[llvm-commits] [PATCH] Adding v*i1 MVTs

[Heikki Kultala]

On 08/04/2012 12:03 AM, Hal Finkel wrote:
> On Fri, 03 Aug 2012 13:50:26 -0700
> Chris Lattner<clattner at apple.com>  wrote:
>
>>
>> On Aug 3, 2012, at 12:45 PM, Hal Finkel<hfinkel at anl.gov>  wrote:
>>
>>> Please review the attached patch which adds v*i1 MVTs. I need these
>>> for properly representing the output of vectorized selects and
>>> boolean operations on the BG/Q supercomputer (v4i1 specifically).
>>> As others have also expressed an interest in these, and in the name
>>> of limiting the number of renumbering patches, this patch adds v2i1
>>> through v64i1.
>>>
>>> Heikki, will this meet your needs as well?
>>>
>>> Micah, this patch obviously affects the same renumbering as your<2
>>> x [i8|i16]>  patch. Maybe we should combine them?
>>
>> I'm not opposed to this, but please only add the specific VTs you
>> need.  We don't want to preemptively add dead code.
>
> Fair enough.
>
> Heikki had specifically mentioned wanting v2 through v16i1. I can
> certainly stick to those (or I can just add v4i1, which I need, and
> let the TCE folks propose their own patch).

Currently TCE supports 8-wide vectors so we need v2i1, v4i1, v8i1, but 
we are planning to experiment with 16-wide vectors which would mean need 
for also v16i1, v16f32.

Adding support for just v2i1, v4i1, v8i1 will make us happy for the next 
couple of months ;)