freedreno/a3xx/compiler: new compiler

The new compiler generates a dependency graph of instructions, including
a few meta-instructions to handle PHI and preserve some extra
information needed for register assignment, etc.

The depth pass assigned a weight/depth to each node (based on sum of
instruction cycles of a given node and all it's dependent nodes), which
is used to schedule instructions.  The scheduling takes into account the
minimum number of cycles/slots between dependent instructions, etc.
Which was something that could not be handled properly with the original
compiler (which was more of a naive TGSI translator than an actual
compiler).

The register assignment is currently split out as a standalone pass.  I
expect that it will be replaced at some point, once I figure out what to
do about relative addressing (which is currently the only thing that
should cause fallback to old compiler).

There are a couple new debug options for FD_MESA_DEBUG env var:

  optmsgs - enable debug prints in optimizer
  optdump - dump instruction graph in .dot format, for example:

http://people.freedesktop.org/~robclark/a3xx/frag-0000.dot.png
http://people.freedesktop.org/~robclark/a3xx/frag-0000.dot

At this point, thanks to proper handling of instruction scheduling, the
new compiler fixes a lot of things that were broken before, and does not
appear to break anything that was working before[1].  So even though it
is not finished, it seems useful to merge it in it's current state.

[1] Not merged in this commit, because I'm not sure if it really belongs
in mesa tree, but the following commit implements a simple shader
emulator, which I've used to compare the output of the new compiler to
the original compiler (ie. run it on all the TGSI shaders dumped out via
ST_DEBUG=tgsi with various games/apps):

https://github.com/freedreno/mesa/commit/163b6306b1660e05ece2f00d264a8393d99b6f12

Signed-off-by: Rob Clark <robclark@freedesktop.org>

17 files changed, 2777 insertions, 209 deletions

diff --git a/src/gallium/drivers/freedreno/Makefile.sources b/src/gallium/drivers/freedreno/Makefile.sources
index bd5d92f..05a12fb 100644
--- a/src/gallium/drivers/freedreno/Makefile.sources
+++ b/src/gallium/drivers/freedreno/Makefile.sources
@@ -43,4 +43,10 @@ a3xx_SOURCES := \
 	a3xx/fd3_util.c \
 	a3xx/fd3_zsa.c \
 	a3xx/disasm-a3xx.c \
+	a3xx/ir3_cp.c \
+	a3xx/ir3_depth.c \
+	a3xx/ir3_dump.c \
+	a3xx/ir3_flatten.c \
+	a3xx/ir3_ra.c \
+	a3xx/ir3_sched.c \
 	a3xx/ir3.c
diff --git a/src/gallium/drivers/freedreno/a3xx/disasm-a3xx.c b/src/gallium/drivers/freedreno/a3xx/disasm-a3xx.c
index 0e45ec5..8c3704b 100644
--- a/src/gallium/drivers/freedreno/a3xx/disasm-a3xx.c
+++ b/src/gallium/drivers/freedreno/a3xx/disasm-a3xx.c
@@ -735,6 +735,14 @@ struct opc_info {
 
 #define GETINFO(instr) (&(opcs[((instr)->opc_cat << NOPC_BITS) | instr_opc(instr)]))
 
+// XXX hack.. probably should move this table somewhere common:
+#include "ir3.h"
+const char *ir3_instr_name(struct ir3_instruction *instr)
+{
+	if (instr->category == -1) return "??meta??";
+	return opcs[(instr->category << NOPC_BITS) | instr->opc].name;
+}
+
 static void print_instr(uint32_t *dwords, int level, int n)
 {
 	instr_t *instr = (instr_t *)dwords;
diff --git a/src/gallium/drivers/freedreno/a3xx/fd3_compiler.c b/src/gallium/drivers/freedreno/a3xx/fd3_compiler.c
index 5ab34e5..da327c9 100644
--- a/src/gallium/drivers/freedreno/a3xx/fd3_compiler.c
+++ b/src/gallium/drivers/freedreno/a3xx/fd3_compiler.c
@@ -50,36 +50,43 @@
 struct fd3_compile_context {
 	const struct tgsi_token *tokens;
 	struct ir3_shader *ir;
-	struct ir3_block *block;
 	struct fd3_shader_stateobj *so;
 
+	struct ir3_block *block;
+	struct ir3_instruction *current_instr;
+
+	/* we need to defer updates to block->outputs[] until the end
+	 * of an instruction (so we don't see new value until *after*
+	 * the src registers are processed)
+	 */
+	struct {
+		struct ir3_instruction *instr, **instrp;
+	} output_updates[16];
+	unsigned num_output_updates;
+
+	/* are we in a sequence of "atomic" instructions?
+	 */
+	bool atomic;
+
+	/* For fragment shaders, from the hw perspective the only
+	 * actual input is r0.xy position register passed to bary.f.
+	 * But TGSI doesn't know that, it still declares things as
+	 * IN[] registers.  So we do all the input tracking normally
+	 * and fix things up after compile_instructions()
+	 */
+	struct ir3_instruction *frag_pos;
+
 	struct tgsi_parse_context parser;
 	unsigned type;
 
 	struct tgsi_shader_info info;
 
-	/* last input dst (for setting (ei) flag): */
-	struct ir3_register *last_input;
-
-	/* last instruction with relative addressing: */
-	struct ir3_instruction *last_rel;
-
 	/* for calculating input/output positions/linkages: */
 	unsigned next_inloc;
 
 	unsigned num_internal_temps;
 	struct tgsi_src_register internal_temps[6];
 
-	/* track registers which need to synchronize w/ "complex alu" cat3
-	 * instruction pipeline:
-	 */
-	regmask_t needs_ss;
-
-	/* track registers which need to synchronize with texture fetch
-	 * pipeline:
-	 */
-	regmask_t needs_sy;
-
 	/* inputs start at r0, temporaries start after last input, and
 	 * outputs start after last temporary.
 	 *
@@ -93,10 +100,8 @@ struct fd3_compile_context {
 	/* idx/slot for last compiler generated immediate */
 	unsigned immediate_idx;
 
-	/* stack of branch instructions that start (potentially nested)
-	 * branch instructions, so that we can fix up the branch targets
-	 * so that we can fix up the branch target on the corresponding
-	 * END instruction
+	/* stack of branch instructions that mark (potentially nested)
+	 * branch if/else/loop/etc
 	 */
 	struct ir3_instruction *branch[16];
 	unsigned int branch_count;
@@ -115,6 +120,7 @@ static void vectorize(struct fd3_compile_context *ctx,
 		int nsrcs, ...);
 static void create_mov(struct fd3_compile_context *ctx,
 		struct tgsi_dst_register *dst, struct tgsi_src_register *src);
+static type_t get_ftype(struct fd3_compile_context *ctx);
 
 static unsigned
 compile_init(struct fd3_compile_context *ctx, struct fd3_shader_stateobj *so,
@@ -125,27 +131,32 @@ compile_init(struct fd3_compile_context *ctx, struct fd3_shader_stateobj *so,
 
 	ctx->tokens = tokens;
 	ctx->ir = so->ir;
-	ctx->block = ir3_block_create(ctx->ir, 0, 0, 0);
 	ctx->so = so;
-	ctx->last_input = NULL;
-	ctx->last_rel = NULL;
 	ctx->next_inloc = 8;
 	ctx->num_internal_temps = 0;
 	ctx->branch_count = 0;
+	ctx->block = NULL;
+	ctx->current_instr = NULL;
+	ctx->num_output_updates = 0;
+	ctx->atomic = false;
 
-	regmask_init(&ctx->needs_ss);
-	regmask_init(&ctx->needs_sy);
 	memset(ctx->base_reg, 0, sizeof(ctx->base_reg));
 
 	tgsi_scan_shader(tokens, &ctx->info);
 
+#define FM(x) (1 << TGSI_FILE_##x)
+	/* optimize can't deal with relative addressing: */
+	if (info->indirect_files & (FM(TEMPORARY) | FM(INPUT) |
+			FM(OUTPUT) | FM(IMMEDIATE) | FM(CONSTANT)))
+		return TGSI_PARSE_ERROR;
+
 	/* Immediates go after constants: */
 	ctx->base_reg[TGSI_FILE_CONSTANT]  = 0;
 	ctx->base_reg[TGSI_FILE_IMMEDIATE] =
 			info->file_max[TGSI_FILE_CONSTANT] + 1;
 
 	/* if full precision and fragment shader, don't clobber
-	 * r0.x w/ bary fetch:
+	 * r0.xy w/ bary fetch:
 	 */
 	if ((so->type == SHADER_FRAGMENT) && !so->half_precision)
 		base = 1;
@@ -202,51 +213,269 @@ struct instr_translater {
 };
 
 static void
-handle_last_rel(struct fd3_compile_context *ctx)
+instr_finish(struct fd3_compile_context *ctx)
 {
-	if (ctx->last_rel) {
-		ctx->last_rel->flags |= IR3_INSTR_UL;
-		ctx->last_rel = NULL;
-	}
+	unsigned i;
+
+	if (ctx->atomic)
+		return;
+
+	for (i = 0; i < ctx->num_output_updates; i++)
+		*(ctx->output_updates[i].instrp) = ctx->output_updates[i].instr;
+
+	ctx->num_output_updates = 0;
+}
+
+/* For "atomic" groups of instructions, for example the four scalar
+ * instructions to perform a vec4 operation.  Basically this just
+ * blocks out handling of output_updates so the next scalar instruction
+ * still sees the result from before the start of the atomic group.
+ *
+ * NOTE: when used properly, this could probably replace get/put_dst()
+ * stuff.
+ */
+static void
+instr_atomic_start(struct fd3_compile_context *ctx)
+{
+	ctx->atomic = true;
+}
+
+static void
+instr_atomic_end(struct fd3_compile_context *ctx)
+{
+	ctx->atomic = false;
+	instr_finish(ctx);
 }
 
 static struct ir3_instruction *
 instr_create(struct fd3_compile_context *ctx, int category, opc_t opc)
 {
-	return ir3_instr_create(ctx->block, category, opc);
+	instr_finish(ctx);
+	return (ctx->current_instr = ir3_instr_create(ctx->block, category, opc));
+}
+
+static struct ir3_instruction *
+instr_clone(struct fd3_compile_context *ctx, struct ir3_instruction *instr)
+{
+	instr_finish(ctx);
+	return (ctx->current_instr = ir3_instr_clone(instr));
+}
+
+static struct ir3_block *
+push_block(struct fd3_compile_context *ctx)
+{
+	struct ir3_block *block;
+	unsigned ntmp, nin, nout;
+
+#define SCALAR_REGS(file) (4 * (ctx->info.file_max[TGSI_FILE_ ## file] + 1))
+
+	/* hmm, give ourselves room to create 4 extra temporaries (vec4):
+	 */
+	ntmp = SCALAR_REGS(TEMPORARY);
+	ntmp += 4 * 4;
+
+	/* for outermost block, 'inputs' are the actual shader INPUT
+	 * register file.  Reads from INPUT registers always go back to
+	 * top block.  For nested blocks, 'inputs' is used to track any
+	 * TEMPORARY file register from one of the enclosing blocks that
+	 * is ready in this block.
+	 */
+	if (!ctx->block) {
+		/* NOTE: fragment shaders actually have two inputs (r0.xy, the
+		 * position)
+		 */
+		nin = SCALAR_REGS(INPUT);
+		if (ctx->type == TGSI_PROCESSOR_FRAGMENT)
+			nin = MAX2(2, nin);
+	} else {
+		nin = ntmp;
+	}
+
+	nout = SCALAR_REGS(OUTPUT);
+
+	block = ir3_block_create(ctx->ir, ntmp, nin, nout);
+
+	block->parent = ctx->block;
+	ctx->block = block;
+
+	return block;
 }
 
 static void
-add_nop(struct fd3_compile_context *ctx, unsigned count)
+pop_block(struct fd3_compile_context *ctx)
 {
-	while (count-- > 0)
-		instr_create(ctx, 0, OPC_NOP);
+	ctx->block = ctx->block->parent;
+	compile_assert(ctx, ctx->block);
 }
 
-static unsigned
-src_flags(struct fd3_compile_context *ctx, struct ir3_register *reg)
+static void
+ssa_dst(struct fd3_compile_context *ctx, struct ir3_instruction *instr,
+		const struct tgsi_dst_register *dst, unsigned chan)
 {
-	unsigned flags = 0;
+	unsigned n = regid(dst->Index, chan);
+	unsigned idx = ctx->num_output_updates;
 
-	if (reg->flags & (IR3_REG_CONST | IR3_REG_IMMED))
-		return flags;
+	compile_assert(ctx, idx < ARRAY_SIZE(ctx->output_updates));
 
-	if (regmask_get(&ctx->needs_ss, reg)) {
-		flags |= IR3_INSTR_SS;
-		regmask_init(&ctx->needs_ss);
+	/* NOTE: defer update of temporaries[idx] or output[idx]
+	 * until instr_finish(), so that if the current instruction
+	 * reads the same TEMP/OUT[] it gets the old value:
+	 *
+	 * bleh.. this might be a bit easier to just figure out
+	 * in instr_finish().  But at that point we've already
+	 * lost information about OUTPUT vs TEMPORARY register
+	 * file..
+	 */
+
+	switch (dst->File) {
+	case TGSI_FILE_OUTPUT:
+		compile_assert(ctx, n < ctx->block->noutputs);
+		ctx->output_updates[idx].instrp = &ctx->block->outputs[n];
+		ctx->output_updates[idx].instr = instr;
+		ctx->num_output_updates++;
+		break;
+	case TGSI_FILE_TEMPORARY:
+		compile_assert(ctx, n < ctx->block->ntemporaries);
+		ctx->output_updates[idx].instrp = &ctx->block->temporaries[n];
+		ctx->output_updates[idx].instr = instr;
+		ctx->num_output_updates++;
+		break;
 	}
+}
+
+static struct ir3_instruction *
+create_output(struct ir3_block *block, struct ir3_instruction *instr,
+		unsigned n)
+{
+	struct ir3_instruction *out;
 
-	if (regmask_get(&ctx->needs_sy, reg)) {
-		flags |= IR3_INSTR_SY;
-		regmask_init(&ctx->needs_sy);
+	out = ir3_instr_create(block, -1, OPC_META_OUTPUT);
+	out->inout.block = block;
+	ir3_reg_create(out, n, 0);
+	if (instr)
+		ir3_reg_create(out, 0, IR3_REG_SSA)->instr = instr;
+
+	return out;
+}
+
+static struct ir3_instruction *
+create_input(struct ir3_block *block, struct ir3_instruction *instr,
+		unsigned n)
+{
+	struct ir3_instruction *in;
+
+	in = ir3_instr_create(block, -1, OPC_META_INPUT);
+	in->inout.block = block;
+	ir3_reg_create(in, n, 0);
+	if (instr)
+		ir3_reg_create(in, 0, IR3_REG_SSA)->instr = instr;
+
+	return in;
+}
+
+static struct ir3_instruction *
+block_input(struct ir3_block *block, unsigned n)
+{
+	/* references to INPUT register file always go back up to
+	 * top level:
+	 */
+	if (block->parent)
+		return block_input(block->parent, n);
+	return block->inputs[n];
+}
+
+/* return temporary in scope, creating if needed meta-input node
+ * to track block inputs
+ */
+static struct ir3_instruction *
+block_temporary(struct ir3_block *block, unsigned n)
+{
+	/* references to TEMPORARY register file, find the nearest
+	 * enclosing block which has already assigned this temporary,
+	 * creating meta-input instructions along the way to keep
+	 * track of block inputs
+	 */
+	if (block->parent && !block->temporaries[n]) {
+		/* if already have input for this block, reuse: */
+		if (!block->inputs[n])
+			block->inputs[n] = block_temporary(block->parent, n);
+
+		/* and create new input to return: */
+		return create_input(block, block->inputs[n], n);
 	}
+	return block->temporaries[n];
+}
 
