From 63432ecfce5415fbf07f1781ec77b5ea3efff599 Mon Sep 17 00:00:00 2001
From: Tom Stellard <tstellar@gmail.com>
Date: Tue, 13 Jul 2010 21:25:27 -0700
Subject: [PATCH] r300/compiler: Enable presubtract sources

The r300 compiler can now emit instructions that select from the presubtract
source.  A peephole optimization has been added to convert instructions like:
ADD Temp[0].x, none.1, -Temp[1].x into the INV (1 - src0) presubtract
operation.
---
 src/gallium/drivers/r300/r300_fs.c            |   1 +
 src/gallium/drivers/r300/r300_vs.c            |   1 +
 .../drivers/dri/r300/compiler/r300_fragprog.c |  76 +++++-
 .../dri/r300/compiler/r300_fragprog_emit.c    |  49 ++++
 .../dri/r300/compiler/r300_fragprog_swizzle.c |  33 ++-
 .../dri/r300/compiler/r500_fragprog_emit.c    |  37 +++
 .../dri/r300/compiler/radeon_compiler.h       |   1 +
 .../dri/r300/compiler/radeon_dataflow.c       | 128 ++++++---
 .../dri/r300/compiler/radeon_optimize.c       | 245 +++++++++++++++++-
 .../dri/r300/compiler/radeon_pair_schedule.c  | 158 ++++++++++-
 .../dri/r300/compiler/radeon_pair_translate.c |  53 ++++
 .../dri/r300/compiler/radeon_program.h        |  11 +-
 .../r300/compiler/radeon_program_constants.h  |  36 ++-
 .../dri/r300/compiler/radeon_program_pair.c   | 185 ++++++++++++-
 .../dri/r300/compiler/radeon_program_pair.h   |  33 ++-
 .../dri/r300/compiler/radeon_program_print.c  |  87 ++++++-
 .../r300/compiler/radeon_remove_constants.c   |  34 ++-
 17 files changed, 1074 insertions(+), 94 deletions(-)

diff --git a/src/gallium/drivers/r300/r300_fs.c b/src/gallium/drivers/r300/r300_fs.c
index b9f4d77dea7..b8dab88ef09 100644
--- a/src/gallium/drivers/r300/r300_fs.c
+++ b/src/gallium/drivers/r300/r300_fs.c
@@ -387,6 +387,7 @@ static void r300_translate_fragment_shader(
     compiler.Base.is_r500 = r300->screen->caps.is_r500;
     compiler.Base.disable_optimizations = DBG_ON(r300, DBG_NO_OPT);
     compiler.Base.has_half_swizzles = TRUE;
+    compiler.Base.has_presub = TRUE;
     compiler.Base.max_temp_regs = compiler.Base.is_r500 ? 128 : 32;
     compiler.Base.max_constants = compiler.Base.is_r500 ? 256 : 32;
     compiler.Base.max_alu_insts = compiler.Base.is_r500 ? 512 : 64;
diff --git a/src/gallium/drivers/r300/r300_vs.c b/src/gallium/drivers/r300/r300_vs.c
index 22c94adbda0..e2b9af9d018 100644
--- a/src/gallium/drivers/r300/r300_vs.c
+++ b/src/gallium/drivers/r300/r300_vs.c
@@ -208,6 +208,7 @@ void r300_translate_vertex_shader(struct r300_context *r300,
     compiler.Base.is_r500 = r300->screen->caps.is_r500;
     compiler.Base.disable_optimizations = DBG_ON(r300, DBG_NO_OPT);
     compiler.Base.has_half_swizzles = FALSE;
+    compiler.Base.has_presub = FALSE;
     compiler.Base.max_temp_regs = 32;
     compiler.Base.max_constants = 256;
     compiler.Base.max_alu_insts = r300->screen->caps.is_r500 ? 1024 : 256;
diff --git a/src/mesa/drivers/dri/r300/compiler/r300_fragprog.c b/src/mesa/drivers/dri/r300/compiler/r300_fragprog.c
index 3c83deffcb5..782671bac01 100644
--- a/src/mesa/drivers/dri/r300/compiler/r300_fragprog.c
+++ b/src/mesa/drivers/dri/r300/compiler/r300_fragprog.c
@@ -31,6 +31,24 @@
 
 #include "../r300_reg.h"
 
+static void presub_string(char out[10], unsigned int inst)
+{
+	switch(inst & 0x600000){
+	case R300_ALU_SRCP_1_MINUS_2_SRC0:
+		sprintf(out, "bias");
+		break;
+	case R300_ALU_SRCP_SRC1_MINUS_SRC0:
+		sprintf(out, "sub");
+		break;
+	case R300_ALU_SRCP_SRC1_PLUS_SRC0:
+		sprintf(out, "add");
+		break;
+	case R300_ALU_SRCP_1_MINUS_SRC0:
+		sprintf(out, "inv ");
+		break;
+	}
+}
+
 /* just some random things... */
 void r300FragmentProgramDump(struct radeon_compiler *c, void *user)
 {
@@ -98,8 +116,8 @@ void r300FragmentProgramDump(struct radeon_compiler *c, void *user)
 
 		for (i = alu_offset;
 		     i <= alu_offset + alu_end; ++i) {
-			char srcc[3][10], dstc[20];
-			char srca[3][10], dsta[20];
+			char srcc[4][10], dstc[20];
+			char srca[4][10], dsta[20];
 			char argc[3][20];
 			char arga[3][20];
 			char flags[5], tmp[10];
@@ -142,6 +160,9 @@ void r300FragmentProgramDump(struct radeon_compiler *c, void *user)
 					flags);
 				strcat(dstc, tmp);
 			}
+			/* Presub */
+			presub_string(srcc[3], code->alu.inst[i].rgb_inst);
+			presub_string(srca[3], code->alu.inst[i].alpha_inst);
 
 			dsta[0] = 0;
 			if (code->alu.inst[i].alpha_addr & R300_ALU_DSTA_REG) {
@@ -160,11 +181,12 @@ void r300FragmentProgramDump(struct radeon_compiler *c, void *user)
 			}
 
 			fprintf(stderr,
-				"%3i: xyz: %3s %3s %3s -> %-20s (%08x)\n"
-				"       w: %3s %3s %3s -> %-20s (%08x)\n", i,
-				srcc[0], srcc[1], srcc[2], dstc,
+				"%3i: xyz: %3s %3s %3s %5s-> %-20s (%08x)\n"
+				"       w: %3s %3s %3s %5s-> %-20s (%08x)\n", i,
+				srcc[0], srcc[1], srcc[2], srcc[3], dstc,
 				code->alu.inst[i].rgb_addr, srca[0], srca[1],
-				srca[2], dsta, code->alu.inst[i].alpha_addr);
+				srca[2], srca[3], dsta,
+				code->alu.inst[i].alpha_addr);
 
 			for (j = 0; j < 3; ++j) {
 				int regc = code->alu.inst[i].rgb_inst >> (j * 7);
@@ -194,6 +216,24 @@ void r300FragmentProgramDump(struct radeon_compiler *c, void *user)
 					}
 				} else if (d < 15) {
 					sprintf(buf, "%s.www", srca[d - 12]);
+				} else if (d < 20 ) {
+					switch(d) {
+					case R300_ALU_ARGC_SRCP_XYZ:
+						sprintf(buf, "srcp.xyz");
+						break;
+					case R300_ALU_ARGC_SRCP_XXX:
+						sprintf(buf, "srcp.xxx");
+						break;
+					case R300_ALU_ARGC_SRCP_YYY:
+						sprintf(buf, "srcp.yyy");
+						break;
+					case R300_ALU_ARGC_SRCP_ZZZ:
+						sprintf(buf, "srcp.zzz");
+						break;
+					case R300_ALU_ARGC_SRCP_WWW:
+						sprintf(buf, "srcp.www");
+						break;
+					}
 				} else if (d == 20) {
 					sprintf(buf, "0.0");
 				} else if (d == 21) {
@@ -231,6 +271,21 @@ void r300FragmentProgramDump(struct radeon_compiler *c, void *user)
 						'x' + (char)(d % 3));
 				} else if (d < 12) {
 					sprintf(buf, "%s.w", srca[d - 9]);
+				} else if (d < 16) {
+					switch(d) {
+					case R300_ALU_ARGA_SRCP_X:
+						sprintf(buf, "srcp.x");
+						break;
+					case R300_ALU_ARGA_SRCP_Y:
+						sprintf(buf, "srcp.y");
+						break;
+					case R300_ALU_ARGA_SRCP_Z:
+						sprintf(buf, "srcp.z");
+						break;
+					case R300_ALU_ARGA_SRCP_W:
+						sprintf(buf, "srcp.w");
+						break;
+					}
 				} else if (d == 16) {
 					sprintf(buf, "0.0");
 				} else if (d == 17) {
@@ -247,11 +302,14 @@ void r300FragmentProgramDump(struct radeon_compiler *c, void *user)
 					buf, (rega & 64) ? "|" : "");
 			}
 
-			fprintf(stderr, "     xyz: %8s %8s %8s    op: %08x\n"
+			fprintf(stderr, "     xyz: %8s %8s %8s    op: %08x %s\n"
 				"       w: %8s %8s %8s    op: %08x\n",
 				argc[0], argc[1], argc[2],
-				code->alu.inst[i].rgb_inst, arga[0], arga[1],
-				arga[2], code->alu.inst[i].alpha_inst);
+				code->alu.inst[i].rgb_inst,
+				code->alu.inst[i].rgb_inst & R300_ALU_INSERT_NOP ?
+				"NOP" : "",
+				arga[0], arga[1],arga[2],
+				code->alu.inst[i].alpha_inst);
 		}
 	}
 }
diff --git a/src/mesa/drivers/dri/r300/compiler/r300_fragprog_emit.c b/src/mesa/drivers/dri/r300/compiler/r300_fragprog_emit.c
index eead2ea4260..3b2b06fc2bd 100644
--- a/src/mesa/drivers/dri/r300/compiler/r300_fragprog_emit.c
+++ b/src/mesa/drivers/dri/r300/compiler/r300_fragprog_emit.c
@@ -164,6 +164,53 @@ static int emit_alu(struct r300_emit_state * emit, struct rc_pair_instruction* i
 		code->alu.inst[ip].alpha_inst |= arg << (7*j);
 	}
 
