r300: Correctly scan for used temporary registers

This fixes a regression introduced by dea8719f0...

6 files changed, 126 insertions, 79 deletions

diff --git a/src/mesa/drivers/dri/r300/r300_fragprog.c b/src/mesa/drivers/dri/r300/r300_fragprog.c
index 8c49e8ada68..6a8ef0ef5fc 100644
--- a/src/mesa/drivers/dri/r300/r300_fragprog.c
+++ b/src/mesa/drivers/dri/r300/r300_fragprog.c
@@ -68,7 +68,7 @@ static void reset_srcreg(struct prog_src_register* reg)
  * be reused.
  */
 static GLboolean transform_TEX(
-	GLcontext *ctx, struct gl_program *p,
+	struct radeon_transform_context *t,
 	struct prog_instruction* orig_inst, void* data)
 {
 	struct r300_fragment_program_compiler *compiler =
@@ -84,11 +84,11 @@ static GLboolean transform_TEX(
 		return GL_FALSE;
 
 	if (inst.Opcode != OPCODE_KIL &&
-	    p->ShadowSamplers & (1 << inst.TexSrcUnit)) {
+	    t->Program->ShadowSamplers & (1 << inst.TexSrcUnit)) {
 		GLuint comparefunc = GL_NEVER + compiler->fp->state.unit[inst.TexSrcUnit].texture_compare_func;
 
 		if (comparefunc == GL_NEVER || comparefunc == GL_ALWAYS) {
-			tgt = radeonAppendInstructions(p, 1);
+			tgt = radeonAppendInstructions(t->Program, 1);
 
 			tgt->Opcode = OPCODE_MOV;
 			tgt->DstReg = inst.DstReg;
@@ -98,7 +98,7 @@ static GLboolean transform_TEX(
 		}
 
 		inst.DstReg.File = PROGRAM_TEMPORARY;
-		inst.DstReg.Index = _mesa_find_free_register(p, PROGRAM_TEMPORARY);
+		inst.DstReg.Index = radeonFindFreeTemporary(t);
 		inst.DstReg.WriteMask = WRITEMASK_XYZW;
 	}
 
@@ -113,7 +113,7 @@ static GLboolean transform_TEX(
 			0
 		};
 
-		int tempreg = _mesa_find_free_register(p, PROGRAM_TEMPORARY);
+		int tempreg = radeonFindFreeTemporary(t);
 		int factor_index;
 
 		tokens[2] = inst.TexSrcUnit;
@@ -121,7 +121,7 @@ static GLboolean transform_TEX(
 			_mesa_add_state_reference(
 				compiler->fp->mesa_program.Base.Parameters, tokens);
 
-		tgt = radeonAppendInstructions(p, 1);
+		tgt = radeonAppendInstructions(t->Program, 1);
 
 		tgt->Opcode = OPCODE_MUL;
 		tgt->DstReg.File = PROGRAM_TEMPORARY;
@@ -140,9 +140,9 @@ static GLboolean transform_TEX(
 	 */
 	if (inst.SrcReg[0].Swizzle != SWIZZLE_NOOP ||
 	    inst.SrcReg[0].Abs || inst.SrcReg[0].NegateBase || inst.SrcReg[0].NegateAbs) {
-		int tempreg = _mesa_find_free_register(p, PROGRAM_TEMPORARY);
+		int tempreg = radeonFindFreeTemporary(t);
 
-		tgt = radeonAppendInstructions(p, 1);
+		tgt = radeonAppendInstructions(t->Program, 1);
 
 		tgt->Opcode = OPCODE_MOV;
 		tgt->DstReg.File = PROGRAM_TEMPORARY;
@@ -157,7 +157,7 @@ static GLboolean transform_TEX(
 	if (inst.Opcode != OPCODE_KIL) {
 		if (inst.DstReg.File != PROGRAM_TEMPORARY ||
 		    inst.DstReg.WriteMask != WRITEMASK_XYZW) {
-			int tempreg = _mesa_find_free_register(p, PROGRAM_TEMPORARY);
+			int tempreg = radeonFindFreeTemporary(t);
 
 			inst.DstReg.File = PROGRAM_TEMPORARY;
 			inst.DstReg.Index = tempreg;
@@ -166,16 +166,16 @@ static GLboolean transform_TEX(
 		}
 	}
 
