SelectionDAGISel.cpp revision bc49100fb2399436fe15bc8892d81f79e6947003
1//===-- SelectionDAGISel.cpp - Implement the SelectionDAGISel class -------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This implements the SelectionDAGISel class. 11// 12//===----------------------------------------------------------------------===// 13 14#define DEBUG_TYPE "isel" 15#include "ScheduleDAGSDNodes.h" 16#include "SelectionDAGBuilder.h" 17#include "FunctionLoweringInfo.h" 18#include "llvm/CodeGen/SelectionDAGISel.h" 19#include "llvm/Analysis/AliasAnalysis.h" 20#include "llvm/Analysis/DebugInfo.h" 21#include "llvm/Constants.h" 22#include "llvm/CallingConv.h" 23#include "llvm/DerivedTypes.h" 24#include "llvm/Function.h" 25#include "llvm/GlobalVariable.h" 26#include "llvm/InlineAsm.h" 27#include "llvm/Instructions.h" 28#include "llvm/Intrinsics.h" 29#include "llvm/IntrinsicInst.h" 30#include "llvm/LLVMContext.h" 31#include "llvm/CodeGen/FastISel.h" 32#include "llvm/CodeGen/GCStrategy.h" 33#include "llvm/CodeGen/GCMetadata.h" 34#include "llvm/CodeGen/MachineFunction.h" 35#include "llvm/CodeGen/MachineFunctionAnalysis.h" 36#include "llvm/CodeGen/MachineFrameInfo.h" 37#include "llvm/CodeGen/MachineInstrBuilder.h" 38#include "llvm/CodeGen/MachineJumpTableInfo.h" 39#include "llvm/CodeGen/MachineModuleInfo.h" 40#include "llvm/CodeGen/MachineRegisterInfo.h" 41#include "llvm/CodeGen/ScheduleHazardRecognizer.h" 42#include "llvm/CodeGen/SchedulerRegistry.h" 43#include "llvm/CodeGen/SelectionDAG.h" 44#include "llvm/Target/TargetRegisterInfo.h" 45#include "llvm/Target/TargetData.h" 46#include "llvm/Target/TargetFrameInfo.h" 47#include "llvm/Target/TargetIntrinsicInfo.h" 48#include "llvm/Target/TargetInstrInfo.h" 49#include "llvm/Target/TargetLowering.h" 50#include "llvm/Target/TargetMachine.h" 51#include "llvm/Target/TargetOptions.h" 52#include "llvm/Support/Compiler.h" 53#include "llvm/Support/Debug.h" 54#include "llvm/Support/ErrorHandling.h" 55#include "llvm/Support/MathExtras.h" 56#include "llvm/Support/Timer.h" 57#include "llvm/Support/raw_ostream.h" 58#include "llvm/ADT/Statistic.h" 59#include <algorithm> 60using namespace llvm; 61 62STATISTIC(NumFastIselFailures, "Number of instructions fast isel failed on"); 63STATISTIC(NumDAGIselRetries,"Number of times dag isel has to try another path"); 64 65static cl::opt<bool> 66EnableFastISelVerbose("fast-isel-verbose", cl::Hidden, 67 cl::desc("Enable verbose messages in the \"fast\" " 68 "instruction selector")); 69static cl::opt<bool> 70EnableFastISelAbort("fast-isel-abort", cl::Hidden, 71 cl::desc("Enable abort calls when \"fast\" instruction fails")); 72static cl::opt<bool> 73SchedLiveInCopies("schedule-livein-copies", cl::Hidden, 74 cl::desc("Schedule copies of livein registers"), 75 cl::init(false)); 76 77#ifndef NDEBUG 78static cl::opt<bool> 79ViewDAGCombine1("view-dag-combine1-dags", cl::Hidden, 80 cl::desc("Pop up a window to show dags before the first " 81 "dag combine pass")); 82static cl::opt<bool> 83ViewLegalizeTypesDAGs("view-legalize-types-dags", cl::Hidden, 84 cl::desc("Pop up a window to show dags before legalize types")); 85static cl::opt<bool> 86ViewLegalizeDAGs("view-legalize-dags", cl::Hidden, 87 cl::desc("Pop up a window to show dags before legalize")); 88static cl::opt<bool> 89ViewDAGCombine2("view-dag-combine2-dags", cl::Hidden, 90 cl::desc("Pop up a window to show dags before the second " 91 "dag combine pass")); 92static cl::opt<bool> 93ViewDAGCombineLT("view-dag-combine-lt-dags", cl::Hidden, 94 cl::desc("Pop up a window to show dags before the post legalize types" 95 " dag combine pass")); 96static cl::opt<bool> 97ViewISelDAGs("view-isel-dags", cl::Hidden, 98 cl::desc("Pop up a window to show isel dags as they are selected")); 99static cl::opt<bool> 100ViewSchedDAGs("view-sched-dags", cl::Hidden, 101 cl::desc("Pop up a window to show sched dags as they are processed")); 102static cl::opt<bool> 103ViewSUnitDAGs("view-sunit-dags", cl::Hidden, 104 cl::desc("Pop up a window to show SUnit dags after they are processed")); 105#else 106static const bool ViewDAGCombine1 = false, 107 ViewLegalizeTypesDAGs = false, ViewLegalizeDAGs = false, 108 ViewDAGCombine2 = false, 109 ViewDAGCombineLT = false, 110 ViewISelDAGs = false, ViewSchedDAGs = false, 111 ViewSUnitDAGs = false; 112#endif 113 114//===---------------------------------------------------------------------===// 115/// 116/// RegisterScheduler class - Track the registration of instruction schedulers. 117/// 118//===---------------------------------------------------------------------===// 119MachinePassRegistry RegisterScheduler::Registry; 120 121//===---------------------------------------------------------------------===// 122/// 123/// ISHeuristic command line option for instruction schedulers. 124/// 125//===---------------------------------------------------------------------===// 126static cl::opt<RegisterScheduler::FunctionPassCtor, false, 127 RegisterPassParser<RegisterScheduler> > 128ISHeuristic("pre-RA-sched", 129 cl::init(&createDefaultScheduler), 130 cl::desc("Instruction schedulers available (before register" 131 " allocation):")); 132 133static RegisterScheduler 134defaultListDAGScheduler("default", "Best scheduler for the target", 135 createDefaultScheduler); 136 137namespace llvm { 138 //===--------------------------------------------------------------------===// 139 /// createDefaultScheduler - This creates an instruction scheduler appropriate 140 /// for the target. 141 ScheduleDAGSDNodes* createDefaultScheduler(SelectionDAGISel *IS, 142 CodeGenOpt::Level OptLevel) { 143 const TargetLowering &TLI = IS->getTargetLowering(); 144 145 if (OptLevel == CodeGenOpt::None) 146 return createFastDAGScheduler(IS, OptLevel); 147 if (TLI.getSchedulingPreference() == TargetLowering::SchedulingForLatency) 148 return createTDListDAGScheduler(IS, OptLevel); 149 assert(TLI.getSchedulingPreference() == 150 TargetLowering::SchedulingForRegPressure && "Unknown sched type!"); 151 return createBURRListDAGScheduler(IS, OptLevel); 152 } 153} 154 155// EmitInstrWithCustomInserter - This method should be implemented by targets 156// that mark instructions with the 'usesCustomInserter' flag. These 157// instructions are special in various ways, which require special support to 158// insert. The specified MachineInstr is created but not inserted into any 159// basic blocks, and this method is called to expand it into a sequence of 160// instructions, potentially also creating new basic blocks and control flow. 161// When new basic blocks are inserted and the edges from MBB to its successors 162// are modified, the method should insert pairs of <OldSucc, NewSucc> into the 163// DenseMap. 164MachineBasicBlock *TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI, 165 MachineBasicBlock *MBB, 166 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const { 167#ifndef NDEBUG 168 dbgs() << "If a target marks an instruction with " 169 "'usesCustomInserter', it must implement " 170 "TargetLowering::EmitInstrWithCustomInserter!"; 171#endif 172 llvm_unreachable(0); 173 return 0; 174} 175 176/// EmitLiveInCopy - Emit a copy for a live in physical register. If the 177/// physical register has only a single copy use, then coalesced the copy 178/// if possible. 179static void EmitLiveInCopy(MachineBasicBlock *MBB, 180 MachineBasicBlock::iterator &InsertPos, 181 unsigned VirtReg, unsigned PhysReg, 182 const TargetRegisterClass *RC, 183 DenseMap<MachineInstr*, unsigned> &CopyRegMap, 184 const MachineRegisterInfo &MRI, 185 const TargetRegisterInfo &TRI, 186 const TargetInstrInfo &TII) { 187 unsigned NumUses = 0; 188 MachineInstr *UseMI = NULL; 189 for (MachineRegisterInfo::use_iterator UI = MRI.use_begin(VirtReg), 190 UE = MRI.use_end(); UI != UE; ++UI) { 191 UseMI = &*UI; 192 if (++NumUses > 1) 193 break; 194 } 195 196 // If the number of uses is not one, or the use is not a move instruction, 197 // don't coalesce. Also, only coalesce away a virtual register to virtual 198 // register copy. 199 bool Coalesced = false; 200 unsigned SrcReg, DstReg, SrcSubReg, DstSubReg; 201 if (NumUses == 1 && 202 TII.isMoveInstr(*UseMI, SrcReg, DstReg, SrcSubReg, DstSubReg) && 203 TargetRegisterInfo::isVirtualRegister(DstReg)) { 204 VirtReg = DstReg; 205 Coalesced = true; 206 } 207 208 // Now find an ideal location to insert the copy. 209 MachineBasicBlock::iterator Pos = InsertPos; 210 while (Pos != MBB->begin()) { 211 MachineInstr *PrevMI = prior(Pos); 212 DenseMap<MachineInstr*, unsigned>::iterator RI = CopyRegMap.find(PrevMI); 213 // copyRegToReg might emit multiple instructions to do a copy. 214 unsigned CopyDstReg = (RI == CopyRegMap.end()) ? 0 : RI->second; 215 if (CopyDstReg && !TRI.regsOverlap(CopyDstReg, PhysReg)) 216 // This is what the BB looks like right now: 217 // r1024 = mov r0 218 // ... 219 // r1 = mov r1024 220 // 221 // We want to insert "r1025 = mov r1". Inserting this copy below the 222 // move to r1024 makes it impossible for that move to be coalesced. 223 // 224 // r1025 = mov r1 225 // r1024 = mov r0 226 // ... 227 // r1 = mov 1024 228 // r2 = mov 1025 229 break; // Woot! Found a good location. 230 --Pos; 231 } 232 233 bool Emitted = TII.copyRegToReg(*MBB, Pos, VirtReg, PhysReg, RC, RC); 234 assert(Emitted && "Unable to issue a live-in copy instruction!\n"); 235 (void) Emitted; 236 237 CopyRegMap.insert(std::make_pair(prior(Pos), VirtReg)); 238 if (Coalesced) { 239 if (&*InsertPos == UseMI) ++InsertPos; 240 MBB->erase(UseMI); 241 } 242} 243 244/// EmitLiveInCopies - If this is the first basic block in the function, 245/// and if it has live ins that need to be copied into vregs, emit the 246/// copies into the block. 247static void EmitLiveInCopies(MachineBasicBlock *EntryMBB, 248 const MachineRegisterInfo &MRI, 249 const TargetRegisterInfo &TRI, 250 const TargetInstrInfo &TII) { 251 if (SchedLiveInCopies) { 252 // Emit the copies at a heuristically-determined location in the block. 253 DenseMap<MachineInstr*, unsigned> CopyRegMap; 254 MachineBasicBlock::iterator InsertPos = EntryMBB->begin(); 255 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(), 256 E = MRI.livein_end(); LI != E; ++LI) 257 if (LI->second) { 258 const TargetRegisterClass *RC = MRI.getRegClass(LI->second); 259 EmitLiveInCopy(EntryMBB, InsertPos, LI->second, LI->first, 260 RC, CopyRegMap, MRI, TRI, TII); 261 } 262 } else { 263 // Emit the copies into the top of the block. 264 for (MachineRegisterInfo::livein_iterator LI = MRI.livein_begin(), 265 E = MRI.livein_end(); LI != E; ++LI) 266 if (LI->second) { 267 const TargetRegisterClass *RC = MRI.getRegClass(LI->second); 268 bool Emitted = TII.copyRegToReg(*EntryMBB, EntryMBB->begin(), 269 LI->second, LI->first, RC, RC); 270 assert(Emitted && "Unable to issue a live-in copy instruction!\n"); 271 (void) Emitted; 272 } 273 } 274} 275 276//===----------------------------------------------------------------------===// 277// SelectionDAGISel code 278//===----------------------------------------------------------------------===// 279 280SelectionDAGISel::SelectionDAGISel(TargetMachine &tm, CodeGenOpt::Level OL) : 281 MachineFunctionPass(&ID), TM(tm), TLI(*tm.getTargetLowering()), 282 FuncInfo(new FunctionLoweringInfo(TLI)), 283 CurDAG(new SelectionDAG(TLI, *FuncInfo)), 284 SDB(new SelectionDAGBuilder(*CurDAG, TLI, *FuncInfo, OL)), 285 GFI(), 286 OptLevel(OL), 287 DAGSize(0) 288{} 289 290SelectionDAGISel::~SelectionDAGISel() { 291 delete SDB; 292 delete CurDAG; 293 delete FuncInfo; 294} 295 296unsigned SelectionDAGISel::MakeReg(EVT VT) { 297 return RegInfo->createVirtualRegister(TLI.getRegClassFor(VT)); 298} 299 300void SelectionDAGISel::getAnalysisUsage(AnalysisUsage &AU) const { 301 AU.addRequired<AliasAnalysis>(); 302 AU.addPreserved<AliasAnalysis>(); 303 AU.addRequired<GCModuleInfo>(); 304 AU.addPreserved<GCModuleInfo>(); 305 MachineFunctionPass::getAnalysisUsage(AU); 306} 307 308bool SelectionDAGISel::runOnMachineFunction(MachineFunction &mf) { 309 Function &Fn = *mf.getFunction(); 310 311 // Do some sanity-checking on the command-line options. 312 assert((!EnableFastISelVerbose || EnableFastISel) && 313 "-fast-isel-verbose requires -fast-isel"); 314 assert((!EnableFastISelAbort || EnableFastISel) && 315 "-fast-isel-abort requires -fast-isel"); 316 317 // Get alias analysis for load/store combining. 318 AA = &getAnalysis<AliasAnalysis>(); 319 320 MF = &mf; 321 const TargetInstrInfo &TII = *TM.getInstrInfo(); 322 const TargetRegisterInfo &TRI = *TM.getRegisterInfo(); 323 324 if (Fn.hasGC()) 325 GFI = &getAnalysis<GCModuleInfo>().getFunctionInfo(Fn); 326 else 327 GFI = 0; 328 RegInfo = &MF->getRegInfo(); 329 DEBUG(dbgs() << "\n\n\n=== " << Fn.getName() << "\n"); 330 331 MachineModuleInfo *MMI = getAnalysisIfAvailable<MachineModuleInfo>(); 332 CurDAG->init(*MF, MMI); 333 FuncInfo->set(Fn, *MF, EnableFastISel); 334 SDB->init(GFI, *AA); 335 336 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) 337 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(I->getTerminator())) 338 // Mark landing pad. 339 FuncInfo->MBBMap[Invoke->getSuccessor(1)]->setIsLandingPad(); 340 341 SelectAllBasicBlocks(Fn, *MF, MMI, TII); 342 343 // If the first basic block in the function has live ins that need to be 344 // copied into vregs, emit the copies into the top of the block before 345 // emitting the code for the block. 