gh-139922: Tail calling for MSVC (VS 2026)

.github/workflows/tail-call.yml

@@ -79,19 +79,17 @@ jobs:
         with:
           python-version: '3.11'
 
-      - name: Native Windows (debug)
+      - name: Native Windows MSVC (PGO)
         if: runner.os == 'Windows' && matrix.architecture != 'ARM64'
         shell: cmd
         run: |
-          choco install llvm --allow-downgrade --no-progress --version ${{ matrix.llvm }}.1.0
-          set PlatformToolset=clangcl
-          set LLVMToolsVersion=${{ matrix.llvm }}.1.0
-          set LLVMInstallDir=C:\Program Files\LLVM
-          call ./PCbuild/build.bat --tail-call-interp -d -p ${{ matrix.architecture }}
-          call ./PCbuild/rt.bat -d -p ${{ matrix.architecture }} -q --multiprocess 0 --timeout 4500 --verbose2 --verbose3
+          choco install visualstudio2026buildtools-preview --pre -allWorkloads
+          set PATH=C:\Program Files\Microsoft Visual Studio\18\Insiders\MSBuild\Current\Bin\;%PATH%
+          ./PCbuild/build.bat --tail-call-interp --pgo -p ${{ matrix.architecture }} "/p:PlatformToolset=v145"
+          ./PCbuild/rt.bat -p ${{ matrix.architecture }} -q --multiprocess 0 --timeout 4500 --verbose2 --verbose3
 
       # No tests (yet):
-      - name: Emulated Windows (release)
+      - name: Emulated Windows Clang (release)
         if: runner.os == 'Windows' && matrix.architecture == 'ARM64'
         shell: cmd
         run: |

Include/internal/pycore_ceval.h

@@ -304,9 +304,9 @@ PyAPI_FUNC(void) _PyEval_MonitorRaise(PyThreadState *tstate, _PyInterpreterFrame
 PyAPI_FUNC(bool) _PyEval_NoToolsForUnwind(PyThreadState *tstate);
 PyAPI_FUNC(int) _PyEval_UnpackIterableStackRef(PyThreadState *tstate, PyObject *v, int argcnt, int argcntafter, _PyStackRef *sp);
 PyAPI_FUNC(void) _PyEval_FrameClearAndPop(PyThreadState *tstate, _PyInterpreterFrame *frame);
-PyAPI_FUNC(PyObject **) _PyObjectArray_FromStackRefArray(_PyStackRef *input, Py_ssize_t nargs, PyObject **scratch);
+PyAPI_FUNC(PyObject **) _PyObjectArray_FromStackRefArray(_PyThreadStateImpl *_tstate, _PyStackRef *input, Py_ssize_t nargs);
 
-PyAPI_FUNC(void) _PyObjectArray_Free(PyObject **array, PyObject **scratch);
+PyAPI_FUNC(void) _PyObjectArray_Free(_PyThreadStateImpl *_tstate, PyObject **array, Py_ssize_t nargs, PyObject **temp_arr);
 
 PyAPI_FUNC(PyObject *) _PyEval_GetANext(PyObject *aiter);
 PyAPI_FUNC(void) _PyEval_LoadGlobalStackRef(PyObject *globals, PyObject *builtins, PyObject *name, _PyStackRef *writeto);

Include/internal/pycore_tstate.h

@@ -21,6 +21,9 @@ struct _gc_thread_state {
 };
 #endif
 
+/* How much scratch space to give stackref to PyObject* conversion. */
+#define MAX_STACKREF_SCRATCH 1024
+
 // Every PyThreadState is actually allocated as a _PyThreadStateImpl. The
 // PyThreadState fields are exposed as part of the C API, although most fields
 // are intended to be private. The _PyThreadStateImpl fields not exposed.
@@ -47,6 +50,8 @@ typedef struct _PyThreadStateImpl {
     struct _qsbr_thread_state *qsbr;  // only used by free-threaded build
     struct llist_node mem_free_queue; // delayed free queue
 
+    PyObject *stackref_scratch[MAX_STACKREF_SCRATCH];
+    int n_stackref_scratch_used;
 #ifdef Py_GIL_DISABLED
     // Stack references for the current thread that exist on the C stack
     struct _PyCStackRef *c_stack_refs;

Misc/NEWS.d/next/Core_and_Builtins/2025-10-11-17-01-21.gh-issue-139922.RUkXyd.rst

@@ -0,0 +1 @@
+Allow building CPython with the tail calling interpreter on Visual Studio 2026 MSVC. This provides a performance gain over the prior interpreter for MSVC. Patch by Ken Jin, Brandt Bucher, and Chris Eibl. With help from the MSVC team including Hulon Jenkins.