+	/* Presubtract */
+	if (inst->RGB.Src[RC_PAIR_PRESUB_SRC].Used) {
+		switch(inst->RGB.Src[RC_PAIR_PRESUB_SRC].Index) {
+		case RC_PRESUB_BIAS:
+			code->alu.inst[ip].rgb_inst |=
+						R300_ALU_SRCP_1_MINUS_2_SRC0;
+			break;
+		case RC_PRESUB_ADD:
+			code->alu.inst[ip].rgb_inst |=
+						R300_ALU_SRCP_SRC1_PLUS_SRC0;
+			break;
+		case RC_PRESUB_SUB:
+			code->alu.inst[ip].rgb_inst |=
+						R300_ALU_SRCP_SRC1_MINUS_SRC0;
+			break;
+		case RC_PRESUB_INV:
+			code->alu.inst[ip].rgb_inst |=
+						R300_ALU_SRCP_1_MINUS_SRC0;
+			break;
+		default:
+			break;
+		}
+	}
+
+	if (inst->Alpha.Src[RC_PAIR_PRESUB_SRC].Used) {
+		switch(inst->Alpha.Src[RC_PAIR_PRESUB_SRC].Index) {
+		case RC_PRESUB_BIAS:
+			code->alu.inst[ip].alpha_inst |=
+						R300_ALU_SRCP_1_MINUS_2_SRC0;
+			break;
+		case RC_PRESUB_ADD:
+			code->alu.inst[ip].alpha_inst |=
+						R300_ALU_SRCP_SRC1_PLUS_SRC0;
+			break;
+		case RC_PRESUB_SUB:
+			code->alu.inst[ip].alpha_inst |=
+						R300_ALU_SRCP_SRC1_MINUS_SRC0;
+			break;
+		case RC_PRESUB_INV:
+			code->alu.inst[ip].alpha_inst |=
+						R300_ALU_SRCP_1_MINUS_SRC0;
+			break;
+		default:
+			break;
+		}
+	}
+
 	if (inst->RGB.Saturate)
 		code->alu.inst[ip].rgb_inst |= R300_ALU_OUTC_CLAMP;
 	if (inst->Alpha.Saturate)
@@ -198,6 +245,8 @@ static int emit_alu(struct r300_emit_state * emit, struct rc_pair_instruction* i
 		emit->node_flags |= R300_W_OUT;
 		c->code->writes_depth = 1;
 	}
+	if (inst->Nop)
+		code->alu.inst[ip].rgb_inst |= R300_ALU_INSERT_NOP;
 
 	return 1;
 }
diff --git a/src/mesa/drivers/dri/r300/compiler/r300_fragprog_swizzle.c b/src/mesa/drivers/dri/r300/compiler/r300_fragprog_swizzle.c
index 5d5de2f1b2a..caa48fe478c 100644
--- a/src/mesa/drivers/dri/r300/compiler/r300_fragprog_swizzle.c
+++ b/src/mesa/drivers/dri/r300/compiler/r300_fragprog_swizzle.c
@@ -44,25 +44,25 @@ struct swizzle_data {
 	unsigned int hash; /**< swizzle value this matches */
 	unsigned int base; /**< base value for hw swizzle */
 	unsigned int stride; /**< difference in base between arg0/1/2 */
+	unsigned int srcp_stride; /**< difference in base between arg0/scrp */
 };
 
 static const struct swizzle_data native_swizzles[] = {
-	{MAKE_SWZ3(X, Y, Z), R300_ALU_ARGC_SRC0C_XYZ, 4},
-	{MAKE_SWZ3(X, X, X), R300_ALU_ARGC_SRC0C_XXX, 4},
-	{MAKE_SWZ3(Y, Y, Y), R300_ALU_ARGC_SRC0C_YYY, 4},
-	{MAKE_SWZ3(Z, Z, Z), R300_ALU_ARGC_SRC0C_ZZZ, 4},
-	{MAKE_SWZ3(W, W, W), R300_ALU_ARGC_SRC0A, 1},
-	{MAKE_SWZ3(Y, Z, X), R300_ALU_ARGC_SRC0C_YZX, 1},
-	{MAKE_SWZ3(Z, X, Y), R300_ALU_ARGC_SRC0C_ZXY, 1},
-	{MAKE_SWZ3(W, Z, Y), R300_ALU_ARGC_SRC0CA_WZY, 1},
-	{MAKE_SWZ3(ONE, ONE, ONE), R300_ALU_ARGC_ONE, 0},
-	{MAKE_SWZ3(ZERO, ZERO, ZERO), R300_ALU_ARGC_ZERO, 0},
-	{MAKE_SWZ3(HALF, HALF, HALF), R300_ALU_ARGC_HALF, 0}
+	{MAKE_SWZ3(X, Y, Z), R300_ALU_ARGC_SRC0C_XYZ, 4, 15},
+	{MAKE_SWZ3(X, X, X), R300_ALU_ARGC_SRC0C_XXX, 4, 15},
+	{MAKE_SWZ3(Y, Y, Y), R300_ALU_ARGC_SRC0C_YYY, 4, 15},
+	{MAKE_SWZ3(Z, Z, Z), R300_ALU_ARGC_SRC0C_ZZZ, 4, 15},
+	{MAKE_SWZ3(W, W, W), R300_ALU_ARGC_SRC0A, 1, 7},
+	{MAKE_SWZ3(Y, Z, X), R300_ALU_ARGC_SRC0C_YZX, 1, 0},
+	{MAKE_SWZ3(Z, X, Y), R300_ALU_ARGC_SRC0C_ZXY, 1, 0},
+	{MAKE_SWZ3(W, Z, Y), R300_ALU_ARGC_SRC0CA_WZY, 1, 0},
+	{MAKE_SWZ3(ONE, ONE, ONE), R300_ALU_ARGC_ONE, 0, 0},
+	{MAKE_SWZ3(ZERO, ZERO, ZERO), R300_ALU_ARGC_ZERO, 0, 0},
+	{MAKE_SWZ3(HALF, HALF, HALF), R300_ALU_ARGC_HALF, 0, 0}
 };
 
 static const int num_native_swizzles = sizeof(native_swizzles)/sizeof(native_swizzles[0]);
 
-
 /**
  * Find a native RGB swizzle that matches the given swizzle.
  * Returns 0 if none found.
@@ -205,7 +205,11 @@ unsigned int r300FPTranslateRGBSwizzle(unsigned int src, unsigned int swizzle)
 		return 0;
 	}
 
-	return sd->base + src*sd->stride;
+	if (src == RC_PAIR_PRESUB_SRC) {
+		return sd->base + sd->srcp_stride;
+	} else {
+		return sd->base + src*sd->stride;
+	}
 }
 
 
@@ -215,6 +219,9 @@ unsigned int r300FPTranslateRGBSwizzle(unsigned int src, unsigned int swizzle)
  */
 unsigned int r300FPTranslateAlphaSwizzle(unsigned int src, unsigned int swizzle)
 {
+	if (src == RC_PAIR_PRESUB_SRC) {
+		return R300_ALU_ARGA_SRCP_X + swizzle;
+	}
 	if (swizzle < 3)
 		return swizzle + 3*src;
 
diff --git a/src/mesa/drivers/dri/r300/compiler/r500_fragprog_emit.c b/src/mesa/drivers/dri/r300/compiler/r500_fragprog_emit.c
index 3220349f26c..54cff9169a4 100644
--- a/src/mesa/drivers/dri/r300/compiler/r500_fragprog_emit.c
+++ b/src/mesa/drivers/dri/r300/compiler/r500_fragprog_emit.c
@@ -260,6 +260,9 @@ static void emit_paired(struct r300_fragment_program_compiler *c, struct rc_pair
 
 	code->inst[ip].inst0 |= (inst->RGB.WriteMask << 11) | (inst->Alpha.WriteMask << 14);
 	code->inst[ip].inst0 |= (inst->RGB.OutputWriteMask << 15) | (inst->Alpha.OutputWriteMask << 18);
+	if (inst->Nop) {
+		code->inst[ip].inst0 |= R500_INST_NOP;
+	}
 	if (inst->Alpha.DepthWriteMask) {
 		code->inst[ip].inst4 |= R500_ALPHA_W_OMASK;
 		c->code->writes_depth = 1;
@@ -275,6 +278,40 @@ static void emit_paired(struct r300_fragment_program_compiler *c, struct rc_pair
 	if (inst->Alpha.Saturate)
 		code->inst[ip].inst0 |= R500_INST_ALPHA_CLAMP;
 
+	/* Set the presubtract operation. */
+	switch(inst->RGB.Src[RC_PAIR_PRESUB_SRC].Index) {
+		case RC_PRESUB_BIAS:
+			code->inst[ip].inst1 |= R500_RGB_SRCP_OP_1_MINUS_2RGB0;
+			break;
+		case RC_PRESUB_SUB:
+			code->inst[ip].inst1 |= R500_RGB_SRCP_OP_RGB1_MINUS_RGB0;
+			break;
+		case RC_PRESUB_ADD:
+			code->inst[ip].inst1 |= R500_RGB_SRCP_OP_RGB1_PLUS_RGB0;
+			break;
+		case RC_PRESUB_INV:
+			code->inst[ip].inst1 |= R500_RGB_SRCP_OP_1_MINUS_RGB0;
+			break;
+		default:
+			break;
+	}
+	switch(inst->Alpha.Src[RC_PAIR_PRESUB_SRC].Index) {
+		case RC_PRESUB_BIAS:
+			code->inst[ip].inst2 |= R500_ALPHA_SRCP_OP_1_MINUS_2A0;
+			break;
+		case RC_PRESUB_SUB:
+			code->inst[ip].inst2 |= R500_ALPHA_SRCP_OP_A1_MINUS_A0;
+			break;
+		case RC_PRESUB_ADD:
+			code->inst[ip].inst2 |= R500_ALPHA_SRCP_OP_A1_PLUS_A0;
+			break;
+		case RC_PRESUB_INV:
+			code->inst[ip].inst2 |= R500_ALPHA_SRCP_OP_1_MINUS_A0;
+			break;
+		default:
+			break;
+	}
+
 	code->inst[ip].inst1 |= R500_RGB_ADDR0(use_source(code, inst->RGB.Src[0]));
 	code->inst[ip].inst1 |= R500_RGB_ADDR1(use_source(code, inst->RGB.Src[1]));
 	code->inst[ip].inst1 |= R500_RGB_ADDR2(use_source(code, inst->RGB.Src[2]));
diff --git a/src/mesa/drivers/dri/r300/compiler/radeon_compiler.h b/src/mesa/drivers/dri/r300/compiler/radeon_compiler.h
index 85c2e43ad67..6d96ac9fdd9 100644
--- a/src/mesa/drivers/dri/r300/compiler/radeon_compiler.h
+++ b/src/mesa/drivers/dri/r300/compiler/radeon_compiler.h
@@ -42,6 +42,7 @@ struct radeon_compiler {
 	/* Hardware specification. */
 	unsigned is_r500:1;
 	unsigned has_half_swizzles:1;
+	unsigned has_presub:1;
 	unsigned disable_optimizations:1;
 	unsigned max_temp_regs;
 	unsigned max_constants;
diff --git a/src/mesa/drivers/dri/r300/compiler/radeon_dataflow.c b/src/mesa/drivers/dri/r300/compiler/radeon_dataflow.c
index 0e6c62541fa..e73700f84a2 100644
--- a/src/mesa/drivers/dri/r300/compiler/radeon_dataflow.c
+++ b/src/mesa/drivers/dri/r300/compiler/radeon_dataflow.c
@@ -29,6 +29,25 @@
 
