Subzero. Moves code around in preparations for 64-bit lowering.

Specifically, it moves

lowerArguments
lowerRet
addProlog
addEpilog

from the x86 lowering template to the concrete lowering implementations.

BUG= https://code.google.com/p/nativeclient/issues/detail?id=4077
R=stichnot@chromium.org

Review URL: https://codereview.chromium.org/1261383002.
diff --git a/src/IceTargetLoweringX8664.cpp b/src/IceTargetLoweringX8664.cpp
index f5d4ead..9056648 100644
--- a/src/IceTargetLoweringX8664.cpp
+++ b/src/IceTargetLoweringX8664.cpp
@@ -243,9 +243,8 @@
   if (Dest) {
     switch (Dest->getType()) {
     case IceType_NUM:
-      llvm_unreachable("Invalid Call dest type");
-      break;
     case IceType_void:
+      llvm::report_fatal_error("Invalid Call dest type");
       break;
     case IceType_i1:
     case IceType_i8:
@@ -339,36 +338,383 @@
   }
 }
 
-void TargetDataX8664::lowerJumpTables() {
-  switch (Ctx->getFlags().getOutFileType()) {
-  case FT_Elf: {
-    ELFObjectWriter *Writer = Ctx->getObjectWriter();
-    for (const JumpTableData &JumpTable : Ctx->getJumpTables())
-      // TODO(jpp): not 386.
-      Writer->writeJumpTable(JumpTable, llvm::ELF::R_386_32);
-  } break;
-  case FT_Asm:
-    // Already emitted from Cfg
-    break;
-  case FT_Iasm: {
-    if (!BuildDefs::dump())
-      return;
-    Ostream &Str = Ctx->getStrEmit();
-    for (const JumpTableData &JT : Ctx->getJumpTables()) {
-      Str << "\t.section\t.rodata." << JT.getFunctionName()
-          << "$jumptable,\"a\",@progbits\n";
-      Str << "\t.align\t" << typeWidthInBytes(getPointerType()) << "\n";
-      Str << InstJumpTable::makeName(JT.getFunctionName(), JT.getId()) << ":";
+void TargetX8664::lowerArguments() {
+  VarList &Args = Func->getArgs();
+  // The first eight vetcor typed arguments (as well as fp arguments) are passed
+  // in %xmm0 through %xmm7 regardless of their position in the argument list.
+  unsigned NumXmmArgs = 0;
+  // The first six integer typed arguments are passed in %rdi, %rsi, %rdx, %rcx,
+  // %r8, and %r9 regardless of their position in the argument list.
+  unsigned NumGprArgs = 0;
 
-      // On X8664 ILP32 pointers are 32-bit hence the use of .long
-      for (intptr_t TargetOffset : JT.getTargetOffsets())
-        Str << "\n\t.long\t" << JT.getFunctionName() << "+" << TargetOffset;
-      Str << "\n";
+  Context.init(Func->getEntryNode());
+  Context.setInsertPoint(Context.getCur());
+
+  for (SizeT i = 0, End = Args.size();
+       i < End && (NumXmmArgs < Traits::X86_MAX_XMM_ARGS ||
+                   NumGprArgs < Traits::X86_MAX_XMM_ARGS);
+       ++i) {
+    Variable *Arg = Args[i];
+    Type Ty = Arg->getType();
+    if ((isVectorType(Ty) || isScalarFloatingType(Ty)) &&
+        NumXmmArgs < Traits::X86_MAX_XMM_ARGS) {
+      // Replace Arg in the argument list with the home register.  Then
+      // generate an instruction in the prolog to copy the home register
+      // to the assigned location of Arg.
+      int32_t RegNum = getRegisterForXmmArgNum(NumXmmArgs);
+      ++NumXmmArgs;
+      Variable *RegisterArg = Func->makeVariable(Ty);
+      if (BuildDefs::dump())
+        RegisterArg->setName(Func, "home_reg:" + Arg->getName(Func));
+      RegisterArg->setRegNum(RegNum);
+      RegisterArg->setIsArg();
+      Arg->setIsArg(false);
+
+      Args[i] = RegisterArg;
+      Context.insert(InstAssign::create(Func, Arg, RegisterArg));
+    } else if (isScalarIntegerType(Ty) &&
+               NumGprArgs < Traits::X86_MAX_GPR_ARGS) {
+      int32_t RegNum = getRegisterForGprArgNum(NumGprArgs);
+      ++NumGprArgs;
+      Variable *RegisterArg = Func->makeVariable(Ty);
+      if (BuildDefs::dump())
+        RegisterArg->setName(Func, "home_reg:" + Arg->getName(Func));
+      RegisterArg->setRegNum(RegNum);
+      RegisterArg->setIsArg();
+      Arg->setIsArg(false);
+
+      Args[i] = RegisterArg;
+      Context.insert(InstAssign::create(Func, Arg, RegisterArg));
     }
-  } break;
   }
 }
 
