| Age | Commit message (Collapse) | Author |
|---|
|
|
|
This commit fixes a subtle bug in how the line buffer was rolling its
contents. Specifically, when ripgrep searches without memory maps,
it uses a "roll" buffer for incremental line oriented search without
needing to read the entire file into memory at once. The roll buffer
works by reading a chunk of bytes from the file into memory, and then
searching everything in that buffer up to the last `\n` byte. The bytes
*after* the last `\n` byte are preserved, since they likely correspond
to *part* of the next line. Once ripgrep is done searching the buffer,
it "rolls" the buffer such that the start of the next line is at the
beginning of the buffer, and then ripgrep reads more data into the
buffer starting at the (possibly) partial end of that line.
The implication of this strategy, necessarily so, is that a buffer must
be big enough to fit a single line in memory. This is because the regex
engine needs a contiguous block of memory to search, so there is no way
to search anything smaller than a single line. So if a file contains a
single line with 7.5 million bytes, then the buffer will grow to be at
least that size. (Many files have super long lines like this, but they
tend to be *binary* files, which ripgrep will detect and stop searching
unless the user forces it with the `-a/--text` flag. So in practice,
they aren't usually a problem. However, in this case, #1335 found a case
where a plain text file had a line with 7.5 million bytes.)
Now, for performance reasons, ripgrep reuses these buffers across its
search. Typically, it will create `N` of these line buffers when it
starts (where `N` is the number of threads it is using), and then reuse
them without creating any new ones as it searches through files.
This means that if you search a file with a very long line, that buffer
will expand to be big enough to store that line. ripgrep never contracts
these buffers, so once it searches the next file, ripgrep will continue
to use this large buffer. While it might be prudent to contract these
buffers in some circumstances, this isn't otherwise inherently a
problem. The memory has already been allocated, and there isn't much
cost to using it, other than the fact that ripgrep hangs on to it and
never gives it back to the OS.
However, the `roll` implementation described above had a really
important bug in it that was impacted by the size of the buffer.
Specifically, it used the following to "roll" the partial line at the
end of the buffer to the beginning:
self.buf.copy_within_str(self.pos.., 0);
Which means that if the buffer is very large, ripgrep will copy
*everything* from `self.pos` (which might be very small, e.g., for small
files) to the end of the buffer, and move it to the beginning of the
buffer. This will happen repeatedly each time the buffer is used to
search small files, which winds up being quite a large slow down if the
line was exceptionally large (say, megabytes).
It turns out that copying everything is completely unnecessary. We only
need to copy the remainder of the last read to the beginning of the
buffer. Everything *after* the last read in the buffer is just free
space that can be filled for the next read. So, all we need to do is
copy just those bytes:
self.buf.copy_within_str(self.pos..self.end, 0);
... which is typically much much smaller than the rest of the buffer.
This was likely also causing small performance losses in other cases as
well. For example, when searching a lot of small files, ripgrep would
likely do a lot more copying than necessary. Although, given that the
default buffer size is 8KB, this extra copying was likely pretty small,
and was thus harder to observe.
Fixes #1335
|
|
|
|
|
|
Use `dyn` for trait objects and use `..=` for inclusive ranges.
|
|
|
|
|
|
|
|
This commit attempts to surface binary filtering in a slightly more
user friendly way. Namely, before, ripgrep would silently stop
searching a file if it detected a NUL byte, even if it had previously
printed a match. This can lead to the user quite reasonably assuming
that there are no more matches, since a partial search is fairly
unintuitive. (ripgrep has this behavior by default because it really
wants to NOT search binary files at all, just like it doesn't search
gitignored or hidden files.)
With this commit, if a match has already been printed and ripgrep detects
a NUL byte, then it will print a warning message indicating that the search
stopped prematurely.
Moreover, this commit adds a new flag, --binary, which causes ripgrep to
stop filtering binary files, but in a way that still avoids dumping
binary data into terminals. That is, the --binary flag makes ripgrep
behave more like grep's default behavior.
For files explicitly specified in a search, e.g., `rg foo some-file`,
then no binary filtering is applied (just like no gitignore and no
hidden file filtering is applied). Instead, ripgrep behaves as if you
gave the --binary flag for all explicitly given files.
This was a fairly invasive change, and potentially increases the UX
complexity of ripgrep around binary files. (Before, there were two
binary modes, where as now there are three.) However, ripgrep is now a
bit louder with warning messages when binary file detection might
otherwise be hiding potential matches, so hopefully this is a net
improvement.
Finally, the `-uuu` convenience now maps to `--no-ignore --hidden
--binary`, since this is closer to the actualy intent of the
`--unrestricted` flag, i.e., to reduce ripgrep's smart filtering. As a
consequence, `rg -uuu foo` should now search roughly the same number of
bytes as `grep -r foo`, and `rg -uuua foo` should search roughly the
same number of bytes as `grep -ra foo`. (The "roughly" weasel word is
used because grep's and ripgrep's binary file detection might differ
somewhat---perhaps based on buffer sizes---which can impact exactly what
is and isn't searched.)
See the numerous tests in tests/binary.rs for intended behavior.
