xref: /llvm-project/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.rsq.ll (revision 9e9907f1cfa424366fba58d9520f9305b537cec9)
1; RUN: llc -mtriple=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
2; RUN: llc -mtriple=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
3
4declare float @llvm.amdgcn.rsq.f32(float) #0
5declare double @llvm.amdgcn.rsq.f64(double) #0
6
7; FUNC-LABEL: {{^}}rsq_f32:
8; SI: v_rsq_f32_e32 {{v[0-9]+}}, {{s[0-9]+}}
9define amdgpu_kernel void @rsq_f32(ptr addrspace(1) %out, float %src) #1 {
10  %rsq = call float @llvm.amdgcn.rsq.f32(float %src) #0
11  store float %rsq, ptr addrspace(1) %out, align 4
12  ret void
13}
14
15; TODO: Really these should be constant folded
16; FUNC-LABEL: {{^}}rsq_f32_constant_4.0
17; SI: v_rsq_f32_e32 {{v[0-9]+}}, 4.0
18define amdgpu_kernel void @rsq_f32_constant_4.0(ptr addrspace(1) %out) #1 {
19  %rsq = call float @llvm.amdgcn.rsq.f32(float 4.0) #0
20  store float %rsq, ptr addrspace(1) %out, align 4
21  ret void
22}
23
24; FUNC-LABEL: {{^}}rsq_f32_constant_100.0
25; SI: v_rsq_f32_e32 {{v[0-9]+}}, 0x42c80000
26define amdgpu_kernel void @rsq_f32_constant_100.0(ptr addrspace(1) %out) #1 {
27  %rsq = call float @llvm.amdgcn.rsq.f32(float 100.0) #0
28  store float %rsq, ptr addrspace(1) %out, align 4
29  ret void
30}
31
32; FUNC-LABEL: {{^}}rsq_f64:
33; SI: v_rsq_f64_e32 {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}
34define amdgpu_kernel void @rsq_f64(ptr addrspace(1) %out, double %src) #1 {
35  %rsq = call double @llvm.amdgcn.rsq.f64(double %src) #0
36  store double %rsq, ptr addrspace(1) %out, align 4
37  ret void
38}
39
40; TODO: Really these should be constant folded
41; FUNC-LABEL: {{^}}rsq_f64_constant_4.0
42; SI: v_rsq_f64_e32 {{v\[[0-9]+:[0-9]+\]}}, 4.0
43define amdgpu_kernel void @rsq_f64_constant_4.0(ptr addrspace(1) %out) #1 {
44  %rsq = call double @llvm.amdgcn.rsq.f64(double 4.0) #0
45  store double %rsq, ptr addrspace(1) %out, align 4
46  ret void
47}
48
49; FUNC-LABEL: {{^}}rsq_f64_constant_100.0
50; SI: v_rsq_f64_e32 {{v\[[0-9]+:[0-9]+\]}}, 0x40590000
51define amdgpu_kernel void @rsq_f64_constant_100.0(ptr addrspace(1) %out) #1 {
52  %rsq = call double @llvm.amdgcn.rsq.f64(double 100.0) #0
53  store double %rsq, ptr addrspace(1) %out, align 4
54  ret void
55}
56
57; FUNC-LABEL: {{^}}rsq_undef_f32:
58; SI-NOT: v_rsq_f32
59define amdgpu_kernel void @rsq_undef_f32(ptr addrspace(1) %out) #1 {
60  %rsq = call float @llvm.amdgcn.rsq.f32(float undef)
61  store float %rsq, ptr addrspace(1) %out, align 4
62  ret void
63}
64
65attributes #0 = { nounwind readnone }
66attributes #1 = { nounwind }
67