diff options
| author | ReinUsesLisp <reinuseslisp@airmail.cc> | 2019-08-31 17:06:00 -0300 |
|---|---|---|
| committer | ReinUsesLisp <reinuseslisp@airmail.cc> | 2019-09-04 01:55:24 -0300 |
| commit | 77ef4fa9078b56c3fcaded3a618cf95fe21e66d4 (patch) | |
| tree | 7d21a7ab8db5c806f93f4d345a887e4f1d2c7a2f /src | |
| parent | 922c7f4e510c12a7e207ba08904c2523d99edd55 (diff) | |
shader/shift: Implement SHR wrapped and clamped variants
Nvidia defaults to wrapped shifts, but this is undefined behaviour on
OpenGL's spec. Explicitly mask/clamp according to what the guest shader
requires.
Diffstat (limited to 'src')
| -rw-r--r-- | src/video_core/engines/shader_bytecode.h | 4 | ||||
| -rw-r--r-- | src/video_core/shader/decode/shift.cpp | 19 |
2 files changed, 17 insertions, 6 deletions
diff --git a/src/video_core/engines/shader_bytecode.h b/src/video_core/engines/shader_bytecode.h index c3678b9eab..bd8c1ada03 100644 --- a/src/video_core/engines/shader_bytecode.h +++ b/src/video_core/engines/shader_bytecode.h @@ -675,6 +675,10 @@ union Instruction { } shift; union { + BitField<39, 1, u64> wrap; + } shr; + + union { BitField<39, 5, u64> shift_amount; BitField<48, 1, u64> negate_b; BitField<49, 1, u64> negate_a; diff --git a/src/video_core/shader/decode/shift.cpp b/src/video_core/shader/decode/shift.cpp index 2ac16eeb06..f6ee68a54c 100644 --- a/src/video_core/shader/decode/shift.cpp +++ b/src/video_core/shader/decode/shift.cpp @@ -17,8 +17,8 @@ u32 ShaderIR::DecodeShift(NodeBlock& bb, u32 pc) { const Instruction instr = {program_code[pc]}; const auto opcode = OpCode::Decode(instr); - const Node op_a = GetRegister(instr.gpr8); - const Node op_b = [&]() { + Node op_a = GetRegister(instr.gpr8); + Node op_b = [&]() { if (instr.is_b_imm) { return Immediate(instr.alu.GetSignedImm20_20()); } else if (instr.is_b_gpr) { @@ -32,16 +32,23 @@ u32 ShaderIR::DecodeShift(NodeBlock& bb, u32 pc) { case OpCode::Id::SHR_C: case OpCode::Id::SHR_R: case OpCode::Id::SHR_IMM: { - const Node value = SignedOperation(OperationCode::IArithmeticShiftRight, - instr.shift.is_signed, PRECISE, op_a, op_b); + if (instr.shr.wrap) { + op_b = Operation(OperationCode::UBitwiseAnd, std::move(op_b), Immediate(0x1f)); + } else { + op_b = Operation(OperationCode::IMax, std::move(op_b), Immediate(0)); + op_b = Operation(OperationCode::IMin, std::move(op_b), Immediate(31)); + } + + Node value = SignedOperation(OperationCode::IArithmeticShiftRight, instr.shift.is_signed, + std::move(op_a), std::move(op_b)); SetInternalFlagsFromInteger(bb, value, instr.generates_cc); - SetRegister(bb, instr.gpr0, value); + SetRegister(bb, instr.gpr0, std::move(value)); break; } case OpCode::Id::SHL_C: case OpCode::Id::SHL_R: case OpCode::Id::SHL_IMM: { - const Node value = Operation(OperationCode::ILogicalShiftLeft, PRECISE, op_a, op_b); + const Node value = Operation(OperationCode::ILogicalShiftLeft, op_a, op_b); SetInternalFlagsFromInteger(bb, value, instr.generates_cc); SetRegister(bb, instr.gpr0, value); break; |