-	tgt = radeonAppendInstructions(p, 1);
+	tgt = radeonAppendInstructions(t->Program, 1);
 	_mesa_copy_instructions(tgt, &inst, 1);
 
 	if (inst.Opcode != OPCODE_KIL &&
-	    p->ShadowSamplers & (1 << inst.TexSrcUnit)) {
+	    t->Program->ShadowSamplers & (1 << inst.TexSrcUnit)) {
 		GLuint comparefunc = GL_NEVER + compiler->fp->state.unit[inst.TexSrcUnit].texture_compare_func;
 		GLuint depthmode = compiler->fp->state.unit[inst.TexSrcUnit].depth_texture_mode;
-		int rcptemp = _mesa_find_free_register(p, PROGRAM_TEMPORARY);
+		int rcptemp = radeonFindFreeTemporary(t);
 
-		tgt = radeonAppendInstructions(p, 3);
+		tgt = radeonAppendInstructions(t->Program, 3);
 
 		tgt[0].Opcode = OPCODE_RCP;
 		tgt[0].DstReg.File = PROGRAM_TEMPORARY;
@@ -222,7 +222,7 @@ static GLboolean transform_TEX(
 			tgt[2].SrcReg[2].Swizzle = SWIZZLE_1111;
 		}
 	} else if (destredirect) {
-		tgt = radeonAppendInstructions(p, 1);
+		tgt = radeonAppendInstructions(t->Program, 1);
 
 		tgt->Opcode = OPCODE_MOV;
 		tgt->DstReg = orig_inst->DstReg;
diff --git a/src/mesa/drivers/dri/r300/r500_fragprog.c b/src/mesa/drivers/dri/r300/r500_fragprog.c
index b46e924ac7b..7ee84947225 100644
--- a/src/mesa/drivers/dri/r300/r500_fragprog.c
+++ b/src/mesa/drivers/dri/r300/r500_fragprog.c
@@ -38,7 +38,7 @@
  *
  */
 static GLboolean transform_TEX(
-	GLcontext *ctx, struct gl_program *p,
+	struct radeon_transform_context *t,
 	struct prog_instruction* orig_inst, void* data)
 {
 	struct r500_fragment_program_compiler *compiler =
@@ -55,11 +55,11 @@ static GLboolean transform_TEX(
 
 	/* ARB_shadow & EXT_shadow_funcs */
 	if (inst.Opcode != OPCODE_KIL &&
-	    p->ShadowSamplers & (1 << inst.TexSrcUnit)) {
+	    t->Program->ShadowSamplers & (1 << inst.TexSrcUnit)) {
 		GLuint comparefunc = GL_NEVER + compiler->fp->state.unit[inst.TexSrcUnit].texture_compare_func;
 
 		if (comparefunc == GL_NEVER || comparefunc == GL_ALWAYS) {
-			tgt = radeonAppendInstructions(p, 1);
+			tgt = radeonAppendInstructions(t->Program, 1);
 
 			tgt->Opcode = OPCODE_MOV;
 			tgt->DstReg.File = inst.DstReg.File;
@@ -71,20 +71,20 @@ static GLboolean transform_TEX(
 		}
 
 		inst.DstReg.File = PROGRAM_TEMPORARY;
-		inst.DstReg.Index = _mesa_find_free_register(p, PROGRAM_TEMPORARY);
+		inst.DstReg.Index = radeonFindFreeTemporary(t);
 		inst.DstReg.WriteMask = WRITEMASK_XYZW;
 	}
 
-	tgt = radeonAppendInstructions(p, 1);
+	tgt = radeonAppendInstructions(t->Program, 1);
 	_mesa_copy_instructions(tgt, &inst, 1);
 
 	if (inst.Opcode != OPCODE_KIL &&
-	    p->ShadowSamplers & (1 << inst.TexSrcUnit)) {
+	    t->Program->ShadowSamplers & (1 << inst.TexSrcUnit)) {
 		GLuint comparefunc = GL_NEVER + compiler->fp->state.unit[inst.TexSrcUnit].texture_compare_func;
 		GLuint depthmode = compiler->fp->state.unit[inst.TexSrcUnit].depth_texture_mode;
-		int rcptemp = _mesa_find_free_register(p, PROGRAM_TEMPORARY);
+		int rcptemp = radeonFindFreeTemporary(t);
 
