77236 - Move GC to after page load (20ms out of 190ms)

Status: RESOLVED DUPLICATE of bug 42321

(Core :: DOM: Core & HTML, defect)

Description

[Hixie (not reading bugmail)]

(found by hyatt during a quantify run of google)

When loading a page, we do a GC in SetNewDocument() (in nsGlobalWindow.cpp).

We'd like to move the GC till after the onload handler fires, effectively
delaying it until we are idle (and the user is reading the page), instead of
doing it during the page loading.

The code currently says that mContext might get released while doing this (and
thus we use a kungFuDeathGrip), but just below it there is a check to see if 
mContext is still there. This seems to point to either a lurking bug or a reason
why we can't move the GC.

Any comments?

Comment 1

[David Hyatt]

Removing this actually slows things down on jrgm's tests.... so this may be an
INVALID.

Comment 2

[John Morrison]

Note: (this may not apply, if hyatt was running with the default page-to-page
(non-zero) delay value, but I thought I'd mention it anyways.) If you are 
running the test with '0' delay in page loads, then the test will never see the 
benefit of making this change. That is because I measure the "onload to onload" 
interval, minus whatever delay value is selected. 

So, if we defer a GC (assuming it is deferable) until after the onload occurs 
_and_ the delay time is greater than the time required for the GC, then we get 
the cost of the GC "for free" since it will occur while no page is actively 
being processed. However, if the delay time is '0', then there is no idle time 
during the test, and the "cost" of the GC is included in the loading time for 
each page.

Comment 3

[David Hyatt]

This is a dup of a futured jst bug.  Reassigning to him so he can dup it. (We 
force a GC during every window create/document create.)

Comment 4

[Hixie (not reading bugmail)]

John: the '0 delay' case is artificial and not found in normal browsing 
situations, so we shouldn't be concerned with whether or not this "fixes" it
for that case. :-)

Comment 5

[John Morrison]

Er, yeah. That was my point: it is artificial. Running the page loader with 
'0' delay is the wrong way to assess the benefit that a real user, who pauses
from page to page, might experience. (But I don't know which way hyatt was 
testing this, so it may still be moot).

Comment 6

[Hixie (not reading bugmail)]

Ok, cool. We're on the same page. Hyatt is indeed doing his tests with a delay,
so that should be ok.

Comment 7

[Asa Dotzler [:asa]]

Bugs targeted at mozilla1.0 without the mozilla1.0 keyword moved to mozilla1.0.1 
(you can query for this string to delete spam or retrieve the list of bugs I've 
moved)

Comment 8

[John Morrison]

jst, you did this already, right? (and it was a win).

Comment 9

[Johnny Stenback (:jst)]

Yup, I did this...

*** This bug has been marked as a duplicate of 42321 ***