-	return flags;
+static struct ir3_instruction *
+create_immed(struct fd3_compile_context *ctx, float val)
+{
+	/* this can happen when registers (or components of a TGSI
+	 * register) are used as src before they have been assigned
+	 * (undefined contents).  To avoid confusing the rest of the
+	 * compiler, and to generally keep things peachy, substitute
+	 * an instruction that sets the src to 0.0.  Or to keep
+	 * things undefined, I could plug in a random number? :-P
+	 *
+	 * NOTE: *don't* use instr_create() here!
+	 */
+	struct ir3_instruction *instr;
+	instr = ir3_instr_create(ctx->block, 1, 0);
+	instr->cat1.src_type = get_ftype(ctx);
+	instr->cat1.dst_type = get_ftype(ctx);
+	ir3_reg_create(instr, 0, 0);
+	ir3_reg_create(instr, 0, IR3_REG_IMMED)->fim_val = val;
+	return instr;
+}
+
+static void
+ssa_src(struct fd3_compile_context *ctx, struct ir3_register *reg,
+		const struct tgsi_src_register *src, unsigned chan)
+{
+	struct ir3_block *block = ctx->block;
+	unsigned n = regid(src->Index, chan);
+
+	switch (src->File) {
+	case TGSI_FILE_INPUT:
+		reg->flags |= IR3_REG_SSA;
+		reg->instr = block_input(ctx->block, n);
+		break;
+	case TGSI_FILE_OUTPUT:
+		/* really this should just happen in case of 'MOV_SAT OUT[n], ..',
+		 * for the following clamp instructions:
+		 */
+		reg->flags |= IR3_REG_SSA;
+		reg->instr = block->outputs[n];
+		/* we don't have to worry about read from an OUTPUT that was
+		 * assigned outside of the current block, because the _SAT
+		 * clamp instructions will always be in the same block as
+		 * the original instruction which wrote the OUTPUT
+		 */
+		compile_assert(ctx, reg->instr);
+		break;
+	case TGSI_FILE_TEMPORARY:
+		reg->flags |= IR3_REG_SSA;
+		reg->instr = block_temporary(ctx->block, n);
+		break;
+	}
+
+	if ((reg->flags & IR3_REG_SSA) && !reg->instr) {
+		/* this can happen when registers (or components of a TGSI
+		 * register) are used as src before they have been assigned
+		 * (undefined contents).  To avoid confusing the rest of the
+		 * compiler, and to generally keep things peachy, substitute
+		 * an instruction that sets the src to 0.0.  Or to keep
+		 * things undefined, I could plug in a random number? :-P
+		 *
+		 * NOTE: *don't* use instr_create() here!
+		 */
+		reg->instr = create_immed(ctx, 0.0);
+	}
 }
 
 static struct ir3_register *
-add_dst_reg(struct fd3_compile_context *ctx, struct ir3_instruction *instr,
-		const struct tgsi_dst_register *dst, unsigned chan)
+add_dst_reg_wrmask(struct fd3_compile_context *ctx,
+		struct ir3_instruction *instr, const struct tgsi_dst_register *dst,
+		unsigned chan, unsigned wrmask)
 {
 	unsigned flags = 0, num = 0;
 	struct ir3_register *reg;
@@ -272,15 +501,55 @@ add_dst_reg(struct fd3_compile_context *ctx, struct ir3_instruction *instr,
 
 	reg = ir3_reg_create(instr, regid(num, chan), flags);
 
-	if (dst->Indirect)
-		ctx->last_rel = instr;
+	/* NOTE: do not call ssa_dst() if atomic.. vectorize()
+	 * itself will call ssa_dst().  This is to filter out
+	 * the (initially bogus) .x component dst which is
+	 * created (but not necessarily used, ie. if the net
+	 * result of the vector operation does not write to
+	 * the .x component)
+	 */
+
+	reg->wrmask = wrmask;
+	if (wrmask == 0x1) {
+		/* normal case */
+		if (!ctx->atomic)
+			ssa_dst(ctx, instr, dst, chan);
+	} else if ((dst->File == TGSI_FILE_TEMPORARY) ||
+			(dst->File == TGSI_FILE_OUTPUT)) {
+		unsigned i;
+
+		/* if instruction writes multiple, we need to create
+		 * some place-holder collect the registers:
+		 */
+		for (i = 0; i < 4; i++) {
+			if (wrmask & (1 << i)) {
+				struct ir3_instruction *collect =
+						ir3_instr_create(ctx->block, -1, OPC_META_FO);
+				collect->fo.off = i;
+				/* unused dst reg: */
+				ir3_reg_create(collect, 0, 0);
+				/* and src reg used to hold original instr */
+				ir3_reg_create(collect, 0, IR3_REG_SSA)->instr = instr;
+				if (!ctx->atomic)
+					ssa_dst(ctx, collect, dst, chan+i);
+			}
+		}
+	}
 
 	return reg;
 }
 
 static struct ir3_register *
-add_src_reg(struct fd3_compile_context *ctx, struct ir3_instruction *instr,
-		const struct tgsi_src_register *src, unsigned chan)
+add_dst_reg(struct fd3_compile_context *ctx, struct ir3_instruction *instr,
+		const struct tgsi_dst_register *dst, unsigned chan)
+{
+	return add_dst_reg_wrmask(ctx, instr, dst, chan, 0x1);
+}
+
+static struct ir3_register *
+add_src_reg_wrmask(struct fd3_compile_context *ctx,
+		struct ir3_instruction *instr, const struct tgsi_src_register *src,
+		unsigned chan, unsigned wrmask)
 {
 	unsigned flags = 0, num = 0;
 	struct ir3_register *reg;
@@ -325,14 +594,49 @@ add_src_reg(struct fd3_compile_context *ctx, struct ir3_instruction *instr,
 
 	reg = ir3_reg_create(instr, regid(num, chan), flags);
 
-	if (src->Indirect)
-		ctx->last_rel = instr;
+	reg->wrmask = wrmask;
+	if (wrmask == 0x1) {
+		/* normal case */
+		ssa_src(ctx, reg, src, chan);
+	} else if ((src->File == TGSI_FILE_TEMPORARY) ||
+			(src->File == TGSI_FILE_OUTPUT) ||
+			(src->File == TGSI_FILE_INPUT)) {
+		struct ir3_instruction *collect;
+		unsigned i;
+
+		/* if instruction reads multiple, we need to create
+		 * some place-holder collect the registers:
+		 */
+		collect = ir3_instr_create(ctx->block, -1, OPC_META_FI);
+		ir3_reg_create(collect, 0, 0);   /* unused dst reg */
+
+		for (i = 0; i < 4; i++) {
+			if (wrmask & (1 << i)) {
+				/* and src reg used point to the original instr */
+				ssa_src(ctx, ir3_reg_create(collect, 0, IR3_REG_SSA),
+						src, chan + i);
+			} else if (wrmask & ~((i << i) - 1)) {
+				/* if any remaining components, then dummy
+				 * placeholder src reg to fill in the blanks:
+				 */
+				ir3_reg_create(collect, 0, 0);
+			}
+		}
 
-	instr->flags |= src_flags(ctx, reg);
+		reg->flags |= IR3_REG_SSA;
+		reg->instr = collect;
+	}
 
 	return reg;
 }
 
+static struct ir3_register *
+add_src_reg(struct fd3_compile_context *ctx, struct ir3_instruction *instr,
+		const struct tgsi_src_register *src, unsigned chan)
+{
+	return add_src_reg_wrmask(ctx, instr, src, chan, 0x1);
+}
+
 static void
 src_from_dst(struct tgsi_src_register *src, struct tgsi_dst_register *dst)
 {
@@ -542,8 +846,6 @@ create_mov(struct fd3_compile_context *ctx, struct tgsi_dst_register *dst,
 
 			add_dst_reg(ctx, instr, dst, i);
 			add_src_reg(ctx, instr, src, src_swiz(src, i));
-		} else {
-			add_nop(ctx, 1);
 		}
 	}
 }
@@ -620,7 +922,8 @@ vectorize(struct fd3_compile_context *ctx, struct ir3_instruction *instr,
 {
 	va_list ap;
 	int i, j, n = 0;
-	bool indirect = dst->Indirect;
+
+	instr_atomic_start(ctx);
 
 	add_dst_reg(ctx, instr, dst, TGSI_SWIZZLE_X);
 
@@ -636,7 +939,6 @@ vectorize(struct fd3_compile_context *ctx, struct ir3_instruction *instr,
 			reg->iim_val = *(int *)&src;
 		} else {
 			reg = add_src_reg(ctx, instr, src, TGSI_SWIZZLE_X);
-			indirect |= src->Indirect;
 		}
 		reg->flags |= flags & ~IR3_REG_NEGATE;
 		if (flags & IR3_REG_NEGATE)
@@ -651,37 +953,32 @@ vectorize(struct fd3_compile_context *ctx, struct ir3_instruction *instr,
 			if (n++ == 0) {
 				cur = instr;
 			} else {
-				cur = ir3_instr_clone(instr);
-				cur->flags &= ~(IR3_INSTR_SY | IR3_INSTR_SS | IR3_INSTR_JP);
+				cur = instr_clone(ctx, instr);
 			}
 
+			ssa_dst(ctx, cur, dst, i);
+
 			/* fix-up dst register component: */
 			cur->regs[0]->num = regid(cur->regs[0]->num >> 2, i);
 
 			/* fix-up src register component: */
 			va_start(ap, nsrcs);
 			for (j = 0; j < nsrcs; j++) {
+				struct ir3_register *reg = cur->regs[j+1];
 				struct tgsi_src_register *src =
 						va_arg(ap, struct tgsi_src_register *);
 				unsigned flags = va_arg(ap, unsigned);
-				if (!(flags & IR3_REG_IMMED)) {
-					cur->regs[j+1]->num =
-							regid(cur->regs[j+1]->num >> 2,
-									src_swiz(src, i));
-					cur->flags |= src_flags(ctx, cur->regs[j+1]);
+				if (reg->flags & IR3_REG_SSA) {
+					ssa_src(ctx, reg, src, src_swiz(src, i));
+				} else if (!(flags & IR3_REG_IMMED)) {
+					reg->num = regid(reg->num >> 2, src_swiz(src, i));
 				}
 			}
 			va_end(ap);
-
-			if (indirect)
-				ctx->last_rel = cur;
 		}
 	}
 
-	/* pad w/ nop's.. at least until we are clever enough to
-	 * figure out if we really need to..
-	 */
-	add_nop(ctx, 4 - n);
+	instr_atomic_end(ctx);
 }
 
 /*
@@ -718,8 +1015,6 @@ trans_arl(const struct instr_translater *t,
 	unsigned chan = src->SwizzleX;
 	compile_assert(ctx, dst->File == TGSI_FILE_ADDRESS);
 
-	handle_last_rel(ctx);
-
 	tmp_src = get_internal_temp_hr(ctx, &tmp_dst);
 
 	/* cov.{f32,f16}s16 Rtmp, Rsrc */
@@ -729,25 +1024,18 @@ trans_arl(const struct instr_translater *t,
 	add_dst_reg(ctx, instr, &tmp_dst, chan)->flags |= IR3_REG_HALF;
 	add_src_reg(ctx, instr, src, chan);
 
-	add_nop(ctx, 3);
-
 	/* shl.b Rtmp, Rtmp, 2 */
 	instr = instr_create(ctx, 2, OPC_SHL_B);
 	add_dst_reg(ctx, instr, &tmp_dst, chan)->flags |= IR3_REG_HALF;
 	add_src_reg(ctx, instr, tmp_src, chan)->flags |= IR3_REG_HALF;
 	ir3_reg_create(instr, 0, IR3_REG_IMMED)->iim_val = 2;
 
-	add_nop(ctx, 3);
-
 	/* mova a0, Rtmp */
 	instr = instr_create(ctx, 1, 0);
 	instr->cat1.src_type = TYPE_S16;
 	instr->cat1.dst_type = TYPE_S16;
 	add_dst_reg(ctx, instr, dst, 0)->flags |= IR3_REG_HALF;
 	add_src_reg(ctx, instr, tmp_src, chan)->flags |= IR3_REG_HALF;
-
-	/* need to ensure 5 instr slots before a0 is used: */
-	add_nop(ctx, 6);
 }
 
 /* texture fetch/sample instructions: */
@@ -756,7 +1044,6 @@ trans_samp(const struct instr_translater *t,
 		struct fd3_compile_context *ctx,
 		struct tgsi_full_instruction *inst)
 {
-	struct ir3_register *r;
 	struct ir3_instruction *instr;
 	struct tgsi_src_register *coord = &inst->Src[0].Register;
 	struct tgsi_src_register *samp  = &inst->Src[1].Register;
@@ -790,10 +1077,8 @@ trans_samp(const struct instr_translater *t,
 		break;
 	}
 
-	if ((tex == TGSI_TEXTURE_3D) || (tex == TGSI_TEXTURE_CUBE)) {
-		add_nop(ctx, 3);
+	if ((tex == TGSI_TEXTURE_3D) || (tex == TGSI_TEXTURE_CUBE))
 		flags |= IR3_INSTR_3D;
-	}
 
 	/* cat5 instruction cannot seem to handle const or relative: */
 	if (is_rel_or_const(coord))
@@ -829,8 +1114,6 @@ trans_samp(const struct instr_translater *t,
 		}
 
 		coord = tmp_src;
-
-		add_nop(ctx, 4 - j);
 	}
 
 	instr = instr_create(ctx, 5, t->opc);
@@ -839,13 +1122,10 @@ trans_samp(const struct instr_translater *t,
 	instr->cat5.tex  = samp->Index;
 	instr->flags |= flags;
 
-	r = add_dst_reg(ctx, instr, &inst->Dst[0].Register, 0);
-	r->wrmask = inst->Dst[0].Register.WriteMask;
+	add_dst_reg_wrmask(ctx, instr, &inst->Dst[0].Register, 0,
+			inst->Dst[0].Register.WriteMask);
 
-	add_src_reg(ctx, instr, coord, coord->SwizzleX)->wrmask = src_wrmask;
-
-	/* after add_src_reg() so we don't set (sy) on sam instr itself! */
-	regmask_set(&ctx->needs_sy, r);
+	add_src_reg_wrmask(ctx, instr, coord, coord->SwizzleX, src_wrmask);
 }
 
 /*
@@ -981,45 +1261,18 @@ trans_cmp(const struct instr_translater *t,
  * Conditional / Flow control
  */
 
-static unsigned
-find_instruction(struct fd3_compile_context *ctx, struct ir3_instruction *instr)
-{
-	unsigned i;
-	for (i = 0; i < ctx->ir->instrs_count; i++)
-		if (ctx->ir->instrs[i] == instr)
-			return i;
-	return ~0;
-}
-
 static void
 push_branch(struct fd3_compile_context *ctx, struct ir3_instruction *instr)
 {
 	ctx->branch[ctx->branch_count++] = instr;
 }
 
-static void
+static struct ir3_instruction *
 pop_branch(struct fd3_compile_context *ctx)
 {
-	struct ir3_instruction *instr;
-
-	/* if we were clever enough, we'd patch this up after the fact,
-	 * and set (jp) flag on whatever the next instruction was, rather
-	 * than inserting an extra nop..
-	 */
-	instr = instr_create(ctx, 0, OPC_NOP);
-	instr->flags |= IR3_INSTR_JP;
-
-	/* pop the branch instruction from the stack and fix up branch target: */
-	instr = ctx->branch[--ctx->branch_count];
-	instr->cat0.immed = ctx->ir->instrs_count - find_instruction(ctx, instr) - 1;
+	return ctx->branch[--ctx->branch_count];
 }
 
