gh-118335: Configure Tier 2 interpreter at build time (#118339)

The code for Tier 2 is now only compiled when configured
with `--enable-experimental-jit[=yes|interpreter]`.

We drop support for `PYTHON_UOPS` and -`Xuops`,
but you can disable the interpreter or JIT
at runtime by setting `PYTHON_JIT=0`.
You can also build it without enabling it by default
using `--enable-experimental-jit=yes-off`;
enable with `PYTHON_JIT=1`.

On Windows, the `build.bat` script supports
`--experimental-jit`, `--experimental-jit-off`,
`--experimental-interpreter`.

In the C code, `_Py_JIT` is defined as before
when the JIT is enabled; the new variable
`_Py_TIER2` is defined when the JIT *or* the
interpreter is enabled. It is actually a bitmask:
1: JIT; 2: default-off; 4: interpreter.

Python/bytecodes.c

@@ -2363,6 +2363,7 @@ dummy_func(
             CHECK_EVAL_BREAKER();
             assert(oparg <= INSTR_OFFSET());
             JUMPBY(-oparg);
+            #ifdef _Py_TIER2
             #if ENABLE_SPECIALIZATION
             _Py_BackoffCounter counter = this_instr[1].counter;
             if (backoff_counter_triggers(counter) && this_instr->op.code == JUMP_BACKWARD) {
@@ -2388,6 +2389,7 @@ dummy_func(
                 ADVANCE_ADAPTIVE_COUNTER(this_instr[1].counter);
             }
             #endif  /* ENABLE_SPECIALIZATION */
+            #endif /* _Py_TIER2 */
         }
 
         pseudo(JUMP) = {
@@ -2401,6 +2403,7 @@ dummy_func(
         };
 
         tier1 inst(ENTER_EXECUTOR, (--)) {
+            #ifdef _Py_TIER2
             int prevoparg = oparg;
             CHECK_EVAL_BREAKER();
             if (this_instr->op.code != ENTER_EXECUTOR ||
@@ -2418,6 +2421,9 @@ dummy_func(
             tstate->previous_executor = Py_None;
             Py_INCREF(executor);
             GOTO_TIER_TWO(executor);
+            #else
+            Py_FatalError("ENTER_EXECUTOR is not supported in this build");
+            #endif /* _Py_TIER2 */
         }
 
         replaced op(_POP_JUMP_IF_FALSE, (cond -- )) {

Python/ceval.c

@@ -755,7 +755,7 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
     _Py_CODEUNIT *next_instr;
     PyObject **stack_pointer;
 
-#ifndef _Py_JIT
+#if defined(_Py_TIER2) && !defined(_Py_JIT)
     /* Tier 2 interpreter state */
     _PyExecutorObject *current_executor = NULL;
     const _PyUOpInstruction *next_uop = NULL;
@@ -959,6 +959,7 @@ resume_with_error:
     goto error;
 
 
+#ifdef _Py_TIER2
 
 // Tier 2 is also here!
 enter_tier_two:
@@ -1113,6 +1114,8 @@ exit_to_trace:
 
 #endif  // _Py_JIT
 
+#endif // _Py_TIER2
+
 }
 
 #if defined(__GNUC__)

Python/generated_cases.c.h

@@ -2492,6 +2492,7 @@
             (void)this_instr;
             next_instr += 1;
             INSTRUCTION_STATS(ENTER_EXECUTOR);
+            #ifdef _Py_TIER2
             int prevoparg = oparg;
             CHECK_EVAL_BREAKER();
             if (this_instr->op.code != ENTER_EXECUTOR ||
@@ -2508,6 +2509,9 @@
             tstate->previous_executor = Py_None;
             Py_INCREF(executor);
             GOTO_TIER_TWO(executor);
+            #else
+            Py_FatalError("ENTER_EXECUTOR is not supported in this build");
+            #endif /* _Py_TIER2 */
             DISPATCH();
         }
 
