Use movss to implement insertelement when elements = 4 and index = 0.

src/IceInstX8632.h

 //===- subzero/src/IceInstX8632.h - Low-level x86 instructions --*- C++ -*-===//
 //
 //                        The Subzero Code Generator
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
 // This file declares the InstX8632 and OperandX8632 classes and
@@ skipping 534 matching lines @@
   virtual ~InstX8632ThreeAddressop() {}
   static const char *Opcode;
 };
 
 typedef InstX8632Inplaceop<InstX8632::Bswap> InstX8632Bswap;
 typedef InstX8632Inplaceop<InstX8632::Neg> InstX8632Neg;
 typedef InstX8632Unaryop<InstX8632::Bsf> InstX8632Bsf;
 typedef InstX8632Unaryop<InstX8632::Bsr> InstX8632Bsr;
 typedef InstX8632Unaryop<InstX8632::Lea> InstX8632Lea;
 typedef InstX8632Unaryop<InstX8632::Movd> InstX8632Movd;
-typedef InstX8632Unaryop<InstX8632::Movss> InstX8632Movss;
 typedef InstX8632Unaryop<InstX8632::Sqrtss> InstX8632Sqrtss;
 typedef InstX8632Binop<InstX8632::Add> InstX8632Add;
 typedef InstX8632Binop<InstX8632::Addps> InstX8632Addps;
 typedef InstX8632Binop<InstX8632::Adc> InstX8632Adc;
 typedef InstX8632Binop<InstX8632::Addss> InstX8632Addss;
 typedef InstX8632Binop<InstX8632::Padd> InstX8632Padd;
 typedef InstX8632Binop<InstX8632::Sub> InstX8632Sub;
 typedef InstX8632Binop<InstX8632::Subps> InstX8632Subps;
 typedef InstX8632Binop<InstX8632::Subss> InstX8632Subss;
 typedef InstX8632Binop<InstX8632::Sbb> InstX8632Sbb;
@@ skipping 13 matching lines @@
 typedef InstX8632Binop<InstX8632::Divps> InstX8632Divps;
 typedef InstX8632Binop<InstX8632::Divss> InstX8632Divss;
 typedef InstX8632Binop<InstX8632::Rol, true> InstX8632Rol;
 typedef InstX8632Binop<InstX8632::Shl, true> InstX8632Shl;
 typedef InstX8632Binop<InstX8632::Psll> InstX8632Psll;
 typedef InstX8632Binop<InstX8632::Shr, true> InstX8632Shr;
 typedef InstX8632Binop<InstX8632::Sar, true> InstX8632Sar;
 typedef InstX8632Binop<InstX8632::Psra> InstX8632Psra;
 typedef InstX8632Binop<InstX8632::Pcmpeq> InstX8632Pcmpeq;
 typedef InstX8632Binop<InstX8632::Pcmpgt> InstX8632Pcmpgt;
+typedef InstX8632Binop<InstX8632::Movss> InstX8632Movss;

[Jim Stichnoth, 2014/07/25 20:07:12]

Could you add a TODO here, or somewhere appropriate, that the Dest (and hence
Src0) operand of the movss instruction should have a vector type?  Also, that
Dest should be a physical register.

The rationale is that (correct me if I'm wrong) the movss instruction is really
only a binary op when used with a vector instruction like insertelement,
otherwise it's more like mov/movsx/movzx.

[wala, 2014/07/25 22:57:41]

Done.

Per our discussion, movss is a binary op only when both source and dest operands
are registers. Any current use of _movss as a non binary op should be made to
use _mov eventually.

 typedef InstX8632Ternop<InstX8632::Idiv> InstX8632Idiv;
 typedef InstX8632Ternop<InstX8632::Div> InstX8632Div;
 typedef InstX8632Ternop<InstX8632::Pinsrw> InstX8632Pinsrw;
 typedef InstX8632Ternop<InstX8632::Shufps> InstX8632Shufps;
 typedef InstX8632ThreeAddressop<InstX8632::Pextrw> InstX8632Pextrw;
 typedef InstX8632ThreeAddressop<InstX8632::Pshufd> InstX8632Pshufd;
 
 // Base class for a lockable x86-32 instruction (emits a locked prefix).
 class InstX8632Lockable : public InstX8632 {
 public:
@@ skipping 574 matching lines @@
 private:
   InstX8632Xchg(Cfg *Func, Operand *Dest, Variable *Source);
   InstX8632Xchg(const InstX8632Xchg &) LLVM_DELETED_FUNCTION;
   InstX8632Xchg &operator=(const InstX8632Xchg &) LLVM_DELETED_FUNCTION;
   virtual ~InstX8632Xchg() {}
 };
 
 } // end of namespace Ice
 
 #endif // SUBZERO_SRC_ICEINSTX8632_H