Merge branch 'main' into warnings-as-error-2

Author: graingert

Misc/NEWS.d/next/Core_and_Builtins/2025-03-10-12-08-57.gh-issue-130887.f823Ih.rst

@@ -0,0 +1 @@
+Optimize the AArch64 code generation for the JIT. Patch by Diego Russo

Misc/NEWS.d/next/Core_and_Builtins/2025-03-25-13-58-25.gh-issue-131719.zKv98a.rst

@@ -0,0 +1 @@
+Fix missing NULL check in ``_PyMem_FreeDelayed`` in :term:`free-threaded <free threading>` build.

Modules/_ctypes/_ctypes.c

@@ -2148,18 +2148,7 @@ static PyObject *CreateSwappedType(ctypes_state *st, PyTypeObject *type,
     if (!swapped_args)
         return NULL;
 
-    if (st->swapped_suffix == NULL) {
-#ifdef WORDS_BIGENDIAN
-        st->swapped_suffix = PyUnicode_InternFromString("_le");
-#else
-        st->swapped_suffix = PyUnicode_InternFromString("_be");
-#endif
-    }
-    if (st->swapped_suffix == NULL) {
-        Py_DECREF(swapped_args);
-        return NULL;
-    }
-
+    assert(st->swapped_suffix != NULL);
     newname = PyUnicode_Concat(name, st->swapped_suffix);
     if (newname == NULL) {
         Py_DECREF(swapped_args);
@@ -5113,12 +5102,7 @@ PyCArrayType_from_ctype(ctypes_state *st, PyObject *itemtype, Py_ssize_t length)
     char name[256];
     PyObject *len;
 
-    if (st->array_cache == NULL) {
-        st->array_cache = PyDict_New();
-        if (st->array_cache == NULL) {
-            return NULL;
-        }
-    }
+    assert(st->array_cache != NULL);
     len = PyLong_FromSsize_t(length);
     if (len == NULL)
         return NULL;
@@ -6099,6 +6083,20 @@ _ctypes_mod_exec(PyObject *mod)
         return -1;
     }
 
+    st->array_cache = PyDict_New();
+    if (st->array_cache == NULL) {
+        return -1;
+    }
+
+#ifdef WORDS_BIGENDIAN
+        st->swapped_suffix = PyUnicode_InternFromString("_le");
+#else
+        st->swapped_suffix = PyUnicode_InternFromString("_be");
+#endif
+    if (st->swapped_suffix == NULL) {
+        return -1;
+    }
+
     if (_ctypes_add_types(mod) < 0) {
         return -1;
     }

Modules/_ctypes/malloc_closure.c

@@ -27,6 +27,16 @@
 
 /******************************************************************/
 
+
+#ifdef Py_GIL_DISABLED
+static PyMutex malloc_closure_lock;
+# define MALLOC_CLOSURE_LOCK()   PyMutex_Lock(&malloc_closure_lock)
+# define MALLOC_CLOSURE_UNLOCK() PyMutex_Unlock(&malloc_closure_lock)
+#else
+# define MALLOC_CLOSURE_LOCK()   ((void)0)
+# define MALLOC_CLOSURE_UNLOCK() ((void)0)
+#endif
+
 typedef union _tagITEM {
     ffi_closure closure;
     union _tagITEM *next;
@@ -110,9 +120,11 @@ void Py_ffi_closure_free(void *p)
     }
 #endif
 #endif
+    MALLOC_CLOSURE_LOCK();
     ITEM *item = (ITEM *)p;
     item->next = free_list;
     free_list = item;
+    MALLOC_CLOSURE_UNLOCK();
 }
 
 /* return one item from the free list, allocating more if needed */
@@ -131,11 +143,15 @@ void *Py_ffi_closure_alloc(size_t size, void** codeloc)
     }
 #endif
 #endif
+    MALLOC_CLOSURE_LOCK();
     ITEM *item;
-    if (!free_list)
+    if (!free_list) {
         more_core();
-    if (!free_list)
+    }
+    if (!free_list) {
+        MALLOC_CLOSURE_UNLOCK();
         return NULL;
+    }
     item = free_list;
     free_list = item->next;
 #ifdef _M_ARM
@@ -144,5 +160,6 @@ void *Py_ffi_closure_alloc(size_t size, void** codeloc)
 #else
     *codeloc = (void *)item;
 #endif
+    MALLOC_CLOSURE_UNLOCK();
     return (void *)item;
 }

Objects/obmalloc.c

@@ -1215,7 +1215,9 @@ void
 _PyMem_FreeDelayed(void *ptr)
 {
     assert(!((uintptr_t)ptr & 0x01));
-    free_delayed((uintptr_t)ptr);
+    if (ptr != NULL) {
+        free_delayed((uintptr_t)ptr);
+    }
 }
 
