1//===-- LegalizeDAG.cpp - Implement SelectionDAG::Legalize ----------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements the SelectionDAG::Legalize method. 11// 12//===----------------------------------------------------------------------===// 13 14#include "llvm/ADT/SetVector.h" 15#include "llvm/ADT/SmallPtrSet.h" 16#include "llvm/ADT/SmallSet.h" 17#include "llvm/ADT/SmallVector.h" 18#include "llvm/ADT/Triple.h" 19#include "llvm/CodeGen/MachineFunction.h" 20#include "llvm/CodeGen/MachineJumpTableInfo.h" 21#include "llvm/CodeGen/SelectionDAG.h" 22#include "llvm/CodeGen/SelectionDAGNodes.h" 23#include "llvm/IR/CallingConv.h" 24#include "llvm/IR/Constants.h" 25#include "llvm/IR/DataLayout.h" 26#include "llvm/IR/DebugInfo.h" 27#include "llvm/IR/DerivedTypes.h" 28#include "llvm/IR/Function.h" 29#include "llvm/IR/LLVMContext.h" 30#include "llvm/Support/Debug.h" 31#include "llvm/Support/ErrorHandling.h" 32#include "llvm/Support/MathExtras.h" 33#include "llvm/Support/raw_ostream.h" 34#include "llvm/Target/TargetFrameLowering.h" 35#include "llvm/Target/TargetLowering.h" 36#include "llvm/Target/TargetMachine.h" 37#include "llvm/Target/TargetSubtargetInfo.h" 38using namespace llvm; 39 40#define DEBUG_TYPE "legalizedag" 41 42namespace { 43 44struct FloatSignAsInt; 45 46//===----------------------------------------------------------------------===// 47/// This takes an arbitrary SelectionDAG as input and 48/// hacks on it until the target machine can handle it. This involves 49/// eliminating value sizes the machine cannot handle (promoting small sizes to 50/// large sizes or splitting up large values into small values) as well as 51/// eliminating operations the machine cannot handle. 52/// 53/// This code also does a small amount of optimization and recognition of idioms 54/// as part of its processing. For example, if a target does not support a 55/// 'setcc' instruction efficiently, but does support 'brcc' instruction, this 56/// will attempt merge setcc and brc instructions into brcc's. 57/// 58class SelectionDAGLegalize { 59 const TargetMachine &TM; 60 const TargetLowering &TLI; 61 SelectionDAG &DAG; 62 63 /// \brief The set of nodes which have already been legalized. We hold a 64 /// reference to it in order to update as necessary on node deletion. 65 SmallPtrSetImpl<SDNode *> &LegalizedNodes; 66 67 /// \brief A set of all the nodes updated during legalization. 68 SmallSetVector<SDNode *, 16> *UpdatedNodes; 69 70 EVT getSetCCResultType(EVT VT) const { 71 return TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); 72 } 73 74 // Libcall insertion helpers. 75 76public: 77 SelectionDAGLegalize(SelectionDAG &DAG, 78 SmallPtrSetImpl<SDNode *> &LegalizedNodes, 79 SmallSetVector<SDNode *, 16> *UpdatedNodes = nullptr) 80 : TM(DAG.getTarget()), TLI(DAG.getTargetLoweringInfo()), DAG(DAG), 81 LegalizedNodes(LegalizedNodes), UpdatedNodes(UpdatedNodes) {} 82 83 /// \brief Legalizes the given operation. 84 void LegalizeOp(SDNode *Node); 85 86private: 87 SDValue OptimizeFloatStore(StoreSDNode *ST); 88 89 void LegalizeLoadOps(SDNode *Node); 90 void LegalizeStoreOps(SDNode *Node); 91 92 /// Some targets cannot handle a variable 93 /// insertion index for the INSERT_VECTOR_ELT instruction. In this case, it 94 /// is necessary to spill the vector being inserted into to memory, perform 95 /// the insert there, and then read the result back. 96 SDValue PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val, SDValue Idx, 97 const SDLoc &dl); 98 SDValue ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val, SDValue Idx, 99 const SDLoc &dl); 100 101 /// Return a vector shuffle operation which 102 /// performs the same shuffe in terms of order or result bytes, but on a type 103 /// whose vector element type is narrower than the original shuffle type. 104 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3> 105 SDValue ShuffleWithNarrowerEltType(EVT NVT, EVT VT, const SDLoc &dl, 106 SDValue N1, SDValue N2, 107 ArrayRef<int> Mask) const; 108 109 bool LegalizeSetCCCondCode(EVT VT, SDValue &LHS, SDValue &RHS, SDValue &CC, 110 bool &NeedInvert, const SDLoc &dl); 111 112 SDValue ExpandLibCall(RTLIB::Libcall LC, SDNode *Node, bool isSigned); 113 SDValue ExpandLibCall(RTLIB::Libcall LC, EVT RetVT, const SDValue *Ops, 114 unsigned NumOps, bool isSigned, const SDLoc &dl); 115 116 std::pair<SDValue, SDValue> ExpandChainLibCall(RTLIB::Libcall LC, 117 SDNode *Node, bool isSigned); 118 SDValue ExpandFPLibCall(SDNode *Node, RTLIB::Libcall Call_F32, 119 RTLIB::Libcall Call_F64, RTLIB::Libcall Call_F80, 120 RTLIB::Libcall Call_F128, 121 RTLIB::Libcall Call_PPCF128); 122 SDValue ExpandIntLibCall(SDNode *Node, bool isSigned, 123 RTLIB::Libcall Call_I8, 124 RTLIB::Libcall Call_I16, 125 RTLIB::Libcall Call_I32, 126 RTLIB::Libcall Call_I64, 127 RTLIB::Libcall Call_I128); 128 void ExpandDivRemLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results); 129 void ExpandSinCosLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results); 130 131 SDValue EmitStackConvert(SDValue SrcOp, EVT SlotVT, EVT DestVT, 132 const SDLoc &dl); 133 SDValue ExpandBUILD_VECTOR(SDNode *Node); 134 SDValue ExpandSCALAR_TO_VECTOR(SDNode *Node); 135 void ExpandDYNAMIC_STACKALLOC(SDNode *Node, 136 SmallVectorImpl<SDValue> &Results); 137 void getSignAsIntValue(FloatSignAsInt &State, const SDLoc &DL, 138 SDValue Value) const; 139 SDValue modifySignAsInt(const FloatSignAsInt &State, const SDLoc &DL, 140 SDValue NewIntValue) const; 141 SDValue ExpandFCOPYSIGN(SDNode *Node) const; 142 SDValue ExpandFABS(SDNode *Node) const; 143 SDValue ExpandLegalINT_TO_FP(bool isSigned, SDValue LegalOp, EVT DestVT, 144 const SDLoc &dl); 145 SDValue PromoteLegalINT_TO_FP(SDValue LegalOp, EVT DestVT, bool isSigned, 146 const SDLoc &dl); 147 SDValue PromoteLegalFP_TO_INT(SDValue LegalOp, EVT DestVT, bool isSigned, 148 const SDLoc &dl); 149 150 SDValue ExpandBITREVERSE(SDValue Op, const SDLoc &dl); 151 SDValue ExpandBSWAP(SDValue Op, const SDLoc &dl); 152 SDValue ExpandBitCount(unsigned Opc, SDValue Op, const SDLoc &dl); 153 154 SDValue ExpandExtractFromVectorThroughStack(SDValue Op); 155 SDValue ExpandInsertToVectorThroughStack(SDValue Op); 156 SDValue ExpandVectorBuildThroughStack(SDNode* Node); 157 158 SDValue ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP); 159 SDValue ExpandConstant(ConstantSDNode *CP); 160 161 // if ExpandNode returns false, LegalizeOp falls back to ConvertNodeToLibcall 162 bool ExpandNode(SDNode *Node); 163 void ConvertNodeToLibcall(SDNode *Node); 164 void PromoteNode(SDNode *Node); 165 166public: 167 // Node replacement helpers 168 void ReplacedNode(SDNode *N) { 169 LegalizedNodes.erase(N); 170 if (UpdatedNodes) 171 UpdatedNodes->insert(N); 172 } 173 void ReplaceNode(SDNode *Old, SDNode *New) { 174 DEBUG(dbgs() << " ... replacing: "; Old->dump(&DAG); 175 dbgs() << " with: "; New->dump(&DAG)); 176 177 assert(Old->getNumValues() == New->getNumValues() && 178 "Replacing one node with another that produces a different number " 179 "of values!"); 180 DAG.ReplaceAllUsesWith(Old, New); 181 if (UpdatedNodes) 182 UpdatedNodes->insert(New); 183 ReplacedNode(Old); 184 } 185 void ReplaceNode(SDValue Old, SDValue New) { 186 DEBUG(dbgs() << " ... replacing: "; Old->dump(&DAG); 187 dbgs() << " with: "; New->dump(&DAG)); 188 189 DAG.ReplaceAllUsesWith(Old, New); 190 if (UpdatedNodes) 191 UpdatedNodes->insert(New.getNode()); 192 ReplacedNode(Old.getNode()); 193 } 194 void ReplaceNode(SDNode *Old, const SDValue *New) { 195 DEBUG(dbgs() << " ... replacing: "; Old->dump(&DAG)); 196 197 DAG.ReplaceAllUsesWith(Old, New); 198 for (unsigned i = 0, e = Old->getNumValues(); i != e; ++i) { 199 DEBUG(dbgs() << (i == 0 ? " with: " 200 : " and: "); 201 New[i]->dump(&DAG)); 202 if (UpdatedNodes) 203 UpdatedNodes->insert(New[i].getNode()); 204 } 205 ReplacedNode(Old); 206 } 207}; 208} 209 210/// Return a vector shuffle operation which 211/// performs the same shuffe in terms of order or result bytes, but on a type 212/// whose vector element type is narrower than the original shuffle type. 213/// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3> 214SDValue SelectionDAGLegalize::ShuffleWithNarrowerEltType( 215 EVT NVT, EVT VT, const SDLoc &dl, SDValue N1, SDValue N2, 216 ArrayRef<int> Mask) const { 217 unsigned NumMaskElts = VT.getVectorNumElements(); 218 unsigned NumDestElts = NVT.getVectorNumElements(); 219 unsigned NumEltsGrowth = NumDestElts / NumMaskElts; 220 221 assert(NumEltsGrowth && "Cannot promote to vector type with fewer elts!"); 222 223 if (NumEltsGrowth == 1) 224 return DAG.getVectorShuffle(NVT, dl, N1, N2, Mask); 225 226 SmallVector<int, 8> NewMask; 227 for (unsigned i = 0; i != NumMaskElts; ++i) { 228 int Idx = Mask[i]; 229 for (unsigned j = 0; j != NumEltsGrowth; ++j) { 230 if (Idx < 0) 231 NewMask.push_back(-1); 232 else 233 NewMask.push_back(Idx * NumEltsGrowth + j); 234 } 235 } 236 assert(NewMask.size() == NumDestElts && "Non-integer NumEltsGrowth?"); 237 assert(TLI.isShuffleMaskLegal(NewMask, NVT) && "Shuffle not legal?"); 238 return DAG.getVectorShuffle(NVT, dl, N1, N2, NewMask); 239} 240 241/// Expands the ConstantFP node to an integer constant or 242/// a load from the constant pool. 243SDValue 244SelectionDAGLegalize::ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP) { 245 bool Extend = false; 246 SDLoc dl(CFP); 247 248 // If a FP immediate is precise when represented as a float and if the 249 // target can do an extending load from float to double, we put it into 250 // the constant pool as a float, even if it's is statically typed as a 251 // double. This shrinks FP constants and canonicalizes them for targets where 252 // an FP extending load is the same cost as a normal load (such as on the x87 253 // fp stack or PPC FP unit). 254 EVT VT = CFP->getValueType(0); 255 ConstantFP *LLVMC = const_cast<ConstantFP*>(CFP->getConstantFPValue()); 256 if (!UseCP) { 257 assert((VT == MVT::f64 || VT == MVT::f32) && "Invalid type expansion"); 258 return DAG.getConstant(LLVMC->getValueAPF().bitcastToAPInt(), dl, 259 (VT == MVT::f64) ? MVT::i64 : MVT::i32); 260 } 261 262 EVT OrigVT = VT; 263 EVT SVT = VT; 264 while (SVT != MVT::f32 && SVT != MVT::f16) { 265 SVT = (MVT::SimpleValueType)(SVT.getSimpleVT().SimpleTy - 1); 266 if (ConstantFPSDNode::isValueValidForType(SVT, CFP->getValueAPF()) && 267 // Only do this if the target has a native EXTLOAD instruction from 268 // smaller type. 269 TLI.isLoadExtLegal(ISD::EXTLOAD, OrigVT, SVT) && 270 TLI.ShouldShrinkFPConstant(OrigVT)) { 271 Type *SType = SVT.getTypeForEVT(*DAG.getContext()); 272 LLVMC = cast<ConstantFP>(ConstantExpr::getFPTrunc(LLVMC, SType)); 273 VT = SVT; 274 Extend = true; 275 } 276 } 277 278 SDValue CPIdx = 279 DAG.getConstantPool(LLVMC, TLI.getPointerTy(DAG.getDataLayout())); 280 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment(); 281 if (Extend) { 282 SDValue Result = DAG.getExtLoad( 283 ISD::EXTLOAD, dl, OrigVT, DAG.getEntryNode(), CPIdx, 284 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), VT, 285 false, false, false, Alignment); 286 return Result; 287 } 288 SDValue Result = 289 DAG.getLoad(OrigVT, dl, DAG.getEntryNode(), CPIdx, 290 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), 291 false, false, false, Alignment); 292 return Result; 293} 294 295/// Expands the Constant node to a load from the constant pool. 296SDValue SelectionDAGLegalize::ExpandConstant(ConstantSDNode *CP) { 297 SDLoc dl(CP); 298 EVT VT = CP->getValueType(0); 299 SDValue CPIdx = DAG.getConstantPool(CP->getConstantIntValue(), 300 TLI.getPointerTy(DAG.getDataLayout())); 301 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment(); 302 SDValue Result = 303 DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx, 304 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), 305 false, false, false, Alignment); 306 return Result; 307} 308 309/// Some target cannot handle a variable insertion index for the 310/// INSERT_VECTOR_ELT instruction. In this case, it 311/// is necessary to spill the vector being inserted into to memory, perform 312/// the insert there, and then read the result back. 313SDValue SelectionDAGLegalize::PerformInsertVectorEltInMemory(SDValue Vec, 314 SDValue Val, 315 SDValue Idx, 316 const SDLoc &dl) { 317 SDValue Tmp1 = Vec; 318 SDValue Tmp2 = Val; 319 SDValue Tmp3 = Idx; 320 321 // If the target doesn't support this, we have to spill the input vector 322 // to a temporary stack slot, update the element, then reload it. This is 323 // badness. We could also load the value into a vector register (either 324 // with a "move to register" or "extload into register" instruction, then 325 // permute it into place, if the idx is a constant and if the idx is 326 // supported by the target. 327 EVT VT = Tmp1.getValueType(); 328 EVT EltVT = VT.getVectorElementType(); 329 EVT IdxVT = Tmp3.getValueType(); 330 EVT PtrVT = TLI.getPointerTy(DAG.getDataLayout()); 331 SDValue StackPtr = DAG.CreateStackTemporary(VT); 332 333 int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex(); 334 335 // Store the vector. 336 SDValue Ch = DAG.getStore( 337 DAG.getEntryNode(), dl, Tmp1, StackPtr, 338 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SPFI), false, 339 false, 0); 340 341 // Truncate or zero extend offset to target pointer type. 342 Tmp3 = DAG.getZExtOrTrunc(Tmp3, dl, PtrVT); 343 // Add the offset to the index. 344 unsigned EltSize = EltVT.getSizeInBits()/8; 345 Tmp3 = DAG.getNode(ISD::MUL, dl, IdxVT, Tmp3, 346 DAG.getConstant(EltSize, dl, IdxVT)); 347 SDValue StackPtr2 = DAG.getNode(ISD::ADD, dl, IdxVT, Tmp3, StackPtr); 348 // Store the scalar value. 349 Ch = DAG.getTruncStore(Ch, dl, Tmp2, StackPtr2, MachinePointerInfo(), EltVT, 350 false, false, 0); 351 // Load the updated vector. 352 return DAG.getLoad(VT, dl, Ch, StackPtr, MachinePointerInfo::getFixedStack( 353 DAG.getMachineFunction(), SPFI), 354 false, false, false, 0); 355} 356 357SDValue SelectionDAGLegalize::ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val, 358 SDValue Idx, 359 const SDLoc &dl) { 360 if (ConstantSDNode *InsertPos = dyn_cast<ConstantSDNode>(Idx)) { 361 // SCALAR_TO_VECTOR requires that the type of the value being inserted 362 // match the element type of the vector being created, except for 363 // integers in which case the inserted value can be over width. 364 EVT EltVT = Vec.getValueType().getVectorElementType(); 365 if (Val.getValueType() == EltVT || 366 (EltVT.isInteger() && Val.getValueType().bitsGE(EltVT))) { 367 SDValue ScVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, 368 Vec.getValueType(), Val); 369 370 unsigned NumElts = Vec.getValueType().getVectorNumElements(); 371 // We generate a shuffle of InVec and ScVec, so the shuffle mask 372 // should be 0,1,2,3,4,5... with the appropriate element replaced with 373 // elt 0 of the RHS. 374 SmallVector<int, 8> ShufOps; 375 for (unsigned i = 0; i != NumElts; ++i) 376 ShufOps.push_back(i != InsertPos->getZExtValue() ? i : NumElts); 377 378 return DAG.getVectorShuffle(Vec.getValueType(), dl, Vec, ScVec, ShufOps); 379 } 380 } 381 return PerformInsertVectorEltInMemory(Vec, Val, Idx, dl); 382} 383 384SDValue SelectionDAGLegalize::OptimizeFloatStore(StoreSDNode* ST) { 385 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr' 386 // FIXME: We shouldn't do this for TargetConstantFP's. 387 // FIXME: move this to the DAG Combiner! Note that we can't regress due 388 // to phase ordering between legalized code and the dag combiner. This 389 // probably means that we need to integrate dag combiner and legalizer 390 // together. 391 // We generally can't do this one for long doubles. 392 SDValue Chain = ST->getChain(); 393 SDValue Ptr = ST->getBasePtr(); 394 unsigned Alignment = ST->getAlignment(); 395 bool isVolatile = ST->isVolatile(); 396 bool isNonTemporal = ST->isNonTemporal(); 397 AAMDNodes AAInfo = ST->getAAInfo(); 398 SDLoc dl(ST); 399 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(ST->getValue())) { 400 if (CFP->getValueType(0) == MVT::f32 && 401 TLI.isTypeLegal(MVT::i32)) { 402 SDValue Con = DAG.getConstant(CFP->getValueAPF(). 403 bitcastToAPInt().zextOrTrunc(32), 404 SDLoc(CFP), MVT::i32); 405 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(), 406 isVolatile, isNonTemporal, Alignment, AAInfo); 407 } 408 409 if (CFP->getValueType(0) == MVT::f64) { 410 // If this target supports 64-bit registers, do a single 64-bit store. 411 if (TLI.isTypeLegal(MVT::i64)) { 412 SDValue Con = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt(). 413 zextOrTrunc(64), SDLoc(CFP), MVT::i64); 414 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(), 415 isVolatile, isNonTemporal, Alignment, AAInfo); 416 } 417 418 if (TLI.isTypeLegal(MVT::i32) && !ST->isVolatile()) { 419 // Otherwise, if the target supports 32-bit registers, use 2 32-bit 420 // stores. If the target supports neither 32- nor 64-bits, this 421 // xform is certainly not worth it. 422 const APInt &IntVal = CFP->getValueAPF().bitcastToAPInt(); 423 SDValue Lo = DAG.getConstant(IntVal.trunc(32), dl, MVT::i32); 424 SDValue Hi = DAG.getConstant(IntVal.lshr(32).trunc(32), dl, MVT::i32); 425 if (DAG.getDataLayout().isBigEndian()) 426 std::swap(Lo, Hi); 427 428 Lo = DAG.getStore(Chain, dl, Lo, Ptr, ST->getPointerInfo(), isVolatile, 429 isNonTemporal, Alignment, AAInfo); 430 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, 431 DAG.getConstant(4, dl, Ptr.getValueType())); 432 Hi = DAG.getStore(Chain, dl, Hi, Ptr, 433 ST->getPointerInfo().getWithOffset(4), 434 isVolatile, isNonTemporal, MinAlign(Alignment, 4U), 435 AAInfo); 436 437 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi); 438 } 439 } 440 } 441 return SDValue(nullptr, 0); 442} 443 444void SelectionDAGLegalize::LegalizeStoreOps(SDNode *Node) { 445 StoreSDNode *ST = cast<StoreSDNode>(Node); 446 SDValue Chain = ST->getChain(); 447 SDValue Ptr = ST->getBasePtr(); 448 SDLoc dl(Node); 449 450 unsigned Alignment = ST->getAlignment(); 451 bool isVolatile = ST->isVolatile(); 452 bool isNonTemporal = ST->isNonTemporal(); 453 AAMDNodes AAInfo = ST->getAAInfo(); 454 455 if (!ST->isTruncatingStore()) { 456 if (SDNode *OptStore = OptimizeFloatStore(ST).getNode()) { 457 ReplaceNode(ST, OptStore); 458 return; 459 } 460 461 { 462 SDValue Value = ST->getValue(); 463 MVT VT = Value.getSimpleValueType(); 464 switch (TLI.getOperationAction(ISD::STORE, VT)) { 465 default: llvm_unreachable("This action is not supported yet!"); 466 case TargetLowering::Legal: { 467 // If this is an unaligned store and the target doesn't support it, 468 // expand it. 469 EVT MemVT = ST->getMemoryVT(); 470 unsigned AS = ST->getAddressSpace(); 471 unsigned Align = ST->getAlignment(); 472 const DataLayout &DL = DAG.getDataLayout(); 473 if (!TLI.allowsMemoryAccess(*DAG.getContext(), DL, MemVT, AS, Align)) { 474 SDValue Result = TLI.expandUnalignedStore(ST, DAG); 475 ReplaceNode(SDValue(ST, 0), Result); 476 } 477 break; 478 } 479 case TargetLowering::Custom: { 480 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG); 481 if (Res && Res != SDValue(Node, 0)) 482 ReplaceNode(SDValue(Node, 0), Res); 483 return; 484 } 485 case TargetLowering::Promote: { 486 MVT NVT = TLI.getTypeToPromoteTo(ISD::STORE, VT); 487 assert(NVT.getSizeInBits() == VT.getSizeInBits() && 488 "Can only promote stores to same size type"); 489 Value = DAG.getNode(ISD::BITCAST, dl, NVT, Value); 490 SDValue Result = 491 DAG.getStore(Chain, dl, Value, Ptr, 492 ST->getPointerInfo(), isVolatile, 493 isNonTemporal, Alignment, AAInfo); 494 ReplaceNode(SDValue(Node, 0), Result); 495 break; 496 } 497 } 498 return; 499 } 500 } else { 501 SDValue Value = ST->getValue(); 502 503 EVT StVT = ST->getMemoryVT(); 504 unsigned StWidth = StVT.getSizeInBits(); 505 auto &DL = DAG.getDataLayout(); 506 507 if (StWidth != StVT.getStoreSizeInBits()) { 508 // Promote to a byte-sized store with upper bits zero if not 509 // storing an integral number of bytes. For example, promote 510 // TRUNCSTORE:i1 X -> TRUNCSTORE:i8 (and X, 1) 511 EVT NVT = EVT::getIntegerVT(*DAG.getContext(), 512 StVT.getStoreSizeInBits()); 513 Value = DAG.getZeroExtendInReg(Value, dl, StVT); 514 SDValue Result = 515 DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(), 516 NVT, isVolatile, isNonTemporal, Alignment, AAInfo); 517 ReplaceNode(SDValue(Node, 0), Result); 518 } else if (StWidth & (StWidth - 1)) { 519 // If not storing a power-of-2 number of bits, expand as two stores. 520 assert(!StVT.isVector() && "Unsupported truncstore!"); 521 unsigned RoundWidth = 1 << Log2_32(StWidth); 522 assert(RoundWidth < StWidth); 523 unsigned ExtraWidth = StWidth - RoundWidth; 524 assert(ExtraWidth < RoundWidth); 525 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) && 526 "Store size not an integral number of bytes!"); 527 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth); 528 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth); 529 SDValue Lo, Hi; 530 unsigned IncrementSize; 531 532 if (DL.isLittleEndian()) { 533 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 X, TRUNCSTORE@+2:i8 (srl X, 16) 534 // Store the bottom RoundWidth bits. 535 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(), 536 RoundVT, 537 isVolatile, isNonTemporal, Alignment, 538 AAInfo); 539 540 // Store the remaining ExtraWidth bits. 541 IncrementSize = RoundWidth / 8; 542 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, 543 DAG.getConstant(IncrementSize, dl, 544 Ptr.getValueType())); 545 Hi = DAG.getNode( 546 ISD::SRL, dl, Value.getValueType(), Value, 547 DAG.getConstant(RoundWidth, dl, 548 TLI.getShiftAmountTy(Value.getValueType(), DL))); 549 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr, 550 ST->getPointerInfo().getWithOffset(IncrementSize), 551 ExtraVT, isVolatile, isNonTemporal, 552 MinAlign(Alignment, IncrementSize), AAInfo); 553 } else { 554 // Big endian - avoid unaligned stores. 