + assert(!new_src.is_ssa || def != new_src.ssa); + + nir_foreach_use_safe(def, use_src) + nir_instr_rewrite_src(use_src->parent_instr, use_src, new_src); + + nir_foreach_if_use_safe(def, use_src) + nir_if_rewrite_condition(use_src->parent_if, new_src); +} + +static bool +is_instr_between(nir_instr *start, nir_instr *end, nir_instr *between) +{ + assert(start->block == end->block); + + if (between->block != start->block) + return false; + + /* Search backwards looking for "between" */ + while (start != end) { + if (between == end) + return true; + + end = nir_instr_prev(end); + assert(end); + } + + return false; +} + +/* Replaces all uses of the given SSA def with the given source but only if + * the use comes after the after_me instruction. This can be useful if you + * are emitting code to fix up the result of some instruction: you can freely + * use the result in that code and then call rewrite_uses_after and pass the + * last fixup instruction as after_me and it will replace all of the uses you + * want without touching the fixup code. + * + * This function assumes that after_me is in the same block as + * def->parent_instr and that after_me comes after def->parent_instr. + */ +void +nir_ssa_def_rewrite_uses_after(nir_ssa_def *def, nir_src new_src, + nir_instr *after_me) +{ + assert(!new_src.is_ssa || def != new_src.ssa); + + nir_foreach_use_safe(def, use_src) { + assert(use_src->parent_instr != def->parent_instr); + /* Since def already dominates all of its uses, the only way a use can + * not be dominated by after_me is if it is between def and after_me in + * the instruction list. + */ + if (!is_instr_between(def->parent_instr, after_me, use_src->parent_instr)) + nir_instr_rewrite_src(use_src->parent_instr, use_src, new_src); + } + + nir_foreach_if_use_safe(def, use_src) + nir_if_rewrite_condition(use_src->parent_if, new_src); +} + +static bool foreach_cf_node(nir_cf_node *node, nir_foreach_block_cb cb, + bool reverse, void *state); + +static inline bool +foreach_if(nir_if *if_stmt, nir_foreach_block_cb cb, bool reverse, void *state) +{ + if (reverse) { + foreach_list_typed_reverse_safe(nir_cf_node, node, node, + &if_stmt->else_list) { + if (!foreach_cf_node(node, cb, reverse, state)) + return false; + } + + foreach_list_typed_reverse_safe(nir_cf_node, node, node, + &if_stmt->then_list) { + if (!foreach_cf_node(node, cb, reverse, state)) + return false; + } + } else { + foreach_list_typed_safe(nir_cf_node, node, node, &if_stmt->then_list) { + if (!foreach_cf_node(node, cb, reverse, state)) + return false; + } + + foreach_list_typed_safe(nir_cf_node, node, node, &if_stmt->else_list) { + if (!foreach_cf_node(node, cb, reverse, state)) + return false; + } + } + + return true; +} + +static inline bool +foreach_loop(nir_loop *loop, nir_foreach_block_cb cb, bool reverse, void *state) +{ + if (reverse) { + foreach_list_typed_reverse_safe(nir_cf_node, node, node, &loop->body) { + if (!foreach_cf_node(node, cb, reverse, state)) + return false; + } + } else { + foreach_list_typed_safe(nir_cf_node, node, node, &loop->body) { + if (!foreach_cf_node(node, cb, reverse, state)) + return false; + } + } + + return true; +} + +static bool +foreach_cf_node(nir_cf_node *node, nir_foreach_block_cb cb, + bool reverse, void *state) +{ + switch (node->type) { + case nir_cf_node_block: + return cb(nir_cf_node_as_block(node), state); + case nir_cf_node_if: + return foreach_if(nir_cf_node_as_if(node), cb, reverse, state); + case nir_cf_node_loop: + return foreach_loop(nir_cf_node_as_loop(node), cb, reverse, state); + break; + + default: + unreachable("Invalid CFG node type"); + break; + } + + return false; +} + +bool +nir_foreach_block_in_cf_node(nir_cf_node *node, nir_foreach_block_cb cb, + void *state) +{ + return foreach_cf_node(node, cb, false, state); +} + +bool +nir_foreach_block(nir_function_impl *impl, nir_foreach_block_cb cb, void *state) +{ + foreach_list_typed_safe(nir_cf_node, node, node, &impl->body) { + if (!foreach_cf_node(node, cb, false, state)) + return false; + } + + return cb(impl->end_block, state); +} + +bool +nir_foreach_block_reverse(nir_function_impl *impl, nir_foreach_block_cb cb, + void *state) +{ + if (!cb(impl->end_block, state)) + return false; + + foreach_list_typed_reverse_safe(nir_cf_node, node, node, &impl->body) { + if (!foreach_cf_node(node, cb, true, state)) + return false; + } + + return true; +} + +nir_if * +nir_block_get_following_if(nir_block *block) +{ + if (exec_node_is_tail_sentinel(&block->cf_node.node)) + return NULL; + + if (nir_cf_node_is_last(&block->cf_node)) + return NULL; + + nir_cf_node *next_node = nir_cf_node_next(&block->cf_node); + + if (next_node->type != nir_cf_node_if) + return NULL; + + return nir_cf_node_as_if(next_node); +} + +nir_loop * +nir_block_get_following_loop(nir_block *block) +{ + if (exec_node_is_tail_sentinel(&block->cf_node.node)) + return NULL; + + if (nir_cf_node_is_last(&block->cf_node)) + return NULL; + + nir_cf_node *next_node = nir_cf_node_next(&block->cf_node); + + if (next_node->type != nir_cf_node_loop) + return NULL; + + return nir_cf_node_as_loop(next_node); +} +static bool +index_block(nir_block *block, void *state) +{ + unsigned *index = state; + block->index = (*index)++; + return true; +} + +void +nir_index_blocks(nir_function_impl *impl) +{ + unsigned index = 0; + + if (impl->valid_metadata & nir_metadata_block_index) + return; + + nir_foreach_block(impl, index_block, &index); + + impl->num_blocks = index; +} + +static bool +index_ssa_def_cb(nir_ssa_def *def, void *state) +{ + unsigned *index = (unsigned *) state; + def->index = (*index)++; + + return true; +} + +static bool +index_ssa_block(nir_block *block, void *state) +{ + nir_foreach_instr(block, instr) + nir_foreach_ssa_def(instr, index_ssa_def_cb, state); + + return true; +} + +/** + * The indices are applied top-to-bottom which has the very nice property + * that, if A dominates B, then A->index <= B->index. + */ +void +nir_index_ssa_defs(nir_function_impl *impl) +{ + unsigned index = 0; + nir_foreach_block(impl, index_ssa_block, &index); + impl->ssa_alloc = index; +} + +static bool +index_instrs_block(nir_block *block, void *state) +{ + unsigned *index = state; + nir_foreach_instr(block, instr) + instr->index = (*index)++; + + return true; +} + +/** + * The indices are applied top-to-bottom which has the very nice property + * that, if A dominates B, then A->index <= B->index. + */ +unsigned +nir_index_instrs(nir_function_impl *impl) +{ + unsigned index = 0; + nir_foreach_block(impl, index_instrs_block, &index); + return index; +} + +nir_intrinsic_op +nir_intrinsic_from_system_value(gl_system_value val) +{ + switch (val) { + case SYSTEM_VALUE_VERTEX_ID: + return nir_intrinsic_load_vertex_id; + case SYSTEM_VALUE_INSTANCE_ID: + return nir_intrinsic_load_instance_id; + case SYSTEM_VALUE_DRAW_ID: + return nir_intrinsic_load_draw_id; + case SYSTEM_VALUE_BASE_INSTANCE: + return nir_intrinsic_load_base_instance; + case SYSTEM_VALUE_VERTEX_ID_ZERO_BASE: + return nir_intrinsic_load_vertex_id_zero_base; + case SYSTEM_VALUE_BASE_VERTEX: + return nir_intrinsic_load_base_vertex; + case SYSTEM_VALUE_INVOCATION_ID: + return nir_intrinsic_load_invocation_id; + case SYSTEM_VALUE_FRONT_FACE: + return nir_intrinsic_load_front_face; + case SYSTEM_VALUE_SAMPLE_ID: + return nir_intrinsic_load_sample_id; + case SYSTEM_VALUE_SAMPLE_POS: + return nir_intrinsic_load_sample_pos; + case SYSTEM_VALUE_SAMPLE_MASK_IN: + return nir_intrinsic_load_sample_mask_in; + case SYSTEM_VALUE_LOCAL_INVOCATION_ID: + return nir_intrinsic_load_local_invocation_id; + case SYSTEM_VALUE_WORK_GROUP_ID: + return nir_intrinsic_load_work_group_id; + case SYSTEM_VALUE_NUM_WORK_GROUPS: + return nir_intrinsic_load_num_work_groups; + case SYSTEM_VALUE_PRIMITIVE_ID: + return nir_intrinsic_load_primitive_id; + case SYSTEM_VALUE_TESS_COORD: + return nir_intrinsic_load_tess_coord; + case SYSTEM_VALUE_TESS_LEVEL_OUTER: + return nir_intrinsic_load_tess_level_outer; + case SYSTEM_VALUE_TESS_LEVEL_INNER: + return nir_intrinsic_load_tess_level_inner; + case SYSTEM_VALUE_VERTICES_IN: + return nir_intrinsic_load_patch_vertices_in; + case SYSTEM_VALUE_HELPER_INVOCATION: + return nir_intrinsic_load_helper_invocation; + default: + unreachable("system value does not directly correspond to intrinsic"); + } +} + +gl_system_value +nir_system_value_from_intrinsic(nir_intrinsic_op intrin) +{ + switch (intrin) { + case nir_intrinsic_load_vertex_id: + return SYSTEM_VALUE_VERTEX_ID; + case nir_intrinsic_load_instance_id: + return SYSTEM_VALUE_INSTANCE_ID; + case nir_intrinsic_load_draw_id: + return SYSTEM_VALUE_DRAW_ID; + case nir_intrinsic_load_base_instance: + return SYSTEM_VALUE_BASE_INSTANCE; + case nir_intrinsic_load_vertex_id_zero_base: + return SYSTEM_VALUE_VERTEX_ID_ZERO_BASE; + case nir_intrinsic_load_base_vertex: + return SYSTEM_VALUE_BASE_VERTEX; + case nir_intrinsic_load_invocation_id: + return SYSTEM_VALUE_INVOCATION_ID; + case nir_intrinsic_load_front_face: + return SYSTEM_VALUE_FRONT_FACE; + case nir_intrinsic_load_sample_id: + return SYSTEM_VALUE_SAMPLE_ID; + case nir_intrinsic_load_sample_pos: + return SYSTEM_VALUE_SAMPLE_POS; + case nir_intrinsic_load_sample_mask_in: + return SYSTEM_VALUE_SAMPLE_MASK_IN; + case nir_intrinsic_load_local_invocation_id: + return SYSTEM_VALUE_LOCAL_INVOCATION_ID; + case nir_intrinsic_load_num_work_groups: + return SYSTEM_VALUE_NUM_WORK_GROUPS; + case nir_intrinsic_load_work_group_id: + return SYSTEM_VALUE_WORK_GROUP_ID; + case nir_intrinsic_load_primitive_id: + return SYSTEM_VALUE_PRIMITIVE_ID; + case nir_intrinsic_load_tess_coord: + return SYSTEM_VALUE_TESS_COORD; + case nir_intrinsic_load_tess_level_outer: + return SYSTEM_VALUE_TESS_LEVEL_OUTER; + case nir_intrinsic_load_tess_level_inner: + return SYSTEM_VALUE_TESS_LEVEL_INNER; + case nir_intrinsic_load_patch_vertices_in: + return SYSTEM_VALUE_VERTICES_IN; + case nir_intrinsic_load_helper_invocation: + return SYSTEM_VALUE_HELPER_INVOCATION; + default: + unreachable("intrinsic doesn't produce a system value"); + } +} diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h new file mode 100644 index 0000000..54e23eb --- /dev/null +++ b/src/compiler/nir/nir.h @@ -0,0 +1,2111 @@ +/* + * Copyright © 2014 Connor Abbott + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +#pragma once + +#include "util/hash_table.h" +#include "glsl/list.h" +#include "GL/gl.h" /* GLenum */ +#include "util/list.h" +#include "util/ralloc.h" +#include "util/set.h" +#include "util/bitset.h" +#include "compiler/nir_types.h" +#include "compiler/shader_enums.h" +#include <stdio.h> + +#include "nir_opcodes.h" + +#ifdef __cplusplus +extern "C" { +#endif + +struct gl_program; +struct gl_shader_program; + +#define NIR_FALSE 0u +#define NIR_TRUE (~0u) + +/** Defines a cast function + * + * This macro defines a cast function from in_type to out_type where + * out_type is some structure type that contains a field of type out_type. + * + * Note that you have to be a bit careful as the generated cast function + * destroys constness. + */ +#define NIR_DEFINE_CAST(name, in_type, out_type, field) \ +static inline out_type * \ +name(const in_type *parent) \ +{ \ + return exec_node_data(out_type, parent, field); \ +} + +struct nir_function; +struct nir_shader; +struct nir_instr; + + +/** + * Description of built-in state associated with a uniform + * + * \sa nir_variable::state_slots + */ +typedef struct { + int tokens[5]; + int swizzle; +} nir_state_slot; + +typedef enum { + nir_var_all = -1, + nir_var_shader_in, + nir_var_shader_out, + nir_var_global, + nir_var_local, + nir_var_uniform, + nir_var_shader_storage, + nir_var_system_value +} nir_variable_mode; + +/** + * Data stored in an nir_constant + */ +union nir_constant_data { + unsigned u[16]; + int i[16]; + float f[16]; + bool b[16]; +}; + +typedef struct nir_constant { + /** + * Value of the constant. + * + * The field used to back the values supplied by the constant is determined + * by the type associated with the \c nir_variable. Constants may be + * scalars, vectors, or matrices. + */ + union nir_constant_data value; + + /* we could get this from the var->type but makes clone *much* easier to + * not have to care about the type. + */ + unsigned num_elements; + + /* Array elements / Structure Fields */ + struct nir_constant **elements; +} nir_constant; + +/** + * \brief Layout qualifiers for gl_FragDepth. + * + * The AMD/ARB_conservative_depth extensions allow gl_FragDepth to be redeclared + * with a layout qualifier. + */ +typedef enum { + nir_depth_layout_none, /**< No depth layout is specified. */ + nir_depth_layout_any, + nir_depth_layout_greater, + nir_depth_layout_less, + nir_depth_layout_unchanged +} nir_depth_layout; + +/** + * Either a uniform, global variable, shader input, or shader output. Based on + * ir_variable - it should be easy to translate between the two. + */ + +typedef struct nir_variable { + struct exec_node node; + + /** + * Declared type of the variable + */ + const struct glsl_type *type; + + /** + * Declared name of the variable + */ + char *name; + + struct nir_variable_data { + + /** + * Is the variable read-only? + * + * This is set for variables declared as \c const, shader inputs, + * and uniforms. + */ + unsigned read_only:1; + unsigned centroid:1; + unsigned sample:1; + unsigned patch:1; + unsigned invariant:1; + + /** + * Storage class of the variable. + * + * \sa nir_variable_mode + */ + nir_variable_mode mode:4; + + /** + * Interpolation mode for shader inputs / outputs + * + * \sa glsl_interp_qualifier + */ + unsigned interpolation:2; + + /** + * \name ARB_fragment_coord_conventions + * @{ + */ + unsigned origin_upper_left:1; + unsigned pixel_center_integer:1; + /*@}*/ + + /** + * Was the location explicitly set in the shader? + * + * If the location is explicitly set in the shader, it \b cannot be changed + * by the linker or by the API (e.g., calls to \c glBindAttribLocation have + * no effect). + */ + unsigned explicit_location:1; + unsigned explicit_index:1; + + /** + * Was an initial binding explicitly set in the shader? + * + * If so, constant_initializer contains an integer nir_constant + * representing the initial binding point. + */ + unsigned explicit_binding:1; + + /** + * Does this variable have an initializer? + * + * This is used by the linker to cross-validiate initializers of global + * variables. + */ + unsigned has_initializer:1; + + /** + * If non-zero, then this variable may be packed along with other variables + * into a single varying slot, so this offset should be applied when + * accessing components. For example, an offset of 1 means that the x + * component of this variable is actually stored in component y of the + * location specified by \c location. + */ + unsigned location_frac:2; + + /** + * Non-zero if this variable was created by lowering a named interface + * block which was not an array. + * + * Note that this variable and \c from_named_ifc_block_array will never + * both be non-zero. + */ + unsigned from_named_ifc_block_nonarray:1; + + /** + * Non-zero if this variable was created by lowering a named interface + * block which was an array. + * + * Note that this variable and \c from_named_ifc_block_nonarray will never + * both be non-zero. + */ + unsigned from_named_ifc_block_array:1; + + /** + * \brief Layout qualifier for gl_FragDepth. + * + * This is not equal to \c ir_depth_layout_none if and only if this + * variable is \c gl_FragDepth and a layout qualifier is specified. + */ + nir_depth_layout depth_layout; + + /** + * Storage location of the base of this variable + * + * The precise meaning of this field depends on the nature of the variable. + * + * - Vertex shader input: one of the values from \c gl_vert_attrib. + * - Vertex shader output: one of the values from \c gl_varying_slot. + * - Geometry shader input: one of the values from \c gl_varying_slot. + * - Geometry shader output: one of the values from \c gl_varying_slot. + * - Fragment shader input: one of the values from \c gl_varying_slot. + * - Fragment shader output: one of the values from \c gl_frag_result. + * - Uniforms: Per-stage uniform slot number for default uniform block. + * - Uniforms: Index within the uniform block definition for UBO members. + * - Non-UBO Uniforms: uniform slot number. + * - Other: This field is not currently used. + * + * If the variable is a uniform, shader input, or shader output, and the + * slot has not been assigned, the value will be -1. + */ + int location; + + /** + * The actual location of the variable in the IR. Only valid for inputs + * and outputs. + */ + unsigned int driver_location; + + /** + * output index for dual source blending. + */ + int index; + + /** + * Initial binding point for a sampler or UBO. + * + * For array types, this represents the binding point for the first element. + */ + int binding; + + /** + * Location an atomic counter is stored at. + */ + unsigned offset; + + /** + * ARB_shader_image_load_store qualifiers. + */ + struct { + bool read_only; /**< "readonly" qualifier. */ + bool write_only; /**< "writeonly" qualifier. */ + bool coherent; + bool _volatile; + bool restrict_flag; + + /** Image internal format if specified explicitly, otherwise GL_NONE. */ + GLenum format; + } image; + + /** + * Highest element accessed with a constant expression array index + * + * Not used for non-array variables. + */ + unsigned max_array_access; + + } data; + + /** + * Built-in state that backs this uniform + * + * Once set at variable creation, \c state_slots must remain invariant. + * This is because, ideally, this array would be shared by all clones of + * this variable in the IR tree. In other words, we'd really like for it + * to be a fly-weight. + * + * If the variable is not a uniform, \c num_state_slots will be zero and + * \c state_slots will be \c NULL. + */ + /*@{*/ + unsigned num_state_slots; /**< Number of state slots used */ + nir_state_slot *state_slots; /**< State descriptors. */ + /*@}*/ + + /** + * Constant expression assigned in the initializer of the variable + */ + nir_constant *constant_initializer; + + /** + * For variables that are in an interface block or are an instance of an + * interface block, this is the \c GLSL_TYPE_INTERFACE type for that block. + * + * \sa ir_variable::location + */ + const struct glsl_type *interface_type; +} nir_variable; + +#define nir_foreach_variable(var, var_list) \ + foreach_list_typed(nir_variable, var, node, var_list) + +typedef struct nir_register { + struct exec_node node; + + unsigned num_components; /** < number of vector components */ + unsigned num_array_elems; /** < size of array (0 for no array) */ + + /** generic register index. */ + unsigned index; + + /** only for debug purposes, can be NULL */ + const char *name; + + /** whether this register is local (per-function) or global (per-shader) */ + bool is_global; + + /** + * If this flag is set to true, then accessing channels >= num_components + * is well-defined, and simply spills over to the next array element. This + * is useful for backends that can do per-component accessing, in + * particular scalar backends. By setting this flag and making + * num_components equal to 1, structures can be packed tightly into + * registers and then registers can be accessed per-component to get to + * each structure member, even if it crosses vec4 boundaries. + */ + bool is_packed; + + /** set of nir_src's where this register is used (read from) */ + struct list_head uses; + + /** set of nir_dest's where this register is defined (written to) */ + struct list_head defs; + + /** set of nir_if's where this register is used as a condition */ + struct list_head if_uses; +} nir_register; + +typedef enum { + nir_instr_type_alu, + nir_instr_type_call, + nir_instr_type_tex, + nir_instr_type_intrinsic, + nir_instr_type_load_const, + nir_instr_type_jump, + nir_instr_type_ssa_undef, + nir_instr_type_phi, + nir_instr_type_parallel_copy, +} nir_instr_type; + +typedef struct nir_instr { + struct exec_node node; + nir_instr_type type; + struct nir_block *block; + + /** generic instruction index. */ + unsigned index; + + /* A temporary for optimization and analysis passes to use for storing + * flags. For instance, DCE uses this to store the "dead/live" info. + */ + uint8_t pass_flags; +} nir_instr; + +static inline nir_instr * +nir_instr_next(nir_instr *instr) +{ + struct exec_node *next = exec_node_get_next(&instr->node); + if (exec_node_is_tail_sentinel(next)) + return NULL; + else + return exec_node_data(nir_instr, next, node); +} + +static inline nir_instr * +nir_instr_prev(nir_instr *instr) +{ + struct exec_node *prev = exec_node_get_prev(&instr->node); + if (exec_node_is_head_sentinel(prev)) + return NULL; + else + return exec_node_data(nir_instr, prev, node); +} + +static inline bool +nir_instr_is_first(nir_instr *instr) +{ + return exec_node_is_head_sentinel(exec_node_get_prev(&instr->node)); +} + +static inline bool +nir_instr_is_last(nir_instr *instr) +{ + return exec_node_is_tail_sentinel(exec_node_get_next(&instr->node)); +} + +typedef struct nir_ssa_def { + /** for debugging only, can be NULL */ + const char* name; + + /** generic SSA definition index. */ + unsigned index; + + /** Index into the live_in and live_out bitfields */ + unsigned live_index; + + nir_instr *parent_instr; + + /** set of nir_instr's where this register is used (read from) */ + struct list_head uses; + + /** set of nir_if's where this register is used as a condition */ + struct list_head if_uses; + + uint8_t num_components; +} nir_ssa_def; + +struct nir_src; + +typedef struct { + nir_register *reg; + struct nir_src *indirect; /** < NULL for no indirect offset */ + unsigned base_offset; + + /* TODO use-def chain goes here */ +} nir_reg_src; + +typedef struct { + nir_instr *parent_instr; + struct list_head def_link; + + nir_register *reg; + struct nir_src *indirect; /** < NULL for no indirect offset */ + unsigned base_offset; + + /* TODO def-use chain goes here */ +} nir_reg_dest; + +struct nir_if; + +typedef struct nir_src { + union { + nir_instr *parent_instr; + struct nir_if *parent_if; + }; + + struct list_head use_link; + + union { + nir_reg_src reg; + nir_ssa_def *ssa; + }; + + bool is_ssa; +} nir_src; + +#define NIR_SRC_INIT (nir_src) { { NULL } } + +#define nir_foreach_use(reg_or_ssa_def, src) \ + list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->uses, use_link) + +#define nir_foreach_use_safe(reg_or_ssa_def, src) \ + list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->uses, use_link) + +#define nir_foreach_if_use(reg_or_ssa_def, src) \ + list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link) + +#define nir_foreach_if_use_safe(reg_or_ssa_def, src) \ + list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link) + +typedef struct { + union { + nir_reg_dest reg; + nir_ssa_def ssa; + }; + + bool is_ssa; +} nir_dest; + +#define NIR_DEST_INIT (nir_dest) { { { NULL } } } + +#define nir_foreach_def(reg, dest) \ + list_for_each_entry(nir_dest, dest, &(reg)->defs, reg.def_link) + +#define nir_foreach_def_safe(reg, dest) \ + list_for_each_entry_safe(nir_dest, dest, &(reg)->defs, reg.def_link) + +static inline nir_src +nir_src_for_ssa(nir_ssa_def *def) +{ + nir_src src = NIR_SRC_INIT; + + src.is_ssa = true; + src.ssa = def; + + return src; +} + +static inline nir_src +nir_src_for_reg(nir_register *reg) +{ + nir_src src = NIR_SRC_INIT; + + src.is_ssa = false; + src.reg.reg = reg; + src.reg.indirect = NULL; + src.reg.base_offset = 0; + + return src; +} + +static inline nir_dest +nir_dest_for_reg(nir_register *reg) +{ + nir_dest dest = NIR_DEST_INIT; + + dest.reg.reg = reg; + + return dest; +} + +void nir_src_copy(nir_src *dest, const nir_src *src, void *instr_or_if); +void nir_dest_copy(nir_dest *dest, const nir_dest *src, nir_instr *instr); + +typedef struct { + nir_src src; + + /** + * \name input modifiers + */ + /*@{*/ + /** + * For inputs interpreted as floating point, flips the sign bit. For + * inputs interpreted as integers, performs the two's complement negation. + */ + bool negate; + + /** + * Clears the sign bit for floating point values, and computes the integer + * absolute value for integers. Note that the negate modifier acts after + * the absolute value modifier, therefore if both are set then all inputs + * will become negative. + */ + bool abs; + /*@}*/ + + /** + * For each input component, says which component of the register it is + * chosen from. Note that which elements of the swizzle are used and which + * are ignored are based on the write mask for most opcodes - for example, + * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and + * a swizzle of {2, x, 1, 0} where x means "don't care." + */ + uint8_t swizzle[4]; +} nir_alu_src; + +typedef struct { + nir_dest dest; + + /** + * \name saturate output modifier + * + * Only valid for opcodes that output floating-point numbers. Clamps the + * output to between 0.0 and 1.0 inclusive. + */ + + bool saturate; + + unsigned write_mask : 4; /* ignored if dest.is_ssa is true */ +} nir_alu_dest; + +typedef enum { + nir_type_invalid = 0, /* Not a valid type */ + nir_type_float, + nir_type_int, + nir_type_uint, + nir_type_bool +} nir_alu_type; + +typedef enum { + NIR_OP_IS_COMMUTATIVE = (1 << 0), + NIR_OP_IS_ASSOCIATIVE = (1 << 1), +} nir_op_algebraic_property; + +typedef struct { + const char *name; + + unsigned num_inputs; + + /** + * The number of components in the output + * + * If non-zero, this is the size of the output and input sizes are + * explicitly given; swizzle and writemask are still in effect, but if + * the output component is masked out, then the input component may + * still be in use. + * + * If zero, the opcode acts in the standard, per-component manner; the + * operation is performed on each component (except the ones that are + * masked out) with the input being taken from the input swizzle for + * that component. + * + * The size of some of the inputs may be given (i.e. non-zero) even + * though output_size is zero; in that case, the inputs with a zero + * size act per-component, while the inputs with non-zero size don't. + */ + unsigned output_size; + + /** + * The type of vector that the instruction outputs. Note that the + * staurate modifier is only allowed on outputs with the float type. + */ + + nir_alu_type output_type; + + /** + * The number of components in each input + */ + unsigned input_sizes[4]; + + /** + * The type of vector that each input takes. Note that negate and + * absolute value are only allowed on inputs with int or float type and + * behave differently on the two. + */ + nir_alu_type input_types[4]; + + nir_op_algebraic_property algebraic_properties; +} nir_op_info; + +extern const nir_op_info nir_op_infos[nir_num_opcodes]; + +typedef struct nir_alu_instr { + nir_instr instr; + nir_op op; + nir_alu_dest dest; + nir_alu_src src[]; +} nir_alu_instr; + +void nir_alu_src_copy(nir_alu_src *dest, const nir_alu_src *src, + nir_alu_instr *instr); +void nir_alu_dest_copy(nir_alu_dest *dest, const nir_alu_dest *src, + nir_alu_instr *instr); + +/* is this source channel used? */ +static inline bool +nir_alu_instr_channel_used(nir_alu_instr *instr, unsigned src, unsigned channel) +{ + if (nir_op_infos[instr->op].input_sizes[src] > 0) + return channel < nir_op_infos[instr->op].input_sizes[src]; + + return (instr->dest.write_mask >> channel) & 1; +} + +/* + * For instructions whose destinations are SSA, get the number of channels + * used for a source + */ +static inline unsigned +nir_ssa_alu_instr_src_components(const nir_alu_instr *instr, unsigned src) +{ + assert(instr->dest.dest.is_ssa); + + if (nir_op_infos[instr->op].input_sizes[src] > 0) + return nir_op_infos[instr->op].input_sizes[src]; + + return instr->dest.dest.ssa.num_components; +} + +typedef enum { + nir_deref_type_var, + nir_deref_type_array, + nir_deref_type_struct +} nir_deref_type; + +typedef struct nir_deref { + nir_deref_type deref_type; + struct nir_deref *child; + const struct glsl_type *type; +} nir_deref; + +typedef struct { + nir_deref deref; + + nir_variable *var; +} nir_deref_var; + +/* This enum describes how the array is referenced. If the deref is + * direct then the base_offset is used. If the deref is indirect then then + * offset is given by base_offset + indirect. If the deref is a wildcard + * then the deref refers to all of the elements of the array at the same + * time. Wildcard dereferences are only ever allowed in copy_var + * intrinsics and the source and destination derefs must have matching + * wildcards. + */ +typedef enum { + nir_deref_array_type_direct, + nir_deref_array_type_indirect, + nir_deref_array_type_wildcard, +} nir_deref_array_type; + +typedef struct { + nir_deref deref; + + nir_deref_array_type deref_array_type; + unsigned base_offset; + nir_src indirect; +} nir_deref_array; + +typedef struct { + nir_deref deref; + + unsigned index; +} nir_deref_struct; + +NIR_DEFINE_CAST(nir_deref_as_var, nir_deref, nir_deref_var, deref) +NIR_DEFINE_CAST(nir_deref_as_array, nir_deref, nir_deref_array, deref) +NIR_DEFINE_CAST(nir_deref_as_struct, nir_deref, nir_deref_struct, deref) + +/* Returns the last deref in the chain. */ +static inline nir_deref * +nir_deref_tail(nir_deref *deref) +{ + while (deref->child) + deref = deref->child; + return deref; +} + +typedef struct { + nir_instr instr; + + unsigned num_params; + nir_deref_var **params; + nir_deref_var *return_deref; + + struct nir_function *callee; +} nir_call_instr; + +#define INTRINSIC(name, num_srcs, src_components, has_dest, dest_components, \ + num_variables, num_indices, flags) \ + nir_intrinsic_##name, + +#define LAST_INTRINSIC(name) nir_last_intrinsic = nir_intrinsic_##name, + +typedef enum { +#include "nir_intrinsics.h" + nir_num_intrinsics = nir_last_intrinsic + 1 +} nir_intrinsic_op; + +#undef INTRINSIC +#undef LAST_INTRINSIC + +/** Represents an intrinsic + * + * An intrinsic is an instruction type for handling things that are + * more-or-less regular operations but don't just consume and produce SSA + * values like ALU operations do. Intrinsics are not for things that have + * special semantic meaning such as phi nodes and parallel copies. + * Examples of intrinsics include variable load/store operations, system + * value loads, and the like. Even though texturing more-or-less falls + * under this category, texturing is its own instruction type because + * trying to represent texturing with intrinsics would lead to a + * combinatorial explosion of intrinsic opcodes. + * + * By having a single instruction type for handling a lot of different + * cases, optimization passes can look for intrinsics and, for the most + * part, completely ignore them. Each intrinsic type also has a few + * possible flags that govern whether or not they can be reordered or + * eliminated. That way passes like dead code elimination can still work + * on intrisics without understanding the meaning of each. + * + * Each intrinsic has some number of constant indices, some number of + * variables, and some number of sources. What these sources, variables, + * and indices mean depends on the intrinsic and is documented with the + * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture + * instructions are the only types of instruction that can operate on + * variables. + */ +typedef struct { + nir_instr instr; + + nir_intrinsic_op intrinsic; + + nir_dest dest; + + /** number of components if this is a vectorized intrinsic + * + * Similarly to ALU operations, some intrinsics are vectorized. + * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0. + * For vectorized intrinsics, the num_components field specifies the + * number of destination components and the number of source components + * for all sources with nir_intrinsic_infos.src_components[i] == 0. + */ + uint8_t num_components; + + int const_index[3]; + + nir_deref_var *variables[2]; + + nir_src src[]; +} nir_intrinsic_instr; + +/** + * \name NIR intrinsics semantic flags + * + * information about what the compiler can do with the intrinsics. + * + * \sa nir_intrinsic_info::flags + */ +typedef enum { + /** + * whether the intrinsic can be safely eliminated if none of its output + * value is not being used. + */ + NIR_INTRINSIC_CAN_ELIMINATE = (1 << 0), + + /** + * Whether the intrinsic can be reordered with respect to any other + * intrinsic, i.e. whether the only reordering dependencies of the + * intrinsic are due to the register reads/writes. + */ + NIR_INTRINSIC_CAN_REORDER = (1 << 1), +} nir_intrinsic_semantic_flag; + +#define NIR_INTRINSIC_MAX_INPUTS 4 + +typedef struct { + const char *name; + + unsigned num_srcs; /** < number of register/SSA inputs */ + + /** number of components of each input register + * + * If this value is 0, the number of components is given by the + * num_components field of nir_intrinsic_instr. + */ + unsigned src_components[NIR_INTRINSIC_MAX_INPUTS]; + + bool has_dest; + + /** number of components of the output register + * + * If this value is 0, the number of components is given by the + * num_components field of nir_intrinsic_instr. + */ + unsigned dest_components; + + /** the number of inputs/outputs that are variables */ + unsigned num_variables; + + /** the number of constant indices used by the intrinsic */ + unsigned num_indices; + + /** semantic flags for calls to this intrinsic */ + nir_intrinsic_semantic_flag flags; +} nir_intrinsic_info; + +extern const nir_intrinsic_info nir_intrinsic_infos[nir_num_intrinsics]; + +/** + * \group texture information + * + * This gives semantic information about textures which is useful to the + * frontend, the backend, and lowering passes, but not the optimizer. + */ + +typedef enum { + nir_tex_src_coord, + nir_tex_src_projector, + nir_tex_src_comparitor, /* shadow comparitor */ + nir_tex_src_offset, + nir_tex_src_bias, + nir_tex_src_lod, + nir_tex_src_ms_index, /* MSAA sample index */ + nir_tex_src_ddx, + nir_tex_src_ddy, + nir_tex_src_sampler_offset, /* < dynamically uniform indirect offset */ + nir_num_tex_src_types +} nir_tex_src_type; + +typedef struct { + nir_src src; + nir_tex_src_type src_type; +} nir_tex_src; + +typedef enum { + nir_texop_tex, /**< Regular texture look-up */ + nir_texop_txb, /**< Texture look-up with LOD bias */ + nir_texop_txl, /**< Texture look-up with explicit LOD */ + nir_texop_txd, /**< Texture look-up with partial derivatvies */ + nir_texop_txf, /**< Texel fetch with explicit LOD */ + nir_texop_txf_ms, /**< Multisample texture fetch */ + nir_texop_txs, /**< Texture size */ + nir_texop_lod, /**< Texture lod query */ + nir_texop_tg4, /**< Texture gather */ + nir_texop_query_levels, /**< Texture levels query */ + nir_texop_texture_samples, /**< Texture samples query */ + nir_texop_samples_identical, /**< Query whether all samples are definitely + * identical. + */ +} nir_texop; + +typedef struct { + nir_instr instr; + + enum glsl_sampler_dim sampler_dim; + nir_alu_type dest_type; + + nir_texop op; + nir_dest dest; + nir_tex_src *src; + unsigned num_srcs, coord_components; + bool is_array, is_shadow; + + /** + * If is_shadow is true, whether this is the old-style shadow that outputs 4 + * components or the new-style shadow that outputs 1 component. + */ + bool is_new_style_shadow; + + /* constant offset - must be 0 if the offset source is used */ + int const_offset[4]; + + /* gather component selector */ + unsigned component : 2; + + /** The sampler index + * + * If this texture instruction has a nir_tex_src_sampler_offset source, + * then the sampler index is given by sampler_index + sampler_offset. + */ + unsigned sampler_index; + + /** The size of the sampler array or 0 if it's not an array */ + unsigned sampler_array_size; + + nir_deref_var *sampler; /* if this is NULL, use sampler_index instead */ +} nir_tex_instr; + +static inline unsigned +nir_tex_instr_dest_size(nir_tex_instr *instr) +{ + switch (instr->op) { + case nir_texop_txs: { + unsigned ret; + switch (instr->sampler_dim) { + case GLSL_SAMPLER_DIM_1D: + case GLSL_SAMPLER_DIM_BUF: + ret = 1; + break; + case GLSL_SAMPLER_DIM_2D: + case GLSL_SAMPLER_DIM_CUBE: + case GLSL_SAMPLER_DIM_MS: + case GLSL_SAMPLER_DIM_RECT: + case GLSL_SAMPLER_DIM_EXTERNAL: + ret = 2; + break; + case GLSL_SAMPLER_DIM_3D: + ret = 3; + break; + default: + unreachable("not reached"); + } + if (instr->is_array) + ret++; + return ret; + } + + case nir_texop_lod: + return 2; + + case nir_texop_texture_samples: + case nir_texop_query_levels: + case nir_texop_samples_identical: + return 1; + + default: + if (instr->is_shadow && instr->is_new_style_shadow) + return 1; + + return 4; + } +} + +/* Returns true if this texture operation queries something about the texture + * rather than actually sampling it. + */ +static inline bool +nir_tex_instr_is_query(nir_tex_instr *instr) +{ + switch (instr->op) { + case nir_texop_txs: + case nir_texop_lod: + case nir_texop_texture_samples: + case nir_texop_query_levels: + return true; + case nir_texop_tex: + case nir_texop_txb: + case nir_texop_txl: + case nir_texop_txd: + case nir_texop_txf: + case nir_texop_txf_ms: + case nir_texop_tg4: + return false; + default: + unreachable("Invalid texture opcode"); + } +} + +static inline unsigned +nir_tex_instr_src_size(nir_tex_instr *instr, unsigned src) +{ + if (instr->src[src].src_type == nir_tex_src_coord) + return instr->coord_components; + + + if (instr->src[src].src_type == nir_tex_src_offset || + instr->src[src].src_type == nir_tex_src_ddx || + instr->src[src].src_type == nir_tex_src_ddy) { + if (instr->is_array) + return instr->coord_components - 1; + else + return instr->coord_components; + } + + return 1; +} + +static inline int +nir_tex_instr_src_index(nir_tex_instr *instr, nir_tex_src_type type) +{ + for (unsigned i = 0; i < instr->num_srcs; i++) + if (instr->src[i].src_type == type) + return (int) i; + + return -1; +} + +typedef struct { + union { + float f[4]; + int32_t i[4]; + uint32_t u[4]; + }; +} nir_const_value; + +typedef struct { + nir_instr instr; + + nir_const_value value; + + nir_ssa_def def; +} nir_load_const_instr; + +typedef enum { + nir_jump_return, + nir_jump_break, + nir_jump_continue, +} nir_jump_type; + +typedef struct { + nir_instr instr; + nir_jump_type type; +} nir_jump_instr; + +/* creates a new SSA variable in an undefined state */ + +typedef struct { + nir_instr instr; + nir_ssa_def def; +} nir_ssa_undef_instr; + +typedef struct { + struct exec_node node; + + /* The predecessor block corresponding to this source */ + struct nir_block *pred; + + nir_src src; +} nir_phi_src; + +#define nir_foreach_phi_src(phi, entry) \ + foreach_list_typed(nir_phi_src, entry, node, &(phi)->srcs) +#define nir_foreach_phi_src_safe(phi, entry) \ + foreach_list_typed_safe(nir_phi_src, entry, node, &(phi)->srcs) + +typedef struct { + nir_instr instr; + + struct exec_list srcs; /** < list of nir_phi_src */ + + nir_dest dest; +} nir_phi_instr; + +typedef struct { + struct exec_node node; + nir_src src; + nir_dest dest; +} nir_parallel_copy_entry; + +#define nir_foreach_parallel_copy_entry(pcopy, entry) \ + foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries) + +typedef struct { + nir_instr instr; + + /* A list of nir_parallel_copy_entry's. The sources of all of the + * entries are copied to the corresponding destinations "in parallel". + * In other words, if we have two entries: a -> b and b -> a, the values + * get swapped. + */ + struct exec_list entries; +} nir_parallel_copy_instr; + +NIR_DEFINE_CAST(nir_instr_as_alu, nir_instr, nir_alu_instr, instr) +NIR_DEFINE_CAST(nir_instr_as_call, nir_instr, nir_call_instr, instr) +NIR_DEFINE_CAST(nir_instr_as_jump, nir_instr, nir_jump_instr, instr) +NIR_DEFINE_CAST(nir_instr_as_tex, nir_instr, nir_tex_instr, instr) +NIR_DEFINE_CAST(nir_instr_as_intrinsic, nir_instr, nir_intrinsic_instr, instr) +NIR_DEFINE_CAST(nir_instr_as_load_const, nir_instr, nir_load_const_instr, instr) +NIR_DEFINE_CAST(nir_instr_as_ssa_undef, nir_instr, nir_ssa_undef_instr, instr) +NIR_DEFINE_CAST(nir_instr_as_phi, nir_instr, nir_phi_instr, instr) +NIR_DEFINE_CAST(nir_instr_as_parallel_copy, nir_instr, + nir_parallel_copy_instr, instr) + +/* + * Control flow + * + * Control flow consists of a tree of control flow nodes, which include + * if-statements and loops. The leaves of the tree are basic blocks, lists of + * instructions that always run start-to-finish. Each basic block also keeps + * track of its successors (blocks which may run immediately after the current + * block) and predecessors (blocks which could have run immediately before the + * current block). Each function also has a start block and an end block which + * all return statements point to (which is always empty). Together, all the + * blocks with their predecessors and successors make up the control flow + * graph (CFG) of the function. There are helpers that modify the tree of + * control flow nodes while modifying the CFG appropriately; these should be + * used instead of modifying the tree directly. + */ + +typedef enum { + nir_cf_node_block, + nir_cf_node_if, + nir_cf_node_loop, + nir_cf_node_function +} nir_cf_node_type; + +typedef struct nir_cf_node { + struct exec_node node; + nir_cf_node_type type; + struct nir_cf_node *parent; +} nir_cf_node; + +typedef struct nir_block { + nir_cf_node cf_node; + + struct exec_list instr_list; /** < list of nir_instr */ + + /** generic block index; generated by nir_index_blocks */ + unsigned index; + + /* + * Each block can only have up to 2 successors, so we put them in a simple + * array - no need for anything more complicated. + */ + struct nir_block *successors[2]; + + /* Set of nir_block predecessors in the CFG */ + struct set *predecessors; + + /* + * this node's immediate dominator in the dominance tree - set to NULL for + * the start block. + */ + struct nir_block *imm_dom; + + /* This node's children in the dominance tree */ + unsigned num_dom_children; + struct nir_block **dom_children; + + /* Set of nir_block's on the dominance frontier of this block */ + struct set *dom_frontier; + + /* + * These two indices have the property that dom_{pre,post}_index for each + * child of this block in the dominance tree will always be between + * dom_pre_index and dom_post_index for this block, which makes testing if + * a given block is dominated by another block an O(1) operation. + */ + unsigned dom_pre_index, dom_post_index; + + /* live in and out for this block; used for liveness analysis */ + BITSET_WORD *live_in; + BITSET_WORD *live_out; +} nir_block; + +static inline nir_instr * +nir_block_first_instr(nir_block *block) +{ + struct exec_node *head = exec_list_get_head(&block->instr_list); + return exec_node_data(nir_instr, head, node); +} + +static inline nir_instr * +nir_block_last_instr(nir_block *block) +{ + struct exec_node *tail = exec_list_get_tail(&block->instr_list); + return exec_node_data(nir_instr, tail, node); +} + +#define nir_foreach_instr(block, instr) \ + foreach_list_typed(nir_instr, instr, node, &(block)->instr_list) +#define nir_foreach_instr_reverse(block, instr) \ + foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list) +#define nir_foreach_instr_safe(block, instr) \ + foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list) +#define nir_foreach_instr_reverse_safe(block, instr) \ + foreach_list_typed_reverse_safe(nir_instr, instr, node, &(block)->instr_list) + +typedef struct nir_if { + nir_cf_node cf_node; + nir_src condition; + + struct exec_list then_list; /** < list of nir_cf_node */ + struct exec_list else_list; /** < list of nir_cf_node */ +} nir_if; + +static inline nir_cf_node * +nir_if_first_then_node(nir_if *if_stmt) +{ + struct exec_node *head = exec_list_get_head(&if_stmt->then_list); + return exec_node_data(nir_cf_node, head, node); +} + +static inline nir_cf_node * +nir_if_last_then_node(nir_if *if_stmt) +{ + struct exec_node *tail = exec_list_get_tail(&if_stmt->then_list); + return exec_node_data(nir_cf_node, tail, node); +} + +static inline nir_cf_node * +nir_if_first_else_node(nir_if *if_stmt) +{ + struct exec_node *head = exec_list_get_head(&if_stmt->else_list); + return exec_node_data(nir_cf_node, head, node); +} + +static inline nir_cf_node * +nir_if_last_else_node(nir_if *if_stmt) +{ + struct exec_node *tail = exec_list_get_tail(&if_stmt->else_list); + return exec_node_data(nir_cf_node, tail, node); +} + +typedef struct { + nir_cf_node cf_node; + + struct exec_list body; /** < list of nir_cf_node */ +} nir_loop; + +static inline nir_cf_node * +nir_loop_first_cf_node(nir_loop *loop) +{ + return exec_node_data(nir_cf_node, exec_list_get_head(&loop->body), node); +} + +static inline nir_cf_node * +nir_loop_last_cf_node(nir_loop *loop) +{ + return exec_node_data(nir_cf_node, exec_list_get_tail(&loop->body), node); +} + +/** + * Various bits of metadata that can may be created or required by + * optimization and analysis passes + */ +typedef enum { + nir_metadata_none = 0x0, + nir_metadata_block_index = 0x1, + nir_metadata_dominance = 0x2, + nir_metadata_live_ssa_defs = 0x4, + nir_metadata_not_properly_reset = 0x8, +} nir_metadata; + +typedef struct { + nir_cf_node cf_node; + + /** pointer to the function of which this is an implementation */ + struct nir_function *function; + + struct exec_list body; /** < list of nir_cf_node */ + + nir_block *end_block; + + /** list for all local variables in the function */ + struct exec_list locals; + + /** array of variables used as parameters */ + unsigned num_params; + nir_variable **params; + + /** variable used to hold the result of the function */ + nir_variable *return_var; + + /** list of local registers in the function */ + struct exec_list registers; + + /** next available local register index */ + unsigned reg_alloc; + + /** next available SSA value index */ + unsigned ssa_alloc; + + /* total number of basic blocks, only valid when block_index_dirty = false */ + unsigned num_blocks; + + nir_metadata valid_metadata; +} nir_function_impl; + +static inline nir_block * +nir_start_block(nir_function_impl *impl) +{ + return (nir_block *) exec_list_get_head(&impl->body); +} + +static inline nir_cf_node * +nir_cf_node_next(nir_cf_node *node) +{ + struct exec_node *next = exec_node_get_next(&node->node); + if (exec_node_is_tail_sentinel(next)) + return NULL; + else + return exec_node_data(nir_cf_node, next, node); +} + +static inline nir_cf_node * +nir_cf_node_prev(nir_cf_node *node) +{ + struct exec_node *prev = exec_node_get_prev(&node->node); + if (exec_node_is_head_sentinel(prev)) + return NULL; + else + return exec_node_data(nir_cf_node, prev, node); +} + +static inline bool +nir_cf_node_is_first(const nir_cf_node *node) +{ + return exec_node_is_head_sentinel(node->node.prev); +} + +static inline bool +nir_cf_node_is_last(const nir_cf_node *node) +{ + return exec_node_is_tail_sentinel(node->node.next); +} + +NIR_DEFINE_CAST(nir_cf_node_as_block, nir_cf_node, nir_block, cf_node) +NIR_DEFINE_CAST(nir_cf_node_as_if, nir_cf_node, nir_if, cf_node) +NIR_DEFINE_CAST(nir_cf_node_as_loop, nir_cf_node, nir_loop, cf_node) +NIR_DEFINE_CAST(nir_cf_node_as_function, nir_cf_node, nir_function_impl, cf_node) + +typedef enum { + nir_parameter_in, + nir_parameter_out, + nir_parameter_inout, +} nir_parameter_type; + +typedef struct { + nir_parameter_type param_type; + const struct glsl_type *type; +} nir_parameter; + +typedef struct nir_function { + struct exec_node node; + + const char *name; + struct nir_shader *shader; + + unsigned num_params; + nir_parameter *params; + const struct glsl_type *return_type; + + /** The implementation of this function. + * + * If the function is only declared and not implemented, this is NULL. + */ + nir_function_impl *impl; +} nir_function; + +typedef struct nir_shader_compiler_options { + bool lower_fdiv; + bool lower_ffma; + bool lower_flrp; + bool lower_fpow; + bool lower_fsat; + bool lower_fsqrt; + bool lower_fmod; + bool lower_bitfield_extract; + bool lower_bitfield_insert; + bool lower_uadd_carry; + bool lower_usub_borrow; + /** lowers fneg and ineg to fsub and isub. */ + bool lower_negate; + /** lowers fsub and isub to fadd+fneg and iadd+ineg. */ + bool lower_sub; + + /* lower {slt,sge,seq,sne} to {flt,fge,feq,fne} + b2f: */ + bool lower_scmp; + + /* Does the native fdot instruction replicate its result for four + * components? If so, then opt_algebraic_late will turn all fdotN + * instructions into fdot_replicatedN instructions. + */ + bool fdot_replicates; + + /** lowers ffract to fsub+ffloor: */ + bool lower_ffract; + + /** + * Does the driver support real 32-bit integers? (Otherwise, integers + * are simulated by floats.) + */ + bool native_integers; +} nir_shader_compiler_options; + +typedef struct nir_shader_info { + const char *name; + + /* Descriptive name provided by the client; may be NULL */ + const char *label; + + /* Number of textures used by this shader */ + unsigned num_textures; + /* Number of uniform buffers used by this shader */ + unsigned num_ubos; + /* Number of atomic buffers used by this shader */ + unsigned num_abos; + /* Number of shader storage buffers used by this shader */ + unsigned num_ssbos; + /* Number of images used by this shader */ + unsigned num_images; + + /* Which inputs are actually read */ + uint64_t inputs_read; + /* Which outputs are actually written */ + uint64_t outputs_written; + /* Which system values are actually read */ + uint64_t system_values_read; + + /* Which patch inputs are actually read */ + uint32_t patch_inputs_read; + /* Which patch outputs are actually written */ + uint32_t patch_outputs_written; + + /* Whether or not this shader ever uses textureGather() */ + bool uses_texture_gather; + + /* Whether or not this shader uses the gl_ClipDistance output */ + bool uses_clip_distance_out; + + /* Whether or not separate shader objects were used */ + bool separate_shader; + + /** Was this shader linked with any transform feedback varyings? */ + bool has_transform_feedback_varyings; + + union { + struct { + /** The number of vertices recieves per input primitive */ + unsigned vertices_in; + + /** The output primitive type (GL enum value) */ + unsigned output_primitive; + + /** The maximum number of vertices the geometry shader might write. */ + unsigned vertices_out; + + /** 1 .. MAX_GEOMETRY_SHADER_INVOCATIONS */ + unsigned invocations; + + /** Whether or not this shader uses EndPrimitive */ + bool uses_end_primitive; + + /** Whether or not this shader uses non-zero streams */ + bool uses_streams; + } gs; + + struct { + bool uses_discard; + + /** + * Whether early fragment tests are enabled as defined by + * ARB_shader_image_load_store. + */ + bool early_fragment_tests; + + /** gl_FragDepth layout for ARB_conservative_depth. */ + enum gl_frag_depth_layout depth_layout; + } fs; + + struct { + unsigned local_size[3]; + } cs; + + struct { + /** The number of vertices in the TCS output patch. */ + unsigned vertices_out; + } tcs; + }; +} nir_shader_info; + +typedef struct nir_shader { + /** list of uniforms (nir_variable) */ + struct exec_list uniforms; + + /** list of inputs (nir_variable) */ + struct exec_list inputs; + + /** list of outputs (nir_variable) */ + struct exec_list outputs; + + /** Set of driver-specific options for the shader. + * + * The memory for the options is expected to be kept in a single static + * copy by the driver. + */ + const struct nir_shader_compiler_options *options; + + /** Various bits of compile-time information about a given shader */ + struct nir_shader_info info; + + /** list of global variables in the shader (nir_variable) */ + struct exec_list globals; + + /** list of system value variables in the shader (nir_variable) */ + struct exec_list system_values; + + struct exec_list functions; /** < list of nir_function */ + + /** list of global register in the shader */ + struct exec_list registers; + + /** next available global register index */ + unsigned reg_alloc; + + /** + * the highest index a load_input_*, load_uniform_*, etc. intrinsic can + * access plus one + */ + unsigned num_inputs, num_uniforms, num_outputs; + + /** The shader stage, such as MESA_SHADER_VERTEX. */ + gl_shader_stage stage; +} nir_shader; + +#define nir_foreach_function(shader, func) \ + foreach_list_typed(nir_function, func, node, &(shader)->functions) + +nir_shader *nir_shader_create(void *mem_ctx, + gl_shader_stage stage, + const nir_shader_compiler_options *options); + +/** creates a register, including assigning it an index and adding it to the list */ +nir_register *nir_global_reg_create(nir_shader *shader); + +nir_register *nir_local_reg_create(nir_function_impl *impl); + +void nir_reg_remove(nir_register *reg); + +/** Adds a variable to the appropreate list in nir_shader */ +void nir_shader_add_variable(nir_shader *shader, nir_variable *var); + +static inline void +nir_function_impl_add_variable(nir_function_impl *impl, nir_variable *var) +{ + assert(var->data.mode == nir_var_local); + exec_list_push_tail(&impl->locals, &var->node); +} + +/** creates a variable, sets a few defaults, and adds it to the list */ +nir_variable *nir_variable_create(nir_shader *shader, + nir_variable_mode mode, + const struct glsl_type *type, + const char *name); +/** creates a local variable and adds it to the list */ +nir_variable *nir_local_variable_create(nir_function_impl *impl, + const struct glsl_type *type, + const char *name); + +/** creates a function and adds it to the shader's list of functions */ +nir_function *nir_function_create(nir_shader *shader, const char *name); + +nir_function_impl *nir_function_impl_create(nir_function *func); + +nir_block *nir_block_create(nir_shader *shader); +nir_if *nir_if_create(nir_shader *shader); +nir_loop *nir_loop_create(nir_shader *shader); + +nir_function_impl *nir_cf_node_get_function(nir_cf_node *node); + +/** requests that the given pieces of metadata be generated */ +void nir_metadata_require(nir_function_impl *impl, nir_metadata required); +/** dirties all but the preserved metadata */ +void nir_metadata_preserve(nir_function_impl *impl, nir_metadata preserved); + +/** creates an instruction with default swizzle/writemask/etc. with NULL registers */ +nir_alu_instr *nir_alu_instr_create(nir_shader *shader, nir_op op); + +nir_jump_instr *nir_jump_instr_create(nir_shader *shader, nir_jump_type type); + +nir_load_const_instr *nir_load_const_instr_create(nir_shader *shader, + unsigned num_components); + +nir_intrinsic_instr *nir_intrinsic_instr_create(nir_shader *shader, + nir_intrinsic_op op); + +nir_call_instr *nir_call_instr_create(nir_shader *shader, + nir_function *callee); + +nir_tex_instr *nir_tex_instr_create(nir_shader *shader, unsigned num_srcs); + +nir_phi_instr *nir_phi_instr_create(nir_shader *shader); + +nir_parallel_copy_instr *nir_parallel_copy_instr_create(nir_shader *shader); + +nir_ssa_undef_instr *nir_ssa_undef_instr_create(nir_shader *shader, + unsigned num_components); + +nir_deref_var *nir_deref_var_create(void *mem_ctx, nir_variable *var); +nir_deref_array *nir_deref_array_create(void *mem_ctx); +nir_deref_struct *nir_deref_struct_create(void *mem_ctx, unsigned field_index); + +nir_deref *nir_copy_deref(void *mem_ctx, nir_deref *deref); + +nir_load_const_instr * +nir_deref_get_const_initializer_load(nir_shader *shader, nir_deref_var *deref); + +/** + * NIR Cursors and Instruction Insertion API + * @{ + * + * A tiny struct representing a point to insert/extract instructions or + * control flow nodes. Helps reduce the combinatorial explosion of possible + * points to insert/extract. + * + * \sa nir_control_flow.h + */ +typedef enum { + nir_cursor_before_block, + nir_cursor_after_block, + nir_cursor_before_instr, + nir_cursor_after_instr, +} nir_cursor_option; + +typedef struct { + nir_cursor_option option; + union { + nir_block *block; + nir_instr *instr; + }; +} nir_cursor; + +static inline nir_cursor +nir_before_block(nir_block *block) +{ + nir_cursor cursor; + cursor.option = nir_cursor_before_block; + cursor.block = block; + return cursor; +} + +static inline nir_cursor +nir_after_block(nir_block *block) +{ + nir_cursor cursor; + cursor.option = nir_cursor_after_block; + cursor.block = block; + return cursor; +} + +static inline nir_cursor +nir_before_instr(nir_instr *instr) +{ + nir_cursor cursor; + cursor.option = nir_cursor_before_instr; + cursor.instr = instr; + return cursor; +} + +static inline nir_cursor +nir_after_instr(nir_instr *instr) +{ + nir_cursor cursor; + cursor.option = nir_cursor_after_instr; + cursor.instr = instr; + return cursor; +} + +static inline nir_cursor +nir_after_block_before_jump(nir_block *block) +{ + nir_instr *last_instr = nir_block_last_instr(block); + if (last_instr && last_instr->type == nir_instr_type_jump) { + return nir_before_instr(last_instr); + } else { + return nir_after_block(block); + } +} + +static inline nir_cursor +nir_before_cf_node(nir_cf_node *node) +{ + if (node->type == nir_cf_node_block) + return nir_before_block(nir_cf_node_as_block(node)); + + return nir_after_block(nir_cf_node_as_block(nir_cf_node_prev(node))); +} + +static inline nir_cursor +nir_after_cf_node(nir_cf_node *node) +{ + if (node->type == nir_cf_node_block) + return nir_after_block(nir_cf_node_as_block(node)); + + return nir_before_block(nir_cf_node_as_block(nir_cf_node_next(node))); +} + +static inline nir_cursor +nir_before_cf_list(struct exec_list *cf_list) +{ + nir_cf_node *first_node = exec_node_data(nir_cf_node, + exec_list_get_head(cf_list), node); + return nir_before_cf_node(first_node); +} + +static inline nir_cursor +nir_after_cf_list(struct exec_list *cf_list) +{ + nir_cf_node *last_node = exec_node_data(nir_cf_node, + exec_list_get_tail(cf_list), node); + return nir_after_cf_node(last_node); +} + +/** + * Insert a NIR instruction at the given cursor. + * + * Note: This does not update the cursor. + */ +void nir_instr_insert(nir_cursor cursor, nir_instr *instr); + +static inline void +nir_instr_insert_before(nir_instr *instr, nir_instr *before) +{ + nir_instr_insert(nir_before_instr(instr), before); +} + +static inline void +nir_instr_insert_after(nir_instr *instr, nir_instr *after) +{ + nir_instr_insert(nir_after_instr(instr), after); +} + +static inline void +nir_instr_insert_before_block(nir_block *block, nir_instr *before) +{ + nir_instr_insert(nir_before_block(block), before); +} + +static inline void +nir_instr_insert_after_block(nir_block *block, nir_instr *after) +{ + nir_instr_insert(nir_after_block(block), after); +} + +static inline void +nir_instr_insert_before_cf(nir_cf_node *node, nir_instr *before) +{ + nir_instr_insert(nir_before_cf_node(node), before); +} + +static inline void +nir_instr_insert_after_cf(nir_cf_node *node, nir_instr *after) +{ + nir_instr_insert(nir_after_cf_node(node), after); +} + +static inline void +nir_instr_insert_before_cf_list(struct exec_list *list, nir_instr *before) +{ + nir_instr_insert(nir_before_cf_list(list), before); +} + +static inline void +nir_instr_insert_after_cf_list(struct exec_list *list, nir_instr *after) +{ + nir_instr_insert(nir_after_cf_list(list), after); +} + +void nir_instr_remove(nir_instr *instr); + +/** @} */ + +typedef bool (*nir_foreach_ssa_def_cb)(nir_ssa_def *def, void *state); +typedef bool (*nir_foreach_dest_cb)(nir_dest *dest, void *state); +typedef bool (*nir_foreach_src_cb)(nir_src *src, void *state); +bool nir_foreach_ssa_def(nir_instr *instr, nir_foreach_ssa_def_cb cb, + void *state); +bool nir_foreach_dest(nir_instr *instr, nir_foreach_dest_cb cb, void *state); +bool nir_foreach_src(nir_instr *instr, nir_foreach_src_cb cb, void *state); + +nir_const_value *nir_src_as_const_value(nir_src src); +bool nir_src_is_dynamically_uniform(nir_src src); +bool nir_srcs_equal(nir_src src1, nir_src src2); +void nir_instr_rewrite_src(nir_instr *instr, nir_src *src, nir_src new_src); +void nir_instr_move_src(nir_instr *dest_instr, nir_src *dest, nir_src *src); +void nir_if_rewrite_condition(nir_if *if_stmt, nir_src new_src); +void nir_instr_rewrite_dest(nir_instr *instr, nir_dest *dest, + nir_dest new_dest); + +void nir_ssa_dest_init(nir_instr *instr, nir_dest *dest, + unsigned num_components, const char *name); +void nir_ssa_def_init(nir_instr *instr, nir_ssa_def *def, + unsigned num_components, const char *name); +void nir_ssa_def_rewrite_uses(nir_ssa_def *def, nir_src new_src); +void nir_ssa_def_rewrite_uses_after(nir_ssa_def *def, nir_src new_src, + nir_instr *after_me); + +/* visits basic blocks in source-code order */ +typedef bool (*nir_foreach_block_cb)(nir_block *block, void *state); +bool nir_foreach_block(nir_function_impl *impl, nir_foreach_block_cb cb, + void *state); +bool nir_foreach_block_reverse(nir_function_impl *impl, nir_foreach_block_cb cb, + void *state); +bool nir_foreach_block_in_cf_node(nir_cf_node *node, nir_foreach_block_cb cb, + void *state); + +/* If the following CF node is an if, this function returns that if. + * Otherwise, it returns NULL. + */ +nir_if *nir_block_get_following_if(nir_block *block); + +nir_loop *nir_block_get_following_loop(nir_block *block); + +void nir_index_local_regs(nir_function_impl *impl); +void nir_index_global_regs(nir_shader *shader); +void nir_index_ssa_defs(nir_function_impl *impl); +unsigned nir_index_instrs(nir_function_impl *impl); + +void nir_index_blocks(nir_function_impl *impl); + +void nir_print_shader(nir_shader *shader, FILE *fp); +void nir_print_instr(const nir_instr *instr, FILE *fp); + +nir_shader * nir_shader_clone(void *mem_ctx, const nir_shader *s); + +#ifdef DEBUG +void nir_validate_shader(nir_shader *shader); +void nir_metadata_set_validation_flag(nir_shader *shader); +void nir_metadata_check_validation_flag(nir_shader *shader); + +#include "util/debug.h" +static inline bool +should_clone_nir(void) +{ + static int should_clone = -1; + if (should_clone < 0) + should_clone = env_var_as_boolean("NIR_TEST_CLONE", false); + + return should_clone; +} +#else +static inline void nir_validate_shader(nir_shader *shader) { (void) shader; } +static inline void nir_metadata_set_validation_flag(nir_shader *shader) { (void) shader; } +static inline void nir_metadata_check_validation_flag(nir_shader *shader) { (void) shader; } +static inline bool should_clone_nir(void) { return false; } +#endif /* DEBUG */ + +#define _PASS(nir, do_pass) do { \ + do_pass \ + nir_validate_shader(nir); \ + if (should_clone_nir()) { \ + nir_shader *clone = nir_shader_clone(ralloc_parent(nir), nir); \ + ralloc_free(nir); \ + nir = clone; \ + } \ +} while (0) + +#define NIR_PASS(progress, nir, pass, ...) _PASS(nir, \ + nir_metadata_set_validation_flag(nir); \ + if (pass(nir, ##__VA_ARGS__)) { \ + progress = true; \ + nir_metadata_check_validation_flag(nir); \ + } \ +) + +#define NIR_PASS_V(nir, pass, ...) _PASS(nir, \ + pass(nir, ##__VA_ARGS__); \ +) + +void nir_calc_dominance_impl(nir_function_impl *impl); +void nir_calc_dominance(nir_shader *shader); + +nir_block *nir_dominance_lca(nir_block *b1, nir_block *b2); +bool nir_block_dominates(nir_block *parent, nir_block *child); + +void nir_dump_dom_tree_impl(nir_function_impl *impl, FILE *fp); +void nir_dump_dom_tree(nir_shader *shader, FILE *fp); + +void nir_dump_dom_frontier_impl(nir_function_impl *impl, FILE *fp); +void nir_dump_dom_frontier(nir_shader *shader, FILE *fp); + +void nir_dump_cfg_impl(nir_function_impl *impl, FILE *fp); +void nir_dump_cfg(nir_shader *shader, FILE *fp); + +int nir_gs_count_vertices(const nir_shader *shader); + +bool nir_split_var_copies(nir_shader *shader); + +void nir_lower_var_copy_instr(nir_intrinsic_instr *copy, void *mem_ctx); +void nir_lower_var_copies(nir_shader *shader); + +bool nir_lower_global_vars_to_local(nir_shader *shader); + +bool nir_lower_locals_to_regs(nir_shader *shader); + +void nir_lower_outputs_to_temporaries(nir_shader *shader); + +void nir_assign_var_locations(struct exec_list *var_list, + unsigned *size, + int (*type_size)(const struct glsl_type *)); + +void nir_lower_io(nir_shader *shader, + nir_variable_mode mode, + int (*type_size)(const struct glsl_type *)); +nir_src *nir_get_io_offset_src(nir_intrinsic_instr *instr); +nir_src *nir_get_io_vertex_index_src(nir_intrinsic_instr *instr); + +void nir_lower_vars_to_ssa(nir_shader *shader); + +bool nir_remove_dead_variables(nir_shader *shader); + +void nir_move_vec_src_uses_to_dest(nir_shader *shader); +bool nir_lower_vec_to_movs(nir_shader *shader); +void nir_lower_alu_to_scalar(nir_shader *shader); +void nir_lower_load_const_to_scalar(nir_shader *shader); + +void nir_lower_phis_to_scalar(nir_shader *shader); + +void nir_lower_samplers(nir_shader *shader, + const struct gl_shader_program *shader_program); + +bool nir_lower_system_values(nir_shader *shader); + +typedef struct nir_lower_tex_options { + /** + * bitmask of (1 << GLSL_SAMPLER_DIM_x) to control for which + * sampler types a texture projector is lowered. + */ + unsigned lower_txp; + + /** + * If true, lower rect textures to 2D, using txs to fetch the + * texture dimensions and dividing the texture coords by the + * texture dims to normalize. + */ + bool lower_rect; + + /** + * To emulate certain texture wrap modes, this can be used + * to saturate the specified tex coord to [0.0, 1.0]. The + * bits are according to sampler #, ie. if, for example: + * + * (conf->saturate_s & (1 << n)) + * + * is true, then the s coord for sampler n is saturated. + * + * Note that clamping must happen *after* projector lowering + * so any projected texture sample instruction with a clamped + * coordinate gets automatically lowered, regardless of the + * 'lower_txp' setting. + */ + unsigned saturate_s; + unsigned saturate_t; + unsigned saturate_r; + + /* Bitmask of samplers that need swizzling. + * + * If (swizzle_result & (1 << sampler_index)), then the swizzle in + * swizzles[sampler_index] is applied to the result of the texturing + * operation. + */ + unsigned swizzle_result; + + /* A swizzle for each sampler. Values 0-3 represent x, y, z, or w swizzles + * while 4 and 5 represent 0 and 1 respectively. + */ + uint8_t swizzles[32][4]; +} nir_lower_tex_options; + +bool nir_lower_tex(nir_shader *shader, + const nir_lower_tex_options *options); + +void nir_lower_idiv(nir_shader *shader); + +void nir_lower_clip_vs(nir_shader *shader, unsigned ucp_enables); +void nir_lower_clip_fs(nir_shader *shader, unsigned ucp_enables); + +void nir_lower_two_sided_color(nir_shader *shader); + +void nir_lower_atomics(nir_shader *shader, + const struct gl_shader_program *shader_program); +void nir_lower_to_source_mods(nir_shader *shader); + +bool nir_lower_gs_intrinsics(nir_shader *shader); + +bool nir_normalize_cubemap_coords(nir_shader *shader); + +void nir_live_ssa_defs_impl(nir_function_impl *impl); +bool nir_ssa_defs_interfere(nir_ssa_def *a, nir_ssa_def *b); + +void nir_convert_to_ssa_impl(nir_function_impl *impl); +void nir_convert_to_ssa(nir_shader *shader); + +/* If phi_webs_only is true, only convert SSA values involved in phi nodes to + * registers. If false, convert all values (even those not involved in a phi + * node) to registers. + */ +void nir_convert_from_ssa(nir_shader *shader, bool phi_webs_only); + +bool nir_opt_algebraic(nir_shader *shader); +bool nir_opt_algebraic_late(nir_shader *shader); +bool nir_opt_constant_folding(nir_shader *shader); + +bool nir_opt_global_to_local(nir_shader *shader); + +bool nir_copy_prop(nir_shader *shader); + +bool nir_opt_cse(nir_shader *shader); + +bool nir_opt_dce(nir_shader *shader); + +bool nir_opt_dead_cf(nir_shader *shader); + +void nir_opt_gcm(nir_shader *shader); + +bool nir_opt_peephole_select(nir_shader *shader); + +bool nir_opt_remove_phis(nir_shader *shader); + +bool nir_opt_undef(nir_shader *shader); + +void nir_sweep(nir_shader *shader); + +nir_intrinsic_op nir_intrinsic_from_system_value(gl_system_value val); +gl_system_value nir_system_value_from_intrinsic(nir_intrinsic_op intrin); + +#ifdef __cplusplus +} /* extern "C" */ +#endif diff --git a/src/compiler/nir/nir_algebraic.py b/src/compiler/nir/nir_algebraic.py new file mode 100644 index 0000000..a30652f --- /dev/null +++ b/src/compiler/nir/nir_algebraic.py @@ -0,0 +1,305 @@ +#! /usr/bin/env python +# +# Copyright (C) 2014 Intel Corporation +# +# Permission is hereby granted, free of charge, to any person obtaining a +# copy of this software and associated documentation files (the "Software"), +# to deal in the Software without restriction, including without limitation +# the rights to use, copy, modify, merge, publish, distribute, sublicense, +# and/or sell copies of the Software, and to permit persons to whom the +# Software is furnished to do so, subject to the following conditions: +# +# The above copyright notice and this permission notice (including the next +# paragraph) shall be included in all copies or substantial portions of the +# Software. +# +# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL +# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +# IN THE SOFTWARE. +# +# Authors: +# Jason Ekstrand (jason@jlekstrand.net) + +import itertools +import struct +import sys +import mako.template +import re + +# Represents a set of variables, each with a unique id +class VarSet(object): + def __init__(self): + self.names = {} + self.ids = itertools.count() + self.immutable = False; + + def __getitem__(self, name): + if name not in self.names: + assert not self.immutable, "Unknown replacement variable: " + name + self.names[name] = self.ids.next() + + return self.names[name] + + def lock(self): + self.immutable = True + +class Value(object): + @staticmethod + def create(val, name_base, varset): + if isinstance(val, tuple): + return Expression(val, name_base, varset) + elif isinstance(val, Expression): + return val + elif isinstance(val, (str, unicode)): + return Variable(val, name_base, varset) + elif isinstance(val, (bool, int, long, float)): + return Constant(val, name_base) + + __template = mako.template.Template(""" +static const ${val.c_type} ${val.name} = { + { ${val.type_enum} }, +% if isinstance(val, Constant): + { ${hex(val)} /* ${val.value} */ }, +% elif isinstance(val, Variable): + ${val.index}, /* ${val.var_name} */ + ${'true' if val.is_constant else 'false'}, + nir_type_${ val.required_type or 'invalid' }, +% elif isinstance(val, Expression): + nir_op_${val.opcode}, + { ${', '.join(src.c_ptr for src in val.sources)} }, +% endif +};""") + + def __init__(self, name, type_str): + self.name = name + self.type_str = type_str + + @property + def type_enum(self): + return "nir_search_value_" + self.type_str + + @property + def c_type(self): + return "nir_search_" + self.type_str + + @property + def c_ptr(self): + return "&{0}.value".format(self.name) + + def render(self): + return self.