Modify the _struct module to use FASTCALL and Argument Clinic

[vstinner]

BPO	29300
Nosy	@vstinner, @larryhastings, @methane, @skrah, @vadmium, @serhiy-storchaka
Files	struct_fastcall.patch struct_fastcall-2.patch struct_fastcall-3.patch struct_fastcall-4.patch struct_fastcall-5.patch struct_fastcall-6.patch struct_fastcall-7.patch cache_struct.patch unpack_buffer.patch cache_struct-2.patch

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = <Date 2017-02-10.11:33:37.652>
created_at = <Date 2017-01-17.16:31:45.553>
labels = ['3.7', 'expert-argument-clinic', 'performance']
title = 'Modify the _struct module to use FASTCALL and Argument Clinic'
updated_at = <Date 2017-02-10.11:33:37.651>
user = 'https://github.com/vstinner'

bugs.python.org fields:

activity = <Date 2017-02-10.11:33:37.651>
actor = 'vstinner'
assignee = 'none'
closed = True
closed_date = <Date 2017-02-10.11:33:37.652>
closer = 'vstinner'
components = ['Argument Clinic']
creation = <Date 2017-01-17.16:31:45.553>
creator = 'vstinner'
dependencies = []
files = ['46319', '46422', '46431', '46432', '46434', '46436', '46488', '46489', '46490', '46492']
hgrepos = []
issue_num = 29300
keywords = ['patch']
message_count = 48.0
messages = ['285663', '285669', '286326', '286327', '286329', '286346', '286348', '286349', '286350', '286351', '286357', '286358', '286364', '286367', '286373', '286645', '286659', '286712', '286756', '286761', '286766', '286767', '286768', '286770', '286771', '286772', '286773', '286774', '286775', '286776', '286781', '286782', '286783', '286784', '286785', '286786', '286789', '286790', '286791', '286792', '286795', '286799', '286800', '286834', '286934', '286943', '287125', '287517']
nosy_count = 7.0
nosy_names = ['vstinner', 'larry', 'methane', 'skrah', 'python-dev', 'martin.panter', 'serhiy.storchaka']
pr_nums = []
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'performance'
url = 'https://bugs.python.org/issue29300'
versions = ['Python 3.7']

[vstinner]

Attached patch modify the _struct module to use FASTCALL and Argument Clinic.

AC requires a summary line, so I duplicated the first sentence of iter_unpack() docstring:

iter_unpack(fmt, buffer, /)
    Return an iterator yielding tuples unpacked from the given bytes.
    
    Return an iterator yielding tuples unpacked from the given bytes
    source according to the format string, like a repeated invocation of
    unpack_from().  Requires that the bytes length be a multiple of the
    format struct size.

[serhiy-storchaka]

Why not just convert the struct module to Argument Clinic?

[vstinner]

Patch version 2 converts most functions and methods to Argument Clinic, except of unpack() and unpack_into() which require "*args" support in Argument Clinic: see the issue bpo-20291.

[vstinner]

I reformatted docstrings, and wrote a summary line.

[serhiy-storchaka]

It looks to me that the Py_buffer converter can be used for all the buffer parameter and special purposed converter can be used for the fmt parameter.

[vstinner]

Patch version 3:

• Use also Argument Clinic for Struct.__init__()
• Rename fmt to format: in the code and docstring. By the way, Struct docstring was wrong: Struct() accepts a 'format' keyword argument, not 'fmt'
• Use Py_buffer converter for the buffer parameter, except for iter_unpack()
• Document that unpack_from() now accepts keywords and write an unit test
• Fix Struct.__init__() error message: the method accepts also Unicode
• Document that Struct accepts str encodable to ASCII and bytes

I would be nice to use Py_buffer format for the "buffer" argument of iter_unpack(), but Argument Clinic calls PyBuffer_Release(&buffer), whereas the buffer should stay alive after the function exit. Is there a way to clone/copy a Py_buffer?