555 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 (srl X, 8), TRUNCSTORE@+2:i8 X 556 // Store the top RoundWidth bits. 557 Hi = DAG.getNode( 558 ISD::SRL, dl, Value.getValueType(), Value, 559 DAG.getConstant(ExtraWidth, dl, 560 TLI.getShiftAmountTy(Value.getValueType(), DL))); 561 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr, ST->getPointerInfo(), 562 RoundVT, isVolatile, isNonTemporal, Alignment, 563 AAInfo); 564 565 // Store the remaining ExtraWidth bits. 566 IncrementSize = RoundWidth / 8; 567 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, 568 DAG.getConstant(IncrementSize, dl, 569 Ptr.getValueType())); 570 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr, 571 ST->getPointerInfo().getWithOffset(IncrementSize), 572 ExtraVT, isVolatile, isNonTemporal, 573 MinAlign(Alignment, IncrementSize), AAInfo); 574 } 575 576 // The order of the stores doesn't matter. 577 SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi); 578 ReplaceNode(SDValue(Node, 0), Result); 579 } else { 580 switch (TLI.getTruncStoreAction(ST->getValue().getValueType(), StVT)) { 581 default: llvm_unreachable("This action is not supported yet!"); 582 case TargetLowering::Legal: { 583 EVT MemVT = ST->getMemoryVT(); 584 unsigned AS = ST->getAddressSpace(); 585 unsigned Align = ST->getAlignment(); 586 // If this is an unaligned store and the target doesn't support it, 587 // expand it. 588 if (!TLI.allowsMemoryAccess(*DAG.getContext(), DL, MemVT, AS, Align)) { 589 SDValue Result = TLI.expandUnalignedStore(ST, DAG); 590 ReplaceNode(SDValue(ST, 0), Result); 591 } 592 break; 593 } 594 case TargetLowering::Custom: { 595 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG); 596 if (Res && Res != SDValue(Node, 0)) 597 ReplaceNode(SDValue(Node, 0), Res); 598 return; 599 } 600 case TargetLowering::Expand: 601 assert(!StVT.isVector() && 602 "Vector Stores are handled in LegalizeVectorOps"); 603 604 // TRUNCSTORE:i16 i32 -> STORE i16 605 assert(TLI.isTypeLegal(StVT) && 606 "Do not know how to expand this store!"); 607 Value = DAG.getNode(ISD::TRUNCATE, dl, StVT, Value); 608 SDValue Result = 609 DAG.getStore(Chain, dl, Value, Ptr, ST->getPointerInfo(), 610 isVolatile, isNonTemporal, Alignment, AAInfo); 611 ReplaceNode(SDValue(Node, 0), Result); 612 break; 613 } 614 } 615 } 616} 617 618void SelectionDAGLegalize::LegalizeLoadOps(SDNode *Node) { 619 LoadSDNode *LD = cast<LoadSDNode>(Node); 620 SDValue Chain = LD->getChain(); // The chain. 621 SDValue Ptr = LD->getBasePtr(); // The base pointer. 622 SDValue Value; // The value returned by the load op. 623 SDLoc dl(Node); 624 625 ISD::LoadExtType ExtType = LD->getExtensionType(); 626 if (ExtType == ISD::NON_EXTLOAD) { 627 MVT VT = Node->getSimpleValueType(0); 628 SDValue RVal = SDValue(Node, 0); 629 SDValue RChain = SDValue(Node, 1); 630 631 switch (TLI.getOperationAction(Node->getOpcode(), VT)) { 632 default: llvm_unreachable("This action is not supported yet!"); 633 case TargetLowering::Legal: { 634 EVT MemVT = LD->getMemoryVT(); 635 unsigned AS = LD->getAddressSpace(); 636 unsigned Align = LD->getAlignment(); 637 const DataLayout &DL = DAG.getDataLayout(); 638 // If this is an unaligned load and the target doesn't support it, 639 // expand it. 640 if (!TLI.allowsMemoryAccess(*DAG.getContext(), DL, MemVT, AS, Align)) { 641 std::tie(RVal, RChain) = TLI.expandUnalignedLoad(LD, DAG); 642 } 643 break; 644 } 645 case TargetLowering::Custom: { 646 if (SDValue Res = TLI.LowerOperation(RVal, DAG)) { 647 RVal = Res; 648 RChain = Res.getValue(1); 649 } 650 break; 651 } 652 case TargetLowering::Promote: { 653 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), VT); 654 assert(NVT.getSizeInBits() == VT.getSizeInBits() && 655 "Can only promote loads to same size type"); 656 657 SDValue Res = DAG.getLoad(NVT, dl, Chain, Ptr, LD->getMemOperand()); 658 RVal = DAG.getNode(ISD::BITCAST, dl, VT, Res); 659 RChain = Res.getValue(1); 660 break; 661 } 662 } 663 if (RChain.getNode() != Node) { 664 assert(RVal.getNode() != Node && "Load must be completely replaced"); 665 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), RVal); 666 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), RChain); 667 if (UpdatedNodes) { 668 UpdatedNodes->insert(RVal.getNode()); 669 UpdatedNodes->insert(RChain.getNode()); 670 } 671 ReplacedNode(Node); 672 } 673 return; 674 } 675 676 EVT SrcVT = LD->getMemoryVT(); 677 unsigned SrcWidth = SrcVT.getSizeInBits(); 678 unsigned Alignment = LD->getAlignment(); 679 bool isVolatile = LD->isVolatile(); 680 bool isNonTemporal = LD->isNonTemporal(); 681 bool isInvariant = LD->isInvariant(); 682 AAMDNodes AAInfo = LD->getAAInfo(); 683 684 if (SrcWidth != SrcVT.getStoreSizeInBits() && 685 // Some targets pretend to have an i1 loading operation, and actually 686 // load an i8. This trick is correct for ZEXTLOAD because the top 7 687 // bits are guaranteed to be zero; it helps the optimizers understand 688 // that these bits are zero. It is also useful for EXTLOAD, since it 689 // tells the optimizers that those bits are undefined. It would be 690 // nice to have an effective generic way of getting these benefits... 691 // Until such a way is found, don't insist on promoting i1 here. 692 (SrcVT != MVT::i1 || 693 TLI.getLoadExtAction(ExtType, Node->getValueType(0), MVT::i1) == 694 TargetLowering::Promote)) { 695 // Promote to a byte-sized load if not loading an integral number of 696 // bytes. For example, promote EXTLOAD:i20 -> EXTLOAD:i24. 697 unsigned NewWidth = SrcVT.getStoreSizeInBits(); 698 EVT NVT = EVT::getIntegerVT(*DAG.getContext(), NewWidth); 699 SDValue Ch; 700 701 // The extra bits are guaranteed to be zero, since we stored them that 702 // way. A zext load from NVT thus automatically gives zext from SrcVT. 703 704 ISD::LoadExtType NewExtType = 705 ExtType == ISD::ZEXTLOAD ? ISD::ZEXTLOAD : ISD::EXTLOAD; 706 707 SDValue Result = 708 DAG.getExtLoad(NewExtType, dl, Node->getValueType(0), 709 Chain, Ptr, LD->getPointerInfo(), 710 NVT, isVolatile, isNonTemporal, isInvariant, Alignment, 711 AAInfo); 712 713 Ch = Result.getValue(1); // The chain. 714 715 if (ExtType == ISD::SEXTLOAD) 716 // Having the top bits zero doesn't help when sign extending. 717 Result = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, 718 Result.getValueType(), 719 Result, DAG.getValueType(SrcVT)); 720 else if (ExtType == ISD::ZEXTLOAD || NVT == Result.getValueType()) 721 // All the top bits are guaranteed to be zero - inform the optimizers. 722 Result = DAG.getNode(ISD::AssertZext, dl, 723 Result.getValueType(), Result, 724 DAG.getValueType(SrcVT)); 725 726 Value = Result; 727 Chain = Ch; 728 } else if (SrcWidth & (SrcWidth - 1)) { 729 // If not loading a power-of-2 number of bits, expand as two loads. 730 assert(!SrcVT.isVector() && "Unsupported extload!"); 731 unsigned RoundWidth = 1 << Log2_32(SrcWidth); 732 assert(RoundWidth < SrcWidth); 733 unsigned ExtraWidth = SrcWidth - RoundWidth; 734 assert(ExtraWidth < RoundWidth); 735 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) && 736 "Load size not an integral number of bytes!"); 737 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth); 738 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth); 739 SDValue Lo, Hi, Ch; 740 unsigned IncrementSize; 741 auto &DL = DAG.getDataLayout(); 742 743 if (DL.isLittleEndian()) { 744 // EXTLOAD:i24 -> ZEXTLOAD:i16 | (shl EXTLOAD@+2:i8, 16) 745 // Load the bottom RoundWidth bits. 746 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0), 747 Chain, Ptr, 748 LD->getPointerInfo(), RoundVT, isVolatile, 749 isNonTemporal, isInvariant, Alignment, AAInfo); 750 751 // Load the remaining ExtraWidth bits. 752 IncrementSize = RoundWidth / 8; 753 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, 754 DAG.getConstant(IncrementSize, dl, 755 Ptr.getValueType())); 756 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr, 757 LD->getPointerInfo().getWithOffset(IncrementSize), 758 ExtraVT, isVolatile, isNonTemporal, isInvariant, 759 MinAlign(Alignment, IncrementSize), AAInfo); 760 761 // Build a factor node to remember that this load is independent of 762 // the other one. 763 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1), 764 Hi.getValue(1)); 765 766 // Move the top bits to the right place. 767 Hi = DAG.getNode( 768 ISD::SHL, dl, Hi.getValueType(), Hi, 769 DAG.getConstant(RoundWidth, dl, 770 TLI.getShiftAmountTy(Hi.getValueType(), DL))); 771 772 // Join the hi and lo parts. 773 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi); 774 } else { 775 // Big endian - avoid unaligned loads. 776 // EXTLOAD:i24 -> (shl EXTLOAD:i16, 8) | ZEXTLOAD@+2:i8 777 // Load the top RoundWidth bits. 778 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr, 779 LD->getPointerInfo(), RoundVT, isVolatile, 780 isNonTemporal, isInvariant, Alignment, AAInfo); 781 782 // Load the remaining ExtraWidth bits. 783 IncrementSize = RoundWidth / 8; 784 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, 785 DAG.getConstant(IncrementSize, dl, 786 Ptr.getValueType())); 787 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, 788 dl, Node->getValueType(0), Chain, Ptr, 789 LD->getPointerInfo().getWithOffset(IncrementSize), 790 ExtraVT, isVolatile, isNonTemporal, isInvariant, 791 MinAlign(Alignment, IncrementSize), AAInfo); 792 793 // Build a factor node to remember that this load is independent of 794 // the other one. 795 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1), 796 Hi.getValue(1)); 797 798 // Move the top bits to the right place. 799 Hi = DAG.getNode( 800 ISD::SHL, dl, Hi.getValueType(), Hi, 801 DAG.getConstant(ExtraWidth, dl, 802 TLI.getShiftAmountTy(Hi.getValueType(), DL))); 803 804 // Join the hi and lo parts. 805 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi); 806 } 807 808 Chain = Ch; 809 } else { 810 bool isCustom = false; 811 switch (TLI.getLoadExtAction(ExtType, Node->getValueType(0), 812 SrcVT.getSimpleVT())) { 813 default: llvm_unreachable("This action is not supported yet!"); 814 case TargetLowering::Custom: 815 isCustom = true; 816 // FALLTHROUGH 817 case TargetLowering::Legal: { 818 Value = SDValue(Node, 0); 819 Chain = SDValue(Node, 1); 820 821 if (isCustom) { 822 if (SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG)) { 823 Value = Res; 824 Chain = Res.getValue(1); 825 } 826 } else { 827 // If this is an unaligned load and the target doesn't support it, 828 // expand it. 829 EVT MemVT = LD->getMemoryVT(); 830 unsigned AS = LD->getAddressSpace(); 831 unsigned Align = LD->getAlignment(); 832 const DataLayout &DL = DAG.getDataLayout(); 833 if (!TLI.allowsMemoryAccess(*DAG.getContext(), DL, MemVT, AS, Align)) { 834 std::tie(Value, Chain) = TLI.expandUnalignedLoad(LD, DAG); 835 } 836 } 837 break; 838 } 839 case TargetLowering::Expand: 840 EVT DestVT = Node->getValueType(0); 841 if (!TLI.isLoadExtLegal(ISD::EXTLOAD, DestVT, SrcVT)) { 842 // If the source type is not legal, see if there is a legal extload to 843 // an intermediate type that we can then extend further. 844 EVT LoadVT = TLI.getRegisterType(SrcVT.getSimpleVT()); 845 if (TLI.isTypeLegal(SrcVT) || // Same as SrcVT == LoadVT? 846 TLI.isLoadExtLegal(ExtType, LoadVT, SrcVT)) { 847 // If we are loading a legal type, this is a non-extload followed by a 848 // full extend. 849 ISD::LoadExtType MidExtType = 850 (LoadVT == SrcVT) ? ISD::NON_EXTLOAD : ExtType; 851 852 SDValue Load = DAG.getExtLoad(MidExtType, dl, LoadVT, Chain, Ptr, 853 SrcVT, LD->getMemOperand()); 854 unsigned ExtendOp = 855 ISD::getExtForLoadExtType(SrcVT.isFloatingPoint(), ExtType); 856 Value = DAG.getNode(ExtendOp, dl, Node->getValueType(0), Load); 857 Chain = Load.getValue(1); 858 break; 859 } 860 861 // Handle the special case of fp16 extloads. EXTLOAD doesn't have the 862 // normal undefined upper bits behavior to allow using an in-reg extend 863 // with the illegal FP type, so load as an integer and do the 864 // from-integer conversion. 865 if (SrcVT.getScalarType() == MVT::f16) { 866 EVT ISrcVT = SrcVT.changeTypeToInteger(); 867 EVT IDestVT = DestVT.changeTypeToInteger(); 868 EVT LoadVT = TLI.getRegisterType(IDestVT.getSimpleVT()); 869 870 SDValue Result = DAG.getExtLoad(ISD::ZEXTLOAD, dl, LoadVT, 871 Chain, Ptr, ISrcVT, 872 LD->getMemOperand()); 873 Value = DAG.getNode(ISD::FP16_TO_FP, dl, DestVT, Result); 874 Chain = Result.getValue(1); 875 break; 876 } 877 } 878 879 assert(!SrcVT.isVector() && 880 "Vector Loads are handled in LegalizeVectorOps"); 881 882 // FIXME: This does not work for vectors on most targets. Sign- 883 // and zero-extend operations are currently folded into extending 884 // loads, whether they are legal or not, and then we end up here 885 // without any support for legalizing them. 886 assert(ExtType != ISD::EXTLOAD && 887 "EXTLOAD should always be supported!"); 888 // Turn the unsupported load into an EXTLOAD followed by an 889 // explicit zero/sign extend inreg. 890 SDValue Result = DAG.getExtLoad(ISD::EXTLOAD, dl, 891 Node->getValueType(0), 892 Chain, Ptr, SrcVT, 893 LD->getMemOperand()); 894 SDValue ValRes; 895 if (ExtType == ISD::SEXTLOAD) 896 ValRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, 897 Result.getValueType(), 898 Result, DAG.getValueType(SrcVT)); 899 else 900 ValRes = DAG.getZeroExtendInReg(Result, dl, SrcVT.getScalarType()); 901 Value = ValRes; 902 Chain = Result.getValue(1); 903 break; 904 } 905 } 906 907 // Since loads produce two values, make sure to remember that we legalized 908 // both of them. 909 if (Chain.getNode() != Node) { 910 assert(Value.getNode() != Node && "Load must be completely replaced"); 911 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Value); 912 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain); 913 if (UpdatedNodes) { 914 UpdatedNodes->insert(Value.getNode()); 915 UpdatedNodes->insert(Chain.getNode()); 916 } 917 ReplacedNode(Node); 918 } 919} 920 921/// Return a legal replacement for the given operation, with all legal operands. 922void SelectionDAGLegalize::LegalizeOp(SDNode *Node) { 923 DEBUG(dbgs() << "\nLegalizing: "; Node->dump(&DAG)); 924 925 if (Node->getOpcode() == ISD::TargetConstant) // Allow illegal target nodes. 926 return; 927 928#ifndef NDEBUG 929 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i) 930 assert((TLI.getTypeAction(*DAG.getContext(), Node->getValueType(i)) == 931 TargetLowering::TypeLegal || 932 TLI.isTypeLegal(Node->getValueType(i))) && 933 "Unexpected illegal type!"); 934 935 for (const SDValue &Op : Node->op_values()) 936 assert((TLI.getTypeAction(*DAG.getContext(), Op.getValueType()) == 937 TargetLowering::TypeLegal || 938 TLI.isTypeLegal(Op.getValueType()) || 939 Op.getOpcode() == ISD::TargetConstant) && 940 "Unexpected illegal type!"); 941#endif 942 943 // Figure out the correct action; the way to query this varies by opcode 944 TargetLowering::LegalizeAction Action = TargetLowering::Legal; 945 bool SimpleFinishLegalizing = true; 946 switch (Node->getOpcode()) { 947 case ISD::INTRINSIC_W_CHAIN: 948 case ISD::INTRINSIC_WO_CHAIN: 949 case ISD::INTRINSIC_VOID: 950 case ISD::STACKSAVE: 951 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other); 952 break; 953 case ISD::GET_DYNAMIC_AREA_OFFSET: 954 Action = TLI.getOperationAction(Node->getOpcode(), 955 Node->getValueType(0)); 956 break; 957 case ISD::VAARG: 958 Action = TLI.getOperationAction(Node->getOpcode(), 959 Node->getValueType(0)); 960 if (Action != TargetLowering::Promote) 961 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other); 962 break; 963 case ISD::FP_TO_FP16: 964 case ISD::SINT_TO_FP: 965 case ISD::UINT_TO_FP: 966 case ISD::EXTRACT_VECTOR_ELT: 967 Action = TLI.getOperationAction(Node->getOpcode(), 968 Node->getOperand(0).getValueType()); 969 break; 970 case ISD::FP_ROUND_INREG: 971 case ISD::SIGN_EXTEND_INREG: { 972 EVT InnerType = cast<VTSDNode>(Node->getOperand(1))->getVT(); 973 Action = TLI.getOperationAction(Node->getOpcode(), InnerType); 974 break; 975 } 976 case ISD::ATOMIC_STORE: { 977 Action = TLI.getOperationAction(Node->getOpcode(), 978 Node->getOperand(2).getValueType()); 979 break; 980 } 981 case ISD::SELECT_CC: 982 case ISD::SETCC: 983 case ISD::BR_CC: { 984 unsigned CCOperand = Node->getOpcode() == ISD::SELECT_CC ? 4 : 985 Node->getOpcode() == ISD::SETCC ? 2 : 986 Node->getOpcode() == ISD::SETCCE ? 3 : 1; 987 unsigned CompareOperand = Node->getOpcode() == ISD::BR_CC ? 2 : 0; 988 MVT OpVT = Node->getOperand(CompareOperand).getSimpleValueType(); 989 ISD::CondCode CCCode = 990 cast<CondCodeSDNode>(Node->getOperand(CCOperand))->get(); 991 Action = TLI.getCondCodeAction(CCCode, OpVT); 992 if (Action == TargetLowering::Legal) { 993 if (Node->getOpcode() == ISD::SELECT_CC) 994 Action = TLI.getOperationAction(Node->getOpcode(), 995 Node->getValueType(0)); 996 else 997 Action = TLI.getOperationAction(Node->getOpcode(), OpVT); 998 } 999 break; 1000 } 1001 case ISD::LOAD: 1002 case ISD::STORE: 1003 // FIXME: Model these properly. LOAD and STORE are complicated, and 1004 // STORE expects the unlegalized operand in some cases. 1005 SimpleFinishLegalizing = false; 1006 break; 1007 case ISD::CALLSEQ_START: 1008 case ISD::CALLSEQ_END: 1009 // FIXME: This shouldn't be necessary. These nodes have special properties 1010 // dealing with the recursive nature of legalization. Removing this 1011 // special case should be done as part of making LegalizeDAG non-recursive. 1012 SimpleFinishLegalizing = false; 1013 break; 1014 case ISD::EXTRACT_ELEMENT: 1015 case ISD::FLT_ROUNDS_: 1016 case ISD::FPOWI: 1017 case ISD::MERGE_VALUES: 1018 case ISD::EH_RETURN: 1019 case ISD::FRAME_TO_ARGS_OFFSET: 1020 case ISD::EH_SJLJ_SETJMP: 1021 case ISD::EH_SJLJ_LONGJMP: 1022 case ISD::EH_SJLJ_SETUP_DISPATCH: 1023 // These operations lie about being legal: when they claim to be legal, 1024 // they should actually be expanded. 1025 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0)); 1026 if (Action == TargetLowering::Legal) 1027 Action = TargetLowering::Expand; 1028 break; 1029 case ISD::INIT_TRAMPOLINE: 1030 case ISD::ADJUST_TRAMPOLINE: 1031 case ISD::FRAMEADDR: 1032 case ISD::RETURNADDR: 1033 // These operations lie about being legal: when they claim to be legal, 1034 // they should actually be custom-lowered. 1035 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0)); 1036 if (Action == TargetLowering::Legal) 1037 Action = TargetLowering::Custom; 1038 break; 1039 case ISD::READCYCLECOUNTER: 1040 // READCYCLECOUNTER returns an i64, even if type legalization might have 1041 // expanded that to several smaller types. 1042 Action = TLI.getOperationAction(Node->getOpcode(), MVT::i64); 1043 break; 1044 case ISD::READ_REGISTER: 1045 case ISD::WRITE_REGISTER: 1046 // Named register is legal in the DAG, but blocked by register name 1047 // selection if not implemented by target (to chose the correct register) 1048 // They'll be converted to Copy(To/From)Reg. 1049 Action = TargetLowering::Legal; 1050 break; 1051 case ISD::DEBUGTRAP: 1052 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0)); 1053 if (Action == TargetLowering::Expand) { 1054 // replace ISD::DEBUGTRAP with ISD::TRAP 1055 SDValue NewVal; 1056 NewVal = DAG.getNode(ISD::TRAP, SDLoc(Node), Node->getVTList(), 1057 Node->getOperand(0)); 1058 ReplaceNode(Node, NewVal.getNode()); 1059 LegalizeOp(NewVal.getNode()); 1060 return; 1061 } 1062 break; 1063 1064 default: 1065 if (Node->getOpcode() >= ISD::BUILTIN_OP_END) { 1066 Action = TargetLowering::Legal; 1067 } else { 1068 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0)); 1069 } 1070 break; 1071 } 1072 1073 if (SimpleFinishLegalizing) { 1074 SDNode *NewNode = Node; 1075 switch (Node->getOpcode()) { 1076 default: break; 1077 case ISD::SHL: 1078 case ISD::SRL: 1079 case ISD::SRA: 1080 case ISD::ROTL: 1081 case ISD::ROTR: 1082 // Legalizing shifts/rotates requires adjusting the shift amount 1083 // to the appropriate width. 1084 if (!Node->getOperand(1).getValueType().isVector()) { 1085 SDValue SAO = 1086 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(), 1087 Node->getOperand(1)); 1088 HandleSDNode Handle(SAO); 1089 LegalizeOp(SAO.getNode()); 1090 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0), 1091 Handle.getValue()); 1092 } 1093 break; 1094 case ISD::SRL_PARTS: 1095 case ISD::SRA_PARTS: 1096 case ISD::SHL_PARTS: 1097 // Legalizing shifts/rotates requires adjusting the shift amount 1098 // to the appropriate width. 1099 if (!Node->getOperand(2).getValueType().isVector()) { 1100 SDValue SAO = 1101 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(), 1102 Node->getOperand(2)); 1103 HandleSDNode Handle(SAO); 1104 LegalizeOp(SAO.getNode()); 1105 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0), 1106 Node->getOperand(1), 1107 Handle.getValue()); 1108 } 1109 break; 1110 } 1111 1112 if (NewNode != Node) { 1113 ReplaceNode(Node, NewNode); 1114 Node = NewNode; 1115 } 1116 switch (Action) { 1117 case TargetLowering::Legal: 1118 return; 1119 case TargetLowering::Custom: { 1120 // FIXME: The handling for custom lowering with multiple results is 1121 // a complete mess. 1122 if (SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG)) { 1123 if (!