[PATCH] D99531: [x86] enhance matching of pmaddwd

Mon Mar 29 13:35:31 PDT 2021

spatel added a comment.

In D99531#2656756 <https://reviews.llvm.org/D99531#2656756>, @craig.topper wrote:

> This seems reasonable. Is it possible we also need to handle the case where all the extract indices from a different subvector of the input? It's hard to say from the test case that we started from since it appears to have a logic mistake that exposed the bug in the first place. It used "sizeof sizeof(__r) in its loop control which caused the loop bounds to be 4. Seems like it should have been sizeof(__r)/sizeof(int16_t).

Right - that source looks unintended.
We could go further in matching, but it gets harder to keep track of the indices and then calculate the offsets, and it doesn't seem as likely to come up in practice. This case seemed plausible from valid source even (!), so that's why I figured we should try to get it.

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D99531/new/

https://reviews.llvm.org/D99531