; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
; RUN: opt -S -passes=loop-idiom < %s | FileCheck %s

; The poison flags should be preserved, as no transform takes place.
define void @test(ptr %p.end, ptr %p.start) {
; CHECK-LABEL: define void @test(
; CHECK-SAME: ptr [[P_END:%.*]], ptr [[P_START:%.*]]) {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[P_END_INT:%.*]] = ptrtoint ptr [[P_END]] to i64
; CHECK-NEXT:    [[P_START_INT:%.*]] = ptrtoint ptr [[P_START]] to i64
; CHECK-NEXT:    [[DIST:%.*]] = sub nuw i64 [[P_END_INT]], [[P_START_INT]]
; CHECK-NEXT:    [[LEN:%.*]] = lshr exact i64 [[DIST]], 5
; CHECK-NEXT:    [[CMP:%.*]] = icmp eq ptr [[P_END]], [[P_START]]
; CHECK-NEXT:    br i1 [[CMP]], label [[EXIT:%.*]], label [[PREHEADER:%.*]]
; CHECK:       preheader:
; CHECK-NEXT:    br label [[LOOP:%.*]]
; CHECK:       loop:
; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ [[LEN]], [[PREHEADER]] ], [ [[IV_DEC:%.*]], [[LOOP]] ]
; CHECK-NEXT:    [[IV_NEG:%.*]] = sub nsw i64 0, [[IV]]
; CHECK-NEXT:    [[SRC:%.*]] = getelementptr inbounds [32 x i8], ptr [[P_END]], i64 [[IV_NEG]]
; CHECK-NEXT:    [[V:%.*]] = load <4 x i64>, ptr [[SRC]], align 8
; CHECK-NEXT:    [[DST:%.*]] = getelementptr inbounds [32 x i8], ptr [[P_START]], i64 [[IV_NEG]]
; CHECK-NEXT:    store <4 x i64> [[V]], ptr [[DST]], align 8
; CHECK-NEXT:    [[IV_DEC]] = add i64 [[IV]], -1
; CHECK-NEXT:    [[CMP2:%.*]] = icmp eq i64 [[IV_DEC]], 0
; CHECK-NEXT:    br i1 [[CMP2]], label [[LOOP_EXIT:%.*]], label [[LOOP]]
; CHECK:       loop.exit:
; CHECK-NEXT:    br label [[EXIT]]
; CHECK:       exit:
; CHECK-NEXT:    ret void
;
entry:
  %p.end.int = ptrtoint ptr %p.end to i64
  %p.start.int = ptrtoint ptr %p.start to i64
  %dist = sub nuw i64 %p.end.int, %p.start.int
  %len = lshr exact i64 %dist, 5
  %cmp = icmp eq ptr %p.end, %p.start
  br i1 %cmp, label %exit, label %preheader

preheader:
  br label %loop

loop:
  %iv = phi i64 [ %len, %preheader ], [ %iv.dec, %loop ]
  %iv.neg = sub nsw i64 0, %iv
  %src = getelementptr inbounds [32 x i8], ptr %p.end, i64 %iv.neg
  %v = load <4 x i64>, ptr %src, align 8
  %dst = getelementptr inbounds [32 x i8], ptr %p.start, i64 %iv.neg
  store <4 x i64> %v, ptr %dst, align 8
  %iv.dec = add i64 %iv, -1
  %cmp2 = icmp eq i64 %iv.dec, 0
  br i1 %cmp2, label %loop.exit, label %loop

loop.exit:
  br label %exit

exit:
  ret void
}