-/* We probably don't really want to translate if/else/endif into branches..
- * the blob driver evaluates both legs of the if and then uses the sel
- * instruction to pick which sides of the branch to "keep".. but figuring
- * that out will take somewhat more compiler smarts.  So hopefully branches
- * don't kill performance too badly.
- */
 static void
 trans_if(const struct instr_translater *t,
 		struct fd3_compile_context *ctx,
@@ -1027,21 +1280,36 @@ trans_if(const struct instr_translater *t,
 {
 	struct ir3_instruction *instr;
 	struct tgsi_src_register *src = &inst->Src[0].Register;
+	struct tgsi_dst_register tmp_dst;
+	struct tgsi_src_register *tmp_src;
 	struct tgsi_src_register constval;
 
 	get_immediate(ctx, &constval, fui(0.0));
+	tmp_src = get_internal_temp(ctx, &tmp_dst);
 
 	if (is_const(src))
 		src = get_unconst(ctx, src);
 
+	/* cmps.f.eq tmp0, b, {0.0} */
 	instr = instr_create(ctx, 2, OPC_CMPS_F);
-	ir3_reg_create(instr, regid(REG_P0, 0), 0);
+	add_dst_reg(ctx, instr, &tmp_dst, 0);
 	add_src_reg(ctx, instr, src, src->SwizzleX);
 	add_src_reg(ctx, instr, &constval, constval.SwizzleX);
 	instr->cat2.condition = IR3_COND_EQ;
 
-	instr = instr_create(ctx, 0, OPC_BR);
+	/* add.s tmp0, tmp0, -1 */
+	instr = instr_create(ctx, 2, OPC_ADD_S);
+	add_dst_reg(ctx, instr, &tmp_dst, TGSI_SWIZZLE_X);
+	add_src_reg(ctx, instr, tmp_src, TGSI_SWIZZLE_X);
+	ir3_reg_create(instr, 0, IR3_REG_IMMED)->iim_val = -1;
+
+	/* meta:flow tmp0 */
+	instr = instr_create(ctx, -1, OPC_META_FLOW);
+	ir3_reg_create(instr, 0, 0);  /* dummy dst */
+	add_src_reg(ctx, instr, tmp_src, TGSI_SWIZZLE_X);
+
 	push_branch(ctx, instr);
+	instr->flow.if_block = push_block(ctx);
 }
 
 static void
@@ -1051,11 +1319,63 @@ trans_else(const struct instr_translater *t,
 {
 	struct ir3_instruction *instr;
 
-	/* for first half of if/else/endif, generate a jump past the else: */
-	instr = instr_create(ctx, 0, OPC_JUMP);
+	pop_block(ctx);
+
+	instr = pop_branch(ctx);
+
+	compile_assert(ctx, (instr->category == -1) &&
+			(instr->opc == OPC_META_FLOW));
 
-	pop_branch(ctx);
 	push_branch(ctx, instr);
+	instr->flow.else_block = push_block(ctx);
+}
+
+static struct ir3_instruction *
+find_temporary(struct ir3_block *block, unsigned n)
+{
+	if (block->parent && !block->temporaries[n])
+		return find_temporary(block->parent, n);
+	return block->temporaries[n];
+}
+
+static struct ir3_instruction *
+find_output(struct ir3_block *block, unsigned n)
+{
+	if (block->parent && !block->outputs[n])
+		return find_output(block->parent, n);
+	return block->outputs[n];
+}
+
+static struct ir3_instruction *
+create_phi(struct fd3_compile_context *ctx, struct ir3_instruction *cond,
+		struct ir3_instruction *a, struct ir3_instruction *b)
+{
+	struct ir3_instruction *phi;
+
+	compile_assert(ctx, cond);
+
+	/* Either side of the condition could be null..  which
+	 * indicates a variable written on only one side of the
+	 * branch.  Normally this should only be variables not
+	 * used outside of that side of the branch.  So we could
+	 * just 'return a ? a : b;' in that case.  But for better
+	 * defined undefined behavior we just stick in imm{0.0}.
+	 * In the common case of a value only used within the
+	 * one side of the branch, the PHI instruction will not
+	 * get scheduled
+	 */
+	if (!a)
+		a = create_immed(ctx, 0.0);
+	if (!b)
+		b = create_immed(ctx, 0.0);
+
+	phi = instr_create(ctx, -1, OPC_META_PHI);
+	ir3_reg_create(phi, 0, 0);  /* dummy dst */
+	ir3_reg_create(phi, 0, IR3_REG_SSA)->instr = cond;
+	ir3_reg_create(phi, 0, IR3_REG_SSA)->instr = a;
+	ir3_reg_create(phi, 0, IR3_REG_SSA)->instr = b;
+
+	return phi;
 }
 
 static void
@@ -1063,7 +1383,120 @@ trans_endif(const struct instr_translater *t,
 		struct fd3_compile_context *ctx,
 		struct tgsi_full_instruction *inst)
 {
-	pop_branch(ctx);
+	struct ir3_instruction *instr;
+	struct ir3_block *ifb, *elseb;
+	struct ir3_instruction **ifout, **elseout;
+	unsigned i, ifnout = 0, elsenout = 0;
+
+	pop_block(ctx);
+
+	instr = pop_branch(ctx);
+
+	compile_assert(ctx, (instr->category == -1) &&
+			(instr->opc == OPC_META_FLOW));
+
+	ifb = instr->flow.if_block;
+	elseb = instr->flow.else_block;
+	/* if there is no else block, the parent block is used for the
+	 * branch-not-taken src of the PHI instructions:
+	 */
+	if (!elseb)
+		elseb = ifb->parent;
+
+	/* count up number of outputs for each block: */
+	for (i = 0; i < ifb->ntemporaries; i++) {
+		if (ifb->temporaries[i])
+			ifnout++;
+		if (elseb->temporaries[i])
+			elsenout++;
+	}
+	for (i = 0; i < ifb->noutputs; i++) {
+		if (ifb->outputs[i])
+			ifnout++;
+		if (elseb->outputs[i])
+			elsenout++;
+	}
+
+	ifout = ir3_alloc(ctx->ir, sizeof(ifb->outputs[0]) * ifnout);
+	if (elseb != ifb->parent)
+		elseout = ir3_alloc(ctx->ir, sizeof(ifb->outputs[0]) * elsenout);
+
+	ifnout = 0;
+	elsenout = 0;
+
+	/* generate PHI instructions for any temporaries written: */
+	for (i = 0; i < ifb->ntemporaries; i++) {
+		struct ir3_instruction *a = ifb->temporaries[i];
+		struct ir3_instruction *b = elseb->temporaries[i];
+
+		/* if temporary written in if-block, or if else block
+		 * is present and temporary written in else-block:
+		 */
+		if (a || ((elseb != ifb->parent) && b)) {
+			struct ir3_instruction *phi;
+
+			/* if only written on one side, find the closest
+			 * enclosing update on other side:
+			 */
+			if (!a)
+				a = find_temporary(ifb, i);
+			if (!b)
+				b = find_temporary(elseb, i);
+
+			ifout[ifnout] = a;
+			a = create_output(ifb, a, ifnout++);
+
+			if (elseb != ifb->parent) {
+				elseout[elsenout] = b;
+				b = create_output(elseb, b, elsenout++);
+			}
+
+			phi = create_phi(ctx, instr, a, b);
+			ctx->block->temporaries[i] = phi;
+		}
+	}
+
+	/* .. and any outputs written: */
+	for (i = 0; i < ifb->noutputs; i++) {
+		struct ir3_instruction *a = ifb->outputs[i];
+		struct ir3_instruction *b = elseb->outputs[i];
+
+		/* if output written in if-block, or if else block
+		 * is present and output written in else-block:
+		 */
+		if (a || ((elseb != ifb->parent) && b)) {
+			struct ir3_instruction *phi;
+
+			/* if only written on one side, find the closest
+			 * enclosing update on other side:
+			 */
+			if (!a)
+				a = find_output(ifb, i);
+			if (!b)
+				b = find_output(elseb, i);
+
+			ifout[ifnout] = a;
+			a = create_output(ifb, a, ifnout++);
+
+			if (elseb != ifb->parent) {
+				elseout[elsenout] = b;
+				b = create_output(elseb, b, elsenout++);
+			}
+
+			phi = create_phi(ctx, instr, a, b);
+			ctx->block->outputs[i] = phi;
+		}
+	}
+
+	ifb->noutputs = ifnout;
+	ifb->outputs = ifout;
+
+	if (elseb != ifb->parent) {
+		elseb->noutputs = elsenout;
+		elseb->outputs = elseout;
+	}
+
+	// TODO maybe we want to compact block->inputs?
 }
 
 /*
@@ -1161,23 +1594,6 @@ instr_cat2(const struct instr_translater *t,
 	put_dst(ctx, inst, dst);
 }
 
-static bool is_mad(opc_t opc)
-{
-	switch (opc) {
-	case OPC_MAD_U16:
-	case OPC_MADSH_U16:
-	case OPC_MAD_S16:
-	case OPC_MADSH_M16:
-	case OPC_MAD_U24:
-	case OPC_MAD_S24:
-	case OPC_MAD_F16:
-	case OPC_MAD_F32:
-		return true;
-	default:
-		return false;
-	}
-}
-
 static void
 instr_cat3(const struct instr_translater *t,
 		struct fd3_compile_context *ctx,
@@ -1217,27 +1633,21 @@ instr_cat4(const struct instr_translater *t,
 	struct tgsi_dst_register *dst = get_dst(ctx, inst);
 	struct tgsi_src_register *src = &inst->Src[0].Register;
 	struct ir3_instruction *instr;
-	unsigned i, n;
+	unsigned i;
 
 	/* seems like blob compiler avoids const as src.. */
 	if (is_const(src))
 		src = get_unconst(ctx, src);
 
-	/* worst case: */
-	add_nop(ctx, 6);
-
 	/* we need to replicate into each component: */
-	for (i = 0, n = 0; i < 4; i++) {
+	for (i = 0; i < 4; i++) {
 		if (dst->WriteMask & (1 << i)) {
-			if (n++)
-				add_nop(ctx, 1);
 			instr = instr_create(ctx, 4, t->opc);
 			add_dst_reg(ctx, instr, dst, i);
 			add_src_reg(ctx, instr, src, src->SwizzleX);
 		}
 	}
 
-	regmask_set(&ctx->needs_ss, instr->regs[0]);
 	put_dst(ctx, inst, dst);
 }
 
@@ -1287,13 +1697,12 @@ decl_semantic(const struct tgsi_declaration_semantic *sem)
 	return fd3_semantic_name(sem->Name, sem->Index);
 }
 
-static int
+static void
 decl_in(struct fd3_compile_context *ctx, struct tgsi_full_declaration *decl)
 {
 	struct fd3_shader_stateobj *so = ctx->so;
 	unsigned base = ctx->base_reg[TGSI_FILE_INPUT];
 	unsigned i, flags = 0;
-	int nop = 0;
 
 	/* I don't think we should get frag shader input without
 	 * semantic info?  Otherwise how do inputs get linked to
@@ -1308,12 +1717,12 @@ decl_in(struct fd3_compile_context *ctx, struct tgsi_full_declaration *decl)
 	for (i = decl->Range.First; i <= decl->Range.Last; i++) {
 		unsigned n = so->inputs_count++;
 		unsigned r = regid(i + base, 0);
-		unsigned ncomp;
+		unsigned ncomp, j;
 
 		/* TODO use ctx->info.input_usage_mask[decl->Range.n] to figure out ncomp: */
 		ncomp = 4;
 
-		DBG("decl in -> r%d", i + base);   // XXX
+		DBG("decl in -> r%d", i + base);
 
 		so->inputs[n].semantic = decl_semantic(&decl->Semantic);
 		so->inputs[n].compmask = (1 << ncomp) - 1;
@@ -1323,33 +1732,33 @@ decl_in(struct fd3_compile_context *ctx, struct tgsi_full_declaration *decl)
 
 		so->total_in += ncomp;
 
-		/* for frag shaders, we need to generate the corresponding bary instr: */
-		if (ctx->type == TGSI_PROCESSOR_FRAGMENT) {
-			unsigned j;
+		for (j = 0; j < ncomp; j++) {
+			struct ir3_instruction *instr;
 
-			for (j = 0; j < ncomp; j++) {
-				struct ir3_instruction *instr;
-				struct ir3_register *dst;
+			if (ctx->type == TGSI_PROCESSOR_FRAGMENT) {
+				struct ir3_register *src;
 
 				instr = instr_create(ctx, 2, OPC_BARY_F);
 
 				/* dst register: */
-				dst = ir3_reg_create(instr, r + j, flags);
-				ctx->last_input = dst;
+				ir3_reg_create(instr, r + j, flags);
 
 				/* input position: */
 				ir3_reg_create(instr, 0, IR3_REG_IMMED)->iim_val =
 						so->inputs[n].inloc + j - 8;
 
 				/* input base (always r0.xy): */
-				ir3_reg_create(instr, regid(0,0), 0)->wrmask = 0x3;
+				src = ir3_reg_create(instr, regid(0,0), IR3_REG_SSA);
+				src->wrmask = 0x3;
+				src->instr = ctx->frag_pos;
+
+			} else {
+				instr = create_input(ctx->block, NULL, (i * 4) + j);
 			}
 
-			nop = 6;
+			ctx->block->inputs[(i * 4) + j] = instr;
 		}
 	}
-
-	return nop;
 }
 
 static void
@@ -1361,9 +1770,9 @@ decl_out(struct fd3_compile_context *ctx, struct tgsi_full_declaration *decl)
 	unsigned name = decl->Semantic.Name;
 	unsigned i;
 
-	compile_assert(ctx, decl->Declaration.Semantic);  // TODO is this ever not true?
+	compile_assert(ctx, decl->Declaration.Semantic);
 
-	DBG("decl out[%d] -> r%d", name, decl->Range.First + base);   // XXX
+	DBG("decl out[%d] -> r%d", name, decl->Range.First + base);
 
 	if (ctx->type == TGSI_PROCESSOR_VERTEX) {
 		switch (name) {
@@ -1396,8 +1805,19 @@ decl_out(struct fd3_compile_context *ctx, struct tgsi_full_declaration *decl)
 
 	for (i = decl->Range.First; i <= decl->Range.Last; i++) {
 		unsigned n = so->outputs_count++;
+		unsigned ncomp, j;
+
+		ncomp = 4;
+
 		so->outputs[n].semantic = decl_semantic(&decl->Semantic);
 		so->outputs[n].regid = regid(i + base, comp);
+
+		/* avoid undefined outputs, stick a dummy mov from imm{0.0},
+		 * which if the output is actually assigned will be over-
+		 * written
+		 */
+		for (j = 0; j < ncomp; j++)
+			ctx->block->outputs[(i * 4) + j] = create_immed(ctx, 0.0);
 	}
 }
 
