Let's improve D's exceptions

[Leandro Motta Barros via Digitalmars-d]

I didn't give much attention to the details. but I like the spirit very
much.

LMB


On Wed, May 13, 2015 at 12:08 PM, Adam D. Ruppe via Digitalmars-d <
digitalmars-d at puremagic.com> wrote:

> Have you ever done:
>
> if(something) {
>    import std.conv;
>    throw new Exception("some error " ~ to!string(some_value));
> }
>
> Don't you hate it?
>
> * having to import std.conv to see data from your exception is a pain
> * it allocates eagerly and thus isn't suitable for a lot of places
> * inspecting the data can be a pain as the string is unstructured
>
> This assumes the data is even bothered to be added. Anyone who has gotten
> a RangeError in D knows important information is often just dropped!
>
> A good solution is to make a new exception subclass for each error type,
> storing details as data members. However, that's a bit of a pain in D
> because of all the work you have to do to make a complete subclass:
>
> * you need to forward the constructors
> * you need to reimplement the toString to print out data members
> * you still have a string member there that is encouraged to be used
>
>
> Finally, with the discussion of @nogc exceptions, it would be nice to
> simplify the constructor and get file/line info separated from it at the
> same time too. Can we tackle all these problems at once?
>
> I propose we do some library changes, one minor compiler change, and a
> culture shift toward better exceptions. Here's my proof of concept:
>
> http://arsdnet.net/dcode/exception.d
>
>
> Read that code and the comments to see what I'm going for and what it
> looks like. Here's the summary:
>
> * The Exception constructor is simplified: no required arguments,
> everything else is done with members. (The Throwable next = null argument
> might come back, but it can also be set after construction anyway so I'd
> prefer not to worry about it in most places either.)
>
> * The string is de-emphasized. It is still there so we don't break all the
> code that uses it now, but it is no longer the main thing and we really
> would prefer that it isn't used at all. Exception messages are instead done
> with the class name and data members. The string may be gotten by the class
> though in cases like converting errno to a message.
>
> * File and line is set at the throw point instead of the construction
> point.
>
> * A mixin template uses reflection to print data to the user instead of
> constructing a string.
>
> * This is allocation-free, except for the exception object itself. User
> code would NOT have to import std.conv, it just stores the data in the
> object.
>
>
>
> Changes:
>
> TO THE COMPILER:
>   * make the call to d_throw add file and line number information. This is
> completely backward compatible and allows the runtime to store it in the
> object if it wants to. Then my "fly" method becomes unnecessary.
>
> TO THE LIBRARY:
>   * Throwable's toString implementation changes a bit to call a couple
> more virtual functions to print details. Very similar to the current code,
> as you can see in my file there, just some overridable hooks.
>
>   * A PrintMembers and perhaps other helper mixins are added to the
> library. Doesn't matter where they go, druntime might do a simpler one for
> its own needs and then Phobos offers a full featured one for user code. Or
> whatever. It'd need to be a bit more complex than my proof of concept to
> cover more data (at least let it try calling toString on types too) but
> this is already pretty useful as is.
>
>
> TO USER CODE:
>   * We start using subclasses with data members instead of string
> arguments, migrating to it incrementally and encouraging people to write
> exceptions this way going forward.
>
>   * Note that this should be completely backward compatible, it won't
> break old code, just encourage a new better way.
>
>
>
> What do you all think? I encourage you to grab my example there and play
> with it a bit to see how you like it too and see if you can think of any
> improvements.
>
> D's exceptions kinda suck right now but they don't have to!
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.puremagic.com/pipermail/digitalmars-d/attachments/20150513/a9ef7623/attachment.html>