[clang] [BoundsSafety][NFC] Introduce logic to replace the LateParsedAttrType placeholder with CountAttributedType (PR #212906)

[Erich Keane via cfe-commits]

erichkeane wrote:

> > The secret I think would be to still do the Profile, but instead of 'filling in' the attribute when you complete it, you replace it. Not uniquing it is a solution, but I would like to avoid doing something particularly 'special' here?
> 
> @erichkeane By "instead of 'filling in' the attribute when you complete it, you replace it" — do you mean creating a fresh CountAttributedType and replacing the incomplete one? My concern with that is that every enclosing type node would then need to be rebuilt. That's why replacing the placeholder needed a TreeTransform, whereas filling in the expression wouldn't: the enclosing nodes keep pointing at the same node and don't change at all.
> 
> On 'special' — I think there's precedent for types that skip Profile, particularly ones holding expressions, since expressions aren't uniqued themselves. getVariableArrayType says so directly: "Since we don't unique expressions, it isn't possible to unique VLA's that have an expression provided for their size." A few other types skip it too.
> 
> Thanks for staying with this, and for the suggestions — I think we're converging on something good and I'll experiment with it.

Ah, right, I forgot it is type modifying.  I think skipping profile/uniquing is acceptable/necessary.

https://github.com/llvm/llvm-project/pull/212906