From 142502ea8d26b17732009b6e981e630c342054f7 Mon Sep 17 00:00:00 2001
From: Guido van Rossum <guido@python.org>
Date: Tue, 20 Feb 2024 12:24:35 -0800
Subject: Tier 2 cleanups and tweaks (#115534)

* Rename `_testinternalcapi.get_{uop,counter}_optimizer` to `new_*_optimizer`
* Use `_PyUOpName()` instead of` _PyOpcode_uop_name[]`
* Add `target` to executor iterator items -- `list(ex)` now returns `(opcode, oparg, target, operand)` quadruples
* Add executor methods `get_opcode()` and `get_oparg()` to get `vmdata.opcode`, `vmdata.oparg`
* Define a helper for printing uops, and unify various places where they are printed
* Add a hack to summarize_stats.py to fix legacy uop names (e.g. `POP_TOP` -> `_POP_TOP`)
* Define helpers in `test_opt.py` for accessing the set or list of opnames of an executor
---
 Lib/test/test_capi/test_opt.py   | 157 +++++++++++++++++++++------------------
 Lib/test/test_monitoring.py      |   2 +-
 Modules/_testinternalcapi.c      |   8 +-
 Python/ceval.c                   |  61 +++++++++------
 Python/optimizer.c               |  77 +++++++++++++++----
 Python/optimizer_analysis.c      |   3 +-
 Python/specialize.c              |  12 +--
 Tools/scripts/summarize_stats.py |   4 +
 8 files changed, 203 insertions(+), 121 deletions(-)

diff --git a/Lib/test/test_capi/test_opt.py b/Lib/test/test_capi/test_opt.py
index d53d6a0..38c6fa4 100644
--- a/Lib/test/test_capi/test_opt.py
+++ b/Lib/test/test_capi/test_opt.py
@@ -32,16 +32,16 @@ def clear_executors(func):
 
 class TestOptimizerAPI(unittest.TestCase):
 
-    def test_get_counter_optimizer_dealloc(self):
+    def test_new_counter_optimizer_dealloc(self):
         # See gh-108727
         def f():
-            _testinternalcapi.get_counter_optimizer()
+            _testinternalcapi.new_counter_optimizer()
 
         f()
 
     def test_get_set_optimizer(self):
         old = _testinternalcapi.get_optimizer()
-        opt = _testinternalcapi.get_counter_optimizer()
+        opt = _testinternalcapi.new_counter_optimizer()
         try:
             _testinternalcapi.set_optimizer(opt)
             self.assertEqual(_testinternalcapi.get_optimizer(), opt)
@@ -62,7 +62,7 @@ class TestOptimizerAPI(unittest.TestCase):
         loop = ns['loop']
 
         for repeat in range(5):
-            opt = _testinternalcapi.get_counter_optimizer()
+            opt = _testinternalcapi.new_counter_optimizer()
             with temporary_optimizer(opt):
                 self.assertEqual(opt.get_count(), 0)
                 with clear_executors(loop):
@@ -90,7 +90,7 @@ class TestOptimizerAPI(unittest.TestCase):
         """), ns, ns)
         long_loop = ns['long_loop']
 
-        opt = _testinternalcapi.get_counter_optimizer()
+        opt = _testinternalcapi.new_counter_optimizer()
         with temporary_optimizer(opt):
             self.assertEqual(opt.get_count(), 0)
             long_loop()
@@ -102,7 +102,7 @@ class TestOptimizerAPI(unittest.TestCase):
             while i < x:
                 i += 1
 
-        opt = _testinternalcapi.get_counter_optimizer()
+        opt = _testinternalcapi.new_counter_optimizer()
         with temporary_optimizer(opt):
             testfunc(1000)
             code, replace_code  = testfunc.__code__, testfunc.__code__.replace()
@@ -123,11 +123,20 @@ def get_first_executor(func):
     return None
 
 
+def iter_opnames(ex):
+    for item in ex:
+        yield item[0]
+
+
+def get_opnames(ex):
+    return set(iter_opnames(ex))
+
+
 class TestExecutorInvalidation(unittest.TestCase):
 
     def setUp(self):
         self.old = _testinternalcapi.get_optimizer()
-        self.opt = _testinternalcapi.get_counter_optimizer()
+        self.opt = _testinternalcapi.new_counter_optimizer()
         _testinternalcapi.set_optimizer(self.opt)
 
     def tearDown(self):
@@ -176,7 +185,7 @@ class TestExecutorInvalidation(unittest.TestCase):
                     pass
         """), ns, ns)
         f = ns['f']
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         with temporary_optimizer(opt):
             f()
         exe = get_first_executor(f)
@@ -189,7 +198,7 @@ class TestExecutorInvalidation(unittest.TestCase):
         def f():
             for _ in range(1000):
                 pass
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         with temporary_optimizer(opt):
             f()
         exe = get_first_executor(f)
@@ -208,13 +217,13 @@ class TestUops(unittest.TestCase):
             while i < x:
                 i += 1
 
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         with temporary_optimizer(opt):
             testfunc(1000)
 
         ex = get_first_executor(testfunc)
         self.assertIsNotNone(ex)
-        uops = {opname for opname, _, _ in ex}
+        uops = get_opnames(ex)
         self.assertIn("_SET_IP", uops)
         self.assertIn("_LOAD_FAST_0", uops)
 
