Subzero, MIPS32: Handling floating point instructions fadd, fsub, fmul, fdiv

src/IceTargetLoweringMIPS32.cpp

Index: src/IceTargetLoweringMIPS32.cpp
diff --git a/src/IceTargetLoweringMIPS32.cpp b/src/IceTargetLoweringMIPS32.cpp
index 70ddef531f88350f82bf5fc12e8d6b01cdca4ba9..269ecdb7a4f02d9ca4301f813a067093287ca5e1 100644
--- a/src/IceTargetLoweringMIPS32.cpp
+++ b/src/IceTargetLoweringMIPS32.cpp
@@ -64,6 +64,14 @@ namespace {
 // The maximum number of arguments to pass in GPR registers.
 constexpr uint32_t MIPS32_MAX_GPR_ARG = 4;
 
+std::array<RegNumT, MIPS32_MAX_GPR_ARG> GPRArgInitializer;
+std::array<RegNumT, MIPS32_MAX_GPR_ARG / 2> I64ArgInitializer;
+
+constexpr uint32_t MIPS32_MAX_FP_ARG = 2;
+
+std::array<RegNumT, MIPS32_MAX_FP_ARG> FP32ArgInitializer;
+std::array<RegNumT, MIPS32_MAX_FP_ARG> FP64ArgInitializer;
+
 const char *getRegClassName(RegClass C) {
   auto ClassNum = static_cast<RegClassMIPS32>(C);
   assert(ClassNum < RCMIPS32_NUM);
@@ -75,6 +83,18 @@ const char *getRegClassName(RegClass C) {
   }
 }
 
+// Stack alignment
+constexpr uint32_t MIPS32_STACK_ALIGNMENT_BYTES = 8;
+
+// Value is in bytes. Return Value adjusted to the next highest multiple of the
+// stack alignment required for the given type.
+uint32_t applyStackAlignmentTy(uint32_t Value, Type Ty) {
+  size_t typeAlignInBytes = typeWidthInBytes(Ty);
+  if (isVectorType(Ty))
+    UnimplementedError(getFlags());
+  return Utils::applyAlignment(Value, typeAlignInBytes);
+}
+
 } // end of anonymous namespace
 
 TargetMIPS32::TargetMIPS32(Cfg *Func) : TargetLowering(Func) {}
@@ -105,6 +125,20 @@ void TargetMIPS32::staticInit(GlobalContext *Ctx) {
   assert(RegisterAliases[RegMIPS32::val][RegMIPS32::val]);
   REGMIPS32_TABLE;
 #undef X
+
+  // TODO(mohit.bhakkad): Change these inits once we provide argument related
+  // field in register tables
+  for (size_t i = 0; i < MIPS32_MAX_GPR_ARG; i++)
+    GPRArgInitializer[i] = RegNumT::fixme(RegMIPS32::Reg_A0 + i);
+
+  for (size_t i = 0; i < MIPS32_MAX_GPR_ARG / 2; i++)
+    I64ArgInitializer[i] = RegNumT::fixme(RegMIPS32::Reg_A0A1 + i);
+
+  for (size_t i = 0; i < MIPS32_MAX_FP_ARG; i++) {
+    FP32ArgInitializer[i] = RegNumT::fixme(RegMIPS32::Reg_F12 + i * 2);
+    FP64ArgInitializer[i] = RegNumT::fixme(RegMIPS32::Reg_F12F13 + i);
+  }
+
   TypeToRegisterSet[IceType_void] = InvalidRegisters;
   TypeToRegisterSet[IceType_i1] = IntegerRegisters;
   TypeToRegisterSet[IceType_i8] = IntegerRegisters;
@@ -129,6 +163,24 @@ void TargetMIPS32::staticInit(GlobalContext *Ctx) {
                           RegMIPS32::getRegName, getRegClassName);
 }
 
+void TargetMIPS32::findMaxStackOutArgsSize() {
+  // MinNeededOutArgsBytes should be updated if the Target ever creates a
+  // high-level InstCall that requires more stack bytes.
+  constexpr size_t MinNeededOutArgsBytes = 16;
+  MaxOutArgsSizeBytes = MinNeededOutArgsBytes;
+  for (CfgNode *Node : Func->getNodes()) {
+    Context.init(Node);
+    while (!Context.atEnd()) {
+      PostIncrLoweringContext PostIncrement(Context);
+      Inst *CurInstr = iteratorToInst(Context.getCur());
+      if (auto *Call = llvm::dyn_cast<InstCall>(CurInstr)) {
+        SizeT OutArgsSizeBytes = getCallStackArgumentsSizeBytes(Call);
+        MaxOutArgsSizeBytes = std::max(MaxOutArgsSizeBytes, OutArgsSizeBytes);
+      }
+    }
+  }
+}
+
 void TargetMIPS32::translateO2() {
   TimerMarker T(TimerStack::TT_O2, Func);
 
