IntrinsicLowering.cpp revision 9a9203b8948cc4ab275ca21d34821589e272c47c
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" 23using namespace llvm; 24 25template <class ArgIt> 26static 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. 40template <class ArgIt> 41static 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 64void 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. 115static 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. 201static Value *LowerCTPOP(Value *V, Instruction *IP) { 202 assert(V->getType()->isInteger() && "Can't ctpop a non-integer type!"); 203 204 static const uint64_t MaskValues[6] = { 205 0x5555555555555555ULL, 0x3333333333333333ULL, 206 0x0F0F0F0F0F0F0F0FULL, 0x00FF00FF00FF00FFULL, 207 0x0000FFFF0000FFFFULL, 0x00000000FFFFFFFFULL 208 }; 209 210 unsigned BitSize = V->getType()->getPrimitiveSizeInBits(); 211 212 for (unsigned i = 1, ct = 0; i != BitSize; i <<= 1, ++ct) { 213 Value *MaskCst = ConstantInt::get(V->getType(), MaskValues[ct]); 214 Value *LHS = BinaryOperator::createAnd(V, MaskCst, "cppop.and1", IP); 215 Value *VShift = BinaryOperator::createLShr(V, 216 ConstantInt::get(V->getType(), i), "ctpop.sh", IP); 217 Value *RHS = BinaryOperator::createAnd(VShift, MaskCst, "cppop.and2", IP); 218 V = BinaryOperator::createAdd(LHS, RHS, "ctpop.step", IP); 219 } 220 221 return CastInst::createIntegerCast(V, Type::Int32Ty, false, "ctpop", IP); 222} 223 224/// LowerCTLZ - Emit the code to lower ctlz of V before the specified 225/// instruction IP. 226static Value *LowerCTLZ(Value *V, Instruction *IP) { 227 228 unsigned BitSize = V->getType()->getPrimitiveSizeInBits(); 229 for (unsigned i = 1; i != BitSize; i <<= 1) { 230 Value *ShVal = ConstantInt::get(V->getType(), i); 231 ShVal = BinaryOperator::createLShr(V, ShVal, "ctlz.sh", IP); 232 V = BinaryOperator::createOr(V, ShVal, "ctlz.step", IP); 233 } 234 235 V = BinaryOperator::createNot(V, "", IP); 236 return LowerCTPOP(V, IP); 237} 238 239/// Convert the llvm.part.select.iX.iY intrinsic. This intrinsic takes 240/// three integer arguments. The first argument is the Value from which the 241/// bits will be selected. It may be of any bit width. The second and third 242/// arguments specify a range of bits to select with the second argument 243/// specifying the low bit and the third argument specifying the high bit. Both 244/// must be type i32. The result is the corresponding selected bits from the 245/// Value in the same width as the Value (first argument). If the low bit index 246/// is higher than the high bit index then the inverse selection is done and 247/// the bits are returned in inverse order. 248/// @brief Lowering of llvm.part.select intrinsic. 249static Instruction *LowerPartSelect(CallInst *CI) { 250 // Make sure we're dealing with a part select intrinsic here 251 Function *F = CI->getCalledFunction(); 252 const FunctionType *FT = F->getFunctionType(); 253 if (!F->isDeclaration() || !FT->getReturnType()->isInteger() || 254 FT->getNumParams() != 3 || !FT->getParamType(0)->isInteger() || 255 !FT->getParamType(1)->isInteger() || !FT->getParamType(2)->isInteger()) 256 return CI; 257 258 // Get the intrinsic implementation function by converting all the . to _ 259 // in the intrinsic's function name and then reconstructing the function 260 // declaration. 