@@ -255,7 +264,7 @@ class TestUops(unittest.TestCase):
         """), ns, ns)
         many_vars = ns["many_vars"]
 
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         with temporary_optimizer(opt):
             ex = get_first_executor(many_vars)
             self.assertIsNone(ex)
@@ -263,7 +272,8 @@ class TestUops(unittest.TestCase):
 
         ex = get_first_executor(many_vars)
         self.assertIsNotNone(ex)
-        self.assertIn(("_LOAD_FAST", 259, 0), list(ex))
+        self.assertTrue(any((opcode, oparg, operand) == ("_LOAD_FAST", 259, 0)
+                            for opcode, oparg, _, operand in list(ex)))
 
     def test_unspecialized_unpack(self):
         # An example of an unspecialized opcode
@@ -277,14 +287,14 @@ class TestUops(unittest.TestCase):
             while i < x:
                 i += 1
 
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
 
         with temporary_optimizer(opt):
             testfunc(20)
 
         ex = get_first_executor(testfunc)
         self.assertIsNotNone(ex)
-        uops = {opname for opname, _, _ in ex}
+        uops = get_opnames(ex)
         self.assertIn("_UNPACK_SEQUENCE", uops)
 
     def test_pop_jump_if_false(self):
@@ -293,13 +303,13 @@ class TestUops(unittest.TestCase):
             while i < n:
                 i += 1
 
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         with temporary_optimizer(opt):
             testfunc(20)
 
         ex = get_first_executor(testfunc)
         self.assertIsNotNone(ex)
-        uops = {opname for opname, _, _ in ex}
+        uops = get_opnames(ex)
         self.assertIn("_GUARD_IS_TRUE_POP", uops)
 
     def test_pop_jump_if_none(self):
@@ -308,13 +318,13 @@ class TestUops(unittest.TestCase):
                 if x is None:
                     x = 0
 
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         with temporary_optimizer(opt):
             testfunc(range(20))
 
         ex = get_first_executor(testfunc)
         self.assertIsNotNone(ex)
-        uops = {opname for opname, _, _ in ex}
+        uops = get_opnames(ex)
         self.assertIn("_GUARD_IS_NOT_NONE_POP", uops)
 
     def test_pop_jump_if_not_none(self):
@@ -324,13 +334,13 @@ class TestUops(unittest.TestCase):
                 if x is not None:
                     x = 0
 
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         with temporary_optimizer(opt):
             testfunc(range(20))
 
         ex = get_first_executor(testfunc)
         self.assertIsNotNone(ex)
-        uops = {opname for opname, _, _ in ex}
+        uops = get_opnames(ex)
         self.assertIn("_GUARD_IS_NONE_POP", uops)
 
     def test_pop_jump_if_true(self):
@@ -339,13 +349,13 @@ class TestUops(unittest.TestCase):
             while not i >= n:
                 i += 1
 
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         with temporary_optimizer(opt):
             testfunc(20)
 
         ex = get_first_executor(testfunc)
         self.assertIsNotNone(ex)
-        uops = {opname for opname, _, _ in ex}
+        uops = get_opnames(ex)
         self.assertIn("_GUARD_IS_FALSE_POP", uops)
 
     def test_jump_backward(self):
@@ -354,13 +364,13 @@ class TestUops(unittest.TestCase):
             while i < n:
                 i += 1
 
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         with temporary_optimizer(opt):
             testfunc(20)
 
         ex = get_first_executor(testfunc)
         self.assertIsNotNone(ex)
-        uops = {opname for opname, _, _ in ex}
+        uops = get_opnames(ex)
         self.assertIn("_JUMP_TO_TOP", uops)
 
     def test_jump_forward(self):
@@ -374,13 +384,13 @@ class TestUops(unittest.TestCase):
                 a += 1
             return a
 
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         with temporary_optimizer(opt):
             testfunc(20)
 
         ex = get_first_executor(testfunc)
         self.assertIsNotNone(ex)
-        uops = {opname for opname, _, _ in ex}
+        uops = get_opnames(ex)
         # Since there is no JUMP_FORWARD instruction,
         # look for indirect evidence: the += operator
         self.assertIn("_BINARY_OP_ADD_INT", uops)
@@ -392,7 +402,7 @@ class TestUops(unittest.TestCase):
                 total += i
             return total
 
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         with temporary_optimizer(opt):
             total = testfunc(20)
             self.assertEqual(total, 190)
@@ -401,7 +411,7 @@ class TestUops(unittest.TestCase):
         self.assertIsNotNone(ex)
         # for i, (opname, oparg) in enumerate(ex):
         #     print(f"{i:4d}: {opname:<20s} {oparg:3d}")
-        uops = {opname for opname, _, _ in ex}
+        uops = get_opnames(ex)
         self.assertIn("_GUARD_NOT_EXHAUSTED_RANGE", uops)
         # Verification that the jump goes past END_FOR
         # is done by manual inspection of the output
@@ -413,7 +423,7 @@ class TestUops(unittest.TestCase):
                 total += i
             return total
 
