spin off of bpo-11549.

http://bugs.python.org/issue11549#msg130955

b) Docstring is now an attribute of Module, FunctionDef and ClassDef, > rather than a first statement. Docstring is a special syntactic
construction, it's not an executable code, so it makes sense to separate it. Otherwise, optimizer would have to take extra care not to introduce, change or remove docstring. For example:

def foo():
"doc" + "string"

Without optimizations foo doesn't have a docstring. After folding, however, the first statement in foo is a string literal. This means that docstring depends on the level of optimizations. Making it an attribute avoids the problem.

I like the change because (IMHO) it makes the code simpler, and becase it also changes the first line of code object. I reviewed the patch: need basic unit tests.

I like this change. Added comments on Rietveld.

Are changes in importlib.h only due to changing first line numbers?

lnotab is changed too.

-    0,0,0,115,12,0,0,0,8,4,4,2,8,8,8,12,
-    8,25,8,13,114,18,0,0,0,99,0,0,0,0,0,0,
+    0,0,0,115,10,0,0,0,12,6,8,8,8,12,8,25,
+    8,13,114,18,0,0,0,99,0,0,0,0,0,0,0,0,

115 is header for bytes type.
next 4 bytes is it's length (little endian, 12->10 bytes)
and everything after that is slided by 2 bytes.

def func(): "doc" + "string"

Currently (Python 2.7-3.6), func.__doc__ is None. I suggest to add an unit test for this corner case, even if the result is going to change in a near future. We need to "specify" the expected behaviour, and make sure that we get the same result if optimizations are enabled or not.

Support adding tests. Tests should cover all cases: module, class, function, coroutine and check also the first line number.

What is the value of co_firstlineno if the function doesn't have any statements?

    def f():
        '''docstring'''

Oh, I misunderstood something.
patched Python 3.7 and system's Python 3.5 shows same output for code below.
I'll check what is actually changed.

inada-n@x250 ~/w/p/ast-docstring> cat -n x.py
1 """module docstring"""
2
3 def func():
4 """func docstring"""
5
6 def func2():
7 """func docstring"""
8 1+1
9
10 print(func.__code__.co_firstlineno)
11 print(func.__code__.co_lnotab)
12 print(func2.__code__.co_firstlineno)
13 print(func2.__code__.co_lnotab)
inada-n@x250 ~/w/p/ast-docstring> ./python x.py
3
b''
6
b'\x00\x02'
inada-n@x250 ~/w/p/ast-docstring> /usr/bin/python3 x.py
3
b''
6
b'\x00\x02'

2017-02-08 10:08 GMT+01:00 INADA Naoki <report@bugs.python.org>:

 6  def func2():
 7      """func docstring"""
 8      1+1

1+1 is replaced with 2 and lone integer literals are removed by the
peephole optimizer. See also the issue bpo-26204.

Oops, I spoke too fast :-) "1+1" is not removed.

"1+1" is replaced with "2" by the peephole optimizer, whereas the compiler ignoring constants comes before the peephole optimizer.

One more time, it would be better to implement constant folding at the AST level ;-)

$ python3
Python 3.5.2 (default, Sep 14 2016, 11:28:32) 
>>> def func():
...  "docstring"
...  1+1
... 
>>> import dis
>>> dis.dis(func)
  3           0 LOAD_CONST               3 (2)
              3 POP_TOP
              4 LOAD_CONST               2 (None)
              7 RETURN_VALUE

This patch affects firstlineno and lnotab of module and class, but not functions.

module:
-<code object <module> at 0x7f053a8f8b70, file "Lib/importlib/_bootstrap.py", line 8>
+<code object <module> at 0x7fdefbf10340, file "Lib/importlib/_bootstrap.py", line 25>
 filename Lib/importlib/_bootstrap.py
-firstlineno: 8
+firstlineno: 25
-lnotab: b'\x04\x11\x04\x02\x08\x08\x08\x07\x04\x02\x04\x03\x10\x04\x0eD\x0e\x15\x0e\x13\x08\x13\x08\x13\x08\x0b\x0e\x08\x08\x0b\x08\x0c\x08\x10\x08$\x0e\x1b\x0ee\x10\x1a\x06\x03\n-\x0e<\x08\x11\x08\x11\x08\x19\x08\x1d\x08\x17\x08\x10\x0eI\x0eM\x0e\r\x08\t\x08\t\n/\x08\x14\x04\x01\x08\x02\x08\x1b\x08\x06\n\x19\x08\x1f\x08\x1b\x12#\x08\x07\x08/'
-  8           0 LOAD_CONST               0 ('Core implementation of import.\n\nThis module is NOT meant to be directly imported! It has been designed such\nthat it can be bootstrapped into Python as the implementation of import. As\nsuch it requires the injection of specific modules and attributes in order to\nwork. One should use importlib as the public-facing version of this module.\n\n')
+lnotab: b'\x04\x00\x04\x02\x08\x08\x08\x07\x04\x02\x04\x03\x10\x04\x0eD\x0e\x15\x0e\x13\x08\x13\x08\x13\x08\x0b\x0e\x08\x08\x0b\x08\x0c\x08\x10\x08$\x0e\x1b\x0ee\x10\x1a\x06\x03\n-\x0e<\x08\x11\x08\x11\x08\x19\x08\x1d\x08\x17\x08\x10\x0eI\x0eM\x0e\r\x08\t\x08\t\n/\x08\x14\x04\x01\x08\x02\x08\x1b\x08\x06\n\x19\x08\x1f\x08\x1b\x12#\x08\x07\x08/'
+ 25           0 LOAD_CONST               0 ('Core implementation of import.\n\nThis module is NOT meant to be directly imported! It has been designed such\nthat it can be bootstrapped into Python as the implementation of import. As\nsuch it requires the injection of specific modules and attributes in order to\nwork. One should use importlib as the public-facing version of this module.\n\n')
               2 STORE_NAME               0 (__doc__)
-
- 25           4 LOAD_CONST               1 (None)
+              4 LOAD_CONST               1 (None)
               6 STORE_GLOBAL             1 (_bootstrap_external)
 

