Change the translation of break/continue into Hydrogen.

Change the translation of break/continue into Hydrogen.

Resolve break and continue when we see them, rather then deferring
them until later.

Committed: http://code.google.com/p/v8/source/detail?r=6878

[Kevin Millikin (Chromium), 2011-02-21 15:41:07]

http://codereview.chromium.org/6541060/diff/4/src/hydrogen.cc
File src/hydrogen.cc (left):

http://codereview.chromium.org/6541060/diff/4/src/hydrogen.cc#oldcode529
src/hydrogen.cc:529: const ZoneList<BreakContinueInfo*>* infos =
break_continue_info();
Before, we kept this zone-allocated list of blocks ending with break or
continue, per subgraph.  We searched it when we built a breakable statement's
graph to find all the breaks or continues, to allocate basic blocks where
necessary, and to set join ids.

Now, break and continue is resolved immediately when seen.  An annoying wrinkle
is that we can't set the join id until after all the predecessors are added (I'd
like to get rid of that, too).

http://codereview.chromium.org/6541060/diff/4/src/hydrogen.cc#oldcode562
src/hydrogen.cc:562: body->ResolveContinue(statement);
This call was actually unnecessary, because all the callers of AppendEndless
already did it.

http://codereview.chromium.org/6541060/diff/4/src/hydrogen.cc#oldcode2116
src/hydrogen.cc:2116: old_subgraph_->AddBreakContinueInfo(subgraph_);
Here we copied all the unhandled breaks and continues to the subgraph captured
by this scope.

This was unsafe: we risked losing break or continue wherever the current
subgraph was changed without using a subgraph scope.

It was also wasteful, potentially copying a list every time we used a subgraph
scope for something other than a breakable target.

http://codereview.chromium.org/6541060/diff/4/src/hydrogen.cc
File src/hydrogen.cc (right):

http://codereview.chromium.org/6541060/diff/4/src/hydrogen.cc#newcode505
src/hydrogen.cc:505: void HSubgraph::ResolveContinue(IterationStatement*
statement,
I have tried to keep these helpers and their use as close as possible to the
original.  There is further refactoring to be done on subgraphs.

http://codereview.chromium.org/6541060/diff/4/src/hydrogen.cc#newcode526
src/hydrogen.cc:526: void HSubgraph::AppendEndless(HSubgraph* body,
Everywhere we used to call ResolveContinue or BundleBreak, we now have the block
already in hand and just pass it in.

[Mads Ager (chromium), 2011-02-22 08:18:58]

LGTM

I would merge the two classes instead of having an info class and a separate
stack structure for them.

http://codereview.chromium.org/6541060/diff/4/src/hydrogen.h
File src/hydrogen.h (right):

http://codereview.chromium.org/6541060/diff/4/src/hydrogen.h#newcode617
src/hydrogen.h:617: class BreakStackEntry BASE_EMBEDDED {
Why is the stack separate from the BreakAndContinueInfo? Can't the break and
continue infos themselves form the stack instead of having an extra structure
for that? When we use these scoping structures elsewhere in the VM we call the
SomethingOrOtherScope. We could call this a BreakAndContinueScope.

[Kevin Millikin (Chromium), 2011-02-22 08:36:35]

New change uploaded.  I renamed BreakStackEntry ==> BreakAndContinueScope.

I might be overgeneralizing at this point, but I wanted to separate the break
stack nesting from the lifetime of the info.

[Mads Ager (chromium), 2011-02-22 08:39:35]

LGTM