__template.render(val=self, + Constant=Constant, + Variable=Variable, + Expression=Expression) + +class Constant(Value): + def __init__(self, val, name): + Value.__init__(self, name, "constant") + self.value = val + + def __hex__(self): + # Even if it's an integer, we still need to unpack as an unsigned + # int. This is because, without C99, we can only assign to the first + # element of a union in an initializer. + if isinstance(self.value, (bool)): + return 'NIR_TRUE' if self.value else 'NIR_FALSE' + if isinstance(self.value, (int, long)): + return hex(struct.unpack('I', struct.pack('i', self.value))[0]) + elif isinstance(self.value, float): + return hex(struct.unpack('I', struct.pack('f', self.value))[0]) + else: + assert False + +_var_name_re = re.compile(r"(?P<const>#)?(?P<name>\w+)(?:@(?P<type>\w+))?") + +class Variable(Value): + def __init__(self, val, name, varset): + Value.__init__(self, name, "variable") + + m = _var_name_re.match(val) + assert m and m.group('name') is not None + + self.var_name = m.group('name') + self.is_constant = m.group('const') is not None + self.required_type = m.group('type') + + if self.required_type is not None: + assert self.required_type in ('float', 'bool', 'int', 'unsigned') + + self.index = varset[self.var_name] + +class Expression(Value): + def __init__(self, expr, name_base, varset): + Value.__init__(self, name_base, "expression") + assert isinstance(expr, tuple) + + self.opcode = expr[0] + self.sources = [ Value.create(src, "{0}_{1}".format(name_base, i), varset) + for (i, src) in enumerate(expr[1:]) ] + + def render(self): + srcs = "\n".join(src.render() for src in self.sources) + return srcs + super(Expression, self).render() + +_optimization_ids = itertools.count() + +condition_list = ['true'] + +class SearchAndReplace(object): + def __init__(self, transform): + self.id = _optimization_ids.next() + + search = transform[0] + replace = transform[1] + if len(transform) > 2: + self.condition = transform[2] + else: + self.condition = 'true' + + if self.condition not in condition_list: + condition_list.append(self.condition) + self.condition_index = condition_list.index(self.condition) + + varset = VarSet() + if isinstance(search, Expression): + self.search = search + else: + self.search = Expression(search, "search{0}".format(self.id), varset) + + varset.lock() + + if isinstance(replace, Value): + self.replace = replace + else: + self.replace = Value.create(replace, "replace{0}".format(self.id), varset) + +_algebraic_pass_template = mako.template.Template(""" +#include "nir.h" +#include "nir_search.h" + +#ifndef NIR_OPT_ALGEBRAIC_STRUCT_DEFS +#define NIR_OPT_ALGEBRAIC_STRUCT_DEFS + +struct transform { + const nir_search_expression *search; + const nir_search_value *replace; + unsigned condition_offset; +}; + +struct opt_state { + void *mem_ctx; + bool progress; + const bool *condition_flags; +}; + +#endif + +% for (opcode, xform_list) in xform_dict.iteritems(): +% for xform in xform_list: + ${xform.search.render()} + ${xform.replace.render()} +% endfor + +static const struct transform ${pass_name}_${opcode}_xforms[] = { +% for xform in xform_list: + { &${xform.search.name}, ${xform.replace.c_ptr}, ${xform.condition_index} }, +% endfor +}; +% endfor + +static bool +${pass_name}_block(nir_block *block, void *void_state) +{ + struct opt_state *state = void_state; + + nir_foreach_instr_safe(block, instr) { + if (instr->type != nir_instr_type_alu) + continue; + + nir_alu_instr *alu = nir_instr_as_alu(instr); + if (!alu->dest.dest.is_ssa) + continue; + + switch (alu->op) { + % for opcode in xform_dict.keys(): + case nir_op_${opcode}: + for (unsigned i = 0; i < ARRAY_SIZE(${pass_name}_${opcode}_xforms); i++) { + const struct transform *xform = &${pass_name}_${opcode}_xforms[i]; + if (state->condition_flags[xform->condition_offset] && + nir_replace_instr(alu, xform->search, xform->replace, + state->mem_ctx)) { + state->progress = true; + break; + } + } + break; + % endfor + default: + break; + } + } + + return true; +} + +static bool +${pass_name}_impl(nir_function_impl *impl, const bool *condition_flags) +{ + struct opt_state state; + + state.mem_ctx = ralloc_parent(impl); + state.progress = false; + state.condition_flags = condition_flags; + + nir_foreach_block(impl, ${pass_name}_block, &state); + + if (state.progress) + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance); + + return state.progress; +} + + +bool +${pass_name}(nir_shader *shader) +{ + bool progress = false; + bool condition_flags[${len(condition_list)}]; + const nir_shader_compiler_options *options = shader->options; + + % for index, condition in enumerate(condition_list): + condition_flags[${index}] = ${condition}; + % endfor + + nir_foreach_function(shader, function) { + if (function->impl) + progress |= ${pass_name}_impl(function->impl, condition_flags); + } + + return progress; +} +""") + +class AlgebraicPass(object): + def __init__(self, pass_name, transforms): + self.xform_dict = {} + self.pass_name = pass_name + + for xform in transforms: + if not isinstance(xform, SearchAndReplace): + xform = SearchAndReplace(xform) + + if xform.search.opcode not in self.xform_dict: + self.xform_dict[xform.search.opcode] = [] + + self.xform_dict[xform.search.opcode].append(xform) + + def render(self): + return _algebraic_pass_template.render(pass_name=self.pass_name, + xform_dict=self.xform_dict, + condition_list=condition_list) diff --git a/src/compiler/nir/nir_array.h b/src/compiler/nir/nir_array.h new file mode 100644 index 0000000..1db4e8c --- /dev/null +++ b/src/compiler/nir/nir_array.h @@ -0,0 +1,96 @@ +/* + * Copyright © 2015 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand (jason@jlekstrand.net) + * + */ + +#pragma once + +#ifdef __cplusplus +extern "C" { +#endif + +typedef struct { + void *mem_ctx; + size_t size; + size_t alloc; + void *data; +} nir_array; + +static inline void +nir_array_init(nir_array *arr, void *mem_ctx) +{ + arr->mem_ctx = mem_ctx; + arr->size = 0; + arr->alloc = 0; + arr->data = NULL; +} + +static inline void +nir_array_fini(nir_array *arr) +{ + if (arr->mem_ctx) + ralloc_free(arr->data); + else + free(arr->data); +} + +#define NIR_ARRAY_INITIAL_SIZE 64 + +/* Increments the size of the array by the given ammount and returns a + * pointer to the beginning of the newly added space. + */ +static inline void * +nir_array_grow(nir_array *arr, size_t additional) +{ + size_t new_size = arr->size + additional; + if (new_size > arr->alloc) { + if (arr->alloc == 0) + arr->alloc = NIR_ARRAY_INITIAL_SIZE; + + while (new_size > arr->alloc) + arr->alloc *= 2; + + if (arr->mem_ctx) + arr->data = reralloc_size(arr->mem_ctx, arr->data, arr->alloc); + else + arr->data = realloc(arr->data, arr->alloc); + } + + void *ptr = (void *)((char *)arr->data + arr->size); + arr->size = new_size; + + return ptr; +} + +#define nir_array_add(arr, type, elem) \ + *(type *)nir_array_grow(arr, sizeof(type)) = (elem) + +#define nir_array_foreach(arr, type, elem) \ + for (type *elem = (type *)(arr)->data; \ + elem < (type *)((char *)(arr)->data + (arr)->size); elem++) + +#ifdef __cplusplus +} /* extern "C" */ +#endif diff --git a/src/compiler/nir/nir_builder.h b/src/compiler/nir/nir_builder.h new file mode 100644 index 0000000..88ba3a1 --- /dev/null +++ b/src/compiler/nir/nir_builder.h @@ -0,0 +1,364 @@ +/* + * Copyright © 2014-2015 Broadcom + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + */ + +#ifndef NIR_BUILDER_H +#define NIR_BUILDER_H + +#include "nir_control_flow.h" + +struct exec_list; + +typedef struct nir_builder { + nir_cursor cursor; + + nir_shader *shader; + nir_function_impl *impl; +} nir_builder; + +static inline void +nir_builder_init(nir_builder *build, nir_function_impl *impl) +{ + memset(build, 0, sizeof(*build)); + build->impl = impl; + build->shader = impl->function->shader; +} + +static inline void +nir_builder_init_simple_shader(nir_builder *build, void *mem_ctx, + gl_shader_stage stage, + const nir_shader_compiler_options *options) +{ + build->shader = nir_shader_create(mem_ctx, stage, options); + nir_function *func = nir_function_create(build->shader, "main"); + build->impl = nir_function_impl_create(func); + build->cursor = nir_after_cf_list(&build->impl->body); +} + +static inline void +nir_builder_instr_insert(nir_builder *build, nir_instr *instr) +{ + nir_instr_insert(build->cursor, instr); + + /* Move the cursor forward. */ + build->cursor = nir_after_instr(instr); +} + +static inline void +nir_builder_cf_insert(nir_builder *build, nir_cf_node *cf) +{ + nir_cf_node_insert(build->cursor, cf); +} + +static inline nir_ssa_def * +nir_build_imm(nir_builder *build, unsigned num_components, nir_const_value value) +{ + nir_load_const_instr *load_const = + nir_load_const_instr_create(build->shader, num_components); + if (!load_const) + return NULL; + + load_const->value = value; + + nir_builder_instr_insert(build, &load_const->instr); + + return &load_const->def; +} + +static inline nir_ssa_def * +nir_imm_float(nir_builder *build, float x) +{ + nir_const_value v; + + memset(&v, 0, sizeof(v)); + v.f[0] = x; + + return nir_build_imm(build, 1, v); +} + +static inline nir_ssa_def * +nir_imm_vec4(nir_builder *build, float x, float y, float z, float w) +{ + nir_const_value v; + + memset(&v, 0, sizeof(v)); + v.f[0] = x; + v.f[1] = y; + v.f[2] = z; + v.f[3] = w; + + return nir_build_imm(build, 4, v); +} + +static inline nir_ssa_def * +nir_imm_int(nir_builder *build, int x) +{ + nir_const_value v; + + memset(&v, 0, sizeof(v)); + v.i[0] = x; + + return nir_build_imm(build, 1, v); +} + +static inline nir_ssa_def * +nir_imm_ivec4(nir_builder *build, int x, int y, int z, int w) +{ + nir_const_value v; + + memset(&v, 0, sizeof(v)); + v.i[0] = x; + v.i[1] = y; + v.i[2] = z; + v.i[3] = w; + + return nir_build_imm(build, 4, v); +} + +static inline nir_ssa_def * +nir_build_alu(nir_builder *build, nir_op op, nir_ssa_def *src0, + nir_ssa_def *src1, nir_ssa_def *src2, nir_ssa_def *src3) +{ + const nir_op_info *op_info = &nir_op_infos[op]; + nir_alu_instr *instr = nir_alu_instr_create(build->shader, op); + if (!instr) + return NULL; + + instr->src[0].src = nir_src_for_ssa(src0); + if (src1) + instr->src[1].src = nir_src_for_ssa(src1); + if (src2) + instr->src[2].src = nir_src_for_ssa(src2); + if (src3) + instr->src[3].src = nir_src_for_ssa(src3); + + /* Guess the number of components the destination temporary should have + * based on our input sizes, if it's not fixed for the op. + */ + unsigned num_components = op_info->output_size; + if (num_components == 0) { + for (unsigned i = 0; i < op_info->num_inputs; i++) { + if (op_info->input_sizes[i] == 0) + num_components = MAX2(num_components, + instr->src[i].src.ssa->num_components); + } + } + assert(num_components != 0); + + /* Make sure we don't swizzle from outside of our source vector (like if a + * scalar value was passed into a multiply with a vector). + */ + for (unsigned i = 0; i < op_info->num_inputs; i++) { + for (unsigned j = instr->src[i].src.ssa->num_components; j < 4; j++) { + instr->src[i].swizzle[j] = instr->src[i].src.ssa->num_components - 1; + } + } + + nir_ssa_dest_init(&instr->instr, &instr->dest.dest, num_components, NULL); + instr->dest.write_mask = (1 << num_components) - 1; + + nir_builder_instr_insert(build, &instr->instr); + + return &instr->dest.dest.ssa; +} + +#define ALU1(op) \ +static inline nir_ssa_def * \ +nir_##op(nir_builder *build, nir_ssa_def *src0) \ +{ \ + return nir_build_alu(build, nir_op_##op, src0, NULL, NULL, NULL); \ +} + +#define ALU2(op) \ +static inline nir_ssa_def * \ +nir_##op(nir_builder *build, nir_ssa_def *src0, nir_ssa_def *src1) \ +{ \ + return nir_build_alu(build, nir_op_##op, src0, src1, NULL, NULL); \ +} + +#define ALU3(op) \ +static inline nir_ssa_def * \ +nir_##op(nir_builder *build, nir_ssa_def *src0, \ + nir_ssa_def *src1, nir_ssa_def *src2) \ +{ \ + return nir_build_alu(build, nir_op_##op, src0, src1, src2, NULL); \ +} + +#define ALU4(op) \ +static inline nir_ssa_def * \ +nir_##op(nir_builder *build, nir_ssa_def *src0, \ + nir_ssa_def *src1, nir_ssa_def *src2, nir_ssa_def *src3) \ +{ \ + return nir_build_alu(build, nir_op_##op, src0, src1, src2, src3); \ +} + +#include "nir_builder_opcodes.h" + +static inline nir_ssa_def * +nir_vec(nir_builder *build, nir_ssa_def **comp, unsigned num_components) +{ + switch (num_components) { + case 4: + return nir_vec4(build, comp[0], comp[1], comp[2], comp[3]); + case 3: + return nir_vec3(build, comp[0], comp[1], comp[2]); + case 2: + return nir_vec2(build, comp[0], comp[1]); + case 1: + return comp[0]; + default: + unreachable("bad component count"); + return NULL; + } +} + +/** + * Similar to nir_fmov, but takes a nir_alu_src instead of a nir_ssa_def. + */ +static inline nir_ssa_def * +nir_fmov_alu(nir_builder *build, nir_alu_src src, unsigned num_components) +{ + nir_alu_instr *mov = nir_alu_instr_create(build->shader, nir_op_fmov); + nir_ssa_dest_init(&mov->instr, &mov->dest.dest, num_components, NULL); + mov->dest.write_mask = (1 << num_components) - 1; + mov->src[0] = src; + nir_builder_instr_insert(build, &mov->instr); + + return &mov->dest.dest.ssa; +} + +static inline nir_ssa_def * +nir_imov_alu(nir_builder *build, nir_alu_src src, unsigned num_components) +{ + nir_alu_instr *mov = nir_alu_instr_create(build->shader, nir_op_imov); + nir_ssa_dest_init(&mov->instr, &mov->dest.dest, num_components, NULL); + mov->dest.write_mask = (1 << num_components) - 1; + mov->src[0] = src; + nir_builder_instr_insert(build, &mov->instr); + + return &mov->dest.dest.ssa; +} + +/** + * Construct an fmov or imov that reswizzles the source's components. + */ +static inline nir_ssa_def * +nir_swizzle(nir_builder *build, nir_ssa_def *src, unsigned swiz[4], + unsigned num_components, bool use_fmov) +{ + nir_alu_src alu_src = { NIR_SRC_INIT }; + alu_src.src = nir_src_for_ssa(src); + for (unsigned i = 0; i < num_components; i++) + alu_src.swizzle[i] = swiz[i]; + + return use_fmov ? nir_fmov_alu(build, alu_src, num_components) : + nir_imov_alu(build, alu_src, num_components); +} + +static inline nir_ssa_def * +nir_channel(nir_builder *b, nir_ssa_def *def, unsigned c) +{ + unsigned swizzle[4] = {c, c, c, c}; + return nir_swizzle(b, def, swizzle, 1, false); +} + +/** + * Turns a nir_src into a nir_ssa_def * so it can be passed to + * nir_build_alu()-based builder calls. + * + * See nir_ssa_for_alu_src() for alu instructions. + */ +static inline nir_ssa_def * +nir_ssa_for_src(nir_builder *build, nir_src src, int num_components) +{ + if (src.is_ssa && src.ssa->num_components == num_components) + return src.ssa; + + nir_alu_src alu = { NIR_SRC_INIT }; + alu.src = src; + for (int j = 0; j < 4; j++) + alu.swizzle[j] = j; + + return nir_imov_alu(build, alu, num_components); +} + +/** + * Similar to nir_ssa_for_src(), but for alu src's, respecting the + * nir_alu_src's swizzle. + */ +static inline nir_ssa_def * +nir_ssa_for_alu_src(nir_builder *build, nir_alu_instr *instr, unsigned srcn) +{ + static uint8_t trivial_swizzle[4] = { 0, 1, 2, 3 }; + nir_alu_src *src = &instr->src[srcn]; + unsigned num_components = nir_ssa_alu_instr_src_components(instr, srcn); + + if (src->src.is_ssa && (src->src.ssa->num_components == num_components) && + !src->abs && !src->negate && + (memcmp(src->swizzle, trivial_swizzle, num_components) == 0)) + return src->src.ssa; + + return nir_imov_alu(build, *src, num_components); +} + +static inline nir_ssa_def * +nir_load_var(nir_builder *build, nir_variable *var) +{ + const unsigned num_components = glsl_get_vector_elements(var->type); + + nir_intrinsic_instr *load = + nir_intrinsic_instr_create(build->shader, nir_intrinsic_load_var); + load->num_components = num_components; + load->variables[0] = nir_deref_var_create(load, var); + nir_ssa_dest_init(&load->instr, &load->dest, num_components, NULL); + nir_builder_instr_insert(build, &load->instr); + return &load->dest.ssa; +} + +static inline void +nir_store_var(nir_builder *build, nir_variable *var, nir_ssa_def *value, + unsigned writemask) +{ + const unsigned num_components = glsl_get_vector_elements(var->type); + + nir_intrinsic_instr *store = + nir_intrinsic_instr_create(build->shader, nir_intrinsic_store_var); + store->num_components = num_components; + store->const_index[0] = writemask; + store->variables[0] = nir_deref_var_create(store, var); + store->src[0] = nir_src_for_ssa(value); + nir_builder_instr_insert(build, &store->instr); +} + +static inline nir_ssa_def * +nir_load_system_value(nir_builder *build, nir_intrinsic_op op, int index) +{ + nir_intrinsic_instr *load = nir_intrinsic_instr_create(build->shader, op); + load->num_components = nir_intrinsic_infos[op].dest_components; + load->const_index[0] = index; + nir_ssa_dest_init(&load->instr, &load->dest, + nir_intrinsic_infos[op].dest_components, NULL); + nir_builder_instr_insert(build, &load->instr); + return &load->dest.ssa; +} + +#endif /* NIR_BUILDER_H */ diff --git a/src/compiler/nir/nir_builder_opcodes_h.py b/src/compiler/nir/nir_builder_opcodes_h.py new file mode 100644 index 0000000..e27206e --- /dev/null +++ b/src/compiler/nir/nir_builder_opcodes_h.py @@ -0,0 +1,38 @@ +#! /usr/bin/env python + +template = """\ +/* Copyright (C) 2015 Broadcom + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + */ + +#ifndef _NIR_BUILDER_OPCODES_ +#define _NIR_BUILDER_OPCODES_ + +% for name, opcode in sorted(opcodes.iteritems()): +ALU${opcode.num_inputs}(${name}); +% endfor + +#endif /* _NIR_BUILDER_OPCODES_ */""" + +from nir_opcodes import opcodes +from mako.template import Template + +print Template(template).render(opcodes=opcodes) diff --git a/src/compiler/nir/nir_clone.c b/src/compiler/nir/nir_clone.c new file mode 100644 index 0000000..5eff743 --- /dev/null +++ b/src/compiler/nir/nir_clone.c @@ -0,0 +1,659 @@ +/* + * Copyright © 2015 Red Hat + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + */ + +#include "nir.h" +#include "nir_control_flow_private.h" + +/* Secret Decoder Ring: + * clone_foo(): + * Allocate and clone a foo. + * __clone_foo(): + * Clone body of foo (ie. parent class, embedded struct, etc) + */ + +typedef struct { + /* maps orig ptr -> cloned ptr: */ + struct hash_table *ptr_table; + + /* List of phi sources. */ + struct list_head phi_srcs; + + /* new shader object, used as memctx for just about everything else: */ + nir_shader *ns; +} clone_state; + +static void +init_clone_state(clone_state *state) +{ + state->ptr_table = _mesa_hash_table_create(NULL, _mesa_hash_pointer, + _mesa_key_pointer_equal); + list_inithead(&state->phi_srcs); +} + +static void +free_clone_state(clone_state *state) +{ + _mesa_hash_table_destroy(state->ptr_table, NULL); +} + +static void * +lookup_ptr(clone_state *state, const void *ptr) +{ + struct hash_entry *entry; + + if (!ptr) + return NULL; + + entry = _mesa_hash_table_search(state->ptr_table, ptr); + assert(entry && "Failed to find pointer!"); + if (!entry) + return NULL; + + return entry->data; +} + +static void +store_ptr(clone_state *state, void *nptr, const void *ptr) +{ + _mesa_hash_table_insert(state->ptr_table, ptr, nptr); +} + +static nir_constant * +clone_constant(clone_state *state, const nir_constant *c, nir_variable *nvar) +{ + nir_constant *nc = ralloc(nvar, nir_constant); + + nc->value = c->value; + nc->num_elements = c->num_elements; + nc->elements = ralloc_array(nvar, nir_constant *, c->num_elements); + for (unsigned i = 0; i < c->num_elements; i++) { + nc->elements[i] = clone_constant(state, c->elements[i], nvar); + } + + return nc; +} + +/* NOTE: for cloning nir_variable's, bypass nir_variable_create to avoid + * having to deal with locals and globals separately: + */ +static nir_variable * +clone_variable(clone_state *state, const nir_variable *var) +{ + nir_variable *nvar = rzalloc(state->ns, nir_variable); + store_ptr(state, nvar, var); + + nvar->type = var->type; + nvar->name = ralloc_strdup(nvar, var->name); + nvar->data = var->data; + nvar->num_state_slots = var->num_state_slots; + nvar->state_slots = ralloc_array(nvar, nir_state_slot, var->num_state_slots); + memcpy(nvar->state_slots, var->state_slots, + var->num_state_slots * sizeof(nir_state_slot)); + if (var->constant_initializer) { + nvar->constant_initializer = + clone_constant(state, var->constant_initializer, nvar); + } + nvar->interface_type = var->interface_type; + + return nvar; +} + +/* clone list of nir_variable: */ +static void +clone_var_list(clone_state *state, struct exec_list *dst, + const struct exec_list *list) +{ + exec_list_make_empty(dst); + foreach_list_typed(nir_variable, var, node, list) { + nir_variable *nvar = clone_variable(state, var); + exec_list_push_tail(dst, &nvar->node); + } +} + +/* NOTE: for cloning nir_register's, bypass nir_global/local_reg_create() + * to avoid having to deal with locals and globals separately: + */ +static nir_register * +clone_register(clone_state *state, const nir_register *reg) +{ + nir_register *nreg = rzalloc(state->ns, nir_register); + store_ptr(state, nreg, reg); + + nreg->num_components = reg->num_components; + nreg->num_array_elems = reg->num_array_elems; + nreg->index = reg->index; + nreg->name = ralloc_strdup(nreg, reg->name); + nreg->is_global = reg->is_global; + nreg->is_packed = reg->is_packed; + + /* reconstructing uses/defs/if_uses handled by nir_instr_insert() */ + list_inithead(&nreg->uses); + list_inithead(&nreg->defs); + list_inithead(&nreg->if_uses); + + return nreg; +} + +/* clone list of nir_register: */ +static void +clone_reg_list(clone_state *state, struct exec_list *dst, + const struct exec_list *list) +{ + exec_list_make_empty(dst); + foreach_list_typed(nir_register, reg, node, list) { + nir_register *nreg = clone_register(state, reg); + exec_list_push_tail(dst, &nreg->node); + } +} + +static void +__clone_src(clone_state *state, void *ninstr_or_if, + nir_src *nsrc, const nir_src *src) +{ + nsrc->is_ssa = src->is_ssa; + if (src->is_ssa) { + nsrc->ssa = lookup_ptr(state, src->ssa); + } else { + nsrc->reg.reg = lookup_ptr(state, src->reg.reg); + if (src->reg.indirect) { + nsrc->reg.indirect = ralloc(ninstr_or_if, nir_src); + __clone_src(state, ninstr_or_if, nsrc->reg.indirect, src->reg.indirect); + } + nsrc->reg.base_offset = src->reg.base_offset; + } +} + +static void +__clone_dst(clone_state *state, nir_instr *ninstr, + nir_dest *ndst, const nir_dest *dst) +{ + ndst->is_ssa = dst->is_ssa; + if (dst->is_ssa) { + nir_ssa_dest_init(ninstr, ndst, dst->ssa.num_components, dst->ssa.name); + store_ptr(state, &ndst->ssa, &dst->ssa); + } else { + ndst->reg.reg = lookup_ptr(state, dst->reg.reg); + if (dst->reg.indirect) { + ndst->reg.indirect = ralloc(ninstr, nir_src); + __clone_src(state, ninstr, ndst->reg.indirect, dst->reg.indirect); + } + ndst->reg.base_offset = dst->reg.base_offset; + } +} + +static nir_deref *clone_deref(clone_state *state, const nir_deref *deref, + nir_instr *ninstr, nir_deref *parent); + +static nir_deref_var * +clone_deref_var(clone_state *state, const nir_deref_var *dvar, + nir_instr *ninstr) +{ + nir_variable *nvar = lookup_ptr(state, dvar->var); + nir_deref_var *ndvar = nir_deref_var_create(ninstr, nvar); + + if (dvar->deref.child) + ndvar->deref.child = clone_deref(state, dvar->deref.child, + ninstr, &ndvar->deref); + + return ndvar; +} + +static nir_deref_array * +clone_deref_array(clone_state *state, const nir_deref_array *darr, + nir_instr *ninstr, nir_deref *parent) +{ + nir_deref_array *ndarr = nir_deref_array_create(parent); + + ndarr->deref.type = darr->deref.type; + if (darr->deref.child) + ndarr->deref.child = clone_deref(state, darr->deref.child, + ninstr, &ndarr->deref); + + ndarr->deref_array_type = darr->deref_array_type; + ndarr->base_offset = darr->base_offset; + if (ndarr->deref_array_type == nir_deref_array_type_indirect) + __clone_src(state, ninstr, &ndarr->indirect, &darr->indirect); + + return ndarr; +} + +static nir_deref_struct * +clone_deref_struct(clone_state *state, const nir_deref_struct *dstr, + nir_instr *ninstr, nir_deref *parent) +{ + nir_deref_struct *ndstr = nir_deref_struct_create(parent, dstr->index); + + ndstr->deref.type = dstr->deref.type; + if (dstr->deref.child) + ndstr->deref.child = clone_deref(state, dstr->deref.child, + ninstr, &ndstr->deref); + + return ndstr; +} + +static nir_deref * +clone_deref(clone_state *state, const nir_deref *dref, + nir_instr *ninstr, nir_deref *parent) +{ + switch (dref->deref_type) { + case nir_deref_type_array: + return &clone_deref_array(state, nir_deref_as_array(dref), + ninstr, parent)->deref; + case nir_deref_type_struct: + return &clone_deref_struct(state, nir_deref_as_struct(dref), + ninstr, parent)->deref; + default: + unreachable("bad deref type"); + return NULL; + } +} + +static nir_alu_instr * +clone_alu(clone_state *state, const nir_alu_instr *alu) +{ + nir_alu_instr *nalu = nir_alu_instr_create(state->ns, alu->op); + + __clone_dst(state, &nalu->instr, &nalu->dest.dest, &alu->dest.dest); + nalu->dest.saturate = alu->dest.saturate; + nalu->dest.write_mask = alu->dest.write_mask; + + for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) { + __clone_src(state, &nalu->instr, &nalu->src[i].src, &alu->src[i].src); + nalu->src[i].negate = alu->src[i].negate; + nalu->src[i].abs = alu->src[i].abs; + memcpy(nalu->src[i].swizzle, alu->src[i].swizzle, + sizeof(nalu->src[i].swizzle)); + } + + return nalu; +} + +static nir_intrinsic_instr * +clone_intrinsic(clone_state *state, const nir_intrinsic_instr *itr) +{ + nir_intrinsic_instr *nitr = + nir_intrinsic_instr_create(state->ns, itr->intrinsic); + + unsigned num_variables = nir_intrinsic_infos[itr->intrinsic].num_variables; + unsigned num_srcs = nir_intrinsic_infos[itr->intrinsic].num_srcs; + + if (nir_intrinsic_infos[itr->intrinsic].has_dest) + __clone_dst(state, &nitr->instr, &nitr->dest, &itr->dest); + + nitr->num_components = itr->num_components; + memcpy(nitr->const_index, itr->const_index, sizeof(nitr->const_index)); + + for (unsigned i = 0; i < num_variables; i++) { + nitr->variables[i] = clone_deref_var(state, itr->variables[i], + &nitr->instr); + } + + for (unsigned i = 0; i < num_srcs; i++) + __clone_src(state, &nitr->instr, &nitr->src[i], &itr->src[i]); + + return nitr; +} + +static nir_load_const_instr * +clone_load_const(clone_state *state, const nir_load_const_instr *lc) +{ + nir_load_const_instr *nlc = + nir_load_const_instr_create(state->ns, lc->def.num_components); + + memcpy(&nlc->value, &lc->value, sizeof(nlc->value)); + + store_ptr(state, &nlc->def, &lc->def); + + return nlc; +} + +static nir_ssa_undef_instr * +clone_ssa_undef(clone_state *state, const nir_ssa_undef_instr *sa) +{ + nir_ssa_undef_instr *nsa = + nir_ssa_undef_instr_create(state->ns, sa->def.num_components); + + store_ptr(state, &nsa->def, &sa->def); + + return nsa; +} + +static nir_tex_instr * +clone_tex(clone_state *state, const nir_tex_instr *tex) +{ + nir_tex_instr *ntex = nir_tex_instr_create(state->ns, tex->num_srcs); + + ntex->sampler_dim = tex->sampler_dim; + ntex->dest_type = tex->dest_type; + ntex->op = tex->op; + __clone_dst(state, &ntex->instr, &ntex->dest, &tex->dest); + for (unsigned i = 0; i < ntex->num_srcs; i++) { + ntex->src[i].src_type = tex->src[i].src_type; + __clone_src(state, &ntex->instr, &ntex->src[i].src, &tex->src[i].src); + } + ntex->coord_components = tex->coord_components; + ntex->is_array = tex->is_array; + ntex->is_shadow = tex->is_shadow; + ntex->is_new_style_shadow = tex->is_new_style_shadow; + memcpy(ntex->const_offset, tex->const_offset, sizeof(ntex->const_offset)); + ntex->component = tex->component; + ntex->sampler_index = tex->sampler_index; + ntex->sampler_array_size = tex->sampler_array_size; + if (tex->sampler) + ntex->sampler = clone_deref_var(state, tex->sampler, &ntex->instr); + + return ntex; +} + +static nir_phi_instr * +clone_phi(clone_state *state, const nir_phi_instr *phi, nir_block *nblk) +{ + nir_phi_instr *nphi = nir_phi_instr_create(state->ns); + + __clone_dst(state, &nphi->instr, &nphi->dest, &phi->dest); + + /* Cloning a phi node is a bit different from other instructions. The + * sources of phi instructions are the only time where we can use an SSA + * def before it is defined. In order to handle this, we just copy over + * the sources from the old phi instruction directly and then fix them up + * in a second pass once all the instrutions in the function have been + * properly cloned. + * + * In order to ensure that the copied sources (which are the same as the + * old phi instruction's sources for now) don't get inserted into the old + * shader's use-def lists, we have to add the phi instruction *before* we + * set up its sources. + */ + nir_instr_insert_after_block(nblk, &nphi->instr); + + foreach_list_typed(nir_phi_src, src, node, &phi->srcs) { + nir_phi_src *nsrc = ralloc(nphi, nir_phi_src); + + /* Just copy the old source for now. */ + memcpy(nsrc, src, sizeof(*src)); + + /* Since we're not letting nir_insert_instr handle use/def stuff for us, + * we have to set the parent_instr manually. It doesn't really matter + * when we do it, so we might as well do it here. + */ + nsrc->src.parent_instr = &nphi->instr; + + /* Stash it in the list of phi sources. We'll walk this list and fix up + * sources at the very end of clone_function_impl. + */ + list_add(&nsrc->src.use_link, &state->phi_srcs); + + exec_list_push_tail(&nphi->srcs, &nsrc->node); + } + + return nphi; +} + +static nir_jump_instr * +clone_jump(clone_state *state, const nir_jump_instr *jmp) +{ + nir_jump_instr *njmp = nir_jump_instr_create(state->ns, jmp->type); + + return njmp; +} + +static nir_call_instr * +clone_call(clone_state *state, const nir_call_instr *call) +{ + nir_function *ncallee = lookup_ptr(state, call->callee); + nir_call_instr *ncall = nir_call_instr_create(state->ns, ncallee); + + for (unsigned i = 0; i < ncall->num_params; i++) + ncall->params[i] = clone_deref_var(state, call->params[i], &ncall->instr); + + ncall->return_deref = clone_deref_var(state, call->return_deref, + &ncall->instr); + + return ncall; +} + +static nir_instr * +clone_instr(clone_state *state, const nir_instr *instr) +{ + switch (instr->type) { + case nir_instr_type_alu: + return &clone_alu(state, nir_instr_as_alu(instr))->instr; + case nir_instr_type_intrinsic: + return &clone_intrinsic(state, nir_instr_as_intrinsic(instr))->instr; + case nir_instr_type_load_const: + return &clone_load_const(state, nir_instr_as_load_const(instr))->instr; + case nir_instr_type_ssa_undef: + return &clone_ssa_undef(state, nir_instr_as_ssa_undef(instr))->instr; + case nir_instr_type_tex: + return &clone_tex(state, nir_instr_as_tex(instr))->instr; + case nir_instr_type_phi: + unreachable("Cannot clone phis with clone_instr"); + case nir_instr_type_jump: + return &clone_jump(state, nir_instr_as_jump(instr))->instr; + case nir_instr_type_call: + return &clone_call(state, nir_instr_as_call(instr))->instr; + case nir_instr_type_parallel_copy: + unreachable("Cannot clone parallel copies"); + default: + unreachable("bad instr type"); + return NULL; + } +} + +static nir_block * +clone_block(clone_state *state, struct exec_list *cf_list, const nir_block *blk) +{ + /* Don't actually create a new block. Just use the one from the tail of + * the list. NIR guarantees that the tail of the list is a block and that + * no two blocks are side-by-side in the IR; It should be empty. + */ + nir_block *nblk = + exec_node_data(nir_block, exec_list_get_tail(cf_list), cf_node.node); + assert(nblk->cf_node.type == nir_cf_node_block); + assert(exec_list_is_empty(&nblk->instr_list)); + + /* We need this for phi sources */ + store_ptr(state, nblk, blk); + + nir_foreach_instr(blk, instr) { + if (instr->type == nir_instr_type_phi) { + /* Phi instructions are a bit of a special case when cloning because + * we don't want inserting the instruction to automatically handle + * use/defs for us. Instead, we need to wait until all the + * blocks/instructions are in so that we can set their sources up. + */ + clone_phi(state, nir_instr_as_phi(instr), nblk); + } else { + nir_instr *ninstr = clone_instr(state, instr); + nir_instr_insert_after_block(nblk, ninstr); + } + } + + return nblk; +} + +static void +clone_cf_list(clone_state *state, struct exec_list *dst, + const struct exec_list *list); + +static nir_if * +clone_if(clone_state *state, struct exec_list *cf_list, const nir_if *i) +{ + nir_if *ni = nir_if_create(state->ns); + + __clone_src(state, ni, &ni->condition, &i->condition); + + nir_cf_node_insert_end(cf_list, &ni->cf_node); + + clone_cf_list(state, &ni->then_list, &i->then_list); + clone_cf_list(state, &ni->else_list, &i->else_list); + + return ni; +} + +static nir_loop * +clone_loop(clone_state *state, struct exec_list *cf_list, const nir_loop *loop) +{ + nir_loop *nloop = nir_loop_create(state->ns); + + nir_cf_node_insert_end(cf_list, &nloop->cf_node); + + clone_cf_list(state, &nloop->body, &loop->body); + + return nloop; +} + +/* clone list of nir_cf_node: */ +static void +clone_cf_list(clone_state *state, struct exec_list *dst, + const struct exec_list *list) +{ + foreach_list_typed(nir_cf_node, cf, node, list) { + switch (cf->type) { + case nir_cf_node_block: + clone_block(state, dst, nir_cf_node_as_block(cf)); + break; + case nir_cf_node_if: + clone_if(state, dst, nir_cf_node_as_if(cf)); + break; + case nir_cf_node_loop: + clone_loop(state, dst, nir_cf_node_as_loop(cf)); + break; + default: + unreachable("bad cf type"); + } + } +} + +static nir_function_impl * +clone_function_impl(clone_state *state, const nir_function_impl *fi, + nir_function *nfxn) +{ + nir_function_impl *nfi = nir_function_impl_create(nfxn); + + clone_var_list(state, &nfi->locals, &fi->locals); + clone_reg_list(state, &nfi->registers, &fi->registers); + nfi->reg_alloc = fi->reg_alloc; + + nfi->num_params = fi->num_params; + nfi->params = ralloc_array(state->ns, nir_variable *, fi->num_params); + for (unsigned i = 0; i < fi->num_params; i++) { + nfi->params[i] = lookup_ptr(state, fi->params[i]); + } + nfi->return_var = lookup_ptr(state, fi->return_var); + + assert(list_empty(&state->phi_srcs)); + + clone_cf_list(state, &nfi->body, &fi->body); + + /* After we've cloned almost everything, we have to walk the list of phi + * sources and fix them up. Thanks to loops, the block and SSA value for a + * phi source may not be defined when we first encounter it. Instead, we + * add it to the phi_srcs list and we fix it up here. + */ + list_for_each_entry_safe(nir_phi_src, src, &state->phi_srcs, src.use_link) { + src->pred = lookup_ptr(state, src->pred); + assert(src->src.is_ssa); + src->src.ssa = lookup_ptr(state, src->src.ssa); + + /* Remove from this list and place in the uses of the SSA def */ + list_del(&src->src.use_link); + list_addtail(&src->src.use_link, &src->src.ssa->uses); + } + assert(list_empty(&state->phi_srcs)); + + /* All metadata is invalidated in the cloning process */ + nfi->valid_metadata = 0; + + return nfi; +} + +static nir_function * +clone_function(clone_state *state, const nir_function *fxn, nir_shader *ns) +{ + assert(ns == state->ns); + nir_function *nfxn = nir_function_create(ns, fxn->name); + + /* Needed for call instructions */ + store_ptr(state, nfxn, fxn); + + nfxn->num_params = fxn->num_params; + nfxn->params = ralloc_array(state->ns, nir_parameter, fxn->num_params); + memcpy(nfxn->params, fxn->params, sizeof(nir_parameter) * fxn->num_params); + + nfxn->return_type = fxn->return_type; + + /* At first glance, it looks like we should clone the function_impl here. + * However, call instructions need to be able to reference at least the + * function and those will get processed as we clone the function_impl's. + * We stop here and do function_impls as a second pass. + */ + + return nfxn; +} + +nir_shader * +nir_shader_clone(void *mem_ctx, const nir_shader *s) +{ + clone_state state; + init_clone_state(&state); + + nir_shader *ns = nir_shader_create(mem_ctx, s->stage, s->options); + state.ns = ns; + + clone_var_list(&state, &ns->uniforms, &s->uniforms); + clone_var_list(&state, &ns->inputs, &s->inputs); + clone_var_list(&state, &ns->outputs, &s->outputs); + clone_var_list(&state, &ns->globals, &s->globals); + clone_var_list(&state, &ns->system_values, &s->system_values); + + /* Go through and clone functions */ + foreach_list_typed(nir_function, fxn, node, &s->functions) + clone_function(&state, fxn, ns); + + /* Only after all functions are cloned can we clone the actual function + * implementations. This is because nir_call_instr's need to reference the + * functions of other functions and we don't know what order the functions + * will have in the list. + */ + nir_foreach_function(s, fxn) { + nir_function *nfxn = lookup_ptr(&state, fxn); + clone_function_impl(&state, fxn->impl, nfxn); + } + + clone_reg_list(&state, &ns->registers, &s->registers); + ns->reg_alloc = s->reg_alloc; + + ns->info = s->info; + ns->info.name = ralloc_strdup(ns, ns->info.name); + if (ns->info.label) + ns->info.label = ralloc_strdup(ns, ns->info.label); + + ns->num_inputs = s->num_inputs; + ns->num_uniforms = s->num_uniforms; + ns->num_outputs = s->num_outputs; + + free_clone_state(&state); + + return ns; +} diff --git a/src/compiler/nir/nir_constant_expressions.h b/src/compiler/nir/nir_constant_expressions.h new file mode 100644 index 0000000..97997f2 --- /dev/null +++ b/src/compiler/nir/nir_constant_expressions.h @@ -0,0 +1,31 @@ +/* + * Copyright © 2014 Connor Abbott + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +#include "nir.h" + +nir_const_value nir_eval_const_opcode(nir_op op, unsigned num_components, + nir_const_value *src); diff --git a/src/compiler/nir/nir_constant_expressions.py b/src/compiler/nir/nir_constant_expressions.py new file mode 100644 index 0000000..32784f6 --- /dev/null +++ b/src/compiler/nir/nir_constant_expressions.py @@ -0,0 +1,336 @@ +#! /usr/bin/python2 +template = """\ +/* + * Copyright (C) 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand (jason@jlekstrand.net) + */ + +#include <math.h> +#include "main/core.h" +#include "util/rounding.h" /* for _mesa_roundeven */ +#include "util/half_float.h" +#include "nir_constant_expressions.h" + +/** + * Evaluate one component of packSnorm4x8. + */ +static uint8_t +pack_snorm_1x8(float x) +{ + /* From section 8.4 of the GLSL 4.30 spec: + * + * packSnorm4x8 + * ------------ + * The conversion for component c of v to fixed point is done as + * follows: + * + * packSnorm4x8: round(clamp(c, -1, +1) * 127.0) + * + * We must first cast the float to an int, because casting a negative + * float to a uint is undefined. + */ + return (uint8_t) (int) + _mesa_roundevenf(CLAMP(x, -1.0f, +1.0f) * 127.0f); +} + +/** + * Evaluate one component of packSnorm2x16. + */ +static uint16_t +pack_snorm_1x16(float x) +{ + /* From section 8.4 of the GLSL ES 3.00 spec: + * + * packSnorm2x16 + * ------------- + * The conversion for component c of v to fixed point is done as + * follows: + * + * packSnorm2x16: round(clamp(c, -1, +1) * 32767.0) + * + * We must first cast the float to an int, because casting a negative + * float to a uint is undefined. + */ + return (uint16_t) (int) + _mesa_roundevenf(CLAMP(x, -1.0f, +1.0f) * 32767.0f); +} + +/** + * Evaluate one component of unpackSnorm4x8. + */ +static float +unpack_snorm_1x8(uint8_t u) +{ + /* From section 8.4 of the GLSL 4.30 spec: + * + * unpackSnorm4x8 + * -------------- + * The conversion for unpacked fixed-point value f to floating point is + * done as follows: + * + * unpackSnorm4x8: clamp(f / 127.0, -1, +1) + */ + return CLAMP((int8_t) u / 127.0f, -1.0f, +1.0f); +} + +/** + * Evaluate one component of unpackSnorm2x16. + */ +static float +unpack_snorm_1x16(uint16_t u) +{ + /* From section 8.4 of the GLSL ES 3.00 spec: + * + * unpackSnorm2x16 + * --------------- + * The conversion for unpacked fixed-point value f to floating point is + * done as follows: + * + * unpackSnorm2x16: clamp(f / 32767.0, -1, +1) + */ + return CLAMP((int16_t) u / 32767.0f, -1.0f, +1.0f); +} + +/** + * Evaluate one component packUnorm4x8. + */ +static uint8_t +pack_unorm_1x8(float x) +{ + /* From section 8.4 of the GLSL 4.30 spec: + * + * packUnorm4x8 + * ------------ + * The conversion for component c of v to fixed point is done as + * follows: + * + * packUnorm4x8: round(clamp(c, 0, +1) * 255.0) + */ + return (uint8_t) (int) + _mesa_roundevenf(CLAMP(x, 0.0f, 1.0f) * 255.0f); +} + +/** + * Evaluate one component packUnorm2x16. + */ +static uint16_t +pack_unorm_1x16(float x) +{ + /* From section 8.4 of the GLSL ES 3.00 spec: + * + * packUnorm2x16 + * ------------- + * The conversion for component c of v to fixed point is done as + * follows: + * + * packUnorm2x16: round(clamp(c, 0, +1) * 65535.0) + */ + return (uint16_t) (int) + _mesa_roundevenf(CLAMP(x, 0.0f, 1.0f) * 65535.0f); +} + +/** + * Evaluate one component of unpackUnorm4x8. + */ +static float +unpack_unorm_1x8(uint8_t u) +{ + /* From section 8.4 of the GLSL 4.30 spec: + * + * unpackUnorm4x8 + * -------------- + * The conversion for unpacked fixed-point value f to floating point is + * done as follows: + * + * unpackUnorm4x8: f / 255.0 + */ + return (float) u / 255.0f; +} + +/** + * Evaluate one component of unpackUnorm2x16. + */ +static float +unpack_unorm_1x16(uint16_t u) +{ + /* From section 8.4 of the GLSL ES 3.00 spec: + * + * unpackUnorm2x16 + * --------------- + * The conversion for unpacked fixed-point value f to floating point is + * done as follows: + * + * unpackUnorm2x16: f / 65535.0 + */ + return (float) u / 65535.0f; +} + +/** + * Evaluate one component of packHalf2x16. + */ +static uint16_t +pack_half_1x16(float x) +{ + return _mesa_float_to_half(x); +} + +/** + * Evaluate one component of unpackHalf2x16. + */ +static float +unpack_half_1x16(uint16_t u) +{ + return _mesa_half_to_float(u); +} + +/* Some typed vector structures to make things like src0.y work */ +% for type in ["float", "int", "uint", "bool"]: +struct ${type}_vec { + ${type} x; + ${type} y; + ${type} z; + ${type} w; +}; +% endfor + +% for name, op in sorted(opcodes.iteritems()): +static nir_const_value +evaluate_${name}(unsigned num_components, nir_const_value *_src) +{ + nir_const_value _dst_val = { { {0, 0, 0, 0} } }; + + ## For each non-per-component input, create a variable srcN that + ## contains x, y, z, and w elements which are filled in with the + ## appropriately-typed values. + % for j in range(op.num_inputs): + % if op.input_sizes[j] == 0: + <% continue %> + % elif "src" + str(j) not in op.const_expr: + ## Avoid unused variable warnings + <% continue %> + %endif + + struct ${op.input_types[j]}_vec src${j} = { + % for k in range(op.input_sizes[j]): + % if op.input_types[j] == "bool": + _src[${j}].u[${k}] != 0, + % else: + _src[${j}].${op.input_types[j][:1]}[${k}], + % endif + % endfor + }; + % endfor + + % if op.output_size == 0: + ## For per-component instructions, we need to iterate over the + ## components and apply the constant expression one component + ## at a time. + for (unsigned _i = 0; _i < num_components; _i++) { + ## For each per-component input, create a variable srcN that + ## contains the value of the current (_i'th) component. + % for j in range(op.num_inputs): + % if op.input_sizes[j] != 0: + <% continue %> + % elif "src" + str(j) not in op.const_expr: + ## Avoid unused variable warnings + <% continue %> + % elif op.input_types[j] == "bool": + bool src${j} = _src[${j}].u[_i] != 0; + % else: + ${op.input_types[j]} src${j} = _src[${j}].${op.input_types[j][:1]}[_i]; + % endif + % endfor + + ## Create an appropriately-typed variable dst and assign the + ## result of the const_expr to it. If const_expr already contains + ## writes to dst, just include const_expr directly. + % if "dst" in op.const_expr: + ${op.output_type} dst; + ${op.const_expr} + % else: + ${op.output_type} dst = ${op.const_expr}; + % endif + + ## Store the current component of the actual destination to the + ## value of dst. + % if op.output_type == "bool": + ## Sanitize the C value to a proper NIR bool + _dst_val.u[_i] = dst ? NIR_TRUE : NIR_FALSE; + % else: + _dst_val.${op.output_type[:1]}[_i] = dst; + % endif + } + % else: + ## In the non-per-component case, create a struct dst with + ## appropriately-typed elements x, y, z, and w and assign the result + ## of the const_expr to all components of dst, or include the + ## const_expr directly if it writes to dst already. + struct ${op.output_type}_vec dst; + + % if "dst" in op.const_expr: + ${op.const_expr} + % else: + ## Splat the value to all components. This way expressions which + ## write the same value to all components don't need to explicitly + ## write to dest. One such example is fnoise which has a + ## const_expr of 0.0f. + dst.x = dst.y = dst.z = dst.w = ${op.const_expr}; + % endif + + ## For each component in the destination, copy the value of dst to + ## the actual destination. + % for k in range(op.output_size): + % if op.output_type == "bool": + ## Sanitize the C value to a proper NIR bool + _dst_val.u[${k}] = dst.${"xyzw"[k]} ? NIR_TRUE : NIR_FALSE; + % else: + _dst_val.${op.output_type[:1]}[${k}] = dst.${"xyzw"[k]}; + % endif + % endfor + % endif + + return _dst_val; +} +% endfor + +nir_const_value +nir_eval_const_opcode(nir_op op, unsigned num_components, + nir_const_value *src) +{ + switch (op) { +% for name in sorted(opcodes.iterkeys()): + case nir_op_${name}: { + return evaluate_${name}(num_components, src); + break; + } +% endfor + default: + unreachable("shouldn't get here"); + } +}""" + +from nir_opcodes import opcodes +from mako.template import Template + +print Template(template).render(opcodes=opcodes) diff --git a/src/compiler/nir/nir_control_flow.c b/src/compiler/nir/nir_control_flow.c new file mode 100644 index 0000000..96395a4 --- /dev/null +++ b/src/compiler/nir/nir_control_flow.c @@ -0,0 +1,808 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +#include "nir_control_flow_private.h" + +/** + * \name Control flow modification + * + * These functions modify the control flow tree while keeping the control flow + * graph up-to-date. The invariants respected are: + * 1. Each then statement, else statement, or loop body must have at least one + * control flow node. + * 2. Each if-statement and loop must have one basic block before it and one + * after. + * 3. Two basic blocks cannot be directly next to each other. + * 4. If a basic block has a jump instruction, there must be only one and it + * must be at the end of the block. + * 5. The CFG must always be connected - this means that we must insert a fake + * CFG edge for loops with no break statement. + * + * The purpose of the second one is so that we have places to insert code during + * GCM, as well as eliminating the possibility of critical edges. + */ +/*@{*/ + +static bool +block_ends_in_jump(nir_block *block) +{ + return !exec_list_is_empty(&block->instr_list) && + nir_block_last_instr(block)->type == nir_instr_type_jump; +} + +static inline void +block_add_pred(nir_block *block, nir_block *pred) +{ + _mesa_set_add(block->predecessors, pred); +} + +static inline void +block_remove_pred(nir_block *block, nir_block *pred) +{ + struct set_entry *entry = _mesa_set_search(block->predecessors, pred); + + assert(entry); + + _mesa_set_remove(block->predecessors, entry); +} + +static void +link_blocks(nir_block *pred, nir_block *succ1, nir_block *succ2) +{ + pred->successors[0] = succ1; + if (succ1 != NULL) + block_add_pred(succ1, pred); + + pred->successors[1] = succ2; + if (succ2 != NULL) + block_add_pred(succ2, pred); +} + +static void +unlink_blocks(nir_block *pred, nir_block *succ) +{ + if (pred->successors[0] == succ) { + pred->successors[0] = pred->successors[1]; + pred->successors[1] = NULL; + } else { + assert(pred->successors[1] == succ); + pred->successors[1] = NULL; + } + + block_remove_pred(succ, pred); +} + +static void +unlink_block_successors(nir_block *block) +{ + if (block->successors[1] != NULL) + unlink_blocks(block, block->successors[1]); + if (block->successors[0] != NULL) + unlink_blocks(block, block->successors[0]); +} + +static void +link_non_block_to_block(nir_cf_node *node, nir_block *block) +{ + if (node->type == nir_cf_node_if) { + /* + * We're trying to link an if to a block after it; this just means linking + * the last block of the then and else branches. + */ + + nir_if *if_stmt = nir_cf_node_as_if(node); + + nir_cf_node *last_then = nir_if_last_then_node(if_stmt); + assert(last_then->type == nir_cf_node_block); + nir_block *last_then_block = nir_cf_node_as_block(last_then); + + nir_cf_node *last_else = nir_if_last_else_node(if_stmt); + assert(last_else->type == nir_cf_node_block); + nir_block *last_else_block = nir_cf_node_as_block(last_else); + + if (!block_ends_in_jump(last_then_block)) { + unlink_block_successors(last_then_block); + link_blocks(last_then_block, block, NULL); + } + + if (!block_ends_in_jump(last_else_block)) { + unlink_block_successors(last_else_block); + link_blocks(last_else_block, block, NULL); + } + } else { + assert(node->type == nir_cf_node_loop); + + /* + * We can only get to this codepath if we're inserting a new loop, or + * at least a loop with no break statements; we can't insert break + * statements into a loop when we haven't inserted it into the CFG + * because we wouldn't know which block comes after the loop + * and therefore, which block should be the successor of the block with + * the break). Therefore, we need to insert a fake edge (see invariant + * #5). + */ + + nir_loop *loop = nir_cf_node_as_loop(node); + + nir_cf_node *last = nir_loop_last_cf_node(loop); + assert(last->type == nir_cf_node_block); + nir_block *last_block = nir_cf_node_as_block(last); + + last_block->successors[1] = block; + block_add_pred(block, last_block); + } +} + +static void +link_block_to_non_block(nir_block *block, nir_cf_node *node) +{ + if (node->type == nir_cf_node_if) { + /* + * We're trying to link a block to an if after it; this just means linking + * the block to the first block of the then and else branches. + */ + + nir_if *if_stmt = nir_cf_node_as_if(node); + + nir_cf_node *first_then = nir_if_first_then_node(if_stmt); + assert(first_then->type == nir_cf_node_block); + nir_block *first_then_block = nir_cf_node_as_block(first_then); + + nir_cf_node *first_else = nir_if_first_else_node(if_stmt); + assert(first_else->type == nir_cf_node_block); + nir_block *first_else_block = nir_cf_node_as_block(first_else); + + unlink_block_successors(block); + link_blocks(block, first_then_block, first_else_block); + } else { + /* + * For similar reasons as the corresponding case in + * link_non_block_to_block(), don't worry about if the loop header has + * any predecessors that need to be unlinked. + */ + + assert(node->type == nir_cf_node_loop); + + nir_loop *loop = nir_cf_node_as_loop(node); + + nir_cf_node *loop_header = nir_loop_first_cf_node(loop); + assert(loop_header->type == nir_cf_node_block); + nir_block *loop_header_block = nir_cf_node_as_block(loop_header); + + unlink_block_successors(block); + link_blocks(block, loop_header_block, NULL); + } + +} + +/** + * Replace a block's successor with a different one. + */ +static void +replace_successor(nir_block *block, nir_block *old_succ, nir_block *new_succ) +{ + if (block->successors[0] == old_succ) { + block->successors[0] = new_succ; + } else { + assert(block->successors[1] == old_succ); + block->successors[1] = new_succ; + } + + block_remove_pred(old_succ, block); + block_add_pred(new_succ, block); +} + +/** + * Takes a basic block and inserts a new empty basic block before it, making its + * predecessors point to the new block. This essentially splits the block into + * an empty header and a body so that another non-block CF node can be inserted + * between the two. Note that this does *not* link the two basic blocks, so + * some kind of cleanup *must* be performed after this call. + */ + +static nir_block * +split_block_beginning(nir_block *block) +{ + nir_block *new_block = nir_block_create(ralloc_parent(block)); + new_block->cf_node.parent = block->cf_node.parent; + exec_node_insert_node_before(&block->cf_node.node, &new_block->cf_node.node); + + struct set_entry *entry; + set_foreach(block->predecessors, entry) { + nir_block *pred = (nir_block *) entry->key; + replace_successor(pred, block, new_block); + } + + /* Any phi nodes must stay part of the new block, or else their + * sourcse will be messed up. This will reverse the order of the phi's, but + * order shouldn't matter. + */ + nir_foreach_instr_safe(block, instr) { + if (instr->type != nir_instr_type_phi) + break; + + exec_node_remove(&instr->node); + instr->block = new_block; + exec_list_push_head(&new_block->instr_list, &instr->node); + } + + return new_block; +} + +static void +rewrite_phi_preds(nir_block *block, nir_block *old_pred, nir_block *new_pred) +{ + nir_foreach_instr_safe(block, instr) { + if (instr->type != nir_instr_type_phi) + break; + + nir_phi_instr *phi = nir_instr_as_phi(instr); + nir_foreach_phi_src(phi, src) { + if (src->pred == old_pred) { + src->pred = new_pred; + break; + } + } + } +} + +static void +insert_phi_undef(nir_block *block, nir_block *pred) +{ + nir_function_impl *impl = nir_cf_node_get_function(&block->cf_node); + nir_foreach_instr(block, instr) { + if (instr->type != nir_instr_type_phi) + break; + + nir_phi_instr *phi = nir_instr_as_phi(instr); + nir_ssa_undef_instr *undef = + nir_ssa_undef_instr_create(ralloc_parent(phi), + phi->dest.ssa.num_components); + nir_instr_insert_before_cf_list(&impl->body, &undef->instr); + nir_phi_src *src = ralloc(phi, nir_phi_src); + src->pred = pred; + src->src.parent_instr = &phi->instr; + src->src.is_ssa = true; + src->src.ssa = &undef->def; + + list_addtail(&src->src.use_link, &undef->def.uses); + + exec_list_push_tail(&phi->srcs, &src->node); + } +} + +/** + * Moves the successors of source to the successors of dest, leaving both + * successors of source NULL. + */ + +static void +move_successors(nir_block *source, nir_block *dest) +{ + nir_block *succ1 = source->successors[0]; + nir_block *succ2 = source->successors[1]; + + if (succ1) { + unlink_blocks(source, succ1); + rewrite_phi_preds(succ1, source, dest); + } + + if (succ2) { + unlink_blocks(source, succ2); + rewrite_phi_preds(succ2, source, dest); + } + + unlink_block_successors(dest); + link_blocks(dest, succ1, succ2); +} + +/* Given a basic block with no successors that has been inserted into the + * control flow tree, gives it the successors it would normally have assuming + * it doesn't end in a jump instruction. Also inserts phi sources with undefs + * if necessary. + */ +static void +block_add_normal_succs(nir_block *block) +{ + if (exec_node_is_tail_sentinel(block->cf_node.node.next)) { + nir_cf_node *parent = block->cf_node.parent; + if (parent->type == nir_cf_node_if) { + nir_cf_node *next = nir_cf_node_next(parent); + assert(next->type == nir_cf_node_block); + nir_block *next_block = nir_cf_node_as_block(next); + + link_blocks(block, next_block, NULL); + } else { + assert(parent->type == nir_cf_node_loop); + nir_loop *loop = nir_cf_node_as_loop(parent); + + nir_cf_node *head = nir_loop_first_cf_node(loop); + assert(head->type == nir_cf_node_block); + nir_block *head_block = nir_cf_node_as_block(head); + + link_blocks(block, head_block, NULL); + insert_phi_undef(head_block, block); + } + } else { + nir_cf_node *next = nir_cf_node_next(&block->cf_node); + if (next->type == nir_cf_node_if) { + nir_if *next_if = nir_cf_node_as_if(next); + + nir_cf_node *first_then = nir_if_first_then_node(next_if); + assert(first_then->type == nir_cf_node_block); + nir_block *first_then_block = nir_cf_node_as_block(first_then); + + nir_cf_node *first_else = nir_if_first_else_node(next_if); + assert(first_else->type == nir_cf_node_block); + nir_block *first_else_block = nir_cf_node_as_block(first_else); + + link_blocks(block, first_then_block, first_else_block); + } else { + assert(next->type == nir_cf_node_loop); + nir_loop *next_loop = nir_cf_node_as_loop(next); + + nir_cf_node *first = nir_loop_first_cf_node(next_loop); + assert(first->type == nir_cf_node_block); + nir_block *first_block = nir_cf_node_as_block(first); + + link_blocks(block, first_block, NULL); + insert_phi_undef(first_block, block); + } + } +} + +static nir_block * +split_block_end(nir_block *block) +{ + nir_block *new_block = nir_block_create(ralloc_parent(block)); + new_block->cf_node.parent = block->cf_node.parent; + exec_node_insert_after(&block->cf_node.node, &new_block->cf_node.node); + + if (block_ends_in_jump(block)) { + /* Figure out what successor block would've had if it didn't have a jump + * instruction, and make new_block have that successor. + */ + block_add_normal_succs(new_block); + } else { + move_successors(block, new_block); + } + + return new_block; +} + +static nir_block * +split_block_before_instr(nir_instr *instr) +{ + assert(instr->type != nir_instr_type_phi); + nir_block *new_block = split_block_beginning(instr->block); + + nir_foreach_instr_safe(instr->block, cur_instr) { + if (cur_instr == instr) + break; + + exec_node_remove(&cur_instr->node); + cur_instr->block = new_block; + exec_list_push_tail(&new_block->instr_list, &cur_instr->node); + } + + return new_block; +} + +/* Splits a basic block at the point specified by the cursor. The "before" and + * "after" arguments are filled out with the blocks resulting from the split + * if non-NULL. Note that the "beginning" of the block is actually interpreted + * as before the first non-phi instruction, and it's illegal to split a block + * before a phi instruction. + */ + +static void +split_block_cursor(nir_cursor cursor, + nir_block **_before, nir_block **_after) +{ + nir_block *before, *after; + switch (cursor.option) { + case nir_cursor_before_block: + after = cursor.block; + before = split_block_beginning(cursor.block); + break; + + case nir_cursor_after_block: + before = cursor.block; + after = split_block_end(cursor.block); + break; + + case nir_cursor_before_instr: + after = cursor.instr->block; + before = split_block_before_instr(cursor.instr); + break; + + case nir_cursor_after_instr: + /* We lower this to split_block_before_instr() so that we can keep the + * after-a-jump-instr case contained to split_block_end(). + */ + if (nir_instr_is_last(cursor.instr)) { + before = cursor.instr->block; + after = split_block_end(cursor.instr->block); + } else { + after = cursor.instr->block; + before = split_block_before_instr(nir_instr_next(cursor.instr)); + } + break; + + default: + unreachable("not reached"); + } + + if (_before) + *_before = before; + if (_after) + *_after = after; +} + +/** + * Inserts a non-basic block between two basic blocks and links them together. + */ + +static void +insert_non_block(nir_block *before, nir_cf_node *node, nir_block *after) +{ + node->parent = before->cf_node.parent; + exec_node_insert_after(&before->cf_node.node, &node->node); + link_block_to_non_block(before, node); + link_non_block_to_block(node, after); +} + +/* walk up the control flow tree to find the innermost enclosed loop */ +static nir_loop * +nearest_loop(nir_cf_node *node) +{ + while (node->type != nir_cf_node_loop) { + node = node->parent; + } + + return nir_cf_node_as_loop(node); +} + +/* + * update the CFG after a jump instruction has been added to the end of a block + */ + +void +nir_handle_add_jump(nir_block *block) +{ + nir_instr *instr = nir_block_last_instr(block); + nir_jump_instr *jump_instr = nir_instr_as_jump(instr); + + unlink_block_successors(block); + + nir_function_impl *impl = nir_cf_node_get_function(&block->cf_node); + nir_metadata_preserve(impl, nir_metadata_none); + + if (jump_instr->type == nir_jump_break || + jump_instr->type == nir_jump_continue) { + nir_loop *loop = nearest_loop(&block->cf_node); + + if (jump_instr->type == nir_jump_continue) { + nir_cf_node *first_node = nir_loop_first_cf_node(loop); + assert(first_node->type == nir_cf_node_block); + nir_block *first_block = nir_cf_node_as_block(first_node); + link_blocks(block, first_block, NULL); + } else { + nir_cf_node *after = nir_cf_node_next(&loop->cf_node); + assert(after->type == nir_cf_node_block); + nir_block *after_block = nir_cf_node_as_block(after); + link_blocks(block, after_block, NULL); + + /* If we inserted a fake link, remove it */ + nir_cf_node *last = nir_loop_last_cf_node(loop); + assert(last->type == nir_cf_node_block); + nir_block *last_block = nir_cf_node_as_block(last); + if (last_block->successors[1] != NULL) + unlink_blocks(last_block, after_block); + } + } else { + assert(jump_instr->type == nir_jump_return); + link_blocks(block, impl->end_block, NULL); + } +} + +static void +remove_phi_src(nir_block *block, nir_block *pred) +{ + nir_foreach_instr(block, instr) { + if (instr->type != nir_instr_type_phi) + break; + + nir_phi_instr *phi = nir_instr_as_phi(instr); + nir_foreach_phi_src_safe(phi, src) { + if (src->pred == pred) { + list_del(&src->src.use_link); + exec_node_remove(&src->node); + } + } + } +} + +/* Removes the successor of a block with a jump, and inserts a fake edge for + * infinite loops. Note that the jump to be eliminated may be free-floating. + */ + +static void +unlink_jump(nir_block *block, nir_jump_type type, bool add_normal_successors) +{ + nir_block *next = block->successors[0]; + + if (block->successors[0]) + remove_phi_src(block->successors[0], block); + if (block->successors[1]) + remove_phi_src(block->successors[1], block); + + unlink_block_successors(block); + if (add_normal_successors) + block_add_normal_succs(block); + + /* If we've just removed a break, and the block we were jumping to (after + * the loop) now has zero predecessors, we've created a new infinite loop. + * + * NIR doesn't allow blocks (other than the start block) to have zero + * predecessors. In particular, dominance assumes all blocks are reachable. + * So, we insert a "fake link" by making successors[1] point after the loop. + * + * Note that we have to do this after unlinking/recreating the block's + * successors. If we removed a "break" at the end of the loop, then + * block == last_block, so block->successors[0] would already be "next", + * and adding a fake link would create two identical successors. Doing + * this afterward works, as we'll have changed block->successors[0] to + * be the top of the loop. + */ + if (type == nir_jump_break && next->predecessors->entries == 0) { + nir_loop *loop = + nir_cf_node_as_loop(nir_cf_node_prev(&next->cf_node)); + + /* insert fake link */ + nir_cf_node *last = nir_loop_last_cf_node(loop); + assert(last->type == nir_cf_node_block); + nir_block *last_block = nir_cf_node_as_block(last); + + last_block->successors[1] = next; + block_add_pred(next, last_block); + } +} + +void +nir_handle_remove_jump(nir_block *block, nir_jump_type type) +{ + unlink_jump(block, type, true); + + nir_function_impl *impl = nir_cf_node_get_function(&block->cf_node); + nir_metadata_preserve(impl, nir_metadata_none); +} + +static void +update_if_uses(nir_cf_node *node) +{ + if (node->type != nir_cf_node_if) + return; + + nir_if *if_stmt = nir_cf_node_as_if(node); + + if_stmt->condition.parent_if = if_stmt; + if (if_stmt->condition.is_ssa) { + list_addtail(&if_stmt->condition.use_link, + &if_stmt->condition.ssa->if_uses); + } else { + list_addtail(&if_stmt->condition.use_link, + &if_stmt->condition.reg.reg->if_uses); + } +} + +/** + * Stitch two basic blocks together into one. The aggregate must have the same + * predecessors as the first and the same successors as the second. + */ + +static void +stitch_blocks(nir_block *before, nir_block *after) +{ + /* + * We move after into before, so we have to deal with up to 2 successors vs. + * possibly a large number of predecessors. + * + * TODO: special case when before is empty and after isn't? + */ + + if (block_ends_in_jump(before)) { + assert(exec_list_is_empty(&after->instr_list)); + if (after->successors[0]) + remove_phi_src(after->successors[0], after); + if (after->successors[1]) + remove_phi_src(after->successors[1], after); + unlink_block_successors(after); + exec_node_remove(&after->cf_node.node); + } else { + move_successors(after, before); + + foreach_list_typed(nir_instr, instr, node, &after->instr_list) { + instr->block = before; + } + + exec_list_append(&before->instr_list, &after->instr_list); + exec_node_remove(&after->cf_node.node); + } +} + +void +nir_cf_node_insert(nir_cursor cursor, nir_cf_node *node) +{ + nir_block *before, *after; + + split_block_cursor(cursor, &before, &after); + + if (node->type == nir_cf_node_block) { + nir_block *block = nir_cf_node_as_block(node); + exec_node_insert_after(&before->cf_node.node, &block->cf_node.node); + block->cf_node.parent = before->cf_node.parent; + /* stitch_blocks() assumes that any block that ends with a jump has + * already been setup with the correct successors, so we need to set + * up jumps here as the block is being inserted. + */ + if (block_ends_in_jump(block)) + nir_handle_add_jump(block); + + stitch_blocks(block, after); + stitch_blocks(before, block); + } else { + update_if_uses(node); + insert_non_block(before, node, after); + } +} + +static bool +replace_ssa_def_uses(nir_ssa_def *def, void *void_impl) +{ + nir_function_impl *impl = void_impl; + void *mem_ctx = ralloc_parent(impl); + + nir_ssa_undef_instr *undef = + nir_ssa_undef_instr_create(mem_ctx, def->num_components); + nir_instr_insert_before_cf_list(&impl->body, &undef->instr); + nir_ssa_def_rewrite_uses(def, nir_src_for_ssa(&undef->def)); + return true; +} + +static void +cleanup_cf_node(nir_cf_node *node, nir_function_impl *impl) +{ + switch (node->type) { + case nir_cf_node_block: { + nir_block *block = nir_cf_node_as_block(node); + /* We need to walk the instructions and clean up defs/uses */ + nir_foreach_instr_safe(block, instr) { + if (instr->type == nir_instr_type_jump) { + nir_jump_type jump_type = nir_instr_as_jump(instr)->type; + unlink_jump(block, jump_type, false); + } else { + nir_foreach_ssa_def(instr, replace_ssa_def_uses, impl); + nir_instr_remove(instr); + } + } + break; + } + + case nir_cf_node_if: { + nir_if *if_stmt = nir_cf_node_as_if(node); + foreach_list_typed(nir_cf_node, child, node, &if_stmt->then_list) + cleanup_cf_node(child, impl); + foreach_list_typed(nir_cf_node, child, node, &if_stmt->else_list) + cleanup_cf_node(child, impl); + + list_del(&if_stmt->condition.use_link); + break; + } + + case nir_cf_node_loop: { + nir_loop *loop = nir_cf_node_as_loop(node); + foreach_list_typed(nir_cf_node, child, node, &loop->body) + cleanup_cf_node(child, impl); + break; + } + case nir_cf_node_function: { + nir_function_impl *impl = nir_cf_node_as_function(node); + foreach_list_typed(nir_cf_node, child, node, &impl->body) + cleanup_cf_node(child, impl); + break; + } + default: + unreachable("Invalid CF node type"); + } +} + +void +nir_cf_extract(nir_cf_list *extracted, nir_cursor begin, nir_cursor end) +{ + nir_block *block_begin, *block_end, *block_before, *block_after; + + /* In the case where begin points to an instruction in some basic block and + * end points to the end of the same basic block, we rely on the fact that + * splitting on an instruction moves earlier instructions into a new basic + * block. If the later instructions were moved instead, then the end cursor + * would be pointing to the same place that begin used to point to, which + * is obviously not what we want. + */ + split_block_cursor(begin, &block_before, &block_begin); + split_block_cursor(end, &block_end, &block_after); + + extracted->impl = nir_cf_node_get_function(&block_begin->cf_node); + exec_list_make_empty(&extracted->list); + + /* Dominance and other block-related information is toast. */ + nir_metadata_preserve(extracted->impl, nir_metadata_none); + + nir_cf_node *cf_node = &block_begin->cf_node; + nir_cf_node *cf_node_end = &block_end->cf_node; + while (true) { + nir_cf_node *next = nir_cf_node_next(cf_node); + + exec_node_remove(&cf_node->node); + cf_node->parent = NULL; + exec_list_push_tail(&extracted->list, &cf_node->node); + + if (cf_node == cf_node_end) + break; + + cf_node = next; + } + + stitch_blocks(block_before, block_after); +} + +void +nir_cf_reinsert(nir_cf_list *cf_list, nir_cursor cursor) +{ + nir_block *before, *after; + + split_block_cursor(cursor, &before, &after); + + foreach_list_typed_safe(nir_cf_node, node, node, &cf_list->list) { + exec_node_remove(&node->node); + node->parent = before->cf_node.parent; + exec_node_insert_node_before(&after->cf_node.node, &node->node); + } + + stitch_blocks(before, + nir_cf_node_as_block(nir_cf_node_next(&before->cf_node))); + stitch_blocks(nir_cf_node_as_block(nir_cf_node_prev(&after->cf_node)), + after); +} + +void +nir_cf_delete(nir_cf_list *cf_list) +{ + foreach_list_typed(nir_cf_node, node, node, &cf_list->list) { + cleanup_cf_node(node, cf_list->impl); + } +} diff --git a/src/compiler/nir/nir_control_flow.h b/src/compiler/nir/nir_control_flow.h new file mode 100644 index 0000000..b71382f --- /dev/null +++ b/src/compiler/nir/nir_control_flow.h @@ -0,0 +1,162 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +#include "nir.h" + +#pragma once + +#ifdef __cplusplus +extern "C" { +#endif + +/** NIR Control Flow Modification + * + * This file contains various API's that make modifying control flow in NIR, + * while maintaining the invariants checked by the validator, much easier. + * There are two parts to this: + * + * 1. Inserting control flow (if's and loops) in various places, for creating + * IR either from scratch or as part of some lowering pass. + * 2. Taking existing pieces of the IR and either moving them around or + * deleting them. + */ + +/** Control flow insertion. */ + +/** puts a control flow node where the cursor is */ +void nir_cf_node_insert(nir_cursor cursor, nir_cf_node *node); + +/** puts a control flow node immediately after another control flow node */ +static inline void +nir_cf_node_insert_after(nir_cf_node *node, nir_cf_node *after) +{ + nir_cf_node_insert(nir_after_cf_node(node), after); +} + +/** puts a control flow node immediately before another control flow node */ +static inline void +nir_cf_node_insert_before(nir_cf_node *node, nir_cf_node *before) +{ + nir_cf_node_insert(nir_before_cf_node(node), before); +} + +/** puts a control flow node at the beginning of a list from an if, loop, or function */ +static inline void +nir_cf_node_insert_begin(struct exec_list *list, nir_cf_node *node) +{ + nir_cf_node_insert(nir_before_cf_list(list), node); +} + +/** puts a control flow node at the end of a list from an if, loop, or function */ +static inline void +nir_cf_node_insert_end(struct exec_list *list, nir_cf_node *node) +{ + nir_cf_node_insert(nir_after_cf_list(list), node); +} + + +/** Control flow motion. + * + * These functions let you take a part of a control flow list (basically + * equivalent to a series of statement in GLSL) and "extract" it from the IR, + * so that it's a free-floating piece of IR that can be either re-inserted + * somewhere else or deleted entirely. A few notes on using it: + * + * 1. Phi nodes are considered attached to the piece of control flow that + * their sources come from. There are three places where phi nodes can + * occur, which are the three places where a block can have multiple + * predecessors: + * + * 1) After an if statement, if neither branch ends in a jump. + * 2) After a loop, if there are multiple break's. + * 3) At the beginning of a loop. + * + * For #1, the phi node is considered to be part of the if, and for #2 and + * #3 the phi node is considered to be part of the loop. This allows us to + * keep phi's intact, but it means that phi nodes cannot be separated from + * the control flow they come from. For example, extracting an if without + * extracting all the phi nodes after it is not allowed, and neither is + * extracting only some of the phi nodes at the beginning of a block. It + * also means that extracting from the beginning of a basic block actually + * means extracting from the first non-phi instruction, since there's no + * situation where extracting phi nodes without extracting what comes + * before them makes any sense. + * + * 2. Phi node sources are guaranteed to remain valid, meaning that they still + * correspond one-to-one with the predecessors of the basic block they're + * part of. In addition, the original sources will be preserved unless they + * correspond to a break or continue that was deleted. However, no attempt + * is made to ensure that SSA form is maintained. In particular, it is + * *not* guaranteed that definitions of SSA values will dominate all their + * uses after all is said and done. Either the caller must ensure that this + * is the case, or it must insert extra phi nodes to restore SSA. + * + * 3. It is invalid to move a piece of IR with a break/continue outside of the + * loop it references. Doing this will result in invalid + * successors/predecessors and phi node sources. + * + * 4. It is invalid to move a piece of IR from one function implementation to + * another. + * + * 5. Extracting a control flow list will leave lots of dangling references to + * and from other pieces of the IR. It also leaves things in a not 100% + * consistent state. This means that some things (e.g. inserting + * instructions) might not work reliably on the extracted control flow. It + * also means that extracting control flow without re-inserting it or + * deleting it is a Bad Thing (tm). + */ + +typedef struct { + struct exec_list list; + nir_function_impl *impl; /* for cleaning up if the list is deleted */ +} nir_cf_list; + +void nir_cf_extract(nir_cf_list *extracted, nir_cursor begin, nir_cursor end); + +void nir_cf_reinsert(nir_cf_list *cf_list, nir_cursor cursor); + +void nir_cf_delete(nir_cf_list *cf_list); + +static inline void +nir_cf_list_extract(nir_cf_list *extracted, struct exec_list *cf_list) +{ + nir_cf_extract(extracted, nir_before_cf_list(cf_list), + nir_after_cf_list(cf_list)); +} + +/** removes a control flow node, doing any cleanup necessary */ +static inline void +nir_cf_node_remove(nir_cf_node *node) +{ + nir_cf_list list; + nir_cf_extract(&list, nir_before_cf_node(node), nir_after_cf_node(node)); + nir_cf_delete(&list); +} + +#ifdef __cplusplus +} +#endif diff --git a/src/compiler/nir/nir_control_flow_private.h b/src/compiler/nir/nir_control_flow_private.h new file mode 100644 index 0000000..f32b57a --- /dev/null +++ b/src/compiler/nir/nir_control_flow_private.h @@ -0,0 +1,37 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +#include "nir_control_flow.h" + +#pragma once + +/* Internal control-flow modification functions used when inserting/removing + * instructions. + */ + +void nir_handle_add_jump(nir_block *block); +void nir_handle_remove_jump(nir_block *block, nir_jump_type type); diff --git a/src/compiler/nir/nir_dominance.c b/src/compiler/nir/nir_dominance.c new file mode 100644 index 0000000..b345b85 --- /dev/null +++ b/src/compiler/nir/nir_dominance.c @@ -0,0 +1,350 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +#include "nir.h" + +/* + * Implements the algorithms for computing the dominance tree and the + * dominance frontier from "A Simple, Fast Dominance Algorithm" by Cooper, + * Harvey, and Kennedy. + */ + +typedef struct { + nir_function_impl *impl; + bool progress; +} dom_state; + +static bool +init_block_cb(nir_block *block, void *_state) +{ + dom_state *state = (dom_state *) _state; + if (block == nir_start_block(state->impl)) + block->imm_dom = block; + else + block->imm_dom = NULL; + block->num_dom_children = 0; + + struct set_entry *entry; + set_foreach(block->dom_frontier, entry) { + _mesa_set_remove(block->dom_frontier, entry); + } + + return true; +} + +static nir_block * +intersect(nir_block *b1, nir_block *b2) +{ + while (b1 != b2) { + /* + * Note, the comparisons here are the opposite of what the paper says + * because we index blocks from beginning -> end (i.e. reverse + * post-order) instead of post-order like they assume. + */ + while (b1->index > b2->index) + b1 = b1->imm_dom; + while (b2->index > b1->index) + b2 = b2->imm_dom; + } + + return b1; +} + +static bool +calc_dominance_cb(nir_block *block, void *_state) +{ + dom_state *state = (dom_state *) _state; + if (block == nir_start_block(state->impl)) + return true; + + nir_block *new_idom = NULL; + struct set_entry *entry; + set_foreach(block->predecessors, entry) { + nir_block *pred = (nir_block *) entry->key; + + if (pred->imm_dom) { + if (new_idom) + new_idom = intersect(pred, new_idom); + else + new_idom = pred; + } + } + + assert(new_idom); + if (block->imm_dom != new_idom) { + block->imm_dom = new_idom; + state->progress = true; + } + + return true; +} + +static bool +calc_dom_frontier_cb(nir_block *block, void *state) +{ + (void) state; + + if (block->predecessors->entries > 1) { + struct set_entry *entry; + set_foreach(block->predecessors, entry) { + nir_block *runner = (nir_block *) entry->key; + while (runner != block->imm_dom) { + _mesa_set_add(runner->dom_frontier, block); + runner = runner->imm_dom; + } + } + } + + return true; +} + +/* + * Compute each node's children in the dominance tree from the immediate + * dominator information. We do this in three stages: + * + * 1. Calculate the number of children each node has + * 2. Allocate arrays, setting the number of children to 0 again + * 3. For each node, add itself to its parent's list of children, using + * num_dom_children as an index - at the end of this step, num_dom_children + * for each node will be the same as it was at the end of step #1. + */ + +static bool +block_count_children(nir_block *block, void *state) +{ + (void) state; + + if (block->imm_dom) + block->imm_dom->num_dom_children++; + + return true; +} + +static bool +block_alloc_children(nir_block *block, void *state) +{ + void *mem_ctx = state; + + block->dom_children = ralloc_array(mem_ctx, nir_block *, + block->num_dom_children); + block->num_dom_children = 0; + + return true; +} + +static bool +block_add_child(nir_block *block, void *state) +{ + (void) state; + + if (block->imm_dom) + block->imm_dom->dom_children[block->imm_dom->num_dom_children++] = block; + + return true; +} + +static void +calc_dom_children(nir_function_impl* impl) +{ + void *mem_ctx = ralloc_parent(impl); + + nir_foreach_block(impl, block_count_children, NULL); + nir_foreach_block(impl, block_alloc_children, mem_ctx); + nir_foreach_block(impl, block_add_child, NULL); +} + +static void +calc_dfs_indicies(nir_block *block, unsigned *index) +{ + block->dom_pre_index = (*index)++; + + for (unsigned i = 0; i < block->num_dom_children; i++) + calc_dfs_indicies(block->dom_children[i], index); + + block->dom_post_index = (*index)++; +} + +void +nir_calc_dominance_impl(nir_function_impl *impl) +{ + if (impl->valid_metadata & nir_metadata_dominance) + return; + + nir_metadata_require(impl, nir_metadata_block_index); + + dom_state state; + state.impl = impl; + state.progress = true; + + nir_foreach_block(impl, init_block_cb, &state); + + while (state.progress) { + state.progress = false; + nir_foreach_block(impl, calc_dominance_cb, &state); + } + + nir_foreach_block(impl, calc_dom_frontier_cb, &state); + + nir_block *start_block = nir_start_block(impl); + start_block->imm_dom = NULL; + + calc_dom_children(impl); + + unsigned dfs_index = 0; + calc_dfs_indicies(start_block, &dfs_index); +} + +void +nir_calc_dominance(nir_shader *shader) +{ + nir_foreach_function(shader, function) { + if (function->impl) + nir_calc_dominance_impl(function->impl); + } +} + +/** + * Computes the least common anscestor of two blocks. If one of the blocks + * is null, the other block is returned. + */ +nir_block * +nir_dominance_lca(nir_block *b1, nir_block *b2) +{ + if (b1 == NULL) + return b2; + + if (b2 == NULL) + return b1; + + assert(nir_cf_node_get_function(&b1->cf_node) == + nir_cf_node_get_function(&b2->cf_node)); + + assert(nir_cf_node_get_function(&b1->cf_node)->valid_metadata & + nir_metadata_dominance); + + return intersect(b1, b2); +} + +/** + * Returns true if parent dominates child + */ +bool +nir_block_dominates(nir_block *parent, nir_block *child) +{ + assert(nir_cf_node_get_function(&parent->cf_node) == + nir_cf_node_get_function(&child->cf_node)); + + assert(nir_cf_node_get_function(&parent->cf_node)->valid_metadata & + nir_metadata_dominance); + + return child->dom_pre_index >= parent->dom_pre_index && + child->dom_post_index <= parent->dom_post_index; +} + +static bool +dump_block_dom(nir_block *block, void *state) +{ + FILE *fp = state; + if (block->imm_dom) + fprintf(fp, "\t%u -> %u\n", block->imm_dom->index, block->index); + return true; +} + +void +nir_dump_dom_tree_impl(nir_function_impl *impl, FILE *fp) +{ + fprintf(fp, "digraph doms_%s {\n", impl->function->name); + nir_foreach_block(impl, dump_block_dom, fp); + fprintf(fp, "}\n\n"); +} + +void +nir_dump_dom_tree(nir_shader *shader, FILE *fp) +{ + nir_foreach_function(shader, function) { + if (function->impl) + nir_dump_dom_tree_impl(function->impl, fp); + } +} + +static bool +dump_block_dom_frontier(nir_block *block, void *state) +{ + FILE *fp = state; + + fprintf(fp, "DF(%u) = {", block->index); + struct set_entry *entry; + set_foreach(block->dom_frontier, entry) { + nir_block *df = (nir_block *) entry->key; + fprintf(fp, "%u, ", df->index); + } + fprintf(fp, "}\n"); + return true; +} + +void +nir_dump_dom_frontier_impl(nir_function_impl *impl, FILE *fp) +{ + nir_foreach_block(impl, dump_block_dom_frontier, fp); +} + +void +nir_dump_dom_frontier(nir_shader *shader, FILE *fp) +{ + nir_foreach_function(shader, function) { + if (function->impl) + nir_dump_dom_frontier_impl(function->impl, fp); + } +} + +static bool +dump_block_succs(nir_block *block, void *state) +{ + FILE *fp = state; + if (block->successors[0]) + fprintf(fp, "\t%u -> %u\n", block->index, block->successors[0]->index); + if (block->successors[1]) + fprintf(fp, "\t%u -> %u\n", block->index, block->successors[1]->index); + return true; +} + +void +nir_dump_cfg_impl(nir_function_impl *impl, FILE *fp) +{ + fprintf(fp, "digraph cfg_%s {\n", impl->function->name); + nir_foreach_block(impl, dump_block_succs, fp); + fprintf(fp, "}\n\n"); +} + +void +nir_dump_cfg(nir_shader *shader, FILE *fp) +{ + nir_foreach_function(shader, function) { + if (function->impl) + nir_dump_cfg_impl(function->impl, fp); + } +} diff --git a/src/compiler/nir/nir_from_ssa.c b/src/compiler/nir/nir_from_ssa.c new file mode 100644 index 0000000..8bc9f24 --- /dev/null +++ b/src/compiler/nir/nir_from_ssa.c @@ -0,0 +1,805 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand (jason@jlekstrand.net) + * + */ + +#include "nir.h" +#include "nir_vla.h" + +/* + * This file implements an out-of-SSA pass as described in "Revisiting + * Out-of-SSA Translation for Correctness, Code Quality, and Efficiency" by + * Boissinot et. al. + */ + +struct from_ssa_state { + void *mem_ctx; + void *dead_ctx; + bool phi_webs_only; + struct hash_table *merge_node_table; + nir_instr *instr; + nir_function_impl *impl; +}; + +/* Returns true if a dominates b */ +static bool +ssa_def_dominates(nir_ssa_def *a, nir_ssa_def *b) +{ + if (a->live_index == 0) { + /* SSA undefs always dominate */ + return true; + } else if (b->live_index < a->live_index) { + return false; + } else if (a->parent_instr->block == b->parent_instr->block) { + return a->live_index <= b->live_index; + } else { + return nir_block_dominates(a->parent_instr->block, + b->parent_instr->block); + } +} + + +/* The following data structure, which I have named merge_set is a way of + * representing a set registers of non-interfering registers. This is + * based on the concept of a "dominence forest" presented in "Fast Copy + * Coalescing and Live-Range Identification" by Budimlic et. al. but the + * implementation concept is taken from "Revisiting Out-of-SSA Translation + * for Correctness, Code Quality, and Efficiency" by Boissinot et. al.. + * + * Each SSA definition is associated with a merge_node and the association + * is represented by a combination of a hash table and the "def" parameter + * in the merge_node structure. The merge_set stores a linked list of + * merge_node's in dominence order of the ssa definitions. (Since the + * liveness analysis pass indexes the SSA values in dominence order for us, + * this is an easy thing to keep up.) It is assumed that no pair of the + * nodes in a given set interfere. Merging two sets or checking for + * interference can be done in a single linear-time merge-sort walk of the + * two lists of nodes. + */ +struct merge_set; + +typedef struct { + struct exec_node node; + struct merge_set *set; + nir_ssa_def *def; +} merge_node; + +typedef struct merge_set { + struct exec_list nodes; + unsigned size; + nir_register *reg; +} merge_set; + +#if 0 +static void +merge_set_dump(merge_set *set, FILE *fp) +{ + nir_ssa_def *dom[set->size]; + int dom_idx = -1; + + foreach_list_typed(merge_node, node, node, &set->nodes) { + while (dom_idx >= 0 && !ssa_def_dominates(dom[dom_idx], node->def)) + dom_idx--; + + for (int i = 0; i <= dom_idx; i++) + fprintf(fp, " "); + + if (node->def->name) + fprintf(fp, "ssa_%d /* %s */\n", node->def->index, node->def->name); + else + fprintf(fp, "ssa_%d\n", node->def->index); + + dom[++dom_idx] = node->def; + } +} +#endif + +static merge_node * +get_merge_node(nir_ssa_def *def, struct from_ssa_state *state) +{ + struct hash_entry *entry = + _mesa_hash_table_search(state->merge_node_table, def); + if (entry) + return entry->data; + + merge_set *set = ralloc(state->dead_ctx, merge_set); + exec_list_make_empty(&set->nodes); + set->size = 1; + set->reg = NULL; + + merge_node *node = ralloc(state->dead_ctx, merge_node); + node->set = set; + node->def = def; + exec_list_push_head(&set->nodes, &node->node); + + _mesa_hash_table_insert(state->merge_node_table, def, node); + + return node; +} + +static bool +merge_nodes_interfere(merge_node *a, merge_node *b) +{ + return nir_ssa_defs_interfere(a->def, b->def); +} + +/* Merges b into a */ +static merge_set * +merge_merge_sets(merge_set *a, merge_set *b) +{ + struct exec_node *an = exec_list_get_head(&a->nodes); + struct exec_node *bn = exec_list_get_head(&b->nodes); + while (!exec_node_is_tail_sentinel(bn)) { + merge_node *a_node = exec_node_data(merge_node, an, node); + merge_node *b_node = exec_node_data(merge_node, bn, node); + + if (exec_node_is_tail_sentinel(an) || + a_node->def->live_index > b_node->def->live_index) { + struct exec_node *next = bn->next; + exec_node_remove(bn); + exec_node_insert_node_before(an, bn); + exec_node_data(merge_node, bn, node)->set = a; + bn = next; + } else { + an = an->next; + } + } + + a->size += b->size; + b->size = 0; + + return a; +} + +/* Checks for any interference between two merge sets + * + * This is an implementation of Algorithm 2 in "Revisiting Out-of-SSA + * Translation for Correctness, Code Quality, and Efficiency" by + * Boissinot et. al. + */ +static bool +merge_sets_interfere(merge_set *a, merge_set *b) +{ + NIR_VLA(merge_node *, dom, a->size + b->size); + int dom_idx = -1; + + struct exec_node *an = exec_list_get_head(&a->nodes); + struct exec_node *bn = exec_list_get_head(&b->nodes); + while (!exec_node_is_tail_sentinel(an) || + !exec_node_is_tail_sentinel(bn)) { + + merge_node *current; + if (exec_node_is_tail_sentinel(an)) { + current = exec_node_data(merge_node, bn, node); + bn = bn->next; + } else if (exec_node_is_tail_sentinel(bn)) { + current = exec_node_data(merge_node, an, node); + an = an->next; + } else { + merge_node *a_node = exec_node_data(merge_node, an, node); + merge_node *b_node = exec_node_data(merge_node, bn, node); + + if (a_node->def->live_index <= b_node->def->live_index) { + current = a_node; + an = an->next; + } else { + current = b_node; + bn = bn->next; + } + } + + while (dom_idx >= 0 && + !ssa_def_dominates(dom[dom_idx]->def, current->def)) + dom_idx--; + + if (dom_idx >= 0 && merge_nodes_interfere(current, dom[dom_idx])) + return true; + + dom[++dom_idx] = current; + } + + return false; +} + +static bool +add_parallel_copy_to_end_of_block(nir_block *block, void *void_state) +{ + struct from_ssa_state *state = void_state; + + bool need_end_copy = false; + if (block->successors[0]) { + nir_instr *instr = nir_block_first_instr(block->successors[0]); + if (instr && instr->type == nir_instr_type_phi) + need_end_copy = true; + } + + if (block->successors[1]) { + nir_instr *instr = nir_block_first_instr(block->successors[1]); + if (instr && instr->type == nir_instr_type_phi) + need_end_copy = true; + } + + if (need_end_copy) { + /* If one of our successors has at least one phi node, we need to + * create a parallel copy at the end of the block but before the jump + * (if there is one). + */ + nir_parallel_copy_instr *pcopy = + nir_parallel_copy_instr_create(state->dead_ctx); + + nir_instr_insert(nir_after_block_before_jump(block), &pcopy->instr); + } + + return true; +} + +static nir_parallel_copy_instr * +get_parallel_copy_at_end_of_block(nir_block *block) +{ + nir_instr *last_instr = nir_block_last_instr(block); + if (last_instr == NULL) + return NULL; + + /* The last instruction may be a jump in which case the parallel copy is + * right before it. + */ + if (last_instr->type == nir_instr_type_jump) + last_instr = nir_instr_prev(last_instr); + + if (last_instr && last_instr->type == nir_instr_type_parallel_copy) + return nir_instr_as_parallel_copy(last_instr); + else + return NULL; +} + +/** Isolate phi nodes with parallel copies + * + * In order to solve the dependency problems with the sources and + * destinations of phi nodes, we first isolate them by adding parallel + * copies to the beginnings and ends of basic blocks. For every block with + * phi nodes, we add a parallel copy immediately following the last phi + * node that copies the destinations of all of the phi nodes to new SSA + * values. We also add a parallel copy to the end of every block that has + * a successor with phi nodes that, for each phi node in each successor, + * copies the corresponding sorce of the phi node and adjust the phi to + * used the destination of the parallel copy. + * + * In SSA form, each value has exactly one definition. What this does is + * ensure that each value used in a phi also has exactly one use. The + * destinations of phis are only used by the parallel copy immediately + * following the phi nodes and. Thanks to the parallel copy at the end of + * the predecessor block, the sources of phi nodes are are the only use of + * that value. This allows us to immediately assign all the sources and + * destinations of any given phi node to the same register without worrying + * about interference at all. We do coalescing to get rid of the parallel + * copies where possible. + * + * Before this pass can be run, we have to iterate over the blocks with + * add_parallel_copy_to_end_of_block to ensure that the parallel copies at + * the ends of blocks exist. We can create the ones at the beginnings as + * we go, but the ones at the ends of blocks need to be created ahead of + * time because of potential back-edges in the CFG. + */ +static bool +isolate_phi_nodes_block(nir_block *block, void *void_state) +{ + struct from_ssa_state *state = void_state; + + nir_instr *last_phi_instr = NULL; + nir_foreach_instr(block, instr) { + /* Phi nodes only ever come at the start of a block */ + if (instr->type != nir_instr_type_phi) + break; + + last_phi_instr = instr; + } + + /* If we don't have any phi's, then there's nothing for us to do. */ + if (last_phi_instr == NULL) + return true; + + /* If we have phi nodes, we need to create a parallel copy at the + * start of this block but after the phi nodes. + */ + nir_parallel_copy_instr *block_pcopy = + nir_parallel_copy_instr_create(state->dead_ctx); + nir_instr_insert_after(last_phi_instr, &block_pcopy->instr); + + nir_foreach_instr(block, instr) { + /* Phi nodes only ever come at the start of a block */ + if (instr->type != nir_instr_type_phi) + break; + + nir_phi_instr *phi = nir_instr_as_phi(instr); + assert(phi->dest.is_ssa); + nir_foreach_phi_src(phi, src) { + nir_parallel_copy_instr *pcopy = + get_parallel_copy_at_end_of_block(src->pred); + assert(pcopy); + + nir_parallel_copy_entry *entry = rzalloc(state->dead_ctx, + nir_parallel_copy_entry); + nir_ssa_dest_init(&pcopy->instr, &entry->dest, + phi->dest.ssa.num_components, src->src.ssa->name); + exec_list_push_tail(&pcopy->entries, &entry->node); + + assert(src->src.is_ssa); + nir_instr_rewrite_src(&pcopy->instr, &entry->src, src->src); + + nir_instr_rewrite_src(&phi->instr, &src->src, + nir_src_for_ssa(&entry->dest.ssa)); + } + + nir_parallel_copy_entry *entry = rzalloc(state->dead_ctx, + nir_parallel_copy_entry); + nir_ssa_dest_init(&block_pcopy->instr, &entry->dest, + phi->dest.ssa.num_components, phi->dest.ssa.name); + exec_list_push_tail(&block_pcopy->entries, &entry->node); + + nir_ssa_def_rewrite_uses(&phi->dest.ssa, + nir_src_for_ssa(&entry->dest.ssa)); + + nir_instr_rewrite_src(&block_pcopy->instr, &entry->src, + nir_src_for_ssa(&phi->dest.ssa)); + } + + return true; +} + +static bool +coalesce_phi_nodes_block(nir_block *block, void *void_state) +{ + struct from_ssa_state *state = void_state; + + nir_foreach_instr(block, instr) { + /* Phi nodes only ever come at the start of a block */ + if (instr->type != nir_instr_type_phi) + break; + + nir_phi_instr *phi = nir_instr_as_phi(instr); + + assert(phi->dest.is_ssa); + merge_node *dest_node = get_merge_node(&phi->dest.ssa, state); + + nir_foreach_phi_src(phi, src) { + assert(src->src.is_ssa); + merge_node *src_node = get_merge_node(src->src.ssa, state); + if (src_node->set != dest_node->set) + merge_merge_sets(dest_node->set, src_node->set); + } + } + + return true; +} + +static void +aggressive_coalesce_parallel_copy(nir_parallel_copy_instr *pcopy, + struct from_ssa_state *state) +{ + nir_foreach_parallel_copy_entry(pcopy, entry) { + if (!entry->src.is_ssa) + continue; + + /* Since load_const instructions are SSA only, we can't replace their + * destinations with registers and, therefore, can't coalesce them. + */ + if (entry->src.ssa->parent_instr->type == nir_instr_type_load_const) + continue; + + /* Don't try and coalesce these */ + if (entry->dest.ssa.num_components != entry->src.ssa->num_components) + continue; + + merge_node *src_node = get_merge_node(entry->src.ssa, state); + merge_node *dest_node = get_merge_node(&entry->dest.ssa, state); + + if (src_node->set == dest_node->set) + continue; + + if (!merge_sets_interfere(src_node->set, dest_node->set)) + merge_merge_sets(src_node->set, dest_node->set); + } +} + +static bool +aggressive_coalesce_block(nir_block *block, void *void_state) +{ + struct from_ssa_state *state = void_state; + + nir_parallel_copy_instr *start_pcopy = NULL; + nir_foreach_instr(block, instr) { + /* Phi nodes only ever come at the start of a block */ + if (instr->type != nir_instr_type_phi) { + if (instr->type != nir_instr_type_parallel_copy) + break; /* The parallel copy must be right after the phis */ + + start_pcopy = nir_instr_as_parallel_copy(instr); + + aggressive_coalesce_parallel_copy(start_pcopy, state); + + break; + } + } + + nir_parallel_copy_instr *end_pcopy = + get_parallel_copy_at_end_of_block(block); + + if (end_pcopy && end_pcopy != start_pcopy) + aggressive_coalesce_parallel_copy(end_pcopy, state); + + return true; +} + +static bool +rewrite_ssa_def(nir_ssa_def *def, void *void_state) +{ + struct from_ssa_state *state = void_state; + nir_register *reg; + + struct hash_entry *entry = + _mesa_hash_table_search(state->merge_node_table, def); + if (entry) { + /* In this case, we're part of a phi web. Use the web's register. */ + merge_node *node = (merge_node *)entry->data; + + /* If it doesn't have a register yet, create one. Note that all of + * the things in the merge set should be the same so it doesn't + * matter which node's definition we use. + */ + if (node->set->reg == NULL) { + node->set->reg = nir_local_reg_create(state->impl); + node->set->reg->name = def->name; + node->set->reg->num_components = def->num_components; + node->set->reg->num_array_elems = 0; + } + + reg = node->set->reg; + } else { + if (state->phi_webs_only) + return true; + + /* We leave load_const SSA values alone. They act as immediates to + * the backend. If it got coalesced into a phi, that's ok. + */ + if (def->parent_instr->type == nir_instr_type_load_const) + return true; + + reg = nir_local_reg_create(state->impl); + reg->name = def->name; + reg->num_components = def->num_components; + reg->num_array_elems = 0; + } + + nir_ssa_def_rewrite_uses(def, nir_src_for_reg(reg)); + assert(list_empty(&def->uses) && list_empty(&def->if_uses)); + + if (def->parent_instr->type == nir_instr_type_ssa_undef) { + /* If it's an ssa_undef instruction, remove it since we know we just got + * rid of all its uses. + */ + nir_instr *parent_instr = def->parent_instr; + nir_instr_remove(parent_instr); + ralloc_steal(state->dead_ctx, parent_instr); + return true; + } + + assert(def->parent_instr->type != nir_instr_type_load_const); + + /* At this point we know a priori that this SSA def is part of a + * nir_dest. We can use exec_node_data to get the dest pointer. + */ + nir_dest *dest = exec_node_data(nir_dest, def, ssa); + + nir_instr_rewrite_dest(state->instr, dest, nir_dest_for_reg(reg)); + + return true; +} + +/* Resolves ssa definitions to registers. While we're at it, we also + * remove phi nodes. + */ +static bool +resolve_registers_block(nir_block *block, void *void_state) +{ + struct from_ssa_state *state = void_state; + + nir_foreach_instr_safe(block, instr) { + state->instr = instr; + nir_foreach_ssa_def(instr, rewrite_ssa_def, state); + + if (instr->type == nir_instr_type_phi) { + nir_instr_remove(instr); + ralloc_steal(state->dead_ctx, instr); + } + } + state->instr = NULL; + + return true; +} + +static void +emit_copy(nir_parallel_copy_instr *pcopy, nir_src src, nir_src dest_src, + void *mem_ctx) +{ + assert(!dest_src.is_ssa && + dest_src.reg.indirect == NULL && + dest_src.reg.base_offset == 0); + + if (src.is_ssa) + assert(src.ssa->num_components >= dest_src.reg.reg->num_components); + else + assert(src.reg.reg->num_components >= dest_src.reg.reg->num_components); + + nir_alu_instr *mov = nir_alu_instr_create(mem_ctx, nir_op_imov); + nir_src_copy(&mov->src[0].src, &src, mov); + mov->dest.dest = nir_dest_for_reg(dest_src.reg.reg); + mov->dest.write_mask = (1 << dest_src.reg.reg->num_components) - 1; + + nir_instr_insert_before(&pcopy->instr, &mov->instr); +} + +/* Resolves a single parallel copy operation into a sequence of mov's + * + * This is based on Algorithm 1 from "Revisiting Out-of-SSA Translation for + * Correctness, Code Quality, and Efficiency" by Boissinot et. al.. + * However, I never got the algorithm to work as written, so this version + * is slightly modified. + * + * The algorithm works by playing this little shell game with the values. + * We start by recording where every source value is and which source value + * each destination value should receive. We then grab any copy whose + * destination is "empty", i.e. not used as a source, and do the following: + * - Find where its source value currently lives + * - Emit the move instruction + * - Set the location of the source value to the destination + * - Mark the location containing the source value + * - Mark the destination as no longer needing to be copied + * + * When we run out of "empty" destinations, we have a cycle and so we + * create a temporary register, copy to that register, and mark the value + * we copied as living in that temporary. Now, the cycle is broken, so we + * can continue with the above steps. + */ +static void +resolve_parallel_copy(nir_parallel_copy_instr *pcopy, + struct from_ssa_state *state) +{ + unsigned num_copies = 0; + nir_foreach_parallel_copy_entry(pcopy, entry) { + /* Sources may be SSA */ + if (!entry->src.is_ssa && entry->src.reg.reg == entry->dest.reg.reg) + continue; + + num_copies++; + } + + if (num_copies == 0) { + /* Hooray, we don't need any copies! */ + nir_instr_remove(&pcopy->instr); + return; + } + + /* The register/source corresponding to the given index */ + NIR_VLA_ZERO(nir_src, values, num_copies * 2); + + /* The current location of a given piece of data. We will use -1 for "null" */ + NIR_VLA_FILL(int, loc, num_copies * 2, -1); + + /* The piece of data that the given piece of data is to be copied from. We will use -1 for "null" */ + NIR_VLA_FILL(int, pred, num_copies * 2, -1); + + /* The destinations we have yet to properly fill */ + NIR_VLA(int, to_do, num_copies * 2); + int to_do_idx = -1; + + /* Now we set everything up: + * - All values get assigned a temporary index + * - Current locations are set from sources + * - Predicessors are recorded from sources and destinations + */ + int num_vals = 0; + nir_foreach_parallel_copy_entry(pcopy, entry) { + /* Sources may be SSA */ + if (!entry->src.is_ssa && entry->src.reg.reg == entry->dest.reg.reg) + continue; + + int src_idx = -1; + for (int i = 0; i < num_vals; ++i) { + if (nir_srcs_equal(values[i], entry->src)) + src_idx = i; + } + if (src_idx < 0) { + src_idx = num_vals++; + values[src_idx] = entry->src; + } + + nir_src dest_src = nir_src_for_reg(entry->dest.reg.reg); + + int dest_idx = -1; + for (int i = 0; i < num_vals; ++i) { + if (nir_srcs_equal(values[i], dest_src)) { + /* Each destination of a parallel copy instruction should be + * unique. A destination may get used as a source, so we still + * have to walk the list. However, the predecessor should not, + * at this point, be set yet, so we should have -1 here. + */ + assert(pred[i] == -1); + dest_idx = i; + } + } + if (dest_idx < 0) { + dest_idx = num_vals++; + values[dest_idx] = dest_src; + } + + loc[src_idx] = src_idx; + pred[dest_idx] = src_idx; + + to_do[++to_do_idx] = dest_idx; + } + + /* Currently empty destinations we can go ahead and fill */ + NIR_VLA(int, ready, num_copies * 2); + int ready_idx = -1; + + /* Mark the ones that are ready for copying. We know an index is a + * destination if it has a predecessor and it's ready for copying if + * it's not marked as containing data. + */ + for (int i = 0; i < num_vals; i++) { + if (pred[i] != -1 && loc[i] == -1) + ready[++ready_idx] = i; + } + + while (to_do_idx >= 0) { + while (ready_idx >= 0) { + int b = ready[ready_idx--]; + int a = pred[b]; + emit_copy(pcopy, values[loc[a]], values[b], state->mem_ctx); + + /* If any other copies want a they can find it at b */ + loc[a] = b; + + /* b has been filled, mark it as not needing to be copied */ + pred[b] = -1; + + /* If a needs to be filled, it's ready for copying now */ + if (pred[a] != -1) + ready[++ready_idx] = a; + } + int b = to_do[to_do_idx--]; + if (pred[b] == -1) + continue; + + /* If we got here, then we don't have any more trivial copies that we + * can do. We have to break a cycle, so we create a new temporary + * register for that purpose. Normally, if going out of SSA after + * register allocation, you would want to avoid creating temporary + * registers. However, we are going out of SSA before register + * allocation, so we would rather not create extra register + * dependencies for the backend to deal with. If it wants, the + * backend can coalesce the (possibly multiple) temporaries. + */ + assert(num_vals < num_copies * 2); + nir_register *reg = nir_local_reg_create(state->impl); + reg->name = "copy_temp"; + reg->num_array_elems = 0; + if (values[b].is_ssa) + reg->num_components = values[b].ssa->num_components; + else + reg->num_components = values[b].reg.reg->num_components; + values[num_vals].is_ssa = false; + values[num_vals].reg.reg = reg; + + emit_copy(pcopy, values[b], values[num_vals], state->mem_ctx); + loc[b] = num_vals; + ready[++ready_idx] = b; + num_vals++; + } + + nir_instr_remove(&pcopy->instr); +} + +/* Resolves the parallel copies in a block. Each block can have at most + * two: One at the beginning, right after all the phi noces, and one at + * the end (or right before the final jump if it exists). + */ +static bool +resolve_parallel_copies_block(nir_block *block, void *void_state) +{ + struct from_ssa_state *state = void_state; + + /* At this point, we have removed all of the phi nodes. If a parallel + * copy existed right after the phi nodes in this block, it is now the + * first instruction. + */ + nir_instr *first_instr = nir_block_first_instr(block); + if (first_instr == NULL) + return true; /* Empty, nothing to do. */ + + if (first_instr->type == nir_instr_type_parallel_copy) { + nir_parallel_copy_instr *pcopy = nir_instr_as_parallel_copy(first_instr); + + resolve_parallel_copy(pcopy, state); + } + + /* It's possible that the above code already cleaned up the end parallel + * copy. However, doing so removed it form the instructions list so we + * won't find it here. Therefore, it's safe to go ahead and just look + * for one and clean it up if it exists. + */ + nir_parallel_copy_instr *end_pcopy = + get_parallel_copy_at_end_of_block(block); + if (end_pcopy) + resolve_parallel_copy(end_pcopy, state); + + return true; +} + +static void +nir_convert_from_ssa_impl(nir_function_impl *impl, bool phi_webs_only) +{ + struct from_ssa_state state; + + state.mem_ctx = ralloc_parent(impl); + state.dead_ctx = ralloc_context(NULL); + state.impl = impl; + state.phi_webs_only = phi_webs_only; + state.merge_node_table = _mesa_hash_table_create(NULL, _mesa_hash_pointer, + _mesa_key_pointer_equal); + + nir_foreach_block(impl, add_parallel_copy_to_end_of_block, &state); + nir_foreach_block(impl, isolate_phi_nodes_block, &state); + + /* Mark metadata as dirty before we ask for liveness analysis */ + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance); + + nir_metadata_require(impl, nir_metadata_live_ssa_defs | + nir_metadata_dominance); + + nir_foreach_block(impl, coalesce_phi_nodes_block, &state); + nir_foreach_block(impl, aggressive_coalesce_block, &state); + + nir_foreach_block(impl, resolve_registers_block, &state); + + nir_foreach_block(impl, resolve_parallel_copies_block, &state); + + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance); + + /* Clean up dead instructions and the hash tables */ + _mesa_hash_table_destroy(state.merge_node_table, NULL); + ralloc_free(state.dead_ctx); +} + +void +nir_convert_from_ssa(nir_shader *shader, bool phi_webs_only) +{ + nir_foreach_function(shader, function) { + if (function->impl) + nir_convert_from_ssa_impl(function->impl, phi_webs_only); + } +} diff --git a/src/compiler/nir/nir_gs_count_vertices.c b/src/compiler/nir/nir_gs_count_vertices.c new file mode 100644 index 0000000..db15d16 --- /dev/null +++ b/src/compiler/nir/nir_gs_count_vertices.c @@ -0,0 +1,93 @@ +/* + * Copyright © 2015 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + */ + +#include "nir.h" +#include "nir_builder.h" + +static nir_intrinsic_instr * +as_intrinsic(nir_instr *instr, nir_intrinsic_op op) +{ + if (instr->type != nir_instr_type_intrinsic) + return NULL; + + nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); + if (intrin->intrinsic != op) + return NULL; + + return intrin; +} + +static nir_intrinsic_instr * +as_set_vertex_count(nir_instr *instr) +{ + return as_intrinsic(instr, nir_intrinsic_set_vertex_count); +} + +/** + * If a geometry shader emits a constant number of vertices, return the + * number of vertices. Otherwise, return -1 (unknown). + * + * This only works if you've used nir_lower_gs_intrinsics() to do vertex + * counting at the NIR level. + */ +int +nir_gs_count_vertices(const nir_shader *shader) +{ + int count = -1; + + nir_foreach_function(shader, function) { + if (!function->impl) + continue; + + /* set_vertex_count intrinsics only appear in predecessors of the + * end block. So we don't need to walk all of them. + */ + struct set_entry *entry; + set_foreach(function->impl->end_block->predecessors, entry) { + nir_block *block = (nir_block *) entry->key; + + nir_foreach_instr_reverse(block, instr) { + nir_intrinsic_instr *intrin = as_set_vertex_count(instr); + if (!intrin) + continue; + + nir_const_value *val = nir_src_as_const_value(intrin->src[0]); + /* We've found a non-constant value. Bail. */ + if (!val) + return -1; + + if (count == -1) + count = val->i[0]; + + /* We've found contradictory set_vertex_count intrinsics. + * This can happen if there are early-returns in main() and + * different paths emit different numbers of vertices. + */ + if (count != val->i[0]) + return -1; + } + } + } + + return count; +} diff --git a/src/compiler/nir/nir_instr_set.c b/src/compiler/nir/nir_instr_set.c new file mode 100644 index 0000000..d3f939f --- /dev/null +++ b/src/compiler/nir/nir_instr_set.c @@ -0,0 +1,519 @@ +/* + * Copyright © 2014 Connor Abbott + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + */ + +#include "nir_instr_set.h" +#include "nir_vla.h" + +#define HASH(hash, data) _mesa_fnv32_1a_accumulate((hash), (data)) + +static uint32_t +hash_src(uint32_t hash, const nir_src *src) +{ + assert(src->is_ssa); + hash = HASH(hash, src->ssa); + return hash; +} + +static uint32_t +hash_alu_src(uint32_t hash, const nir_alu_src *src, unsigned num_components) +{ + hash = HASH(hash, src->abs); + hash = HASH(hash, src->negate); + + for (unsigned i = 0; i < num_components; i++) + hash = HASH(hash, src->swizzle[i]); + + hash = hash_src(hash, &src->src); + return hash; +} + +static uint32_t +hash_alu(uint32_t hash, const nir_alu_instr *instr) +{ + hash = HASH(hash, instr->op); + hash = HASH(hash, instr->dest.dest.ssa.num_components); + + if (nir_op_infos[instr->op].algebraic_properties & NIR_OP_IS_COMMUTATIVE) { + assert(nir_op_infos[instr->op].num_inputs == 2); + uint32_t hash0 = hash_alu_src(hash, &instr->src[0], + nir_ssa_alu_instr_src_components(instr, 0)); + uint32_t hash1 = hash_alu_src(hash, &instr->src[1], + nir_ssa_alu_instr_src_components(instr, 1)); + /* For commutative operations, we need some commutative way of + * combining the hashes. One option would be to XOR them but that + * means that anything with two identical sources will hash to 0 and + * that's common enough we probably don't want the guaranteed + * collision. Either addition or multiplication will also work. + */ + hash = hash0 * hash1; + } else { + for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) { + hash = hash_alu_src(hash, &instr->src[i], + nir_ssa_alu_instr_src_components(instr, i)); + } + } + + return hash; +} + +static uint32_t +hash_load_const(uint32_t hash, const nir_load_const_instr *instr) +{ + hash = HASH(hash, instr->def.num_components); + + hash = _mesa_fnv32_1a_accumulate_block(hash, instr->value.f, + instr->def.num_components + * sizeof(instr->value.f[0])); + + return hash; +} + +static int +cmp_phi_src(const void *data1, const void *data2) +{ + nir_phi_src *src1 = *(nir_phi_src **)data1; + nir_phi_src *src2 = *(nir_phi_src **)data2; + return src1->pred - src2->pred; +} + +static uint32_t +hash_phi(uint32_t hash, const nir_phi_instr *instr) +{ + hash = HASH(hash, instr->instr.block); + + /* sort sources by predecessor, since the order shouldn't matter */ + unsigned num_preds = instr->instr.block->predecessors->entries; + NIR_VLA(nir_phi_src *, srcs, num_preds); + unsigned i = 0; + nir_foreach_phi_src(instr, src) { + srcs[i++] = src; + } + + qsort(srcs, num_preds, sizeof(nir_phi_src *), cmp_phi_src); + + for (i = 0; i < num_preds; i++) { + hash = hash_src(hash, &srcs[i]->src); + hash = HASH(hash, srcs[i]->pred); + } + + return hash; +} + +static uint32_t +hash_intrinsic(uint32_t hash, const nir_intrinsic_instr *instr) +{ + const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; + hash = HASH(hash, instr->intrinsic); + + if (info->has_dest) + hash = HASH(hash, instr->dest.ssa.num_components); + + assert(info->num_variables == 0); + + hash = _mesa_fnv32_1a_accumulate_block(hash, instr->const_index, + info->num_indices + * sizeof(instr->const_index[0])); + return hash; +} + +static uint32_t +hash_tex(uint32_t hash, const nir_tex_instr *instr) +{ + hash = HASH(hash, instr->op); + hash = HASH(hash, instr->num_srcs); + + for (unsigned i = 0; i < instr->num_srcs; i++) { + hash = HASH(hash, instr->src[i].src_type); + hash = hash_src(hash, &instr->src[i].src); + } + + hash = HASH(hash, instr->coord_components); + hash = HASH(hash, instr->sampler_dim); + hash = HASH(hash, instr->is_array); + hash = HASH(hash, instr->is_shadow); + hash = HASH(hash, instr->is_new_style_shadow); + hash = HASH(hash, instr->const_offset); + unsigned component = instr->component; + hash = HASH(hash, component); + hash = HASH(hash, instr->sampler_index); + hash = HASH(hash, instr->sampler_array_size); + + assert(!instr->sampler); + + return hash; +} + +/* Computes a hash of an instruction for use in a hash table. Note that this + * will only work for instructions where instr_can_rewrite() returns true, and + * it should return identical hashes for two instructions that are the same + * according nir_instrs_equal(). + */ + +static uint32_t +hash_instr(const void *data) +{ + const nir_instr *instr = data; + uint32_t hash = _mesa_fnv32_1a_offset_bias; + + switch (instr->type) { + case nir_instr_type_alu: + hash = hash_alu(hash, nir_instr_as_alu(instr)); + break; + case nir_instr_type_load_const: + hash = hash_load_const(hash, nir_instr_as_load_const(instr)); + break; + case nir_instr_type_phi: + hash = hash_phi(hash, nir_instr_as_phi(instr)); + break; + case nir_instr_type_intrinsic: + hash = hash_intrinsic(hash, nir_instr_as_intrinsic(instr)); + break; + case nir_instr_type_tex: + hash = hash_tex(hash, nir_instr_as_tex(instr)); + break; + default: + unreachable("Invalid instruction type"); + } + + return hash; +} + +bool +nir_srcs_equal(nir_src src1, nir_src src2) +{ + if (src1.is_ssa) { + if (src2.is_ssa) { + return src1.ssa == src2.ssa; + } else { + return false; + } + } else { + if (src2.is_ssa) { + return false; + } else { + if ((src1.reg.indirect == NULL) != (src2.reg.indirect == NULL)) + return false; + + if (src1.reg.indirect) { + if (!nir_srcs_equal(*src1.reg.indirect, *src2.reg.indirect)) + return false; + } + + return src1.reg.reg == src2.reg.reg && + src1.reg.base_offset == src2.reg.base_offset; + } + } +} + +static bool +nir_alu_srcs_equal(const nir_alu_instr *alu1, const nir_alu_instr *alu2, + unsigned src1, unsigned src2) +{ + if (alu1->src[src1].abs != alu2->src[src2].abs || + alu1->src[src1].negate != alu2->src[src2].negate) + return false; + + for (unsigned i = 0; i < nir_ssa_alu_instr_src_components(alu1, src1); i++) { + if (alu1->src[src1].swizzle[i] != alu2->src[src2].swizzle[i]) + return false; + } + + return nir_srcs_equal(alu1->src[src1].src, alu2->src[src2].src); +} + +/* Returns "true" if two instructions are equal. Note that this will only + * work for the subset of instructions defined by instr_can_rewrite(). Also, + * it should only return "true" for instructions that hash_instr() will return + * the same hash for (ignoring collisions, of course). + */ + +static bool +nir_instrs_equal(const nir_instr *instr1, const nir_instr *instr2) +{ + if (instr1->type != instr2->type) + return false; + + switch (instr1->type) { + case nir_instr_type_alu: { + nir_alu_instr *alu1 = nir_instr_as_alu(instr1); + nir_alu_instr *alu2 = nir_instr_as_alu(instr2); + + if (alu1->op != alu2->op) + return false; + + /* TODO: We can probably acutally do something more inteligent such + * as allowing different numbers and taking a maximum or something + * here */ + if (alu1->dest.dest.ssa.num_components != alu2->dest.dest.ssa.num_components) + return false; + + if (nir_op_infos[alu1->op].algebraic_properties & NIR_OP_IS_COMMUTATIVE) { + assert(nir_op_infos[alu1->op].num_inputs == 2); + return (nir_alu_srcs_equal(alu1, alu2, 0, 0) && + nir_alu_srcs_equal(alu1, alu2, 1, 1)) || + (nir_alu_srcs_equal(alu1, alu2, 0, 1) && + nir_alu_srcs_equal(alu1, alu2, 1, 0)); + } else { + for (unsigned i = 0; i < nir_op_infos[alu1->op].num_inputs; i++) { + if (!nir_alu_srcs_equal(alu1, alu2, i, i)) + return false; + } + } + return true; + } + case nir_instr_type_tex: { + nir_tex_instr *tex1 = nir_instr_as_tex(instr1); + nir_tex_instr *tex2 = nir_instr_as_tex(instr2); + + if (tex1->op != tex2->op) + return false; + + if (tex1->num_srcs != tex2->num_srcs) + return false; + for (unsigned i = 0; i < tex1->num_srcs; i++) { + if (tex1->src[i].src_type != tex2->src[i].src_type || + !nir_srcs_equal(tex1->src[i].src, tex2->src[i].src)) { + return false; + } + } + + if (tex1->coord_components != tex2->coord_components || + tex1->sampler_dim != tex2->sampler_dim || + tex1->is_array != tex2->is_array || + tex1->is_shadow != tex2->is_shadow || + tex1->is_new_style_shadow != tex2->is_new_style_shadow || + memcmp(tex1->const_offset, tex2->const_offset, + sizeof(tex1->const_offset)) != 0 || + tex1->component != tex2->component || + tex1->sampler_index != tex2->sampler_index || + tex1->sampler_array_size != tex2->sampler_array_size) { + return false; + } + + /* Don't support un-lowered sampler derefs currently. */ + assert(!tex1->sampler && !tex2->sampler); + + return true; + } + case nir_instr_type_load_const: { + nir_load_const_instr *load1 = nir_instr_as_load_const(instr1); + nir_load_const_instr *load2 = nir_instr_as_load_const(instr2); + + if (load1->def.num_components != load2->def.num_components) + return false; + + return memcmp(load1->value.f, load2->value.f, + load1->def.num_components * sizeof(*load2->value.f)) == 0; + } + case nir_instr_type_phi: { + nir_phi_instr *phi1 = nir_instr_as_phi(instr1); + nir_phi_instr *phi2 = nir_instr_as_phi(instr2); + + if (phi1->instr.block != phi2->instr.block) + return false; + + nir_foreach_phi_src(phi1, src1) { + nir_foreach_phi_src(phi2, src2) { + if (src1->pred == src2->pred) { + if (!nir_srcs_equal(src1->src, src2->src)) + return false; + + break; + } + } + } + + return true; + } + case nir_instr_type_intrinsic: { + nir_intrinsic_instr *intrinsic1 = nir_instr_as_intrinsic(instr1); + nir_intrinsic_instr *intrinsic2 = nir_instr_as_intrinsic(instr2); + const nir_intrinsic_info *info = + &nir_intrinsic_infos[intrinsic1->intrinsic]; + + if (intrinsic1->intrinsic != intrinsic2->intrinsic || + intrinsic1->num_components != intrinsic2->num_components) + return false; + + if (info->has_dest && intrinsic1->dest.ssa.num_components != + intrinsic2->dest.ssa.num_components) + return false; + + for (unsigned i = 0; i < info->num_srcs; i++) { + if (!nir_srcs_equal(intrinsic1->src[i], intrinsic2->src[i])) + return false; + } + + assert(info->num_variables == 0); + + for (unsigned i = 0; i < info->num_indices; i++) { + if (intrinsic1->const_index[i] != intrinsic2->const_index[i]) + return false; + } + + return true; + } + case nir_instr_type_call: + case nir_instr_type_jump: + case nir_instr_type_ssa_undef: + case nir_instr_type_parallel_copy: + default: + unreachable("Invalid instruction type"); + } + + return false; +} + +static bool +src_is_ssa(nir_src *src, void *data) +{ + (void) data; + return src->is_ssa; +} + +static bool +dest_is_ssa(nir_dest *dest, void *data) +{ + (void) data; + return dest->is_ssa; +} + +/* This function determines if uses of an instruction can safely be rewritten + * to use another identical instruction instead. Note that this function must + * be kept in sync with hash_instr() and nir_instrs_equal() -- only + * instructions that pass this test will be handed on to those functions, and + * conversely they must handle everything that this function returns true for. + */ + +static bool +instr_can_rewrite(nir_instr *instr) +{ + /* We only handle SSA. */ + if (!nir_foreach_dest(instr, dest_is_ssa, NULL) || + !nir_foreach_src(instr, src_is_ssa, NULL)) + return false; + + switch (instr->type) { + case nir_instr_type_alu: + case nir_instr_type_load_const: + case nir_instr_type_phi: + return true; + case nir_instr_type_tex: { + nir_tex_instr *tex = nir_instr_as_tex(instr); + + /* Don't support un-lowered sampler derefs currently. */ + if (tex->sampler) + return false; + + return true; + } + case nir_instr_type_intrinsic: { + const nir_intrinsic_info *info = + &nir_intrinsic_infos[nir_instr_as_intrinsic(instr)->intrinsic]; + return (info->flags & NIR_INTRINSIC_CAN_ELIMINATE) && + (info->flags & NIR_INTRINSIC_CAN_REORDER) && + info->num_variables == 0; /* not implemented yet */ + } + case nir_instr_type_call: + case nir_instr_type_jump: + case nir_instr_type_ssa_undef: + return false; + case nir_instr_type_parallel_copy: + default: + unreachable("Invalid instruction type"); + } + + return false; +} + +static nir_ssa_def * +nir_instr_get_dest_ssa_def(nir_instr *instr) +{ + switch (instr->type) { + case nir_instr_type_alu: + assert(nir_instr_as_alu(instr)->dest.dest.is_ssa); + return &nir_instr_as_alu(instr)->dest.dest.ssa; + case nir_instr_type_load_const: + return &nir_instr_as_load_const(instr)->def; + case nir_instr_type_phi: + assert(nir_instr_as_phi(instr)->dest.is_ssa); + return &nir_instr_as_phi(instr)->dest.ssa; + case nir_instr_type_intrinsic: + assert(nir_instr_as_intrinsic(instr)->dest.is_ssa); + return &nir_instr_as_intrinsic(instr)->dest.ssa; + case nir_instr_type_tex: + assert(nir_instr_as_tex(instr)->dest.is_ssa); + return &nir_instr_as_tex(instr)->dest.ssa; + default: + unreachable("We never ask for any of these"); + } +} + +static bool +cmp_func(const void *data1, const void *data2) +{ + return nir_instrs_equal(data1, data2); +} + +struct set * +nir_instr_set_create(void *mem_ctx) +{ + return _mesa_set_create(mem_ctx, hash_instr, cmp_func); +} + +void +nir_instr_set_destroy(struct set *instr_set) +{ + _mesa_set_destroy(instr_set, NULL); +} + +bool +nir_instr_set_add_or_rewrite(struct set *instr_set, nir_instr *instr) +{ + if (!instr_can_rewrite(instr)) + return false; + + struct set_entry *entry = _mesa_set_search(instr_set, instr); + if (entry) { + nir_ssa_def *def = nir_instr_get_dest_ssa_def(instr); + nir_ssa_def *new_def = + nir_instr_get_dest_ssa_def((nir_instr *) entry->key); + nir_ssa_def_rewrite_uses(def, nir_src_for_ssa(new_def)); + return true; + } + + _mesa_set_add(instr_set, instr); + return false; +} + +void +nir_instr_set_remove(struct set *instr_set, nir_instr *instr) +{ + if (!instr_can_rewrite(instr)) + return; + + struct set_entry *entry = _mesa_set_search(instr_set, instr); + if (entry) + _mesa_set_remove(instr_set, entry); +} + diff --git a/src/compiler/nir/nir_instr_set.h b/src/compiler/nir/nir_instr_set.h new file mode 100644 index 0000000..939e8dd --- /dev/null +++ b/src/compiler/nir/nir_instr_set.h @@ -0,0 +1,62 @@ +/* + * Copyright © 2014 Connor Abbott + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + */ + +#pragma once + +#include "nir.h" + +/** + * This file defines functions for creating, destroying, and manipulating an + * "instruction set," which is an abstraction for finding duplicate + * instructions using a hash set. Note that the question of whether an + * instruction is actually a duplicate (e.g. whether it has any side effects) + * is handled transparently. The user can pass any instruction to + * nir_instr_set_add_or_rewrite() and nir_instr_set_remove(), and if the + * instruction isn't safe to rewrite or isn't supported, it's silently + * removed. + */ + +/*@{*/ + +/** Creates an instruction set, using a given ralloc mem_ctx */ +struct set *nir_instr_set_create(void *mem_ctx); + +/** Destroys an instruction set. */ +void nir_instr_set_destroy(struct set *instr_set); + +/** + * Adds an instruction to an instruction set if it doesn't exist, or if it + * does already exist, rewrites all uses of it to point to the other + * already-inserted instruction. Returns 'true' if the uses of the instruction + * were rewritten. + */ +bool nir_instr_set_add_or_rewrite(struct set *instr_set, nir_instr *instr); + +/** + * Removes an instruction from an instruction set, so that other instructions + * won't be merged with it. + */ +void nir_instr_set_remove(struct set *instr_set, nir_instr *instr); + +/*@}*/ + diff --git a/src/compiler/nir/nir_intrinsics.c b/src/compiler/nir/nir_intrinsics.c new file mode 100644 index 0000000..a7c868c --- /dev/null +++ b/src/compiler/nir/nir_intrinsics.c @@ -0,0 +1,49 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +#include "nir.h" + +#define OPCODE(name) nir_intrinsic_##name + +#define INTRINSIC(_name, _num_srcs, _src_components, _has_dest, \ + _dest_components, _num_variables, _num_indices, _flags) \ +{ \ + .name = #_name, \ + .num_srcs = _num_srcs, \ + .src_components = _src_components, \ + .has_dest = _has_dest, \ + .dest_components = _dest_components, \ + .num_variables = _num_variables, \ + .num_indices = _num_indices, \ + .flags = _flags \ +}, + +#define LAST_INTRINSIC(name) + +const nir_intrinsic_info nir_intrinsic_infos[nir_num_intrinsics] = { +#include "nir_intrinsics.h" +};
\ No newline at end of file diff --git a/src/compiler/nir/nir_intrinsics.h b/src/compiler/nir/nir_intrinsics.h new file mode 100644 index 0000000..62eead4 --- /dev/null +++ b/src/compiler/nir/nir_intrinsics.h @@ -0,0 +1,316 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +/** + * This header file defines all the available intrinsics in one place. It + * expands to a list of macros of the form: + * + * INTRINSIC(name, num_srcs, src_components, has_dest, dest_components, + * num_variables, num_indices, flags) + * + * Which should correspond one-to-one with the nir_intrinsic_info structure. It + * is included in both ir.h to create the nir_intrinsic enum (with members of + * the form nir_intrinsic_(name)) and and in opcodes.c to create + * nir_intrinsic_infos, which is a const array of nir_intrinsic_info structures + * for each intrinsic. + */ + +#define ARR(...) { __VA_ARGS__ } + + +INTRINSIC(load_var, 0, ARR(), true, 0, 1, 0, NIR_INTRINSIC_CAN_ELIMINATE) +INTRINSIC(store_var, 1, ARR(0), false, 0, 1, 1, 0) +INTRINSIC(copy_var, 0, ARR(), false, 0, 2, 0, 0) + +/* + * Interpolation of input. The interp_var_at* intrinsics are similar to the + * load_var intrinsic acting an a shader input except that they interpolate + * the input differently. The at_sample and at_offset intrinsics take an + * aditional source that is a integer sample id or a vec2 position offset + * respectively. + */ + +INTRINSIC(interp_var_at_centroid, 0, ARR(0), true, 0, 1, 0, + NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) +INTRINSIC(interp_var_at_sample, 1, ARR(1), true, 0, 1, 0, + NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) +INTRINSIC(interp_var_at_offset, 1, ARR(2), true, 0, 1, 0, + NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) + +/* + * Ask the driver for the size of a given buffer. It takes the buffer index + * as source. + */ +INTRINSIC(get_buffer_size, 1, ARR(1), true, 1, 0, 0, + NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) + +/* + * a barrier is an intrinsic with no inputs/outputs but which can't be moved + * around/optimized in general + */ +#define BARRIER(name) INTRINSIC(name, 0, ARR(), false, 0, 0, 0, 0) + +BARRIER(barrier) +BARRIER(discard) + +/* + * Memory barrier with semantics analogous to the memoryBarrier() GLSL + * intrinsic. + */ +BARRIER(memory_barrier) + +/* + * Shader clock intrinsic with semantics analogous to the clock2x32ARB() + * GLSL intrinsic. + * The latter can be used as code motion barrier, which is currently not + * feasible with NIR. + */ +INTRINSIC(shader_clock, 0, ARR(), true, 1, 0, 0, NIR_INTRINSIC_CAN_ELIMINATE) + +/* + * Memory barrier with semantics analogous to the compute shader + * groupMemoryBarrier(), memoryBarrierAtomicCounter(), memoryBarrierBuffer(), + * memoryBarrierImage() and memoryBarrierShared() GLSL intrinsics. + */ +BARRIER(group_memory_barrier) +BARRIER(memory_barrier_atomic_counter) +BARRIER(memory_barrier_buffer) +BARRIER(memory_barrier_image) +BARRIER(memory_barrier_shared) + +/** A conditional discard, with a single boolean source. */ +INTRINSIC(discard_if, 1, ARR(1), false, 0, 0, 0, 0) + +/** + * Basic Geometry Shader intrinsics. + * + * emit_vertex implements GLSL's EmitStreamVertex() built-in. It takes a single + * index, which is the stream ID to write to. + * + * end_primitive implements GLSL's EndPrimitive() built-in. + */ +INTRINSIC(emit_vertex, 0, ARR(), false, 0, 0, 1, 0) +INTRINSIC(end_primitive, 0, ARR(), false, 0, 0, 1, 0) + +/** + * Geometry Shader intrinsics with a vertex count. + * + * Alternatively, drivers may implement these intrinsics, and use + * nir_lower_gs_intrinsics() to convert from the basic intrinsics. + * + * These maintain a count of the number of vertices emitted, as an additional + * unsigned integer source. + */ +INTRINSIC(emit_vertex_with_counter, 1, ARR(1), false, 0, 0, 1, 0) +INTRINSIC(end_primitive_with_counter, 1, ARR(1), false, 0, 0, 1, 0) +INTRINSIC(set_vertex_count, 1, ARR(1), false, 0, 0, 0, 0) + +/* + * Atomic counters + * + * The *_var variants take an atomic_uint nir_variable, while the other, + * lowered, variants take a constant buffer index and register offset. + */ + +#define ATOMIC(name, flags) \ + INTRINSIC(atomic_counter_##name##_var, 0, ARR(), true, 1, 1, 0, flags) \ + INTRINSIC(atomic_counter_##name, 1, ARR(1), true, 1, 0, 1, flags) + +ATOMIC(inc, 0) +ATOMIC(dec, 0) +ATOMIC(read, NIR_INTRINSIC_CAN_ELIMINATE) + +/* + * Image load, store and atomic intrinsics. + * + * All image intrinsics take an image target passed as a nir_variable. Image + * variables contain a number of memory and layout qualifiers that influence + * the semantics of the intrinsic. + * + * All image intrinsics take a four-coordinate vector and a sample index as + * first two sources, determining the location within the image that will be + * accessed by the intrinsic. Components not applicable to the image target + * in use are undefined. Image store takes an additional four-component + * argument with the value to be written, and image atomic operations take + * either one or two additional scalar arguments with the same meaning as in + * the ARB_shader_image_load_store specification. + */ +INTRINSIC(image_load, 2, ARR(4, 1), true, 4, 1, 0, + NIR_INTRINSIC_CAN_ELIMINATE) +INTRINSIC(image_store, 3, ARR(4, 1, 4), false, 0, 1, 0, 0) +INTRINSIC(image_atomic_add, 3, ARR(4, 1, 1), true, 1, 1, 0, 0) +INTRINSIC(image_atomic_min, 3, ARR(4, 1, 1), true, 1, 1, 0, 0) +INTRINSIC(image_atomic_max, 3, ARR(4, 1, 1), true, 1, 1, 0, 0) +INTRINSIC(image_atomic_and, 3, ARR(4, 1, 1), true, 1, 1, 0, 0) +INTRINSIC(image_atomic_or, 3, ARR(4, 1, 1), true, 1, 1, 0, 0) +INTRINSIC(image_atomic_xor, 3, ARR(4, 1, 1), true, 1, 1, 0, 0) +INTRINSIC(image_atomic_exchange, 3, ARR(4, 1, 1), true, 1, 1, 0, 0) +INTRINSIC(image_atomic_comp_swap, 4, ARR(4, 1, 1, 1), true, 1, 1, 0, 0) +INTRINSIC(image_size, 0, ARR(), true, 4, 1, 0, + NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) +INTRINSIC(image_samples, 0, ARR(), true, 1, 1, 0, + NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) + +/* + * SSBO atomic intrinsics + * + * All of the SSBO atomic memory operations read a value from memory, + * compute a new value using one of the operations below, write the new + * value to memory, and return the original value read. + * + * All operations take 3 sources except CompSwap that takes 4. These + * sources represent: + * + * 0: The SSBO buffer index. + * 1: The offset into the SSBO buffer of the variable that the atomic + * operation will operate on. + * 2: The data parameter to the atomic function (i.e. the value to add + * in ssbo_atomic_add, etc). + * 3: For CompSwap only: the second data parameter. + */ +INTRINSIC(ssbo_atomic_add, 3, ARR(1, 1, 1), true, 1, 0, 0, 0) +INTRINSIC(ssbo_atomic_imin, 3, ARR(1, 1, 1), true, 1, 0, 0, 0) +INTRINSIC(ssbo_atomic_umin, 3, ARR(1, 1, 1), true, 1, 0, 0, 0) +INTRINSIC(ssbo_atomic_imax, 3, ARR(1, 1, 1), true, 1, 0, 0, 0) +INTRINSIC(ssbo_atomic_umax, 3, ARR(1, 1, 1), true, 1, 0, 0, 0) +INTRINSIC(ssbo_atomic_and, 3, ARR(1, 1, 1), true, 1, 0, 0, 0) +INTRINSIC(ssbo_atomic_or, 3, ARR(1, 1, 1), true, 1, 0, 0, 0) +INTRINSIC(ssbo_atomic_xor, 3, ARR(1, 1, 1), true, 1, 0, 0, 0) +INTRINSIC(ssbo_atomic_exchange, 3, ARR(1, 1, 1), true, 1, 0, 0, 0) +INTRINSIC(ssbo_atomic_comp_swap, 4, ARR(1, 1, 1, 1), true, 1, 0, 0, 0) + +/* + * CS shared variable atomic intrinsics + * + * All of the shared variable atomic memory operations read a value from + * memory, compute a new value using one of the operations below, write the + * new value to memory, and return the original value read. + * + * All operations take 2 sources except CompSwap that takes 3. These + * sources represent: + * + * 0: The offset into the shared variable storage region that the atomic + * operation will operate on. + * 1: The data parameter to the atomic function (i.e. the value to add + * in shared_atomic_add, etc). + * 2: For CompSwap only: the second data parameter. + */ +INTRINSIC(shared_atomic_add, 2, ARR(1, 1), true, 1, 0, 0, 0) +INTRINSIC(shared_atomic_imin, 2, ARR(1, 1), true, 1, 0, 0, 0) +INTRINSIC(shared_atomic_umin, 2, ARR(1, 1), true, 1, 0, 0, 0) +INTRINSIC(shared_atomic_imax, 2, ARR(1, 1), true, 1, 0, 0, 0) +INTRINSIC(shared_atomic_umax, 2, ARR(1, 1), true, 1, 0, 0, 0) +INTRINSIC(shared_atomic_and, 2, ARR(1, 1), true, 1, 0, 0, 0) +INTRINSIC(shared_atomic_or, 2, ARR(1, 1), true, 1, 0, 0, 0) +INTRINSIC(shared_atomic_xor, 2, ARR(1, 1), true, 1, 0, 0, 0) +INTRINSIC(shared_atomic_exchange, 2, ARR(1, 1), true, 1, 0, 0, 0) +INTRINSIC(shared_atomic_comp_swap, 3, ARR(1, 1, 1), true, 1, 0, 0, 0) + +#define SYSTEM_VALUE(name, components, num_indices) \ + INTRINSIC(load_##name, 0, ARR(), true, components, 0, num_indices, \ + NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) + +SYSTEM_VALUE(front_face, 1, 0) +SYSTEM_VALUE(vertex_id, 1, 0) +SYSTEM_VALUE(vertex_id_zero_base, 1, 0) +SYSTEM_VALUE(base_vertex, 1, 0) +SYSTEM_VALUE(instance_id, 1, 0) +SYSTEM_VALUE(base_instance, 1, 0) +SYSTEM_VALUE(draw_id, 1, 0) +SYSTEM_VALUE(sample_id, 1, 0) +SYSTEM_VALUE(sample_pos, 2, 0) +SYSTEM_VALUE(sample_mask_in, 1, 0) +SYSTEM_VALUE(primitive_id, 1, 0) +SYSTEM_VALUE(invocation_id, 1, 0) +SYSTEM_VALUE(tess_coord, 3, 0) +SYSTEM_VALUE(tess_level_outer, 4, 0) +SYSTEM_VALUE(tess_level_inner, 2, 0) +SYSTEM_VALUE(patch_vertices_in, 1, 0) +SYSTEM_VALUE(local_invocation_id, 3, 0) +SYSTEM_VALUE(work_group_id, 3, 0) +SYSTEM_VALUE(user_clip_plane, 4, 1) /* const_index[0] is user_clip_plane[idx] */ +SYSTEM_VALUE(num_work_groups, 3, 0) +SYSTEM_VALUE(helper_invocation, 1, 0) + +/* + * Load operations pull data from some piece of GPU memory. All load + * operations operate in terms of offsets into some piece of theoretical + * memory. Loads from externally visible memory (UBO and SSBO) simply take a + * byte offset as a source. Loads from opaque memory (uniforms, inputs, etc.) + * take a base+offset pair where the base (const_index[0]) gives the location + * of the start of the variable being loaded and and the offset source is a + * offset into that variable. + * + * Some load operations such as UBO/SSBO load and per_vertex loads take an + * additional source to specify which UBO/SSBO/vertex to load from. + * + * The exact address type depends on the lowering pass that generates the + * load/store intrinsics. Typically, this is vec4 units for things such as + * varying slots and float units for fragment shader inputs. UBO and SSBO + * offsets are always in bytes. + */ + +#define LOAD(name, srcs, indices, flags) \ + INTRINSIC(load_##name, srcs, ARR(1, 1, 1, 1), true, 0, 0, indices, flags) + +/* src[] = { offset }. const_index[] = { base } */ +LOAD(uniform, 1, 1, NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) +/* src[] = { buffer_index, offset }. No const_index */ +LOAD(ubo, 2, 0, NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) +/* src[] = { offset }. const_index[] = { base } */ +LOAD(input, 1, 1, NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) +/* src[] = { vertex, offset }. const_index[] = { base } */ +LOAD(per_vertex_input, 2, 1, NIR_INTRINSIC_CAN_ELIMINATE | NIR_INTRINSIC_CAN_REORDER) +/* src[] = { buffer_index, offset }. No const_index */ +LOAD(ssbo, 2, 0, NIR_INTRINSIC_CAN_ELIMINATE) +/* src[] = { offset }. const_index[] = { base } */ +LOAD(output, 1, 1, NIR_INTRINSIC_CAN_ELIMINATE) +/* src[] = { vertex, offset }. const_index[] = { base } */ +LOAD(per_vertex_output, 2, 1, NIR_INTRINSIC_CAN_ELIMINATE) +/* src[] = { offset }. const_index[] = { base } */ +LOAD(shared, 1, 1, NIR_INTRINSIC_CAN_ELIMINATE) + +/* + * Stores work the same way as loads, except now the first source is the value + * to store and the second (and possibly third) source specify where to store + * the value. SSBO and shared memory stores also have a write mask as + * const_index[0]. + */ + +#define STORE(name, srcs, indices, flags) \ + INTRINSIC(store_##name, srcs, ARR(0, 1, 1, 1), false, 0, 0, indices, flags) + +/* src[] = { value, offset }. const_index[] = { base, write_mask } */ +STORE(output, 2, 2, 0) +/* src[] = { value, vertex, offset }. const_index[] = { base, write_mask } */ +STORE(per_vertex_output, 3, 2, 0) +/* src[] = { value, block_index, offset }. const_index[] = { write_mask } */ +STORE(ssbo, 3, 1, 0) +/* src[] = { value, offset }. const_index[] = { base, write_mask } */ +STORE(shared, 2, 2, 0) + +LAST_INTRINSIC(store_shared) diff --git a/src/compiler/nir/nir_liveness.c b/src/compiler/nir/nir_liveness.c new file mode 100644 index 0000000..05f79d7 --- /dev/null +++ b/src/compiler/nir/nir_liveness.c @@ -0,0 +1,297 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand (jason@jlekstrand.net) + */ + +#include "nir.h" +#include "nir_worklist.h" +#include "nir_vla.h" + +/* + * Basic liveness analysis. This works only in SSA form. + * + * This liveness pass treats phi nodes as being melded to the space between + * blocks so that the destinations of a phi are in the livein of the block + * in which it resides and the sources are in the liveout of the + * corresponding block. By formulating the liveness information in this + * way, we ensure that the definition of any variable dominates its entire + * live range. This is true because the only