GH-98686: Quicken everything (GH-98687)

Author: brandtbucher

Include/cpython/code.h

@@ -70,7 +70,6 @@ typedef struct {
     PyObject *co_exceptiontable;   /* Byte string encoding exception handling  \
                                       table */                                 \
     int co_flags;                  /* CO_..., see below */                     \
-    short co_warmup;                 /* Warmup counter for quickening */       \
     short _co_linearray_entry_size;  /* Size of each entry in _co_linearray */ \
                                                                                \
     /* The rest are not so impactful on performance. */                        \

Include/internal/pycore_code.h

@@ -91,28 +91,8 @@ typedef struct {
 
 #define INLINE_CACHE_ENTRIES_FOR_ITER CACHE_ENTRIES(_PyForIterCache)
 
-#define QUICKENING_WARMUP_DELAY 8
-
-/* We want to compare to zero for efficiency, so we offset values accordingly */
-#define QUICKENING_INITIAL_WARMUP_VALUE (-QUICKENING_WARMUP_DELAY)
-
-void _PyCode_Quicken(PyCodeObject *code);
-
-static inline void
-_PyCode_Warmup(PyCodeObject *code)
-{
-    if (code->co_warmup != 0) {
-        code->co_warmup++;
-        if (code->co_warmup == 0) {
-            _PyCode_Quicken(code);
-        }
-    }
-}
-
 extern uint8_t _PyOpcode_Adaptive[256];
 
-extern Py_ssize_t _Py_QuickenedCount;
-
 // Borrowed references to common callables:
 struct callable_cache {
     PyObject *isinstance;
@@ -252,10 +232,10 @@ extern void _Py_Specialize_UnpackSequence(PyObject *seq, _Py_CODEUNIT *instr,
                                           int oparg);
 extern void _Py_Specialize_ForIter(PyObject *iter, _Py_CODEUNIT *instr);
 
-/* Deallocator function for static codeobjects used in deepfreeze.py */
-extern void _PyStaticCode_Dealloc(PyCodeObject *co);
-/* Function to intern strings of codeobjects */
-extern int _PyStaticCode_InternStrings(PyCodeObject *co);
+/* Finalizer function for static codeobjects used in deepfreeze.py */
+extern void _PyStaticCode_Fini(PyCodeObject *co);
+/* Function to intern strings of codeobjects and quicken the bytecode */
+extern int _PyStaticCode_Init(PyCodeObject *co);
 
 #ifdef Py_STATS
 
@@ -397,8 +377,8 @@ write_location_entry_start(uint8_t *ptr, int code, int length)
 
 /* With a 16-bit counter, we have 12 bits for the counter value, and 4 bits for the backoff */
 #define ADAPTIVE_BACKOFF_BITS 4
-/* The initial counter value is 31 == 2**ADAPTIVE_BACKOFF_START - 1 */
-#define ADAPTIVE_BACKOFF_START 5
+/* The initial counter value is 1 == 2**ADAPTIVE_BACKOFF_START - 1 */
+#define ADAPTIVE_BACKOFF_START 1
 
 #define MAX_BACKOFF_VALUE (16 - ADAPTIVE_BACKOFF_BITS)
 

Include/internal/pycore_opcode.h

@@ -327,9 +327,6 @@ def pseudo_op(name, op, real_ops):
         "FOR_ITER_LIST",
         "FOR_ITER_RANGE",
     ],
-    "JUMP_BACKWARD": [
-        "JUMP_BACKWARD_QUICK",
-    ],
     "LOAD_ATTR": [
         "LOAD_ATTR_ADAPTIVE",
         # These potentially push [NULL, bound method] onto the stack.
@@ -358,9 +355,6 @@ def pseudo_op(name, op, real_ops):
         "LOAD_GLOBAL_BUILTIN",
         "LOAD_GLOBAL_MODULE",
     ],
-    "RESUME": [
-        "RESUME_QUICK",
-    ],
     "STORE_ATTR": [
         "STORE_ATTR_ADAPTIVE",
         "STORE_ATTR_INSTANCE_VALUE",

Include/opcode.h

@@ -73,7 +73,6 @@ def get_pooled_int(value):
     fd_deltas = [0] * repcount
     getallocatedblocks = sys.getallocatedblocks
     gettotalrefcount = sys.gettotalrefcount
-    _getquickenedcount = sys._getquickenedcount
     fd_count = os_helper.fd_count
     # initialize variables to make pyflakes quiet
     rc_before = alloc_before = fd_before = 0
@@ -93,7 +92,7 @@ def get_pooled_int(value):
         support.gc_collect()
 
         # Read memory statistics immediately after the garbage collection
-        alloc_after = getallocatedblocks() - _getquickenedcount()
+        alloc_after = getallocatedblocks()
         rc_after = gettotalrefcount()
         fd_after = fd_count()
 

Lib/opcode.py

@@ -580,7 +580,7 @@ def testfunction_kw(self, *, kw):
     return self
 
 
-QUICKENING_WARMUP_DELAY = 8
+ADAPTIVE_WARMUP_DELAY = 2
 
 
 class TestPEP590(unittest.TestCase):
@@ -771,7 +771,7 @@ def f(num): return num + 1
         assert_equal(11, f(num))
         function_setvectorcall(f)
         # make sure specializer is triggered by running > 50 times
-        for _ in range(10 * QUICKENING_WARMUP_DELAY):
+        for _ in range(10 * ADAPTIVE_WARMUP_DELAY):
             assert_equal("overridden", f(num))
 
     def test_setvectorcall_load_attr_specialization_skip(self):
@@ -787,7 +787,7 @@ def __getattribute__(self, attr):
         function_setvectorcall(X.__getattribute__)
         # make sure specialization doesn't trigger
         # when vectorcall is overridden
-        for _ in range(QUICKENING_WARMUP_DELAY):
+        for _ in range(ADAPTIVE_WARMUP_DELAY):
             assert_equal("overridden", x.a)
 
     def test_setvectorcall_load_attr_specialization_deopt(self):
@@ -803,12 +803,12 @@ def get_a(x):
         assert_equal = self.assertEqual
         x = X()
         # trigger LOAD_ATTR_GETATTRIBUTE_OVERRIDDEN specialization
-        for _ in range(QUICKENING_WARMUP_DELAY):
+        for _ in range(ADAPTIVE_WARMUP_DELAY):
             assert_equal("a", get_a(x))
         function_setvectorcall(X.__getattribute__)
         # make sure specialized LOAD_ATTR_GETATTRIBUTE_OVERRIDDEN
         # gets deopted due to overridden vectorcall
-        for _ in range(QUICKENING_WARMUP_DELAY):
+        for _ in range(ADAPTIVE_WARMUP_DELAY):
             assert_equal("overridden", get_a(x))
 
     @requires_limited_api