X8632 Templatization completed.

src/IceInstX86BaseImpl.h

+//===- subzero/src/IceInstX86BaseImpl.h - Generic 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 implements the InstX86Base and OperandX86Base template classes that
+// are shared accross all x86 targets.

[jvoung (off chromium), 2015/07/06 18:58:45]

accross -> across

[John, 2015/07/06 22:30:09]

Done.

+//
+//===----------------------------------------------------------------------===//
+
+#ifndef SUBZERO_SRC_ICEINSTX86BASEIMPL_H
+#define SUBZERO_SRC_ICEINSTX86BASEIMPL_H
+
+#include "IceInstX86Base.h"
+
+#include "IceAssemblerX86Base.h"
+#include "IceCfg.h"
+#include "IceCfgNode.h"
+#include "IceDefs.h"
+#include "IceInst.h"
+#include "IceOperand.h"
+#include "IceTargetLowering.h"
+
+namespace Ice {
+
+namespace X86Internal {
+
+template <class Machine>
+const char *InstX86Base<Machine>::getWidthString(Type Ty) {
+  return Traits::TypeAttributes[Ty].WidthString;
+}
+
+template <class Machine>
+const char *InstX86Base<Machine>::getFldString(Type Ty) {
+  return Traits::TypeAttributes[Ty].FldString;
+}
+
+template <class Machine>
+typename InstX86Base<Machine>::Traits::Cond::BrCond
+InstX86Base<Machine>::getOppositeCondition(typename Traits::Cond::BrCond Cond) {
+  return Traits::InstBrAttributes[Cond].Opposite;
+}
+
+template <class Machine>
+InstX86BaseFakeRMW<Machine>::InstX86BaseFakeRMW(Cfg *Func, Operand *Data,
+                                                Operand *Addr,
+                                                InstArithmetic::OpKind Op,
+                                                Variable *Beacon)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::FakeRMW, 3, nullptr),
+      Op(Op) {
+  this->addSource(Data);
+  this->addSource(Addr);
+  this->addSource(Beacon);
+}
+
+template <class Machine>
+InstX86BaseAdjustStack<Machine>::InstX86BaseAdjustStack(Cfg *Func, SizeT Amount,
+                                                        Variable *Esp)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Adjuststack, 1, Esp),
+      Amount(Amount) {
+  this->addSource(Esp);
+}
+
+template <class Machine>
+InstX86BaseMul<Machine>::InstX86BaseMul(Cfg *Func, Variable *Dest,
+                                        Variable *Source1, Operand *Source2)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Mul, 2, Dest) {
+  this->addSource(Source1);
+  this->addSource(Source2);
+}
+
+template <class Machine>
+InstX86BaseShld<Machine>::InstX86BaseShld(Cfg *Func, Variable *Dest,
+                                          Variable *Source1, Variable *Source2)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Shld, 3, Dest) {
+  this->addSource(Dest);
+  this->addSource(Source1);
+  this->addSource(Source2);
+}
+
+template <class Machine>
+InstX86BaseShrd<Machine>::InstX86BaseShrd(Cfg *Func, Variable *Dest,
+                                          Variable *Source1, Variable *Source2)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Shrd, 3, Dest) {
+  this->addSource(Dest);
+  this->addSource(Source1);
+  this->addSource(Source2);
+}
+
+template <class Machine>
+InstX86BaseLabel<Machine>::InstX86BaseLabel(
+    Cfg *Func, typename InstX86Base<Machine>::Traits::TargetLowering *Target)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Label, 0, nullptr),
+      Number(Target->makeNextLabelNumber()) {}
+
+template <class Machine>
+IceString InstX86BaseLabel<Machine>::getName(const Cfg *Func) const {
+  return ".L" + Func->getFunctionName() + "$local$__" + std::to_string(Number);
+}
+
+template <class Machine>
+InstX86BaseBr<Machine>::InstX86BaseBr(
+    Cfg *Func, const CfgNode *TargetTrue, const CfgNode *TargetFalse,
+    const InstX86BaseLabel<Machine> *Label,
+    typename InstX86Base<Machine>::Traits::Cond::BrCond Condition)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Br, 0, nullptr),
+      Condition(Condition), TargetTrue(TargetTrue), TargetFalse(TargetFalse),
+      Label(Label) {}
+
+template <class Machine>
+bool InstX86BaseBr<Machine>::optimizeBranch(const CfgNode *NextNode) {
+  // If there is no next block, then there can be no fallthrough to
+  // optimize.
+  if (NextNode == nullptr)
+    return false;
+  // Intra-block conditional branches can't be optimized.
+  if (Label)
+    return false;
+  // If there is no fallthrough node, such as a non-default case label
+  // for a switch instruction, then there is no opportunity to
+  // optimize.
+  if (getTargetFalse() == nullptr)
+    return false;
+
+  // Unconditional branch to the next node can be removed.
+  if (Condition == InstX86Base<Machine>::Traits::Cond::Br_None &&
+      getTargetFalse() == NextNode) {
+    assert(getTargetTrue() == nullptr);
+    this->setDeleted();
+    return true;
+  }
+  // If the fallthrough is to the next node, set fallthrough to nullptr
+  // to indicate.
+  if (getTargetFalse() == NextNode) {
+    TargetFalse = nullptr;
+    return true;
+  }
+  // If TargetTrue is the next node, and TargetFalse is not nullptr
+  // (which was already tested above), then invert the branch
+  // condition, swap the targets, and set new fallthrough to nullptr.
+  if (getTargetTrue() == NextNode) {
+    assert(Condition != InstX86Base<Machine>::Traits::Cond::Br_None);
+    Condition = this->getOppositeCondition(Condition);
+    TargetTrue = getTargetFalse();
+    TargetFalse = nullptr;
+    return true;
+  }
+  return false;
+}
+
+template <class Machine>
+bool InstX86BaseBr<Machine>::repointEdge(CfgNode *OldNode, CfgNode *NewNode) {
+  if (TargetFalse == OldNode) {
+    TargetFalse = NewNode;
+    return true;
+  } else if (TargetTrue == OldNode) {
+    TargetTrue = NewNode;
+    return true;
+  }
+  return false;
+}
+
+template <class Machine>
+InstX86BaseJmp<Machine>::InstX86BaseJmp(Cfg *Func, Operand *Target)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Jmp, 1, nullptr) {
+  this->addSource(Target);
+}
+
+template <class Machine>
+InstX86BaseCall<Machine>::InstX86BaseCall(Cfg *Func, Variable *Dest,
+                                          Operand *CallTarget)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Call, 1, Dest) {
+  this->HasSideEffects = true;
+  this->addSource(CallTarget);
+}
+
+template <class Machine>
+InstX86BaseCmov<Machine>::InstX86BaseCmov(
+    Cfg *Func, Variable *Dest, Operand *Source,
+    typename InstX86Base<Machine>::Traits::Cond::BrCond Condition)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Cmov, 2, Dest),
+      Condition(Condition) {
+  // The final result is either the original Dest, or Source, so mark
+  // both as sources.
+  this->addSource(Dest);
+  this->addSource(Source);
+}
+
+template <class Machine>
+InstX86BaseCmpps<Machine>::InstX86BaseCmpps(
+    Cfg *Func, Variable *Dest, Operand *Source,
+    typename InstX86Base<Machine>::Traits::Cond::CmppsCond Condition)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Cmpps, 2, Dest),
+      Condition(Condition) {
+  this->addSource(Dest);
+  this->addSource(Source);
+}
+
+template <class Machine>
+InstX86BaseCmpxchg<Machine>::InstX86BaseCmpxchg(Cfg *Func, Operand *DestOrAddr,
+                                                Variable *Eax,
+                                                Variable *Desired, bool Locked)
+    : InstX86BaseLockable<Machine>(Func, InstX86Base<Machine>::Cmpxchg, 3,
+                                   llvm::dyn_cast<Variable>(DestOrAddr),
+                                   Locked) {
+  assert(Eax->getRegNum() ==
+         InstX86Base<Machine>::Traits::RegisterSet::Reg_eax);
+  this->addSource(DestOrAddr);
+  this->addSource(Eax);
+  this->addSource(Desired);
+}
+
+template <class Machine>
+InstX86BaseCmpxchg8b<Machine>::InstX86BaseCmpxchg8b(
+    Cfg *Func, typename InstX86Base<Machine>::Traits::X86OperandMem *Addr,
+    Variable *Edx, Variable *Eax, Variable *Ecx, Variable *Ebx, bool Locked)
+    : InstX86BaseLockable<Machine>(Func, InstX86Base<Machine>::Cmpxchg, 5,
+                                   nullptr, Locked) {
+  assert(Edx->getRegNum() ==
+         InstX86Base<Machine>::Traits::RegisterSet::Reg_edx);
+  assert(Eax->getRegNum() ==
+         InstX86Base<Machine>::Traits::RegisterSet::Reg_eax);
+  assert(Ecx->getRegNum() ==
+         InstX86Base<Machine>::Traits::RegisterSet::Reg_ecx);
+  assert(Ebx->getRegNum() ==
+         InstX86Base<Machine>::Traits::RegisterSet::Reg_ebx);
+  this->addSource(Addr);
+  this->addSource(Edx);
+  this->addSource(Eax);
+  this->addSource(Ecx);
+  this->addSource(Ebx);
+}
+
+template <class Machine>
+InstX86BaseCvt<Machine>::InstX86BaseCvt(Cfg *Func, Variable *Dest,
+                                        Operand *Source, CvtVariant Variant)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Cvt, 1, Dest),
+      Variant(Variant) {
+  this->addSource(Source);
+}
+
+template <class Machine>
+InstX86BaseIcmp<Machine>::InstX86BaseIcmp(Cfg *Func, Operand *Src0,
+                                          Operand *Src1)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Icmp, 2, nullptr) {
+  this->addSource(Src0);
+  this->addSource(Src1);
+}
+
+template <class Machine>
+InstX86BaseUcomiss<Machine>::InstX86BaseUcomiss(Cfg *Func, Operand *Src0,
+                                                Operand *Src1)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Ucomiss, 2, nullptr) {
+  this->addSource(Src0);
+  this->addSource(Src1);
+}
+
+template <class Machine>
+InstX86BaseUD2<Machine>::InstX86BaseUD2(Cfg *Func)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::UD2, 0, nullptr) {}
+
+template <class Machine>
+InstX86BaseTest<Machine>::InstX86BaseTest(Cfg *Func, Operand *Src1,
+                                          Operand *Src2)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Test, 2, nullptr) {
+  this->addSource(Src1);
+  this->addSource(Src2);
+}
+
+template <class Machine>
+InstX86BaseMfence<Machine>::InstX86BaseMfence(Cfg *Func)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Mfence, 0, nullptr) {
+  this->HasSideEffects = true;
+}
+
+template <class Machine>
+InstX86BaseStore<Machine>::InstX86BaseStore(
+    Cfg *Func, Operand *Value,
+    typename InstX86Base<Machine>::Traits::X86Operand *Mem)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Store, 2, nullptr) {
+  this->addSource(Value);
+  this->addSource(Mem);
+}
+
+template <class Machine>
+InstX86BaseStoreP<Machine>::InstX86BaseStoreP(
+    Cfg *Func, Variable *Value,
+    typename InstX86Base<Machine>::Traits::X86OperandMem *Mem)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::StoreP, 2, nullptr) {
+  this->addSource(Value);
+  this->addSource(Mem);
+}
+
+template <class Machine>
+InstX86BaseStoreQ<Machine>::InstX86BaseStoreQ(
+    Cfg *Func, Variable *Value,
+    typename InstX86Base<Machine>::Traits::X86OperandMem *Mem)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::StoreQ, 2, nullptr) {
+  this->addSource(Value);
+  this->addSource(Mem);
+}
+
+template <class Machine>
+InstX86BaseNop<Machine>::InstX86BaseNop(Cfg *Func,
+                                        InstX86BaseNop::NopVariant Variant)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Nop, 0, nullptr),
+      Variant(Variant) {}
+
+template <class Machine>
+InstX86BaseFld<Machine>::InstX86BaseFld(Cfg *Func, Operand *Src)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Fld, 1, nullptr) {
+  this->addSource(Src);
+}
+
+template <class Machine>
+InstX86BaseFstp<Machine>::InstX86BaseFstp(Cfg *Func, Variable *Dest)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Fstp, 0, Dest) {}
+
+template <class Machine>
+InstX86BasePop<Machine>::InstX86BasePop(Cfg *Func, Variable *Dest)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Pop, 0, Dest) {
+  // A pop instruction affects the stack pointer and so it should not
+  // be allowed to be automatically dead-code eliminated.  (The
+  // corresponding push instruction doesn't need this treatment
+  // because it has no dest variable and therefore won't be dead-code
+  // eliminated.)  This is needed for late-stage liveness analysis
+  // (e.g. asm-verbose mode).
+  this->HasSideEffects = true;
+}
+
+template <class Machine>
+InstX86BasePush<Machine>::InstX86BasePush(Cfg *Func, Variable *Source)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Push, 1, nullptr) {
+  this->addSource(Source);
+}
+
+template <class Machine>
+InstX86BaseRet<Machine>::InstX86BaseRet(Cfg *Func, Variable *Source)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Ret, Source ? 1 : 0,
+                           nullptr) {
+  if (Source)
+    this->addSource(Source);
+}
+
+template <class Machine>
+InstX86BaseSetcc<Machine>::InstX86BaseSetcc(
+    Cfg *Func, Variable *Dest,
+    typename InstX86Base<Machine>::Traits::Cond::BrCond Cond)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Setcc, 0, Dest),
+      Condition(Cond) {}
+
+template <class Machine>
+InstX86BaseXadd<Machine>::InstX86BaseXadd(Cfg *Func, Operand *Dest,
+                                          Variable *Source, bool Locked)
+    : InstX86BaseLockable<Machine>(Func, InstX86Base<Machine>::Xadd, 2,
+                                   llvm::dyn_cast<Variable>(Dest), Locked) {
+  this->addSource(Dest);
+  this->addSource(Source);
+}
+
+template <class Machine>
+InstX86BaseXchg<Machine>::InstX86BaseXchg(Cfg *Func, Operand *Dest,
+                                          Variable *Source)
+    : InstX86Base<Machine>(Func, InstX86Base<Machine>::Xchg, 2,
+                           llvm::dyn_cast<Variable>(Dest)) {
+  this->addSource(Dest);
+  this->addSource(Source);
+}
+
+// ======================== Dump routines ======================== //
+
+template <class Machine>
+void InstX86Base<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  Str << "[" << Traits::TargetName << "] ";
+  Inst::dump(Func);
+}
+
+template <class Machine>
+void InstX86BaseFakeRMW<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  Type Ty = getData()->getType();
+  Str << "rmw " << InstArithmetic::getOpName(getOp()) << " " << Ty << " *";
+  getAddr()->dump(Func);
+  Str << ", ";
+  getData()->dump(Func);
+  Str << ", beacon=";
+  getBeacon()->dump(Func);
+}
+
+template <class Machine>
+void InstX86BaseLabel<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  Str << getName(Func) << ":";
+}
+
+template <class Machine>
+void InstX86BaseLabel<Machine>::emitIAS(const Cfg *Func) const {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  Asm->BindLocalLabel(Number);
+}
+
+template <class Machine>
+void InstX86BaseLabel<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  Str << getName(Func) << ":";
+}
+
+template <class Machine>
+void InstX86BaseBr<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  Str << "\t";
+
+  if (Condition == InstX86Base<Machine>::Traits::Cond::Br_None) {
+    Str << "jmp";
+  } else {
+    Str << InstX86Base<Machine>::Traits::InstBrAttributes[Condition].EmitString;
+  }
+
+  if (Label) {
+    Str << "\t" << Label->getName(Func);
+  } else {
+    if (Condition == InstX86Base<Machine>::Traits::Cond::Br_None) {
+      Str << "\t" << getTargetFalse()->getAsmName();
+    } else {
+      Str << "\t" << getTargetTrue()->getAsmName();
+      if (getTargetFalse()) {
+        Str << "\n\tjmp\t" << getTargetFalse()->getAsmName();
+      }
+    }
+  }
+}
+
+template <class Machine>
+void InstX86BaseBr<Machine>::emitIAS(const Cfg *Func) const {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  if (Label) {
+    class Label *L = Asm->GetOrCreateLocalLabel(Label->getNumber());
+    // In all these cases, local Labels should only be used for Near.
+    const bool Near = true;
+    if (Condition == InstX86Base<Machine>::Traits::Cond::Br_None) {
+      Asm->jmp(L, Near);
+    } else {
+      Asm->j(Condition, L, Near);
+    }
+  } else {
+    // Pessimistically assume it's far. This only affects Labels that
+    // are not Bound.
+    const bool Near = false;
+    if (Condition == InstX86Base<Machine>::Traits::Cond::Br_None) {
+      class Label *L =
+          Asm->GetOrCreateCfgNodeLabel(getTargetFalse()->getIndex());
+      assert(!getTargetTrue());
+      Asm->jmp(L, Near);
+    } else {
+      class Label *L =
+          Asm->GetOrCreateCfgNodeLabel(getTargetTrue()->getIndex());
+      Asm->j(Condition, L, Near);
+      if (getTargetFalse()) {
+        class Label *L2 =
+            Asm->GetOrCreateCfgNodeLabel(getTargetFalse()->getIndex());
+        Asm->jmp(L2, Near);
+      }
+    }
+  }
+}
+
+template <class Machine>
+void InstX86BaseBr<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  Str << "br ";
+
+  if (Condition == InstX86Base<Machine>::Traits::Cond::Br_None) {
+    Str << "label %"
+        << (Label ? Label->getName(Func) : getTargetFalse()->getName());
+    return;
+  }
+
+  Str << InstX86Base<Machine>::Traits::InstBrAttributes[Condition]
+             .DisplayString;
+  if (Label) {
+    Str << ", label %" << Label->getName(Func);
+  } else {
+    Str << ", label %" << getTargetTrue()->getName();
+    if (getTargetFalse()) {
+      Str << ", label %" << getTargetFalse()->getName();
+    }
+  }
+}
+
+template <class Machine>
+void InstX86BaseJmp<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 1);
+  Str << "\tjmp\t*";
+  getJmpTarget()->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseJmp<Machine>::emitIAS(const Cfg *Func) const {
+  // Note: Adapted (mostly copied) from InstX86BaseCall<Machine>::emitIAS().
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  Operand *Target = getJmpTarget();
+  if (const auto Var = llvm::dyn_cast<Variable>(Target)) {
+    if (Var->hasReg()) {
+      Asm->jmp(InstX86Base<Machine>::Traits::RegisterSet::getEncodedGPR(
+          Var->getRegNum()));
+    } else {
+      // The jmp instruction with a memory operand should be possible
+      // to encode, but it isn't a valid sandboxed instruction, and
+      // there shouldn't be a register allocation issue to jump
+      // through a scratch register, so we don't really need to bother
+      // implementing it.
+      llvm::report_fatal_error("Assembler can't jmp to memory operand");
+    }
+  } else if (const auto Mem = llvm::dyn_cast<
+                 typename InstX86Base<Machine>::Traits::X86OperandMem>(
+                 Target)) {
+    (void)Mem;
+    assert(Mem->getSegmentRegister() ==
+           InstX86Base<Machine>::Traits::X86OperandMem::DefaultSegment);
+    llvm::report_fatal_error("Assembler can't jmp to memory operand");
+  } else if (const auto CR = llvm::dyn_cast<ConstantRelocatable>(Target)) {
+    assert(CR->getOffset() == 0 && "We only support jumping to a function");
+    Asm->jmp(CR);
+  } else if (const auto Imm = llvm::dyn_cast<ConstantInteger32>(Target)) {
+    // NaCl trampoline calls refer to an address within the sandbox directly.
+    // This is usually only needed for non-IRT builds and otherwise not
+    // very portable or stable. Usually this is only done for "calls"
+    // and not jumps.
+    // TODO(jvoung): Support this when there is a lowering that
+    // actually triggers this case.
+    (void)Imm;
+    llvm::report_fatal_error("Unexpected jmp to absolute address");
+  } else {
+    llvm::report_fatal_error("Unexpected operand type");
+  }
+}
+
+template <class Machine>
+void InstX86BaseJmp<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  Str << "jmp ";
+  getJmpTarget()->dump(Func);
+}
+
+template <class Machine>
+void InstX86BaseCall<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 1);
+  Str << "\tcall\t";
+  if (const auto CI = llvm::dyn_cast<ConstantInteger32>(getCallTarget())) {
+    // Emit without a leading '$'.
+    Str << CI->getValue();
+  } else if (const auto CallTarget =
+                 llvm::dyn_cast<ConstantRelocatable>(getCallTarget())) {