+void TargetX8664::lowerRet(const InstRet *Inst) {
+  Variable *Reg = nullptr;
+  if (Inst->hasRetValue()) {
+    Operand *Src0 = legalize(Inst->getRetValue());
+    // TODO(jpp): this is not needed.
+    if (Src0->getType() == IceType_i64) {
+      Variable *eax =
+          legalizeToReg(loOperand(Src0), Traits::RegisterSet::Reg_eax);
+      Variable *edx =
+          legalizeToReg(hiOperand(Src0), Traits::RegisterSet::Reg_edx);
+      Reg = eax;
+      Context.insert(InstFakeUse::create(Func, edx));
+    } else if (isScalarFloatingType(Src0->getType())) {
+      _fld(Src0);
+    } else if (isVectorType(Src0->getType())) {
+      Reg = legalizeToReg(Src0, Traits::RegisterSet::Reg_xmm0);
+    } else {
+      _mov(Reg, Src0, Traits::RegisterSet::Reg_eax);
+    }
+  }
+  // Add a ret instruction even if sandboxing is enabled, because
+  // addEpilog explicitly looks for a ret instruction as a marker for
+  // where to insert the frame removal instructions.
+  _ret(Reg);
+  // Add a fake use of esp to make sure esp stays alive for the entire
+  // function.  Otherwise post-call esp adjustments get dead-code
+  // eliminated.  TODO: Are there more places where the fake use
+  // should be inserted?  E.g. "void f(int n){while(1) g(n);}" may not
+  // have a ret instruction.
+  Variable *esp =
+      Func->getTarget()->getPhysicalRegister(Traits::RegisterSet::Reg_esp);
+  Context.insert(InstFakeUse::create(Func, esp));
+}
+
+void TargetX8664::addProlog(CfgNode *Node) {
+  // Stack frame layout:
+  //
+  // +------------------------+
+  // | 1. return address      |
+  // +------------------------+
+  // | 2. preserved registers |
+  // +------------------------+
+  // | 3. padding             |
+  // +------------------------+
+  // | 4. global spill area   |
+  // +------------------------+
+  // | 5. padding             |
+  // +------------------------+
+  // | 6. local spill area    |
+  // +------------------------+
+  // | 7. padding             |
+  // +------------------------+
+  // | 8. allocas             |
+  // +------------------------+
+  //
+  // The following variables record the size in bytes of the given areas:
+  //  * X86_RET_IP_SIZE_BYTES:  area 1
+  //  * PreservedRegsSizeBytes: area 2
+  //  * SpillAreaPaddingBytes:  area 3
+  //  * GlobalsSize:            area 4
+  //  * GlobalsAndSubsequentPaddingSize: areas 4 - 5
+  //  * LocalsSpillAreaSize:    area 6
+  //  * SpillAreaSizeBytes:     areas 3 - 7
+
+  // Determine stack frame offsets for each Variable without a
+  // register assignment.  This can be done as one variable per stack
+  // slot.  Or, do coalescing by running the register allocator again
+  // with an infinite set of registers (as a side effect, this gives
+  // variables a second chance at physical register assignment).
+  //
+  // A middle ground approach is to leverage sparsity and allocate one
+  // block of space on the frame for globals (variables with
+  // multi-block lifetime), and one block to share for locals
+  // (single-block lifetime).
+
+  Context.init(Node);
+  Context.setInsertPoint(Context.getCur());
+
+  llvm::SmallBitVector CalleeSaves =
+      getRegisterSet(RegSet_CalleeSave, RegSet_None);
+  RegsUsed = llvm::SmallBitVector(CalleeSaves.size());
