i965: Add new pass to split vectors into scalar variables

Combined with the previous pass, this lets other optimization passes
do their work thanks to ir_tree_grafting.  Still have regression in
instruction count with INTEL_NEW_FS, but register count is even
better.

1 files changed, 388 insertions, 0 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_fs_vector_splitting.cpp b/src/mesa/drivers/dri/i965/brw_fs_vector_splitting.cpp
new file mode 100644
index 0000000..d4da86b
--- /dev/null
+++ b/src/mesa/drivers/dri/i965/brw_fs_vector_splitting.cpp
@@ -0,0 +1,388 @@
+/*
+ * Copyright © 2010 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.
+ */
+
+/**
+ * \file brw_wm_vector_splitting.cpp
+ *
+ * If a vector is only ever referenced by its components, then
+ * split those components out to individual variables so they can be
+ * handled normally by other optimization passes.
+ *
+ * This skips vectors in uniforms and varyings, which need to be
+ * accessible as vectors for their access by the GL.  Also, vector
+ * results of non-variable-derefs in assignments aren't handled
+ * because to do so we would have to store the vector result to a
+ * temporary in order to unload each channel, and to do so would just
+ * loop us back to where we started.  For the 965, this is exactly the
+ * behavior we want for the results of texture lookups, but probably not for
+ */
+
+extern "C" {
+#include "main/core.h"
+#include "intel_context.h"
+}
+#include "../glsl/ir.h"
+#include "../glsl/ir_visitor.h"
+#include "../glsl/ir_print_visitor.h"
+#include "../glsl/ir_rvalue_visitor.h"
+#include "../glsl/glsl_types.h"
+
+static bool debug = false;
+
+class variable_entry : public exec_node
+{
+public:
+   variable_entry(ir_variable *var)
+   {
+      this->var = var;
+      this->whole_vector_access = 0;
+      this->declaration = false;
+      this->mem_ctx = NULL;
+   }
+
+   ir_variable *var; /* The key: the variable's pointer. */
+
+   /** Number of times the variable is referenced, including assignments. */
+   unsigned whole_vector_access;
+
+   bool declaration; /* If the variable had a decl in the instruction stream */
+
+   ir_variable *components[4];
+
+   /** talloc_parent(this->var) -- the shader's talloc context. */
+   void *mem_ctx;
+};
+
+class ir_vector_reference_visitor : public ir_hierarchical_visitor {
+public:
+   ir_vector_reference_visitor(void)
+   {
+      this->mem_ctx = talloc_new(NULL);
+      this->variable_list.make_empty();
+   }
+
+   ~ir_vector_reference_visitor(void)
+   {
+      talloc_free(mem_ctx);
+   }
+
+   virtual ir_visitor_status visit(ir_variable *);
+   virtual ir_visitor_status visit(ir_dereference_variable *);
+   virtual ir_visitor_status visit_enter(ir_swizzle *);
+   virtual ir_visitor_status visit_enter(ir_assignment *);
+   virtual ir_visitor_status visit_enter(ir_function_signature *);
+
+   variable_entry *get_variable_entry(ir_variable *var);
+
+   /* List of variable_entry */
+   exec_list variable_list;
+
+   void *mem_ctx;
+};
+
+variable_entry *
+ir_vector_reference_visitor::get_variable_entry(ir_variable *var)
+{
+   assert(var);
+
+   if (!var->type->is_vector())
+      return NULL;
+
+   switch (var->mode) {
+   case ir_var_uniform:
+   case ir_var_in:
+   case ir_var_out:
+   case ir_var_inout:
+      /* Can't split varyings or uniforms.  Function in/outs won't get split
+       * either, so don't care about the ambiguity.
