1//===-- TargetLowering.cpp - Implement the TargetLowering class -----------===// 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 implements the TargetLowering class. 11// 12//===----------------------------------------------------------------------===// 13 14#include "llvm/Target/TargetLowering.h" 15#include "llvm/ADT/BitVector.h" 16#include "llvm/ADT/STLExtras.h" 17#include "llvm/CodeGen/Analysis.h" 18#include "llvm/CodeGen/MachineFrameInfo.h" 19#include "llvm/CodeGen/MachineFunction.h" 20#include "llvm/CodeGen/MachineJumpTableInfo.h" 21#include "llvm/CodeGen/SelectionDAG.h" 22#include "llvm/IR/DataLayout.h" 23#include "llvm/IR/DerivedTypes.h" 24#include "llvm/IR/GlobalVariable.h" 25#include "llvm/IR/LLVMContext.h" 26#include "llvm/MC/MCAsmInfo.h" 27#include "llvm/MC/MCExpr.h" 28#include "llvm/Support/CommandLine.h" 29#include "llvm/Support/ErrorHandling.h" 30#include "llvm/Support/MathExtras.h" 31#include "llvm/Target/TargetLoweringObjectFile.h" 32#include "llvm/Target/TargetMachine.h" 33#include "llvm/Target/TargetRegisterInfo.h" 34#include "llvm/Target/TargetSubtargetInfo.h" 35#include <cctype> 36using namespace llvm; 37 38/// NOTE: The TargetMachine owns TLOF. 39TargetLowering::TargetLowering(const TargetMachine &tm) 40 : TargetLoweringBase(tm) {} 41 42const char *TargetLowering::getTargetNodeName(unsigned Opcode) const { 43 return nullptr; 44} 45 46/// Check whether a given call node is in tail position within its function. If 47/// so, it sets Chain to the input chain of the tail call. 48bool TargetLowering::isInTailCallPosition(SelectionDAG &DAG, SDNode *Node, 49 SDValue &Chain) const { 50 const Function *F = DAG.getMachineFunction().getFunction(); 51 52 // Conservatively require the attributes of the call to match those of 53 // the return. Ignore noalias because it doesn't affect the call sequence. 54 AttributeSet CallerAttrs = F->getAttributes(); 55 if (AttrBuilder(CallerAttrs, AttributeSet::ReturnIndex) 56 .removeAttribute(Attribute::NoAlias).hasAttributes()) 57 return false; 58 59 // It's not safe to eliminate the sign / zero extension of the return value. 60 if (CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt) || 61 CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt)) 62 return false; 63 64 // Check if the only use is a function return node. 65 return isUsedByReturnOnly(Node, Chain); 66} 67 68/// \brief Set CallLoweringInfo attribute flags based on a call instruction 69/// and called function attributes. 70void TargetLowering::ArgListEntry::setAttributes(ImmutableCallSite *CS, 71 unsigned AttrIdx) { 72 isSExt = CS->paramHasAttr(AttrIdx, Attribute::SExt); 73 isZExt = CS->paramHasAttr(AttrIdx, Attribute::ZExt); 74 isInReg = CS->paramHasAttr(AttrIdx, Attribute::InReg); 75 isSRet = CS->paramHasAttr(AttrIdx, Attribute::StructRet); 76 isNest = CS->paramHasAttr(AttrIdx, Attribute::Nest); 77 isByVal = CS->paramHasAttr(AttrIdx, Attribute::ByVal); 78 isInAlloca = CS->paramHasAttr(AttrIdx, Attribute::InAlloca); 79 isReturned = CS->paramHasAttr(AttrIdx, Attribute::Returned); 80 Alignment = CS->getParamAlignment(AttrIdx); 81} 82 83/// Generate a libcall taking the given operands as arguments and returning a 84/// result of type RetVT. 85std::pair<SDValue, SDValue> 86TargetLowering::makeLibCall(SelectionDAG &DAG, 87 RTLIB::Libcall LC, EVT RetVT, 88 ArrayRef<SDValue> Ops, 89 bool isSigned, SDLoc dl, 90 bool doesNotReturn, 91 bool isReturnValueUsed) const { 92 TargetLowering::ArgListTy Args; 93 Args.reserve(Ops.size()); 94 95 TargetLowering::ArgListEntry Entry; 96 for (SDValue Op : Ops) { 97 Entry.Node = Op; 98 Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext()); 99 Entry.isSExt = shouldSignExtendTypeInLibCall(Op.getValueType(), isSigned); 100 Entry.isZExt = !shouldSignExtendTypeInLibCall(Op.getValueType(), isSigned); 101 Args.push_back(Entry); 102 } 103 104 markInRegArguments(DAG, Args); 105 106 if (LC == RTLIB::UNKNOWN_LIBCALL) 107 report_fatal_error("Unsupported library call operation!"); 108 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC), 109 getPointerTy(DAG.getDataLayout())); 110 111 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext()); 112 TargetLowering::CallLoweringInfo CLI(DAG); 113 bool signExtend = shouldSignExtendTypeInLibCall(RetVT, isSigned); 114 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode()) 115 .setCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0) 116 .setNoReturn(doesNotReturn).setDiscardResult(!isReturnValueUsed) 117 .setSExtResult(signExtend).setZExtResult(!signExtend); 118 return LowerCallTo(CLI); 119} 120 121/// SoftenSetCCOperands - Soften the operands of a comparison. This code is 122/// shared among BR_CC, SELECT_CC, and SETCC handlers. 123void TargetLowering::softenSetCCOperands(SelectionDAG &DAG, EVT VT, 124 SDValue &NewLHS, SDValue &NewRHS, 125 ISD::CondCode &CCCode, 126 SDLoc dl) const { 127 assert((VT == MVT::f32 || VT == MVT::f64 || VT == MVT::f128) 128 && "Unsupported setcc type!"); 129 130 // Expand into one or more soft-fp libcall(s). 131 RTLIB::Libcall LC1 = RTLIB::UNKNOWN_LIBCALL, LC2 = RTLIB::UNKNOWN_LIBCALL; 132 bool ShouldInvertCC = false; 133 switch (CCCode) { 134 case ISD::SETEQ: 135 case ISD::SETOEQ: 136 LC1 = (VT == MVT::f32) ? RTLIB::OEQ_F32 : 137 (VT == MVT::f64) ? RTLIB::OEQ_F64 : RTLIB::OEQ_F128; 138 break; 139 case ISD::SETNE: 140 case ISD::SETUNE: 141 LC1 = (VT == MVT::f32) ? RTLIB::UNE_F32 : 142 (VT == MVT::f64) ? RTLIB::UNE_F64 : RTLIB::UNE_F128; 143 break; 144 case ISD::SETGE: 145 case ISD::SETOGE: 146 LC1 = (VT == MVT::f32) ? RTLIB::OGE_F32 : 147 (VT == MVT::f64) ? RTLIB::OGE_F64 : RTLIB::OGE_F128; 148 break; 149 case ISD::SETLT: 150 case ISD::SETOLT: 151 LC1 = (VT == MVT::f32) ? RTLIB::OLT_F32 : 152 (VT == MVT::f64) ? RTLIB::OLT_F64 : RTLIB::OLT_F128; 153 break; 154 case ISD::SETLE: 155 case ISD::SETOLE: 156 LC1 = (VT == MVT::f32) ? RTLIB::OLE_F32 : 157 (VT == MVT::f64) ? RTLIB::OLE_F64 : RTLIB::OLE_F128; 158 break; 159 case ISD::SETGT: 160 case ISD::SETOGT: 161 LC1 = (VT == MVT::f32) ? RTLIB::OGT_F32 : 162 (VT == MVT::f64) ? RTLIB::OGT_F64 : RTLIB::OGT_F128; 163 break; 164 case ISD::SETUO: 165 LC1 = (VT == MVT::f32) ? RTLIB::UO_F32 : 166 (VT == MVT::f64) ? RTLIB::UO_F64 : RTLIB::UO_F128; 167 break; 168 case ISD::SETO: 169 LC1 = (VT == MVT::f32) ? RTLIB::O_F32 : 170 (VT == MVT::f64) ? RTLIB::O_F64 : RTLIB::O_F128; 171 break; 172 case ISD::SETONE: 173 // SETONE = SETOLT | SETOGT 174 LC1 = (VT == MVT::f32) ? RTLIB::OLT_F32 : 175 (VT == MVT::f64) ? RTLIB::OLT_F64 : RTLIB::OLT_F128; 176 LC2 = (VT == MVT::f32) ? RTLIB::OGT_F32 : 177 (VT == MVT::f64) ? RTLIB::OGT_F64 : RTLIB::OGT_F128; 178 break; 179 case ISD::SETUEQ: 180 LC1 = (VT == MVT::f32) ? RTLIB::UO_F32 : 181 (VT == MVT::f64) ? RTLIB::UO_F64 : RTLIB::UO_F128; 182 LC2 = (VT == MVT::f32) ? RTLIB::OEQ_F32 : 183 (VT == MVT::f64) ? RTLIB::OEQ_F64 : RTLIB::OEQ_F128; 184 break; 185 default: 186 // Invert CC for unordered comparisons 187 ShouldInvertCC = true; 188 switch (CCCode) { 189 case ISD::SETULT: 190 LC1 = (VT == MVT::f32) ? RTLIB::OGE_F32 : 191 (VT == MVT::f64) ? RTLIB::OGE_F64 : RTLIB::OGE_F128; 192 break; 193 case ISD::SETULE: 194 LC1 = (VT == MVT::f32) ? RTLIB::OGT_F32 : 195 (VT == MVT::f64) ? RTLIB::OGT_F64 : RTLIB::OGT_F128; 196 break; 197 case ISD::SETUGT: 198 LC1 = (VT == MVT::f32) ? RTLIB::OLE_F32 : 199 (VT == MVT::f64) ? RTLIB::OLE_F64 : RTLIB::OLE_F128; 200 break; 201 case ISD::SETUGE: 202 LC1 = (VT == MVT::f32) ? RTLIB::OLT_F32 : 203 (VT == MVT::f64) ? RTLIB::OLT_F64 : RTLIB::OLT_F128; 204 break; 205 default: llvm_unreachable("Do not know how to soften this setcc!"); 206 } 207 } 208 209 // Use the target specific return value for comparions lib calls. 210 EVT RetVT = getCmpLibcallReturnType(); 211 SDValue Ops[2] = {NewLHS, NewRHS}; 212 NewLHS = makeLibCall(DAG, LC1, RetVT, Ops, false /*sign irrelevant*/, 213 dl).first; 214 NewRHS = DAG.getConstant(0, dl, RetVT); 215 216 CCCode = getCmpLibcallCC(LC1); 217 if (ShouldInvertCC) 218 CCCode = getSetCCInverse(CCCode, /*isInteger=*/true); 219 220 if (LC2 != RTLIB::UNKNOWN_LIBCALL) { 221 SDValue Tmp = DAG.getNode( 222 ISD::SETCC, dl, 223 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), RetVT), 224 NewLHS, NewRHS, DAG.getCondCode(CCCode)); 225 NewLHS = makeLibCall(DAG, LC2, RetVT, Ops, false/*sign irrelevant*/, 226 dl).first; 227 NewLHS = DAG.getNode( 228 ISD::SETCC, dl, 229 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), RetVT), 230 NewLHS, NewRHS, DAG.getCondCode(getCmpLibcallCC(LC2))); 231 NewLHS = DAG.getNode(ISD::OR, dl, Tmp.getValueType(), Tmp, NewLHS); 232 NewRHS = SDValue(); 233 } 234} 235 236/// getJumpTableEncoding - Return the entry encoding for a jump table in the 237/// current function. The returned value is a member of the 238/// MachineJumpTableInfo::JTEntryKind enum. 239unsigned TargetLowering::getJumpTableEncoding() const { 240 // In non-pic modes, just use the address of a block. 241 if (getTargetMachine().getRelocationModel() != Reloc::PIC_) 242 return MachineJumpTableInfo::EK_BlockAddress; 243 244 // In PIC mode, if the target supports a GPRel32 directive, use it. 245 if (getTargetMachine().getMCAsmInfo()->getGPRel32Directive() != nullptr) 246 return MachineJumpTableInfo::EK_GPRel32BlockAddress; 247 248 // Otherwise, use a label difference. 249 return MachineJumpTableInfo::EK_LabelDifference32; 250} 251 252SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table, 253 SelectionDAG &DAG) const { 254 // If our PIC model is GP relative, use the global offset table as the base. 255 unsigned JTEncoding = getJumpTableEncoding(); 256 257 if ((JTEncoding == MachineJumpTableInfo::EK_GPRel64BlockAddress) || 258 (JTEncoding == MachineJumpTableInfo::EK_GPRel32BlockAddress)) 259 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy(DAG.getDataLayout())); 260 261 return Table; 262} 263 264/// getPICJumpTableRelocBaseExpr - This returns the relocation base for the 265/// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an 266/// MCExpr. 267const MCExpr * 268TargetLowering::getPICJumpTableRelocBaseExpr(const MachineFunction *MF, 269 unsigned JTI,MCContext &Ctx) const{ 270 // The normal PIC reloc base is the label at the start of the jump table. 271 return MCSymbolRefExpr::create(MF->getJTISymbol(JTI, Ctx), Ctx); 272} 273 274bool 275TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { 276 // Assume that everything is safe in static mode. 277 if (getTargetMachine().getRelocationModel() == Reloc::Static) 278 return true; 279 280 // In dynamic-no-pic mode, assume that known defined values are safe. 281 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC && 282 GA && GA->getGlobal()->isStrongDefinitionForLinker()) 283 return true; 284 285 // Otherwise assume nothing is safe. 286 return false; 287} 288 289//===----------------------------------------------------------------------===// 290// Optimization Methods 291//===----------------------------------------------------------------------===// 292 293/// ShrinkDemandedConstant - Check to see if the specified operand of the 294/// specified instruction is a constant integer. If so, check to see if there 295/// are any bits set in the constant that are not demanded. If so, shrink the 296/// constant and return true. 297bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op, 298 const APInt &Demanded) { 299 SDLoc dl(Op); 300 301 // FIXME: ISD::SELECT, ISD::SELECT_CC 302 switch (Op.getOpcode()) { 303 default: break; 304 case ISD::XOR: 305 case ISD::AND: 306 case ISD::OR: { 307 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)); 308 if (!C) return false; 309 310 if (Op.getOpcode() == ISD::XOR && 311 (C->getAPIntValue() | (~Demanded)).isAllOnesValue()) 312 return false; 313 314 // if we can expand it to have all bits set, do it 315 if (C->getAPIntValue().intersects(~Demanded)) { 316 EVT VT = Op.getValueType(); 317 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0), 318 DAG.getConstant(Demanded & 319 C->getAPIntValue(), 320 dl, VT)); 321 return CombineTo(Op, New); 322 } 323 324 break; 325 } 326 } 327 328 return false; 329} 330 331/// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the 332/// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening 333/// cast, but it could be generalized for targets with other types of 334/// implicit widening casts. 335bool 336TargetLowering::TargetLoweringOpt::ShrinkDemandedOp(SDValue Op, 337 unsigned BitWidth, 338 const APInt &Demanded, 339 SDLoc dl) { 340 assert(Op.getNumOperands() == 2 && 341 "ShrinkDemandedOp only supports binary operators!"); 342 assert(Op.getNode()->getNumValues() == 1 && 343 "ShrinkDemandedOp only supports nodes with one result!"); 344 345 // Early return, as this function cannot handle vector types. 346 if (Op.getValueType().isVector()) 347 return false; 348 349 // Don't do this if the node has another user, which may require the 350 // full value. 351 if (!Op.getNode()->hasOneUse()) 352 return false; 353 354 // Search for the smallest integer type with free casts to and from 355 // Op's type. For expedience, just check power-of-2 integer types. 356 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 357 unsigned DemandedSize = BitWidth - Demanded.countLeadingZeros(); 358 unsigned SmallVTBits = DemandedSize; 359 if (!isPowerOf2_32(SmallVTBits)) 360 SmallVTBits = NextPowerOf2(SmallVTBits); 361 for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) { 362 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits); 363 if (TLI.isTruncateFree(Op.getValueType(), SmallVT) && 364 TLI.isZExtFree(SmallVT, Op.getValueType())) { 365 // We found a type with free casts. 366 SDValue X = DAG.getNode(Op.getOpcode(), dl, SmallVT, 367 DAG.getNode(ISD::TRUNCATE, dl, SmallVT, 368 Op.getNode()->getOperand(0)), 369 DAG.getNode(ISD::TRUNCATE, dl, SmallVT, 370 Op.getNode()->getOperand(1))); 371 bool NeedZext = DemandedSize > SmallVTBits; 372 SDValue Z = DAG.getNode(NeedZext ? ISD::ZERO_EXTEND : ISD::ANY_EXTEND, 373 dl, Op.getValueType(), X); 374 return CombineTo(Op, Z); 375 } 376 } 377 return false; 378} 379 380/// SimplifyDemandedBits - Look at Op. At this point, we know that only the 381/// DemandedMask bits of the result of Op are ever used downstream. If we can 382/// use this information to simplify Op, create a new simplified DAG node and 383/// return true, returning the original and new nodes in Old and New. Otherwise, 384/// analyze the expression and return a mask of KnownOne and KnownZero bits for 385/// the expression (used to simplify the caller). The KnownZero/One bits may 386/// only be accurate for those bits in the DemandedMask. 387bool TargetLowering::SimplifyDemandedBits(SDValue Op, 388 const APInt &DemandedMask, 389 APInt &KnownZero, 390 APInt &KnownOne, 391 TargetLoweringOpt &TLO, 392 unsigned Depth) const { 393 unsigned BitWidth = DemandedMask.getBitWidth(); 394 assert(Op.getValueType().getScalarType().getSizeInBits() == BitWidth && 395 "Mask size mismatches value type size!"); 396 APInt NewMask = DemandedMask; 397 SDLoc dl(Op); 398 auto &DL = TLO.DAG.getDataLayout(); 399 400 // Don't know anything. 