[mypyc] Add more comments about overlapping error values

mypyc/ir/class_ir.py

@@ -180,7 +180,12 @@ def __init__(
         self.attrs_with_defaults: set[str] = set()
 
         # Attributes that are always initialized in __init__ or class body
-        # (inferred in mypyc.analysis.attrdefined using interprocedural analysis)
+        # (inferred in mypyc.analysis.attrdefined using interprocedural analysis).
+        # These can never raise AttributeError when accessed. If an attribute
+        # is *not* always initialized, we normally use the error value for
+        # an undefined value. If the attribute byte has an overlapping error value
+        # (the error_overlap attribute is true for the RType), we use a bitmap
+        # to track if the attribute is defined instead (see bitmap_attrs).
         self._always_initialized_attrs: set[str] = set()
 
         # Attributes that are sometimes initialized in __init__
@@ -191,8 +196,8 @@ def __init__(
 
         # Definedness of these attributes is backed by a bitmap. Index in the list
         # indicates the bit number. Includes inherited attributes. We need the
-        # bitmap for types such as native ints that can't have a dedicated error
-        # value that doesn't overlap a valid value. The bitmap is used if the
+        # bitmap for types such as native ints (i64 etc.) that can't have a dedicated
+        # error value that doesn't overlap a valid value. The bitmap is used if the
         # value of an attribute is the same as the error value.
         self.bitmap_attrs: list[str] = []
 

mypyc/ir/rtypes.py

@@ -58,8 +58,21 @@ class RType:
     # to checking for error value as the return value of a function.
     #
     # For example, no i64 value can be reserved for error value, so we
-    # pick an arbitrary value (e.g. -113) to signal error, but this is
-    # also a valid non-error value.
+    # pick an arbitrary value (-113) to signal error, but this is
+    # also a valid non-error value. The chosen value is rare as a
+    # normal, non-error value, so most of the time we can avoid calling
+    # PyErr_Occurred() when checking for errors raised by called
+    # functions.
+    #
+    # This also means that if an attribute with this type might be
+    # undefined, we can't just rely on the error value to signal this.
+    # Instead, we add a bitfield to keep track whether attributes with
+    # "error overlap" have a value. If there is no value, AttributeError
+    # is raised on attribute read. Parameters with default values also
+    # use the bitfield trick to indicate whether the caller passed a
+    # value. (If we can determine that an attribute is "always defined",
+    # we never raise an AttributeError and don't need the bitfield
+    # entry.)
     error_overlap = False
 
     @abstractmethod

mypyc/test-data/run-async.test

@@ -457,6 +457,19 @@ def run(x: object) -> object: ...
 async def sleep(t: float) -> None: ...
 
 [case testRunAsyncMiscTypesInEnvironment]
+# Here we test that values of various kinds of types can be spilled to the
+# environment. In particular, types with "overlapping error values" such as
+# i64 can be tricky, since they require extra work to support undefined
+# attribute values (which raise AttributeError when accessed). For these,
+# the object struct has a bitfield which keeps track of whether certain
+# attributes have an assigned value.
+#
+# In practice we mark these attributes as "always defined", which causes these
+# checks to be skipped on attribute access, and thus we don't require the
+# bitfield to exist.
+#
+# See the comment of RType.error_overlap for more information.
+
 import asyncio
 
 from mypy_extensions import i64, i32, i16, u8