* error_already_set::what() is now constructed lazily

Prior to this commit throwing error_already_set was expensive due to the
eager construction of the error string (which required traversing the
Python stack). See #1853 for more context and an alternative take on the
issue.

Note that error_already_set no longer inherits from std::runtime_error
because the latter has no default constructor.

* Do not attempt to normalize if no exception occurred

This is not supported on PyPy-2.7 5.8.0.

* Extract exception name via tp_name

This is faster than dynamically looking up __name__ via GetAttrString.
Note though that the runtime of the code throwing an error_already_set
will be dominated by stack unwinding so the improvement will not be
noticeable.

Before:

396 ns ± 0.913 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)

After:

277 ns ± 0.549 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)

Benchmark:

const std::string foo() {
    PyErr_SetString(PyExc_KeyError, "");
    const std::string &s = py::detail::error_string();
    PyErr_Clear();
    return s;
}

PYBIND11_MODULE(foo, m) {
    m.def("foo", &::foo);
}

* Reverted error_already_set to subclass std::runtime_error

* Revert "Extract exception name via tp_name"

The implementation of __name__ is slightly more complex than that.
It handles the module name prefix, and heap-allocated types. We could
port it to pybind11 later on but for now it seems like an overkill.

This reverts commit f1435c7e6b068a1ed13ebd3db597ea3bb15aa398.

* Cosmit following @YannickJadoul's comments

Note that detail::error_string() no longer calls PyException_SetTraceback
as it is unncessary for pretty-printing the exception.

* Fixed PyPy build

* Moved normalization to error_already_set ctor

* Fix merge bugs

* Fix more merge errors

* Improve formatting

* Improve error message in rare case

* Revert back if statements

* Fix clang-tidy

* Try removing mutable

* Does build_mode release fix it

* Set to Debug to expose segfault

* Fix remove set error string

* Do not run error_string() more than once

* Trying setting the tracebackk to the value

* guard if m_type is null

* Try to debug PGI

* One last try for PGI

* Does reverting this fix PyPy

* Reviewer suggestions

* Remove unnecessary initialization

* Add noexcept move and explicit fail throw

* Optimize error_string creation

* Fix typo

* Revert noexcept

* Fix merge conflict error

* Abuse assignment operator

* Revert operator abuse

* See if we still need debug

* Remove unnecessary mutable

* Report "FATAL failure building pybind11::error_already_set error_string" and terminate process.

* Try specifying noexcept again

* Try explicit ctor

* default ctor is noexcept too

* Apply reviewer suggestions, simplify code, and make helper method private

* Remove unnecessary include

* Clang-Tidy fix

* detail::obj_class_name(), fprintf with [STDERR], [STDOUT] tags, polish comments

* consistently check m_lazy_what.empty() also in production builds

* Make a comment slightly less ambiguous.

* Bug fix: Remove `what();` from `restore()`.

It sure would need to be guarded by `if (m_type)`, otherwise `what()` fails and masks that no error was set (see update unit test). But since `error_already_set` is copyable, there is no point in releasing m_type, m_value, m_trace, therefore we can just as well avoid the runtime overhead of force-building `m_lazy_what`, it may never be used.

* Replace extremely opaque (unhelpful) error message with a truthful reflection of what we know.

* Fix clang-tidy error [performance-move-constructor-init].

* Make expected error message less specific.

* Various changes.

* bug fix: error_string(PyObject **, ...)

* Putting back the two eager PyErr_NormalizeException() calls.

* Change error_already_set() to call pybind11_fail() if the Python error indicator not set. The net result is that a std::runtime_error is thrown instead of error_already_set, but all tests pass as is.

* Remove mutable (fixes oversight in the previous commit).

* Normalize the exception only locally in error_string(). Python 3.6 & 3.7 test failures expected. This is meant for benchmarking, to determine if it is worth the trouble looking into the failures.

* clang-tidy: use auto

* Use `gil_scoped_acquire_local` in `error_already_set` destructor. See long comment.

* For Python < 3.8: `PyErr_NormalizeException` before `PyErr_WriteUnraisable`

* Go back to replacing the held Python exception with then normalized exception, if & when needed. Consistently document the side-effect.

* Slightly rewording comment. (There were also other failures.)

