diff --git a/src/compile.cpp b/src/compile.cpp
index 29eb342760..93f5ca543d 100644
--- a/src/compile.cpp
+++ b/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);
@@ -785,6 +785,10 @@ class Context {
     {
       if (size == 8) {
         switch(op) {
+        case Absolute:
+          assert(t, resultSize == 8);
+          return local::getThunk(t, absoluteLongThunk);
+
         case FloatNegate:
           assert(t, resultSize == 8);
           return local::getThunk(t, negateDoubleThunk);
@@ -819,12 +823,16 @@ class Context {
         assert(t, size == 4);
 
         switch(op) {
+        case Absolute:
+          assert(t, resultSize == 4);
+          return local::getThunk(t, absoluteIntThunk);
+
         case FloatNegate:
-          assert(t, size == 4);
+          assert(t, resultSize == 4);
           return local::getThunk(t, negateFloatThunk);
 
         case FloatAbsolute:
-          assert(t, size == 4);
+          assert(t, resultSize == 4);
           return local::getThunk(t, absoluteFloatThunk);
 
         case Float2Float:
@@ -2160,6 +2168,18 @@ absoluteFloat(uint32_t a)
   return floatToBits(fabsf(bitsToFloat(a)));
 }
 
+int64_t
+absoluteLong(int64_t a)
+{
+  return a > 0 ? a : -a;
+}
+
+int64_t
+absoluteInt(int32_t a)
+{
+  return a > 0 ? a : -a;
+}
+
 int64_t
 divideLong(int64_t b, int64_t a)
 {
@@ -5878,7 +5898,7 @@ 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 +7489,7 @@ 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 +8249,7 @@ 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);
 
diff --git a/src/compiler.cpp b/src/compiler.cpp
index a4e9540b1b..311d4a288c 100644
--- a/src/compiler.cpp
+++ b/src/compiler.cpp
@@ -1476,12 +1476,19 @@ pickTarget(Context* c, Read* read, bool intersectRead,
 
   Value* value = read->value;
 
-  uint32_t registerMask = (value->type == ValueGeneral
-                           ? c->arch->generalRegisterMask() : ~0);
+  uint32_t registerMask
+    = (value->type == ValueFloat ? ~0 : c->arch->generalRegisterMask());
 
   SiteMask mask(~0, registerMask, AnyFrameIndex);
   read->intersect(&mask);
 
+  if (value->type == ValueFloat) {
+    uint32_t floatMask = mask.registerMask & c->arch->floatRegisterMask();
+    if (floatMask) {
+      mask.registerMask = floatMask;
+    }
+  }
+
   Target best;
 
   Value* successor = read->successor();
@@ -1792,8 +1799,21 @@ class RegisterSite: public Site {
   }
 
   virtual bool matchNextWord(Context* c, Site* s, unsigned) {
-    return s->type(c) == RegisterOperand
-      and s->registerSize(c) == BytesPerWord;
+    assert(c, number != NoRegister);
+
+    if (s->type(c) != RegisterOperand) {
+      return false;
+    }
+
+    RegisterSite* rs = static_cast<RegisterSite*>(s);
+    unsigned size = rs->registerSize(c);
+    if (size > BytesPerWord) {
+      assert(c, number != NoRegister);
+      return number == rs->number;
+    } else {
+      uint32_t mask = c->arch->generalRegisterMask();
+      return ((1 << number) & mask) and ((1 << rs->number) & mask);
+    }
   }
 
   virtual void acquire(Context* c, Value* v) {
@@ -1876,6 +1896,9 @@ class RegisterSite: public Site {
   }
 
   virtual Site* makeNextWord(Context* c, unsigned) {
+    assert(c, number != NoRegister);
+    assert(c, ((1 << number) & c->arch->generalRegisterMask()));
+
     return freeRegisterSite(c, c->arch->generalRegisterMask());    
   }
 
@@ -1883,8 +1906,16 @@ class RegisterSite: public Site {
     return SiteMask(1 << RegisterOperand, mask_, NoFrameIndex);
   }
 