401 KnownZero = KnownOne = APInt(BitWidth, 0); 402 403 // Other users may use these bits. 404 if (!Op.getNode()->hasOneUse()) { 405 if (Depth != 0) { 406 // If not at the root, Just compute the KnownZero/KnownOne bits to 407 // simplify things downstream. 408 TLO.DAG.computeKnownBits(Op, KnownZero, KnownOne, Depth); 409 return false; 410 } 411 // If this is the root being simplified, allow it to have multiple uses, 412 // just set the NewMask to all bits. 413 NewMask = APInt::getAllOnesValue(BitWidth); 414 } else if (DemandedMask == 0) { 415 // Not demanding any bits from Op. 416 if (Op.getOpcode() != ISD::UNDEF) 417 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType())); 418 return false; 419 } else if (Depth == 6) { // Limit search depth. 420 return false; 421 } 422 423 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut; 424 switch (Op.getOpcode()) { 425 case ISD::Constant: 426 // We know all of the bits for a constant! 427 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue(); 428 KnownZero = ~KnownOne; 429 return false; // Don't fall through, will infinitely loop. 430 case ISD::AND: 431 // If the RHS is a constant, check to see if the LHS would be zero without 432 // using the bits from the RHS. Below, we use knowledge about the RHS to 433 // simplify the LHS, here we're using information from the LHS to simplify 434 // the RHS. 435 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 436 APInt LHSZero, LHSOne; 437 // Do not increment Depth here; that can cause an infinite loop. 438 TLO.DAG.computeKnownBits(Op.getOperand(0), LHSZero, LHSOne, Depth); 439 // If the LHS already has zeros where RHSC does, this and is dead. 440 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask)) 441 return TLO.CombineTo(Op, Op.getOperand(0)); 442 // If any of the set bits in the RHS are known zero on the LHS, shrink 443 // the constant. 444 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask)) 445 return true; 446 } 447 448 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 449 KnownOne, TLO, Depth+1)) 450 return true; 451 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 452 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask, 453 KnownZero2, KnownOne2, TLO, Depth+1)) 454 return true; 455 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 456 457 // If all of the demanded bits are known one on one side, return the other. 458 // These bits cannot contribute to the result of the 'and'. 459 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask)) 460 return TLO.CombineTo(Op, Op.getOperand(0)); 461 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask)) 462 return TLO.CombineTo(Op, Op.getOperand(1)); 463 // If all of the demanded bits in the inputs are known zeros, return zero. 464 if ((NewMask & (KnownZero|KnownZero2)) == NewMask) 465 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, dl, Op.getValueType())); 466 // If the RHS is a constant, see if we can simplify it. 467 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask)) 468 return true; 469 // If the operation can be done in a smaller type, do so. 470 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 471 return true; 472 473 // Output known-1 bits are only known if set in both the LHS & RHS. 474 KnownOne &= KnownOne2; 475 // Output known-0 are known to be clear if zero in either the LHS | RHS. 476 KnownZero |= KnownZero2; 477 break; 478 case ISD::OR: 479 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 480 KnownOne, TLO, Depth+1)) 481 return true; 482 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 483 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask, 484 KnownZero2, KnownOne2, TLO, Depth+1)) 485 return true; 486 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 487 488 // If all of the demanded bits are known zero on one side, return the other. 489 // These bits cannot contribute to the result of the 'or'. 490 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask)) 491 return TLO.CombineTo(Op, Op.getOperand(0)); 492 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask)) 493 return TLO.CombineTo(Op, Op.getOperand(1)); 494 // If all of the potentially set bits on one side are known to be set on 495 // the other side, just use the 'other' side. 496 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask)) 497 return TLO.CombineTo(Op, Op.getOperand(0)); 498 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask)) 499 return TLO.CombineTo(Op, Op.getOperand(1)); 500 // If the RHS is a constant, see if we can simplify it. 501 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 502 return true; 503 // If the operation can be done in a smaller type, do so. 504 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 505 return true; 506 507 // Output known-0 bits are only known if clear in both the LHS & RHS. 508 KnownZero &= KnownZero2; 509 // Output known-1 are known to be set if set in either the LHS | RHS. 510 KnownOne |= KnownOne2; 511 break; 512 case ISD::XOR: 513 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 514 KnownOne, TLO, Depth+1)) 515 return true; 516 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 517 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2, 518 KnownOne2, TLO, Depth+1)) 519 return true; 520 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 521 522 // If all of the demanded bits are known zero on one side, return the other. 523 // These bits cannot contribute to the result of the 'xor'. 524 if ((KnownZero & NewMask) == NewMask) 525 return TLO.CombineTo(Op, Op.getOperand(0)); 526 if ((KnownZero2 & NewMask) == NewMask) 527 return TLO.CombineTo(Op, Op.getOperand(1)); 528 // If the operation can be done in a smaller type, do so. 529 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 530 return true; 531 532 // If all of the unknown bits are known to be zero on one side or the other 533 // (but not both) turn this into an *inclusive* or. 534 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0 535 if ((NewMask & ~KnownZero & ~KnownZero2) == 0) 536 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(), 537 Op.getOperand(0), 538 Op.getOperand(1))); 539 540 // Output known-0 bits are known if clear or set in both the LHS & RHS. 541 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2); 542 // Output known-1 are known to be set if set in only one of the LHS, RHS. 543 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2); 544 545 // If all of the demanded bits on one side are known, and all of the set 546 // bits on that side are also known to be set on the other side, turn this 547 // into an AND, as we know the bits will be cleared. 548 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2 549 // NB: it is okay if more bits are known than are requested 550 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known on one side 551 if (KnownOne == KnownOne2) { // set bits are the same on both sides 552 EVT VT = Op.getValueType(); 553 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, dl, VT); 554 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT, 555 Op.getOperand(0), ANDC)); 556 } 557 } 558 559 // If the RHS is a constant, see if we can simplify it. 560 // for XOR, we prefer to force bits to 1 if they will make a -1. 561 // if we can't force bits, try to shrink constant 562 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 563 APInt Expanded = C->getAPIntValue() | (~NewMask); 564 // if we can expand it to have all bits set, do it 565 if (Expanded.isAllOnesValue()) { 566 if (Expanded != C->getAPIntValue()) { 567 EVT VT = Op.getValueType(); 568 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0), 569 TLO.DAG.getConstant(Expanded, dl, VT)); 570 return TLO.CombineTo(Op, New); 571 } 572 // if it already has all the bits set, nothing to change 573 // but don't shrink either! 574 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) { 575 return true; 576 } 577 } 578 579 KnownZero = KnownZeroOut; 580 KnownOne = KnownOneOut; 581 break; 582 case ISD::SELECT: 583 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero, 584 KnownOne, TLO, Depth+1)) 585 return true; 586 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2, 587 KnownOne2, TLO, Depth+1)) 588 return true; 589 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 590 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 591 592 // If the operands are constants, see if we can simplify them. 593 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 594 return true; 595 596 // Only known if known in both the LHS and RHS. 597 KnownOne &= KnownOne2; 598 KnownZero &= KnownZero2; 599 break; 600 case ISD::SELECT_CC: 601 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero, 602 KnownOne, TLO, Depth+1)) 603 return true; 604 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2, 605 KnownOne2, TLO, Depth+1)) 606 return true; 607 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 608 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 609 610 // If the operands are constants, see if we can simplify them. 611 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 612 return true; 613 614 // Only known if known in both the LHS and RHS. 615 KnownOne &= KnownOne2; 616 KnownZero &= KnownZero2; 617 break; 618 case ISD::SHL: 619 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 620 unsigned ShAmt = SA->getZExtValue(); 621 SDValue InOp = Op.getOperand(0); 622 623 // If the shift count is an invalid immediate, don't do anything. 624 if (ShAmt >= BitWidth) 625 break; 626 627 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a 628 // single shift. We can do this if the bottom bits (which are shifted 629 // out) are never demanded. 630 if (InOp.getOpcode() == ISD::SRL && 631 isa<ConstantSDNode>(InOp.getOperand(1))) { 632 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) { 633 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue(); 634 unsigned Opc = ISD::SHL; 635 int Diff = ShAmt-C1; 636 if (Diff < 0) { 637 Diff = -Diff; 638 Opc = ISD::SRL; 639 } 640 641 SDValue NewSA = 642 TLO.DAG.getConstant(Diff, dl, Op.getOperand(1).getValueType()); 643 EVT VT = Op.getValueType(); 644 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, 645 InOp.getOperand(0), NewSA)); 646 } 647 } 648 649 if (SimplifyDemandedBits(InOp, NewMask.lshr(ShAmt), 650 KnownZero, KnownOne, TLO, Depth+1)) 651 return true; 652 653 // Convert (shl (anyext x, c)) to (anyext (shl x, c)) if the high bits 654 // are not demanded. This will likely allow the anyext to be folded away. 655 if (InOp.getNode()->getOpcode() == ISD::ANY_EXTEND) { 656 SDValue InnerOp = InOp.getNode()->getOperand(0); 657 EVT InnerVT = InnerOp.getValueType(); 658 unsigned InnerBits = InnerVT.getSizeInBits(); 659 if (ShAmt < InnerBits && NewMask.lshr(InnerBits) == 0 && 660 isTypeDesirableForOp(ISD::SHL, InnerVT)) { 661 EVT ShTy = getShiftAmountTy(InnerVT, DL); 662 if (!APInt(BitWidth, ShAmt).isIntN(ShTy.getSizeInBits())) 663 ShTy = InnerVT; 664 SDValue NarrowShl = 665 TLO.DAG.getNode(ISD::SHL, dl, InnerVT, InnerOp, 666 TLO.DAG.getConstant(ShAmt, dl, ShTy)); 667 return 668 TLO.CombineTo(Op, 669 TLO.DAG.getNode(ISD::ANY_EXTEND, dl, Op.getValueType(), 670 NarrowShl)); 671 } 672 // Repeat the SHL optimization above in cases where an extension 673 // intervenes: (shl (anyext (shr x, c1)), c2) to 674 // (shl (anyext x), c2-c1). This requires that the bottom c1 bits 675 // aren't demanded (as above) and that the shifted upper c1 bits of 676 // x aren't demanded. 677 if (InOp.hasOneUse() && 678 InnerOp.getOpcode() == ISD::SRL && 679 InnerOp.hasOneUse() && 680 isa<ConstantSDNode>(InnerOp.getOperand(1))) { 681 uint64_t InnerShAmt = cast<ConstantSDNode>(InnerOp.getOperand(1)) 682 ->getZExtValue(); 683 if (InnerShAmt < ShAmt && 684 InnerShAmt < InnerBits && 685 NewMask.lshr(InnerBits - InnerShAmt + ShAmt) == 0 && 686 NewMask.trunc(ShAmt) == 0) { 687 SDValue NewSA = 688 TLO.DAG.getConstant(ShAmt - InnerShAmt, dl, 689 Op.getOperand(1).getValueType()); 690 EVT VT = Op.getValueType(); 691 SDValue NewExt = TLO.DAG.getNode(ISD::ANY_EXTEND, dl, VT, 692 InnerOp.getOperand(0)); 693 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl, VT, 694 NewExt, NewSA)); 695 } 696 } 697 } 698 699 KnownZero <<= SA->getZExtValue(); 700 KnownOne <<= SA->getZExtValue(); 701 // low bits known zero. 702 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue()); 703 } 704 break; 705 case ISD::SRL: 706 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 707 EVT VT = Op.getValueType(); 708 unsigned ShAmt = SA->getZExtValue(); 709 unsigned VTSize = VT.getSizeInBits(); 710 SDValue InOp = Op.getOperand(0); 711 712 // If the shift count is an invalid immediate, don't do anything. 713 if (ShAmt >= BitWidth) 714 break; 715 716 APInt InDemandedMask = (NewMask << ShAmt); 717 718 // If the shift is exact, then it does demand the low bits (and knows that 719 // they are zero). 720 if (cast<BinaryWithFlagsSDNode>(Op)->Flags.hasExact()) 721 InDemandedMask |= APInt::getLowBitsSet(BitWidth, ShAmt); 722 723 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a 724 // single shift. We can do this if the top bits (which are shifted out) 725 // are never demanded. 726 if (InOp.getOpcode() == ISD::SHL && 727 isa<ConstantSDNode>(InOp.getOperand(1))) { 728 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) { 729 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue(); 730 unsigned Opc = ISD::SRL; 731 int Diff = ShAmt-C1; 732 if (Diff < 0) { 733 Diff = -Diff; 734 Opc = ISD::SHL; 735 } 736 737 SDValue NewSA = 738 TLO.DAG.getConstant(Diff, dl, Op.getOperand(1).getValueType()); 739 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, 740 InOp.getOperand(0), NewSA)); 741 } 742 } 743 744 // Compute the new bits that are at the top now. 745 if (SimplifyDemandedBits(InOp, InDemandedMask, 746 KnownZero, KnownOne, TLO, Depth+1)) 747 return true; 748 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 749 KnownZero = KnownZero.lshr(ShAmt); 750 KnownOne = KnownOne.lshr(ShAmt); 751 752 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt); 753 KnownZero |= HighBits; // High bits known zero. 754 } 755 break; 756 case ISD::SRA: 757 // If this is an arithmetic shift right and only the low-bit is set, we can 758 // always convert this into a logical shr, even if the shift amount is 759 // variable. The low bit of the shift cannot be an input sign bit unless 760 // the shift amount is >= the size of the datatype, which is undefined. 761 if (NewMask == 1) 762 return TLO.CombineTo(Op, 763 TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(), 764 Op.getOperand(0), Op.getOperand(1))); 765 766 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 767 EVT VT = Op.getValueType(); 768 unsigned ShAmt = SA->getZExtValue(); 769 770 // If the shift count is an invalid immediate, don't do anything. 