diff options
author | bunnei <bunneidev@gmail.com> | 2015-09-07 21:59:24 +0200 |
---|---|---|
committer | bunnei <bunneidev@gmail.com> | 2015-09-07 21:59:24 +0200 |
commit | 47221ab1e62d7bca6a0e430401afb1e0b32cdce7 (patch) | |
tree | a2427dc7c5ba6c94fd8ac5562c0fa27e13c862df | |
parent | Merge pull request #1052 from yuriks/vertex-disasm (diff) | |
parent | Shader JIT: Use SCALE constant from emitter (diff) | |
download | yuzu-47221ab1e62d7bca6a0e430401afb1e0b32cdce7.tar yuzu-47221ab1e62d7bca6a0e430401afb1e0b32cdce7.tar.gz yuzu-47221ab1e62d7bca6a0e430401afb1e0b32cdce7.tar.bz2 yuzu-47221ab1e62d7bca6a0e430401afb1e0b32cdce7.tar.lz yuzu-47221ab1e62d7bca6a0e430401afb1e0b32cdce7.tar.xz yuzu-47221ab1e62d7bca6a0e430401afb1e0b32cdce7.tar.zst yuzu-47221ab1e62d7bca6a0e430401afb1e0b32cdce7.zip |
-rw-r--r-- | src/video_core/shader/shader.h | 12 | ||||
-rw-r--r-- | src/video_core/shader/shader_jit_x64.cpp | 26 |
2 files changed, 22 insertions, 16 deletions
diff --git a/src/video_core/shader/shader.h b/src/video_core/shader/shader.h index bac51ddd8..1c6fa592c 100644 --- a/src/video_core/shader/shader.h +++ b/src/video_core/shader/shader.h @@ -289,13 +289,13 @@ struct UnitState { DebugData<Debug> debug; - static int InputOffset(const SourceRegister& reg) { + static size_t InputOffset(const SourceRegister& reg) { switch (reg.GetRegisterType()) { case RegisterType::Input: - return (int)offsetof(UnitState::Registers, input) + reg.GetIndex()*sizeof(Math::Vec4<float24>); + return offsetof(UnitState::Registers, input) + reg.GetIndex()*sizeof(Math::Vec4<float24>); case RegisterType::Temporary: - return (int)offsetof(UnitState::Registers, temporary) + reg.GetIndex()*sizeof(Math::Vec4<float24>); + return offsetof(UnitState::Registers, temporary) + reg.GetIndex()*sizeof(Math::Vec4<float24>); default: UNREACHABLE(); @@ -303,13 +303,13 @@ struct UnitState { } } - static int OutputOffset(const DestRegister& reg) { + static size_t OutputOffset(const DestRegister& reg) { switch (reg.GetRegisterType()) { case RegisterType::Output: - return (int)offsetof(UnitState::Registers, output) + reg.GetIndex()*sizeof(Math::Vec4<float24>); + return offsetof(UnitState::Registers, output) + reg.GetIndex()*sizeof(Math::Vec4<float24>); case RegisterType::Temporary: - return (int)offsetof(UnitState::Registers, temporary) + reg.GetIndex()*sizeof(Math::Vec4<float24>); + return offsetof(UnitState::Registers, temporary) + reg.GetIndex()*sizeof(Math::Vec4<float24>); default: UNREACHABLE(); diff --git a/src/video_core/shader/shader_jit_x64.cpp b/src/video_core/shader/shader_jit_x64.cpp index b9a0b19e3..0c02976ac 100644 --- a/src/video_core/shader/shader_jit_x64.cpp +++ b/src/video_core/shader/shader_jit_x64.cpp @@ -144,7 +144,7 @@ static const u8 NO_DEST_REG_MASK = 0xf; */ void JitCompiler::Compile_SwizzleSrc(Instruction instr, unsigned src_num, SourceRegister src_reg, X64Reg dest) { X64Reg src_ptr; - int src_offset; + size_t src_offset; if (src_reg.GetRegisterType() == RegisterType::FloatUniform) { src_ptr = UNIFORMS; @@ -154,6 +154,9 @@ void