From 6900046aad77b3b01feb30af9e50375224747b7d Mon Sep 17 00:00:00 2001 From: Maciej Cencora Date: Mon, 20 Apr 2009 07:31:44 -0600 Subject: [PATCH] r300: fix register-negate branch merge regression --- .../drivers/dri/r300/r300_fragprog_swizzle.c | 16 +++++------ .../drivers/dri/r300/radeon_program_alu.c | 27 +++---------------- .../drivers/dri/r300/radeon_program_pair.c | 8 ++---- 3 files changed, 12 insertions(+), 39 deletions(-) diff --git a/src/mesa/drivers/dri/r300/r300_fragprog_swizzle.c b/src/mesa/drivers/dri/r300/r300_fragprog_swizzle.c index 191853ac1f2..6f0d733d463 100644 --- a/src/mesa/drivers/dri/r300/r300_fragprog_swizzle.c +++ b/src/mesa/drivers/dri/r300/r300_fragprog_swizzle.c @@ -122,7 +122,7 @@ GLboolean r300FPIsNativeSwizzle(GLuint opcode, struct prog_src_register reg) if (GET_SWZ(reg.Swizzle, j) != SWIZZLE_NIL) relevant |= 1 << j; - if ((reg.Negate & relevant) && (reg.Negate & relevant) != relevant) + if ((reg.Negate & relevant) && ((reg.Negate & relevant) != relevant)) return GL_FALSE; if (!lookup_native_swizzle(reg.Swizzle)) @@ -144,7 +144,6 @@ void r300FPBuildSwizzle(struct nqssadce_state *s, struct prog_dst_register dst, const struct swizzle_data *best_swizzle = 0; GLuint best_matchcount = 0; GLuint best_matchmask = 0; - GLboolean rgbnegate; int i, comp; for(i = 0; i < num_native_swizzles; ++i) { @@ -158,6 +157,11 @@ void r300FPBuildSwizzle(struct nqssadce_state *s, struct prog_dst_register dst, if (swz == SWIZZLE_NIL) continue; if (swz == GET_SWZ(sd->hash, comp)) { + /* check if the negate bit of current component + * is the same for already matched components */ + if (matchmask && (!!(src.Negate & matchmask) != !!(src.Negate & (1 << comp)))) + continue; + matchcount++; matchmask |= 1 << comp; } @@ -171,13 +175,6 @@ void r300FPBuildSwizzle(struct nqssadce_state *s, struct prog_dst_register dst, } } - if ((src.Negate & best_matchmask) != 0) { - best_matchmask &= src.Negate; - rgbnegate = !src.Negate; - } else { - rgbnegate = src.Negate; - } - struct prog_instruction *inst; _mesa_insert_instructions(s->Program, s->IP, 1); @@ -186,6 +183,7 @@ void r300FPBuildSwizzle(struct nqssadce_state *s, struct prog_dst_register dst, inst->DstReg = dst; inst->DstReg.WriteMask &= (best_matchmask | WRITEMASK_W); inst->SrcReg[0] = src; + inst->SrcReg[0].Negate = (best_matchmask & src.Negate) ? NEGATE_XYZW : NEGATE_NONE; /* Note: We rely on NqSSA/DCE to set unused swizzle components to NIL */ dst.WriteMask &= ~inst->DstReg.WriteMask; diff --git a/src/mesa/drivers/dri/r300/radeon_program_alu.c b/src/mesa/drivers/dri/r300/radeon_program_alu.c index ebc5c913b25..8283723bad7 100644 --- a/src/mesa/drivers/dri/r300/radeon_program_alu.c +++ b/src/mesa/drivers/dri/r300/radeon_program_alu.c @@ -81,19 +81,6 @@ static struct prog_instruction *emit3(struct gl_program* p, return fpi; } -static void set_swizzle(struct prog_src_register *SrcReg, int coordinate, int swz) -{ - SrcReg->Swizzle &= ~(7 << (3*coordinate)); - SrcReg->Swizzle |= swz << (3*coordinate); -} - -static