@@ -3432,6 +3436,7 @@
             CHECK_EVAL_BREAKER();
             assert(oparg <= INSTR_OFFSET());
             JUMPBY(-oparg);
+            #ifdef _Py_TIER2
             #if ENABLE_SPECIALIZATION
             _Py_BackoffCounter counter = this_instr[1].counter;
             if (backoff_counter_triggers(counter) && this_instr->op.code == JUMP_BACKWARD) {
@@ -3457,6 +3462,7 @@
                 ADVANCE_ADAPTIVE_COUNTER(this_instr[1].counter);
             }
             #endif  /* ENABLE_SPECIALIZATION */
+            #endif /* _Py_TIER2 */
             DISPATCH();
         }
 

Python/instrumentation.c

@@ -1705,10 +1705,12 @@ instrument_lock_held(PyCodeObject *code, PyInterpreterState *interp)
         );
         return 0;
     }
+#ifdef _Py_TIER2
     if (code->co_executors != NULL) {
         _PyCode_Clear_Executors(code);
     }
     _Py_Executors_InvalidateDependency(interp, code, 1);
+#endif
     int code_len = (int)Py_SIZE(code);
     /* Exit early to avoid creating instrumentation
      * data for potential statically allocated code
@@ -1946,7 +1948,9 @@ _PyMonitoring_SetEvents(int tool_id, _PyMonitoringEventSet events)
         goto done;
     }
     set_global_version(tstate, new_version);
+#ifdef _Py_TIER2
     _Py_Executors_InvalidateAll(interp, 1);
+#endif
     res = instrument_all_executing_code_objects(interp);
 done:
     _PyEval_StartTheWorld(interp);
@@ -1986,7 +1990,9 @@ _PyMonitoring_SetLocalEvents(PyCodeObject *code, int tool_id, _PyMonitoringEvent
         code->_co_instrumentation_version -= MONITORING_VERSION_INCREMENT;
     }
 
+#ifdef _Py_TIER2
     _Py_Executors_InvalidateDependency(interp, code, 1);
+#endif
 
     res = instrument_lock_held(code, interp);
 

Python/optimizer.c

@@ -1,3 +1,5 @@
+#ifdef _Py_TIER2
+
 #include "Python.h"
 #include "opcode.h"
 #include "pycore_interp.h"
@@ -1622,3 +1624,5 @@ _Py_Executors_InvalidateAll(PyInterpreterState *interp, int is_invalidation)
         }
     }
 }
+
+#endif /* _Py_TIER2 */

Python/optimizer_analysis.c

@@ -1,3 +1,5 @@
+#ifdef _Py_TIER2
+
 /*
  * This file contains the support code for CPython's uops optimizer.
  * It also performs some simple optimizations.
@@ -603,3 +605,5 @@ _Py_uop_analyze_and_optimize(
     OPT_STAT_INC(optimizer_successes);
     return length;
 }
+
+#endif /* _Py_TIER2 */

Python/optimizer_symbols.c

@@ -1,3 +1,4 @@
+#ifdef _Py_TIER2
 
 #include "Python.h"
 
@@ -506,3 +507,5 @@ fail:
     Py_XDECREF(val_43);
     return NULL;
 }
+
+#endif /* _Py_TIER2 */

Python/pylifecycle.c

@@ -624,9 +624,11 @@ static int
 builtins_dict_watcher(PyDict_WatchEvent event, PyObject *dict, PyObject *key, PyObject *new_value)
 {
     PyInterpreterState *interp = _PyInterpreterState_GET();
+#ifdef _Py_TIER2
     if (interp->rare_events.builtin_dict < _Py_MAX_ALLOWED_BUILTINS_MODIFICATIONS) {
         _Py_Executors_InvalidateAll(interp, 1);
     }
+#endif
     RARE_EVENT_INTERP_INC(interp, builtin_dict);
     return 0;
 }
@@ -1272,30 +1274,30 @@ init_interp_main(PyThreadState *tstate)
     }
 
