From 28066ed012b51f2171866669a2972bc7ee293565 Mon Sep 17 00:00:00 2001
From: Maciej Cencora <m.cencora@gmail.com>
Date: Sat, 4 Jul 2009 12:56:31 +0200
Subject: r300: update state parameters only once per rendering operation

---
 src/mesa/drivers/dri/r300/r300_state.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

(limited to 'src/mesa/drivers/dri/r300/r300_state.c')

diff --git a/src/mesa/drivers/dri/r300/r300_state.c b/src/mesa/drivers/dri/r300/r300_state.c
index c0eda977db..67bd6ec33b 100644
--- a/src/mesa/drivers/dri/r300/r300_state.c
+++ b/src/mesa/drivers/dri/r300/r300_state.c
@@ -1093,7 +1093,7 @@ r300FetchStateParameter(GLcontext * ctx,
  * Update R300's own internal state parameters.
  * For now just STATE_R300_WINDOW_DIMENSION
  */
-void r300UpdateStateParameters(GLcontext * ctx, GLuint new_state)
+static void r300UpdateStateParameters(GLcontext * ctx, GLuint new_state)
 {
 	struct r300_fragment_program *fp;
 	struct gl_program_parameter_list *paramList;
@@ -1111,6 +1111,8 @@ void r300UpdateStateParameters(GLcontext * ctx, GLuint new_state)
 	if (!paramList)
 		return;
 
+	_mesa_load_state_parameters(ctx, paramList);
+
 	for (i = 0; i < paramList->NumParameters; i++) {
 		if (paramList->Parameters[i].Type == PROGRAM_STATE_VAR) {
 			r300FetchStateParameter(ctx,
@@ -2323,8 +2325,6 @@ static void r300InvalidateState(GLcontext * ctx, GLuint new_state)
 		R300_STATECHANGE(r300, cb);
 	}
 
-	r300UpdateStateParameters(ctx, new_state);
-
 	r300->radeon.NewGLState |= new_state;
 }
 
-- 
cgit v1.2.3


From 7829b7a1b85dd8e6c31189e7f3dec91d71d134c3 Mon Sep 17 00:00:00 2001
From: Maciej Cencora <m.cencora@gmail.com>
Date: Sat, 4 Jul 2009 13:26:49 +0200
Subject: r300: cache translated fragment programs

---
 src/mesa/drivers/dri/r300/r300_context.h         |  10 +-
 src/mesa/drivers/dri/r300/r300_fragprog_common.c | 169 ++++++++++++-----------
 src/mesa/drivers/dri/r300/r300_fragprog_common.h |   6 +-
 src/mesa/drivers/dri/r300/r300_shader.c          |  51 ++++++-
 src/mesa/drivers/dri/r300/r300_state.c           |  59 ++++----
 src/mesa/drivers/dri/r300/r300_swtcl.c           |   2 +-
 6 files changed, 173 insertions(+), 124 deletions(-)

(limited to 'src/mesa/drivers/dri/r300/r300_state.c')

diff --git a/src/mesa/drivers/dri/r300/r300_context.h b/src/mesa/drivers/dri/r300/r300_context.h
index 026c33c67c..fc436e1100 100644
--- a/src/mesa/drivers/dri/r300/r300_context.h
+++ b/src/mesa/drivers/dri/r300/r300_context.h
@@ -546,7 +546,7 @@ struct r500_fragment_program_code {
 * to render with that program.
 */
 struct r300_fragment_program {
-	struct gl_fragment_program Base;
+	struct gl_program *Base;
 
 	GLboolean translated;
 	GLboolean error;
@@ -559,6 +559,13 @@ struct r300_fragment_program {
 
 	GLboolean writes_depth;
 	GLuint optimization;
+
+	struct r300_fragment_program *next;
+};
+
+struct r300_fragment_program_cont {
+	struct gl_fragment_program Base;
+	struct r300_fragment_program *progs;
 };
 
 struct r300_fragment_program_compiler {
@@ -633,6 +640,7 @@ struct r300_context {
 	struct r300_hw_state hw;
 
 	struct r300_vertex_program *selected_vp;
+	struct r300_fragment_program *selected_fp;
 
 	/* Vertex buffers
 	 */
diff --git a/src/mesa/drivers/dri/r300/r300_fragprog_common.c b/src/mesa/drivers/dri/r300/r300_fragprog_common.c
index 1644e6f42a..b25cf24007 100644
--- a/src/mesa/drivers/dri/r300/r300_fragprog_common.c
+++ b/src/mesa/drivers/dri/r300/r300_fragprog_common.c
@@ -67,7 +67,7 @@ static void nqssadce_init(struct nqssadce_state* s)
  */
 static void insert_WPOS_trailer(struct r300_fragment_program_compiler *compiler)
 {
-	GLuint InputsRead = compiler->fp->Base.Base.InputsRead;
+	GLuint InputsRead = compiler->fp->Base->InputsRead;
 
 	if (!(InputsRead & FRAG_BIT_WPOS))
 		return;
@@ -168,7 +168,7 @@ static GLuint build_func(GLuint comparefunc)
  */
 static void build_state(
 	r300ContextPtr r300,
-	struct r300_fragment_program *fp,
+	struct gl_fragment_program *fp,
 	struct r300_fragment_program_external_state *state)
 {
 	int unit;
@@ -176,7 +176,7 @@ static void build_state(
 	_mesa_bzero(state, sizeof(*state));
 
 	for(unit = 0; unit < 16; ++unit) {
-		if (fp->Base.Base.ShadowSamplers & (1 << unit)) {
+		if (fp->Base.ShadowSamplers & (1 << unit)) {
 			struct gl_texture_object* tex = r300->radeon.glCtx->Texture.Unit[unit]._Current;
 
 			state->unit[unit].depth_texture_mode = build_dtm(tex->DepthMode);
@@ -185,96 +185,107 @@ static void build_state(
 	}
 }
 
-void r300TranslateFragmentShader(GLcontext *ctx, struct gl_fragment_program *fp)
+void r300TranslateFragmentShader(GLcontext *ctx, struct r300_fragment_program *fp)
 {
 	r300ContextPtr r300 = R300_CONTEXT(ctx);
-	struct r300_fragment_program *r300_fp = (struct r300_fragment_program *)fp;
-	struct r300_fragment_program_external_state state;
+	struct r300_fragment_program_compiler compiler;
+
+	compiler.r300 = r300;
+	compiler.fp = fp;
+	compiler.code = &fp->code;
+	compiler.program = fp->Base;
+
+	if (RADEON_DEBUG & DEBUG_PIXEL) {
+		fflush(stdout);
+		_mesa_printf("Fragment Program: Initial program:\n");
+		_mesa_print_program(compiler.program);
+		fflush(stdout);
+	}
 
-	build_state(r300, r300_fp, &state);
-	if (_mesa_memcmp(&r300_fp->state, &state, sizeof(state))) {
-		/* TODO: cache compiled programs */
-		r300_fp->translated = GL_FALSE;
-		_mesa_memcpy(&r300_fp->state, &state, sizeof(state));
+	insert_WPOS_trailer(&compiler);
+
+	if (r300->radeon.radeonScreen->chip_family >= CHIP_FAMILY_RV515) {
+		struct radeon_program_transformation transformations[] = {
+			{ &r500_transform_TEX, &compiler },
+			{ &radeonTransformALU, 0 },
+			{ &radeonTransformDeriv, 0 },
+			{ &radeonTransformTrigScale, 0 }
+		};
+		radeonLocalTransform(ctx, compiler.program, 4, transformations);
+	} else {
+		struct radeon_program_transformation transformations[] = {
+			{ &r300_transform_TEX, &compiler },
+			{ &radeonTransformALU, 0 },
+			{ &radeonTransformTrigSimple, 0 }
+		};
+		radeonLocalTransform(ctx, compiler.program, 3, transformations);
 	}
 
-	if (!r300_fp->translated) {
-		struct r300_fragment_program_compiler compiler;
+	if (RADEON_DEBUG & DEBUG_PIXEL) {
+		_mesa_printf("Fragment Program: After native rewrite:\n");
+		_mesa_print_program(compiler.program);
+		fflush(stdout);
+	}
 
-		compiler.r300 = r300;
-		compiler.fp = r300_fp;
-		compiler.code = &r300_fp->code;
-		compiler.program = _mesa_clone_program(ctx, &fp->Base);
+	if (r300->radeon.radeonScreen->chip_family >= CHIP_FAMILY_RV515) {
+		struct radeon_nqssadce_descr nqssadce = {
+			.Init = &nqssadce_init,
+			.IsNativeSwizzle = &r500FPIsNativeSwizzle,
+			.BuildSwizzle = &r500FPBuildSwizzle,
+			.RewriteDepthOut = GL_TRUE
+		};
+		radeonNqssaDce(ctx, compiler.program, &nqssadce);
+	} else {
+		struct radeon_nqssadce_descr nqssadce = {
+			.Init = &nqssadce_init,
+			.IsNativeSwizzle = &r300FPIsNativeSwizzle,
+			.BuildSwizzle = &r300FPBuildSwizzle,
+			.RewriteDepthOut = GL_TRUE
+		};
+		radeonNqssaDce(ctx, compiler.program, &nqssadce);
+	}
 
-		if (RADEON_DEBUG & DEBUG_PIXEL) {
-			fflush(stdout);
-			_mesa_printf("Fragment Program: Initial program:\n");
-			_mesa_print_program(compiler.program);
-			fflush(stdout);
-		}
+	if (RADEON_DEBUG & DEBUG_PIXEL) {
+		_mesa_printf("Compiler: after NqSSA-DCE:\n");
+		_mesa_print_program(compiler.program);
+		fflush(stdout);
+	}
 
-		insert_WPOS_trailer(&compiler);
-
-		if (r300->radeon.radeonScreen->chip_family >= CHIP_FAMILY_RV515) {
-			struct radeon_program_transformation transformations[] = {
-				{ &r500_transform_TEX, &compiler },
-				{ &radeonTransformALU, 0 },
-				{ &radeonTransformDeriv, 0 },
-				{ &radeonTransformTrigScale, 0 }
-			};
-			radeonLocalTransform(ctx, compiler.program, 4, transformations);
-		} else {
-			struct radeon_program_transformation transformations[] = {
-				{ &r300_transform_TEX, &compiler },
-				{ &radeonTransformALU, 0 },
-				{ &radeonTransformTrigSimple, 0 }
-			};
-			radeonLocalTransform(ctx, compiler.program, 3, transformations);
-		}
+	if (!r300->vtbl.BuildFragmentProgramHwCode(&compiler))
+		fp->error = GL_TRUE;
 
-		if (RADEON_DEBUG & DEBUG_PIXEL) {
-			_mesa_printf("Fragment Program: After native rewrite:\n");
-			_mesa_print_program(compiler.program);
-			fflush(stdout);
-		}
+	fp->translated = GL_TRUE;
 
-		if (r300->radeon.radeonScreen->chip_family >= CHIP_FAMILY_RV515) {
-			struct radeon_nqssadce_descr nqssadce = {
-				.Init = &nqssadce_init,
-				.IsNativeSwizzle = &r500FPIsNativeSwizzle,
-				.BuildSwizzle = &r500FPBuildSwizzle,
-				.RewriteDepthOut = GL_TRUE
-			};
-			radeonNqssaDce(ctx, compiler.program, &nqssadce);
-		} else {
-			struct radeon_nqssadce_descr nqssadce = {
-				.Init = &nqssadce_init,
-				.IsNativeSwizzle = &r300FPIsNativeSwizzle,
-				.BuildSwizzle = &r300FPBuildSwizzle,
-				.RewriteDepthOut = GL_TRUE
-			};
-			radeonNqssaDce(ctx, compiler.program, &nqssadce);
-		}
+	if (fp->error || (RADEON_DEBUG & DEBUG_PIXEL))
+		r300->vtbl.FragmentProgramDump(&fp->code);
+}
 
