IntrinsicLowering.cpp revision c4c966012901691fff21eed02d72a3de44dd47f1
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#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 37/// ReplaceCallWith - This function is used when we want to lower an intrinsic 38/// call to a call of an external function. This handles hard cases such as 39/// when there was already a prototype for the external function, and if that 40/// prototype doesn't match the arguments we expect to pass in. 41template <class ArgIt> 42static CallInst *ReplaceCallWith(const char *NewFn, CallInst *CI, 43 ArgIt ArgBegin, ArgIt ArgEnd, 44 const Type *RetTy, Constant *&FCache) { 45 if (!FCache) { 46 // If we haven't already looked up this function, check to see if the 47 // program already contains a function with this name. 48 Module *M = CI->getParent()->getParent()->getParent(); 49 // Get or insert the definition now. 50 std::vector<const Type *> ParamTys; 51 for (ArgIt I = ArgBegin; I != ArgEnd; ++I) 52 ParamTys.push_back((*I)->getType()); 53 FCache = M->getOrInsertFunction(NewFn, 54 FunctionType::get(RetTy, ParamTys, false)); 55 } 56 57 SmallVector<Value *, 8> Args(ArgBegin, ArgEnd); 58 CallInst *NewCI = new CallInst(FCache, Args.begin(), Args.end(), 59 CI->getName(), CI); 60 if (!CI->use_empty()) 61 CI->replaceAllUsesWith(NewCI); 62 return NewCI; 63} 64 65void IntrinsicLowering::AddPrototypes(Module &M) { 66 for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I) 67 if (I->isDeclaration() && !I->use_empty()) 68 switch (I->getIntrinsicID()) { 69 default: break; 70 case Intrinsic::setjmp: 71 EnsureFunctionExists(M, "setjmp", I->arg_begin(), I->arg_end(), 72 Type::Int32Ty); 73 break; 74 case Intrinsic::longjmp: 75 EnsureFunctionExists(M, "longjmp", I->arg_begin(), I->arg_end(), 76 Type::VoidTy); 77 break; 78 case Intrinsic::siglongjmp: 79 EnsureFunctionExists(M, "abort", I->arg_end(), I->arg_end(), 80 Type::VoidTy); 81 break; 82 case Intrinsic::memcpy_i32: 83 case Intrinsic::memcpy_i64: 84 M.getOrInsertFunction("memcpy", PointerType::get(Type::Int8Ty), 85 PointerType::get(Type::Int8Ty), 86 PointerType::get(Type::Int8Ty), 87 TD.getIntPtrType(), (Type *)0); 88 break; 89 case Intrinsic::memmove_i32: 90 case Intrinsic::memmove_i64: 91 M.getOrInsertFunction("memmove", PointerType::get(Type::Int8Ty), 92 PointerType::get(Type::Int8Ty), 93 PointerType::get(Type::Int8Ty), 94 TD.getIntPtrType(), (Type *)0); 95 break; 96 case Intrinsic::memset_i32: 97 case Intrinsic::memset_i64: 98 M.getOrInsertFunction("memset", PointerType::get(Type::Int8Ty), 99 PointerType::get(Type::Int8Ty), Type::Int32Ty, 100 TD.getIntPtrType(), (Type *)0); 101 break; 102 case Intrinsic::sqrt: 103 switch((int)I->arg_begin()->getType()->getTypeID()) { 104 case Type::FloatTyID: 105 EnsureFunctionExists(M, "sqrtf", I->arg_begin(), I->arg_end(), 106 Type::FloatTy); 107 case Type::DoubleTyID: 108 EnsureFunctionExists(M, "sqrt", I->arg_begin(), I->arg_end(), 109 Type::DoubleTy); 110 case Type::X86_FP80TyID: 111 case Type::FP128TyID: 112 case Type::PPC_FP128TyID: 113 EnsureFunctionExists(M, "sqrtl", I->arg_begin(), I->arg_end(), 114 I->arg_begin()->getType()); 115 } 116 break; 117 case Intrinsic::sin: 118 switch((int)I->arg_begin()->getType()->getTypeID()) { 119 case Type::FloatTyID: 120 EnsureFunctionExists(M, "sinf", I->arg_begin(), I->arg_end(), 121 Type::FloatTy); 122 case Type::DoubleTyID: 123 EnsureFunctionExists(M, "sin", I->arg_begin(), I->arg_end(), 124 Type::DoubleTy); 125 case Type::X86_FP80TyID: 126 case Type::FP128TyID: 127 case Type::PPC_FP128TyID: 128 EnsureFunctionExists(M, "sinl", I->arg_begin(), I->arg_end(), 129 I->arg_begin()->getType()); 130 } 131 break; 132 case Intrinsic::cos: 133 switch((int)I->arg_begin()->getType()->getTypeID()) { 134 case Type::FloatTyID: 135 EnsureFunctionExists(M, "cosf", I->arg_begin(), I->arg_end(), 136 Type::FloatTy); 137 case Type::DoubleTyID: 138 EnsureFunctionExists(M, "cos", I->arg_begin(), I->arg_end(), 139 Type::DoubleTy); 140 case Type::X86_FP80TyID: 141 case Type::FP128TyID: 142 case Type::PPC_FP128TyID: 143 EnsureFunctionExists(M, "cosl", I->arg_begin(), I->arg_end(), 144 