Re: optimizing OOPL's

I just finished writing a short paper w/a longer version due shortly
on profile-based optimization for C++ programs. Our goal was to
accurately predict the destination of indirect function calls, mainly
to avoid misprediction penalties on speculative or heavily pipelined
architectures.


It turns out that for the programs we examined, profile-based static
prediction was very successful (90% prediction rate), although
introduction of some dynamic methods improved this (albeit
w/additional hardware). Even on a machine like the Alpha, this should
show a 5-10% performance improvement for some programs, and probably
more if you than inlined the functions.


Note that this isn't based on type inference -- Barbera Ryder has
shown that type inference in C++ is NP hard in general -- moreover,
you need the entire set of classes to infer anything meaningful
anyway.


Dirk Grunwald Asst. Prof, Univ. of Colorado at Boulder
--

Return to the comp.compilers page.
Search the comp.compilers archives again.