[PATCH] D89903: [CodeGen] Crash instead of generating broken code with self-capturing __block var

[ille via Phabricator via cfe-commits]

ille added a comment.

> We do not actually support allocation failure for a lot of things around blocks. I don't think the copy-helper functions even have a way to propagate out a failure in copying a field. I have never seen any code in the wild that would handle Block_copy returning a null pointer. Effectively it is assumed to not happen.

Fair – although if that's the case, perhaps xnu should not be using blocks.  I suppose failures of small allocations like this are rare enough that the issue hasn't come up in practice, and same with the null blocks under ARC.

> It seems somewhat unlikely to me that anyone would actually write code like your example without copying the block and potentially triggering the `__block` variable to be moved to the heap, which is why I think pre-moving the variable might be acceptable.

Also fair.  I can implement this, but I do think there should be an associated warning flag.

> With all that said, I agree that crashing and/or just not drilling into the variable is not acceptable.

The patch I submitted switches from not drilling into the variable to crashing.  Are you saying I should submit a more thorough fix rather than trying to land this first?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D89903/new/

https://reviews.llvm.org/D89903