+    CallTarget->emitWithoutPrefix(Func->getTarget());
+  } else {
+    Str << "*";
+    getCallTarget()->emit(Func);
+  }
+  Func->getTarget()->resetStackAdjustment();
+}
+
+template <class Machine>
+void InstX86BaseCall<Machine>::emitIAS(const Cfg *Func) const {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  Operand *Target = getCallTarget();
+  if (const auto Var = llvm::dyn_cast<Variable>(Target)) {
+    if (Var->hasReg()) {
+      Asm->call(InstX86Base<Machine>::Traits::RegisterSet::getEncodedGPR(
+          Var->getRegNum()));
+    } else {
+      Asm->call(
+          static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+              Func->getTarget())
+              ->stackVarToAsmOperand(Var));
+    }
+  } else if (const auto Mem = llvm::dyn_cast<
+                 typename InstX86Base<Machine>::Traits::X86OperandMem>(
+                 Target)) {
+    assert(Mem->getSegmentRegister() ==
+           InstX86Base<Machine>::Traits::X86OperandMem::DefaultSegment);
+    Asm->call(Mem->toAsmAddress(Asm));
+  } else if (const auto CR = llvm::dyn_cast<ConstantRelocatable>(Target)) {
+    assert(CR->getOffset() == 0 && "We only support calling a function");
+    Asm->call(CR);
+  } else if (const auto Imm = llvm::dyn_cast<ConstantInteger32>(Target)) {
+    Asm->call(Immediate(Imm->getValue()));
+  } else {
+    llvm_unreachable("Unexpected operand type");
+  }
+  Func->getTarget()->resetStackAdjustment();
+}
+
+template <class Machine>
+void InstX86BaseCall<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  if (this->getDest()) {
+    this->dumpDest(Func);
+    Str << " = ";
+  }
+  Str << "call ";
+  getCallTarget()->dump(Func);
+}
+
+// The ShiftHack parameter is used to emit "cl" instead of "ecx" for
+// shift instructions, in order to be syntactically valid.  The
+// this->Opcode parameter needs to be char* and not IceString because of
+// template issues.
+template <class Machine>
+void InstX86Base<Machine>::emitTwoAddress(const char *Opcode, const Inst *Inst,
+                                          const Cfg *Func, bool ShiftHack) {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(Inst->getSrcSize() == 2);
+  Operand *Dest = Inst->getDest();
+  if (Dest == nullptr)
+    Dest = Inst->getSrc(0);
+  assert(Dest == Inst->getSrc(0));
+  Operand *Src1 = Inst->getSrc(1);
+  Str << "\t" << Opcode << InstX86Base<Machine>::getWidthString(Dest->getType())
+      << "\t";
+  const auto ShiftReg = llvm::dyn_cast<Variable>(Src1);
+  if (ShiftHack && ShiftReg &&
+      ShiftReg->getRegNum() ==
+          InstX86Base<Machine>::Traits::RegisterSet::Reg_ecx)
+    Str << "%cl";
+  else
+    Src1->emit(Func);
+  Str << ", ";
+  Dest->emit(Func);
+}
+
+template <class Machine>
+void emitIASOpTyGPR(const Cfg *Func, Type Ty, const Operand *Op,
+                    const typename InstX86Base<
+                        Machine>::Traits::Assembler::GPREmitterOneOp &Emitter) {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  if (const auto Var = llvm::dyn_cast<Variable>(Op)) {
+    if (Var->hasReg()) {
+      // We cheat a little and use GPRRegister even for byte operations.
+      typename InstX86Base<Machine>::Traits::RegisterSet::GPRRegister VarReg =
+          InstX86Base<Machine>::Traits::RegisterSet::getEncodedByteRegOrGPR(
+              Ty, Var->getRegNum());
+      (Asm->*(Emitter.Reg))(Ty, VarReg);
+    } else {
+      typename InstX86Base<Machine>::Traits::Address StackAddr(
+          static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+              Func->getTarget())
+              ->stackVarToAsmOperand(Var));
+      (Asm->*(Emitter.Addr))(Ty, StackAddr);
+    }
+  } else if (const auto Mem = llvm::dyn_cast<
+                 typename InstX86Base<Machine>::Traits::X86OperandMem>(Op)) {
+    Mem->emitSegmentOverride(Asm);
+    (Asm->*(Emitter.Addr))(Ty, Mem->toAsmAddress(Asm));
+  } else {
+    llvm_unreachable("Unexpected operand type");
+  }
+}
+
+template <class Machine, bool VarCanBeByte, bool SrcCanBeByte>
+void emitIASRegOpTyGPR(
+    const Cfg *Func, Type Ty, const Variable *Var, const Operand *Src,
+    const typename InstX86Base<Machine>::Traits::Assembler::GPREmitterRegOp
+        &Emitter) {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  assert(Var->hasReg());
+  // We cheat a little and use GPRRegister even for byte operations.
+  typename InstX86Base<Machine>::Traits::RegisterSet::GPRRegister VarReg =
+      VarCanBeByte
+          ? InstX86Base<Machine>::Traits::RegisterSet::getEncodedByteRegOrGPR(
+                Ty, Var->getRegNum())
+          : InstX86Base<Machine>::Traits::RegisterSet::getEncodedGPR(
+                Var->getRegNum());
+  if (const auto SrcVar = llvm::dyn_cast<Variable>(Src)) {
+    if (SrcVar->hasReg()) {
+      typename InstX86Base<Machine>::Traits::RegisterSet::GPRRegister SrcReg =
+          SrcCanBeByte
+              ? InstX86Base<Machine>::Traits::RegisterSet::
+                    getEncodedByteRegOrGPR(Ty, SrcVar->getRegNum())
+              : InstX86Base<Machine>::Traits::RegisterSet::getEncodedGPR(
+                    SrcVar->getRegNum());
+      (Asm->*(Emitter.GPRGPR))(Ty, VarReg, SrcReg);
+    } else {
+      typename InstX86Base<Machine>::Traits::Address SrcStackAddr =
+          static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+              Func->getTarget())
+              ->stackVarToAsmOperand(SrcVar);
+      (Asm->*(Emitter.GPRAddr))(Ty, VarReg, SrcStackAddr);
+    }
+  } else if (const auto Mem = llvm::dyn_cast<
+                 typename InstX86Base<Machine>::Traits::X86OperandMem>(Src)) {
+    Mem->emitSegmentOverride(Asm);
+    (Asm->*(Emitter.GPRAddr))(Ty, VarReg, Mem->toAsmAddress(Asm));
+  } else if (const auto Imm = llvm::dyn_cast<ConstantInteger32>(Src)) {
+    (Asm->*(Emitter.GPRImm))(Ty, VarReg, Immediate(Imm->getValue()));
+  } else if (const auto Reloc = llvm::dyn_cast<ConstantRelocatable>(Src)) {
+    AssemblerFixup *Fixup = Asm->createFixup(llvm::ELF::R_386_32, Reloc);
+    (Asm->*(Emitter.GPRImm))(Ty, VarReg, Immediate(Reloc->getOffset(), Fixup));
+  } else if (const auto Split = llvm::dyn_cast<
+                 typename InstX86Base<Machine>::Traits::VariableSplit>(Src)) {
+    (Asm->*(Emitter.GPRAddr))(Ty, VarReg, Split->toAsmAddress(Func));
+  } else {
+    llvm_unreachable("Unexpected operand type");
+  }
+}
+
+template <class Machine>
+void emitIASAddrOpTyGPR(
+    const Cfg *Func, Type Ty,
+    const typename InstX86Base<Machine>::Traits::Address &Addr,
+    const Operand *Src,
+    const typename InstX86Base<Machine>::Traits::Assembler::GPREmitterAddrOp
+        &Emitter) {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  // Src can only be Reg or Immediate.
+  if (const auto SrcVar = llvm::dyn_cast<Variable>(Src)) {
+    assert(SrcVar->hasReg());
+    typename InstX86Base<Machine>::Traits::RegisterSet::GPRRegister SrcReg =
+        InstX86Base<Machine>::Traits::RegisterSet::getEncodedByteRegOrGPR(
+            Ty, SrcVar->getRegNum());
+    (Asm->*(Emitter.AddrGPR))(Ty, Addr, SrcReg);
+  } else if (const auto Imm = llvm::dyn_cast<ConstantInteger32>(Src)) {
+    (Asm->*(Emitter.AddrImm))(Ty, Addr, Immediate(Imm->getValue()));
+  } else if (const auto Reloc = llvm::dyn_cast<ConstantRelocatable>(Src)) {
+    AssemblerFixup *Fixup = Asm->createFixup(llvm::ELF::R_386_32, Reloc);
+    (Asm->*(Emitter.AddrImm))(Ty, Addr, Immediate(Reloc->getOffset(), Fixup));
+  } else {
+    llvm_unreachable("Unexpected operand type");
+  }
+}
+
+template <class Machine>
+void emitIASAsAddrOpTyGPR(
+    const Cfg *Func, Type Ty, const Operand *Op0, const Operand *Op1,
+    const typename InstX86Base<Machine>::Traits::Assembler::GPREmitterAddrOp
+        &Emitter) {
+  if (const auto Op0Var = llvm::dyn_cast<Variable>(Op0)) {
+    assert(!Op0Var->hasReg());
+    typename InstX86Base<Machine>::Traits::Address StackAddr(
+        static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+            Func->getTarget())
+            ->stackVarToAsmOperand(Op0Var));
+    emitIASAddrOpTyGPR<Machine>(Func, Ty, StackAddr, Op1, Emitter);
+  } else if (const auto Op0Mem = llvm::dyn_cast<
+                 typename InstX86Base<Machine>::Traits::X86OperandMem>(Op0)) {
+    typename InstX86Base<Machine>::Traits::Assembler *Asm =
+        Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+    Op0Mem->emitSegmentOverride(Asm);
+    emitIASAddrOpTyGPR<Machine>(Func, Ty, Op0Mem->toAsmAddress(Asm), Op1,
+                                Emitter);
+  } else if (const auto Split = llvm::dyn_cast<
+                 typename InstX86Base<Machine>::Traits::VariableSplit>(Op0)) {
+    emitIASAddrOpTyGPR<Machine>(Func, Ty, Split->toAsmAddress(Func), Op1,
+                                Emitter);
+  } else {
+    llvm_unreachable("Unexpected operand type");
+  }
+}
+
+template <class Machine>
+void InstX86Base<Machine>::emitIASGPRShift(
+    const Cfg *Func, Type Ty, const Variable *Var, const Operand *Src,
+    const typename InstX86Base<Machine>::Traits::Assembler::GPREmitterShiftOp
+        &Emitter) {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  // Technically, the Dest Var can be mem as well, but we only use Reg.
+  // We can extend this to check Dest if we decide to use that form.
+  assert(Var->hasReg());
+  // We cheat a little and use GPRRegister even for byte operations.
+  typename InstX86Base<Machine>::Traits::RegisterSet::GPRRegister VarReg =
+      InstX86Base<Machine>::Traits::RegisterSet::getEncodedByteRegOrGPR(
+          Ty, Var->getRegNum());
+  // Src must be reg == ECX or an Imm8.
+  // This is asserted by the assembler.
+  if (const auto SrcVar = llvm::dyn_cast<Variable>(Src)) {
+    assert(SrcVar->hasReg());
+    typename InstX86Base<Machine>::Traits::RegisterSet::GPRRegister SrcReg =
+        InstX86Base<Machine>::Traits::RegisterSet::getEncodedByteRegOrGPR(
+            Ty, SrcVar->getRegNum());
+    (Asm->*(Emitter.GPRGPR))(Ty, VarReg, SrcReg);
+  } else if (const auto Imm = llvm::dyn_cast<ConstantInteger32>(Src)) {
+    (Asm->*(Emitter.GPRImm))(Ty, VarReg, Immediate(Imm->getValue()));
+  } else {
+    llvm_unreachable("Unexpected operand type");
+  }
+}
+
+template <class Machine>
+void emitIASGPRShiftDouble(
+    const Cfg *Func, const Variable *Dest, const Operand *Src1Op,
+    const Operand *Src2Op,
+    const typename InstX86Base<Machine>::Traits::Assembler::GPREmitterShiftD
+        &Emitter) {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  // Dest can be reg or mem, but we only use the reg variant.
+  assert(Dest->hasReg());
+  typename InstX86Base<Machine>::Traits::RegisterSet::GPRRegister DestReg =
+      InstX86Base<Machine>::Traits::RegisterSet::getEncodedGPR(
+          Dest->getRegNum());
+  // SrcVar1 must be reg.
+  const auto SrcVar1 = llvm::cast<Variable>(Src1Op);
+  assert(SrcVar1->hasReg());
+  typename InstX86Base<Machine>::Traits::RegisterSet::GPRRegister SrcReg =
+      InstX86Base<Machine>::Traits::RegisterSet::getEncodedGPR(
+          SrcVar1->getRegNum());
+  Type Ty = SrcVar1->getType();
+  // Src2 can be the implicit CL register or an immediate.
+  if (const auto Imm = llvm::dyn_cast<ConstantInteger32>(Src2Op)) {
+    (Asm->*(Emitter.GPRGPRImm))(Ty, DestReg, SrcReg,
+                                Immediate(Imm->getValue()));
+  } else {
+    assert(llvm::cast<Variable>(Src2Op)->getRegNum() ==
+           InstX86Base<Machine>::Traits::RegisterSet::Reg_ecx);
+    (Asm->*(Emitter.GPRGPR))(Ty, DestReg, SrcReg);
+  }
+}
+
+template <class Machine>
+void emitIASXmmShift(
+    const Cfg *Func, Type Ty, const Variable *Var, const Operand *Src,
+    const typename InstX86Base<Machine>::Traits::Assembler::XmmEmitterShiftOp
+        &Emitter) {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  assert(Var->hasReg());
+  typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister VarReg =
+      InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm(
+          Var->getRegNum());
+  if (const auto SrcVar = llvm::dyn_cast<Variable>(Src)) {
+    if (SrcVar->hasReg()) {
+      typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister SrcReg =
+          InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm(
+              SrcVar->getRegNum());
+      (Asm->*(Emitter.XmmXmm))(Ty, VarReg, SrcReg);
+    } else {
+      typename InstX86Base<Machine>::Traits::Address SrcStackAddr =
+          static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+              Func->getTarget())
+              ->stackVarToAsmOperand(SrcVar);
+      (Asm->*(Emitter.XmmAddr))(Ty, VarReg, SrcStackAddr);
+    }
+  } else if (const auto Mem = llvm::dyn_cast<
+                 typename InstX86Base<Machine>::Traits::X86OperandMem>(Src)) {
+    assert(Mem->getSegmentRegister() ==
+           InstX86Base<Machine>::Traits::X86OperandMem::DefaultSegment);
+    (Asm->*(Emitter.XmmAddr))(Ty, VarReg, Mem->toAsmAddress(Asm));
+  } else if (const auto Imm = llvm::dyn_cast<ConstantInteger32>(Src)) {
+    (Asm->*(Emitter.XmmImm))(Ty, VarReg, Immediate(Imm->getValue()));
+  } else {
+    llvm_unreachable("Unexpected operand type");
+  }
+}
+
+template <class Machine>
+void emitIASRegOpTyXMM(
+    const Cfg *Func, Type Ty, const Variable *Var, const Operand *Src,
+    const typename InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp
+        &Emitter) {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  assert(Var->hasReg());
+  typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister VarReg =
+      InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm(
+          Var->getRegNum());
+  if (const auto SrcVar = llvm::dyn_cast<Variable>(Src)) {
+    if (SrcVar->hasReg()) {
+      typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister SrcReg =
+          InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm(
+              SrcVar->getRegNum());
+      (Asm->*(Emitter.XmmXmm))(Ty, VarReg, SrcReg);
+    } else {
+      typename InstX86Base<Machine>::Traits::Address SrcStackAddr =
+          static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+              Func->getTarget())
+              ->stackVarToAsmOperand(SrcVar);
+      (Asm->*(Emitter.XmmAddr))(Ty, VarReg, SrcStackAddr);
+    }
+  } else if (const auto Mem = llvm::dyn_cast<
+                 typename InstX86Base<Machine>::Traits::X86OperandMem>(Src)) {
+    assert(Mem->getSegmentRegister() ==
+           InstX86Base<Machine>::Traits::X86OperandMem::DefaultSegment);
+    (Asm->*(Emitter.XmmAddr))(Ty, VarReg, Mem->toAsmAddress(Asm));
+  } else if (const auto Imm = llvm::dyn_cast<Constant>(Src)) {
+    (Asm->*(Emitter.XmmAddr))(
+        Ty, VarReg,
+        InstX86Base<Machine>::Traits::Address::ofConstPool(Asm, Imm));
+  } else {
+    llvm_unreachable("Unexpected operand type");
+  }
+}
+
+template <class Machine, typename DReg_t, typename SReg_t,
+          DReg_t (*destEnc)(int32_t), SReg_t (*srcEnc)(int32_t)>
+void emitIASCastRegOp(const Cfg *Func, Type DispatchTy, const Variable *Dest,
+                      const Operand *Src,
+                      const typename InstX86Base<Machine>::Traits::Assembler::
+                          template CastEmitterRegOp<DReg_t, SReg_t> &Emitter) {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  assert(Dest->hasReg());
+  DReg_t DestReg = destEnc(Dest->getRegNum());
+  if (const auto SrcVar = llvm::dyn_cast<Variable>(Src)) {
+    if (SrcVar->hasReg()) {
+      SReg_t SrcReg = srcEnc(SrcVar->getRegNum());
+      (Asm->*(Emitter.RegReg))(DispatchTy, DestReg, SrcReg);
+    } else {
+      typename InstX86Base<Machine>::Traits::Address SrcStackAddr =
+          static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+              Func->getTarget())
+              ->stackVarToAsmOperand(SrcVar);
+      (Asm->*(Emitter.RegAddr))(DispatchTy, DestReg, SrcStackAddr);
+    }
+  } else if (const auto Mem = llvm::dyn_cast<
+                 typename InstX86Base<Machine>::Traits::X86OperandMem>(Src)) {
+    Mem->emitSegmentOverride(Asm);
+    (Asm->*(Emitter.RegAddr))(DispatchTy, DestReg, Mem->toAsmAddress(Asm));
+  } else {
+    llvm_unreachable("Unexpected operand type");
+  }
+}
+
+template <class Machine, typename DReg_t, typename SReg_t,
+          DReg_t (*destEnc)(int32_t), SReg_t (*srcEnc)(int32_t)>
+void emitIASThreeOpImmOps(
+    const Cfg *Func, Type DispatchTy, const Variable *Dest, const Operand *Src0,
+    const Operand *Src1,
+    const typename InstX86Base<Machine>::Traits::Assembler::
+        template ThreeOpImmEmitter<DReg_t, SReg_t> Emitter) {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  // This only handles Dest being a register, and Src1 being an immediate.
+  assert(Dest->hasReg());
+  DReg_t DestReg = destEnc(Dest->getRegNum());
+  Immediate Imm(llvm::cast<ConstantInteger32>(Src1)->getValue());
+  if (const auto SrcVar = llvm::dyn_cast<Variable>(Src0)) {
+    if (SrcVar->hasReg()) {
+      SReg_t SrcReg = srcEnc(SrcVar->getRegNum());
+      (Asm->*(Emitter.RegRegImm))(DispatchTy, DestReg, SrcReg, Imm);
+    } else {
+      typename InstX86Base<Machine>::Traits::Address SrcStackAddr =
+          static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+              Func->getTarget())
+              ->stackVarToAsmOperand(SrcVar);
+      (Asm->*(Emitter.RegAddrImm))(DispatchTy, DestReg, SrcStackAddr, Imm);
+    }
+  } else if (const auto Mem = llvm::dyn_cast<
+                 typename InstX86Base<Machine>::Traits::X86OperandMem>(Src0)) {
+    Mem->emitSegmentOverride(Asm);
+    (Asm->*(Emitter.RegAddrImm))(DispatchTy, DestReg, Mem->toAsmAddress(Asm),
+                                 Imm);
+  } else {
+    llvm_unreachable("Unexpected operand type");
+  }
+}
+
+template <class Machine>
+void emitIASMovlikeXMM(
+    const Cfg *Func, const Variable *Dest, const Operand *Src,
+    const typename InstX86Base<Machine>::Traits::Assembler::XmmEmitterMovOps
+        Emitter) {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  if (Dest->hasReg()) {
+    typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister DestReg =
+        InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm(
+            Dest->getRegNum());
+    if (const auto SrcVar = llvm::dyn_cast<Variable>(Src)) {
+      if (SrcVar->hasReg()) {
+        (Asm->*(Emitter.XmmXmm))(
+            DestReg, InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm(
+                         SrcVar->getRegNum()));
+      } else {
+        typename InstX86Base<Machine>::Traits::Address StackAddr(
+            static_cast<typename InstX86Base<Machine>::Traits::TargetLowering
+                            *>(Func->getTarget())
+                ->stackVarToAsmOperand(SrcVar));
+        (Asm->*(Emitter.XmmAddr))(DestReg, StackAddr);
+      }
+    } else if (const auto SrcMem = llvm::dyn_cast<
+                   typename InstX86Base<Machine>::Traits::X86OperandMem>(Src)) {
+      assert(SrcMem->getSegmentRegister() ==
+             InstX86Base<Machine>::Traits::X86OperandMem::DefaultSegment);
+      (Asm->*(Emitter.XmmAddr))(DestReg, SrcMem->toAsmAddress(Asm));
+    } else {
+      llvm_unreachable("Unexpected operand type");
+    }
+  } else {
+    typename InstX86Base<Machine>::Traits::Address StackAddr(
+        static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+            Func->getTarget())
+            ->stackVarToAsmOperand(Dest));
+    // Src must be a register in this case.
+    const auto SrcVar = llvm::cast<Variable>(Src);
+    assert(SrcVar->hasReg());
+    (Asm->*(Emitter.AddrXmm))(
+        StackAddr, InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm(
+                       SrcVar->getRegNum()));
+  }
+}
+
+#if 0 
+TODO(jpp)

