LinearScanner hotspot.
[oota-llvm.git] / lib / CodeGen / IntrinsicLowering.cpp
index 634f959f3c4781684989879d73a896770fb6c3f7..b5a03fc54d53e4aff1595a63a7803cf73067123b 100644 (file)
@@ -99,24 +99,30 @@ void DefaultIntrinsicLowering::AddPrototypes(Module &M) {
         EnsureFunctionExists(M, "abort", I->arg_end(), I->arg_end(),
                              Type::VoidTy);
         break;
-      case Intrinsic::memcpy:
+      case Intrinsic::memcpy_i32:
+      case Intrinsic::memcpy_i64:
         EnsureFunctionExists(M, "memcpy", I->arg_begin(), --I->arg_end(),
                              I->arg_begin()->getType());
         break;
-      case Intrinsic::memmove:
+      case Intrinsic::memmove_i32:
+      case Intrinsic::memmove_i64:
         EnsureFunctionExists(M, "memmove", I->arg_begin(), --I->arg_end(),
                              I->arg_begin()->getType());
         break;
-      case Intrinsic::memset:
+      case Intrinsic::memset_i32:
+      case Intrinsic::memset_i64:
         M.getOrInsertFunction("memset", PointerType::get(Type::SByteTy),
                               PointerType::get(Type::SByteTy),
-                              Type::IntTy, (--(--I->arg_end()))->getType(), 0);
+                              Type::IntTy, (--(--I->arg_end()))->getType(),
+                              (Type *)0);
         break;
-      case Intrinsic::isunordered:
+      case Intrinsic::isunordered_f32:
+      case Intrinsic::isunordered_f64:
         EnsureFunctionExists(M, "isunordered", I->arg_begin(), I->arg_end(),
                              Type::BoolTy);
         break;
-      case Intrinsic::sqrt:
+      case Intrinsic::sqrt_f32:
+      case Intrinsic::sqrt_f64:
         if(I->arg_begin()->getType() == Type::FloatTy)
           EnsureFunctionExists(M, "sqrtf", I->arg_begin(), I->arg_end(),
                                Type::FloatTy);
@@ -127,11 +133,105 @@ void DefaultIntrinsicLowering::AddPrototypes(Module &M) {
       }
 }
 
