emitIAS for Shld and Shrd and the ternary and three-address ops.

src/assembler_ia32.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.
 //
 // Modified by the Subzero authors.
 //
 //===- subzero/src/assembler_ia32.h - Assembler for x86-32 ----------------===//
 //
 //                        The Subzero Code Generator
 //
@@ skipping 364 matching lines @@
                                                  const Address &);
   typedef void (AssemblerX86::*TypedEmitGPRImm)(Type, GPRRegister,
                                                 const Immediate &);
   struct GPREmitterRegOp {
     TypedEmitGPRGPR GPRGPR;
     TypedEmitGPRAddr GPRAddr;
     TypedEmitGPRImm GPRImm;
   };
 
   struct GPREmitterShiftOp {
+    // Technically, Addr/GPR and Addr/Imm are also allowed, but */Addr are not.
+    // In practice, we always normalize the Dest to a Register first.
     TypedEmitGPRGPR GPRGPR;
     TypedEmitGPRImm GPRImm;
-    // Technically, Addr/GPR and Addr/Imm are also allowed, but */Addr are not.
-    // In practice, we always normalize the Dest to a Register first.
+  };
+
+  typedef void (AssemblerX86::*TypedEmitGPRGPRImm)(Type, GPRRegister,
+                                                   GPRRegister,
+                                                   const Immediate &);
+  struct GPREmitterShiftD {
+    // Technically AddrGPR and AddrGPRImm are also allowed, but in practice
+    // we always normalize Dest to a Register first.
+    TypedEmitGPRGPR GPRGPR;
+    TypedEmitGPRGPRImm GPRGPRImm;
   };
 
   typedef void (AssemblerX86::*TypedEmitAddrGPR)(Type, const Address &,
                                                  GPRRegister);
   typedef void (AssemblerX86::*TypedEmitAddrImm)(Type, const Address &,
                                                  const Immediate &);
   struct GPREmitterAddrOp {
     TypedEmitAddrGPR AddrGPR;
     TypedEmitAddrImm AddrImm;
   };
@@ skipping 27 matching lines @@
 
   // Cross Xmm/GPR cast instructions.
   template <typename DReg_t, typename SReg_t> struct CastEmitterRegOp {
     typedef void (AssemblerX86::*TypedEmitRegs)(Type, DReg_t, SReg_t);
     typedef void (AssemblerX86::*TypedEmitAddr)(Type, DReg_t, const Address &);
 
     TypedEmitRegs RegReg;
     TypedEmitAddr RegAddr;
   };
 
+  // Three operand (potentially) cross Xmm/GPR instructions.
+  // The last operand must be an immediate.
+  template <typename DReg_t, typename SReg_t> struct ThreeOpImmEmitter {
+    typedef void (AssemblerX86::*TypedEmitRegRegImm)(Type, DReg_t, SReg_t,
+                                                     const Immediate &);
+    typedef void (AssemblerX86::*TypedEmitRegAddrImm)(Type, DReg_t,
+                                                      const Address &,
+                                                      const Immediate &);
+
+    TypedEmitRegRegImm RegRegImm;
+    TypedEmitRegAddrImm RegAddrImm;
+  };
+
   /*
    * Emit Machine Instructions.
    */
   void call(GPRRegister reg);
   void call(const Address &address);
   void call(Label *label);
   void call(const ConstantRelocatable *label);
 
   static const intptr_t kCallExternalLabelSize = 5;
 
@@ skipping 117 matching lines @@
   void unpckhpd(XmmRegister dst, XmmRegister src);
 
   void set1ps(XmmRegister dst, GPRRegister tmp, const Immediate &imm);
   void shufps(XmmRegister dst, XmmRegister src, const Immediate &mask);
 
   void minpd(XmmRegister dst, XmmRegister src);
   void maxpd(XmmRegister dst, XmmRegister src);
   void sqrtpd(XmmRegister dst);
   void shufpd(XmmRegister dst, XmmRegister src, const Immediate &mask);
 
