[PATCH] [X86] Skip concat_vectors when lowering vector broadcast

Robert Lougher rob.lougher at gmail.com
Wed Dec 11 13:03:25 PST 2013

Hi Cameron,

On 11 December 2013 19:42, Cameron McInally <cameron.mcinally at nyu.edu> wrote:
> Hey Rob,
> On Wed, Dec 11, 2013 at 12:54 PM, Robert Lougher <rob.lougher at gmail.com> wrote:

>> The attached patch fixes this by skipping the concat_vectors during
>> pattern recognition.  In this case, once the concat_vectors is
>> skipped, we get a BUILD_VECTOR, and the pattern matches.
> Yes, this makes sense to me. I've been carrying this change locally
> since late 2012 and have not had problems IIRC. Also IIRC, the
> broadcast showed a slight performance gain over the shuffle and insert
> sequence. Our compiler is currently on LLVM 3.3 though.

I'm glad to see that other people had the same idea!

> If there are no other comments on a better solution, I will commit
> this patch for you.

Thank you.  Let's see what others think.


> -Cameron

More information about the llvm-commits mailing list