Notify collector about lazily deoptimized code objects.

Notify collector about lazily deoptimized code objects.

All slots that were recorded on these objects during incremental marking should be ignored as they are no longer valid.

To filter such invalidated slots out during slots buffers iteration we set all markbits under the invalidated code object to 1 after the code space was swept and before slots buffers are processed.

R=erik.corry@gmail.com
BUG=v8:1713
TEST=test/mjsunit/regress/regress-1713.js

Committed: http://code.google.com/p/v8/source/detail?r=9402

[Erik Corry, 2011-09-22 09:10:47]

LGTM

http://codereview.chromium.org/7983045/diff/5001/src/mark-compact.cc
File src/mark-compact.cc (right):

http://codereview.chromium.org/7983045/diff/5001/src/mark-compact.cc#newcode2896
src/mark-compact.cc:2896: Page* p = Page::FromAddress(addr);
This doesn't work for large objects.
FromAnyPointerAddress works for large pointer objects, but not large code pages.
 Large data pages don't have slots on them.

http://codereview.chromium.org/7983045/diff/5001/src/mark-compact.cc#newcode2898
src/mark-compact.cc:2898: // First check owners identity because old pointer and
old data spaces
owners -> owner's

[Erik Corry, 2011-09-22 11:43:16]

LEBTM

http://codereview.chromium.org/7983045/diff/6001/src/mark-compact.cc
File src/mark-compact.cc (right):

http://codereview.chromium.org/7983045/diff/6001/src/mark-compact.cc#newcode2906
src/mark-compact.cc:2906: if (p->owner()->identity() != CODE_SPACE) return
false;
This will return immediately for large code pages.  Please add comment as to why
this is OK.

[Vyacheslav Egorov (Chromium), 2011-09-22 15:30:11]

I am going to commit this.

http://codereview.chromium.org/7983045/diff/8004/src/spaces.cc
File src/spaces.cc (right):

http://codereview.chromium.org/7983045/diff/8004/src/spaces.cc#newcode1936
src/spaces.cc:1936: if (!p->WasSwept()) {
The same regression test revealed a bug here.

We should not use ShouldBeSweptLazily here because it skips evacuation
candidates. 

But _new_ evacuation candidates are already selected when we call
PrepareForMarkCompact (happens only with --always-compact) and they _have_ to be
swept before the GC.