1 //===-- ARMLowOverheadLoops.cpp - CodeGen Low-overhead Loops ---*- C++ -*-===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 /// \file 9 /// Finalize v8.1-m low-overhead loops by converting the associated pseudo 10 /// instructions into machine operations. 11 /// The expectation is that the loop contains three pseudo instructions: 12 /// - t2*LoopStart - placed in the preheader or pre-preheader. The do-loop 13 /// form should be in the preheader, whereas the while form should be in the 14 /// preheaders only predecessor. 15 /// - t2LoopDec - placed within in the loop body. 16 /// - t2LoopEnd - the loop latch terminator. 17 /// 18 /// In addition to this, we also look for the presence of the VCTP instruction, 19 /// which determines whether we can generated the tail-predicated low-overhead 20 /// loop form. 21 /// 22 /// Assumptions and Dependencies: 23 /// Low-overhead loops are constructed and executed using a setup instruction: 24 /// DLS, WLS, DLSTP or WLSTP and an instruction that loops back: LE or LETP. 25 /// WLS(TP) and LE(TP) are branching instructions with a (large) limited range 26 /// but fixed polarity: WLS can only branch forwards and LE can only branch 27 /// backwards. These restrictions mean that this pass is dependent upon block 28 /// layout and block sizes, which is why it's the last pass to run. The same is 29 /// true for ConstantIslands, but this pass does not increase the size of the 30 /// basic blocks, nor does it change the CFG. Instructions are mainly removed 31 /// during the transform and pseudo instructions are replaced by real ones. In 32 /// some cases, when we have to revert to a 'normal' loop, we have to introduce 33 /// multiple instructions for a single pseudo (see RevertWhile and 34 /// RevertLoopEnd). To handle this situation, t2WhileLoopStartLR and t2LoopEnd 35 /// are defined to be as large as this maximum sequence of replacement 36 /// instructions. 37 /// 38 /// A note on VPR.P0 (the lane mask): 39 /// VPT, VCMP, VPNOT and VCTP won't overwrite VPR.P0 when they update it in a 40 /// "VPT Active" context (which includes low-overhead loops and vpt blocks). 41 /// They will simply "and" the result of their calculation with the current 42 /// value of VPR.P0. You can think of it like this: 43 /// \verbatim 44 /// if VPT active: ; Between a DLSTP/LETP, or for predicated instrs 45 /// VPR.P0 &= Value 46 /// else 47 /// VPR.P0 = Value 48 /// \endverbatim 49 /// When we're inside the low-overhead loop (between DLSTP and LETP), we always 50 /// fall in the "VPT active" case, so we can consider that all VPR writes by 51 /// one of those instruction is actually a "and". 52 //===----------------------------------------------------------------------===// 53 54 #include "ARM.h" 55 #include "ARMBaseInstrInfo.h" 56 #include "ARMBasicBlockInfo.h" 57 #include "ARMSubtarget.h" 58 #include "MVETailPredUtils.h" 59 #include "Thumb2InstrInfo.h" 60 #include "llvm/ADT/SetVector.h" 61 #include "llvm/CodeGen/LivePhysRegs.h" 62 #include "llvm/CodeGen/MachineFrameInfo.h" 63 #include "llvm/CodeGen/MachineFunctionPass.h" 64 #include "llvm/CodeGen/MachineLoopInfo.h" 65 #include "llvm/CodeGen/MachineLoopUtils.h" 66 #include "llvm/CodeGen/MachineRegisterInfo.h" 67 #include "llvm/CodeGen/Passes.h" 68 #include "llvm/CodeGen/ReachingDefAnalysis.h" 69 #include "llvm/MC/MCInstrDesc.h" 70 71 using namespace llvm; 72 73 #define DEBUG_TYPE "arm-low-overhead-loops" 74 #define ARM_LOW_OVERHEAD_LOOPS_NAME "ARM Low Overhead Loops pass" 75 76 static cl::opt<bool> 77 DisableTailPredication("arm-loloops-disable-tailpred", cl::Hidden, 78 cl::desc("Disable tail-predication