https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90151
Bug ID: 90151
Summary: 554.roms_r regression on x86_64 at -O2 and generic
march/mtune
Product: gcc
Version: 9.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: ipa
Assignee: unassigned at gcc dot gnu.org
Reporter: jamborm at gcc dot gnu.org
CC: hubicka at gcc dot gnu.org, marxin at gcc dot gnu.org
Blocks: 26163
Target Milestone: ---
Host: x86_64-linux
Target: x86_64-linux
My apologies for discovering and reporting this so late, but my
measurements show that 554.roms_r at -O2 and with generic x86_64
march/mtune regressed in performance when compiled with GCC 8 or GCC 9
compared to GCC 7. On an Intel Sandy Bridge I see 4% slowdown, on an
AMD Zen CPU I see 6.8%, LNT reports almost 4% on Kaby Lake and 5.5% on
Zen: https://lnt.opensuse.org/db_default/v4/SPEC/spec_report/branch
I have bisected this down to r254885 which is Honza's update of
profile info in IPA-CP:
Author: hubicka
Date: Fri Nov 17 17:41:10 2017
New Revision: 254885
URL: https://gcc.gnu.org/viewcvs?rev=254885&root=gcc&view=rev
Log:
* ipa-cp.c (update_profiling_info): Handle conversion to local
profile.
* tree-cfg.c (execute_fixup_cfg): Do fixup same way as inliner does.
Modified:
trunk/gcc/ChangeLog
trunk/gcc/ipa-cp.c
trunk/gcc/tree-cfg.c
I compared the number of samples in the hottest functions and arrived
at the following table:
| Object | Function | r254884 | r254885 | diff
| % |
|--------------+------------------------------------+---------+---------+-------+--------|
| roms_r_base | __step2d_mod_MOD_step2d_tile | 407180 | 409257 | 2077
| +0.51 |
| roms_r_base | __pre_step3d_mod_MOD_pre_step3d | 118119 | 134705 | 16586
| +14.04 |
| roms_r_base | __t3dmix_mod_MOD_t3dmix2 | 85499 | 101650 | 16151
| +18.89 |
| roms_r_base | __step3d_t_mod_MOD_step3d_t | 85003 | 104599 | 19596
| +23.05 |
| roms_r_base | __rho_eos_mod_MOD_rho_eos_tile | 74216 | 74746 | 530
| +0.71 |
| roms_r_base | __step3d_uv_mod_MOD_step3d_uv_tile | 66393 | 67117 | 724
| +1.09 |
| roms_r_base | __rhs3d_mod_MOD_rhs3d | 62354 | 73321 | 10967
| +17.59 |
| roms_r_base | __lmd_skpp_mod_MOD_lmd_skpp | 59767 | 70798 | 11031
| +18.46 |
| libm-2.29.so | __ieee754_exp_fma | 54324 | 56546 | 2222
| +4.09 |
| roms_r_base | __prsgrd_mod_MOD_prsgrd | 48439 | 56413 | 7974
| +16.46 |
| roms_r_base | __uv3dmix_mod_MOD_uv3dmix2 | 45255 | 52950 | 7695
| +17.00 |
| roms_r_base | __lmd_vmix_mod_MOD_lmd_vmix | 45069 | 46098 | 1029
| +2.28 |
| libm-2.29.so | __ieee754_pow_fma | 39869 | 40731 | 862
| +2.16 |
When I looked at what happens to the compilation unit with
__step3d_t_mod_MOD_step3d_t, I discovered that no IPA-CP is taking
place. In fact all IPA dumps from both revisions are exactly the
same, but profile counts of BB in tree dumps that immediately follow
are vastly different:
$ diff -u0 1/step3d_t.fppized.f90.092t.ccp2 2/step3d_t.fppized.f90.092t.ccp2 |
head -38
--- 1/step3d_t.fppized.f90.092t.ccp2 2019-04-18 18:34:39.725703893 +0200
+++ 2/step3d_t.fppized.f90.092t.ccp2 2019-04-18 18:53:23.999336873 +0200
@@ -1064 +1064 @@
- <bb 2> [local count: 8]:
+ <bb 2> [local count: 10000]:
@@ -1227 +1227 @@
- <bb 3> [local count: 4]:
+ <bb 3> [local count: 5000]:
@@ -1229 +1229 @@
- <bb 4> [local count: 8]:
+ <bb 4> [local count: 10000]:
@@ -1264 +1264 @@
- <bb 5> [local count: 4]:
+ <bb 5> [local count: 5000]:
@@ -1266 +1266 @@
- <bb 6> [local count: 8]:
+ <bb 6> [local count: 10000]:
@@ -1314 +1314 @@
- <bb 7> [local count: 4]:
+ <bb 7> [local count: 5000]:
@@ -1316 +1316 @@
- <bb 8> [local count: 8]:
+ <bb 8> [local count: 10000]:
@@ -1347 +1347 @@
- <bb 9> [local count: 4]:
+ <bb 9> [local count: 5000]:
@@ -1349 +1349 @@
- <bb 10> [local count: 8]:
+ <bb 10> [local count: 10000]:
@@ -1380 +1380 @@
- <bb 11> [local count: 4]:
+ <bb 11> [local count: 5000]:
@@ -1382 +1382 @@
- <bb 12> [local count: 8]:
+ <bb 12> [local count: 10000]:
@@ -1407 +1407 @@
- <bb 13> [local count: 4]:
+ <bb 13> [local count: 5000]:
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=26163
[Bug 26163] [meta-bug] missed optimization in SPEC (2k17, 2k and 2k6 and 95)