771 if (ShAmt >= BitWidth) 772 break; 773 774 APInt InDemandedMask = (NewMask << ShAmt); 775 776 // If the shift is exact, then it does demand the low bits (and knows that 777 // they are zero). 778 if (cast<BinaryWithFlagsSDNode>(Op)->Flags.hasExact()) 779 InDemandedMask |= APInt::getLowBitsSet(BitWidth, ShAmt); 780 781 // If any of the demanded bits are produced by the sign extension, we also 782 // demand the input sign bit. 783 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt); 784 if (HighBits.intersects(NewMask)) 785 InDemandedMask |= APInt::getSignBit(VT.getScalarType().getSizeInBits()); 786 787 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask, 788 KnownZero, KnownOne, TLO, Depth+1)) 789 return true; 790 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 791 KnownZero = KnownZero.lshr(ShAmt); 792 KnownOne = KnownOne.lshr(ShAmt); 793 794 // Handle the sign bit, adjusted to where it is now in the mask. 795 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt); 796 797 // If the input sign bit is known to be zero, or if none of the top bits 798 // are demanded, turn this into an unsigned shift right. 799 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) { 800 SDNodeFlags Flags; 801 Flags.setExact(cast<BinaryWithFlagsSDNode>(Op)->Flags.hasExact()); 802 return TLO.CombineTo(Op, 803 TLO.DAG.getNode(ISD::SRL, dl, VT, Op.getOperand(0), 804 Op.getOperand(1), &Flags)); 805 } 806 807 int Log2 = NewMask.exactLogBase2(); 808 if (Log2 >= 0) { 809 // The bit must come from the sign. 810 SDValue NewSA = 811 TLO.DAG.getConstant(BitWidth - 1 - Log2, dl, 812 Op.getOperand(1).getValueType()); 813 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT, 814 Op.getOperand(0), NewSA)); 815 } 816 817 if (KnownOne.intersects(SignBit)) 818 // New bits are known one. 819 KnownOne |= HighBits; 820 } 821 break; 822 case ISD::SIGN_EXTEND_INREG: { 823 EVT ExVT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 824 825 APInt MsbMask = APInt::getHighBitsSet(BitWidth, 1); 826 // If we only care about the highest bit, don't bother shifting right. 827 if (MsbMask == NewMask) { 828 unsigned ShAmt = ExVT.getScalarType().getSizeInBits(); 829 SDValue InOp = Op.getOperand(0); 830 unsigned VTBits = Op->getValueType(0).getScalarType().getSizeInBits(); 831 bool AlreadySignExtended = 832 TLO.DAG.ComputeNumSignBits(InOp) >= VTBits-ShAmt+1; 833 // However if the input is already sign extended we expect the sign 834 // extension to be dropped altogether later and do not simplify. 835 if (!AlreadySignExtended) { 836 // Compute the correct shift amount type, which must be getShiftAmountTy 837 // for scalar types after legalization. 838 EVT ShiftAmtTy = Op.getValueType(); 839 if (TLO.LegalTypes() && !ShiftAmtTy.isVector()) 840 ShiftAmtTy = getShiftAmountTy(ShiftAmtTy, DL); 841 842 SDValue ShiftAmt = TLO.DAG.getConstant(BitWidth - ShAmt, dl, 843 ShiftAmtTy); 844 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl, 845 Op.getValueType(), InOp, 846 ShiftAmt)); 847 } 848 } 849 850 // Sign extension. Compute the demanded bits in the result that are not 851 // present in the input. 852 APInt NewBits = 853 APInt::getHighBitsSet(BitWidth, 854 BitWidth - ExVT.getScalarType().getSizeInBits()); 855 856 // If none of the extended bits are demanded, eliminate the sextinreg. 857 if ((NewBits & NewMask) == 0) 858 return TLO.CombineTo(Op, Op.getOperand(0)); 859 860 APInt InSignBit = 861 APInt::getSignBit(ExVT.getScalarType().getSizeInBits()).zext(BitWidth); 862 APInt InputDemandedBits = 863 APInt::getLowBitsSet(BitWidth, 864 ExVT.getScalarType().getSizeInBits()) & 865 NewMask; 866 867 // Since the sign extended bits are demanded, we know that the sign 868 // bit is demanded. 869 InputDemandedBits |= InSignBit; 870 871 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits, 872 KnownZero, KnownOne, TLO, Depth+1)) 873 return true; 874 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 875 876 // If the sign bit of the input is known set or clear, then we know the 877 // top bits of the result. 878 879 // If the input sign bit is known zero, convert this into a zero extension. 880 if (KnownZero.intersects(InSignBit)) 881 return TLO.CombineTo(Op, 882 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,ExVT)); 883 884 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set 885 KnownOne |= NewBits; 886 KnownZero &= ~NewBits; 887 } else { // Input sign bit unknown 888 KnownZero &= ~NewBits; 889 KnownOne &= ~NewBits; 890 } 891 break; 892 } 893 case ISD::BUILD_PAIR: { 894 EVT HalfVT = Op.getOperand(0).getValueType(); 895 unsigned HalfBitWidth = HalfVT.getScalarSizeInBits(); 896 897 APInt MaskLo = NewMask.getLoBits(HalfBitWidth).trunc(HalfBitWidth); 898 APInt MaskHi = NewMask.getHiBits(HalfBitWidth).trunc(HalfBitWidth); 899 900 APInt KnownZeroLo, KnownOneLo; 901 APInt KnownZeroHi, KnownOneHi; 902 903 if (SimplifyDemandedBits(Op.getOperand(0), MaskLo, KnownZeroLo, 904 KnownOneLo, TLO, Depth + 1)) 905 return true; 906 907 if (SimplifyDemandedBits(Op.getOperand(1), MaskHi, KnownZeroHi, 908 KnownOneHi, TLO, Depth + 1)) 909 return true; 910 911 KnownZero = KnownZeroLo.zext(BitWidth) | 912 KnownZeroHi.zext(BitWidth).shl(HalfBitWidth); 913 914 KnownOne = KnownOneLo.zext(BitWidth) | 915 KnownOneHi.zext(BitWidth).shl(HalfBitWidth); 916 break; 917 } 918 case ISD::ZERO_EXTEND: { 919 unsigned OperandBitWidth = 920 Op.getOperand(0).getValueType().getScalarType().getSizeInBits(); 921 APInt InMask = NewMask.trunc(OperandBitWidth); 922 923 // If none of the top bits are demanded, convert this into an any_extend. 924 APInt NewBits = 925 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask; 926 if (!NewBits.intersects(NewMask)) 927 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl, 928 Op.getValueType(), 929 Op.getOperand(0))); 930 931 if (SimplifyDemandedBits(Op.getOperand(0), InMask, 932 KnownZero, KnownOne, TLO, Depth+1)) 933 return true; 934 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 935 KnownZero = KnownZero.zext(BitWidth); 936 KnownOne = KnownOne.zext(BitWidth); 937 KnownZero |= NewBits; 938 break; 939 } 940 case ISD::SIGN_EXTEND: { 941 EVT InVT = Op.getOperand(0).getValueType(); 942 unsigned InBits = InVT.getScalarType().getSizeInBits(); 943 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits); 944 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits); 945 APInt NewBits = ~InMask & NewMask; 946 947 // If none of the top bits are demanded, convert this into an any_extend. 948 if (NewBits == 0) 949 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl, 950 Op.getValueType(), 951 Op.getOperand(0))); 952 953 // Since some of the sign extended bits are demanded, we know that the sign 954 // bit is demanded. 955 APInt InDemandedBits = InMask & NewMask; 956 InDemandedBits |= InSignBit; 957 InDemandedBits = InDemandedBits.trunc(InBits); 958 959 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero, 960 KnownOne, TLO, Depth+1)) 961 return true; 962 KnownZero = KnownZero.zext(BitWidth); 963 KnownOne = KnownOne.zext(BitWidth); 964 965 // If the sign bit is known zero, convert this to a zero extend. 966 if (KnownZero.intersects(InSignBit)) 967 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, 968 Op.getValueType(), 969 Op.getOperand(0))); 970 971 // If the sign bit is known one, the top bits match. 972 if (KnownOne.intersects(InSignBit)) { 973 KnownOne |= NewBits; 974 assert((KnownZero & NewBits) == 0); 975 } else { // Otherwise, top bits aren't known. 976 assert((KnownOne & NewBits) == 0); 977 assert((KnownZero & NewBits) == 0); 978 } 979 break; 980 } 981 case ISD::ANY_EXTEND: { 982 unsigned OperandBitWidth = 983 Op.getOperand(0).getValueType().getScalarType().getSizeInBits(); 984 APInt InMask = NewMask.trunc(OperandBitWidth); 985 if (SimplifyDemandedBits(Op.getOperand(0), InMask, 986 KnownZero, KnownOne, TLO, Depth+1)) 987 return true; 988 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 989 KnownZero = KnownZero.zext(BitWidth); 990 KnownOne = KnownOne.zext(BitWidth); 991 break; 992 } 993 case ISD::TRUNCATE: { 994 // Simplify the input, using demanded bit information, and compute the known 995 // zero/one bits live out. 996 unsigned OperandBitWidth = 997 Op.getOperand(0).getValueType().getScalarType().getSizeInBits(); 998 APInt TruncMask = NewMask.zext(OperandBitWidth); 999 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask, 1000 KnownZero, KnownOne, TLO, Depth+1)) 1001 return true; 1002 KnownZero = KnownZero.trunc(BitWidth); 1003 KnownOne = KnownOne.trunc(BitWidth); 1004 1005 // If the input is only used by this truncate, see if we can shrink it based 1006 // on the known demanded bits. 1007 if (Op.getOperand(0).getNode()->hasOneUse()) { 1008 SDValue In = Op.getOperand(0); 1009 switch (In.getOpcode()) { 1010 default: break; 1011 case ISD::SRL: 1012 // Shrink SRL by a constant if none of the high bits shifted in are 1013 // demanded. 1014 if (TLO.LegalTypes() && 1015 !isTypeDesirableForOp(ISD::SRL, Op.getValueType())) 1016 // Do not turn (vt1 truncate (vt2 srl)) into (vt1 srl) if vt1 is 1017 // undesirable. 1018 break; 1019 ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1)); 1020 if (!ShAmt) 1021 break; 1022 SDValue Shift = In.getOperand(1); 1023 if (TLO.LegalTypes()) { 1024 uint64_t ShVal = ShAmt->getZExtValue(); 1025 Shift = TLO.DAG.getConstant(ShVal, dl, 1026 getShiftAmountTy(Op.getValueType(), DL)); 1027 } 1028 1029 APInt HighBits = APInt::getHighBitsSet(OperandBitWidth, 1030 OperandBitWidth - BitWidth); 1031 HighBits = HighBits.lshr(ShAmt->getZExtValue()).trunc(BitWidth); 1032 1033 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) { 1034 // None of the shifted in bits are needed. Add a truncate of the 1035 // shift input, then shift it. 1036 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl, 1037 Op.getValueType(), 1038 In.getOperand(0)); 1039 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, 1040 Op.getValueType(), 1041 NewTrunc, 1042 Shift)); 1043 } 1044 break; 1045 } 1046 } 1047 1048 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1049 break; 1050 } 1051 case ISD::AssertZext: { 1052 // AssertZext demands all of the high bits, plus any of the low bits 1053 // demanded by its users. 1054 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 1055 APInt InMask = APInt::getLowBitsSet(BitWidth, 1056 VT.getSizeInBits()); 1057 if (SimplifyDemandedBits(Op.getOperand(0), ~InMask | NewMask, 1058 KnownZero, KnownOne, TLO, Depth+1)) 1059 return true; 1060 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1061 1062 KnownZero |= ~InMask & NewMask; 1063 break; 1064 } 1065 case ISD::BITCAST: 1066 // If this is an FP->Int bitcast and if the sign bit is the only 1067 // thing demanded, turn this into a FGETSIGN. 1068 if (!TLO.LegalOperations() && 1069 !Op.getValueType().isVector() && 1070 !Op.getOperand(0).getValueType().isVector() && 1071 NewMask == APInt::getSignBit(Op.getValueType().getSizeInBits()) && 1072 Op.getOperand(0).getValueType().isFloatingPoint()) { 1073 bool OpVTLegal = isOperationLegalOrCustom(ISD::FGETSIGN, Op.getValueType()); 1074 bool i32Legal = isOperationLegalOrCustom(ISD::FGETSIGN, MVT::i32); 1075 if ((OpVTLegal || i32Legal) && Op.getValueType().isSimple() && 1076 Op.getOperand(0).getValueType() != MVT::f128) { 1077 // Cannot eliminate/lower SHL for f128 yet. 1078 EVT Ty = OpVTLegal ? Op.getValueType() : MVT::i32; 1079 // Make a FGETSIGN + SHL to move the sign bit into the appropriate 1080 // place. We expect the SHL to be eliminated by other optimizations. 1081 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, dl, Ty, Op.getOperand(0)); 1082 unsigned OpVTSizeInBits = Op.getValueType().getSizeInBits(); 1083 if (!OpVTLegal && OpVTSizeInBits > 32) 1084 Sign = TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), Sign); 1085 unsigned ShVal = Op.getValueType().getSizeInBits()-1; 1086 SDValue ShAmt = TLO.DAG.getConstant(ShVal, dl, Op.getValueType()); 1087 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl, 1088 Op.getValueType(), 1089 Sign, ShAmt)); 1090 } 1091 } 1092 break; 1093 case ISD::ADD: 1094 case ISD::MUL: 1095 case ISD::SUB: { 1096 // Add, Sub, and Mul don't demand any bits in positions beyond that 1097 // of the highest bit demanded of them. 1098 APInt LoMask = APInt::getLowBitsSet(BitWidth, 1099 BitWidth - NewMask.countLeadingZeros()); 1100 if (SimplifyDemandedBits(Op.getOperand(0), LoMask, KnownZero2, 1101 KnownOne2, TLO, Depth+1)) 1102 return true; 1103 if (SimplifyDemandedBits(Op.getOperand(1), LoMask, KnownZero2, 1104 KnownOne2, TLO, Depth+1)) 1105 return true; 1106 // See if the operation should be performed at a smaller bit width. 1107 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 1108 return true; 1109 } 1110 // FALL THROUGH 1111 default: 1112 // Just use computeKnownBits to compute output bits. 1113 TLO.DAG.computeKnownBits(Op, KnownZero, KnownOne, Depth); 1114 break; 1115 } 1116 1117 // If we know the value of all of the demanded bits, return this as a 1118 // constant. 1119 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { 1120 // Avoid folding to a constant if any OpaqueConstant is involved. 1121 const SDNode *N = Op.getNode(); 1122 for (SDNodeIterator I = SDNodeIterator::begin(N), 1123 E = SDNodeIterator::end(N); I != E; ++I) { 1124 SDNode *Op = *I; 1125 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) 1126 if (C->isOpaque()) 1127 return false; 1128 } 1129 return TLO.CombineTo(Op, 1130 TLO.DAG.getConstant(KnownOne, dl, Op.getValueType())); 1131 } 1132 1133 return false; 1134} 1135 1136/// computeKnownBitsForTargetNode - Determine which of the bits specified 1137/// in Mask are known to be either zero or one and return them in the 1138/// KnownZero/KnownOne bitsets. 1139void TargetLowering::computeKnownBitsForTargetNode(const SDValue Op, 1140 APInt &KnownZero, 1141 APInt &KnownOne, 1142 const SelectionDAG &DAG, 1143 unsigned Depth) const { 1144 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || 1145 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || 1146 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || 1147 Op.getOpcode() == ISD::INTRINSIC_VOID) && 1148 "Should use MaskedValueIsZero if you don't know whether Op" 1149 " is a target node!"); 1150 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); 1151} 1152 1153/// ComputeNumSignBitsForTargetNode - This method can be implemented by 1154/// targets that want to expose additional information about sign bits to the 1155/// DAG Combiner. 