+  VarList SortedSpilledVariables, VariablesLinkedToSpillSlots;
+  size_t GlobalsSize = 0;
+  // If there is a separate locals area, this represents that area.
+  // Otherwise it counts any variable not counted by GlobalsSize.
+  SpillAreaSizeBytes = 0;
+  // If there is a separate locals area, this specifies the alignment
+  // for it.
+  uint32_t LocalsSlotsAlignmentBytes = 0;
+  // The entire spill locations area gets aligned to largest natural
+  // alignment of the variables that have a spill slot.
+  uint32_t SpillAreaAlignmentBytes = 0;
+  // A spill slot linked to a variable with a stack slot should reuse
+  // that stack slot.
+  std::function<bool(Variable *)> TargetVarHook =
+      [&VariablesLinkedToSpillSlots](Variable *Var) {
+        if (auto *SpillVar =
+                llvm::dyn_cast<typename Traits::SpillVariable>(Var)) {
+          assert(Var->getWeight().isZero());
+          if (SpillVar->getLinkedTo() && !SpillVar->getLinkedTo()->hasReg()) {
+            VariablesLinkedToSpillSlots.push_back(Var);
+            return true;
+          }
+        }
+        return false;
+      };
+
+  // Compute the list of spilled variables and bounds for GlobalsSize, etc.
+  getVarStackSlotParams(SortedSpilledVariables, RegsUsed, &GlobalsSize,
+                        &SpillAreaSizeBytes, &SpillAreaAlignmentBytes,
+                        &LocalsSlotsAlignmentBytes, TargetVarHook);
+  uint32_t LocalsSpillAreaSize = SpillAreaSizeBytes;
+  SpillAreaSizeBytes += GlobalsSize;
+
+  // Add push instructions for preserved registers.
+  uint32_t NumCallee = 0;
+  size_t PreservedRegsSizeBytes = 0;
+  for (SizeT i = 0; i < CalleeSaves.size(); ++i) {
+    if (CalleeSaves[i] && RegsUsed[i]) {
+      ++NumCallee;
+      PreservedRegsSizeBytes += typeWidthInBytes(IceType_i64);
+      _push(getPhysicalRegister(i));
+    }
+  }
+  Ctx->statsUpdateRegistersSaved(NumCallee);
+
+  // Generate "push ebp; mov ebp, esp"
+  if (IsEbpBasedFrame) {
+    assert((RegsUsed & getRegisterSet(RegSet_FramePointer, RegSet_None))
+               .count() == 0);
+    PreservedRegsSizeBytes += typeWidthInBytes(IceType_i64);
+    Variable *ebp = getPhysicalRegister(Traits::RegisterSet::Reg_ebp);
+    Variable *esp = getPhysicalRegister(Traits::RegisterSet::Reg_esp);
+    _push(ebp);
+    _mov(ebp, esp);
+    // Keep ebp live for late-stage liveness analysis
+    // (e.g. asm-verbose mode).
+    Context.insert(InstFakeUse::create(Func, ebp));
+  }
+
+  // Align the variables area. SpillAreaPaddingBytes is the size of
+  // the region after the preserved registers and before the spill areas.
+  // LocalsSlotsPaddingBytes is the amount of padding between the globals
+  // and locals area if they are separate.
+  assert(SpillAreaAlignmentBytes <= Traits::X86_STACK_ALIGNMENT_BYTES);
+  assert(LocalsSlotsAlignmentBytes <= SpillAreaAlignmentBytes);
+  uint32_t SpillAreaPaddingBytes = 0;
+  uint32_t LocalsSlotsPaddingBytes = 0;
+  alignStackSpillAreas(Traits::X86_RET_IP_SIZE_BYTES + PreservedRegsSizeBytes,
+                       SpillAreaAlignmentBytes, GlobalsSize,
+                       LocalsSlotsAlignmentBytes, &SpillAreaPaddingBytes,
+                       &LocalsSlotsPaddingBytes);
+  SpillAreaSizeBytes += SpillAreaPaddingBytes + LocalsSlotsPaddingBytes;
+  uint32_t GlobalsAndSubsequentPaddingSize =