I->arg_begin()->getType()); 145 } 146 break; 147 case Intrinsic::pow: 148 switch((int)I->arg_begin()->getType()->getTypeID()) { 149 case Type::FloatTyID: 150 EnsureFunctionExists(M, "powf", I->arg_begin(), I->arg_end(), 151 Type::FloatTy); 152 case Type::DoubleTyID: 153 EnsureFunctionExists(M, "pow", I->arg_begin(), I->arg_end(), 154 Type::DoubleTy); 155 case Type::X86_FP80TyID: 156 case Type::FP128TyID: 157 case Type::PPC_FP128TyID: 158 EnsureFunctionExists(M, "powl", I->arg_begin(), I->arg_end(), 159 I->arg_begin()->getType()); 160 } 161 break; 162 } 163} 164 165/// LowerBSWAP - Emit the code to lower bswap of V before the specified 166/// instruction IP. 167static Value *LowerBSWAP(Value *V, Instruction *IP) { 168 assert(V->getType()->isInteger() && "Can't bswap a non-integer type!"); 169 170 unsigned BitSize = V->getType()->getPrimitiveSizeInBits(); 171 172 switch(BitSize) { 173 default: assert(0 && "Unhandled type size of value to byteswap!"); 174 case 16: { 175 Value *Tmp1 = BinaryOperator::createShl(V, 176 ConstantInt::get(V->getType(),8),"bswap.2",IP); 177 Value *Tmp2 = BinaryOperator::createLShr(V, 178 ConstantInt::get(V->getType(),8),"bswap.1",IP); 179 V = BinaryOperator::createOr(Tmp1, Tmp2, "bswap.i16", IP); 180 break; 181 } 182 case 32: { 183 Value *Tmp4 = BinaryOperator::createShl(V, 184 ConstantInt::get(V->getType(),24),"bswap.4", IP); 185 Value *Tmp3 = BinaryOperator::createShl(V, 186 ConstantInt::get(V->getType(),8),"bswap.3",IP); 187 Value *Tmp2 = BinaryOperator::createLShr(V, 188 ConstantInt::get(V->getType(),8),"bswap.2",IP); 189 Value *Tmp1 = BinaryOperator::createLShr(V, 190 ConstantInt::get(V->getType(),24),"bswap.1", IP); 191 Tmp3 = BinaryOperator::createAnd(Tmp3, 192 ConstantInt::get(Type::Int32Ty, 0xFF0000), 193 "bswap.and3", IP); 194 Tmp2 = BinaryOperator::createAnd(Tmp2, 195 ConstantInt::get(Type::Int32Ty, 0xFF00), 196 "bswap.and2", IP); 197 Tmp4 = BinaryOperator::createOr(Tmp4, Tmp3, "bswap.or1", IP); 198 Tmp2 = BinaryOperator::createOr(Tmp2, Tmp1, "bswap.or2", IP); 199 V = BinaryOperator::createOr(Tmp4, Tmp2, "bswap.i32", IP); 200 break; 201 } 202 case 64: { 203 Value *Tmp8 = BinaryOperator::createShl(V, 204 ConstantInt::get(V->getType(),56),"bswap.8", IP); 205 Value *Tmp7 = BinaryOperator::createShl(V, 206 ConstantInt::get(V->getType(),40),"bswap.7", IP); 207 Value *Tmp6 = BinaryOperator::createShl(V, 208 ConstantInt::get(V->getType(),24),"bswap.6", IP); 209 Value *Tmp5 = BinaryOperator::createShl(V, 210 ConstantInt::get(V->getType(),8),"bswap.5", IP); 211 Value* Tmp4 = BinaryOperator::createLShr(V, 212 ConstantInt::get(V->getType(),8),"bswap.4", IP); 213 Value* Tmp3 = BinaryOperator::createLShr(V, 214 ConstantInt::get(V->getType(),24),"bswap.3", IP); 215 Value* Tmp2 = BinaryOperator::createLShr(V, 216 ConstantInt::get(V->getType(),40),"bswap.2", IP); 217 Value* Tmp1 = BinaryOperator::createLShr(V, 218 ConstantInt::get(V->getType(),56),"bswap.1", IP); 219 Tmp7 = BinaryOperator::createAnd(Tmp7, 220 ConstantInt::get(Type::Int64Ty, 221 0xFF000000000000ULL), 222 "bswap.and7", IP); 223 Tmp6 = BinaryOperator::createAnd(Tmp6, 224 ConstantInt::get(Type::Int64Ty, 0xFF0000000000ULL), 225 "bswap.and6", IP); 226 Tmp5 = BinaryOperator::createAnd(Tmp5, 227 ConstantInt::get(Type::Int64Ty, 0xFF00000000ULL), 228 "bswap.and5", IP); 229 Tmp4 = BinaryOperator::createAnd(Tmp4, 230 ConstantInt::get(Type::Int64Ty, 0xFF000000ULL), 231 "bswap.and4", IP); 232 Tmp3 = BinaryOperator::createAnd(Tmp3, 233 ConstantInt::get(Type::Int64Ty, 0xFF0000ULL), 234 "bswap.and3", IP); 235 Tmp2 = BinaryOperator::createAnd(Tmp2, 236 ConstantInt::get(Type::Int64Ty, 0xFF00ULL), 237 "bswap.and2", IP); 238 Tmp8 = BinaryOperator::createOr(Tmp8, Tmp7, "bswap.or1", IP); 239 Tmp6 = BinaryOperator::createOr(Tmp6, Tmp5, "bswap.or2", IP); 240 Tmp4 = BinaryOperator::createOr(Tmp4, Tmp3, "bswap.or3", IP); 241 Tmp2 = BinaryOperator::createOr(Tmp2, Tmp1, "bswap.or4", IP); 242 Tmp8 = BinaryOperator::createOr(Tmp8, Tmp6, "bswap.or5", IP); 243 Tmp4 = BinaryOperator::createOr(Tmp4, Tmp2, "bswap.or6", IP); 244 V = BinaryOperator::createOr(Tmp8, Tmp4, "bswap.i64", IP); 245 break; 246 } 247 } 248 return