A fast wc clone in Rust.
-% cw --help
cw 0.5.0
Thomas Hurst <[email protected]>
Count Words - word, line, character and byte count
USAGE:
cw [FLAGS] [OPTIONS] [input]...
FLAGS:
-c, --bytes Count bytes
-m, --chars Count UTF-8 characters instead of bytes
-h, --help Prints help information
-l, --lines Count lines
-L, --max-line-length Count bytes (default) or characters (-m) of the longest line
-V, --version Prints version information
-w, --words Count words
OPTIONS:
--files0-from <files0_from> Read input from the NUL-terminated list of filenames in the given file.
--files-from <files_from> Read input from the newline-terminated list of filenames in the given file.
--threads <threads> Number of counting threads to spawn [default: 1]
ARGS:
<input>... Input files
-% cw Dickens_Charles_Pickwick_Papers.xml
3449440 51715840 341152640 Dickens_Charles_Pickwick_Papers.xml
Counts of multiple files may be accelerated by use of the --threads option:
'xargs <files cw --threads=12' ran
2.01 ± 0.03 times faster than 'xargs <files cw --threads=4'
7.07 ± 0.09 times faster than 'xargs <files cw'
11.55 ± 0.15 times faster than 'xargs <files wc'
17.31 ± 0.23 times faster than 'xargs <files gwc'
Line counts are optimized using the bytecount crate:
'cw -l Dickens_Charles_Pickwick_Papers.xml' ran
3.44 ± 0.04 times faster than 'wc -l Dickens_Charles_Pickwick_Papers.xml'
4.17 ± 0.05 times faster than 'gwc -l Dickens_Charles_Pickwick_Papers.xml'
Line counts with line length are optimized using the memchr crate:
'cw -lL Dickens_Charles_Pickwick_Papers.xml' ran
1.73 ± 0.01 times faster than 'wc -lL Dickens_Charles_Pickwick_Papers.xml'
15.07 ± 0.07 times faster than 'gwc -lL Dickens_Charles_Pickwick_Papers.xml'
Note without -m cw only operates on bytes, and it never cares about your locale.
'cw Dickens_Charles_Pickwick_Papers.xml' ran
1.45 ± 0.01 times faster than 'wc Dickens_Charles_Pickwick_Papers.xml'
2.05 ± 0.00 times faster than 'gwc Dickens_Charles_Pickwick_Papers.xml'
-m enables UTF-8 processing, with a fast-path for just character length, again
using bytecount:
'cw -m Dickens_Charles_Pickwick_Papers.xml' ran
30.21 ± 0.39 times faster than 'gwc -m Dickens_Charles_Pickwick_Papers.xml'
70.36 ± 0.91 times faster than 'wc -m Dickens_Charles_Pickwick_Papers.xml'
'cw -m test-utf-8.html' ran
84.74 ± 1.12 times faster than 'wc -m test-utf-8.html'
124.21 ± 1.64 times faster than 'gwc -m test-utf-8.html'
And another path for character and line length:
'cw -mlL Dickens_Charles_Pickwick_Papers.xml' ran
3.88 ± 0.01 times faster than 'gwc -mlL Dickens_Charles_Pickwick_Papers.xml'
9.05 ± 0.02 times faster than 'wc -mlL Dickens_Charles_Pickwick_Papers.xml'
'cw -mlL test-utf-8.html' ran
9.42 ± 0.01 times faster than 'wc -mlL test-utf-8.html'
18.95 ± 0.03 times faster than 'gwc -mlL test-utf-8.html'
And a slow path for everything else:
'cw -mLlw Dickens_Charles_Pickwick_Papers.xml' ran
1.35 ± 0.00 times faster than 'gwc -mLlw Dickens_Charles_Pickwick_Papers.xml'
3.15 ± 0.00 times faster than 'wc -mLlw Dickens_Charles_Pickwick_Papers.xml'
These tests are on FreeBSD 12 on a 2.1GHz Westmere Xeon. gwc is from GNU
coreutils 8.30 - note its performance here is rather pessimised in some areas by
FreeBSD's rather weak memchr implementation. YMMV.
For best results build with:
cargo build --release --features runtime-dispatch-simd
This enables SIMD optimizations for line and character counting. It has no effect if you count anything else.
- Test suite.
- Factor internals out into a library. (#1)
- Improve multibyte support.
- Possibly implement locale.
- Replace clap/structopt with something lighter.
uwc focuses on following Unicode rules as precisely as possible, taking into account less-common newlines, counting graphemes as well as codepoints, and following Unicode word-boundary rules precisely.
The cost of this is currently a great deal of performance, with counts on my benchmark file taking over a minute.
cw was originally called rwc until I noticed this existed. It's quite old and doesn't appear to compile.
A little library that only does plain newline counting, along with a binary
called lc. Version 0.2 will use the same algorithm as cw.