lib/CodeGen/IntrinsicLowering.cpp

   1 //===-- IntrinsicLowering.cpp - Intrinsic Lowering default implementation -===//
   2 //
   3 //                     The LLVM Compiler Infrastructure
   4 //
   5 // This file was developed by the LLVM research group and is distributed under
   6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
   7 //
   8 //===----------------------------------------------------------------------===//
   9 //
  10 // This file implements the IntrinsicLowering class.
  11 //
  12 //===----------------------------------------------------------------------===//
  13
  14 #include "llvm/Constants.h"
  15 #include "llvm/DerivedTypes.h"
  16 #include "llvm/Module.h"
  17 #include "llvm/Instructions.h"
  18 #include "llvm/Type.h"
  19 #include "llvm/CodeGen/IntrinsicLowering.h"
  20 #include "llvm/Support/Streams.h"
  21 #include "llvm/Target/TargetData.h"
  22 #include "llvm/ADT/SmallVector.h"
  23 using namespace llvm;
  24
  25 template <class ArgIt>
  26 static void EnsureFunctionExists(Module &M, const char *Name,
  27                                  ArgIt ArgBegin, ArgIt ArgEnd,
  28                                  const Type *RetTy) {
  29   // Insert a correctly-typed definition now.
  30   std::vector<const Type *> ParamTys;
  31   for (ArgIt I = ArgBegin; I != ArgEnd; ++I)
  32     ParamTys.push_back(I->getType());
  33   M.getOrInsertFunction(Name, FunctionType::get(RetTy, ParamTys, false));
  34 }
  35
  36 /// ReplaceCallWith - This function is used when we want to lower an intrinsic
  37 /// call to a call of an external function.  This handles hard cases such as
  38 /// when there was already a prototype for the external function, and if that
  39 /// prototype doesn't match the arguments we expect to pass in.
  40 template <class ArgIt>
  41 static CallInst *ReplaceCallWith(const char *NewFn, CallInst *CI,
  42                                  ArgIt ArgBegin, ArgIt ArgEnd,
  43                                  const Type *RetTy, Constant *&FCache) {
  44   if (!FCache) {
  45     // If we haven't already looked up this function, check to see if the
  46     // program already contains a function with this name.
  47     Module *M = CI->getParent()->getParent()->getParent();
  48     // Get or insert the definition now.
  49     std::vector<const Type *> ParamTys;
  50     for (ArgIt I = ArgBegin; I != ArgEnd; ++I)
  51       ParamTys.push_back((*I)->getType());
  52     FCache = M->getOrInsertFunction(NewFn,
  53                                     FunctionType::get(RetTy, ParamTys, false));
  54   }
  55
  56   SmallVector<Value*, 8> Operands(ArgBegin, ArgEnd);
  57   CallInst *NewCI = new CallInst(FCache, &Operands[0], Operands.size(),
  58                                  CI->getName(), CI);
  59   if (!CI->use_empty())
  60     CI->replaceAllUsesWith(NewCI);
  61   return NewCI;
  62 }
  63
  64 void IntrinsicLowering::AddPrototypes(Module &M) {
  65   for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
  66     if (I->isDeclaration() && !I->use_empty())
  67       switch (I->getIntrinsicID()) {
  68       default: break;
  69       case Intrinsic::setjmp:
  70         EnsureFunctionExists(M, "setjmp", I->arg_begin(), I->arg_end(),
  71                              Type::Int32Ty);
  72         break;
  73       case Intrinsic::longjmp:
