more define-inlinable-related breakage

[Andy Wingo <wingo@pobox.com>]

I was doing some bisecting.  I started at v2.0.0, did a full clean and
build then went to df1297956211b7353155c9b54d7e9c22d05ce493 and built
without a "clean".  However I got an error:

      GUILEC web/request.go
    ;;; note: source file /home/wingo/src/guile/module/web/uri.scm
    ;;;       newer than compiled /home/wingo/src/guile/module/web/uri.go
    [...]
     986: 1 [declare-uri-header! "Content-Location"]
    In module/ice-9/boot-9.scm:
     119: 0 [#<procedure f79050 at module/ice-9/boot-9.scm:110:6 (thrown-k . args)> unbound-variable ...]

    module/ice-9/boot-9.scm:118:20: In procedure #<procedure f79050 at module/ice-9/boot-9.scm:110:6 (thrown-k . args)>:
    module/ice-9/boot-9.scm:118:20: In procedure module-lookup: Unbound variable: %uri?-procedure

This indicates that (web request) uses (web http) which calls uri? on
something.  (web http) is up-to-date, and so the .go file is used.  When
the .go was compiled, `uri?' expanded to reference `%uri?-procedure',
but 531c9f1dc51c4801c4d031ee80a31f15285a6b85 changed the names of the
private procedures to have a space instead of a dash, so the new uri.scm
is incompatible with the old web/http.go.

Note that this situation would be even worse if we generated a unique
name for the uri? procedure when uri.scm was compiled, as has been
discussed in http://thread.gmane.org/gmane.lisp.guile.devel/11722.

What should we do here?  Here are a few options:

  1) Require people to recompile all the time.  Sucks.

  2) Implement some sort of proper dependency management.  Tricky,
     because installing a new version of package A could force a
     recompile of all dependent packages; tough to get right on a
     package-management level.

  3) Declare that private names referenced by macro expansions are
     actually API that should not break.

Regards,

Andy
-- 
http://wingolog.org/