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
 240
 241 void IntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
 242   Function *Callee = CI->getCalledFunction();
 243   assert(Callee && "Cannot lower an indirect call!");
 244
 245   switch (Callee->getIntrinsicID()) {
 246   case Intrinsic::not_intrinsic:
 247     cerr << "Cannot lower a call to a non-intrinsic function '"
 248          << Callee->getName() << "'!\n";
 249     abort();
 250   default:
 251     cerr << "Error: Code generator does not support intrinsic function '"
 252          << Callee->getName() << "'!\n";
 253     abort();
 254
 255     // The setjmp/longjmp intrinsics should only exist in the code if it was
 256     // never optimized (ie, right out of the CFE), or if it has been hacked on
 257     // by the lowerinvoke pass.  In both cases, the right thing to do is to
 258     // convert the call to an explicit setjmp or longjmp call.
 259   case Intrinsic::setjmp: {
 260     static Constant *SetjmpFCache = 0;
 261     Value *V = ReplaceCallWith("setjmp", CI, CI->op_begin()+1, CI->op_end(),
 262                                Type::Int32Ty, SetjmpFCache);
 263     if (CI->getType() != Type::VoidTy)
 264       CI->replaceAllUsesWith(V);
 265     break;
 266   }
 267   case Intrinsic::sigsetjmp:
 268      if (CI->getType() != Type::VoidTy)
 269        CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
 270      break;
 271
 272   case Intrinsic::longjmp: {
 273     static Constant *LongjmpFCache = 0;
 274     ReplaceCallWith("longjmp", CI, CI->op_begin()+1, CI->op_end(),
 275                     Type::VoidTy, LongjmpFCache);
 276     break;
 277   }
 278
 279   case Intrinsic::siglongjmp: {
 280     // Insert the call to abort
 281     static Constant *AbortFCache = 0;
 282     ReplaceCallWith("abort", CI, CI->op_end(), CI->op_end(),
 283                     Type::VoidTy, AbortFCache);
 284     break;
 285   }
 286   case Intrinsic::ctpop_i8:
 287   case Intrinsic::ctpop_i16:
 288   case Intrinsic::ctpop_i32:
 289   case Intrinsic::ctpop_i64:
 290     CI->replaceAllUsesWith(LowerCTPOP(CI->getOperand(1), CI));
 291     break;
 292
 293   case Intrinsic::bswap_i16:
 294   case Intrinsic::bswap_i32:
 295   case Intrinsic::bswap_i64:
 296     CI->replaceAllUsesWith(LowerBSWAP(CI->getOperand(1), CI));
 297     break;
 298
 299   case Intrinsic::ctlz_i8:
 300   case Intrinsic::ctlz_i16:
 301   case Intrinsic::ctlz_i32:
 302   case Intrinsic::ctlz_i64:
 303     CI->replaceAllUsesWith(LowerCTLZ(CI->getOperand(1), CI));
 304     break;
 305
 306   case Intrinsic::cttz_i8:
 307   case Intrinsic::cttz_i16:
 308   case Intrinsic::cttz_i32:
 309   case Intrinsic::cttz_i64: {
 310     // cttz(x) -> ctpop(~X & (X-1))
 311     Value *Src = CI->getOperand(1);
 312     Value *NotSrc = BinaryOperator::createNot(Src, Src->getName()+".not", CI);
 313     Value *SrcM1  = ConstantInt::get(Src->getType(), 1);
 314     SrcM1 = BinaryOperator::createSub(Src, SrcM1, "", CI);
 315     Src = LowerCTPOP(BinaryOperator::createAnd(NotSrc, SrcM1, "", CI), CI);
 316     CI->replaceAllUsesWith(Src);
 317     break;
 318   }
 319
 320   case Intrinsic::stacksave:
 321   case Intrinsic::stackrestore: {
 322     static bool Warned = false;
 323     if (!Warned)
 324       cerr << "WARNING: this target does not support the llvm.stack"
 325            << (Callee->getIntrinsicID() == Intrinsic::stacksave ?
 326                "save" : "restore") << " intrinsic.\n";
 327     Warned = true;
 328     if (Callee->getIntrinsicID() == Intrinsic::stacksave)
 329       CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
 330     break;
 331   }
 332
 333   case Intrinsic::returnaddress:
 334   case Intrinsic::frameaddress:
 335     cerr << "WARNING: this target does not support the llvm."
 336          << (Callee->getIntrinsicID() == Intrinsic::returnaddress ?
