| commit | 2f4dfc04c243fad9b331a7135c5a91aedcc6bc41 | |
| author | Guilherme Ottoni <ottoni@fb.com> | |
| Wed, 10 Jun 2015 02:07:29 +0000 (9 19:07 -0700) | ||
| committer | hhvm-bot <hhvm-bot@fb.com> | |
| Wed, 10 Jun 2015 02:32:30 +0000 (9 19:32 -0700) | ||
| tree | ebc49424bb6c4898884424ba1302bb03c3b1a430 | treesnapshot (tar.gz zip) |
| parent | 9d4e1eb536386c84dc3be8a78d7749b7bb9f28cf | commitdiff |
Retune hotcfg thresholds
Summary: These haven't been retuned since we added support for pruning region
arcs due to incompatible types. I tried a number of experiments, and
it seems that the best thing right now is to set
JitPGOMinBlockCountPercent and JitPGOMinArcProbability to zero.
Having these values be larger than zero can increase the total JITed
code size (due to more side exits), and it looks like the benefits we
saw from them before were mostly from pruning bad arcs from the
region (for which we have a better solution now).
Reviewed By: @jdelong
Differential Revision: D2141147
Summary: These haven't been retuned since we added support for pruning region
arcs due to incompatible types. I tried a number of experiments, and
it seems that the best thing right now is to set
JitPGOMinBlockCountPercent and JitPGOMinArcProbability to zero.
Having these values be larger than zero can increase the total JITed
code size (due to more side exits), and it looks like the benefits we
saw from them before were mostly from pruning bad arcs from the
region (for which we have a better solution now).
Reviewed By: @jdelong
Differential Revision: D2141147
| hphp/runtime/base/runtime-option.h | diffblobblamehistory |