X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fmesa%2Fshader%2Fslang%2Fslang_emit.c;h=500112b6f67156340a1233c20867b2fbce9d2f4d;hb=e8d7db31e2a6784c765911233cb3d888f612837f;hp=cbc71f36f6bbe15bb85a1e45620e5979ccd133f4;hpb=dad97b4688ea25caf15cae66194db6ddbb98e936;p=mesa.git diff --git a/src/mesa/shader/slang/slang_emit.c b/src/mesa/shader/slang/slang_emit.c index cbc71f36f6b..500112b6f67 100644 --- a/src/mesa/shader/slang/slang_emit.c +++ b/src/mesa/shader/slang/slang_emit.c @@ -1,8 +1,8 @@ /* * Mesa 3-D graphics library - * Version: 6.5.3 * - * Copyright (C) 2005-2007 Brian Paul All Rights Reserved. + * Copyright (C) 2005-2008 Brian Paul All Rights Reserved. + * Copyright (C) 2008 VMware, Inc. All Rights Reserved. * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the "Software"), @@ -36,15 +36,16 @@ ***/ -#include "imports.h" -#include "context.h" -#include "macros.h" -#include "program.h" -#include "prog_instruction.h" -#include "prog_parameter.h" -#include "prog_print.h" +#include "main/imports.h" +#include "main/context.h" +#include "main/macros.h" +#include "shader/program.h" +#include "shader/prog_instruction.h" +#include "shader/prog_parameter.h" +#include "shader/prog_print.h" #include "slang_builtin.h" #include "slang_emit.h" +#include "slang_mem.h" #define PEEPHOLE_OPTIMIZATIONS 1 @@ -59,6 +60,8 @@ typedef struct struct gl_program **Subroutines; GLuint NumSubroutines; + GLuint MaxInstructions; /**< size of prog->Instructions[] buffer */ + /* code-gen options */ GLboolean EmitHighLevelInstructions; GLboolean EmitCondCodes; @@ -78,7 +81,7 @@ new_subroutine(slang_emit_info *emitInfo, GLuint *id) _mesa_realloc(emitInfo->Subroutines, n * sizeof(struct gl_program), (n + 1) * sizeof(struct gl_program)); - emitInfo->Subroutines[n] = _mesa_new_program(ctx, emitInfo->prog->Target, 0); + emitInfo->Subroutines[n] = ctx->Driver.NewProgram(ctx, emitInfo->prog->Target, 0); emitInfo->Subroutines[n]->Parameters = emitInfo->prog->Parameters; emitInfo->NumSubroutines++; *id = n; @@ -107,51 +110,100 @@ writemask_to_swizzle(GLuint writemask) /** - * Swizzle a swizzle. That is, return swz2(swz1) + * Convert a swizzle mask to a writemask. + * Note that the slang_ir_storage->Swizzle field can represent either a + * swizzle mask or a writemask, depending on how it's used. For example, + * when we parse "direction.yz" alone, we don't know whether .yz is a + * writemask or a swizzle. In this case, we encode ".yz" in store->Swizzle + * as a swizzle mask (.yz?? actually). Later, if direction.yz is used as + * an R-value, we use store->Swizzle as-is. Otherwise, if direction.yz is + * used as an L-value, we convert it to a writemask. */ static GLuint -swizzle_swizzle(GLuint swz1, GLuint swz2) +swizzle_to_writemask(GLuint swizzle) +{ + GLuint i, writemask = 0x0; + for (i = 0; i < 4; i++) { + GLuint swz = GET_SWZ(swizzle, i); + if (swz <= SWIZZLE_W) { + writemask |= (1 << swz); + } + } + return writemask; +} + + +/** + * Swizzle a swizzle (function composition). + * That is, return swz2(swz1), or said another way: swz1.szw2 + * Example: swizzle_swizzle(".zwxx", ".xxyw") yields ".zzwx" + */ +GLuint +_slang_swizzle_swizzle(GLuint swz1, GLuint swz2) { GLuint i, swz, s[4]; for (i = 0; i < 4; i++) { GLuint c = GET_SWZ(swz2, i); - s[i] = GET_SWZ(swz1, c); + if (c <= SWIZZLE_W) + s[i] = GET_SWZ(swz1, c); + else + s[i] = c; } swz = MAKE_SWIZZLE4(s[0], s[1], s[2], s[3]); return swz; } -slang_ir_storage * -_slang_new_ir_storage(enum register_file file, GLint index, GLint size) +/** + * Return the default swizzle mask for accessing a variable of the + * given size (in floats). If size = 1, comp is used to identify + * which component [0..3] of the register holds the variable. + */ +GLuint +_slang_var_swizzle(GLint size, GLint comp) { - slang_ir_storage *st; - st = (slang_ir_storage *) _mesa_calloc(sizeof(slang_ir_storage)); - if (st) { - st->File = file; - st->Index = index; - st->Size = size; - st->Swizzle = SWIZZLE_NOOP; - } - return st; + switch (size) { + case 1: + return MAKE_SWIZZLE4(comp, comp, comp, comp); + case 2: + return MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_NIL, SWIZZLE_NIL); + case 3: + return MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_NIL); + default: + return SWIZZLE_XYZW; + } } + /** - * Allocate temporary storage for an intermediate result (such as for - * a multiply or add, etc. + * Allocate storage for the given node (if it hasn't already been allocated). + * + * Typically this is temporary storage for an intermediate result (such as + * for a multiply or add, etc). + * + * If n->Store does not exist it will be created and will be of the size + * specified by defaultSize. */ static GLboolean -alloc_temp_storage(slang_emit_info *emitInfo, slang_ir_node *n, GLint size) +alloc_node_storage(slang_emit_info *emitInfo, slang_ir_node *n, + GLint defaultSize) { assert(!n->Var); - assert(!n->Store); - assert(size > 0); - n->Store = _slang_new_ir_storage(PROGRAM_TEMPORARY, -1, size); - if (!_slang_alloc_temp(emitInfo->vt, n->Store)) { - slang_info_log_error(emitInfo->log, - "Ran out of registers, too many temporaries"); - return GL_FALSE; + if (!n->Store) { + assert(defaultSize > 0); + n->Store = _slang_new_ir_storage(PROGRAM_TEMPORARY, -1, defaultSize); + } + + /* now allocate actual register(s). I.e. set n->Store->Index >= 0 */ + if (n->Store->Index < 0) { + if (!_slang_alloc_temp(emitInfo->vt, n->Store)) { + slang_info_log_error(emitInfo->log, + "Ran out of registers, too many temporaries"); + _slang_free(n->Store); + n->Store = NULL; + return GL_FALSE; + } } return GL_TRUE; } @@ -162,7 +214,7 @@ alloc_temp_storage(slang_emit_info *emitInfo, slang_ir_node *n, GLint size) * Otherwise, no-op. */ static void -free_temp_storage(slang_var_table *vt, slang_ir_node *n) +free_node_storage(slang_var_table *vt, slang_ir_node *n) { if (n->Store->File == PROGRAM_TEMPORARY && n->Store->Index >= 0 && @@ -170,35 +222,115 @@ free_temp_storage(slang_var_table *vt, slang_ir_node *n) if (_slang_is_temp(vt, n->Store)) { _slang_free_temp(vt, n->Store); n->Store->Index = -1; - n->Store->Size = -1; - _mesa_free(n->Store); - n->Store = NULL; + n->Store = NULL; /* XXX this may not be needed */ } } } +/** + * Helper function to allocate a short-term temporary. + * Free it with _slang_free_temp(). + */ +static GLboolean +alloc_local_temp(slang_emit_info *emitInfo, slang_ir_storage *temp, GLint size) +{ + assert(size >= 1); + assert(size <= 4); + _mesa_bzero(temp, sizeof(*temp)); + temp->Size = size; + temp->File = PROGRAM_TEMPORARY; + temp->Index = -1; + return _slang_alloc_temp(emitInfo->vt, temp); +} + + +/** + * Remove any SWIZZLE_NIL terms from given swizzle mask. + * For a swizzle like .z??? generate .zzzz (replicate single component). + * Else, for .wx?? generate .wxzw (insert default component for the position). + */ +static GLuint +fix_swizzle(GLuint swizzle) +{ + GLuint c0 = GET_SWZ(swizzle, 0), + c1 = GET_SWZ(swizzle, 1), + c2 = GET_SWZ(swizzle, 2), + c3 = GET_SWZ(swizzle, 3); + if (c1 == SWIZZLE_NIL && c2 == SWIZZLE_NIL && c3 == SWIZZLE_NIL) { + /* smear first component across all positions */ + c1 = c2 = c3 = c0; + } + else { + /* insert default swizzle components */ + if (c0 == SWIZZLE_NIL) + c0 = SWIZZLE_X; + if (c1 == SWIZZLE_NIL) + c1 = SWIZZLE_Y; + if (c2 == SWIZZLE_NIL) + c2 = SWIZZLE_Z; + if (c3 == SWIZZLE_NIL) + c3 = SWIZZLE_W; + } + return MAKE_SWIZZLE4(c0, c1, c2, c3); +} + + + /** * Convert IR storage to an instruction dst register. */ static void -storage_to_dst_reg(struct prog_dst_register *dst, const slang_ir_storage *st, - GLuint writemask) +storage_to_dst_reg(struct prog_dst_register *dst, const slang_ir_storage *st) { - assert(st->Index >= 0); - dst->File = st->File; - dst->Index = st->Index; + const GLboolean relAddr = st->RelAddr; + const GLint size = st->Size; + GLint index = st->Index; + GLuint swizzle = st->Swizzle; + + assert(index >= 0); + /* if this is storage relative to some parent storage, walk up the tree */ + while (st->Parent) { + st = st->Parent; + assert(st->Index >= 0); + index += st->Index; + swizzle = _slang_swizzle_swizzle(st->Swizzle, swizzle); + } + assert(st->File != PROGRAM_UNDEFINED); - assert(st->Size >= 1); - assert(st->Size <= 4); - if (st->Size == 1) { - GLuint comp = GET_SWZ(st->Swizzle, 0); - assert(comp < 4); - dst->WriteMask = WRITEMASK_X << comp; + dst->File = st->File; + + assert(index >= 0); + dst->Index = index; + + assert(size >= 1); + assert(size <= 4); + + if (swizzle != SWIZZLE_XYZW) { + dst->WriteMask = swizzle_to_writemask(swizzle); } else { + GLuint