This is the mail archive of the
mailing list for the GCC project.
Re: [PATCH, GCC/LRA] Teach LRA to not use same register value for multiple output operands of an insn
- From: Thomas Preudhomme <thomas dot preudhomme at foss dot arm dot com>
- To: Mike Stump <mikestump at comcast dot net>
- Cc: gcc-patches at gcc dot gnu dot org, Vladimir Makarov <vmakarov at redhat dot com>
- Date: Thu, 14 Jul 2016 13:23:11 +0100
- Subject: Re: [PATCH, GCC/LRA] Teach LRA to not use same register value for multiple output operands of an insn
- Authentication-results: sourceware.org; auth=none
- References: <3064206.UDVhgfly4a@e108577-lin> <9A948B64-0BFE-48EA-A781-F177787174AE@comcast.net> <6390626.drAJ9tODO2@e108577-lin>
On Thursday 14 July 2016 10:32:48 Thomas Preudhomme wrote:
> On Friday 08 July 2016 09:05:55 Mike Stump wrote:
> > On Jul 8, 2016, at 8:07 AM, Thomas Preudhomme
> <firstname.lastname@example.org> wrote:
> > > While investigating the root cause a testsuite regression for the
> > > ARM/embedded-5-branch GCC in gcc.dg/vect/slp-perm-5.c, we found that the
> > > bug seems to also affect trunk.
> > Hum... If in 6.x, and safe to back port to 6, a back port would be
> > nice...
> > I use LRA in 6.x, and seems like I'd be susceptible to this sort of thing,
> > but, I didn't test it.
> It should affect 6.x as well yes since there is no special protection
> against it in the code being modified. However I only ever managed to
> reproduce this on the ARM/embedded-5-branch.
I've created PR71878 for this problem with more details on how to reproduce
the issue. it only requires to backport a single commit to gcc-5-branch, build
gcc for arm-none-eabi (make all-gcc is enough) and compile the testcase
attached in the PR with the given options.
Please let me know if you need any more details.