WalkUtils: handle unsafe_ref_cast which casts between AnyObject and a class

[eeckstein]

We need to ignore such casts because otherwise the constructed walking path wouldn't contain the right existential field kind. Fixes a miscompile caused by redundant load elimination.

rdar://132364917

[eeckstein]

@swift-ci test

[eeckstein]

@swift-ci benchmark

[atrick]

Approved for 5.10 and 6.0 CCC, because the change seems conservative.

But I don't understand what problem is being fixed. Surely this test already reported "may alias" before this change, because the use-def walk from both ref_element_addr instructions would have found the same root.
Was the problem that the def-use walk did not match the use-def walk because the path did not include the correct type? Did it optimistically assume that AnyObject cannot have a field of type s? I just wonder how that failure to match the type happened and how it ended up being optimistic instead of pessimistic.

 CHECK-LABEL: Address escape information for test_anyobject_cast:
// CHECK:       pair 0 - 1
// CHECK-NEXT:    %3 = ref_element_addr %2 : $X, #X.s
// CHECK-NEXT:    %5 = ref_element_addr %2 : $X, #X.s
// CHECK-NEXT:  may alias
// CHECK:       End function test_anyobject_cast
sil @test_anyobject_cast : $@convention(thin) () -> () {
bb0:
  %0 = alloc_ref $X
  %1 = init_existential_ref %0 : $X : $X, $AnyObject
  %2 = unchecked_ref_cast %1 : $AnyObject to $X
  %3 = ref_element_addr %2 : $X, #X.s
  fix_lifetime %3 : $*Str
  %5 = ref_element_addr %2 : $X, #X.s
  fix_lifetime %5 : $*Str
  %r = tuple ()
  return %r : $()
}

[eeckstein]

This test did not report "may alias" before this change. The reason was a mismatch of init- and open existential instructions (unchecked_ref_cast was not considered as potentially "open" the AnyObject existential).
The EscapeInfo walkers push an existential field onto the walking path when walking up an open_existential_ref instruction. And try to pop it when seeing an init_existential_ref. But if there is no such field type in the path, it's considered a path mismatch which aborts the walk and the result is a no-escape.

[atrick]

Thanks. I understand. What is the purpose of creating a path component for init_existential_ref %0 : $X : $X, $AnyObject if it's not because of the type change? Converting to AnyObject from a class can't change the representation right? Then you could continue ignoring unchecked_ref_cast

[eeckstein]

Yes, ignoring creating AnyObject existentials would be a possibility. Currently we don't treat AnyObject different than other existentials.