writemask; + switch (size) { + case 1: + writemask = WRITEMASK_X << GET_SWZ(st->Swizzle, 0); + break; + case 2: + writemask = WRITEMASK_XY; + break; + case 3: + writemask = WRITEMASK_XYZ; + break; + case 4: + writemask = WRITEMASK_XYZW; + break; + default: + ; /* error would have been caught above */ + } dst->WriteMask = writemask; } + + dst->RelAddr = relAddr; } @@ -208,30 +340,69 @@ storage_to_dst_reg(struct prog_dst_register *dst, const slang_ir_storage *st, static void storage_to_src_reg(struct prog_src_register *src, const slang_ir_storage *st) { - static const GLuint defaultSwizzle[4] = { - MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_X, SWIZZLE_X, SWIZZLE_X), - MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_W), - MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_W), - MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Z, SWIZZLE_W) - }; + const GLboolean relAddr = st->RelAddr; + GLint index = st->Index; + GLuint swizzle = st->Swizzle; + + /* if this is storage relative to some parent storage, walk up the tree */ + assert(index >= 0); + while (st->Parent) { + st = st->Parent; + if (st->Index < 0) { + /* an error should have been reported already */ + return; + } + assert(st->Index >= 0); + index += st->Index; + swizzle = _slang_swizzle_swizzle(fix_swizzle(st->Swizzle), swizzle); + } + assert(st->File >= 0); +#if 1 /* XXX temporary */ + if (st->File == PROGRAM_UNDEFINED) { + slang_ir_storage *st0 = (slang_ir_storage *) st; + st0->File = PROGRAM_TEMPORARY; + } +#endif assert(st->File < PROGRAM_UNDEFINED); - assert(st->Size >= 1); - assert(st->Size <= 4); src->File = st->File; - src->Index = st->Index; - if (st->Swizzle != SWIZZLE_NOOP) - src->Swizzle = st->Swizzle; - else - src->Swizzle = defaultSwizzle[st->Size - 1]; /*XXX really need this?*/ - assert(GET_SWZ(src->Swizzle, 0) <= 3); - assert(GET_SWZ(src->Swizzle, 1) <= 3); - assert(GET_SWZ(src->Swizzle, 2) <= 3); - assert(GET_SWZ(src->Swizzle, 3) <= 3); + assert(index >= 0); + src->Index = index; + + swizzle = fix_swizzle(swizzle); + assert(GET_SWZ(swizzle, 0) <= SWIZZLE_W); + assert(GET_SWZ(swizzle, 1) <= SWIZZLE_W); + assert(GET_SWZ(swizzle, 2) <= SWIZZLE_W); + assert(GET_SWZ(swizzle, 3) <= SWIZZLE_W); + src->Swizzle = swizzle; + + src->RelAddr = relAddr; } +/* + * Setup storage pointing to a scalar constant/literal. + */ +static void +constant_to_storage(slang_emit_info *emitInfo, + GLfloat val, + slang_ir_storage *store) +{ + GLuint swizzle; + GLint reg; + GLfloat value[4]; + + value[0] = val; + reg = _mesa_add_unnamed_constant(emitInfo->prog->Parameters, + value, 1, &swizzle); + + memset(store, 0, sizeof(*store)); + store->File = PROGRAM_CONSTANT; + store->Index = reg; + store->Swizzle = swizzle; +} + /** * Add new instruction at end of given program. @@ -251,9 +422,17 @@ new_instruction(slang_emit_info *emitInfo, gl_inst_opcode opcode) _mesa_print_instruction(prog->Instructions + prog->NumInstructions - 1); } #endif - prog->Instructions = _mesa_realloc_instructions(prog->Instructions, - prog->NumInstructions, - prog->NumInstructions + 1); + assert(prog->NumInstructions <= emitInfo->MaxInstructions); + + if (prog->NumInstructions == emitInfo->MaxInstructions) { + /* grow the instruction buffer */ + emitInfo->MaxInstructions += 20; + prog->Instructions = + _mesa_realloc_instructions(prog->Instructions, + prog->NumInstructions, + emitInfo->MaxInstructions); + } + inst = prog->Instructions + prog->NumInstructions; prog->NumInstructions++; _mesa_init_instructions(inst, 1); @@ -267,6 +446,166 @@ new_instruction(slang_emit_info *emitInfo, gl_inst_opcode opcode) } +static struct prog_instruction * +emit_arl_load(slang_emit_info *emitInfo, + enum register_file file, GLint index, GLuint swizzle) +{ + struct prog_instruction *inst = new_instruction(emitInfo, OPCODE_ARL); + inst->SrcReg[0].File = file; + inst->SrcReg[0].Index = index; + inst->SrcReg[0].Swizzle = swizzle; + inst->DstReg.File = PROGRAM_ADDRESS; + inst->DstReg.Index = 0; + inst->DstReg.WriteMask = WRITEMASK_X; + return inst; +} + + +/** + * Emit a new instruction with given opcode, operands. + * At this point the instruction may have multiple indirect register + * loads/stores. We convert those into ARL loads and address-relative + * operands. See comments inside. + * At some point in the future we could directly emit indirectly addressed + * registers in Mesa GPU instructions. + */ +static struct prog_instruction * +emit_instruction(slang_emit_info *emitInfo, + gl_inst_opcode opcode, + const slang_ir_storage *dst, + const slang_ir_storage *src0, + const slang_ir_storage *src1, + const slang_ir_storage *src2) +{ + struct prog_instruction *inst; + GLuint numIndirect = 0; + const slang_ir_storage *src[3]; + slang_ir_storage newSrc[3], newDst; + GLuint i; + GLboolean isTemp[3]; + + isTemp[0] = isTemp[1] = isTemp[2] = GL_FALSE; + + src[0] = src0; + src[1] = src1; + src[2] = src2; + + /* count up how many operands are indirect loads */ + for (i = 0; i < 3; i++) { + if (src[i] && src[i]->IsIndirect) + numIndirect++; + } + if (dst && dst->IsIndirect) + numIndirect++; + + /* Take special steps for indirect register loads. + * If we had multiple address registers this would be simpler. + * For example, this GLSL code: + * x[i] = y[j] + z[k]; + * would translate into something like: + * ARL ADDR.x, i; + * ARL ADDR.y, j; + * ARL ADDR.z, k; + * ADD TEMP[ADDR.x+5], TEMP[ADDR.y+9], TEMP[ADDR.z+4]; + * But since we currently only have one address register we have to do this: + * ARL ADDR.x, i; + * MOV t1, TEMP[ADDR.x+9]; + * ARL ADDR.x, j; + * MOV t2, TEMP[ADDR.x+4]; + * ARL ADDR.x, k; + * ADD TEMP[ADDR.x+5], t1, t2; + * The code here figures this out... + */ + if (numIndirect > 0) { + for (i = 0; i < 3; i++) { + if (src[i] && src[i]->IsIndirect) { + /* load the ARL register with the indirect register */ + emit_arl_load(emitInfo, + src[i]->IndirectFile, + src[i]->IndirectIndex, + src[i]->IndirectSwizzle); + + if (numIndirect > 1) { + /* Need to load src[i] into a temporary register */ + slang_ir_storage srcRelAddr; + alloc_local_temp(emitInfo, &newSrc[i], src[i]->Size); + isTemp[i] = GL_TRUE; + + /* set RelAddr flag on src register */ + srcRelAddr = *src[i]; + srcRelAddr.RelAddr = GL_TRUE; + srcRelAddr.IsIndirect = GL_FALSE; /* not really needed */ + + /* MOV newSrc, srcRelAddr; */ + inst = emit_instruction(emitInfo, + OPCODE_MOV, + &newSrc[i], + &srcRelAddr, + NULL, + NULL); + + src[i] = &newSrc[i]; + } + else { + /* just rewrite the src[i] storage to be ARL-relative */ + newSrc[i] = *src[i]; + newSrc[i].RelAddr = GL_TRUE; + newSrc[i].IsIndirect = GL_FALSE; /* not really needed */ + src[i] = &newSrc[i]; + } + } + } + } + + /* Take special steps for indirect dest register write */ + if (dst && dst->IsIndirect) { + /* load the ARL register with the indirect register */ + emit_arl_load(emitInfo, + dst->IndirectFile, + dst->IndirectIndex, + dst->IndirectSwizzle); + newDst = *dst; + newDst.RelAddr = GL_TRUE; + newDst.IsIndirect = GL_FALSE; + dst = &newDst; + } + + /* OK, emit the instruction and its dst, src regs */ + inst = new_instruction(emitInfo, opcode); + if (!inst) + return NULL; + + if (dst) + storage_to_dst_reg(&inst->DstReg, dst); + + for (i = 0; i < 3; i++) { + if (src[i]) + storage_to_src_reg(&inst->SrcReg[i], src[i]); + } + + /* Free any temp registers that we allocated above */ + for (i = 0; i < 3; i++) { + if (isTemp[i]) + _slang_free_temp(emitInfo->vt, &newSrc[i]); + } + + return inst; +} + + + +/** + * Put a comment on the given instruction. + */ +static void +inst_comment(struct prog_instruction *inst, const char *comment) +{ + if (inst) + inst->Comment = _mesa_strdup(comment); +} + + + /** * Return pointer to last instruction in program. */ @@ -380,6 +719,9 @@ instruction_annotation(gl_inst_opcode opcode, char *dstAnnot, case OPCODE_MUL: operator = "*"; break; + case OPCODE_DP2: + operator = "DP2"; + break; case OPCODE_DP3: operator = "DP3"; break; @@ -420,12 +762,10 @@ instruction_annotation(gl_inst_opcode opcode, char *dstAnnot, * Emit an instruction that's just a comment. */ static struct prog_instruction * -emit_comment(slang_emit_info *emitInfo, const char *s) +emit_comment(slang_emit_info *emitInfo, const char *comment) { struct prog_instruction *inst = new_instruction(emitInfo, OPCODE_NOP); - if (inst) { - inst->Comment = _mesa_strdup(s); - } + inst_comment(inst, comment); return inst; } @@ -437,16 +777,13 @@ emit_comment(slang_emit_info *emitInfo, const char *s) static struct prog_instruction * emit_arith(slang_emit_info *emitInfo, slang_ir_node *n) { - struct prog_instruction *inst; const slang_ir_info *info = _slang_ir_info(n->Opcode); - char *srcAnnot[3], *dstAnnot; + struct prog_instruction *inst; GLuint i; assert(info); assert(info->InstOpcode != OPCODE_NOP); - srcAnnot[0] = srcAnnot[1] = srcAnnot[2] = dstAnnot = NULL; - #if PEEPHOLE_OPTIMIZATIONS /* Look for MAD opportunity */ if (info->NumParams == 2 && @@ -455,69 +792,67 @@ emit_arith(slang_emit_info *emitInfo, slang_ir_node *n) emit(emitInfo, n->Children[0]->Children[0]); /* A */ emit(emitInfo, n->Children[0]->Children[1]); /* B */ emit(emitInfo, n->Children[1]); /* C */ - /* generate MAD instruction */ - inst = new_instruction(emitInfo, OPCODE_MAD); - /* operands: A, B, C: */ - storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Children[0]->Store); - storage_to_src_reg(&inst->SrcReg[1], n->Children[0]->Children[1]->Store); - storage_to_src_reg(&inst->SrcReg[2], n->Children[1]->Store); - free_temp_storage(emitInfo->vt, n->Children[0]->Children[0]); - free_temp_storage(emitInfo->vt, n->Children[0]->Children[1]); - free_temp_storage(emitInfo->vt, n->Children[1]); - } - else if (info->NumParams == 2 && - n->Opcode == IR_ADD && n->Children[1]->Opcode == IR_MUL) { + alloc_node_storage(emitInfo, n, -1); /* dest */ + + inst = emit_instruction(emitInfo, + OPCODE_MAD, + n->Store, + n->Children[0]->Children[0]->Store, + n->Children[0]->Children[1]->Store, + n->Children[1]->Store); + + free_node_storage(emitInfo->vt, n->Children[0]->Children[0]); + free_node_storage(emitInfo->vt, n->Children[0]->Children[1]); + free_node_storage(emitInfo->vt, n->Children[1]); + return inst; + } + + if (info->NumParams == 2 && + n->Opcode == IR_ADD && n->Children[1]->Opcode == IR_MUL) { /* found pattern IR_ADD(A, IR_MUL(B, C)) */ emit(emitInfo, n->Children[0]); /* A */ emit(emitInfo, n->Children[1]->Children[0]); /* B */ emit(emitInfo, n->Children[1]->Children[1]); /* C */ - /* generate MAD instruction */ - inst = new_instruction(emitInfo, OPCODE_MAD); - /* operands: B, C, A */ - storage_to_src_reg(&inst->SrcReg[0], n->Children[1]->Children[0]->Store); - storage_to_src_reg(&inst->SrcReg[1], n->Children[1]->Children[1]->Store); - storage_to_src_reg(&inst->SrcReg[2], n->Children[0]->Store); - free_temp_storage(emitInfo->vt, n->Children[1]->Children[0]); - free_temp_storage(emitInfo->vt, n->Children[1]->Children[1]); - free_temp_storage(emitInfo->vt, n->Children[0]); + alloc_node_storage(emitInfo, n, -1); /* dest */ + + inst = emit_instruction(emitInfo, + OPCODE_MAD, + n->Store, + n->Children[1]->Children[0]->Store, + n->Children[1]->Children[1]->Store, + n->Children[0]->Store); + + free_node_storage(emitInfo->vt, n->Children[1]->Children[0]); + free_node_storage(emitInfo->vt, n->Children[1]->Children[1]); + free_node_storage(emitInfo->vt, n->Children[0]); + return inst; } - else #endif - { - /* normal case */ - - /* gen code for children */ - for (i = 0; i < info->NumParams; i++) - emit(emitInfo, n->Children[i]); - - /* gen this instruction and src registers */ - inst = new_instruction(emitInfo, info->InstOpcode); - for (i = 0; i < info->NumParams; i++) - storage_to_src_reg(&inst->SrcReg[i], n->Children[i]->Store); - /* annotation */ - for (i = 0; i < info->NumParams; i++) - srcAnnot[i] = storage_annotation(n->Children[i], emitInfo->prog); - - /* free temps */ - for (i = 0; i < info->NumParams; i++) - free_temp_storage(emitInfo->vt, n->Children[i]); + /* gen code for children, may involve temp allocation */ + for (i = 0; i < info->NumParams; i++) { + emit(emitInfo, n->Children[i]); + if (!n->Children[i] || !n->Children[i]->Store) { + /* error recovery */ + return NULL; + } } /* result storage */ - if (!n->Store) { - /* XXX this size isn't correct, it depends on the operands */ - if (!alloc_temp_storage(emitInfo, n, info->ResultSize)) - return NULL; - } - storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask); + alloc_node_storage(emitInfo, n, -1); - dstAnnot = storage_annotation(n, emitInfo->prog); + inst = emit_instruction(emitInfo, + info->InstOpcode, + n->Store, /* dest */ + (info->NumParams > 0 ? n->Children[0]->Store : NULL), + (info->NumParams > 1 ? n->Children[1]->Store : NULL), + (info->NumParams > 2 ? n->Children[2]->Store : NULL) + ); - inst->Comment = instruction_annotation(inst->Opcode, dstAnnot, srcAnnot[0], - srcAnnot[1], srcAnnot[2]); + /* free temps */ + for (i = 0; i < info->NumParams; i++) + free_node_storage(emitInfo->vt, n->Children[i]); - /*_mesa_print_instruction(inst);*/ return inst; } @@ -529,7 +864,7 @@ emit_arith(slang_emit_info *emitInfo, slang_ir_node *n) static struct prog_instruction * emit_compare(slang_emit_info *emitInfo, slang_ir_node *n) { - struct prog_instruction *inst; + struct prog_instruction *inst = NULL; GLint size; assert(n->Opcode == IR_EQUAL || n->Opcode == IR_NOTEQUAL); @@ -538,36 +873,39 @@ emit_compare(slang_emit_info *emitInfo, slang_ir_node *n) emit(emitInfo, n->Children[0]); emit(emitInfo, n->Children[1]); - assert(n->Children[0]->Store->Size == n->Children[1]->Store->Size); - size = n->Children[0]->Store->Size; + if (n->Children[0]->Store->Size != n->Children[1]->Store->Size) { + slang_info_log_error(emitInfo->log, "invalid operands to == or !="); + return NULL; + } - if (size == 1) { - gl_inst_opcode opcode; + /* final result is 1 bool */ + if (!alloc_node_storage(emitInfo, n, 1)) + return NULL; - if (!n->Store) { - if (!alloc_temp_storage(emitInfo, n, 1)) /* 1 bool */ - return NULL; - } + size = n->Children[0]->Store->Size; - opcode = n->Opcode == IR_EQUAL ? OPCODE_SEQ : OPCODE_SNE; - inst = new_instruction(emitInfo, opcode); - storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store); - storage_to_src_reg(&inst->SrcReg[1], n->Children[1]->Store); - storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask); + if (size == 1) { + gl_inst_opcode opcode = n->Opcode == IR_EQUAL ? OPCODE_SEQ : OPCODE_SNE; + inst = emit_instruction(emitInfo, + opcode, + n->Store, /* dest */ + n->Children[0]->Store, + n->Children[1]->Store, + NULL); } else if (size <= 4) { - static const GLfloat zero[4] = { 0, 0, 0, 0 }; - GLuint zeroSwizzle, swizzle; - GLint zeroReg = _mesa_add_unnamed_constant(emitInfo->prog->Parameters, - zero, 4, &zeroSwizzle); + /* compare two vectors. + * Unfortunately, there's no instruction to compare vectors and + * return a scalar result. Do it with some compare and dot product + * instructions... + */ + GLuint swizzle; gl_inst_opcode dotOp; - - assert(zeroReg >= 0); + slang_ir_storage tempStore; - assert(!n->Store); - if (!n->Store) { - if (!alloc_temp_storage(emitInfo, n, size)) /* 'size' bools */ - return NULL; + if (!alloc_local_temp(emitInfo, &tempStore, 4)) { + return NULL; + /* out of temps */ } if (size == 4) { @@ -580,68 +918,115 @@ emit_compare(slang_emit_info *emitInfo, slang_ir_node *n) } else { assert(size == 2); - dotOp = OPCODE_DP3; + dotOp = OPCODE_DP3; /* XXX use OPCODE_DP2 eventually */ swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, SWIZZLE_Y, SWIZZLE_Y); } - /* Compute equality, inequality (tmp1 = (A ?= B)) */ - inst = new_instruction(emitInfo, OPCODE_SNE); - storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store); - storage_to_src_reg(&inst->SrcReg[1], n->Children[1]->Store); - storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask); - inst->Comment = _mesa_strdup("Compare values"); - - /* Compute tmp2 = DOT(tmp1, tmp1) (reduction) */ - inst = new_instruction(emitInfo, dotOp); - storage_to_src_reg(&inst->SrcReg[0], n->Store); - storage_to_src_reg(&inst->SrcReg[1], n->Store); + /* Compute inequality (temp = (A != B)) */ + inst = emit_instruction(emitInfo, + OPCODE_SNE, + &tempStore, + n->Children[0]->Store, + n->Children[1]->Store, + NULL); + inst_comment(inst, "Compare values"); + + /* Compute val = DOT(temp, temp) (reduction) */ + inst = emit_instruction(emitInfo, + dotOp, + n->Store, + &tempStore, + &tempStore, + NULL); inst->SrcReg[0].Swizzle = inst->SrcReg[1].Swizzle = swizzle; /*override*/ - free_temp_storage(emitInfo->vt, n); /* free tmp1 */ - if (!alloc_temp_storage(emitInfo, n, 1)) /* alloc tmp2 */ - return NULL; - storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask); - inst->Comment = _mesa_strdup("Reduce vec to bool"); + inst_comment(inst, "Reduce vec to bool"); + + _slang_free_temp(emitInfo->vt, &tempStore); /* free temp */ if (n->Opcode == IR_EQUAL) { - /* compute tmp2.x = !tmp2.x via tmp2.x = (tmp2.x == 0) */ - inst = new_instruction(emitInfo, OPCODE_SEQ); - storage_to_src_reg(&inst->SrcReg[0], n->Store); - inst->SrcReg[1].File = PROGRAM_CONSTANT; - inst->SrcReg[1].Index = zeroReg; - inst->SrcReg[1].Swizzle = zeroSwizzle; - storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask); - inst->Comment = _mesa_strdup("Invert true/false"); + /* compute val = !val.x with SEQ val, val, 0; */ + slang_ir_storage zero; + constant_to_storage(emitInfo, 0.0, &zero); + inst = emit_instruction(emitInfo, + OPCODE_SEQ, + n->Store, /* dest */ + n->Store, + &zero, + NULL); + inst_comment(inst, "Invert true/false"); } } else { - /* size > 4, struct compare */ -#if 0 + /* size > 4, struct or array compare. + * XXX this won't work reliably for structs with padding!! + */ GLint i, num = (n->Children[0]->Store->Size + 3) / 4; - /*printf("BEGIN COMPARE size %d\n", num);*/ + slang_ir_storage accTemp, sneTemp; + + if (!alloc_local_temp(emitInfo, &accTemp, 4)) + return NULL; + + if (!alloc_local_temp(emitInfo, &sneTemp, 4)) + return NULL; + for (i = 0; i < num; i++) { - inst = new_instruction(emitInfo, opcode); - inst->SrcReg[0].File = n->Children[0]->Store->File; - inst->SrcReg[0].Index = n->Children[0]->Store->Index + i; - inst->SrcReg[1].File = n->Children[1]->Store->File; - inst->SrcReg[1].Index = n->Children[1]->Store->Index + i; - inst->DstReg.File = n->Store->File; - inst->DstReg.Index = n->Store->Index; - - inst->CondUpdate = 1; /* update cond code */ - if (i > 0) { - inst->DstReg.CondMask = COND_NE; /* update if !=0 */ + slang_ir_storage srcStore0 = *n->Children[0]->Store; + slang_ir_storage srcStore1 = *n->Children[1]->Store; + srcStore0.Index += i; + srcStore1.Index += i; + + if (i == 0) { + /* SNE accTemp, left[i], right[i] */ + inst = emit_instruction(emitInfo, OPCODE_SNE, + &accTemp, /* dest */ + &srcStore0, + &srcStore1, + NULL); + inst_comment(inst, "Begin struct/array comparison"); + } + else { + /* SNE sneTemp, left[i], right[i] */ + inst = emit_instruction(emitInfo, OPCODE_SNE, + &sneTemp, /* dest */ + &srcStore0, + &srcStore1, + NULL); + /* ADD accTemp, accTemp, sneTemp; # like logical-OR */ + inst = emit_instruction(emitInfo, OPCODE_ADD, + &accTemp, /* dest */ + &accTemp, + &sneTemp, + NULL); } - /*_mesa_print_instruction(inst);*/ } - storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask); -#endif - _mesa_problem(NULL, "struct comparison not implemented yet"); - inst = NULL; + + /* compute accTemp.x || accTemp.y || accTemp.z || accTemp.w with DOT4 */ + inst = emit_instruction(emitInfo, OPCODE_DP4, + n->Store, + &accTemp, + &accTemp, + NULL); + inst_comment(inst, "End struct/array comparison"); + + if (n->Opcode == IR_EQUAL) { + /* compute tmp.x = !tmp.x via tmp.x = (tmp.x == 0) */ + slang_ir_storage zero; + constant_to_storage(emitInfo, 0.0, &zero); + inst = emit_instruction(emitInfo, OPCODE_SEQ, + n->Store, /* dest */ + n->Store, + &zero, + NULL); + inst_comment(inst, "Invert true/false"); + } + + _slang_free_temp(emitInfo->vt, &accTemp); + _slang_free_temp(emitInfo->vt, &sneTemp); } /* free temps */ - free_temp_storage(emitInfo->vt, n->Children[0]); - free_temp_storage(emitInfo->vt, n->Children[1]); + free_node_storage(emitInfo->vt, n->Children[0]); + free_node_storage(emitInfo->vt, n->Children[1]); return inst; } @@ -655,6 +1040,7 @@ static struct prog_instruction * emit_clamp(slang_emit_info *emitInfo, slang_ir_node *n) { struct prog_instruction *inst; + slang_ir_node tmpNode; assert(n->Opcode == IR_CLAMP); /* ch[0] = value @@ -693,24 +1079,34 @@ emit_clamp(slang_emit_info *emitInfo, slang_ir_node *n) } #endif - if (!n->Store) - if (!alloc_temp_storage(emitInfo, n, n->Children[0]->Store->Size)) - return NULL; + if (!alloc_node_storage(emitInfo, n, n->Children[0]->Store->Size)) + return NULL; emit(emitInfo, n->Children[1]); emit(emitInfo, n->Children[2]); + /* Some GPUs don't allow reading from output registers. So if the + * dest for this clamp() is an output reg, we can't use that reg for + * the intermediate result. Use a temp register instead. + */ + _mesa_bzero(&tmpNode, sizeof(tmpNode)); + alloc_node_storage(emitInfo, &tmpNode, n->Store->Size); + /* tmp = max(ch[0], ch[1]) */ - inst = new_instruction(emitInfo, OPCODE_MAX); - storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask); - storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store); - storage_to_src_reg(&inst->SrcReg[1], n->Children[1]->Store); + inst = emit_instruction(emitInfo, OPCODE_MAX, + tmpNode.Store, /* dest */ + n->Children[0]->Store, + n->Children[1]->Store, + NULL); + + /* n->dest = min(tmp, ch[2]) */ + inst = emit_instruction(emitInfo, OPCODE_MIN, + n->Store, /* dest */ + tmpNode.Store, + n->Children[2]->Store, + NULL); - /* tmp = min(tmp, ch[2]) */ - inst = new_instruction(emitInfo, OPCODE_MIN); - storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask); - storage_to_src_reg(&inst->SrcReg[0], n->Store); - storage_to_src_reg(&inst->SrcReg[1], n->Children[2]->Store); + free_node_storage(emitInfo->vt, &tmpNode); return inst; } @@ -727,13 +1123,15 @@ emit_negation(slang_emit_info *emitInfo, slang_ir_node *n) emit(emitInfo, n->Children[0]); - if (!n->Store) - if (!alloc_temp_storage(emitInfo, n, n->Children[0]->Store->Size)) - return NULL; + if (!alloc_node_storage(emitInfo, n, n->Children[0]->Store->Size)) + return NULL; - inst = new_instruction(emitInfo, OPCODE_MOV); - storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask); - storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store); + inst = emit_instruction(emitInfo, + OPCODE_MOV, + n->Store, /* dest */ + n->Children[0]->Store, + NULL, + NULL); inst->SrcReg[0].NegateBase = NEGATE_XYZW; return inst; } @@ -758,21 +1156,27 @@ emit_label(slang_emit_info *emitInfo, const slang_ir_node *n) /** - * Emit code for an inlined function call (subroutine). + * Emit code for a function call. + * Note that for each time a function is called, we emit the function's + * body code again because the set of available registers may be different. */ static struct prog_instruction * -emit_func(slang_emit_info *emitInfo, slang_ir_node *n) +emit_fcall(slang_emit_info *emitInfo, slang_ir_node *n) { struct gl_program *progSave; struct prog_instruction *inst; GLuint subroutineId; + GLuint maxInstSave; - assert(n->Opcode == IR_FUNC); + assert(n->Opcode == IR_CALL); assert(n->Label); /* save/push cur program */ + maxInstSave = emitInfo->MaxInstructions; progSave = emitInfo->prog; + emitInfo->prog = new_subroutine(emitInfo, &subroutineId); + emitInfo->MaxInstructions = emitInfo->prog->NumInstructions; _slang_label_set_location(n->Label, emitInfo->prog->NumInstructions, emitInfo->prog); @@ -784,7 +1188,7 @@ emit_func(slang_emit_info *emitInfo, slang_ir_node *n) * really just a NOP to attach the label to. */ inst = new_instruction(emitInfo, OPCODE_BGNSUB); - inst->Comment = _mesa_strdup(n->Label->Name); + inst_comment(inst, n->Label->Name); } /* body of function: */ @@ -799,17 +1203,18 @@ emit_func(slang_emit_info *emitInfo, slang_ir_node *n) if (emitInfo->EmitBeginEndSub) { inst = new_instruction(emitInfo, OPCODE_ENDSUB); - inst->Comment = _mesa_strdup(n->Label->Name); + inst_comment(inst, n->Label->Name); } /* pop/restore cur program */ emitInfo->prog = progSave; + emitInfo->MaxInstructions = maxInstSave; /* emit the function call */ inst = new_instruction(emitInfo, OPCODE_CAL); /* The branch target is just the subroutine number (changed later) */ inst->BranchTarget = subroutineId; - inst->Comment = _mesa_strdup(n->Label->Name); + inst_comment(inst, n->Label->Name); assert(inst->BranchTarget >= 0); return inst; @@ -835,12 +1240,18 @@ emit_return(slang_emit_info *emitInfo, slang_ir_node *n) static struct prog_instruction * emit_kill(slang_emit_info *emitInfo) { + struct gl_fragment_program *fp; struct prog_instruction *inst; /* NV-KILL - discard fragment depending on condition code. * Note that ARB-KILL depends on sign of vector operand. */ inst = new_instruction(emitInfo, OPCODE_KIL_NV); - inst->DstReg.CondMask = COND_TR; /* always branch */ + inst->DstReg.CondMask = COND_TR; /* always kill */ + + assert(emitInfo->prog->Target == GL_FRAGMENT_PROGRAM_ARB); + fp = (struct gl_fragment_program *) emitInfo->prog; + fp->UsesKill = GL_TRUE; + return inst; } @@ -849,81 +1260,117 @@ static struct prog_instruction * emit_tex(slang_emit_info *emitInfo, slang_ir_node *n) { struct prog_instruction *inst; - - (void) emit(emitInfo, n->Children[1]); + gl_inst_opcode opcode; if (n->Opcode == IR_TEX) { - inst = new_instruction(emitInfo, OPCODE_TEX); + opcode = OPCODE_TEX; } else if (n->Opcode == IR_TEXB) { - inst = new_instruction(emitInfo, OPCODE_TXB); + opcode = OPCODE_TXB; } else { assert(n->Opcode == IR_TEXP); - inst = new_instruction(emitInfo, OPCODE_TXP); + opcode = OPCODE_TXP; } - if (!n->Store) - if (!alloc_temp_storage(emitInfo, n, 