@@ -136,6 +188,8 @@ void TargetMIPS32::translateO2() {
   // https://code.google.com/p/nativeclient/issues/detail?id=4094
   genTargetHelperCalls();
 
+  findMaxStackOutArgsSize();
+
   // Merge Alloca instructions, and lay out the stack.
   static constexpr bool SortAndCombineAllocas = false;
   Func->processAllocas(SortAndCombineAllocas);
@@ -237,6 +291,8 @@ void TargetMIPS32::translateOm1() {
   // TODO: share passes with X86?
   genTargetHelperCalls();
 
+  findMaxStackOutArgsSize();
+
   // Do not merge Alloca instructions, and lay out the stack.
   static constexpr bool SortAndCombineAllocas = false;
   Func->processAllocas(SortAndCombineAllocas);
@@ -403,88 +459,450 @@ void TargetMIPS32::emitVariable(const Variable *Var) const {
     Str << "($" << getRegName(getFrameOrStackReg(), FrameSPTy);
     Str << ")";
   }
+  // UnimplementedError(getFlags());

[Jim Stichnoth, 2016/06/13 20:14:19]

remove this

+}
+
+TargetMIPS32::CallingConv::CallingConv()
+    : GPRegsUsed(RegMIPS32::Reg_NUM),
+      GPRArgs(GPRArgInitializer.rbegin(), GPRArgInitializer.rend()),
+      I64Args(I64ArgInitializer.rbegin(), I64ArgInitializer.rend()),
+      VFPRegsUsed(RegMIPS32::Reg_NUM),
+      FP32Args(FP32ArgInitializer.rbegin(), FP32ArgInitializer.rend()),
+      FP64Args(FP64ArgInitializer.rbegin(), FP64ArgInitializer.rend()) {}
+
+// In MIPS O32 abi FP argument registers can be used only if first argument is
+// of type float/double. UseFPRegs flag is used to care of that. Also FP arg
+// registers can be used only for first 2 arguments, so we require argument
+// number to make register allocation decisions.
+bool TargetMIPS32::CallingConv::argInReg(Type Ty, uint32_t ArgNo,
+                                         RegNumT *Reg) {
+  if (isScalarIntegerType(Ty))
+    return argInGPR(Ty, Reg);
+  if (isScalarFloatingType(Ty)) {
+    if (ArgNo == 0) {
+      UseFPRegs = true;
+      return argInVFP(Ty, Reg);
+    }
+    if (UseFPRegs && ArgNo == 1) {
+      UseFPRegs = false;
+      return argInVFP(Ty, Reg);
+    }
+    return argInGPR(Ty, Reg);
+  }
   UnimplementedError(getFlags());
+  return false;
+}
+
+bool TargetMIPS32::CallingConv::argInGPR(Type Ty, RegNumT *Reg) {
+  CfgVector<RegNumT> *Source;
+
+  switch (Ty) {
+  default: {
+    UnimplementedError(getFlags());
+    return false;
+  } break;
+  case IceType_i32:
+  case IceType_f32: {
+    Source = &GPRArgs;
+  } break;
+  case IceType_i64:
+  case IceType_f64: {
+    Source = &I64Args;
+  } break;
+  }
+
+  discardUnavailableGPRsAndTheirAliases(Source);
+
+  if (Source->empty()) {
+    GPRegsUsed.set();
+    return false;
+  }
+
+  *Reg = Source->back();
+  // Note that we don't Source->pop_back() here. This is intentional. Notice how
+  // we mark all of Reg's aliases as Used. So, for the next argument,
+  // Source->back() is marked as unavailable, and it is thus implicitly popped
+  // from the stack.
+  GPRegsUsed |= RegisterAliases[*Reg];
+  return true;
+}
+
+inline void TargetMIPS32::CallingConv::discardNextGPRAndItsAliases(
+    CfgVector<RegNumT> *Regs) {
+  GPRegsUsed |= RegisterAliases[Regs->back()];
+  Regs->pop_back();
+}
+
+// GPR are not packed when passing parameters. Thus, a function foo(i32, i64,
+// i32) will have the first argument in a0, the second in a2-a3, and the third
+// on the stack. To model this behavior, whenever we pop a register from Regs,
+// we remove all of its aliases from the pool of available GPRs. This has the
+// effect of computing the "closure" on the GPR registers.
+void TargetMIPS32::CallingConv::discardUnavailableGPRsAndTheirAliases(
+    CfgVector<RegNumT> *Regs) {
+  while (!Regs->empty() && GPRegsUsed[Regs->back()]) {
+    discardNextGPRAndItsAliases(Regs);
+  }
+}
+
+bool TargetMIPS32::CallingConv::argInVFP(Type Ty, RegNumT *Reg) {
+  CfgVector<RegNumT> *Source;
+
+  switch (Ty) {
+  default: {
+    UnimplementedError(getFlags());
+    return false;
+  } break;
+  case IceType_f32: {
+    Source = &FP32Args;
+  } break;
+  case IceType_f64: {
+    Source = &FP64Args;
+  } break;
+  }
+
+  discardUnavailableVFPRegsAndTheirAliases(Source);
+
+  if (Source->empty()) {
+    VFPRegsUsed.set();
+    return false;
+  }
+
+  *Reg = Source->back();
+  VFPRegsUsed |= RegisterAliases[*Reg];
+
+  // In MIPS O32 abi if fun arguments are (f32, i32) then one can not use reg_a0
+  // for second argument even though it's free. f32 arg goes in reg_f12, i32 arg
+  // goes in reg_a1. Similarly if arguments are (f64, i32) second argument goes
+  // in reg_a3 and a0, a1 are not used.
+  Source = &GPRArgs;
+  // Discard one GPR reg for f32(4 bytes), two for f64(4 + 4 bytes)
+  discardNextGPRAndItsAliases(Source);
+  if (Ty == IceType_f64)
+    discardNextGPRAndItsAliases(Source);
+
+  return true;
+}
+
+void TargetMIPS32::CallingConv::discardUnavailableVFPRegsAndTheirAliases(
+    CfgVector<RegNumT> *Regs) {
+  while (!Regs->empty() && VFPRegsUsed[Regs->back()]) {
+    Regs->pop_back();
+  }
 }
 
