review for 7086585: make Java field injection more flexible
tom.rodriguez at oracle.com
Fri Sep 9 09:31:28 PDT 2011
On Sep 8, 2011, at 6:56 PM, Vladimir Kozlov wrote:
> Thank you, Tom
> Very nice cleanup.
Thanks. I think there's another possible round of this to simplify the field layout and javaClasses some more but I didn't want to tangle it up with the injection changes.
> Can we use an arrays instead of switch to map type to allocation type in FieldAllocationCount::update()?
I considered doing that. I'll go ahead and add it.
> Can we factor out next sequence since it is used in two places?:
> + RawFieldDescriptor* field = RawFieldDescriptor::from_field_array(fields(), n);
> + field->set_access_flags(access_flags.as_short());
> + field->set_name_index(name_index);
> + field->set_signature_index(signature_index);
> + field->set_initval_index(constantvalue_index);
> + field->set_generic_signature_index(generic_signature_index);
I'm not sure how to factor that in a way that would improve it much. Any suggestions?
> Tom Rodriguez wrote:
>> 1577 lines changed: 621 ins; 612 del; 344 mod; 65541 unchg
>> 7086585: make Java field injection more flexible
>> Classes which are special to the JVM often need fields for the VM's
>> use. Currently the mechanism for this is adhoc or relies on the Java
>> code declaring the needed fields. We need a mechanism for flexibly
>> injecting needed fields. The basic idea is to built a standard
>> mechanism for declaring injected fields and to inject them during
>> class file parsing so that they get laid out and tracked like normal
>> fields. This is also means that they can simply be looked up when CDS
>> is used so we don't have to rely on hardcoding them. To do this I
>> hide the encoding of the fields array and replaced all existing
>> iteration with a new set of *FieldStream classes. JavaFieldStream
>> only iterates true Java fields so the behavious of existing code is
>> completely unchanged.
>> The new fields use a special access flag JVM_ACC_INTERNAL to describe
>> themselves and instead of having the name_index and signature_index
>> refer to the constant pool they are vmSymbols::SID so they are looked
>> up directly from the vmSymbols. All of this is hidden inside the
>> RawFieldDescriptor class which is responsible for decoding fields from
>> the field array.
>> I also cleaned up the field layout code a bit. I eliminated the
>> _ALIGNED_ AllocationType since there's no longer any difference in how
>> we align double word types. I turned FieldAllocationCount into a real
>> class and moved some of the layout code into it.
>> I deleted the nonexistent MethodHandle::vmslot field. I moved
>> ClassLoader::parallelCapable into java_lang_ClassLoader. I deleted
>> some dead field injection like the Reference.discovered field which
>> has existed since 1.5, and the existing logic for java.lang.Class
>> fields since it now uses the field injection machinery.
>> Tested with runthese, the sajdi tests, the tmtools suite, the nsk
>> regression, jdb, jdi, jvmti, hprof and jit subsuites along with
>> vm/mlvm, but with and without the JDK changes that remove these fields
>> and rely on injection to introduce them. They will be removed under
>> 7088481 and those changes are http://cr.openjdk.java.net/~never/7088481.
More information about the hotspot-compiler-dev