 #include "radeon_program.h"
 
+static void reads_normal_callback(
+	rc_read_write_chan_fn cb,
+	struct rc_instruction * fullinst,
+	struct rc_src_register src,
+	void * userdata)
+{
+	unsigned int refmask = 0;
+	unsigned int chan;
+	for(chan = 0; chan < 4; chan++) {
+		refmask |= 1 << GET_SWZ(src.Swizzle, chan);
+	}
+	refmask &= RC_MASK_XYZW;
+
+	if (refmask)
+		cb(userdata, fullinst, src.File, src.Index, refmask);
+
+	if (refmask && src.RelAddr)
+		cb(userdata, fullinst, RC_FILE_ADDRESS, 0, RC_MASK_X);
+}
 
 static void reads_normal(struct rc_instruction * fullinst, rc_read_write_chan_fn cb, void * userdata)
 {
@@ -36,47 +55,81 @@ static void reads_normal(struct rc_instruction * fullinst, rc_read_write_chan_fn
 	const struct rc_opcode_info * opcode = rc_get_opcode_info(inst->Opcode);
 
 	for(unsigned int src = 0; src < opcode->NumSrcRegs; ++src) {
-		unsigned int refmask = 0;
 
 		if (inst->SrcReg[src].File == RC_FILE_NONE)
 			return;
 
-		for(unsigned int chan = 0; chan < 4; ++chan)
-			refmask |= 1 << GET_SWZ(inst->SrcReg[src].Swizzle, chan);
-
-		refmask &= RC_MASK_XYZW;
-
-		if (refmask)
-			cb(userdata, fullinst, inst->SrcReg[src].File, inst->SrcReg[src].Index, refmask);
-
-		if (refmask && inst->SrcReg[src].RelAddr)
-			cb(userdata, fullinst, RC_FILE_ADDRESS, 0, RC_MASK_X);
+		if (inst->SrcReg[src].File == RC_FILE_PRESUB) {
+			unsigned int i;
+			unsigned int srcp_regs = rc_presubtract_src_reg_count(
+							inst->PreSub.Opcode);
+			for( i = 0; i < srcp_regs; i++) {
+				reads_normal_callback(cb, fullinst,
+						inst->PreSub.SrcReg[i],
+						userdata);
+			}
+		} else {
+			reads_normal_callback(cb, fullinst,
+						inst->SrcReg[src], userdata);
+		}
 	}
 }
 
-static void reads_pair(struct rc_instruction * fullinst,  rc_read_write_mask_fn cb, void * userdata)
+static void pair_get_src_refmasks(unsigned int * refmasks,
+					struct rc_pair_instruction * inst,
+					unsigned int swz, unsigned int src)
 {
-	struct rc_pair_instruction * inst = &fullinst->U.P;
-	unsigned int refmasks[3] = { 0, 0, 0 };
-
-	if (inst->RGB.Opcode != RC_OPCODE_NOP) {
-		const struct rc_opcode_info * opcode = rc_get_opcode_info(inst->RGB.Opcode);
+	if (swz >= 4)
+		return;
+
+	if (swz == RC_SWIZZLE_X || swz == RC_SWIZZLE_Y || swz == RC_SWIZZLE_Z) {
+		if(src == RC_PAIR_PRESUB_SRC) {
+			unsigned int i;
+			int srcp_regs =
+				rc_presubtract_src_reg_count(
+				inst->RGB.Src[src].Index);
+			for(i = 0; i < srcp_regs; i++) {
+				refmasks[i] |= 1 << swz;
+			}
+		}
+		else {
+			refmasks[src] |= 1 << swz;
+		}
+	}
 
-		for(unsigned int arg = 0; arg < opcode->NumSrcRegs; ++arg) {
-			for(unsigned int chan = 0; chan < 3; ++chan) {
-				unsigned int swz = GET_SWZ(inst->RGB.Arg[arg].Swizzle, chan);
-				if (swz < 4)
-					refmasks[inst->RGB.Arg[arg].Source] |= 1 << swz;
+	if (swz == RC_SWIZZLE_W) {
+		if (src == RC_PAIR_PRESUB_SRC) {
+			unsigned int i;
+			int srcp_regs = rc_presubtract_src_reg_count(
+					inst->Alpha.Src[src].Index);
+			for(i = 0; i < srcp_regs; i++) {
+				refmasks[i] |= 1 << swz;
 			}
 		}
+		else {
+			refmasks[src] |= 1 << swz;
+		}
 	}
+}
 
-	if (inst->Alpha.Opcode != RC_OPCODE_NOP) {
-		const struct rc_opcode_info * opcode = rc_get_opcode_info(inst->Alpha.Opcode);
+static void reads_pair(struct rc_instruction * fullinst,  rc_read_write_mask_fn cb, void * userdata)
+{
+	struct rc_pair_instruction * inst = &fullinst->U.P;
+	unsigned int refmasks[3] = { 0, 0, 0 };
 
-		for(unsigned int arg = 0; arg < opcode->NumSrcRegs; ++arg) {
-			if (inst->Alpha.Arg[arg].Swizzle < 4)
-				refmasks[inst->Alpha.Arg[arg].Source] |= 1 << inst->Alpha.Arg[arg].Swizzle;
+	unsigned int arg;
+
+	for(arg = 0; arg < 3; ++arg) {
+		unsigned int chan;
+		for(chan = 0; chan < 3; ++chan) {
+			unsigned int swz_rgb =
+				GET_SWZ(inst->RGB.Arg[arg].Swizzle, chan);
+			unsigned int swz_alpha =
+				GET_SWZ(inst->Alpha.Arg[arg].Swizzle, chan);
+			pair_get_src_refmasks(refmasks, inst, swz_rgb,
+						inst->RGB.Arg[arg].Source);
+			pair_get_src_refmasks(refmasks, inst, swz_alpha,
+						inst->Alpha.Arg[arg].Source);
 		}
 	}
 
@@ -212,10 +265,25 @@ static void remap_normal_instruction(struct rc_instruction * fullinst,
 		rc_register_file file = inst->SrcReg[src].File;
 		unsigned int index = inst->SrcReg[src].Index;
 
-		cb(userdata, fullinst, &file, &index);
+		if (file == RC_FILE_PRESUB) {
+			unsigned int i;
+			unsigned int srcp_srcs = rc_presubtract_src_reg_count(
+						inst->PreSub.Opcode);
+			for(i = 0; i < srcp_srcs; i++) {
+				file = inst->PreSub.SrcReg[i].File;
+				index = inst->PreSub.SrcReg[i].Index;
+				cb(userdata, fullinst, &file, &index);
+				inst->PreSub.SrcReg[i].File = file;
+				inst->PreSub.SrcReg[i].Index = index;
+			}
 
-		inst->SrcReg[src].File = file;
-		inst->SrcReg[src].Index = index;
+		}
+		else {
+			cb(userdata, fullinst, &file, &index);
+
+			inst->SrcReg[src].File = file;
+			inst->SrcReg[src].Index = index;
+		}
 	}
 }
 
diff --git a/src/mesa/drivers/dri/r300/compiler/radeon_optimize.c b/src/mesa/drivers/dri/r300/compiler/radeon_optimize.c
index 8e994671064..e01ba85aa56 100644
--- a/src/mesa/drivers/dri/r300/compiler/radeon_optimize.c
+++ b/src/mesa/drivers/dri/r300/compiler/radeon_optimize.c
@@ -30,6 +30,13 @@
 #include "radeon_compiler.h"
 #include "radeon_swizzle.h"
 
+struct peephole_state {
+	struct rc_instruction * Inst;
+	/** Stores a bitmask of the components that are still "alive" (i.e.
+	 * they have not been written to since Inst was executed.)
+	 */
+	unsigned int WriteMask;
+};
 
 static struct rc_src_register chain_srcregs(struct rc_src_register outer, struct rc_src_register inner)
 {
@@ -54,7 +61,7 @@ static struct rc_src_register chain_srcregs(struct rc_src_register outer, struct
 	return combine;
 }
 
-struct peephole_state {
+struct copy_propagate_state {
 	struct radeon_compiler * C;
 	struct rc_instruction * Mov;
 	unsigned int Conflict:1;
@@ -84,10 +91,10 @@ struct peephole_state {
  * @param index The index of the source register.
  * @param mask The components of the source register that are being read from.
  */
-static void peephole_scan_read(void * data, struct rc_instruction * inst,
+static void copy_propagate_scan_read(void * data, struct rc_instruction * inst,
 		rc_register_file file, unsigned int index, unsigned int mask)
 {
-	struct peephole_state * s = data;
+	struct copy_propagate_state * s = data;
 
