From f533e6b0edf84d7a498fe10f45ed34f53eac7425 Mon Sep 17 00:00:00 2001 From: Matt Arsenault Date: Wed, 15 Aug 2018 21:46:27 +0000 Subject: [PATCH] AMDGPU: Fold fneg into fmed3 llvm-svn: 339821 --- llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp | 11 +++ llvm/test/CodeGen/AMDGPU/llvm.amdgcn.fmed3.ll | 86 ++++++++++++++++++- 2 files changed, 93 insertions(+), 4 deletions(-) diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp index 57cd19943ba6..530162668102 100644 --- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp +++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp @@ -562,6 +562,7 @@ static bool fnegFoldsIntoOp(unsigned Opc) { case AMDGPUISD::FMUL_LEGACY: case AMDGPUISD::FMIN_LEGACY: case AMDGPUISD::FMAX_LEGACY: + case AMDGPUISD::FMED3: return true; default: return false; @@ -3608,6 +3609,16 @@ SDValue AMDGPUTargetLowering::performFNegCombine(SDNode *N, DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res)); return Res; } + case AMDGPUISD::FMED3: { + SDValue Ops[3]; + for (unsigned I = 0; I < 3; ++I) + Ops[I] = DAG.getNode(ISD::FNEG, SL, VT, N0->getOperand(I), N0->getFlags()); + + SDValue Res = DAG.getNode(AMDGPUISD::FMED3, SL, VT, Ops, N0->getFlags()); + if (!N0.hasOneUse()) + DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res)); + return Res; + } case ISD::FP_EXTEND: case ISD::FTRUNC: case ISD::FRINT: diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.fmed3.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.fmed3.ll index 230e625ad45b..05b074bfe2d4 100644 --- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.fmed3.ll +++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.fmed3.ll @@ -4,8 +4,8 @@ ; GCN-LABEL: {{^}}test_fmed3: ; GCN: v_med3_f32 v{{[0-9]+}}, s{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} define amdgpu_kernel void @test_fmed3(float addrspace(1)* %out, float %src0, float %src1, float %src2) #1 { - %mad = call float @llvm.amdgcn.fmed3.f32(float %src0, float %src1, float %src2) - store float %mad, float addrspace(1)* %out + %med3 = call float @llvm.amdgcn.fmed3.f32(float %src0, float %src1, float %src2) + store float %med3, float addrspace(1)* %out ret void } @@ -16,8 +16,86 @@ define amdgpu_kernel void @test_fmed3_srcmods(float addrspace(1)* %out, float %s %src1.fabs = call float @llvm.fabs.f32(float %src1) %src2.fabs = call float @llvm.fabs.f32(float %src2) %src2.fneg.fabs = fsub float -0.0, %src2.fabs - %mad = call float @llvm.amdgcn.fmed3.f32(float %src0.fneg, float %src1.fabs, float %src2.fneg.fabs) - store float %mad, float addrspace(1)* %out + %med3 = call float @llvm.amdgcn.fmed3.f32(float %src0.fneg, float %src1.fabs, float %src2.fneg.fabs) + store float %med3, float addrspace(1)* %out + ret void +} + +; GCN-LABEL: {{^}}test_fneg_fmed3: +; GCN: v_med3_f32 v{{[0-9]+}}, -s{{[0-9]+}}, -v{{[0-9]+}}, -v{{[0-9]+}} +define amdgpu_kernel void @test_fneg_fmed3(float addrspace(1)* %out, float %src0, float %src1, float %src2) #1 { + %med3 = call float @llvm.amdgcn.fmed3.f32(float %src0, float %src1, float %src2) + %neg.med3 = fsub float -0.0, %med3 + store float %neg.med3, float addrspace(1)* %out + ret void +} + +; GCN-LABEL: {{^}}test_fneg_fmed3_multi_use: +; GCN: v_med3_f32 [[MED3:v[0-9]+]], -s{{[0-9]+}}, -v{{[0-9]+}}, -v{{[0-9]+}} +; GCN: