From f03ed0e3b63c5c851ef0d170fdfb0dbba5f523e9 Mon Sep 17 00:00:00 2001
From: David Hoppenbrouwers <david@salt-inc.org>
Date: Thu, 7 Jul 2022 13:20:41 +0200
Subject: [PATCH] Skip rep movsb in copy_backward if possible

There is currently no measureable performance difference in benchmarks
but it likely will make a difference in real workloads.
---
 src/mem/x86_64.rs | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/src/mem/x86_64.rs b/src/mem/x86_64.rs
index a1015f612..e9c1c56d5 100644
--- a/src/mem/x86_64.rs
+++ b/src/mem/x86_64.rs
@@ -73,15 +73,21 @@ pub unsafe fn copy_backward(dest: *mut u8, src: *const u8, count: usize) {
     // We can't separate this block due to std/cld
     asm!(
         "std",
+        "test %ecx, %ecx",
+        "jz 1f",
         "rep movsb",
+        "1:",
         "sub $7, %rsi",
         "sub $7, %rdi",
         "mov {qword_count}, %rcx",
         "rep movsq",
+        "test {pre_byte_count:e}, {pre_byte_count:e}",
+        "jz 1f",
         "add $7, %rsi",
         "add $7, %rdi",
         "mov {pre_byte_count:e}, %ecx",
         "rep movsb",
+        "1:",
         "cld",
         pre_byte_count = in(reg) pre_byte_count,
         qword_count = in(reg) qword_count,