 	/* XXX This could probably be handled better. */
 	if (file == RC_FILE_ADDRESS) {
@@ -123,10 +130,10 @@ static void peephole_scan_read(void * data, struct rc_instruction * inst,
 	}
 }
 
-static void peephole_scan_write(void * data, struct rc_instruction * inst,
+static void copy_propagate_scan_write(void * data, struct rc_instruction * inst,
 		rc_register_file file, unsigned int index, unsigned int mask)
 {
-	struct peephole_state * s = data;
+	struct copy_propagate_state * s = data;
 
 	if (s->BranchDepth < 0)
 		return;
@@ -146,9 +153,9 @@ static void peephole_scan_write(void * data, struct rc_instruction * inst,
 	}
 }
 
-static void peephole(struct radeon_compiler * c, struct rc_instruction * inst_mov)
+static void copy_propagate(struct radeon_compiler * c, struct rc_instruction * inst_mov)
 {
-	struct peephole_state s;
+	struct copy_propagate_state s;
 
 	if (inst_mov->U.I.DstReg.File != RC_FILE_TEMPORARY ||
 	    inst_mov->U.I.DstReg.RelAddr ||
@@ -170,14 +177,23 @@ static void peephole(struct radeon_compiler * c, struct rc_instruction * inst_mo
 	for(struct rc_instruction * inst = inst_mov->Next;
 	    inst != &c->Program.Instructions;
 	    inst = inst->Next) {
+		const struct rc_opcode_info * info = rc_get_opcode_info(inst->U.I.Opcode);
 		/* XXX In the future we might be able to make the optimizer
 		 * smart enough to handle loops. */
 		if(inst->U.I.Opcode == RC_OPCODE_BGNLOOP
 				|| inst->U.I.Opcode == RC_OPCODE_ENDLOOP){
 			return;
 		}
-		rc_for_all_reads_mask(inst, peephole_scan_read, &s);
-		rc_for_all_writes_mask(inst, peephole_scan_write, &s);
+
+		/* It is possible to do copy propigation in this situation,
+		 * just not right now, see peephole_add_presub_inv() */
+		if (inst->U.I.PreSub.Opcode != RC_PRESUB_NONE &&
+						info->NumSrcRegs > 2) {
+			return;
+		}
+
+		rc_for_all_reads_mask(inst, copy_propagate_scan_read, &s);
+		rc_for_all_writes_mask(inst, copy_propagate_scan_write, &s);
 		if (s.Conflict)
 			return;
 
@@ -206,7 +222,6 @@ static void peephole(struct radeon_compiler * c, struct rc_instruction * inst_mo
 	    inst != &c->Program.Instructions;
 	    inst = inst->Next) {
 		const struct rc_opcode_info * opcode = rc_get_opcode_info(inst->U.I.Opcode);
-
 		for(unsigned int src = 0; src < opcode->NumSrcRegs; ++src) {
 			if (inst->U.I.SrcReg[src].File == RC_FILE_TEMPORARY &&
 			    inst->U.I.SrcReg[src].Index == s.Mov->U.I.DstReg.Index) {
@@ -217,8 +232,11 @@ static void peephole(struct radeon_compiler * c, struct rc_instruction * inst_mo
 					refmask |= (1 << swz) & RC_MASK_XYZW;
 				}
 
-				if ((refmask & s.MovMask) == refmask)
+				if ((refmask & s.MovMask) == refmask) {
 					inst->U.I.SrcReg[src] = chain_srcregs(inst->U.I.SrcReg[src], s.Mov->U.I.SrcReg[0]);
+					if (s.Mov->U.I.SrcReg[0].File == RC_FILE_PRESUB)
+						inst->U.I.PreSub = s.Mov->U.I.PreSub;
+				}
 			}
 		}
 
@@ -283,7 +301,6 @@ static int is_src_uniform_constant(struct rc_src_register src,
 	return 1;
 }
 
-
 static void constant_folding_mad(struct rc_instruction * inst)
 {
 	rc_swizzle swz;
@@ -379,7 +396,6 @@ static void constant_folding_add(struct rc_instruction * inst)
 	}
 }
 
-
 /**
  * Replace 0.0, 1.0 and 0.5 immediate constants by their
  * respective swizzles. Simplify instructions like ADD dst, src, 0;
@@ -454,6 +470,204 @@ static void constant_folding(struct radeon_compiler * c, struct rc_instruction *
 		constant_folding_add(inst);
 }
 
+/**
+ * This function returns a writemask that indicates wich components are
+ * read by src and also written by dst.
+ */
+static unsigned int src_reads_dst_mask(struct rc_src_register src,
+						struct rc_dst_register dst)
+{
+	unsigned int mask = 0;
+	unsigned int i;
+	if (dst.File != src.File || dst.Index != src.Index) {
+		return 0;
+	}
+
+	for(i = 0; i < 4; i++) {
+		mask |= 1 << GET_SWZ(src.Swizzle, i);
+	}
+	mask &= RC_MASK_XYZW;
+
+	return mask;
+}
+
+/* Return 1 if the source registers has a constant swizzle (e.g. 0, 0.5, 1.0)
+ * in any of its channels.  Return 0 otherwise. */
+static int src_has_const_swz(struct rc_src_register src) {
+	int chan;
+	for(chan = 0; chan < 4; chan++) {
+		unsigned int swz = GET_SWZ(src.Swizzle, chan);
+		if (swz == RC_SWIZZLE_ZERO || swz == RC_SWIZZLE_HALF
+						|| swz == RC_SWIZZLE_ONE) {
+			return 1;
+		}
+	}
+	return 0;
+}
+
+static void peephole_scan_write(void * data, struct rc_instruction * inst,
+		rc_register_file file, unsigned int index, unsigned int mask)
+{
+	struct peephole_state * s = data;
+	if(s->Inst->U.I.DstReg.File == file
+	   && s->Inst->U.I.DstReg.Index == index) {
+		unsigned int common_mask = s->WriteMask & mask;
+		s->WriteMask &= ~common_mask;
+	}
+}
+
+/**
+ * PRESUB_INV: ADD TEMP[0], none.1, -TEMP[1]
+ * Use the presubtract 1 - src0 for all readers of TEMP[0].  The first source
+ * of the add instruction must have the constatnt 1 swizzle.  This function
+ * does not check const registers to see if their value is 1.0, so it should
+ * be called after the constant_folding optimization.
+ * @return 
+ * 	0 if the ADD instruction is still part of the program.
+ * 	1 if the ADD instruction is no longer part of the program.
+ */
+static int peephole_add_presub_inv(
+	struct radeon_compiler * c,
+	struct rc_instruction * inst_add)
+{
+	unsigned int i, swz, mask;
+	unsigned int can_remove = 0;
+	unsigned int cant_sub = 0;
+	struct rc_instruction * inst;
+	struct peephole_state s;
+
+	if (inst_add->U.I.SaturateMode)
+		return 0;
+
+	mask = inst_add->U.I.DstReg.WriteMask;
+
+	/* Check if src0 is 1. */
+	/* XXX It would be nice to use is_src_uniform_constant here, but that
+	 * function only works if the register's file is RC_FILE_NONE */
+	for(i = 0; i < 4; i++ ) {
+		swz = GET_SWZ(inst_add->U.I.SrcReg[0].Swizzle, i);
+		if(((1 << i) & inst_add->U.I.DstReg.WriteMask)
+						&& swz != RC_SWIZZLE_ONE) {
+			return 0;
+		}
+	}
+
+	/* Check src1. */
+	if ((inst_add->U.I.SrcReg[1].Negate & inst_add->U.I.DstReg.WriteMask) !=
+						inst_add->U.I.DstReg.WriteMask
+		|| inst_add->U.I.SrcReg[1].Abs
+		|| (inst_add->U.I.SrcReg[1].File != RC_FILE_TEMPORARY
+			&& inst_add->U.I.SrcReg[1].File != RC_FILE_CONSTANT)
+		|| src_has_const_swz(inst_add->U.I.SrcReg[1])) {
+
+		return 0;
+	}
+
+	/* Setup the peephole_state information. */
+	s.Inst = inst_add;
+	s.WriteMask = inst_add->U.I.DstReg.WriteMask;
+
+	/* For all instructions that read inst_add->U.I.DstReg before it is
+	 * written again, use the 1 - src0 presubtact instead. */
+	for(inst = inst_add->Next; inst != &c->Program.Instructions;
+							inst = inst->Next) {
+		const struct rc_opcode_info * info =
+					rc_get_opcode_info(inst->U.I.Opcode);
+
+		for(i = 0; i < info->NumSrcRegs; i++) {
+			if(inst_add->U.I.DstReg.WriteMask !=
+					src_reads_dst_mask(inst->U.I.SrcReg[i],
+						inst_add->U.I.DstReg)) {
+				continue;
+			}
+			if (cant_sub) {
+				can_remove = 0;
+				break;
+			}
+			/* XXX: There are some situations where instructions
+			 * with more than 2 src registers can use the
+			 * presubtract select, but to keep things simple we
+			 * will disable presubtract on these instructions for
+			 * now. Note: This if statement should not be pulled
+			 * outside of the loop, because it only applies to
+			 * instructions that could potentially use the
+			 * presubtract source. */
+			if (info->NumSrcRegs > 2) {
+				can_remove = 0;
+				break;
+			}
+
+			/* We can't use more than one presubtract value in an
+			 * instruction, unless the two prsubtract operations
+			 * are the same and read from the same registers. */
+			if (inst->U.I.PreSub.Opcode != RC_PRESUB_NONE) {
+				if (inst->U.I.PreSub.Opcode != RC_PRESUB_INV
+					|| inst->U.I.PreSub.SrcReg[0].File !=
+						inst_add->U.I.SrcReg[1].File
+					|| inst->U.I.PreSub.SrcReg[0].Index !=
+						inst_add->U.I.SrcReg[1].Index) {
+
+					can_remove = 0;
+					break;
+				}
+			}
+			/* We must be careful not to modify inst_add, since it
+			 * is possible it will remain part of the program. */
+			inst->U.I.PreSub.SrcReg[0] = inst_add->U.I.SrcReg[1];
+			inst->U.I.PreSub.SrcReg[0].Negate = 0;
+			inst->U.I.PreSub.Opcode = RC_PRESUB_INV;
+			inst->U.I.SrcReg[i] = chain_srcregs(inst->U.I.SrcReg[i],
+						inst->U.I.PreSub.SrcReg[0]);
+
+			inst->U.I.SrcReg[i].File = RC_FILE_PRESUB;
+			inst->U.I.SrcReg[i].Index = RC_PRESUB_INV;
+			can_remove = 1;
+		}
+		if(!can_remove)
+			break;
+		rc_for_all_writes_mask(inst, peephole_scan_write, &s);
+		/* If all components of inst_add's destination register have
+		 * been written to by subsequent instructions, the original
+		 * value of the destination register is no longer valid and
+		 * we can't keep doing substitutions. */
+		if (!s.WriteMask){
+			break;
+		}
+		/* Make this instruction doesn't write to the presubtract source. */
+		if (inst->U.I.DstReg.WriteMask &
+				src_reads_dst_mask(inst_add->U.I.SrcReg[1],
+							inst->U.I.DstReg)
+				|| info->IsFlowControl) {
+			cant_sub = 1;
+		}
+	}
+	if(can_remove) {
+		rc_remove_instruction(inst_add);
+		return 1;
+	}
+	return 0;
+}
+
+/**
+ * @return
+ * 	0 if inst is still part of the program.
+ * 	1 if inst is no longer part of the program.
+ */
+static int peephole(struct radeon_compiler * c, struct rc_instruction * inst)
+{
+	switch(inst->U.I.Opcode){
+	case RC_OPCODE_ADD:
+		if (c->has_presub) {
+			if(peephole_add_presub_inv(c, inst))
+				return 1;
+		}
+		break;
+	default:
+		break;
+	}
+	return 0;
+}
+
 void rc_optimize(struct radeon_compiler * c, void *user)
 {
 	struct rc_instruction * inst = c->Program.Instructions.Next;
@@ -463,8 +677,11 @@ void rc_optimize(struct radeon_compiler * c, void *user)
 
