1; RUN: llc < %s -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -start-before=amdgpu-isel -stop-after=amdgpu-isel -verify-machineinstrs | FileCheck %s --check-prefixes=CHECK
2
3; This caused failure in infinite cycle in Selection DAG (combine) due to missing insert_subvector.
4;
5; CHECK-LABEL: name: test1
6; CHECK: GLOBAL_LOAD_DWORDX4
7; CHECK: GLOBAL_LOAD_DWORDX4
8; CHECK: GLOBAL_STORE_DWORDX4
9define protected amdgpu_kernel void @test1() local_unnamed_addr !kernel_arg_addr_space !0 !kernel_arg_access_qual !1 !kernel_arg_type !2 !kernel_arg_base_type !2 !kernel_arg_type_qual !3 !kernel_arg_name !4 {
10entry:
11  %tmp = load <3 x i64>, <3 x i64> addrspace(4)* undef, align 16, !invariant.load !5
12  %srcA.load2 = extractelement <3 x i64> %tmp, i32 0
13  %tmp1 = inttoptr i64 %srcA.load2 to double addrspace(1)*
14  %tmp2 = getelementptr inbounds double, double addrspace(1)* %tmp1, i64 undef
15  %tmp3 = bitcast double addrspace(1)* %tmp2 to <3 x double> addrspace(1)*
16  %tmp4 = load <3 x double>, <3 x double> addrspace(1)* %tmp3, align 8, !tbaa !6
17  %tmp5 = extractelement <3 x double> %tmp4, i32 1
18  %tmp6 = insertelement <3 x double> undef, double %tmp5, i32 1
19  %tmp7 = insertelement <3 x double> %tmp6, double undef, i32 2
20  %tmp8 = load <3 x double>, <3 x double> addrspace(1)* undef, align 8, !tbaa !6
21  %tmp9 = extractelement <3 x double> %tmp8, i32 2
22  %tmp10 = insertelement <3 x double> undef, double %tmp9, i32 2
23  %tmp11 = fcmp olt <3 x double> %tmp10, %tmp7
24  %tmp12 = select <3 x i1> %tmp11, <3 x double> zeroinitializer, <3 x double> <double 1.000000e+00, double 1.000000e+00, double 1.000000e+00>
25  %tmp13 = extractelement <3 x double> %tmp12, i64 1
26  %tmp14 = insertelement <2 x double> undef, double %tmp13, i32 1
27  store <2 x double> %tmp14, <2 x double> addrspace(1)* undef, align 8, !tbaa !6
28  ret void
29}
30
31; This caused failure in Selection DAG due to lack of insert_subvector implementation.
32;
33; CHECK-LABEL: name: test2
34; CHECK: GLOBAL_LOAD_DWORDX2
35; CHECK: GLOBAL_LOAD_DWORDX2
36; CHECK: GLOBAL_STORE_DWORDX2
37define protected amdgpu_kernel void @test2() local_unnamed_addr !kernel_arg_addr_space !0 !kernel_arg_access_qual !1 !kernel_arg_type !2 !kernel_arg_base_type !2 !kernel_arg_type_qual !3 !kernel_arg_name !4 {
38entry:
39  %tmp = load <3 x i64>, <3 x i64> addrspace(4)* undef, align 16, !invariant.load !5
40  %srcA.load2 = extractelement <3 x i64> %tmp, i32 0
41  %tmp1 = inttoptr i64 %srcA.load2 to double addrspace(1)*
42  %tmp2 = getelementptr inbounds double, double addrspace(1)* %tmp1, i64 undef
43  %tmp3 = bitcast double addrspace(1)* %tmp2 to <3 x double> addrspace(1)*
44  %tmp4 = load <3 x double>, <3 x double> addrspace(1)* %tmp3, align 8, !tbaa !6
45  %tmp5 = extractelement <3 x double> %tmp4, i32 1
46  %tmp6 = insertelement <3 x double> undef, double %tmp5, i32 1
47  %tmp7 = insertelement <3 x double> %tmp6, double undef, i32 2
48  %tmp8 = load <3 x double>, <3 x double> addrspace(1)* undef, align 8, !tbaa !6
49  %tmp9 = extractelement <3 x double> %tmp8, i32 1
50  %tmp10 = insertelement <3 x double> undef, double %tmp9, i32 1
51  %tmp11 = insertelement <3 x double> %tmp10, double undef, i32 2
52  %tmp12 = fcmp olt <3 x double> %tmp11, %tmp7
53  %tmp13 = select <3 x i1> %tmp12, <3 x double> zeroinitializer, <3 x double> <double 1.000000e+00, double 1.000000e+00, double 1.000000e+00>
54  %tmp14 = extractelement <3 x double> %tmp13, i64 2
55  store double %tmp14, double addrspace(1)* undef, align 8, !tbaa !6
56  ret void
57}
58
59!0 = !{i32 1, i32 1, i32 1}
60!1 = !{!"none", !"none", !"none"}
61!2 = !{!"double*", !"double*", !"double*"}
62!3 = !{!"", !"", !""}
63!4 = !{!"srcA", !"srcB", !"dst"}
64!5 = !{}
65!6 = !{!7, !7, i64 0}
66!7 = !{!"double", !8, i64 0}
67!8 = !{!"omnipotent char", !9, i64 0}
68!9 = !{!"Simple C/C++ TBAA"}
69