4)) - return NULL; + /* emit code for the texcoord operand */ + (void) emit(emitInfo, n->Children[1]); - storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask); + /* alloc storage for result of texture fetch */ + if (!alloc_node_storage(emitInfo, n, 4)) + return NULL; - /* Child[1] is the coord */ - assert(n->Children[1]->Store->File != PROGRAM_UNDEFINED); - assert(n->Children[1]->Store->Index >= 0); - storage_to_src_reg(&inst->SrcReg[0], n->Children[1]->Store); + /* emit TEX instruction; Child[1] is the texcoord */ + inst = emit_instruction(emitInfo, + opcode, + n->Store, + n->Children[1]->Store, + NULL, + NULL); /* Child[0] is the sampler (a uniform which'll indicate the texture unit) */ assert(n->Children[0]->Store); + /* Store->Index is the sampler index */ + assert(n->Children[0]->Store->Index >= 0); + /* Store->Size is the texture target */ assert(n->Children[0]->Store->Size >= TEXTURE_1D_INDEX); + assert(n->Children[0]->Store->Size <= TEXTURE_RECT_INDEX); - inst->Sampler = n->Children[0]->Store->Index; /* i.e. uniform's index */ inst->TexSrcTarget = n->Children[0]->Store->Size; - inst->TexSrcUnit = 27; /* Dummy value; the TexSrcUnit will be computed at - * link time, using the sampler uniform's value. - */ + inst->TexSrcUnit = n->Children[0]->Store->Index; /* i.e. uniform's index */ + return inst; } +/** + * Assignment/copy + */ static struct prog_instruction * -emit_move(slang_emit_info *emitInfo, slang_ir_node *n) +emit_copy(slang_emit_info *emitInfo, slang_ir_node *n) { struct prog_instruction *inst; + assert(n->Opcode == IR_COPY); + /* lhs */ emit(emitInfo, n->Children[0]); + if (!n->Children[0]->Store || n->Children[0]->Store->Index < 0) { + /* an error should have been already recorded */ + return NULL; + } /* rhs */ assert(n->Children[1]); inst = emit(emitInfo, n->Children[1]); - if (!n->Children[1]->Store) { - slang_info_log_error(emitInfo->log, "invalid assignment"); + if (!n->Children[1]->Store || n->Children[1]->Store->Index < 0) { + if (!emitInfo->log->text) { + slang_info_log_error(emitInfo->log, "invalid assignment"); + } return NULL; } + assert(n->Children[1]->Store->Index >= 0); + /*assert(n->Children[0]->Store->Size == n->Children[1]->Store->Size);*/ + n->Store = n->Children[0]->Store; + if (n->Store->File == PROGRAM_SAMPLER) { + /* no code generated for sampler assignments, + * just copy the sampler index at compile time. + */ + n->Store->Index = n->Children[1]->Store->Index; + return NULL; + } + #if PEEPHOLE_OPTIMIZATIONS if (inst && _slang_is_temp(emitInfo->vt, n->Children[1]->Store) && (inst->DstReg.File == n->Children[1]->Store->File) && - (inst->DstReg.Index == n->Children[1]->Store->Index)) { + (inst->DstReg.Index == n->Children[1]->Store->Index) && + !n->Children[0]->Store->IsIndirect && + n->Children[0]->Store->Size <= 4) { /* Peephole optimization: * The Right-Hand-Side has its results in a temporary place. * Modify the RHS (and the prev instruction) to store its results * in the destination specified by n->Children[0]. * Then, this MOVE is a no-op. + * Ex: + * MUL tmp, x, y; + * MOV a, tmp; + * becomes: + * MUL a, x, y; */ if (n->Children[1]->Opcode != IR_SWIZZLE) _slang_free_temp(emitInfo->vt, n->Children[1]->Store); *n->Children[1]->Store = *n->Children[0]->Store; + /* fixup the previous instruction (which stored the RHS result) */ assert(n->Children[0]->Store->Index >= 0); - storage_to_dst_reg(&inst->DstReg, n->Children[0]->Store, n->Writemask); + + storage_to_dst_reg(&inst->DstReg, n->Children[0]->Store); return inst; } else @@ -934,15 +1381,16 @@ emit_move(slang_emit_info *emitInfo, slang_ir_node *n) slang_ir_storage dstStore = *n->Children[0]->Store; slang_ir_storage srcStore = *n->Children[1]->Store; GLint size = srcStore.Size; - ASSERT(n->Children[0]->Writemask == WRITEMASK_XYZW); ASSERT(n->Children[1]->Store->Swizzle == SWIZZLE_NOOP); dstStore.Size = 4; srcStore.Size = 4; while (size >= 4) { - inst = new_instruction(emitInfo, OPCODE_MOV); - inst->Comment = _mesa_strdup("IR_MOVE block"); - storage_to_dst_reg(&inst->DstReg, &dstStore, n->Writemask); - storage_to_src_reg(&inst->SrcReg[0], &srcStore); + inst = emit_instruction(emitInfo, OPCODE_MOV, + &dstStore, + &srcStore, + NULL, + NULL); + inst_comment(inst, "IR_COPY block"); srcStore.Index++; dstStore.Index++; size -= 4; @@ -951,16 +1399,18 @@ emit_move(slang_emit_info *emitInfo, slang_ir_node *n) else { /* single register move */ char *srcAnnot, *dstAnnot; - inst = new_instruction(emitInfo, OPCODE_MOV); assert(n->Children[0]->Store->Index >= 0); - storage_to_dst_reg(&inst->DstReg, n->Children[0]->Store, n->Writemask); - storage_to_src_reg(&inst->SrcReg[0], n->Children[1]->Store); + inst = emit_instruction(emitInfo, OPCODE_MOV, + n->Children[0]->Store, /* dest */ + n->Children[1]->Store, + NULL, + NULL); dstAnnot = storage_annotation(n->Children[0], emitInfo->prog); srcAnnot = storage_annotation(n->Children[1], emitInfo->prog); inst->Comment = instruction_annotation(inst->Opcode, dstAnnot, srcAnnot, NULL, NULL); } - free_temp_storage(emitInfo->vt, n->Children[1]); + free_node_storage(emitInfo->vt, n->Children[1]); return inst; } } @@ -983,6 +1433,11 @@ emit_cond(slang_emit_info *emitInfo, slang_ir_node *n) /* emit code for the expression */ inst = emit(emitInfo, n->Children[0]); + if (!n->Children[0]->Store) { + /* error recovery */ + return NULL; + } + assert(n->Children[0]->Store); /*assert(n->Children[0]->Store->Size == 1);*/ @@ -1004,14 +1459,16 @@ emit_cond(slang_emit_info *emitInfo, slang_ir_node *n) * is normally generated for the expression "i". * Generate a move instruction just to set condition codes. */ - if (!alloc_temp_storage(emitInfo, n, 1)) + if (!alloc_node_storage(emitInfo, n, 1)) return NULL; - inst = new_instruction(emitInfo, OPCODE_MOV); + inst = emit_instruction(emitInfo, OPCODE_MOV, + n->Store, /* dest */ + n->Children[0]->Store, + NULL, + NULL); inst->CondUpdate = GL_TRUE; - storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask); - storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store); + inst_comment(inst, "COND expr"); _slang_free_temp(emitInfo->vt, n->Store); - inst->Comment = _mesa_strdup("COND expr"); return inst; } } @@ -1029,32 +1486,52 @@ emit_cond(slang_emit_info *emitInfo, slang_ir_node *n) static struct prog_instruction * emit_not(slang_emit_info *emitInfo, slang_ir_node *n) { - GLfloat zero = 0.0; - slang_ir_storage st; + static const struct { + gl_inst_opcode op, opNot; + } operators[] = { + { OPCODE_SLT, OPCODE_SGE }, + { OPCODE_SLE, OPCODE_SGT }, + { OPCODE_SGT, OPCODE_SLE }, + { OPCODE_SGE, OPCODE_SLT }, + { OPCODE_SEQ, OPCODE_SNE }, + { OPCODE_SNE, OPCODE_SEQ }, + { 0, 0 } + }; struct prog_instruction *inst; - - /* need zero constant */ - st.File = PROGRAM_CONSTANT; - st.Size = 1; - st.Index = _mesa_add_unnamed_constant(emitInfo->prog->Parameters, &zero, - 1, &st.Swizzle); + slang_ir_storage zero; + GLuint i; /* child expr */ - (void) emit(emitInfo, n->Children[0]); - /* XXXX if child instr is SGT convert to SLE, if SEQ, SNE, etc */ + inst = emit(emitInfo, n->Children[0]); - if (!n->Store) - if (!alloc_temp_storage(emitInfo, n, n->Children[0]->Store->Size)) - return NULL; +#if PEEPHOLE_OPTIMIZATIONS + if (inst) { + /* if the prev instruction was a comparison instruction, invert it */ + for (i = 0; operators[i].op; i++) { + if (inst->Opcode == operators[i].op) { + inst->Opcode = operators[i].opNot; + n->Store = n->Children[0]->Store; + return inst; + } + } + } +#endif + + /* else, invert using SEQ (v = v == 0) */ + if (!alloc_node_storage(emitInfo, n, n->Children[0]->Store->Size)) + return NULL; - inst = new_instruction(emitInfo, OPCODE_SEQ); - storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask); - storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store); - storage_to_src_reg(&inst->SrcReg[1], &st); + constant_to_storage(emitInfo, 0.0, &zero); + inst = emit_instruction(emitInfo, + OPCODE_SEQ, + n->Store, + n->Children[0]->Store, + &zero, + NULL); + inst_comment(inst, "NOT"); - free_temp_storage(emitInfo->vt, n->Children[0]); + free_node_storage(emitInfo->vt, n->Children[0]); - inst->Comment = _mesa_strdup("NOT"); return inst; } @@ -1063,24 +1540,35 @@ static struct prog_instruction * emit_if(slang_emit_info *emitInfo, slang_ir_node *n) { struct gl_program *prog = emitInfo->prog; - struct prog_instruction *ifInst, *inst; GLuint ifInstLoc, elseInstLoc = 0; GLuint condWritemask = 0; - inst = emit(emitInfo, n->Children[0]); /* the condition */ - if (emitInfo->EmitCondCodes) { - assert(inst); - condWritemask = inst->DstReg.WriteMask; + /* emit