+/// LowerBSWAP - Emit the code to lower bswap of V before the specified
+/// instruction IP.
+static Value *LowerBSWAP(Value *V, Instruction *IP) {
+  assert(V->getType()->isInteger() && "Can't bswap a non-integer type!");
+
+  const Type *DestTy = V->getType();
+  
+  // Force to unsigned so that the shift rights are logical.
+  if (DestTy->isSigned())
+    V = new CastInst(V, DestTy->getUnsignedVersion(), V->getName(), IP);
+
+  unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
+  
+  switch(BitSize) {
+  default: assert(0 && "Unhandled type size of value to byteswap!");
+  case 16: {
+    Value *Tmp1 = new ShiftInst(Instruction::Shl, V,
+                                ConstantInt::get(Type::UByteTy,8),"bswap.2",IP);
+    Value *Tmp2 = new ShiftInst(Instruction::Shr, V,
+                                ConstantInt::get(Type::UByteTy,8),"bswap.1",IP);
+    V = BinaryOperator::createOr(Tmp1, Tmp2, "bswap.i16", IP);
+    break;
+  }
+  case 32: {
+    Value *Tmp4 = new ShiftInst(Instruction::Shl, V,
+                              ConstantInt::get(Type::UByteTy,24),"bswap.4", IP);
+    Value *Tmp3 = new ShiftInst(Instruction::Shl, V,
+                                ConstantInt::get(Type::UByteTy,8),"bswap.3",IP);
+    Value *Tmp2 = new ShiftInst(Instruction::Shr, V,
+                                ConstantInt::get(Type::UByteTy,8),"bswap.2",IP);
+    Value *Tmp1 = new ShiftInst(Instruction::Shr, V,
+                              ConstantInt::get(Type::UByteTy,24),"bswap.1", IP);
+    Tmp3 = BinaryOperator::createAnd(Tmp3, 
+                                     ConstantInt::get(Type::UIntTy, 0xFF0000),
+                                     "bswap.and3", IP);
+    Tmp2 = BinaryOperator::createAnd(Tmp2, 
+                                     ConstantInt::get(Type::UIntTy, 0xFF00),
+                                     "bswap.and2", IP);
+    Tmp4 = BinaryOperator::createOr(Tmp4, Tmp3, "bswap.or1", IP);
+    Tmp2 = BinaryOperator::createOr(Tmp2, Tmp1, "bswap.or2", IP);
+    V = BinaryOperator::createOr(Tmp4, Tmp3, "bswap.i32", IP);
+    break;
+  }
+  case 64: {
+    Value *Tmp8 = new ShiftInst(Instruction::Shl, V,
+                              ConstantInt::get(Type::UByteTy,56),"bswap.8", IP);
+    Value *Tmp7 = new ShiftInst(Instruction::Shl, V,
+                              ConstantInt::get(Type::UByteTy,40),"bswap.7", IP);
+    Value *Tmp6 = new ShiftInst(Instruction::Shl, V,
+                              ConstantInt::get(Type::UByteTy,24),"bswap.6", IP);
+    Value *Tmp5 = new ShiftInst(Instruction::Shl, V,
+                                ConstantInt::get(Type::UByteTy,8),"bswap.5",IP);
+    Value *Tmp4 = new ShiftInst(Instruction::Shr, V,
+                                ConstantInt::get(Type::UByteTy,8),"bswap.4",IP);
+    Value *Tmp3 = new ShiftInst(Instruction::Shr, V,
+                              ConstantInt::get(Type::UByteTy,24),"bswap.3", IP);
+    Value *Tmp2 = new ShiftInst(Instruction::Shr, V,
+                              ConstantInt::get(Type::UByteTy,40),"bswap.2", IP);
+    Value *Tmp1 = new ShiftInst(Instruction::Shr, V,
+                              ConstantInt::get(Type::UByteTy,56),"bswap.1", IP);
+    Tmp7 = BinaryOperator::createAnd(Tmp7,
+                             ConstantInt::get(Type::ULongTy, 
+                               0xFF000000000000ULL),
+                             "bswap.and7", IP);
+    Tmp6 = BinaryOperator::createAnd(Tmp6,
+                             ConstantInt::get(Type::ULongTy, 0xFF0000000000ULL),
+                             "bswap.and6", IP);
+    Tmp5 = BinaryOperator::createAnd(Tmp5,
+                             ConstantInt::get(Type::ULongTy, 0xFF00000000ULL),
+                             "bswap.and5", IP);
+    Tmp4 = BinaryOperator::createAnd(Tmp4,
+                             ConstantInt::get(Type::ULongTy, 0xFF000000ULL),
+                             "bswap.and4", IP);
+    Tmp3 = BinaryOperator::createAnd(Tmp3,
+                             ConstantInt::get(Type::ULongTy, 0xFF0000ULL),
+                             "bswap.and3", IP);
+    Tmp2 = BinaryOperator::createAnd(Tmp2,
+                             ConstantInt::get(Type::ULongTy, 0xFF00ULL),
+                             "bswap.and2", IP);
+    Tmp8 = BinaryOperator::createOr(Tmp8, Tmp7, "bswap.or1", IP);
+    Tmp6 = BinaryOperator::createOr(Tmp6, Tmp5, "bswap.or2", IP);
+    Tmp4 = BinaryOperator::createOr(Tmp4, Tmp3, "bswap.or3", IP);
+    Tmp2 = BinaryOperator::createOr(Tmp2, Tmp1, "bswap.or4", IP);
+    Tmp8 = BinaryOperator::createOr(Tmp8, Tmp6, "bswap.or5", IP);
+    Tmp4 = BinaryOperator::createOr(Tmp4, Tmp2, "bswap.or6", IP);
+    V = BinaryOperator::createOr(Tmp8, Tmp4, "bswap.i64", IP);
+    break;
+  }
+  }
+  
+  if (V->getType() != DestTy)
+    V = new CastInst(V, DestTy, V->getName(), IP);
+  return V;
+}
+
 /// LowerCTPOP - Emit the code to lower ctpop of V before the specified
-/// instruction.
+/// instruction IP.
 static Value *LowerCTPOP(Value *V, Instruction *IP) {
   assert(V->getType()->isInteger() && "Can't ctpop a non-integer type!");
-  unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
 
   static const uint64_t MaskValues[6] = {
     0x5555555555555555ULL, 0x3333333333333333ULL,
@@ -145,12 +245,13 @@ static Value *LowerCTPOP(Value *V, Instruction *IP) {
   if (DestTy->isSigned())
     V = new CastInst(V, DestTy->getUnsignedVersion(), V->getName(), IP);
 