346 EmitLiveInCopies(MF->begin(), *RegInfo, TRI, TII); 347 348 // Add function live-ins to entry block live-in set. 349 for (MachineRegisterInfo::livein_iterator I = RegInfo->livein_begin(), 350 E = RegInfo->livein_end(); I != E; ++I) 351 MF->begin()->addLiveIn(I->first); 352 353#ifndef NDEBUG 354 assert(FuncInfo->CatchInfoFound.size() == FuncInfo->CatchInfoLost.size() && 355 "Not all catch info was assigned to a landing pad!"); 356#endif 357 358 FuncInfo->clear(); 359 360 return true; 361} 362 363/// SetDebugLoc - Update MF's and SDB's DebugLocs if debug information is 364/// attached with this instruction. 365static void SetDebugLoc(Instruction *I, SelectionDAGBuilder *SDB, 366 FastISel *FastIS, MachineFunction *MF) { 367 DebugLoc DL = I->getDebugLoc(); 368 if (DL.isUnknown()) return; 369 370 SDB->setCurDebugLoc(DL); 371 372 if (FastIS) 373 FastIS->setCurDebugLoc(DL); 374 375 // If the function doesn't have a default debug location yet, set 376 // it. This is kind of a hack. 377 if (MF->getDefaultDebugLoc().isUnknown()) 378 MF->setDefaultDebugLoc(DL); 379} 380 381/// ResetDebugLoc - Set MF's and SDB's DebugLocs to Unknown. 382static void ResetDebugLoc(SelectionDAGBuilder *SDB, FastISel *FastIS) { 383 SDB->setCurDebugLoc(DebugLoc()); 384 if (FastIS) 385 FastIS->setCurDebugLoc(DebugLoc()); 386} 387 388void SelectionDAGISel::SelectBasicBlock(BasicBlock *LLVMBB, 389 BasicBlock::iterator Begin, 390 BasicBlock::iterator End, 391 bool &HadTailCall) { 392 SDB->setCurrentBasicBlock(BB); 393 394 // Lower all of the non-terminator instructions. If a call is emitted 395 // as a tail call, cease emitting nodes for this block. 396 for (BasicBlock::iterator I = Begin; I != End && !SDB->HasTailCall; ++I) { 397 SetDebugLoc(I, SDB, 0, MF); 398 399 if (!isa<TerminatorInst>(I)) { 400 SDB->visit(*I); 401 402 // Set the current debug location back to "unknown" so that it doesn't 403 // spuriously apply to subsequent instructions. 404 ResetDebugLoc(SDB, 0); 405 } 406 } 407 408 if (!SDB->HasTailCall) { 409 // Ensure that all instructions which are used outside of their defining 410 // blocks are available as virtual registers. Invoke is handled elsewhere. 411 for (BasicBlock::iterator I = Begin; I != End; ++I) 412 if (!isa<PHINode>(I) && !isa<InvokeInst>(I)) 413 SDB->CopyToExportRegsIfNeeded(I); 414 415 // Handle PHI nodes in successor blocks. 416 if (End == LLVMBB->end()) { 417 HandlePHINodesInSuccessorBlocks(LLVMBB); 418 419 // Lower the terminator after the copies are emitted. 420 SetDebugLoc(LLVMBB->getTerminator(), SDB, 0, MF); 421 SDB->visit(*LLVMBB->getTerminator()); 422 ResetDebugLoc(SDB, 0); 423 } 424 } 425 426 // Make sure the root of the DAG is up-to-date. 427 CurDAG->setRoot(SDB->getControlRoot()); 428 429 // Final step, emit the lowered DAG as machine code. 430 CodeGenAndEmitDAG(); 431 HadTailCall = SDB->HasTailCall; 432 SDB->clear(); 433} 434 435namespace { 436/// WorkListRemover - This class is a DAGUpdateListener that removes any deleted 437/// nodes from the worklist. 438class SDOPsWorkListRemover : public SelectionDAG::DAGUpdateListener { 439 SmallVector<SDNode*, 128> &Worklist; 440 SmallPtrSet<SDNode*, 128> &InWorklist; 441public: 442 SDOPsWorkListRemover(SmallVector<SDNode*, 128> &wl, 443 SmallPtrSet<SDNode*, 128> &inwl) 444 : Worklist(wl), InWorklist(inwl) {} 445 446 void RemoveFromWorklist(SDNode *N) { 447 if (!InWorklist.erase(N)) return; 448 449 SmallVector<SDNode*, 128>::iterator I = 450 std::find(Worklist.begin(), Worklist.end(), N); 451 assert(I != Worklist.end() && "Not in worklist"); 452 453 *I = Worklist.back(); 454 Worklist.pop_back(); 455 } 456 457 virtual void NodeDeleted(SDNode *N, SDNode *E) { 458 RemoveFromWorklist(N); 459 } 460 461 virtual void NodeUpdated(SDNode *N) { 462 // Ignore updates. 463 } 464}; 465} 466 467/// TrivialTruncElim - Eliminate some trivial nops that can result from 468/// ShrinkDemandedOps: (trunc (ext n)) -> n. 469static bool TrivialTruncElim(SDValue Op, 470 TargetLowering::TargetLoweringOpt &TLO) { 471 SDValue N0 = Op.getOperand(0); 472 EVT VT = Op.getValueType(); 473 if ((N0.getOpcode() == ISD::ZERO_EXTEND || 474 N0.getOpcode() == ISD::SIGN_EXTEND || 475 N0.getOpcode() == ISD::ANY_EXTEND) && 476 N0.getOperand(0).getValueType() == VT) { 477 return TLO.CombineTo(Op, N0.getOperand(0)); 478 } 479 return false; 480} 481 482/// ShrinkDemandedOps - A late transformation pass that shrink expressions 483/// using TargetLowering::TargetLoweringOpt::ShrinkDemandedOp. It converts 484/// x+y to (VT)((SmallVT)x+(SmallVT)y) if the casts are free. 485void SelectionDAGISel::ShrinkDemandedOps() { 486 SmallVector<SDNode*, 128> Worklist; 487 SmallPtrSet<SDNode*, 128> InWorklist; 488 489 // Add all the dag nodes to the worklist. 490 Worklist.reserve(CurDAG->allnodes_size()); 491 for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(), 492 E = CurDAG->allnodes_end(); I != E; ++I) { 493 Worklist.push_back(I); 494 InWorklist.insert(I); 495 } 496 497 TargetLowering::TargetLoweringOpt TLO(*CurDAG, true); 498 while (!Worklist.empty()) { 499 SDNode *N = Worklist.pop_back_val(); 500 InWorklist.erase(N); 501 502 if (N->use_empty() && N != CurDAG->getRoot().getNode()) { 503 // Deleting this node may make its operands dead, add them to the worklist 504 // if they aren't already there. 505 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 506 if (InWorklist.insert(N->getOperand(i).getNode())) 507 Worklist.push_back(N->getOperand(i).getNode()); 508 509 CurDAG->DeleteNode(N); 510 continue; 511 } 512 513 // Run ShrinkDemandedOp on scalar binary operations. 514 if (N->getNumValues() != 1 || 515 !N->getValueType(0).isSimple() || !N->getValueType(0).isInteger()) 516 continue; 517 518 unsigned BitWidth = N->getValueType(0).getScalarType().getSizeInBits(); 519 APInt Demanded = APInt::getAllOnesValue(BitWidth); 520 APInt KnownZero, KnownOne; 521 if (!TLI.SimplifyDemandedBits(SDValue(N, 0), Demanded, 522 KnownZero, KnownOne, TLO) && 523 (N->getOpcode() != ISD::TRUNCATE || 524 !TrivialTruncElim(SDValue(N, 0), TLO))) 525 continue; 526 527 // Revisit the node. 528 assert(!InWorklist.count(N) && "Already in worklist"); 529 Worklist.push_back(N); 530 InWorklist.insert(N); 531 532 // Replace the old value with the new one. 533 DEBUG(errs() << "\nShrinkDemandedOps replacing "; 534 TLO.Old.getNode()->dump(CurDAG); 535 errs() << "\nWith: "; 536 TLO.New.getNode()->dump(CurDAG); 537 errs() << '\n'); 538 539 if (InWorklist.insert(TLO.New.getNode())) 540 Worklist.push_back(TLO.New.getNode()); 541 542 SDOPsWorkListRemover DeadNodes(Worklist, InWorklist); 543 CurDAG->ReplaceAllUsesOfValueWith(TLO.Old, TLO.New, &DeadNodes); 544 545 if (!TLO.Old.getNode()->use_empty()) continue; 546 547 for (unsigned i = 0, e = TLO.Old.getNode()->getNumOperands(); 548 i != e; ++i) { 549 SDNode *OpNode = TLO.Old.getNode()->getOperand(i).getNode(); 550 if (OpNode->hasOneUse()) { 551 // Add OpNode to the end of the list to revisit. 552 DeadNodes.RemoveFromWorklist(OpNode); 553 Worklist.push_back(OpNode); 554 InWorklist.insert(OpNode); 555 } 556 } 557 558 DeadNodes.RemoveFromWorklist(TLO.Old.getNode()); 559 CurDAG->DeleteNode(TLO.Old.getNode()); 560 } 561} 562 563void SelectionDAGISel::ComputeLiveOutVRegInfo() { 564 SmallPtrSet<SDNode*, 128> VisitedNodes; 565 SmallVector<SDNode*, 128> Worklist; 566 567 Worklist.push_back(CurDAG->getRoot().getNode()); 568 569 APInt Mask; 570 APInt KnownZero; 571 APInt KnownOne; 572 573 do { 574 SDNode *N = Worklist.pop_back_val(); 575 576 // If we've already seen this node, ignore it. 577 if (!VisitedNodes.insert(N)) 578 continue; 579 580 // Otherwise, add all chain operands to the worklist. 581 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 582 if (N->getOperand(i).getValueType() == MVT::Other) 583 Worklist.push_back(N->getOperand(i).getNode()); 584 585 // If this is a CopyToReg with a vreg dest, process it. 586 if (N->getOpcode() != ISD::CopyToReg) 587 continue; 588 589 unsigned DestReg = cast<RegisterSDNode>(N->getOperand(1))->getReg(); 590 if (!TargetRegisterInfo::isVirtualRegister(DestReg)) 591 continue; 592 593 // Ignore non-scalar or non-integer values. 594 SDValue Src = N->getOperand(2); 595 EVT SrcVT = Src.getValueType(); 596 if (!SrcVT.isInteger() || SrcVT.isVector()) 597 continue; 598 599 unsigned NumSignBits = CurDAG->ComputeNumSignBits(Src); 600 Mask = APInt::getAllOnesValue(SrcVT.getSizeInBits()); 601 CurDAG->ComputeMaskedBits(Src, Mask, KnownZero, KnownOne); 602 603 // Only install this information if it tells us something. 604 if (NumSignBits != 1 || KnownZero != 0 || KnownOne != 0) { 605 DestReg -= TargetRegisterInfo::FirstVirtualRegister; 606 if (DestReg >= FuncInfo->LiveOutRegInfo.size()) 607 FuncInfo->LiveOutRegInfo.resize(DestReg+1); 608 FunctionLoweringInfo::LiveOutInfo &LOI = 609 FuncInfo->LiveOutRegInfo[DestReg]; 610 LOI.NumSignBits = NumSignBits; 611 LOI.KnownOne = KnownOne; 612 LOI.KnownZero = KnownZero; 613 } 614 } while (!Worklist.empty()); 615} 616 617void SelectionDAGISel::CodeGenAndEmitDAG() { 618 std::string GroupName; 619 if (TimePassesIsEnabled) 620 GroupName = "Instruction Selection and Scheduling"; 621 std::string BlockName; 622 if (ViewDAGCombine1 || ViewLegalizeTypesDAGs || ViewLegalizeDAGs || 623 ViewDAGCombine2 || ViewDAGCombineLT || ViewISelDAGs || ViewSchedDAGs || 624 ViewSUnitDAGs) 625 BlockName = MF->getFunction()->getNameStr() + ":" + 626 BB->getBasicBlock()->getNameStr(); 627 628 DEBUG(dbgs() << "Initial selection DAG:\n"); 629 DEBUG(CurDAG->dump()); 630 631 if (ViewDAGCombine1) CurDAG->viewGraph("dag-combine1 input for " + BlockName); 632 633 // Run the DAG combiner in pre-legalize mode. 634 if (TimePassesIsEnabled) { 635 NamedRegionTimer T("DAG Combining 1", GroupName); 636 CurDAG->Combine(Unrestricted, *AA, OptLevel); 637 } else { 638 CurDAG->Combine(Unrestricted, *AA, OptLevel); 639 } 640 641 DEBUG(dbgs() << "Optimized lowered selection DAG:\n"); 642 DEBUG(CurDAG->dump()); 643 644 // Second step, hack on the DAG until it only uses operations and types that 645 // the target supports. 646 if (ViewLegalizeTypesDAGs) CurDAG->viewGraph("legalize-types input for " + 647 BlockName); 648 649 bool Changed; 650 if (TimePassesIsEnabled) { 651 NamedRegionTimer T("Type Legalization", GroupName); 652 Changed = CurDAG->LegalizeTypes(); 653 } else { 654 Changed = CurDAG->LegalizeTypes(); 655 } 656 657 DEBUG(dbgs() << "Type-legalized selection DAG:\n"); 658 DEBUG(CurDAG->dump()); 659 660 if (Changed) { 661 if (ViewDAGCombineLT) 662 CurDAG->viewGraph("dag-combine-lt input for " + BlockName); 663 664 // Run the DAG combiner in post-type-legalize mode. 665 if (TimePassesIsEnabled) { 666 NamedRegionTimer T("DAG Combining after legalize types", GroupName); 667 CurDAG->Combine(NoIllegalTypes, *AA, OptLevel); 668 } else { 669 CurDAG->Combine(NoIllegalTypes, *AA, OptLevel); 670 } 671 672 DEBUG(dbgs() << "Optimized type-legalized selection DAG:\n"); 673 DEBUG(CurDAG->dump()); 674 } 675 676 if (TimePassesIsEnabled) { 677 NamedRegionTimer T("Vector Legalization", GroupName); 678 Changed = CurDAG->LegalizeVectors(); 679 } else { 680 Changed = CurDAG->LegalizeVectors(); 681 } 682 683 if (Changed) { 684 if (TimePassesIsEnabled) { 685 NamedRegionTimer T("Type Legalization 2", GroupName); 686 CurDAG->LegalizeTypes(); 687 } else { 688 CurDAG->LegalizeTypes(); 689 } 690 691 if (ViewDAGCombineLT) 692 CurDAG->viewGraph("dag-combine-lv input for " + BlockName); 693 694 // Run the DAG combiner in post-type-legalize mode. 695 if (TimePassesIsEnabled) { 696 NamedRegionTimer T("DAG Combining after legalize vectors", GroupName); 697 CurDAG->Combine(NoIllegalOperations, *AA, OptLevel); 698 } else { 699 CurDAG->Combine(NoIllegalOperations, *AA, OptLevel); 700 } 701 702 DEBUG(dbgs() << "Optimized vector-legalized selection DAG:\n"); 703 DEBUG(CurDAG->dump()); 704 } 705 706 if (ViewLegalizeDAGs) CurDAG->viewGraph("legalize input for " + BlockName); 707 708 if (TimePassesIsEnabled) { 709 NamedRegionTimer T("DAG Legalization", GroupName); 710 CurDAG->Legalize(OptLevel); 711 } else { 712 CurDAG->Legalize(OptLevel); 713 } 714 715 DEBUG(dbgs() << "Legalized selection DAG:\n"); 716 DEBUG(CurDAG->dump()); 717 718 if (ViewDAGCombine2) CurDAG->viewGraph("dag-combine2 input for " + BlockName); 719 720 // Run the DAG combiner in post-legalize mode. 721 if (TimePassesIsEnabled) { 722 NamedRegionTimer T("DAG Combining 2", GroupName); 723 CurDAG->Combine(NoIllegalOperations, *AA, OptLevel); 724 } else { 725 CurDAG->Combine(NoIllegalOperations, *AA, OptLevel); 726 } 727 728 DEBUG(dbgs() << "Optimized legalized selection DAG:\n"); 729 DEBUG(CurDAG->dump()); 730 731 if (OptLevel != CodeGenOpt::None) { 732 ShrinkDemandedOps(); 733 ComputeLiveOutVRegInfo(); 734 } 735 736 if (ViewISelDAGs) CurDAG->viewGraph("isel input for " + BlockName); 737 738 // Third, instruction select all of the operations to machine code, adding the 739 // code to the MachineBasicBlock. 