pythonGH-139922: Tail calling for MSVC (VS 2026) (pythonGH-143068)

chris-eibl · Fidget-Spinner · brandtbucher · web-flow · commit be3c1316406b · 2025-12-22T23:01:34.000Z
Co-authored-by: Ken Jin &lt;28750310+Fidget-Spinner@users.noreply.github.com&gt;
Co-authored-by: Brandt Bucher &lt;brandt@python.org&gt;
Co-authored-by: Hugo van Kemenade &lt;1324225+hugovk@users.noreply.github.com&gt;
diff --git a/.github/workflows/tail-call.yml b/.github/workflows/tail-call.yml
@@ -79,19 +79,17 @@ jobs:
         with:
           python-version: '3.11'
 
-      - name: Native Windows (debug)
+      - name: Native Windows MSVC (release)
         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 --no-progress -y --force --params "--add Microsoft.VisualStudio.Component.VC.Tools.x86.x64  --locale en-US --passive"
+          $env:PATH = "C:\Program Files (x86)\Microsoft Visual Studio\18\BuildTools\MSBuild\Current\bin;$env:PATH"
+          ./PCbuild/build.bat --tail-call-interp -c Release -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: |
diff --git a/Include/internal/pycore_ceval.h b/Include/internal/pycore_ceval.h
@@ -415,6 +415,17 @@ _Py_VectorCall_StackRefSteal(
     int total_args,
     _PyStackRef kwnames);
 
+PyAPI_FUNC(PyObject*)
+_Py_VectorCallInstrumentation_StackRefSteal(
+    _PyStackRef callable,
+    _PyStackRef* arguments,
+    int total_args,
+    _PyStackRef kwnames,
+    bool call_instrumentation,
+    _PyInterpreterFrame* frame,
+    _Py_CODEUNIT* this_instr,
+    PyThreadState* tstate);
+
 PyAPI_FUNC(PyObject *)
 _Py_BuiltinCallFast_StackRefSteal(
     _PyStackRef callable,
@@ -464,6 +475,11 @@ _Py_assert_within_stack_bounds(
     _PyInterpreterFrame *frame, _PyStackRef *stack_pointer,
     const char *filename, int lineno);
 
+// Like PyMapping_GetOptionalItem, but returns the PyObject* instead of taking
+// it as an out parameter. This helps MSVC's escape analysis when used with
+// tail calling.
+PyAPI_FUNC(PyObject*) _PyMapping_GetOptionalItem2(PyObject* obj, PyObject* key, int* err);
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/Misc/NEWS.d/next/Core_and_Builtins/2025-10-11-17-01-21.gh-issue-139922.RUkXyd.rst b/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.
diff --git a/Objects/abstract.c b/Objects/abstract.c
@@ -224,6 +224,14 @@ PyMapping_GetOptionalItem(PyObject *obj, PyObject *key, PyObject **result)
     return 0;
 }
 
+PyObject*
+_PyMapping_GetOptionalItem2(PyObject *obj, PyObject *key, int *err)
+{
+    PyObject* result;
+    *err = PyMapping_GetOptionalItem(obj, key, &result);
+    return result;
+}
+
 int
 PyObject_SetItem(PyObject *o, PyObject *key, PyObject *value)
 {
diff --git a/PCbuild/pythoncore.vcxproj b/PCbuild/pythoncore.vcxproj
@@ -600,7 +600,9 @@
     <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" />
diff --git a/Python/bytecodes.c b/Python/bytecodes.c
@@ -1507,8 +1507,8 @@ dummy_func(
         }
 
         inst(LOAD_BUILD_CLASS, ( -- bc)) {
-            PyObject *bc_o;
-            int err = PyMapping_GetOptionalItem(BUILTINS(), &_Py_ID(__build_class__), &bc_o);
+            int err;
+            PyObject *bc_o = _PyMapping_GetOptionalItem2(BUILTINS(), &_Py_ID(__build_class__), &err);
             ERROR_IF(err < 0);
             if (bc_o == NULL) {
                 _PyErr_SetString(tstate, PyExc_NameError,
@@ -1711,8 +1711,9 @@ dummy_func(
 
