reduce memory for javadoc

[Jonathan Gibbons]

Andrew,

I agree that any change would have to be tested on a variety of
platforms.  Historically, the limit used to be 512m, but has crept up
over recent times.  My investigations have been to understand
why we have been forced to increase this number; as a result
we found out that a memory leak was causing javadoc to retain
references to the source code buffers for all the files it was reading.

Since javadoc is all Java code, it is reasonable to believe the problem,
and the fix, would apply to all platforms, which is why I posted the
update here.

However, additional investigations show that the performance does suffer
if we reduce the Xmx value. So the good news is that we can now reduce
the value; the question is, do we want to?

FWIW, I'm still trying to understand why we need even 512m. But that will
be another story.

-- Jon

On 04/11/2013 03:26 AM, Andrew Hughes wrote:
> ----- Original Message -----
>> Build folk,
>>
>> As a result of these two changesets now in tl/langtools, we can now
>> reduce the memory requirements for javadoc in
>> (root)common/makefiles/javadoc/Javadoc.gmk.
>>
>> changeset:   1678:3f3cc8d3f13c
>> tag:         tip
>> user:        jjg
>> date:        Mon Apr 08 11:57:37 2013 -0700
>> summary:     8011677: EndPosTables should avoid hidden references to Parser
>>
>> changeset:   1677:b402b93cbe38
>> user:        jjg
>> date:        Mon Apr 08 11:54:26 2013 -0700
>> summary:     8011676: Instances of Tokens.Comment should not be defined
>> in inner classes
>>
>>
>> I have done builds with the memory limited to 512m, down from 1024m.  I
>> did try further, with 384m, but that was too much of a reduction.
>>
>> $ hg diff
>> diff -r 11c057460b91 common/makefiles/javadoc/Javadoc.gmk
>> --- a/common/makefiles/javadoc/Javadoc.gmk    Fri Apr 05 14:49:54 2013 -0700
>> +++ b/common/makefiles/javadoc/Javadoc.gmk    Tue Apr 09 18:42:37 2013 -0700
>> @@ -47,7 +47,7 @@
>>    BUILD_NUMBER=$(JDK_BUILD_NUMBER)
>>
>>    JAVADOC_CMD = $(JAVA) \
>> -              -Xmx1024m \
>> +              -Xmx512m \
>>              -Djava.awt.headless=true \
>>              $(NEW_JAVADOC) \
>>              -bootclasspath $(JDK_OUTPUTDIR)/classes
>>
>> -- Jon
>>
> I'd be wary of changing this because, in my experience, it varies depending
> on VM and architecture.  What was tested for these figures?