summaryrefslogtreecommitdiffstats
path: root/src/amd/compiler/aco_optimizer.cpp
diff options
context:
space:
mode:
authorDaniel Schürmann <[email protected]>2019-10-17 15:06:48 +0200
committerRhys Perry <[email protected]>2019-10-17 16:21:19 +0000
commit4b458b3e8f5fb01c23677a905554c070a8809fba (patch)
treefaebde64da374c3afb8b8cfc441051cc4d7a4f31 /src/amd/compiler/aco_optimizer.cpp
parent045f05a2f68c301c850a38a511c8e254080fce9a (diff)
aco: don't combine minmax3 if there is a neg or abs modifier in between
This fixes a graphical corruption in HotS. No pipelinedb changes other than that. Reviewed-by: Rhys Perry <[email protected]>
Diffstat (limited to 'src/amd/compiler/aco_optimizer.cpp')
-rw-r--r--src/amd/compiler/aco_optimizer.cpp30
1 files changed, 1 insertions, 29 deletions
diff --git a/src/amd/compiler/aco_optimizer.cpp b/src/amd/compiler/aco_optimizer.cpp
index 6d600d55254..501e48162be 100644
--- a/src/amd/compiler/aco_optimizer.cpp
+++ b/src/amd/compiler/aco_optimizer.cpp
@@ -1483,34 +1483,6 @@ void create_vop3_for_op3(opt_ctx& ctx, aco_opcode opcode, aco_ptr<Instruction>&
instr.reset(new_instr);
}
-bool combine_minmax3(opt_ctx& ctx, aco_ptr<Instruction>& instr, aco_opcode new_op)
-{
- uint32_t omod_clamp = ctx.info[instr->definitions[0].tempId()].label &
- (label_omod_success | label_clamp_success);
-
- for (unsigned swap = 0; swap < 2; swap++) {
- Operand operands[3];
- bool neg[3], abs[3], opsel[3], clamp, inbetween_neg, inbetween_abs;
- unsigned omod;
- if (match_op3_for_vop3(ctx, instr->opcode, instr->opcode, instr.get(), swap,
- "012", operands, neg, abs, opsel,
- &clamp, &omod, &inbetween_neg, &inbetween_abs, NULL)) {
- ctx.uses[instr->operands[swap].tempId()]--;
- neg[1] ^= inbetween_neg;
- neg[2] ^= inbetween_neg;
- abs[1] |= inbetween_abs;
- abs[2] |= inbetween_abs;
- create_vop3_for_op3(ctx, new_op, instr, operands, neg, abs, opsel, clamp, omod);
- if (omod_clamp & label_omod_success)
- ctx.info[instr->definitions[0].tempId()].set_omod_success(instr.get());
- if (omod_clamp & label_clamp_success)
- ctx.info[instr->definitions[0].tempId()].set_clamp_success(instr.get());
- return true;
- }
- }
- return false;
-}
-
bool combine_three_valu_op(opt_ctx& ctx, aco_ptr<Instruction>& instr, aco_opcode op2, aco_opcode new_op, const char *shuffle, uint8_t ops)
{
uint32_t omod_clamp = ctx.info[instr->definitions[0].tempId()].label &
@@ -2202,7 +2174,7 @@ void combine_instruction(opt_ctx &ctx, aco_ptr<Instruction>& instr)
bool some_gfx9_only;
if (get_minmax_info(instr->opcode, &min, &max, &min3, &max3, &med3, &some_gfx9_only) &&
(!some_gfx9_only || ctx.program->chip_class >= GFX9)) {
- if (combine_minmax3(ctx, instr, instr->opcode == min ? min3 : max3)) ;
+ if (combine_three_valu_op(ctx, instr, instr->opcode, instr->opcode == min ? min3 : max3, "012", 1 | 2));
else combine_clamp(ctx, instr, min, max, med3);
}
}