Start adding an integrated assembler.

src/IceInstX8632.cpp

Index: src/IceInstX8632.cpp
diff --git a/src/IceInstX8632.cpp b/src/IceInstX8632.cpp
index 373ab59e47503c4b3535edaaf87bc17a4aa015f5..fdd14fa579879bee919f5c56443e5eeeb25f93b4 100644
--- a/src/IceInstX8632.cpp
+++ b/src/IceInstX8632.cpp
@@ -12,6 +12,7 @@
 //
 //===----------------------------------------------------------------------===//
 
+#include "assembler_ia32.h"
 #include "IceCfg.h"
 #include "IceCfgNode.h"
 #include "IceInst.h"
@@ -63,6 +64,51 @@ const char *InstX8632SegmentRegNames[] = {
 #undef X
 };
 
+x86::Condition convertToAsmCondition(InstX8632::BrCond Cond) {
+  return x86::Condition(Cond);
+}
+
+x86::Register convertToAsmGPR(int32_t RegNum) {
+  assert(TargetX8632::Reg_eax <= RegNum && RegNum <= TargetX8632::Reg_edi);
+  return x86::Register(RegNum);
+}
+
+x86::ByteRegister convertToAsmByteRegister(int32_t RegNum) {
+  if (RegNum == TargetX8632::Reg_ah) {
+    return x86::AH;
+  }
+  assert(TargetX8632::Reg_eax <= RegNum && RegNum <= TargetX8632::Reg_ebx);
+  return x86::ByteRegister(RegNum);
+}
+
+x86::Immediate convertIntImmToAsmImm(const ConstantInteger *CI) {
+  int32_t V = static_cast<int32_t>(CI->getValue());
+  // Check that it's really 32-bit.
+  // Sometimes Subzero stores negative values as sign-extended 64-bit
+  // and sometimes it's zero-extended to 64-bit.
+  // Should we just make a ConstantInteger32?

[Jim Stichnoth, 2014/09/09 14:58:41]

I was also thinking that maybe constant integers need to be enhanced somewhat. 
At a minimum, adding a flag indicating whether it is intended to be signed or
unsigned (or unknown / don't care?) so that when getting its value, validation
against width can be done.

[jvoung (off chromium), 2014/09/15 17:19:15]

Hmm, we could add signed/unsigned, but I'm not yet sure how that would be used.
Constants read in from the bitcode file are always sign-extended to the constant
pool's storage type. Constants created from within subzero might be sign or zero
extended. E.g., what happens when adding two constants in the addressing mode
optimization and making a new constant.

Regardless, it seemed like if we had a 32-bit constant pool as well, then the
check here could be free. Lots of constants that we use internally are 32-bit or
less. Constants read in by the bitcode reader could be classified at the time
it's read, and by the time we emit instructions/bytes, the lowering should have
split 64-bit constants into 32-bit ones. Maybe that's a bit too x86 specific...
For ARM, lowering would split a constant into upper16 and lower16 instead, or it
might want to know that a constant offset fits in 12-bits for load/store.

Anyway, I tried splitting the constant pool more, to see what that would look
like:
https://codereview.chromium.org/569033002/

Does that seem worthwhile?

+  assert((CI->getValue() >> 32) == 0 || (CI->getValue() >> 32) == 0xffffffff);
+  return x86::Immediate(V);
+}
+
+x86::XmmRegister convertToAsmXMMReg(int32_t RegNum) {
+  assert(TargetX8632::Reg_xmm0 <= RegNum && RegNum <= TargetX8632::Reg_xmm7);
+  return x86::XmmRegister(RegNum - TargetX8632::Reg_xmm0);
+}
+
+// Convert a float or double immediate to an Address Operand.
+// Modifies SymbolicOffset to contain the symbol.
+x86::Address convertFloatImmToAsmAddr(const Constant *Imm,
+                                      IceString &SymbolicOffset) {
+  std::string Buffer;
+  llvm::raw_string_ostream StrBuf(Buffer);
+  Type Ty = Imm->getType();
+  assert(llvm::isa<ConstantFloat>(Imm) || llvm::isa<ConstantDouble>(Imm));
+  StrBuf << "L$" << Ty << "$" << Imm->getPoolEntryID();
+  SymbolicOffset = StrBuf.str();
+  return x86::Address::Absolute(0);
+}
+
 } // end of anonymous namespace
 
 const char *InstX8632::getWidthString(Type Ty) {
@@ -285,12 +331,207 @@ InstX8632Xchg::InstX8632Xchg(Cfg *Func, Operand *Dest, Variable *Source)
 
