IntrinsicLowering.cpp revision dc770929cb2f97397970e2942b746839fc387992
1//===-- IntrinsicLowering.cpp - Intrinsic Lowering default implementation -===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// 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#include "llvm/ADT/STLExtras.h" 24using namespace llvm; 25 26template <class ArgIt> 27static void EnsureFunctionExists(Module &M, const char *Name, 28 ArgIt ArgBegin, ArgIt ArgEnd, 29 const Type *RetTy) { 30 // Insert a correctly-typed definition now. 31 std::vector<const Type *> ParamTys; 32 for (ArgIt I = ArgBegin; I != ArgEnd; ++I) 33 ParamTys.push_back(I->getType()); 34 M.getOrInsertFunction(Name, FunctionType::get(RetTy, ParamTys, false)); 35} 36 37static void EnsureFPIntrinsicsExist(Module &M, Module::iterator I, 38 const char *FName, 39 const char *DName, const char *LDName) { 40 // Insert definitions for all the floating point types. 41 switch((int)I->arg_begin()->getType()->getTypeID()) { 42 case Type::FloatTyID: 43 EnsureFunctionExists(M, FName, I->arg_begin(), I->arg_end(), 44 Type::FloatTy); 45 case Type::DoubleTyID: 46 EnsureFunctionExists(M, DName, I->arg_begin(), I->arg_end(), 47 Type::DoubleTy); 48 case Type::X86_FP80TyID: 49 case Type::FP128TyID: 50 case Type::PPC_FP128TyID: 51 EnsureFunctionExists(M, LDName, I->arg_begin(), I->arg_end(), 52 I->arg_begin()->getType()); 53 } 54} 55 56/// ReplaceCallWith - This function is used when we want to lower an intrinsic 57/// call to a call of an external function. This handles hard cases such as 58/// when there was already a prototype for the external function, and if that 59/// prototype doesn't match the arguments we expect to pass in. 60template <class ArgIt> 61static CallInst *ReplaceCallWith(const char *NewFn, CallInst *CI, 62 ArgIt ArgBegin, ArgIt ArgEnd, 63 const Type *RetTy, Constant *&FCache) { 64 if (!FCache) { 65 // If we haven't already looked up this function, check to see if the 66 // program already contains a function with this name. 67 Module *M = CI->getParent()->getParent()->getParent(); 68 // Get or insert the definition now. 69 std::vector<const Type *> ParamTys; 70 for (ArgIt I = ArgBegin; I != ArgEnd; ++I) 71 ParamTys.push_back((*I)->getType()); 72 FCache = M->getOrInsertFunction(NewFn, 73 FunctionType::get(RetTy, ParamTys, false)); 74 } 75 76 SmallVector<Value *, 8> Args(ArgBegin, ArgEnd); 77 CallInst *NewCI = CallInst::Create(FCache, Args.begin(), Args.end(), 78 CI->getName(), CI); 79 if (!CI->use_empty()) 80 CI->replaceAllUsesWith(NewCI); 81 return NewCI; 82} 83 84void IntrinsicLowering::AddPrototypes(Module &M) { 85 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) 86 if (I->isDeclaration() && !I->use_empty()) 87 switch (I->getIntrinsicID()) { 88 default: break; 89 case Intrinsic::setjmp: 90 EnsureFunctionExists(M, "setjmp", I->arg_begin(), I->arg_end(), 91 Type::Int32Ty); 92 break; 93 case Intrinsic::longjmp: 94 EnsureFunctionExists(M, "longjmp", I->arg_begin(), I->arg_end(), 95 Type::VoidTy); 96 break; 97 case Intrinsic::siglongjmp: 98 EnsureFunctionExists(M, "abort", I->arg_end(), I->arg_end(), 99 Type::VoidTy); 100 break; 101 case Intrinsic::memcpy_i32: 102 case Intrinsic::memcpy_i64: 103 M.getOrInsertFunction("memcpy", PointerType::getUnqual(Type::Int8Ty), 104 PointerType::getUnqual(Type::Int8Ty), 105 PointerType::getUnqual(Type::Int8Ty), 106 TD.getIntPtrType(), (Type *)0); 107 break; 108 case Intrinsic::memmove_i32: 109 case Intrinsic::memmove_i64: 110 M.getOrInsertFunction("memmove", PointerType::getUnqual(Type::Int8Ty), 111 PointerType::getUnqual(Type::Int8Ty), 112 PointerType::getUnqual(Type::Int8Ty), 113 TD.getIntPtrType(), (Type *)0); 114 break; 115 case Intrinsic::memset_i32: 116 case Intrinsic::memset_i64: 117 M.getOrInsertFunction("memset", PointerType::getUnqual(Type::Int8Ty), 118 PointerType::getUnqual(Type::Int8Ty), 119 Type::Int32Ty, 120 TD.getIntPtrType(), (Type *)0); 121 break; 122 case Intrinsic::sqrt: 123 EnsureFPIntrinsicsExist(M, I, "sqrtf", "sqrt", "sqrtl"); 124 break; 125 case Intrinsic::sin: 126 EnsureFPIntrinsicsExist(M, I, "sinf", "sin", "sinl"); 127 break; 128 case Intrinsic::cos: 129 EnsureFPIntrinsicsExist(M, I, "cosf", "cos", "cosl"); 130 break; 131 case Intrinsic::pow: 132 EnsureFPIntrinsicsExist(M, I, "powf", "pow", "powl"); 133 break; 134 case Intrinsic::log: 135 EnsureFPIntrinsicsExist(M, I, "logf", "log", "logl"); 136 break; 137 case Intrinsic::log2: 138 