bpo-41324 Add a minimal decimal capsule API (#21519)

Author: Stefan Krah

Doc/c-api/concrete.rst

@@ -115,3 +115,4 @@ Other Objects
    coro.rst
    contextvars.rst
    datetime.rst
+   decimal.rst

Doc/c-api/decimal.rst

@@ -0,0 +1,231 @@
+.. sectionauthor:: Stefan Krah
+
+.. highlight:: c
+
+
+Decimal capsule API
+===================
+
+Capsule API functions can be used in the same manner as regular library
+functions, provided that the API has been initialized.
+
+
+Initialize
+----------
+
+Typically, a C extension module that uses the decimal API will do these
+steps in its init function:
+
+.. code-block::
+
+    #include "pydecimal.h"
+
+    static int decimal_initialized = 0;
+    if (!decimal_initialized) {
+        if (import_decimal() < 0) {
+            return NULL;
+        }
+
+        decimal_initialized = 1;
+    }
+
+
+Type checking, predicates, accessors
+------------------------------------
+
+.. c:function:: int PyDec_TypeCheck(const PyObject *dec)
+
+   Return 1 if ``dec`` is a Decimal, 0 otherwise.  This function does not set
+   any exceptions.
+
+
+.. c:function:: int PyDec_IsSpecial(const PyObject *dec)
+
+   Return 1 if ``dec`` is ``NaN``, ``sNaN`` or ``Infinity``, 0 otherwise.
+
+   Set TypeError and return -1 if ``dec`` is not a Decimal.  It is guaranteed that
+   this is the only failure mode, so if ``dec`` has already been type-checked, no
+   errors can occur and the function can be treated as a simple predicate.
+
+
+.. c:function:: int PyDec_IsNaN(const PyObject *dec)
+
+   Return 1 if ``dec`` is ``NaN`` or ``sNaN``, 0 otherwise.
+
+   Set TypeError and return -1 if ``dec`` is not a Decimal.  It is guaranteed that
+   this is the only failure mode, so if ``dec`` has already been type-checked, no
+   errors can occur and the function can be treated as a simple predicate.
+
+
+.. c:function:: int PyDec_IsInfinite(const PyObject *dec)
+
+   Return 1 if ``dec`` is ``Infinity``, 0 otherwise.
+
+   Set TypeError and return -1 if ``dec`` is not a Decimal.  It is guaranteed that
+   this is the only failure mode, so if ``dec`` has already been type-checked, no
+   errors can occur and the function can be treated as a simple predicate.
+
+
+.. c:function:: int64_t PyDec_GetDigits(const PyObject *dec)
+
+   Return the number of digits in the coefficient.  For ``Infinity``, the
+   number of digits is always zero.  Typically, the same applies to ``NaN``
+   and ``sNaN``, but both of these can have a payload that is equivalent to
+   a coefficient.  Therefore, ``NaNs`` can have a nonzero return value.
+
+   Set TypeError and return -1 if ``dec`` is not a Decimal.  It is guaranteed that
+   this is the only failure mode, so if ``dec`` has already been type-checked, no
+   errors can occur and the function can be treated as a simple accessor.
+
+
+Exact conversions between decimals and primitive C types
+--------------------------------------------------------
+
+This API supports conversions for decimals with a coefficient up to 38 digits.
+
+Data structures
+~~~~~~~~~~~~~~~
+
+The conversion functions use the following status codes and data structures:
+
+.. code-block::
+
+   /* status cases for getting a triple */
+   enum mpd_triple_class {
+     MPD_TRIPLE_NORMAL,
+     MPD_TRIPLE_INF,
+     MPD_TRIPLE_QNAN,
+     MPD_TRIPLE_SNAN,
+     MPD_TRIPLE_ERROR,
+   };
+
+   typedef struct {
+     enum mpd_triple_class tag;
+     uint8_t sign;
+     uint64_t hi;
+     uint64_t lo;
+     int64_t exp;
+   } mpd_uint128_triple_t;
+
+The status cases are explained below.  ``sign`` is 0 for positive and 1 for negative.
+``((uint128_t)hi << 64) + lo`` is the coefficient, ``exp`` is the exponent.
+
+The data structure is called "triple" because the decimal triple (sign, coeff, exp)
+is an established term and (``hi``, ``lo``) represents a single ``uint128_t`` coefficient.
+
+
+Functions