 // ======================== Dump routines ======================== //
 
+namespace {
+
+void emitIASBytes(Ostream &Str, const x86::AssemblerX86 *Asm,
+                  intptr_t StartPosition,
+                  const IceString &SymbolicOffset = IceString("")) {
+  intptr_t EndPosition = Asm->GetPosition();
+  if (!SymbolicOffset.empty()) {
+    const intptr_t OffsetSize = 4;
+    EndPosition -= OffsetSize;
+  }
+  for (intptr_t i = 0; i < EndPosition - StartPosition; ++i) {
+    Str << "\t.byte "
+        << static_cast<uint32_t>(Asm->LoadBuffer<uint8_t>(StartPosition + i))
+        << "\n";
+  }
+  if (!SymbolicOffset.empty()) {
+    int32_t Disp = Asm->LoadBuffer<int32_t>(EndPosition);
+    if (Disp)
+      Str << "\t.long " << SymbolicOffset << " + " << Disp << "\n";
+    else
+      Str << "\t.long " << SymbolicOffset << "\n";
+  }
+}
+
+void emitIASVarOperandInt(const Cfg *Func, const Variable *Var,
+                          const Operand *Src,
+                          x86::AssemblerX86::EmitTyRegReg EmitRegReg,
+                          x86::AssemblerX86::EmitTyRegAddr EmitRegAddr,
+                          x86::AssemblerX86::EmitTyRegImm EmitRegImm) {
+  // For now, we assume Var is reg, since most of the time we use:
+  //     _add(T, foo)
+  //     _mov(Var, T)
+  // where T is a register.
+  //
+  // Otherwise, if Var can be mem, use the emitIASTwoOperandInt function..
+  x86::AssemblerX86 *Asm = Func->getAssembler<x86::AssemblerX86>();
+  intptr_t StartPosition = Asm->GetPosition();
+  IceString SymbolicOffset = "";
+  Type Ty = Var->getType();
+  // Classify Var.
+  x86::Register VarReg;
+  assert(Var->hasReg());
+  if (typeWidthInBytes(Ty) == 1) {
+    // We fudge a bit, and use Register for ByteRegister.
+    // This avoids having to add another assembler method,
+    // just for ByteRegisters. The register encoding is the same.
+    VarReg = convertToAsmGPR(convertToAsmByteRegister(Var->getRegNum()));
+  } else {
+    VarReg = convertToAsmGPR(Var->getRegNum());
+  }
+
+  // Classify Src.
+  if (const Variable *SrcVar = llvm::dyn_cast<Variable>(Src)) {
+    if (SrcVar->hasReg()) {
+      x86::Register SrcReg;
+      if (typeWidthInBytes(Ty) == 1) {
+        SrcReg = convertToAsmGPR(convertToAsmByteRegister(SrcVar->getRegNum()));
+      } else {
+        SrcReg = convertToAsmGPR(SrcVar->getRegNum());
+      }
+      (Asm->*EmitRegReg)(Ty, VarReg, SrcReg);
+    } else {
+      x86::Address SrcAddr = static_cast<TargetX8632 *>(Func->getTarget())
+                                 ->stackVarToAsmOperand(SrcVar);
+      (Asm->*EmitRegAddr)(Ty, VarReg, SrcAddr);
+    }
+  } else if (const OperandX8632Mem *Mem =
+                 llvm::dyn_cast<OperandX8632Mem>(Src)) {
+    x86::Address SrcAddr = Mem->convertToAsmAddress(SymbolicOffset);
+    (Asm->*EmitRegAddr)(Ty, VarReg, SrcAddr);
+  } else if (const ConstantInteger *CI = llvm::dyn_cast<ConstantInteger>(Src)) {
+    x86::Immediate Imm = convertIntImmToAsmImm(CI);
+    (Asm->*EmitRegImm)(Ty, VarReg, Imm);
+  }
+  Ostream &Str = Func->getContext()->getStrEmit();
+  emitIASBytes(Str, Asm, StartPosition, SymbolicOffset);
+}
+
+void emitIASTwoOperandInt(const Cfg *Func, const Operand *Src0,
+                          const Operand *Src1,
+                          x86::AssemblerX86::EmitTyRegReg EmitRegReg,
+                          x86::AssemblerX86::EmitTyRegAddr EmitRegAddr,
+                          x86::AssemblerX86::EmitTyRegImm EmitRegImm,
+                          x86::AssemblerX86::EmitTyAddrReg EmitAddrReg,
+                          x86::AssemblerX86::EmitTyAddrImm EmitAddrImm) {
+  x86::AssemblerX86 *Asm = Func->getAssembler<x86::AssemblerX86>();
+  intptr_t StartPosition = Asm->GetPosition();
+  IceString SymbolicOffset = "";
+  Type Ty = Src0->getType();
+
+  // Classify Src0 (Reg or Mem).
+  // Could use a lambda/functional to curry the first parameter.
+  bool Src0IsReg = false;
+  x86::Register Src0Reg = x86::kNoRegister;
+  x86::Address Src0Addr = x86::Address::Absolute(0);
+  if (const Variable *Src0Var = llvm::dyn_cast<Variable>(Src0)) {
+    if (Src0Var->hasReg()) {
+      Src0IsReg = true;
+      if (typeWidthInBytes(Ty) == 1) {
