Patch to fix vselect compilation on KNL.

[Nadav Rotem]

LGTM!

On Dec 4, 2013, at 6:06 AM, Cameron McInally <cameron.mcinally at nyu.edu> wrote:

> Hey Aleksey,
> 
> On Tue, Dec 3, 2013 at 3:15 PM, Bader, Aleksey A
> <aleksey.a.bader at intel.com> wrote:
>> Hi LLVM developers,
>> 
>> 
>> 
>> I need your help with reviewing and committing the patch.
>> 
>> 
>> 
>> Build of that program:
>> 
>> define <16 x i32> @test() {
>> 
>> entry:
>> 
>>  %0 = icmp slt <16 x i32> undef, undef
>> 
>>  %1 = select <16 x i1> %0, <16 x i32> undef, <16 x i32> zeroinitializer
>> 
>>  ret <16 x i32> %1
>> 
>> }
>> 
>> Fails on KNL because of optimization in PerformSELECTCombine, which replaces
>> (x < y) ? a : 0 => (x < y) & a.
>> 
>> It seems to be profitable only if we keep comparison result in the same
>> register group as ‘a’ (i.e. for vector architectures without dedicated mask
>> registers).
>> 
>> There is no need in this optimization on KNL because vselect can be
>> implemented as masked move (single instruction).
> 
> This LGTM. I was just looking at the same issue, so thanks for this. :)
> 
> -Cameron
> 
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20131204/bf4da3aa/attachment.html>