condition expression code */ + { + struct prog_instruction *inst; + inst = emit(emitInfo, n->Children[0]); + if (emitInfo->EmitCondCodes) { + if (!inst) { + /* error recovery */ + return NULL; + } + condWritemask = inst->DstReg.WriteMask; + } } + if (!n->Children[0]->Store) + return NULL; + #if 0 assert(n->Children[0]->Store->Size == 1); /* a bool! */ #endif ifInstLoc = prog->NumInstructions; if (emitInfo->EmitHighLevelInstructions) { - ifInst = new_instruction(emitInfo, OPCODE_IF); if (emitInfo->EmitCondCodes) { + /* IF condcode THEN ... */ + struct prog_instruction *ifInst; + ifInst = new_instruction(emitInfo, OPCODE_IF); ifInst->DstReg.CondMask = COND_NE; /* if cond is non-zero */ /* only test the cond code (1 of 4) that was updated by the * previous instruction. @@ -1088,15 +1576,19 @@ emit_if(slang_emit_info *emitInfo, slang_ir_node *n) ifInst->DstReg.CondSwizzle = writemask_to_swizzle(condWritemask); } else { - /* test reg.x */ - storage_to_src_reg(&ifInst->SrcReg[0], n->Children[0]->Store); + /* IF src[0] THEN ... */ + emit_instruction(emitInfo, OPCODE_IF, + NULL, /* dst */ + n->Children[0]->Store, /* op0 */ + NULL, + NULL); } } else { /* conditional jump to else, or endif */ - ifInst = new_instruction(emitInfo, OPCODE_BRA); + struct prog_instruction *ifInst = new_instruction(emitInfo, OPCODE_BRA); ifInst->DstReg.CondMask = COND_EQ; /* BRA if cond is zero */ - ifInst->Comment = _mesa_strdup("if zero"); + inst_comment(ifInst, "if zero"); ifInst->DstReg.CondSwizzle = writemask_to_swizzle(condWritemask); } @@ -1113,18 +1605,15 @@ emit_if(slang_emit_info *emitInfo, slang_ir_node *n) /* jump to endif instruction */ struct prog_instruction *inst; inst = new_instruction(emitInfo, OPCODE_BRA); - inst->Comment = _mesa_strdup("else"); + inst_comment(inst, "else"); inst->DstReg.CondMask = COND_TR; /* always branch */ } - ifInst = prog->Instructions + ifInstLoc; - ifInst->BranchTarget = prog->NumInstructions; - + prog->Instructions[ifInstLoc].BranchTarget = prog->NumInstructions; emit(emitInfo, n->Children[2]); } else { /* no else body */ - ifInst = prog->Instructions + ifInstLoc; - ifInst->BranchTarget = prog->NumInstructions /*+ 1*/; + prog->Instructions[ifInstLoc].BranchTarget = prog->NumInstructions; } if (emitInfo->EmitHighLevelInstructions) { @@ -1132,9 +1621,7 @@ emit_if(slang_emit_info *emitInfo, slang_ir_node *n) } if (n->Children[2]) { - struct prog_instruction *elseInst; - elseInst = prog->Instructions + elseInstLoc; - elseInst->BranchTarget = prog->NumInstructions; + prog->Instructions[elseInstLoc].BranchTarget = prog->NumInstructions; } return NULL; } @@ -1144,7 +1631,7 @@ static struct prog_instruction * emit_loop(slang_emit_info *emitInfo, slang_ir_node *n) { struct gl_program *prog = emitInfo->prog; - struct prog_instruction *beginInst, *endInst; + struct prog_instruction *endInst; GLuint beginInstLoc, tailInstLoc, endInstLoc; slang_ir_node *ir; @@ -1180,8 +1667,7 @@ emit_loop(slang_emit_info *emitInfo, slang_ir_node *n) if (emitInfo->EmitHighLevelInstructions) { /* BGNLOOP's BranchTarget points to the ENDLOOP inst */ - beginInst = prog->Instructions + beginInstLoc; - beginInst->BranchTarget = prog->NumInstructions - 1; + prog->Instructions[beginInstLoc].BranchTarget = prog->NumInstructions -1; } /* Done emitting loop code. Now walk over the loop's linked list of @@ -1192,22 +1678,16 @@ emit_loop(slang_emit_info *emitInfo, slang_ir_node *n) struct prog_instruction *inst = prog->Instructions + ir->InstLocation; assert(inst->BranchTarget < 0); if (ir->Opcode == IR_BREAK || - ir->Opcode == IR_BREAK_IF_FALSE || ir->Opcode == IR_BREAK_IF_TRUE) { assert(inst->Opcode == OPCODE_BRK || - inst->Opcode == OPCODE_BRK0 || - inst->Opcode == OPCODE_BRK1 || inst->Opcode == OPCODE_BRA); /* go to instruction after end of loop */ inst->BranchTarget = endInstLoc + 1; } else { assert(ir->Opcode == IR_CONT || - ir->Opcode == IR_CONT_IF_FALSE || ir->Opcode == IR_CONT_IF_TRUE); assert(inst->Opcode == OPCODE_CONT || - inst->Opcode == OPCODE_CONT0 || - inst->Opcode == OPCODE_CONT1 || inst->Opcode == OPCODE_BRA); /* go to instruction at tail of loop */ inst->BranchTarget = endInstLoc; @@ -1259,16 +1739,12 @@ emit_cont_break(slang_emit_info *emitInfo, slang_ir_node *n) * Either OPCODE_CONT, OPCODE_BRK or OPCODE_BRA will be emitted. */ static struct prog_instruction * -emit_cont_break_if(slang_emit_info *emitInfo, slang_ir_node *n, - GLboolean breakTrue) +emit_cont_break_if_true(slang_emit_info *emitInfo, slang_ir_node *n) { - gl_inst_opcode opcode; struct prog_instruction *inst; assert(n->Opcode == IR_CONT_IF_TRUE || - n->Opcode == IR_CONT_IF_FALSE || - n->Opcode == IR_BREAK_IF_TRUE || - n->Opcode == IR_BREAK_IF_FALSE); + n->Opcode == IR_BREAK_IF_TRUE); /* evaluate condition expr, setting cond codes */ inst = emit(emitInfo, n->Children[0]); @@ -1281,169 +1757,365 @@ emit_cont_break_if(slang_emit_info *emitInfo, slang_ir_node *n, /* opcode selection */ if (emitInfo->EmitHighLevelInstructions) { + const gl_inst_opcode opcode + = (n->Opcode == IR_CONT_IF_TRUE) ? OPCODE_CONT : OPCODE_BRK; if (emitInfo->EmitCondCodes) { - if (n->Opcode == IR_CONT_IF_TRUE || - n->Opcode == IR_CONT_IF_FALSE) - opcode = OPCODE_CONT; - else - opcode = OPCODE_BRK; + /* Get the writemask from the previous instruction which set + * the condcodes. Use that writemask as the CondSwizzle. + */ + const GLuint condWritemask = inst->DstReg.WriteMask; + inst = new_instruction(emitInfo, opcode); + inst->DstReg.CondMask = COND_NE; + inst->DstReg.CondSwizzle = writemask_to_swizzle(condWritemask); + return inst; } else { - if (n->Opcode == IR_CONT_IF_TRUE) - opcode = OPCODE_CONT1; - else if (n->Opcode == IR_CONT_IF_FALSE) - opcode = OPCODE_CONT0; - else if (n->Opcode == IR_BREAK_IF_TRUE) - opcode = OPCODE_BRK1; - else if (n->Opcode == IR_BREAK_IF_FALSE) - opcode = OPCODE_BRK0; + /* IF reg + * BRK/CONT; + * ENDIF + */ + GLint ifInstLoc; + ifInstLoc = emitInfo->prog->NumInstructions; + inst = emit_instruction(emitInfo, OPCODE_IF, + NULL, /* dest */ + n->Children[0]->Store, + NULL, + NULL); + n->InstLocation = emitInfo->prog->NumInstructions; + + inst = new_instruction(emitInfo, opcode); + inst = new_instruction(emitInfo, OPCODE_ENDIF); + + emitInfo->prog->Instructions[ifInstLoc].BranchTarget + = emitInfo->prog->NumInstructions; + return inst; } } else { - opcode = OPCODE_BRA; + const GLuint condWritemask = inst->DstReg.WriteMask; + assert(emitInfo->EmitCondCodes); + inst = new_instruction(emitInfo, OPCODE_BRA); + inst->DstReg.CondMask = COND_NE; + inst->DstReg.CondSwizzle = writemask_to_swizzle(condWritemask); + return inst; } +} - inst = new_instruction(emitInfo, opcode); - if (emitInfo->EmitCondCodes) { - inst->DstReg.CondMask = breakTrue ? COND_NE : COND_EQ; - } - else { - /* BRK0, BRK1, CONT0, CONT1 uses SrcReg[0] as the condition */ - storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store); - } + +static struct prog_instruction * +emit_swizzle(slang_emit_info *emitInfo, slang_ir_node *n) +{ + struct prog_instruction *inst; + + inst = emit(emitInfo, n->Children[0]); + +#if 0 + assert(n->Store->Parent); + /* Apply this node's swizzle to parent's storage */ + GLuint swizzle = n->Store->Swizzle; + _slang_copy_ir_storage(n->Store, n->Store->Parent); + n->Store->Swizzle = _slang_swizzle_swizzle(n->Store->Swizzle, swizzle); + assert(!n->Store->Parent); +#endif return inst; } - /** - * Remove any SWIZZLE_NIL terms from given swizzle mask (smear prev term). - * Ex: fix_swizzle("zyNN") -> "zyyy" + * Dereference array element: element == array[index] + * This basically involves emitting code for computing the array index + * and updating the node/element's storage info. */ -static GLuint -fix_swizzle(GLuint swizzle) +static struct prog_instruction * +emit_array_element(slang_emit_info *emitInfo, slang_ir_node *n) { - GLuint swz[4], i; - for (i = 0; i < 4; i++) { - swz[i] = GET_SWZ(swizzle, i); - if (swz[i] == SWIZZLE_NIL) { - swz[i] = swz[i - 1]; + slang_ir_storage *arrayStore, *indexStore; + const int elemSize = n->Store->Size; /* number of floats */ + const GLint elemSizeVec = (elemSize + 3) / 4; /* number of vec4 */ + struct prog_instruction *inst; + + assert(n->Opcode == IR_ELEMENT); + assert(elemSize > 0); + + /* special case for built-in state variables, like light state */ + { + slang_ir_storage *root = n->Store; + assert(!root->Parent); + while (root->Parent) + root = root->Parent; + + if (root->File == PROGRAM_STATE_VAR) { + GLboolean direct; + GLint index = + _slang_alloc_statevar(n, emitInfo->prog->Parameters, &direct); + if (index < 0) { + /* error */ + return