1156unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op, 1157 const SelectionDAG &, 1158 unsigned Depth) const { 1159 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || 1160 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || 1161 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || 1162 Op.getOpcode() == ISD::INTRINSIC_VOID) && 1163 "Should use ComputeNumSignBits if you don't know whether Op" 1164 " is a target node!"); 1165 return 1; 1166} 1167 1168/// ValueHasExactlyOneBitSet - Test if the given value is known to have exactly 1169/// one bit set. This differs from computeKnownBits in that it doesn't need to 1170/// determine which bit is set. 1171/// 1172static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) { 1173 // A left-shift of a constant one will have exactly one bit set, because 1174 // shifting the bit off the end is undefined. 1175 if (Val.getOpcode() == ISD::SHL) 1176 if (ConstantSDNode *C = 1177 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0))) 1178 if (C->getAPIntValue() == 1) 1179 return true; 1180 1181 // Similarly, a right-shift of a constant sign-bit will have exactly 1182 // one bit set. 1183 if (Val.getOpcode() == ISD::SRL) 1184 if (ConstantSDNode *C = 1185 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0))) 1186 if (C->getAPIntValue().isSignBit()) 1187 return true; 1188 1189 // More could be done here, though the above checks are enough 1190 // to handle some common cases. 1191 1192 // Fall back to computeKnownBits to catch other known cases. 1193 EVT OpVT = Val.getValueType(); 1194 unsigned BitWidth = OpVT.getScalarType().getSizeInBits(); 1195 APInt KnownZero, KnownOne; 1196 DAG.computeKnownBits(Val, KnownZero, KnownOne); 1197 return (KnownZero.countPopulation() == BitWidth - 1) && 1198 (KnownOne.countPopulation() == 1); 1199} 1200 1201bool TargetLowering::isConstTrueVal(const SDNode *N) const { 1202 if (!N) 1203 return false; 1204 1205 const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N); 1206 if (!CN) { 1207 const BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N); 1208 if (!BV) 1209 return false; 1210 1211 BitVector UndefElements; 1212 CN = BV->getConstantSplatNode(&UndefElements); 1213 // Only interested in constant splats, and we don't try to handle undef 1214 // elements in identifying boolean constants. 1215 if (!CN || UndefElements.none()) 1216 return false; 1217 } 1218 1219 switch (getBooleanContents(N->getValueType(0))) { 1220 case UndefinedBooleanContent: 1221 return CN->getAPIntValue()[0]; 1222 case ZeroOrOneBooleanContent: 1223 return CN->isOne(); 1224 case ZeroOrNegativeOneBooleanContent: 1225 return CN->isAllOnesValue(); 1226 } 1227 1228 llvm_unreachable("Invalid boolean contents"); 1229} 1230 1231bool TargetLowering::isConstFalseVal(const SDNode *N) const { 1232 if (!N) 1233 return false; 1234 1235 const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N); 1236 if (!CN) { 1237 const BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N); 1238 if (!BV) 1239 return false; 1240 1241 BitVector UndefElements; 1242 CN = BV->getConstantSplatNode(&UndefElements); 1243 // Only interested in constant splats, and we don't try to handle undef 1244 // elements in identifying boolean constants. 1245 if (!CN || UndefElements.none()) 1246 return false; 1247 } 1248 1249 if (getBooleanContents(N->getValueType(0)) == UndefinedBooleanContent) 1250 return !CN->getAPIntValue()[0]; 1251 1252 return CN->isNullValue(); 1253} 1254 1255/// SimplifySetCC - Try to simplify a setcc built with the specified operands 1256/// and cc. If it is unable to simplify it, return a null SDValue. 1257SDValue 1258TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1, 1259 ISD::CondCode Cond, bool foldBooleans, 1260 DAGCombinerInfo &DCI, SDLoc dl) const { 1261 SelectionDAG &DAG = DCI.DAG; 1262 1263 // These setcc operations always fold. 1264 switch (Cond) { 1265 default: break; 1266 case ISD::SETFALSE: 1267 case ISD::SETFALSE2: return DAG.getConstant(0, dl, VT); 1268 case ISD::SETTRUE: 1269 case ISD::SETTRUE2: { 1270 TargetLowering::BooleanContent Cnt = 1271 getBooleanContents(N0->getValueType(0)); 1272 return DAG.getConstant( 1273 Cnt == TargetLowering::ZeroOrNegativeOneBooleanContent ? -1ULL : 1, dl, 1274 VT); 1275 } 1276 } 1277 1278 // Ensure that the constant occurs on the RHS, and fold constant 1279 // comparisons. 1280 ISD::CondCode SwappedCC = ISD::getSetCCSwappedOperands(Cond); 1281 if (isa<ConstantSDNode>(N0.getNode()) && 1282 (DCI.isBeforeLegalizeOps() || 1283 isCondCodeLegal(SwappedCC, N0.getSimpleValueType()))) 1284 return DAG.getSetCC(dl, VT, N1, N0, SwappedCC); 1285 1286 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) { 1287 const APInt &C1 = N1C->getAPIntValue(); 1288 1289 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an 1290 // equality comparison, then we're just comparing whether X itself is 1291 // zero. 1292 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) && 1293 N0.getOperand(0).getOpcode() == ISD::CTLZ && 1294 N0.getOperand(1).getOpcode() == ISD::Constant) { 1295 const APInt &ShAmt 1296 = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); 1297 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 1298 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) { 1299 if ((C1 == 0) == (Cond == ISD::SETEQ)) { 1300 // (srl (ctlz x), 5) == 0 -> X != 0 1301 // (srl (ctlz x), 5) != 1 -> X != 0 1302 Cond = ISD::SETNE; 1303 } else { 1304 // (srl (ctlz x), 5) != 0 -> X == 0 1305 // (srl (ctlz x), 5) == 1 -> X == 0 1306 Cond = ISD::SETEQ; 1307 } 1308 SDValue Zero = DAG.getConstant(0, dl, N0.getValueType()); 1309 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0), 1310 Zero, Cond); 1311 } 1312 } 1313 1314 SDValue CTPOP = N0; 1315 // Look through truncs that don't change the value of a ctpop. 1316 if (N0.hasOneUse() && N0.getOpcode() == ISD::TRUNCATE) 1317 CTPOP = N0.getOperand(0); 1318 1319 if (CTPOP.hasOneUse() && CTPOP.getOpcode() == ISD::CTPOP && 1320 (N0 == CTPOP || N0.getValueType().getSizeInBits() > 1321 Log2_32_Ceil(CTPOP.getValueType().getSizeInBits()))) { 1322 EVT CTVT = CTPOP.getValueType(); 1323 SDValue CTOp = CTPOP.getOperand(0); 1324 1325 // (ctpop x) u< 2 -> (x & x-1) == 0 1326 // (ctpop x) u> 1 -> (x & x-1) != 0 1327 if ((Cond == ISD::SETULT && C1 == 2) || (Cond == ISD::SETUGT && C1 == 1)){ 1328 SDValue Sub = DAG.getNode(ISD::SUB, dl, CTVT, CTOp, 1329 DAG.getConstant(1, dl, CTVT)); 1330 SDValue And = DAG.getNode(ISD::AND, dl, CTVT, CTOp, Sub); 1331 ISD::CondCode CC = Cond == ISD::SETULT ? ISD::SETEQ : ISD::SETNE; 1332 return DAG.getSetCC(dl, VT, And, DAG.getConstant(0, dl, CTVT), CC); 1333 } 1334 1335 // TODO: (ctpop x) == 1 -> x && (x & x-1) == 0 iff ctpop is illegal. 1336 } 1337 1338 // (zext x) == C --> x == (trunc C) 1339 // (sext x) == C --> x == (trunc C) 1340 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 1341 DCI.isBeforeLegalize() && N0->hasOneUse()) { 1342 unsigned MinBits = N0.getValueSizeInBits(); 1343 SDValue PreExt; 1344 bool Signed = false; 1345 if (N0->getOpcode() == ISD::ZERO_EXTEND) { 1346 // ZExt 1347 MinBits = N0->getOperand(0).getValueSizeInBits(); 1348 PreExt = N0->getOperand(0); 1349 } else if (N0->getOpcode() == ISD::AND) { 1350 // DAGCombine turns costly ZExts into ANDs 1351 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0->getOperand(1))) 1352 if ((C->getAPIntValue()+1).isPowerOf2()) { 1353 MinBits = C->getAPIntValue().countTrailingOnes(); 1354 PreExt = N0->getOperand(0); 1355 } 1356 } else if (N0->getOpcode() == ISD::SIGN_EXTEND) { 1357 // SExt 1358 MinBits = N0->getOperand(0).getValueSizeInBits(); 1359 PreExt = N0->getOperand(0); 1360 Signed = true; 1361 } else if (LoadSDNode *LN0 = dyn_cast<LoadSDNode>(N0)) { 1362 // ZEXTLOAD / SEXTLOAD 1363 if (LN0->getExtensionType() == ISD::ZEXTLOAD) { 1364 MinBits = LN0->getMemoryVT().getSizeInBits(); 1365 PreExt = N0; 1366 } else if (LN0->getExtensionType() == ISD::SEXTLOAD) { 1367 Signed = true; 1368 MinBits = LN0->getMemoryVT().getSizeInBits(); 1369 PreExt = N0; 1370 } 1371 } 1372 1373 // Figure out how many bits we need to preserve this constant. 1374 unsigned ReqdBits = Signed ? 1375 C1.getBitWidth() - C1.getNumSignBits() + 1 : 1376 C1.getActiveBits(); 1377 1378 // Make sure we're not losing bits from the constant. 1379 if (MinBits > 0 && 1380 MinBits < C1.getBitWidth() && 1381 MinBits >= ReqdBits) { 1382 EVT MinVT = EVT::getIntegerVT(*DAG.getContext(), MinBits); 1383 if (isTypeDesirableForOp(ISD::SETCC, MinVT)) { 1384 // Will get folded away. 1385 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MinVT, PreExt); 1386 SDValue C = DAG.getConstant(C1.trunc(MinBits), dl, MinVT); 1387 return DAG.getSetCC(dl, VT, Trunc, C, Cond); 1388 } 1389 } 1390 } 1391 1392 // If the LHS is '(and load, const)', the RHS is 0, 1393 // the test is for equality or unsigned, and all 1 bits of the const are 1394 // in the same partial word, see if we can shorten the load. 1395 if (DCI.isBeforeLegalize() && 1396 !ISD::isSignedIntSetCC(Cond) && 1397 N0.getOpcode() == ISD::AND && C1 == 0 && 1398 N0.getNode()->hasOneUse() && 1399 isa<LoadSDNode>(N0.getOperand(0)) && 1400 N0.getOperand(0).getNode()->hasOneUse() && 1401 isa<ConstantSDNode>(N0.getOperand(1))) { 1402 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0)); 1403 APInt bestMask; 1404 unsigned bestWidth = 0, bestOffset = 0; 1405 if (!Lod->isVolatile() && Lod->isUnindexed()) { 1406 unsigned origWidth = N0.getValueType().getSizeInBits(); 1407 unsigned maskWidth = origWidth; 1408 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to 1409 // 8 bits, but have to be careful... 1410 if (Lod->getExtensionType() != ISD::NON_EXTLOAD) 1411 origWidth = Lod->getMemoryVT().getSizeInBits(); 1412 const APInt &Mask = 1413 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); 1414 for (unsigned width = origWidth / 2; width>=8; width /= 2) { 1415 APInt newMask = APInt::getLowBitsSet(maskWidth, width); 1416 for (unsigned offset=0; offset<origWidth/width; offset++) { 1417 if ((newMask & Mask) == Mask) { 1418 if (!DAG.getDataLayout().isLittleEndian()) 1419 bestOffset = (origWidth/width - offset - 1) * (width/8); 1420 else 1421 bestOffset = (uint64_t)offset * (width/8); 1422 bestMask = Mask.lshr(offset * (width/8) * 8); 1423 bestWidth = width; 1424 break; 1425 } 1426 newMask = newMask << width; 1427 } 1428 } 1429 } 1430 if (bestWidth) { 1431 EVT newVT = EVT::getIntegerVT(*DAG.getContext(), bestWidth); 1432 if (newVT.isRound()) { 1433 EVT PtrType = Lod->getOperand(1).getValueType(); 1434 SDValue Ptr = Lod->getBasePtr(); 1435 if (bestOffset != 0) 1436 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(), 1437 DAG.getConstant(bestOffset, dl, PtrType)); 1438 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset); 1439 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr, 1440 Lod->getPointerInfo().getWithOffset(bestOffset), 1441 false, false, false, NewAlign); 1442 return DAG.getSetCC(dl, VT, 1443 DAG.getNode(ISD::AND, dl, newVT, NewLoad, 1444 DAG.getConstant(bestMask.trunc(bestWidth), 1445 dl, newVT)), 1446 DAG.getConstant(0LL, dl, newVT), Cond); 1447 } 1448 } 1449 } 1450 1451 // If the LHS is a ZERO_EXTEND, perform the comparison on the input. 1452 if (N0.getOpcode() == ISD::ZERO_EXTEND) { 1453 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits(); 1454 1455 // If the comparison constant has bits in the upper part, the 1456 // zero-extended value could never match. 1457 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(), 1458 C1.getBitWidth() - InSize))) { 1459 switch (Cond) { 1460 case ISD::SETUGT: 1461 case ISD::SETUGE: 1462 case ISD::SETEQ: return DAG.getConstant(0, dl, VT); 1463 case ISD::SETULT: 1464 case ISD::SETULE: 1465 case ISD::SETNE: return DAG.getConstant(1, dl, VT); 1466 case ISD::SETGT: 1467 case ISD::SETGE: 1468 // True if the sign bit of C1 is set. 1469 return DAG.getConstant(C1.isNegative(), dl, VT); 1470 case ISD::SETLT: 1471 case ISD::SETLE: 1472 // True if the sign bit of C1 isn't set. 1473 return DAG.getConstant(C1.isNonNegative(), dl, VT); 1474 default: 1475 break; 1476 } 1477 } 1478 1479 // Otherwise, we can perform the comparison with the low bits. 1480 switch (Cond) { 1481 case ISD::SETEQ: 1482 case ISD::SETNE: 1483 case ISD::SETUGT: 1484 case ISD::SETUGE: 1485 case ISD::SETULT: 1486 case ISD::SETULE: { 1487 EVT newVT = N0.getOperand(0).getValueType(); 1488 if (DCI.isBeforeLegalizeOps() || 1489 (isOperationLegal(ISD::SETCC, newVT) && 1490 getCondCodeAction(Cond, newVT.getSimpleVT()) == Legal)) { 1491 EVT NewSetCCVT = 1492 getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), newVT); 1493 SDValue NewConst = DAG.getConstant(C1.trunc(InSize), dl, newVT); 1494 1495 SDValue NewSetCC = DAG.getSetCC(dl, NewSetCCVT, N0.getOperand(0), 1496 NewConst, Cond); 1497 return DAG.getBoolExtOrTrunc(NewSetCC, dl, VT, N0.getValueType()); 1498 } 1499 break; 1500 } 1501 default: 1502 break; // todo, be more careful with signed comparisons 1503 } 1504 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG && 1505 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 1506 EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT(); 1507 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits(); 1508 EVT ExtDstTy = N0.getValueType(); 1509 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits(); 1510 1511 // If the constant doesn't fit into the number of bits for the source of 1512 // the sign extension, it is impossible for both sides to be equal. 1513 if (C1.getMinSignedBits() > ExtSrcTyBits) 1514 return DAG.getConstant(Cond == ISD::SETNE, dl, VT); 1515 1516 SDValue ZextOp; 1517 EVT Op0Ty = N0.getOperand(0).getValueType(); 1518 if (Op0Ty == ExtSrcTy) { 1519 ZextOp = N0.getOperand(0); 1520 } else { 1521 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits); 1522 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0), 1523 DAG.getConstant(Imm, dl, Op0Ty)); 1524 } 1525 if (!DCI.isCalledByLegalizer()) 1526 DCI.AddToWorklist(ZextOp.getNode()); 1527 // Otherwise, make this a use of a zext. 1528 return DAG.getSetCC(dl, VT, ZextOp, 1529 DAG.getConstant(C1 & APInt::getLowBitsSet( 1530 ExtDstTyBits, 1531 ExtSrcTyBits), 1532 dl, ExtDstTy), 1533 Cond); 1534 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) && 1535 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 1536 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC 1537 if (N0.getOpcode() == ISD::SETCC && 1538 isTypeLegal(VT) && VT.bitsLE(N0.getValueType())) { 1539 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getAPIntValue() != 1); 1540 if (TrueWhenTrue) 1541 return DAG.getNode(ISD::TRUNCATE, dl, VT, N0); 1542 // Invert the condition. 1543 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get(); 1544 CC = ISD::getSetCCInverse(CC, 1545 N0.getOperand(0).getValueType().isInteger()); 1546 if (DCI.isBeforeLegalizeOps() || 1547 isCondCodeLegal(CC, N0.getOperand(0).getSimpleValueType())) 1548 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC); 1549 } 1550 1551 if ((N0.getOpcode() == ISD::XOR || 1552 (N0.getOpcode() == ISD::AND && 1553 N0.getOperand(0).getOpcode() == ISD::XOR && 1554 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) && 1555 isa<ConstantSDNode>(N0.getOperand(1)) && 1556 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) { 1557 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We 1558 // can only do this if the top bits are known zero. 1559 unsigned BitWidth = N0.getValueSizeInBits(); 1560 if (DAG.MaskedValueIsZero(N0, 1561 APInt::getHighBitsSet(BitWidth, 1562 BitWidth-1))) { 1563 // Okay, get the un-inverted input value. 