261 std::string Name(F->getName()); 262 for (unsigned i = 4; i < Name.length(); ++i) 263 if (Name[i] == '.') 264 Name[i] = '_'; 265 Module* M = F->getParent(); 266 F = cast<Function>(M->getOrInsertFunction(Name, FT)); 267 F->setLinkage(GlobalValue::WeakLinkage); 268 269 // If we haven't defined the impl function yet, do so now 270 if (F->isDeclaration()) { 271 272 // Get the arguments to the function 273 Function::arg_iterator args = F->arg_begin(); 274 Value* Val = args++; Val->setName("Val"); 275 Value* Lo = args++; Lo->setName("Lo"); 276 Value* Hi = args++; Hi->setName("High"); 277 278 // We want to select a range of bits here such that [Hi, Lo] is shifted 279 // down to the low bits. However, it is quite possible that Hi is smaller 280 // than Lo in which case the bits have to be reversed. 281 282 // Create the blocks we will need for the two cases (forward, reverse) 283 BasicBlock* CurBB = new BasicBlock("entry", F); 284 BasicBlock *RevSize = new BasicBlock("revsize", CurBB->getParent()); 285 BasicBlock *FwdSize = new BasicBlock("fwdsize", CurBB->getParent()); 286 BasicBlock *Compute = new BasicBlock("compute", CurBB->getParent()); 287 BasicBlock *Reverse = new BasicBlock("reverse", CurBB->getParent()); 288 BasicBlock *RsltBlk = new BasicBlock("result", CurBB->getParent()); 289 290 // Cast Hi and Lo to the size of Val so the widths are all the same 291 if (Hi->getType() != Val->getType()) 292 Hi = CastInst::createIntegerCast(Hi, Val->getType(), false, 293 "tmp", CurBB); 294 if (Lo->getType() != Val->getType()) 295 Lo = CastInst::createIntegerCast(Lo, Val->getType(), false, 296 "tmp", CurBB); 297 298 // Compute a few things that both cases will need, up front. 299 Constant* Zero = ConstantInt::get(Val->getType(), 0); 300 Constant* One = ConstantInt::get(Val->getType(), 1); 301 Constant* AllOnes = ConstantInt::getAllOnesValue(Val->getType()); 302 303 // Compare the Hi and Lo bit positions. This is used to determine 304 // which case we have (forward or reverse) 305 ICmpInst *Cmp = new ICmpInst(ICmpInst::ICMP_ULT, Hi, Lo, "less",CurBB); 306 new BranchInst(RevSize, FwdSize, Cmp, CurBB); 307 308 // First, copmute the number of bits in the forward case. 309 Instruction* FBitSize = 310 BinaryOperator::createSub(Hi, Lo,"fbits", FwdSize); 311 new BranchInst(Compute, FwdSize); 312 313 // Second, compute the number of bits in the reverse case. 314 Instruction* RBitSize = 315 BinaryOperator::createSub(Lo, Hi, "rbits", RevSize); 316 new BranchInst(Compute, RevSize); 317 318 // Now, compute the bit range. Start by getting the bitsize and the shift 319 // amount (either Hi or Lo) from PHI nodes. Then we compute a mask for 320 // the number of bits we want in the range. We shift the bits down to the 321 // least significant bits, apply the mask to zero out unwanted high bits, 322 // and we have computed the "forward" result. It may still need to be 323 // reversed. 