+        Src0Reg =
+            convertToAsmGPR(convertToAsmByteRegister(Src0Var->getRegNum()));
+      } else {
+        Src0Reg = convertToAsmGPR(Src0Var->getRegNum());
+      }
+    } else {
+      Src0Addr = static_cast<TargetX8632 *>(Func->getTarget())
+                     ->stackVarToAsmOperand(Src0Var);
+    }
+  } else if (const OperandX8632Mem *Mem =
+                 llvm::dyn_cast<OperandX8632Mem>(Src0)) {
+    Src0Addr = Mem->convertToAsmAddress(SymbolicOffset);
+  }
+
+  // Classify Src1 (Reg, Mem, or Imm).
+  if (const Variable *Src1Var = llvm::dyn_cast<Variable>(Src1)) {
+    if (Src1Var->hasReg()) {
+      x86::Register Src1Reg;
+      if (typeWidthInBytes(Ty) == 1) {
+        Src1Reg =
+            convertToAsmGPR(convertToAsmByteRegister(Src1Var->getRegNum()));
+      } else {
+        Src1Reg = convertToAsmGPR(Src1Var->getRegNum());
+      }
+      if (Src0IsReg)
+        (Asm->*EmitRegReg)(Ty, Src0Reg, Src1Reg);
+      else
+        (Asm->*EmitAddrReg)(Ty, Src0Addr, Src1Reg);
+    } else {
+      x86::Address Src1Addr = static_cast<TargetX8632 *>(Func->getTarget())
+                                  ->stackVarToAsmOperand(Src1Var);
+      assert(Src0IsReg);
+      (Asm->*EmitRegAddr)(Ty, Src0Reg, Src1Addr);
+    }
+  } else if (const OperandX8632Mem *Mem =
+                 llvm::dyn_cast<OperandX8632Mem>(Src1)) {
+    x86::Address Src1Addr = Mem->convertToAsmAddress(SymbolicOffset);
+    assert(Src0IsReg);
+    (Asm->*EmitRegAddr)(Ty, Src0Reg, Src1Addr);
+  } else if (const ConstantInteger *CI =
+                 llvm::dyn_cast<ConstantInteger>(Src1)) {
+    x86::Immediate Imm = convertIntImmToAsmImm(CI);
+    if (Src0IsReg)
+      (Asm->*EmitRegImm)(Ty, Src0Reg, Imm);
+    else
+      (Asm->*EmitAddrImm)(Ty, Src0Addr, Imm);
+    // TODO(jvoung): Record the buffer position containing the relocation,
+    // instead of just assuming it's the last 4 bytes. That would not be
+    // true if Src0 is Memory w/ a symbolic displacement, and Src1
+    // is an immediate. The immediate would take up the last few bytes
+    // and the memory operand's displacement would come before those imm bytes.
+    //
+    // Perhaps the "Address" Operand type could indicate that it's a
+    // relocatable, and then the emitOperand() methods will check
+    // and create Fixups inline while the buffer is in the correct position.
+    assert(SymbolicOffset.empty());
+  }
+  Ostream &Str = Func->getContext()->getStrEmit();
+  emitIASBytes(Str, Asm, StartPosition, SymbolicOffset);
+}
+
+void emitIASVarOperandXMM(const Cfg *Func, const Variable *Var,
+                          const Operand *Src,
+                          x86::AssemblerX86::EmitXmmXmm EmitXmmXmm,
+                          x86::AssemblerX86::EmitXmmAddr EmitXmmAddr) {
+  x86::AssemblerX86 *Asm = Func->getAssembler<x86::AssemblerX86>();
+  intptr_t StartPosition = Asm->GetPosition();
+  assert(Var->hasReg());
+  x86::XmmRegister VarReg = convertToAsmXMMReg(Var->getRegNum());
+  IceString SymbolicOffset = "";
+  if (const Variable *SrcVar = llvm::dyn_cast<Variable>(Src)) {
+    if (SrcVar->hasReg()) {
+      x86::XmmRegister SrcReg = convertToAsmXMMReg(SrcVar->getRegNum());
+      (Asm->*EmitXmmXmm)(VarReg, SrcReg);
+    } else {
+      x86::Address SrcStackAddr = static_cast<TargetX8632 *>(Func->getTarget())
+                                      ->stackVarToAsmOperand(SrcVar);
+      (Asm->*EmitXmmAddr)(VarReg, SrcStackAddr);
+    }
+  } else if (const OperandX8632Mem *Mem =
+                 llvm::dyn_cast<OperandX8632Mem>(Src)) {
+    x86::Address SrcAddr = Mem->convertToAsmAddress(SymbolicOffset);
+    (Asm->*EmitXmmAddr)(VarReg, SrcAddr);
+  } else if (const Constant *Imm = llvm::dyn_cast<Constant>(Src)) {
+    (Asm->*EmitXmmAddr)(VarReg, convertFloatImmToAsmAddr(Imm, SymbolicOffset));
+  } else {
+    llvm_unreachable("Unexpected operand type");
+  }
+  Ostream &Str = Func->getContext()->getStrEmit();
+  emitIASBytes(Str, Asm, StartPosition, SymbolicOffset);
+}
+
+} // end of anonymous namespace
+
 void InstX8632::dump(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrDump();
   Str << "[X8632] ";
   Inst::dump(Func);
 }
 
