[issue42126] Optimize BUILD_CONST_KEY_MAP for distinct keys

[Inada Naoki]

Inada Naoki  added the comment:

OK. Microbenchmark don't justify adding complexity to the eval loop and the 
compiler.

--
resolution:  -> rejected
stage: patch review -> resolved
status: open -> closed

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue42126] Optimize BUILD_CONST_KEY_MAP for distinct keys

[Mark Shannon]

Mark Shannon  added the comment:

It sounds like the root cause is that annotations are repeatedly being 
computed. We should address the underlying issue, IMO.

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue42126] Optimize BUILD_CONST_KEY_MAP for distinct keys

[Inada Naoki]

Inada Naoki  added the comment:

It is difficult to estimate. Real world applications has different code-style 
than stdlib. And pyperformance didn't include whole real world applications too.
Some application may use dict display in heavy loop. But it is difficult to say 
how much exactly.

One example from pyperformance. It uses inner function and function annotations 
heavily.
BUILD_CONST_KEY_MAP is used to build annotation dict, instead of dict display.
https://github.com/tornadoweb/tornado/blob/5eb7bb8efb4e1be8a2cec1744c857a8dadd43af9/tornado/gen.py

```
591  32 LOAD_CONST   1 ('Future')
 34 LOAD_CONST   2 ('None')
 36 LOAD_CONST   3 (('future', 'return'))
 38 BUILD_CONST_KEY_MAP  5
 40 LOAD_CLOSURE 3 (quiet_exceptions)
 42 BUILD_TUPLE  1
 44 LOAD_CONST   4 ()
 46 LOAD_CONST   5 
('with_timeout..error_callback')
 48 MAKE_FUNCTION   12 (annotations, closure)
 50 STORE_DEREF  0 (error_callback)
```

I have not posted pyperformance result because this pull request is based on 
#41835 and I don't measure each performance.
But #41835 + this is 16% faster than master on bm_tornado_http.

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue42126] Optimize BUILD_CONST_KEY_MAP for distinct keys

[Mark Shannon]

Mark Shannon  added the comment:

Is this worthwhile?

Statically, BUILD_CONST_KEY_MAP represents 0.14% of all bytecode instructions 
in the standard library, but only 0.03% of the bytecode instructions in 
functions.
Which suggests that dynamically only 1 in 3000 instructions executed is 
BUILD_CONST_KEY_MAP. The implication is that making BUILD_CONST_KEY_MAP more 
complex is just as likely to slow things down as speed them up.

Do you have any evidence that this will make a measurable difference?

--
nosy: +Mark.Shannon

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue42126] Optimize BUILD_CONST_KEY_MAP for distinct keys

[Inada Naoki]

Inada Naoki  added the comment:

$ ./python -m pyperf timeit --compare-to ~/pyenv/versions/3.10-dev/bin/python 
'{}'
/home/inada-n/pyenv/versions/3.10-dev/bin/python: . 23.5 ns 
+- 0.2 ns
/home/inada-n/work/python/cpython/python: . 22.4 ns +- 0.1 
ns

Mean +- std dev: [/home/inada-n/pyenv/versions/3.10-dev/bin/python] 23.5 ns +- 
0.2 ns -> [/home/inada-n/work/python/cpython/python] 22.4 ns +- 0.1 ns: 1.05x f
aster (-5%)

$ ./python -m pyperf timeit --compare-to ~/pyenv/versions/3.10-dev/bin/python 
'{1:1,2:2}'
/home/inada-n/pyenv/versions/3.10-dev/bin/python: . 83.5 ns 
+- 0.3 ns
/home/inada-n/work/python/cpython/python: . 96.7 ns +- 0.2 
ns

Mean +- std dev: [/home/inada-n/pyenv/versions/3.10-dev/bin/python] 83.5 ns +- 
0.3 ns -> [/home/inada-n/work/python/cpython/python] 96.7 ns +- 0.2 ns: 1.16x s
lower (+16%)