V; 249} 250 251/// LowerCTPOP - Emit the code to lower ctpop of V before the specified 252/// instruction IP. 253static Value *LowerCTPOP(Value *V, Instruction *IP) { 254 assert(V->getType()->isInteger() && "Can't ctpop a non-integer type!"); 255 256 static const uint64_t MaskValues[6] = { 257 0x5555555555555555ULL, 0x3333333333333333ULL, 258 0x0F0F0F0F0F0F0F0FULL, 0x00FF00FF00FF00FFULL, 259 0x0000FFFF0000FFFFULL, 0x00000000FFFFFFFFULL 260 }; 261 262 unsigned BitSize = V->getType()->getPrimitiveSizeInBits(); 263 unsigned WordSize = (BitSize + 63) / 64; 264 Value *Count = ConstantInt::get(V->getType(), 0); 265 266 for (unsigned n = 0; n < WordSize; ++n) { 267 Value *PartValue = V; 268 for (unsigned i = 1, ct = 0; i < (BitSize>64 ? 64 : BitSize); 269 i <<= 1, ++ct) { 270 Value *MaskCst = ConstantInt::get(V->getType(), MaskValues[ct]); 271 Value *LHS = BinaryOperator::createAnd( 272 PartValue, MaskCst, "cppop.and1", IP); 273 Value *VShift = BinaryOperator::createLShr(PartValue, 274 ConstantInt::get(V->getType(), i), "ctpop.sh", IP); 275 Value *RHS = BinaryOperator::createAnd(VShift, MaskCst, "cppop.and2", IP); 276 PartValue = BinaryOperator::createAdd(LHS, RHS, "ctpop.step", IP); 277 } 278 Count = BinaryOperator::createAdd(PartValue, Count, "ctpop.part", IP); 279 if (BitSize > 64) { 280 V = BinaryOperator::createLShr(V, ConstantInt::get(V->getType(), 64), 281 "ctpop.part.sh", IP); 282 BitSize -= 64; 283 } 284 } 285 286 return Count; 287} 288 289/// LowerCTLZ - Emit the code to lower ctlz of V before the specified 290/// instruction IP. 291static Value *LowerCTLZ(Value *V, Instruction *IP) { 292 293 unsigned BitSize = V->getType()->getPrimitiveSizeInBits(); 294 for (unsigned i = 1; i < BitSize; i <<= 1) { 295 Value *ShVal = ConstantInt::get(V->getType(), i); 296 ShVal = BinaryOperator::createLShr(V, ShVal, "ctlz.sh", IP); 297 V = BinaryOperator::createOr(V, ShVal, "ctlz.step", IP); 298 } 299 300 V = BinaryOperator::createNot(V, "", IP); 301 return LowerCTPOP(V, IP); 302} 303 304/// Convert the llvm.part.select.iX.iY intrinsic. This intrinsic takes 305/// three integer arguments. The first argument is the Value from which the 306/// bits will be selected. It may be of any bit width. The second and third 307/// arguments specify a range of bits to select with the second argument 308/// specifying the low bit and the third argument specifying the high bit. Both 309/// must be type i32. The result is the corresponding selected bits from the 310/// Value in the same width as the Value (first argument). If the low bit index 311/// is higher than the high bit index then the inverse selection is done and 312/// the bits are returned in inverse order. 313/// @brief Lowering of llvm.part.select intrinsic. 314static Instruction *LowerPartSelect(CallInst *CI) { 315 // Make sure we're dealing with a part select intrinsic here 316 Function *F = CI->getCalledFunction(); 317 const FunctionType *FT = F->getFunctionType(); 318 if (!F->isDeclaration() || !FT->getReturnType()->isInteger() || 319 FT->getNumParams() != 3 || !FT->getParamType(0)->isInteger() || 320 !FT->getParamType(1)->isInteger() || !FT->getParamType(2)->isInteger()) 321 return CI; 322 323 // Get the intrinsic implementation function by converting all the . to _ 324 // in the intrinsic's function name and then reconstructing the function 325 // declaration. 326 std::string Name(F->getName()); 327 for (unsigned i = 4; i < Name.length(); ++i) 328 if (Name[i] == '.') 329 Name[i] = '_'; 330 Module* M = F->getParent(); 331 F = cast<Function>(M->getOrInsertFunction(Name, FT)); 332 F->setLinkage(GlobalValue::WeakLinkage); 333 334 // If we haven't defined the impl function yet, do so now 335 if (F->isDeclaration()) { 336 337 // Get the arguments to the function 338 Function::arg_iterator args = F->arg_begin(); 339 Value* Val = args++; Val->setName("Val"); 340 Value* Lo = args++; Lo->setName("Lo"); 341 Value* Hi = args++; Hi->setName("High"); 342 343 // We want to select a range of bits here such that [Hi, Lo] is shifted 344 // down to the low bits. However, it is quite possible that Hi is smaller 345 // than Lo in which case the bits have to be reversed. 