740 if (TimePassesIsEnabled) { 741 NamedRegionTimer T("Instruction Selection", GroupName); 742 DoInstructionSelection(); 743 } else { 744 DoInstructionSelection(); 745 } 746 747 DEBUG(dbgs() << "Selected selection DAG:\n"); 748 DEBUG(CurDAG->dump()); 749 750 if (ViewSchedDAGs) CurDAG->viewGraph("scheduler input for " + BlockName); 751 752 // Schedule machine code. 753 ScheduleDAGSDNodes *Scheduler = CreateScheduler(); 754 if (TimePassesIsEnabled) { 755 NamedRegionTimer T("Instruction Scheduling", GroupName); 756 Scheduler->Run(CurDAG, BB, BB->end()); 757 } else { 758 Scheduler->Run(CurDAG, BB, BB->end()); 759 } 760 761 if (ViewSUnitDAGs) Scheduler->viewGraph(); 762 763 // Emit machine code to BB. This can change 'BB' to the last block being 764 // inserted into. 765 if (TimePassesIsEnabled) { 766 NamedRegionTimer T("Instruction Creation", GroupName); 767 BB = Scheduler->EmitSchedule(&SDB->EdgeMapping); 768 } else { 769 BB = Scheduler->EmitSchedule(&SDB->EdgeMapping); 770 } 771 772 // Free the scheduler state. 773 if (TimePassesIsEnabled) { 774 NamedRegionTimer T("Instruction Scheduling Cleanup", GroupName); 775 delete Scheduler; 776 } else { 777 delete Scheduler; 778 } 779 780 DEBUG(dbgs() << "Selected machine code:\n"); 781 DEBUG(BB->dump()); 782} 783 784void SelectionDAGISel::DoInstructionSelection() { 785 DEBUG(errs() << "===== Instruction selection begins:\n"); 786 787 PreprocessISelDAG(); 788 789 // Select target instructions for the DAG. 790 { 791 // Number all nodes with a topological order and set DAGSize. 792 DAGSize = CurDAG->AssignTopologicalOrder(); 793 794 // Create a dummy node (which is not added to allnodes), that adds 795 // a reference to the root node, preventing it from being deleted, 796 // and tracking any changes of the root. 797 HandleSDNode Dummy(CurDAG->getRoot()); 798 ISelPosition = SelectionDAG::allnodes_iterator(CurDAG->getRoot().getNode()); 799 ++ISelPosition; 800 801 // The AllNodes list is now topological-sorted. Visit the 802 // nodes by starting at the end of the list (the root of the 803 // graph) and preceding back toward the beginning (the entry 804 // node). 805 while (ISelPosition != CurDAG->allnodes_begin()) { 806 SDNode *Node = --ISelPosition; 807 // Skip dead nodes. DAGCombiner is expected to eliminate all dead nodes, 808 // but there are currently some corner cases that it misses. Also, this 809 // makes it theoretically possible to disable the DAGCombiner. 810 if (Node->use_empty()) 811 continue; 812 813 SDNode *ResNode = Select(Node); 814 815 // FIXME: This is pretty gross. 'Select' should be changed to not return 816 // anything at all and this code should be nuked with a tactical strike. 817 818 // If node should not be replaced, continue with the next one. 819 if (ResNode == Node || Node->getOpcode() == ISD::DELETED_NODE) 820 continue; 821 // Replace node. 822 if (ResNode) 823 ReplaceUses(Node, ResNode); 824 825 // If after the replacement this node is not used any more, 826 // remove this dead node. 827 if (Node->use_empty()) { // Don't delete EntryToken, etc. 828 ISelUpdater ISU(ISelPosition); 829 CurDAG->RemoveDeadNode(Node, &ISU); 830 } 831 } 832 833 CurDAG->setRoot(Dummy.getValue()); 834 } 835 DEBUG(errs() << "===== Instruction selection ends:\n"); 836 837 PostprocessISelDAG(); 838} 839 840 841void SelectionDAGISel::SelectAllBasicBlocks(Function &Fn, 842 MachineFunction &MF, 843 MachineModuleInfo *MMI, 844 const TargetInstrInfo &TII) { 845 // Initialize the Fast-ISel state, if needed. 846 FastISel *FastIS = 0; 847 if (EnableFastISel) 848 FastIS = TLI.createFastISel(MF, FuncInfo->ValueMap, FuncInfo->MBBMap, 849 FuncInfo->StaticAllocaMap 850#ifndef NDEBUG 851 , FuncInfo->CatchInfoLost 852#endif 853 ); 854 855 // Iterate over all basic blocks in the function. 856 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) { 857 BasicBlock *LLVMBB = &*I; 858 BB = FuncInfo->MBBMap[LLVMBB]; 859 860 BasicBlock::iterator const Begin = LLVMBB->begin(); 861 BasicBlock::iterator const End = LLVMBB->end(); 862 BasicBlock::iterator BI = Begin; 863 864 // Lower any arguments needed in this block if this is the entry block. 865 bool SuppressFastISel = false; 866 if (LLVMBB == &Fn.getEntryBlock()) { 867 LowerArguments(LLVMBB); 868 869 // If any of the arguments has the byval attribute, forgo 870 // fast-isel in the entry block. 871 if (FastIS) { 872 unsigned j = 1; 873 for (Function::arg_iterator I = Fn.arg_begin(), E = Fn.arg_end(); 874 I != E; ++I, ++j) 875 if (Fn.paramHasAttr(j, Attribute::ByVal)) { 876 if (EnableFastISelVerbose || EnableFastISelAbort) 877 dbgs() << "FastISel skips entry block due to byval argument\n"; 878 SuppressFastISel = true; 879 break; 880 } 881 } 882 } 883 884 if (MMI && BB->isLandingPad()) { 885 // Add a label to mark the beginning of the landing pad. Deletion of the 886 // landing pad can thus be detected via the MachineModuleInfo. 887 MCSymbol *Label = MMI->addLandingPad(BB); 888 889 const TargetInstrDesc &II = TII.get(TargetOpcode::EH_LABEL); 890 BuildMI(BB, SDB->getCurDebugLoc(), II).addSym(Label); 891 892 // Mark exception register as live in. 893 unsigned Reg = TLI.getExceptionAddressRegister(); 894 if (Reg) BB->addLiveIn(Reg); 895 896 // Mark exception selector register as live in. 897 Reg = TLI.getExceptionSelectorRegister(); 898 if (Reg) BB->addLiveIn(Reg); 899 900 // FIXME: Hack around an exception handling flaw (PR1508): the personality 901 // function and list of typeids logically belong to the invoke (or, if you 902 // like, the basic block containing the invoke), and need to be associated 903 // with it in the dwarf exception handling tables. Currently however the 904 // information is provided by an intrinsic (eh.selector) that can be moved 905 // to unexpected places by the optimizers: if the unwind edge is critical, 906 // then breaking it can result in the intrinsics being in the successor of 907 // the landing pad, not the landing pad itself. This results 908 // in exceptions not being caught because no typeids are associated with 909 // the invoke. This may not be the only way things can go wrong, but it 910 // is the only way we try to work around for the moment. 911 BranchInst *Br = dyn_cast<BranchInst>(LLVMBB->getTerminator()); 912 913 if (Br && Br->isUnconditional()) { // Critical edge? 914 BasicBlock::iterator I, E; 915 for (I = LLVMBB->begin(), E = --LLVMBB->end(); I != E; ++I) 916 if (isa<EHSelectorInst>(I)) 917 break; 918 919 if (I == E) 920 // No catch info found - try to extract some from the successor. 921 CopyCatchInfo(Br->getSuccessor(0), LLVMBB, MMI, *FuncInfo); 922 } 923 } 924 925 // Before doing SelectionDAG ISel, see if FastISel has been requested. 926 if (FastIS && !SuppressFastISel) { 927 // Emit code for any incoming arguments. This must happen before 928 // beginning FastISel on the entry block. 929 if (LLVMBB == &Fn.getEntryBlock()) { 930 CurDAG->setRoot(SDB->getControlRoot()); 931 CodeGenAndEmitDAG(); 932 SDB->clear(); 933 } 934 FastIS->startNewBlock(BB); 935 // Do FastISel on as many instructions as possible. 936 for (; BI != End; ++BI) { 937 // Just before the terminator instruction, insert instructions to 938 // feed PHI nodes in successor blocks. 939 if (isa<TerminatorInst>(BI)) 940 if (!HandlePHINodesInSuccessorBlocksFast(LLVMBB, FastIS)) { 941 ++NumFastIselFailures; 942 ResetDebugLoc(SDB, FastIS); 943 if (EnableFastISelVerbose || EnableFastISelAbort) { 944 dbgs() << "FastISel miss: "; 945 BI->dump(); 946 } 947 assert(!EnableFastISelAbort && 948 "FastISel didn't handle a PHI in a successor"); 949 break; 950 } 951 952 SetDebugLoc(BI, SDB, FastIS, &MF); 953 954 // Try to select the instruction with FastISel. 955 if (FastIS->SelectInstruction(BI)) { 956 ResetDebugLoc(SDB, FastIS); 957 continue; 958 } 959 960 // Clear out the debug location so that it doesn't carry over to 961 // unrelated instructions. 962 ResetDebugLoc(SDB, FastIS); 963 964 // Then handle certain instructions as single-LLVM-Instruction blocks. 965 if (isa<CallInst>(BI)) { 966 ++NumFastIselFailures; 967 if (EnableFastISelVerbose || EnableFastISelAbort) { 968 dbgs() << "FastISel missed call: "; 969 BI->dump(); 970 } 971 972 if (!BI->getType()->isVoidTy()) { 973 unsigned &R = FuncInfo->ValueMap[BI]; 974 if (!R) 975 R = FuncInfo->CreateRegForValue(BI); 976 } 977 978 bool HadTailCall = false; 979 SelectBasicBlock(LLVMBB, BI, llvm::next(BI), HadTailCall); 980 981 // If the call was emitted as a tail call, we're done with the block. 982 if (HadTailCall) { 983 BI = End; 984 break; 985 } 986 987 // If the instruction was codegen'd with multiple blocks, 988 // inform the FastISel object where to resume inserting. 989 FastIS->setCurrentBlock(BB); 990 continue; 991 } 992 993 // Otherwise, give up on FastISel for the rest of the block. 994 // For now, be a little lenient about non-branch terminators. 995 if (!isa<TerminatorInst>(BI) || isa<BranchInst>(BI)) { 996 ++NumFastIselFailures; 997 if (EnableFastISelVerbose || EnableFastISelAbort) { 998 dbgs() << "FastISel miss: "; 999 BI->dump(); 1000 } 1001 if (EnableFastISelAbort) 1002 // The "fast" selector couldn't handle something and bailed. 1003 // For the purpose of debugging, just abort. 1004 llvm_unreachable("FastISel didn't select the entire block"); 1005 } 1006 break; 1007 } 1008 } 1009 1010 // Run SelectionDAG instruction selection on the remainder of the block 1011 // not handled by FastISel. If FastISel is not run, this is the entire 1012 // block. 1013 if (BI != End) { 1014 bool HadTailCall; 1015 SelectBasicBlock(LLVMBB, BI, End, HadTailCall); 1016 } 1017 1018 FinishBasicBlock(); 1019 } 1020 1021 delete FastIS; 1022} 1023 1024void 1025SelectionDAGISel::FinishBasicBlock() { 1026 1027 DEBUG(dbgs() << "Target-post-processed machine code:\n"); 1028 DEBUG(BB->dump()); 1029 1030 DEBUG(dbgs() << "Total amount of phi nodes to update: " 1031 << SDB->PHINodesToUpdate.size() << "\n"); 1032 DEBUG(for (unsigned i = 0, e = SDB->PHINodesToUpdate.size(); i != e; ++i) 1033 dbgs() << "Node " << i << " : (" 1034 << SDB->PHINodesToUpdate[i].first 1035 << ", " << SDB->PHINodesToUpdate[i].second << ")\n"); 1036 1037 // Next, now that we know what the last MBB the LLVM BB expanded is, update 1038 // PHI nodes in successors. 1039 if (SDB->SwitchCases.empty() && 1040 SDB->JTCases.empty() && 1041 SDB->BitTestCases.empty()) { 1042 for (unsigned i = 0, e = SDB->PHINodesToUpdate.size(); i != e; ++i) { 1043 MachineInstr *PHI = SDB->PHINodesToUpdate[i].first; 1044 assert(PHI->isPHI() && 1045 "This is not a machine PHI node that we are updating!"); 1046 if (!BB->isSuccessor(PHI->getParent())) 1047 continue; 1048 PHI->addOperand(MachineOperand::CreateReg(SDB->PHINodesToUpdate[i].second, 1049 false)); 1050 PHI->addOperand(MachineOperand::CreateMBB(BB)); 1051 } 1052 SDB->PHINodesToUpdate.clear(); 1053 return; 1054 } 1055 1056 for (unsigned i = 0, e = SDB->BitTestCases.size(); i != e; ++i) { 1057 // Lower header first, if it wasn't already lowered 1058 if (!SDB->BitTestCases[i].Emitted) { 1059 // Set the current basic block to the mbb we wish to insert the code into 1060 BB = SDB->BitTestCases[i].Parent; 1061 SDB->setCurrentBasicBlock(BB); 1062 // Emit the code 1063 SDB->visitBitTestHeader(SDB->BitTestCases[i]); 1064 CurDAG->setRoot(SDB->getRoot()); 1065 CodeGenAndEmitDAG(); 1066 SDB->clear(); 1067 } 1068 1069 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size(); j != ej; ++j) { 1070 // Set the current basic block to the mbb we wish to insert the code into 1071 BB = SDB->BitTestCases[i].Cases[j].ThisBB; 1072 SDB->setCurrentBasicBlock(BB); 1073 // Emit the code 1074 if (j+1 != ej) 1075 SDB->visitBitTestCase(SDB->BitTestCases[i].Cases[j+1].ThisBB, 1076 SDB->BitTestCases[i].Reg, 1077 SDB->BitTestCases[i].Cases[j]); 1078 else 1079 SDB->visitBitTestCase(SDB->BitTestCases[i].Default, 1080 SDB->BitTestCases[i].Reg, 1081 SDB->BitTestCases[i].Cases[j]); 1082 1083 1084 CurDAG->setRoot(SDB->getRoot()); 1085 CodeGenAndEmitDAG(); 1086 SDB->clear(); 1087 } 1088 1089 // Update PHI Nodes 1090 for (unsigned pi = 0, pe = SDB->PHINodesToUpdate.size(); pi != pe; ++pi) { 1091 MachineInstr *PHI = SDB->PHINodesToUpdate[pi].first; 1092 MachineBasicBlock *PHIBB = PHI->getParent(); 1093 assert(PHI->isPHI() && 1094 "This is not a machine PHI node that we are updating!"); 1095 // This is "default" BB. We have two jumps to it. From "header" BB and 1096 // from last "case" BB. 1097 if (PHIBB == SDB->BitTestCases[i].Default) { 1098 PHI->addOperand(MachineOperand:: 1099 CreateReg(SDB->PHINodesToUpdate[pi].second, false)); 1100 PHI->addOperand(MachineOperand::CreateMBB(SDB->BitTestCases[i].Parent)); 1101 PHI->addOperand(MachineOperand:: 1102 CreateReg(SDB->PHINodesToUpdate[pi].second, false)); 1103 PHI->addOperand(MachineOperand::CreateMBB(SDB->BitTestCases[i].Cases. 