[PATCH][X86] Fix assertion failure caused by a wrong folding of vector shifts by immediate count (regression introduced by my r198113).

[Andrea Di Biagio]

Committed revision 199213.

Thanks Nadav.

On Tue, Jan 14, 2014 at 1:05 AM, Nadav Rotem <nrotem at apple.com> wrote:
> LGTM.
>
> On Jan 13, 2014, at 8:24 AM, Andrea Di Biagio <andrea.dibiagio at gmail.com> wrote:
>
>> Hi,
>>
>> This patch fixes an assertion failure caused by the algorithm
>> introduced by my r198113. Revision 198113 introduced an algorithm that
>> tries to fold a vector shift by immediate count into a build_vector if
>> the input vector to the shift is a known vector of constants.
>>
>> However, the algorithm only works under the assumption that the input
>> vector type and the shift type are exactly the same.
>> This patch conservatively disables the folding of the vector shift by
>> immediate count if the input vector type and the shift valuetype are
>> not the same.
>>
>> Thanks to Patrik Hägglund H for providing the test case.
>>
>> Please let me know if ok to submit.
>>
>> Thanks,
>> Andrea Di Biagio
>> SN Systems - Sony Computer Entertainment Group
>> <patch.diff>
>