  74         EnsureFunctionExists(M, "longjmp", I->arg_begin(), I->arg_end(),
  75                              Type::VoidTy);
  76         break;
  77       case Intrinsic::siglongjmp:
  78         EnsureFunctionExists(M, "abort", I->arg_end(), I->arg_end(),
  79                              Type::VoidTy);
  80         break;
  81       case Intrinsic::memcpy_i32:
  82       case Intrinsic::memcpy_i64:
  83         M.getOrInsertFunction("memcpy", PointerType::get(Type::Int8Ty),
  84                               PointerType::get(Type::Int8Ty),
  85                               PointerType::get(Type::Int8Ty),
  86                               TD.getIntPtrType(), (Type *)0);
  87         break;
  88       case Intrinsic::memmove_i32:
  89       case Intrinsic::memmove_i64:
  90         M.getOrInsertFunction("memmove", PointerType::get(Type::Int8Ty),
  91                               PointerType::get(Type::Int8Ty),
  92                               PointerType::get(Type::Int8Ty),
  93                               TD.getIntPtrType(), (Type *)0);
  94         break;
  95       case Intrinsic::memset_i32:
  96       case Intrinsic::memset_i64:
  97         M.getOrInsertFunction("memset", PointerType::get(Type::Int8Ty),
  98                               PointerType::get(Type::Int8Ty), Type::Int32Ty,
  99                               TD.getIntPtrType(), (Type *)0);
 100         break;
 101       case Intrinsic::sqrt_f32:
 102       case Intrinsic::sqrt_f64:
 103         if(I->arg_begin()->getType() == Type::FloatTy)
 104           EnsureFunctionExists(M, "sqrtf", I->arg_begin(), I->arg_end(),
 105                                Type::FloatTy);
 106         else
 107           EnsureFunctionExists(M, "sqrt", I->arg_begin(), I->arg_end(),
 108                                Type::DoubleTy);
 109         break;
 110       }
 111 }
 112
 113 /// LowerBSWAP - Emit the code to lower bswap of V before the specified
 114 /// instruction IP.
 115 static Value *LowerBSWAP(Value *V, Instruction *IP) {
 116   assert(V->getType()->isInteger() && "Can't bswap a non-integer type!");
 117
 118   unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
 119
 120   switch(BitSize) {
 121   default: assert(0 && "Unhandled type size of value to byteswap!");
 122   case 16: {
 123     Value *Tmp1 = BinaryOperator::createShl(V,
 124                                 ConstantInt::get(V->getType(),8),"bswap.2",IP);
 125     Value *Tmp2 = BinaryOperator::createLShr(V,
 126                                 ConstantInt::get(V->getType(),8),"bswap.1",IP);
 127     V = BinaryOperator::createOr(Tmp1, Tmp2, "bswap.i16", IP);
 128     break;
 129   }
 130   case 32: {
 131     Value *Tmp4 = BinaryOperator::createShl(V,
 132                               ConstantInt::get(V->getType(),24),"bswap.4", IP);
 133     Value *Tmp3 = BinaryOperator::createShl(V,
 134                               ConstantInt::get(V->getType(),8),"bswap.3",IP);
 135     Value *Tmp2 = BinaryOperator::createLShr(V,
 136                               ConstantInt::get(V->getType(),8),"bswap.2",IP);
 137     Value *Tmp1 = BinaryOperator::createLShr(V,
 138                               ConstantInt::get(V->getType(),24),"bswap.1", IP);
 139     Tmp3 = BinaryOperator::createAnd(Tmp3,
 140                                      ConstantInt::get(Type::Int32Ty, 0xFF0000),
 141                                      "bswap.and3", IP);
