Fix weak consistency bug in guard-aquire/guard-release

Change-Id: Iac6ff2df14ce5c900a0d4af90f475100dc5ad75f
Reviewed-on: http://gfwr801.rchland.ibm.com:8080/gerrit/78
Tested-by: Jenkins Server
Reviewed-by: Nicholas E. Bofferding <bofferdn@us.ibm.com>

1 files changed, 16 insertions, 31 deletions

diff --git a/src/libc++/builtins.C b/src/libc++/builtins.C
index a336c46c3..04ea744f9 100755
--- a/src/libc++/builtins.C
+++ b/src/libc++/builtins.C
@@ -1,6 +1,8 @@
 #include <stdint.h>
 #include <stdlib.h>
 
+#include <arch/ppc.H>
+
 void* operator new(size_t s)
 {
     return malloc(s);
@@ -33,50 +35,33 @@ void operator delete[](void* p)
 
 extern "C" int __cxa_guard_acquire(volatile uint64_t* gv)
 {
-    // 0 -> uninitialized
-    // 1 -> locked
-    // 2 -> unlocked and initialized
+    // States:
+    //     0 -> uninitialized
+    //     1 -> locked
+    //     2 -> unlocked and initialized
     
     uint32_t v = __sync_val_compare_and_swap((volatile uint32_t*)gv, 0, 1);
     if (v == 0)
 	return 1;
     if (v == 2)
 	return 0;
+    
+    // Wait for peer thread to perform initialization (state 2).
     while(2 != *(volatile uint32_t*)gv);
+    
+    // Instruction barrier to ensure value is set before later loads execute.
+    isync(); 
+    
     return 0;
-/*
-    register volatile void* guard = gv;
-    register uint32_t c = 0;
-
-    asm volatile(
-	"__cxa_guard_acquire_begin:"
-	"lwarx %0,0,%1;"	// Load guard with reserve
-	"cmpi 0,%0,0;"		// Compare with 0
-	"bne+ __cxa_guard_acquire_finish;"	// != 0, goto "finished"
-	"li %0, 1;"		// Set to 1.
-	"stwcx. %0,0,%1;"	// Store against reserve
-	"bne- __cxa_guard_acquire_begin;"	// goto begin if failed store.
-	"li %0, 3;"		// Set to 3 -> success in lock
-	"__cxa_guard_acquire_finish:" 
-	    : "+r" (c) : "r" (guard): "memory","cc"
-    );
-    while (2 > c)
-    {
-	asm volatile("lwz %0, 0(%1);" : "=r" (c) : "r" (guard));
-    }
-    return (3 == c ? 1 : 0);  // 3 means success in lock, return 1 (obtained)
-			      // 2 means initialized, return 0
-*/
 }
 
 extern "C" void __cxa_guard_release(volatile uint64_t* gv)
 {
+    // Memory barrier to ensure all preceeding writes have completed before
+    // releasing guard.
+    lwsync();
+
     (*(volatile uint32_t*)gv) = 2;
-    /*
-    register volatile void* guard = gv;
-    register uint32_t c = 2;
-    asm volatile("stw %0, 0(%1)" :: "r"(c) , "r" (guard): "memory");
-    */
     return;
 }