[PATCH][AVX512] Fix miscompile for unpack

[Adam Nemet]

Ping^2?

On Sep 2, 2014, at 10:55 AM, Adam Nemet <anemet at apple.com> wrote:

> Ping?
> 
> On Aug 13, 2014, at 11:48 PM, Adam Nemet <anemet at apple.com> wrote:
> 
>> Hi Elena,
>> 
>> r189189 implemented AVX512 unpack by essentially performing a 256-bit unpack
>> between the low and the high 256 bits of src1 into the low part of the
>> destination and another unpack of the low and high 256 bits of src2 into the
>> high part of the destination.
>> 
>> I don't think that's how unpack works.  AVX512 unpack simply has more 128-bit
>> lanes but other than it works the same way as AVX.  So in each 128-bit lane, we're
>> always interleaving certain parts of both operands rather different parts of
>> one of the operands.
>> 
>> E.g. for this:
>> __v16sf a = { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 };
>> __v16sf b = { 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31 };
>> __v16sf c = __builtin_shufflevector(a, b, 0, 8, 1, 9, 4, 12, 5, 13, 16,
>>                                            24, 17, 25, 20, 28, 21, 29);
>> 
>> we generated punpcklps (notice how the elements of a and b are not interleaved
>> in the shuffle).  In turn, c was set to this:
>> 
>> 0 16 1 17 4 20 5 21 8 24 9 25 12 28 13 29
>> 
>> Obviously this should have just returned the mask vector of the shuffle
>> vector.
>> 
>> I mostly reverted this change and made sure the original AVX code worked
>> for 512-bit vectors as well.
>> 
>> Also updated the tests because they matched the logic from the code.
>> 
>> Please let me know if this looks good.
>> 
>> Thanks,
>> Adam
>> 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: AVX512-Fix-miscompile-for-unpack.patch
Type: application/octet-stream
Size: 7925 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20140910/c638a8b0/attachment.obj>
-------------- next part --------------
>