452499 - TM: jit doesn't speed up this testcase for some reason

Status: RESOLVED WORKSFORME

(Core :: JavaScript Engine, defect, P3)

Description

[Boris Zbarsky [:bzbarsky]]

Attachment: js shell testcase

On the attached testcase we do no better than interpreted, at least in the JS shell.  The log says we fall off the jit on the inner for loop:

  Abort recording (line 44, pc 88): JSOP_ENTERBLOCK.

but we should still be jitting the inner loop and getting a benefit from that, no?  In any case, this testcase is a little slower in optimized JS shell with -j than without -j.

In a debug shell it's 2x slower, but that doesn't matter as much, of course.

Word of warning: this generates a _huge_ log file with TRACEMONKEY=verbose (over about 10 megs a second, and this slows it down enough that we hit 4 gigs before the script finishes).

Comment 1

[Damon Sicore (:damons)]

blocking1.9.1, P3.  Assigning to bkap.

Comment 2

[Blake Kaplan (:mrbkap) (inactive)]

For what it's worth, here are some new numbers:

mrbkap@grasshopper(...opt-obj)$ ./js ~/tmp/test.js 
Time taken: 18971

mrbkap@grasshopper(...opt-obj)$ ./js -j ~/tmp/test.js 
Time taken: 12541

so we've definitely gotten better.

Comment 3

[Blake Kaplan (:mrbkap) (inactive)]

Does this still need to block? Also, I'm not sure if I'm really the right assignee for this bug.

Comment 4

[Boris Zbarsky [:bzbarsky]]

It needed to block while -j was slower than interpreter, I think.

Comment 5

[Robert Sayre]

a similar speed increase for me, so I don't think this blocks

Comment 6

[Ryan VanderMeulen [:RyanVM]]

Current js shell:
Interp: 21016.560 ms
-j:     21446.513 ms
-m:     2939.768 ms
-m -n:  2167.218 ms

JM+TI is an order of magnitude faster than plain interp and tracejit. Seems plenty sped up for the sake of this bug!