 		constant_folding(c, cur);
 
+		if(peephole(c, cur))
+			continue;
+
 		if (cur->U.I.Opcode == RC_OPCODE_MOV) {
-			peephole(c, cur);
+			copy_propagate(c, cur);
 			/* cur may no longer be part of the program */
 		}
 	}
diff --git a/src/mesa/drivers/dri/r300/compiler/radeon_pair_schedule.c b/src/mesa/drivers/dri/r300/compiler/radeon_pair_schedule.c
index 8e232bb2436..32c54fd74bc 100644
--- a/src/mesa/drivers/dri/r300/compiler/radeon_pair_schedule.c
+++ b/src/mesa/drivers/dri/r300/compiler/radeon_pair_schedule.c
@@ -279,11 +279,118 @@ static int destructive_merge_instructions(
 		struct rc_pair_instruction * rgb,
 		struct rc_pair_instruction * alpha)
 {
+	const struct rc_opcode_info * opcode;
 	assert(rgb->Alpha.Opcode == RC_OPCODE_NOP);
 	assert(alpha->RGB.Opcode == RC_OPCODE_NOP);
 
+	/* Presubtract registers need to be merged first so that registers
+	 * needed by the presubtract operation can be placed in src0 and/or
+	 * src1. */
+
+	/* Merge the rgb presubtract registers. */
+	const struct rc_opcode_info * rgb_info =
+					rc_get_opcode_info(rgb->RGB.Opcode);
+	if (alpha->RGB.Src[RC_PAIR_PRESUB_SRC].Used) {
+		unsigned int srcp_src;
+		unsigned int srcp_regs;
+		if (rgb->RGB.Src[RC_PAIR_PRESUB_SRC].Used)
+			return 0;
+		srcp_regs = rc_presubtract_src_reg_count(
+				alpha->RGB.Src[RC_PAIR_PRESUB_SRC].Index);
+		for(srcp_src = 0; srcp_src < srcp_regs; srcp_src++) {
+			unsigned int arg;
+			int free_source;
+			struct radeon_pair_instruction_source srcp =
+						alpha->RGB.Src[srcp_src];
+			struct radeon_pair_instruction_source temp;
+			/* 2nd arg of 1 means this is an rgb source.
+			 * 3rd arg of 0 means this is not an alpha source. */
+			free_source = rc_pair_alloc_source(rgb, 1, 0,
+							srcp.File, srcp.Index);
+			/* If free_source == srcp_src, then either the
+			 * presubtract source is already in the correct place. */
+			if (free_source == srcp_src)
+				continue;
+			/* If free_source < 0 then there are no free source
+			 * slots. */
+			if (free_source < 0)
+				return 0;
+			/* Shuffle the sources, so we can put the
+			 * presubtract source in the correct place. */
+			for (arg = 0; arg < rgb_info->NumSrcRegs; arg++) {
+				/*If this arg does not read from an rgb source,
+				 * do nothing. */
+				if (rc_source_type_that_arg_reads(
+					rgb->RGB.Arg[arg].Source,
+					rgb->RGB.Arg[arg].Swizzle, 3)
+							!= RC_PAIR_SOURCE_RGB) {
+					continue;
+				}
+				if (rgb->RGB.Arg[arg].Source == srcp_src)
+					rgb->RGB.Arg[arg].Source = free_source;
+				/* We need to do this just in case register
+				 * is one of the sources already, but in the
+				 * wrong spot. */
+				else if(rgb->RGB.Arg[arg].Source == free_source)
+					rgb->RGB.Arg[arg].Source = srcp_src;
+			}
+			temp = rgb->RGB.Src[srcp_src];
+			rgb->RGB.Src[srcp_src] = rgb->RGB.Src[free_source];
+			rgb->RGB.Src[free_source] = temp;
+		}
+	}
+
+	/* Merge the alpha presubtract registers */
+	if (alpha->Alpha.Src[RC_PAIR_PRESUB_SRC].Used) {
+		unsigned int srcp_src;
+		unsigned int srcp_regs;
+		if(rgb->Alpha.Src[RC_PAIR_PRESUB_SRC].Used)
+			return 0;
+
+		srcp_regs = rc_presubtract_src_reg_count(
+			alpha->Alpha.Src[RC_PAIR_PRESUB_SRC].Index);
+		for(srcp_src = 0; srcp_src < srcp_regs; srcp_src++) {
+			unsigned int arg;
+			int free_source;
+			struct radeon_pair_instruction_source srcp =
+						alpha->Alpha.Src[srcp_src];
+			struct radeon_pair_instruction_source temp;
+			/* 2nd arg of 0 means this is not an rgb source.
+			 * 3rd arg of 1 means this is an alpha source. */
+			free_source = rc_pair_alloc_source(rgb, 0, 1,
+							srcp.File, srcp.Index);
+			/* If free_source == srcp_src, then either the
+			 * presubtract source is already in the correct place. */
+			if (free_source == srcp_src)
+				continue;
+			/* If free_source < 0 then there are no free source
+			 * slots. */
+			if (free_source < 0)
+				return 0;
+			/* Shuffle the sources, so we can put the
+			 * presubtract source in the correct place. */
+			for(arg = 0; arg < rgb_info->NumSrcRegs; arg++) {
+				/*If this arg does not read from an alpha
+				 * source, do nothing. */
+				if (rc_source_type_that_arg_reads(
+					rgb->RGB.Arg[arg].Source,
+					rgb->RGB.Arg[arg].Swizzle, 3)
+						!= RC_PAIR_SOURCE_ALPHA) {
+					continue;
+				}
+				if (rgb->RGB.Arg[arg].Source == srcp_src)
+					rgb->RGB.Arg[arg].Source = free_source;
+				else if (rgb->RGB.Arg[arg].Source == free_source)
+					rgb->RGB.Arg[arg].Source = srcp_src;
+			}
+			temp = rgb->Alpha.Src[srcp_src];
+			rgb->Alpha.Src[srcp_src] = rgb->Alpha.Src[free_source];
+			rgb->Alpha.Src[free_source] = temp;
+		}
+	}
+
 	/* Copy alpha args into rgb */
-	const struct rc_opcode_info * opcode = rc_get_opcode_info(alpha->Alpha.Opcode);
+	opcode = rc_get_opcode_info(alpha->Alpha.Opcode);
 
 	for(unsigned int arg = 0; arg < opcode->NumSrcRegs; ++arg) {
 		unsigned int srcrgb = 0;
@@ -351,7 +458,52 @@ static int merge_instructions(struct rc_pair_instruction * rgb, struct rc_pair_i
 	return 0;
 }
 
+static void presub_nop(struct rc_instruction * emitted) {
+	int prev_rgb_index, prev_alpha_index, i, num_src;
 
+	/* We don't need a nop if the previous instruction is a TEX. */
+	if (emitted->Prev->Type != RC_INSTRUCTION_PAIR) {
+		return;
+	}
+	if (emitted->Prev->U.P.RGB.WriteMask)
+		prev_rgb_index = emitted->Prev->U.P.RGB.DestIndex;
+	else
+		prev_rgb_index = -1;
+	if (emitted->Prev->U.P.Alpha.WriteMask)
+		prev_alpha_index = emitted->Prev->U.P.Alpha.DestIndex;
+	else
+		prev_alpha_index = 1;
+
+	/* Check the previous rgb instruction */
+	if (emitted->U.P.RGB.Src[RC_PAIR_PRESUB_SRC].Used) {
+		num_src = rc_presubtract_src_reg_count(
+				emitted->U.P.RGB.Src[RC_PAIR_PRESUB_SRC].Index);
+		for (i = 0; i < num_src; i++) {
+			unsigned int index = emitted->U.P.RGB.Src[i].Index;
+			if (emitted->U.P.RGB.Src[i].File == RC_FILE_TEMPORARY
+			    && (index  == prev_rgb_index
+				|| index == prev_alpha_index)) {
+				emitted->Prev->U.P.Nop = 1;
+				return;
+			}
+		}
+	}
+
+	/* Check the previous alpha instruction. */
+	if (!emitted->U.P.Alpha.Src[RC_PAIR_PRESUB_SRC].Used)
+		return;
+
+	num_src = rc_presubtract_src_reg_count(
+				emitted->U.P.Alpha.Src[RC_PAIR_PRESUB_SRC].Index);
+	for (i = 0; i < num_src; i++) {
+		unsigned int index = emitted->U.P.Alpha.Src[i].Index;
+		if(emitted->U.P.Alpha.Src[i].File == RC_FILE_TEMPORARY
+		   && (index == prev_rgb_index || index == prev_alpha_index)) {
+			emitted->Prev->U.P.Nop = 1;
+			return;
+		}
+	}
+}
 /**
  * Find a good ALU instruction or pair of ALU instruction and emit it.
  *
@@ -408,6 +560,10 @@ static void emit_one_alu(struct schedule_state *s, struct rc_instruction * befor
 		commit_alu_instruction(s, sinst);
 	success: ;
 	}
+	/* If the instruction we just emitted uses a presubtract value, and
+	 * the presubtract sources were written by the previous intstruction,
+	 * the previous instruction needs a nop. */
+	presub_nop(before->Prev);
 }
 
 static void scan_read(void * data, struct rc_instruction * inst,
diff --git a/src/mesa/drivers/dri/r300/compiler/radeon_pair_translate.c b/src/mesa/drivers/dri/r300/compiler/radeon_pair_translate.c
index 9fe39344f8e..4cdb7ea748e 100644
--- a/src/mesa/drivers/dri/r300/compiler/radeon_pair_translate.c
+++ b/src/mesa/drivers/dri/r300/compiler/radeon_pair_translate.c
@@ -127,6 +127,18 @@ static void classify_instruction(struct rc_sub_instruction * inst,
 	}
 }
 
