thrift-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Jens Geyer (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (THRIFT-4484) C++ codegen invalid for optional self-membership
Date Wed, 31 Jan 2018 08:09:00 GMT

    [ https://issues.apache.org/jira/browse/THRIFT-4484?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16346405#comment-16346405
] 

Jens Geyer commented on THRIFT-4484:
------------------------------------

The more interesting question is whether it makes sense to do that. 

Let's say you have an instance A that points to itself and then gets serialized. 
Now you de-serialize the nbits and get what ... an A that contains another copy of A = *two
instances*. 

And that is a well-known limitation of this feature.

I made another proposal a while ago that essentially introduces a new {{graph<T>}}
container type, THRIFT-2005. That can still be implemented, if anyone needs it. It does not
interfere with the other feature, as it buildson top of {{list<T>}}  but it very likely
will need some overhaul to compile.

 

> C++ codegen invalid for optional self-membership
> ------------------------------------------------
>
>                 Key: THRIFT-4484
>                 URL: https://issues.apache.org/jira/browse/THRIFT-4484
>             Project: Thrift
>          Issue Type: Bug
>          Components: C++ - Compiler
>    Affects Versions: 0.11.0
>         Environment: Thrift 0.10.0 tested, but I don't see a change in 0.11.0. Fedora
25. gcc 6.4.1. x86_64.
>            Reporter: Craig Ringer
>            Priority: Minor
>
> Support was added for self-referential objects in in [https://github.com/apache/thrift/pull/84] "Tree/Recursive
struct support in thrift".
>  
> The tests cover objects that are co-recursive, objects that have lists of themselves,
etc. But there's no test for optional self-containment e.g.
> {code}
> struct RecSelf {
>    1: i16 item
>    2: optional RecSelf 
>  }
> {code}
> This works fine for languages like Java and Go. But in C++ it generates nonsensical code
that cannot compile because it contains a by-value member of its self and a separate {{isset}}
member.
> For example, from opentracing jaeger's IDL:
> {code}
> struct Downstream {
>     1: required string serviceName
>     2: required string serverRole
>     3: required string host
>     4: required string port
>     5: required Transport transport
>     6: optional Downstream downstream
> }
> {code}
> code-generation produces
> {code}
> class Downstream {
>  public:
>  
>   /* blah elided blah */
>   virtual ~Downstream() throw();
>   std::string serviceName;
>   std::string serverRole;
>   std::string host;
>   std::string port;
>   Transport::type transport;
>   Downstream downstream;
>   _Downstream__isset __isset;
>   /* blah elided blah */
> };
> {code}
> Compilation fails with
> {code}
> tracetest_types.h:64:14: error: field ‘downstream’ has incomplete type ‘jaegertracing::crossdock::thrift::Downstream’
>    Downstream downstream;
>               ^~~~~~~~~~
> tracetest_types.h:47:7: note: definition of ‘class jaegertracing::crossdock::thrift::Downstream’
is not complete until the closing brace
>  class Downstream {
>        ^~~~~~~~~~
> {code}
> I'd argue that the {{__isset}} model is not ideal, and a {{std::expected}}-like "optional"
or "maybe" construct would be a lot better. But presumably there are historical reasons for
that.
> The simplest correct solution would be to generate
> {code}
> class Downstream {
>   /* ... */
>   std::shared_ptr<Downstream> downstream;
>   /* ... */
> };
> {code}
> instead.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Mime
View raw message