Without explicitly overriding __reduce_ex__ pickling was failing for
protocols < 2:

    _________________________ test_strings_pickle __________________________

        def test_strings_pickle():
            bs = b("мир")
            us = u("май")

            #from pickletools import dis
            for proto in range(0, pickle.HIGHEST_PROTOCOL):
    >           p_bs = pickle.dumps(bs, proto)

    golang/golang_str_test.py:282:
    _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

    self = b'\xd0\xbc\xd0\xb8\xd1\x80', proto = 0

        def _reduce_ex(self, proto):
    >       assert proto < 2
    E       RecursionError: maximum recursion depth exceeded in comparison

    /usr/lib/python3.9/copyreg.py:56: RecursionError

See added comments for details.

Even though bstr is semantically array of bytes, while ustr is array of
unicode characters, iterating them _both_ yields unicode characters.
This goes in line with Go approach described in "Strings, bytes, runes
and characters in Go"[1] and allows for both ustr _and_ bstr to be used
as strings in unicode world.

Even though this diverges (just a bit) from str/py2 str behaviur, and
diverges more from bytes/py3 behaviour, I have not hit any problem in
practice due to this divergence. In other words the semantics of
bytestring used in Go - to iterate them as unicode characters - is
sound. For the reference it is the authors of Go who originally invented
UTF-8 - see [2] for details.

See also [3] for our discussion with Jérome on this topic.

[1] https://blog.golang.org/strings
[2] https://www.cl.cam.ac.uk/~mgk25/ucs/UTF-8-Plan9-paper.pdf
[3] nexedi/zodbtools!13 (comment 81646)

Implement access to bstr/ustr by [index] and by slice. Result of such
[index] access - similarly to standard str - returns the same bstr/ustr
type with one character:

  - ustr[i] returns ustr with one unicode character taken from i'th character of original string, while
  - bstr[i] returns bstr with one byte taken from i'th byte of original bytestring.

This follows str/unicode semantics on both py2/py3, bytes semantic on
py2, but diverges from bytes semantics on py3. I originally tried to
follow bytes/py3 semantic - for bstr to return an integer instead of
1-byte character, but later found several compatibility breakages due to
it. I contemplated about this divergence for a long time and finally
took decision to follow strings semantics for both ustr and bstr. This
preserves backward compatibility with Python2 and also allows for bstr
to be practically drop-in replacement for str type.

To get an ordinal corresponding to retrieved character, one can use
standard `ord`, e.g. as in `ord(bstr[i])`. This will always return an
integer for all bstr/ustr/str/unicode. Similarly to standard `chr` and
`unichr`, we also provide two utility functions - `uchr` and `bbyte` to
create 1-character and 1-byte ustr/bstr correspondingly.

Verify that it works as expected, and that memoryview(ustr) is rejected,
because ustr is semantically array of unicode characters, not bytes.

No change to the code - just add tests for current status which is
already working as expected.

And to convert them to bstr/ustr decoding buffer data as if it was
bytes. This is needed if e.g. we have data in mmap or numpy.ndarray, and
want to convert the data to string. The conversion is always explicit via
explicit call to b/u. And for bstr/ustr constructors, we preserver their
behaviour to match unicode constructor not to convert automatically, but
instead to stringify the object, e.g. as shown below:

    In [1]: bdata = b'hello 123'

    In [2]: mview = memoryview(bdata)

    In [3]: str(mview)
    Out[3]: '<memory at 0x7fb226b26700>'	# NOTE _not_ b'hello 123'

bytearray was introduced in Python as a mutable version of bytes. It has
all strings methods (e.g. .capitalize() .islower(), etc), and it also
supports % formatting. In other words it has all attributes of being a
byte-string, with the only difference from bytes in that bytearray is
mutable. In other words bytearray is handy to have when a string is
being incrementally constructed step by step without hitting overhead of
many bytes objects creation/destruction.