+static void src_uses(struct rc_src_register src, unsigned int * rgb,
+							unsigned int * alpha)
+{
+	int j;
+	for(j = 0; j < 4; ++j) {
+		unsigned int swz = GET_SWZ(src.Swizzle, j);
+		if (swz < 3)
+			*rgb = 1;
+		else if (swz < 4)
+			*alpha = 1;
+	}
+}
 
 /**
  * Fill the given ALU instruction's opcodes and source operands into the given pair,
@@ -158,12 +170,51 @@ static void set_pair_instruction(struct r300_fragment_program_compiler *c,
 	const struct rc_opcode_info * opcode = rc_get_opcode_info(inst->Opcode);
 	int i;
 
+	/* Presubtract handling:
+	 * We need to make sure that the values used by the presubtract
+	 * operation end up in src0 or src1. */
+	if(inst->PreSub.Opcode != RC_PRESUB_NONE) {
+		/* rc_pair_alloc_source() will fill in data for
+		 * pair->{RGB,ALPHA}.Src[RC_PAIR_PRESUB_SRC] */
+		int j;
+		for(j = 0; j < 3; j++) {
+			int src_regs;
+			if(inst->SrcReg[j].File != RC_FILE_PRESUB)
+				continue;
+
+			src_regs = rc_presubtract_src_reg_count(
+							inst->PreSub.Opcode);
+			for(i = 0; i < src_regs; i++) {
+				unsigned int rgb = 0;
+				unsigned int alpha = 0;
+				src_uses(inst->SrcReg[j], &rgb, &alpha);
+				if(rgb) {
+					pair->RGB.Src[i].File =
+						inst->PreSub.SrcReg[i].File;
+					pair->RGB.Src[i].Index =
+						inst->PreSub.SrcReg[i].Index;
+					pair->RGB.Src[i].Used = 1;
+				}
+				if(alpha) {
+					pair->Alpha.Src[i].File =
+						inst->PreSub.SrcReg[i].File;
+					pair->Alpha.Src[i].Index =
+						inst->PreSub.SrcReg[i].Index;
+					pair->Alpha.Src[i].Used = 1;
+				}
+			}
+		}
+	}
+
 	for(i = 0; i < opcode->NumSrcRegs; ++i) {
 		int source;
 		if (needrgb && !istranscendent) {
 			unsigned int srcrgb = 0;
 			unsigned int srcalpha = 0;
 			int j;
+			/* We don't care about the alpha channel here.  We only
+			 * want the part of the swizzle that writes to rgb,
+			 * since we are creating an rgb instruction. */
 			for(j = 0; j < 3; ++j) {
 				unsigned int swz = GET_SWZ(inst->SrcReg[i].Swizzle, j);
 				if (swz < 3)
@@ -173,6 +224,7 @@ static void set_pair_instruction(struct r300_fragment_program_compiler *c,
 			}
 			source = rc_pair_alloc_source(pair, srcrgb, srcalpha,
 							inst->SrcReg[i].File, inst->SrcReg[i].Index);
+			assert(source != -1);
 			pair->RGB.Arg[i].Source = source;
 			pair->RGB.Arg[i].Swizzle = inst->SrcReg[i].Swizzle & 0x1ff;
 			pair->RGB.Arg[i].Abs = inst->SrcReg[i].Abs;
@@ -188,6 +240,7 @@ static void set_pair_instruction(struct r300_fragment_program_compiler *c,
 				srcalpha = 1;
 			source = rc_pair_alloc_source(pair, srcrgb, srcalpha,
 							inst->SrcReg[i].File, inst->SrcReg[i].Index);
+			assert(source != -1);
 			pair->Alpha.Arg[i].Source = source;
 			pair->Alpha.Arg[i].Swizzle = swz;
 			pair->Alpha.Arg[i].Abs = inst->SrcReg[i].Abs;
diff --git a/src/mesa/drivers/dri/r300/compiler/radeon_program.h b/src/mesa/drivers/dri/r300/compiler/radeon_program.h
index 5582f56d921..f0a77d7b539 100644
--- a/src/mesa/drivers/dri/r300/compiler/radeon_program.h
+++ b/src/mesa/drivers/dri/r300/compiler/radeon_program.h
@@ -39,7 +39,7 @@
 struct radeon_compiler;
 
 struct rc_src_register {
-	unsigned int File:3;
+	unsigned int File:4;
 
 	/** Negative values may be used for relative addressing. */
 	signed int Index:(RC_REGISTER_INDEX_BITS+1);
@@ -64,6 +64,11 @@ struct rc_dst_register {
 	unsigned int WriteMask:4;
 };
 
+struct rc_presub_instruction {
+	rc_presubtract_op Opcode;
+	struct rc_src_register SrcReg[2];
+};
+
 /**
  * Instructions are maintained by the compiler in a doubly linked list
  * of these structures.
@@ -108,6 +113,10 @@ struct rc_sub_instruction {
 	/** True if tex instruction should do shadow comparison */
 	unsigned int TexShadow:1;
 	/*@}*/
+
+	/** This holds information about the presubtract operation used by
+	 * this instruction. */
+	struct rc_presub_instruction PreSub;
 };
 
 typedef enum {
diff --git a/src/mesa/drivers/dri/r300/compiler/radeon_program_constants.h b/src/mesa/drivers/dri/r300/compiler/radeon_program_constants.h
index 2ddf60b6774..9dcd44c522d 100644
--- a/src/mesa/drivers/dri/r300/compiler/radeon_program_constants.h
+++ b/src/mesa/drivers/dri/r300/compiler/radeon_program_constants.h
@@ -79,7 +79,13 @@ typedef enum {
 	/**
 	 * Indicates a special register, see RC_SPECIAL_xxx.
 	 */
-	RC_FILE_SPECIAL
+	RC_FILE_SPECIAL,
+
+	/**
+	 * Indicates this register should use the result of the presubtract
+	 * operation.
+	 */
+	RC_FILE_PRESUB
 } rc_register_file;
 
 enum {
@@ -147,4 +153,32 @@ typedef enum {
 	RC_ALURESULT_W
 } rc_write_aluresult;
 
+typedef enum {
+	RC_PRESUB_NONE = 0,
+
+	/** 1 - 2 * src0 */
+	RC_PRESUB_BIAS,
+
+	/** src1 - src0 */
+	RC_PRESUB_SUB,
+
+	/** src1 + src0 */
+	RC_PRESUB_ADD,
+
+	/** 1 - src0 */
+	RC_PRESUB_INV
+} rc_presubtract_op;
+
+static inline int rc_presubtract_src_reg_count(rc_presubtract_op op){
+	switch(op){
+	case RC_PRESUB_BIAS:
+	case RC_PRESUB_INV:
+		return 1;
+	case RC_PRESUB_ADD:
+	case RC_PRESUB_SUB:
+		return 2;
+	default:
+		return 0;
+	}
+}
 #endif /* RADEON_PROGRAM_CONSTANTS_H */
diff --git a/src/mesa/drivers/dri/r300/compiler/radeon_program_pair.c b/src/mesa/drivers/dri/r300/compiler/radeon_program_pair.c
index ee839596aab..5a50584b725 100644
--- a/src/mesa/drivers/dri/r300/compiler/radeon_program_pair.c
+++ b/src/mesa/drivers/dri/r300/compiler/radeon_program_pair.c
@@ -38,26 +38,52 @@ int rc_pair_alloc_source(struct rc_pair_instruction *pair,
 {
 	int candidate = -1;
 	int candidate_quality = -1;
+	unsigned int alpha_used = 0;
+	unsigned int rgb_used = 0;
 	int i;
 
 	if ((!rgb && !alpha) || file == RC_FILE_NONE)
 		return 0;
 
+	if (rgb && pair->RGB.Src[RC_PAIR_PRESUB_SRC].Used) {
+		if (file == RC_FILE_PRESUB) {
+			if (index != pair->RGB.Src[RC_PAIR_PRESUB_SRC].Index) {
+				return -1;
+			}
+		} else {
+			rgb_used++;
+		}
+	}
+
+	if (alpha && pair->Alpha.Src[RC_PAIR_PRESUB_SRC].Used) {
+		if (file == RC_FILE_PRESUB) {
+			if (index != pair->Alpha.Src[RC_PAIR_PRESUB_SRC].Index) {
+				return -1;
+			}
+		} else {
+			alpha_used++;
+		}
+	}
+
 	for(i = 0; i < 3; ++i) {
 		int q = 0;
 		if (rgb) {
 			if (pair->RGB.Src[i].Used) {
 				if (pair->RGB.Src[i].File != file ||
-				    pair->RGB.Src[i].Index != index)
+				    pair->RGB.Src[i].Index != index) {
+					rgb_used++;
 					continue;
+				}
 				q++;
 			}
 		}
 		if (alpha) {
 			if (pair->Alpha.Src[i].Used) {
 				if (pair->Alpha.Src[i].File != file ||
-				    pair->Alpha.Src[i].Index != index)
+				    pair->Alpha.Src[i].Index != index) {
+					alpha_used++;
 					continue;
+				}
 				q++;
 			}
 		}
@@ -66,19 +92,156 @@ int rc_pair_alloc_source(struct rc_pair_instruction *pair,
 			candidate = i;
 		}
 	}
+	if (candidate < 0 || (rgb && rgb_used > 2) || (alpha && alpha_used > 2))
+		return -1;
 