[jvoung (off chromium), 2015/07/06 18:58:45]

Can this TODO be addressed as part of this CL (seems like it has something to do
with instantiating static data)?

[John, 2015/07/06 22:30:09]

It had been handled, I just forgot to remove it.

+// In-place ops
+template <> const char *InstX86BaseBswap::Opcode = "bswap";
+template <> const char *InstX86BaseNeg::Opcode = "neg";
+// Unary ops
+template <> const char *InstX86BaseBsf::Opcode = "bsf";
+template <> const char *InstX86BaseBsr::Opcode = "bsr";
+template <> const char *InstX86BaseLea::Opcode = "lea";
+template <> const char *InstX86BaseMovd::Opcode = "movd";
+template <> const char *InstX86BaseMovsx::Opcode = "movs";
+template <> const char *InstX86BaseMovzx::Opcode = "movz";
+template <> const char *InstX86BaseSqrtss::Opcode = "sqrtss";
+template <> const char *InstX86BaseCbwdq::Opcode = "cbw/cwd/cdq";
+// Mov-like ops
+template <> const char *InstX86BaseMov::Opcode = "mov";
+template <> const char *InstX86BaseMovp::Opcode = "movups";
+template <> const char *InstX86BaseMovq::Opcode = "movq";
+// Binary ops
+template <> const char *InstX86BaseAdd::Opcode = "add";
+template <> const char *InstX86BaseAddRMW::Opcode = "add";
+template <> const char *InstX86BaseAddps::Opcode = "addps";
+template <> const char *InstX86BaseAdc::Opcode = "adc";
+template <> const char *InstX86BaseAdcRMW::Opcode = "adc";
+template <> const char *InstX86BaseAddss::Opcode = "addss";
+template <> const char *InstX86BasePadd::Opcode = "padd";
+template <> const char *InstX86BaseSub::Opcode = "sub";
+template <> const char *InstX86BaseSubRMW::Opcode = "sub";
+template <> const char *InstX86BaseSubps::Opcode = "subps";
+template <> const char *InstX86BaseSubss::Opcode = "subss";
+template <> const char *InstX86BaseSbb::Opcode = "sbb";
+template <> const char *InstX86BaseSbbRMW::Opcode = "sbb";
+template <> const char *InstX86BasePsub::Opcode = "psub";
+template <> const char *InstX86BaseAnd::Opcode = "and";
+template <> const char *InstX86BaseAndRMW::Opcode = "and";
+template <> const char *InstX86BasePand::Opcode = "pand";
+template <> const char *InstX86BasePandn::Opcode = "pandn";
+template <> const char *InstX86BaseOr::Opcode = "or";
+template <> const char *InstX86BaseOrRMW::Opcode = "or";
+template <> const char *InstX86BasePor::Opcode = "por";
+template <> const char *InstX86BaseXor::Opcode = "xor";
+template <> const char *InstX86BaseXorRMW::Opcode = "xor";
+template <> const char *InstX86BasePxor::Opcode = "pxor";
+template <> const char *InstX86BaseImul::Opcode = "imul";
+template <> const char *InstX86BaseMulps::Opcode = "mulps";
+template <> const char *InstX86BaseMulss::Opcode = "mulss";
+template <> const char *InstX86BasePmull::Opcode = "pmull";
+template <> const char *InstX86BasePmuludq::Opcode = "pmuludq";
+template <> const char *InstX86BaseDiv::Opcode = "div";
+template <> const char *InstX86BaseDivps::Opcode = "divps";
+template <> const char *InstX86BaseIdiv::Opcode = "idiv";
+template <> const char *InstX86BaseDivss::Opcode = "divss";
+template <> const char *InstX86BaseRol::Opcode = "rol";
+template <> const char *InstX86BaseShl::Opcode = "shl";
+template <> const char *InstX86BasePsll::Opcode = "psll";
+template <> const char *InstX86BaseShr::Opcode = "shr";
+template <> const char *InstX86BaseSar::Opcode = "sar";
+template <> const char *InstX86BasePsra::Opcode = "psra";
+template <> const char *InstX86BasePsrl::Opcode = "psrl";
+template <> const char *InstX86BasePcmpeq::Opcode = "pcmpeq";
+template <> const char *InstX86BasePcmpgt::Opcode = "pcmpgt";
+template <> const char *InstX86BaseMovssRegs::Opcode = "movss";
+// Ternary ops
+template <> const char *InstX86BaseInsertps::Opcode = "insertps";
+template <> const char *InstX86BaseShufps::Opcode = "shufps";
+template <> const char *InstX86BasePinsr::Opcode = "pinsr";
+template <> const char *InstX86BaseBlendvps::Opcode = "blendvps";
+template <> const char *InstX86BasePblendvb::Opcode = "pblendvb";
+// Three address ops
+template <> const char *InstX86BasePextr::Opcode = "pextr";
+template <> const char *InstX86BasePshufd::Opcode = "pshufd";
+
+// Inplace GPR ops
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterOneOp InstX86BaseBswap::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::bswap, nullptr /* only a reg form exists */
+};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterOneOp InstX86BaseNeg::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::neg, &InstX86Base<Machine>::Traits::Assembler::neg};

[jvoung (off chromium), 2015/07/06 18:58:45]

wonder why clang-format doesn't move stuff to the next line (spilling past 80
col)

Oh, because of #if 0?