So, since bytearray is also a bytestring, similarly to bytes, let's add
support to interoperate with bytearray to bstr and ustr:

- b/u and bstr/ustr now accept bytearray as argument and treat it as bytestring.
- bytearray() constructor, similarly to bytes() and unicode()
  constructors, now also accepts bstr/ustr and create bytearray object
  corresponding to byte-stream of input.

For the latter point to work we need to patch bytearray.__init__() a bit,
since, contrary to bytes.__init__(), it does not pay attention to
whether provided argument has __bytes__ method or not.

Both bstr and ustr constructors mimic constructor of unicode(= str on py3) -
an object is either stringified, or decoded if it provides buffer
interface, or the constructor is invoked with optional encoding and
errors argument:

    # py2
    class unicode(basestring)
     |  unicode(object='') -> unicode object
     |  unicode(string[, encoding[, errors]]) -> unicode object

    # py3
    class str(object)
     |  str(object='') -> str
     |  str(bytes_or_buffer[, encoding[, errors]]) -> str

Stringification of all bstr/ustr / unicode/bytes is handled
automatically with the meaning to convert to created type via b or u.

We follow unicode semantic for both ustr _and_ bstr, because bstr/ustr
are intended to be used as strings.

So that e.g. `bstr == <any string type>` works. We want `bstr == ustr`
to work because we intend those types to be interoperable. We also want
e.g. `bstr == "a_string"` to work because we want bstr to be
interoperable with standard strings. In general we want to have full
automatic interoperability with all string types, so that e.g. `bstr == X`
works for X being all bstr, ustr, unicode, bytes (and later bytearray).

For now we add support only for comparison operators. But later, we
will be adding support for e.g. +, string methods, etc - and in all
those operations we will be following the same approach: to have
automatic interoperability with all string types out of the box.

The text added to README reflects this.

The patch to unicode.tp_richcompare on py2 illustrates our approach to
adjust builtin types when absolutely needed. In this particular case
original builtin unicode.__eq__(unicode, bstr) is always returning False
for non-ASCII bstr even despite bstr having .__unicode__() method. Our
adjustment is non-intrusive - we adjust unicode behaviour only wrt bstr
and it stays exactly the same as before wrt all other types.

We anyway do that with care and add a test that verifies that behaviour
of what we patched stays unaffected when used outside of bstr/ustr
context.

_patch_slot(typ, slotname, func) installs func into typ's
dict[slotname]. For example in the next patch we will need to adjust
unicode.__eq__ on py2 not to reject bstr with always assuming that
`unicode == bstr` is False. We will do it via patching unicode.__eq__ to
first check rhs or whether it is bstr and handling that with our code,
while tailing to original unicode.__eq__ for all other types.

Document explicitly which types b/u accept and how they are handled.
Change bstr/ustr docstrings to also be more explicit.

Documentation changes only.

In other words casting to bytes/unicode preserves pygolang string to
remain pygolang string.

Without the changes to bstr/ustr added test fails as e.g.

    >       assert bytes  (bs) is bs
    E       AssertionError: assert b'\xd0\xbc\xd0\xb8\xd1\x80' is b'\xd0\xbc\xd0\xb8\xd1\x80'
    E        +  where b'\xd0\xbc\xd0\xb8\xd1\x80' = bytes(b'\xd0\xbc\xd0\xb8\xd1\x80')

in other words bytes(bstr) was creating a copy and changing type to bytes.

Extend current coverage for b/u tests more explicitly verifying
resulting type (`type(·) is ...` instead of `isinstance(·, ...)`),
verifying unicode(bstr)->ustr and bytes(ustr)->bstr, and str() of both
bstr and ustr.

Move the check for "no custom attributes" from test_qq to generic
test_strings_basic, because now verified string types are publicly
accessible, not only via qq.

Small cosmetics in benchmarks - by reusing hereby introduced xbytes()
utility.

No change for the code itself - the tests just add verification to
current status.