-	if (candidate >= 0) {
-		if (rgb) {
-			pair->RGB.Src[candidate].Used = 1;
-			pair->RGB.Src[candidate].File = file;
-			pair->RGB.Src[candidate].Index = index;
+	/* candidate >= 0 */
+
+	/* Even if we have a presub src, the above loop needs to run,
+	 * because we still need to make sure there is a free source.
+	 */
+	if (file == RC_FILE_PRESUB)
+		candidate = RC_PAIR_PRESUB_SRC;
+
+	if (rgb) {
+		pair->RGB.Src[candidate].Used = 1;
+		pair->RGB.Src[candidate].File = file;
+		pair->RGB.Src[candidate].Index = index;
+		if (candidate == RC_PAIR_PRESUB_SRC) {
+			/* For registers with the RC_FILE_PRESUB file,
+			 * the index stores the presubtract op. */
+			int src_regs = rc_presubtract_src_reg_count(index);
+			for(i = 0; i < src_regs; i++) {
+				pair->RGB.Src[i].Used = 1;
+			}
 		}
-		if (alpha) {
-			pair->Alpha.Src[candidate].Used = 1;
-			pair->Alpha.Src[candidate].File = file;
-			pair->Alpha.Src[candidate].Index = index;
+	}
+	if (alpha) {
+		pair->Alpha.Src[candidate].Used = 1;
+		pair->Alpha.Src[candidate].File = file;
+		pair->Alpha.Src[candidate].Index = index;
+		if (candidate == RC_PAIR_PRESUB_SRC) {
+			/* For registers with the RC_FILE_PRESUB file,
+			 * the index stores the presubtract op. */
+			int src_regs = rc_presubtract_src_reg_count(index);
+			for(i=0; i < src_regs; i++) {
+				pair->Alpha.Src[i].Used = 1;
+			}
 		}
 	}
 
 	return candidate;
 }
+
+static void pair_foreach_source_callback(
+	struct rc_pair_instruction * pair,
+	void * data,
+	rc_pair_foreach_src_fn cb,
+	unsigned int swz,
+	unsigned int src)
+{
+	/* swz > 3 means that the swizzle is either not used, or a constant
+	 * swizzle (e.g. 0, 1, 0.5). */
+	if(swz > 3)
+		return;
+
+	if(swz == RC_SWIZZLE_W) {
+		if (src == RC_PAIR_PRESUB_SRC) {
+			unsigned int i;
+			unsigned int src_count = rc_presubtract_src_reg_count(
+				pair->Alpha.Src[RC_PAIR_PRESUB_SRC].Index);
+			for(i = 0; i < src_count; i++) {
+				cb(data, &pair->Alpha.Src[i]);
+			}
+		} else {
+			cb(data, &pair->Alpha.Src[src]);
+		}
+	} else {
+		if (src == RC_PAIR_PRESUB_SRC) {
+			unsigned int i;
+			unsigned int src_count = rc_presubtract_src_reg_count(
+				pair->RGB.Src[RC_PAIR_PRESUB_SRC].Index);
+			for(i = 0; i < src_count; i++) {
+				cb(data, &pair->RGB.Src[i]);
+			}
+		}
+		else {
+			cb(data, &pair->RGB.Src[src]);
+		}
+	}
+}
+
+void rc_pair_foreach_source_that_alpha_reads(
+	struct rc_pair_instruction * pair,
+	void * data,
+	rc_pair_foreach_src_fn cb)
+{
+	unsigned int i;
+	const struct rc_opcode_info * info =
+				rc_get_opcode_info(pair->Alpha.Opcode);
+	for(i = 0; i < info->NumSrcRegs; i++) {
+		pair_foreach_source_callback(pair, data, cb,
+					GET_SWZ(pair->Alpha.Arg[i].Swizzle, 0),
+					pair->Alpha.Arg[i].Source);
+	}
+}
+
+void rc_pair_foreach_source_that_rgb_reads(
+	struct rc_pair_instruction * pair,
+	void * data,
+	rc_pair_foreach_src_fn cb)
+{
+	unsigned int i;
+	const struct rc_opcode_info * info =
+				rc_get_opcode_info(pair->RGB.Opcode);
+	for(i = 0; i < info->NumSrcRegs; i++) {
+		unsigned int chan;
+		unsigned int swz = RC_SWIZZLE_UNUSED;
+		/* Find a swizzle that is either X,Y,Z,or W.  We assume here
+		 * that if one channel swizzles X,Y, or Z, then none of the
+		 * other channels swizzle W, and vice-versa. */
+		for(chan = 0; chan < 4; chan++) {
+			swz = GET_SWZ(pair->RGB.Arg[i].Swizzle, chan);
+			if(swz == RC_SWIZZLE_X || swz == RC_SWIZZLE_Y
+			|| swz == RC_SWIZZLE_Z || swz == RC_SWIZZLE_W)
+				continue;
+		}
+		pair_foreach_source_callback(pair, data, cb,
+					swz,
+					pair->RGB.Arg[i].Source);
+	}
+}
+
+/*return 0 for rgb, 1 for alpha -1 for error. */
+
+rc_pair_source_type rc_source_type_that_arg_reads(
+	unsigned int source,
+	unsigned int swizzle,
+	unsigned int channels)
+{
+	unsigned int chan;
+	unsigned int swz = RC_SWIZZLE_UNUSED;
+	int isRGB = 0;
+	int isAlpha = 0;
+	/* Find a swizzle that is either X,Y,Z,or W.  We assume here
+	 * that if one channel swizzles X,Y, or Z, then none of the
+	 * other channels swizzle W, and vice-versa. */
+	for(chan = 0; chan < channels; chan++) {
+		swz = GET_SWZ(swizzle, chan);
+		if (swz == RC_SWIZZLE_W) {
+			isAlpha = 1;
+		} else if (swz == RC_SWIZZLE_X || swz == RC_SWIZZLE_Y
+						|| swz == RC_SWIZZLE_Z) {
+			isRGB = 1;
+		}
+	}
+	assert(!isRGB || !isAlpha);
+
+	if(!isRGB && !isAlpha)
+		return RC_PAIR_SOURCE_NONE;
+
+	if (isRGB)
+		return RC_PAIR_SOURCE_RGB;
+	/*isAlpha*/
+	return RC_PAIR_SOURCE_ALPHA;
+}
diff --git a/src/mesa/drivers/dri/r300/compiler/radeon_program_pair.h b/src/mesa/drivers/dri/r300/compiler/radeon_program_pair.h
index ef5a0347009..e0061e454bf 100644
--- a/src/mesa/drivers/dri/r300/compiler/radeon_program_pair.h
+++ b/src/mesa/drivers/dri/r300/compiler/radeon_program_pair.h
@@ -49,6 +49,11 @@ struct radeon_compiler;
  * see \ref rc_pair_translate
  */
 
+/* For rgb and alpha instructions when arg[n].Source = RC_PAIR_PRESUB_SRC, then
+ * the presubtract value will be used, and
+ * {RGB,Alpha}.Src[RC_PAIR_PRESUB_SRC].File will be set to RC_FILE_PRESUB.
+ */
+#define RC_PAIR_PRESUB_SRC 3
 
 struct radeon_pair_instruction_source {
 	unsigned int Used:1;
@@ -64,7 +69,7 @@ struct radeon_pair_instruction_rgb {
 	unsigned int OutputWriteMask:3;
 	unsigned int Saturate:1;
 
-	struct radeon_pair_instruction_source Src[3];
+	struct radeon_pair_instruction_source Src[4];
 
 	struct {
 		unsigned int Source:2;
@@ -83,7 +88,7 @@ struct radeon_pair_instruction_alpha {
 	unsigned int DepthWriteMask:1;
 	unsigned int Saturate:1;
 
-	struct radeon_pair_instruction_source Src[3];
+	struct radeon_pair_instruction_source Src[4];
 
 	struct {
 		unsigned int Source:2;
@@ -99,8 +104,17 @@ struct rc_pair_instruction {
 
 	unsigned int WriteALUResult:2;
 	unsigned int ALUResultCompare:3;
+	unsigned int Nop:1;
 };
 
+typedef void (*rc_pair_foreach_src_fn)
+			(void *, struct radeon_pair_instruction_source *);
+
+typedef enum {
+	RC_PAIR_SOURCE_NONE = 0,
+	RC_PAIR_SOURCE_RGB,
+	RC_PAIR_SOURCE_ALPHA
+} rc_pair_source_type;
 
 /**
  * General helper functions for dealing with the paired instruction format.
@@ -109,6 +123,21 @@ struct rc_pair_instruction {
 int rc_pair_alloc_source(struct rc_pair_instruction *pair,
 	unsigned int rgb, unsigned int alpha,
 	rc_register_file file, unsigned int index);
+
+void rc_pair_foreach_source_that_alpha_reads(
+	struct rc_pair_instruction * pair,
+	void * data,
+	rc_pair_foreach_src_fn cb);
+
+void rc_pair_foreach_source_that_rgb_reads(
+	struct rc_pair_instruction * pair,
+	void * data,
+	rc_pair_foreach_src_fn cb);
+
+rc_pair_source_type rc_source_type_that_arg_reads(
+	unsigned int source,
+	unsigned int swizzle,
+	unsigned int channels);
 /*@}*/
 
 
diff --git a/src/mesa/drivers/dri/r300/compiler/radeon_program_print.c b/src/mesa/drivers/dri/r300/compiler/radeon_program_print.c
index a356e94e032..01612195810 100644
--- a/src/mesa/drivers/dri/r300/compiler/radeon_program_print.c
+++ b/src/mesa/drivers/dri/r300/compiler/radeon_program_print.c
@@ -38,6 +38,24 @@ static const char * textarget_to_string(rc_texture_target target)
 	}
 }
 