-		if (RADEON_DEBUG & DEBUG_PIXEL) {
-			_mesa_printf("Compiler: after NqSSA-DCE:\n");
-			_mesa_print_program(compiler.program);
-			fflush(stdout);
-		}
+struct r300_fragment_program *r300SelectFragmentShader(GLcontext *ctx)
+{
+	r300ContextPtr r300 = R300_CONTEXT(ctx);
+	struct r300_fragment_program_cont *fp_list;
+	struct r300_fragment_program *fp;
+	struct r300_fragment_program_external_state state;
+
+	fp_list = (struct r300_fragment_program_cont *)ctx->FragmentProgram._Current;
+	build_state(r300, ctx->FragmentProgram._Current, &state);
 
-		if (!r300->vtbl.BuildFragmentProgramHwCode(&compiler))
-			r300_fp->error = GL_TRUE;
+	fp = fp_list->progs;
+	while (fp) {
+	       if (_mesa_memcmp(&fp->state, &state, sizeof(state)) == 0) {
+	               return r300->selected_fp = fp;
+	       }
+	       fp = fp->next;
+	}
 
-		/* Subtle: Rescue any parameters that have been added during transformations */
-		_mesa_free_parameter_list(fp->Base.Parameters);
-		fp->Base.Parameters = compiler.program->Parameters;
-		compiler.program->Parameters = 0;
+	fp = _mesa_calloc(sizeof(struct r300_fragment_program));
 
-		_mesa_reference_program(ctx, &compiler.program, NULL);
+	fp->state = state;
+	fp->translated = GL_FALSE;
+	fp->Base = _mesa_clone_program(ctx, &ctx->FragmentProgram._Current->Base);
 
-		r300_fp->translated = GL_TRUE;
+	fp->next = fp_list->progs;
+	fp_list->progs = fp;
 
-		if (r300_fp->error || (RADEON_DEBUG & DEBUG_PIXEL))
-			r300->vtbl.FragmentProgramDump(&r300_fp->code);
-	}
+	return r300->selected_fp = fp;
 }
diff --git a/src/mesa/drivers/dri/r300/r300_fragprog_common.h b/src/mesa/drivers/dri/r300/r300_fragprog_common.h
index 85ea86fecb..5e103ee408 100644
--- a/src/mesa/drivers/dri/r300/r300_fragprog_common.h
+++ b/src/mesa/drivers/dri/r300/r300_fragprog_common.h
@@ -30,6 +30,10 @@
 
 #include "main/mtypes.h"
 
-extern void r300TranslateFragmentShader(GLcontext *ctx, struct gl_fragment_program *fp);
+#include "r300_context.h"
+
+extern void r300TranslateFragmentShader(GLcontext *ctx, struct r300_fragment_program *fp);
+
+struct r300_fragment_program *r300SelectFragmentShader(GLcontext *ctx);
 
 #endif
diff --git a/src/mesa/drivers/dri/r300/r300_shader.c b/src/mesa/drivers/dri/r300/r300_shader.c
index 0133b83796..c9a429943a 100644
--- a/src/mesa/drivers/dri/r300/r300_shader.c
+++ b/src/mesa/drivers/dri/r300/r300_shader.c
@@ -32,11 +32,34 @@
 #include "r300_context.h"
 #include "r300_fragprog_common.h"
 
+static void freeFragProgCache(GLcontext *ctx, struct r300_fragment_program_cont *cache)
+{
+	struct r300_fragment_program *tmp, *fp = cache->progs;
+
+	while (fp) {
+		tmp = fp->next;
+		_mesa_reference_program(ctx, &fp->Base, NULL);
+		_mesa_free(fp);
+		fp = tmp;
+	}
+}
+
+static void freeVertProgCache(GLcontext *ctx, struct r300_vertex_program_cont *cache)
+{
+	struct r300_vertex_program *tmp, *vp = cache->progs;
+
+	while (vp) {
+		tmp = vp->next;
+		_mesa_free(vp);
+		vp = tmp;
+	}
+}
+
 static struct gl_program *r300NewProgram(GLcontext * ctx, GLenum target,
 					 GLuint id)
 {
 	struct r300_vertex_program_cont *vp;
-	struct r300_fragment_program *fp;
+	struct r300_fragment_program_cont *fp;
 
 	switch (target) {
 	case GL_VERTEX_STATE_PROGRAM_NV:
@@ -47,7 +70,7 @@ static struct gl_program *r300NewProgram(GLcontext * ctx, GLenum target,
 
 	case GL_FRAGMENT_PROGRAM_NV:
 	case GL_FRAGMENT_PROGRAM_ARB:
-		fp = CALLOC_STRUCT(r300_fragment_program);
+		fp = CALLOC_STRUCT(r300_fragment_program_cont);
 		return _mesa_init_fragment_program(ctx, &fp->Base, target, id);
 
 	default:
@@ -59,21 +82,35 @@ static struct gl_program *r300NewProgram(GLcontext * ctx, GLenum target,
 
 static void r300DeleteProgram(GLcontext * ctx, struct gl_program *prog)
 {
+	struct r300_vertex_program_cont *vp = (struct r300_vertex_program_cont *)prog;
+	struct r300_fragment_program_cont *fp = (struct r300_fragment_program_cont *)prog;
+
+	switch (prog->Target) {
+		case GL_VERTEX_PROGRAM_ARB:
+			freeVertProgCache(ctx, vp);
+			break;
+		case GL_FRAGMENT_PROGRAM_ARB:
+			freeFragProgCache(ctx, fp);
+			break;
+	}
+
 	_mesa_delete_program(ctx, prog);
 }
 
 static void
 r300ProgramStringNotify(GLcontext * ctx, GLenum target, struct gl_program *prog)
 {
-	struct r300_vertex_program_cont *vp = (void *)prog;
-	struct r300_fragment_program *r300_fp = (struct r300_fragment_program *)prog;
+	struct r300_vertex_program_cont *vp = (struct r300_vertex_program_cont *)prog;
+	struct r300_fragment_program_cont *fp = (struct r300_fragment_program_cont *)prog;
 
 	switch (target) {
 	case GL_VERTEX_PROGRAM_ARB:
+		freeVertProgCache(ctx, vp);
 		vp->progs = NULL;
 		break;
 	case GL_FRAGMENT_PROGRAM_ARB:
-		r300_fp->translated = GL_FALSE;
+		freeFragProgCache(ctx, fp);
+		fp->progs = NULL;
 		break;
 	}
 
@@ -85,9 +122,9 @@ static GLboolean
 r300IsProgramNative(GLcontext * ctx, GLenum target, struct gl_program *prog)
 {
 	if (target == GL_FRAGMENT_PROGRAM_ARB) {
-		struct r300_fragment_program *fp = (struct r300_fragment_program *)prog;
+		struct r300_fragment_program *fp = r300SelectFragmentShader(ctx);
 		if (!fp->translated)
-			r300TranslateFragmentShader(ctx, &fp->Base);
+			r300TranslateFragmentShader(ctx, fp);
 
 		return !fp->error;
 	} else
diff --git a/src/mesa/drivers/dri/r300/r300_state.c b/src/mesa/drivers/dri/r300/r300_state.c
index 67bd6ec33b..e0996383eb 100644
--- a/src/mesa/drivers/dri/r300/r300_state.c
+++ b/src/mesa/drivers/dri/r300/r300_state.c
@@ -452,9 +452,9 @@ static void r300SetPolygonOffsetState(GLcontext * ctx, GLboolean state)
 
 static GLboolean current_fragment_program_writes_depth(GLcontext* ctx)
 {
-	struct r300_fragment_program *fp = (struct r300_fragment_program *) ctx->FragmentProgram._Current;
+	r300ContextPtr r300 = R300_CONTEXT(ctx);
 
-	return (fp && fp->writes_depth);
+	return ctx->FragmentProgram._Current && r300->selected_fp->writes_depth;
 }
 
 static void r300SetEarlyZState(GLcontext * ctx)
@@ -1095,18 +1095,17 @@ r300FetchStateParameter(GLcontext * ctx,
  */
 static void r300UpdateStateParameters(GLcontext * ctx, GLuint new_state)
 {
-	struct r300_fragment_program *fp;
+	r300ContextPtr rmesa = R300_CONTEXT(ctx);
 	struct gl_program_parameter_list *paramList;
 	GLuint i;
 
 	if (!(new_state & (_NEW_BUFFERS | _NEW_PROGRAM | _NEW_PROGRAM_CONSTANTS)))
 		return;
 
-	fp = (struct r300_fragment_program *)ctx->FragmentProgram._Current;
-	if (!fp)
+	if (!ctx->FragmentProgram._Current || !rmesa->selected_fp)
 		return;
 
-	paramList = fp->Base.Base.Parameters;
+	paramList = rmesa->selected_fp->Base->Parameters;
 
 	if (!paramList)
 		return;
@@ -1227,8 +1226,7 @@ static void r300SetupFragmentShaderTextures(GLcontext *ctx, int *tmu_mappings)
 {
 	r300ContextPtr r300 = R300_CONTEXT(ctx);
 	int i;
-	struct r300_fragment_program *fp = (struct r300_fragment_program *) ctx->FragmentProgram._Current;
-	struct r300_fragment_program_code *code = &fp->code.r300;
+	struct r300_fragment_program_code *code = &r300->selected_fp->code.r300;
 
 	R300_STATECHANGE(r300, fpt);
 
@@ -1268,9 +1266,9 @@ static void r300SetupFragmentShaderTextures(GLcontext *ctx, int *tmu_mappings)
 
 static void r500SetupFragmentShaderTextures(GLcontext *ctx, int *tmu_mappings)
 {
+	r300ContextPtr r300 = R300_CONTEXT(ctx);
 	int i;
-	struct r300_fragment_program *fp = (struct r300_fragment_program *) ctx->FragmentProgram._Current;
-	struct r500_fragment_program_code *code = &fp->code.r500;
+	struct r500_fragment_program_code *code = &r300->selected_fp->code.r500;
 
 	/* find all the texture instructions and relocate the texture units */
 	for (i = 0; i < code->inst_end + 1; i++) {
@@ -1318,8 +1316,6 @@ static void r300SetupTextures(GLcontext * ctx)
 	int hw_tmu = 0;
 	int last_hw_tmu = -1;	/* -1 translates into no setup costs for fields */
 	int tmu_mappings[R300_MAX_TEXTURE_UNITS] = { -1, };
-	struct r300_fragment_program *fp = (struct r300_fragment_program *)
-	    (char *)ctx->FragmentProgram._Current;
 
 	R300_STATECHANGE(r300, txe);
 	R300_STATECHANGE(r300, tex.filter);
@@ -1422,7 +1418,7 @@ static void r300SetupTextures(GLcontext * ctx)
 	    cmdpacket0(r300->radeon.radeonScreen, R300_TX_BORDER_COLOR_0, last_hw_tmu + 1);
 
 	if (r300->radeon.radeonScreen->chip_family < CHIP_FAMILY_RV515) {
-		if (fp->Base.UsesKill && last_hw_tmu < 0) {
+		if (ctx->FragmentProgram._Current->UsesKill && last_hw_tmu < 0) {
 			// The KILL operation requires the first texture unit
 			// to be enabled.
 			r300->hw.txe.cmd[R300_TXE_ENABLE] |= 1;
@@ -1464,7 +1460,7 @@ static void r300SetupRSUnit(GLcontext * ctx)
 	else
 		RENDERINPUTS_COPY(OutputsWritten.index_bitset, r300->render_inputs_bitset);
 
