Inline two argument shuffle operations on IA32 and X64.

runtime/vm/intermediate_language.h

 // 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.
 
 #ifndef VM_INTERMEDIATE_LANGUAGE_H_
 #define VM_INTERMEDIATE_LANGUAGE_H_
 
 #include "vm/allocation.h"
 #include "vm/ast.h"
 #include "vm/growable_array.h"
@@ skipping 90 matching lines @@
   V(_Float32x4, _reciprocalSqrt, Float32x4ReciprocalSqrt, 157194268)           \
   V(_Float32x4, _reciprocal, Float32x4Reciprocal, 157194268)                   \
   V(_Float32x4, _negate, Float32x4Negate, 157194268)                           \
   V(_Float32x4, _abs, Float32x4Absolute, 157194268)                            \
   V(_Float32x4, _clamp, Float32x4Clamp, 1553901850)                            \
   V(_Float32x4, withX, Float32x4WithX, 1816943014)                             \
   V(_Float32x4, withY, Float32x4WithY, 820404963)                              \
   V(_Float32x4, withZ, Float32x4WithZ, 881355277)                              \
   V(_Float32x4, withW, Float32x4WithW, 441497035)                              \
   V(_Float32x4, _toUint32x4, Float32x4ToUint32x4, 802289205)                   \
+  V(_Float32x4, withZWInXY, Float32x4WithZWInXY, 465519415)                    \
+  V(_Float32x4, interleaveXY, Float32x4InterleaveXY, 465519415)                \
+  V(_Float32x4, interleaveZW, Float32x4InterleaveZW, 465519415)                \
+  V(_Float32x4, interleaveXYPairs, Float32x4InterleaveXYPairs, 465519415)      \
+  V(_Float32x4, interleaveZWPairs, Float32x4InterleaveZWPairs, 465519415)      \
   V(Uint32x4, Uint32x4.bool, Uint32x4BoolConstructor, 487876159)               \
   V(_Uint32x4, get:flagX, Uint32x4GetFlagX, 782547529)                         \
   V(_Uint32x4, get:flagY, Uint32x4GetFlagY, 782547529)                         \
   V(_Uint32x4, get:flagZ, Uint32x4GetFlagZ, 782547529)                         \
   V(_Uint32x4, get:flagW, Uint32x4GetFlagW, 782547529)                         \
   V(_Uint32x4, select, Uint32x4Select, 810336099)                              \
   V(_Uint32x4, withFlagX, Uint32x4WithFlagX, 25547408)                         \
   V(_Uint32x4, withFlagY, Uint32x4WithFlagY, 1176493005)                       \
   V(_Uint32x4, withFlagZ, Uint32x4WithFlagZ, 1237443319)                       \
   V(_Uint32x4, withFlagW, Uint32x4WithFlagW, 797585077)                        \
@@ skipping 485 matching lines @@
   M(Float32x4Zero)                                                             \
   M(Float32x4Splat)                                                            \
   M(Float32x4Comparison)                                                       \
   M(Float32x4MinMax)                                                           \
   M(Float32x4Scale)                                                            \
   M(Float32x4Sqrt)                                                             \
   M(Float32x4ZeroArg)                                                          \
   M(Float32x4Clamp)                                                            \
   M(Float32x4With)                                                             \
   M(Float32x4ToUint32x4)                                                       \
+  M(Float32x4TwoArgShuffle)                                                    \
   M(MaterializeObject)                                                         \
   M(Uint32x4BoolConstructor)                                                   \
   M(Uint32x4GetFlag)                                                           \
   M(Uint32x4Select)                                                            \
   M(Uint32x4SetFlag)                                                           \
   M(Uint32x4ToFloat32x4)                                                       \
   M(BinaryUint32x4Op)                                                          \
 
