1; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=SI -check-prefix=GCN -check-prefix=FUNC %s 2; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=VI -check-prefix=GCN -check-prefix=FUNC %s 3; RUN: llc -march=r600 -mcpu=redwood < %s | FileCheck -enable-var-scope -check-prefix=R600 -check-prefix=FUNC %s 4 5; FUNC-LABEL: {{^}}s_fneg_f32: 6; R600: -PV 7 8; GCN: s_load_dword [[VAL:s[0-9]+]] 9; GCN: s_xor_b32 [[NEG_VAL:s[0-9]+]], [[VAL]], 0x80000000 10; GCN: v_mov_b32_e32 v{{[0-9]+}}, [[NEG_VAL]] 11define amdgpu_kernel void @s_fneg_f32(float addrspace(1)* %out, float %in) { 12 %fneg = fsub float -0.000000e+00, %in 13 store float %fneg, float addrspace(1)* %out 14 ret void 15} 16 17; FUNC-LABEL: {{^}}s_fneg_v2f32: 18; R600: -PV 19; R600: -PV 20 21; GCN: s_xor_b32 {{s[0-9]+}}, {{s[0-9]+}}, 0x80000000 22; GCN: s_xor_b32 {{s[0-9]+}}, {{s[0-9]+}}, 0x80000000 23define amdgpu_kernel void @s_fneg_v2f32(<2 x float> addrspace(1)* nocapture %out, <2 x float> %in) { 24 %fneg = fsub <2 x float> <float -0.000000e+00, float -0.000000e+00>, %in 25 store <2 x float> %fneg, <2 x float> addrspace(1)* %out 26 ret void 27} 28 29; FUNC-LABEL: {{^}}s_fneg_v4f32: 30; R600: -PV 31; R600: -T 32; R600: -PV 33; R600: -PV 34 35; GCN: s_xor_b32 {{s[0-9]+}}, {{s[0-9]+}}, 0x80000000 36; GCN: s_xor_b32 {{s[0-9]+}}, {{s[0-9]+}}, 0x80000000 37; GCN: s_xor_b32 {{s[0-9]+}}, {{s[0-9]+}}, 0x80000000 38; GCN: s_xor_b32 {{s[0-9]+}}, {{s[0-9]+}}, 0x80000000 39define amdgpu_kernel void @s_fneg_v4f32(<4 x float> addrspace(1)* nocapture %out, <4 x float> %in) { 40 %fneg = fsub <4 x float> <float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00>, %in 41 store <4 x float> %fneg, <4 x float> addrspace(1)* %out 42 ret void 43} 44 45; DAGCombiner will transform: 46; (fneg (f32 bitcast (i32 a))) => (f32 bitcast (xor (i32 a), 0x80000000)) 47; unless the target returns true for isNegFree() 48 49; FUNC-LABEL: {{^}}fsub0_f32: 50 51; GCN: v_sub_f32_e64 v{{[0-9]}}, 0, s{{[0-9]+$}} 52 53; R600-NOT: XOR 54; R600: -KC0[2].Z 55define amdgpu_kernel void @fsub0_f32(float addrspace(1)* %out, i32 %in) { 56 %bc = bitcast i32 %in to float 57 %fsub = fsub float 0.0, %bc 58 store float %fsub, float addrspace(1)* %out 59 ret void 60} 61; FUNC-LABEL: {{^}}fneg_free_f32: 62; SI: s_load_dword [[NEG_VALUE:s[0-9]+]], s[{{[0-9]+:[0-9]+}}], 0xb 63; VI: s_load_dword [[NEG_VALUE:s[0-9]+]], s[{{[0-9]+:[0-9]+}}], 0x2c 64 65; GCN: s_xor_b32 [[RES:s[0-9]+]], [[NEG_VALUE]], 0x80000000 66; GCN: v_mov_b32_e32 [[V_RES:v[0-9]+]], [[RES]] 67; GCN: buffer_store_dword [[V_RES]] 68 69; R600-NOT: XOR 70; R600: -PV.W 71define amdgpu_kernel void @fneg_free_f32(float addrspace(1)* %out, i32 %in) { 72 %bc = bitcast i32 %in to float 73 %fsub = fsub float -0.0, %bc 74 store float %fsub, float addrspace(1)* %out 75 ret void 76} 77 78; FUNC-LABEL: {{^}}fneg_fold_f32: 79; SI: s_load_dword [[NEG_VALUE:s[0-9]+]], s[{{[0-9]+:[0-9]+}}], 0xb 80; VI: s_load_dword [[NEG_VALUE:s[0-9]+]], s[{{[0-9]+:[0-9]+}}], 0x2c 81; GCN-NOT: xor 82; GCN: v_mul_f32_e64 v{{[0-9]+}}, -[[NEG_VALUE]], [[NEG_VALUE]] 83define amdgpu_kernel void @fneg_fold_f32(float addrspace(1)* %out, float %in) { 84 %fsub = fsub float -0.0, %in 85 %fmul = fmul float %fsub, %in 86 store float %fmul, float addrspace(1)* %out 87 ret void 88} 89 90; Make sure we turn some integer operations back into fabs 91; FUNC-LABEL: {{^}}bitpreserve_fneg_f32: 92; GCN: v_mul_f32_e64 v{{[0-9]+}}, s{{[0-9]+}}, -4.0 93define amdgpu_kernel void @bitpreserve_fneg_f32(float addrspace(1)* %out, float %in) { 94 %in.bc = bitcast float %in to i32 95 %int.abs = xor i32 %in.bc, 2147483648 96 %bc = bitcast i32 %int.abs to float 97 %fadd = fmul float %bc, 4.0 98 store float %fadd, float addrspace(1)* %out 99 ret void 100} 101