Objects/abstract.c

@@ -205,8 +205,14 @@ PyObject_GetItem(PyObject *o, PyObject *key)
     return type_error("'%.200s' object is not subscriptable", o);
 }
 
+// MSVC fails during a tail call release build with loads of
+// error C4737: Unable to perform required tail call.
+// without using Py_NO_INLINE here, but PGO works fine.
+#if defined(_MSC_VER) && !defined(__clang__) && _Py_TAIL_CALL_INTERP && !defined(_Py_USING_PGO)
+Py_NO_INLINE
+#endif
 int
-PyMapping_GetOptionalItem(PyObject *obj, PyObject *key, PyObject **result)
+PyMapping_GetOptionalItem(PyObject *obj, PyObject *key, PyObject **restrict result)
 {
     if (PyDict_CheckExact(obj)) {
         return PyDict_GetItemRef(obj, key, result);

PCbuild/pythoncore.vcxproj

@@ -598,7 +598,9 @@
     <ClCompile Include="..\Python\bltinmodule.c" />
     <ClCompile Include="..\Python\bootstrap_hash.c" />
     <ClCompile Include="..\Python\brc.c" />
-    <ClCompile Include="..\Python\ceval.c" />
+    <ClCompile Include="..\Python\ceval.c">
+      <AdditionalOptions Condition="'$(UseTailCallInterp)' == 'true' and $(PlatformToolset) != 'ClangCL'">/std:clatest %(AdditionalOptions)</AdditionalOptions>
+    </ClCompile>
     <ClCompile Include="..\Python\codecs.c" />
     <ClCompile Include="..\Python\codegen.c" />
     <ClCompile Include="..\Python\compile.c" />