346 347 // Create the blocks we will need for the two cases (forward, reverse) 348 BasicBlock* CurBB = new BasicBlock("entry", F); 349 BasicBlock *RevSize = new BasicBlock("revsize", CurBB->getParent()); 350 BasicBlock *FwdSize = new BasicBlock("fwdsize", CurBB->getParent()); 351 BasicBlock *Compute = new BasicBlock("compute", CurBB->getParent()); 352 BasicBlock *Reverse = new BasicBlock("reverse", CurBB->getParent()); 353 BasicBlock *RsltBlk = new BasicBlock("result", CurBB->getParent()); 354 355 // Cast Hi and Lo to the size of Val so the widths are all the same 356 if (Hi->getType() != Val->getType()) 357 Hi = CastInst::createIntegerCast(Hi, Val->getType(), false, 358 "tmp", CurBB); 359 if (Lo->getType() != Val->getType()) 360 Lo = CastInst::createIntegerCast(Lo, Val->getType(), false, 361 "tmp", CurBB); 362 363 // Compute a few things that both cases will need, up front. 364 Constant* Zero = ConstantInt::get(Val->getType(), 0); 365 Constant* One = ConstantInt::get(Val->getType(), 1); 366 Constant* AllOnes = ConstantInt::getAllOnesValue(Val->getType()); 367 368 // Compare the Hi and Lo bit positions. This is used to determine 369 // which case we have (forward or reverse) 370 ICmpInst *Cmp = new ICmpInst(ICmpInst::ICMP_ULT, Hi, Lo, "less",CurBB); 371 new BranchInst(RevSize, FwdSize, Cmp, CurBB); 372 373 // First, copmute the number of bits in the forward case. 374 Instruction* FBitSize = 375 BinaryOperator::createSub(Hi, Lo,"fbits", FwdSize); 376 new BranchInst(Compute, FwdSize); 377 378 // Second, compute the number of bits in the reverse case. 379 Instruction* RBitSize = 380 BinaryOperator::createSub(Lo, Hi, "rbits", RevSize); 381 new BranchInst(Compute, RevSize); 382 383 // Now, compute the bit range. Start by getting the bitsize and the shift 384 // amount (either Hi or Lo) from PHI nodes. Then we compute a mask for 385 // the number of bits we want in the range. We shift the bits down to the 386 // least significant bits, apply the mask to zero out unwanted high bits, 387 // and we have computed the "forward" result. It may still need to be 388 // reversed. 389 390 // Get the BitSize from one of the two subtractions 391 PHINode *BitSize = new PHINode(Val->getType(), "bits", Compute); 392 BitSize->reserveOperandSpace(2); 393 BitSize->addIncoming(FBitSize, FwdSize); 394 BitSize->addIncoming(RBitSize, RevSize); 395 396 // Get the ShiftAmount as the smaller of Hi/Lo 397 PHINode *ShiftAmt = new PHINode(Val->getType(), "shiftamt", Compute); 398 ShiftAmt->reserveOperandSpace(2); 399 ShiftAmt->addIncoming(Lo, FwdSize); 400 ShiftAmt->addIncoming(Hi, RevSize); 401 402 // Increment the bit size 403 Instruction *BitSizePlusOne = 404 BinaryOperator::createAdd(BitSize, One, "bits", Compute); 405 406 // Create a Mask to zero out the high order bits. 407 Instruction* Mask = 408 BinaryOperator::createShl(AllOnes, BitSizePlusOne, "mask", Compute); 409 Mask = BinaryOperator::createNot(Mask, "mask", Compute); 410 411 // Shift the bits down and apply the mask 412 Instruction* FRes = 413 BinaryOperator::createLShr(Val, ShiftAmt, "fres", Compute); 414 FRes = BinaryOperator::createAnd(FRes, Mask, "fres", Compute); 415 new BranchInst(Reverse, RsltBlk, Cmp, Compute); 416 417 // In the Reverse block we have the mask already in FRes but we must reverse 418 // it by shifting FRes bits right and putting them in RRes by shifting them 419 // in from left. 420 421 // First set up our loop counters 422 PHINode *Count = new PHINode(Val->getType(), "count", Reverse); 423 Count->reserveOperandSpace(2); 424 Count->addIncoming(BitSizePlusOne, Compute); 425 426 // Next, get the value that we are shifting. 427 PHINode *BitsToShift = new PHINode(Val->getType(), "val", Reverse); 428 BitsToShift->reserveOperandSpace(2); 429 BitsToShift->addIncoming(FRes, Compute); 430 431 // Finally, get the result of the last computation 432 PHINode *RRes = new PHINode(Val->getType(), "rres", Reverse); 433 RRes->reserveOperandSpace(2); 434 RRes->addIncoming(Zero, Compute); 435 436 // Decrement the counter 437 Instruction *Decr = BinaryOperator::createSub(Count, One, "decr", Reverse); 438 Count->addIncoming(Decr, Reverse); 439 440 // Compute the Bit that we want to move 441 Instruction *Bit = 442 BinaryOperator::createAnd(BitsToShift, One, "bit", Reverse); 443 444 // Compute the new value for next iteration. 