https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99083
Uroš Bizjak <ubizjak at gmail dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|RESOLVED |REOPENED
Keywords| |patch
Assignee|unassigned at gcc dot gnu.org |ubizjak at gmail dot com
Resolution|FIXED |---
--- Comment #13 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Martin Jambor from comment #12)
> For the record, I have benchmarked the patches from comment #4 and comment
> #10 on top of commit 6b1633378b7 (for which I already have unpatched
> benchmark results) and the regression of 519.lbm_r compiled with -O2 LTO
> dropped from 62% to 8%.
>
> The -Ofast -march=native -flto vs. non-LTO regression also dropped from 8%
> to about 5% (GCC 10 also has non-LTO 2.5% faster than LTO, but at least both
> times improved vs. GCC 10).
>
> The only notable regression brought about the patch was 538.imagick_r when
> compiled at -Ofast -march=native without LTO, which was 6% slower with the
> patch.
>
> All of the measurements were done on a Zen2 machine.
>
> Thank you for reverting the patch, now we need to look for LNT to pick up
> the changes.
The complete patch that presumably corrects HONOR_REG_ALLOC_ORDER usage is at
[1],
but IIUC the above measurements, there is still a regression of 8% vs unpatched
compiler. With the complete patch [1], ira_better_spill_reload_regno_p change
should be a NO-OP, but the new default also disables the internal calculations
in assign_hard_reg, please see [2] for reasoning.
Based on the above benchmarks, it looks that disabling the internal
calculations in assign_hard_reg is harmful even for HONOR_REG_ALLOC_ORDER
targets, at least patched x86 compiler shows this effect. Maybe Vlad could
comment this part.
Let's reopen this PR to keep the discussions in one place.
[1] https://gcc.gnu.org/pipermail/gcc-patches/2021-February/565640.html
[2] https://gcc.gnu.org/pipermail/gcc-patches/2021-February/565699.html
