https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118999
Bug ID: 118999
Summary: AArch64: Switching off early scheduling causes
regressions in Snappy workload for -mcpu=neoverse-v2
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: jschmitz at gcc dot gnu.org
Target Milestone: ---
Created attachment 60573
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=60573&action=edit
Script to reproduce snappy regression
The commit that switched off early scheduling for AArch64
(https://gcc.gnu.org/g:c5db3f50bdf34ea96fd193a2a66d686401053bd2) causes changes
in performance for the Snappy workload for -mcpu=neoverse-v2, including runtime
increases of up to 20%.
In the attachment is a script to reproduce the regressions. It builds GCC from
commit c5db3f50 and runs Snappy with and without the -fschedule-insns option
(in addition to the other flags for which we saw the regression). Use it like
this:
parentdir=<path/to/gcc_src/parentdir> ./snappy_script.sh
As of today, we observed the following runtime changes (for Ofast_VLA; values
are percentages; positive values mean that running Snappy WITHOUT
-fschedule-insns has longer runtime than WITH -fschedule-insns):
BM_UFlat/5/2 -2.12766
BM_UValidate/1/1 12.9032
BM_UValidate/1/2 13.6905
BM_UValidate/2/1 8.21918
BM_UValidate/2/2 8.33333
BM_UValidate/3/1 5.88235
BM_UValidate/3/2 6.12245
BM_UValidate/5/1 12.5
BM_UValidate/5/2 6.10329
BM_UValidate/6/1 18.4906
BM_UValidate/6/2 15.8458
BM_UValidate/7/1 20.3024
BM_UValidate/7/2 16.3934
BM_UValidate/8/1 9.34066
BM_UValidate/8/2 9.49367
BM_UValidate/9/1 8.51852
BM_UValidate/9/2 9.42623
BM_UValidateMedley 2.24829
BM_UIOVecSource/6/1 3.21285
BM_UIOVecSource/7/1 4.2654
BM_UIOVecSource/11/1 2.32558
BM_UIOVecSink/0 21.1726
BM_UIOVecSink/3 4.83871
BM_UFlatSink/11/1 2.02808
BM_ZFlat/6/1 2.03252
BM_ZFlat/7/1 4.2654
In the past, we have also seen regressions in other tests, such as UFlat/3/2
and UFlat/3/1.