+  unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
   for (unsigned i = 1, ct = 0; i != BitSize; i <<= 1, ++ct) {
     Value *MaskCst =
-      ConstantExpr::getCast(ConstantUInt::get(Type::ULongTy,
-                                              MaskValues[ct]), V->getType());
+      ConstantExpr::getCast(ConstantInt::get(Type::ULongTy, MaskValues[ct]),
+                                             V->getType());
     Value *LHS = BinaryOperator::createAnd(V, MaskCst, "cppop.and1", IP);
-    Value *VShift = new ShiftInst(Instruction::Shr, V, 
+    Value *VShift = new ShiftInst(Instruction::Shr, V,
                       ConstantInt::get(Type::UByteTy, i), "ctpop.sh", IP);
     Value *RHS = BinaryOperator::createAnd(VShift, MaskCst, "cppop.and2", IP);
     V = BinaryOperator::createAdd(LHS, RHS, "ctpop.step", IP);
@@ -161,6 +262,31 @@ static Value *LowerCTPOP(Value *V, Instruction *IP) {
   return V;
 }
 
+/// LowerCTLZ - Emit the code to lower ctlz of V before the specified
+/// instruction IP.
+static Value *LowerCTLZ(Value *V, Instruction *IP) {
+  const Type *DestTy = V->getType();
+
+  // Force to unsigned so that the shift rights are logical.
+  if (DestTy->isSigned())
+    V = new CastInst(V, DestTy->getUnsignedVersion(), V->getName(), IP);
+
+  unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
+  for (unsigned i = 1; i != BitSize; i <<= 1) {
+    Value *ShVal = ConstantInt::get(Type::UByteTy, i);
+    ShVal = new ShiftInst(Instruction::Shr, V, ShVal, "ctlz.sh", IP);
+    V = BinaryOperator::createOr(V, ShVal, "ctlz.step", IP);
+  }
+
+  if (V->getType() != DestTy)
+    V = new CastInst(V, DestTy, V->getName(), IP);
+
+  V = BinaryOperator::createNot(V, "", IP);
+  return LowerCTPOP(V, IP);
+}
+
+
+
 void DefaultIntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
   Function *Callee = CI->getCalledFunction();
   assert(Callee && "Cannot lower an indirect call!");
@@ -206,51 +332,30 @@ void DefaultIntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
                     AbortFCache);
     break;
   }
-  case Intrinsic::ctpop:
+  case Intrinsic::ctpop_i8:
+  case Intrinsic::ctpop_i16:
+  case Intrinsic::ctpop_i32:
+  case Intrinsic::ctpop_i64:
     CI->replaceAllUsesWith(LowerCTPOP(CI->getOperand(1), CI));
     break;
 
-  case Intrinsic::ctlz: {
-    Value *Src = CI->getOperand(1);
-    Value* SA;
-    switch (CI->getOperand(0)->getType()->getTypeID())
-    {
-    case Type::LongTyID:
-    case Type::ULongTyID:
-      SA = ConstantUInt::get(Type::UByteTy, 32);
-      Src = BinaryOperator::createOr(Src, new ShiftInst(Instruction::Shr, Src,
-                                                        SA, "", CI), "", CI);
-    case Type::IntTyID:
-    case Type::UIntTyID:
-      SA = ConstantUInt::get(Type::UByteTy, 16);
-      Src = BinaryOperator::createOr(Src, new ShiftInst(Instruction::Shr,
-                                                        Src, SA, "", CI),
-                                     "", CI);
-    case Type::ShortTyID:
-    case Type::UShortTyID:
-      SA = ConstantUInt::get(Type::UByteTy, 8);
-      Src = BinaryOperator::createOr(Src, new ShiftInst(Instruction::Shr,
-                                                        Src, SA, "", CI),
-                                     "", CI);
-    default:
-      SA = ConstantUInt::get(Type::UByteTy, 1);
-      Src = BinaryOperator::createOr(Src, new ShiftInst(Instruction::Shr, Src,
-                                                        SA, "", CI), "", CI);
-      SA = ConstantUInt::get(Type::UByteTy, 2);
-      Src = BinaryOperator::createOr(Src, new ShiftInst(Instruction::Shr, Src,
-                                                        SA, "", CI), "", CI);
-      SA = ConstantUInt::get(Type::UByteTy, 4);
-      Src = BinaryOperator::createOr(Src, new ShiftInst(Instruction::Shr, Src,
-                                                        SA, "", CI), "", CI);
-    };
-    Src = BinaryOperator::createNot(Src, "", CI);
-
-
-    Src = LowerCTPOP(Src, CI);
-    CI->replaceAllUsesWith(Src);
+  case Intrinsic::bswap_i16:
+  case Intrinsic::bswap_i32:
+  case Intrinsic::bswap_i64:
+    CI->replaceAllUsesWith(LowerBSWAP(CI->getOperand(1), CI));
     break;
-  }
-  case Intrinsic::cttz: {
+    
+  case Intrinsic::ctlz_i8:
+  case Intrinsic::ctlz_i16:
+  case Intrinsic::ctlz_i32:
+  case Intrinsic::ctlz_i64:
+    CI->replaceAllUsesWith(LowerCTLZ(CI->getOperand(1), CI));
+    break;
+
+  case Intrinsic::cttz_i8:
+  case Intrinsic::cttz_i16:
+  case Intrinsic::cttz_i32:
+  case Intrinsic::cttz_i64: {
     // cttz(x) -> ctpop(~X & (X-1))
     Value *Src = CI->getOperand(1);
     Value *NotSrc = BinaryOperator::createNot(Src, Src->getName()+".not", CI);
@@ -261,6 +366,19 @@ void DefaultIntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
     break;
   }
 