 142     Tmp2 = BinaryOperator::createAnd(Tmp2,
 143                                      ConstantInt::get(Type::Int32Ty, 0xFF00),
 144                                      "bswap.and2", IP);
 145     Tmp4 = BinaryOperator::createOr(Tmp4, Tmp3, "bswap.or1", IP);
 146     Tmp2 = BinaryOperator::createOr(Tmp2, Tmp1, "bswap.or2", IP);
 147     V = BinaryOperator::createOr(Tmp4, Tmp3, "bswap.i32", IP);
 148     break;
 149   }
 150   case 64: {
 151     Value *Tmp8 = BinaryOperator::createShl(V,
 152                               ConstantInt::get(V->getType(),56),"bswap.8", IP);
 153     Value *Tmp7 = BinaryOperator::createShl(V,
 154                               ConstantInt::get(V->getType(),40),"bswap.7", IP);
 155     Value *Tmp6 = BinaryOperator::createShl(V,
 156                               ConstantInt::get(V->getType(),24),"bswap.6", IP);
 157     Value *Tmp5 = BinaryOperator::createShl(V,
 158                               ConstantInt::get(V->getType(),8),"bswap.5", IP);
 159     Value* Tmp4 = BinaryOperator::createLShr(V,
 160                               ConstantInt::get(V->getType(),8),"bswap.4", IP);
 161     Value* Tmp3 = BinaryOperator::createLShr(V,
 162                               ConstantInt::get(V->getType(),24),"bswap.3", IP);
 163     Value* Tmp2 = BinaryOperator::createLShr(V,
 164                               ConstantInt::get(V->getType(),40),"bswap.2", IP);
 165     Value* Tmp1 = BinaryOperator::createLShr(V,
 166                               ConstantInt::get(V->getType(),56),"bswap.1", IP);
 167     Tmp7 = BinaryOperator::createAnd(Tmp7,
 168                              ConstantInt::get(Type::Int64Ty,
 169                                0xFF000000000000ULL),
 170                              "bswap.and7", IP);
 171     Tmp6 = BinaryOperator::createAnd(Tmp6,
 172                              ConstantInt::get(Type::Int64Ty, 0xFF0000000000ULL),
 173                              "bswap.and6", IP);
 174     Tmp5 = BinaryOperator::createAnd(Tmp5,
 175                              ConstantInt::get(Type::Int64Ty, 0xFF00000000ULL),
 176                              "bswap.and5", IP);
 177     Tmp4 = BinaryOperator::createAnd(Tmp4,
 178                              ConstantInt::get(Type::Int64Ty, 0xFF000000ULL),
 179                              "bswap.and4", IP);
 180     Tmp3 = BinaryOperator::createAnd(Tmp3,
 181                              ConstantInt::get(Type::Int64Ty, 0xFF0000ULL),
 182                              "bswap.and3", IP);
 183     Tmp2 = BinaryOperator::createAnd(Tmp2,
 184                              ConstantInt::get(Type::Int64Ty, 0xFF00ULL),
 185                              "bswap.and2", IP);
 186     Tmp8 = BinaryOperator::createOr(Tmp8, Tmp7, "bswap.or1", IP);
 187     Tmp6 = BinaryOperator::createOr(Tmp6, Tmp5, "bswap.or2", IP);
 188     Tmp4 = BinaryOperator::createOr(Tmp4, Tmp3, "bswap.or3", IP);
 189     Tmp2 = BinaryOperator::createOr(Tmp2, Tmp1, "bswap.or4", IP);
 190     Tmp8 = BinaryOperator::createOr(Tmp8, Tmp6, "bswap.or5", IP);
 191     Tmp4 = BinaryOperator::createOr(Tmp4, Tmp2, "bswap.or6", IP);
 192     V = BinaryOperator::createOr(Tmp8, Tmp4, "bswap.i64", IP);
 193     break;
 194   }
 195   }
 196   return V;
 197 }
 198
 199 /// LowerCTPOP - Emit the code to lower ctpop of V before the specified
 200 /// instruction IP.
