test/Transforms/LoopRotate/convergent-controlled.ll

; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
; RUN: opt < %s -passes=loop-rotate -S | FileCheck %s

; LoopRotate restructures a loop so the exit test appears at the end of the
; loop body rather than the beginning. This test ensures the pass does not
; rotate a loop whose header contains a convergence.loop token that guards a
; convergent operation. Rotating the header would place the exit test before
; the convergence token, which breaks the convergence contract and enables
; later passes to collapse the loop into straight-line code.
;
; This is one of three sibling tests (LoopRotate, IndVarSimplify,
; SimpleLoopUnswitch) that each guard against a pass moving or removing the
; exit condition of a loop with a convergent operation, which can change the
; set of active threads inside the loop. See https://godbolt.org/z/K1a8xW1TE
; for the problematic pass pipeline and
; https://github.com/llvm/llvm-project/issues/180621 for full context.

define void @convergent_loop(i32 %tid, ptr %array) #0 {
; CHECK-LABEL: @convergent_loop(
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[TOK_ENTRY:%.*]] = call token @llvm.experimental.convergence.entry()
; CHECK-NEXT:    br label [[FOR_COND:%.*]]
; CHECK:       for.cond:
; CHECK-NEXT:    [[I:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[INC:%.*]], [[FOR_BODY:%.*]] ]
; CHECK-NEXT:    [[TOK_LOOP:%.*]] = call token @llvm.experimental.convergence.loop() [ "convergencectrl"(token [[TOK_ENTRY]]) ]
; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i32 [[I]], 8
; CHECK-NEXT:    br i1 [[CMP]], label [[FOR_BODY]], label [[FOR_END_LOOPEXIT:%.*]]
; CHECK:       for.body:
; CHECK-NEXT:    [[CMP1:%.*]] = icmp eq i32 [[I]], [[TID:%.*]]
; CHECK-NEXT:    [[INC]] = add nuw i32 [[I]], 1
; CHECK-NEXT:    br i1 [[CMP1]], label [[IF_THEN:%.*]], label [[FOR_COND]]
; CHECK:       if.then:
; CHECK-NEXT:    [[WAVE:%.*]] = call i32 @llvm.dx.wave.reduce.umax.i32(i32 [[TID]]) [ "convergencectrl"(token [[TOK_LOOP]]) ]
; CHECK-NEXT:    [[GEP:%.*]] = getelementptr inbounds i32, ptr [[ARRAY:%.*]], i32 [[TID]]
; CHECK-NEXT:    store i32 [[WAVE]], ptr [[GEP]], align 4
; CHECK-NEXT:    br label [[FOR_END:%.*]]
; CHECK:       for.end.loopexit:
; CHECK-NEXT:    br label [[FOR_END]]
; CHECK:       for.end:
; CHECK-NEXT:    ret void
;
entry:
  %tok.entry = call token @llvm.experimental.convergence.entry()
  br label %for.cond

for.cond:
  %i = phi i32 [ 0, %entry ], [ %inc, %for.body ]
  %tok.loop = call token @llvm.experimental.convergence.loop() [ "convergencectrl"(token %tok.entry) ]
  %cmp = icmp ult i32 %i, 8
  br i1 %cmp, label %for.body, label %for.end

for.body:
  %cmp1 = icmp eq i32 %i, %tid
  %inc = add nuw i32 %i, 1
  br i1 %cmp1, label %if.then, label %for.cond

if.then:
  %wave = call i32 @llvm.dx.wave.reduce.umax.i32(i32 %tid) [ "convergencectrl"(token %tok.loop) ]
  %gep = getelementptr inbounds i32, ptr %array, i32 %tid
  store i32 %wave, ptr %gep, align 4
  br label %for.end

for.end:
  ret void
}

declare token @llvm.experimental.convergence.entry() #1
declare token @llvm.experimental.convergence.loop() #1
declare i32 @llvm.dx.wave.reduce.umax.i32(i32) #1

attributes #0 = { convergent mustprogress nounwind "no-trapping-math"="true" }
attributes #1 = { convergent }