+void InstX8632::emitIAS(const Cfg *Func) const { emit(Func); }
+
 void InstX8632Label::emit(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrEmit();
   Str << getName(Func) << ":\n";
@@ -355,6 +596,43 @@ void InstX8632Call::emit(const Cfg *Func) const {
   Func->getTarget()->resetStackAdjustment();
 }
 
+void InstX8632Call::emitIAS(const Cfg *Func) const {
+  Ostream &Str = Func->getContext()->getStrEmit();
+  x86::AssemblerX86 *Asm = Func->getAssembler<x86::AssemblerX86>();
+  intptr_t StartPosition = Asm->GetPosition();
+  Operand *Target = getCallTarget();
+  IceString SymbolicOffset = "";
+  if (Variable *Var = llvm::dyn_cast<Variable>(Target)) {
+    if (Var->hasReg()) {
+      Asm->call(convertToAsmGPR(Var->getRegNum()));
+    } else {
+      Asm->call(static_cast<TargetX8632 *>(Func->getTarget())
+                    ->stackVarToAsmOperand(Var));
+    }
+  } else if (OperandX8632Mem *Mem = llvm::dyn_cast<OperandX8632Mem>(Target)) {
+    Asm->call(Mem->convertToAsmAddress(SymbolicOffset));
+  } else if (ConstantRelocatable *CR =
+                 llvm::dyn_cast<ConstantRelocatable>(Target)) {
+    assert(CR->getOffset() == 0 && "We only support calling a function");
+    SymbolicOffset = CR->getName();
+    Asm->call(CR);
+  }
+  if (!SymbolicOffset.empty()) {
+    // TODO(jvoung): This .byte and .long hack doesn't work, since we need
+    // a pc-rel relocation, and we also need the section contents to be
+    // -4 instead of 0.
+    //
+    // Still, we have at least filled the assembler buffer so that the
+    // instruction sizes/positions are correct for jumps.
+    //
+    // For now, fall back to the regular .s emission.
+    emit(Func);
+  } else {
+    emitIASBytes(Str, Asm, StartPosition, SymbolicOffset);
+  }
+  Func->getTarget()->resetStackAdjustment();
+}
+
 void InstX8632Call::dump(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrDump();
   if (getDest()) {
@@ -478,6 +756,37 @@ template <> void InstX8632Sqrtss::emit(const Cfg *Func) const {
   Str << "\n";
 }
 
+template <> void InstX8632Sqrtss::emitIAS(const Cfg *Func) const {
+  Type Ty = getDest()->getType();
+  assert(getSrcSize() == 1);
+  assert(Ty == IceType_f32 || Ty == IceType_f64);
+  if (Ty == IceType_f64) {
+    emitIASVarOperandXMM(Func, getDest(), getSrc(0), &x86::AssemblerX86::sqrtsd,
+                         &x86::AssemblerX86::sqrtsd);
+  } else {
+    emitIASVarOperandXMM(Func, getDest(), getSrc(0), &x86::AssemblerX86::sqrtss,
+                         &x86::AssemblerX86::sqrtss);
+  }
+}
+
+template <> void InstX8632Adc::emitIAS(const Cfg *Func) const {
+  assert(getSrcSize() == 2);
+  emitIASVarOperandInt(Func, getDest(), getSrc(1), &x86::AssemblerX86::adc,
+                       &x86::AssemblerX86::adc, &x86::AssemblerX86::adc);
+}
+
+template <> void InstX8632Add::emitIAS(const Cfg *Func) const {
+  assert(getSrcSize() == 2);
+  emitIASVarOperandInt(Func, getDest(), getSrc(1), &x86::AssemblerX86::add,
+                       &x86::AssemblerX86::add, &x86::AssemblerX86::add);
+}
+
+template <> void InstX8632Addps::emitIAS(const Cfg *Func) const {
+  assert(getSrcSize() == 2);
+  emitIASVarOperandXMM(Func, getDest(), getSrc(1), &x86::AssemblerX86::addps,
+                       &x86::AssemblerX86::addps);
+}
+
 template <> void InstX8632Addss::emit(const Cfg *Func) const {
   char buf[30];
   snprintf(buf, llvm::array_lengthof(buf), "add%s",
@@ -485,6 +794,24 @@ template <> void InstX8632Addss::emit(const Cfg *Func) const {
   emitTwoAddress(buf, this, Func);
 }
 
+template <> void InstX8632Addss::emitIAS(const Cfg *Func) const {
+  Type Ty = getDest()->getType();
+  assert(getSrcSize() == 2);
+  if (Ty == IceType_f64) {
+    emitIASVarOperandXMM(Func, getDest(), getSrc(1), &x86::AssemblerX86::addsd,
+                         &x86::AssemblerX86::addsd);
+  } else {
+    emitIASVarOperandXMM(Func, getDest(), getSrc(1), &x86::AssemblerX86::addss,
+                         &x86::AssemblerX86::addss);
+  }
+}
+
+template <> void InstX8632And::emitIAS(const Cfg *Func) const {
+  assert(getSrcSize() == 2);
+  emitIASVarOperandInt(Func, getDest(), getSrc(1), &x86::AssemblerX86::_and,
+                       &x86::AssemblerX86::_and, &x86::AssemblerX86::_and);
+}
+
 template <> void InstX8632Padd::emit(const Cfg *Func) const {
   char buf[30];
   snprintf(buf, llvm::array_lengthof(buf), "padd%s",
@@ -509,6 +836,24 @@ template <> void InstX8632Pmull::emit(const Cfg *Func) const {
   emitTwoAddress(buf, this, Func);
 }
 
