[Sbcl-devel] Proposal for typed read syntax (was Re: Print-readably bug)

[Kevin R. <kp...@ma...>]

On Aug 4, 2009, at 14:07, Tobias C. Rittweiler wrote:

> Nikodemus Siivola <nik...@ra...> writes:
>
>> While ~ is explicitly reserved, I somehow feel that using sharpmacros
>> is preferable, as they tend to stand out better from source text. #"
>> seems sanish to me -- as long as the code that does the printing
>> checks the readtable before using it.
>
> How about #T? T for Type. This generalizes nicely. So for example
>
>     #T(simple-base-string "foobar"),
>
> or
>
>     #T(simple-array fixnum (3) #(1 2 3))
>
> Or even...
>
>     #THE(simple-base-string "foobar")

This structure, especially in the second case, makes it difficult to  
implement because you can't just READ-DELIMITED-LIST (unless you want  
to read once and then reconstruct as the specified type). Also, giving  
a special interpretation to the last element of the list is *somewhat*  
unidiomatic.

I propose:

   #T(simple-base-string "foobar")
   #T((simple-array fixnum (3)) #(1 2 3))

Since the type is not flattened in the list, the occurrence of the  
value can be predicted without lookahead, and the read procedure is  
straightforward:

(defun read-sharp-t (...)
   Expect #\(
   Read type specifier
   Read value with appropriate context
   Expect #\)
   )

Two concrete definitions, both designed to be usefully user-extensible:

1. #T binds SB-EXT:*READ-COERCE* to the type specifier. The  
implementations of ", #(, and #A would check *READ-COERCE* to  
determine what type of array to allocate.

2. #T invokes a generic function (sb-ext:read-typed stream type- 
specifier-name type-specifier-args).

Option 1 is simpler, in that the dispatch is handled by whatever read- 
macro occurs inside the #T. (To avoid silent maybe-misbehavior, it  
would be good for the #T implementation to check that the value is of  
the given type.)

Option 2 is more powerful, in that the contained read syntax can be  
completely arbitrary.

Both of these could be implemented as compatibility libraries in other  
Lisp implementations, though option 1 would require overriding more  
syntax, or special peeking cases in the definition of #T.

Option 2 has the useful property that it can be used as a hook for  
user-defined read  syntax identified by symbol (benefiting from the  
package system) rather than by readtable and character; new kinds of  
readable object could be introduced without requiring the user to  
choose readtables. (Think of it as #S extended beyond structures.)

-- 
Kevin Reid                                  <http://switchb.org/kpreid/>