This patch by Antoine Pitrou optimizes the bytecode for conditional branches by
merging the following "POP_TOP" instruction into the conditional jump. For
example, the list comprehension "[x for x in l if not x]" produced the
following bytecode:
1 0 BUILD_LIST 0
3 LOAD_FAST 0 (.0)
>> 6 FOR_ITER 23 (to 32)
9 STORE_FAST 1 (x)
12 LOAD_FAST 1 (x)
15 JUMP_IF_TRUE 10 (to 28)
18 POP_TOP
19 LOAD_FAST 1 (x)
22 LIST_APPEND 2
25 JUMP_ABSOLUTE 6
>> 28 POP_TOP
29 JUMP_ABSOLUTE 6
>> 32 RETURN_VALUE
but after the patch it produces the following bytecode:
1 0 BUILD_LIST 0
3 LOAD_FAST 0 (.0)
>> 6 FOR_ITER 18 (to 27)
9 STORE_FAST 1 (x)
12 LOAD_FAST 1 (x)
15 POP_JUMP_IF_TRUE 6
18 LOAD_FAST 1 (x)
21 LIST_APPEND 2
24 JUMP_ABSOLUTE 6
>> 27 RETURN_VALUE
Notice that not only the code is shorter, but the conditional jump
(POP_JUMP_IF_TRUE) jumps right to the start of the loop instead of going through
the JUMP_ABSOLUTE at the end. "continue" statements are helped
similarly.
Furthermore, the old jump opcodes (JUMP_IF_FALSE, JUMP_IF_TRUE) have been
replaced by two new opcodes:
- JUMP_IF_TRUE_OR_POP, which jumps if true and pops otherwise
- JUMP_IF_FALSE_OR_POP, which jumps if false and pops otherwise
on compilers that support it (notably: gcc, SunPro, icc), the bytecode
evaluation loop is compiled with a new dispatch mechanism which gives
speedups of up to 20%, depending on the system, on various benchmarks.
Jeffrey Yasskin
Antoine Pitrou