 void TargetMIPS32::lowerArguments() {
   VarList &Args = Func->getArgs();
-  // We are only handling integer registers for now. The Mips o32 ABI is
-  // somewhat complex but will be implemented in its totality through follow
-  // on patches.
-  //
-  unsigned NumGPRRegsUsed = 0;
-  // For each register argument, 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.
+  TargetMIPS32::CallingConv CC;
+
+  // For each register argument, 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.
   Context.init(Func->getEntryNode());
   Context.setInsertPoint(Context.getCur());
+
   for (SizeT I = 0, E = Args.size(); I < E; ++I) {
     Variable *Arg = Args[I];
     Type Ty = Arg->getType();
-    // TODO(rkotler): handle float/vector types.
-    if (isVectorType(Ty)) {
-      UnimplementedError(getFlags());
+    RegNumT RegNum;
+    if (!CC.argInReg(Ty, I, &RegNum)) {
       continue;
     }
-    if (isFloatingType(Ty)) {
-      UnimplementedError(getFlags());
-      continue;
+    Variable *RegisterArg = Func->makeVariable(Ty);
+    if (BuildDefs::dump()) {
+      RegisterArg->setName(Func, "home_reg:" + Arg->getName());
     }
-    if (Ty == IceType_i64) {
-      if (NumGPRRegsUsed >= MIPS32_MAX_GPR_ARG)
-        continue;
-      auto RegLo = RegNumT::fixme(RegMIPS32::Reg_A0 + NumGPRRegsUsed);
-      auto RegHi = RegNumT::fixme(RegLo + 1);
-      ++NumGPRRegsUsed;
-      // Always start i64 registers at an even register, so this may end
-      // up padding away a register.
-      if (RegLo % 2 != 0) {
-        RegLo = RegNumT::fixme(RegLo + 1);
-        ++NumGPRRegsUsed;
-      }
-      // If this leaves us without room to consume another register,
-      // leave any previously speculatively consumed registers as consumed.
-      if (NumGPRRegsUsed >= MIPS32_MAX_GPR_ARG)
-        continue;
-      // RegHi = RegNumT::fixme(RegMIPS32::Reg_A0 + NumGPRRegsUsed);
-      ++NumGPRRegsUsed;
-      Variable *RegisterArg = Func->makeVariable(Ty);
-      auto *RegisterArg64On32 = llvm::cast<Variable64On32>(RegisterArg);
-      if (BuildDefs::dump())
-        RegisterArg64On32->setName(Func, "home_reg:" + Arg->getName());
-      RegisterArg64On32->initHiLo(Func);
-      RegisterArg64On32->setIsArg();
-      RegisterArg64On32->getLo()->setRegNum(RegLo);
-      RegisterArg64On32->getHi()->setRegNum(RegHi);
-      Arg->setIsArg(false);
-      Args[I] = RegisterArg64On32;
-      Context.insert<InstAssign>(Arg, RegisterArg);
-      continue;
-    } else {
-      assert(Ty == IceType_i32);
-      if (NumGPRRegsUsed >= MIPS32_MAX_GPR_ARG)
-        continue;
-      const auto RegNum = RegNumT::fixme(RegMIPS32::Reg_A0 + NumGPRRegsUsed);
-      ++NumGPRRegsUsed;
-      Variable *RegisterArg = Func->makeVariable(Ty);
-      if (BuildDefs::dump()) {
-        RegisterArg->setName(Func, "home_reg:" + Arg->getName());
-      }
-      RegisterArg->setRegNum(RegNum);
-      RegisterArg->setIsArg();
-      Arg->setIsArg(false);
-      Args[I] = RegisterArg;
-      Context.insert<InstAssign>(Arg, RegisterArg);
+    RegisterArg->setIsArg();
+    Arg->setIsArg(false);
+    Args[I] = RegisterArg;
+    switch (Ty) {
+    default: { RegisterArg->setRegNum(RegNum); } break;
+    case IceType_i64: {
+      auto *RegisterArg64 = llvm::cast<Variable64On32>(RegisterArg);
+      RegisterArg64->initHiLo(Func);
+      RegisterArg64->getLo()->setRegNum(
+          RegNumT::fixme(RegMIPS32::getI64PairFirstGPRNum(RegNum)));
+      RegisterArg64->getHi()->setRegNum(
+          RegNumT::fixme(RegMIPS32::getI64PairSecondGPRNum(RegNum)));
+    } break;
     }
+    Context.insert<InstAssign>(Arg, RegisterArg);
   }
 }
 