 #ifdef Py_GIL_DISABLED

Objects/typeobject.c

@@ -574,14 +574,16 @@ _PyType_GetMRO(PyTypeObject *self)
 static inline void
 set_tp_mro(PyTypeObject *self, PyObject *mro, int initial)
 {
-    assert(PyTuple_CheckExact(mro));
-    if (self->tp_flags & _Py_TPFLAGS_STATIC_BUILTIN) {
-        // XXX tp_mro can probably be statically allocated for each
-        // static builtin type.
-        assert(initial);
-        assert(self->tp_mro == NULL);
-        /* Other checks are done via set_tp_bases. */
-        _Py_SetImmortal(mro);
+    if (mro != NULL) {
+        assert(PyTuple_CheckExact(mro));
+        if (self->tp_flags & _Py_TPFLAGS_STATIC_BUILTIN) {
+            // XXX tp_mro can probably be statically allocated for each
+            // static builtin type.
+            assert(initial);
+            assert(self->tp_mro == NULL);
+            /* Other checks are done via set_tp_bases. */
+            _Py_SetImmortal(mro);
+        }
     }
     self->tp_mro = mro;
 }

Tools/jit/_stencils.py

@@ -209,7 +209,24 @@ def pad(self, alignment: int) -> None:
             self.disassembly.append(f"{offset:x}: {' '.join(['00'] * padding)}")
         self.body.extend([0] * padding)
 
-    def remove_jump(self, *, alignment: int = 1) -> None:
+    def add_nops(self, nop: bytes, alignment: int) -> None:
+        """Add NOPs until there is alignment. Fail if it is not possible."""
+        offset = len(self.body)
+        nop_size = len(nop)
+
+        # Calculate the gap to the next multiple of alignment.
+        gap = -offset % alignment
+        if gap:
+            if gap % nop_size == 0:
+                count = gap // nop_size
+                self.body.extend(nop * count)
+            else:
+                raise ValueError(
+                    f"Cannot add nops of size '{nop_size}' to a body with "
+                    f"offset '{offset}' to align with '{alignment}'"
+                )
+
+    def remove_jump(self) -> None:
         """Remove a zero-length continuation jump, if it exists."""
         hole = max(self.holes, key=lambda hole: hole.offset)
         match hole:
@@ -244,7 +261,7 @@ def remove_jump(self, *, alignment: int = 1) -> None:
                 jump = b"\x00\x00\x00\x14"
             case _:
                 return
-        if self.body[offset:] == jump and offset % alignment == 0:
+        if self.body[offset:] == jump:
             self.body = self.body[:offset]
             self.holes.remove(hole)
 
@@ -266,10 +283,7 @@ class StencilGroup:
     _trampolines: set[int] = dataclasses.field(default_factory=set, init=False)
 
     def process_relocations(
-        self,
-        known_symbols: dict[str, int],
-        *,
-        alignment: int = 1,
+        self, known_symbols: dict[str, int], *, alignment: int = 1, nop: bytes = b""
     ) -> None:
         """Fix up all GOT and internal relocations for this stencil group."""
         for hole in self.code.holes.copy():
@@ -289,8 +303,8 @@ def process_relocations(
                 self._trampolines.add(ordinal)
                 hole.addend = ordinal
                 hole.symbol = None
-        self.code.remove_jump(alignment=alignment)
-        self.code.pad(alignment)
+        self.code.remove_jump()
+        self.code.add_nops(nop=nop, alignment=alignment)
         self.data.pad(8)
         for stencil in [self.code, self.data]:
             for hole in stencil.holes:

Tools/jit/_targets.py

@@ -44,6 +44,15 @@ class _Target(typing.Generic[_S, _R]):
     verbose: bool = False
     known_symbols: dict[str, int] = dataclasses.field(default_factory=dict)
 
+    def _get_nop(self) -> bytes:
+        if re.fullmatch(r"aarch64-.*", self.triple):
+            nop = b"\x1f\x20\x03\xD5"
+        elif re.fullmatch(r"x86_64-.*|i686.*", self.triple):
+            nop = b"\x90"
+        else:
+            raise ValueError(f"NOP not defined for {self.triple}")
+        return nop
+
     def _compute_digest(self, out: pathlib.Path) -> str:
         hasher = hashlib.sha256()
         hasher.update(self.triple.encode())
@@ -172,7 +181,9 @@ async def _build_stencils(self) -> dict[str, _stencils.StencilGroup]:
         stencil_groups = {task.get_name(): task.result() for task in tasks}
         for stencil_group in stencil_groups.values():
             stencil_group.process_relocations(
-                known_symbols=self.known_symbols, alignment=self.alignment
+                known_symbols=self.known_symbols,
+                alignment=self.alignment,
+                nop=self._get_nop(),
             )
         return stencil_groups