llvm dev - Feb 2016 - Schedule pass for a subtarget?

[resending to correct llvm-dev address]

Hi,

I’d like to start enabling the loop data prefetch pass for Cyclone (an ARM64
subtarget).  AFAICT, you can’t really add passes specifically to a subtarget but
only to a target.  This makes sense because in the new world, you could have
different subtargets per function.  I just want to make sure my understanding is
correct.

My plan is to add the pass for the target and then exit the pass early based on
a TTI hook (getPrefetchDistance is not set).

Currently, PowerPC gets around this problem by checking the vendor part of the
triple to decide whether the pass should be added.  Obviously there is no
guarantee that there is a one-to-one mapping between vendor and subtarget.

Please let me know if there is more preferred way of doing this.

Thanks,
Adam

This is fine. There are other passes in aarch64 you can use as examples.

On Sat, Feb 20, 2016, 9:34 PM Adam Nemet <anemet at apple.com> wrote:
> [resending to correct llvm-dev address]
>
> Hi,
>
> I’d like to start enabling the loop data prefetch pass for Cyclone (an
> ARM64 subtarget).  AFAICT, you can’t really add passes specifically to a
> subtarget but only to a target.  This makes sense because in the new world,
> you could have different subtargets per function.  I just want to make sure
> my understanding is correct.
>
> My plan is to add the pass for the target and then exit the pass early
> based on a TTI hook (getPrefetchDistance is not set).
>
> Currently, PowerPC gets around this problem by checking the vendor part of
> the triple to decide whether the pass should be added.  Obviously there is
> no guarantee that there is a one-to-one mapping between vendor and
> subtarget.
>
> Please let me know if there is more preferred way of doing this.
>
> Thanks,
> Adam
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160221/6759af6a/attachment.html>

Thanks, Eric.  After talking to Tim, we found one somewhat similar example, the
Thumb2SizeReduction pass in ARM.   ARM schedules this pass unconditionally and
then the pass bails if the subtarget only supports Thumb1, etc.

Adam
> On Feb 20, 2016, at 10:20 PM, Eric Christopher <echristo at
gmail.com> wrote:
> 
> This is fine. There are other passes in aarch64 you can use as examples.
> 
> 
> On Sat, Feb 20, 2016, 9:34 PM Adam Nemet <anemet at apple.com
<mailto:anemet at apple.com>> wrote:
> [resending to correct llvm-dev address]
> 
> Hi,
> 
> I’d like to start enabling the loop data prefetch pass for Cyclone (an
ARM64 subtarget).  AFAICT, you can’t really add passes specifically to a
subtarget but only to a target.  This makes sense because in the new world, you
could have different subtargets per function.  I just want to make sure my
understanding is correct.
> 
> My plan is to add the pass for the target and then exit the pass early
based on a TTI hook (getPrefetchDistance is not set).
> 
> Currently, PowerPC gets around this problem by checking the vendor part of
the triple to decide whether the pass should be added.  Obviously there is no
guarantee that there is a one-to-one mapping between vendor and subtarget.
> 
> Please let me know if there is more preferred way of doing this.
> 
> Thanks,
> Adam
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20160222/d0bbfe5c/attachment.html>