RFR: Bulk backport to sh/jdk8u

[Aleksey Shipilev]

On 09/28/2017 11:05 AM, Andrew Haley wrote:
> On 27/09/17 20:09, Aleksey Shipilev wrote:
>> On 09/27/2017 07:33 PM, Andrew Haley wrote:
>>> 5111   // During evacuation and evacuation only, we can have the stores of cset-values
>>> 5112   // to non-cset destinations. Everything else is covered by storeval barriers.
>>> 5113   mov(tmp1, ShenandoahHeap::evacuation_in_progress_addr());
>>> 5114   ldrw(tmp1, Address(tmp1));
>>> 5115   cbnzw(tmp1, done);
>>>
>>> Looks reasonable, but in jdk8 is evacuation_in_progress not a field in the
>>> thread?
>>
>> Yes, we can poll it off TLS, like the write barrier does. But for the checking/debugging code like
>> this, it makes more sense to test the closest source of truth: the global heap variable, instead of
>> going via optimized TLS flag. If TLS flag setter fails, then prior WB/storeval silently fails, and
>> then this checking code will capture the inconsistency. IIRC, we did experience TLS setting failure
>> once, this is when we beefed up on store checks!
> 
> OK, but such subtle reasoning demands a comment in the code.

Noted for later, we would need to rehash all the checking code to capture this.

-Aleksey