Remove old shim code

Author: markshannon

Include/internal/pycore_ceval.h

@@ -123,12 +123,19 @@ _PyEval_EvalFrame(PyThreadState *tstate, _PyInterpreterFrame *frame, int throwfl
     return tstate->interp->eval_frame(tstate, frame, throwflag);
 }
 
-#if defined(_Py_TIER2) && !defined(_Py_JIT)
+#ifdef _Py_TIER2
+#ifdef _Py_JIT
+_Py_CODEUNIT *_Py_LazyJitTrampoline(
+    struct _PyExecutorObject *current_executor, _PyInterpreterFrame *frame,
+    _PyStackRef *stack_pointer, PyThreadState *tstate
+);
+#else
 _Py_CODEUNIT *_PyTier2Interpreter(
     struct _PyExecutorObject *current_executor, _PyInterpreterFrame *frame,
     _PyStackRef *stack_pointer, PyThreadState *tstate
 );
 #endif
+#endif
 
 extern _PyJitEntryFuncPtr _Py_jit_entry;
 

Include/internal/pycore_optimizer.h

@@ -82,7 +82,6 @@ typedef struct _PyExecutorObject {
     uint32_t code_size;
     size_t jit_size;
     void *jit_code;
-    void *jit_side_entry;
     _PyExitData exits[1];
 } _PyExecutorObject;
 

Python/jit.c

@@ -494,10 +494,6 @@ _PyJIT_Compile(_PyExecutorObject *executor, const _PyUOpInstruction trace[], siz
     size_t code_size = 0;
     size_t data_size = 0;
     jit_state state = {0};
-    group = &shim;
-    code_size += group->code_size;
-    data_size += group->data_size;
-    combine_symbol_mask(group->trampoline_mask, state.trampolines.mask);
     for (size_t i = 0; i < length; i++) {
         const _PyUOpInstruction *instruction = &trace[i];
         group = &stencil_groups[instruction->opcode];
@@ -539,13 +535,6 @@ _PyJIT_Compile(_PyExecutorObject *executor, const _PyUOpInstruction trace[], siz
     unsigned char *code = memory;
     state.trampolines.mem = memory + code_size;
     unsigned char *data = memory + code_size + state.trampolines.size + code_padding;
-    // Compile the shim, which handles converting between the native
-    // calling convention and the calling convention used by jitted code
-    // (which may be different for efficiency reasons).
-    group = &shim;
-    group->emit(code, data, executor, NULL, &state);
-    code += group->code_size;
-    data += group->data_size;
     assert(trace[0].opcode == _START_EXECUTOR || trace[0].opcode == _COLD_EXIT);
     for (size_t i = 0; i < length; i++) {
         const _PyUOpInstruction *instruction = &trace[i];
@@ -566,7 +555,6 @@ _PyJIT_Compile(_PyExecutorObject *executor, const _PyUOpInstruction trace[], siz
         return -1;
     }
     executor->jit_code = memory;
-    executor->jit_side_entry = memory + shim.code_size;
     executor->jit_size = total_size;
     return 0;
 }
@@ -615,19 +603,23 @@ compile_trampoline(void)
     return (_PyJitEntryFuncPtr)memory;
 }
 
+static PyMutex lazy_jit_mutex = { 0 };
 
 _Py_CODEUNIT *
 _Py_LazyJitTrampoline(
     _PyExecutorObject *executor, _PyInterpreterFrame *frame, _PyStackRef *stack_pointer, PyThreadState *tstate
 ) {
-    assert(_Py_jit_entry == _Py_LazyJitTrampoline);
-    PyInterpreterState *interp = PyInterpreterState_Get();
-    _PyJitEntryFuncPtr trampoline = compile_trampoline();
-    if (trampoline == NULL) {
-        Py_FatalError("Cannot allocate core JIT code");
+    PyMutex_Lock(&lazy_jit_mutex);
+    if (_Py_jit_entry == _Py_LazyJitTrampoline) {
+        _PyJitEntryFuncPtr trampoline = compile_trampoline();
+        if (trampoline == NULL) {
+            PyMutex_Unlock(&lazy_jit_mutex);
+            Py_FatalError("Cannot allocate core JIT code");
+        }
+        _Py_jit_entry = trampoline;
     }
-    _Py_jit_entry = trampoline;
-    return trampoline(executor, frame, stack_pointer, tstate);
+    PyMutex_Unlock(&lazy_jit_mutex);
+    return _Py_jit_entry(executor, frame, stack_pointer, tstate);
 }
 