1564 SDValue Val; 1565 if (N0.getOpcode() == ISD::XOR) 1566 Val = N0.getOperand(0); 1567 else { 1568 assert(N0.getOpcode() == ISD::AND && 1569 N0.getOperand(0).getOpcode() == ISD::XOR); 1570 // ((X^1)&1)^1 -> X & 1 1571 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(), 1572 N0.getOperand(0).getOperand(0), 1573 N0.getOperand(1)); 1574 } 1575 1576 return DAG.getSetCC(dl, VT, Val, N1, 1577 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); 1578 } 1579 } else if (N1C->getAPIntValue() == 1 && 1580 (VT == MVT::i1 || 1581 getBooleanContents(N0->getValueType(0)) == 1582 ZeroOrOneBooleanContent)) { 1583 SDValue Op0 = N0; 1584 if (Op0.getOpcode() == ISD::TRUNCATE) 1585 Op0 = Op0.getOperand(0); 1586 1587 if ((Op0.getOpcode() == ISD::XOR) && 1588 Op0.getOperand(0).getOpcode() == ISD::SETCC && 1589 Op0.getOperand(1).getOpcode() == ISD::SETCC) { 1590 // (xor (setcc), (setcc)) == / != 1 -> (setcc) != / == (setcc) 1591 Cond = (Cond == ISD::SETEQ) ? ISD::SETNE : ISD::SETEQ; 1592 return DAG.getSetCC(dl, VT, Op0.getOperand(0), Op0.getOperand(1), 1593 Cond); 1594 } 1595 if (Op0.getOpcode() == ISD::AND && 1596 isa<ConstantSDNode>(Op0.getOperand(1)) && 1597 cast<ConstantSDNode>(Op0.getOperand(1))->getAPIntValue() == 1) { 1598 // If this is (X&1) == / != 1, normalize it to (X&1) != / == 0. 1599 if (Op0.getValueType().bitsGT(VT)) 1600 Op0 = DAG.getNode(ISD::AND, dl, VT, 1601 DAG.getNode(ISD::TRUNCATE, dl, VT, Op0.getOperand(0)), 1602 DAG.getConstant(1, dl, VT)); 1603 else if (Op0.getValueType().bitsLT(VT)) 1604 Op0 = DAG.getNode(ISD::AND, dl, VT, 1605 DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op0.getOperand(0)), 1606 DAG.getConstant(1, dl, VT)); 1607 1608 return DAG.getSetCC(dl, VT, Op0, 1609 DAG.getConstant(0, dl, Op0.getValueType()), 1610 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); 1611 } 1612 if (Op0.getOpcode() == ISD::AssertZext && 1613 cast<VTSDNode>(Op0.getOperand(1))->getVT() == MVT::i1) 1614 return DAG.getSetCC(dl, VT, Op0, 1615 DAG.getConstant(0, dl, Op0.getValueType()), 1616 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); 1617 } 1618 } 1619 1620 APInt MinVal, MaxVal; 1621 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits(); 1622 if (ISD::isSignedIntSetCC(Cond)) { 1623 MinVal = APInt::getSignedMinValue(OperandBitSize); 1624 MaxVal = APInt::getSignedMaxValue(OperandBitSize); 1625 } else { 1626 MinVal = APInt::getMinValue(OperandBitSize); 1627 MaxVal = APInt::getMaxValue(OperandBitSize); 1628 } 1629 1630 // Canonicalize GE/LE comparisons to use GT/LT comparisons. 1631 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) { 1632 if (C1 == MinVal) return DAG.getConstant(1, dl, VT); // X >= MIN --> true 1633 // X >= C0 --> X > (C0 - 1) 1634 APInt C = C1 - 1; 1635 ISD::CondCode NewCC = (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT; 1636 if ((DCI.isBeforeLegalizeOps() || 1637 isCondCodeLegal(NewCC, VT.getSimpleVT())) && 1638 (!N1C->isOpaque() || (N1C->isOpaque() && C.getBitWidth() <= 64 && 1639 isLegalICmpImmediate(C.getSExtValue())))) { 1640 return DAG.getSetCC(dl, VT, N0, 1641 DAG.getConstant(C, dl, N1.getValueType()), 1642 NewCC); 1643 } 1644 } 1645 1646 if (Cond == ISD::SETLE || Cond == ISD::SETULE) { 1647 if (C1 == MaxVal) return DAG.getConstant(1, dl, VT); // X <= MAX --> true 1648 // X <= C0 --> X < (C0 + 1) 1649 APInt C = C1 + 1; 1650 ISD::CondCode NewCC = (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT; 1651 if ((DCI.isBeforeLegalizeOps() || 1652 isCondCodeLegal(NewCC, VT.getSimpleVT())) && 1653 (!N1C->isOpaque() || (N1C->isOpaque() && C.getBitWidth() <= 64 && 1654 isLegalICmpImmediate(C.getSExtValue())))) { 1655 return DAG.getSetCC(dl, VT, N0, 1656 DAG.getConstant(C, dl, N1.getValueType()), 1657 NewCC); 1658 } 1659 } 1660 1661 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal) 1662 return DAG.getConstant(0, dl, VT); // X < MIN --> false 1663 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal) 1664 return DAG.getConstant(1, dl, VT); // X >= MIN --> true 1665 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal) 1666 return DAG.getConstant(0, dl, VT); // X > MAX --> false 1667 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal) 1668 return DAG.getConstant(1, dl, VT); // X <= MAX --> true 1669 1670 // Canonicalize setgt X, Min --> setne X, Min 1671 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal) 1672 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE); 1673 // Canonicalize setlt X, Max --> setne X, Max 1674 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal) 1675 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE); 1676 1677 // If we have setult X, 1, turn it into seteq X, 0 1678 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1) 1679 return DAG.getSetCC(dl, VT, N0, 1680 DAG.getConstant(MinVal, dl, N0.getValueType()), 1681 ISD::SETEQ); 1682 // If we have setugt X, Max-1, turn it into seteq X, Max 1683 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1) 1684 return DAG.getSetCC(dl, VT, N0, 1685 DAG.getConstant(MaxVal, dl, N0.getValueType()), 1686 ISD::SETEQ); 1687 1688 // If we have "setcc X, C0", check to see if we can shrink the immediate 1689 // by changing cc. 1690 1691 // SETUGT X, SINTMAX -> SETLT X, 0 1692 if (Cond == ISD::SETUGT && 1693 C1 == APInt::getSignedMaxValue(OperandBitSize)) 1694 return DAG.getSetCC(dl, VT, N0, 1695 DAG.getConstant(0, dl, N1.getValueType()), 1696 ISD::SETLT); 1697 1698 // SETULT X, SINTMIN -> SETGT X, -1 1699 if (Cond == ISD::SETULT && 1700 C1 == APInt::getSignedMinValue(OperandBitSize)) { 1701 SDValue ConstMinusOne = 1702 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize), dl, 1703 N1.getValueType()); 1704 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT); 1705 } 1706 1707 // Fold bit comparisons when we can. 1708 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 1709 (VT == N0.getValueType() || 1710 (isTypeLegal(VT) && VT.bitsLE(N0.getValueType()))) && 1711 N0.getOpcode() == ISD::AND) { 1712 auto &DL = DAG.getDataLayout(); 1713 if (ConstantSDNode *AndRHS = 1714 dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 1715 EVT ShiftTy = DCI.isBeforeLegalize() 1716 ? getPointerTy(DL) 1717 : getShiftAmountTy(N0.getValueType(), DL); 1718 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3 1719 // Perform the xform if the AND RHS is a single bit. 1720 if (AndRHS->getAPIntValue().isPowerOf2()) { 1721 return DAG.getNode(ISD::TRUNCATE, dl, VT, 1722 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0, 1723 DAG.getConstant(AndRHS->getAPIntValue().logBase2(), dl, 1724 ShiftTy))); 1725 } 1726 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getAPIntValue()) { 1727 // (X & 8) == 8 --> (X & 8) >> 3 1728 // Perform the xform if C1 is a single bit. 1729 if (C1.isPowerOf2()) { 1730 return DAG.getNode(ISD::TRUNCATE, dl, VT, 1731 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0, 1732 DAG.getConstant(C1.logBase2(), dl, 1733 ShiftTy))); 1734 } 1735 } 1736 } 1737 } 1738 1739 if (C1.getMinSignedBits() <= 64 && 1740 !isLegalICmpImmediate(C1.getSExtValue())) { 1741 // (X & -256) == 256 -> (X >> 8) == 1 1742 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 1743 N0.getOpcode() == ISD::AND && N0.hasOneUse()) { 1744 if (ConstantSDNode *AndRHS = 1745 dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 1746 const APInt &AndRHSC = AndRHS->getAPIntValue(); 1747 if ((-AndRHSC).isPowerOf2() && (AndRHSC & C1) == C1) { 1748 unsigned ShiftBits = AndRHSC.countTrailingZeros(); 1749 auto &DL = DAG.getDataLayout(); 1750 EVT ShiftTy = DCI.isBeforeLegalize() 1751 ? getPointerTy(DL) 1752 : getShiftAmountTy(N0.getValueType(), DL); 1753 EVT CmpTy = N0.getValueType(); 1754 SDValue Shift = DAG.getNode(ISD::SRL, dl, CmpTy, N0.getOperand(0), 1755 DAG.getConstant(ShiftBits, dl, 1756 ShiftTy)); 1757 SDValue CmpRHS = DAG.getConstant(C1.lshr(ShiftBits), dl, CmpTy); 1758 return DAG.getSetCC(dl, VT, Shift, CmpRHS, Cond); 1759 } 1760 } 1761 } else if (Cond == ISD::SETULT || Cond == ISD::SETUGE || 1762 Cond == ISD::SETULE || Cond == ISD::SETUGT) { 1763 bool AdjOne = (Cond == ISD::SETULE || Cond == ISD::SETUGT); 1764 // X < 0x100000000 -> (X >> 32) < 1 1765 // X >= 0x100000000 -> (X >> 32) >= 1 1766 // X <= 0x0ffffffff -> (X >> 32) < 1 1767 // X > 0x0ffffffff -> (X >> 32) >= 1 1768 unsigned ShiftBits; 1769 APInt NewC = C1; 1770 ISD::CondCode NewCond = Cond; 1771 if (AdjOne) { 1772 ShiftBits = C1.countTrailingOnes(); 1773 NewC = NewC + 1; 1774 NewCond = (Cond == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE; 1775 } else { 1776 ShiftBits = C1.countTrailingZeros(); 1777 } 1778 NewC = NewC.lshr(ShiftBits); 1779 if (ShiftBits && NewC.getMinSignedBits() <= 64 && 1780 isLegalICmpImmediate(NewC.getSExtValue())) { 1781 auto &DL = DAG.getDataLayout(); 1782 EVT ShiftTy = DCI.isBeforeLegalize() 1783 ? getPointerTy(DL) 1784 : getShiftAmountTy(N0.getValueType(), DL); 1785 EVT CmpTy = N0.getValueType(); 1786 SDValue Shift = DAG.getNode(ISD::SRL, dl, CmpTy, N0, 1787 DAG.getConstant(ShiftBits, dl, ShiftTy)); 1788 SDValue CmpRHS = DAG.getConstant(NewC, dl, CmpTy); 1789 return DAG.getSetCC(dl, VT, Shift, CmpRHS, NewCond); 1790 } 1791 } 1792 } 1793 } 1794 1795 if (isa<ConstantFPSDNode>(N0.getNode())) { 1796 // Constant fold or commute setcc. 1797 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl); 1798 if (O.getNode()) return O; 1799 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) { 1800 // If the RHS of an FP comparison is a constant, simplify it away in 1801 // some cases. 1802 if (CFP->getValueAPF().isNaN()) { 1803 // If an operand is known to be a nan, we can fold it. 1804 switch (ISD::getUnorderedFlavor(Cond)) { 1805 default: llvm_unreachable("Unknown flavor!"); 1806 case 0: // Known false. 1807 return DAG.getConstant(0, dl, VT); 1808 case 1: // Known true. 1809 return DAG.getConstant(1, dl, VT); 1810 case 2: // Undefined. 1811 return DAG.getUNDEF(VT); 1812 } 1813 } 1814 1815 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the 1816 // constant if knowing that the operand is non-nan is enough. We prefer to 1817 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to 1818 // materialize 0.0. 1819 if (Cond == ISD::SETO || Cond == ISD::SETUO) 1820 return DAG.getSetCC(dl, VT, N0, N0, Cond); 1821 1822 // If the condition is not legal, see if we can find an equivalent one 1823 // which is legal. 1824 if (!isCondCodeLegal(Cond, N0.getSimpleValueType())) { 1825 // If the comparison was an awkward floating-point == or != and one of 1826 // the comparison operands is infinity or negative infinity, convert the 1827 // condition to a less-awkward <= or >=. 1828 if (CFP->getValueAPF().isInfinity()) { 1829 if (CFP->getValueAPF().isNegative()) { 1830 if (Cond == ISD::SETOEQ && 1831 isCondCodeLegal(ISD::SETOLE, N0.getSimpleValueType())) 1832 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLE); 1833 if (Cond == ISD::SETUEQ && 1834 isCondCodeLegal(ISD::SETOLE, N0.getSimpleValueType())) 1835 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULE); 1836 if (Cond == ISD::SETUNE && 1837 isCondCodeLegal(ISD::SETUGT, N0.getSimpleValueType())) 1838 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGT); 1839 if (Cond == ISD::SETONE && 1840 isCondCodeLegal(ISD::SETUGT, N0.getSimpleValueType())) 1841 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGT); 1842 } else { 1843 if (Cond == ISD::SETOEQ && 1844 isCondCodeLegal(ISD::SETOGE, N0.getSimpleValueType())) 1845 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGE); 1846 if (Cond == ISD::SETUEQ && 1847 isCondCodeLegal(ISD::SETOGE, N0.getSimpleValueType())) 1848 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGE); 1849 if (Cond == ISD::SETUNE && 1850 isCondCodeLegal(ISD::SETULT, N0.getSimpleValueType())) 1851 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULT); 1852 if (Cond == ISD::SETONE && 1853 isCondCodeLegal(ISD::SETULT, N0.getSimpleValueType())) 1854 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLT); 1855 } 1856 } 1857 } 1858 } 1859 1860 if (N0 == N1) { 1861 // The sext(setcc()) => setcc() optimization relies on the appropriate 1862 // constant being emitted. 1863 uint64_t EqVal = 0; 1864 switch (getBooleanContents(N0.getValueType())) { 1865 case UndefinedBooleanContent: 1866 case ZeroOrOneBooleanContent: 1867 EqVal = ISD::isTrueWhenEqual(Cond); 1868 break; 1869 case ZeroOrNegativeOneBooleanContent: 1870 EqVal = ISD::isTrueWhenEqual(Cond) ? -1 : 0; 1871 break; 1872 } 1873 1874 // We can always fold X == X for integer setcc's. 1875 if (N0.getValueType().isInteger()) { 1876 return DAG.getConstant(EqVal, dl, VT); 1877 } 1878 unsigned UOF = ISD::getUnorderedFlavor(Cond); 1879 if (UOF == 2) // FP operators that are undefined on NaNs. 1880 return DAG.getConstant(EqVal, dl, VT); 1881 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond))) 1882 return DAG.getConstant(EqVal, dl, VT); 1883 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO 1884 // if it is not already. 1885 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO; 1886 if (NewCond != Cond && (DCI.isBeforeLegalizeOps() || 1887 getCondCodeAction(NewCond, N0.getSimpleValueType()) == Legal)) 1888 return DAG.getSetCC(dl, VT, N0, N1, NewCond); 1889 } 1890 1891 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 1892 N0.getValueType().isInteger()) { 1893 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB || 1894 N0.getOpcode() == ISD::XOR) { 1895 // Simplify (X+Y) == (X+Z) --> Y == Z 1896 if (N0.getOpcode() == N1.getOpcode()) { 1897 if (N0.getOperand(0) == N1.getOperand(0)) 1898 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond); 1899 if (N0.getOperand(1) == N1.getOperand(1)) 1900 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond); 1901 if (DAG.isCommutativeBinOp(N0.getOpcode())) { 1902 // If X op Y == Y op X, try other combinations. 1903 if (N0.getOperand(0) == N1.getOperand(1)) 1904 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0), 1905 Cond); 1906 if (N0.getOperand(1) == N1.getOperand(0)) 1907 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1), 1908 Cond); 1909 } 1910 } 1911 1912 // If RHS is a legal immediate value for a compare instruction, we need 1913 // to be careful about increasing register pressure needlessly. 1914 bool LegalRHSImm = false; 1915 1916 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) { 1917 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 1918 // Turn (X+C1) == C2 --> X == C2-C1 1919 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) { 1920 return DAG.getSetCC(dl, VT, N0.getOperand(0), 1921 DAG.getConstant(RHSC->getAPIntValue()- 1922 LHSR->getAPIntValue(), 1923 dl, N0.getValueType()), Cond); 1924 } 1925 1926 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0. 