Thanks Alex Schroeder for reporting this crash. The scenario:
* Edit a file like say this repo's Readme.
* The second line is empty and there's a '+' to insert a drawing.
Click on that.
* Resize the window so just the first line of text and the drawing are
visible.
* Close the window.
* Reopen lines.love, it will reopen the same file.
* Click on the left margin to the left of the drawing.
Before this commit these steps yielded the following crash:
Error: bad argument #1 to 'len' (string expected, got nil)
text.lua:626: in function 'pos_at_end_of_screen_line'
edit.lua:298: in function 'mouse_press'
There were two distinct problems here:
1. State.screen_bottom1 is not required to point to a text line, it
could just as well be a drawing. I have been sloppy in handling that.
2. The bug was partially masked (the need to close and reopen the
window) by a second bug: inserting a drawing was not invalidating the
cache I save of starty coordinates for each line. (I've inserted and
deleted starty invalidations a few times in the past, but it looks
like I'd never had one in this particular location edit.draw before.)
How did these issues get missed for years?
- Even though I use lines.love on a daily basis, it turns out I don't
actually create line drawings all that often.
- When I do, I'm still living in files that are mostly text with only
an occasional drawing.
- I keep my windows fairly large.
Between these 3 patterns, the odds of running into a drawing as the
first or bottom-most line on the screen were fairly small. And then I
had to interact with it. I suspect I tend to interact with drawings
after centering them vertically.
---
Bug #1 in particular has some interesting past history.
* Near the start of the project, when I implemented line-wrapping I
started saving screen_bottom, the bottom-most line displayed on
screen. I did this so I could scroll down easily just by assigning
`screen_top = screen_bottom`. (On the other hand, scrolling up still
required some work. I should perhaps get rid of it and just compute
scrolls from scratch each time.)
* Also near the start of the project, I supported selecting text by a
complex state machine spanning keypress, mouse press and mouse
release:
mouse click (press and immediate release) moves cursor
mouse drag (press and much later release) creates selection
shift-click selects from current cursor to click location
shift-movement creates/grows a selection
* On 2023-06-01, inscript reported a bug. Opening a window with just a
little bit of text (lots of unused space in the window), selecting all
the text and then clicking below all the text would crash the editor.
To fix this I added code at the bottom of edit.mouse_press which
computed the final visible line+pos location and used that in the
cursor-move/text-selection state machine. It did this computation
based on.. screen_bottom. But I didn't notice that screen_bottom could
be a drawing (which has no pos). This commit's bug/regression was
created.
* On 2023-09-20, Matt Wynne encountered a crash which got me to realize
I need code at the bottom of edit.mouse_release symmetric to the code
at the bottom of edit.mouse_press. I still didn't notice that
screen_bottom could be a drawing.
So in fixing inscript's bug report, I introduced (at least) 2
regressions, because I either had no idea or quickly forgot that
screen_bottom could point at a drawing.
While I created regressions, the underlying mental bug feels new. I just
never focused on the fact that screen_bottom could point at a drawing.
This past history makes me suspicious of my mouse_press/mouse_release
code. I think I'm going to get rid of screen_bottom entirely as a
concept. I'll still have to be careful though about the remaining
locations and which of them are allowed to point at drawings:
- cursor and selection are not allowed to point at drawings
- screen_top and screen_bottom are allowed to point at drawings
I sometimes copy between these 4 location variables. Auditing shows no
gaps where cursor could ever end up pointing at a drawing. It's just
when I started using screen_bottom for a whole new purpose (in
the mouse_press/release state machine) that I went wrong.
I should also try getting rid of starty entirely. Is it _really_ needed
for a responsive editor? I think I introduced it back when I didn't know
what I was doing with LÖVE and was profligately creating text objects
willy-nilly just to compute widths.
Getting rid of these two fairly global bits of mutable state will
hopefully make lines much more robust when the next person tries it out
in 6 months :-/ X-(
Thanks everyone for the conversation around this bug:
https://merveilles.town/@akkartik/112567862542495637
---
Bug #2 has some complexity as well, and might lead to some follow-on
cleanup.
