Replace `println!` calls on the results paths with `writeln!` on
`io::stdout()` and propagate the `io::Error` to `main`, where
`ErrorKind::BrokenPipe` is turned into a quiet `exit(0)`. This matches
the conventional behavior of command-line tools when their output pipe
is closed early (e.g. `hyperfine ... | head -n 0`), and fixes panics in
all run-output paths: the per-benchmark header and results, the speed
comparison summary, and `-`-style exports to stdout. The warnings and
"Note" outputs to stderr are converted the same way.
---------
Co-authored-by: Mathjk <Mathjk@users.noreply.github.com>
Co-authored-by: David Peter <mail@david-peter.de>
This implementation of --conclude is basically identical
to the existing implementation for --prepare. I have not
adapted any existing tests. All tests which test --conclude
functionality are new.
The --conclude commands will run in the same order in which
they are supplied on the cli in the same manner as --prepare
and there must either be one --conclude command, or else the
same number of --concludes as there are commands.
This adds a new "raw executor" (next to the "shell executor") that
allows hyperfine to execute commands directly without any intermediate
shell.
The command line is split into tokens (using the `shell-words` crate),
and according to POSIX rules. The first token is taken as the executable,
and the rest as arguments.
The new executor is enabled by default. In order to select the shell
executor, users will have to pass `--shell=default`.
This allows us to reduce measurement noise and to benchmark very quick
commands. It also decreases the time to run benchmarks, as we don't need
the calibration phase.
Also, it allows one to make sure that the executed command is not
implemented as a shell builtin. For example `hyperfine true`
and `hyperfine --shell=default true` return different times due
to the fact that `bash` executes `true` as a NOP.
Co-authored: Ciprian Dorin Craciun <ciprian@volution.ro>
David Peter