[llvm-commits] [PATCH] Add missing X86 FMA patterns to folding tables

[Craig Topper]

Sure we can go ahead with the current patch. I can commit it later tonight
if you would like.

On Tue, Oct 30, 2012 at 1:55 PM, Cameron McInally
<cameron.mcinally at nyu.edu>wrote:

> On Tue, Oct 30, 2012 at 4:10 PM, Craig Topper <craig.topper at gmail.com>wrote:
>
>> That's true too. I've been meaning to go through and remove the
>> alignments from all the AVX patterns in X86InstrSSE.td and in
>> X86InstrInfo.cpp. I believe FMA4 already doesn't alignments in the td file.
>
>
> Yup, the memop pattern fragment references the unaligned memory target
> flag, namely Subtarget->hasVectorUAMem(), and checks the load's actual
> alignment. Removing the alignment checks in the folding table will show a
> good number of new folds in practice.
>
> We could extend the memop pattern fragments further for small performance
> gains on specific architectures, like SandyBridge. But, I think I'll have
> to save that for another day.
>
> Would you mind if we go ahead with the last patch and handle the alignment
> issues separately? I am under a strict code sharing policy and project
> creep like such can cause trouble.
>
> Also, I do not have commit access for trunk. Just FYI.
>
> Thanks again,
> Cameron
>
>


-- 
~Craig
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20121030/8bd63ffb/attachment.html>