#include "radeon_compiler.h"
-#include "shader/prog_parameter.h"
-#include "shader/prog_print.h"
-#include "shader/prog_statevars.h"
+#include <stdio.h>
-#include "radeon_nqssadce.h"
+#include "radeon_dataflow.h"
#include "radeon_program_alu.h"
#include "r300_fragprog.h"
#include "r300_fragprog_swizzle.h"
#include "r500_fragprog.h"
-static void nqssadce_init(struct nqssadce_state* s)
+static void dataflow_outputs_mark_use(void * userdata, void * data,
+ void (*callback)(void *, unsigned int, unsigned int))
{
- s->Outputs[FRAG_RESULT_COLOR].Sourced = WRITEMASK_XYZW;
- s->Outputs[FRAG_RESULT_DEPTH].Sourced = WRITEMASK_W;
+ struct r300_fragment_program_compiler * c = userdata;
+ callback(data, c->OutputColor, RC_MASK_XYZW);
+ callback(data, c->OutputDepth, RC_MASK_W);
}
-/**
- * Transform the program to support fragment.position.
- *
- * Introduce a small fragment at the start of the program that will be
- * the only code that directly reads the FRAG_ATTRIB_WPOS input.
- * All other code pieces that reference that input will be rewritten
- * to read from a newly allocated temporary.
- *
- */
-static void insert_WPOS_trailer(struct r300_fragment_program_compiler *compiler)
-{
- GLuint InputsRead = compiler->program->InputsRead;
-
- if (!(InputsRead & FRAG_BIT_WPOS)) {
- compiler->code->wpos_attr = FRAG_ATTRIB_MAX;
- return;
- }
-
- static gl_state_index tokens[STATE_LENGTH] = {
- STATE_INTERNAL, STATE_R300_WINDOW_DIMENSION, 0, 0, 0
- };
- 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->program->InputsRead = InputsRead;
- compiler->code->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;
-
- fpi[i].DstReg.File = PROGRAM_TEMPORARY;
- fpi[i].DstReg.Index = tempregi;
- fpi[i].DstReg.WriteMask = WRITEMASK_W;
- fpi[i].DstReg.CondMask = COND_TR;
-
- fpi[i].SrcReg[0].File = PROGRAM_INPUT;
- fpi[i].SrcReg[0].Index = compiler->code->wpos_attr;
- fpi[i].SrcReg[0].Swizzle = SWIZZLE_WWWW;
- i++;
-
- fpi[i].Opcode = OPCODE_MUL;
-
- fpi[i].DstReg.File = PROGRAM_TEMPORARY;
- fpi[i].DstReg.Index = tempregi;
- fpi[i].DstReg.WriteMask = WRITEMASK_XYZ;
- fpi[i].DstReg.CondMask = COND_TR;
-
- fpi[i].SrcReg[0].File = PROGRAM_INPUT;
- fpi[i].SrcReg[0].Index = compiler->code->wpos_attr;
- fpi[i].SrcReg[0].Swizzle = SWIZZLE_XYZW;
-
- fpi[i].SrcReg[1].File = PROGRAM_TEMPORARY;
- fpi[i].SrcReg[1].Index = tempregi;
- fpi[i].SrcReg[1].Swizzle = SWIZZLE_WWWW;
- i++;
-
- /* viewport transformation */
- window_index = _mesa_add_state_reference(compiler->program->Parameters, tokens);
-
- fpi[i].Opcode = OPCODE_MAD;
-
- fpi[i].DstReg.File = PROGRAM_TEMPORARY;
- fpi[i].DstReg.Index = tempregi;
- fpi[i].DstReg.WriteMask = WRITEMASK_XYZ;
- fpi[i].DstReg.CondMask = COND_TR;
-
- fpi[i].SrcReg[0].File = PROGRAM_TEMPORARY;
- fpi[i].SrcReg[0].Index = tempregi;
- fpi[i].SrcReg[0].Swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_ZERO);
-
- fpi[i].SrcReg[1].File = PROGRAM_STATE_VAR;
- fpi[i].SrcReg[1].Index = window_index;
- fpi[i].SrcReg[1].Swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_ZERO);
-
- fpi[i].SrcReg[2].File = PROGRAM_STATE_VAR;
- fpi[i].SrcReg[2].Index = window_index;
- fpi[i].SrcReg[2].Swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_ZERO);
- i++;
-
- for (; i < compiler->program->NumInstructions; ++i) {
- int reg;
- for (reg = 0; reg < 3; reg++) {
- if (fpi[i].SrcReg[reg].File == PROGRAM_INPUT &&
- fpi[i].SrcReg[reg].Index == FRAG_ATTRIB_WPOS) {
- fpi[i].SrcReg[reg].File = PROGRAM_TEMPORARY;
- fpi[i].SrcReg[reg].Index = tempregi;
- }
- }
- }
-}
-
-
-/**
- * Rewrite fragment.fogcoord to use a texture coordinate slot.
