gh-81057: Move the Extension Modules Cache to _PyRuntimeState (gh-99355)

We also move the closely related max_module_number and add comments documenting the group of struct members.

https://github.com/python/cpython/issues/81057

Include/internal/pycore_import.h

@@ -5,6 +5,23 @@
 extern "C" {
 #endif
 
+
+struct _import_runtime_state {
+    /* The most recent value assigned to a PyModuleDef.m_base.m_index.
+       This is incremented each time PyModuleDef_Init() is called,
+       which is just about every time an extension module is imported.
+       See PyInterpreterState.modules_by_index for more info. */
+    Py_ssize_t last_module_index;
+    /* A dict mapping (filename, name) to PyModuleDef for modules.
+       Only legacy (single-phase init) extension modules are added
+       and only if they support multiple initialization (m_size >- 0)
+       or are imported in the main interpreter.
+       This is initialized lazily in _PyImport_FixupExtensionObject().
+       Modules are added there and looked up in _imp.find_extension(). */
+    PyObject *extensions;
+};
+
+
 #ifdef HAVE_FORK
 extern PyStatus _PyImport_ReInitLock(void);
 #endif

Include/internal/pycore_interp.h

@@ -123,6 +123,25 @@ struct _is {
 
     // sys.modules dictionary
     PyObject *modules;
+    /* This is the list of module objects for all legacy (single-phase init)
+       extension modules ever loaded in this process (i.e. imported
+       in this interpreter or in any other).  Py_None stands in for
+       modules that haven't actually been imported in this interpreter.
+
+       A module's index (PyModuleDef.m_base.m_index) is used to look up
+       the corresponding module object for this interpreter, if any.
+       (See PyState_FindModule().)  When any extension module
+       is initialized during import, its moduledef gets initialized by
+       PyModuleDef_Init(), and the first time that happens for each
+       PyModuleDef, its index gets set to the current value of
+       a global counter (see _PyRuntimeState.imports.last_module_index).
+       The entry for that index in this interpreter remains unset until
+       the module is actually imported here.  (Py_None is used as
+       a placeholder.)  Note that multi-phase init modules always get
+       an index for which there will never be a module set.
+
+       This is initialized lazily in _PyState_AddModule(), which is also
+       where modules get added. */
     PyObject *modules_by_index;
     // Dictionary of the sys module
     PyObject *sysdict;

Include/internal/pycore_runtime.h

@@ -11,6 +11,7 @@ extern "C" {
 #include "pycore_atomic.h"          /* _Py_atomic_address */
 #include "pycore_gil.h"             // struct _gil_runtime_state
 #include "pycore_global_objects.h"  // struct _Py_global_objects
+#include "pycore_import.h"          // struct _import_runtime_state
 #include "pycore_interp.h"          // PyInterpreterState
 #include "pycore_unicodeobject.h"   // struct _Py_unicode_runtime_ids
 
@@ -115,6 +116,7 @@ typedef struct pyruntimestate {
     void (*exitfuncs[NEXITFUNCS])(void);
     int nexitfuncs;
 
+    struct _import_runtime_state imports;
     struct _ceval_runtime_state ceval;
     struct _gilstate_runtime_state gilstate;
     struct _getargs_runtime_state getargs;