[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [PATCH v4 11/14] qapi/introspect.py: add type hint annotations

From: John Snow
Subject: Re: [PATCH v4 11/14] qapi/introspect.py: add type hint annotations
Date: Mon, 8 Feb 2021 16:39:12 -0500
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:78.0) Gecko/20100101 Thunderbird/78.6.0

On 2/5/21 8:42 AM, Markus Armbruster wrote:
John Snow <jsnow@redhat.com> writes:

On 2/3/21 10:15 AM, Markus Armbruster wrote:
John Snow <jsnow@redhat.com> writes:

Signed-off-by: John Snow <jsnow@redhat.com>
   scripts/qapi/introspect.py | 115 ++++++++++++++++++++++++++-----------
   scripts/qapi/mypy.ini      |   5 --
   scripts/qapi/schema.py     |   2 +-
   3 files changed, 82 insertions(+), 40 deletions(-)

diff --git a/scripts/qapi/introspect.py b/scripts/qapi/introspect.py
index 60ec326d2c7..b7f2a6cf260 100644
--- a/scripts/qapi/introspect.py
+++ b/scripts/qapi/introspect.py
@@ -30,10 +30,19 @@
   from .gen import QAPISchemaMonolithicCVisitor
   from .schema import (
+    QAPISchema,
+    QAPISchemaEntity,
+    QAPISchemaEnumMember,
+    QAPISchemaFeature,
+    QAPISchemaObjectType,
+    QAPISchemaObjectTypeMember,
+    QAPISchemaVariant,
+    QAPISchemaVariants,
+from .source import QAPISourceInfo
# This module constructs a tree data structure that is used to
@@ -57,6 +66,8 @@
     # generate the introspection information for QEMU. It behaves similarly
     # to a JSON value.
     # A complexity over JSON is that our values may or may not be annotated.
     # Un-annotated values may be:
     #     Scalar: str, bool, None.
     #     Non-scalar: List, Dict
     # _value = Union[str, bool, None, Dict[str, TreeValue], List[TreeValue]]
     # With optional annotations, the type of all values is:
     # TreeValue = Union[_value, Annotated[_value]]
     # Sadly, mypy does not support recursive types, so we must approximate 
     _stub = Any
     _scalar = Union[str, bool, None]
     _nonscalar = Union[Dict[str, _stub], List[_stub]]
   _value = Union[_scalar, _nonscalar]
   TreeValue = Union[_value, 'Annotated[_value]']

I'm once again terminally confused about when to use _lower_case and
when to use CamelCase for such variables.

That's my fault for not using them consistently.


TitleCase: Classes, Real Type Names :tm:
lowercase: instance names (and certain built-in types like str/bool/int)
UPPERCASE: "Constants". This is an extremely loose idea in Python.

I use the "_" prefix for any of the above categories to indicate something not intended to be used outside of the current scope. These types won't be accessible outside the module by default.

TypeVars I use "T", "U", "V", etc unless I bind them to another type; then I use e.g. NodeT instead.

When it comes to things like type aliases, I believe I instinctively used lowercase because I am not creating a new Real Type and wanted some visual distinction from a real class name. (aliases created in this way cannot be used with isinstance and hold no significance to mypy.)

That's why I used _stub, _scalar, _nonscalar, and _value for those types there. Then I disregarded my own convention and used TreeValue; perhaps that ought to be tree_value for consistency as it's not a Real Type :tm:

...but then we have the SchemaInfo type aliases, which I named using the same type name as they use in QAPI to help paint the association (and pick up 'git grep' searchers.)

Not fantastically consistent, sorry. Feel free to express a preference, I clearly don't have a universally applied one.

(Current leaning: rename TreeValue to tree_value, but leave everything else as it is.)

The reader has to connect _stub = Any back "we must approximate this".
Hmm... "we approximate with Any"?

I can try to be more explicit about it.

+# This is a (strict) alias for an arbitrary object non-scalar, as above:
+_DObject = Dict[str, object]

Sounds greek :)

Admittedly it is still not explained well ... until the next patch. I'm
going to leave it alone for now until you have a chance to respond to
these walls of text.

You explain it some futher down.

It's almost the Dict part of _nonscalar, but not quite: object vs. Any.

I naively expect something closer to

     _scalar = ...
     _object = Dict[str, _stub]
     _nonscalar = Union[_object, List[_stub]

and (still naively) expect _object to be good enough to serve as type
annotation for dicts representing JSON objects.

"_object" would be good, except ... I am trying to avoid using that word
because what does it mean? Python object? JSON object? Here at the
boundary between two worlds, nothing makes sense.

Naming is hard.

Yep. We can skip this debate by just naming the incoming types SchemaInfo and similar... (cont'd below)

We talked about these names in review of v2.  Let me try again.

introspect.py needs to generate (a suitable C representation of) an
instance of QAPI type '[SchemaInfo]'.

Its current choice of "suitable C representation" is "a QLitQObject
initializer with #if and comments".  This is a "lose" representation:
QLitQObject can encode pretty much anything, not just instances of

C code converts this QLitQObject to a SchemaInfoList object[*].
SchemaInfoList is the C type for QAPI type '[SchemaInfo]'.  Automated
tests ensure this conversion cannot fail, i.e. the "lose" QLitQObject
actually encodes a '[SchemaInfo]'.

introspect.py separates concerns: it first builds an abstract
representation of "set of QObject with #if and comments", then generates
C code from that.