+template <> void InstX8632Sbb::emitIAS(const Cfg *Func) const {
+  assert(getSrcSize() == 2);
+  emitIASVarOperandInt(Func, getDest(), getSrc(1), &x86::AssemblerX86::sbb,
+                       &x86::AssemblerX86::sbb, &x86::AssemblerX86::sbb);
+}
+
+template <> void InstX8632Sub::emitIAS(const Cfg *Func) const {
+  assert(getSrcSize() == 2);
+  emitIASVarOperandInt(Func, getDest(), getSrc(1), &x86::AssemblerX86::sub,
+                       &x86::AssemblerX86::sub, &x86::AssemblerX86::sub);
+}
+
+template <> void InstX8632Subps::emitIAS(const Cfg *Func) const {
+  assert(getSrcSize() == 2);
+  emitIASVarOperandXMM(Func, getDest(), getSrc(1), &x86::AssemblerX86::subps,
+                       &x86::AssemblerX86::subps);
+}
+
 template <> void InstX8632Subss::emit(const Cfg *Func) const {
   char buf[30];
   snprintf(buf, llvm::array_lengthof(buf), "sub%s",
@@ -516,6 +861,18 @@ template <> void InstX8632Subss::emit(const Cfg *Func) const {
   emitTwoAddress(buf, this, Func);
 }
 
+template <> void InstX8632Subss::emitIAS(const Cfg *Func) const {
+  Type Ty = getDest()->getType();
+  assert(getSrcSize() == 2);
+  if (Ty == IceType_f64) {
+    emitIASVarOperandXMM(Func, getDest(), getSrc(1), &x86::AssemblerX86::subsd,
+                         &x86::AssemblerX86::subsd);
+  } else {
+    emitIASVarOperandXMM(Func, getDest(), getSrc(1), &x86::AssemblerX86::subss,
+                         &x86::AssemblerX86::subss);
+  }
+}
+
 template <> void InstX8632Psub::emit(const Cfg *Func) const {
   char buf[30];
   snprintf(buf, llvm::array_lengthof(buf), "psub%s",
@@ -523,6 +880,12 @@ template <> void InstX8632Psub::emit(const Cfg *Func) const {
   emitTwoAddress(buf, this, Func);
 }
 
+template <> void InstX8632Mulps::emitIAS(const Cfg *Func) const {
+  assert(getSrcSize() == 2);
+  emitIASVarOperandXMM(Func, getDest(), getSrc(1), &x86::AssemblerX86::mulps,
+                       &x86::AssemblerX86::mulps);
+}
+
 template <> void InstX8632Mulss::emit(const Cfg *Func) const {
   char buf[30];
   snprintf(buf, llvm::array_lengthof(buf), "mul%s",
@@ -530,12 +893,30 @@ template <> void InstX8632Mulss::emit(const Cfg *Func) const {
   emitTwoAddress(buf, this, Func);
 }
 
+template <> void InstX8632Mulss::emitIAS(const Cfg *Func) const {
+  Type Ty = getDest()->getType();
+  assert(getSrcSize() == 2);
+  if (Ty == IceType_f64) {
+    emitIASVarOperandXMM(Func, getDest(), getSrc(1), &x86::AssemblerX86::mulsd,
+                         &x86::AssemblerX86::mulsd);
+  } else {
+    emitIASVarOperandXMM(Func, getDest(), getSrc(1), &x86::AssemblerX86::mulss,
+                         &x86::AssemblerX86::mulss);
+  }
+}
+
 template <> void InstX8632Pmuludq::emit(const Cfg *Func) const {
   assert(getSrc(0)->getType() == IceType_v4i32 &&
          getSrc(1)->getType() == IceType_v4i32);
   emitTwoAddress(Opcode, this, Func);
 }
 
+template <> void InstX8632Divps::emitIAS(const Cfg *Func) const {
+  assert(getSrcSize() == 2);
+  emitIASVarOperandXMM(Func, getDest(), getSrc(1), &x86::AssemblerX86::divps,
+                       &x86::AssemblerX86::divps);
+}
+
 template <> void InstX8632Divss::emit(const Cfg *Func) const {
   char buf[30];
   snprintf(buf, llvm::array_lengthof(buf), "div%s",
@@ -543,6 +924,18 @@ template <> void InstX8632Divss::emit(const Cfg *Func) const {
   emitTwoAddress(buf, this, Func);
 }
 
+template <> void InstX8632Divss::emitIAS(const Cfg *Func) const {
+  Type Ty = getDest()->getType();
+  assert(getSrcSize() == 2);
+  if (Ty == IceType_f64) {
+    emitIASVarOperandXMM(Func, getDest(), getSrc(1), &x86::AssemblerX86::divsd,
+                         &x86::AssemblerX86::divsd);
+  } else {
+    emitIASVarOperandXMM(Func, getDest(), getSrc(1), &x86::AssemblerX86::divss,
+                         &x86::AssemblerX86::divss);
+  }
+}
+
 template <> void InstX8632Div::emit(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrEmit();
   assert(getSrcSize() == 3);
@@ -640,6 +1033,34 @@ template <> void InstX8632Cbwdq::emit(const Cfg *Func) const {
   }
 }
 
