Disable arm-gepopt by default

[James Molloy]

Hi Michael,

I'm not too keen on disabling this without understanding what it's doing
wrong. It seems like a large hammer.

If this is the pass I'm thinking of, it was introduced by Hao to help
remove many redundant multiplies formed by complex GEPs, and did manifest
in spec, gobmk was the main culprit iirc.

It will also affect in order cores more than out of order ones- would it be
ok to understand whether the regression is a deficiency in the pass easily
fixed or if it is something more fundamental?

James
On Fri, 17 Apr 2015 at 19:48, Michael Zolotukhin <mzolotukhin at apple.com>
wrote:

> Hi,
>
> Recently I’ve found that LLVM regressed on a TwoFish test (from geekbench)
> when comparing relatively new compiler and a year-old one. The performance
> regression is pretty big - about 20%, so I decided to investigate it
> further and understand what change exposed that. Triaging led me to r222328
> and r222331, which introduced a new ARM64-specific pass for hoisting
> constant offsets from addresses. It claimed some gains on SPECS, but when I
> turned it off, I saw no regressions there (I checked SPEC2006/ref-input
> set, built with PGO and LTO).
>
> Moreover, when the patch was committed, it effectively disabled two
> existing regression tests, which is not good I think.
>
> Is it ok to disable this pass by default and enable the tests back? We can
> turn it back on when we are confident that all tests pass, and there are no
> new regressions. Any objections?
>
> Thanks,
> Michael
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150418/b0f28286/attachment.html>