EnsureFPIntrinsicsExist(M, I, "log2f", "log2", "log2l"); 139 break; 140 case Intrinsic::log10: 141 EnsureFPIntrinsicsExist(M, I, "log10f", "log10", "log10l"); 142 break; 143 case Intrinsic::exp: 144 EnsureFPIntrinsicsExist(M, I, "expf", "exp", "expl"); 145 break; 146 case Intrinsic::exp2: 147 EnsureFPIntrinsicsExist(M, I, "exp2f", "exp2", "exp2l"); 148 break; 149 } 150} 151 152/// LowerBSWAP - Emit the code to lower bswap of V before the specified 153/// instruction IP. 154static Value *LowerBSWAP(Value *V, Instruction *IP) { 155 assert(V->getType()->isInteger() && "Can't bswap a non-integer type!"); 156 157 unsigned BitSize = V->getType()->getPrimitiveSizeInBits(); 158 159 switch(BitSize) { 160 default: assert(0 && "Unhandled type size of value to byteswap!"); 161 case 16: { 162 Value *Tmp1 = BinaryOperator::CreateShl(V, 163 ConstantInt::get(V->getType(),8),"bswap.2",IP); 164 Value *Tmp2 = BinaryOperator::CreateLShr(V, 165 ConstantInt::get(V->getType(),8),"bswap.1",IP); 166 V = BinaryOperator::CreateOr(Tmp1, Tmp2, "bswap.i16", IP); 167 break; 168 } 169 case 32: { 170 Value *Tmp4 = BinaryOperator::CreateShl(V, 171 ConstantInt::get(V->getType(),24),"bswap.4", IP); 172 Value *Tmp3 = BinaryOperator::CreateShl(V, 173 ConstantInt::get(V->getType(),8),"bswap.3",IP); 174 Value *Tmp2 = BinaryOperator::CreateLShr(V, 175 ConstantInt::get(V->getType(),8),"bswap.2",IP); 176 Value *Tmp1 = BinaryOperator::CreateLShr(V, 177 ConstantInt::get(V->getType(),24),"bswap.1", IP); 178 Tmp3 = BinaryOperator::CreateAnd(Tmp3, 179 ConstantInt::get(Type::Int32Ty, 0xFF0000), 180 "bswap.and3", IP); 181 Tmp2 = BinaryOperator::CreateAnd(Tmp2, 182 ConstantInt::get(Type::Int32Ty, 0xFF00), 183 "bswap.and2", IP); 184 Tmp4 = BinaryOperator::CreateOr(Tmp4, Tmp3, "bswap.or1", IP); 185 Tmp2 = BinaryOperator::CreateOr(Tmp2, Tmp1, "bswap.or2", IP); 186 V = BinaryOperator::CreateOr(Tmp4, Tmp2, "bswap.i32", IP); 187 break; 188 } 189 case 64: { 190 Value *Tmp8 = BinaryOperator::CreateShl(V, 191 ConstantInt::get(V->getType(),56),"bswap.8", IP); 192 Value *Tmp7 = BinaryOperator::CreateShl(V, 193 ConstantInt::get(V->getType(),40),"bswap.7", IP); 194 Value *Tmp6 = BinaryOperator::CreateShl(V, 195 ConstantInt::get(V->getType(),24),"bswap.6", IP); 196 Value *Tmp5 = BinaryOperator::CreateShl(V, 197 ConstantInt::get(V->getType(),8),"bswap.5", IP); 198 Value* Tmp4 = BinaryOperator::CreateLShr(V, 199 ConstantInt::get(V->getType(),8),"bswap.4", IP); 200 Value* Tmp3 = BinaryOperator::CreateLShr(V, 201 ConstantInt::get(V->getType(),24),"bswap.3", IP); 202 Value* Tmp2 = BinaryOperator::CreateLShr(V, 203 ConstantInt::get(V->getType(),40),"bswap.2", IP); 204 Value* Tmp1 = BinaryOperator::CreateLShr(V, 205 ConstantInt::get(V->getType(),56),"bswap.1", IP); 206 Tmp7 = BinaryOperator::CreateAnd(Tmp7, 207 ConstantInt::get(Type::Int64Ty, 208 0xFF000000000000ULL), 209 "bswap.and7", IP); 210 Tmp6 = BinaryOperator::CreateAnd(Tmp6, 211 ConstantInt::get(Type::Int64Ty, 0xFF0000000000ULL), 212 "bswap.and6", IP); 213 Tmp5 = BinaryOperator::CreateAnd(Tmp5, 214 ConstantInt::get(Type::Int64Ty, 0xFF00000000ULL), 215 "bswap.and5", IP); 216 Tmp4 = BinaryOperator::CreateAnd(Tmp4, 217 ConstantInt::get(Type::Int64Ty, 0xFF000000ULL), 218 "bswap.and4", IP); 219 Tmp3 = BinaryOperator::CreateAnd(Tmp3, 220 ConstantInt::get(Type::Int64Ty, 0xFF0000ULL), 221 "bswap.and3", IP); 222 Tmp2 = BinaryOperator::CreateAnd(Tmp2, 223 ConstantInt::get(Type::Int64Ty, 0xFF00ULL), 224 "bswap.and2", IP); 225 Tmp8 = BinaryOperator::CreateOr(Tmp8, Tmp7, "bswap.or1", IP); 226 Tmp6 = BinaryOperator::CreateOr(Tmp6, Tmp5, "bswap.or2", IP); 227 Tmp4 = BinaryOperator::CreateOr(Tmp4, Tmp3, "bswap.or3", IP); 228 Tmp2 = BinaryOperator::CreateOr(Tmp2, Tmp1, "bswap.or4", IP); 229 Tmp8 = BinaryOperator::CreateOr(Tmp8, Tmp6, "bswap.or5", IP); 230 Tmp4 = BinaryOperator::CreateOr(Tmp4, Tmp2, "bswap.or6", IP); 231 V = BinaryOperator::CreateOr(Tmp8, Tmp4, "bswap.i64", IP); 232 break; 233 } 234 } 235 return V; 236} 237 238/// LowerCTPOP - Emit the code to lower ctpop of V before the specified 239/// instruction IP. 240static Value *LowerCTPOP(Value *V, Instruction *IP) { 241 assert(V->getType()->isInteger() && "Can't ctpop a non-integer type!"); 242 243 static const uint64_t MaskValues[6] = { 244 0x5555555555555555ULL, 0x3333333333333333ULL, 245 0x0F0F0F0F0F0F0F0FULL, 0x00FF00FF00FF00FFULL, 246 0x0000FFFF0000FFFFULL, 0x00000000FFFFFFFFULL 247 }; 248 249 unsigned BitSize = V->getType()->getPrimitiveSizeInBits(); 250 unsigned WordSize = (BitSize + 63) / 64; 251 Value *Count = ConstantInt::get(V->getType(), 0); 252 253 for (unsigned n = 0; n < WordSize; ++n) { 254 Value *PartValue = V; 255 for (unsigned i = 1, ct = 0; i < (BitSize>64 ? 