Workaround a problem where clang and gcc behavior wrt the creation

Workaround a problem where clang and gcc behavior wrt the creation of temp objects differs.
Since the ctor/dtor make pepper calls this results
in observable differences in program behavior.

The workaround is to use different lifetimes for the 
objects involved. I will leave the bug open until this is better understood.

BUG= http://code.google.com/p/nativeclient/issues/detail?id=2293

Committed: http://src.chromium.org/viewvc/chrome?view=rev&revision=103737

[robertm, 2011-09-30 19:17:06]

PTAL

[dmichael (off chromium), 2011-09-30 19:32:55]

I don't think the test should need to be changed, if you can make your
stdin/stdout files accept either valid sequence of events.

If we really believe it should be perfectly consistent at the proxy level, then
the client side of the proxy should keep its own reference count and only talk
to the browser for transitions from 0->1 or 1->0. The Chrome proxy already does
this, so we'll get it if/when we transition to the Chrome IPC proxy for NaCl.

[robertm, 2011-10-03 15:28:25]

On 2011/09/30 19:32:55, dmichael wrote:
> I don't think the test should need to be changed, if you can make your
> stdin/stdout files accept either valid sequence of events.

If we accept either sequence we will stop noticing problems like this in the
future. Since there is significant performance impact of 
having an rpc, I do not think we should go this route. 



> If we really believe it should be perfectly consistent at the proxy level,
then
> the client side of the proxy should keep its own reference count and only talk
> to the browser for transitions from 0->1 or 1->0. The Chrome proxy already
does
> this, so we'll get it if/when we transition to the Chrome IPC proxy for NaCl.

[dmichael (off chromium), 2011-10-03 16:19:11]

On 2011/10/03 15:28:25, robertm wrote:
> On 2011/09/30 19:32:55, dmichael wrote:
> > I don't think the test should need to be changed, if you can make your
> > stdin/stdout files accept either valid sequence of events.
> 
> If we accept either sequence we will stop noticing problems like this in the
> future. Since there is significant performance impact of 
> having an rpc, I do not think we should go this route. 
> 
I thought you were able to do that just by changing the stdin/stdout files. I
guess the stdout file would still differ?

This change you've made sort of hides the difference too, by making it easier
for Clang to do the elision, I guess. But that's fine.

LGTM
> 
> 
> > If we really believe it should be perfectly consistent at the proxy level,
> then
> > the client side of the proxy should keep its own reference count and only
talk
> > to the browser for transitions from 0->1 or 1->0. The Chrome proxy already
> does
> > this, so we'll get it if/when we transition to the Chrome IPC proxy for
NaCl.

[commit-bot: I haz the power, 2011-10-03 16:26:31]

CQ is trying the patch. Follow status at
https://chromium-status.appspot.com/cq/robertm@google.com/8102006/1

[commit-bot: I haz the power, 2011-10-03 17:46:40]

Change committed as 103737