-	InputsRead = ctx->FragmentProgram._Current->Base.InputsRead;
+	InputsRead = r300->selected_fp->Base->InputsRead;
 
 	R300_STATECHANGE(r300, ri);
 	R300_STATECHANGE(r300, rc);
@@ -1581,7 +1577,7 @@ static void r500SetupRSUnit(GLcontext * ctx)
 	else
 		RENDERINPUTS_COPY(OutputsWritten.index_bitset, r300->render_inputs_bitset);
 
-	InputsRead = ctx->FragmentProgram._Current->Base.InputsRead;
+	InputsRead = r300->selected_fp->Base->InputsRead;
 
 	R300_STATECHANGE(r300, ri);
 	R300_STATECHANGE(r300, rc);
@@ -2038,21 +2034,18 @@ static void r300ResetHwState(r300ContextPtr r300)
 
 void r300UpdateShaders(r300ContextPtr rmesa)
 {
-	GLcontext *ctx;
+	GLcontext *ctx = rmesa->radeon.glCtx;
 	struct r300_fragment_program *fp;
-	int i;
-
-	ctx = rmesa->radeon.glCtx;
-	fp = (struct r300_fragment_program *) ctx->FragmentProgram._Current;
 
 	/* should only happenen once, just after context is created */
 	/* TODO: shouldn't we fallback to sw here? */
-	if (!fp) {
+	if (!ctx->FragmentProgram._Current) {
 		_mesa_fprintf(stderr, "No ctx->FragmentProgram._Current!!\n");
 		return;
 	}
 
 	if (rmesa->radeon.NewGLState && rmesa->options.hw_tcl_enabled) {
+		int i;
 		for (i = _TNL_FIRST_MAT; i <= _TNL_LAST_MAT; i++) {
 			rmesa->temp_attrib[i] =
 			    TNL_CONTEXT(ctx)->vb.AttribPtr[i];
@@ -2071,8 +2064,9 @@ void r300UpdateShaders(r300ContextPtr rmesa)
 		r300SwitchFallback(ctx, R300_FALLBACK_VERTEX_PROGRAM, rmesa->selected_vp->error);
 	}
 
-	if (!fp->translated || rmesa->radeon.NewGLState)
-		r300TranslateFragmentShader(ctx, ctx->FragmentProgram._Current);
+	fp = r300SelectFragmentShader(ctx);
+	if (!fp->translated)
+		r300TranslateFragmentShader(ctx, fp);
 
 	r300SwitchFallback(ctx, R300_FALLBACK_FRAGMENT_PROGRAM, fp->error);
 
@@ -2104,7 +2098,7 @@ static const GLfloat *get_fragmentprogram_constant(GLcontext *ctx,
 static void r300SetupPixelShader(GLcontext *ctx)
 {
 	r300ContextPtr rmesa = R300_CONTEXT(ctx);
-	struct r300_fragment_program *fp = (struct r300_fragment_program *) ctx->FragmentProgram._Current;
+	struct r300_fragment_program *fp = rmesa->selected_fp;
 	struct r300_fragment_program_code *code;
 	int i, k;
 
@@ -2150,8 +2144,7 @@ static void r300SetupPixelShader(GLcontext *ctx)
 	R300_STATECHANGE(rmesa, fpp);
 	rmesa->hw.fpp.cmd[R300_FPP_CMD_0] = cmdpacket0(rmesa->radeon.radeonScreen, R300_PFS_PARAM_0_X, code->const_nr * 4);
 	for (i = 0; i < code->const_nr; i++) {
-		const GLfloat *constant = get_fragmentprogram_constant(ctx,
-			&fp->Base.Base, code->constant[i]);
+		const GLfloat *constant = get_fragmentprogram_constant(ctx, fp->Base, code->constant[i]);
 		rmesa->hw.fpp.cmd[R300_FPP_PARAM_0 + 4 * i + 0] = r300PackFloat24(constant[0]);
 		rmesa->hw.fpp.cmd[R300_FPP_PARAM_0 + 4 * i + 1] = r300PackFloat24(constant[1]);
 		rmesa->hw.fpp.cmd[R300_FPP_PARAM_0 + 4 * i + 2] = r300PackFloat24(constant[2]);
@@ -2176,7 +2169,7 @@ static void r300SetupPixelShader(GLcontext *ctx)
 static void r500SetupPixelShader(GLcontext *ctx)
 {
 	r300ContextPtr rmesa = R300_CONTEXT(ctx);
-	struct r300_fragment_program *fp = (struct r300_fragment_program *) ctx->FragmentProgram._Current;
+	struct r300_fragment_program *fp = rmesa->selected_fp;
 	int i;
 	struct r500_fragment_program_code *code;
 
@@ -2212,8 +2205,7 @@ static void r500SetupPixelShader(GLcontext *ctx)
 
 	R300_STATECHANGE(rmesa, r500fp_const);
 	for (i = 0; i < code->const_nr; i++) {
-		const GLfloat *constant = get_fragmentprogram_constant(ctx,
-			&fp->Base.Base, code->constant[i]);
+		const GLfloat *constant = get_fragmentprogram_constant(ctx, fp->Base, code->constant[i]);
 		rmesa->hw.r500fp_const.cmd[R300_FPP_PARAM_0 + 4 * i + 0] = r300PackFloat32(constant[0]);
 		rmesa->hw.r500fp_const.cmd[R300_FPP_PARAM_0 + 4 * i + 1] = r300PackFloat32(constant[1]);
 		rmesa->hw.r500fp_const.cmd[R300_FPP_PARAM_0 + 4 * i + 2] = r300PackFloat32(constant[2]);
@@ -2276,20 +2268,17 @@ void r300SetupVAP(GLcontext *ctx, GLuint InputsRead, GLuint OutputsWritten)
 
 	rmesa->hw.vic.cmd[R300_VIC_CNTL_0] = r300VAPInputCntl0(ctx, InputsRead);
 	rmesa->hw.vic.cmd[R300_VIC_CNTL_1] = r300VAPInputCntl1(ctx, InputsRead);
-	rmesa->hw.vof.cmd[R300_VOF_CNTL_0] = r300VAPOutputCntl0(ctx, OutputsWritten, ctx->FragmentProgram._Current->Base.InputsRead);
-	rmesa->hw.vof.cmd[R300_VOF_CNTL_1] = r300VAPOutputCntl1(ctx, OutputsWritten, ctx->FragmentProgram._Current->Base.InputsRead);
+	rmesa->hw.vof.cmd[R300_VOF_CNTL_0] = r300VAPOutputCntl0(ctx, OutputsWritten, rmesa->selected_fp->Base->InputsRead);
+	rmesa->hw.vof.cmd[R300_VOF_CNTL_1] = r300VAPOutputCntl1(ctx, OutputsWritten, rmesa->selected_fp->Base->InputsRead);
 }
 
 void r300UpdateShaderStates(r300ContextPtr rmesa)
 {
 	GLcontext *ctx;
 	ctx = rmesa->radeon.glCtx;
-	struct r300_fragment_program *r300_fp;
-
-	r300_fp = (struct r300_fragment_program *) ctx->FragmentProgram._Current;
 
 	/* should only happenen once, just after context is created */
-	if (!r300_fp)
+	if (!ctx->FragmentProgram._Current)
 		return;
 
 	r300SetEarlyZState(ctx);
diff --git a/src/mesa/drivers/dri/r300/r300_swtcl.c b/src/mesa/drivers/dri/r300/r300_swtcl.c
index ce4179208e..db4ccce6f1 100644
--- a/src/mesa/drivers/dri/r300/r300_swtcl.c
+++ b/src/mesa/drivers/dri/r300/r300_swtcl.c
@@ -76,7 +76,7 @@ void r300ChooseSwtclVertexFormat(GLcontext *ctx, GLuint *_InputsRead,  GLuint *_
 	GLuint InputsRead = 0;
 	GLuint OutputsWritten = 0;
 	int num_attrs = 0;
-	GLuint fp_reads = ctx->FragmentProgram._Current->Base.InputsRead;
+	GLuint fp_reads = rmesa->selected_fp->Base->InputsRead;
 	struct vertex_attribute *attrs = rmesa->vbuf.attribs;
 
 	rmesa->swtcl.coloroffset = rmesa->swtcl.specoffset = 0;
-- 
cgit v1.2.3


From 37c319f62f59d7750dd172034e43dfd489f572cc Mon Sep 17 00:00:00 2001
From: Maciej Cencora <m.cencora@gmail.com>
Date: Fri, 3 Jul 2009 20:14:24 +0200
Subject: r300: implement proper IsProgramNative check for vertex programs

---
 src/mesa/drivers/dri/r300/r300_context.h  |  1 -
 src/mesa/drivers/dri/r300/r300_shader.c   |  9 +++++++--
 src/mesa/drivers/dri/r300/r300_state.c    |  8 ++++++--
 src/mesa/drivers/dri/r300/r300_vertprog.c | 13 ++++++-------
 src/mesa/drivers/dri/r300/r300_vertprog.h |  4 ++++
 5 files changed, 23 insertions(+), 12 deletions(-)

(limited to 'src/mesa/drivers/dri/r300/r300_state.c')

diff --git a/src/mesa/drivers/dri/r300/r300_context.h b/src/mesa/drivers/dri/r300/r300_context.h
index 30b6e75039..44211a45b3 100644
--- a/src/mesa/drivers/dri/r300/r300_context.h
+++ b/src/mesa/drivers/dri/r300/r300_context.h
@@ -673,7 +673,6 @@ extern GLboolean r300CreateContext(const __GLcontextModes * glVisual,
 				   __DRIcontextPrivate * driContextPriv,
 				   void *sharedContextPrivate);
 
-extern void r300SelectVertexShader(r300ContextPtr r300);
 extern void r300InitShaderFuncs(struct dd_function_table *functions);
 
 extern void r300InitShaderFunctions(r300ContextPtr r300);
diff --git a/src/mesa/drivers/dri/r300/r300_shader.c b/src/mesa/drivers/dri/r300/r300_shader.c
index 7206379ea0..854eb5d80a 100644
--- a/src/mesa/drivers/dri/r300/r300_shader.c
+++ b/src/mesa/drivers/dri/r300/r300_shader.c
@@ -128,8 +128,13 @@ r300IsProgramNative(GLcontext * ctx, GLenum target, struct gl_program *prog)
 			r300TranslateFragmentShader(ctx, fp);
 
 		return !fp->error;
-	} else
-		return GL_TRUE;
+	} else {
+		struct r300_vertex_program *vp = r300SelectVertexShader(ctx);
+		if (!vp->translated)
+			r300TranslateVertexShader(vp);
+
+		return !vp->error;
+	}
 }
 
 void r300InitShaderFuncs(struct dd_function_table *functions)
diff --git a/src/mesa/drivers/dri/r300/r300_state.c b/src/mesa/drivers/dri/r300/r300_state.c
index e0996383eb..8c228ab5e7 100644
--- a/src/mesa/drivers/dri/r300/r300_state.c
+++ b/src/mesa/drivers/dri/r300/r300_state.c
@@ -2045,6 +2045,7 @@ void r300UpdateShaders(r300ContextPtr rmesa)
 	}
 
 	if (rmesa->radeon.NewGLState && rmesa->options.hw_tcl_enabled) {
+		struct r300_vertex_program *vp;
 		int i;
 		for (i = _TNL_FIRST_MAT; i <= _TNL_LAST_MAT; i++) {
 			rmesa->temp_attrib[i] =
@@ -2060,8 +2061,11 @@ void r300UpdateShaders(r300ContextPtr rmesa)
 			    rmesa->temp_attrib[i];
 		}
 