- * Note that fogcoord is forced into an X001 pattern, and this enforcement
- * is done here.
- *
- * See also the counterpart rewriting for vertex programs.
- */
-static void rewriteFog(struct r300_fragment_program_compiler *compiler)
-{
- struct rX00_fragment_program_code *code = compiler->code;
- GLuint InputsRead = compiler->program->InputsRead;
- int i;
-
- if (!(InputsRead & FRAG_BIT_FOGC)) {
- code->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;
- compiler->program->InputsRead = InputsRead;
- code->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 = code->fog_attr;
- inst->SrcReg[i].Swizzle = combine_swizzles(
- MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_ZERO, SWIZZLE_ZERO, SWIZZLE_ONE),
- inst->SrcReg[i].Swizzle);
- }
- }
- ++inst;
- }
- }
-}
-
-
static void rewrite_depth_out(struct r300_fragment_program_compiler * c)
{
struct rc_instruction *rci;
for (rci = c->Base.Program.Instructions.Next; rci != &c->Base.Program.Instructions; rci = rci->Next) {
- struct prog_instruction * inst = &rci->I;
+ struct rc_sub_instruction * inst = &rci->I;
- if (inst->DstReg.File != PROGRAM_OUTPUT || inst->DstReg.Index != FRAG_RESULT_DEPTH)
+ if (inst->DstReg.File != RC_FILE_OUTPUT || inst->DstReg.Index != c->OutputDepth)
continue;
- if (inst->DstReg.WriteMask & WRITEMASK_Z) {
- inst->DstReg.WriteMask = WRITEMASK_W;
+ if (inst->DstReg.WriteMask & RC_MASK_Z) {
+ inst->DstReg.WriteMask = RC_MASK_W;
} else {
inst->DstReg.WriteMask = 0;
continue;
}
switch (inst->Opcode) {
- case OPCODE_FRC:
- case OPCODE_MOV:
- inst->SrcReg[0] = lmul_swizzle(SWIZZLE_ZZZZ, inst->SrcReg[0]);
+ case RC_OPCODE_FRC:
+ case RC_OPCODE_MOV:
+ inst->SrcReg[0] = lmul_swizzle(RC_SWIZZLE_ZZZZ, inst->SrcReg[0]);
break;
- case OPCODE_ADD:
- case OPCODE_MAX:
- case OPCODE_MIN:
- case OPCODE_MUL:
- inst->SrcReg[0] = lmul_swizzle(SWIZZLE_ZZZZ, inst->SrcReg[0]);
- inst->SrcReg[1] = lmul_swizzle(SWIZZLE_ZZZZ, inst->SrcReg[1]);
+ case RC_OPCODE_ADD:
+ case RC_OPCODE_MAX:
+ case RC_OPCODE_MIN:
+ case RC_OPCODE_MUL:
+ inst->SrcReg[0] = lmul_swizzle(RC_SWIZZLE_ZZZZ, inst->SrcReg[0]);
+ inst->SrcReg[1] = lmul_swizzle(RC_SWIZZLE_ZZZZ, inst->SrcReg[1]);
break;
- case OPCODE_CMP:
- case OPCODE_MAD:
- inst->SrcReg[0] = lmul_swizzle(SWIZZLE_ZZZZ, inst->SrcReg[0]);
