Re: PATCH: interpreter.cc bug patch and performance improvements

[Chris Sears <cbsears_sf at yahoo dot com>]

Tom

> I read through your patch.

thanks.

> Chris>   1) in array loads and stores, pushing the SAVE_PC into
> Chris>      the NULLCHECK throw block.
> 
> I think we can't do this, because on many platforms NULLCHECK is
> simply defined away -- we rely on the segv handler to generate the
> NullPointerException.  In this situation we still have to have the
> saved PC correct.

Sorry, I didn't know about the SEGV case.  So in that case, define
NULLARRAYCHECK to be SAVE_PC. I've appended a modified patch.
But since I don't have that configuration, I can't test it.

> Do you know whether this change in isolation really does help
> performance?  And if so, how much?

The whole patch isn't that much.  But with the bug and the fact
that I'm stuck with the cache problem, I figured I'd send it in.
>From cm3 644 up to 662.

> One idea might be for us to have NULLCHECK expand to SAVE_PC in some
> situations.  We'd have to make sure this is always correct though --
> if the PC is incremented before the NULLCHECK it would not be.

I should have read before I typed.  This  isn't a problem for  the
array access bytecodes.  BTW, shouldn't it be possible to save pc
rather than pc-1 and let the handler handle it?

> I consider your patch to be below the "needs paperwork" limit, since
> much of the patch is mechanical substitution.  However, if you're
> planning to send more libgcj patches, it would be best to get started
> on the paperwork soon.  Sometimes that step takes a long time :-(.
> Paperwork is required if you go over some minimum patch
> size/quantity.

Sorry, I'm unfamiliar with the paperwork.  But I will look at the
gnu site and try to figure it out.

thanks,

Chris


__________________________________________________
Do You Yahoo!?
Send FREE video emails in Yahoo! Mail!
http://promo.yahoo.com/videomail/

Attachment: interpret.cc.PATCH
Description: interpret.cc.PATCH