-  virtual SiteMask nextWordMask(Context*, unsigned) {
-    return SiteMask(1 << RegisterOperand, ~0, NoFrameIndex);
+  virtual SiteMask nextWordMask(Context* c, unsigned) {
+    assert(c, number != NoRegister);
+
+    if (registerSize(c) > BytesPerWord) {
+      return SiteMask
+        (1 << RegisterOperand, number, NoFrameIndex);
+    } else {
+      return SiteMask
+        (1 << RegisterOperand, c->arch->generalRegisterMask(), NoFrameIndex);
+    }
   }
 
   virtual unsigned registerSize(Context* c) {
@@ -1965,12 +1996,14 @@ class MemorySite: public Site {
     assert(c, acquired);
 
     if (mask.typeMask & (1 << MemoryOperand)) {
-      if (base == c->arch->stack()) {
-        assert(c, index == NoRegister);
-        return mask.frameIndex == AnyFrameIndex
-          or (mask.frameIndex != NoFrameIndex
-              and static_cast<int>(frameIndexToOffset(c, mask.frameIndex))
-              == offset);
+      if (mask.frameIndex >= 0) {
+        if (base == c->arch->stack()) {
+          assert(c, index == NoRegister);
+          return static_cast<int>(frameIndexToOffset(c, mask.frameIndex))
+            == offset;
+        } else {
+          return false;
+        }
       } else {
         return true;
       }
@@ -2287,13 +2320,17 @@ acceptMatch(Context* c, Site* s, Read*, const SiteMask& mask)
 
 Site*
 pickSourceSite(Context* c, Read* read, Site* target = 0,
-               unsigned* cost = 0, uint8_t typeMask = ~0,
+               unsigned* cost = 0, SiteMask* extraMask = 0,
                bool intersectRead = true, bool includeBuddies = true,
                bool includeNextWord = true,
                bool (*accept)(Context*, Site*, Read*, const SiteMask&)
                = acceptMatch)
 {
-  SiteMask mask(typeMask, ~0, AnyFrameIndex);
+  SiteMask mask;
+
+  if (extraMask) {
+    mask = intersect(mask, *extraMask);
+  }
 
   if (intersectRead) {
     read->intersect(&mask);
@@ -2444,7 +2481,7 @@ pickSiteOrMove(Context* c, Read* read, bool intersectRead,
                bool includeNextWord, unsigned registerReserveCount = 0)
 {
   Site* s = pickSourceSite
-    (c, read, 0, 0, ~0, intersectRead, true, includeNextWord);
+    (c, read, 0, 0, 0, intersectRead, true, includeNextWord);
   
   if (s) {
     return s;
@@ -3533,22 +3570,51 @@ maybeMove(Context* c, BinaryOperation type, unsigned srcSize,
   }
 }
 
-void
-pickSiteOrMove(Context* c, Value* src, Value* dst)
+Site*
+pickMatchOrMove(Context* c, Read* r, Site* nextWord, unsigned index,
+                bool intersectRead)
+{
+  Site* s = pickSite(c, r->value, nextWord, index, true);
+  SiteMask mask;
+  if (intersectRead) {
+    r->intersect(&mask);
+  }
+  if (s and s->match(c, mask)) {
+    return s;
+  }
+
+  return pickSiteOrMove
+    (c, r->value, intersect(mask, nextWord->nextWordMask(c, index)),
+     true, true);
+}
+
+Site*
+pickSiteOrMove(Context* c, Value* src, Value* dst, Site* nextWord,
+               unsigned index)
 {
   if (live(dst)) {
     Read* read = live(src);
-    Site* s = pickSourceSite(c, read, 0, 0, ~0, false, true, true);
+    Site* s;
+    if (nextWord) {
+      s = pickMatchOrMove(c, read, nextWord, index, false);
+    } else {
+      s = pickSourceSite(c, read, 0, 0, 0, false, true, true);
 
-    if (s == 0 or s->isVolatile(c)) {
-      maybeMove(c, read, false, true);
+      if (s == 0 or s->isVolatile(c)) {
+        s = maybeMove(c, read, false, true);
+      }
     }
+    assert(c, s);
 
     addBuddy(src, dst);
 
     if (src->source->isVolatile(c)) {
       removeSite(c, src, src->source);
     }
+
+    return s;
+  } else {
+    return 0;
   }
 }
 
@@ -3645,6 +3711,13 @@ class MoveEvent: public Event {
         {
           apply(c, Move, srcSelectSize, src->source, src->source,
                 dstSize, dst->target, dst->target);
+
+          if (live(dst) == 0) {
+            removeSite(c, dst, dst->target);
+            if (dstSize > BytesPerWord) {
+              removeSite(c, dst->nextWord, dst->nextWord->target);
+            }
+          }
         } else {
           maybeMove(c, Move, BytesPerWord, BytesPerWord, src,
                     BytesPerWord, dst, dstLowMask);
@@ -3654,9 +3727,9 @@ class MoveEvent: public Event {
           }
         }
       } else {
-        pickSiteOrMove(c, src, dst);
+        Site* low = pickSiteOrMove(c, src, dst, 0, 0);
         if (dstSize > BytesPerWord) {
-          pickSiteOrMove(c, src->nextWord, dst->nextWord);
+          pickSiteOrMove(c, src->nextWord, dst->nextWord, low, 1);
         }
       }
     } else if (srcSelectSize <= BytesPerWord and dstSize <= BytesPerWord) {
@@ -3714,7 +3787,7 @@ class MoveEvent: public Event {
 
         low->thaw(c, dst);
       } else {
-        pickSiteOrMove(c, src, dst);
+        pickSiteOrMove(c, src, dst, 0, 0);
       }
     }
 
