.unwrap() is in effect an assert(). If it is applied mistakenly, the
program crashes and there isn't a good error.
I would like it to be used as a last resort. In these cases there are
nicer ways to do it that handle a missing result properly.
Issue #10194 reports Cobra completions do
set -l args (commandline -opc)
eval $args[1] __complete $args[2..] (commandline -ct | string escape)
The intent behind "eval" is to expand variables and tildes in "$args".
Fair enough. Several of our own completions do the same, see the next commit.
The problem with "commandline -o" + "eval" is that the former already
removes quotes that are relevant for "eval". This becomes a problem if $args
contains quoted () or {}, for example this command will wrongly execute a
command substituion:
git --work-tree='(launch-missiles)' <TAB>
It is possible to escape the string the tokens before running eval, but
then there will be no expansion of variables etc. The problem is that
"commandline -o" only unescapes tokens so they end up in a weird state
somewhere in-between what the user typed and the expanded version.
Remove the need for "eval" by introducing "commandline -x" which expands
things like variables and braces. This enables custom completion scripts to
be aware of shell variables without eval, see the added test for completions
to "make -C $var/some/dir ".
This means that essentially all third party scripts should migrate from
"commandline -o" to "commandline -x". For example
set -l tokens
if commandline -x >/dev/null 2>&1
set tokens (commandline -xpc)
else
set tokens (commandline -opc)
end
Since this is mainly used for completions, the expansion skips command
substitutions. They are passed through as-is (instead of cancelling or
expanding to nothing) to make custom completion scripts work reasonably well
in the common case. Of course there are cases where we would want to expand
command substitutions here, so I'm not sure.
Move all qmark tests to `scoped_test()` sections with explicitly set feature
flags. We already test the default qmark behavior in the functionality tests.
It seems the logic for calculating the cursor position was not ported correctly,
because the correct place to insert it is at the cursor_pos regardless of
range.start, going by the parameters submitted to the function and the expected
result.
This allows running a fish built from `cargo build` *and* built via
cmake.
In future, we should make this an optional thing that's removed from
installed builds.
NCurses headers contain this conditional "#define cur_term":
print "#elif @cf_cv_enable_reentrant@"
print "NCURSES_WRAPPED_VAR(TERMINAL *, cur_term);"
print "#define cur_term NCURSES_PUBLIC_VAR(cur_term())"
print "#else"
OpenSUSE Tumbleweed uses this configuration option; For reentrancy, cur_term
is a function. If the NCurses autoconf variable @NCURSES_WRAP_PREFIX@
is not changed from its default, the function is called _nc_cur_term.
I'm not sure if we have a need to support non-default @NCURSES_WRAP_PREFIX@
but if we do there are various ways;
- search for the symbol with the cur_term suffix
- figure out the prefix based on the local curses installation,
for example by looking at the header files.
Fixes#10243
This is run in the current locale, without resetting to en_US.UTF-8
like our integration tests do.
So if you want to check for a specific message you need to check the
localized version.
This was previously limited to Linux predicated on the existence
of certain headers, but Rust just exposes those functions unconditionally. So
remove the check and just perform the mtime hack on Linux and Android.
Make sure to also look for the error part that occurs after the last format
specifier.
Still not great because it won't fail if there's unexpected output at the
beginning or end of the string.
Commit 5f849d0 changed control-C to print an inverted ^C and then a newline.
The original motivation was
> In bash if you type something and press ctrl-c then the content of the line
> is preserved and the cursor is moved to a new line. In fish the ctrl-c just
> clears the line. For me the behaviour of bash is a bit better, because it
> allows me to type something then press ctrl-c and I have the typed string
> in the log for further reference.
This sounds like a valid use case in some scenarios but I think that most
abandoned commands are noise. After all, the user erased them. Also, now that
we have undo that can be used to get back a limited set of canceled commands.
I believe the original motivation for existing behavior (in other shells) was
that TERM=dumb does not support erasing characters. Similarly, other shells
like to leave behind other artifacts, for example when using tab-completion
or in their interactive menus but we generally don't.
Control-C is the obvious way to quickly clear a multi-line commandline.
IPython does the same. For the other behavior we have Alt-# although that's
probably not very well-known.
Restore the old Control-C behavior of simply clearing the command line.
Our unused __fish_cancel_commandline still prints the ^C. For folks who
have explicitly bound ^C to that, it's probably better to keep the existing
behavior, so let's leave this one.
Previous attempt at #4713 fizzled.
Closes#10213
This function is a hotspot, but it has inefficient codegen:
1. For whatever reason, the chars() iterator of wstr is slower
than that of a slice. Use the slice.
2. Unnecessary overflow checks were preventing vectorization.
Switch to a more optimized implementation.
This improves aliases benchmark time by about 9%.
Since none of the compiles(xxx) calls are to particularly complex code, we can
just use `rsconf` directly to test for the presence of the symbols or headers as
needed.
Note that it seems at least some of the previous detection was not working
correctly; in particular HAVE_PIPE2 was evaluating to false on my WSL install
where pipe2(2) was available (caught because it revealed some compilation errors
in that conditional compilation path after porting).
I kept the cfg names and the tests themselves mostly as-is, though we might want
to change that to conform with the rust convention of lowercase cfg names and
decide whether we want to prefix all these with have_, fish_, or nothing at all.
Also the posix_spawn() test should probably check for the symbol `posix_spawn()`
rather than the header `spawn.h` since we don't use it via the header but rather
via the symbol (but in reality they're almost certainly going to give the same
result).
NB: I only encountered this when rewriting the cfg detection, which means that
the previous detection wasn't correct since I have pipe2 on Linux but didn't run
into this build error before.
I had originally created a safe `set_locale()` wrapper and clippy-disallowed
`libc::setlocale()` but almost all our uses of `libc::setlocale()` are in a loop
where it makes much more sense to just obtain the lock outright then call
`setlocale()` repeatedly rather than lock it in the wrapper function each time.
No need to use cfg_attr and have to worry about syncing the preconditions for
the cfg_attr with the preconditions for where `slice_contains_slice()` is used
in the codebase, just mark it as `allow(unused)` with a comment.
Use Rust for executables
Drops the C++ entry points and restructures the Rust package into a
library and three binary crates.
Renames the fish-rust package to fish.
At least on Ubuntu, "fish_indent" is built before "fish".
Make sure export CURSES_LIBRARY_LIST to all binaries to make sure
that "cached-curses-libnames" is populated.
Closes#10198
Keep running tests serially to avoid breaking assumptions.
I think many of these tests can run in parallel and/or don't need test_init().
Use the safe variant everywhere, to get it done faster.
This would misname `\e\x7F` as "backspace":
bind -k backspace 'do something'
bind \e\x7F 'do something'
because it would check if there was any key *in there*.
This was probably meant for continuous mode, but it simply doesn't
work right. It's preferable to not give a key when one would work over
giving one when it's not correct.
This implements input and input_common FFI pieces in input_ffi.rs, and
simultaneously ports bind.rs. This was done as a single commit because
builtin_bind would have required a substantial amount of work to use the input
ffi.
The existing subsequence search commonly returns false positives.
Support globs, to allow searching for disconnected substrings in a better way.
Closes#10143Closes#10131
The "#[bench]" attribute is not allowed in stable Rust, so keep it behind
a new feature flag. Run on nightly Rust with
$ cargo bench --features=bechmark
test tests::encoding::bench::bench_convert_ascii ... bench: 125,988 ns/iter (+/- 1,128) = 1040 MB/s
Drop support for history file version 1.
ParseExecutionContext no longer contains an OperationContext because in my
first implementation, ParseExecutionContext didn't have interior mutability.
We should probably try to add it back.
Add a few to-do style comments. Search for "todo!" and "PORTING".
Co-authored-by: Xiretza <xiretza@xiretza.xyz>
(complete, wildcard, expand, history, history/file)
Co-authored-by: Henrik Hørlück Berg <36937807+henrikhorluck@users.noreply.github.com>
(builtins/set)
This reduces noise in the upcoming "Port execution" commit.
I accidentally made IoStreams a "class" instead of a "struct". Would be
easy to correct that but this will be deleted soon, so I don't think we care.
These printed "Unknown error while evaluating command substitution".
Now they print something like
```
fish: for: status: cannot overwrite read-only variable
for status in foo; end
^~~~~^
in command substitution
fish: Invalid arguments
echo (for status in foo; end)
^~~~~~~~~~~~~~~~~~~~~~~^
```
for `echo (for status in foo; end)`
This is, of course, still not *great*. Mostly the `fish: Invalid
arguments` is basically entirely redundant.