@@ -1410,8 +1830,19 @@ decl_samp(struct fd3_compile_context *ctx, struct tgsi_full_declaration *decl)
 static void
 compile_instructions(struct fd3_compile_context *ctx)
 {
-	struct ir3_shader *ir = ctx->ir;
-	int nop = 0;
+	push_block(ctx);
+
+	/* for fragment shader, we have a single input register (r0.xy)
+	 * which is used as the base for bary.f varying fetch instrs:
+	 */
+	if (ctx->type == TGSI_PROCESSOR_FRAGMENT) {
+		struct ir3_instruction *instr;
+		instr = ir3_instr_create(ctx->block, -1, OPC_META_FI);
+		ir3_reg_create(instr, 0, 0);
+		ir3_reg_create(instr, 0, IR3_REG_SSA);    /* r0.x */
+		ir3_reg_create(instr, 0, IR3_REG_SSA);    /* r0.y */
+		ctx->frag_pos = instr;
+	}
 
 	while (!tgsi_parse_end_of_tokens(&ctx->parser)) {
 		tgsi_parse_token(&ctx->parser);
@@ -1423,7 +1854,7 @@ compile_instructions(struct fd3_compile_context *ctx)
 			if (decl->Declaration.File == TGSI_FILE_OUTPUT) {
 				decl_out(ctx, decl);
 			} else if (decl->Declaration.File == TGSI_FILE_INPUT) {
-				nop = decl_in(ctx, decl);
+				decl_in(ctx, decl);
 			} else if (decl->Declaration.File == TGSI_FILE_SAMPLER) {
 				decl_samp(ctx, decl);
 			}
@@ -1446,9 +1877,6 @@ compile_instructions(struct fd3_compile_context *ctx)
 			unsigned opc = inst->Instruction.Opcode;
 			const struct instr_translater *t = &translaters[opc];
 
-			add_nop(ctx, nop);
-			nop = 0;
-
 			if (t->fxn) {
 				t->fxn(t, ctx, inst);
 				ctx->num_internal_temps = 0;
@@ -1468,6 +1896,8 @@ compile_instructions(struct fd3_compile_context *ctx)
 				break;
 			}
 
+			instr_finish(ctx);
+
 			break;
 		}
 		default:
@@ -1475,13 +1905,38 @@ compile_instructions(struct fd3_compile_context *ctx)
 		}
 	}
 
-	if (ir->instrs_count > 0)
-		ir->instrs[0]->flags |= IR3_INSTR_SS | IR3_INSTR_SY;
+	/* fixup actual inputs for frag shader: */
+	if (ctx->type == TGSI_PROCESSOR_FRAGMENT) {
+		struct ir3_instruction *instr;
+
+		ctx->block->ninputs = 2;
 
-	if (ctx->last_input)
-		ctx->last_input->flags |= IR3_REG_EI;
+		/* r0.x */
+		instr = create_input(ctx->block, NULL, 0);
+		ctx->block->inputs[0] = instr;
+		ctx->frag_pos->regs[1]->instr = instr;
 
-	handle_last_rel(ctx);
+		/* r0.y */
+		instr = create_input(ctx->block, NULL, 1);
+		ctx->block->inputs[1] = instr;
+		ctx->frag_pos->regs[2]->instr = instr;
+	}
+}
+
+static void
+compile_dump(struct fd3_compile_context *ctx)
+{
+	const char *name = (ctx->so->type == SHADER_VERTEX) ? "vert" : "frag";
+	static unsigned n = 0;
+	char fname[16];
+	FILE *f;
+	snprintf(fname, sizeof(fname), "%s-%04u.dot", name, n++);
+	f = fopen(fname, "w");
+	if (!f)
+		return;
+	ir3_block_depth(ctx->block);
+	ir3_shader_dump(ctx->ir, name, ctx->block, f);
+	fclose(f);
 }
 
 int
@@ -1489,6 +1944,8 @@ fd3_compile_shader(struct fd3_shader_stateobj *so,
 		const struct tgsi_token *tokens)
 {
 	struct fd3_compile_context ctx;
+	unsigned i, actual_in;
+	int ret = 0;
 
 	assert(!so->ir);
 
@@ -1496,12 +1953,91 @@ fd3_compile_shader(struct fd3_shader_stateobj *so,
 
 	assert(so->ir);
 
-	if (compile_init(&ctx, so, tokens) != TGSI_PARSE_OK)
-		return -1;
+	if (compile_init(&ctx, so, tokens) != TGSI_PARSE_OK) {
+		ret = -1;
+		goto out;
+	}
 
 	compile_instructions(&ctx);
 
+	if (fd_mesa_debug & FD_DBG_OPTDUMP)
+		compile_dump(&ctx);
+
+	ret = ir3_block_flatten(ctx.block);
+	if (ret < 0)
+		goto out;
+	if ((ret > 0) && (fd_mesa_debug & FD_DBG_OPTDUMP))
+		compile_dump(&ctx);
+
+	ir3_block_cp(ctx.block);
+
+	if (fd_mesa_debug & FD_DBG_OPTDUMP)
+		compile_dump(&ctx);
+
+	ir3_block_depth(ctx.block);
+
+	if (fd_mesa_debug & FD_DBG_OPTMSGS) {
+		printf("AFTER DEPTH:\n");
+		ir3_dump_instr_list(ctx.block->head);
+	}
+
+	ir3_block_sched(ctx.block);
+
+	if (fd_mesa_debug & FD_DBG_OPTMSGS) {
+		printf("AFTER SCHED:\n");
+		ir3_dump_instr_list(ctx.block->head);
+	}
+
+	ret = ir3_block_ra(ctx.block, so->type);
+	if (ret)
+		goto out;
+
+	if (fd_mesa_debug & FD_DBG_OPTMSGS) {
+		printf("AFTER RA:\n");
+		ir3_dump_instr_list(ctx.block->head);
+	}
+
+	/* fixup input/outputs: */
+	for (i = 0; i < so->outputs_count; i++) {
+		so->outputs[i].regid = ctx.block->outputs[i*4]->regs[0]->num;
+		/* preserve hack for depth output.. tgsi writes depth to .z,
+		 * but what we give the hw is the scalar register:
+		 */
+		if ((ctx.type == TGSI_PROCESSOR_FRAGMENT) &&
+			(sem2name(so->outputs[i].semantic) == TGSI_SEMANTIC_POSITION))
+			so->outputs[i].regid += 2;
+	}
+	/* Note that some or all channels of an input may be unused: */
+	actual_in = 0;
+	for (i = 0; i < so->inputs_count; i++) {
+		unsigned j, regid = ~0, compmask = 0;
+		for (j = 0; j < 4; j++) {
+			struct ir3_instruction *in = ctx.block->inputs[(i*4) + j];
+			if (in) {
+				compmask |= (1 << j);
+				regid = in->regs[0]->num - j;
+				actual_in++;
+			}
+		}
+		so->inputs[i].regid = regid;
+		so->inputs[i].compmask = compmask;
+	}
+
+	/* fragment shader always gets full vec4's even if it doesn't
+	 * fetch all components, but vertex shader we need to update
+	 * with the actual number of components fetch, otherwise thing
+	 * will hang due to mismaptch between VFD_DECODE's and
+	 * TOTALATTRTOVS
+	 */
+	if (so->type == SHADER_VERTEX)
+		so->total_in = actual_in;
+
+out:
+	if (ret) {
+		ir3_shader_destroy(so->ir);
+		so->ir = NULL;
+	}
 	compile_free(&ctx);
 
-	return 0;
+	return ret;
 }
diff --git a/src/gallium/drivers/freedreno/a3xx/fd3_compiler_old.c b/src/gallium/drivers/freedreno/a3xx/fd3_compiler_old.c
index b5715be..9ace26e 100644
--- a/src/gallium/drivers/freedreno/a3xx/fd3_compiler_old.c
+++ b/src/gallium/drivers/freedreno/a3xx/fd3_compiler_old.c
@@ -1161,23 +1161,6 @@ instr_cat2(const struct instr_translater *t,
 	put_dst(ctx, inst, dst);
 }
 
-static bool is_mad(opc_t opc)
-{
-	switch (opc) {
-	case OPC_MAD_U16:
-	case OPC_MADSH_U16:
-	case OPC_MAD_S16:
-	case OPC_MADSH_M16:
-	case OPC_MAD_U24:
-	case OPC_MAD_S24:
-	case OPC_MAD_F16:
-	case OPC_MAD_F32:
-		return true;
-	default:
-		return false;
-	}
-}
-
 static void
 instr_cat3(const struct instr_translater *t,
 		struct fd3_compile_context *ctx,
diff --git a/src/gallium/drivers/freedreno/a3xx/fd3_program.c b/src/gallium/drivers/freedreno/a3xx/fd3_program.c
index d177614..793e703 100644
--- a/src/gallium/drivers/freedreno/a3xx/fd3_program.c
+++ b/src/gallium/drivers/freedreno/a3xx/fd3_program.c
@@ -80,9 +80,9 @@ fixup_vp_regfootprint(struct fd3_shader_stateobj *so)
 {
 	unsigned i;
 	for (i = 0; i < so->inputs_count; i++)
-		so->info.max_reg = MAX2(so->info.max_reg, so->inputs[i].regid >> 2);
+		so->info.max_reg = MAX2(so->info.max_reg, (so->inputs[i].regid + 3) >> 2);
 	for (i = 0; i < so->outputs_count; i++)
-		so->info.max_reg = MAX2(so->info.max_reg, so->outputs[i].regid >> 2);
+		so->info.max_reg = MAX2(so->info.max_reg, (so->outputs[i].regid + 3) >> 2);
 }
 
 static struct fd3_shader_stateobj *
diff --git a/src/gallium/drivers/freedreno/a3xx/instr-a3xx.h b/src/gallium/drivers/freedreno/a3xx/instr-a3xx.h
index b0f7834..a79998e 100644
--- a/src/gallium/drivers/freedreno/a3xx/instr-a3xx.h
+++ b/src/gallium/drivers/freedreno/a3xx/instr-a3xx.h
@@ -190,6 +190,22 @@ typedef enum {
 	OPC_LDC_4 = 30,
 	OPC_LDLV = 31,
 
+	/* meta instructions (category -1): */
+	/* placeholder instr to mark inputs/outputs: */
+	OPC_META_INPUT = 0,
+	OPC_META_OUTPUT = 1,
+	/* The "fan-in" and "fan-out" instructions are used for keeping
+	 * track of instructions that write to multiple dst registers
+	 * (fan-out) like texture sample instructions, or read multiple
+	 * consecutive scalar registers (fan-in) (bary.f, texture samp)
+	 */
+	OPC_META_FO = 2,
+	OPC_META_FI = 3,
+	/* branches/flow control */
+	OPC_META_FLOW = 4,
+	OPC_META_PHI = 5,
+
+
 } opc_t;
 
 typedef enum {
@@ -643,4 +659,21 @@ static inline uint32_t instr_opc(instr_t *instr)
 	}
 }
 
+static inline bool is_mad(opc_t opc)
+{
+	switch (opc) {
+	case OPC_MAD_U16:
+	case OPC_MADSH_U16:
+	case OPC_MAD_S16:
+	case OPC_MADSH_M16:
+	case OPC_MAD_U24:
+	case OPC_MAD_S24:
+	case OPC_MAD_F16:
+	case OPC_MAD_F32:
+		return true;
+	default:
+		return false;
+	}
+}
+
 #endif /* INSTR_A3XX_H_ */
diff --git a/src/gallium/drivers/freedreno/a3xx/ir3.c b/src/gallium/drivers/freedreno/a3xx/ir3.c
index 2a06d42..2a68a8e 100644
--- a/src/gallium/drivers/freedreno/a3xx/ir3.c
+++ b/src/gallium/drivers/freedreno/a3xx/ir3.c
@@ -36,7 +36,7 @@
 /* simple allocator to carve allocations out of an up-front allocated heap,
  * so that we can free everything easily in one shot.
  */
-static void * ir3_alloc(struct ir3_shader *shader, int sz)
+void * ir3_alloc(struct ir3_shader *shader, int sz)
 {
 	void *ptr = &shader->heap[shader->heap_idx];
 	shader->heap_idx += align(sz, 4);
diff --git a/src/gallium/drivers/freedreno/a3xx/ir3.h b/src/gallium/drivers/freedreno/a3xx/ir3.h
index 896bec1..ccd3b0b 100644
--- a/src/gallium/drivers/freedreno/a3xx/ir3.h
+++ b/src/gallium/drivers/freedreno/a3xx/ir3.h
@@ -65,6 +65,11 @@ struct ir3_register {
 		 * that the shader needs no more input:
 		 */
 		IR3_REG_EI     = 0x200,
+		/* meta-flags, for intermediate stages of IR, ie.
+		 * before register assignment is done:
+		 */
+		IR3_REG_SSA    = 0x1000,   /* 'instr' is ptr to assigning instr */
+		IR3_REG_IA     = 0x2000,   /* meta-input dst is "assigned" */
 	} flags;
 	union {
 		/* normal registers:
@@ -77,6 +82,10 @@ struct ir3_register {
 		float   fim_val;
 		/* relative: */
 		int offset;
+		/* for IR3_REG_SSA, src registers contain ptr back to
+		 * assigning instruction.
+		 */
+		struct ir3_instruction *instr;
 	};
 
 	/* used for cat5 instructions, but also for internal/IR level
@@ -139,6 +148,10 @@ struct ir3_instruction {
 		IR3_INSTR_P     = 0x080,
 		IR3_INSTR_S     = 0x100,
 		IR3_INSTR_S2EN  = 0x200,
+		/* meta-flags, for intermediate stages of IR, ie.
+		 * before register assignment is done:
+		 */
+		IR3_INSTR_MARK  = 0x1000,
 	} flags;
 	int repeat;
 	unsigned regs_count;
@@ -171,7 +184,33 @@ struct ir3_instruction {
 			int offset;
 			int iim_val;
 		} cat6;
+		/* for meta-instructions, just used to hold extra data
+		 * before instruction scheduling, etc
+		 */
+		struct {
+			int off;              /* component/offset */
+		} fo;
+		struct {
+			struct ir3_block *if_block, *else_block;
+		} flow;
+		struct {
+			struct ir3_block *block;
+		} inout;
 	};
+
+	/* transient values used during various algorithms: */
+	union {
+		/* The instruction depth is the max dependency distance to output.
+		 *
+		 * You can also think of it as the "cost", if we did any sort of
+		 * optimization for register footprint.  Ie. a value that is  just
+		 * result of moving a const to a reg would have a low cost,  so to
+		 * it could make sense to duplicate the instruction at various
+		 * points where the result is needed to reduce register footprint.
+		 */
+		unsigned depth;
+	};
+	struct ir3_instruction *next;
 #ifdef DEBUG
 	uint32_t serialno;
 #endif
@@ -201,6 +240,7 @@ struct ir3_shader * ir3_shader_create(void);
 void ir3_shader_destroy(struct ir3_shader *shader);
 void * ir3_shader_assemble(struct ir3_shader *shader,
 		struct ir3_shader_info *info);
+void * ir3_alloc(struct ir3_shader *shader, int sz);
 
 struct ir3_block * ir3_block_create(struct ir3_shader *shader,
 		unsigned ntmp, unsigned nin, unsigned nout);
@@ -208,11 +248,44 @@ struct ir3_block * ir3_block_create(struct ir3_shader *shader,
 struct ir3_instruction * ir3_instr_create(struct ir3_block *block,
 		int category, opc_t opc);
 struct ir3_instruction * ir3_instr_clone(struct ir3_instruction *instr);
+const char *ir3_instr_name(struct ir3_instruction *instr);
 
 struct ir3_register * ir3_reg_create(struct ir3_instruction *instr,
 		int num, int flags);
 
 
+static inline bool ir3_instr_check_mark(struct ir3_instruction *instr)
+{
+	if (instr->flags & IR3_INSTR_MARK)
+		return true;  /* already visited */
+	instr->flags ^= IR3_INSTR_MARK;
+	return false;
+}
+
+static inline void ir3_shader_clear_mark(struct ir3_shader *shader)
+{
+	/* TODO would be nice to drop the instruction array.. for
+	 * new compiler, _clear_mark() is all we use it for, and
+	 * we could probably manage a linked list instead..
+	 */
+	unsigned i;
+	for (i = 0; i < shader->instrs_count; i++) {
+		struct ir3_instruction *instr = shader->instrs[i];
+		instr->flags &= ~IR3_INSTR_MARK;
+	}
+}
+
+static inline int ir3_instr_regno(struct ir3_instruction *instr,
+		struct ir3_register *reg)
+{
+	unsigned i;
+	for (i = 0; i < instr->regs_count; i++)
+		if (reg == instr->regs[i])
+			return i;
+	return -1;
+}
+
+
 /* comp:
  *   0 - x
  *   1 - y
@@ -254,6 +327,15 @@ static inline bool is_input(struct ir3_instruction *instr)
 	return (instr->category == 2) && (instr->opc == OPC_BARY_F);
 }
 
