(XS) RFR: 8056183: os::is_MP() always reports true when NMT is enabled

[David Holmes]

Thanks Coleen (and Bertrand!)

I'll push once I have some final testing completed.

David

On 9/09/2014 9:58 PM, Coleen Phillimore wrote:
>
> Looks good.  Good comment.
> Coleen
>
> On 9/9/14, 1:50 AM, David Holmes wrote:
>> webrev:
>>
>> http://cr.openjdk.java.net/~dholmes/8056183/webrev/
>>
>> This is a confidential bug so here is the background:
>>
>> The NMT code uses atomics very early in the initialization process,
>> before _processor_count has been initialized, and so it would hit the
>> assertion in is_MP that _processor_count > 0. To workaround that it
>> was assumed that if NMT were present then we could just assume a MP
>> system and have is_MP return true. That change was made under 8046598.
>>
>> However that is an invalid assumption on some platforms and will lead
>> to incorrect code generation. The real fix for the original assertion
>> failure is to just remove the assertion as weakening it to check >= 0
>> would serve no point. Further we already assert that _processor_count
>> is > 0 when we actually initialize it. Having is_MP return false
>> during the early stages of initialization is okay because the VM is
>> single-threaded at that point. However, we must be vigilant to ensure
>> that we don't make any long term decisions (such as code generation
>> choices) based on is_MP during this pre-initialization phase.
>>
>> So this change reverts the change from 8046598 and removes the assertion.
>>
>> Testing: JPRT
>>
>> Pushing to hs-rt/hotspot
>>
>> Thanks,
>> David
>