[PATCH] D22859: [X86][SSE] Enable target shuffle combining to combine multiple shuffle inputs.

[Simon Pilgrim via llvm-commits]

RKSimon added a comment.

In https://reviews.llvm.org/D22859#504618, @andreadb wrote:

> Hi Simon,
>
> Overall the patch lgtm.
>  My major concern was related to the complexity of your new algorithm. 'combineX86ShufflesRecursively' is now called for every "used" shuffle operand. Have you noticed any particular cases where this function is now expensive?
>  However, given that we already limit the maximum recursion depth to 8, I don't expect this to be much of an issue.
>
> -Andrea


Thanks Andrea.

I haven't found anything to indicate that the search has become notably more costly to builds - between the limits on depth and hasOneUse(), many shuffles already early out. If anybody can create a repro then I can investigate ways to manipulate the max search depth but this hasn't proved to be necessary yet.


Repository:
  rL LLVM

https://reviews.llvm.org/D22859