An alternative is to simply skip the error message, but that requires some
more scaffolding (describe_with_prefix adds some error messages on its
own, so we can't simply say "don't add the prefix if we don't have a
message")
(cherry picked from commit 1b5eec2af6)
* wildcard: Remove file size from the description
We no longer add descriptions for normal file completions, so this was
only ever reached if this was a command completion, and then it was
only added if the file wasn't a regular file... in which case it can't
be an executable.
So this was dead.
* Make possible_link() a maybe
This gives us the full information, not just "no" or "maybe"
* wildcard: Rationalize file/command completions
This keeps the entry_t as long as possible, and asks it, so especially
on systems with working d_type we can get by without a single stat in
most cases.
Then it guts file_get_desc, because that is only used for command
completions - we have been disabling file descriptions for *years*,
and so this is never called there.
That means we have no need to print descriptions about e.g. broken symlinks, because those are not executable.
Put together, what this means is that we, in most cases, only do
an *access(2)* call instead of a stat, because that might be checking
more permissions.
So we have the following constellations:
- If we have d_type:
- We need a stat() for every _symlink_ to get the type (e.g. dir or regular)
(this is for most symlinks, if we want to know if it's a dir or executable)
- We need an access() for every file for executables
- If we do not have d_type:
- We need a stat() for every file
- We need an lstat() for every file if we do descriptions
(i.e. just for command completion)
- We need an access() for every file for executables
As opposed to the current way, where every file gets one lstat whether
with d_type or not, and an additional stat() for links, *and* an
access.
So we go from two syscalls to one for executables.
* Some more comments
* rust link option
* rust remove size
* rust accessovaganza
* Check for .dll first for WSL
This saves quite a few checks if e.g. System32 is in $PATH (which it
is if you inherit windows paths, IIRC).
Note: Our WSL check currently fails for WSL2, where this would
be *more* important because of how abysmal the filesystem performance
on that is.
Given an env like
foo
bar=baz
we would set "foo" to empty due to a typo.
The typo is pointed out by a PORTING comment.
Luckily I don't think we ever hit this case because that would mean our
parent process has a serious bug. Rust's std::env::vars_os() skips env
lines that don't contain a "=" char. This seems like a reasonable behavior
for us too. Do that.
This makes it so expand_intermediate_segment knows about the case
where it's last, only followed by a "/".
When it is, it can do without the file_id for finding links (we don't
resolve the files we get here), which allows us to remove a stat()
call.
This speeds up the case of `...*/` by quite a bit.
If that last component was a directory with 1000 subdirectories we
could skip 1000 stat calls!
One slight weirdness: We refuse to add links to directories that we already visited, even if they are the last component and we don't actually follow them. That means we can't do the fast path here either, but we do know if something is a link (if we get d_type), so it still works in common cases.
This can be bound like `bind \cl clear-screen`, and is, by default
In contrast to the current way it doesn't need the external `clear`
command that was always awkward.
Also it will clear the screen and first draw the old prompt to remove
flicker.
Then it will immediately trigger a repaint, so the prompt will be overwritten.
This fixes the following deadlock. The C++ functions path_get_config and
path_get_data lazily determine paths and then cache those in a C++ static
variable. The path determination requires inspecting the environment stack.
If these functions are first called while the environment stack is locked
(in this case, when fetching the $history variable) we can get a deadlock.
The fix is to call them eagerly during env_init. This can be removed once
the corresponding C++ functions are removed.
This issue caused fish_config to fail to report colors and themes.
Add a test.
This uses "screen.reset_line" to move the cursor without informing the
reader's machinery (because that deals with positions *in the
commandline*), but then only repainted "if needed" - meaning if the
reader thought anything changed.
That could lead to a situation where the cursor stays at column 0
until you do something, e.g. in
```fish
bind -m insert u undo
```
when you press alt+u - because the *escape* calls repaint-mode, which
puts the cursor in column 0, and then the undo doesn't, which keeps it
there.
Of course this binding should also `repaint-mode`, because it changes
the mode.
Some changes might be ergonomic:
1. Make repaint-mode the default if the mode changed (we would need to
skip it for bracketed-paste)
2. Make triggering the repaint easier - do we need to set
force_exec_prompt_and_repaint to false here as well?
Anyway, this
Fixes#7910
This is a sensible thing to do, and fixes some cases where we're
state-dependent.
E.g. this fixes the case in the pager where some things are bold and
some aren't, because that bolding is (rather awkwardly) implicitly
triggered when we have a background, and so we don't notice we need to
re-do that bolding after we moved to the next line because we think we
still have the same color.
Fixes#9617
This adopts the Rust postfork code, bridging it from C++ exec module.
We use direct function calls for the bridge, rather than cxx/autocxx, so that we
can be sure that no memory allocations or other shenanigans are happening.
This implements the "postfork" code in Rust, including calling fork(),
exec(), and all the bits that have to happen in between. postfork lives
in the fork_exec module.
It is not yet adopted.
- wildcard_match is now closer to the original that is linked in a comment, as
pointer-arithmetic translates very poorly. The act of calling wildcard
patterns wc or wildcard is kinda confusing when wc elsewhere is widechar.
I sometimes find myself doing something like this:
- Look for a commandline that includes "echo" (as an example)
- Type echo, press up a few times
- I can't immediately find what I'm looking for
- Press ctrl-r to open up the history pager
- It uses the current commandline as the search string,
so now I'm looking for "echo foobar"
This makes it so if the search string already is in use, that's what
the history-pager picks as the initial search string.
We don't change anything about compilation-setup, we just immediately jump to
Rust, making the eventual final swap to a Rust entrypoint very easy.
There are some string-usage and format-string differences that are generally
quite messy.
- `libc::setlinebuf` is not available through Rust's libc it appears.
- autocxx fails to generate bindings using `*mut FILE`, instead go through
`void*`
- rust_main needs `parse_util_detect_errors_in_ast`, which is _partially_
ported, instead add FFI interop for C++.
- We need to set the filename if we are sourcing a file
This used to be assigned to the job, but that was removed in
f30ce21aaa.
Since then this was vestigial. It could have technically errored out,
but we should be catching that where we use the actual modes, not here.
Similar to `time`, except that one is more common as a command.
Note that this will also allow `builtin and`, which is somewhat
useless, but then it is also useless outside of a pipeline.
Addition to #9985
This used to print all codepoints outside of the ASCII range (i.e.
above 0x80) in \uXXXX or \UYYYYYYYY notation.
That's quite awkward, considering that this is about keys that are
being pressed, and many keyboards have actual symbols for these on
them - I have an "ö" key, so I would like to use `bind ö` and not
`bind \u00F6`. So we go by iswgraph.
On a slightly different note, `\e` was written as `\c[ (or \e)`. I do
not believe anyone really uses `\c[` (the `[` would need to
be escaped!), and it's confusing and unnecessary to even mention that.
This allows e.g. `foo | command time`, while still rejecting `foo | time`.
(this should really be done in the ast itself, but tbh most of
parse_util kinda should)
Fixes#9985
These are both clearly behind early returns, there is no need to check it again.
This isn't a case where we're doing logic gymnastics to see that it
can't be run without no_exec() being handled, this is
```c++
if (no_exec()) return;
// ..
// ..
// ..
if (no_exec()) foo;
```
We have already run waccess with X_OK. We already *know* the file is
executable.
There is no reason to check again.
Restores some of the speedup from the fast_waccess hack that was
removed to fix#9699.
- Add test to verify piped string replace exit code
Ensure fields parsing error messages are the same.
Note: C++ relied upon the value of the parsed value even when `errno` was set,
that is defined behaviour we should not rely on, and cannot easilt be replicated from Rust.
Therefore the Rust version will change the following error behaviour from:
```shell
> string split --fields=a "" abc
string split: Invalid fields value 'a'
> string split --fields=1a "" abc
string split: 1a: invalid integer
```
To:
```shell
> string split --fields=a "" abc
string split: a: invalid integer
> string split --fields=1a "" abc
string split: 1a: invalid integer
```
This adopts the new function store, replacing the C++ version.
It also reimplements builtin_function in Rust, as these was too coupled to
the function store to handle in a separate commit.
We could end up overflowing if we print out something that's a multiple of the
chunk size, which would then finish printing in the chunk-printing, but not
break out early.
This confirmed that a file existed via access(file, F_OK).
But we already *know* that it does because this is the expansion for
the "trailing slash" - by definition all wildcard components up to
here have already been checked.
And it's not checking for directoryness either because it does F_OK.