         inst(LOAD_FROM_DICT_OR_GLOBALS, (mod_or_class_dict -- v)) {
             PyObject *name = GETITEM(FRAME_CO_NAMES, oparg);
-            PyObject *v_o;
-            int err = PyMapping_GetOptionalItem(PyStackRef_AsPyObjectBorrow(mod_or_class_dict), name, &v_o);
+            int err;
+            PyObject *v_o = _PyMapping_GetOptionalItem2(PyStackRef_AsPyObjectBorrow(mod_or_class_dict), name, &err);
+
             PyStackRef_CLOSE(mod_or_class_dict);
             ERROR_IF(err < 0);
             if (v_o == NULL) {
@@ -1735,11 +1736,11 @@ dummy_func(
                 else {
                     /* Slow-path if globals or builtins is not a dict */
                     /* namespace 1: globals */
-                    int err = PyMapping_GetOptionalItem(GLOBALS(), name, &v_o);
+                    v_o = _PyMapping_GetOptionalItem2(GLOBALS(), name, &err);
                     ERROR_IF(err < 0);
                     if (v_o == NULL) {
                         /* namespace 2: builtins */
-                        int err = PyMapping_GetOptionalItem(BUILTINS(), name, &v_o);
+                        v_o = _PyMapping_GetOptionalItem2(BUILTINS(), name, &err);
                         ERROR_IF(err < 0);
                         if (v_o == NULL) {
                             _PyEval_FormatExcCheckArg(
@@ -1898,14 +1899,14 @@ dummy_func(
         }
 
         inst(LOAD_FROM_DICT_OR_DEREF, (class_dict_st -- value)) {
-            PyObject *value_o;
             PyObject *name;
             PyObject *class_dict = PyStackRef_AsPyObjectBorrow(class_dict_st);
 
             assert(class_dict);
             assert(oparg >= 0 && oparg < _PyFrame_GetCode(frame)->co_nlocalsplus);
             name = PyTuple_GET_ITEM(_PyFrame_GetCode(frame)->co_localsplusnames, oparg);
-            int err = PyMapping_GetOptionalItem(class_dict, name, &value_o);
+            int err;
+            PyObject* value_o = _PyMapping_GetOptionalItem2(class_dict, name, &err);
             if (err < 0) {
                 ERROR_NO_POP();
             }
@@ -2074,14 +2075,14 @@ dummy_func(
         }
 
