r300g: add a new debug option which disables compiler optimizations

Those are:
- dead-code elimination
- constant folding
- peephole (mainly copy propagation)
- register allocation

There are some bugs which I need to track down.

Also fix up the descriptions of all the debug options.

src/gallium/drivers/r300/r300_debug.c

@@ -27,24 +27,25 @@
 #include <stdio.h>
 
 static const struct debug_named_value debug_options[] = {
-    { "fp", DBG_FP, "Fragment program handling (for debugging)" },
-    { "vp", DBG_VP, "Vertex program handling (for debugging)" },
-    { "draw", DBG_DRAW, "Draw calls (for debugging)" },
-    { "swtcl", DBG_SWTCL, "SWTCL-specific info (for debugging)" },
-    { "rsblock", DBG_RS_BLOCK, "Rasterizer registers (for debugging)" },
-    { "psc", DBG_PSC, "Vertex stream registers (for debugging)" },
-    { "tex", DBG_TEX, "Textures (for debugging)" },
-    { "texalloc", DBG_TEXALLOC, "Texture allocation (for debugging)" },
-    { "fall", DBG_FALL, "Fallbacks (for debugging)" },
-    { "rs", DBG_RS, "Rasterizer (for debugging)" },
-    { "fb", DBG_FB, "Framebuffer (for debugging)" },
-    { "cbzb", DBG_CBZB, "Fast color clear info (for debugging)" },
-    { "fakeocc", DBG_FAKE_OCC, "Use fake occlusion queries (for debugging)" },
-    { "anisohq", DBG_ANISOHQ, "High quality anisotropic filtering (for benchmarking)" },
-    { "notiling", DBG_NO_TILING, "Disable tiling (for benchmarking)" },
-    { "noimmd", DBG_NO_IMMD, "Disable immediate mode (for benchmarking)" },
-    { "stats", DBG_STATS, "Gather statistics" },
-    { "hyperz", DBG_HYPERZ, "HyperZ (for debugging)" },
+    { "fp", DBG_FP, "Log fragment program compilation" },
+    { "vp", DBG_VP, "Log bertex program compilation" },
+    { "draw", DBG_DRAW, "Log draw calls" },
+    { "swtcl", DBG_SWTCL, "Log SWTCL-specific info" },
+    { "rsblock", DBG_RS_BLOCK, "Log rasterizer registers" },
+    { "psc", DBG_PSC, "Log vertex stream registers" },
+    { "tex", DBG_TEX, "Log basic info about textures" },
+    { "texalloc", DBG_TEXALLOC, "Log texture mipmap tree info" },
+    { "fall", DBG_FALL, "Log fallbacks" },
+    { "rs", DBG_RS, "Log rasterizer" },
+    { "fb", DBG_FB, "Log framebuffer" },
+    { "cbzb", DBG_CBZB, "Log fast color clear info" },
+    { "stats", DBG_STATS, "Log emission statistics" },
+    { "hyperz", DBG_HYPERZ, "Log HyperZ info" },
+    { "fakeocc", DBG_FAKE_OCC, "Use fake occlusion queries" },
+    { "anisohq", DBG_ANISOHQ, "Use high quality anisotropic filtering" },
+    { "notiling", DBG_NO_TILING, "Disable tiling" },
+    { "noimmd", DBG_NO_IMMD, "Disable immediate mode" },
+    { "noopt", DBG_NO_OPT, "Disable shader optimizations" },
 