1104 back().ThisBB)); 1105 } 1106 // One of "cases" BB. 1107 for (unsigned j = 0, ej = SDB->BitTestCases[i].Cases.size(); 1108 j != ej; ++j) { 1109 MachineBasicBlock* cBB = SDB->BitTestCases[i].Cases[j].ThisBB; 1110 if (cBB->isSuccessor(PHIBB)) { 1111 PHI->addOperand(MachineOperand:: 1112 CreateReg(SDB->PHINodesToUpdate[pi].second, false)); 1113 PHI->addOperand(MachineOperand::CreateMBB(cBB)); 1114 } 1115 } 1116 } 1117 } 1118 SDB->BitTestCases.clear(); 1119 1120 // If the JumpTable record is filled in, then we need to emit a jump table. 1121 // Updating the PHI nodes is tricky in this case, since we need to determine 1122 // whether the PHI is a successor of the range check MBB or the jump table MBB 1123 for (unsigned i = 0, e = SDB->JTCases.size(); i != e; ++i) { 1124 // Lower header first, if it wasn't already lowered 1125 if (!SDB->JTCases[i].first.Emitted) { 1126 // Set the current basic block to the mbb we wish to insert the code into 1127 BB = SDB->JTCases[i].first.HeaderBB; 1128 SDB->setCurrentBasicBlock(BB); 1129 // Emit the code 1130 SDB->visitJumpTableHeader(SDB->JTCases[i].second, SDB->JTCases[i].first); 1131 CurDAG->setRoot(SDB->getRoot()); 1132 CodeGenAndEmitDAG(); 1133 SDB->clear(); 1134 } 1135 1136 // Set the current basic block to the mbb we wish to insert the code into 1137 BB = SDB->JTCases[i].second.MBB; 1138 SDB->setCurrentBasicBlock(BB); 1139 // Emit the code 1140 SDB->visitJumpTable(SDB->JTCases[i].second); 1141 CurDAG->setRoot(SDB->getRoot()); 1142 CodeGenAndEmitDAG(); 1143 SDB->clear(); 1144 1145 // Update PHI Nodes 1146 for (unsigned pi = 0, pe = SDB->PHINodesToUpdate.size(); pi != pe; ++pi) { 1147 MachineInstr *PHI = SDB->PHINodesToUpdate[pi].first; 1148 MachineBasicBlock *PHIBB = PHI->getParent(); 1149 assert(PHI->isPHI() && 1150 "This is not a machine PHI node that we are updating!"); 1151 // "default" BB. We can go there only from header BB. 1152 if (PHIBB == SDB->JTCases[i].second.Default) { 1153 PHI->addOperand 1154 (MachineOperand::CreateReg(SDB->PHINodesToUpdate[pi].second, false)); 1155 PHI->addOperand 1156 (MachineOperand::CreateMBB(SDB->JTCases[i].first.HeaderBB)); 1157 } 1158 // JT BB. Just iterate over successors here 1159 if (BB->isSuccessor(PHIBB)) { 1160 PHI->addOperand 1161 (MachineOperand::CreateReg(SDB->PHINodesToUpdate[pi].second, false)); 1162 PHI->addOperand(MachineOperand::CreateMBB(BB)); 1163 } 1164 } 1165 } 1166 SDB->JTCases.clear(); 1167 1168 // If the switch block involved a branch to one of the actual successors, we 1169 // need to update PHI nodes in that block. 1170 for (unsigned i = 0, e = SDB->PHINodesToUpdate.size(); i != e; ++i) { 1171 MachineInstr *PHI = SDB->PHINodesToUpdate[i].first; 1172 assert(PHI->isPHI() && 1173 "This is not a machine PHI node that we are updating!"); 1174 if (BB->isSuccessor(PHI->getParent())) { 1175 PHI->addOperand(MachineOperand::CreateReg(SDB->PHINodesToUpdate[i].second, 1176 false)); 1177 PHI->addOperand(MachineOperand::CreateMBB(BB)); 1178 } 1179 } 1180 1181 // If we generated any switch lowering information, build and codegen any 1182 // additional DAGs necessary. 1183 for (unsigned i = 0, e = SDB->SwitchCases.size(); i != e; ++i) { 1184 // Set the current basic block to the mbb we wish to insert the code into 1185 MachineBasicBlock *ThisBB = BB = SDB->SwitchCases[i].ThisBB; 1186 SDB->setCurrentBasicBlock(BB); 1187 1188 // Emit the code 1189 SDB->visitSwitchCase(SDB->SwitchCases[i]); 1190 CurDAG->setRoot(SDB->getRoot()); 1191 CodeGenAndEmitDAG(); 1192 1193 // Handle any PHI nodes in successors of this chunk, as if we were coming 1194 // from the original BB before switch expansion. Note that PHI nodes can 1195 // occur multiple times in PHINodesToUpdate. We have to be very careful to 1196 // handle them the right number of times. 1197 while ((BB = SDB->SwitchCases[i].TrueBB)) { // Handle LHS and RHS. 1198 // If new BB's are created during scheduling, the edges may have been 1199 // updated. That is, the edge from ThisBB to BB may have been split and 1200 // BB's predecessor is now another block. 1201 DenseMap<MachineBasicBlock*, MachineBasicBlock*>::iterator EI = 1202 SDB->EdgeMapping.find(BB); 1203 if (EI != SDB->EdgeMapping.end()) 1204 ThisBB = EI->second; 1205 1206 // BB may have been removed from the CFG if a branch was constant folded. 1207 if (ThisBB->isSuccessor(BB)) { 1208 for (MachineBasicBlock::iterator Phi = BB->begin(); 1209 Phi != BB->end() && Phi->isPHI(); 1210 ++Phi) { 1211 // This value for this PHI node is recorded in PHINodesToUpdate. 1212 for (unsigned pn = 0; ; ++pn) { 1213 assert(pn != SDB->PHINodesToUpdate.size() && 1214 "Didn't find PHI entry!"); 1215 if (SDB->PHINodesToUpdate[pn].first == Phi) { 1216 Phi->addOperand(MachineOperand:: 1217 CreateReg(SDB->PHINodesToUpdate[pn].second, 1218 false)); 1219 Phi->addOperand(MachineOperand::CreateMBB(ThisBB)); 1220 break; 1221 } 1222 } 1223 } 1224 } 1225 1226 // Don't process RHS if same block as LHS. 1227 if (BB == SDB->SwitchCases[i].FalseBB) 1228 SDB->SwitchCases[i].FalseBB = 0; 1229 1230 // If we haven't handled the RHS, do so now. Otherwise, we're done. 1231 SDB->SwitchCases[i].TrueBB = SDB->SwitchCases[i].FalseBB; 1232 SDB->SwitchCases[i].FalseBB = 0; 1233 } 1234 assert(SDB->SwitchCases[i].TrueBB == 0 && SDB->SwitchCases[i].FalseBB == 0); 1235 SDB->clear(); 1236 } 1237 SDB->SwitchCases.clear(); 1238 1239 SDB->PHINodesToUpdate.clear(); 1240} 1241 1242 1243/// Create the scheduler. If a specific scheduler was specified 1244/// via the SchedulerRegistry, use it, otherwise select the 1245/// one preferred by the target. 1246/// 1247ScheduleDAGSDNodes *SelectionDAGISel::CreateScheduler() { 1248 RegisterScheduler::FunctionPassCtor Ctor = RegisterScheduler::getDefault(); 1249 1250 if (!Ctor) { 1251 Ctor = ISHeuristic; 1252 RegisterScheduler::setDefault(Ctor); 1253 } 1254 1255 return Ctor(this, OptLevel); 1256} 1257 1258ScheduleHazardRecognizer *SelectionDAGISel::CreateTargetHazardRecognizer() { 1259 return new ScheduleHazardRecognizer(); 1260} 1261 1262//===----------------------------------------------------------------------===// 1263// Helper functions used by the generated instruction selector. 1264//===----------------------------------------------------------------------===// 1265// Calls to these methods are generated by tblgen. 1266 1267/// CheckAndMask - The isel is trying to match something like (and X, 255). If 1268/// the dag combiner simplified the 255, we still want to match. RHS is the 1269/// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value 1270/// specified in the .td file (e.g. 255). 1271bool SelectionDAGISel::CheckAndMask(SDValue LHS, ConstantSDNode *RHS, 1272 int64_t DesiredMaskS) const { 1273 const APInt &ActualMask = RHS->getAPIntValue(); 1274 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS); 1275 1276 // If the actual mask exactly matches, success! 1277 if (ActualMask == DesiredMask) 1278 return true; 1279 1280 // If the actual AND mask is allowing unallowed bits, this doesn't match. 1281 if (ActualMask.intersects(~DesiredMask)) 1282 return false; 1283 1284 // Otherwise, the DAG Combiner may have proven that the value coming in is 1285 // either already zero or is not demanded. Check for known zero input bits. 1286 APInt NeededMask = DesiredMask & ~ActualMask; 1287 if (CurDAG->MaskedValueIsZero(LHS, NeededMask)) 1288 return true; 1289 1290 // TODO: check to see if missing bits are just not demanded. 1291 1292 // Otherwise, this pattern doesn't match. 1293 return false; 1294} 1295 1296/// CheckOrMask - The isel is trying to match something like (or X, 255). If 1297/// the dag combiner simplified the 255, we still want to match. RHS is the 1298/// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value 1299/// specified in the .td file (e.g. 255). 1300bool SelectionDAGISel::CheckOrMask(SDValue LHS, ConstantSDNode *RHS, 1301 int64_t DesiredMaskS) const { 1302 const APInt &ActualMask = RHS->getAPIntValue(); 1303 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS); 1304 1305 // If the actual mask exactly matches, success! 1306 if (ActualMask == DesiredMask) 1307 return true; 1308 1309 // If the actual AND mask is allowing unallowed bits, this doesn't match. 1310 if (ActualMask.intersects(~DesiredMask)) 1311 return false; 1312 1313 // Otherwise, the DAG Combiner may have proven that the value coming in is 1314 // either already zero or is not demanded. Check for known zero input bits. 1315 APInt NeededMask = DesiredMask & ~ActualMask; 1316 1317 APInt KnownZero, KnownOne; 1318 CurDAG->ComputeMaskedBits(LHS, NeededMask, KnownZero, KnownOne); 1319 1320 // If all the missing bits in the or are already known to be set, match! 1321 if ((NeededMask & KnownOne) == NeededMask) 1322 return true; 1323 1324 // TODO: check to see if missing bits are just not demanded. 1325 1326 // Otherwise, this pattern doesn't match. 1327 return false; 1328} 1329 1330 1331/// SelectInlineAsmMemoryOperands - Calls to this are automatically generated 1332/// by tblgen. Others should not call it. 1333void SelectionDAGISel:: 1334SelectInlineAsmMemoryOperands(std::vector<SDValue> &Ops) { 1335 std::vector<SDValue> InOps; 1336 std::swap(InOps, Ops); 1337 1338 Ops.push_back(InOps[0]); // input chain. 1339 Ops.push_back(InOps[1]); // input asm string. 1340 1341 unsigned i = 2, e = InOps.size(); 1342 if (InOps[e-1].getValueType() == MVT::Flag) 1343 --e; // Don't process a flag operand if it is here. 1344 1345 while (i != e) { 1346 unsigned Flags = cast<ConstantSDNode>(InOps[i])->getZExtValue(); 1347 if ((Flags & 7) != 4 /*MEM*/) { 1348 // Just skip over this operand, copying the operands verbatim. 1349 Ops.insert(Ops.end(), InOps.begin()+i, 1350 InOps.begin()+i+InlineAsm::getNumOperandRegisters(Flags) + 1); 1351 i += InlineAsm::getNumOperandRegisters(Flags) + 1; 1352 } else { 1353 assert(InlineAsm::getNumOperandRegisters(Flags) == 1 && 1354 "Memory operand with multiple values?"); 1355 // Otherwise, this is a memory operand. Ask the target to select it. 1356 std::vector<SDValue> SelOps; 1357 if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps)) { 1358 llvm_report_error("Could not match memory address. Inline asm" 1359 " failure!"); 1360 } 1361 1362 // Add this to the output node. 1363 Ops.push_back(CurDAG->getTargetConstant(4/*MEM*/ | (SelOps.size()<< 3), 1364 MVT::i32)); 1365 Ops.insert(Ops.end(), SelOps.begin(), SelOps.end()); 1366 i += 2; 1367 } 1368 } 1369 1370 // Add the flag input back if present. 1371 if (e != InOps.size()) 1372 Ops.push_back(InOps.back()); 1373} 1374 1375/// findFlagUse - Return use of EVT::Flag value produced by the specified 1376/// SDNode. 1377/// 1378static SDNode *findFlagUse(SDNode *N) { 1379 unsigned FlagResNo = N->getNumValues()-1; 1380 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) { 1381 SDUse &Use = I.getUse(); 1382 if (Use.getResNo() == FlagResNo) 1383 return Use.getUser(); 1384 } 1385 return NULL; 1386} 1387 1388/// findNonImmUse - Return true if "Use" is a non-immediate use of "Def". 1389/// This function recursively traverses up the operand chain, ignoring 1390/// certain nodes. 1391static bool findNonImmUse(SDNode *Use, SDNode* Def, SDNode *ImmedUse, 1392 SDNode *Root, SmallPtrSet<SDNode*, 16> &Visited, 1393 bool IgnoreChains) { 1394 // The NodeID's are given uniques ID's where a node ID is guaranteed to be 1395 // greater than all of its (recursive) operands. If we scan to a point where 1396 // 'use' is smaller than the node we're scanning for, then we know we will 1397 // never find it. 1398 // 1399 // The Use may be -1 (unassigned) if it is a newly allocated node. This can 1400 // happen because we scan down to newly selected nodes in the case of flag 1401 // uses. 1402 if ((Use->getNodeId() < Def->getNodeId() && Use->getNodeId() != -1)) 1403 return false; 1404 1405 // Don't revisit nodes if we already scanned it and didn't fail, we know we 1406 // won't fail if we scan it again. 1407 if (!Visited.insert(Use)) 1408 return false; 1409 1410 for (unsigned i = 0, e = Use->getNumOperands(); i != e; ++i) { 1411 // Ignore chain uses, they are validated by HandleMergeInputChains. 1412 if (Use->getOperand(i).getValueType() == MVT::Other && IgnoreChains) 1413 continue; 1414 1415 SDNode *N = Use->getOperand(i).getNode(); 1416 if (N == Def) { 1417 if (Use == ImmedUse || Use == Root) 1418 continue; // We are not looking for immediate use. 1419 assert(N != Root); 1420 return true; 1421 } 1422 1423 // Traverse up the operand chain. 