Adapt fragmentation heuristics for over reserved pages.

Adapt fragmentation heuristics for over reserved pages.

This adapts the heuristics that detect fragmented pages to reduce memory
footprint for spaces with over reserved memory. This minimizes external
fragmentation caused by pages that cannot be released to the OS because
of just a few live objects on them.

R=erik.corry@gmail.com
TEST=cctest/test-heap/ReleaseOverReservedPages

Committed: https://code.google.com/p/v8/source/detail?r=11901

[Erik Corry, 2012-06-22 08:52:54]

LGTM with modifications

https://chromiumcodereview.appspot.com/10629004/diff/1/src/mark-compact.cc
File src/mark-compact.cc (right):

https://chromiumcodereview.appspot.com/10629004/diff/1/src/mark-compact.cc#ne...
src/mark-compact.cc:539: // about choosing pages to free.  We expect that halve
empty pages
halve empty -> half-empty

https://chromiumcodereview.appspot.com/10629004/diff/1/src/mark-compact.cc#ne...
src/mark-compact.cc:548: // try to free all mostly empty pages.  We expect that
empty pages
that empty pages -> that almost empty pages
?

https://chromiumcodereview.appspot.com/10629004/diff/1/src/mark-compact.cc#ne...
src/mark-compact.cc:552: freeness_threshold = 80;
I think we shouldn't set this so high.  We might have 1/3 over-allocation and no
pages are 80% free, but that shouldn't exclude them all from consideration.  We
pick the best 'n' below anyway.

https://chromiumcodereview.appspot.com/10629004/diff/1/test/cctest/test-heap.cc
File test/cctest/test-heap.cc (right):

https://chromiumcodereview.appspot.com/10629004/diff/1/test/cctest/test-heap....
test/cctest/test-heap.cc:1929: // Triggering subsequent GCs should cause at
least halve of the pages
halve -> half

[Michael Starzinger, 2012-06-22 09:14:23]

Comments addressed. Landing ...

https://chromiumcodereview.appspot.com/10629004/diff/1/src/mark-compact.cc
File src/mark-compact.cc (right):

https://chromiumcodereview.appspot.com/10629004/diff/1/src/mark-compact.cc#ne...
src/mark-compact.cc:539: // about choosing pages to free.  We expect that halve
empty pages
On 2012/06/22 08:52:54, Erik Corry wrote:
> halve empty -> half-empty

Done.

https://chromiumcodereview.appspot.com/10629004/diff/1/src/mark-compact.cc#ne...
src/mark-compact.cc:548: // try to free all mostly empty pages.  We expect that
empty pages
On 2012/06/22 08:52:54, Erik Corry wrote:
> that empty pages -> that almost empty pages
> ?

Done.

https://chromiumcodereview.appspot.com/10629004/diff/1/src/mark-compact.cc#ne...
src/mark-compact.cc:552: freeness_threshold = 80;
On 2012/06/22 08:52:54, Erik Corry wrote:
> I think we shouldn't set this so high.  We might have 1/3 over-allocation and
no
> pages are 80% free, but that shouldn't exclude them all from consideration. 
We
> pick the best 'n' below anyway.

Done. Reverted it back to a constant of 50 like before. Less magic constants.

https://chromiumcodereview.appspot.com/10629004/diff/1/test/cctest/test-heap.cc
File test/cctest/test-heap.cc (right):

https://chromiumcodereview.appspot.com/10629004/diff/1/test/cctest/test-heap....
test/cctest/test-heap.cc:1929: // Triggering subsequent GCs should cause at
least halve of the pages
On 2012/06/22 08:52:54, Erik Corry wrote:
> halve -> half

Done.