     /* must be last */
     DEBUG_NAMED_VALUE_END

src/gallium/drivers/r300/r300_fs.c

@@ -385,6 +385,7 @@ static void r300_translate_fragment_shader(
     compiler.code = &shader->code;
     compiler.state = shader->compare_state;
     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.max_temp_regs = compiler.Base.is_r500 ? 128 : 32;
     compiler.Base.max_constants = compiler.Base.is_r500 ? 256 : 32;

src/gallium/drivers/r300/r300_screen.h

@@ -97,6 +97,7 @@ r300_winsys_screen(struct pipe_screen *screen) {
 #define DBG_NO_TILING   (1 << 17)
 #define DBG_NO_IMMD     (1 << 18)
 #define DBG_FAKE_OCC    (1 << 19)
+#define DBG_NO_OPT	(1 << 20)
 /* Statistics. */
 #define DBG_STATS       (1 << 24)
 /*@}*/

src/gallium/drivers/r300/r300_vs.c

@@ -205,6 +205,7 @@ void r300_translate_vertex_shader(struct r300_context *r300,
     compiler.code = &vs->code;
     compiler.UserData = vs;
     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.max_temp_regs = 32;
     compiler.Base.max_constants = 256;

src/mesa/drivers/dri/r300/compiler/r3xx_fragprog.c

@@ -94,6 +94,7 @@ void r3xx_compile_fragment_program(struct r300_fragment_program_compiler* c)
 {
 	int is_r500 = c->Base.is_r500;
 	int kill_consts = c->Base.remove_unused_constants;
+	int opt = !c->Base.disable_optimizations;
 
 	/* Lists of instruction transformations. */
 	struct radeon_program_transformation rewrite_tex[] = {
@@ -128,9 +129,9 @@ void r3xx_compile_fragment_program(struct r300_fragment_program_compiler* c)
 		{"transform TEX",		1, 1,		rc_local_transform,		rewrite_tex},
 		{"native rewrite",		1, is_r500,	rc_local_transform,		native_rewrite_r500},
 		{"native rewrite",		1, !is_r500,	rc_local_transform,		native_rewrite_r300},
-		{"deadcode",			1, 1,		rc_dataflow_deadcode,		dataflow_outputs_mark_use},
+		{"deadcode",			1, opt,		rc_dataflow_deadcode,		dataflow_outputs_mark_use},
 		{"emulate loops",		1, !is_r500,	rc_emulate_loops,		NULL},
-		{"dataflow optimize",		1, 1,		rc_optimize,			NULL},
+		{"dataflow optimize",		1, opt,		rc_optimize,			NULL},
 		{"dataflow swizzles",		1, 1,		rc_dataflow_swizzles,		NULL},
 		{"dead constants",		1, kill_consts, rc_remove_unused_constants,	&c->code->constants_remap_table},
 		/* This pass makes it easier for the scheduler to group TEX
@@ -139,7 +140,7 @@ void r3xx_compile_fragment_program(struct r300_fragment_program_compiler* c)
 		{"register rename",		1, !is_r500,	rc_rename_regs,			NULL},
 		{"pair translate",		1, 1,		rc_pair_translate,		NULL},
 		{"pair scheduling",		1, 1,		rc_pair_schedule,		NULL},
-		{"register allocation",		1, 1,		rc_pair_regalloc,		NULL},
+		{"register allocation",		1, opt,		rc_pair_regalloc,		NULL},
 		{"final code validation",	0, 1,		rc_validate_final_shader,	NULL},
 		{"machine code generation",	0, is_r500,	r500BuildFragmentProgramHwCode,	NULL},
 		{"machine code generation",	0, !is_r500,	r300BuildFragmentProgramHwCode,	NULL},

src/mesa/drivers/dri/r300/compiler/r3xx_vertprog.c

@@ -991,6 +991,7 @@ void r3xx_compile_vertex_program(struct r300_vertex_program_compiler *c)
 {
 	int is_r500 = c->Base.is_r500;
 	int kill_consts = c->Base.remove_unused_constants;
+	int opt = !c->Base.disable_optimizations;
 
