From e10a34635628fab3bd383e2d17c89dcd6be75b15 Mon Sep 17 00:00:00 2001
From: Benedikt Rascher-Friesenhausen
<benediktrascherfriesenhausen+git@gmail.com>
Date: Sun, 7 Oct 2018 10:25:19 +0200
Subject: [PATCH] Optimise `memcmp` for speed
I saw that in other parts of the `string` module iterations over `usize` were
used to increase iteration speed. In this patch I apply the same logic to
`memcmp`. With this change I measured a 7x speedup for `memcmp` on a ~1MB
buffer (comparing two buffers with the same content) on my machine (i7-7500U),
but I did not do any real world benchmarking for the change. The increase in
speed comes with the tradeoff of both increased complexity and larger generated
assembly code for the function.
I tested the correctness of the implementation by generating two randomly filled
buffers and comparing the `memcmp` result of the old implementation against this
new one.
I ran the tests and currently currently three of them fail:
- netdb (fails to run)
- stdio/rename (fails to verify)
- unistd/pipe (fails to verify)
They do so though regardless of this change, so I don't think they are related.
---
src/header/string/mod.rs | 30 ++++++++++++++++++++++++------
1 file changed, 24 insertions(+), 6 deletions(-)
diff --git a/src/header/string/mod.rs b/src/header/string/mod.rs
index 9a6e4ddd..0efe60f8 100644
--- a/src/header/string/mod.rs
+++ b/src/header/string/mod.rs
@@ -73,14 +73,32 @@ pub unsafe extern "C" fn memchr(s: *const c_void, c: c_int, n: usize) -> *mut c_
#[no_mangle]
pub unsafe extern "C" fn memcmp(s1: *const c_void, s2: *const c_void, n: usize) -> c_int {
- let mut i = 0;
- while i < n {
- let a = *(s1 as *const u8).offset(i as isize);
- let b = *(s2 as *const u8).offset(i as isize);
- if a != b {
+ let (div, rem) = (n / mem::size_of::<usize>(), n % mem::size_of::<usize>());
+ let mut a = s1 as *const usize;
+ let mut b = s2 as *const usize;
+ for _ in 0..div {
+ if *a != *b {
+ for i in 0..mem::size_of::<usize>() {
+ let c = *(a as *const u8).offset(i as isize);
+ let d = *(b as *const u8).offset(i as isize);
+ if c != d {
+ return c as i32 - d as i32;
+ }
+ }
+ unreachable!()
+ }
+ a = a.offset(1);
+ b = b.offset(1);
+ }
+
+ let mut a = a as *const u8;
+ let mut b = b as *const u8;
+ for _ in 0..rem {
+ if *a != *b {
return a as i32 - b as i32;
}
- i += 1;
+ a = a.offset(1);
+ b = b.offset(1);
}
0
}
--
GitLab