+template <> void InstX8632Cbwdq::emitIAS(const Cfg *Func) const {
+  Ostream &Str = Func->getContext()->getStrEmit();
+  x86::AssemblerX86 *Asm = Func->getAssembler<x86::AssemblerX86>();
+  intptr_t StartPosition = Asm->GetPosition();
+  assert(getSrcSize() == 1);
+  Operand *Src0 = getSrc(0);
+  assert(llvm::isa<Variable>(Src0));
+  assert(llvm::cast<Variable>(Src0)->getRegNum() == TargetX8632::Reg_eax);
+  switch (Src0->getType()) {
+  default:
+    llvm_unreachable("unexpected source type!");
+    break;
+  case IceType_i8:
+    assert(getDest()->getRegNum() == TargetX8632::Reg_eax);
+    Asm->cbw();
+    break;
+  case IceType_i16:
+    assert(getDest()->getRegNum() == TargetX8632::Reg_edx);
+    Asm->cwd();
+    break;
+  case IceType_i32:
+    assert(getDest()->getRegNum() == TargetX8632::Reg_edx);
+    Asm->cdq();
+    break;
+  }
+  emitIASBytes(Str, Asm, StartPosition);
+}
+
 void InstX8632Mul::emit(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrEmit();
   assert(getSrcSize() == 2);
@@ -723,6 +1144,23 @@ void InstX8632Cmov::emit(const Cfg *Func) const {
   Str << "\n";
 }
 
+void InstX8632Cmov::emitIAS(const Cfg *Func) const {
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(Condition != Br_None);
+  assert(getDest()->hasReg());
+  assert(getSrcSize() == 2);
+  const Variable *Src = llvm::cast<Variable>(getSrc(1));
+  // Only need the 32-bit register form right now.
+  assert(Src->hasReg());
+  assert(Src->getType() == IceType_i32);
+  x86::AssemblerX86 *Asm = Func->getAssembler<x86::AssemblerX86>();
+  intptr_t StartPosition = Asm->GetPosition();
+  Asm->cmov(convertToAsmCondition(Condition),
+            convertToAsmGPR(getDest()->getRegNum()),
+            convertToAsmGPR(Src->getRegNum()));
+  emitIASBytes(Str, Asm, StartPosition);
+}
+
 void InstX8632Cmov::dump(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrDump();
   Str << "cmov" << InstX8632BrAttributes[Condition].DisplayString << ".";
@@ -745,6 +1183,28 @@ void InstX8632Cmpps::emit(const Cfg *Func) const {
   Str << "\n";
 }
 
+void InstX8632Cmpps::emitIAS(const Cfg *Func) const {
+  Ostream &Str = Func->getContext()->getStrEmit();
+  x86::AssemblerX86 *Asm = Func->getAssembler<x86::AssemblerX86>();
+  intptr_t StartPosition = Asm->GetPosition();
+  assert(getSrcSize() == 2);
+  assert(Condition < Cmpps_Invalid);
+  // Assuming there isn't any load folding for cmpps, and vector constants
+  // are not allowed in PNaCl.
+  assert(llvm::isa<Variable>(getSrc(1)));
+  const Variable *SrcVar = llvm::cast<Variable>(getSrc(1));
+  if (SrcVar->hasReg()) {
+    Asm->cmpps(convertToAsmXMMReg(getDest()->getRegNum()),
+               convertToAsmXMMReg(SrcVar->getRegNum()), Condition);
+  } else {
+    x86::Address SrcStackAddr = static_cast<TargetX8632 *>(Func->getTarget())
+                                    ->stackVarToAsmOperand(SrcVar);
+    Asm->cmpps(convertToAsmXMMReg(getDest()->getRegNum()), SrcStackAddr,
+               Condition);
+  }
+  emitIASBytes(Str, Asm, StartPosition);
+}
+
 void InstX8632Cmpps::dump(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrDump();
   assert(Condition < Cmpps_Invalid);
@@ -831,6 +1291,13 @@ void InstX8632Icmp::emit(const Cfg *Func) const {
   Str << "\n";
 }
 
+void InstX8632Icmp::emitIAS(const Cfg *Func) const {
+  assert(getSrcSize() == 2);
+  emitIASTwoOperandInt(Func, getSrc(0), getSrc(1), &x86::AssemblerX86::cmp,
+                       &x86::AssemblerX86::cmp, &x86::AssemblerX86::cmp,
+                       &x86::AssemblerX86::cmp, &x86::AssemblerX86::cmp);
+}
+
 void InstX8632Icmp::dump(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrDump();
   Str << "cmp." << getSrc(0)->getType() << " ";
@@ -848,6 +1315,22 @@ void InstX8632Ucomiss::emit(const Cfg *Func) const {
   Str << "\n";
 }
 
+void InstX8632Ucomiss::emitIAS(const Cfg *Func) const {
+  assert(getSrcSize() == 2);
+  // Currently src0 is always a variable by convention, to avoid having
+  // two memory operands.
+  assert(llvm::isa<Variable>(getSrc(0)));
+  const Variable *Src0 = llvm::cast<Variable>(getSrc(0));
+  Type Ty = Src0->getType();
+  if (Ty == IceType_f64) {
+    emitIASVarOperandXMM(Func, Src0, getSrc(1), &x86::AssemblerX86::ucomisd,
+                         &x86::AssemblerX86::ucomisd);
+  } else {
+    emitIASVarOperandXMM(Func, Src0, getSrc(1), &x86::AssemblerX86::ucomiss,
+                         &x86::AssemblerX86::ucomiss);
+  }
+}
+
 void InstX8632Ucomiss::dump(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrDump();
   Str << "ucomiss." << getSrc(0)->getType() << " ";
@@ -1018,6 +1501,11 @@ template <> void InstX8632Mov::emit(const Cfg *Func) const {
   }
 }
 
