llvm dev - Dec 2015 - asan for allocas on powerpc64

On 24 November 2015 at 13:06, Maxim Ostapenko
<m.ostapenko at partner.samsung.com> wrote:
> On 23/11/15 16:46, Jay Foad wrote:
>>
>> In LowerGET_DYNAMIC_AREA_OFFSET() you're calling
>> MFI->getMaxCallFrameSize(), but it looks like that doesn't
return useful
>> information until after the PrologEpilogInserter's
>> PEI::calculateCallsInformation() has run.
>>
>> So maybe the lowering has to be done as part of frame index
elimination?
>> (I'm not too familiar with this code.)
>
>
> Yeah, this sounds like a good idea. I'm attaching another draft patch
that
> tries to lower getDynamicAreaOffset during frame index elimination. There
> are still some directions for improvements, thought, such as combining
> getDynamicAreaOffset with other instructions, but anyway, could you please
> check it on your PPC box?
Yes, this fixes the test case. Thanks!

Jay.

Hi Jay, Maxim,

Thanks for working on this. Please post this patch to reviews.llvm.org for
review, but to save some time, one initial comment:

Please make the new ISD nodes Expand by default in TargetLoweringBase.cpp (in
TargetLoweringBase::initActions).

 -Hal

----- Original Message -----
> From: "Jay Foad" <jay.foad at gmail.com>
> To: "Maxim Ostapenko" <m.ostapenko at partner.samsung.com>
> Cc: "Kostya Serebryany" <kcc at google.com>, "Yury
Gribov" <y.gribov at samsung.com>, "LLVM Dev" <llvm-dev
at lists.llvm.org>,
> hfinkel at anl.gov
> Sent: Tuesday, November 24, 2015 7:32:40 AM
> Subject: Re: [llvm-dev] asan for allocas on powerpc64
> 
> On 24 November 2015 at 13:06, Maxim Ostapenko
> <m.ostapenko at partner.samsung.com> wrote:
> > On 23/11/15 16:46, Jay Foad wrote:
> >>
> >> In LowerGET_DYNAMIC_AREA_OFFSET() you're calling
> >> MFI->getMaxCallFrameSize(), but it looks like that doesn't
return
> >> useful
> >> information until after the PrologEpilogInserter's
> >> PEI::calculateCallsInformation() has run.
> >>
> >> So maybe the lowering has to be done as part of frame index
> >> elimination?
> >> (I'm not too familiar with this code.)
> >
> >
> > Yeah, this sounds like a good idea. I'm attaching another draft
> > patch that
> > tries to lower getDynamicAreaOffset during frame index elimination.
> > There
> > are still some directions for improvements, thought, such as
> > combining
> > getDynamicAreaOffset with other instructions, but anyway, could you
> > please
> > check it on your PPC box?
> 
> Yes, this fixes the test case. Thanks!
> 
> Jay.
> 
-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory

On 24 November 2015 at 22:32, Hal Finkel <hfinkel at anl.gov>
wrote:
> Please post this patch to reviews.llvm.org for review,
Maxim, Yury,

Thanks for getting D14983 committed. Are you also planning to push the
changes to lib/Transforms/Instrumentation/AddressSanitizer.cpp, to
actually use the new intrinsic?

Thanks,
Jay.