void set_negate_base(struct prog_src_register *SrcReg, int coordinate, int negate) -{ - /* XXX note sure about this negation logic here */ - SrcReg->Negate &= ~(1 << coordinate); - SrcReg->Negate |= (negate << coordinate); -} - static struct prog_dst_register dstreg(int file, int index) { struct prog_dst_register dst; @@ -197,17 +184,9 @@ static void transform_DPH(struct radeon_transform_context* t, struct prog_instruction* inst) { struct prog_src_register src0 = inst->SrcReg[0]; - if (src0.Negate) { - if (src0.Abs) { - int tempreg = radeonFindFreeTemporary(t); - emit1(t->Program, OPCODE_MOV, 0, dstreg(PROGRAM_TEMPORARY, tempreg), src0); - src0 = srcreg(src0.File, src0.Index); - } else { - src0.Negate ^= NEGATE_XYZW; - } - } - set_swizzle(&src0, 3, SWIZZLE_ONE); - set_negate_base(&src0, 3, 0); + src0.Negate &= ~NEGATE_W; + src0.Swizzle &= ~(7 << (3 * 3)); + src0.Swizzle |= SWIZZLE_ONE << (3 * 3); emit2(t->Program, OPCODE_DP4, inst->SaturateMode, inst->DstReg, src0, inst->SrcReg[1]); } diff --git a/src/mesa/drivers/dri/r300/radeon_program_pair.c b/src/mesa/drivers/dri/r300/radeon_program_pair.c index ecc82ff8a81..4aa2319a454 100644 --- a/src/mesa/drivers/dri/r300/radeon_program_pair.c +++ b/src/mesa/drivers/dri/r300/radeon_program_pair.c @@ -721,7 +721,6 @@ static GLboolean fill_instruction_into_pair(struct pair_state *s, struct radeon_ if (pairinst->NeedRGB && !pairinst->IsTranscendent) { GLboolean srcrgb = GL_FALSE; GLboolean srcalpha = GL_FALSE; - GLuint negatebase = 0; int j; for(j = 0; j < 3; ++j) { GLuint swz = GET_SWZ(inst->SrcReg[i].Swizzle, j); @@ -729,8 +728,6 @@ static GLboolean fill_instruction_into_pair(struct pair_state *s, struct radeon_ srcrgb = GL_TRUE; else if (swz < 4) srcalpha = GL_TRUE; - if (swz != SWIZZLE_NIL && GET_BIT(inst->SrcReg[i].Negate, j)) - negatebase = 1; } source = alloc_pair_source(s, pair, inst->SrcReg[i], srcrgb, srcalpha); if (source < 0) @@ -738,12 +735,11 @@ static GLboolean fill_instruction_into_pair(struct pair_state *s, struct radeon_ pair->RGB.Arg[i].Source = source; pair->RGB.Arg[i].Swizzle = inst->SrcReg[i].Swizzle & 0x1ff; pair->RGB.Arg[i].Abs = inst->SrcReg[i].Abs; - pair->RGB.Arg[i].Negate = (negatebase & ~pair->RGB.Arg[i].Abs) ^ inst->SrcReg[i].Negate; + pair->RGB.Arg[i].Negate = !!(inst->SrcReg[i].Negate & (NEGATE_X | NEGATE_Y | NEGATE_Z)); } if (pairinst->NeedAlpha) { GLboolean srcrgb = GL_FALSE; GLboolean srcalpha = GL_FALSE; - GLuint negatebase = GET_BIT(inst->SrcReg[i].Negate, pairinst->IsTranscendent ? 0 : 3); GLuint swz = GET_SWZ(inst->SrcReg[i].Swizzle, pairinst->IsTranscendent ? 0 : 3); if (swz < 3) srcrgb = GL_TRUE; @@ -755,7 +751,7 @@ static GLboolean fill_instruction_into_pair(struct pair_state *s, struct radeon_ pair->Alpha.Arg[i].Source = source; pair->Alpha.Arg[i].Swizzle = swz; pair->Alpha.Arg[i].Abs = inst->SrcReg[i].Abs; - pair->Alpha.Arg[i].Negate = (negatebase & ~pair->RGB.Arg[i].Abs) ^ inst->SrcReg[i].Negate; + pair->Alpha.Arg[i].Negate = !!(inst->SrcReg[i].Negate & NEGATE_W); } } -- 2.30.2