[PATCH] D16986: [LICM] Don't assert on volatile accesses

[Matthew Simpson via llvm-commits]

mssimpso added a comment.

Ed,

It's not surprising to me that TBAA would produce more precise results (and thus an additional alias set in this case). The essential piece is that, in the test case, it's claiming that the address of b cannot be null, which is correct I think. That allows LICM to remove load2 and change load3 to load from null. The alias set containing null cannot be marked volatile since it doesn't contain any loads or stores. But now null is used by load3, which is volatile. So your original patch sounded like a reasonable solution to me! Hopefully someone more knowledgeable will chime in.

In your patch, the check for isVolatile is done in IsUnordered.


http://reviews.llvm.org/D16986