[issue32477] Move jumps optimization from the peepholer to the compiler

[Serhiy Storchaka]

Change by Serhiy Storchaka :


--
versions: +Python 3.9 -Python 3.8

___
Python tracker 

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

[issue32477] Move jumps optimization from the peepholer to the compiler

[Serhiy Storchaka]

Serhiy Storchaka  added the comment:

PR 14116 is based on the part of PR 5077. It serves three functions:

* Finally fixes issue1875.

* Simplifies the code. Removes special cases for "if 0" and "while 1". 31 
insertions, 80 deletions in Python/compile.c + Python/peephole.c.

* However such optimization is now applied in more general cases, not just for 
"if" and "while".

--

___
Python tracker 

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

[issue32477] Move jumps optimization from the peepholer to the compiler

[Serhiy Storchaka]

Change by Serhiy Storchaka :


--
pull_requests: +13966
pull_request: https://github.com/python/cpython/pull/14116

___
Python tracker 

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

[issue32477] Move jumps optimization from the peepholer to the compiler

[Serhiy Storchaka]

Serhiy Storchaka  added the comment:

Opened issue37213 for the regression in the peepholer.

--

___
Python tracker 

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

[issue32477] Move jumps optimization from the peepholer to the compiler

[Serhiy Storchaka]

Serhiy Storchaka  added the comment:

After analyzing the difference between bytecodes generated by the current 
peepholer and the new optimizer I have found that he peepholer do not optimizes 
jumps in case of some multiline expressions. For example:

[x
 for x in a if x]

  1   0 BUILD_LIST   0
  2 LOAD_FAST0 (.0)
>>4 FOR_ITER12 (to 18)

  2   6 STORE_FAST   1 (x)
  8 LOAD_FAST1 (x)
 10 POP_JUMP_IF_FALSE   16

  1  12 LOAD_FAST1 (x)
 14 LIST_APPEND  2
>>   16 JUMP_ABSOLUTE4
>>   18 RETURN_VALUE

if x:
if (y and
z):
foo()
else:
bar()

  1   0 LOAD_NAME0 (x)
  2 POP_JUMP_IF_FALSE   20

  2   4 LOAD_NAME1 (y)
  6 POP_JUMP_IF_FALSE   18

  3   8 LOAD_NAME2 (z)

  2  10 POP_JUMP_IF_FALSE   18

  4  12 LOAD_NAME3 (foo)
 14 CALL_FUNCTION0
 16 POP_TOP
>>   18 JUMP_FORWARD 6 (to 26)

  6 >>   20 LOAD_NAME4 (bar)
 22 CALL_FUNCTION0
 24 POP_TOP
>>   26 LOAD_CONST   0 (None)
 28 RETURN_VALUE

It preserves jumps to jump instructions. I do not know the cause. All works 
with single-line expressions.

The new optimizer does not have this bug.

--

___
Python tracker 

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

[issue32477] Move jumps optimization from the peepholer to the compiler

[Serhiy Storchaka]

Serhiy Storchaka  added the comment:

Simplified PR 5077. It uses now NEXT_BLOCK() to guarantee that a new block is 
always used after jumps. NEXT_BLOCK() was defined, but not used before.

bpo-35193 and bpo-9566 demonstrate that the code of peephole.c is complex and 
error-prone. Moving it to the upper level makes it more error-proof and general.

--
versions: +Python 3.8 -Python 3.7

___
Python tracker 

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

[issue32477] Move jumps optimization from the peepholer to the compiler

[Antoine Pitrou]

Antoine Pitrou  added the comment:

Well, on the other hand, the change you're proposing is hardly necessary, 
unless it actually demonstrates a noticeable improvement on some workload.  So 
perhaps we should leave the peepholer alone until someone has a better idea.

--

___
Python tracker 

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

[issue32477] Move jumps optimization from the peepholer to the compiler

[Serhiy Storchaka]

Serhiy Storchaka  added the comment:

> I think it's valuable to have a separate C module for optimizations, instead 
> of cramming them in compiler.c.

This would require major rewriting. PR 5077 adds just two functions in 
compile.c. But they use 4 structures defined locally in this file. If move 
these two functions in a separate file we will need to move these 4 structures 
and related definitions to the common header file. And maybe it will be in vain 
if once all these optimizations will be moved to other parts of the compiler. 
Some of jumps optimization already are performed on instructions emitting 
stage. Dead code elimination perhaps could be performed at the stage of 
assembling the bytecode from basic blocks. Constant tuples folding could be 
implemented by at least three ways besides the current implementation, I'm 
testing what of them is the simplest.

--

___
Python tracker 

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

[issue32477] Move jumps optimization from the peepholer to the compiler

[Antoine Pitrou]

Antoine Pitrou  added the comment:

> But some peephole optimizations will work with the intermediate fixed-size 
> representations used by the compiler, before generating the concrete bytecode.

Then perhaps the goal should be to make the peephole operate on that 
intermediate representation?

I think it's valuable to have a separate C module for optimizations, instead of 
cramming them in compiler.c.

--

___
Python tracker 

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

[issue32477] Move jumps optimization from the peepholer to the compiler

[Serhiy Storchaka]

Serhiy Storchaka  added the comment:

Actually moving any of remaining optimizations (including two optimizations 
that are moved by this issue and two other that are left) slightly increases 
the complexity. But instead the optimizations become more general. At the end, 
after moving all optimizations, we could drop the auxiliary code in the 
peepholer (building the table of basic blocks, removing NOPs, fixing up jump 
targets) and the net complexity will be reduced. This can also reduce the time 
and memory consumption of compilation.

Thus getting rid of the peepholer optimizer working with the bytecode is my 
goal. But some peephole optimizations will work with the intermediate 
fixed-size representations used by the compiler, before generating the concrete 
bytecode.

--

___
Python tracker 

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

[issue32477] Move jumps optimization from the peepholer to the compiler

[Raymond Hettinger]

Raymond Hettinger  added the comment:

> This is a step to getting rid of the peepholer.

There is no goal to get rid of the peepholer optimizer.  Please don't invent 
this as a requirement.

If some optimization are more easily performed upstream, then go ahead and move 
them upstream.  The goal is to have a net reduction in complexity, not just the 
elimination of peephole.c just because you decided it should disappear.

--
nosy: +rhettinger

___
Python tracker 

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

[issue32477] Move jumps optimization from the peepholer to the compiler

[Serhiy Storchaka]

Change by Serhiy Storchaka :


--
nosy: +inada.naoki, pitrou

___
Python tracker 

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

[issue32477] Move jumps optimization from the peepholer to the compiler

[Serhiy Storchaka]

Change by Serhiy Storchaka :


--
keywords: +patch
pull_requests: +4952
stage:  -> patch review

___
Python tracker 

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

[issue32477] Move jumps optimization from the peepholer to the compiler

[Serhiy Storchaka]

New submission from Serhiy Storchaka :

The proposed patch moves jumps optimization from the peepholer to the compiler. 
The optimization is performed for lists of instructions before generating 
concrete bytecode and is not restricted by the size of bytecode instructions. 
It allows to optimize more long chains of jumps.

This is a step to getting rid of the peepholer.

--
components: Interpreter Core
messages: 309351
nosy: benjamin.peterson, brett.cannon, ncoghlan, serhiy.storchaka, yselivanov
priority: normal
severity: normal
status: open
title: Move jumps optimization from the peepholer to the compiler
type: performance
versions: Python 3.7

___
Python tracker 

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