[PATCH] D106053: [CodeGen] Remove pending AssertZext AssertSext in promoting FP_TO_INT

[Xiang Zhang via Phabricator via llvm-commits]

xiangzhangllvm added a comment.

@efriedma We have a job blocked by this patch, so I hope we can quickly got the key point.

Let me duplicate my previous statement:

  Or we can see in this way. 
  In the edge calculation, we just need to load 4 elements of v8f32, but for performance reason, we usually load full (8) elements. So some of its loaded element has no meaning (may over flow).
  So the shuffle don't select these element, but the shuffle may zero these elements then for other use. (very just like this small reproduce case )
  
  +  After appending AssertZext, the shuffle optimization will stop zero these elements, and directly use them.   That is wrong.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D106053/new/

https://reviews.llvm.org/D106053