Fixes #306, Fixes #855
|
|
This commit adds a new encoding feature where the -E/--encoding flag
will now accept a value of 'none'. When given this value, all encoding
related machinery is disabled and ripgrep will search the raw bytes of
the file, including the BOM if it's present.
Closes #1207, Closes #1208
|
|
This commit causes grep-searcher to use byte strings internally for its
line buffer support. We manage to remove a use of `unsafe` by doing this
(by pushing it down into `bstr`).
We stop short of using byte strings everywhere else because we rely
heavily on the `impl ops::Index<[u8]> for grep_matcher::Match` impl,
which isn't available for byte strings. (It is premature to make bstr a
public dep of a core crate like grep-matcher, but maybe some day.)
|
|
|
|
This undoes the patch to stop using bytecount on big-endian
architectures. In particular, we bump our bytecount dependency to the
latest release, which has a fix.
This reverts commit a4868b88351318182eed3b801d0c97a106a7d38f.
Fixes #1144 (again), Closes #1194
|
|
|
|
This patches out bytecount's "fast" vectorized algorithm on big-endian
machines, where it has been observed to fail. Going forward, bytecount
should probably fix this on their end, but for now, we take a small
performance hit on big-endian machines.
Fixes #1144
|
|
This commit improves the CRLF hack to be more robust. In particular, in
addition to rewriting `$` as `(?:\r??$)`, we now strip `\r` from the end
of a match if and only if the regex has an ending line anchor required for
a match. This doesn't quite make the hack 100% correct, but should fix most
use cases in practice. An example of a regex that will still be incorrect
is `foo|bar$`, since the analysis isn't quite sophisticated enough to
determine that a `\r` can be safely stripped from any match. Even if we
fix that, regexes like `foo\r|bar$` still won't be handled correctly. Alas,
more work on this front should really be focused on enabling this in the
regex engine itself.
The specific cause of this bug was that grep-searcher was sneakily
stripping CRLF from matching lines when it really shouldn't have. We remove
that code now, and instead rely on better match semantics provided at a
lower level.
Fixes #1095
|
|
This fixes a bug where a BOM prefix was included. While this was somewhat
intentional in order to have a faithful "UTF8 passthru" option, in
practice, this causes problems such as breaking patterns like `^` in a
really non-obvious way.
The actual fix was to add a new API to encoding_rs_io, which this commit
brings in.
Fixes #1163
|
|
|
|
bytecount now uses runtime dispatch for enabling SIMD, which means we can
no longer need the avx-accel features. We remove it from ripgrep since the
next release will be a minor version bump, but leave them as no-ops for
the crates that previously used it.
|
|
We also increase the MSRV to 1.32, the current stable release, which sets
the stage for migrating to Rust 2018.
|
|
Looks like the deny(missing_docs) lint got a bit stronger.
|
|
This commit bumps the version of encoding_rs to use the latest release.
This appears to fix a panic in UTF-16 decoding.
Fixes #1089
|
|
This update includes a work-around for a presumed bug in encoding_rs
that causes a panic:
https://github.com/hsivonen/encoding_rs/issues/34
Specifically, to reproduce this in ripgrep, one can run the following:
$ curl -LO https://cache.ruby-lang.org/pub/ruby/2.5/ruby-2.5.1.tar.gz
$ tar xf ruby-2.5.1.tar.gz
$ rg ZZZZZ ruby-2.5.1/test/rexml/data/t63-2.svg
thread 'main' panicked at 'index out of bounds: the len is 1 but the index is 1'
Fixes #1052
|
|
I don't think every change here is needed, but this ensures we're using
the latest version of every direct dependency.
|
|
We initially did not have this impl because the first revision of the Sink
trait was much more complicated. In particular, each method was
parameterized over a Matcher. But not every Sink impl actually needs a
Matcher, and it is just as easy to borrow a Matcher explicitly, so the
added parameterization wasn't holding its own.
This does permit Sink implementations to be used as trait objects. One
key use case here is to reduce compile times, since there is quite a bit
of code inside grep-searcher that is parameterized on Sink. Unfortunately,
that code is *also* parameterized on Matcher, and the various printers in
grep-printer are also parameterized on Matcher, which means Sink trait
objects are necessary but no sufficient for a major reduction in compile
times. Unfortunately, the path to making Matcher object safe isn't quite
clear. Extension traits maybe? There's also stuff in the Serde ecosystem
that might help, but the type shenanigans can get pretty gnarly.
|
|
|
|
This prepares them for an initial 0.1.0 release.
|
|
libripgrep is not any one library, but rather, a collection of libraries
that roughly separate the following key distinct phases in a grep
implementation:
1. Pattern matching (e.g., by a regex engine).
2. Searching a file using a pattern matcher.
3. Printing results.
Ultimately, both (1) and (3) are defined by de-coupled interfaces, of
which there may be multiple implementations. Namely, (1) is satisfied by
the `Matcher` trait in the `grep-matcher` crate and (3) is satisfied by
the `Sink` trait in the `grep2` crate. The searcher (2) ties everything
together and finds results using a matcher and reports those results
using a `Sink` implementation.
Closes #162
|
