1; RUN: opt -mtriple=amdgcn-- -O3 -S -enable-new-pm=0 %s | FileCheck %s
2
3; This fails with the new pass manager:
4; https://bugs.llvm.org/show_bug.cgi?id=48819
5
6; Check that loop unswitch happened and condition hoisted out of the loop.
7; Condition is uniform so all targets should perform unswitching.
8
9; CHECK-LABEL: {{^}}define amdgpu_kernel void @uniform_unswitch
10; CHECK: entry:
11; CHECK-NEXT: [[LOOP_COND:%[a-z0-9]+]] = icmp
12; CHECK-NEXT: [[IF_COND:%[a-z0-9]+]] = icmp eq i32 %x, 123456
13; CHECK-NEXT: and i1 [[LOOP_COND]], [[IF_COND]]
14; CHECK-NEXT: br i1
15
16define amdgpu_kernel void @uniform_unswitch(i32 * nocapture %out, i32 %n, i32 %x) {
17entry:
18  %cmp6 = icmp sgt i32 %n, 0
19  br i1 %cmp6, label %for.body.lr.ph, label %for.cond.cleanup
20
21for.body.lr.ph:                                   ; preds = %entry
22  %cmp1 = icmp eq i32 %x, 123456
23  br label %for.body
24
25for.cond.cleanup.loopexit:                        ; preds = %for.inc
26  br label %for.cond.cleanup
27
28for.cond.cleanup:                                 ; preds = %for.cond.cleanup.loopexit, %entry
29  ret void
30
31for.body:                                         ; preds = %for.inc, %for.body.lr.ph
32  %i.07 = phi i32 [ 0, %for.body.lr.ph ], [ %inc, %for.inc ]
33  br i1 %cmp1, label %if.then, label %for.inc
34
35if.then:                                          ; preds = %for.body
36  %arrayidx = getelementptr inbounds i32, i32 * %out, i32 %i.07
37  store i32 %i.07, i32 * %arrayidx, align 4
38  br label %for.inc
39
40for.inc:                                          ; preds = %for.body, %if.then
41  %inc = add nuw nsw i32 %i.07, 1
42  %exitcond = icmp eq i32 %inc, %n
43  br i1 %exitcond, label %for.cond.cleanup.loopexit, label %for.body
44}
45
46; Check that loop unswitch does not happen if condition is divergent.
47
48; CHECK-LABEL: {{^}}define amdgpu_kernel void @divergent_unswitch
49; CHECK: entry:
50; CHECK: icmp
51; CHECK: [[IF_COND:%[a-z0-9]+]] = icmp {{.*}} 567890
52; CHECK: br label
53; CHECK: br i1 [[IF_COND]]
54
55define amdgpu_kernel void @divergent_unswitch(i32 * nocapture %out, i32 %n) {
56entry:
57  %cmp9 = icmp sgt i32 %n, 0
58  br i1 %cmp9, label %for.body.lr.ph, label %for.cond.cleanup
59
60for.body.lr.ph:                                   ; preds = %entry
61  %call = tail call i32 @llvm.amdgcn.workitem.id.x() #0
62  %cmp2 = icmp eq i32 %call, 567890
63  br label %for.body
64
65for.cond.cleanup.loopexit:                        ; preds = %for.inc
66  br label %for.cond.cleanup
67
68for.cond.cleanup:                                 ; preds = %for.cond.cleanup.loopexit, %entry
69  ret void
70
71for.body:                                         ; preds = %for.inc, %for.body.lr.ph
72  %i.010 = phi i32 [ 0, %for.body.lr.ph ], [ %inc, %for.inc ]
73  br i1 %cmp2, label %if.then, label %for.inc
74
75if.then:                                          ; preds = %for.body
76  %arrayidx = getelementptr inbounds i32, i32 * %out, i32 %i.010
77  store i32 %i.010, i32 * %arrayidx, align 4
78  br label %for.inc
79
80for.inc:                                          ; preds = %for.body, %if.then
81  %inc = add nuw nsw i32 %i.010, 1
82  %exitcond = icmp eq i32 %inc, %n
83  br i1 %exitcond, label %for.cond.cleanup.loopexit, label %for.body
84}
85
86declare i32 @llvm.amdgcn.workitem.id.x() #0
87
88attributes #0 = { nounwind readnone }
89