|
150
|
1 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
|
|
|
2 ; RUN: not llc -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s 2>&1 | FileCheck -check-prefix=ERROR %s
|
|
|
3
|
|
|
4 ; ERROR: error: <unknown>:0:0: in function rcp_legacy_f32 void (float addrspace(1)*, float): intrinsic not supported on subtarget
|
|
|
5
|
|
|
6 declare float @llvm.amdgcn.rcp.legacy(float) #0
|
|
|
7
|
|
|
8 ; GCN-LABEL: {{^}}rcp_legacy_f32:
|
|
|
9 ; GCN: v_rcp_legacy_f32_e32 {{v[0-9]+}}, {{s[0-9]+}}
|
|
|
10 define amdgpu_kernel void @rcp_legacy_f32(float addrspace(1)* %out, float %src) #1 {
|
|
|
11 %rcp = call float @llvm.amdgcn.rcp.legacy(float %src) #0
|
|
|
12 store float %rcp, float addrspace(1)* %out, align 4
|
|
|
13 ret void
|
|
|
14 }
|
|
|
15
|
|
|
16 ; TODO: Really these should be constant folded
|
|
|
17 ; GCN-LABEL: {{^}}rcp_legacy_f32_constant_4.0
|
|
|
18 ; GCN: v_rcp_legacy_f32_e32 {{v[0-9]+}}, 4.0
|
|
|
19 define amdgpu_kernel void @rcp_legacy_f32_constant_4.0(float addrspace(1)* %out) #1 {
|
|
|
20 %rcp = call float @llvm.amdgcn.rcp.legacy(float 4.0) #0
|
|
|
21 store float %rcp, float addrspace(1)* %out, align 4
|
|
|
22 ret void
|
|
|
23 }
|
|
|
24
|
|
|
25 ; GCN-LABEL: {{^}}rcp_legacy_f32_constant_100.0
|
|
|
26 ; GCN: v_rcp_legacy_f32_e32 {{v[0-9]+}}, 0x42c80000
|
|
|
27 define amdgpu_kernel void @rcp_legacy_f32_constant_100.0(float addrspace(1)* %out) #1 {
|
|
|
28 %rcp = call float @llvm.amdgcn.rcp.legacy(float 100.0) #0
|
|
|
29 store float %rcp, float addrspace(1)* %out, align 4
|
|
|
30 ret void
|
|
|
31 }
|
|
|
32
|
|
|
33 ; GCN-LABEL: {{^}}rcp_legacy_undef_f32:
|
|
|
34 ; GCN-NOT: v_rcp_legacy_f32
|
|
|
35 define amdgpu_kernel void @rcp_legacy_undef_f32(float addrspace(1)* %out) #1 {
|
|
|
36 %rcp = call float @llvm.amdgcn.rcp.legacy(float undef)
|
|
|
37 store float %rcp, float addrspace(1)* %out, align 4
|
|
|
38 ret void
|
|
|
39 }
|
|
|
40
|
|
|
41 attributes #0 = { nounwind readnone }
|
|
|
42 attributes #1 = { nounwind }
|
