[PATCH] - Improve widening of 3 element binary vector operations that don't trap

[Nadav Rotem]

On Aug 19, 2013, at 2:23 PM, Redmond, Paul <paul.redmond at intel.com> wrote:

> 
> 
> On 2013-08-19 3:04 PM, "Nadav Rotem" <nrotem at apple.com> wrote:
> 
>> 
>> On Aug 19, 2013, at 11:30 AM, Redmond, Paul <paul.redmond at intel.com>
>> wrote:
>> 
>>> Hi Nadav,
>>> 
>>> On 2013-08-19 1:00 PM, "Nadav Rotem" <nrotem at apple.com> wrote:
>>> 
>>>> Hi Paul, 
>>>> 
>>>> This patch looks good, but I am a little worried about denormals.  With
>>>> this patch we will execute vector operations on garbage at the fourth
>>>> vector element.  One possible solution would be to mask out the last
>>>> element. Does that sound right to you ?
>>> 
>>> That's a good question. I don't have a good answer right now. What about
>>> handling the floating point operations in WidenVecRes_BinaryCanTrap for
>>> now? (only the integral types are widened)
>>> 
>> 
>> I am okay with this solution.
> 
> I've committed this in r188699.
> 
> However, I just looked at this code again and it seems the denormal
> problem existed before (or perhaps isn't a problem?). For example,
> consider a v3f32 add on x86: In WidenVecRes_BinaryCanTrap WidenVT will be
> v4f32 and operation will be widened (since canOpTrap() is false) There
> seems to be an assumption that unused elements are already masked out.
> 

Hi Paul, 

I looked at the code and I agree that the problem existed before this patch and I think its a bug :)   I suspect that we don’t handle denormals because no one needed to optimize non-power-of-two vector code before. 

Thanks,
Nadav    

> Thoughts?
> 
> paul
> 
>> 
>> Thanks,
>> Nadav

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20130819/26e1066b/attachment.html>