514066 - TM: Make short functions in LIR.cpp inline [nanojit]

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Andreas Gal :gal]

A bunch of short methods of LIns are implemented in LIR.cpp. We should make those inline. The prolog/epilog overhead is killing us according to shark.

Comment 1

[Nicholas Nethercote [inactive]]

I've seen it for imm64f(), maybe imm64(), not sure about others.

Comment 2

[Andreas Gal :gal]

Functions that contain less code than the call to them should generally be inline (return oprnd1() i.e.)

Comment 3

[Nicholas Nethercote [inactive]]

This depends on bug 512824 because the patch for bug 512824 inlines some of these functions, so work on this should wait until that's landed to avoid conflicts.

Comment 4

[Nicholas Nethercote [inactive]]

Attachment: patch

This patch inlines all remaining functions in class LIns except for staticSanityCheck() and isFloat().  isFloat() contains a switch, and is currently unused in TM anyway.

Stats:

- 'js' code size: 1,430,580 bytes -> 1,426,184 bytes

- instructions executed for 3d-raytrace: 244,466,449 -> 242,067,356

- SunSpider overall results:
  ** TOTAL **: 1.006x as fast  897.7ms +/- 0.1%  892.1ms +/- 0.1%   significant

Inlining FTW!

Comment 5

[Andreas Gal :gal]

Comment on attachment 399646 [details] [diff] [review]
patch

Cool.

Comment 6

[Nicholas Nethercote [inactive]]

http://hg.mozilla.org/tracemonkey/rev/c253aa1674c3

Comment 7

[Nicholas Nethercote [inactive]]

Follow-up bustage fix:

http://hg.mozilla.org/tracemonkey/rev/0bca8ad422bb

Comment 8

[Blake Kaplan (:mrbkap) (inactive)]

http://hg.mozilla.org/mozilla-central/rev/c253aa1674c3

Comment 9

[Blake Kaplan (:mrbkap) (inactive)]

http://hg.mozilla.org/mozilla-central/rev/0bca8ad422bb

Comment 10

[Robert Sayre]

http://hg.mozilla.org/releases/mozilla-1.9.2/rev/0c5d2f477c71