[Python-Dev] Re: opcode performance measurements

[Jeremy Hylton]

>>>>> "SM" == Skip Montanaro <skip@pobox.com> writes:

  SM> Interesting results.  I've been working on my
  SM> {TRACK,UNTRACK}_GLOBAL opcode implementations.  I have an
  SM> optimizer filter that sets up tracking for all
  SM> LOAD_GLOBAL,{LOAD_ATTR}* combinations.  It's still not quite
  SM> working and will only be a proof of concept by devday if I do
  SM> get it working, but I expect most of these expensive opcode
  SM> combinations to collapse into a LOAD_FAST, with the addition of
  SM> a TRACK_GLOBAL/UNTRACK_GLOBAL pair executed at function start
  SM> and end, respectively.

I won't have any implementation done at all, but should have finished
the design for LOAD_FAST-style access to globals and module
attributes.  I also have some ideas about Python bytecode specializer
that would work essentially like a JIT but generated specialized
bytecode instead of machine code.

Jeremy

PS Skip-- Sorry the PEP isn't clear, but the only dictionary lookups
that need to occur are at function creation time.  MAKE_FUNCTION would
need to lookup the offsets of the globals used by the functions, so
that a LOAD_FAST_GLOBAL opcode would take an int argument.