tac

tac is a high-performance, simd-accelerated, cross-platform rewrite of the GNU tac utility from Coreutils, released under a BSD-compatible (MIT) license. tac reads input from a file (or from stdin, but see below) and then prints it line-by-line backwards.

This tac implementation uses simd-acceleration for new line detection (read more about that here) and utilizes memory-mapped files on all supported operating systems. It is additionally written in rust for maximum integrity and safety.

Who needs a faster tac anyway?

Good question. Try grepping through a multi-gigabyte web access log file in reverse chronological order (tac --line-buffered access.log | grep foo) and then get back to me.

Usage

Usage: tac [OPTIONS] [FILE1..]
Write each FILE to standard output, last line first.
Reads from stdin if FILE is - or not specified.

Options:
  -h --help        Print this help text and exit
  -v --version     Print version and exit.
  --line-buffered  Always flush output after each line.

tac reads lines from any combination of stdin and/or zero or more files and writes the lines to the output in reverse order.

Example

$ echo -e "hello\nworld" | tac
world
hello

Installation

tac may be installed via cargo, the rust package manager:

cargo install tac

Help is humbly requested in getting tac into the package managers for various platforms. It's really a time-consuming task, especially for someone that only interacts with the various packaging tools once in a blue moon as opposed to on a daily basis.

Currently arm64 (aarch64) NEON acceleration is only available with the nightly compiler. If using it, provide the --features nightly flag to cargo to enable support.

Implementation Notes

This implementation of tac uses SIMD instruction sets (AVX2, NEON) to accelerate the detection of new lines. The usage of memory-mapped files additionally boosts performance by avoiding slowdowns caused by context switches when reading from the input if speculative execution mitigations are enabled. It is significantly (2.55x if mitigations disabled, more otherwise) faster than the version of tac that ships with GNU Coreutils, in addition to being more liberally licensed.

To obtain maximum performance:

• Try not to pipe input into tac. e.g. instead of running cat /usr/share/dict/words | tac, run tac /usr/share/dict/words directly. Because tac by definition must reach the end-of-file before it can emit its input with the lines reversed, if you use tac's stdin interface (e.g. cat foo | tac), it must buffer all stdin input before it can begin to process the results. tac will try to buffer in memory, but once it exceeds a certain high-water mark (currently 4 MiB), it switches to disk-based buffering (because it can't know how large the input is or if it will end up exceeding the available free memory).
• Always try to place tac at the start of a pipeline where possible. Even if you can guarantee that the input to tac will not exceed the in-memory buffering limit (see above), tac is almost certainly faster than any other command in your pipeline, and if you are going to reverse the output, you will benefit most if you reverse it from the start, unless you are always going to run the command to completion. For example, instead of running grep foo /var/log/nginx/access.log | tac, run tac /var/log/nginx/access.log | grep foo. This will (significantly) reduce the amount of time/work before the first n matches are reported (because the file is first quickly reversed then searched in the desired order, vs slowly searched in its entirety and only then are the results reversed).
• Use line-buffered output mode (tac --line-buffered) if tac is piping into another command rather than writing to the tty directly. This gives you "live" streaming of results and lets you terminate much sooner if you're only looking for the first n matches. e.g. tac --line-buffered access.log | grep foo will print its first match much, much sooner than tac access.log | grep foo would.
• In the same vein, if you are chaining the output of n utilities, make sure that all commands up to n - 1 are all using line-buffered mode unless you don't care about latency and only care about throughput. For example, to print the first two matches for some grep pattern: tac --line-buffered access.log | grep --line-buffered foo | head -n2.

License and Copyright

tac is released under the terms of the MIT public license. Copyright Mahmoud Al-Qudsi 2017-2021. All rights not assigned by the MIT license are reserved.

As an open source project, tac would not exist without the tireless efforts of its various contributors - see CONTRIBUTORS.md for full details.