Nikilite/suyu
1
1
Fork 0
forked from suyu/suyu
Code Pull requests Activity

lower_int64_to_int32: Add 64-bit atomic fallbacks

Browse source
This commit is contained in:
ameerj 2022-01-29 14:06:03 -05:00
parent ad58d7eae7
commit 90a0506d56
3 changed files with 76 additions and 11 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

18
src/shader_recompiler/frontend/ir/opcodes.inc
View file

@ -363,15 +363,15 @@ OPCODE(GlobalAtomicAnd64, U64, U64,
OPCODE(GlobalAtomicOr64, U64, U64, U64, ) OPCODE(GlobalAtomicOr64, U64, U64, U64, )
OPCODE(GlobalAtomicXor64, U64, U64, U64, ) OPCODE(GlobalAtomicXor64, U64, U64, U64, )
OPCODE(GlobalAtomicExchange64, U64, U64, U64, ) OPCODE(GlobalAtomicExchange64, U64, U64, U64, )
OPCODE(GlobalAtomicIAdd32x2, U32x2, U64, U32x2, ) OPCODE(GlobalAtomicIAdd32x2, U32x2, U32x2, U32x2, )
OPCODE(GlobalAtomicSMin32x2, U32x2, U64, U32x2, ) OPCODE(GlobalAtomicSMin32x2, U32x2, U32x2, U32x2, )
OPCODE(GlobalAtomicUMin32x2, U32x2, U64, U32x2, ) OPCODE(GlobalAtomicUMin32x2, U32x2, U32x2, U32x2, )
OPCODE(GlobalAtomicSMax32x2, U32x2, U64, U32x2, ) OPCODE(GlobalAtomicSMax32x2, U32x2, U32x2, U32x2, )
OPCODE(GlobalAtomicUMax32x2, U32x2, U64, U32x2, ) OPCODE(GlobalAtomicUMax32x2, U32x2, U32x2, U32x2, )
OPCODE(GlobalAtomicAnd32x2, U32x2, U64, U32x2, ) OPCODE(GlobalAtomicAnd32x2, U32x2, U32x2, U32x2, )
OPCODE(GlobalAtomicOr32x2, U32x2, U64, U32x2, ) OPCODE(GlobalAtomicOr32x2, U32x2, U32x2, U32x2, )
OPCODE(GlobalAtomicXor32x2, U32x2, U64, U32x2, ) OPCODE(GlobalAtomicXor32x2, U32x2, U32x2, U32x2, )
OPCODE(GlobalAtomicExchange32x2, U32x2, U64, U32x2, ) OPCODE(GlobalAtomicExchange32x2, U32x2, U32x2, U32x2, )
OPCODE(GlobalAtomicAddF32, F32, U64, F32, ) OPCODE(GlobalAtomicAddF32, F32, U64, F32, )
OPCODE(GlobalAtomicAddF16x2, U32, U64, F16x2, ) OPCODE(GlobalAtomicAddF16x2, U32, U64, F16x2, )
OPCODE(GlobalAtomicAddF32x2, U32, U64, F32x2, ) OPCODE(GlobalAtomicAddF32x2, U32, U64, F32x2, )

49
src/shader_recompiler/ir_opt/global_memory_to_storage_buffer_pass.cpp
View file