+static const char * presubtract_op_to_string(rc_presubtract_op op)
+{
+	switch(op) {
+	case RC_PRESUB_NONE:
+		return "NONE";
+	case RC_PRESUB_BIAS:
+		return "(1 - 2 * src0)";
+	case RC_PRESUB_SUB:
+		return "(src1 - src0)";
+	case RC_PRESUB_ADD:
+		return "(src1 + src0)";
+	case RC_PRESUB_INV:
+		return "(1 - src0)";
+	default:
+		return "BAD_PRESUBTRACT_OP";
+	}
+}
+
 static void rc_print_comparefunc(FILE * f, const char * lhs, rc_compare_func func, const char * rhs)
 {
 	if (func == RC_COMPARE_FUNC_NEVER) {
@@ -125,7 +143,43 @@ static void rc_print_swizzle(FILE * f, unsigned int swizzle, unsigned int negate
 	}
 }
 
-static void rc_print_src_register(FILE * f, struct rc_src_register src)
+static void rc_print_presub_instruction(FILE * f,
+					struct rc_presub_instruction inst)
+{
+	fprintf(f,"(");
+	switch(inst.Opcode){
+	case RC_PRESUB_BIAS:
+		fprintf(f, "1 - 2 * ");
+		rc_print_register(f, inst.SrcReg[0].File,
+				inst.SrcReg[0].Index,inst.SrcReg[0].RelAddr);
+		break;
+	case RC_PRESUB_SUB:
+		rc_print_register(f, inst.SrcReg[0].File,
+				inst.SrcReg[0].Index,inst.SrcReg[0].RelAddr);
+		fprintf(f, " - ");
+		rc_print_register(f, inst.SrcReg[1].File,
+				inst.SrcReg[1].Index,inst.SrcReg[1].RelAddr);
+		break;
+	case RC_PRESUB_ADD:
+		rc_print_register(f, inst.SrcReg[0].File,
+				inst.SrcReg[0].Index,inst.SrcReg[0].RelAddr);
+		fprintf(f, " + ");
+		rc_print_register(f, inst.SrcReg[1].File,
+				inst.SrcReg[1].Index,inst.SrcReg[1].RelAddr);
+		break;
+	case RC_PRESUB_INV:
+		fprintf(f, "1 - ");
+		rc_print_register(f, inst.SrcReg[0].File,
+				inst.SrcReg[0].Index,inst.SrcReg[0].RelAddr);
+		break;
+	default:
+		break;
+	}
+	fprintf(f, ")");
+}
+
+static void rc_print_src_register(FILE * f, struct rc_instruction * inst,
+						struct rc_src_register src)
 {
 	int trivial_negate = (src.Negate == RC_MASK_NONE || src.Negate == RC_MASK_XYZW);
 
@@ -134,7 +188,10 @@ static void rc_print_src_register(FILE * f, struct rc_src_register src)
 	if (src.Abs)
 		fprintf(f, "|");
 
-	rc_print_register(f, src.File, src.Index, src.RelAddr);
+	if(src.File == RC_FILE_PRESUB)
+		rc_print_presub_instruction(f, inst->U.I.PreSub);
+	else
+		rc_print_register(f, src.File, src.Index, src.RelAddr);
 
 	if (src.Abs && !trivial_negate)
 		fprintf(f, "|");
@@ -198,7 +255,7 @@ static void rc_print_normal_instruction(FILE * f, struct rc_instruction * inst,
 		if (reg > 0)
 			fprintf(f, ",");
 		fprintf(f, " ");
-		rc_print_src_register(f, inst->U.I.SrcReg[reg]);
+		rc_print_src_register(f, inst, inst->U.I.SrcReg[reg]);
 	}
 
 	if (opcode->HasTexture) {
@@ -247,6 +304,16 @@ static void rc_print_pair_instruction(FILE * f, struct rc_instruction * fullinst
 			printedsrc = 1;
 		}
 	}
+	if(inst->RGB.Src[RC_PAIR_PRESUB_SRC].Used) {
+		fprintf(f, ", srcp.xyz = %s",
+			presubtract_op_to_string(
+					inst->RGB.Src[RC_PAIR_PRESUB_SRC].Index));
+	}
+	if(inst->Alpha.Src[RC_PAIR_PRESUB_SRC].Used) {
+		fprintf(f, ", srcp.w = %s",
+			presubtract_op_to_string(
+					inst->Alpha.Src[RC_PAIR_PRESUB_SRC].Index));
+	}
 	fprintf(f, "\n");
 
 	if (inst->RGB.Opcode != RC_OPCODE_NOP) {
@@ -272,7 +339,12 @@ static void rc_print_pair_instruction(FILE * f, struct rc_instruction * fullinst
 		for(unsigned int arg = 0; arg < opcode->NumSrcRegs; ++arg) {
 			const char* abs = inst->RGB.Arg[arg].Abs ? "|" : "";
 			const char* neg = inst->RGB.Arg[arg].Negate ? "-" : "";
-			fprintf(f, ", %s%ssrc%i.%c%c%c%s", neg, abs, inst->RGB.Arg[arg].Source,
+			fprintf(f, ", %s%ssrc", neg, abs);
+			if(inst->RGB.Arg[arg].Source == RC_PAIR_PRESUB_SRC)
+				fprintf(f,"p");
+			else
+				fprintf(f,"%d", inst->RGB.Arg[arg].Source);
+			fprintf(f,".%c%c%c%s",
 				rc_swizzle_char(GET_SWZ(inst->RGB.Arg[arg].Swizzle, 0)),
 				rc_swizzle_char(GET_SWZ(inst->RGB.Arg[arg].Swizzle, 1)),
 				rc_swizzle_char(GET_SWZ(inst->RGB.Arg[arg].Swizzle, 2)),
@@ -300,7 +372,12 @@ static void rc_print_pair_instruction(FILE * f, struct rc_instruction * fullinst
 		for(unsigned int arg = 0; arg < opcode->NumSrcRegs; ++arg) {
 			const char* abs = inst->Alpha.Arg[arg].Abs ? "|" : "";
 			const char* neg = inst->Alpha.Arg[arg].Negate ? "-" : "";
-			fprintf(f, ", %s%ssrc%i.%c%s", neg, abs, inst->Alpha.Arg[arg].Source,
+			fprintf(f, ", %s%ssrc", neg, abs);
+			if(inst->Alpha.Arg[arg].Source == RC_PAIR_PRESUB_SRC)
+				fprintf(f,"p");
+			else
+				fprintf(f,"%d", inst->Alpha.Arg[arg].Source);
+			fprintf(f,".%c%s",
 				rc_swizzle_char(inst->Alpha.Arg[arg].Swizzle), abs);
 		}
 		fprintf(f, "\n");
diff --git a/src/mesa/drivers/dri/r300/compiler/radeon_remove_constants.c b/src/mesa/drivers/dri/r300/compiler/radeon_remove_constants.c
index 9281feecfa1..facea382f4e 100644
--- a/src/mesa/drivers/dri/r300/compiler/radeon_remove_constants.c
+++ b/src/mesa/drivers/dri/r300/compiler/radeon_remove_constants.c
@@ -27,6 +27,16 @@
 
 #include "radeon_remove_constants.h"
 
+static void remap_regs(void * userdata, struct rc_instruction * inst,
+			rc_register_file * pfile, unsigned int * pindex)
+{
+        unsigned *inv_remap_table = userdata;
+
+        if (*pfile == RC_FILE_CONSTANT) {
+                *pindex = inv_remap_table[*pindex];
+        }
+}
+
 void rc_remove_unused_constants(struct radeon_compiler *c, void *user)
 {
 	unsigned **out_remap_table = (unsigned**)user;
@@ -51,6 +61,10 @@ void rc_remove_unused_constants(struct radeon_compiler *c, void *user)
              inst != &c->Program.Instructions; inst = inst->Next) {
                 const struct rc_opcode_info *opcode = rc_get_opcode_info(inst->U.I.Opcode);
 
+                /* XXX: This loop and the if statement after it should be
+                 * replaced by a call to one of the rc_for_all_reads_* functions.
+                 * The reason it does not use one of those functions now is
+                 * because none of them have RelAddr as an argument. */
                 for (unsigned i = 0; i < opcode->NumSrcRegs; i++) {
                         if (inst->U.I.SrcReg[i].File == RC_FILE_CONSTANT) {
                                 if (inst->U.I.SrcReg[i].RelAddr) {
@@ -60,6 +74,18 @@ void rc_remove_unused_constants(struct radeon_compiler *c, void *user)
                                 }
                         }
                 }
+                if (inst->U.I.PreSub.Opcode != RC_PRESUB_NONE) {
+			unsigned int i;
+			unsigned int srcp_regs = rc_presubtract_src_reg_count(
+							inst->U.I.PreSub.Opcode);
+			for( i = 0; i < srcp_regs; i++) {
+                                if (inst->U.I.PreSub.SrcReg[i].File ==
+                                                        RC_FILE_CONSTANT) {
+                                        const_used[
+                                            inst->U.I.PreSub.SrcReg[i].Index] = 1;
+                                }
+                        }
+		}
         }
 
         /* Pass 2: If there is relative addressing, mark all externals as used. */
@@ -100,13 +126,7 @@ void rc_remove_unused_constants(struct radeon_compiler *c, void *user)
         if (!is_identity) {
                 for (struct rc_instruction *inst = c->Program.Instructions.Next;
                      inst != &c->Program.Instructions; inst = inst->Next) {
-                        const struct rc_opcode_info *opcode = rc_get_opcode_info(inst->U.I.Opcode);
-
-                        for (unsigned i = 0; i < opcode->NumSrcRegs; i++) {
-                                if (inst->U.I.SrcReg[i].File == RC_FILE_CONSTANT) {
-                                        inst->U.I.SrcReg[i].Index = inv_remap_table[inst->U.I.SrcReg[i].Index];
-                                }
-                        }
+                        rc_remap_registers(inst, remap_regs, inv_remap_table);
                 }
 
 	}
-- 
2.30.2