1 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s 2 3 declare float @llvm.amdgcn.rcp.f32(float) #0 4 declare double @llvm.amdgcn.rcp.f64(double) #0 5 6 declare double @llvm.sqrt.f64(double) #0 7 declare float @llvm.sqrt.f32(float) #0 8 9 ; FUNC-LABEL: {{^}}rcp_undef_f32: 10 ; SI-NOT: v_rcp_f32 11 define void @rcp_undef_f32(float addrspace(1)* %out) #1 { 12 %rcp = call float @llvm.amdgcn.rcp.f32(float undef) 13 store float %rcp, float addrspace(1)* %out, align 4 14 ret void 15 } 16 17 ; FUNC-LABEL: {{^}}safe_no_fp32_denormals_rcp_f32: 18 ; SI: v_rcp_f32_e32 [[RESULT:v[0-9]+]], s{{[0-9]+}} 19 ; SI-NOT: [[RESULT]] 20 ; SI: buffer_store_dword [[RESULT]] 21 define void @safe_no_fp32_denormals_rcp_f32(float addrspace(1)* %out, float %src) #1 { 22 %rcp = fdiv float 1.0, %src 23 store float %rcp, float addrspace(1)* %out, align 4 24 ret void 25 } 26 27 ; FUNC-LABEL: {{^}}safe_f32_denormals_rcp_pat_f32: 28 ; SI: v_rcp_f32_e32 [[RESULT:v[0-9]+]], s{{[0-9]+}} 29 ; SI-NOT: [[RESULT]] 30 ; SI: buffer_store_dword [[RESULT]] 31 define void @safe_f32_denormals_rcp_pat_f32(float addrspace(1)* %out, float %src) #4 { 32 %rcp = fdiv float 1.0, %src 33 store float %rcp, float addrspace(1)* %out, align 4 34 ret void 35 } 36 37 ; FUNC-LABEL: {{^}}unsafe_f32_denormals_rcp_pat_f32: 38 ; SI: v_div_scale_f32 39 define void @unsafe_f32_denormals_rcp_pat_f32(float addrspace(1)* %out, float %src) #3 { 40 %rcp = fdiv float 1.0, %src 41 store float %rcp, float addrspace(1)* %out, align 4 42 ret void 43 } 44 45 ; FUNC-LABEL: {{^}}safe_rsq_rcp_pat_f32: 46 ; SI: v_sqrt_f32_e32 47 ; SI: v_rcp_f32_e32 48 define void @safe_rsq_rcp_pat_f32(float addrspace(1)* %out, float %src) #1 { 49 %sqrt = call float @llvm.sqrt.f32(float %src) 50 %rcp = call float @llvm.amdgcn.rcp.f32(float %sqrt) 51 store float %rcp, float addrspace(1)* %out, align 4 52 ret void 53 } 54 55 ; FUNC-LABEL: {{^}}unsafe_rsq_rcp_pat_f32: 56 ; SI: v_rsq_f32_e32 57 define void @unsafe_rsq_rcp_pat_f32(float addrspace(1)* %out, float %src) #2 { 58 %sqrt = call float @llvm.sqrt.f32(float %src) 59 %rcp = call float @llvm.amdgcn.rcp.f32(float %sqrt) 60 store float %rcp, float addrspace(1)* %out, align 4 61 ret void 62 } 63 64 ; FUNC-LABEL: {{^}}rcp_f64: 65 ; SI: v_rcp_f64_e32 [[RESULT:v\[[0-9]+:[0-9]+\]]], s{{\[[0-9]+:[0-9]+\]}} 66 ; SI-NOT: [[RESULT]] 67 ; SI: buffer_store_dwordx2 [[RESULT]] 68 define void @rcp_f64(double addrspace(1)* %out, double %src) #1 { 69 %rcp = call double @llvm.amdgcn.rcp.f64(double %src) 70 store double %rcp, double addrspace(1)* %out, align 8 71 ret void 72 } 73 74 ; FUNC-LABEL: {{^}}unsafe_rcp_f64: 75 ; SI: v_rcp_f64_e32 [[RESULT:v\[[0-9]+:[0-9]+\]]], s{{\[[0-9]+:[0-9]+\]}} 76 ; SI-NOT: [[RESULT]] 77 ; SI: buffer_store_dwordx2 [[RESULT]] 78 define void @unsafe_rcp_f64(double addrspace(1)* %out, double %src) #2 { 79 %rcp = call double @llvm.amdgcn.rcp.f64(double %src) 80 store double %rcp, double addrspace(1)* %out, align 8 81 ret void 82 } 83 84 ; FUNC-LABEL: {{^}}rcp_pat_f64: 85 ; SI: v_div_scale_f64 86 define void @rcp_pat_f64(double addrspace(1)* %out, double %src) #1 { 87 %rcp = fdiv double 1.0, %src 88 store double %rcp, double addrspace(1)* %out, align 8 89 ret void 90 } 91 92 ; FUNC-LABEL: {{^}}unsafe_rcp_pat_f64: 93 ; SI: v_rcp_f64_e32 [[RESULT:v\[[0-9]+:[0-9]+\]]], s{{\[[0-9]+:[0-9]+\]}} 94 ; SI-NOT: [[RESULT]] 95 ; SI: buffer_store_dwordx2 [[RESULT]] 96 define void @unsafe_rcp_pat_f64(double addrspace(1)* %out, double %src) #2 { 97 %rcp = fdiv double 1.0, %src 98 store double %rcp, double addrspace(1)* %out, align 8 99 ret void 100 } 101 102 ; FUNC-LABEL: {{^}}safe_rsq_rcp_pat_f64: 103 ; SI-NOT: v_rsq_f64_e32 104 ; SI: v_sqrt_f64 105 ; SI: v_rcp_f64 106 define void @safe_rsq_rcp_pat_f64(double addrspace(1)* %out, double %src) #1 { 107 %sqrt = call double @llvm.sqrt.f64(double %src) 108 %rcp = call double @llvm.amdgcn.rcp.f64(double %sqrt) 109 store double %rcp, double addrspace(1)* %out, align 8 110 ret void 111 } 112 113 ; FUNC-LABEL: {{^}}unsafe_rsq_rcp_pat_f64: 114 ; SI: v_rsq_f64_e32 [[RESULT:v\[[0-9]+:[0-9]+\]]], s{{\[[0-9]+:[0-9]+\]}} 115 ; SI-NOT: [[RESULT]] 116 ; SI: buffer_store_dwordx2 [[RESULT]] 117 define void @unsafe_rsq_rcp_pat_f64(double addrspace(1)* %out, double %src) #2 { 118 %sqrt = call double @llvm.sqrt.f64(double %src) 119 %rcp = call double @llvm.amdgcn.rcp.f64(double %sqrt) 120 store double %rcp, double addrspace(1)* %out, align 8 121 ret void 122 } 123 124 attributes #0 = { nounwind readnone } 125 attributes #1 = { nounwind "unsafe-fp-math"="false" "target-features"="-fp32-denormals" } 126 attributes #2 = { nounwind "unsafe-fp-math"="true" "target-features"="-fp32-denormals" } 127 attributes #3 = { nounwind "unsafe-fp-math"="false" "target-features"="+fp32-denormals" } 128 attributes #4 = { nounwind "unsafe-fp-math"="true" "target-features"="+fp32-denormals" } 129
