[PATCH] D99565: [X86] Support replacing aligned vector moves with unaligned moves when avx is enabled.

[Craig Topper via Phabricator via cfe-commits]

craig.topper added a comment.

The only use I could really see for this is to prevent a developers code from crashing when it’s distributed to someone else. For paranoia because it’s possible you have a bug and got lucky with alignment in your internal testing before you shipped.

If you need this your code has undefined behavior which should be fixed. You should not use this to make a known runtime exception go away.

Your code could still be miscompiled. For example, llvm really likes to replace ADD with OR when the bits don’t overlap. So it would be very easy to have your pointer arithmetic miscompiled because llvm believes a pointer is aligned but really isn’t.

Your code would not be portable to SSE. If your application does dynamic dispatch most of your users may get the AVX code path, but the smaller percentage on older hardware or cheaper hardware that doesn’t have AVX still get the exceptions.

I think you need to be very careful with how this feature is communicated.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D99565/new/

https://reviews.llvm.org/D99565