This will remove one `access()` per result, which will cut the number
of syscalls needed for a glob that ends in a "/" in half.
This brings us on-par with e.g. `ls` (which uses statx while we use
newfstatat, but that should have about the same results)
Fixes#9891.
Remove the following C++ functions/methods, which have no callers:
fallback.cpp:
- wcstod_l
proc.cpp:
- job_t::get_processes
wutil.cpp:
- fish_wcstoll
- fish_wcstoull
Also drop unused configure checks/defines:
- HAVE_WCSTOD_L
- HAVE_USELOCALE
Remove the following C++ functions/methods, which have all been ported to Rust and no longer have any callers in C++:
common.cpp:
- assert_is_locked/ASSERT_IS_LOCKED
path.cpp:
- path_make_canonical
wutil.cpp:
- wreadlink
- fish_iswgraph
- file_id_t::older_than
This makes `fish -c begin` fail with a status of 127 - it already
printed a syntax error so that was weird. (127 was the status for
syntax errors when piping to fish, so we stay consistent with that)
We allow multiple `-c` commands, and this will return the regular
status if the last `-c` succeeded.
This is fundamentally an extremely weird situation but this is the
simple targeted fix - we did nothing, unsuccessfully, so we should
fail.
Things to consider in future:
1. Return something better than 127 - that's the status for "unknown
command"!
2. Fail after a `-c` failed, potentially even checking all of them
before executing the first?
Fixes#9888
This also cleans up and removes unnecessary usage of FFI-oriented `feature_metadata_t`,
which is only used from Rust code after `builtins/status` was ported.
Note this is slightly incomplete - the FD is not moved into the parser, and so
will be freed at the end of each directory change. The FD saved in the parser is
never actually used in existing code, so this doesn't break anything, but will
need to be corrected once the parser is ported.
Prior to this commit, FLOG used the ffi bridge to get the output fd. Invert
this: have fish set the output fd within main. This allows FLOG to be used in
pure Rust tests.
After accidentally running a command that includes a pasted password, I want
to delete command from history. Today we need to recall or type (part of)
that command and type "history delete". Let's maybe add a shortcut to do
this from the history pager.
The current shortcut is Shift+Delete. I don't think that's very discoverable,
maybe we should use Delete instead (but only if the cursor is at the end of
the commandline, otherwise delete a char).
Closes#9454
The tentative binding for the upcoming "history-pager-delete" is
bind -k sdc history-pager-delete or backward-delete-char
When Shift+Delete is pressed while the history pager is active,
"history-pager-delete" succeeds. In this case, the "or" needs to kick the
"backward-delete-char" out of the input queue.
After doing so, it continues reading, but interprets the input as
single-char binding. This breaks when the next key emits a multi-char sequence,
like the arrow keys.
Fix this by reading a full sequence, which means we need to run "read_char()"
instead of "read_ch()" (confusing, right?).
I'm still working on writing a test. Somehow this only reproduces in the
history pager where Shift+Delete followed by down arrow emits "[B" (since
we swallowed the leading escape char). Confusingly, it doesn't do that in
the commandline or the completion search field.
Prior to this change, parser_t exposed an environment_t, and Rust had to go
through that. But because we have implemented Environment in Rust, it is
better to just expose the native Environment from parser_t. Make that
change and update call sites.
We can't just call the Rust version of `fish_setlocale()` without also either
calling the C++ version of `fish_setlocale()` or removing all `src/complete.cpp`
variables that are initialized and aliasing them to their new rust counterparts.
Since we're not interested in keeping the C++ code around, just call the C++
version of the function via ffi until we don't have *any* C++ code referencing
`src/common.h` at all.
Note that *not* doing this and then calling the rust version of
`fish_setlocale()` instead of the C++ version will cause errant behavior and
random segfaults as the C++ code will try to read and use uninitialized values
(including uninitialized pointers) that have only had their rust counterparts
init.
Either add rust wrappers for C++ functions called via ffi or port some pure code
from C++ to rust to provide support for the upcoming `env_dispatch` rewrite.
The global variables are moved (not copied) from C++ to rust and exported as
extern C integers. On the rust side they are accessed only with atomic semantics
but regular int access is preserved from the C++ side (until that code is also
ported).
Historically fish has used the functions `fish_wcstol`, `fish_wcstoi`, and
`fish_wcstoul` (and some long long variants) for most integer conversions.
These have semantics that are deliberately different from the libc
functions, such as consuming trailing whitespace, and disallowing `-` in
unsigned versions.
fish has started to drift away from these semantics; some divergence from
C++ has crept in.
Rename the existing `fish_wcs*` functions in Rust to remove the fish
prefix, to express that they attempt to mirror libc semantics; then
introduce `fish_` wrappers which are ported from C++. Also fix some
miscellaneous bugs which have crept in, such as missing range checks.
This implements the primary environment stack, and other environments such
as the null and snapshot environments, in Rust. These are used to implement
the push and pop from block scoped commands such as `for` and `begin`, and
also function calls.
owning_null_terminated_array is used for environment variables, where we need to
provide envp for child processes. This switches the implementation from C++ to
Rust.
We retain the C++ owning_null_terminated_array_t; it simply wraps the Rust
version now.
init_curses() is/can be called more than once, in which case the previous
ncurses terminal state is leaked and a new one is allocated.
`del_curterm(cur_term)` is supposed to be called prior to calling `setupterm()`
if `setupterm()` is being used to reinit the default `TERMINAL *cur_term`.
The new asan exit handlers are called to get proper ASAN leak reports (as
calling _exit(0) skips the LSAN reporting stage and exits with success every
time).
They are no-ops when not compiled for ASAN.
This ports some signal setup and handling bits to Rust.
The signal handling machinery requires walking over the list of known signals;
that's not supported by the Signal type. Rather than duplicate the list of
signals yet again, switch back to a table, as we had in C++.
This also adds two further pieces which were neglected by the Signal struct:
1. Localize signal descriptions
2. Support for integers as the signal name
This allows the rust code to free up C++ resources allocated for a callback even
when the callback isn't executed (as opposed to requiring the callback to run
and at the end of the callback cleaning up all allocated resources).
Also add type-erased destructor registration to callback_t. This allows for
freeing variables allocated by the callback for debounce_t's
perform_with_callback() that don't end up having their completion called due to
a timeout.
Largely routine but for the trampolines in iothread.h and iothread.cpp which
were a real PITA to get correct w/ all their variants.
Integration is complete with all old code ripped out and the tests using the
rust version of the code.
There are many places where we want to treat a missing variable the same as
a variable with an empty value.
In C++ we handle this by branching on maybe_t<env_var_t>::missing_or_empty().
If it returns false, we go on to access maybe_t<env_var_t>::value() aka
operator*.
In Rust, Environment::get() will return an Option<EnvVar>.
We could define a MissingOrEmpty trait and implement it for Option<EnvVar>.
However that will still leave us with ugly calls to Option::unwrap()
(by convention Rust does use shorthands like *).
Let's add a variable getter that returns none for empty variables.
Except for the indent visitor bits.
Tests for parse_util_detect_errors* are not ported yet because they depend
on expand.h (and operation_context.h which depends on env.h).
The translation is fairly direct though it adds some duplication, for example
there are multiple "match" statements that mimic function overloading.
Rust has no overloading, and we cannot have generic methods in the Node trait
(due to a Rust limitation, the error is like "cannot be made into an object")
so we include the type name in method names.
Give clients like "indent_visitor_t" a Rust companion ("IndentVisitor")
that takes care of the AST traversal while the AST consumption remains
in C++ for now. In future, "IndentVisitor" should absorb the entirety of
"indent_visitor_t". This pattern requires that "fish_indent" be exposed
includable header to the CXX bridge.
Alternatively, we could define FFI wrappers for recursive AST traversal.
Rust requires we separate the AST visitors for "mut" and "const"
scenarios. Take this opportunity to concretize both visitors:
The only client that requires mutable access is the populator. To match the
structure of the C++ populator which makes heavy use of function overloading,
we need to add a bunch of functions to the trait. Since there is no other
mutable visit, this seems acceptable.
The "const" visitors never use "will_visit_fields_of()" or
"did_visit_fields_of()", so remove them (though this is debatable).
Like in the C++ implementation, the AST nodes themselves are largely defined
via macros. Union fields like "Statement" and "ArgumentOrRedirection"
do currently not use macros but may in future.
This commit also introduces a precedent for a type that is defined in one
CXX bridge and used in another one - "ParseErrorList". To make this work
we need to manually define "ExternType".
