MIPS: port Merge experimental/gc branch to the bleeding_edge.

MIPS: port Merge experimental/gc branch to the bleeding_edge.

Simplified based on Michael's change Refactor how embedded pointers are visited. (9597)

Ported r9328 (bdc13b7)

BUG=
TEST=

Committed: http://code.google.com/p/v8/source/detail?r=9600

[Paul Lind, 2011-10-03 05:48:47]

Hi - We finally have a working version with the new gc. The fundamental issue is
the mips use of pointers inlined in the code, in a pair of consecutive
instructions, lui & ori (like arm movt/movw).

In the old gc, all access to these inlined pointers came through
CodeIterateBody, and RelocInfo.

In the new gc the Slots and SlotsBuffer presented a problem as all pointers are
updated with indirect access (Object **). 

Our solution adds a wrapper class within SlotsBuffer that should not impose any
performance impact on the other architectures.

Open to your suggestions and better ideas, of course.

The associated change in http://codereview.chromium.org/8112008 adds mips ports
for all commits since the initial new-gc, and with that, this code passes all
tests. Because of changes since the initial new-gc, this commit will not build
without all the other one.

[danno, 2011-10-04 07:59:41]

Thanks, Paul. I am uncomfortable with the MIPS-specific #ifdefs in the shared
code. I don't have a better solution off the top of my head, so I'm adding Erik
as a GC guru to help figure out the right way to do this.

[Paul Lind, 2011-10-05 04:02:26]

Hi Erik, Danno - 

I understand your concern about the MIPS ifdefs. I've uploaded a patch from
Daniel Kalmar that cleans this up a bit, by passing a boolean (could be an emum
for readability) thru some of the Visitors, which lets us remove some ifdefs.

Another thing that we could do is substitute the pre-processor symbol
V8_TARGET_ARCH_MIPS for something a little more generic sounding like
V8_ENABLE_GC_INDIRECT_POINTERS, and then define that var within one of the mips
header files.

However, that doesn't really address the issue of ifdefs in the code. It just
hides the bad-guy a little bit.

Another thing I have prototyped, but did not upload, is moving the two
mips-specific functions SlotsBuffer::ObjectSlot::GetPointer() and
SlotsBuffer::ObjectSlot::SetPointer(Object* target) from mark-compact.cc down
into src/mips/assembler-mips-inl.h.  

I then added the generic version of these same functions to the other
architecture's assembler-xxx-inl.h, and just leave prototypes in class
ObjectSlot in mark-compact.h. All ifdefs are gone at that point. 

If this sounds reasonable, let me know and I can upload those changes.

thanks,

paul

[Paul Lind, 2011-10-11 04:23:48]

Rebased on r9574, 10-Oct. No functional changes.

[danno, 2011-10-11 16:05:07]

Paul, this is better, but perhaps it needs to be even more generalized, i.e.
switching between architectures that store object pointers inline directly and
those that encode them in some indirect way, rather than a MIPS-specific
decision.

Adding michael to review as soon as new GC lands, I'd like him to work with you
to find a reasonable solution so that we can get this landed ASAP.

[Paul Lind, 2011-10-12 04:22:11]

Danno, I generalized it a bit, hopefully in the direction you intended. Still
use an ifdef, but now on sym USE_GC_INDIRECT_POINTERS, which is defined for MIPS
only, in constants-mips.h. Moved our implementation of the ObjectSlot class
methods to assembler-mips-inl.h. Add some appropriate comments around class
ObjectSlot in mark-compact.h.

[Michael Starzinger, 2011-10-12 14:25:18]

Hello Paul! I looked at the proposed changes to ObjectSlot for the
mark-and-compact collector. I think it is better to completely overhaul the way
embedded object pointers are visited. Please see the following CL for a proposal
on how this might look like and let me know if that simplifies the MIPS port (I
think it will).

The theory is to completely get rid of RelocInfo::target_object_address(),
because MIPS cannot provide this in a sensible way. The visitors will just use
target_object() and set_target_object() to do their job. The impact on other
architectures should be minimal because it will still use a simple indirection.

http://codereview.chromium.org/8245007/

This way we can avoid #ifdefs and avoid doubling the size of every ObjectSlot,
only the ones for embedded object pointers will take two words that way.

[Paul Lind, 2011-10-13 06:56:24]

Michael, thank-you!! Your change in r9597 provides enormous simplification for
mips. None of our arch-independent changes are needed anymore, and are now
removed. I figured there had to be a simpler way, but I was not seeing it. 

I think the rest of this patch should be pretty non-controversial. 

As before, I have updated the omnibus change in
http://codereview.chromium.org/8112008/, which you can download after this, to
bring the mips port up to date with all changes since new-gc was landed. 

With this, we pass all but one test. We do fail mozilla regress-355497, which a
colleague tells me is failing on Arm also. We have not yet found the cause.

[danno, 2011-10-13 07:35:31]

lgtm