v8: don't busy loop in cpu profiler thread - backport

[tunniclm]

As described in issue #9439, the profiler consumes 100% of a core on 0.12 when enabled via the CpuProfiler API. This was already improved in 0.10 but the fix did not carry forward correctly to 0.12 due to changes in V8.

I submitted a pull request in #9439 for a fix, but a change has now landed in V8 master for this issue and should probably be backported instead.

The V8 change is https://codereview.chromium.org/1118533003, worked under issue https://code.google.com/p/v8/issues/detail?id=3967.

[misterdjules]

@tunniclm Thank you! Could you please create a PR against the v0.12 branch that backports this change? There are instructions on how to do so in the wiki. If they're incomplete, please feel free to suggest improvements.

Adding this to the 0.12.3 milestone, and closing #9439 in favor of this one.

[tunniclm]

Note that this backport does not address the problem on Windows, where the coarse scheduler granularity makes it more difficult to make the trade-off between CPU usage and precision.

I would like to investigate some alternatives on Windows (eg: using Sleep(0), SwitchToThread(), or a more significant change involving locking)--it is unclear to me what potential scheduling delays would be introduced by each of these approaches, however.

[misterdjules]

@tunniclm Thank you for backporting the change from upstream. What do you mean by:

Note that this backport does not address the problem on Windows

? Do you mean that on Windows, the profiler still consumes 100%, or that it doesn't but the precision is not as good as before?

EDIT: Just took a look at the patch, and on Windows it is still busy looping, so I guess it still takes 100% of the CPU.

[tunniclm]

@misterdjules : Yes, on Windows the code busy loops and consumes a core.

[misterdjules]

Fixed by 80cdae8, thank you @tunniclm 👍