Bug #99781 PGO build improvement
Submitted: 4 Jun 2020 20:02 Modified: 9 Feb 2021 7:09
Reporter: Dmitriy Philimonov Email Updates:
Status: Closed Impact on me:
None 
Category:MySQL Server: Compiling Severity:S3 (Non-critical)
Version:8.0.20 OS:Any
Assigned to: CPU Architecture:Any
Tags: Contribution, PGO

[4 Jun 2020 20:02] Dmitriy Philimonov 
Description:
During our research about applying profile guided optimizations towards MySQL Server, we indeed achieved excellent results.
We noticed that:
  1. GCC's -fprofile-update=atomic usually generates thread-safe code for multithreaded applications. Using this flag makes
     profile data more precise and reliable. Additional info could be found in the official documentation:
     https://gcc.gnu.org/onlinedocs/gcc/Instrumentation-Options.html
  2. GCC10+ -fprofile-partial-training improves performance even more (+5%, OLTP_RO, 1-4 threads) comparing with usual PGO.
  3. We detected that using ld.gold while compiling instrumented binary for aarch64 platform leads to linker internal error,
     meanwhile, the second build, which uses generated profile, is successful. At least the version 1.16 (Ubuntu 20.04) is
     broken. It makes sense to turn it off for a while.

How to repeat:
* Append recommended flags to the build scripts, try the latest stable GCC (10.1 to the date).
 * Usage of both FPROFILE_GENERATE=1 && USE_LD_GOLD on aarch64 leads to compiling error.

Suggested fix:
  1. Append -fprofile-update=atomic for GNU compiler
  2. Append -fprofile-partial-training for GNU compiler, if its version is greater than '10'
  3. Turn off USE_LD_GOLD if FPROFILE_GENERATE is turned on and system processor is aarch64
[5 Jun 2020 6:35] MySQL Verification Team 
Hello Dmitriy,

Thank you for the report and contribution.

regards,
Umesh
[2 Jul 2020 12:41] Tor Didriksen 
Posted by developer:
 
8.0.19 and up has:
Bug #96698 	Use gold as default on x86_64 only
which enables the gold linker by default for x86_64 only.
[9 Feb 2021 7:09] Paul DuBois 
Posted by developer:
 
Fixed in 8.0.24.

Compiler options for using profile guided optimization with GCC were
improved to include -fprofile-partial-training and
-fprofile-update=prefer-atomic when appropriate. Thanks to Dmitriy
Philimonov for the suggestion.