Why "QObject with #if and comments", and not "QLitQObject with #if and
comments"?  Because QLitQObject is *one* way to represent QObject, and
we don't care which way outside C code generation.

A QObject represents a JSON value.  We could just as well say "JSON
value with #if and comments".

So, the abstract representation of "JSON value with #if and comments" is
what we're trying to type.  If you'd rather say "QObject with #if and
comments", that's fine.

Our abstract representation is a tree, where

* JSON null / QNull is represented as Python None

* JSON string / QString as str

* JSON true and false / QBool as bool

* JSON number / QNum is not implemented

* JSON object / QDict is dict mapping string keys to sub-trees

* JSON array / QList is list of sub-trees

* #if and comment tacked to a sub-tree is represented by wrapping the
   subtree in Annotated

   Wrapping a sub-tree that is already wrapped seems mostly useless, but
   the code doesn't care.

   Wrapping dictionary values makes no sense.  The code doesn't care, and
   gives you GIGO.

   Making the code reject these two feels out of scope.  If you want to
   anyway, I won't object unless it gets in the way of "in scope" stuff
   (right now it doesn't seem to).

Let me stress once again: this is *not* an abstract representation of a
'SchemaInfo'.  Such a representation would also work, and you might like
it better, but it's simply not what we have.  Evidence: _tree_to_qlit()
works fine for *any* tree, not just for trees that encode instances of

... as long as you don't feel that's incorrect to do. We are free to name those structures SchemaInfo but type _tree_to_qlit() in terms of generic Dict objects, leaving us without a middle-abstract thing to name at all.

Based on your review of the "dummy types" patch, I'm going to assume that's fine.

Since each (sub-)tree represents a JSON value / QObject, possibly with
annotations, I'd like to propose a few "obvious" (hahaha) names:

* an unannotated QObject: QObject

* an annotated QObject: AnnotatedQObject

* a possibly annotated QObject: PossiblyAnnotatedQObject

   Too long.  Rename QObject to BareQObject, then call this one QObject.

This gives us:

     _BareQObject = Union[None, str, bool, Dict[str, Any], List[Any]]
     _AnnotatedQObject = Annotated[_QObject]
     _QObject = Union[_BareQObject, _AnnotatedQObject]

Feel free to replace QObject by JsonValue in these names if you like
that better.  I think I'd slightly prefer JsonValue right now.

Now back to _DObject:

(See patch 12/14 for A More Betterer Understanding of what _DObject is
used for. It will contribute to A Greater Understanding.)

Anyway, to your questions;

(1) _DObject was my shorthand garbage way of saying "This is a Python
Dict that represents a JSON object". Hence Dict-Object, "DObject". I
have also derisively called this a "dictly-typed" object at times.

In the naming system I proposed, this is BareQDict, with an additional
complication: we actually have two different types for the same thing,
an anonymous one within _BareQObject, and a named one.

(2) Dict[str, Any] and Dict[str, object] are similar, but do have a

The former is the anonymous one, the latter the named one.

Kinda-sorta. I am talking about pure mypy here, and the differences between typing two things this way.

Though I think you're right: I used the "Any" form for the anonymous type (inherent to the structure of a JSON compound type) and the "object" form for the named forms (The SchemaInfo objects we build in the visitors to pass to the generator later).

semantic difference. I alluded to it by calling this a "(strict) alias";
which does not help you understand any of the following points:

Whenever you use "Any", it basically turns off type-checking at that
boundary; it is the gradually typed boundary type. Avoid it whenever
reasonably possible.

Example time:

def foo(thing: Any) -> None:
      print(thing.value)  # Sure, I guess! We'll believe you.

def foo(thing: object) -> None:
      print(thing.value)  # BZZT, Python object has no value prop.

Use "Any" when you really just cannot constrain the type, because you're
out of bourbon or you've decided to give in to the darkness inside your

Use "object" when the type of the value actually doesn't matter, because
you are only passing it on to something else later that will do its own
type analysis or introspection on the object.

For introspect.py, 'object' is actually a really good type when we can
use it, because we interrogate the type exhaustively upon receipt in

That leaves one question you would almost assuredly ask as a followup:

"Why didn't you use object for the stub type to begin with?"

Let's say we define _stub as `object` instead, the Python object. When
_tree_to_qlit recurses on non-scalar structures, the held value there is
only known as "object" and not as str/bool/None, which causes a typing
error at that point.

Moving the stub to "Any" tells mypy to ... not worry about what type we
actually passed here. I gave in to the darkness in my heart. It's just
too annoying without real recursion.

May I have an abridged version of this in the comments?  It might look
quaint in ten years, when we're all fluent in Python type annotations.
But right now, at least some readers aren't, and they can use a bit of

Yeah, I'm sympathetic to that.... though I'm not sure what to write or where. I can add some reference points in the commit message, like this one:


maybe in conjunction with the named type aliases patch this is actually sufficient?

[*] Actually, we take a shortcut and convert straight to QObject, but
that's just laziness.  See qmp_query_qmp_schema()'s "Minor hack:"


reply via email to

[Prev in Thread] Current Thread [Next in Thread]