RFR: Concurrent code cache evac should not update ptrs outside of safepoint

[Aleksey Shipilev]

On 07/02/2017 11:06 AM, Andrew Haley wrote:
> On 01/07/17 09:47, Roman Kennke wrote:
> 
>> Only evacuating concurrently, without patching, seems a reasonable
>> compromise for now. Aleksey suggested to record all code cache pop
>> locations, this way we could very quickly scan+update oops during
>> stw, without having to parse the code cache.
> 
> OK, I can see how that makes sense.  It's certainly attractive not to
> have to do arch-dependent tweaks: Shenandoah already has quite enough
> of those.

Yes, we don't want to have platform-specific things outside of usual
compiler-related stuff with Shenandoah. We have considered patching only aligned
code cache addresses on x86, but backed away understanding that would open the
door insomnia-inducing intermittent bugs. There are other ways to optimize this,
like the thing Roman mentioned, and we would like to try them first.

Thanks,
-Aleksey