For the fmt/format argument: s_init() uses custom code to accept Unicode encodable to ASCII and bytes objects. Using Argument Clinic to check fmt type would require to write a converter. I would prefer to not make too many changes in a single patch, and I don't know how to write such converter. I suggest to do that later and keep the existing code. It seems like all functions getting a format ends in s_init() to check the format argument.

[serhiy-storchaka]

The converter for the format parameter already exists. It is cache_struct(). You just need to change its interface.

• Fix Struct.__init__() error message: the method accepts also Unicode
• Document that Struct accepts str encodable to ASCII and bytes

I think this should be backported to older versions. Maybe there is an open issue for this.

I would be nice to use Py_buffer format for the "buffer" argument of iter_unpack(), but Argument Clinic calls PyBuffer_Release(&buffer), whereas the buffer should stay alive after the function exit. Is there a way to clone/copy a Py_buffer?

I agreed that the Py_buffer converter is not applicable here. Just object is good.

[vstinner]

The converter for the format parameter already exists. It is cache_struct(). You just need to change its interface.

The code to convert the format string is in s_init(). The code increases the reference counter, I don't see how to produce the Py_DECREF() code.

Again, I would prefer to do that later. The change is already big enough :-)

I think this should be backported to older versions.

Ok, will do once this change is accepted.

[serhiy-storchaka]

static int
cache_struct_converter(PyObject *arg, PyObject **s_object)
{
    if (arg == NULL) {
        Py_DECREF();
        return 1;
    }
    *s_object = cache_struct(arg); // actually inline this
    if (*s_object == NULL)
        return 0;
    return Py_CLEANUP_SUPPORTED;
}

[serhiy-storchaka]

See also ascii_buffer_converter in Modules/binascii.c for example.

[vstinner]

Thank you for your reviews Serhiy :-) Here is a new patch to take in account your latest comments.

I didn't change the API anymore: Struct.unpack_from() accepts keywords, unpack_from() doesn't.

[vstinner]

See also ascii_buffer_converter in Modules/binascii.c for example.

Nice. I didn't know the "def cleanup" feature. I used it in my patch version 4.

[serhiy-storchaka]

I said not about the converter that converts format to bytes (it is needed only in Struct.__init__), but about the converter that converts format to the Struct object. It is used in all module-level functions.

[vstinner]

New patch version 5 without the "format_converter".

Serhiy Storchaka: I said not about the converter that converts format to bytes (it is needed only in Struct.__init__),

For me, it makes sense to check format type earlier than Struct.__init__(), but since you like the current code, I'm also ok to leave this part of the code unchanged :-)

Serhiy: but about the converter that converts format to the Struct object. It is used in all module-level functions.

Sorry, but I dislike this change. Even if multiple functions do the same, for me creating a Struct object is not something part of the argument parsing. I prefer to keep the following 3 lines in each C function *body* to ease debugging:

    PyObject *s_object = cache_struct(format);
    if (s_object == NULL)
        return NULL;

IMHO my struct_fastcall-5.patch is already big enough. I prefer incremental changes.

I have no strong opinion on cache_struct(), just a matter of taste :-) And others may like your change!

Feel free to write your own patch on top of mine, once mine is merged ;-)

[vstinner]

Updated patch, version 6:

• Rename "obj" variable to "iter"
• Fix unpack_from() signature: only the first parameter is positional only

[vstinner]

Serhiy: Would you mind to review my latest patch, struct_fastcall-6.patch? It should address all your remarks, except your proposal to write a converter for cache_struct(): see my previous comment about this idea. Thanks in advance ;)

[serhiy-storchaka]

It looks to me that the type of the self parameter can be changed from PyObject* to PyStructObject*. This will make the patch larger but the final code simpler.

class Struct "PyStructObject *" "&PyStructType"

[vstinner]

It looks to me that the type of the self parameter can be changed from PyObject* to PyStructObject*. This will make the patch larger but the final code simpler.

