Mr. Wolfgang Hoschek wrote:
> See, these days many computations formerly termed compute-intensive are
> actually bounded by the memory subsystem. (Easily run 10 times as fast
> when fetching from Level-1-Cache instead of SDRAM, not only in C but
> also in Java.) That's the reason the PentiumIII introduces new helper
> instructions for prefetching memory and hiding memory latency.
AMD's K6-2 and Athlon have some instructions for
prefetching, which are PREFETCH and PREFETCHW.
> But if Java jitters do not generate code
> exploiting CPU specific instructions, the value remains limited.
`MOV EAX,EDX' is an instruction specific to x86
and all JIT compilers will generate it. :)
Sorry, it's a joke.
I know there are some efforts to exploit particular
instructions which are not common to most processors.
Java Grande Forum (*1) Numerics Working Group (*2)
(JGNWG) has been trying to allow JVM and JIT developers
to use PowerPC's fmac (fused multiply and accumulate)
instruction. In May 1998, They proposed an
`associativefp' modifier, but the proposal is refused by
Sun. But a `strictfp' modifier which is introduced into
JDK 1.2 is an important result of the discussion.
(*1) http://www.javagrande.org/
(*2) http://math.nist.gov/javanumerics/
JIT compiler may utilize even MMX instructions of
x86. Of course, current JITs cannot utilize them
efficiently, but I have a paper written by Intel
researchers. The paper describes the method for JIT to
exploit MMX instructions.
> > I don't know about the linux jit. IBM research
> > may have put out some papers about jit methodology...
> > but I doubt they would go into that level of detail.....
IBM JIT which is attached to Linux/Windows JDK 1.1.8 is
developed by IBM Tokyo Research Lab.(TRL)
researchers. Some papers I've read state optimization
techniques which implemented to the JIT and effects of
each methods in detail.
Kazuyuki SHUDO Happy Hacking!
Muraoka Lab., School of Sci. & Eng., Waseda Univ.
----------------------------------------------------------------------
To UNSUBSCRIBE, email to [EMAIL PROTECTED]
with a subject of "unsubscribe". Trouble? Contact [EMAIL PROTECTED]
