aboutsummaryrefslogtreecommitdiff
path: root/doc/notes.txt
blob: 3f2538859ba2938016c3dcaae197bb4be46538b2 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
Definitely need to rewrite huge swathes of code in asm.
Candidates are:

fujichat.c: main(), handle_keystroke(), and the protocol parsing
routines

rs232dev.c: the whole thing could be rewritten in asm for speed,
but the size isn't so bad.

uIP itself: have to investigate. Some of the code is pretty
hairy. Before rewriting in asm, should make it so we can
compile without TCP listen support, since we don't use it.


Need to re-do the R: handler stuff. At least one driver (the
plain Atari 850 one) actually crashes when loaded from either
the DOS menu or aexec. Apparently it *only* works if it's
loaded as autorun.sys. So, something like this:

- Initial autorun.sys on disk loads a program called makeauto.
[done]

- makeauto asks for the serial driver, then concatenates the
driver + menu.com into a new autorun.sys (renaming the old
one to autorun.old or something), then asks user to reboot.
[done]

- Any time fujiconf loads and can't find an R: handler, it tells the
user and runs makeauto for him (or anyway asks, with default Yes). The
"set R: driver" stuff in fujiconf needs to move to makeauto (and be
changed so it appends files instead copy).

- FujiChat itself still won't contain any R: detection or
anything, beyond the RS232 Not Initialized error message
[done]

- While I'm at it, fix the filename paths so they *all* begin with
the D: prefix.
[done]

- Possibly menu.com has a new menu item, "RS232 Setup", and warns
the user "You must run RS232 Setup before using FujiChat"

20081124 bkw: Add /peek, /poke, /fgcolor, /bgcolor commands.

Also, now that we're parsing numerics from the server, it'd be
nice to do the alternate-nick thing, for when the nick is in use.

20081124 bkw: Keep a time-of-day clock (don't use RTCLOK to store
it though). At connect, send /TIME to the server, parse the results.
The config file will need a timezone setting.

20081121 bkw: Feature idea: auto-away when attract mode kicks in,
with default away message (but don't announce to channel). Un-away
as soon as a key is hit. Possibly let the user set a timeout
instead of letting the OS default to 9 minutes.
[done, using default OS timeout only]

Ideas for logging:

Will anyone give a shit about logging to disk? If so, it wouldn't
be impossible... but we'd likely drop packets (or does uip_stop()
gracefully handle this, maybe with a callback?)

/log [on|off|flush|file filename|search <term>]

Keep the disk file open? Log stuff to memory, then just before
the buffer fills, write it out to disk.

Logging to memory, looks like we have a little over 5K for this.
When we get ready to log a message that would put us past the end
of the buffer, memmove() the buffer down by some amount (say 1/4
of its size).

Log viewing is a mode. While enabled, handle_keystroke() will do
the --more-- thing. Traffic indicator's still active, and we
still log stuff to memory, but *don't* print it. If the log
has to be shifted down, also shift down the currently-viewing
pointer. If the view pointer points too low to move, set it
to the bottom of the log buffer if it isn't already (user will
get confused, but what you gonna do?). Forward/backward searches
should work. Filtering by nick should work.

When entering logviewer mode, save the current end-of-log address.
Any time the log moves while viewing, also subtract from this
address. More stuff can be coming in & being logged above this
address... on exiting logviewer mode, anything aboved the saved
address should be printed to the screen immediately.

In this model of "scrollback" support, we actually don't scroll
per se, we're *still* using stdio to print stuff. It might be
that on entering logviewer mode, we should run through the log,
counting characters, so we'll know where each screens' worth
starts (for going back/up in the viewer). Either that, or we
could print the damn thing backwards?

At a price of 3 bytes per message, we can store timestamps. Not
sure if it's worth it. Actually maybe we can just store a single
byte in the buffer every N minutes, to give a rough guide to
how long ago the next bunch of messages were.