NULL; + } + if (direct) { + n->Store->Index = index; + return NULL; /* all done */ + } } } - return MAKE_SWIZZLE4(swz[0], swz[1], swz[2], swz[3]); -} + /* do codegen for array itself */ + emit(emitInfo, n->Children[0]); + arrayStore = n->Children[0]->Store; -/** - * Return the number of components actually named by the swizzle. - * Recall that swizzles may have undefined/don't-care values. - */ -static GLuint -swizzle_size(GLuint swizzle) -{ - GLuint size = 0, i; - for (i = 0; i < 4; i++) { - GLuint swz = GET_SWZ(swizzle, i); - size += (swz >= 0 && swz <= 3); + /* The initial array element storage is the array's storage, + * then modified below. + */ + _slang_copy_ir_storage(n->Store, arrayStore); + + + if (n->Children[1]->Opcode == IR_FLOAT) { + /* Constant array index */ + const GLint element = (GLint) n->Children[1]->Value[0]; + + /* this element's storage is the array's storage, plus constant offset */ + n->Store->Index += elemSizeVec * element; + } + else { + /* Variable array index */ + + /* do codegen for array index expression */ + emit(emitInfo, n->Children[1]); + indexStore = n->Children[1]->Store; + + if (indexStore->IsIndirect) { + /* need to put the array index into a temporary since we can't + * directly support a[b[i]] constructs. + */ + + + /*indexStore = tempstore();*/ + } + + + if (elemSize > 4) { + /* need to multiply array index by array element size */ + struct prog_instruction *inst; + slang_ir_storage *indexTemp; + slang_ir_storage elemSizeStore; + + /* allocate 1 float indexTemp */ + indexTemp = _slang_new_ir_storage(PROGRAM_TEMPORARY, -1, 1); + _slang_alloc_temp(emitInfo->vt, indexTemp); + + /* allocate a constant containing the element size */ + constant_to_storage(emitInfo, (float) elemSizeVec, &elemSizeStore); + + /* multiply array index by element size */ + inst = emit_instruction(emitInfo, + OPCODE_MUL, + indexTemp, /* dest */ + indexStore, /* the index */ + &elemSizeStore, + NULL); + + indexStore = indexTemp; + } + + if (arrayStore->IsIndirect) { + /* ex: in a[i][j], a[i] (the arrayStore) is indirect */ + /* Need to add indexStore to arrayStore->Indirect store */ + slang_ir_storage indirectArray; + slang_ir_storage *indexTemp; + + _slang_init_ir_storage(&indirectArray, + arrayStore->IndirectFile, + arrayStore->IndirectIndex, + 1, + arrayStore->IndirectSwizzle); + + /* allocate 1 float indexTemp */ + indexTemp = _slang_new_ir_storage(PROGRAM_TEMPORARY, -1, 1); + _slang_alloc_temp(emitInfo->vt, indexTemp); + + inst = emit_instruction(emitInfo, + OPCODE_ADD, + indexTemp, /* dest */ + indexStore, /* the index */ + &indirectArray, /* indirect array base */ + NULL); + + indexStore = indexTemp; + } + + /* update the array element storage info */ + n->Store->IsIndirect = GL_TRUE; + n->Store->IndirectFile = indexStore->File; + n->Store->IndirectIndex = indexStore->Index; + n->Store->IndirectSwizzle = indexStore->Swizzle; } - return size; + + n->Store->Size = elemSize; + n->Store->Swizzle = _slang_var_swizzle(elemSize, 0); + + return NULL; /* no instruction */ } +/** + * Resolve storage for accessing a structure field. + */ static struct prog_instruction * -emit_swizzle(slang_emit_info *emitInfo, slang_ir_node *n) +emit_struct_field(slang_emit_info *emitInfo, slang_ir_node *n) { - GLuint swizzle; - struct prog_instruction *inst; + slang_ir_storage *root = n->Store; + GLint fieldOffset, fieldSize; - inst = emit(emitInfo, n->Children[0]); + assert(n->Opcode == IR_FIELD); -#ifdef DEBUG - { - GLuint s = n->Children[0]->Store->Swizzle; - assert(GET_SWZ(s, 0) != SWIZZLE_NIL); - assert(GET_SWZ(s, 1) != SWIZZLE_NIL); - assert(GET_SWZ(s, 2) != SWIZZLE_NIL); - assert(GET_SWZ(s, 3) != SWIZZLE_NIL); + assert(!root->Parent); + while (root->Parent) + root = root->Parent; + + /* If this is the field of a state var, allocate constant/uniform + * storage for it now if we haven't already. + * Note that we allocate storage (uniform/constant slots) for state + * variables here rather than at declaration time so we only allocate + * space for the ones that we actually use! + */ + if (root->File == PROGRAM_STATE_VAR) { + GLboolean direct; + GLint index = _slang_alloc_statevar(n, emitInfo->prog->Parameters, &direct); + if (index < 0) { + slang_info_log_error(emitInfo->log, "Error parsing state variable"); + return NULL; + } + if (direct) { + root->Index = index; + return NULL; /* all done */ + } } -#endif - /* For debug: n->Var = n->Children[0]->Var; */ - /* "pull-up" the child's storage info, applying our swizzle info */ - n->Store->File = n->Children[0]->Store->File; - n->Store->Index = n->Children[0]->Store->Index; - n->Store->Size = swizzle_size(n->Store->Swizzle); -#if 0 - printf("Emit Swizzle %s reg %d chSize %d mySize %d\n", - _mesa_swizzle_string(n->Store->Swizzle, 0, 0), - n->Store->Index, n->Children[0]->Store->Size, - n->Store->Size); -#endif + /* do codegen for struct */ + emit(emitInfo, n->Children[0]); + assert(n->Children[0]->Store->Index >= 0); - /* apply this swizzle to child's swizzle to get composed swizzle */ - swizzle = fix_swizzle(n->Store->Swizzle); /* remove the don't care terms */ - n->Store->Swizzle = swizzle_swizzle(n->Children[0]->Store->Swizzle, - swizzle); - return inst; + fieldOffset = n->Store->Index; + fieldSize = n->Store->Size; + + _slang_copy_ir_storage(n->Store, n->Children[0]->Store); + + n->Store->Index = n->Children[0]->Store->Index + fieldOffset / 4; + n->Store->Size = fieldSize; + + switch (fieldSize) { + case 1: + { + GLint swz = fieldOffset % 4; + n->Store->Swizzle = MAKE_SWIZZLE4(swz, swz, swz, swz); + } + break; + case 2: + n->Store->Swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, + SWIZZLE_NIL, SWIZZLE_NIL); + break; + case 3: + n->Store->Swizzle = MAKE_SWIZZLE4(SWIZZLE_X, SWIZZLE_Y, + SWIZZLE_Z, SWIZZLE_NIL); + break; + default: + n->Store->Swizzle = SWIZZLE_XYZW; + } + + assert(n->Store->Index >= 0); + + return NULL; /* no instruction */ } /** - * Dereference array element. Just resolve storage for the array - * element represented by this node. + * Emit code for a variable declaration. + * This usually doesn't result in any code generation, but just + * memory allocation. */ static struct prog_instruction * -emit_array_element(slang_emit_info *emitInfo, slang_ir_node *n) +emit_var_decl(slang_emit_info *emitInfo, slang_ir_node *n) { assert(n->Store); assert(n->Store->File != PROGRAM_UNDEFINED); assert(n->Store->Size > 0); - - if (n->Store->File == PROGRAM_STATE_VAR) { - n->Store->Index = _slang_alloc_statevar(n, emitInfo->prog->Parameters); - return NULL; - } - - if (n->Children[1]->Opcode == IR_FLOAT) { - /* Constant index */ - const GLint arrayAddr = n->Children[0]->Store->Index; - const GLint index = (GLint) n->Children[1]->Value[0]; - n->Store->Index = arrayAddr + index; + /*assert(n->Store->Index < 0);*/ + + if (!n->Var || n->Var->isTemp) { + /* a nameless/temporary variable, will be freed after first use */ + /*NEW*/ + if (n->Store->Index < 0 && !_slang_alloc_temp(emitInfo->vt, n->Store)) { + slang_info_log_error(emitInfo->log, + "Ran out of registers, too many temporaries"); + return NULL; + } } else { - /* Variable index - PROBLEM */ - const GLint arrayAddr = n->Children[0]->Store->Index; - const GLint index = 0; - _mesa_problem(NULL, "variable array indexes not supported yet!"); - n->Store->Index = arrayAddr + index; + /* a regular variable */ + _slang_add_variable(emitInfo->vt, n->Var); + if (!