[New-bugs-announce] [issue27255] More opcode predictions

[Serhiy Storchaka]

New submission from Serhiy Storchaka:

Currently the PREDICT() macros is used for predicticting following pair of opcodes in the ceval loop:

LIST_APPEND JUMP_ABSOLUTE
SET_ADD JUMP_ABSOLUTE
MAP_ADD JUMP_ABSOLUTE
COMPARE_OP POP_JUMP_IF_FALSE
COMPARE_OP POP_JUMP_IF_TRUE
GET_ITER FOR_ITER
FOR_ITER STORE_FAST (if continue iterating)
FOR_ITER UNPACK_SEQUENCE (if continue iterating)
WITH_CLEANUP_START WITH_CLEANUP_FINISH
WITH_CLEANUP_FINISH END_FINALLY

I collected static and dynamic opcode statistics. Static statistics is a statistics about what opcodes are used in code objects after running the compileall script on the Lib directory. Dynamic statistics is is a statistics about what opcodes are executed when run all Python tests. Resulting numbers are close, but there are some reasonable differences (opcodes used in loops have larger numbers in dynamic statistics, and opcodes used for creating classes and functions have larger numbers in static statistics). In following lines the fist number is from dynamic statistics and the second number is from static statistics.

LIST_APPEND, SET_ADD and MAP_ADD are always followed by JUMP_ABSOLUTE.
COMPARE_OP is followed by POP_JUMP_IF_FALSE in 83% (78%) and by POP_JUMP_IF_TRUE in 13% (5%).
GET_ITER is followed by FOR_ITER in 72% (80%) and CALL_FUNCTION in 28% (28%).
FOR_ITER is followed by STORE_FAST in 50% (80%), UNPACK_SEQUENCE in 21% (18%) and POP_BLOCK in 20% (0%).
WITH_CLEANUP_START is always followed by WITH_CLEANUP_FINISH.
WITH_CLEANUP_FINISH is always followed by END_FINALLY

According to this statistic existing predictions are correct.

But the statistics suggests other predictions.

JUMP_ABSOLUTE is followed by FOR_ITER in 77% (0%).
UNPACK_SEQUENCE is followed by STORE_FAST in 97% (94%).
SETUP_LOOP is followed by LOAD_FAST in 81% (52%) and LOAD_GLOBAL in 18% (31%).
BUILD_SLICE is followed by BINARY_SUBSCR in 99% (87%).
ROT_TWO is followed by STORE_FAST in 85% (40%).
LOAD_CLOSURE is followed by BUILD_TUPLE in 94% (68%).
SETUP_WITH is followed by POP_TOP in 94% (54%) and STORE_FAST in 6% (44%).
GET_YIELD_FROM_ITER, GET_AITER, GET_ANEXT and GET_AWAITABLE are always followed by LOAD_CONST.
DUP_TOP_TWO is always followed by BINARY_SUBSCR.
BEFORE_ASYNC_WITH is always followed by GET_AWAITABLE.
All INPLACE_ instructions almost always are followed by STORE_FAST.

Proposed patch adds predictions of following pairs:

DUP_TOP_TWO BINARY_SUBSCR
INPLACE_POWER STORE_FAST
INPLACE_MULTIPLY STORE_FAST
INPLACE_MATRIX_MULTIPLY STORE_FAST
INPLACE_TRUE_DIVIDE STORE_FAST
INPLACE_FLOOR_DIVIDE STORE_FAST
INPLACE_MODULO STORE_FAST
INPLACE_ADD STORE_FAST
INPLACE_SUBTRACT STORE_FAST
INPLACE_LSHIFT STORE_FAST
INPLACE_RSHIFT STORE_FAST
INPLACE_AND STORE_FAST
INPLACE_XOR STORE_FAST
INPLACE_OR STORE_FAST
GET_AITER LOAD_CONST
GET_ANEXT LOAD_CONST
GET_AWAITABLE LOAD_CONST
UNPACK_SEQUENCE STORE_FAST
LOAD_CLOSURE BUILD_TUPLE
GET_ITER CALL_FUNCTION
GET_YIELD_FROM_ITER LOAD_CONST
FOR_ITER POP_BLOCK
BEFORE_ASYNC_WITH GET_AWAITABLE
SETUP_WITH POP_TOP
SETUP_WITH STORE_FAST
BUILD_SLICE BINARY_SUBSCR

Note that the effect of this change is not very significant since the PREDICT() macros is no-op if computed GOTOs are used.

----------
components: Interpreter Core
files: more_opcode_predicts.patch
keywords: patch
messages: 267719
nosy: serhiy.storchaka
priority: normal
severity: normal
stage: patch review
status: open
title: More opcode predictions
type: enhancement
versions: Python 3.6
Added file: http://bugs.python.org/file43283/more_opcode_predicts.patch

_______________________________________
Python tracker <report at bugs.python.org>
<http://bugs.python.org/issue27255>
_______________________________________