1 ; RUN: llc -march=r600 -mcpu=SI -mattr=-promote-alloca -verify-machineinstrs < %s | FileCheck -check-prefix=SI-ALLOCA -check-prefix=SI %s
2 ; RUN: llc -march=r600 -mcpu=SI -mattr=+promote-alloca -verify-machineinstrs < %s | FileCheck -check-prefix=SI-PROMOTE -check-prefix=SI %s
5 declare void @llvm.AMDGPU.barrier.local() noduplicate nounwind
7 ; SI-LABEL: @private_access_f64_alloca:
9 ; SI-ALLOCA: BUFFER_STORE_DWORDX2
10 ; FIXME: We should be able to use BUFFER_LOAD_DWORDX2
11 ; SI-ALLOCA: BUFFER_LOAD_DWORD
12 ; SI-ALLOCA: BUFFER_LOAD_DWORD
14 ; SI-PROMOTE: DS_WRITE_B64
15 ; SI-PROMOTE: DS_READ_B64
16 define void @private_access_f64_alloca(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in, i32 %b) nounwind {
17 %val = load double addrspace(1)* %in, align 8
18 %array = alloca double, i32 16, align 8
19 %ptr = getelementptr double* %array, i32 %b
20 store double %val, double* %ptr, align 8
21 call void @llvm.AMDGPU.barrier.local() noduplicate nounwind
22 %result = load double* %ptr, align 8
23 store double %result, double addrspace(1)* %out, align 8
27 ; SI-LABEL: @private_access_v2f64_alloca:
29 ; SI-ALLOCA: BUFFER_STORE_DWORDX4
30 ; FIXME: We should be able to use BUFFER_LOAD_DWORDX4
31 ; SI-ALLOCA: BUFFER_LOAD_DWORD
32 ; SI-ALLOCA: BUFFER_LOAD_DWORD
33 ; SI-ALLOCA: BUFFER_LOAD_DWORD
34 ; SI-ALLOCA: BUFFER_LOAD_DWORD
36 ; SI-PROMOTE: DS_WRITE_B32
37 ; SI-PROMOTE: DS_WRITE_B32
38 ; SI-PROMOTE: DS_WRITE_B32
39 ; SI-PROMOTE: DS_WRITE_B32
40 ; SI-PROMOTE: DS_READ_B32
41 ; SI-PROMOTE: DS_READ_B32
42 ; SI-PROMOTE: DS_READ_B32
43 ; SI-PROMOTE: DS_READ_B32
44 define void @private_access_v2f64_alloca(<2 x double> addrspace(1)* noalias %out, <2 x double> addrspace(1)* noalias %in, i32 %b) nounwind {
45 %val = load <2 x double> addrspace(1)* %in, align 16
46 %array = alloca <2 x double>, i32 16, align 16
47 %ptr = getelementptr <2 x double>* %array, i32 %b
48 store <2 x double> %val, <2 x double>* %ptr, align 16
49 call void @llvm.AMDGPU.barrier.local() noduplicate nounwind
50 %result = load <2 x double>* %ptr, align 16
51 store <2 x double> %result, <2 x double> addrspace(1)* %out, align 16
55 ; SI-LABEL: @private_access_i64_alloca:
57 ; SI-ALLOCA: BUFFER_STORE_DWORDX2
58 ; FIXME: We should be able to use BUFFER_LOAD_DWORDX2
59 ; SI-ALLOCA: BUFFER_LOAD_DWORD
60 ; SI-ALLOCA: BUFFER_LOAD_DWORD
62 ; SI-PROMOTE: DS_WRITE_B64
63 ; SI-PROMOTE: DS_READ_B64
64 define void @private_access_i64_alloca(i64 addrspace(1)* noalias %out, i64 addrspace(1)* noalias %in, i32 %b) nounwind {
65 %val = load i64 addrspace(1)* %in, align 8
66 %array = alloca i64, i32 16, align 8
67 %ptr = getelementptr i64* %array, i32 %b
68 store i64 %val, i64* %ptr, align 8
69 call void @llvm.AMDGPU.barrier.local() noduplicate nounwind
70 %result = load i64* %ptr, align 8
71 store i64 %result, i64 addrspace(1)* %out, align 8
75 ; SI-LABEL: @private_access_v2i64_alloca:
77 ; SI-ALLOCA: BUFFER_STORE_DWORDX4
78 ; FIXME: We should be able to use BUFFER_LOAD_DWORDX4
79 ; SI-ALLOCA: BUFFER_LOAD_DWORD
80 ; SI-ALLOCA: BUFFER_LOAD_DWORD
81 ; SI-ALLOCA: BUFFER_LOAD_DWORD
82 ; SI-ALLOCA: BUFFER_LOAD_DWORD
84 ; SI-PROMOTE: DS_WRITE_B32
85 ; SI-PROMOTE: DS_WRITE_B32
86 ; SI-PROMOTE: DS_WRITE_B32
87 ; SI-PROMOTE: DS_WRITE_B32
88 ; SI-PROMOTE: DS_READ_B32
89 ; SI-PROMOTE: DS_READ_B32
90 ; SI-PROMOTE: DS_READ_B32
91 ; SI-PROMOTE: DS_READ_B32
92 define void @private_access_v2i64_alloca(<2 x i64> addrspace(1)* noalias %out, <2 x i64> addrspace(1)* noalias %in, i32 %b) nounwind {
93 %val = load <2 x i64> addrspace(1)* %in, align 16
94 %array = alloca <2 x i64>, i32 16, align 16
95 %ptr = getelementptr <2 x i64>* %array, i32 %b
96 store <2 x i64> %val, <2 x i64>* %ptr, align 16
97 call void @llvm.AMDGPU.barrier.local() noduplicate nounwind
98 %result = load <2 x i64>* %ptr, align 16
99 store <2 x i64> %result, <2 x i64> addrspace(1)* %out, align 16