I have another patch here for review that optimizes the performance of
glyph drawing, primarily when using non-unit kerning values, and fixes a
few other minor issues. It's dependent on the earlier patch from me that
stores unicode codepoints in a Rune type, typedef'd to uint_least32_t.
This patch is a pretty big change to xdraws so your scrutiny is
appreciated.
First, some performance numbers. I used Yu-Jie Lin termfps.sh shell
script to benchmark before and after, and you can find it in the
attachments. On my Kaveri A10 7850k machine, I get the following
results:
Before Patch
============
1) Font: "Liberation Mono:pixelsize=12:antialias=false:autohint=false"
cwscale: 1.0, chscale: 1.0
For 273x83 100 frames.
Elapsed time : 1.553
Frames/second: 64.352
Chars /second: 1,458,159
2) Font: "Inconsolata:pixelsize=14:antialias=true:autohint=true"
cwscale: 1.001, chscale: 1.001
For 239x73 100 frames.
Elapsed time : 159.286
Frames/second: 0.627
Chars /second: 10,953
After Patch
===========
3) Font: "Liberation Mono:pixelsize=12:antialias=false:autohint=false"
cwscale: 1.0, chscale: 1.0
For 273x83 100 frames.
Elapsed time : 1.544
Frames/second: 64.728
Chars /second: 1,466,690
4) Font: "Inconsolata:pixelsize=14:antialias=true:autohint=true"
cwscale: 1.001, chscale: 1.001
For 239x73 100 frames.
Elapsed time : 1.955
Frames/second: 51.146
Chars /second: 892,361
As you can see, while the improvements for fonts with unit-kerning is
marginal, there's a huge ~81x performance increase with the patch when
using kerning values other than 1.0.
So what does the patch do?
The `xdraws' function would render each glyph one at a time if non-unit
kerning values were configured, and this was the primary cause of the
slow down. Xft provides a handful of functions which allow you to render
multiple characters or glyphs at time, each with a unique <x,y> position,
so it was simply a matter of massaging the data into a format that would
allow us to use one of these functions.
I've split `xdraws' up into two functions. In the first pass with
`xmakeglyphfontspecs' it will iterate over all of the glyphs in a given
row and it will build up an array of corresponding XftGlyphFontSpec
records. Much of the old logic for resolving fonts for glyphs using Xft
and fontconfig went into this function.
The second pass is done with `xrenderglyphfontspecs' which contains the
old logic for determining colors, clearing the background, and finally
rendering the array of XftGlyphFontSpec records.
There's a couple of other things that have been improved by this patch.
For instance, the UTF-32 codepoints in the Line's were being re-encoded
back into UTF-8 strings to be passed to `xdraws' which in turn would then
decode back to UTF-32 to verify that the Font contained a matching glyph
for the code point. Next, the UTF-8 string was being passed to
`XftDrawStringUtf8' which internally mallocs a scratch buffer and decodes
back to UTF-32 and does the lookup of the glyphs all over again.
This patch gets rid of all of this redundant round-trip encoding and
decoding of characters to be rendered and only looks up the glyph index
once (per font) during the font resolution phase. So this is probably
what's responsible for the marginal improvements seen when kerning values
are kept to 1.0.
I imagine there are other performance improvements here too, not seen in
the above benchmarks, if the user has lots of non-ASCII code plane characters
on the screen, or several different fonts are being utilized during
screen redraw.
Anyway, if you see any problems, please let me know and I can fix them.
When user clicks LMB, one character is selected, but will not be copied
to selection until the user moves cursor a bit. Therefore, the character
should not be highlighted as selected yet.
Before the patch, the trick was not to mark line as dirty to avoid
highlighting it. However, if user has already selected something and
clicks in line that contains selection, selclear sets the line as dirty
and one character is highlighted when it should not.
This patch replaces dirty trick with explicit check for sel.mode inside
selected().
This patch also prevents sel.mode from increasing beyond 2. It is almost
impossible, but sel.mode may overflow if mouse is moved around for too
long while selecting.
st.c:1321:2: warning: ignoring return value of function declared with warn_unused_result attribute [-Wunused-result]
system(cmd);
^~~~~~ ~~~
Debatable whether an error here should case exit(EXIT_FAILURE). Just
preserving the existing behaviour for now.
Not always is desirable to create a pseudo terminal, and some times
we want to open a terminal emulator over a tty line. With this new
patch is possible to do someting like:
$ st -l /dev/ttyS0 115200
Without this option was needed to launch another terminal emulator
over st (for example minicom, picocom, cu, ...).