1424 if (findNonImmUse(N, Def, ImmedUse, Root, Visited, IgnoreChains)) 1425 return true; 1426 } 1427 return false; 1428} 1429 1430/// IsProfitableToFold - Returns true if it's profitable to fold the specific 1431/// operand node N of U during instruction selection that starts at Root. 1432bool SelectionDAGISel::IsProfitableToFold(SDValue N, SDNode *U, 1433 SDNode *Root) const { 1434 if (OptLevel == CodeGenOpt::None) return false; 1435 return N.hasOneUse(); 1436} 1437 1438/// IsLegalToFold - Returns true if the specific operand node N of 1439/// U can be folded during instruction selection that starts at Root. 1440bool SelectionDAGISel::IsLegalToFold(SDValue N, SDNode *U, SDNode *Root, 1441 bool IgnoreChains) const { 1442 if (OptLevel == CodeGenOpt::None) return false; 1443 1444 // If Root use can somehow reach N through a path that that doesn't contain 1445 // U then folding N would create a cycle. e.g. In the following 1446 // diagram, Root can reach N through X. If N is folded into into Root, then 1447 // X is both a predecessor and a successor of U. 1448 // 1449 // [N*] // 1450 // ^ ^ // 1451 // / \ // 1452 // [U*] [X]? // 1453 // ^ ^ // 1454 // \ / // 1455 // \ / // 1456 // [Root*] // 1457 // 1458 // * indicates nodes to be folded together. 1459 // 1460 // If Root produces a flag, then it gets (even more) interesting. Since it 1461 // will be "glued" together with its flag use in the scheduler, we need to 1462 // check if it might reach N. 1463 // 1464 // [N*] // 1465 // ^ ^ // 1466 // / \ // 1467 // [U*] [X]? // 1468 // ^ ^ // 1469 // \ \ // 1470 // \ | // 1471 // [Root*] | // 1472 // ^ | // 1473 // f | // 1474 // | / // 1475 // [Y] / // 1476 // ^ / // 1477 // f / // 1478 // | / // 1479 // [FU] // 1480 // 1481 // If FU (flag use) indirectly reaches N (the load), and Root folds N 1482 // (call it Fold), then X is a predecessor of FU and a successor of 1483 // Fold. But since Fold and FU are flagged together, this will create 1484 // a cycle in the scheduling graph. 1485 1486 // If the node has flags, walk down the graph to the "lowest" node in the 1487 // flagged set. 1488 EVT VT = Root->getValueType(Root->getNumValues()-1); 1489 while (VT == MVT::Flag) { 1490 SDNode *FU = findFlagUse(Root); 1491 if (FU == NULL) 1492 break; 1493 Root = FU; 1494 VT = Root->getValueType(Root->getNumValues()-1); 1495 1496 // If our query node has a flag result with a use, we've walked up it. If 1497 // the user (which has already been selected) has a chain or indirectly uses 1498 // the chain, our WalkChainUsers predicate will not consider it. Because of 1499 // this, we cannot ignore chains in this predicate. 1500 IgnoreChains = false; 1501 } 1502 1503 1504 SmallPtrSet<SDNode*, 16> Visited; 1505 return !findNonImmUse(Root, N.getNode(), U, Root, Visited, IgnoreChains); 1506} 1507 1508SDNode *SelectionDAGISel::Select_INLINEASM(SDNode *N) { 1509 std::vector<SDValue> Ops(N->op_begin(), N->op_end()); 1510 SelectInlineAsmMemoryOperands(Ops); 1511 1512 std::vector<EVT> VTs; 1513 VTs.push_back(MVT::Other); 1514 VTs.push_back(MVT::Flag); 1515 SDValue New = CurDAG->getNode(ISD::INLINEASM, N->getDebugLoc(), 1516 VTs, &Ops[0], Ops.size()); 1517 New->setNodeId(-1); 1518 return New.getNode(); 1519} 1520 1521SDNode *SelectionDAGISel::Select_UNDEF(SDNode *N) { 1522 return CurDAG->SelectNodeTo(N, TargetOpcode::IMPLICIT_DEF,N->getValueType(0)); 1523} 1524 1525/// GetVBR - decode a vbr encoding whose top bit is set. 1526ALWAYS_INLINE static uint64_t 1527GetVBR(uint64_t Val, const unsigned char *MatcherTable, unsigned &Idx) { 1528 assert(Val >= 128 && "Not a VBR"); 1529 Val &= 127; // Remove first vbr bit. 1530 1531 unsigned Shift = 7; 1532 uint64_t NextBits; 1533 do { 1534 NextBits = MatcherTable[Idx++]; 1535 Val |= (NextBits&127) << Shift; 1536 Shift += 7; 1537 } while (NextBits & 128); 1538 1539 return Val; 1540} 1541 1542 1543/// UpdateChainsAndFlags - When a match is complete, this method updates uses of 1544/// interior flag and chain results to use the new flag and chain results. 1545void SelectionDAGISel:: 1546UpdateChainsAndFlags(SDNode *NodeToMatch, SDValue InputChain, 1547 const SmallVectorImpl<SDNode*> &ChainNodesMatched, 1548 SDValue InputFlag, 1549 const SmallVectorImpl<SDNode*> &FlagResultNodesMatched, 1550 bool isMorphNodeTo) { 1551 SmallVector<SDNode*, 4> NowDeadNodes; 1552 1553 ISelUpdater ISU(ISelPosition); 1554 1555 // Now that all the normal results are replaced, we replace the chain and 1556 // flag results if present. 1557 if (!ChainNodesMatched.empty()) { 1558 assert(InputChain.getNode() != 0 && 1559 "Matched input chains but didn't produce a chain"); 1560 // Loop over all of the nodes we matched that produced a chain result. 1561 // Replace all the chain results with the final chain we ended up with. 1562 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) { 1563 SDNode *ChainNode = ChainNodesMatched[i]; 1564 1565 // If this node was already deleted, don't look at it. 1566 if (ChainNode->getOpcode() == ISD::DELETED_NODE) 1567 continue; 1568 1569 // Don't replace the results of the root node if we're doing a 1570 // MorphNodeTo. 1571 if (ChainNode == NodeToMatch && isMorphNodeTo) 1572 continue; 1573 1574 SDValue ChainVal = SDValue(ChainNode, ChainNode->getNumValues()-1); 1575 if (ChainVal.getValueType() == MVT::Flag) 1576 ChainVal = ChainVal.getValue(ChainVal->getNumValues()-2); 1577 assert(ChainVal.getValueType() == MVT::Other && "Not a chain?"); 1578 CurDAG->ReplaceAllUsesOfValueWith(ChainVal, InputChain, &ISU); 1579 1580 // If the node became dead and we haven't already seen it, delete it. 1581 if (ChainNode->use_empty() && 1582 !std::count(NowDeadNodes.begin(), NowDeadNodes.end(), ChainNode)) 1583 NowDeadNodes.push_back(ChainNode); 1584 } 1585 } 1586 1587 // If the result produces a flag, update any flag results in the matched 1588 // pattern with the flag result. 1589 if (InputFlag.getNode() != 0) { 1590 // Handle any interior nodes explicitly marked. 1591 for (unsigned i = 0, e = FlagResultNodesMatched.size(); i != e; ++i) { 1592 SDNode *FRN = FlagResultNodesMatched[i]; 1593 1594 // If this node was already deleted, don't look at it. 1595 if (FRN->getOpcode() == ISD::DELETED_NODE) 1596 continue; 1597 1598 assert(FRN->getValueType(FRN->getNumValues()-1) == MVT::Flag && 1599 "Doesn't have a flag result"); 1600 CurDAG->ReplaceAllUsesOfValueWith(SDValue(FRN, FRN->getNumValues()-1), 1601 InputFlag, &ISU); 1602 1603 // If the node became dead and we haven't already seen it, delete it. 1604 if (FRN->use_empty() && 1605 !std::count(NowDeadNodes.begin(), NowDeadNodes.end(), FRN)) 1606 NowDeadNodes.push_back(FRN); 1607 } 1608 } 1609 1610 if (!NowDeadNodes.empty()) 1611 CurDAG->RemoveDeadNodes(NowDeadNodes, &ISU); 1612 1613 DEBUG(errs() << "ISEL: Match complete!\n"); 1614} 1615 1616enum ChainResult { 1617 CR_Simple, 1618 CR_InducesCycle, 1619 CR_LeadsToInteriorNode 1620}; 1621 1622/// WalkChainUsers - Walk down the users of the specified chained node that is 1623/// part of the pattern we're matching, looking at all of the users we find. 1624/// This determines whether something is an interior node, whether we have a 1625/// non-pattern node in between two pattern nodes (which prevent folding because 1626/// it would induce a cycle) and whether we have a TokenFactor node sandwiched 1627/// between pattern nodes (in which case the TF becomes part of the pattern). 1628/// 1629/// The walk we do here is guaranteed to be small because we quickly get down to 1630/// already selected nodes "below" us. 1631static ChainResult 1632WalkChainUsers(SDNode *ChainedNode, 1633 SmallVectorImpl<SDNode*> &ChainedNodesInPattern, 1634 SmallVectorImpl<SDNode*> &InteriorChainedNodes) { 1635 ChainResult Result = CR_Simple; 1636 1637 for (SDNode::use_iterator UI = ChainedNode->use_begin(), 1638 E = ChainedNode->use_end(); UI != E; ++UI) { 1639 // Make sure the use is of the chain, not some other value we produce. 1640 if (UI.getUse().getValueType() != MVT::Other) continue; 1641 1642 SDNode *User = *UI; 1643 1644 // If we see an already-selected machine node, then we've gone beyond the 1645 // pattern that we're selecting down into the already selected chunk of the 1646 // DAG. 1647 if (User->isMachineOpcode() || 1648 User->getOpcode() == ISD::HANDLENODE) // Root of the graph. 1649 continue; 1650 1651 if (User->getOpcode() == ISD::CopyToReg || 1652 User->getOpcode() == ISD::CopyFromReg || 1653 User->getOpcode() == ISD::INLINEASM || 1654 User->getOpcode() == ISD::EH_LABEL) { 1655 // If their node ID got reset to -1 then they've already been selected. 1656 // Treat them like a MachineOpcode. 1657 if (User->getNodeId() == -1) 1658 continue; 1659 } 1660 1661 // If we have a TokenFactor, we handle it specially. 1662 if (User->getOpcode() != ISD::TokenFactor) { 1663 // If the node isn't a token factor and isn't part of our pattern, then it 1664 // must be a random chained node in between two nodes we're selecting. 1665 // This happens when we have something like: 1666 // x = load ptr 1667 // call 1668 // y = x+4 1669 // store y -> ptr 1670 // Because we structurally match the load/store as a read/modify/write, 1671 // but the call is chained between them. We cannot fold in this case 1672 // because it would induce a cycle in the graph. 1673 if (!std::count(ChainedNodesInPattern.begin(), 1674 ChainedNodesInPattern.end(), User)) 1675 return CR_InducesCycle; 1676 1677 // Otherwise we found a node that is part of our pattern. For example in: 1678 // x = load ptr 1679 // y = x+4 1680 // store y -> ptr 1681 // This would happen when we're scanning down from the load and see the 1682 // store as a user. Record that there is a use of ChainedNode that is 1683 // part of the pattern and keep scanning uses. 1684 Result = CR_LeadsToInteriorNode; 1685 InteriorChainedNodes.push_back(User); 1686 continue; 1687 } 1688 1689 // If we found a TokenFactor, there are two cases to consider: first if the 1690 // TokenFactor is just hanging "below" the pattern we're matching (i.e. no 1691 // uses of the TF are in our pattern) we just want to ignore it. Second, 1692 // the TokenFactor can be sandwiched in between two chained nodes, like so: 1693 // [Load chain] 1694 // ^ 1695 // | 1696 // [Load] 1697 // ^ ^ 1698 // | \ DAG's like cheese 1699 // / \ do you? 1700 // / | 1701 // [TokenFactor] [Op] 1702 // ^ ^ 1703 // | | 1704 // \ / 1705 // \ / 1706 // [Store] 1707 // 1708 // In this case, the TokenFactor becomes part of our match and we rewrite it 1709 // as a new TokenFactor. 1710 // 1711 // To distinguish these two cases, do a recursive walk down the uses. 1712 switch (WalkChainUsers(User, ChainedNodesInPattern, InteriorChainedNodes)) { 1713 case CR_Simple: 1714 // If the uses of the TokenFactor are just already-selected nodes, ignore 1715 // it, it is "below" our pattern. 1716 continue; 1717 case CR_InducesCycle: 1718 // If the uses of the TokenFactor lead to nodes that are not part of our 1719 // pattern that are not selected, folding would turn this into a cycle, 1720 // bail out now. 1721 return CR_InducesCycle; 1722 case CR_LeadsToInteriorNode: 1723 break; // Otherwise, keep processing. 1724 } 1725 1726 // Okay, we know we're in the interesting interior case. The TokenFactor 1727 // is now going to be considered part of the pattern so that we rewrite its 1728 // uses (it may have uses that are not part of the pattern) with the 1729 // ultimate chain result of the generated code. We will also add its chain 1730 // inputs as inputs to the ultimate TokenFactor we create. 1731 Result = CR_LeadsToInteriorNode; 1732 ChainedNodesInPattern.push_back(User); 1733 InteriorChainedNodes.push_back(User); 1734 continue; 1735 } 1736 1737 return Result; 1738} 1739 1740/// HandleMergeInputChains - This implements the OPC_EmitMergeInputChains 1741/// operation for when the pattern matched at least one node with a chains. The 1742/// input vector contains a list of all of the chained nodes that we match. We 1743/// must determine if this is a valid thing to cover (i.e. matching it won't 1744/// induce cycles in the DAG) and if so, creating a TokenFactor node. that will 1745/// be used as the input node chain for the generated nodes. 1746static SDValue 1747HandleMergeInputChains(SmallVectorImpl<SDNode*> &ChainNodesMatched, 1748 SelectionDAG *CurDAG) { 1749 // Walk all of the chained nodes we've matched, recursively scanning down the 1750 // users of the chain result. This adds any TokenFactor nodes that are caught 1751 // in between chained nodes to the chained and interior nodes list. 1752 SmallVector<SDNode*, 3> InteriorChainedNodes; 1753 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) { 1754 if (WalkChainUsers(ChainNodesMatched[i], ChainNodesMatched, 1755 InteriorChainedNodes) == CR_InducesCycle) 1756 return SDValue(); // Would induce a cycle. 