+static inline bool is_meta(struct ir3_instruction *instr)
+{
+	/* TODO how should we count PHI (and maybe fan-in/out) which
+	 * might actually contribute some instructions to the final
+	 * result?
+	 */
+	return (instr->category == -1);
+}
+
 static inline bool is_gpr(struct ir3_register *reg)
 {
 	return !(reg->flags & (IR3_REG_CONST | IR3_REG_IMMED));
@@ -262,13 +344,39 @@ static inline bool is_gpr(struct ir3_register *reg)
 /* TODO combine is_gpr()/reg_gpr().. */
 static inline bool reg_gpr(struct ir3_register *r)
 {
-	if (r->flags & (IR3_REG_CONST | IR3_REG_IMMED | IR3_REG_RELATIV))
+	if (r->flags & (IR3_REG_CONST | IR3_REG_IMMED | IR3_REG_RELATIV | IR3_REG_SSA))
 		return false;
 	if ((reg_num(r) == REG_A0) || (reg_num(r) == REG_P0))
 		return false;
 	return true;
 }
 
+/* dump: */
+#include <stdio.h>
+void ir3_shader_dump(struct ir3_shader *shader, const char *name,
+		struct ir3_block *block /* XXX maybe 'block' ptr should move to ir3_shader? */,
+		FILE *f);
+void ir3_dump_instr_single(struct ir3_instruction *instr);
+void ir3_dump_instr_list(struct ir3_instruction *instr);
+
+/* flatten if/else: */
+int ir3_block_flatten(struct ir3_block *block);
+
+/* depth calculation: */
+int ir3_delayslots(struct ir3_instruction *assigner,
+		struct ir3_instruction *consumer, unsigned n);
+void ir3_block_depth(struct ir3_block *block);
+
+/* copy-propagate: */
+void ir3_block_cp(struct ir3_block *block);
+
+/* scheduling: */
+void ir3_block_sched(struct ir3_block *block);
+
+/* register assignment: */
+int ir3_block_ra(struct ir3_block *block, enum shader_t type);
+
+
 #ifndef ARRAY_SIZE
 #  define ARRAY_SIZE(arr) (sizeof(arr) / sizeof((arr)[0]))
 #endif
diff --git a/src/gallium/drivers/freedreno/a3xx/ir3_cp.c b/src/gallium/drivers/freedreno/a3xx/ir3_cp.c
new file mode 100644
index 0000000..81f6c90
--- /dev/null
+++ b/src/gallium/drivers/freedreno/a3xx/ir3_cp.c
@@ -0,0 +1,155 @@
+/* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
+
+/*
+ * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
+ *
+ * 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 "ir3.h"
+
+/*
+ * Copy Propagate:
+ *
+ * TODO probably want some sort of visitor sort of interface to
+ * avoid duplicating the same graph traversal logic everywhere..
+ *
+ */
+
+static void block_cp(struct ir3_block *block);
+static struct ir3_instruction * instr_cp(struct ir3_instruction *instr, bool keep);
+
+static bool is_eligible_mov(struct ir3_instruction *instr)
+{
+	if ((instr->category == 1) &&
+			(instr->cat1.src_type == instr->cat1.dst_type)) {
+		struct ir3_register *src = instr->regs[1];
+		if ((src->flags & IR3_REG_SSA) &&
+				/* TODO: propagate abs/neg modifiers if possible */
+				!(src->flags & (IR3_REG_ABS | IR3_REG_NEGATE)))
+			return true;
+	}
+	return false;
+}
+
+static void walk_children(struct ir3_instruction *instr, bool keep)
+{
+	unsigned i;
+
+	/* walk down the graph from each src: */
+	for (i = 1; i < instr->regs_count; i++) {
+		struct ir3_register *src = instr->regs[i];
+		if (src->flags & IR3_REG_SSA)
+			src->instr = instr_cp(src->instr, keep);
+	}
+}
+
+static struct ir3_instruction *
+instr_cp_fanin(struct ir3_instruction *instr)
+{
+	unsigned i;
+
+	/* we need to handle fanin specially, to detect cases
+	 * when we need to keep a mov
+	 */
+
+	for (i = 1; i < instr->regs_count; i++) {
+		struct ir3_register *src = instr->regs[i];
+		if (src->flags & IR3_REG_SSA) {
+			struct ir3_instruction *cand =
+					instr_cp(src->instr, false);
+
+			/* if the candidate is a fanout, then keep
+			 * the move.
+			 *
+			 * This is a bit, um, fragile, but it should
+			 * catch the extra mov's that the front-end
+			 * puts in for us already in these cases.
+			 */
+			if (is_meta(cand) && (cand->opc == OPC_META_FO))
+				cand = instr_cp(src->instr, true);
+
+			src->instr = cand;
+		}
+	}
+
+	walk_children(instr, false);
+
+	return instr;
+
+}
+
+static struct ir3_instruction *
+instr_cp(struct ir3_instruction *instr, bool keep)
+{
+	/* if we've already visited this instruction, bail now: */
+	if (ir3_instr_check_mark(instr))
+		return instr;
+
+	if (is_meta(instr) && (instr->opc == OPC_META_FI))
+		return instr_cp_fanin(instr);
+
+	if (is_eligible_mov(instr) && !keep) {
+		struct ir3_register *src = instr->regs[1];
+		return instr_cp(src->instr, false);
+	}
+
+	walk_children(instr, false);
+
+	return instr;
+}
+
+static void block_cp(struct ir3_block *block)
+{
+	unsigned i, j;
+
+	for (i = 0; i < block->noutputs; i++) {
+		if (block->outputs[i]) {
+			struct ir3_instruction *out =
+					instr_cp(block->outputs[i], false);
+
+			/* To deal with things like this:
+			 *
+			 *   43: MOV OUT[2], TEMP[5]
+			 *   44: MOV OUT[0], TEMP[5]
+			 *
+			 * we need to ensure that no two outputs point to
+			 * the same instruction
+			 */
+			for (j = 0; j < i; j++) {
+				if (block->outputs[j] == out) {
+					out = instr_cp(block->outputs[i], true);
+					break;
+				}
+			}
+
+			block->outputs[i] = out;
+		}
+	}
+}
+
+void ir3_block_cp(struct ir3_block *block)
+{
+	ir3_shader_clear_mark(block->shader);
+	block_cp(block);
+}
diff --git a/src/gallium/drivers/freedreno/a3xx/ir3_depth.c b/src/gallium/drivers/freedreno/a3xx/ir3_depth.c
new file mode 100644
index 0000000..580ae08
--- /dev/null
+++ b/src/gallium/drivers/freedreno/a3xx/ir3_depth.c
@@ -0,0 +1,156 @@
+/* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
+
+/*
+ * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
+ *
+ * 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 "util/u_math.h"
+
+#include "ir3.h"
+
+/*
+ * Instruction Depth:
+ *
+ * Calculates weighted instruction depth, ie. the sum of # of needed
+ * instructions plus delay slots back to original input (ie INPUT or
+ * CONST).  That is to say, an instructions depth is:
+ *
+ *   depth(instr) {
+ *     d = 0;
+ *     // for each src register:
+ *     foreach (src in instr->regs[1..n])
+ *       d = max(d, delayslots(src->instr, n) + depth(src->instr));
+ *     return d + 1;
+ *   }
+ *
+ * After an instruction's depth is calculated, it is inserted into the
+ * blocks depth sorted list, which is used by the scheduling pass.
+ */
+
+/* calculate required # of delay slots between the instruction that
+ * assigns a value and the one that consumes
+ */
+int ir3_delayslots(struct ir3_instruction *assigner,
+		struct ir3_instruction *consumer, unsigned n)
+{
+	/* worst case is cat1-3 (alu) -> cat4/5 needing 6 cycles, normal
+	 * alu -> alu needs 3 cycles, cat4 -> alu and texture fetch
+	 * handled with sync bits
+	 */
+
+	if (is_meta(assigner))
+		return 0;
+
+	/* handled via sync flags: */
+	if (is_sfu(assigner) || is_tex(assigner))
+		return 0;
+
+	/* assigner must be alu: */
+	if (is_sfu(consumer) || is_tex(consumer)) {
+		return 8;
+	} else if ((consumer->category == 3) &&
+			is_mad(consumer->opc) && (n == 2)) {
+		/* special case, 3rd src to cat3 not required on first cycle */
+		return 2;
+	} else {
+		return 5;
+	}
+}
+
+static void insert_by_depth(struct ir3_instruction *instr)
+{
+	struct ir3_block *block = instr->block;
+	struct ir3_instruction *n = block->head;
+	struct ir3_instruction *p = NULL;
+
+	while (n && (n != instr) && (n->depth > instr->depth)) {
+		p = n;
+		n = n->next;
+	}
+
+	instr->next = n;
+	if (p)
+		p->next = instr;
+	else
+		block->head = instr;
+}
+
+static void ir3_instr_depth(struct ir3_instruction *instr)
+{
+	unsigned i;
+
+	/* if we've already visited this instruction, bail now: */
+	if (ir3_instr_check_mark(instr))
+		return;
+
+	instr->depth = 0;
+
+	for (i = 1; i < instr->regs_count; i++) {
+		struct ir3_register *src = instr->regs[i];
+		if (src->flags & IR3_REG_SSA) {
+			unsigned sd;
+
+			/* visit child to compute it's depth: */
+			ir3_instr_depth(src->instr);
+
+			sd = ir3_delayslots(src->instr, instr, i-1) +
+					src->instr->depth;
+
+			instr->depth = MAX2(instr->depth, sd);
+		}
+	}
+
+	/* meta-instructions don't add cycles, other than PHI.. which
+	 * might translate to a real instruction..
+	 *
+	 * well, not entirely true, fan-in/out, etc might need to need
+	 * to generate some extra mov's in edge cases, etc.. probably
+	 * we might want to do depth calculation considering the worst
+	 * case for these??
+	 */
+	if (!is_meta(instr))
+		instr->depth++;
+
+	insert_by_depth(instr);
+}
+
+void ir3_block_depth(struct ir3_block *block)
+{
+	unsigned i;
+
+	block->head = NULL;
+
+	ir3_shader_clear_mark(block->shader);
+	for (i = 0; i < block->noutputs; i++)
+		if (block->outputs[i])
+			ir3_instr_depth(block->outputs[i]);
+
+	/* at this point, any unvisited input is unused: */
+	for (i = 0; i < block->ninputs; i++) {
+		struct ir3_instruction *in = block->inputs[i];
+		if (in && !ir3_instr_check_mark(in))
+			block->inputs[i] = NULL;
+	}
+}
diff --git a/src/gallium/drivers/freedreno/a3xx/ir3_dump.c b/src/gallium/drivers/freedreno/a3xx/ir3_dump.c
new file mode 100644
index 0000000..0afd048
--- /dev/null
+++ b/src/gallium/drivers/freedreno/a3xx/ir3_dump.c
@@ -0,0 +1,416 @@
+/* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
+
+/*
+ * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
+ *
+ * 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 <stdarg.h>
+
+#include "ir3.h"
+
+#define PTRID(x) ((unsigned long)(x))
+
+struct ir3_dump_ctx {
+	FILE *f;
+	bool verbose;
+};
+
+static void dump_instr_name(struct ir3_dump_ctx *ctx,
+		struct ir3_instruction *instr)
+{
+	/* for debugging: */
+	if (ctx->verbose) {
+#ifdef DEBUG
+		fprintf(ctx->f, "%04u:", instr->serialno);
+#endif
+		fprintf(ctx->f, "%03u: ", instr->depth);
+	}
+
+	if (is_meta(instr)) {
+		switch(instr->opc) {
+		case OPC_META_PHI:
+			fprintf(ctx->f, "&#934;");
+			break;
+		default:
+			/* shouldn't hit here.. just for debugging: */
+			switch (instr->opc) {
+			case OPC_META_INPUT:  fprintf(ctx->f, "_meta:in");   break;
+			case OPC_META_OUTPUT: fprintf(ctx->f, "_meta:out");  break;
+			case OPC_META_FO:     fprintf(ctx->f, "_meta:fo");   break;
+			case OPC_META_FI:     fprintf(ctx->f, "_meta:fi");   break;
+			case OPC_META_FLOW:   fprintf(ctx->f, "_meta:flow"); break;
+			case OPC_META_PHI:    fprintf(ctx->f, "_meta:phi");  break;
+
+			default: fprintf(ctx->f, "_meta:%d", instr->opc); break;
+			}
+			break;
+		}
+	} else if (instr->category == 1) {
+		static const char *type[] = {
+				[TYPE_F16] = "f16",
+				[TYPE_F32] = "f32",
+				[TYPE_U16] = "u16",
+				[TYPE_U32] = "u32",
+				[TYPE_S16] = "s16",
+				[TYPE_S32] = "s32",
+				[TYPE_U8]  = "u8",
+				[TYPE_S8]  = "s8",
+		};
+		if (instr->cat1.src_type == instr->cat1.dst_type)
+			fprintf(ctx->f, "mov");
+		else
+			fprintf(ctx->f, "cov");
+		fprintf(ctx->f, ".%s%s", type[instr->cat1.src_type], type[instr->cat1.dst_type]);
+	} else {
+		fprintf(ctx->f, "%s", ir3_instr_name(instr));
+		if (instr->flags & IR3_INSTR_3D)
+			fprintf(ctx->f, ".3d");
+		if (instr->flags & IR3_INSTR_A)
+			fprintf(ctx->f, ".a");
+		if (instr->flags & IR3_INSTR_O)
+			fprintf(ctx->f, ".o");
+		if (instr->flags & IR3_INSTR_P)
+			fprintf(ctx->f, ".p");
+		if (instr->flags & IR3_INSTR_S)
+			fprintf(ctx->f, ".s");
+		if (instr->flags & IR3_INSTR_S2EN)
+			fprintf(ctx->f, ".s2en");
+	}
+}
+
+static void dump_reg_name(struct ir3_dump_ctx *ctx,
+		struct ir3_register *reg)
+{
+	if ((reg->flags & IR3_REG_ABS) && (reg->flags & IR3_REG_NEGATE))
+		fprintf(ctx->f, "(absneg)");
+	else if (reg->flags & IR3_REG_NEGATE)
+		fprintf(ctx->f, "(neg)");
+	else if (reg->flags & IR3_REG_ABS)
+		fprintf(ctx->f, "(abs)");
+
+	if (reg->flags & IR3_REG_IMMED) {
+		fprintf(ctx->f, "imm[%f,%d,0x%x]", reg->fim_val, reg->iim_val, reg->iim_val);
+	} else if (reg->flags & IR3_REG_SSA) {
+		if (ctx->verbose) {
+			fprintf(ctx->f, "_[");
+			dump_instr_name(ctx, reg->instr);
+			fprintf(ctx->f, "]");
+		}
+	} else {
+		if (reg->flags & IR3_REG_HALF)
+			fprintf(ctx->f, "h");
+		if (reg->flags & IR3_REG_CONST)
+			fprintf(ctx->f, "c%u.%c", reg_num(reg), "xyzw"[reg_comp(reg)]);
+		else
+			fprintf(ctx->f, "r%u.%c", reg_num(reg), "xyzw"[reg_comp(reg)]);
+	}
+}
+
+static void ir3_instr_dump(struct ir3_dump_ctx *ctx,
+		struct ir3_instruction *instr);
+static void ir3_block_dump(struct ir3_dump_ctx *ctx,
+		struct ir3_block *block, const char *name);
+
+static void dump_instr(struct ir3_dump_ctx *ctx,
+		struct ir3_instruction *instr)
+{
+	/* if we've already visited this instruction, bail now: */
