gh-140544: store pointer to interpreter state as a thread local for fast access (#140573)

Include/internal/pycore_freelist.h

@@ -17,15 +17,16 @@ extern "C" {
 static inline struct _Py_freelists *
 _Py_freelists_GET(void)
 {
-    PyThreadState *tstate = _PyThreadState_GET();
 #ifdef Py_DEBUG
-    _Py_EnsureTstateNotNULL(tstate);
+    _Py_AssertHoldsTstate();
 #endif
 
 #ifdef Py_GIL_DISABLED
+    PyThreadState *tstate = _PyThreadState_GET();
     return &((_PyThreadStateImpl*)tstate)->freelists;
 #else
-    return &tstate->interp->object_state.freelists;
+    PyInterpreterState *interp = _PyInterpreterState_GET();
+    return &interp->object_state.freelists;
 #endif
 }
 

Include/internal/pycore_pystate.h

@@ -91,6 +91,7 @@ _Py_ThreadCanHandleSignals(PyInterpreterState *interp)
 
 #if defined(HAVE_THREAD_LOCAL) && !defined(Py_BUILD_CORE_MODULE)
 extern _Py_thread_local PyThreadState *_Py_tss_tstate;
+extern _Py_thread_local PyInterpreterState *_Py_tss_interp;
 #endif
 
 #ifndef NDEBUG
@@ -204,11 +205,15 @@ _Py_EnsureFuncTstateNotNULL(const char *func, PyThreadState *tstate)
    See also PyInterpreterState_Get()
    and _PyGILState_GetInterpreterStateUnsafe(). */
 static inline PyInterpreterState* _PyInterpreterState_GET(void) {
-    PyThreadState *tstate = _PyThreadState_GET();
 #ifdef Py_DEBUG
+    PyThreadState *tstate = _PyThreadState_GET();
     _Py_EnsureTstateNotNULL(tstate);
 #endif
-    return tstate->interp;
+#if !defined(Py_BUILD_CORE_MODULE)
+    return _Py_tss_interp;
+#else
+    return _PyThreadState_GET()->interp;
+#endif
 }
 
 

Misc/NEWS.d/next/Core_and_Builtins/2025-10-25-07-25-52.gh-issue-140544.lwjtQe.rst

@@ -0,0 +1 @@
+Speed up accessing interpreter state by caching it in a thread local variable. Patch by Kumar Aditya.

Python/pystate.c

@@ -78,6 +78,10 @@ _Py_thread_local PyThreadState *_Py_tss_tstate = NULL;
    also known as a "gilstate." */
 _Py_thread_local PyThreadState *_Py_tss_gilstate = NULL;
 
+/* The interpreter of the attached thread state,
+   and is same as tstate->interp. */
+_Py_thread_local PyInterpreterState *_Py_tss_interp = NULL;
+
 static inline PyThreadState *
 current_fast_get(void)
 {
@@ -89,12 +93,15 @@ current_fast_set(_PyRuntimeState *Py_UNUSED(runtime), PyThreadState *tstate)
 {
     assert(tstate != NULL);
     _Py_tss_tstate = tstate;
+    assert(tstate->interp != NULL);
+    _Py_tss_interp = tstate->interp;
 }
 
 static inline void
 current_fast_clear(_PyRuntimeState *Py_UNUSED(runtime))
 {
     _Py_tss_tstate = NULL;
+    _Py_tss_interp = NULL;
 }
 
 #define tstate_verify_not_active(tstate) \
@@ -1281,9 +1288,8 @@ _PyInterpreterState_RequireIDRef(PyInterpreterState *interp, int required)
 PyInterpreterState*
 PyInterpreterState_Get(void)
 {
-    PyThreadState *tstate = current_fast_get();
-    _Py_EnsureTstateNotNULL(tstate);
-    PyInterpreterState *interp = tstate->interp;
+    _Py_AssertHoldsTstate();
+    PyInterpreterState *interp = _Py_tss_interp;
     if (interp == NULL) {
         Py_FatalError("no current interpreter");
     }

Tools/c-analyzer/cpython/ignored.tsv

@@ -194,6 +194,7 @@ Python/pyfpe.c	-	PyFPE_counter	-
 Python/import.c	-	pkgcontext	-
 Python/pystate.c	-	_Py_tss_tstate	-
 Python/pystate.c	-	_Py_tss_gilstate	-
+Python/pystate.c	-	_Py_tss_interp	-
 
 ##-----------------------
 ## should be const