1757 } 1758 1759 // Okay, we have walked all the matched nodes and collected TokenFactor nodes 1760 // that we are interested in. Form our input TokenFactor node. 1761 SmallVector<SDValue, 3> InputChains; 1762 for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) { 1763 // Add the input chain of this node to the InputChains list (which will be 1764 // the operands of the generated TokenFactor) if it's not an interior node. 1765 SDNode *N = ChainNodesMatched[i]; 1766 if (N->getOpcode() != ISD::TokenFactor) { 1767 if (std::count(InteriorChainedNodes.begin(),InteriorChainedNodes.end(),N)) 1768 continue; 1769 1770 // Otherwise, add the input chain. 1771 SDValue InChain = ChainNodesMatched[i]->getOperand(0); 1772 assert(InChain.getValueType() == MVT::Other && "Not a chain"); 1773 InputChains.push_back(InChain); 1774 continue; 1775 } 1776 1777 // If we have a token factor, we want to add all inputs of the token factor 1778 // that are not part of the pattern we're matching. 1779 for (unsigned op = 0, e = N->getNumOperands(); op != e; ++op) { 1780 if (!std::count(ChainNodesMatched.begin(), ChainNodesMatched.end(), 1781 N->getOperand(op).getNode())) 1782 InputChains.push_back(N->getOperand(op)); 1783 } 1784 } 1785 1786 SDValue Res; 1787 if (InputChains.size() == 1) 1788 return InputChains[0]; 1789 return CurDAG->getNode(ISD::TokenFactor, ChainNodesMatched[0]->getDebugLoc(), 1790 MVT::Other, &InputChains[0], InputChains.size()); 1791} 1792 1793/// MorphNode - Handle morphing a node in place for the selector. 1794SDNode *SelectionDAGISel:: 1795MorphNode(SDNode *Node, unsigned TargetOpc, SDVTList VTList, 1796 const SDValue *Ops, unsigned NumOps, unsigned EmitNodeInfo) { 1797 // It is possible we're using MorphNodeTo to replace a node with no 1798 // normal results with one that has a normal result (or we could be 1799 // adding a chain) and the input could have flags and chains as well. 1800 // In this case we need to shift the operands down. 1801 // FIXME: This is a horrible hack and broken in obscure cases, no worse 1802 // than the old isel though. 1803 int OldFlagResultNo = -1, OldChainResultNo = -1; 1804 1805 unsigned NTMNumResults = Node->getNumValues(); 1806 if (Node->getValueType(NTMNumResults-1) == MVT::Flag) { 1807 OldFlagResultNo = NTMNumResults-1; 1808 if (NTMNumResults != 1 && 1809 Node->getValueType(NTMNumResults-2) == MVT::Other) 1810 OldChainResultNo = NTMNumResults-2; 1811 } else if (Node->getValueType(NTMNumResults-1) == MVT::Other) 1812 OldChainResultNo = NTMNumResults-1; 1813 1814 // Call the underlying SelectionDAG routine to do the transmogrification. Note 1815 // that this deletes operands of the old node that become dead. 1816 SDNode *Res = CurDAG->MorphNodeTo(Node, ~TargetOpc, VTList, Ops, NumOps); 1817 1818 // MorphNodeTo can operate in two ways: if an existing node with the 1819 // specified operands exists, it can just return it. Otherwise, it 1820 // updates the node in place to have the requested operands. 1821 if (Res == Node) { 1822 // If we updated the node in place, reset the node ID. To the isel, 1823 // this should be just like a newly allocated machine node. 1824 Res->setNodeId(-1); 1825 } 1826 1827 unsigned ResNumResults = Res->getNumValues(); 1828 // Move the flag if needed. 1829 if ((EmitNodeInfo & OPFL_FlagOutput) && OldFlagResultNo != -1 && 1830 (unsigned)OldFlagResultNo != ResNumResults-1) 1831 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldFlagResultNo), 1832 SDValue(Res, ResNumResults-1)); 1833 1834 if ((EmitNodeInfo & OPFL_FlagOutput) != 0) 1835 --ResNumResults; 1836 1837 // Move the chain reference if needed. 1838 if ((EmitNodeInfo & OPFL_Chain) && OldChainResultNo != -1 && 1839 (unsigned)OldChainResultNo != ResNumResults-1) 1840 CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldChainResultNo), 1841 SDValue(Res, ResNumResults-1)); 1842 1843 // Otherwise, no replacement happened because the node already exists. Replace 1844 // Uses of the old node with the new one. 1845 if (Res != Node) 1846 CurDAG->ReplaceAllUsesWith(Node, Res); 1847 1848 return Res; 1849} 1850 1851/// CheckPatternPredicate - Implements OP_CheckPatternPredicate. 1852ALWAYS_INLINE static bool 1853CheckSame(const unsigned char *MatcherTable, unsigned &MatcherIndex, 1854 SDValue N, const SmallVectorImpl<SDValue> &RecordedNodes) { 1855 // Accept if it is exactly the same as a previously recorded node. 1856 unsigned RecNo = MatcherTable[MatcherIndex++]; 1857 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame"); 1858 return N == RecordedNodes[RecNo]; 1859} 1860 1861/// CheckPatternPredicate - Implements OP_CheckPatternPredicate. 1862ALWAYS_INLINE static bool 1863CheckPatternPredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex, 1864 SelectionDAGISel &SDISel) { 1865 return SDISel.CheckPatternPredicate(MatcherTable[MatcherIndex++]); 1866} 1867 1868/// CheckNodePredicate - Implements OP_CheckNodePredicate. 1869ALWAYS_INLINE static bool 1870CheckNodePredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex, 1871 SelectionDAGISel &SDISel, SDNode *N) { 1872 return SDISel.CheckNodePredicate(N, MatcherTable[MatcherIndex++]); 1873} 1874 1875ALWAYS_INLINE static bool 1876CheckOpcode(const unsigned char *MatcherTable, unsigned &MatcherIndex, 1877 SDNode *N) { 1878 uint16_t Opc = MatcherTable[MatcherIndex++]; 1879 Opc |= (unsigned short)MatcherTable[MatcherIndex++] << 8; 1880 return N->getOpcode() == Opc; 1881} 1882 1883ALWAYS_INLINE static bool 1884CheckType(const unsigned char *MatcherTable, unsigned &MatcherIndex, 1885 SDValue N, const TargetLowering &TLI) { 1886 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++]; 1887 if (N.getValueType() == VT) return true; 1888 1889 // Handle the case when VT is iPTR. 1890 return VT == MVT::iPTR && N.getValueType() == TLI.getPointerTy(); 1891} 1892 1893ALWAYS_INLINE static bool 1894CheckChildType(const unsigned char *MatcherTable, unsigned &MatcherIndex, 1895 SDValue N, const TargetLowering &TLI, 1896 unsigned ChildNo) { 1897 if (ChildNo >= N.getNumOperands()) 1898 return false; // Match fails if out of range child #. 1899 return ::CheckType(MatcherTable, MatcherIndex, N.getOperand(ChildNo), TLI); 1900} 1901 1902 1903ALWAYS_INLINE static bool 1904CheckCondCode(const unsigned char *MatcherTable, unsigned &MatcherIndex, 1905 SDValue N) { 1906 return cast<CondCodeSDNode>(N)->get() == 1907 (ISD::CondCode)MatcherTable[MatcherIndex++]; 1908} 1909 1910ALWAYS_INLINE static bool 1911CheckValueType(const unsigned char *MatcherTable, unsigned &MatcherIndex, 1912 SDValue N, const TargetLowering &TLI) { 1913 MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++]; 1914 if (cast<VTSDNode>(N)->getVT() == VT) 1915 return true; 1916 1917 // Handle the case when VT is iPTR. 1918 return VT == MVT::iPTR && cast<VTSDNode>(N)->getVT() == TLI.getPointerTy(); 1919} 1920 1921ALWAYS_INLINE static bool 1922CheckInteger(const unsigned char *MatcherTable, unsigned &MatcherIndex, 1923 SDValue N) { 1924 int64_t Val = MatcherTable[MatcherIndex++]; 1925 if (Val & 128) 1926 Val = GetVBR(Val, MatcherTable, MatcherIndex); 1927 1928 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N); 1929 return C != 0 && C->getSExtValue() == Val; 1930} 1931 1932ALWAYS_INLINE static bool 1933CheckAndImm(const unsigned char *MatcherTable, unsigned &MatcherIndex, 1934 SDValue N, SelectionDAGISel &SDISel) { 1935 int64_t Val = MatcherTable[MatcherIndex++]; 1936 if (Val & 128) 1937 Val = GetVBR(Val, MatcherTable, MatcherIndex); 1938 1939 if (N->getOpcode() != ISD::AND) return false; 1940 1941 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1)); 1942 return C != 0 && SDISel.CheckAndMask(N.getOperand(0), C, Val); 1943} 1944 1945ALWAYS_INLINE static bool 1946CheckOrImm(const unsigned char *MatcherTable, unsigned &MatcherIndex, 1947 SDValue N, SelectionDAGISel &SDISel) { 1948 int64_t Val = MatcherTable[MatcherIndex++]; 1949 if (Val & 128) 1950 Val = GetVBR(Val, MatcherTable, MatcherIndex); 1951 1952 if (N->getOpcode() != ISD::OR) return false; 1953 1954 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1)); 1955 return C != 0 && SDISel.CheckOrMask(N.getOperand(0), C, Val); 1956} 1957 1958/// IsPredicateKnownToFail - If we know how and can do so without pushing a 1959/// scope, evaluate the current node. If the current predicate is known to 1960/// fail, set Result=true and return anything. If the current predicate is 1961/// known to pass, set Result=false and return the MatcherIndex to continue 1962/// with. If the current predicate is unknown, set Result=false and return the 1963/// MatcherIndex to continue with. 1964static unsigned IsPredicateKnownToFail(const unsigned char *Table, 1965 unsigned Index, SDValue N, 1966 bool &Result, SelectionDAGISel &SDISel, 1967 SmallVectorImpl<SDValue> &RecordedNodes){ 1968 switch (Table[Index++]) { 1969 default: 1970 Result = false; 1971 return Index-1; // Could not evaluate this predicate. 1972 case SelectionDAGISel::OPC_CheckSame: 1973 Result = !::CheckSame(Table, Index, N, RecordedNodes); 1974 return Index; 1975 case SelectionDAGISel::OPC_CheckPatternPredicate: 1976 Result = !::CheckPatternPredicate(Table, Index, SDISel); 1977 return Index; 1978 case SelectionDAGISel::OPC_CheckPredicate: 1979 Result = !::CheckNodePredicate(Table, Index, SDISel, N.getNode()); 1980 return Index; 1981 case SelectionDAGISel::OPC_CheckOpcode: 1982 Result = !::CheckOpcode(Table, Index, N.getNode()); 1983 return Index; 1984 case SelectionDAGISel::OPC_CheckType: 1985 Result = !::CheckType(Table, Index, N, SDISel.TLI); 1986 return Index; 1987 case SelectionDAGISel::OPC_CheckChild0Type: 1988 case SelectionDAGISel::OPC_CheckChild1Type: 1989 case SelectionDAGISel::OPC_CheckChild2Type: 1990 case SelectionDAGISel::OPC_CheckChild3Type: 1991 case SelectionDAGISel::OPC_CheckChild4Type: 1992 case SelectionDAGISel::OPC_CheckChild5Type: 1993 case SelectionDAGISel::OPC_CheckChild6Type: 1994 case SelectionDAGISel::OPC_CheckChild7Type: 1995 Result = !::CheckChildType(Table, Index, N, SDISel.TLI, 1996 Table[Index-1] - SelectionDAGISel::OPC_CheckChild0Type); 1997 return Index; 1998 case SelectionDAGISel::OPC_CheckCondCode: 1999 Result = !::CheckCondCode(Table, Index, N); 2000 return Index; 2001 case SelectionDAGISel::OPC_CheckValueType: 2002 Result = !::CheckValueType(Table, Index, N, SDISel.TLI); 2003 return Index; 2004 case SelectionDAGISel::OPC_CheckInteger: 2005 Result = !::CheckInteger(Table, Index, N); 2006 return Index; 2007 case SelectionDAGISel::OPC_CheckAndImm: 2008 Result = !::CheckAndImm(Table, Index, N, SDISel); 2009 return Index; 2010 case SelectionDAGISel::OPC_CheckOrImm: 2011 Result = !::CheckOrImm(Table, Index, N, SDISel); 2012 return Index; 2013 } 2014} 2015 2016 2017struct MatchScope { 2018 /// FailIndex - If this match fails, this is the index to continue with. 2019 unsigned FailIndex; 2020 2021 /// NodeStack - The node stack when the scope was formed. 2022 SmallVector<SDValue, 4> NodeStack; 2023 2024 /// NumRecordedNodes - The number of recorded nodes when the scope was formed. 2025 unsigned NumRecordedNodes; 2026 2027 /// NumMatchedMemRefs - The number of matched memref entries. 2028 unsigned NumMatchedMemRefs; 2029 2030 /// InputChain/InputFlag - The current chain/flag 2031 SDValue InputChain, InputFlag; 2032 2033 /// HasChainNodesMatched - True if the ChainNodesMatched list is non-empty. 2034 bool HasChainNodesMatched, HasFlagResultNodesMatched; 2035}; 2036 2037SDNode *SelectionDAGISel:: 2038SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable, 2039 unsigned TableSize) { 2040 // FIXME: Should these even be selected? Handle these cases in the caller? 2041 switch (NodeToMatch->getOpcode()) { 2042 default: 2043 break; 2044 case ISD::EntryToken: // These nodes remain the same. 2045 case ISD::BasicBlock: 2046 case ISD::Register: 2047 //case ISD::VALUETYPE: 2048 //case ISD::CONDCODE: 2049 case ISD::HANDLENODE: 2050 case ISD::TargetConstant: 2051 case ISD::TargetConstantFP: 2052 case ISD::TargetConstantPool: 2053 case ISD::TargetFrameIndex: 2054 case ISD::TargetExternalSymbol: 2055 case ISD::TargetBlockAddress: 2056 case ISD::TargetJumpTable: 2057 case ISD::TargetGlobalTLSAddress: 2058 case ISD::TargetGlobalAddress: 2059 case ISD::TokenFactor: 2060 case ISD::CopyFromReg: 2061 case ISD::CopyToReg: 2062 case ISD::EH_LABEL: 2063 NodeToMatch->setNodeId(-1); // Mark selected. 2064 return 0; 2065 case ISD::AssertSext: 2066 case ISD::AssertZext: 2067 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, 0), 2068 NodeToMatch->getOperand(0)); 2069 return 0; 2070 case ISD::INLINEASM: return Select_INLINEASM(NodeToMatch); 2071 case ISD::UNDEF: return Select_UNDEF(NodeToMatch); 2072 } 2073 2074 assert(!NodeToMatch->isMachineOpcode() && "Node already selected!"); 2075 2076 // Set up the node stack with NodeToMatch as the only node on the stack. 2077 SmallVector<SDValue, 8> NodeStack; 2078 SDValue N = SDValue(NodeToMatch, 0); 2079 NodeStack.push_back(N); 2080 2081 // MatchScopes - Scopes used when matching, if a match failure happens, this 2082 // indicates where to continue checking. 