1; RUN: opt < %s -O2 -force-vector-interleave=2 -force-vector-width=4 -debug-only=loop-vectorize -stats -S 2>&1 | FileCheck %s
2; REQUIRES: asserts
3
4; Loop from "rotated"
5; CHECK: LV: Loop hints: force=enabled
6; Loop from "nonrotated"
7; CHECK: LV: Loop hints: force=enabled
8; No more loops in the module
9; CHECK-NOT: LV: Loop hints: force=
10; In total only 1 loop should be rotated.
11; CHECK: 1 loop-rotate
12
13target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
14target triple = "x86_64-unknown-linux-gnu"
15
16; See http://reviews.llvm.org/D3348 for details.
17
18;
19; Test #1
20;
21; Ensure that "llvm.loop.vectorize.enable" metadata was not lost prior to LoopVectorize pass.
22; In past LoopRotate was clearing that metadata.
23;
24; The source C code is:
25; void rotated(float *a, int size)
26; {
27;   int t = 0;
28;   #pragma omp simd
29;   for (int i = 0; i < size; ++i) {
30;     a[i] = a[i-5] * a[i+2];
31;     ++t;
32;   }
33;}
34
35define void @rotated(float* nocapture %a, i64 %size) {
36entry:
37  %cmp1 = icmp sgt i64 %size, 0
38  br i1 %cmp1, label %for.header, label %for.end
39
40for.header:
41  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ]
42  %cmp2 = icmp sgt i64 %indvars.iv, %size
43  br i1 %cmp2, label %for.end, label %for.body
44
45for.body:
46
47  %0 = add nsw i64 %indvars.iv, -5
48  %arrayidx = getelementptr inbounds float, float* %a, i64 %0
49  %1 = load float, float* %arrayidx, align 4, !llvm.mem.parallel_loop_access !1
50  %2 = add nsw i64 %indvars.iv, 2
51  %arrayidx2 = getelementptr inbounds float, float* %a, i64 %2
52  %3 = load float, float* %arrayidx2, align 4, !llvm.mem.parallel_loop_access !1
53  %mul = fmul float %1, %3
54  %arrayidx4 = getelementptr inbounds float, float* %a, i64 %indvars.iv
55  store float %mul, float* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !1
56
57  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
58  br label %for.header, !llvm.loop !1
59
60for.end:
61  ret void
62}
63
64!1 = !{!1, !2}
65!2 = !{!"llvm.loop.vectorize.enable", i1 true}
66
67;
68; Test #2
69;
70; Ensure that "llvm.loop.vectorize.enable" metadata was not lost even
71; if loop was not rotated (see http://reviews.llvm.org/D3348#comment-4).
72;
73define i32 @nonrotated(i32 %a) {
74entry:
75  br label %loop_cond
76loop_cond:
77  %indx = phi i32 [ 1, %entry ], [ %inc, %loop_inc ]
78  %cmp = icmp ne i32 %indx, %a
79  br i1 %cmp, label %return, label %loop_inc
80loop_inc:
81  %inc = add i32 %indx, 1
82  br label %loop_cond, !llvm.loop !3
83return:
84  ret i32 0
85}
86
87!3 = !{!3, !4}
88!4 = !{!"llvm.loop.vectorize.enable", i1 true}
89