way that the definition of an + * SSA value may not dominate a use is if the use is in a phi node and the + * uses in phi no are in the live-out of the corresponding predecessor + * block but not in the live-in of the block containing the phi node. + */ + +struct live_ssa_defs_state { + unsigned num_ssa_defs; + unsigned bitset_words; + + nir_block_worklist worklist; +}; + +static bool +index_ssa_def(nir_ssa_def *def, void *void_state) +{ + struct live_ssa_defs_state *state = void_state; + + if (def->parent_instr->type == nir_instr_type_ssa_undef) + def->live_index = 0; + else + def->live_index = state->num_ssa_defs++; + + return true; +} + +static bool +index_ssa_definitions_block(nir_block *block, void *state) +{ + nir_foreach_instr(block, instr) + nir_foreach_ssa_def(instr, index_ssa_def, state); + + return true; +} + +/* Initialize the liveness data to zero and add the given block to the + * worklist. + */ +static bool +init_liveness_block(nir_block *block, void *void_state) +{ + struct live_ssa_defs_state *state = void_state; + + block->live_in = reralloc(block, block->live_in, BITSET_WORD, + state->bitset_words); + memset(block->live_in, 0, state->bitset_words * sizeof(BITSET_WORD)); + + block->live_out = reralloc(block, block->live_out, BITSET_WORD, + state->bitset_words); + memset(block->live_out, 0, state->bitset_words * sizeof(BITSET_WORD)); + + nir_block_worklist_push_head(&state->worklist, block); + + return true; +} + +static bool +set_src_live(nir_src *src, void *void_live) +{ + BITSET_WORD *live = void_live; + + if (!src->is_ssa) + return true; + + if (src->ssa->live_index == 0) + return true; /* undefined variables are never live */ + + BITSET_SET(live, src->ssa->live_index); + + return true; +} + +static bool +set_ssa_def_dead(nir_ssa_def *def, void *void_live) +{ + BITSET_WORD *live = void_live; + + BITSET_CLEAR(live, def->live_index); + + return true; +} + +/** Propagates the live in of succ across the edge to the live out of pred + * + * Phi nodes exist "between" blocks and all the phi nodes at the start of a + * block act "in parallel". When we propagate from the live_in of one + * block to the live out of the other, we have to kill any writes from phis + * and make live any sources. + * + * Returns true if updating live out of pred added anything + */ +static bool +propagate_across_edge(nir_block *pred, nir_block *succ, + struct live_ssa_defs_state *state) +{ + NIR_VLA(BITSET_WORD, live, state->bitset_words); + memcpy(live, succ->live_in, state->bitset_words * sizeof *live); + + nir_foreach_instr(succ, instr) { + if (instr->type != nir_instr_type_phi) + break; + nir_phi_instr *phi = nir_instr_as_phi(instr); + + assert(phi->dest.is_ssa); + set_ssa_def_dead(&phi->dest.ssa, live); + } + + nir_foreach_instr(succ, instr) { + if (instr->type != nir_instr_type_phi) + break; + nir_phi_instr *phi = nir_instr_as_phi(instr); + + nir_foreach_phi_src(phi, src) { + if (src->pred == pred) { + set_src_live(&src->src, live); + break; + } + } + } + + BITSET_WORD progress = 0; + for (unsigned i = 0; i < state->bitset_words; ++i) { + progress |= live[i] & ~pred->live_out[i]; + pred->live_out[i] |= live[i]; + } + return progress != 0; +} + +void +nir_live_ssa_defs_impl(nir_function_impl *impl) +{ + struct live_ssa_defs_state state; + + /* We start at 1 because we reserve the index value of 0 for ssa_undef + * instructions. Those are never live, so their liveness information + * can be compacted into a single bit. + */ + state.num_ssa_defs = 1; + nir_foreach_block(impl, index_ssa_definitions_block, &state); + + nir_block_worklist_init(&state.worklist, impl->num_blocks, NULL); + + /* We now know how many unique ssa definitions we have and we can go + * ahead and allocate live_in and live_out sets and add all of the + * blocks to the worklist. + */ + state.bitset_words = BITSET_WORDS(state.num_ssa_defs); + nir_foreach_block(impl, init_liveness_block, &state); + + /* We're now ready to work through the worklist and update the liveness + * sets of each of the blocks. By the time we get to this point, every + * block in the function implementation has been pushed onto the + * worklist in reverse order. As long as we keep the worklist + * up-to-date as we go, everything will get covered. + */ + while (!nir_block_worklist_is_empty(&state.worklist)) { + /* We pop them off in the reverse order we pushed them on. This way + * the first walk of the instructions is backwards so we only walk + * once in the case of no control flow. + */ + nir_block *block = nir_block_worklist_pop_head(&state.worklist); + + memcpy(block->live_in, block->live_out, + state.bitset_words * sizeof(BITSET_WORD)); + + nir_if *following_if = nir_block_get_following_if(block); + if (following_if) + set_src_live(&following_if->condition, block->live_in); + + nir_foreach_instr_reverse(block, instr) { + /* Phi nodes are handled seperately so we want to skip them. Since + * we are going backwards and they are at the beginning, we can just + * break as soon as we see one. + */ + if (instr->type == nir_instr_type_phi) + break; + + nir_foreach_ssa_def(instr, set_ssa_def_dead, block->live_in); + nir_foreach_src(instr, set_src_live, block->live_in); + } + + /* Walk over all of the predecessors of the current block updating + * their live in with the live out of this one. If anything has + * changed, add the predecessor to the work list so that we ensure + * that the new information is used. + */ + struct set_entry *entry; + set_foreach(block->predecessors, entry) { + nir_block *pred = (nir_block *)entry->key; + if (propagate_across_edge(pred, block, &state)) + nir_block_worklist_push_tail(&state.worklist, pred); + } + } + + nir_block_worklist_fini(&state.worklist); +} + +static bool +src_does_not_use_def(nir_src *src, void *def) +{ + return !src->is_ssa || src->ssa != (nir_ssa_def *)def; +} + +static bool +search_for_use_after_instr(nir_instr *start, nir_ssa_def *def) +{ + /* Only look for a use strictly after the given instruction */ + struct exec_node *node = start->node.next; + while (!exec_node_is_tail_sentinel(node)) { + nir_instr *instr = exec_node_data(nir_instr, node, node); + if (!nir_foreach_src(instr, src_does_not_use_def, def)) + return true; + node = node->next; + } + return false; +} + +/* Returns true if def is live at instr assuming that def comes before + * instr in a pre DFS search of the dominance tree. + */ +static bool +nir_ssa_def_is_live_at(nir_ssa_def *def, nir_instr *instr) +{ + if (BITSET_TEST(instr->block->live_out, def->live_index)) { + /* Since def dominates instr, if def is in the liveout of the block, + * it's live at instr + */ + return true; + } else { + if (BITSET_TEST(instr->block->live_in, def->live_index) || + def->parent_instr->block == instr->block) { + /* In this case it is either live coming into instr's block or it + * is defined in the same block. In this case, we simply need to + * see if it is used after instr. + */ + return search_for_use_after_instr(instr, def); + } else { + return false; + } + } +} + +bool +nir_ssa_defs_interfere(nir_ssa_def *a, nir_ssa_def *b) +{ + if (a->parent_instr == b->parent_instr) { + /* Two variables defined at the same time interfere assuming at + * least one isn't dead. + */ + return true; + } else if (a->live_index == 0 || b->live_index == 0) { + /* If either variable is an ssa_undef, then there's no interference */ + return false; + } else if (a->live_index < b->live_index) { + return nir_ssa_def_is_live_at(a, b->parent_instr); + } else { + return nir_ssa_def_is_live_at(b, a->parent_instr); + } +} diff --git a/src/compiler/nir/nir_lower_alu_to_scalar.c b/src/compiler/nir/nir_lower_alu_to_scalar.c new file mode 100644 index 0000000..0a27e66 --- /dev/null +++ b/src/compiler/nir/nir_lower_alu_to_scalar.c @@ -0,0 +1,210 @@ +/* + * Copyright © 2014-2015 Broadcom + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + */ + +#include "nir.h" +#include "nir_builder.h" + +/** @file nir_lower_alu_to_scalar.c + * + * Replaces nir_alu_instr operations with more than one channel used in the + * arguments with individual per-channel operations. + */ + +static void +nir_alu_ssa_dest_init(nir_alu_instr *instr, unsigned num_components) +{ + nir_ssa_dest_init(&instr->instr, &instr->dest.dest, num_components, NULL); + instr->dest.write_mask = (1 << num_components) - 1; +} + +static void +lower_reduction(nir_alu_instr *instr, nir_op chan_op, nir_op merge_op, + nir_builder *builder) +{ + unsigned num_components = nir_op_infos[instr->op].input_sizes[0]; + + nir_ssa_def *last = NULL; + for (unsigned i = 0; i < num_components; i++) { + nir_alu_instr *chan = nir_alu_instr_create(builder->shader, chan_op); + nir_alu_ssa_dest_init(chan, 1); + nir_alu_src_copy(&chan->src[0], &instr->src[0], chan); + chan->src[0].swizzle[0] = chan->src[0].swizzle[i]; + if (nir_op_infos[chan_op].num_inputs > 1) { + assert(nir_op_infos[chan_op].num_inputs == 2); + nir_alu_src_copy(&chan->src[1], &instr->src[1], chan); + chan->src[1].swizzle[0] = chan->src[1].swizzle[i]; + } + + nir_builder_instr_insert(builder, &chan->instr); + + if (i == 0) { + last = &chan->dest.dest.ssa; + } else { + last = nir_build_alu(builder, merge_op, + last, &chan->dest.dest.ssa, NULL, NULL); + } + } + + assert(instr->dest.write_mask == 1); + nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(last)); + nir_instr_remove(&instr->instr); +} + +static void +lower_alu_instr_scalar(nir_alu_instr *instr, nir_builder *b) +{ + unsigned num_src = nir_op_infos[instr->op].num_inputs; + unsigned i, chan; + + assert(instr->dest.dest.is_ssa); + assert(instr->dest.write_mask != 0); + + b->cursor = nir_before_instr(&instr->instr); + +#define LOWER_REDUCTION(name, chan, merge) \ + case name##2: \ + case name##3: \ + case name##4: \ + lower_reduction(instr, chan, merge, b); \ + return; + + switch (instr->op) { + case nir_op_vec4: + case nir_op_vec3: + case nir_op_vec2: + /* We don't need to scalarize these ops, they're the ones generated to + * group up outputs into a value that can be SSAed. + */ + return; + + case nir_op_unpack_unorm_4x8: + case nir_op_unpack_snorm_4x8: + case nir_op_unpack_unorm_2x16: + case nir_op_unpack_snorm_2x16: + /* There is no scalar version of these ops, unless we were to break it + * down to bitshifts and math (which is definitely not intended). + */ + return; + + case nir_op_unpack_half_2x16: + /* We could split this into unpack_half_2x16_split_[xy], but should + * we? + */ + return; + + case nir_op_fdph: { + nir_ssa_def *sum[4]; + for (unsigned i = 0; i < 3; i++) { + sum[i] = nir_fmul(b, nir_channel(b, instr->src[0].src.ssa, + instr->src[0].swizzle[i]), + nir_channel(b, instr->src[1].src.ssa, + instr->src[1].swizzle[i])); + } + sum[3] = nir_channel(b, instr->src[1].src.ssa, instr->src[1].swizzle[3]); + + nir_ssa_def *val = nir_fadd(b, nir_fadd(b, sum[0], sum[1]), + nir_fadd(b, sum[2], sum[3])); + + nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(val)); + nir_instr_remove(&instr->instr); + return; + } + + LOWER_REDUCTION(nir_op_fdot, nir_op_fmul, nir_op_fadd); + LOWER_REDUCTION(nir_op_ball_fequal, nir_op_feq, nir_op_iand); + LOWER_REDUCTION(nir_op_ball_iequal, nir_op_ieq, nir_op_iand); + LOWER_REDUCTION(nir_op_bany_fnequal, nir_op_fne, nir_op_ior); + LOWER_REDUCTION(nir_op_bany_inequal, nir_op_ine, nir_op_ior); + LOWER_REDUCTION(nir_op_fall_equal, nir_op_seq, nir_op_fand); + LOWER_REDUCTION(nir_op_fany_nequal, nir_op_sne, nir_op_for); + + default: + break; + } + + if (instr->dest.dest.ssa.num_components == 1) + return; + + unsigned num_components = instr->dest.dest.ssa.num_components; + nir_ssa_def *comps[] = { NULL, NULL, NULL, NULL }; + + for (chan = 0; chan < 4; chan++) { + if (!(instr->dest.write_mask & (1 << chan))) + continue; + + nir_alu_instr *lower = nir_alu_instr_create(b->shader, instr->op); + for (i = 0; i < num_src; i++) { + /* We only handle same-size-as-dest (input_sizes[] == 0) or scalar + * args (input_sizes[] == 1). + */ + assert(nir_op_infos[instr->op].input_sizes[i] < 2); + unsigned src_chan = (nir_op_infos[instr->op].input_sizes[i] == 1 ? + 0 : chan); + + nir_alu_src_copy(&lower->src[i], &instr->src[i], lower); + for (int j = 0; j < 4; j++) + lower->src[i].swizzle[j] = instr->src[i].swizzle[src_chan]; + } + + nir_alu_ssa_dest_init(lower, 1); + lower->dest.saturate = instr->dest.saturate; + comps[chan] = &lower->dest.dest.ssa; + + nir_builder_instr_insert(b, &lower->instr); + } + + nir_ssa_def *vec = nir_vec(b, comps, num_components); + + nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, nir_src_for_ssa(vec)); + + nir_instr_remove(&instr->instr); +} + +static bool +lower_alu_to_scalar_block(nir_block *block, void *builder) +{ + nir_foreach_instr_safe(block, instr) { + if (instr->type == nir_instr_type_alu) + lower_alu_instr_scalar(nir_instr_as_alu(instr), builder); + } + + return true; +} + +static void +nir_lower_alu_to_scalar_impl(nir_function_impl *impl) +{ + nir_builder builder; + nir_builder_init(&builder, impl); + + nir_foreach_block(impl, lower_alu_to_scalar_block, &builder); +} + +void +nir_lower_alu_to_scalar(nir_shader *shader) +{ + nir_foreach_function(shader, function) { + if (function->impl) + nir_lower_alu_to_scalar_impl(function->impl); + } +} diff --git a/src/compiler/nir/nir_lower_atomics.c b/src/compiler/nir/nir_lower_atomics.c new file mode 100644 index 0000000..2cbc1b7 --- /dev/null +++ b/src/compiler/nir/nir_lower_atomics.c @@ -0,0 +1,166 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +#include "glsl/ir_uniform.h" +#include "nir.h" +#include "main/config.h" +#include <assert.h> + +typedef struct { + const struct gl_shader_program *shader_program; + nir_shader *shader; +} lower_atomic_state; + +/* + * replace atomic counter intrinsics that use a variable with intrinsics + * that directly store the buffer index and byte offset + */ + +static void +lower_instr(nir_intrinsic_instr *instr, + lower_atomic_state *state) +{ + nir_intrinsic_op op; + switch (instr->intrinsic) { + case nir_intrinsic_atomic_counter_read_var: + op = nir_intrinsic_atomic_counter_read; + break; + + case nir_intrinsic_atomic_counter_inc_var: + op = nir_intrinsic_atomic_counter_inc; + break; + + case nir_intrinsic_atomic_counter_dec_var: + op = nir_intrinsic_atomic_counter_dec; + break; + + default: + return; + } + + if (instr->variables[0]->var->data.mode != nir_var_uniform && + instr->variables[0]->var->data.mode != nir_var_shader_storage) + return; /* atomics passed as function arguments can't be lowered */ + + void *mem_ctx = ralloc_parent(instr); + unsigned uniform_loc = instr->variables[0]->var->data.location; + + nir_intrinsic_instr *new_instr = nir_intrinsic_instr_create(mem_ctx, op); + new_instr->const_index[0] = + state->shader_program->UniformStorage[uniform_loc].opaque[state->shader->stage].index; + + nir_load_const_instr *offset_const = nir_load_const_instr_create(mem_ctx, 1); + offset_const->value.u[0] = instr->variables[0]->var->data.offset; + + nir_instr_insert_before(&instr->instr, &offset_const->instr); + + nir_ssa_def *offset_def = &offset_const->def; + + nir_deref *tail = &instr->variables[0]->deref; + while (tail->child != NULL) { + assert(tail->child->deref_type == nir_deref_type_array); + nir_deref_array *deref_array = nir_deref_as_array(tail->child); + tail = tail->child; + + unsigned child_array_elements = tail->child != NULL ? + glsl_get_aoa_size(tail->type) : 1; + + offset_const->value.u[0] += deref_array->base_offset * + child_array_elements * ATOMIC_COUNTER_SIZE; + + if (deref_array->deref_array_type == nir_deref_array_type_indirect) { + nir_load_const_instr *atomic_counter_size = + nir_load_const_instr_create(mem_ctx, 1); + atomic_counter_size->value.u[0] = child_array_elements * ATOMIC_COUNTER_SIZE; + nir_instr_insert_before(&instr->instr, &atomic_counter_size->instr); + + nir_alu_instr *mul = nir_alu_instr_create(mem_ctx, nir_op_imul); + nir_ssa_dest_init(&mul->instr, &mul->dest.dest, 1, NULL); + mul->dest.write_mask = 0x1; + nir_src_copy(&mul->src[0].src, &deref_array->indirect, mul); + mul->src[1].src.is_ssa = true; + mul->src[1].src.ssa = &atomic_counter_size->def; + nir_instr_insert_before(&instr->instr, &mul->instr); + + nir_alu_instr *add = nir_alu_instr_create(mem_ctx, nir_op_iadd); + nir_ssa_dest_init(&add->instr, &add->dest.dest, 1, NULL); + add->dest.write_mask = 0x1; + add->src[0].src.is_ssa = true; + add->src[0].src.ssa = &mul->dest.dest.ssa; + add->src[1].src.is_ssa = true; + add->src[1].src.ssa = offset_def; + nir_instr_insert_before(&instr->instr, &add->instr); + + offset_def = &add->dest.dest.ssa; + } + } + + new_instr->src[0].is_ssa = true; + new_instr->src[0].ssa = offset_def; + + if (instr->dest.is_ssa) { + nir_ssa_dest_init(&new_instr->instr, &new_instr->dest, + instr->dest.ssa.num_components, NULL); + nir_ssa_def_rewrite_uses(&instr->dest.ssa, + nir_src_for_ssa(&new_instr->dest.ssa)); + } else { + nir_dest_copy(&new_instr->dest, &instr->dest, mem_ctx); + } + + nir_instr_insert_before(&instr->instr, &new_instr->instr); + nir_instr_remove(&instr->instr); +} + +static bool +lower_block(nir_block *block, void *state) +{ + nir_foreach_instr_safe(block, instr) { + if (instr->type == nir_instr_type_intrinsic) + lower_instr(nir_instr_as_intrinsic(instr), + (lower_atomic_state *) state); + } + + return true; +} + +void +nir_lower_atomics(nir_shader *shader, + const struct gl_shader_program *shader_program) +{ + lower_atomic_state state = { + .shader = shader, + .shader_program = shader_program, + }; + + nir_foreach_function(shader, function) { + if (function->impl) { + nir_foreach_block(function->impl, lower_block, (void *) &state); + nir_metadata_preserve(function->impl, nir_metadata_block_index | + nir_metadata_dominance); + } + } +} diff --git a/src/compiler/nir/nir_lower_clip.c b/src/compiler/nir/nir_lower_clip.c new file mode 100644 index 0000000..0ca6a28 --- /dev/null +++ b/src/compiler/nir/nir_lower_clip.c @@ -0,0 +1,339 @@ +/* + * Copyright © 2015 Red Hat + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Rob Clark <robclark@freedesktop.org> + */ + +#include "nir.h" +#include "nir_builder.h" + +#define MAX_CLIP_PLANES 8 + +/* Generates the lowering code for user-clip-planes, generating CLIPDIST + * from UCP[n] + CLIPVERTEX or POSITION. Additionally, an optional pass + * for fragment shaders to insert conditional kill's based on the inter- + * polated CLIPDIST + * + * NOTE: should be run after nir_lower_outputs_to_temporaries() (or at + * least in scenarios where you can count on each output written once + * and only once). + */ + + +static nir_variable * +create_clipdist_var(nir_shader *shader, unsigned drvloc, + bool output, gl_varying_slot slot) +{ + nir_variable *var = rzalloc(shader, nir_variable); + + var->data.driver_location = drvloc; + var->type = glsl_vec4_type(); + var->data.mode = output ? nir_var_shader_out : nir_var_shader_in; + var->name = ralloc_asprintf(var, "clipdist_%d", drvloc); + var->data.index = 0; + var->data.location = slot; + + if (output) { + exec_list_push_tail(&shader->outputs, &var->node); + shader->num_outputs++; /* TODO use type_size() */ + } + else { + exec_list_push_tail(&shader->inputs, &var->node); + shader->num_inputs++; /* TODO use type_size() */ + } + return var; +} + +static void +store_clipdist_output(nir_builder *b, nir_variable *out, nir_ssa_def **val) +{ + nir_intrinsic_instr *store; + + store = nir_intrinsic_instr_create(b->shader, nir_intrinsic_store_output); + store->num_components = 4; + store->const_index[0] = out->data.driver_location; + store->const_index[1] = 0xf; /* wrmask */ + store->src[0].ssa = nir_vec4(b, val[0], val[1], val[2], val[3]); + store->src[0].is_ssa = true; + store->src[1] = nir_src_for_ssa(nir_imm_int(b, 0)); + nir_builder_instr_insert(b, &store->instr); +} + +static void +load_clipdist_input(nir_builder *b, nir_variable *in, nir_ssa_def **val) +{ + nir_intrinsic_instr *load; + + load = nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_input); + load->num_components = 4; + load->const_index[0] = in->data.driver_location; + load->src[0] = nir_src_for_ssa(nir_imm_int(b, 0)); + nir_ssa_dest_init(&load->instr, &load->dest, 4, NULL); + nir_builder_instr_insert(b, &load->instr); + + val[0] = nir_channel(b, &load->dest.ssa, 0); + val[1] = nir_channel(b, &load->dest.ssa, 1); + val[2] = nir_channel(b, &load->dest.ssa, 2); + val[3] = nir_channel(b, &load->dest.ssa, 3); +} + +struct find_output_state +{ + unsigned drvloc; + nir_ssa_def *def; +}; + +static bool +find_output_in_block(nir_block *block, void *void_state) +{ + struct find_output_state *state = void_state; + nir_foreach_instr(block, instr) { + + if (instr->type == nir_instr_type_intrinsic) { + nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr); + if ((intr->intrinsic == nir_intrinsic_store_output) && + intr->const_index[0] == state->drvloc) { + assert(state->def == NULL); + assert(intr->src[0].is_ssa); + assert(nir_src_as_const_value(intr->src[1])); + state->def = intr->src[0].ssa; + +#if !defined(DEBUG) + /* for debug builds, scan entire shader to assert + * if output is written multiple times. For release + * builds just assume all is well and bail when we + * find first: + */ + return false; +#endif + } + } + } + + return true; +} + +/* TODO: maybe this would be a useful helper? + * NOTE: assumes each output is written exactly once (and unconditionally) + * so if needed nir_lower_outputs_to_temporaries() + */ +static nir_ssa_def * +find_output(nir_shader *shader, unsigned drvloc) +{ + struct find_output_state state = { + .drvloc = drvloc, + }; + + nir_foreach_function(shader, function) { + if (function->impl) { + nir_foreach_block_reverse(function->impl, + find_output_in_block, &state); + } + } + + return state.def; +} + +/* + * VS lowering + */ + +static void +lower_clip_vs(nir_function_impl *impl, unsigned ucp_enables, + nir_ssa_def *cv, nir_variable **out) +{ + nir_ssa_def *clipdist[MAX_CLIP_PLANES]; + nir_builder b; + + nir_builder_init(&b, impl); + + /* NIR should ensure that, even in case of loops/if-else, there + * should be only a single predecessor block to end_block, which + * makes the perfect place to insert the clipdist calculations. + * + * NOTE: in case of early return's, these would have to be lowered + * to jumps to end_block predecessor in a previous pass. Not sure + * if there is a good way to sanity check this, but for now the + * users of this pass don't support sub-routines. + */ + assert(impl->end_block->predecessors->entries == 1); + b.cursor = nir_after_cf_list(&impl->body); + + for (int plane = 0; plane < MAX_CLIP_PLANES; plane++) { + if (ucp_enables & (1 << plane)) { + nir_ssa_def *ucp = + nir_load_system_value(&b, nir_intrinsic_load_user_clip_plane, plane); + + /* calculate clipdist[plane] - dot(ucp, cv): */ + clipdist[plane] = nir_fdot4(&b, ucp, cv); + } + else { + /* 0.0 == don't-clip == disabled: */ + clipdist[plane] = nir_imm_float(&b, 0.0); + } + } + + if (ucp_enables & 0x0f) + store_clipdist_output(&b, out[0], &clipdist[0]); + if (ucp_enables & 0xf0) + store_clipdist_output(&b, out[1], &clipdist[4]); + + nir_metadata_preserve(impl, nir_metadata_dominance); +} + +/* ucp_enables is bitmask of enabled ucp's. Actual ucp values are + * passed in to shader via user_clip_plane system-values + */ +void +nir_lower_clip_vs(nir_shader *shader, unsigned ucp_enables) +{ + int clipvertex = -1; + int position = -1; + int maxloc = -1; + nir_ssa_def *cv; + nir_variable *out[2] = { NULL }; + + if (!ucp_enables) + return; + + /* find clipvertex/position outputs: */ + nir_foreach_variable(var, &shader->outputs) { + int loc = var->data.driver_location; + + /* keep track of last used driver-location.. we'll be + * appending CLIP_DIST0/CLIP_DIST1 after last existing + * output: + */ + maxloc = MAX2(maxloc, loc); + + switch (var->data.location) { + case VARYING_SLOT_POS: + position = loc; + break; + case VARYING_SLOT_CLIP_VERTEX: + clipvertex = loc; + break; + case VARYING_SLOT_CLIP_DIST0: + case VARYING_SLOT_CLIP_DIST1: + /* if shader is already writing CLIPDIST, then + * there should be no user-clip-planes to deal + * with. + */ + return; + } + } + + if (clipvertex != -1) + cv = find_output(shader, clipvertex); + else if (position != -1) + cv = find_output(shader, position); + else + return; + + /* insert CLIPDIST outputs: */ + if (ucp_enables & 0x0f) + out[0] = + create_clipdist_var(shader, ++maxloc, true, VARYING_SLOT_CLIP_DIST0); + if (ucp_enables & 0xf0) + out[1] = + create_clipdist_var(shader, ++maxloc, true, VARYING_SLOT_CLIP_DIST1); + + nir_foreach_function(shader, function) { + if (!strcmp(function->name, "main")) + lower_clip_vs(function->impl, ucp_enables, cv, out); + } +} + +/* + * FS lowering + */ + +static void +lower_clip_fs(nir_function_impl *impl, unsigned ucp_enables, + nir_variable **in) +{ + nir_ssa_def *clipdist[MAX_CLIP_PLANES]; + nir_builder b; + + nir_builder_init(&b, impl); + b.cursor = nir_before_cf_list(&impl->body); + + if (ucp_enables & 0x0f) + load_clipdist_input(&b, in[0], &clipdist[0]); + if (ucp_enables & 0xf0) + load_clipdist_input(&b, in[1], &clipdist[4]); + + for (int plane = 0; plane < MAX_CLIP_PLANES; plane++) { + if (ucp_enables & (1 << plane)) { + nir_intrinsic_instr *discard; + nir_ssa_def *cond; + + cond = nir_flt(&b, clipdist[plane], nir_imm_float(&b, 0.0)); + + discard = nir_intrinsic_instr_create(b.shader, + nir_intrinsic_discard_if); + discard->src[0] = nir_src_for_ssa(cond); + nir_builder_instr_insert(&b, &discard->instr); + } + } +} + +/* insert conditional kill based on interpolated CLIPDIST + */ +void +nir_lower_clip_fs(nir_shader *shader, unsigned ucp_enables) +{ + nir_variable *in[2]; + int maxloc = -1; + + if (!ucp_enables) + return; + + nir_foreach_variable(var, &shader->inputs) { + int loc = var->data.driver_location; + + /* keep track of last used driver-location.. we'll be + * appending CLIP_DIST0/CLIP_DIST1 after last existing + * input: + */ + maxloc = MAX2(maxloc, loc); + } + + /* The shader won't normally have CLIPDIST inputs, so we + * must add our own: + */ + /* insert CLIPDIST outputs: */ + if (ucp_enables & 0x0f) + in[0] = + create_clipdist_var(shader, ++maxloc, false, + VARYING_SLOT_CLIP_DIST0); + if (ucp_enables & 0xf0) + in[1] = + create_clipdist_var(shader, ++maxloc, false, + VARYING_SLOT_CLIP_DIST1); + + nir_foreach_function(shader, function) { + if (!strcmp(function->name, "main")) + lower_clip_fs(function->impl, ucp_enables, in); + } +} diff --git a/src/compiler/nir/nir_lower_global_vars_to_local.c b/src/compiler/nir/nir_lower_global_vars_to_local.c new file mode 100644 index 0000000..7b4cd4e --- /dev/null +++ b/src/compiler/nir/nir_lower_global_vars_to_local.c @@ -0,0 +1,113 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand (jason@jlekstrand.net) + * + */ + +/* + * This lowering pass detects when a global variable is only being used by + * one function and makes it local to that function + */ + +#include "nir.h" + +struct global_to_local_state { + nir_function_impl *impl; + /* A hash table keyed on variable pointers that stores the unique + * nir_function_impl that uses the given variable. If a variable is + * used in multiple functions, the data for the given key will be NULL. + */ + struct hash_table *var_func_table; +}; + +static bool +mark_global_var_uses_block(nir_block *block, void *void_state) +{ + struct global_to_local_state *state = void_state; + + nir_foreach_instr(block, instr) { + if (instr->type != nir_instr_type_intrinsic) + continue; + + nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); + unsigned num_vars = nir_intrinsic_infos[intrin->intrinsic].num_variables; + + for (unsigned i = 0; i < num_vars; i++) { + nir_variable *var = intrin->variables[i]->var; + if (var->data.mode != nir_var_global) + continue; + + struct hash_entry *entry = + _mesa_hash_table_search(state->var_func_table, var); + + if (entry) { + if (entry->data != state->impl) + entry->data = NULL; + } else { + _mesa_hash_table_insert(state->var_func_table, var, state->impl); + } + } + } + + return true; +} + +bool +nir_lower_global_vars_to_local(nir_shader *shader) +{ + struct global_to_local_state state; + bool progress = false; + + state.var_func_table = _mesa_hash_table_create(NULL, _mesa_hash_pointer, + _mesa_key_pointer_equal); + + nir_foreach_function(shader, function) { + if (function->impl) { + state.impl = function->impl; + nir_foreach_block(function->impl, mark_global_var_uses_block, &state); + } + } + + struct hash_entry *entry; + hash_table_foreach(state.var_func_table, entry) { + nir_variable *var = (void *)entry->key; + nir_function_impl *impl = entry->data; + + assert(var->data.mode == nir_var_global); + + if (impl != NULL) { + exec_node_remove(&var->node); + var->data.mode = nir_var_local; + exec_list_push_tail(&impl->locals, &var->node); + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance | + nir_metadata_live_ssa_defs); + progress = true; + } + } + + _mesa_hash_table_destroy(state.var_func_table, NULL); + + return progress; +} diff --git a/src/compiler/nir/nir_lower_gs_intrinsics.c b/src/compiler/nir/nir_lower_gs_intrinsics.c new file mode 100644 index 0000000..fdff165 --- /dev/null +++ b/src/compiler/nir/nir_lower_gs_intrinsics.c @@ -0,0 +1,219 @@ +/* + * Copyright © 2015 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + */ + +#include "nir.h" +#include "nir_builder.h" + +/** + * \file nir_lower_gs_intrinsics.c + * + * Geometry Shaders can call EmitVertex()/EmitStreamVertex() to output an + * arbitrary number of vertices. However, the shader must declare the maximum + * number of vertices that it will ever output - further attempts to emit + * vertices result in undefined behavior according to the GLSL specification. + * + * Drivers might use this maximum number of vertices to allocate enough space + * to hold the geometry shader's output. Some drivers (such as i965) need to + * implement "safety checks" which ensure that the shader hasn't emitted too + * many vertices, to avoid overflowing that space and trashing other memory. + * + * The count of emitted vertices can also be useful in buffer offset + * calculations, so drivers know where to write the GS output. + * + * However, for simple geometry shaders that emit a statically determinable + * number of vertices, this extra bookkeeping is unnecessary and inefficient. + * By tracking the vertex count in NIR, we allow constant folding/propagation + * and dead control flow optimizations to eliminate most of it where possible. + * + * This pass introduces a new global variable which stores the current vertex + * count (initialized to 0), and converts emit_vertex/end_primitive intrinsics + * to their *_with_counter variants. emit_vertex is also wrapped in a safety + * check to avoid buffer overflows. Finally, it adds a set_vertex_count + * intrinsic at the end of the program, informing the driver of the final + * vertex count. + */ + +struct state { + nir_builder *builder; + nir_variable *vertex_count_var; + bool progress; +}; + +/** + * Replace emit_vertex intrinsics with: + * + * if (vertex_count < max_vertices) { + * emit_vertex_with_counter vertex_count ... + * vertex_count += 1 + * } + */ +static void +rewrite_emit_vertex(nir_intrinsic_instr *intrin, struct state *state) +{ + nir_builder *b = state->builder; + + /* Load the vertex count */ + b->cursor = nir_before_instr(&intrin->instr); + nir_ssa_def *count = nir_load_var(b, state->vertex_count_var); + + nir_ssa_def *max_vertices = nir_imm_int(b, b->shader->info.gs.vertices_out); + + /* Create: if (vertex_count < max_vertices) and insert it. + * + * The new if statement needs to be hooked up to the control flow graph + * before we start inserting instructions into it. + */ + nir_if *if_stmt = nir_if_create(b->shader); + if_stmt->condition = nir_src_for_ssa(nir_ilt(b, count, max_vertices)); + nir_builder_cf_insert(b, &if_stmt->cf_node); + + /* Fill out the new then-block */ + b->cursor = nir_after_cf_list(&if_stmt->then_list); + + nir_intrinsic_instr *lowered = + nir_intrinsic_instr_create(b->shader, + nir_intrinsic_emit_vertex_with_counter); + lowered->const_index[0] = intrin->const_index[0]; + lowered->src[0] = nir_src_for_ssa(count); + nir_builder_instr_insert(b, &lowered->instr); + + /* Increment the vertex count by 1 */ + nir_store_var(b, state->vertex_count_var, + nir_iadd(b, count, nir_imm_int(b, 1)), + 0x1); /* .x */ + + nir_instr_remove(&intrin->instr); + + state->progress = true; +} + +/** + * Replace end_primitive with end_primitive_with_counter. + */ +static void +rewrite_end_primitive(nir_intrinsic_instr *intrin, struct state *state) +{ + nir_builder *b = state->builder; + + b->cursor = nir_before_instr(&intrin->instr); + nir_ssa_def *count = nir_load_var(b, state->vertex_count_var); + + nir_intrinsic_instr *lowered = + nir_intrinsic_instr_create(b->shader, + nir_intrinsic_end_primitive_with_counter); + lowered->const_index[0] = intrin->const_index[0]; + lowered->src[0] = nir_src_for_ssa(count); + nir_builder_instr_insert(b, &lowered->instr); + + nir_instr_remove(&intrin->instr); + + state->progress = true; +} + +static bool +rewrite_intrinsics(nir_block *block, void *closure) +{ + struct state *state = closure; + + nir_foreach_instr_safe(block, instr) { + if (instr->type != nir_instr_type_intrinsic) + continue; + + nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); + switch (intrin->intrinsic) { + case nir_intrinsic_emit_vertex: + rewrite_emit_vertex(intrin, state); + break; + case nir_intrinsic_end_primitive: + rewrite_end_primitive(intrin, state); + break; + default: + /* not interesting; skip this */ + break; + } + } + + return true; +} + +/** + * Add a set_vertex_count intrinsic at the end of the program + * (representing the final vertex count). + */ +static void +append_set_vertex_count(nir_block *end_block, struct state *state) +{ + nir_builder *b = state->builder; + nir_shader *shader = state->builder->shader; + + /* Insert the new intrinsic in all of the predecessors of the end block, + * but before any jump instructions (return). + */ + struct set_entry *entry; + set_foreach(end_block->predecessors, entry) { + nir_block *pred = (nir_block *) entry->key; + b->cursor = nir_after_block_before_jump(pred); + + nir_ssa_def *count = nir_load_var(b, state->vertex_count_var); + + nir_intrinsic_instr *set_vertex_count = + nir_intrinsic_instr_create(shader, nir_intrinsic_set_vertex_count); + set_vertex_count->src[0] = nir_src_for_ssa(count); + + nir_builder_instr_insert(b, &set_vertex_count->instr); + } +} + +bool +nir_lower_gs_intrinsics(nir_shader *shader) +{ + struct state state; + state.progress = false; + + /* Create the counter variable */ + nir_variable *var = rzalloc(shader, nir_variable); + var->data.mode = nir_var_global; + var->type = glsl_uint_type(); + var->name = "vertex_count"; + var->constant_initializer = rzalloc(shader, nir_constant); /* initialize to 0 */ + + exec_list_push_tail(&shader->globals, &var->node); + state.vertex_count_var = var; + + nir_foreach_function(shader, function) { + if (function->impl) { + nir_builder b; + nir_builder_init(&b, function->impl); + state.builder = &b; + + nir_foreach_block(function->impl, rewrite_intrinsics, &state); + + /* This only works because we have a single main() function. */ + append_set_vertex_count(function->impl->end_block, &state); + + nir_metadata_preserve(function->impl, 0); + } + } + + return state.progress; +} diff --git a/src/compiler/nir/nir_lower_idiv.c b/src/compiler/nir/nir_lower_idiv.c new file mode 100644 index 0000000..a084ad9 --- /dev/null +++ b/src/compiler/nir/nir_lower_idiv.c @@ -0,0 +1,151 @@ +/* + * Copyright © 2015 Red Hat + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Rob Clark <robclark@freedesktop.org> + */ + +#include "nir.h" +#include "nir_builder.h" + +/* Lowers idiv/udiv/umod + * Based on NV50LegalizeSSA::handleDIV() + * + * Note that this is probably not enough precision for compute shaders. + * Perhaps we want a second higher precision (looping) version of this? + * Or perhaps we assume if you can do compute shaders you can also + * branch out to a pre-optimized shader library routine.. + */ + +static void +convert_instr(nir_builder *bld, nir_alu_instr *alu) +{ + nir_ssa_def *numer, *denom, *af, *bf, *a, *b, *q, *r; + nir_op op = alu->op; + bool is_signed; + + if ((op != nir_op_idiv) && + (op != nir_op_udiv) && + (op != nir_op_umod)) + return; + + is_signed = (op == nir_op_idiv); + + bld->cursor = nir_before_instr(&alu->instr); + + numer = nir_ssa_for_alu_src(bld, alu, 0); + denom = nir_ssa_for_alu_src(bld, alu, 1); + + if (is_signed) { + af = nir_i2f(bld, numer); + bf = nir_i2f(bld, denom); + af = nir_fabs(bld, af); + bf = nir_fabs(bld, bf); + a = nir_iabs(bld, numer); + b = nir_iabs(bld, denom); + } else { + af = nir_u2f(bld, numer); + bf = nir_u2f(bld, denom); + a = numer; + b = denom; + } + + /* get first result: */ + bf = nir_frcp(bld, bf); + bf = nir_isub(bld, bf, nir_imm_int(bld, 2)); /* yes, really */ + q = nir_fmul(bld, af, bf); + + if (is_signed) { + q = nir_f2i(bld, q); + } else { + q = nir_f2u(bld, q); + } + + /* get error of first result: */ + r = nir_imul(bld, q, b); + r = nir_isub(bld, a, r); + r = nir_u2f(bld, r); + r = nir_fmul(bld, r, bf); + r = nir_f2u(bld, r); + + /* add quotients: */ + q = nir_iadd(bld, q, r); + + /* correction: if modulus >= divisor, add 1 */ + r = nir_imul(bld, q, b); + r = nir_isub(bld, a, r); + + r = nir_uge(bld, r, b); + r = nir_b2i(bld, r); + + q = nir_iadd(bld, q, r); + if (is_signed) { + /* fix the sign: */ + r = nir_ixor(bld, numer, denom); + r = nir_ushr(bld, r, nir_imm_int(bld, 31)); + r = nir_i2b(bld, r); + b = nir_ineg(bld, q); + q = nir_bcsel(bld, r, b, q); + } + + if (op == nir_op_umod) { + /* division result in q */ + r = nir_imul(bld, q, b); + q = nir_isub(bld, a, r); + } + + assert(alu->dest.dest.is_ssa); + nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, nir_src_for_ssa(q)); +} + +static bool +convert_block(nir_block *block, void *state) +{ + nir_builder *b = state; + + nir_foreach_instr_safe(block, instr) { + if (instr->type == nir_instr_type_alu) + convert_instr(b, nir_instr_as_alu(instr)); + } + + return true; +} + +static void +convert_impl(nir_function_impl *impl) +{ + nir_builder b; + nir_builder_init(&b, impl); + + nir_foreach_block(impl, convert_block, &b); + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance); +} + +void +nir_lower_idiv(nir_shader *shader) +{ + nir_foreach_function(shader, function) { + if (function->impl) + convert_impl(function->impl); + } +} diff --git a/src/compiler/nir/nir_lower_io.c b/src/compiler/nir/nir_lower_io.c new file mode 100644 index 0000000..80c5151 --- /dev/null +++ b/src/compiler/nir/nir_lower_io.c @@ -0,0 +1,350 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * Jason Ekstrand (jason@jlekstrand.net) + * + */ + +/* + * This lowering pass converts references to input/output variables with + * loads/stores to actual input/output intrinsics. + */ + +#include "nir.h" +#include "nir_builder.h" + +struct lower_io_state { + nir_builder builder; + void *mem_ctx; + int (*type_size)(const struct glsl_type *type); + nir_variable_mode mode; +}; + +void +nir_assign_var_locations(struct exec_list *var_list, unsigned *size, + int (*type_size)(const struct glsl_type *)) +{ + unsigned location = 0; + + nir_foreach_variable(var, var_list) { + /* + * UBO's have their own address spaces, so don't count them towards the + * number of global uniforms + */ + if ((var->data.mode == nir_var_uniform || var->data.mode == nir_var_shader_storage) && + var->interface_type != NULL) + continue; + + var->data.driver_location = location; + location += type_size(var->type); + } + + *size = location; +} + +/** + * Returns true if we're processing a stage whose inputs are arrays indexed + * by a vertex number (such as geometry shader inputs). + */ +static bool +is_per_vertex_input(struct lower_io_state *state, nir_variable *var) +{ + gl_shader_stage stage = state->builder.shader->stage; + + return var->data.mode == nir_var_shader_in && !var->data.patch && + (stage == MESA_SHADER_TESS_CTRL || + stage == MESA_SHADER_TESS_EVAL || + stage == MESA_SHADER_GEOMETRY); +} + +static bool +is_per_vertex_output(struct lower_io_state *state, nir_variable *var) +{ + gl_shader_stage stage = state->builder.shader->stage; + return var->data.mode == nir_var_shader_out && !var->data.patch && + stage == MESA_SHADER_TESS_CTRL; +} + +static nir_ssa_def * +get_io_offset(nir_builder *b, nir_deref_var *deref, + nir_ssa_def **vertex_index, + int (*type_size)(const struct glsl_type *)) +{ + nir_deref *tail = &deref->deref; + + /* For per-vertex input arrays (i.e. geometry shader inputs), keep the + * outermost array index separate. Process the rest normally. + */ + if (vertex_index != NULL) { + tail = tail->child; + assert(tail->deref_type == nir_deref_type_array); + nir_deref_array *deref_array = nir_deref_as_array(tail); + + nir_ssa_def *vtx = nir_imm_int(b, deref_array->base_offset); + if (deref_array->deref_array_type == nir_deref_array_type_indirect) { + vtx = nir_iadd(b, vtx, nir_ssa_for_src(b, deref_array->indirect, 1)); + } + *vertex_index = vtx; + } + + /* Just emit code and let constant-folding go to town */ + nir_ssa_def *offset = nir_imm_int(b, 0); + + while (tail->child != NULL) { + const struct glsl_type *parent_type = tail->type; + tail = tail->child; + + if (tail->deref_type == nir_deref_type_array) { + nir_deref_array *deref_array = nir_deref_as_array(tail); + unsigned size = type_size(tail->type); + + offset = nir_iadd(b, offset, + nir_imm_int(b, size * deref_array->base_offset)); + + if (deref_array->deref_array_type == nir_deref_array_type_indirect) { + nir_ssa_def *mul = + nir_imul(b, nir_imm_int(b, size), + nir_ssa_for_src(b, deref_array->indirect, 1)); + + offset = nir_iadd(b, offset, mul); + } + } else if (tail->deref_type == nir_deref_type_struct) { + nir_deref_struct *deref_struct = nir_deref_as_struct(tail); + + unsigned field_offset = 0; + for (unsigned i = 0; i < deref_struct->index; i++) { + field_offset += type_size(glsl_get_struct_field(parent_type, i)); + } + offset = nir_iadd(b, offset, nir_imm_int(b, field_offset)); + } + } + + return offset; +} + +static nir_intrinsic_op +load_op(struct lower_io_state *state, + nir_variable_mode mode, bool per_vertex) +{ + nir_intrinsic_op op; + switch (mode) { + case nir_var_shader_in: + op = per_vertex ? nir_intrinsic_load_per_vertex_input : + nir_intrinsic_load_input; + break; + case nir_var_shader_out: + op = per_vertex ? nir_intrinsic_load_per_vertex_output : + nir_intrinsic_load_output; + break; + case nir_var_uniform: + op = nir_intrinsic_load_uniform; + break; + default: + unreachable("Unknown variable mode"); + } + return op; +} + +static bool +nir_lower_io_block(nir_block *block, void *void_state) +{ + struct lower_io_state *state = void_state; + + nir_builder *b = &state->builder; + + nir_foreach_instr_safe(block, instr) { + if (instr->type != nir_instr_type_intrinsic) + continue; + + nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); + + if (intrin->intrinsic != nir_intrinsic_load_var && + intrin->intrinsic != nir_intrinsic_store_var) + continue; + + nir_variable_mode mode = intrin->variables[0]->var->data.mode; + + if (state->mode != nir_var_all && state->mode != mode) + continue; + + if (mode != nir_var_shader_in && + mode != nir_var_shader_out && + mode != nir_var_uniform) + continue; + + b->cursor = nir_before_instr(instr); + + switch (intrin->intrinsic) { + case nir_intrinsic_load_var: { + bool per_vertex = + is_per_vertex_input(state, intrin->variables[0]->var) || + is_per_vertex_output(state, intrin->variables[0]->var); + + nir_ssa_def *offset; + nir_ssa_def *vertex_index; + + offset = get_io_offset(b, intrin->variables[0], + per_vertex ? &vertex_index : NULL, + state->type_size); + + nir_intrinsic_instr *load = + nir_intrinsic_instr_create(state->mem_ctx, + load_op(state, mode, per_vertex)); + load->num_components = intrin->num_components; + + load->const_index[0] = + intrin->variables[0]->var->data.driver_location; + + if (per_vertex) + load->src[0] = nir_src_for_ssa(vertex_index); + + load->src[per_vertex ? 1 : 0] = nir_src_for_ssa(offset); + + if (intrin->dest.is_ssa) { + nir_ssa_dest_init(&load->instr, &load->dest, + intrin->num_components, NULL); + nir_ssa_def_rewrite_uses(&intrin->dest.ssa, + nir_src_for_ssa(&load->dest.ssa)); + } else { + nir_dest_copy(&load->dest, &intrin->dest, state->mem_ctx); + } + + nir_instr_insert_before(&intrin->instr, &load->instr); + nir_instr_remove(&intrin->instr); + break; + } + + case nir_intrinsic_store_var: { + assert(mode == nir_var_shader_out); + + nir_ssa_def *offset; + nir_ssa_def *vertex_index; + + bool per_vertex = + is_per_vertex_output(state, intrin->variables[0]->var); + + offset = get_io_offset(b, intrin->variables[0], + per_vertex ? &vertex_index : NULL, + state->type_size); + + nir_intrinsic_op store_op = + per_vertex ? nir_intrinsic_store_per_vertex_output : + nir_intrinsic_store_output; + + nir_intrinsic_instr *store = nir_intrinsic_instr_create(state->mem_ctx, + store_op); + store->num_components = intrin->num_components; + + nir_src_copy(&store->src[0], &intrin->src[0], store); + + store->const_index[0] = + intrin->variables[0]->var->data.driver_location; + + /* Copy the writemask */ + store->const_index[1] = intrin->const_index[0]; + + if (per_vertex) + store->src[1] = nir_src_for_ssa(vertex_index); + + store->src[per_vertex ? 2 : 1] = nir_src_for_ssa(offset); + + nir_instr_insert_before(&intrin->instr, &store->instr); + nir_instr_remove(&intrin->instr); + break; + } + + default: + break; + } + } + + return true; +} + +static void +nir_lower_io_impl(nir_function_impl *impl, + nir_variable_mode mode, + int (*type_size)(const struct glsl_type *)) +{ + struct lower_io_state state; + + nir_builder_init(&state.builder, impl); + state.mem_ctx = ralloc_parent(impl); + state.mode = mode; + state.type_size = type_size; + + nir_foreach_block(impl, nir_lower_io_block, &state); + + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance); +} + +void +nir_lower_io(nir_shader *shader, nir_variable_mode mode, + int (*type_size)(const struct glsl_type *)) +{ + nir_foreach_function(shader, function) { + if (function->impl) + nir_lower_io_impl(function->impl, mode, type_size); + } +} + +/** + * Return the offset soruce for a load/store intrinsic. + */ +nir_src * +nir_get_io_offset_src(nir_intrinsic_instr *instr) +{ + switch (instr->intrinsic) { + case nir_intrinsic_load_input: + case nir_intrinsic_load_output: + case nir_intrinsic_load_uniform: + return &instr->src[0]; + case nir_intrinsic_load_per_vertex_input: + case nir_intrinsic_load_per_vertex_output: + case nir_intrinsic_store_output: + return &instr->src[1]; + case nir_intrinsic_store_per_vertex_output: + return &instr->src[2]; + default: + return NULL; + } +} + +/** + * Return the vertex index source for a load/store per_vertex intrinsic. + */ +nir_src * +nir_get_io_vertex_index_src(nir_intrinsic_instr *instr) +{ + switch (instr->intrinsic) { + case nir_intrinsic_load_per_vertex_input: + case nir_intrinsic_load_per_vertex_output: + return &instr->src[0]; + case nir_intrinsic_store_per_vertex_output: + return &instr->src[1]; + default: + return NULL; + } +} diff --git a/src/compiler/nir/nir_lower_load_const_to_scalar.c b/src/compiler/nir/nir_lower_load_const_to_scalar.c new file mode 100644 index 0000000..1eeed13 --- /dev/null +++ b/src/compiler/nir/nir_lower_load_const_to_scalar.c @@ -0,0 +1,89 @@ +/* + * Copyright © 2015 Broadcom + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + */ + +#include "util/macros.h" +#include "nir.h" +#include "nir_builder.h" + +/** @file nir_lower_load_const_to_scalar.c + * + * Replaces vector nir_load_const instructions with a series of loads and a + * vec[234] to reconstruct the original vector (on the assumption that + * nir_lower_alu_to_scalar() will then be used to split it up). + * + * This gives NIR a chance to CSE more operations on a scalar shader, when the + * same value was used in different vector contant loads. + */ + +static void +lower_load_const_instr_scalar(nir_load_const_instr *lower) +{ + if (lower->def.num_components == 1) + return; + + nir_builder b; + nir_builder_init(&b, nir_cf_node_get_function(&lower->instr.block->cf_node)); + b.cursor = nir_before_instr(&lower->instr); + + /* Emit the individual loads. */ + nir_ssa_def *loads[4]; + for (unsigned i = 0; i < lower->def.num_components; i++) { + nir_load_const_instr *load_comp = nir_load_const_instr_create(b.shader, 1); + load_comp->value.u[0] = lower->value.u[i]; + nir_builder_instr_insert(&b, &load_comp->instr); + loads[i] = &load_comp->def; + } + + /* Batch things back together into a vector. */ + nir_ssa_def *vec = nir_vec(&b, loads, lower->def.num_components); + + /* Replace the old load with a reference to our reconstructed vector. */ + nir_ssa_def_rewrite_uses(&lower->def, nir_src_for_ssa(vec)); + nir_instr_remove(&lower->instr); +} + +static bool +lower_load_const_to_scalar_block(nir_block *block, void *data) +{ + nir_foreach_instr_safe(block, instr) { + if (instr->type == nir_instr_type_load_const) + lower_load_const_instr_scalar(nir_instr_as_load_const(instr)); + } + + return true; +} + +static void +nir_lower_load_const_to_scalar_impl(nir_function_impl *impl) +{ + nir_foreach_block(impl, lower_load_const_to_scalar_block, NULL); +} + +void +nir_lower_load_const_to_scalar(nir_shader *shader) +{ + nir_foreach_function(shader, function) { + if (function->impl) + nir_lower_load_const_to_scalar_impl(function->impl); + } +} diff --git a/src/compiler/nir/nir_lower_locals_to_regs.c b/src/compiler/nir/nir_lower_locals_to_regs.c new file mode 100644 index 0000000..51b0fa7 --- /dev/null +++ b/src/compiler/nir/nir_lower_locals_to_regs.c @@ -0,0 +1,396 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand (jason@jlekstrand.net) + * + */ + +#include "nir.h" +#include "nir_array.h" + +struct locals_to_regs_state { + nir_shader *shader; + nir_function_impl *impl; + + /* A hash table mapping derefs to registers */ + struct hash_table *regs_table; + + /* A growing array of derefs that we have encountered. There is exactly + * one element of this array per element in the hash table. This is + * used to make adding register initialization code deterministic. + */ + nir_array derefs_array; + + bool progress; +}; + +/* The following two functions implement a hash and equality check for + * variable dreferences. When the hash or equality function encounters an + * array, it ignores the offset and whether it is direct or indirect + * entirely. + */ +static uint32_t +hash_deref(const void *void_deref) +{ + uint32_t hash = _mesa_fnv32_1a_offset_bias; + + const nir_deref_var *deref_var = void_deref; + hash = _mesa_fnv32_1a_accumulate(hash, deref_var->var); + + for (const nir_deref *deref = deref_var->deref.child; + deref; deref = deref->child) { + if (deref->deref_type == nir_deref_type_struct) { + const nir_deref_struct *deref_struct = nir_deref_as_struct(deref); + hash = _mesa_fnv32_1a_accumulate(hash, deref_struct->index); + } + } + + return hash; +} + +static bool +derefs_equal(const void *void_a, const void *void_b) +{ + const nir_deref_var *a_var = void_a; + const nir_deref_var *b_var = void_b; + + if (a_var->var != b_var->var) + return false; + + for (const nir_deref *a = a_var->deref.child, *b = b_var->deref.child; + a != NULL; a = a->child, b = b->child) { + if (a->deref_type != b->deref_type) + return false; + + if (a->deref_type == nir_deref_type_struct) { + if (nir_deref_as_struct(a)->index != nir_deref_as_struct(b)->index) + return false; + } + /* Do nothing for arrays. They're all the same. */ + + assert((a->child == NULL) == (b->child == NULL)); + if((a->child == NULL) != (b->child == NULL)) + return false; + } + + return true; +} + +static nir_register * +get_reg_for_deref(nir_deref_var *deref, struct locals_to_regs_state *state) +{ + uint32_t hash = hash_deref(deref); + + struct hash_entry *entry = + _mesa_hash_table_search_pre_hashed(state->regs_table, hash, deref); + if (entry) + return entry->data; + + unsigned array_size = 1; + nir_deref *tail = &deref->deref; + while (tail->child) { + if (tail->child->deref_type == nir_deref_type_array) + array_size *= glsl_get_length(tail->type); + tail = tail->child; + } + + assert(glsl_type_is_vector(tail->type) || glsl_type_is_scalar(tail->type)); + + nir_register *reg = nir_local_reg_create(state->impl); + reg->num_components = glsl_get_vector_elements(tail->type); + reg->num_array_elems = array_size > 1 ? array_size : 0; + + _mesa_hash_table_insert_pre_hashed(state->regs_table, hash, deref, reg); + nir_array_add(&state->derefs_array, nir_deref_var *, deref); + + return reg; +} + +static nir_src +get_deref_reg_src(nir_deref_var *deref, nir_instr *instr, + struct locals_to_regs_state *state) +{ + nir_src src; + + src.is_ssa = false; + src.reg.reg = get_reg_for_deref(deref, state); + src.reg.base_offset = 0; + src.reg.indirect = NULL; + + /* It is possible for a user to create a shader that has an array with a + * single element and then proceed to access it indirectly. Indirectly + * accessing a non-array register is not allowed in NIR. In order to + * handle this case we just convert it to a direct reference. + */ + if (src.reg.reg->num_array_elems == 0) + return src; + + nir_deref *tail = &deref->deref; + while (tail->child != NULL) { + const struct glsl_type *parent_type = tail->type; + tail = tail->child; + + if (tail->deref_type != nir_deref_type_array) + continue; + + nir_deref_array *deref_array = nir_deref_as_array(tail); + + src.reg.base_offset *= glsl_get_length(parent_type); + src.reg.base_offset += deref_array->base_offset; + + if (src.reg.indirect) { + nir_load_const_instr *load_const = + nir_load_const_instr_create(state->shader, 1); + load_const->value.u[0] = glsl_get_length(parent_type); + nir_instr_insert_before(instr, &load_const->instr); + + nir_alu_instr *mul = nir_alu_instr_create(state->shader, nir_op_imul); + mul->src[0].src = *src.reg.indirect; + mul->src[1].src.is_ssa = true; + mul->src[1].src.ssa = &load_const->def; + mul->dest.write_mask = 1; + nir_ssa_dest_init(&mul->instr, &mul->dest.dest, 1, NULL); + nir_instr_insert_before(instr, &mul->instr); + + src.reg.indirect->is_ssa = true; + src.reg.indirect->ssa = &mul->dest.dest.ssa; + } + + if (deref_array->deref_array_type == nir_deref_array_type_indirect) { + if (src.reg.indirect == NULL) { + src.reg.indirect = ralloc(state->shader, nir_src); + nir_src_copy(src.reg.indirect, &deref_array->indirect, + state->shader); + } else { + nir_alu_instr *add = nir_alu_instr_create(state->shader, + nir_op_iadd); + add->src[0].src = *src.reg.indirect; + nir_src_copy(&add->src[1].src, &deref_array->indirect, add); + add->dest.write_mask = 1; + nir_ssa_dest_init(&add->instr, &add->dest.dest, 1, NULL); + nir_instr_insert_before(instr, &add->instr); + + src.reg.indirect->is_ssa = true; + src.reg.indirect->ssa = &add->dest.dest.ssa; + } + } + } + + return src; +} + +static bool +lower_locals_to_regs_block(nir_block *block, void *void_state) +{ + struct locals_to_regs_state *state = void_state; + + nir_foreach_instr_safe(block, instr) { + if (instr->type != nir_instr_type_intrinsic) + continue; + + nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); + + switch (intrin->intrinsic) { + case nir_intrinsic_load_var: { + if (intrin->variables[0]->var->data.mode != nir_var_local) + continue; + + nir_alu_instr *mov = nir_alu_instr_create(state->shader, nir_op_imov); + mov->src[0].src = get_deref_reg_src(intrin->variables[0], + &intrin->instr, state); + mov->dest.write_mask = (1 << intrin->num_components) - 1; + if (intrin->dest.is_ssa) { + nir_ssa_dest_init(&mov->instr, &mov->dest.dest, + intrin->num_components, NULL); + nir_ssa_def_rewrite_uses(&intrin->dest.ssa, + nir_src_for_ssa(&mov->dest.dest.ssa)); + } else { + nir_dest_copy(&mov->dest.dest, &intrin->dest, &mov->instr); + } + nir_instr_insert_before(&intrin->instr, &mov->instr); + + nir_instr_remove(&intrin->instr); + state->progress = true; + break; + } + + case nir_intrinsic_store_var: { + if (intrin->variables[0]->var->data.mode != nir_var_local) + continue; + + nir_src reg_src = get_deref_reg_src(intrin->variables[0], + &intrin->instr, state); + + nir_alu_instr *mov = nir_alu_instr_create(state->shader, nir_op_imov); + nir_src_copy(&mov->src[0].src, &intrin->src[0], mov); + mov->dest.write_mask = intrin->const_index[0]; + mov->dest.dest.is_ssa = false; + mov->dest.dest.reg.reg = reg_src.reg.reg; + mov->dest.dest.reg.base_offset = reg_src.reg.base_offset; + mov->dest.dest.reg.indirect = reg_src.reg.indirect; + + nir_instr_insert_before(&intrin->instr, &mov->instr); + + nir_instr_remove(&intrin->instr); + state->progress = true; + break; + } + + case nir_intrinsic_copy_var: + unreachable("There should be no copies whatsoever at this point"); + break; + + default: + continue; + } + } + + return true; +} + +static nir_block * +compute_reg_usedef_lca(nir_register *reg) +{ + nir_block *lca = NULL; + + list_for_each_entry(nir_dest, def_dest, ®->defs, reg.def_link) + lca = nir_dominance_lca(lca, def_dest->reg.parent_instr->block); + + list_for_each_entry(nir_src, use_src, ®->uses, use_link) + lca = nir_dominance_lca(lca, use_src->parent_instr->block); + + list_for_each_entry(nir_src, use_src, ®->if_uses, use_link) { + nir_cf_node *prev_node = nir_cf_node_prev(&use_src->parent_if->cf_node); + assert(prev_node->type == nir_cf_node_block); + lca = nir_dominance_lca(lca, nir_cf_node_as_block(prev_node)); + } + + return lca; +} + +static void +insert_constant_initializer(nir_deref_var *deref_head, nir_deref *deref_tail, + nir_block *block, + struct locals_to_regs_state *state) +{ + if (deref_tail->child) { + switch (deref_tail->child->deref_type) { + case nir_deref_type_array: { + unsigned array_elems = glsl_get_length(deref_tail->type); + + nir_deref_array arr_deref; + arr_deref.deref = *deref_tail->child; + arr_deref.deref_array_type = nir_deref_array_type_direct; + + nir_deref *old_child = deref_tail->child; + deref_tail->child = &arr_deref.deref; + for (unsigned i = 0; i < array_elems; i++) { + arr_deref.base_offset = i; + insert_constant_initializer(deref_head, &arr_deref.deref, + block, state); + } + deref_tail->child = old_child; + return; + } + + case nir_deref_type_struct: + insert_constant_initializer(deref_head, deref_tail->child, + block, state); + return; + + default: + unreachable("Invalid deref child type"); + } + } + + assert(deref_tail->child == NULL); + + nir_load_const_instr *load = + nir_deref_get_const_initializer_load(state->shader, deref_head); + nir_instr_insert_before_block(block, &load->instr); + + nir_src reg_src = get_deref_reg_src(deref_head, &load->instr, state); + + nir_alu_instr *mov = nir_alu_instr_create(state->shader, nir_op_imov); + mov->src[0].src = nir_src_for_ssa(&load->def); + mov->dest.write_mask = (1 << load->def.num_components) - 1; + mov->dest.dest.is_ssa = false; + mov->dest.dest.reg.reg = reg_src.reg.reg; + mov->dest.dest.reg.base_offset = reg_src.reg.base_offset; + mov->dest.dest.reg.indirect = reg_src.reg.indirect; + + nir_instr_insert_after(&load->instr, &mov->instr); + state->progress = true; +} + +static bool +nir_lower_locals_to_regs_impl(nir_function_impl *impl) +{ + struct locals_to_regs_state state; + + state.shader = impl->function->shader; + state.impl = impl; + state.progress = false; + state.regs_table = _mesa_hash_table_create(NULL, hash_deref, derefs_equal); + nir_array_init(&state.derefs_array, NULL); + + nir_metadata_require(impl, nir_metadata_dominance); + + nir_foreach_block(impl, lower_locals_to_regs_block, &state); + + nir_array_foreach(&state.derefs_array, nir_deref_var *, deref_ptr) { + nir_deref_var *deref = *deref_ptr; + struct hash_entry *deref_entry = + _mesa_hash_table_search(state.regs_table, deref); + assert(deref_entry && deref_entry->key == deref); + nir_register *reg = (nir_register *)deref_entry->data; + + if (deref->var->constant_initializer == NULL) + continue; + + nir_block *usedef_lca = compute_reg_usedef_lca(reg); + + insert_constant_initializer(deref, &deref->deref, usedef_lca, &state); + } + + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance); + + nir_array_fini(&state.derefs_array); + _mesa_hash_table_destroy(state.regs_table, NULL); + + return state.progress; +} + +bool +nir_lower_locals_to_regs(nir_shader *shader) +{ + bool progress = false; + + nir_foreach_function(shader, function) { + if (function->impl) + progress = nir_lower_locals_to_regs_impl(function->impl) || progress; + } + + return progress; +} diff --git a/src/compiler/nir/nir_lower_outputs_to_temporaries.c b/src/compiler/nir/nir_lower_outputs_to_temporaries.c new file mode 100644 index 0000000..71b06b8 --- /dev/null +++ b/src/compiler/nir/nir_lower_outputs_to_temporaries.c @@ -0,0 +1,133 @@ +/* + * Copyright © 2015 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + */ + +/* + * Implements a pass that lowers output variables to a temporary plus an + * output variable with a single copy at each exit point of the shader. + * This way the output variable is only ever written. + * + * Because valid NIR requires that output variables are never read, this + * pass is more of a helper for NIR producers and must be run before the + * shader is ever validated. + */ + +#include "nir.h" + +struct lower_outputs_state { + nir_shader *shader; + struct exec_list old_outputs; +}; + +static void +emit_output_copies(nir_cursor cursor, struct lower_outputs_state *state) +{ + assert(exec_list_length(&state->shader->outputs) == + exec_list_length(&state->old_outputs)); + + foreach_two_lists(out_node, &state->shader->outputs, + temp_node, &state->old_outputs) { + nir_variable *output = exec_node_data(nir_variable, out_node, node); + nir_variable *temp = exec_node_data(nir_variable, temp_node, node); + + nir_intrinsic_instr *copy = + nir_intrinsic_instr_create(state->shader, nir_intrinsic_copy_var); + copy->variables[0] = nir_deref_var_create(copy, output); + copy->variables[1] = nir_deref_var_create(copy, temp); + + nir_instr_insert(cursor, ©->instr); + } +} + +static bool +emit_output_copies_block(nir_block *block, void *state) +{ + nir_foreach_instr(block, instr) { + if (instr->type != nir_instr_type_intrinsic) + continue; + + nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); + if (intrin->intrinsic == nir_intrinsic_emit_vertex) + emit_output_copies(nir_before_instr(&intrin->instr), state); + } + + return true; +} + +void +nir_lower_outputs_to_temporaries(nir_shader *shader) +{ + struct lower_outputs_state state; + + if (shader->stage == MESA_SHADER_TESS_CTRL) + return; + + state.shader = shader; + exec_list_move_nodes_to(&shader->outputs, &state.old_outputs); + + /* Walk over all of the outputs turn each output into a temporary and + * make a new variable for the actual output. + */ + nir_foreach_variable(var, &state.old_outputs) { + nir_variable *output = ralloc(shader, nir_variable); + memcpy(output, var, sizeof *output); + + /* The orignal is now the temporary */ + nir_variable *temp = var; + + /* Reparent the name to the new variable */ + ralloc_steal(output, output->name); + + /* Give the output a new name with @out-temp appended */ + temp->name = ralloc_asprintf(var, "%s@out-temp", output->name); + temp->data.mode = nir_var_global; + temp->constant_initializer = NULL; + + exec_list_push_tail(&shader->outputs, &output->node); + } + + nir_foreach_function(shader, function) { + if (function->impl == NULL) + continue; + + if (shader->stage == MESA_SHADER_GEOMETRY) { + /* For geometry shaders, we have to emit the output copies right + * before each EmitVertex call. + */ + nir_foreach_block(function->impl, emit_output_copies_block, &state); + } else if (strcmp(function->name, "main") == 0) { + /* For all other shader types, we need to do the copies right before + * the jumps to the end block. + */ + struct set_entry *block_entry; + set_foreach(function->impl->end_block->predecessors, block_entry) { + struct nir_block *block = (void *)block_entry->key; + emit_output_copies(nir_after_block_before_jump(block), &state); + } + } + + nir_metadata_preserve(function->impl, nir_metadata_block_index | + nir_metadata_dominance); + } + + exec_list_append(&shader->globals, &state.old_outputs); +} diff --git a/src/compiler/nir/nir_lower_phis_to_scalar.c b/src/compiler/nir/nir_lower_phis_to_scalar.c new file mode 100644 index 0000000..dd2abcf --- /dev/null +++ b/src/compiler/nir/nir_lower_phis_to_scalar.c @@ -0,0 +1,293 @@ +/* + * Copyright © 2015 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand (jason@jlekstrand.net) + * + */ + +#include "nir.h" + +/* + * Implements a pass that lowers vector phi nodes to scalar phi nodes when + * we don't think it will hurt anything. + */ + +struct lower_phis_to_scalar_state { + void *mem_ctx; + void *dead_ctx; + + /* Hash table marking which phi nodes are scalarizable. The key is + * pointers to phi instructions and the entry is either NULL for not + * scalarizable or non-null for scalarizable. + */ + struct hash_table *phi_table; +}; + +static bool +should_lower_phi(nir_phi_instr *phi, struct lower_phis_to_scalar_state *state); + +static bool +is_phi_src_scalarizable(nir_phi_src *src, + struct lower_phis_to_scalar_state *state) +{ + /* Don't know what to do with non-ssa sources */ + if (!src->src.is_ssa) + return false; + + nir_instr *src_instr = src->src.ssa->parent_instr; + switch (src_instr->type) { + case nir_instr_type_alu: { + nir_alu_instr *src_alu = nir_instr_as_alu(src_instr); + + /* ALU operations with output_size == 0 should be scalarized. We + * will also see a bunch of vecN operations from scalarizing ALU + * operations and, since they can easily be copy-propagated, they + * are ok too. + */ + return nir_op_infos[src_alu->op].output_size == 0 || + src_alu->op == nir_op_vec2 || + src_alu->op == nir_op_vec3 || + src_alu->op == nir_op_vec4; + } + + case nir_instr_type_phi: + /* A phi is scalarizable if we're going to lower it */ + return should_lower_phi(nir_instr_as_phi(src_instr), state); + + case nir_instr_type_load_const: + case nir_instr_type_ssa_undef: + /* These are trivially scalarizable */ + return true; + + case nir_instr_type_intrinsic: { + nir_intrinsic_instr *src_intrin = nir_instr_as_intrinsic(src_instr); + + switch (src_intrin->intrinsic) { + case nir_intrinsic_load_var: + return src_intrin->variables[0]->var->data.mode == nir_var_shader_in || + src_intrin->variables[0]->var->data.mode == nir_var_uniform; + + case nir_intrinsic_interp_var_at_centroid: + case nir_intrinsic_interp_var_at_sample: + case nir_intrinsic_interp_var_at_offset: + case nir_intrinsic_load_uniform: + case nir_intrinsic_load_ubo: + case nir_intrinsic_load_ssbo: + case nir_intrinsic_load_input: + return true; + default: + break; + } + } + + default: + /* We can't scalarize this type of instruction */ + return false; + } +} + +/** + * Determines if the given phi node should be lowered. The only phi nodes + * we will scalarize at the moment are those where all of the sources are + * scalarizable. + * + * The reason for this comes down to coalescing. Since phi sources can't + * swizzle, swizzles on phis have to be resolved by inserting a mov right + * before the phi. The choice then becomes between movs to pick off + * components for a scalar phi or potentially movs to recombine components + * for a vector phi. The problem is that the movs generated to pick off + * the components are almost uncoalescable. We can't coalesce them in NIR + * because we need them to pick off components and we can't coalesce them + * in the backend because the source register is a vector and the + * destination is a scalar that may be used at other places in the program. + * On the other hand, if we have a bunch of scalars going into a vector + * phi, the situation is much better. In this case, if the SSA def is + * generated in the predecessor block to the corresponding phi source, the + * backend code will be an ALU op into a temporary and then a mov into the + * given vector component; this move can almost certainly be coalesced + * away. + */ +static bool +should_lower_phi(nir_phi_instr *phi, struct lower_phis_to_scalar_state *state) +{ + /* Already scalar */ + if (phi->dest.ssa.num_components == 1) + return false; + + struct hash_entry *entry = _mesa_hash_table_search(state->phi_table, phi); + if (entry) + return entry->data != NULL; + + /* Insert an entry and mark it as scalarizable for now. That way + * we don't recurse forever and a cycle in the dependence graph + * won't automatically make us fail to scalarize. + */ + entry = _mesa_hash_table_insert(state->phi_table, phi, (void *)(intptr_t)1); + + bool scalarizable = true; + + nir_foreach_phi_src(phi, src) { + scalarizable = is_phi_src_scalarizable(src, state); + if (!scalarizable) + break; + } + + /* The hash table entry for 'phi' may have changed while recursing the + * dependence graph, so we need to reset it */ + entry = _mesa_hash_table_search(state->phi_table, phi); + assert(entry); + + entry->data = (void *)(intptr_t)scalarizable; + + return scalarizable; +} + +static bool +lower_phis_to_scalar_block(nir_block *block, void *void_state) +{ + struct lower_phis_to_scalar_state *state = void_state; + + /* Find the last phi node in the block */ + nir_phi_instr *last_phi = NULL; + nir_foreach_instr(block, instr) { + if (instr->type != nir_instr_type_phi) + break; + + last_phi = nir_instr_as_phi(instr); + } + + /* We have to handle the phi nodes in their own pass due to the way + * we're modifying the