gh-119396: Optimize PyUnicode_FromFormat() UTF-8 decoder

[vstinner]

Add unicode_decode_utf8_writer() to write directly characters into a _PyUnicodeWriter writer: avoid the creation of a temporary string. Optimize PyUnicode_FromFormat() by using the new
unicode_decode_utf8_writer().

Rename unicode_fromformat_write_cstr() to
unicode_fromformat_write_utf8().

Microbenchmark on the code:

return PyUnicode_FromFormat(
    "%s %s %s %s %s.",
    "format", "multiple", "utf8", "short", "strings");

Result: 620 ns +- 8 ns -> 382 ns +- 2 ns: 1.62x faster.

• Issue: [C API] Add an efficient public PyUnicodeWriter API #119182

• Issue: Optimize _PyUnicodeWriter implementation #119396

[vstinner]

Benchmark:

diff --git a/Modules/_testcapimodule.c b/Modules/_testcapimodule.c
index f99ebf0dde..0752b2b1d2 100644
--- a/Modules/_testcapimodule.c
+++ b/Modules/_testcapimodule.c
@@ -3312,6 +3312,14 @@ function_set_warning(PyObject *Py_UNUSED(module), PyObject *Py_UNUSED(args))
     Py_RETURN_NONE;
 }
 
+static PyObject *
+bench(PyObject *Py_UNUSED(module), PyObject *Py_UNUSED(args))
+{
+    return PyUnicode_FromFormat(
+        "%s %s %s %s %s.",
+        "format", "multiple", "utf8", "short", "strings");
+}
+
 static PyMethodDef TestMethods[] = {
     {"set_errno",               set_errno,                       METH_VARARGS},
     {"test_config",             test_config,                     METH_NOARGS},
@@ -3454,6 +3462,7 @@ static PyMethodDef TestMethods[] = {
     {"check_pyimport_addmodule", check_pyimport_addmodule, METH_VARARGS},
     {"test_weakref_capi", test_weakref_capi, METH_NOARGS},
     {"function_set_warning", function_set_warning, METH_NOARGS},
+    {"bench", bench, METH_NOARGS},
     {NULL, NULL} /* sentinel */
 };

Command:

./python -m venv env
env/bin/python -m pip install pyperf
env/bin/python -m pyperf timeit -s 'import _testcapi; func=_testcapi.bench' 'func()' -v -o ref.json

Result, Python built with gcc -O3:

620 ns +- 8 ns -> 382 ns +- 2 ns: 1.62x faster

[vstinner]

Oh, there was a performance regression on b"abc".decode(): I fixed it.

Benchmark:

import pyperf
import _testcapi
runner = pyperf.Runner()

utf8 = b'abc'
runner.bench_func('abc', utf8.decode)

utf8 = 'abcé'.encode()
runner.bench_func('abc + UTF-8', utf8.decode)

utf8 = 'éabc'.encode()
runner.bench_func('UTF-8 + abc', utf8.decode)

utf8 = b'x' * (1024 * 1024)
runner.bench_func('ASCII 1 MiB', utf8.decode)

utf8 = ('x' * (1024 * 1024) + 'é').encode()
runner.bench_func('ASCII 1 MiB + UTF-8', utf8.decode)

utf8 = ('é' + 'x' * (1024 * 1024)).encode()
runner.bench_func('UTF-8 + ASCII 1 MiB', utf8.decode)

utf8 = ('€' + 'x' * (1024 * 1024)).encode()
runner.bench_func('UTF-8 euro + ASCII 1 MiB', utf8.decode)

Results, Python built with gcc -O3, CPU isolation.

+---------------------+---------+-----------------------+
| Benchmark           | ref     | change                |
+=====================+=========+=======================+
| abc                 | 73.7 ns | 74.7 ns: 1.01x slower |
+---------------------+---------+-----------------------+
| abc + UTF-8         | 167 ns  | 172 ns: 1.03x slower  |
+---------------------+---------+-----------------------+
| ASCII 1 MiB         | 118 us  | 118 us: 1.00x faster  |
+---------------------+---------+-----------------------+
| ASCII 1 MiB + UTF-8 | 1.08 ms | 1.07 ms: 1.00x faster |
+---------------------+---------+-----------------------+
| UTF-8 + ASCII 1 MiB | 572 us  | 570 us: 1.00x faster  |
+---------------------+---------+-----------------------+
| Geometric mean      | (ref)   | 1.00x slower          |
+---------------------+---------+-----------------------+

Benchmark hidden because not significant (2): UTF-8 + abc, UTF-8 euro + ASCII 1 MiB

=> There is no significant impact on bytes.decode() performance (no slow down).

[vstinner]

cc @serhiy-storchaka

[serhiy-storchaka]

LGTM.

[vstinner]

I enabled automerge. Thanks for the review @serhiy-storchaka.