RFR (L): 8024545: make develop and notproduct flag values available in product builds

[Christian Thalinger]

On Sep 12, 2013, at 4:30 PM, Vladimir Kozlov <vladimir.kozlov at oracle.com> wrote:

> Christian,
> 
> Can you put _flags first to avoid padding between _addr and _value in 32-bit VM?

Yes, good point.

> 
> It would be nice enumerate flag's type and use int instead of strings:
> 
> enum {
>  bool_flag = 1,
>  intx_flag = 2,
> 
> + #define RUNTIME_PRODUCT_FLAG_STRUCT(     type, name, value, doc) { type##_flag, XSTR(name), &name,              VALUE(value), NOT_PRODUCT_ARG(doc) Flag::Flags(Flag::DEFAULT | Flag::KIND_PRODUCT) },

I was thinking of putting the type into the flags, too.  That would save another pointer word.  Should I give it a shot in this change or a separate one?

-- Chris

> 
> Thanks,
> Vladimir
> 
> On 9/12/13 9:58 AM, Christian Thalinger wrote:
>> 
>> On Sep 11, 2013, at 5:34 PM, David Holmes <david.holmes at oracle.com> wrote:
>> 
>>> Hi Chris,
>>> 
>>> There seems to be an awful lot "infrastructure" work needed to make something that sounds simple happen. :( All the changes to the scoping and the set/get methods tends to obscure the core change. My only suggestion here is that the "set" methods could perhaps factor this:
>>> 
>>> +   if (is_constant_in_binary()) {
>>> +     fatal(err_msg("flag is constant: %s", _name));
>>> 
>>> into a check_writable() method so that it isn't duplicated so much.
>> 
>> Good point.  I made that change:
>> 
>> http://cr.openjdk.java.net/~twisti/8024545/webrev/src/share/vm/runtime/globals.cpp.udiff.html
>> 
>>> 
>>> I also wonder whether a ConstFlag sub/superclass would simplify this at all?
>> 
>> Maybe it would but then we need more infrastructure to read the additional entries.  Might be a wash.  SA only knows about offsets in structs and the flags array is statically defined.
>> 
>>> 
>>> That aside I'm curious why the minimal VM size change is only 22K when client is 32K?
>> 
>> The 32-bit product build also contains the server compiler.
>> 
>> -- Chris
>> 
>>> 
>>> Thanks,
>>> David
>>> 
>>> On 11/09/2013 10:56 AM, Christian Thalinger wrote:
>>>> http://cr.openjdk.java.net/~twisti/8024545/webrev/
>>>> 
>>>> 8024545: make develop and notproduct flag values available in product builds
>>>> Reviewed-by:
>>>> 
>>>> Right now the internal flag table only contains flags which are defined in a product build. This does not include develop and notproduct flags. Sometimes it is useful to have access to these values for post-mortem core file analysis or to read these values for compiler settings for a Java-based compiler.
>>>> 
>>>> This change enables develop and notproduct flag values to be read by the serviceability agent. The binary size is increased by 42k for a 64-bit product build and by 32k for a 32-bit product build.
>>>> 
>>>> Before:
>>>> 
>>>> $ java -cp /java/re/jdk/8/latest/binaries/linux-x64/lib/sa-jdi.jar sun.jvm.hotspot.CLHSDB 9399
>>>> Attaching to process 9399, please wait...
>>>> hsdb> flags -nd
>>>> InitialHeapSize = 495006528 5
>>>> MaxHeapSize = 7920943104 5
>>>> UseCompressedKlassPointers = true 5
>>>> UseCompressedOops = true 5
>>>> UseParallelGC = true 5
>>>> hsdb> flags InlineMathNatives
>>>> Couldn't find flag: InlineMathNatives
>>>> 
>>>> After:
>>>> 
>>>> $ java -cp $JAVA_HOME/lib/sa-jdi.jar sun.jvm.hotspot.CLHSDB 3726
>>>> Attaching to process 3726, please wait...
>>>> hsdb> flags -nd
>>>> InitialHeapSize = 495006528 5
>>>> MaxHeapSize = 7920943104 5
>>>> UseCompressedKlassPointers = true 5
>>>> UseCompressedOops = true 5
>>>> UseParallelGC = true 5
>>>> hsdb> flags InlineMathNatives
>>>> InlineMathNatives = true 0
>>>> 
>>>> This patch has one behavioral difference; when printing flags with e.g. PrintFlagsFinal in a debug build it prints "develop" for develop flags:
>>>> 
>>>>     uintx AdaptiveSizePolicyGCTimeLimitThreshold    = 5               {develop}
>>>> 
>>>> The output for product builds is unchanged.
>>>> 
>>