Lower insertelement and extractelement.

src/IceInstX8632.cpp

 //===- subzero/src/IceInstX8632.cpp - X86-32 instruction implementation ---===//
 //
 //                        The Subzero Code Generator
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
 // This file implements the InstX8632 and OperandX8632 classes,
@@ skipping 24 matching lines @@
   };
 const size_t InstX8632BrAttributesSize =
     llvm::array_lengthof(InstX8632BrAttributes);
 
 const struct TypeX8632Attributes_ {
   const char *CvtString;   // i (integer), s (single FP), d (double FP)
   const char *SdSsString;  // ss, sd, or <blank>
   const char *PackString;  // b, w, d, or <blank>
   const char *WidthString; // {byte,word,dword,qword} ptr
 } TypeX8632Attributes[] = {
-#define X(tag, cvt, sdss, pack, width)                                         \
+#define X(tag, elementty, cvt, sdss, pack, width)                              \
   { cvt, "" sdss, pack, width }                                                \
   ,
     ICETYPEX8632_TABLE
 #undef X
   };
 const size_t TypeX8632AttributesSize =
     llvm::array_lengthof(TypeX8632Attributes);
 
 const char *InstX8632SegmentRegNames[] = {
 #define X(val, name) name,
@@ skipping 249 matching lines @@
     return false;
   if (getDest()->hasReg() && getDest()->getRegNum() == Src->getRegNum()) {
     return true;
   }
   if (!getDest()->hasReg() && !Src->hasReg() &&
       Dest->getStackOffset() == Src->getStackOffset())
     return true;
   return false;
 }
 
-InstX8632Pshufd::InstX8632Pshufd(Cfg *Func, Variable *Dest, Operand *Source1,
-                                 Operand *Source2)
-    : InstX8632(Func, InstX8632::Pshufd, 2, Dest) {
-  addSource(Source1);
-  addSource(Source2);
-}
-
-InstX8632Shufps::InstX8632Shufps(Cfg *Func, Variable *Dest, Operand *Source1,
-                                 Operand *Source2)
-    : InstX8632(Func, InstX8632::Shufps, 3, Dest) {
-  addSource(Dest);
-  addSource(Source1);
-  addSource(Source2);
-}
-
 InstX8632Ret::InstX8632Ret(Cfg *Func, Variable *Source)
     : InstX8632(Func, InstX8632::Ret, Source ? 1 : 0, NULL) {
   if (Source)
     addSource(Source);
 }
 
 InstX8632Xadd::InstX8632Xadd(Cfg *Func, Operand *Dest, Variable *Source,
                              bool Locked)
     : InstX8632Lockable(Func, InstX8632::Xadd, 2,
                         llvm::dyn_cast<Variable>(Dest), Locked) {
@@ skipping 107 matching lines @@
     if (ShiftReg && ShiftReg->getRegNum() == TargetX8632::Reg_ecx) {
       Str << "cl";
       EmittedSrc1 = true;
     }
   }
   if (!EmittedSrc1)
     Inst->getSrc(1)->emit(Func);
   Str << "\n";
 }
 
+
+// Unary ops
 template <> const char *InstX8632Bsf::Opcode = "bsf";
 template <> const char *InstX8632Bsr::Opcode = "bsr";
+template <> const char *InstX8632Lea::Opcode = "lea";
+template <> const char *InstX8632Movd::Opcode = "movd";
+template <> const char *InstX8632Movss::Opcode = "movss";
 template <> const char *InstX8632Sqrtss::Opcode = "sqrtss";
+// Binary ops
 template <> const char *InstX8632Add::Opcode = "add";
 template <> const char *InstX8632Addps::Opcode = "addps";
 template <> const char *InstX8632Adc::Opcode = "adc";
 template <> const char *InstX8632Addss::Opcode = "addss";
 template <> const char *InstX8632Padd::Opcode = "padd";
 template <> const char *InstX8632Sub::Opcode = "sub";
 template <> const char *InstX8632Subps::Opcode = "subps";
 template <> const char *InstX8632Subss::Opcode = "subss";
 template <> const char *InstX8632Sbb::Opcode = "sbb";
 template <> const char *InstX8632Psub::Opcode = "psub";
@@ skipping 12 matching lines @@
 template <> const char *InstX8632Divps::Opcode = "divps";
 template <> const char *InstX8632Idiv::Opcode = "idiv";
 template <> const char *InstX8632Divss::Opcode = "divss";
 template <> const char *InstX8632Shl::Opcode = "shl";
 template <> const char *InstX8632Psll::Opcode = "psll";
 template <> const char *InstX8632Shr::Opcode = "shr";
 template <> const char *InstX8632Sar::Opcode = "sar";
 template <> const char *InstX8632Psra::Opcode = "psra";
 template <> const char *InstX8632Pcmpeq::Opcode = "pcmpeq";
 template <> const char *InstX8632Pcmpgt::Opcode = "pcmpgt";
+// Ternary ops
+template <> const char *InstX8632Shufps::Opcode = "shufps";
+template <> const char *InstX8632Pinsrw::Opcode = "pinsrw";
+// Three address ops
+template <> const char *InstX8632Pextrw::Opcode = "pextrw";
+template <> const char *InstX8632Pshufd::Opcode = "pshufd";
 