+      GlobalsSize + LocalsSlotsPaddingBytes;
+
+  // Align esp if necessary.
+  if (NeedsStackAlignment) {
+    uint32_t StackOffset =
+        Traits::X86_RET_IP_SIZE_BYTES + PreservedRegsSizeBytes;
+    uint32_t StackSize =
+        Traits::applyStackAlignment(StackOffset + SpillAreaSizeBytes);
+    SpillAreaSizeBytes = StackSize - StackOffset;
+  }
+
+  // Generate "sub esp, SpillAreaSizeBytes"
+  if (SpillAreaSizeBytes)
+    _sub(getPhysicalRegister(Traits::RegisterSet::Reg_esp),
+         Ctx->getConstantInt32(SpillAreaSizeBytes));
+  Ctx->statsUpdateFrameBytes(SpillAreaSizeBytes);
+
+  resetStackAdjustment();
+
+  // Fill in stack offsets for stack args, and copy args into registers
+  // for those that were register-allocated.  Args are pushed right to
+  // left, so Arg[0] is closest to the stack/frame pointer.
+  Variable *FramePtr = getPhysicalRegister(getFrameOrStackReg());
+  size_t BasicFrameOffset =
+      PreservedRegsSizeBytes + Traits::X86_RET_IP_SIZE_BYTES;
+  if (!IsEbpBasedFrame)
+    BasicFrameOffset += SpillAreaSizeBytes;
+
+  const VarList &Args = Func->getArgs();
+  size_t InArgsSizeBytes = 0;
+  unsigned NumXmmArgs = 0;
+  unsigned NumGPRArgs = 0;
+  for (Variable *Arg : Args) {
+    // Skip arguments passed in registers.
+    if (isVectorType(Arg->getType()) && NumXmmArgs < Traits::X86_MAX_XMM_ARGS) {
+      ++NumXmmArgs;
+      continue;
+    }
+    if (isScalarFloatingType(Arg->getType()) &&
+        NumXmmArgs < Traits::X86_MAX_XMM_ARGS) {
+      ++NumXmmArgs;
+      continue;
+    }
+    if (isScalarIntegerType(Arg->getType()) &&
+        NumGPRArgs < Traits::X86_MAX_GPR_ARGS) {
+      ++NumGPRArgs;
+      continue;
+    }
+    finishArgumentLowering(Arg, FramePtr, BasicFrameOffset, InArgsSizeBytes);
+  }
+
+  // Fill in stack offsets for locals.
+  assignVarStackSlots(SortedSpilledVariables, SpillAreaPaddingBytes,
+                      SpillAreaSizeBytes, GlobalsAndSubsequentPaddingSize,
+                      IsEbpBasedFrame);
+  // Assign stack offsets to variables that have been linked to spilled
+  // variables.
+  for (Variable *Var : VariablesLinkedToSpillSlots) {
+    Variable *Linked =
+        (llvm::cast<typename Traits::SpillVariable>(Var))->getLinkedTo();
+    Var->setStackOffset(Linked->getStackOffset());
+  }
+  this->HasComputedFrame = true;
+
+  if (BuildDefs::dump() && Func->isVerbose(IceV_Frame)) {
+    OstreamLocker L(Func->getContext());
+    Ostream &Str = Func->getContext()->getStrDump();
+
+    Str << "Stack layout:\n";
+    uint32_t EspAdjustmentPaddingSize =
+        SpillAreaSizeBytes - LocalsSpillAreaSize -
+        GlobalsAndSubsequentPaddingSize - SpillAreaPaddingBytes;
+    Str << " in-args = " << InArgsSizeBytes << " bytes\n"
+        << " return address = " << Traits::X86_RET_IP_SIZE_BYTES << " bytes\n"
+        << " preserved registers = " << PreservedRegsSizeBytes << " bytes\n"
+        << " spill area padding = " << SpillAreaPaddingBytes << " bytes\n"
+        << " globals spill area = " << GlobalsSize << " bytes\n"
+        << " globals-locals spill areas intermediate padding = "
+        << GlobalsAndSubsequentPaddingSize - GlobalsSize << " bytes\n"
+        << " locals spill area = " << LocalsSpillAreaSize << " bytes\n"