2083 SmallVector<MatchScope, 8> MatchScopes; 2084 2085 // RecordedNodes - This is the set of nodes that have been recorded by the 2086 // state machine. 2087 SmallVector<SDValue, 8> RecordedNodes; 2088 2089 // MatchedMemRefs - This is the set of MemRef's we've seen in the input 2090 // pattern. 2091 SmallVector<MachineMemOperand*, 2> MatchedMemRefs; 2092 2093 // These are the current input chain and flag for use when generating nodes. 2094 // Various Emit operations change these. For example, emitting a copytoreg 2095 // uses and updates these. 2096 SDValue InputChain, InputFlag; 2097 2098 // ChainNodesMatched - If a pattern matches nodes that have input/output 2099 // chains, the OPC_EmitMergeInputChains operation is emitted which indicates 2100 // which ones they are. The result is captured into this list so that we can 2101 // update the chain results when the pattern is complete. 2102 SmallVector<SDNode*, 3> ChainNodesMatched; 2103 SmallVector<SDNode*, 3> FlagResultNodesMatched; 2104 2105 DEBUG(errs() << "ISEL: Starting pattern match on root node: "; 2106 NodeToMatch->dump(CurDAG); 2107 errs() << '\n'); 2108 2109 // Determine where to start the interpreter. Normally we start at opcode #0, 2110 // but if the state machine starts with an OPC_SwitchOpcode, then we 2111 // accelerate the first lookup (which is guaranteed to be hot) with the 2112 // OpcodeOffset table. 2113 unsigned MatcherIndex = 0; 2114 2115 if (!OpcodeOffset.empty()) { 2116 // Already computed the OpcodeOffset table, just index into it. 2117 if (N.getOpcode() < OpcodeOffset.size()) 2118 MatcherIndex = OpcodeOffset[N.getOpcode()]; 2119 DEBUG(errs() << " Initial Opcode index to " << MatcherIndex << "\n"); 2120 2121 } else if (MatcherTable[0] == OPC_SwitchOpcode) { 2122 // Otherwise, the table isn't computed, but the state machine does start 2123 // with an OPC_SwitchOpcode instruction. Populate the table now, since this 2124 // is the first time we're selecting an instruction. 2125 unsigned Idx = 1; 2126 while (1) { 2127 // Get the size of this case. 2128 unsigned CaseSize = MatcherTable[Idx++]; 2129 if (CaseSize & 128) 2130 CaseSize = GetVBR(CaseSize, MatcherTable, Idx); 2131 if (CaseSize == 0) break; 2132 2133 // Get the opcode, add the index to the table. 2134 uint16_t Opc = MatcherTable[Idx++]; 2135 Opc |= (unsigned short)MatcherTable[Idx++] << 8; 2136 if (Opc >= OpcodeOffset.size()) 2137 OpcodeOffset.resize((Opc+1)*2); 2138 OpcodeOffset[Opc] = Idx; 2139 Idx += CaseSize; 2140 } 2141 2142 // Okay, do the lookup for the first opcode. 2143 if (N.getOpcode() < OpcodeOffset.size()) 2144 MatcherIndex = OpcodeOffset[N.getOpcode()]; 2145 } 2146 2147 while (1) { 2148 assert(MatcherIndex < TableSize && "Invalid index"); 2149#ifndef NDEBUG 2150 unsigned CurrentOpcodeIndex = MatcherIndex; 2151#endif 2152 BuiltinOpcodes Opcode = (BuiltinOpcodes)MatcherTable[MatcherIndex++]; 2153 switch (Opcode) { 2154 case OPC_Scope: { 2155 // Okay, the semantics of this operation are that we should push a scope 2156 // then evaluate the first child. However, pushing a scope only to have 2157 // the first check fail (which then pops it) is inefficient. If we can 2158 // determine immediately that the first check (or first several) will 2159 // immediately fail, don't even bother pushing a scope for them. 2160 unsigned FailIndex; 2161 2162 while (1) { 2163 unsigned NumToSkip = MatcherTable[MatcherIndex++]; 2164 if (NumToSkip & 128) 2165 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex); 2166 // Found the end of the scope with no match. 2167 if (NumToSkip == 0) { 2168 FailIndex = 0; 2169 break; 2170 } 2171 2172 FailIndex = MatcherIndex+NumToSkip; 2173 2174 unsigned MatcherIndexOfPredicate = MatcherIndex; 2175 (void)MatcherIndexOfPredicate; // silence warning. 2176 2177 // If we can't evaluate this predicate without pushing a scope (e.g. if 2178 // it is a 'MoveParent') or if the predicate succeeds on this node, we 2179 // push the scope and evaluate the full predicate chain. 2180 bool Result; 2181 MatcherIndex = IsPredicateKnownToFail(MatcherTable, MatcherIndex, N, 2182 Result, *this, RecordedNodes); 2183 if (!Result) 2184 break; 2185 2186 DEBUG(errs() << " Skipped scope entry (due to false predicate) at " 2187 << "index " << MatcherIndexOfPredicate 2188 << ", continuing at " << FailIndex << "\n"); 2189 ++NumDAGIselRetries; 2190 2191 // Otherwise, we know that this case of the Scope is guaranteed to fail, 2192 // move to the next case. 2193 MatcherIndex = FailIndex; 2194 } 2195 2196 // If the whole scope failed to match, bail. 2197 if (FailIndex == 0) break; 2198 2199 // Push a MatchScope which indicates where to go if the first child fails 2200 // to match. 2201 MatchScope NewEntry; 2202 NewEntry.FailIndex = FailIndex; 2203 NewEntry.NodeStack.append(NodeStack.begin(), NodeStack.end()); 2204 NewEntry.NumRecordedNodes = RecordedNodes.size(); 2205 NewEntry.NumMatchedMemRefs = MatchedMemRefs.size(); 2206 NewEntry.InputChain = InputChain; 2207 NewEntry.InputFlag = InputFlag; 2208 NewEntry.HasChainNodesMatched = !ChainNodesMatched.empty(); 2209 NewEntry.HasFlagResultNodesMatched = !FlagResultNodesMatched.empty(); 2210 MatchScopes.push_back(NewEntry); 2211 continue; 2212 } 2213 case OPC_RecordNode: 2214 // Remember this node, it may end up being an operand in the pattern. 2215 RecordedNodes.push_back(N); 2216 continue; 2217 2218 case OPC_RecordChild0: case OPC_RecordChild1: 2219 case OPC_RecordChild2: case OPC_RecordChild3: 2220 case OPC_RecordChild4: case OPC_RecordChild5: 2221 case OPC_RecordChild6: case OPC_RecordChild7: { 2222 unsigned ChildNo = Opcode-OPC_RecordChild0; 2223 if (ChildNo >= N.getNumOperands()) 2224 break; // Match fails if out of range child #. 2225 2226 RecordedNodes.push_back(N->getOperand(ChildNo)); 2227 continue; 2228 } 2229 case OPC_RecordMemRef: 2230 MatchedMemRefs.push_back(cast<MemSDNode>(N)->getMemOperand()); 2231 continue; 2232 2233 case OPC_CaptureFlagInput: 2234 // If the current node has an input flag, capture it in InputFlag. 2235 if (N->getNumOperands() != 0 && 2236 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag) 2237 InputFlag = N->getOperand(N->getNumOperands()-1); 2238 continue; 2239 2240 case OPC_MoveChild: { 2241 unsigned ChildNo = MatcherTable[MatcherIndex++]; 2242 if (ChildNo >= N.getNumOperands()) 2243 break; // Match fails if out of range child #. 2244 N = N.getOperand(ChildNo); 2245 NodeStack.push_back(N); 2246 continue; 2247 } 2248 2249 case OPC_MoveParent: 2250 // Pop the current node off the NodeStack. 2251 NodeStack.pop_back(); 2252 assert(!NodeStack.empty() && "Node stack imbalance!"); 2253 N = NodeStack.back(); 2254 continue; 2255 2256 case OPC_CheckSame: 2257 if (!::CheckSame(MatcherTable, MatcherIndex, N, RecordedNodes)) break; 2258 continue; 2259 case OPC_CheckPatternPredicate: 2260 if (!::CheckPatternPredicate(MatcherTable, MatcherIndex, *this)) break; 2261 continue; 2262 case OPC_CheckPredicate: 2263 if (!::CheckNodePredicate(MatcherTable, MatcherIndex, *this, 2264 N.getNode())) 2265 break; 2266 continue; 2267 case OPC_CheckComplexPat: { 2268 unsigned CPNum = MatcherTable[MatcherIndex++]; 2269 unsigned RecNo = MatcherTable[MatcherIndex++]; 2270 assert(RecNo < RecordedNodes.size() && "Invalid CheckComplexPat"); 2271 if (!CheckComplexPattern(NodeToMatch, RecordedNodes[RecNo], CPNum, 2272 RecordedNodes)) 2273 break; 2274 continue; 2275 } 2276 case OPC_CheckOpcode: 2277 if (!::CheckOpcode(MatcherTable, MatcherIndex, N.getNode())) break; 2278 continue; 2279 2280 case OPC_CheckType: 2281 if (!::CheckType(MatcherTable, MatcherIndex, N, TLI)) break; 2282 continue; 2283 2284 case OPC_SwitchOpcode: { 2285 unsigned CurNodeOpcode = N.getOpcode(); 2286 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart; 2287 unsigned CaseSize; 2288 while (1) { 2289 // Get the size of this case. 2290 CaseSize = MatcherTable[MatcherIndex++]; 2291 if (CaseSize & 128) 2292 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex); 2293 if (CaseSize == 0) break; 2294 2295 uint16_t Opc = MatcherTable[MatcherIndex++]; 2296 Opc |= (unsigned short)MatcherTable[MatcherIndex++] << 8; 2297 2298 // If the opcode matches, then we will execute this case. 2299 if (CurNodeOpcode == Opc) 2300 break; 2301 2302 // Otherwise, skip over this case. 2303 MatcherIndex += CaseSize; 2304 } 2305 2306 // If no cases matched, bail out. 2307 if (CaseSize == 0) break; 2308 2309 // Otherwise, execute the case we found. 2310 DEBUG(errs() << " OpcodeSwitch from " << SwitchStart 2311 << " to " << MatcherIndex << "\n"); 2312 continue; 2313 } 2314 2315 case OPC_SwitchType: { 2316 MVT::SimpleValueType CurNodeVT = N.getValueType().getSimpleVT().SimpleTy; 2317 unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart; 2318 unsigned CaseSize; 2319 while (1) { 2320 // Get the size of this case. 2321 CaseSize = MatcherTable[MatcherIndex++]; 2322 if (CaseSize & 128) 2323 CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex); 2324 if (CaseSize == 0) break; 2325 2326 MVT::SimpleValueType CaseVT = 2327 (MVT::SimpleValueType)MatcherTable[MatcherIndex++]; 2328 if (CaseVT == MVT::iPTR) 2329 CaseVT = TLI.getPointerTy().SimpleTy; 2330 2331 // If the VT matches, then we will execute this case. 2332 if (CurNodeVT == CaseVT) 2333 break; 2334 2335 // Otherwise, skip over this case. 2336 MatcherIndex += CaseSize; 2337 } 2338 2339 // If no cases matched, bail out. 2340 if (CaseSize == 0) break; 2341 2342 // Otherwise, execute the case we found. 2343 DEBUG(errs() << " TypeSwitch[" << EVT(CurNodeVT).getEVTString() 2344 << "] from " << SwitchStart << " to " << MatcherIndex<<'\n'); 2345 continue; 2346 } 2347 case OPC_CheckChild0Type: case OPC_CheckChild1Type: 2348 case OPC_CheckChild2Type: case OPC_CheckChild3Type: 2349 case OPC_CheckChild4Type: case OPC_CheckChild5Type: 2350 case OPC_CheckChild6Type: case OPC_CheckChild7Type: 2351 if (!::CheckChildType(MatcherTable, MatcherIndex, N, TLI, 2352 Opcode-OPC_CheckChild0Type)) 2353 break; 2354 continue; 2355 case OPC_CheckCondCode: 2356 if (!::CheckCondCode(MatcherTable, MatcherIndex, N)) break; 2357 continue; 2358 case OPC_CheckValueType: 2359 if (!::CheckValueType(MatcherTable, MatcherIndex, N, TLI)) break; 2360 continue; 2361 case OPC_CheckInteger: 2362 if (!::CheckInteger(MatcherTable, MatcherIndex, N)) break; 2363 continue; 2364 case OPC_CheckAndImm: 2365 if (!::CheckAndImm(MatcherTable, MatcherIndex, N, *this)) break; 2366 continue; 2367 case OPC_CheckOrImm: 2368 if (!::CheckOrImm(MatcherTable, MatcherIndex, N, *this)) break; 2369 continue; 2370 2371 case OPC_CheckFoldableChainNode: { 2372 assert(NodeStack.size() != 1 && "No parent node"); 2373 // Verify that all intermediate nodes between the root and this one have 2374 // a single use. 2375 bool HasMultipleUses = false; 2376 for (unsigned i = 1, e = NodeStack.size()-1; i != e; ++i) 2377 if (!NodeStack[i].hasOneUse()) { 2378 HasMultipleUses = true; 2379 break; 2380 } 2381 if (HasMultipleUses) break; 2382 2383 // Check to see that the target thinks this is profitable to fold and that 2384 // we can fold it without inducing cycles in the graph. 2385 if (!IsProfitableToFold(N, NodeStack[NodeStack.size()-2].getNode(), 2386 NodeToMatch) || 2387 !IsLegalToFold(N, NodeStack[NodeStack.size()-2].getNode(), 2388 NodeToMatch, true/*We validate our own chains*/)) 2389 break; 2390 2391 continue; 2392 } 2393 case OPC_EmitInteger: { 2394 MVT::SimpleValueType VT = 2395 (MVT::SimpleValueType)MatcherTable[MatcherIndex++]; 2396 int64_t Val = MatcherTable[MatcherIndex++]; 2397 if (Val & 128) 2398 Val = GetVBR(Val, MatcherTable, MatcherIndex); 2399 RecordedNodes.push_back(CurDAG->getTargetConstant(Val, VT)); 2400 continue; 2401 } 2402 case OPC_EmitRegister: { 2403 MVT::SimpleValueType VT = 2404 (MVT::SimpleValueType)MatcherTable[MatcherIndex++]; 2405 unsigned RegNo = MatcherTable[MatcherIndex++]; 2406 RecordedNodes.push_back(CurDAG->getRegister(RegNo, VT)); 2407 continue; 2408 } 2409 2410 case OPC_EmitConvertToTarget: { 2411 // Convert from IMM/FPIMM to target version. 2412 unsigned RecNo = MatcherTable[MatcherIndex++]; 2413 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame"); 2414 SDValue Imm = RecordedNodes[RecNo]; 2415 2416 if (Imm->getOpcode() == ISD::Constant) { 2417 int64_t Val = cast<ConstantSDNode>(Imm)->getZExtValue(); 2418 Imm = CurDAG->getTargetConstant(Val, Imm.getValueType()); 2419 } else if (Imm->getOpcode() == ISD::ConstantFP) { 2420 const ConstantFP *Val=cast<ConstantFPSDNode>(Imm)->getConstantFPValue(); 2421 Imm = CurDAG->getTargetConstantFP(*Val, Imm.getValueType()); 2422 } 2423 2424 RecordedNodes.push_back(Imm); 2425 continue; 2426 } 2427 2428 case OPC_EmitMergeInputChains1_0: // OPC_EmitMergeInputChains, 1, 0 2429 case OPC_EmitMergeInputChains1_1: { // OPC_EmitMergeInputChains, 1, 1 2430 // These are space-optimized forms of OPC_EmitMergeInputChains. 2431 assert(InputChain.getNode() == 0 && 2432 "EmitMergeInputChains should be the first chain producing node"); 2433 assert(ChainNodesMatched.empty() && 2434 "Should only have one EmitMergeInputChains per match"); 2435 2436 // Read all of the chained nodes. 