652303 - TI+JM is slower than TM on this array testcase

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Boris Zbarsky [:bzbarsky]]

Attachment: Shell testcase

See attached shell testcase.  I see times around 61ms in TM and 73ms in TI+JM.

Comment 1

[Jan de Mooij [:jandem]]

Are you on 64-bit? I see these numbers (JM tip):

32-bit:

-m -n: 59 ms
-j:    71 ms
-m:    97 ms

64-bit:

-m -n: 99 ms
-j:    82 ms
-m:    154 ms

The problem is that we only inline jsop_mod on 32-bit, on 64-bit we always call a stub.. I will file a bug, we should try to fix this in TM for Firefox 6.

Comment 2

[Jan de Mooij [:jandem]]

Filed bug 652321.

Comment 3

[Jan de Mooij [:jandem]]

With the patch in bug 652321 applied to JM:

64-bit:

-m -n: 64 ms
-j:    81 ms
-m:    101 ms

Comment 4

[Boris Zbarsky [:bzbarsky]]

> Are you on 64-bit?

Yes (note that the bug is filed on Mac).

Comment 5

[David Mandelin [:dmandelin]]

Before landing, we need to make sure that this doesn't regress Mac when TI is pref'd off. If that checks out, this can be unhooked from LandTI.

Comment 6

[Jan de Mooij [:jandem]]

This regressed terribly in the last month. From 97 ms (comment 1) to 408 ms.

parseInt returns doubles and this disables the inline paths for JSOP_MOD/JSOP_GETELEM.

Comment 7

[Jan de Mooij [:jandem]]

With the patch in bug 662766 there's a large perf win here compared to the TM-branch, even without -n. I'll close this bug after landing the patch.

Comment 8

[Jan de Mooij [:jandem]]

Some 64-bit numbers:

JM tip:

js -m -n    : 64 ms
d8          : 77 ms
js -j       : 82 ms
js -m -j -p : 82 ms
js -m       : 111 ms

TM tip:

js -j       : 144 ms
js -m -j -p : 144 ms
js -m       : 355 ms

32-bit is a bit faster (59 ms with -m -n). Not much more to do here.