         inst(SETUP_ANNOTATIONS, (--)) {
-            PyObject *ann_dict;
             if (LOCALS() == NULL) {
                 _PyErr_Format(tstate, PyExc_SystemError,
                               "no locals found when setting up annotations");
                 ERROR_IF(true);
             }
             /* check if __annotations__ in locals()... */
-            int err = PyMapping_GetOptionalItem(LOCALS(), &_Py_ID(__annotations__), &ann_dict);
+            int err;
+            PyObject* ann_dict = _PyMapping_GetOptionalItem2(LOCALS(), &_Py_ID(__annotations__), &err);
             ERROR_IF(err < 0);
             if (ann_dict == NULL) {
                 ann_dict = PyDict_New();
@@ -2185,8 +2186,12 @@ 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;
+            {
+                // scope to tell MSVC that stack is not escaping
+                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) {
@@ -2245,8 +2250,13 @@ 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;
+            {
+                // scope to tell MSVC that method_found_ptr is not escaping
+                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();
             }
@@ -3472,10 +3482,14 @@ 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;
+            {
+                // scope to tell MSVC that stack is not escaping
+                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);
@@ -3707,36 +3721,18 @@ dummy_func(
                 frame->return_offset = INSTRUCTION_SIZE;
                 DISPATCH_INLINED(new_frame);
             }
-            /* Callable is not a normal Python function */
-            STACKREFS_TO_PYOBJECTS(arguments, total_args, args_o);
-            if (CONVERSION_FAILED(args_o)) {
-                DECREF_INPUTS();
-                ERROR_IF(true);
-            }
-            PyObject *res_o = PyObject_Vectorcall(
-                callable_o, args_o,
-                total_args | PY_VECTORCALL_ARGUMENTS_OFFSET,
-                NULL);
-            STACKREFS_TO_PYOBJECTS_CLEANUP(args_o);
-            if (opcode == INSTRUMENTED_CALL) {
-                PyObject *arg = total_args == 0 ?
-                    &_PyInstrumentation_MISSING : PyStackRef_AsPyObjectBorrow(arguments[0]);
-                if (res_o == NULL) {
-                    _Py_call_instrumentation_exc2(
-                        tstate, PY_MONITORING_EVENT_C_RAISE,
-                        frame, this_instr, callable_o, arg);
-                }
-                else {
-                    int err = _Py_call_instrumentation_2args(
-                        tstate, PY_MONITORING_EVENT_C_RETURN,
-                        frame, this_instr, callable_o, arg);
-                    if (err < 0) {
-                        Py_CLEAR(res_o);
-                    }
-                }
-            }
-            assert((res_o != NULL) ^ (_PyErr_Occurred(tstate) != NULL));
-            DECREF_INPUTS();
+            PyObject* res_o = _Py_VectorCallInstrumentation_StackRefSteal(
+                callable,
+                arguments,
+                total_args,
+                PyStackRef_NULL,
+                opcode == INSTRUMENTED_CALL,
+                frame,
+                this_instr,
+                tstate);
+            DEAD(args);
+            DEAD(self_or_null);
+            DEAD(callable);
             ERROR_IF(res_o == NULL);
             res = PyStackRef_FromPyObjectSteal(res_o);
         }
@@ -4587,35 +4583,19 @@ dummy_func(
                 frame->return_offset = INSTRUCTION_SIZE;
                 DISPATCH_INLINED(new_frame);
             }
-            /* Callable is not a normal Python function */
-            STACKREFS_TO_PYOBJECTS(arguments, total_args, args_o);
-            if (CONVERSION_FAILED(args_o)) {
-                DECREF_INPUTS();
-                ERROR_IF(true);
-            }
-            PyObject *res_o = PyObject_Vectorcall(
-                callable_o, args_o,
-                positional_args | PY_VECTORCALL_ARGUMENTS_OFFSET,
-                kwnames_o);
-            STACKREFS_TO_PYOBJECTS_CLEANUP(args_o);
-            if (opcode == INSTRUMENTED_CALL_KW) {
-                PyObject *arg = total_args == 0 ?
-                    &_PyInstrumentation_MISSING : PyStackRef_AsPyObjectBorrow(arguments[0]);
-                if (res_o == NULL) {
-                    _Py_call_instrumentation_exc2(
-                        tstate, PY_MONITORING_EVENT_C_RAISE,
-                        frame, this_instr, callable_o, arg);
-                }
-                else {
-                    int err = _Py_call_instrumentation_2args(
-                        tstate, PY_MONITORING_EVENT_C_RETURN,
-                        frame, this_instr, callable_o, arg);
-                    if (err < 0) {
-                        Py_CLEAR(res_o);
-                    }
-                }
-            }
-            DECREF_INPUTS();
+            PyObject* res_o = _Py_VectorCallInstrumentation_StackRefSteal(
+                callable,
+                arguments,
+                total_args,
+                kwnames,
+                opcode == INSTRUMENTED_CALL_KW,
+                frame,
+                this_instr,
+                tstate);
+            DEAD(kwnames);
+            DEAD(args);
+            DEAD(self_or_null);
+            DEAD(callable);
             ERROR_IF(res_o == NULL);
             res = PyStackRef_FromPyObjectSteal(res_o);
         }
diff --git a/Python/ceval.c b/Python/ceval.c
@@ -1071,6 +1071,65 @@ _Py_VectorCall_StackRefSteal(
     return res;
 }
 