@ -92,6 +92,15 @@ bool IsGlobalMemory(const IR::Inst& inst) {
case IR::Opcode::GlobalAtomicOr64: case IR::Opcode::GlobalAtomicOr64:
case IR::Opcode::GlobalAtomicXor64: case IR::Opcode::GlobalAtomicXor64:
case IR::Opcode::GlobalAtomicExchange64: case IR::Opcode::GlobalAtomicExchange64:
case IR::Opcode::GlobalAtomicIAdd32x2:
case IR::Opcode::GlobalAtomicSMin32x2:
case IR::Opcode::GlobalAtomicUMin32x2:
case IR::Opcode::GlobalAtomicSMax32x2:
case IR::Opcode::GlobalAtomicUMax32x2:
case IR::Opcode::GlobalAtomicAnd32x2:
case IR::Opcode::GlobalAtomicOr32x2:
case IR::Opcode::GlobalAtomicXor32x2:
case IR::Opcode::GlobalAtomicExchange32x2:
case IR::Opcode::GlobalAtomicAddF32: case IR::Opcode::GlobalAtomicAddF32:
case IR::Opcode::GlobalAtomicAddF16x2: case IR::Opcode::GlobalAtomicAddF16x2:
case IR::Opcode::GlobalAtomicAddF32x2: case IR::Opcode::GlobalAtomicAddF32x2:
@ -135,6 +144,15 @@ bool IsGlobalMemoryWrite(const IR::Inst& inst) {
case IR::Opcode::GlobalAtomicOr64: case IR::Opcode::GlobalAtomicOr64:
case IR::Opcode::GlobalAtomicXor64: case IR::Opcode::GlobalAtomicXor64:
case IR::Opcode::GlobalAtomicExchange64: case IR::Opcode::GlobalAtomicExchange64:
case IR::Opcode::GlobalAtomicIAdd32x2:
case IR::Opcode::GlobalAtomicSMin32x2:
case IR::Opcode::GlobalAtomicUMin32x2:
case IR::Opcode::GlobalAtomicSMax32x2:
case IR::Opcode::GlobalAtomicUMax32x2:
case IR::Opcode::GlobalAtomicAnd32x2:
case IR::Opcode::GlobalAtomicOr32x2:
case IR::Opcode::GlobalAtomicXor32x2:
case IR::Opcode::GlobalAtomicExchange32x2:
case IR::Opcode::GlobalAtomicAddF32: case IR::Opcode::GlobalAtomicAddF32:
case IR::Opcode::GlobalAtomicAddF16x2: case IR::Opcode::GlobalAtomicAddF16x2:
case IR::Opcode::GlobalAtomicAddF32x2: case IR::Opcode::GlobalAtomicAddF32x2:
@ -199,6 +217,8 @@ IR::Opcode GlobalToStorage(IR::Opcode opcode) {
return IR::Opcode::StorageAtomicOr32; return IR::Opcode::StorageAtomicOr32;
case IR::Opcode::GlobalAtomicXor32: case IR::Opcode::GlobalAtomicXor32:
return IR::Opcode::StorageAtomicXor32; return IR::Opcode::StorageAtomicXor32;
case IR::Opcode::GlobalAtomicExchange32:
return IR::Opcode::StorageAtomicExchange32;
case IR::Opcode::GlobalAtomicIAdd64: case IR::Opcode::GlobalAtomicIAdd64:
return IR::Opcode::StorageAtomicIAdd64; return IR::Opcode::StorageAtomicIAdd64;
case IR::Opcode::GlobalAtomicSMin64: case IR::Opcode::GlobalAtomicSMin64:
@ -215,10 +235,26 @@ IR::Opcode GlobalToStorage(IR::Opcode opcode) {
return IR::Opcode::StorageAtomicOr64; return IR::Opcode::StorageAtomicOr64;
case IR::Opcode::GlobalAtomicXor64: case IR::Opcode::GlobalAtomicXor64:
return IR::Opcode::StorageAtomicXor64; return IR::Opcode::StorageAtomicXor64;
case IR::Opcode::GlobalAtomicExchange32:
return IR::Opcode::StorageAtomicExchange32;
case IR::Opcode::GlobalAtomicExchange64: case IR::Opcode::GlobalAtomicExchange64:
return IR::Opcode::StorageAtomicExchange64; return IR::Opcode::StorageAtomicExchange64;
case IR::Opcode::GlobalAtomicIAdd32x2:
return IR::Opcode::StorageAtomicIAdd32x2;
case IR::Opcode::GlobalAtomicSMin32x2:
return IR::Opcode::StorageAtomicSMin32x2;
case IR::Opcode::GlobalAtomicUMin32x2:
return IR::Opcode::StorageAtomicUMin32x2;
case IR::Opcode::GlobalAtomicSMax32x2:
return IR::Opcode::StorageAtomicSMax32x2;
case IR::Opcode::GlobalAtomicUMax32x2:
return IR::Opcode::StorageAtomicUMax32x2;
case IR::Opcode::GlobalAtomicAnd32x2:
return IR::Opcode::StorageAtomicAnd32x2;
case IR::Opcode::GlobalAtomicOr32x2:
return IR::Opcode::StorageAtomicOr32x2;
case IR::Opcode::GlobalAtomicXor32x2:
return IR::Opcode::StorageAtomicXor32x2;
case IR::Opcode::GlobalAtomicExchange32x2:
return IR::Opcode::StorageAtomicExchange32x2;
case IR::Opcode::GlobalAtomicAddF32: case IR::Opcode::GlobalAtomicAddF32:
return IR::Opcode::StorageAtomicAddF32; return IR::Opcode::StorageAtomicAddF32;
case IR::Opcode::GlobalAtomicAddF16x2: case IR::Opcode::GlobalAtomicAddF16x2:
@ -454,6 +490,15 @@ void Replace(IR::Block& block, IR::Inst& inst, const IR::U32& storage_index,
case IR::Opcode::GlobalAtomicOr64: case IR::Opcode::GlobalAtomicOr64:
case IR::Opcode::GlobalAtomicXor64: case IR::Opcode::GlobalAtomicXor64:
case IR::Opcode::GlobalAtomicExchange64: case IR::Opcode::GlobalAtomicExchange64:
case IR::Opcode::GlobalAtomicIAdd32x2:
case IR::Opcode::GlobalAtomicSMin32x2:
case IR::Opcode::GlobalAtomicUMin32x2:
case IR::Opcode::GlobalAtomicSMax32x2:
case IR::Opcode::GlobalAtomicUMax32x2:
case IR::Opcode::GlobalAtomicAnd32x2:
case IR::Opcode::GlobalAtomicOr32x2:
case IR::Opcode::GlobalAtomicXor32x2:
case IR::Opcode::GlobalAtomicExchange32x2:
case IR::Opcode::GlobalAtomicAddF32: case IR::Opcode::GlobalAtomicAddF32:
case IR::Opcode::GlobalAtomicAddF16x2: case IR::Opcode::GlobalAtomicAddF16x2:
case IR::Opcode::GlobalAtomicAddF32x2: case IR::Opcode::GlobalAtomicAddF32x2:

