This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH 2/2][vect]Make vect-epilogues-nomask=1 default

On 8/23/19 10:50 AM, Andre Vieira (lists) wrote:
> I am not sure whether this is the right thing to do, we might want to
> enable this only for certain targets and if so, I also don't know how to
> tackle the testing issue where all the tests that check the number of
> times a dump mentions something will now depend on whether that test
> enables epilogue vectorization or not. I was trying to avoid disabling
> the option for all tests, since if we do that then we aren't really
> testing it. Maybe we ought to enable the option for all tests,
> regardless of the targets default. For testing purposes.
In general we try to drive things with proper costing rather than
enabling/disabling transformations based on the target.  If we need to
adjust cost models, then we're better off getting the bits in now to
give developers a chance to make those adjustments.

WRT testing.  One approach would be to somehow indicate in the dump file
that we're dealing with an epilogue by prefixing the messages.  THen the
scanner could distinguish between main loop vectorization and epilogue
vectorization.  I'm not sure how invasive that would be.

I think the approach you've taken here is fine for now.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]