[John, 2015/07/06 22:30:09]

Maybe?... :)

+
+// Unary GPR ops
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterRegOp InstX86BaseBsf::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::bsf, &InstX86Base<Machine>::Traits::Assembler::bsf, nullptr};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterRegOp InstX86BaseBsr::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::bsr, &InstX86Base<Machine>::Traits::Assembler::bsr, nullptr};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterRegOp InstX86BaseLea::Emitter = {
+    /* reg/reg and reg/imm are illegal */ nullptr, &InstX86Base<Machine>::Traits::Assembler::lea,
+    nullptr};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterRegOp InstX86BaseMovsx::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::movsx, &InstX86Base<Machine>::Traits::Assembler::movsx, nullptr};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterRegOp InstX86BaseMovzx::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::movzx, &InstX86Base<Machine>::Traits::Assembler::movzx, nullptr};
+
+// Unary XMM ops
+template <>
+const InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp InstX86BaseSqrtss::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::sqrtss, &InstX86Base<Machine>::Traits::Assembler::sqrtss};
+
+// Binary GPR ops
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterRegOp InstX86BaseAdd::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::add, &InstX86Base<Machine>::Traits::Assembler::add,
+    &InstX86Base<Machine>::Traits::Assembler::add};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterAddrOp InstX86BaseAddRMW::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::add, &InstX86Base<Machine>::Traits::Assembler::add};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterRegOp InstX86BaseAdc::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::adc, &InstX86Base<Machine>::Traits::Assembler::adc,
+    &InstX86Base<Machine>::Traits::Assembler::adc};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterAddrOp InstX86BaseAdcRMW::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::adc, &InstX86Base<Machine>::Traits::Assembler::adc};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterRegOp InstX86BaseAnd::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::And, &InstX86Base<Machine>::Traits::Assembler::And,
+    &InstX86Base<Machine>::Traits::Assembler::And};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterAddrOp InstX86BaseAndRMW::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::And, &InstX86Base<Machine>::Traits::Assembler::And};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterRegOp InstX86BaseOr::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::Or, &InstX86Base<Machine>::Traits::Assembler::Or,
+    &InstX86Base<Machine>::Traits::Assembler::Or};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterAddrOp InstX86BaseOrRMW::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::Or, &InstX86Base<Machine>::Traits::Assembler::Or};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterRegOp InstX86BaseSbb::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::sbb, &InstX86Base<Machine>::Traits::Assembler::sbb,
+    &InstX86Base<Machine>::Traits::Assembler::sbb};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterAddrOp InstX86BaseSbbRMW::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::sbb, &InstX86Base<Machine>::Traits::Assembler::sbb};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterRegOp InstX86BaseSub::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::sub, &InstX86Base<Machine>::Traits::Assembler::sub,
+    &InstX86Base<Machine>::Traits::Assembler::sub};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterAddrOp InstX86BaseSubRMW::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::sub, &InstX86Base<Machine>::Traits::Assembler::sub};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterRegOp InstX86BaseXor::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::Xor, &InstX86Base<Machine>::Traits::Assembler::Xor,
+    &InstX86Base<Machine>::Traits::Assembler::Xor};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterAddrOp InstX86BaseXorRMW::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::Xor, &InstX86Base<Machine>::Traits::Assembler::Xor};
+
+// Binary Shift GPR ops
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterShiftOp InstX86BaseRol::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::rol, &InstX86Base<Machine>::Traits::Assembler::rol};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterShiftOp InstX86BaseSar::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::sar, &InstX86Base<Machine>::Traits::Assembler::sar};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterShiftOp InstX86BaseShl::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::shl, &InstX86Base<Machine>::Traits::Assembler::shl};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::GPREmitterShiftOp InstX86BaseShr::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::shr, &InstX86Base<Machine>::Traits::Assembler::shr};
+
+// Binary XMM ops
+template <>
+const InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp InstX86BaseAddss::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::addss, &InstX86Base<Machine>::Traits::Assembler::addss};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp InstX86BaseAddps::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::addps, &InstX86Base<Machine>::Traits::Assembler::addps};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp InstX86BaseDivss::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::divss, &InstX86Base<Machine>::Traits::Assembler::divss};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp InstX86BaseDivps::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::divps, &InstX86Base<Machine>::Traits::Assembler::divps};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp InstX86BaseMulss::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::mulss, &InstX86Base<Machine>::Traits::Assembler::mulss};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp InstX86BaseMulps::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::mulps, &InstX86Base<Machine>::Traits::Assembler::mulps};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp InstX86BasePadd::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::padd, &InstX86Base<Machine>::Traits::Assembler::padd};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp InstX86BasePand::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::pand, &InstX86Base<Machine>::Traits::Assembler::pand};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp InstX86BasePandn::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::pandn, &InstX86Base<Machine>::Traits::Assembler::pandn};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp InstX86BasePcmpeq::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::pcmpeq, &InstX86Base<Machine>::Traits::Assembler::pcmpeq};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp InstX86BasePcmpgt::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::pcmpgt, &InstX86Base<Machine>::Traits::Assembler::pcmpgt};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp InstX86BasePmull::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::pmull, &InstX86Base<Machine>::Traits::Assembler::pmull};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp InstX86BasePmuludq::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::pmuludq, &InstX86Base<Machine>::Traits::Assembler::pmuludq};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp InstX86BasePor::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::por, &InstX86Base<Machine>::Traits::Assembler::por};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp InstX86BasePsub::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::psub, &InstX86Base<Machine>::Traits::Assembler::psub};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp InstX86BasePxor::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::pxor, &InstX86Base<Machine>::Traits::Assembler::pxor};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp InstX86BaseSubss::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::subss, &InstX86Base<Machine>::Traits::Assembler::subss};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp InstX86BaseSubps::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::subps, &InstX86Base<Machine>::Traits::Assembler::subps};
+
+// Binary XMM Shift ops
+template <>
+const InstX86Base<Machine>::Traits::Assembler::XmmEmitterShiftOp InstX86BasePsll::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::psll, &InstX86Base<Machine>::Traits::Assembler::psll,
+    &InstX86Base<Machine>::Traits::Assembler::psll};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::XmmEmitterShiftOp InstX86BasePsra::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::psra, &InstX86Base<Machine>::Traits::Assembler::psra,
+    &InstX86Base<Machine>::Traits::Assembler::psra};
+template <>
+const InstX86Base<Machine>::Traits::Assembler::XmmEmitterShiftOp InstX86BasePsrl::Emitter = {
+    &InstX86Base<Machine>::Traits::Assembler::psrl, &InstX86Base<Machine>::Traits::Assembler::psrl,
+    &InstX86Base<Machine>::Traits::Assembler::psrl};
+#endif // 0
+
+template <class Machine>
+void InstX86BaseSqrtss<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 1);
+  Type Ty = this->getSrc(0)->getType();
+  assert(isScalarFloatingType(Ty));
+  Str << "\tsqrt" << InstX86Base<Machine>::Traits::TypeAttributes[Ty].SdSsString
+      << "\t";
+  this->getSrc(0)->emit(Func);
+  Str << ", ";
+  this->getDest()->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseAddss<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  char buf[30];
+  snprintf(
+      buf, llvm::array_lengthof(buf), "add%s",
+      InstX86Base<Machine>::Traits::TypeAttributes[this->getDest()->getType()]
+          .SdSsString);
+  this->emitTwoAddress(buf, this, Func);
+}
+
+template <class Machine>
+void InstX86BasePadd<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  char buf[30];
+  snprintf(
+      buf, llvm::array_lengthof(buf), "padd%s",
+      InstX86Base<Machine>::Traits::TypeAttributes[this->getDest()->getType()]
+          .PackString);
+  this->emitTwoAddress(buf, this, Func);
+}
+
+template <class Machine>
+void InstX86BasePmull<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  char buf[30];
+  bool TypesAreValid = this->getDest()->getType() == IceType_v4i32 ||
+                       this->getDest()->getType() == IceType_v8i16;
+  bool InstructionSetIsValid =
+      this->getDest()->getType() == IceType_v8i16 ||
+      static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+          Func->getTarget())
+              ->getInstructionSet() >= InstX86Base<Machine>::Traits::SSE4_1;
+  (void)TypesAreValid;
+  (void)InstructionSetIsValid;
+  assert(TypesAreValid);
+  assert(InstructionSetIsValid);
+  snprintf(
+      buf, llvm::array_lengthof(buf), "pmull%s",
+      InstX86Base<Machine>::Traits::TypeAttributes[this->getDest()->getType()]
+          .PackString);
+  this->emitTwoAddress(buf, this, Func);
+}
+
+template <class Machine>
+void InstX86BasePmull<Machine>::emitIAS(const Cfg *Func) const {
+  Type Ty = this->getDest()->getType();
+  bool TypesAreValid = Ty == IceType_v4i32 || Ty == IceType_v8i16;
+  bool InstructionSetIsValid =
+      Ty == IceType_v8i16 ||
+      static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+          Func->getTarget())
+              ->getInstructionSet() >= InstX86Base<Machine>::Traits::SSE4_1;
+  (void)TypesAreValid;
+  (void)InstructionSetIsValid;
+  assert(TypesAreValid);
+  assert(InstructionSetIsValid);
+  assert(this->getSrcSize() == 2);
+  Type ElementTy = typeElementType(Ty);
+  emitIASRegOpTyXMM<Machine>(Func, ElementTy, this->getDest(), this->getSrc(1),
+                             this->Emitter);
+}
+
+template <class Machine>
+void InstX86BaseSubss<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  char buf[30];
+  snprintf(
+      buf, llvm::array_lengthof(buf), "sub%s",
+      InstX86Base<Machine>::Traits::TypeAttributes[this->getDest()->getType()]
+          .SdSsString);
+  this->emitTwoAddress(buf, this, Func);
+}
+
+template <class Machine>
+void InstX86BasePsub<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  char buf[30];
+  snprintf(
+      buf, llvm::array_lengthof(buf), "psub%s",
+      InstX86Base<Machine>::Traits::TypeAttributes[this->getDest()->getType()]
+          .PackString);
+  this->emitTwoAddress(buf, this, Func);
+}
+
+template <class Machine>
+void InstX86BaseMulss<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  char buf[30];
+  snprintf(
+      buf, llvm::array_lengthof(buf), "mul%s",
+      InstX86Base<Machine>::Traits::TypeAttributes[this->getDest()->getType()]
+          .SdSsString);
+  this->emitTwoAddress(buf, this, Func);
+}
+
+template <class Machine>
+void InstX86BasePmuludq<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  assert(this->getSrc(0)->getType() == IceType_v4i32 &&
+         this->getSrc(1)->getType() == IceType_v4i32);
+  this->emitTwoAddress(this->Opcode, this, Func);
+}
+
+template <class Machine>
+void InstX86BaseDivss<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  char buf[30];
+  snprintf(
+      buf, llvm::array_lengthof(buf), "div%s",
+      InstX86Base<Machine>::Traits::TypeAttributes[this->getDest()->getType()]
+          .SdSsString);
+  this->emitTwoAddress(buf, this, Func);
+}
+
+template <class Machine>
+void InstX86BaseDiv<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 3);
+  Operand *Src1 = this->getSrc(1);
+  Str << "\t" << this->Opcode << this->getWidthString(Src1->getType()) << "\t";
+  Src1->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseDiv<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 3);
+  const Operand *Src = this->getSrc(1);
+  Type Ty = Src->getType();
+  static const typename InstX86Base<Machine>::Traits::Assembler::GPREmitterOneOp
+      Emitter = {&InstX86Base<Machine>::Traits::Assembler::div,
+                 &InstX86Base<Machine>::Traits::Assembler::div};
+  emitIASOpTyGPR<Machine>(Func, Ty, Src, Emitter);
+}
+
+template <class Machine>
+void InstX86BaseIdiv<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 3);
+  Operand *Src1 = this->getSrc(1);
+  Str << "\t" << this->Opcode << this->getWidthString(Src1->getType()) << "\t";
+  Src1->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseIdiv<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 3);
+  const Operand *Src = this->getSrc(1);
+  Type Ty = Src->getType();
+  static const typename InstX86Base<Machine>::Traits::Assembler::GPREmitterOneOp
+      Emitter = {&InstX86Base<Machine>::Traits::Assembler::idiv,
+                 &InstX86Base<Machine>::Traits::Assembler::idiv};
+  emitIASOpTyGPR<Machine>(Func, Ty, Src, Emitter);
+}
+
+// pblendvb and blendvps take xmm0 as a final implicit argument.
+template <class Machine>
+void emitVariableBlendInst(const char *Opcode, const Inst *Inst,
+                           const Cfg *Func) {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(Inst->getSrcSize() == 3);
+  assert(llvm::cast<Variable>(Inst->getSrc(2))->getRegNum() ==
+         InstX86Base<Machine>::Traits::RegisterSet::Reg_xmm0);
+  Str << "\t" << Opcode << "\t";
+  Inst->getSrc(1)->emit(Func);
+  Str << ", ";
+  Inst->getDest()->emit(Func);
+}
+
+template <class Machine>
+void emitIASVariableBlendInst(
+    const Inst *Inst, const Cfg *Func,
+    const typename InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp
+        &Emitter) {
+  assert(Inst->getSrcSize() == 3);
+  assert(llvm::cast<Variable>(Inst->getSrc(2))->getRegNum() ==
+         InstX86Base<Machine>::Traits::RegisterSet::Reg_xmm0);
+  const Variable *Dest = Inst->getDest();
+  const Operand *Src = Inst->getSrc(1);
+  emitIASRegOpTyXMM<Machine>(Func, Dest->getType(), Dest, Src, Emitter);
+}
+
+template <class Machine>
+void InstX86BaseBlendvps<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  assert(static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+             Func->getTarget())
+             ->getInstructionSet() >= InstX86Base<Machine>::Traits::SSE4_1);
+  emitVariableBlendInst<Machine>(this->Opcode, this, Func);
+}
+
+template <class Machine>
+void InstX86BaseBlendvps<Machine>::emitIAS(const Cfg *Func) const {
