On 9/28/23 12:52, Vineet Gupta wrote:
On 9/28/23 05:53, Jeff Law wrote:
Vineet -- assuming Vlad's patch goes in, the other obvious candidate
for this would be the mvconst_internal define_insn_and_split where
we'd probably want to reject the insn as a whole once IRA has started.
Good point, although currently we've kind of papered over it with
-fsched-pressure, but I'm sure there are way more cases that this will
improve still.
I will spin up a full multilib test with that, hopefully with no
fallout :-)
I have the results finally. This is testsuite neutral. Same results
before/after
========= Summary of gcc testsuite =========
| # of unexpected case / # of unique
unexpected case
| gcc | g++ | gfortran |
rv32imac/ ilp32/ medlow | 168 / 70 | 13 / 6 | 67 / 12 |
rv32imafdc/ ilp32d/ medlow | 168 / 70 | 13 / 6 | 24 / 4 |
rv64imac/ lp64/ medlow | 161 / 70 | 9 / 3 | 67 / 12 |
rv64imafdc/ lp64d/ medlow | 160 / 69 | 5 / 2 | 6 / 1 |
But the SPEC runs are not affected at all, if anything it seems to be
way under noise for 5 workloads.
Not sure if we still want to add the gate, your call
511.povray_r
< 3028749801452
---
> 3028749851039
521.wrf_r
< 3897783090826
---
> 3897783089025
523.xalancbmk_r
< 1062577057227
---
> 1062577053403
527.cam4_r
< 2141572063843
---
> 2141572063800
Thx,
-Vineet
