r300: always emit output insts after all KIL insts

2 files changed, 46 insertions, 3 deletions

diff --git a/src/mesa/drivers/dri/r300/r300_state.c b/src/mesa/drivers/dri/r300/r300_state.c
index 8095538ff9..6b79aa4313 100644
--- a/src/mesa/drivers/dri/r300/r300_state.c
+++ b/src/mesa/drivers/dri/r300/r300_state.c
@@ -472,7 +472,9 @@ static void r300SetEarlyZState(GLcontext * ctx)
 
 	if (ctx->Color.AlphaEnabled && ctx->Color.AlphaFunc != GL_ALWAYS)
 		topZ = R300_ZTOP_DISABLE;
-	if (current_fragment_program_writes_depth(ctx))
+	else if (current_fragment_program_writes_depth(ctx))
+		topZ = R300_ZTOP_DISABLE;
+	else if (ctx->FragmentProgram._Current && ctx->FragmentProgram._Current->UsesKill)
 		topZ = R300_ZTOP_DISABLE;
 
 	if (topZ != r300->hw.zstencil_format.cmd[2]) {
diff --git a/src/mesa/drivers/dri/r300/radeon_program_pair.c b/src/mesa/drivers/dri/r300/radeon_program_pair.c
index 4aa2319a45..2e21f7bf66 100644
--- a/src/mesa/drivers/dri/r300/radeon_program_pair.c
+++ b/src/mesa/drivers/dri/r300/radeon_program_pair.c
@@ -47,6 +47,7 @@
 
 struct pair_state_instruction {
 	GLuint IsTex:1; /**< Is a texture instruction */
+	GLuint IsOutput:1; /**< Is output instruction */
 	GLuint NeedRGB:1; /**< Needs the RGB ALU */
 	GLuint NeedAlpha:1; /**< Needs the Alpha ALU */
 	GLuint IsTranscendent:1; /**< Is a special transcendent instruction */
@@ -123,6 +124,7 @@ struct pair_state {
 	GLboolean Debug;
 	GLboolean Verbose;
 	void *UserData;
+	GLubyte NumKillInsts;
 
 	/**
 	 * Translate Mesa registers to hardware registers
@@ -149,6 +151,11 @@ struct pair_state {
 	struct pair_state_instruction *ReadyTEX;
 
 	/**
+	 * Linked list of deferred instructions
+	 */
+	struct pair_state_instruction *DeferredInsts;
+
+	/**
 	 * Pool of @ref reg_value structures for fast allocation.
 	 */
 	struct reg_value *ValuePool;
@@ -231,7 +238,9 @@ static void instruction_ready(struct pair_state *s, int ip)
 	if (s->Verbose)
 		_mesa_printf("instruction_ready(%i)\n", ip);
 
-	if (pairinst->IsTex)
+	if (s->NumKillInsts > 0 && pairinst->IsOutput)
+		add_pairinst_to_list(&s->DeferredInsts, pairinst);
+	else if (pairinst->IsTex)
 		add_pairinst_to_list(&s->ReadyTEX, pairinst);
 	else if (!pairinst->NeedAlpha)
 		add_pairinst_to_list(&s->ReadyRGB, pairinst);
@@ -339,6 +348,8 @@ static void classify_instruction(struct pair_state *s,
 		error("Unknown opcode %d\n", inst->Opcode);
 		break;
 	}
+
+	pairinst->IsOutput = (inst->DstReg.File == PROGRAM_OUTPUT);
 }
 
 
@@ -602,8 +613,11 @@ static void emit_all_tex(struct pair_state *s)
 		struct prog_instruction *inst = s->Program->Instructions + ip;
 		commit_instruction(s, ip);
 
-		if (inst->Opcode != OPCODE_KIL)
+		if (inst->Opcode == OPCODE_KIL)
+			--s->NumKillInsts;
+		else
 			inst->DstReg.Index = get_hw_reg(s, inst->DstReg.File, inst->DstReg.Index);
+
 		inst->SrcReg[0].Index = get_hw_reg(s, inst->SrcReg[0].File, inst->SrcReg[0].Index);
 
 		if (s->Debug) {
@@ -861,6 +875,17 @@ static void emit_alu(struct pair_state *s)
 	s->Error = s->Error || !s->Handler->EmitPaired(s->UserData, &pair);
 }
 
+static GLubyte countKillInsts(struct gl_program *prog)
+{
+	GLubyte i, count = 0;
+
+	for (i = 0; i < prog->NumInstructions; ++i) {
+		if (prog->Instructions[i].Opcode == OPCODE_KIL)
+			++count;
+	}
+
+	return count;
+}
 
 GLboolean radeonPairProgram(GLcontext *ctx, struct gl_program *program,
 	const struct radeon_pair_handler* handler, void *userdata)
@@ -874,6 +899,7 @@ GLboolean radeonPairProgram(GLcontext *ctx, struct gl_program *program,
 	s.UserData = userdata;
 	s.Debug = (RADEON_DEBUG & DEBUG_PIXEL) ? GL_TRUE : GL_FALSE;
 	s.Verbose = GL_FALSE && s.Debug;
+	s.NumKillInsts = countKillInsts(program);
 
 	s.Instructions = (struct pair_state_instruction*)_mesa_calloc(
 		sizeof(struct pair_state_instruction)*s.Program->NumInstructions);
@@ -892,6 +918,21 @@ GLboolean radeonPairProgram(GLcontext *ctx, struct gl_program *program,
 		if (s.ReadyTEX)
 			emit_all_tex(&s);
 
+		if (!s.NumKillInsts) {
+			struct pair_state_instruction *pairinst = s.DeferredInsts;
+			while (pairinst) {
+				if (!pairinst->NeedAlpha)
+					add_pairinst_to_list(&s.ReadyRGB, pairinst);
+				else if (!pairinst->NeedRGB)
+					add_pairinst_to_list(&s.ReadyAlpha, pairinst);
+				else
+					add_pairinst_to_list(&s.ReadyFullALU, pairinst);
+
+				pairinst = pairinst->NextReady;
+			}
+			s.DeferredInsts = NULL;
+		}
+
 		while(s.ReadyFullALU || s.ReadyRGB || s.ReadyAlpha)
 			emit_alu(&s);
 	}