 template <> void InstX8632Sqrtss::emit(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrEmit();
   assert(getSrcSize() == 1);
   Type Ty = getSrc(0)->getType();
   assert(Ty == IceType_f32 || Ty == IceType_f64);
   Str << "\tsqrt" << TypeX8632Attributes[Ty].SdSsString << "\t";
   getDest()->emit(Func);
   Str << ", ";
   getSrc(0)->emit(Func);
@@ skipping 47 matching lines @@
   emitTwoAddress(Opcode, this, Func);
 }
 
 template <> void InstX8632Divss::emit(const Cfg *Func) const {
   char buf[30];
   snprintf(buf, llvm::array_lengthof(buf), "div%s",
            TypeX8632Attributes[getDest()->getType()].SdSsString);
   emitTwoAddress(buf, this, Func);
 }
 
+template <> void InstX8632Div::emit(const Cfg *Func) const {
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(getSrcSize() == 3);
+  Str << "\t" << Opcode << "\t";
+  getSrc(1)->emit(Func);
+  Str << "\n";
+}
+
+template <> void InstX8632Idiv::emit(const Cfg *Func) const {
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(getSrcSize() == 3);
+  Str << "\t" << Opcode << "\t";
+  getSrc(1)->emit(Func);
+  Str << "\n";
+}
+
 template <> void InstX8632Imul::emit(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrEmit();
   assert(getSrcSize() == 2);
   if (getDest()->getType() == IceType_i8) {
     // The 8-bit version of imul only allows the form "imul r/m8".
     Variable *Src0 = llvm::dyn_cast<Variable>(getSrc(0));
     assert(Src0 && Src0->getRegNum() == TargetX8632::Reg_eax);
     Str << "\timul\t";
     getSrc(1)->emit(Func);
     Str << "\n";
@@ skipping 292 matching lines @@
 }
 
 void InstX8632StoreQ::dump(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrDump();
   Str << "storeq." << getSrc(0)->getType() << " ";
   getSrc(1)->dump(Func);
   Str << ", ";
   getSrc(0)->dump(Func);
 }
 
+template <> void InstX8632Lea::emit(const Cfg *Func) const {
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(getSrcSize() == 1);
+  assert(getDest()->hasReg());
+  Str << "\tlea\t";
+  getDest()->emit(Func);
+  Str << ", ";
+  Operand *Src0 = getSrc(0);
+  if (Variable *VSrc0 = llvm::dyn_cast<Variable>(Src0)) {
+    Type Ty = VSrc0->getType();
+    // lea on x86-32 doesn't accept mem128 operands, so cast VSrc0 to an
+    // acceptable type.
+    VSrc0->asType(isVectorType(Ty) ? IceType_i32 : Ty).emit(Func);
+  } else {
+    Src0->emit(Func);
+  }
+  Str << "\n";
+}
+
 void InstX8632Mov::emit(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrEmit();
   assert(getSrcSize() == 1);
   Operand *Src = getSrc(0);
   // The llvm-mc assembler using Intel syntax has a bug in which "mov
   // reg, RelocatableConstant" does not generate the right instruction
   // with a relocation.  To work around, we emit "lea reg,
   // [RelocatableConstant]".  Also, the lowering and legalization is
   // changed to allow relocatable constants only in Assign and Call
   // instructions or in Mem operands.  TODO(stichnot): remove LEAHACK
   // once a proper emitter is used.
   bool UseLeaHack = llvm::isa<ConstantRelocatable>(Src);
   Str << "\t";
   if (UseLeaHack)
     Str << "lea";
   else
     Str << "mov" << TypeX8632Attributes[getDest()->getType()].SdSsString;
   Str << "\t";
   // For an integer truncation operation, src is wider than dest.
   // Ideally, we use a mov instruction whose data width matches the
   // narrower dest.  This is a problem if e.g. src is a register like
   // esi or si where there is no 8-bit version of the register.  To be
   // safe, we instead widen the dest to match src.  This works even
   // for stack-allocated dest variables because typeWidthOnStack()
   // pads to a 4-byte boundary even if only a lower portion is used.
+  // TODO: This assert disallows usages such as copying a floating point
+  // value between a vector and a scalar (which movss is used for).
+  // Clean this up.
   assert(Func->getTarget()->typeWidthInBytesOnStack(getDest()->getType()) ==
          Func->getTarget()->typeWidthInBytesOnStack(Src->getType()));
   getDest()->asType(Src->getType()).emit(Func);
   Str << ", ";
   Src->emit(Func);
   Str << "\n";
 }
 
 void InstX8632Mov::dump(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrDump();
@@ skipping 153 matching lines @@
   emitTwoAddress(buf, this, Func);
 }
 
 template <> void InstX8632Pcmpgt::emit(const Cfg *Func) const {
   char buf[30];
   snprintf(buf, llvm::array_lengthof(buf), "pcmpgt%s",
            TypeX8632Attributes[getDest()->getType()].PackString);
   emitTwoAddress(buf, this, Func);
 }
 