In 2020 in edc7aaab (golang: Teach qq to be usable with both bytes and
str format whatever type qq argument is) I added custom bytes- and
unicode- like types for qq to return instead of str with the idea for
qq's result to be interoperable with both bytes and unicode. Citing that patch:

    qq is used to quote strings or byte-strings. The following example
    illustrates the problem we are currently hitting in zodbtools with
    Python3:

        >>> "hello %s" % qq("мир")
        'hello "мир"'

        >>> b"hello %s" % qq("мир")
        Traceback (most recent call last):
          File "<stdin>", line 1, in <module>
        TypeError: %b requires a bytes-like object, or an object that implements __bytes__, not 'str'

        >>> "hello %s" % qq(b("мир"))
        'hello "мир"'

        >>> b"hello %s" % qq(b("мир"))
        Traceback (most recent call last):
          File "<stdin>", line 1, in <module>
        TypeError: %b requires a bytes-like object, or an object that implements __bytes__, not 'str'

    i.e. one way or another if type of format string and what qq returns do not
    match it creates a TypeError.

    We want qq(obj) to be useable with both string and bytestring format.

    For that let's teach qq to return special str- and bytes- derived types that
    know how to automatically convert to str->bytes and bytes->str via b/u
    correspondingly. This way formatting works whatever types combination it was
    for format and for qq, and the whole result has the same type as format.

    For now we teach only qq to use new types and don't generally expose
    _str and _unicode to be returned by b and u yet. However we might do so
    in the future after incrementally gaining a bit more experience.

So two years later I gained that experience and found that having string
type, that can interoperate with both bytes and unicode, is generally
useful. It is useful for practical backward compatibility with Python2
and for simplicity of programming avoiding constant stream of
encode/decode noise. Thus the day to expose Pygolang string types for
general use has come.

This patch does the first small step: it exposes bytes- and unicode-
like types (now named as bstr and ustr) publicly. It switches b and u to
return bstr and ustr correspondingly instead of bytes and unicode. This
is change in behaviour, but hopefully it should not break anything as
there are not many b/u users currently and bstr and ustr are intended to
be drop-in replacements for standard string types.

Next patches will enhance bstr/ustr step by step to be actually drop-in
replacements for standard string types for real.

See nexedi/zodbtools!13 (comment 81646)
for preliminary discussion from 2019.

See also "Python 3 Losses: Nexedi Perspective"[1] and associated "cost
overview"[2] for related presentation by Jean-Paul from 2018.

[1] https://www.nexedi.com/NXD-Presentation.Multicore.PyconFR.2018?portal_skin=CI_slideshow#/20/1
[2] https://www.nexedi.com/NXD-Presentation.Multicore.PyconFR.2018?portal_skin=CI_slideshow#/20

Since the beginning (9bf03d9c "py.bench: New command to benchmark python
code similarly to `go test -bench`") py.bench was automatically
discovering benchmarks in bench_*.py files only. This was inherited from
wendelin.core which keeps its benchmarks in those files.

However in pygolang, following Go convention(*), we already have several
benchmarks that reside together with tests in same *_test.py files. And
currently just running py.bench does not discover them.

-> Let's fix this and teach py.bench to automatically discover
benchmarks in the test files by default as well.

Pytest's default is to look for tests in test_*.py and *_test.py (+).
Add those patterns and also keep bench_*.py for backward compatibility.

Before this patch running py.bench inside pygolang repository does not
run any benchmark at all. After the patch py.bench runs all the
benchmarks by default:

    (z-dev) kirr@deca:~/src/tools/go/pygolang$ py.bench
    ========================= test session starts ==========================
    platform linux2 -- Python 2.7.18, pytest-4.6.11, py-1.10.0, pluggy-0.13.1
    rootdir: /home/kirr/src/tools/go/pygolang
    plugins: timeout-1.4.2, profiling-1.7.0, mock-2.0.0
    collected 18 items

    pymod: golang/golang_str_test.py
    Benchmarkstddecode              2000000 0.756 µs/op
    Benchmarkudecode                20000   74.359 µs/op
    Benchmarkstdencode              3000000 0.327 µs/op
    Benchmarkbencode                40000   32.613 µs/op

