1483542 - Performance regression in Octane/Box2D due to porting Compare IC to CacheIR

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, defect)

Description

[Matthew Gaudet (he/him) [:mgaudet]]

+++ This bug was initially created as a clone of Bug #1483189 +++

In Bug 1341261, the most recent CacheIR patches lead to a large (45% performance regression). This is visible on https://arewefastyet.com/#machine=29&view=single&suite=octane&subtest=Box2D. 

Bug 1483189 fixed one regression, however the fixes there appear not to resolve the Box2D regression.

Comment 1

[Matthew Gaudet (he/him) [:mgaudet]]

Bisect points at 

    
    Matthews-MacBook-Pro:build_pt7_OPT.OBJ mgaudet$ hg bisect -g
    The first bad revision is:
    changeset:   481101:87fabc832341
    user:        Matthew Gaudet <[e-mail]>
    date:        Fri Jun 08 16:46:23 2018 -0400
    summary:     Bug 1341261: [Part 6] Teach the Baseline Inspector about the new CacheIR ops r=tcampbell
    


Hilariously, I get all the perf back with this: 


    diff --git a/js/src/jit/BaselineInspector.cpp b/js/src/jit/BaselineInspector.cpp
    --- a/js/src/jit/BaselineInspector.cpp
    +++ b/js/src/jit/BaselineInspector.cpp
    @@ -567,16 +567,17 @@ CompatibleType(MCompare::CompareType fir
             return second;
    
         return first;
     }
    
     MCompare::CompareType
     BaselineInspector::expectedCompareType(jsbytecode* pc)
     {
    +    return MCompare::Compare_Unknown;
         ICStub* first = monomorphicStub(pc);

Will try to investigate more while flying.

Comment 2

[Matthew Gaudet (he/him) [:mgaudet]]

Attachment: Correct the expected compare type computation for CacheIR stubs in some cases

Comment 3

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/mozilla-inbound/rev/cbfd7962458a
Correct the expected compare type computation for CacheIR stubs in some cases r=tcampbell

Comment 4

[Eliza Balazs [:ebalazs_]]

https://hg.mozilla.org/mozilla-central/rev/cbfd7962458a