Python/bytecodes.c

@@ -1982,7 +1982,7 @@ dummy_func(
                 ERROR_IF(true);
             }
             PyObject *str_o = _PyUnicode_JoinArray(&_Py_STR(empty), pieces_o, oparg);
-            STACKREFS_TO_PYOBJECTS_CLEANUP(pieces_o);
+            STACKREFS_TO_PYOBJECTS_CLEANUP(pieces_o, oparg);
             DECREF_INPUTS();
             ERROR_IF(str_o == NULL);
             str = PyStackRef_FromPyObjectSteal(str_o);
@@ -2107,7 +2107,7 @@ dummy_func(
                     values_o, 2,
                     values_o+1, 2,
                     oparg);
-            STACKREFS_TO_PYOBJECTS_CLEANUP(values_o);
+            STACKREFS_TO_PYOBJECTS_CLEANUP(values_o, oparg*2);
             DECREF_INPUTS();
             ERROR_IF(map_o == NULL);
             map = PyStackRef_FromPyObjectStealMortal(map_o);
@@ -2225,8 +2225,11 @@ dummy_func(
             }
             // we make no attempt to optimize here; specializations should
             // handle any case whose performance we care about
-            PyObject *stack[] = {class, self};
-            PyObject *super = PyObject_Vectorcall(global_super, stack, oparg & 2, NULL);
+            PyObject *super;
+            {
+                PyObject *stack[] = {class, self};
+                super = PyObject_Vectorcall(global_super, stack, oparg & 2, NULL);
+            }
             if (opcode == INSTRUMENTED_LOAD_SUPER_ATTR) {
                 PyObject *arg = oparg & 2 ? class : &_PyInstrumentation_MISSING;
                 if (super == NULL) {
@@ -2285,8 +2288,12 @@ dummy_func(
             PyObject *name = GETITEM(FRAME_CO_NAMES, oparg >> 2);
             PyTypeObject *cls = (PyTypeObject *)class;
             int method_found = 0;
-            PyObject *attr_o = _PySuper_Lookup(cls, self, name,
-                                   Py_TYPE(self)->tp_getattro == PyObject_GenericGetAttr ? &method_found : NULL);
+            PyObject *attr_o;
+            {
+                int *method_found_ptr = &method_found;
+                attr_o = _PySuper_Lookup(cls, self, name,
+                    Py_TYPE(self)->tp_getattro == PyObject_GenericGetAttr ? method_found_ptr : NULL);
+            }
             if (attr_o == NULL) {
                 ERROR_NO_POP();
             }
@@ -3509,10 +3516,13 @@ dummy_func(
             }
             assert(PyStackRef_IsTaggedInt(lasti));
             (void)lasti; // Shut up compiler warning if asserts are off
-            PyObject *stack[5] = {NULL, PyStackRef_AsPyObjectBorrow(exit_self), exc, val_o, tb};
-            int has_self = !PyStackRef_IsNull(exit_self);
-            PyObject *res_o = PyObject_Vectorcall(exit_func_o, stack + 2 - has_self,
-                    (3 + has_self) | PY_VECTORCALL_ARGUMENTS_OFFSET, NULL);
+            PyObject* res_o;
+            {
+                PyObject *stack[5] = {NULL, PyStackRef_AsPyObjectBorrow(exit_self), exc, val_o, tb};
+                int has_self = !PyStackRef_IsNull(exit_self);
+                res_o = PyObject_Vectorcall(exit_func_o, stack + 2 - has_self,
+                        (3 + has_self) | PY_VECTORCALL_ARGUMENTS_OFFSET, NULL);
+            }
             Py_XDECREF(original_tb);
             ERROR_IF(res_o == NULL);
             res = PyStackRef_FromPyObjectSteal(res_o);
@@ -3754,7 +3764,7 @@ dummy_func(
                 callable_o, args_o,
                 total_args | PY_VECTORCALL_ARGUMENTS_OFFSET,
                 NULL);
-            STACKREFS_TO_PYOBJECTS_CLEANUP(args_o);
+            STACKREFS_TO_PYOBJECTS_CLEANUP(args_o, total_args);
             if (opcode == INSTRUMENTED_CALL) {
                 PyObject *arg = total_args == 0 ?
                     &_PyInstrumentation_MISSING : PyStackRef_AsPyObjectBorrow(arguments[0]);
@@ -3909,7 +3919,7 @@ dummy_func(
                 callable_o, args_o,
                 total_args | PY_VECTORCALL_ARGUMENTS_OFFSET,
                 NULL);
-            STACKREFS_TO_PYOBJECTS_CLEANUP(args_o);
+            STACKREFS_TO_PYOBJECTS_CLEANUP(args_o, total_args);
             assert((res_o != NULL) ^ (_PyErr_Occurred(tstate) != NULL));
             DECREF_INPUTS();
             ERROR_IF(res_o == NULL);
@@ -4192,7 +4202,7 @@ dummy_func(
                 ERROR_IF(true);
             }
             PyObject *res_o = tp->tp_vectorcall((PyObject *)tp, args_o, total_args, NULL);
-            STACKREFS_TO_PYOBJECTS_CLEANUP(args_o);
+            STACKREFS_TO_PYOBJECTS_CLEANUP(args_o, total_args);
             DECREF_INPUTS();
             ERROR_IF(res_o == NULL);
             res = PyStackRef_FromPyObjectSteal(res_o);
@@ -4263,7 +4273,7 @@ dummy_func(
                 PyCFunction_GET_SELF(callable_o),
                 args_o,
                 total_args);
-            STACKREFS_TO_PYOBJECTS_CLEANUP(args_o);
+            STACKREFS_TO_PYOBJECTS_CLEANUP(args_o, total_args);
             assert((res_o != NULL) ^ (_PyErr_Occurred(tstate) != NULL));
             DECREF_INPUTS();
             ERROR_IF(res_o == NULL);
@@ -4299,7 +4309,7 @@ dummy_func(
                 ERROR_IF(true);
             }
             PyObject *res_o = cfunc(PyCFunction_GET_SELF(callable_o), args_o, total_args, NULL);
-            STACKREFS_TO_PYOBJECTS_CLEANUP(args_o);
+            STACKREFS_TO_PYOBJECTS_CLEANUP(args_o, total_args);
             assert((res_o != NULL) ^ (_PyErr_Occurred(tstate) != NULL));
             DECREF_INPUTS();
             ERROR_IF(res_o == NULL);
@@ -4479,7 +4489,7 @@ dummy_func(
             PyCFunctionFastWithKeywords cfunc =
                 _PyCFunctionFastWithKeywords_CAST(meth->ml_meth);
             PyObject *res_o = cfunc(self, (args_o + 1), nargs, NULL);
-            STACKREFS_TO_PYOBJECTS_CLEANUP(args_o);
+            STACKREFS_TO_PYOBJECTS_CLEANUP(args_o, total_args);
             assert((res_o != NULL) ^ (_PyErr_Occurred(tstate) != NULL));
             DECREF_INPUTS();
             ERROR_IF(res_o == NULL);
@@ -4558,7 +4568,7 @@ dummy_func(
             }
             PyCFunctionFast cfunc = _PyCFunctionFast_CAST(meth->ml_meth);
             PyObject *res_o = cfunc(self, (args_o + 1), nargs);
-            STACKREFS_TO_PYOBJECTS_CLEANUP(args_o);
+            STACKREFS_TO_PYOBJECTS_CLEANUP(args_o, total_args);
             assert((res_o != NULL) ^ (_PyErr_Occurred(tstate) != NULL));
             DECREF_INPUTS();
             ERROR_IF(res_o == NULL);
@@ -4657,7 +4667,7 @@ dummy_func(
                 callable_o, args_o,
                 positional_args | PY_VECTORCALL_ARGUMENTS_OFFSET,
                 kwnames_o);
-            STACKREFS_TO_PYOBJECTS_CLEANUP(args_o);
+            STACKREFS_TO_PYOBJECTS_CLEANUP(args_o, total_args);
             if (opcode == INSTRUMENTED_CALL_KW) {
                 PyObject *arg = total_args == 0 ?
                     &_PyInstrumentation_MISSING : PyStackRef_AsPyObjectBorrow(arguments[0]);
@@ -4814,7 +4824,7 @@ dummy_func(
                 positional_args | PY_VECTORCALL_ARGUMENTS_OFFSET,
                 kwnames_o);
             PyStackRef_CLOSE(kwnames);
-            STACKREFS_TO_PYOBJECTS_CLEANUP(args_o);
+            STACKREFS_TO_PYOBJECTS_CLEANUP(args_o, total_args);
             assert((res_o != NULL) ^ (_PyErr_Occurred(tstate) != NULL));
             DECREF_INPUTS();
             ERROR_IF(res_o == NULL);

