From 6d3c0bee68dbcb02ef343f1505b33cfdc2c80246 Mon Sep 17 00:00:00 2001
From: Nicolas Boichat <nicolas@boichat.ch>
Date: Tue, 18 Mar 2025 20:03:47 +0100
Subject: [PATCH] seq: Buffer writes to stdout

Use a BufWriter to wrap stdout: reduces the numbers of system calls,
improves performance drastically (2x in some cases).

Also document use cases in src/uu/seq/BENCHMARKING.md, and the
optimization we have just done here.
---
 src/uu/seq/BENCHMARKING.md | 33 ++++++++++++++++++++++++++++++++-
 src/uu/seq/src/seq.rs      |  6 +++---
 2 files changed, 35 insertions(+), 4 deletions(-)

diff --git a/src/uu/seq/BENCHMARKING.md b/src/uu/seq/BENCHMARKING.md
index a633d509c..89cc47861 100644
--- a/src/uu/seq/BENCHMARKING.md
+++ b/src/uu/seq/BENCHMARKING.md
@@ -19,7 +19,38 @@ Finally, you can compare the performance of the two versions of `seq`
 by running, for example,
 
 ```shell
-hyperfine "seq 1000000" "target/release/seq 1000000"
+hyperfine -L seq seq,target/release/seq "{seq} 1000000"
 ```
 
+## Interesting test cases
+
+Performance characteristics may vary a lot depending on the parameters,
+and if custom formatting is required. In particular, it does appear
+that the GNU implementation is heavily optimized for positive integer
+outputs (which is probably the most common use case for `seq`).
+
+Specifying a format or fixed width will slow down the
+execution a lot (~15-20 times on GNU `seq`):
+```shell
+hyperfine -L seq seq,target/release/seq "{seq} -f%g 1000000"
+hyperfine -L seq seq,target/release/seq "{seq} -w 1000000"
+```
+
+Floating point increments, or any negative bound, also degrades the
+performance (~10-15 times on GNU `seq`):
+```shell
+hyperfine -L seq seq,./target/release/seq "{seq} 0 0.000001 1"
+hyperfine -L seq seq,./target/release/seq "{seq} -100 1 1000000"
+```
+
+## Optimizations
+
+### Buffering stdout
+
+The original `uutils` implementation of `seq` did unbuffered writes
+to stdout, causing a large number of system calls (and therefore a large amount
+of system time). Simply wrapping `stdout` in a `BufWriter` increased performance
+by about 2 times for a floating point increment test case, leading to similar
+performance compared with GNU `seq`.
+
 [0]: https://github.com/sharkdp/hyperfine
diff --git a/src/uu/seq/src/seq.rs b/src/uu/seq/src/seq.rs
index a6b5e32ea..08b989815 100644
--- a/src/uu/seq/src/seq.rs
+++ b/src/uu/seq/src/seq.rs
@@ -4,7 +4,7 @@
 // file that was distributed with this source code.
 // spell-checker:ignore (ToDO) bigdecimal extendedbigdecimal numberparse hexadecimalfloat
 use std::ffi::OsString;
-use std::io::{stdout, ErrorKind, Write};
+use std::io::{stdout, BufWriter, ErrorKind, Write};
 
 use clap::{Arg, ArgAction, Command};
 use num_traits::{ToPrimitive, Zero};
@@ -262,8 +262,8 @@ fn print_seq(
     padding: usize,
     format: Option<&Format<num_format::Float>>,
 ) -> std::io::Result<()> {
-    let stdout = stdout();
-    let mut stdout = stdout.lock();
+    let stdout = stdout().lock();
+    let mut stdout = BufWriter::new(stdout);
     let (first, increment, last) = range;
     let mut value = first;
     let padding = if pad {