445 Instruction *NewVal = 446 BinaryOperator::createLShr(BitsToShift, One, "rshift", Reverse); 447 BitsToShift->addIncoming(NewVal, Reverse); 448 449 // Shift the bit into the low bits of the result. 450 Instruction *NewRes = 451 BinaryOperator::createShl(RRes, One, "lshift", Reverse); 452 NewRes = BinaryOperator::createOr(NewRes, Bit, "addbit", Reverse); 453 RRes->addIncoming(NewRes, Reverse); 454 455 // Terminate loop if we've moved all the bits. 456 ICmpInst *Cond = 457 new ICmpInst(ICmpInst::ICMP_EQ, Decr, Zero, "cond", Reverse); 458 new BranchInst(RsltBlk, Reverse, Cond, Reverse); 459 460 // Finally, in the result block, select one of the two results with a PHI 461 // node and return the result; 462 CurBB = RsltBlk; 463 PHINode *BitSelect = new PHINode(Val->getType(), "part_select", CurBB); 464 BitSelect->reserveOperandSpace(2); 465 BitSelect->addIncoming(FRes, Compute); 466 BitSelect->addIncoming(NewRes, Reverse); 467 new ReturnInst(BitSelect, CurBB); 468 } 469 470 // Return a call to the implementation function 471 Value *Args[] = { 472 CI->getOperand(1), 473 CI->getOperand(2), 474 CI->getOperand(3) 475 }; 476 return new CallInst(F, Args, array_endof(Args), CI->getName(), CI); 477} 478 479/// Convert the llvm.part.set.iX.iY.iZ intrinsic. This intrinsic takes 480/// four integer arguments (iAny %Value, iAny %Replacement, i32 %Low, i32 %High) 481/// The first two arguments can be any bit width. The result is the same width 482/// as %Value. The operation replaces bits between %Low and %High with the value 483/// in %Replacement. If %Replacement is not the same width, it is truncated or 484/// zero extended as appropriate to fit the bits being replaced. If %Low is 485/// greater than %High then the inverse set of bits are replaced. 486/// @brief Lowering of llvm.bit.part.set intrinsic. 487static Instruction *LowerPartSet(CallInst *CI) { 488 // Make sure we're dealing with a part select intrinsic here 489 Function *F = CI->getCalledFunction(); 490 const FunctionType *FT = F->getFunctionType(); 491 if (!F->isDeclaration() || !FT->getReturnType()->isInteger() || 492 FT->getNumParams() != 4 || !FT->getParamType(0)->isInteger() || 493 !FT->getParamType(1)->isInteger() || !FT->getParamType(2)->isInteger() || 494 !FT->getParamType(3)->isInteger()) 495 return CI; 496 497 // Get the intrinsic implementation function by converting all the . to _ 498 // in the intrinsic's function name and then reconstructing the function 499 // declaration. 500 std::string Name(F->getName()); 501 for (unsigned i = 4; i < Name.length(); ++i) 502 if (Name[i] == '.') 503 Name[i] = '_'; 504 Module* M = F->getParent(); 505 F = cast<Function>(M->getOrInsertFunction(Name, FT)); 506 F->setLinkage(GlobalValue::WeakLinkage); 507 508 // If we haven't defined the impl function yet, do so now 509 if (F->isDeclaration()) { 510 // Get the arguments for the function. 511 Function::arg_iterator args = F->arg_begin(); 512 Value* Val = args++; Val->setName("Val"); 513 Value* Rep = args++; Rep->setName("Rep"); 514 Value* Lo = args++; Lo->setName("Lo"); 515 Value* Hi = args++; Hi->setName("Hi"); 516 517 // Get some types we need 518 const IntegerType* ValTy = cast<IntegerType>(Val->getType()); 519 const IntegerType* RepTy = cast<IntegerType>(Rep->getType()); 520 uint32_t ValBits = ValTy->getBitWidth(); 521 uint32_t RepBits = RepTy->getBitWidth(); 522 523 // Constant Definitions 524 ConstantInt* RepBitWidth = ConstantInt::get(Type::Int32Ty, RepBits); 525 ConstantInt* RepMask = ConstantInt::getAllOnesValue(RepTy); 526 ConstantInt* ValMask = ConstantInt::getAllOnesValue(ValTy); 527 ConstantInt* One = ConstantInt::get(Type::Int32Ty, 1); 528 ConstantInt* ValOne = ConstantInt::get(ValTy, 1); 529 ConstantInt* Zero = ConstantInt::get(Type::Int32Ty, 0); 530 ConstantInt* ValZero = ConstantInt::get(ValTy, 0); 531 532 // Basic blocks we fill in below. 