236
|
1 ; RUN: opt %loadPolly -polly-opt-isl -polly-codegen -polly-vectorizer=polly -polly-prevect-width=8 -S -dce < %s | FileCheck %s
|
|
2 ;
|
|
3 ; void foo(long n, float A[restrict][n], float B[restrict][n],
|
|
4 ; float C[restrict][n], float D[restrict][n]) {
|
|
5 ; for (long i = 0; i < 8; i++)
|
|
6 ; for (long j = 0; j < 8; j++)
|
|
7 ; A[i][j] += B[i][0] + C[i][2 * j] + D[j][0];
|
|
8 ; }
|
|
9 ;
|
|
10
|
|
11 ; CHECK: shufflevector
|
|
12 ; CHECK: insertelement
|
|
13 ; CHECK: insertelement
|
|
14 ; CHECK: insertelement
|
|
15 ; CHECK: insertelement
|
|
16 ; CHECK: insertelement
|
|
17 ; CHECK: insertelement
|
|
18 ; CHECK: insertelement
|
|
19 ; CHECK: insertelement
|
|
20 ; CHECK: store <8 x float>
|
|
21
|
|
22 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
|
|
23
|
|
24 define void @foo(i64 %n, float* noalias %A, float* noalias %B, float* noalias %C, float* noalias %D) {
|
|
25 bb:
|
|
26 br label %bb3
|
|
27
|
|
28 bb3: ; preds = %bb25, %bb
|
|
29 %i.0 = phi i64 [ 0, %bb ], [ %tmp26, %bb25 ]
|
|
30 %exitcond2 = icmp ne i64 %i.0, 8
|
|
31 br i1 %exitcond2, label %bb4, label %bb27
|
|
32
|
|
33 bb4: ; preds = %bb3
|
|
34 br label %bb5
|
|
35
|
|
36 bb5: ; preds = %bb22, %bb4
|
|
37 %j.0 = phi i64 [ 0, %bb4 ], [ %tmp23, %bb22 ]
|
|
38 %exitcond = icmp ne i64 %j.0, 8
|
|
39 br i1 %exitcond, label %bb6, label %bb24
|
|
40
|
|
41 bb6: ; preds = %bb5
|
|
42 %tmp = mul nsw i64 %i.0, %n
|
|
43 %tmp7 = getelementptr inbounds float, float* %B, i64 %tmp
|
|
44 %tmp8 = load float, float* %tmp7, align 4
|
|
45 %tmp9 = shl nsw i64 %j.0, 1
|
|
46 %tmp10 = mul nsw i64 %i.0, %n
|
|
47 %.sum = add i64 %tmp10, %tmp9
|
|
48 %tmp11 = getelementptr inbounds float, float* %C, i64 %.sum
|
|
49 %tmp12 = load float, float* %tmp11, align 4
|
|
50 %tmp13 = fadd float %tmp8, %tmp12
|
|
51 %tmp14 = mul nsw i64 %j.0, %n
|
|
52 %tmp15 = getelementptr inbounds float, float* %D, i64 %tmp14
|
|
53 %tmp16 = load float, float* %tmp15, align 4
|
|
54 %tmp17 = fadd float %tmp13, %tmp16
|
|
55 %tmp18 = mul nsw i64 %i.0, %n
|
|
56 %.sum1 = add i64 %tmp18, %j.0
|
|
57 %tmp19 = getelementptr inbounds float, float* %A, i64 %.sum1
|
|
58 %tmp20 = load float, float* %tmp19, align 4
|
|
59 %tmp21 = fadd float %tmp20, %tmp17
|
|
60 store float %tmp21, float* %tmp19, align 4
|
|
61 br label %bb22
|
|
62
|
|
63 bb22: ; preds = %bb6
|
|
64 %tmp23 = add nsw i64 %j.0, 1
|
|
65 br label %bb5
|
|
66
|
|
67 bb24: ; preds = %bb5
|
|
68 br label %bb25
|
|
69
|
|
70 bb25: ; preds = %bb24
|
|
71 %tmp26 = add nsw i64 %i.0, 1
|
|
72 br label %bb3
|
|
73
|
|
74 bb27: ; preds = %bb3
|
|
75 ret void
|
|
76 }
|