1; RUN: llc -march=amdgcn -mcpu=SI -mattr=-promote-alloca -verify-machineinstrs < %s | FileCheck -check-prefix=SI-ALLOCA -check-prefix=SI %s
2; RUN: llc -march=amdgcn -mcpu=SI -mattr=+promote-alloca -verify-machineinstrs < %s | FileCheck -check-prefix=SI-PROMOTE -check-prefix=SI %s
3; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-promote-alloca -verify-machineinstrs < %s | FileCheck -check-prefix=SI-ALLOCA -check-prefix=SI %s
4; RUN: llc -march=amdgcn -mcpu=tonga -mattr=+promote-alloca -verify-machineinstrs < %s | FileCheck -check-prefix=SI-PROMOTE -check-prefix=SI %s
5
6
7declare void @llvm.AMDGPU.barrier.local() convergent nounwind
8
9; SI-LABEL: {{^}}private_access_f64_alloca:
10
11; SI-ALLOCA: buffer_store_dwordx2
12; SI-ALLOCA: buffer_load_dwordx2
13
14; SI-PROMOTE: ds_write_b64
15; SI-PROMOTE: ds_read_b64
16define void @private_access_f64_alloca(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in, i32 %b) nounwind {
17  %val = load double, double addrspace(1)* %in, align 8
18  %array = alloca double, i32 16, align 8
19  %ptr = getelementptr double, double* %array, i32 %b
20  store double %val, double* %ptr, align 8
21  call void @llvm.AMDGPU.barrier.local() convergent nounwind
22  %result = load double, double* %ptr, align 8
23  store double %result, double addrspace(1)* %out, align 8
24  ret void
25}
26
27; SI-LABEL: {{^}}private_access_v2f64_alloca:
28
29; SI-ALLOCA: buffer_store_dwordx4
30; SI-ALLOCA: buffer_load_dwordx4
31
32; SI-PROMOTE: ds_write_b64
33; SI-PROMOTE: ds_write_b64
34; SI-PROMOTE: ds_read_b64
35; SI-PROMOTE: ds_read_b64
36define void @private_access_v2f64_alloca(<2 x double> addrspace(1)* noalias %out, <2 x double> addrspace(1)* noalias %in, i32 %b) nounwind {
37  %val = load <2 x double>, <2 x double> addrspace(1)* %in, align 16
38  %array = alloca <2 x double>, i32 16, align 16
39  %ptr = getelementptr <2 x double>, <2 x double>* %array, i32 %b
40  store <2 x double> %val, <2 x double>* %ptr, align 16
41  call void @llvm.AMDGPU.barrier.local() convergent nounwind
42  %result = load <2 x double>, <2 x double>* %ptr, align 16
43  store <2 x double> %result, <2 x double> addrspace(1)* %out, align 16
44  ret void
45}
46
47; SI-LABEL: {{^}}private_access_i64_alloca:
48
49; SI-ALLOCA: buffer_store_dwordx2
50; SI-ALLOCA: buffer_load_dwordx2
51
52; SI-PROMOTE: ds_write_b64
53; SI-PROMOTE: ds_read_b64
54define void @private_access_i64_alloca(i64 addrspace(1)* noalias %out, i64 addrspace(1)* noalias %in, i32 %b) nounwind {
55  %val = load i64, i64 addrspace(1)* %in, align 8
56  %array = alloca i64, i32 16, align 8
57  %ptr = getelementptr i64, i64* %array, i32 %b
58  store i64 %val, i64* %ptr, align 8
59  call void @llvm.AMDGPU.barrier.local() convergent nounwind
60  %result = load i64, i64* %ptr, align 8
61  store i64 %result, i64 addrspace(1)* %out, align 8
62  ret void
63}
64
65; SI-LABEL: {{^}}private_access_v2i64_alloca:
66
67; SI-ALLOCA: buffer_store_dwordx4
68; SI-ALLOCA: buffer_load_dwordx4
69
70; SI-PROMOTE: ds_write_b64
71; SI-PROMOTE: ds_write_b64
72; SI-PROMOTE: ds_read_b64
73; SI-PROMOTE: ds_read_b64
74define void @private_access_v2i64_alloca(<2 x i64> addrspace(1)* noalias %out, <2 x i64> addrspace(1)* noalias %in, i32 %b) nounwind {
75  %val = load <2 x i64>, <2 x i64> addrspace(1)* %in, align 16
76  %array = alloca <2 x i64>, i32 16, align 16
77  %ptr = getelementptr <2 x i64>, <2 x i64>* %array, i32 %b
78  store <2 x i64> %val, <2 x i64>* %ptr, align 16
79  call void @llvm.AMDGPU.barrier.local() convergent nounwind
80  %result = load <2 x i64>, <2 x i64>* %ptr, align 16
81  store <2 x i64> %result, <2 x i64> addrspace(1)* %out, align 16
82  ret void
83}
84