There is one annoyance with CXX: functions that take explicit lifetime
parameters require to be marked as unsafe. This makes little sense
because functions that return `&Foo` with implicit lifetime can be
misused the same way on the C++ side.
One notable change is that we cannot directly port "find_block_open_keyword()"
(which is used to compute an error) because it relies on the stack of visited
nodes. We cannot modify a stack of node references while we do the "mut"
walk. Happily, an idiomatic solution is easy: we can tell the AST visitor
to backtrack to the parent node and create the error there.
Since "node_t::accept_base" is no longer a template we don't need the
"node_visitation_t" trampoline anymore.
The added copying at the FFI boundary makes things slower (memcpy dominates
the profile) but it's not unusable, which is good news:
$ hyperfine ./fish.{old,new}" -c 'source ../share/completions/git.fish'"
Benchmark 1: ./fish.old -c 'source ../share/completions/git.fish'
Time (mean ± σ): 195.5 ms ± 2.9 ms [User: 190.1 ms, System: 4.4 ms]
Range (min … max): 193.2 ms … 205.1 ms 15 runs
Benchmark 2: ./fish.new -c 'source ../share/completions/git.fish'
Time (mean ± σ): 677.5 ms ± 62.0 ms [User: 665.4 ms, System: 10.0 ms]
Range (min … max): 611.7 ms … 805.5 ms 10 runs
Summary
'./fish.old -c 'source ../share/completions/git.fish'' ran
3.47 ± 0.32 times faster than './fish.new -c 'source ../share/completions/git.fish''
Leftovers:
- Enum variants are still snakecase; I didn't get around to changing this yet.
- "ast_type_to_string()" still returns a snakecase name. This could be
changed since it's not user visible.
This is basically a subset of type, so we might as well.
To be clear this is `command -s` and friends, if you do `command grep` that's
handled as a keyword.
One issue here is that we can't get "one path or not" because I don't
know how to translate a maybe_t? Do we need to make it a shared_ptr instead?
Vi visual mode selection highlighting behaves unexpectedly when the selection
foreground and background in the highlight spec don't match. The following
unexpected behaviors are:
* The foreground color is not being applied when defined by the
`fish_color_selection` variable.
* `set_color` options (e.g., `--bold`) would not be applied under the cursor
when selection begins in the middle of the command line or when the cursor
moves forward after visually selecting text backward.
With this change, visual selection respects the foreground color and any
`set_color` options are applied consistently regardless of where visual
selection begins and the position of the cursor during selection.
Most of it is duplicated, hence untested.
Functions like mbrtowc are not exposed by the libc crate, so declare them
ourselves.
Since we don't know the definition of C macros, add two big hacks to make
this work:
1. Replace MB_LEN_MAX and mbstate_t with values (resp types) that should
be large enough for any implementation.
2. Detect the definition of MB_CUR_MAX in the build script. This requires
more changes for each new libc. We could also use this approach for 1.
Additionally, this commit brings a small behavior change to
read_unquoted_escape(): we cannot decode surrogate code points like \UDE01
into a Rust char, so use � (\UFFFD, replacement character) instead.
Previously, we added such code points to a wcstring; looks like they were
ignored when printed.
wcs2string converts a wide string to a narrow one. The result is
null-terminated and may also contain interior null-characters.
std::string allows this.
Rust's null-terminated string, CString, does not like interior null-characters.
This means we will need to use Vec<u8> or OsString for the places where we
use interior null-characters.
On the other hand, we want to use CString for places that require a
null-terminator, because other Rust types don't guarantee the null-terminator.
Turns out there is basically no overlap between the two use cases, so make
it two functions. Their equivalents in Rust will have the same name, so
we'll only need to adjust the type when porting.
This can be triggered on linux with:
```js
import { spawn } from 'child_process';
const shell = spawn('/home/alfa/dev/fish-shell/build-c++/fish', []);
```
Under node 19.8.1.
*No clue* how that happens, but since this is a workaround we shall
skip it.
Another from the "why are we asserting instead of doing something
sensible" department.
The alternative is to make exit() and return() compute their own exit
code, but tbh I don't want any *other* builtin to hit this either?
Fixes#9659
This shows some of the ugliness of the rust borrow checker when it comes to
safely implementing any sort of recursive access and the need to be overly
explicit about which types are actually used across threads and which aren't.
We're forced to use an `Arc` for `ItemMaker` (née `item_maker_t`) because
there's no other way to make it clear that its lifetime will last longer than
the FdMonitor's. But once we've created an `Arc<T>` we can't call
`Arc::get_mut()` to get an `&mut T` once we've created even a single weak
reference to the Arc (because that weak ref could be upgraded to a strong ref at
any time). This means we need to finish configuring any non-atomic properties
(such as `ItemMaker::always_exit`) before we initialize the callback (which
needs an `Arc<ItemMaker>` to do its thing).
Because rust doesn't like self-referential types and because of the fact that we
now need to create both the `ItemMaker` and the `FdMonitorItem` separately
before we set the callback (at which point it becomes impossible to get a
mutable reference to the `ItemMaker`), `ItemMaker::item` is dropped from the
struct and we instead have the "constructor" for `ItemMaker` take a reference to
an `FdMonitor` instance and directly add itself to the monitor's set, meaning we
don't need to move the item out of the `ItemMaker` in order to add it to the
`FdMonitor` set later.
CXX does not allow generic types like maybe_t. When porting a C++ function
that returns maybe_t to Rust, we return std::unique_ptr instead. Let's make
the transition more seamless by allowing to convert back to maybe_t implicitly.
* wutil: Rewrite `wrealpath` in Rust
* Reduce use of FFI types in `wrealpath`
* Addressed PR comments regarding allocation
* Replace let binding assignment with regular comparison
More ugliness with types that cxx bridge can't recognize as being POD. Using
pointers to get/set `termios` values with an assert to make sure we're using
identical definitions on both sides (in cpp from the system headers and in rust
from the libc crate as exported).
I don't know why cxx bridge doesn't allow `SharedPtr<OpaqueRustType>` but we can
work around it in C++ by converting a `Box<T>` to a `shared_ptr<T>` then convert
it back when it needs to be destructed. I can't find a clean way of doing it
from the cxx bridge wrapper so for now it needs to be done manually in the C++
code.
Types/values that are drop-in ready over ffi are renamed to match the old cpp
names but for types that now differ due to ffi difficulties I've left the `_ffi`
in the function names to indicate that this isn't the "correct" way of using the
types/methods.
The way cxx bridge works, it doesn't recognize any types from another module as
being shared cxx bridge types with generations native to both C++ and Rust,
meaning every module that was going to use function pointers would have to
define its own `c_void` type (because cxx bridge doesn't recognize any of
libc::c_void, std::ffi::c_void, or autocxx::c_void).
FFI on other platforms has long used the equivalent of `uint8_t *` as an
alternative to `void *` for code where `void` was not available or was
undesirable for some reason. We can join the club - this way we can always use
`* {const|mut} u8` in our rust code and `uint8_t *` in our C++ code to pass
around parameters or values over the C abi.
I needed to rename some types already ported to rust so they don't clash with
their still-extant cpp counterparts. Helper ffi functions added to avoid needing
to dynamically allocate an FdMonitorItem for every fd (we use dozens per basic
prompt).
I ported some functions from cpp to rust that are used only in the backend but
without removing their existing cpp counterparts so cpp code can continue to use
their version of them (`wperror` and `make_detached_pthread`).
I ran into issues porting line-by-line logic because rust inverts the behavior
of `std::remove_if(..)` by making it (basically) `Vec::retain_if(..)` so I
replaced bools with an explict enum to make everything clearer.
I'll port the cpp tests for this separately, for now they're using ffi.
Porting closures was ugly. It's nothing hard, but it's very ugly as now each
capturing lambda has been changed into an explicit struct that contains its
parameters (that needs to be dynamically allocated), a standalone callback
(member) function to replace the lambda contents, and a separate trampoline
function to call it from rust over the shared C abi (not really relevant to
x86_64 w/ its single calling convention but probably needed on other platforms).
I don't like that `fd_monitor.rs` has its own `c_void`. I couldn't find a way to
move that to `ffi.rs` but still get cxx bridge to consider it a shared POD.
Every time I moved it to a different module, it would consider it to be an
opaque rust type instead. I worry this means we're going to have multiple
`c_void1`, `c_void2`, etc. types as we continue to port code to use function
pointers.
Also, rust treats raw pointers as foreign so you can't do `impl Send for * const
Foo` even if `Foo` is from the same module. That necessitated a wrapper type
(`void_ptr`) that implements `Send` and `Sync` so we can move stuff between
threads.