ICCCM mandates the use of real timestamps to interact with the
selection, to rule out race conditions if the clients are run at
different speeds. I have implemented the low hanging fruit, putting the
timestamps into text selection. Also, ICCCM mandates a check for whether
XSetSelectionOwner() worked. Not sure my version is correct, though.
tmoveto resets CURSOR_WRAPNEXT.
Simple testcase:
for i in $(seq 1 200); do
printf '\t.';
usleep 100000;
printf '\t@';
usleep 100000;
done
In st executing this script causes @ and . to overwrite each other in
the last column.
XFilterEvent usually filters KeyPress events according to input method.
At this point the window is not mapped. The only events that we process
are ConfigureNotify and MapNotify. They should not be filtered by input
method.
strsep() is not a POSIX function, and it means that every system
needs different defines to expose it. If the prototype of strsep
is not exposed then an ugly int/pointer is done and it might mean
a crash. The best solution?, to remove the strsep and make a custom
loop. If C programmers cannot do this kind of loops without calling
a library function, then maybe we should move all the suckless
software to Java.
Some programs can only deal with XA_STRING, and it makes impossible st
be able of copying to them. This patch makes st answer also to XA_STRING,
althought it sends utf8 strings. It is not a problem because moderm
applications must support utf8.
Thanks to Alex Pilon <alp@alexpilon.ca>!
Now there is a distinction between the primary and clipboard selection. With
Mod + Shift + c/v the clipboard is handled. The old Insert behavious does
reside.
The unicode long is added to the cache. So when fontconfig does fall back to
the default font (where there is no easy way to find this out from the
pattern) it isn't reloaded.
Use the terminfo delay syntax ($<x>) in our flash capability to avoid
hardcoding a fixed delay in redraw() when called from tsetmode() with
DECSCNM.
We need to turn on the npc capability so that delays are made with
xon/xoff instead of padding characters.
When MODE_INSERT is set we'd shift characters on the same
line forward before inserting our character in tputc().
This did not account for wide characters where width != 1.
This patch makes it so we shift the correct amount.
In tputc(), when a character wasn't large enough to fit
on the current line, we would call tnewline() to place it on
the next line. Unfortunately, we weren't resetting our glyph
pointer and this caused memory corruption when a
wide character (width == 2) was being written. This patch
resets our glyph pointer after calls to tnewline().
If blinktimeout is set to a value greater than 1000, pselect will
receive a timeout argument with tv_nsec greater than 1E9 (1 sec), and
fail, making st crash. This patch just ensures that the timespec
structure is correctly filled with a value properly decomposed between
tv_sec and tv_nsec.
Reported by JasonWoof on IRC. Thanks!
Trailing whitespaces are trimmed when copying from normal selection and
rectangular selection on lines that have their last character included
or on the left of the selection. It leads to inconsistent behaviors when
copying the exact same text from the left and right window in
applications with vertical splits.
This patch solves this issue by always trimming the selection.
- POSIX states the SHELL environment variable "... shall represent a
pathname of the user's preferred command language interpreter." As
such, st should check for its presence when deciding what shell to
use; just as HOME can be defined to override one's passwd-defined home
directory, a user should also be able to override their passwd-defined
shell using the SHELL environment variable.
The XBell() call currently used when a bell is recieved sends a message
to the X server, but if the X server doesn't know how to sound it,
it just gets ignored and I have not been able to find anywhere in x.org's
code a way to configure the action that the server does.
However, if you use XkbBell() then you can have a process listening for
the XkbBellNotifyEvent that is produced and either alert you visually or
play an audio file or whatever you want as your notification. You have
to include one more header file but the function seems to be compiled as
part of Xlib, at least on my installation.
CustaiCo
SI (0x0F or ^O) means Shift In, and it selects G1 charset definition,
and SO (0x0E or ^N) means Shift Out, and it selects G0 charset
definition, but st was doing just the inverse.
St runs an interactive shell and not a login shell, and it means
that profile is not loaded. The default terminal configuration
in some system is not the correct for st, but since profile is
not loaded there is no way of getting a script configures the
correct values.
St doesn't update the utmp files, this is the job of another
suckless tool, utmp. Utmp also opens a login shell (it is the
logical behaviour when you create a new user record) it is a
good option execute utmp and then get a correct input in
utmp, wtmp and lastlog file, and execute the content of the
profile.
When getting selected text, lines that were wrapped because of length
ought not include the wrapping newline in the selection.
This comes up, for example, when copying a bash command that is long
enough to wrap from the console and pasting it back into the console.
The extra newline breaks it.