* Add 1-line comment for obj_class_name()

* Benchmark code, with results in this commit message.

          function                   #calls  test time [s]  μs / call
master    pure_unwind                729540      1.061      14.539876
          err_set_unwind_err_clear   681476      1.040      15.260282
          err_set_error_already_set  508038      1.049      20.640525
          error_already_set_restore  555578      1.052      18.933288
          pr1895_original_foo        244113      1.050      43.018168
                                                                       PR / master
PR #1895  pure_unwind                736981      1.054      14.295685       98.32%
          err_set_unwind_err_clear   685820      1.045      15.237399       99.85%
          err_set_error_already_set  661374      1.046      15.811879       76.61%
          error_already_set_restore  669881      1.048      15.645176       82.63%
          pr1895_original_foo        318243      1.059      33.290806       77.39%

master @ commit ad146b2a

Running tests in directory "/usr/local/google/home/rwgk/forked/pybind11/tests":
============================= test session starts ==============================
platform linux -- Python 3.9.10, pytest-6.2.3, py-1.10.0, pluggy-0.13.1 -- /usr/bin/python3
cachedir: .pytest_cache
rootdir: /usr/local/google/home/rwgk/forked/pybind11/tests, configfile: pytest.ini
collecting ... collected 5 items

test_perf_error_already_set.py::test_perf[pure_unwind]
PERF pure_unwind,729540,1.061,14.539876
PASSED
test_perf_error_already_set.py::test_perf[err_set_unwind_err_clear]
PERF err_set_unwind_err_clear,681476,1.040,15.260282
PASSED
test_perf_error_already_set.py::test_perf[err_set_error_already_set]
PERF err_set_error_already_set,508038,1.049,20.640525
PASSED
test_perf_error_already_set.py::test_perf[error_already_set_restore]
PERF error_already_set_restore,555578,1.052,18.933288
PASSED
test_perf_error_already_set.py::test_perf[pr1895_original_foo]
PERF pr1895_original_foo,244113,1.050,43.018168
PASSED

============================== 5 passed in 12.38s ==============================

pr1895 @ commit 8dff51d12e4af11aff415ee966070368fe606664

Running tests in directory "/usr/local/google/home/rwgk/forked/pybind11/tests":
============================= test session starts ==============================
platform linux -- Python 3.9.10, pytest-6.2.3, py-1.10.0, pluggy-0.13.1 -- /usr/bin/python3
cachedir: .pytest_cache
rootdir: /usr/local/google/home/rwgk/forked/pybind11/tests, configfile: pytest.ini
collecting ... collected 5 items

test_perf_error_already_set.py::test_perf[pure_unwind]
PERF pure_unwind,736981,1.054,14.295685
PASSED
test_perf_error_already_set.py::test_perf[err_set_unwind_err_clear]
PERF err_set_unwind_err_clear,685820,1.045,15.237399
PASSED
test_perf_error_already_set.py::test_perf[err_set_error_already_set]
PERF err_set_error_already_set,661374,1.046,15.811879
PASSED
test_perf_error_already_set.py::test_perf[error_already_set_restore]
PERF error_already_set_restore,669881,1.048,15.645176
PASSED
test_perf_error_already_set.py::test_perf[pr1895_original_foo]
PERF pr1895_original_foo,318243,1.059,33.290806
PASSED

============================== 5 passed in 12.40s ==============================

clang++ -o pybind11/tests/test_perf_error_already_set.os -c -std=c++17 -fPIC -fvisibility=hidden -Os -flto -Wall -Wextra -Wconversion -Wcast-qual -Wdeprecated -Wnon-virtual-dtor -Wunused-result -isystem /usr/include/python3.9 -isystem /usr/include/eigen3 -DPYBIND11_STRICT_ASSERTS_CLASS_HOLDER_VS_TYPE_CASTER_MIX -DPYBIND11_TEST_BOOST -Ipybind11/include -I/usr/local/google/home/rwgk/forked/pybind11/include -I/usr/local/google/home/rwgk/clone/pybind11/include /usr/local/google/home/rwgk/forked/pybind11/tests/test_perf_error_already_set.cpp

clang++ -o lib/pybind11_tests.so -shared -fPIC -Os -flto -shared ...

Debian clang version 13.0.1-3+build2
Target: x86_64-pc-linux-gnu
Thread model: posix

* Changing call_repetitions_target_elapsed_secs to 0.1 for regular unit testing.

* Adding in `recursion_depth`

* Optimized ctor

* Fix silly bug in recurse_first_then_call()

* Add tests that have equivalent PyErr_Fetch(), PyErr_Restore() but no try-catch.

* Add call_error_string to tests. Sample only recursion_depth 0, 100.

* Show lazy-what speed-up in percent.

* Include real_work in benchmarks.

* Replace all PyErr_SetString() with generate_python_exception_with_traceback()

* Better organization of test loops.

* Add test_error_already_set_copy_move

* Fix bug in newly added test (discovered by clang-tidy): actually use move ctor

* MSVC detects the unreachable return

* change test_perf_error_already_set.py back to quick mode

* Inherit from std::exception (instead of std::runtime_error, which does not make sense anymore with the lazy what)

* Special handling under Windows.

* print with leading newline

* Removing test_perf_error_already_set (copies are under https://github.com/rwgk/rwgk_tbx/commit/7765113fbb659e1ea004c5ba24fb578244bc6cfd).