class:
-<code object _ModuleLock at 0x7f053ab61db0, file "Lib/importlib/_bootstrap.py", line 51>
+<code object _ModuleLock at 0x7fdefc61c580, file "Lib/importlib/_bootstrap.py", line 51>
 filename Lib/importlib/_bootstrap.py
 firstlineno: 51
-lnotab: b'\x08\x04\x04\x02\x08\x08\x08\x0c\x08\x19\x08\r'
+lnotab: b'\x0c\x06\x08\x08\x08\x0c\x08\x19\x08\r'
  51           0 LOAD_NAME                0 (__name__)
               2 STORE_NAME               1 (__module__)
               4 LOAD_CONST               0 ('_ModuleLock')
               6 STORE_NAME               2 (__qualname__)
-
- 55           8 LOAD_CONST               1 ('A recursive lock implementation which is able to detect deadlocks\n    (e.g. thread 1 trying to take locks A then B, and thread 2 trying to\n    take locks B then A).\n    ')
+              8 LOAD_CONST               1 ('A recursive lock implementation which is able to detect deadlocks\n    (e.g. thread 1 trying to take locks A then B, and thread 2 trying to\n    take locks B then A).\n    ')
              10 STORE_NAME               3 (__doc__)

So what's new entry may be:

+* ``Module``, ``FunctionDef``, ``AsyncFunctionDef``, and
+ ``ClassDef`` AST nodes now have a new ``docstring`` attribute.
+ The first statement in their body is not considered as a docstring anymore.
+ This affects ``co_firstlineno`` and ``co_lnotab`` attribute of code object
+ for module and class.
+ (Contributed by Eugene Toder and INADA Naoki in :issue:`29463`.)
+

Now I doubt about this patch is really good.

docstring of Module and Class generates two bytecode.
So it's a real, executed statement.

LOAD_CONST 0 ("docstring")
STORE_NAME 0 ("__doc__")

Therefore we loss the possibility to set a breakpoint on the docstring? It doesn't look a great lost.

Just terminology ... strictly speaking what you've done here is "add a *field* to the nodes Module, FunctionDef and ClassDef", rather than add an *attribute* -- that is, when one is consistent with the terms used in the ast module (https://docs.python.org/3/library/ast.html#node-classes) or Wang (https://docs.python.org/devguide/compiler.html#wang97).

Thanks. I don't familiar with language frontend.
I'll check NEWS entry.

This issue broke the ast API. Copy of the following comment on the PR:
#46 (comment)
---
Carreau: Thanks for this ! Improvement to the AST are welcome !

Would it have been possible to make the docstring optional ? (It's already breaking things, like IPython).

Should I comment on upstream bpo ?
---

I created the issue bpo-29622 to fix this.

thank you for your work on the AST, I know many developers are looking forward to improvement and stabilisation with the hope of having it stable (and documented) in the stdlib at some point.

The recent change in PR 46 now change (at least) the constructor of ast.Module to take a second mandatory parameter (the docstring).

I know the ast is autogenerated and "use at your own risk". But IPython for example, use mod = ast.Module([nodes]), with the second mandatory parameter added to Module that make it incompatible with current Python 3.7.
Well it's long until it's released, and we can patch things, but I'm sure we are not the only one in this case, and we'd like older version of IPython to still be compatible with Python 3.7, so if Module()'s second parameter (the docstring) could be optional, that would be great.

I would be happy if it raise a deprecation warning that it will be required in the future.

I'm of course speaking about Module because that's the first error I encountered, but I'm guessing it applies to other changed AST nodes.

Thanks.

OK, let's continue on bpo-29522, and close this issue.

s/ bpo-29522 / bpo-29622 /

New changeset 4c78c52 by INADA Naoki in branch 'master':
bpo-29622: Make AST constructor to accept less than enough number of positional arguments (GH-249)
4c78c52

New changeset cb41b27 by Victor Stinner (INADA Naoki) in branch 'master':
bpo-29463: Add docstring field to some AST nodes. (#46)
cb41b27

New changeset 2641ee5 by Serhiy Storchaka in branch '3.7':
bpo-32911: Revert bpo-29463. (GH-7121)
2641ee5

New changeset 73cbe7a by Serhiy Storchaka in branch 'master':
bpo-32911: Revert bpo-29463. (GH-7121) (GH-7197)
73cbe7a