-		r300SelectVertexShader(rmesa);
-		r300SwitchFallback(ctx, R300_FALLBACK_VERTEX_PROGRAM, rmesa->selected_vp->error);
+		vp = r300SelectVertexShader(ctx);
+		if (!vp->translated)
+			r300TranslateVertexShader(vp);
+
+		r300SwitchFallback(ctx, R300_FALLBACK_VERTEX_PROGRAM, vp->error);
 	}
 
 	fp = r300SelectFragmentShader(ctx);
diff --git a/src/mesa/drivers/dri/r300/r300_vertprog.c b/src/mesa/drivers/dri/r300/r300_vertprog.c
index 25c42814ec..326183bda4 100644
--- a/src/mesa/drivers/dri/r300/r300_vertprog.c
+++ b/src/mesa/drivers/dri/r300/r300_vertprog.c
@@ -1027,7 +1027,7 @@ static void t_inputs_outputs(struct r300_vertex_program *vp)
 	}
 }
 
-static void r300TranslateVertexShader(struct r300_vertex_program *vp)
+void r300TranslateVertexShader(struct r300_vertex_program *vp)
 {
 	struct prog_instruction *vpi = vp->Base->Base.Instructions;
 	int i;
@@ -1524,7 +1524,6 @@ static struct r300_vertex_program *build_program(GLcontext *ctx,
 
 	assert(prog->NumInstructions);
 	vp->num_temporaries = prog->NumTemporaries;
-	r300TranslateVertexShader(vp);
 
 	return vp;
 }
@@ -1538,9 +1537,9 @@ static void add_outputs(struct r300_vertex_program_key *key, GLint vert)
 	key->OutputsAdded |= 1 << vert;
 }
 
-void r300SelectVertexShader(r300ContextPtr r300)
+struct r300_vertex_program * r300SelectVertexShader(GLcontext *ctx)
 {
-	GLcontext *ctx = ctx = r300->radeon.glCtx;
+	r300ContextPtr r300 = R300_CONTEXT(ctx);
 	GLuint InputsRead;
 	struct r300_vertex_program_key wanted_key = { 0 };
 	GLint i;
@@ -1596,14 +1595,14 @@ void r300SelectVertexShader(r300ContextPtr r300)
 	for (vp = vpc->progs; vp; vp = vp->next)
 		if (_mesa_memcmp(&vp->key, &wanted_key, sizeof(wanted_key))
 		    == 0) {
-			r300->selected_vp = vp;
-			return;
+			return r300->selected_vp = vp;
 		}
 
 	vp = build_program(ctx, &wanted_key, &vpc->mesa_program, wpos_idx);
 	vp->next = vpc->progs;
 	vpc->progs = vp;
-	r300->selected_vp = vp;
+
+	return r300->selected_vp = vp;
 }
 
 #define bump_vpu_count(ptr, new_count)   do { \
diff --git a/src/mesa/drivers/dri/r300/r300_vertprog.h b/src/mesa/drivers/dri/r300/r300_vertprog.h
index b552e3fb1b..2dab11c337 100644
--- a/src/mesa/drivers/dri/r300/r300_vertprog.h
+++ b/src/mesa/drivers/dri/r300/r300_vertprog.h
@@ -34,4 +34,8 @@
 
 void r300SetupVertexProgram(r300ContextPtr rmesa);
 
+struct r300_vertex_program * r300SelectVertexShader(GLcontext *ctx);
+
+void r300TranslateVertexShader(struct r300_vertex_program *vp);
+
 #endif
-- 
cgit v1.2.3


From 7360f83364b407e949529eeca9f8c421d2da3ded Mon Sep 17 00:00:00 2001
From: Maciej Cencora <m.cencora@gmail.com>
Date: Sat, 4 Jul 2009 13:17:35 +0200
Subject: r300: move fragment program selection before vertex program selection

Prepare for wpos and fogc handling rewrite.
---
 src/mesa/drivers/dri/r300/r300_state.c | 44 +++++++++++++++++++---------------
 1 file changed, 25 insertions(+), 19 deletions(-)

(limited to 'src/mesa/drivers/dri/r300/r300_state.c')

diff --git a/src/mesa/drivers/dri/r300/r300_state.c b/src/mesa/drivers/dri/r300/r300_state.c
index 8c228ab5e7..344f021acf 100644
--- a/src/mesa/drivers/dri/r300/r300_state.c
+++ b/src/mesa/drivers/dri/r300/r300_state.c
@@ -2035,7 +2035,6 @@ static void r300ResetHwState(r300ContextPtr r300)
 void r300UpdateShaders(r300ContextPtr rmesa)
 {
 	GLcontext *ctx = rmesa->radeon.glCtx;
-	struct r300_fragment_program *fp;
 
 	/* should only happenen once, just after context is created */
 	/* TODO: shouldn't we fallback to sw here? */
@@ -2044,21 +2043,34 @@ void r300UpdateShaders(r300ContextPtr rmesa)
 		return;
 	}
 
-	if (rmesa->radeon.NewGLState && rmesa->options.hw_tcl_enabled) {
+	{
+		struct r300_fragment_program *fp;
+
+		fp = r300SelectFragmentShader(ctx);
+		if (!fp->translated)
+			r300TranslateFragmentShader(ctx, fp);
+
+		r300SwitchFallback(ctx, R300_FALLBACK_FRAGMENT_PROGRAM, fp->error);
+	}
+
+	if (rmesa->options.hw_tcl_enabled) {
 		struct r300_vertex_program *vp;
-		int i;
-		for (i = _TNL_FIRST_MAT; i <= _TNL_LAST_MAT; i++) {
-			rmesa->temp_attrib[i] =
-			    TNL_CONTEXT(ctx)->vb.AttribPtr[i];
-			TNL_CONTEXT(ctx)->vb.AttribPtr[i] =
-			    &rmesa->dummy_attrib[i];
-		}
 
-		_tnl_UpdateFixedFunctionProgram(ctx);
+		if (rmesa->radeon.NewGLState) {
+			int i;
+			for (i = _TNL_FIRST_MAT; i <= _TNL_LAST_MAT; i++) {
+				rmesa->temp_attrib[i] =
+				    TNL_CONTEXT(ctx)->vb.AttribPtr[i];
+				TNL_CONTEXT(ctx)->vb.AttribPtr[i] =
+				    &rmesa->dummy_attrib[i];
+			}
+
+			_tnl_UpdateFixedFunctionProgram(ctx);
 
-		for (i = _TNL_FIRST_MAT; i <= _TNL_LAST_MAT; i++) {
-			TNL_CONTEXT(ctx)->vb.AttribPtr[i] =
-			    rmesa->temp_attrib[i];
+			for (i = _TNL_FIRST_MAT; i <= _TNL_LAST_MAT; i++) {
+				TNL_CONTEXT(ctx)->vb.AttribPtr[i] =
+				    rmesa->temp_attrib[i];
+			}
 		}
 
 		vp = r300SelectVertexShader(ctx);
@@ -2068,12 +2080,6 @@ void r300UpdateShaders(r300ContextPtr rmesa)
 		r300SwitchFallback(ctx, R300_FALLBACK_VERTEX_PROGRAM, vp->error);
 	}
 
-	fp = r300SelectFragmentShader(ctx);
-	if (!fp->translated)
-		r300TranslateFragmentShader(ctx, fp);
-
-	r300SwitchFallback(ctx, R300_FALLBACK_FRAGMENT_PROGRAM, fp->error);
-
 	r300UpdateStateParameters(ctx, _NEW_PROGRAM | _NEW_PROGRAM_CONSTANTS);
 	rmesa->radeon.NewGLState = 0;
 }
-- 
cgit v1.2.3


From df5fe747fa08f63b949ba0fd6628060831b562ec Mon Sep 17 00:00:00 2001
From: Maciej Cencora <m.cencora@gmail.com>
Date: Sun, 5 Jul 2009 02:03:32 +0200
Subject: r300: bind vertex program to fragment program

---
 src/mesa/drivers/dri/r300/r300_context.h  |   4 +-
 src/mesa/drivers/dri/r300/r300_draw.c     |   4 +-
 src/mesa/drivers/dri/r300/r300_state.c    |   4 +-
 src/mesa/drivers/dri/r300/r300_vertprog.c | 205 ++++++++++++++----------------
 4 files changed, 103 insertions(+), 114 deletions(-)

(limited to 'src/mesa/drivers/dri/r300/r300_state.c')

diff --git a/src/mesa/drivers/dri/r300/r300_context.h b/src/mesa/drivers/dri/r300/r300_context.h
index 44211a45b3..6f99e49050 100644
--- a/src/mesa/drivers/dri/r300/r300_context.h
+++ b/src/mesa/drivers/dri/r300/r300_context.h
@@ -409,9 +409,7 @@ struct r300_vertex_program {
 	struct r300_vertex_program *next;
 
 	struct r300_vertex_program_key {
-		GLuint InputsRead;
-		GLuint OutputsWritten;
-		GLuint OutputsAdded;
+		GLuint FpReads;
 	} key;
 	
 	struct r300_vertex_shader_hw_code {
diff --git a/src/mesa/drivers/dri/r300/r300_draw.c b/src/mesa/drivers/dri/r300/r300_draw.c
index 4e8b62f186..5420293b91 100644
--- a/src/mesa/drivers/dri/r300/r300_draw.c
+++ b/src/mesa/drivers/dri/r300/r300_draw.c
@@ -332,7 +332,7 @@ static void r300SetVertexFormat(GLcontext *ctx, const struct gl_client_array *ar
 	{
 		int i, tmp;
 
-		tmp = r300->selected_vp->key.InputsRead;
+		tmp = r300->selected_vp->Base->Base.InputsRead;
 		i = 0;
 		vbuf->num_attribs = 0;
 		while (tmp) {
@@ -428,7 +428,7 @@ static GLboolean r300TryDrawPrims(GLcontext *ctx,
 	if (r300->fallback)
 		return GL_FALSE;
 
-	r300SetupVAP(ctx, r300->selected_vp->key.InputsRead, r300->selected_vp->key.OutputsWritten);
+	r300SetupVAP(ctx, r300->selected_vp->Base->Base.InputsRead, r300->selected_vp->Base->Base.OutputsWritten);
 
 	r300UpdateShaderStates(r300);
 
diff --git a/src/mesa/drivers/dri/r300/r300_state.c b/src/mesa/drivers/dri/r300/r300_state.c
index 344f021acf..0f3198e792 100644
--- a/src/mesa/drivers/dri/r300/r300_state.c
+++ b/src/mesa/drivers/dri/r300/r300_state.c
@@ -1456,7 +1456,7 @@ static void r300SetupRSUnit(GLcontext * ctx)
 	hw_tcl_on = r300->options.hw_tcl_enabled;
 
 	if (hw_tcl_on)
-		OutputsWritten.vp_outputs = r300->selected_vp->key.OutputsWritten;
+		OutputsWritten.vp_outputs = r300->selected_vp->Base->Base.OutputsWritten;
 	else
 		RENDERINPUTS_COPY(OutputsWritten.index_bitset, r300->render_inputs_bitset);
 
