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 V;
 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 void IntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
 240   Function *Callee = CI->getCalledFunction();
 241   assert(Callee && "Cannot lower an indirect call!");
 242
 243   switch (Callee->getIntrinsicID()) {
 244   case Intrinsic::not_intrinsic:
 245     cerr << "Cannot lower a call to a non-intrinsic function '"
 246          << Callee->getName() << "'!\n";
 247     abort();
 248   default:
 249     cerr << "Error: Code generator does not support intrinsic function '"
 250          << Callee->getName() << "'!\n";
 251     abort();
 252
 253     // The setjmp/longjmp intrinsics should only exist in the code if it was
 254     // never optimized (ie, right out of the CFE), or if it has been hacked on
 255     // by the lowerinvoke pass.  In both cases, the right thing to do is to
 256     // convert the call to an explicit setjmp or longjmp call.
 257   case Intrinsic::setjmp: {
 258     static Constant *SetjmpFCache = 0;
 259     Value *V = ReplaceCallWith("setjmp", CI, CI->op_begin()+1, CI->op_end(),
 260                                Type::Int32Ty, SetjmpFCache);
 261     if (CI->getType() != Type::VoidTy)
 262       CI->replaceAllUsesWith(V);
 263     break;
 264   }
 265   case Intrinsic::sigsetjmp:
 266      if (CI->getType() != Type::VoidTy)
 267        CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
 268      break;
 269
 270   case Intrinsic::longjmp: {
 271     static Constant *LongjmpFCache = 0;
 272     ReplaceCallWith("longjmp", CI, CI->op_begin()+1, CI->op_end(),
 273                     Type::VoidTy, LongjmpFCache);
 274     break;
 275   }
 276
 277   case Intrinsic::siglongjmp: {
 278     // Insert the call to abort
 279     static Constant *AbortFCache = 0;
 280     ReplaceCallWith("abort", CI, CI->op_end(), CI->op_end(),
 281                     Type::VoidTy, AbortFCache);
 282     break;
 283   }
 284   case Intrinsic::ctpop:
 285     CI->replaceAllUsesWith(LowerCTPOP(CI->getOperand(1), CI));
 286     break;
 287
 288   case Intrinsic::bswap:
 289     CI->replaceAllUsesWith(LowerBSWAP(CI->getOperand(1), CI));
 290     break;
 291
 292   case Intrinsic::ctlz:
 293     CI->replaceAllUsesWith(LowerCTLZ(CI->getOperand(1), CI));
 294     break;
 295
 296   case Intrinsic::cttz: {
 297     // cttz(x) -> ctpop(~X & (X-1))
 298     Value *Src = CI->getOperand(1);
 299     Value *NotSrc = BinaryOperator::createNot(Src, Src->getName()+".not", CI);
 300     Value *SrcM1  = ConstantInt::get(Src->getType(), 1);
 301     SrcM1 = BinaryOperator::createSub(Src, SrcM1, "", CI);
 302     Src = LowerCTPOP(BinaryOperator::createAnd(NotSrc, SrcM1, "", CI), CI);
 303     CI->replaceAllUsesWith(Src);
 304     break;
 305   }
 306
 307   case Intrinsic::stacksave:
 308   case Intrinsic::stackrestore: {
 309     static bool Warned = false;
 310     if (!Warned)
 311       cerr << "WARNING: this target does not support the llvm.stack"
 312            << (Callee->getIntrinsicID() == Intrinsic::stacksave ?
 313                "save" : "restore") << " intrinsic.\n";
 314     Warned = true;
 315     if (Callee->getIntrinsicID() == Intrinsic::stacksave)
 316       CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
 317     break;
 318   }
 319
 320   case Intrinsic::returnaddress:
 321   case Intrinsic::frameaddress:
 322     cerr << "WARNING: this target does not support the llvm."
 323          << (Callee->getIntrinsicID() == Intrinsic::returnaddress ?
