From: William D Clinger
Date: Thu, 14 Jun 2007 20:18:30 -0400
Subject: [Formal] raise semantics no longer makes sense

Submitter: William D Clinger
Issue type: Defect
Priority: Major
Component: I/O
Report version: 5.94
Summary: raise semantics no longer makes sense

Full description of issue:

The semantics of raise error handling, which is the
default, was designed for the i/o system of the 5.91
draft, and no longer makes sense now that arbitrary
mixing of binary and textual i/o is impossible and
the i/o system has been rewritten to allow transcoding
in bulk.

The problematic semantics is described in the second
paragraph of the specification of &i/o-decoding,
make-i/o-decoding-error, i/o-decoding-error?, and
i/o-decoding-error-transcoder:

    Exceptions of the type raised by the operations
    described in this section are continuable.  When
    such an exception is raised, the port's position
    is at the beginning of the invalid encoding.  If
    the exception handler returns, it should return
    a character or string representing the decoded
    text starting at the port's current position, and
    the exception handler must update the port's
    position to point past the error.

There are several problems with that paragraph.  The
"must" should be "should".  Second, the exception
handler cannot determine the bits involved in the
invalid encoding, since binary input is not allowed
on textual ports.  Hence little is gained by
requiring the position of the port to be "at the
beginning of the invalid encoding."  Furthermore it
is generally impossible for the exception handler
to "update the port's position to point past the
error", since operations such as get-char would
just re-raise the exception, and only a few textual
ports are required to support the set-port-position!
operation.

In summary, the 5.94 semantics can be implemented,
but it makes bulk transcoding more difficult, and
is not very useful in the wake of changes made to
the i/o system since the 5.91 draft.

Here are several possible repairs:

1.  For buffered i/o and bulk transcoding, it makes
sense to allow the exception to be raised when the
invalid encoding is encountered or generated, which
can come before (input) or after (output) the get-char
or put-char operation that a naive programmer might
think of as the cause of the exception.

2.  The exception handler for &i/o-decoding might
be given the invalid encoding as a bytevector that
is packaged up within the &i/o-decoding condition.

3.  The port might already have been updated to
point past the error when the exception handler
gains control.

4.  The exception might be made non-continuable.
This would greatly reduce the usefulness of the
raise error handling mode, however, so the raise
mode should no longer be the default if it is made
non-continuable.

I don't like any of those choices, especially the
fourth, but all of them would improve upon the 5.94
draft.  Note that the four choices above are
independent: any subset of them could be adopted.

The editors should keep in mind that whatever
semantics they adopt should make sense for fairly
arbitrary transcoders, not just for the three
standard codecs.


RESPONSE:

Suggestions #3 and #4 have been adopted.