+template <> void InstX8632Pextrw::emit(const Cfg *Func) const {
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(getSrcSize() == 2);
+  Str << "\t" << Opcode << "\t";
+  Variable *Dest = getDest();
+  assert(Dest->hasReg() && Dest->getType() == IceType_i16);
+  // pextrw takes r32 dest.
+  Dest->asType(IceType_i32).emit(Func);
+  Str << ", ";
+  getSrc(0)->emit(Func);
+  Str << ", ";
+  getSrc(1)->emit(Func);
+  Str << "\n";
+}
+
+template <> void InstX8632Pinsrw::emit(const Cfg *Func) const {
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(getSrcSize() == 3);
+  Str << "\t" << Opcode << "\t";
+  getDest()->emit(Func);
+  Str << ", ";
+  Operand *Src1 = getSrc(1);
+  if (Variable *VSrc1 = llvm::dyn_cast<Variable>(Src1)) {
+    // If src1 is a register, it should be r32.
+    VSrc1->asType(VSrc1->hasReg() ? IceType_i32 : IceType_i16).emit(Func);
+  } else {
+    Src1->emit(Func);
+  }
+  Str << ", ";
+  getSrc(2)->emit(Func);
+  Str << "\n";
+}
+
 void InstX8632Pop::emit(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrEmit();
   assert(getSrcSize() == 0);
   Str << "\tpop\t";
   getDest()->emit(Func);
   Str << "\n";
 }
 
 void InstX8632Pop::dump(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrDump();
@@ skipping 52 matching lines @@
 
 template <> void InstX8632Psra::emit(const Cfg *Func) const {
   assert(getDest()->getType() == IceType_v8i16 ||
          getDest()->getType() == IceType_v4i32);
   char buf[30];
   snprintf(buf, llvm::array_lengthof(buf), "psra%s",
            TypeX8632Attributes[getDest()->getType()].PackString);
   emitTwoAddress(buf, this, Func);
 }
 
-void InstX8632Pshufd::emit(const Cfg *Func) const {
-  Ostream &Str = Func->getContext()->getStrEmit();
-  assert(getSrcSize() == 2);
-  Str << "\tpshufd\t";
-  getDest()->emit(Func);
-  Str << ", ";
-  getSrc(0)->emit(Func);
-  Str << ", ";
-  getSrc(1)->emit(Func);
-  Str << "\n";
-}
-
-void InstX8632Pshufd::dump(const Cfg *Func) const {
-  Ostream &Str = Func->getContext()->getStrDump();
-  dumpDest(Func);
-  Str << " = pshufd." << getDest()->getType() << " ";
-  dumpSources(Func);
-}
-
 void InstX8632Ret::emit(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrEmit();
   Str << "\tret\n";
 }
 
 void InstX8632Ret::dump(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrDump();
   Type Ty = (getSrcSize() == 0 ? IceType_void : getSrc(0)->getType());
   Str << "ret." << Ty << " ";
   dumpSources(Func);
 }
 
-void InstX8632Shufps::emit(const Cfg *Func) const {
-  Ostream &Str = Func->getContext()->getStrEmit();
-  assert(getSrcSize() == 3);
-  Str << "\tshufps\t";
-  getDest()->emit(Func);
-  Str << ", ";
-  getSrc(1)->emit(Func);
-  Str << ", ";
-  getSrc(2)->emit(Func);
-  Str << "\n";
-}
-
-void InstX8632Shufps::dump(const Cfg *Func) const {
-  Ostream &Str = Func->getContext()->getStrDump();
-  dumpDest(Func);
-  Str << " = shufps." << getDest()->getType() << " ";
-  dumpSources(Func);
-}
-
 void InstX8632Xadd::emit(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrEmit();
   if (Locked) {
     Str << "\tlock";
   }
   Str << "\txadd\t";
   getSrc(0)->emit(Func);
   Str << ", ";
   getSrc(1)->emit(Func);
   Str << "\n";
@@ skipping 151 matching lines @@
   default:
     Str << "???";
     break;
   }
   Str << "(";
   Var->dump(Func);
   Str << ")";
 }
 
 } // end of namespace Ice