[racket-dev] expand, local-expand, and syntax-procedure-converted-arguments-property

At Wed, 10 Jul 2013 23:15:46 -0400, Ryan Culpepper wrote:
> On 07/10/2013 09:04 PM, Asumu Takikawa wrote:
> > Hi all,
> >
> > I'm currently trying to fix the Typed Racket unit tests. I think I've
> > narrowed down the issue to a certain syntax property for keyword
> > functions.
> >
> > The issue is illustrated by the following example:
> >
> >    #lang racket
> >
> >    (require racket/file
> >             (for-syntax racket/file
> >                         racket/keyword-transform))
> >
> >    ;; the property is #f
> >    (begin-for-syntax
> >     (displayln
> >       (syntax-case (expand-syntax #'(copy-directory/files 1 2)) ()
> >         [(let-values (((temp1) _)
> >                       ((temp2) _))
> >            (if _
> >                (#%plain-app1 copy-directory/files15 e1 ...)
> >                (#%plain-app2 copy-directory/files17 e2 ...)))
> >          (syntax-procedure-converted-arguments-property 
> #'copy-directory/files15)])))
> >
> >    ;; property is syntax
> >    (begin-for-syntax
> >     (displayln
> >       (syntax-case (local-expand #'(copy-directory/files 1 2) 'expression 
> null) ()
> >         [(let-values (((temp1) _)
> >                       ((temp2) _))
> >            (if _
> >                (#%plain-app1 copy-directory/files15 e1 ...)
> >                (#%plain-app2 copy-directory/files17 e2 ...)))
> >          (syntax-procedure-converted-arguments-property 
> #'copy-directory/files15)])))
> >
> > There are two syntax-time computations here. Both are expanding an
> > application of a keyword function (one with local-expand, one with
> > expand) and looking at the resulting syntax.
> >
> > The key point here is that I want to find the property looked up by
> > `syntax-procedure-converted-arguments-property` on an output identifier
> > because Typed Racket needs it to type-check the expansion.
> >
> > Unfortunately, as the comments indicate, only the second piece of code
> > can find the property. The reason appears to be that the property key is
> > actually a private `gensym`ed symbol and the two pieces of code appear
> > to get separate instantiations of the kw.rkt module (perhaps at different
> > phases).
> >
> > To check that, if I modify kw.rkt to use a plain symbol, both of the
> > snippets above return the same property value.
> >
> > Anyone have any idea how I can keep using `expand` but still be able to
> > look up the property?
> 
> To get information about a phase-0 '#%app' expansion, you need to call 
> the phase-1 version of 'syntax-procedure-converted-arguments-property'. 
> That's going to require a bit of phase-crossing trickery, because the 
> identifier you want to query is a phase-0 (dynamic) value, and you want 
> the result as a phase-0 value, but the phase-1 function naturally 
> consumes and produces phase-1 values.
> 
> One solution is to use 'quote-syntax', 'eval', and 'phase1-eval' all 
> together. Use 'eval' with 'quote-syntax' to convert the phase-0 
> identifier to a phase-1 identifier. Use 'phase1-eval' to run the 
> computation at phase 1 and capture the phase-1 result as a phase-0 value 
> (also using 'quote-syntax').

I haven't thought about this much, but would it make more sense to move
the property key to a cross-phase persistent module (with a `protected'
provide) in this case?