Re: Inliner for Java

[Jeff Sturm <jsturm at one-point dot com>]

On Fri, 9 Aug 2002, Andrew Haley wrote:
> I admit that I'm unsure who would be the best person to review this.

I'm not a reviewer, but I did a bit of experimenting with this patch.

First off, I think you lost the "return false;" at the end of
java_post_options.  My compiler ICEd until I replaced it.

The inlining performance seemed about as good as the old RTL inliner:

a) no inlining of constructors, because they don't have DECL_INLINE set
b) no inlining from bytecode, since the jcf frontend doesn't produce
trees
c) inlining takes place in declaration order only, i.e. only preceding
methods are inlined
d) no inlining across translation units

a) and c) seem easy to fix; I patched my own tree to do just that.  b)
is hard until the jcf frontend is rewritten, right?.

I don't know what is needed for d).  When resolve_class finds another
source file, it creates a tree representation, but my attempts to use it
resulted either in bad code or an ICE.  My hunch is that some of the
patch_* functions are not applied to those trees, but I haven't
investigated much.

Regardless, this patch finally provides a solid basis for gcj inliner
improvements, and in my opinion ought to go in (with the java_post_options
fix).

Jeff