 
 #define FORWARD_DECLARATION(type) class type##Instr;
@@ skipping 266 matching lines @@
   friend class Float32x4ShuffleInstr;
   friend class Float32x4ConstructorInstr;
   friend class Float32x4ComparisonInstr;
   friend class Float32x4MinMaxInstr;
   friend class Float32x4ScaleInstr;
   friend class Float32x4SqrtInstr;
   friend class Float32x4ZeroArgInstr;
   friend class Float32x4ClampInstr;
   friend class Float32x4WithInstr;
   friend class Float32x4ToUint32x4Instr;
+  friend class Float32x4TwoArgShuffleInstr;
   friend class Uint32x4BoolConstructorInstr;
   friend class Uint32x4GetFlagInstr;
   friend class Uint32x4SetFlagInstr;
   friend class Uint32x4SelectInstr;
   friend class Uint32x4ToFloat32x4Instr;
   friend class BinaryUint32x4OpInstr;
   friend class BinaryMintOpInstr;
   friend class BinarySmiOpInstr;
   friend class UnarySmiOpInstr;
   friend class ShiftMintOpInstr;
@@ skipping 4593 matching lines @@
 
   virtual bool MayThrow() const { return false; }
 
  private:
   const MethodRecognizer::Kind op_kind_;
 
   DISALLOW_COPY_AND_ASSIGN(Uint32x4GetFlagInstr);
 };
 
 
+class Float32x4TwoArgShuffleInstr : public TemplateDefinition<2> {
+ public:
+  Float32x4TwoArgShuffleInstr(MethodRecognizer::Kind op_kind, Value* left,
+                              Value* right, intptr_t deopt_id)
+      : op_kind_(op_kind) {
+    SetInputAt(0, left);
+    SetInputAt(1, right);
+    deopt_id_ = deopt_id;
+  }
+
+  Value* left() const { return inputs_[0]; }
+  Value* right() const { return inputs_[1]; }
+
+  MethodRecognizer::Kind op_kind() const { return op_kind_; }
+
+  virtual void PrintOperandsTo(BufferFormatter* f) const;
+
+  virtual bool CanDeoptimize() const { return false; }
+
+  virtual Representation representation() const {
+    return kUnboxedFloat32x4;
+  }
+
+  virtual Representation RequiredInputRepresentation(intptr_t idx) const {
+    ASSERT((idx == 0) || (idx == 1));
+    return kUnboxedFloat32x4;
+  }
+
+  virtual intptr_t DeoptimizationTarget() const {
+    // Direct access since this instruction cannot deoptimize, and the deopt-id
+    // was inherited from another instruction that could deoptimize.
+    return deopt_id_;
+  }
+
+  DECLARE_INSTRUCTION(Float32x4TwoArgShuffle)
+  virtual CompileType ComputeType() const;
+
+  virtual bool AllowsCSE() const { return true; }
+  virtual EffectSet Effects() const { return EffectSet::None(); }
+  virtual EffectSet Dependencies() const { return EffectSet::None(); }
+  virtual bool AttributesEqual(Instruction* other) const {
+    return op_kind() == other->AsFloat32x4TwoArgShuffle()->op_kind();
+  }
+
+  virtual bool MayThrow() const { return false; }
+
+ private:
+  const MethodRecognizer::Kind op_kind_;
+
+  DISALLOW_COPY_AND_ASSIGN(Float32x4TwoArgShuffleInstr);
+};
+
+
 class Uint32x4SelectInstr : public TemplateDefinition<3> {
  public:
   Uint32x4SelectInstr(Value* mask, Value* trueValue, Value* falseValue,
                       intptr_t deopt_id) {
     SetInputAt(0, mask);
     SetInputAt(1, trueValue);
     SetInputAt(2, falseValue);
     deopt_id_ = deopt_id;
   }
 
@@ skipping 1031 matching lines @@
   ForwardInstructionIterator* current_iterator_;
 
  private:
   DISALLOW_COPY_AND_ASSIGN(FlowGraphVisitor);
 };
 
 
 }  // namespace dart
 
 #endif  // VM_INTERMEDIATE_LANGUAGE_H_