1; RUN: opt < %s -loop-vectorize -S -mcpu=pwr9 -interleave-small-loop-scalar-reduction=true 2>&1 | FileCheck %s
2; RUN: opt < %s -passes='loop-vectorize' -S -mcpu=pwr9 -interleave-small-loop-scalar-reduction=true 2>&1 | FileCheck %s
3
4; CHECK-LABEL: vector.body
5; CHECK: load double, double*
6; CHECK-NEXT: load double, double*
7; CHECK-NEXT: load double, double*
8; CHECK-NEXT: load double, double*
9
10; CHECK: fmul fast double
11; CHECK-NEXT: fmul fast double
12; CHECK-NEXT: fmul fast double
13; CHECK-NEXT: fmul fast double
14
15; CHECK: fadd fast double
16; CHECK-NEXT: fadd fast double
17; CHECK-NEXT: fadd fast double
18; CHECK-NEXT: fadd fast double
19
20target datalayout = "e-m:e-i64:64-n32:64"
21target triple = "powerpc64le-unknown-linux-gnu"
22
23define dso_local void @test(i32*** %arg, double** %arg1) align 2 {
24bb:
25  %tpm15 = load i32**, i32*** %arg, align 8
26  %tpm19 = load double*, double** %arg1, align 8
27  br label %bb22
28bb22:                                             ; preds = %bb33, %bb
29  %tpm26 = add i64 0, 1
30  %tpm10 = alloca i32, align 8
31  %tpm27 = getelementptr inbounds i32, i32* %tpm10, i64 %tpm26
32  %tpm28 = getelementptr inbounds i32*, i32** %tpm15, i64 0
33  %tpm29 = load i32*, i32** %tpm28, align 8
34  %tpm17 = alloca double, align 8
35  %tpm32 = getelementptr inbounds double, double* %tpm17, i64 %tpm26
36  br label %bb40
37bb33:                                             ; preds = %bb40
38  %tpm35 = getelementptr inbounds double, double* %tpm19, i64 0
39  %tpm37 = fsub fast double 0.000000e+00, %tpm50
40  store double %tpm37, double* %tpm35, align 8
41  br label %bb22
42bb40:                                             ; preds = %bb40, %bb22
43  %tpm41 = phi i32* [ %tpm51, %bb40 ], [ %tpm27, %bb22 ]
44  %tpm42 = phi double* [ %tpm52, %bb40 ], [ %tpm32, %bb22 ]
45  %tpm43 = phi double [ %tpm50, %bb40 ], [ 0.000000e+00, %bb22 ]
46  %tpm44 = load double, double* %tpm42, align 8
47  %tpm45 = load i32, i32* %tpm41, align 4
48  %tpm46 = zext i32 %tpm45 to i64
49  %tpm47 = getelementptr inbounds double, double* %tpm19, i64 %tpm46
50  %tpm48 = load double, double* %tpm47, align 8
51  %tpm49 = fmul fast double %tpm48, %tpm44
52  %tpm50 = fadd fast double %tpm49, %tpm43
53  %tpm51 = getelementptr inbounds i32, i32* %tpm41, i64 1
54  %tpm52 = getelementptr inbounds double, double* %tpm42, i64 1
55  %tpm53 = icmp eq i32* %tpm51, %tpm29
56  br i1 %tpm53, label %bb33, label %bb40
57}
58