+        << " esp alignment padding = " << EspAdjustmentPaddingSize
+        << " bytes\n";
+
+    Str << "Stack details:\n"
+        << " esp adjustment = " << SpillAreaSizeBytes << " bytes\n"
+        << " spill area alignment = " << SpillAreaAlignmentBytes << " bytes\n"
+        << " locals spill area alignment = " << LocalsSlotsAlignmentBytes
+        << " bytes\n"
+        << " is ebp based = " << IsEbpBasedFrame << "\n";
+  }
+}
+
+void TargetX8664::addEpilog(CfgNode *Node) {
+  InstList &Insts = Node->getInsts();
+  InstList::reverse_iterator RI, E;
+  for (RI = Insts.rbegin(), E = Insts.rend(); RI != E; ++RI) {
+    if (llvm::isa<typename Traits::Insts::Ret>(*RI))
+      break;
+  }
+  if (RI == E)
+    return;
+
+  // Convert the reverse_iterator position into its corresponding
+  // (forward) iterator position.
+  InstList::iterator InsertPoint = RI.base();
+  --InsertPoint;
+  Context.init(Node);
+  Context.setInsertPoint(InsertPoint);
+
+  Variable *esp = getPhysicalRegister(Traits::RegisterSet::Reg_esp);
+  if (IsEbpBasedFrame) {
+    Variable *ebp = getPhysicalRegister(Traits::RegisterSet::Reg_ebp);
+    // For late-stage liveness analysis (e.g. asm-verbose mode),
+    // adding a fake use of esp before the assignment of esp=ebp keeps
+    // previous esp adjustments from being dead-code eliminated.
+    Context.insert(InstFakeUse::create(Func, esp));
+    _mov(esp, ebp);
+    _pop(ebp);
+  } else {
+    // add esp, SpillAreaSizeBytes
+    if (SpillAreaSizeBytes)
+      _add(esp, Ctx->getConstantInt32(SpillAreaSizeBytes));
+  }
+
+  // Add pop instructions for preserved registers.
+  llvm::SmallBitVector CalleeSaves =
+      getRegisterSet(RegSet_CalleeSave, RegSet_None);
+  for (SizeT i = 0; i < CalleeSaves.size(); ++i) {
+    SizeT j = CalleeSaves.size() - i - 1;
+    if (j == Traits::RegisterSet::Reg_ebp && IsEbpBasedFrame)
+      continue;
+    if (CalleeSaves[j] && RegsUsed[j]) {
+      _pop(getPhysicalRegister(j));
+    }
+  }
+
+  if (!Ctx->getFlags().getUseSandboxing())
+    return;
+  // Change the original ret instruction into a sandboxed return sequence.
+  // t:ecx = pop
+  // bundle_lock
+  // and t, ~31
+  // jmp *t
+  // bundle_unlock
+  // FakeUse <original_ret_operand>
+  Variable *T_ecx = makeReg(IceType_i32, Traits::RegisterSet::Reg_ecx);
+  _pop(T_ecx);
+  lowerIndirectJump(T_ecx);
+  if (RI->getSrcSize()) {
+    Variable *RetValue = llvm::cast<Variable>(RI->getSrc(0));
+    Context.insert(InstFakeUse::create(Func, RetValue));
+  }
+  RI->setDeleted();
+}
+
+void TargetX8664::emitJumpTable(const Cfg *Func,
+                                const InstJumpTable *JumpTable) const {
+  if (!BuildDefs::dump())
+    return;
+  Ostream &Str = Ctx->getStrEmit();
+  IceString MangledName = Ctx->mangleName(Func->getFunctionName());
+  Str << "\t.section\t.rodata." << MangledName
+      << "$jumptable,\"a\",@progbits\n";
+  Str << "\t.align\t" << typeWidthInBytes(getPointerType()) << "\n";
+  Str << InstJumpTable::makeName(MangledName, JumpTable->getId()) << ":";
+
+  // On X8664 ILP32 pointers are 32-bit hence the use of .long
+  for (SizeT I = 0; I < JumpTable->getNumTargets(); ++I)
+    Str << "\n\t.long\t" << JumpTable->getTarget(I)->getAsmName();
+  Str << "\n";
+}
+
 namespace {
 template <typename T> struct PoolTypeConverter {};
 