+  assert(static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+             Func->getTarget())
+             ->getInstructionSet() >= InstX86Base<Machine>::Traits::SSE4_1);
+  static const typename InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp
+      Emitter = {&InstX86Base<Machine>::Traits::Assembler::blendvps,
+                 &InstX86Base<Machine>::Traits::Assembler::blendvps};
+  emitIASVariableBlendInst<Machine>(this, Func, Emitter);
+}
+
+template <class Machine>
+void InstX86BasePblendvb<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  assert(static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+             Func->getTarget())
+             ->getInstructionSet() >= InstX86Base<Machine>::Traits::SSE4_1);
+  emitVariableBlendInst<Machine>(this->Opcode, this, Func);
+}
+
+template <class Machine>
+void InstX86BasePblendvb<Machine>::emitIAS(const Cfg *Func) const {
+  assert(static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+             Func->getTarget())
+             ->getInstructionSet() >= InstX86Base<Machine>::Traits::SSE4_1);
+  static const typename InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp
+      Emitter = {&InstX86Base<Machine>::Traits::Assembler::pblendvb,
+                 &InstX86Base<Machine>::Traits::Assembler::pblendvb};
+  emitIASVariableBlendInst<Machine>(this, Func, Emitter);
+}
+
+template <class Machine>
+void InstX86BaseImul<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 2);
+  Variable *Dest = this->getDest();
+  if (isByteSizedArithType(Dest->getType())) {
+    // The 8-bit version of imul only allows the form "imul r/m8".
+    const auto Src0Var = llvm::dyn_cast<Variable>(this->getSrc(0));
+    (void)Src0Var;
+    assert(Src0Var &&
+           Src0Var->getRegNum() ==
+               InstX86Base<Machine>::Traits::RegisterSet::Reg_eax);
+    Str << "\timulb\t";
+    this->getSrc(1)->emit(Func);
+  } else if (llvm::isa<Constant>(this->getSrc(1))) {
+    Str << "\timul" << this->getWidthString(Dest->getType()) << "\t";
+    this->getSrc(1)->emit(Func);
+    Str << ", ";
+    this->getSrc(0)->emit(Func);
+    Str << ", ";
+    Dest->emit(Func);
+  } else {
+    this->emitTwoAddress("imul", this, Func);
+  }
+}
+
+template <class Machine>
+void InstX86BaseImul<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 2);
+  const Variable *Var = this->getDest();
+  Type Ty = Var->getType();
+  const Operand *Src = this->getSrc(1);
+  if (isByteSizedArithType(Ty)) {
+    // The 8-bit version of imul only allows the form "imul r/m8".
+    const auto Src0Var = llvm::dyn_cast<Variable>(this->getSrc(0));
+    (void)Src0Var;
+    assert(Src0Var &&
+           Src0Var->getRegNum() ==
+               InstX86Base<Machine>::Traits::RegisterSet::Reg_eax);
+    static const typename InstX86Base<
+        Machine>::Traits::Assembler::GPREmitterOneOp Emitter = {
+        &InstX86Base<Machine>::Traits::Assembler::imul,
+        &InstX86Base<Machine>::Traits::Assembler::imul};
+    emitIASOpTyGPR<Machine>(Func, Ty, this->getSrc(1), Emitter);
+  } else {
+    // We only use imul as a two-address instruction even though
+    // there is a 3 operand version when one of the operands is a constant.
+    assert(Var == this->getSrc(0));
+    static const typename InstX86Base<
+        Machine>::Traits::Assembler::GPREmitterRegOp Emitter = {
+        &InstX86Base<Machine>::Traits::Assembler::imul,
+        &InstX86Base<Machine>::Traits::Assembler::imul,
+        &InstX86Base<Machine>::Traits::Assembler::imul};
+    emitIASRegOpTyGPR<Machine>(Func, Ty, Var, Src, Emitter);
+  }
+}
+
+template <class Machine>
+void InstX86BaseInsertps<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 3);
+  assert(static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+             Func->getTarget())
+             ->getInstructionSet() >= InstX86Base<Machine>::Traits::SSE4_1);
+  const Variable *Dest = this->getDest();
+  assert(Dest == this->getSrc(0));
+  Type Ty = Dest->getType();
+  static const typename InstX86Base<Machine>::Traits::Assembler::
+      template ThreeOpImmEmitter<
+          typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister,
+          typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister>
+          Emitter = {&InstX86Base<Machine>::Traits::Assembler::insertps,
+                     &InstX86Base<Machine>::Traits::Assembler::insertps};
+  emitIASThreeOpImmOps<
+      Machine, typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister,
+      typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister,
+      InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm,
+      InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm>(
+      Func, Ty, Dest, this->getSrc(1), this->getSrc(2), Emitter);
+}
+
+template <class Machine>
+void InstX86BaseCbwdq<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 1);
+  Operand *Src0 = this->getSrc(0);
+  assert(llvm::isa<Variable>(Src0));
+  assert(llvm::cast<Variable>(Src0)->getRegNum() ==
+         InstX86Base<Machine>::Traits::RegisterSet::Reg_eax);
+  switch (Src0->getType()) {
+  default:
+    llvm_unreachable("unexpected source type!");
+    break;
+  case IceType_i8:
+    assert(this->getDest()->getRegNum() ==
+           InstX86Base<Machine>::Traits::RegisterSet::Reg_eax);
+    Str << "\tcbtw";
+    break;
+  case IceType_i16:
+    assert(this->getDest()->getRegNum() ==
+           InstX86Base<Machine>::Traits::RegisterSet::Reg_edx);
+    Str << "\tcwtd";
+    break;
+  case IceType_i32:
+    assert(this->getDest()->getRegNum() ==
+           InstX86Base<Machine>::Traits::RegisterSet::Reg_edx);
+    Str << "\tcltd";
+    break;
+  }
+}
+
+template <class Machine>
+void InstX86BaseCbwdq<Machine>::emitIAS(const Cfg *Func) const {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  assert(this->getSrcSize() == 1);
+  Operand *Src0 = this->getSrc(0);
+  assert(llvm::isa<Variable>(Src0));
+  assert(llvm::cast<Variable>(Src0)->getRegNum() ==
+         InstX86Base<Machine>::Traits::RegisterSet::Reg_eax);
+  switch (Src0->getType()) {
+  default:
+    llvm_unreachable("unexpected source type!");
+    break;
+  case IceType_i8:
+    assert(this->getDest()->getRegNum() ==
+           InstX86Base<Machine>::Traits::RegisterSet::Reg_eax);
+    Asm->cbw();
+    break;
+  case IceType_i16:
+    assert(this->getDest()->getRegNum() ==
+           InstX86Base<Machine>::Traits::RegisterSet::Reg_edx);
+    Asm->cwd();
+    break;
+  case IceType_i32:
+    assert(this->getDest()->getRegNum() ==
+           InstX86Base<Machine>::Traits::RegisterSet::Reg_edx);
+    Asm->cdq();
+    break;
+  }
+}
+
+template <class Machine>
+void InstX86BaseMul<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 2);
+  assert(llvm::isa<Variable>(this->getSrc(0)));
+  assert(llvm::cast<Variable>(this->getSrc(0))->getRegNum() ==
+         InstX86Base<Machine>::Traits::RegisterSet::Reg_eax);
+  assert(
+      this->getDest()->getRegNum() ==
+      InstX86Base<Machine>::Traits::RegisterSet::Reg_eax); // TODO: allow edx?
+  Str << "\tmul" << this->getWidthString(this->getDest()->getType()) << "\t";
+  this->getSrc(1)->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseMul<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 2);
+  assert(llvm::isa<Variable>(this->getSrc(0)));
+  assert(llvm::cast<Variable>(this->getSrc(0))->getRegNum() ==
+         InstX86Base<Machine>::Traits::RegisterSet::Reg_eax);
+  assert(
+      this->getDest()->getRegNum() ==
+      InstX86Base<Machine>::Traits::RegisterSet::Reg_eax); // TODO: allow edx?
+  const Operand *Src = this->getSrc(1);
+  Type Ty = Src->getType();
+  static const typename InstX86Base<Machine>::Traits::Assembler::GPREmitterOneOp
+      Emitter = {&InstX86Base<Machine>::Traits::Assembler::mul,
+                 &InstX86Base<Machine>::Traits::Assembler::mul};
+  emitIASOpTyGPR<Machine>(Func, Ty, Src, Emitter);
+}
+
+template <class Machine>
+void InstX86BaseMul<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  this->dumpDest(Func);
+  Str << " = mul." << this->getDest()->getType() << " ";
+  this->dumpSources(Func);
+}
+
+template <class Machine>
+void InstX86BaseShld<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  Variable *Dest = this->getDest();
+  assert(this->getSrcSize() == 3);
+  assert(Dest == this->getSrc(0));
+  Str << "\tshld" << this->getWidthString(Dest->getType()) << "\t";
+  if (const auto ShiftReg = llvm::dyn_cast<Variable>(this->getSrc(2))) {
+    (void)ShiftReg;
+    assert(ShiftReg->getRegNum() ==
+           InstX86Base<Machine>::Traits::RegisterSet::Reg_ecx);
+    Str << "%cl";
+  } else {
+    this->getSrc(2)->emit(Func);
+  }
+  Str << ", ";
+  this->getSrc(1)->emit(Func);
+  Str << ", ";
+  Dest->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseShld<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 3);
+  assert(this->getDest() == this->getSrc(0));
+  const Variable *Dest = this->getDest();
+  const Operand *Src1 = this->getSrc(1);
+  const Operand *Src2 = this->getSrc(2);
+  static const typename InstX86Base<
+      Machine>::Traits::Assembler::GPREmitterShiftD Emitter = {
+      &InstX86Base<Machine>::Traits::Assembler::shld,
+      &InstX86Base<Machine>::Traits::Assembler::shld};
+  emitIASGPRShiftDouble<Machine>(Func, Dest, Src1, Src2, Emitter);
+}
+
+template <class Machine>
+void InstX86BaseShld<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  this->dumpDest(Func);
+  Str << " = shld." << this->getDest()->getType() << " ";
+  this->dumpSources(Func);
+}
+
+template <class Machine>
+void InstX86BaseShrd<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  Variable *Dest = this->getDest();
+  assert(this->getSrcSize() == 3);
+  assert(Dest == this->getSrc(0));
+  Str << "\tshrd" << this->getWidthString(Dest->getType()) << "\t";
+  if (const auto ShiftReg = llvm::dyn_cast<Variable>(this->getSrc(2))) {
+    (void)ShiftReg;
+    assert(ShiftReg->getRegNum() ==
+           InstX86Base<Machine>::Traits::RegisterSet::Reg_ecx);
+    Str << "%cl";
+  } else {
+    this->getSrc(2)->emit(Func);
+  }
+  Str << ", ";
+  this->getSrc(1)->emit(Func);
+  Str << ", ";
+  Dest->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseShrd<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 3);
+  assert(this->getDest() == this->getSrc(0));
+  const Variable *Dest = this->getDest();
+  const Operand *Src1 = this->getSrc(1);
+  const Operand *Src2 = this->getSrc(2);
+  static const typename InstX86Base<
+      Machine>::Traits::Assembler::GPREmitterShiftD Emitter = {
+      &InstX86Base<Machine>::Traits::Assembler::shrd,
+      &InstX86Base<Machine>::Traits::Assembler::shrd};
+  emitIASGPRShiftDouble<Machine>(Func, Dest, Src1, Src2, Emitter);
+}
+
+template <class Machine>
+void InstX86BaseShrd<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  this->dumpDest(Func);
+  Str << " = shrd." << this->getDest()->getType() << " ";
+  this->dumpSources(Func);
+}
+
+template <class Machine>
+void InstX86BaseCmov<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  Variable *Dest = this->getDest();
+  Str << "\t";
+  assert(Condition != InstX86Base<Machine>::Traits::Cond::Br_None);
+  assert(this->getDest()->hasReg());
+  Str << "cmov"
+      << InstX86Base<Machine>::Traits::InstBrAttributes[Condition].DisplayString
+      << this->getWidthString(Dest->getType()) << "\t";
+  this->getSrc(1)->emit(Func);
+  Str << ", ";
+  Dest->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseCmov<Machine>::emitIAS(const Cfg *Func) const {
+  assert(Condition != InstX86Base<Machine>::Traits::Cond::Br_None);
+  assert(this->getDest()->hasReg());
+  assert(this->getSrcSize() == 2);
+  Operand *Src = this->getSrc(1);
+  Type SrcTy = Src->getType();
+  assert(SrcTy == IceType_i16 || SrcTy == IceType_i32);
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  if (const auto *SrcVar = llvm::dyn_cast<Variable>(Src)) {
+    if (SrcVar->hasReg()) {
+      Asm->cmov(SrcTy, Condition,
+                InstX86Base<Machine>::Traits::RegisterSet::getEncodedGPR(
+                    this->getDest()->getRegNum()),
+                InstX86Base<Machine>::Traits::RegisterSet::getEncodedGPR(
+                    SrcVar->getRegNum()));
+    } else {
+      Asm->cmov(
+          SrcTy, Condition,
+          InstX86Base<Machine>::Traits::RegisterSet::getEncodedGPR(
+              this->getDest()->getRegNum()),
+          static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+              Func->getTarget())
+              ->stackVarToAsmOperand(SrcVar));
+    }
+  } else if (const auto Mem = llvm::dyn_cast<
+                 typename InstX86Base<Machine>::Traits::X86OperandMem>(Src)) {
+    assert(Mem->getSegmentRegister() ==
+           InstX86Base<Machine>::Traits::X86OperandMem::DefaultSegment);
+    Asm->cmov(SrcTy, Condition,
+              InstX86Base<Machine>::Traits::RegisterSet::getEncodedGPR(
+                  this->getDest()->getRegNum()),
+              Mem->toAsmAddress(Asm));
+  } else {
+    llvm_unreachable("Unexpected operand type");
+  }
+}
+
+template <class Machine>
+void InstX86BaseCmov<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  Str << "cmov"
+      << InstX86Base<Machine>::Traits::InstBrAttributes[Condition].DisplayString
+      << ".";
+  Str << this->getDest()->getType() << " ";
+  this->dumpDest(Func);
+  Str << ", ";
+  this->dumpSources(Func);
+}
+
+template <class Machine>
+void InstX86BaseCmpps<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 2);
+  assert(Condition < InstX86Base<Machine>::Traits::Cond::Cmpps_Invalid);
+  Str << "\t";
+  Str << "cmp"
+      << InstX86Base<Machine>::Traits::InstCmppsAttributes[Condition].EmitString
+      << "ps"
+      << "\t";
+  this->getSrc(1)->emit(Func);
+  Str << ", ";
+  this->getDest()->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseCmpps<Machine>::emitIAS(const Cfg *Func) const {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  assert(this->getSrcSize() == 2);
+  assert(Condition < InstX86Base<Machine>::Traits::Cond::Cmpps_Invalid);
+  // Assuming there isn't any load folding for cmpps, and vector constants
+  // are not allowed in PNaCl.
+  assert(llvm::isa<Variable>(this->getSrc(1)));
+  const auto SrcVar = llvm::cast<Variable>(this->getSrc(1));
+  if (SrcVar->hasReg()) {
+    Asm->cmpps(InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm(
+                   this->getDest()->getRegNum()),
+               InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm(
+                   SrcVar->getRegNum()),
+               Condition);
+  } else {
+    typename InstX86Base<Machine>::Traits::Address SrcStackAddr =
+        static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+            Func->getTarget())
+            ->stackVarToAsmOperand(SrcVar);
+    Asm->cmpps(InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm(
+                   this->getDest()->getRegNum()),
+               SrcStackAddr, Condition);
+  }
+}
+
+template <class Machine>
+void InstX86BaseCmpps<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  assert(Condition < InstX86Base<Machine>::Traits::Cond::Cmpps_Invalid);
+  this->dumpDest(Func);
+  Str << " = cmp"
+      << InstX86Base<Machine>::Traits::InstCmppsAttributes[Condition].EmitString
+      << "ps"
+      << "\t";
+  this->dumpSources(Func);
+}
+
+template <class Machine>
+void InstX86BaseCmpxchg<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 3);
+  if (this->Locked) {
+    Str << "\tlock";
+  }
+  Str << "\tcmpxchg" << this->getWidthString(this->getSrc(0)->getType())
+      << "\t";
+  this->getSrc(2)->emit(Func);
+  Str << ", ";
+  this->getSrc(0)->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseCmpxchg<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 3);
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  Type Ty = this->getSrc(0)->getType();
+  const auto Mem =
+      llvm::cast<typename InstX86Base<Machine>::Traits::X86OperandMem>(
+          this->getSrc(0));
+  assert(Mem->getSegmentRegister() ==
+         InstX86Base<Machine>::Traits::X86OperandMem::DefaultSegment);
+  const typename InstX86Base<Machine>::Traits::Address Addr =
+      Mem->toAsmAddress(Asm);
+  const auto VarReg = llvm::cast<Variable>(this->getSrc(2));
+  assert(VarReg->hasReg());
