[gforth] Performance anomality with dynamic superinstructions on MIPSel

[David Kuehling]

Hi,

I'm using a recent gforth revision from git (6ec9915f6277de) and noticed
that running gforth --dynamic produces pretty extreme performance
degradation (about a factor of 5) for the benchmark I was running [1].
This happens on Loongson-2f MIPS (debian squeeze mipsel, 32bit).  Note
that on MIPS dynamic superinstructions aren't enabled by default as they
may violate load delay slot requirements on some very old MIPS CPUs.

The minimum code I could come up with that clearly shows the anomaly is:

  time gforth-fast -r 600M \
      -e '30000000 :noname 1- DUP 0> IF RECURSE THEN ; EXECUTE BYE' 
  user  0m1.680s

vs.

  time gforth-fast --dynamic -r 600M \
      -e '30000000 :noname 1- DUP 0> IF RECURSE THEN ; EXECUTE BYE'
  user  0m12.529s

I.e. a degradation by a factor of 7.

Any ideas how to proceed further?  This could be a side effect of the
BTB errata of Loongson2f [2] maybe doing speculative loads to invalid
addresses causing instruction stalls or cache flushes.  But then how
could the micro-benchmark shown above ever cause a BTB prediction miss?
(the Loongson2 BTB has 16 entries).  Any ideas how to explain the result
without invoking CPU bugs?

cheers,

David

[1] http://svn.code.sf.net/p/fkt/code/trunk/benchmark.fs
[2] https://sourceware.org/ml/binutils/2009-11/msg00387.html
-- 
GnuPG public key: http://dvdkhlng.users.sourceforge.net/dk2.gpg
Fingerprint: B63B 6AF2 4EEB F033 46F7  7F1D 935E 6F08 E457 205F

pgpPQkpsPWgAx.pgp
Description: PGP signature