 Type TargetMIPS32::stackSlotType() { return IceType_i32; }
 
+// Helper function for addProlog().
+//
+// This assumes Arg is an argument passed on the stack. This sets the frame
+// offset for Arg and updates InArgsSizeBytes according to Arg's width. For an
+// I64 arg that has been split into Lo and Hi components, it calls itself
+// recursively on the components, taking care to handle Lo first because of the
+// little-endian architecture. Lastly, this function generates an instruction
+// to copy Arg into its assigned register if applicable.
+void TargetMIPS32::finishArgumentLowering(Variable *Arg, Variable *FramePtr,
+                                          size_t BasicFrameOffset,
+                                          size_t *InArgsSizeBytes) {
+  const Type Ty = Arg->getType();
+  *InArgsSizeBytes = applyStackAlignmentTy(*InArgsSizeBytes, Ty);
+
+  if (auto *Arg64On32 = llvm::dyn_cast<Variable64On32>(Arg)) {
+    Variable *const Lo = Arg64On32->getLo();
+    Variable *const Hi = Arg64On32->getHi();
+    finishArgumentLowering(Lo, FramePtr, BasicFrameOffset, InArgsSizeBytes);
+    finishArgumentLowering(Hi, FramePtr, BasicFrameOffset, InArgsSizeBytes);
+    return;
+  }
+  assert(Ty != IceType_i64);
+
+  const int32_t ArgStackOffset = BasicFrameOffset + *InArgsSizeBytes;
+  *InArgsSizeBytes += typeWidthInBytesOnStack(Ty);
+
+  if (!Arg->hasReg()) {
+    Arg->setStackOffset(ArgStackOffset);
+    return;
+  }
+
+  // If the argument variable has been assigned a register, we need to copy the
+  // value from the stack slot.
+  Variable *Parameter = Func->makeVariable(Ty);
+  Parameter->setMustNotHaveReg();
+  Parameter->setStackOffset(ArgStackOffset);
+  _mov(Arg, Parameter);
+}
+
 void TargetMIPS32::addProlog(CfgNode *Node) {
-  (void)Node;
+  // Stack frame layout:
+  //
+  // +------------------------+
+  // | 1. preserved registers |
+  // +------------------------+
+  // | 2. padding             |
+  // +------------------------+
+  // | 3. global spill area   |
+  // +------------------------+
+  // | 4. padding             |
+  // +------------------------+
+  // | 5. local spill area    |
+  // +------------------------+
+  // | 6. padding             |
+  // +------------------------+
+  // | 7. allocas             |
+  // +------------------------+
+  // | 8. padding             |
+  // +------------------------+
+  // | 9. out args           |
+  // +------------------------+ <--- StackPointer
+  //
+  // The following variables record the size in bytes of the given areas:
+  //  * PreservedRegsSizeBytes: area 1
+  //  * SpillAreaPaddingBytes:  area 2
+  //  * GlobalsSize:            area 3
+  //  * GlobalsAndSubsequentPaddingSize: areas 3 - 4
+  //  * LocalsSpillAreaSize:    area 5
+  //  * SpillAreaSizeBytes:     areas 2 - 9
+  //  * maxOutArgsSizeBytes():  area 9
+
+  Context.init(Node);
+  Context.setInsertPoint(Context.getCur());
+
+  SmallBitVector CalleeSaves = getRegisterSet(RegSet_CalleeSave, RegSet_None);
+  RegsUsed = SmallBitVector(CalleeSaves.size());
+
+  VarList SortedSpilledVariables;
+
+  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;
+  // For now, we don't have target-specific variables that need special
+  // treatment (no stack-slot-linked SpillVariable type).
+  std::function<bool(Variable *)> TargetVarHook = [](Variable *Var) {
+    static constexpr bool AssignStackSlot = false;
+    static constexpr bool DontAssignStackSlot = !AssignStackSlot;
+    if (llvm::isa<Variable64On32>(Var)) {
+      return DontAssignStackSlot;
+    }
+    return AssignStackSlot;
+  };
+
+  // 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;
+
+  PreservedGPRs.reserve(CalleeSaves.size());
+
+  // Consider FP and RA as callee-save / used as needed.
+  if (UsesFramePointer) {
+    if (RegsUsed[RegMIPS32::Reg_FP]) {
+      llvm::report_fatal_error("Frame pointer has been used.");
+    }
+    CalleeSaves[RegMIPS32::Reg_FP] = true;
+    RegsUsed[RegMIPS32::Reg_FP] = true;
+  }
+  if (!MaybeLeafFunc) {
+    CalleeSaves[RegMIPS32::Reg_RA] = true;
+    RegsUsed[RegMIPS32::Reg_RA] = true;
+  }
+
+  // Make two passes over the used registers. The first pass records all the
+  // used registers -- and their aliases. Then, we figure out which GPR
+  // registers should be saved.
+  SmallBitVector ToPreserve(RegMIPS32::Reg_NUM);
+  for (SizeT i = 0; i < CalleeSaves.size(); ++i) {
+    if (CalleeSaves[i] && RegsUsed[i]) {
+      ToPreserve |= RegisterAliases[i];
+    }
+  }
+
+  uint32_t NumCallee = 0;
+  size_t PreservedRegsSizeBytes = 0;
+
+  // RegClasses is a tuple of
+  //
+  // <First Register in Class, Last Register in Class, Vector of Save Registers>
+  //
+  // We use this tuple to figure out which register we should save/restore
+  // during
+  // prolog/epilog.
+  using RegClassType = std::tuple<uint32_t, uint32_t, VarList *>;
