Experimental implementation: Exposing SIMD instructions into JavaScript

src/x64/lithium-gap-resolver-x64.cc

 // Copyright 2011 the V8 project authors. All rights reserved.
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions are
 // met:
 //
 //     * Redistributions of source code must retain the above copyright
 //       notice, this list of conditions and the following disclaimer.
 //     * Redistributions in binary form must reproduce the above
 //       copyright notice, this list of conditions and the following
 //       disclaimer in the documentation and/or other materials provided
@@ skipping 226 matching lines @@
     }
   } else if (source->IsDoubleStackSlot()) {
     Operand src = cgen_->ToOperand(source);
     if (destination->IsDoubleRegister()) {
       __ movsd(cgen_->ToDoubleRegister(destination), src);
     } else {
       ASSERT(destination->IsDoubleStackSlot());
       __ movsd(xmm0, src);
       __ movsd(cgen_->ToOperand(destination), xmm0);
     }
+  } else if (source->IsFloat32x4Register()) {
+    XMMRegister src = cgen_->ToFloat32x4Register(source);
+    if (destination->IsFloat32x4Register()) {
+      __ movaps(cgen_->ToFloat32x4Register(destination), src);
+    } else {
+      ASSERT(destination->IsFloat32x4StackSlot());
+      __ movups(cgen_->ToOperand(destination), src);
+    }
+  } else if (source->IsFloat32x4StackSlot()) {
+    Operand src = cgen_->ToOperand(source);
+    if (destination->IsFloat32x4Register()) {
+      __ movups(cgen_->ToFloat32x4Register(destination), src);
+    } else {
+      ASSERT(destination->IsFloat32x4StackSlot());
+      __ movups(xmm0, src);
+      __ movups(cgen_->ToOperand(destination), xmm0);
+    }
+  } else if (source->IsInt32x4Register()) {
+    XMMRegister src = cgen_->ToInt32x4Register(source);
+    if (destination->IsInt32x4Register()) {
+      __ movaps(cgen_->ToInt32x4Register(destination), src);
+    } else {
+      ASSERT(destination->IsInt32x4StackSlot());
+      __ movups(cgen_->ToOperand(destination), src);
+    }
+  } else if (source->IsInt32x4StackSlot()) {
+    Operand src = cgen_->ToOperand(source);
+    if (destination->IsInt32x4Register()) {
+      __ movups(cgen_->ToInt32x4Register(destination), src);
+    } else {
+      ASSERT(destination->IsInt32x4StackSlot());
+      __ movups(xmm0, src);
+      __ movups(cgen_->ToOperand(destination), xmm0);
+    }
   } else {
     UNREACHABLE();
   }
 
   moves_[index].Eliminate();
 }
 
 
 void LGapResolver::EmitSwap(int index) {
   LOperand* source = moves_[index].source();
@@ skipping 21 matching lines @@
   } else if ((source->IsStackSlot() && destination->IsStackSlot()) ||
       (source->IsDoubleStackSlot() && destination->IsDoubleStackSlot())) {
     // Swap two stack slots or two double stack slots.
     Operand src = cgen_->ToOperand(source);
     Operand dst = cgen_->ToOperand(destination);
     __ movsd(xmm0, src);
     __ movq(kScratchRegister, dst);
     __ movsd(dst, xmm0);
     __ movq(src, kScratchRegister);
 
+  } else if ((source->IsFloat32x4StackSlot() &&
+              destination->IsFloat32x4StackSlot()) ||
+             ((source->IsInt32x4StackSlot() &&
+               destination->IsInt32x4StackSlot()))) {
+    // Swap two XMM stack slots.
+    Operand src = cgen_->ToOperand(source);
+    Operand dst = cgen_->ToOperand(destination);
+    __ movups(xmm0, src);
+    __ movq(kScratchRegister, dst);
+    __ movq(src, kScratchRegister);
+    __ movq(kScratchRegister, Operand(dst, kDoubleSize));
+    __ movq(Operand(src, kDoubleSize), kScratchRegister);
+    __ movups(dst, xmm0);
+
   } else if (source->IsDoubleRegister() && destination->IsDoubleRegister()) {
     // Swap two double registers.
     XMMRegister source_reg = cgen_->ToDoubleRegister(source);
     XMMRegister destination_reg = cgen_->ToDoubleRegister(destination);
     __ movaps(xmm0, source_reg);
     __ movaps(source_reg, destination_reg);
     __ movaps(destination_reg, xmm0);
 
