[llvm] r214042 - [x86] Add a much more powerful framework for combining x86 shuffle
peter_cooper at apple.com
Sun Jul 27 20:11:10 PDT 2014
Thanks for the great explanation!
Sorry to call out Atom. That's not actually a fair question as yeah, they will be crazy expensive there. But I don't expect any regressions based on what you said.
The level of HW you've gone back to is all I tend to care about too. I'm not seeing anyone ask for my 2004 P4 to be dusted off to benchmark llvm :)
Sent from my iPhone
> On Jul 27, 2014, at 6:41 PM, Chandler Carruth <chandlerc at gmail.com> wrote:
>> On Sun, Jul 27, 2014 at 6:37 PM, Chandler Carruth <chandlerc at gmail.com> wrote:
>> and can't be
> Ah, with one exception: if you used the builtins or llvm intrinsics directly to force the emission of pshufd, pshuflw, and pshufhw in long chains, you could trigger this. But you can't even trigger it really from _mm_shuffle_epi32, _mm_shufflelo_epi16, or _mm_shufflehi_epi16 because those lower through generic shuffles in LLVM.... =/
> Anyways, do let me know if you're seeing regressions due to this.. I'd be fascinated with how, and its easy to put it too behind the same flag.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the llvm-commits