+  const RegClassType RegClass = RegClassType(
+      RegMIPS32::Reg_GPR_First, RegMIPS32::Reg_GPR_Last, &PreservedGPRs);
+  const uint32_t FirstRegInClass = std::get<0>(RegClass);
+  const uint32_t LastRegInClass = std::get<1>(RegClass);
+  VarList *const PreservedRegsInClass = std::get<2>(RegClass);
+  for (uint32_t Reg = LastRegInClass; Reg > FirstRegInClass; Reg--) {
+    if (!ToPreserve[Reg]) {
+      continue;
+    }
+    ++NumCallee;
+    Variable *PhysicalRegister = getPhysicalRegister(RegNumT::fromInt(Reg));
+    PreservedRegsSizeBytes +=
+        typeWidthInBytesOnStack(PhysicalRegister->getType());
+    PreservedRegsInClass->push_back(PhysicalRegister);
+  }
+
+  Ctx->statsUpdateRegistersSaved(NumCallee);
+
+  // 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 <= MIPS32_STACK_ALIGNMENT_BYTES);
+  (void)MIPS32_STACK_ALIGNMENT_BYTES;
+  assert(LocalsSlotsAlignmentBytes <= SpillAreaAlignmentBytes);
+  uint32_t SpillAreaPaddingBytes = 0;
+  uint32_t LocalsSlotsPaddingBytes = 0;
+  alignStackSpillAreas(PreservedRegsSizeBytes, SpillAreaAlignmentBytes,
+                       GlobalsSize, LocalsSlotsAlignmentBytes,
+                       &SpillAreaPaddingBytes, &LocalsSlotsPaddingBytes);
+  SpillAreaSizeBytes += SpillAreaPaddingBytes + LocalsSlotsPaddingBytes;
+  uint32_t GlobalsAndSubsequentPaddingSize =
+      GlobalsSize + LocalsSlotsPaddingBytes;
+
+  if (MaybeLeafFunc)
+    MaxOutArgsSizeBytes = 0;
+
+  // Adds the out args space to the stack, and align SP if necessary.
+  uint32_t TotalStackSizeBytes = PreservedRegsSizeBytes + SpillAreaSizeBytes;
+
+  // TODO(sagar.thakur): Combine fixed alloca and maximum out argument size with
+  // TotalStackSizeBytes once lowerAlloca is implemented and leaf function
+  // information is generated by lowerCall.
+
+  // Generate "addiu sp, sp, -TotalStackSizeBytes"
+  if (TotalStackSizeBytes) {
+    // Use the scratch register if needed to legalize the immediate.
+    Variable *SP = getPhysicalRegister(RegMIPS32::Reg_SP);
+    _addiu(SP, SP, -(TotalStackSizeBytes));
+  }
+
+  Ctx->statsUpdateFrameBytes(TotalStackSizeBytes);
+
+  if (!PreservedGPRs.empty()) {
+    uint32_t StackOffset = TotalStackSizeBytes;
+    for (Variable *Var : *PreservedRegsInClass) {
+      Variable *PhysicalRegister = getPhysicalRegister(Var->getRegNum());
+      StackOffset -= typeWidthInBytesOnStack(PhysicalRegister->getType());
+      Variable *SP = getPhysicalRegister(RegMIPS32::Reg_SP);
+      OperandMIPS32Mem *MemoryLocation = OperandMIPS32Mem::create(
+          Func, IceType_i32, SP,
+          llvm::cast<ConstantInteger32>(Ctx->getConstantInt32(StackOffset)));
+      _sw(PhysicalRegister, MemoryLocation);
+    }
+  }
+
+  Variable *FP = getPhysicalRegister(RegMIPS32::Reg_FP);
+
+  // Generate "mov FP, SP" if needed.
+  if (UsesFramePointer) {
+    Variable *SP = getPhysicalRegister(RegMIPS32::Reg_SP);
+    _mov(FP, SP);
+    // Keep FP live for late-stage liveness analysis (e.g. asm-verbose mode).
+    Context.insert<InstFakeUse>(FP);
+  }
+
+  // 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.
+  const VarList &Args = Func->getArgs();
+  size_t InArgsSizeBytes = 0;
+  TargetMIPS32::CallingConv CC;
+  uint32_t ArgNo = 0;
+
+  for (Variable *Arg : Args) {
+    RegNumT DummyReg;
+    const Type Ty = Arg->getType();
+    // Skip arguments passed in registers.
+    if (CC.argInReg(Ty, ArgNo, &DummyReg)) {
+      ArgNo++;
+      continue;
+    } else {
+      finishArgumentLowering(Arg, FP, TotalStackSizeBytes, &InArgsSizeBytes);
+    }
+  }
+
+  // Fill in stack offsets for locals.
+  assignVarStackSlots(SortedSpilledVariables, SpillAreaPaddingBytes,
+                      SpillAreaSizeBytes, GlobalsAndSubsequentPaddingSize,
+                      UsesFramePointer);
+  this->HasComputedFrame = true;
+
+  if (BuildDefs::dump() && Func->isVerbose(IceV_Frame)) {
+    OstreamLocker _(Func->getContext());
+    Ostream &Str = Func->getContext()->getStrDump();
+
+    Str << "Stack layout:\n";
+    uint32_t SPAdjustmentPaddingSize =
+        SpillAreaSizeBytes - LocalsSpillAreaSize -
+        GlobalsAndSubsequentPaddingSize - SpillAreaPaddingBytes -
+        MaxOutArgsSizeBytes;
+    Str << " in-args = " << InArgsSizeBytes << " 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"
+        << " SP alignment padding = " << SPAdjustmentPaddingSize << " bytes\n";
+
+    Str << "Stack details:\n"
+        << " SP adjustment = " << SpillAreaSizeBytes << " bytes\n"
+        << " spill area alignment = " << SpillAreaAlignmentBytes << " bytes\n"
+        << " outgoing args size = " << MaxOutArgsSizeBytes << " bytes\n"
+        << " locals spill area alignment = " << LocalsSlotsAlignmentBytes
+        << " bytes\n"
+        << " is FP based = " << 1 << "\n";
+  }
   return;
-  UnimplementedError(getFlags());
 }
 