+	if (ir3_instr_check_mark(instr))
+		return;
+
+	/* some meta-instructions need to be handled specially: */
+	if (is_meta(instr)) {
+		if ((instr->opc == OPC_META_FO) ||
+				(instr->opc == OPC_META_FI)) {
+			unsigned i;
+			for (i = 1; i < instr->regs_count; i++) {
+				struct ir3_register *reg = instr->regs[i];
+				if (reg->flags & IR3_REG_SSA)
+					dump_instr(ctx, reg->instr);
+			}
+		} else if (instr->opc == OPC_META_FLOW) {
+			struct ir3_register *reg = instr->regs[1];
+			ir3_block_dump(ctx, instr->flow.if_block, "if");
+			if (instr->flow.else_block)
+				ir3_block_dump(ctx, instr->flow.else_block, "else");
+			if (reg->flags & IR3_REG_SSA)
+				dump_instr(ctx, reg->instr);
+		} else if (instr->opc == OPC_META_PHI) {
+			/* treat like a normal instruction: */
+			ir3_instr_dump(ctx, instr);
+		}
+	} else {
+		ir3_instr_dump(ctx, instr);
+	}
+}
+
+/* arrarraggh!  if link is to something outside of the current block, we
+ * need to defer emitting the link until the end of the block, since the
+ * edge triggers pre-creation of the node it links to inside the cluster,
+ * even though it is meant to be outside..
+ */
+static struct {
+	char buf[40960];
+	unsigned n;
+} edge_buf;
+
+/* helper to print or defer: */
+static void printdef(struct ir3_dump_ctx *ctx,
+		bool defer, const char *fmt, ...)
+{
+	va_list ap;
+	va_start(ap, fmt);
+	if (defer) {
+		unsigned n = edge_buf.n;
+		n += vsnprintf(&edge_buf.buf[n], sizeof(edge_buf.buf) - n,
+				fmt, ap);
+		edge_buf.n = n;
+	} else {
+		vfprintf(ctx->f, fmt, ap);
+	}
+	va_end(ap);
+}
+
+static void dump_link2(struct ir3_dump_ctx *ctx,
+		struct ir3_instruction *instr, const char *target, bool defer)
+{
+	/* some meta-instructions need to be handled specially: */
+	if (is_meta(instr)) {
+		if (instr->opc == OPC_META_INPUT) {
+			printdef(ctx, defer, "input%lx:<in%u>:w -> %s",
+					PTRID(instr->inout.block),
+					instr->regs[0]->num, target);
+		} else if (instr->opc == OPC_META_FO) {
+			struct ir3_register *reg = instr->regs[1];
+			dump_link2(ctx, reg->instr, target, defer);
+			printdef(ctx, defer, "[label=\".%c\"]",
+					"xyzw"[instr->fo.off & 0x3]);
+		} else if (instr->opc == OPC_META_FI) {
+			unsigned i;
+
+			/* recursively dump all parents and links */
+			for (i = 1; i < instr->regs_count; i++) {
+				struct ir3_register *reg = instr->regs[i];
+				if (reg->flags & IR3_REG_SSA) {
+					dump_link2(ctx, reg->instr, target, defer);
+					printdef(ctx, defer, "[label=\".%c\"]",
+							"xyzw"[(i - 1) & 0x3]);
+				}
+			}
+		} else if (instr->opc == OPC_META_OUTPUT) {
+			printdef(ctx, defer, "output%lx:<out%u>:w -> %s",
+					PTRID(instr->inout.block),
+					instr->regs[0]->num, target);
+		} else if (instr->opc == OPC_META_PHI) {
+			/* treat like a normal instruction: */
+			printdef(ctx, defer, "instr%lx:<dst0> -> %s", PTRID(instr), target);
+		}
+	} else {
+		printdef(ctx, defer, "instr%lx:<dst0> -> %s", PTRID(instr), target);
+	}
+}
+
+static void dump_link(struct ir3_dump_ctx *ctx,
+		struct ir3_instruction *instr,
+		struct ir3_block *block, const char *target)
+{
+	bool defer = instr->block != block;
+	dump_link2(ctx, instr, target, defer);
+	printdef(ctx, defer, "\n");
+}
+
+static struct ir3_register *follow_flow(struct ir3_register *reg)
+{
+	if (reg->flags & IR3_REG_SSA) {
+		struct ir3_instruction *instr = reg->instr;
+		/* go with the flow.. */
+		if (is_meta(instr) && (instr->opc == OPC_META_FLOW))
+			return instr->regs[1];
+	}
+	return reg;
+}
+
+static void ir3_instr_dump(struct ir3_dump_ctx *ctx,
+		struct ir3_instruction *instr)
+{
+	unsigned i;
+
+	fprintf(ctx->f, "instr%lx [shape=record,style=filled,fillcolor=lightgrey,label=\"{",
+			PTRID(instr));
+	dump_instr_name(ctx, instr);
+
+	/* destination register: */
+	fprintf(ctx->f, "|<dst0>");
+
+	/* source register(s): */
+	for (i = 1; i < instr->regs_count; i++) {
+		struct ir3_register *reg = follow_flow(instr->regs[i]);
+
+		fprintf(ctx->f, "|");
+
+		if (reg->flags & IR3_REG_SSA)
+			fprintf(ctx->f, "<src%u> ", (i - 1));
+
+		dump_reg_name(ctx, reg);
+	}
+
+	fprintf(ctx->f, "}\"];\n");
+
+	/* and recursively dump dependent instructions: */
+	for (i = 1; i < instr->regs_count; i++) {
+		struct ir3_register *reg = instr->regs[i];
+		char target[32];  /* link target */
+
+		if (!(reg->flags & IR3_REG_SSA))
+			continue;
+
+		snprintf(target, sizeof(target), "instr%lx:<src%u>",
+				PTRID(instr), (i - 1));
+
+		dump_instr(ctx, reg->instr);
+		dump_link(ctx, follow_flow(reg)->instr, instr->block, target);
+	}
+}
+
+static void ir3_block_dump(struct ir3_dump_ctx *ctx,
+		struct ir3_block *block, const char *name)
+{
+	unsigned i, n;
+
+	n = edge_buf.n;
+
+	fprintf(ctx->f, "subgraph cluster%lx {\n", PTRID(block));
+	fprintf(ctx->f, "label=\"%s\";\n", name);
+
+	/* draw inputs: */
+	fprintf(ctx->f, "input%lx [shape=record,label=\"inputs", PTRID(block));
+	for (i = 0; i < block->ninputs; i++)
+		if (block->inputs[i])
+			fprintf(ctx->f, "|<in%u> i%u.%c", i, (i >> 2), "xyzw"[i & 0x3]);
+	fprintf(ctx->f, "\"];\n");
+
+	/* draw instruction graph: */
+	for (i = 0; i < block->noutputs; i++)
+		dump_instr(ctx, block->outputs[i]);
+
+	/* draw outputs: */
+	fprintf(ctx->f, "output%lx [shape=record,label=\"outputs", PTRID(block));
+	for (i = 0; i < block->noutputs; i++)
+		fprintf(ctx->f, "|<out%u> o%u.%c", i, (i >> 2), "xyzw"[i & 0x3]);
+	fprintf(ctx->f, "\"];\n");
+
+	/* and links to outputs: */
+	for (i = 0; i < block->noutputs; i++) {
+		char target[32];  /* link target */
+
+		/* NOTE: there could be outputs that are never assigned,
+		 * so skip them
+		 */
+		if (!block->outputs[i])
+			continue;
+
+		snprintf(target, sizeof(target), "output%lx:<out%u>:e",
+				PTRID(block), i);
+
+		dump_link(ctx, block->outputs[i], block, target);
+	}
+
+	fprintf(ctx->f, "}\n");
+
+	/* and links to inputs: */
+	if (block->parent) {
+		for (i = 0; i < block->ninputs; i++) {
+			char target[32];  /* link target */
+
+			if (!block->inputs[i])
+				continue;
+
+			dump_instr(ctx, block->inputs[i]);
+
+			snprintf(target, sizeof(target), "input%lx:<in%u>:e",
+					PTRID(block), i);
+
+			dump_link(ctx, block->inputs[i], block, target);
+		}
+	}
+
+	/* dump deferred edges: */
+	if (edge_buf.n > n) {
+		fprintf(ctx->f, "%*s", edge_buf.n - n, &edge_buf.buf[n]);
+		edge_buf.n = n;
+	}
+}
+
+void ir3_shader_dump(struct ir3_shader *shader, const char *name,
+		struct ir3_block *block /* XXX maybe 'block' ptr should move to ir3_shader? */,
+		FILE *f)
+{
+	struct ir3_dump_ctx ctx = {
+			.f = f,
+	};
+	ir3_shader_clear_mark(shader);
+	fprintf(ctx.f, "digraph G {\n");
+	fprintf(ctx.f, "rankdir=RL;\n");
+	fprintf(ctx.f, "nodesep=0.25;\n");
+	fprintf(ctx.f, "ranksep=1.5;\n");
+	ir3_block_dump(&ctx, block, name);
+	fprintf(ctx.f, "}\n");
+}
+
+/*
+ * For Debugging:
+ */
+
+void
+ir3_dump_instr_single(struct ir3_instruction *instr)
+{
+	struct ir3_dump_ctx ctx = {
+			.f = stdout,
+			.verbose = true,
+	};
+	unsigned i;
+
+	dump_instr_name(&ctx, instr);
+	for (i = 0; i < instr->regs_count; i++) {
+		struct ir3_register *reg = instr->regs[i];
+		printf(i ? ", " : " ");
+		dump_reg_name(&ctx, reg);
+	}
+	printf("\n");
+}
+
+void
+ir3_dump_instr_list(struct ir3_instruction *instr)
+{
+	unsigned n = 0;
+
+	while (instr) {
+		ir3_dump_instr_single(instr);
+		if (!is_meta(instr))
+			n++;
+		instr = instr->next;
+	}
+	printf("%u instructions\n", n);
+}
diff --git a/src/gallium/drivers/freedreno/a3xx/ir3_flatten.c b/src/gallium/drivers/freedreno/a3xx/ir3_flatten.c
new file mode 100644
index 0000000..3cf76db
--- /dev/null
+++ b/src/gallium/drivers/freedreno/a3xx/ir3_flatten.c
@@ -0,0 +1,140 @@
+/* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
+
+/*
+ * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
+ *
+ * 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 <stdarg.h>
+
+#include "ir3.h"
+
+/*
+ * Flatten: flatten out legs of if/else, etc
+ *
+ * TODO probably should use some heuristic to decide to not flatten
+ * if one side of the other is too large / deeply nested / whatever?
+ */
+
+struct ir3_flatten_ctx {
+	struct ir3_block *block;
+	unsigned cnt;
+};
+
+static struct ir3_register *unwrap(struct ir3_register *reg)
+{
+
+	if (reg->flags & IR3_REG_SSA) {
+		struct ir3_instruction *instr = reg->instr;
+		if (is_meta(instr)) {
+			switch (instr->opc) {
+			case OPC_META_OUTPUT:
+			case OPC_META_FLOW:
+				return instr->regs[1];
+			default:
+				break;
+			}
+		}
+	}
+	return reg;
+}
+
+static void ir3_instr_flatten(struct ir3_flatten_ctx *ctx,
+		struct ir3_instruction *instr)
+{
+	unsigned i;
+
+	/* if we've already visited this instruction, bail now: */
+	if (ir3_instr_check_mark(instr))
+		return;
+
+	instr->block = ctx->block;
+
+	/* TODO: maybe some threshold to decide whether to
+	 * flatten or not??
+	 */
+	if (is_meta(instr)) {
+		if (instr->opc == OPC_META_PHI) {
+			struct ir3_register *cond, *t, *f;
+
+			/* convert the PHI instruction to sel.{f16,f32} */
+			instr->category = 3;
+
+			/* instruction type based on dst size: */
+			if (instr->regs[0]->flags & IR3_REG_HALF)
+				instr->opc = OPC_SEL_F16;
+			else
+				instr->opc = OPC_SEL_F32;
+
+			/* swap around src register order, to match what
+			 * hw expects:
+			 */
+			cond = instr->regs[1];
+			t    = instr->regs[2];  /* true val */
+			f    = instr->regs[3];  /* false val */
+
+			instr->regs[1] = unwrap(f);
+			instr->regs[2] = unwrap(cond);
+			instr->regs[3] = unwrap(t);
+
+			ctx->cnt++;
+		} else if ((instr->opc == OPC_META_INPUT) &&
+				(instr->regs_count == 2)) {
+			type_t ftype;
+
+			if (instr->regs[0]->flags & IR3_REG_HALF)
+				ftype = TYPE_F16;
+			else
+				ftype = TYPE_F32;
+
+			/* convert meta:input to mov: */
+			instr->category = 1;
+			instr->cat1.src_type = ftype;
+			instr->cat1.dst_type = ftype;
+		}
+	}
+
+	/* recursively visit children: */
+	for (i = 1; i < instr->regs_count; i++) {
+		struct ir3_register *src = instr->regs[i];
+		if (src->flags & IR3_REG_SSA)
+			ir3_instr_flatten(ctx, src->instr);
+	}
+}
+
+/* return >= 0 is # of phi's flattened, < 0 is error */
+int ir3_block_flatten(struct ir3_block *block)
+{
+	struct ir3_flatten_ctx ctx = {
+			.block = block,
+	};
+	unsigned i;
+
+	ir3_shader_clear_mark(block->shader);
+	for(i = 0; i < block->noutputs; i++)
+		if (block->outputs[i])
+			ir3_instr_flatten(&ctx, block->outputs[i]);
+
+	return ctx.cnt;
+}
diff --git a/src/gallium/drivers/freedreno/a3xx/ir3_ra.c b/src/gallium/drivers/freedreno/a3xx/ir3_ra.c
new file mode 100644
index 0000000..6c868e2
--- /dev/null
+++ b/src/gallium/drivers/freedreno/a3xx/ir3_ra.c
@@ -0,0 +1,580 @@
+/* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
+
+/*
+ * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
+ *
+ * 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 "pipe/p_shader_tokens.h"
+#include "util/u_math.h"
+
+#include "ir3.h"
+#include "ir3_visitor.h"
+
+/*
+ * Register Assignment:
+ *
+ * NOTE: currently only works on a single basic block.. need to think
+ * about how multiple basic blocks are going to get scheduled.  But
+ * I think I want to re-arrange how blocks work, ie. get rid of the
+ * block nesting thing..
+ *
+ * NOTE: we could do register coalescing (eliminate moves) as part of
+ * the RA step.. OTOH I think we need to do scheduling before register
+ * assignment.  And if we remove a mov that effects scheduling (unless
+ * we leave a placeholder nop, which seems lame), so I'm not really
+ * sure how practical this is to do both in a single stage.  But OTOH
+ * I'm not really sure a sane way for the CP stage to realize when it
+ * cannot remove a mov due to multi-register constraints..
+ *
+ */
+
+struct ir3_ra_ctx {
+	struct ir3_block *block;
+	enum shader_t type;
+	int cnt;
+	bool error;
+};
+
+struct ir3_ra_assignment {
+	int8_t  off;        /* offset of instruction dst within range */
+	uint8_t num;        /* number of components for the range */
+};
+
+static void ra_assign(struct ir3_ra_ctx *ctx,
+		struct ir3_instruction *assigner, int num);
+static struct ir3_ra_assignment ra_calc(struct ir3_instruction *instr);
+
+/*
+ * Register Allocation:
+ */
+
+#define REG(n, wm) (struct ir3_register){ \
+		/*.flags  = ((so)->half_precision) ? IR3_REG_HALF : 0,*/ \
+		.num    = (n), \
+		.wrmask = TGSI_WRITEMASK_ ## wm, \
+	}
+
+/* check that the register exists, is a GPR and is not special (a0/p0) */