_slang_alloc_var(emitInfo->vt, n->Store)) { + slang_info_log_error(emitInfo->log, + "Ran out of registers, too many variables"); + return NULL; + } + /* + printf("IR_VAR_DECL %s %d store %p\n", + (char*) n->Var->a_name, n->Store->Index, (void*) n->Store); + */ + assert(n->Var->store == n->Store); + } + if (emitInfo->EmitComments) { + /* emit NOP with comment describing the variable's storage location */ + char s[1000]; + sprintf(s, "TEMP[%d]%s = variable %s (size %d)", + n->Store->Index, + _mesa_swizzle_string(n->Store->Swizzle, 0, GL_FALSE), + (n->Var ? (char *) n->Var->a_name : "anonymous"), + n->Store->Size); + emit_comment(emitInfo, s); } - return NULL; /* no instruction */ + return NULL; } /** - * Resolve storage for accessing a structure field. + * Emit code for a reference to a variable. + * Actually, no code is generated but we may do some memory allocation. + * In particular, state vars (uniforms) are allocated on an as-needed basis. */ static struct prog_instruction * -emit_struct_field(slang_emit_info *emitInfo, slang_ir_node *n) +emit_var_ref(slang_emit_info *emitInfo, slang_ir_node *n) { - if (n->Store->File == PROGRAM_STATE_VAR) { - n->Store->Index = _slang_alloc_statevar(n, emitInfo->prog->Parameters); - } - else { - GLint offset = n->FieldOffset / 4; - assert(n->Children[0]->Store->Index >= 0); - n->Store->Index = n->Children[0]->Store->Index + offset; - if (n->Store->Size == 1) { - GLint swz = n->FieldOffset % 4; - n->Store->Swizzle = MAKE_SWIZZLE4(swz, swz, swz, swz); - } - else { - n->Store->Swizzle = SWIZZLE_XYZW; + assert(n->Store); + assert(n->Store->File != PROGRAM_UNDEFINED); + + if (n->Store->File == PROGRAM_STATE_VAR && n->Store->Index < 0) { + GLboolean direct; + GLint index = _slang_alloc_statevar(n, emitInfo->prog->Parameters, &direct); + if (index < 0) { + /* error */ + char s[100]; + snprintf(s, sizeof(s), "Undefined variable '%s'", + (char *) n->Var->a_name); + slang_info_log_error(emitInfo->log, s); + return NULL; } + + n->Store->Index = index; } - return NULL; /* no instruction */ + else if (n->Store->File == PROGRAM_UNIFORM) { + /* mark var as used */ + _mesa_use_uniform(emitInfo->prog->Parameters, (char *) n->Var->a_name); + } + + if (n->Store->Index < 0) { + /* probably ran out of registers */ + return NULL; + } + assert(n->Store->Size > 0); + + return NULL; } @@ -1454,12 +2126,18 @@ emit(slang_emit_info *emitInfo, slang_ir_node *n) if (!n) return NULL; + if (emitInfo->log->error_flag) { + return NULL; + } + switch (n->Opcode) { case IR_SEQ: /* sequence of two sub-trees */ assert(n->Children[0]); assert(n->Children[1]); emit(emitInfo, n->Children[0]); + if (emitInfo->log->error_flag) + return NULL; inst = emit(emitInfo, n->Children[1]); #if 0 assert(!n->Store); @@ -1476,64 +2154,14 @@ emit(slang_emit_info *emitInfo, slang_ir_node *n) case IR_VAR_DECL: /* Variable declaration - allocate a register for it */ - assert(n->Store); - assert(n->Store->File != PROGRAM_UNDEFINED); - assert(n->Store->Size > 0); - /*assert(n->Store->Index < 0);*/ - if (!n->Var || n->Var->isTemp) { - /* a nameless/temporary variable, will be freed after first use */ - /*NEW*/ - if (n->Store->Index < 0 && !_slang_alloc_temp(emitInfo->vt, n->Store)) { - slang_info_log_error(emitInfo->log, - "Ran out of registers, too many temporaries"); - return NULL; - } - } - else { - /* a regular variable */ - _slang_add_variable(emitInfo->vt, n->Var); - if (!_slang_alloc_var(emitInfo->vt, n->Store)) { - slang_info_log_error(emitInfo->log, - "Ran out of registers, too many variables"); - return NULL; - } - /* - printf("IR_VAR_DECL %s %d store %p\n", - (char*) n->Var->a_name, n->Store->Index, (void*) n->Store); - */ - assert(n->Var->aux == n->Store); - } - if (emitInfo->EmitComments) { - /* emit NOP with comment describing the variable's storage location */ - char s[1000]; - sprintf(s, "TEMP[%d]%s = variable %s (size %d)", - n->Store->Index, - _mesa_swizzle_string(n->Store->Swizzle, 0, GL_FALSE), - (n->Var ? (char *) n->Var->a_name : "anonymous"), - n->Store->Size); - inst = emit_comment(emitInfo, s); - return inst; - } - return NULL; + inst = emit_var_decl(emitInfo, n); + return inst; case IR_VAR: /* Reference to a variable * Storage should have already been resolved/allocated. */ - assert(n->Store); - assert(n->Store->File != PROGRAM_UNDEFINED); - - if (n->Store->File == PROGRAM_STATE_VAR && - n->Store->Index < 0) { - n->Store->Index = _slang_alloc_statevar(n, emitInfo->prog->Parameters); - } - - if (n->Store->Index < 0) { - printf("#### VAR %s not allocated!\n", (char*)n->Var->a_name); - } - assert(n->Store->Index >= 0); - assert(n->Store->Size > 0); - break; + return emit_var_ref(emitInfo, n); case IR_ELEMENT: return emit_array_element(emitInfo, n); @@ -1542,42 +2170,36 @@ emit(slang_emit_info *emitInfo, slang_ir_node *n) case IR_SWIZZLE: return emit_swizzle(emitInfo, n); - case IR_I_TO_F: - /* just move */ - emit(emitInfo, n->Children[0]); - inst = new_instruction(emitInfo, OPCODE_MOV); - if (!n->Store) { - if (!alloc_temp_storage(emitInfo, n, 1)) - return NULL; - } - storage_to_dst_reg(&inst->DstReg, n->Store, n->Writemask); - storage_to_src_reg(&inst->SrcReg[0], n->Children[0]->Store); - if (emitInfo->EmitComments) - inst->Comment = _mesa_strdup("int to float"); - return NULL; - /* Simple arithmetic */ /* unary */ + case IR_MOVE: case IR_RSQ: case IR_RCP: case IR_FLOOR: case IR_FRAC: case IR_F_TO_I: + case IR_I_TO_F: case IR_ABS: case IR_SIN: case IR_COS: case IR_DDX: case IR_DDY: + case IR_EXP: + case IR_EXP2: + case IR_LOG2: case IR_NOISE1: case IR_NOISE2: case IR_NOISE3: case IR_NOISE4: + case IR_NRM4: + case IR_NRM3: /* binary */ case IR_ADD: case IR_SUB: case IR_MUL: case IR_DOT4: case IR_DOT3: + case IR_DOT2: case IR_CROSS: case IR_MIN: case IR_MAX: @@ -1588,8 +2210,6 @@ emit(slang_emit_info *emitInfo, slang_ir_node *n) case IR_SLE: case IR_SLT: case IR_POW: - case IR_EXP: - case IR_EXP2: /* trinary operators */ case IR_LRP: return emit_arith(emitInfo, n); @@ -1618,8 +2238,8 @@ emit(slang_emit_info *emitInfo, slang_ir_node *n) } return NULL; - case IR_MOVE: - return emit_move(emitInfo, n); + case IR_COPY: + return emit_copy(emitInfo, n); case IR_COND: return emit_cond(emitInfo, n); @@ -1633,10 +2253,10 @@ emit(slang_emit_info *emitInfo, slang_ir_node *n) case IR_KILL: return emit_kill(emitInfo); - case IR_FUNC: - /* new variable scope for subroutines/function calls*/ + case IR_CALL: + /* new variable scope for subroutines/function calls */ _slang_push_var_table(emitInfo->vt); - inst = emit_func(emitInfo, n); + inst = emit_fcall(emitInfo, n); _slang_pop_var_table(emitInfo->vt); return inst; @@ -1645,12 +2265,9 @@ emit(slang_emit_info *emitInfo, slang_ir_node *n) case IR_LOOP: return emit_loop(emitInfo, n); - case IR_BREAK_IF_FALSE: - case IR_CONT_IF_FALSE: - return emit_cont_break_if(emitInfo, n, GL_FALSE); case IR_BREAK_IF_TRUE: case IR_CONT_IF_TRUE: - return emit_cont_break_if(emitInfo, n, GL_TRUE); + return emit_cont_break_if_true(emitInfo, n); case IR_BREAK: /* fall-through */ case IR_CONT: @@ -1696,7 +2313,7 @@ _slang_resolve_subroutines(slang_emit_info *emitInfo) total += emitInfo->Subroutines[i]->NumInstructions; } - /* adjust BrancTargets within the functions */ + /* adjust BranchTargets within the functions */ for (i = 0; i < emitInfo->NumSubroutines; i++) { struct gl_program *sub = emitInfo->Subroutines[i]; GLuint j; @@ -1712,6 +2329,7 @@ _slang_resolve_subroutines(slang_emit_info *emitInfo) mainP->Instructions = _mesa_realloc_instructions(mainP->Instructions, mainP->NumInstructions, total); + mainP->NumInstructions = total; for (i = 0; i < emitInfo->NumSubroutines; i++) { struct gl_program *sub = emitInfo->Subroutines[i]; _mesa_copy_instructions(mainP->Instructions + subroutineLoc[i], @@ -1719,12 +2337,18 @@ _slang_resolve_subroutines(slang_emit_info *emitInfo) sub->NumInstructions); /* delete subroutine code */ sub->Parameters = NULL; /* prevent double-free */ - _mesa_delete_program(ctx, sub); + _mesa_reference_program(ctx, &emitInfo->Subroutines[i], NULL); } - mainP->NumInstructions = total; + + /* free subroutine list */ + if (emitInfo->Subroutines) { + _mesa_free(emitInfo->Subroutines); + emitInfo->Subroutines = NULL; + } + emitInfo->NumSubroutines = 0; /* Examine CAL instructions. - * At this point, the BranchTarget field of the CAL instructions is + * At this point, the BranchTarget field of the CAL instruction is * the number/id of the subroutine to call (an index into the * emitInfo->Subroutines list). * Translate that into an actual instruction location now. @@ -1751,22 +2375,37 @@ _slang_emit_code(slang_ir_node *n, slang_var_table *vt, GET_CURRENT_CONTEXT(ctx); GLboolean success; slang_emit_info emitInfo; + GLuint maxUniforms; emitInfo.log = log; emitInfo.vt = vt; emitInfo.prog = prog; emitInfo.Subroutines = NULL; emitInfo.NumSubroutines = 0; + emitInfo.MaxInstructions = prog->NumInstructions; emitInfo.EmitHighLevelInstructions = ctx->Shader.EmitHighLevelInstructions; emitInfo.EmitCondCodes = ctx->Shader.EmitCondCodes; emitInfo.EmitComments = ctx->Shader.EmitComments; - emitInfo.EmitBeginEndSub = 0; /* XXX for compiler debug only */ + emitInfo.EmitBeginEndSub = GL_TRUE; if (!emitInfo.EmitCondCodes) { emitInfo.EmitHighLevelInstructions = GL_TRUE; } + /* Check uniform/constant limits */ + if (prog->Target == GL_FRAGMENT_PROGRAM_ARB) { + maxUniforms = ctx->Const.FragmentProgram.MaxUniformComponents / 4; + } + else { + assert(prog->Target == GL_VERTEX_PROGRAM_ARB); + maxUniforms = ctx->Const.VertexProgram.MaxUniformComponents / 4; + } + if (prog->Parameters->NumParameters > maxUniforms) { + slang_info_log_error(log, "Constant/uniform register limit exceeded"); + return GL_FALSE; + } + (void) emit(&emitInfo, n); /* finish up by adding the END opcode to program */