From: Niels Möller <nisse@lysator.liu.se>
Date: Tue, 13 Jan 2015 16:43:07 +0000 (+0100)
Subject: Explain why x86_64 fat initialization doesn't need any memory barrier.
X-Git-Tag: nettle_3.1rc1~46^2~15
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=44f2687d2cd8f6d2163a43f7299e2b903f6e4552;p=thirdparty%2Fnettle.git

Explain why x86_64 fat initialization doesn't need any memory barrier.
---

diff --git a/x86_64/fat/fat.c b/x86_64/fat/fat.c
index 7deb45ae..fc022b9e 100644
--- a/x86_64/fat/fat.c
+++ b/x86_64/fat/fat.c
@@ -150,10 +150,10 @@ fat_init (void)
       _aes_encrypt_vec = _nettle_aes_encrypt_x86_64;
       _aes_decrypt_vec = _nettle_aes_decrypt_x86_64;
     }
-  /* FIXME: We ought to use some thread-aware memory barrier before
-     setting the initialized flag. For now, just do another cpuinfo
-     call to get some synchronization. */
-  _nettle_cpuid (1, cpuid_data);
+
+  /* The x86_64 architecture should always make stores visible in the
+     right order to other processors (except for non-temporal stores
+     and the like). So we don't need any memory barrier. */
   initialized = 1;
 }