2437 unsigned RecNo = Opcode == OPC_EmitMergeInputChains1_1; 2438 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame"); 2439 ChainNodesMatched.push_back(RecordedNodes[RecNo].getNode()); 2440 2441 // FIXME: What if other value results of the node have uses not matched 2442 // by this pattern? 2443 if (ChainNodesMatched.back() != NodeToMatch && 2444 !RecordedNodes[RecNo].hasOneUse()) { 2445 ChainNodesMatched.clear(); 2446 break; 2447 } 2448 2449 // Merge the input chains if they are not intra-pattern references. 2450 InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG); 2451 2452 if (InputChain.getNode() == 0) 2453 break; // Failed to merge. 2454 continue; 2455 } 2456 2457 case OPC_EmitMergeInputChains: { 2458 assert(InputChain.getNode() == 0 && 2459 "EmitMergeInputChains should be the first chain producing node"); 2460 // This node gets a list of nodes we matched in the input that have 2461 // chains. We want to token factor all of the input chains to these nodes 2462 // together. However, if any of the input chains is actually one of the 2463 // nodes matched in this pattern, then we have an intra-match reference. 2464 // Ignore these because the newly token factored chain should not refer to 2465 // the old nodes. 2466 unsigned NumChains = MatcherTable[MatcherIndex++]; 2467 assert(NumChains != 0 && "Can't TF zero chains"); 2468 2469 assert(ChainNodesMatched.empty() && 2470 "Should only have one EmitMergeInputChains per match"); 2471 2472 // Read all of the chained nodes. 2473 for (unsigned i = 0; i != NumChains; ++i) { 2474 unsigned RecNo = MatcherTable[MatcherIndex++]; 2475 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame"); 2476 ChainNodesMatched.push_back(RecordedNodes[RecNo].getNode()); 2477 2478 // FIXME: What if other value results of the node have uses not matched 2479 // by this pattern? 2480 if (ChainNodesMatched.back() != NodeToMatch && 2481 !RecordedNodes[RecNo].hasOneUse()) { 2482 ChainNodesMatched.clear(); 2483 break; 2484 } 2485 } 2486 2487 // If the inner loop broke out, the match fails. 2488 if (ChainNodesMatched.empty()) 2489 break; 2490 2491 // Merge the input chains if they are not intra-pattern references. 2492 InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG); 2493 2494 if (InputChain.getNode() == 0) 2495 break; // Failed to merge. 2496 2497 continue; 2498 } 2499 2500 case OPC_EmitCopyToReg: { 2501 unsigned RecNo = MatcherTable[MatcherIndex++]; 2502 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame"); 2503 unsigned DestPhysReg = MatcherTable[MatcherIndex++]; 2504 2505 if (InputChain.getNode() == 0) 2506 InputChain = CurDAG->getEntryNode(); 2507 2508 InputChain = CurDAG->getCopyToReg(InputChain, NodeToMatch->getDebugLoc(), 2509 DestPhysReg, RecordedNodes[RecNo], 2510 InputFlag); 2511 2512 InputFlag = InputChain.getValue(1); 2513 continue; 2514 } 2515 2516 case OPC_EmitNodeXForm: { 2517 unsigned XFormNo = MatcherTable[MatcherIndex++]; 2518 unsigned RecNo = MatcherTable[MatcherIndex++]; 2519 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame"); 2520 RecordedNodes.push_back(RunSDNodeXForm(RecordedNodes[RecNo], XFormNo)); 2521 continue; 2522 } 2523 2524 case OPC_EmitNode: 2525 case OPC_MorphNodeTo: { 2526 uint16_t TargetOpc = MatcherTable[MatcherIndex++]; 2527 TargetOpc |= (unsigned short)MatcherTable[MatcherIndex++] << 8; 2528 unsigned EmitNodeInfo = MatcherTable[MatcherIndex++]; 2529 // Get the result VT list. 2530 unsigned NumVTs = MatcherTable[MatcherIndex++]; 2531 SmallVector<EVT, 4> VTs; 2532 for (unsigned i = 0; i != NumVTs; ++i) { 2533 MVT::SimpleValueType VT = 2534 (MVT::SimpleValueType)MatcherTable[MatcherIndex++]; 2535 if (VT == MVT::iPTR) VT = TLI.getPointerTy().SimpleTy; 2536 VTs.push_back(VT); 2537 } 2538 2539 if (EmitNodeInfo & OPFL_Chain) 2540 VTs.push_back(MVT::Other); 2541 if (EmitNodeInfo & OPFL_FlagOutput) 2542 VTs.push_back(MVT::Flag); 2543 2544 // This is hot code, so optimize the two most common cases of 1 and 2 2545 // results. 2546 SDVTList VTList; 2547 if (VTs.size() == 1) 2548 VTList = CurDAG->getVTList(VTs[0]); 2549 else if (VTs.size() == 2) 2550 VTList = CurDAG->getVTList(VTs[0], VTs[1]); 2551 else 2552 VTList = CurDAG->getVTList(VTs.data(), VTs.size()); 2553 2554 // Get the operand list. 2555 unsigned NumOps = MatcherTable[MatcherIndex++]; 2556 SmallVector<SDValue, 8> Ops; 2557 for (unsigned i = 0; i != NumOps; ++i) { 2558 unsigned RecNo = MatcherTable[MatcherIndex++]; 2559 if (RecNo & 128) 2560 RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex); 2561 2562 assert(RecNo < RecordedNodes.size() && "Invalid EmitNode"); 2563 Ops.push_back(RecordedNodes[RecNo]); 2564 } 2565 2566 // If there are variadic operands to add, handle them now. 2567 if (EmitNodeInfo & OPFL_VariadicInfo) { 2568 // Determine the start index to copy from. 2569 unsigned FirstOpToCopy = getNumFixedFromVariadicInfo(EmitNodeInfo); 2570 FirstOpToCopy += (EmitNodeInfo & OPFL_Chain) ? 1 : 0; 2571 assert(NodeToMatch->getNumOperands() >= FirstOpToCopy && 2572 "Invalid variadic node"); 2573 // Copy all of the variadic operands, not including a potential flag 2574 // input. 2575 for (unsigned i = FirstOpToCopy, e = NodeToMatch->getNumOperands(); 2576 i != e; ++i) { 2577 SDValue V = NodeToMatch->getOperand(i); 2578 if (V.getValueType() == MVT::Flag) break; 2579 Ops.push_back(V); 2580 } 2581 } 2582 2583 // If this has chain/flag inputs, add them. 2584 if (EmitNodeInfo & OPFL_Chain) 2585 Ops.push_back(InputChain); 2586 if ((EmitNodeInfo & OPFL_FlagInput) && InputFlag.getNode() != 0) 2587 Ops.push_back(InputFlag); 2588 2589 // Create the node. 2590 SDNode *Res = 0; 2591 if (Opcode != OPC_MorphNodeTo) { 2592 // If this is a normal EmitNode command, just create the new node and 2593 // add the results to the RecordedNodes list. 2594 Res = CurDAG->getMachineNode(TargetOpc, NodeToMatch->getDebugLoc(), 2595 VTList, Ops.data(), Ops.size()); 2596 2597 // Add all the non-flag/non-chain results to the RecordedNodes list. 2598 for (unsigned i = 0, e = VTs.size(); i != e; ++i) { 2599 if (VTs[i] == MVT::Other || VTs[i] == MVT::Flag) break; 2600 RecordedNodes.push_back(SDValue(Res, i)); 2601 } 2602 2603 } else { 2604 Res = MorphNode(NodeToMatch, TargetOpc, VTList, Ops.data(), Ops.size(), 2605 EmitNodeInfo); 2606 } 2607 2608 // If the node had chain/flag results, update our notion of the current 2609 // chain and flag. 2610 if (EmitNodeInfo & OPFL_FlagOutput) { 2611 InputFlag = SDValue(Res, VTs.size()-1); 2612 if (EmitNodeInfo & OPFL_Chain) 2613 InputChain = SDValue(Res, VTs.size()-2); 2614 } else if (EmitNodeInfo & OPFL_Chain) 2615 InputChain = SDValue(Res, VTs.size()-1); 2616 2617 // If the OPFL_MemRefs flag is set on this node, slap all of the 2618 // accumulated memrefs onto it. 2619 // 2620 // FIXME: This is vastly incorrect for patterns with multiple outputs 2621 // instructions that access memory and for ComplexPatterns that match 2622 // loads. 2623 if (EmitNodeInfo & OPFL_MemRefs) { 2624 MachineSDNode::mmo_iterator MemRefs = 2625 MF->allocateMemRefsArray(MatchedMemRefs.size()); 2626 std::copy(MatchedMemRefs.begin(), MatchedMemRefs.end(), MemRefs); 2627 cast<MachineSDNode>(Res) 2628 ->setMemRefs(MemRefs, MemRefs + MatchedMemRefs.size()); 2629 } 2630 2631 DEBUG(errs() << " " 2632 << (Opcode == OPC_MorphNodeTo ? "Morphed" : "Created") 2633 << " node: "; Res->dump(CurDAG); errs() << "\n"); 2634 2635 // If this was a MorphNodeTo then we're completely done! 2636 if (Opcode == OPC_MorphNodeTo) { 2637 // Update chain and flag uses. 2638 UpdateChainsAndFlags(NodeToMatch, InputChain, ChainNodesMatched, 2639 InputFlag, FlagResultNodesMatched, true); 2640 return Res; 2641 } 2642 2643 continue; 2644 } 2645 2646 case OPC_MarkFlagResults: { 2647 unsigned NumNodes = MatcherTable[MatcherIndex++]; 2648 2649 // Read and remember all the flag-result nodes. 2650 for (unsigned i = 0; i != NumNodes; ++i) { 2651 unsigned RecNo = MatcherTable[MatcherIndex++]; 2652 if (RecNo & 128) 2653 RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex); 2654 2655 assert(RecNo < RecordedNodes.size() && "Invalid CheckSame"); 2656 FlagResultNodesMatched.push_back(RecordedNodes[RecNo].getNode()); 2657 } 2658 continue; 2659 } 2660 2661 case OPC_CompleteMatch: { 2662 // The match has been completed, and any new nodes (if any) have been 2663 // created. Patch up references to the matched dag to use the newly 2664 // created nodes. 2665 unsigned NumResults = MatcherTable[MatcherIndex++]; 2666 2667 for (unsigned i = 0; i != NumResults; ++i) { 2668 unsigned ResSlot = MatcherTable[MatcherIndex++]; 2669 if (ResSlot & 128) 2670 ResSlot = GetVBR(ResSlot, MatcherTable, MatcherIndex); 2671 2672 assert(ResSlot < RecordedNodes.size() && "Invalid CheckSame"); 2673 SDValue Res = RecordedNodes[ResSlot]; 2674 2675 assert(i < NodeToMatch->getNumValues() && 2676 NodeToMatch->getValueType(i) != MVT::Other && 2677 NodeToMatch->getValueType(i) != MVT::Flag && 2678 "Invalid number of results to complete!"); 2679 assert((NodeToMatch->getValueType(i) == Res.getValueType() || 2680 NodeToMatch->getValueType(i) == MVT::iPTR || 2681 Res.getValueType() == MVT::iPTR || 2682 NodeToMatch->getValueType(i).getSizeInBits() == 2683 Res.getValueType().getSizeInBits()) && 2684 "invalid replacement"); 2685 CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, i), Res); 2686 } 2687 2688 // If the root node defines a flag, add it to the flag nodes to update 2689 // list. 2690 if (NodeToMatch->getValueType(NodeToMatch->getNumValues()-1) == MVT::Flag) 2691 FlagResultNodesMatched.push_back(NodeToMatch); 2692 2693 // Update chain and flag uses. 2694 UpdateChainsAndFlags(NodeToMatch, InputChain, ChainNodesMatched, 2695 InputFlag, FlagResultNodesMatched, false); 2696 2697 assert(NodeToMatch->use_empty() && 2698 "Didn't replace all uses of the node?"); 2699 2700 // FIXME: We just return here, which interacts correctly with SelectRoot 2701 // above. We should fix this to not return an SDNode* anymore. 2702 return 0; 2703 } 2704 } 2705 2706 // If the code reached this point, then the match failed. See if there is 2707 // another child to try in the current 'Scope', otherwise pop it until we 2708 // find a case to check. 2709 DEBUG(errs() << " Match failed at index " << CurrentOpcodeIndex << "\n"); 2710 ++NumDAGIselRetries; 2711 while (1) { 2712 if (MatchScopes.empty()) { 2713 CannotYetSelect(NodeToMatch); 2714 return 0; 2715 } 2716 2717 // Restore the interpreter state back to the point where the scope was 2718 // formed. 2719 MatchScope &LastScope = MatchScopes.back(); 2720 RecordedNodes.resize(LastScope.NumRecordedNodes); 2721 NodeStack.clear(); 2722 NodeStack.append(LastScope.NodeStack.begin(), LastScope.NodeStack.end()); 2723 N = NodeStack.back(); 2724 2725 if (LastScope.NumMatchedMemRefs != MatchedMemRefs.size()) 2726 MatchedMemRefs.resize(LastScope.NumMatchedMemRefs); 2727 MatcherIndex = LastScope.FailIndex; 2728 2729 DEBUG(errs() << " Continuing at " << MatcherIndex << "\n"); 2730 2731 InputChain = LastScope.InputChain; 2732 InputFlag = LastScope.InputFlag; 2733 if (!LastScope.HasChainNodesMatched) 2734 ChainNodesMatched.clear(); 2735 if (!LastScope.HasFlagResultNodesMatched) 2736 FlagResultNodesMatched.clear(); 2737 2738 // Check to see what the offset is at the new MatcherIndex. If it is zero 2739 // we have reached the end of this scope, otherwise we have another child 2740 // in the current scope to try. 2741 unsigned NumToSkip = MatcherTable[MatcherIndex++]; 2742 if (NumToSkip & 128) 2743 NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex); 2744 2745 // If we have another child in this scope to match, update FailIndex and 2746 // try it. 2747 if (NumToSkip != 0) { 2748 LastScope.FailIndex = MatcherIndex+NumToSkip; 2749 break; 2750 } 2751 2752 // End of this scope, pop it and try the next child in the containing 2753 // scope. 2754 MatchScopes.pop_back(); 2755 } 2756 } 2757} 2758 2759 2760 2761void SelectionDAGISel::CannotYetSelect(SDNode *N) { 2762 std::string msg; 2763 raw_string_ostream Msg(msg); 2764 Msg << "Cannot yet select: "; 2765 2766 if (N->getOpcode() != ISD::INTRINSIC_W_CHAIN && 2767 N->getOpcode() != ISD::INTRINSIC_WO_CHAIN && 2768 N->getOpcode() != ISD::INTRINSIC_VOID) { 2769 N->printrFull(Msg, CurDAG); 2770 } else { 2771 bool HasInputChain = N->getOperand(0).getValueType() == MVT::Other; 2772 unsigned iid = 2773 cast<ConstantSDNode>(N->getOperand(HasInputChain))->getZExtValue(); 2774 if (iid < Intrinsic::num_intrinsics) 2775 Msg << "intrinsic %" << Intrinsic::getName((Intrinsic::ID)iid); 2776 else if (const TargetIntrinsicInfo *TII = TM.getIntrinsicInfo()) 2777 Msg << "target intrinsic %" << TII->getName(iid); 2778 else 2779 Msg << "unknown intrinsic #" << iid; 2780 } 2781 llvm_report_error(Msg.str()); 2782} 2783 2784char SelectionDAGISel::ID = 0; 2785