+template <> void InstX8632Mov::emitIAS(const Cfg *Func) const {
+  // Could be movss/sd, or movb/movw/movl.
+  emit(Func);
+}
+
 template <> void InstX8632Movp::emit(const Cfg *Func) const {
   // TODO(wala,stichnot): movups works with all vector operands, but
   // there exist other instructions (movaps, movdqa, movdqu) that may
@@ -1044,6 +1532,43 @@ template <> void InstX8632Movq::emit(const Cfg *Func) const {
   Str << "\n";
 }
 
+static void emitIASSignZeroExt(const Cfg *Func, const Variable *Dest,
+                               const Operand *Src0,
+                               x86::AssemblerX86::EmitRegByteR EmitByte,
+                               x86::AssemblerX86::EmitRegReg EmitWord,
+                               x86::AssemblerX86::EmitRegAddr EmitAddress) {
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(Dest->hasReg());
+  x86::Register DestReg = convertToAsmGPR(Dest->getRegNum());
+  x86::AssemblerX86 *Asm = Func->getAssembler<x86::AssemblerX86>();
+  intptr_t StartPosition = Asm->GetPosition();
+  IceString SymbolicOffset = "";
+  Type SrcType = Src0->getType();
+  // Source is ByteReg, WordReg, or stack slot, or other memory operand.
+  if (const Variable *SrcVar = llvm::dyn_cast<Variable>(Src0)) {
+    if (SrcVar->hasReg()) {
+      if (SrcType == IceType_i8 || SrcType == IceType_i1) {
+        (Asm->*EmitByte)(DestReg,
+                         convertToAsmByteRegister(SrcVar->getRegNum()));
+      } else {
+        assert(SrcType == IceType_i16);
+        (Asm->*EmitWord)(DestReg, convertToAsmGPR(SrcVar->getRegNum()));
+      }
+    } else {
+      x86::Address StackAddr = static_cast<TargetX8632 *>(Func->getTarget())
+                                   ->stackVarToAsmOperand(SrcVar);
+      (Asm->*EmitAddress)(DestReg, StackAddr);
+    }
+  } else if (const OperandX8632Mem *Mem =
+                 llvm::dyn_cast<OperandX8632Mem>(Src0)) {
+    x86::Address SrcAddr = Mem->convertToAsmAddress(SymbolicOffset);
+    (Asm->*EmitAddress)(DestReg, SrcAddr);
+  } else {
+    llvm_unreachable("Unexpected operand type for Movzx");
+  }
+  emitIASBytes(Str, Asm, StartPosition, SymbolicOffset);
+}
+
 void InstX8632Movsx::emit(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrEmit();
   assert(getSrcSize() == 1);
@@ -1054,6 +1579,20 @@ void InstX8632Movsx::emit(const Cfg *Func) const {
   Str << "\n";
 }
 
+void InstX8632Movsx::emitIAS(const Cfg *Func) const {
+  assert(getSrcSize() == 1);
+  const Operand *Src0 = getSrc(0);
+  Type SrcType = Src0->getType();
+  if (SrcType == IceType_i8 || SrcType == IceType_i1) {
+    emitIASSignZeroExt(Func, getDest(), Src0, &x86::AssemblerX86::movsxb,
+                       &x86::AssemblerX86::movsxw, &x86::AssemblerX86::movsxb);
+  } else {
+    assert(SrcType == IceType_i16);
+    emitIASSignZeroExt(Func, getDest(), Src0, &x86::AssemblerX86::movsxb,
+                       &x86::AssemblerX86::movsxw, &x86::AssemblerX86::movsxw);
+  }
+}
+
 void InstX8632Movsx::dump(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrDump();
   Str << "movsx." << getDest()->getType() << "." << getSrc(0)->getType();
@@ -1073,6 +1612,20 @@ void InstX8632Movzx::emit(const Cfg *Func) const {
   Str << "\n";
 }
 
+void InstX8632Movzx::emitIAS(const Cfg *Func) const {
+  assert(getSrcSize() == 1);
+  const Operand *Src0 = getSrc(0);
+  Type SrcType = Src0->getType();
+  if (SrcType == IceType_i8 || SrcType == IceType_i1) {
+    emitIASSignZeroExt(Func, getDest(), Src0, &x86::AssemblerX86::movzxb,
+                       &x86::AssemblerX86::movzxw, &x86::AssemblerX86::movzxb);
+  } else {
+    assert(SrcType == IceType_i16);
+    emitIASSignZeroExt(Func, getDest(), Src0, &x86::AssemblerX86::movzxb,
+                       &x86::AssemblerX86::movzxw, &x86::AssemblerX86::movzxw);
+  }
+}
+
 void InstX8632Movzx::dump(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrDump();
   Str << "movzx." << getDest()->getType() << "." << getSrc(0)->getType();
@@ -1088,11 +1641,26 @@ void InstX8632Nop::emit(const Cfg *Func) const {
   Str << "\tnop\t# variant = " << Variant << "\n";
 }
 
+void InstX8632Nop::emitIAS(const Cfg *Func) const {
+  Ostream &Str = Func->getContext()->getStrEmit();
+  x86::AssemblerX86 *Asm = Func->getAssembler<x86::AssemblerX86>();
+  intptr_t StartPosition = Asm->GetPosition();
+  // TODO: Emit the right code for the variant.
+  Asm->nop();
+  emitIASBytes(Str, Asm, StartPosition);
+}
+
 void InstX8632Nop::dump(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrDump();
   Str << "nop (variant = " << Variant << ")";
 }
 