linked list of instructions. + */ + nir_foreach_instr_safe(block, instr) { + if (instr->type != nir_instr_type_phi) + break; + + nir_phi_instr *phi = nir_instr_as_phi(instr); + + if (!should_lower_phi(phi, state)) + continue; + + /* Create a vecN operation to combine the results. Most of these + * will be redundant, but copy propagation should clean them up for + * us. No need to add the complexity here. + */ + nir_op vec_op; + switch (phi->dest.ssa.num_components) { + case 2: vec_op = nir_op_vec2; break; + case 3: vec_op = nir_op_vec3; break; + case 4: vec_op = nir_op_vec4; break; + default: unreachable("Invalid number of components"); + } + + nir_alu_instr *vec = nir_alu_instr_create(state->mem_ctx, vec_op); + nir_ssa_dest_init(&vec->instr, &vec->dest.dest, + phi->dest.ssa.num_components, NULL); + vec->dest.write_mask = (1 << phi->dest.ssa.num_components) - 1; + + for (unsigned i = 0; i < phi->dest.ssa.num_components; i++) { + nir_phi_instr *new_phi = nir_phi_instr_create(state->mem_ctx); + nir_ssa_dest_init(&new_phi->instr, &new_phi->dest, 1, NULL); + + vec->src[i].src = nir_src_for_ssa(&new_phi->dest.ssa); + + nir_foreach_phi_src(phi, src) { + /* We need to insert a mov to grab the i'th component of src */ + nir_alu_instr *mov = nir_alu_instr_create(state->mem_ctx, + nir_op_imov); + nir_ssa_dest_init(&mov->instr, &mov->dest.dest, 1, NULL); + mov->dest.write_mask = 1; + nir_src_copy(&mov->src[0].src, &src->src, state->mem_ctx); + mov->src[0].swizzle[0] = i; + + /* Insert at the end of the predecessor but before the jump */ + nir_instr *pred_last_instr = nir_block_last_instr(src->pred); + if (pred_last_instr && pred_last_instr->type == nir_instr_type_jump) + nir_instr_insert_before(pred_last_instr, &mov->instr); + else + nir_instr_insert_after_block(src->pred, &mov->instr); + + nir_phi_src *new_src = ralloc(new_phi, nir_phi_src); + new_src->pred = src->pred; + new_src->src = nir_src_for_ssa(&mov->dest.dest.ssa); + + exec_list_push_tail(&new_phi->srcs, &new_src->node); + } + + nir_instr_insert_before(&phi->instr, &new_phi->instr); + } + + nir_instr_insert_after(&last_phi->instr, &vec->instr); + + nir_ssa_def_rewrite_uses(&phi->dest.ssa, + nir_src_for_ssa(&vec->dest.dest.ssa)); + + ralloc_steal(state->dead_ctx, phi); + nir_instr_remove(&phi->instr); + + /* We're using the safe iterator and inserting all the newly + * scalarized phi nodes before their non-scalarized version so that's + * ok. However, we are also inserting vec operations after all of + * the last phi node so once we get here, we can't trust even the + * safe iterator to stop properly. We have to break manually. + */ + if (instr == &last_phi->instr) + break; + } + + return true; +} + +static void +lower_phis_to_scalar_impl(nir_function_impl *impl) +{ + struct lower_phis_to_scalar_state state; + + state.mem_ctx = ralloc_parent(impl); + state.dead_ctx = ralloc_context(NULL); + state.phi_table = _mesa_hash_table_create(state.dead_ctx, _mesa_hash_pointer, + _mesa_key_pointer_equal); + + nir_foreach_block(impl, lower_phis_to_scalar_block, &state); + + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance); + + ralloc_free(state.dead_ctx); +} + +/** A pass that lowers vector phi nodes to scalar + * + * This pass loops through the blocks and lowers looks for vector phi nodes + * it can lower to scalar phi nodes. Not all phi nodes are lowered. For + * instance, if one of the sources is a non-scalarizable vector, then we + * don't bother lowering because that would generate hard-to-coalesce movs. + */ +void +nir_lower_phis_to_scalar(nir_shader *shader) +{ + nir_foreach_function(shader, function) { + if (function->impl) + lower_phis_to_scalar_impl(function->impl); + } +} diff --git a/src/compiler/nir/nir_lower_samplers.c b/src/compiler/nir/nir_lower_samplers.c new file mode 100644 index 0000000..9c91212 --- /dev/null +++ b/src/compiler/nir/nir_lower_samplers.c @@ -0,0 +1,187 @@ +/* + * Copyright (C) 2005-2007 Brian Paul All Rights Reserved. + * Copyright (C) 2008 VMware, Inc. All Rights Reserved. + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER + * DEALINGS IN THE SOFTWARE. + */ + +#include "nir.h" +#include "nir_builder.h" +#include "program/hash_table.h" +#include "glsl/ir_uniform.h" + +#include "main/compiler.h" +#include "main/mtypes.h" +#include "program/prog_parameter.h" +#include "program/program.h" + +/* Calculate the sampler index based on array indicies and also + * calculate the base uniform location for struct members. + */ +static void +calc_sampler_offsets(nir_deref *tail, nir_tex_instr *instr, + unsigned *array_elements, nir_ssa_def **indirect, + nir_builder *b, unsigned *location) +{ + if (tail->child == NULL) + return; + + switch (tail->child->deref_type) { + case nir_deref_type_array: { + nir_deref_array *deref_array = nir_deref_as_array(tail->child); + + assert(deref_array->deref_array_type != nir_deref_array_type_wildcard); + + calc_sampler_offsets(tail->child, instr, array_elements, + indirect, b, location); + instr->sampler_index += deref_array->base_offset * *array_elements; + + if (deref_array->deref_array_type == nir_deref_array_type_indirect) { + nir_ssa_def *mul = + nir_imul(b, nir_imm_int(b, *array_elements), + nir_ssa_for_src(b, deref_array->indirect, 1)); + + nir_instr_rewrite_src(&instr->instr, &deref_array->indirect, + NIR_SRC_INIT); + + if (*indirect) { + *indirect = nir_iadd(b, *indirect, mul); + } else { + *indirect = mul; + } + } + + *array_elements *= glsl_get_length(tail->type); + break; + } + + case nir_deref_type_struct: { + nir_deref_struct *deref_struct = nir_deref_as_struct(tail->child); + *location += glsl_get_record_location_offset(tail->type, deref_struct->index); + calc_sampler_offsets(tail->child, instr, array_elements, + indirect, b, location); + break; + } + + default: + unreachable("Invalid deref type"); + break; + } +} + +static void +lower_sampler(nir_tex_instr *instr, const struct gl_shader_program *shader_program, + gl_shader_stage stage, nir_builder *builder) +{ + if (instr->sampler == NULL) + return; + + instr->sampler_index = 0; + unsigned location = instr->sampler->var->data.location; + unsigned array_elements = 1; + nir_ssa_def *indirect = NULL; + + builder->cursor = nir_before_instr(&instr->instr); + calc_sampler_offsets(&instr->sampler->deref, instr, &array_elements, + &indirect, builder, &location); + + if (indirect) { + /* First, we have to resize the array of texture sources */ + nir_tex_src *new_srcs = rzalloc_array(instr, nir_tex_src, + instr->num_srcs + 1); + + for (unsigned i = 0; i < instr->num_srcs; i++) { + new_srcs[i].src_type = instr->src[i].src_type; + nir_instr_move_src(&instr->instr, &new_srcs[i].src, + &instr->src[i].src); + } + + ralloc_free(instr->src); + instr->src = new_srcs; + + /* Now we can go ahead and move the source over to being a + * first-class texture source. + */ + instr->src[instr->num_srcs].src_type = nir_tex_src_sampler_offset; + instr->num_srcs++; + nir_instr_rewrite_src(&instr->instr, + &instr->src[instr->num_srcs - 1].src, + nir_src_for_ssa(indirect)); + + instr->sampler_array_size = array_elements; + } + + if (location > shader_program->NumUniformStorage - 1 || + !shader_program->UniformStorage[location].opaque[stage].active) { + assert(!"cannot return a sampler"); + return; + } + + instr->sampler_index += + shader_program->UniformStorage[location].opaque[stage].index; + + instr->sampler = NULL; +} + +typedef struct { + nir_builder builder; + const struct gl_shader_program *shader_program; + gl_shader_stage stage; +} lower_state; + +static bool +lower_block_cb(nir_block *block, void *_state) +{ + lower_state *state = (lower_state *) _state; + + nir_foreach_instr(block, instr) { + if (instr->type == nir_instr_type_tex) { + nir_tex_instr *tex_instr = nir_instr_as_tex(instr); + lower_sampler(tex_instr, state->shader_program, state->stage, + &state->builder); + } + } + + return true; +} + +static void +lower_impl(nir_function_impl *impl, const struct gl_shader_program *shader_program, + gl_shader_stage stage) +{ + lower_state state; + + nir_builder_init(&state.builder, impl); + state.shader_program = shader_program; + state.stage = stage; + + nir_foreach_block(impl, lower_block_cb, &state); +} + +void +nir_lower_samplers(nir_shader *shader, + const struct gl_shader_program *shader_program) +{ + nir_foreach_function(shader, function) { + if (function->impl) + lower_impl(function->impl, shader_program, shader->stage); + } +} diff --git a/src/compiler/nir/nir_lower_system_values.c b/src/compiler/nir/nir_lower_system_values.c new file mode 100644 index 0000000..2bd787d --- /dev/null +++ b/src/compiler/nir/nir_lower_system_values.c @@ -0,0 +1,98 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +#include "nir.h" +#include "nir_builder.h" + +struct lower_system_values_state { + nir_builder builder; + bool progress; +}; + +static bool +convert_block(nir_block *block, void *void_state) +{ + struct lower_system_values_state *state = void_state; + + nir_builder *b = &state->builder; + + nir_foreach_instr_safe(block, instr) { + if (instr->type != nir_instr_type_intrinsic) + continue; + + nir_intrinsic_instr *load_var = nir_instr_as_intrinsic(instr); + + if (load_var->intrinsic != nir_intrinsic_load_var) + continue; + + nir_variable *var = load_var->variables[0]->var; + if (var->data.mode != nir_var_system_value) + continue; + + b->cursor = nir_after_instr(&load_var->instr); + + nir_intrinsic_op sysval_op = + nir_intrinsic_from_system_value(var->data.location); + nir_ssa_def *sysval = nir_load_system_value(b, sysval_op, 0); + + nir_ssa_def_rewrite_uses(&load_var->dest.ssa, nir_src_for_ssa(sysval)); + nir_instr_remove(&load_var->instr); + + state->progress = true; + } + + return true; +} + +static bool +convert_impl(nir_function_impl *impl) +{ + struct lower_system_values_state state; + + state.progress = false; + nir_builder_init(&state.builder, impl); + + nir_foreach_block(impl, convert_block, &state); + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance); + return state.progress; +} + +bool +nir_lower_system_values(nir_shader *shader) +{ + bool progress = false; + + nir_foreach_function(shader, function) { + if (function->impl) + progress = convert_impl(function->impl) || progress; + } + + exec_list_make_empty(&shader->system_values); + + return progress; +} diff --git a/src/compiler/nir/nir_lower_tex.c b/src/compiler/nir/nir_lower_tex.c new file mode 100644 index 0000000..ae24fb2 --- /dev/null +++ b/src/compiler/nir/nir_lower_tex.c @@ -0,0 +1,355 @@ +/* + * Copyright © 2015 Broadcom + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + */ + +/* + * This lowering pass supports (as configured via nir_lower_tex_options) + * various texture related conversions: + * + texture projector lowering: converts the coordinate division for + * texture projection to be done in ALU instructions instead of + * asking the texture operation to do so. + * + lowering RECT: converts the un-normalized RECT texture coordinates + * to normalized coordinates with txs plus ALU instructions + * + saturate s/t/r coords: to emulate certain texture clamp/wrap modes, + * inserts instructions to clamp specified coordinates to [0.0, 1.0]. + * Note that this automatically triggers texture projector lowering if + * needed, since clamping must happen after projector lowering. + */ + +#include "nir.h" +#include "nir_builder.h" + +typedef struct { + nir_builder b; + const nir_lower_tex_options *options; + bool progress; +} lower_tex_state; + +static void +project_src(nir_builder *b, nir_tex_instr *tex) +{ + /* Find the projector in the srcs list, if present. */ + unsigned proj_index; + for (proj_index = 0; proj_index < tex->num_srcs; proj_index++) { + if (tex->src[proj_index].src_type == nir_tex_src_projector) + break; + } + if (proj_index == tex->num_srcs) + return; + + b->cursor = nir_before_instr(&tex->instr); + + nir_ssa_def *inv_proj = + nir_frcp(b, nir_ssa_for_src(b, tex->src[proj_index].src, 1)); + + /* Walk through the sources projecting the arguments. */ + for (unsigned i = 0; i < tex->num_srcs; i++) { + switch (tex->src[i].src_type) { + case nir_tex_src_coord: + case nir_tex_src_comparitor: + break; + default: + continue; + } + nir_ssa_def *unprojected = + nir_ssa_for_src(b, tex->src[i].src, nir_tex_instr_src_size(tex, i)); + nir_ssa_def *projected = nir_fmul(b, unprojected, inv_proj); + + /* Array indices don't get projected, so make an new vector with the + * coordinate's array index untouched. + */ + if (tex->is_array && tex->src[i].src_type == nir_tex_src_coord) { + switch (tex->coord_components) { + case 4: + projected = nir_vec4(b, + nir_channel(b, projected, 0), + nir_channel(b, projected, 1), + nir_channel(b, projected, 2), + nir_channel(b, unprojected, 3)); + break; + case 3: + projected = nir_vec3(b, + nir_channel(b, projected, 0), + nir_channel(b, projected, 1), + nir_channel(b, unprojected, 2)); + break; + case 2: + projected = nir_vec2(b, + nir_channel(b, projected, 0), + nir_channel(b, unprojected, 1)); + break; + default: + unreachable("bad texture coord count for array"); + break; + } + } + + nir_instr_rewrite_src(&tex->instr, + &tex->src[i].src, + nir_src_for_ssa(projected)); + } + + /* Now move the later tex sources down the array so that the projector + * disappears. + */ + nir_instr_rewrite_src(&tex->instr, &tex->src[proj_index].src, + NIR_SRC_INIT); + for (unsigned i = proj_index + 1; i < tex->num_srcs; i++) { + tex->src[i-1].src_type = tex->src[i].src_type; + nir_instr_move_src(&tex->instr, &tex->src[i-1].src, &tex->src[i].src); + } + tex->num_srcs--; +} + +static nir_ssa_def * +get_texture_size(nir_builder *b, nir_tex_instr *tex) +{ + b->cursor = nir_before_instr(&tex->instr); + + /* RECT textures should not be array: */ + assert(!tex->is_array); + + nir_tex_instr *txs; + + txs = nir_tex_instr_create(b->shader, 1); + txs->op = nir_texop_txs; + txs->sampler_dim = GLSL_SAMPLER_DIM_RECT; + txs->sampler_index = tex->sampler_index; + txs->dest_type = nir_type_int; + + /* only single src, the lod: */ + txs->src[0].src = nir_src_for_ssa(nir_imm_int(b, 0)); + txs->src[0].src_type = nir_tex_src_lod; + + nir_ssa_dest_init(&txs->instr, &txs->dest, 2, NULL); + nir_builder_instr_insert(b, &txs->instr); + + return nir_i2f(b, &txs->dest.ssa); +} + +static void +lower_rect(nir_builder *b, nir_tex_instr *tex) +{ + nir_ssa_def *txs = get_texture_size(b, tex); + nir_ssa_def *scale = nir_frcp(b, txs); + + /* Walk through the sources normalizing the requested arguments. */ + for (unsigned i = 0; i < tex->num_srcs; i++) { + if (tex->src[i].src_type != nir_tex_src_coord) + continue; + + nir_ssa_def *coords = + nir_ssa_for_src(b, tex->src[i].src, tex->coord_components); + nir_instr_rewrite_src(&tex->instr, + &tex->src[i].src, + nir_src_for_ssa(nir_fmul(b, coords, scale))); + } + + tex->sampler_dim = GLSL_SAMPLER_DIM_2D; +} + +static void +saturate_src(nir_builder *b, nir_tex_instr *tex, unsigned sat_mask) +{ + b->cursor = nir_before_instr(&tex->instr); + + /* Walk through the sources saturating the requested arguments. */ + for (unsigned i = 0; i < tex->num_srcs; i++) { + if (tex->src[i].src_type != nir_tex_src_coord) + continue; + + nir_ssa_def *src = + nir_ssa_for_src(b, tex->src[i].src, tex->coord_components); + + /* split src into components: */ + nir_ssa_def *comp[4]; + + for (unsigned j = 0; j < tex->coord_components; j++) + comp[j] = nir_channel(b, src, j); + + /* clamp requested components, array index does not get clamped: */ + unsigned ncomp = tex->coord_components; + if (tex->is_array) + ncomp--; + + for (unsigned j = 0; j < ncomp; j++) { + if ((1 << j) & sat_mask) { + if (tex->sampler_dim == GLSL_SAMPLER_DIM_RECT) { + /* non-normalized texture coords, so clamp to texture + * size rather than [0.0, 1.0] + */ + nir_ssa_def *txs = get_texture_size(b, tex); + comp[j] = nir_fmax(b, comp[j], nir_imm_float(b, 0.0)); + comp[j] = nir_fmin(b, comp[j], nir_channel(b, txs, j)); + } else { + comp[j] = nir_fsat(b, comp[j]); + } + } + } + + /* and move the result back into a single vecN: */ + src = nir_vec(b, comp, tex->coord_components); + + nir_instr_rewrite_src(&tex->instr, + &tex->src[i].src, + nir_src_for_ssa(src)); + } +} + +static nir_ssa_def * +get_zero_or_one(nir_builder *b, nir_alu_type type, uint8_t swizzle_val) +{ + nir_const_value v; + + memset(&v, 0, sizeof(v)); + + if (swizzle_val == 4) { + v.u[0] = v.u[1] = v.u[2] = v.u[3] = 0; + } else { + assert(swizzle_val == 5); + if (type == nir_type_float) + v.f[0] = v.f[1] = v.f[2] = v.f[3] = 1.0; + else + v.u[0] = v.u[1] = v.u[2] = v.u[3] = 1; + } + + return nir_build_imm(b, 4, v); +} + +static void +swizzle_result(nir_builder *b, nir_tex_instr *tex, const uint8_t swizzle[4]) +{ + assert(tex->dest.is_ssa); + + b->cursor = nir_after_instr(&tex->instr); + + nir_ssa_def *swizzled; + if (tex->op == nir_texop_tg4) { + if (swizzle[tex->component] < 4) { + /* This one's easy */ + tex->component = swizzle[tex->component]; + return; + } else { + swizzled = get_zero_or_one(b, tex->dest_type, swizzle[tex->component]); + } + } else { + assert(nir_tex_instr_dest_size(tex) == 4); + if (swizzle[0] < 4 && swizzle[1] < 4 && + swizzle[2] < 4 && swizzle[3] < 4) { + unsigned swiz[4] = { swizzle[0], swizzle[1], swizzle[2], swizzle[3] }; + /* We have no 0's or 1's, just emit a swizzling MOV */ + swizzled = nir_swizzle(b, &tex->dest.ssa, swiz, 4, false); + } else { + nir_ssa_def *srcs[4]; + for (unsigned i = 0; i < 4; i++) { + if (swizzle[i] < 4) { + srcs[i] = nir_channel(b, &tex->dest.ssa, swizzle[i]); + } else { + srcs[i] = get_zero_or_one(b, tex->dest_type, swizzle[i]); + } + } + swizzled = nir_vec(b, srcs, 4); + } + } + + nir_ssa_def_rewrite_uses_after(&tex->dest.ssa, nir_src_for_ssa(swizzled), + swizzled->parent_instr); +} + +static bool +nir_lower_tex_block(nir_block *block, void *void_state) +{ + lower_tex_state *state = void_state; + nir_builder *b = &state->b; + + nir_foreach_instr_safe(block, instr) { + if (instr->type != nir_instr_type_tex) + continue; + + nir_tex_instr *tex = nir_instr_as_tex(instr); + bool lower_txp = !!(state->options->lower_txp & (1 << tex->sampler_dim)); + + /* mask of src coords to saturate (clamp): */ + unsigned sat_mask = 0; + + if ((1 << tex->sampler_index) & state->options->saturate_r) + sat_mask |= (1 << 2); /* .z */ + if ((1 << tex->sampler_index) & state->options->saturate_t) + sat_mask |= (1 << 1); /* .y */ + if ((1 << tex->sampler_index) & state->options->saturate_s) + sat_mask |= (1 << 0); /* .x */ + + /* If we are clamping any coords, we must lower projector first + * as clamping happens *after* projection: + */ + if (lower_txp || sat_mask) { + project_src(b, tex); + state->progress = true; + } + + if ((tex->sampler_dim == GLSL_SAMPLER_DIM_RECT) && + state->options->lower_rect) { + lower_rect(b, tex); + state->progress = true; + } + + if (sat_mask) { + saturate_src(b, tex, sat_mask); + state->progress = true; + } + + if (((1 << tex->sampler_index) & state->options->swizzle_result) && + !nir_tex_instr_is_query(tex) && + !(tex->is_shadow && tex->is_new_style_shadow)) { + swizzle_result(b, tex, state->options->swizzles[tex->sampler_index]); + state->progress = true; + } + } + + return true; +} + +static void +nir_lower_tex_impl(nir_function_impl *impl, lower_tex_state *state) +{ + nir_builder_init(&state->b, impl); + + nir_foreach_block(impl, nir_lower_tex_block, state); + + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance); +} + +bool +nir_lower_tex(nir_shader *shader, const nir_lower_tex_options *options) +{ + lower_tex_state state; + state.options = options; + state.progress = false; + + nir_foreach_function(shader, function) { + if (function->impl) + nir_lower_tex_impl(function->impl, &state); + } + + return state.progress; +} diff --git a/src/compiler/nir/nir_lower_to_source_mods.c b/src/compiler/nir/nir_lower_to_source_mods.c new file mode 100644 index 0000000..6c4e1f0 --- /dev/null +++ b/src/compiler/nir/nir_lower_to_source_mods.c @@ -0,0 +1,196 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand (jason@jlekstrand.net) + * + */ + +#include "nir.h" + +/* + * This pass lowers the neg, abs, and sat operations to source modifiers on + * ALU operations to make things nicer for the backend. It's just much + * easier to not have them when we're doing optimizations. + */ + +static bool +nir_lower_to_source_mods_block(nir_block *block, void *state) +{ + nir_foreach_instr(block, instr) { + if (instr->type != nir_instr_type_alu) + continue; + + nir_alu_instr *alu = nir_instr_as_alu(instr); + + for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) { + if (!alu->src[i].src.is_ssa) + continue; + + if (alu->src[i].src.ssa->parent_instr->type != nir_instr_type_alu) + continue; + + nir_alu_instr *parent = nir_instr_as_alu(alu->src[i].src.ssa->parent_instr); + + if (parent->dest.saturate) + continue; + + switch (nir_op_infos[alu->op].input_types[i]) { + case nir_type_float: + if (parent->op != nir_op_fmov) + continue; + break; + case nir_type_int: + if (parent->op != nir_op_imov) + continue; + break; + default: + continue; + } + + /* We can only do a rewrite if the source we are copying is SSA. + * Otherwise, moving the read might invalidly reorder reads/writes + * on a register. + */ + if (!parent->src[0].src.is_ssa) + continue; + + nir_instr_rewrite_src(instr, &alu->src[i].src, parent->src[0].src); + if (alu->src[i].abs) { + /* abs trumps both neg and abs, do nothing */ + } else { + alu->src[i].negate = (alu->src[i].negate != parent->src[0].negate); + alu->src[i].abs |= parent->src[0].abs; + } + + for (int j = 0; j < 4; ++j) { + if (!nir_alu_instr_channel_used(alu, i, j)) + continue; + alu->src[i].swizzle[j] = parent->src[0].swizzle[alu->src[i].swizzle[j]]; + } + + if (list_empty(&parent->dest.dest.ssa.uses) && + list_empty(&parent->dest.dest.ssa.if_uses)) + nir_instr_remove(&parent->instr); + } + + switch (alu->op) { + case nir_op_fsat: + alu->op = nir_op_fmov; + alu->dest.saturate = true; + break; + case nir_op_ineg: + alu->op = nir_op_imov; + alu->src[0].negate = !alu->src[0].negate; + break; + case nir_op_fneg: + alu->op = nir_op_fmov; + alu->src[0].negate = !alu->src[0].negate; + break; + case nir_op_iabs: + alu->op = nir_op_imov; + alu->src[0].abs = true; + alu->src[0].negate = false; + break; + case nir_op_fabs: + alu->op = nir_op_fmov; + alu->src[0].abs = true; + alu->src[0].negate = false; + break; + default: + break; + } + + /* We've covered sources. Now we're going to try and saturate the + * destination if we can. + */ + + if (!alu->dest.dest.is_ssa) + continue; + + /* We can only saturate float destinations */ + if (nir_op_infos[alu->op].output_type != nir_type_float) + continue; + + if (!list_empty(&alu->dest.dest.ssa.if_uses)) + continue; + + bool all_children_are_sat = true; + nir_foreach_use(&alu->dest.dest.ssa, child_src) { + assert(child_src->is_ssa); + nir_instr *child = child_src->parent_instr; + if (child->type != nir_instr_type_alu) { + all_children_are_sat = false; + continue; + } + + nir_alu_instr *child_alu = nir_instr_as_alu(child); + if (child_alu->src[0].negate || child_alu->src[0].abs) { + all_children_are_sat = false; + continue; + } + + if (child_alu->op != nir_op_fsat && + !(child_alu->op == nir_op_fmov && child_alu->dest.saturate)) { + all_children_are_sat = false; + continue; + } + } + + if (!all_children_are_sat) + continue; + + alu->dest.saturate = true; + + nir_foreach_use(&alu->dest.dest.ssa, child_src) { + assert(child_src->is_ssa); + nir_instr *child = child_src->parent_instr; + assert(child->type == nir_instr_type_alu); + nir_alu_instr *child_alu = nir_instr_as_alu(child); + + child_alu->op = nir_op_fmov; + child_alu->dest.saturate = false; + /* We could propagate the dest of our instruction to the + * destinations of the uses here. However, one quick round of + * copy propagation will clean that all up and then we don't have + * the complexity. + */ + } + } + + return true; +} + +static void +nir_lower_to_source_mods_impl(nir_function_impl *impl) +{ + nir_foreach_block(impl, nir_lower_to_source_mods_block, NULL); +} + +void +nir_lower_to_source_mods(nir_shader *shader) +{ + nir_foreach_function(shader, function) { + if (function->impl) + nir_lower_to_source_mods_impl(function->impl); + } +} diff --git a/src/compiler/nir/nir_lower_two_sided_color.c b/src/compiler/nir/nir_lower_two_sided_color.c new file mode 100644 index 0000000..1294cb8 --- /dev/null +++ b/src/compiler/nir/nir_lower_two_sided_color.c @@ -0,0 +1,212 @@ +/* + * Copyright © 2015 Red Hat + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Rob Clark <robclark@freedesktop.org> + */ + +#include "nir.h" +#include "nir_builder.h" + +#define MAX_COLORS 2 /* VARYING_SLOT_COL0/COL1 */ + +typedef struct { + nir_builder b; + nir_shader *shader; + nir_variable *face; + struct { + nir_variable *front; /* COLn */ + nir_variable *back; /* BFCn */ + } colors[MAX_COLORS]; + int colors_count; +} lower_2side_state; + + +/* Lowering pass for fragment shaders to emulated two-sided-color. For + * each COLOR input, a corresponding BCOLOR input is created, and bcsel + * instruction used to select front or back color based on FACE. + */ + +static nir_variable * +create_input(nir_shader *shader, unsigned drvloc, gl_varying_slot slot) +{ + nir_variable *var = rzalloc(shader, nir_variable); + + var->data.driver_location = drvloc; + var->type = glsl_vec4_type(); + var->data.mode = nir_var_shader_in; + var->name = ralloc_asprintf(var, "in_%d", drvloc); + var->data.index = 0; + var->data.location = slot; + + exec_list_push_tail(&shader->inputs, &var->node); + + shader->num_inputs++; /* TODO use type_size() */ + + return var; +} + +static nir_ssa_def * +load_input(nir_builder *b, nir_variable *in) +{ + nir_intrinsic_instr *load; + + load = nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_input); + load->num_components = 4; + load->const_index[0] = in->data.driver_location; + load->src[0] = nir_src_for_ssa(nir_imm_int(b, 0)); + nir_ssa_dest_init(&load->instr, &load->dest, 4, NULL); + nir_builder_instr_insert(b, &load->instr); + + return &load->dest.ssa; +} + +static int +setup_inputs(lower_2side_state *state) +{ + int maxloc = -1; + + /* find color/face inputs: */ + nir_foreach_variable(var, &state->shader->inputs) { + int loc = var->data.driver_location; + + /* keep track of last used driver-location.. we'll be + * appending BCLr/FACE after last existing input: + */ + maxloc = MAX2(maxloc, loc); + + switch (var->data.location) { + case VARYING_SLOT_COL0: + case VARYING_SLOT_COL1: + assert(state->colors_count < ARRAY_SIZE(state->colors)); + state->colors[state->colors_count].front = var; + state->colors_count++; + break; + case VARYING_SLOT_FACE: + state->face = var; + break; + } + } + + /* if we don't have any color inputs, nothing to do: */ + if (state->colors_count == 0) + return -1; + + /* if we don't already have one, insert a FACE input: */ + if (!state->face) { + state->face = create_input(state->shader, ++maxloc, VARYING_SLOT_FACE); + state->face->data.interpolation = INTERP_QUALIFIER_FLAT; + } + + /* add required back-face color inputs: */ + for (int i = 0; i < state->colors_count; i++) { + gl_varying_slot slot; + + if (state->colors[i].front->data.location == VARYING_SLOT_COL0) + slot = VARYING_SLOT_BFC0; + else + slot = VARYING_SLOT_BFC1; + + state->colors[i].back = create_input(state->shader, ++maxloc, slot); + } + + return 0; +} + +static bool +nir_lower_two_sided_color_block(nir_block *block, void *void_state) +{ + lower_2side_state *state = void_state; + nir_builder *b = &state->b; + + nir_foreach_instr_safe(block, instr) { + if (instr->type != nir_instr_type_intrinsic) + continue; + + nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr); + + if (intr->intrinsic != nir_intrinsic_load_input) + continue; + + int idx; + for (idx = 0; idx < state->colors_count; idx++) { + unsigned drvloc = + state->colors[idx].front->data.driver_location; + if (intr->const_index[0] == drvloc) { + assert(nir_src_as_const_value(intr->src[0])); + break; + } + } + + if (idx == state->colors_count) + continue; + + /* replace load_input(COLn) with + * bcsel(load_input(FACE), load_input(COLn), load_input(BFCn)) + */ + b->cursor = nir_before_instr(&intr->instr); + nir_ssa_def *face = nir_channel(b, load_input(b, state->face), 0); + nir_ssa_def *front = load_input(b, state->colors[idx].front); + nir_ssa_def *back = load_input(b, state->colors[idx].back); + nir_ssa_def *cond = nir_flt(b, face, nir_imm_float(b, 0.0)); + nir_ssa_def *color = nir_bcsel(b, cond, back, front); + + assert(intr->dest.is_ssa); + nir_ssa_def_rewrite_uses(&intr->dest.ssa, nir_src_for_ssa(color)); + } + + return true; +} + +static void +nir_lower_two_sided_color_impl(nir_function_impl *impl, + lower_2side_state *state) +{ + nir_builder *b = &state->b; + + nir_builder_init(b, impl); + + nir_foreach_block(impl, nir_lower_two_sided_color_block, state); + + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance); +} + +void +nir_lower_two_sided_color(nir_shader *shader) +{ + lower_2side_state state = { + .shader = shader, + }; + + if (shader->stage != MESA_SHADER_FRAGMENT) + return; + + if (setup_inputs(&state) != 0) + return; + + nir_foreach_function(shader, function) { + if (function->impl) + nir_lower_two_sided_color_impl(function->impl, &state); + } + +} diff --git a/src/compiler/nir/nir_lower_var_copies.c b/src/compiler/nir/nir_lower_var_copies.c new file mode 100644 index 0000000..8cb3edd --- /dev/null +++ b/src/compiler/nir/nir_lower_var_copies.c @@ -0,0 +1,190 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand (jason@jlekstrand.net) + * + */ + +#include "nir.h" +#include "compiler/nir_types.h" + +/* + * Lowers all copy intrinsics to sequences of load/store intrinsics. + */ + +/* Walks down the deref chain and returns the next deref in the chain whose + * child is a wildcard. In other words, given the chain a[1].foo[*].bar, + * this function will return the deref to foo. Calling it a second time + * with the [*].bar, it will return NULL. + */ +static nir_deref * +deref_next_wildcard_parent(nir_deref *deref) +{ + for (nir_deref *tail = deref; tail->child; tail = tail->child) { + if (tail->child->deref_type != nir_deref_type_array) + continue; + + nir_deref_array *arr = nir_deref_as_array(tail->child); + + if (arr->deref_array_type == nir_deref_array_type_wildcard) + return tail; + } + + return NULL; +} + +/* This function recursively walks the given deref chain and replaces the + * given copy instruction with an equivalent sequence load/store + * operations. + * + * @copy_instr The copy instruction to replace; new instructions will be + * inserted before this one + * + * @dest_head The head of the destination variable deref chain + * + * @src_head The head of the source variable deref chain + * + * @dest_tail The current tail of the destination variable deref chain; + * this is used for recursion and external callers of this + * function should call it with tail == head + * + * @src_tail The current tail of the source variable deref chain; + * this is used for recursion and external callers of this + * function should call it with tail == head + * + * @state The current variable lowering state + */ +static void +emit_copy_load_store(nir_intrinsic_instr *copy_instr, + nir_deref_var *dest_head, nir_deref_var *src_head, + nir_deref *dest_tail, nir_deref *src_tail, void *mem_ctx) +{ + /* Find the next pair of wildcards */ + nir_deref *src_arr_parent = deref_next_wildcard_parent(src_tail); + nir_deref *dest_arr_parent = deref_next_wildcard_parent(dest_tail); + + if (src_arr_parent || dest_arr_parent) { + /* Wildcards had better come in matched pairs */ + assert(dest_arr_parent && dest_arr_parent); + + nir_deref_array *src_arr = nir_deref_as_array(src_arr_parent->child); + nir_deref_array *dest_arr = nir_deref_as_array(dest_arr_parent->child); + + unsigned length = glsl_get_length(src_arr_parent->type); + /* The wildcards should represent the same number of elements */ + assert(length == glsl_get_length(dest_arr_parent->type)); + assert(length > 0); + + /* Walk over all of the elements that this wildcard refers to and + * call emit_copy_load_store on each one of them */ + src_arr->deref_array_type = nir_deref_array_type_direct; + dest_arr->deref_array_type = nir_deref_array_type_direct; + for (unsigned i = 0; i < length; i++) { + src_arr->base_offset = i; + dest_arr->base_offset = i; + emit_copy_load_store(copy_instr, dest_head, src_head, + &dest_arr->deref, &src_arr->deref, mem_ctx); + } + src_arr->deref_array_type = nir_deref_array_type_wildcard; + dest_arr->deref_array_type = nir_deref_array_type_wildcard; + } else { + /* In this case, we have no wildcards anymore, so all we have to do + * is just emit the load and store operations. */ + src_tail = nir_deref_tail(src_tail); + dest_tail = nir_deref_tail(dest_tail); + + assert(src_tail->type == dest_tail->type); + + unsigned num_components = glsl_get_vector_elements(src_tail->type); + + nir_intrinsic_instr *load = + nir_intrinsic_instr_create(mem_ctx, nir_intrinsic_load_var); + load->num_components = num_components; + load->variables[0] = nir_deref_as_var(nir_copy_deref(load, &src_head->deref)); + nir_ssa_dest_init(&load->instr, &load->dest, num_components, NULL); + + nir_instr_insert_before(©_instr->instr, &load->instr); + + nir_intrinsic_instr *store = + nir_intrinsic_instr_create(mem_ctx, nir_intrinsic_store_var); + store->num_components = num_components; + store->const_index[0] = (1 << num_components) - 1; + store->variables[0] = nir_deref_as_var(nir_copy_deref(store, &dest_head->deref)); + + store->src[0].is_ssa = true; + store->src[0].ssa = &load->dest.ssa; + + nir_instr_insert_before(©_instr->instr, &store->instr); + } +} + +/* Lowers a copy instruction to a sequence of load/store instructions + * + * The new instructions are placed before the copy instruction in the IR. + */ +void +nir_lower_var_copy_instr(nir_intrinsic_instr *copy, void *mem_ctx) +{ + assert(copy->intrinsic == nir_intrinsic_copy_var); + emit_copy_load_store(copy, copy->variables[0], copy->variables[1], + ©->variables[0]->deref, + ©->variables[1]->deref, mem_ctx); +} + +static bool +lower_var_copies_block(nir_block *block, void *mem_ctx) +{ + nir_foreach_instr_safe(block, instr) { + if (instr->type != nir_instr_type_intrinsic) + continue; + + nir_intrinsic_instr *copy = nir_instr_as_intrinsic(instr); + if (copy->intrinsic != nir_intrinsic_copy_var) + continue; + + nir_lower_var_copy_instr(copy, mem_ctx); + + nir_instr_remove(©->instr); + ralloc_free(copy); + } + + return true; +} + +static void +lower_var_copies_impl(nir_function_impl *impl) +{ + nir_foreach_block(impl, lower_var_copies_block, ralloc_parent(impl)); +} + +/* Lowers every copy_var instruction in the program to a sequence of + * load/store instructions. + */ +void +nir_lower_var_copies(nir_shader *shader) +{ + nir_foreach_function(shader, function) { + if (function->impl) + lower_var_copies_impl(function->impl); + } +} diff --git a/src/compiler/nir/nir_lower_vars_to_ssa.c b/src/compiler/nir/nir_lower_vars_to_ssa.c new file mode 100644 index 0000000..75d31ff --- /dev/null +++ b/src/compiler/nir/nir_lower_vars_to_ssa.c @@ -0,0 +1,973 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand (jason@jlekstrand.net) + * + */ + +#include "nir.h" +#include "nir_builder.h" +#include "nir_vla.h" + + +struct deref_node { + struct deref_node *parent; + const struct glsl_type *type; + + bool lower_to_ssa; + + /* Only valid for things that end up in the direct list. + * Note that multiple nir_deref_vars may correspond to this node, but they + * will all be equivalent, so any is as good as the other. + */ + nir_deref_var *deref; + struct exec_node direct_derefs_link; + + struct set *loads; + struct set *stores; + struct set *copies; + + nir_ssa_def **def_stack; + nir_ssa_def **def_stack_tail; + + struct deref_node *wildcard; + struct deref_node *indirect; + struct deref_node *children[0]; +}; + +struct lower_variables_state { + nir_shader *shader; + void *dead_ctx; + nir_function_impl *impl; + + /* A hash table mapping variables to deref_node data */ + struct hash_table *deref_var_nodes; + + /* A hash table mapping fully-qualified direct dereferences, i.e. + * dereferences with no indirect or wildcard array dereferences, to + * deref_node data. + * + * At the moment, we only lower loads, stores, and copies that can be + * trivially lowered to loads and stores, i.e. copies with no indirects + * and no wildcards. If a part of a variable that is being loaded from + * and/or stored into is also involved in a copy operation with + * wildcards, then we lower that copy operation to loads and stores, but + * otherwise we leave copies with wildcards alone. Since the only derefs + * used in these loads, stores, and trivial copies are ones with no + * wildcards and no indirects, these are precisely the derefs that we + * can actually consider lowering. + */ + struct exec_list direct_deref_nodes; + + /* Controls whether get_deref_node will add variables to the + * direct_deref_nodes table. This is turned on when we are initially + * scanning for load/store instructions. It is then turned off so we + * don't accidentally change the direct_deref_nodes table while we're + * iterating throug it. + */ + bool add_to_direct_deref_nodes; + + /* A hash table mapping phi nodes to deref_state data */ + struct hash_table *phi_table; +}; + +static struct deref_node * +deref_node_create(struct deref_node *parent, + const struct glsl_type *type, nir_shader *shader) +{ + size_t size = sizeof(struct deref_node) + + glsl_get_length(type) * sizeof(struct deref_node *); + + struct deref_node *node = rzalloc_size(shader, size); + node->type = type; + node->parent = parent; + node->deref = NULL; + exec_node_init(&node->direct_derefs_link); + + return node; +} + +/* Returns the deref node associated with the given variable. This will be + * the root of the tree representing all of the derefs of the given variable. + */ +static struct deref_node * +get_deref_node_for_var(nir_variable *var, struct lower_variables_state *state) +{ + struct deref_node *node; + + struct hash_entry *var_entry = + _mesa_hash_table_search(state->deref_var_nodes, var); + + if (var_entry) { + return var_entry->data; + } else { + node = deref_node_create(NULL, var->type, state->dead_ctx); + _mesa_hash_table_insert(state->deref_var_nodes, var, node); + return node; + } +} + +/* Gets the deref_node for the given deref chain and creates it if it + * doesn't yet exist. If the deref is fully-qualified and direct and + * state->add_to_direct_deref_nodes is true, it will be added to the hash + * table of of fully-qualified direct derefs. + */ +static struct deref_node * +get_deref_node(nir_deref_var *deref, struct lower_variables_state *state) +{ + bool is_direct = true; + + /* Start at the base of the chain. */ + struct deref_node *node = get_deref_node_for_var(deref->var, state); + assert(deref->deref.type == node->type); + + for (nir_deref *tail = deref->deref.child; tail; tail = tail->child) { + switch (tail->deref_type) { + case nir_deref_type_struct: { + nir_deref_struct *deref_struct = nir_deref_as_struct(tail); + + assert(deref_struct->index < glsl_get_length(node->type)); + + if (node->children[deref_struct->index] == NULL) + node->children[deref_struct->index] = + deref_node_create(node, tail->type, state->dead_ctx); + + node = node->children[deref_struct->index]; + break; + } + + case nir_deref_type_array: { + nir_deref_array *arr = nir_deref_as_array(tail); + + switch (arr->deref_array_type) { + case nir_deref_array_type_direct: + /* This is possible if a loop unrolls and generates an + * out-of-bounds offset. We need to handle this at least + * somewhat gracefully. + */ + if (arr->base_offset >= glsl_get_length(node->type)) + return NULL; + + if (node->children[arr->base_offset] == NULL) + node->children[arr->base_offset] = + deref_node_create(node, tail->type, state->dead_ctx); + + node = node->children[arr->base_offset]; + break; + + case nir_deref_array_type_indirect: + if (node->indirect == NULL) + node->indirect = deref_node_create(node, tail->type, + state->dead_ctx); + + node = node->indirect; + is_direct = false; + break; + + case nir_deref_array_type_wildcard: + if (node->wildcard == NULL) + node->wildcard = deref_node_create(node, tail->type, + state->dead_ctx); + + node = node->wildcard; + is_direct = false; + break; + + default: + unreachable("Invalid array deref type"); + } + break; + } + default: + unreachable("Invalid deref type"); + } + } + + assert(node); + + /* Only insert if it isn't already in the list. */ + if (is_direct && state->add_to_direct_deref_nodes && + node->direct_derefs_link.next == NULL) { + node->deref = deref; + assert(deref->var != NULL); + exec_list_push_tail(&state->direct_deref_nodes, + &node->direct_derefs_link); + } + + return node; +} + +/* \sa foreach_deref_node_match */ +static bool +foreach_deref_node_worker(struct deref_node *node, nir_deref *deref, + bool (* cb)(struct deref_node *node, + struct lower_variables_state *state), + struct lower_variables_state *state) +{ + if (deref->child == NULL) { + return cb(node, state); + } else { + switch (deref->child->deref_type) { + case nir_deref_type_array: { + nir_deref_array *arr = nir_deref_as_array(deref->child); + assert(arr->deref_array_type == nir_deref_array_type_direct); + if (node->children[arr->base_offset] && + !foreach_deref_node_worker(node->children[arr->base_offset], + deref->child, cb, state)) + return false; + + if (node->wildcard && + !foreach_deref_node_worker(node->wildcard, + deref->child, cb, state)) + return false; + + return true; + } + + case nir_deref_type_struct: { + nir_deref_struct *str = nir_deref_as_struct(deref->child); + return foreach_deref_node_worker(node->children[str->index], + deref->child, cb, state); + } + + default: + unreachable("Invalid deref child type"); + } + } +} + +/* Walks over every "matching" deref_node and calls the callback. A node + * is considered to "match" if either refers to that deref or matches up t + * a wildcard. In other words, the following would match a[6].foo[3].bar: + * + * a[6].foo[3].bar + * a[*].foo[3].bar + * a[6].foo[*].bar + * a[*].foo[*].bar + * + * The given deref must be a full-length and fully qualified (no wildcards + * or indirects) deref chain. + */ +static bool +foreach_deref_node_match(nir_deref_var *deref, + bool (* cb)(struct deref_node *node, + struct lower_variables_state *state), + struct lower_variables_state *state) +{ + nir_deref_var var_deref = *deref; + var_deref.deref.child = NULL; + struct deref_node *node = get_deref_node(&var_deref, state); + + if (node == NULL) + return false; + + return foreach_deref_node_worker(node, &deref->deref, cb, state); +} + +/* \sa deref_may_be_aliased */ +static bool +deref_may_be_aliased_node(struct deref_node *node, nir_deref *deref, + struct lower_variables_state *state) +{ + if (deref->child == NULL) { + return false; + } else { + switch (deref->child->deref_type) { + case nir_deref_type_array: { + nir_deref_array *arr = nir_deref_as_array(deref->child); + if (arr->deref_array_type == nir_deref_array_type_indirect) + return true; + + /* If there is an indirect at this level, we're aliased. */ + if (node->indirect) + return true; + + assert(arr->deref_array_type == nir_deref_array_type_direct); + + if (node->children[arr->base_offset] && + deref_may_be_aliased_node(node->children[arr->base_offset], + deref->child, state)) + return true; + + if (node->wildcard && + deref_may_be_aliased_node(node->wildcard, deref->child, state)) + return true; + + return false; + } + + case nir_deref_type_struct: { + nir_deref_struct *str = nir_deref_as_struct(deref->child); + if (node->children[str->index]) { + return deref_may_be_aliased_node(node->children[str->index], + deref->child, state); + } else { + return false; + } + } + + default: + unreachable("Invalid nir_deref child type"); + } + } +} + +/* Returns true if there are no indirects that can ever touch this deref. + * + * For example, if the given deref is a[6].foo, then any uses of a[i].foo + * would cause this to return false, but a[i].bar would not affect it + * because it's a different structure member. A var_copy involving of + * a[*].bar also doesn't affect it because that can be lowered to entirely + * direct load/stores. + * + * We only support asking this question about fully-qualified derefs. + * Obviously, it's pointless to ask this about indirects, but we also + * rule-out wildcards. Handling Wildcard dereferences would involve + * checking each array index to make sure that there aren't any indirect + * references. + */ +static bool +deref_may_be_aliased(nir_deref_var *deref, + struct lower_variables_state *state) +{ + return deref_may_be_aliased_node(get_deref_node_for_var(deref->var, state), + &deref->deref, state); +} + +static void +register_load_instr(nir_intrinsic_instr *load_instr, + struct lower_variables_state *state) +{ + struct deref_node *node = get_deref_node(load_instr->variables[0], state); + if (node == NULL) + return; + + if (node->loads == NULL) + node->loads = _mesa_set_create(state->dead_ctx, _mesa_hash_pointer, + _mesa_key_pointer_equal); + + _mesa_set_add(node->loads, load_instr); +} + +static void +register_store_instr(nir_intrinsic_instr *store_instr, + struct lower_variables_state *state) +{ + struct deref_node *node = get_deref_node(store_instr->variables[0], state); + if (node == NULL) + return; + + if (node->stores == NULL) + node->stores = _mesa_set_create(state->dead_ctx, _mesa_hash_pointer, + _mesa_key_pointer_equal); + + _mesa_set_add(node->stores, store_instr); +} + +static void +register_copy_instr(nir_intrinsic_instr *copy_instr, + struct lower_variables_state *state) +{ + for (unsigned idx = 0; idx < 2; idx++) { + struct deref_node *node = + get_deref_node(copy_instr->variables[idx], state); + + if (node == NULL) + continue; + + if (node->copies == NULL) + node->copies = _mesa_set_create(state->dead_ctx, _mesa_hash_pointer, + _mesa_key_pointer_equal); + + _mesa_set_add(node->copies, copy_instr); + } +} + +/* Registers all variable uses in the given block. */ +static bool +register_variable_uses_block(nir_block *block, void *void_state) +{ + struct lower_variables_state *state = void_state; + + nir_foreach_instr_safe(block, instr) { + if (instr->type != nir_instr_type_intrinsic) + continue; + + nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); + + switch (intrin->intrinsic) { + case nir_intrinsic_load_var: + register_load_instr(intrin, state); + break; + + case nir_intrinsic_store_var: + register_store_instr(intrin, state); + break; + + case nir_intrinsic_copy_var: + register_copy_instr(intrin, state); + break; + + default: + continue; + } + } + + return true; +} + +/* Walks over all of the copy instructions to or from the given deref_node + * and lowers them to load/store intrinsics. + */ +static bool +lower_copies_to_load_store(struct deref_node *node, + struct lower_variables_state *state) +{ + if (!node->copies) + return true; + + struct set_entry *copy_entry; + set_foreach(node->copies, copy_entry) { + nir_intrinsic_instr *copy = (void *)copy_entry->key; + + nir_lower_var_copy_instr(copy, state->shader); + + for (unsigned i = 0; i < 2; ++i) { + struct deref_node *arg_node = + get_deref_node(copy->variables[i], state); + + /* Only bother removing copy entries for other nodes */ + if (arg_node == NULL || arg_node == node) + continue; + + struct set_entry *arg_entry = _mesa_set_search(arg_node->copies, copy); + assert(arg_entry); + _mesa_set_remove(node->copies, arg_entry); + } + + nir_instr_remove(©->instr); + } + + node->copies = NULL; + + return true; +} + +/** Pushes an SSA def onto the def stack for the given node + * + * Each node is potentially associated with a stack of SSA definitions. + * This stack is used for determining what SSA definition reaches a given + * point in the program for variable renaming. The stack is always kept in + * dominance-order with at most one SSA def per block. If the SSA + * definition on the top of the stack is in the same block as the one being + * pushed, the top element is replaced. + */ +static void +def_stack_push(struct deref_node *node, nir_ssa_def *def, + struct lower_variables_state *state) +{ + if (node->def_stack == NULL) { + node->def_stack = ralloc_array(state->dead_ctx, nir_ssa_def *, + state->impl->num_blocks); + node->def_stack_tail = node->def_stack - 1; + } + + if (node->def_stack_tail >= node->def_stack) { + nir_ssa_def *top_def = *node->def_stack_tail; + + if (def->parent_instr->block == top_def->parent_instr->block) { + /* They're in the same block, just replace the top */ + *node->def_stack_tail = def; + return; + } + } + + *(++node->def_stack_tail) = def; +} + +/* Pop the top of the def stack if it's in the given block */ +static void +def_stack_pop_if_in_block(struct deref_node *node, nir_block *block) +{ + /* If we're popping, then we have presumably pushed at some time in the + * past so this should exist. + */ + assert(node->def_stack != NULL); + + /* The stack is already empty. Do nothing. */ + if (node->def_stack_tail < node->def_stack) + return; + + nir_ssa_def *def = *node->def_stack_tail; + if (def->parent_instr->block == block) + node->def_stack_tail--; +} + +/** Retrieves the SSA definition on the top of the stack for the given + * node, if one exists. If the stack is empty, then we return the constant + * initializer (if it exists) or an SSA undef. + */ +static nir_ssa_def * +get_ssa_def_for_block(struct deref_node *node, nir_block *block, + struct lower_variables_state *state) +{ + /* If we have something on the stack, go ahead and return it. We're + * assuming that the top of the stack dominates the given block. + */ + if (node->def_stack && node->def_stack_tail >= node->def_stack) + return *node->def_stack_tail; + + /* If we got here then we don't have a definition that dominates the + * given block. This means that we need to add an undef and use that. + */ + nir_ssa_undef_instr *undef = + nir_ssa_undef_instr_create(state->shader, + glsl_get_vector_elements(node->type)); + nir_instr_insert_before_cf_list(&state->impl->body, &undef->instr); + def_stack_push(node, &undef->def, state); + return &undef->def; +} + +/* Given a block and one of its predecessors, this function fills in the + * souces of the phi nodes to take SSA defs from the given predecessor. + * This function must be called exactly once per block/predecessor pair. + */ +static void +add_phi_sources(nir_block *block, nir_block *pred, + struct lower_variables_state *state) +{ + nir_foreach_instr(block, instr) { + if (instr->type != nir_instr_type_phi) + break; + + nir_phi_instr *phi = nir_instr_as_phi(instr); + + struct hash_entry *entry = + _mesa_hash_table_search(state->phi_table, phi); + if (!entry) + continue; + + struct deref_node *node = entry->data; + + nir_phi_src *src = ralloc(phi, nir_phi_src); + src->pred = pred; + src->src.parent_instr = &phi->instr; + src->src.is_ssa = true; + src->src.ssa = get_ssa_def_for_block(node, pred, state); + + list_addtail(&src->src.use_link, &src->src.ssa->uses); + + exec_list_push_tail(&phi->srcs, &src->node); + } +} + +/* Performs variable renaming by doing a DFS of the dominance tree + * + * This algorithm is very similar to the one outlined in "Efficiently + * Computing Static Single Assignment Form and the Control Dependence + * Graph" by Cytron et. al. The primary difference is that we only put one + * SSA def on the stack per block. + */ +static bool +rename_variables_block(nir_block *block, struct lower_variables_state *state) +{ + nir_builder b; + nir_builder_init(&b, state->impl); + + nir_foreach_instr_safe(block, instr) { + if (instr->type == nir_instr_type_phi) { + nir_phi_instr *phi = nir_instr_as_phi(instr); + + struct hash_entry *entry = + _mesa_hash_table_search(state->phi_table, phi); + + /* This can happen if we already have phi nodes in the program + * that were not created in this pass. + */ + if (!entry) + continue; + + struct deref_node *node = entry->data; + + def_stack_push(node, &phi->dest.ssa, state); + } else if (instr->type == nir_instr_type_intrinsic) { + nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); + + switch (intrin->intrinsic) { + case nir_intrinsic_load_var: { + struct deref_node *node = + get_deref_node(intrin->variables[0], state); + + if (node == NULL) { + /* If we hit this path then we are referencing an invalid + * value. Most likely, we unrolled something and are + * reading past the end of some array. In any case, this + * should result in an undefined value. + */ + nir_ssa_undef_instr *undef = + nir_ssa_undef_instr_create(state->shader, + intrin->num_components); + + nir_instr_insert_before(&intrin->instr, &undef->instr); + nir_instr_remove(&intrin->instr); + + nir_ssa_def_rewrite_uses(&intrin->dest.ssa, + nir_src_for_ssa(&undef->def)); + continue; + } + + if (!node->lower_to_ssa) + continue; + + nir_alu_instr *mov = nir_alu_instr_create(state->shader, + nir_op_imov); + mov->src[0].src.is_ssa = true; + mov->src[0].src.ssa = get_ssa_def_for_block(node, block, state); + for (unsigned i = intrin->num_components; i < 4; i++) + mov->src[0].swizzle[i] = 0; + + assert(intrin->dest.is_ssa); + + mov->dest.write_mask = (1 << intrin->num_components) - 1; + nir_ssa_dest_init(&mov->instr, &mov->dest.dest, + intrin->num_components, NULL); + + nir_instr_insert_before(&intrin->instr, &mov->instr); + nir_instr_remove(&intrin->instr); + + nir_ssa_def_rewrite_uses(&intrin->dest.ssa, + nir_src_for_ssa(&mov->dest.dest.ssa)); + break; + } + + case nir_intrinsic_store_var: { + struct deref_node *node = + get_deref_node(intrin->variables[0], state); + + if (node == NULL) { + /* Probably an out-of-bounds array store. That should be a + * no-op. */ + nir_instr_remove(&intrin->instr); + continue; + } + + if (!node->lower_to_ssa) + continue; + + assert(intrin->num_components == + glsl_get_vector_elements(node->type)); + + assert(intrin->src[0].is_ssa); + + nir_ssa_def *new_def; + b.cursor = nir_before_instr(&intrin->instr); + + if (intrin->const_index[0] == (1 << intrin->num_components) - 1) { + /* Whole variable store - just copy the source. Note that + * intrin->num_components and intrin->src[0].ssa->num_components + * may differ. + */ + unsigned swiz[4]; + for (unsigned i = 0; i < 4; i++) + swiz[i] = i < intrin->num_components ? i : 0; + + new_def = nir_swizzle(&b, intrin->src[0].ssa, swiz, + intrin->num_components, false); + } else { + nir_ssa_def *old_def = get_ssa_def_for_block(node, block, state); + /* For writemasked store_var intrinsics, we combine the newly + * written values with the existing contents of unwritten + * channels, creating a new SSA value for the whole vector. + */ + nir_ssa_def *srcs[4]; + for (unsigned i = 0; i < intrin->num_components; i++) { + if (intrin->const_index[0] & (1 << i)) { + srcs[i] = nir_channel(&b, intrin->src[0].ssa, i); + } else { + srcs[i] = nir_channel(&b, old_def, i); + } + } + new_def = nir_vec(&b, srcs, intrin->num_components); + } + + assert(new_def->num_components == intrin->num_components); + + def_stack_push(node, new_def, state); + + /* We'll wait to remove the instruction until the next pass + * where we pop the node we just pushed back off the stack. + */ + break; + } + + default: + break; + } + } + } + + if (block->successors[0]) + add_phi_sources(block->successors[0], block, state); + if (block->successors[1]) + add_phi_sources(block->successors[1], block, state); + + for (unsigned i = 0; i < block->num_dom_children; ++i) + rename_variables_block(block->dom_children[i], state); + + /* Now we iterate over the instructions and pop off any SSA defs that we + * pushed in the first loop. + */ + nir_foreach_instr_safe(block, instr) { + if (instr->type == nir_instr_type_phi) { + nir_phi_instr *phi = nir_instr_as_phi(instr); + + struct hash_entry *entry = + _mesa_hash_table_search(state->phi_table, phi); + + /* This can happen if we already have phi nodes in the program + * that were not created in this pass. + */ + if (!entry) + continue; + + struct deref_node *node = entry->data; + + def_stack_pop_if_in_block(node, block); + } else if (instr->type == nir_instr_type_intrinsic) { + nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); + + if (intrin->intrinsic != nir_intrinsic_store_var) + continue; + + struct deref_node *node = get_deref_node(intrin->variables[0], state); + if (!node) + continue; + + if (!node->lower_to_ssa) + continue; + + def_stack_pop_if_in_block(node, block); + nir_instr_remove(&intrin->instr); + } + } + + return true; +} + +/* Inserts phi nodes for all variables marked lower_to_ssa + * + * This is the same algorithm as presented in "Efficiently Computing Static + * Single Assignment Form and the Control Dependence Graph" by Cytron et. + * al. + */ +static void +insert_phi_nodes(struct lower_variables_state *state) +{ + NIR_VLA_ZERO(unsigned, work, state->impl->num_blocks); + NIR_VLA_ZERO(unsigned, has_already, state->impl->num_blocks); + + /* + * Since the work flags already prevent us from inserting a node that has + * ever been inserted into W, we don't need to use a set to represent W. + * Also, since no block can ever be inserted into W more than once, we know + * that the maximum size of W is the number of basic blocks in the + * function. So all we need to handle W is an array and a pointer to the + * next element to be inserted and the next element to be removed. + */ + NIR_VLA(nir_block *, W, state->impl->num_blocks); + + unsigned w_start, w_end; + unsigned iter_count = 0; + + foreach_list_typed(struct deref_node, node, direct_derefs_link, + &state->direct_deref_nodes) { + if (node->stores == NULL) + continue; + + if (!node->lower_to_ssa) + continue; + + w_start = w_end = 0; + iter_count++; + + struct set_entry *store_entry; + set_foreach(node->stores, store_entry) { + nir_intrinsic_instr *store = (nir_intrinsic_instr *)store_entry->key; + if (work[store->instr.block->index] < iter_count) + W[w_end++] = store->instr.block; + work[store->instr.block->index] = iter_count; + } + + while (w_start != w_end) { + nir_block *cur = W[w_start++]; + struct set_entry *dom_entry; + set_foreach(cur->dom_frontier, dom_entry) { + nir_block *next = (nir_block *) dom_entry->key; + + /* + * If there's more than one return statement, then the end block + * can be a join point for some definitions. However, there are + * no instructions in the end block, so nothing would use those + * phi nodes. Of course, we couldn't place those phi nodes + * anyways due to the restriction of having no instructions in the + * end block... + */ + if (next == state->impl->end_block) + continue; + + if (has_already[next->index] < iter_count) { + nir_phi_instr *phi = nir_phi_instr_create(state->shader); + nir_ssa_dest_init(&phi->instr, &phi->dest, + glsl_get_vector_elements(node->type), NULL); + nir_instr_insert_before_block(next, &phi->instr); + + _mesa_hash_table_insert(state->phi_table, phi, node); + + has_already[next->index] = iter_count; + if (work[next->index] < iter_count) { + work[next->index] = iter_count; + W[w_end++] = next; + } + } + } + } + } +} + + +/** Implements a pass to lower variable uses to SSA values + * + * This path walks the list of instructions and tries to lower as many + * local variable load/store operations to SSA defs and uses as it can. + * The process involves four passes: + * + * 1) Iterate over all of the instructions and mark where each local + * variable deref is used in a load, store, or copy. While we're at + * it, we keep track of all of the fully-qualified (no wildcards) and + * fully-direct references we see and store them in the + * direct_deref_nodes hash table. + * + * 2) Walk over the the list of fully-qualified direct derefs generated in + * the previous pass. For each deref, we determine if it can ever be + * aliased, i.e. if there is an indirect reference anywhere that may + * refer to it. If it cannot be aliased, we mark it for lowering to an + * SSA value. At this point, we lower any var_copy instructions that + * use the given deref to load/store operations and, if the deref has a + * constant initializer, we go ahead and add a load_const value at the + * beginning of the function with the initialized value. + * + * 3) Walk over the list of derefs we plan to lower to SSA values and + * insert phi nodes as needed. + * + * 4) Perform "variable renaming" by replacing the load/store instructions + * with SSA definitions and SSA uses. + */ +static bool +nir_lower_vars_to_ssa_impl(nir_function_impl *impl) +{ + struct lower_variables_state state; + + state.shader = impl->function->shader; + state.dead_ctx = ralloc_context(state.shader); + state.impl = impl; + + state.deref_var_nodes = _mesa_hash_table_create(state.dead_ctx, + _mesa_hash_pointer, + _mesa_key_pointer_equal); + exec_list_make_empty(&state.direct_deref_nodes); + state.phi_table = _mesa_hash_table_create(state.dead_ctx, + _mesa_hash_pointer, + _mesa_key_pointer_equal); + + /* Build the initial deref structures and direct_deref_nodes table */ + state.add_to_direct_deref_nodes = true; + nir_foreach_block(impl, register_variable_uses_block, &state); + + bool progress = false; + + nir_metadata_require(impl, nir_metadata_block_index); + + /* We're about to iterate through direct_deref_nodes. Don't modify it. */ + state.add_to_direct_deref_nodes = false; + + foreach_list_typed_safe(struct deref_node, node, direct_derefs_link, + &state.direct_deref_nodes) { + nir_deref_var *deref = node->deref; + + if (deref->var->data.mode != nir_var_local) { + exec_node_remove(&node->direct_derefs_link); + continue; + } + + if (deref_may_be_aliased(deref, &state)) { + exec_node_remove(&node->direct_derefs_link); + continue; + } + + node->lower_to_ssa = true; + progress = true; + + if (deref->var->constant_initializer) { + nir_load_const_instr *load = + nir_deref_get_const_initializer_load(state.shader, deref); + nir_ssa_def_init(&load->instr, &load->def, + glsl_get_vector_elements(node->type), NULL); + nir_instr_insert_before_cf_list(&impl->body, &load->instr); + def_stack_push(node, &load->def, &state); + } + + foreach_deref_node_match(deref, lower_copies_to_load_store, &state); + } + + if (!progress) + return false; + + nir_metadata_require(impl, nir_metadata_dominance); + + /* We may have lowered some copy instructions to load/store + * instructions. The uses from the copy instructions hav already been + * removed but we need to rescan to ensure that the uses from the newly + * added load/store instructions are registered. We need this + * information for phi node insertion below. + */ + nir_foreach_block(impl, register_variable_uses_block, &state); + + insert_phi_nodes(&state); + rename_variables_block(nir_start_block(impl), &state); + + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance); + + ralloc_free(state.dead_ctx); + + return progress; +} + +void +nir_lower_vars_to_ssa(nir_shader *shader) +{ + nir_foreach_function(shader, function) { + if (function->impl) + nir_lower_vars_to_ssa_impl(function->impl); + } +} diff --git a/src/compiler/nir/nir_lower_vec_to_movs.c b/src/compiler/nir/nir_lower_vec_to_movs.c new file mode 100644 index 0000000..06d6279 --- /dev/null +++ b/src/compiler/nir/nir_lower_vec_to_movs.c @@ -0,0 +1,310 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand (jason@jlekstrand.net) + * + */ + +#include "nir.h" + +/* + * Implements a simple pass that lowers vecN instructions to a series of + * moves with partial writes. + */ + +struct vec_to_movs_state { + nir_function_impl *impl; + bool progress; +}; + +static bool +src_matches_dest_reg(nir_dest *dest, nir_src *src) +{ + if (dest->is_ssa || src->is_ssa) + return false; + + return (dest->reg.reg == src->reg.reg && + dest->reg.base_offset == src->reg.base_offset && + !dest->reg.indirect && + !src->reg.indirect); +} + +/** + * For a given starting writemask channel and corresponding source index in + * the vec instruction, insert a MOV to the vec instruction's dest of all the + * writemask channels that get read from the same src reg. + * + * Returns the writemask of our MOV, so the parent loop calling this knows + * which ones have been processed. + */ +static unsigned +insert_mov(nir_alu_instr *vec, unsigned start_idx, nir_shader *shader) +{ + assert(start_idx < nir_op_infos[vec->op].num_inputs); + + nir_alu_instr *mov = nir_alu_instr_create(shader, nir_op_imov); + nir_alu_src_copy(&mov->src[0], &vec->src[start_idx], mov); + nir_alu_dest_copy(&mov->dest, &vec->dest, mov); + + mov->dest.write_mask = (1u << start_idx); + mov->src[0].swizzle[start_idx] = vec->src[start_idx].swizzle[0]; + mov->src[0].negate = vec->src[start_idx].negate; + mov->src[0].abs = vec->src[start_idx].abs; + + for (unsigned i = start_idx + 1; i < 4; i++) { + if (!(vec->dest.write_mask & (1 << i))) + continue; + + if (nir_srcs_equal(vec->src[i].src, vec->src[start_idx].src) && + vec->src[i].negate == vec->src[start_idx].negate && + vec->src[i].abs == vec->src[start_idx].abs) { + mov->dest.write_mask |= (1 << i); + mov->src[0].swizzle[i] = vec->src[i].swizzle[0]; + } + } + + /* In some situations (if the vecN is involved in a phi-web), we can end + * up with a mov from a register to itself. Some of those channels may end + * up doing nothing and there's no reason to have them as part of the mov. + */ + if (src_matches_dest_reg(&mov->dest.dest, &mov->src[0].src) && + !mov->src[0].abs && !mov->src[0].negate) { + for (unsigned i = 0; i < 4; i++) { + if (mov->src[0].swizzle[i] == i) { + mov->dest.write_mask &= ~(1 << i); + } + } + } + + /* Only emit the instruction if it actually does something */ + if (mov->dest.write_mask) { + nir_instr_insert_before(&vec->instr, &mov->instr); + } else { + ralloc_free(mov); + } + + return mov->dest.write_mask; +} + +static bool +has_replicated_dest(nir_alu_instr *alu) +{ + return alu->op == nir_op_fdot_replicated2 || + alu->op == nir_op_fdot_replicated3 || + alu->op == nir_op_fdot_replicated4 || + alu->op == nir_op_fdph_replicated; +} + +/* Attempts to coalesce the "move" from the given source of the vec to the + * destination of the instruction generating the value. If, for whatever + * reason, we cannot coalesce the mmove, it does nothing and returns 0. We + * can then call insert_mov as normal. + */ +static unsigned +try_coalesce(nir_alu_instr *vec, unsigned start_idx, nir_shader *shader) +{ + assert(start_idx < nir_op_infos[vec->op].num_inputs); + + /* We will only even try if the source is SSA */ + if (!vec->src[start_idx].src.is_ssa) + return 0; + + assert(vec->src[start_idx].src.ssa); + + /* If we are going to do a reswizzle, then the vecN operation must be the + * only use of the source value. We also can't have any source modifiers. + */ + nir_foreach_use(vec->src[start_idx].src.ssa, src) { + if (src->parent_instr != &vec->instr) + return 0; + + nir_alu_src *alu_src = exec_node_data(nir_alu_src, src, src); + if (alu_src->abs || alu_src->negate) + return 0; + } + + if (!list_empty(&vec->src[start_idx].src.ssa->if_uses)) + return 0; + + if (vec->src[start_idx].src.ssa->parent_instr->type != nir_instr_type_alu) + return 0; + + nir_alu_instr *src_alu = + nir_instr_as_alu(vec->src[start_idx].src.ssa->parent_instr); + + if (has_replicated_dest(src_alu)) { + /* The fdot instruction is special: It replicates its result to all + * components. This means that we can always rewrite its destination + * and we don't need to swizzle anything. + */ + } else { + /* We only care about being able to re-swizzle the instruction if it is + * something that we can reswizzle. It must be per-component. The one + * exception to this is the fdotN instructions which implicitly splat + * their result out to all channels. + */ + if (nir_op_infos[src_alu->op].output_size != 0) + return 0; + + /* If we are going to reswizzle the instruction, we can't have any + * non-per-component sources either. + */ + for (unsigned j = 0; j < nir_op_infos[src_alu->op].num_inputs; j++) + if (nir_op_infos[src_alu->op].input_sizes[j] != 0) + return 0; + } + + /* Stash off all of the ALU instruction's swizzles. */ + uint8_t swizzles[4][4]; + for (unsigned j = 0; j < nir_op_infos[src_alu->op].num_inputs; j++) + for (unsigned i = 0; i < 4; i++) + swizzles[j][i] = src_alu->src[j].swizzle[i]; + + unsigned write_mask = 0; + for (unsigned i = start_idx; i < 4; i++) { + if (!(vec->dest.write_mask & (1 << i))) + continue; + + if (!vec->src[i].src.is_ssa || + vec->src[i].src.ssa != &src_alu->dest.dest.ssa) + continue; + + /* At this point, the give vec source matchese up with the ALU + * instruction so we can re-swizzle that component to match. + */ + write_mask |= 1 << i; + if (has_replicated_dest(src_alu)) { + /* Since the destination is a single replicated value, we don't need + * to do any reswizzling + */ + } else { + for (unsigned j = 0; j < nir_op_infos[src_alu->op].num_inputs; j++) + src_alu->src[j].swizzle[i] = swizzles[j][vec->src[i].swizzle[0]]; + } + + /* Clear the no longer needed vec source */ + nir_instr_rewrite_src(&vec->instr, &vec->src[i].src, NIR_SRC_INIT); + } + + nir_instr_rewrite_dest(&src_alu->instr, &src_alu->dest.dest, vec->dest.dest); + src_alu->dest.write_mask = write_mask; + + return write_mask; +} + +static bool +lower_vec_to_movs_block(nir_block *block, void *void_state) +{ + struct vec_to_movs_state *state = void_state; + nir_function_impl *impl = state->impl; + nir_shader *shader = impl->function->shader; + + nir_foreach_instr_safe(block, instr) { + if (instr->type != nir_instr_type_alu) + continue; + + nir_alu_instr *vec = nir_instr_as_alu(instr); + + switch (vec->op) { + case nir_op_vec2: + case nir_op_vec3: + case nir_op_vec4: + break; + default: + continue; /* The loop */ + } + + if (vec->dest.dest.is_ssa) { + /* Since we insert multiple MOVs, we have a register destination. */ + nir_register *reg = nir_local_reg_create(impl); + reg->num_components = vec->dest.dest.ssa.num_components; + + nir_ssa_def_rewrite_uses(&vec->dest.dest.ssa, nir_src_for_reg(reg)); + + nir_instr_rewrite_dest(&vec->instr, &vec->dest.dest, + nir_dest_for_reg(reg)); + } + + unsigned finished_write_mask = 0; + + /* First, emit a MOV for all the src channels that are in the + * destination reg, in case other values we're populating in the dest + * might overwrite them. + */ + for (unsigned i = 0; i < 4; i++) { + if (!