+PyObject*
+_Py_VectorCallInstrumentation_StackRefSteal(
+    _PyStackRef callable,
+    _PyStackRef* arguments,
+    int total_args,
+    _PyStackRef kwnames,
+    bool call_instrumentation,
+    _PyInterpreterFrame* frame,
+    _Py_CODEUNIT* this_instr,
+    PyThreadState* tstate)
+{
+    PyObject* res;
+    STACKREFS_TO_PYOBJECTS(arguments, total_args, args_o);
+    if (CONVERSION_FAILED(args_o)) {
+        res = NULL;
+        goto cleanup;
+    }
+    PyObject* callable_o = PyStackRef_AsPyObjectBorrow(callable);
+    PyObject* kwnames_o = PyStackRef_AsPyObjectBorrow(kwnames);
+    int positional_args = total_args;
+    if (kwnames_o != NULL) {
+        positional_args -= (int)PyTuple_GET_SIZE(kwnames_o);
+    }
+    res = PyObject_Vectorcall(
+        callable_o, args_o,
+        positional_args | PY_VECTORCALL_ARGUMENTS_OFFSET,
+        kwnames_o);
+    STACKREFS_TO_PYOBJECTS_CLEANUP(args_o);
+    if (call_instrumentation) {
+        PyObject* arg = total_args == 0 ?
+            &_PyInstrumentation_MISSING : PyStackRef_AsPyObjectBorrow(arguments[0]);
+        if (res == NULL) {
+            _Py_call_instrumentation_exc2(
+                tstate, PY_MONITORING_EVENT_C_RAISE,
+                frame, this_instr, callable_o, arg);
+        }
+        else {
+            int err = _Py_call_instrumentation_2args(
+                tstate, PY_MONITORING_EVENT_C_RETURN,
+                frame, this_instr, callable_o, arg);
+            if (err < 0) {
+                Py_CLEAR(res);
+            }
+        }
+    }
+    assert((res != NULL) ^ (PyErr_Occurred() != NULL));
+cleanup:
+    PyStackRef_XCLOSE(kwnames);
+    // arguments is a pointer into the GC visible stack,
+    // so we must NULL out values as we clear them.
+    for (int i = total_args - 1; i >= 0; i--) {
+        _PyStackRef tmp = arguments[i];
+        arguments[i] = PyStackRef_NULL;
+        PyStackRef_CLOSE(tmp);
+    }
+    PyStackRef_CLOSE(callable);
+    return res;
+}
+
 PyObject *
 _Py_BuiltinCallFast_StackRefSteal(
     _PyStackRef callable,
diff --git a/Python/ceval_macros.h b/Python/ceval_macros.h
@@ -87,16 +87,19 @@
 #   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);
+#   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 DISPATCH_TABLE_VAR instruction_funcptr_table
 #   define DISPATCH_TABLE instruction_funcptr_handler_table
 #   define TRACING_DISPATCH_TABLE instruction_funcptr_tracing_table
-#   define TARGET(op) Py_PRESERVE_NONE_CC PyObject *_TAIL_CALL_##op(TAIL_CALL_PARAMS)
+#   define TARGET(op) Py_NO_INLINE PyObject *Py_PRESERVE_NONE_CC _TAIL_CALL_##op(TAIL_CALL_PARAMS)
 
 #   define DISPATCH_GOTO() \
         do { \
diff --git a/Python/executor_cases.c.h b/Python/executor_cases.c.h
diff --git a/Python/generated_cases.c.h b/Python/generated_cases.c.h
diff --git a/Python/opcode_targets.h b/Python/opcode_targets.h
diff --git a/Tools/cases_generator/target_generator.py b/Tools/cases_generator/target_generator.py

Original file line number	Diff line number	Diff line change
`@@ -0,0 +1 @@`
	`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.`
Original file line number	Diff line number	Diff line change
`@@ -224,6 +224,14 @@ PyMapping_GetOptionalItem(PyObject obj, PyObject key, PyObject **result)`
`224`	`224`	`return 0;`
`225`	`225`	`}`
`226`	`226`
	`227`	`+PyObject*`
	`228`	`+_PyMapping_GetOptionalItem2(PyObject obj, PyObject key, int *err)`
	`229`	`+{`
	`230`	`+ PyObject* result;`
	`231`	`+ *err = PyMapping_GetOptionalItem(obj, key, &result);`
	`232`	`+ return result;`
	`233`	`+}`
	`234`	`+`
`227`	`235`	`int`
`228`	`236`	`PyObject_SetItem(PyObject o, PyObject key, PyObject *value)`
`229`	`237`	`{`