324 325 // Get the BitSize from one of the two subtractions 326 PHINode *BitSize = new PHINode(Val->getType(), "bits", Compute); 327 BitSize->reserveOperandSpace(2); 328 BitSize->addIncoming(FBitSize, FwdSize); 329 BitSize->addIncoming(RBitSize, RevSize); 330 331 // Get the ShiftAmount as the smaller of Hi/Lo 332 PHINode *ShiftAmt = new PHINode(Val->getType(), "shiftamt", Compute); 333 ShiftAmt->reserveOperandSpace(2); 334 ShiftAmt->addIncoming(Lo, FwdSize); 335 ShiftAmt->addIncoming(Hi, RevSize); 336 337 // Increment the bit size 338 Instruction *BitSizePlusOne = 339 BinaryOperator::createAdd(BitSize, One, "bits", Compute); 340 341 // Create a Mask to zero out the high order bits. 342 Instruction* Mask = 343 BinaryOperator::createShl(AllOnes, BitSizePlusOne, "mask", Compute); 344 Mask = BinaryOperator::createNot(Mask, "mask", Compute); 345 346 // Shift the bits down and apply the mask 347 Instruction* FRes = 348 BinaryOperator::createLShr(Val, ShiftAmt, "fres", Compute); 349 FRes = BinaryOperator::createAnd(FRes, Mask, "fres", Compute); 350 new BranchInst(Reverse, RsltBlk, Cmp, Compute); 351 352 // In the Reverse block we have the mask already in FRes but we must reverse 353 // it by shifting FRes bits right and putting them in RRes by shifting them 354 // in from left. 355 356 // First set up our loop counters 357 PHINode *Count = new PHINode(Val->getType(), "count", Reverse); 358 Count->reserveOperandSpace(2); 359 Count->addIncoming(BitSizePlusOne, Compute); 360 361 // Next, get the value that we are shifting. 362 PHINode *BitsToShift = new PHINode(Val->getType(), "val", Reverse); 363 BitsToShift->reserveOperandSpace(2); 364 BitsToShift->addIncoming(FRes, Compute); 365 366 // Finally, get the result of the last computation 367 PHINode *RRes = new PHINode(Val->getType(), "rres", Reverse); 368 RRes->reserveOperandSpace(2); 369 RRes->addIncoming(Zero, Compute); 370 371 // Decrement the counter 372 Instruction *Decr = BinaryOperator::createSub(Count, One, "decr", Reverse); 373 Count->addIncoming(Decr, Reverse); 374 375 // Compute the Bit that we want to move 376 Instruction *Bit = 377 BinaryOperator::createAnd(BitsToShift, One, "bit", Reverse); 378 379 // Compute the new value for next iteration. 380 Instruction *NewVal = 381 BinaryOperator::createLShr(BitsToShift, One, "rshift", Reverse); 382 BitsToShift->addIncoming(NewVal, Reverse); 383 384 // Shift the bit into the low bits of the result. 385 Instruction *NewRes = 386 BinaryOperator::createShl(RRes, One, "lshift", Reverse); 387 NewRes = BinaryOperator::createOr(NewRes, Bit, "addbit", Reverse); 388 RRes->addIncoming(NewRes, Reverse); 389 390 // Terminate loop if we've moved all the bits. 391 ICmpInst *Cond = 392 new ICmpInst(ICmpInst::ICMP_EQ, Decr, Zero, "cond", Reverse); 393 new BranchInst(RsltBlk, Reverse, Cond, Reverse); 394 395 // Finally, in the result block, select one of the two results with a PHI 396 // node and return the result; 397 CurBB = RsltBlk; 398 PHINode *BitSelect = new PHINode(Val->getType(), "part_select", CurBB); 399 BitSelect->reserveOperandSpace(2); 400 