Request for review 7177409: Perf regression in JVM_GetClassDeclaredFields after generic signature changes

[David Holmes]

HI Jiangli,

On 20/06/2012 3:32 AM, Jiangli Zhou wrote:
> Please review the following webrev that address the performance
> regression in JVM_GetClassDeclaredFields with very large class.
>
> http://cr.openjdk.java.net/~jiangli/7177409/webrev.00/
>
> In fieldDescriptor::generic_signature(), it now returns NULL immediately
> if the access_flags indicates the field has no generic signature. This
> avoids iterating though the fields unnecessarily. Mikael Gerdin has
> confirmed the change resolves the performance issue for
> JVM_GetClassDeclaredFields (thanks, Mikael!).

Your change fixes the problem but I can't help but feel that there is an 
API problem here, and a usage problem in that this fragment in 
reflection.cpp:

   if (java_lang_reflect_Field::has_signature_field() &&
       fd->generic_signature() != NULL) {
     Symbol*  gs = fd->generic_signature();

should really be:

   if (java_lang_reflect_Field::has_signature_field() &&
       fd->has_generic_signature()) {
     Symbol*  gs = fd->generic_signature();

If the query is basically constant time the performance would not be 
impacted. In fact the current code seems to be relying on the C compiler 
to reuse the result of generic_signature() so that it doesn't actually 
call it twice. If it is calling it twice then an API change (for methods 
too) would yield an even better performance improvement.

Maybe an RFE ...

David