-		tgt = radeonAppendInstructions(p, 3);
+		tgt = radeonAppendInstructions(t->Program, 3);
 
 		tgt[0].Opcode = OPCODE_RCP;
 		tgt[0].DstReg.File = PROGRAM_TEMPORARY;
@@ -131,7 +131,7 @@ static GLboolean transform_TEX(
 			tgt[2].SrcReg[2].Swizzle = SWIZZLE_1111;
 		}
 	} else if (destredirect) {
-		tgt = radeonAppendInstructions(p, 1);
+		tgt = radeonAppendInstructions(t->Program, 1);
 
 		tgt->Opcode = OPCODE_MOV;
 		tgt->DstReg = orig_inst->DstReg;
diff --git a/src/mesa/drivers/dri/r300/radeon_program.c b/src/mesa/drivers/dri/r300/radeon_program.c
index 3112339f81c..da5e7aefce5 100644
--- a/src/mesa/drivers/dri/r300/radeon_program.c
+++ b/src/mesa/drivers/dri/r300/radeon_program.c
@@ -46,29 +46,30 @@
  * one instruction at a time.
  */
 void radeonLocalTransform(
-	GLcontext *ctx,
+	GLcontext *Ctx,
 	struct gl_program *program,
 	int num_transformations,
 	struct radeon_program_transformation* transformations)
 {
-	struct prog_instruction *source;
-	int numinstructions;
+	struct radeon_transform_context ctx;
 	int ip;
 
-	source = program->Instructions;
-	numinstructions = program->NumInstructions;
+	ctx.Ctx = Ctx;
+	ctx.Program = program;
+	ctx.OldInstructions = program->Instructions;
+	ctx.OldNumInstructions = program->NumInstructions;
 
 	program->Instructions = 0;
 	program->NumInstructions = 0;
 
-	for(ip = 0; ip < numinstructions; ++ip) {
-		struct prog_instruction *instr = source + ip;
+	for(ip = 0; ip < ctx.OldNumInstructions; ++ip) {
+		struct prog_instruction *instr = ctx.OldInstructions + ip;
 		int i;
 
 		for(i = 0; i < num_transformations; ++i) {
 			struct radeon_program_transformation* t = transformations + i;
 
-			if (t->function(ctx, program, instr, t->userData))
+			if (t->function(&ctx, instr, t->userData))
 				break;
 		}
 
@@ -78,7 +79,40 @@ void radeonLocalTransform(
 		}
 	}
 
-	_mesa_free_instructions(source, numinstructions);
+	_mesa_free_instructions(ctx.OldInstructions, ctx.OldNumInstructions);
+}
+
+
+static void scan_instructions(GLboolean* used, const struct prog_instruction* insts, GLuint count)
+{
+	GLuint i;
+	for (i = 0; i < count; i++) {
+		const struct prog_instruction *inst = insts + i;
+		const GLuint n = _mesa_num_inst_src_regs(inst->Opcode);
+		GLuint k;
+
+		for (k = 0; k < n; k++) {
+			if (inst->SrcReg[k].File == PROGRAM_TEMPORARY)
+				used[inst->SrcReg[k].Index] = GL_TRUE;
+		}
+	}
+}
+
+GLint radeonFindFreeTemporary(struct radeon_transform_context *t)
+{
+	GLboolean used[MAX_PROGRAM_TEMPS];
+	GLuint i;
+
+	_mesa_memset(used, 0, sizeof(used));
+	scan_instructions(used, t->Program->Instructions, t->Program->NumInstructions);
+	scan_instructions(used, t->OldInstructions, t->OldNumInstructions);
+
+	for (i = 0; i < MAX_PROGRAM_TEMPS; i++) {
+		if (!used[i])
+			return i;
+	}
+
+	return -1;
 }
 
 
diff --git a/src/mesa/drivers/dri/r300/radeon_program.h b/src/mesa/drivers/dri/r300/radeon_program.h
index 012104fa5ae..ba76bc47cfb 100644
--- a/src/mesa/drivers/dri/r300/radeon_program.h
+++ b/src/mesa/drivers/dri/r300/radeon_program.h
@@ -49,6 +49,19 @@ enum {
 #define SWIZZLE_1111 MAKE_SWIZZLE4(SWIZZLE_ONE, SWIZZLE_ONE, SWIZZLE_ONE, SWIZZLE_ONE)
 