+  const typename InstX86Base<Machine>::Traits::RegisterSet::GPRRegister Reg =
+      InstX86Base<Machine>::Traits::RegisterSet::getEncodedGPR(
+          VarReg->getRegNum());
+  Asm->cmpxchg(Ty, Addr, Reg, this->Locked);
+}
+
+template <class Machine>
+void InstX86BaseCmpxchg<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  if (this->Locked) {
+    Str << "lock ";
+  }
+  Str << "cmpxchg." << this->getSrc(0)->getType() << " ";
+  this->dumpSources(Func);
+}
+
+template <class Machine>
+void InstX86BaseCmpxchg8b<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 5);
+  if (this->Locked) {
+    Str << "\tlock";
+  }
+  Str << "\tcmpxchg8b\t";
+  this->getSrc(0)->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseCmpxchg8b<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 5);
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  const auto Mem =
+      llvm::cast<typename InstX86Base<Machine>::Traits::X86OperandMem>(
+          this->getSrc(0));
+  assert(Mem->getSegmentRegister() ==
+         InstX86Base<Machine>::Traits::X86OperandMem::DefaultSegment);
+  const typename InstX86Base<Machine>::Traits::Address Addr =
+      Mem->toAsmAddress(Asm);
+  Asm->cmpxchg8b(Addr, this->Locked);
+}
+
+template <class Machine>
+void InstX86BaseCmpxchg8b<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  if (this->Locked) {
+    Str << "lock ";
+  }
+  Str << "cmpxchg8b ";
+  this->dumpSources(Func);
+}
+
+template <class Machine>
+void InstX86BaseCvt<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 1);
+  Str << "\tcvt";
+  if (isTruncating())
+    Str << "t";
+  Str << InstX86Base<Machine>::Traits::TypeAttributes[this->getSrc(0)
+                                                          ->getType()]
+             .CvtString << "2"
+      << InstX86Base<
+             Machine>::Traits::TypeAttributes[this->getDest()->getType()]
+             .CvtString << "\t";
+  this->getSrc(0)->emit(Func);
+  Str << ", ";
+  this->getDest()->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseCvt<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 1);
+  const Variable *Dest = this->getDest();
+  const Operand *Src = this->getSrc(0);
+  Type DestTy = Dest->getType();
+  Type SrcTy = Src->getType();
+  switch (Variant) {
+  case Si2ss: {
+    assert(isScalarIntegerType(SrcTy));
+    assert(typeWidthInBytes(SrcTy) <= 4);
+    assert(isScalarFloatingType(DestTy));
+    static const typename InstX86Base<Machine>::Traits::Assembler::
+        template CastEmitterRegOp<
+            typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister,
+            typename InstX86Base<Machine>::Traits::RegisterSet::GPRRegister>
+            Emitter = {&InstX86Base<Machine>::Traits::Assembler::cvtsi2ss,
+                       &InstX86Base<Machine>::Traits::Assembler::cvtsi2ss};
+    emitIASCastRegOp<
+        Machine,
+        typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister,
+        typename InstX86Base<Machine>::Traits::RegisterSet::GPRRegister,
+        InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm,
+        InstX86Base<Machine>::Traits::RegisterSet::getEncodedGPR>(
+        Func, DestTy, Dest, Src, Emitter);
+    return;
+  }
+  case Tss2si: {
+    assert(isScalarFloatingType(SrcTy));
+    assert(isScalarIntegerType(DestTy));
+    assert(typeWidthInBytes(DestTy) <= 4);
+    static const typename InstX86Base<Machine>::Traits::Assembler::
+        template CastEmitterRegOp<
+            typename InstX86Base<Machine>::Traits::RegisterSet::GPRRegister,
+            typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister>
+            Emitter = {&InstX86Base<Machine>::Traits::Assembler::cvttss2si,
+                       &InstX86Base<Machine>::Traits::Assembler::cvttss2si};
+    emitIASCastRegOp<
+        Machine,
+        typename InstX86Base<Machine>::Traits::RegisterSet::GPRRegister,
+        typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister,
+        InstX86Base<Machine>::Traits::RegisterSet::getEncodedGPR,
+        InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm>(
+        Func, SrcTy, Dest, Src, Emitter);
+    return;
+  }
+  case Float2float: {
+    assert(isScalarFloatingType(SrcTy));
+    assert(isScalarFloatingType(DestTy));
+    assert(DestTy != SrcTy);
+    static const typename InstX86Base<
+        Machine>::Traits::Assembler::XmmEmitterRegOp Emitter = {
+        &InstX86Base<Machine>::Traits::Assembler::cvtfloat2float,
+        &InstX86Base<Machine>::Traits::Assembler::cvtfloat2float};
+    emitIASRegOpTyXMM<Machine>(Func, SrcTy, Dest, Src, Emitter);
+    return;
+  }
+  case Dq2ps: {
+    assert(isVectorIntegerType(SrcTy));
+    assert(isVectorFloatingType(DestTy));
+    static const typename InstX86Base<
+        Machine>::Traits::Assembler::XmmEmitterRegOp Emitter = {
+        &InstX86Base<Machine>::Traits::Assembler::cvtdq2ps,
+        &InstX86Base<Machine>::Traits::Assembler::cvtdq2ps};
+    emitIASRegOpTyXMM<Machine>(Func, DestTy, Dest, Src, Emitter);
+    return;
+  }
+  case Tps2dq: {
+    assert(isVectorFloatingType(SrcTy));
+    assert(isVectorIntegerType(DestTy));
+    static const typename InstX86Base<
+        Machine>::Traits::Assembler::XmmEmitterRegOp Emitter = {
+        &InstX86Base<Machine>::Traits::Assembler::cvttps2dq,
+        &InstX86Base<Machine>::Traits::Assembler::cvttps2dq};
+    emitIASRegOpTyXMM<Machine>(Func, DestTy, Dest, Src, Emitter);
+    return;
+  }
+  }
+}
+
+template <class Machine>
+void InstX86BaseCvt<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  this->dumpDest(Func);
+  Str << " = cvt";
+  if (isTruncating())
+    Str << "t";
+  Str << InstX86Base<Machine>::Traits::TypeAttributes[this->getSrc(0)
+                                                          ->getType()]
+             .CvtString << "2"
+      << InstX86Base<
+             Machine>::Traits::TypeAttributes[this->getDest()->getType()]
+             .CvtString << " ";
+  this->dumpSources(Func);
+}
+
+template <class Machine>
+void InstX86BaseIcmp<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 2);
+  Str << "\tcmp" << this->getWidthString(this->getSrc(0)->getType()) << "\t";
+  this->getSrc(1)->emit(Func);
+  Str << ", ";
+  this->getSrc(0)->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseIcmp<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 2);
+  const Operand *Src0 = this->getSrc(0);
+  const Operand *Src1 = this->getSrc(1);
+  Type Ty = Src0->getType();
+  static const typename InstX86Base<Machine>::Traits::Assembler::GPREmitterRegOp
+      RegEmitter = {&InstX86Base<Machine>::Traits::Assembler::cmp,
+                    &InstX86Base<Machine>::Traits::Assembler::cmp,
+                    &InstX86Base<Machine>::Traits::Assembler::cmp};
+  static const typename InstX86Base<
+      Machine>::Traits::Assembler::GPREmitterAddrOp AddrEmitter = {
+      &InstX86Base<Machine>::Traits::Assembler::cmp,
+      &InstX86Base<Machine>::Traits::Assembler::cmp};
+  if (const auto SrcVar0 = llvm::dyn_cast<Variable>(Src0)) {
+    if (SrcVar0->hasReg()) {
+      emitIASRegOpTyGPR<Machine>(Func, Ty, SrcVar0, Src1, RegEmitter);
+      return;
+    }
+  }
+  emitIASAsAddrOpTyGPR<Machine>(Func, Ty, Src0, Src1, AddrEmitter);
+}
+
+template <class Machine>
+void InstX86BaseIcmp<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  Str << "cmp." << this->getSrc(0)->getType() << " ";
+  this->dumpSources(Func);
+}
+
+template <class Machine>
+void InstX86BaseUcomiss<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 2);
+  Str << "\tucomi"
+      << InstX86Base<Machine>::Traits::TypeAttributes[this->getSrc(0)
+                                                          ->getType()]
+             .SdSsString << "\t";
+  this->getSrc(1)->emit(Func);
+  Str << ", ";
+  this->getSrc(0)->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseUcomiss<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 2);
+  // Currently src0 is always a variable by convention, to avoid having
+  // two memory operands.
+  assert(llvm::isa<Variable>(this->getSrc(0)));
+  const auto Src0Var = llvm::cast<Variable>(this->getSrc(0));
+  Type Ty = Src0Var->getType();
+  static const typename InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp
+      Emitter = {&InstX86Base<Machine>::Traits::Assembler::ucomiss,
+                 &InstX86Base<Machine>::Traits::Assembler::ucomiss};
+  emitIASRegOpTyXMM<Machine>(Func, Ty, Src0Var, this->getSrc(1), Emitter);
+}
+
+template <class Machine>
+void InstX86BaseUcomiss<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  Str << "ucomiss." << this->getSrc(0)->getType() << " ";
+  this->dumpSources(Func);
+}
+
+template <class Machine>
+void InstX86BaseUD2<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 0);
+  Str << "\tud2";
+}
+
+template <class Machine>
+void InstX86BaseUD2<Machine>::emitIAS(const Cfg *Func) const {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  Asm->ud2();
+}
+
+template <class Machine>
+void InstX86BaseUD2<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  Str << "ud2";
+}
+
+template <class Machine>
+void InstX86BaseTest<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 2);
+  Str << "\ttest" << this->getWidthString(this->getSrc(0)->getType()) << "\t";
+  this->getSrc(1)->emit(Func);
+  Str << ", ";
+  this->getSrc(0)->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseTest<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 2);
+  const Operand *Src0 = this->getSrc(0);
+  const Operand *Src1 = this->getSrc(1);
+  Type Ty = Src0->getType();
+  // The Reg/Addr form of test is not encodeable.
+  static const typename InstX86Base<Machine>::Traits::Assembler::GPREmitterRegOp
+      RegEmitter = {&InstX86Base<Machine>::Traits::Assembler::test, nullptr,
+                    &InstX86Base<Machine>::Traits::Assembler::test};
+  static const typename InstX86Base<
+      Machine>::Traits::Assembler::GPREmitterAddrOp AddrEmitter = {
+      &InstX86Base<Machine>::Traits::Assembler::test,
+      &InstX86Base<Machine>::Traits::Assembler::test};
+  if (const auto SrcVar0 = llvm::dyn_cast<Variable>(Src0)) {
+    if (SrcVar0->hasReg()) {
+      emitIASRegOpTyGPR<Machine>(Func, Ty, SrcVar0, Src1, RegEmitter);
+      return;
+    }
+  }
+  llvm_unreachable("Nothing actually generates this so it's untested");
+  emitIASAsAddrOpTyGPR<Machine>(Func, Ty, Src0, Src1, AddrEmitter);
+}
+
+template <class Machine>
+void InstX86BaseTest<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  Str << "test." << this->getSrc(0)->getType() << " ";
+  this->dumpSources(Func);
+}
+
+template <class Machine>
+void InstX86BaseMfence<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 0);
+  Str << "\tmfence";
+}
+
+template <class Machine>
+void InstX86BaseMfence<Machine>::emitIAS(const Cfg *Func) const {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  Asm->mfence();
+}
+
+template <class Machine>
+void InstX86BaseMfence<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  Str << "mfence";
+}
+
+template <class Machine>
+void InstX86BaseStore<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 2);
+  Type Ty = this->getSrc(0)->getType();
+  Str << "\tmov" << this->getWidthString(Ty)
+      << InstX86Base<Machine>::Traits::TypeAttributes[Ty].SdSsString << "\t";
+  this->getSrc(0)->emit(Func);
+  Str << ", ";
+  this->getSrc(1)->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseStore<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 2);
+  const Operand *Dest = this->getSrc(1);
+  const Operand *Src = this->getSrc(0);
+  Type DestTy = Dest->getType();
+  if (isScalarFloatingType(DestTy)) {
+    // Src must be a register, since Dest is a Mem operand of some kind.
+    const auto SrcVar = llvm::cast<Variable>(Src);
+    assert(SrcVar->hasReg());
+    typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister SrcReg =
+        InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm(
+            SrcVar->getRegNum());
+    typename InstX86Base<Machine>::Traits::Assembler *Asm =
+        Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+    if (const auto DestVar = llvm::dyn_cast<Variable>(Dest)) {
+      assert(!DestVar->hasReg());
+      typename InstX86Base<Machine>::Traits::Address StackAddr(
+          static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+              Func->getTarget())
+              ->stackVarToAsmOperand(DestVar));
+      Asm->movss(DestTy, StackAddr, SrcReg);
+    } else {
+      const auto DestMem =
+          llvm::cast<typename InstX86Base<Machine>::Traits::X86OperandMem>(
+              Dest);
+      assert(DestMem->getSegmentRegister() ==
+             InstX86Base<Machine>::Traits::X86OperandMem::DefaultSegment);
+      Asm->movss(DestTy, DestMem->toAsmAddress(Asm), SrcReg);
+    }
+    return;
+  } else {
+    assert(isScalarIntegerType(DestTy));
+    static const typename InstX86Base<
+        Machine>::Traits::Assembler::GPREmitterAddrOp GPRAddrEmitter = {
+        &InstX86Base<Machine>::Traits::Assembler::mov,
+        &InstX86Base<Machine>::Traits::Assembler::mov};
+    emitIASAsAddrOpTyGPR<Machine>(Func, DestTy, Dest, Src, GPRAddrEmitter);
+  }
+}
+
+template <class Machine>
+void InstX86BaseStore<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  Str << "mov." << this->getSrc(0)->getType() << " ";
+  this->getSrc(1)->dump(Func);
+  Str << ", ";
+  this->getSrc(0)->dump(Func);
+}
+
+template <class Machine>
+void InstX86BaseStoreP<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 2);
+  Str << "\tmovups\t";
+  this->getSrc(0)->emit(Func);
+  Str << ", ";
+  this->getSrc(1)->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseStoreP<Machine>::emitIAS(const Cfg *Func) const {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  assert(this->getSrcSize() == 2);
+  const auto SrcVar = llvm::cast<Variable>(this->getSrc(0));
+  const auto DestMem =
+      llvm::cast<typename InstX86Base<Machine>::Traits::X86OperandMem>(
+          this->getSrc(1));
+  assert(DestMem->getSegmentRegister() ==
+         InstX86Base<Machine>::Traits::X86OperandMem::DefaultSegment);
+  assert(SrcVar->hasReg());
+  Asm->movups(DestMem->toAsmAddress(Asm),
+              InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm(
+                  SrcVar->getRegNum()));
+}
+
+template <class Machine>
+void InstX86BaseStoreP<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  Str << "storep." << this->getSrc(0)->getType() << " ";
+  this->getSrc(1)->dump(Func);
+  Str << ", ";
+  this->getSrc(0)->dump(Func);
+}
+
+template <class Machine>
+void InstX86BaseStoreQ<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 2);
+  assert(this->getSrc(1)->getType() == IceType_i64 ||
+         this->getSrc(1)->getType() == IceType_f64);
+  Str << "\tmovq\t";
+  this->getSrc(0)->emit(Func);
+  Str << ", ";
+  this->getSrc(1)->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseStoreQ<Machine>::emitIAS(const Cfg *Func) const {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  assert(this->getSrcSize() == 2);
+  const auto SrcVar = llvm::cast<Variable>(this->getSrc(0));
+  const auto DestMem =
+      llvm::cast<typename InstX86Base<Machine>::Traits::X86OperandMem>(
+          this->getSrc(1));
+  assert(DestMem->getSegmentRegister() ==
+         InstX86Base<Machine>::Traits::X86OperandMem::DefaultSegment);
+  assert(SrcVar->hasReg());
+  Asm->movq(DestMem->toAsmAddress(Asm),
+            InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm(
+                SrcVar->getRegNum()));
+}
+
+template <class Machine>
+void InstX86BaseStoreQ<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  Str << "storeq." << this->getSrc(0)->getType() << " ";
+  this->getSrc(1)->dump(Func);
+  Str << ", ";
+  this->getSrc(0)->dump(Func);
+}
+
+template <class Machine>
+void InstX86BaseLea<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 1);
+  assert(this->getDest()->hasReg());
+  Str << "\tleal\t";
+  Operand *Src0 = this->getSrc(0);
+  if (const auto Src0Var = llvm::dyn_cast<Variable>(Src0)) {
+    Type Ty = Src0Var->getType();
+    // lea on x86-32 doesn't accept mem128 operands, so cast VSrc0 to an
+    // acceptable type.
+    Src0Var->asType(isVectorType(Ty) ? IceType_i32 : Ty)->emit(Func);
+  } else {
+    Src0->emit(Func);
+  }
+  Str << ", ";
+  this->getDest()->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseMov<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 1);