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         with temporary_optimizer(opt):
             a = list(range(20))
             total = testfunc(a)
@@ -423,7 +433,7 @@ class TestUops(unittest.TestCase):
         self.assertIsNotNone(ex)
         # for i, (opname, oparg) in enumerate(ex):
         #     print(f"{i:4d}: {opname:<20s} {oparg:3d}")
-        uops = {opname for opname, _, _ in ex}
+        uops = get_opnames(ex)
         self.assertIn("_GUARD_NOT_EXHAUSTED_LIST", uops)
         # Verification that the jump goes past END_FOR
         # is done by manual inspection of the output
@@ -435,7 +445,7 @@ class TestUops(unittest.TestCase):
                 total += i
             return total
 
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         with temporary_optimizer(opt):
             a = tuple(range(20))
             total = testfunc(a)
@@ -445,7 +455,7 @@ class TestUops(unittest.TestCase):
         self.assertIsNotNone(ex)
         # for i, (opname, oparg) in enumerate(ex):
         #     print(f"{i:4d}: {opname:<20s} {oparg:3d}")
-        uops = {opname for opname, _, _ in ex}
+        uops = get_opnames(ex)
         self.assertIn("_GUARD_NOT_EXHAUSTED_TUPLE", uops)
         # Verification that the jump goes past END_FOR
         # is done by manual inspection of the output
@@ -455,7 +465,7 @@ class TestUops(unittest.TestCase):
             for x in it:
                 pass
 
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         with temporary_optimizer(opt):
             a = [1, 2, 3]
             it = iter(a)
@@ -471,13 +481,13 @@ class TestUops(unittest.TestCase):
             for i in range(n):
                 dummy(i)
 
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         with temporary_optimizer(opt):
             testfunc(20)
 
         ex = get_first_executor(testfunc)
         self.assertIsNotNone(ex)
-        uops = {opname for opname, _, _ in ex}
+        uops = get_opnames(ex)
         self.assertIn("_PUSH_FRAME", uops)
         self.assertIn("_BINARY_OP_ADD_INT", uops)
 
@@ -489,13 +499,13 @@ class TestUops(unittest.TestCase):
                 else:
                     i = 1
 
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         with temporary_optimizer(opt):
             testfunc(20)
 
         ex = get_first_executor(testfunc)
         self.assertIsNotNone(ex)
-        uops = {opname for opname, _, _ in ex}
+        uops = get_opnames(ex)
         self.assertIn("_GUARD_IS_FALSE_POP", uops)
 
     def test_for_iter_tier_two(self):
@@ -517,7 +527,7 @@ class TestUops(unittest.TestCase):
                     x += 1000*i + j
             return x
 
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         with temporary_optimizer(opt):
             x = testfunc(10, 10)
 
@@ -525,7 +535,7 @@ class TestUops(unittest.TestCase):
 
         ex = get_first_executor(testfunc)
         self.assertIsNotNone(ex)
-        uops = {opname for opname, _, _ in ex}
+        uops = get_opnames(ex)
         self.assertIn("_FOR_ITER_TIER_TWO", uops)
 
     def test_confidence_score(self):
@@ -546,14 +556,14 @@ class TestUops(unittest.TestCase):
                     bits += 1
             return bits
 
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         with temporary_optimizer(opt):
             x = testfunc(20)
 
         self.assertEqual(x, 40)
         ex = get_first_executor(testfunc)
         self.assertIsNotNone(ex)
-        ops = [opname for opname, _, _ in ex]
+        ops = list(iter_opnames(ex))
         #Since branch is 50/50 the trace could go either way.
         count = ops.count("_GUARD_IS_TRUE_POP") + ops.count("_GUARD_IS_FALSE_POP")
         self.assertLessEqual(count, 2)
@@ -562,7 +572,7 @@ class TestUopsOptimization(unittest.TestCase):
 
     def _run_with_optimizer(self, testfunc, arg):
         res = None
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         with temporary_optimizer(opt):
             res = testfunc(arg)
 
@@ -582,8 +592,8 @@ class TestUopsOptimization(unittest.TestCase):
         res, ex = self._run_with_optimizer(testfunc, 32)
         self.assertIsNotNone(ex)
         self.assertEqual(res, 63)
-        binop_count = [opname for opname, _, _ in ex if opname == "_BINARY_OP_ADD_INT"]
-        guard_both_int_count = [opname for opname, _, _ in ex if opname == "_GUARD_BOTH_INT"]
+        binop_count = [opname for opname in iter_opnames(ex) if opname == "_BINARY_OP_ADD_INT"]
+        guard_both_int_count = [opname for opname in iter_opnames(ex) if opname == "_GUARD_BOTH_INT"]
         self.assertGreaterEqual(len(binop_count), 3)
         self.assertLessEqual(len(guard_both_int_count), 1)
 
@@ -598,7 +608,7 @@ class TestUopsOptimization(unittest.TestCase):
                 num += 1
             return a
 
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         res = None
         with temporary_optimizer(opt):
             res = testfunc(32)
@@ -606,8 +616,8 @@ class TestUopsOptimization(unittest.TestCase):
         ex = get_first_executor(testfunc)
         self.assertIsNotNone(ex)
         self.assertEqual(res, 124)
-        binop_count = [opname for opname, _, _ in ex if opname == "_BINARY_OP_ADD_INT"]
-        guard_both_int_count = [opname for opname, _, _ in ex if opname == "_GUARD_BOTH_INT"]
+        binop_count = [opname for opname in iter_opnames(ex) if opname == "_BINARY_OP_ADD_INT"]
+        guard_both_int_count = [opname for opname in iter_opnames(ex) if opname == "_GUARD_BOTH_INT"]
         self.assertGreaterEqual(len(binop_count), 3)
         self.assertLessEqual(len(guard_both_int_count), 1)
 
