When `define' binds a keyword procedure at the top level or module
level, then it actually binds syntax that checks (i.e., warns
if arith or keywords are bad) and optimizes (i.e., avoids checking
keywords or allocating keyword-argument lists) first-order uses of
the procedure.
Checking and conversion are currently only performed for top-
and module-level bindings, because some macros detect the difference
between a variable or syntax definition --- notably `class' and
`unit'. A related problem in `class' is that the expansion
produces multiple definitions (for direct access to the fast
version of the function), which `class' turns into fields instead
of methods.
For example, if you make a multi-column table with a
`racketblock' in each column, then the columns size
to fit the code --- instead of forcing the table width
to match the page width and forcing each column to take
an equal share width.
For HTML, the style used in the output of `racketblock' now
disables line wrapping, and the Rkt text styles inherit
line-wrap behavior. This doesn't solve the general problem
of code overflowing the horizontal space, but it makes the
failure mode usefully better.
A new 'vertical-inset nested-flow style is used by
`defproc', `defform', etc. It has no effect for HTML,
but it introduces suitable vertical space for Latex output
(in case you want to use `defproc' in a SIGPLAN-format
paper, for example).
This case doesn't appear necessary, since LWs are constructed in an
expansion step that occurs after all of the meta-function names
(including the current one) are bound.
This is triggered by an environment variable that points at a file that
is used to cache polling results. If the variable is not specified just
use all mirror links as usual. This way, for random builds when people
try things out there is no problem, while the nightly builds (and my
manual builds when there are changes) do the right thing.
The file holds information about the polls, and verifies that the
download exists -- if not, the link is not shown. If it is, then it
tries to get the size of the remote file (via HEAD or via FTP), and if
it doesn't match our download information, then drop it too. If no size
informaion is available, include it, but re-poll after a few days. Even
if the size matches, a re-poll will be done after a month, so stale
mirrors will not stick.
This takes the advice from
http://files.stairways.com/other/ftp-list-specs-info.txt
further: search for the date by an explicit occurrence of a known month
name. This means that we won't see files with bad names (they'd be
filtered out of the result), but the filtered out entries are ones that
would not be usable with `ftp-make-file-seconds'.
When the month is found, and the entry is a file, look for a number
preceding the month, and if found, return it as the file size string.
This is a minor change in the API. (But it's probably better to either
revise it further, or eventually make it irrelevant by exposing the
interesting functionality via `net/url'.)
This function was using 2002 when there's no year... Updated it to use
the last occurrence of the guessed date, as described at
http://files.stairways.com/other/ftp-list-specs-info.txt
This function still looks pretty bogus -- the RFC does *not* say
anything about the format of response to `LIST', so it's whatever
semi-random thing the server does. (The above link looks like an
attempt to fix it, but I didn't see anything more official than that.)
From some looking around, it looks like ftp clients just try a bunch of
patterns against the text. Add also warnings in the documentation about
this.
RFC 2640 specifies that all transport strings are assumed UTF-8 if they
can be parsed as such. To make this complete it should really send a
FEAT and check that the result has "UTF8", but it's kind of redundant
since there's not much else to do if it isn't.
Also, switch to a more correct line endings of CRLF when sending
messages to the server.
