2016-06-16 9:26 GMT+03:00 Jeff Law <l...@redhat.com>:
> On 06/15/2016 05:22 AM, Richard Biener wrote:
>>
>>
>> You look at TREE_TYPE of LOOP_VINFO_NITERS (loop_vinfo) - I don't think
>> this is meaningful (if then only by accident). I think you should look at
>> the
>> control IV itself, possibly it's value-range, to determine the smallest
>> possible
>> type to use.
>
> Can we get an IV that's created after VRP? If so, then we have to be
> prepared for the case where there's no range information on the IV. At
> which point I think using type min/max of the IV is probably the right
> fallback. But I do think we should be looking at range info much more
> systematically.
>
> I can't see how TREE_TYPE of the NITERS makes sense either.
I need to build a vector {niters, ..., niters} and compare to it. Why doesn't
it make sense to choose the same type for IV? I agree that choosing a smaller
type may be beneficial. Shouldn't I look at nb_iterations_upper_bound then
to check if NITERS can be casted to a smaller type?
Thanks,
Ilya
>
>> Finally we have a related missed optimization opportunity, namely avoiding
>> peeling for gaps if we mask the last load of the group (profitability
>> depends
>> on the overhead of such masking of course as it would be done in the main
>> vectorized loop).
>
> I think that's a specific instance of a more general question -- what
> transformations can be avoided by masking and can we generate costs to
> select between those transformations and masking. Seems like a follow-up
> item rather than a requirement for this work to go forward to me.
>
> Jeff
