1 ; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mattr=-fp32-denormals -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -check-prefix=SI-UNSAFE -check-prefix=SI %s
2 ; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mattr=-fp32-denormals -verify-machineinstrs < %s | FileCheck -check-prefix=SI-SAFE -check-prefix=SI %s
4 declare i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
5 declare float @llvm.sqrt.f32(float) nounwind readnone
6 declare double @llvm.sqrt.f64(double) nounwind readnone
8 ; SI-LABEL: {{^}}rsq_f32:
11 define amdgpu_kernel void @rsq_f32(float addrspace(1)* noalias %out, float addrspace(1)* noalias %in) nounwind {
12 %val = load float, float addrspace(1)* %in, align 4
13 %sqrt = call float @llvm.sqrt.f32(float %val) nounwind readnone
14 %div = fdiv float 1.0, %sqrt
15 store float %div, float addrspace(1)* %out, align 4
19 ; SI-LABEL: {{^}}rsq_f64:
20 ; SI-UNSAFE: v_rsq_f64_e32
21 ; SI-SAFE: v_sqrt_f64_e32
23 define amdgpu_kernel void @rsq_f64(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) nounwind {
24 %val = load double, double addrspace(1)* %in, align 4
25 %sqrt = call double @llvm.sqrt.f64(double %val) nounwind readnone
26 %div = fdiv double 1.0, %sqrt
27 store double %div, double addrspace(1)* %out, align 4
31 ; SI-LABEL: {{^}}rsq_f32_sgpr:
32 ; SI: v_rsq_f32_e32 {{v[0-9]+}}, {{s[0-9]+}}
34 define amdgpu_kernel void @rsq_f32_sgpr(float addrspace(1)* noalias %out, float %val) nounwind {
35 %sqrt = call float @llvm.sqrt.f32(float %val) nounwind readnone
36 %div = fdiv float 1.0, %sqrt
37 store float %div, float addrspace(1)* %out, align 4
41 ; Recognize that this is rsqrt(a) * rcp(b) * c,
42 ; not 1 / ( 1 / sqrt(a)) * rcp(b) * c.
44 ; SI-LABEL: @rsqrt_fmul
45 ; SI-DAG: buffer_load_dword [[A:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
46 ; SI-DAG: buffer_load_dword [[B:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
47 ; SI-DAG: buffer_load_dword [[C:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:8
49 ; SI-UNSAFE-DAG: v_rsq_f32_e32 [[RSQA:v[0-9]+]], [[A]]
50 ; SI-UNSAFE-DAG: v_rcp_f32_e32 [[RCPB:v[0-9]+]], [[B]]
51 ; SI-UNSAFE-DAG: v_mul_f32_e32 [[TMP:v[0-9]+]], [[RCPB]], [[RSQA]]
52 ; SI-UNSAFE: v_mul_f32_e32 [[RESULT:v[0-9]+]], [[C]], [[TMP]]
53 ; SI-UNSAFE: buffer_store_dword [[RESULT]]
55 ; SI-SAFE-NOT: v_rsq_f32
58 define amdgpu_kernel void @rsqrt_fmul(float addrspace(1)* %out, float addrspace(1)* %in) {
59 %tid = call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
60 %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
61 %gep.0 = getelementptr float, float addrspace(1)* %in, i32 %tid
62 %gep.1 = getelementptr float, float addrspace(1)* %gep.0, i32 1
63 %gep.2 = getelementptr float, float addrspace(1)* %gep.0, i32 2
65 %a = load volatile float, float addrspace(1)* %gep.0
66 %b = load volatile float, float addrspace(1)* %gep.1
67 %c = load volatile float, float addrspace(1)* %gep.2
