834108 - avoid O(n^2) chunk-searching in LifoAlloc::ensureUnusedApproximate

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Luke Wagner [:luke]]

Attachment: fix

IM calls ensureUnusedApproximate after all dynamic allocations (and lots of other places) to ensure there is still 16KB free in the LifoAlloc.  The problem is that, when a new chunk needs to be allocated, ensureUnusedApproximateSlow (true to its name) uses LifoAllocScope which does a LifoAlloc::release which does an O(n) search to find the chunk containing the given mark.  Thus, for n bytes of allocation, there is O(n^2) chunk-searching.

Really, all the code wants to do is keep 'latest' unchanged, so this patch does that.  This change cuts the compilation time of sqlite.js in OdinMonkey almost in half.  There seems to be a slight speedup in --v8-suite run from the SS harness.

Comment 1

[Luke Wagner [:luke]]

https://hg.mozilla.org/integration/mozilla-inbound/rev/cf6169428098

Comment 2

[Paul Wright]

Looks like this bug *might* have blown up PdfJS and, to a lesser degree, Splay on 32-bit AWFY.

Comment 3

[Paul Wright]

Don't know about 64-bit on AWFY, since it has been down for quite some time.

Comment 4

[Sean Stangl [:sstangl]]

(In reply to Paul Wright from comment #2)
> Looks like this bug *might* have blown up PdfJS and, to a lesser degree,
> Splay on 32-bit AWFY.

It appears to have been a temporary fluke.

Comment 5

[Paul Wright]

(In reply to Sean Stangl from comment #4)
> (In reply to Paul Wright from comment #2)
> > Looks like this bug *might* have blown up PdfJS and, to a lesser degree,
> > Splay on 32-bit AWFY.
> 
> It appears to have been a temporary fluke.

Yep, false alarm... My bad.  The 64-bit issue is real, though.

Comment 6

[Ryan VanderMeulen [:RyanVM]]

https://hg.mozilla.org/mozilla-central/rev/cf6169428098