(Res.getNode() != Node || Res.getResNo() != 0)) 1124 return; 1125 1126 if (Node->getNumValues() == 1) { 1127 // We can just directly replace this node with the lowered value. 1128 ReplaceNode(SDValue(Node, 0), Res); 1129 return; 1130 } 1131 1132 SmallVector<SDValue, 8> ResultVals; 1133 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i) 1134 ResultVals.push_back(Res.getValue(i)); 1135 ReplaceNode(Node, ResultVals.data()); 1136 return; 1137 } 1138 } 1139 // FALL THROUGH 1140 case TargetLowering::Expand: 1141 if (ExpandNode(Node)) 1142 return; 1143 // FALL THROUGH 1144 case TargetLowering::LibCall: 1145 ConvertNodeToLibcall(Node); 1146 return; 1147 case TargetLowering::Promote: 1148 PromoteNode(Node); 1149 return; 1150 } 1151 } 1152 1153 switch (Node->getOpcode()) { 1154 default: 1155#ifndef NDEBUG 1156 dbgs() << "NODE: "; 1157 Node->dump( &DAG); 1158 dbgs() << "\n"; 1159#endif 1160 llvm_unreachable("Do not know how to legalize this operator!"); 1161 1162 case ISD::CALLSEQ_START: 1163 case ISD::CALLSEQ_END: 1164 break; 1165 case ISD::LOAD: { 1166 return LegalizeLoadOps(Node); 1167 } 1168 case ISD::STORE: { 1169 return LegalizeStoreOps(Node); 1170 } 1171 } 1172} 1173 1174SDValue SelectionDAGLegalize::ExpandExtractFromVectorThroughStack(SDValue Op) { 1175 SDValue Vec = Op.getOperand(0); 1176 SDValue Idx = Op.getOperand(1); 1177 SDLoc dl(Op); 1178 1179 // Before we generate a new store to a temporary stack slot, see if there is 1180 // already one that we can use. There often is because when we scalarize 1181 // vector operations (using SelectionDAG::UnrollVectorOp for example) a whole 1182 // series of EXTRACT_VECTOR_ELT nodes are generated, one for each element in 1183 // the vector. If all are expanded here, we don't want one store per vector 1184 // element. 1185 1186 // Caches for hasPredecessorHelper 1187 SmallPtrSet<const SDNode *, 32> Visited; 1188 SmallVector<const SDNode *, 16> Worklist; 1189 Worklist.push_back(Idx.getNode()); 1190 SDValue StackPtr, Ch; 1191 for (SDNode::use_iterator UI = Vec.getNode()->use_begin(), 1192 UE = Vec.getNode()->use_end(); UI != UE; ++UI) { 1193 SDNode *User = *UI; 1194 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(User)) { 1195 if (ST->isIndexed() || ST->isTruncatingStore() || 1196 ST->getValue() != Vec) 1197 continue; 1198 1199 // Make sure that nothing else could have stored into the destination of 1200 // this store. 1201 if (!ST->getChain().reachesChainWithoutSideEffects(DAG.getEntryNode())) 1202 continue; 1203 1204 // If the index is dependent on the store we will introduce a cycle when 1205 // creating the load (the load uses the index, and by replacing the chain 1206 // we will make the index dependent on the load). 1207 if (SDNode::hasPredecessorHelper(ST, Visited, Worklist)) 1208 continue; 1209 1210 StackPtr = ST->getBasePtr(); 1211 Ch = SDValue(ST, 0); 1212 break; 1213 } 1214 } 1215 1216 if (!Ch.getNode()) { 1217 // Store the value to a temporary stack slot, then LOAD the returned part. 1218 StackPtr = DAG.CreateStackTemporary(Vec.getValueType()); 1219 Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, 1220 MachinePointerInfo(), false, false, 0); 1221 } 1222 1223 // Add the offset to the index. 1224 unsigned EltSize = 1225 Vec.getValueType().getVectorElementType().getSizeInBits()/8; 1226 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx, 1227 DAG.getConstant(EltSize, SDLoc(Vec), Idx.getValueType())); 1228 1229 Idx = DAG.getZExtOrTrunc(Idx, dl, TLI.getPointerTy(DAG.getDataLayout())); 1230 StackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx, StackPtr); 1231 1232 SDValue NewLoad; 1233 1234 if (Op.getValueType().isVector()) 1235 NewLoad = DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr, 1236 MachinePointerInfo(), false, false, false, 0); 1237 else 1238 NewLoad = DAG.getExtLoad( 1239 ISD::EXTLOAD, dl, Op.getValueType(), Ch, StackPtr, MachinePointerInfo(), 1240 Vec.getValueType().getVectorElementType(), false, false, false, 0); 1241 1242 // Replace the chain going out of the store, by the one out of the load. 1243 DAG.ReplaceAllUsesOfValueWith(Ch, SDValue(NewLoad.getNode(), 1)); 1244 1245 // We introduced a cycle though, so update the loads operands, making sure 1246 // to use the original store's chain as an incoming chain. 1247 SmallVector<SDValue, 6> NewLoadOperands(NewLoad->op_begin(), 1248 NewLoad->op_end()); 1249 NewLoadOperands[0] = Ch; 1250 NewLoad = 1251 SDValue(DAG.UpdateNodeOperands(NewLoad.getNode(), NewLoadOperands), 0); 1252 return NewLoad; 1253} 1254 1255SDValue SelectionDAGLegalize::ExpandInsertToVectorThroughStack(SDValue Op) { 1256 assert(Op.getValueType().isVector() && "Non-vector insert subvector!"); 1257 1258 SDValue Vec = Op.getOperand(0); 1259 SDValue Part = Op.getOperand(1); 1260 SDValue Idx = Op.getOperand(2); 1261 SDLoc dl(Op); 1262 1263 // Store the value to a temporary stack slot, then LOAD the returned part. 1264 1265 SDValue StackPtr = DAG.CreateStackTemporary(Vec.getValueType()); 1266 int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex(); 1267 MachinePointerInfo PtrInfo = 1268 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI); 1269 1270 // First store the whole vector. 1271 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, PtrInfo, 1272 false, false, 0); 1273 1274 // Then store the inserted part. 1275 1276 // Add the offset to the index. 1277 unsigned EltSize = 1278 Vec.getValueType().getVectorElementType().getSizeInBits()/8; 1279 1280 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx, 1281 DAG.getConstant(EltSize, SDLoc(Vec), Idx.getValueType())); 1282 Idx = DAG.getZExtOrTrunc(Idx, dl, TLI.getPointerTy(DAG.getDataLayout())); 1283 1284 SDValue SubStackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx, 1285 StackPtr); 1286 1287 // Store the subvector. 1288 Ch = DAG.getStore(Ch, dl, Part, SubStackPtr, 1289 MachinePointerInfo(), false, false, 0); 1290 1291 // Finally, load the updated vector. 1292 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr, PtrInfo, 1293 false, false, false, 0); 1294} 1295 1296SDValue SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node) { 1297 // We can't handle this case efficiently. Allocate a sufficiently 1298 // aligned object on the stack, store each element into it, then load 1299 // the result as a vector. 1300 // Create the stack frame object. 1301 EVT VT = Node->getValueType(0); 1302 EVT EltVT = VT.getVectorElementType(); 1303 SDLoc dl(Node); 1304 SDValue FIPtr = DAG.CreateStackTemporary(VT); 1305 int FI = cast<FrameIndexSDNode>(FIPtr.getNode())->getIndex(); 1306 MachinePointerInfo PtrInfo = 1307 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI); 1308 1309 // Emit a store of each element to the stack slot. 1310 SmallVector<SDValue, 8> Stores; 1311 unsigned TypeByteSize = EltVT.getSizeInBits() / 8; 1312 // Store (in the right endianness) the elements to memory. 1313 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) { 1314 // Ignore undef elements. 1315 if (Node->getOperand(i).isUndef()) continue; 1316 1317 unsigned Offset = TypeByteSize*i; 1318 1319 SDValue Idx = DAG.getConstant(Offset, dl, FIPtr.getValueType()); 1320 Idx = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr, Idx); 1321 1322 // If the destination vector element type is narrower than the source 1323 // element type, only store the bits necessary. 1324 if (EltVT.bitsLT(Node->getOperand(i).getValueType().getScalarType())) { 1325 Stores.push_back(DAG.getTruncStore(DAG.getEntryNode(), dl, 1326 Node->getOperand(i), Idx, 1327 PtrInfo.getWithOffset(Offset), 1328 EltVT, false, false, 0)); 1329 } else 1330 Stores.push_back(DAG.getStore(DAG.getEntryNode(), dl, 1331 Node->getOperand(i), Idx, 1332 PtrInfo.getWithOffset(Offset), 1333 false, false, 0)); 1334 } 1335 1336 SDValue StoreChain; 1337 if (!Stores.empty()) // Not all undef elements? 1338 StoreChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores); 1339 else 1340 StoreChain = DAG.getEntryNode(); 1341 1342 // Result is a load from the stack slot. 1343 return DAG.getLoad(VT, dl, StoreChain, FIPtr, PtrInfo, 1344 false, false, false, 0); 1345} 1346 1347namespace { 1348/// Keeps track of state when getting the sign of a floating-point value as an 1349/// integer. 1350struct FloatSignAsInt { 1351 EVT FloatVT; 1352 SDValue Chain; 1353 SDValue FloatPtr; 1354 SDValue IntPtr; 1355 MachinePointerInfo IntPointerInfo; 1356 MachinePointerInfo FloatPointerInfo; 1357 SDValue IntValue; 1358 APInt SignMask; 1359 uint8_t SignBit; 1360}; 1361} 1362 1363/// Bitcast a floating-point value to an integer value. Only bitcast the part 1364/// containing the sign bit if the target has no integer value capable of 1365/// holding all bits of the floating-point value. 1366void SelectionDAGLegalize::getSignAsIntValue(FloatSignAsInt &State, 1367 const SDLoc &DL, 1368 SDValue Value) const { 1369 EVT FloatVT = Value.getValueType(); 1370 unsigned NumBits = FloatVT.getSizeInBits(); 1371 State.FloatVT = FloatVT; 1372 EVT IVT = EVT::getIntegerVT(*DAG.getContext(), NumBits); 1373 // Convert to an integer of the same size. 1374 if (TLI.isTypeLegal(IVT)) { 1375 State.IntValue = DAG.getNode(ISD::BITCAST, DL, IVT, Value); 1376 State.SignMask = APInt::getSignBit(NumBits); 1377 State.SignBit = NumBits - 1; 1378 return; 1379 } 1380 1381 auto &DataLayout = DAG.getDataLayout(); 1382 // Store the float to memory, then load the sign part out as an integer. 1383 MVT LoadTy = TLI.getRegisterType(*DAG.getContext(), MVT::i8); 1384 // First create a temporary that is aligned for both the load and store. 1385 SDValue StackPtr = DAG.CreateStackTemporary(FloatVT, LoadTy); 1386 int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex(); 1387 // Then store the float to it. 1388 State.FloatPtr = StackPtr; 1389 MachineFunction &MF = DAG.getMachineFunction(); 1390 State.FloatPointerInfo = MachinePointerInfo::getFixedStack(MF, FI); 1391 State.Chain = DAG.getStore(DAG.getEntryNode(), DL, Value, State.FloatPtr, 1392 State.FloatPointerInfo, false, false, 0); 1393 1394 SDValue IntPtr; 1395 if (DataLayout.isBigEndian()) { 1396 assert(FloatVT.isByteSized() && "Unsupported floating point type!"); 1397 // Load out a legal integer with the same sign bit as the float. 1398 IntPtr = StackPtr; 1399 State.IntPointerInfo = State.FloatPointerInfo; 1400 } else { 1401 // Advance the pointer so that the loaded byte will contain the sign bit. 1402 unsigned ByteOffset = (FloatVT.getSizeInBits() / 8) - 1; 1403 IntPtr = DAG.getNode(ISD::ADD, DL, StackPtr.getValueType(), StackPtr, 1404 DAG.getConstant(ByteOffset, DL, StackPtr.getValueType())); 1405 State.IntPointerInfo = MachinePointerInfo::getFixedStack(MF, FI, 1406 ByteOffset); 1407 } 1408 1409 State.IntPtr = IntPtr; 1410 State.IntValue = DAG.getExtLoad(ISD::EXTLOAD, DL, LoadTy, State.Chain, 1411 IntPtr, State.IntPointerInfo, MVT::i8, 1412 false, false, false, 0); 1413 State.SignMask = APInt::getOneBitSet(LoadTy.getSizeInBits(), 7); 1414 State.SignBit = 7; 1415} 1416 1417/// Replace the integer value produced by getSignAsIntValue() with a new value 1418/// and cast the result back to a floating-point type. 1419SDValue SelectionDAGLegalize::modifySignAsInt(const FloatSignAsInt &State, 1420 const SDLoc &DL, 1421 SDValue NewIntValue) const { 1422 if (!State.Chain) 1423 return DAG.getNode(ISD::BITCAST, DL, State.FloatVT, NewIntValue); 1424 1425 // Override the part containing the sign bit in the value stored on the stack. 1426 SDValue Chain = DAG.getTruncStore(State.Chain, DL, NewIntValue, State.IntPtr, 1427 State.IntPointerInfo, MVT::i8, false, false, 1428 0); 1429 return DAG.getLoad(State.FloatVT, DL, Chain, State.FloatPtr, 1430 State.FloatPointerInfo, false, false, false, 0); 1431} 1432 1433SDValue SelectionDAGLegalize::ExpandFCOPYSIGN(SDNode *Node) const { 1434 SDLoc DL(Node); 1435 SDValue Mag = Node->getOperand(0); 1436 SDValue Sign = Node->getOperand(1); 1437 1438 // Get sign bit into an integer value. 1439 FloatSignAsInt SignAsInt; 1440 getSignAsIntValue(SignAsInt, DL, Sign); 1441 1442 EVT IntVT = SignAsInt.IntValue.getValueType(); 1443 SDValue SignMask = DAG.getConstant(SignAsInt.SignMask, DL, IntVT); 1444 SDValue SignBit = DAG.getNode(ISD::AND, DL, IntVT, SignAsInt.IntValue, 1445 SignMask); 1446 1447 // If FABS is legal transform FCOPYSIGN(x, y) => sign(x) ? -FABS(x) : FABS(X) 1448 EVT FloatVT = Mag.getValueType(); 1449 if (TLI.isOperationLegalOrCustom(ISD::FABS, FloatVT) && 1450 TLI.isOperationLegalOrCustom(ISD::FNEG, FloatVT)) { 1451 SDValue AbsValue = DAG.getNode(ISD::FABS, DL, FloatVT, Mag); 1452 SDValue NegValue = DAG.getNode(ISD::FNEG, DL, FloatVT, AbsValue); 1453 SDValue Cond = DAG.getSetCC(DL, getSetCCResultType(IntVT), SignBit, 1454 DAG.getConstant(0, DL, IntVT), ISD::SETNE); 1455 return DAG.getSelect(DL, FloatVT, Cond, NegValue, AbsValue); 1456 } 1457 1458 // Transform Mag value to integer, and clear the sign bit. 1459 FloatSignAsInt MagAsInt; 1460 getSignAsIntValue(MagAsInt, DL, Mag); 1461 EVT MagVT = MagAsInt.IntValue.getValueType(); 1462 SDValue ClearSignMask = DAG.getConstant(~MagAsInt.SignMask, DL, MagVT); 1463 SDValue ClearedSign = DAG.getNode(ISD::AND, DL, MagVT, MagAsInt.IntValue, 1464 ClearSignMask); 1465 1466 // Get the signbit at the right position for MagAsInt. 1467 int ShiftAmount = SignAsInt.SignBit - MagAsInt.SignBit; 1468 if (SignBit.getValueSizeInBits() > ClearedSign.getValueSizeInBits()) { 1469 if (ShiftAmount > 0) { 1470 SDValue ShiftCnst = DAG.getConstant(ShiftAmount, DL, IntVT); 1471 SignBit = DAG.getNode(ISD::SRL, DL, IntVT, SignBit, ShiftCnst); 1472 } else if (ShiftAmount < 0) { 1473 SDValue ShiftCnst = DAG.getConstant(-ShiftAmount, DL, IntVT); 1474 SignBit = DAG.getNode(ISD::SHL, DL, IntVT, SignBit, ShiftCnst); 1475 } 1476 SignBit = DAG.getNode(ISD::TRUNCATE, DL, MagVT, SignBit); 1477 } else if (SignBit.getValueSizeInBits() < ClearedSign.getValueSizeInBits()) { 1478 SignBit = DAG.getNode(ISD::ZERO_EXTEND, DL, MagVT, SignBit); 1479 if (ShiftAmount > 0) { 1480 SDValue ShiftCnst = DAG.getConstant(ShiftAmount, DL, MagVT); 1481 SignBit = DAG.getNode(ISD::SRL, DL, MagVT, SignBit, ShiftCnst); 1482 } else if (ShiftAmount < 0) { 1483 SDValue ShiftCnst = DAG.getConstant(-ShiftAmount, DL, MagVT); 1484 SignBit = DAG.getNode(ISD::SHL, DL, MagVT, SignBit, ShiftCnst); 1485 } 1486 } 1487 1488 // Store the part with the modified sign and convert back to float. 1489 SDValue CopiedSign = DAG.getNode(ISD::OR, DL, MagVT, ClearedSign, SignBit); 1490 return modifySignAsInt(MagAsInt, DL, CopiedSign); 1491} 1492 1493SDValue SelectionDAGLegalize::ExpandFABS(SDNode *Node) const { 1494 SDLoc DL(Node); 1495 SDValue Value = Node->getOperand(0); 1496 1497 // Transform FABS(x) => FCOPYSIGN(x, 0.0) if FCOPYSIGN is legal. 1498 EVT FloatVT = Value.getValueType(); 1499 if (TLI.isOperationLegalOrCustom(ISD::FCOPYSIGN, FloatVT)) { 1500 SDValue Zero = DAG.getConstantFP(0.0, DL, FloatVT); 1501 return DAG.getNode(ISD::FCOPYSIGN, DL, FloatVT, Value, Zero); 1502 } 1503 1504 // Transform value to integer, clear the sign bit and transform back. 1505 FloatSignAsInt ValueAsInt; 1506 getSignAsIntValue(ValueAsInt, DL, Value); 1507 EVT IntVT = ValueAsInt.IntValue.getValueType(); 1508 SDValue ClearSignMask = DAG.getConstant(~ValueAsInt.SignMask, DL, IntVT); 1509 SDValue ClearedSign = DAG.getNode(ISD::AND, DL, IntVT, ValueAsInt.IntValue, 1510 ClearSignMask); 1511 return modifySignAsInt(ValueAsInt, DL, ClearedSign); 1512} 1513 1514void SelectionDAGLegalize::ExpandDYNAMIC_STACKALLOC(SDNode* Node, 1515 SmallVectorImpl<SDValue> &Results) { 1516 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore(); 1517 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and" 1518 " not tell us which reg is the stack pointer!"); 1519 SDLoc dl(Node); 1520 EVT VT = Node->getValueType(0); 1521 SDValue Tmp1 = SDValue(Node, 0); 1522 SDValue Tmp2 = SDValue(Node, 1); 1523 SDValue Tmp3 = Node->getOperand(2); 1524 SDValue Chain = Tmp1.getOperand(0); 1525 1526 // Chain the dynamic stack allocation so that it doesn't modify the stack 1527 // pointer when other instructions are using the stack. 1528 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true), dl); 1529 1530 SDValue Size = Tmp2.getOperand(1); 1531 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT); 1532 Chain = SP.getValue(1); 1533 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue(); 1534 unsigned StackAlign = 1535 DAG.getSubtarget().getFrameLowering()->getStackAlignment(); 1536 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value 1537 if (Align > StackAlign) 1538 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1, 1539 DAG.getConstant(-(uint64_t)Align, dl, VT)); 1540 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain 1541 1542 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true), 1543 DAG.getIntPtrConstant(0, dl, true), SDValue(), dl); 1544 1545 Results.push_back(Tmp1); 1546 Results.push_back(Tmp2); 1547} 1548 1549/// Legalize a SETCC with given LHS and RHS and condition code CC on the current 1550/// target. 1551/// 1552/// If the SETCC has been legalized using AND / OR, then the legalized node 1553/// will be stored in LHS. RHS and CC will be set to SDValue(). NeedInvert 1554/// will be set to false. 1555/// 1556/// If the SETCC has been legalized by using getSetCCSwappedOperands(), 1557/// then the values of LHS and RHS will be swapped, CC will be set to the 1558/// new condition, and NeedInvert will be set to false. 1559/// 1560/// If the SETCC has been legalized using the inverse condcode, then LHS and 1561/// RHS will be unchanged, CC will set to the inverted condcode, and NeedInvert 1562/// will be set to true. The caller must invert the result of the SETCC with 1563/// SelectionDAG::getLogicalNOT() or take equivalent action to swap the effect 1564/// of a true/false result. 1565/// 1566/// \returns true if the SetCC has been legalized, false if it hasn't. 1567bool SelectionDAGLegalize::LegalizeSetCCCondCode(EVT VT, SDValue &LHS, 1568 SDValue &RHS, SDValue &CC, 1569 bool &NeedInvert, 1570 const SDLoc &dl) { 1571 MVT OpVT = LHS.getSimpleValueType(); 1572 ISD::CondCode CCCode = cast<CondCodeSDNode>(CC)->get(); 1573 NeedInvert = false; 1574 switch (TLI.getCondCodeAction(CCCode, OpVT)) { 1575 default: llvm_unreachable("Unknown condition code action!"); 1576 case TargetLowering::Legal: 1577 // Nothing to do. 1578 break; 1579 case TargetLowering::Expand: { 1580 ISD::CondCode InvCC = ISD::getSetCCSwappedOperands(CCCode); 1581 if (TLI.isCondCodeLegal(InvCC, OpVT)) { 1582 std::swap(LHS, RHS); 1583 CC = DAG.getCondCode(InvCC); 1584 return true; 1585 } 1586 ISD::CondCode CC1 = ISD::SETCC_INVALID, CC2 = ISD::SETCC_INVALID; 1587 unsigned Opc = 0; 1588 switch (CCCode) { 1589 default: llvm_unreachable("Don't know how to expand this condition!"); 1590 case ISD::SETO: 1591 assert(TLI.getCondCodeAction(ISD::SETOEQ, OpVT) 1592 == TargetLowering::Legal 1593 && "If SETO is expanded, SETOEQ must be legal!"); 1594 CC1 = ISD::SETOEQ; CC2 = ISD::SETOEQ; Opc = ISD::AND; break; 1595 case ISD::SETUO: 1596 assert(TLI.getCondCodeAction(ISD::SETUNE, OpVT) 1597 == TargetLowering::Legal 1598 && "If SETUO is expanded, SETUNE must be legal!"); 1599 CC1 = ISD::SETUNE; CC2 = ISD::SETUNE; Opc = ISD::OR; break; 1600 case ISD::SETOEQ: 1601 case ISD::SETOGT: 1602 case ISD::SETOGE: 1603 case ISD::SETOLT: 1604 case ISD::SETOLE: 1605 case ISD::SETONE: 1606 case ISD::SETUEQ: 1607 case ISD::SETUNE: 1608 case ISD::SETUGT: 1609 case ISD::SETUGE: 1610 case ISD::SETULT: 1611 case ISD::SETULE: 1612 // If we are floating point, assign and break, otherwise fall through. 1613 if (!OpVT.isInteger()) { 1614 // We can use the 4th bit to tell if we are the unordered 1615 // or ordered version of the opcode. 1616 CC2 = ((unsigned)CCCode & 0x8U) ? ISD::SETUO : ISD::SETO; 1617 Opc = ((unsigned)CCCode & 0x8U) ? ISD::OR : ISD::AND; 1618 CC1 = (ISD::CondCode)(((int)CCCode & 0x7) | 0x10); 1619 break; 1620 } 1621 // Fallthrough if we are unsigned integer. 1622 case ISD::SETLE: 1623 case ISD::SETGT: 1624 case ISD::SETGE: 1625 case ISD::SETLT: 1626 // We only support using the inverted operation, which is computed above 1627 // and not a different manner of supporting expanding these cases. 1628 llvm_unreachable("Don't know how to expand this condition!"); 1629 case ISD::SETNE: 1630 case ISD::SETEQ: 1631 // Try inverting the result of the inverse condition. 1632 InvCC = CCCode == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ; 1633 if (TLI.isCondCodeLegal(InvCC, OpVT)) { 1634 CC = DAG.getCondCode(InvCC); 1635 NeedInvert = true; 1636 return true; 1637 } 1638 // If inverting the condition didn't work then we have no means to expand 1639 // the condition. 1640 llvm_unreachable("Don't know how to expand this condition!"); 1641 } 1642 1643 SDValue SetCC1, SetCC2; 1644 if (CCCode != ISD::SETO && CCCode != ISD::SETUO) { 1645 // If we aren't the ordered or unorder operation, 1646 // then the pattern is (LHS CC1 RHS) Opc (LHS CC2 RHS). 1647 SetCC1 = DAG.getSetCC(dl, VT, LHS, RHS, CC1); 1648 SetCC2 = DAG.getSetCC(dl, VT, LHS, RHS, CC2); 1649 } else { 1650 // Otherwise, the pattern is (LHS CC1 LHS) Opc (RHS CC2 RHS) 1651 SetCC1 = DAG.getSetCC(dl, VT, LHS, LHS, CC1); 1652 SetCC2 = DAG.getSetCC(dl, VT, RHS, RHS, CC2); 1653 } 1654 LHS = DAG.getNode(Opc, dl, VT, SetCC1, SetCC2); 1655 RHS = SDValue(); 1656 CC = SDValue(); 1657 return true; 1658 } 1659 } 1660 return false; 1661} 1662 1663/// Emit a store/load combination to the stack. This stores 1664/// SrcOp to a stack slot of type SlotVT, truncating it if needed. It then does 1665/// a load from the stack slot to DestVT, extending it if needed. 