+  case Intrinsic::stacksave:
+  case Intrinsic::stackrestore: {
+    static bool Warned = false;
+    if (!Warned)
+      std::cerr << "WARNING: this target does not support the llvm.stack"
+       << (Callee->getIntrinsicID() == Intrinsic::stacksave ?
+           "save" : "restore") << " intrinsic.\n";
+    Warned = true;
+    if (Callee->getIntrinsicID() == Intrinsic::stacksave)
+      CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
+    break;
+  }
+    
   case Intrinsic::returnaddress:
   case Intrinsic::frameaddress:
     std::cerr << "WARNING: this target does not support the llvm."
@@ -275,17 +393,22 @@ void DefaultIntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
 
   case Intrinsic::pcmarker:
     break;    // Simply strip out pcmarker on unsupported architectures
+  case Intrinsic::readcyclecounter: {
+    std::cerr << "WARNING: this target does not support the llvm.readcyclecoun"
+              << "ter intrinsic.  It is being lowered to a constant 0\n";
+    CI->replaceAllUsesWith(ConstantInt::get(Type::ULongTy, 0));
+    break;
+  }
 
   case Intrinsic::dbg_stoppoint:
   case Intrinsic::dbg_region_start:
   case Intrinsic::dbg_region_end:
-  case Intrinsic::dbg_declare:
   case Intrinsic::dbg_func_start:
-    if (CI->getType() != Type::VoidTy)
-      CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
+  case Intrinsic::dbg_declare:
     break;    // Simply strip out debugging intrinsics
 
-  case Intrinsic::memcpy: {
+  case Intrinsic::memcpy_i32:
+  case Intrinsic::memcpy_i64: {
     // The memcpy intrinsic take an extra alignment argument that the memcpy
     // libc function does not.
     static Function *MemcpyFCache = 0;
@@ -293,7 +416,8 @@ void DefaultIntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
                     (*(CI->op_begin()+1))->getType(), MemcpyFCache);
     break;
   }
-  case Intrinsic::memmove: {
+  case Intrinsic::memmove_i32: 
+  case Intrinsic::memmove_i64: {
     // The memmove intrinsic take an extra alignment argument that the memmove
     // libc function does not.
     static Function *MemmoveFCache = 0;
@@ -301,7 +425,8 @@ void DefaultIntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
                     (*(CI->op_begin()+1))->getType(), MemmoveFCache);
     break;
   }
-  case Intrinsic::memset: {
+  case Intrinsic::memset_i32:
+  case Intrinsic::memset_i64: {
     // The memset intrinsic take an extra alignment argument that the memset
     // libc function does not.
     static Function *MemsetFCache = 0;
@@ -309,7 +434,8 @@ void DefaultIntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
                     (*(CI->op_begin()+1))->getType(), MemsetFCache);
     break;
   }
-  case Intrinsic::isunordered: {
+  case Intrinsic::isunordered_f32:
+  case Intrinsic::isunordered_f64: {
     Value *L = CI->getOperand(1);
     Value *R = CI->getOperand(2);
 
@@ -320,7 +446,8 @@ void DefaultIntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
                              "isunordered", CI));
     break;
   }
-  case Intrinsic::sqrt: {
+  case Intrinsic::sqrt_f32:
+  case Intrinsic::sqrt_f64: {
     static Function *sqrtFCache = 0;
     static Function *sqrtfFCache = 0;
     if(CI->getType() == Type::FloatTy)