Previously, when decoding UTF-8, we would first run through the
array to compute the correct size, then allocate a buffer of that size,
then run through the array again to fill the buffer, and then copy it
into a std::wstring. With this fix we can copy it into the string
directly, reducing allocations and only requiring a single pass.
This narrows the range of Unicode codepoints fish reserves for its own
use from U+E000 thru U+F8FE (6399 codepoints) to U+F600 thru U+F73F (320
codepoints). This is still not ideal since fish shouldn't be using any
Unicode private-use codepoints but it's a step in the right direction.
This partially addresses issue #2684.
This was used to cache a narrow string representation
of commands, so that if certain system calls returned errors
after fork, we could output error messages without allocating
memory. But in practice these errors are very uncommon, as are
commands that have wide characters. It is simpler to do a best-effort
output of the wide string, instead of caching a narrow string
unconditionally.
Prior to this fix, read_ni would use parse_util_detect_errors
to lint the script to run, and then parser_t::eval() to execute it.
Both functions would parse the script into a parse tree. This allows
us to re-use the parse tree, improving perfomance.
Introduces a new template moved_ref which is like an rvalue reference.
This allows passing around objects while being explicit that the
receiver may acquire ownership. This will help reduce some allocations.
pcre2_substitute() now sets the output buffer length to PCRE2_UNSET (~0)
if the output buffer is determined to be too small. This change keeps
track of the buffer size separately where pcre2 can't touch it.
A better fix would be to let pcre2 tell fish what size buffer it needs.
This can be done with PCRE2_SUBSTITUTE_OVERFLOW_LENGTH, but this
requires pcre2 10.21 or later (released January 12), which may be too
new to introduce as a dependency at this point.
Fixes#2743
Expand globs to zero arguments (nullglob) only for set, for and count.
The warning about failing globs, and setting the accompanying $status,
now happens regardless of mode, interactive or not.
It is assumed that the above commands are the common cases where
nullglob behaviour is desirable.
More importantly, doing this with `set` is a real feature enabler,
since the resulting empty array can be passed on to any command.
The previous behaviour was actually all nullglob (since commit
cab115c8b9), but this was undocumented;
the failglob warning was still printed in interactive mode,
and the documentation was bragging about failglob behaviour.
The argv argument may be modified on calls to exchange within the function and should not be const qualified (it's not true from the caller's point of view).
On arm, wchar_t is unsigned, and C++11 and newer disallow implicit
narrowing conversions inside braces. Use an explicit conversion to
fix the build on GCC 6 and up, which defaults to C++11.
This changes the default escape timeout for the default keybindings (emacs
mode) to 300ms and the default for vi keybindings to 10ms.
I couldn't resist fixing a few nits in the fish_vi_key_bindings.fish file
since I was touching it to set the escape timeout.
All versions of fish prior to this change silently discarded anything written
to stderr while source a config.fish file. Apparently just to avoid having
the source command display an error if the file did not exist. This can mask
real problems. So instead this change explicitly checks whether the file is
readable and silently skips sourcing it if it isn't.
Resolves issue #2702.
Introduce a "fish_escape_delay_ms" variable to allow the user to configure the
delay used when seeing a bare escape before assuming no other characters will
be received that might match a bound character sequence. This is primarily
useful for vi mode so that a bare escape character (i.e., keystroke) can
switch to vi "insert" to "normal" mode in a timely fashion.
Rather than storing short and long options separately, using
a complicated set of invariants, store them in a single string
and use an explicit type complete_option_type_t to track how they
are interpreted.
This was a "cache" of dubious value that was also very confusing.
The idea was to express in one place all of the short options that
were allowed for a command, in a big string. But it's simpler to
just construct that on-demand by walking the list of
complete_entry_opt_t.
Also remove some other dead code as part of cleanup.
The fix for #2075 inadvertently started unescaping the strings emitted
from `commandline -b`. Only strings emitted with the `-o` flag are
supposed to be unescaped.
Fixes#2210.
If you have a prompt preceded by a new line, you'll get a line full of spaces instead of an empty line above your prompt. This doesn't make a difference in normal usage, but copying and pasting your terminal log becomes a pain. This commit clears that line, making it an actual empty line.
The random builtin command may or may not produce values with a truly
random distribution. So make the documentation reflect that reality. Also,
make the command consistent with similar shells (e.g., bash, zsh) which
produce a range of [0..32767].
Resolves issue #1272.
Before this change, `fish ./test.fish` would fully resolve the
relative paths and symlinks of test.fish, as reported by `status -f`.
However `source` would not. With this change, both cases return relative
paths. `realpath` may be used by scripts to resolve them.
