RFR: JDK-8277822: Remove debug-only heap overrun checks in os::malloc and friends [v3]

[David Holmes]

On Tue, 18 Jan 2022 07:09:08 GMT, David Holmes <dholmes at openjdk.org> wrote:

>> Hi David,
>> 
>> I understand your concern. I have had this patch in our internal tests enabled for several months now, but am fine with you guys doing more tests.
>> 
>> My arguments for this:
>> 
>> 1) using NMT in debug builds in place of the old hard wired fences should be less overhead, not more. NMT is quite frugal. NMT=summary does:
>>   - increase the counter per NMT category (basically atomic-incing a number in a static array)
>>   - add and populate a malloc header and a footer
>> That's it. We did the same thing before. The old code atomically increased two global counters (`inc_stat_counter`), then added a header and a footer. Only the old header/footer had been much larger than what NMT does. So we use less memory, not more, in debug.
>> 
>> 2) Another argument for this is that NMT is important. It gets used by support, typically in stressful situations. I rather have it thoroughly tested as part of our internal tests. I am not sure that we have ironed out every single NMT issue. NMT feels very stable to me, but there may be lingering issues. But then we should fix them.
>> 
>> All that said, I'd be happy for you to give this patch a spin in your CI. I can hold off the push for a bit.
>> 
>> Cheers, Thomas
>
> Thanks for the additional info Thomas. I will run it through our CI and report back.

I'm seeing a timeout with langtools tier1 test `tools/javac/failover/CheckAttributedTree.java` on macosx-aarch64. It timed out after 9m. I'll try a few re-runs.

-------------

PR: https://git.openjdk.java.net/jdk/pull/6554