When I click on the button to insert a new drawing, the mouse_release
hook triggers and moves the cursor below the new drawing. This is
desirable, but I'd never noticed this happy accident. It stops working
when I invalidate starty for all lines (which gets recomputed and cached
for all visible lines on every frame).
Fixing this caused a couple of unit tests start crashing for 2 reasons
that required their own minor fixes:
- My emulated mouse press and release didn't have an intervening
frame and so mouse_release no longer receives starty. Now I've added
a call to edit.draw() between press and release.
This might actually bite someone for real someday, if they're
running on a slow computer or something like that. I've tried to
click really fast but I can't seem to put mouse_press and release in
the same frame (assuming 30 frames per second)
- My tests' window dimensions often violate my constraint that the
screen always have one line of text for showing the cursor. They're
unrealistically small or have a really wide aspect ratio (width 2x
of height). I suspect lines.love will itself crash in those
situations, but hopefully they're unrealistic. Hmm, I wonder what
would happen if someone maximized in a 16:9 screen, that's almost
2x.. Anyways, I've cleaned a couple of tests up, but might need to
fix up others at some point. I'd have to rejigger all my brittle
line-wrapping tests if I modify the screen width :-/ X-(
scenario:
- create a long wrapping line
- tap past end of first screen line
Before this commit the cursor would be positioned not quite at the end
of the screen line but one character before. In effect there was no way
to position cursor at end of a wrapping line.
I'm not sure how this bug has lasted so long. It was introduced in
commit 8d3adfa36 back in June 2022, which was itself billed as a bugfix
for "clicking past end of screen line". But when I go back to it this
bug exists even back then. How did I miss it?! I wrote a test back then
-- and the test was wrong, has always been wrong.
This is all quite hacky. Many of my tests are unfortunately brittle to
changes in text rendering. Fortunately there's only one test that
currently requires a hacky special case (and a second test I tweaked
slightly to be more robust).
I can't think of a better approach. It doesn't help to standardize the
font, because version changes still come with changes to text-shaping
algorithms even if the font itself is unchanged. I could base all my
assertions on the widths of individual characters, but that would make
the tests much less readable and not express intent as clearly. So here
we are, with hopefully just a few hacky special cases (there might be a
few more as LÖVE v12 advances towards publication, and in further
versions).
Matt Wynne pointed out that snap.love would crash when a node went off
screen. While debugging it I noticed that selection1 was being set when
it shouldn't be.
Turns out I introduced a bug when I fixed the inscript bug back in June
(commit 9656e13774). One invariant I want to preserve is: selection1
should be unset after a mouse click (press and release without
intervening drag). This invariant was violated in my bugfix back in
June. I was concerned only with selection back then, and I didn't
realize I was breaking the mouse click case (in a fairly subtle way; you
can have selection set, and when it's set identically to the cursor
everything looks the same).
I think there might still be an issue in snap.love after this fix. I
noticed screen_bottom1.pos was nil, and as far as I recall that should
never happen.
To fix this I have to first stop incrementally updating screen_bottom1
in the middle of a frame. Now it always has a good value from the end of
a frame.
I'm also running into some limitations in the test I'd ideally like to
write (that are documented in a comment), but I still get some sort of
automated test for this bugfix.
It's a hack:
- if you start selecting from below final line the start of the
selection is the most recent click even if it was forever ago
- (the crash we're currently fixing) if you start up and immediately
select all then click below final line => crash. recent_mouse was
never set.
- getting rid of it breaks no tests (except the crash we're currently
fixing)
Disquieting that none of my tests caught these. On the other hand, I
also haven't noticed any issues in practice. Perhaps cache invalidation
is often unnecessary.
The published version of lines.love was broken for almost an hour. The
cursor would render one position to the right of where it really is. To
fix it, this commit rolls back 26ba6e4e5a. There doesn't seem a good
way to test it.
I've been sloppy about this so far, and outside of tests I can't find
any examples where it matters, but it matters in a potential fork where
I'm rendering multiple columns of text.
It's unfortunate that my tests have this level of brittleness. What I'd
really like to assert in many of these changed lines is that the text
stays inside the margins and that more text would overflow margins.