+  } else if (source->IsFloat32x4Register() &&
+             destination->IsFloat32x4Register()) {
+    // Swap two XMM registers.
+    XMMRegister source_reg = cgen_->ToFloat32x4Register(source);
+    XMMRegister destination_reg = cgen_->ToFloat32x4Register(destination);
+    __ movaps(xmm0, source_reg);
+    __ movaps(source_reg, destination_reg);
+    __ movaps(destination_reg, xmm0);
+
+  } else if (source->IsInt32x4Register() &&
+             destination->IsInt32x4Register()) {
+    // Swap two XMM registers.
+    XMMRegister source_reg = cgen_->ToInt32x4Register(source);
+    XMMRegister destination_reg = cgen_->ToInt32x4Register(destination);
+    __ movaps(xmm0, source_reg);
+    __ movaps(source_reg, destination_reg);
+    __ movaps(destination_reg, xmm0);
+
   } else if (source->IsDoubleRegister() || destination->IsDoubleRegister()) {
     // Swap a double register and a double stack slot.
     ASSERT((source->IsDoubleRegister() && destination->IsDoubleStackSlot()) ||
            (source->IsDoubleStackSlot() && destination->IsDoubleRegister()));
     XMMRegister reg = cgen_->ToDoubleRegister(source->IsDoubleRegister()
                                                   ? source
                                                   : destination);
     LOperand* other = source->IsDoubleRegister() ? destination : source;
     ASSERT(other->IsDoubleStackSlot());
     Operand other_operand = cgen_->ToOperand(other);
     __ movsd(xmm0, other_operand);
     __ movsd(other_operand, reg);
     __ movaps(reg, xmm0);
 
+  } else if (source->IsFloat32x4Register() ||
+             destination->IsFloat32x4Register()) {
+    // Swap a xmm register and a xmm stack slot.
+    ASSERT((source->IsFloat32x4Register() &&
+            destination->IsFloat32x4StackSlot()) ||
+           (source->IsFloat32x4StackSlot() &&
+            destination->IsFloat32x4Register()));
+    XMMRegister reg = cgen_->ToFloat32x4Register(source->IsFloat32x4Register()
+                                                     ? source
+                                                     : destination);
+    LOperand* other = source->IsFloat32x4Register() ? destination : source;
+    ASSERT(other->IsFloat32x4StackSlot());
+    Operand other_operand = cgen_->ToOperand(other);
+    __ movups(xmm0, other_operand);
+    __ movups(other_operand, reg);
+    __ movups(reg, xmm0);
+
+  } else if (source->IsInt32x4Register() ||
+             destination->IsInt32x4Register()) {
+    // Swap a xmm register and a xmm stack slot.
+    ASSERT((source->IsInt32x4Register() &&
+            destination->IsInt32x4StackSlot()) ||
+           (source->IsInt32x4StackSlot() &&
+            destination->IsInt32x4Register()));
+    XMMRegister reg = cgen_->ToInt32x4Register(source->IsInt32x4Register()
+                                                     ? source
+                                                     : destination);
+    LOperand* other = source->IsInt32x4Register() ? destination : source;
+    ASSERT(other->IsInt32x4StackSlot());
+    Operand other_operand = cgen_->ToOperand(other);
+    __ movups(xmm0, other_operand);
+    __ movups(other_operand, reg);
+    __ movups(reg, xmm0);
+
   } else {
     // No other combinations are possible.
     UNREACHABLE();
   }
 
   // The swap of source and destination has executed a move from source to
   // destination.
   moves_[index].Eliminate();
 
   // Any unperformed (including pending) move with a source of either
   // this move's source or destination needs to have their source
   // changed to reflect the state of affairs after the swap.
   for (int i = 0; i < moves_.length(); ++i) {
     LMoveOperands other_move = moves_[i];
     if (other_move.Blocks(source)) {
       moves_[i].set_source(destination);
     } else if (other_move.Blocks(destination)) {
       moves_[i].set_source(source);
     }
   }
 }
 
 #undef __
 
 } }  // namespace v8::internal
 
 #endif  // V8_TARGET_ARCH_X64