64 : BitSize); 256 i <<= 1, ++ct) { 257 Value *MaskCst = ConstantInt::get(V->getType(), MaskValues[ct]); 258 Value *LHS = BinaryOperator::CreateAnd( 259 PartValue, MaskCst, "cppop.and1", IP); 260 Value *VShift = BinaryOperator::CreateLShr(PartValue, 261 ConstantInt::get(V->getType(), i), "ctpop.sh", IP); 262 Value *RHS = BinaryOperator::CreateAnd(VShift, MaskCst, "cppop.and2", IP); 263 PartValue = BinaryOperator::CreateAdd(LHS, RHS, "ctpop.step", IP); 264 } 265 Count = BinaryOperator::CreateAdd(PartValue, Count, "ctpop.part", IP); 266 if (BitSize > 64) { 267 V = BinaryOperator::CreateLShr(V, ConstantInt::get(V->getType(), 64), 268 "ctpop.part.sh", IP); 269 BitSize -= 64; 270 } 271 } 272 273 return Count; 274} 275 276/// LowerCTLZ - Emit the code to lower ctlz of V before the specified 277/// instruction IP. 278static Value *LowerCTLZ(Value *V, Instruction *IP) { 279 280 unsigned BitSize = V->getType()->getPrimitiveSizeInBits(); 281 for (unsigned i = 1; i < BitSize; i <<= 1) { 282 Value *ShVal = ConstantInt::get(V->getType(), i); 283 ShVal = BinaryOperator::CreateLShr(V, ShVal, "ctlz.sh", IP); 284 V = BinaryOperator::CreateOr(V, ShVal, "ctlz.step", IP); 285 } 286 287 V = BinaryOperator::CreateNot(V, "", IP); 288 return LowerCTPOP(V, IP); 289} 290 291/// Convert the llvm.part.select.iX.iY intrinsic. This intrinsic takes 292/// three integer arguments. The first argument is the Value from which the 293/// bits will be selected. It may be of any bit width. The second and third 294/// arguments specify a range of bits to select with the second argument 295/// specifying the low bit and the third argument specifying the high bit. Both 296/// must be type i32. The result is the corresponding selected bits from the 297/// Value in the same width as the Value (first argument). If the low bit index 298/// is higher than the high bit index then the inverse selection is done and 299/// the bits are returned in inverse order. 300/// @brief Lowering of llvm.part.select intrinsic. 301static Instruction *LowerPartSelect(CallInst *CI) { 302 // Make sure we're dealing with a part select intrinsic here 303 Function *F = CI->getCalledFunction(); 304 const FunctionType *FT = F->getFunctionType(); 305 if (!F->isDeclaration() || !FT->getReturnType()->isInteger() || 306 FT->getNumParams() != 3 || !FT->getParamType(0)->isInteger() || 307 !FT->getParamType(1)->isInteger() || !FT->getParamType(2)->isInteger()) 308 return CI; 309 310 // Get the intrinsic implementation function by converting all the . to _ 311 // in the intrinsic's function name and then reconstructing the function 312 // declaration. 313 std::string Name(F->getName()); 314 for (unsigned i = 4; i < Name.length(); ++i) 315 if (Name[i] == '.') 316 Name[i] = '_'; 317 Module* M = F->getParent(); 318 F = cast<Function>(M->getOrInsertFunction(Name, FT)); 319 F->setLinkage(GlobalValue::WeakLinkage); 320 321 // If we haven't defined the impl function yet, do so now 322 if (F->isDeclaration()) { 323 324 // Get the arguments to the function 325 Function::arg_iterator args = F->arg_begin(); 326 Value* Val = args++; Val->setName("Val"); 327 Value* Lo = args++; Lo->setName("Lo"); 328 Value* Hi = args++; Hi->setName("High"); 329 330 // We want to select a range of bits here such that [Hi, Lo] is shifted 331 // down to the low bits. However, it is quite possible that Hi is smaller 332 // than Lo in which case the bits have to be reversed. 333 334 // Create the blocks we will need for the two cases (forward, reverse) 335 BasicBlock* CurBB = BasicBlock::Create("entry", F); 336 BasicBlock *RevSize = BasicBlock::Create("revsize", CurBB->getParent()); 337 BasicBlock *FwdSize = BasicBlock::Create("fwdsize", CurBB->getParent()); 338 BasicBlock *Compute = BasicBlock::Create("compute", CurBB->getParent()); 339 BasicBlock *Reverse = BasicBlock::Create("reverse", CurBB->getParent()); 340 BasicBlock *RsltBlk = BasicBlock::Create("result", CurBB->getParent()); 341 342 // Cast Hi and Lo to the size of Val so the widths are all the same 343 if (Hi->getType() != Val->getType()) 344 Hi = CastInst::CreateIntegerCast(Hi, Val->getType(), false, 345 "tmp", CurBB); 346 if (Lo->getType() != Val->getType()) 347 Lo = CastInst::CreateIntegerCast(Lo, Val->getType(), false, 348 "tmp", CurBB); 349 350 // Compute a few things that both cases will need, up front. 