Similiarly, changes behavior when trimming whitespace from the end of a
physical line to only do so if the line does not wrap. Otherwise we are
trimming whitespace from the middle of a logical line, which may change
its meaning.
Signed-off-by: Roberto E. Vargas Caballero <k0ga@shike2.com>
This makes any sequence of identical delimiters be considered a single
word in word-snapping mode. This seems more coherent for this mode and
is similar to what xterm does.
Signed-off-by: Roberto E. Vargas Caballero <k0ga@shike2.com>
This simplifies getbuttoninfo() and bpress(), and fixes a bug which made word
snapping behave incorrectly when a delimiter was at the beginning or end of
line.
Signed-off-by: Roberto E. Vargas Caballero <k0ga@shike2.com>
The 'left shift from one' notation of power of two integers is more
expressive than the result.
Signed-off-by: Alexander Huemer <alexander.huemer@xx.vu>
Signed-off-by: Roberto E. Vargas Caballero <k0ga@shike2.com>
We already have a csihandle() function, where is located code about
CSI sequences, so it is logical do the same with ESC sequences.
This change helps to simplify tcontrol(), which has a complex flow
and should be rewritten.
DEL character is not thecnically talking a C0 control character,
although it has some common properties with them, so it is useful
for us consider it as C0. Before this patch DEL (\177), was not
ignored as it ought to be.
Man page was repeating -f option, the second time instead of -i,
and this option was lost in usage() message. This patch also indent
the output of usage().
VT102ID is the sequence that the terminal returns when it is inquired
to identify itself. This value should be configurable in the same
way that another st parameters.
ISCONTROL chechks if a value is between 0 and 0x1f or
between 0x80 and 0x9f. Char signess depends of architecture
and compiler, so in some environment the second case is
always false (and wrong), Techo() calls ISCONTROL with a
char variable, whose type cannot be changed because tpuc()
expects a pointer to char, so the solution is to insert a
cast in the call to ISCONTROL.
tclearregion() was clearing regions using spaces and the current
attributes of the terminal. It was correct with all the modes excepct
underline, because they didn't affect the space character, but in
the case of underline it was a problem. A easy way of seeing this
problem is writing this in the last line of the terminal:
tput smul ; echo first; tput rmul; echo second; echo third
Fist was underlined, and second and third were not underlined, but
the spaces at the right of second was underlined becuause in the
previous scrool underline mode was set.
Master proccess was not showing any error message when the child
died with an error, and it was very confusing for the user (for
example with incorrect -e command).
One blinking mode is good enough, and two is too much. The best aproach
is emulate the fast blinking with the slow blinking, that it is more
used.
It is removed the flag ATTR_FASTBLINK because it has not a different
meaning of ATTR_BLINK, so it is not needed.
Signed-off-by: Roberto E. Vargas Caballero <k0ga@shike2.com>
According to ECMA-48¹ 8.3.117, an attribute value of 21 is "doubly
underlined", while 22 is "normal colour or normal intensity (neither
bold nor faint)".
Additionally, 25 is "steady (not blinking)", which likely means neither
slow blink nor fast blink.
¹: http://www.ecma-international.org/publications/files/ECMA-ST/Ecma-048.pdf
Signed-off-by: Roberto E. Vargas Caballero <k0ga@shike2.com>
XFilterEvent need to be called against every event, otherwise it would
missing some message in the xim protocol and misbehave on some im server.
Signed-off-by: Roberto E. Vargas Caballero <k0ga@shike2.com>
Faint text is implemented by allocating a new color at one-half
intensity of each of the r, g, b components, or if the text bold at the
same time, it is not made lighter.
Signed-off-by: Roberto E. Vargas Caballero <k0ga@shike2.com>
Signed-off-by: Christoph Lohmann <20h@r-36.net>
There were a few occurrences of strcmp and strlen being called on Glyph.c[],
which is not always null-terminated (this actually depends on the last values in
the buffer s in ttyread()). This patch replace all the calls to strcmp with a
test on c[0] directly or a call to tlinelen, and the one to strlen with utf8len.
I also took the opportunity to refactor getsel and tdumpline.
Signed-off-by: Roberto E. Vargas Caballero <k0ga@shike2.com>
Implement crossed-out text with an XftDrawRect call, similar to how
underline is implemented. The line is drawn at 2/3 of the font ascent,
which seems to work nicely in practice.
Signed-off-by: Roberto E. Vargas Caballero <k0ga@shike2.com>
Implement invisible mode by setting the foreground color to be the same
as the background color. Not rendering anything would also be an
alternative, but this seems less likely to cause surprises in
conjunction with any hacks.
