diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp --- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp +++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp @@ -8288,22 +8288,24 @@ SDValue Arg = Op.getOperand(0); SDValue TrigVal; - // TODO: Should this propagate fast-math-flags? + // Propagate fast-math flags so that the multiply we introduce can be folded + // if Arg is already the result of a multiply by constant. + auto Flags = Op->getFlags(); SDValue OneOver2Pi = DAG.getConstantFP(0.5 * numbers::inv_pi, DL, VT); if (Subtarget->hasTrigReducedRange()) { - SDValue MulVal = DAG.getNode(ISD::FMUL, DL, VT, Arg, OneOver2Pi); - TrigVal = DAG.getNode(AMDGPUISD::FRACT, DL, VT, MulVal); + SDValue MulVal = DAG.getNode(ISD::FMUL, DL, VT, Arg, OneOver2Pi, Flags); + TrigVal = DAG.getNode(AMDGPUISD::FRACT, DL, VT, MulVal, Flags); } else { - TrigVal = DAG.getNode(ISD::FMUL, DL, VT, Arg, OneOver2Pi); + TrigVal = DAG.getNode(ISD::FMUL, DL, VT, Arg, OneOver2Pi, Flags); } switch (Op.getOpcode()) { case ISD::FCOS: - return DAG.getNode(AMDGPUISD::COS_HW, SDLoc(Op), VT, TrigVal); + return DAG.getNode(AMDGPUISD::COS_HW, SDLoc(Op), VT, TrigVal, Flags); case ISD::FSIN: - return DAG.getNode(AMDGPUISD::SIN_HW, SDLoc(Op), VT, TrigVal); + return DAG.getNode(AMDGPUISD::SIN_HW, SDLoc(Op), VT, TrigVal, Flags); default: llvm_unreachable("Wrong trig opcode"); } diff --git a/llvm/test/CodeGen/AMDGPU/llvm.sin.ll b/llvm/test/CodeGen/AMDGPU/llvm.sin.ll --- a/llvm/test/CodeGen/AMDGPU/llvm.sin.ll +++ b/llvm/test/CodeGen/AMDGPU/llvm.sin.ll @@ -52,7 +52,8 @@ } ; FUNC-LABEL: {{^}}fmf_sin_3x_f32: -; GCN: v_mul_f32 +; GCN-NOT: v_add_f32 +; GCN: 0x3ef47644 ; GCN: v_mul_f32 ; SICIVI: v_fract_f32 ; GFX9-NOT: v_fract_f32 @@ -95,7 +96,8 @@ } ; FUNC-LABEL: {{^}}fmf_sin_2x_f32: -; GCN: v_add_f32 +; GCN-NOT: v_add_f32 +; GCN: 0x3ea2f983 ; GCN: v_mul_f32 ; SICIVI: v_fract_f32 ; GFX9-NOT: v_fract_f32 @@ -137,8 +139,8 @@ } ; FUNC-LABEL: {{^}}fmf_sin_cancel_f32: -; GCN: v_mul_f32 -; GCN: v_mul_f32 +; GCN-NOT: v_add_f32 +; GCN-NOT: v_mul_f32 ; SICIVI: v_fract_f32 ; GFX9-NOT: v_fract_f32 ; GCN: v_sin_f32