 201 static Value *LowerCTPOP(Value *V, Instruction *IP) {
 202   assert(V->getType()->isInteger() && "Can't ctpop a non-integer type!");
 203
 204   static const uint64_t MaskValues[6] = {
 205     0x5555555555555555ULL, 0x3333333333333333ULL,
 206     0x0F0F0F0F0F0F0F0FULL, 0x00FF00FF00FF00FFULL,
 207     0x0000FFFF0000FFFFULL, 0x00000000FFFFFFFFULL
 208   };
 209
 210   unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
 211
 212   for (unsigned i = 1, ct = 0; i != BitSize; i <<= 1, ++ct) {
 213     Value *MaskCst = ConstantInt::get(V->getType(), MaskValues[ct]);
 214     Value *LHS = BinaryOperator::createAnd(V, MaskCst, "cppop.and1", IP);
 215     Value *VShift = BinaryOperator::createLShr(V,
 216                       ConstantInt::get(V->getType(), i), "ctpop.sh", IP);
 217     Value *RHS = BinaryOperator::createAnd(VShift, MaskCst, "cppop.and2", IP);
 218     V = BinaryOperator::createAdd(LHS, RHS, "ctpop.step", IP);
 219   }
 220
 221   return CastInst::createIntegerCast(V, Type::Int32Ty, false, "ctpop", IP);
 222 }
 223
 224 /// LowerCTLZ - Emit the code to lower ctlz of V before the specified
 225 /// instruction IP.
 226 static Value *LowerCTLZ(Value *V, Instruction *IP) {
 227
 228   unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
 229   for (unsigned i = 1; i != BitSize; i <<= 1) {
 230     Value *ShVal = ConstantInt::get(V->getType(), i);
 231     ShVal = BinaryOperator::createLShr(V, ShVal, "ctlz.sh", IP);
 232     V = BinaryOperator::createOr(V, ShVal, "ctlz.step", IP);
 233   }
 234
 235   V = BinaryOperator::createNot(V, "", IP);
 236   return LowerCTPOP(V, IP);
 237 }
 238
 239 /// Convert the llvm.bit.part_select.iX.iY.iZ intrinsic. This intrinsic takes
 240 /// three integer operands of arbitrary bit width. The first operand is the
 241 /// value from which to select the bits. The second and third operands define a
 242 /// range of bits to select.  The result is the bits selected and has a
 243 /// corresponding width of Left-Right (second operand - third operand).
 244 /// @see IEEE 1666-2005, System C, Section 7.2.6, pg 175.
 245 /// @brief Lowering of llvm.bit.part_select intrinsic.
 246 static Instruction *LowerBitPartSelect(CallInst *CI) {
 247   // Make sure we're dealing with a part select intrinsic here
 248   Function *F = CI->getCalledFunction();
 249   const FunctionType *FT = F->getFunctionType();
 250   if (!F->isDeclaration() || !FT->getReturnType()->isInteger() ||
 251       FT->getNumParams() != 3 || !FT->getParamType(0)->isInteger() ||
 252       !FT->getParamType(1)->isInteger() || !FT->getParamType(2)->isInteger())
 253     return CI;
 254
 255   // Get the intrinsic implementation function by converting all the . to _
 256   // in the intrinsic's function name and then reconstructing the function
 257   // declaration.
 258   std::string Name(F->getName());
 259   for (unsigned i = 4; i < Name.length(); ++i)
 260     if (Name[i] == '.')
 261       Name[i] = '_';
 262   Module* M = F->getParent();
 263   F = cast<Function>(M->getOrInsertFunction(Name, FT));
 264   F->setLinkage(GlobalValue::InternalLinkage);
 265
 266   // If we haven't defined the impl function yet, do so now
 267   if (F->isDeclaration()) {
 268
 269     // Get the arguments to the function
 270     Value* Val = F->getOperand(0);
 271     Value* Left = F->getOperand(1);
 272     Value* Right = F->getOperand(2);
 273
 274     // We want to select a range of bits here such that [Left, Right] is shifted
 275     // down to the low bits. However, it is quite possible that Left is smaller
 276     // than Right in which case the bits have to be reversed.