Signed-off-by: Roberto E. Vargas Caballero <k0ga@shike2.com>
Faint, invisible, struck and fast blink are added as glyph attributes.
Since there's an edit here, let's take the opportunity to reorder them
so that they correspond to the two's power of the corresponding escape
code. (just for neatness, let's hope that property never gets used for
anything.)
Signed-off-by: Roberto E. Vargas Caballero <k0ga@shike2.com>
This macro was not correct in some cases, and it was used only in
one place, where we did'nt get any benefit in performance of in size,
so the macro is removed and ceilf is used instead of it. The only
function needed from math.h is ceilf, so this patch defines the
prototype of it instead of including math.h.
Commit 5edeec1 introduced a wrong factor for nanosecond computation, the correct
value is 1E6. Time and timeout values are 10 times less than they should be and
this cause high CPU usage.
Reported by pyroh on IRC. Thanks!
Signed-off-by: Roberto E. Vargas Caballero <k0ga@shike2.com>
This patch replaces the gettimeofday()/timeval-system with
uses of clock_gettime() with a monolithic clock and timespec-structs.
gettimeofday() is not accurate and prone to jumps and POSIX.1-2008
marks it as obsolete. Read more here [0].
The patch should speak for itself and decreases the binary
size for me by almost 200K(!).
[0]: http://blog.habets.pp.se/2010/09/gettimeofday-should-never-be-used-to-measure-time
Signed-off-by: Roberto E. Vargas Caballero <k0ga@shike2.com>
Refactor the SNAP_WORD part in selsnap, and fix a bug that caused the word
delimiters to be ignored if it was at the very beginning or end of a wrapped
line.
Signed-off-by: Roberto E. Vargas Caballero <k0ga@shike2.com>
selsort computes the wrong normalized coordinates when rectangular
selection is enabled, causing rectangular selection to only work
when going toward either the top left corner, or the bottom right
one.
Signed-off-by: Roberto E. Vargas Caballero <k0ga@shike2.com>
Currently, selection is expanded to the end of the line over line breaks only in
regular selection mode, when the line in not empty and when going down and/or
right. This covers all the cases including word selection mode, with the
exception of rectangular selection because it would make this mode too rigid.
This adjustment is made in selsort so I renamed it to selnormalize to better
reflect what it does now.
Signed-off-by: Roberto E. Vargas Caballero <k0ga@shike2.com>
By the recommendation of FRIGN I refactored xsetcolorname to remove the
unnecessary r, g, b variables when allocating a new color. Colors are
now freed and set to the new color. A die() should not happen here. Oth‐
erwise it is easy for applications to kill st. St should be resilent to
malicious input.
Second this patch standardises the naming of »color«. There is no
»colour« here. Maybe in some parts of the world.
I mainly improved the slightly off algorithm used to load colours in the 256-colour-space and
removed unnecessary local values (r,g,b,colour).
"colour" is not necessary as a punchbag for XftColorAlloc[Value,Name], as they don't mess with
the result-adress until they are absolutely sure everything worked out[0].
Being at it, I changed the error-returns for AllocValue to dies (just like in xloadcols()), as
a failure is most likely an OOM-situation you better catch early.
In case of an invalid name everything stays the same.
[0]: http://www.opensource.apple.com/source/X11libs/X11libs-40/libXft/libXft-2.1.13/src/xftcolor.c
Signed-off-by: Christoph Lohmann <20h@r-36.net>
Similar to xterm or urxvt holding shift before selecting text with the mouse
allows to override copying text. For example in tmux with "mode-mouse on" or
vim (compiled with --with-x), mc, htop, etc.
forceselmod in config.h sets the modifier to use this mode, by default
ShiftMask.
Signed-off-by: Hiltjo Posthuma <hiltjo@codemadness.org>
ATTR_GFX was used long time ago to detect when terminal was in
graphic mode. Today graphic mode is implemented using a charset
pointer, so ATTR_GFX is not needed anymore because graphic
condition can be detected directly checking if current charset
is GRAPHICS C0.
This patch fixes the bug introduced in
8f11e1cd03
To reproduce the bug:
1. Save cursor: printf '\e[s'
2. Load cursor: printf '\e[u'
3. Resize st window.
4. Load cursor again: printf '\e[u'
The patch 53105cf modified how control codes were detected, because
it tried to handle also C1 control codes (0x80-0x9f), that have
upper bit to 1, so they are multi byte character in utf8.
Code was checking the value of width in order to known that after
decoding the unicode point had a width of 1 byte, but it as incorrect
because this width is the columnb width.
