Night Owl 25/nopv25.iso/040A/INTERCAL.ZIP/BUGS

Text File | 1996-06-22 | 3.7 KB | 78 lines

            BUGS

There are still some known bugs in INTERCAL (and no doubt many we don't even
suspect as yet).  The following list accumulates current bug reports and
TO DO agenda items through the current release.

ESR = Eric S. Raymond
SS  = Steve Swales
LLH = Louis Howell
BR = Brian Raiter

1) (ESR) INTERCAL would be intrinsically a crock even if it worked right.

2) (SS) The system library syslib.i does not appear to work 100% as advertised.
   (ESR) Louis has corrected several bugs.  The gods alone know what others
   are lurking in there...

3) (LLH) The lexical analyzer is not completely independent of newlines.
   There is a conflict between the fact that newlines should be
   ignored almost everywhere, and the facts that the error mechanism
   wants to know about line numbers and the splat mechanism wants
   statements from the original program.

   The way these issues are currently resolved by the compiler is
   a bit over-complicated, but seems to work properly in almost
   every respect.  I see only two problems:  First, splat operation
   prints only the text line on which a typo was detected, not the
   whole statement.  Second, for debugging purposes the text lines
   are printed as comments in the degenerated C code.  These
   comments are only properly aligned with the degenerated code
   if the program has exactly one statement per line.  (See
   pit/tests/testsplat.i for examples.)

   Neither of these problems is a big deal, which is why I haven't
   worried about them too much.  The best solution I can see is
   to make "phase 1" of the lexer into a preprocessor able to
   correctly separate statements.  It could then either add or
   remove newlines to put the program into one statement per line
   format, or put markers between statements for the main lexer
   to detect.  Either way, it would be a good bit of work for a
   relatively small gain, and anyone trying it should make sure
   the result isn't buggier than the current version.

4) (BR) I have hacked up a fix to the problem with line numbers being off
   due to statements sharing/spanning lines. The parser now stores the
   starting line number for all tuples, not just the splatted
   ones. Not only does this give synchronized lines in the comments in
   the degenerated C code, but run-time error messages are just as
   reliable as the parse/compile-time ones. The line number in the
   error message is actually the line number of the next statement
   (e.g., if the next statement is on the same line, the line number
   of both statements will be used).

   Note that I said "just as reliable," as opposed to "fully
   reliable". Since the lexer doesn't know where the line ends until
   after it's already parsed the preamble (DO-PLEASE combo, line
   label, oh-oh-seven number, etc) of the following line, it actually
   stores as the starting line number the first non-whitespace of the
   statement proper. So if a statement has its preamble on one line
   and the rest of it on the next, the preamble will be beheaded.

   Still, it's a vast improvement, and as LLH points out, it's not
   going to get much better without writing a preprocessing lexer that
   can actually separate out the statements before any real parsing
   begins. Until someone wants to do that, this will do nicely, I think.

   As a side bonus, splatting now shows the entire statement. (Er,
   except when the statement both spans and shares a line. If a multi-
   line statement ends on the same line as the beginning of the next
   statement, the last part will get truncated. Again, not a case we
   want to be encouraging people to explore anyway.)

            TO DO

1. (ESR) Add more optimization templates, esp. idioms for +, -, *, /.

2. (ESR) Forget this @!%$#! crock and take a long vacation.