Fixes#2643
My PR #2578 had the unexpected side-effect of altering the tty modes of
commands run via "fish -c command" or "fish scriptname". This change fixes
that; albeit incompletely. The correct solution is to unconditionally set
shell tty modes if stdin is attached to a tty and restore the appropriate
modes whenever an external command is run -- regardless of the path used to
run the external command. The proper fix should be done as part of addressing
issues #2315 and #1041.
Resolves issue #2619
Increase the delay between seeing an escape character and giving up on
whether additional characters that match a key binding are seen. I'm
setting the value to 500 ms to match the readline library. We don't need
such a large window for sequences transmitted by a terminal (even over ssh
where network delays can be a problem). However, we can't expect humans to
reliably press the escape key followed by another key with an inter-char
delay of less than ~250 ms based on my testing and research. A value of
500 ms provides a nice experience even for people using "fish_vi_mode"
bindings as a half second to switch from insert to normal mode is still
fast enough that most people won't even notice.
Resolves#1356
While investigating issue #2619 my first thought was that the problem
had something to do with the "is_interactive_session" global variable.
That preliminary conclusion appears to be wrong (i.e., the problem
lies elsewhere). However, that hypothesis caused me to look at function
"fish_parse_opt" and other mentions of "is_interactive_session".
I decided to take the opportunity to simplify and improve the style of
"fish_parse_opt" since I just spent an hour reviewing the code that
references "is_interactive_session". For example, the "has_cmd" variable
isn't really needed. And there is inconsistent whitespace not to mention
confusion about bool's versus int's and zero versus NULL.
Rather than returning a list of productions and an index,
return the relevant production directly from the rule function.
Also introduce a tag value (replacing production_idx) which tracks
information like command decorations, etc. with more clarity.
When replacing the existing fish process with a new process it is
important to restore the temrinal modes to what they were when fish
started running. We don't want any tweaks done for the benefit of fish
(e.g., disabling ICRNL mode) to bleed thru to an "exec"ed command.
Resolves#2609
If stdio is dead due to EPIPE, there's no great reason to spew a stack dump.
This will still write an error to stderr if stdout dies. This might be
undesirable, but changing that should be considered separately.
It is critical that we ensure our interactive tty modes are in effect at
the earliest possible moment. This achieves that goal and is harmless if
stdin is not tied to a tty. The reason for doing this is to ensure that
\r characters are not converted to \n if we're running on the slave side
of a pty controlled by a program like tmux that is stuffing keystrokes
into the pty before we issue our first prompt.
The special token "normal" should not be in the basic sixteen color table
because a) it is not a color, and b) it is special cased with the result of
resetting the terminal colors (usually via a ANSI X3.64 CSI [0m sequence).
This adds support for the ANSI x3.64 "bright" colors in the basic sixteen
color palette. This is especially useful when trying to use the base colors
as a background color. The "bright" variants tend to be more useful as
background colors compared to the non-bright variants.
This also fixes a bug in so far as palette number 7 is actually grey and
not white whereas palette number 15 is white. At least on the terminal
emulators on which I've tested this change (Ubuntu xterm & uxterm, Mac
OS X Terminal & iTerm2).
Resolves issue #1464.
We identify when the universal variable file has changed out from under us by
comparing a bunch of fields from its stat: inode, device, size, high-precision
timestamp, generation. Linux aggressively reuses inodes, and the size may be
the same by coincidence (which is the case in the tests). Also, Linux
officially has nanosecond precision, but in practice it seems to only uses
millisecond precision for storing mtimes. Thus if there are three or more
updates within a millisecond, every field we check may be the same, and we are
vulnerable to the ABA problem. I believe this explains the occasional test
failures.
The solution is to manually set the nanosecond field of the mtime timestamp to
something unlikely to be duplicated, like a random number, or better yet, the
current time (with nanosecond precision). This is more in the spirit of the
timestamp, and it means we're around a million times less likely to collide.
This seems to fix the tests.
input_mapping_execute, when passed false for allow_commands, will return
R_NULL. However currently it does this unconditionally, even if we don't
have any commands. This defeats our read-ahead optimization, so we
always read and process one byte at a time. This caused pasting to be
much slower.
Fixes#2215
If we are cd'ing into a directory, and the directory has only one
child which is itself a directory, the autosuggestion should
descend as far as it can.
Fixes#2531
Previously 'set -ql' would only look for variables in the
immediate local scope. This was not very useful. It's also
arguably surprising, since a 'set -l' in a function, followed
by a 'set -ql' in a child block, would fail. There was also no
way to check for a function-scoped variable, since omitting the
scope would also pull in global variables.
We could revisit this and introduce an explicit function scope.
Fixes#2502