+  Operand *Src = this->getSrc(0);
+  Type SrcTy = Src->getType();
+  Type DestTy = this->getDest()->getType();
+  Str << "\tmov"
+      << (!isScalarFloatingType(DestTy)
+              ? this->getWidthString(SrcTy)
+              : InstX86Base<Machine>::Traits::TypeAttributes[DestTy].SdSsString)
+      << "\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(DestTy) ==
+         Func->getTarget()->typeWidthInBytesOnStack(SrcTy));
+  Src->emit(Func);
+  Str << ", ";
+  this->getDest()->asType(SrcTy)->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseMov<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 1);
+  const Variable *Dest = this->getDest();
+  const Operand *Src = this->getSrc(0);
+  Type DestTy = Dest->getType();
+  Type SrcTy = Src->getType();
+  // Mov can be used for GPRs or XMM registers. Also, the type does not
+  // necessarily match (Mov can be used for bitcasts). However, when
+  // the type does not match, one of the operands must be a register.
+  // Thus, the strategy is to find out if Src or Dest are a register,
+  // then use that register's type to decide on which emitter set to use.
+  // The emitter set will include reg-reg movs, but that case should
+  // be unused when the types don't match.
+  static const typename InstX86Base<Machine>::Traits::Assembler::XmmEmitterRegOp
+      XmmRegEmitter = {&InstX86Base<Machine>::Traits::Assembler::movss,
+                       &InstX86Base<Machine>::Traits::Assembler::movss};
+  static const typename InstX86Base<Machine>::Traits::Assembler::GPREmitterRegOp
+      GPRRegEmitter = {&InstX86Base<Machine>::Traits::Assembler::mov,
+                       &InstX86Base<Machine>::Traits::Assembler::mov,
+                       &InstX86Base<Machine>::Traits::Assembler::mov};
+  static const typename InstX86Base<
+      Machine>::Traits::Assembler::GPREmitterAddrOp GPRAddrEmitter = {
+      &InstX86Base<Machine>::Traits::Assembler::mov,
+      &InstX86Base<Machine>::Traits::Assembler::mov};
+  // 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(this->getDest()->getType()) ==
+      Func->getTarget()->typeWidthInBytesOnStack(Src->getType()));
+  if (Dest->hasReg()) {
+    if (isScalarFloatingType(DestTy)) {
+      emitIASRegOpTyXMM<Machine>(Func, DestTy, Dest, Src, XmmRegEmitter);
+      return;
+    } else {
+      assert(isScalarIntegerType(DestTy));
+      // Widen DestTy for truncation (see above note). We should only do this
+      // when both Src and Dest are integer types.
+      if (isScalarIntegerType(SrcTy)) {
+        DestTy = SrcTy;
+      }
+      emitIASRegOpTyGPR<Machine>(Func, DestTy, Dest, Src, GPRRegEmitter);
+      return;
+    }
+  } else {
+    // Dest must be Stack and Src *could* be a register. Use Src's type
+    // to decide on the emitters.
+    typename InstX86Base<Machine>::Traits::Address StackAddr(
+        static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+            Func->getTarget())
+            ->stackVarToAsmOperand(Dest));
+    if (isScalarFloatingType(SrcTy)) {
+      // Src must be a register.
+      const auto SrcVar = llvm::cast<Variable>(Src);
+      assert(SrcVar->hasReg());
+      typename InstX86Base<Machine>::Traits::Assembler *Asm =
+          Func->getAssembler<
+              typename InstX86Base<Machine>::Traits::Assembler>();
+      Asm->movss(SrcTy, StackAddr,
+                 InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm(
+                     SrcVar->getRegNum()));
+      return;
+    } else {
+      // Src can be a register or immediate.
+      assert(isScalarIntegerType(SrcTy));
+      emitIASAddrOpTyGPR<Machine>(Func, SrcTy, StackAddr, Src, GPRAddrEmitter);
+      return;
+    }
+    return;
+  }
+}
+
+template <class Machine>
+void InstX86BaseMovd<Machine>::emitIAS(const Cfg *Func) const {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  assert(this->getSrcSize() == 1);
+  const Variable *Dest = this->getDest();
+  const auto SrcVar = llvm::cast<Variable>(this->getSrc(0));
+  // For insert/extract element (one of Src/Dest is an Xmm vector and
+  // the other is an int type).
+  if (SrcVar->getType() == IceType_i32) {
+    assert(isVectorType(Dest->getType()));
+    assert(Dest->hasReg());
+    typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister DestReg =
+        InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm(
+            Dest->getRegNum());
+    if (SrcVar->hasReg()) {
+      Asm->movd(DestReg,
+                InstX86Base<Machine>::Traits::RegisterSet::getEncodedGPR(
+                    SrcVar->getRegNum()));
+    } else {
+      typename InstX86Base<Machine>::Traits::Address StackAddr(
+          static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+              Func->getTarget())
+              ->stackVarToAsmOperand(SrcVar));
+      Asm->movd(DestReg, StackAddr);
+    }
+  } else {
+    assert(isVectorType(SrcVar->getType()));
+    assert(SrcVar->hasReg());
+    assert(Dest->getType() == IceType_i32);
+    typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister SrcReg =
+        InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm(
+            SrcVar->getRegNum());
+    if (Dest->hasReg()) {
+      Asm->movd(InstX86Base<Machine>::Traits::RegisterSet::getEncodedGPR(
+                    Dest->getRegNum()),
+                SrcReg);
+    } else {
+      typename InstX86Base<Machine>::Traits::Address StackAddr(
+          static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+              Func->getTarget())
+              ->stackVarToAsmOperand(Dest));
+      Asm->movd(StackAddr, SrcReg);
+    }
+  }
+}
+
+template <class Machine>
+void InstX86BaseMovp<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  // TODO(wala,stichnot): movups works with all vector operands, but
+  // there exist other instructions (movaps, movdqa, movdqu) that may
+  // perform better, depending on the data type and alignment of the
+  // operands.
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 1);
+  Str << "\tmovups\t";
+  this->getSrc(0)->emit(Func);
+  Str << ", ";
+  this->getDest()->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseMovp<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 1);
+  assert(isVectorType(this->getDest()->getType()));
+  const Variable *Dest = this->getDest();
+  const Operand *Src = this->getSrc(0);
+  static const typename InstX86Base<
+      Machine>::Traits::Assembler::XmmEmitterMovOps Emitter = {
+      &InstX86Base<Machine>::Traits::Assembler::movups,
+      &InstX86Base<Machine>::Traits::Assembler::movups,
+      &InstX86Base<Machine>::Traits::Assembler::movups};
+  emitIASMovlikeXMM<Machine>(Func, Dest, Src, Emitter);
+}
+
+template <class Machine>
+void InstX86BaseMovq<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 1);
+  assert(this->getDest()->getType() == IceType_i64 ||
+         this->getDest()->getType() == IceType_f64);
+  Str << "\tmovq\t";
+  this->getSrc(0)->emit(Func);
+  Str << ", ";
+  this->getDest()->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseMovq<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 1);
+  assert(this->getDest()->getType() == IceType_i64 ||
+         this->getDest()->getType() == IceType_f64);
+  const Variable *Dest = this->getDest();
+  const Operand *Src = this->getSrc(0);
+  static const typename InstX86Base<
+      Machine>::Traits::Assembler::XmmEmitterMovOps Emitter = {
+      &InstX86Base<Machine>::Traits::Assembler::movq,
+      &InstX86Base<Machine>::Traits::Assembler::movq,
+      &InstX86Base<Machine>::Traits::Assembler::movq};
+  emitIASMovlikeXMM<Machine>(Func, Dest, Src, Emitter);
+}
+
+template <class Machine>
+void InstX86BaseMovssRegs<Machine>::emitIAS(const Cfg *Func) const {
+  // This is Binop variant is only intended to be used for reg-reg moves
+  // where part of the Dest register is untouched.
+  assert(this->getSrcSize() == 2);
+  const Variable *Dest = this->getDest();
+  assert(Dest == this->getSrc(0));
+  const auto SrcVar = llvm::cast<Variable>(this->getSrc(1));
+  assert(Dest->hasReg() && SrcVar->hasReg());
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  Asm->movss(IceType_f32,
+             InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm(
+                 Dest->getRegNum()),
+             InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm(
+                 SrcVar->getRegNum()));
+}
+
+template <class Machine>
+void InstX86BaseMovsx<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 1);
+  const Variable *Dest = this->getDest();
+  const Operand *Src = this->getSrc(0);
+  // Dest must be a > 8-bit register, but Src can be 8-bit. In practice
+  // we just use the full register for Dest to avoid having an
+  // OperandSizeOverride prefix. It also allows us to only dispatch on SrcTy.
+  Type SrcTy = Src->getType();
+  assert(typeWidthInBytes(Dest->getType()) > 1);
+  assert(typeWidthInBytes(Dest->getType()) > typeWidthInBytes(SrcTy));
+  emitIASRegOpTyGPR<Machine, false, true>(Func, SrcTy, Dest, Src,
+                                          this->Emitter);
+}
+
+template <class Machine>
+void InstX86BaseMovzx<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 1);
+  const Variable *Dest = this->getDest();
+  const Operand *Src = this->getSrc(0);
+  Type SrcTy = Src->getType();
+  assert(typeWidthInBytes(Dest->getType()) > 1);
+  assert(typeWidthInBytes(Dest->getType()) > typeWidthInBytes(SrcTy));
+  emitIASRegOpTyGPR<Machine, false, true>(Func, SrcTy, Dest, Src,
+                                          this->Emitter);
+}
+
+template <class Machine>
+void InstX86BaseNop<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  // TODO: Emit the right code for each variant.
+  Str << "\tnop\t# variant = " << Variant;
+}
+
+template <class Machine>
+void InstX86BaseNop<Machine>::emitIAS(const Cfg *Func) const {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  // TODO: Emit the right code for the variant.
+  Asm->nop();
+}
+
+template <class Machine>
+void InstX86BaseNop<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  Str << "nop (variant = " << Variant << ")";
+}
+
+template <class Machine>
+void InstX86BaseFld<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 1);
+  Type Ty = this->getSrc(0)->getType();
+  SizeT Width = typeWidthInBytes(Ty);
+  const auto Var = llvm::dyn_cast<Variable>(this->getSrc(0));
+  if (Var && Var->hasReg()) {
+    // This is a physical xmm register, so we need to spill it to a
+    // temporary stack slot.
+    Str << "\tsubl\t$" << Width << ", %esp"
+        << "\n";
+    Str << "\tmov"
+        << InstX86Base<Machine>::Traits::TypeAttributes[Ty].SdSsString << "\t";
+    Var->emit(Func);
+    Str << ", (%esp)\n";
+    Str << "\tfld" << this->getFldString(Ty) << "\t"
+        << "(%esp)\n";
+    Str << "\taddl\t$" << Width << ", %esp";
+    return;
+  }
+  Str << "\tfld" << this->getFldString(Ty) << "\t";
+  this->getSrc(0)->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseFld<Machine>::emitIAS(const Cfg *Func) const {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  assert(this->getSrcSize() == 1);
+  const Operand *Src = this->getSrc(0);
+  Type Ty = Src->getType();
+  if (const auto Var = llvm::dyn_cast<Variable>(Src)) {
+    if (Var->hasReg()) {
+      // This is a physical xmm register, so we need to spill it to a
+      // temporary stack slot.
+      Immediate Width(typeWidthInBytes(Ty));
+      Asm->sub(IceType_i32,
+               InstX86Base<Machine>::Traits::RegisterSet::Encoded_Reg_esp,
+               Width);
+      typename InstX86Base<Machine>::Traits::Address StackSlot =
+          typename InstX86Base<Machine>::Traits::Address(
+              InstX86Base<Machine>::Traits::RegisterSet::Encoded_Reg_esp, 0);
+      Asm->movss(Ty, StackSlot,
+                 InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm(
+                     Var->getRegNum()));
+      Asm->fld(Ty, StackSlot);
+      Asm->add(IceType_i32,
+               InstX86Base<Machine>::Traits::RegisterSet::Encoded_Reg_esp,
+               Width);
+    } else {
+      typename InstX86Base<Machine>::Traits::Address StackAddr(
+          static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+              Func->getTarget())
+              ->stackVarToAsmOperand(Var));
+      Asm->fld(Ty, StackAddr);
+    }
+  } else if (const auto Mem = llvm::dyn_cast<
+                 typename InstX86Base<Machine>::Traits::X86OperandMem>(Src)) {
+    assert(Mem->getSegmentRegister() ==
+           InstX86Base<Machine>::Traits::X86OperandMem::DefaultSegment);
+    Asm->fld(Ty, Mem->toAsmAddress(Asm));
+  } else if (const auto Imm = llvm::dyn_cast<Constant>(Src)) {
+    Asm->fld(Ty, InstX86Base<Machine>::Traits::Address::ofConstPool(Asm, Imm));
+  } else {
+    llvm_unreachable("Unexpected operand type");
+  }
+}
+
+template <class Machine>
+void InstX86BaseFld<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  Str << "fld." << this->getSrc(0)->getType() << " ";
+  this->dumpSources(Func);
+}
+
+template <class Machine>
+void InstX86BaseFstp<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 0);
+  // TODO(jvoung,stichnot): Utilize this by setting Dest to nullptr to
+  // "partially" delete the fstp if the Dest is unused.
+  // Even if Dest is unused, the fstp should be kept for the SideEffects
+  // of popping the stack.
+  if (!this->getDest()) {
+    Str << "\tfstp\tst(0)";
+    return;
+  }
+  Type Ty = this->getDest()->getType();
+  size_t Width = typeWidthInBytes(Ty);
+  if (!this->getDest()->hasReg()) {
+    Str << "\tfstp" << this->getFldString(Ty) << "\t";
+    this->getDest()->emit(Func);
+    return;
+  }
+  // Dest is a physical (xmm) register, so st(0) needs to go through
+  // memory.  Hack this by creating a temporary stack slot, spilling
+  // st(0) there, loading it into the xmm register, and deallocating
+  // the stack slot.
+  Str << "\tsubl\t$" << Width << ", %esp\n";
+  Str << "\tfstp" << this->getFldString(Ty) << "\t"
+      << "(%esp)\n";
+  Str << "\tmov" << InstX86Base<Machine>::Traits::TypeAttributes[Ty].SdSsString
+      << "\t"
+      << "(%esp), ";
+  this->getDest()->emit(Func);
+  Str << "\n";
+  Str << "\taddl\t$" << Width << ", %esp";
+}
+
+template <class Machine>
+void InstX86BaseFstp<Machine>::emitIAS(const Cfg *Func) const {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  assert(this->getSrcSize() == 0);
+  const Variable *Dest = this->getDest();
+  // TODO(jvoung,stichnot): Utilize this by setting Dest to nullptr to
+  // "partially" delete the fstp if the Dest is unused.
+  // Even if Dest is unused, the fstp should be kept for the SideEffects
+  // of popping the stack.
+  if (!Dest) {
+    Asm->fstp(InstX86Base<Machine>::Traits::RegisterSet::getEncodedSTReg(0));
+    return;
+  }
+  Type Ty = Dest->getType();
+  if (!Dest->hasReg()) {
+    typename InstX86Base<Machine>::Traits::Address StackAddr(
+        static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+            Func->getTarget())
+            ->stackVarToAsmOperand(Dest));
+    Asm->fstp(Ty, StackAddr);
+  } else {
+    // Dest is a physical (xmm) register, so st(0) needs to go through
+    // memory.  Hack this by creating a temporary stack slot, spilling
+    // st(0) there, loading it into the xmm register, and deallocating
+    // the stack slot.
+    Immediate Width(typeWidthInBytes(Ty));
+    Asm->sub(IceType_i32,
+             InstX86Base<Machine>::Traits::RegisterSet::Encoded_Reg_esp, Width);
+    typename InstX86Base<Machine>::Traits::Address StackSlot =
+        typename InstX86Base<Machine>::Traits::Address(
+            InstX86Base<Machine>::Traits::RegisterSet::Encoded_Reg_esp, 0);
+    Asm->fstp(Ty, StackSlot);
+    Asm->movss(Ty, InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm(
+                       Dest->getRegNum()),
+               StackSlot);
+    Asm->add(IceType_i32,
+             InstX86Base<Machine>::Traits::RegisterSet::Encoded_Reg_esp, Width);
+  }
+}
+
+template <class Machine>
+void InstX86BaseFstp<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  this->dumpDest(Func);
+  Str << " = fstp." << this->getDest()->getType() << ", st(0)";
+}
+
+template <class Machine>
+void InstX86BasePcmpeq<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  char buf[30];
