Change the way we handle backward jumps in the code generator. Keep...

Change the way we handle backward jumps in the code generator.  Keep
the frame states of all the forward-entering frames that had merge
code generated.  At a backward branch, check if there is no merge code
or if we can reuse an existing merge block.

Committed: http://code.google.com/p/v8/source/detail?r=1502

[Kevin Millikin (Chromium), 2009-03-05 18:05:34]

We can do a similar thing at backward unconditional jumps.  It would reduce code
size, but I'm not sure the effect on compilation time (depending on how many
merge blocks there are to compare against (usually not many) vs. how long it
takes to compute and emit the code).

The ARM implementation will need to be updated as part of this change.  It will
be a straight port (or a bit simpler).  I'll do it tomorrow morning and send it
out for review.

[Kasper Lund, 2009-03-09 11:50:52]

LGTM. You should run some of our tests through valgrind.

[Kevin Millikin (Chromium), 2009-03-12 09:39:54]

On 2009/03/09 11:50:52, Kasper Lund wrote:
> LGTM. You should run some of our tests through valgrind.

Good idea, I will do that.  ARM port uploaded.

[Kevin Millikin (Chromium), 2009-03-12 13:49:30]

On 2009/03/12 09:39:54, Kevin Millikin wrote:
> On 2009/03/09 11:50:52, Kasper Lund wrote:
> > LGTM. You should run some of our tests through valgrind.
> 
> Good idea, I will do that.  ARM port uploaded.

Once more, with feeling.  It's passing the V8 tests under valgrind.

[Kasper Lund, 2009-03-12 14:45:00]

LGTM.

http://codereview.chromium.org/40169/diff/3015/2009
File src/jump-target.cc (right):

http://codereview.chromium.org/40169/diff/3015/2009#newcode75
Line 75: delete reaching_frames_[i];
Do you want to NULL them out too in case the JumpTarget is re-used?

[Kevin Millikin (Chromium), 2009-03-12 14:53:51]

http://codereview.chromium.org/40169/diff/3015/2009
File src/jump-target.cc (right):

http://codereview.chromium.org/40169/diff/3015/2009#newcode75
Line 75: delete reaching_frames_[i];
On 2009/03/12 14:45:00, Kasper Lund wrote:
> Do you want to NULL them out too in case the JumpTarget is re-used?

Maybe it's safer (in the face of changes).  Right now, the entire list is
cleared in the call to Reset.