I like the idea, but I prefer to do in a separated change, once the first big one is merged. Otherwise, the first patch will do too many unrelated things at the same time IMHO.

[serhiy-storchaka]

This doesn't make the patch much bigger. Usually such kind of changes are made in one patch.

Here is a patch that also changes the type of the self parameter to PyStructObject*.

struct_fastcall-6.patch LGTM, but I prefer struct_fastcall-7.patch.

[vstinner]

• if (PyObject_GetBuffer(input, &vbuf, PyBUF_SIMPLE) < 0)

•    return NULL;
  

Oh by the way, I forgot to mention a subtle change. PyObject_GetBuffer(PyBUF_SIMPLE) is less strict that PyArg_Parse("y#") / "buffer" converter of Argument Clinic: getargs.c also checks that the buffer is contiguous, extract of getbuffer():

    if (!PyBuffer_IsContiguous(view, 'C')) {
        PyBuffer_Release(view);
        *errmsg = "contiguous buffer";
        return -1;
    }

I don't know well the buffer protocol. I don't know any object which provide a non-contiguous buffer. At least, I can say that the last time I looked at this dark part of Python, the documentation was between tiny and non-existent :-/ The buffer protocol is complex but not well documented :-(

[python-dev]

New changeset f3ff4a3ce77c by Victor Stinner in branch 'default':
Issue bpo-29300: Convert _struct module to Argument Clinic
https://hg.python.org/cpython/rev/f3ff4a3ce77c

[vstinner]

I also prefer Serhiy's struct_fastcall-7.patch over my struct_fastcall-6.patch.

Serhiy Storchaka: "This doesn't make the patch much bigger. Usually such kind of changes are made in one patch."

Well, it's just that you prefer to reduce the number of commits and so fold tiny changes into a single big commit, whereas I now prefer multiple tiny commits. I have not strong preferences between the two ways to commit, so I pushed your patch.

Here is the full commit message, since Roundbot bot only shows the first line:
---
Issue bpo-29300: Convert _struct module to Argument Clinic

• The struct module now requires contiguous buffers.
• Convert most functions and methods of the _struct module to Argument Clinic
• Use "Py_buffer" type for the "buffer" argument. Argument Clinic is
  responsible to create and release the Py_buffer object.
• Use "PyStructObject *" type for self to avoid explicit conversions.
• Add an unit test on the _struct.Struct.unpack_from() method to test passing
  arguments as keywords.
• Rephrase docstrings.
• Rename "fmt" argument to "format" in docstrings and the documentation.

As a side effect, functions and methods which used METH_VARARGS calling
convention like struct.pack() now use the METH_FASTCALL calling convention
which avoids the creation of temporary tuple to pass positional arguments and
so is faster. For example, struct.pack("i", 1) becomes 1.56x faster (-36%)::

    $ ./python -m perf timeit \
        -s 'import struct; pack=struct.pack' 'pack("i", 1)' \
        --compare-to=../default-ref/python
    Median +- std dev: 119 ns +- 1 ns -> 76.8 ns +- 0.4 ns: 1.56x faster (-36%)
    Significant (t=295.91)

Patch co-written with Serhiy Storchaka.
---

So yeah, the "side effect" is that struct.pack("i", 1) becomes 1.56x faster (-36%). Ok, maybe it was my main goal ;-) I also mentioned the "new" (?) contiguous requirement on buffers.

Thanks Serhiy for your multiple reviews, very useful as usual. I like the final result: better documentation, better docstrings and better code!

[vadmium]

Shouldn’t the top-level unpack() parameter be called “buffer” like the other functions and methods, not “inputstr”?

[serhiy-storchaka]

Propose for your consideration a patch that uses Argument Clinic for getting possible cached struct object from the format. This simplifies the implementation of module level functions.