@@ -622,7 +632,7 @@ class TestUopsOptimization(unittest.TestCase):
                 num += 1
             return x
 
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         res = None
         with temporary_optimizer(opt):
             res = testfunc(32)
@@ -630,8 +640,8 @@ class TestUopsOptimization(unittest.TestCase):
         ex = get_first_executor(testfunc)
         self.assertIsNotNone(ex)
         self.assertEqual(res, 124)
-        binop_count = [opname for opname, _, _ in ex if opname == "_BINARY_OP_ADD_INT"]
-        guard_both_int_count = [opname for opname, _, _ in ex if opname == "_GUARD_BOTH_INT"]
+        binop_count = [opname for opname in iter_opnames(ex) if opname == "_BINARY_OP_ADD_INT"]
+        guard_both_int_count = [opname for opname in iter_opnames(ex) if opname == "_GUARD_BOTH_INT"]
         self.assertGreaterEqual(len(binop_count), 3)
         self.assertLessEqual(len(guard_both_int_count), 1)
 
@@ -648,7 +658,7 @@ class TestUopsOptimization(unittest.TestCase):
 
         res, ex = self._run_with_optimizer(testfunc, 64)
         self.assertIsNotNone(ex)
-        binop_count = [opname for opname, _, _ in ex if opname == "_BINARY_OP_ADD_INT"]
+        binop_count = [opname for opname in iter_opnames(ex) if opname == "_BINARY_OP_ADD_INT"]
         self.assertGreaterEqual(len(binop_count), 3)
 
     def test_call_py_exact_args(self):
@@ -660,7 +670,7 @@ class TestUopsOptimization(unittest.TestCase):
 
         res, ex = self._run_with_optimizer(testfunc, 32)
         self.assertIsNotNone(ex)
-        uops = {opname for opname, _, _ in ex}
+        uops = get_opnames(ex)
         self.assertIn("_PUSH_FRAME", uops)
         self.assertIn("_BINARY_OP_ADD_INT", uops)
         self.assertNotIn("_CHECK_PEP_523", uops)
@@ -675,7 +685,7 @@ class TestUopsOptimization(unittest.TestCase):
         res, ex = self._run_with_optimizer(testfunc, 32)
         self.assertEqual(res, 62)
         self.assertIsNotNone(ex)
-        uops = {opname for opname, _, _ in ex}
+        uops = get_opnames(ex)
         self.assertNotIn("_GUARD_BOTH_INT", uops)
 
     def test_int_value_numbering(self):
@@ -693,9 +703,9 @@ class TestUopsOptimization(unittest.TestCase):
         res, ex = self._run_with_optimizer(testfunc, 32)
         self.assertEqual(res, 4)
         self.assertIsNotNone(ex)
-        uops = {opname for opname, _, _ in ex}
+        uops = get_opnames(ex)
         self.assertIn("_GUARD_BOTH_INT", uops)
-        guard_count = [opname for opname, _, _ in ex if opname == "_GUARD_BOTH_INT"]
+        guard_count = [opname for opname in iter_opnames(ex) if opname == "_GUARD_BOTH_INT"]
         self.assertEqual(len(guard_count), 1)
 
     def test_comprehension(self):
@@ -706,7 +716,7 @@ class TestUopsOptimization(unittest.TestCase):
         res, ex = self._run_with_optimizer(testfunc, 32)
         self.assertEqual(res, list(range(32)))
         self.assertIsNotNone(ex)
-        uops = {opname for opname, _, _ in ex}
+        uops = get_opnames(ex)
         self.assertNotIn("_BINARY_OP_ADD_INT", uops)
 
     def test_call_py_exact_args_disappearing(self):
@@ -717,7 +727,7 @@ class TestUopsOptimization(unittest.TestCase):
             for i in range(n):
                 dummy(i)
 
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         # Trigger specialization
         testfunc(8)
         with temporary_optimizer(opt):
@@ -752,18 +762,21 @@ class TestUopsOptimization(unittest.TestCase):
                         pass
             return None
 
+        def get_opnames(ex):
+            return {item[0] for item in ex}
+
         def testfunc(n):
             for i in range(n):
                 x = range(i)
             return x
 
-        opt = _testinternalcapi.get_uop_optimizer()
+        opt = _testinternalcapi.new_uop_optimizer()
         _testinternalcapi.set_optimizer(opt)
         testfunc(64)
 