How bout this for a lightly compressed format: we're only storing
7-bit printable ascii, in 8-bit bytes, so...

1st byte:

0-31 - No message, just timestamp. Next message starts at
next byte in buffer. The value is the minutes-1 since the
last timestamp. Basically, every minute of wall-clock time,
a timestamp of 0 is added to the buffer if someone has said
something in the past minute. If not, there will already
be a timestamp at the end of buffer, so just increment it
(unless it's at the max value, in which case add a zero
timestamp after it). We can use these to e.g. print
"(42 minutes silence)" or such. I'd like to be able to
do "N minutes ago" type messages, but to do that I have to
run through the buffer backwards to calculate what they
should be (then forwards to actually read messages from it).

Anything else: 1st char of nick, or 0x80 meaning "same nick as last
message" (or 0x81 meaning "my nick"?). Some care has to be taken when
memmoving the buffer: blocks of messages from the same user need to be
treated as a unit (kept entirely, or disposed of entirely).  For this
to make sense, such blocks have to be kept short (say, less than the
move-amount).

Next bytes are the rest of the nick, with bit 7 set for the last
one. In case of a 1-char nick, 1st and only char has bit 7 set.

After last nick char comes the message.
For the first byte, if the top bit is set, the bottom 7 bits are:

0 - empty message
1-31 - stuff like join/quit/part (don't bother to store part,
quit, or kick messages? or store normally?)
Anything else - normal message character (the only one, if the
top bit is set).

Further message bytes are stored with the top bit of the last byte
being set.

So if someone joins, says hello, and parts (<c> means 'c' with high
bit set:

Use<r>\x01\x80Hello\x80\x02

(Assuming \x01 is the join code and \x02 is the part code)

This is 13 bytes. If we stored what a normal client stores in its
log, it might look like:

User has joined\n
<User> Hello\n
User has left\n

Which is 16+13+14 = 43 bytes. Most clients would actually store the
channel name with the joins/parts; I left them out to make the
comparison more fair (we don't penalize them for storing info that
we don't store, only for storing the same info less densely).


20081120 bkw: smallest 0.4 binary, with all FEAT_* turned off except
for FEAT_LOW_RAM_BUFFERS, the BSS ends at $920E. With COL80E relocated
to $9A00, that gives us 1K for stack. Not great, but might work. It's
still too big to use with E80 (by about 3K, counting the stack).

COL80E loads at $7A00-$7F80, but doesn't really seem to use any RAM
from $8000-A000 with BASIC disabled. If the code could be relocated so
it loads at $9A00, we'd have a fighting chance of squeezing fujichat to
fit. Currently fujichat loads $2E00-8D3B... meaning there'd be a little
over 3K for stack & data/bss variables. Need to examine cc65's map file,
see how high the data/bss stuff goes. Moving the input and output buffers
to low RAM is going to help a lot. It also might be possible to move
the stack to low RAM (it's supposed to sit right above the heap, but
fujichat isn't using malloc(), which unless I'm mistaken means no heap
usage... though some of the libraries it links with might be using it).

Try compiling uip without UDP checksums, with fewer TCP/UDP conns,
asm optimize anything that doesn't compile efficiently?

It looks like TCP listen support could be made a compile-time option
that could be disabled. IRC doesn't need to listen on ports except
for DCCs, and we probably can't ever make those work anyway. It
won't save much (guesstimate is 50-75 bytes)...

Each TCP conn is supposed to take 30 bytes. For now we only need one,
but we have 5, so wasting 120 bytes there.

[so far, changing max. tcp conns, listen ports, and udp conns to 1,
and disabling udp cksums, has saved 325 bytes. Worth it, but need to
make sure the app still works!]

[also, make some of FujiChat's features able to be removed at
compile-time. Removing them all saves 963 bytes, and makes the BSS
end at $96cb! If COL80E were relocated to $9A00, we'd have 800-odd
bytes for the heap/stack, which still ain't enough... Possibly
the stack could be shrunk by 256 bytes and relocated to $2700?]

