Re: binary compatibility ABI

[Bryce McKinlay <bryce at mckinlay dot net dot nz>]

On Thursday, Sep 11, 2003, at 02:16 Pacific/Auckland, Andrew Haley 
wrote:

> Jeff Sturm writes:
> > > I think that keeping the initialization check/call at the call (or
> > > field access) site gives us more opportunity to optimize it out at
> > > compile time. eg: hoisting it out of loops, eliminating it if class 
> > > is
> > > already known to be initialized in the current method, keeping a 
> > > local
> > > cache, etc.
> >
> > Very good point.  Libgcj hides a lot of class inits now, in 'new' or
> > 'instanceof' calls for example.  The more these can be exposed the 
> > easier
> > they are to eliminate.

Yeah - although actually 'instanceof' (and checkcast) do not cause 
initialization.

> Well, kinda sorta.  The first time a class is accessed it has to be
> initialized, but as long as we have to go through an indirection to
> invoke a method we might as well rewrite that indirection once the
> class is initialized.

The problem I see with having mutable method tables is the memory 
ordering problem - it seems that on an MP architecture we would have to 
put a read barrier before each call or else the method table could be 
seen as updated before class initialization is seen as complete.

> > (As an aside, I know there's been discussions before about static
> > vs.  local boolean flags to track class initializations.  What
> > about using TLS on platforms that support it (i.e. x86/linux)?

Interesting idea. I was thinking an initial implementation would just 
use a small cache in the stack for whatever classes are 
used/initialized by the current method. But if thread-local storage can 
be used efficiently, then obviously that is even better.

> > Each thread would check a thread-local boolean variable before
> > calling _Jv_InitClass, so it would have to attempt class
> > initialization at most once per class.  I can't think of a reason
> > that would violate Java's memory model.)
>
> But TLS is slow.

The current implementations may be slow, but the point of TLS is that 
its supposed to be fast - ie it avoids locks and having to worry about 
memory ordering, etc.

Regards

Bryce.