[Qemu-devel] Problem with discriminated unions with enum prefixes

[Qemu-devel] Problem with discriminated unions with enum prefixes

[Daniel P. Berrange]

In my LUKS encryption series, I have a discriminated union for
storing options for different encryption formats. See qapi/crypto.json
in this file:

  https://lists.gnu.org/archive/html/qemu-devel/2016-02/msg03187.html

You'll notice I have the 'prefix' line for the enum commented out. When
I uncomment this, I discovered that the discriminated union visitor does
not deal with prefixes.

To save having to wade through that big series, here is a quick
reproducer

diff --git a/qapi-schema.json b/qapi-schema.json
index 8d04897..a58648d 100644
--- a/qapi-schema.json
+++ b/qapi-schema.json
@@ -4083,3 +4083,22 @@
 ##
 { 'enum': 'ReplayMode',
   'data': [ 'none', 'record', 'play' ] }
+
+
+{ 'enum': 'QDemoType',
+  'prefix': 'QDEMO_TYPE',
+  'data': ['foo', 'bar' ] }
+
+{ 'struct': 'QDemoBase',
+  'data': { 'type': 'QDemoType' } }
+
+{ 'struct': 'QDemoFoo',
+  'data': { 'eek': 'int' } }
+
+{ 'struct': 'QDemoBar',
+  'data': { 'wizz': 'str' } }
+
+{ 'union': 'QDemo',
+  'base': 'QDemoBase',
+  'discriminator': 'type',
+  'data': { 'foo': 'QDemoFoo', 'bar': 'QDemoBar' } }


Apply that and then try to build and it'll fail with:

qapi-visit.c: In function ‘visit_type_QDemo’:
qapi-visit.c:7596:10: error: ‘Q_DEMO_TYPE_FOO’ undeclared (first use in this function)
     case Q_DEMO_TYPE_FOO:
          ^
qapi-visit.c:7596:10: note: each undeclared identifier is reported only once for each function it appears in
qapi-visit.c:7599:10: error: ‘Q_DEMO_TYPE_BAR’ undeclared (first use in this function)
     case Q_DEMO_TYPE_BAR:
          ^

The issue is that we used the 'QDEMO_TYPE' custom prefix for generating the
enum, but we didn't use the prefix in the union visitor.

I know we had had previous discussions with Markus strongly wanting to kill
off the support for enum prefixes. So before I waste time trying to fix
this union visitor code to handle prefixes, I figure we should decide if
we actually want to fix it, or go with Markus' plan to kill custom prefixes
on enums.

Per previous discussions, I think the ability to have custom prefixes is
quite desirable, to get more natural enum constant names. At the end of
the day though, the default enum naming is far from the worst bit of
QEMU, so I'm not ultimately too bothered either way. We either make
custom enum prefixes work everything they need to, or remove them.

Regards,
Daniel
-- 
|: http://berrange.com      -o-    http://www.flickr.com/photos/dberrange/ :|
|: http://libvirt.org              -o-             http://virt-manager.org :|
|: http://autobuild.org       -o-         http://search.cpan.org/~danberr/ :|
|: http://entangle-photo.org       -o-       http://live.gnome.org/gtk-vnc :|

Re: [Qemu-devel] Problem with discriminated unions with enum prefixes

[Eric Blake]

[-- Attachment #1: Type: text/plain, Size: 2359 bytes --]

On 02/16/2016 10:35 AM, Daniel P. Berrange wrote:
> In my LUKS encryption series, I have a discriminated union for
> storing options for different encryption formats. See qapi/crypto.json
> in this file:
> 
>   https://lists.gnu.org/archive/html/qemu-devel/2016-02/msg03187.html
> 
> You'll notice I have the 'prefix' line for the enum commented out. When
> I uncomment this, I discovered that the discriminated union visitor does
> not deal with prefixes.
> 

> Apply that and then try to build and it'll fail with:
> 
> qapi-visit.c: In function ‘visit_type_QDemo’:
> qapi-visit.c:7596:10: error: ‘Q_DEMO_TYPE_FOO’ undeclared (first use in this function)
>      case Q_DEMO_TYPE_FOO:
>           ^
> qapi-visit.c:7596:10: note: each undeclared identifier is reported only once for each function it appears in
> qapi-visit.c:7599:10: error: ‘Q_DEMO_TYPE_BAR’ undeclared (first use in this function)
>      case Q_DEMO_TYPE_BAR:
>           ^
> 
> The issue is that we used the 'QDEMO_TYPE' custom prefix for generating the
> enum, but we didn't use the prefix in the union visitor.