351 Constant* Zero = ConstantInt::get(Val->getType(), 0); 352 Constant* One = ConstantInt::get(Val->getType(), 1); 353 Constant* AllOnes = ConstantInt::getAllOnesValue(Val->getType()); 354 355 // Compare the Hi and Lo bit positions. This is used to determine 356 // which case we have (forward or reverse) 357 ICmpInst *Cmp = new ICmpInst(ICmpInst::ICMP_ULT, Hi, Lo, "less",CurBB); 358 BranchInst::Create(RevSize, FwdSize, Cmp, CurBB); 359 360 // First, copmute the number of bits in the forward case. 361 Instruction* FBitSize = 362 BinaryOperator::CreateSub(Hi, Lo,"fbits", FwdSize); 363 BranchInst::Create(Compute, FwdSize); 364 365 // Second, compute the number of bits in the reverse case. 366 Instruction* RBitSize = 367 BinaryOperator::CreateSub(Lo, Hi, "rbits", RevSize); 368 BranchInst::Create(Compute, RevSize); 369 370 // Now, compute the bit range. Start by getting the bitsize and the shift 371 // amount (either Hi or Lo) from PHI nodes. Then we compute a mask for 372 // the number of bits we want in the range. We shift the bits down to the 373 // least significant bits, apply the mask to zero out unwanted high bits, 374 // and we have computed the "forward" result. It may still need to be 375 // reversed. 376 377 // Get the BitSize from one of the two subtractions 378 PHINode *BitSize = PHINode::Create(Val->getType(), "bits", Compute); 379 BitSize->reserveOperandSpace(2); 380 BitSize->addIncoming(FBitSize, FwdSize); 381 BitSize->addIncoming(RBitSize, RevSize); 382 383 // Get the ShiftAmount as the smaller of Hi/Lo 384 PHINode *ShiftAmt = PHINode::Create(Val->getType(), "shiftamt", Compute); 385 ShiftAmt->reserveOperandSpace(2); 386 ShiftAmt->addIncoming(Lo, FwdSize); 387 ShiftAmt->addIncoming(Hi, RevSize); 388 389 // Increment the bit size 390 Instruction *BitSizePlusOne = 391 BinaryOperator::CreateAdd(BitSize, One, "bits", Compute); 392 393 // Create a Mask to zero out the high order bits. 394 Instruction* Mask = 395 BinaryOperator::CreateShl(AllOnes, BitSizePlusOne, "mask", Compute); 396 Mask = BinaryOperator::CreateNot(Mask, "mask", Compute); 397 398 // Shift the bits down and apply the mask 399 Instruction* FRes = 400 BinaryOperator::CreateLShr(Val, ShiftAmt, "fres", Compute); 401 FRes = BinaryOperator::CreateAnd(FRes, Mask, "fres", Compute); 402 BranchInst::Create(Reverse, RsltBlk, Cmp, Compute); 403 404 // In the Reverse block we have the mask already in FRes but we must reverse 405 // it by shifting FRes bits right and putting them in RRes by shifting them 406 // in from left. 407 408 // First set up our loop counters 409 PHINode *Count = PHINode::Create(Val->getType(), "count", Reverse); 410 Count->reserveOperandSpace(2); 411 Count->addIncoming(BitSizePlusOne, Compute); 412 413 // Next, get the value that we are shifting. 414 PHINode *BitsToShift = PHINode::Create(Val->getType(), "val", Reverse); 415 BitsToShift->reserveOperandSpace(2); 416 BitsToShift->addIncoming(FRes, Compute); 417 418 // Finally, get the result of the last computation 419 PHINode *RRes = PHINode::Create(Val->getType(), "rres", Reverse); 420 RRes->reserveOperandSpace(2); 421 RRes->addIncoming(Zero, Compute); 422 423 // Decrement the counter 424 Instruction *Decr = BinaryOperator::CreateSub(Count, One, "decr", Reverse); 425 Count->addIncoming(Decr, Reverse); 426 427 // Compute the Bit that we want to move 428 Instruction *Bit = 429 BinaryOperator::CreateAnd(BitsToShift, One, "bit", Reverse); 430 431 // Compute the new value for next iteration. 432 Instruction *NewVal = 433 BinaryOperator::CreateLShr(BitsToShift, One, "rshift", Reverse); 434 BitsToShift->addIncoming(NewVal, Reverse); 435 436 // Shift the bit into the low bits of the result. 