Currently, codegen may spent some time in SDISel passes even if an entire

function is successfully handled by fast-isel. That's because function
arguments are *always* handled by SDISel. Introduce FastLowerArguments to
allow each target to provide hook to handle formal argument lowering.

As a proof-of-concept, add ARMFastIsel::FastLowerArguments to handle
functions with 4 or fewer scalar integer (i8, i16, or i32) arguments. It
completely eliminates the need for SDISel for trivial functions.

rdar://13163905


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@174855 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/CodeGen/FastISel.h

@@ -90,6 +90,11 @@ public:
 
   /// getCurDebugLoc() - Return current debug location information.
   DebugLoc getCurDebugLoc() const { return DL; }
+  
+  /// LowerArguments - Do "fast" instruction selection for function arguments
+  /// and append machine instructions to the current block. Return true if
+  /// it is successful.
+  bool LowerArguments();
 
   /// SelectInstruction - Do "fast" instruction selection for the given
   /// LLVM IR instruction, and append generated machine instructions to
@@ -160,6 +165,11 @@ protected:
   ///
   virtual bool
   TargetSelectInstruction(const Instruction *I) = 0;
+  
+  /// FastLowerArguments - This method is called by target-independent code to
+  /// do target specific argument lowering. It returns true if it was
+  /// successful.
+  virtual bool FastLowerArguments();
 
   /// FastEmit_r - This method is called by target-independent code
   /// to request that an instruction with the given type and opcode

lib/CodeGen/SelectionDAG/FastISel.cpp

@@ -87,6 +87,27 @@ void FastISel::startNewBlock() {
   LastLocalValue = EmitStartPt;
 }
 
+bool FastISel::LowerArguments() {
+  if (!FuncInfo.CanLowerReturn)
+    // Fallback to SDISel argument lowering code to deal with sret pointer
+    // parameter.
+    return false;
+  
+  if (!FastLowerArguments())
+    return false;
+
+  // Enter non-dead arguments into ValueMap for uses in non-entry BBs.
+  for (Function::const_arg_iterator I = FuncInfo.Fn->arg_begin(),
+         E = FuncInfo.Fn->arg_end(); I != E; ++I) {
+    if (!I->use_empty()) {
+      DenseMap<const Value *, unsigned>::iterator VI = LocalValueMap.find(I);
+      assert(VI != LocalValueMap.end() && "Missed an argument?");
+      FuncInfo.ValueMap[I] = VI->second;
+    }
+  }
+  return true;
+}
+
 void FastISel::flushLocalValueMap() {
   LocalValueMap.clear();
   LastLocalValue = EmitStartPt;
@@ -836,7 +857,8 @@ FastISel::SelectInstruction(const Instruction *I) {
 void
 FastISel::FastEmitBranch(MachineBasicBlock *MSucc, DebugLoc DL) {
 
-  if (FuncInfo.MBB->getBasicBlock()->size() > 1 && FuncInfo.MBB->isLayoutSuccessor(MSucc)) {
+  if (FuncInfo.MBB->getBasicBlock()->size() > 1 &&
+      FuncInfo.MBB->isLayoutSuccessor(MSucc)) {
     // For more accurate line information if this is the only instruction
     // in the block then emit it, otherwise we have the unconditional
     // fall-through case, which needs no instructions.
@@ -1067,6 +1089,10 @@ FastISel::FastISel(FunctionLoweringInfo &funcInfo,
 
 FastISel::~FastISel() {}
 
+bool FastISel::FastLowerArguments() {
+  return false;
+}
+
 unsigned FastISel::FastEmit_(MVT, MVT,
                              unsigned) {
   return 0;

lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp

@@ -1032,10 +1032,6 @@ void SelectionDAGISel::SelectAllBasicBlocks(const Function &Fn) {
     if (FuncInfo->MBB->isLandingPad())
       PrepareEHLandingPad();
 