- inst->SrcReg[1] = lmul_swizzle(SWIZZLE_ZZZZ, inst->SrcReg[1]);
- inst->SrcReg[2] = lmul_swizzle(SWIZZLE_ZZZZ, inst->SrcReg[2]);
+ case RC_OPCODE_CMP:
+ case RC_OPCODE_MAD:
+ inst->SrcReg[0] = lmul_swizzle(RC_SWIZZLE_ZZZZ, inst->SrcReg[0]);
+ inst->SrcReg[1] = lmul_swizzle(RC_SWIZZLE_ZZZZ, inst->SrcReg[1]);
+ inst->SrcReg[2] = lmul_swizzle(RC_SWIZZLE_ZZZZ, inst->SrcReg[2]);
break;
default:
// Scalar instructions needn't be reswizzled
void r3xx_compile_fragment_program(struct r300_fragment_program_compiler* c)
{
- if (c->Base.Debug) {
- fflush(stdout);
- _mesa_printf("Fragment Program: Initial program:\n");
- _mesa_print_program(c->program);
- fflush(stdout);
- }
-
- insert_WPOS_trailer(c);
-
- rewriteFog(c);
-
- rc_mesa_to_rc_program(&c->Base, c->program);
-
rewrite_depth_out(c);
if (c->is_r500) {
{ &radeonTransformTrigScale, 0 }
};
radeonLocalTransform(&c->Base, 4, transformations);
+
+ c->Base.SwizzleCaps = &r500_swizzle_caps;
} else {
struct radeon_program_transformation transformations[] = {
{ &r300_transform_TEX, c },
{ &radeonTransformTrigSimple, 0 }
};
radeonLocalTransform(&c->Base, 3, transformations);
+
+ c->Base.SwizzleCaps = &r300_swizzle_caps;
}
if (c->Base.Debug) {
- _mesa_printf("Fragment Program: After native rewrite:\n");
+ fprintf(stderr, "Fragment Program: After native rewrite:\n");
rc_print_program(&c->Base.Program);
- fflush(stdout);
+ fflush(stderr);
}
- if (c->is_r500) {
- struct radeon_nqssadce_descr nqssadce = {
- .Init = &nqssadce_init,
- .IsNativeSwizzle = &r500FPIsNativeSwizzle,
- .BuildSwizzle = &r500FPBuildSwizzle
- };
- radeonNqssaDce(&c->Base, &nqssadce, 0);
- } else {
- struct radeon_nqssadce_descr nqssadce = {
- .Init = &nqssadce_init,
- .IsNativeSwizzle = &r300FPIsNativeSwizzle,
- .BuildSwizzle = &r300FPBuildSwizzle
- };
- radeonNqssaDce(&c->Base, &nqssadce, 0);
+ rc_dataflow_deadcode(&c->Base, &dataflow_outputs_mark_use, c);
+
+ if (c->Base.Debug) {
+ fprintf(stderr, "Fragment Program: After deadcode:\n");
+ rc_print_program(&c->Base.Program);
+ fflush(stderr);
}
+ rc_dataflow_swizzles(&c->Base);
+
if (c->Base.Debug) {
- _mesa_printf("Compiler: after NqSSA-DCE:\n");
+ fprintf(stderr, "Compiler: after dataflow passes:\n");
rc_print_program(&c->Base.Program);
- fflush(stdout);
+ fflush(stderr);
}
if (c->is_r500) {
projects / mesa.git / blobdiff