Experimental: introduce a simple mechanism to allow jump targets with...

Experimental: introduce a simple mechanism to allow jump targets with
backward control flows to indicate at their binding site that only the
topmost N elements need to be mergable.  Aliasing and constants lower
in the frame will be as expected based on the forward control flows.

Use the mechanism for deferred code snippets, which all fit the simple
mechanism.

Committed: http://code.google.com/p/v8/source/detail?r=1297

[Kevin Millikin (Chromium), 2009-02-18 08:58:26]

There's lots of cheesiness here, but it's simple and works.

We should figure out a way to enforce it (in other words, catch failures early)
in debug builds.  The default argument is annoying, but fits the bill.

[Kevin Millikin (Chromium), 2009-02-18 12:04:08]

I created DeferredCode::BindExit and overloaded it for the various arg counts
(0, 1, and it turns out 3).  It seemed reasonable then to hide the exit label
from manipulation.

There are some new fiddly bits to avoid spilling the entire frame in two of the
deferred code types, so maybe you should take a quick look at the update.

[Kasper Lund, 2009-02-18 12:11:17]

LGTM.

http://codereview.chromium.org/21447/diff/32/39
File src/codegen-ia32.cc (right):

http://codereview.chromium.org/21447/diff/32/39#newcode1475
Line 1475: // generator()->frame()->SpillAll();
Remove?

http://codereview.chromium.org/21447/diff/32/33
File src/codegen.h (right):

http://codereview.chromium.org/21447/diff/32/33#newcode105
Line 105: // JumpTarget* exit() { return &exit_; }
Remove?

[Kevin Millikin (Chromium), 2009-02-18 12:26:22]

http://codereview.chromium.org/21447/diff/32/39
File src/codegen-ia32.cc (right):

http://codereview.chromium.org/21447/diff/32/39#newcode1475
Line 1475: // generator()->frame()->SpillAll();
On 2009/02/18 12:11:17, Kasper Lund wrote:
> Remove?

Yes.

http://codereview.chromium.org/21447/diff/32/33
File src/codegen.h (right):

http://codereview.chromium.org/21447/diff/32/33#newcode105
Line 105: // JumpTarget* exit() { return &exit_; }
On 2009/02/18 12:11:17, Kasper Lund wrote:
> Remove?

Yes.