From 9df32c34af9f82afeedb199e8bdfcc39e0af44d7 Mon Sep 17 00:00:00 2001
From: simonmar <unknown>
Date: Thu, 24 Nov 2005 14:21:33 +0000
Subject: [PATCH] [project @ 2005-11-24 14:21:33 by simonmar] unlockClosure()
 requires a write barrier for the compiler - write barriers
 aren't required for the CPU, but gcc re-orders non-aliasing
 writes unless we use an explicit barrier.

This only just showed up when we started compiling the RTS with -O2.
---
 ghc/includes/SMP.h |   38 ++++++++++++++++++++++++++++++++++----
 1 file changed, 34 insertions(+), 4 deletions(-)

diff --git a/ghc/includes/SMP.h b/ghc/includes/SMP.h
index cc95941..d8f2aa1 100644
--- a/ghc/includes/SMP.h
+++ b/ghc/includes/SMP.h
@@ -25,7 +25,7 @@
 
 /* 
  * XCHG - the atomic exchange instruction.  Used for locking closures
- * during updates (see LOCK_CLOSURE below) and the MVar primops.
+ * during updates (see lockClosure() below) and the MVar primops.
  *
  * NB: the xchg instruction is implicitly locked, so we do not need
  * a lock prefix here. 
@@ -57,14 +57,43 @@ cas(StgVolatilePtr p, StgWord o, StgWord n)
     return o;
 }
 
+/*
+ * Write barrier - ensure that all preceding writes have happened
+ * before all following writes.  
+ *
+ * We need to tell both the compiler AND the CPU about the barrier.
+ * This is a brute force solution; better results might be obtained by
+ * using volatile type declarations to get fine-grained ordering
+ * control in C, and optionally a memory barrier instruction on CPUs
+ * that require it (not x86 or x86_64).
+ */
+INLINE_HEADER void
+wb(void) {
+#if i386_HOST_ARCH || x86_64_HOST_ARCH
+    __asm__ __volatile__ ("" : : : "memory");
+#else
+#error memory barriers unimplemented on this architecture
+#endif
+}
+
+/*
+ * Locking/unlocking closures
+ *
+ * This is used primarily in the implementation of MVars.
+ */
+#define SPIN_COUNT 4000
+
 INLINE_HEADER StgInfoTable *
 lockClosure(StgClosure *p)
 {
 #if i386_HOST_ARCH || x86_64_HOST_ARCH
     StgWord info;
     do {
-	info = xchg((P_)&p->header.info, (W_)&stg_WHITEHOLE_info);
-	if (info != (W_)&stg_WHITEHOLE_info) return (StgInfoTable *)info;
+	nat i = 0;
+	do {
+	    info = xchg((P_)&p->header.info, (W_)&stg_WHITEHOLE_info);
+	    if (info != (W_)&stg_WHITEHOLE_info) return (StgInfoTable *)info;
+	} while (++i < SPIN_COUNT);
 	yieldThread();
     } while (1);
 #else
@@ -76,7 +105,8 @@ INLINE_HEADER void
 unlockClosure(StgClosure *p, StgInfoTable *info)
 {
 #if i386_HOST_ARCH || x86_64_HOST_ARCH
-    // This is safe enough, because lockClosure() does the memory barrier:
+    // This is a strictly ordered write, so we need a wb():
+    wb();
     p->header.info = info;
 #else
     RELEASE_SM_LOCK;
-- 
1.7.10.4