@@ -1573,7 +1573,7 @@ static void r500SetupRSUnit(GLcontext * ctx)
 	hw_tcl_on = r300->options.hw_tcl_enabled;
 
 	if (hw_tcl_on)
-		OutputsWritten.vp_outputs = r300->selected_vp->key.OutputsWritten;
+		OutputsWritten.vp_outputs = r300->selected_vp->Base->Base.OutputsWritten;
 	else
 		RENDERINPUTS_COPY(OutputsWritten.index_bitset, r300->render_inputs_bitset);
 
diff --git a/src/mesa/drivers/dri/r300/r300_vertprog.c b/src/mesa/drivers/dri/r300/r300_vertprog.c
index 326183bda4..0b04830350 100644
--- a/src/mesa/drivers/dri/r300/r300_vertprog.c
+++ b/src/mesa/drivers/dri/r300/r300_vertprog.c
@@ -961,10 +961,14 @@ static void t_inputs_outputs(struct r300_vertex_program *vp)
 {
 	int i;
 	int cur_reg;
+	GLuint OutputsWritten, InputsRead;
+
+	OutputsWritten = vp->Base->Base.OutputsWritten;
+	InputsRead = vp->Base->Base.InputsRead;
 
 	cur_reg = -1;
 	for (i = 0; i < VERT_ATTRIB_MAX; i++) {
-		if (vp->key.InputsRead & (1 << i))
+		if (InputsRead & (1 << i))
 			vp->inputs[i] = ++cur_reg;
 		else
 			vp->inputs[i] = -1;
@@ -974,13 +978,13 @@ static void t_inputs_outputs(struct r300_vertex_program *vp)
 	for (i = 0; i < VERT_RESULT_MAX; i++)
 		vp->outputs[i] = -1;
 
-	assert(vp->key.OutputsWritten & (1 << VERT_RESULT_HPOS));
+	assert(OutputsWritten & (1 << VERT_RESULT_HPOS));
 
-	if (vp->key.OutputsWritten & (1 << VERT_RESULT_HPOS)) {
+	if (OutputsWritten & (1 << VERT_RESULT_HPOS)) {
 		vp->outputs[VERT_RESULT_HPOS] = cur_reg++;
 	}
 
-	if (vp->key.OutputsWritten & (1 << VERT_RESULT_PSIZ)) {
+	if (OutputsWritten & (1 << VERT_RESULT_PSIZ)) {
 		vp->outputs[VERT_RESULT_PSIZ] = cur_reg++;
 	}
 
@@ -990,39 +994,39 @@ static void t_inputs_outputs(struct r300_vertex_program *vp)
 	 * pretend it does by skipping output index reg so the colors
 	 * get written into appropriate output vectors.
 	 */
-	if (vp->key.OutputsWritten & (1 << VERT_RESULT_COL0)) {
+	if (OutputsWritten & (1 << VERT_RESULT_COL0)) {
 		vp->outputs[VERT_RESULT_COL0] = cur_reg++;
-	} else if (vp->key.OutputsWritten & (1 << VERT_RESULT_BFC0) ||
-		vp->key.OutputsWritten & (1 << VERT_RESULT_BFC1)) {
+	} else if (OutputsWritten & (1 << VERT_RESULT_BFC0) ||
+		OutputsWritten & (1 << VERT_RESULT_BFC1)) {
 		cur_reg++;
 	}
 
-	if (vp->key.OutputsWritten & (1 << VERT_RESULT_COL1)) {
+	if (OutputsWritten & (1 << VERT_RESULT_COL1)) {
 		vp->outputs[VERT_RESULT_COL1] = cur_reg++;
-	} else if (vp->key.OutputsWritten & (1 << VERT_RESULT_BFC0) ||
-		vp->key.OutputsWritten & (1 << VERT_RESULT_BFC1)) {
+	} else if (OutputsWritten & (1 << VERT_RESULT_BFC0) ||
+		OutputsWritten & (1 << VERT_RESULT_BFC1)) {
 		cur_reg++;
 	}
 
-	if (vp->key.OutputsWritten & (1 << VERT_RESULT_BFC0)) {
+	if (OutputsWritten & (1 << VERT_RESULT_BFC0)) {
 		vp->outputs[VERT_RESULT_BFC0] = cur_reg++;
-	} else if (vp->key.OutputsWritten & (1 << VERT_RESULT_BFC1)) {
+	} else if (OutputsWritten & (1 << VERT_RESULT_BFC1)) {
 		cur_reg++;
 	}
 
-	if (vp->key.OutputsWritten & (1 << VERT_RESULT_BFC1)) {
+	if (OutputsWritten & (1 << VERT_RESULT_BFC1)) {
 		vp->outputs[VERT_RESULT_BFC1] = cur_reg++;
-	} else if (vp->key.OutputsWritten & (1 << VERT_RESULT_BFC0)) {
+	} else if (OutputsWritten & (1 << VERT_RESULT_BFC0)) {
 		cur_reg++;
 	}
 
 	for (i = VERT_RESULT_TEX0; i <= VERT_RESULT_TEX7; i++) {
-		if (vp->key.OutputsWritten & (1 << i)) {
+		if (OutputsWritten & (1 << i)) {
 			vp->outputs[i] = cur_reg++;
 		}
 	}
 
-	if (vp->key.OutputsWritten & (1 << VERT_RESULT_FOGC)) {
+	if (OutputsWritten & (1 << VERT_RESULT_FOGC)) {
 		vp->outputs[VERT_RESULT_FOGC] = cur_reg++;
 	}
 }
@@ -1255,6 +1259,8 @@ static void insert_wpos(struct r300_vertex_program *vp, struct gl_program *prog,
 	++vpi;
 
 	vpi->Opcode = OPCODE_END;
+
+	prog->OutputsWritten |= 1 << (VERT_RESULT_TEX0 + vp->wpos_idx);
 }
 
 static void pos_as_texcoord(struct r300_vertex_program *vp,
@@ -1448,6 +1454,63 @@ static void translateInsts(struct gl_program *prog)
 	}
 }
 
+#define ADD_OUTPUT(fp_attr, vp_result) \
+	do { \
+		if ((FpReads & (1 << (fp_attr))) && !(prog->OutputsWritten & (1 << (vp_result)))) { \
+			OutputsAdded |= 1 << (vp_result); \
+			count++; \
+		} \
+	} while (0)
+
+static void addArtificialOutputs(GLcontext *ctx, struct gl_program *prog)
+{
+	r300ContextPtr r300 = R300_CONTEXT(ctx);
+	GLuint OutputsAdded, FpReads;
+	int i, count;
+
+	OutputsAdded = count = 0;
+	FpReads = r300->selected_fp->Base->InputsRead;
+
+	ADD_OUTPUT(FRAG_ATTRIB_COL0, VERT_RESULT_COL0);
+	ADD_OUTPUT(FRAG_ATTRIB_COL1, VERT_RESULT_COL1);
+
+	for (i = 0; i < 7; ++i) {
+		ADD_OUTPUT(FRAG_ATTRIB_TEX0 + i, VERT_RESULT_TEX0 + i);
+	}
+
+	/* Some outputs may be artificially added, to match the inputs of the fragment program.
+	 * Issue 16 of vertex program spec says that all vertex attributes that are unwritten by
+	 * vertex program are undefined, so just use MOV [vertex_result], CONST[0]
+	 */
+	if (count > 0) {
+		struct prog_instruction *inst;
+
+		_mesa_insert_instructions(prog, prog->NumInstructions - 1, count);
+		inst = &prog->Instructions[prog->NumInstructions - 1 - count];
+
+		for (i = 0; i < VERT_RESULT_MAX; ++i) {
+			if (OutputsAdded & (1 << i)) {
+				inst->Opcode = OPCODE_MOV;
+
+				inst->DstReg.File = PROGRAM_OUTPUT;
+				inst->DstReg.Index = i;
+				inst->DstReg.WriteMask = WRITEMASK_XYZW;
+				inst->DstReg.CondMask = COND_TR;
+
+				inst->SrcReg[0].File = PROGRAM_CONSTANT;
+				inst->SrcReg[0].Index = 0;
+				inst->SrcReg[0].Swizzle = SWIZZLE_XYZW;
+
+				++inst;
+			}
+		}
+
+		prog->OutputsWritten |= OutputsAdded;
+	}
+}
+
+#undef ADD_OUTPUT
+
 static struct r300_vertex_program *build_program(GLcontext *ctx,
 						 struct r300_vertex_program_key *wanted_key,
 						 const struct gl_vertex_program *mesa_vp,
@@ -1477,42 +1540,7 @@ static struct r300_vertex_program *build_program(GLcontext *ctx,
 		pos_as_texcoord(vp, prog);
 	}
 
-	/* Some outputs may be artificially added, to match the inputs of the fragment program.
-	 * Issue 16 of vertex program spec says that all vertex attributes that are unwritten by
-	 * vertex program are undefined, so just use MOV [vertex_result], CONST[0]
-	 */
-	{
-		int i, count = 0;
-		for (i = 0; i < VERT_RESULT_MAX; ++i) {
-			if (vp->key.OutputsAdded & (1 << i)) {
-				++count;
-			}
-		}
-
-		if (count > 0) {
-			struct prog_instruction *inst;
-
-			_mesa_insert_instructions(prog, prog->NumInstructions - 1, count);
-			inst = &prog->Instructions[prog->NumInstructions - 1 - count];
-
-			for (i = 0; i < VERT_RESULT_MAX; ++i) {
-				if (vp->key.OutputsAdded & (1 << i)) {
-					inst->Opcode = OPCODE_MOV;
-
-					inst->DstReg.File = PROGRAM_OUTPUT;
-					inst->DstReg.Index = i;
-					inst->DstReg.WriteMask = WRITEMASK_XYZW;
-					inst->DstReg.CondMask = COND_TR;
-
-					inst->SrcReg[0].File = PROGRAM_CONSTANT;
-					inst->SrcReg[0].Index = 0;
-					inst->SrcReg[0].Swizzle = SWIZZLE_XYZW;
-
-					++inst;
-				}
-			}
-		}
-	}
+	addArtificialOutputs(ctx, prog);
 
 	translateInsts(prog);
 
@@ -1528,76 +1556,39 @@ static struct r300_vertex_program *build_program(GLcontext *ctx,
 	return vp;
 }
 
-static void add_outputs(struct r300_vertex_program_key *key, GLint vert)
-{
-	if (key->OutputsWritten & (1 << vert))
-		return;
-
-	key->OutputsWritten |= 1 << vert;
-	key->OutputsAdded |= 1 << vert;
-}
-
 struct r300_vertex_program * r300SelectVertexShader(GLcontext *ctx)
 {
 	r300ContextPtr r300 = R300_CONTEXT(ctx);
-	GLuint InputsRead;
 	struct r300_vertex_program_key wanted_key = { 0 };
-	GLint i;
 	struct r300_vertex_program_cont *vpc;
 	struct r300_vertex_program *vp;
 	GLint wpos_idx;
 
 	vpc = (struct r300_vertex_program_cont *)ctx->VertexProgram._Current;
-	wanted_key.InputsRead = vpc->mesa_program.Base.InputsRead;
-	wanted_key.OutputsWritten = vpc->mesa_program.Base.OutputsWritten;
-	InputsRead = ctx->FragmentProgram._Current->Base.InputsRead;
-
-	wpos_idx = -1;
-	if (InputsRead & FRAG_BIT_WPOS) {
-		for (i = 0; i < ctx->Const.MaxTextureUnits; i++)
-			if (!(InputsRead & (FRAG_BIT_TEX0 << i)))
-				break;
+	wanted_key.FpReads = r300->selected_fp->Base->InputsRead;
 
-		if (i == ctx->Const.MaxTextureUnits) {
-			fprintf(stderr, "\tno free texcoord found\n");
-			_mesa_exit(-1);
+	for (vp = vpc->progs; vp; vp = vp->next)
+		if (_mesa_memcmp(&vp->key, &wanted_key, sizeof(wanted_key))
+		    == 0) {
+			return r300->selected_vp = vp;
 		}
 