437 Instruction *NewRes = 438 BinaryOperator::CreateShl(RRes, One, "lshift", Reverse); 439 NewRes = BinaryOperator::CreateOr(NewRes, Bit, "addbit", Reverse); 440 RRes->addIncoming(NewRes, Reverse); 441 442 // Terminate loop if we've moved all the bits. 443 ICmpInst *Cond = 444 new ICmpInst(ICmpInst::ICMP_EQ, Decr, Zero, "cond", Reverse); 445 BranchInst::Create(RsltBlk, Reverse, Cond, Reverse); 446 447 // Finally, in the result block, select one of the two results with a PHI 448 // node and return the result; 449 CurBB = RsltBlk; 450 PHINode *BitSelect = PHINode::Create(Val->getType(), "part_select", CurBB); 451 BitSelect->reserveOperandSpace(2); 452 BitSelect->addIncoming(FRes, Compute); 453 BitSelect->addIncoming(NewRes, Reverse); 454 ReturnInst::Create(BitSelect, CurBB); 455 } 456 457 // Return a call to the implementation function 458 Value *Args[] = { 459 CI->getOperand(1), 460 CI->getOperand(2), 461 CI->getOperand(3) 462 }; 463 return CallInst::Create(F, Args, array_endof(Args), CI->getName(), CI); 464} 465 466/// Convert the llvm.part.set.iX.iY.iZ intrinsic. This intrinsic takes 467/// four integer arguments (iAny %Value, iAny %Replacement, i32 %Low, i32 %High) 468/// The first two arguments can be any bit width. The result is the same width 469/// as %Value. The operation replaces bits between %Low and %High with the value 470/// in %Replacement. If %Replacement is not the same width, it is truncated or 471/// zero extended as appropriate to fit the bits being replaced. If %Low is 472/// greater than %High then the inverse set of bits are replaced. 473/// @brief Lowering of llvm.bit.part.set intrinsic. 474static Instruction *LowerPartSet(CallInst *CI) { 475 // Make sure we're dealing with a part select intrinsic here 476 Function *F = CI->getCalledFunction(); 477 const FunctionType *FT = F->getFunctionType(); 478 if (!F->isDeclaration() || !FT->getReturnType()->isInteger() || 479 FT->getNumParams() != 4 || !FT->getParamType(0)->isInteger() || 480 !FT->getParamType(1)->isInteger() || !FT->getParamType(2)->isInteger() || 481 !FT->getParamType(3)->isInteger()) 482 return CI; 483 484 // Get the intrinsic implementation function by converting all the . to _ 485 // in the intrinsic's function name and then reconstructing the function 486 // declaration. 487 std::string Name(F->getName()); 488 for (unsigned i = 4; i < Name.length(); ++i) 489 if (Name[i] == '.') 490 Name[i] = '_'; 491 Module* M = F->getParent(); 492 F = cast<Function>(M->getOrInsertFunction(Name, FT)); 493 F->setLinkage(GlobalValue::WeakLinkage); 494 495 // If we haven't defined the impl function yet, do so now 496 if (F->isDeclaration()) { 497 // Get the arguments for the function. 498 Function::arg_iterator args = F->arg_begin(); 499 Value* Val = args++; Val->setName("Val"); 500 Value* Rep = args++; Rep->setName("Rep"); 501 Value* Lo = args++; Lo->setName("Lo"); 502 Value* Hi = args++; Hi->setName("Hi"); 503 504 // Get some types we need 505 const IntegerType* ValTy = cast<IntegerType>(Val->getType()); 506 const IntegerType* RepTy = cast<IntegerType>(Rep->getType()); 507 uint32_t ValBits = ValTy->getBitWidth(); 508 uint32_t RepBits = RepTy->getBitWidth(); 509 510 // Constant Definitions 511 ConstantInt* RepBitWidth = ConstantInt::get(Type::Int32Ty, RepBits); 512 ConstantInt* RepMask = ConstantInt::getAllOnesValue(RepTy); 513 ConstantInt* ValMask = ConstantInt::getAllOnesValue(ValTy); 514 ConstantInt* One = ConstantInt::get(Type::Int32Ty, 1); 515 ConstantInt* ValOne = ConstantInt::get(ValTy, 1); 516 ConstantInt* Zero = ConstantInt::get(Type::Int32Ty, 0); 517 ConstantInt* ValZero = ConstantInt::get(ValTy, 0); 518 519 // Basic blocks we fill in below. 520 BasicBlock* entry = BasicBlock::Create("entry", F, 0); 521 BasicBlock* large = BasicBlock::Create("large", F, 0); 522 BasicBlock* small = BasicBlock::Create("small", F, 0); 523 BasicBlock* reverse = BasicBlock::Create("reverse", F, 0); 524 BasicBlock* result = BasicBlock::Create("result", F, 0); 525 526 // BASIC BLOCK: entry 527 // First, get the number of bits that we're placing as an i32 528 ICmpInst* is_forward = 529 new ICmpInst(ICmpInst::ICMP_ULT, Lo, Hi, "", entry); 530 SelectInst* Hi_pn = SelectInst::Create(is_forward, Hi, Lo, "", entry); 531 SelectInst* Lo_pn = SelectInst::Create(is_forward, Lo, Hi, "", entry); 532 BinaryOperator* NumBits = BinaryOperator::CreateSub(Hi_pn, Lo_pn, "",entry); 533 NumBits = BinaryOperator::CreateAdd(NumBits, One, "", entry); 534 // Now, convert Lo and Hi to ValTy bit width 535 if (ValBits > 32) { 536 Lo = new ZExtInst(Lo_pn, ValTy, "", entry); 537 } else if (ValBits < 32) { 538 Lo = new TruncInst(Lo_pn, ValTy, "", entry); 539 } 540 // Determine if the replacement bits are larger than the number of bits we 541 // are replacing and deal with it. 542 ICmpInst* is_large = 543 new ICmpInst(ICmpInst::ICMP_ULT, NumBits, RepBitWidth, "", entry); 544 BranchInst::Create(large, small, is_large, entry); 545 546 // BASIC BLOCK: large 547 Instruction* MaskBits = 548 BinaryOperator::CreateSub(RepBitWidth, NumBits, "", large); 549 MaskBits = CastInst::CreateIntegerCast(MaskBits, RepMask->getType(), 550 false, "", large); 551 BinaryOperator* Mask1 = 552 BinaryOperator::CreateLShr(RepMask, MaskBits, "", large); 553 BinaryOperator* Rep2 = BinaryOperator::CreateAnd(Mask1, Rep, "", large); 554 BranchInst::Create(small, large); 555 556 // BASIC BLOCK: small 557 PHINode* Rep3 = PHINode::Create(RepTy, "", small); 558 Rep3->reserveOperandSpace(2); 559 Rep3->addIncoming(Rep2, large); 560 Rep3->addIncoming(Rep, entry); 561 Value* Rep4 = Rep3; 562 if (ValBits > RepBits) 563 Rep4 = new ZExtInst(Rep3, ValTy, "", small); 564 else if (ValBits < RepBits) 565 Rep4 = new TruncInst(Rep3, ValTy, "", small); 566 BranchInst::Create(result, reverse, is_forward, small); 567 568 // BASIC BLOCK: reverse (reverses the bits of the replacement) 569 // Set up our loop counter as a PHI so we can decrement on each iteration. 570 // We will loop for the number of bits in the replacement value. 571 PHINode *Count = PHINode::Create(Type::Int32Ty, "count", reverse); 572 Count->reserveOperandSpace(2); 573 Count->addIncoming(NumBits, small); 574 575 // Get the value that we are shifting bits out of as a PHI because 576 // we'll change this with each iteration. 577 PHINode *BitsToShift = PHINode::Create(Val->getType(), "val", reverse); 578 BitsToShift->reserveOperandSpace(2); 579 BitsToShift->addIncoming(Rep4, small); 580 581 // Get the result of the last computation or zero on first iteration 582 PHINode *RRes = PHINode::Create(Val->getType(), "rres", reverse); 583 RRes->reserveOperandSpace(2); 584 RRes->addIncoming(ValZero, small); 585 586 // Decrement the loop counter by one 587 Instruction *Decr = BinaryOperator::CreateSub(Count, One, "", reverse); 588 Count->addIncoming(Decr, reverse); 589 590 // Get the bit that we want to move into the result 591 Value *Bit = BinaryOperator::CreateAnd(BitsToShift, ValOne, "", reverse); 592 593 // Compute the new value of the bits to shift for the next iteration. 594 Value *NewVal = BinaryOperator::CreateLShr(BitsToShift, ValOne,"", reverse); 595 BitsToShift->addIncoming(NewVal, reverse); 596 597 // Shift the bit we extracted into the low bit of the result. 598 Instruction *NewRes = BinaryOperator::CreateShl(RRes, ValOne, "", reverse); 599 NewRes = BinaryOperator::CreateOr(NewRes, Bit, "", reverse); 600 RRes->addIncoming(NewRes, reverse); 601 602 // Terminate loop if we've moved all the bits. 603 ICmpInst *Cond = new ICmpInst(ICmpInst::ICMP_EQ, Decr, Zero, "", reverse); 604 BranchInst::Create(result, reverse, Cond, reverse); 605 606 // BASIC BLOCK: result 607 PHINode *Rplcmnt = PHINode::Create(Val->getType(), "", result); 608 Rplcmnt->reserveOperandSpace(2); 609 Rplcmnt->addIncoming(NewRes, reverse); 610 Rplcmnt->addIncoming(Rep4, small); 611 Value* t0 = CastInst::CreateIntegerCast(NumBits,ValTy,false,"",result); 612 Value* t1 = BinaryOperator::CreateShl(ValMask, Lo, "", result); 613 Value* t2 = BinaryOperator::CreateNot(t1, "", result); 614 Value* t3 = BinaryOperator::CreateShl(t1, t0, "", result); 615 Value* t4 = BinaryOperator::CreateOr(t2, t3, "", result); 616 Value* t5 = BinaryOperator::CreateAnd(t4, Val, "", result); 617 Value* t6 = BinaryOperator::CreateShl(Rplcmnt, Lo, "", result); 618 Value* Rslt = BinaryOperator::CreateOr(t5, t6, "part_set", result); 619 ReturnInst::Create(Rslt, result); 620 } 621 622 // Return a call to the implementation function 623 Value *Args[] = { 624 CI->getOperand(1), 625 CI->getOperand(2), 626 CI->getOperand(3), 627 CI->getOperand(4) 628 }; 629 return CallInst::Create(F, Args, array_endof(Args), CI->getName(), CI); 630} 631 632static void ReplaceFPIntrinsicWithCall(CallInst *CI, Constant *FCache, 633 Constant *DCache, Constant *LDCache, 634 const char *Fname, const char *Dname, 635 const char *LDname) { 636 switch (CI->getOperand(1)->getType()->getTypeID()) { 637 default: assert(0 && "Invalid type in intrinsic"); abort(); 638 case Type::FloatTyID: 639 ReplaceCallWith(Fname, CI, CI->op_begin()+1, CI->op_end(), 640 Type::FloatTy, FCache); 641 break; 642 case Type::DoubleTyID: 643 ReplaceCallWith(Dname, CI, CI->op_begin()+1, CI->op_end(), 644 Type::DoubleTy, DCache); 645 break; 646 case Type::X86_FP80TyID: 647 case Type::FP128TyID: 648 case Type::PPC_FP128TyID: 649 ReplaceCallWith(LDname, CI, CI->op_begin()+1, CI->op_end(), 650 CI->getOperand(1)->getType(), LDCache); 651 break; 652 } 653} 654 655void IntrinsicLowering::LowerIntrinsicCall(CallInst *CI) { 656 Function *Callee = CI->getCalledFunction(); 657 assert(Callee && "Cannot lower an indirect call!"); 658 659 switch (Callee->getIntrinsicID()) { 660 case Intrinsic::not_intrinsic: 661 cerr << "Cannot lower a call to a non-intrinsic function '" 662 << Callee->getName() << "'!