 void
@@ -637,7 +629,6 @@ _PyJIT_Free(_PyExecutorObject *executor)
     size_t size = executor->jit_size;
     if (memory) {
         executor->jit_code = NULL;
-        executor->jit_side_entry = NULL;
         executor->jit_size = 0;
         if (jit_free(memory, size)) {
             PyErr_FormatUnraisable("Exception ignored while "

Python/optimizer.c

@@ -1238,7 +1238,6 @@ make_executor_from_uops(_PyUOpInstruction *buffer, int length, const _PyBloomFil
 #endif
 #ifdef _Py_JIT
     executor->jit_code = NULL;
-    executor->jit_side_entry = NULL;
     executor->jit_size = 0;
     // This is initialized to true so we can prevent the executor
     // from being immediately detected as cold and invalidated.
@@ -1490,7 +1489,6 @@ _PyExecutor_GetColdExecutor(void)
     ((_PyUOpInstruction *)cold->trace)->opcode = _COLD_EXIT;
 #ifdef _Py_JIT
     cold->jit_code = NULL;
-    cold->jit_side_entry = NULL;
     cold->jit_size = 0;
     // This is initialized to true so we can prevent the executor
     // from being immediately detected as cold and invalidated.

Tools/jit/_targets.py

@@ -191,8 +191,6 @@ async def _build_stencils(self) -> dict[str, _stencils.StencilGroup]:
         with tempfile.TemporaryDirectory() as tempdir:
             work = pathlib.Path(tempdir).resolve()
             async with asyncio.TaskGroup() as group:
-                coro = self._compile("shim", TOOLS_JIT / "shim.c", work)
-                tasks.append(group.create_task(coro, name="shim"))
                 coro = self._compile("trampoline", TOOLS_JIT / "trampoline.c", work)
                 tasks.append(group.create_task(coro, name="trampoline"))
                 template = TOOLS_JIT_TEMPLATE_C.read_text()

Tools/jit/_writer.py

@@ -22,12 +22,11 @@ def _dump_footer(
     yield "    symbol_mask trampoline_mask;"
     yield "} StencilGroup;"
     yield ""
-    yield f"static const StencilGroup shim = {groups['shim'].as_c('shim')};"
     yield f"static const StencilGroup trampoline = {groups['trampoline'].as_c('trampoline')};"
     yield ""
     yield "static const StencilGroup stencil_groups[MAX_UOP_ID + 1] = {"
     for opname, group in sorted(groups.items()):
-        if opname in ("shim", "trampoline"):
+        if opname == "trampoline":
             continue
         yield f"    [{opname}] = {group.as_c(opname)},"
     yield "};"

Tools/jit/shim.c

@@ -47,11 +47,11 @@
 #define CURRENT_TARGET() (_target)
 
 #undef TIER2_TO_TIER2
-#define TIER2_TO_TIER2(EXECUTOR)                                            \
+#define TIER2_TO_TIER2(EXECUTOR)                                           \
 do {                                                                       \
     OPT_STAT_INC(traces_executed);                                         \
     _PyExecutorObject *_executor = (EXECUTOR);                             \
-    jit_func_preserve_none jitted = _executor->jit_side_entry;             \
+    jit_func_preserve_none jitted = _executor->jit_code;                   \
     __attribute__((musttail)) return jitted(frame, stack_pointer, tstate); \
 } while (0)