-		wanted_key.OutputsWritten |= 1 << (VERT_RESULT_TEX0 + i);
-		wpos_idx = i;
-	}
-
-	if (vpc->mesa_program.IsPositionInvariant) {
-		wanted_key.InputsRead |= (1 << VERT_ATTRIB_POS);
-		wanted_key.OutputsWritten |= (1 << VERT_RESULT_HPOS);
-	} else {
-		add_outputs(&wanted_key, VERT_RESULT_HPOS);
-	}
-
-	if (InputsRead & FRAG_BIT_COL0) {
-		add_outputs(&wanted_key, VERT_RESULT_COL0);
-	}
+	wpos_idx = -1;
+	if (wanted_key.FpReads & FRAG_BIT_WPOS) {
+		GLint i;
 
-	if (InputsRead & FRAG_BIT_COL1) {
-		add_outputs(&wanted_key, VERT_RESULT_COL1);
-	}
+		for (i = 0; i < ctx->Const.MaxTextureUnits; i++)
+			if (!(wanted_key.FpReads & (FRAG_BIT_TEX(i))))
+				break;
 
-	if (InputsRead & FRAG_BIT_FOGC) {
-		add_outputs(&wanted_key, VERT_RESULT_FOGC);
-	}
+			if (i == ctx->Const.MaxTextureUnits) {
+				fprintf(stderr, "\tno free texcoord found\n");
+				_mesa_exit(-1);
+			}
 
-	for (i = 0; i < ctx->Const.MaxTextureUnits; i++) {
-		if (InputsRead & (FRAG_BIT_TEX0 << i)) {
-			add_outputs(&wanted_key, VERT_RESULT_TEX0 + i);
-		}
+			wpos_idx = i;
 	}
 
-	for (vp = vpc->progs; vp; vp = vp->next)
-		if (_mesa_memcmp(&vp->key, &wanted_key, sizeof(wanted_key))
-		    == 0) {
-			return r300->selected_vp = vp;
-		}
-
 	vp = build_program(ctx, &wanted_key, &vpc->mesa_program, wpos_idx);
 	vp->next = vpc->progs;
 	vpc->progs = vp;
@@ -1663,8 +1654,8 @@ void r300SetupVertexProgram(r300ContextPtr rmesa)
 	r300EmitVertexProgram(rmesa, R300_PVS_CODE_START, &(prog->hw_code));
 	inst_count = (prog->hw_code.length / 4) - 1;
 
-	r300VapCntl(rmesa, _mesa_bitcount(prog->key.InputsRead),
-				 _mesa_bitcount(prog->key.OutputsWritten), prog->num_temporaries);
+	r300VapCntl(rmesa, _mesa_bitcount(prog->Base->Base.InputsRead),
+				 _mesa_bitcount(prog->Base->Base.OutputsWritten), prog->num_temporaries);
 
 	R300_STATECHANGE(rmesa, pvs);
 	rmesa->hw.pvs.cmd[R300_PVS_CNTL_1] = (0 << R300_PVS_FIRST_INST_SHIFT) | (inst_count << R300_PVS_XYZW_VALID_INST_SHIFT) |
-- 
cgit v1.2.3


From f79ef95df4f19124c24e59583bf9fb1e347d8f2b Mon Sep 17 00:00:00 2001
From: Maciej Cencora <m.cencora@gmail.com>
Date: Sun, 5 Jul 2009 02:32:51 +0200
Subject: r300: rewrite FOGC and HPOS attribs handling

Rewrite vertex and fragment programs so that we don't have to do any hacks on lower level.
---
 src/mesa/drivers/dri/r300/r300_context.h         |  8 ++-
 src/mesa/drivers/dri/r300/r300_emit.c            |  9 ---
 src/mesa/drivers/dri/r300/r300_fragprog_common.c | 62 ++++++++++++++++++-
 src/mesa/drivers/dri/r300/r300_shader.c          |  3 +-
 src/mesa/drivers/dri/r300/r300_state.c           | 53 ----------------
 src/mesa/drivers/dri/r300/r300_swtcl.c           | 38 +++++-------
 src/mesa/drivers/dri/r300/r300_vertprog.c        | 79 ++++++++++++++----------
 7 files changed, 130 insertions(+), 122 deletions(-)

(limited to 'src/mesa/drivers/dri/r300/r300_state.c')

diff --git a/src/mesa/drivers/dri/r300/r300_context.h b/src/mesa/drivers/dri/r300/r300_context.h
index 6f99e49050..32104b8158 100644
--- a/src/mesa/drivers/dri/r300/r300_context.h
+++ b/src/mesa/drivers/dri/r300/r300_context.h
@@ -410,6 +410,8 @@ struct r300_vertex_program {
 
 	struct r300_vertex_program_key {
 		GLuint FpReads;
+		GLuint FogAttr;
+		GLuint WPosAttr;
 	} key;
 	
 	struct r300_vertex_shader_hw_code {
@@ -425,7 +427,6 @@ struct r300_vertex_program {
 
 	int pos_end;
 	int num_temporaries;	/* Number of temp vars used by program */
-	int wpos_idx;
 	int inputs[VERT_ATTRIB_MAX];
 	int outputs[VERT_RESULT_MAX];
 };
@@ -560,6 +561,11 @@ struct r300_fragment_program {
 	GLuint optimization;
 
 	struct r300_fragment_program *next;
+
+	/* attribute that we are sending the WPOS in */
+	gl_frag_attrib wpos_attr;
+	/* attribute that we are sending the fog coordinate in */
+	gl_frag_attrib fog_attr;
 };
 
 struct r300_fragment_program_cont {
diff --git a/src/mesa/drivers/dri/r300/r300_emit.c b/src/mesa/drivers/dri/r300/r300_emit.c
index c3817721dc..707d1284ed 100644
--- a/src/mesa/drivers/dri/r300/r300_emit.c
+++ b/src/mesa/drivers/dri/r300/r300_emit.c
@@ -116,15 +116,6 @@ GLuint r300VAPOutputCntl1(GLcontext * ctx, GLuint vp_writes, GLuint fp_reads)
 		}
 	}
 
-	if (fp_reads & FRAG_BIT_WPOS) {
-		ret |= (4 << (3 * first_free_texcoord));
-		++first_free_texcoord;
-	}
-
-	if (vp_writes & (1 << VERT_RESULT_FOGC) && fp_reads & FRAG_BIT_FOGC) {
-		ret |= 4 << (3 * first_free_texcoord);
-	}
-
 	if (first_free_texcoord > 8) {
 		fprintf(stderr, "\tout of free texcoords\n");
 		_mesa_exit(-1);
diff --git a/src/mesa/drivers/dri/r300/r300_fragprog_common.c b/src/mesa/drivers/dri/r300/r300_fragprog_common.c
index b25cf24007..e90be9b7f8 100644
--- a/src/mesa/drivers/dri/r300/r300_fragprog_common.c
+++ b/src/mesa/drivers/dri/r300/r300_fragprog_common.c
@@ -69,8 +69,10 @@ static void insert_WPOS_trailer(struct r300_fragment_program_compiler *compiler)
 {
 	GLuint InputsRead = compiler->fp->Base->InputsRead;
 
-	if (!(InputsRead & FRAG_BIT_WPOS))
+	if (!(InputsRead & FRAG_BIT_WPOS)) {
+		compiler->fp->wpos_attr = FRAG_ATTRIB_MAX;
 		return;
+	}
 
 	static gl_state_index tokens[STATE_LENGTH] = {
 		STATE_INTERNAL, STATE_R300_WINDOW_DIMENSION, 0, 0, 0
@@ -78,10 +80,23 @@ static void insert_WPOS_trailer(struct r300_fragment_program_compiler *compiler)
 	struct prog_instruction *fpi;
 	GLuint window_index;
 	int i = 0;
+
+	for (i = FRAG_ATTRIB_TEX0; i <= FRAG_ATTRIB_TEX7; ++i)
+	{
+		if (!(InputsRead & (1 << i))) {
+			InputsRead &= ~(1 << FRAG_ATTRIB_WPOS);
+			InputsRead |= 1 << i;
+			compiler->fp->Base->InputsRead = InputsRead;
+			compiler->fp->wpos_attr = i;
+			break;
+		}
+	}
+
 	GLuint tempregi = _mesa_find_free_register(compiler->program, PROGRAM_TEMPORARY);
 
 	_mesa_insert_instructions(compiler->program, 0, 3);
 	fpi = compiler->program->Instructions;
+	i = 0;
 
 	/* perspective divide */
 	fpi[i].Opcode = OPCODE_RCP;
@@ -92,7 +107,7 @@ static void insert_WPOS_trailer(struct r300_fragment_program_compiler *compiler)
 	fpi[i].DstReg.CondMask = COND_TR;
 
 	fpi[i].SrcReg[0].File = PROGRAM_INPUT;
-	fpi[i].SrcReg[0].Index = FRAG_ATTRIB_WPOS;
+	fpi[i].SrcReg[0].Index = compiler->fp->wpos_attr;
 	fpi[i].SrcReg[0].Swizzle = SWIZZLE_WWWW;
 	i++;
 
@@ -104,7 +119,7 @@ static void insert_WPOS_trailer(struct r300_fragment_program_compiler *compiler)
 	fpi[i].DstReg.CondMask = COND_TR;
 
 	fpi[i].SrcReg[0].File = PROGRAM_INPUT;
-	fpi[i].SrcReg[0].Index = FRAG_ATTRIB_WPOS;
+	fpi[i].SrcReg[0].Index = compiler->fp->wpos_attr;
 	fpi[i].SrcReg[0].Swizzle = SWIZZLE_XYZW;
 
 	fpi[i].SrcReg[1].File = PROGRAM_TEMPORARY;
@@ -147,6 +162,45 @@ static void insert_WPOS_trailer(struct r300_fragment_program_compiler *compiler)
 	}
 }
 
+static void rewriteFog(struct r300_fragment_program_compiler *compiler)
+{
+	struct r300_fragment_program *fp = compiler->fp;
+	GLuint InputsRead;
+	int i;
+
+	InputsRead = fp->Base->InputsRead;
+
+	if (!(InputsRead & FRAG_BIT_FOGC)) {
+		fp->fog_attr = FRAG_ATTRIB_MAX;
+		return;
+	}
+
+	for (i = FRAG_ATTRIB_TEX0; i <= FRAG_ATTRIB_TEX7; ++i)
+	{
+		if (!(InputsRead & (1 << i))) {
+			InputsRead &= ~(1 << FRAG_ATTRIB_FOGC);
+			InputsRead |= 1 << i;
+			fp->Base->InputsRead = InputsRead;
+			fp->fog_attr = i;
+			break;
+		}
+	}
+
+	{
+		struct prog_instruction *inst;
+
+		inst = compiler->program->Instructions;
+		while (inst->Opcode != OPCODE_END) {
+			const int src_regs = _mesa_num_inst_src_regs(inst->Opcode);
+			for (i = 0; i < src_regs; ++i) {
+				if (inst->SrcReg[i].File == PROGRAM_INPUT && inst->SrcReg[i].Index == FRAG_ATTRIB_FOGC)
+					inst->SrcReg[i].Index = fp->fog_attr;
+			}
+			++inst;
+		}
+	}
+}
+
 static GLuint build_dtm(GLuint depthmode)
 {
 	switch(depthmode) {
@@ -204,6 +258,8 @@ void r300TranslateFragmentShader(GLcontext *ctx, struct r300_fragment_program *f
 
 	insert_WPOS_trailer(&compiler);
 
