[llvm] r248250 - [X86][SSE] Match zero/any extension shuffles that don't start from the first element

[Jeroen Ketema via llvm-commits]

Hi Simon,

Thanks for your quick fix! I can confirm that this fixes the issues I 
was as seeing. I'll have a look at the AVX2 issues I was seeing to see 
if they're still there and create tickets for those.

Thanks,

  Jeroen

On 24/09/15 20:37, Simon Pilgrim wrote:
> Thanks Jeroen - I have a candidate fix for the shuffle you mentioned and will commit shortly - if you have other shuffle bugs (whether you think they are related to this patch or not) please can you raise them on bugzilla and CC me?
>
> Simon.
>
>> On 24 Sep 2015, at 16:02, Jeroen Ketema <jeroen at codeplay.com> wrote:
>>
>>
>> Hi Simon,
>>
>> I also see failures when I enable AVX2, but it might be that these are not be related to your changes. I hadn't run with AVX2 enabled for a long time. It might be worth looking in to these issues once the below issue is fixed.
>>
>> Jeroen
>>
>> On 24/09/15 09:44, Simon Pilgrim wrote:
>>> Looking at this now - was this targeting AVX2 or something earlier?
>>>
>>>> On 23 Sep 2015, at 17:29, Jeroen Ketema <jeroen at codeplay.com> wrote:
>>>>
>>>>
>>>> Hi Simon,
>>>>
>>>> This change breaks a number of shuffles for me. For example,
>>>>
>>>> shufflevector <16 x i16> %0, <16 x i16> undef, <16 x i32> <i32 10, i32 3, i32 13, i32 3, i32 8, i32 14, i32 13, i32 14, i32 5, i32 7, i32 12, i32 2, i32 13, i32 6, i32 3, i32 6>
>>>>
>>>> Now shuffles to
>>>>
>>>> 10, 3, 13, 3, 8, 14, 13, 14, 5, 7, 10, 2, 11, 6, 3, 6
>>>>
>>>> so the 10th and 12th element are wrong.
>>>>
>>>> Best,
>>>>
>>>> Jeroen
>>>>
>>>> On 22/09/15 09:16, llvm-commits at lists.llvm.org (Simon Pilgrim via llvm-commits) wrote:
>>>>> Author: rksimon
>>>>> Date: Tue Sep 22 03:16:08 2015
>>>>> New Revision: 248250
>>>>>
>>>>> URL: http://llvm.org/viewvc/llvm-project?rev=248250&view=rev
>>>>> Log:
>>>>> [X86][SSE] Match zero/any extension shuffles that don't start from the first element
>>>>>
>>>>> This patch generalizes the lowering of shuffles as zero extensions to allow extensions that don't start from the first element. It now recognises extensions starting anywhere in the lower 128-bits or at the start of any higher 128-bit lane.
>>>>>
>>>>> The motivation was to reduce the number of high cost pshufb calls, but it also improves the SSE2 case as well.
>>>>>
>>>>> Differential Revision: http://reviews.llvm.org/D12561
>>>>>
>>>>> Modified:
>>>>>      llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
>>>>>      llvm/trunk/test/CodeGen/X86/machine-cp.ll
>>>>>      llvm/trunk/test/CodeGen/X86/vec_cast2.ll
>>>>>      llvm/trunk/test/CodeGen/X86/vec_int_to_fp.ll
>>>>>      llvm/trunk/test/CodeGen/X86/vector-sext.ll
>>>>>      llvm/trunk/test/CodeGen/X86/vector-shuffle-256-v16.ll
>>>>>      llvm/trunk/test/CodeGen/X86/vector-shuffle-256-v32.ll
>>>>>      llvm/trunk/test/CodeGen/X86/vector-shuffle-sse4a.ll
>>>>>      llvm/trunk/test/CodeGen/X86/vector-zext.ll
>>>>>