[fast-isel] Add X86FastIsel::FastLowerArguments to handle functions with 6 or

fewer scalar integer (i32 or i64) arguments. It completely eliminates the need
for SDISel for trivial functions.

Also, add the new llc -fast-isel-abort-args option, which is similar to
-fast-isel-abort option, but for formal argument lowering.

llvm-svn: 176052

lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp

@@ -145,6 +145,10 @@ EnableFastISelVerbose("fast-isel-verbose", cl::Hidden,
 static cl::opt<bool>
 EnableFastISelAbort("fast-isel-abort", cl::Hidden,
           cl::desc("Enable abort calls when \"fast\" instruction fails"));
+static cl::opt<bool>
+EnableFastISelAbortArgs("fast-isel-abort-args", cl::Hidden,
+          cl::desc("Enable abort calls when \"fast\" instruction fails to "
+                   "lower formal arguments"));
 
 static cl::opt<bool>
 UseMBPI("use-mbpi",
@@ -1048,6 +1052,12 @@ void SelectionDAGISel::SelectAllBasicBlocks(const Function &Fn) {
       if (LLVMBB == &Fn.getEntryBlock()) {
         // Lower any arguments needed in this block if this is the entry block.
         if (!FastIS->LowerArguments()) {
+
+          if (EnableFastISelAbortArgs)
+            // The "fast" selector couldn't lower these arguments.  For the
+            // purpose of debugging, just abort.
+            llvm_unreachable("FastISel didn't lower all arguments");
+
           // Call target indepedent SDISel argument lowering code if the target
           // specific routine is not successful.
           LowerArguments(LLVMBB);

lib/Target/X86/X86FastISel.cpp

@@ -75,6 +75,8 @@ public:
   virtual bool TryToFoldLoad(MachineInstr *MI, unsigned OpNo,
                              const LoadInst *LI);
 
+  virtual bool FastLowerArguments();
+
 #include "X86GenFastISel.inc"
 
 private:
@@ -1520,6 +1522,77 @@ bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) {
   }
 }
 
+bool X86FastISel::FastLowerArguments() {
+  if (!FuncInfo.CanLowerReturn)
+    return false;
+
+  const Function *F = FuncInfo.Fn;
+  if (F->isVarArg())
+    return false;
+
+  CallingConv::ID CC = F->getCallingConv();
+  if (CC != CallingConv::C)
+    return false;
+  
+  if (!Subtarget->is64Bit())
+    return false;
+  
+  // Only handle simple cases. i.e. Up to 6 i32/i64 scalar arguments.
+  unsigned Idx = 1;
+  for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
+       I != E; ++I, ++Idx) {
+    if (Idx > 6)
+      return false;
+
+    if (F->getAttributes().hasAttribute(Idx, Attribute::ByVal) ||
+        F->getAttributes().hasAttribute(Idx, Attribute::InReg) ||
+        F->getAttributes().hasAttribute(Idx, Attribute::StructRet) ||
+        F->getAttributes().hasAttribute(Idx, Attribute::Nest))
+      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::i32:
+    case MVT::i64:
+      break;
+    default:
+      return false;
+    }
+  }
+
+  static const uint16_t GPR32ArgRegs[] = {
+    X86::EDI, X86::ESI, X86::EDX, X86::ECX, X86::R8D, X86::R9D
+  };
+  static const uint16_t GPR64ArgRegs[] = {
+    X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8 , X86::R9
+  };
+
+  Idx = 0;
+  const TargetRegisterClass *RC32 = TLI.getRegClassFor(MVT::i32);
+  const TargetRegisterClass *RC64 = TLI.getRegClassFor(MVT::i64);
+  for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
+       I != E; ++I, ++Idx) {
+    if (I->use_empty())
+      continue;
+    bool is32Bit = TLI.getValueType(I->getType()) == MVT::i32;
+    const TargetRegisterClass *RC = is32Bit ? RC32 : RC64;
+    unsigned SrcReg = is32Bit ? GPR32ArgRegs[Idx] : GPR64ArgRegs[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;
+}
+
 bool X86FastISel::X86SelectCall(const Instruction *I) {
   const CallInst *CI = cast<CallInst>(I);
   const Value *Callee = CI->getCalledValue();

test/CodeGen/X86/fast-isel-args.ll

@@ -0,0 +1,25 @@
+; RUN: llc < %s -fast-isel -fast-isel-abort -fast-isel-abort-args -verify-machineinstrs -mtriple=x86_64-apple-darwin10
+
+; Just make sure these don't abort when lowering the arguments.
+define i32 @t1(i32 %a, i32 %b, i32 %c) {
+entry:
+  %add = add nsw i32 %b, %a
+  %add1 = add nsw i32 %add, %c
+  ret i32 %add1
+}
+
+define i64 @t2(i64 %a, i64 %b, i64 %c) {
+entry:
+  %add = add nsw i64 %b, %a
+  %add1 = add nsw i64 %add, %c
+  ret i64 %add1
+}
+
+define i64 @t3(i32 %a, i64 %b, i32 %c) #2 {
+entry:
+  %conv = sext i32 %a to i64
+  %add = add nsw i64 %conv, %b
+  %conv1 = sext i32 %c to i64
+  %add2 = add nsw i64 %add, %conv1
+  ret i64 %add2
+}