1; RUN: llc -march=amdgcn -mcpu=SI -mattr=-fp32-denormals -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -check-prefix=SI-UNSAFE -check-prefix=SI %s
2; RUN: llc -march=amdgcn -mcpu=SI -mattr=-fp32-denormals -verify-machineinstrs < %s | FileCheck -check-prefix=SI-SAFE -check-prefix=SI %s
3
4declare i32 @llvm.r600.read.tidig.x() nounwind readnone
5declare float @llvm.sqrt.f32(float) nounwind readnone
6declare double @llvm.sqrt.f64(double) nounwind readnone
7
8; SI-LABEL: {{^}}rsq_f32:
9; SI: v_rsq_f32_e32
10; SI: s_endpgm
11define void @rsq_f32(float addrspace(1)* noalias %out, float addrspace(1)* noalias %in) nounwind {
12  %val = load 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
16  ret void
17}
18
19; SI-LABEL: {{^}}rsq_f64:
20; SI-UNSAFE: v_rsq_f64_e32
21; SI-SAFE: v_sqrt_f64_e32
22; SI: s_endpgm
23define void @rsq_f64(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) nounwind {
24  %val = load 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
28  ret void
29}
30
31; SI-LABEL: {{^}}rsq_f32_sgpr:
32; SI: v_rsq_f32_e32 {{v[0-9]+}}, {{s[0-9]+}}
33; SI: s_endpgm
34define 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
38  ret void
39}
40
41; Recognize that this is rsqrt(a) * rcp(b) * c,
42; not 1 / ( 1 / sqrt(a)) * rcp(b) * c.
43
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
48
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]+]], [[TMP]], [[C]]
53; SI-UNSAFE: buffer_store_dword [[RESULT]]
54
55; SI-SAFE-NOT: v_rsq_f32
56
57; SI: s_endpgm
58define void @rsqrt_fmul(float addrspace(1)* %out, float addrspace(1)* %in) {
59  %tid = call i32 @llvm.r600.read.tidig.x() nounwind readnone
60  %out.gep = getelementptr float addrspace(1)* %out, i32 %tid
61  %gep.0 = getelementptr float addrspace(1)* %in, i32 %tid
62  %gep.1 = getelementptr float addrspace(1)* %gep.0, i32 1
63  %gep.2 = getelementptr float addrspace(1)* %gep.0, i32 2
64
65  %a = load float addrspace(1)* %gep.0
66  %b = load float addrspace(1)* %gep.1
67  %c = load float addrspace(1)* %gep.2
68
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
73  ret void
74}
75