From ca11b53896365c948426974cb90e8f71c70d123b Mon Sep 17 00:00:00 2001
From: Simon South <simon@simonsouth.net>
Date: Sun, 31 May 2020 20:36:43 -0400
Subject: [PATCH] Guard floating-point opcodes with explicit memory barrier
---
src/interp/engine/interp-inlining.h | 11 ++++++++++-
1 file changed, 10 insertions(+), 1 deletion(-)
diff --git a/src/interp/engine/interp-inlining.h b/src/interp/engine/interp-inlining.h
index 3339b0e..4ee5c5a 100644
--- a/src/interp/engine/interp-inlining.h
+++ b/src/interp/engine/interp-inlining.h
@@ -78,8 +78,17 @@
4.3, we need to insert a label, and ensure its address
is taken (to stop it being optimised out). However,
this reduces performance on PowerPC by approx 1 - 2%.
+
+ With gcc 5 and newer an asm statement with a "memory"
+ clobber argument explicitly sets a memory barrier for the
+ compiler, preventing it from reordering memory accesses
+ in a way that breaks decaching.
*/
-#if (__GNUC__ == 4) && (__GNUC_MINOR__ >= 3)
+#if (__GNUC__ > 4)
+#define DEF_GUARD_TABLE(level) /* none */
+#define GUARD(opcode, level) __asm__("" ::: "memory");
+#define GUARD_TBLS /* none */
+#elif (__GNUC__ == 4) && (__GNUC_MINOR__ >= 3)
#define DEF_GUARD_TABLE(level) DEF_HANDLER_TABLE(level, GUARD)
#define GUARD(opcode, level) label(opcode, level, GUARD)
#define GUARD_TBLS , HNDLR_TBLS(GUARD)
--
2.26.2