 void TargetMIPS32::addEpilog(CfgNode *Node) {
@@ -707,17 +1125,6 @@ void TargetMIPS32::lowerArithmetic(const InstArithmetic *Instr) {
     UnimplementedLoweringError(this, Instr);
     return;
   }
-  switch (Instr->getOp()) {
-  default:
-    break;
-  case InstArithmetic::Fadd:
-  case InstArithmetic::Fsub:
-  case InstArithmetic::Fmul:
-  case InstArithmetic::Fdiv:
-  case InstArithmetic::Frem:
-    UnimplementedLoweringError(this, Instr);
-    return;
-  }
 
   // At this point Dest->getType() is non-i64 scalar
 
@@ -796,13 +1203,54 @@ void TargetMIPS32::lowerArithmetic(const InstArithmetic *Instr) {
     _mov(Dest, T);
     return;
   }
-  case InstArithmetic::Fadd:
+  case InstArithmetic::Fadd: {
+    if (DestTy == IceType_f32) {
+      _add_s(T, Src0R, Src1R);
+      _mov_s(Dest, T);

[Jim Stichnoth, 2016/06/13 20:14:19]

From this, I am seeing invalid asm code being generated:

<stdin>:9:5: error: invalid operand for instruction
        sw      $f12, 8($sp)
                ^

I think maybe swc1 should be emitted instead?