+  void pshufd(Type Ty, XmmRegister dst, XmmRegister src, const Immediate &mask);
+  void pshufd(Type Ty, XmmRegister dst, const Address &src,
+              const Immediate &mask);
+  void shufps(Type Ty, XmmRegister dst, XmmRegister src, const Immediate &mask);
+  void shufps(Type Ty, XmmRegister dst, const Address &src,
+              const Immediate &mask);
+
   void cvtdq2ps(Type, XmmRegister dst, XmmRegister src);
   void cvtdq2ps(Type, XmmRegister dst, const Address &src);
 
   void cvttps2dq(Type, XmmRegister dst, XmmRegister src);
   void cvttps2dq(Type, XmmRegister dst, const Address &src);
 
   void cvtsi2ss(Type DestTy, XmmRegister dst, GPRRegister src);
   void cvtsi2ss(Type DestTy, XmmRegister dst, const Address &src);
 
   void cvtfloat2float(Type SrcTy, XmmRegister dst, XmmRegister src);
@@ skipping 14 matching lines @@
   void xorpd(XmmRegister dst, const Address &src);
   void xorpd(XmmRegister dst, XmmRegister src);
   void xorps(XmmRegister dst, const Address &src);
   void xorps(XmmRegister dst, XmmRegister src);
 
   void andpd(XmmRegister dst, const Address &src);
   void andpd(XmmRegister dst, XmmRegister src);
 
   void orpd(XmmRegister dst, XmmRegister src);
 
-  void pextrd(GPRRegister dst, XmmRegister src, const Immediate &imm);
+  void insertps(Type Ty, XmmRegister dst, XmmRegister src,
+                const Immediate &imm);
+  void insertps(Type Ty, XmmRegister dst, const Address &src,
+                const Immediate &imm);
+
+  void pinsr(Type Ty, XmmRegister dst, GPRRegister src, const Immediate &imm);
+  void pinsr(Type Ty, XmmRegister dst, const Address &src,
+             const Immediate &imm);
+
+  void pextr(Type Ty, GPRRegister dst, XmmRegister src, const Immediate &imm);
+  void pextr(Type Ty, GPRRegister dst, const Address &src,
+             const Immediate &imm);
+
   void pmovsxdq(XmmRegister dst, XmmRegister src);
 
   void pcmpeq(Type Ty, XmmRegister dst, XmmRegister src);
   void pcmpeq(Type Ty, XmmRegister dst, const Address &src);
   void pcmpgt(Type Ty, XmmRegister dst, XmmRegister src);
   void pcmpgt(Type Ty, XmmRegister dst, const Address &src);
 
   enum RoundingMode {
     kRoundToNearest = 0x0,
     kRoundDown = 0x1,
@@ skipping 90 matching lines @@
   void shl(Type Ty, const Address &operand, GPRRegister shifter);
 
   void shr(Type Ty, GPRRegister reg, const Immediate &imm);
   void shr(Type Ty, GPRRegister operand, GPRRegister shifter);
   void shr(Type Ty, const Address &operand, GPRRegister shifter);
 
   void sar(Type Ty, GPRRegister reg, const Immediate &imm);
   void sar(Type Ty, GPRRegister operand, GPRRegister shifter);
   void sar(Type Ty, const Address &address, GPRRegister shifter);
 
-  void shld(GPRRegister dst, GPRRegister src);
-  void shld(GPRRegister dst, GPRRegister src, const Immediate &imm);
-  void shld(const Address &operand, GPRRegister src);
-  void shrd(GPRRegister dst, GPRRegister src);
-  void shrd(GPRRegister dst, GPRRegister src, const Immediate &imm);
-  void shrd(const Address &dst, GPRRegister src);
+  void shld(Type Ty, GPRRegister dst, GPRRegister src);
+  void shld(Type Ty, GPRRegister dst, GPRRegister src, const Immediate &imm);
+  void shld(Type Ty, const Address &operand, GPRRegister src);
+  void shrd(Type Ty, GPRRegister dst, GPRRegister src);
+  void shrd(Type Ty, GPRRegister dst, GPRRegister src, const Immediate &imm);
+  void shrd(Type Ty, const Address &dst, GPRRegister src);
 
   void neg(Type Ty, GPRRegister reg);
   void neg(Type Ty, const Address &addr);
   void notl(GPRRegister reg);
 
   void bsf(Type Ty, GPRRegister dst, GPRRegister src);
   void bsf(Type Ty, GPRRegister dst, const Address &src);
   void bsr(Type Ty, GPRRegister dst, GPRRegister src);
   void bsr(Type Ty, GPRRegister dst, const Address &src);
 
@@ skipping 102 matching lines @@
 inline void AssemblerX86::EmitFixup(AssemblerFixup *fixup) {
   buffer_.EmitFixup(fixup);
 }
 
 inline void AssemblerX86::EmitOperandSizeOverride() { EmitUint8(0x66); }
 
 } // end of namespace x86
 } // end of namespace Ice
 
 #endif // SUBZERO_SRC_ASSEMBLER_IA32_H_