The code in fd_monitor_t has been split into two objects, one that is used by
the caller and a separate one associated with the background thread (this is
made nice and clean by rust's ownership model). Objects not needed under the
lock (i.e. accessed by the background thread exclusively) were moved to the
separate `BackgroundFdMonitor` type.
Keeps the location of original function definition, and also stores
where it was copied. `functions` and `type` show both locations,
instead of none. It also retains the line numbers in the stack trace.
By default, fish does not complete files that have leading dots, unless the
wildcard itself has a leading dot. However this also affected completions;
for example `git add` would not offer `.gitlab-ci.yml` because it has a
leading dot.
Relax this for custom completions. Default file expansion still
suppresses leading dots, but now custom completions can create
leading-dot completions and they will be offered.
Fixes#3707.
When we draw the prompt, we move the cursor to the actual
position *we* think it is by issuing a carriage return (via
`move(0,0)`), and then going forward until we hit the spot.
This helps when the terminal and fish disagree on the width of the
prompt, because we are now definitely in the correct place, so we can
only overwrite a bit of the prompt (if it renders longer than we
expected) or leave space after the prompt. Both of these are benign in
comparison to staircase effects we would otherwise get.
Unfortunately, midnight commander ("mc") tries to extract the last
line of the prompt, and does so in a way that is overly naive - it
resets everything to 0 when it sees a `\r`, and doesn't account for
cursor movement. In effect it's playing a terminal, but not committing
to the bit.
Since this has been an open request in mc for quite a while, we hack
around it, by checking the $MC_SID environment variable.
If we see it, we skip the clearing. We end up most likely doing
relative movement from where we think we are, and in most cases it
should be *fine*.
This is early work but I guess there's no harm in pushing it?
Some thoughts on the conventions:
Types that live only inside Rust follow Rust naming convention
("FeatureMetadata").
Types that live on both sides of the language boundary follow the existing
naming ("feature_flag_t").
The alternative is to define a type alias ("using feature_flag_t =
rust::FeatureFlag") but that doesn't seem to be supported in "[cxx::bridge]"
blocks. We could put it in a header ("future_feature_flags.h").
"feature_metadata_t" is a variant of "FeatureMetadata" that can cross
the language boundary. This has the advantage that we can avoid tainting
"FeatureMetadata" with "CxxString" and such. This is an experimental approach,
probably not what we should do in general.
The initial port of feature flags requires a global initialization. Since
fish_indent accesses feature flags, let's make sure to initialize them here.
In future, we can stop initializing things fish_indent doesn't need (like
the topic monitor) but that's no big deal. Global initialization should
always be a benign addition.
The original implementation without the test took me 3 hours (first time
seriously looking into this)
The functions take "wcharz_t" for smooth integration with existing C++ callers.
This is at the expense of Rust callers, which would prefer "&wstr". Would be
nice to declare a function parameter that accepts both but I don't think
that really works since "wcharz_t" drops the lifetime annotation.
This works around an autocxx limitations where different types cannot
have the same name even if they live in different namespace.
ast::job_t conflicts with job_t.
This translated ctrl-k to "\v", which is a "vertical tab", and ctrl-l
to "\f" and ctrl-g to "\a".
There is no "vertical tab" or "alarm" or "\f" *key*, so these
shouldn't be translated. Just drop these and call them `\ck` and such.
(vertical tab specifically is utterly useless and I would be okay with
dropping it entirely, I have never seen it used anywhere)
Commit 3b30d92b6 (Commit transient edit when closing pager, 2022-08-31)
inadvertently introduced two regressions to history search:
1. It made Escape keeps the selected history entry,
instead of restoring the commandline before history search.
2. It made history search commands add undo entries.
Fix both of this issues.
Inadvertently broken in a2d816710f,
this made `cd .` no longer offer `cd ../` (same for general file completions
like `ls .`, which only offers dotfiles)
This meant we didn't actually do our weird en/decoding scheme for e.g.
a C locale, which meant that, when you then switch to a proper locale
the previous variables were broken.
I don't know how to test this automatically - none of my attempts seem
to ever *fail* with the old code, here's what you'd do manually:
- Run fish with an actual C locale (LC_ALL=C
fish_allow_singlebyte_locale=1 fish)
- `set -gx foo 💩`
- `set -e LC_ALL`
- `echo $foo` outputs "💩" if it works and "ð⏎" if it's broken.
Fixes#2613
This means cleaning out old universal variables is now just:
```fish
abbr --erase (abbr --list)
```
which makes upgrading much easier.
Note that this erases the currently defined variable and/or any
universal. It doesn't stop at the former because that makes it *easy*
to remove the universals (no running `abbr --erase` twice), and it
doesn't care about globals because, well, they would be gone on
restart anyway.
Fixes#9468.
Like I mentioned in #9089, 12 entries is a bit few.
So, instead, we do like we do for completions before disclosing and
pick half the screen (but at least X, in this case 12).
This avoids filling the entire screen, and will avoid an unsightly "X
more entries" (which requires scrolling down to fully disclose)
because it matches what the pager does.
Note: For multiline commands we can be pushed further upwards, and in
case of a multi-column layout we could fit more lines. That would
require asking the pager to fit as many as possible and give us back
the index of the last matching entry and rewinding the history search.
That's gonna be left as an exercise for later if it turns out to be necessary.
This now means `abbr --add` has two modes:
```fish
abbr --add name --function foo --regex regex
```
```fish
abbr --add name --regex regex replacement
```
This is because `--function` was seen to be confusing as a boolean flag.
Unfortunately print_hints was true *by default* - so for all builtins
that didn't pass it it would now be false instead.
This resulted in the trailer missing, which includes the line number
and context. So if you ran a script that includes `bind -M` the error
message would now just be "bind: -M: option requires an argument",
with no indication as to where.
This reverts commit 8a50d47a46.
The print_hints variable was always false, so just remove it.
This caused a cascade of other changes where the parser_t variable
becomes unused, so remove it from the call sites.
No functional change expected here.
When we insert characters that don't yet have highlighting, we use the
highlighting to the left, unless there is nothing to our left. The logic to
check if we are the leftmost character uses an overly loose comparison. Let's
make it more specific.
No functional change.
When there are multiple event handlers for a single event, we would print
the same log statement twice. Let's add the function name to make this
less confusing.
This would print
```
abbr -a -- dotdot --regex ^\\.\\.+\$ --function multicd
```
which expands "dotdot" to "--regex ^\\.\\.+\$...".
Instead, we move the name to right before the replacement, and move
the `--` before that:
```
abbr -a --regex ^\\.\\.+\$ --function -- dotdot multicd
```
It might be possible to improve that, but this at least round-trips.
Historical behavior is to stop option parsing at the first non-option argument.
Since we have added more options, it seemed impractical to keep that behavior.
However people are using options in their abbr expansions ("abbr e emacs
-nw"). To support this, we ignore options. However, we only ignore them
if they are not valid "abbr" options. Let's ignore all options in the
expansion definition, which is a small price to pay to keep most existing
configurations working.
Fixes#9410
This does not fix other cases which used to work, like
abbr x -unknown
Those are hopefully not used by anyone, so I don't think we need to maintain
support for that.
Enhances abbreviations with extra features
- global abbreviations
- trigger on regex match as alternative to literal match
- the ability to expand abbreviations with a user-defined function
- the ability to set cursor position after expansion
Also default the marker to '%'. So you may write:
abbr -a L --position anywhere --set-cursor "% | less"
or set an explicit marker:
abbr -a L --position anywhere --set-cursor=! "! | less"
This renames abbreviation triggers from `--trigger-on entry` and
`--trigger-on exec` to `--on-space` and `--on-enter`. These names are less
precise, as abbreviations trigger on any character that terminates a word
or any key binding that triggers exec, but they're also more human friendly
and that's a better tradeoff.
set-cursor enables abbreviations to specify the cursor location after
expansion, by passing in a string which is expected to be found in the
expansion. For example you may create an abbreviation like `L!`:
abbr L! --position anywhere --set-cursor ! "! | less"
and the cursor will be positioned where the "!" is after expansion, with
the "| less" appearing to its right.
This adds support for the `--function` option of abbreviations, so that the
expansion of an abbreviation may be generated dynamically via a fish
function.
Prior to this change, abbreviations were stored as fish variables, often
universal. However we intend to add additional features to abbreviations
which would be very awkward to shoe-horn into variables.
Re-implement abbreviations using a builtin, managing them internally.
Existing abbreviations stored in universal variables are still imported,
for compatibility. However new abbreviations will need to be added to a
function. A follow-up commit will add it.
