Clinic was previously pretty confused about what things accepted "bytes", "bytearray", "buffer", "robuffer", and "rwbuffer". This is because the documentation itself was somewhat confusing.

The documentation was recently cleaned up on these in trunk (including one final fix from me this morning). This patch cleans up Clinic's understanding of how the types map to the format units.

My notes on how the format units map to types is copy & pasted below.

--

the documentation for format units is a little confused, and it is therefore confusing

inside:

getbuffer (buffer)
y* s* z*
uses PyObject_GetBuffer(arg, view, PyBUF_SIMPLE)
requires PyBuffer_IsContiguous(view, 'C')
doesn't check specifically for bytes objects
handles bytes objects
accepts bytesarray

convertbuffer (robuffer)
y y# s# z#
handles "read-only buffer object",
all it really does is ensure it doesn't have a pb_releasebuffer,
then call getbuffer() above

so it doesn't accept bytesarray, *just because*

doesn't accept bytes
s z es es#

requires writeable buffer (rwbuffer)
w*

actually checks / specifically handles bytes objects
S

actually checks / specifically handles bytearray objects
Y

actually checks / specifically handles bytes and bytearray objects
c et et#

Huh. Why didn't it attach my patch? Here it is.

Updated the patch a little, to make it enforce compatible "types" parameters. The previous patch had some functions that would permit passing in "types='blah de blah'" or other nonsense.

Added few comments on Rietveld.

New patch revision, including the new API change (the "types" argument to a constructor must now be a set of strings).

The new API change looks doubtful for me. It is more verbose. And less readable. It makes harder to search in sources, because {'str', 'robuffer'} is the same as {'robuffer', 'str'}. If Argument Clinic would a tool with programmatic API, sets would make sense, but for now the main interface of Argument Clinic is text inclusions in C files.

If you want to convert the types parameter into a set, is it possible to continue to accept strings? Many user-friendly APIs accept either a sequence of string names or a string containing space-separated names. For example namedtuple, Enum.

In the case of namedtuple and Enum, the parameter represents a sequence of strings--order is significant. With the 'types' parameter for converters, the internal model was always meant to be a *set* of strings. The order was explicitly *not* significant. So types="str robuffer" and types="robuffer str" should be identical. Your "harder to search in sources" was already true.

Argument Clinic isn't something very many people will learn. Lots of people will read the conversions, but few will ever write a conversion, and they won't do it often and they'll forget the funny details. So we shouldn't make the API quirky and succinct at the expense of readability. The best API will be one that makes the code easier to read.

I think requiring types to be a set of strings makes it easier to read because it makes the semantics obvious. If you see
types="str robuffer"
you aren't sure if order is significant, but if you see
types={'str', 'robuffer'}
you are certain it is not.

Latest patch, with another round of lovely comments from Serhiy.

Many people will see Argument Clinic syntax in declarations in C files. My recent patches (for _io, _ssl and _codecs modules) adds many types=. But I agree that there are many reasons to make this change.

The patch LGTM. You can factorize it more if you like.

New changeset 4f74452a11e5 by Larry Hastings in branch 'default':
Issue bpo-23935: Argument Clinic's understanding of format units
https://hg.python.org/cpython/rev/4f74452a11e5

While cleaning this up I noticed that es# and et# should both require "nullable". Fixed that too. I sure hope it's all correct now!