+  snprintf(
+      buf, llvm::array_lengthof(buf), "pcmpeq%s",
+      InstX86Base<Machine>::Traits::TypeAttributes[this->getDest()->getType()]
+          .PackString);
+  this->emitTwoAddress(buf, this, Func);
+}
+
+template <class Machine>
+void InstX86BasePcmpgt<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  char buf[30];
+  snprintf(
+      buf, llvm::array_lengthof(buf), "pcmpgt%s",
+      InstX86Base<Machine>::Traits::TypeAttributes[this->getDest()->getType()]
+          .PackString);
+  this->emitTwoAddress(buf, this, Func);
+}
+
+template <class Machine>
+void InstX86BasePextr<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 2);
+  // pextrb and pextrd are SSE4.1 instructions.
+  assert(this->getSrc(0)->getType() == IceType_v8i16 ||
+         this->getSrc(0)->getType() == IceType_v8i1 ||
+         static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+             Func->getTarget())
+                 ->getInstructionSet() >= InstX86Base<Machine>::Traits::SSE4_1);
+  Str << "\t" << this->Opcode
+      << InstX86Base<Machine>::Traits::TypeAttributes[this->getSrc(0)
+                                                          ->getType()]
+             .PackString << "\t";
+  this->getSrc(1)->emit(Func);
+  Str << ", ";
+  this->getSrc(0)->emit(Func);
+  Str << ", ";
+  Variable *Dest = this->getDest();
+  // pextrw must take a register dest. There is an SSE4.1 version that takes
+  // a memory dest, but we aren't using it. For uniformity, just restrict
+  // them all to have a register dest for now.
+  assert(Dest->hasReg());
+  Dest->asType(IceType_i32)->emit(Func);
+}
+
+template <class Machine>
+void InstX86BasePextr<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 2);
+  // pextrb and pextrd are SSE4.1 instructions.
+  const Variable *Dest = this->getDest();
+  Type DispatchTy = Dest->getType();
+  assert(DispatchTy == IceType_i16 ||
+         static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+             Func->getTarget())
+                 ->getInstructionSet() >= InstX86Base<Machine>::Traits::SSE4_1);
+  // pextrw must take a register dest. There is an SSE4.1 version that takes
+  // a memory dest, but we aren't using it. For uniformity, just restrict
+  // them all to have a register dest for now.
+  assert(Dest->hasReg());
+  // pextrw's Src(0) must be a register (both SSE4.1 and SSE2).
+  assert(llvm::cast<Variable>(this->getSrc(0))->hasReg());
+  static const typename InstX86Base<Machine>::Traits::Assembler::
+      template ThreeOpImmEmitter<
+          typename InstX86Base<Machine>::Traits::RegisterSet::GPRRegister,
+          typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister>
+          Emitter = {&InstX86Base<Machine>::Traits::Assembler::pextr, nullptr};
+  emitIASThreeOpImmOps<
+      Machine, typename InstX86Base<Machine>::Traits::RegisterSet::GPRRegister,
+      typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister,
+      InstX86Base<Machine>::Traits::RegisterSet::getEncodedGPR,
+      InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm>(
+      Func, DispatchTy, Dest, this->getSrc(0), this->getSrc(1), Emitter);
+}
+
+template <class Machine>
+void InstX86BasePinsr<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 3);
+  // pinsrb and pinsrd are SSE4.1 instructions.
+  assert(this->getDest()->getType() == IceType_v8i16 ||
+         this->getDest()->getType() == IceType_v8i1 ||
+         static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+             Func->getTarget())
+                 ->getInstructionSet() >= InstX86Base<Machine>::Traits::SSE4_1);
+  Str << "\t" << this->Opcode
+      << InstX86Base<
+             Machine>::Traits::TypeAttributes[this->getDest()->getType()]
+             .PackString << "\t";
+  this->getSrc(2)->emit(Func);
+  Str << ", ";
+  Operand *Src1 = this->getSrc(1);
+  if (const auto Src1Var = llvm::dyn_cast<Variable>(Src1)) {
+    // If src1 is a register, it should always be r32.
+    if (Src1Var->hasReg()) {
+      Src1Var->asType(IceType_i32)->emit(Func);
+    } else {
+      Src1Var->emit(Func);
+    }
+  } else {
+    Src1->emit(Func);
+  }
+  Str << ", ";
+  this->getDest()->emit(Func);
+}
+
+template <class Machine>
+void InstX86BasePinsr<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 3);
+  assert(this->getDest() == this->getSrc(0));
+  // pinsrb and pinsrd are SSE4.1 instructions.
+  const Operand *Src0 = this->getSrc(1);
+  Type DispatchTy = Src0->getType();
+  assert(DispatchTy == IceType_i16 ||
+         static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+             Func->getTarget())
+                 ->getInstructionSet() >= InstX86Base<Machine>::Traits::SSE4_1);
+  // If src1 is a register, it should always be r32 (this should fall out
+  // from the encodings for ByteRegs overlapping the encodings for r32),
+  // but we have to trust the regalloc to not choose "ah", where it
+  // doesn't overlap.
+  static const typename InstX86Base<Machine>::Traits::Assembler::
+      template ThreeOpImmEmitter<
+          typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister,
+          typename InstX86Base<Machine>::Traits::RegisterSet::GPRRegister>
+          Emitter = {&InstX86Base<Machine>::Traits::Assembler::pinsr,
+                     &InstX86Base<Machine>::Traits::Assembler::pinsr};
+  emitIASThreeOpImmOps<
+      Machine, typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister,
+      typename InstX86Base<Machine>::Traits::RegisterSet::GPRRegister,
+      InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm,
+      InstX86Base<Machine>::Traits::RegisterSet::getEncodedGPR>(
+      Func, DispatchTy, this->getDest(), Src0, this->getSrc(2), Emitter);
+}
+
+template <class Machine>
+void InstX86BasePshufd<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 2);
+  const Variable *Dest = this->getDest();
+  Type Ty = Dest->getType();
+  static const typename InstX86Base<Machine>::Traits::Assembler::
+      template ThreeOpImmEmitter<
+          typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister,
+          typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister>
+          Emitter = {&InstX86Base<Machine>::Traits::Assembler::pshufd,
+                     &InstX86Base<Machine>::Traits::Assembler::pshufd};
+  emitIASThreeOpImmOps<
+      Machine, typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister,
+      typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister,
+      InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm,
+      InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm>(
+      Func, Ty, Dest, this->getSrc(0), this->getSrc(1), Emitter);
+}
+
+template <class Machine>
+void InstX86BaseShufps<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 3);
+  const Variable *Dest = this->getDest();
+  assert(Dest == this->getSrc(0));
+  Type Ty = Dest->getType();
+  static const typename InstX86Base<Machine>::Traits::Assembler::
+      template ThreeOpImmEmitter<
+          typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister,
+          typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister>
+          Emitter = {&InstX86Base<Machine>::Traits::Assembler::shufps,
+                     &InstX86Base<Machine>::Traits::Assembler::shufps};
+  emitIASThreeOpImmOps<
+      Machine, typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister,
+      typename InstX86Base<Machine>::Traits::RegisterSet::XmmRegister,
+      InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm,
+      InstX86Base<Machine>::Traits::RegisterSet::getEncodedXmm>(
+      Func, Ty, Dest, this->getSrc(1), this->getSrc(2), Emitter);
+}
+
+template <class Machine>
+void InstX86BasePop<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 0);
+  Str << "\tpop\t";
+  this->getDest()->emit(Func);
+}
+
+template <class Machine>
+void InstX86BasePop<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 0);
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  if (this->getDest()->hasReg()) {
+    Asm->popl(InstX86Base<Machine>::Traits::RegisterSet::getEncodedGPR(
+        this->getDest()->getRegNum()));
+  } else {
+    Asm->popl(
+        static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+            Func->getTarget())
+            ->stackVarToAsmOperand(this->getDest()));
+  }
+}
+
+template <class Machine>
+void InstX86BasePop<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  this->dumpDest(Func);
+  Str << " = pop." << this->getDest()->getType() << " ";
+}
+
+template <class Machine>
+void InstX86BaseAdjustStack<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  Str << "\tsubl\t$" << Amount << ", %esp";
+  Func->getTarget()->updateStackAdjustment(Amount);
+}
+
+template <class Machine>
+void InstX86BaseAdjustStack<Machine>::emitIAS(const Cfg *Func) const {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  Asm->sub(IceType_i32,
+           InstX86Base<Machine>::Traits::RegisterSet::Encoded_Reg_esp,
+           Immediate(Amount));
+  Func->getTarget()->updateStackAdjustment(Amount);
+}
+
+template <class Machine>
+void InstX86BaseAdjustStack<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  Str << "esp = sub.i32 esp, " << Amount;
+}
+
+template <class Machine>
+void InstX86BasePush<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(this->getSrcSize() == 1);
+  // Push is currently only used for saving GPRs.
+  const auto Var = llvm::cast<Variable>(this->getSrc(0));
+  assert(Var->hasReg());
+  Str << "\tpush\t";
+  Var->emit(Func);
+}
+
+template <class Machine>
+void InstX86BasePush<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 1);
+  // Push is currently only used for saving GPRs.
+  const auto Var = llvm::cast<Variable>(this->getSrc(0));
+  assert(Var->hasReg());
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  Asm->pushl(InstX86Base<Machine>::Traits::RegisterSet::getEncodedGPR(
+      Var->getRegNum()));
+}
+
+template <class Machine>
+void InstX86BasePush<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  Str << "push." << this->getSrc(0)->getType() << " ";
+  this->dumpSources(Func);
+}
+
+template <class Machine>
+void InstX86BasePsll<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  assert(this->getDest()->getType() == IceType_v8i16 ||
+         this->getDest()->getType() == IceType_v8i1 ||
+         this->getDest()->getType() == IceType_v4i32 ||
+         this->getDest()->getType() == IceType_v4i1);
+  char buf[30];
+  snprintf(
+      buf, llvm::array_lengthof(buf), "psll%s",
+      InstX86Base<Machine>::Traits::TypeAttributes[this->getDest()->getType()]
+          .PackString);
+  this->emitTwoAddress(buf, this, Func);
+}
+
+template <class Machine>
+void InstX86BasePsra<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  assert(this->getDest()->getType() == IceType_v8i16 ||
+         this->getDest()->getType() == IceType_v8i1 ||
+         this->getDest()->getType() == IceType_v4i32 ||
+         this->getDest()->getType() == IceType_v4i1);
+  char buf[30];
+  snprintf(
+      buf, llvm::array_lengthof(buf), "psra%s",
+      InstX86Base<Machine>::Traits::TypeAttributes[this->getDest()->getType()]
+          .PackString);
+  this->emitTwoAddress(buf, this, Func);
+}
+
+template <class Machine>
+void InstX86BasePsrl<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  char buf[30];
+  snprintf(
+      buf, llvm::array_lengthof(buf), "psrl%s",
+      InstX86Base<Machine>::Traits::TypeAttributes[this->getDest()->getType()]
+          .PackString);
+  this->emitTwoAddress(buf, this, Func);
+}
+
+template <class Machine>
+void InstX86BaseRet<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  Str << "\tret";
+}
+
+template <class Machine>
+void InstX86BaseRet<Machine>::emitIAS(const Cfg *Func) const {
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  Asm->ret();
+}
+
+template <class Machine>
+void InstX86BaseRet<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  Type Ty =
+      (this->getSrcSize() == 0 ? IceType_void : this->getSrc(0)->getType());
+  Str << "ret." << Ty << " ";
+  this->dumpSources(Func);
+}
+
+template <class Machine>
+void InstX86BaseSetcc<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  Str << "\tset"
+      << InstX86Base<Machine>::Traits::InstBrAttributes[Condition].DisplayString
+      << "\t";
+  this->Dest->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseSetcc<Machine>::emitIAS(const Cfg *Func) const {
+  assert(Condition != InstX86Base<Machine>::Traits::Cond::Br_None);
+  assert(this->getDest()->getType() == IceType_i1);
+  assert(this->getSrcSize() == 0);
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  if (this->getDest()->hasReg())
+    Asm->setcc(Condition,
+               InstX86Base<Machine>::Traits::RegisterSet::getEncodedByteReg(
+                   this->getDest()->getRegNum()));
+  else
+    Asm->setcc(
+        Condition,
+        static_cast<typename InstX86Base<Machine>::Traits::TargetLowering *>(
+            Func->getTarget())
+            ->stackVarToAsmOperand(this->getDest()));
+  return;
+}
+
+template <class Machine>
+void InstX86BaseSetcc<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  Str << "setcc."
+      << InstX86Base<Machine>::Traits::InstBrAttributes[Condition].DisplayString
+      << " ";
+  this->dumpDest(Func);
+}
+
+template <class Machine>
+void InstX86BaseXadd<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  if (this->Locked) {
+    Str << "\tlock";
+  }
+  Str << "\txadd" << this->getWidthString(this->getSrc(0)->getType()) << "\t";
+  this->getSrc(1)->emit(Func);
+  Str << ", ";
+  this->getSrc(0)->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseXadd<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 2);
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  Type Ty = this->getSrc(0)->getType();
+  const auto Mem =
+      llvm::cast<typename InstX86Base<Machine>::Traits::X86OperandMem>(
+          this->getSrc(0));
+  assert(Mem->getSegmentRegister() ==
+         InstX86Base<Machine>::Traits::X86OperandMem::DefaultSegment);
+  const typename InstX86Base<Machine>::Traits::Address Addr =
+      Mem->toAsmAddress(Asm);
+  const auto VarReg = llvm::cast<Variable>(this->getSrc(1));
+  assert(VarReg->hasReg());
+  const typename InstX86Base<Machine>::Traits::RegisterSet::GPRRegister Reg =
+      InstX86Base<Machine>::Traits::RegisterSet::getEncodedGPR(
+          VarReg->getRegNum());
+  Asm->xadd(Ty, Addr, Reg, this->Locked);
+}
+
+template <class Machine>
+void InstX86BaseXadd<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  if (this->Locked) {
+    Str << "lock ";
+  }
+  Type Ty = this->getSrc(0)->getType();
+  Str << "xadd." << Ty << " ";
+  this->dumpSources(Func);
+}
+
+template <class Machine>
+void InstX86BaseXchg<Machine>::emit(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrEmit();
+  Str << "\txchg" << this->getWidthString(this->getSrc(0)->getType()) << "\t";
+  this->getSrc(1)->emit(Func);
+  Str << ", ";
+  this->getSrc(0)->emit(Func);
+}
+
+template <class Machine>
+void InstX86BaseXchg<Machine>::emitIAS(const Cfg *Func) const {
+  assert(this->getSrcSize() == 2);
+  typename InstX86Base<Machine>::Traits::Assembler *Asm =
+      Func->getAssembler<typename InstX86Base<Machine>::Traits::Assembler>();
+  Type Ty = this->getSrc(0)->getType();
+  const auto Mem =
+      llvm::cast<typename InstX86Base<Machine>::Traits::X86OperandMem>(
+          this->getSrc(0));
+  assert(Mem->getSegmentRegister() ==
+         InstX86Base<Machine>::Traits::X86OperandMem::DefaultSegment);
+  const typename InstX86Base<Machine>::Traits::Address Addr =
+      Mem->toAsmAddress(Asm);
+  const auto VarReg = llvm::cast<Variable>(this->getSrc(1));
+  assert(VarReg->hasReg());
+  const typename InstX86Base<Machine>::Traits::RegisterSet::GPRRegister Reg =
+      InstX86Base<Machine>::Traits::RegisterSet::getEncodedGPR(
+          VarReg->getRegNum());
+  Asm->xchg(Ty, Addr, Reg);
+}
+
+template <class Machine>
+void InstX86BaseXchg<Machine>::dump(const Cfg *Func) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Func->getContext()->getStrDump();
+  Type Ty = this->getSrc(0)->getType();
+  Str << "xchg." << Ty << " ";
+  this->dumpSources(Func);
+}
+
+} // end of namespace X86Internal
+
+} // end of namespace Ice
+
+#endif // SUBZERO_SRC_ICEINSTX86BASEIMPL_H