Now that abbr is a built-in, remove the abbr function; but leave the
abbr.fish file so that stale files from past installs do not override
the abbr builtin.
This allows adjusting a pattern string so that it matches an entire
string, by wrapping the regex in a group like ^(?:...)$
This is a workaround for the fact that PCRE2_ENDANCHORED is unavailable
on PCRE2 prior to 2017, so we have to adjust the pattern instead.
Also introduce an overload of match() which creates its own
match_data_t.
We have had multiple crashes for relative CDPATH entries. Commit 5e274066e
(Always return absolute path in path_get_cdpath, 2019-10-17) tried to fix
all of them but it failed to do justice to its title. Let's fix this to
actually return absolute paths, always. Take care to to normalize the path
because it is used for autosuggestions. The normalization is mostly relevant
for CDPATH=. (the default) but it doesn't hurt others.
Closes#9407
wopterr was a feature to allow wgetopt to emit error messages; but we do
not use this and never will. Remove its support. No functional change
expected here.
We wrongly highlight this as prefix when actually the trailing slash should
invalidate it. Turns out path normalization drops the slash, so let's
sidestep that.
Fixes#9394
The "flag" field enables an option to discover which flag it was invoked
with. However in practice none of our options use multiple flags so this
parameter was always nullptr. Remove it and fix up all the builtins to
stop passing this.
No functional change here.
I believe this should be identical to the previous code and handle the same
cases (I'm guessing going by the comment that this came from a C codebase
without `bool` types).
The problem with the previous code is that it tripped up the `clangd` analyzer
into thinking `assert()` expressions can/should be simplified via DeMorgan's to
improve readability (because it was seeing the fully expanded macro).
The tty_ownership test was sometimes failing. In this test,
`fish_test_helper` creates a child and transfers the tty to it,
"abandoning" the tty. In some cases, the child was running before the
parent; the child claims the tty. When the parent tries to transfer it to
the child, it get SIGTTIN and stops. Fix this by ignoring SIGTTIN and
SIGTTOU.
This only affects macOS and BSDs.
The stack overflow tests are too slow without this.
This is because the tests are essentially quadratic: with 500 jobs, and
each job attempts to reap all jobs.
Inside a comment we offer plain file completions (or command completions if
the comment is in command position). However these completions are broken
because they don't consider any of the surrounding characters. For example
with a command line
echo # comment
^ cursor
we suggest file completions and insert them as
echo # comsomefile ment
Providing completions inside comments does not seem useful and it can be
misleading. Let's remove the completions; this should communicate better that
we are in a free-form comment that's not subject to fish syntax.
Closes#9320
When unsetting, the scope indicates the scope that was *removed* not
set, so the warning is incorrectly triggered. If anything, the confusion
is now removed or we emit a warning that the variable is still present
in another scope (but don't do that!).
Closes#9338.
This fixes#9321
IEEE Std 1003.1-2017 Issue 6 added optional error condition
[EINVAL] for if no conversion could be performed.
Switch back to wcstoimax/wcstoumax: do not work around the old FreeBSD
8 issue.
Add a test for printf '%d %d' 1 2 3
This addresses a long-standing TODO where `complete -C` output isn't
deduplicated.
With this patch, the same deduplication and sort procedure that is run on actual
pager completions is also executed for `complete -C` completions (with a `-C`
payload specified).
This makes it possible to use `complete -C` to test what completions will
actually be generated by the completions pager instead of it displaying
something completely divorced from reality, improving the productivity of fish
completions developers.
Note that completions that wouldn't be shown in the pager are also omitted from
the results, e.g. `test/buildroot/` and `test/fish_expand_test/` are omitted
from the check matches in `checks/complete_directories.fish` because even if
they were generated, the pager wouldn't have shown them. This again makes
reasoning about and debugging completions much easier and more sane.
This reverts commit 1c92d4c5db and
reintroduces support for trivially copyable `maybe_t` impls but with a
GCC version check to disable the optimization for GNU GCC compiler
versions 9 and below.
GCC 8.3.0 armhf builds seem to have a problem with the trivially
copyable `maybe_t` impl that introduces odd heisenbugs that cause the
tests to fail. GDB reveals that `maybe_t` function parameters received
in the callee differ from what was passed-in by the caller.
This behavior appears to be (but has not been confirmed as) a
platform-specific compiler bug. Under the same system (32-bit Debian 10
armhf), compiling with clang 7.0.1 does not result in any bugs and
causes all the tests to pass while compiling with GCC 10.2 under 32-bit
Debian 11 armhf also doesn't run into any problems, so just expand the
existing GCC version check that gates support for trivially copyable
`maybe_t` impls to encompass both the troublesome GCC 8 version and the
untested GCC 9 version.
This reverts commit 9d303a74e3.
This reverts commit 0305c842e6.
9d303a7 broke 32-bit armhf builds for unknown reasons, specifically in
settings where a trivial copy of `maybe_t<int>` was performed. A caller
would pass a literal int in the place of a `maybe_t<int>` parameter and
the callee would see a populated `maybe_t` but with a value of `0`
rather than the actual value that was passed in. It was too painful to
debug to a resolution under qemu.
Fixes ommitted newline char shown after complete -n'(foo)'
Also axes the 'contains syntax errors' line before the error.
Update tests
before
> complete -n'(foo)'
complete: Condition '(foo)' contained a syntax error
complete: Command substitutions not allowed⏎
after
> complete -n'(foo)'
complete: -n '(foo)': command substitutions not allowed here
This is a salvage of the "no functional changes" part of #9221, and cherry-picks
storing completion entries in a vector instead of a linked list. The legacy
"reverse intuitive" group ordering is kept by iterating in reverse order.
Tests pass but don't actually cover group order, which needs another test.
Makes it possible to retrieve the currently executing command line as
opposed to the currently executing command (`status current-command`).
Closes#8905.
There should be no functional changes in this commit.
The global variable `$_` set in the parser variables by `reader.cpp` and
read by the `status` builtin was deprecated in fish 2.0 but kept around
internally because there's no good way to store/share/forward parser
variables.
A new enum is added that identifies the status variable and they are
stored in a private array in the parser. There is no need for
synchronization because they are only set during job init and never
thereafter. This is currently asserted via ASSERT_IS_MAIN_THREAD() but
that assert can be dropped in the interest of making the parser possible
to clone and use from worker threads.
The old `$_` global variable is still kept for backwards compatibility,
though it will be dropped in a future release.
As the user is typing an argument, fish continually checks if the input is
the prefix of a valid file path. If yes, the input is underlined.
The same prefix-logic is used for all tokens on the command line, even for
"finished" tokens. This means we highlight any token that happens to be
a prefix of a valid file path. We actually want this to only apply to the
token that the user is currently typing.
Let's use the prefix-logic only for tokens adjacent to the cursor. This should
better match user expectations (and reduce IO traffic). I don't think this is
the perfect criteria but I don't know how else we can determine if a token is
"unfinished".
When visiting the "cd" node, we mark invalid paths as error, but don't
underline valid paths. This works fine most of the time because we later
underline paths (for any command, not just "cd").
However the latter check fails to honor CDPATH. Let's correct that, which
also allows to simplify the logic.
The next commit wants to move the "Underline every valid path" logic into the
visit() methods. The logic currently polls the cancel checker before checking
each path. If that's valid, it should probably have the same behavior inside
visit(). Since we currently can't cancel an AST-visitation, the next best
thing seems to suspend all IO operations, the rest should be very fast anyway.
I'm not sure if the motivation is strong enough; a conceivable alternative
would be to stop using the cancel checker altogether for highlighting.
When passing a value of type maybe_t<size_t>, clangd complains:
Parameter 'cursor' is passed by value and only copied once; consider
moving it to avoid unnecessary copies (fix available)
We get this warning because maybe_t<size_t> is not trivially copyable
because it has a user-defined destructor and copy-constructor. Let's remove
them if the contained type is trivially copyable, to avoid such warnings.
No functional change.
The destructor is equivalent to the compiler-generated one. The user-defined
destructor prevents maybe_t<size_t> from bearing the predicate "trivially
copyable". Let's remove it. No functional change.
It seems to have originally been thought that the only possible way a stack
overflow could happen is via function calls, but there are other possibilities.
Issue #9302 reports how `eval` can be abused to recursively execute a string
substitution ad infinitum, triggering a stack overflow in fish.
This patch extends the stack overflow check to also check the current
`eval_level` against a new constant `FISH_MAX_EVAL_DEPTH`, currently set to a
conservative but hopefully still fair limit of 500. For future reference, with
the default stack size for the main/foreground thread of 8 MiB, we actually have
room for a stack depth around 2800, but that's only with extremely minimal state
stored in each stack frame.