533 BasicBlock* entry = new BasicBlock("entry", F, 0); 534 BasicBlock* large = new BasicBlock("large", F, 0); 535 BasicBlock* small = new BasicBlock("small", F, 0); 536 BasicBlock* reverse = new BasicBlock("reverse", F, 0); 537 BasicBlock* result = new BasicBlock("result", F, 0); 538 539 // BASIC BLOCK: entry 540 // First, get the number of bits that we're placing as an i32 541 ICmpInst* is_forward = 542 new ICmpInst(ICmpInst::ICMP_ULT, Lo, Hi, "", entry); 543 SelectInst* Hi_pn = new SelectInst(is_forward, Hi, Lo, "", entry); 544 SelectInst* Lo_pn = new SelectInst(is_forward, Lo, Hi, "", entry); 545 BinaryOperator* NumBits = BinaryOperator::createSub(Hi_pn, Lo_pn, "",entry); 546 NumBits = BinaryOperator::createAdd(NumBits, One, "", entry); 547 // Now, convert Lo and Hi to ValTy bit width 548 if (ValBits > 32) { 549 Lo = new ZExtInst(Lo_pn, ValTy, "", entry); 550 } else if (ValBits < 32) { 551 Lo = new TruncInst(Lo_pn, ValTy, "", entry); 552 } 553 // Determine if the replacement bits are larger than the number of bits we 554 // are replacing and deal with it. 555 ICmpInst* is_large = 556 new ICmpInst(ICmpInst::ICMP_ULT, NumBits, RepBitWidth, "", entry); 557 new BranchInst(large, small, is_large, entry); 558 559 // BASIC BLOCK: large 560 Instruction* MaskBits = 561 BinaryOperator::createSub(RepBitWidth, NumBits, "", large); 562 MaskBits = CastInst::createIntegerCast(MaskBits, RepMask->getType(), 563 false, "", large); 564 BinaryOperator* Mask1 = 565 BinaryOperator::createLShr(RepMask, MaskBits, "", large); 566 BinaryOperator* Rep2 = BinaryOperator::createAnd(Mask1, Rep, "", large); 567 new BranchInst(small, large); 568 569 // BASIC BLOCK: small 570 PHINode* Rep3 = new PHINode(RepTy, "", small); 571 Rep3->reserveOperandSpace(2); 572 Rep3->addIncoming(Rep2, large); 573 Rep3->addIncoming(Rep, entry); 574 Value* Rep4 = Rep3; 575 if (ValBits > RepBits) 576 Rep4 = new ZExtInst(Rep3, ValTy, "", small); 577 else if (ValBits < RepBits) 578 Rep4 = new TruncInst(Rep3, ValTy, "", small); 579 new BranchInst(result, reverse, is_forward, small); 580 581 // BASIC BLOCK: reverse (reverses the bits of the replacement) 582 // Set up our loop counter as a PHI so we can decrement on each iteration. 583 // We will loop for the number of bits in the replacement value. 584 PHINode *Count = new PHINode(Type::Int32Ty, "count", reverse); 585 Count->reserveOperandSpace(2); 586 Count->addIncoming(NumBits, small); 587 588 // Get the value that we are shifting bits out of as a PHI because 589 // we'll change this with each iteration. 590 PHINode *BitsToShift = new PHINode(Val->getType(), "val", reverse); 591 BitsToShift->reserveOperandSpace(2); 592 BitsToShift->addIncoming(Rep4, small); 593 594 // Get the result of the last computation or zero on first iteration 595 PHINode *RRes = new PHINode(Val->getType(), "rres", reverse); 596 RRes->reserveOperandSpace(2); 597 RRes->addIncoming(ValZero, small); 598 599 // Decrement the loop counter by one 600 Instruction *Decr = BinaryOperator::createSub(Count, One, "", reverse); 601 Count->addIncoming(Decr, reverse); 602 603 // Get the bit that we want to move into the result 604 Value *Bit = BinaryOperator::createAnd(BitsToShift, ValOne, "", reverse); 605 606 // Compute the new value of the bits to shift for the next iteration. 607 Value *NewVal = BinaryOperator::createLShr(BitsToShift, ValOne,"", reverse); 608 BitsToShift->addIncoming(NewVal, reverse); 609 610 // Shift the bit we extracted into the low bit of the result. 611 Instruction *NewRes = BinaryOperator::createShl(RRes, ValOne, "", reverse); 612 NewRes = BinaryOperator::createOr(NewRes, Bit, "", reverse); 613 RRes->addIncoming(NewRes, reverse); 614 615 // Terminate loop if we've moved all the bits. 616 ICmpInst *Cond = new ICmpInst(ICmpInst::ICMP_EQ, Decr, Zero, "", reverse); 617 new BranchInst(result, reverse, Cond, reverse); 618 619 // BASIC BLOCK: result 620 PHINode *Rplcmnt = new PHINode(Val->getType(), "", result); 621 Rplcmnt->reserveOperandSpace(2); 622 Rplcmnt->addIncoming(NewRes, reverse); 623 Rplcmnt->addIncoming(Rep4, small); 624 Value* t0 = CastInst::createIntegerCast(NumBits,ValTy,false,"",result); 625 Value* t1 = BinaryOperator::createShl(ValMask, Lo, "", result); 626 Value* t2 = BinaryOperator::createNot(t1, "", result); 627 Value* t3 = BinaryOperator::createShl(t1, t0, "", result); 628 Value* t4 = BinaryOperator::createOr(t2, t3, "", result); 629 Value* t5 = BinaryOperator::createAnd(t4, Val, "", result); 630 Value* t6 = BinaryOperator::createShl(Rplcmnt, Lo, "", result); 631 Value* Rslt = BinaryOperator::createOr(t5, t6, "part_set", result); 632 new ReturnInst(Rslt, result); 633 } 634 635 // Return a call to the implementation function 636 Value *Args[] = { 637 CI->getOperand(1), 638 CI->getOperand(2), 639 CI->getOperand(3), 640 CI->getOperand(4) 641 }; 642 return new CallInst(F, Args, array_endof(Args), CI->getName(), CI); 643} 644 645 646void IntrinsicLowering::LowerIntrinsicCall(CallInst *CI) { 647 Function *Callee = CI->getCalledFunction(); 648 assert(Callee && "Cannot lower an indirect call!"); 649 650 switch (Callee->getIntrinsicID()) { 651 case Intrinsic::not_intrinsic: 652 cerr << "Cannot lower a call to a non-intrinsic function '" 653 << Callee->getName() << "'!\n"; 654 abort(); 655 default: 656 cerr << "Error: Code generator does not support intrinsic function '" 657 << Callee->getName() << "'!\n"; 658 abort(); 659 660 // The setjmp/longjmp intrinsics should only exist in the code if it was 661 // never optimized (ie, right out of the CFE), or if it has been hacked on 662 // by the lowerinvoke pass. In both cases, the right thing to do is to 663 // convert the call to an explicit setjmp or longjmp call. 664 case Intrinsic::setjmp: { 665 static Constant *SetjmpFCache = 0; 666 Value *V = ReplaceCallWith("setjmp", CI, CI->op_begin()+1, CI->op_end(), 667 Type::Int32Ty, SetjmpFCache); 668 if (CI->getType() != Type::VoidTy) 669 CI->replaceAllUsesWith(V); 670 break; 671 } 672 case Intrinsic::sigsetjmp: 673 if (CI->getType() != Type::VoidTy) 674 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType())); 675 break; 676 677 case Intrinsic::longjmp: { 678 static Constant *LongjmpFCache = 0; 679 ReplaceCallWith("longjmp", CI, CI->op_begin()+1, CI->op_end(), 680 Type::VoidTy, LongjmpFCache); 681 break; 682 } 683 684 case Intrinsic::siglongjmp: { 685 // Insert the call to abort 686 static Constant *AbortFCache = 0; 687 ReplaceCallWith("abort", CI, CI->op_end(), CI->op_end(), 688 Type::VoidTy, AbortFCache); 689 break; 690 } 691 case Intrinsic::ctpop: 692 CI->replaceAllUsesWith(LowerCTPOP(CI->getOperand(1), CI)); 693 break; 694 695 case Intrinsic::bswap: 696 CI->replaceAllUsesWith(LowerBSWAP(CI->getOperand(1), CI)); 697 break; 698 699 case Intrinsic::ctlz: 700 CI->replaceAllUsesWith(LowerCTLZ(CI->getOperand(1), CI)); 701 break; 702 703 case Intrinsic::cttz: { 704 // cttz(x) -> ctpop(~X & (X-1)) 705 Value *Src = CI->getOperand(1); 706 Value *NotSrc = BinaryOperator::createNot(Src, Src->getName()+".not", CI); 707 Value *SrcM1 = ConstantInt::get(Src->getType(), 1); 708 SrcM1 = BinaryOperator::createSub(Src, SrcM1, "", CI); 709 Src = LowerCTPOP(BinaryOperator::createAnd(NotSrc, SrcM1, "", CI), CI); 710 CI->replaceAllUsesWith(Src); 711 break; 712 } 713 714 case Intrinsic::part_select: 715 CI->replaceAllUsesWith(LowerPartSelect(CI)); 716 break; 717 718 case Intrinsic::part_set: 719 CI->replaceAllUsesWith(LowerPartSet(CI)); 720 break; 721 722 case Intrinsic::stacksave: 723 case Intrinsic::stackrestore: { 724 static bool Warned = false; 725 if (!Warned) 726 cerr << "WARNING: this target does not support the llvm.stack" 727 << (Callee->getIntrinsicID() == Intrinsic::stacksave ? 728 "save" : "restore") << " intrinsic.\n"; 729 Warned = true; 730 if (Callee->getIntrinsicID() == Intrinsic::stacksave) 731 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType())); 732 break; 733 } 734 735 case Intrinsic::returnaddress: 736 case Intrinsic::frameaddress: 737 cerr << "WARNING: this target does not support the llvm." 738 << (Callee->getIntrinsicID() == Intrinsic::returnaddress ? 739 "return" : "frame") << "address intrinsic.