         ex = get_first_executor(testfunc)
         assert ex is not None
-        uops = {opname for opname, _, _ in ex}
+        uops = get_opnames(ex)
         assert "_LOAD_GLOBAL_BUILTINS" not in uops
         assert "_LOAD_CONST_INLINE_BORROW_WITH_NULL" in uops
         """))
@@ -779,8 +792,8 @@ class TestUopsOptimization(unittest.TestCase):
         res, ex = self._run_with_optimizer(testfunc, 32)
         self.assertAlmostEqual(res, 4.2)
         self.assertIsNotNone(ex)
-        uops = {opname for opname, _, _ in ex}
-        guard_both_float_count = [opname for opname, _, _ in ex if opname == "_GUARD_BOTH_FLOAT"]
+        uops = get_opnames(ex)
+        guard_both_float_count = [opname for opname in iter_opnames(ex) if opname == "_GUARD_BOTH_FLOAT"]
         self.assertLessEqual(len(guard_both_float_count), 1)
         # TODO gh-115506: this assertion may change after propagating constants.
         # We'll also need to verify that propagation actually occurs.
@@ -796,8 +809,8 @@ class TestUopsOptimization(unittest.TestCase):
         res, ex = self._run_with_optimizer(testfunc, 32)
         self.assertAlmostEqual(res, -2.2)
         self.assertIsNotNone(ex)
-        uops = {opname for opname, _, _ in ex}
-        guard_both_float_count = [opname for opname, _, _ in ex if opname == "_GUARD_BOTH_FLOAT"]
+        uops = get_opnames(ex)
+        guard_both_float_count = [opname for opname in iter_opnames(ex) if opname == "_GUARD_BOTH_FLOAT"]
         self.assertLessEqual(len(guard_both_float_count), 1)
         # TODO gh-115506: this assertion may change after propagating constants.
         # We'll also need to verify that propagation actually occurs.
@@ -813,8 +826,8 @@ class TestUopsOptimization(unittest.TestCase):
         res, ex = self._run_with_optimizer(testfunc, 32)
         self.assertAlmostEqual(res, 2 ** 32)
         self.assertIsNotNone(ex)
-        uops = {opname for opname, _, _ in ex}
-        guard_both_float_count = [opname for opname, _, _ in ex if opname == "_GUARD_BOTH_FLOAT"]
+        uops = get_opnames(ex)
+        guard_both_float_count = [opname for opname in iter_opnames(ex) if opname == "_GUARD_BOTH_FLOAT"]
         self.assertLessEqual(len(guard_both_float_count), 1)
         # TODO gh-115506: this assertion may change after propagating constants.
         # We'll also need to verify that propagation actually occurs.
@@ -833,8 +846,8 @@ class TestUopsOptimization(unittest.TestCase):
         res, ex = self._run_with_optimizer(testfunc, 32)
         self.assertTrue(res)
         self.assertIsNotNone(ex)
-        uops = {opname for opname, _, _ in ex}
-        guard_both_float_count = [opname for opname, _, _ in ex if opname == "_GUARD_BOTH_FLOAT"]
+        uops = get_opnames(ex)
+        guard_both_float_count = [opname for opname in iter_opnames(ex) if opname == "_GUARD_BOTH_FLOAT"]
         self.assertLessEqual(len(guard_both_float_count), 1)
         self.assertIn("_COMPARE_OP_FLOAT", uops)
 
@@ -851,8 +864,8 @@ class TestUopsOptimization(unittest.TestCase):
         res, ex = self._run_with_optimizer(testfunc, 32)
         self.assertTrue(res)
         self.assertIsNotNone(ex)
-        uops = {opname for opname, _, _ in ex}
-        guard_both_float_count = [opname for opname, _, _ in ex if opname == "_GUARD_BOTH_INT"]
+        uops = get_opnames(ex)
+        guard_both_float_count = [opname for opname in iter_opnames(ex) if opname == "_GUARD_BOTH_INT"]
         self.assertLessEqual(len(guard_both_float_count), 1)
         self.assertIn("_COMPARE_OP_INT", uops)
 
@@ -869,8 +882,8 @@ class TestUopsOptimization(unittest.TestCase):
         res, ex = self._run_with_optimizer(testfunc, 32)
         self.assertTrue(res)
         self.assertIsNotNone(ex)
-        uops = {opname for opname, _, _ in ex}
-        guard_both_float_count = [opname for opname, _, _ in ex if opname == "_GUARD_BOTH_UNICODE"]
+        uops = get_opnames(ex)
+        guard_both_float_count = [opname for opname in iter_opnames(ex) if opname == "_GUARD_BOTH_UNICODE"]
         self.assertLessEqual(len(guard_both_float_count), 1)
         self.assertIn("_COMPARE_OP_STR", uops)
 
diff --git a/Lib/test/test_monitoring.py b/Lib/test/test_monitoring.py
index 60b6326..58aa4bc 100644
--- a/Lib/test/test_monitoring.py
+++ b/Lib/test/test_monitoring.py
@@ -1799,7 +1799,7 @@ class TestOptimizer(MonitoringTestBase, unittest.TestCase):
     def setUp(self):
         import _testinternalcapi
         self.old_opt = _testinternalcapi.get_optimizer()
-        opt = _testinternalcapi.get_counter_optimizer()
+        opt = _testinternalcapi.new_counter_optimizer()
         _testinternalcapi.set_optimizer(opt)
         super(TestOptimizer, self).setUp()
 
diff --git a/Modules/_testinternalcapi.c b/Modules/_testinternalcapi.c
index bcc431a..0d23b18 100644
--- a/Modules/_testinternalcapi.c
+++ b/Modules/_testinternalcapi.c
@@ -960,13 +960,13 @@ iframe_getlasti(PyObject *self, PyObject *frame)
 }
 
 static PyObject *
-get_counter_optimizer(PyObject *self, PyObject *arg)
+new_counter_optimizer(PyObject *self, PyObject *arg)
 {
     return PyUnstable_Optimizer_NewCounter();
 }
 
