1; RUN: opt %loadPolly -polly-stmt-granularity=bb -polly-mse -polly-print-scops -disable-output < %s | FileCheck %s
2; RUN: opt %loadNPMPolly -polly-stmt-granularity=bb "-passes=scop(print<polly-mse>)" -disable-output < %s | FileCheck %s
3;
4; Verify that the accesses are correctly expanded
5;
6; Original source code :
7;
8; #define Ni 2000
9; #define Nj 3000
10;
11; void mse(double A[Ni], double B[Nj], double C[Nj], double D[Nj]) {
12;   int i,j;
13;   for (j = 0; j < Ni; j++) {
14;     for (int i = 0; i<Nj; i++)
15;       B[i] = i;
16;
17;     for (int i = 0; i<Nj; i++)
18;       D[i] = i;
19;
20;     A[j] = B[j];
21;     C[j] = D[j];
22;   }
23; }
24;
25; Check that expanded SAI are created
26;
27; CHECK: double MemRef_B_Stmt_for_body4_expanded[10000][10000]; // Element size 8
28; CHECK: double MemRef_D_Stmt_for_body9_expanded[10000][10000]; // Element size 8
29; CHECK: i64 MemRef_A_Stmt_for_end15_expanded[10000]; // Element size 8
30; CHECK: i64 MemRef_C_Stmt_for_end15_expanded[10000]; // Element size 8
31;
32; Check that the memory accesses are modified
33; CHECK: new: { Stmt_for_body4[i0, i1] -> MemRef_B_Stmt_for_body4_expanded[i0, i1] };
34; CHECK: new: { Stmt_for_body9[i0, i1] -> MemRef_D_Stmt_for_body9_expanded[i0, i1] };
35; CHECK: new: { Stmt_for_end15[i0] -> MemRef_B_Stmt_for_body4_expanded[i0, i0] };
36; CHECK: new: { Stmt_for_end15[i0] -> MemRef_A_Stmt_for_end15_expanded[i0] };
37; CHECK: new: { Stmt_for_end15[i0] -> MemRef_D_Stmt_for_body9_expanded[i0, i0] };
38; CHECK: new: { Stmt_for_end15[i0] -> MemRef_C_Stmt_for_end15_expanded[i0] };
39;
40target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
41target triple = "x86_64-unknown-linux-gnu"
42
43define void @mse(double* %A, double* %B, double* %C, double* %D) {
44entry:
45  br label %entry.split
46
47entry.split:                                      ; preds = %entry
48  br label %for.body
49
50for.body:                                         ; preds = %entry.split, %for.end15
51  %indvars.iv7 = phi i64 [ 0, %entry.split ], [ %indvars.iv.next8, %for.end15 ]
52  br label %for.body4
53
54for.body4:                                        ; preds = %for.body, %for.body4
55  %indvars.iv = phi i64 [ 0, %for.body ], [ %indvars.iv.next, %for.body4 ]
56  %0 = trunc i64 %indvars.iv to i32
57  %conv = sitofp i32 %0 to double
58  %arrayidx = getelementptr inbounds double, double* %B, i64 %indvars.iv
59  store double %conv, double* %arrayidx, align 8
60  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
61  %exitcond = icmp ne i64 %indvars.iv.next, 10000
62  br i1 %exitcond, label %for.body4, label %for.end
63
64for.end:                                          ; preds = %for.body4
65  br label %for.body9
66
67for.body9:                                        ; preds = %for.end, %for.body9
68  %indvars.iv4 = phi i64 [ 0, %for.end ], [ %indvars.iv.next5, %for.body9 ]
69  %1 = trunc i64 %indvars.iv4 to i32
70  %conv10 = sitofp i32 %1 to double
71  %arrayidx12 = getelementptr inbounds double, double* %D, i64 %indvars.iv4
72  store double %conv10, double* %arrayidx12, align 8
73  %indvars.iv.next5 = add nuw nsw i64 %indvars.iv4, 1
74  %exitcond6 = icmp ne i64 %indvars.iv.next5, 10000
75  br i1 %exitcond6, label %for.body9, label %for.end15
76
77for.end15:                                        ; preds = %for.body9
78  %arrayidx17 = getelementptr inbounds double, double* %B, i64 %indvars.iv7
79  %2 = bitcast double* %arrayidx17 to i64*
80  %3 = load i64, i64* %2, align 8
81  %arrayidx19 = getelementptr inbounds double, double* %A, i64 %indvars.iv7
82  %4 = bitcast double* %arrayidx19 to i64*
83  store i64 %3, i64* %4, align 8
84  %arrayidx21 = getelementptr inbounds double, double* %D, i64 %indvars.iv7
85  %5 = bitcast double* %arrayidx21 to i64*
86  %6 = load i64, i64* %5, align 8
87  %arrayidx23 = getelementptr inbounds double, double* %C, i64 %indvars.iv7
88  %7 = bitcast double* %arrayidx23 to i64*
89  store i64 %6, i64* %7, align 8
90  %indvars.iv.next8 = add nuw nsw i64 %indvars.iv7, 1
91  %exitcond9 = icmp ne i64 %indvars.iv.next8, 10000
92  br i1 %exitcond9, label %for.body, label %for.end26
93
94for.end26:                                        ; preds = %for.end15
95  ret void
96}
97