+       */
+      return NULL;
+   case ir_var_auto:
+   case ir_var_temporary:
+      break;
+   }
+
+   foreach_iter(exec_list_iterator, iter, this->variable_list) {
+      variable_entry *entry = (variable_entry *)iter.get();
+      if (entry->var == var)
+	 return entry;
+   }
+
+   variable_entry *entry = new(mem_ctx) variable_entry(var);
+   this->variable_list.push_tail(entry);
+   return entry;
+}
+
+
+ir_visitor_status
+ir_vector_reference_visitor::visit(ir_variable *ir)
+{
+   variable_entry *entry = this->get_variable_entry(ir);
+
+   if (entry)
+      entry->declaration = true;
+
+   return visit_continue;
+}
+
+ir_visitor_status
+ir_vector_reference_visitor::visit(ir_dereference_variable *ir)
+{
+   ir_variable *const var = ir->var;
+   variable_entry *entry = this->get_variable_entry(var);
+
+   if (entry)
+      entry->whole_vector_access++;
+
+   return visit_continue;
+}
+
+ir_visitor_status
+ir_vector_reference_visitor::visit_enter(ir_swizzle *ir)
+{
+   /* Don't descend into a vector ir_dereference_variable below. */
+   if (ir->val->as_dereference_variable() && ir->type->is_scalar())
+      return visit_continue_with_parent;
+
+   return visit_continue;
+}
+
+ir_visitor_status
+ir_vector_reference_visitor::visit_enter(ir_assignment *ir)
+{
+   if (ir->lhs->as_dereference_variable() &&
+       ir->rhs->as_dereference_variable() &&
+       !ir->condition) {
+      /* We'll split copies of a vector to copies of channels, so don't
+       * descend to the ir_dereference_variables.
+       */
+      return visit_continue_with_parent;
+   }
+   if (ir->lhs->as_dereference_variable() &&
+       is_power_of_two(ir->write_mask) &&
+       !ir->condition) {
+      /* If we're writing just a channel, then channel-splitting the LHS is OK.
+       */
+      ir->rhs->accept(this);
+      return visit_continue_with_parent;
+   }
+   return visit_continue;
+}
+
+ir_visitor_status
+ir_vector_reference_visitor::visit_enter(ir_function_signature *ir)
+{
+   /* We don't want to descend into the function parameters and
+    * split them, so just accept the body here.
+    */
+   visit_list_elements(this, &ir->body);
+   return visit_continue_with_parent;
+}
+
+class ir_vector_splitting_visitor : public ir_rvalue_visitor {
+public:
+   ir_vector_splitting_visitor(exec_list *vars)
+   {
+      this->variable_list = vars;
+   }
+
+   virtual ir_visitor_status visit_leave(ir_assignment *);
+
+   void handle_rvalue(ir_rvalue **rvalue);
+   struct variable_entry *get_splitting_entry(ir_variable *var);
+
+   exec_list *variable_list;
+   void *mem_ctx;
+};
+
+struct variable_entry *
+ir_vector_splitting_visitor::get_splitting_entry(ir_variable *var)
+{
+   assert(var);
+
+   if (!var->type->is_vector())
+      return NULL;
+
+   foreach_iter(exec_list_iterator, iter, *this->variable_list) {
+      variable_entry *entry = (variable_entry *)iter.get();
+      if (entry->var == var) {
+	 return entry;
+      }
+   }
+
+   return NULL;
+}
+
+void
+ir_vector_splitting_visitor::handle_rvalue(ir_rvalue **rvalue)
+{
+   if (!*rvalue)
+      return;
+
+   ir_swizzle *swiz = (*rvalue)->as_swizzle();
+   if (!swiz || !swiz->type->is_scalar())
+      return;
+
+   ir_dereference_variable *deref_var = swiz->val->as_dereference_variable();
+   if (!deref_var)
+      return;
+
+   variable_entry *entry = get_splitting_entry(deref_var->var);
+   if (!entry)
+      return;
+
+   ir_variable *var = entry->components[swiz->mask.x];
+   *rvalue = new(entry->mem_ctx) ir_dereference_variable(var);
+}
+
+ir_visitor_status
+ir_vector_splitting_visitor::visit_leave(ir_assignment *ir)