 static PyObject *
-get_uop_optimizer(PyObject *self, PyObject *arg)
+new_uop_optimizer(PyObject *self, PyObject *arg)
 {
     return PyUnstable_Optimizer_NewUOpOptimizer();
 }
@@ -1711,8 +1711,8 @@ static PyMethodDef module_functions[] = {
     {"get_optimizer", get_optimizer,  METH_NOARGS, NULL},
     {"set_optimizer", set_optimizer,  METH_O, NULL},
     {"get_executor", _PyCFunction_CAST(get_executor),  METH_FASTCALL, NULL},
-    {"get_counter_optimizer", get_counter_optimizer, METH_NOARGS, NULL},
-    {"get_uop_optimizer", get_uop_optimizer, METH_NOARGS, NULL},
+    {"new_counter_optimizer", new_counter_optimizer, METH_NOARGS, NULL},
+    {"new_uop_optimizer", new_uop_optimizer, METH_NOARGS, NULL},
     {"add_executor_dependency", add_executor_dependency, METH_VARARGS, NULL},
     {"invalidate_executors", invalidate_executors, METH_O, NULL},
     {"pending_threadfunc", _PyCFunction_CAST(pending_threadfunc),
diff --git a/Python/ceval.c b/Python/ceval.c
index b7a5d62..06c136a 100644
--- a/Python/ceval.c
+++ b/Python/ceval.c
@@ -649,7 +649,10 @@ static const _Py_CODEUNIT _Py_INTERPRETER_TRAMPOLINE_INSTRUCTIONS[] = {
 
 extern const struct _PyCode_DEF(8) _Py_InitCleanup;
 
-extern const char *_PyUOpName(int index);
+#ifdef Py_DEBUG
+extern void _PyUOpPrint(const _PyUOpInstruction *uop);
+#endif
+
 
 /* Disable unused label warnings.  They are handy for debugging, even
    if computed gotos aren't used. */
@@ -1006,14 +1009,14 @@ enter_tier_two:
     assert(next_uop->opcode == _START_EXECUTOR || next_uop->opcode == _COLD_EXIT);
     for (;;) {
         uopcode = next_uop->opcode;
-        DPRINTF(3,
-                "%4d: uop %s, oparg %d, operand %" PRIu64 ", target %d, stack_level %d\n",
-                (int)(next_uop - (current_executor == NULL ? next_uop : current_executor->trace)),
-                _PyUOpName(uopcode),
-                next_uop->oparg,
-                next_uop->operand,
-                next_uop->target,
+#ifdef Py_DEBUG
+        if (lltrace >= 3) {
+            printf("%4d uop: ", (int)(next_uop - (current_executor == NULL ? next_uop : current_executor->trace)));
+            _PyUOpPrint(next_uop);
+            printf(" stack_level=%d\n",
                 (int)(stack_pointer - _PyFrame_Stackbase(frame)));
+        }
+#endif
         next_uop++;
         OPT_STAT_INC(uops_executed);
         UOP_STAT_INC(uopcode, execution_count);
@@ -1028,9 +1031,9 @@ enter_tier_two:
             default:
 #ifdef Py_DEBUG
             {
-                fprintf(stderr, "Unknown uop %d, oparg %d, operand %" PRIu64 " @ %d\n",
-                        next_uop[-1].opcode, next_uop[-1].oparg, next_uop[-1].operand,
-                        (int)(next_uop - (current_executor == NULL ? next_uop : current_executor->trace) - 1));
+                printf("Unknown uop: ");
+                _PyUOpPrint(&next_uop[-1]);
+                printf(" @ %d\n", (int)(next_uop - current_executor->trace - 1));
                 Py_FatalError("Unknown uop");
             }
 #else
@@ -1058,10 +1061,15 @@ pop_2_error_tier_two:
 pop_1_error_tier_two:
     STACK_SHRINK(1);
 error_tier_two:
-    DPRINTF(2, "Error: [UOp %d (%s), oparg %d, operand %" PRIu64 ", target %d @ %d -> %s]\n",
-            uopcode, _PyUOpName(uopcode), next_uop[-1].oparg, next_uop[-1].operand, next_uop[-1].target,
-            (int)(next_uop - current_executor->trace - 1),
-            _PyOpcode_OpName[frame->instr_ptr->op.code]);
+#ifdef Py_DEBUG
+    if (lltrace >= 2) {
+        printf("Error: [UOp ");
+        _PyUOpPrint(&next_uop[-1]);
+        printf(" @ %d -> %s]\n",
+               (int)(next_uop - current_executor->trace - 1),
+               _PyOpcode_OpName[frame->instr_ptr->op.code]);
+    }
+#endif
     OPT_HIST(trace_uop_execution_counter, trace_run_length_hist);
     frame->return_offset = 0;  // Don't leave this random
     _PyFrame_SetStackPointer(frame, stack_pointer);
@@ -1072,9 +1080,14 @@ error_tier_two:
 // Jump here from DEOPT_IF()
 deoptimize:
     next_instr = next_uop[-1].target + _PyCode_CODE(_PyFrame_GetCode(frame));
-    DPRINTF(2, "DEOPT: [UOp %d (%s), oparg %d, operand %" PRIu64 ", target %d -> %s]\n",
-            uopcode, _PyUOpName(uopcode), next_uop[-1].oparg, next_uop[-1].operand, next_uop[-1].target,
-            _PyOpcode_OpName[next_instr->op.code]);
+#ifdef Py_DEBUG
+    if (lltrace >= 2) {
+        printf("DEOPT: [UOp ");
+        _PyUOpPrint(&next_uop[-1]);
+        printf(" -> %s]\n",
+               _PyOpcode_OpName[frame->instr_ptr->op.code]);
+    }
+#endif
     OPT_HIST(trace_uop_execution_counter, trace_run_length_hist);
     UOP_STAT_INC(uopcode, miss);
     Py_DECREF(current_executor);
@@ -1088,9 +1101,15 @@ side_exit:
     uint32_t exit_index = next_uop[-1].exit_index;
     assert(exit_index < current_executor->exit_count);
     _PyExitData *exit = &current_executor->exits[exit_index];
-    DPRINTF(2, "SIDE EXIT: [UOp %d (%s), oparg %d, operand %" PRIu64 ", exit %u, temp %d, target %d -> %s]\n",
-            uopcode, _PyUOpName(uopcode), next_uop[-1].oparg, next_uop[-1].operand, exit_index, exit->temperature,
-            exit->target, _PyOpcode_OpName[_PyCode_CODE(_PyFrame_GetCode(frame))[exit->target].op.code]);
+#ifdef Py_DEBUG
+    if (lltrace >= 2) {
+        printf("SIDE EXIT: [UOp ");
+        _PyUOpPrint(&next_uop[-1]);
+        printf(", exit %u, temp %d, target %d -> %s]\n",
+               exit_index, exit->temperature, exit->target,
+               _PyOpcode_OpName[frame->instr_ptr->op.code]);
+    }
+#endif
     Py_INCREF(exit->executor);
     tstate->previous_executor = (PyObject *)current_executor;
     GOTO_TIER_TWO(exit->executor);
diff --git a/Python/optimizer.c b/Python/optimizer.c
index df8f0ed..74708be 100644
--- a/Python/optimizer.c
+++ b/Python/optimizer.c
@@ -262,8 +262,22 @@ is_valid(PyObject *self, PyObject *Py_UNUSED(ignored))
     return PyBool_FromLong(((_PyExecutorObject *)self)->vm_data.valid);
 }
 