 324              "return" : "frame") << "address intrinsic.\n";
 325     CI->replaceAllUsesWith(ConstantPointerNull::get(
 326                                             cast<PointerType>(CI->getType())));
 327     break;
 328
 329   case Intrinsic::prefetch:
 330     break;    // Simply strip out prefetches on unsupported architectures
 331
 332   case Intrinsic::pcmarker:
 333     break;    // Simply strip out pcmarker on unsupported architectures
 334   case Intrinsic::readcyclecounter: {
 335     cerr << "WARNING: this target does not support the llvm.readcyclecoun"
 336          << "ter intrinsic.  It is being lowered to a constant 0\n";
 337     CI->replaceAllUsesWith(ConstantInt::get(Type::Int64Ty, 0));
 338     break;
 339   }
 340
 341   case Intrinsic::dbg_stoppoint:
 342   case Intrinsic::dbg_region_start:
 343   case Intrinsic::dbg_region_end:
 344   case Intrinsic::dbg_func_start:
 345   case Intrinsic::dbg_declare:
 346   case Intrinsic::eh_exception:
 347   case Intrinsic::eh_selector:
 348   case Intrinsic::eh_filter:
 349     break;    // Simply strip out debugging and eh intrinsics
 350
 351   case Intrinsic::memcpy_i32:
 352   case Intrinsic::memcpy_i64: {
 353     static Constant *MemcpyFCache = 0;
 354     Value *Size = CI->getOperand(3);
 355     const Type *IntPtr = TD.getIntPtrType();
 356     if (Size->getType()->getPrimitiveSizeInBits() <
 357         IntPtr->getPrimitiveSizeInBits())
 358       Size = new ZExtInst(Size, IntPtr, "", CI);
 359     else if (Size->getType()->getPrimitiveSizeInBits() >
 360              IntPtr->getPrimitiveSizeInBits())
 361       Size = new TruncInst(Size, IntPtr, "", CI);
 362     Value *Ops[3];
 363     Ops[0] = CI->getOperand(1);
 364     Ops[1] = CI->getOperand(2);
 365     Ops[2] = Size;
 366     ReplaceCallWith("memcpy", CI, Ops, Ops+3, CI->getOperand(1)->getType(),
 367                     MemcpyFCache);
 368     break;
 369   }
 370   case Intrinsic::memmove_i32:
 371   case Intrinsic::memmove_i64: {
 372     static Constant *MemmoveFCache = 0;
 373     Value *Size = CI->getOperand(3);
 374     const Type *IntPtr = TD.getIntPtrType();
 375     if (Size->getType()->getPrimitiveSizeInBits() <
 376         IntPtr->getPrimitiveSizeInBits())
 377       Size = new ZExtInst(Size, IntPtr, "", CI);
 378     else if (Size->getType()->getPrimitiveSizeInBits() >
 379              IntPtr->getPrimitiveSizeInBits())
 380       Size = new TruncInst(Size, IntPtr, "", CI);
 381     Value *Ops[3];
 382     Ops[0] = CI->getOperand(1);
 383     Ops[1] = CI->getOperand(2);
 384     Ops[2] = Size;
 385     ReplaceCallWith("memmove", CI, Ops, Ops+3, CI->getOperand(1)->getType(),
 386                     MemmoveFCache);
 387     break;
 388   }
 389   case Intrinsic::memset_i32:
 390   case Intrinsic::memset_i64: {
 391     static Constant *MemsetFCache = 0;
 392     Value *Size = CI->getOperand(3);
 393     const Type *IntPtr = TD.getIntPtrType();
 394     if (Size->getType()->getPrimitiveSizeInBits() <
 395         IntPtr->getPrimitiveSizeInBits())
 396       Size = new ZExtInst(Size, IntPtr, "", CI);
 397     else if (Size->getType()->getPrimitiveSizeInBits() >
 398              IntPtr->getPrimitiveSizeInBits())
 399       Size = new TruncInst(Size, IntPtr, "", CI);
 400     Value *Ops[3];
 401     Ops[0] = CI->getOperand(1);
 402     // Extend the amount to i32.
 403     Ops[1] = new ZExtInst(CI->getOperand(2), Type::Int32Ty, "", CI);
 404     Ops[2] = Size;
 405     ReplaceCallWith("memset", CI, Ops, Ops+3, CI->getOperand(1)->getType(),
 406                     MemsetFCache);
 407     break;
 408   }
 409   case Intrinsic::sqrt_f32: {
 410     static Constant *sqrtfFCache = 0;
 411     ReplaceCallWith("sqrtf", CI, CI->op_begin()+1, CI->op_end(),
 412                     Type::FloatTy, sqrtfFCache);
 413     break;
 414   }
 415   case Intrinsic::sqrt_f64: {
 416     static Constant *sqrtFCache = 0;
 417     ReplaceCallWith("sqrt", CI, CI->op_begin()+1, CI->op_end(),
 418                     Type::DoubleTy, sqrtFCache);
 419     break;
 420   }
 421   }
 422
 423   assert(CI->use_empty() &&
 424          "Lowering should have eliminated any uses of the intrinsic call!");
 425   CI->eraseFromParent();
 426 }