 277
 278     // Create the blocks we will need for the two cases (forward, reverse)
 279     BasicBlock* CurBB   = new BasicBlock("entry", F);
 280     BasicBlock *RevSize = new BasicBlock("revsize", CurBB->getParent());
 281     BasicBlock *FwdSize = new BasicBlock("fwdsize", CurBB->getParent());
 282     BasicBlock *Compute = new BasicBlock("compute", CurBB->getParent());
 283     BasicBlock *Reverse = new BasicBlock("reverse", CurBB->getParent());
 284     BasicBlock *RsltBlk = new BasicBlock("result",  CurBB->getParent());
 285
 286     // Cast Left and Right to the size of Val so the widths are all the same
 287     if (Left->getType() != Val->getType())
 288       Left = CastInst::createIntegerCast(Left, Val->getType(), false,
 289                                          "tmp", CurBB);
 290     if (Right->getType() != Val->getType())
 291       Right = CastInst::createIntegerCast(Right, Val->getType(), false,
 292                                           "tmp", CurBB);
 293
 294     // Compute a few things that both cases will need, up front.
 295     Constant* Zero = ConstantInt::get(Val->getType(), 0);
 296     Constant* One = ConstantInt::get(Val->getType(), 1);
 297     Constant* AllOnes = ConstantInt::getAllOnesValue(Val->getType());
 298
 299     // Compare the Left and Right bit positions. This is used to determine
 300     // which case we have (forward or reverse)
 301     ICmpInst *Cmp = new ICmpInst(ICmpInst::ICMP_ULT, Left, Right, "less",CurBB);
 302     new BranchInst(RevSize, FwdSize, Cmp, CurBB);
 303
 304     // First, copmute the number of bits in the forward case.
 305     Instruction* FBitSize =
 306       BinaryOperator::createSub(Left, Right,"fbits", FwdSize);
 307     new BranchInst(Compute, FwdSize);
 308
 309     // Second, compute the number of bits in the reverse case.
 310     Instruction* RBitSize =
 311       BinaryOperator::createSub(Right, Left, "rbits", RevSize);
 312     new BranchInst(Compute, RevSize);
 313
 314     // Now, compute the bit range. Start by getting the bitsize and the shift
 315     // amount (either Left or Right) from PHI nodes. Then we compute a mask for
 316     // the number of bits we want in the range. We shift the bits down to the
 317     // least significant bits, apply the mask to zero out unwanted high bits,
 318     // and we have computed the "forward" result. It may still need to be
 319     // reversed.
 320
 321     // Get the BitSize from one of the two subtractions
 322     PHINode *BitSize = new PHINode(Val->getType(), "bits", Compute);
 323     BitSize->reserveOperandSpace(2);
 324     BitSize->addIncoming(FBitSize, FwdSize);
 325     BitSize->addIncoming(RBitSize, RevSize);
 326
 327     // Get the ShiftAmount as the smaller of Left/Right
 328     PHINode *ShiftAmt = new PHINode(Val->getType(), "shiftamt", Compute);
 329     ShiftAmt->reserveOperandSpace(2);
 330     ShiftAmt->addIncoming(Right, FwdSize);
 331     ShiftAmt->addIncoming(Left, RevSize);
 332
 333     // Increment the bit size
 334     Instruction *BitSizePlusOne =
 335       BinaryOperator::createAdd(BitSize, One, "bits", Compute);
 336
 337     // Create a Mask to zero out the high order bits.
 338     Instruction* Mask =
 339       BinaryOperator::createShl(AllOnes, BitSizePlusOne, "mask", Compute);
 340     Mask = BinaryOperator::createNot(Mask, "mask", Compute);
 341
 342     // Shift the bits down and apply the mask
 343     Instruction* FRes =
 344       BinaryOperator::createLShr(Val, ShiftAmt, "fres", Compute);
 345     FRes = BinaryOperator::createAnd(FRes, Mask, "fres", Compute);
 346     new BranchInst(Reverse, RsltBlk, Cmp, Compute);
 347
 348     // In the Reverse block we have the mask already in FRes but we must reverse
 349     // it by shifting FRes bits right and putting them in RRes by shifting them
 350     // in from left.
 351
 352     // First set up our loop counters
 353     PHINode *Count = new PHINode(Val->getType(), "count", Reverse);
 354     Count->reserveOperandSpace(2);
 355     Count->addIncoming(BitSizePlusOne, Compute);
 356
 357     // Next, get the value that we are shifting.
