Re: [rvm-research] SHA-256 very slow on Jikes?

[Eliot M. <mo...@cs...>]

On 7/29/2010 1:33 PM, Yang Wang wrote:
> Hmmm. I've found one reason: Putting "MessageDigest.getInstance()" out of the loop makes Jikes significant faster:
>
> If that function call is in the loop (for both Jikes and HotSpot), Jikes is 40 times slower than HotSpot
>
> If that function call is out of the loop (still for both), Jikes is only about 20% slower than HotSpot.

Sounds like inlining is affecting this.  Which continues
to point the finger at available registers, one way or the
other, or the register allocation algorithm.

You can get Jikes RVM to dump the code for that method,
which might confirm why it is slow (but not why HotSpot
is faster).  In looking at the open source code for
HotSpot, I see that it has these relevant flags:
   -XX:+UnlockDiagnosticVMOptions
   -XX:+PrintAssembly
But using them causes the VM to try to load a dll that
is not present -- presumably I would have to build the
VM myself with suitable "make" options to get that
dll that support disassembly.

Cheers -- Eliot Moss