Python/ceval.c

@@ -960,10 +960,12 @@ extern void _PyUOpPrint(const _PyUOpInstruction *uop);
 
 
 PyObject **
-_PyObjectArray_FromStackRefArray(_PyStackRef *input, Py_ssize_t nargs, PyObject **scratch)
+_PyObjectArray_FromStackRefArray(_PyThreadStateImpl *_tstate, _PyStackRef *restrict input, Py_ssize_t nargs)
 {
     PyObject **result;
-    if (nargs > MAX_STACKREF_SCRATCH) {
+    /* +1 because vectorcall might use -1 to write self */
+    PyObject **target = _tstate->stackref_scratch + _tstate->n_stackref_scratch_used + nargs + 1;
+    if (target > _tstate->stackref_scratch + MAX_STACKREF_SCRATCH) {
         // +1 in case PY_VECTORCALL_ARGUMENTS_OFFSET is set.
         result = PyMem_Malloc((nargs + 1) * sizeof(PyObject *));
         if (result == NULL) {
@@ -972,7 +974,9 @@ _PyObjectArray_FromStackRefArray(_PyStackRef *input, Py_ssize_t nargs, PyObject
         result++;
     }
     else {
-        result = scratch;
+        result = _tstate->stackref_scratch + _tstate->n_stackref_scratch_used;
+        _tstate->n_stackref_scratch_used += (int)nargs + 1;
+        assert(_tstate->n_stackref_scratch_used < MAX_STACKREF_SCRATCH);
     }
     for (int i = 0; i < nargs; i++) {
         result[i] = PyStackRef_AsPyObjectBorrow(input[i]);
@@ -981,10 +985,15 @@ _PyObjectArray_FromStackRefArray(_PyStackRef *input, Py_ssize_t nargs, PyObject
 }
 