in the ARM LowOverheadLoop pass"), 79 cl::init(false)); 80 81 static cl::opt<bool> 82 DisableOmitDLS("arm-disable-omit-dls", cl::Hidden, 83 cl::desc("Disable omitting 'dls lr, lr' instructions"), 84 cl::init(false)); 85 86 static bool isVectorPredicated(MachineInstr *MI) { 87 int PIdx = llvm::findFirstVPTPredOperandIdx(*MI); 88 return PIdx != -1 && MI->getOperand(PIdx + 1).getReg() == ARM::VPR; 89 } 90 91 static bool isVectorPredicate(MachineInstr *MI) { 92 return MI->findRegisterDefOperandIdx(ARM::VPR, /*TRI=*/nullptr) != -1; 93 } 94 95 static bool hasVPRUse(MachineInstr &MI) { 96 return MI.findRegisterUseOperandIdx(ARM::VPR, /*TRI=*/nullptr) != -1; 97 } 98 99 static bool isDomainMVE(MachineInstr *MI) { 100 uint64_t Domain = MI->getDesc().TSFlags & ARMII::DomainMask; 101 return Domain == ARMII::DomainMVE; 102 } 103 104 static int getVecSize(const MachineInstr &MI) { 105 const MCInstrDesc &MCID = MI.getDesc(); 106 uint64_t Flags = MCID.TSFlags; 107 return (Flags & ARMII::VecSize) >> ARMII::VecSizeShift; 108 } 109 110 static bool shouldInspect(MachineInstr &MI) { 111 if (MI.isDebugInstr()) 112 return false; 113 return isDomainMVE(&MI) || isVectorPredicate(&MI) || hasVPRUse(MI); 114 } 115 116 static bool isHorizontalReduction(const MachineInstr &MI) { 117 const MCInstrDesc &MCID = MI.getDesc(); 118 uint64_t Flags = MCID.TSFlags; 119 return (Flags & ARMII::HorizontalReduction) != 0; 120 } 121 122 namespace { 123 124 using InstSet = SmallPtrSetImpl<MachineInstr *>; 125 126 class PostOrderLoopTraversal { 127 MachineLoop &ML; 128 MachineLoopInfo &MLI; 129 SmallPtrSet<MachineBasicBlock*, 4> Visited; 130 SmallVector<MachineBasicBlock*, 4> Order; 131 132 public: 133 PostOrderLoopTraversal(MachineLoop &ML, MachineLoopInfo &MLI) 134 : ML(ML), MLI(MLI) { } 135 136 const SmallVectorImpl<MachineBasicBlock*> &getOrder() const { 137 return Order; 138 } 139 140 // Visit all the blocks within the loop, as well as exit blocks and any 141 // blocks properly dominating the header. 142 void ProcessLoop() { 143 std::function<void(MachineBasicBlock *)> Search = 144 [this, &Search](MachineBasicBlock *MBB) -> void { 145 if (!Visited.insert(MBB).second) 146 return; 147 148 for (auto *Succ : MBB->successors()) { 149 if (!ML.contains(Succ)) 150 continue; 151 Search(Succ); 152 } 153 Order.push_back(MBB); 154 }; 155 156 // Insert exit blocks. 157 SmallVector<MachineBasicBlock*, 2> ExitBlocks; 158 ML.getExitBlocks(ExitBlocks); 159 append_range(Order, ExitBlocks); 160 161 // Then add the loop body. 162 Search(ML.getHeader()); 163 164 // Then try the preheader and its predecessors. 165 std::function<void(MachineBasicBlock*)> GetPredecessor = 166 [this, &GetPredecessor] (MachineBasicBlock *MBB) -> void { 167 Order.push_back(MBB); 168 if (MBB->pred_size() == 1) 169 GetPredecessor(*MBB->pred_begin()); 170 }; 171 172 if (auto *Preheader = ML.getLoopPreheader()) 173 GetPredecessor(Preheader); 174 else if (auto *Preheader = MLI.findLoopPreheader(&ML, true, true)) 175 GetPredecessor(Preheader); 176 } 177 }; 178 179 class VPTBlock { 180 SmallVector<MachineInstr *, 4> Insts; 181 182 public: 183 VPTBlock(MachineInstr *MI) { Insts.push_back(MI); } 184 185 // Have we found an instruction within the block which defines the vpr? If 186 // so, not all the instructions in the block will have the same predicate. 187 bool hasUniformPredicate() { return getDivergent() == nullptr; } 188 189 // If it exists, return the first internal instruction which modifies the 190 // VPR. 191 MachineInstr *getDivergent() { 192 SmallVectorImpl<MachineInstr *> &Insts = getInsts(); 193 for (unsigned i = 1; i < Insts.size(); ++i) { 194 MachineInstr *Next = Insts[i]; 195 if (isVectorPredicate(Next)) 196 return Next; // Found an instruction altering the vpr. 197 } 198 return nullptr; 199 } 200 201 void insert(MachineInstr *MI) { 202 Insts.push_back(MI); 203 // VPT/VPST + 4 predicated instructions. 