    pymod: golang/golang_test.py
    Benchmarkpyx_select_nogil       500000  2.051 µs/op
    Benchmarkpyx_go_nogil           90000   12.177 µs/op
    Benchmarkpyx_chan_nogil         600000  1.826 µs/op
    Benchmarkgo                     80000   13.267 µs/op
    Benchmarkchan                   500000  2.076 µs/op
    Benchmarkselect                 300000  3.835 µs/op
    Benchmarkdef                    30000000        0.035 µs/op
    Benchmarkfunc_def               40000   29.387 µs/op
    Benchmarkcall                   30000000        0.043 µs/op
    Benchmarkfunc_call              2000000 0.819 µs/op
    Benchmarktry_finally            20000000        0.096 µs/op
    Benchmarkdefer                  600000  1.755 µs/op

    pymod: golang/sync_test.py
    Benchmarkworkgroup_empty        40000   25.807 µs/op
    Benchmarkworkgroup_raise        40000   31.637 µs/op                     [100%]

    =========================== warnings summary ===========================

(*) see https://pkg.go.dev/cmd/go#hdr-Test_packages
(+) see https://docs.pytest.org/en/7.1.x/reference/reference.html#confval-python_files

/reviewed-by @jerome
/reviewed-on nexedi/pygolang!20

We recently moved our custom UTF-8 encoding/decoding routines to Cython.
Now we can start taking speedup advantage on C level to make our own
UTF-8 decoder a bit less horribly slow on py2:

    name       old time/op  new time/op  delta
    stddecode   752ns ± 0%   743ns ± 0%   -1.19%  (p=0.000 n=9+10)
    udecode     216µs ± 0%    75µs ± 0%  -65.19%  (p=0.000 n=9+10)
    stdencode   328ns ± 2%   327ns ± 1%     ~     (p=0.252 n=10+9)
    bencode    34.1µs ± 1%  32.1µs ± 1%   -5.92%  (p=0.000 n=10+10)

So it is ~ 3x speedup for u(), but still significantly slower compared
to std unicode.decode('utf-8').

Only low-hanging fruit here to make _utf_decode_rune a bit more prompt,
since it sits in the most inner loop. In the future
_utf8_decode_surrogateescape might be reworked as well to avoid
constructing resulting unicode via py-level list of py-unicode character
objects. And similarly for _utf8_encode_surrogateescape.

On py3 the performance of std and u/b decode/encode is approximately the same.

/trusted-by @jerome
/reviewed-on nexedi/pygolang!19

Error rune (u+fffd) is returned by _utf8_decode_rune to indicate an
error in decoding. But the error rune itself is valid unicode codepoint:

   >>> x = u"�"
   >>> x
   u'\ufffd'
   >>> x.encode('utf-8')
   '\xef\xbf\xbd'

This way only (r=_rune_error, size=1) should be treated by the caller as
utf8 decoding error.

But e.g. strconv.quote was not careful to also inspect the size, and this way
was quoting � into just "\xef" instead of "\xef\xbf\xbd".
_utf8_decode_surrogateescape was also subject to similar error.

-> Fix it.

Without the fix e.g. added test for strconv.quote fails as

    >           assert quote(tin) == tquoted
    E           assert '"\xef"' == '"�"'
    E             - "\xef"
    E             + "�"

/reviewed-by @jerome
/reviewed-at nexedi/pygolang!18

So that those routines could be just called and do what is expected
without the caller caring whether it is py2 or py3. We will soon need to
use those routines from several callsites, and having that py2/py3
conditioning being spread over all usage places would be inconvenient.

/reviewed-by @jerome
/reviewed-at nexedi/pygolang!18

- Move _utf8_decode_rune, _utf8_decode_surrogateescape, _utf8_encode_surrogateescape out from strconv into _golang_str
- Factor _bstr/_ustr code into pyb/pyu. _bstr/_ustr become plain wrappers over pyb/pyu.
- work-around emerged golang ↔ strconv dependency with at-runtime import.