@@ -507,21 +853,34 @@
   }
 }
 
-void TargetX8664::emitJumpTable(const Cfg *Func,
-                                const InstJumpTable *JumpTable) const {
-  if (!BuildDefs::dump())
-    return;
-  Ostream &Str = Ctx->getStrEmit();
-  IceString MangledName = Ctx->mangleName(Func->getFunctionName());
-  Str << "\t.section\t.rodata." << MangledName
-      << "$jumptable,\"a\",@progbits\n";
-  Str << "\t.align\t" << typeWidthInBytes(getPointerType()) << "\n";
-  Str << InstJumpTable::makeName(MangledName, JumpTable->getId()) << ":";
+void TargetDataX8664::lowerJumpTables() {
+  switch (Ctx->getFlags().getOutFileType()) {
+  case FT_Elf: {
+    ELFObjectWriter *Writer = Ctx->getObjectWriter();
+    for (const JumpTableData &JumpTable : Ctx->getJumpTables())
+      // TODO(jpp): not 386.
+      Writer->writeJumpTable(JumpTable, llvm::ELF::R_386_32);
+  } break;
+  case FT_Asm:
+    // Already emitted from Cfg
+    break;
+  case FT_Iasm: {
+    if (!BuildDefs::dump())
+      return;
+    Ostream &Str = Ctx->getStrEmit();
+    for (const JumpTableData &JT : Ctx->getJumpTables()) {
+      Str << "\t.section\t.rodata." << JT.getFunctionName()
+          << "$jumptable,\"a\",@progbits\n";
+      Str << "\t.align\t" << typeWidthInBytes(getPointerType()) << "\n";
+      Str << InstJumpTable::makeName(JT.getFunctionName(), JT.getId()) << ":";
 
-  // On X8664 ILP32 pointers are 32-bit hence the use of .long
-  for (SizeT I = 0; I < JumpTable->getNumTargets(); ++I)
-    Str << "\n\t.long\t" << JumpTable->getTarget(I)->getAsmName();
-  Str << "\n";
+      // On X8664 ILP32 pointers are 32-bit hence the use of .long
+      for (intptr_t TargetOffset : JT.getTargetOffsets())
+        Str << "\n\t.long\t" << JT.getFunctionName() << "+" << TargetOffset;
+      Str << "\n";
+    }
+  } break;
+  }
 }
 
 void TargetDataX8664::lowerGlobals(const VariableDeclarationList &Vars,