20
src/shader_recompiler/ir_opt/lower_int64_to_int32.cpp
View file

@ -199,6 +199,26 @@ void Lower(IR::Block& block, IR::Inst& inst) {
return ShiftRightLogical64To32(block, inst); return ShiftRightLogical64To32(block, inst);
case IR::Opcode::ShiftRightArithmetic64: case IR::Opcode::ShiftRightArithmetic64:
return ShiftRightArithmetic64To32(block, inst); return ShiftRightArithmetic64To32(block, inst);
case IR::Opcode::SharedAtomicExchange64:
return inst.ReplaceOpcode(IR::Opcode::SharedAtomicExchange32x2);
case IR::Opcode::GlobalAtomicIAdd64:
return inst.ReplaceOpcode(IR::Opcode::GlobalAtomicIAdd32x2);
case IR::Opcode::GlobalAtomicSMin64:
return inst.ReplaceOpcode(IR::Opcode::GlobalAtomicSMin32x2);
case IR::Opcode::GlobalAtomicUMin64:
return inst.ReplaceOpcode(IR::Opcode::GlobalAtomicUMin32x2);
case IR::Opcode::GlobalAtomicSMax64:
return inst.ReplaceOpcode(IR::Opcode::GlobalAtomicSMax32x2);
case IR::Opcode::GlobalAtomicUMax64:
return inst.ReplaceOpcode(IR::Opcode::GlobalAtomicUMax32x2);
case IR::Opcode::GlobalAtomicAnd64:
return inst.ReplaceOpcode(IR::Opcode::GlobalAtomicAnd32x2);
case IR::Opcode::GlobalAtomicOr64:
return inst.ReplaceOpcode(IR::Opcode::GlobalAtomicOr32x2);
case IR::Opcode::GlobalAtomicXor64:
return inst.ReplaceOpcode(IR::Opcode::GlobalAtomicXor32x2);
case IR::Opcode::GlobalAtomicExchange64:
return inst.ReplaceOpcode(IR::Opcode::GlobalAtomicExchange32x2);
default: default:
break; break;
} }