+      return;
+    }
+    if (DestTy == IceType_f64) {
+      _add_d(T, Src0R, Src1R);
+      _mov_d(Dest, T);
+      return;
+    }
     break;
+  }
   case InstArithmetic::Fsub:
+    if (DestTy == IceType_f32) {
+      _sub_s(T, Src0R, Src1R);
+      _mov_s(Dest, T);
+      return;
+    }
+    if (DestTy == IceType_f64) {
+      _sub_d(T, Src0R, Src1R);
+      _mov_d(Dest, T);
+      return;
+    }
     break;
   case InstArithmetic::Fmul:
+    if (DestTy == IceType_f32) {
+      _mul_s(T, Src0R, Src1R);
+      _mov_s(Dest, T);
+      return;
+    }
+    if (DestTy == IceType_f64) {
+      _mul_d(T, Src0R, Src1R);
+      _mov_d(Dest, T);
+      return;
+    }
     break;
   case InstArithmetic::Fdiv:
+    if (DestTy == IceType_f32) {
+      _div_s(T, Src0R, Src1R);
+      _mov_s(Dest, T);
+      return;
+    }
+    if (DestTy == IceType_f64) {
+      _div_d(T, Src0R, Src1R);
+      _mov_d(Dest, T);
+      return;
+    }
     break;
   case InstArithmetic::Frem:

[Jim Stichnoth, 2016/06/13 20:14:19]

For now, avoid a liveness validation error by adding FakeUses of the
temporaries:

    Context.insert<InstFakeUse>(Src0R);
    Context.insert<InstFakeUse>(Src1R);

The Om1 register allocator doesn't like it when infinite-weight temporaries have
a definition but no uses.  (Technically that's not really an error, but it may
reflect what may be "sloppy" lowering so we decided to retain that error
reporting.)

     break;
@@ -934,10 +1382,10 @@ void TargetMIPS32::lowerBr(const InstBr *Instr) {
 
 void TargetMIPS32::lowerCall(const InstCall *Instr) {
   // TODO(rkotler): assign arguments to registers and stack. Also reserve stack.
-  if (Instr->getNumArgs()) {
-    UnimplementedLoweringError(this, Instr);
-    return;
-  }
+  // if (Instr->getNumArgs()) {

[Jim Stichnoth, 2016/06/13 20:14:19]

remove this

+  //  UnimplementedLoweringError(this, Instr);
+  //  return;
+  //}
   // Generate the call instruction.  Assign its result to a temporary with high
   // register allocation weight.
   Variable *Dest = Instr->getDest();
@@ -962,19 +1410,22 @@ void TargetMIPS32::lowerCall(const InstCall *Instr) {
       ReturnRegHi = I32Reg(RegMIPS32::Reg_V1);
       break;
     case IceType_f32:
+      ReturnReg = makeReg(Dest->getType(), RegMIPS32::Reg_F0);
+      break;
     case IceType_f64:
-      UnimplementedLoweringError(this, Instr);
-      return;
+      ReturnReg = makeReg(Dest->getType(), RegMIPS32::Reg_F0F1);
+      break;
     case IceType_v4i1:
     case IceType_v8i1:
     case IceType_v16i1:
     case IceType_v16i8:
     case IceType_v8i16:
     case IceType_v4i32:
-    case IceType_v4f32:
+    case IceType_v4f32: {
       UnimplementedLoweringError(this, Instr);
       return;
     }
+    }
   }
   Operand *CallTarget = Instr->getCallTarget();
   // Allow ConstantRelocatable to be left alone as a direct call,
@@ -998,23 +1449,43 @@ void TargetMIPS32::lowerCall(const InstCall *Instr) {
 