 358     PHINode *BitsToShift   = new PHINode(Val->getType(), "val", Reverse);
 359     BitsToShift->reserveOperandSpace(2);
 360     BitsToShift->addIncoming(FRes, Compute);
 361
 362     // Finally, get the result of the last computation
 363     PHINode *RRes  = new PHINode(Val->getType(), "rres", Reverse);
 364     RRes->reserveOperandSpace(2);
 365     RRes->addIncoming(Zero, Compute);
 366
 367     // Decrement the counter
 368     Instruction *Decr = BinaryOperator::createSub(Count, One, "decr", Reverse);
 369     Count->addIncoming(Decr, Reverse);
 370
 371     // Compute the Bit that we want to move
 372     Instruction *Bit =
 373       BinaryOperator::createAnd(BitsToShift, One, "bit", Reverse);
 374
 375     // Compute the new value for next iteration.
 376     Instruction *NewVal =
 377       BinaryOperator::createLShr(BitsToShift, One, "rshift", Reverse);
 378     BitsToShift->addIncoming(NewVal, Reverse);
 379
 380     // Shift the bit into the low bits of the result.
 381     Instruction *NewRes =
 382       BinaryOperator::createShl(RRes, One, "lshift", Reverse);
 383     NewRes = BinaryOperator::createOr(NewRes, Bit, "addbit", Reverse);
 384     RRes->addIncoming(NewRes, Reverse);
 385
 386     // Terminate loop if we've moved all the bits.
 387     ICmpInst *Cond =
 388       new ICmpInst(ICmpInst::ICMP_EQ, Decr, Zero, "cond", Reverse);
 389     new BranchInst(RsltBlk, Reverse, Cond, Reverse);
 390
 391     // Finally, in the result block, select one of the two results with a PHI
 392     // node and return the result;
 393     CurBB = RsltBlk;
 394     PHINode *BitSelect = new PHINode(Val->getType(), "part_select", CurBB);
 395     BitSelect->reserveOperandSpace(2);
 396     BitSelect->addIncoming(FRes, Compute);
 397     BitSelect->addIncoming(NewRes, Reverse);
 398     new ReturnInst(BitSelect, CurBB);
 399   }
 400
 401   // Return a call to the implementation function
 402   Value *Args[3];
 403   Args[0] = CI->getOperand(0);
 404   Args[1] = CI->getOperand(1);
 405   Args[2] = CI->getOperand(2);
 406   return new CallInst(F, Args, 3, CI->getName(), CI);
 407 }
 408
 409
 410 void IntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
 411   Function *Callee = CI->getCalledFunction();
 412   assert(Callee && "Cannot lower an indirect call!");
 413
 414   switch (Callee->getIntrinsicID()) {
 415   case Intrinsic::not_intrinsic:
 416     cerr << "Cannot lower a call to a non-intrinsic function '"
 417          << Callee->getName() << "'!\n";
 418     abort();
 419   default:
 420     cerr << "Error: Code generator does not support intrinsic function '"
 421          << Callee->getName() << "'!\n";
 422     abort();
 423
 424     // The setjmp/longjmp intrinsics should only exist in the code if it was
 425     // never optimized (ie, right out of the CFE), or if it has been hacked on
 426     // by the lowerinvoke pass.  In both cases, the right thing to do is to
 427     // convert the call to an explicit setjmp or longjmp call.