+template <> void InstX8632Or::emitIAS(const Cfg *Func) const {
+  assert(getSrcSize() == 2);
+  emitIASVarOperandInt(Func, getDest(), getSrc(1), &x86::AssemblerX86::_or,
+                       &x86::AssemblerX86::_or, &x86::AssemblerX86::_or);
+}
+
 void InstX8632Fld::emit(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrEmit();
   assert(getSrcSize() == 1);
@@ -1222,11 +1790,27 @@ template <> void InstX8632Pinsr::emit(const Cfg *Func) const {
 void InstX8632Pop::emit(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrEmit();
   assert(getSrcSize() == 0);
+  assert(getDest()->getType() == IceType_i32);
   Str << "\tpop\t";
   getDest()->emit(Func);
   Str << "\n";
 }
 
+void InstX8632Pop::emitIAS(const Cfg *Func) const {
+  Ostream &Str = Func->getContext()->getStrEmit();
+  assert(getSrcSize() == 0);
+  assert(getDest()->getType() == IceType_i32);
+  x86::AssemblerX86 *Asm = Func->getAssembler<x86::AssemblerX86>();
+  intptr_t StartPosition = Asm->GetPosition();
+  if (getDest()->hasReg()) {
+    Asm->popl(convertToAsmGPR(getDest()->getRegNum()));
+  } else {
+    Asm->popl(static_cast<TargetX8632 *>(Func->getTarget())
+                  ->stackVarToAsmOperand(getDest()));
+  }
+  emitIASBytes(Str, Asm, StartPosition);
+}
+
 void InstX8632Pop::dump(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrDump();
   dumpDest(Func);
@@ -1311,6 +1895,14 @@ void InstX8632Ret::emit(const Cfg *Func) const {
   Str << "\tret\n";
 }
 
+void InstX8632Ret::emitIAS(const Cfg *Func) const {
+  Ostream &Str = Func->getContext()->getStrEmit();
+  x86::AssemblerX86 *Asm = Func->getAssembler<x86::AssemblerX86>();
+  intptr_t StartPosition = Asm->GetPosition();
+  Asm->ret();
+  emitIASBytes(Str, Asm, StartPosition);
+}
+
 void InstX8632Ret::dump(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrDump();
   Type Ty = (getSrcSize() == 0 ? IceType_void : getSrc(0)->getType());
@@ -1356,6 +1948,12 @@ void InstX8632Xchg::dump(const Cfg *Func) const {
   dumpSources(Func);
 }
 
+template <> void InstX8632Xor::emitIAS(const Cfg *Func) const {
+  assert(getSrcSize() == 2);
+  emitIASVarOperandInt(Func, getDest(), getSrc(1), &x86::AssemblerX86::_xor,
+                       &x86::AssemblerX86::_xor, &x86::AssemblerX86::_xor);
+}
+
 void OperandX8632::dump(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrDump();
   Str << "<OperandX8632>";
@@ -1447,6 +2045,49 @@ void OperandX8632Mem::dump(const Cfg *Func) const {
   Str << "]";
 }
 
+x86::Address
+OperandX8632Mem::convertToAsmAddress(IceString &SymbolicOffset) const {
+  int32_t Disp = 0;
+  // Determine the offset (is it relocatable?)
+  if (getOffset()) {
+    if (ConstantInteger *CI = llvm::dyn_cast<ConstantInteger>(getOffset())) {
+      Disp = static_cast<int32_t>(CI->getValue());
+      // Check that it's really 32-bit.
+      // Sometimes Subzero stores negative values as sign-extended 64-bit
+      // and sometimes it's zero-extended to 64-bit.
+      // Should we just make a ConstantInteger32?
+      assert((CI->getValue() >> 32) == 0 ||
+             (CI->getValue() >> 32) == 0xffffffff);
+    } else if (ConstantRelocatable *CR =
+                   llvm::dyn_cast<ConstantRelocatable>(getOffset())) {
+      SymbolicOffset = CR->getName();
+      // TODO(jvoung): this isn't really tested yet, since none of the
+      // arithmetic, etc. ops have the addressing mode optimization
+      // for ConstantRelocatable.
+      assert(CR->getOffset() == 0);
+      // Check that the int64_t only really has 32-bits.
+      // Should we just the displacement 32-bit?
+      assert((CR->getOffset() >> 32) == 0 ||
+             (CR->getOffset() >> 32) == 0xffffffff);
+      Disp = CR->getOffset();
+    }
+  }
+
+  // Now convert to the various possible forms.
+  if (getBase() && getIndex()) {
+    return x86::Address(convertToAsmGPR(getBase()->getRegNum()),
+                        convertToAsmGPR(getIndex()->getRegNum()),
+                        x86::ScaleFactor(getShift()), Disp);
+  } else if (getBase()) {
+    return x86::Address(convertToAsmGPR(getBase()->getRegNum()), Disp);
+  } else if (getIndex()) {
+    return x86::Address(convertToAsmGPR(getIndex()->getRegNum()),
+                        x86::ScaleFactor(getShift()), Disp);
+  } else {
+    return x86::Address::Absolute(Disp);
+  }
+}
+
 void VariableSplit::emit(const Cfg *Func) const {
   Ostream &Str = Func->getContext()->getStrEmit();
   assert(Var->getLocalUseNode() == NULL ||