704174 - nsTArray: Investigate the benefits and costs of boxed mHdr

Status: RESOLVED INCOMPLETE

(Core :: XPCOM, defect)

Description

[Benoit Jacob [:bjacob] (mostly away)]

Attachment: count empty nsTArrays

nsTArray only stores a pointer |mHdr| to a Header, which contains all the data: mLength, mCapacity, mIsAutoArray, followed by the actual storage of array elements.

The benefit is that empty arrays use only sizeof(void*) bytes, as their |mHdr| pointer points to a shared sEmptyHdr object.

The possible drawbacks are:
 - more complex source code
 - more complex array indexing (have to add sizeof(Header) bytes to address array elements)
 - one more dereference needed to address mLength, mCapacity, mIsAutoArray.

Justin has commented on this: bug 677571 comment 46 and bug 677571 comment 47.

Here's a patch counting the maximum (peak) number of empty nsTArrays that we reach during Firefox's lifetime.

Some results:

empty session:                                9181
nytimes.com:                                 16056
arstechnica.com:                             19162
nytimes.com + arstechnica.com:               25120
youtube.com:                                 19403
nytimes.com + arstechnica.com + youtube.com: 35350

So the conclusion is that the peak number of empty nsTArrays is, in practice, as a very rough approximation, roughly given by this formula:

    9000 * (1 + max_number_of_nondefault_tabs)

Of course this can vary depending on what the page does. This is just out of a sample of 3 mainstream pages.

Anyway, 9000 empty nsTArrays per tab.

The mHdr boxing saves us 8 bytes per empty nsTArray, so the conclusion is that mHdr boxing saves ~ 70k per tab.

Comment 1

[Benjamin Smedberg]

Most of the source code simply dereferences mHdr without any branches, which is pretty readable and should be fast in compiled code.

Are there any next steps for this bug? 70k per tab seems significant to me.

Comment 2

[Benoit Jacob [:bjacob] (mostly away)]

Not necessarily; I just wanted it to be a place to gather data and make sure that this has been considered. Feel free to close.

Comment 3

[Nathan Froyd [:froydnj]]

Closing as per comment 2, though I think Benoit has replicated this work in other bugs. ;)