 428   case Intrinsic::setjmp: {
 429     static Constant *SetjmpFCache = 0;
 430     Value *V = ReplaceCallWith("setjmp", CI, CI->op_begin()+1, CI->op_end(),
 431                                Type::Int32Ty, SetjmpFCache);
 432     if (CI->getType() != Type::VoidTy)
 433       CI->replaceAllUsesWith(V);
 434     break;
 435   }
 436   case Intrinsic::sigsetjmp:
 437      if (CI->getType() != Type::VoidTy)
 438        CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
 439      break;
 440
 441   case Intrinsic::longjmp: {
 442     static Constant *LongjmpFCache = 0;
 443     ReplaceCallWith("longjmp", CI, CI->op_begin()+1, CI->op_end(),
 444                     Type::VoidTy, LongjmpFCache);
 445     break;
 446   }
 447
 448   case Intrinsic::siglongjmp: {
 449     // Insert the call to abort
 450     static Constant *AbortFCache = 0;
 451     ReplaceCallWith("abort", CI, CI->op_end(), CI->op_end(),
 452                     Type::VoidTy, AbortFCache);
 453     break;
 454   }
 455   case Intrinsic::ctpop:
 456     CI->replaceAllUsesWith(LowerCTPOP(CI->getOperand(1), CI));
 457     break;
 458
 459   case Intrinsic::bswap:
 460     CI->replaceAllUsesWith(LowerBSWAP(CI->getOperand(1), CI));
 461     break;
 462
 463   case Intrinsic::ctlz:
 464     CI->replaceAllUsesWith(LowerCTLZ(CI->getOperand(1), CI));
 465     break;
 466
 467   case Intrinsic::cttz: {
 468     // cttz(x) -> ctpop(~X & (X-1))
 469     Value *Src = CI->getOperand(1);
 470     Value *NotSrc = BinaryOperator::createNot(Src, Src->getName()+".not", CI);
 471     Value *SrcM1  = ConstantInt::get(Src->getType(), 1);
 472     SrcM1 = BinaryOperator::createSub(Src, SrcM1, "", CI);
 473     Src = LowerCTPOP(BinaryOperator::createAnd(NotSrc, SrcM1, "", CI), CI);
 474     CI->replaceAllUsesWith(Src);
 475     break;
 476   }
 477
 478   case Intrinsic::part_select:
 479     CI->replaceAllUsesWith(LowerBitPartSelect(CI));
 480     break;
 481
 482   case Intrinsic::stacksave:
 483   case Intrinsic::stackrestore: {
 484     static bool Warned = false;
 485     if (!Warned)
 486       cerr << "WARNING: this target does not support the llvm.stack"
 487            << (Callee->getIntrinsicID() == Intrinsic::stacksave ?
 488                "save" : "restore") << " intrinsic.\n";
 489     Warned = true;
 490     if (Callee->getIntrinsicID() == Intrinsic::stacksave)
 491       CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
 492     break;
 493   }
 494
 495   case Intrinsic::returnaddress:
 496   case Intrinsic::frameaddress:
 497     cerr << "WARNING: this target does not support the llvm."
 498          << (Callee->getIntrinsicID() == Intrinsic::returnaddress ?
 499              "return" : "frame") << "address intrinsic.\n";
 500     CI->replaceAllUsesWith(ConstantPointerNull::get(
 501                                             cast<PointerType>(CI->getType())));
 502     break;
 503
 504   case Intrinsic::prefetch:
 505     break;    // Simply strip out prefetches on unsupported architectures
 506
 507   case Intrinsic::pcmarker:
 508     break;    // Simply strip out pcmarker on unsupported architectures
 509   case Intrinsic::readcyclecounter: {
 510     cerr << "WARNING: this target does not support the llvm.readcyclecoun"
 511          << "ter intrinsic.  It is being lowered to a constant 0\n";
 512     CI->replaceAllUsesWith(ConstantInt::get(Type::Int64Ty, 0));
 513     break;
 514   }
 515
 516   case Intrinsic::dbg_stoppoint:
