Another four years, and here I am again.
Perhaps it's simply bad luck to name a programming language after a spoken
human language. After all, it didn't seem to work too well for the Plain
English folks.
Well anyway. Lately I've been trying to put my languages and things online.
The rationale is straightfoward enough; Javascript and HTML5 are mature
enough to make it possible now, and people by their nature are lazy. If, to
try something out, you need to install this, that, and the other on your
system and configure them, chances are you'll be like, ehh. Don't feel like
it right now. But if it's just on a web page, you don't have that obstacle.
And my thoughts turned to my DOS stuff, because one of the more burdensome
things to ask the user to install is a system emulator.
I had come across jpc some time earlier, but I'd really prefer something in
Javascript, just because of the poor security track record Java has had
lately. (Guys, the whole idea was that you didn't have to trust a Java
applet, because it would be running in a virtual machine that didn't allow...
yeah, anyway) I looked a bit more and found jslinux, but it isn't open
source. So I looked a bit more and eventually found v86. It's not quite
as mature as the others, but for simple things it seems to work fine, and
as a nice bonus it's BSD-licensed, too.
So I started preparing a FreeDOS boot disk image with the Shelta distribution
on it. To my surprise, it was flakey. Sometimes the compilers would work,
and sometimes they'd hang. Sometimes the bootstrapped compiler seemed to
work better than the Nasm one, and sometimes it didn't. The problems
persisted after I re-assembled the source with the latest version of Nasm,
which did generate slightly different code, but only slightly (only two
instructions in sheltan.com were changed, and the code was the same size.)
This was all somewhat surprising. The thing is, it worked perfectly in DOSBox
(which is probably what I was developing it under in 2009.) And the same
flakiness was happening under both QEMU and v86, leading me to believe it was
a problem with FreeDOS. I tried FreeDOS version 1.1 as well as 1.0 -- no
difference in behaviour. I briefly looked for a web-enabled DOSBox, and it
turns out there are two: jdosbox, which is Java, and a gigantic (~4M) jarfile
besides, and jsdosbox, which is a GWT hack applied to jdosbox, which is
frankly really unappealing.
So I turned my attention back to FreeDOS, determined to get Shelta working
under it under v86.
After some experimentation, I figured it must be INT 21h/ah=07h that FreeDOS
was having problems with. It's the legacy DOS function for reading a
character from standard input. Shelta uses it because it's significantly
fewer instructions than the "normal" DOS input function.
Thing is, INT 21h/ah=07h doesn't have a way to indicate EOF. That's why
Shelta has traditionally insisted that Shelta source code end with a zero
byte (ASCII NUL character,) so that it can tell when the source has ended.
If the Shelta compiler wasn't consistently reading that character, that could
explain the hanging.
But after more experimentation, I couldn't tell what was wrong with
FreeDOS'es implementation of this function. I even looked at the FreeDOS
source code, but it wasn't too helpful; it seems to be implemented in terms
of the "normal" DOS read function.
It wasn't until I tried replacing the check for the NUL character with a
different character that it worked. Yes -- FreeDOS sometimes has trouble
reading NUL characters using this function. I'm not sure why. Since the
function is for reading from standard input, and in the absence of redirection
standard input means the keyboard, and on the keyboard you can type a NUL
(Ctrl+@), and since it works on both MS-DOS whence I originally developed
Shelta and DOSBox -- this is probably a bug in FreeDOS. It might be confused
about standard input not being open in "binary mode" or something. I'd submit
a bug report, but this is possibly the only piece of software on the planet
that is affected...
Anyway, solving the problem after realizing this was simple. Instead of
insisting Shelta programs end with a NUL, we now insist that they end with
everybody's favourite ASCII character, the VERTICAL TAB. And we apologize
to all those Shelta users who have huge code bases written in Shelta in which
vertical tabs are used to critical effect in some way I can't imagine. For
the rest of us, the SHELTA.BAT batchfile handles all this for us, so the change
is pretty much transparent.
The new check for vertical-tab-as-EOF does add a few bytes to the compilers,
but the Nasm version is still under 512 bytes.
And with those small changes, the Shelta language is now at version 1.2.
Happy bootstrapping -- in your browser!
Chris Pressey
November 30, 2013
Vancouver, BC