cross-module inlining (was Re: using gcj for a different language - is it possible?)

Mon Feb 9 17:48:00 GMT 2004

Jeff Sturm writes:
 > On Mon, 9 Feb 2004, Andrew Haley wrote:
 > > > I've been thinking a little about this since it can restrict inlining,
 > > > escape analysis, and any other sort of useful IPA gcj might perform.
 > >
 > > If you want to do IPA, compile the files at the same time, and this
 > > gives gcj carte blanche to do any optimizations it can find, including
 > > full inlining. Separate compilation, however, implies binary
 > > compatibility.
 > 
 > Yes. Well, almost. I don't think gcj should have carte blanche
 > when building a library, unless it continues to disallow class
 > replacement at runtime.
Well, compiling everything together rather implies that, because calls
can be made directly between classes. Anyone who really wants to
replace an individual class in a library at runtime had better not
compile a whole library with -O3. But that's up to them.
 > > > Similarly in Java, why not leave the most aggressive optimizations
 > > > for the main executable (non-PIC) and stick to conservative
 > > > binary-compatible optimizations for libraries? Does binary
 > > > compatibility really matter for non-library code?
 > >
 > > That's up to the user, surely. I don't think we need to have any
 > > particular policy here.
 > 
 > Going back to the OP:
 > 
 > "We're sort of moving away from doing inlining by default, since that
 > breaks binary compatibility."
That was what I was disagreeing with. We have never inlined by
default, only with -O3, and only in a single compilation.
 > Sounds like the preferred policy will be indirect dispatch (binary
 > compatibility) everywhere.
The default policy, yes. By default we comply with the JLS.
 > My simple suggestion is that indirect dispatch is really only
 > needed for libraries, and by default more aggressive IPA can be
 > done for classes compiled into the main executable.
No, that's wrong. You build an executable from a number of objects.
You don't want to have to recompile all its callers whan a class
changes. This is true whether the class is in a library or not.
Surely our default should be compliance with the JLS.
Andrew.