refine memory barrier implementation and usage

src/compile.cpp

@@ -235,7 +235,7 @@ methodForIp(MyThread* t, void* ip)
   // we must use a version of the method tree at least as recent as the
   // compiled form of the method containing the specified address (see
   // compile(MyThread*, Allocator*, BootContext*, object)):
-  memoryBarrier();
+  loadMemoryBarrier();
 
   return treeQuery(t, methodTree(t), reinterpret_cast<intptr_t>(ip),
                    methodTreeSentinal(t), compareIpToMethodBounds);
@@ -5878,7 +5878,7 @@ resolveNative(MyThread* t, object method)
     // methodCompiled, since we don't want them using the slow calling
     // convention on a function that expects the fast calling
     // convention:
-    memoryBarrier();
+    storeStoreMemoryBarrier();
 
     methodCompiled(t, method) = reinterpret_cast<uintptr_t>(function);
   }
@@ -7469,7 +7469,7 @@ findCallNode(MyThread* t, void* address)
   // we must use a version of the call table at least as recent as the
   // compiled form of the method containing the specified address (see
   // compile(MyThread*, Allocator*, BootContext*, object)):
-  memoryBarrier();
+  loadMemoryBarrier();
 
   MyProcessor* p = processor(t);
   object table = p->callTable;
@@ -8229,7 +8229,7 @@ compile(MyThread* t, Allocator* allocator, BootContext* bootContext,
          reinterpret_cast<intptr_t>(compiled), clone, methodTreeSentinal(t),
          compareIpToMethodBounds);
 
-      memoryBarrier();
+      storeStoreMemoryBarrier();
 
       methodCompiled(t, method) = reinterpret_cast<intptr_t>(compiled);
 

src/machine.cpp

@@ -2334,11 +2334,11 @@ enter(Thread* t, Thread::State s)
 #ifdef USE_ATOMIC_OPERATIONS
 #  define INCREMENT atomicIncrement
 #  define ACQUIRE_LOCK ACQUIRE_RAW(t, t->m->stateLock)
-#  define BARRIER memoryBarrier()
+#  define STORE_LOAD_MEMORY_BARRIER storeLoadMemoryBarrier()
 #else
 #  define INCREMENT(pointer, value) *(pointer) += value;
 #  define ACQUIRE_LOCK
-#  define BARRIER
+#  define STORE_LOAD_MEMORY_BARRIER
 
   ACQUIRE_RAW(t, t->m->stateLock);
 #endif // not USE_ATOMIC_OPERATIONS
@@ -2366,7 +2366,7 @@ enter(Thread* t, Thread::State s)
     t->state = Thread::ExclusiveState;
     t->m->exclusive = t;
     
-    BARRIER;
+    STORE_LOAD_MEMORY_BARRIER;
 
     while (t->m->activeCount > 1) {
       t->m->stateLock->wait(t->systemThread, 0);

src/x86.cpp

@@ -113,23 +113,7 @@ class MyBlock: public Assembler::Block {
   unsigned size;
 };
 
-class Context {
- public:
-  Context(System* s, Allocator* a, Zone* zone):
-    s(s), zone(zone), client(0), code(s, a, 1024), tasks(0), result(0),
-    firstBlock(new (zone->allocate(sizeof(MyBlock))) MyBlock(0)),
-    lastBlock(firstBlock)
-  { }
-
-  System* s;
-  Zone* zone;
-  Assembler::Client* client;
-  Vector code;
-  Task* tasks;
-  uint8_t* result;
-  MyBlock* firstBlock;
-  MyBlock* lastBlock;
-};
+class Context;
 
 typedef void (*OperationType)(Context*);
 
@@ -163,6 +147,25 @@ class ArchitectureContext {
    * OperandTypeCount];
 };
 
+class Context {
+ public:
+  Context(System* s, Allocator* a, Zone* zone, ArchitectureContext* ac):
+    s(s), zone(zone), client(0), code(s, a, 1024), tasks(0), result(0),
+    firstBlock(new (zone->allocate(sizeof(MyBlock))) MyBlock(0)),
+    lastBlock(firstBlock), ac(ac)
+  { }
+
+  System* s;
+  Zone* zone;
+  Assembler::Client* client;
+  Vector code;
+  Task* tasks;
+  uint8_t* result;
+  MyBlock* firstBlock;
+  MyBlock* lastBlock;
+  ArchitectureContext* ac;
+};
+
 void NO_RETURN
 abort(Context* c)
 {
@@ -620,6 +623,27 @@ void
 ignore(Context*)
 { }
 
+void
+storeLoadBarrier(Context* c)
+{
+  if (useSSE(c->ac)) {
+    // mfence:
+    c->code.append(0x0f);
+    c->code.append(0xae);
+    c->code.append(0xf0);
+  } else {
+    // lock addq $0x0,(%rsp):
+    c->code.append(0xf0);
+    if (BytesPerWord == 8) {
+      c->code.append(0x48);
+    }
+    c->code.append(0x83);
+    c->code.append(0x04);
+    c->code.append(0x24);
+    c->code.append(0x00);    
+  }
+}
+
 void
 unconditional(Context* c, unsigned jump, Assembler::Constant* a)
 {
@@ -2503,7 +2527,7 @@ populateTables(ArchitectureContext* c)
   zo[Return] = return_;
   zo[LoadBarrier] = ignore;
   zo[StoreStoreBarrier] = ignore;
-  zo[StoreLoadBarrier] = ignore;
+  zo[StoreLoadBarrier] = storeLoadBarrier;
 
   uo[index(c, Call, C)] = CAST1(callC);
   uo[index(c, Call, R)] = CAST1(callR);
@@ -3225,7 +3249,7 @@ class MyArchitecture: public Assembler::Architecture {
 class MyAssembler: public Assembler {
  public:
   MyAssembler(System* s, Allocator* a, Zone* zone, MyArchitecture* arch):
-    c(s, a, zone), arch_(arch)
+    c(s, a, zone, &(arch->c)), arch_(arch)
   { }
 
   virtual void setClient(Client* client) {

src/x86.h

@@ -169,7 +169,7 @@ memoryBarrier()
 inline void
 storeStoreMemoryBarrier()
 {
-  memoryBarrier();
+  __asm__ __volatile__("": : :"memory");
 }
 
 inline void
@@ -181,13 +181,13 @@ storeLoadMemoryBarrier()
 inline void
 loadMemoryBarrier()
 {
-  memoryBarrier();
+  __asm__ __volatile__("": : :"memory");
 }
 
 inline void
 syncInstructionCache(const void*, unsigned)
 {
-  // ignore
+  __asm__ __volatile__("": : :"memory");
 }
 
 #ifdef USE_ATOMIC_OPERATIONS