|
236
|
1 ; RUN: opt %loadPolly -polly-invariant-load-hoisting=true -polly-print-scops -disable-output < %s | FileCheck %s
|
|
|
2 ; RUN: opt %loadPolly -polly-invariant-load-hoisting=true -polly-overflow-tracking=never -polly-codegen -S < %s | FileCheck %s --check-prefix=IR
|
|
|
3 ;
|
|
|
4 ; As (p + q) can overflow we have to check that we load from
|
|
|
5 ; I[p + q] only if it does not.
|
|
|
6 ;
|
|
|
7 ; CHECK: Invariant Accesses: {
|
|
|
8 ; CHECK-NEXT: ReadAccess := [Reduction Type: NONE] [Scalar: 0]
|
|
|
9 ; CHECK-NEXT: [N, p, q] -> { Stmt_for_body[i0] -> MemRef_I[p + q] };
|
|
|
10 ; CHECK-NEXT: Execution Context: [N, p, q] -> { : N > 0 and -2147483648 - p <= q <= 2147483647 - p }
|
|
|
11 ; CHECK-NEXT: ReadAccess := [Reduction Type: NONE] [Scalar: 0]
|
|
|
12 ; CHECK-NEXT: [N, p, q] -> { Stmt_for_body[i0] -> MemRef_tmp1[0] };
|
|
|
13 ; CHECK-NEXT: Execution Context: [N, p, q] -> { : N > 0 }
|
|
|
14 ; CHECK-NEXT: }
|
|
|
15 ;
|
|
|
16 ; IR: polly.preload.merge:
|
|
|
17 ; IR-NEXT: %polly.preload.tmp1.merge = phi i32* [ %polly.access.I.load, %polly.preload.exec ], [ null, %polly.preload.cond ]
|
|
|
18 ; IR-NEXT: store i32* %polly.preload.tmp1.merge, i32** %tmp1.preload.s2a
|
|
|
19 ; IR-NEXT: %12 = sext i32 %N to i64
|
|
|
20 ; IR-NEXT: %13 = icmp sge i64 %12, 1
|
|
|
21 ; IR-NEXT: %14 = sext i32 %q to i64
|
|
|
22 ; IR-NEXT: %15 = sext i32 %p to i64
|
|
|
23 ; IR-NEXT: %16 = add nsw i64 %15, %14
|
|
|
24 ; IR-NEXT: %17 = icmp sle i64 %16, 2147483647
|
|
|
25 ; IR-NEXT: %18 = and i1 %13, %17
|
|
|
26 ; IR-NEXT: %19 = sext i32 %q to i64
|
|
|
27 ; IR-NEXT: %20 = sext i32 %p to i64
|
|
|
28 ; IR-NEXT: %21 = add nsw i64 %20, %19
|
|
|
29 ; IR-NEXT: %22 = icmp sge i64 %21, -2147483648
|
|
|
30 ; IR-NEXT: %23 = and i1 %18, %22
|
|
|
31 ; IR-NEXT: %polly.preload.cond.result1 = and i1 %23, true
|
|
|
32 ; IR-NEXT: br label %polly.preload.cond2
|
|
|
33 ;
|
|
|
34 ; IR: polly.preload.cond2:
|
|
|
35 ; IR-NEXT: br i1 %polly.preload.cond.result1
|
|
|
36 ;
|
|
|
37 ; IR: polly.preload.exec4:
|
|
|
38 ; IR-NEXT: %polly.access.polly.preload.tmp1.merge = getelementptr i32, i32* %polly.preload.tmp1.merge, i64 0
|
|
|
39 ; IR-NEXT: %polly.access.polly.preload.tmp1.merge.load = load i32, i32* %polly.access.polly.preload.tmp1.merge, align 4
|
|
|
40 ;
|
|
|
41 ; void f(int **I, int *A, int N, int p, int q) {
|
|
|
42 ; for (int i = 0; i < N; i++)
|
|
|
43 ; A[i] = *(I[p + q]);
|
|
|
44 ; }
|
|
|
45 ;
|
|
|
46 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
|
|
|
47
|
|
|
48 define void @f(i32** %I, i32* %A, i32 %N, i32 %p, i32 %q) {
|
|
|
49 entry:
|
|
|
50 %tmp = sext i32 %N to i64
|
|
|
51 br label %for.cond
|
|
|
52
|
|
|
53 for.cond: ; preds = %for.inc, %entry
|
|
|
54 %indvars.iv = phi i64 [ %indvars.iv.next, %for.inc ], [ 0, %entry ]
|
|
|
55 %cmp = icmp slt i64 %indvars.iv, %tmp
|
|
|
56 br i1 %cmp, label %for.body, label %for.end
|
|
|
57
|
|
|
58 for.body: ; preds = %for.cond
|
|
|
59 %add = add i32 %p, %q
|
|
|
60 %idxprom = sext i32 %add to i64
|
|
|
61 %arrayidx = getelementptr inbounds i32*, i32** %I, i64 %idxprom
|
|
|
62 %tmp1 = load i32*, i32** %arrayidx, align 8
|
|
|
63 %tmp2 = load i32, i32* %tmp1, align 4
|
|
|
64 %arrayidx2 = getelementptr inbounds i32, i32* %A, i64 %indvars.iv
|
|
|
65 store i32 %tmp2, i32* %arrayidx2, align 4
|
|
|
66 br label %for.inc
|
|
|
67
|
|
|
68 for.inc: ; preds = %for.body
|
|
|
69 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
|
|
|
70 br label %for.cond
|
|
|
71
|
|
|
72 for.end: ; preds = %for.cond
|
|
|
73 ret void
|
|
|
74 }
|
