83
|
1 ; RUN: llc -march=amdgcn -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
|
|
2 ; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
|
77
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
3 ; RUN: llc -march=r600 -mcpu=cypress -verify-machineinstrs < %s | FileCheck -check-prefix=EG -check-prefix=FUNC %s
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
4
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
5 declare float @llvm.AMDGPU.rsq.f32(float) nounwind readnone
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
6
|
83
|
7 ; FUNC-LABEL: {{^}}rsq_f32:
|
|
8 ; SI: v_rsq_f32_e32 {{v[0-9]+}}, {{s[0-9]+}}
|
77
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
9 ; EG: RECIPSQRT_IEEE
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
10 define void @rsq_f32(float addrspace(1)* %out, float %src) nounwind {
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
11 %rsq = call float @llvm.AMDGPU.rsq.f32(float %src) nounwind readnone
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
12 store float %rsq, float addrspace(1)* %out, align 4
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
13 ret void
|
Kaito Tokumori <e105711@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
14 }
|
83
|
15
|
|
16 ; TODO: Really these should be constant folded
|
|
17 ; FUNC-LABEL: {{^}}rsq_f32_constant_4.0
|
|
18 ; SI: v_rsq_f32_e32 {{v[0-9]+}}, 4.0
|
|
19 ; EG: RECIPSQRT_IEEE
|
|
20 define void @rsq_f32_constant_4.0(float addrspace(1)* %out) nounwind {
|
|
21 %rsq = call float @llvm.AMDGPU.rsq.f32(float 4.0) nounwind readnone
|
|
22 store float %rsq, float addrspace(1)* %out, align 4
|
|
23 ret void
|
|
24 }
|
|
25
|
|
26 ; FUNC-LABEL: {{^}}rsq_f32_constant_100.0
|
|
27 ; SI: v_rsq_f32_e32 {{v[0-9]+}}, 0x42c80000
|
|
28 ; EG: RECIPSQRT_IEEE
|
|
29 define void @rsq_f32_constant_100.0(float addrspace(1)* %out) nounwind {
|
|
30 %rsq = call float @llvm.AMDGPU.rsq.f32(float 100.0) nounwind readnone
|
|
31 store float %rsq, float addrspace(1)* %out, align 4
|
|
32 ret void
|
|
33 }
|