To achieve this we have to switch to a model of the screen in termbox that
is closer to the underlying terminal.
Before:
a screen is a grid of characters
writing out of bounds does nothing
After:
a screen is a scrolling raster of characters
writing out of bounds wraps to next line and scrolls if necessary
To move to the new model, it was essential that I migrate my fake screen
at the same time to mimic it. This is why the first attempt (commit 3824)
failed (commit 3858). This is also why this commit can't be split into
smaller pieces.
The fake screen now 'scrolls' by rotating screen lines from top to bottom.
There's still no notion of a scrollback buffer.
The newer model is richer; it permits repl-like apps that upstream termbox
can't do easily. It also permits us to simply use `printf` or `cout` to
write to the screen, and everything mostly works as you would expect. Exceptions:
a) '\n' won't do what you expect. You need to explicitly print both '\n'
and '\r'.
b) backspace won't do what you expect. It only moves the cursor back,
without erasing the previous character. It does not wrap.
Both behaviors exactly mimic my existing terminal's emulation of vt100.
The catch: it's easy to accidentally scroll in apps. Out-of-bounds prints
didn't matter before, but they're bugs now. To help track them down, use
the `save-top-idx`, `assert-no-scroll` pair of helpers.
An important trick is to wrap the cursor before rather after printing
a character. Otherwise we end up scrolling every time we print to the
bottom-right character. This means that the cursor position can be invalid
at the start of a print, and we need to handle that.
In the process we also lose the ability to hide and show the screen. We
have to show the prints happening. Seems apt for a "white-box" platform
like Mu.
Always start with an untouched screen that can scroll on printing "\r\n".
We can still clear the screen as needed.
Also drop support for hiding the cursor.
Now it's much more apparent why things are slow. You can see each repaint
happening. Already I fixed one performance bug -- in clear-rest-of-screen.
Since this subverts Mu's fake screen there may be bugs.
Another salubrious side effect: I've finally internalized that switching
to raw mode doesn't have to clear the screen. That was just an artifact
of how termbox abstracted operations. Now I can conceive of using termbox
to build a repl as well.
(I was inspired to poke into termbox internals by
http://viewsourcecode.org/snaptoken/kilo and
https://github.com/antirez/linenoise)
I'd messed up termbox in commit 3443; it was weird how it failed though.
The terminal got really sluggish to switch between windows when the
edit/ app was running. And it stopped clearing the screen properly.
Commit 3191 stopped defining _XOPEN_SOURCE when building termbox/ to get
Mu to build on OpenBSD. However, that had the side effect of not
declaring the prototype for wcwidth() on some versions of Linux. Ugh.
Just hack around this morass.
In general the direction we want to go in Mu is fewer feature #defines.
At least explicit ones. Should be an internal detail of the underlying
platform our code shouldn't have to worry about. If headers don't
cooperate, just start explicitly declaring prototypes and damn the
consequences.
Looks like the _XOPEN_SOURCE #define isn't needed in termbox anymore, at
least after I removed some features from it that I don't need. All it
was doing is hiding SIGWINCH and likely other names as well.
On slow networks sometimes escape sequences were being partially
consumed, causing junk to be added to the editor when you pressed arrow
keys and so on. Now we have a way to wait.
Empirically seems to work if I page-up and then scroll back up using
up-arrow. Before I'd consistently get junk even on my local machine. Now
I no longer do.
If we still see problems I'll increase the wait time and see if the
increase helps. Then we'll know more about this approach.
For some reason porting the termbox-go implementation was still leaving
some gunk from git on screen when I ran my usual test:
$ mkdir lesson; cd lesson; git init; mu edit.mu
Then hit F4, generating messages from git on the initial commit.
Then hit ctrl-l to clear all git gunk.
Mu still isn't so optimized that I can be profligate with spare cycles.
Instead we'll follow termbox-go's example and create a new API routine
called tb_sync() (after discussion with termbox's author). Now we only
need use tb_sync() on ctrl-l. For everything else use the optimized
render.
Now I think I've eradicated all signs of "cursor thrashing" during
refresh, in spite of how slow mu is as an interpreted language. We only
render the whole screen in some situations, and only if there's no more
input, and even then we only refresh the parts of the screen that
changed.
Thanks Jack and Caleb Couch for providing the impetus behind commits
2109-2113.
I've been lazy about writing tests for all this, but it's still good to
know I could, if I wanted to.
Termbox had been taking shortcuts when it thinks the screen hasn't
changed, which doesn't work if some other process messes up the screen.
The Go version has a Sync method in addition to Flush/tb_present for
precisely this eventuality. But it feels like an unnecessary
optimization given C's general speed. Just drop it altogether.
---
This took me a long time to track down, and interestingly I found myself
writing a new tracing primitive before I remembered how to selectively
trace just certain layers during manual tests. I'm scared of generating
traces not because of performance but because of the visual noise. Be
aware of this. I'm going to clean up $log now.
Maybe I should also stop using $print..
It took me a long time to fix termbox because the escape codes it was
seeing seemed all wrong. Had to stop calling tb_shutdown/printf and put
in the extra 3 lines to log to a file. Then everything cleared up.
Weird.
Lots mixed into this commit:
some off-by-one errors in display.cc
a new transform to translate jump labels that I'd somehow never gotten around to supporting
Kartik Agaram