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

Christian Thalinger christian.thalinger at oracle.com
Wed Sep 25 14:17:13 PDT 2013


Oh well.  It turned out that SunStudio cannot handle strings properly in the static table.  So I chose another approach which actually uses less memory:  for develop and notproduct flags define both a const variable and a static variable with CONST_ as a prefix.  The address of the latter is then used in the flags table.  No need for an additional _value field.

With that the binary size increase is:

linux_i486_compiler2: 24k
linux_i486_minimal1: 17k

I hope this is the last iteration:

http://cr.openjdk.java.net/~twisti/8024545/webrev/

The gist of this iteration is at the bottom of this page:

http://cr.openjdk.java.net/~twisti/8024545/webrev/src/share/vm/runtime/globals.hpp.udiff.html

On Sep 19, 2013, at 6:34 PM, Christian Thalinger <christian.thalinger at oracle.com> wrote:

> Thank you, again, Vladimir.
> 
> On Sep 19, 2013, at 4:51 PM, Vladimir Kozlov <vladimir.kozlov at oracle.com> wrote:
> 
>> Good.
>> 
>> Vladimir
>> 
>> On 9/19/13 4:08 PM, Christian Thalinger wrote:
>>> JPRT found some problems.  Apparently something does #define C2 'E' on Solaris so I had to prefix all kind flags with KIND_:
>>> 
>>> +     // flag kind
>>> +     KIND_PRODUCT            = 1 << 4,
>>> +     KIND_MANAGEABLE         = 1 << 5,
>>> +     KIND_DIAGNOSTIC         = 1 << 6,
>>> +     KIND_EXPERIMENTAL       = 1 << 7,
>>> +     KIND_NOT_PRODUCT        = 1 << 8,
>>> +     KIND_DEVELOP            = 1 << 9,
>>> +     KIND_PLATFORM_DEPENDENT = 1 << 10,
>>> +     KIND_READ_WRITE         = 1 << 11,
>>> +     KIND_C1                 = 1 << 12,
>>> +     KIND_C2                 = 1 << 13,
>>> +     KIND_ARCH               = 1 << 14,
>>> +     KIND_SHARK              = 1 << 15,
>>> +     KIND_LP64_PRODUCT       = 1 << 16,
>>> +     KIND_COMMERCIAL         = 1 << 17,
>>> 
>>> Two other problems found by Visual Studio were a bool return:
>>> 
>>> + bool Flag::get_bool() const {
>>> +   if (is_constant_in_binary()) {
>>> +     return _value != 0;
>>> +   } else {
>>> +     return *((bool*) _addr);
>>> +   }
>>> + }
>>> 
>>> and the size_t changes at the bottom of this file:
>>> 
>>> http://cr.openjdk.java.net/~twisti/8024545/webrev/src/share/vm/runtime/globals.cpp.udiff.html
>>> 
>>> Everything else (except the renames) are the same.  I tried to create an incremental webrev but I failed.  Sorry.
>>> 
>>> On Sep 12, 2013, at 6:00 PM, Christian Thalinger <christian.thalinger at oracle.com> wrote:
>>> 
>>>> 
>>>> On Sep 12, 2013, at 5:24 PM, Vladimir Kozlov <vladimir.kozlov at oracle.com> wrote:
>>>> 
>>>>> On 9/12/13 5:11 PM, Christian Thalinger wrote:
>>>>>> 
>>>>>> 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?
>>>>> 
>>>>> Do it in separate changes after you push this.
>>>>> And if you remove _type field later you don't need to move _flags now.
>>>>> So you can push your current changes as it is. They are good.
>>>> 
>>>> Thank you, Vladimir.  -- Chris
>>>> 
>>>>> 
>>>>> Thanks,
>>>>> Vladimir
>>>>> 
>>>>>> 
>>>>>> -- 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.
>>>>>>>>>> 
>>>>>>>> 
>>>>>> 
>>>> 
>>> 
> 



More information about the hotspot-dev mailing list