|Cache access at compile time firstname.lastname@example.org (1999-11-16)|
|Re: Cache access at compile time Daniel.Vogelheim@post.rwth-aachen.de (Daniel Vogelheim) (1999-11-21)|
|Re: Cache access at compile time Sid-Ahmed-Ali.TOUATI@inria.fr (Sid Ahmed Ali TOUATI) (1999-11-28)|
|Re: Cache access at compile time email@example.com (Dan Truong) (1999-12-07)|
|From:||Sid Ahmed Ali TOUATI <Sid-Ahmed-Ali.TOUATI@inria.fr>|
|Date:||28 Nov 1999 01:31:06 -0500|
> > Can anyone give me pointers to way to ascertain the cache access
> > patterns at compile time.
> The Universitaet Saarbruecken has done research to predict cache
> behaviour of programs in order to guarantee upper time bounds in hard
> real time systems.
> If I remember correctly they analyzed SPARC binaries and given a
> particular cache architecture classified all memory accesses into
> "always hits", "don't know", or "don't know first iteration, always
> hits afterwards" (for loops). The restrictions for this system were
> quite severe ( ? only static addresses could be predicted ? ), but
> they seemed to work allright for the real time people.
> If this sounds like what you want, you can check their papers at:
it was a very interesting work. But the cache behavior prediction was
oriented mainly to instruction cache and scalar acces. A more
interesting problem is to predict the data cache behavior for array
acces in loops (the most serious bottleneck). Ising subscripts will
make the abstract interpretation bit harder. Ferdinand's thesis talk
about this problem and proposed a solution, but its approach was very
restrictive and conservative : I do not see experimental results for
this work (predicting cache miss for array access in a loop).
Return to the
Search the comp.compilers archives again.