Moved routines belong to the main part of golang strings processing
-> their home should be in _golang_str.pyx

/reviewed-by @jerome
/reviewed-at nexedi/pygolang!18

We are going to significantly extend py-strings related functionality soon
- to the point where amount of strings related code will be
approximately the same compared to the amount of all other
python-related code inside golang module.

-> First move everything related to py strings to dedicated
_golang_str.pyx as a preparatory step.

Keep that new file included from _golang.pyx instead of being real new
module, because we want strings functionality to be provided by golang
main namespace itself, and to ease internal code interdependencies.

Plain code movement.

/reviewed-by @jerome
/reviewed-at nexedi/pygolang!18

On Python2 without .tp_print printing _pystr crashes as:

    pygolang$ ./golang/testprog/golang_test_str.py
    Traceback (most recent call last):
      File "./golang/testprog/golang_test_str.py", line 39, in <module>
        main()
      File "./golang/testprog/golang_test_str.py", line 34, in main
        print("print(qq(b)):", qq(sb))
    RuntimeError: print recursion

See added comments for details.

Add convenient utility to read whole file and return its content
similarly to Go. The code is taken from wendelin.core:

https://lab.nexedi.com/nexedi/wendelin.core/blob/wendelin.core-2.0.alpha1-18-g38dde766/wcfs/client/wcfs_misc.cpp#L246-281

Provide os/signal package that can be used to setup signal delivery to nogil
channels. This way for user code signal handling becomes regular handling of a
signalling channel instead of being something special or limited to only-main
python thread. The rationale for why we need it is explained below:

There are several problems with regular python's stdlib signal module:

1. Python2 does not call signal handler from under blocked lock.acquire.
   This means that if the main thread is blocked waiting on a semaphore,
   signal delivery will be delayed indefinitely, similarly to e.g. problem
   described in nxdtest!14 (comment 147527)
   where raising KeyboardInterrupt is delayed after SIGINT for many,
   potentially unbounded, seconds until ~semaphore wait finishes.

   Note that Python3 does not have this problem wrt stdlib locks and
   semaphores, but read below for the next point.

2. all pygolang communication operations (channels send/recv, sync.Mutex,
   sync.RWMutex, sync.Sema, sync.WaitGroup, sync.WorkGroup, ...) run with
   GIL released, but if blocked do not handle EINTR and do not schedule
   python signal handler to run (on main thread).

   Even if we could theoretically adjust this behaviour of pygolang at python
   level to match Python3, there are also C++ and pyx/nogil worlds. And we want gil
   and nogil worlds to interoperate (see https://pypi.org/project/pygolang/#cython-nogil-api),
   so that e.g. if completely nogil code happens to run on the main thread,
   signal handling is still possible, even if that signal handling was setup at
   python level.

With signals delivered to nogil channels both nogil world and python
world can setup signal handlers and to be notified of them irregardles
of whether main python thread is currently blocked in nogil wait or not.

/reviewed-on !17

To convert an object to str of current python.
It will be handy to use __pystr when implementing __str__ methods.

/reviewed-on nexedi/pygolang!17

Provide C++ package "os" with File, Pipe, etc similarly to what is
provided on Go side. The package works through IO methods provided by
runtimes.

We need IO facility because os/signal package will need to use
pipe in cooperative IO mode in its receiving-loop goroutine.

os.h and os.cpp are based on drafts from wendelin.core:

https://lab.nexedi.com/nexedi/wendelin.core/blob/wendelin.core-2.0.alpha1-18-g38dde766/wcfs/client/wcfs_misc.h
https://lab.nexedi.com/nexedi/wendelin.core/blob/wendelin.core-2.0.alpha1-18-g38dde766/wcfs/client/wcfs_misc.cpp

/reviewed-on !17

Else as https://github.com/python-greenlet/greenlet/pull/285
demonstrates there can be segmentation faults and crashes due to
exceptions from one greenlet propagating to C stack of another greenlet.