I'm not entirely sure why we don't check `eval_depth` regardless of block type;
it can't be for performance reasons since it's just a simple integer comparison
- and a ridiculously easily one for the branch predictor handle, at that - but
maybe it's to try and support non-recursive nested execution blocks of greater
than `FISH_MAX_STACK_DEPTH`? But even without recursion, the stack can still
overflow so may be we should just bump the limit up some (to 500 like the new
`FISH_MAX_EVAL_DEPTH`?) and check it all the time?
Closes#9302.
A `block_t` instance is allocated for each live block type in memory when
executing a script or snippet of fish code. While many of the items in a
`block_t` class are specific to a particular type of block, the overhead of
`maybe_t<event_t>` that's unused except in the relatively extremely rare case of
an event block is more significant than the rest, given that 88 out of the 216
bytes of a `block_t` are set aside for this field that is rarely used.
This patch reorders the `block_t` members by order of decreasing alignment,
bringing down the size to 208 bytes, then changes `maybe_t<event_t>` to
`shared_ptr<event_t>` instead of allocating room for the event on the stack.
This brings down the runtime memory size of a `block_t` to 136 bytes for a 37%
reduction in size.
I would like to investigate using inheritance and virtual methods to have a
`block_t` only include the values that actually make sense for the block rather
than always allocating some sort of storage for them and then only sometimes
using it. In addition to further reducing the memory, I think this could also be
a safer and saner approach overall, as it would make it very clear when and
where we can expect each block_type_type_t-dependent member to be present and
hold a value.
This is a false positive as a result of disabling TLS support in LSAN due to an
incompatibility with newer versions of glibc.
Also remove the older workaround (because it didn't work).
When there are multiple screens worth of output and `history` is writing to the
pager, pressing Ctrl-C at the end of a screen doesn't exit the pager (`q` is
needed for that) but previously caused fish to emit an error ("write:
Interrupted system call) until we starting silently handling SIGINT in
`fd_output_stream_t::append()`.
This patch makes `history` detect when the `append()` call returns with an error
and causes it to end early rather than repeatedly trying (and failing) to write
to the output stream.
If EINTR caused by SIGINT is encountered while writing to the
`fd_output_stream_t` output fd, mark the output stream as errored and return
false to the caller but do not visibly complain.
Addressing the outstanding TODO notwithstanding, this is needed to avoid
littering the tty with spurious errors when the user hits Ctrl-C to abort a
long-running builtin's output (w/ the primary example being `history`).
Up to now, in normal locales \x was essentially the same as \X, except
that it errored if given a value > 0x7f.
That's kind of annoying and useless.
A subtle change is that `\xHH` now represents the character (if any)
encoded by the byte value "HH", so even for values <= 0x7f if that's
not the same as the ASCII value we would diverge.
I do not believe anyone has ever run fish on a system where that
distinction matters. It isn't a thing for UTF-8, it isn't a thing for
ASCII, it isn't a thing for UTF-16, it isn't a thing for any extended
ASCII scheme - ISO8859-X, it isn't a thing for SHIFT-JIS.
I am reasonably certain we are making that same assumption in other
places.
Fixes#1352
Closes#9240.
Squash of the following commits (in reverse-chronological order):
commit 03b5cab3dc40eca9d50a9df07a8a32524338a807
Author: Mahmoud Al-Qudsi <mqudsi@neosmart.net>
Date: Sun Sep 25 15:09:04 2022 -0500
Handle differently declared posix_spawnxxx_t on macOS
On macOS, posix_spawnattr_t and posix_spawn_file_actions_t are declared as void
pointers, so we can't use maybe_t's bool operator to test if it has a value.
commit aed83b8bb308120c0f287814d108b5914593630a
Author: Mahmoud Al-Qudsi <mqudsi@neosmart.net>
Date: Sun Sep 25 14:48:46 2022 -0500
Update maybe_t tests to reflect dynamic bool conversion
maybe_t<T> is now bool-convertible only if T _isn't_ already bool-convertible.
commit 2b5a12ca97b46f96b1c6b56a41aafcbdb0dfddd6
Author: Mahmoud Al-Qudsi <mqudsi@neosmart.net>
Date: Sun Sep 25 14:34:03 2022 -0500
Make maybe_t a little harder to misuse
We've had a few bugs over the years stemming from accidental misuse of maybe_t
with bool-convertible types. This patch disables maybe_t's bool operator if the
type T is already bool convertible, forcing the (barely worth mentioning) need
to use maybe_t::has_value() instead.
This patch both removes maybe_t's bool conversion for bool-convertible types and
updates the existing codebase to use the explicit `has_value()` method in place
of existing implicit bool conversions.
The parent commit made the destructor of the DIR* member close it if necessary
(i.e. only if it's not null). This means that we can use the same logic in
the move constructor (where the source DIR* is null) and for move assignment
(where it might not be).
No functional change.
dir_iter_t closes its DIR* member in two places: the move assignment and
the destructor. Simplify this by closing it in the destructor of the DIR*
member which is called in both places. Use std::unique_ptr, which is shorter
than a dedicated wrapper class. Conveniently, it calls the deleter only if
the pointer is not-null. Unfortunately, std::unique_ptr requires explicit
conversion to DIR* when interacting with C APIs but it's probably still
better than a wrapper class.
This means that the noncopyable_t annotation is now implied due to the
unique_ptr member.
Additionally, we could probably remove the user-declared move constructor
and move assignment (the compiler-generated ones should be good enough). To
be safe, keep them around since they also erase the fd (though I hope we
don't rely on that behavior anywhere).
We should perhaps remove the user-declared destructor entirely but
dir_iter_t::entry_t also has one, I'm not sure why. Maybe there's a good
reason, like code size.
No functional change.
This was recently converted to a while-loop. However, we only
loop in a specific case when (by hitting "continue") so a
loop condition is not necessary.
No functional change.
We forgot to decode (i.e. turn into nice wchar_t codepoints)
"byte_literal" escape sequences. This meant that e.g.
```fish
string match ö \Xc3\Xb6
math 5 \X2b 5
```
didn't work, but `math 5 \x2b 5` did, and would print the wonderful
error:
```
math: Error: Missing operator
'5 + 5'
^
```
So, instead, we decode eagerly.
descend_unique_hierarchy is used for the cd autosuggestion: if a directory
contains exactly one subdirectory and no other entries, then propose that
as part of the cd autosuggestion.
This had a bug: if the subdirectory is a symlink to the parent, we would
chase that, going around the loop suggesting a longer path until we hit
PATH_MAX.
Fix this by using the new API which provides the inode "for free," and
track whether we've seen this inode before. This is technically too
conservative since the inode may be for a directory on a different device,
but devices are not available for free so this would incur a cost. In
practice encountering the same inode twice with different devices in a
unique hierarchy is unlikely, and should it happen the consequences are
merely cosmetic: we fail to suggest a longer path.
This introduces dir_iter_t, a new class for iterating the contents of a
directory. dir_iter_t encapsulates the logic that tries to avoid using
stat() to determine the type of a file, when possible.
While we hardcode the return values for the rest of our builtins, the `return`
builtin bubbles up whatever the user returned in their fish script, allowing
invalid return values such as negative numbers to make it into our C++ side of
things.
In creating a `proc_status_t` from the return code of a builtin, we invoke
W_EXITCODE() which is a macro that shifts left the return code by some amount,
and left-shifting a negative integer is undefined behavior.
Aside from causing us to land in UB territory, it also can cause some negative
return values to map to a "successful" exit code of 0, which was probably not
the fish script author's intention.
This patch also adds error logging to help catch any inadvertent additions of
cases where a builtin returns a negative value (should one forget that unix
return codes are always positive) and an assertion protecting against UB.
This was always the case if HAVE_TEXT wasn't defined, but if it was then we were
coercing the result of `_C()` to a `const wchar_t *` pointer, because we were
returning the address of a constant zero-length wchar_t pointer. This reserves a
local static `wcstring` variable that we can return as the "no text" sentinel
and bubbles back the `wcstring` reference rather than decomposing it into a
pointer.
This is a prerequisite for a bigger change I'm working on.
It's gone from 136 bytes to a 128 bytes by rearranging the items in order of
decreasing alignment requirements. While this reduces the memory consumption
slightly (by around 6%) for each completion we have in-memory, that translates
to only around ~8KiB of savings for a command with 1000 possible completions,
which is nice but ultimately not that big of a deal.
