xref: /llvm-project/llvm/test/Transforms/SimpleLoopUnswitch/AMDGPU/uniform-unswitch.ll (revision 29441e4f5fa5f5c7709f7cf180815ba97f611297)
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 3
2; RUN: opt -mtriple=amdgcn-- -passes='default<O3>' -S %s | FileCheck %s
3
4; Check that loop unswitch happened and condition hoisted out of the loop.
5; Condition is uniform so even targets with divergence should perform unswitching.
6
7; This fails with the new pass manager:
8; https://bugs.llvm.org/show_bug.cgi?id=48819
9; The correct behaviour (allow uniform non-trivial branches to be
10; unswitched on all targets) requires access to the function-level
11; divergence analysis from a loop transform, which is currently not
12; supported in the new pass manager.
13
14; SHOULDBE-LABEL: {{^}}define amdgpu_kernel void @uniform_unswitch
15; SHOULDBE: entry:
16; SHOULDBE-NEXT: [[LOOP_COND:%[a-z0-9]+]] = icmp
17; SHOULDBE-NEXT: [[IF_COND:%[a-z0-9]+]] = icmp eq i32 %x, 123456
18; SHOULDBE-NEXT: and i1 [[LOOP_COND]], [[IF_COND]]
19; SHOULDBE-NEXT: br i1
20
21define amdgpu_kernel void @uniform_unswitch(ptr nocapture %out, i32 %n, i32 %x) {
22; CHECK-LABEL: define amdgpu_kernel void @uniform_unswitch(
23; CHECK-SAME: ptr writeonly captures(none) [[OUT:%.*]], i32 [[N:%.*]], i32 [[X:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {
24; CHECK-NEXT:  entry:
25; CHECK-NEXT:    [[OUT_GLOBAL:%.*]] = addrspacecast ptr [[OUT]] to ptr addrspace(1)
26; CHECK-NEXT:    [[CMP6:%.*]] = icmp sgt i32 [[N]], 0
27; CHECK-NEXT:    br i1 [[CMP6]], label [[FOR_BODY_LR_PH:%.*]], label [[FOR_COND_CLEANUP:%.*]]
28; CHECK:       for.body.lr.ph:
29; CHECK-NEXT:    [[CMP1:%.*]] = icmp eq i32 [[X]], 123456
30; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
31; CHECK:       for.cond.cleanup:
32; CHECK-NEXT:    ret void
33; CHECK:       for.body:
34; CHECK-NEXT:    [[I_07:%.*]] = phi i32 [ 0, [[FOR_BODY_LR_PH]] ], [ [[INC:%.*]], [[FOR_INC:%.*]] ]
35; CHECK-NEXT:    br i1 [[CMP1]], label [[IF_THEN:%.*]], label [[FOR_INC]]
36; CHECK:       if.then:
37; CHECK-NEXT:    [[TMP0:%.*]] = zext nneg i32 [[I_07]] to i64
38; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds nuw i32, ptr addrspace(1) [[OUT_GLOBAL]], i64 [[TMP0]]
39; CHECK-NEXT:    store i32 [[I_07]], ptr addrspace(1) [[ARRAYIDX]], align 4
40; CHECK-NEXT:    br label [[FOR_INC]]
41; CHECK:       for.inc:
42; CHECK-NEXT:    [[INC]] = add nuw nsw i32 [[I_07]], 1
43; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp eq i32 [[INC]], [[N]]
44; CHECK-NEXT:    br i1 [[EXITCOND]], label [[FOR_COND_CLEANUP]], label [[FOR_BODY]]
45;
46entry:
47  %cmp6 = icmp sgt i32 %n, 0
48  br i1 %cmp6, label %for.body.lr.ph, label %for.cond.cleanup
49
50for.body.lr.ph:                                   ; preds = %entry
51  %cmp1 = icmp eq i32 %x, 123456
52  br label %for.body
53
54for.cond.cleanup.loopexit:                        ; preds = %for.inc
55  br label %for.cond.cleanup
56
57for.cond.cleanup:                                 ; preds = %for.cond.cleanup.loopexit, %entry
58  ret void
59
60for.body:                                         ; preds = %for.inc, %for.body.lr.ph
61  %i.07 = phi i32 [ 0, %for.body.lr.ph ], [ %inc, %for.inc ]
62  br i1 %cmp1, label %if.then, label %for.inc
63
64if.then:                                          ; preds = %for.body
65  %arrayidx = getelementptr inbounds i32, ptr %out, i32 %i.07
66  store i32 %i.07, ptr %arrayidx, align 4
67  br label %for.inc
68
69for.inc:                                          ; preds = %for.body, %if.then
70  %inc = add nuw nsw i32 %i.07, 1
71  %exitcond = icmp eq i32 %inc, %n
72  br i1 %exitcond, label %for.cond.cleanup.loopexit, label %for.body
73}
74
75declare i32 @llvm.amdgcn.workitem.id.x() #0
76
77attributes #0 = { nounwind readnone }
78