[Python-Dev] Dynamic bytecode analysis

[Damien Morton]

So I enhanced the PVM dynamic execution profiler to keep track of quads of
instructions rather than just pairs. Running it on PyStone, I get the
following results:

Note: score is frequency * len(trace)

http://www.bitfurnace.com/python/dxstats-traces.txt
score    trace
 7204142 LOAD_FAST
 3701796 LOAD_FAST, LOAD_FAST
 3453243 COMPARE_OP, JUMP_IF_FALSE, POP_TOP
 2701110 LOAD_GLOBAL
 2603086 JUMP_IF_FALSE, POP_TOP
 2402196 COMPARE_OP, JUMP_IF_FALSE
 2302782 LOAD_CONST
 2251111 STORE_FAST
 2200676 LOAD_GLOBAL, COMPARE_OP, JUMP_IF_FALSE, POP_TOP
 2101580 LOAD_FAST, LOAD_CONST
 2100896 STORE_FAST, LOAD_FAST
 2000188 COMPARE_OP, JUMP_IF_FALSE, POP_TOP, LOAD_GLOBAL
 1800639 LOAD_FAST, LOAD_CONST, BINARY_ADD
 1650507 LOAD_GLOBAL, COMPARE_OP, JUMP_IF_FALSE
 1650177 JUMP_IF_FALSE, POP_TOP, LOAD_GLOBAL
 1600986 LOAD_GLOBAL, LOAD_FAST
 1600492 LOAD_FAST, LOAD_GLOBAL, COMPARE_OP, JUMP_IF_FALSE
 1500536 LOAD_CONST, BINARY_ADD
 1500339 STORE_FAST, LOAD_FAST, LOAD_FAST
 1500282 LOAD_FAST, STORE_ATTR
 1452043 POP_TOP
 1400524 LOAD_FAST, LOAD_CONST, BINARY_ADD, STORE_FAST
 1400056 LOAD_FAST, LOAD_FAST, BINARY_SUBSCR, LOAD_FAST
 1351596 JUMP_IF_FALSE
 1300164 POP_TOP, LOAD_GLOBAL
 1251127 COMPARE_OP
 1203536 COMPARE_OP, JUMP_IF_FALSE, POP_TOP, LOAD_FAST
...

certainly, making JUMPs consume the top-of-stack still seems worthwhile

other possibilities
LOAD_FAST+FAST == LOAD_FAST, LOAD_FAST
JUMP_IF_CMP == COMPARE_OP, JUMP_IF_FALSE -- consumes two items off the stack
LOAD_FAST+CONST == LOAD_FAST, LOAD_CONST
ADD_CONST == LOAD_CONST, BINARY_ADD
FAST_SUBSCR == LOAD_FAST, LOAD_FAST, BINARY_SUBSCR

If anyone has any suggestions of a better app than PyStone to profile, I
would be grateful.