E80 is probably a lost cause. It seems to take 12K, which seems like an
awful lot. Screen is hard-wired to $6E70, but it seems to not use any
RAM from $A000-BFFF with BASIC disabled. moving the whole thing up 8K
(assuming I can relocate the code) gives us a start address of $8E70,
or 3K lower than COL80E, or 130-odd bytes(!)  above the end of cc65's
code segment (ouch!)... however maybe I could leave it unrelocated, and
use a linker script to move all the data/bss/heap/stack/etc (everything
but the code, and maybe some of the code too) to the $A000-BFFF area
(8K seems like a good bit, the stack is 2K).

Fast E40 driver doesn't have to be a full E: replacement! Instead, we only
need to hook the "put 1 byte" routine, and do the following:

Check & see if byte-to-print is printable or an EOL. If it's neither,
jump to the OS's print-one-byte routine.

If the screen is about to scroll (if we're on bottom line and byte is
0x9b, or if we're on bottom line + last column and byte is printable),
we capture the top line (first 40 bytes) of screen RAM and store in
scrollback buffer.  We can probably detect when the line ends in >1
spaces, and not store those.

If a screen scroll is needed, do it ourselves, *fast*. This may turn
out to be too messy or bloated to deal with, but I think it might not
be so bad.

If the char is printable, print it, *fast*. Fast E40 doesn't have to
support margins other than 0 and 39, which should help. Also we can
make assumptions about where the screen RAM is if need be, and use
multiple ZP pointers into it (using the FP zero page area).

However much scrollback RAM we have, we have to subtract 1K from it for
a viewing buffer. This allows us to keep printing to the main screen
RAM while it's not showing, and buffering it as it scrolls, without the
viewing buffer changing while we look at it... also when done viewing,
we toggle back to displaying main screen RAM instantly.

I'm thinking there have to be separate builds for Fast E40 and whatever
80-col driver I get working. The 80 version won't have any support for
scrollback at all, not least because there's no RAM for the buffer!

Also if there's ever going to be a separate edit buffer "window",
or multiple channel windows, fast E40 is where they'll live.

It might also be possible to do word-wrap. Not sure whether that's better
done in the driver or in the app. Doing it in the app means it'll work
in 80-column mode, too.

...potential new feature: should be able to do ping times by sending the
low 2 RTCLOK bytes in numeric form, then when the response comes back,
decode back into 2 bytes, subtract from current RTCLOK, convert to
10ths of a second, and print.

I *really* wish I could do CSLIP. Maybe later. Not delved too deeply
into the guts of uIP... maybe the serial driver could transparently
do the comp/decomp, and speak plain SLIP to uIP? (would this be
too slow, cause drops? Even if it doesn't make us drop packets, would
the CPU overhead negate any benefit to be gained from it?)

Also on the drawing board, keyboard buffering during rs232 I/O. Fairly
straightforward, I think. Don't actually cgetc() or anything, just
save the contents of CH in a buffer if it's not $FF, and set it
to $FF to clear the keystroke. In fact, there's a prototype of it
in keybuftest.c (just needs to be integrated).

Another idea: a status line wouldn't be so hard to do, even without a
fancy E: driver. Could write directly into the first 40 bytes of screen
RAM. I don't want to give up a line of display though... but instead of
a full status line, it'd be cool to display a little up or down arrow
in the upper-right corner during rs232 sends/receives (save & restore
the character that really goes there). The screen codes for the up
and down arrows are $5C and $5D ($DC and $DD for inverse).

/j or /join without a channel should rejoin the current channel, if
FujiChat thinks it's in a channel. This would give us an easy way
to rejoin the channel if kicked.

There needs to be an extra-dumb terminal program for use with manual
login systems. It should fit in 1K or so, load on top of fujichat's
input/output/msg buffer space, and RTS back to fujichat when it's
done. Later on there should be a chatscript engine that loads in
the same space and works the same way.