1067703 - Huge regressions on misc-typedobj-write-struct-field-* tests on AWFY

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, defect, P5)

Description

[Guilherme Lima]

According to the changelog we can get on AWFY, the responsible patch is from Bug 1063488.

http://arewefastyet.com/#machine=12&view=single&suite=misc&subtest=typedobj-write-struct-field-standard
http://arewefastyet.com/#machine=12&view=single&suite=misc&subtest=typedobj-write-struct-field-typedobj
http://arewefastyet.com/#machine=11&view=single&suite=misc&subtest=typedobj-write-struct-field-standard
http://arewefastyet.com/#machine=11&view=single&suite=misc&subtest=typedobj-write-struct-field-typedobj

Reading comments 3 and 5 from bug 1063488, Brian Hackett knew about a possible regression. So this might be a Wontfix?

Comment 1

[Brian Hackett [Laid off!]]

(In reply to Guilherme Lima from comment #0)
> According to the changelog we can get on AWFY, the responsible patch is from
> Bug 1063488.
> 
> http://arewefastyet.com/#machine=12&view=single&suite=misc&subtest=typedobj-
> write-struct-field-standard
> http://arewefastyet.com/#machine=12&view=single&suite=misc&subtest=typedobj-
> write-struct-field-typedobj
> http://arewefastyet.com/#machine=11&view=single&suite=misc&subtest=typedobj-
> write-struct-field-standard
> http://arewefastyet.com/#machine=11&view=single&suite=misc&subtest=typedobj-
> write-struct-field-typedobj
> 
> Reading comments 3 and 5 from bug 1063488, Brian Hackett knew about a
> possible regression. So this might be a Wontfix?

Yes, we were optimizing this test in an unsound way, and are no longer able to hoist most of the contents of the main loop in these microbenchmarks.  I don't think this is a WONTFIX, as by making GVN and LICM smarter we should be able to hoist these accesses in a sound way.  But it would be nice to see some real code or larger benchmarks affected by this problem.

Comment 2

[Guilherme Lima]

Bug 1425687 just made the tests faster than they were before (80% improvement).