 /**
+ * Transformation context that is passed to local transformations.
+ *
+ * Care must be taken with some operations during transformation,
+ * e.g. finding new temporary registers must use @ref radeonFindFreeTemporary
+ */
+struct radeon_transform_context {
+	GLcontext *Ctx;
+	struct gl_program *Program;
+	struct prog_instruction *OldInstructions;
+	GLuint OldNumInstructions;
+};
+
+/**
  * A transformation that can be passed to \ref radeonLocalTransform.
  *
  * The function will be called once for each instruction.
@@ -60,8 +73,7 @@ enum {
  */
 struct radeon_program_transformation {
 	GLboolean (*function)(
-		GLcontext*,
-		struct gl_program*,
+		struct radeon_transform_context*,
 		struct prog_instruction*,
 		void*);
 	void *userData;
@@ -73,6 +85,10 @@ void radeonLocalTransform(
 	int num_transformations,
 	struct radeon_program_transformation* transformations);
 
+/**
+ * Find a usable free temporary register during program transformation
+ */
+GLint radeonFindFreeTemporary(struct radeon_transform_context *ctx);
 
 struct prog_instruction *radeonAppendInstructions(struct gl_program *program, int count);
 
diff --git a/src/mesa/drivers/dri/r300/radeon_program_alu.c b/src/mesa/drivers/dri/r300/radeon_program_alu.c
index 3104d07facd..d6d016d7c12 100644
--- a/src/mesa/drivers/dri/r300/radeon_program_alu.c
+++ b/src/mesa/drivers/dri/r300/radeon_program_alu.c
@@ -148,24 +148,24 @@ static struct prog_src_register scalar(struct prog_src_register reg)
 	return swizzle(reg, SWIZZLE_X, SWIZZLE_X, SWIZZLE_X, SWIZZLE_X);
 }
 
-static void transform_ABS(struct gl_program* p,
+static void transform_ABS(struct radeon_transform_context* t,
 	struct prog_instruction* inst)
 {
 	struct prog_src_register src = inst->SrcReg[0];
 	src.Abs = 1;
 	src.NegateBase = 0;
 	src.NegateAbs = 0;
-	emit1(p, OPCODE_MOV, inst->DstReg, src);
+	emit1(t->Program, OPCODE_MOV, inst->DstReg, src);
 }
 
-static void transform_DPH(struct gl_program* p,
+static void transform_DPH(struct radeon_transform_context* t,
 	struct prog_instruction* inst)
 {
 	struct prog_src_register src0 = inst->SrcReg[0];
 	if (src0.NegateAbs) {
 		if (src0.Abs) {
-			int tempreg = _mesa_find_free_register(p, PROGRAM_TEMPORARY);
-			emit1(p, OPCODE_MOV, dstreg(PROGRAM_TEMPORARY, tempreg), src0);
+			int tempreg = radeonFindFreeTemporary(t);
+			emit1(t->Program, OPCODE_MOV, dstreg(PROGRAM_TEMPORARY, tempreg), src0);
 			src0 = srcreg(src0.File, src0.Index);
 		} else {
 			src0.NegateAbs = 0;
@@ -174,70 +174,70 @@ static void transform_DPH(struct gl_program* p,
 	}
 	set_swizzle(&src0, 3, SWIZZLE_ONE);
 	set_negate_base(&src0, 3, 0);
-	emit2(p, OPCODE_DP4, inst->DstReg, src0, inst->SrcReg[1]);
+	emit2(t->Program, OPCODE_DP4, inst->DstReg, src0, inst->SrcReg[1]);
 }
 
-static void transform_FLR(struct gl_program* p,
+static void transform_FLR(struct radeon_transform_context* t,
 	struct prog_instruction* inst)
 {
-	int tempreg = _mesa_find_free_register(p, PROGRAM_TEMPORARY);
-	emit1(p, OPCODE_FRC, dstreg(PROGRAM_TEMPORARY, tempreg), inst->SrcReg[0]);
-	emit2(p, OPCODE_ADD, inst->DstReg, inst->SrcReg[0], negate(srcreg(PROGRAM_TEMPORARY, tempreg)));
+	int tempreg = radeonFindFreeTemporary(t);
+	emit1(t->Program, OPCODE_FRC, dstreg(PROGRAM_TEMPORARY, tempreg), inst->SrcReg[0]);
+	emit2(t->Program, OPCODE_ADD, inst->DstReg, inst->SrcReg[0], negate(srcreg(PROGRAM_TEMPORARY, tempreg)));
 }
 
