cpython/InternalDocs/string_interning.md
Petr Viktorin 6f1d448bc1
gh-113993: Allow interned strings to be mortal, and fix related issues (GH-120520)
* Add an InternalDocs file describing how interning should work and how to use it.

* Add internal functions to *explicitly* request what kind of interning is done:
  - `_PyUnicode_InternMortal`
  - `_PyUnicode_InternImmortal`
  - `_PyUnicode_InternStatic`

* Switch uses of `PyUnicode_InternInPlace` to those.

* Disallow using `_Py_SetImmortal` on strings directly.
  You should use `_PyUnicode_InternImmortal` instead:
  - Strings should be interned before immortalization, otherwise you're possibly
    interning a immortalizing copy.
  - `_Py_SetImmortal` doesn't handle the `SSTATE_INTERNED_MORTAL` to
    `SSTATE_INTERNED_IMMORTAL` update, and those flags can't be changed in
    backports, as they are now part of public API and version-specific ABI.

* Add private `_only_immortal` argument for `sys.getunicodeinternedsize`, used in refleak test machinery.

* Make sure the statically allocated string singletons are unique. This means these sets are now disjoint:
  - `_Py_ID`
  - `_Py_STR` (including the empty string)
  - one-character latin-1 singletons

  Now, when you intern a singleton, that exact singleton will be interned.

* Add a `_Py_LATIN1_CHR` macro, use it instead of `_Py_ID`/`_Py_STR` for one-character latin-1 singletons everywhere (including Clinic).

* Intern `_Py_STR` singletons at startup.

* For free-threaded builds, intern `_Py_LATIN1_CHR` singletons at startup.

* Beef up the tests. Cover internal details (marked with `@cpython_only`).

* Add lots of assertions

Co-Authored-By: Eric Snow <ericsnowcurrently@gmail.com>
2024-06-21 17:19:31 +02:00

4.7 KiB

String interning

Interned strings are conceptually part of an interpreter-global set of interned strings, meaning that:

  • no two interned strings have the same content (across an interpreter);
  • two interned strings can be safely compared using pointer equality (Python is).

This is used to optimize dict and attribute lookups, among other things.

Python uses three different mechanisms to intern strings:

  • Singleton strings marked in C source with _Py_STR and _Py_ID macros. These are statically allocated, and collected using make regen-global-objects (Tools/build/generate_global_objects.py), which generates code for declaration, initialization and finalization.

    The difference between the two kinds is not important. (A _Py_ID string is a valid C name, with which we can refer to it; a _Py_STR may e.g. contain non-identifier characters, so it needs a separate C-compatible name.)

    The empty string is in this category (as _Py_STR(empty)).

    These singletons are interned in a runtime-global lookup table, _PyRuntime.cached_objects.interned_strings (INTERNED_STRINGS), at runtime initialization.

  • The 256 possible one-character latin-1 strings are singletons, which can be retrieved with _Py_LATIN1_CHR(c), are stored in runtime-global arrays, _PyRuntime.static_objects.strings.ascii and _PyRuntime.static_objects.strings.latin1.

    These are NOT interned at startup in the normal build. In the free-threaded build, they are; this avoids modifying the global lookup table after threads are started.

    Interning a one-char latin-1 string will always intern the corresponding singleton.

  • All other strings are allocated dynamically, and have their _PyUnicode_STATE(s).statically_allocated flag set to zero. When interned, such strings are added to an interpreter-wide dict, PyInterpreterState.cached_objects.interned_strings.

    The key and value of each entry in this dict reference the same object.

The three sets of singletons (_Py_STR, _Py_ID, _Py_LATIN1_CHR) are disjoint. If you have such a singleton, it (and no other copy) will be interned.

Immortality and reference counting

Invariant: Every immortal string is interned, except the one-char latin-1 singletons (which might but might not be interned).

In practice, this means that you must not use _Py_SetImmortal on a string. (If you know it's already immortal, don't immortalize it; if you know it's not interned you might be immortalizing a redundant copy; if it's interned and mortal it needs extra processing in _PyUnicode_InternImmortal.)

The converse is not true: interned strings can be mortal. For mortal interned strings:

  • the 2 references from the interned dict (key & value) are excluded from their refcount
  • the deallocator (unicode_dealloc) removes the string from the interned dict
  • at shutdown, when the interned dict is cleared, the references are added back

As with any type, you should only immortalize strings that will live until interpreter shutdown. We currently also immortalize strings contained in code objects and similar, specifically in the compiler and in marshal. These are “close enough” to immortal: even in use cases like hot reloading or eval-ing user input, the number of distinct identifiers and string constants expected to stay low.

Internal API

We have the following internal API for interning:

  • _PyUnicode_InternMortal: just intern the string
  • _PyUnicode_InternImmortal: intern, and immortalize the result
  • _PyUnicode_InternStatic: intern a static singleton (_Py_STR, _Py_ID or one-byte). Not for general use.

All take an interpreter state, and a pointer to a PyObject* which they modify in place.

The functions take ownership of (“steal”) the reference to their argument, and update the argument with a new reference. This means:

  • They're “reference neutral”.
  • They must not be called with a borrowed reference.

State

The intern state (retrieved by PyUnicode_CHECK_INTERNED(s); stored in _PyUnicode_STATE(s).interned) can be:

  • SSTATE_NOT_INTERNED (defined as 0, which is useful in a boolean context)
  • SSTATE_INTERNED_MORTAL (1)
  • SSTATE_INTERNED_IMMORTAL (2)
  • SSTATE_INTERNED_IMMORTAL_STATIC (3)

The valid transitions between these states are:

  • For dynamically allocated strings:

    • 0 -> 1 (_PyUnicode_InternMortal)
    • 1 -> 2 or 0 -> 2 (_PyUnicode_InternImmortal)

    Using _PyUnicode_InternStatic on these is an error; the other cases don't change the state.

  • One-char latin-1 singletons can be interned (0 -> 3) using any interning function; after that the functions don't change the state.

  • Other statically allocated strings are interned (0 -> 3) at runtime init; after that all interning functions don't change the state.