Re: Proposal for CNI/JNI problems

[Per Bothner <per at bothner dot com>]

Bernd Kreimeier <bk@lokigames.com> writes:

> compile time additional information about object memory layout,
> for a JNI source?

Well, in some ways you have more information, because you all
the low-level details needed for JNI.  But in other ways it is much
harder [to compile JNI to the CNI ABI] because JNI is all function
calls and it may be difficult to infer the programmer's intention.
For example, you have a JNI function that gets a jfieldID, and
other functions to actually use the jfieldID to access a field.
The problem is the compiler needs to follow the chain of
getting the jclass -> getting the jfieldID -> getting the actual field
before it can figure out the actual field.  It need to replace all
of that with a direct field access in CNI.  Following that chain
can be difficult when (e.g.) a jfieldID is cached.  It may require
some global analysis (i.e. more than a single method).

> In other words, the semantics of JNI code prohibits all
> CNI-like optimizations due to parsing complexity issues?

Not the parsing;  the global analysis mentioned above.

> If a hypothetical G++ "emit-jni" option would generate JNI C 

G++ emits assembly code, not C. -femit-jni would emit binary
code *equivalent* to JNI C (i.e. *as if* compiled from JNI C),
but not actually C.

> No doubt, JNI ABI compatibility would make CNI a much more
> viable option. In this case, the major concerns seem to be
> Hans Boehm's points (3) and (5), quote:

Those are certainly important concerns.  We may have to settle
for conditional compilation in tricky cases.  The goal, though,
would be that *most* native methods would not need conditional
compilation, or that the -femit-jni at least greatly reduces
the uglyness of conditional compilation.  If we can't figure out
how to do that, then there isn't much point in the idea.
-- 
	--Per Bothner
per@bothner.com   http://www.bothner.com/~per/