\n"; 663 abort(); 664 default: 665 cerr << "Error: Code generator does not support intrinsic function '" 666 << Callee->getName() << "'!\n"; 667 abort(); 668 669 // The setjmp/longjmp intrinsics should only exist in the code if it was 670 // never optimized (ie, right out of the CFE), or if it has been hacked on 671 // by the lowerinvoke pass. In both cases, the right thing to do is to 672 // convert the call to an explicit setjmp or longjmp call. 673 case Intrinsic::setjmp: { 674 static Constant *SetjmpFCache = 0; 675 Value *V = ReplaceCallWith("setjmp", CI, CI->op_begin()+1, CI->op_end(), 676 Type::Int32Ty, SetjmpFCache); 677 if (CI->getType() != Type::VoidTy) 678 CI->replaceAllUsesWith(V); 679 break; 680 } 681 case Intrinsic::sigsetjmp: 682 if (CI->getType() != Type::VoidTy) 683 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType())); 684 break; 685 686 case Intrinsic::longjmp: { 687 static Constant *LongjmpFCache = 0; 688 ReplaceCallWith("longjmp", CI, CI->op_begin()+1, CI->op_end(), 689 Type::VoidTy, LongjmpFCache); 690 break; 691 } 692 693 case Intrinsic::siglongjmp: { 694 // Insert the call to abort 695 static Constant *AbortFCache = 0; 696 ReplaceCallWith("abort", CI, CI->op_end(), CI->op_end(), 697 Type::VoidTy, AbortFCache); 698 break; 699 } 700 case Intrinsic::ctpop: 701 CI->replaceAllUsesWith(LowerCTPOP(CI->getOperand(1), CI)); 702 break; 703 704 case Intrinsic::bswap: 705 CI->replaceAllUsesWith(LowerBSWAP(CI->getOperand(1), CI)); 706 break; 707 708 case Intrinsic::ctlz: 709 CI->replaceAllUsesWith(LowerCTLZ(CI->getOperand(1), CI)); 710 break; 711 712 case Intrinsic::cttz: { 713 // cttz(x) -> ctpop(~X & (X-1)) 714 Value *Src = CI->getOperand(1); 715 Value *NotSrc = BinaryOperator::CreateNot(Src, Src->getName()+".not", CI); 716 Value *SrcM1 = ConstantInt::get(Src->getType(), 1); 717 SrcM1 = BinaryOperator::CreateSub(Src, SrcM1, "", CI); 718 Src = LowerCTPOP(BinaryOperator::CreateAnd(NotSrc, SrcM1, "", CI), CI); 719 CI->replaceAllUsesWith(Src); 720 break; 721 } 722 723 case Intrinsic::part_select: 724 CI->replaceAllUsesWith(LowerPartSelect(CI)); 725 break; 726 727 case Intrinsic::part_set: 728 CI->replaceAllUsesWith(LowerPartSet(CI)); 729 break; 730 731 case Intrinsic::stacksave: 732 case Intrinsic::stackrestore: { 733 static bool Warned = false; 734 if (!Warned) 735 cerr << "WARNING: this target does not support the llvm.stack" 736 << (Callee->getIntrinsicID() == Intrinsic::stacksave ? 737 "save" : "restore") << " intrinsic.\n"; 738 Warned = true; 739 if (Callee->getIntrinsicID() == Intrinsic::stacksave) 740 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType())); 741 break; 742 } 743 744 case Intrinsic::returnaddress: 745 case Intrinsic::frameaddress: 746 cerr << "WARNING: this target does not support the llvm." 747 << (Callee->getIntrinsicID() == Intrinsic::returnaddress ? 748 "return" : "frame") << "address intrinsic.\n"; 749 CI->replaceAllUsesWith(ConstantPointerNull::get( 750 cast<PointerType>(CI->getType()))); 751 break; 752 753 case Intrinsic::prefetch: 754 break; // Simply strip out prefetches on unsupported architectures 755 756 case Intrinsic::pcmarker: 757 break; // Simply strip out pcmarker on unsupported architectures 758 case Intrinsic::readcyclecounter: { 759 cerr << "WARNING: this target does not support the llvm.readcyclecoun" 760 << "ter intrinsic. It is being lowered to a constant 0\n"; 761 CI->replaceAllUsesWith(ConstantInt::get(Type::Int64Ty, 0)); 762 break; 763 } 764 765 case Intrinsic::dbg_stoppoint: 766 case Intrinsic::dbg_region_start: 767 case Intrinsic::dbg_region_end: 768 case Intrinsic::dbg_func_start: 769 case Intrinsic::dbg_declare: 770 break; // Simply strip out debugging intrinsics 771 772 case Intrinsic::eh_exception: 773 case Intrinsic::eh_selector_i32: 774 case Intrinsic::eh_selector_i64: 775 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType())); 776 break; 777 778 case Intrinsic::eh_typeid_for_i32: 779 case Intrinsic::eh_typeid_for_i64: 780 // Return something different to eh_selector. 