From: Nick Terrell <terrelln@fb.com>
Date: Wed, 13 May 2020 00:51:16 +0000 (-0700)
Subject: [greedy] Fix performance instability
X-Git-Tag: v1.4.5^2~16^2
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=70c80e19e6c167a0d7eb30a8604fe70a8f81fb8d;p=thirdparty%2Fzstd.git

[greedy] Fix performance instability
---

diff --git a/lib/compress/zstd_lazy.c b/lib/compress/zstd_lazy.c
index 67ed07cba..4cf5c88b5 100644
--- a/lib/compress/zstd_lazy.c
+++ b/lib/compress/zstd_lazy.c
@@ -681,6 +681,12 @@ ZSTD_compressBlock_lazy_generic(
     }
 
     /* Match Loop */
+#if defined(__GNUC__) && defined(__x86_64__)
+    /* I've measured random a 5% speed loss on levels 5 & 6 (greedy) when the
+     * code alignment is perturbed. To fix the instability align the loop on 32-bytes.
+     */
+    __asm__(".p2align 5");
+#endif
     while (ip < ilimit) {
         size_t matchLength=0;
         size_t offset=0;
@@ -952,6 +958,12 @@ size_t ZSTD_compressBlock_lazy_extDict_generic(
     ip += (ip == prefixStart);
 
     /* Match Loop */
+#if defined(__GNUC__) && defined(__x86_64__)
+    /* I've measured random a 5% speed loss on levels 5 & 6 (greedy) when the
+     * code alignment is perturbed. To fix the instability align the loop on 32-bytes.
+     */
+    __asm__(".p2align 5");
+#endif
     while (ip < ilimit) {
         size_t matchLength=0;
         size_t offset=0;