@@ -5065,16 +5138,7 @@ readSource(Context* c, Read* r)
 
   Value* high = r->high(c);
   if (high) {
-    Site* s = pickSite(c, r->value, high->source, 0, true);
-    SiteMask mask;
-    r->intersect(&mask);
-    if (s and s->match(c, mask)) {
-      return s;
-    } else {
-      return pickSiteOrMove
-        (c, r->value, intersect(mask, high->source->nextWordMask(c, 0)),
-         true, true);
-    }
+    return pickMatchOrMove(c, r, high->source, 0, true);
   } else {
     return pickSiteOrMove(c, r, true, true);
   }
@@ -5210,13 +5274,14 @@ resolveSourceSites(Context* c, Event* e, SiteRecordList* frozen, Site** sites)
     Read* r = live(v);
 
     if (r and sites[el.localIndex] == 0) {
-      const uint32_t mask = (1 << RegisterOperand) | (1 << MemoryOperand);
+      SiteMask mask((1 << RegisterOperand) | (1 << MemoryOperand),
+                    c->arch->generalRegisterMask(), AnyFrameIndex);
 
       Site* s = pickSourceSite
-        (c, r, 0, 0, mask, true, false, true, acceptForResolve);
+        (c, r, 0, 0, &mask, true, false, true, acceptForResolve);
       if (s == 0) {
         s = pickSourceSite
-          (c, r, 0, 0, mask, false, false, true, acceptForResolve);
+          (c, r, 0, 0, &mask, false, false, true, acceptForResolve);
       }
 
       if (s) {
@@ -5247,15 +5312,16 @@ resolveTargetSites(Context* c, Event* e, SiteRecordList* frozen, Site** sites)
     Read* r = live(v);
 
     if (r and sites[el.localIndex] == 0) {
-      const uint32_t mask = (1 << RegisterOperand) | (1 << MemoryOperand);
+      SiteMask mask((1 << RegisterOperand) | (1 << MemoryOperand),
+                    c->arch->generalRegisterMask(), AnyFrameIndex);
 
       Site* s = pickSourceSite
-        (c, r, 0, 0, mask, true, true, true, acceptForResolve);
+        (c, r, 0, 0, &mask, true, true, true, acceptForResolve);
       if (s == 0) {
         s = pickSourceSite
-          (c, r, 0, 0, mask, false, true, true, acceptForResolve);
+          (c, r, 0, 0, &mask, false, true, true, acceptForResolve);
         if (s == 0) {
-          s = maybeMove(c, r, false, true, ResolveRegisterReserveCount);
+          s = maybeMove(c, v, mask, false, true, ResolveRegisterReserveCount);
         }
       }
 
diff --git a/src/machine.cpp b/src/machine.cpp
index 9abe61301d..4fa9aaf9b5 100644
--- a/src/machine.cpp
+++ b/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);
diff --git a/src/powerpc.cpp b/src/powerpc.cpp
index 63ee6da3b4..9cec9a1cdd 100644
--- a/src/powerpc.cpp
+++ b/src/powerpc.cpp
@@ -2065,6 +2065,9 @@ class MyArchitecture: public Assembler::Architecture {
       *aTypeMask = (1 << RegisterOperand);
       break;
 
+    case Absolute:
+    case FloatAbsolute:
+    case FloatSquareRoot:
     case FloatNegate:
     case Float2Float:
     case Float2Int:
@@ -2096,9 +2099,9 @@ class MyArchitecture: public Assembler::Architecture {
   }
 
   virtual void planMove
-  (unsigned size, uint8_t* srcTypeMask, uint64_t* srcRegisterMask,
+  (unsigned, uint8_t* srcTypeMask, uint64_t* srcRegisterMask,
    uint8_t* tmpTypeMask, uint64_t* tmpRegisterMask,
-   uint8_t dstTypeMask, uint64_t dstRegisterMask)
+   uint8_t dstTypeMask, uint64_t)
   {
     *srcTypeMask = ~0;
     *srcRegisterMask = ~static_cast<uint64_t>(0);
diff --git a/src/powerpc.h b/src/powerpc.h
index 1ef1437e0f..9a55c13266 100644
--- a/src/powerpc.h
+++ b/src/powerpc.h
@@ -122,7 +122,7 @@ atomicCompareAndSwap32(uint32_t* p, uint32_t old, uint32_t new_)
 inline bool
 atomicCompareAndSwap(uintptr_t* p, uintptr_t old, uintptr_t new_)
 {
-  return atomicCompareAndSwap32(p, old, new_);
+  return atomicCompareAndSwap32(reinterpret_cast<uint32_t*>(p), old, new_);
 }
 #endif // USE_ATOMIC_OPERATIONS
 