+{
+   ir_dereference_variable *lhs_deref = ir->lhs->as_dereference_variable();
+   ir_dereference_variable *rhs_deref = ir->rhs->as_dereference_variable();
+   variable_entry *lhs = lhs_deref ? get_splitting_entry(lhs_deref->var) : NULL;
+   variable_entry *rhs = rhs_deref ? get_splitting_entry(rhs_deref->var) : NULL;
+
+   if (lhs_deref && rhs_deref && (lhs || rhs) && !ir->condition) {
+      /* Straight assignment of vector variables. */
+      for (unsigned int i = 0; i < ir->rhs->type->vector_elements; i++) {
+	 ir_dereference *new_lhs;
+	 ir_rvalue *new_rhs;
+	 void *mem_ctx = lhs ? lhs->mem_ctx : rhs->mem_ctx;
+	 unsigned int writemask;
+
+	 if (lhs) {
+	    new_lhs = new(mem_ctx) ir_dereference_variable(lhs->components[i]);
+	    writemask = (ir->write_mask >> i) & 1;
+	 } else {
+	    new_lhs = ir->lhs->clone(mem_ctx, NULL);
+	    writemask = ir->write_mask & (1 << i);
+	 }
+
+	 if (rhs) {
+	    new_rhs = new(mem_ctx) ir_dereference_variable(rhs->components[i]);
+	 } else {
+	    new_rhs = new(mem_ctx) ir_swizzle(ir->rhs->clone(mem_ctx, NULL),
+					      i, i, i, i, 1);
+	 }
+
+	 ir->insert_before(new(mem_ctx) ir_assignment(new_lhs,
+						      new_rhs,
+						      NULL, writemask));
+      }
+      ir->remove();
+   } else if (lhs) {
+      int elem = -1;
+
+      switch (ir->write_mask) {
+      case (1 << 0):
+	 elem = 0;
+	 break;
+      case (1 << 1):
+	 elem = 1;
+	 break;
+      case (1 << 2):
+	 elem = 2;
+	 break;
+      case (1 << 3):
+	 elem = 3;
+	 break;
+      default:
+	 ir->print();
+	 assert(!"not reached: non-channelwise dereference of LHS.");
+      }
+
+      ir->lhs = new(mem_ctx) ir_dereference_variable(lhs->components[elem]);
+      ir->write_mask = (1 << 0);
+
+      handle_rvalue(&ir->rhs);
+      ir->rhs = new(mem_ctx) ir_swizzle(ir->rhs,
+					elem, elem, elem, elem, 1);
+   } else {
+      handle_rvalue(&ir->rhs);
+   }
+
+   handle_rvalue(&ir->condition);
+
+   return visit_continue;
+}
+
+extern "C" {
+bool
+brw_do_vector_splitting(exec_list *instructions)
+{
+   ir_vector_reference_visitor refs;
+
+   visit_list_elements(&refs, instructions);
+
+   /* Trim out variables we can't split. */
+   foreach_iter(exec_list_iterator, iter, refs.variable_list) {
+      variable_entry *entry = (variable_entry *)iter.get();
+
+      if (debug) {
+	 printf("vector %s@%p: decl %d, whole_access %d\n",
+		entry->var->name, (void *) entry->var, entry->declaration,
+		entry->whole_vector_access);
+      }
+
+      if (!entry->declaration || entry->whole_vector_access) {
+	 entry->remove();
+      }
+   }
+
+   if (refs.variable_list.is_empty())
+      return false;
+
+   void *mem_ctx = talloc_new(NULL);
+
+   /* Replace the decls of the vectors to be split with their split
+    * components.
+    */
+   foreach_iter(exec_list_iterator, iter, refs.variable_list) {
+      variable_entry *entry = (variable_entry *)iter.get();
+      const struct glsl_type *type;
+      type = glsl_type::get_instance(entry->var->type->base_type, 1, 1);
+
+      entry->mem_ctx = talloc_parent(entry->var);
+
+      for (unsigned int i = 0; i < entry->var->type->vector_elements; i++) {
+	 const char *name = talloc_asprintf(mem_ctx, "%s_%c",
+					    entry->var->name,
+					    "xyzw"[i]);
+
+	 entry->components[i] = new(entry->mem_ctx) ir_variable(type, name,
+								ir_var_temporary);
+	 entry->var->insert_before(entry->components[i]);
+      }
+
+      entry->var->remove();
+   }
+
+   ir_vector_splitting_visitor split(&refs.variable_list);
+   visit_list_elements(&split, instructions);
+
+   talloc_free(mem_ctx);
+
+   return true;
+}
+}