-static void transform_POW(struct gl_program* p,
+static void transform_POW(struct radeon_transform_context* t,
 	struct prog_instruction* inst)
 {
-	int tempreg = _mesa_find_free_register(p, PROGRAM_TEMPORARY);
+	int tempreg = radeonFindFreeTemporary(t);
 	struct prog_dst_register tempdst = dstreg(PROGRAM_TEMPORARY, tempreg);
 	struct prog_src_register tempsrc = srcreg(PROGRAM_TEMPORARY, tempreg);
 	tempdst.WriteMask = WRITEMASK_W;
 	tempsrc.Swizzle = SWIZZLE_WWWW;
 
-	emit1(p, OPCODE_LG2, tempdst, scalar(inst->SrcReg[0]));
-	emit2(p, OPCODE_MUL, tempdst, tempsrc, scalar(inst->SrcReg[1]));
-	emit1(p, OPCODE_EX2, inst->DstReg, tempsrc);
+	emit1(t->Program, OPCODE_LG2, tempdst, scalar(inst->SrcReg[0]));
+	emit2(t->Program, OPCODE_MUL, tempdst, tempsrc, scalar(inst->SrcReg[1]));
+	emit1(t->Program, OPCODE_EX2, inst->DstReg, tempsrc);
 }
 
-static void transform_SGE(struct gl_program* p,
+static void transform_SGE(struct radeon_transform_context* t,
 	struct prog_instruction* inst)
 {
-	int tempreg = _mesa_find_free_register(p, PROGRAM_TEMPORARY);
+	int tempreg = radeonFindFreeTemporary(t);
 
-	emit2(p, OPCODE_ADD, dstreg(PROGRAM_TEMPORARY, tempreg), inst->SrcReg[0], negate(inst->SrcReg[1]));
-	emit3(p, OPCODE_CMP, inst->DstReg, srcreg(PROGRAM_TEMPORARY, tempreg), builtin_zero, builtin_one);
+	emit2(t->Program, OPCODE_ADD, dstreg(PROGRAM_TEMPORARY, tempreg), inst->SrcReg[0], negate(inst->SrcReg[1]));
+	emit3(t->Program, OPCODE_CMP, inst->DstReg, srcreg(PROGRAM_TEMPORARY, tempreg), builtin_zero, builtin_one);
 }
 
-static void transform_SLT(struct gl_program* p,
+static void transform_SLT(struct radeon_transform_context* t,
 	struct prog_instruction* inst)
 {
-	int tempreg = _mesa_find_free_register(p, PROGRAM_TEMPORARY);
+	int tempreg = radeonFindFreeTemporary(t);
 
-	emit2(p, OPCODE_ADD, dstreg(PROGRAM_TEMPORARY, tempreg), inst->SrcReg[0], negate(inst->SrcReg[1]));
-	emit3(p, OPCODE_CMP, inst->DstReg, srcreg(PROGRAM_TEMPORARY, tempreg), builtin_one, builtin_zero);
+	emit2(t->Program, OPCODE_ADD, dstreg(PROGRAM_TEMPORARY, tempreg), inst->SrcReg[0], negate(inst->SrcReg[1]));
+	emit3(t->Program, OPCODE_CMP, inst->DstReg, srcreg(PROGRAM_TEMPORARY, tempreg), builtin_one, builtin_zero);
 }
 
-static void transform_SUB(struct gl_program* p,
+static void transform_SUB(struct radeon_transform_context* t,
 	struct prog_instruction* inst)
 {
-	emit2(p, OPCODE_ADD, inst->DstReg, inst->SrcReg[0], negate(inst->SrcReg[1]));
+	emit2(t->Program, OPCODE_ADD, inst->DstReg, inst->SrcReg[0], negate(inst->SrcReg[1]));
 }
 
