diff --git a/External/FEXCore/Source/Interface/Core/JIT/x86_64/JIT.cpp b/External/FEXCore/Source/Interface/Core/JIT/x86_64/JIT.cpp index f3291131e..1c96b2636 100644 --- a/External/FEXCore/Source/Interface/Core/JIT/x86_64/JIT.cpp +++ b/External/FEXCore/Source/Interface/Core/JIT/x86_64/JIT.cpp @@ -2847,7 +2847,9 @@ void *JITCore::CompileCode([[maybe_unused]] FEXCore::IR::IRListView const // Scalar switch (Op->Header.ElementSize) { case 4: { - vrcpss(GetDst(Node), GetSrc(Op->Header.Args[0].ID()), GetSrc(Op->Header.Args[0].ID())); + mov(eax, 0x3f800000); // 1.0f + vmovd(xmm15, eax); + vdivss(GetDst(Node), xmm15, GetSrc(Op->Header.Args[0].ID())); break; } default: LogMan::Msg::A("Unknown Element Size: %d", Op->Header.ElementSize); break; @@ -2857,7 +2859,10 @@ void *JITCore::CompileCode([[maybe_unused]] FEXCore::IR::IRListView const // Vector switch (Op->Header.ElementSize) { case 4: { - vrcpps(GetDst(Node), GetSrc(Op->Header.Args[0].ID())); + mov(eax, 0x3f800000); // 1.0f + vmovd(xmm15, eax); + pshufd(xmm15, xmm15, 0); + vdivps(GetDst(Node), xmm15, GetSrc(Op->Header.Args[0].ID())); break; } default: LogMan::Msg::A("Unknown Element Size: %d", Op->Header.ElementSize); break; diff --git a/External/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp b/External/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp index afddaf6c2..ac0049bc8 100644 --- a/External/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp +++ b/External/FEXCore/Source/Interface/Core/OpcodeDispatcher.cpp @@ -5630,6 +5630,7 @@ void InstallOpcodeHandlers() { {0x50, 1, &OpDispatchBuilder::MOVMSKOp<4>}, {0x51, 1, &OpDispatchBuilder::VectorUnaryOp}, {0x52, 1, &OpDispatchBuilder::VectorUnaryOp}, + {0x53, 1, &OpDispatchBuilder::VectorUnaryOp}, {0x54, 1, &OpDispatchBuilder::VectorALUOp}, {0x55, 1, &OpDispatchBuilder::ANDNOp}, {0x56, 1, &OpDispatchBuilder::VectorALUOp}, @@ -5826,6 +5827,7 @@ void InstallOpcodeHandlers() { {0x2D, 1, &OpDispatchBuilder::CVTFPR_To_GPR<4, true>}, {0x51, 1, &OpDispatchBuilder::VectorUnaryOp}, {0x52, 1, &OpDispatchBuilder::VectorUnaryOp}, + {0x53, 1, &OpDispatchBuilder::VectorUnaryOp}, {0x58, 1, &OpDispatchBuilder::VectorScalarALUOp}, {0x59, 1, &OpDispatchBuilder::VectorScalarALUOp}, {0x5A, 1, &OpDispatchBuilder::Scalar_CVT_Float_To_Float<8, 4>}, diff --git a/unittests/ASM/REP/F3_53.asm b/unittests/ASM/REP/F3_53.asm new file mode 100644 index 000000000..ee21277db --- /dev/null +++ b/unittests/ASM/REP/F3_53.asm @@ -0,0 +1,31 @@ +%ifdef CONFIG +{ + "RegData": { + "XMM0": ["0x3f8000003f800000", "0x3f8000003f800000"], + "XMM1": ["0x408000003e800000", "0x4080000040800000"] + }, + "MemoryRegions": { + "0x100000000": "4096" + } +} +%endif + +mov rdx, 0xe0000000 + +mov rax, 0x3f8000003f800000 ; 1.0 +mov [rdx + 8 * 0], rax +mov rax, 0x3f8000003f800000 +mov [rdx + 8 * 1], rax + +mov rax, 0x4080000040800000 ; 4.0 +mov [rdx + 8 * 2], rax +mov rax, 0x4080000040800000 +mov [rdx + 8 * 3], rax + +movapd xmm0, [rdx + 8 * 0] +movapd xmm1, [rdx + 8 * 2] + +rcpss xmm0, xmm0 +rcpss xmm1, [rdx + 8 * 2] + +hlt diff --git a/unittests/ASM/TwoByte/0F_53.asm b/unittests/ASM/TwoByte/0F_53.asm new file mode 100644 index 000000000..168ca896c --- /dev/null +++ b/unittests/ASM/TwoByte/0F_53.asm @@ -0,0 +1,30 @@ +%ifdef CONFIG +{ + "RegData": { + "XMM0": ["0x3f8000003f800000", "0x3f8000003f800000"], + "XMM1": ["0x3e8000003e800000", "0x3e8000003e800000"] + }, + "MemoryRegions": { + "0x100000000": "4096" + } +} +%endif + +mov rdx, 0xe0000000 + +mov rax, 0x3f8000003f800000 ; 1.0 +mov [rdx + 8 * 0], rax +mov rax, 0x3f8000003f800000 +mov [rdx + 8 * 1], rax + +mov rax, 0x4080000040800000 ; 4.0 +mov [rdx + 8 * 2], rax +mov rax, 0x4080000040800000 +mov [rdx + 8 * 3], rax + +movapd xmm0, [rdx + 8 * 0] + +rcpps xmm0, xmm0 +rcpps xmm1, [rdx + 8 * 2] + +hlt