38 lines · plain
1; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=GCN %s2 3; GCN-LABEL: {{^}}fold_mul_neg:4; GCN: load_dword [[V:v[0-9]+]]5; GCN: v_or_b32_e32 [[NEG:v[0-9]]], 0x80000000, [[V]]6; GCN: store_dword [[NEG]]7 8define amdgpu_kernel void @fold_mul_neg(ptr addrspace(1) %arg) {9 %tid = tail call i32 @llvm.amdgcn.workitem.id.x()10 %gep = getelementptr inbounds float, ptr addrspace(1) %arg, i32 %tid11 %v = load float, ptr addrspace(1) %gep, align 412 %cmp = fcmp fast ogt float %v, 0.000000e+0013 %sel = select i1 %cmp, float -1.000000e+00, float 1.000000e+0014 %mul = fmul fast float %v, %sel15 store float %mul, ptr addrspace(1) %gep, align 416 ret void17}18 19; GCN-LABEL: {{^}}fold_mul_abs:20; GCN: load_dword [[V:v[0-9]+]]21; GCN: v_and_b32_e32 [[ABS:v[0-9]]], 0x7fffffff, [[V]]22; GCN: store_dword [[ABS]]23 24define amdgpu_kernel void @fold_mul_abs(ptr addrspace(1) %arg) {25 %tid = tail call i32 @llvm.amdgcn.workitem.id.x()26 %gep = getelementptr inbounds float, ptr addrspace(1) %arg, i32 %tid27 %v = load float, ptr addrspace(1) %gep, align 428 %cmp = fcmp fast olt float %v, 0.000000e+0029 %sel = select i1 %cmp, float -1.000000e+00, float 1.000000e+0030 %mul = fmul fast float %v, %sel31 store float %mul, ptr addrspace(1) %gep, align 432 ret void33}34 35declare i32 @llvm.amdgcn.workitem.id.x() #036 37attributes #0 = { nounwind readnone speculatable }38