As a last resort try to allocate out of the smaller size size-class in

As a last resort try to allocate out of the smaller size size-class in
the free list allocator.

BUG=
R=mstarzinger@chromium.org

Committed: https://code.google.com/p/v8/source/detail?r=15146

[Hannes Payer (out of office), 2013-06-14 09:03:08]

This change will enable us to use 8K of memory for the first page of the
property cell space.

[danno, 2013-06-14 09:23:36]

fundamentally ok, but I think Michael should give final approval

[Michael Starzinger, 2013-06-14 09:44:53]

https://codereview.chromium.org/17058002/diff/1/src/spaces.cc
File src/spaces.cc (right):

https://codereview.chromium.org/17058002/diff/1/src/spaces.cc#newcode2147
src/spaces.cc:2147: return NULL;
This is not necessarily correct. The check only verifies that the top-most
free-list entry is big enough. If the top-most free-list entry happens to be on
an evacuation candidate we cannot allocate it and will take a subsequent entry
which in turn might be smaller.

As discussed offline: Let's either duplicate the above functionality and do the
size check after evacuation candidates have been skipped, or call
PichNodeFromList and return the entry back if size doesn't fit.

https://codereview.chromium.org/17058002/diff/1/src/spaces.cc#newcode2245
src/spaces.cc:2245: if ((owner_->ConstantAllocationSize() && size_in_bytes <=
kSmallListMax) ||
This change supersedes the previous optimization to also look at the smallest
bucket for constant-size spaces. Let's take out this optimization (plus the
ConstantAllocationSize predicate) again.

[Hannes Payer (out of office), 2013-06-14 10:40:54]

https://codereview.chromium.org/17058002/diff/1/src/spaces.cc
File src/spaces.cc (right):

https://codereview.chromium.org/17058002/diff/1/src/spaces.cc#newcode2147
src/spaces.cc:2147: return NULL;
On 2013/06/14 09:44:53, Michael Starzinger wrote:
> This is not necessarily correct. The check only verifies that the top-most
> free-list entry is big enough. If the top-most free-list entry happens to be
on
> an evacuation candidate we cannot allocate it and will take a subsequent entry
> which in turn might be smaller.
> 
> As discussed offline: Let's either duplicate the above functionality and do
the
> size check after evacuation candidates have been skipped, or call
> PichNodeFromList and return the entry back if size doesn't fit.

Done. I ended up not duplicating the code, since I do not think we will go to
the linear size allocation approach anytime soon. Given the block back using the
existing API is cleaner.

https://codereview.chromium.org/17058002/diff/1/src/spaces.cc#newcode2245
src/spaces.cc:2245: if ((owner_->ConstantAllocationSize() && size_in_bytes <=
kSmallListMax) ||
On 2013/06/14 09:44:53, Michael Starzinger wrote:
> This change supersedes the previous optimization to also look at the smallest
> bucket for constant-size spaces. Let's take out this optimization (plus the
> ConstantAllocationSize predicate) again.

Done.

[Michael Starzinger, 2013-06-14 11:13:05]

LGTM with a nit. Feel free to add the additional assertions we discussed
offline.

https://codereview.chromium.org/17058002/diff/8001/src/spaces.cc
File src/spaces.cc (right):

https://codereview.chromium.org/17058002/diff/8001/src/spaces.cc#newcode2121
src/spaces.cc:2121: available_ -= reinterpret_cast<FreeSpace*>(node)->Size();
The cast seems obsolete.

https://codereview.chromium.org/17058002/diff/8001/src/spaces.cc#newcode2127
src/spaces.cc:2127: *node_size = reinterpret_cast<FreeSpace*>(node)->Size();
Likewise.

[Hannes Payer (out of office), 2013-06-14 11:18:25]

https://codereview.chromium.org/17058002/diff/8001/src/spaces.cc
File src/spaces.cc (right):

https://codereview.chromium.org/17058002/diff/8001/src/spaces.cc#newcode2121
src/spaces.cc:2121: available_ -= reinterpret_cast<FreeSpace*>(node)->Size();
On 2013/06/14 11:13:05, Michael Starzinger wrote:
> The cast seems obsolete.

The free space map is not done at deserialization time.

https://codereview.chromium.org/17058002/diff/8001/src/spaces.cc#newcode2127
src/spaces.cc:2127: *node_size = reinterpret_cast<FreeSpace*>(node)->Size();
On 2013/06/14 11:13:05, Michael Starzinger wrote:
> Likewise.

The free space map is not done at deserialization time.

[Hannes Payer (out of office), 2013-06-14 11:22:30]

Committed patchset #5 manually as r15146 (presubmit successful).