v_mul_f32_e32 v{{[0-9]+}}, -4.0, [[MED3]] +define amdgpu_kernel void @test_fneg_fmed3_multi_use(float addrspace(1)* %out, float %src0, float %src1, float %src2) #1 { + %med3 = call float @llvm.amdgcn.fmed3.f32(float %src0, float %src1, float %src2) + %neg.med3 = fsub float -0.0, %med3 + %med3.user = fmul float %med3, 4.0 + store volatile float %med3.user, float addrspace(1)* %out + store volatile float %neg.med3, float addrspace(1)* %out + ret void +} + +; GCN-LABEL: {{^}}test_fabs_fmed3: +; GCN: v_med3_f32 [[MED3:v[0-9]+]], s{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} +; GCN: v_and_b32_e32 v{{[0-9]+}}, 0x7fffffff, [[MED3]] +define amdgpu_kernel void @test_fabs_fmed3(float addrspace(1)* %out, float %src0, float %src1, float %src2) #1 { + %med3 = call float @llvm.amdgcn.fmed3.f32(float %src0, float %src1, float %src2) + %fabs.med3 = call float @llvm.fabs.f32(float %med3) + store float %fabs.med3, float addrspace(1)* %out + ret void +} + +; GCN-LABEL: {{^}}test_fneg_fmed3_rr_0: +; GCN: v_bfrev_b32_e32 [[NEG0:v[0-9]+]], 1 +; GCN: v_med3_f32 v{{[0-9]+}}, -s{{[0-9]+}}, -v{{[0-9]+}}, [[NEG0]] +define amdgpu_kernel void @test_fneg_fmed3_rr_0(float addrspace(1)* %out, float %src0, float %src1) #1 { + %med3 = call float @llvm.amdgcn.fmed3.f32(float %src0, float %src1, float 0.0) + %neg.med3 = fsub float -0.0, %med3 + store float %neg.med3, float addrspace(1)* %out + ret void +} + +; FIXME: Worse off from folding this +; GCN-LABEL: {{^}}test_fneg_fmed3_rr_0_foldable_user: +; GCN: v_bfrev_b32_e32 [[NEG0:v[0-9]+]], 1 +; GCN: v_med3_f32 [[MED3:v[0-9]+]], -s{{[0-9]+}}, -v{{[0-9]+}}, [[NEG0]] +; GCN: v_mul_f32_e32 v{{[0-9]+}}, s{{[0-9]+}}, [[MED3]] +define amdgpu_kernel void @test_fneg_fmed3_rr_0_foldable_user(float addrspace(1)* %out, float %src0, float %src1, float %mul.arg) #1 { + %med3 = call float @llvm.amdgcn.fmed3.f32(float %src0, float %src1, float 0.0) + %neg.med3 = fsub float -0.0, %med3 + %mul = fmul float %neg.med3, %mul.arg + store float %mul, float addrspace(1)* %out + ret void +} + +; GCN-LABEL: {{^}}test_fneg_fmed3_r_inv2pi_0: +; GCN-DAG: v_bfrev_b32_e32 [[NEG0:v[0-9]+]], 1 +; GCN-DAG: v_mov_b32_e32 [[NEG_INV:v[0-9]+]], 0xbe22f983 +; GCN: v_med3_f32 v{{[0-9]+}}, -s{{[0-9]+}}, [[NEG_INV]], [[NEG0]] +define amdgpu_kernel void @test_fneg_fmed3_r_inv2pi_0(float addrspace(1)* %out, float %src0) #1 { + %med3 = call float @llvm.amdgcn.fmed3.f32(float %src0, float 0x3FC45F3060000000, float 0.0) + %neg.med3 = fsub float -0.0, %med3 + store float %neg.med3, float addrspace(1)* %out + ret void +} + +; GCN-LABEL: {{^}}test_fneg_fmed3_r_inv2pi_0_foldable_user: +; GCN-DAG: v_bfrev_b32_e32 [[NEG0:v[0-9]+]], 1 +; GCN-DAG: v_mov_b32_e32 [[NEG_INV:v[0-9]+]], 0xbe22f983 +; GCN: v_med3_f32 [[MED3:v[0-9]+]], -s{{[0-9]+}}, [[NEG_INV]], [[NEG0]] +; GCN: v_mul_f32_e32 v{{[0-9]+}}, s{{[0-9]+}}, [[MED3]] +define amdgpu_kernel void @test_fneg_fmed3_r_inv2pi_0_foldable_user(float addrspace(1)* %out, float %src0, float %mul.arg) #1 { + %med3 = call float @llvm.amdgcn.fmed3.f32(float %src0, float 0x3FC45F3060000000, float 0.0) + %neg.med3 = fsub float -0.0, %med3 + %mul = fmul float %neg.med3, %mul.arg + store float %mul, float addrspace(1)* %out ret void } -- 2.34.1