We have two issues with our qapi union layout:
1) Even though the QMP wire format spells the tag 'type', the
C code spells it 'kind', requiring some hacks in the generator.
2) The C struct uses an anonymous union, which places all tag
values in the same namespace as all non-variant members. This
leads to spurious collisions if a tag value matches a non-variant
member's name.
This patch is the back end for a series that converts to a
saner qapi union layout. Now that all clients have been
converted to use 'type' and 'obj->u.value', we can drop the
temporary parallel support for 'kind' and 'obj->value'.
Given a simple union qapi type:
{ 'union':'Foo', 'data': { 'a':'int', 'b':'bool' } }
this is the overall effect, when compared to the state before
this series of patches:
| struct Foo {
|- FooKind kind;
|- union { /* union tag is @kind */
|+ FooKind type;
|+ union { /* union tag is @type */
| void *data;
| int64_t a;
| bool b;
|- };
|+ } u;
| };
The testsuite still contains some examples of artificial restrictions
(see flat-union-clash-type.json, for example) that are no longer
technically necessary, now that there is no longer a collision between
enum tag values and non-variant member names; but fixing this will be
done in later patches, in part because some further changes are required
to keep QAPISchema*.check() from asserting. Also, a later patch will
add a reservation for the member name 'u' to avoid a collision between a
user's non-variant names and our internal choice of C union name.
Note, however, that we do not rename the generated enum, which
is still 'FooKind'. A further patch could generate implicit
enums as 'FooType', but while the generator already reserved
the '*Kind' namespace (commit
4dc2e69), there are already QMP
constructs with '*Type' naming, which means changing our
reservation namespace would have lots of churn to C code to
deal with a forced name change.
Signed-off-by: Eric Blake <eblake@redhat.com>
Message-Id: <
1445898903-12082-23-git-send-email-eblake@redhat.com>
[Commit message tweaked]
Signed-off-by: Markus Armbruster <armbru@redhat.com>
if base:
ret += gen_struct_fields([], base)
else:
- # TODO As a hack, we emit both 'kind' and 'type'. Ultimately, we
- # want to use only 'type', but the conversion is large enough to
- # require staging over several commits.
- ret += mcgen('''
- union {
- %(c_type)s kind;
- %(c_type)s type;
- };
-''',
- c_type=c_name(variants.tag_member.type.name))
-
- # TODO As a hack, we emit the union twice, once as an anonymous union
- # and once as a named union. Ultimately, we want to use only the
- # named union version (as it avoids conflicts between tag values as
- # branch names competing with non-variant QMP names), but the conversion
- # is large enough to require staging over several commits.
- tmp = ''
+ ret += gen_struct_field(variants.tag_member.name,
+ variants.tag_member.type,
+ False)
+
# FIXME: What purpose does data serve, besides preventing a union that
# has a branch named 'data'? We use it in qapi-visit.py to decide
# whether to bypass the switch statement if visiting the discriminator
# should not be any data leaks even without a data pointer. Or, if
# 'data' is merely added to guarantee we don't have an empty union,
# shouldn't we enforce that at .json parse time?
- tmp += mcgen('''
+ ret += mcgen('''
union { /* union tag is @%(c_name)s */
void *data;
''',
for var in variants.variants:
# Ugly special case for simple union TODO get rid of it
typ = var.simple_union_type() or var.type
- tmp += mcgen('''
+ ret += mcgen('''
%(c_type)s %(c_name)s;
''',
c_type=typ.c_type(),
c_name=c_name(var.name))
- ret += tmp
- ret += ' ' + '\n '.join(tmp.split('\n'))
ret += mcgen('''
} u;
- };
};
''')