1; RUN: opt -mattr=+avx512f --loop-vectorize -S < %s | llc -mattr=+avx512f | FileCheck %s
2; RUN: opt -mattr=+avx512vl,+prefer-256-bit --loop-vectorize -S < %s | llc -mattr=+avx512f | FileCheck %s --check-prefix=CHECK-PREFER-AVX256
3
4target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
5target triple = "x86_64-apple-macosx10.9.0"
6
7; Verify that we generate 512-bit wide vectors for a basic integer memset
8; loop.
9
10; CHECK-LABEL: f:
11; CHECK: vmovdqu64 %zmm{{.}},
12; CHECK-NOT: %ymm
13; CHECK: epilog
14; CHECK: %ymm
15
16; Verify that we don't generate 512-bit wide vectors when subtarget feature says not to
17
18; CHECK-PREFER-AVX256-LABEL: f:
19; CHECK-PREFER-AVX256: vmovdqu %ymm{{.}},
20; CHECK-PREFER-AVX256-NOT: %zmm
21
22define void @f(i32* %a, i32 %n) {
23entry:
24  %cmp4 = icmp sgt i32 %n, 0
25  br i1 %cmp4, label %for.body.preheader, label %for.end
26
27for.body.preheader:                               ; preds = %entry
28  br label %for.body
29
30for.body:                                         ; preds = %for.body.preheader, %for.body
31  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %for.body.preheader ]
32  %arrayidx = getelementptr inbounds i32, i32* %a, i64 %indvars.iv
33  store i32 %n, i32* %arrayidx, align 4
34  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
35  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
36  %exitcond = icmp eq i32 %lftr.wideiv, %n
37  br i1 %exitcond, label %for.end.loopexit, label %for.body
38
39for.end.loopexit:                                 ; preds = %for.body
40  br label %for.end
41
42for.end:                                          ; preds = %for.end.loopexit, %entry
43  ret void
44}
45
46; Verify that the "prefer-vector-width=256" attribute prevents the use of 512-bit
47; vectors
48
49; CHECK-LABEL: g:
50; CHECK: vmovdqu %ymm{{.}},
51; CHECK-NOT: %zmm
52
53; CHECK-PREFER-AVX256-LABEL: g:
54; CHECK-PREFER-AVX256: vmovdqu %ymm{{.}},
55; CHECK-PREFER-AVX256-NOT: %zmm
56
57define void @g(i32* %a, i32 %n) "prefer-vector-width"="256" {
58entry:
59  %cmp4 = icmp sgt i32 %n, 0
60  br i1 %cmp4, label %for.body.preheader, label %for.end
61
62for.body.preheader:                               ; preds = %entry
63  br label %for.body
64
65for.body:                                         ; preds = %for.body.preheader, %for.body
66  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %for.body.preheader ]
67  %arrayidx = getelementptr inbounds i32, i32* %a, i64 %indvars.iv
68  store i32 %n, i32* %arrayidx, align 4
69  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
70  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
71  %exitcond = icmp eq i32 %lftr.wideiv, %n
72  br i1 %exitcond, label %for.end.loopexit, label %for.body
73
74for.end.loopexit:                                 ; preds = %for.body
75  br label %for.end
76
77for.end:                                          ; preds = %for.end.loopexit, %entry
78  ret void
79}
80
81; Verify that the "prefer-vector-width=512" attribute override the subtarget
82; vectors
83
84; CHECK-LABEL: h:
85; CHECK: vmovdqu64 %zmm{{.}},
86; CHECK-NOT: %ymm
87; CHECK: epilog
88; CHECK: %ymm
89
90; CHECK-PREFER-AVX256-LABEL: h:
91; CHECK-PREFER-AVX256: vmovdqu64 %zmm{{.}},
92; CHECK-PREFER-AVX256-NOT: %ymm
93; CHECK-PREFER-AVX256: epilog
94; CHECK-PREFER-AVX256: %ymm
95
96define void @h(i32* %a, i32 %n) "prefer-vector-width"="512" {
97entry:
98  %cmp4 = icmp sgt i32 %n, 0
99  br i1 %cmp4, label %for.body.preheader, label %for.end
100
101for.body.preheader:                               ; preds = %entry
102  br label %for.body
103
104for.body:                                         ; preds = %for.body.preheader, %for.body
105  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %for.body.preheader ]
106  %arrayidx = getelementptr inbounds i32, i32* %a, i64 %indvars.iv
107  store i32 %n, i32* %arrayidx, align 4
108  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
109  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
110  %exitcond = icmp eq i32 %lftr.wideiv, %n
111  br i1 %exitcond, label %for.end.loopexit, label %for.body
112
113for.end.loopexit:                                 ; preds = %for.body
114  br label %for.end
115
116for.end:                                          ; preds = %for.end.loopexit, %entry
117  ret void
118}
119