 	/* Lists of instruction transformations. */
 	struct radeon_program_transformation alu_rewrite_r500[] = {
@@ -1029,12 +1030,12 @@ void r3xx_compile_vertex_program(struct r300_vertex_program_compiler *c)
 		{"native rewrite",		1, is_r500,	rc_local_transform,		alu_rewrite_r500},
 		{"native rewrite",		1, !is_r500,	rc_local_transform,		alu_rewrite_r300},
 		{"emulate modifiers",		1, !is_r500,	rc_local_transform,		emulate_modifiers},
-		{"deadcode",			1, 1,		rc_dataflow_deadcode,		dataflow_outputs_mark_used},
-		{"dataflow optimize",		1, 1,		rc_optimize,			NULL},
+		{"deadcode",			1, opt,		rc_dataflow_deadcode,		dataflow_outputs_mark_used},
+		{"dataflow optimize",		1, opt,		rc_optimize,			NULL},
 		/* This pass must be done after optimizations. */
 		{"source conflict resolve",	1, 1,		rc_local_transform,		resolve_src_conflicts},
 		{"dataflow swizzles",		1, 1,		rc_dataflow_swizzles,		NULL},
-		{"register allocation",		1, 1,		allocate_temporary_registers,	NULL},
+		{"register allocation",		1, opt,		allocate_temporary_registers,	NULL},
 		{"dead constants",		1, kill_consts, rc_remove_unused_constants,	&c->code->constants_remap_table},
 		{"final code validation",	0, 1,		rc_validate_final_shader,	NULL},
 		{"machine code generation",	0, 1,		translate_vertex_program,	NULL},

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 disable_optimizations:1;
 	unsigned max_temp_regs;
 	unsigned max_constants;
 	int max_alu_insts;

src/mesa/drivers/dri/r300/r300_blit.c

@@ -89,6 +89,7 @@ static void create_vertex_program(struct r300_context *r300)
     compiler.SetHwInputOutput = vp_ins_outs;
     compiler.code = &r300->blit.vp_code;
     compiler.Base.is_r500 = r300->radeon.radeonScreen->chip_family >= CHIP_FAMILY_RV515;
+    compiler.Base.disable_optimizations = 0;
     compiler.Base.has_half_swizzles = 0;
     compiler.Base.max_temp_regs = 32;
     compiler.Base.max_constants = 256;
@@ -124,6 +125,7 @@ static void create_fragment_program(struct r300_context *r300)
     compiler.OutputDepth = FRAG_RESULT_DEPTH;
     compiler.enable_shadow_ambient = GL_TRUE;
     compiler.Base.is_r500 = (r300->radeon.radeonScreen->chip_family >= CHIP_FAMILY_RV515);
+    compiler.Base.disable_optimizations = 0;
     compiler.Base.has_half_swizzles = 1;
     compiler.Base.max_temp_regs = (compiler.Base.is_r500) ? 128 : 32;
     compiler.Base.max_constants = compiler.Base.is_r500 ? 256 : 32;

src/mesa/drivers/dri/r300/r300_fragprog_common.c

@@ -220,6 +220,7 @@ static void translate_fragment_program(GLcontext *ctx, struct r300_fragment_prog
 	compiler.state = fp->state;
 	compiler.enable_shadow_ambient = GL_TRUE;
 	compiler.Base.is_r500 = (r300->radeon.radeonScreen->chip_family >= CHIP_FAMILY_RV515) ? GL_TRUE : GL_FALSE;
+	compiler.Base.disable_optimizations = 0;
 	compiler.Base.has_half_swizzles = 1;
 	compiler.Base.max_temp_regs = (compiler.Base.is_r500) ? 128 : 32;
 	compiler.Base.max_constants = compiler.Base.is_r500 ? 256 : 32;

src/mesa/drivers/dri/r300/r300_vertprog.c

@@ -245,6 +245,7 @@ static struct r300_vertex_program *build_program(GLcontext *ctx,
 	compiler.RequiredOutputs = compute_required_outputs(vp->Base, vp->key.FpReads);
 	compiler.SetHwInputOutput = &t_inputs_outputs;
 	compiler.Base.is_r500 = R300_CONTEXT(ctx)->radeon.radeonScreen->chip_family >= CHIP_FAMILY_RV515;
+	compiler.Base.disable_optimizations = 0;
 	compiler.Base.has_half_swizzles = 0;
 	compiler.Base.max_temp_regs = 32;
 	compiler.Base.max_constants = 256;