Should be a quick fix, if we want to keep prefixes.  I'll go ahead and
post it, at least for discussion purposes.

> 
> I know we had had previous discussions with Markus strongly wanting to kill
> off the support for enum prefixes. So before I waste time trying to fix
> this union visitor code to handle prefixes, I figure we should decide if
> we actually want to fix it, or go with Markus' plan to kill custom prefixes
> on enums.

I'm still on the fence which way to go; we've definitely improved the
code base so that inadvertent collisions due to odd heuristics are less
likely to occur, but every special case we have to carry (custom prefix
being one of them) results in more code to maintain and test.

> 
> Per previous discussions, I think the ability to have custom prefixes is
> quite desirable, to get more natural enum constant names. At the end of
> the day though, the default enum naming is far from the worst bit of
> QEMU, so I'm not ultimately too bothered either way. We either make
> custom enum prefixes work everything they need to, or remove them.
> 
> Regards,
> Daniel
> 

-- 
Eric Blake   eblake redhat com    +1-919-301-3266
Libvirt virtualization library http://libvirt.org


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 604 bytes --]

Re: [Qemu-devel] Problem with discriminated unions with enum prefixes

[Markus Armbruster]

Eric Blake <eblake@redhat.com> writes:

> On 02/16/2016 10:35 AM, Daniel P. Berrange wrote:
>> In my LUKS encryption series, I have a discriminated union for
>> storing options for different encryption formats. See qapi/crypto.json
>> in this file:
>> 
>>   https://lists.gnu.org/archive/html/qemu-devel/2016-02/msg03187.html
>> 
>> You'll notice I have the 'prefix' line for the enum commented out. When
>> I uncomment this, I discovered that the discriminated union visitor does
>> not deal with prefixes.
>> 
>
>> Apply that and then try to build and it'll fail with:
>> 
>> qapi-visit.c: In function ‘visit_type_QDemo’:
>> qapi-visit.c:7596:10: error: ‘Q_DEMO_TYPE_FOO’ undeclared (first use in this function)
>>      case Q_DEMO_TYPE_FOO:
>>           ^
>> qapi-visit.c:7596:10: note: each undeclared identifier is reported only once for each function it appears in
>> qapi-visit.c:7599:10: error: ‘Q_DEMO_TYPE_BAR’ undeclared (first use in this function)
>>      case Q_DEMO_TYPE_BAR:
>>           ^
>> 
>> The issue is that we used the 'QDEMO_TYPE' custom prefix for generating the
>> enum, but we didn't use the prefix in the union visitor.
>
> Should be a quick fix, if we want to keep prefixes.  I'll go ahead and
> post it, at least for discussion purposes.
>
>> 
>> I know we had had previous discussions with Markus strongly wanting to kill
>> off the support for enum prefixes. So before I waste time trying to fix
>> this union visitor code to handle prefixes, I figure we should decide if
>> we actually want to fix it, or go with Markus' plan to kill custom prefixes
>> on enums.
>
> I'm still on the fence which way to go; we've definitely improved the
> code base so that inadvertent collisions due to odd heuristics are less
> likely to occur, but every special case we have to carry (custom prefix
> being one of them) results in more code to maintain and test.

Yes.

As so many things, QAPI started out relatively simple & stupid, if
saddled with a certain amount of accidental complexity due to hasty
design and implementation.  Then features got bolted on left and right.
That's normal; useful software expands to fill available space.

We've been working hard on reducing the accidental complexity.  I
believe we should also try to control the complexity due to features.
Features need to earn their keep.

>> Per previous discussions, I think the ability to have custom prefixes is
>> quite desirable, to get more natural enum constant names. At the end of
>> the day though, the default enum naming is far from the worst bit of
>> QEMU, so I'm not ultimately too bothered either way. We either make
>> custom enum prefixes work everything they need to, or remove them.

Keeping an unloved feature half-broken while we deliberate whether to
keep it is the worst of all worlds.  Since the fix is a one-liner (plus
tests), let's just apply it to unblock Dan as soon as possible.