[PATCH] D53306: [X86] Stop promoting integer loads to vXi64

[Sam McCall via Phabricator via llvm-commits]

sammccall added a comment.

This broke our continuous integration, it appears to cause miscompiles in tensorflow/compiler when running on X86 CPU:

Specifically the `PoolGradTest.testMaxPool` test here: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/compiler/tests/pooling_ops_test.py#L545

Unfortunately I'm not familiar enough with the LLVM backend nor tensorflow/XLA to work out how to reproduce/reduce this.


Repository:
  rL LLVM

https://reviews.llvm.org/D53306