1; RUN: opt < %s -scalarrepl -S | FileCheck %s
2; PR3290
3target datalayout = "E-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64"
4
5;; Store of integer to whole alloca struct.
6define i32 @test1(i64 %V) nounwind {
7; CHECK: test1
8; CHECK-NOT: alloca
9	%X = alloca {{i32, i32}}
10	%Y = bitcast {{i32,i32}}* %X to i64*
11	store i64 %V, i64* %Y
12
13	%A = getelementptr {{i32,i32}}* %X, i32 0, i32 0, i32 0
14	%B = getelementptr {{i32,i32}}* %X, i32 0, i32 0, i32 1
15	%a = load i32* %A
16	%b = load i32* %B
17	%c = add i32 %a, %b
18	ret i32 %c
19}
20
21;; Store of integer to whole struct/array alloca.
22define float @test2(i128 %V) nounwind {
23; CHECK: test2
24; CHECK-NOT: alloca
25	%X = alloca {[4 x float]}
26	%Y = bitcast {[4 x float]}* %X to i128*
27	store i128 %V, i128* %Y
28
29	%A = getelementptr {[4 x float]}* %X, i32 0, i32 0, i32 0
30	%B = getelementptr {[4 x float]}* %X, i32 0, i32 0, i32 3
31	%a = load float* %A
32	%b = load float* %B
33	%c = fadd float %a, %b
34	ret float %c
35}
36
37;; Load of whole alloca struct as integer
38define i64 @test3(i32 %a, i32 %b) nounwind {
39; CHECK: test3
40; CHECK-NOT: alloca
41	%X = alloca {{i32, i32}}
42
43	%A = getelementptr {{i32,i32}}* %X, i32 0, i32 0, i32 0
44	%B = getelementptr {{i32,i32}}* %X, i32 0, i32 0, i32 1
45        store i32 %a, i32* %A
46        store i32 %b, i32* %B
47
48	%Y = bitcast {{i32,i32}}* %X to i64*
49        %Z = load i64* %Y
50	ret i64 %Z
51}
52
53;; load of integer from whole struct/array alloca.
54define i128 @test4(float %a, float %b) nounwind {
55; CHECK: test4
56; CHECK-NOT: alloca
57	%X = alloca {[4 x float]}
58	%A = getelementptr {[4 x float]}* %X, i32 0, i32 0, i32 0
59	%B = getelementptr {[4 x float]}* %X, i32 0, i32 0, i32 3
60	store float %a, float* %A
61	store float %b, float* %B
62
63      	%Y = bitcast {[4 x float]}* %X to i128*
64	%V = load i128* %Y
65	ret i128 %V
66}
67
68;; If the elements of a struct or array alloca contain padding, SROA can still
69;; split up the alloca as long as there is no padding between the elements.
70%padded = type { i16, i8 }
71define void @test5([4 x %padded]* %p, [4 x %padded]* %q) {
72entry:
73; CHECK: test5
74; CHECK-NOT: i128
75  %var = alloca [4 x %padded], align 4
76  %vari8 = bitcast [4 x %padded]* %var to i8*
77  %pi8 = bitcast [4 x %padded]* %p to i8*
78  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %vari8, i8* %pi8, i32 16, i32 4, i1 false)
79  %qi8 = bitcast [4 x %padded]* %q to i8*
80  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %qi8, i8* %vari8, i32 16, i32 4, i1 false)
81  ret void
82}
83
84;; Check that an array alloca can be split up when it is also accessed with
85;; a load or store as a homogeneous structure with the same element type and
86;; number of elements as the array.
87%homogeneous = type { <8 x i16>, <8 x i16>, <8 x i16> }
88%wrapped_array = type { [3 x <8 x i16>] }
89define void @test6(i8* %p, %wrapped_array* %arr) {
90entry:
91; CHECK: test6
92; CHECK: store <8 x i16>
93; CHECK: store <8 x i16>
94; CHECK: store <8 x i16>
95  %var = alloca %wrapped_array, align 16
96  %res = call %homogeneous @test6callee(i8* %p)
97  %varcast = bitcast %wrapped_array* %var to %homogeneous*
98  store %homogeneous %res, %homogeneous* %varcast
99  %tmp1 = bitcast %wrapped_array* %arr to i8*
100  %tmp2 = bitcast %wrapped_array* %var to i8*
101  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %tmp1, i8* %tmp2, i32 48, i32 16, i1 false)
102  ret void
103}
104
105declare %homogeneous @test6callee(i8* nocapture) nounwind
106
107declare void @llvm.memcpy.p0i8.p0i8.i32(i8* nocapture, i8* nocapture, i32, i32, i1) nounwind
108