[python-committers] Adding a (small) feature to 3.4 for Argument Clinic: inspect.Signature supporting simple named constants for default values

[Larry Hastings]

Serhiy Storchaka ran into a ticklish problem with Argument Clinic and 
inspect.Signature information for builtins.

Consider pattern_match() in Modules/_sre.c.  This implements the match 
method on a pattern object; in other words, re.compile().match().  The 
third parameter, endpos, defaults to PY_SSIZE_T_MAX in C.  What should 
inspect.Signature() report as the default value for endpos?  And how 
should it get that value?

Before you answer, consider how inspect.Signature works for builtins.  
Argument Clinic hides a signature for the function as the first line of 
the docstring; the initialization of the code object strips that off and 
puts it in a separate member called __text_signature__.  
inspect.Signature pulls that out string and passes it in to ast.parse, 
then walks the tree it gets back, pulls out the arguments and their 
default values and goes on from there.

We can't turn PY_SSIZE_T_MAX into an integer at Argument Clinic 
preprocessing time, because this could be done on a completely different 
architecture than the computer where Python is running. We can't stuff 
it in at compile time because the macro could devolve into an arbitrary 
expression (with | or + or something) so while that might work here 
that's not a general solution.  We can't do it at runtime becuase the 
docstring is a static string.

The best solution seems to be: allow simple symbolic constants as 
default values.  For example, for endpos we'd use sys.maxsize.  The code 
in inspect.Signature would have to support getting an Attribute node, 
and look up the first field ("sys" in this case) in sys.modules.  You 
could then specify PY_SSIZE_T_MAX as the default for generated C code, 
and life would be a dream.

I've posted a prototype patch on the tracker:

    http://bugs.python.org/issue20144

It need tests and such but I think the basic technology is fine.


The thing is, I feel like this is borderline between bug fix and new 
feature.  But without adding this, we would make a lot of the Argument 
Clinic conversions pretty messy.  So I want to check it in.  I just 
don't want to piss everybody off in the process.

Can you guys live with this?



/arry

p.s.

For what it's worth, the documentation for match() dodges this problem 
by outright lying.  It claims that the prototype for the function is:

     match(string[, pos[, endpos]])

which is a lie.  pattern_match() parses its arguments by calling 
PyArg_ParseTupleAndKeywords() with a format string of "O|nn".  Which 
means, for example, you could call:

     match("abc", endpos=5)

The documentation suggests this is invalid but it works fine.  So my 
feeling is, this is a legitimate problem, and those who came before me 
swept it under the rug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/python-committers/attachments/20140106/bf9706ce/attachment.html>