Don't emit code for instructions that are hiding behind an HSoftDeoptimize

Don't emit code for instructions that are hiding behind an HSoftDeoptimize

Committed: http://code.google.com/p/v8/source/detail?r=13412

[Jakob Kummerow, 2012-12-07 14:23:44]

Please take a look.

[danno, 2013-01-10 12:30:05]

https://codereview.chromium.org/11377135/diff/10001/src/hydrogen-instructions.h
File src/hydrogen-instructions.h (right):

https://codereview.chromium.org/11377135/diff/10001/src/hydrogen-instructions...
src/hydrogen-instructions.h:1032: set_representation(Representation::Tagged());
Can you add a comment here that explains why you use HType::Smi, i.e. explain
that you want to minimize the expense of inserted changes?

https://codereview.chromium.org/11377135/diff/10001/src/hydrogen.cc
File src/hydrogen.cc (right):

https://codereview.chromium.org/11377135/diff/10001/src/hydrogen.cc#newcode1263
src/hydrogen.cc:1263: void HGraph::NullifyUnreachableInstructions() {
Can you skip this pass all together if there are no soft deopts and a situation
hasn't occurred during MarkAsDeoptimizingRecursively() where
all_predecessors_deoptimizing is true?

[Jakob Kummerow, 2013-01-16 11:35:02]

PTAL.

https://codereview.chromium.org/11377135/diff/10001/src/hydrogen-instructions.h
File src/hydrogen-instructions.h (right):

https://codereview.chromium.org/11377135/diff/10001/src/hydrogen-instructions...
src/hydrogen-instructions.h:1032: set_representation(Representation::Tagged());
On 2013/01/10 12:30:05, danno wrote:
> Can you add a comment here that explains why you use HType::Smi, i.e. explain
> that you want to minimize the expense of inserted changes?

Done.

https://codereview.chromium.org/11377135/diff/10001/src/hydrogen.cc
File src/hydrogen.cc (right):

https://codereview.chromium.org/11377135/diff/10001/src/hydrogen.cc#newcode1263
src/hydrogen.cc:1263: void HGraph::NullifyUnreachableInstructions() {
On 2013/01/10 12:30:05, danno wrote:
> Can you skip this pass all together if there are no soft deopts and a
situation
> hasn't occurred during MarkAsDeoptimizingRecursively() where
> all_predecessors_deoptimizing is true? 

I can skip the entire phase when there are no soft deopts -- done.

MarkAsDeoptimizingRecursively, on the other hand, does not currently look at
predecessors, and even if it did, that wouldn't allow us to skip the second
pass, because all_predecessors_deoptimizing is just one of the conditions that
set nullify = true.

[danno, 2013-01-16 12:20:03]

lgtm