1666/// The resultant code need not be legal. 1667SDValue SelectionDAGLegalize::EmitStackConvert(SDValue SrcOp, EVT SlotVT, 1668 EVT DestVT, const SDLoc &dl) { 1669 // Create the stack frame object. 1670 unsigned SrcAlign = DAG.getDataLayout().getPrefTypeAlignment( 1671 SrcOp.getValueType().getTypeForEVT(*DAG.getContext())); 1672 SDValue FIPtr = DAG.CreateStackTemporary(SlotVT, SrcAlign); 1673 1674 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(FIPtr); 1675 int SPFI = StackPtrFI->getIndex(); 1676 MachinePointerInfo PtrInfo = 1677 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SPFI); 1678 1679 unsigned SrcSize = SrcOp.getValueType().getSizeInBits(); 1680 unsigned SlotSize = SlotVT.getSizeInBits(); 1681 unsigned DestSize = DestVT.getSizeInBits(); 1682 Type *DestType = DestVT.getTypeForEVT(*DAG.getContext()); 1683 unsigned DestAlign = DAG.getDataLayout().getPrefTypeAlignment(DestType); 1684 1685 // Emit a store to the stack slot. Use a truncstore if the input value is 1686 // later than DestVT. 1687 SDValue Store; 1688 1689 if (SrcSize > SlotSize) 1690 Store = DAG.getTruncStore(DAG.getEntryNode(), dl, SrcOp, FIPtr, 1691 PtrInfo, SlotVT, false, false, SrcAlign); 1692 else { 1693 assert(SrcSize == SlotSize && "Invalid store"); 1694 Store = DAG.getStore(DAG.getEntryNode(), dl, SrcOp, FIPtr, 1695 PtrInfo, false, false, SrcAlign); 1696 } 1697 1698 // Result is a load from the stack slot. 1699 if (SlotSize == DestSize) 1700 return DAG.getLoad(DestVT, dl, Store, FIPtr, PtrInfo, 1701 false, false, false, DestAlign); 1702 1703 assert(SlotSize < DestSize && "Unknown extension!"); 1704 return DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT, Store, FIPtr, 1705 PtrInfo, SlotVT, false, false, false, DestAlign); 1706} 1707 1708SDValue SelectionDAGLegalize::ExpandSCALAR_TO_VECTOR(SDNode *Node) { 1709 SDLoc dl(Node); 1710 // Create a vector sized/aligned stack slot, store the value to element #0, 1711 // then load the whole vector back out. 1712 SDValue StackPtr = DAG.CreateStackTemporary(Node->getValueType(0)); 1713 1714 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(StackPtr); 1715 int SPFI = StackPtrFI->getIndex(); 1716 1717 SDValue Ch = DAG.getTruncStore( 1718 DAG.getEntryNode(), dl, Node->getOperand(0), StackPtr, 1719 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SPFI), 1720 Node->getValueType(0).getVectorElementType(), false, false, 0); 1721 return DAG.getLoad( 1722 Node->getValueType(0), dl, Ch, StackPtr, 1723 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SPFI), false, 1724 false, false, 0); 1725} 1726 1727static bool 1728ExpandBVWithShuffles(SDNode *Node, SelectionDAG &DAG, 1729 const TargetLowering &TLI, SDValue &Res) { 1730 unsigned NumElems = Node->getNumOperands(); 1731 SDLoc dl(Node); 1732 EVT VT = Node->getValueType(0); 1733 1734 // Try to group the scalars into pairs, shuffle the pairs together, then 1735 // shuffle the pairs of pairs together, etc. until the vector has 1736 // been built. This will work only if all of the necessary shuffle masks 1737 // are legal. 1738 1739 // We do this in two phases; first to check the legality of the shuffles, 1740 // and next, assuming that all shuffles are legal, to create the new nodes. 1741 for (int Phase = 0; Phase < 2; ++Phase) { 1742 SmallVector<std::pair<SDValue, SmallVector<int, 16> >, 16> IntermedVals, 1743 NewIntermedVals; 1744 for (unsigned i = 0; i < NumElems; ++i) { 1745 SDValue V = Node->getOperand(i); 1746 if (V.isUndef()) 1747 continue; 1748 1749 SDValue Vec; 1750 if (Phase) 1751 Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, V); 1752 IntermedVals.push_back(std::make_pair(Vec, SmallVector<int, 16>(1, i))); 1753 } 1754 1755 while (IntermedVals.size() > 2) { 1756 NewIntermedVals.clear(); 1757 for (unsigned i = 0, e = (IntermedVals.size() & ~1u); i < e; i += 2) { 1758 // This vector and the next vector are shuffled together (simply to 1759 // append the one to the other). 1760 SmallVector<int, 16> ShuffleVec(NumElems, -1); 1761 1762 SmallVector<int, 16> FinalIndices; 1763 FinalIndices.reserve(IntermedVals[i].second.size() + 1764 IntermedVals[i+1].second.size()); 1765 1766 int k = 0; 1767 for (unsigned j = 0, f = IntermedVals[i].second.size(); j != f; 1768 ++j, ++k) { 1769 ShuffleVec[k] = j; 1770 FinalIndices.push_back(IntermedVals[i].second[j]); 1771 } 1772 for (unsigned j = 0, f = IntermedVals[i+1].second.size(); j != f; 1773 ++j, ++k) { 1774 ShuffleVec[k] = NumElems + j; 1775 FinalIndices.push_back(IntermedVals[i+1].second[j]); 1776 } 1777 1778 SDValue Shuffle; 1779 if (Phase) 1780 Shuffle = DAG.getVectorShuffle(VT, dl, IntermedVals[i].first, 1781 IntermedVals[i+1].first, 1782 ShuffleVec); 1783 else if (!TLI.isShuffleMaskLegal(ShuffleVec, VT)) 1784 return false; 1785 NewIntermedVals.push_back( 1786 std::make_pair(Shuffle, std::move(FinalIndices))); 1787 } 1788 1789 // If we had an odd number of defined values, then append the last 1790 // element to the array of new vectors. 1791 if ((IntermedVals.size() & 1) != 0) 1792 NewIntermedVals.push_back(IntermedVals.back()); 1793 1794 IntermedVals.swap(NewIntermedVals); 1795 } 1796 1797 assert(IntermedVals.size() <= 2 && IntermedVals.size() > 0 && 1798 "Invalid number of intermediate vectors"); 1799 SDValue Vec1 = IntermedVals[0].first; 1800 SDValue Vec2; 1801 if (IntermedVals.size() > 1) 1802 Vec2 = IntermedVals[1].first; 1803 else if (Phase) 1804 Vec2 = DAG.getUNDEF(VT); 1805 1806 SmallVector<int, 16> ShuffleVec(NumElems, -1); 1807 for (unsigned i = 0, e = IntermedVals[0].second.size(); i != e; ++i) 1808 ShuffleVec[IntermedVals[0].second[i]] = i; 1809 for (unsigned i = 0, e = IntermedVals[1].second.size(); i != e; ++i) 1810 ShuffleVec[IntermedVals[1].second[i]] = NumElems + i; 1811 1812 if (Phase) 1813 Res = DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec); 1814 else if (!TLI.isShuffleMaskLegal(ShuffleVec, VT)) 1815 return false; 1816 } 1817 1818 return true; 1819} 1820 1821/// Expand a BUILD_VECTOR node on targets that don't 1822/// support the operation, but do support the resultant vector type. 1823SDValue SelectionDAGLegalize::ExpandBUILD_VECTOR(SDNode *Node) { 1824 unsigned NumElems = Node->getNumOperands(); 1825 SDValue Value1, Value2; 1826 SDLoc dl(Node); 1827 EVT VT = Node->getValueType(0); 1828 EVT OpVT = Node->getOperand(0).getValueType(); 1829 EVT EltVT = VT.getVectorElementType(); 1830 1831 // If the only non-undef value is the low element, turn this into a 1832 // SCALAR_TO_VECTOR node. If this is { X, X, X, X }, determine X. 1833 bool isOnlyLowElement = true; 1834 bool MoreThanTwoValues = false; 1835 bool isConstant = true; 1836 for (unsigned i = 0; i < NumElems; ++i) { 1837 SDValue V = Node->getOperand(i); 1838 if (V.isUndef()) 1839 continue; 1840 if (i > 0) 1841 isOnlyLowElement = false; 1842 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V)) 1843 isConstant = false; 1844 1845 if (!Value1.getNode()) { 1846 Value1 = V; 1847 } else if (!Value2.getNode()) { 1848 if (V != Value1) 1849 Value2 = V; 1850 } else if (V != Value1 && V != Value2) { 1851 MoreThanTwoValues = true; 1852 } 1853 } 1854 1855 if (!Value1.getNode()) 1856 return DAG.getUNDEF(VT); 1857 1858 if (isOnlyLowElement) 1859 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Node->getOperand(0)); 1860 1861 // If all elements are constants, create a load from the constant pool. 1862 if (isConstant) { 1863 SmallVector<Constant*, 16> CV; 1864 for (unsigned i = 0, e = NumElems; i != e; ++i) { 1865 if (ConstantFPSDNode *V = 1866 dyn_cast<ConstantFPSDNode>(Node->getOperand(i))) { 1867 CV.push_back(const_cast<ConstantFP *>(V->getConstantFPValue())); 1868 } else if (ConstantSDNode *V = 1869 dyn_cast<ConstantSDNode>(Node->getOperand(i))) { 1870 if (OpVT==EltVT) 1871 CV.push_back(const_cast<ConstantInt *>(V->getConstantIntValue())); 1872 else { 1873 // If OpVT and EltVT don't match, EltVT is not legal and the 1874 // element values have been promoted/truncated earlier. Undo this; 1875 // we don't want a v16i8 to become a v16i32 for example. 1876 const ConstantInt *CI = V->getConstantIntValue(); 1877 CV.push_back(ConstantInt::get(EltVT.getTypeForEVT(*DAG.getContext()), 1878 CI->getZExtValue())); 1879 } 1880 } else { 1881 assert(Node->getOperand(i).isUndef()); 1882 Type *OpNTy = EltVT.getTypeForEVT(*DAG.getContext()); 1883 CV.push_back(UndefValue::get(OpNTy)); 1884 } 1885 } 1886 Constant *CP = ConstantVector::get(CV); 1887 SDValue CPIdx = 1888 DAG.getConstantPool(CP, TLI.getPointerTy(DAG.getDataLayout())); 1889 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment(); 1890 return DAG.getLoad( 1891 VT, dl, DAG.getEntryNode(), CPIdx, 1892 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false, 1893 false, false, Alignment); 1894 } 1895 1896 SmallSet<SDValue, 16> DefinedValues; 1897 for (unsigned i = 0; i < NumElems; ++i) { 1898 if (Node->getOperand(i).isUndef()) 1899 continue; 1900 DefinedValues.insert(Node->getOperand(i)); 1901 } 1902 1903 if (TLI.shouldExpandBuildVectorWithShuffles(VT, DefinedValues.size())) { 1904 if (!MoreThanTwoValues) { 1905 SmallVector<int, 8> ShuffleVec(NumElems, -1); 1906 for (unsigned i = 0; i < NumElems; ++i) { 1907 SDValue V = Node->getOperand(i); 1908 if (V.isUndef()) 1909 continue; 1910 ShuffleVec[i] = V == Value1 ? 0 : NumElems; 1911 } 1912 if (TLI.isShuffleMaskLegal(ShuffleVec, Node->getValueType(0))) { 1913 // Get the splatted value into the low element of a vector register. 1914 SDValue Vec1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value1); 1915 SDValue Vec2; 1916 if (Value2.getNode()) 1917 Vec2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value2); 1918 else 1919 Vec2 = DAG.getUNDEF(VT); 1920 1921 // Return shuffle(LowValVec, undef, <0,0,0,0>) 1922 return DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec); 1923 } 1924 } else { 1925 SDValue Res; 1926 if (ExpandBVWithShuffles(Node, DAG, TLI, Res)) 1927 return Res; 1928 } 1929 } 1930 1931 // Otherwise, we can't handle this case efficiently. 1932 return ExpandVectorBuildThroughStack(Node); 1933} 1934 1935// Expand a node into a call to a libcall. If the result value 1936// does not fit into a register, return the lo part and set the hi part to the 1937// by-reg argument. If it does fit into a single register, return the result 1938// and leave the Hi part unset. 1939SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, SDNode *Node, 1940 bool isSigned) { 1941 TargetLowering::ArgListTy Args; 1942 TargetLowering::ArgListEntry Entry; 1943 for (const SDValue &Op : Node->op_values()) { 1944 EVT ArgVT = Op.getValueType(); 1945 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext()); 1946 Entry.Node = Op; 1947 Entry.Ty = ArgTy; 1948 Entry.isSExt = isSigned; 1949 Entry.isZExt = !isSigned; 1950 Args.push_back(Entry); 1951 } 1952 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC), 1953 TLI.getPointerTy(DAG.getDataLayout())); 1954 1955 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext()); 1956 1957 // By default, the input chain to this libcall is the entry node of the 1958 // function. If the libcall is going to be emitted as a tail call then 1959 // TLI.isUsedByReturnOnly will change it to the right chain if the return 1960 // node which is being folded has a non-entry input chain. 1961 SDValue InChain = DAG.getEntryNode(); 1962 1963 // isTailCall may be true since the callee does not reference caller stack 1964 // frame. Check if it's in the right position and that the return types match. 1965 SDValue TCChain = InChain; 1966 const Function *F = DAG.getMachineFunction().getFunction(); 1967 bool isTailCall = 1968 TLI.isInTailCallPosition(DAG, Node, TCChain) && 1969 (RetTy == F->getReturnType() || F->getReturnType()->isVoidTy()); 1970 if (isTailCall) 1971 InChain = TCChain; 1972 1973 TargetLowering::CallLoweringInfo CLI(DAG); 1974 CLI.setDebugLoc(SDLoc(Node)).setChain(InChain) 1975 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args)) 1976 .setTailCall(isTailCall).setSExtResult(isSigned).setZExtResult(!isSigned); 1977 1978 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI); 1979 1980 if (!CallInfo.second.getNode()) 1981 // It's a tailcall, return the chain (which is the DAG root). 1982 return DAG.getRoot(); 1983 1984 return CallInfo.first; 1985} 1986 1987/// Generate a libcall taking the given operands as arguments 1988/// and returning a result of type RetVT. 1989SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, EVT RetVT, 1990 const SDValue *Ops, unsigned NumOps, 1991 bool isSigned, const SDLoc &dl) { 1992 TargetLowering::ArgListTy Args; 1993 Args.reserve(NumOps); 1994 1995 TargetLowering::ArgListEntry Entry; 1996 for (unsigned i = 0; i != NumOps; ++i) { 1997 Entry.Node = Ops[i]; 1998 Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext()); 1999 Entry.isSExt = isSigned; 2000 Entry.isZExt = !isSigned; 2001 Args.push_back(Entry); 2002 } 2003 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC), 2004 TLI.getPointerTy(DAG.getDataLayout())); 2005 2006 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext()); 2007 2008 TargetLowering::CallLoweringInfo CLI(DAG); 2009 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode()) 2010 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args)) 2011 .setSExtResult(isSigned).setZExtResult(!isSigned); 2012 2013 std::pair<SDValue,SDValue> CallInfo = TLI.LowerCallTo(CLI); 2014 2015 return CallInfo.first; 2016} 2017 2018// Expand a node into a call to a libcall. Similar to 2019// ExpandLibCall except that the first operand is the in-chain. 2020std::pair<SDValue, SDValue> 2021SelectionDAGLegalize::ExpandChainLibCall(RTLIB::Libcall LC, 2022 SDNode *Node, 2023 bool isSigned) { 2024 SDValue InChain = Node->getOperand(0); 2025 2026 TargetLowering::ArgListTy Args; 2027 TargetLowering::ArgListEntry Entry; 2028 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i) { 2029 EVT ArgVT = Node->getOperand(i).getValueType(); 2030 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext()); 2031 Entry.Node = Node->getOperand(i); 2032 Entry.Ty = ArgTy; 2033 Entry.isSExt = isSigned; 2034 Entry.isZExt = !isSigned; 2035 Args.push_back(Entry); 2036 } 2037 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC), 2038 TLI.getPointerTy(DAG.getDataLayout())); 2039 2040 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext()); 2041 2042 TargetLowering::CallLoweringInfo CLI(DAG); 2043 CLI.setDebugLoc(SDLoc(Node)).setChain(InChain) 2044 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args)) 2045 .setSExtResult(isSigned).setZExtResult(!isSigned); 2046 2047 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI); 2048 2049 return CallInfo; 2050} 2051 2052SDValue SelectionDAGLegalize::ExpandFPLibCall(SDNode* Node, 2053 RTLIB::Libcall Call_F32, 2054 RTLIB::Libcall Call_F64, 2055 RTLIB::Libcall Call_F80, 2056 RTLIB::Libcall Call_F128, 2057 RTLIB::Libcall Call_PPCF128) { 2058 RTLIB::Libcall LC; 2059 switch (Node->getSimpleValueType(0).SimpleTy) { 2060 default: llvm_unreachable("Unexpected request for libcall!"); 2061 case MVT::f32: LC = Call_F32; break; 2062 case MVT::f64: LC = Call_F64; break; 2063 case MVT::f80: LC = Call_F80; break; 2064 case MVT::f128: LC = Call_F128; break; 2065 case MVT::ppcf128: LC = Call_PPCF128; break; 2066 } 2067 return ExpandLibCall(LC, Node, false); 2068} 2069 2070SDValue SelectionDAGLegalize::ExpandIntLibCall(SDNode* Node, bool isSigned, 2071 RTLIB::Libcall Call_I8, 2072 RTLIB::Libcall Call_I16, 2073 RTLIB::Libcall Call_I32, 2074 RTLIB::Libcall Call_I64, 2075 RTLIB::Libcall Call_I128) { 2076 RTLIB::Libcall LC; 2077 switch (Node->getSimpleValueType(0).SimpleTy) { 2078 default: llvm_unreachable("Unexpected request for libcall!"); 2079 case MVT::i8: LC = Call_I8; break; 2080 case MVT::i16: LC = Call_I16; break; 2081 case MVT::i32: LC = Call_I32; break; 2082 case MVT::i64: LC = Call_I64; break; 2083 case MVT::i128: LC = Call_I128; break; 2084 } 2085 return ExpandLibCall(LC, Node, isSigned); 2086} 2087 2088/// Issue libcalls to __{u}divmod to compute div / rem pairs. 2089void 2090SelectionDAGLegalize::ExpandDivRemLibCall(SDNode *Node, 2091 SmallVectorImpl<SDValue> &Results) { 2092 unsigned Opcode = Node->getOpcode(); 2093 bool isSigned = Opcode == ISD::SDIVREM; 2094 2095 RTLIB::Libcall LC; 2096 switch (Node->getSimpleValueType(0).SimpleTy) { 2097 default: llvm_unreachable("Unexpected request for libcall!"); 2098 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break; 2099 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break; 2100 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break; 2101 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break; 2102 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break; 2103 } 2104 2105 // The input chain to this libcall is the entry node of the function. 2106 // Legalizing the call will automatically add the previous call to the 2107 // dependence. 2108 SDValue InChain = DAG.getEntryNode(); 2109 2110 EVT RetVT = Node->getValueType(0); 2111 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext()); 2112 2113 TargetLowering::ArgListTy Args; 2114 TargetLowering::ArgListEntry Entry; 2115 for (const SDValue &Op : Node->op_values()) { 2116 EVT ArgVT = Op.getValueType(); 2117 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext()); 2118 Entry.Node = Op; 2119 Entry.Ty = ArgTy; 2120 Entry.isSExt = isSigned; 2121 Entry.isZExt = !isSigned; 2122 Args.push_back(Entry); 2123 } 2124 2125 // Also pass the return address of the remainder. 2126 SDValue FIPtr = DAG.CreateStackTemporary(RetVT); 2127 Entry.Node = FIPtr; 2128 Entry.Ty = RetTy->getPointerTo(); 2129 Entry.isSExt = isSigned; 2130 Entry.isZExt = !isSigned; 2131 Args.push_back(Entry); 2132 2133 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC), 2134 TLI.getPointerTy(DAG.getDataLayout())); 2135 2136 SDLoc dl(Node); 2137 TargetLowering::CallLoweringInfo CLI(DAG); 2138 CLI.setDebugLoc(dl).setChain(InChain) 2139 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args)) 2140 .setSExtResult(isSigned).setZExtResult(!isSigned); 2141 2142 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI); 2143 2144 // Remainder is loaded back from the stack frame. 2145 SDValue Rem = DAG.getLoad(RetVT, dl, CallInfo.second, FIPtr, 2146 MachinePointerInfo(), false, false, false, 0); 2147 Results.push_back(CallInfo.first); 2148 Results.push_back(Rem); 2149} 2150 2151/// Return true if sincos libcall is available. 2152static bool isSinCosLibcallAvailable(SDNode *Node, const TargetLowering &TLI) { 2153 RTLIB::Libcall LC; 2154 switch (Node->getSimpleValueType(0).SimpleTy) { 2155 default: llvm_unreachable("Unexpected request for libcall!"); 2156 case MVT::f32: LC = RTLIB::SINCOS_F32; break; 2157 case MVT::f64: LC = RTLIB::SINCOS_F64; break; 2158 case MVT::f80: LC = RTLIB::SINCOS_F80; break; 2159 case MVT::f128: LC = RTLIB::SINCOS_F128; break; 2160 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break; 2161 } 2162 return TLI.getLibcallName(LC) != nullptr; 2163} 2164 2165/// Return true if sincos libcall is available and can be used to combine sin 2166/// and cos. 2167static bool canCombineSinCosLibcall(SDNode *Node, const TargetLowering &TLI, 2168 const TargetMachine &TM) { 2169 if (!isSinCosLibcallAvailable(Node, TLI)) 2170 return false; 2171 // GNU sin/cos functions set errno while sincos does not. Therefore 2172 // combining sin and cos is only safe if unsafe-fpmath is enabled. 2173 if (TM.getTargetTriple().isGNUEnvironment() && !TM.Options.UnsafeFPMath) 2174 return false; 2175 return true; 2176} 2177 2178/// Only issue sincos libcall if both sin and cos are needed. 2179static bool useSinCos(SDNode *Node) { 2180 unsigned OtherOpcode = Node->getOpcode() == ISD::FSIN 2181 ? ISD::FCOS : ISD::FSIN; 2182 2183 SDValue Op0 = Node->getOperand(0); 2184 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(), 2185 UE = Op0.getNode()->use_end(); UI != UE; ++UI) { 2186 SDNode *User = *UI; 2187 if (User == Node) 2188 continue; 2189 // The other user might have been turned into sincos already. 2190 if (User->getOpcode() == OtherOpcode || User->getOpcode() == ISD::FSINCOS) 2191 return true; 2192 } 2193 return false; 2194} 2195 2196/// Issue libcalls to sincos to compute sin / cos pairs. 2197void 2198SelectionDAGLegalize::ExpandSinCosLibCall(SDNode *Node, 2199 SmallVectorImpl<SDValue> &Results) { 2200 RTLIB::Libcall LC; 2201 switch (Node->getSimpleValueType(0).SimpleTy) { 2202 default: llvm_unreachable("Unexpected request for libcall!"); 2203 case MVT::f32: LC = RTLIB::SINCOS_F32; break; 2204 case MVT::f64: LC = RTLIB::SINCOS_F64; break; 2205 case MVT::f80: LC = RTLIB::SINCOS_F80; break; 2206 case MVT::f128: LC = RTLIB::SINCOS_F128; break; 2207 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break; 2208 } 2209 2210 // The input chain to this libcall is the entry node of the function. 2211 // Legalizing the call will automatically add the previous call to the 2212 // dependence. 2213 SDValue InChain = DAG.getEntryNode(); 2214 2215 EVT RetVT = Node->getValueType(0); 2216 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext()); 2217 2218 TargetLowering::ArgListTy Args; 2219 TargetLowering::ArgListEntry Entry; 2220 2221 // Pass the argument. 