69 %x = call float @llvm.sqrt.f32(float %a)
70 %y = fmul float %x, %b
71 %z = fdiv float %c, %y
72 store float %z, float addrspace(1)* %out.gep
76 ; SI-LABEL: {{^}}neg_rsq_f32:
77 ; SI-SAFE: v_sqrt_f32_e32 [[SQRT:v[0-9]+]], v{{[0-9]+}}
78 ; SI-SAFE: v_rcp_f32_e64 [[RSQ:v[0-9]+]], -[[SQRT]]
79 ; SI-SAFE: buffer_store_dword [[RSQ]]
81 ; SI-UNSAFE: v_rsq_f32_e32 [[RSQ:v[0-9]+]], v{{[0-9]+}}
82 ; SI-UNSAFE: v_xor_b32_e32 [[NEG_RSQ:v[0-9]+]], 0x80000000, [[RSQ]]
83 ; SI-UNSAFE: buffer_store_dword [[NEG_RSQ]]
84 define amdgpu_kernel void @neg_rsq_f32(float addrspace(1)* noalias %out, float addrspace(1)* noalias %in) nounwind {
85 %val = load float, float addrspace(1)* %in, align 4
86 %sqrt = call float @llvm.sqrt.f32(float %val)
87 %div = fdiv float -1.0, %sqrt
88 store float %div, float addrspace(1)* %out, align 4
92 ; SI-LABEL: {{^}}neg_rsq_f64:
93 ; SI-SAFE: v_sqrt_f64_e32
94 ; SI-SAFE: v_div_scale_f64
96 ; SI-UNSAFE: v_sqrt_f64_e32 [[SQRT:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}
97 ; SI-UNSAFE: v_rcp_f64_e64 [[RCP:v\[[0-9]+:[0-9]+\]]], -[[SQRT]]
98 ; SI-UNSAFE: buffer_store_dwordx2 [[RCP]]
99 define amdgpu_kernel void @neg_rsq_f64(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) nounwind {
100 %val = load double, double addrspace(1)* %in, align 4
101 %sqrt = call double @llvm.sqrt.f64(double %val)
102 %div = fdiv double -1.0, %sqrt
103 store double %div, double addrspace(1)* %out, align 4
107 ; SI-LABEL: {{^}}neg_rsq_neg_f32:
108 ; SI-SAFE: v_sqrt_f32_e64 [[SQRT:v[0-9]+]], -v{{[0-9]+}}
109 ; SI-SAFE: v_rcp_f32_e64 [[RSQ:v[0-9]+]], -[[SQRT]]
110 ; SI-SAFE: buffer_store_dword [[RSQ]]
112 ; SI-UNSAFE: v_rsq_f32_e64 [[RSQ:v[0-9]+]], -v{{[0-9]+}}
113 ; SI-UNSAFE: v_xor_b32_e32 [[NEG_RSQ:v[0-9]+]], 0x80000000, [[RSQ]]
114 ; SI-UNSAFE: buffer_store_dword [[NEG_RSQ]]
115 define amdgpu_kernel void @neg_rsq_neg_f32(float addrspace(1)* noalias %out, float addrspace(1)* noalias %in) nounwind {
116 %val = load float, float addrspace(1)* %in, align 4
117 %val.fneg = fsub float -0.0, %val
118 %sqrt = call float @llvm.sqrt.f32(float %val.fneg)
119 %div = fdiv float -1.0, %sqrt
120 store float %div, float addrspace(1)* %out, align 4
124 ; SI-LABEL: {{^}}neg_rsq_neg_f64:
125 ; SI-SAFE: v_sqrt_f64_e64 v{{\[[0-9]+:[0-9]+\]}}, -v{{\[[0-9]+:[0-9]+\]}}
126 ; SI-SAFE: v_div_scale_f64
128 ; SI-UNSAFE: v_sqrt_f64_e64 [[SQRT:v\[[0-9]+:[0-9]+\]]], -v{{\[[0-9]+:[0-9]+\]}}
129 ; SI-UNSAFE: v_rcp_f64_e64 [[RCP:v\[[0-9]+:[0-9]+\]]], -[[SQRT]]
130 ; SI-UNSAFE: buffer_store_dwordx2 [[RCP]]
131 define amdgpu_kernel void @neg_rsq_neg_f64(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) nounwind {
132 %val = load double, double addrspace(1)* %in, align 4
133 %val.fneg = fsub double -0.0, %val
134 %sqrt = call double @llvm.sqrt.f64(double %val.fneg)
135 %div = fdiv double -1.0, %sqrt
136 store double %div, double addrspace(1)* %out, align 4