204 assert(Insts.size() <= 5 && "Too many instructions in VPT block!"); 205 } 206 207 bool containsVCTP() const { return llvm::any_of(Insts, isVCTP); } 208 209 unsigned size() const { return Insts.size(); } 210 SmallVectorImpl<MachineInstr *> &getInsts() { return Insts; } 211 }; 212 213 // Represent the current state of the VPR and hold all instances which 214 // represent a VPT block, which is a list of instructions that begins with a 215 // VPT/VPST and has a maximum of four proceeding instructions. All 216 // instructions within the block are predicated upon the vpr and we allow 217 // instructions to define the vpr within in the block too. 218 class VPTState { 219 friend struct LowOverheadLoop; 220 221 SmallVector<VPTBlock, 4> Blocks; 222 SetVector<MachineInstr *> CurrentPredicates; 223 std::map<MachineInstr *, SetVector<MachineInstr *>> PredicatedInsts; 224 225 void CreateVPTBlock(MachineInstr *MI) { 226 assert((CurrentPredicates.size() || MI->getParent()->isLiveIn(ARM::VPR)) 227 && "Can't begin VPT without predicate"); 228 Blocks.emplace_back(MI); 229 // The execution of MI is predicated upon the current set of instructions 230 // that are AND'ed together to form the VPR predicate value. In the case 231 // that MI is a VPT, CurrentPredicates will also just be MI. 232 PredicatedInsts[MI] = CurrentPredicates; 233 } 234 235 void addInst(MachineInstr *MI) { 236 Blocks.back().insert(MI); 237 PredicatedInsts[MI] = CurrentPredicates; 238 } 239 240 void addPredicate(MachineInstr *MI) { 241 LLVM_DEBUG(dbgs() << "ARM Loops: Adding VPT Predicate: " << *MI); 242 CurrentPredicates.insert(MI); 243 } 244 245 void resetPredicate(MachineInstr *MI) { 246 LLVM_DEBUG(dbgs() << "ARM Loops: Resetting VPT Predicate: " << *MI); 247 CurrentPredicates.clear(); 248 CurrentPredicates.insert(MI); 249 } 250 251 public: 252 // Return whether the given instruction is predicated upon a VCTP. 253 bool isPredicatedOnVCTP(MachineInstr *MI, bool Exclusive = false) { 254 SetVector<MachineInstr *> &Predicates = PredicatedInsts[MI]; 255 if (Exclusive && Predicates.size() != 1) 256 return false; 257 // We do not know how to convert an else predicate of a VCTP. 258 if (getVPTInstrPredicate(*MI) == ARMVCC::Else) 259 return false; 260 return llvm::any_of(Predicates, isVCTP); 261 } 262 263 // Is the VPST, controlling the block entry, predicated upon a VCTP. 264 bool isEntryPredicatedOnVCTP(VPTBlock &Block, bool Exclusive = false) { 265 SmallVectorImpl<MachineInstr *> &Insts = Block.getInsts(); 266 return isPredicatedOnVCTP(Insts.front(), Exclusive); 267 } 268 269 // If this block begins with a VPT, we can check whether it's using 270 // at least one predicated input(s), as well as possible loop invariant 271 // which would result in it being implicitly predicated. 272 bool hasImplicitlyValidVPT(VPTBlock &Block, ReachingDefAnalysis &RDA) { 273 SmallVectorImpl<MachineInstr *> &Insts = Block.getInsts(); 274 MachineInstr *VPT = Insts.front(); 275 assert(isVPTOpcode(VPT->getOpcode()) && 276 "Expected VPT block to begin with VPT/VPST"); 277 278 if (VPT->getOpcode() == ARM::MVE_VPST) 279 return false; 280 281 // If the VPT block does not define something that is an "output", then 282 // the tail-predicated version will just perform a subset of the original 283 // vpt block, where the last lanes should not be used. 284 if (isVPTOpcode(VPT->getOpcode()) && 