[Bug target/107093] AVX512 mask operations not simplified in fully masked loop

["rguenther at suse dot de" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107093

--- Comment #9 from rguenther at suse dot de <rguenther at suse dot de> ---
On Tue, 11 Oct 2022, crazylht at gmail dot com wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107093
> 
> --- Comment #8 from Hongtao.liu <crazylht at gmail dot com> ---
> 
> > 
> > One downside for a fully masked body is that we're using masked stores
> > which usually have higher latency due to the "merge" semantics which
> > means an extra memory input + merge operation.  Not sure if modern
> > uArchs can optimize the all-ones mask case, the vectorizer, for
> Also I guess mask store won't be store forward even load is inside the mask
> store.

I guess the masking of the store is resolved in the load-store unit
and not by splitting the operation into a load, modify, store because
that cannot easily hide exceptions.  So yes, a masked store in the
store buffer likely cannot act as forwarding source (though the
actual mask should be fully resolved there) since the actual
merging will take place later.