2222 Entry.Node = Node->getOperand(0); 2223 Entry.Ty = RetTy; 2224 Entry.isSExt = false; 2225 Entry.isZExt = false; 2226 Args.push_back(Entry); 2227 2228 // Pass the return address of sin. 2229 SDValue SinPtr = DAG.CreateStackTemporary(RetVT); 2230 Entry.Node = SinPtr; 2231 Entry.Ty = RetTy->getPointerTo(); 2232 Entry.isSExt = false; 2233 Entry.isZExt = false; 2234 Args.push_back(Entry); 2235 2236 // Also pass the return address of the cos. 2237 SDValue CosPtr = DAG.CreateStackTemporary(RetVT); 2238 Entry.Node = CosPtr; 2239 Entry.Ty = RetTy->getPointerTo(); 2240 Entry.isSExt = false; 2241 Entry.isZExt = false; 2242 Args.push_back(Entry); 2243 2244 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC), 2245 TLI.getPointerTy(DAG.getDataLayout())); 2246 2247 SDLoc dl(Node); 2248 TargetLowering::CallLoweringInfo CLI(DAG); 2249 CLI.setDebugLoc(dl).setChain(InChain) 2250 .setCallee(TLI.getLibcallCallingConv(LC), 2251 Type::getVoidTy(*DAG.getContext()), Callee, std::move(Args)); 2252 2253 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI); 2254 2255 Results.push_back(DAG.getLoad(RetVT, dl, CallInfo.second, SinPtr, 2256 MachinePointerInfo(), false, false, false, 0)); 2257 Results.push_back(DAG.getLoad(RetVT, dl, CallInfo.second, CosPtr, 2258 MachinePointerInfo(), false, false, false, 0)); 2259} 2260 2261/// This function is responsible for legalizing a 2262/// INT_TO_FP operation of the specified operand when the target requests that 2263/// we expand it. At this point, we know that the result and operand types are 2264/// legal for the target. 2265SDValue SelectionDAGLegalize::ExpandLegalINT_TO_FP(bool isSigned, SDValue Op0, 2266 EVT DestVT, 2267 const SDLoc &dl) { 2268 // TODO: Should any fast-math-flags be set for the created nodes? 2269 2270 if (Op0.getValueType() == MVT::i32 && TLI.isTypeLegal(MVT::f64)) { 2271 // simple 32-bit [signed|unsigned] integer to float/double expansion 2272 2273 // Get the stack frame index of a 8 byte buffer. 2274 SDValue StackSlot = DAG.CreateStackTemporary(MVT::f64); 2275 2276 // word offset constant for Hi/Lo address computation 2277 SDValue WordOff = DAG.getConstant(sizeof(int), dl, 2278 StackSlot.getValueType()); 2279 // set up Hi and Lo (into buffer) address based on endian 2280 SDValue Hi = StackSlot; 2281 SDValue Lo = DAG.getNode(ISD::ADD, dl, StackSlot.getValueType(), 2282 StackSlot, WordOff); 2283 if (DAG.getDataLayout().isLittleEndian()) 2284 std::swap(Hi, Lo); 2285 2286 // if signed map to unsigned space 2287 SDValue Op0Mapped; 2288 if (isSigned) { 2289 // constant used to invert sign bit (signed to unsigned mapping) 2290 SDValue SignBit = DAG.getConstant(0x80000000u, dl, MVT::i32); 2291 Op0Mapped = DAG.getNode(ISD::XOR, dl, MVT::i32, Op0, SignBit); 2292 } else { 2293 Op0Mapped = Op0; 2294 } 2295 // store the lo of the constructed double - based on integer input 2296 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, 2297 Op0Mapped, Lo, MachinePointerInfo(), 2298 false, false, 0); 2299 // initial hi portion of constructed double 2300 SDValue InitialHi = DAG.getConstant(0x43300000u, dl, MVT::i32); 2301 // store the hi of the constructed double - biased exponent 2302 SDValue Store2 = DAG.getStore(Store1, dl, InitialHi, Hi, 2303 MachinePointerInfo(), 2304 false, false, 0); 2305 // load the constructed double 2306 SDValue Load = DAG.getLoad(MVT::f64, dl, Store2, StackSlot, 2307 MachinePointerInfo(), false, false, false, 0); 2308 // FP constant to bias correct the final result 2309 SDValue Bias = DAG.getConstantFP(isSigned ? 2310 BitsToDouble(0x4330000080000000ULL) : 2311 BitsToDouble(0x4330000000000000ULL), 2312 dl, MVT::f64); 2313 // subtract the bias 2314 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Load, Bias); 2315 // final result 2316 SDValue Result; 2317 // handle final rounding 2318 if (DestVT == MVT::f64) { 2319 // do nothing 2320 Result = Sub; 2321 } else if (DestVT.bitsLT(MVT::f64)) { 2322 Result = DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub, 2323 DAG.getIntPtrConstant(0, dl)); 2324 } else if (DestVT.bitsGT(MVT::f64)) { 2325 Result = DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub); 2326 } 2327 return Result; 2328 } 2329 assert(!isSigned && "Legalize cannot Expand SINT_TO_FP for i64 yet"); 2330 // Code below here assumes !isSigned without checking again. 2331 2332 // Implementation of unsigned i64 to f64 following the algorithm in 2333 // __floatundidf in compiler_rt. This implementation has the advantage 2334 // of performing rounding correctly, both in the default rounding mode 2335 // and in all alternate rounding modes. 2336 // TODO: Generalize this for use with other types. 2337 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f64) { 2338 SDValue TwoP52 = 2339 DAG.getConstant(UINT64_C(0x4330000000000000), dl, MVT::i64); 2340 SDValue TwoP84PlusTwoP52 = 2341 DAG.getConstantFP(BitsToDouble(UINT64_C(0x4530000000100000)), dl, 2342 MVT::f64); 2343 SDValue TwoP84 = 2344 DAG.getConstant(UINT64_C(0x4530000000000000), dl, MVT::i64); 2345 2346 SDValue Lo = DAG.getZeroExtendInReg(Op0, dl, MVT::i32); 2347 SDValue Hi = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0, 2348 DAG.getConstant(32, dl, MVT::i64)); 2349 SDValue LoOr = DAG.getNode(ISD::OR, dl, MVT::i64, Lo, TwoP52); 2350 SDValue HiOr = DAG.getNode(ISD::OR, dl, MVT::i64, Hi, TwoP84); 2351 SDValue LoFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, LoOr); 2352 SDValue HiFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, HiOr); 2353 SDValue HiSub = DAG.getNode(ISD::FSUB, dl, MVT::f64, HiFlt, 2354 TwoP84PlusTwoP52); 2355 return DAG.getNode(ISD::FADD, dl, MVT::f64, LoFlt, HiSub); 2356 } 2357 2358 // Implementation of unsigned i64 to f32. 2359 // TODO: Generalize this for use with other types. 2360 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f32) { 2361 // For unsigned conversions, convert them to signed conversions using the 2362 // algorithm from the x86_64 __floatundidf in compiler_rt. 2363 if (!isSigned) { 2364 SDValue Fast = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Op0); 2365 2366 SDValue ShiftConst = DAG.getConstant( 2367 1, dl, TLI.getShiftAmountTy(Op0.getValueType(), DAG.getDataLayout())); 2368 SDValue Shr = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0, ShiftConst); 2369 SDValue AndConst = DAG.getConstant(1, dl, MVT::i64); 2370 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0, AndConst); 2371 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And, Shr); 2372 2373 SDValue SignCvt = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Or); 2374 SDValue Slow = DAG.getNode(ISD::FADD, dl, MVT::f32, SignCvt, SignCvt); 2375 2376 // TODO: This really should be implemented using a branch rather than a 2377 // select. We happen to get lucky and machinesink does the right 2378 // thing most of the time. This would be a good candidate for a 2379 //pseudo-op, or, even better, for whole-function isel. 2380 SDValue SignBitTest = DAG.getSetCC(dl, getSetCCResultType(MVT::i64), 2381 Op0, DAG.getConstant(0, dl, MVT::i64), ISD::SETLT); 2382 return DAG.getSelect(dl, MVT::f32, SignBitTest, Slow, Fast); 2383 } 2384 2385 // Otherwise, implement the fully general conversion. 2386 2387 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0, 2388 DAG.getConstant(UINT64_C(0xfffffffffffff800), dl, MVT::i64)); 2389 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And, 2390 DAG.getConstant(UINT64_C(0x800), dl, MVT::i64)); 2391 SDValue And2 = DAG.getNode(ISD::AND, dl, MVT::i64, Op0, 2392 DAG.getConstant(UINT64_C(0x7ff), dl, MVT::i64)); 2393 SDValue Ne = DAG.getSetCC(dl, getSetCCResultType(MVT::i64), And2, 2394 DAG.getConstant(UINT64_C(0), dl, MVT::i64), 2395 ISD::SETNE); 2396 SDValue Sel = DAG.getSelect(dl, MVT::i64, Ne, Or, Op0); 2397 SDValue Ge = DAG.getSetCC(dl, getSetCCResultType(MVT::i64), Op0, 2398 DAG.getConstant(UINT64_C(0x0020000000000000), dl, 2399 MVT::i64), 2400 ISD::SETUGE); 2401 SDValue Sel2 = DAG.getSelect(dl, MVT::i64, Ge, Sel, Op0); 2402 EVT SHVT = TLI.getShiftAmountTy(Sel2.getValueType(), DAG.getDataLayout()); 2403 2404 SDValue Sh = DAG.getNode(ISD::SRL, dl, MVT::i64, Sel2, 2405 DAG.getConstant(32, dl, SHVT)); 2406 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sh); 2407 SDValue Fcvt = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Trunc); 2408 SDValue TwoP32 = 2409 DAG.getConstantFP(BitsToDouble(UINT64_C(0x41f0000000000000)), dl, 2410 MVT::f64); 2411 SDValue Fmul = DAG.getNode(ISD::FMUL, dl, MVT::f64, TwoP32, Fcvt); 2412 SDValue Lo = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sel2); 2413 SDValue Fcvt2 = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Lo); 2414 SDValue Fadd = DAG.getNode(ISD::FADD, dl, MVT::f64, Fmul, Fcvt2); 2415 return DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Fadd, 2416 DAG.getIntPtrConstant(0, dl)); 2417 } 2418 2419 SDValue Tmp1 = DAG.getNode(ISD::SINT_TO_FP, dl, DestVT, Op0); 2420 2421 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(Op0.getValueType()), 2422 Op0, 2423 DAG.getConstant(0, dl, Op0.getValueType()), 2424 ISD::SETLT); 2425 SDValue Zero = DAG.getIntPtrConstant(0, dl), 2426 Four = DAG.getIntPtrConstant(4, dl); 2427 SDValue CstOffset = DAG.getSelect(dl, Zero.getValueType(), 2428 SignSet, Four, Zero); 2429 2430 // If the sign bit of the integer is set, the large number will be treated 2431 // as a negative number. To counteract this, the dynamic code adds an 2432 // offset depending on the data type. 2433 uint64_t FF; 2434 switch (Op0.getSimpleValueType().SimpleTy) { 2435 default: llvm_unreachable("Unsupported integer type!"); 2436 case MVT::i8 : FF = 0x43800000ULL; break; // 2^8 (as a float) 2437 case MVT::i16: FF = 0x47800000ULL; break; // 2^16 (as a float) 2438 case MVT::i32: FF = 0x4F800000ULL; break; // 2^32 (as a float) 2439 case MVT::i64: FF = 0x5F800000ULL; break; // 2^64 (as a float) 2440 } 2441 if (DAG.getDataLayout().isLittleEndian()) 2442 FF <<= 32; 2443 Constant *FudgeFactor = ConstantInt::get( 2444 Type::getInt64Ty(*DAG.getContext()), FF); 2445 2446 SDValue CPIdx = 2447 DAG.getConstantPool(FudgeFactor, TLI.getPointerTy(DAG.getDataLayout())); 2448 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment(); 2449 CPIdx = DAG.getNode(ISD::ADD, dl, CPIdx.getValueType(), CPIdx, CstOffset); 2450 Alignment = std::min(Alignment, 4u); 2451 SDValue FudgeInReg; 2452 if (DestVT == MVT::f32) 2453 FudgeInReg = DAG.getLoad( 2454 MVT::f32, dl, DAG.getEntryNode(), CPIdx, 2455 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false, 2456 false, false, Alignment); 2457 else { 2458 SDValue Load = DAG.getExtLoad( 2459 ISD::EXTLOAD, dl, DestVT, DAG.getEntryNode(), CPIdx, 2460 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32, 2461 false, false, false, Alignment); 2462 HandleSDNode Handle(Load); 2463 LegalizeOp(Load.getNode()); 2464 FudgeInReg = Handle.getValue(); 2465 } 2466 2467 return DAG.getNode(ISD::FADD, dl, DestVT, Tmp1, FudgeInReg); 2468} 2469 2470/// This function is responsible for legalizing a 2471/// *INT_TO_FP operation of the specified operand when the target requests that 2472/// we promote it. At this point, we know that the result and operand types are 2473/// legal for the target, and that there is a legal UINT_TO_FP or SINT_TO_FP 2474/// operation that takes a larger input. 2475SDValue SelectionDAGLegalize::PromoteLegalINT_TO_FP(SDValue LegalOp, EVT DestVT, 2476 bool isSigned, 2477 const SDLoc &dl) { 2478 // First step, figure out the appropriate *INT_TO_FP operation to use. 2479 EVT NewInTy = LegalOp.getValueType(); 2480 2481 unsigned OpToUse = 0; 2482 2483 // Scan for the appropriate larger type to use. 2484 while (1) { 2485 NewInTy = (MVT::SimpleValueType)(NewInTy.getSimpleVT().SimpleTy+1); 2486 assert(NewInTy.isInteger() && "Ran out of possibilities!"); 2487 2488 // If the target supports SINT_TO_FP of this type, use it. 2489 if (TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, NewInTy)) { 2490 OpToUse = ISD::SINT_TO_FP; 2491 break; 2492 } 2493 if (isSigned) continue; 2494 2495 // If the target supports UINT_TO_FP of this type, use it. 2496 if (TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, NewInTy)) { 2497 OpToUse = ISD::UINT_TO_FP; 2498 break; 2499 } 2500 2501 // Otherwise, try a larger type. 2502 } 2503 2504 // Okay, we found the operation and type to use. Zero extend our input to the 2505 // desired type then run the operation on it. 2506 return DAG.getNode(OpToUse, dl, DestVT, 2507 DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND, 2508 dl, NewInTy, LegalOp)); 2509} 2510 2511/// This function is responsible for legalizing a 2512/// FP_TO_*INT operation of the specified operand when the target requests that 2513/// we promote it. At this point, we know that the result and operand types are 2514/// legal for the target, and that there is a legal FP_TO_UINT or FP_TO_SINT 2515/// operation that returns a larger result. 2516SDValue SelectionDAGLegalize::PromoteLegalFP_TO_INT(SDValue LegalOp, EVT DestVT, 2517 bool isSigned, 2518 const SDLoc &dl) { 2519 // First step, figure out the appropriate FP_TO*INT operation to use. 2520 EVT NewOutTy = DestVT; 2521 2522 unsigned OpToUse = 0; 2523 2524 // Scan for the appropriate larger type to use. 2525 while (1) { 2526 NewOutTy = (MVT::SimpleValueType)(NewOutTy.getSimpleVT().SimpleTy+1); 2527 assert(NewOutTy.isInteger() && "Ran out of possibilities!"); 2528 2529 // A larger signed type can hold all unsigned values of the requested type, 2530 // so using FP_TO_SINT is valid 2531 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewOutTy)) { 2532 OpToUse = ISD::FP_TO_SINT; 2533 break; 2534 } 2535 2536 // However, if the value may be < 0.0, we *must* use some FP_TO_SINT. 2537 if (!isSigned && TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewOutTy)) { 2538 OpToUse = ISD::FP_TO_UINT; 2539 break; 2540 } 2541 2542 // Otherwise, try a larger type. 2543 } 2544 2545 2546 // Okay, we found the operation and type to use. 2547 SDValue Operation = DAG.getNode(OpToUse, dl, NewOutTy, LegalOp); 2548 2549 // Truncate the result of the extended FP_TO_*INT operation to the desired 2550 // size. 2551 return DAG.getNode(ISD::TRUNCATE, dl, DestVT, Operation); 2552} 2553 2554/// Open code the operations for BITREVERSE. 2555SDValue SelectionDAGLegalize::ExpandBITREVERSE(SDValue Op, const SDLoc &dl) { 2556 EVT VT = Op.getValueType(); 2557 EVT SHVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout()); 2558 unsigned Sz = VT.getScalarSizeInBits(); 2559 2560 SDValue Tmp, Tmp2; 2561 Tmp = DAG.getConstant(0, dl, VT); 2562 for (unsigned I = 0, J = Sz-1; I < Sz; ++I, --J) { 2563 if (I < J) 2564 Tmp2 = 2565 DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(J - I, dl, SHVT)); 2566 else 2567 Tmp2 = 2568 DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(I - J, dl, SHVT)); 2569 2570 APInt Shift(Sz, 1); 2571 Shift = Shift.shl(J); 2572 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(Shift, dl, VT)); 2573 Tmp = DAG.getNode(ISD::OR, dl, VT, Tmp, Tmp2); 2574 } 2575 2576 return Tmp; 2577} 2578 2579/// Open code the operations for BSWAP of the specified operation. 2580SDValue SelectionDAGLegalize::ExpandBSWAP(SDValue Op, const SDLoc &dl) { 2581 EVT VT = Op.getValueType(); 2582 EVT SHVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout()); 2583 SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5, Tmp6, Tmp7, Tmp8; 2584 switch (VT.getSimpleVT().SimpleTy) { 2585 default: llvm_unreachable("Unhandled Expand type in BSWAP!"); 2586 case MVT::i16: 2587 Tmp2 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, dl, SHVT)); 2588 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT)); 2589 return DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2); 2590 case MVT::i32: 2591 Tmp4 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, dl, SHVT)); 2592 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, dl, SHVT)); 2593 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT)); 2594 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, dl, SHVT)); 2595 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, 2596 DAG.getConstant(0xFF0000, dl, VT)); 2597 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(0xFF00, dl, VT)); 2598 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3); 2599 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1); 2600 return DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2); 2601 case MVT::i64: 2602 Tmp8 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(56, dl, SHVT)); 2603 Tmp7 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(40, dl, SHVT)); 2604 Tmp6 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, dl, SHVT)); 2605 Tmp5 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, dl, SHVT)); 2606 Tmp4 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT)); 2607 Tmp3 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, dl, SHVT)); 2608 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(40, dl, SHVT)); 2609 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(56, dl, SHVT)); 2610 Tmp7 = DAG.getNode(ISD::AND, dl, VT, Tmp7, 2611 DAG.getConstant(255ULL<<48, dl, VT)); 2612 Tmp6 = DAG.getNode(ISD::AND, dl, VT, Tmp6, 2613 DAG.getConstant(255ULL<<40, dl, VT)); 2614 Tmp5 = DAG.getNode(ISD::AND, dl, VT, Tmp5, 2615 DAG.getConstant(255ULL<<32, dl, VT)); 2616 Tmp4 = DAG.getNode(ISD::AND, dl, VT, Tmp4, 2617 DAG.getConstant(255ULL<<24, dl, VT)); 2618 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, 2619 DAG.getConstant(255ULL<<16, dl, VT)); 2620 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, 2621 DAG.getConstant(255ULL<<8 , dl, VT)); 2622 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp7); 2623 Tmp6 = DAG.getNode(ISD::OR, dl, VT, Tmp6, Tmp5); 2624 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3); 2625 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1); 2626 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp6); 2627 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2); 2628 return DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp4); 2629 } 2630} 2631 2632/// Expand the specified bitcount instruction into operations. 2633SDValue SelectionDAGLegalize::ExpandBitCount(unsigned Opc, SDValue Op, 2634 const SDLoc &dl) { 2635 switch (Opc) { 2636 default: llvm_unreachable("Cannot expand this yet!"); 2637 case ISD::CTPOP: { 2638 EVT VT = Op.getValueType(); 2639 EVT ShVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout()); 2640 unsigned Len = VT.getSizeInBits(); 2641 2642 assert(VT.isInteger() && Len <= 128 && Len % 8 == 0 && 2643 "CTPOP not implemented for this type."); 2644 2645 // This is the "best" algorithm from 2646 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel 2647 2648 SDValue Mask55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), 2649 dl, VT); 2650 SDValue Mask33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), 2651 dl, VT); 2652 SDValue Mask0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), 2653 dl, VT); 2654 SDValue Mask01 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x01)), 2655 dl, VT); 2656 2657 // v = v - ((v >> 1) & 0x55555555...) 2658 Op = DAG.getNode(ISD::SUB, dl, VT, Op, 2659 DAG.getNode(ISD::AND, dl, VT, 2660 DAG.getNode(ISD::SRL, dl, VT, Op, 2661 DAG.getConstant(1, dl, ShVT)), 2662 Mask55)); 2663 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...) 2664 Op = DAG.getNode(ISD::ADD, dl, VT, 2665 DAG.getNode(ISD::AND, dl, VT, Op, Mask33), 2666 DAG.getNode(ISD::AND, dl, VT, 2667 DAG.getNode(ISD::SRL, dl, VT, Op, 2668 DAG.getConstant(2, dl, ShVT)), 2669 Mask33)); 2670 // v = (v + (v >> 4)) & 0x0F0F0F0F... 2671 Op = DAG.getNode(ISD::AND, dl, VT, 2672 DAG.getNode(ISD::ADD, dl, VT, Op, 2673 DAG.getNode(ISD::SRL, dl, VT, Op, 2674 DAG.getConstant(4, dl, ShVT))), 2675 Mask0F); 2676 // v = (v * 0x01010101...) >> (Len - 8) 2677 Op = DAG.getNode(ISD::SRL, dl, VT, 2678 DAG.getNode(ISD::MUL, dl, VT, Op, Mask01), 2679 DAG.getConstant(Len - 8, dl, ShVT)); 2680 2681 return Op; 2682 } 2683 case ISD::CTLZ_ZERO_UNDEF: 2684 // This trivially expands to CTLZ. 2685 return DAG.getNode(ISD::CTLZ, dl, Op.getValueType(), Op); 2686 case ISD::CTLZ: { 2687 EVT VT = Op.getValueType(); 2688 unsigned len = VT.getSizeInBits(); 2689 2690 if (TLI.isOperationLegalOrCustom(ISD::CTLZ_ZERO_UNDEF, VT)) { 2691 EVT SetCCVT = getSetCCResultType(VT); 2692 SDValue CTLZ = DAG.getNode(ISD::CTLZ_ZERO_UNDEF, dl, VT, Op); 2693 SDValue Zero = DAG.getConstant(0, dl, VT); 2694 SDValue SrcIsZero = DAG.getSetCC(dl, SetCCVT, Op, Zero, ISD::SETEQ); 2695 return DAG.getNode(ISD::SELECT, dl, VT, SrcIsZero, 2696 DAG.getConstant(len, dl, VT), CTLZ); 2697 } 2698 2699 // for now, we do this: 2700 // x = x | (x >> 1); 2701 // x = x | (x >> 2); 2702 // ... 2703 // x = x | (x >>16); 2704 // x = x | (x >>32); // for 64-bit input 2705 // return popcount(~x); 2706 // 2707 // Ref: "Hacker's Delight" by Henry Warren 2708 EVT ShVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout()); 2709 for (unsigned i = 0; (1U << i) <= (len / 2); ++i) { 2710 SDValue Tmp3 = DAG.getConstant(1ULL << i, dl, ShVT); 2711 Op = DAG.getNode(ISD::OR, dl, VT, Op, 2712 DAG.getNode(ISD::SRL, dl, VT, Op, Tmp3)); 2713 } 2714 Op = DAG.getNOT(dl, Op, VT); 2715 return DAG.getNode(ISD::CTPOP, dl, VT, Op); 2716 } 2717 case ISD::CTTZ_ZERO_UNDEF: 2718 // This trivially expands to CTTZ. 2719 return DAG.getNode(ISD::CTTZ, dl, Op.getValueType(), Op); 2720 case ISD::CTTZ: { 2721 // for now, we use: { return popcount(~x & (x - 1)); } 2722 // unless the target has ctlz but not ctpop, in which case we use: 2723 // { return 32 - nlz(~x & (x-1)); } 2724 // Ref: "Hacker's Delight" by Henry Warren 2725 EVT VT = Op.getValueType(); 2726 SDValue Tmp3 = DAG.getNode(ISD::AND, dl, VT, 2727 DAG.getNOT(dl, Op, VT), 2728 DAG.getNode(ISD::SUB, dl, VT, Op, 2729 DAG.getConstant(1, dl, VT))); 2730 // If ISD::CTLZ is legal and CTPOP isn't, then do that instead. 