(vec->dest.write_mask & (1 << i))) + continue; + + if (src_matches_dest_reg(&vec->dest.dest, &vec->src[i].src)) { + finished_write_mask |= insert_mov(vec, i, shader); + break; + } + } + + /* Now, emit MOVs for all the other src channels. */ + for (unsigned i = 0; i < 4; i++) { + if (!(vec->dest.write_mask & (1 << i))) + continue; + + if (!(finished_write_mask & (1 << i))) + finished_write_mask |= try_coalesce(vec, i, shader); + + if (!(finished_write_mask & (1 << i))) + finished_write_mask |= insert_mov(vec, i, shader); + } + + nir_instr_remove(&vec->instr); + ralloc_free(vec); + state->progress = true; + } + + return true; +} + +static bool +nir_lower_vec_to_movs_impl(nir_function_impl *impl) +{ + struct vec_to_movs_state state = { impl, false }; + + nir_foreach_block(impl, lower_vec_to_movs_block, &state); + + if (state.progress) { + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance); + } + + return state.progress; +} + +bool +nir_lower_vec_to_movs(nir_shader *shader) +{ + bool progress = false; + + nir_foreach_function(shader, function) { + if (function->impl) + progress = nir_lower_vec_to_movs_impl(function->impl) || progress; + } + + return progress; +} diff --git a/src/compiler/nir/nir_metadata.c b/src/compiler/nir/nir_metadata.c new file mode 100644 index 0000000..61aae73 --- /dev/null +++ b/src/compiler/nir/nir_metadata.c @@ -0,0 +1,90 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand (jason@jlekstrand.net) + */ + +#include "nir.h" + +/* + * Handles management of the metadata. + */ + +void +nir_metadata_require(nir_function_impl *impl, nir_metadata required) +{ +#define NEEDS_UPDATE(X) ((required & ~impl->valid_metadata) & (X)) + + if (NEEDS_UPDATE(nir_metadata_block_index)) + nir_index_blocks(impl); + if (NEEDS_UPDATE(nir_metadata_dominance)) + nir_calc_dominance_impl(impl); + if (NEEDS_UPDATE(nir_metadata_live_ssa_defs)) + nir_live_ssa_defs_impl(impl); + +#undef NEEDS_UPDATE + + impl->valid_metadata |= required; +} + +void +nir_metadata_preserve(nir_function_impl *impl, nir_metadata preserved) +{ + impl->valid_metadata &= preserved; +} + +#ifdef DEBUG +/** + * Make sure passes properly invalidate metadata (part 1). + * + * Call this before running a pass to set a bogus metadata flag, which will + * only be preserved if the pass forgets to call nir_metadata_preserve(). + */ +void +nir_metadata_set_validation_flag(nir_shader *shader) +{ + nir_foreach_function(shader, function) { + if (function->impl) { + function->impl->valid_metadata |= nir_metadata_not_properly_reset; + } + } +} + +/** + * Make sure passes properly invalidate metadata (part 2). + * + * Call this after a pass makes progress to verify that the bogus metadata set by + * the earlier function was properly thrown away. Note that passes may not call + * nir_metadata_preserve() if they don't actually make any changes at all. + */ +void +nir_metadata_check_validation_flag(nir_shader *shader) +{ + nir_foreach_function(shader, function) { + if (function->impl) { + assert(!(function->impl->valid_metadata & + nir_metadata_not_properly_reset)); + } + } +} +#endif diff --git a/src/compiler/nir/nir_move_vec_src_uses_to_dest.c b/src/compiler/nir/nir_move_vec_src_uses_to_dest.c new file mode 100644 index 0000000..b5186e6 --- /dev/null +++ b/src/compiler/nir/nir_move_vec_src_uses_to_dest.c @@ -0,0 +1,197 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand (jason@jlekstrand.net) + * + */ + +#include "nir.h" + +/* + * Implements a pass that tries to move uses vecN sources to their + * destinations. This is kind of like an inverse copy-propagation pass. + * For instance, if you have + * + * ssa_1 = vec4(a, b, c, d) + * ssa_2 = fadd(a, b) + * + * This will be turned into + * + * ssa_1 = vec4(a, b, c, d) + * ssa_2 = fadd(ssa_1.x, ssa_1.y) + * + * While this is "worse" because it adds a bunch of unneeded dependencies, it + * actually makes it much easier for vec4-based backends to coalesce the MOV's + * that result from the vec4 operation because it doesn't have to worry about + * quite as many reads. + */ + +/* Returns true if the given SSA def dominates the instruction. An SSA def is + * considered to *not* dominate the instruction that defines it. + */ +static bool +ssa_def_dominates_instr(nir_ssa_def *def, nir_instr *instr) +{ + if (instr->index <= def->parent_instr->index) { + return false; + } else if (def->parent_instr->block == instr->block) { + return def->parent_instr->index < instr->index; + } else { + return nir_block_dominates(def->parent_instr->block, instr->block); + } +} + +static bool +move_vec_src_uses_to_dest_block(nir_block *block, void *shader) +{ + nir_foreach_instr(block, instr) { + if (instr->type != nir_instr_type_alu) + continue; + + nir_alu_instr *vec = nir_instr_as_alu(instr); + + switch (vec->op) { + case nir_op_vec2: + case nir_op_vec3: + case nir_op_vec4: + break; + default: + continue; /* The loop */ + } + + /* Can't handle non-SSA vec operations */ + if (!vec->dest.dest.is_ssa) + continue; + + /* Can't handle saturation */ + if (vec->dest.saturate) + continue; + + /* First, mark all of the sources we are going to consider for rewriting + * to the destination + */ + int srcs_remaining = 0; + for (unsigned i = 0; i < nir_op_infos[vec->op].num_inputs; i++) { + /* We can't rewrite a source if it's not in SSA form */ + if (!vec->src[i].src.is_ssa) + continue; + + /* We can't rewrite a source if it has modifiers */ + if (vec->src[i].abs || vec->src[i].negate) + continue; + + srcs_remaining |= 1 << i; + } + + /* We can't actually do anything with this instruction */ + if (srcs_remaining == 0) + continue; + + for (unsigned i; i = ffs(srcs_remaining) - 1, srcs_remaining;) { + int8_t swizzle[4] = { -1, -1, -1, -1 }; + + for (unsigned j = i; j < nir_op_infos[vec->op].num_inputs; j++) { + if (vec->src[j].src.ssa != vec->src[i].src.ssa) + continue; + + /* Mark the given chanle as having been handled */ + srcs_remaining &= ~(1 << j); + + /* Mark the appropreate channel as coming from src j */ + swizzle[vec->src[j].swizzle[0]] = j; + } + + nir_foreach_use_safe(vec->src[i].src.ssa, use) { + if (use->parent_instr == &vec->instr) + continue; + + /* We need to dominate the use if we are going to rewrite it */ + if (!ssa_def_dominates_instr(&vec->dest.dest.ssa, use->parent_instr)) + continue; + + /* For now, we'll just rewrite ALU instructions */ + if (use->parent_instr->type != nir_instr_type_alu) + continue; + + assert(use->is_ssa); + + nir_alu_instr *use_alu = nir_instr_as_alu(use->parent_instr); + + /* Figure out which source we're actually looking at */ + nir_alu_src *use_alu_src = exec_node_data(nir_alu_src, use, src); + unsigned src_idx = use_alu_src - use_alu->src; + assert(src_idx < nir_op_infos[use_alu->op].num_inputs); + + bool can_reswizzle = true; + for (unsigned j = 0; j < 4; j++) { + if (!nir_alu_instr_channel_used(use_alu, src_idx, j)) + continue; + + if (swizzle[use_alu_src->swizzle[j]] == -1) { + can_reswizzle = false; + break; + } + } + + if (!can_reswizzle) + continue; + + /* At this point, we have determined that the given use can be + * reswizzled to actually use the destination of the vecN operation. + * Go ahead and rewrite it as needed. + */ + nir_instr_rewrite_src(use->parent_instr, use, + nir_src_for_ssa(&vec->dest.dest.ssa)); + for (unsigned j = 0; j < 4; j++) { + if (!nir_alu_instr_channel_used(use_alu, src_idx, j)) + continue; + + use_alu_src->swizzle[j] = swizzle[use_alu_src->swizzle[j]]; + } + } + } + } + + return true; +} + +static void +nir_move_vec_src_uses_to_dest_impl(nir_shader *shader, nir_function_impl *impl) +{ + nir_metadata_require(impl, nir_metadata_dominance); + + nir_index_instrs(impl); + nir_foreach_block(impl, move_vec_src_uses_to_dest_block, shader); + + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance); +} + +void +nir_move_vec_src_uses_to_dest(nir_shader *shader) +{ + nir_foreach_function(shader, function) { + if (function->impl) + nir_move_vec_src_uses_to_dest_impl(shader, function->impl); + } +} diff --git a/src/compiler/nir/nir_normalize_cubemap_coords.c b/src/compiler/nir/nir_normalize_cubemap_coords.c new file mode 100644 index 0000000..9c15eb8 --- /dev/null +++ b/src/compiler/nir/nir_normalize_cubemap_coords.c @@ -0,0 +1,120 @@ +/* + * Copyright © 2015 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand <jason@jlekstrand.net> + */ + +#include "nir.h" +#include "nir_builder.h" + +/** + * This file implements a NIR lowering pass to perform the normalization of + * the cubemap coordinates to have the largest magnitude component be -1.0 + * or 1.0. This is based on the old GLSL IR based pass by Eric. + */ + +struct normalize_cubemap_state { + nir_builder b; + bool progress; +}; + +static bool +normalize_cubemap_coords_block(nir_block *block, void *void_state) +{ + struct normalize_cubemap_state *state = void_state; + nir_builder *b = &state->b; + + nir_foreach_instr(block, instr) { + if (instr->type != nir_instr_type_tex) + continue; + + nir_tex_instr *tex = nir_instr_as_tex(instr); + if (tex->sampler_dim != GLSL_SAMPLER_DIM_CUBE) + continue; + + b->cursor = nir_before_instr(&tex->instr); + + for (unsigned i = 0; i < tex->num_srcs; i++) { + if (tex->src[i].src_type != nir_tex_src_coord) + continue; + + nir_ssa_def *orig_coord = + nir_ssa_for_src(b, tex->src[i].src, nir_tex_instr_src_size(tex, i)); + assert(orig_coord->num_components >= 3); + + nir_ssa_def *abs = nir_fabs(b, orig_coord); + nir_ssa_def *norm = nir_fmax(b, nir_channel(b, abs, 0), + nir_fmax(b, nir_channel(b, abs, 1), + nir_channel(b, abs, 2))); + + nir_ssa_def *normalized = nir_fmul(b, orig_coord, nir_frcp(b, norm)); + + /* Array indices don't have to be normalized, so make a new vector + * with the coordinate's array index untouched. + */ + if (tex->coord_components == 4) { + normalized = nir_vec4(b, + nir_channel(b, normalized, 0), + nir_channel(b, normalized, 1), + nir_channel(b, normalized, 2), + nir_channel(b, orig_coord, 3)); + } + + nir_instr_rewrite_src(&tex->instr, + &tex->src[i].src, + nir_src_for_ssa(normalized)); + + state->progress = true; + } + } + + return true; +} + +static bool +normalize_cubemap_coords_impl(nir_function_impl *impl) +{ + struct normalize_cubemap_state state; + nir_builder_init(&state.b, impl); + state.progress = false; + + nir_foreach_block(impl, normalize_cubemap_coords_block, &state); + + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance); + + return state.progress; +} + +bool +nir_normalize_cubemap_coords(nir_shader *shader) +{ + bool progress = false; + + nir_foreach_function(shader, function) { + if (function->impl) + progress = normalize_cubemap_coords_impl(function->impl) || progress; + } + + return progress; +} diff --git a/src/compiler/nir/nir_opcodes.py b/src/compiler/nir/nir_opcodes.py new file mode 100644 index 0000000..e79810c --- /dev/null +++ b/src/compiler/nir/nir_opcodes.py @@ -0,0 +1,668 @@ +#! /usr/bin/env python +# +# Copyright (C) 2014 Connor Abbott +# +# Permission is hereby granted, free of charge, to any person obtaining a +# copy of this software and associated documentation files (the "Software"), +# to deal in the Software without restriction, including without limitation +# the rights to use, copy, modify, merge, publish, distribute, sublicense, +# and/or sell copies of the Software, and to permit persons to whom the +# Software is furnished to do so, subject to the following conditions: +# +# The above copyright notice and this permission notice (including the next +# paragraph) shall be included in all copies or substantial portions of the +# Software. +# +# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL +# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +# IN THE SOFTWARE. +# +# Authors: +# Connor Abbott (cwabbott0@gmail.com) + + +# Class that represents all the information we have about the opcode +# NOTE: this must be kept in sync with nir_op_info + +class Opcode(object): + """Class that represents all the information we have about the opcode + NOTE: this must be kept in sync with nir_op_info + """ + def __init__(self, name, output_size, output_type, input_sizes, + input_types, algebraic_properties, const_expr): + """Parameters: + + - name is the name of the opcode (prepend nir_op_ for the enum name) + - all types are strings that get nir_type_ prepended to them + - input_types is a list of types + - algebraic_properties is a space-seperated string, where nir_op_is_ is + prepended before each entry + - const_expr is an expression or series of statements that computes the + constant value of the opcode given the constant values of its inputs. + + Constant expressions are formed from the variables src0, src1, ..., + src(N-1), where N is the number of arguments. The output of the + expression should be stored in the dst variable. Per-component input + and output variables will be scalars and non-per-component input and + output variables will be a struct with fields named x, y, z, and w + all of the correct type. Input and output variables can be assumed + to already be of the correct type and need no conversion. In + particular, the conversion from the C bool type to/from NIR_TRUE and + NIR_FALSE happens automatically. + + For per-component instructions, the entire expression will be + executed once for each component. For non-per-component + instructions, the expression is expected to store the correct values + in dst.x, dst.y, etc. If "dst" does not exist anywhere in the + constant expression, an assignment to dst will happen automatically + and the result will be equivalent to "dst = <expression>" for + per-component instructions and "dst.x = dst.y = ... = <expression>" + for non-per-component instructions. + """ + assert isinstance(name, str) + assert isinstance(output_size, int) + assert isinstance(output_type, str) + assert isinstance(input_sizes, list) + assert isinstance(input_sizes[0], int) + assert isinstance(input_types, list) + assert isinstance(input_types[0], str) + assert isinstance(algebraic_properties, str) + assert isinstance(const_expr, str) + assert len(input_sizes) == len(input_types) + assert 0 <= output_size <= 4 + for size in input_sizes: + assert 0 <= size <= 4 + if output_size != 0: + assert size != 0 + self.name = name + self.num_inputs = len(input_sizes) + self.output_size = output_size + self.output_type = output_type + self.input_sizes = input_sizes + self.input_types = input_types + self.algebraic_properties = algebraic_properties + self.const_expr = const_expr + +# helper variables for strings +tfloat = "float" +tint = "int" +tbool = "bool" +tuint = "uint" + +commutative = "commutative " +associative = "associative " + +# global dictionary of opcodes +opcodes = {} + +def opcode(name, output_size, output_type, input_sizes, input_types, + algebraic_properties, const_expr): + assert name not in opcodes + opcodes[name] = Opcode(name, output_size, output_type, input_sizes, + input_types, algebraic_properties, const_expr) + +def unop_convert(name, in_type, out_type, const_expr): + opcode(name, 0, out_type, [0], [in_type], "", const_expr) + +def unop(name, ty, const_expr): + opcode(name, 0, ty, [0], [ty], "", const_expr) + +def unop_horiz(name, output_size, output_type, input_size, input_type, + const_expr): + opcode(name, output_size, output_type, [input_size], [input_type], "", + const_expr) + +def unop_reduce(name, output_size, output_type, input_type, prereduce_expr, + reduce_expr, final_expr): + def prereduce(src): + return "(" + prereduce_expr.format(src=src) + ")" + def final(src): + return final_expr.format(src="(" + src + ")") + def reduce_(src0, src1): + return reduce_expr.format(src0=src0, src1=src1) + src0 = prereduce("src0.x") + src1 = prereduce("src0.y") + src2 = prereduce("src0.z") + src3 = prereduce("src0.w") + unop_horiz(name + "2", output_size, output_type, 2, input_type, + final(reduce_(src0, src1))) + unop_horiz(name + "3", output_size, output_type, 3, input_type, + final(reduce_(reduce_(src0, src1), src2))) + unop_horiz(name + "4", output_size, output_type, 4, input_type, + final(reduce_(reduce_(src0, src1), reduce_(src2, src3)))) + + +# These two move instructions differ in what modifiers they support and what +# the negate modifier means. Otherwise, they are identical. +unop("fmov", tfloat, "src0") +unop("imov", tint, "src0") + +unop("ineg", tint, "-src0") +unop("fneg", tfloat, "-src0") +unop("inot", tint, "~src0") # invert every bit of the integer +unop("fnot", tfloat, "(src0 == 0.0f) ? 1.0f : 0.0f") +unop("fsign", tfloat, "(src0 == 0.0f) ? 0.0f : ((src0 > 0.0f) ? 1.0f : -1.0f)") +unop("isign", tint, "(src0 == 0) ? 0 : ((src0 > 0) ? 1 : -1)") +unop("iabs", tint, "(src0 < 0) ? -src0 : src0") +unop("fabs", tfloat, "fabsf(src0)") +unop("fsat", tfloat, "(src0 > 1.0f) ? 1.0f : ((src0 <= 0.0f) ? 0.0f : src0)") +unop("frcp", tfloat, "1.0f / src0") +unop("frsq", tfloat, "1.0f / sqrtf(src0)") +unop("fsqrt", tfloat, "sqrtf(src0)") +unop("fexp2", tfloat, "exp2f(src0)") +unop("flog2", tfloat, "log2f(src0)") +unop_convert("f2i", tfloat, tint, "src0") # Float-to-integer conversion. +unop_convert("f2u", tfloat, tuint, "src0") # Float-to-unsigned conversion +unop_convert("i2f", tint, tfloat, "src0") # Integer-to-float conversion. +# Float-to-boolean conversion +unop_convert("f2b", tfloat, tbool, "src0 != 0.0f") +# Boolean-to-float conversion +unop_convert("b2f", tbool, tfloat, "src0 ? 1.0f : 0.0f") +# Int-to-boolean conversion +unop_convert("i2b", tint, tbool, "src0 != 0") +unop_convert("b2i", tbool, tint, "src0 ? 1 : 0") # Boolean-to-int conversion +unop_convert("u2f", tuint, tfloat, "src0") # Unsigned-to-float conversion. + +# Unary floating-point rounding operations. + + +unop("ftrunc", tfloat, "truncf(src0)") +unop("fceil", tfloat, "ceilf(src0)") +unop("ffloor", tfloat, "floorf(src0)") +unop("ffract", tfloat, "src0 - floorf(src0)") +unop("fround_even", tfloat, "_mesa_roundevenf(src0)") + + +# Trigonometric operations. + + +unop("fsin", tfloat, "sinf(src0)") +unop("fcos", tfloat, "cosf(src0)") + + +# Partial derivatives. + + +unop("fddx", tfloat, "0.0f") # the derivative of a constant is 0. +unop("fddy", tfloat, "0.0f") +unop("fddx_fine", tfloat, "0.0f") +unop("fddy_fine", tfloat, "0.0f") +unop("fddx_coarse", tfloat, "0.0f") +unop("fddy_coarse", tfloat, "0.0f") + + +# Floating point pack and unpack operations. + +def pack_2x16(fmt): + unop_horiz("pack_" + fmt + "_2x16", 1, tuint, 2, tfloat, """ +dst.x = (uint32_t) pack_fmt_1x16(src0.x); +dst.x |= ((uint32_t) pack_fmt_1x16(src0.y)) << 16; +""".replace("fmt", fmt)) + +def pack_4x8(fmt): + unop_horiz("pack_" + fmt + "_4x8", 1, tuint, 4, tfloat, """ +dst.x = (uint32_t) pack_fmt_1x8(src0.x); +dst.x |= ((uint32_t) pack_fmt_1x8(src0.y)) << 8; +dst.x |= ((uint32_t) pack_fmt_1x8(src0.z)) << 16; +dst.x |= ((uint32_t) pack_fmt_1x8(src0.w)) << 24; +""".replace("fmt", fmt)) + +def unpack_2x16(fmt): + unop_horiz("unpack_" + fmt + "_2x16", 2, tfloat, 1, tuint, """ +dst.x = unpack_fmt_1x16((uint16_t)(src0.x & 0xffff)); +dst.y = unpack_fmt_1x16((uint16_t)(src0.x << 16)); +""".replace("fmt", fmt)) + +def unpack_4x8(fmt): + unop_horiz("unpack_" + fmt + "_4x8", 4, tfloat, 1, tuint, """ +dst.x = unpack_fmt_1x8((uint8_t)(src0.x & 0xff)); +dst.y = unpack_fmt_1x8((uint8_t)((src0.x >> 8) & 0xff)); +dst.z = unpack_fmt_1x8((uint8_t)((src0.x >> 16) & 0xff)); +dst.w = unpack_fmt_1x8((uint8_t)(src0.x >> 24)); +""".replace("fmt", fmt)) + + +pack_2x16("snorm") +pack_4x8("snorm") +pack_2x16("unorm") +pack_4x8("unorm") +pack_2x16("half") +unpack_2x16("snorm") +unpack_4x8("snorm") +unpack_2x16("unorm") +unpack_4x8("unorm") +unpack_2x16("half") + + +# Lowered floating point unpacking operations. + + +unop_horiz("unpack_half_2x16_split_x", 1, tfloat, 1, tuint, + "unpack_half_1x16((uint16_t)(src0.x & 0xffff))") +unop_horiz("unpack_half_2x16_split_y", 1, tfloat, 1, tuint, + "unpack_half_1x16((uint16_t)(src0.x >> 16))") + + +# Bit operations, part of ARB_gpu_shader5. + + +unop("bitfield_reverse", tuint, """ +/* we're not winning any awards for speed here, but that's ok */ +dst = 0; +for (unsigned bit = 0; bit < 32; bit++) + dst |= ((src0 >> bit) & 1) << (31 - bit); +""") +unop("bit_count", tuint, """ +dst = 0; +for (unsigned bit = 0; bit < 32; bit++) { + if ((src0 >> bit) & 1) + dst++; +} +""") + +unop_convert("ufind_msb", tuint, tint, """ +dst = -1; +for (int bit = 31; bit > 0; bit--) { + if ((src0 >> bit) & 1) { + dst = bit; + break; + } +} +""") + +unop("ifind_msb", tint, """ +dst = -1; +for (int bit = 31; bit >= 0; bit--) { + /* If src0 < 0, we're looking for the first 0 bit. + * if src0 >= 0, we're looking for the first 1 bit. + */ + if ((((src0 >> bit) & 1) && (src0 >= 0)) || + (!((src0 >> bit) & 1) && (src0 < 0))) { + dst = bit; + break; + } +} +""") + +unop("find_lsb", tint, """ +dst = -1; +for (unsigned bit = 0; bit < 32; bit++) { + if ((src0 >> bit) & 1) { + dst = bit; + break; + } +} +""") + + +for i in xrange(1, 5): + for j in xrange(1, 5): + unop_horiz("fnoise{0}_{1}".format(i, j), i, tfloat, j, tfloat, "0.0f") + +def binop_convert(name, out_type, in_type, alg_props, const_expr): + opcode(name, 0, out_type, [0, 0], [in_type, in_type], alg_props, const_expr) + +def binop(name, ty, alg_props, const_expr): + binop_convert(name, ty, ty, alg_props, const_expr) + +def binop_compare(name, ty, alg_props, const_expr): + binop_convert(name, tbool, ty, alg_props, const_expr) + +def binop_horiz(name, out_size, out_type, src1_size, src1_type, src2_size, + src2_type, const_expr): + opcode(name, out_size, out_type, [src1_size, src2_size], [src1_type, src2_type], + "", const_expr) + +def binop_reduce(name, output_size, output_type, src_type, prereduce_expr, + reduce_expr, final_expr): + def final(src): + return final_expr.format(src= "(" + src + ")") + def reduce_(src0, src1): + return reduce_expr.format(src0=src0, src1=src1) + def prereduce(src0, src1): + return "(" + prereduce_expr.format(src0=src0, src1=src1) + ")" + src0 = prereduce("src0.x", "src1.x") + src1 = prereduce("src0.y", "src1.y") + src2 = prereduce("src0.z", "src1.z") + src3 = prereduce("src0.w", "src1.w") + opcode(name + "2", output_size, output_type, + [2, 2], [src_type, src_type], commutative, + final(reduce_(src0, src1))) + opcode(name + "3", output_size, output_type, + [3, 3], [src_type, src_type], commutative, + final(reduce_(reduce_(src0, src1), src2))) + opcode(name + "4", output_size, output_type, + [4, 4], [src_type, src_type], commutative, + final(reduce_(reduce_(src0, src1), reduce_(src2, src3)))) + +binop("fadd", tfloat, commutative + associative, "src0 + src1") +binop("iadd", tint, commutative + associative, "src0 + src1") +binop("fsub", tfloat, "", "src0 - src1") +binop("isub", tint, "", "src0 - src1") + +binop("fmul", tfloat, commutative + associative, "src0 * src1") +# low 32-bits of signed/unsigned integer multiply +binop("imul", tint, commutative + associative, "src0 * src1") +# high 32-bits of signed integer multiply +binop("imul_high", tint, commutative, + "(int32_t)(((int64_t) src0 * (int64_t) src1) >> 32)") +# high 32-bits of unsigned integer multiply +binop("umul_high", tuint, commutative, + "(uint32_t)(((uint64_t) src0 * (uint64_t) src1) >> 32)") + +binop("fdiv", tfloat, "", "src0 / src1") +binop("idiv", tint, "", "src0 / src1") +binop("udiv", tuint, "", "src0 / src1") + +# returns a boolean representing the carry resulting from the addition of +# the two unsigned arguments. + +binop_convert("uadd_carry", tuint, tuint, commutative, "src0 + src1 < src0") + +# returns a boolean representing the borrow resulting from the subtraction +# of the two unsigned arguments. + +binop_convert("usub_borrow", tuint, tuint, "", "src0 < src1") + +binop("fmod", tfloat, "", "src0 - src1 * floorf(src0 / src1)") +binop("umod", tuint, "", "src1 == 0 ? 0 : src0 % src1") + +# +# Comparisons +# + + +# these integer-aware comparisons return a boolean (0 or ~0) + +binop_compare("flt", tfloat, "", "src0 < src1") +binop_compare("fge", tfloat, "", "src0 >= src1") +binop_compare("feq", tfloat, commutative, "src0 == src1") +binop_compare("fne", tfloat, commutative, "src0 != src1") +binop_compare("ilt", tint, "", "src0 < src1") +binop_compare("ige", tint, "", "src0 >= src1") +binop_compare("ieq", tint, commutative, "src0 == src1") +binop_compare("ine", tint, commutative, "src0 != src1") +binop_compare("ult", tuint, "", "src0 < src1") +binop_compare("uge", tuint, "", "src0 >= src1") + +# integer-aware GLSL-style comparisons that compare floats and ints + +binop_reduce("ball_fequal", 1, tbool, tfloat, "{src0} == {src1}", + "{src0} && {src1}", "{src}") +binop_reduce("bany_fnequal", 1, tbool, tfloat, "{src0} != {src1}", + "{src0} || {src1}", "{src}") +binop_reduce("ball_iequal", 1, tbool, tint, "{src0} == {src1}", + "{src0} && {src1}", "{src}") +binop_reduce("bany_inequal", 1, tbool, tint, "{src0} != {src1}", + "{src0} || {src1}", "{src}") + +# non-integer-aware GLSL-style comparisons that return 0.0 or 1.0 + +binop_reduce("fall_equal", 1, tfloat, tfloat, "{src0} == {src1}", + "{src0} && {src1}", "{src} ? 1.0f : 0.0f") +binop_reduce("fany_nequal", 1, tfloat, tfloat, "{src0} != {src1}", + "{src0} || {src1}", "{src} ? 1.0f : 0.0f") + +# These comparisons for integer-less hardware return 1.0 and 0.0 for true +# and false respectively + +binop("slt", tfloat, "", "(src0 < src1) ? 1.0f : 0.0f") # Set on Less Than +binop("sge", tfloat, "", "(src0 >= src1) ? 1.0f : 0.0f") # Set on Greater or Equal +binop("seq", tfloat, commutative, "(src0 == src1) ? 1.0f : 0.0f") # Set on Equal +binop("sne", tfloat, commutative, "(src0 != src1) ? 1.0f : 0.0f") # Set on Not Equal + + +binop("ishl", tint, "", "src0 << src1") +binop("ishr", tint, "", "src0 >> src1") +binop("ushr", tuint, "", "src0 >> src1") + +# bitwise logic operators +# +# These are also used as boolean and, or, xor for hardware supporting +# integers. + + +binop("iand", tuint, commutative + associative, "src0 & src1") +binop("ior", tuint, commutative + associative, "src0 | src1") +binop("ixor", tuint, commutative + associative, "src0 ^ src1") + + +# floating point logic operators +# +# These use (src != 0.0) for testing the truth of the input, and output 1.0 +# for true and 0.0 for false + +binop("fand", tfloat, commutative, + "((src0 != 0.0f) && (src1 != 0.0f)) ? 1.0f : 0.0f") +binop("for", tfloat, commutative, + "((src0 != 0.0f) || (src1 != 0.0f)) ? 1.0f : 0.0f") +binop("fxor", tfloat, commutative, + "(src0 != 0.0f && src1 == 0.0f) || (src0 == 0.0f && src1 != 0.0f) ? 1.0f : 0.0f") + +binop_reduce("fdot", 1, tfloat, tfloat, "{src0} * {src1}", "{src0} + {src1}", + "{src}") + +binop_reduce("fdot_replicated", 4, tfloat, tfloat, + "{src0} * {src1}", "{src0} + {src1}", "{src}") + +opcode("fdph", 1, tfloat, [3, 4], [tfloat, tfloat], "", + "src0.x * src1.x + src0.y * src1.y + src0.z * src1.z + src1.w") +opcode("fdph_replicated", 4, tfloat, [3, 4], [tfloat, tfloat], "", + "src0.x * src1.x + src0.y * src1.y + src0.z * src1.z + src1.w") + +binop("fmin", tfloat, "", "fminf(src0, src1)") +binop("imin", tint, commutative + associative, "src1 > src0 ? src0 : src1") +binop("umin", tuint, commutative + associative, "src1 > src0 ? src0 : src1") +binop("fmax", tfloat, "", "fmaxf(src0, src1)") +binop("imax", tint, commutative + associative, "src1 > src0 ? src1 : src0") +binop("umax", tuint, commutative + associative, "src1 > src0 ? src1 : src0") + +# Saturated vector add for 4 8bit ints. +binop("usadd_4x8", tint, commutative + associative, """ +dst = 0; +for (int i = 0; i < 32; i += 8) { + dst |= MIN2(((src0 >> i) & 0xff) + ((src1 >> i) & 0xff), 0xff) << i; +} +""") + +# Saturated vector subtract for 4 8bit ints. +binop("ussub_4x8", tint, "", """ +dst = 0; +for (int i = 0; i < 32; i += 8) { + int src0_chan = (src0 >> i) & 0xff; + int src1_chan = (src1 >> i) & 0xff; + if (src0_chan > src1_chan) + dst |= (src0_chan - src1_chan) << i; +} +""") + +# vector min for 4 8bit ints. +binop("umin_4x8", tint, commutative + associative, """ +dst = 0; +for (int i = 0; i < 32; i += 8) { + dst |= MIN2((src0 >> i) & 0xff, (src1 >> i) & 0xff) << i; +} +""") + +# vector max for 4 8bit ints. +binop("umax_4x8", tint, commutative + associative, """ +dst = 0; +for (int i = 0; i < 32; i += 8) { + dst |= MAX2((src0 >> i) & 0xff, (src1 >> i) & 0xff) << i; +} +""") + +# unorm multiply: (a * b) / 255. +binop("umul_unorm_4x8", tint, commutative + associative, """ +dst = 0; +for (int i = 0; i < 32; i += 8) { + int src0_chan = (src0 >> i) & 0xff; + int src1_chan = (src1 >> i) & 0xff; + dst |= ((src0_chan * src1_chan) / 255) << i; +} +""") + +binop("fpow", tfloat, "", "powf(src0, src1)") + +binop_horiz("pack_half_2x16_split", 1, tuint, 1, tfloat, 1, tfloat, + "pack_half_1x16(src0.x) | (pack_half_1x16(src1.x) << 16)") + +# bfm implements the behavior of the first operation of the SM5 "bfi" assembly +# and that of the "bfi1" i965 instruction. That is, it has undefined behavior +# if either of its arguments are 32. +binop_convert("bfm", tuint, tint, "", """ +int bits = src0, offset = src1; +if (offset < 0 || bits < 0 || offset > 31 || bits > 31 || offset + bits > 32) + dst = 0; /* undefined */ +else + dst = ((1u << bits) - 1) << offset; +""") + +opcode("ldexp", 0, tfloat, [0, 0], [tfloat, tint], "", """ +dst = ldexpf(src0, src1); +/* flush denormals to zero. */ +if (!isnormal(dst)) + dst = copysignf(0.0f, src0); +""") + +# Combines the first component of each input to make a 2-component vector. + +binop_horiz("vec2", 2, tuint, 1, tuint, 1, tuint, """ +dst.x = src0.x; +dst.y = src1.x; +""") + +def triop(name, ty, const_expr): + opcode(name, 0, ty, [0, 0, 0], [ty, ty, ty], "", const_expr) +def triop_horiz(name, output_size, src1_size, src2_size, src3_size, const_expr): + opcode(name, output_size, tuint, + [src1_size, src2_size, src3_size], + [tuint, tuint, tuint], "", const_expr) + +triop("ffma", tfloat, "src0 * src1 + src2") + +triop("flrp", tfloat, "src0 * (1 - src2) + src1 * src2") + +# Conditional Select +# +# A vector conditional select instruction (like ?:, but operating per- +# component on vectors). There are two versions, one for floating point +# bools (0.0 vs 1.0) and one for integer bools (0 vs ~0). + + +triop("fcsel", tfloat, "(src0 != 0.0f) ? src1 : src2") +opcode("bcsel", 0, tuint, [0, 0, 0], + [tbool, tuint, tuint], "", "src0 ? src1 : src2") + +# SM5 bfi assembly +triop("bfi", tuint, """ +unsigned mask = src0, insert = src1, base = src2; +if (mask == 0) { + dst = base; +} else { + unsigned tmp = mask; + while (!(tmp & 1)) { + tmp >>= 1; + insert <<= 1; + } + dst = (base & ~mask) | (insert & mask); +} +""") + +# SM5 ubfe/ibfe assembly +opcode("ubfe", 0, tuint, + [0, 0, 0], [tuint, tint, tint], "", """ +unsigned base = src0; +int offset = src1, bits = src2; +if (bits == 0) { + dst = 0; +} else if (bits < 0 || offset < 0) { + dst = 0; /* undefined */ +} else if (offset + bits < 32) { + dst = (base << (32 - bits - offset)) >> (32 - bits); +} else { + dst = base >> offset; +} +""") +opcode("ibfe", 0, tint, + [0, 0, 0], [tint, tint, tint], "", """ +int base = src0; +int offset = src1, bits = src2; +if (bits == 0) { + dst = 0; +} else if (bits < 0 || offset < 0) { + dst = 0; /* undefined */ +} else if (offset + bits < 32) { + dst = (base << (32 - bits - offset)) >> (32 - bits); +} else { + dst = base >> offset; +} +""") + +# GLSL bitfieldExtract() +opcode("ubitfield_extract", 0, tuint, + [0, 0, 0], [tuint, tint, tint], "", """ +unsigned base = src0; +int offset = src1, bits = src2; +if (bits == 0) { + dst = 0; +} else if (bits < 0 || offset < 0 || offset + bits > 32) { + dst = 0; /* undefined per the spec */ +} else { + dst = (base >> offset) & ((1ull << bits) - 1); +} +""") +opcode("ibitfield_extract", 0, tint, + [0, 0, 0], [tint, tint, tint], "", """ +int base = src0; +int offset = src1, bits = src2; +if (bits == 0) { + dst = 0; +} else if (offset < 0 || bits < 0 || offset + bits > 32) { + dst = 0; +} else { + dst = (base << (32 - offset - bits)) >> offset; /* use sign-extending shift */ +} +""") + +# Combines the first component of each input to make a 3-component vector. + +triop_horiz("vec3", 3, 1, 1, 1, """ +dst.x = src0.x; +dst.y = src1.x; +dst.z = src2.x; +""") + +def quadop_horiz(name, output_size, src1_size, src2_size, src3_size, + src4_size, const_expr): + opcode(name, output_size, tuint, + [src1_size, src2_size, src3_size, src4_size], + [tuint, tuint, tuint, tuint], + "", const_expr) + +opcode("bitfield_insert", 0, tuint, [0, 0, 0, 0], + [tuint, tuint, tint, tint], "", """ +unsigned base = src0, insert = src1; +int offset = src2, bits = src3; +if (bits == 0) { + dst = 0; +} else if (offset < 0 || bits < 0 || bits + offset > 32) { + dst = 0; +} else { + unsigned mask = ((1ull << bits) - 1) << offset; + dst = (base & ~mask) | ((insert << bits) & mask); +} +""") + +quadop_horiz("vec4", 4, 1, 1, 1, 1, """ +dst.x = src0.x; +dst.y = src1.x; +dst.z = src2.x; +dst.w = src3.x; +""") + + diff --git a/src/compiler/nir/nir_opcodes_c.py b/src/compiler/nir/nir_opcodes_c.py new file mode 100644 index 0000000..7049c5b --- /dev/null +++ b/src/compiler/nir/nir_opcodes_c.py @@ -0,0 +1,55 @@ +#! /usr/bin/env python +# +# Copyright (C) 2014 Connor Abbott +# +# Permission is hereby granted, free of charge, to any person obtaining a +# copy of this software and associated documentation files (the "Software"), +# to deal in the Software without restriction, including without limitation +# the rights to use, copy, modify, merge, publish, distribute, sublicense, +# and/or sell copies of the Software, and to permit persons to whom the +# Software is furnished to do so, subject to the following conditions: +# +# The above copyright notice and this permission notice (including the next +# paragraph) shall be included in all copies or substantial portions of the +# Software. +# +# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL +# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +# IN THE SOFTWARE. +# +# Authors: +# Connor Abbott (cwabbott0@gmail.com) + +from nir_opcodes import opcodes +from mako.template import Template + +template = Template(""" +#include "nir.h" + +const nir_op_info nir_op_infos[nir_num_opcodes] = { +% for name, opcode in sorted(opcodes.iteritems()): +{ + .name = "${name}", + .num_inputs = ${opcode.num_inputs}, + .output_size = ${opcode.output_size}, + .output_type = ${"nir_type_" + opcode.output_type}, + .input_sizes = { + ${ ", ".join(str(size) for size in opcode.input_sizes) } + }, + .input_types = { + ${ ", ".join("nir_type_" + type for type in opcode.input_types) } + }, + .algebraic_properties = + ${ "0" if opcode.algebraic_properties == "" else " | ".join( + "NIR_OP_IS_" + prop.upper() for prop in + opcode.algebraic_properties.strip().split(" ")) } +}, +% endfor +}; +""") + +print template.render(opcodes=opcodes) diff --git a/src/compiler/nir/nir_opcodes_h.py b/src/compiler/nir/nir_opcodes_h.py new file mode 100644 index 0000000..be15a96 --- /dev/null +++ b/src/compiler/nir/nir_opcodes_h.py @@ -0,0 +1,47 @@ +#! /usr/bin/env python + +template = """\ +/* Copyright (C) 2014 Connor Abbott + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + */ + +#ifndef _NIR_OPCODES_ +#define _NIR_OPCODES_ + +<% opcode_names = sorted(opcodes.iterkeys()) %> + +typedef enum { +% for name in opcode_names: + nir_op_${name}, +% endfor + nir_last_opcode = nir_op_${opcode_names[-1]}, + nir_num_opcodes = nir_last_opcode + 1 +} nir_op; + +#endif /* _NIR_OPCODES_ */""" + +from nir_opcodes import opcodes +from mako.template import Template + +print Template(template).render(opcodes=opcodes) diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py new file mode 100644 index 0000000..7745b76 --- /dev/null +++ b/src/compiler/nir/nir_opt_algebraic.py @@ -0,0 +1,285 @@ +#! /usr/bin/env python +# +# Copyright (C) 2014 Intel Corporation +# +# Permission is hereby granted, free of charge, to any person obtaining a +# copy of this software and associated documentation files (the "Software"), +# to deal in the Software without restriction, including without limitation +# the rights to use, copy, modify, merge, publish, distribute, sublicense, +# and/or sell copies of the Software, and to permit persons to whom the +# Software is furnished to do so, subject to the following conditions: +# +# The above copyright notice and this permission notice (including the next +# paragraph) shall be included in all copies or substantial portions of the +# Software. +# +# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL +# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING +# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS +# IN THE SOFTWARE. +# +# Authors: +# Jason Ekstrand (jason@jlekstrand.net) + +import nir_algebraic + +# Convenience variables +a = 'a' +b = 'b' +c = 'c' +d = 'd' + +# Written in the form (<search>, <replace>) where <search> is an expression +# and <replace> is either an expression or a value. An expression is +# defined as a tuple of the form (<op>, <src0>, <src1>, <src2>, <src3>) +# where each source is either an expression or a value. A value can be +# either a numeric constant or a string representing a variable name. +# +# Variable names are specified as "[#]name[@type]" where "#" inicates that +# the given variable will only match constants and the type indicates that +# the given variable will only match values from ALU instructions with the +# given output type. +# +# For constants, you have to be careful to make sure that it is the right +# type because python is unaware of the source and destination types of the +# opcodes. + +optimizations = [ + (('fneg', ('fneg', a)), a), + (('ineg', ('ineg', a)), a), + (('fabs', ('fabs', a)), ('fabs', a)), + (('fabs', ('fneg', a)), ('fabs', a)), + (('iabs', ('iabs', a)), ('iabs', a)), + (('iabs', ('ineg', a)), ('iabs', a)), + (('fadd', a, 0.0), a), + (('iadd', a, 0), a), + (('usadd_4x8', a, 0), a), + (('usadd_4x8', a, ~0), ~0), + (('fadd', ('fmul', a, b), ('fmul', a, c)), ('fmul', a, ('fadd', b, c))), + (('iadd', ('imul', a, b), ('imul', a, c)), ('imul', a, ('iadd', b, c))), + (('fadd', ('fneg', a), a), 0.0), + (('iadd', ('ineg', a), a), 0), + (('iadd', ('ineg', a), ('iadd', a, b)), b), + (('iadd', a, ('iadd', ('ineg', a), b)), b), + (('fadd', ('fneg', a), ('fadd', a, b)), b), + (('fadd', a, ('fadd', ('fneg', a), b)), b), + (('fmul', a, 0.0), 0.0), + (('imul', a, 0), 0), + (('umul_unorm_4x8', a, 0), 0), + (('umul_unorm_4x8', a, ~0), a), + (('fmul', a, 1.0), a), + (('imul', a, 1), a), + (('fmul', a, -1.0), ('fneg', a)), + (('imul', a, -1), ('ineg', a)), + (('ffma', 0.0, a, b), b), + (('ffma', a, 0.0, b), b), + (('ffma', a, b, 0.0), ('fmul', a, b)), + (('ffma', a, 1.0, b), ('fadd', a, b)), + (('ffma', 1.0, a, b), ('fadd', a, b)), + (('flrp', a, b, 0.0), a), + (('flrp', a, b, 1.0), b), + (('flrp', a, a, b), a), + (('flrp', 0.0, a, b), ('fmul', a, b)), + (('flrp', a, b, c), ('fadd', ('fmul', c, ('fsub', b, a)), a), 'options->lower_flrp'), + (('ffract', a), ('fsub', a, ('ffloor', a)), 'options->lower_ffract'), + (('fadd', ('fmul', a, ('fadd', 1.0, ('fneg', c))), ('fmul', b, c)), ('flrp', a, b, c), '!options->lower_flrp'), + (('fadd', a, ('fmul', c, ('fadd', b, ('fneg', a)))), ('flrp', a, b, c), '!options->lower_flrp'), + (('ffma', a, b, c), ('fadd', ('fmul', a, b), c), 'options->lower_ffma'), + (('fadd', ('fmul', a, b), c), ('ffma', a, b, c), '!options->lower_ffma'), + # Comparison simplifications + (('inot', ('flt', a, b)), ('fge', a, b)), + (('inot', ('fge', a, b)), ('flt', a, b)), + (('inot', ('feq', a, b)), ('fne', a, b)), + (('inot', ('fne', a, b)), ('feq', a, b)), + (('inot', ('ilt', a, b)), ('ige', a, b)), + (('inot', ('ige', a, b)), ('ilt', a, b)), + (('inot', ('ieq', a, b)), ('ine', a, b)), + (('inot', ('ine', a, b)), ('ieq', a, b)), + (('fge', ('fneg', ('fabs', a)), 0.0), ('feq', a, 0.0)), + (('bcsel', ('flt', a, b), a, b), ('fmin', a, b)), + (('bcsel', ('flt', a, b), b, a), ('fmax', a, b)), + (('bcsel', ('inot', 'a@bool'), b, c), ('bcsel', a, c, b)), + (('bcsel', a, ('bcsel', a, b, c), d), ('bcsel', a, b, d)), + (('fmin', a, a), a), + (('fmax', a, a), a), + (('imin', a, a), a), + (('imax', a, a), a), + (('umin', a, a), a), + (('umax', a, a), a), + (('fmin', ('fmax', a, 0.0), 1.0), ('fsat', a), '!options->lower_fsat'), + (('fmax', ('fmin', a, 1.0), 0.0), ('fsat', a), '!options->lower_fsat'), + (('fsat', a), ('fmin', ('fmax', a, 0.0), 1.0), 'options->lower_fsat'), + (('fsat', ('fsat', a)), ('fsat', a)), + (('fmin', ('fmax', ('fmin', ('fmax', a, 0.0), 1.0), 0.0), 1.0), ('fmin', ('fmax', a, 0.0), 1.0)), + (('ior', ('flt', a, b), ('flt', a, c)), ('flt', a, ('fmax', b, c))), + (('ior', ('flt', a, c), ('flt', b, c)), ('flt', ('fmin', a, b), c)), + (('ior', ('fge', a, b), ('fge', a, c)), ('fge', a, ('fmin', b, c))), + (('ior', ('fge', a, c), ('fge', b, c)), ('fge', ('fmax', a, b), c)), + (('slt', a, b), ('b2f', ('flt', a, b)), 'options->lower_scmp'), + (('sge', a, b), ('b2f', ('fge', a, b)), 'options->lower_scmp'), + (('seq', a, b), ('b2f', ('feq', a, b)), 'options->lower_scmp'), + (('sne', a, b), ('b2f', ('fne', a, b)), 'options->lower_scmp'), + (('fne', ('fneg', a), a), ('fne', a, 0.0)), + (('feq', ('fneg', a), a), ('feq', a, 0.0)), + # Emulating booleans + (('imul', ('b2i', a), ('b2i', b)), ('b2i', ('iand', a, b))), + (('fmul', ('b2f', a), ('b2f', b)), ('b2f', ('iand', a, b))), + (('fsat', ('fadd', ('b2f', a), ('b2f', b))), ('b2f', ('ior', a, b))), + (('iand', 'a@bool', 1.0), ('b2f', a)), + (('flt', ('fneg', ('b2f', a)), 0), a), # Generated by TGSI KILL_IF. + (('flt', ('fsub', 0.0, ('b2f', a)), 0), a), # Generated by TGSI KILL_IF. + # Comparison with the same args. Note that these are not done for + # the float versions because NaN always returns false on float + # inequalities. + (('ilt', a, a), False), + (('ige', a, a), True), + (('ieq', a, a), True), + (('ine', a, a), False), + (('ult', a, a), False), + (('uge', a, a), True), + # Logical and bit operations + (('fand', a, 0.0), 0.0), + (('iand', a, a), a), + (('iand', a, ~0), a), + (('iand', a, 0), 0), + (('ior', a, a), a), + (('ior', a, 0), a), + (('fxor', a, a), 0.0), + (('ixor', a, a), 0), + (('inot', ('inot', a)), a), + # DeMorgan's Laws + (('iand', ('inot', a), ('inot', b)), ('inot', ('ior', a, b))), + (('ior', ('inot', a), ('inot', b)), ('inot', ('iand', a, b))), + # Shift optimizations + (('ishl', 0, a), 0), + (('ishl', a, 0), a), + (('ishr', 0, a), 0), + (('ishr', a, 0), a), + (('ushr', 0, a), 0), + (('ushr', a, 0), a), + # Exponential/logarithmic identities + (('fexp2', ('flog2', a)), a), # 2^lg2(a) = a + (('flog2', ('fexp2', a)), a), # lg2(2^a) = a + (('fpow', a, b), ('fexp2', ('fmul', ('flog2', a), b)), 'options->lower_fpow'), # a^b = 2^(lg2(a)*b) + (('fexp2', ('fmul', ('flog2', a), b)), ('fpow', a, b), '!options->lower_fpow'), # 2^(lg2(a)*b) = a^b + (('fpow', a, 1.0), a), + (('fpow', a, 2.0), ('fmul', a, a)), + (('fpow', a, 4.0), ('fmul', ('fmul', a, a), ('fmul', a, a))), + (('fpow', 2.0, a), ('fexp2', a)), + (('fpow', ('fpow', a, 2.2), 0.454545), a), + (('fpow', ('fabs', ('fpow', a, 2.2)), 0.454545), ('fabs', a)), + (('fsqrt', ('fexp2', a)), ('fexp2', ('fmul', 0.5, a))), + (('frcp', ('fexp2', a)), ('fexp2', ('fneg', a))), + (('frsq', ('fexp2', a)), ('fexp2', ('fmul', -0.5, a))), + (('flog2', ('fsqrt', a)), ('fmul', 0.5, ('flog2', a))), + (('flog2', ('frcp', a)), ('fneg', ('flog2', a))), + (('flog2', ('frsq', a)), ('fmul', -0.5, ('flog2', a))), + (('flog2', ('fpow', a, b)), ('fmul', b, ('flog2', a))), + (('fadd', ('flog2', a), ('flog2', b)), ('flog2', ('fmul', a, b))), + (('fadd', ('flog2', a), ('fneg', ('flog2', b))), ('flog2', ('fdiv', a, b))), + (('fmul', ('fexp2', a), ('fexp2', b)), ('fexp2', ('fadd', a, b))), + # Division and reciprocal + (('fdiv', 1.0, a), ('frcp', a)), + (('fdiv', a, b), ('fmul', a, ('frcp', b)), 'options->lower_fdiv'), + (('frcp', ('frcp', a)), a), + (('frcp', ('fsqrt', a)), ('frsq', a)), + (('fsqrt', a), ('frcp', ('frsq', a)), 'options->lower_fsqrt'), + (('frcp', ('frsq', a)), ('fsqrt', a), '!options->lower_fsqrt'), + # Boolean simplifications + (('ieq', 'a@bool', True), a), + (('ine', 'a@bool', True), ('inot', a)), + (('ine', 'a@bool', False), a), + (('ieq', 'a@bool', False), ('inot', 'a')), + (('bcsel', a, True, False), ('ine', a, 0)), + (('bcsel', a, False, True), ('ieq', a, 0)), + (('bcsel', True, b, c), b), + (('bcsel', False, b, c), c), + # The result of this should be hit by constant propagation and, in the + # next round of opt_algebraic, get picked up by one of the above two. + (('bcsel', '#a', b, c), ('bcsel', ('ine', 'a', 0), b, c)), + + (('bcsel', a, b, b), b), + (('fcsel', a, b, b), b), + + # Conversions + (('i2b', ('b2i', a)), a), + (('f2i', ('ftrunc', a)), ('f2i', a)), + (('f2u', ('ftrunc', a)), ('f2u', a)), + + # Subtracts + (('fsub', a, ('fsub', 0.0, b)), ('fadd', a, b)), + (('isub', a, ('isub', 0, b)), ('iadd', a, b)), + (('ussub_4x8', a, 0), a), + (('ussub_4x8', a, ~0), 0), + (('fsub', a, b), ('fadd', a, ('fneg', b)), 'options->lower_sub'), + (('isub', a, b), ('iadd', a, ('ineg', b)), 'options->lower_sub'), + (('fneg', a), ('fsub', 0.0, a), 'options->lower_negate'), + (('ineg', a), ('isub', 0, a), 'options->lower_negate'), + (('fadd', a, ('fsub', 0.0, b)), ('fsub', a, b)), + (('iadd', a, ('isub', 0, b)), ('isub', a, b)), + (('fabs', ('fsub', 0.0, a)), ('fabs', a)), + (('iabs', ('isub', 0, a)), ('iabs', a)), + + # Misc. lowering + (('fmod', a, b), ('fsub', a, ('fmul', b, ('ffloor', ('fdiv', a, b)))), 'options->lower_fmod'), + (('uadd_carry', a, b), ('b2i', ('ult', ('iadd', a, b), a)), 'options->lower_uadd_carry'), + (('usub_borrow', a, b), ('b2i', ('ult', a, b)), 'options->lower_usub_borrow'), + + (('bitfield_insert', 'base', 'insert', 'offset', 'bits'), + ('bcsel', ('ilt', 31, 'bits'), 'insert', + ('bfi', ('bfm', 'bits', 'offset'), 'insert', 'base')), + 'options->lower_bitfield_insert'), + + (('ibitfield_extract', 'value', 'offset', 'bits'), + ('bcsel', ('ilt', 31, 'bits'), 'value', + ('ibfe', 'value', 'offset', 'bits')), + 'options->lower_bitfield_extract'), + + (('ubitfield_extract', 'value', 'offset', 'bits'), + ('bcsel', ('ult', 31, 'bits'), 'value', + ('ubfe', 'value', 'offset', 'bits')), + 'options->lower_bitfield_extract'), +] + +# Add optimizations to handle the case where the result of a ternary is +# compared to a constant. This way we can take things like +# +# (a ? 0 : 1) > 0 +# +# and turn it into +# +# a ? (0 > 0) : (1 > 0) +# +# which constant folding will eat for lunch. The resulting ternary will +# further get cleaned up by the boolean reductions above and we will be +# left with just the original variable "a". +for op in ['flt', 'fge', 'feq', 'fne', + 'ilt', 'ige', 'ieq', 'ine', 'ult', 'uge']: + optimizations += [ + ((op, ('bcsel', 'a', '#b', '#c'), '#d'), + ('bcsel', 'a', (op, 'b', 'd'), (op, 'c', 'd'))), + ((op, '#d', ('bcsel', a, '#b', '#c')), + ('bcsel', 'a', (op, 'd', 'b'), (op, 'd', 'c'))), + ] + +# This section contains "late" optimizations that should be run after the +# regular optimizations have finished. Optimizations should go here if +# they help code generation but do not necessarily produce code that is +# more easily optimizable. +late_optimizations = [ + (('flt', ('fadd', a, b), 0.0), ('flt', a, ('fneg', b))), + (('fge', ('fadd', a, b), 0.0), ('fge', a, ('fneg', b))), + (('feq', ('fadd', a, b), 0.0), ('feq', a, ('fneg', b))), + (('fne', ('fadd', a, b), 0.0), ('fne', a, ('fneg', b))), + (('fdot2', a, b), ('fdot_replicated2', a, b), 'options->fdot_replicates'), + (('fdot3', a, b), ('fdot_replicated3', a, b), 'options->fdot_replicates'), + (('fdot4', a, b), ('fdot_replicated4', a, b), 'options->fdot_replicates'), + (('fdph', a, b), ('fdph_replicated', a, b), 'options->fdot_replicates'), +] + +print nir_algebraic.AlgebraicPass("nir_opt_algebraic", optimizations).render() +print nir_algebraic.AlgebraicPass("nir_opt_algebraic_late", + late_optimizations).render() diff --git a/src/compiler/nir/nir_opt_constant_folding.c b/src/compiler/nir/nir_opt_constant_folding.c new file mode 100644 index 0000000..28a73f8 --- /dev/null +++ b/src/compiler/nir/nir_opt_constant_folding.c @@ -0,0 +1,201 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand (jason@jlekstrand.net) + * + */ + +#include "nir_constant_expressions.h" +#include <math.h> + +/* + * Implements SSA-based constant folding. + */ + +struct constant_fold_state { + void *mem_ctx; + nir_function_impl *impl; + bool progress; +}; + +static bool +constant_fold_alu_instr(nir_alu_instr *instr, void *mem_ctx) +{ + nir_const_value src[4]; + + if (!instr->dest.dest.is_ssa) + return false; + + for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) { + if (!instr->src[i].src.is_ssa) + return false; + + nir_instr *src_instr = instr->src[i].src.ssa->parent_instr; + + if (src_instr->type != nir_instr_type_load_const) + return false; + nir_load_const_instr* load_const = nir_instr_as_load_const(src_instr); + + for (unsigned j = 0; j < nir_ssa_alu_instr_src_components(instr, i); + j++) { + src[i].u[j] = load_const->value.u[instr->src[i].swizzle[j]]; + } + + /* We shouldn't have any source modifiers in the optimization loop. */ + assert(!instr->src[i].abs && !instr->src[i].negate); + } + + /* We shouldn't have any saturate modifiers in the optimization loop. */ + assert(!instr->dest.saturate); + + nir_const_value dest = + nir_eval_const_opcode(instr->op, instr->dest.dest.ssa.num_components, + src); + + nir_load_const_instr *new_instr = + nir_load_const_instr_create(mem_ctx, + instr->dest.dest.ssa.num_components); + + new_instr->value = dest; + + nir_instr_insert_before(&instr->instr, &new_instr->instr); + + nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, + nir_src_for_ssa(&new_instr->def)); + + nir_instr_remove(&instr->instr); + ralloc_free(instr); + + return true; +} + +static bool +constant_fold_deref(nir_instr *instr, nir_deref_var *deref) +{ + bool progress = false; + + for (nir_deref *tail = deref->deref.child; tail; tail = tail->child) { + if (tail->deref_type != nir_deref_type_array) + continue; + + nir_deref_array *arr = nir_deref_as_array(tail); + + if (arr->deref_array_type == nir_deref_array_type_indirect && + arr->indirect.is_ssa && + arr->indirect.ssa->parent_instr->type == nir_instr_type_load_const) { + nir_load_const_instr *indirect = + nir_instr_as_load_const(arr->indirect.ssa->parent_instr); + + arr->base_offset += indirect->value.u[0]; + + /* Clear out the source */ + nir_instr_rewrite_src(instr, &arr->indirect, nir_src_for_ssa(NULL)); + + arr->deref_array_type = nir_deref_array_type_direct; + + progress = true; + } + } + + return progress; +} + +static bool +constant_fold_intrinsic_instr(nir_intrinsic_instr *instr) +{ + bool progress = false; + + unsigned num_vars = nir_intrinsic_infos[instr->intrinsic].num_variables; + for (unsigned i = 0; i < num_vars; i++) { + progress |= constant_fold_deref(&instr->instr, instr->variables[i]); + } + + return progress; +} + +static bool +constant_fold_tex_instr(nir_tex_instr *instr) +{ + if (instr->sampler) + return constant_fold_deref(&instr->instr, instr->sampler); + else + return false; +} + +static bool +constant_fold_block(nir_block *block, void *void_state) +{ + struct constant_fold_state *state = void_state; + + nir_foreach_instr_safe(block, instr) { + switch (instr->type) { + case nir_instr_type_alu: + state->progress |= constant_fold_alu_instr(nir_instr_as_alu(instr), + state->mem_ctx); + break; + case nir_instr_type_intrinsic: + state->progress |= + constant_fold_intrinsic_instr(nir_instr_as_intrinsic(instr)); + break; + case nir_instr_type_tex: + state->progress |= constant_fold_tex_instr(nir_instr_as_tex(instr)); + break; + default: + /* Don't know how to constant fold */ + break; + } + } + + return true; +} + +static bool +nir_opt_constant_folding_impl(nir_function_impl *impl) +{ + struct constant_fold_state state; + + state.mem_ctx = ralloc_parent(impl); + state.impl = impl; + state.progress = false; + + nir_foreach_block(impl, constant_fold_block, &state); + + if (state.progress) + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance); + + return state.progress; +} + +bool +nir_opt_constant_folding(nir_shader *shader) +{ + bool progress = false; + + nir_foreach_function(shader, function) { + if (function->impl) + progress |= nir_opt_constant_folding_impl(function->impl); + } + + return progress; +} diff --git a/src/compiler/nir/nir_opt_copy_propagate.c b/src/compiler/nir/nir_opt_copy_propagate.c new file mode 100644 index 0000000..d99f78d --- /dev/null +++ b/src/compiler/nir/nir_opt_copy_propagate.c @@ -0,0 +1,290 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +#include "nir.h" +#include <main/imports.h> + +/** + * SSA-based copy propagation + */ + +static bool is_move(nir_alu_instr *instr) +{ + if (instr->op != nir_op_fmov && + instr->op != nir_op_imov) + return false; + + if (instr->dest.saturate) + return false; + + /* we handle modifiers in a separate pass */ + + if (instr->src[0].abs || instr->src[0].negate) + return false; + + if (!instr->src[0].src.is_ssa) + return false; + + return true; + +} + +static bool is_vec(nir_alu_instr *instr) +{ + for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) { + if (!instr->src[i].src.is_ssa) + return false; + + /* we handle modifiers in a separate pass */ + if (instr->src[i].abs || instr->src[i].negate) + return false; + } + + return instr->op == nir_op_vec2 || + instr->op == nir_op_vec3 || + instr->op == nir_op_vec4; +} + +static bool +is_swizzleless_move(nir_alu_instr *instr) +{ + if (is_move(instr)) { + for (unsigned i = 0; i < 4; i++) { + if (!((instr->dest.write_mask >> i) & 1)) + break; + if (instr->src[0].swizzle[i] != i) + return false; + } + return true; + } else if (is_vec(instr)) { + nir_ssa_def *def = NULL; + for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) { + if (instr->src[i].swizzle[0] != i) + return false; + + if (def == NULL) { + def = instr->src[i].src.ssa; + } else if (instr->src[i].src.ssa != def) { + return false; + } + } + return true; + } else { + return false; + } +} + +static bool +copy_prop_src(nir_src *src, nir_instr *parent_instr, nir_if *parent_if) +{ + if (!src->is_ssa) { + if (src->reg.indirect) + return copy_prop_src(src, parent_instr, parent_if); + return false; + } + + nir_instr *src_instr = src->ssa->parent_instr; + if (src_instr->type != nir_instr_type_alu) + return false; + + nir_alu_instr *alu_instr = nir_instr_as_alu(src_instr); + if (!is_swizzleless_move(alu_instr)) + return false; + + /* Don't let copy propagation land us with a phi that has more + * components in its source than it has in its destination. That badly + * messes up out-of-ssa. + */ + if (parent_instr && parent_instr->type == nir_instr_type_phi) { + nir_phi_instr *phi = nir_instr_as_phi(parent_instr); + assert(phi->dest.is_ssa); + if (phi->dest.ssa.num_components != + alu_instr->src[0].src.ssa->num_components) + return false; + } + + if (parent_instr) { + nir_instr_rewrite_src(parent_instr, src, + nir_src_for_ssa(alu_instr->src[0].src.ssa)); + } else { + assert(src == &parent_if->condition); + nir_if_rewrite_condition(parent_if, + nir_src_for_ssa(alu_instr->src[0].src.ssa)); + } + + return true; +} + +static bool +copy_prop_alu_src(nir_alu_instr *parent_alu_instr, unsigned index) +{ + nir_alu_src *src = &parent_alu_instr->src[index]; + if (!src->src.is_ssa) { + if (src->src.reg.indirect) + return copy_prop_src(src->src.reg.indirect, &parent_alu_instr->instr, + NULL); + return false; + } + + nir_instr *src_instr = src->src.ssa->parent_instr; + if (src_instr->type != nir_instr_type_alu) + return false; + + nir_alu_instr *alu_instr = nir_instr_as_alu(src_instr); + if (!is_move(alu_instr) && !is_vec(alu_instr)) + return false; + + nir_ssa_def *def; + unsigned new_swizzle[4] = {0, 0, 0, 0}; + + if (alu_instr->op == nir_op_fmov || + alu_instr->op == nir_op_imov) { + for (unsigned i = 0; i < 4; i++) + new_swizzle[i] = alu_instr->src[0].swizzle[src->swizzle[i]]; + def = alu_instr->src[0].src.ssa; + } else { + def = NULL; + + for (unsigned i = 0; i < 4; i++) { + if (!nir_alu_instr_channel_used(parent_alu_instr, index, i)) + continue; + + nir_ssa_def *new_def = alu_instr->src[src->swizzle[i]].src.ssa; + if (def == NULL) + def = new_def; + else { + if (def != new_def) + return false; + } + new_swizzle[i] = alu_instr->src[src->swizzle[i]].swizzle[0]; + } + } + + for (unsigned i = 0; i < 4; i++) + src->swizzle[i] = new_swizzle[i]; + + nir_instr_rewrite_src(&parent_alu_instr->instr, &src->src, + nir_src_for_ssa(def)); + + return true; +} + +typedef struct { + nir_instr *parent_instr; + bool progress; +} copy_prop_state; + +static bool +copy_prop_src_cb(nir_src *src, void *_state) +{ + copy_prop_state *state = (copy_prop_state *) _state; + while (copy_prop_src(src, state->parent_instr, NULL)) + state->progress = true; + + return true; +} + +static bool +copy_prop_instr(nir_instr *instr) +{ + if (instr->type == nir_instr_type_alu) { + nir_alu_instr *alu_instr = nir_instr_as_alu(instr); + bool progress = false; + + for (unsigned i = 0; i < nir_op_infos[alu_instr->op].num_inputs; i++) + while (copy_prop_alu_src(alu_instr, i)) + progress = true; + + if (!alu_instr->dest.dest.is_ssa && alu_instr->dest.dest.reg.indirect) + while (copy_prop_src(alu_instr->dest.dest.reg.indirect, instr, NULL)) + progress = true; + + return progress; + } + + copy_prop_state state; + state.parent_instr = instr; + state.progress = false; + nir_foreach_src(instr, copy_prop_src_cb, &state); + + return state.progress; +} + +static bool +copy_prop_if(nir_if *if_stmt) +{ + return copy_prop_src(&if_stmt->condition, NULL, if_stmt); +} + +static bool +copy_prop_block(nir_block *block, void *_state) +{ + bool *progress = (bool *) _state; + + nir_foreach_instr(block, instr) { + if (copy_prop_instr(instr)) + *progress = true; + } + + if (block->cf_node.node.next != NULL && /* check that we aren't the end node */ + !nir_cf_node_is_last(&block->cf_node) && + nir_cf_node_next(&block->cf_node)->type == nir_cf_node_if) { + nir_if *if_stmt = nir_cf_node_as_if(nir_cf_node_next(&block->cf_node)); + if (copy_prop_if(if_stmt)) + *progress = true; + } + + return true; +} + +static bool +nir_copy_prop_impl(nir_function_impl *impl) +{ + bool progress = false; + + nir_foreach_block(impl, copy_prop_block, &progress); + + if (progress) { + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance); + } + + return progress; +} + +bool +nir_copy_prop(nir_shader *shader) +{ + bool progress = false; + + nir_foreach_function(shader, function) { + if (function->impl && nir_copy_prop_impl(function->impl)) + progress = true; + } + + return progress; +} diff --git a/src/compiler/nir/nir_opt_cse.c b/src/compiler/nir/nir_opt_cse.c new file mode 100644 index 0000000..364fb02 --- /dev/null +++ b/src/compiler/nir/nir_opt_cse.c @@ -0,0 +1,93 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand (jason@jlekstrand.net) + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +#include "nir_instr_set.h" + +/* + * Implements common subexpression elimination + */ + +/* + * Visits and CSE's the given block and all its descendants in the dominance + * tree recursively. Note that the instr_set is guaranteed to only ever + * contain instructions that dominate the current block. + */ + +static bool +cse_block(nir_block *block, struct set *instr_set) +{ + bool progress = false; + + nir_foreach_instr_safe(block, instr) { + if (nir_instr_set_add_or_rewrite(instr_set, instr)) { + progress = true; + nir_instr_remove(instr); + } + } + + for (unsigned i = 0; i < block->num_dom_children; i++) { + nir_block *child = block->dom_children[i]; + progress |= cse_block(child, instr_set); + } + + nir_foreach_instr(block, instr) + nir_instr_set_remove(instr_set, instr); + + return progress; +} + +static bool +nir_opt_cse_impl(nir_function_impl *impl) +{ + struct set *instr_set = nir_instr_set_create(NULL); + + nir_metadata_require(impl, nir_metadata_dominance); + + bool progress = cse_block(nir_start_block(impl), instr_set); + + if (progress) + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance); + + nir_instr_set_destroy(instr_set); + return progress; +} + +bool +nir_opt_cse(nir_shader *shader) +{ + bool progress = false; + + nir_foreach_function(shader, function) { + if (function->impl) + progress |= nir_opt_cse_impl(function->impl); + } + + return progress; +} + diff --git a/src/compiler/nir/nir_opt_dce.c b/src/compiler/nir/nir_opt_dce.c new file mode 100644 index 0000000..32436c1 --- /dev/null +++ b/src/compiler/nir/nir_opt_dce.c @@ -0,0 +1,183 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +#include "nir.h" + +/* SSA-based mark-and-sweep dead code elimination */ + +typedef struct { + struct exec_node node; + nir_instr *instr; +} worklist_elem; + +static void +worklist_push(struct exec_list *worklist, nir_instr *instr) +{ + worklist_elem *elem = ralloc(worklist, worklist_elem); + elem->instr = instr; + instr->pass_flags = 1; + exec_list_push_tail(worklist, &elem->node); +} + +static nir_instr * +worklist_pop(struct exec_list *worklist) +{ + struct exec_node *node = exec_list_pop_head(worklist); + worklist_elem *elem = exec_node_data(worklist_elem, node, node); + return elem->instr; +} + +static bool +mark_live_cb(nir_src *src, void *_state) +{ + struct exec_list *worklist = (struct exec_list *) _state; + + if (src->is_ssa && !src->ssa->parent_instr->pass_flags) { + worklist_push(worklist, src->ssa->parent_instr); + } + + return true; +} + +static void +init_instr(nir_instr *instr, struct exec_list *worklist) +{ + nir_alu_instr *alu_instr; + nir_intrinsic_instr *intrin_instr; + nir_tex_instr *tex_instr; + + /* We use the pass_flags to store the live/dead information. In DCE, we + * just treat it as a zero/non-zerl boolean for whether or not the + * instruction is live. + */ + instr->pass_flags = 0; + + switch (instr->type) { + case nir_instr_type_call: + case nir_instr_type_jump: + worklist_push(worklist, instr); + break; + + case nir_instr_type_alu: + alu_instr = nir_instr_as_alu(instr); + if (!alu_instr->dest.dest.is_ssa) + worklist_push(worklist, instr); + break; + + case nir_instr_type_intrinsic: + intrin_instr = nir_instr_as_intrinsic(instr); + if (nir_intrinsic_infos[intrin_instr->intrinsic].flags & + NIR_INTRINSIC_CAN_ELIMINATE) { + if (nir_intrinsic_infos[intrin_instr->intrinsic].has_dest && + !intrin_instr->dest.is_ssa) { + worklist_push(worklist, instr); + } + } else { + worklist_push(worklist, instr); + } + break; + + case nir_instr_type_tex: + tex_instr = nir_instr_as_tex(instr); + if (!tex_instr->dest.is_ssa) + worklist_push(worklist, instr); + break; + + default: + break; + } +} + +static bool +init_block_cb(nir_block *block, void *_state) +{ + struct exec_list *worklist = (struct exec_list *) _state; + + nir_foreach_instr(block, instr) + init_instr(instr, worklist); + + nir_if *following_if = nir_block_get_following_if(block); + if (following_if) { + if (following_if->condition.is_ssa && + !following_if->condition.ssa->parent_instr->pass_flags) + worklist_push(worklist, following_if->condition.ssa->parent_instr); + } + + return true; +} + +static bool +delete_block_cb(nir_block *block, void *_state) +{ + bool *progress = (bool *) _state; + + nir_foreach_instr_safe(block, instr) { + if (!instr->pass_flags) { + nir_instr_remove(instr); + *progress = true; + } + } + + return true; +} + +static bool +nir_opt_dce_impl(nir_function_impl *impl) +{ + struct exec_list *worklist = ralloc(NULL, struct exec_list); + exec_list_make_empty(worklist); + + nir_foreach_block(impl, init_block_cb, worklist); + + while (!exec_list_is_empty(worklist)) { + nir_instr *instr = worklist_pop(worklist); + nir_foreach_src(instr, mark_live_cb, worklist); + } + + ralloc_free(worklist); + + bool progress = false; + nir_foreach_block(impl, delete_block_cb, &progress); + + if (progress) + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance); + + return progress; +} + +bool +nir_opt_dce(nir_shader *shader) +{ + bool progress = false; + nir_foreach_function(shader, function) { + if (function->impl && nir_opt_dce_impl(function->impl)) + progress = true; + } + + return progress; +} diff --git a/src/compiler/nir/nir_opt_dead_cf.c b/src/compiler/nir/nir_opt_dead_cf.c new file mode 100644 index 0000000..4cc6798 --- /dev/null +++ b/src/compiler/nir/nir_opt_dead_cf.c @@ -0,0 +1,358 @@ +/* + * Copyright © 2014 Connor Abbott + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +#include "nir.h" +#include "nir_control_flow.h" + +/* + * This file implements an optimization that deletes statically + * unreachable/dead code. In NIR, one way this can happen if if an if + * statement has a constant condition: + * + * if (true) { + * ... + * } + * + * We delete the if statement and paste the contents of the always-executed + * branch into the surrounding control flow, possibly removing more code if + * the branch had a jump at the end. + * + * Another way is that control flow can end in a jump so that code after it + * never gets executed. In particular, this can happen after optimizing + * something like: + * + * if (true) { + * ... + * break; + * } + * ... + * + * We also consider the case where both branches of an if end in a jump, e.g.: + * + * if (...) { + * break; + * } else { + * continue; + * } + * ... + * + * Finally, we also handle removing useless loops, i.e. loops with no side + * effects and without any definitions that are used elsewhere. This case is a + * little different from the first two in that the code is actually run (it + * just never does anything), but there are similar issues with needing to + * be careful with restarting after deleting the cf_node (see dead_cf_list()) + * so this is a convenient place to remove them. + */ + +static void +remove_after_cf_node(nir_cf_node *node) +{ + nir_cf_node *end = node; + while (!nir_cf_node_is_last(end)) + end = nir_cf_node_next(end); + + nir_cf_list list; + nir_cf_extract(&list, nir_after_cf_node(node), nir_after_cf_node(end)); + nir_cf_delete(&list); +} + +static void +opt_constant_if(nir_if *if_stmt, bool condition) +{ + /* First, we need to remove any phi nodes after the if by rewriting uses to + * point to the correct source. + */ + nir_block *after = nir_cf_node_as_block(nir_cf_node_next(&if_stmt->cf_node)); + nir_block *last_block = + nir_cf_node_as_block(condition ? nir_if_last_then_node(if_stmt) + : nir_if_last_else_node(if_stmt)); + + nir_foreach_instr_safe(after, instr) { + if (instr->type != nir_instr_type_phi) + break; + + nir_phi_instr *phi = nir_instr_as_phi(instr); + nir_ssa_def *def = NULL; + nir_foreach_phi_src(phi, phi_src) { + if (phi_src->pred != last_block) + continue; + + assert(phi_src->src.is_ssa); + def = phi_src->src.ssa; + } + + assert(def); + assert(phi->dest.is_ssa); + nir_ssa_def_rewrite_uses(&phi->dest.ssa, nir_src_for_ssa(def)); + nir_instr_remove(instr); + } + + /* The control flow list we're about to paste in may include a jump at the + * end, and in that case we have to delete the rest of the control flow + * list after the if since it's unreachable and the validator will balk if + * we don't. + */ + + if (!exec_list_is_empty(&last_block->instr_list)) { + nir_instr *last_instr = nir_block_last_instr(last_block); + if (last_instr->type == nir_instr_type_jump) + remove_after_cf_node(&if_stmt->cf_node); + } + + /* Finally, actually paste in the then or else branch and delete the if. */ + struct exec_list *cf_list = condition ? &if_stmt->then_list + : &if_stmt->else_list; + + nir_cf_list list; + nir_cf_extract(&list, nir_before_cf_list(cf_list), + nir_after_cf_list(cf_list)); + nir_cf_reinsert(&list, nir_after_cf_node(&if_stmt->cf_node)); + nir_cf_node_remove(&if_stmt->cf_node); +} + +static bool +block_has_no_side_effects(nir_block *block, void *state) +{ + (void) state; + + nir_foreach_instr(block, instr) { + if (instr->type == nir_instr_type_call) + return false; + + /* Return instructions can cause us to skip over other side-effecting + * instructions after the loop, so consider them to have side effects + * here. + */ + + if (instr->type == nir_instr_type_jump && + nir_instr_as_jump(instr)->type == nir_jump_return) + return false; + + if (instr->type != nir_instr_type_intrinsic) + continue; + + nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); + if (!nir_intrinsic_infos[intrin->intrinsic].flags & + NIR_INTRINSIC_CAN_ELIMINATE) + return false; + } + + return true; +} + +static bool +def_not_live_out(nir_ssa_def *def, void *state) +{ + nir_block *after = state; + + return !BITSET_TEST(after->live_in, def->live_index); +} + +/* + * Test if a loop is dead. A loop is dead if: + * + * 1) It has no side effects (i.e. intrinsics which could possibly affect the + * state of the program aside from producing an SSA value, indicated by a lack + * of NIR_INTRINSIC_CAN_ELIMINATE). + * + * 2) It has no phi nodes after it, since those indicate values inside the + * loop being used after the loop. + * + * 3) If there are no phi nodes after the loop, then the only way a value + * defined inside the loop can be used outside the loop is if its definition + * dominates the block after the loop. If none of the definitions that + * dominate the loop exit are used outside the loop, then the loop is dead + * and it can be deleted. + */ + +static bool +loop_is_dead(nir_loop *loop) +{ + nir_block *before = nir_cf_node_as_block(nir_cf_node_prev(&loop->cf_node)); + nir_block *after = nir_cf_node_as_block(nir_cf_node_next(&loop->cf_node)); + + if (!exec_list_is_empty(&after->instr_list) && + nir_block_first_instr(after)->type == nir_instr_type_phi) + return false; + + if (!nir_foreach_block_in_cf_node(&loop->cf_node, block_has_no_side_effects, + NULL)) + return false; + + nir_function_impl *impl = nir_cf_node_get_function(&loop->cf_node); + nir_metadata_require(impl, nir_metadata_live_ssa_defs | + nir_metadata_dominance); + + for (nir_block *cur = after->imm_dom; cur != before; cur = cur->imm_dom) { + nir_foreach_instr(cur, instr) { + if (!nir_foreach_ssa_def(instr, def_not_live_out, after)) + return false; + } + } + + return true; +} + +static bool +dead_cf_block(nir_block *block) +{ + nir_if *following_if = nir_block_get_following_if(block); + if (following_if) { + nir_const_value *const_value = + nir_src_as_const_value(following_if->condition); + + if (!const_value) + return false; + + opt_constant_if(following_if, const_value->u[0] != 0); + return true; + } + + nir_loop *following_loop = nir_block_get_following_loop(block); + if (!following_loop) + return false; + + if (!loop_is_dead(following_loop)) + return false; + + nir_cf_node_remove(&following_loop->cf_node); + return true; +} + +static bool +ends_in_jump(nir_block *block) +{ + if (exec_list_is_empty(&block->instr_list)) + return false; + + nir_instr *instr = nir_block_last_instr(block); + return instr->type == nir_instr_type_jump; +} + +static bool +dead_cf_list(struct exec_list *list, bool *list_ends_in_jump) +{ + bool progress = false; + *list_ends_in_jump = false; + + nir_cf_node *prev = NULL; + + foreach_list_typed(nir_cf_node, cur, node, list) { + switch (cur->type) { + case nir_cf_node_block: { + nir_block *block = nir_cf_node_as_block(cur); + if (dead_cf_block(block)) { + /* We just deleted the if or loop after this block, so we may have + * deleted the block before or after it -- which one is an + * implementation detail. Therefore, to recover the place we were + * at, we have to use the previous cf_node. + */ + + if (prev) { + cur = nir_cf_node_next(prev); + } else { + cur = exec_node_data(nir_cf_node, exec_list_get_head(list), + node); + } + + block = nir_cf_node_as_block(cur); + + progress = true; + } + + if (ends_in_jump(block)) { + *list_ends_in_jump = true; + + if (!exec_node_is_tail_sentinel(cur->node.next)) { + remove_after_cf_node(cur); + return true; + } + } + + break; + } + + case nir_cf_node_if: { + nir_if *if_stmt = nir_cf_node_as_if(cur); + bool then_ends_in_jump, else_ends_in_jump; + progress |= dead_cf_list(&if_stmt->then_list, &then_ends_in_jump); + progress |= dead_cf_list(&if_stmt->else_list, &else_ends_in_jump); + + if (then_ends_in_jump && else_ends_in_jump) { + *list_ends_in_jump = true; + nir_block *next = nir_cf_node_as_block(nir_cf_node_next(cur)); + if (!exec_list_is_empty(&next->instr_list) || + !exec_node_is_tail_sentinel(next->cf_node.node.next)) { + remove_after_cf_node(cur); + return true; + } + } + + break; + } + + case nir_cf_node_loop: { + nir_loop *loop = nir_cf_node_as_loop(cur); + bool dummy; + progress |= dead_cf_list(&loop->body, &dummy); + + break; + } + + default: + unreachable("unknown cf node type"); + } + + prev = cur; + } + + return progress; +} + +static bool +opt_dead_cf_impl(nir_function_impl *impl) +{ + bool dummy; + bool progress = dead_cf_list(&impl->body, &dummy); + + if (progress) + nir_metadata_preserve(impl, nir_metadata_none); + + return progress; +} + +bool +nir_opt_dead_cf(nir_shader *shader) +{ + bool progress = false; + + nir_foreach_function(shader, function) + if (function->impl) + progress |= opt_dead_cf_impl(function->impl); + + return progress; +} diff --git a/src/compiler/nir/nir_opt_gcm.c b/src/compiler/nir/nir_opt_gcm.c new file mode 100644 index 0000000..a8779ce --- /dev/null +++ b/src/compiler/nir/nir_opt_gcm.c @@ -0,0 +1,494 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand (jason@jlekstrand.net) + * + */ + +#include "nir.h" + +/* + * Implements Global Code Motion. A description of GCM can be found in + * "Global Code Motion; Global Value Numbering" by Cliff Click. + * Unfortunately, the algorithm presented in the paper is broken in a + * number of ways. The algorithm used here differs substantially from the + * one in the paper but it is, in my opinion, much easier to read and + * verify correcness. + */ + +struct gcm_block_info { + /* Number of loops this block is inside */ + unsigned loop_depth; + + /* The last instruction inserted into this block. This is used as we + * traverse the instructions and insert them back into the program to + * put them in the right order. + */ + nir_instr *last_instr; +}; + +/* Flags used in the instr->pass_flags field for various instruction states */ +enum { + GCM_INSTR_PINNED = (1 << 0), + GCM_INSTR_SCHEDULED_EARLY = (1 << 1), + GCM_INSTR_SCHEDULED_LATE = (1 << 2), + GCM_INSTR_PLACED = (1 << 3), +}; + +struct gcm_state { + nir_function_impl *impl; + nir_instr *instr; + + /* The list of non-pinned instructions. As we do the late scheduling, + * we pull non-pinned instructions out of their blocks and place them in + * this list. This saves us from having linked-list problems when we go + * to put instructions back in their blocks. + */ + struct exec_list instrs; + + struct gcm_block_info *blocks; +}; + +/* Recursively walks the CFG and builds the block_info structure */ +static void +gcm_build_block_info(struct exec_list *cf_list, struct gcm_state *state, + unsigned loop_depth) +{ + foreach_list_typed(nir_cf_node, node, node, cf_list) { + switch (node->type) { + case nir_cf_node_block: { + nir_block *block = nir_cf_node_as_block(node); + state->blocks[block->index].loop_depth = loop_depth; + break; + } + case nir_cf_node_if: { + nir_if *if_stmt = nir_cf_node_as_if(node); + gcm_build_block_info(&if_stmt->then_list, state, loop_depth); + gcm_build_block_info(&if_stmt->else_list, state, loop_depth); + break; + } + case nir_cf_node_loop: { + nir_loop *loop = nir_cf_node_as_loop(node); + gcm_build_block_info(&loop->body, state, loop_depth + 1); + break; + } + default: + unreachable("Invalid CF node type"); + } + } +} + +/* Walks the instruction list and marks immovable instructions as pinned + * + * This function also serves to initialize the instr->pass_flags field. + * After this is completed, all instructions' pass_flags fields will be set + * to either GCM_INSTR_PINNED or 0. + */ +static bool +gcm_pin_instructions_block(nir_block *block, void *void_state) +{ + struct gcm_state *state = void_state; + + nir_foreach_instr_safe(block, instr) { + switch (instr->type) { + case nir_instr_type_alu: + switch (nir_instr_as_alu(instr)->op) { + case nir_op_fddx: + case nir_op_fddy: + case nir_op_fddx_fine: + case nir_op_fddy_fine: + case nir_op_fddx_coarse: + case nir_op_fddy_coarse: + /* These can only go in uniform control flow; pin them for now */ + instr->pass_flags = GCM_INSTR_PINNED; + break; + + default: + instr->pass_flags = 0; + break; + } + break; + + case nir_instr_type_tex: + switch (nir_instr_as_tex(instr)->op) { + case nir_texop_tex: + case nir_texop_txb: + case nir_texop_lod: + /* These two take implicit derivatives so they need to be pinned */ + instr->pass_flags = GCM_INSTR_PINNED; + break; + + default: + instr->pass_flags = 0; + break; + } + break; + + case nir_instr_type_load_const: + instr->pass_flags = 0; + break; + + case nir_instr_type_intrinsic: { + const nir_intrinsic_info *info = + &nir_intrinsic_infos[nir_instr_as_intrinsic(instr)->intrinsic]; + + if ((info->flags & NIR_INTRINSIC_CAN_ELIMINATE) && + (info->flags & NIR_INTRINSIC_CAN_REORDER)) { + instr->pass_flags = 0; + } else { + instr->pass_flags = GCM_INSTR_PINNED; + } + break; + } + + case nir_instr_type_jump: + case nir_instr_type_ssa_undef: + case nir_instr_type_phi: + instr->pass_flags = GCM_INSTR_PINNED; + break; + + default: + unreachable("Invalid instruction type in GCM"); + } + + if (!