No test here. I've tried to do it, but with gevent (contrary to plain
greenlets), spawning new task only schedules corresponding greenlet to
run in the end of current event loop cycle instead of switching to
created greenlet immediately. With this delaying, it was hard for me to
develop corresponding test in a reasonable time.

Hopefully having the test I've done for greenlet itself + hereby
protection is good enough.

/reviewed-on nexedi/pygolang!17

This package provides special kind of atomic that is automatically reset
to zero after fork in child. This kind of atomic will be used in os
package to implement IO that does not deadlock in Close after fork.

/reviewed-on nexedi/pygolang!17

This package provides wrappers to OS system calls.
Minimal functionality that will be used to implement os and os/signal pacakges.

/reviewed-on nexedi/pygolang!17

Package os will need to access runtime operations to implement IO.

/reviewed-on nexedi/pygolang!17

We will soon need to use common functionality(*) from both runtimes and
other packages. The "other packages" are all C++ and it is handy to keep
common functionality in C++ as well. While we could also maintain
`extern "C"` interface, it duplicates the work. Let's switch everything
to C++ to ease further maintenance.

(*) e.g. package internal/syscall from both runtimes and from package os.

/reviewed-on !17

Without defined std::hash it is not possible to use channels as keys in
dict or set. We will be using set<chan> in os/signal package
implementation.

/reviewed-on nexedi/pygolang!17

I forgot to include it in ad00be70 (libgolang: Introduce runtimes).

/reviewed-on !17

/reviewed-on !17

    $ check-manifest
    lists of files in version control and sdist do not match!
    missing from sdist:
      .nxdtest

/reviewed-on !17

See https://foss.heptapod.net/pypy/pypy/-/issues/3610 for stdin case.
-m started to behave as CPython.

On py3 'ps1' and b'ps1' are different keys and it was failing on PyPy3 as follows:

            # -i after stdin (also tests interactive mode as -i forces interactive even on non-tty)
            d = {
                b'hellopy': b(hellopy),
                b'ps1':     b'' # cpython emits prompt to stderr
            }
            if is_pypy and not is_gpython:
                d['ps1'] = b'>>>> ' # native pypy emits prompt to stdout and >>>> instead of >>>
            _ = pyout(['-i'], stdin=b'import hello\n', cwd=testdata)
    >       assert _ == b"%(ps1)shello\nworld\n['']\n%(ps1)s"           % d
    E       assert b">>>> hello\...\n['']\n>>>> " == b"hello\nworld\n['']\n"
    E         At index 0 diff: b'>' != b'h'
    E         Full diff:
    E         - b"hello\nworld\n['']\n"
    E         + b">>>> hello\nworld\n['']\n>>>> "
    E         ?   +++++                    +++++

    gpython/gpython_test.py:200: AssertionError

Fix it.
Amends e205dbf6 (gpython: Implement -i (interactive inspect after program run) + promised interactive-mode tests)

When _BaseCtx is setting up cancel propagation it locks a parent,
checks for parent.err != nil, and, if it is, calls
ctx._cancel(parent.err) _with_ _holding_ parent.mu. Since _cancel
internally also goes through parents and locks them, this was deadlocking
on the second call to parent.mu.lock().

-> Fix it by calling ctx._cancel(err) in the constructor outside of
parent lock.

The bug was there from the beginning - from e9567c7b (context: New
package that mirrors Go's context).

/trusted-by @jerome
/reviewed-on nexedi/pygolang!16

- _propagateCancel is used only in _BaseCtx constructor -> inline it
  there. Being run in the constructor makes it clear that this code
  works on new _BaseCtx object with empty set of children.

- since _cancelFrom interacts with the code moved from _propagateCancel,
  also move it to be close to cancel propagation setup.

No functional changes, just plain code movement.

/trusted-by @jerome
/reviewed-on !16

The function that verifies "no exception/panic" case was prepared but not called at all.
This oversight was there from 5146eb0b "Add support for defer & recover".