1927 if (N0.getOpcode() == ISD::XOR) 1928 // If we know that all of the inverted bits are zero, don't bother 1929 // performing the inversion. 1930 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue())) 1931 return 1932 DAG.getSetCC(dl, VT, N0.getOperand(0), 1933 DAG.getConstant(LHSR->getAPIntValue() ^ 1934 RHSC->getAPIntValue(), 1935 dl, N0.getValueType()), 1936 Cond); 1937 } 1938 1939 // Turn (C1-X) == C2 --> X == C1-C2 1940 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) { 1941 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) { 1942 return 1943 DAG.getSetCC(dl, VT, N0.getOperand(1), 1944 DAG.getConstant(SUBC->getAPIntValue() - 1945 RHSC->getAPIntValue(), 1946 dl, N0.getValueType()), 1947 Cond); 1948 } 1949 } 1950 1951 // Could RHSC fold directly into a compare? 1952 if (RHSC->getValueType(0).getSizeInBits() <= 64) 1953 LegalRHSImm = isLegalICmpImmediate(RHSC->getSExtValue()); 1954 } 1955 1956 // Simplify (X+Z) == X --> Z == 0 1957 // Don't do this if X is an immediate that can fold into a cmp 1958 // instruction and X+Z has other uses. It could be an induction variable 1959 // chain, and the transform would increase register pressure. 1960 if (!LegalRHSImm || N0.getNode()->hasOneUse()) { 1961 if (N0.getOperand(0) == N1) 1962 return DAG.getSetCC(dl, VT, N0.getOperand(1), 1963 DAG.getConstant(0, dl, N0.getValueType()), Cond); 1964 if (N0.getOperand(1) == N1) { 1965 if (DAG.isCommutativeBinOp(N0.getOpcode())) 1966 return DAG.getSetCC(dl, VT, N0.getOperand(0), 1967 DAG.getConstant(0, dl, N0.getValueType()), 1968 Cond); 1969 if (N0.getNode()->hasOneUse()) { 1970 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!"); 1971 auto &DL = DAG.getDataLayout(); 1972 // (Z-X) == X --> Z == X<<1 1973 SDValue SH = DAG.getNode( 1974 ISD::SHL, dl, N1.getValueType(), N1, 1975 DAG.getConstant(1, dl, 1976 getShiftAmountTy(N1.getValueType(), DL))); 1977 if (!DCI.isCalledByLegalizer()) 1978 DCI.AddToWorklist(SH.getNode()); 1979 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond); 1980 } 1981 } 1982 } 1983 } 1984 1985 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB || 1986 N1.getOpcode() == ISD::XOR) { 1987 // Simplify X == (X+Z) --> Z == 0 1988 if (N1.getOperand(0) == N0) 1989 return DAG.getSetCC(dl, VT, N1.getOperand(1), 1990 DAG.getConstant(0, dl, N1.getValueType()), Cond); 1991 if (N1.getOperand(1) == N0) { 1992 if (DAG.isCommutativeBinOp(N1.getOpcode())) 1993 return DAG.getSetCC(dl, VT, N1.getOperand(0), 1994 DAG.getConstant(0, dl, N1.getValueType()), Cond); 1995 if (N1.getNode()->hasOneUse()) { 1996 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!"); 1997 auto &DL = DAG.getDataLayout(); 1998 // X == (Z-X) --> X<<1 == Z 1999 SDValue SH = DAG.getNode( 2000 ISD::SHL, dl, N1.getValueType(), N0, 2001 DAG.getConstant(1, dl, getShiftAmountTy(N0.getValueType(), DL))); 2002 if (!DCI.isCalledByLegalizer()) 2003 DCI.AddToWorklist(SH.getNode()); 2004 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond); 2005 } 2006 } 2007 } 2008 2009 // Simplify x&y == y to x&y != 0 if y has exactly one bit set. 2010 // Note that where y is variable and is known to have at most 2011 // one bit set (for example, if it is z&1) we cannot do this; 2012 // the expressions are not equivalent when y==0. 2013 if (N0.getOpcode() == ISD::AND) 2014 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) { 2015 if (ValueHasExactlyOneBitSet(N1, DAG)) { 2016 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true); 2017 if (DCI.isBeforeLegalizeOps() || 2018 isCondCodeLegal(Cond, N0.getSimpleValueType())) { 2019 SDValue Zero = DAG.getConstant(0, dl, N1.getValueType()); 2020 return DAG.getSetCC(dl, VT, N0, Zero, Cond); 2021 } 2022 } 2023 } 2024 if (N1.getOpcode() == ISD::AND) 2025 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) { 2026 if (ValueHasExactlyOneBitSet(N0, DAG)) { 2027 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true); 2028 if (DCI.isBeforeLegalizeOps() || 2029 isCondCodeLegal(Cond, N1.getSimpleValueType())) { 2030 SDValue Zero = DAG.getConstant(0, dl, N0.getValueType()); 2031 return DAG.getSetCC(dl, VT, N1, Zero, Cond); 2032 } 2033 } 2034 } 2035 } 2036 2037 // Fold away ALL boolean setcc's. 2038 SDValue Temp; 2039 if (N0.getValueType() == MVT::i1 && foldBooleans) { 2040 switch (Cond) { 2041 default: llvm_unreachable("Unknown integer setcc!"); 2042 case ISD::SETEQ: // X == Y -> ~(X^Y) 2043 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1); 2044 N0 = DAG.getNOT(dl, Temp, MVT::i1); 2045 if (!DCI.isCalledByLegalizer()) 2046 DCI.AddToWorklist(Temp.getNode()); 2047 break; 2048 case ISD::SETNE: // X != Y --> (X^Y) 2049 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1); 2050 break; 2051 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y 2052 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y 2053 Temp = DAG.getNOT(dl, N0, MVT::i1); 2054 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp); 2055 if (!DCI.isCalledByLegalizer()) 2056 DCI.AddToWorklist(Temp.getNode()); 2057 break; 2058 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X 2059 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X 2060 Temp = DAG.getNOT(dl, N1, MVT::i1); 2061 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp); 2062 if (!DCI.isCalledByLegalizer()) 2063 DCI.AddToWorklist(Temp.getNode()); 2064 break; 2065 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y 2066 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y 2067 Temp = DAG.getNOT(dl, N0, MVT::i1); 2068 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp); 2069 if (!DCI.isCalledByLegalizer()) 2070 DCI.AddToWorklist(Temp.getNode()); 2071 break; 2072 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X 2073 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X 2074 Temp = DAG.getNOT(dl, N1, MVT::i1); 2075 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp); 2076 break; 2077 } 2078 if (VT != MVT::i1) { 2079 if (!DCI.isCalledByLegalizer()) 2080 DCI.AddToWorklist(N0.getNode()); 2081 // FIXME: If running after legalize, we probably can't do this. 2082 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0); 2083 } 2084 return N0; 2085 } 2086 2087 // Could not fold it. 2088 return SDValue(); 2089} 2090 2091/// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the 2092/// node is a GlobalAddress + offset. 2093bool TargetLowering::isGAPlusOffset(SDNode *N, const GlobalValue *&GA, 2094 int64_t &Offset) const { 2095 if (isa<GlobalAddressSDNode>(N)) { 2096 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N); 2097 GA = GASD->getGlobal(); 2098 Offset += GASD->getOffset(); 2099 return true; 2100 } 2101 2102 if (N->getOpcode() == ISD::ADD) { 2103 SDValue N1 = N->getOperand(0); 2104 SDValue N2 = N->getOperand(1); 2105 if (isGAPlusOffset(N1.getNode(), GA, Offset)) { 2106 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2); 2107 if (V) { 2108 Offset += V->getSExtValue(); 2109 return true; 2110 } 2111 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) { 2112 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1); 2113 if (V) { 2114 Offset += V->getSExtValue(); 2115 return true; 2116 } 2117 } 2118 } 2119 2120 return false; 2121} 2122 2123SDValue TargetLowering::PerformDAGCombine(SDNode *N, 2124 DAGCombinerInfo &DCI) const { 2125 // Default implementation: no optimization. 2126 return SDValue(); 2127} 2128 2129//===----------------------------------------------------------------------===// 2130// Inline Assembler Implementation Methods 2131//===----------------------------------------------------------------------===// 2132 2133TargetLowering::ConstraintType 2134TargetLowering::getConstraintType(StringRef Constraint) const { 2135 unsigned S = Constraint.size(); 2136 2137 if (S == 1) { 2138 switch (Constraint[0]) { 2139 default: break; 2140 case 'r': return C_RegisterClass; 2141 case 'm': // memory 2142 case 'o': // offsetable 2143 case 'V': // not offsetable 2144 return C_Memory; 2145 case 'i': // Simple Integer or Relocatable Constant 2146 case 'n': // Simple Integer 2147 case 'E': // Floating Point Constant 2148 case 'F': // Floating Point Constant 2149 case 's': // Relocatable Constant 2150 case 'p': // Address. 2151 case 'X': // Allow ANY value. 2152 case 'I': // Target registers. 2153 case 'J': 2154 case 'K': 2155 case 'L': 2156 case 'M': 2157 case 'N': 2158 case 'O': 2159 case 'P': 2160 case '<': 2161 case '>': 2162 return C_Other; 2163 } 2164 } 2165 2166 if (S > 1 && Constraint[0] == '{' && Constraint[S-1] == '}') { 2167 if (S == 8 && Constraint.substr(1, 6) == "memory") // "{memory}" 2168 return C_Memory; 2169 return C_Register; 2170 } 2171 return C_Unknown; 2172} 2173 2174/// LowerXConstraint - try to replace an X constraint, which matches anything, 2175/// with another that has more specific requirements based on the type of the 2176/// corresponding operand. 2177const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const{ 2178 if (ConstraintVT.isInteger()) 2179 return "r"; 2180 if (ConstraintVT.isFloatingPoint()) 2181 return "f"; // works for many targets 2182 return nullptr; 2183} 2184 2185/// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 2186/// vector. If it is invalid, don't add anything to Ops. 2187void TargetLowering::LowerAsmOperandForConstraint(SDValue Op, 2188 std::string &Constraint, 2189 std::vector<SDValue> &Ops, 2190 SelectionDAG &DAG) const { 2191 2192 if (Constraint.length() > 1) return; 2193 2194 char ConstraintLetter = Constraint[0]; 2195 switch (ConstraintLetter) { 2196 default: break; 2197 case 'X': // Allows any operand; labels (basic block) use this. 2198 if (Op.getOpcode() == ISD::BasicBlock) { 2199 Ops.push_back(Op); 2200 return; 2201 } 2202 // fall through 2203 case 'i': // Simple Integer or Relocatable Constant 2204 case 'n': // Simple Integer 2205 case 's': { // Relocatable Constant 2206 // These operands are interested in values of the form (GV+C), where C may 2207 // be folded in as an offset of GV, or it may be explicitly added. Also, it 2208 // is possible and fine if either GV or C are missing. 2209 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op); 2210 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op); 2211 2212 // If we have "(add GV, C)", pull out GV/C 2213 if (Op.getOpcode() == ISD::ADD) { 2214 C = dyn_cast<ConstantSDNode>(Op.getOperand(1)); 2215 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0)); 2216 if (!C || !GA) { 2217 C = dyn_cast<ConstantSDNode>(Op.getOperand(0)); 2218 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1)); 2219 } 2220 if (!C || !GA) 2221 C = nullptr, GA = nullptr; 2222 } 2223 2224 // If we find a valid operand, map to the TargetXXX version so that the 2225 // value itself doesn't get selected. 2226 if (GA) { // Either &GV or &GV+C 2227 if (ConstraintLetter != 'n') { 2228 int64_t Offs = GA->getOffset(); 2229 if (C) Offs += C->getZExtValue(); 2230 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(), 2231 C ? SDLoc(C) : SDLoc(), 2232 Op.getValueType(), Offs)); 2233 } 2234 return; 2235 } 2236 if (C) { // just C, no GV. 2237 // Simple constants are not allowed for 's'. 2238 if (ConstraintLetter != 's') { 2239 // gcc prints these as sign extended. Sign extend value to 64 bits 2240 // now; without this it would get ZExt'd later in 2241 // ScheduleDAGSDNodes::EmitNode, which is very generic. 2242 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(), 2243 SDLoc(C), MVT::i64)); 2244 } 2245 return; 2246 } 2247 break; 2248 } 2249 } 2250} 2251 2252std::pair<unsigned, const TargetRegisterClass *> 2253TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *RI, 2254 StringRef Constraint, 2255 MVT VT) const { 2256 if (Constraint.empty() || Constraint[0] != '{') 2257 return std::make_pair(0u, static_cast<TargetRegisterClass*>(nullptr)); 2258 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?"); 2259 2260 // Remove the braces from around the name. 2261 StringRef RegName(Constraint.data()+1, Constraint.size()-2); 2262 2263 std::pair<unsigned, const TargetRegisterClass*> R = 2264 std::make_pair(0u, static_cast<const TargetRegisterClass*>(nullptr)); 2265 2266 // Figure out which register class contains this reg. 2267 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(), 2268 E = RI->regclass_end(); RCI != E; ++RCI) { 2269 const TargetRegisterClass *RC = *RCI; 2270 2271 // If none of the value types for this register class are valid, we 2272 // can't use it. For example, 64-bit reg classes on 32-bit targets. 2273 if (!isLegalRC(RC)) 2274 continue; 2275 2276 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end(); 2277 I != E; ++I) { 2278 if (RegName.equals_lower(RI->getName(*I))) { 2279 std::pair<unsigned, const TargetRegisterClass*> S = 2280 std::make_pair(*I, RC); 2281 2282 // If this register class has the requested value type, return it, 2283 // otherwise keep searching and return the first class found 2284 // if no other is found which explicitly has the requested type. 2285 if (RC->hasType(VT)) 2286 return S; 2287 else if (!R.second) 2288 R = S; 2289 } 2290 } 2291 } 2292 2293 return R; 2294} 2295 2296//===----------------------------------------------------------------------===// 2297// Constraint Selection. 2298 2299/// isMatchingInputConstraint - Return true of this is an input operand that is 2300/// a matching constraint like "4". 2301bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const { 2302 assert(!ConstraintCode.empty() && "No known constraint!"); 2303 return isdigit(static_cast<unsigned char>(ConstraintCode[0])); 2304} 2305 2306/// getMatchedOperand - If this is an input matching constraint, this method 2307/// returns the output operand it matches. 2308unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const { 2309 assert(!ConstraintCode.empty() && "No known constraint!"); 2310 return atoi(ConstraintCode.c_str()); 2311} 2312 2313/// ParseConstraints - Split up the constraint string from the inline 2314/// assembly value into the specific constraints and their prefixes, 2315/// and also tie in the associated operand values. 2316/// If this returns an empty vector, and if the constraint string itself 2317/// isn't empty, there was an error parsing. 2318TargetLowering::AsmOperandInfoVector 2319TargetLowering::ParseConstraints(const DataLayout &DL, 2320 const TargetRegisterInfo *TRI, 2321 ImmutableCallSite CS) const { 2322 /// ConstraintOperands - Information about all of the constraints. 2323 AsmOperandInfoVector ConstraintOperands; 2324 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue()); 2325 unsigned maCount = 0; // Largest number of multiple alternative constraints. 2326 2327 // Do a prepass over the constraints, canonicalizing them, and building up the 2328 // ConstraintOperands list. 2329 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst. 2330 unsigned ResNo = 0; // ResNo - The result number of the next output. 