1029422 - [PGO]Text rendering 50% slowdown on PGO build, regardless of font-family

Status: RESOLVED WORKSFORME

(Core :: Web Painting, defect)

Description

[Alice0775 White]

Attachment: index.html using "MS PGothic"

Build Identifier:
https://hg.mozilla.org/mozilla-central/rev/366b5c0c02d3
Mozilla/5.0 (Windows NT 6.1; WOW64; rv:33.0) Gecko/20100101 Firefox/33.0 ID:20140623030201

Steps To Reproduce:
1. Make sure "MS PGothic" is in your PC
2. Open attached

Actual Results:
Nightly 33.0a1 is 50 % slower than Aurora32.0a2

Expected Results:
Should be same as Aurora32.0a2

Comment 1

[Alice0775 White]

This problem seems to happen only in PGO build(regular nightly and PGO build of tinderbox).

Regression window(m-i PGO build)
Good about 20fps:
https://hg.mozilla.org/integration/mozilla-inbound/rev/bff872c9d4b2
Mozilla/5.0 (Windows NT 6.1; WOW64; rv:33.0) Gecko/20100101 Firefox/33.0 ID:20140619013007
Bad about 10 fps:
https://hg.mozilla.org/integration/mozilla-inbound/rev/e589c195f61d
Mozilla/5.0 (Windows NT 6.1; WOW64; rv:33.0) Gecko/20100101 Firefox/33.0 ID:20140619043005
Pushlog:
http://hg.mozilla.org/integration/mozilla-inbound/pushloghtml?fromchange=bff872c9d4b2&tochange=e589c195f61d

Comment 2

[Alice0775 White]

Attachment: index.html using Arial

Comment 3

[Alice0775 White]

This problem seems to happen regardless of font-family.

Comment 4

[Alice0775 White]

Suspect
	4df3e4664b11	Jan de Mooij — Bug 1026438 part 5 - Make irregexp work with Latin1 strings.

Comment 5

[Jan de Mooij [:jandem]]

(In reply to Alice0775 White from comment #4)
> Suspect
> 	4df3e4664b11	Jan de Mooij — Bug 1026438 part 5 - Make irregexp work with
> Latin1 strings.

I really doubt it; most of the changes I'm making won't have any effect until we turn on Latin1 strings by default.

Also, a profile on OS X shows most of the time here is in graphics, and the website doesn't use regular expressions at all AFAICS. Nicolas, the only graphics-related bug I see in that range is bug 1027365, could it be related somehow?

If needed we can do some PGO Try pushes to bisect this, but it will take a while.

Comment 6

[Alice0775 White]

:jandem
remove blocking bug #..

Comment 7

[Nicolas Silva [:nical]]

(In reply to Jan de Mooij [:jandem] from comment #5)
> (In reply to Alice0775 White from comment #4)
> Also, a profile on OS X shows most of the time here is in graphics, and the
> website doesn't use regular expressions at all AFAICS. Nicolas, the only
> graphics-related bug I see in that range is bug 1027365, could it be related
> somehow?

Bug 1027365 only affects video compositing, and more importantly it doesn't change anything on the configurations we already ship.

Comment 8

[Alice0775 White]

I would like to request to PGO try build of each changesets to find out regression.
[e-mail]
[e-mail]
[e-mail]
[e-mail]
[e-mail]

Comment 9

[Alice0775 White]

and [e-mail]

Comment 10

[Jonathan Watt [:jwatt]]

My commit (b1f5029982f2) definitely could not have changed the performance.

Comment 11

[Jan de Mooij [:jandem]]

Alice, I started a Win32 PGO build for revision b1f5029982f2:

https://tbpl.mozilla.org/?tree=Try&rev=163a89a8037a

PGO builds take a long time, but when it's done can you try it? Then we can narrow it down further.

Comment 12

[Brian Smith (:briansmith, :bsmith, use NEEDINFO?)]

My changes didn't have anything to do with text rendering or graphics or anything except networking.

Comment 13

[Makoto Kato [:m_kato]]

My landing is to fix compiler warning.  If slow by me, JS benchmark becomes slow too.

But since this depends on PGO, we may have to get all data per commit.

Comment 14

[Nicolas B. Pierron [:nbp]]

Both patches should be on Aurora (which is making PGO builds), can we reproduce this on Aurora after these patches?  My guess is that they are extremely unlikely to cause any performance issue.

d2152031a5d1 Bug 1022232
2bc6f38f9cff Bug 1011745

Comment 15

[Jan de Mooij [:jandem]]

Did you have a chance to test the PGO build in comment 11?