 337              "return" : "frame") << "address intrinsic.\n";
 338     CI->replaceAllUsesWith(ConstantPointerNull::get(
 339                                             cast<PointerType>(CI->getType())));
 340     break;
 341
 342   case Intrinsic::prefetch:
 343     break;    // Simply strip out prefetches on unsupported architectures
 344
 345   case Intrinsic::pcmarker:
 346     break;    // Simply strip out pcmarker on unsupported architectures
 347   case Intrinsic::readcyclecounter: {
 348     cerr << "WARNING: this target does not support the llvm.readcyclecoun"
 349          << "ter intrinsic.  It is being lowered to a constant 0\n";
 350     CI->replaceAllUsesWith(ConstantInt::get(Type::Int64Ty, 0));
 351     break;
 352   }
 353
 354   case Intrinsic::dbg_stoppoint:
 355   case Intrinsic::dbg_region_start:
 356   case Intrinsic::dbg_region_end:
 357   case Intrinsic::dbg_func_start:
 358   case Intrinsic::dbg_declare:
 359   case Intrinsic::eh_exception:
 360   case Intrinsic::eh_selector:
 361   case Intrinsic::eh_filter:
 362     break;    // Simply strip out debugging and eh intrinsics
 363
 364   case Intrinsic::memcpy_i32:
 365   case Intrinsic::memcpy_i64: {
 366     static Constant *MemcpyFCache = 0;
 367     Value *Size = CI->getOperand(3);
 368     const Type *IntPtr = TD.getIntPtrType();
 369     if (Size->getType()->getPrimitiveSizeInBits() <
 370         IntPtr->getPrimitiveSizeInBits())
 371       Size = new ZExtInst(Size, IntPtr, "", CI);
 372     else if (Size->getType()->getPrimitiveSizeInBits() >
 373              IntPtr->getPrimitiveSizeInBits())
 374       Size = new TruncInst(Size, IntPtr, "", CI);
 375     Value *Ops[3];
 376     Ops[0] = CI->getOperand(1);
 377     Ops[1] = CI->getOperand(2);
 378     Ops[2] = Size;
 379     ReplaceCallWith("memcpy", CI, Ops, Ops+3, CI->getOperand(1)->getType(),
 380                     MemcpyFCache);
 381     break;
 382   }
 383   case Intrinsic::memmove_i32:
 384   case Intrinsic::memmove_i64: {
 385     static Constant *MemmoveFCache = 0;
 386     Value *Size = CI->getOperand(3);
 387     const Type *IntPtr = TD.getIntPtrType();
 388     if (Size->getType()->getPrimitiveSizeInBits() <
 389         IntPtr->getPrimitiveSizeInBits())
 390       Size = new ZExtInst(Size, IntPtr, "", CI);
 391     else if (Size->getType()->getPrimitiveSizeInBits() >
 392              IntPtr->getPrimitiveSizeInBits())
 393       Size = new TruncInst(Size, IntPtr, "", CI);
 394     Value *Ops[3];
 395     Ops[0] = CI->getOperand(1);
 396     Ops[1] = CI->getOperand(2);
 397     Ops[2] = Size;
 398     ReplaceCallWith("memmove", CI, Ops, Ops+3, CI->getOperand(1)->getType(),
 399                     MemmoveFCache);
 400     break;
 401   }
 402   case Intrinsic::memset_i32:
 403   case Intrinsic::memset_i64: {
 404     static Constant *MemsetFCache = 0;
 405     Value *Size = CI->getOperand(3);
 406     const Type *IntPtr = TD.getIntPtrType();
 407     if (Size->getType()->getPrimitiveSizeInBits() <
 408         IntPtr->getPrimitiveSizeInBits())
 409       Size = new ZExtInst(Size, IntPtr, "", CI);
 410     else if (Size->getType()->getPrimitiveSizeInBits() >
 411              IntPtr->getPrimitiveSizeInBits())
 412       Size = new TruncInst(Size, IntPtr, "", CI);
 413     Value *Ops[3];
 414     Ops[0] = CI->getOperand(1);
 415     // Extend the amount to i32.
 416     Ops[1] = new ZExtInst(CI->getOperand(2), Type::Int32Ty, "", CI);
 417     Ops[2] = Size;
 418     ReplaceCallWith("memset", CI, Ops, Ops+3, CI->getOperand(1)->getType(),
 419                     MemsetFCache);
 420     break;
 421   }
 422   case Intrinsic::sqrt_f32: {
 423     static Constant *sqrtfFCache = 0;
 424     ReplaceCallWith("sqrtf", CI, CI->op_begin()+1, CI->op_end(),
 425                     Type::FloatTy, sqrtfFCache);
 426     break;
 427   }
 428   case Intrinsic::sqrt_f64: {
 429     static Constant *sqrtFCache = 0;
 430     ReplaceCallWith("sqrt", CI, CI->op_begin()+1, CI->op_end(),
 431                     Type::DoubleTy, sqrtFCache);
 432     break;
 433   }
 434   }
 435
 436   assert(CI->use_empty() &&
 437          "Lowering should have eliminated any uses of the intrinsic call!");
 438   CI->eraseFromParent();
 439 }