On Wed, Aug 31, 2011 at 10:35 AM, Duncan Sands <baldr...@free.fr> wrote:
> Hi Artem,
>
> On 31/08/11 10:27, Artem Shinkarov wrote:
>>
>> On Wed, Aug 31, 2011 at 12:51 AM, Chris Lattner<clatt...@apple.com>
>> wrote:
>>>
>>> On Aug 30, 2011, at 10:01 AM, Artem Shinkarov wrote:
>>>>>>
>>>>>> The patch at the moment lacks of some examples, but mainly it works
>>>>>> fine for me. It would be nice if i386 gurus could look into the way I
>>>>>> am doing the expansion.
>>>>>>
>>>>>> Middle-end parts seems to be more or less fine, they have not changed
>>>>>> much from the previous time.
>>>>>
>>>>> +@code{__builtin_shuffle (vec, mask)} and
>>>>> +@code{__builtin_shuffle (vec0, vec1, mask)}. Both functions construct
>>>>>
>>>>> the latter would be __builtin_shuffle2.
>>>>
>>>> Why??
>>>> That was the syntax we agreed on that elegantly handles both cases in
>>>> one place.
>>>
>>> If you're going to add vector shuffling builtins, you might consider
>>> adding the same builtin that clang has for compatibility:
>>>
>>> http://clang.llvm.org/docs/LanguageExtensions.html#__builtin_shufflevector
>>>
>>> It should be straight-forward to map it into the same IR.
>>>
>>> -Chris
>>>
>>
>> Chris
>>
>> I am trying to use OpenCL syntax here which says that the mask for
>> shuffling is a vector. Also I didn't really get from the clang
>> description if the indexes could be non-constnants? If not, then I
>> have a problem here, because I want to support this.
>
> probably it maps directly to the LLVM shufflevector instruction, see
> http://llvm.org/docs/LangRef.html#i_shufflevector
> That requires the shuffle mask to be constant.
I see. I think it's not worth copying LLVM builtins that merely map
its internal IL.
Richard.
> Ciao, Duncan.
>