     // Turn on experimental tier 2 (uops-based) optimizer
+    // This is also needed when the JIT is enabled
+#ifdef _Py_TIER2
     if (is_main_interp) {
-#ifndef _Py_JIT
-        // No JIT, maybe use the tier two interpreter:
-        char *envvar = Py_GETENV("PYTHON_UOPS");
-        int enabled = envvar != NULL && *envvar > '0';
-        if (_Py_get_xoption(&config->xoptions, L"uops") != NULL) {
-            enabled = 1;
+        int enabled = 1;
+#if _Py_TIER2 & 2
+        enabled = 0;
+#endif
+        char *env = Py_GETENV("PYTHON_JIT");
+        if (env && *env != '\0') {
+            // PYTHON_JIT=0|1 overrides the default
+            enabled = *env != '0';
         }
         if (enabled) {
-#else
-        // Always enable tier two for JIT builds (ignoring the environment
-        // variable and command-line option above):
-        if (true) {
-#endif
             PyObject *opt = PyUnstable_Optimizer_NewUOpOptimizer();
             if (opt == NULL) {
                 return _PyStatus_ERR("can't initialize optimizer");
             }
             if (PyUnstable_SetOptimizer((_PyOptimizerObject *)opt)) {
-                return _PyStatus_ERR("can't initialize optimizer");
+                return _PyStatus_ERR("can't install optimizer");
             }
             Py_DECREF(opt);
         }
     }
+#endif
 
     if (!is_main_interp) {
         // The main interpreter is handled in Py_Main(), for now.
@@ -1655,10 +1657,12 @@ finalize_modules(PyThreadState *tstate)
 {
     PyInterpreterState *interp = tstate->interp;
 
+#ifdef _Py_TIER2
     // Invalidate all executors and turn off tier 2 optimizer
     _Py_Executors_InvalidateAll(interp, 0);
     _PyOptimizerObject *old = _Py_SetOptimizer(interp, NULL);
     Py_XDECREF(old);
+#endif
 
     // Stop watching __builtin__ modifications
     PyDict_Unwatch(0, interp->builtins);

Python/pystate.c

@@ -653,8 +653,10 @@ init_interpreter(PyInterpreterState *interp,
     }
     interp->sys_profile_initialized = false;
     interp->sys_trace_initialized = false;
+#ifdef _Py_TIER2
     (void)_Py_SetOptimizer(interp, NULL);
     interp->executor_list_head = NULL;
+#endif
     if (interp != &runtime->_main_interpreter) {
         /* Fix the self-referential, statically initialized fields. */
         interp->dtoa = (struct _dtoa_state)_dtoa_state_INIT(interp);
@@ -806,9 +808,11 @@ interpreter_clear(PyInterpreterState *interp, PyThreadState *tstate)
         tstate->_status.cleared = 0;
     }
 
+#ifdef _Py_TIER2
     _PyOptimizerObject *old = _Py_SetOptimizer(interp, NULL);
     assert(old != NULL);
     Py_DECREF(old);
+#endif
 
     /* It is possible that any of the objects below have a finalizer
        that runs Python code or otherwise relies on a thread state
@@ -2821,9 +2825,11 @@ _PyInterpreterState_SetEvalFrameFunc(PyInterpreterState *interp,
     if (eval_frame == interp->eval_frame) {
         return;
     }
+#ifdef _Py_TIER2
     if (eval_frame != NULL) {
         _Py_Executors_InvalidateAll(interp, 1);
     }
+#endif
     RARE_EVENT_INC(set_eval_frame_func);
     interp->eval_frame = eval_frame;
 }

Python/sysmodule.c

@@ -2165,8 +2165,10 @@ static PyObject *
 sys__clear_internal_caches_impl(PyObject *module)
 /*[clinic end generated code: output=0ee128670a4966d6 input=253e741ca744f6e8]*/
 {
+#ifdef _Py_TIER2
     PyInterpreterState *interp = _PyInterpreterState_GET();
     _Py_Executors_InvalidateAll(interp, 0);
+#endif
     PyType_ClearCache();
     Py_RETURN_NONE;
 }