781 CI->replaceAllUsesWith(ConstantInt::get(CI->getType(), 1)); 782 break; 783 784 case Intrinsic::var_annotation: 785 break; // Strip out annotate intrinsic 786 787 case Intrinsic::memcpy_i32: 788 case Intrinsic::memcpy_i64: { 789 static Constant *MemcpyFCache = 0; 790 Value *Size = CI->getOperand(3); 791 const Type *IntPtr = TD.getIntPtrType(); 792 if (Size->getType()->getPrimitiveSizeInBits() < 793 IntPtr->getPrimitiveSizeInBits()) 794 Size = new ZExtInst(Size, IntPtr, "", CI); 795 else if (Size->getType()->getPrimitiveSizeInBits() > 796 IntPtr->getPrimitiveSizeInBits()) 797 Size = new TruncInst(Size, IntPtr, "", CI); 798 Value *Ops[3]; 799 Ops[0] = CI->getOperand(1); 800 Ops[1] = CI->getOperand(2); 801 Ops[2] = Size; 802 ReplaceCallWith("memcpy", CI, Ops, Ops+3, CI->getOperand(1)->getType(), 803 MemcpyFCache); 804 break; 805 } 806 case Intrinsic::memmove_i32: 807 case Intrinsic::memmove_i64: { 808 static Constant *MemmoveFCache = 0; 809 Value *Size = CI->getOperand(3); 810 const Type *IntPtr = TD.getIntPtrType(); 811 if (Size->getType()->getPrimitiveSizeInBits() < 812 IntPtr->getPrimitiveSizeInBits()) 813 Size = new ZExtInst(Size, IntPtr, "", CI); 814 else if (Size->getType()->getPrimitiveSizeInBits() > 815 IntPtr->getPrimitiveSizeInBits()) 816 Size = new TruncInst(Size, IntPtr, "", CI); 817 Value *Ops[3]; 818 Ops[0] = CI->getOperand(1); 819 Ops[1] = CI->getOperand(2); 820 Ops[2] = Size; 821 ReplaceCallWith("memmove", CI, Ops, Ops+3, CI->getOperand(1)->getType(), 822 MemmoveFCache); 823 break; 824 } 825 case Intrinsic::memset_i32: 826 case Intrinsic::memset_i64: { 827 static Constant *MemsetFCache = 0; 828 Value *Size = CI->getOperand(3); 829 const Type *IntPtr = TD.getIntPtrType(); 830 if (Size->getType()->getPrimitiveSizeInBits() < 831 IntPtr->getPrimitiveSizeInBits()) 832 Size = new ZExtInst(Size, IntPtr, "", CI); 833 else if (Size->getType()->getPrimitiveSizeInBits() > 834 IntPtr->getPrimitiveSizeInBits()) 835 Size = new TruncInst(Size, IntPtr, "", CI); 836 Value *Ops[3]; 837 Ops[0] = CI->getOperand(1); 838 // Extend the amount to i32. 839 Ops[1] = new ZExtInst(CI->getOperand(2), Type::Int32Ty, "", CI); 840 Ops[2] = Size; 841 ReplaceCallWith("memset", CI, Ops, Ops+3, CI->getOperand(1)->getType(), 842 MemsetFCache); 843 break; 844 } 845 case Intrinsic::sqrt: { 846 static Constant *sqrtFCache = 0; 847 static Constant *sqrtDCache = 0; 848 static Constant *sqrtLDCache = 0; 849 ReplaceFPIntrinsicWithCall(CI, sqrtFCache, sqrtDCache, sqrtLDCache, 850 "sqrtf", "sqrt", "sqrtl"); 851 break; 852 } 853 case Intrinsic::log: { 854 static Constant *logFCache = 0; 855 static Constant *logDCache = 0; 856 static Constant *logLDCache = 0; 857 ReplaceFPIntrinsicWithCall(CI, logFCache, logDCache, logLDCache, 858 "logf", "log", "logl"); 859 break; 860 } 861 case Intrinsic::log2: { 862 static Constant *log2FCache = 0; 863 static Constant *log2DCache = 0; 864 static Constant *log2LDCache = 0; 865 ReplaceFPIntrinsicWithCall(CI, log2FCache, log2DCache, log2LDCache, 866 "log2f", "log2", "log2l"); 867 break; 868 } 869 case Intrinsic::log10: { 870 static Constant *log10FCache = 0; 871 static Constant *log10DCache = 0; 872 static Constant *log10LDCache = 0; 873 ReplaceFPIntrinsicWithCall(CI, log10FCache, log10DCache, log10LDCache, 874 "log10f", "log10", "log10l"); 875 break; 876 } 877 case Intrinsic::exp: { 878 static Constant *expFCache = 0; 879 static Constant *expDCache = 0; 880 static Constant *expLDCache = 0; 881 ReplaceFPIntrinsicWithCall(CI, expFCache, expDCache, expLDCache, 882 "expf", "exp", "expl"); 883 break; 884 } 885 case Intrinsic::exp2: { 886 static Constant *exp2FCache = 0; 887 static Constant *exp2DCache = 0; 888 static Constant *exp2LDCache = 0; 889 ReplaceFPIntrinsicWithCall(CI, exp2FCache, exp2DCache, exp2LDCache, 890 "exp2f", "exp2", "exp2l"); 891 break; 892 } 893 case Intrinsic::pow: { 894 static Constant *powFCache = 0; 895 static Constant *powDCache = 0; 896 static Constant *powLDCache = 0; 897 ReplaceFPIntrinsicWithCall(CI, powFCache, powDCache, powLDCache, 898 "powf", "pow", "powl"); 899 break; 900 } 901 case Intrinsic::flt_rounds: 902 // Lower to "round to the nearest" 903 if (CI->getType() != Type::VoidTy) 904 CI->replaceAllUsesWith(ConstantInt::get(CI->getType(), 1)); 905 break; 906 } 907 908 assert(CI->use_empty() && 909 "Lowering should have eliminated any uses of the intrinsic call!"); 910 CI->eraseFromParent(); 911} 912