+	rewriteFog(&compiler);
+
 	if (r300->radeon.radeonScreen->chip_family >= CHIP_FAMILY_RV515) {
 		struct radeon_program_transformation transformations[] = {
 			{ &r500_transform_TEX, &compiler },
diff --git a/src/mesa/drivers/dri/r300/r300_shader.c b/src/mesa/drivers/dri/r300/r300_shader.c
index 854eb5d80a..62228a3786 100644
--- a/src/mesa/drivers/dri/r300/r300_shader.c
+++ b/src/mesa/drivers/dri/r300/r300_shader.c
@@ -66,8 +66,7 @@ static struct gl_program *r300NewProgram(GLcontext * ctx, GLenum target,
 	case GL_VERTEX_STATE_PROGRAM_NV:
 	case GL_VERTEX_PROGRAM_ARB:
 		vp = CALLOC_STRUCT(r300_vertex_program_cont);
-		return _mesa_init_vertex_program(ctx, &vp->mesa_program,
-						 target, id);
+		return _mesa_init_vertex_program(ctx, &vp->mesa_program, target, id);
 
 	case GL_FRAGMENT_PROGRAM_NV:
 	case GL_FRAGMENT_PROGRAM_ARB:
diff --git a/src/mesa/drivers/dri/r300/r300_state.c b/src/mesa/drivers/dri/r300/r300_state.c
index 0f3198e792..bae2f0f3cb 100644
--- a/src/mesa/drivers/dri/r300/r300_state.c
+++ b/src/mesa/drivers/dri/r300/r300_state.c
@@ -1519,29 +1519,6 @@ static void r300SetupRSUnit(GLcontext * ctx)
 		++fp_reg;
 	}
 
-	if (InputsRead & FRAG_BIT_WPOS) {
-		r300->hw.ri.cmd[R300_RI_INTERP_0 + tex_ip] |= R300_RS_SEL_S(0) | R300_RS_SEL_T(1) | R300_RS_SEL_R(2) | R300_RS_SEL_Q(3) | R300_RS_TEX_PTR(rs_tex_count);
-		r300->hw.rr.cmd[R300_RR_INST_0 + tex_ip] |= R300_RS_INST_TEX_ID(tex_ip) | R300_RS_INST_TEX_CN_WRITE | R300_RS_INST_TEX_ADDR(fp_reg);
-		InputsRead &= ~FRAG_BIT_WPOS;
-		rs_tex_count += 4;
-		++tex_ip;
-		++fp_reg;
-	}
-
-	if (InputsRead & FRAG_BIT_FOGC) {
-		if (R300_OUTPUTS_WRITTEN_TEST(OutputsWritten, VERT_RESULT_FOGC, _TNL_ATTRIB_FOG)) {
-			r300->hw.ri.cmd[R300_RI_INTERP_0 + tex_ip] |= R300_RS_SEL_S(0) | R300_RS_SEL_T(R300_RS_SEL_K0) | R300_RS_SEL_R(R300_RS_SEL_K0);
-			r300->hw.ri.cmd[R300_RI_INTERP_0 + tex_ip] |= R300_RS_SEL_Q(R300_RS_SEL_K1) | R300_RS_TEX_PTR(rs_tex_count);
-			r300->hw.rr.cmd[R300_RR_INST_0 + tex_ip] |= R300_RS_INST_TEX_ID(tex_ip) | R300_RS_INST_TEX_CN_WRITE | R300_RS_INST_TEX_ADDR(fp_reg);
-			InputsRead &= ~FRAG_BIT_FOGC;
-			rs_tex_count += 4;
-			++tex_ip;
-			++fp_reg;
-		} else {
-			WARN_ONCE("fragprog wants fogc, vp doesn't provide it\n");
-		}
-	}
-
 	/* Setup default color if no color or tex was set */
 	if (rs_tex_count == 0 && col_ip == 0) {
 		r300->hw.rr.cmd[R300_RR_INST_0] = R300_RS_INST_COL_ID(0) | R300_RS_INST_COL_ADDR(0);
@@ -1640,36 +1617,6 @@ static void r500SetupRSUnit(GLcontext * ctx)
 		++fp_reg;
 	}
 
-	if (InputsRead & FRAG_BIT_WPOS) {
-		r300->hw.ri.cmd[R300_RI_INTERP_0 + tex_ip] |= ((rs_tex_count + 0) << R500_RS_IP_TEX_PTR_S_SHIFT) |
-			((rs_tex_count + 1) << R500_RS_IP_TEX_PTR_T_SHIFT) |
-			((rs_tex_count + 2) << R500_RS_IP_TEX_PTR_R_SHIFT) |
-			((rs_tex_count + 3) << R500_RS_IP_TEX_PTR_Q_SHIFT);
-
-		r300->hw.rr.cmd[R300_RR_INST_0 + tex_ip] |= R500_RS_INST_TEX_ID(tex_ip) | R500_RS_INST_TEX_CN_WRITE | R500_RS_INST_TEX_ADDR(fp_reg);
-		InputsRead &= ~FRAG_BIT_WPOS;
-		rs_tex_count += 4;
-		++tex_ip;
-		++fp_reg;
-	}
-
-	if (InputsRead & FRAG_BIT_FOGC) {
-		if (R300_OUTPUTS_WRITTEN_TEST(OutputsWritten, VERT_RESULT_FOGC, _TNL_ATTRIB_FOG)) {
-			r300->hw.ri.cmd[R300_RI_INTERP_0 + tex_ip] |= (rs_tex_count << R500_RS_IP_TEX_PTR_S_SHIFT) |
-				(R500_RS_IP_PTR_K0 << R500_RS_IP_TEX_PTR_T_SHIFT) |
-				(R500_RS_IP_PTR_K0 << R500_RS_IP_TEX_PTR_R_SHIFT) |
-				(R500_RS_IP_PTR_K1 << R500_RS_IP_TEX_PTR_Q_SHIFT);
-
-			r300->hw.rr.cmd[R300_RR_INST_0 + tex_ip] |= R500_RS_INST_TEX_ID(tex_ip) | R500_RS_INST_TEX_CN_WRITE | R500_RS_INST_TEX_ADDR(fp_reg);
-			InputsRead &= ~FRAG_BIT_FOGC;
-			rs_tex_count += 4;
-			++tex_ip;
-			++fp_reg;
-		} else {
-			WARN_ONCE("fragprog wants fogc, vp doesn't provide it\n");
-		}
-	}
-
 	/* Setup default color if no color or tex was set */
 	if (rs_tex_count == 0 && col_ip == 0) {
 		r300->hw.rr.cmd[R300_RR_INST_0] = R500_RS_INST_COL_ID(0) | R500_RS_INST_COL_ADDR(0);
diff --git a/src/mesa/drivers/dri/r300/r300_swtcl.c b/src/mesa/drivers/dri/r300/r300_swtcl.c
index db4ccce6f1..d73508d36e 100644
--- a/src/mesa/drivers/dri/r300/r300_swtcl.c
+++ b/src/mesa/drivers/dri/r300/r300_swtcl.c
@@ -192,31 +192,23 @@ void r300ChooseSwtclVertexFormat(GLcontext *ctx, GLuint *_InputsRead,  GLuint *_
 		}
 	}
 
-	/* RS can't put fragment position on the pixel stack, so stuff it in texcoord if needed */
-	if (fp_reads & FRAG_BIT_WPOS) {
-		if (first_free_tex >= ctx->Const.MaxTextureUnits) {
-			fprintf(stderr, "\tout of free texcoords to write w pos\n");
-			_mesa_exit(-1);
-		}
+	if (rmesa->selected_fp->wpos_attr != FRAG_ATTRIB_MAX) {
+		int tex_id = rmesa->selected_fp->wpos_attr - FRAG_ATTRIB_TEX0;
 
-		InputsRead |= 1 << (VERT_ATTRIB_TEX0 + first_free_tex);
-		OutputsWritten |= 1 << (VERT_RESULT_TEX0 + first_free_tex);
-		EMIT_ATTR( _TNL_ATTRIB_POS, EMIT_4F );
-		ADD_ATTR(VERT_ATTRIB_POS, R300_DATA_TYPE_FLOAT_4, SWTCL_OVM_TEX(first_free_tex), SWIZZLE_XYZW, MASK_XYZW, 0);
-		++first_free_tex;
+		VB->AttribPtr[VERT_ATTRIB_TEX0 + tex_id] = VB->AttribPtr[VERT_ATTRIB_POS];
+		RENDERINPUTS_SET(tnl->render_inputs_bitset, _TNL_ATTRIB_TEX0 + tex_id);
 	}
 
-	if (fp_reads & FRAG_BIT_FOGC) {
-		if (first_free_tex >= ctx->Const.MaxTextureUnits) {
-			fprintf(stderr, "\tout of free texcoords to write fog coordinate\n");
-			_mesa_exit(-1);
-		}
+	if (rmesa->selected_fp->fog_attr != FRAG_ATTRIB_MAX) {
+		int tex_id = rmesa->selected_fp->fog_attr - FRAG_ATTRIB_TEX0;
 
-		InputsRead |= 1 << VERT_ATTRIB_FOG;
-		OutputsWritten |= 1 << VERT_RESULT_FOGC;
-		GLuint swiz = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_ZERO, SWIZZLE_ZERO, SWIZZLE_ZERO);
-		EMIT_ATTR( _TNL_ATTRIB_FOG, EMIT_1F );
-		ADD_ATTR(VERT_ATTRIB_FOG, R300_DATA_TYPE_FLOAT_1, SWTCL_OVM_TEX(first_free_tex), swiz, MASK_XYZW, 0);
+		VB->AttribPtr[VERT_ATTRIB_TEX0 + tex_id] = VB->AttribPtr[VERT_ATTRIB_FOG];
+		RENDERINPUTS_SET(tnl->render_inputs_bitset, _TNL_ATTRIB_TEX0 + tex_id);
+	}
+
+	if (first_free_tex >= ctx->Const.MaxTextureUnits) {
+		fprintf(stderr, "\tout of free texcoords to write fog coordinate\n");
+		_mesa_exit(-1);
 	}
 
 	R300_NEWPRIM(rmesa);
@@ -497,11 +489,13 @@ void r300RenderStart(GLcontext *ctx)
 	r300ContextPtr rmesa = R300_CONTEXT( ctx );
 
 	r300ChooseRenderState(ctx);
+
+	r300UpdateShaders(rmesa);
+
 	r300PrepareVertices(ctx);
 
 	r300ValidateBuffers(ctx);
 
-	r300UpdateShaders(rmesa);
 	r300UpdateShaderStates(rmesa);
 
 	r300EmitCacheFlush(rmesa);
diff --git a/src/mesa/drivers/dri/r300/r300_vertprog.c b/src/mesa/drivers/dri/r300/r300_vertprog.c
index 0b04830350..6f9b4db40f 100644
--- a/src/mesa/drivers/dri/r300/r300_vertprog.c
+++ b/src/mesa/drivers/dri/r300/r300_vertprog.c
@@ -1223,8 +1223,7 @@ void r300TranslateVertexShader(struct r300_vertex_program *vp)
 	}
 }
 
-static void insert_wpos(struct r300_vertex_program *vp, struct gl_program *prog,
-			GLuint temp_index)
+static void insert_wpos(struct gl_program *prog, GLuint temp_index, int tex_id)
 {
 	struct prog_instruction *vpi;
 
@@ -1248,7 +1247,7 @@ static void insert_wpos(struct r300_vertex_program *vp, struct gl_program *prog,
 	vpi->Opcode = OPCODE_MOV;
 
 	vpi->DstReg.File = PROGRAM_OUTPUT;
-	vpi->DstReg.Index = VERT_RESULT_TEX0 + vp->wpos_idx;
+	vpi->DstReg.Index = VERT_RESULT_TEX0 + tex_id;
 	vpi->DstReg.WriteMask = WRITEMASK_XYZW;
 	vpi->DstReg.CondMask = COND_TR;
 
@@ -1259,12 +1258,9 @@ static void insert_wpos(struct r300_vertex_program *vp, struct gl_program *prog,
 	++vpi;
 
 	vpi->Opcode = OPCODE_END;
-
-	prog->OutputsWritten |= 1 << (VERT_RESULT_TEX0 + vp->wpos_idx);
 }
 