+~~~~~~~~~
+
+.. c:function:: mpd_uint128_triple_t PyDec_AsUint128Triple(const PyObject *dec)
+
+   Convert a decimal to a triple.  As above, it is guaranteed that the only
+   Python failure mode is a TypeError, checks can be omitted if the type is
+   known.
+
+   For simplicity, the usage of the function and all special cases are
+   explained in code form and comments:
+
+.. code-block::
+
+    triple = PyDec_AsUint128Triple(dec);
+    switch (triple.tag) {
+    case MPD_TRIPLE_QNAN:
+        /*
+         * Success: handle a quiet NaN.
+         *   1) triple.sign is 0 or 1.
+         *   2) triple.exp is always 0.
+         *   3) If triple.hi or triple.lo are nonzero, the NaN has a payload.
+         */
+        break;
+
+    case MPD_TRIPLE_SNAN:
+        /*
+         * Success: handle a signaling NaN.
+         *   1) triple.sign is 0 or 1.
+         *   2) triple.exp is always 0.
+         *   3) If triple.hi or triple.lo are nonzero, the sNaN has a payload.
+         */
+        break;
+
+    case MPD_TRIPLE_INF:
+        /*
+         * Success: handle Infinity.
+         *   1) triple.sign is 0 or 1.
+         *   2) triple.exp is always 0.
+         *   3) triple.hi and triple.lo are always zero.
+         */
+        break;
+
+    case MPD_TRIPLE_NORMAL:
+        /* Success: handle a finite value. */
+        break;
+
+    case MPD_TRIPLE_ERROR:
+        /* TypeError check: can be omitted if the type of dec is known. */
+        if (PyErr_Occurred()) {
+            return NULL;
+        }
+
+        /* Too large for conversion.  PyDec_AsUint128Triple() does not set an
+           exception so applications can choose themselves.  Typically this
+           would be a ValueError. */
+        PyErr_SetString(PyExc_ValueError,
+            "value out of bounds for a uint128 triple");
+        return NULL;
+    }
+
+.. c:function:: PyObject *PyDec_FromUint128Triple(const mpd_uint128_triple_t *triple)
+
+   Create a decimal from a triple.  The following rules must be observed for
+   initializing the triple:
+
+   1) ``triple.sign`` must always be 0 (for positive) or 1 (for negative).
+
+   2) ``MPD_TRIPLE_QNAN``: ``triple.exp`` must be 0.  If ``triple.hi`` or ``triple.lo``
+      are nonzero,  create a ``NaN`` with a payload.
+
+   3) ``MPD_TRIPLE_SNAN``: ``triple.exp`` must be 0. If ``triple.hi`` or ``triple.lo``
+      are nonzero,  create an ``sNaN`` with a payload.
+
+   4) ``MPD_TRIPLE_INF``: ``triple.exp``, ``triple.hi`` and ``triple.lo`` must be zero.
+
+   5) ``MPD_TRIPLE_NORMAL``: ``MPD_MIN_ETINY + 38 < triple.exp < MPD_MAX_EMAX - 38``.
+      ``triple.hi`` and ``triple.lo`` can be chosen freely.
+
+   6) ``MPD_TRIPLE_ERROR``: It is always an error to set this tag.
+
+
+   If one of the above conditions is not met, the function returns ``NaN`` if
+   the ``InvalidOperation`` trap is not set in the thread local context.  Otherwise,
+   it sets the ``InvalidOperation`` exception and returns NULL.
+
+   Additionally, though extremely unlikely give the small allocation sizes,
+   the function can set ``MemoryError`` and return ``NULL``.
+
+
+Advanced API
+------------
+
+This API enables the use of ``libmpdec`` functions.  Since Python is compiled with
+hidden symbols, the API requires an external libmpdec and the ``mpdecimal.h``
+header.
+
+
+Functions
+~~~~~~~~~
+
+.. c:function:: PyObject *PyDec_Alloc(void)
+
+   Return a new decimal that can be used in the ``result`` position of ``libmpdec``
+   functions.
+
+.. c:function:: mpd_t *PyDec_Get(PyObject *v)
+
+   Get a pointer to the internal ``mpd_t`` of the decimal.  Decimals are immutable,
+   so this function must only be used on a new Decimal that has been created by
+   PyDec_Alloc().
+
+.. c:function:: const mpd_t *PyDec_GetConst(const PyObject *v)
+
+   Get a pointer to the constant internal ``mpd_t`` of the decimal.