285 all_of(Block.getInsts(), [](const MachineInstr *MI) { 286 return !MI->mayStore() && !MI->mayLoad() && 287 !isHorizontalReduction(*MI) && !isVCTP(MI); 288 })) 289 return true; 290 291 auto IsOperandPredicated = [&](MachineInstr *MI, unsigned Idx) { 292 MachineInstr *Op = RDA.getMIOperand(MI, MI->getOperand(Idx)); 293 return Op && PredicatedInsts.count(Op) && isPredicatedOnVCTP(Op); 294 }; 295 296 auto IsOperandInvariant = [&](MachineInstr *MI, unsigned Idx) { 297 MachineOperand &MO = MI->getOperand(Idx); 298 if (!MO.isReg() || !MO.getReg()) 299 return true; 300 301 SmallPtrSet<MachineInstr *, 2> Defs; 302 RDA.getGlobalReachingDefs(MI, MO.getReg(), Defs); 303 if (Defs.empty()) 304 return true; 305 306 for (auto *Def : Defs) 307 if (Def->getParent() == VPT->getParent()) 308 return false; 309 return true; 310 }; 311 312 // Check that at least one of the operands is directly predicated on a 313 // vctp and allow an invariant value too. 314 return (IsOperandPredicated(VPT, 1) || IsOperandPredicated(VPT, 2)) && 315 (IsOperandPredicated(VPT, 1) || IsOperandInvariant(VPT, 1)) && 316 (IsOperandPredicated(VPT, 2) || IsOperandInvariant(VPT, 2)); 317 } 318 319 bool isValid(ReachingDefAnalysis &RDA) { 320 // All predication within the loop should be based on vctp. If the block 321 // isn't predicated on entry, check whether the vctp is within the block 322 // and that all other instructions are then predicated on it. 323 for (auto &Block : Blocks) { 324 if (isEntryPredicatedOnVCTP(Block, false) && 325 !any_of(drop_begin(Block.getInsts()), [](const MachineInstr *MI) { 326 return getVPTInstrPredicate(*MI) == ARMVCC::Else; 327 })) 328 continue; 329 if (hasImplicitlyValidVPT(Block, RDA)) 330 continue; 331 332 SmallVectorImpl<MachineInstr *> &Insts = Block.getInsts(); 333 // We don't know how to convert a block with just a VPT;VCTP into 334 // anything valid once we remove the VCTP. For now just bail out. 335 assert(isVPTOpcode(Insts.front()->getOpcode()) && 336 "Expected VPT block to start with a VPST or VPT!"); 337 if (Insts.size() == 2 && Insts.front()->getOpcode() != ARM::MVE_VPST && 338 isVCTP(Insts.back())) 339 return false; 340 341 for (auto *MI : Insts) { 342 // Check that any internal VCTPs are 'Then' predicated. 343 if (isVCTP(MI) && getVPTInstrPredicate(*MI) != ARMVCC::Then) 344 return false; 345 // Skip other instructions that build up the predicate. 346 if (MI->getOpcode() == ARM::MVE_VPST || isVectorPredicate(MI)) 347 continue; 348 // Check that any other instructions are predicated upon a vctp. 349 // TODO: We could infer when VPTs are implicitly predicated on the 350 // vctp (when the operands are predicated). 351 if (!isPredicatedOnVCTP(MI)) { 352 LLVM_DEBUG(dbgs() << "ARM Loops: Can't convert: " << *MI); 353 return false; 354 } 355 } 356 } 357 return true; 358 } 359 }; 360 361 struct LowOverheadLoop { 362 363 MachineLoop &ML; 364 MachineBasicBlock *Preheader = nullptr; 365 MachineLoopInfo &MLI; 366 ReachingDefAnalysis &RDA; 367 const TargetRegisterInfo &TRI; 368 const ARMBaseInstrInfo &TII; 369 MachineFunction *MF = nullptr; 370 MachineBasicBlock::iterator StartInsertPt; 371 MachineBasicBlock *StartInsertBB = nullptr; 372 MachineInstr *Start = nullptr; 373 MachineInstr *Dec = nullptr; 374 MachineInstr *End = nullptr; 375 MachineOperand TPNumElements; 376 SmallVector<MachineInstr *, 4> VCTPs; 377 SmallPtrSet<MachineInstr *, 4> ToRemove; 378 SmallPtrSet<MachineInstr *, 4> BlockMasksToRecompute; 379 SmallPtrSet<MachineInstr *, 4> DoubleWidthResultInstrs; 380 SmallPtrSet<MachineInstr *, 4> VMOVCopies; 381 bool Revert = false; 382 bool CannotTailPredicate = false; 383 