BitSelect->addIncoming(FRes, Compute); 401 BitSelect->addIncoming(NewRes, Reverse); 402 new ReturnInst(BitSelect, CurBB); 403 } 404 405 // Return a call to the implementation function 406 Value *Args[3]; 407 Args[0] = CI->getOperand(1); 408 Args[1] = CI->getOperand(2); 409 Args[2] = CI->getOperand(3); 410 return new CallInst(F, Args, 3, CI->getName(), CI); 411} 412 413/// Convert the llvm.part.set.iX.iY.iZ intrinsic. This intrinsic takes 414/// four integer arguments (iAny %Value, iAny %Replacement, i32 %Low, i32 %High) 415/// The first two arguments can be any bit width. The result is the same width 416/// as %Value. The operation replaces bits between %Low and %High with the value 417/// in %Replacement. If %Replacement is not the same width, it is truncated or 418/// zero extended as appropriate to fit the bits being replaced. If %Low is 419/// greater than %High then the inverse set of bits are replaced. 420/// @brief Lowering of llvm.bit.part.set intrinsic. 421static Instruction *LowerPartSet(CallInst *CI) { 422 // Make sure we're dealing with a part select intrinsic here 423 Function *F = CI->getCalledFunction(); 424 const FunctionType *FT = F->getFunctionType(); 425 if (!F->isDeclaration() || !FT->getReturnType()->isInteger() || 426 FT->getNumParams() != 4 || !FT->getParamType(0)->isInteger() || 427 !FT->getParamType(1)->isInteger() || !FT->getParamType(2)->isInteger() || 428 !FT->getParamType(3)->isInteger()) 429 return CI; 430 431 // Get the intrinsic implementation function by converting all the . to _ 432 // in the intrinsic's function name and then reconstructing the function 433 // declaration. 434 std::string Name(F->getName()); 435 for (unsigned i = 4; i < Name.length(); ++i) 436 if (Name[i] == '.') 437 Name[i] = '_'; 438 Module* M = F->getParent(); 439 F = cast<Function>(M->getOrInsertFunction(Name, FT)); 440 F->setLinkage(GlobalValue::WeakLinkage); 441 442 // If we haven't defined the impl function yet, do so now 443 if (F->isDeclaration()) { 444 // Note: the following code is based on code generated by llvm2cpp with 445 // the following input. This is just *one* example of a generated function. 446 // The functions vary by bit width of result and first two arguments. 447 // The generated code has been changed to deal with any bit width not just 448 // the 32/64 bitwidths used in the above sample. 449 // 450 // define i64 @part_set(i64 %Val, i32 %Rep, i32 %Lo, i32 %Hi) { 451 // entry: 452 // %is_forward = icmp ult i32 %Lo, %Hi 453 // %Lo.pn = select i1 %is_forward, i32 %Hi, i32 %Lo 454 // %Hi.pn = select i1 %is_forward, i32 %Lo, i32 %Hi 455 // %iftmp.16.0 = sub i32 %Lo.pn, %Hi.pn 456 // icmp ult i32 %iftmp.16.0, 32 457 // br i1 %1, label %cond_true11, label %cond_next19 458 // cond_true11: 459 // %tmp13 = sub i32 32, %iftmp.16.0 460 // %tmp14 = lshr i32 -1, %tmp13 461 // %tmp16 = and i32 %tmp14, %Rep 462 // br label %cond_next19 463 // cond_next19: 464 // %iftmp.17.0 = phi i32 [ %tmp16, %cond_true11 ], [ %Rep, %entry ] 465 // %tmp2021 = zext i32 %iftmp.17.0 to i64 466 // icmp ugt i32 %Lo, %Hi 467 // br i1 %2, label %cond_next60, label %cond_true24 468 // cond_true24: 469 // %tmp25.cast = zext i32 %Hi to i64 470 // %tmp26 = lshr i64 -1, %tmp25.cast 471 // %tmp27.cast = zext i32 %Lo to i64 472 // %tmp28 = shl i64 %tmp26, %tmp27.cast 473 // %tmp28not = xor i64 %tmp28, -1 474 // %tmp31 = shl i64 %tmp2021, %tmp27.cast 475 // %tmp34 = and i64 %tmp28not, %Val 476 // %Val_addr.064 = or i64 %tmp31, %tmp34 477 // ret i64 %Val_addr.064 478 // cond_next60: 479 // %tmp39.cast = zext i32 %Lo to i64 480 // %tmp40 = shl i64 -1, %tmp39.cast 481 // %tmp41.cast = zext i32 %Hi to i64 482 // %tmp42 = shl i64 -1, %tmp41.cast 483 // %tmp45.demorgan = or i64 %tmp42, %tmp40 484 // %tmp45 = xor i64 %tmp45.demorgan, -1 485 // %tmp47 = and i64 %tmp45, %Val 486 // %tmp50 = shl i64 %tmp2021, %tmp39.cast 487 // %tmp52 = sub i32 32, %Hi 488 // %tmp52.cast = zext i32 %tmp52 to i64 489 // %tmp54 = lshr i64 %tmp2021, %tmp52.cast 490 // %tmp57 = or i64 %tmp50, %tmp47 491 // %Val_addr.0 = or i64 %tmp57, %tmp54 492 // ret i64 %Val_addr.0 493 // } 494 495 // Get the arguments for the function. 496 Function::arg_iterator args = F->arg_begin(); 497 Value* Val = args++; Val->setName("Val"); 498 Value* Rep = args++; Rep->setName("Rep"); 499 Value* Lo = args++; Lo->setName("Lo"); 500 Value* Hi = args++; Hi->setName("Hi"); 501 502 // Get some types we need 503 const IntegerType* ValTy = cast<IntegerType>(Val->getType()); 504 const IntegerType* RepTy = cast<IntegerType>(Rep->getType()); 505 uint32_t ValBits = ValTy->getBitWidth(); 506 uint32_t RepBits = RepTy->getBitWidth(); 507 508 // Constant Definitions 509 ConstantInt* RepBitWidth = ConstantInt::get(Type::Int32Ty, RepBits); 510 ConstantInt* RepMask = ConstantInt::getAllOnesValue(RepTy); 511 ConstantInt* ValMask = ConstantInt::getAllOnesValue(ValTy); 512 513 BasicBlock* entry = new BasicBlock("entry",F,0); 514 BasicBlock* large = new BasicBlock("large",F,0); 515 BasicBlock* small = new BasicBlock("small",F,0); 516 BasicBlock* forward = new BasicBlock("forward",F,0); 517 BasicBlock* reverse = new BasicBlock("reverse",F,0); 518 519 // Block entry (entry) 520 // First, get the number of bits that we're placing as an i32 521 ICmpInst* is_forward = 522 new ICmpInst(ICmpInst::ICMP_ULT, Lo, Hi, "", entry); 523 SelectInst* Lo_pn = new SelectInst(is_forward, Hi, Lo, "", entry); 524 SelectInst* Hi_pn = new SelectInst(is_forward, Lo, Hi, "", entry); 525 BinaryOperator* NumBits = BinaryOperator::createSub(Lo_pn, Hi_pn, "",entry); 526 // Now, convert Lo and Hi to ValTy bit width 527 if (ValBits > 32) { 528 Hi = new ZExtInst(Hi, ValTy, "", entry); 529 Lo = new ZExtInst(Lo, ValTy, "", entry); 530 } else if (ValBits < 32) { 531 Hi = new TruncInst(Hi, ValTy, "", entry); 532 Lo = new TruncInst(Lo, ValTy, "", entry); 533 } 534 // Determine if the replacement bits are larger than the number of bits we 535 // are replacing and deal with it. 536 ICmpInst* is_large = 537 new ICmpInst(ICmpInst::ICMP_ULT, NumBits, RepBitWidth, "", entry); 538 new BranchInst(large, small, is_large, entry); 539 540 // Block "large" 541 Instruction* MaskBits = 542 BinaryOperator::createSub(RepBitWidth, NumBits, "", large); 