Fix bug in ParallelMoveResolver::EmitSwap: implement swaps of FPU spill slots.

runtime/vm/flow_graph_compiler_x64.cc

 // Copyright (c) 2013, the Dart project authors.  Please see the AUTHORS file
 // for details. All rights reserved. Use of this source code is governed by a
 // BSD-style license that can be found in the LICENSE file.
 
 #include "vm/globals.h"  // Needed here to get TARGET_ARCH_X64.
 #if defined(TARGET_ARCH_X64)
 
 #include "vm/flow_graph_compiler.h"
 
 #include "lib/error.h"
@@ skipping 1578 matching lines @@
     }
   } else if (source.IsFpuRegister()) {
     if (destination.IsFpuRegister()) {
       // Optimization manual recommends using MOVAPS for register
       // to register moves.
       __ movaps(destination.fpu_reg(), source.fpu_reg());
     } else {
       if (destination.IsDoubleStackSlot()) {
         __ movsd(destination.ToStackSlotAddress(), source.fpu_reg());
       } else {
-        ASSERT(destination.IsFloat32x4StackSlot() ||
-               destination.IsUint32x4StackSlot());
+        ASSERT(destination.IsQuadStackSlot());
         __ movups(destination.ToStackSlotAddress(), source.fpu_reg());
       }
     }
   } else if (source.IsDoubleStackSlot()) {
     if (destination.IsFpuRegister()) {
       __ movsd(destination.fpu_reg(), source.ToStackSlotAddress());
     } else {
       ASSERT(destination.IsDoubleStackSlot());
       __ movsd(XMM0, source.ToStackSlotAddress());
       __ movsd(destination.ToStackSlotAddress(), XMM0);
     }
-  } else if (source.IsFloat32x4StackSlot() || source.IsUint32x4StackSlot()) {
+  } else if (source.IsQuadStackSlot()) {
     if (destination.IsFpuRegister()) {
       __ movups(destination.fpu_reg(), source.ToStackSlotAddress());
     } else {
-      ASSERT(destination.IsFloat32x4StackSlot() ||
-             destination.IsUint32x4StackSlot());
+      ASSERT(destination.IsQuadStackSlot());
       __ movups(XMM0, source.ToStackSlotAddress());
       __ movups(destination.ToStackSlotAddress(), XMM0);
     }
   } else {
     ASSERT(source.IsConstant());
     if (destination.IsRegister()) {
       const Object& constant = source.constant();
       if (constant.IsSmi() && (Smi::Cast(constant).Value() == 0)) {
         __ xorq(destination.reg(), destination.reg());
       } else {
@@ skipping 21 matching lines @@
   } else if (source.IsStackSlot() && destination.IsRegister()) {
     Exchange(destination.reg(), source.ToStackSlotAddress());
   } else if (source.IsStackSlot() && destination.IsStackSlot()) {
     Exchange(destination.ToStackSlotAddress(), source.ToStackSlotAddress());
   } else if (source.IsFpuRegister() && destination.IsFpuRegister()) {
     __ movaps(XMM0, source.fpu_reg());
     __ movaps(source.fpu_reg(), destination.fpu_reg());
     __ movaps(destination.fpu_reg(), XMM0);
   } else if (source.IsFpuRegister() || destination.IsFpuRegister()) {
     ASSERT(destination.IsDoubleStackSlot() ||
-           destination.IsFloat32x4StackSlot() ||
-           destination.IsUint32x4StackSlot() ||
+           destination.IsQuadStackSlot() ||
            source.IsDoubleStackSlot() ||
-           source.IsFloat32x4StackSlot() ||
-           source.IsUint32x4StackSlot());
+           source.IsQuadStackSlot());
     bool double_width = destination.IsDoubleStackSlot() ||
                         source.IsDoubleStackSlot();
     XmmRegister reg = source.IsFpuRegister() ? source.fpu_reg()
                                              : destination.fpu_reg();
     Address slot_address = source.IsFpuRegister()
         ? destination.ToStackSlotAddress()
         : source.ToStackSlotAddress();
 
     if (double_width) {
       __ movsd(XMM0, slot_address);
       __ movsd(slot_address, reg);
     } else {
       __ movups(XMM0, slot_address);
       __ movups(slot_address, reg);
     }
     __ movaps(reg, XMM0);
+  } else if (source.IsDoubleStackSlot() && destination.IsDoubleStackSlot()) {
+    const Address& source_slot_address = source.ToStackSlotAddress();
+    const Address& destination_slot_address = destination.ToStackSlotAddress();
+
+    ScratchFpuRegisterScope ensure_scratch(this);
+    __ movsd(XMM0, source_slot_address);
+    __ movsd(ensure_scratch.reg(), destination_slot_address);
+    __ movsd(destination_slot_address, XMM0);
+    __ movsd(source_slot_address, ensure_scratch.reg());
+  } else if (source.IsQuadStackSlot() && destination.IsQuadStackSlot()) {
+    const Address& source_slot_address = source.ToStackSlotAddress();
+    const Address& destination_slot_address = destination.ToStackSlotAddress();
+
+    ScratchFpuRegisterScope ensure_scratch(this);
+    __ movups(XMM0, source_slot_address);
+    __ movups(ensure_scratch.reg(), destination_slot_address);
+    __ movups(destination_slot_address, XMM0);
+    __ movups(source_slot_address, ensure_scratch.reg());
   } else {
     UNREACHABLE();
   }
 
   // The swap of source and destination has executed a move from source to
   // destination.
   move->Eliminate();
 
   // Any unperformed (including pending) move with a source of either
   // this move's source or destination needs to have their source
@@ skipping 23 matching lines @@
 void ParallelMoveResolver::Exchange(Register reg, const Address& mem) {
   __ Exchange(reg, mem);
 }
 
 
 void ParallelMoveResolver::Exchange(const Address& mem1, const Address& mem2) {
   __ Exchange(mem1, mem2);
 }
 
 
+void ParallelMoveResolver::SpillScratch(Register reg) {
+  __ pushq(reg);
+}
+
+
+void ParallelMoveResolver::RestoreScratch(Register reg) {
+  __ popq(reg);
+}
+
+
+void ParallelMoveResolver::SpillFpuScratch(FpuRegister reg) {
+  __ subq(RSP, Immediate(kFpuRegisterSize));
+  __ movups(Address(RSP, 0), reg);
+}
+
+
+void ParallelMoveResolver::RestoreFpuScratch(FpuRegister reg) {
+  __ movups(reg, Address(RSP, 0));
+  __ addq(RSP, Immediate(kFpuRegisterSize));
+}
+
+
 #undef __
 
 }  // namespace dart
 
 #endif  // defined TARGET_ARCH_X64