question about gcj -fjni code

[Tom Tromey <tromey at redhat dot com>]

I'm looking into a strange libgcj runtime failure, and I've found I
don't understand some of the code gcj generates.  I'm hoping someone
can help me out.

This is on x86 Red Hat Linux 6.2.

I compile a file like so:

    gcj -fPIC -fjni -S -g0 gnu/java/awt/peer/gtk/GtkComponentPeer.java 


In gcj, with -fjni, native methods are compiled as stubs that make
calls into the runtime.  These stubs are generated by
expr.c:build_jni_stub.  I suspect my problem lies there.


Here is a small piece of the output that looks ok:

	    cmpl	$0, meth.31@GOTOFF(%ebx)
	    je	.L135
	    movl	meth.31@GOTOFF(%ebx), %eax
	    movl	%eax, -12(%ebp)
     [ ... ]
	    .align 4
	    .type	meth.31,@object
	    .size	meth.31,4
    meth.31:
	    .long	0


In another method, I get very similar code:

	    cmpl	$0, meth.5@GOTOFF(%ebx)
	    je	.L8
	    movl	meth.5@GOTOFF(%ebx), %eax
	    movl	%eax, -16(%ebp)
    [ ... ]

However, `meth.5' doesn't appear to be defined in the .s file.
At runtime this manifests as the variable having some random value,
which causes a crash.  The value ought to be initialized to 0.

Why doesn't this cause an error from the assembler?

The `meth' variables are constructed like so:

  meth_var = build_decl (VAR_DECL, get_identifier ("meth"), ptr_type_node);
  TREE_STATIC (meth_var) = 1;
  TREE_PUBLIC (meth_var) = 0;
  DECL_EXTERNAL (meth_var) = 0;
  DECL_CONTEXT (meth_var) = method;
  make_decl_rtl (meth_var, NULL);

I tried setting TREE_USED and DECL_INITIAL on the VAR_DECL, but that
didn't have an effect.

I'm hoping I'm missing something obvious.

Tom