+static PyObject *
+get_opcode(PyObject *self, PyObject *Py_UNUSED(ignored))
+{
+    return PyLong_FromUnsignedLong(((_PyExecutorObject *)self)->vm_data.opcode);
+}
+
+static PyObject *
+get_oparg(PyObject *self, PyObject *Py_UNUSED(ignored))
+{
+    return PyLong_FromUnsignedLong(((_PyExecutorObject *)self)->vm_data.oparg);
+}
+
 static PyMethodDef executor_methods[] = {
     { "is_valid", is_valid, METH_NOARGS, NULL },
+    { "get_opcode", get_opcode, METH_NOARGS, NULL },
+    { "get_oparg", get_oparg, METH_NOARGS, NULL },
     { NULL, NULL },
 };
 
@@ -282,9 +296,30 @@ uop_dealloc(_PyExecutorObject *self) {
 const char *
 _PyUOpName(int index)
 {
+    if (index < 0 || index > MAX_UOP_ID) {
+        return NULL;
+    }
     return _PyOpcode_uop_name[index];
 }
 
+#ifdef Py_DEBUG
+void
+_PyUOpPrint(const _PyUOpInstruction *uop)
+{
+    const char *name = _PyUOpName(uop->opcode);
+    if (name == NULL) {
+        printf("<uop %d>", uop->opcode);
+    }
+    else {
+        printf("%s", name);
+    }
+    printf(" (%d, target=%d, operand=%" PRIx64 ")",
+           uop->oparg,
+           uop->target,
+           (uint64_t)uop->operand);
+}
+#endif
+
 static Py_ssize_t
 uop_len(_PyExecutorObject *self)
 {
@@ -312,14 +347,21 @@ uop_item(_PyExecutorObject *self, Py_ssize_t index)
         Py_DECREF(oname);
         return NULL;
     }
+    PyObject *target = PyLong_FromUnsignedLong(self->trace[index].target);
+    if (oparg == NULL) {
+        Py_DECREF(oparg);
+        Py_DECREF(oname);
+        return NULL;
+    }
     PyObject *operand = PyLong_FromUnsignedLongLong(self->trace[index].operand);
     if (operand == NULL) {
+        Py_DECREF(target);
         Py_DECREF(oparg);
         Py_DECREF(oname);
         return NULL;
     }
-    PyObject *args[3] = { oname, oparg, operand };
-    return _PyTuple_FromArraySteal(args, 3);
+    PyObject *args[4] = { oname, oparg, target, operand };
+    return _PyTuple_FromArraySteal(args, 4);
 }
 
 PySequenceMethods uop_as_sequence = {
@@ -390,19 +432,29 @@ BRANCH_TO_GUARD[4][2] = {
 #endif
 