Comment 16

[Alice0775 White]

The tryserver build fixes the problem. :)

https://hg.mozilla.org/try/rev/163a89a8037a
Mozilla/5.0 (Windows NT 6.1; WOW64; rv:33.0) Gecko/20100101 Firefox/33.0 ID:20140624110337

Comment 17

[Alice0775 White]

Today's Nightly33.0a1 seemed to recover to the level of the previous performance.
https://hg.mozilla.org/mozilla-central/rev/83c09fe3a658
Mozilla/5.0 (Windows NT 6.1; WOW64; rv:33.0) Gecko/20100101 Firefox/33.0 ID:20140701030202


Progression window(m-i PGO)
Bad(<10fps):
https://hg.mozilla.org/integration/mozilla-inbound/rev/58bf8348c8f1
Mozilla/5.0 (Windows NT 6.1; WOW64; rv:33.0) Gecko/20100101 Firefox/33.0 ID:20140630043006
Good(>20fps):
https://hg.mozilla.org/integration/mozilla-inbound/rev/b9954289176d
Mozilla/5.0 (Windows NT 6.1; WOW64; rv:33.0) Gecko/20100101 Firefox/33.0 ID:20140630073013
Progression pushlog:
http://hg.mozilla.org/integration/mozilla-inbound/pushloghtml?fromchange=58bf8348c8f1&tochange=b9954289176d

Comment 18

[Alice0775 White]

It seems that Bug1028866 fixed the performance regression of PGO.

Comment 19

[Alice0775 White]

The PGO regression appears again in tinderbox-builds/mozilla-inbound-win32-pgo build.

Regression window
Good:
https://hg.mozilla.org/integration/mozilla-inbound/rev/ab2b0fb5cfff
Mozilla/5.0 (Windows NT 6.1; WOW64; rv:33.0) Gecko/20100101 Firefox/33.0 ID:20140714103003
Bad:
https://hg.mozilla.org/integration/mozilla-inbound/rev/0f25f3ccb7b3
Mozilla/5.0 (Windows NT 6.1; WOW64; rv:33.0) Gecko/20100101 Firefox/33.0 ID:20140714133007

Pushlog:
http://hg.mozilla.org/integration/mozilla-inbound/pushloghtml?fromchange=ab2b0fb5cfff&tochange=0f25f3ccb7b3

Suspected: 
0f25f3ccb7b3	Jan de Mooij — Bug 1037869 - Fix remaining Latin1 string issues. r=terrence

Comment 20

[Alice0775 White]

Attachment: bug1029422en.html

In my slow pc,  the bad build is 50% slower than Good build.

Steps To Reproduce:
1. Open attached
2. Wait until an alert box pops up

Actual Results:
38000 msec

Expected Results:
17000 msec

Comment 21

[Jan de Mooij [:jandem]]

Alice, would you mind trying this build:

http://ftp.mozilla.org/pub/mozilla.org/firefox/try-builds/[e-mail]-8e89e911813c/try-win32/

What numbers do you get? (This is with the PGO patch in bug 1030706.)

Comment 22

[Alice0775 White]

(In reply to Jan de Mooij [:jandem] from comment #21)
> Alice, would you mind trying this build:
> 
> http://ftp.mozilla.org/pub/mozilla.org/firefox/try-builds/jandemooij@gmail.
> com-8e89e911813c/try-win32/
> 
> What numbers do you get? (This is with the PGO patch in bug 1030706.)

I got 17354 ms
The try build fixes the performance problem :)

https://hg.mozilla.org/try/rev/8e89e911813c
Mozilla/5.0 (Windows NT 6.1; WOW64; rv:33.0) Gecko/20100101 Firefox/33.0 ID:20140715031024

Comment 23

[Jan de Mooij [:jandem]]

(In reply to Alice0775 White from comment #22)
> I got 17354 ms
> The try build fixes the performance problem :)

Thanks for checking!

Comment 25

[Alice0775 White]

The performance is back.
I cannot reproduce the performance problem in the following today's Nightly33.0a1.

https://hg.mozilla.org/mozilla-central/rev/a74600665875
Mozilla/5.0 (Windows NT 6.1; WOW64; rv:33.0) Gecko/20100101 Firefox/33.0 ID:20140717030339