-    // Lower any arguments needed in this block if this is the entry block.
-    if (LLVMBB == &Fn.getEntryBlock())
-      LowerArguments(LLVMBB);
-
     // Before doing SelectionDAG ISel, see if FastISel has been requested.
     if (FastIS) {
       FastIS->startNewBlock();
@@ -1043,9 +1039,15 @@ void SelectionDAGISel::SelectAllBasicBlocks(const Function &Fn) {
       // Emit code for any incoming arguments. This must happen before
       // beginning FastISel on the entry block.
       if (LLVMBB == &Fn.getEntryBlock()) {
-        CurDAG->setRoot(SDB->getControlRoot());
-        SDB->clear();
-        CodeGenAndEmitDAG();
+        // Lower any arguments needed in this block if this is the entry block.
+        if (!FastIS->LowerArguments()) {
+          // Call target indepedent SDISel argument lowering code if the target
+          // specific routine is not successful.
+          LowerArguments(LLVMBB);
+          CurDAG->setRoot(SDB->getControlRoot());
+          SDB->clear();
+          CodeGenAndEmitDAG();
+        }
 
         // If we inserted any instructions at the beginning, make a note of
         // where they are, so we can be sure to emit subsequent instructions
@@ -1156,6 +1158,10 @@ void SelectionDAGISel::SelectAllBasicBlocks(const Function &Fn) {
       }
 
       FastIS->recomputeInsertPt();
+    } else {
+      // Lower any arguments needed in this block if this is the entry block.
+      if (LLVMBB == &Fn.getEntryBlock())
+        LowerArguments(LLVMBB);
     }
 
     if (Begin != BI)

lib/Target/ARM/ARMFastISel.cpp

@@ -146,6 +146,7 @@ class ARMFastISel : public FastISel {
     virtual unsigned TargetMaterializeAlloca(const AllocaInst *AI);
     virtual bool TryToFoldLoad(MachineInstr *MI, unsigned OpNo,
                                const LoadInst *LI);
+    virtual bool FastLowerArguments();
   private:
   #include "ARMGenFastISel.inc"
 
@@ -2884,6 +2885,79 @@ unsigned ARMFastISel::ARMLowerPICELF(const GlobalValue *GV,
   return DestReg2;
 }
 
+bool ARMFastISel::FastLowerArguments() {
+  if (!FuncInfo.CanLowerReturn)
+    return false;
+
+  const Function *F = FuncInfo.Fn;
+  if (F->isVarArg())
+    return false;
+
+  CallingConv::ID CC = F->getCallingConv();
+  switch (CC) {
+  default:
+    return false;
+  case CallingConv::Fast:
+  case CallingConv::C:
+  case CallingConv::ARM_AAPCS_VFP:
+  case CallingConv::ARM_AAPCS:
+  case CallingConv::ARM_APCS:
+    break;
+  }
+
+  // Only handle simple cases. i.e. Up to 4 i8/i16/i32 scalar arguments
+  // which are passed in r0 - r3.
+  unsigned Idx = 1;
+  for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
+       I != E; ++I, ++Idx) {
+    if (Idx > 4)
+      return false;
+
+    if (F->getAttributes().hasAttribute(Idx, Attribute::InReg) ||
+        F->getAttributes().hasAttribute(Idx, Attribute::StructRet) ||
+        F->getAttributes().hasAttribute(Idx, Attribute::ByVal))
+      return false;
+
+    Type *ArgTy = I->getType();
+    if (ArgTy->isStructTy() || ArgTy->isArrayTy() || ArgTy->isVectorTy())
+      return false;
+
+    EVT ArgVT = TLI.getValueType(ArgTy);
+    switch (ArgVT.getSimpleVT().SimpleTy) {
+    case MVT::i8:
+    case MVT::i16:
+    case MVT::i32:
+      break;
+    default:
+      return false;
+    }
+  }
+
+
+  static const uint16_t GPRArgRegs[] = {
+    ARM::R0, ARM::R1, ARM::R2, ARM::R3
+  };
+
+  const TargetRegisterClass *RC = TLI.getRegClassFor(MVT::i32);
+  Idx = 0;
+  for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
+       I != E; ++I, ++Idx) {
+    if (I->use_empty())
+      continue;
+    unsigned SrcReg = GPRArgRegs[Idx];
+    unsigned DstReg = FuncInfo.MF->addLiveIn(SrcReg, RC);
+    // FIXME: Unfortunately it's necessary to emit a copy from the livein copy.
+    // Without this, EmitLiveInCopies may eliminate the livein if its only
+    // use is a bitcast (which isn't turned into an instruction).
+    unsigned ResultReg = createResultReg(RC);
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
+            ResultReg).addReg(DstReg, getKillRegState(true));
+    UpdateValueMap(I, ResultReg);
+  }
+
+  return true;
+}
+
 namespace llvm {
   FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo,
                                 const TargetLibraryInfo *libInfo) {