1; RUN: opt -codegenprepare -disable-cgp-branch-opts -S < %s | FileCheck %s
2target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
3target triple = "x86_64-unknown-linux-gnu"
4
5; The first cast should be sunk into block2, in order that the
6; instruction selector can form an efficient
7; i64 * i64 -> i128 multiplication.
8define i128 @sink(i64* %mem1, i64* %mem2) {
9; CHECK-LABEL: block1:
10; CHECK-NEXT: load
11block1:
12  %l1 = load i64, i64* %mem1
13  %s1 = sext i64 %l1 to i128
14  br label %block2
15
16; CHECK-NEXT: sext
17; CHECK-NEXT: load
18; CHECK-NEXT: sext
19block2:
20  %l2 = load i64, i64* %mem2
21  %s2 = sext i64 %l2 to i128
22  %res = mul i128 %s1, %s2
23  ret i128 %res
24}
25
26; The first cast should be hoisted into block1, in order that the
27; instruction selector can form an extend-load.
28define i64 @hoist(i32* %mem1, i32* %mem2) {
29; CHECK-LABEL: block1:
30; CHECK-NEXT: load
31; CHECK-NEXT: sext
32block1:
33  %l1 = load i32, i32* %mem1
34  br label %block2
35
36; CHECK-NEXT: load
37; CHECK-NEXT: sext
38block2:
39  %s1 = sext i32 %l1 to i64
40  %l2 = load i32, i32* %mem2
41  %s2 = sext i32 %l2 to i64
42  %res = mul i64 %s1, %s2
43  ret i64 %res
44}
45
46; Make sure the cast sink logic and OptimizeExtUses don't end up in an infinite
47; loop.
48define i128 @use_ext_source() {
49block1:
50  %v1 = or i64 undef, undef
51  %v2 = zext i64 %v1 to i128
52  br i1 undef, label %block2, label %block3
53
54block2:
55  %v3 = add i64 %v1, 1
56  %v4 = zext i64 %v3 to i128
57  br label %block3
58
59block3:
60  %res = phi i128 [ %v2, %block1 ], [ %v4, %block2 ]
61  ret i128 %res
62}
63