+static struct ir3_register * reg_check(struct ir3_instruction *instr, unsigned n)
+{
+	if ((n < instr->regs_count) && reg_gpr(instr->regs[n]))
+		return instr->regs[n];
+	return NULL;
+}
+
+static int output_base(struct ir3_ra_ctx *ctx)
+{
+	/* ugg, for fragment shader we need to have input at r0.x
+	 * (or at least if there is a way to configure it, I can't
+	 * see how because the blob driver always uses r0.x (ie.
+	 * all zeros)
+	 */
+	if (ctx->type == SHADER_FRAGMENT)
+		return 2;
+	return 0;
+}
+
+/* live means read before written */
+static void compute_liveregs(struct ir3_ra_ctx *ctx,
+		struct ir3_instruction *instr, regmask_t *liveregs)
+{
+	struct ir3_block *block = instr->block;
+	regmask_t written;
+	unsigned i, j;
+
+	regmask_init(liveregs);
+	regmask_init(&written);
+
+	for (instr = instr->next; instr; instr = instr->next) {
+		struct ir3_register *r;
+
+		if (is_meta(instr))
+			continue;
+
+		/* check first src's read: */
+		for (j = 1; j < instr->regs_count; j++) {
+			r = reg_check(instr, j);
+			if (r)
+				regmask_set_if_not(liveregs, r, &written);
+		}
+
+		/* then dst written (if assigned already): */
+		if (instr->flags & IR3_INSTR_MARK) {
+			r = reg_check(instr, 0);
+			if (r)
+				regmask_set(&written, r);
+		}
+	}
+
+	/* be sure to account for output registers too: */
+	for (i = 0; i < block->noutputs; i++) {
+		struct ir3_register reg = REG(output_base(ctx) + i, X);
+		regmask_set_if_not(liveregs, &reg, &written);
+	}
+}
+
+/* calculate registers that are clobbered before last use of 'assigner'.
+ * This needs to be done backwards, although it could possibly be
+ * combined into compute_liveregs().  (Ie. compute_liveregs() could
+ * reverse the list, then do this part backwards reversing the list
+ * again back to original order.)  Otoh, probably I should try to
+ * construct a proper interference graph instead.
+ *
+ * XXX this need to follow the same recursion path that is used for
+ * to rename/assign registers (ie. ra_assign_src()).. this is a bit
+ * ugly right now, maybe refactor into node iterator sort of things
+ * that iterates nodes in the correct order?
+ */
+static bool compute_clobbers(struct ir3_ra_ctx *ctx,
+		struct ir3_instruction *instr, struct ir3_instruction *assigner,
+		regmask_t *liveregs)
+{
+	unsigned i;
+	bool live = false, was_live = false;
+
+	if (instr == NULL) {
+		struct ir3_block *block = ctx->block;
+
+		/* if at the end, check outputs: */
+		for (i = 0; i < block->noutputs; i++)
+			if (block->outputs[i] == assigner)
+				return true;
+		return false;
+	}
+
+	for (i = 1; i < instr->regs_count; i++) {
+		struct ir3_register *reg = instr->regs[i];
+		if ((reg->flags & IR3_REG_SSA) && (reg->instr == assigner)) {
+			if (is_meta(instr)) {
+				switch (instr->opc) {
+				case OPC_META_INPUT:
+					// TODO
+					assert(0);
+					break;
+				case OPC_META_FO:
+				case OPC_META_FI:
+					was_live |= compute_clobbers(ctx, instr->next,
+							instr, liveregs);
+					break;
+				default:
+					break;
+				}
+			}
+			live = true;
+			break;
+		}
+	}
+
+	was_live |= compute_clobbers(ctx, instr->next, assigner, liveregs);
+
+	if (was_live && (instr->regs_count > 0) &&
+			(instr->flags & IR3_INSTR_MARK) &&
+			!is_meta(instr))
+		regmask_set(liveregs, instr->regs[0]);
+
+	return live || was_live;
+}
+
+static int find_available(regmask_t *liveregs, int size)
+{
+	unsigned i;
+	for (i = 0; i < MAX_REG - size; i++) {
+		if (!regmask_get(liveregs, &REG(i, X))) {
+			unsigned start = i++;
+			for (; (i < MAX_REG) && ((i - start) < size); i++)
+				if (regmask_get(liveregs, &REG(i, X)))
+					break;
+			if ((i - start) >= size)
+				return start;
+		}
+	}
+	assert(0);
+	return -1;
+}
+
+static int alloc_block(struct ir3_ra_ctx *ctx,
+		struct ir3_instruction *instr, int size)
+{
+	if (!instr) {
+		/* special case, allocating shader outputs.  At this
+		 * point, nothing is allocated, just start the shader
+		 * outputs at r0.x and let compute_liveregs() take
+		 * care of the rest from here:
+		 */
+		return 0;
+	} else {
+		regmask_t liveregs;
+		compute_liveregs(ctx, instr, &liveregs);
+
+		// XXX XXX XXX XXX XXX XXX XXX XXX XXX
+		// XXX hack.. maybe ra_calc should give us a list of
+		// instrs to compute_clobbers() on?
+		if (is_meta(instr) && (instr->opc == OPC_META_INPUT) &&
+				(instr->regs_count == 1)) {
+			unsigned i, base = instr->regs[0]->num & ~0x3;
+			for (i = 0; i < 4; i++) {
+				struct ir3_instruction *in = ctx->block->inputs[base + i];
+				if (in)
+					compute_clobbers(ctx, in->next, in, &liveregs);
+			}
+		} else
+		// XXX XXX XXX XXX XXX XXX XXX XXX XXX
+		compute_clobbers(ctx, instr->next, instr, &liveregs);
+		return find_available(&liveregs, size);
+	}
+}
+
+/*
+ * Constraint Calculation:
+ */
+
+struct ra_calc_visitor {
+	struct ir3_visitor base;
+	struct ir3_ra_assignment a;
+};
+
+static inline struct ra_calc_visitor *ra_calc_visitor(struct ir3_visitor *v)
+{
+	return (struct ra_calc_visitor *)v;
+}
+
+/* calculate register assignment for the instruction.  If the register
+ * written by this instruction is required to be part of a range, to
+ * handle other (input/output/sam/bary.f/etc) contiguous register range
+ * constraints, that is calculated handled here.
+ */
+static void ra_calc_dst(struct ir3_visitor *v,
+		struct ir3_instruction *instr, struct ir3_register *reg)
+{
+	struct ra_calc_visitor *c = ra_calc_visitor(v);
+	if (is_tex(instr)) {
+		c->a.off = 0;
+		c->a.num = 4;
+	} else {
+		c->a.off = 0;
+		c->a.num = 1;
+	}
+}
+
+static void
+ra_calc_dst_shader_input(struct ir3_visitor *v,
+		struct ir3_instruction *instr, struct ir3_register *reg)
+{
+	struct ra_calc_visitor *c = ra_calc_visitor(v);
+	struct ir3_block *block = instr->block;
+	struct ir3_register *dst = instr->regs[0];
+	unsigned base = dst->num & ~0x3;
+	unsigned i, num = 0;
+
+	assert(!(dst->flags & IR3_REG_IA));
+
+	/* check what input components we need: */
+	for (i = 0; i < 4; i++) {
+		unsigned idx = base + i;
+		if ((idx < block->ninputs) && block->inputs[idx])
+			num = i + 1;
+	}
+
+	c->a.off = dst->num - base;
+	c->a.num = num;
+}
+
+static void ra_calc_src_fanin(struct ir3_visitor *v,
+		struct ir3_instruction *instr, struct ir3_register *reg)
+{
+	struct ra_calc_visitor *c = ra_calc_visitor(v);
+	unsigned srcn = ir3_instr_regno(instr, reg) - 1;
+	c->a.off -= srcn;
+	c->a.num += srcn;
+	c->a.num = MAX2(c->a.num, instr->regs_count - 1);
+}
+
+static const struct ir3_visitor_funcs calc_visitor_funcs = {
+		.instr = ir3_visit_instr,
+		.dst_shader_input = ra_calc_dst_shader_input,
+		.dst_fanout = ra_calc_dst,
+		.dst_fanin = ra_calc_dst,
+		.dst = ra_calc_dst,
+		.src_fanout = ir3_visit_reg,
+		.src_fanin = ra_calc_src_fanin,
+		.src = ir3_visit_reg,
+};
+
+static struct ir3_ra_assignment ra_calc(struct ir3_instruction *assigner)
+{
+	struct ra_calc_visitor v = {
+			.base.funcs = &calc_visitor_funcs,
+	};
+
+	ir3_visit_instr(&v.base, assigner);
+
+	return v.a;
+}
+
+/*
+ * Register Assignment:
+ */
+
+struct ra_assign_visitor {
+	struct ir3_visitor base;
+	struct ir3_ra_ctx *ctx;
+	int num;
+};
+
+static inline struct ra_assign_visitor *ra_assign_visitor(struct ir3_visitor *v)
+{
+	return (struct ra_assign_visitor *)v;
+}
+
+static void ra_assign_reg(struct ir3_visitor *v,
+		struct ir3_instruction *instr, struct ir3_register *reg)
+{
+	struct ra_assign_visitor *a = ra_assign_visitor(v);
+	reg->flags &= ~IR3_REG_SSA;
+	reg->num = a->num;
+}
+
+static void ra_assign_dst_shader_input(struct ir3_visitor *v,
+		struct ir3_instruction *instr, struct ir3_register *reg)
+{
+	struct ra_assign_visitor *a = ra_assign_visitor(v);
+	unsigned i, base = reg->num & ~0x3;
+	int off = base - reg->num;
+
+	ra_assign_reg(v, instr, reg);
+	reg->flags |= IR3_REG_IA;
+
+	/* trigger assignment of all our companion input components: */
+	for (i = 0; i < 4; i++) {
+		struct ir3_instruction *in = instr->block->inputs[i+base];
+		if (in && is_meta(in) && (in->opc == OPC_META_INPUT))
+			ra_assign(a->ctx, in, a->num + off + i);
+	}
+}
+
+static void ra_assign_dst_fanout(struct ir3_visitor *v,
+		struct ir3_instruction *instr, struct ir3_register *reg)
+{
+	struct ra_assign_visitor *a = ra_assign_visitor(v);
+	struct ir3_register *src = instr->regs[1];
+	ra_assign_reg(v, instr, reg);
+	if (src->flags & IR3_REG_SSA)
+		ra_assign(a->ctx, src->instr, a->num - instr->fo.off);
+}
+
+static void ra_assign_src_fanout(struct ir3_visitor *v,
+		struct ir3_instruction *instr, struct ir3_register *reg)
+{
+	struct ra_assign_visitor *a = ra_assign_visitor(v);
+	ra_assign_reg(v, instr, reg);
+	ra_assign(a->ctx, instr, a->num + instr->fo.off);
+}
+
+
+static void ra_assign_src_fanin(struct ir3_visitor *v,
+		struct ir3_instruction *instr, struct ir3_register *reg)
+{
+	struct ra_assign_visitor *a = ra_assign_visitor(v);
+	unsigned j, srcn = ir3_instr_regno(instr, reg) - 1;
+	ra_assign_reg(v, instr, reg);
+	ra_assign(a->ctx, instr, a->num - srcn);
+	for (j = 1; j < instr->regs_count; j++) {
+		struct ir3_register *reg = instr->regs[j];
+		if (reg->flags & IR3_REG_SSA)  /* could be renamed already */
+			ra_assign(a->ctx, reg->instr, a->num - srcn + j - 1);
+	}
+}
+
+static const struct ir3_visitor_funcs assign_visitor_funcs = {
+		.instr = ir3_visit_instr,
+		.dst_shader_input = ra_assign_dst_shader_input,
+		.dst_fanout = ra_assign_dst_fanout,
+		.dst_fanin = ra_assign_reg,
+		.dst = ra_assign_reg,
+		.src_fanout = ra_assign_src_fanout,
+		.src_fanin = ra_assign_src_fanin,
+		.src = ra_assign_reg,
+};
+
+static void ra_assign(struct ir3_ra_ctx *ctx,
+		struct ir3_instruction *assigner, int num)
+{
+	struct ra_assign_visitor v = {
+			.base.funcs = &assign_visitor_funcs,
+			.ctx = ctx,
+			.num = num,
+	};
+
+	/* if we've already visited this instruction, bail now: */
+	if (ir3_instr_check_mark(assigner)) {
+		debug_assert(assigner->regs[0]->num == num);
+		if (assigner->regs[0]->num != num) {
+			/* impossible situation, should have been resolved
+			 * at an earlier stage by inserting extra mov's:
+			 */
+			ctx->error = true;
+		}
+		return;
+	}
+
+	ir3_visit_instr(&v.base, assigner);
+}
+
+/*
+ *
+ */
+
+static void ir3_instr_ra(struct ir3_ra_ctx *ctx,
+		struct ir3_instruction *instr)
+{
+	struct ir3_ra_assignment a;
+	unsigned num;
+
+	/* skip over nop's */
+	if (instr->regs_count == 0)
+		return;
+
+	/* if we've already visited this instruction, bail now: */
+	if (instr->flags & IR3_INSTR_MARK)
+		return;
+
+	/* allocate register(s): */
+	a = ra_calc(instr);
+	num = alloc_block(ctx, instr, a.num) + a.off;
+
+	ra_assign(ctx, instr, num);
+}
+
+/* flatten into shader: */
+// XXX this should probably be somewhere else:
+static void legalize(struct ir3_ra_ctx *ctx, struct ir3_block *block)
+{
+	struct ir3_instruction *n;
+	struct ir3_shader *shader = block->shader;
+	struct ir3_instruction *end =
+			ir3_instr_create(block, 0, OPC_END);
+	struct ir3_instruction *last_input = NULL;
+	regmask_t needs_ss;
+	regmask_t needs_sy;
+
+	regmask_init(&needs_ss);
+	regmask_init(&needs_sy);
+
+	shader->instrs_count = 0;
+
+	for (n = block->head; n; n = n->next) {
+		unsigned i;
+
+		if (is_meta(n))
+			continue;
+
+		for (i = 1; i < n->regs_count; i++) {
+			struct ir3_register *reg = n->regs[i];
+
+			if (is_gpr(reg)) {
+
+				/* TODO: we probably only need (ss) for alu
+				 * instr consuming sfu result.. need to make
+				 * some tests for both this and (sy)..
+				 */
+				if (regmask_get(&needs_ss, reg)) {
+					n->flags |= IR3_INSTR_SS;
+					regmask_init(&needs_ss);
+				}
+
+				if (regmask_get(&needs_sy, reg)) {
+					n->flags |= IR3_INSTR_SY;
+					regmask_init(&needs_sy);
+				}
+			}
+		}
+
+		shader->instrs[shader->instrs_count++] = n;
+
+		if (is_sfu(n))
+			regmask_set(&needs_ss, n->regs[0]);
+		if (is_tex(n))
+			regmask_set(&needs_sy, n->regs[0]);
+		if (is_input(n))
+			last_input = n;
+	}
+
+	if (last_input)
+		last_input->regs[0]->flags |= IR3_REG_EI;
+
+	shader->instrs[shader->instrs_count++] = end;
+
+	shader->instrs[0]->flags |= IR3_INSTR_SS | IR3_INSTR_SY;
+}
+
+static int block_ra(struct ir3_ra_ctx *ctx, struct ir3_block *block)
+{
+	struct ir3_instruction *n;
+
+	if (!block->parent) {
+		unsigned i;
+		int base, off = output_base(ctx);
+
+		base = alloc_block(ctx, NULL, block->noutputs + off);
+
+		for (i = 0; i < block->noutputs; i++)
+			if (block->outputs[i])
