Turn propagate deoptimizing mark into a proper HPhase.

Turn propagate deoptimizing mark into a proper HPhase.

Also turn the recursion on the domination chain into a loop with
an explicit stack, to avoid possible stack overflow here.

R=dslomov@chromium.org

Committed: https://code.google.com/p/v8/source/detail?r=15660

[Dmitry Lomov (no reviews), 2013-07-15 07:51:26]

lgtm with a comment

https://codereview.chromium.org/19150002/diff/1/src/hydrogen-deoptimizing-mark.h
File src/hydrogen-deoptimizing-mark.h (right):

https://codereview.chromium.org/19150002/diff/1/src/hydrogen-deoptimizing-mar...
src/hydrogen-deoptimizing-mark.h:47: INLINE(void MarkAsDeoptimizing());
Why INLINE? Looks like those functions are not hot at all. 
Let's remove INLINE unless I am missing something.

[Dmitry Lomov (no reviews), 2013-07-15 09:13:54]

On 2013/07/15 07:51:26, Dmitry Lomov (chromium) wrote:
> lgtm with a comment
> 
>
https://codereview.chromium.org/19150002/diff/1/src/hydrogen-deoptimizing-mark.h
> File src/hydrogen-deoptimizing-mark.h (right):
> 
>
https://codereview.chromium.org/19150002/diff/1/src/hydrogen-deoptimizing-mar...
> src/hydrogen-deoptimizing-mark.h:47: INLINE(void MarkAsDeoptimizing());
> Why INLINE? Looks like those functions are not hot at all. 
> Let's remove INLINE unless I am missing something.

Oops not lgtm yet.

[Dmitry Lomov (no reviews), 2013-07-15 09:14:27]

https://codereview.chromium.org/19150002/diff/1/src/hydrogen-deoptimizing-mar...
File src/hydrogen-deoptimizing-mark.cc (right):

https://codereview.chromium.org/19150002/diff/1/src/hydrogen-deoptimizing-mar...
src/hydrogen-deoptimizing-mark.cc:36: for (;;) {
Can we make the terminating condition for this loop more clear? block here is
"inlined" top of the stack, so how about having 'block != NULL' as a loop guard?

[Benedikt Meurer, 2013-07-15 09:22:53]

PTAL

https://codereview.chromium.org/19150002/diff/1/src/hydrogen-deoptimizing-mar...
File src/hydrogen-deoptimizing-mark.cc (right):

https://codereview.chromium.org/19150002/diff/1/src/hydrogen-deoptimizing-mar...
src/hydrogen-deoptimizing-mark.cc:36: for (;;) {
On 2013/07/15 09:14:27, Dmitry Lomov (chromium) wrote:
> Can we make the terminating condition for this loop more clear? block here is
> "inlined" top of the stack, so how about having 'block != NULL' as a loop
guard?

Done.

https://codereview.chromium.org/19150002/diff/1/src/hydrogen-deoptimizing-mark.h
File src/hydrogen-deoptimizing-mark.h (right):

https://codereview.chromium.org/19150002/diff/1/src/hydrogen-deoptimizing-mar...
src/hydrogen-deoptimizing-mark.h:47: INLINE(void MarkAsDeoptimizing());
On 2013/07/15 07:51:26, Dmitry Lomov (chromium) wrote:
> Why INLINE? Looks like those functions are not hot at all. 
> Let's remove INLINE unless I am missing something.

Done.

[Dmitry Lomov (no reviews), 2013-07-15 09:41:42]

https://codereview.chromium.org/19150002/diff/7001/src/hydrogen-deoptimizing-...
File src/hydrogen-deoptimizing-mark.cc (right):

https://codereview.chromium.org/19150002/diff/7001/src/hydrogen-deoptimizing-...
src/hydrogen-deoptimizing-mark.cc:44: for (int i = dominated_blocks->length();
--i > 0; ) {
Ah comparison with side effects - my eyes hurt!
Can we pretty please have 
 for (int i = dominated_blocks->length() - 1; i > 0; i--) 
or maybe even:
 block = dominated_blocks->at(0);
 for (int i = 1; i < dominated_blocks->length(); i++)
The latter form makes it really clear that we are skipping zero block
(it does not matter in which order we push them on the stack right?)

[Benedikt Meurer, 2013-07-15 09:45:44]

:-)

https://codereview.chromium.org/19150002/diff/7001/src/hydrogen-deoptimizing-...
File src/hydrogen-deoptimizing-mark.cc (right):

https://codereview.chromium.org/19150002/diff/7001/src/hydrogen-deoptimizing-...
src/hydrogen-deoptimizing-mark.cc:44: for (int i = dominated_blocks->length();
--i > 0; ) {
On 2013/07/15 09:41:42, Dmitry Lomov (chromium) wrote:
> Ah comparison with side effects - my eyes hurt!
> Can we pretty please have 
>  for (int i = dominated_blocks->length() - 1; i > 0; i--) 
> or maybe even:
>  block = dominated_blocks->at(0);
>  for (int i = 1; i < dominated_blocks->length(); i++)
> The latter form makes it really clear that we are skipping zero block
> (it does not matter in which order we push them on the stack right?)

Done.

[Dmitry Lomov (no reviews), 2013-07-15 09:52:21]

lgtm - thanks!

[Benedikt Meurer, 2013-07-15 09:53:06]

Committed patchset #3 manually as r15660 (presubmit successful).