|
236
|
1 ; RUN: opt %loadPolly -polly-mse -polly-print-scops -disable-output < %s | FileCheck %s
|
|
|
2 ; RUN: opt %loadNPMPolly "-passes=scop(print<polly-mse>)" -disable-output < %s | FileCheck %s
|
|
150
|
3 ;
|
|
|
4 ; Verify that the accesses are correctly expanded for MemoryKind::Array
|
|
|
5 ;
|
|
|
6 ; Original source code :
|
|
|
7 ;
|
|
|
8 ; #define Ni 2000
|
|
|
9 ; #define Nj 3000
|
|
236
|
10 ;
|
|
150
|
11 ; double mse(double A[Ni], double B[Nj]) {
|
|
|
12 ; int i;
|
|
|
13 ; double tmp = 6;
|
|
|
14 ; for (i = 0; i < Ni; i++) {
|
|
|
15 ; for (int j = 0; j<Nj; j++) {
|
|
|
16 ; B[j] = j;
|
|
|
17 ; }
|
|
236
|
18 ; A[i] = B[i];
|
|
150
|
19 ; }
|
|
|
20 ; return tmp;
|
|
|
21 ; }
|
|
|
22 ;
|
|
|
23 ; Check if the expanded SAI are created
|
|
|
24 ;
|
|
|
25 ; CHECK: double MemRef_B_Stmt_for_body3_expanded[2000][3000]; // Element size 8
|
|
|
26 ;
|
|
|
27 ; Check if the memory accesses are modified
|
|
|
28 ;
|
|
|
29 ; CHECK: new: { Stmt_for_body3[i0, i1] -> MemRef_B_Stmt_for_body3_expanded[i0, i1] };
|
|
|
30 ; CHECK: new: { Stmt_for_end[i0] -> MemRef_B_Stmt_for_body3_expanded[i0, i0] };
|
|
|
31 ;
|
|
|
32 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
|
|
|
33 target triple = "x86_64-unknown-linux-gnu"
|
|
|
34
|
|
|
35 define double @mse(double* %A, double* %B) {
|
|
|
36 entry:
|
|
|
37 br label %entry.split
|
|
|
38
|
|
|
39 entry.split: ; preds = %entry
|
|
|
40 br label %for.body
|
|
|
41
|
|
|
42 for.body: ; preds = %entry.split, %for.end
|
|
|
43 %indvars.iv3 = phi i64 [ 0, %entry.split ], [ %indvars.iv.next4, %for.end ]
|
|
|
44 br label %for.body3
|
|
|
45
|
|
|
46 for.body3: ; preds = %for.body, %for.body3
|
|
|
47 %indvars.iv = phi i64 [ 0, %for.body ], [ %indvars.iv.next, %for.body3 ]
|
|
|
48 %0 = trunc i64 %indvars.iv to i32
|
|
|
49 %conv = sitofp i32 %0 to double
|
|
|
50 %arrayidx = getelementptr inbounds double, double* %B, i64 %indvars.iv
|
|
|
51 store double %conv, double* %arrayidx, align 8
|
|
|
52 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
|
|
|
53 %exitcond = icmp ne i64 %indvars.iv.next, 3000
|
|
|
54 br i1 %exitcond, label %for.body3, label %for.end
|
|
|
55
|
|
|
56 for.end: ; preds = %for.body3
|
|
|
57 %arrayidx5 = getelementptr inbounds double, double* %B, i64 %indvars.iv3
|
|
|
58 %1 = bitcast double* %arrayidx5 to i64*
|
|
|
59 %2 = load i64, i64* %1, align 8
|
|
|
60 %arrayidx7 = getelementptr inbounds double, double* %A, i64 %indvars.iv3
|
|
|
61 %3 = bitcast double* %arrayidx7 to i64*
|
|
|
62 store i64 %2, i64* %3, align 8
|
|
|
63 %indvars.iv.next4 = add nuw nsw i64 %indvars.iv3, 1
|
|
|
64 %exitcond5 = icmp ne i64 %indvars.iv.next4, 2000
|
|
|
65 br i1 %exitcond5, label %for.body, label %for.end10
|
|
|
66
|
|
|
67 for.end10: ; preds = %for.end
|
|
|
68 ret double 6.000000e+00
|
|
|
69 }
|