   // Assign the result of the call to Dest.
   if (ReturnReg) {
-    if (ReturnRegHi) {
-      assert(Dest->getType() == IceType_i64);
+    auto *Zero = getZero();
+    switch (Dest->getType()) {
+    default:
+      UnimplementedLoweringError(this, Instr);
+    case IceType_i1:
+    case IceType_i8:
+    case IceType_i16:
+    case IceType_i32: {
+      _addu(Dest, Zero, ReturnReg);
+      break;
+    }
+    case IceType_i64: {
       auto *Dest64On32 = llvm::cast<Variable64On32>(Dest);
       Variable *DestLo = Dest64On32->getLo();
       Variable *DestHi = Dest64On32->getHi();
-      _mov(DestLo, ReturnReg);
-      _mov(DestHi, ReturnRegHi);
-    } else {
-      assert(Dest->getType() == IceType_i32 || Dest->getType() == IceType_i16 ||
-             Dest->getType() == IceType_i8 || Dest->getType() == IceType_i1 ||
-             isVectorType(Dest->getType()));
-      if (isFloatingType(Dest->getType()) || isVectorType(Dest->getType())) {
-        UnimplementedLoweringError(this, Instr);
-        return;
-      } else {
-        _mov(Dest, ReturnReg);
-      }
+      _addu(DestLo, Zero, ReturnReg);
+      _addu(DestHi, Zero, ReturnRegHi);
+      break;
+    }
+    case IceType_f32: {
+      _mov_s(Dest, ReturnReg);
+      break;
+    }
+    case IceType_f64: {
+      _mov_d(Dest, ReturnReg);
+      break;
+    }
+    case IceType_v4i1:
+    case IceType_v8i1:
+    case IceType_v16i1:
+    case IceType_v16i8:
+    case IceType_v8i16:
+    case IceType_v4i32:
+    case IceType_v4f32: {
+      UnimplementedLoweringError(this, Instr);
+      return;
+    }
     }
   }
 }
@@ -1423,7 +1894,24 @@ void TargetMIPS32::lowerRet(const InstRet *Instr) {
   Variable *Reg = nullptr;
   if (Instr->hasRetValue()) {
     Operand *Src0 = Instr->getRetValue();
+
     switch (Src0->getType()) {
+    default:
+      UnimplementedLoweringError(this, Instr);

[Jim Stichnoth, 2016/06/13 20:14:19]

You probably want a "break;" after this, otherwise it will fallthrough to the
f32 case when running with -skip-unimplemented.

+    case IceType_f32: {
+      if (auto *Src0V = llvm::dyn_cast<Variable>(Src0)) {
+        Reg = makeReg(Src0V->getType(), RegMIPS32::Reg_F0);
+        _mov_s(Reg, Src0V);
+      }
+      break;
+    }
+    case IceType_f64: {
+      if (auto *Src0V = llvm::dyn_cast<Variable>(Src0)) {
+        Reg = makeReg(Src0V->getType(), RegMIPS32::Reg_F0F1);
+        _mov_d(Reg, Src0V);
+      }
+      break;
+    }
     case IceType_i1:
     case IceType_i8:
     case IceType_i16:
@@ -1442,9 +1930,16 @@ void TargetMIPS32::lowerRet(const InstRet *Instr) {
       Context.insert<InstFakeUse>(R1);
       break;
     }
-
-    default:
+    case IceType_v4i1:
+    case IceType_v8i1:
+    case IceType_v16i1:
+    case IceType_v16i8:
+    case IceType_v8i16:
+    case IceType_v4i32:
+    case IceType_v4f32: {
       UnimplementedLoweringError(this, Instr);
+      break;
+    }
     }
   }
   _ret(getPhysicalRegister(RegMIPS32::Reg_RA), Reg);
@@ -1551,13 +2046,13 @@ void TargetDataMIPS32::lowerGlobals(const VariableDeclarationList &Vars,
 void TargetDataMIPS32::lowerConstants() {
   if (getFlags().getDisableTranslation())
     return;
-  UnimplementedError(getFlags());
+  // UnimplementedError(getFlags());

[Jim Stichnoth, 2016/06/13 20:14:19]

remove this

 }
 
 void TargetDataMIPS32::lowerJumpTables() {
   if (getFlags().getDisableTranslation())
     return;
-  UnimplementedError(getFlags());
+  // UnimplementedError(getFlags());

[Jim Stichnoth, 2016/06/13 20:14:19]

remove this

 }
 
 // Helper for legalize() to emit the right code to lower an operand to a
@@ -1565,8 +2060,11 @@ void TargetDataMIPS32::lowerJumpTables() {
 Variable *TargetMIPS32::copyToReg(Operand *Src, RegNumT RegNum) {
   Type Ty = Src->getType();
   Variable *Reg = makeReg(Ty, RegNum);
-  if (isVectorType(Ty) || isFloatingType(Ty)) {
+  if (isVectorType(Ty)) {
     UnimplementedError(getFlags());
+  } else if (isFloatingType(Ty)) {
+    (Ty == IceType_f32) ? _mov_s(Reg, llvm::dyn_cast<Variable>(Src))
+                        : _mov_d(Reg, llvm::dyn_cast<Variable>(Src));
   } else {
     // Mov's Src operand can really only be the flexible second operand type
     // or a register. Users should guarantee that.