Marmot papers

Tom Tromey tromey@cygnus.com
Thu Oct 26 11:51:00 GMT 2000


Bryce> It is also interesting that they were able to reduce the cost
Bryce> of safety checks (type checking, array bounds) to an average of
Bryce> around 4% of execution time - checks that are currently a major
Bryce> performance drain on gcj.

Did this work require whole-program analysis?  Or could they reduce
the cost while still keeping separate compilation?

Bryce> Also, despite a very efficient lock implementation, they found
Bryce> that synchronization elimination optimizations in the compiler
Bryce> could improve execution time by as much as 30%.

One of IBM Jalapeno (I think) papers about escape analysis has a
similar finding.  Their approach required whole-program analysis,
though, which is currently outside our reach.  Implementing a thinner
synchronization approach is still on our to-do list, though.  That
should gain something.

Tom


More information about the Java mailing list