[vadmium]

FYI Victor, you can make non-C-contiguous buffers by slicing memoryview:

>>> struct.unpack(">L", memoryview(b"1234")[::-1])
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
BufferError: memoryview: underlying buffer is not C-contiguous

Can also use the built-in _testbuffer module to create stranger buffers.

[vstinner]

Martin> Shouldn’t the top-level unpack() parameter be called “buffer” like the other functions and methods, not “inputstr”?

Hum. I reopen the issue.

Attached patch renames unpack() "inputstr" argument to "buffer" and uses the Py_buffer type for it. I had to fix unit tests which passed str instead of bytes to buffer. Before this error was missed because the unit test is written to test the format string value, not the type of arguments.

[serhiy-storchaka]

Oh by the way, I forgot to mention a subtle change.
PyObject_GetBuffer(PyBUF_SIMPLE) is less strict that PyArg_Parse("y#") /
"buffer" converter of Argument Clinic: getargs.c also checks that the
buffer is contiguous, extract of getbuffer():

We already made such changes in the past. The difference is subtle and I have
doubts that choosing any of ways was deliberate.

[serhiy-storchaka]

So yeah, the "side effect" is that struct.pack("i", 1) becomes 1.56x faster
(-36%). Ok, maybe it was my main goal ;-) I also mentioned the "new" (?)
contiguous requirement on buffers.

struct.pack() always was faster than int.to_bytes(). I wanted to speed up
int.to_bytes(), and after converting to Argument Clinic in bpo-20185 it have
became faster than struct.pack(). But after converting the struct module to
Argument Clinic struct.pack() is faster than int.to_bytes() again! Now I need
to find other ways to make int.to_bytes() even faster to win this chase.

[serhiy-storchaka]

unpack_buffer.patch LGTM. Please backport test changes and other changes discussed before to other branches.

[vstinner]

Serhiy Storchaka: "But after converting the struct module to Argument Clinic struct.pack() is faster than int.to_bytes() again!"

Sorry about that ;-)

Serhiy Storchaka: "Now I need to find other ways to make int.to_bytes() even faster to win this chase."

I ran a microbenchmark:

$ ./python -m perf timeit -s 'to_bytes=int.to_bytes' 'to_bytes(1, 4, "little")'

Reference: ~154 ns

Replace int_to_bytes_impl() body with:
PyBytes_FromStringAndSize("1", 1)
=> ~120 ns (-34 ns)

Replace int_to_bytes() body with:
return int_to_bytes_impl(self, 4, NULL, 1);
=> ~76 ns (-44 ns)

_PyArg_ParseStackAndKeywords() with _PyArg_Parser{"nU|$p:to_bytes"} takes 44 ns on a total of 154 ns. 29% of the runtime is spent on parsing arguments.

If you want to optimize further int.to_bytes(), IMHO we should explore the issue bpo-29419: "Argument Clinic: inline PyArg_UnpackTuple and PyArg_ParseStack(AndKeyword)?".

[python-dev]

New changeset 32380d41e788 by Victor Stinner in branch '3.5':
Issue bpo-29300: test_struct tests unpack_from() with keywords
https://hg.python.org/cpython/rev/32380d41e788

[python-dev]

New changeset faa1e4f4b156 by Victor Stinner in branch 'default':
Rename struct.unpack() 2nd parameter to "buffer"
https://hg.python.org/cpython/rev/faa1e4f4b156

[vstinner]

Serhiy Storchaka: "We already made such changes in the past. The difference is subtle and I have
doubts that choosing any of ways was deliberate."

Right. IMHO it's safe to make sure that the buffer is contiguous. I'm quite sure that the code doesn't support non-contiguous buffers.

Serhiy Storchaka: "Please backport test changes and other changes discussed before to other branches."

Done. Sorry, I forgot this part.

Serhiy Storchaka: "unpack_buffer.patch LGTM."

