Use separate 'counters' and 'executors' arrays

(The latter as yet unused.)
python · Dec 12, 2023 · 65c41cb · 65c41cb
1 parent f4027de
commit 65c41cb
Show file tree

Hide file tree

Showing 3 changed files with 10 additions and 6 deletions.
diff --git a/Include/internal/pycore_uops.h b/Include/internal/pycore_uops.h
@@ -21,7 +21,9 @@ typedef struct {
 
 typedef struct {
     _PyExecutorObject base;
-    uintptr_t *extra;  // An array of uintptr_t of size base.ob_base.ob_size
+    // Auxiliary arrays, allocated after trace[base.ob_size]
+    uint16_t *counters;  // An array of counters
+    _PyExecutorObject **executors;  // An array of executors
     _PyUOpInstruction trace[1];
 } _PyUOpExecutorObject;
 

diff --git a/Python/ceval.c b/Python/ceval.c
@@ -1108,7 +1108,7 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
     _PyFrame_SetStackPointer(frame, stack_pointer);
     // Increment side exit counter for this uop
     int pc = next_uop - 1 - current_executor->trace;
-    uintptr_t *pcounter = current_executor->extra + pc;
+    uint16_t *pcounter = current_executor->counters + pc;
     *pcounter += 1;
     if (*pcounter == 16 &&  // TODO: use resume_threshold
         tstate->interp->optimizer != &_PyOptimizer_Default &&
@@ -1120,7 +1120,7 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
         DPRINTF(2, "--> %s @ %d in %p has %d side exits\n",
                 _PyUOpName(uopcode), pc, current_executor, (int)(*pcounter));
         DPRINTF(2, "    T1: %s\n", _PyOpcode_OpName[opcode]);
-        // The counter will cycle around in 2**64 executions :-)
+        // The counter will cycle around once the 16 bits overflow
         int optimized = _PyOptimizer_Anywhere(frame, src, dest, stack_pointer);
         if (optimized < 0) {
             goto error_tier_two;

diff --git a/Python/optimizer.c b/Python/optimizer.c
@@ -403,7 +403,7 @@ PyTypeObject _PyUOpExecutor_Type = {
     PyVarObject_HEAD_INIT(&PyType_Type, 0)
     .tp_name = "uop_executor",
     .tp_basicsize = sizeof(_PyUOpExecutorObject) - sizeof(_PyUOpInstruction),
-    .tp_itemsize = sizeof(_PyUOpInstruction) + sizeof(uintptr_t),
+    .tp_itemsize = sizeof(_PyUOpInstruction) + sizeof(uint16_t) + sizeof(_PyExecutorObject *),
     .tp_flags = Py_TPFLAGS_DEFAULT | Py_TPFLAGS_DISALLOW_INSTANTIATION,
     .tp_dealloc = (destructor)uop_dealloc,
     .tp_as_sequence = &uop_as_sequence,
@@ -835,8 +835,10 @@ make_executor_from_uops(_PyUOpInstruction *buffer, _PyBloomFilter *dependencies)
     if (executor == NULL) {
         return NULL;
     }
-    executor->extra = (uintptr_t *)(executor->trace + length);
-    memset(executor->extra, 0, sizeof(uintptr_t) * length);
+    executor->counters = (uint16_t *)(&executor->trace[length]);
+    memset(executor->counters, 0, sizeof(uint16_t) * length);
+    executor->executors = (_PyExecutorObject **)(&executor->counters[length]);
+    memset(executor->executors, 0, sizeof(_PyExecutorObject *) * length);
     int dest = length - 1;
     /* Scan backwards, so that we see the destinations of jumps before the jumps themselves. */
     for (int i = _Py_UOP_MAX_TRACE_LENGTH-1; i >= 0; i--) {