The bigger benefit is that a single `complete_entry_t` might now fit in a cache
line, hopefully making the process of testing completions for matches more
cache friendly (and maybe even faster).
The impact here depends on the command and how much output it
produces.
It's possible to get up to 1.5x - `string upper` being a good example,
or a no-op `string match '*'`.
But the more the command actually needs to do, the less of an effect
this has.
This basically immediately issues a "write()" if it's to a pipe or the
terminal.
That means we can reduce syscalls and improve performance, even by
doing something like
```c++
streams.out.append(somewcstring + L"\n");
```
instead of
```c++
streams.out.append(somewcstring);
streams.out.push_back(L'\n');
```
Some benchmarks of the
```fish
for i in (string repeat -n 2000 \n)
$thing
end
```
variety:
1. `set` (printing variables) sped up 1.75x
2. `builtin -n` 1.60x
3. `jobs` 1.25x (with 3 jobs)
4. `functions` 1.20x
5. `math 1 + 1` 1.1x
6. `pwd` 1.1x
Piping yields similar results, there is no real difference when
outputting to a command substitution.
This writes the output once per argument instead of once per format or
escaped char.
An egregious case:
```fish
printf (string repeat -n 200 \\x7f)%s\n (string repeat -n 2000 aaa\n)
```
Has been sped up by ~20x by reducing write() calls from 40000 to 200.
Even a simple
```fish
printf %s\n (string repeat -n 2000 aaa\n)
```
should now be ~1.2x faster by issuing 2000 instead of 4000 write
calls (the `\n` was written separately!).
This at least halves the number of "write()" calls we do if it goes to
a pipe or the terminal, or reduces them by 75% if there is a
description.
This makes
```fish
complete -c foo -xa "(seq 50000)"
complete -C"foo "
```
faster by 1.33x.
This uses wreaddir_resolving, which tries to use the dirent d_type
field if it exists. In that way, it can skip the `stat` to determine
if the given file is a directory.
This allows `cd` completions to skip stat in most cases:
```fish
strace -Ce newfstatat fish --no-config -c 'complete -C"cd /tmp/completion_test/"' >/dev/null
```
prints before:
```
% time seconds usecs/call calls errors syscall
------ ----------- ----------- --------- --------- ----------------
100,00 0,002627 2 1033 4 newfstatat
```
after:
```
% time seconds usecs/call calls errors syscall
------ ----------- ----------- --------- --------- ----------------
100,00 0,000054 1 31 3 newfstatat
```
for a directory with 1000 subdirectories.
(just `fish --no-config -c exit` does 26 newfstatat)
This should improve the situation with slow filesystems like fuse or
network fsen.
In case we have no d_type, we use `stat`, which would yield about the
same results.
The worst case is that we need directories *and* descriptions or the
"executable" flag (which we don't currently check for cd, if I read
this right?).
This flag determines whether or not more shortopt switches will be offered up as
potential completions (vs only the payload for the last-parsed shortopt switch).
Previously, it was being stomped before it was determined whether or not two
`complete` rules with different `result_mode.requires_param` values were
actually resolved against the current command line or not, and the last
evaluated completion rule would win out.
There are two changes here:
* `last_option_requires_param` is only assigned if all associated conditions for
a potential completion are also met, and
* If already assigned by a conflicting rule (which can only be user/developer
error), `last_option_requires_param` is allowed to change from true to false
but not the other way around (i.e. in case of a conflict, generate both
payloads and other shortopt completions)
The first change is immediately noticeable and affects many of our own
completions, see the discussion in #9221 for an example regarding `git` where
`-c` has any of about a million different possible meanings depending on which
completion preconditions have been met. The second change should only happen if
a dev/user mistakenly enters a `complete -c ...` rule for the same shortopt more
than once, both with conditions matching, sometimes requiring an argument and
not sometimes not. It should be a rare occurence.
This reverts commit 3d8f98c395.
In addition to the issues mentioned on the GitHub page for this commit,
it also broke the CentOS 7 build.
Note one can locally test the CentOS 7 build via:
./docker/docker_run_tests.sh ./docker/centos7.Dockerfile
Be more careful with sign extension issues stemming from the differences in how
an untyped literal is promoted to an integer vs how a typed (and signed) `char`
is promoted to an integer.
Also convert some `const[expr] static xxx` to `const[expr] xxx` where it makes
sense to let the compiler deduce on its own whether or not to allocate storage
for a constant variable rather than imposing our view that it should have STATIC
storage set aside for it.
A few call sites were not making use of the `XXX_LEN` definitions and were
calling `strlen(XXX)` - these have been updated to use `const_strlen(XXX)`
instead.
I'm not sure if any toolchains will have raise any issues with these changes...
CI will tell!
The optimization takes references to strings which are stored in a vector,
and stores those references in a set; but the strings are simultaneously
being moved within the vector, which may invalidate those references.
It's probably safe if you work through which particular strings are being
moved, but as a matter of principle we shouldn't take references to elements
of a vector while the vector is being rearranged, absenet a clear improvement
on a benchmark.
This reverts commit d5561623aa.
Whenever the command line changes, we redraw it with the previously computed
syntax highlighting. At the same time we start recomputing highlighting in
a background thread.
On some systems, the highlighting computation is slow, so the stale syntax
highlighting is visible.
The stale highlighting was computed for an old commandline. When the user
had inserted or deleted some characters in the middle, then the highlighting
is wrong for the characters to the right. This is because the characters
to the right have shifted but the highlighting hasn't. Fix this by also
shifting highlighting.
This means that text that was alrady highlighted will use the same
highlighting until a new one is computed. Newly inserted text uses the color
left of the cursor.
This is implemented by giving editable_line_t ownership of the highlighting.
It is able to perfectly sync text and highlighting; they will invariably
have the same length.
Fixes#9180
While its true that we only ever call this with temporaries, there is no
fundamental reason for this restriction. Taking by value is simpler and
more flexible. I think it does not change the generated code.
No functional change.
The idea for this function was that it stands as the one place that modifies
the text without push_edit. In practice I don't think it helps.
No functional change.
In theory this does less work so we should generally use this style.
In practice it looks uglier so I'm not sure. Maybe wait for stdlib ranges...
No functional change.
It turns out there *is* an obviously portable way... except it's
not-so-obviously not portable after all.
POSIX specifies that sigqueue(2) can be used to validate pid and signo
separately, returning EINVAL in the specific case of an invalid or unsupported
signal number. This would be perfect... if only it were actually implemented.
It seems that the WSLv1 implementation of pselect(2) does not check for
undelivered signals after the temporary sigmask is un-applied from the thread in
question.
When fish runs with job control enabled, it transfers ownership of the
tty to a child process, and then reclaims the tty after the process
exits. If job control is disabled then fish does not transfer or reclaim
the tty.
It may happen that the child process creates a pgroup and then transfers
the tty to it. In that case fish will not attempt to reclaim the tty, as
fish did not transfer it. Then when fish reads from stdin it will
receive SIGTTIN instead of data.
Fix this by unconditionally claiming the tty in readline().
Fixes#9181
This errored out *later* because the result was infinite or NaN, but
it didn't actually stop evaluation.
I'm not sure if there is a way to get floating point math to turn an
infinity back into something that doesn't depend on a literal
infinity, but division by zero conceptually isn't a thing we can
support.
There's entire branches of maths dedicated to figuring out what
dividing by "basically zero" means and we don't have to get into it.
This is essentially the inverse of `string pad`.
Where that adds characters to get up to the specified width,
this adds an ellipsis to a string if it goes over a specific maximum width.
The char can be given, but defaults to our ellipsis string.
("…" if the locale can handle it and "..." otherwise)
If the ellipsis string is empty, it just truncates.
For arguments given via argv, it goes line-by-line,
because otherwise length makes no sense.
If "--no-newline" is given, it adds an ellipsis instead and removes all subsequent lines.
Like pad and `length --visible`, it goes by visible width,
skipping recognized escape sequences, as those have no influence on width.
The default target width is the shortest of the given widths that is non-zero.
If the ellipsis is already wider than the target width,
we truncate instead. This is safer overall, so we don't e.g. move into a new line.
This is especially important given our default ellipsis might be width 3.
When selecting items in the pager, only the latest of those items is kept
in the edit history, as so-called transient edit. Each new transient edit
evicts any old transient edit (via undo).
If the pager is closed by a command that performs another transient edit
(like history-token-search-backward) we thus inadvertently undo (= remove)
the token inserted by the pager. Fix this by closing a transient edit
session when closing the pager. Token search will start its own session.
Fixes#9160
