Handlify the stub cache lookup and patching for CallIC and KeyedCallIC.

Handlify the stub cache lookup and patching for CallIC and KeyedCallIC.


R=ulan@chromium.org,vegorov@chromium.org
BUG=
TEST=

Committed: http://code.google.com/p/v8/source/detail?r=9729

[Kevin Millikin (Chromium), 2011-10-19 17:30:59]

Slava, could you also take a look at this?  You can inspect the handle uses, but
I'd also like high-level comments on the direction this is going.  Thanks.

[Vyacheslav Egorov (Chromium), 2011-10-20 10:12:23]

I like the direction.

I do not have any comments except some nits.

http://codereview.chromium.org/8357010/diff/1/src/ic.cc
File src/ic.cc (right):

http://codereview.chromium.org/8357010/diff/1/src/ic.cc#newcode1897
src/ic.cc:1897: 
add empty line

http://codereview.chromium.org/8357010/diff/1/src/stub-cache.cc
File src/stub-cache.cc (right):

http://codereview.chromium.org/8357010/diff/1/src/stub-cache.cc#newcode935
src/stub-cache.cc:935: return reinterpret_cast<Code*>(code);
I wonder how this can be callled during marking phase.

Also marking phase does not corrupt maps anymore so we should convert to checked
casts here.

http://codereview.chromium.org/8357010/diff/1/src/stub-cache.cc#newcode1056
src/stub-cache.cc:1056: MaybeObject* StubCache::TryComputeCallMiss(int argc,
I wonder why ComputeCallMiss is not defined in terms on TryComputeCallMiss?

Are we going to kill TryComputeCallMiss once the stub compiler is fully
handlified? If so please state this in the comments.

http://codereview.chromium.org/8357010/diff/1/src/stub-cache.h
File src/stub-cache.h (right):

http://codereview.chromium.org/8357010/diff/1/src/stub-cache.h#newcode800
src/stub-cache.h:800: MUST_USE_RESULT MaybeObject* CompileCallField(JSObject*
object,
Should not we prefix unhandlified functions' names with Try?

[ulan, 2011-10-20 10:20:43]

LGTM

[Kevin Millikin (Chromium), 2011-10-20 10:57:08]

Thanks for the comments.

http://codereview.chromium.org/8357010/diff/1/src/ic.cc
File src/ic.cc (right):

http://codereview.chromium.org/8357010/diff/1/src/ic.cc#newcode1897
src/ic.cc:1897: 
On 2011/10/20 10:12:23, Vyacheslav Egorov wrote:
> add empty line

Done.

http://codereview.chromium.org/8357010/diff/1/src/stub-cache.cc
File src/stub-cache.cc (right):

http://codereview.chromium.org/8357010/diff/1/src/stub-cache.cc#newcode935
src/stub-cache.cc:935: return reinterpret_cast<Code*>(code);
On 2011/10/20 10:12:23, Vyacheslav Egorov wrote:
> I wonder how this can be callled during marking phase.
> 
> Also marking phase does not corrupt maps anymore so we should convert to
checked
> casts here.

Good question, and good suggestion to switch to checked casts.  I'll
investigate.

http://codereview.chromium.org/8357010/diff/1/src/stub-cache.cc#newcode1056
src/stub-cache.cc:1056: MaybeObject* StubCache::TryComputeCallMiss(int argc,
On 2011/10/20 10:12:23, Vyacheslav Egorov wrote:
> I wonder why ComputeCallMiss is not defined in terms on TryComputeCallMiss?
> 
> Are we going to kill TryComputeCallMiss once the stub compiler is fully
> handlified? If so please state this in the comments.

Yeah, Try... will be eliminated next.  I'll comment it.

For this change, I just wanted ComputeCallMiss to be able to do GC at a finer
granularity than wrapping CALL_HEAP_FUNCTION around TryComputeCallMiss (it has
two possible allocation failures).

http://codereview.chromium.org/8357010/diff/1/src/stub-cache.h
File src/stub-cache.h (right):

http://codereview.chromium.org/8357010/diff/1/src/stub-cache.h#newcode800
src/stub-cache.h:800: MUST_USE_RESULT MaybeObject* CompileCallField(JSObject*
object,
On 2011/10/20 10:12:23, Vyacheslav Egorov wrote:
> Should not we prefix unhandlified functions' names with Try?

Maybe.  They're going to be changed next, so we'll be left with the non-Try
version.  For now, we just wanted to make the smallest impact.  (The only reason
I had Try... was because I couldn't overload on the argument type there.)

[Kevin Millikin (Chromium), 2011-10-20 11:36:54]

New patch set uploaded with one significant change:

Before: we bailed out of compiling a specialized stub if the function was not
compiled and did not patch the IC (staying in the uninitialized or
premonomorphic state).  Then, we would specialize on the next call, after the
function was compiled, if the IC was ever hit again.

Now: we don't rely on this, so we can immediately specialize for lazily compiled
functions by always calling indirectly.  When Crankshaft is enabled we did this
anyway.