(instr->pass_flags & GCM_INSTR_PINNED)) { + /* If this is an unpinned instruction, go ahead and pull it out of + * the program and put it on the instrs list. This has a couple + * of benifits. First, it makes the scheduling algorithm more + * efficient because we can avoid walking over basic blocks and + * pinned instructions. Second, it keeps us from causing linked + * list confusion when we're trying to put everything in its + * proper place at the end of the pass. + * + * Note that we don't use nir_instr_remove here because that also + * cleans up uses and defs and we want to keep that information. + */ + exec_node_remove(&instr->node); + exec_list_push_tail(&state->instrs, &instr->node); + } + } + + return true; +} + +static void +gcm_schedule_early_instr(nir_instr *instr, struct gcm_state *state); + +/** Update an instructions schedule for the given source + * + * This function is called iteratively as we walk the sources of an + * instruction. It ensures that the given source instruction has been + * scheduled and then update this instruction's block if the source + * instruction is lower down the tree. + */ +static bool +gcm_schedule_early_src(nir_src *src, void *void_state) +{ + struct gcm_state *state = void_state; + nir_instr *instr = state->instr; + + assert(src->is_ssa); + + gcm_schedule_early_instr(src->ssa->parent_instr, void_state); + + /* While the index isn't a proper dominance depth, it does have the + * property that if A dominates B then A->index <= B->index. Since we + * know that this instruction must have been dominated by all of its + * sources at some point (even if it's gone through value-numbering), + * all of the sources must lie on the same branch of the dominance tree. + * Therefore, we can just go ahead and just compare indices. + */ + if (instr->block->index < src->ssa->parent_instr->block->index) + instr->block = src->ssa->parent_instr->block; + + /* We need to restore the state instruction because it may have been + * changed through the gcm_schedule_early_instr call above. Since we + * may still be iterating through sources and future calls to + * gcm_schedule_early_src for the same instruction will still need it. + */ + state->instr = instr; + + return true; +} + +/** Schedules an instruction early + * + * This function performs a recursive depth-first search starting at the + * given instruction and proceeding through the sources to schedule + * instructions as early as they can possibly go in the dominance tree. + * The instructions are "scheduled" by updating their instr->block field. + */ +static void +gcm_schedule_early_instr(nir_instr *instr, struct gcm_state *state) +{ + if (instr->pass_flags & GCM_INSTR_SCHEDULED_EARLY) + return; + + instr->pass_flags |= GCM_INSTR_SCHEDULED_EARLY; + + /* Pinned instructions are already scheduled so we don't need to do + * anything. Also, bailing here keeps us from ever following the + * sources of phi nodes which can be back-edges. + */ + if (instr->pass_flags & GCM_INSTR_PINNED) + return; + + /* Start with the instruction at the top. As we iterate over the + * sources, it will get moved down as needed. + */ + instr->block = nir_start_block(state->impl); + state->instr = instr; + + nir_foreach_src(instr, gcm_schedule_early_src, state); +} + +static void +gcm_schedule_late_instr(nir_instr *instr, struct gcm_state *state); + +/** Schedules the instruction associated with the given SSA def late + * + * This function works by first walking all of the uses of the given SSA + * definition, ensuring that they are scheduled, and then computing the LCA + * (least common ancestor) of its uses. It then schedules this instruction + * as close to the LCA as possible while trying to stay out of loops. + */ +static bool +gcm_schedule_late_def(nir_ssa_def *def, void *void_state) +{ + struct gcm_state *state = void_state; + + nir_block *lca = NULL; + + nir_foreach_use(def, use_src) { + nir_instr *use_instr = use_src->parent_instr; + + gcm_schedule_late_instr(use_instr, state); + + /* Phi instructions are a bit special. SSA definitions don't have to + * dominate the sources of the phi nodes that use them; instead, they + * have to dominate the predecessor block corresponding to the phi + * source. We handle this by looking through the sources, finding + * any that are usingg this SSA def, and using those blocks instead + * of the one the phi lives in. + */ + if (use_instr->type == nir_instr_type_phi) { + nir_phi_instr *phi = nir_instr_as_phi(use_instr); + + nir_foreach_phi_src(phi, phi_src) { + if (phi_src->src.ssa == def) + lca = nir_dominance_lca(lca, phi_src->pred); + } + } else { + lca = nir_dominance_lca(lca, use_instr->block); + } + } + + nir_foreach_if_use(def, use_src) { + nir_if *if_stmt = use_src->parent_if; + + /* For if statements, we consider the block to be the one immediately + * preceding the if CF node. + */ + nir_block *pred_block = + nir_cf_node_as_block(nir_cf_node_prev(&if_stmt->cf_node)); + + lca = nir_dominance_lca(lca, pred_block); + } + + /* Some instructions may never be used. We'll just leave them scheduled + * early and let dead code clean them up. + */ + if (lca == NULL) + return true; + + /* We know have the LCA of all of the uses. If our invariants hold, + * this is dominated by the block that we chose when scheduling early. + * We now walk up the dominance tree and pick the lowest block that is + * as far outside loops as we can get. + */ + nir_block *best = lca; + while (lca != def->parent_instr->block) { + assert(lca); + if (state->blocks[lca->index].loop_depth < + state->blocks[best->index].loop_depth) + best = lca; + lca = lca->imm_dom; + } + def->parent_instr->block = best; + + return true; +} + +/** Schedules an instruction late + * + * This function performs a depth-first search starting at the given + * instruction and proceeding through its uses to schedule instructions as + * late as they can reasonably go in the dominance tree. The instructions + * are "scheduled" by updating their instr->block field. + * + * The name of this function is actually a bit of a misnomer as it doesn't + * schedule them "as late as possible" as the paper implies. Instead, it + * first finds the lates possible place it can schedule the instruction and + * then possibly schedules it earlier than that. The actual location is as + * far down the tree as we can go while trying to stay out of loops. + */ +static void +gcm_schedule_late_instr(nir_instr *instr, struct gcm_state *state) +{ + if (instr->pass_flags & GCM_INSTR_SCHEDULED_LATE) + return; + + instr->pass_flags |= GCM_INSTR_SCHEDULED_LATE; + + /* Pinned instructions are already scheduled so we don't need to do + * anything. Also, bailing here keeps us from ever following phi nodes + * which can be back-edges. + */ + if (instr->pass_flags & GCM_INSTR_PINNED) + return; + + nir_foreach_ssa_def(instr, gcm_schedule_late_def, state); +} + +static void +gcm_place_instr(nir_instr *instr, struct gcm_state *state); + +static bool +gcm_place_instr_def(nir_ssa_def *def, void *state) +{ + nir_foreach_use(def, use_src) + gcm_place_instr(use_src->parent_instr, state); + + return false; +} + +/** Places an instrution back into the program + * + * The earlier passes of GCM simply choose blocks for each instruction and + * otherwise leave them alone. This pass actually places the instructions + * into their chosen blocks. + * + * To do so, we use a standard post-order depth-first search linearization + * algorithm. We walk over the uses of the given instruction and ensure + * that they are placed and then place this instruction. Because we are + * working on multiple blocks at a time, we keep track of the last inserted + * instruction per-block in the state structure's block_info array. When + * we insert an instruction in a block we insert it before the last + * instruction inserted in that block rather than the last instruction + * inserted globally. + */ +static void +gcm_place_instr(nir_instr *instr, struct gcm_state *state) +{ + if (instr->pass_flags & GCM_INSTR_PLACED) + return; + + instr->pass_flags |= GCM_INSTR_PLACED; + + /* Phi nodes are our once source of back-edges. Since right now we are + * only doing scheduling within blocks, we don't need to worry about + * them since they are always at the top. Just skip them completely. + */ + if (instr->type == nir_instr_type_phi) { + assert(instr->pass_flags & GCM_INSTR_PINNED); + return; + } + + nir_foreach_ssa_def(instr, gcm_place_instr_def, state); + + if (instr->pass_flags & GCM_INSTR_PINNED) { + /* Pinned instructions have an implicit dependence on the pinned + * instructions that come after them in the block. Since the pinned + * instructions will naturally "chain" together, we only need to + * explicitly visit one of them. + */ + for (nir_instr *after = nir_instr_next(instr); + after; + after = nir_instr_next(after)) { + if (after->pass_flags & GCM_INSTR_PINNED) { + gcm_place_instr(after, state); + break; + } + } + } + + struct gcm_block_info *block_info = &state->blocks[instr->block->index]; + if (!(instr->pass_flags & GCM_INSTR_PINNED)) { + exec_node_remove(&instr->node); + + if (block_info->last_instr) { + exec_node_insert_node_before(&block_info->last_instr->node, + &instr->node); + } else { + /* Schedule it at the end of the block */ + nir_instr *jump_instr = nir_block_last_instr(instr->block); + if (jump_instr && jump_instr->type == nir_instr_type_jump) { + exec_node_insert_node_before(&jump_instr->node, &instr->node); + } else { + exec_list_push_tail(&instr->block->instr_list, &instr->node); + } + } + } + + block_info->last_instr = instr; +} + +static void +opt_gcm_impl(nir_function_impl *impl) +{ + struct gcm_state state; + + state.impl = impl; + state.instr = NULL; + exec_list_make_empty(&state.instrs); + state.blocks = rzalloc_array(NULL, struct gcm_block_info, impl->num_blocks); + + nir_metadata_require(impl, nir_metadata_block_index | + nir_metadata_dominance); + + gcm_build_block_info(&impl->body, &state, 0); + nir_foreach_block(impl, gcm_pin_instructions_block, &state); + + foreach_list_typed(nir_instr, instr, node, &state.instrs) + gcm_schedule_early_instr(instr, &state); + + foreach_list_typed(nir_instr, instr, node, &state.instrs) + gcm_schedule_late_instr(instr, &state); + + while (!exec_list_is_empty(&state.instrs)) { + nir_instr *instr = exec_node_data(nir_instr, + state.instrs.tail_pred, node); + gcm_place_instr(instr, &state); + } + + ralloc_free(state.blocks); +} + +void +nir_opt_gcm(nir_shader *shader) +{ + nir_foreach_function(shader, function) { + if (function->impl) + opt_gcm_impl(function->impl); + } +} diff --git a/src/compiler/nir/nir_opt_global_to_local.c b/src/compiler/nir/nir_opt_global_to_local.c new file mode 100644 index 0000000..bccb45b --- /dev/null +++ b/src/compiler/nir/nir_opt_global_to_local.c @@ -0,0 +1,102 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +#include "nir.h" + +static bool +global_to_local(nir_register *reg) +{ + nir_function_impl *impl = NULL; + + assert(reg->is_global); + + nir_foreach_def(reg, def_dest) { + nir_instr *instr = def_dest->reg.parent_instr; + nir_function_impl *instr_impl = + nir_cf_node_get_function(&instr->block->cf_node); + if (impl != NULL) { + if (impl != instr_impl) + return false; + } else { + impl = instr_impl; + } + } + + nir_foreach_use(reg, use_src) { + nir_instr *instr = use_src->parent_instr; + nir_function_impl *instr_impl = + nir_cf_node_get_function(&instr->block->cf_node); + if (impl != NULL) { + if (impl != instr_impl) + return false; + } else { + impl = instr_impl; + } + } + + nir_foreach_if_use(reg, use_src) { + nir_if *if_stmt = use_src->parent_if; + nir_function_impl *if_impl = nir_cf_node_get_function(&if_stmt->cf_node); + if (impl != NULL) { + if (impl != if_impl) + return false; + } else { + impl = if_impl; + } + } + + if (impl == NULL) { + /* this instruction is never used/defined, delete it */ + nir_reg_remove(reg); + return true; + } + + /* + * if we've gotten to this point, the register is always used/defined in + * the same implementation so we can move it to be local to that + * implementation. + */ + + exec_node_remove(®->node); + exec_list_push_tail(&impl->registers, ®->node); + reg->index = impl->reg_alloc++; + reg->is_global = false; + return true; +} + +bool +nir_opt_global_to_local(nir_shader *shader) +{ + bool progress = false; + + foreach_list_typed_safe(nir_register, reg, node, &shader->registers) { + if (global_to_local(reg)) + progress = true; + } + + return progress; +} diff --git a/src/compiler/nir/nir_opt_peephole_select.c b/src/compiler/nir/nir_opt_peephole_select.c new file mode 100644 index 0000000..0fc658d --- /dev/null +++ b/src/compiler/nir/nir_opt_peephole_select.c @@ -0,0 +1,256 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand (jason@jlekstrand.net) + * + */ + +#include "nir.h" +#include "nir_control_flow.h" + +/* + * Implements a small peephole optimization that looks for + * + * if (cond) { + * <empty> + * } else { + * <empty> + * } + * phi + * ... + * phi + * + * and replaces it with a series of selects. It can also handle the case + * where, instead of being empty, the if may contain some move operations + * whose only use is one of the following phi nodes. This happens all the + * time when the SSA form comes from a conditional assignment with a + * swizzle. + */ + +struct peephole_select_state { + void *mem_ctx; + bool progress; +}; + +static bool +block_check_for_allowed_instrs(nir_block *block) +{ + nir_foreach_instr(block, instr) { + switch (instr->type) { + case nir_instr_type_intrinsic: { + nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); + + switch (intrin->intrinsic) { + case nir_intrinsic_load_var: + switch (intrin->variables[0]->var->data.mode) { + case nir_var_shader_in: + case nir_var_uniform: + break; + + default: + return false; + } + break; + + default: + return false; + } + + break; + } + + case nir_instr_type_load_const: + break; + + case nir_instr_type_alu: { + nir_alu_instr *mov = nir_instr_as_alu(instr); + switch (mov->op) { + case nir_op_fmov: + case nir_op_imov: + case nir_op_fneg: + case nir_op_ineg: + case nir_op_fabs: + case nir_op_iabs: + case nir_op_vec2: + case nir_op_vec3: + case nir_op_vec4: + /* It must be a move-like operation. */ + break; + default: + return false; + } + + /* Can't handle saturate */ + if (mov->dest.saturate) + return false; + + /* It must be SSA */ + if (!mov->dest.dest.is_ssa) + return false; + + /* It cannot have any if-uses */ + if (!list_empty(&mov->dest.dest.ssa.if_uses)) + return false; + + /* The only uses of this definition must be phi's in the successor */ + nir_foreach_use(&mov->dest.dest.ssa, use) { + if (use->parent_instr->type != nir_instr_type_phi || + use->parent_instr->block != block->successors[0]) + return false; + } + break; + } + + default: + return false; + } + } + + return true; +} + +static bool +nir_opt_peephole_select_block(nir_block *block, void *void_state) +{ + struct peephole_select_state *state = void_state; + + /* If the block is empty, then it certainly doesn't have any phi nodes, + * so we can skip it. This also ensures that we do an early skip on the + * end block of the function which isn't actually attached to the CFG. + */ + if (exec_list_is_empty(&block->instr_list)) + return true; + + if (nir_cf_node_is_first(&block->cf_node)) + return true; + + nir_cf_node *prev_node = nir_cf_node_prev(&block->cf_node); + if (prev_node->type != nir_cf_node_if) + return true; + + nir_if *if_stmt = nir_cf_node_as_if(prev_node); + nir_cf_node *then_node = nir_if_first_then_node(if_stmt); + nir_cf_node *else_node = nir_if_first_else_node(if_stmt); + + /* We can only have one block in each side ... */ + if (nir_if_last_then_node(if_stmt) != then_node || + nir_if_last_else_node(if_stmt) != else_node) + return true; + + nir_block *then_block = nir_cf_node_as_block(then_node); + nir_block *else_block = nir_cf_node_as_block(else_node); + + /* ... and those blocks must only contain "allowed" instructions. */ + if (!block_check_for_allowed_instrs(then_block) || + !block_check_for_allowed_instrs(else_block)) + return true; + + /* At this point, we know that the previous CFG node is an if-then + * statement containing only moves to phi nodes in this block. We can + * just remove that entire CF node and replace all of the phi nodes with + * selects. + */ + + nir_block *prev_block = nir_cf_node_as_block(nir_cf_node_prev(prev_node)); + assert(prev_block->cf_node.type == nir_cf_node_block); + + /* First, we move the remaining instructions from the blocks to the + * block before. We have already guaranteed that this is safe by + * calling block_check_for_allowed_instrs() + */ + nir_foreach_instr_safe(then_block, instr) { + exec_node_remove(&instr->node); + instr->block = prev_block; + exec_list_push_tail(&prev_block->instr_list, &instr->node); + } + + nir_foreach_instr_safe(else_block, instr) { + exec_node_remove(&instr->node); + instr->block = prev_block; + exec_list_push_tail(&prev_block->instr_list, &instr->node); + } + + nir_foreach_instr_safe(block, instr) { + if (instr->type != nir_instr_type_phi) + break; + + nir_phi_instr *phi = nir_instr_as_phi(instr); + nir_alu_instr *sel = nir_alu_instr_create(state->mem_ctx, nir_op_bcsel); + nir_src_copy(&sel->src[0].src, &if_stmt->condition, sel); + /* Splat the condition to all channels */ + memset(sel->src[0].swizzle, 0, sizeof sel->src[0].swizzle); + + assert(exec_list_length(&phi->srcs) == 2); + nir_foreach_phi_src(phi, src) { + assert(src->pred == then_block || src->pred == else_block); + assert(src->src.is_ssa); + + unsigned idx = src->pred == then_block ? 1 : 2; + nir_src_copy(&sel->src[idx].src, &src->src, sel); + } + + nir_ssa_dest_init(&sel->instr, &sel->dest.dest, + phi->dest.ssa.num_components, phi->dest.ssa.name); + sel->dest.write_mask = (1 << phi->dest.ssa.num_components) - 1; + + nir_ssa_def_rewrite_uses(&phi->dest.ssa, + nir_src_for_ssa(&sel->dest.dest.ssa)); + + nir_instr_insert_before(&phi->instr, &sel->instr); + nir_instr_remove(&phi->instr); + } + + nir_cf_node_remove(&if_stmt->cf_node); + state->progress = true; + + return true; +} + +static bool +nir_opt_peephole_select_impl(nir_function_impl *impl) +{ + struct peephole_select_state state; + + state.mem_ctx = ralloc_parent(impl); + state.progress = false; + + nir_foreach_block(impl, nir_opt_peephole_select_block, &state); + + if (state.progress) + nir_metadata_preserve(impl, nir_metadata_none); + + return state.progress; +} + +bool +nir_opt_peephole_select(nir_shader *shader) +{ + bool progress = false; + + nir_foreach_function(shader, function) { + if (function->impl) + progress |= nir_opt_peephole_select_impl(function->impl); + } + + return progress; +} diff --git a/src/compiler/nir/nir_opt_remove_phis.c b/src/compiler/nir/nir_opt_remove_phis.c new file mode 100644 index 0000000..6461837 --- /dev/null +++ b/src/compiler/nir/nir_opt_remove_phis.c @@ -0,0 +1,130 @@ +/* + * Copyright © 2015 Connor Abbott + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +#include "nir.h" + +/* + * This is a pass for removing phi nodes that look like: + * a = phi(b, b, b, ...) + * + * Note that we can't ignore undef sources here, or else we may create a + * situation where the definition of b isn't dominated by its uses. We're + * allowed to do this since the definition of b must dominate all of the + * phi node's predecessors, which means it must dominate the phi node as well + * as all of the phi node's uses. In essence, the phi node acts as a copy + * instruction. b can't be another phi node in the same block, since the only + * time when phi nodes can source other phi nodes defined in the same block is + * at the loop header, and in that case one of the sources of the phi has to + * be from before the loop and that source can't be b. + */ + +static bool +remove_phis_block(nir_block *block, void *state) +{ + bool *progress = state; + + nir_foreach_instr_safe(block, instr) { + if (instr->type != nir_instr_type_phi) + break; + + nir_phi_instr *phi = nir_instr_as_phi(instr); + + nir_ssa_def *def = NULL; + bool srcs_same = true; + + nir_foreach_phi_src(phi, src) { + assert(src->src.is_ssa); + + /* For phi nodes at the beginning of loops, we may encounter some + * sources from backedges that point back to the destination of the + * same phi, i.e. something like: + * + * a = phi(a, b, ...) + * + * We can safely ignore these sources, since if all of the normal + * sources point to the same definition, then that definition must + * still dominate the phi node, and the phi will still always take + * the value of that definition. + */ + if (src->src.ssa == &phi->dest.ssa) + continue; + + if (def == NULL) { + def = src->src.ssa; + } else { + if (src->src.ssa != def) { + srcs_same = false; + break; + } + } + } + + if (!srcs_same) + continue; + + /* We must have found at least one definition, since there must be at + * least one forward edge. + */ + assert(def != NULL); + + assert(phi->dest.is_ssa); + nir_ssa_def_rewrite_uses(&phi->dest.ssa, nir_src_for_ssa(def)); + nir_instr_remove(instr); + + *progress = true; + } + + return true; +} + +static bool +remove_phis_impl(nir_function_impl *impl) +{ + bool progress = false; + + nir_foreach_block(impl, remove_phis_block, &progress); + + if (progress) { + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance); + } + + return progress; +} + +bool +nir_opt_remove_phis(nir_shader *shader) +{ + bool progress = false; + + nir_foreach_function(shader, function) + if (function->impl) + progress = remove_phis_impl(function->impl) || progress; + + return progress; +} + diff --git a/src/compiler/nir/nir_opt_undef.c b/src/compiler/nir/nir_opt_undef.c new file mode 100644 index 0000000..374564d --- /dev/null +++ b/src/compiler/nir/nir_opt_undef.c @@ -0,0 +1,104 @@ +/* + * Copyright © 2015 Broadcom + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + */ + +#include "nir.h" + +/** @file nir_opt_undef.c + * + * Handles optimization of operations involving ssa_undef. For now, we just + * make sure that csels between undef and some other value just give the other + * value (on the assumption that the condition's going to be choosing the + * defined value). This reduces work after if flattening when each side of + * the if is defining a variable. + * + * Some day, we may find some use for making other operations consuming an + * undef arg output undef, but I don't know of any cases currently. + */ + +static bool +opt_undef_alu(nir_alu_instr *instr) +{ + if (instr->op != nir_op_bcsel && instr->op != nir_op_fcsel) + return false; + + assert(instr->dest.dest.is_ssa); + + for (int i = 1; i <= 2; i++) { + if (!instr->src[i].src.is_ssa) + continue; + + nir_instr *parent = instr->src[i].src.ssa->parent_instr; + if (parent->type != nir_instr_type_ssa_undef) + continue; + + /* We can't just use nir_alu_src_copy, because we need the def/use + * updated. + */ + nir_instr_rewrite_src(&instr->instr, &instr->src[0].src, + instr->src[i == 1 ? 2 : 1].src); + nir_alu_src_copy(&instr->src[0], &instr->src[i == 1 ? 2 : 1], + ralloc_parent(instr)); + + nir_src empty_src; + memset(&empty_src, 0, sizeof(empty_src)); + nir_instr_rewrite_src(&instr->instr, &instr->src[1].src, empty_src); + nir_instr_rewrite_src(&instr->instr, &instr->src[2].src, empty_src); + instr->op = nir_op_imov; + + return true; + } + + return false; +} + +static bool +opt_undef_block(nir_block *block, void *data) +{ + bool *progress = data; + + nir_foreach_instr_safe(block, instr) { + if (instr->type == nir_instr_type_alu) + if (opt_undef_alu(nir_instr_as_alu(instr))) + (*progress) = true; + } + + return true; +} + +bool +nir_opt_undef(nir_shader *shader) +{ + bool progress = false; + + nir_foreach_function(shader, function) { + if (function->impl) { + nir_foreach_block(function->impl, opt_undef_block, &progress); + if (progress) + nir_metadata_preserve(function->impl, + nir_metadata_block_index | + nir_metadata_dominance); + } + } + + return progress; +} diff --git a/src/compiler/nir/nir_print.c b/src/compiler/nir/nir_print.c new file mode 100644 index 0000000..48ecb48 --- /dev/null +++ b/src/compiler/nir/nir_print.c @@ -0,0 +1,1069 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +#include "nir.h" +#include "compiler/shader_enums.h" +#include <stdio.h> +#include <stdlib.h> + +static void +print_tabs(unsigned num_tabs, FILE *fp) +{ + for (unsigned i = 0; i < num_tabs; i++) + fprintf(fp, "\t"); +} + +typedef struct { + FILE *fp; + nir_shader *shader; + /** map from nir_variable -> printable name */ + struct hash_table *ht; + + /** set of names used so far for nir_variables */ + struct set *syms; + + /* an index used to make new non-conflicting names */ + unsigned index; +} print_state; + +static void +print_register(nir_register *reg, print_state *state) +{ + FILE *fp = state->fp; + if (reg->name != NULL) + fprintf(fp, "/* %s */ ", reg->name); + if (reg->is_global) + fprintf(fp, "gr%u", reg->index); + else + fprintf(fp, "r%u", reg->index); +} + +static const char *sizes[] = { "error", "vec1", "vec2", "vec3", "vec4" }; + +static void +print_register_decl(nir_register *reg, print_state *state) +{ + FILE *fp = state->fp; + fprintf(fp, "decl_reg %s ", sizes[reg->num_components]); + if (reg->is_packed) + fprintf(fp, "(packed) "); + print_register(reg, state); + if (reg->num_array_elems != 0) + fprintf(fp, "[%u]", reg->num_array_elems); + fprintf(fp, "\n"); +} + +static void +print_ssa_def(nir_ssa_def *def, print_state *state) +{ + FILE *fp = state->fp; + if (def->name != NULL) + fprintf(fp, "/* %s */ ", def->name); + fprintf(fp, "%s ssa_%u", sizes[def->num_components], def->index); +} + +static void +print_ssa_use(nir_ssa_def *def, print_state *state) +{ + FILE *fp = state->fp; + if (def->name != NULL) + fprintf(fp, "/* %s */ ", def->name); + fprintf(fp, "ssa_%u", def->index); +} + +static void print_src(nir_src *src, print_state *state); + +static void +print_reg_src(nir_reg_src *src, print_state *state) +{ + FILE *fp = state->fp; + print_register(src->reg, state); + if (src->reg->num_array_elems != 0) { + fprintf(fp, "[%u", src->base_offset); + if (src->indirect != NULL) { + fprintf(fp, " + "); + print_src(src->indirect, state); + } + fprintf(fp, "]"); + } +} + +static void +print_reg_dest(nir_reg_dest *dest, print_state *state) +{ + FILE *fp = state->fp; + print_register(dest->reg, state); + if (dest->reg->num_array_elems != 0) { + fprintf(fp, "[%u", dest->base_offset); + if (dest->indirect != NULL) { + fprintf(fp, " + "); + print_src(dest->indirect, state); + } + fprintf(fp, "]"); + } +} + +static void +print_src(nir_src *src, print_state *state) +{ + if (src->is_ssa) + print_ssa_use(src->ssa, state); + else + print_reg_src(&src->reg, state); +} + +static void +print_dest(nir_dest *dest, print_state *state) +{ + if (dest->is_ssa) + print_ssa_def(&dest->ssa, state); + else + print_reg_dest(&dest->reg, state); +} + +static void +print_alu_src(nir_alu_instr *instr, unsigned src, print_state *state) +{ + FILE *fp = state->fp; + + if (instr->src[src].negate) + fprintf(fp, "-"); + if (instr->src[src].abs) + fprintf(fp, "abs("); + + print_src(&instr->src[src].src, state); + + bool print_swizzle = false; + for (unsigned i = 0; i < 4; i++) { + if (!nir_alu_instr_channel_used(instr, src, i)) + continue; + + if (instr->src[src].swizzle[i] != i) { + print_swizzle = true; + break; + } + } + + if (print_swizzle) { + fprintf(fp, "."); + for (unsigned i = 0; i < 4; i++) { + if (!nir_alu_instr_channel_used(instr, src, i)) + continue; + + fprintf(fp, "%c", "xyzw"[instr->src[src].swizzle[i]]); + } + } + + if (instr->src[src].abs) + fprintf(fp, ")"); +} + +static void +print_alu_dest(nir_alu_dest *dest, print_state *state) +{ + FILE *fp = state->fp; + /* we're going to print the saturate modifier later, after the opcode */ + + print_dest(&dest->dest, state); + + if (!dest->dest.is_ssa && + dest->write_mask != (1 << dest->dest.reg.reg->num_components) - 1) { + fprintf(fp, "."); + for (unsigned i = 0; i < 4; i++) + if ((dest->write_mask >> i) & 1) + fprintf(fp, "%c", "xyzw"[i]); + } +} + +static void +print_alu_instr(nir_alu_instr *instr, print_state *state) +{ + FILE *fp = state->fp; + + print_alu_dest(&instr->dest, state); + + fprintf(fp, " = %s", nir_op_infos[instr->op].name); + if (instr->dest.saturate) + fprintf(fp, ".sat"); + fprintf(fp, " "); + + for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) { + if (i != 0) + fprintf(fp, ", "); + + print_alu_src(instr, i, state); + } +} + +static void +print_constant(nir_constant *c, const struct glsl_type *type, print_state *state) +{ + FILE *fp = state->fp; + unsigned total_elems = glsl_get_components(type); + unsigned i; + + switch (glsl_get_base_type(type)) { + case GLSL_TYPE_UINT: + case GLSL_TYPE_INT: + case GLSL_TYPE_BOOL: + for (i = 0; i < total_elems; i++) { + if (i > 0) fprintf(fp, ", "); + fprintf(fp, "0x%08x", c->value.u[i]); + } + break; + + case GLSL_TYPE_FLOAT: + for (i = 0; i < total_elems; i++) { + if (i > 0) fprintf(fp, ", "); + fprintf(fp, "%f", c->value.f[i]); + } + break; + + case GLSL_TYPE_STRUCT: + for (i = 0; i < c->num_elements; i++) { + if (i > 0) fprintf(fp, ", "); + fprintf(fp, "{ "); + print_constant(c->elements[i], glsl_get_struct_field(type, i), state); + fprintf(fp, " }"); + } + break; + + case GLSL_TYPE_ARRAY: + for (i = 0; i < c->num_elements; i++) { + if (i > 0) fprintf(fp, ", "); + fprintf(fp, "{ "); + print_constant(c->elements[i], glsl_get_array_element(type), state); + fprintf(fp, " }"); + } + break; + + default: + unreachable("not reached"); + } +} + +static void +print_var_decl(nir_variable *var, print_state *state) +{ + FILE *fp = state->fp; + + fprintf(fp, "decl_var "); + + const char *const cent = (var->data.centroid) ? "centroid " : ""; + const char *const samp = (var->data.sample) ? "sample " : ""; + const char *const patch = (var->data.patch) ? "patch " : ""; + const char *const inv = (var->data.invariant) ? "invariant " : ""; + const char *const mode[] = { "shader_in ", "shader_out ", "", "", + "uniform ", "shader_storage", "system " }; + + fprintf(fp, "%s%s%s%s%s%s ", + cent, samp, patch, inv, mode[var->data.mode], + glsl_interp_qualifier_name(var->data.interpolation)); + + glsl_print_type(var->type, fp); + + struct set_entry *entry = NULL; + if (state->syms) + entry = _mesa_set_search(state->syms, var->name); + + char *name; + + if (entry != NULL) { + /* we have a collision with another name, append an @ + a unique index */ + name = ralloc_asprintf(state->syms, "%s@%u", var->name, state->index++); + } else { + name = var->name; + } + + fprintf(fp, " %s", name); + + if (var->data.mode == nir_var_shader_in || + var->data.mode == nir_var_shader_out || + var->data.mode == nir_var_uniform || + var->data.mode == nir_var_shader_storage) { + const char *loc = NULL; + char buf[4]; + + switch (state->shader->stage) { + case MESA_SHADER_VERTEX: + if (var->data.mode == nir_var_shader_in) + loc = gl_vert_attrib_name(var->data.location); + else if (var->data.mode == nir_var_shader_out) + loc = gl_varying_slot_name(var->data.location); + break; + case MESA_SHADER_GEOMETRY: + if ((var->data.mode == nir_var_shader_in) || + (var->data.mode == nir_var_shader_out)) + loc = gl_varying_slot_name(var->data.location); + break; + case MESA_SHADER_FRAGMENT: + if (var->data.mode == nir_var_shader_in) + loc = gl_varying_slot_name(var->data.location); + else if (var->data.mode == nir_var_shader_out) + loc = gl_frag_result_name(var->data.location); + break; + case MESA_SHADER_TESS_CTRL: + case MESA_SHADER_TESS_EVAL: + case MESA_SHADER_COMPUTE: + default: + /* TODO */ + break; + } + + if (!loc) { + snprintf(buf, sizeof(buf), "%u", var->data.location); + loc = buf; + } + + fprintf(fp, " (%s, %u)", loc, var->data.driver_location); + } + + if (var->constant_initializer) { + fprintf(fp, " = { "); + print_constant(var->constant_initializer, var->type, state); + fprintf(fp, " }"); + } + + fprintf(fp, "\n"); + + if (state->syms) { + _mesa_set_add(state->syms, name); + _mesa_hash_table_insert(state->ht, var, name); + } +} + +static void +print_var(nir_variable *var, print_state *state) +{ + FILE *fp = state->fp; + const char *name; + if (state->ht) { + struct hash_entry *entry = _mesa_hash_table_search(state->ht, var); + + assert(entry != NULL); + name = entry->data; + } else { + name = var->name; + } + + fprintf(fp, "%s", name); +} + +static void +print_deref_var(nir_deref_var *deref, print_state *state) +{ + print_var(deref->var, state); +} + +static void +print_deref_array(nir_deref_array *deref, print_state *state) +{ + FILE *fp = state->fp; + fprintf(fp, "["); + switch (deref->deref_array_type) { + case nir_deref_array_type_direct: + fprintf(fp, "%u", deref->base_offset); + break; + case nir_deref_array_type_indirect: + if (deref->base_offset != 0) + fprintf(fp, "%u + ", deref->base_offset); + print_src(&deref->indirect, state); + break; + case nir_deref_array_type_wildcard: + fprintf(fp, "*"); + break; + } + fprintf(fp, "]"); +} + +static void +print_deref_struct(nir_deref_struct *deref, const struct glsl_type *parent_type, + print_state *state) +{ + FILE *fp = state->fp; + fprintf(fp, ".%s", glsl_get_struct_elem_name(parent_type, deref->index)); +} + +static void +print_deref(nir_deref_var *deref, print_state *state) +{ + nir_deref *tail = &deref->deref; + nir_deref *pretail = NULL; + while (tail != NULL) { + switch (tail->deref_type) { + case nir_deref_type_var: + assert(pretail == NULL); + assert(tail == &deref->deref); + print_deref_var(deref, state); + break; + + case nir_deref_type_array: + assert(pretail != NULL); + print_deref_array(nir_deref_as_array(tail), state); + break; + + case nir_deref_type_struct: + assert(pretail != NULL); + print_deref_struct(nir_deref_as_struct(tail), + pretail->type, state); + break; + + default: + unreachable("Invalid deref type"); + } + + pretail = tail; + tail = pretail->child; + } +} + +static void +print_intrinsic_instr(nir_intrinsic_instr *instr, print_state *state) +{ + unsigned num_srcs = nir_intrinsic_infos[instr->intrinsic].num_srcs; + FILE *fp = state->fp; + + if (nir_intrinsic_infos[instr->intrinsic].has_dest) { + print_dest(&instr->dest, state); + fprintf(fp, " = "); + } + + fprintf(fp, "intrinsic %s (", nir_intrinsic_infos[instr->intrinsic].name); + + for (unsigned i = 0; i < num_srcs; i++) { + if (i != 0) + fprintf(fp, ", "); + + print_src(&instr->src[i], state); + } + + fprintf(fp, ") ("); + + unsigned num_vars = nir_intrinsic_infos[instr->intrinsic].num_variables; + + for (unsigned i = 0; i < num_vars; i++) { + if (i != 0) + fprintf(fp, ", "); + + print_deref(instr->variables[i], state); + } + + fprintf(fp, ") ("); + + unsigned num_indices = nir_intrinsic_infos[instr->intrinsic].num_indices; + + for (unsigned i = 0; i < num_indices; i++) { + if (i != 0) + fprintf(fp, ", "); + + fprintf(fp, "%d", instr->const_index[i]); + } + + fprintf(fp, ")"); + + if (!state->shader) + return; + + struct exec_list *var_list = NULL; + + switch (instr->intrinsic) { + case nir_intrinsic_load_uniform: + var_list = &state->shader->uniforms; + break; + case nir_intrinsic_load_input: + case nir_intrinsic_load_per_vertex_input: + var_list = &state->shader->inputs; + break; + case nir_intrinsic_load_output: + case nir_intrinsic_store_output: + case nir_intrinsic_store_per_vertex_output: + var_list = &state->shader->outputs; + break; + default: + return; + } + + nir_foreach_variable(var, var_list) { + if ((var->data.driver_location == instr->const_index[0]) && + var->name) { + fprintf(fp, "\t/* %s */", var->name); + break; + } + } +} + +static void +print_tex_instr(nir_tex_instr *instr, print_state *state) +{ + FILE *fp = state->fp; + + print_dest(&instr->dest, state); + + fprintf(fp, " = "); + + switch (instr->op) { + case nir_texop_tex: + fprintf(fp, "tex "); + break; + case nir_texop_txb: + fprintf(fp, "txb "); + break; + case nir_texop_txl: + fprintf(fp, "txl "); + break; + case nir_texop_txd: + fprintf(fp, "txd "); + break; + case nir_texop_txf: + fprintf(fp, "txf "); + break; + case nir_texop_txf_ms: + fprintf(fp, "txf_ms "); + break; + case nir_texop_txs: + fprintf(fp, "txs "); + break; + case nir_texop_lod: + fprintf(fp, "lod "); + break; + case nir_texop_tg4: + fprintf(fp, "tg4 "); + break; + case nir_texop_query_levels: + fprintf(fp, "query_levels "); + break; + case nir_texop_texture_samples: + fprintf(fp, "texture_samples "); + break; + case nir_texop_samples_identical: + fprintf(fp, "samples_identical "); + break; + default: + unreachable("Invalid texture operation"); + break; + } + + for (unsigned i = 0; i < instr->num_srcs; i++) { + print_src(&instr->src[i].src, state); + + fprintf(fp, " "); + + switch(instr->src[i].src_type) { + case nir_tex_src_coord: + fprintf(fp, "(coord)"); + break; + case nir_tex_src_projector: + fprintf(fp, "(projector)"); + break; + case nir_tex_src_comparitor: + fprintf(fp, "(comparitor)"); + break; + case nir_tex_src_offset: + fprintf(fp, "(offset)"); + break; + case nir_tex_src_bias: + fprintf(fp, "(bias)"); + break; + case nir_tex_src_lod: + fprintf(fp, "(lod)"); + break; + case nir_tex_src_ms_index: + fprintf(fp, "(ms_index)"); + break; + case nir_tex_src_ddx: + fprintf(fp, "(ddx)"); + break; + case nir_tex_src_ddy: + fprintf(fp, "(ddy)"); + break; + case nir_tex_src_sampler_offset: + fprintf(fp, "(sampler_offset)"); + break; + + default: + unreachable("Invalid texture source type"); + break; + } + + fprintf(fp, ", "); + } + + bool has_nonzero_offset = false; + for (unsigned i = 0; i < 4; i++) { + if (instr->const_offset[i] != 0) { + has_nonzero_offset = true; + break; + } + } + + if (has_nonzero_offset) { + fprintf(fp, "[%i %i %i %i] (offset), ", + instr->const_offset[0], instr->const_offset[1], + instr->const_offset[2], instr->const_offset[3]); + } + + if (instr->op == nir_texop_tg4) { + fprintf(fp, "%u (gather_component), ", instr->component); + } + + if (instr->sampler) { + print_deref(instr->sampler, state); + } else { + fprintf(fp, "%u", instr->sampler_index); + } + + fprintf(fp, " (sampler)"); +} + +static void +print_call_instr(nir_call_instr *instr, print_state *state) +{ + FILE *fp = state->fp; + + fprintf(fp, "call %s ", instr->callee->name); + + for (unsigned i = 0; i < instr->num_params; i++) { + if (i != 0) + fprintf(fp, ", "); + + print_deref(instr->params[i], state); + } + + if (instr->return_deref != NULL) { + if (instr->num_params != 0) + fprintf(fp, ", "); + fprintf(fp, "returning "); + print_deref(instr->return_deref, state); + } +} + +static void +print_load_const_instr(nir_load_const_instr *instr, print_state *state) +{ + FILE *fp = state->fp; + + print_ssa_def(&instr->def, state); + + fprintf(fp, " = load_const ("); + + for (unsigned i = 0; i < instr->def.num_components; i++) { + if (i != 0) + fprintf(fp, ", "); + + /* + * we don't really know the type of the constant (if it will be used as a + * float or an int), so just print the raw constant in hex for fidelity + * and then print the float in a comment for readability. + */ + + fprintf(fp, "0x%08x /* %f */", instr->value.u[i], instr->value.f[i]); + } + + fprintf(fp, ")"); +} + +static void +print_jump_instr(nir_jump_instr *instr, print_state *state) +{ + FILE *fp = state->fp; + + switch (instr->type) { + case nir_jump_break: + fprintf(fp, "break"); + break; + + case nir_jump_continue: + fprintf(fp, "continue"); + break; + + case nir_jump_return: + fprintf(fp, "return"); + break; + } +} + +static void +print_ssa_undef_instr(nir_ssa_undef_instr* instr, print_state *state) +{ + FILE *fp = state->fp; + print_ssa_def(&instr->def, state); + fprintf(fp, " = undefined"); +} + +static void +print_phi_instr(nir_phi_instr *instr, print_state *state) +{ + FILE *fp = state->fp; + print_dest(&instr->dest, state); + fprintf(fp, " = phi "); + nir_foreach_phi_src(instr, src) { + if (&src->node != exec_list_get_head(&instr->srcs)) + fprintf(fp, ", "); + + fprintf(fp, "block_%u: ", src->pred->index); + print_src(&src->src, state); + } +} + +static void +print_parallel_copy_instr(nir_parallel_copy_instr *instr, print_state *state) +{ + FILE *fp = state->fp; + nir_foreach_parallel_copy_entry(instr, entry) { + if (&entry->node != exec_list_get_head(&instr->entries)) + fprintf(fp, "; "); + + print_dest(&entry->dest, state); + fprintf(fp, " = "); + print_src(&entry->src, state); + } +} + +static void +print_instr(const nir_instr *instr, print_state *state, unsigned tabs) +{ + FILE *fp = state->fp; + print_tabs(tabs, fp); + + switch (instr->type) { + case nir_instr_type_alu: + print_alu_instr(nir_instr_as_alu(instr), state); + break; + + case nir_instr_type_call: + print_call_instr(nir_instr_as_call(instr), state); + break; + + case nir_instr_type_intrinsic: + print_intrinsic_instr(nir_instr_as_intrinsic(instr), state); + break; + + case nir_instr_type_tex: + print_tex_instr(nir_instr_as_tex(instr), state); + break; + + case nir_instr_type_load_const: + print_load_const_instr(nir_instr_as_load_const(instr), state); + break; + + case nir_instr_type_jump: + print_jump_instr(nir_instr_as_jump(instr), state); + break; + + case nir_instr_type_ssa_undef: + print_ssa_undef_instr(nir_instr_as_ssa_undef(instr), state); + break; + + case nir_instr_type_phi: + print_phi_instr(nir_instr_as_phi(instr), state); + break; + + case nir_instr_type_parallel_copy: + print_parallel_copy_instr(nir_instr_as_parallel_copy(instr), state); + break; + + default: + unreachable("Invalid instruction type"); + break; + } +} + +static int +compare_block_index(const void *p1, const void *p2) +{ + const nir_block *block1 = *((const nir_block **) p1); + const nir_block *block2 = *((const nir_block **) p2); + + return (int) block1->index - (int) block2->index; +} + +static void print_cf_node(nir_cf_node *node, print_state *state, + unsigned tabs); + +static void +print_block(nir_block *block, print_state *state, unsigned tabs) +{ + FILE *fp = state->fp; + + print_tabs(tabs, fp); + fprintf(fp, "block block_%u:\n", block->index); + + /* sort the predecessors by index so we consistently print the same thing */ + + nir_block **preds = + malloc(block->predecessors->entries * sizeof(nir_block *)); + + struct set_entry *entry; + unsigned i = 0; + set_foreach(block->predecessors, entry) { + preds[i++] = (nir_block *) entry->key; + } + + qsort(preds, block->predecessors->entries, sizeof(nir_block *), + compare_block_index); + + print_tabs(tabs, fp); + fprintf(fp, "/* preds: "); + for (unsigned i = 0; i < block->predecessors->entries; i++) { + fprintf(fp, "block_%u ", preds[i]->index); + } + fprintf(fp, "*/\n"); + + free(preds); + + nir_foreach_instr(block, instr) { + print_instr(instr, state, tabs); + fprintf(fp, "\n"); + } + + print_tabs(tabs, fp); + fprintf(fp, "/* succs: "); + for (unsigned i = 0; i < 2; i++) + if (block->successors[i]) { + fprintf(fp, "block_%u ", block->successors[i]->index); + } + fprintf(fp, "*/\n"); +} + +static void +print_if(nir_if *if_stmt, print_state *state, unsigned tabs) +{ + FILE *fp = state->fp; + + print_tabs(tabs, fp); + fprintf(fp, "if "); + print_src(&if_stmt->condition, state); + fprintf(fp, " {\n"); + foreach_list_typed(nir_cf_node, node, node, &if_stmt->then_list) { + print_cf_node(node, state, tabs + 1); + } + print_tabs(tabs, fp); + fprintf(fp, "} else {\n"); + foreach_list_typed(nir_cf_node, node, node, &if_stmt->else_list) { + print_cf_node(node, state, tabs + 1); + } + print_tabs(tabs, fp); + fprintf(fp, "}\n"); +} + +static void +print_loop(nir_loop *loop, print_state *state, unsigned tabs) +{ + FILE *fp = state->fp; + + print_tabs(tabs, fp); + fprintf(fp, "loop {\n"); + foreach_list_typed(nir_cf_node, node, node, &loop->body) { + print_cf_node(node, state, tabs + 1); + } + print_tabs(tabs, fp); + fprintf(fp, "}\n"); +} + +static void +print_cf_node(nir_cf_node *node, print_state *state, unsigned int tabs) +{ + switch (node->type) { + case nir_cf_node_block: + print_block(nir_cf_node_as_block(node), state, tabs); + break; + + case nir_cf_node_if: + print_if(nir_cf_node_as_if(node), state, tabs); + break; + + case nir_cf_node_loop: + print_loop(nir_cf_node_as_loop(node), state, tabs); + break; + + default: + unreachable("Invalid CFG node type"); + } +} + +static void +print_function_impl(nir_function_impl *impl, print_state *state) +{ + FILE *fp = state->fp; + + fprintf(fp, "\nimpl %s ", impl->function->name); + + for (unsigned i = 0; i < impl->num_params; i++) { + if (i != 0) + fprintf(fp, ", "); + + print_var(impl->params[i], state); + } + + if (impl->return_var != NULL) { + if (impl->num_params != 0) + fprintf(fp, ", "); + fprintf(fp, "returning "); + print_var(impl->return_var, state); + } + + fprintf(fp, "{\n"); + + nir_foreach_variable(var, &impl->locals) { + fprintf(fp, "\t"); + print_var_decl(var, state); + } + + foreach_list_typed(nir_register, reg, node, &impl->registers) { + fprintf(fp, "\t"); + print_register_decl(reg, state); + } + + nir_index_blocks(impl); + + foreach_list_typed(nir_cf_node, node, node, &impl->body) { + print_cf_node(node, state, 1); + } + + fprintf(fp, "\tblock block_%u:\n}\n\n", impl->end_block->index); +} + +static void +print_function(nir_function *function, print_state *state) +{ + FILE *fp = state->fp; + + fprintf(fp, "decl_function %s ", function->name); + + for (unsigned i = 0; i < function->num_params; i++) { + if (i != 0) + fprintf(fp, ", "); + + switch (function->params[i].param_type) { + case nir_parameter_in: + fprintf(fp, "in "); + break; + case nir_parameter_out: + fprintf(fp, "out "); + break; + case nir_parameter_inout: + fprintf(fp, "inout "); + break; + default: + unreachable("Invalid parameter type"); + } + + glsl_print_type(function->params[i].type, fp); + } + + if (function->return_type != NULL) { + if (function->num_params != 0) + fprintf(fp, ", "); + fprintf(fp, "returning "); + glsl_print_type(function->return_type, fp); + } + + fprintf(fp, "\n"); + + if (function->impl != NULL) { + print_function_impl(function->impl, state); + return; + } +} + +static void +init_print_state(print_state *state, nir_shader *shader, FILE *fp) +{ + state->fp = fp; + state->shader = shader; + state->ht = _mesa_hash_table_create(NULL, _mesa_hash_pointer, + _mesa_key_pointer_equal); + state->syms = _mesa_set_create(NULL, _mesa_key_hash_string, + _mesa_key_string_equal); + state->index = 0; +} + +static void +destroy_print_state(print_state *state) +{ + _mesa_hash_table_destroy(state->ht, NULL); + _mesa_set_destroy(state->syms, NULL); +} + +void +nir_print_shader(nir_shader *shader, FILE *fp) +{ + print_state state; + init_print_state(&state, shader, fp); + + fprintf(fp, "shader: %s\n", gl_shader_stage_name(shader->stage)); + + if (shader->info.name) + fprintf(fp, "name: %s\n", shader->info.name); + + if (shader->info.label) + fprintf(fp, "label: %s\n", shader->info.label); + + fprintf(fp, "inputs: %u\n", shader->num_inputs); + fprintf(fp, "outputs: %u\n", shader->num_outputs); + fprintf(fp, "uniforms: %u\n", shader->num_uniforms); + + nir_foreach_variable(var, &shader->uniforms) { + print_var_decl(var, &state); + } + + nir_foreach_variable(var, &shader->inputs) { + print_var_decl(var, &state); + } + + nir_foreach_variable(var, &shader->outputs) { + print_var_decl(var, &state); + } + + nir_foreach_variable(var, &shader->globals) { + print_var_decl(var, &state); + } + + nir_foreach_variable(var, &shader->system_values) { + print_var_decl(var, &state); + } + + foreach_list_typed(nir_register, reg, node, &shader->registers) { + print_register_decl(reg, &state); + } + + foreach_list_typed(nir_function, func, node, &shader->functions) { + print_function(func, &state); + } + + destroy_print_state(&state); +} + +void +nir_print_instr(const nir_instr *instr, FILE *fp) +{ + print_state state = { + .fp = fp, + }; + print_instr(instr, &state, 0); + +} diff --git a/src/compiler/nir/nir_remove_dead_variables.c b/src/compiler/nir/nir_remove_dead_variables.c new file mode 100644 index 0000000..db754e5 --- /dev/null +++ b/src/compiler/nir/nir_remove_dead_variables.c @@ -0,0 +1,141 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +#include "nir.h" + +static void +add_var_use_intrinsic(nir_intrinsic_instr *instr, struct set *live) +{ + unsigned num_vars = nir_intrinsic_infos[instr->intrinsic].num_variables; + for (unsigned i = 0; i < num_vars; i++) { + nir_variable *var = instr->variables[i]->var; + _mesa_set_add(live, var); + } +} + +static void +add_var_use_call(nir_call_instr *instr, struct set *live) +{ + if (instr->return_deref != NULL) { + nir_variable *var = instr->return_deref->var; + _mesa_set_add(live, var); + } + + for (unsigned i = 0; i < instr->num_params; i++) { + nir_variable *var = instr->params[i]->var; + _mesa_set_add(live, var); + } +} + +static void +add_var_use_tex(nir_tex_instr *instr, struct set *live) +{ + if (instr->sampler != NULL) { + nir_variable *var = instr->sampler->var; + _mesa_set_add(live, var); + } +} + +static bool +add_var_use_block(nir_block *block, void *state) +{ + struct set *live = state; + + nir_foreach_instr(block, instr) { + switch(instr->type) { + case nir_instr_type_intrinsic: + add_var_use_intrinsic(nir_instr_as_intrinsic(instr), live); + break; + + case nir_instr_type_call: + add_var_use_call(nir_instr_as_call(instr), live); + break; + + case nir_instr_type_tex: + add_var_use_tex(nir_instr_as_tex(instr), live); + break; + + default: + break; + } + } + + return true; +} + +static void +add_var_use_shader(nir_shader *shader, struct set *live) +{ + nir_foreach_function(shader, function) { + if (function->impl) { + nir_foreach_block(function->impl, add_var_use_block, live); + } + } +} + +static bool +remove_dead_vars(struct exec_list *var_list, struct set *live) +{ + bool progress = false; + + foreach_list_typed_safe(nir_variable, var, node, var_list) { + struct set_entry *entry = _mesa_set_search(live, var); + if (entry == NULL) { + exec_node_remove(&var->node); + ralloc_free(var); + progress = true; + } + } + + return progress; +} + +bool +nir_remove_dead_variables(nir_shader *shader) +{ + bool progress = false; + struct set *live = + _mesa_set_create(NULL, _mesa_hash_pointer, _mesa_key_pointer_equal); + + add_var_use_shader(shader, live); + + progress = remove_dead_vars(&shader->globals, live) || progress; + + nir_foreach_function(shader, function) { + if (function->impl) { + if (remove_dead_vars(&function->impl->locals, live)) { + nir_metadata_preserve(function->impl, nir_metadata_block_index | + nir_metadata_dominance | + nir_metadata_live_ssa_defs); + progress = true; + } + } + } + + _mesa_set_destroy(live, NULL); + return progress; +} diff --git a/src/compiler/nir/nir_search.c b/src/compiler/nir/nir_search.c new file mode 100644 index 0000000..56d7e81 --- /dev/null +++ b/src/compiler/nir/nir_search.c @@ -0,0 +1,379 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand (jason@jlekstrand.net) + * + */ + +#include "nir_search.h" + +struct match_state { + unsigned variables_seen; + nir_alu_src variables[NIR_SEARCH_MAX_VARIABLES]; +}; + +static bool +match_expression(const nir_search_expression *expr, nir_alu_instr *instr, + unsigned num_components, const uint8_t *swizzle, + struct match_state *state); + +static const uint8_t identity_swizzle[] = { 0, 1, 2, 3 }; + +static bool alu_instr_is_bool(nir_alu_instr *instr); + +static bool +src_is_bool(nir_src src) +{ + if (!src.is_ssa) + return false; + if (src.ssa->parent_instr->type != nir_instr_type_alu) + return false; + return alu_instr_is_bool(nir_instr_as_alu(src.ssa->parent_instr)); +} + +static bool +alu_instr_is_bool(nir_alu_instr *instr) +{ + switch (instr->op) { + case nir_op_iand: + case nir_op_ior: + case nir_op_ixor: + return src_is_bool(instr->src[0].src) && src_is_bool(instr->src[1].src); + case nir_op_inot: + return src_is_bool(instr->src[0].src); + default: + return nir_op_infos[instr->op].output_type == nir_type_bool; + } +} + +static bool +match_value(const nir_search_value *value, nir_alu_instr *instr, unsigned src, + unsigned num_components, const uint8_t *swizzle, + struct match_state *state) +{ + uint8_t new_swizzle[4]; + + /* If the source is an explicitly sized source, then we need to reset + * both the number of components and the swizzle. + */ + if (nir_op_infos[instr->op].input_sizes[src] != 0) { + num_components = nir_op_infos[instr->op].input_sizes[src]; + swizzle = identity_swizzle; + } + + for (unsigned i = 0; i < num_components; ++i) + new_swizzle[i] = instr->src[src].swizzle[swizzle[i]]; + + switch (value->type) { + case nir_search_value_expression: + if (!instr->src[src].src.is_ssa) + return false; + + if (instr->src[src].src.ssa->parent_instr->type != nir_instr_type_alu) + return false; + + return match_expression(nir_search_value_as_expression(value), + nir_instr_as_alu(instr->src[src].src.ssa->parent_instr), + num_components, new_swizzle, state); + + case nir_search_value_variable: { + nir_search_variable *var = nir_search_value_as_variable(value); + assert(var->variable < NIR_SEARCH_MAX_VARIABLES); + + if (state->variables_seen & (1 << var->variable)) { + if (!nir_srcs_equal(state->variables[var->variable].src, + instr->src[src].src)) + return false; + + assert(!instr->src[src].abs && !instr->src[src].negate); + + for (unsigned i = 0; i < num_components; ++i) { + if (state->variables[var->variable].swizzle[i] != new_swizzle[i]) + return false; + } + + return true; + } else { + if (var->is_constant && + instr->src[src].src.ssa->parent_instr->type != nir_instr_type_load_const) + return false; + + if (var->type != nir_type_invalid) { + if (instr->src[src].src.ssa->parent_instr->type != nir_instr_type_alu) + return false; + + nir_alu_instr *src_alu = + nir_instr_as_alu(instr->src[src].src.ssa->parent_instr); + + if (nir_op_infos[src_alu->op].output_type != var->type && + !