Re: Fibonacci and performance

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

On 30 Apr 2001, Tom Tromey wrote:
> But Hans is saying that in AG's scenario the check of the method-local
> flag could be reordered with respect to access to class variables,
> unless the flag is volatile.

Reordered by the compiler, or hardware?

>     if (! method-local-flag) { _Jv_InitClass (...); m-l-f = true; }
> 
> Will the compiler really pull a class variable access before this?
> I don't see how it could do that.

Me either.  The compiler can't predict the side effects of the function
call, so it must reload global data following the call.

If _Jv_InitClass is properly synchronized, there is no danger here of
reordering by the processor.  That's where we probably have a bug:  the
short-circuit test isn't synchronized.

> Maybe we could introduce our own java-specific tree code representing
> class initialization, which is only expanded ("lowered") to generic
> tree codes later on.

That's what I was thinking.  By using higher-level tree nodes during
parsing we could eliminate a lot of cruft anyway.  Then run tree-based
optimizations before expanding.  (Maybe we could even generate .class and
.s output in one pass with a little more rework.)

Jeff