diff --git a/src/thunks.cpp b/src/thunks.cpp
index bab06da664..f55c0018a9 100644
--- a/src/thunks.cpp
+++ b/src/thunks.cpp
@@ -22,6 +22,8 @@ THUNK(divideFloat)
 THUNK(moduloFloat)
 THUNK(negateFloat)
 THUNK(absoluteFloat)
+THUNK(absoluteLong)
+THUNK(absoluteInt)
 THUNK(divideLong)
 THUNK(divideInt)
 THUNK(moduloLong)
diff --git a/src/x86.cpp b/src/x86.cpp
index 7e565dc179..b8290389e4 100644
--- a/src/x86.cpp
+++ b/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)
 {
@@ -954,12 +978,12 @@ sseMoveRR(Context* c, unsigned aSize, Assembler::Register* a,
       opcode(c, 0xf3);
       maybeRex(c, 4, a, b);
       opcode(c, 0x0f, 0x10);
-      modrm(c, 0xc0, b, a);
+      modrm(c, 0xc0, a, b);
     } else {
       opcode(c, 0xf2);
-      maybeRex(c, 4, a, b);
+      maybeRex(c, 8, a, b);
       opcode(c, 0x0f, 0x10);
-      modrm(c, 0xc0, b, a);
+      modrm(c, 0xc0, a, b);
     } 
   } else if (floatReg(a)) {
     opcode(c, 0x66);
@@ -1094,7 +1118,6 @@ sseMoveMR(Context* c, unsigned aSize, Assembler::Memory* a,
           unsigned bSize UNUSED, Assembler::Register* b)
 {
   assert(c, aSize >= 4);
-  assert(c, aSize == bSize);
 
   if (BytesPerWord == 4 and aSize == 8) {
     opcode(c, 0xf3);
@@ -2503,7 +2526,7 @@ 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);
@@ -2896,9 +2919,13 @@ class MyArchitecture: public Assembler::Architecture {
       break;
 
     case FloatAbsolute:
-      *aTypeMask = (1 << RegisterOperand);
-      *aRegisterMask = (static_cast<uint64_t>(FloatRegisterMask) << 32)
-        | FloatRegisterMask;
+      if (useSSE(&c)) {
+        *aTypeMask = (1 << RegisterOperand);
+        *aRegisterMask = (static_cast<uint64_t>(FloatRegisterMask) << 32)
+          | FloatRegisterMask;
+      } else {
+        *thunk = true;
+      }
       break;  
   
     case FloatNegate:
@@ -2912,9 +2939,13 @@ class MyArchitecture: public Assembler::Architecture {
       break;
 
     case FloatSquareRoot:
-      *aTypeMask = (1 << RegisterOperand) | (1 << MemoryOperand);
-      *aRegisterMask = (static_cast<uint64_t>(FloatRegisterMask) << 32)
-        | FloatRegisterMask;
+      if (useSSE(&c)) {
+        *aTypeMask = (1 << RegisterOperand) | (1 << MemoryOperand);
+        *aRegisterMask = (static_cast<uint64_t>(FloatRegisterMask) << 32)
+          | FloatRegisterMask;
+      } else {
+        *thunk = true;
+      }
       break;
 
     case Float2Float:
@@ -3225,7 +3256,7 @@ 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) {
diff --git a/src/x86.h b/src/x86.h
index 0528d33695..4ae08f6859 100644
--- a/src/x86.h
+++ b/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
@@ -233,9 +233,9 @@ inline bool
 atomicCompareAndSwap(uintptr_t* p, uintptr_t old, uintptr_t new_)
 {
 #ifdef ARCH_x86_32
-  return atomicCompareAndSwap32(p, old, new_);
+  return atomicCompareAndSwap32(reinterpret_cast<uint32_t*>(p), old, new_);
 #elif defined ARCH_x86_64
-  return atomicCompareAndSwap64(p, old, new_);
+  return atomicCompareAndSwap64(reinterpret_cast<uint64_t*>(p), old, new_);
 #endif // ARCH_x86_64
 }
 #endif // USE_ATOMIC_OPERATIONS