[PATCH][AArch64] Enable sign check optimization by CSE

[Jiangning Liu]

Hi Sergey,

Thanks for your update!

Unfortunately, with the latest patch, I can see "miscompare" failures for
445.gobmk and 482.sphinx3 in spec2006.

Can you verify it in your local box?

I think you can "run" them with qemu following
http://lists.cs.uiuc.edu/pipermail/llvmdev/2014-April/072085.html, if you
don't really have aarch64 hardware.

Thanks,
-Jiangning



2014-08-13 16:51 GMT+08:00 Sergey Dmitrouk <sdmitrouk at accesssoftek.com>:

> Hi Jiangning,
>
> > The 2nd compare instruction should not be under the same condition as the
> > 1st one, I think.
>
> It should for currently supported transformations, otherwise we can try
> to adjust wrong instruction.  Allowing second compare be followed by
> arbitrary instruction is a separate case that needs changes in at least
> two places (need to take into account that we can't touch it).  I can do
> this if you want to have it now, but it'll increase size of the patch
> even more, so I just added TODO item for it at the top for now.
>
> >    Can you try attached small case? With your patch, it may trigger
> >    segmentation fault.
>
> I missed handling of "cmp; ...; b.??" in adjustCmp(), it assumed the
> next instruction to be Bcc.  Thank you, I reduced that code and added
> it to tests.  I also added hidden command-line option to enable/disable
> the pass to omit recompilation just to disable it.
>
> Thanks,
> Sergey
>
> On Tue, Aug 12, 2014 at 06:20:20PM -0700, Jiangning Liu wrote:
> >    Hi Sergey,
> >    The 2nd compare instruction should not be under the same condition as
> the
> >    1st one, I think.
> >    Can you try attached small case? With your patch, it may trigger
> >    segmentation fault.
> >    Thanks,
> >    -Jiangning
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20140815/a9f2e0a0/attachment.html>