-static void transform_SWZ(struct gl_program* p,
+static void transform_SWZ(struct radeon_transform_context* t,
 	struct prog_instruction* inst)
 {
-	emit1(p, OPCODE_MOV, inst->DstReg, inst->SrcReg[0]);
+	emit1(t->Program, OPCODE_MOV, inst->DstReg, inst->SrcReg[0]);
 }
 
-static void transform_XPD(struct gl_program* p,
+static void transform_XPD(struct radeon_transform_context* t,
 	struct prog_instruction* inst)
 {
-	int tempreg = _mesa_find_free_register(p, PROGRAM_TEMPORARY);
+	int tempreg = radeonFindFreeTemporary(t);
 
-	emit2(p, OPCODE_MUL, dstreg(PROGRAM_TEMPORARY, tempreg),
+	emit2(t->Program, OPCODE_MUL, dstreg(PROGRAM_TEMPORARY, tempreg),
 		swizzle(inst->SrcReg[0], SWIZZLE_Z, SWIZZLE_X, SWIZZLE_Y, SWIZZLE_W),
 		swizzle(inst->SrcReg[1], SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_X, SWIZZLE_W));
-	emit3(p, OPCODE_MAD, inst->DstReg,
+	emit3(t->Program, OPCODE_MAD, inst->DstReg,
 		swizzle(inst->SrcReg[0], SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_X, SWIZZLE_W),
 		swizzle(inst->SrcReg[1], SWIZZLE_Z, SWIZZLE_X, SWIZZLE_Y, SWIZZLE_W),
 		negate(srcreg(PROGRAM_TEMPORARY, tempreg)));
@@ -257,22 +257,20 @@ static void transform_XPD(struct gl_program* p,
  *
  * @todo add LIT here as well?
  */
-GLboolean radeonTransformALU(
-	GLcontext* ctx,
-	struct gl_program* prog,
+GLboolean radeonTransformALU(struct radeon_transform_context* t,
 	struct prog_instruction* inst,
 	void* unused)
 {
 	switch(inst->Opcode) {
-	case OPCODE_ABS: transform_ABS(prog, inst); return GL_TRUE;
-	case OPCODE_DPH: transform_DPH(prog, inst); return GL_TRUE;
-	case OPCODE_FLR: transform_FLR(prog, inst); return GL_TRUE;
-	case OPCODE_POW: transform_POW(prog, inst); return GL_TRUE;
-	case OPCODE_SGE: transform_SGE(prog, inst); return GL_TRUE;
-	case OPCODE_SLT: transform_SLT(prog, inst); return GL_TRUE;
-	case OPCODE_SUB: transform_SUB(prog, inst); return GL_TRUE;
-	case OPCODE_SWZ: transform_SWZ(prog, inst); return GL_TRUE;
-	case OPCODE_XPD: transform_XPD(prog, inst); return GL_TRUE;
+	case OPCODE_ABS: transform_ABS(t, inst); return GL_TRUE;
+	case OPCODE_DPH: transform_DPH(t, inst); return GL_TRUE;
+	case OPCODE_FLR: transform_FLR(t, inst); return GL_TRUE;
+	case OPCODE_POW: transform_POW(t, inst); return GL_TRUE;
+	case OPCODE_SGE: transform_SGE(t, inst); return GL_TRUE;
+	case OPCODE_SLT: transform_SLT(t, inst); return GL_TRUE;
+	case OPCODE_SUB: transform_SUB(t, inst); return GL_TRUE;
+	case OPCODE_SWZ: transform_SWZ(t, inst); return GL_TRUE;
+	case OPCODE_XPD: transform_XPD(t, inst); return GL_TRUE;
 	default:
 		return GL_FALSE;
 	}
diff --git a/src/mesa/drivers/dri/r300/radeon_program_alu.h b/src/mesa/drivers/dri/r300/radeon_program_alu.h
index f5beb9f8c30..858c5ed0b8c 100644
--- a/src/mesa/drivers/dri/r300/radeon_program_alu.h
+++ b/src/mesa/drivers/dri/r300/radeon_program_alu.h
@@ -31,8 +31,7 @@
 #include "radeon_program.h"
 
 GLboolean radeonTransformALU(
-	GLcontext*,
-	struct gl_program*,
+	struct radeon_transform_context *t,
 	struct prog_instruction*,
 	void*);