2331 2332 for (InlineAsm::ConstraintInfo &CI : IA->ParseConstraints()) { 2333 ConstraintOperands.emplace_back(std::move(CI)); 2334 AsmOperandInfo &OpInfo = ConstraintOperands.back(); 2335 2336 // Update multiple alternative constraint count. 2337 if (OpInfo.multipleAlternatives.size() > maCount) 2338 maCount = OpInfo.multipleAlternatives.size(); 2339 2340 OpInfo.ConstraintVT = MVT::Other; 2341 2342 // Compute the value type for each operand. 2343 switch (OpInfo.Type) { 2344 case InlineAsm::isOutput: 2345 // Indirect outputs just consume an argument. 2346 if (OpInfo.isIndirect) { 2347 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++)); 2348 break; 2349 } 2350 2351 // The return value of the call is this value. As such, there is no 2352 // corresponding argument. 2353 assert(!CS.getType()->isVoidTy() && 2354 "Bad inline asm!"); 2355 if (StructType *STy = dyn_cast<StructType>(CS.getType())) { 2356 OpInfo.ConstraintVT = 2357 getSimpleValueType(DL, STy->getElementType(ResNo)); 2358 } else { 2359 assert(ResNo == 0 && "Asm only has one result!"); 2360 OpInfo.ConstraintVT = getSimpleValueType(DL, CS.getType()); 2361 } 2362 ++ResNo; 2363 break; 2364 case InlineAsm::isInput: 2365 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++)); 2366 break; 2367 case InlineAsm::isClobber: 2368 // Nothing to do. 2369 break; 2370 } 2371 2372 if (OpInfo.CallOperandVal) { 2373 llvm::Type *OpTy = OpInfo.CallOperandVal->getType(); 2374 if (OpInfo.isIndirect) { 2375 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy); 2376 if (!PtrTy) 2377 report_fatal_error("Indirect operand for inline asm not a pointer!"); 2378 OpTy = PtrTy->getElementType(); 2379 } 2380 2381 // Look for vector wrapped in a struct. e.g. { <16 x i8> }. 2382 if (StructType *STy = dyn_cast<StructType>(OpTy)) 2383 if (STy->getNumElements() == 1) 2384 OpTy = STy->getElementType(0); 2385 2386 // If OpTy is not a single value, it may be a struct/union that we 2387 // can tile with integers. 2388 if (!OpTy->isSingleValueType() && OpTy->isSized()) { 2389 unsigned BitSize = DL.getTypeSizeInBits(OpTy); 2390 switch (BitSize) { 2391 default: break; 2392 case 1: 2393 case 8: 2394 case 16: 2395 case 32: 2396 case 64: 2397 case 128: 2398 OpInfo.ConstraintVT = 2399 MVT::getVT(IntegerType::get(OpTy->getContext(), BitSize), true); 2400 break; 2401 } 2402 } else if (PointerType *PT = dyn_cast<PointerType>(OpTy)) { 2403 unsigned PtrSize = DL.getPointerSizeInBits(PT->getAddressSpace()); 2404 OpInfo.ConstraintVT = MVT::getIntegerVT(PtrSize); 2405 } else { 2406 OpInfo.ConstraintVT = MVT::getVT(OpTy, true); 2407 } 2408 } 2409 } 2410 2411 // If we have multiple alternative constraints, select the best alternative. 2412 if (!ConstraintOperands.empty()) { 2413 if (maCount) { 2414 unsigned bestMAIndex = 0; 2415 int bestWeight = -1; 2416 // weight: -1 = invalid match, and 0 = so-so match to 5 = good match. 2417 int weight = -1; 2418 unsigned maIndex; 2419 // Compute the sums of the weights for each alternative, keeping track 2420 // of the best (highest weight) one so far. 2421 for (maIndex = 0; maIndex < maCount; ++maIndex) { 2422 int weightSum = 0; 2423 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size(); 2424 cIndex != eIndex; ++cIndex) { 2425 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex]; 2426 if (OpInfo.Type == InlineAsm::isClobber) 2427 continue; 2428 2429 // If this is an output operand with a matching input operand, 2430 // look up the matching input. If their types mismatch, e.g. one 2431 // is an integer, the other is floating point, or their sizes are 2432 // different, flag it as an maCantMatch. 2433 if (OpInfo.hasMatchingInput()) { 2434 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput]; 2435 if (OpInfo.ConstraintVT != Input.ConstraintVT) { 2436 if ((OpInfo.ConstraintVT.isInteger() != 2437 Input.ConstraintVT.isInteger()) || 2438 (OpInfo.ConstraintVT.getSizeInBits() != 2439 Input.ConstraintVT.getSizeInBits())) { 2440 weightSum = -1; // Can't match. 2441 break; 2442 } 2443 } 2444 } 2445 weight = getMultipleConstraintMatchWeight(OpInfo, maIndex); 2446 if (weight == -1) { 2447 weightSum = -1; 2448 break; 2449 } 2450 weightSum += weight; 2451 } 2452 // Update best. 2453 if (weightSum > bestWeight) { 2454 bestWeight = weightSum; 2455 bestMAIndex = maIndex; 2456 } 2457 } 2458 2459 // Now select chosen alternative in each constraint. 2460 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size(); 2461 cIndex != eIndex; ++cIndex) { 2462 AsmOperandInfo& cInfo = ConstraintOperands[cIndex]; 2463 if (cInfo.Type == InlineAsm::isClobber) 2464 continue; 2465 cInfo.selectAlternative(bestMAIndex); 2466 } 2467 } 2468 } 2469 2470 // Check and hook up tied operands, choose constraint code to use. 2471 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size(); 2472 cIndex != eIndex; ++cIndex) { 2473 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex]; 2474 2475 // If this is an output operand with a matching input operand, look up the 2476 // matching input. If their types mismatch, e.g. one is an integer, the 2477 // other is floating point, or their sizes are different, flag it as an 2478 // error. 2479 if (OpInfo.hasMatchingInput()) { 2480 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput]; 2481 2482 if (OpInfo.ConstraintVT != Input.ConstraintVT) { 2483 std::pair<unsigned, const TargetRegisterClass *> MatchRC = 2484 getRegForInlineAsmConstraint(TRI, OpInfo.ConstraintCode, 2485 OpInfo.ConstraintVT); 2486 std::pair<unsigned, const TargetRegisterClass *> InputRC = 2487 getRegForInlineAsmConstraint(TRI, Input.ConstraintCode, 2488 Input.ConstraintVT); 2489 if ((OpInfo.ConstraintVT.isInteger() != 2490 Input.ConstraintVT.isInteger()) || 2491 (MatchRC.second != InputRC.second)) { 2492 report_fatal_error("Unsupported asm: input constraint" 2493 " with a matching output constraint of" 2494 " incompatible type!"); 2495 } 2496 } 2497 } 2498 } 2499 2500 return ConstraintOperands; 2501} 2502 2503/// getConstraintGenerality - Return an integer indicating how general CT 2504/// is. 2505static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) { 2506 switch (CT) { 2507 case TargetLowering::C_Other: 2508 case TargetLowering::C_Unknown: 2509 return 0; 2510 case TargetLowering::C_Register: 2511 return 1; 2512 case TargetLowering::C_RegisterClass: 2513 return 2; 2514 case TargetLowering::C_Memory: 2515 return 3; 2516 } 2517 llvm_unreachable("Invalid constraint type"); 2518} 2519 2520/// Examine constraint type and operand type and determine a weight value. 2521/// This object must already have been set up with the operand type 2522/// and the current alternative constraint selected. 2523TargetLowering::ConstraintWeight 2524 TargetLowering::getMultipleConstraintMatchWeight( 2525 AsmOperandInfo &info, int maIndex) const { 2526 InlineAsm::ConstraintCodeVector *rCodes; 2527 if (maIndex >= (int)info.multipleAlternatives.size()) 2528 rCodes = &info.Codes; 2529 else 2530 rCodes = &info.multipleAlternatives[maIndex].Codes; 2531 ConstraintWeight BestWeight = CW_Invalid; 2532 2533 // Loop over the options, keeping track of the most general one. 2534 for (unsigned i = 0, e = rCodes->size(); i != e; ++i) { 2535 ConstraintWeight weight = 2536 getSingleConstraintMatchWeight(info, (*rCodes)[i].c_str()); 2537 if (weight > BestWeight) 2538 BestWeight = weight; 2539 } 2540 2541 return BestWeight; 2542} 2543 2544/// Examine constraint type and operand type and determine a weight value. 2545/// This object must already have been set up with the operand type 2546/// and the current alternative constraint selected. 2547TargetLowering::ConstraintWeight 2548 TargetLowering::getSingleConstraintMatchWeight( 2549 AsmOperandInfo &info, const char *constraint) const { 2550 ConstraintWeight weight = CW_Invalid; 2551 Value *CallOperandVal = info.CallOperandVal; 2552 // If we don't have a value, we can't do a match, 2553 // but allow it at the lowest weight. 2554 if (!CallOperandVal) 2555 return CW_Default; 2556 // Look at the constraint type. 2557 switch (*constraint) { 2558 case 'i': // immediate integer. 2559 case 'n': // immediate integer with a known value. 2560 if (isa<ConstantInt>(CallOperandVal)) 2561 weight = CW_Constant; 2562 break; 2563 case 's': // non-explicit intregal immediate. 2564 if (isa<GlobalValue>(CallOperandVal)) 2565 weight = CW_Constant; 2566 break; 2567 case 'E': // immediate float if host format. 2568 case 'F': // immediate float. 2569 if (isa<ConstantFP>(CallOperandVal)) 2570 weight = CW_Constant; 2571 break; 2572 case '<': // memory operand with autodecrement. 2573 case '>': // memory operand with autoincrement. 2574 case 'm': // memory operand. 2575 case 'o': // offsettable memory operand 2576 case 'V': // non-offsettable memory operand 2577 weight = CW_Memory; 2578 break; 2579 case 'r': // general register. 2580 case 'g': // general register, memory operand or immediate integer. 2581 // note: Clang converts "g" to "imr". 2582 if (CallOperandVal->getType()->isIntegerTy()) 2583 weight = CW_Register; 2584 break; 2585 case 'X': // any operand. 2586 default: 2587 weight = CW_Default; 2588 break; 2589 } 2590 return weight; 2591} 2592 2593/// ChooseConstraint - If there are multiple different constraints that we 2594/// could pick for this operand (e.g. "imr") try to pick the 'best' one. 2595/// This is somewhat tricky: constraints fall into four classes: 2596/// Other -> immediates and magic values 2597/// Register -> one specific register 2598/// RegisterClass -> a group of regs 2599/// Memory -> memory 2600/// Ideally, we would pick the most specific constraint possible: if we have 2601/// something that fits into a register, we would pick it. The problem here 2602/// is that if we have something that could either be in a register or in 2603/// memory that use of the register could cause selection of *other* 2604/// operands to fail: they might only succeed if we pick memory. Because of 2605/// this the heuristic we use is: 2606/// 2607/// 1) If there is an 'other' constraint, and if the operand is valid for 2608/// that constraint, use it. This makes us take advantage of 'i' 2609/// constraints when available. 2610/// 2) Otherwise, pick the most general constraint present. This prefers 2611/// 'm' over 'r', for example. 2612/// 2613static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo, 2614 const TargetLowering &TLI, 2615 SDValue Op, SelectionDAG *DAG) { 2616 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options"); 2617 unsigned BestIdx = 0; 2618 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown; 2619 int BestGenerality = -1; 2620 2621 // Loop over the options, keeping track of the most general one. 2622 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) { 2623 TargetLowering::ConstraintType CType = 2624 TLI.getConstraintType(OpInfo.Codes[i]); 2625 2626 // If this is an 'other' constraint, see if the operand is valid for it. 2627 // For example, on X86 we might have an 'rI' constraint. If the operand 2628 // is an integer in the range [0..31] we want to use I (saving a load 2629 // of a register), otherwise we must use 'r'. 2630 if (CType == TargetLowering::C_Other && Op.getNode()) { 2631 assert(OpInfo.Codes[i].size() == 1 && 2632 "Unhandled multi-letter 'other' constraint"); 2633 std::vector<SDValue> ResultOps; 2634 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i], 2635 ResultOps, *DAG); 2636 if (!ResultOps.empty()) { 2637 BestType = CType; 2638 BestIdx = i; 2639 break; 2640 } 2641 } 2642 2643 // Things with matching constraints can only be registers, per gcc 2644 // documentation. This mainly affects "g" constraints. 2645 if (CType == TargetLowering::C_Memory && OpInfo.hasMatchingInput()) 2646 continue; 2647 2648 // This constraint letter is more general than the previous one, use it. 2649 int Generality = getConstraintGenerality(CType); 2650 if (Generality > BestGenerality) { 2651 BestType = CType; 2652 BestIdx = i; 2653 BestGenerality = Generality; 2654 } 2655 } 2656 2657 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx]; 2658 OpInfo.ConstraintType = BestType; 2659} 2660 2661/// ComputeConstraintToUse - Determines the constraint code and constraint 2662/// type to use for the specific AsmOperandInfo, setting 2663/// OpInfo.ConstraintCode and OpInfo.ConstraintType. 2664void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo, 2665 SDValue Op, 2666 SelectionDAG *DAG) const { 2667 assert(!OpInfo.Codes.empty() && "Must have at least one constraint"); 2668 2669 // Single-letter constraints ('r') are very common. 2670 if (OpInfo.Codes.size() == 1) { 2671 OpInfo.ConstraintCode = OpInfo.Codes[0]; 2672 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode); 2673 } else { 2674 ChooseConstraint(OpInfo, *this, Op, DAG); 2675 } 2676 2677 // 'X' matches anything. 2678 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) { 2679 // Labels and constants are handled elsewhere ('X' is the only thing 2680 // that matches labels). For Functions, the type here is the type of 2681 // the result, which is not what we want to look at; leave them alone. 2682 Value *v = OpInfo.CallOperandVal; 2683 if (isa<BasicBlock>(v) || isa<ConstantInt>(v) || isa<Function>(v)) { 2684 OpInfo.CallOperandVal = v; 2685 return; 2686 } 2687 2688 // Otherwise, try to resolve it to something we know about by looking at 2689 // the actual operand type. 2690 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) { 2691 OpInfo.ConstraintCode = Repl; 2692 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode); 2693 } 2694 } 2695} 2696 2697/// \brief Given an exact SDIV by a constant, create a multiplication 2698/// with the multiplicative inverse of the constant. 2699static SDValue BuildExactSDIV(const TargetLowering &TLI, SDValue Op1, APInt d, 2700 SDLoc dl, SelectionDAG &DAG, 2701 std::vector<SDNode *> &Created) { 2702 assert(d != 0 && "Division by zero!"); 2703 2704 // Shift the value upfront if it is even, so the LSB is one. 2705 unsigned ShAmt = d.countTrailingZeros(); 2706 if (ShAmt) { 2707 // TODO: For UDIV use SRL instead of SRA. 2708 SDValue Amt = 2709 DAG.getConstant(ShAmt, dl, TLI.getShiftAmountTy(Op1.getValueType(), 2710 DAG.getDataLayout())); 2711 SDNodeFlags Flags; 2712 Flags.setExact(true); 2713 Op1 = DAG.getNode(ISD::SRA, dl, Op1.getValueType(), Op1, Amt, &Flags); 2714 Created.push_back(Op1.getNode()); 2715 d = d.ashr(ShAmt); 2716 } 2717 2718 // Calculate the multiplicative inverse, using Newton's method. 2719 APInt t, xn = d; 2720 while ((t = d*xn) != 1) 2721 xn *= APInt(d.getBitWidth(), 2) - t; 2722 2723 SDValue Op2 = DAG.getConstant(xn, dl, Op1.getValueType()); 2724 SDValue Mul = DAG.getNode(ISD::MUL, dl, Op1.getValueType(), Op1, Op2); 2725 Created.push_back(Mul.getNode()); 2726 return Mul; 2727} 2728 2729SDValue TargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor, 2730 SelectionDAG &DAG, 2731 std::vector<SDNode *> *Created) const { 2732 AttributeSet Attr = DAG.getMachineFunction().getFunction()->getAttributes(); 2733 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 2734 if (TLI.isIntDivCheap(N->getValueType(0), Attr)) 2735 return SDValue(N,0); // Lower SDIV as SDIV 2736 return SDValue(); 2737} 2738 2739/// \brief Given an ISD::SDIV node expressing a divide by constant, 2740/// return a DAG expression to select that will generate the same value by 2741/// multiplying by a magic number. 