 517   case Intrinsic::dbg_region_start:
 518   case Intrinsic::dbg_region_end:
 519   case Intrinsic::dbg_func_start:
 520   case Intrinsic::dbg_declare:
 521   case Intrinsic::eh_exception:
 522   case Intrinsic::eh_selector:
 523   case Intrinsic::eh_filter:
 524     break;    // Simply strip out debugging and eh intrinsics
 525
 526   case Intrinsic::memcpy_i32:
 527   case Intrinsic::memcpy_i64: {
 528     static Constant *MemcpyFCache = 0;
 529     Value *Size = CI->getOperand(3);
 530     const Type *IntPtr = TD.getIntPtrType();
 531     if (Size->getType()->getPrimitiveSizeInBits() <
 532         IntPtr->getPrimitiveSizeInBits())
 533       Size = new ZExtInst(Size, IntPtr, "", CI);
 534     else if (Size->getType()->getPrimitiveSizeInBits() >
 535              IntPtr->getPrimitiveSizeInBits())
 536       Size = new TruncInst(Size, IntPtr, "", CI);
 537     Value *Ops[3];
 538     Ops[0] = CI->getOperand(1);
 539     Ops[1] = CI->getOperand(2);
 540     Ops[2] = Size;
 541     ReplaceCallWith("memcpy", CI, Ops, Ops+3, CI->getOperand(1)->getType(),
 542                     MemcpyFCache);
 543     break;
 544   }
 545   case Intrinsic::memmove_i32:
 546   case Intrinsic::memmove_i64: {
 547     static Constant *MemmoveFCache = 0;
 548     Value *Size = CI->getOperand(3);
 549     const Type *IntPtr = TD.getIntPtrType();
 550     if (Size->getType()->getPrimitiveSizeInBits() <
 551         IntPtr->getPrimitiveSizeInBits())
 552       Size = new ZExtInst(Size, IntPtr, "", CI);
 553     else if (Size->getType()->getPrimitiveSizeInBits() >
 554              IntPtr->getPrimitiveSizeInBits())
 555       Size = new TruncInst(Size, IntPtr, "", CI);
 556     Value *Ops[3];
 557     Ops[0] = CI->getOperand(1);
 558     Ops[1] = CI->getOperand(2);
 559     Ops[2] = Size;
 560     ReplaceCallWith("memmove", CI, Ops, Ops+3, CI->getOperand(1)->getType(),
 561                     MemmoveFCache);
 562     break;
 563   }
 564   case Intrinsic::memset_i32:
 565   case Intrinsic::memset_i64: {
 566     static Constant *MemsetFCache = 0;
 567     Value *Size = CI->getOperand(3);
 568     const Type *IntPtr = TD.getIntPtrType();
 569     if (Size->getType()->getPrimitiveSizeInBits() <
 570         IntPtr->getPrimitiveSizeInBits())
 571       Size = new ZExtInst(Size, IntPtr, "", CI);
 572     else if (Size->getType()->getPrimitiveSizeInBits() >
 573              IntPtr->getPrimitiveSizeInBits())
 574       Size = new TruncInst(Size, IntPtr, "", CI);
 575     Value *Ops[3];
 576     Ops[0] = CI->getOperand(1);
 577     // Extend the amount to i32.
 578     Ops[1] = new ZExtInst(CI->getOperand(2), Type::Int32Ty, "", CI);
 579     Ops[2] = Size;
 580     ReplaceCallWith("memset", CI, Ops, Ops+3, CI->getOperand(1)->getType(),
 581                     MemsetFCache);
 582     break;
 583   }
 584   case Intrinsic::sqrt_f32: {
 585     static Constant *sqrtfFCache = 0;
 586     ReplaceCallWith("sqrtf", CI, CI->op_begin()+1, CI->op_end(),
 587                     Type::FloatTy, sqrtfFCache);
 588     break;
 589   }
 590   case Intrinsic::sqrt_f64: {
 591     static Constant *sqrtFCache = 0;
 592     ReplaceCallWith("sqrt", CI, CI->op_begin()+1, CI->op_end(),
 593                     Type::DoubleTy, sqrtFCache);
 594     break;
 595   }
 596   }
 597
 598   assert(CI->use_empty() &&
 599          "Lowering should have eliminated any uses of the intrinsic call!");
 600   CI->eraseFromParent();
 601 }