543 MaskBits = CastInst::createIntegerCast(MaskBits, RepMask->getType(), 544 false, "", large); 545 BinaryOperator* Mask1 = 546 BinaryOperator::createLShr(RepMask, MaskBits, "", large); 547 BinaryOperator* Rep2 = BinaryOperator::createAnd(Mask1, Rep, "", large); 548 new BranchInst(small, large); 549 550 // Block "small" 551 PHINode* Rep3 = new PHINode(RepTy, "", small); 552 Rep3->reserveOperandSpace(2); 553 Rep3->addIncoming(Rep2, large); 554 Rep3->addIncoming(Rep, entry); 555 Value* Rep4 = Rep3; 556 if (ValBits > RepBits) 557 Rep4 = new ZExtInst(Rep3, ValTy, "", small); 558 else if (ValBits < RepBits) 559 Rep4 = new TruncInst(Rep3, ValTy, "", small); 560 ICmpInst* is_reverse = 561 new ICmpInst(ICmpInst::ICMP_UGT, Lo, Hi, "", small); 562 new BranchInst(reverse, forward, is_reverse, small); 563 564 // Block "forward" 565 Value* t1 = BinaryOperator::createLShr(ValMask, Hi, "", forward); 566 Value* t2 = BinaryOperator::createShl(t1, Lo, "", forward); 567 Value* nott2 = BinaryOperator::createXor(t2, ValMask, "", forward); 568 Value* t3 = BinaryOperator::createShl(Rep4, Lo, "", forward); 569 Value* t4 = BinaryOperator::createAnd(nott2, Val, "", forward); 570 Value* FRslt = BinaryOperator::createOr(t3, t4, "", forward); 571 new ReturnInst(FRslt, forward); 572 573 // Block "reverse" 574 Value* t5 = BinaryOperator::createShl(ValMask, Lo, "", reverse); 575 Value* t6 = BinaryOperator::createShl(ValMask, Hi, "", reverse); 576 Value* t7 = BinaryOperator::createOr(t6, t5, "", reverse); 577 Value* t8 = BinaryOperator::createXor(t7, ValMask, "", reverse); 578 Value* t9 = BinaryOperator::createAnd(t8, Val, "", reverse); 579 Value* t10 = BinaryOperator::createShl(Rep4, Lo, "", reverse); 580 if (32 < ValBits) 581 RepBitWidth = 582 cast<ConstantInt>(ConstantExpr::getZExt(RepBitWidth, ValTy)); 583 else if (32 > ValBits) 584 RepBitWidth = 585 cast<ConstantInt>(ConstantExpr::getTrunc(RepBitWidth, ValTy)); 586 Value* t11 = BinaryOperator::createSub(RepBitWidth, Hi, "", reverse); 587 Value* t13 = BinaryOperator::createLShr(Rep4, t11, "",reverse); 588 Value* t14 = BinaryOperator::createOr(t10, t9, "", reverse); 589 Value* RRslt = BinaryOperator::createOr(t14, t13, "", reverse); 590 new ReturnInst(RRslt, reverse); 591 } 592 593 // Return a call to the implementation function 594 Value *Args[3]; 595 Args[0] = CI->getOperand(1); 596 Args[1] = CI->getOperand(2); 597 Args[2] = CI->getOperand(3); 598 Args[3] = CI->getOperand(4); 599 return new CallInst(F, Args, 4, CI->getName(), CI); 600} 601 602 603void IntrinsicLowering::LowerIntrinsicCall(CallInst *CI) { 604 Function *Callee = CI->getCalledFunction(); 605 assert(Callee && "Cannot lower an indirect call!"); 606 607 switch (Callee->getIntrinsicID()) { 608 case Intrinsic::not_intrinsic: 609 cerr << "Cannot lower a call to a non-intrinsic function '" 610 << Callee->getName() << "'!\n"; 611 abort(); 612 default: 613 cerr << "Error: Code generator does not support intrinsic function '" 614 << Callee->getName() << "'!