2742/// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide". 2743SDValue TargetLowering::BuildSDIV(SDNode *N, const APInt &Divisor, 2744 SelectionDAG &DAG, bool IsAfterLegalization, 2745 std::vector<SDNode *> *Created) const { 2746 assert(Created && "No vector to hold sdiv ops."); 2747 2748 EVT VT = N->getValueType(0); 2749 SDLoc dl(N); 2750 2751 // Check to see if we can do this. 2752 // FIXME: We should be more aggressive here. 2753 if (!isTypeLegal(VT)) 2754 return SDValue(); 2755 2756 // If the sdiv has an 'exact' bit we can use a simpler lowering. 2757 if (cast<BinaryWithFlagsSDNode>(N)->Flags.hasExact()) 2758 return BuildExactSDIV(*this, N->getOperand(0), Divisor, dl, DAG, *Created); 2759 2760 APInt::ms magics = Divisor.magic(); 2761 2762 // Multiply the numerator (operand 0) by the magic value 2763 // FIXME: We should support doing a MUL in a wider type 2764 SDValue Q; 2765 if (IsAfterLegalization ? isOperationLegal(ISD::MULHS, VT) : 2766 isOperationLegalOrCustom(ISD::MULHS, VT)) 2767 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0), 2768 DAG.getConstant(magics.m, dl, VT)); 2769 else if (IsAfterLegalization ? isOperationLegal(ISD::SMUL_LOHI, VT) : 2770 isOperationLegalOrCustom(ISD::SMUL_LOHI, VT)) 2771 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT), 2772 N->getOperand(0), 2773 DAG.getConstant(magics.m, dl, VT)).getNode(), 1); 2774 else 2775 return SDValue(); // No mulhs or equvialent 2776 // If d > 0 and m < 0, add the numerator 2777 if (Divisor.isStrictlyPositive() && magics.m.isNegative()) { 2778 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0)); 2779 Created->push_back(Q.getNode()); 2780 } 2781 // If d < 0 and m > 0, subtract the numerator. 2782 if (Divisor.isNegative() && magics.m.isStrictlyPositive()) { 2783 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0)); 2784 Created->push_back(Q.getNode()); 2785 } 2786 auto &DL = DAG.getDataLayout(); 2787 // Shift right algebraic if shift value is nonzero 2788 if (magics.s > 0) { 2789 Q = DAG.getNode( 2790 ISD::SRA, dl, VT, Q, 2791 DAG.getConstant(magics.s, dl, getShiftAmountTy(Q.getValueType(), DL))); 2792 Created->push_back(Q.getNode()); 2793 } 2794 // Extract the sign bit and add it to the quotient 2795 SDValue T = 2796 DAG.getNode(ISD::SRL, dl, VT, Q, 2797 DAG.getConstant(VT.getScalarSizeInBits() - 1, dl, 2798 getShiftAmountTy(Q.getValueType(), DL))); 2799 Created->push_back(T.getNode()); 2800 return DAG.getNode(ISD::ADD, dl, VT, Q, T); 2801} 2802 2803/// \brief Given an ISD::UDIV node expressing a divide by constant, 2804/// return a DAG expression to select that will generate the same value by 2805/// multiplying by a magic number. 2806/// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide". 2807SDValue TargetLowering::BuildUDIV(SDNode *N, const APInt &Divisor, 2808 SelectionDAG &DAG, bool IsAfterLegalization, 2809 std::vector<SDNode *> *Created) const { 2810 assert(Created && "No vector to hold udiv ops."); 2811 2812 EVT VT = N->getValueType(0); 2813 SDLoc dl(N); 2814 auto &DL = DAG.getDataLayout(); 2815 2816 // Check to see if we can do this. 2817 // FIXME: We should be more aggressive here. 2818 if (!isTypeLegal(VT)) 2819 return SDValue(); 2820 2821 // FIXME: We should use a narrower constant when the upper 2822 // bits are known to be zero. 2823 APInt::mu magics = Divisor.magicu(); 2824 2825 SDValue Q = N->getOperand(0); 2826 2827 // If the divisor is even, we can avoid using the expensive fixup by shifting 2828 // the divided value upfront. 2829 if (magics.a != 0 && !Divisor[0]) { 2830 unsigned Shift = Divisor.countTrailingZeros(); 2831 Q = DAG.getNode( 2832 ISD::SRL, dl, VT, Q, 2833 DAG.getConstant(Shift, dl, getShiftAmountTy(Q.getValueType(), DL))); 2834 Created->push_back(Q.getNode()); 2835 2836 // Get magic number for the shifted divisor. 2837 magics = Divisor.lshr(Shift).magicu(Shift); 2838 assert(magics.a == 0 && "Should use cheap fixup now"); 2839 } 2840 2841 // Multiply the numerator (operand 0) by the magic value 2842 // FIXME: We should support doing a MUL in a wider type 2843 if (IsAfterLegalization ? isOperationLegal(ISD::MULHU, VT) : 2844 isOperationLegalOrCustom(ISD::MULHU, VT)) 2845 Q = DAG.getNode(ISD::MULHU, dl, VT, Q, DAG.getConstant(magics.m, dl, VT)); 2846 else if (IsAfterLegalization ? isOperationLegal(ISD::UMUL_LOHI, VT) : 2847 isOperationLegalOrCustom(ISD::UMUL_LOHI, VT)) 2848 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT), Q, 2849 DAG.getConstant(magics.m, dl, VT)).getNode(), 1); 2850 else 2851 return SDValue(); // No mulhu or equvialent 2852 2853 Created->push_back(Q.getNode()); 2854 2855 if (magics.a == 0) { 2856 assert(magics.s < Divisor.getBitWidth() && 2857 "We shouldn't generate an undefined shift!"); 2858 return DAG.getNode( 2859 ISD::SRL, dl, VT, Q, 2860 DAG.getConstant(magics.s, dl, getShiftAmountTy(Q.getValueType(), DL))); 2861 } else { 2862 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q); 2863 Created->push_back(NPQ.getNode()); 2864 NPQ = DAG.getNode( 2865 ISD::SRL, dl, VT, NPQ, 2866 DAG.getConstant(1, dl, getShiftAmountTy(NPQ.getValueType(), DL))); 2867 Created->push_back(NPQ.getNode()); 2868 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q); 2869 Created->push_back(NPQ.getNode()); 2870 return DAG.getNode( 2871 ISD::SRL, dl, VT, NPQ, 2872 DAG.getConstant(magics.s - 1, dl, 2873 getShiftAmountTy(NPQ.getValueType(), DL))); 2874 } 2875} 2876 2877bool TargetLowering:: 2878verifyReturnAddressArgumentIsConstant(SDValue Op, SelectionDAG &DAG) const { 2879 if (!isa<ConstantSDNode>(Op.getOperand(0))) { 2880 DAG.getContext()->emitError("argument to '__builtin_return_address' must " 2881 "be a constant integer"); 2882 return true; 2883 } 2884 2885 return false; 2886} 2887 2888//===----------------------------------------------------------------------===// 2889// Legalization Utilities 2890//===----------------------------------------------------------------------===// 2891 2892bool TargetLowering::expandMUL(SDNode *N, SDValue &Lo, SDValue &Hi, EVT HiLoVT, 2893 SelectionDAG &DAG, SDValue LL, SDValue LH, 2894 SDValue RL, SDValue RH) const { 2895 EVT VT = N->getValueType(0); 2896 SDLoc dl(N); 2897 2898 bool HasMULHS = isOperationLegalOrCustom(ISD::MULHS, HiLoVT); 2899 bool HasMULHU = isOperationLegalOrCustom(ISD::MULHU, HiLoVT); 2900 bool HasSMUL_LOHI = isOperationLegalOrCustom(ISD::SMUL_LOHI, HiLoVT); 2901 bool HasUMUL_LOHI = isOperationLegalOrCustom(ISD::UMUL_LOHI, HiLoVT); 2902 if (HasMULHU || HasMULHS || HasUMUL_LOHI || HasSMUL_LOHI) { 2903 unsigned OuterBitSize = VT.getSizeInBits(); 2904 unsigned InnerBitSize = HiLoVT.getSizeInBits(); 2905 unsigned LHSSB = DAG.ComputeNumSignBits(N->getOperand(0)); 2906 unsigned RHSSB = DAG.ComputeNumSignBits(N->getOperand(1)); 2907 2908 // LL, LH, RL, and RH must be either all NULL or all set to a value. 2909 assert((LL.getNode() && LH.getNode() && RL.getNode() && RH.getNode()) || 2910 (!LL.getNode() && !LH.getNode() && !RL.getNode() && !RH.getNode())); 2911 2912 if (!LL.getNode() && !RL.getNode() && 2913 isOperationLegalOrCustom(ISD::TRUNCATE, HiLoVT)) { 2914 LL = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, N->getOperand(0)); 2915 RL = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, N->getOperand(1)); 2916 } 2917 2918 if (!LL.getNode()) 2919 return false; 2920 2921 APInt HighMask = APInt::getHighBitsSet(OuterBitSize, InnerBitSize); 2922 if (DAG.MaskedValueIsZero(N->getOperand(0), HighMask) && 2923 DAG.MaskedValueIsZero(N->getOperand(1), HighMask)) { 2924 // The inputs are both zero-extended. 2925 if (HasUMUL_LOHI) { 2926 // We can emit a umul_lohi. 2927 Lo = DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(HiLoVT, HiLoVT), LL, 2928 RL); 2929 Hi = SDValue(Lo.getNode(), 1); 2930 return true; 2931 } 2932 if (HasMULHU) { 2933 // We can emit a mulhu+mul. 2934 Lo = DAG.getNode(ISD::MUL, dl, HiLoVT, LL, RL); 2935 Hi = DAG.getNode(ISD::MULHU, dl, HiLoVT, LL, RL); 2936 return true; 2937 } 2938 } 2939 if (LHSSB > InnerBitSize && RHSSB > InnerBitSize) { 2940 // The input values are both sign-extended. 2941 if (HasSMUL_LOHI) { 2942 // We can emit a smul_lohi. 2943 Lo = DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(HiLoVT, HiLoVT), LL, 2944 RL); 2945 Hi = SDValue(Lo.getNode(), 1); 2946 return true; 2947 } 2948 if (HasMULHS) { 2949 // We can emit a mulhs+mul. 2950 Lo = DAG.getNode(ISD::MUL, dl, HiLoVT, LL, RL); 2951 Hi = DAG.getNode(ISD::MULHS, dl, HiLoVT, LL, RL); 2952 return true; 2953 } 2954 } 2955 2956 if (!LH.getNode() && !RH.getNode() && 2957 isOperationLegalOrCustom(ISD::SRL, VT) && 2958 isOperationLegalOrCustom(ISD::TRUNCATE, HiLoVT)) { 2959 auto &DL = DAG.getDataLayout(); 2960 unsigned ShiftAmt = VT.getSizeInBits() - HiLoVT.getSizeInBits(); 2961 SDValue Shift = DAG.getConstant(ShiftAmt, dl, getShiftAmountTy(VT, DL)); 2962 LH = DAG.getNode(ISD::SRL, dl, VT, N->getOperand(0), Shift); 2963 LH = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, LH); 2964 RH = DAG.getNode(ISD::SRL, dl, VT, N->getOperand(1), Shift); 2965 RH = DAG.getNode(ISD::TRUNCATE, dl, HiLoVT, RH); 2966 } 2967 2968 if (!LH.getNode()) 2969 return false; 2970 2971 if (HasUMUL_LOHI) { 2972 // Lo,Hi = umul LHS, RHS. 2973 SDValue UMulLOHI = DAG.getNode(ISD::UMUL_LOHI, dl, 2974 DAG.getVTList(HiLoVT, HiLoVT), LL, RL); 2975 Lo = UMulLOHI; 2976 Hi = UMulLOHI.getValue(1); 2977 RH = DAG.getNode(ISD::MUL, dl, HiLoVT, LL, RH); 2978 LH = DAG.getNode(ISD::MUL, dl, HiLoVT, LH, RL); 2979 Hi = DAG.getNode(ISD::ADD, dl, HiLoVT, Hi, RH); 2980 Hi = DAG.getNode(ISD::ADD, dl, HiLoVT, Hi, LH); 2981 return true; 2982 } 2983 if (HasMULHU) { 2984 Lo = DAG.getNode(ISD::MUL, dl, HiLoVT, LL, RL); 2985 Hi = DAG.getNode(ISD::MULHU, dl, HiLoVT, LL, RL); 2986 RH = DAG.getNode(ISD::MUL, dl, HiLoVT, LL, RH); 2987 LH = DAG.getNode(ISD::MUL, dl, HiLoVT, LH, RL); 2988 Hi = DAG.getNode(ISD::ADD, dl, HiLoVT, Hi, RH); 2989 Hi = DAG.getNode(ISD::ADD, dl, HiLoVT, Hi, LH); 2990 return true; 2991 } 2992 } 2993 return false; 2994} 2995 2996bool TargetLowering::expandFP_TO_SINT(SDNode *Node, SDValue &Result, 2997 SelectionDAG &DAG) const { 2998 EVT VT = Node->getOperand(0).getValueType(); 2999 EVT NVT = Node->getValueType(0); 3000 SDLoc dl(SDValue(Node, 0)); 3001 3002 // FIXME: Only f32 to i64 conversions are supported. 3003 if (VT != MVT::f32 || NVT != MVT::i64) 3004 return false; 3005 3006 // Expand f32 -> i64 conversion 3007 // This algorithm comes from compiler-rt's implementation of fixsfdi: 3008 // https://github.com/llvm-mirror/compiler-rt/blob/master/lib/builtins/fixsfdi.c 3009 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), 3010 VT.getSizeInBits()); 3011 SDValue ExponentMask = DAG.getConstant(0x7F800000, dl, IntVT); 3012 SDValue ExponentLoBit = DAG.getConstant(23, dl, IntVT); 3013 SDValue Bias = DAG.getConstant(127, dl, IntVT); 3014 SDValue SignMask = DAG.getConstant(APInt::getSignBit(VT.getSizeInBits()), dl, 3015 IntVT); 3016 SDValue SignLowBit = DAG.getConstant(VT.getSizeInBits() - 1, dl, IntVT); 3017 SDValue MantissaMask = DAG.getConstant(0x007FFFFF, dl, IntVT); 3018 3019 SDValue Bits = DAG.getNode(ISD::BITCAST, dl, IntVT, Node->getOperand(0)); 3020 3021 auto &DL = DAG.getDataLayout(); 3022 SDValue ExponentBits = DAG.getNode( 3023 ISD::SRL, dl, IntVT, DAG.getNode(ISD::AND, dl, IntVT, Bits, ExponentMask), 3024 DAG.getZExtOrTrunc(ExponentLoBit, dl, getShiftAmountTy(IntVT, DL))); 3025 SDValue Exponent = DAG.getNode(ISD::SUB, dl, IntVT, ExponentBits, Bias); 3026 3027 SDValue Sign = DAG.getNode( 3028 ISD::SRA, dl, IntVT, DAG.getNode(ISD::AND, dl, IntVT, Bits, SignMask), 3029 DAG.getZExtOrTrunc(SignLowBit, dl, getShiftAmountTy(IntVT, DL))); 3030 Sign = DAG.getSExtOrTrunc(Sign, dl, NVT); 3031 3032 SDValue R = DAG.getNode(ISD::OR, dl, IntVT, 3033 DAG.getNode(ISD::AND, dl, IntVT, Bits, MantissaMask), 3034 DAG.getConstant(0x00800000, dl, IntVT)); 3035 3036 R = DAG.getZExtOrTrunc(R, dl, NVT); 3037 3038 R = DAG.getSelectCC( 3039 dl, Exponent, ExponentLoBit, 3040 DAG.getNode(ISD::SHL, dl, NVT, R, 3041 DAG.getZExtOrTrunc( 3042 DAG.getNode(ISD::SUB, dl, IntVT, Exponent, ExponentLoBit), 3043 dl, getShiftAmountTy(IntVT, DL))), 3044 DAG.getNode(ISD::SRL, dl, NVT, R, 3045 DAG.getZExtOrTrunc( 3046 DAG.getNode(ISD::SUB, dl, IntVT, ExponentLoBit, Exponent), 3047 dl, getShiftAmountTy(IntVT, DL))), 3048 ISD::SETGT); 3049 3050 SDValue Ret = DAG.getNode(ISD::SUB, dl, NVT, 3051 DAG.getNode(ISD::XOR, dl, NVT, R, Sign), 3052 Sign); 3053 3054 Result = DAG.getSelectCC(dl, Exponent, DAG.getConstant(0, dl, IntVT), 3055 DAG.getConstant(0, dl, NVT), Ret, ISD::SETLT); 3056 return true; 3057} 3058 3059//===----------------------------------------------------------------------===// 3060// Implementation of Emulated TLS Model 3061//===----------------------------------------------------------------------===// 3062 3063SDValue TargetLowering::LowerToTLSEmulatedModel(const GlobalAddressSDNode *GA, 3064 SelectionDAG &DAG) const { 3065 // Access to address of TLS varialbe xyz is lowered to a function call: 3066 // __emutls_get_address( address of global variable named "__emutls_v.xyz" ) 3067 EVT PtrVT = getPointerTy(DAG.getDataLayout()); 3068 PointerType *VoidPtrType = Type::getInt8PtrTy(*DAG.getContext()); 3069 SDLoc dl(GA); 3070 3071 ArgListTy Args; 3072 ArgListEntry Entry; 3073 std::string NameString = ("__emutls_v." + GA->getGlobal()->getName()).str(); 3074 Module *VariableModule = const_cast<Module*>(GA->getGlobal()->getParent()); 3075 StringRef EmuTlsVarName(NameString); 3076 GlobalVariable *EmuTlsVar = VariableModule->getNamedGlobal(EmuTlsVarName); 3077 assert(EmuTlsVar && "Cannot find EmuTlsVar "); 3078 Entry.Node = DAG.getGlobalAddress(EmuTlsVar, dl, PtrVT); 3079 Entry.Ty = VoidPtrType; 3080 Args.push_back(Entry); 3081 3082 SDValue EmuTlsGetAddr = DAG.getExternalSymbol("__emutls_get_address", PtrVT); 3083 3084 TargetLowering::CallLoweringInfo CLI(DAG); 3085 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode()); 3086 CLI.setCallee(CallingConv::C, VoidPtrType, EmuTlsGetAddr, std::move(Args), 0); 3087 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI); 3088 3089 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls. 3090 // At last for X86 targets, maybe good for other targets too? 3091 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 3092 MFI->setAdjustsStack(true); // Is this only for X86 target? 3093 MFI->setHasCalls(true); 3094 3095 assert((GA->getOffset() == 0) && 3096 "Emulated TLS must have zero offset in GlobalAddressSDNode"); 3097 return CallResult.first; 3098} 3099