2731 if (!TLI.isOperationLegalOrCustom(ISD::CTPOP, VT) && 2732 TLI.isOperationLegalOrCustom(ISD::CTLZ, VT)) 2733 return DAG.getNode(ISD::SUB, dl, VT, 2734 DAG.getConstant(VT.getSizeInBits(), dl, VT), 2735 DAG.getNode(ISD::CTLZ, dl, VT, Tmp3)); 2736 return DAG.getNode(ISD::CTPOP, dl, VT, Tmp3); 2737 } 2738 } 2739} 2740 2741bool SelectionDAGLegalize::ExpandNode(SDNode *Node) { 2742 SmallVector<SDValue, 8> Results; 2743 SDLoc dl(Node); 2744 SDValue Tmp1, Tmp2, Tmp3, Tmp4; 2745 bool NeedInvert; 2746 switch (Node->getOpcode()) { 2747 case ISD::CTPOP: 2748 case ISD::CTLZ: 2749 case ISD::CTLZ_ZERO_UNDEF: 2750 case ISD::CTTZ: 2751 case ISD::CTTZ_ZERO_UNDEF: 2752 Tmp1 = ExpandBitCount(Node->getOpcode(), Node->getOperand(0), dl); 2753 Results.push_back(Tmp1); 2754 break; 2755 case ISD::BITREVERSE: 2756 Results.push_back(ExpandBITREVERSE(Node->getOperand(0), dl)); 2757 break; 2758 case ISD::BSWAP: 2759 Results.push_back(ExpandBSWAP(Node->getOperand(0), dl)); 2760 break; 2761 case ISD::FRAMEADDR: 2762 case ISD::RETURNADDR: 2763 case ISD::FRAME_TO_ARGS_OFFSET: 2764 Results.push_back(DAG.getConstant(0, dl, Node->getValueType(0))); 2765 break; 2766 case ISD::FLT_ROUNDS_: 2767 Results.push_back(DAG.getConstant(1, dl, Node->getValueType(0))); 2768 break; 2769 case ISD::EH_RETURN: 2770 case ISD::EH_LABEL: 2771 case ISD::PREFETCH: 2772 case ISD::VAEND: 2773 case ISD::EH_SJLJ_LONGJMP: 2774 // If the target didn't expand these, there's nothing to do, so just 2775 // preserve the chain and be done. 2776 Results.push_back(Node->getOperand(0)); 2777 break; 2778 case ISD::READCYCLECOUNTER: 2779 // If the target didn't expand this, just return 'zero' and preserve the 2780 // chain. 2781 Results.append(Node->getNumValues() - 1, 2782 DAG.getConstant(0, dl, Node->getValueType(0))); 2783 Results.push_back(Node->getOperand(0)); 2784 break; 2785 case ISD::EH_SJLJ_SETJMP: 2786 // If the target didn't expand this, just return 'zero' and preserve the 2787 // chain. 2788 Results.push_back(DAG.getConstant(0, dl, MVT::i32)); 2789 Results.push_back(Node->getOperand(0)); 2790 break; 2791 case ISD::ATOMIC_LOAD: { 2792 // There is no libcall for atomic load; fake it with ATOMIC_CMP_SWAP. 2793 SDValue Zero = DAG.getConstant(0, dl, Node->getValueType(0)); 2794 SDVTList VTs = DAG.getVTList(Node->getValueType(0), MVT::Other); 2795 SDValue Swap = DAG.getAtomicCmpSwap( 2796 ISD::ATOMIC_CMP_SWAP, dl, cast<AtomicSDNode>(Node)->getMemoryVT(), VTs, 2797 Node->getOperand(0), Node->getOperand(1), Zero, Zero, 2798 cast<AtomicSDNode>(Node)->getMemOperand(), 2799 cast<AtomicSDNode>(Node)->getOrdering(), 2800 cast<AtomicSDNode>(Node)->getOrdering(), 2801 cast<AtomicSDNode>(Node)->getSynchScope()); 2802 Results.push_back(Swap.getValue(0)); 2803 Results.push_back(Swap.getValue(1)); 2804 break; 2805 } 2806 case ISD::ATOMIC_STORE: { 2807 // There is no libcall for atomic store; fake it with ATOMIC_SWAP. 2808 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl, 2809 cast<AtomicSDNode>(Node)->getMemoryVT(), 2810 Node->getOperand(0), 2811 Node->getOperand(1), Node->getOperand(2), 2812 cast<AtomicSDNode>(Node)->getMemOperand(), 2813 cast<AtomicSDNode>(Node)->getOrdering(), 2814 cast<AtomicSDNode>(Node)->getSynchScope()); 2815 Results.push_back(Swap.getValue(1)); 2816 break; 2817 } 2818 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: { 2819 // Expanding an ATOMIC_CMP_SWAP_WITH_SUCCESS produces an ATOMIC_CMP_SWAP and 2820 // splits out the success value as a comparison. Expanding the resulting 2821 // ATOMIC_CMP_SWAP will produce a libcall. 2822 SDVTList VTs = DAG.getVTList(Node->getValueType(0), MVT::Other); 2823 SDValue Res = DAG.getAtomicCmpSwap( 2824 ISD::ATOMIC_CMP_SWAP, dl, cast<AtomicSDNode>(Node)->getMemoryVT(), VTs, 2825 Node->getOperand(0), Node->getOperand(1), Node->getOperand(2), 2826 Node->getOperand(3), cast<MemSDNode>(Node)->getMemOperand(), 2827 cast<AtomicSDNode>(Node)->getSuccessOrdering(), 2828 cast<AtomicSDNode>(Node)->getFailureOrdering(), 2829 cast<AtomicSDNode>(Node)->getSynchScope()); 2830 2831 SDValue ExtRes = Res; 2832 SDValue LHS = Res; 2833 SDValue RHS = Node->getOperand(1); 2834 2835 EVT AtomicType = cast<AtomicSDNode>(Node)->getMemoryVT(); 2836 EVT OuterType = Node->getValueType(0); 2837 switch (TLI.getExtendForAtomicOps()) { 2838 case ISD::SIGN_EXTEND: 2839 LHS = DAG.getNode(ISD::AssertSext, dl, OuterType, Res, 2840 DAG.getValueType(AtomicType)); 2841 RHS = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, OuterType, 2842 Node->getOperand(2), DAG.getValueType(AtomicType)); 2843 ExtRes = LHS; 2844 break; 2845 case ISD::ZERO_EXTEND: 2846 LHS = DAG.getNode(ISD::AssertZext, dl, OuterType, Res, 2847 DAG.getValueType(AtomicType)); 2848 RHS = DAG.getNode(ISD::ZERO_EXTEND, dl, OuterType, Node->getOperand(2)); 2849 ExtRes = LHS; 2850 break; 2851 case ISD::ANY_EXTEND: 2852 LHS = DAG.getZeroExtendInReg(Res, dl, AtomicType); 2853 RHS = DAG.getNode(ISD::ZERO_EXTEND, dl, OuterType, Node->getOperand(2)); 2854 break; 2855 default: 2856 llvm_unreachable("Invalid atomic op extension"); 2857 } 2858 2859 SDValue Success = 2860 DAG.getSetCC(dl, Node->getValueType(1), LHS, RHS, ISD::SETEQ); 2861 2862 Results.push_back(ExtRes.getValue(0)); 2863 Results.push_back(Success); 2864 Results.push_back(Res.getValue(1)); 2865 break; 2866 } 2867 case ISD::DYNAMIC_STACKALLOC: 2868 ExpandDYNAMIC_STACKALLOC(Node, Results); 2869 break; 2870 case ISD::MERGE_VALUES: 2871 for (unsigned i = 0; i < Node->getNumValues(); i++) 2872 Results.push_back(Node->getOperand(i)); 2873 break; 2874 case ISD::UNDEF: { 2875 EVT VT = Node->getValueType(0); 2876 if (VT.isInteger()) 2877 Results.push_back(DAG.getConstant(0, dl, VT)); 2878 else { 2879 assert(VT.isFloatingPoint() && "Unknown value type!"); 2880 Results.push_back(DAG.getConstantFP(0, dl, VT)); 2881 } 2882 break; 2883 } 2884 case ISD::FP_ROUND: 2885 case ISD::BITCAST: 2886 Tmp1 = EmitStackConvert(Node->getOperand(0), Node->getValueType(0), 2887 Node->getValueType(0), dl); 2888 Results.push_back(Tmp1); 2889 break; 2890 case ISD::FP_EXTEND: 2891 Tmp1 = EmitStackConvert(Node->getOperand(0), 2892 Node->getOperand(0).getValueType(), 2893 Node->getValueType(0), dl); 2894 Results.push_back(Tmp1); 2895 break; 2896 case ISD::SIGN_EXTEND_INREG: { 2897 // NOTE: we could fall back on load/store here too for targets without 2898 // SAR. However, it is doubtful that any exist. 2899 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT(); 2900 EVT VT = Node->getValueType(0); 2901 EVT ShiftAmountTy = TLI.getShiftAmountTy(VT, DAG.getDataLayout()); 2902 if (VT.isVector()) 2903 ShiftAmountTy = VT; 2904 unsigned BitsDiff = VT.getScalarType().getSizeInBits() - 2905 ExtraVT.getScalarType().getSizeInBits(); 2906 SDValue ShiftCst = DAG.getConstant(BitsDiff, dl, ShiftAmountTy); 2907 Tmp1 = DAG.getNode(ISD::SHL, dl, Node->getValueType(0), 2908 Node->getOperand(0), ShiftCst); 2909 Tmp1 = DAG.getNode(ISD::SRA, dl, Node->getValueType(0), Tmp1, ShiftCst); 2910 Results.push_back(Tmp1); 2911 break; 2912 } 2913 case ISD::FP_ROUND_INREG: { 2914 // The only way we can lower this is to turn it into a TRUNCSTORE, 2915 // EXTLOAD pair, targeting a temporary location (a stack slot). 2916 2917 // NOTE: there is a choice here between constantly creating new stack 2918 // slots and always reusing the same one. We currently always create 2919 // new ones, as reuse may inhibit scheduling. 2920 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT(); 2921 Tmp1 = EmitStackConvert(Node->getOperand(0), ExtraVT, 2922 Node->getValueType(0), dl); 2923 Results.push_back(Tmp1); 2924 break; 2925 } 2926 case ISD::SINT_TO_FP: 2927 case ISD::UINT_TO_FP: 2928 Tmp1 = ExpandLegalINT_TO_FP(Node->getOpcode() == ISD::SINT_TO_FP, 2929 Node->getOperand(0), Node->getValueType(0), dl); 2930 Results.push_back(Tmp1); 2931 break; 2932 case ISD::FP_TO_SINT: 2933 if (TLI.expandFP_TO_SINT(Node, Tmp1, DAG)) 2934 Results.push_back(Tmp1); 2935 break; 2936 case ISD::FP_TO_UINT: { 2937 SDValue True, False; 2938 EVT VT = Node->getOperand(0).getValueType(); 2939 EVT NVT = Node->getValueType(0); 2940 APFloat apf(DAG.EVTToAPFloatSemantics(VT), 2941 APInt::getNullValue(VT.getSizeInBits())); 2942 APInt x = APInt::getSignBit(NVT.getSizeInBits()); 2943 (void)apf.convertFromAPInt(x, false, APFloat::rmNearestTiesToEven); 2944 Tmp1 = DAG.getConstantFP(apf, dl, VT); 2945 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(VT), 2946 Node->getOperand(0), 2947 Tmp1, ISD::SETLT); 2948 True = DAG.getNode(ISD::FP_TO_SINT, dl, NVT, Node->getOperand(0)); 2949 // TODO: Should any fast-math-flags be set for the FSUB? 2950 False = DAG.getNode(ISD::FP_TO_SINT, dl, NVT, 2951 DAG.getNode(ISD::FSUB, dl, VT, 2952 Node->getOperand(0), Tmp1)); 2953 False = DAG.getNode(ISD::XOR, dl, NVT, False, 2954 DAG.getConstant(x, dl, NVT)); 2955 Tmp1 = DAG.getSelect(dl, NVT, Tmp2, True, False); 2956 Results.push_back(Tmp1); 2957 break; 2958 } 2959 case ISD::VAARG: 2960 Results.push_back(DAG.expandVAArg(Node)); 2961 Results.push_back(Results[0].getValue(1)); 2962 break; 2963 case ISD::VACOPY: 2964 Results.push_back(DAG.expandVACopy(Node)); 2965 break; 2966 case ISD::EXTRACT_VECTOR_ELT: 2967 if (Node->getOperand(0).getValueType().getVectorNumElements() == 1) 2968 // This must be an access of the only element. Return it. 2969 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0), 2970 Node->getOperand(0)); 2971 else 2972 Tmp1 = ExpandExtractFromVectorThroughStack(SDValue(Node, 0)); 2973 Results.push_back(Tmp1); 2974 break; 2975 case ISD::EXTRACT_SUBVECTOR: 2976 Results.push_back(ExpandExtractFromVectorThroughStack(SDValue(Node, 0))); 2977 break; 2978 case ISD::INSERT_SUBVECTOR: 2979 Results.push_back(ExpandInsertToVectorThroughStack(SDValue(Node, 0))); 2980 break; 2981 case ISD::CONCAT_VECTORS: { 2982 Results.push_back(ExpandVectorBuildThroughStack(Node)); 2983 break; 2984 } 2985 case ISD::SCALAR_TO_VECTOR: 2986 Results.push_back(ExpandSCALAR_TO_VECTOR(Node)); 2987 break; 2988 case ISD::INSERT_VECTOR_ELT: 2989 Results.push_back(ExpandINSERT_VECTOR_ELT(Node->getOperand(0), 2990 Node->getOperand(1), 2991 Node->getOperand(2), dl)); 2992 break; 2993 case ISD::VECTOR_SHUFFLE: { 2994 SmallVector<int, 32> NewMask; 2995 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask(); 2996 2997 EVT VT = Node->getValueType(0); 2998 EVT EltVT = VT.getVectorElementType(); 2999 SDValue Op0 = Node->getOperand(0); 3000 SDValue Op1 = Node->getOperand(1); 3001 if (!TLI.isTypeLegal(EltVT)) { 3002 3003 EVT NewEltVT = TLI.getTypeToTransformTo(*DAG.getContext(), EltVT); 3004 3005 // BUILD_VECTOR operands are allowed to be wider than the element type. 3006 // But if NewEltVT is smaller that EltVT the BUILD_VECTOR does not accept 3007 // it. 3008 if (NewEltVT.bitsLT(EltVT)) { 3009 3010 // Convert shuffle node. 3011 // If original node was v4i64 and the new EltVT is i32, 3012 // cast operands to v8i32 and re-build the mask. 3013 3014 // Calculate new VT, the size of the new VT should be equal to original. 3015 EVT NewVT = 3016 EVT::getVectorVT(*DAG.getContext(), NewEltVT, 3017 VT.getSizeInBits() / NewEltVT.getSizeInBits()); 3018 assert(NewVT.bitsEq(VT)); 3019 3020 // cast operands to new VT 3021 Op0 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op0); 3022 Op1 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op1); 3023 3024 // Convert the shuffle mask 3025 unsigned int factor = 3026 NewVT.getVectorNumElements()/VT.getVectorNumElements(); 3027 3028 // EltVT gets smaller 3029 assert(factor > 0); 3030 3031 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) { 3032 if (Mask[i] < 0) { 3033 for (unsigned fi = 0; fi < factor; ++fi) 3034 NewMask.push_back(Mask[i]); 3035 } 3036 else { 3037 for (unsigned fi = 0; fi < factor; ++fi) 3038 NewMask.push_back(Mask[i]*factor+fi); 3039 } 3040 } 3041 Mask = NewMask; 3042 VT = NewVT; 3043 } 3044 EltVT = NewEltVT; 3045 } 3046 unsigned NumElems = VT.getVectorNumElements(); 3047 SmallVector<SDValue, 16> Ops; 3048 for (unsigned i = 0; i != NumElems; ++i) { 3049 if (Mask[i] < 0) { 3050 Ops.push_back(DAG.getUNDEF(EltVT)); 3051 continue; 3052 } 3053 unsigned Idx = Mask[i]; 3054 if (Idx < NumElems) 3055 Ops.push_back(DAG.getNode( 3056 ISD::EXTRACT_VECTOR_ELT, dl, EltVT, Op0, 3057 DAG.getConstant(Idx, dl, TLI.getVectorIdxTy(DAG.getDataLayout())))); 3058 else 3059 Ops.push_back(DAG.getNode( 3060 ISD::EXTRACT_VECTOR_ELT, dl, EltVT, Op1, 3061 DAG.getConstant(Idx - NumElems, dl, 3062 TLI.getVectorIdxTy(DAG.getDataLayout())))); 3063 } 3064 3065 Tmp1 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops); 3066 // We may have changed the BUILD_VECTOR type. Cast it back to the Node type. 3067 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0), Tmp1); 3068 Results.push_back(Tmp1); 3069 break; 3070 } 3071 case ISD::EXTRACT_ELEMENT: { 3072 EVT OpTy = Node->getOperand(0).getValueType(); 3073 if (cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue()) { 3074 // 1 -> Hi 3075 Tmp1 = DAG.getNode(ISD::SRL, dl, OpTy, Node->getOperand(0), 3076 DAG.getConstant(OpTy.getSizeInBits() / 2, dl, 3077 TLI.getShiftAmountTy( 3078 Node->getOperand(0).getValueType(), 3079 DAG.getDataLayout()))); 3080 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0), Tmp1); 3081 } else { 3082 // 0 -> Lo 3083 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0), 3084 Node->getOperand(0)); 3085 } 3086 Results.push_back(Tmp1); 3087 break; 3088 } 3089 case ISD::STACKSAVE: 3090 // Expand to CopyFromReg if the target set 3091 // StackPointerRegisterToSaveRestore. 3092 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) { 3093 Results.push_back(DAG.getCopyFromReg(Node->getOperand(0), dl, SP, 3094 Node->getValueType(0))); 3095 Results.push_back(Results[0].getValue(1)); 3096 } else { 3097 Results.push_back(DAG.getUNDEF(Node->getValueType(0))); 3098 Results.push_back(Node->getOperand(0)); 3099 } 3100 break; 3101 case ISD::STACKRESTORE: 3102 // Expand to CopyToReg if the target set 3103 // StackPointerRegisterToSaveRestore. 3104 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) { 3105 Results.push_back(DAG.getCopyToReg(Node->getOperand(0), dl, SP, 3106 Node->getOperand(1))); 3107 } else { 3108 Results.push_back(Node->getOperand(0)); 3109 } 3110 break; 3111 case ISD::GET_DYNAMIC_AREA_OFFSET: 3112 Results.push_back(DAG.getConstant(0, dl, Node->getValueType(0))); 3113 Results.push_back(Results[0].getValue(0)); 3114 break; 3115 case ISD::FCOPYSIGN: 3116 Results.push_back(ExpandFCOPYSIGN(Node)); 3117 break; 3118 case ISD::FNEG: 3119 // Expand Y = FNEG(X) -> Y = SUB -0.0, X 3120 Tmp1 = DAG.getConstantFP(-0.0, dl, Node->getValueType(0)); 3121 // TODO: If FNEG has fast-math-flags, propagate them to the FSUB. 3122 Tmp1 = DAG.getNode(ISD::FSUB, dl, Node->getValueType(0), Tmp1, 3123 Node->getOperand(0)); 3124 Results.push_back(Tmp1); 3125 break; 3126 case ISD::FABS: 3127 Results.push_back(ExpandFABS(Node)); 3128 break; 3129 case ISD::SMIN: 3130 case ISD::SMAX: 3131 case ISD::UMIN: 3132 case ISD::UMAX: { 3133 // Expand Y = MAX(A, B) -> Y = (A > B) ? A : B 3134 ISD::CondCode Pred; 3135 switch (Node->getOpcode()) { 3136 default: llvm_unreachable("How did we get here?"); 3137 case ISD::SMAX: Pred = ISD::SETGT; break; 3138 case ISD::SMIN: Pred = ISD::SETLT; break; 3139 case ISD::UMAX: Pred = ISD::SETUGT; break; 3140 case ISD::UMIN: Pred = ISD::SETULT; break; 3141 } 3142 Tmp1 = Node->getOperand(0); 3143 Tmp2 = Node->getOperand(1); 3144 Tmp1 = DAG.getSelectCC(dl, Tmp1, Tmp2, Tmp1, Tmp2, Pred); 3145 Results.push_back(Tmp1); 3146 break; 3147 } 3148 3149 case ISD::FSIN: 3150 case ISD::FCOS: { 3151 EVT VT = Node->getValueType(0); 3152 // Turn fsin / fcos into ISD::FSINCOS node if there are a pair of fsin / 3153 // fcos which share the same operand and both are used. 3154 if ((TLI.isOperationLegalOrCustom(ISD::FSINCOS, VT) || 3155 canCombineSinCosLibcall(Node, TLI, TM)) 3156 && useSinCos(Node)) { 3157 SDVTList VTs = DAG.getVTList(VT, VT); 3158 Tmp1 = DAG.getNode(ISD::FSINCOS, dl, VTs, Node->getOperand(0)); 3159 if (Node->getOpcode() == ISD::FCOS) 3160 Tmp1 = Tmp1.getValue(1); 3161 Results.push_back(Tmp1); 3162 } 3163 break; 3164 } 3165 case ISD::FMAD: 3166 llvm_unreachable("Illegal fmad should never be formed"); 3167 3168 case ISD::FP16_TO_FP: 3169 if (Node->getValueType(0) != MVT::f32) { 3170 // We can extend to types bigger than f32 in two steps without changing 3171 // the result. Since "f16 -> f32" is much more commonly available, give 3172 // CodeGen the option of emitting that before resorting to a libcall. 3173 SDValue Res = 3174 DAG.getNode(ISD::FP16_TO_FP, dl, MVT::f32, Node->getOperand(0)); 3175 Results.push_back( 3176 DAG.getNode(ISD::FP_EXTEND, dl, Node->getValueType(0), Res)); 3177 } 3178 break; 3179 case ISD::FP_TO_FP16: 3180 if (!TLI.useSoftFloat() && TM.Options.UnsafeFPMath) { 3181 SDValue Op = Node->getOperand(0); 3182 MVT SVT = Op.getSimpleValueType(); 3183 if ((SVT == MVT::f64 || SVT == MVT::f80) && 3184 TLI.isOperationLegalOrCustom(ISD::FP_TO_FP16, MVT::f32)) { 3185 // Under fastmath, we can expand this node into a fround followed by 3186 // a float-half conversion. 3187 SDValue FloatVal = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Op, 3188 DAG.getIntPtrConstant(0, dl)); 3189 Results.push_back( 3190 DAG.getNode(ISD::FP_TO_FP16, dl, Node->getValueType(0), FloatVal)); 3191 } 3192 } 3193 break; 3194 case ISD::ConstantFP: { 3195 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Node); 3196 // Check to see if this FP immediate is already legal. 3197 // If this is a legal constant, turn it into a TargetConstantFP node. 3198 if (!TLI.isFPImmLegal(CFP->getValueAPF(), Node->getValueType(0))) 3199 Results.push_back(ExpandConstantFP(CFP, true)); 3200 break; 3201 } 3202 case ISD::Constant: { 3203 ConstantSDNode *CP = cast<ConstantSDNode>(Node); 3204 Results.push_back(ExpandConstant(CP)); 3205 break; 3206 } 3207 case ISD::FSUB: { 3208 EVT VT = Node->getValueType(0); 3209 if (TLI.isOperationLegalOrCustom(ISD::FADD, VT) && 3210 TLI.isOperationLegalOrCustom(ISD::FNEG, VT)) { 3211 const SDNodeFlags *Flags = &cast<BinaryWithFlagsSDNode>(Node)->Flags; 3212 Tmp1 = DAG.getNode(ISD::FNEG, dl, VT, Node->getOperand(1)); 3213 Tmp1 = DAG.getNode(ISD::FADD, dl, VT, Node->getOperand(0), Tmp1, Flags); 3214 Results.push_back(Tmp1); 3215 } 3216 break; 3217 } 3218 case ISD::SUB: { 3219 EVT VT = Node->getValueType(0); 3220 assert(TLI.isOperationLegalOrCustom(ISD::ADD, VT) && 3221 TLI.isOperationLegalOrCustom(ISD::XOR, VT) && 3222 "Don't know how to expand this subtraction!"); 3223 Tmp1 = DAG.getNode(ISD::XOR, dl, VT, Node->getOperand(1), 3224 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), dl, 3225 VT)); 3226 Tmp1 = DAG.getNode(ISD::ADD, dl, VT, Tmp1, DAG.getConstant(1, dl, VT)); 3227 Results.push_back(DAG.getNode(ISD::ADD, dl, VT, Node->getOperand(0), Tmp1)); 3228 break; 3229 } 3230 case ISD::UREM: 3231 case ISD::SREM: { 3232 EVT VT = Node->getValueType(0); 3233 bool isSigned = Node->getOpcode() == ISD::SREM; 3234 unsigned DivOpc = isSigned ? ISD::SDIV : ISD::UDIV; 3235 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM; 3236 Tmp2 = Node->getOperand(0); 3237 Tmp3 = Node->getOperand(1); 3238 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT)) { 3239 SDVTList VTs = DAG.getVTList(VT, VT); 3240 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Tmp2, Tmp3).getValue(1); 3241 Results.push_back(Tmp1); 3242 } else if (TLI.isOperationLegalOrCustom(DivOpc, VT)) { 3243 // X % Y -> X-X/Y*Y 3244 Tmp1 = DAG.getNode(DivOpc, dl, VT, Tmp2, Tmp3); 3245 Tmp1 = DAG.getNode(ISD::MUL, dl, VT, Tmp1, Tmp3); 3246 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, Tmp2, Tmp1); 3247 Results.push_back(Tmp1); 3248 } 3249 break; 3250 } 3251 case ISD::UDIV: 3252 case ISD::SDIV: { 3253 bool isSigned = Node->getOpcode() == ISD::SDIV; 3254 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM; 3255 EVT VT = Node->getValueType(0); 3256 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT)) { 3257 SDVTList VTs = DAG.getVTList(VT, VT); 3258 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Node->getOperand(0), 3259 Node->getOperand(1)); 3260 Results.push_back(Tmp1); 3261 } 3262 break; 3263 } 3264 case ISD::MULHU: 3265 case ISD::MULHS: { 3266 unsigned ExpandOpcode = Node->getOpcode() == ISD::MULHU ? ISD::UMUL_LOHI : 3267 ISD::SMUL_LOHI; 3268 EVT VT = Node->getValueType(0); 3269 SDVTList VTs = DAG.getVTList(VT, VT); 3270 assert(TLI.isOperationLegalOrCustom(ExpandOpcode, VT) && 3271 "If this wasn't legal, it shouldn't have been created!"); 3272 Tmp1 = DAG.getNode(ExpandOpcode, dl, VTs, Node->getOperand(0), 3273 Node->getOperand(1)); 3274 Results.push_back(Tmp1.getValue(1)); 3275 break; 3276 } 3277 case ISD::MUL: { 3278 EVT VT = Node->getValueType(0); 3279 SDVTList VTs = DAG.getVTList(VT, VT); 3280 // See if multiply or divide can be lowered using two-result operations. 3281 // We just need the low half of the multiply; try both the signed 3282 // and unsigned forms. If the target supports both SMUL_LOHI and 3283 // UMUL_LOHI, form a preference by checking which forms of plain 3284 // MULH it supports. 