[PATCH] D49403: More aggressively complete RecordTypes with Function Pointers

[John McCall via Phabricator via cfe-commits]

rjmccall added a comment.

Sorry, let me be more precise.  The semantics of LLVM IR are that the element type of a pointer matters when doing specific operations on pointers: loads, stores, GEPs, calls, and so on.  Many of these operations have been at least partially updated — in some combination of the APIs for constructing them and the IR assembly listings — to require the "element" type to be given separately, but that's not really instrumental here.  What's instrumental is that, for almost all of these operations, the only thing that matters about the element type is its basic layout.  For example, it is perfectly legal in IR to load a `float` by loading any type that happens to be the same size as a `float` and then bitcasting that value to `float` — well, maybe it's not okay to load it as a `struct` with internal padding, because I think the padding bits might take on unspecified values, but anything else.  It is also legal in IR to do pointer arithmetic by using any combination of GEPs that happens to yield the right offset.  The only pointer operations where the specific details of the element type actually have semantic weight beyond their layout are calls.

This general lack of meaning is part of the reason we'd like to remove element types entirely.

I'm sorry, I'm not interested in taking the patch.


https://reviews.llvm.org/D49403