* Avoid gil and scope overhead if there is nothing to release.

* Restore default move ctor. "member function" instead of "function" (note that "method" is Python terminology).

* Delete error_already_set copy ctor.

* Make restore() non-const again to resolve clang-tidy failure (still experimenting).

* Bring back error_already_set copy ctor, to see if that resolves the 4 MSVC test failures.

* Add noexcept to error_already_set copy & move ctors (as suggested by @skylion007 IIUC).

* Trying one-by-one noexcept copy ctor for old compilers.

* Add back test covering copy ctor. Add another simple test that exercises the copy ctor.

* Exclude more older compilers from using the noexcept = default ctors. (The tests in the previous commit exposed that those are broken.)

* Factor out & reuse gil_scoped_acquire_local as gil_scoped_acquire_simple

* Guard gil_scoped_acquire_simple by _Py_IsFinalizing() check.

* what() GIL safety

* clang-tidy & Python 3.6 fixes

* Use `gil_scoped_acquire` in dtor, copy ctor, `what()`. Remove `_Py_IsFinalizing()` checks (they are racy: https://github.com/python/cpython/pull/28525).

* Remove error_scope from copy ctor.

* Add `error_scope` to `get_internals()`, to cover the situation that `get_internals()` is called from the `error_already_set` dtor while a new Python error is in flight already. Also backing out `gil_scoped_acquire_simple` change.

* Add `FlakyException` tests with failure triggers in `__init__` and `__str__`

THIS IS STILL A WORK IN PROGRESS. This commit is only an important resting point.

This commit is a first attempt at addressing the observation that `PyErr_NormalizeException()` completely replaces the original exception if `__init__` fails. This can be very confusing even in small applications, and extremely confusing in large ones.

* Tweaks to resolve Py 3.6 and PyPy CI failures.

* Normalize Python exception immediately in error_already_set ctor.

For background see: https://github.com/pybind/pybind11/pull/1895#issuecomment-1135304081

* Fix oversights based on CI failures (copy & move ctor initialization).

* Move @pytest.mark.xfail("env.PYPY") after @pytest.mark.parametrize(...)

* Use @pytest.mark.skipif (xfail does not work for segfaults, of course).

* Remove unused obj_class_name_or() function (it was added only under this PR).

* Remove already obsolete C++ comments and code that were added only under this PR.

* Slightly better (newly added) comments.

* Factor out detail::error_fetch_and_normalize. Preparation for producing identical results from error_already_set::what() and detail::error_string(). Note that this is a very conservative refactoring. It would be much better to first move detail::error_string into detail/error_string.h

* Copy most of error_string() code to new error_fetch_and_normalize::complete_lazy_error_string()

* Remove all error_string() code from detail/type_caster_base.h. Note that this commit includes a subtle bug fix: previously error_string() restored the Python error, which will upset pybind11_fail(). This never was a problem in practice because the two PyType_Ready() calls in detail/class.h do not usually fail.

* Return const std::string& instead of const char * and move error_string() to pytypes.h

* Remove gil_scope_acquire from error_fetch_and_normalize, add back to error_already_set

* Better handling of FlakyException __str__ failure.

* Move error_fetch_and_normalize::complete_lazy_error_string() implementation from pybind11.h to pytypes.h

* Add error_fetch_and_normalize::release_py_object_references() and use from error_already_set dtor.

* Use shared_ptr for m_fetched_error => 1. non-racy, copy ctor that does not need the GIL; 2. enables guard against duplicate restore() calls.

* Add comments.

* Trivial renaming of a newly introduced member function.

* Workaround for PyPy

* Bug fix (oversight). Only valgrind got this one.

* Use shared_ptr custom deleter for m_fetched_error in error_already_set. This enables removing the dtor, copy ctor, move ctor completely.

* Further small simplification. With the GIL held, simply deleting the raw_ptr takes care of everything.

* IWYU cleanup

```
iwyu version: include-what-you-use 0.17 based on Debian clang version 13.0.1-3+build2
```

Command used:

```
iwyu -c -std=c++17 -DPYBIND11_TEST_BOOST -Iinclude/pybind11 -I/usr/include/python3.9 -I/usr/include/eigen3 include/pybind11/pytypes.cpp
```

pytypes.cpp is a temporary file: `#include "pytypes.h"`

The raw output is very long and noisy.

I decided to use `#include <cstddef>` instead of `#include <cstdio>` for `std::size_t` (iwyu sticks to the manual choice).

I ignored all iwyu suggestions that are indirectly covered by `#include <Python.h>`.

I manually verified that all added includes are actually needed.

Co-authored-by: Aaron Gokaslan <skylion.aaron@gmail.com>
Co-authored-by: Ralf W. Grosse-Kunstleve <rwgk@google.com>