 void
-_PyObjectArray_Free(PyObject **array, PyObject **scratch)
+_PyObjectArray_Free(_PyThreadStateImpl *_tstate, PyObject **array, Py_ssize_t nargs, PyObject **temp_arr)
 {
-    if (array != scratch) {
-        PyMem_Free(array);
+    /* -1 because we +1 previously */
+    if (array == temp_arr) {
+        _tstate->n_stackref_scratch_used -= ((int)nargs + 1);
+        assert(_tstate->n_stackref_scratch_used >= 0);
+    }
+    else {
+        PyMem_Free(array-1);
     }
 }
 

Python/ceval_macros.h

@@ -87,13 +87,15 @@
 #   elif defined(_MSC_VER) && (_MSC_VER < 1950)
 #       error "You need at least VS 2026 / PlatformToolset v145 for tail calling."
 #   endif
-
-    // Note: [[clang::musttail]] works for GCC 15, but not __attribute__((musttail)) at the moment.
-#   define Py_MUSTTAIL [[clang::musttail]]
-#   define Py_PRESERVE_NONE_CC __attribute__((preserve_none))
-    Py_PRESERVE_NONE_CC typedef PyObject* (*py_tail_call_funcptr)(TAIL_CALL_PARAMS);
-
-#   define TARGET(op) Py_PRESERVE_NONE_CC PyObject *_TAIL_CALL_##op(TAIL_CALL_PARAMS)
+#   if defined(_MSC_VER) && !defined(__clang__)
+#      define Py_MUSTTAIL [[msvc::musttail]]
+#      define Py_PRESERVE_NONE_CC __preserve_none
+#   else
+#       define Py_MUSTTAIL __attribute__((musttail))
+#       define Py_PRESERVE_NONE_CC __attribute__((preserve_none))
+#   endif
+    typedef PyObject *(Py_PRESERVE_NONE_CC *py_tail_call_funcptr)(TAIL_CALL_PARAMS);
+#   define TARGET(op) Py_NO_INLINE PyObject *Py_PRESERVE_NONE_CC _TAIL_CALL_##op(TAIL_CALL_PARAMS)
 #   define DISPATCH_GOTO() \
         do { \
             Py_MUSTTAIL return (((py_tail_call_funcptr *)instruction_funcptr_table)[opcode])(TAIL_CALL_ARGS); \
@@ -401,17 +403,14 @@ do {                                                   \
 
 /* Stackref macros */
 
-/* How much scratch space to give stackref to PyObject* conversion. */
-#define MAX_STACKREF_SCRATCH 10
 
 #define STACKREFS_TO_PYOBJECTS(ARGS, ARG_COUNT, NAME) \
-    /* +1 because vectorcall might use -1 to write self */ \
-    PyObject *NAME##_temp[MAX_STACKREF_SCRATCH+1]; \
-    PyObject **NAME = _PyObjectArray_FromStackRefArray(ARGS, ARG_COUNT, NAME##_temp + 1);
+    _PyThreadStateImpl *_tstate = (_PyThreadStateImpl *)tstate; \
+    PyObject **NAME##_temp = _tstate->stackref_scratch + _tstate->n_stackref_scratch_used;                                                  \
+    PyObject **NAME = _PyObjectArray_FromStackRefArray(_tstate, ARGS, ARG_COUNT);
 
-#define STACKREFS_TO_PYOBJECTS_CLEANUP(NAME) \
-    /* +1 because we +1 previously */ \
-    _PyObjectArray_Free(NAME - 1, NAME##_temp);
+#define STACKREFS_TO_PYOBJECTS_CLEANUP(NAME, ARG_COUNT) \
+    _PyObjectArray_Free(_tstate, NAME, ARG_COUNT, NAME##_temp);
 
 #define CONVERSION_FAILED(NAME) ((NAME) == NULL)