JitCompiler::Compile_SwizzleSrc(Instruction instr, unsigned src_num, Source src_offset = UnitState<false>::InputOffset(src_reg); } + int src_offset_disp = (int)src_offset; + ASSERT_MSG(src_offset == src_offset_disp, "Source register offset too large for int type"); + unsigned operand_desc_id; if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MAD || instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MADI) { @@ -163,7 +166,7 @@ void JitCompiler::Compile_SwizzleSrc(Instruction instr, unsigned src_num, Source operand_desc_id = instr.mad.operand_desc_id; // Load the source - MOVAPS(dest, MDisp(src_ptr, src_offset)); + MOVAPS(dest, MDisp(src_ptr, src_offset_disp)); } else { operand_desc_id = instr.common.operand_desc_id; @@ -173,13 +176,13 @@ void JitCompiler::Compile_SwizzleSrc(Instruction instr, unsigned src_num, Source if (src_num == offset_src && instr.common.address_register_index != 0) { switch (instr.common.address_register_index) { case 1: // address offset 1 - MOVAPS(dest, MComplex(src_ptr, ADDROFFS_REG_0, 1, src_offset)); + MOVAPS(dest, MComplex(src_ptr, ADDROFFS_REG_0, SCALE_1, src_offset_disp)); break; case 2: // address offset 2 - MOVAPS(dest, MComplex(src_ptr, ADDROFFS_REG_1, 1, src_offset)); + MOVAPS(dest, MComplex(src_ptr, ADDROFFS_REG_1, SCALE_1, src_offset_disp)); break; - case 3: // adddress offet 3 - MOVAPS(dest, MComplex(src_ptr, LOOPCOUNT_REG, 1, src_offset)); + case 3: // address offset 3 + MOVAPS(dest, MComplex(src_ptr, LOOPCOUNT_REG, SCALE_1, src_offset_disp)); break; default: UNREACHABLE(); @@ -187,7 +190,7 @@ void JitCompiler::Compile_SwizzleSrc(Instruction instr, unsigned src_num, Source } } else { // Load the source - MOVAPS(dest, MDisp(src_ptr, src_offset)); + MOVAPS(dest, MDisp(src_ptr, src_offset_disp)); } } @@ -224,14 +227,17 @@ void JitCompiler::Compile_DestEnable(Instruction instr,X64Reg src) { SwizzlePattern swiz = { g_state.vs.swizzle_data[operand_desc_id] }; + int dest_offset_disp = (int)UnitState<false>::OutputOffset(dest); + ASSERT_MSG(dest_offset_disp == UnitState<false>::OutputOffset(dest), "Destinaton offset too large for int type"); + // If all components are enabled, write the result to the destination register if (swiz.dest_mask == NO_DEST_REG_MASK) { // Store dest back to memory - MOVAPS(MDisp(REGISTERS, UnitState<false>::OutputOffset(dest)), src); + MOVAPS(MDisp(REGISTERS, dest_offset_disp), src); } else { // Not all components are enabled, so mask the result when storing to the destination register... - MOVAPS(SCRATCH, MDisp(REGISTERS, UnitState<false>::OutputOffset(dest))); + MOVAPS(SCRATCH, MDisp(REGISTERS, dest_offset_disp)); if (Common::GetCPUCaps().sse4_1) { u8 mask = ((swiz.dest_mask & 1) << 3) | ((swiz.dest_mask & 8) >> 3) | ((swiz.dest_mask & 2) << 1) | ((swiz.dest_mask & 4) >> 1); @@ -250,7 +256,7 @@ void JitCompiler::Compile_DestEnable(Instruction instr,X64Reg src) { } // Store dest back to memory - MOVAPS(MDisp(REGISTERS, UnitState<false>::OutputOffset(dest)), SCRATCH); + MOVAPS(MDisp(REGISTERS, dest_offset_disp), SCRATCH); } } |