 
+// Beware: Macro arg order differs from struct member order
+#ifdef Py_DEBUG
 #define ADD_TO_TRACE(OPCODE, OPARG, OPERAND, TARGET) \
-    DPRINTF(2, \
-            "  ADD_TO_TRACE(%s, %d, %" PRIu64 ", %d)\n", \
-            _PyUOpName(OPCODE), \
-            (OPARG), \
-            (uint64_t)(OPERAND), \
-            TARGET); \
     assert(trace_length < max_length); \
     trace[trace_length].opcode = (OPCODE); \
     trace[trace_length].oparg = (OPARG); \
+    trace[trace_length].target = (TARGET); \
     trace[trace_length].operand = (OPERAND); \
+    if (lltrace >= 2) { \
+        printf("%4d ADD_TO_TRACE: ", trace_length); \
+        _PyUOpPrint(&trace[trace_length]); \
+        printf("\n"); \
+    } \
+    trace_length++;
+#else
+#define ADD_TO_TRACE(OPCODE, OPARG, OPERAND, TARGET) \
+    assert(trace_length < max_length); \
+    trace[trace_length].opcode = (OPCODE); \
+    trace[trace_length].oparg = (OPARG); \
     trace[trace_length].target = (TARGET); \
+    trace[trace_length].operand = (OPERAND); \
     trace_length++;
+#endif
 
 #define INSTR_IP(INSTR, CODE) \
     ((uint32_t)((INSTR) - ((_Py_CODEUNIT *)(CODE)->co_code_adaptive)))
@@ -890,12 +942,9 @@ make_executor_from_uops(_PyUOpInstruction *buffer, const _PyBloomFilter *depende
     if (lltrace >= 2) {
         printf("Optimized executor (length %d):\n", length);
         for (int i = 0; i < length; i++) {
-            printf("%4d %s(%d, %d, %" PRIu64 ")\n",
-                   i,
-                   _PyUOpName(executor->trace[i].opcode),
-                   executor->trace[i].oparg,
-                   executor->trace[i].target,
-                   executor->trace[i].operand);
+            printf("%4d OPTIMIZED: ", i);
+            _PyUOpPrint(&executor->trace[i]);
+            printf("\n");
         }
     }
 #endif
diff --git a/Python/optimizer_analysis.c b/Python/optimizer_analysis.c
index b104d2f..e7fb1e3 100644
--- a/Python/optimizer_analysis.c
+++ b/Python/optimizer_analysis.c
@@ -44,6 +44,7 @@
 #define MAX_ABSTRACT_FRAME_DEPTH (TRACE_STACK_SIZE + 2)
 
 #ifdef Py_DEBUG
+    extern const char *_PyUOpName(int index);
     static const char *const DEBUG_ENV = "PYTHON_OPT_DEBUG";
     static inline int get_lltrace(void) {
         char *uop_debug = Py_GETENV(DEBUG_ENV);
@@ -632,7 +633,7 @@ uop_redundancy_eliminator(
         _Py_UOpsSymType **stack_pointer = ctx->frame->stack_pointer;
 
         DPRINTF(3, "Abstract interpreting %s:%d ",
-                _PyOpcode_uop_name[opcode],
+                _PyUOpName(opcode),
                 oparg);
         switch (opcode) {
 #include "tier2_redundancy_eliminator_cases.c.h"
diff --git a/Python/specialize.c b/Python/specialize.c
index 2256d79..871979d 100644
--- a/Python/specialize.c
+++ b/Python/specialize.c
@@ -17,6 +17,7 @@
 
 #include <stdlib.h> // rand()
 
+extern const char *_PyUOpName(int index);
 
 /* For guidance on adding or extending families of instructions see
  * ./adaptive.md
@@ -246,17 +247,12 @@ print_optimization_stats(FILE *out, OptimizationStats *stats)
             stats->optimizer_failure_reason_no_memory);
 
     const char* const* names;
-    for (int i = 0; i < 512; i++) {
-        if (i < 256) {
-            names = _PyOpcode_OpName;
-        } else {
-            names = _PyOpcode_uop_name;
-        }
+    for (int i = 0; i <= MAX_UOP_ID; i++) {
         if (stats->opcode[i].execution_count) {
-            fprintf(out, "uops[%s].execution_count : %" PRIu64 "\n", names[i], stats->opcode[i].execution_count);
+            fprintf(out, "uops[%s].execution_count : %" PRIu64 "\n", _PyUOpName(i), stats->opcode[i].execution_count);
         }
         if (stats->opcode[i].miss) {
-            fprintf(out, "uops[%s].specialization.miss : %" PRIu64 "\n", names[i], stats->opcode[i].miss);
+            fprintf(out, "uops[%s].specialization.miss : %" PRIu64 "\n", _PyUOpName(i), stats->opcode[i].miss);
         }
     }
 
diff --git a/Tools/scripts/summarize_stats.py b/Tools/scripts/summarize_stats.py
index 5bc39fc..6b60b59 100644
--- a/Tools/scripts/summarize_stats.py
+++ b/Tools/scripts/summarize_stats.py
@@ -102,6 +102,10 @@ def load_raw_data(input: Path) -> RawData:
                             file=sys.stderr,
                         )
                         continue
+                    # Hack to handle older data files where some uops
+                    # are missing an underscore prefix in their name
+                    if key.startswith("uops[") and key[5:6] != "_":
+                        key = "uops[_" + key[5:]
                     stats[key.strip()] += int(value)
             stats["__nfiles__"] += 1
 
-- 
cgit v0.12