+				ra_assign(ctx, block->outputs[i], base + i + off);
+
+		if (ctx->type == SHADER_FRAGMENT) {
+			for (i = 0; i < block->ninputs; i++)
+				if (block->inputs[i])
+					ra_assign(ctx, block->inputs[i], base + i);
+		} else {
+			for (i = 0; i < block->ninputs; i++)
+				if (block->inputs[i])
+					ir3_instr_ra(ctx, block->inputs[i]);
+		}
+	}
+
+	/* then loop over instruction list and assign registers:
+	 */
+	n = block->head;
+	while (n) {
+		ir3_instr_ra(ctx, n);
+		if (ctx->error)
+			return -1;
+		n = n->next;
+	}
+
+	legalize(ctx, block);
+
+	return 0;
+}
+
+int ir3_block_ra(struct ir3_block *block, enum shader_t type)
+{
+	struct ir3_ra_ctx ctx = {
+			.block = block,
+			.type = type,
+	};
+	ir3_shader_clear_mark(block->shader);
+	return block_ra(&ctx, block);
+}
diff --git a/src/gallium/drivers/freedreno/a3xx/ir3_sched.c b/src/gallium/drivers/freedreno/a3xx/ir3_sched.c
new file mode 100644
index 0000000..5ac33ab
--- /dev/null
+++ b/src/gallium/drivers/freedreno/a3xx/ir3_sched.c
@@ -0,0 +1,289 @@
+/* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
+
+/*
+ * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
+ *
+ * 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 "util/u_math.h"
+
+#include "ir3.h"
+
+/*
+ * Instruction Scheduling:
+ *
+ * Using the depth sorted list from depth pass, attempt to recursively
+ * schedule deepest unscheduled path.  The first instruction that cannot
+ * be scheduled, returns the required delay slots it needs, at which
+ * point we return back up to the top and attempt to schedule by next
+ * highest depth.  After a sufficient number of instructions have been
+ * scheduled, return back to beginning of list and start again.  If you
+ * reach the end of depth sorted list without being able to insert any
+ * instruction, insert nop's.  Repeat until no more unscheduled
+ * instructions.
+ */
+
+struct ir3_sched_ctx {
+	struct ir3_instruction *scheduled;
+	unsigned cnt;
+};
+
+static struct ir3_instruction *
+deepest(struct ir3_instruction **srcs, unsigned nsrcs)
+{
+	struct ir3_instruction *d = NULL;
+	unsigned i = 0, id = 0;
+
+	while ((i < nsrcs) && !(d = srcs[id = i]))
+		i++;
+
+	if (!d)
+		return NULL;
+
+	for (; i < nsrcs; i++)
+		if (srcs[i] && (srcs[i]->depth > d->depth))
+			d = srcs[id = i];
+
+	srcs[id] = NULL;
+
+	return d;
+}
+
+static unsigned distance(struct ir3_sched_ctx *ctx,
+		struct ir3_instruction *instr, unsigned maxd)
+{
+	struct ir3_instruction *n = ctx->scheduled;
+	unsigned d = 0;
+	while (n && (n != instr) && (d < maxd)) {
+		if (!is_meta(n))
+			d++;
+		n = n->next;
+	}
+	return d;
+}
+
+/* TODO maybe we want double linked list? */
+static struct ir3_instruction * prev(struct ir3_instruction *instr)
+{
+	struct ir3_instruction *p = instr->block->head;
+	while (p && (p->next != instr))
+		p = p->next;
+	return p;
+}
+
+static void schedule(struct ir3_sched_ctx *ctx,
+		struct ir3_instruction *instr, bool remove)
+{
+	struct ir3_block *block = instr->block;
+
+	/* maybe there is a better way to handle this than just stuffing
+	 * a nop.. ideally we'd know about this constraint in the
+	 * scheduling and depth calculation..
+	 */
+	if (ctx->scheduled && is_sfu(ctx->scheduled) && is_sfu(instr))
+		schedule(ctx, ir3_instr_create(block, 0, OPC_NOP), false);
+
+	/* remove from depth list:
+	 */
+	if (remove) {
+		struct ir3_instruction *p = prev(instr);
+
+		/* NOTE: this can happen for inputs which are not
+		 * read.. in that case there is no need to schedule
+		 * the input, so just bail:
+		 */
+		if (instr != (p ? p->next : block->head))
+			return;
+
+		if (p)
+			p->next = instr->next;
+		else
+			block->head = instr->next;
+	}
+
+	instr->flags |= IR3_INSTR_MARK;
+
+	instr->next = ctx->scheduled;
+	ctx->scheduled = instr;
+
+	ctx->cnt++;
+}
+
+/*
+ * Delay-slot calculation.  Follows fanin/fanout.
+ */
+
+static unsigned delay_calc2(struct ir3_sched_ctx *ctx,
+		struct ir3_instruction *assigner,
+		struct ir3_instruction *consumer, unsigned srcn)
+{
+	unsigned delay = 0;
+
+	if (is_meta(assigner)) {
+		unsigned i;
+		for (i = 1; i < assigner->regs_count; i++) {
+			struct ir3_register *reg = assigner->regs[i];
+			if (reg->flags & IR3_REG_SSA) {
+				unsigned d = delay_calc2(ctx, reg->instr,
+						consumer, srcn);
+				delay = MAX2(delay, d);
+			}
+		}
+	} else {
+		delay = ir3_delayslots(assigner, consumer, srcn);
+		delay -= distance(ctx, assigner, delay);
+	}
+
+	return delay;
+}
+
+static unsigned delay_calc(struct ir3_sched_ctx *ctx,
+		struct ir3_instruction *instr)
+{
+	unsigned i, delay = 0;
+
+	for (i = 1; i < instr->regs_count; i++) {
+		struct ir3_register *reg = instr->regs[i];
+		if (reg->flags & IR3_REG_SSA) {
+			unsigned d = delay_calc2(ctx, reg->instr,
+					instr, i - 1);
+			delay = MAX2(delay, d);
+		}
+	}
+
+	return delay;
+}
+
+/* A negative return value signals that an instruction has been newly
+ * scheduled, return back up to the top of the stack (to block_sched())
+ */
+static int trysched(struct ir3_sched_ctx *ctx,
+		struct ir3_instruction *instr)
+{
+	struct ir3_instruction *srcs[ARRAY_SIZE(instr->regs) - 1];
+	struct ir3_instruction *src;
+	unsigned i, delay, nsrcs = 0;
+
+	/* if already scheduled: */
+	if (instr->flags & IR3_INSTR_MARK)
+		return 0;
+
+	/* figure out our src's: */
+	for (i = 1; i < instr->regs_count; i++) {
+		struct ir3_register *reg = instr->regs[i];
+		if (reg->flags & IR3_REG_SSA)
+			srcs[nsrcs++] = reg->instr;
+	}
+
+	/* for each src register in sorted order:
+	 */
+	delay = 0;
+	while ((src = deepest(srcs, nsrcs))) {
+		delay = trysched(ctx, src);
+		if (delay)
+			return delay;
+	}
+
+	/* all our dependents are scheduled, figure out if
+	 * we have enough delay slots to schedule ourself:
+	 */
+	delay = delay_calc(ctx, instr);
+
+	if (!delay) {
+		schedule(ctx, instr, true);
+		return -1;
+	}
+
+	return delay;
+}
+
+static struct ir3_instruction * reverse(struct ir3_instruction *instr)
+{
+	struct ir3_instruction *reversed = NULL;
+	while (instr) {
+		struct ir3_instruction *next = instr->next;
+		instr->next = reversed;
+		reversed = instr;
+		instr = next;
+	}
+	return reversed;
+}
+
+static void block_sched(struct ir3_sched_ctx *ctx, struct ir3_block *block)
+{
+	struct ir3_instruction *instr;
+
+	/* schedule all the shader input's (meta-instr) first so that
+	 * the RA step sees that the input registers contain a value
+	 * from the start of the shader:
+	 */
+	if (!block->parent) {
+		unsigned i;
+		for (i = 0; i < block->ninputs; i++) {
+			struct ir3_instruction *in = block->inputs[i];
+			if (in)
+				schedule(ctx, in, true);
+		}
+	}
+
+	while ((instr = block->head)) {
+		/* NOTE: always grab next *before* trysched(), in case the
+		 * instruction is actually scheduled (and therefore moved
+		 * from depth list into scheduled list)
+		 */
+		struct ir3_instruction *next = instr->next;
+		int cnt = trysched(ctx, instr);
+		/* -1 is signal to return up stack, but to us means same as 0: */
+		cnt = MAX2(0, cnt);
+		cnt += ctx->cnt;
+		instr = next;
+
+		/* if deepest remaining instruction cannot be scheduled, try
+		 * the increasingly more shallow instructions until needed
+		 * number of delay slots is filled:
+		 */
+		while (instr && (cnt > ctx->cnt)) {
+			next = instr->next;
+			trysched(ctx, instr);
+			instr = next;
+		}
+
+		/* and if we run out of instructions that can be scheduled,
+		 * then it is time for nop's:
+		 */
+		while (cnt > ctx->cnt)
+			schedule(ctx, ir3_instr_create(block, 0, OPC_NOP), false);
+	}
+
+	/* at this point, scheduled list is in reverse order, so fix that: */
+	block->head = reverse(ctx->scheduled);
+}
+
+void ir3_block_sched(struct ir3_block *block)
+{
+	struct ir3_sched_ctx ctx = {0};
+	ir3_shader_clear_mark(block->shader);
+	block_sched(&ctx, block);
+}
diff --git a/src/gallium/drivers/freedreno/a3xx/ir3_visitor.h b/src/gallium/drivers/freedreno/a3xx/ir3_visitor.h
new file mode 100644
index 0000000..1c60d16
--- /dev/null
+++ b/src/gallium/drivers/freedreno/a3xx/ir3_visitor.h
@@ -0,0 +1,154 @@
+/* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
+
+/*
+ * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
+ *
+ * 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>
+ */
+
+#ifndef IR3_VISITOR_H_
+#define IR3_VISITOR_H_
+
+/**
+ * Visitor which follows dst to src relationships between instructions,
+ * first visiting the dst (writer) instruction, followed by src (reader)
+ * instruction(s).
+ *
+ * TODO maybe we want multiple different visitors to walk the
+ * graph in different ways?
+ */
+
+struct ir3_visitor;
+
+typedef void (*ir3_visit_instr_func)(struct ir3_visitor *v,
+		struct ir3_instruction *instr);
+
+typedef void (*ir3_visit_reg_func)(struct ir3_visitor *v,
+		struct ir3_instruction *instr, struct ir3_register *reg);
+
+struct ir3_visitor_funcs {
+	ir3_visit_instr_func instr;  // TODO do we need??
+
+	ir3_visit_reg_func dst_shader_input;
+	ir3_visit_reg_func dst_block_input;
+	ir3_visit_reg_func dst_fanout;
+	ir3_visit_reg_func dst_fanin;
+	ir3_visit_reg_func dst;
+
+	ir3_visit_reg_func src_block_input;
+	ir3_visit_reg_func src_fanout;
+	ir3_visit_reg_func src_fanin;
+	ir3_visit_reg_func src;
+};
+
+struct ir3_visitor {
+	const struct ir3_visitor_funcs *funcs;
+	bool error;
+};
+
+#include "util/u_debug.h"
+
+static void visit_instr_dst(struct ir3_visitor *v,
+		struct ir3_instruction *instr)
+{
+	struct ir3_register *reg = instr->regs[0];
+
+	if (is_meta(instr)) {
+		switch (instr->opc) {
+		case OPC_META_INPUT:
+			if (instr->regs_count == 1)
+				v->funcs->dst_shader_input(v, instr, reg);
+			else
+				v->funcs->dst_block_input(v, instr, reg);
+			return;
+		case OPC_META_FO:
+			v->funcs->dst_fanout(v, instr, reg);
+			return;
+		case OPC_META_FI:
+			v->funcs->dst_fanin(v, instr, reg);
+			return;
+		default:
+			break;
+
+		}
+	}
+
+	v->funcs->dst(v, instr, reg);
+}
+
+static void visit_instr_src(struct ir3_visitor *v,
+		struct ir3_instruction *instr, struct ir3_register *reg)
+{
+	if (is_meta(instr)) {
+		switch (instr->opc) {
+		case OPC_META_INPUT:
+			/* shader-input does not have a src, only block input: */
+			debug_assert(instr->regs_count == 2);
+			v->funcs->src_block_input(v, instr, reg);
+			return;
+		case OPC_META_FO:
+			v->funcs->src_fanout(v, instr, reg);
+			return;
+		case OPC_META_FI:
+			v->funcs->src_fanin(v, instr, reg);
+			return;
+		default:
+			break;
+
+		}
+	}
+
+	v->funcs->src(v, instr, reg);
+}
+
+static void ir3_visit_instr(struct ir3_visitor *v,
+		struct ir3_instruction *instr)
+{
+	struct ir3_instruction *n;
+
+	/* visit instruction that assigns value: */
+	if (instr->regs_count > 0)
+		visit_instr_dst(v, instr);
+
+	/* and of any following instructions which read that value: */
+	n = instr->next;
+	while (n && !v->error) {
+		unsigned i;
+
+		for (i = 1; i < n->regs_count; i++) {
+			struct ir3_register *reg = n->regs[i];
+			if ((reg->flags & IR3_REG_SSA) && (reg->instr == instr))
+				visit_instr_src(v, n, reg);
+		}
+
+		n = n->next;
+	}
+}
+
+static void ir3_visit_reg(struct ir3_visitor *v,
+		struct ir3_instruction *instr, struct ir3_register *reg)
+{
+	/* no-op */
+}
+
+#endif /* IR3_VISITOR_H_ */
diff --git a/src/gallium/drivers/freedreno/freedreno_screen.c b/src/gallium/drivers/freedreno/freedreno_screen.c
index e32f970..f6affd6 100644
--- a/src/gallium/drivers/freedreno/freedreno_screen.c
+++ b/src/gallium/drivers/freedreno/freedreno_screen.c
@@ -68,6 +68,8 @@ static const struct debug_named_value debug_options[] = {
 		{"binning",   FD_DBG_BINNING,  "Enable hw binning"},
 		{"dbinning",  FD_DBG_DBINNING, "Disable hw binning"},
 		{"optimize",  FD_DBG_OPTIMIZE, "Enable optimization passes in compiler"},
+		{"optmsgs",   FD_DBG_OPTMSGS,  "Enable optimizater debug messages"},
+		{"optdump",   FD_DBG_OPTDUMP,  "Dump shader DAG to .dot files"},
 		DEBUG_NAMED_VALUE_END
 };
 
diff --git a/src/gallium/drivers/freedreno/freedreno_util.h b/src/gallium/drivers/freedreno/freedreno_util.h
index c126208..43026c5 100644
--- a/src/gallium/drivers/freedreno/freedreno_util.h
+++ b/src/gallium/drivers/freedreno/freedreno_util.h
@@ -64,6 +64,8 @@ enum adreno_stencil_op fd_stencil_op(unsigned op);
 #define FD_DBG_BINNING  0x0100
 #define FD_DBG_DBINNING 0x0200
 #define FD_DBG_OPTIMIZE 0x0400
+#define FD_DBG_OPTMSGS  0x0800
+#define FD_DBG_OPTDUMP  0x1000
 
 extern int fd_mesa_debug;
 extern bool fd_binning_enabled;