eden-emu/eden
15
3
Fork 4
Code Issues3 Pull requests7 Projects Releases2 Packages Wiki Activity Actions2

glasm: Catch more register leaks

Browse source 
Add support for null registers. These are used when an instruction has
no usages.

This comes handy when an instruction is only used for its CC value, with
the caveat of having to invalidate all pseudo-instructions before
defining the instruction itself in the register allocator. This commits
changes this.

Workaround a bug on Nvidia's condition codes conditional execution using
branches.
This commit is contained in:
ReinUsesLisp 2021-05-25 02:22:21 -03:00 • committed by ameerj
parent a5ce5f4a65
commit 143db4b112
8 changed files with 114 additions and 41 deletions
src/shader_recompiler/backend/glasm
emit_glasm_integer.cpp

58
src/shader_recompiler/backend/glasm/emit_glasm_integer.cpp
View file

@ -9,6 +9,17 @@
namespace Shader::Backend::GLASM {
void EmitIAdd32(EmitContext& ctx, IR::Inst& inst, ScalarS32 a, ScalarS32 b) {
const std::array flags{
inst.GetAssociatedPseudoOperation(IR::Opcode::GetZeroFromOp),
inst.GetAssociatedPseudoOperation(IR::Opcode::GetSignFromOp),
inst.GetAssociatedPseudoOperation(IR::Opcode::GetCarryFromOp),
inst.GetAssociatedPseudoOperation(IR::Opcode::GetOverflowFromOp),
};
for (IR::Inst* const flag_inst : flags) {
if (flag_inst) {
flag_inst->Invalidate();
}
}
const bool cc{inst.HasAssociatedPseudoOperation()};
const std::string_view cc_mod{cc ? ".CC" : ""};
if (cc) {
@ -19,20 +30,22 @@ void EmitIAdd32(EmitContext& ctx, IR::Inst& inst, ScalarS32 a, ScalarS32 b) {
if (!cc) {
return;
}
static constexpr std::array<std::string_view, 4> masks{"EQ", "SF", "CF", "OF"};
const std::array flags{
inst.GetAssociatedPseudoOperation(IR::Opcode::GetZeroFromOp),
inst.GetAssociatedPseudoOperation(IR::Opcode::GetSignFromOp),
inst.GetAssociatedPseudoOperation(IR::Opcode::GetCarryFromOp),
inst.GetAssociatedPseudoOperation(IR::Opcode::GetOverflowFromOp),
};
for (size_t i = 0; i < flags.size(); ++i) {
if (flags[i]) {
const auto flag_ret{ctx.reg_alloc.Define(*flags[i])};
ctx.Add("MOV.S {},0;"
"MOV.S {}({}.x),-1;",
flag_ret, flag_ret, masks[i]);
flags[i]->Invalidate();
static constexpr std::array<std::string_view, 4> masks{"", "SF", "CF", "OF"};
for (size_t flag_index = 0; flag_index < flags.size(); ++flag_index) {
if (!flags[flag_index]) {
continue;
}
const auto flag_ret{ctx.reg_alloc.Define(*flags[flag_index])};
if (flag_index == 0) {
ctx.Add("SEQ.S {}.x,{}.x,0;", flag_ret, ret);
} else {
// We could use conditional execution here, but it's broken on Nvidia's compiler
ctx.Add("IF {}.x;"
"MOV.S {}.x,-1;"
"ELSE;"
"MOV.S {}.x,0;"
"ENDIF;",
masks[flag_index], flag_ret, flag_ret);
}
}
}
@ -136,6 +149,17 @@ void EmitBitFieldSExtract(EmitContext& ctx, IR::Inst& inst, ScalarS32 base, Scal
void EmitBitFieldUExtract(EmitContext& ctx, IR::Inst& inst, ScalarU32 base, ScalarU32 offset,
ScalarU32 count) {
const auto zero = inst.GetAssociatedPseudoOperation(IR::Opcode::GetZeroFromOp);
const auto sign = inst.GetAssociatedPseudoOperation(IR::Opcode::GetSignFromOp);
if (zero) {
zero->Invalidate();
}
if (sign) {
sign->Invalidate();
}
if (zero || sign) {
ctx.reg_alloc.InvalidateConditionCodes();
}
const Register ret{ctx.reg_alloc.Define(inst)};
if (count.type != Type::Register && offset.type != Type::Register) {
ctx.Add("BFE.U {},{{{},{},0,0}},{};", ret, count, offset, base);
@ -145,13 +169,11 @@ void EmitBitFieldUExtract(EmitContext& ctx, IR::Inst& inst, ScalarU32 base, Scal
"BFE.U {},RC,{};",
count, offset, ret, base);
}
if (const auto zero = inst.GetAssociatedPseudoOperation(IR::Opcode::GetZeroFromOp)) {
if (zero) {
ctx.Add("SEQ.S {},{},0;", *zero, ret);
zero->Invalidate();
}
if (const auto sign = inst.GetAssociatedPseudoOperation(IR::Opcode::GetSignFromOp)) {
if (sign) {
ctx.Add("SLT.S {},{},0;", *sign, ret);
sign->Invalidate();
}
}