Incorrect MachineSink into clobbering loop

[nikic]

; RUN: llc -mtriple=x86_64-- < %s
declare void @clobber()

define void @test(i1 %c, i64* %p, i64* noalias %p2) nounwind {
entry:
  %val = load i64, i64* %p, align 8
  br label %loop

loop:
  switch i8 undef, label %unreachable [
    i8 0, label %latch
    i8 1, label %split.1
    i8 2, label %split.2
    i8 3, label %split.3
  ]

unreachable:
  unreachable

split.3:
  br i1 %c, label %clobber, label %sink

split.1:
  br label %latch

split.2:
  br label %latch

clobber:
  call void @clobber()
  br label %sink

sink:
  store i64 %val, i64* %p2, align 8
  br label %latch

latch:
  %phi = phi i64 [ 0, %sink ], [ 0, %split.2 ], [ 1, %split.1 ], [ 0, %loop ]
  %phi.live = add i64 %phi, 0
  br label %loop
}

The load is sunk into split.3, but may then be clobbered by following the split.3 -> clobber -> sink -> latch -> loop -> split.3 chain.

This is due to the functionality introduced in https://reviews.llvm.org/D86864, which uses incorrect post-dominance reasoning.

[nikic]

Candidate patch: https://reviews.llvm.org/D120330

[nikic]

/cherry-pick e075bf6 6fde043

[llvmbot]

/branch llvmbot/llvm-project/issue53990