3285 bool HasSMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::SMUL_LOHI, VT); 3286 bool HasUMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::UMUL_LOHI, VT); 3287 bool HasMULHS = TLI.isOperationLegalOrCustom(ISD::MULHS, VT); 3288 bool HasMULHU = TLI.isOperationLegalOrCustom(ISD::MULHU, VT); 3289 unsigned OpToUse = 0; 3290 if (HasSMUL_LOHI && !HasMULHS) { 3291 OpToUse = ISD::SMUL_LOHI; 3292 } else if (HasUMUL_LOHI && !HasMULHU) { 3293 OpToUse = ISD::UMUL_LOHI; 3294 } else if (HasSMUL_LOHI) { 3295 OpToUse = ISD::SMUL_LOHI; 3296 } else if (HasUMUL_LOHI) { 3297 OpToUse = ISD::UMUL_LOHI; 3298 } 3299 if (OpToUse) { 3300 Results.push_back(DAG.getNode(OpToUse, dl, VTs, Node->getOperand(0), 3301 Node->getOperand(1))); 3302 break; 3303 } 3304 3305 SDValue Lo, Hi; 3306 EVT HalfType = VT.getHalfSizedIntegerVT(*DAG.getContext()); 3307 if (TLI.isOperationLegalOrCustom(ISD::ZERO_EXTEND, VT) && 3308 TLI.isOperationLegalOrCustom(ISD::ANY_EXTEND, VT) && 3309 TLI.isOperationLegalOrCustom(ISD::SHL, VT) && 3310 TLI.isOperationLegalOrCustom(ISD::OR, VT) && 3311 TLI.expandMUL(Node, Lo, Hi, HalfType, DAG)) { 3312 Lo = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Lo); 3313 Hi = DAG.getNode(ISD::ANY_EXTEND, dl, VT, Hi); 3314 SDValue Shift = 3315 DAG.getConstant(HalfType.getSizeInBits(), dl, 3316 TLI.getShiftAmountTy(HalfType, DAG.getDataLayout())); 3317 Hi = DAG.getNode(ISD::SHL, dl, VT, Hi, Shift); 3318 Results.push_back(DAG.getNode(ISD::OR, dl, VT, Lo, Hi)); 3319 } 3320 break; 3321 } 3322 case ISD::SADDO: 3323 case ISD::SSUBO: { 3324 SDValue LHS = Node->getOperand(0); 3325 SDValue RHS = Node->getOperand(1); 3326 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::SADDO ? 3327 ISD::ADD : ISD::SUB, dl, LHS.getValueType(), 3328 LHS, RHS); 3329 Results.push_back(Sum); 3330 EVT ResultType = Node->getValueType(1); 3331 EVT OType = getSetCCResultType(Node->getValueType(0)); 3332 3333 SDValue Zero = DAG.getConstant(0, dl, LHS.getValueType()); 3334 3335 // LHSSign -> LHS >= 0 3336 // RHSSign -> RHS >= 0 3337 // SumSign -> Sum >= 0 3338 // 3339 // Add: 3340 // Overflow -> (LHSSign == RHSSign) && (LHSSign != SumSign) 3341 // Sub: 3342 // Overflow -> (LHSSign != RHSSign) && (LHSSign != SumSign) 3343 // 3344 SDValue LHSSign = DAG.getSetCC(dl, OType, LHS, Zero, ISD::SETGE); 3345 SDValue RHSSign = DAG.getSetCC(dl, OType, RHS, Zero, ISD::SETGE); 3346 SDValue SignsMatch = DAG.getSetCC(dl, OType, LHSSign, RHSSign, 3347 Node->getOpcode() == ISD::SADDO ? 3348 ISD::SETEQ : ISD::SETNE); 3349 3350 SDValue SumSign = DAG.getSetCC(dl, OType, Sum, Zero, ISD::SETGE); 3351 SDValue SumSignNE = DAG.getSetCC(dl, OType, LHSSign, SumSign, ISD::SETNE); 3352 3353 SDValue Cmp = DAG.getNode(ISD::AND, dl, OType, SignsMatch, SumSignNE); 3354 Results.push_back(DAG.getBoolExtOrTrunc(Cmp, dl, ResultType, ResultType)); 3355 break; 3356 } 3357 case ISD::UADDO: 3358 case ISD::USUBO: { 3359 SDValue LHS = Node->getOperand(0); 3360 SDValue RHS = Node->getOperand(1); 3361 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::UADDO ? 3362 ISD::ADD : ISD::SUB, dl, LHS.getValueType(), 3363 LHS, RHS); 3364 Results.push_back(Sum); 3365 3366 EVT ResultType = Node->getValueType(1); 3367 EVT SetCCType = getSetCCResultType(Node->getValueType(0)); 3368 ISD::CondCode CC 3369 = Node->getOpcode() == ISD::UADDO ? ISD::SETULT : ISD::SETUGT; 3370 SDValue SetCC = DAG.getSetCC(dl, SetCCType, Sum, LHS, CC); 3371 3372 Results.push_back(DAG.getBoolExtOrTrunc(SetCC, dl, ResultType, ResultType)); 3373 break; 3374 } 3375 case ISD::UMULO: 3376 case ISD::SMULO: { 3377 EVT VT = Node->getValueType(0); 3378 EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits() * 2); 3379 SDValue LHS = Node->getOperand(0); 3380 SDValue RHS = Node->getOperand(1); 3381 SDValue BottomHalf; 3382 SDValue TopHalf; 3383 static const unsigned Ops[2][3] = 3384 { { ISD::MULHU, ISD::UMUL_LOHI, ISD::ZERO_EXTEND }, 3385 { ISD::MULHS, ISD::SMUL_LOHI, ISD::SIGN_EXTEND }}; 3386 bool isSigned = Node->getOpcode() == ISD::SMULO; 3387 if (TLI.isOperationLegalOrCustom(Ops[isSigned][0], VT)) { 3388 BottomHalf = DAG.getNode(ISD::MUL, dl, VT, LHS, RHS); 3389 TopHalf = DAG.getNode(Ops[isSigned][0], dl, VT, LHS, RHS); 3390 } else if (TLI.isOperationLegalOrCustom(Ops[isSigned][1], VT)) { 3391 BottomHalf = DAG.getNode(Ops[isSigned][1], dl, DAG.getVTList(VT, VT), LHS, 3392 RHS); 3393 TopHalf = BottomHalf.getValue(1); 3394 } else if (TLI.isTypeLegal(WideVT)) { 3395 LHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, LHS); 3396 RHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, RHS); 3397 Tmp1 = DAG.getNode(ISD::MUL, dl, WideVT, LHS, RHS); 3398 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1, 3399 DAG.getIntPtrConstant(0, dl)); 3400 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1, 3401 DAG.getIntPtrConstant(1, dl)); 3402 } else { 3403 // We can fall back to a libcall with an illegal type for the MUL if we 3404 // have a libcall big enough. 3405 // Also, we can fall back to a division in some cases, but that's a big 3406 // performance hit in the general case. 3407 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL; 3408 if (WideVT == MVT::i16) 3409 LC = RTLIB::MUL_I16; 3410 else if (WideVT == MVT::i32) 3411 LC = RTLIB::MUL_I32; 3412 else if (WideVT == MVT::i64) 3413 LC = RTLIB::MUL_I64; 3414 else if (WideVT == MVT::i128) 3415 LC = RTLIB::MUL_I128; 3416 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Cannot expand this operation!"); 3417 3418 // The high part is obtained by SRA'ing all but one of the bits of low 3419 // part. 3420 unsigned LoSize = VT.getSizeInBits(); 3421 SDValue HiLHS = 3422 DAG.getNode(ISD::SRA, dl, VT, RHS, 3423 DAG.getConstant(LoSize - 1, dl, 3424 TLI.getPointerTy(DAG.getDataLayout()))); 3425 SDValue HiRHS = 3426 DAG.getNode(ISD::SRA, dl, VT, LHS, 3427 DAG.getConstant(LoSize - 1, dl, 3428 TLI.getPointerTy(DAG.getDataLayout()))); 3429 3430 // Here we're passing the 2 arguments explicitly as 4 arguments that are 3431 // pre-lowered to the correct types. This all depends upon WideVT not 3432 // being a legal type for the architecture and thus has to be split to 3433 // two arguments. 3434 SDValue Args[] = { LHS, HiLHS, RHS, HiRHS }; 3435 SDValue Ret = ExpandLibCall(LC, WideVT, Args, 4, isSigned, dl); 3436 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret, 3437 DAG.getIntPtrConstant(0, dl)); 3438 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret, 3439 DAG.getIntPtrConstant(1, dl)); 3440 // Ret is a node with an illegal type. Because such things are not 3441 // generally permitted during this phase of legalization, make sure the 3442 // node has no more uses. The above EXTRACT_ELEMENT nodes should have been 3443 // folded. 3444 assert(Ret->use_empty() && 3445 "Unexpected uses of illegally type from expanded lib call."); 3446 } 3447 3448 if (isSigned) { 3449 Tmp1 = DAG.getConstant( 3450 VT.getSizeInBits() - 1, dl, 3451 TLI.getShiftAmountTy(BottomHalf.getValueType(), DAG.getDataLayout())); 3452 Tmp1 = DAG.getNode(ISD::SRA, dl, VT, BottomHalf, Tmp1); 3453 TopHalf = DAG.getSetCC(dl, getSetCCResultType(VT), TopHalf, Tmp1, 3454 ISD::SETNE); 3455 } else { 3456 TopHalf = DAG.getSetCC(dl, getSetCCResultType(VT), TopHalf, 3457 DAG.getConstant(0, dl, VT), ISD::SETNE); 3458 } 3459 Results.push_back(BottomHalf); 3460 Results.push_back(TopHalf); 3461 break; 3462 } 3463 case ISD::BUILD_PAIR: { 3464 EVT PairTy = Node->getValueType(0); 3465 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, PairTy, Node->getOperand(0)); 3466 Tmp2 = DAG.getNode(ISD::ANY_EXTEND, dl, PairTy, Node->getOperand(1)); 3467 Tmp2 = DAG.getNode( 3468 ISD::SHL, dl, PairTy, Tmp2, 3469 DAG.getConstant(PairTy.getSizeInBits() / 2, dl, 3470 TLI.getShiftAmountTy(PairTy, DAG.getDataLayout()))); 3471 Results.push_back(DAG.getNode(ISD::OR, dl, PairTy, Tmp1, Tmp2)); 3472 break; 3473 } 3474 case ISD::SELECT: 3475 Tmp1 = Node->getOperand(0); 3476 Tmp2 = Node->getOperand(1); 3477 Tmp3 = Node->getOperand(2); 3478 if (Tmp1.getOpcode() == ISD::SETCC) { 3479 Tmp1 = DAG.getSelectCC(dl, Tmp1.getOperand(0), Tmp1.getOperand(1), 3480 Tmp2, Tmp3, 3481 cast<CondCodeSDNode>(Tmp1.getOperand(2))->get()); 3482 } else { 3483 Tmp1 = DAG.getSelectCC(dl, Tmp1, 3484 DAG.getConstant(0, dl, Tmp1.getValueType()), 3485 Tmp2, Tmp3, ISD::SETNE); 3486 } 3487 Results.push_back(Tmp1); 3488 break; 3489 case ISD::BR_JT: { 3490 SDValue Chain = Node->getOperand(0); 3491 SDValue Table = Node->getOperand(1); 3492 SDValue Index = Node->getOperand(2); 3493 3494 EVT PTy = TLI.getPointerTy(DAG.getDataLayout()); 3495 3496 const DataLayout &TD = DAG.getDataLayout(); 3497 unsigned EntrySize = 3498 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(TD); 3499 3500 Index = DAG.getNode(ISD::MUL, dl, Index.getValueType(), Index, 3501 DAG.getConstant(EntrySize, dl, Index.getValueType())); 3502 SDValue Addr = DAG.getNode(ISD::ADD, dl, Index.getValueType(), 3503 Index, Table); 3504 3505 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8); 3506 SDValue LD = DAG.getExtLoad( 3507 ISD::SEXTLOAD, dl, PTy, Chain, Addr, 3508 MachinePointerInfo::getJumpTable(DAG.getMachineFunction()), MemVT, 3509 false, false, false, 0); 3510 Addr = LD; 3511 if (TM.isPositionIndependent()) { 3512 // For PIC, the sequence is: 3513 // BRIND(load(Jumptable + index) + RelocBase) 3514 // RelocBase can be JumpTable, GOT or some sort of global base. 3515 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, 3516 TLI.getPICJumpTableRelocBase(Table, DAG)); 3517 } 3518 Tmp1 = DAG.getNode(ISD::BRIND, dl, MVT::Other, LD.getValue(1), Addr); 3519 Results.push_back(Tmp1); 3520 break; 3521 } 3522 case ISD::BRCOND: 3523 // Expand brcond's setcc into its constituent parts and create a BR_CC 3524 // Node. 3525 Tmp1 = Node->getOperand(0); 3526 Tmp2 = Node->getOperand(1); 3527 if (Tmp2.getOpcode() == ISD::SETCC) { 3528 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other, 3529 Tmp1, Tmp2.getOperand(2), 3530 Tmp2.getOperand(0), Tmp2.getOperand(1), 3531 Node->getOperand(2)); 3532 } else { 3533 // We test only the i1 bit. Skip the AND if UNDEF. 3534 Tmp3 = (Tmp2.isUndef()) ? Tmp2 : 3535 DAG.getNode(ISD::AND, dl, Tmp2.getValueType(), Tmp2, 3536 DAG.getConstant(1, dl, Tmp2.getValueType())); 3537 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other, Tmp1, 3538 DAG.getCondCode(ISD::SETNE), Tmp3, 3539 DAG.getConstant(0, dl, Tmp3.getValueType()), 3540 Node->getOperand(2)); 3541 } 3542 Results.push_back(Tmp1); 3543 break; 3544 case ISD::SETCC: { 3545 Tmp1 = Node->getOperand(0); 3546 Tmp2 = Node->getOperand(1); 3547 Tmp3 = Node->getOperand(2); 3548 bool Legalized = LegalizeSetCCCondCode(Node->getValueType(0), Tmp1, Tmp2, 3549 Tmp3, NeedInvert, dl); 3550 3551 if (Legalized) { 3552 // If we expanded the SETCC by swapping LHS and RHS, or by inverting the 3553 // condition code, create a new SETCC node. 3554 if (Tmp3.getNode()) 3555 Tmp1 = DAG.getNode(ISD::SETCC, dl, Node->getValueType(0), 3556 Tmp1, Tmp2, Tmp3); 3557 3558 // If we expanded the SETCC by inverting the condition code, then wrap 3559 // the existing SETCC in a NOT to restore the intended condition. 3560 if (NeedInvert) 3561 Tmp1 = DAG.getLogicalNOT(dl, Tmp1, Tmp1->getValueType(0)); 3562 3563 Results.push_back(Tmp1); 3564 break; 3565 } 3566 3567 // Otherwise, SETCC for the given comparison type must be completely 3568 // illegal; expand it into a SELECT_CC. 3569 EVT VT = Node->getValueType(0); 3570 int TrueValue; 3571 switch (TLI.getBooleanContents(Tmp1->getValueType(0))) { 3572 case TargetLowering::ZeroOrOneBooleanContent: 3573 case TargetLowering::UndefinedBooleanContent: 3574 TrueValue = 1; 3575 break; 3576 case TargetLowering::ZeroOrNegativeOneBooleanContent: 3577 TrueValue = -1; 3578 break; 3579 } 3580 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, VT, Tmp1, Tmp2, 3581 DAG.getConstant(TrueValue, dl, VT), 3582 DAG.getConstant(0, dl, VT), 3583 Tmp3); 3584 Results.push_back(Tmp1); 3585 break; 3586 } 3587 case ISD::SELECT_CC: { 3588 Tmp1 = Node->getOperand(0); // LHS 3589 Tmp2 = Node->getOperand(1); // RHS 3590 Tmp3 = Node->getOperand(2); // True 3591 Tmp4 = Node->getOperand(3); // False 3592 EVT VT = Node->getValueType(0); 3593 SDValue CC = Node->getOperand(4); 3594 ISD::CondCode CCOp = cast<CondCodeSDNode>(CC)->get(); 3595 3596 if (TLI.isCondCodeLegal(CCOp, Tmp1.getSimpleValueType())) { 3597 // If the condition code is legal, then we need to expand this 3598 // node using SETCC and SELECT. 3599 EVT CmpVT = Tmp1.getValueType(); 3600 assert(!TLI.isOperationExpand(ISD::SELECT, VT) && 3601 "Cannot expand ISD::SELECT_CC when ISD::SELECT also needs to be " 3602 "expanded."); 3603 EVT CCVT = 3604 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), CmpVT); 3605 SDValue Cond = DAG.getNode(ISD::SETCC, dl, CCVT, Tmp1, Tmp2, CC); 3606 Results.push_back(DAG.getSelect(dl, VT, Cond, Tmp3, Tmp4)); 3607 break; 3608 } 3609 3610 // SELECT_CC is legal, so the condition code must not be. 3611 bool Legalized = false; 3612 // Try to legalize by inverting the condition. This is for targets that 3613 // might support an ordered version of a condition, but not the unordered 3614 // version (or vice versa). 3615 ISD::CondCode InvCC = ISD::getSetCCInverse(CCOp, 3616 Tmp1.getValueType().isInteger()); 3617 if (TLI.isCondCodeLegal(InvCC, Tmp1.getSimpleValueType())) { 3618 // Use the new condition code and swap true and false 3619 Legalized = true; 3620 Tmp1 = DAG.getSelectCC(dl, Tmp1, Tmp2, Tmp4, Tmp3, InvCC); 3621 } else { 3622 // If The inverse is not legal, then try to swap the arguments using 3623 // the inverse condition code. 3624 ISD::CondCode SwapInvCC = ISD::getSetCCSwappedOperands(InvCC); 3625 if (TLI.isCondCodeLegal(SwapInvCC, Tmp1.getSimpleValueType())) { 3626 // The swapped inverse condition is legal, so swap true and false, 3627 // lhs and rhs. 3628 Legalized = true; 3629 Tmp1 = DAG.getSelectCC(dl, Tmp2, Tmp1, Tmp4, Tmp3, SwapInvCC); 3630 } 3631 } 3632 3633 if (!Legalized) { 3634 Legalized = LegalizeSetCCCondCode( 3635 getSetCCResultType(Tmp1.getValueType()), Tmp1, Tmp2, CC, NeedInvert, 3636 dl); 3637 3638 assert(Legalized && "Can't legalize SELECT_CC with legal condition!"); 3639 3640 // If we expanded the SETCC by inverting the condition code, then swap 3641 // the True/False operands to match. 3642 if (NeedInvert) 3643 std::swap(Tmp3, Tmp4); 3644 3645 // If we expanded the SETCC by swapping LHS and RHS, or by inverting the 3646 // condition code, create a new SELECT_CC node. 3647 if (CC.getNode()) { 3648 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0), 3649 Tmp1, Tmp2, Tmp3, Tmp4, CC); 3650 } else { 3651 Tmp2 = DAG.getConstant(0, dl, Tmp1.getValueType()); 3652 CC = DAG.getCondCode(ISD::SETNE); 3653 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0), Tmp1, 3654 Tmp2, Tmp3, Tmp4, CC); 3655 } 3656 } 3657 Results.push_back(Tmp1); 3658 break; 3659 } 3660 case ISD::BR_CC: { 3661 Tmp1 = Node->getOperand(0); // Chain 3662 Tmp2 = Node->getOperand(2); // LHS 3663 Tmp3 = Node->getOperand(3); // RHS 3664 Tmp4 = Node->getOperand(1); // CC 3665 3666 bool Legalized = LegalizeSetCCCondCode(getSetCCResultType( 3667 Tmp2.getValueType()), Tmp2, Tmp3, Tmp4, NeedInvert, dl); 3668 (void)Legalized; 3669 assert(Legalized && "Can't legalize BR_CC with legal condition!"); 3670 3671 // If we expanded the SETCC by inverting the condition code, then wrap 3672 // the existing SETCC in a NOT to restore the intended condition. 3673 if (NeedInvert) 3674 Tmp4 = DAG.getNOT(dl, Tmp4, Tmp4->getValueType(0)); 3675 3676 // If we expanded the SETCC by swapping LHS and RHS, create a new BR_CC 3677 // node. 3678 if (Tmp4.getNode()) { 3679 Tmp1 = DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), Tmp1, 3680 Tmp4, Tmp2, Tmp3, Node->getOperand(4)); 3681 } else { 3682 Tmp3 = DAG.getConstant(0, dl, Tmp2.getValueType()); 3683 Tmp4 = DAG.getCondCode(ISD::SETNE); 3684 Tmp1 = DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), Tmp1, Tmp4, 3685 Tmp2, Tmp3, Node->getOperand(4)); 3686 } 3687 Results.push_back(Tmp1); 3688 break; 3689 } 3690 case ISD::BUILD_VECTOR: 3691 Results.push_back(ExpandBUILD_VECTOR(Node)); 3692 break; 3693 case ISD::SRA: 3694 case ISD::SRL: 3695 case ISD::SHL: { 3696 // Scalarize vector SRA/SRL/SHL. 3697 EVT VT = Node->getValueType(0); 3698 assert(VT.isVector() && "Unable to legalize non-vector shift"); 3699 assert(TLI.isTypeLegal(VT.getScalarType())&& "Element type must be legal"); 3700 unsigned NumElem = VT.getVectorNumElements(); 3701 3702 SmallVector<SDValue, 8> Scalars; 3703 for (unsigned Idx = 0; Idx < NumElem; Idx++) { 3704 SDValue Ex = DAG.getNode( 3705 ISD::EXTRACT_VECTOR_ELT, dl, VT.getScalarType(), Node->getOperand(0), 3706 DAG.getConstant(Idx, dl, TLI.getVectorIdxTy(DAG.getDataLayout()))); 3707 SDValue Sh = DAG.getNode( 3708 ISD::EXTRACT_VECTOR_ELT, dl, VT.getScalarType(), Node->getOperand(1), 3709 DAG.getConstant(Idx, dl, TLI.getVectorIdxTy(DAG.getDataLayout()))); 3710 Scalars.push_back(DAG.getNode(Node->getOpcode(), dl, 3711 VT.getScalarType(), Ex, Sh)); 3712 } 3713 SDValue Result = 3714 DAG.getNode(ISD::BUILD_VECTOR, dl, Node->getValueType(0), Scalars); 3715 ReplaceNode(SDValue(Node, 0), Result); 3716 break; 3717 } 3718 case ISD::GLOBAL_OFFSET_TABLE: 3719 case ISD::GlobalAddress: 3720 case ISD::GlobalTLSAddress: 3721 case ISD::ExternalSymbol: 3722 case ISD::ConstantPool: 3723 case ISD::JumpTable: 3724 case ISD::INTRINSIC_W_CHAIN: 3725 case ISD::INTRINSIC_WO_CHAIN: 3726 case ISD::INTRINSIC_VOID: 3727 // FIXME: Custom lowering for these operations shouldn't return null! 3728 break; 3729 } 3730 3731 // Replace the original node with the legalized result. 3732 if (Results.empty()) 3733 return false; 3734 3735 ReplaceNode(Node, Results.data()); 3736 return true; 3737} 3738 3739void SelectionDAGLegalize::ConvertNodeToLibcall(SDNode *Node) { 3740 SmallVector<SDValue, 8> Results; 3741 SDLoc dl(Node); 3742 SDValue Tmp1, Tmp2, Tmp3, Tmp4; 3743 unsigned Opc = Node->getOpcode(); 3744 switch (Opc) { 3745 case ISD::ATOMIC_FENCE: { 3746 // If the target didn't lower this, lower it to '__sync_synchronize()' call 3747 // FIXME: handle "fence singlethread" more efficiently. 3748 TargetLowering::ArgListTy Args; 3749 3750 TargetLowering::CallLoweringInfo CLI(DAG); 3751 CLI.setDebugLoc(dl) 3752 .setChain(Node->getOperand(0)) 3753 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()), 3754 DAG.getExternalSymbol("__sync_synchronize", 3755 TLI.getPointerTy(DAG.getDataLayout())), 3756 std::move(Args)); 3757 3758 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI); 3759 3760 Results.push_back(CallResult.second); 3761 break; 3762 } 3763 // By default, atomic intrinsics are marked Legal and lowered. Targets 3764 // which don't support them directly, however, may want libcalls, in which 3765 // case they mark them Expand, and we get here. 3766 case ISD::ATOMIC_SWAP: 3767 case ISD::ATOMIC_LOAD_ADD: 3768 case ISD::ATOMIC_LOAD_SUB: 3769 case ISD::ATOMIC_LOAD_AND: 3770 case ISD::ATOMIC_LOAD_OR: 3771 case ISD::ATOMIC_LOAD_XOR: 3772 case ISD::ATOMIC_LOAD_NAND: 3773 case ISD::ATOMIC_LOAD_MIN: 3774 case ISD::ATOMIC_LOAD_MAX: 3775 case ISD::ATOMIC_LOAD_UMIN: 3776 case ISD::ATOMIC_LOAD_UMAX: 3777 case ISD::ATOMIC_CMP_SWAP: { 3778 MVT VT = cast<AtomicSDNode>(Node)->getMemoryVT().getSimpleVT(); 3779 RTLIB::Libcall LC = RTLIB::getSYNC(Opc, VT); 3780 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unexpected atomic op or value type!"); 3781 3782 std::pair<SDValue, SDValue> Tmp = ExpandChainLibCall(LC, Node, false); 3783 Results.push_back(Tmp.first); 3784 Results.push_back(Tmp.second); 3785 break; 3786 } 3787 case ISD::TRAP: { 3788 // If this operation is not supported, lower it to 'abort()' call 3789 TargetLowering::ArgListTy Args; 3790 TargetLowering::CallLoweringInfo CLI(DAG); 3791 CLI.setDebugLoc(dl) 3792 .setChain(Node->getOperand(0)) 3793 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()), 3794 DAG.getExternalSymbol("abort", 3795 TLI.getPointerTy(DAG.getDataLayout())), 3796 std::move(Args)); 3797 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI); 3798 3799 Results.push_back(CallResult.second); 3800 break; 3801 } 3802 case ISD::FMINNUM: 3803 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMIN_F32, RTLIB::FMIN_F64, 3804 RTLIB::FMIN_F80, RTLIB::FMIN_F128, 3805 RTLIB::FMIN_PPCF128)); 3806 break; 3807 case ISD::FMAXNUM: 3808 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMAX_F32, RTLIB::FMAX_F64, 3809 RTLIB::FMAX_F80, RTLIB::FMAX_F128, 3810 RTLIB::FMAX_PPCF128)); 3811 break; 3812 case ISD::FSQRT: 3813 Results.push_back(ExpandFPLibCall(Node, RTLIB::SQRT_F32, RTLIB::SQRT_F64, 3814 RTLIB::SQRT_F80, RTLIB::SQRT_F128, 3815 RTLIB::SQRT_PPCF128)); 3816 break; 3817 case ISD::FSIN: 3818 Results.push_back(ExpandFPLibCall(Node, RTLIB::SIN_F32, RTLIB::SIN_F64, 3819 RTLIB::SIN_F80, RTLIB::SIN_F128, 3820 RTLIB::SIN_PPCF128)); 3821 break; 3822 case ISD::FCOS: 3823 Results.push_back(ExpandFPLibCall(Node, RTLIB::COS_F32, RTLIB::COS_F64, 3824 RTLIB::COS_F80, RTLIB::COS_F128, 3825 RTLIB::COS_PPCF128)); 3826 break; 3827 case ISD::FSINCOS: 3828 // Expand into sincos libcall. 