Thanks for the review, it's now merged. It was a regression in unpack() docstring, Python 3.5 docstring contains "unpack(fmt, buffer)".

[vstinner]

Martin Panter: """FYI Victor, you can make non-C-contiguous buffers by slicing memoryview:

>>> struct.unpack(">L", memoryview(b"1234")[::-1])
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
BufferError: memoryview: underlying buffer is not C-contiguous"""

Oh, it means that the Argument Clinic change doesn't add new checks on the buffer? In Python 3.6, memory_getbuf() raises an exception in the following code:

    if (!REQ_STRIDES(flags)) {
        if (!MV_C_CONTIGUOUS(baseflags)) {
            PyErr_SetString(PyExc_BufferError,
                "memoryview: underlying buffer is not C-contiguous");
            return -1;
        }
        view->strides = NULL;
    }

I undersrtand that memory_getbuf() is smart enough to raise an exception becaues the buffer is not contiguous. But a weaker implementation of getbuffer may not implement such check, whereas getbuffer() double check that the buffer is C-contiguous. Am I right?

[vstinner]

I like the overall cache_struct.patch change, but I have questions: see my review.

[skrah]

PyBUF_SIMPLE implies C-contiguous for a conforming buffer provider.

[vstinner]

Stefan Krah: "PyBUF_SIMPLE implies C-contiguous for a conforming buffer provider."

Oh ok, thanks for the confirmation. So my change didn't add new checks and my commit message is wrong :-) Only non-conforming objects are impacted, but in such case, it's more a bug in the object than a bug in struct. I guess that non-conforming non-contiguous objects are very rare in the wild ;-)

The new code is more strict, so is safe in all corner cases.

[serhiy-storchaka]

Dear Roundup Robot, please don't close issues.

[vstinner]

Serhiy Storchaka: "Dear Roundup Robot, please don't close issues."

For what it's worth, I reported issues of this robot to python-dev:

https://mail.python.org/pipermail/python-dev/2017-February/147317.html

https://mail.python.org/pipermail/python-dev/2017-February/147325.html

[vstinner]

cache_struct-2.patch LGTM.

[python-dev]

New changeset 9245894af223 by Serhiy Storchaka in branch 'default':
Issue bpo-29300: Use Argument Clinic for getting struct object from the format.
https://hg.python.org/cpython/rev/9245894af223

[python-dev]

New changeset e21cda70a3a13eb6e6238e436a5c0e2c4e4bebef by Serhiy Storchaka in branch 'master':
Issue bpo-29300: Use Argument Clinic for getting struct object from the format.
e21cda7

[vstinner]

@serhiy: Can we please now close this issue? Or is there still something to do?

[vstinner]

I'm not aware of any pending issues, buildbots are happy, I'm happy, I close the issue :-) Don't hesitate to reopen it if I missed something.

[python-dev]

New changeset e552e185f3087204d326409e8631eb33dd0e7958 by Victor Stinner in branch 'master':
Issue bpo-29300: Convert _struct module to Argument Clinic
e552e18

[python-dev]

New changeset a88eb4fa9e7fdf1a1050786223044c6bb7949784 by Victor Stinner in branch 'master':
Issue bpo-29300: test_struct tests unpack_from() with keywords
a88eb4f

New changeset dcd4b1af2c59b0aae33cbac00d9f6fb47782ac57 by Victor Stinner in branch 'master':
Rename struct.unpack() 2nd parameter to "buffer"
dcd4b1a

[python-dev]

New changeset a88eb4fa9e7fdf1a1050786223044c6bb7949784 by Victor Stinner in branch '3.5':
Issue bpo-29300: test_struct tests unpack_from() with keywords
a88eb4f

[python-dev]

New changeset a88eb4fa9e7fdf1a1050786223044c6bb7949784 by Victor Stinner in branch '3.6':
Issue bpo-29300: test_struct tests unpack_from() with keywords
a88eb4f