\n"; 615 abort(); 616 617 // The setjmp/longjmp intrinsics should only exist in the code if it was 618 // never optimized (ie, right out of the CFE), or if it has been hacked on 619 // by the lowerinvoke pass. In both cases, the right thing to do is to 620 // convert the call to an explicit setjmp or longjmp call. 621 case Intrinsic::setjmp: { 622 static Constant *SetjmpFCache = 0; 623 Value *V = ReplaceCallWith("setjmp", CI, CI->op_begin()+1, CI->op_end(), 624 Type::Int32Ty, SetjmpFCache); 625 if (CI->getType() != Type::VoidTy) 626 CI->replaceAllUsesWith(V); 627 break; 628 } 629 case Intrinsic::sigsetjmp: 630 if (CI->getType() != Type::VoidTy) 631 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType())); 632 break; 633 634 case Intrinsic::longjmp: { 635 static Constant *LongjmpFCache = 0; 636 ReplaceCallWith("longjmp", CI, CI->op_begin()+1, CI->op_end(), 637 Type::VoidTy, LongjmpFCache); 638 break; 639 } 640 641 case Intrinsic::siglongjmp: { 642 // Insert the call to abort 643 static Constant *AbortFCache = 0; 644 ReplaceCallWith("abort", CI, CI->op_end(), CI->op_end(), 645 Type::VoidTy, AbortFCache); 646 break; 647 } 648 case Intrinsic::ctpop: 649 CI->replaceAllUsesWith(LowerCTPOP(CI->getOperand(1), CI)); 650 break; 651 652 case Intrinsic::bswap: 653 CI->replaceAllUsesWith(LowerBSWAP(CI->getOperand(1), CI)); 654 break; 655 656 case Intrinsic::ctlz: 657 CI->replaceAllUsesWith(LowerCTLZ(CI->getOperand(1), CI)); 658 break; 659 660 case Intrinsic::cttz: { 661 // cttz(x) -> ctpop(~X & (X-1)) 662 Value *Src = CI->getOperand(1); 663 Value *NotSrc = BinaryOperator::createNot(Src, Src->getName()+".not", CI); 664 Value *SrcM1 = ConstantInt::get(Src->getType(), 1); 665 SrcM1 = BinaryOperator::createSub(Src, SrcM1, "", CI); 666 Src = LowerCTPOP(BinaryOperator::createAnd(NotSrc, SrcM1, "", CI), CI); 667 CI->replaceAllUsesWith(Src); 668 break; 669 } 670 671 case Intrinsic::part_select: 672 CI->replaceAllUsesWith(LowerPartSelect(CI)); 673 break; 674 675 case Intrinsic::part_set: 676 CI->replaceAllUsesWith(LowerPartSet(CI)); 677 break; 678 679 case Intrinsic::stacksave: 680 case Intrinsic::stackrestore: { 681 static bool Warned = false; 682 if (!Warned) 683 cerr << "WARNING: this target does not support the llvm.stack" 684 << (Callee->getIntrinsicID() == Intrinsic::stacksave ? 685 "save" : "restore") << " intrinsic.\n"; 686 Warned = true; 687 if (Callee->getIntrinsicID() == Intrinsic::stacksave) 688 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType())); 689 break; 690 } 691 692 case Intrinsic::returnaddress: 693 case Intrinsic::frameaddress: 694 cerr << "WARNING: this target does not support the llvm." 695 << (Callee->getIntrinsicID() == Intrinsic::returnaddress ? 696 "return" : "frame") << "address intrinsic.