\n"; 740 CI->replaceAllUsesWith(ConstantPointerNull::get( 741 cast<PointerType>(CI->getType()))); 742 break; 743 744 case Intrinsic::prefetch: 745 break; // Simply strip out prefetches on unsupported architectures 746 747 case Intrinsic::pcmarker: 748 break; // Simply strip out pcmarker on unsupported architectures 749 case Intrinsic::readcyclecounter: { 750 cerr << "WARNING: this target does not support the llvm.readcyclecoun" 751 << "ter intrinsic. It is being lowered to a constant 0\n"; 752 CI->replaceAllUsesWith(ConstantInt::get(Type::Int64Ty, 0)); 753 break; 754 } 755 756 case Intrinsic::dbg_stoppoint: 757 case Intrinsic::dbg_region_start: 758 case Intrinsic::dbg_region_end: 759 case Intrinsic::dbg_func_start: 760 case Intrinsic::dbg_declare: 761 break; // Simply strip out debugging intrinsics 762 763 case Intrinsic::eh_exception: 764 case Intrinsic::eh_selector_i32: 765 case Intrinsic::eh_selector_i64: 766 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType())); 767 break; 768 769 case Intrinsic::eh_typeid_for_i32: 770 case Intrinsic::eh_typeid_for_i64: 771 // Return something different to eh_selector. 772 CI->replaceAllUsesWith(ConstantInt::get(CI->getType(), 1)); 773 break; 774 775 case Intrinsic::var_annotation: 776 break; // Strip out annotate intrinsic 777 778 case Intrinsic::memcpy_i32: 779 case Intrinsic::memcpy_i64: { 780 static Constant *MemcpyFCache = 0; 781 Value *Size = CI->getOperand(3); 782 const Type *IntPtr = TD.getIntPtrType(); 783 if (Size->getType()->getPrimitiveSizeInBits() < 784 IntPtr->getPrimitiveSizeInBits()) 785 Size = new ZExtInst(Size, IntPtr, "", CI); 786 else if (Size->getType()->getPrimitiveSizeInBits() > 787 IntPtr->getPrimitiveSizeInBits()) 788 Size = new TruncInst(Size, IntPtr, "", CI); 789 Value *Ops[3]; 790 Ops[0] = CI->getOperand(1); 791 Ops[1] = CI->getOperand(2); 792 Ops[2] = Size; 793 ReplaceCallWith("memcpy", CI, Ops, Ops+3, CI->getOperand(1)->getType(), 794 MemcpyFCache); 795 break; 796 } 797 case Intrinsic::memmove_i32: 798 case Intrinsic::memmove_i64: { 799 static Constant *MemmoveFCache = 0; 800 Value *Size = CI->getOperand(3); 801 const Type *IntPtr = TD.getIntPtrType(); 802 if (Size->getType()->getPrimitiveSizeInBits() < 803 IntPtr->getPrimitiveSizeInBits()) 804 Size = new ZExtInst(Size, IntPtr, "", CI); 805 else if (Size->getType()->getPrimitiveSizeInBits() > 806 IntPtr->getPrimitiveSizeInBits()) 807 Size = new TruncInst(Size, IntPtr, "", CI); 808 Value *Ops[3]; 809 Ops[0] = CI->getOperand(1); 810 Ops[1] = CI->getOperand(2); 811 Ops[2] = Size; 812 ReplaceCallWith("memmove", CI, Ops, Ops+3, CI->getOperand(1)->getType(), 813 MemmoveFCache); 814 break; 815 } 816 case Intrinsic::memset_i32: 817 case Intrinsic::memset_i64: { 818 static Constant *MemsetFCache = 0; 819 Value *Size = CI->getOperand(3); 820 const Type *IntPtr = TD.getIntPtrType(); 821 if (Size->getType()->getPrimitiveSizeInBits() < 822 IntPtr->getPrimitiveSizeInBits()) 823 Size = new ZExtInst(Size, IntPtr, "", CI); 824 else if (Size->getType()->getPrimitiveSizeInBits() > 825 IntPtr->getPrimitiveSizeInBits()) 826 Size = new TruncInst(Size, IntPtr, "", CI); 827 Value *Ops[3]; 828 Ops[0] = CI->getOperand(1); 829 // Extend the amount to i32. 830 Ops[1] = new ZExtInst(CI->getOperand(2), Type::Int32Ty, "", CI); 831 Ops[2] = Size; 832 ReplaceCallWith("memset", CI, Ops, Ops+3, CI->getOperand(1)->getType(), 833 MemsetFCache); 834 break; 835 } 836 case Intrinsic::sqrt: { 837 static Constant *sqrtfFCache = 0; 838 static Constant *sqrtFCache = 0; 839 static Constant *sqrtLDCache = 0; 840 switch (CI->getOperand(1)->getType()->getTypeID()) { 841 default: assert(0 && "Invalid type in sqrt"); abort(); 842 case Type::FloatTyID: 843 ReplaceCallWith("sqrtf", CI, CI->op_begin()+1, CI->op_end(), 844 Type::FloatTy, sqrtfFCache); 845 break; 846 case Type::DoubleTyID: 847 ReplaceCallWith("sqrt", CI, CI->op_begin()+1, CI->op_end(), 848 Type::DoubleTy, sqrtFCache); 849 break; 850 case Type::X86_FP80TyID: 851 case Type::FP128TyID: 852 case Type::PPC_FP128TyID: 853 ReplaceCallWith("sqrtl", CI, CI->op_begin()+1, CI->op_end(), 854 CI->getOperand(1)->getType(), sqrtLDCache); 855 break; 856 } 857 break; 858 } 859 } 860 861 assert(CI->use_empty() && 862 "Lowering should have eliminated any uses of the intrinsic call!"); 863 CI->eraseFromParent(); 864} 865