$ ./python -m pyperf timeit --compare-to ~/pyenv/versions/3.10-dev/bin/python 
'{1:1,2:2,3:3}'
/home/inada-n/pyenv/versions/3.10-dev/bin/python: . 108 ns 
+- 0 ns
/home/inada-n/work/python/cpython/python: . 110 ns +- 1 ns

Mean +- std dev: [/home/inada-n/pyenv/versions/3.10-dev/bin/python] 108 ns +- 0 
ns -> [/home/inada-n/work/python/cpython/python] 110 ns +- 1 ns: 1.01x slower
(+1%)

$ ./python -m pyperf timeit --compare-to ~/pyenv/versions/3.10-dev/bin/python 
'{1:1,2:2,3:3,4:4}'
/home/inada-n/pyenv/versions/3.10-dev/bin/python: . 134 ns 
+- 1 ns
/home/inada-n/work/python/cpython/python: . 122 ns +- 1 ns

Mean +- std dev: [/home/inada-n/pyenv/versions/3.10-dev/bin/python] 134 ns +- 1 
ns -> [/home/inada-n/work/python/cpython/python] 122 ns +- 1 ns: 1.10x faster
(-9%)

$ ./python -m pyperf timeit --compare-to ~/pyenv/versions/3.10-dev/bin/python 
'{1:1,2:2,3:3,4:4,5:5}'
/home/inada-n/pyenv/versions/3.10-dev/bin/python: . 168 ns 
+- 1 ns
/home/inada-n/work/python/cpython/python: . 112 ns +- 1 ns

Mean +- std dev: [/home/inada-n/pyenv/versions/3.10-dev/bin/python] 168 ns +- 1 
ns -> [/home/inada-n/work/python/cpython/python] 112 ns +- 1 ns: 1.50x faster
(-33%)

$ ./python -m pyperf timeit --compare-to ~/pyenv/versions/3.10-dev/bin/python 
'{1:1,2:2,3:3,4:4,5:5,6:6}'
/home/inada-n/pyenv/versions/3.10-dev/bin/python: . 223 ns 
+- 1 ns
/home/inada-n/work/python/cpython/python: . 150 ns +- 4 ns

Mean +- std dev: [/home/inada-n/pyenv/versions/3.10-dev/bin/python] 223 ns +- 1 
ns -> [/home/inada-n/work/python/cpython/python] 150 ns +- 4 ns: 1.48x faster
(-33%)

$ ./python -m pyperf timeit --compare-to ~/pyenv/versions/3.10-dev/bin/python 
'{1:1,2:2,3:3,4:4,5:5,6:6,7:7,8:8}'
/home/inada-n/pyenv/versions/3.10-dev/bin/python: . 273 ns 
+- 1 ns
/home/inada-n/work/python/cpython/python: . 177 ns +- 1 ns

Mean +- std dev: [/home/inada-n/pyenv/versions/3.10-dev/bin/python] 273 ns +- 1 
ns -> [/home/inada-n/work/python/cpython/python] 177 ns +- 1 ns: 1.54x faster
(-35%)

--

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue42126] Optimize BUILD_CONST_KEY_MAP for distinct keys

[Inada Naoki]

Change by Inada Naoki :


--
keywords: +patch
pull_requests: +21842
stage:  -> patch review
pull_request: https://github.com/python/cpython/pull/22911

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com

[issue42126] Optimize BUILD_CONST_KEY_MAP for distinct keys

[Inada Naoki]

New submission from Inada Naoki :

BUILD_CONST_KEY_MAP can be optimized based on #41835 optimization.

1. compiler checks keys tuple is distinct.
2. Add distinct flag to BUILD_CONST_KEY_MAP oparg

To be considered:

* Should we use new opcode, instead of flag in oparg?
* Is this technique safe? Wrong co_consts can make invalid dict instance.

--
assignee: methane
components: Interpreter Core
messages: 379415
nosy: methane
priority: normal
severity: normal
status: open
title: Optimize BUILD_CONST_KEY_MAP for distinct keys
type: performance
versions: Python 3.10

___
Python tracker 

___
___
Python-bugs-list mailing list
Unsubscribe: 
https://mail.python.org/mailman/options/python-bugs-list/archive%40mail-archive.com