\n"; 697 CI->replaceAllUsesWith(ConstantPointerNull::get( 698 cast<PointerType>(CI->getType()))); 699 break; 700 701 case Intrinsic::prefetch: 702 break; // Simply strip out prefetches on unsupported architectures 703 704 case Intrinsic::pcmarker: 705 break; // Simply strip out pcmarker on unsupported architectures 706 case Intrinsic::readcyclecounter: { 707 cerr << "WARNING: this target does not support the llvm.readcyclecoun" 708 << "ter intrinsic. It is being lowered to a constant 0\n"; 709 CI->replaceAllUsesWith(ConstantInt::get(Type::Int64Ty, 0)); 710 break; 711 } 712 713 case Intrinsic::dbg_stoppoint: 714 case Intrinsic::dbg_region_start: 715 case Intrinsic::dbg_region_end: 716 case Intrinsic::dbg_func_start: 717 case Intrinsic::dbg_declare: 718 case Intrinsic::eh_exception: 719 case Intrinsic::eh_selector: 720 case Intrinsic::eh_filter: 721 break; // Simply strip out debugging and eh intrinsics 722 723 case Intrinsic::memcpy_i32: 724 case Intrinsic::memcpy_i64: { 725 static Constant *MemcpyFCache = 0; 726 Value *Size = CI->getOperand(3); 727 const Type *IntPtr = TD.getIntPtrType(); 728 if (Size->getType()->getPrimitiveSizeInBits() < 729 IntPtr->getPrimitiveSizeInBits()) 730 Size = new ZExtInst(Size, IntPtr, "", CI); 731 else if (Size->getType()->getPrimitiveSizeInBits() > 732 IntPtr->getPrimitiveSizeInBits()) 733 Size = new TruncInst(Size, IntPtr, "", CI); 734 Value *Ops[3]; 735 Ops[0] = CI->getOperand(1); 736 Ops[1] = CI->getOperand(2); 737 Ops[2] = Size; 738 ReplaceCallWith("memcpy", CI, Ops, Ops+3, CI->getOperand(1)->getType(), 739 MemcpyFCache); 740 break; 741 } 742 case Intrinsic::memmove_i32: 743 case Intrinsic::memmove_i64: { 744 static Constant *MemmoveFCache = 0; 745 Value *Size = CI->getOperand(3); 746 const Type *IntPtr = TD.getIntPtrType(); 747 if (Size->getType()->getPrimitiveSizeInBits() < 748 IntPtr->getPrimitiveSizeInBits()) 749 Size = new ZExtInst(Size, IntPtr, "", CI); 750 else if (Size->getType()->getPrimitiveSizeInBits() > 751 IntPtr->getPrimitiveSizeInBits()) 752 Size = new TruncInst(Size, IntPtr, "", CI); 753 Value *Ops[3]; 754 Ops[0] = CI->getOperand(1); 755 Ops[1] = CI->getOperand(2); 756 Ops[2] = Size; 757 ReplaceCallWith("memmove", CI, Ops, Ops+3, CI->getOperand(1)->getType(), 758 MemmoveFCache); 759 break; 760 } 761 case Intrinsic::memset_i32: 762 case Intrinsic::memset_i64: { 763 static Constant *MemsetFCache = 0; 764 Value *Size = CI->getOperand(3); 765 const Type *IntPtr = TD.getIntPtrType(); 766 if (Size->getType()->getPrimitiveSizeInBits() < 767 IntPtr->getPrimitiveSizeInBits()) 768 Size = new ZExtInst(Size, IntPtr, "", CI); 769 else if (Size->getType()->getPrimitiveSizeInBits() > 770 IntPtr->getPrimitiveSizeInBits()) 771 Size = new TruncInst(Size, IntPtr, "", CI); 772 Value *Ops[3]; 773 Ops[0] = CI->getOperand(1); 774 // Extend the amount to i32. 775 Ops[1] = new ZExtInst(CI->getOperand(2), Type::Int32Ty, "", CI); 776 Ops[2] = Size; 777 ReplaceCallWith("memset", CI, Ops, Ops+3, CI->getOperand(1)->getType(), 778 MemsetFCache); 779 break; 780 } 781 case Intrinsic::sqrt_f32: { 782 static Constant *sqrtfFCache = 0; 783 ReplaceCallWith("sqrtf", CI, CI->op_begin()+1, CI->op_end(), 784 Type::FloatTy, sqrtfFCache); 785 break; 786 } 787 case Intrinsic::sqrt_f64: { 788 static Constant *sqrtFCache = 0; 789 ReplaceCallWith("sqrt", CI, CI->op_begin()+1, CI->op_end(), 790 Type::DoubleTy, sqrtFCache); 791 break; 792 } 793 } 794 795 assert(CI->use_empty() && 796 "Lowering should have eliminated any uses of the intrinsic call!"); 797 CI->eraseFromParent(); 798} 799