3829 ExpandSinCosLibCall(Node, Results); 3830 break; 3831 case ISD::FLOG: 3832 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG_F32, RTLIB::LOG_F64, 3833 RTLIB::LOG_F80, RTLIB::LOG_F128, 3834 RTLIB::LOG_PPCF128)); 3835 break; 3836 case ISD::FLOG2: 3837 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG2_F32, RTLIB::LOG2_F64, 3838 RTLIB::LOG2_F80, RTLIB::LOG2_F128, 3839 RTLIB::LOG2_PPCF128)); 3840 break; 3841 case ISD::FLOG10: 3842 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG10_F32, RTLIB::LOG10_F64, 3843 RTLIB::LOG10_F80, RTLIB::LOG10_F128, 3844 RTLIB::LOG10_PPCF128)); 3845 break; 3846 case ISD::FEXP: 3847 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP_F32, RTLIB::EXP_F64, 3848 RTLIB::EXP_F80, RTLIB::EXP_F128, 3849 RTLIB::EXP_PPCF128)); 3850 break; 3851 case ISD::FEXP2: 3852 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP2_F32, RTLIB::EXP2_F64, 3853 RTLIB::EXP2_F80, RTLIB::EXP2_F128, 3854 RTLIB::EXP2_PPCF128)); 3855 break; 3856 case ISD::FTRUNC: 3857 Results.push_back(ExpandFPLibCall(Node, RTLIB::TRUNC_F32, RTLIB::TRUNC_F64, 3858 RTLIB::TRUNC_F80, RTLIB::TRUNC_F128, 3859 RTLIB::TRUNC_PPCF128)); 3860 break; 3861 case ISD::FFLOOR: 3862 Results.push_back(ExpandFPLibCall(Node, RTLIB::FLOOR_F32, RTLIB::FLOOR_F64, 3863 RTLIB::FLOOR_F80, RTLIB::FLOOR_F128, 3864 RTLIB::FLOOR_PPCF128)); 3865 break; 3866 case ISD::FCEIL: 3867 Results.push_back(ExpandFPLibCall(Node, RTLIB::CEIL_F32, RTLIB::CEIL_F64, 3868 RTLIB::CEIL_F80, RTLIB::CEIL_F128, 3869 RTLIB::CEIL_PPCF128)); 3870 break; 3871 case ISD::FRINT: 3872 Results.push_back(ExpandFPLibCall(Node, RTLIB::RINT_F32, RTLIB::RINT_F64, 3873 RTLIB::RINT_F80, RTLIB::RINT_F128, 3874 RTLIB::RINT_PPCF128)); 3875 break; 3876 case ISD::FNEARBYINT: 3877 Results.push_back(ExpandFPLibCall(Node, RTLIB::NEARBYINT_F32, 3878 RTLIB::NEARBYINT_F64, 3879 RTLIB::NEARBYINT_F80, 3880 RTLIB::NEARBYINT_F128, 3881 RTLIB::NEARBYINT_PPCF128)); 3882 break; 3883 case ISD::FROUND: 3884 Results.push_back(ExpandFPLibCall(Node, RTLIB::ROUND_F32, 3885 RTLIB::ROUND_F64, 3886 RTLIB::ROUND_F80, 3887 RTLIB::ROUND_F128, 3888 RTLIB::ROUND_PPCF128)); 3889 break; 3890 case ISD::FPOWI: 3891 Results.push_back(ExpandFPLibCall(Node, RTLIB::POWI_F32, RTLIB::POWI_F64, 3892 RTLIB::POWI_F80, RTLIB::POWI_F128, 3893 RTLIB::POWI_PPCF128)); 3894 break; 3895 case ISD::FPOW: 3896 Results.push_back(ExpandFPLibCall(Node, RTLIB::POW_F32, RTLIB::POW_F64, 3897 RTLIB::POW_F80, RTLIB::POW_F128, 3898 RTLIB::POW_PPCF128)); 3899 break; 3900 case ISD::FDIV: 3901 Results.push_back(ExpandFPLibCall(Node, RTLIB::DIV_F32, RTLIB::DIV_F64, 3902 RTLIB::DIV_F80, RTLIB::DIV_F128, 3903 RTLIB::DIV_PPCF128)); 3904 break; 3905 case ISD::FREM: 3906 Results.push_back(ExpandFPLibCall(Node, RTLIB::REM_F32, RTLIB::REM_F64, 3907 RTLIB::REM_F80, RTLIB::REM_F128, 3908 RTLIB::REM_PPCF128)); 3909 break; 3910 case ISD::FMA: 3911 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMA_F32, RTLIB::FMA_F64, 3912 RTLIB::FMA_F80, RTLIB::FMA_F128, 3913 RTLIB::FMA_PPCF128)); 3914 break; 3915 case ISD::FADD: 3916 Results.push_back(ExpandFPLibCall(Node, RTLIB::ADD_F32, RTLIB::ADD_F64, 3917 RTLIB::ADD_F80, RTLIB::ADD_F128, 3918 RTLIB::ADD_PPCF128)); 3919 break; 3920 case ISD::FMUL: 3921 Results.push_back(ExpandFPLibCall(Node, RTLIB::MUL_F32, RTLIB::MUL_F64, 3922 RTLIB::MUL_F80, RTLIB::MUL_F128, 3923 RTLIB::MUL_PPCF128)); 3924 break; 3925 case ISD::FP16_TO_FP: 3926 if (Node->getValueType(0) == MVT::f32) { 3927 Results.push_back(ExpandLibCall(RTLIB::FPEXT_F16_F32, Node, false)); 3928 } 3929 break; 3930 case ISD::FP_TO_FP16: { 3931 RTLIB::Libcall LC = 3932 RTLIB::getFPROUND(Node->getOperand(0).getValueType(), MVT::f16); 3933 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unable to expand fp_to_fp16"); 3934 Results.push_back(ExpandLibCall(LC, Node, false)); 3935 break; 3936 } 3937 case ISD::FSUB: 3938 Results.push_back(ExpandFPLibCall(Node, RTLIB::SUB_F32, RTLIB::SUB_F64, 3939 RTLIB::SUB_F80, RTLIB::SUB_F128, 3940 RTLIB::SUB_PPCF128)); 3941 break; 3942 case ISD::SREM: 3943 Results.push_back(ExpandIntLibCall(Node, true, 3944 RTLIB::SREM_I8, 3945 RTLIB::SREM_I16, RTLIB::SREM_I32, 3946 RTLIB::SREM_I64, RTLIB::SREM_I128)); 3947 break; 3948 case ISD::UREM: 3949 Results.push_back(ExpandIntLibCall(Node, false, 3950 RTLIB::UREM_I8, 3951 RTLIB::UREM_I16, RTLIB::UREM_I32, 3952 RTLIB::UREM_I64, RTLIB::UREM_I128)); 3953 break; 3954 case ISD::SDIV: 3955 Results.push_back(ExpandIntLibCall(Node, true, 3956 RTLIB::SDIV_I8, 3957 RTLIB::SDIV_I16, RTLIB::SDIV_I32, 3958 RTLIB::SDIV_I64, RTLIB::SDIV_I128)); 3959 break; 3960 case ISD::UDIV: 3961 Results.push_back(ExpandIntLibCall(Node, false, 3962 RTLIB::UDIV_I8, 3963 RTLIB::UDIV_I16, RTLIB::UDIV_I32, 3964 RTLIB::UDIV_I64, RTLIB::UDIV_I128)); 3965 break; 3966 case ISD::SDIVREM: 3967 case ISD::UDIVREM: 3968 // Expand into divrem libcall 3969 ExpandDivRemLibCall(Node, Results); 3970 break; 3971 case ISD::MUL: 3972 Results.push_back(ExpandIntLibCall(Node, false, 3973 RTLIB::MUL_I8, 3974 RTLIB::MUL_I16, RTLIB::MUL_I32, 3975 RTLIB::MUL_I64, RTLIB::MUL_I128)); 3976 break; 3977 } 3978 3979 // Replace the original node with the legalized result. 3980 if (!Results.empty()) 3981 ReplaceNode(Node, Results.data()); 3982} 3983 3984// Determine the vector type to use in place of an original scalar element when 3985// promoting equally sized vectors. 3986static MVT getPromotedVectorElementType(const TargetLowering &TLI, 3987 MVT EltVT, MVT NewEltVT) { 3988 unsigned OldEltsPerNewElt = EltVT.getSizeInBits() / NewEltVT.getSizeInBits(); 3989 MVT MidVT = MVT::getVectorVT(NewEltVT, OldEltsPerNewElt); 3990 assert(TLI.isTypeLegal(MidVT) && "unexpected"); 3991 return MidVT; 3992} 3993 3994void SelectionDAGLegalize::PromoteNode(SDNode *Node) { 3995 SmallVector<SDValue, 8> Results; 3996 MVT OVT = Node->getSimpleValueType(0); 3997 if (Node->getOpcode() == ISD::UINT_TO_FP || 3998 Node->getOpcode() == ISD::SINT_TO_FP || 3999 Node->getOpcode() == ISD::SETCC || 4000 Node->getOpcode() == ISD::EXTRACT_VECTOR_ELT || 4001 Node->getOpcode() == ISD::INSERT_VECTOR_ELT) { 4002 OVT = Node->getOperand(0).getSimpleValueType(); 4003 } 4004 if (Node->getOpcode() == ISD::BR_CC) 4005 OVT = Node->getOperand(2).getSimpleValueType(); 4006 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), OVT); 4007 SDLoc dl(Node); 4008 SDValue Tmp1, Tmp2, Tmp3; 4009 switch (Node->getOpcode()) { 4010 case ISD::CTTZ: 4011 case ISD::CTTZ_ZERO_UNDEF: 4012 case ISD::CTLZ: 4013 case ISD::CTLZ_ZERO_UNDEF: 4014 case ISD::CTPOP: 4015 // Zero extend the argument. 4016 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0)); 4017 if (Node->getOpcode() == ISD::CTTZ) { 4018 // The count is the same in the promoted type except if the original 4019 // value was zero. This can be handled by setting the bit just off 4020 // the top of the original type. 4021 auto TopBit = APInt::getOneBitSet(NVT.getSizeInBits(), 4022 OVT.getSizeInBits()); 4023 Tmp1 = DAG.getNode(ISD::OR, dl, NVT, Tmp1, 4024 DAG.getConstant(TopBit, dl, NVT)); 4025 } 4026 // Perform the larger operation. For CTPOP and CTTZ_ZERO_UNDEF, this is 4027 // already the correct result. 4028 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1); 4029 if (Node->getOpcode() == ISD::CTLZ || 4030 Node->getOpcode() == ISD::CTLZ_ZERO_UNDEF) { 4031 // Tmp1 = Tmp1 - (sizeinbits(NVT) - sizeinbits(Old VT)) 4032 Tmp1 = DAG.getNode(ISD::SUB, dl, NVT, Tmp1, 4033 DAG.getConstant(NVT.getSizeInBits() - 4034 OVT.getSizeInBits(), dl, NVT)); 4035 } 4036 Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, OVT, Tmp1)); 4037 break; 4038 case ISD::BSWAP: { 4039 unsigned DiffBits = NVT.getSizeInBits() - OVT.getSizeInBits(); 4040 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0)); 4041 Tmp1 = DAG.getNode(ISD::BSWAP, dl, NVT, Tmp1); 4042 Tmp1 = DAG.getNode( 4043 ISD::SRL, dl, NVT, Tmp1, 4044 DAG.getConstant(DiffBits, dl, 4045 TLI.getShiftAmountTy(NVT, DAG.getDataLayout()))); 4046 Results.push_back(Tmp1); 4047 break; 4048 } 4049 case ISD::FP_TO_UINT: 4050 case ISD::FP_TO_SINT: 4051 Tmp1 = PromoteLegalFP_TO_INT(Node->getOperand(0), Node->getValueType(0), 4052 Node->getOpcode() == ISD::FP_TO_SINT, dl); 4053 Results.push_back(Tmp1); 4054 break; 4055 case ISD::UINT_TO_FP: 4056 case ISD::SINT_TO_FP: 4057 Tmp1 = PromoteLegalINT_TO_FP(Node->getOperand(0), Node->getValueType(0), 4058 Node->getOpcode() == ISD::SINT_TO_FP, dl); 4059 Results.push_back(Tmp1); 4060 break; 4061 case ISD::VAARG: { 4062 SDValue Chain = Node->getOperand(0); // Get the chain. 4063 SDValue Ptr = Node->getOperand(1); // Get the pointer. 4064 4065 unsigned TruncOp; 4066 if (OVT.isVector()) { 4067 TruncOp = ISD::BITCAST; 4068 } else { 4069 assert(OVT.isInteger() 4070 && "VAARG promotion is supported only for vectors or integer types"); 4071 TruncOp = ISD::TRUNCATE; 4072 } 4073 4074 // Perform the larger operation, then convert back 4075 Tmp1 = DAG.getVAArg(NVT, dl, Chain, Ptr, Node->getOperand(2), 4076 Node->getConstantOperandVal(3)); 4077 Chain = Tmp1.getValue(1); 4078 4079 Tmp2 = DAG.getNode(TruncOp, dl, OVT, Tmp1); 4080 4081 // Modified the chain result - switch anything that used the old chain to 4082 // use the new one. 4083 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Tmp2); 4084 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain); 4085 if (UpdatedNodes) { 4086 UpdatedNodes->insert(Tmp2.getNode()); 4087 UpdatedNodes->insert(Chain.getNode()); 4088 } 4089 ReplacedNode(Node); 4090 break; 4091 } 4092 case ISD::AND: 4093 case ISD::OR: 4094 case ISD::XOR: { 4095 unsigned ExtOp, TruncOp; 4096 if (OVT.isVector()) { 4097 ExtOp = ISD::BITCAST; 4098 TruncOp = ISD::BITCAST; 4099 } else { 4100 assert(OVT.isInteger() && "Cannot promote logic operation"); 4101 ExtOp = ISD::ANY_EXTEND; 4102 TruncOp = ISD::TRUNCATE; 4103 } 4104 // Promote each of the values to the new type. 4105 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0)); 4106 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1)); 4107 // Perform the larger operation, then convert back 4108 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2); 4109 Results.push_back(DAG.getNode(TruncOp, dl, OVT, Tmp1)); 4110 break; 4111 } 4112 case ISD::SELECT: { 4113 unsigned ExtOp, TruncOp; 4114 if (Node->getValueType(0).isVector() || 4115 Node->getValueType(0).getSizeInBits() == NVT.getSizeInBits()) { 4116 ExtOp = ISD::BITCAST; 4117 TruncOp = ISD::BITCAST; 4118 } else if (Node->getValueType(0).isInteger()) { 4119 ExtOp = ISD::ANY_EXTEND; 4120 TruncOp = ISD::TRUNCATE; 4121 } else { 4122 ExtOp = ISD::FP_EXTEND; 4123 TruncOp = ISD::FP_ROUND; 4124 } 4125 Tmp1 = Node->getOperand(0); 4126 // Promote each of the values to the new type. 4127 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1)); 4128 Tmp3 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(2)); 4129 // Perform the larger operation, then round down. 4130 Tmp1 = DAG.getSelect(dl, NVT, Tmp1, Tmp2, Tmp3); 4131 if (TruncOp != ISD::FP_ROUND) 4132 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1); 4133 else 4134 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1, 4135 DAG.getIntPtrConstant(0, dl)); 4136 Results.push_back(Tmp1); 4137 break; 4138 } 4139 case ISD::VECTOR_SHUFFLE: { 4140 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask(); 4141 4142 // Cast the two input vectors. 4143 Tmp1 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(0)); 4144 Tmp2 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(1)); 4145 4146 // Convert the shuffle mask to the right # elements. 4147 Tmp1 = ShuffleWithNarrowerEltType(NVT, OVT, dl, Tmp1, Tmp2, Mask); 4148 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OVT, Tmp1); 4149 Results.push_back(Tmp1); 4150 break; 4151 } 4152 case ISD::SETCC: { 4153 unsigned ExtOp = ISD::FP_EXTEND; 4154 if (NVT.isInteger()) { 4155 ISD::CondCode CCCode = 4156 cast<CondCodeSDNode>(Node->getOperand(2))->get(); 4157 ExtOp = isSignedIntSetCC(CCCode) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; 4158 } 4159 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0)); 4160 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1)); 4161 Results.push_back(DAG.getNode(ISD::SETCC, dl, Node->getValueType(0), 4162 Tmp1, Tmp2, Node->getOperand(2))); 4163 break; 4164 } 4165 case ISD::BR_CC: { 4166 unsigned ExtOp = ISD::FP_EXTEND; 4167 if (NVT.isInteger()) { 4168 ISD::CondCode CCCode = 4169 cast<CondCodeSDNode>(Node->getOperand(1))->get(); 4170 ExtOp = isSignedIntSetCC(CCCode) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; 4171 } 4172 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(2)); 4173 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(3)); 4174 Results.push_back(DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), 4175 Node->getOperand(0), Node->getOperand(1), 4176 Tmp1, Tmp2, Node->getOperand(4))); 4177 break; 4178 } 4179 case ISD::FADD: 4180 case ISD::FSUB: 4181 case ISD::FMUL: 4182 case ISD::FDIV: 4183 case ISD::FREM: 4184 case ISD::FMINNUM: 4185 case ISD::FMAXNUM: 4186 case ISD::FPOW: { 4187 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0)); 4188 Tmp2 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(1)); 4189 Tmp3 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2, 4190 Node->getFlags()); 4191 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT, 4192 Tmp3, DAG.getIntPtrConstant(0, dl))); 4193 break; 4194 } 4195 case ISD::FMA: { 4196 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0)); 4197 Tmp2 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(1)); 4198 Tmp3 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(2)); 4199 Results.push_back( 4200 DAG.getNode(ISD::FP_ROUND, dl, OVT, 4201 DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2, Tmp3), 4202 DAG.getIntPtrConstant(0, dl))); 4203 break; 4204 } 4205 case ISD::FCOPYSIGN: 4206 case ISD::FPOWI: { 4207 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0)); 4208 Tmp2 = Node->getOperand(1); 4209 Tmp3 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2); 4210 4211 // fcopysign doesn't change anything but the sign bit, so 4212 // (fp_round (fcopysign (fpext a), b)) 4213 // is as precise as 4214 // (fp_round (fpext a)) 4215 // which is a no-op. Mark it as a TRUNCating FP_ROUND. 4216 const bool isTrunc = (Node->getOpcode() == ISD::FCOPYSIGN); 4217 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT, 4218 Tmp3, DAG.getIntPtrConstant(isTrunc, dl))); 4219 break; 4220 } 4221 case ISD::FFLOOR: 4222 case ISD::FCEIL: 4223 case ISD::FRINT: 4224 case ISD::FNEARBYINT: 4225 case ISD::FROUND: 4226 case ISD::FTRUNC: 4227 case ISD::FNEG: 4228 case ISD::FSQRT: 4229 case ISD::FSIN: 4230 case ISD::FCOS: 4231 case ISD::FLOG: 4232 case ISD::FLOG2: 4233 case ISD::FLOG10: 4234 case ISD::FABS: 4235 case ISD::FEXP: 4236 case ISD::FEXP2: { 4237 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0)); 4238 Tmp2 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1); 4239 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT, 4240 Tmp2, DAG.getIntPtrConstant(0, dl))); 4241 break; 4242 } 4243 case ISD::BUILD_VECTOR: { 4244 MVT EltVT = OVT.getVectorElementType(); 4245 MVT NewEltVT = NVT.getVectorElementType(); 4246 4247 // Handle bitcasts to a different vector type with the same total bit size 4248 // 4249 // e.g. v2i64 = build_vector i64:x, i64:y => v4i32 4250 // => 4251 // v4i32 = concat_vectors (v2i32 (bitcast i64:x)), (v2i32 (bitcast i64:y)) 4252 4253 assert(NVT.isVector() && OVT.getSizeInBits() == NVT.getSizeInBits() && 4254 "Invalid promote type for build_vector"); 4255 assert(NewEltVT.bitsLT(EltVT) && "not handled"); 4256 4257 MVT MidVT = getPromotedVectorElementType(TLI, EltVT, NewEltVT); 4258 4259 SmallVector<SDValue, 8> NewOps; 4260 for (unsigned I = 0, E = Node->getNumOperands(); I != E; ++I) { 4261 SDValue Op = Node->getOperand(I); 4262 NewOps.push_back(DAG.getNode(ISD::BITCAST, SDLoc(Op), MidVT, Op)); 4263 } 4264 4265 SDLoc SL(Node); 4266 SDValue Concat = DAG.getNode(ISD::CONCAT_VECTORS, SL, NVT, NewOps); 4267 SDValue CvtVec = DAG.getNode(ISD::BITCAST, SL, OVT, Concat); 4268 Results.push_back(CvtVec); 4269 break; 4270 } 4271 case ISD::EXTRACT_VECTOR_ELT: { 4272 MVT EltVT = OVT.getVectorElementType(); 4273 MVT NewEltVT = NVT.getVectorElementType(); 4274 4275 // Handle bitcasts to a different vector type with the same total bit size. 4276 // 4277 // e.g. v2i64 = extract_vector_elt x:v2i64, y:i32 4278 // => 4279 // v4i32:castx = bitcast x:v2i64 4280 // 4281 // i64 = bitcast 4282 // (v2i32 build_vector (i32 (extract_vector_elt castx, (2 * y))), 4283 // (i32 (extract_vector_elt castx, (2 * y + 1))) 4284 // 4285 4286 assert(NVT.isVector() && OVT.getSizeInBits() == NVT.getSizeInBits() && 4287 "Invalid promote type for extract_vector_elt"); 4288 assert(NewEltVT.bitsLT(EltVT) && "not handled"); 4289 4290 MVT MidVT = getPromotedVectorElementType(TLI, EltVT, NewEltVT); 4291 unsigned NewEltsPerOldElt = MidVT.getVectorNumElements(); 4292 4293 SDValue Idx = Node->getOperand(1); 4294 EVT IdxVT = Idx.getValueType(); 4295 SDLoc SL(Node); 4296 SDValue Factor = DAG.getConstant(NewEltsPerOldElt, SL, IdxVT); 4297 SDValue NewBaseIdx = DAG.getNode(ISD::MUL, SL, IdxVT, Idx, Factor); 4298 4299 SDValue CastVec = DAG.getNode(ISD::BITCAST, SL, NVT, Node->getOperand(0)); 4300 4301 SmallVector<SDValue, 8> NewOps; 4302 for (unsigned I = 0; I < NewEltsPerOldElt; ++I) { 4303 SDValue IdxOffset = DAG.getConstant(I, SL, IdxVT); 4304 SDValue TmpIdx = DAG.getNode(ISD::ADD, SL, IdxVT, NewBaseIdx, IdxOffset); 4305 4306 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, NewEltVT, 4307 CastVec, TmpIdx); 4308 NewOps.push_back(Elt); 4309 } 4310 4311 SDValue NewVec = DAG.getNode(ISD::BUILD_VECTOR, SL, MidVT, NewOps); 4312 4313 Results.push_back(DAG.getNode(ISD::BITCAST, SL, EltVT, NewVec)); 4314 break; 4315 } 4316 case ISD::INSERT_VECTOR_ELT: { 4317 MVT EltVT = OVT.getVectorElementType(); 4318 MVT NewEltVT = NVT.getVectorElementType(); 4319 4320 // Handle bitcasts to a different vector type with the same total bit size 4321 // 4322 // e.g. v2i64 = insert_vector_elt x:v2i64, y:i64, z:i32 4323 // => 4324 // v4i32:castx = bitcast x:v2i64 4325 // v2i32:casty = bitcast y:i64 4326 // 4327 // v2i64 = bitcast 4328 // (v4i32 insert_vector_elt 4329 // (v4i32 insert_vector_elt v4i32:castx, 4330 // (extract_vector_elt casty, 0), 2 * z), 4331 // (extract_vector_elt casty, 1), (2 * z + 1)) 4332 4333 assert(NVT.isVector() && OVT.getSizeInBits() == NVT.getSizeInBits() && 4334 "Invalid promote type for insert_vector_elt"); 4335 assert(NewEltVT.bitsLT(EltVT) && "not handled"); 4336 4337 MVT MidVT = getPromotedVectorElementType(TLI, EltVT, NewEltVT); 4338 unsigned NewEltsPerOldElt = MidVT.getVectorNumElements(); 4339 4340 SDValue Val = Node->getOperand(1); 4341 SDValue Idx = Node->getOperand(2); 4342 EVT IdxVT = Idx.getValueType(); 4343 SDLoc SL(Node); 4344 4345 SDValue Factor = DAG.getConstant(NewEltsPerOldElt, SDLoc(), IdxVT); 4346 SDValue NewBaseIdx = DAG.getNode(ISD::MUL, SL, IdxVT, Idx, Factor); 4347 4348 SDValue CastVec = DAG.getNode(ISD::BITCAST, SL, NVT, Node->getOperand(0)); 4349 SDValue CastVal = DAG.getNode(ISD::BITCAST, SL, MidVT, Val); 4350 4351 SDValue NewVec = CastVec; 4352 for (unsigned I = 0; I < NewEltsPerOldElt; ++I) { 4353 SDValue IdxOffset = DAG.getConstant(I, SL, IdxVT); 4354 SDValue InEltIdx = DAG.getNode(ISD::ADD, SL, IdxVT, NewBaseIdx, IdxOffset); 4355 4356 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, NewEltVT, 4357 CastVal, IdxOffset); 4358 4359 NewVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, SL, NVT, 4360 NewVec, Elt, InEltIdx); 4361 } 4362 4363 Results.push_back(DAG.getNode(ISD::BITCAST, SL, OVT, NewVec)); 4364 break; 4365 } 4366 case ISD::SCALAR_TO_VECTOR: { 4367 MVT EltVT = OVT.getVectorElementType(); 4368 MVT NewEltVT = NVT.getVectorElementType(); 4369 4370 // Handle bitcasts to different vector type with the smae total bit size. 4371 // 4372 // e.g. v2i64 = scalar_to_vector x:i64 4373 // => 4374 // concat_vectors (v2i32 bitcast x:i64), (v2i32 undef) 4375 // 4376 4377 MVT MidVT = getPromotedVectorElementType(TLI, EltVT, NewEltVT); 4378 SDValue Val = Node->getOperand(0); 4379 SDLoc SL(Node); 4380 4381 SDValue CastVal = DAG.getNode(ISD::BITCAST, SL, MidVT, Val); 4382 SDValue Undef = DAG.getUNDEF(MidVT); 4383 4384 SmallVector<SDValue, 8> NewElts; 4385 NewElts.push_back(CastVal); 4386 for (unsigned I = 1, NElts = OVT.getVectorNumElements(); I != NElts; ++I) 4387 NewElts.push_back(Undef); 4388 4389 SDValue Concat = DAG.getNode(ISD::CONCAT_VECTORS, SL, NVT, NewElts); 4390 SDValue CvtVec = DAG.getNode(ISD::BITCAST, SL, OVT, Concat); 4391 Results.push_back(CvtVec); 4392 break; 4393 } 4394 } 4395 4396 // Replace the original node with the legalized result. 4397 if (!Results.empty()) 4398 ReplaceNode(Node, Results.data()); 4399} 4400 4401/// This is the entry point for the file. 4402void SelectionDAG::Legalize() { 4403 AssignTopologicalOrder(); 4404 4405 SmallPtrSet<SDNode *, 16> LegalizedNodes; 4406 SelectionDAGLegalize Legalizer(*this, LegalizedNodes); 4407 4408 // Visit all the nodes. We start in topological order, so that we see 4409 // nodes with their original operands intact. Legalization can produce 4410 // new nodes which may themselves need to be legalized. Iterate until all 4411 // nodes have been legalized. 4412 for (;;) { 4413 bool AnyLegalized = false; 4414 for (auto NI = allnodes_end(); NI != allnodes_begin();) { 4415 --NI; 4416 4417 SDNode *N = &*NI; 4418 if (N->use_empty() && N != getRoot().getNode()) { 4419 ++NI; 4420 DeleteNode(N); 4421 continue; 4422 } 4423 4424 if (LegalizedNodes.insert(N).second) { 4425 AnyLegalized = true; 4426 Legalizer.LegalizeOp(N); 4427 4428 if (N->use_empty() && N != getRoot().getNode()) { 4429 ++NI; 4430 DeleteNode(N); 4431 } 4432 } 4433 } 4434 if (!AnyLegalized) 4435 break; 4436 4437 } 4438 4439 // Remove dead nodes now. 4440 RemoveDeadNodes(); 4441} 4442 4443bool SelectionDAG::LegalizeOp(SDNode *N, 4444 SmallSetVector<SDNode *, 16> &UpdatedNodes) { 4445 SmallPtrSet<SDNode *, 16> LegalizedNodes; 4446 SelectionDAGLegalize Legalizer(*this, LegalizedNodes, &UpdatedNodes); 4447 4448 // Directly insert the node in question, and legalize it. This will recurse 4449 // as needed through operands. 4450 LegalizedNodes.insert(N); 4451 Legalizer.LegalizeOp(N); 4452 4453 return LegalizedNodes.count(N); 4454} 4455