gh-134584: Specialize POP_TOP by reference and type in JIT (GH-135761)

Author: Fidget-Spinner

Include/internal/pycore_uop_ids.h

@@ -2392,6 +2392,46 @@ def testfunc(n):
         assert ex is not None
         """))
 
+    def test_pop_top_specialize_none(self):
+        def testfunc(n):
+            for _ in range(n):
+                global_identity(None)
+
+        testfunc(TIER2_THRESHOLD)
+
+        ex = get_first_executor(testfunc)
+        self.assertIsNotNone(ex)
+        uops = get_opnames(ex)
+
+        self.assertIn("_POP_TOP_NOP", uops)
+
+    def test_pop_top_specialize_int(self):
+        def testfunc(n):
+            for _ in range(n):
+                global_identity(100000)
+
+        testfunc(TIER2_THRESHOLD)
+
+        ex = get_first_executor(testfunc)
+        self.assertIsNotNone(ex)
+        uops = get_opnames(ex)
+
+        self.assertIn("_POP_TOP_INT", uops)
+
+    def test_pop_top_specialize_float(self):
+        def testfunc(n):
+            for _ in range(n):
+                global_identity(1e6)
+
+        testfunc(TIER2_THRESHOLD)
+
+        ex = get_first_executor(testfunc)
+        self.assertIsNotNone(ex)
+        uops = get_opnames(ex)
+
+        self.assertIn("_POP_TOP_FLOAT", uops)
+
+
 
 def global_identity(x):
     return x

Include/internal/pycore_uop_metadata.h

@@ -0,0 +1 @@
+Specialize :opcode:`POP_TOP` in the JIT compiler by specializing for reference lifetime and type. This will also enable easier top of stack caching in the JIT compiler.

Lib/test/test_capi/test_opt.py

@@ -344,6 +344,27 @@ dummy_func(
             PyStackRef_XCLOSE(value);
         }
 
+        op(_POP_TOP_NOP, (value --)) {
+            assert(PyStackRef_IsNull(value) || (!PyStackRef_RefcountOnObject(value)) ||
+                _Py_IsImmortal((PyStackRef_AsPyObjectBorrow(value))));
+            DEAD(value);
+        }
+
+        op(_POP_TOP_INT, (value --)) {
+            assert(PyLong_CheckExact(PyStackRef_AsPyObjectBorrow(value)));
+            PyStackRef_CLOSE_SPECIALIZED(value, _PyLong_ExactDealloc);
+        }
+
+        op(_POP_TOP_FLOAT, (value --)) {
+            assert(PyFloat_CheckExact(PyStackRef_AsPyObjectBorrow(value)));
+            PyStackRef_CLOSE_SPECIALIZED(value, _PyFloat_ExactDealloc);
+        }
+
+        op(_POP_TOP_UNICODE, (value --)) {
+            assert(PyUnicode_CheckExact(PyStackRef_AsPyObjectBorrow(value)));
+            PyStackRef_CLOSE_SPECIALIZED(value, _PyUnicode_ExactDealloc);
+        }
+
         tier2 op(_POP_TWO, (nos, tos --)) {
             PyStackRef_CLOSE(tos);
             PyStackRef_CLOSE(nos);

Misc/NEWS.d/next/Core_and_Builtins/2025-06-20-14-50-44.gh-issue-134584.3CJdAI.rst

@@ -345,7 +345,7 @@ remove_globals(_PyInterpreterFrame *frame, _PyUOpInstruction *buffer,
 #define sym_new_tuple _Py_uop_sym_new_tuple
 #define sym_tuple_getitem _Py_uop_sym_tuple_getitem
 #define sym_tuple_length _Py_uop_sym_tuple_length
-#define sym_is_immortal _Py_uop_sym_is_immortal
+#define sym_is_immortal _Py_uop_symbol_is_immortal
 #define sym_is_compact_int _Py_uop_sym_is_compact_int
 #define sym_new_compact_int _Py_uop_sym_new_compact_int
 #define sym_new_truthiness _Py_uop_sym_new_truthiness

Python/bytecodes.c

@@ -34,7 +34,7 @@ typedef struct _Py_UOpsAbstractFrame _Py_UOpsAbstractFrame;
 #define sym_new_tuple _Py_uop_sym_new_tuple
 #define sym_tuple_getitem _Py_uop_sym_tuple_getitem
 #define sym_tuple_length _Py_uop_sym_tuple_length
-#define sym_is_immortal _Py_uop_sym_is_immortal
+#define sym_is_immortal _Py_uop_symbol_is_immortal
 #define sym_new_compact_int _Py_uop_sym_new_compact_int
 #define sym_is_compact_int _Py_uop_sym_is_compact_int
 #define sym_new_truthiness _Py_uop_sym_new_truthiness
@@ -534,15 +534,15 @@ dummy_func(void) {
     }
 
     op(_LOAD_CONST_INLINE, (ptr/4 -- value)) {
-        value = PyJitRef_Borrow(sym_new_const(ctx, ptr));
+        value = sym_new_const(ctx, ptr);
     }
 
     op(_LOAD_CONST_INLINE_BORROW, (ptr/4 -- value)) {
         value = PyJitRef_Borrow(sym_new_const(ctx, ptr));
     }
 
     op(_POP_TOP_LOAD_CONST_INLINE, (ptr/4, pop -- value)) {
-        value = PyJitRef_Borrow(sym_new_const(ctx, ptr));
+        value = sym_new_const(ctx, ptr);
     }
 
     op(_POP_TOP_LOAD_CONST_INLINE_BORROW, (ptr/4, pop -- value)) {
@@ -561,6 +561,24 @@ dummy_func(void) {
         value = PyJitRef_Borrow(sym_new_const(ctx, ptr));
     }
 
+    op(_POP_TOP, (value -- )) {
+        PyTypeObject *typ = sym_get_type(value);
+        if (PyJitRef_IsBorrowed(value) ||
+            sym_is_immortal(PyJitRef_Unwrap(value)) ||
+            sym_is_null(value)) {
+            REPLACE_OP(this_instr, _POP_TOP_NOP, 0, 0);
+        }
+        else if (typ == &PyLong_Type) {
+            REPLACE_OP(this_instr, _POP_TOP_INT, 0, 0);
+        }
+        else if (typ == &PyFloat_Type) {
+            REPLACE_OP(this_instr, _POP_TOP_FLOAT, 0, 0);
+        }
+        else if (typ == &PyUnicode_Type) {
+            REPLACE_OP(this_instr, _POP_TOP_UNICODE, 0, 0);
+        }
+    }
+
     op(_COPY, (bottom, unused[oparg-1] -- bottom, unused[oparg-1], top)) {
         assert(oparg > 0);
         top = bottom;
@@ -803,7 +821,9 @@ dummy_func(void) {
     }
 
     op(_RETURN_VALUE, (retval -- res)) {
-        JitOptRef temp = retval;
+        // We wrap and unwrap the value to mimic PyStackRef_MakeHeapSafe
+        // in bytecodes.c
+        JitOptRef temp = PyJitRef_Wrap(PyJitRef_Unwrap(retval));
         DEAD(retval);
         SAVE_STACK();
         ctx->frame->stack_pointer = stack_pointer;