Once a sequence is completed term.esc must return to 0, so
instead of repeating this expression in all the cases is
better put it at the end of the block.
From http://www.vt100.net/docs/vt510-rm/chapter4:
*The VT510 ignores all following characters until it receives a
SUB, ST, or any other C1 control character.
So OSC, PM and APC sequence ends with a SUB (it cancels the sequence
and show a question mark as error), ST or any another C1 (8 bits)
code, or their C0 (7 bits) equivalent sequences (at this moment we
do not handle C1 codes, but we should). But it is also said that:
Cancel CAN
1/8 Immediately cancels an escape sequence, control sequence,
or device control string in progress. In this case, the
VT510 does not display any error character.
Escape ESC
1/11 Introduces an escape sequence. ESC also cancels any escape
sequence, control sequence, or device control string in
progress.
Currently tputc handles the case of too long control string waiting for
the end of control string.
Another case is when there is ESC character is encountered but is not
followed by '\\'. In this case st stops processing control string,
but ESC character is ignored.
After this patch st processes ESC characters in control strings properly.
Test case:
printf '\e]0;abc\e[1mBOLD\e[0m'
Also ^[\ is actually processed in the code that handles ST.
According to ECMA-048 ST stands for STRING TERMINATOR and is used to
close control strings.
Thanks to Yuri Karaban for suggesting this!
These changes make -g correspond to <cols>x<rows> and honor it so non-tiling
window managers can work with the size hints afterwards. It also adds a -i
flag to force the window size. This is needed so -g keeps being useful in dwm.
The large and repeated expression used in memmove to indirect
the line can be simplified using a pointer, that makes more
clear where begins and where ends the movement.
Current CSI parsing code uses strtol to parse arguments and allows them
to be negative. Negative argument is not properly handled in tdeletechar
and tinsertblank and results in memory corruption in memmove.
Reproduce with printf '\e[-500@'
Patch also removes special handling for corner case and simplifies
the code.
Removed
term.dirty[term.c.y] = 1
because tclearregion sets dirty flag.
tscrollup and tscrolldown do not use tsetdirt, but their code is
equivalent to
tsetdirt(orig, term.bot-n);
tsetdirt(orig+n, term.bot);
tclearregion also marks cleared lines as dirty.
In tscrolldown it sets lines from term.bot-n+1 to term.bot dirty, and in
tscrollup it sets lines from orig to orig+n-1 dirty.
In both functions all lines from orig to term.bot are effectively set
dirty, but in tscrolldown lines from orig+n to term.bot are set dirty
twice, and in tscrollup lines from orig to term.bot-n are set dirty
twice.
These patches make it clear which lines are set dirty and sets them
dirty once in each funciton.
techo compares signed char to '\x20'. Any character with code less then
'\x20' is treated as control character. This way characters with MSB
set to 1 are considered control characters too.
Also this patch makes techo display DEL character as ^?.
To reprocuce the bug, enable echo mode using printf '\e[12l',
then type DEL character or any non-ASCII character.
I found the SERRNO Macro slightly confusing, since you have to look
it up, if you don't know it already. A web search showed it does
not seem to be any kind of standard. Also there was no reason in
the commit log when it was introduced in 2009. As you can see it
also leads to new patches, which don't use this macro (probably the
author did not know about it).
I don't like this alt screen thing, but when
allowaltscreen == 0, the cursor is still saved
and restored after calling 'less' (or 'man').
This patch makes allowaltscreen == 0 usable.
This patch replaces current utf decoder with a new one, which is ~50
lines shorter and should be easier to understand. Parsing 5 and 6
sequences, if necessary, requires trivial modification of UTF_SIZ
constant and utfbyte, utfmask, utfmin, utfmax arrays.
This sequence print the current line. It is different to the
'printer on' sequence, where all the characters that arrive to the
terminal are printer. Here only the ascii characters are printed.
Christoph Lohmann
suigin
noname
Jochen Sprickerhof
mvdan@mvdan.cc
alp@alexpilon.ca
Roberto E. Vargas Caballero
sin
Markus Wichmann
Omar Sandoval
LemonBoy
Ivan Delalande
Nils Reuße
Rian Hunter
Rian Hunter
Eric Pruitt
czarkoff@gmail.com
Quentin Rameau
CustaiCo
Ben Hendrickson
Alexander Huemer
Quentin Carbonneaux
Michael Forney
Weng Xuetian
Anders Eurenius
FRIGN
Troy Sankey
Silvan Jegen
Hiltjo Posthuma
Balazs Kezes
Alexander
Markus Teich
Toaster Toaster
Rafa Garcia Gallego
Damian Okrasa