-static void pos_as_texcoord(struct r300_vertex_program *vp,
-			    struct gl_program *prog)
+static void pos_as_texcoord(struct gl_program *prog, int tex_id)
 {
 	struct prog_instruction *vpi;
 	GLuint tempregi = prog->NumTemporaries;
@@ -1278,7 +1274,37 @@ static void pos_as_texcoord(struct r300_vertex_program *vp,
 		}
 	}
 
-	insert_wpos(vp, prog, tempregi);
+	insert_wpos(prog, tempregi, tex_id);
+
+	prog->OutputsWritten |= 1 << (VERT_RESULT_TEX0 + tex_id);
+}
+
+/**
+ @TODO
+  We can put X001 swizzle only if input components are directly mapped from output components.
+  For some insts we need to skip source swizzles and add: MOV OUTPUT[fog_attr].yzw, CONST[0].0001
+ */
+static void fog_as_texcoord(struct gl_program *prog, int tex_id)
+{
+	struct prog_instruction *vpi;
+	int i;
+
+	vpi = prog->Instructions;
+	while (vpi->Opcode != OPCODE_END) {
+		if (vpi->DstReg.File == PROGRAM_OUTPUT && vpi->DstReg.Index == VERT_RESULT_FOGC) {
+			vpi->DstReg.Index = VERT_RESULT_TEX0 + tex_id;
+			vpi->DstReg.WriteMask = WRITEMASK_XYZW;
+
+			for (i = 0; i < _mesa_num_inst_src_regs(vpi->Opcode); ++i) {
+				vpi->SrcReg[i].Swizzle = combineSwizzles(vpi->SrcReg[i].Swizzle, SWIZZLE_X, SWIZZLE_ZERO, SWIZZLE_ZERO, SWIZZLE_ONE);
+			}
+		}
+
+		++vpi;
+	}
+
+	prog->OutputsWritten &= ~(1 << VERT_RESULT_FOGC);
+	prog->OutputsWritten |= 1 << (VERT_RESULT_TEX0 + tex_id);
 }
 
 static int translateABS(struct gl_program *prog, int pos)
@@ -1513,16 +1539,15 @@ static void addArtificialOutputs(GLcontext *ctx, struct gl_program *prog)
 
 static struct r300_vertex_program *build_program(GLcontext *ctx,
 						 struct r300_vertex_program_key *wanted_key,
-						 const struct gl_vertex_program *mesa_vp,
-						 GLint wpos_idx)
+						 const struct gl_vertex_program *mesa_vp)
 {
+	r300ContextPtr r300 = R300_CONTEXT(ctx);
 	struct r300_vertex_program *vp;
 	struct gl_program *prog;
 
 	vp = _mesa_calloc(sizeof(*vp));
 	vp->Base = (struct gl_vertex_program *) _mesa_clone_program(ctx, &mesa_vp->Base);
 	_mesa_memcpy(&vp->key, wanted_key, sizeof(vp->key));
-	vp->wpos_idx = wpos_idx;
 
 	prog = &vp->Base->Base;
 
@@ -1532,12 +1557,16 @@ static struct r300_vertex_program *build_program(GLcontext *ctx,
 		fflush(stdout);
 	}
 
-	if (mesa_vp->IsPositionInvariant) {
+	if (vp->Base->IsPositionInvariant) {
 		_mesa_insert_mvp_code(ctx, vp->Base);
 	}
 
-	if (wpos_idx > -1) {
-		pos_as_texcoord(vp, prog);
+	if (r300->selected_fp->wpos_attr != FRAG_ATTRIB_MAX) {
+		pos_as_texcoord(&vp->Base->Base, r300->selected_fp->wpos_attr - FRAG_ATTRIB_TEX0);
+	}
+
+	if (r300->selected_fp->fog_attr != FRAG_ATTRIB_MAX) {
+		fog_as_texcoord(&vp->Base->Base, r300->selected_fp->fog_attr - FRAG_ATTRIB_TEX0);
 	}
 
 	addArtificialOutputs(ctx, prog);
@@ -1562,34 +1591,20 @@ struct r300_vertex_program * r300SelectVertexShader(GLcontext *ctx)
 	struct r300_vertex_program_key wanted_key = { 0 };
 	struct r300_vertex_program_cont *vpc;
 	struct r300_vertex_program *vp;
-	GLint wpos_idx;
 
 	vpc = (struct r300_vertex_program_cont *)ctx->VertexProgram._Current;
 	wanted_key.FpReads = r300->selected_fp->Base->InputsRead;
+	wanted_key.FogAttr = r300->selected_fp->fog_attr;
+	wanted_key.WPosAttr = r300->selected_fp->wpos_attr;
 
-	for (vp = vpc->progs; vp; vp = vp->next)
+	for (vp = vpc->progs; vp; vp = vp->next) {
 		if (_mesa_memcmp(&vp->key, &wanted_key, sizeof(wanted_key))
 		    == 0) {
 			return r300->selected_vp = vp;
 		}
-
-	wpos_idx = -1;
-	if (wanted_key.FpReads & FRAG_BIT_WPOS) {
-		GLint i;
-
-		for (i = 0; i < ctx->Const.MaxTextureUnits; i++)
-			if (!(wanted_key.FpReads & (FRAG_BIT_TEX(i))))
-				break;
-
-			if (i == ctx->Const.MaxTextureUnits) {
-				fprintf(stderr, "\tno free texcoord found\n");
-				_mesa_exit(-1);
-			}
-
-			wpos_idx = i;
 	}
 
-	vp = build_program(ctx, &wanted_key, &vpc->mesa_program, wpos_idx);
+	vp = build_program(ctx, &wanted_key, &vpc->mesa_program);
 	vp->next = vpc->progs;
 	vpc->progs = vp;
 
-- 
cgit v1.2.3


From bdc8a95fc906dcc617995479fbffbf8fdd1f4e34 Mon Sep 17 00:00:00 2001
From: Maciej Cencora <m.cencora@gmail.com>
Date: Sat, 4 Jul 2009 23:38:59 +0200
Subject: r300: removed unnecessary params

We don't have check which attributes are used by fragment program - it's already done by NQSSADCE.
---
 src/mesa/drivers/dri/r300/r300_emit.c  | 10 +++++-----
 src/mesa/drivers/dri/r300/r300_emit.h  |  4 ++--
 src/mesa/drivers/dri/r300/r300_state.c |  4 ++--
 3 files changed, 9 insertions(+), 9 deletions(-)

(limited to 'src/mesa/drivers/dri/r300/r300_state.c')

diff --git a/src/mesa/drivers/dri/r300/r300_emit.c b/src/mesa/drivers/dri/r300/r300_emit.c
index 707d1284ed..feb3370f37 100644
--- a/src/mesa/drivers/dri/r300/r300_emit.c
+++ b/src/mesa/drivers/dri/r300/r300_emit.c
@@ -81,17 +81,17 @@ GLuint r300VAPInputCntl1(GLcontext * ctx, GLuint InputsRead)
 	return vic_1;
 }
 
-GLuint r300VAPOutputCntl0(GLcontext * ctx, GLuint vp_writes, GLuint fp_reads)
+GLuint r300VAPOutputCntl0(GLcontext * ctx, GLuint vp_writes)
 {
 	GLuint ret = 0;
 
 	if (vp_writes & (1 << VERT_RESULT_HPOS))
 		ret |= R300_VAP_OUTPUT_VTX_FMT_0__POS_PRESENT;
 
-	if (vp_writes & (1 << VERT_RESULT_COL0) && fp_reads & FRAG_BIT_COL0)
+	if (vp_writes & (1 << VERT_RESULT_COL0))
 		ret |= R300_VAP_OUTPUT_VTX_FMT_0__COLOR_0_PRESENT;
 
-	if (vp_writes & (1 << VERT_RESULT_COL1) && fp_reads & FRAG_BIT_COL1)
+	if (vp_writes & (1 << VERT_RESULT_COL1))
 		ret |= R300_VAP_OUTPUT_VTX_FMT_0__COLOR_1_PRESENT;
 
 	/* Two sided lighting works only if all 4 colors are written */
@@ -105,12 +105,12 @@ GLuint r300VAPOutputCntl0(GLcontext * ctx, GLuint vp_writes, GLuint fp_reads)
 	return ret;
 }
 
-GLuint r300VAPOutputCntl1(GLcontext * ctx, GLuint vp_writes, GLuint fp_reads)
+GLuint r300VAPOutputCntl1(GLcontext * ctx, GLuint vp_writes)
 {
 	GLuint i, ret = 0, first_free_texcoord = 0;
 
 	for (i = 0; i < ctx->Const.MaxTextureUnits; i++) {
-		if (vp_writes & (1 << (VERT_RESULT_TEX0 + i)) && fp_reads & FRAG_BIT_TEX(i)) {
+		if (vp_writes & (1 << (VERT_RESULT_TEX0 + i))) {
 			ret |= (4 << (3 * first_free_texcoord));
 			++first_free_texcoord;
 		}
diff --git a/src/mesa/drivers/dri/r300/r300_emit.h b/src/mesa/drivers/dri/r300/r300_emit.h
index 2fb8b82d3a..3f8c60ffae 100644
--- a/src/mesa/drivers/dri/r300/r300_emit.h
+++ b/src/mesa/drivers/dri/r300/r300_emit.h
@@ -225,7 +225,7 @@ extern void r300EmitCacheFlush(r300ContextPtr rmesa);
 
 extern GLuint r300VAPInputCntl0(GLcontext * ctx, GLuint InputsRead);
 extern GLuint r300VAPInputCntl1(GLcontext * ctx, GLuint InputsRead);
-extern GLuint r300VAPOutputCntl0(GLcontext * ctx, GLuint vp_writes, GLuint fp_reads);
-extern GLuint r300VAPOutputCntl1(GLcontext * ctx, GLuint vp_writes, GLuint fp_reads);
+extern GLuint r300VAPOutputCntl0(GLcontext * ctx, GLuint vp_writes);
+extern GLuint r300VAPOutputCntl1(GLcontext * ctx, GLuint vp_writes);
 
 #endif
diff --git a/src/mesa/drivers/dri/r300/r300_state.c b/src/mesa/drivers/dri/r300/r300_state.c
index bae2f0f3cb..4d504d14e3 100644
--- a/src/mesa/drivers/dri/r300/r300_state.c
+++ b/src/mesa/drivers/dri/r300/r300_state.c
@@ -2225,8 +2225,8 @@ void r300SetupVAP(GLcontext *ctx, GLuint InputsRead, GLuint OutputsWritten)
 
 	rmesa->hw.vic.cmd[R300_VIC_CNTL_0] = r300VAPInputCntl0(ctx, InputsRead);
 	rmesa->hw.vic.cmd[R300_VIC_CNTL_1] = r300VAPInputCntl1(ctx, InputsRead);
-	rmesa->hw.vof.cmd[R300_VOF_CNTL_0] = r300VAPOutputCntl0(ctx, OutputsWritten, rmesa->selected_fp->Base->InputsRead);
-	rmesa->hw.vof.cmd[R300_VOF_CNTL_1] = r300VAPOutputCntl1(ctx, OutputsWritten, rmesa->selected_fp->Base->InputsRead);
+	rmesa->hw.vof.cmd[R300_VOF_CNTL_0] = r300VAPOutputCntl0(ctx, OutputsWritten);
+	rmesa->hw.vof.cmd[R300_VOF_CNTL_1] = r300VAPOutputCntl1(ctx, OutputsWritten);
 }
 
 void r300UpdateShaderStates(r300ContextPtr rmesa)
-- 
cgit v1.2.3