[Python-Dev] Opcode frequency

[Antoine Pitrou]

Hi,

> Maciej Fijalkowski did an opcode analysis for PyPy,
> it also shows the relative frequency of opcodes following a
> specifc one:
> 
> http://codespeak.net/svn/user/fijal/opcodes.txt

Nice, but we have to be careful here: what is the tested workload?
For example, I find it hard to believe that CALL_FUNCTION_VAR_KW would 
always (99%) be followed by STORE_FAST.

I'm not even sure we're talking about the same VM/compiler. What are
CALL_LIKELY_BUILTIN and LOOKUP_METHOD? :-)

> Might it make sense to add more PREDICT()ions based
> on this, e.g. for BUILD_SLICE -> BINARY_SUBSCR?

This particular one sounds fine, yes.

Another approach is to create opcodes which better reflect actual usage.
For example part of the following patch (*) was to create opcodes fusing
JUMP_IF_{TRUE,FALSE} and POP_TOP. This is because in most cases, you want
to pop the comparison result regardless of its value (true or false), which
nowadays produces the following kind of code:

    JUMP_IF_FALSE <branch_for_false>
    POP_TOP
    [... do stuff and then ...]
    JUMP_FORWARD <branch_for_false_after_pop_top>
branch_for_false:
    POP_TOP
branch_for_false_after_pop_top:
    [etc.]

With a JUMP_IF_FALSE which also pops the pop of the stack, this gives:

    JUMP_IF_FALSE <branch_for_false>
    [... do stuff and then ...]
branch_for_false:
    [etc.]

Prediction doesn't help you in this case.

(*) http://bugs.python.org/issue2459

Regards

Antoine.