[Turbofan] Support 128 bit moves/swaps for x64.

src/compiler/x64/code-generator-x64.cc

Index: src/compiler/x64/code-generator-x64.cc
diff --git a/src/compiler/x64/code-generator-x64.cc b/src/compiler/x64/code-generator-x64.cc
index 4dc75592ec2b189a5581a5ce1f89399cd40a02b2..3826e17c144081a198c1d619373c45a937a442d9 100644
--- a/src/compiler/x64/code-generator-x64.cc
+++ b/src/compiler/x64/code-generator-x64.cc
@@ -2359,24 +2359,47 @@ void CodeGenerator::AssembleMove(InstructionOperand* source,
     }
   } else if (source->IsFPRegister()) {
     XMMRegister src = g.ToDoubleRegister(source);
+    MachineRepresentation rep = LocationOperand::cast(source)->representation();
     if (destination->IsFPRegister()) {
       XMMRegister dst = g.ToDoubleRegister(destination);
-      __ Movapd(dst, src);
+      if (rep != MachineRepresentation::kSimd128) {

[Benedikt Meurer, 2016/07/10 17:21:06]

Can't we always do Movapd here?

[bbudge, 2016/07/10 22:45:22]

You're right. Replaced original code.

+        __ Movapd(dst, src);
+      } else {
+        DCHECK_EQ(MachineRepresentation::kSimd128, rep);
+        __ movups(dst, src);
+      }
     } else {
       DCHECK(destination->IsFPStackSlot());
       Operand dst = g.ToOperand(destination);
-      __ Movsd(dst, src);
+      if (rep != MachineRepresentation::kSimd128) {
+        __ Movsd(dst, src);
+      } else {
+        DCHECK_EQ(MachineRepresentation::kSimd128, rep);
+        __ movups(dst, src);

[Benedikt Meurer, 2016/07/10 17:21:06]

You should introduce Movups, that uses AVX instructions if available (see what
we do for Movapd).

[bbudge, 2016/07/10 22:45:22]

Done.

+      }
     }
   } else if (source->IsFPStackSlot()) {
     DCHECK(destination->IsFPRegister() || destination->IsFPStackSlot());
     Operand src = g.ToOperand(source);
+    MachineRepresentation rep = LocationOperand::cast(source)->representation();
     if (destination->IsFPRegister()) {
       XMMRegister dst = g.ToDoubleRegister(destination);
-      __ Movsd(dst, src);
+      if (rep != MachineRepresentation::kSimd128) {
+        __ Movsd(dst, src);
+      } else {
+        DCHECK_EQ(MachineRepresentation::kSimd128, rep);
+        __ movups(dst, src);
+      }
     } else {
       Operand dst = g.ToOperand(destination);
-      __ Movsd(kScratchDoubleReg, src);
-      __ Movsd(dst, kScratchDoubleReg);
+      if (rep != MachineRepresentation::kSimd128) {
+        __ Movsd(kScratchDoubleReg, src);
+        __ Movsd(dst, kScratchDoubleReg);
+      } else {
+        DCHECK_EQ(MachineRepresentation::kSimd128, rep);
+        __ movups(kScratchDoubleReg, src);
+        __ movups(dst, kScratchDoubleReg);
+      }
     }
   } else {
     UNREACHABLE();
@@ -2408,31 +2431,60 @@ void CodeGenerator::AssembleSwap(InstructionOperand* source,
   } else if ((source->IsStackSlot() && destination->IsStackSlot()) ||
              (source->IsFPStackSlot() && destination->IsFPStackSlot())) {
     // Memory-memory.
-    Register tmp = kScratchRegister;
     Operand src = g.ToOperand(source);
     Operand dst = g.ToOperand(destination);
-    __ movq(tmp, dst);
-    __ pushq(src);
-    frame_access_state()->IncreaseSPDelta(1);
-    src = g.ToOperand(source);
-    __ movq(src, tmp);
-    frame_access_state()->IncreaseSPDelta(-1);
-    dst = g.ToOperand(destination);
-    __ popq(dst);
+    MachineRepresentation rep = LocationOperand::cast(source)->representation();
+    if (rep != MachineRepresentation::kSimd128) {
+      Register tmp = kScratchRegister;
+      __ movq(tmp, dst);
+      __ pushq(src);
+      frame_access_state()->IncreaseSPDelta(1);
+      src = g.ToOperand(source);
+      __ movq(src, tmp);
+      frame_access_state()->IncreaseSPDelta(-1);
+      dst = g.ToOperand(destination);
+      __ popq(dst);
+    } else {
+      DCHECK_EQ(MachineRepresentation::kSimd128, rep);
+      // Use the XOR trick to swap without a temporary.
+      __ movups(kScratchDoubleReg, src);
+      __ xorps(kScratchDoubleReg, dst);  // scratch contains src ^ dst.
+      __ movups(src, kScratchDoubleReg);
+      __ xorps(kScratchDoubleReg, dst);  // scratch contains src.
+      __ movups(dst, kScratchDoubleReg);
+      __ xorps(kScratchDoubleReg, src);  // scratch contains dst.
+      __ movups(src, kScratchDoubleReg);
+    }
   } else if (source->IsFPRegister() && destination->IsFPRegister()) {
     // XMM register-register swap.
     XMMRegister src = g.ToDoubleRegister(source);
     XMMRegister dst = g.ToDoubleRegister(destination);
-    __ Movapd(kScratchDoubleReg, src);
-    __ Movapd(src, dst);
-    __ Movapd(dst, kScratchDoubleReg);
+    MachineRepresentation rep = LocationOperand::cast(source)->representation();
+    if (rep != MachineRepresentation::kSimd128) {
+      __ Movapd(kScratchDoubleReg, src);
+      __ Movapd(src, dst);
+      __ Movapd(dst, kScratchDoubleReg);
+    } else {
+      DCHECK_EQ(MachineRepresentation::kSimd128, rep);
+      __ movups(kScratchDoubleReg, src);
+      __ movups(src, dst);
+      __ movups(dst, kScratchDoubleReg);
+    }
   } else if (source->IsFPRegister() && destination->IsFPStackSlot()) {
     // XMM register-memory swap.
     XMMRegister src = g.ToDoubleRegister(source);
     Operand dst = g.ToOperand(destination);
-    __ Movsd(kScratchDoubleReg, src);
-    __ Movsd(src, dst);
-    __ Movsd(dst, kScratchDoubleReg);
+    MachineRepresentation rep = LocationOperand::cast(source)->representation();
+    if (rep != MachineRepresentation::kSimd128) {
+      __ Movsd(kScratchDoubleReg, src);
+      __ Movsd(src, dst);
+      __ Movsd(dst, kScratchDoubleReg);
+    } else {
+      DCHECK_EQ(MachineRepresentation::kSimd128, rep);
+      __ movups(kScratchDoubleReg, src);
+      __ movups(src, dst);
+      __ movups(dst, kScratchDoubleReg);
+    }
   } else {
     // No other combinations are possible.
     UNREACHABLE();