(var->type == nir_type_bool && alu_instr_is_bool(src_alu))) + return false; + } + + state->variables_seen |= (1 << var->variable); + state->variables[var->variable].src = instr->src[src].src; + state->variables[var->variable].abs = false; + state->variables[var->variable].negate = false; + + for (unsigned i = 0; i < 4; ++i) { + if (i < num_components) + state->variables[var->variable].swizzle[i] = new_swizzle[i]; + else + state->variables[var->variable].swizzle[i] = 0; + } + + return true; + } + } + + case nir_search_value_constant: { + nir_search_constant *const_val = nir_search_value_as_constant(value); + + if (!instr->src[src].src.is_ssa) + return false; + + if (instr->src[src].src.ssa->parent_instr->type != nir_instr_type_load_const) + return false; + + nir_load_const_instr *load = + nir_instr_as_load_const(instr->src[src].src.ssa->parent_instr); + + switch (nir_op_infos[instr->op].input_types[src]) { + case nir_type_float: + for (unsigned i = 0; i < num_components; ++i) { + if (load->value.f[new_swizzle[i]] != const_val->data.f) + return false; + } + return true; + case nir_type_int: + case nir_type_uint: + case nir_type_bool: + for (unsigned i = 0; i < num_components; ++i) { + if (load->value.i[new_swizzle[i]] != const_val->data.i) + return false; + } + return true; + default: + unreachable("Invalid alu source type"); + } + } + + default: + unreachable("Invalid search value type"); + } +} + +static bool +match_expression(const nir_search_expression *expr, nir_alu_instr *instr, + unsigned num_components, const uint8_t *swizzle, + struct match_state *state) +{ + if (instr->op != expr->opcode) + return false; + + assert(!instr->dest.saturate); + assert(nir_op_infos[instr->op].num_inputs > 0); + + /* If we have an explicitly sized destination, we can only handle the + * identity swizzle. While dot(vec3(a, b, c).zxy) is a valid + * expression, we don't have the information right now to propagate that + * swizzle through. We can only properly propagate swizzles if the + * instruction is vectorized. + */ + if (nir_op_infos[instr->op].output_size != 0) { + for (unsigned i = 0; i < num_components; i++) { + if (swizzle[i] != i) + return false; + } + } + + /* Stash off the current variables_seen bitmask. This way we can + * restore it prior to matching in the commutative case below. + */ + unsigned variables_seen_stash = state->variables_seen; + + bool matched = true; + for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) { + if (!match_value(expr->srcs[i], instr, i, num_components, + swizzle, state)) { + matched = false; + break; + } + } + + if (matched) + return true; + + if (nir_op_infos[instr->op].algebraic_properties & NIR_OP_IS_COMMUTATIVE) { + assert(nir_op_infos[instr->op].num_inputs == 2); + + /* Restore the variables_seen bitmask. If we don't do this, then we + * could end up with an erroneous failure due to variables found in the + * first match attempt above not matching those in the second. + */ + state->variables_seen = variables_seen_stash; + + if (!match_value(expr->srcs[0], instr, 1, num_components, + swizzle, state)) + return false; + + return match_value(expr->srcs[1], instr, 0, num_components, + swizzle, state); + } else { + return false; + } +} + +static nir_alu_src +construct_value(const nir_search_value *value, nir_alu_type type, + unsigned num_components, struct match_state *state, + nir_instr *instr, void *mem_ctx) +{ + switch (value->type) { + case nir_search_value_expression: { + const nir_search_expression *expr = nir_search_value_as_expression(value); + + if (nir_op_infos[expr->opcode].output_size != 0) + num_components = nir_op_infos[expr->opcode].output_size; + + nir_alu_instr *alu = nir_alu_instr_create(mem_ctx, expr->opcode); + nir_ssa_dest_init(&alu->instr, &alu->dest.dest, num_components, NULL); + alu->dest.write_mask = (1 << num_components) - 1; + alu->dest.saturate = false; + + for (unsigned i = 0; i < nir_op_infos[expr->opcode].num_inputs; i++) { + /* If the source is an explicitly sized source, then we need to reset + * the number of components to match. + */ + if (nir_op_infos[alu->op].input_sizes[i] != 0) + num_components = nir_op_infos[alu->op].input_sizes[i]; + + alu->src[i] = construct_value(expr->srcs[i], + nir_op_infos[alu->op].input_types[i], + num_components, + state, instr, mem_ctx); + } + + nir_instr_insert_before(instr, &alu->instr); + + nir_alu_src val; + val.src = nir_src_for_ssa(&alu->dest.dest.ssa); + val.negate = false; + val.abs = false, + memcpy(val.swizzle, identity_swizzle, sizeof val.swizzle); + + return val; + } + + case nir_search_value_variable: { + const nir_search_variable *var = nir_search_value_as_variable(value); + assert(state->variables_seen & (1 << var->variable)); + + nir_alu_src val = { NIR_SRC_INIT }; + nir_alu_src_copy(&val, &state->variables[var->variable], mem_ctx); + + assert(!var->is_constant); + + return val; + } + + case nir_search_value_constant: { + const nir_search_constant *c = nir_search_value_as_constant(value); + nir_load_const_instr *load = nir_load_const_instr_create(mem_ctx, 1); + + switch (type) { + case nir_type_float: + load->def.name = ralloc_asprintf(mem_ctx, "%f", c->data.f); + load->value.f[0] = c->data.f; + break; + case nir_type_int: + load->def.name = ralloc_asprintf(mem_ctx, "%d", c->data.i); + load->value.i[0] = c->data.i; + break; + case nir_type_uint: + case nir_type_bool: + load->value.u[0] = c->data.u; + break; + default: + unreachable("Invalid alu source type"); + } + + nir_instr_insert_before(instr, &load->instr); + + nir_alu_src val; + val.src = nir_src_for_ssa(&load->def); + val.negate = false; + val.abs = false, + memset(val.swizzle, 0, sizeof val.swizzle); + + return val; + } + + default: + unreachable("Invalid search value type"); + } +} + +nir_alu_instr * +nir_replace_instr(nir_alu_instr *instr, const nir_search_expression *search, + const nir_search_value *replace, void *mem_ctx) +{ + uint8_t swizzle[4] = { 0, 0, 0, 0 }; + + for (unsigned i = 0; i < instr->dest.dest.ssa.num_components; ++i) + swizzle[i] = i; + + assert(instr->dest.dest.is_ssa); + + struct match_state state; + state.variables_seen = 0; + + if (!match_expression(search, instr, instr->dest.dest.ssa.num_components, + swizzle, &state)) + return NULL; + + /* Inserting a mov may be unnecessary. However, it's much easier to + * simply let copy propagation clean this up than to try to go through + * and rewrite swizzles ourselves. + */ + nir_alu_instr *mov = nir_alu_instr_create(mem_ctx, nir_op_imov); + mov->dest.write_mask = instr->dest.write_mask; + nir_ssa_dest_init(&mov->instr, &mov->dest.dest, + instr->dest.dest.ssa.num_components, NULL); + + mov->src[0] = construct_value(replace, nir_op_infos[instr->op].output_type, + instr->dest.dest.ssa.num_components, &state, + &instr->instr, mem_ctx); + nir_instr_insert_before(&instr->instr, &mov->instr); + + nir_ssa_def_rewrite_uses(&instr->dest.dest.ssa, + nir_src_for_ssa(&mov->dest.dest.ssa)); + + /* We know this one has no more uses because we just rewrote them all, + * so we can remove it. The rest of the matched expression, however, we + * don't know so much about. We'll just let dead code clean them up. + */ + nir_instr_remove(&instr->instr); + + return mov; +} diff --git a/src/compiler/nir/nir_search.h b/src/compiler/nir/nir_search.h new file mode 100644 index 0000000..7d47792 --- /dev/null +++ b/src/compiler/nir/nir_search.h @@ -0,0 +1,99 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand (jason@jlekstrand.net) + * + */ + +#ifndef _NIR_SEARCH_ +#define _NIR_SEARCH_ + +#include "nir.h" + +#define NIR_SEARCH_MAX_VARIABLES 16 + +typedef enum { + nir_search_value_expression, + nir_search_value_variable, + nir_search_value_constant, +} nir_search_value_type; + +typedef struct { + nir_search_value_type type; +} nir_search_value; + +typedef struct { + nir_search_value value; + + /** The variable index; Must be less than NIR_SEARCH_MAX_VARIABLES */ + unsigned variable; + + /** Indicates that the given variable must be a constant + * + * This is only alloed in search expressions and indicates that the + * given variable is only allowed to match constant values. + */ + bool is_constant; + + /** Indicates that the given variable must have a certain type + * + * This is only allowed in search expressions and indicates that the + * given variable is only allowed to match values that come from an ALU + * instruction with the given output type. A type of nir_type_void + * means it can match any type. + * + * Note: A variable that is both constant and has a non-void type will + * never match anything. + */ + nir_alu_type type; +} nir_search_variable; + +typedef struct { + nir_search_value value; + + union { + uint32_t u; + int32_t i; + float f; + } data; +} nir_search_constant; + +typedef struct { + nir_search_value value; + + nir_op opcode; + const nir_search_value *srcs[4]; +} nir_search_expression; + +NIR_DEFINE_CAST(nir_search_value_as_variable, nir_search_value, + nir_search_variable, value) +NIR_DEFINE_CAST(nir_search_value_as_constant, nir_search_value, + nir_search_constant, value) +NIR_DEFINE_CAST(nir_search_value_as_expression, nir_search_value, + nir_search_expression, value) + +nir_alu_instr * +nir_replace_instr(nir_alu_instr *instr, const nir_search_expression *search, + const nir_search_value *replace, void *mem_ctx); + +#endif /* _NIR_SEARCH_ */ diff --git a/src/compiler/nir/nir_split_var_copies.c b/src/compiler/nir/nir_split_var_copies.c new file mode 100644 index 0000000..6fdaefa --- /dev/null +++ b/src/compiler/nir/nir_split_var_copies.c @@ -0,0 +1,285 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand (jason@jlekstrand.net) + * + */ + +#include "nir.h" + +/* + * Implements "copy splitting" which is similar to structure splitting only + * it works on copy operations rather than the datatypes themselves. The + * GLSL language allows you to copy one variable to another an entire + * structure (which may contain arrays or other structures) at a time. + * Normally, in a language such as C this would be handled by a "structure + * splitting" pass that breaks up the structures. Unfortunately for us, + * structures used in inputs or outputs can't be split. Therefore, + * regardlesss of what we do, we have to be able to copy to/from + * structures. + * + * The primary purpose of structure splitting is to allow you to better + * optimize variable access and lower things to registers where you can. + * The primary issue here is that, if you lower the copy to a bunch of + * loads and stores, you loose a lot of information about the copy + * operation that you would like to keep around. To solve this problem, we + * have a "copy splitting" pass that, instead of splitting the structures + * or lowering the copy into loads and storres, splits the copy operation + * into a bunch of copy operations one for each leaf of the structure tree. + * If an intermediate array is encountered, it is referenced with a + * wildcard reference to indicate that the entire array is to be copied. + * + * As things become direct, array copies may be able to be losslessly + * lowered to having fewer and fewer wildcards. However, until that + * happens we want to keep the information about the arrays intact. + * + * Prior to the copy splitting pass, there are no wildcard references but + * there may be incomplete references where the tail of the deref chain is + * an array or a structure and not a specific element. After the copy + * splitting pass has completed, every variable deref will be a full-length + * dereference pointing to a single leaf in the structure type tree with + * possibly a few wildcard array dereferences. + */ + +struct split_var_copies_state { + void *mem_ctx; + void *dead_ctx; + bool progress; +}; + +/* Recursively constructs deref chains to split a copy instruction into + * multiple (if needed) copy instructions with full-length deref chains. + * External callers of this function should pass the tail and head of the + * deref chains found as the source and destination of the copy instruction + * into this function. + * + * \param old_copy The copy instruction we are splitting + * \param dest_head The head of the destination deref chain we are building + * \param src_head The head of the source deref chain we are building + * \param dest_tail The tail of the destination deref chain we are building + * \param src_tail The tail of the source deref chain we are building + * \param state The current split_var_copies_state object + */ +static void +split_var_copy_instr(nir_intrinsic_instr *old_copy, + nir_deref *dest_head, nir_deref *src_head, + nir_deref *dest_tail, nir_deref *src_tail, + struct split_var_copies_state *state) +{ + assert(src_tail->type == dest_tail->type); + + /* Make sure these really are the tails of the deref chains */ + assert(dest_tail->child == NULL); + assert(src_tail->child == NULL); + + switch (glsl_get_base_type(src_tail->type)) { + case GLSL_TYPE_ARRAY: { + /* Make a wildcard dereference */ + nir_deref_array *deref = nir_deref_array_create(state->dead_ctx); + deref->deref.type = glsl_get_array_element(src_tail->type); + deref->deref_array_type = nir_deref_array_type_wildcard; + + /* Set the tail of both as the newly created wildcard deref. It is + * safe to use the same wildcard in both places because a) we will be + * copying it before we put it in an actual instruction and b) + * everything that will potentially add another link in the deref + * chain will also add the same thing to both chains. + */ + src_tail->child = &deref->deref; + dest_tail->child = &deref->deref; + + split_var_copy_instr(old_copy, dest_head, src_head, + dest_tail->child, src_tail->child, state); + + /* Set it back to the way we found it */ + src_tail->child = NULL; + dest_tail->child = NULL; + break; + } + + case GLSL_TYPE_STRUCT: + /* This is the only part that actually does any interesting + * splitting. For array types, we just use wildcards and resolve + * them later. For structure types, we need to emit one copy + * instruction for every structure element. Because we may have + * structs inside structs, we just recurse and let the next level + * take care of any additional structures. + */ + for (unsigned i = 0; i < glsl_get_length(src_tail->type); i++) { + nir_deref_struct *deref = nir_deref_struct_create(state->dead_ctx, i); + deref->deref.type = glsl_get_struct_field(src_tail->type, i); + + /* Set the tail of both as the newly created structure deref. It + * is safe to use the same wildcard in both places because a) we + * will be copying it before we put it in an actual instruction + * and b) everything that will potentially add another link in the + * deref chain will also add the same thing to both chains. + */ + src_tail->child = &deref->deref; + dest_tail->child = &deref->deref; + + split_var_copy_instr(old_copy, dest_head, src_head, + dest_tail->child, src_tail->child, state); + } + /* Set it back to the way we found it */ + src_tail->child = NULL; + dest_tail->child = NULL; + break; + + case GLSL_TYPE_UINT: + case GLSL_TYPE_INT: + case GLSL_TYPE_FLOAT: + case GLSL_TYPE_BOOL: + if (glsl_type_is_matrix(src_tail->type)) { + nir_deref_array *deref = nir_deref_array_create(state->dead_ctx); + deref->deref.type = glsl_get_column_type(src_tail->type); + deref->deref_array_type = nir_deref_array_type_wildcard; + + /* Set the tail of both as the newly created wildcard deref. It + * is safe to use the same wildcard in both places because a) we + * will be copying it before we put it in an actual instruction + * and b) everything that will potentially add another link in the + * deref chain will also add the same thing to both chains. + */ + src_tail->child = &deref->deref; + dest_tail->child = &deref->deref; + + split_var_copy_instr(old_copy, dest_head, src_head, + dest_tail->child, src_tail->child, state); + + /* Set it back to the way we found it */ + src_tail->child = NULL; + dest_tail->child = NULL; + } else { + /* At this point, we have fully built our deref chains and can + * actually add the new copy instruction. + */ + nir_intrinsic_instr *new_copy = + nir_intrinsic_instr_create(state->mem_ctx, nir_intrinsic_copy_var); + + /* We need to make copies because a) this deref chain actually + * belongs to the copy instruction and b) the deref chains may + * have some of the same links due to the way we constructed them + */ + nir_deref *src = nir_copy_deref(new_copy, src_head); + nir_deref *dest = nir_copy_deref(new_copy, dest_head); + + new_copy->variables[0] = nir_deref_as_var(dest); + new_copy->variables[1] = nir_deref_as_var(src); + + /* Emit the copy instruction after the old instruction. We'll + * remove the old one later. + */ + nir_instr_insert_after(&old_copy->instr, &new_copy->instr); + state->progress = true; + } + break; + + case GLSL_TYPE_SAMPLER: + case GLSL_TYPE_IMAGE: + case GLSL_TYPE_ATOMIC_UINT: + case GLSL_TYPE_INTERFACE: + default: + unreachable("Cannot copy these types"); + } +} + +static bool +split_var_copies_block(nir_block *block, void *void_state) +{ + struct split_var_copies_state *state = void_state; + + nir_foreach_instr_safe(block, instr) { + if (instr->type != nir_instr_type_intrinsic) + continue; + + nir_intrinsic_instr *intrinsic = nir_instr_as_intrinsic(instr); + if (intrinsic->intrinsic != nir_intrinsic_copy_var) + continue; + + nir_deref *dest_head = &intrinsic->variables[0]->deref; + nir_deref *src_head = &intrinsic->variables[1]->deref; + nir_deref *dest_tail = nir_deref_tail(dest_head); + nir_deref *src_tail = nir_deref_tail(src_head); + + switch (glsl_get_base_type(src_tail->type)) { + case GLSL_TYPE_ARRAY: + case GLSL_TYPE_STRUCT: + split_var_copy_instr(intrinsic, dest_head, src_head, + dest_tail, src_tail, state); + nir_instr_remove(&intrinsic->instr); + ralloc_steal(state->dead_ctx, instr); + break; + case GLSL_TYPE_FLOAT: + case GLSL_TYPE_INT: + case GLSL_TYPE_UINT: + case GLSL_TYPE_BOOL: + if (glsl_type_is_matrix(src_tail->type)) { + split_var_copy_instr(intrinsic, dest_head, src_head, + dest_tail, src_tail, state); + nir_instr_remove(&intrinsic->instr); + ralloc_steal(state->dead_ctx, instr); + } + break; + default: + unreachable("Invalid type"); + break; + } + } + + return true; +} + +static bool +split_var_copies_impl(nir_function_impl *impl) +{ + struct split_var_copies_state state; + + state.mem_ctx = ralloc_parent(impl); + state.dead_ctx = ralloc_context(NULL); + state.progress = false; + + nir_foreach_block(impl, split_var_copies_block, &state); + + ralloc_free(state.dead_ctx); + + if (state.progress) { + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance); + } + + return state.progress; +} + +bool +nir_split_var_copies(nir_shader *shader) +{ + bool progress = false; + + nir_foreach_function(shader, function) { + if (function->impl) + progress = split_var_copies_impl(function->impl) || progress; + } + + return progress; +} diff --git a/src/compiler/nir/nir_sweep.c b/src/compiler/nir/nir_sweep.c new file mode 100644 index 0000000..0710bdb --- /dev/null +++ b/src/compiler/nir/nir_sweep.c @@ -0,0 +1,173 @@ +/* + * Copyright © 2015 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + */ + +#include "nir.h" + +/** + * \file nir_sweep.c + * + * The nir_sweep() pass performs a mark and sweep pass over a nir_shader's associated + * memory - anything still connected to the program will be kept, and any dead memory + * we dropped on the floor will be freed. + * + * The expectation is that drivers should call this when finished compiling the shader + * (after any optimization, lowering, and so on). However, it's also fine to call it + * earlier, and even many times, trading CPU cycles for memory savings. + */ + +#define steal_list(mem_ctx, type, list) \ + foreach_list_typed(type, obj, node, list) { ralloc_steal(mem_ctx, obj); } + +static void sweep_cf_node(nir_shader *nir, nir_cf_node *cf_node); + +static bool +sweep_src_indirect(nir_src *src, void *nir) +{ + if (!src->is_ssa && src->reg.indirect) + ralloc_steal(nir, src->reg.indirect); + + return true; +} + +static bool +sweep_dest_indirect(nir_dest *dest, void *nir) +{ + if (!dest->is_ssa && dest->reg.indirect) + ralloc_steal(nir, dest->reg.indirect); + + return true; +} + +static void +sweep_block(nir_shader *nir, nir_block *block) +{ + ralloc_steal(nir, block); + + nir_foreach_instr(block, instr) { + ralloc_steal(nir, instr); + + nir_foreach_src(instr, sweep_src_indirect, nir); + nir_foreach_dest(instr, sweep_dest_indirect, nir); + } +} + +static void +sweep_if(nir_shader *nir, nir_if *iff) +{ + ralloc_steal(nir, iff); + + foreach_list_typed(nir_cf_node, cf_node, node, &iff->then_list) { + sweep_cf_node(nir, cf_node); + } + + foreach_list_typed(nir_cf_node, cf_node, node, &iff->else_list) { + sweep_cf_node(nir, cf_node); + } +} + +static void +sweep_loop(nir_shader *nir, nir_loop *loop) +{ + ralloc_steal(nir, loop); + + foreach_list_typed(nir_cf_node, cf_node, node, &loop->body) { + sweep_cf_node(nir, cf_node); + } +} + +static void +sweep_cf_node(nir_shader *nir, nir_cf_node *cf_node) +{ + switch (cf_node->type) { + case nir_cf_node_block: + sweep_block(nir, nir_cf_node_as_block(cf_node)); + break; + case nir_cf_node_if: + sweep_if(nir, nir_cf_node_as_if(cf_node)); + break; + case nir_cf_node_loop: + sweep_loop(nir, nir_cf_node_as_loop(cf_node)); + break; + default: + unreachable("Invalid CF node type"); + } +} + +static void +sweep_impl(nir_shader *nir, nir_function_impl *impl) +{ + ralloc_steal(nir, impl); + + ralloc_steal(nir, impl->params); + ralloc_steal(nir, impl->return_var); + steal_list(nir, nir_variable, &impl->locals); + steal_list(nir, nir_register, &impl->registers); + + foreach_list_typed(nir_cf_node, cf_node, node, &impl->body) { + sweep_cf_node(nir, cf_node); + } + + sweep_block(nir, impl->end_block); + + /* Wipe out all the metadata, if any. */ + nir_metadata_preserve(impl, nir_metadata_none); +} + +static void +sweep_function(nir_shader *nir, nir_function *f) +{ + ralloc_steal(nir, f); + ralloc_steal(nir, f->params); + + if (f->impl) + sweep_impl(nir, f->impl); +} + +void +nir_sweep(nir_shader *nir) +{ + void *rubbish = ralloc_context(NULL); + + /* First, move ownership of all the memory to a temporary context; assume dead. */ + ralloc_adopt(rubbish, nir); + + ralloc_steal(nir, (char *)nir->info.name); + if (nir->info.label) + ralloc_steal(nir, (char *)nir->info.label); + + /* Variables and registers are not dead. Steal them back. */ + steal_list(nir, nir_variable, &nir->uniforms); + steal_list(nir, nir_variable, &nir->inputs); + steal_list(nir, nir_variable, &nir->outputs); + steal_list(nir, nir_variable, &nir->globals); + steal_list(nir, nir_variable, &nir->system_values); + steal_list(nir, nir_register, &nir->registers); + + /* Recurse into functions, stealing their contents back. */ + foreach_list_typed(nir_function, func, node, &nir->functions) { + sweep_function(nir, func); + } + + /* Free everything we didn't steal back. */ + ralloc_free(rubbish); +} diff --git a/src/compiler/nir/nir_to_ssa.c b/src/compiler/nir/nir_to_ssa.c new file mode 100644 index 0000000..44a5054 --- /dev/null +++ b/src/compiler/nir/nir_to_ssa.c @@ -0,0 +1,536 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +#include "nir.h" +#include <stdlib.h> +#include <unistd.h> + +/* + * Implements the classic to-SSA algorithm described by Cytron et. al. in + * "Efficiently Computing Static Single Assignment Form and the Control + * Dependence Graph." + */ + +/* inserts a phi node of the form reg = phi(reg, reg, reg, ...) */ + +static void +insert_trivial_phi(nir_register *reg, nir_block *block, void *mem_ctx) +{ + nir_phi_instr *instr = nir_phi_instr_create(mem_ctx); + + instr->dest.reg.reg = reg; + struct set_entry *entry; + set_foreach(block->predecessors, entry) { + nir_block *pred = (nir_block *) entry->key; + + nir_phi_src *src = ralloc(instr, nir_phi_src); + src->pred = pred; + src->src.is_ssa = false; + src->src.reg.base_offset = 0; + src->src.reg.indirect = NULL; + src->src.reg.reg = reg; + exec_list_push_tail(&instr->srcs, &src->node); + } + + nir_instr_insert_before_block(block, &instr->instr); +} + +static void +insert_phi_nodes(nir_function_impl *impl) +{ + void *mem_ctx = ralloc_parent(impl); + + unsigned *work = calloc(impl->num_blocks, sizeof(unsigned)); + unsigned *has_already = calloc(impl->num_blocks, sizeof(unsigned)); + + /* + * Since the work flags already prevent us from inserting a node that has + * ever been inserted into W, we don't need to use a set to represent W. + * Also, since no block can ever be inserted into W more than once, we know + * that the maximum size of W is the number of basic blocks in the + * function. So all we need to handle W is an array and a pointer to the + * next element to be inserted and the next element to be removed. + */ + nir_block **W = malloc(impl->num_blocks * sizeof(nir_block *)); + unsigned w_start, w_end; + + unsigned iter_count = 0; + + nir_index_blocks(impl); + + foreach_list_typed(nir_register, reg, node, &impl->registers) { + if (reg->num_array_elems != 0) + continue; + + w_start = w_end = 0; + iter_count++; + + nir_foreach_def(reg, dest) { + nir_instr *def = dest->reg.parent_instr; + if (work[def->block->index] < iter_count) + W[w_end++] = def->block; + work[def->block->index] = iter_count; + } + + while (w_start != w_end) { + nir_block *cur = W[w_start++]; + struct set_entry *entry; + set_foreach(cur->dom_frontier, entry) { + nir_block *next = (nir_block *) entry->key; + + /* + * If there's more than one return statement, then the end block + * can be a join point for some definitions. However, there are + * no instructions in the end block, so nothing would use those + * phi nodes. Of course, we couldn't place those phi nodes + * anyways due to the restriction of having no instructions in the + * end block... + */ + if (next == impl->end_block) + continue; + + if (has_already[next->index] < iter_count) { + insert_trivial_phi(reg, next, mem_ctx); + has_already[next->index] = iter_count; + if (work[next->index] < iter_count) { + work[next->index] = iter_count; + W[w_end++] = next; + } + } + } + } + } + + free(work); + free(has_already); + free(W); +} + +typedef struct { + nir_ssa_def **stack; + int index; + unsigned num_defs; /** < used to add indices to debug names */ +#ifndef NDEBUG + unsigned stack_size; +#endif +} reg_state; + +typedef struct { + reg_state *states; + void *mem_ctx; + nir_instr *parent_instr; + nir_if *parent_if; + nir_function_impl *impl; + + /* map from SSA value -> original register */ + struct hash_table *ssa_map; +} rewrite_state; + +static nir_ssa_def *get_ssa_src(nir_register *reg, rewrite_state *state) +{ + unsigned index = reg->index; + + if (state->states[index].index == -1) { + /* + * We're using an undefined register, create a new undefined SSA value + * to preserve the information that this source is undefined + */ + nir_ssa_undef_instr *instr = + nir_ssa_undef_instr_create(state->mem_ctx, reg->num_components); + + /* + * We could just insert the undefined instruction before the instruction + * we're rewriting, but we could be rewriting a phi source in which case + * we can't do that, so do the next easiest thing - insert it at the + * beginning of the program. In the end, it doesn't really matter where + * the undefined instructions are because they're going to be ignored + * in the backend. + */ + nir_instr_insert_before_cf_list(&state->impl->body, &instr->instr); + return &instr->def; + } + + return state->states[index].stack[state->states[index].index]; +} + +static bool +rewrite_use(nir_src *src, void *_state) +{ + rewrite_state *state = (rewrite_state *) _state; + + if (src->is_ssa) + return true; + + unsigned index = src->reg.reg->index; + + if (state->states[index].stack == NULL) + return true; + + nir_ssa_def *def = get_ssa_src(src->reg.reg, state); + if (state->parent_instr) + nir_instr_rewrite_src(state->parent_instr, src, nir_src_for_ssa(def)); + else + nir_if_rewrite_condition(state->parent_if, nir_src_for_ssa(def)); + + return true; +} + +static bool +rewrite_def_forwards(nir_dest *dest, void *_state) +{ + rewrite_state *state = (rewrite_state *) _state; + + if (dest->is_ssa) + return true; + + nir_register *reg = dest->reg.reg; + unsigned index = reg->index; + + if (state->states[index].stack == NULL) + return true; + + char *name = NULL; + if (dest->reg.reg->name) + name = ralloc_asprintf(state->mem_ctx, "%s_%u", dest->reg.reg->name, + state->states[index].num_defs); + + list_del(&dest->reg.def_link); + nir_ssa_dest_init(state->parent_instr, dest, reg->num_components, name); + + /* push our SSA destination on the stack */ + state->states[index].index++; + assert(state->states[index].index < state->states[index].stack_size); + state->states[index].stack[state->states[index].index] = &dest->ssa; + state->states[index].num_defs++; + + _mesa_hash_table_insert(state->ssa_map, &dest->ssa, reg); + + return true; +} + +static void +rewrite_alu_instr_forward(nir_alu_instr *instr, rewrite_state *state) +{ + state->parent_instr = &instr->instr; + + nir_foreach_src(&instr->instr, rewrite_use, state); + + if (instr->dest.dest.is_ssa) + return; + + nir_register *reg = instr->dest.dest.reg.reg; + unsigned index = reg->index; + + if (state->states[index].stack == NULL) + return; + + unsigned write_mask = instr->dest.write_mask; + if (write_mask != (1 << instr->dest.dest.reg.reg->num_components) - 1) { + /* + * Calculate the number of components the final instruction, which for + * per-component things is the number of output components of the + * instruction and non-per-component things is the number of enabled + * channels in the write mask. + */ + unsigned num_components; + if (nir_op_infos[instr->op].output_size == 0) { + unsigned temp = (write_mask & 0x5) + ((write_mask >> 1) & 0x5); + num_components = (temp & 0x3) + ((temp >> 2) & 0x3); + } else { + num_components = nir_op_infos[instr->op].output_size; + } + + char *name = NULL; + if (instr->dest.dest.reg.reg->name) + name = ralloc_asprintf(state->mem_ctx, "%s_%u", + reg->name, state->states[index].num_defs); + + instr->dest.write_mask = (1 << num_components) - 1; + list_del(&instr->dest.dest.reg.def_link); + nir_ssa_dest_init(&instr->instr, &instr->dest.dest, num_components, name); + + if (nir_op_infos[instr->op].output_size == 0) { + /* + * When we change the output writemask, we need to change the + * swizzles for per-component inputs too + */ + for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) { + if (nir_op_infos[instr->op].input_sizes[i] != 0) + continue; + + unsigned new_swizzle[4] = {0, 0, 0, 0}; + + /* + * We keep two indices: + * 1. The index of the original (non-SSA) component + * 2. The index of the post-SSA, compacted, component + * + * We need to map the swizzle component at index 1 to the swizzle + * component at index 2. + */ + + unsigned ssa_index = 0; + for (unsigned index = 0; index < 4; index++) { + if (!((write_mask >> index) & 1)) + continue; + + new_swizzle[ssa_index] = instr->src[i].swizzle[index]; + ssa_index++; + } + + for (unsigned j = 0; j < 4; j++) + instr->src[i].swizzle[j] = new_swizzle[j]; + } + } + + nir_op op; + switch (reg->num_components) { + case 2: op = nir_op_vec2; break; + case 3: op = nir_op_vec3; break; + case 4: op = nir_op_vec4; break; + default: unreachable("not reached"); + } + + nir_alu_instr *vec = nir_alu_instr_create(state->mem_ctx, op); + + vec->dest.dest.reg.reg = reg; + vec->dest.write_mask = (1 << reg->num_components) - 1; + + nir_ssa_def *old_src = get_ssa_src(reg, state); + nir_ssa_def *new_src = &instr->dest.dest.ssa; + + unsigned ssa_index = 0; + for (unsigned i = 0; i < reg->num_components; i++) { + vec->src[i].src.is_ssa = true; + if ((write_mask >> i) & 1) { + vec->src[i].src.ssa = new_src; + if (nir_op_infos[instr->op].output_size == 0) + vec->src[i].swizzle[0] = ssa_index; + else + vec->src[i].swizzle[0] = i; + ssa_index++; + } else { + vec->src[i].src.ssa = old_src; + vec->src[i].swizzle[0] = i; + } + } + + nir_instr_insert_after(&instr->instr, &vec->instr); + + state->parent_instr = &vec->instr; + rewrite_def_forwards(&vec->dest.dest, state); + } else { + rewrite_def_forwards(&instr->dest.dest, state); + } +} + +static void +rewrite_phi_instr(nir_phi_instr *instr, rewrite_state *state) +{ + state->parent_instr = &instr->instr; + rewrite_def_forwards(&instr->dest, state); +} + +static void +rewrite_instr_forward(nir_instr *instr, rewrite_state *state) +{ + if (instr->type == nir_instr_type_alu) { + rewrite_alu_instr_forward(nir_instr_as_alu(instr), state); + return; + } + + if (instr->type == nir_instr_type_phi) { + rewrite_phi_instr(nir_instr_as_phi(instr), state); + return; + } + + state->parent_instr = instr; + + nir_foreach_src(instr, rewrite_use, state); + nir_foreach_dest(instr, rewrite_def_forwards, state); +} + +static void +rewrite_phi_sources(nir_block *block, nir_block *pred, rewrite_state *state) +{ + nir_foreach_instr(block, instr) { + if (instr->type != nir_instr_type_phi) + break; + + nir_phi_instr *phi_instr = nir_instr_as_phi(instr); + + state->parent_instr = instr; + + nir_foreach_phi_src(phi_instr, src) { + if (src->pred == pred) { + rewrite_use(&src->src, state); + break; + } + } + } +} + +static bool +rewrite_def_backwards(nir_dest *dest, void *_state) +{ + rewrite_state *state = (rewrite_state *) _state; + + if (!dest->is_ssa) + return true; + + struct hash_entry *entry = + _mesa_hash_table_search(state->ssa_map, &dest->ssa); + + if (!entry) + return true; + + nir_register *reg = (nir_register *) entry->data; + unsigned index = reg->index; + + state->states[index].index--; + assert(state->states[index].index >= -1); + + return true; +} + +static void +rewrite_instr_backwards(nir_instr *instr, rewrite_state *state) +{ + nir_foreach_dest(instr, rewrite_def_backwards, state); +} + +static void +rewrite_block(nir_block *block, rewrite_state *state) +{ + /* This will skip over any instructions after the current one, which is + * what we want because those instructions (vector gather, conditional + * select) will already be in SSA form. + */ + nir_foreach_instr_safe(block, instr) { + rewrite_instr_forward(instr, state); + } + + if (block != state->impl->end_block && + !nir_cf_node_is_last(&block->cf_node) && + nir_cf_node_next(&block->cf_node)->type == nir_cf_node_if) { + nir_if *if_stmt = nir_cf_node_as_if(nir_cf_node_next(&block->cf_node)); + state->parent_instr = NULL; + state->parent_if = if_stmt; + rewrite_use(&if_stmt->condition, state); + } + + if (block->successors[0]) + rewrite_phi_sources(block->successors[0], block, state); + if (block->successors[1]) + rewrite_phi_sources(block->successors[1], block, state); + + for (unsigned i = 0; i < block->num_dom_children; i++) + rewrite_block(block->dom_children[i], state); + + nir_foreach_instr_reverse(block, instr) { + rewrite_instr_backwards(instr, state); + } +} + +static void +remove_unused_regs(nir_function_impl *impl, rewrite_state *state) +{ + foreach_list_typed_safe(nir_register, reg, node, &impl->registers) { + if (state->states[reg->index].stack != NULL) + exec_node_remove(®->node); + } +} + +static void +init_rewrite_state(nir_function_impl *impl, rewrite_state *state) +{ + state->impl = impl; + state->mem_ctx = ralloc_parent(impl); + state->ssa_map = _mesa_hash_table_create(NULL, _mesa_hash_pointer, + _mesa_key_pointer_equal); + state->states = ralloc_array(NULL, reg_state, impl->reg_alloc); + + foreach_list_typed(nir_register, reg, node, &impl->registers) { + assert(reg->index < impl->reg_alloc); + if (reg->num_array_elems > 0) { + state->states[reg->index].stack = NULL; + } else { + /* + * Calculate a conservative estimate of the stack size based on the + * number of definitions there are. Note that this function *must* be + * called after phi nodes are inserted so we can count phi node + * definitions too. + */ + unsigned stack_size = list_length(®->defs); + + state->states[reg->index].stack = ralloc_array(state->states, + nir_ssa_def *, + stack_size); +#ifndef NDEBUG + state->states[reg->index].stack_size = stack_size; +#endif + state->states[reg->index].index = -1; + state->states[reg->index].num_defs = 0; + } + } +} + +static void +destroy_rewrite_state(rewrite_state *state) +{ + _mesa_hash_table_destroy(state->ssa_map, NULL); + ralloc_free(state->states); +} + +void +nir_convert_to_ssa_impl(nir_function_impl *impl) +{ + nir_metadata_require(impl, nir_metadata_dominance); + + insert_phi_nodes(impl); + + rewrite_state state; + init_rewrite_state(impl, &state); + + rewrite_block(nir_start_block(impl), &state); + + remove_unused_regs(impl, &state); + + nir_metadata_preserve(impl, nir_metadata_block_index | + nir_metadata_dominance); + + destroy_rewrite_state(&state); +} + +void +nir_convert_to_ssa(nir_shader *shader) +{ + nir_foreach_function(shader, function) { + if (function->impl) + nir_convert_to_ssa_impl(function->impl); + } +} diff --git a/src/compiler/nir/nir_validate.c b/src/compiler/nir/nir_validate.c new file mode 100644 index 0000000..e4db68d --- /dev/null +++ b/src/compiler/nir/nir_validate.c @@ -0,0 +1,1071 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Connor Abbott (cwabbott0@gmail.com) + * + */ + +#include "nir.h" +#include <assert.h> + +/* + * This file checks for invalid IR indicating a bug somewhere in the compiler. + */ + +/* Since this file is just a pile of asserts, don't bother compiling it if + * we're not building a debug build. + */ +#ifdef DEBUG + +/* + * Per-register validation state. + */ + +typedef struct { + /* + * equivalent to the uses and defs in nir_register, but built up by the + * validator. At the end, we verify that the sets have the same entries. + */ + struct set *uses, *if_uses, *defs; + nir_function_impl *where_defined; /* NULL for global registers */ +} reg_validate_state; + +typedef struct { + /* + * equivalent to the uses in nir_ssa_def, but built up by the validator. + * At the end, we verify that the sets have the same entries. + */ + struct set *uses, *if_uses; + nir_function_impl *where_defined; +} ssa_def_validate_state; + +typedef struct { + /* map of register -> validation state (struct above) */ + struct hash_table *regs; + + /* the current shader being validated */ + nir_shader *shader; + + /* the current instruction being validated */ + nir_instr *instr; + + /* the current basic block being validated */ + nir_block *block; + + /* the current if statement being validated */ + nir_if *if_stmt; + + /* the current loop being visited */ + nir_loop *loop; + + /* the parent of the current cf node being visited */ + nir_cf_node *parent_node; + + /* the current function implementation being validated */ + nir_function_impl *impl; + + /* map of SSA value -> function implementation where it is defined */ + struct hash_table *ssa_defs; + + /* bitset of ssa definitions we have found; used to check uniqueness */ + BITSET_WORD *ssa_defs_found; + + /* bitset of registers we have currently found; used to check uniqueness */ + BITSET_WORD *regs_found; + + /* map of local variable -> function implementation where it is defined */ + struct hash_table *var_defs; +} validate_state; + +static void validate_src(nir_src *src, validate_state *state); + +static void +validate_reg_src(nir_src *src, validate_state *state) +{ + assert(src->reg.reg != NULL); + + struct hash_entry *entry; + entry = _mesa_hash_table_search(state->regs, src->reg.reg); + assert(entry); + + reg_validate_state *reg_state = (reg_validate_state *) entry->data; + + if (state->instr) { + _mesa_set_add(reg_state->uses, src); + } else { + assert(state->if_stmt); + _mesa_set_add(reg_state->if_uses, src); + } + + if (!src->reg.reg->is_global) { + assert(reg_state->where_defined == state->impl && + "using a register declared in a different function"); + } + + assert((src->reg.reg->num_array_elems == 0 || + src->reg.base_offset < src->reg.reg->num_array_elems) && + "definitely out-of-bounds array access"); + + if (src->reg.indirect) { + assert(src->reg.reg->num_array_elems != 0); + assert((src->reg.indirect->is_ssa || + src->reg.indirect->reg.indirect == NULL) && + "only one level of indirection allowed"); + validate_src(src->reg.indirect, state); + } +} + +static void +validate_ssa_src(nir_src *src, validate_state *state) +{ + assert(src->ssa != NULL); + + struct hash_entry *entry = _mesa_hash_table_search(state->ssa_defs, src->ssa); + + assert(entry); + + ssa_def_validate_state *def_state = (ssa_def_validate_state *)entry->data; + + assert(def_state->where_defined == state->impl && + "using an SSA value defined in a different function"); + + if (state->instr) { + _mesa_set_add(def_state->uses, src); + } else { + assert(state->if_stmt); + _mesa_set_add(def_state->if_uses, src); + } + + /* TODO validate that the use is dominated by the definition */ +} + +static void +validate_src(nir_src *src, validate_state *state) +{ + if (state->instr) + assert(src->parent_instr == state->instr); + else + assert(src->parent_if == state->if_stmt); + + if (src->is_ssa) + validate_ssa_src(src, state); + else + validate_reg_src(src, state); +} + +static void +validate_alu_src(nir_alu_instr *instr, unsigned index, validate_state *state) +{ + nir_alu_src *src = &instr->src[index]; + + unsigned num_components; + if (src->src.is_ssa) + num_components = src->src.ssa->num_components; + else { + if (src->src.reg.reg->is_packed) + num_components = 4; /* can't check anything */ + else + num_components = src->src.reg.reg->num_components; + } + for (unsigned i = 0; i < 4; i++) { + assert(src->swizzle[i] < 4); + + if (nir_alu_instr_channel_used(instr, index, i)) + assert(src->swizzle[i] < num_components); + } + + validate_src(&src->src, state); +} + +static void +validate_reg_dest(nir_reg_dest *dest, validate_state *state) +{ + assert(dest->reg != NULL); + + assert(dest->parent_instr == state->instr); + + struct hash_entry *entry2; + entry2 = _mesa_hash_table_search(state->regs, dest->reg); + + assert(entry2); + + reg_validate_state *reg_state = (reg_validate_state *) entry2->data; + _mesa_set_add(reg_state->defs, dest); + + if (!dest->reg->is_global) { + assert(reg_state->where_defined == state->impl && + "writing to a register declared in a different function"); + } + + assert((dest->reg->num_array_elems == 0 || + dest->base_offset < dest->reg->num_array_elems) && + "definitely out-of-bounds array access"); + + if (dest->indirect) { + assert(dest->reg->num_array_elems != 0); + assert((dest->indirect->is_ssa || dest->indirect->reg.indirect == NULL) && + "only one level of indirection allowed"); + validate_src(dest->indirect, state); + } +} + +static void +validate_ssa_def(nir_ssa_def *def, validate_state *state) +{ + assert(def->index < state->impl->ssa_alloc); + assert(!BITSET_TEST(state->ssa_defs_found, def->index)); + BITSET_SET(state->ssa_defs_found, def->index); + + assert(def->parent_instr == state->instr); + + assert(def->num_components <= 4); + + list_validate(&def->uses); + list_validate(&def->if_uses); + + ssa_def_validate_state *def_state = ralloc(state->ssa_defs, + ssa_def_validate_state); + def_state->where_defined = state->impl; + def_state->uses = _mesa_set_create(def_state, _mesa_hash_pointer, + _mesa_key_pointer_equal); + def_state->if_uses = _mesa_set_create(def_state, _mesa_hash_pointer, + _mesa_key_pointer_equal); + _mesa_hash_table_insert(state->ssa_defs, def, def_state); +} + +static void +validate_dest(nir_dest *dest, validate_state *state) +{ + if (dest->is_ssa) + validate_ssa_def(&dest->ssa, state); + else + validate_reg_dest(&dest->reg, state); +} + +static void +validate_alu_dest(nir_alu_dest *dest, validate_state *state) +{ + unsigned dest_size = + dest->dest.is_ssa ? dest->dest.ssa.num_components + : dest->dest.reg.reg->num_components; + bool is_packed = !dest->dest.is_ssa && dest->dest.reg.reg->is_packed; + /* + * validate that the instruction doesn't write to components not in the + * register/SSA value + */ + assert(is_packed || !(dest->write_mask & ~((1 << dest_size) - 1))); + + /* validate that saturate is only ever used on instructions with + * destinations of type float + */ + nir_alu_instr *alu = nir_instr_as_alu(state->instr); + assert(nir_op_infos[alu->op].output_type == nir_type_float || + !dest->saturate); + + validate_dest(&dest->dest, state); +} + +static void +validate_alu_instr(nir_alu_instr *instr, validate_state *state) +{ + assert(instr->op < nir_num_opcodes); + + for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) { + validate_alu_src(instr, i, state); + } + + validate_alu_dest(&instr->dest, state); +} + +static void +validate_deref_chain(nir_deref *deref, validate_state *state) +{ + assert(deref->child == NULL || ralloc_parent(deref->child) == deref); + + nir_deref *parent = NULL; + while (deref != NULL) { + switch (deref->deref_type) { + case nir_deref_type_array: + assert(deref->type == glsl_get_array_element(parent->type)); + if (nir_deref_as_array(deref)->deref_array_type == + nir_deref_array_type_indirect) + validate_src(&nir_deref_as_array(deref)->indirect, state); + break; + + case nir_deref_type_struct: + assert(deref->type == + glsl_get_struct_field(parent->type, + nir_deref_as_struct(deref)->index)); + break; + + case nir_deref_type_var: + break; + + default: + assert(!"Invalid deref type"); + break; + } + + parent = deref; + deref = deref->child; + } +} + +static void +validate_var_use(nir_variable *var, validate_state *state) +{ + if (var->data.mode == nir_var_local) { + struct hash_entry *entry = _mesa_hash_table_search(state->var_defs, var); + + assert(entry); + assert((nir_function_impl *) entry->data == state->impl); + } +} + +static void +validate_deref_var(void *parent_mem_ctx, nir_deref_var *deref, validate_state *state) +{ + assert(deref != NULL); + assert(ralloc_parent(deref) == parent_mem_ctx); + assert(deref->deref.type == deref->var->type); + + validate_var_use(deref->var, state); + + validate_deref_chain(&deref->deref, state); +} + +static void +validate_intrinsic_instr(nir_intrinsic_instr *instr, validate_state *state) +{ + unsigned num_srcs = nir_intrinsic_infos[instr->intrinsic].num_srcs; + for (unsigned i = 0; i < num_srcs; i++) { + unsigned components_read = + nir_intrinsic_infos[instr->intrinsic].src_components[i]; + if (components_read == 0) + components_read = instr->num_components; + + assert(components_read > 0); + + if (instr->src[i].is_ssa) { + assert(components_read <= instr->src[i].ssa->num_components); + } else if (!instr->src[i].reg.reg->is_packed) { + assert(components_read <= instr->src[i].reg.reg->num_components); + } + + validate_src(&instr->src[i], state); + } + + unsigned num_vars = nir_intrinsic_infos[instr->intrinsic].num_variables; + for (unsigned i = 0; i < num_vars; i++) { + validate_deref_var(instr, instr->variables[i], state); + } + + if (nir_intrinsic_infos[instr->intrinsic].has_dest) { + unsigned components_written = + nir_intrinsic_infos[instr->intrinsic].dest_components; + if (components_written == 0) + components_written = instr->num_components; + + assert(components_written > 0); + + if (instr->dest.is_ssa) { + assert(components_written <= instr->dest.ssa.num_components); + } else if (!instr->dest.reg.reg->is_packed) { + assert(components_written <= instr->dest.reg.reg->num_components); + } + + validate_dest(&instr->dest, state); + } + + switch (instr->intrinsic) { + case nir_intrinsic_load_var: { + const struct glsl_type *type = + nir_deref_tail(&instr->variables[0]->deref)->type; + assert(glsl_type_is_vector_or_scalar(type) || + (instr->variables[0]->var->data.mode == nir_var_uniform && + glsl_get_base_type(type) == GLSL_TYPE_SUBROUTINE)); + assert(instr->num_components == glsl_get_vector_elements(type)); + break; + } + case nir_intrinsic_store_var: { + const struct glsl_type *type = + nir_deref_tail(&instr->variables[0]->deref)->type; + assert(glsl_type_is_vector_or_scalar(type) || + (instr->variables[0]->var->data.mode == nir_var_uniform && + glsl_get_base_type(type) == GLSL_TYPE_SUBROUTINE)); + assert(instr->num_components == glsl_get_vector_elements(type)); + assert(instr->variables[0]->var->data.mode != nir_var_shader_in && + instr->variables[0]->var->data.mode != nir_var_uniform && + instr->variables[0]->var->data.mode != nir_var_shader_storage); + assert((instr->const_index[0] & ~((1 << instr->num_components) - 1)) == 0); + break; + } + case nir_intrinsic_copy_var: + assert(nir_deref_tail(&instr->variables[0]->deref)->type == + nir_deref_tail(&instr->variables[1]->deref)->type); + assert(instr->variables[0]->var->data.mode != nir_var_shader_in && + instr->variables[0]->var->data.mode != nir_var_uniform && + instr->variables[0]->var->data.mode != nir_var_shader_storage); + break; + default: + break; + } +} + +static void +validate_tex_instr(nir_tex_instr *instr, validate_state *state) +{ + bool src_type_seen[nir_num_tex_src_types]; + for (unsigned i = 0; i < nir_num_tex_src_types; i++) + src_type_seen[i] = false; + + for (unsigned i = 0; i < instr->num_srcs; i++) { + assert(!src_type_seen[instr->src[i].src_type]); + src_type_seen[instr->src[i].src_type] = true; + validate_src(&instr->src[i].src, state); + } + + if (instr->sampler != NULL) + validate_deref_var(instr, instr->sampler, state); + + validate_dest(&instr->dest, state); +} + +static void +validate_call_instr(nir_call_instr *instr, validate_state *state) +{ + if (instr->return_deref == NULL) + assert(glsl_type_is_void(instr->callee->return_type)); + else + assert(instr->return_deref->deref.type == instr->callee->return_type); + + assert(instr->num_params == instr->callee->num_params); + + for (unsigned i = 0; i < instr->num_params; i++) { + assert(instr->callee->params[i].type == instr->params[i]->deref.type); + validate_deref_var(instr, instr->params[i], state); + } + + validate_deref_var(instr, instr->return_deref, state); +} + +static void +validate_load_const_instr(nir_load_const_instr *instr, validate_state *state) +{ + validate_ssa_def(&instr->def, state); +} + +static void +validate_ssa_undef_instr(nir_ssa_undef_instr *instr, validate_state *state) +{ + validate_ssa_def(&instr->def, state); +} + +static void +validate_phi_instr(nir_phi_instr *instr, validate_state *state) +{ + /* + * don't validate the sources until we get to them from their predecessor + * basic blocks, to avoid validating an SSA use before its definition. + */ + + validate_dest(&instr->dest, state); + + exec_list_validate(&instr->srcs); + assert(exec_list_length(&instr->srcs) == + state->block->predecessors->entries); +} + +static void +validate_instr(nir_instr *instr, validate_state *state) +{ + assert(instr->block == state->block); + + state->instr = instr; + + switch (instr->type) { + case nir_instr_type_alu: + validate_alu_instr(nir_instr_as_alu(instr), state); + break; + + case nir_instr_type_call: + validate_call_instr(nir_instr_as_call(instr), state); + break; + + case nir_instr_type_intrinsic: + validate_intrinsic_instr(nir_instr_as_intrinsic(instr), state); + break; + + case nir_instr_type_tex: + validate_tex_instr(nir_instr_as_tex(instr), state); + break; + + case nir_instr_type_load_const: + validate_load_const_instr(nir_instr_as_load_const(instr), state); + break; + + case nir_instr_type_phi: + validate_phi_instr(nir_instr_as_phi(instr), state); + break; + + case nir_instr_type_ssa_undef: + validate_ssa_undef_instr(nir_instr_as_ssa_undef(instr), state); + break; + + case nir_instr_type_jump: + break; + + default: + assert(!"Invalid ALU instruction type"); + break; + } + + state->instr = NULL; +} + +static void +validate_phi_src(nir_phi_instr *instr, nir_block *pred, validate_state *state) +{ + state->instr = &instr->instr; + + assert(instr->dest.is_ssa); + + exec_list_validate(&instr->srcs); + nir_foreach_phi_src(instr, src) { + if (src->pred == pred) { + assert(src->src.is_ssa); + assert(src->src.ssa->num_components == + instr->dest.ssa.num_components); + + validate_src(&src->src, state); + state->instr = NULL; + return; + } + } + + abort(); +} + +static void +validate_phi_srcs(nir_block *block, nir_block *succ, validate_state *state) +{ + nir_foreach_instr(succ, instr) { + if (instr->type != nir_instr_type_phi) + break; + + validate_phi_src(nir_instr_as_phi(instr), block, state); + } +} + +static void validate_cf_node(nir_cf_node *node, validate_state *state); + +static void +validate_block(nir_block *block, validate_state *state) +{ + assert(block->cf_node.parent == state->parent_node); + + state->block = block; + + exec_list_validate(&block->instr_list); + nir_foreach_instr(block, instr) { + if (instr->type == nir_instr_type_phi) { + assert(instr == nir_block_first_instr(block) || + nir_instr_prev(instr)->type == nir_instr_type_phi); + } + + if (instr->type == nir_instr_type_jump) { + assert(instr == nir_block_last_instr(block)); + } + + validate_instr(instr, state); + } + + assert(block->successors[0] != NULL); + assert(block->successors[0] != block->successors[1]); + + for (unsigned i = 0; i < 2; i++) { + if (block->successors[i] != NULL) { + struct set_entry *entry = + _mesa_set_search(block->successors[i]->predecessors, block); + assert(entry); + + validate_phi_srcs(block, block->successors[i], state); + } + } + + struct set_entry *entry; + set_foreach(block->predecessors, entry) { + const nir_block *pred = entry->key; + assert(pred->successors[0] == block || + pred->successors[1] == block); + } + + if (!exec_list_is_empty(&block->instr_list) && + nir_block_last_instr(block)->type == nir_instr_type_jump) { + assert(block->successors[1] == NULL); + nir_jump_instr *jump = nir_instr_as_jump(nir_block_last_instr(block)); + switch (jump->type) { + case nir_jump_break: { + nir_block *after = + nir_cf_node_as_block(nir_cf_node_next(&state->loop->cf_node)); + assert(block->successors[0] == after); + break; + } + + case nir_jump_continue: { + nir_block *first = + nir_cf_node_as_block(nir_loop_first_cf_node(state->loop)); + assert(block->successors[0] == first); + break; + } + + case nir_jump_return: + assert(block->successors[0] == state->impl->end_block); + break; + + default: + unreachable("bad jump type"); + } + } else { + nir_cf_node *next = nir_cf_node_next(&block->cf_node); + if (next == NULL) { + switch (state->parent_node->type) { + case nir_cf_node_loop: { + nir_block *first = + nir_cf_node_as_block(nir_loop_first_cf_node(state->loop)); + assert(block->successors[0] == first); + /* due to the hack for infinite loops, block->successors[1] may + * point to the block after the loop. + */ + break; + } + + case nir_cf_node_if: { + nir_block *after = + nir_cf_node_as_block(nir_cf_node_next(state->parent_node)); + assert(block->successors[0] == after); + assert(block->successors[1] == NULL); + break; + } + + case nir_cf_node_function: + assert(block->successors[0] == state->impl->end_block); + assert(block->successors[1] == NULL); + break; + + default: + unreachable("unknown control flow node type"); + } + } else { + if (next->type == nir_cf_node_if) { + nir_if *if_stmt = nir_cf_node_as_if(next); + assert(&block->successors[0]->cf_node == + nir_if_first_then_node(if_stmt)); + assert(&block->successors[1]->cf_node == + nir_if_first_else_node(if_stmt)); + } else { + assert(next->type == nir_cf_node_loop); + nir_loop *loop = nir_cf_node_as_loop(next); + assert(&block->successors[0]->cf_node == + nir_loop_first_cf_node(loop)); + assert(block->successors[1] == NULL); + } + } + } +} + +static void +validate_if(nir_if *if_stmt, validate_state *state) +{ + state->if_stmt = if_stmt; + + assert(!exec_node_is_head_sentinel(if_stmt->cf_node.node.prev)); + nir_cf_node *prev_node = nir_cf_node_prev(&if_stmt->cf_node); + assert(prev_node->type == nir_cf_node_block); + + assert(!exec_node_is_tail_sentinel(if_stmt->cf_node.node.next)); + nir_cf_node *next_node = nir_cf_node_next(&if_stmt->cf_node); + assert(next_node->type == nir_cf_node_block); + + validate_src(&if_stmt->condition, state); + + assert(!exec_list_is_empty(&if_stmt->then_list)); + assert(!exec_list_is_empty(&if_stmt->else_list)); + + nir_cf_node *old_parent = state->parent_node; + state->parent_node = &if_stmt->cf_node; + + exec_list_validate(&if_stmt->then_list); + foreach_list_typed(nir_cf_node, cf_node, node, &if_stmt->then_list) { + validate_cf_node(cf_node, state); + } + + exec_list_validate(&if_stmt->else_list); + foreach_list_typed(nir_cf_node, cf_node, node, &if_stmt->else_list) { + validate_cf_node(cf_node, state); + } + + state->parent_node = old_parent; + state->if_stmt = NULL; +} + +static void +validate_loop(nir_loop *loop, validate_state *state) +{ + assert(!exec_node_is_head_sentinel(loop->cf_node.node.prev)); + nir_cf_node *prev_node = nir_cf_node_prev(&loop->cf_node); + assert(prev_node->type == nir_cf_node_block); + + assert(!exec_node_is_tail_sentinel(loop->cf_node.node.next)); + nir_cf_node *next_node = nir_cf_node_next(&loop->cf_node); + assert(next_node->type == nir_cf_node_block); + + assert(!exec_list_is_empty(&loop->body)); + + nir_cf_node *old_parent = state->parent_node; + state->parent_node = &loop->cf_node; + nir_loop *old_loop = state->loop; + state->loop = loop; + + exec_list_validate(&loop->body); + foreach_list_typed(nir_cf_node, cf_node, node, &loop->body) { + validate_cf_node(cf_node, state); + } + + state->parent_node = old_parent; + state->loop = old_loop; +} + +static void +validate_cf_node(nir_cf_node *node, validate_state *state) +{ + assert(node->parent == state->parent_node); + + switch (node->type) { + case nir_cf_node_block: + validate_block(nir_cf_node_as_block(node), state); + break; + + case nir_cf_node_if: + validate_if(nir_cf_node_as_if(node), state); + break; + + case nir_cf_node_loop: + validate_loop(nir_cf_node_as_loop(node), state); + break; + + default: + unreachable("Invalid CF node type"); + } +} + +static void +prevalidate_reg_decl(nir_register *reg, bool is_global, validate_state *state) +{ + assert(reg->is_global == is_global); + + if (is_global) + assert(reg->index < state->shader->reg_alloc); + else + assert(reg->index < state->impl->reg_alloc); + assert(!BITSET_TEST(state->regs_found, reg->index)); + BITSET_SET(state->regs_found, reg->index); + + list_validate(®->uses); + list_validate(®->defs); + list_validate(®->if_uses); + + reg_validate_state *reg_state = ralloc(state->regs, reg_validate_state); + reg_state->uses = _mesa_set_create(reg_state, _mesa_hash_pointer, + _mesa_key_pointer_equal); + reg_state->if_uses = _mesa_set_create(reg_state, _mesa_hash_pointer, + _mesa_key_pointer_equal); + reg_state->defs = _mesa_set_create(reg_state, _mesa_hash_pointer, + _mesa_key_pointer_equal); + + reg_state->where_defined = is_global ? NULL : state->impl; + + _mesa_hash_table_insert(state->regs, reg, reg_state); +} + +static void +postvalidate_reg_decl(nir_register *reg, validate_state *state) +{ + struct hash_entry *entry = _mesa_hash_table_search(state->regs, reg); + + reg_validate_state *reg_state = (reg_validate_state *) entry->data; + + nir_foreach_use(reg, src) { + struct set_entry *entry = _mesa_set_search(reg_state->uses, src); + assert(entry); + _mesa_set_remove(reg_state->uses, entry); + } + + if (reg_state->uses->entries != 0) { + printf("extra entries in register uses:\n"); + struct set_entry *entry; + set_foreach(reg_state->uses, entry) + printf("%p\n", entry->key); + + abort(); + } + + nir_foreach_if_use(reg, src) { + struct set_entry *entry = _mesa_set_search(reg_state->if_uses, src); + assert(entry); + _mesa_set_remove(reg_state->if_uses, entry); + } + + if (reg_state->if_uses->entries != 0) { + printf("extra entries in register if_uses:\n"); + struct set_entry *entry; + set_foreach(reg_state->if_uses, entry) + printf("%p\n", entry->key); + + abort(); + } + + nir_foreach_def(reg, src) { + struct set_entry *entry = _mesa_set_search(reg_state->defs, src); + assert(entry); + _mesa_set_remove(reg_state->defs, entry); + } + + if (reg_state->defs->entries != 0) { + printf("extra entries in register defs:\n"); + struct set_entry *entry; + set_foreach(reg_state->defs, entry) + printf("%p\n", entry->key); + + abort(); + } +} + +static void +validate_var_decl(nir_variable *var, bool is_global, validate_state *state) +{ + assert(is_global != (var->data.mode == nir_var_local)); + + /* + * TODO validate some things ir_validate.cpp does (requires more GLSL type + * support) + */ + + if (!is_global) { + _mesa_hash_table_insert(state->var_defs, var, state->impl); + } +} + +static bool +postvalidate_ssa_def(nir_ssa_def *def, void *void_state) +{ + validate_state *state = void_state; + + struct hash_entry *entry = _mesa_hash_table_search(state->ssa_defs, def); + ssa_def_validate_state *def_state = (ssa_def_validate_state *)entry->data; + + nir_foreach_use(def, src) { + struct set_entry *entry = _mesa_set_search(def_state->uses, src); + assert(entry); + _mesa_set_remove(def_state->uses, entry); + } + + if (def_state->uses->entries != 0) { + printf("extra entries in register uses:\n"); + struct set_entry *entry; + set_foreach(def_state->uses, entry) + printf("%p\n", entry->key); + + abort(); + } + + nir_foreach_if_use(def, src) { + struct set_entry *entry = _mesa_set_search(def_state->if_uses, src); + assert(entry); + _mesa_set_remove(def_state->if_uses, entry); + } + + if (def_state->if_uses->entries != 0) { + printf("extra entries in register uses:\n"); + struct set_entry *entry; + set_foreach(def_state->if_uses, entry) + printf("%p\n", entry->key); + + abort(); + } + + return true; +} + +static bool +postvalidate_ssa_defs_block(nir_block *block, void *state) +{ + nir_foreach_instr(block, instr) + nir_foreach_ssa_def(instr, postvalidate_ssa_def, state); + + return true; +} + +static void +validate_function_impl(nir_function_impl *impl, validate_state *state) +{ + assert(impl->function->impl == impl); + assert(impl->cf_node.parent == NULL); + + assert(impl->num_params == impl->function->num_params); + for (unsigned i = 0; i < impl->num_params; i++) + assert(impl->params[i]->type == impl->function->params[i].type); + + if (glsl_type_is_void(impl->function->return_type)) + assert(impl->return_var == NULL); + else + assert(impl->return_var->type == impl->function->return_type); + + assert(exec_list_is_empty(&impl->end_block->instr_list)); + assert(impl->end_block->successors[0] == NULL); + assert(impl->end_block->successors[1] == NULL); + + state->impl = impl; + state->parent_node = &impl->cf_node; + + exec_list_validate(&impl->locals); + nir_foreach_variable(var, &impl->locals) { + validate_var_decl(var, false, state); + } + + state->regs_found = realloc(state->regs_found, + BITSET_WORDS(impl->reg_alloc) * + sizeof(BITSET_WORD)); + memset(state->regs_found, 0, BITSET_WORDS(impl->reg_alloc) * + sizeof(BITSET_WORD)); + exec_list_validate(&impl->registers); + foreach_list_typed(nir_register, reg, node, &impl->registers) { + prevalidate_reg_decl(reg, false, state); + } + + state->ssa_defs_found = realloc(state->ssa_defs_found, + BITSET_WORDS(impl->ssa_alloc) * + sizeof(BITSET_WORD)); + memset(state->ssa_defs_found, 0, BITSET_WORDS(impl->ssa_alloc) * + sizeof(BITSET_WORD)); + exec_list_validate(&impl->body); + foreach_list_typed(nir_cf_node, node, node, &impl->body) { + validate_cf_node(node, state); + } + + foreach_list_typed(nir_register, reg, node, &impl->registers) { + postvalidate_reg_decl(reg, state); + } + + nir_foreach_block(impl, postvalidate_ssa_defs_block, state); +} + +static void +validate_function(nir_function *func, validate_state *state) +{ + if (func->impl != NULL) { + assert(func->impl->function == func); + validate_function_impl(func->impl, state); + } +} + +static void +init_validate_state(validate_state *state) +{ + state->regs = _mesa_hash_table_create(NULL, _mesa_hash_pointer, + _mesa_key_pointer_equal); + state->ssa_defs = _mesa_hash_table_create(NULL, _mesa_hash_pointer, + _mesa_key_pointer_equal); + state->ssa_defs_found = NULL; + state->regs_found = NULL; + state->var_defs = _mesa_hash_table_create(NULL, _mesa_hash_pointer, + _mesa_key_pointer_equal); + state->loop = NULL; +} + +static void +destroy_validate_state(validate_state *state) +{ + _mesa_hash_table_destroy(state->regs, NULL); + _mesa_hash_table_destroy(state->ssa_defs, NULL); + free(state->ssa_defs_found); + free(state->regs_found); + _mesa_hash_table_destroy(state->var_defs, NULL); +} + +void +nir_validate_shader(nir_shader *shader) +{ + validate_state state; + init_validate_state(&state); + + state.shader = shader; + + exec_list_validate(&shader->uniforms); + nir_foreach_variable(var, &shader->uniforms) { + validate_var_decl(var, true, &state); + } + + exec_list_validate(&shader->inputs); + nir_foreach_variable(var, &shader->inputs) { + validate_var_decl(var, true, &state); + } + + exec_list_validate(&shader->outputs); + nir_foreach_variable(var, &shader->outputs) { + validate_var_decl(var, true, &state); + } + + exec_list_validate(&shader->globals); + nir_foreach_variable(var, &shader->globals) { + validate_var_decl(var, true, &state); + } + + exec_list_validate(&shader->system_values); + nir_foreach_variable(var, &shader->system_values) { + validate_var_decl(var, true, &state); + } + + state.regs_found = realloc(state.regs_found, + BITSET_WORDS(shader->reg_alloc) * + sizeof(BITSET_WORD)); + memset(state.regs_found, 0, BITSET_WORDS(shader->reg_alloc) * + sizeof(BITSET_WORD)); + exec_list_validate(&shader->registers); + foreach_list_typed(nir_register, reg, node, &shader->registers) { + prevalidate_reg_decl(reg, true, &state); + } + + exec_list_validate(&shader->functions); + foreach_list_typed(nir_function, func, node, &shader->functions) { + validate_function(func, &state); + } + + foreach_list_typed(nir_register, reg, node, &shader->registers) { + postvalidate_reg_decl(reg, &state); + } + + destroy_validate_state(&state); +} + +#endif /* NDEBUG */ diff --git a/src/compiler/nir/nir_vla.h b/src/compiler/nir/nir_vla.h new file mode 100644 index 0000000..7537833 --- /dev/null +++ b/src/compiler/nir/nir_vla.h @@ -0,0 +1,54 @@ +/************************************************************************** + * + * Copyright 2015 VMware, Inc. + * All Rights Reserved. + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the + * "Software"), to deal in the Software without restriction, including + * without limitation the rights to use, copy, modify, merge, publish, + * distribute, sub license, and/or sell copies of the Software, and to + * permit persons to whom the Software is furnished to do so, subject to + * the following conditions: + * + * The above copyright notice and this permission notice (including the + * next paragraph) shall be included in all copies or substantial portions + * of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS + * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. + * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR + * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, + * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * + **************************************************************************/ + +#pragma once + + +#include "c99_alloca.h" + + +/* Declare a variable length array, with no initialization */ +#define NIR_VLA(_type, _name, _length) \ + _type *_name = alloca((_length) * sizeof *_name) + + +/* Declare a variable length array, and initialize it with the given byte. + * + * _length is evaluated twice, so expressions with side-effects must be + * avoided. + */ +#define NIR_VLA_FILL(_type, _name, _length, _byte) \ + _type *_name = memset(alloca((_length) * sizeof *_name), _byte, (_length) * sizeof *_name) + + +/* Declare a variable length array, and zero it. + * + * Just like NIR_VLA_FILL, _length is evaluated twice, so expressions with + * side-effects must be avoided. + */ +#define NIR_VLA_ZERO(_type, _name, _length) \ + NIR_VLA_FILL(_type, _name, _length, 0) diff --git a/src/compiler/nir/nir_worklist.c b/src/compiler/nir/nir_worklist.c new file mode 100644 index 0000000..3087a1d --- /dev/null +++ b/src/compiler/nir/nir_worklist.c @@ -0,0 +1,144 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand (jason@jlekstrand.net) + * + */ + +#include "nir_worklist.h" + +void +nir_block_worklist_init(nir_block_worklist *w, unsigned num_blocks, + void *mem_ctx) +{ + w->size = num_blocks; + w->count = 0; + w->start = 0; + + w->blocks_present = rzalloc_array(mem_ctx, BITSET_WORD, + BITSET_WORDS(num_blocks)); + w->blocks = ralloc_array(mem_ctx, nir_block *, num_blocks); +} + +void +nir_block_worklist_fini(nir_block_worklist *w) +{ + ralloc_free(w->blocks_present); + ralloc_free(w->blocks); +} + +static bool +worklist_add_block(nir_block *block, void *w) +{ + nir_block_worklist_push_tail(w, block); + + return true; +} + +void +nir_block_worklist_add_all(nir_block_worklist *w, nir_function_impl *impl) +{ + nir_foreach_block(impl, worklist_add_block, w); +} + +void +nir_block_worklist_push_head(nir_block_worklist *w, nir_block *block) +{ + /* Pushing a block we already have is a no-op */ + if (BITSET_TEST(w->blocks_present, block->index)) + return; + + assert(w->count < w->size); + + if (w->start == 0) + w->start = w->size - 1; + else + w->start--; + + w->count++; + + w->blocks[w->start] = block; + BITSET_SET(w->blocks_present, block->index); +} + +nir_block * +nir_block_worklist_peek_head(const nir_block_worklist *w) +{ + assert(w->count > 0); + + return w->blocks[w->start]; +} + +nir_block * +nir_block_worklist_pop_head(nir_block_worklist *w) +{ + assert(w->count > 0); + + unsigned head = w->start; + + w->start = (w->start + 1) % w->size; + w->count--; + + BITSET_CLEAR(w->blocks_present, w->blocks[head]->index); + return w->blocks[head]; +} + +void +nir_block_worklist_push_tail(nir_block_worklist *w, nir_block *block) +{ + /* Pushing a block we already have is a no-op */ + if (BITSET_TEST(w->blocks_present, block->index)) + return; + + assert(w->count < w->size); + + w->count++; + + unsigned tail = (w->start + w->count - 1) % w->size; + + w->blocks[tail] = block; + BITSET_SET(w->blocks_present, block->index); +} + +nir_block * +nir_block_worklist_peek_tail(const nir_block_worklist *w) +{ + assert(w->count > 0); + + unsigned tail = (w->start + w->count - 1) % w->size; + + return w->blocks[tail]; +} + +nir_block * +nir_block_worklist_pop_tail(nir_block_worklist *w) +{ + assert(w->count > 0); + + unsigned tail = (w->start + w->count - 1) % w->size; + + w->count--; + + BITSET_CLEAR(w->blocks_present, w->blocks[tail]->index); + return w->blocks[tail]; +} diff --git a/src/compiler/nir/nir_worklist.h b/src/compiler/nir/nir_worklist.h new file mode 100644 index 0000000..829bff2 --- /dev/null +++ b/src/compiler/nir/nir_worklist.h @@ -0,0 +1,91 @@ +/* + * Copyright © 2014 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + * + * Authors: + * Jason Ekstrand (jason@jlekstrand.net) + * + */ + +#pragma once + +#ifndef _NIR_WORKLIST_ +#define _NIR_WORKLIST_ + +#include "nir.h" + +#ifdef __cplusplus +extern "C" { +#endif + +/** Represents a double-ended queue of unique blocks + * + * The worklist datastructure guarantees that eacy block is in the queue at + * most once. Pushing a block onto either end of the queue is a no-op if + * the block is already in the queue. In order for this to work, the + * caller must ensure that the blocks are properly indexed. + */ +typedef struct { + /* The total size of the worklist */ + unsigned size; + + /* The number of blocks currently in the worklist */ + unsigned count; + + /* The offset in the array of blocks at which the list starts */ + unsigned start; + + /* A bitset of all of the blocks currently present in the worklist */ + BITSET_WORD *blocks_present; + + /* The actual worklist */ + nir_block **blocks; +} nir_block_worklist; + +void nir_block_worklist_init(nir_block_worklist *w, unsigned num_blocks, + void *mem_ctx); +void nir_block_worklist_fini(nir_block_worklist *w); + +void nir_block_worklist_add_all(nir_block_worklist *w, nir_function_impl *impl); + +static inline bool +nir_block_worklist_is_empty(const nir_block_worklist *w) +{ + return w->count == 0; +} + +void nir_block_worklist_push_head(nir_block_worklist *w, nir_block *block); + +nir_block *nir_block_worklist_peek_head(const nir_block_worklist *w); + +nir_block *nir_block_worklist_pop_head(nir_block_worklist *w); + +void nir_block_worklist_push_tail(nir_block_worklist *w, nir_block *block); + +nir_block *nir_block_worklist_peek_tail(const nir_block_worklist *w); + +nir_block *nir_block_worklist_pop_tail(nir_block_worklist *w); + +#ifdef __cplusplus +} /* extern "C" */ +#endif + +#endif /* _NIR_WORKLIST_ */ diff --git a/src/compiler/nir/tests/control_flow_tests.cpp b/src/compiler/nir/tests/control_flow_tests.cpp new file mode 100644 index 0000000..b9379ef --- /dev/null +++ b/src/compiler/nir/tests/control_flow_tests.cpp @@ -0,0 +1,148 @@ +/* + * Copyright © 2015 Intel Corporation + * + * Permission is hereby granted, free of charge, to any person obtaining a + * copy of this software and associated documentation files (the "Software"), + * to deal in the Software without restriction, including without limitation + * the rights to use, copy, modify, merge, publish, distribute, sublicense, + * and/or sell copies of the Software, and to permit persons to whom the + * Software is furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice (including the next + * paragraph) shall be included in all copies or substantial portions of the + * Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER + * DEALINGS IN THE SOFTWARE. + */ +#include <gtest/gtest.h> +#include "nir.h" +#include "nir_builder.h" + +class nir_cf_test : public ::testing::Test { +protected: + nir_cf_test(); + ~nir_cf_test(); + + nir_builder b; +}; + +nir_cf_test::nir_cf_test() +{ + static const nir_shader_compiler_options options = { }; + nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_VERTEX, &options); +} + +nir_cf_test::~nir_cf_test() +{ + ralloc_free(b.shader); +} + +TEST_F(nir_cf_test, delete_break_in_loop) +{ + /* Create IR: + * + * while (...) { break; } + */ + nir_loop *loop = nir_loop_create(b.shader); + nir_cf_node_insert(nir_after_cf_list(&b.impl->body), &loop->cf_node); + + b.cursor = nir_after_cf_list(&loop->body); + + nir_jump_instr *jump = nir_jump_instr_create(b.shader, nir_jump_break); + nir_builder_instr_insert(&b, &jump->instr); + + /* At this point, we should have: + * + * impl main { + * block block_0: + * // preds: + * // succs: block_1 + * loop { + * block block_1: + * // preds: block_0 + * break + * // succs: block_2 + * } + * block block_2: + * // preds: block_1 + * // succs: block_3 + * block block_3: + * } + */ + nir_block *block_0 = nir_start_block(b.impl); + nir_block *block_1 = nir_cf_node_as_block(nir_loop_first_cf_node(loop)); + nir_block *block_2 = nir_cf_node_as_block(nir_cf_node_next(&loop->cf_node)); + nir_block *block_3 = b.impl->end_block; + ASSERT_EQ(nir_cf_node_block, block_0->cf_node.type); + ASSERT_EQ(nir_cf_node_block, block_1->cf_node.type); + ASSERT_EQ(nir_cf_node_block, block_2->cf_node.type); + ASSERT_EQ(nir_cf_node_block, block_3->cf_node.type); + + /* Verify the successors and predecessors. */ + EXPECT_EQ(block_1, block_0->successors[0]); + EXPECT_EQ(NULL, block_0->successors[1]); + EXPECT_EQ(block_2, block_1->successors[0]); + EXPECT_EQ(NULL, block_1->successors[1]); + EXPECT_EQ(block_3, block_2->successors[0]); + EXPECT_EQ(NULL, block_2->successors[1]); + EXPECT_EQ(NULL, block_3->successors[0]); + EXPECT_EQ(NULL, block_3->successors[1]); + EXPECT_EQ(0, block_0->predecessors->entries); + EXPECT_EQ(1, block_1->predecessors->entries); + EXPECT_EQ(1, block_2->predecessors->entries); + EXPECT_EQ(1, block_3->predecessors->entries); + EXPECT_TRUE(_mesa_set_search(block_1->predecessors, block_0)); + EXPECT_TRUE(_mesa_set_search(block_2->predecessors, block_1)); + EXPECT_TRUE(_mesa_set_search(block_3->predecessors, block_2)); + + nir_print_shader(b.shader, stderr); + + /* Now remove the break. */ + nir_instr_remove(&jump->instr); + + nir_print_shader(b.shader, stderr); + + /* At this point, we should have: + * + * impl main { + * block block_0: + * // preds: + * // succs: block_1 + * loop { + * block block_1: + * // preds: block_0 block_1 + * // succs: block_1 + * } + * block block_2: + * // preds: block_1 + * // succs: block_3 + * block block_3: + * } + * + * Re-verify the predecessors and successors. + */ + EXPECT_EQ(block_1, block_0->successors[0]); + EXPECT_EQ(NULL, block_0->successors[1]); + EXPECT_EQ(block_1, block_1->successors[0]); /* back to itself */ + EXPECT_EQ(block_2, block_1->successors[1]); /* fake successor */ + EXPECT_EQ(block_3, block_2->successors[0]); + EXPECT_EQ(NULL, block_2->successors[1]); + EXPECT_EQ(NULL, block_3->successors[0]); + EXPECT_EQ(NULL, block_3->successors[1]); + EXPECT_EQ(0, block_0->predecessors->entries); + EXPECT_EQ(2, block_1->predecessors->entries); + EXPECT_EQ(1, block_2->predecessors->entries); + EXPECT_EQ(1, block_3->predecessors->entries); + EXPECT_TRUE(_mesa_set_search(block_1->predecessors, block_0)); + EXPECT_TRUE(_mesa_set_search(block_1->predecessors, block_1)); + EXPECT_TRUE(_mesa_set_search(block_2->predecessors, block_1)); + EXPECT_TRUE(_mesa_set_search(block_3->predecessors, block_2)); + + nir_metadata_require(b.impl, nir_metadata_dominance); +} |