Bug ID: JDK-8194740 UseSubwordForMaxVector causes performance regression

Type: Bug
Component: hotspot
Sub-Component: compiler
Affected Version: 9,10

Priority: P3
Status: Resolved
Resolution: Fixed

Submitted: 2018-01-08
Updated: 2019-09-13
Resolved: 2018-07-09

JDK 11	JDK 12
11 b22Fixed	12Fixed

While investigating JDK-8194659, I've noticed that disabling UseSubwordForMaxVector (which is enabled by default) improves performance. 

The optimization reduces performance of the SPECjvm2008-MPEG benchmark by up to 3.5% and also affects performance of other benchmarks (see JDK-8194659 for details).

Testing passed with one timeout.
09-07-2018
I have to update change based on latest jdk/jdk11: ``` diff -r c9bf46f4d555 src/hotspot/share/opto/loopTransform.cpp --- a/src/hotspot/share/opto/loopTransform.cpp +++ b/src/hotspot/share/opto/loopTransform.cpp @@ -861,7 +861,9 @@ // Check for being too big if (body_size > (uint)_local_loop_unroll_limit) { - if ((UseSubwordForMaxVector \|\| xors_in_loop >= 4) && body_size < (uint)LoopUnrollLimit * 4) return true; + if ((cl->is_subword_loop() \|\| xors_in_loop >= 4) && body_size < (uint)LoopUnrollLimit * 4) { + return true; + } // Normal case: loop too big return false; } diff -r c9bf46f4d555 src/hotspot/share/opto/loopnode.hpp --- a/src/hotspot/share/opto/loopnode.hpp +++ b/src/hotspot/share/opto/loopnode.hpp @@ -75,7 +75,8 @@ HasRangeChecks=8192, IsMultiversioned=16384, StripMined=32768, - ProfileTripFailed=65536}; + SubwordLoop=65536, + ProfileTripFailed=131072}; char _unswitch_count; enum { _unswitch_max=3 }; char _postloop_flags; @@ -99,6 +100,7 @@ bool partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; } bool is_strip_mined() const { return _loop_flags & StripMined; } bool is_profile_trip_failed() const { return _loop_flags & ProfileTripFailed; } + bool is_subword_loop() const { return _loop_flags & SubwordLoop; } void mark_partial_peel_failed() { _loop_flags \|= PartialPeelFailed; } void mark_has_reductions() { _loop_flags \|= HasReductions; } @@ -112,6 +114,7 @@ void mark_strip_mined() { _loop_flags \|= StripMined; } void clear_strip_mined() { _loop_flags &= ~StripMined; } void mark_profile_trip_failed() { _loop_flags \|= ProfileTripFailed; } + void mark_subword_loop() { _loop_flags \|= SubwordLoop; } int unswitch_max() { return _unswitch_max; } int unswitch_count() { return _unswitch_count; } diff -r c9bf46f4d555 src/hotspot/share/opto/superword.cpp --- a/src/hotspot/share/opto/superword.cpp +++ b/src/hotspot/share/opto/superword.cpp @@ -376,6 +376,7 @@ if (same_type) { max_vector = cur_max_vector; flag_small_bt = true; + cl->mark_subword_loop(); } } } ```
09-07-2018
After discussion with Performance group we decided to accept this fix.
09-07-2018
I've submitted performance testing.
12-06-2018
Updated webrev: http://cr.openjdk.java.net/~vdeshpande/SubwordFix/webrev.01/
11-06-2018
[~kvn] SPECjvm2008 is currently broken in dev-submit due to a change in Java fonts. I'm afraid we have to wait for it to to be fixed. I will then submit testing.
11-06-2018
[~thartmann] Tobias, can you rerun performance testing with suggested fix? Thanks! Changes seems fine to me.
08-06-2018
Webrev: http://cr.openjdk.java.net/~vdeshpande/SubwordFix/webrev.00/
07-06-2018
[~vdeshpande], you've implemented JDK-8175096. Could you please have a look?
08-01-2018
ILW = Small performance regression, 3 benchmarks, -XX:-UseSubwordForMaxVector = MMM = P3
08-01-2018

Duplicate :	JDK-8194660 - SPECjvm2008-SOR.large regressions in 10-b31 on SPARC
Duplicate :	JDK-8194659 - SPECjvm2008-MPEG performance regressions in 10-b31
Relates :	JDK-8209516 - DSA generateKeyPair performance regression in jdk11b22 (SPARC)
Relates :	JDK-8175096 - Analyse subword in the loop to set maximum vector size.