VPTState VPTstate; 384 385 LowOverheadLoop(MachineLoop &ML, MachineLoopInfo &MLI, 386 ReachingDefAnalysis &RDA, const TargetRegisterInfo &TRI, 387 const ARMBaseInstrInfo &TII) 388 : ML(ML), MLI(MLI), RDA(RDA), TRI(TRI), TII(TII), 389 TPNumElements(MachineOperand::CreateImm(0)) { 390 MF = ML.getHeader()->getParent(); 391 if (auto *MBB = ML.getLoopPreheader()) 392 Preheader = MBB; 393 else if (auto *MBB = MLI.findLoopPreheader(&ML, true, true)) 394 Preheader = MBB; 395 } 396 397 // If this is an MVE instruction, check that we know how to use tail 398 // predication with it. Record VPT blocks and return whether the 399 // instruction is valid for tail predication. 400 bool ValidateMVEInst(MachineInstr *MI); 401 402 void AnalyseMVEInst(MachineInstr *MI) { 403 CannotTailPredicate = !ValidateMVEInst(MI); 404 } 405 406 bool IsTailPredicationLegal() const { 407 // For now, let's keep things really simple and only support a single 408 // block for tail predication. 409 return !Revert && FoundAllComponents() && !VCTPs.empty() && 410 !CannotTailPredicate && ML.getNumBlocks() == 1; 411 } 412 413 // Given that MI is a VCTP, check that is equivalent to any other VCTPs 414 // found. 415 bool AddVCTP(MachineInstr *MI); 416 417 // Check that the predication in the loop will be equivalent once we 418 // perform the conversion. Also ensure that we can provide the number 419 // of elements to the loop start instruction. 420 bool ValidateTailPredicate(); 421 422 // Check that any values available outside of the loop will be the same 423 // after tail predication conversion. 424 bool ValidateLiveOuts(); 425 426 // Check the branch targets are within range and we satisfy our 427 // restrictions. 428 void Validate(ARMBasicBlockUtils *BBUtils); 429 430 bool FoundAllComponents() const { 431 return Start && Dec && End; 432 } 433 434 SmallVectorImpl<VPTBlock> &getVPTBlocks() { return VPTstate.Blocks; } 435 436 // Return the operand for the loop start instruction. This will be the loop 437 // iteration count, or the number of elements if we're tail predicating. 438 MachineOperand &getLoopStartOperand() { 439 if (IsTailPredicationLegal()) 440 return TPNumElements; 441 return Start->getOperand(1); 442 } 443 444 unsigned getStartOpcode() const { 445 bool IsDo = isDoLoopStart(*Start); 446 if (!IsTailPredicationLegal()) 447 return IsDo ? ARM::t2DLS : ARM::t2WLS; 448 449 return VCTPOpcodeToLSTP(VCTPs.back()->getOpcode(), IsDo); 450 } 451 452 void dump() const { 453 if (Start) dbgs() << "ARM Loops: Found Loop Start: " << *Start; 454 if (Dec) dbgs() << "ARM Loops: Found Loop Dec: " << *Dec; 455 if (End) dbgs() << "ARM Loops: Found Loop End: " << *End; 456 if (!VCTPs.empty()) { 457 dbgs() << "ARM Loops: Found VCTP(s):\n"; 458 for (auto *MI : VCTPs) 459 dbgs() << " - " << *MI; 460 } 461 if (!FoundAllComponents()) 462 dbgs() << "ARM Loops: Not a low-overhead loop.\n"; 463 else if (!(Start && Dec && End)) 464 dbgs() << "ARM Loops: Failed to find all loop components.\n"; 465 } 466 }; 467 468 class ARMLowOverheadLoops : public MachineFunctionPass { 469 MachineFunction *MF = nullptr; 470 MachineLoopInfo *MLI = nullptr; 471 ReachingDefAnalysis *RDA = nullptr; 472 const ARMBaseInstrInfo *TII = nullptr; 473 MachineRegisterInfo *MRI = nullptr; 474 const TargetRegisterInfo *TRI = nullptr; 475 std::unique_ptr<ARMBasicBlockUtils> BBUtils = nullptr; 476 477 public: 478 static char ID; 479 480 ARMLowOverheadLoops() : MachineFunctionPass(ID) { } 481 482 void getAnalysisUsage(AnalysisUsage &AU) const override { 483 AU.setPreservesCFG(); 484 AU.addRequired<MachineLoopInfoWrapperPass>(); 485 AU.addRequired<ReachingDefAnalysis>(); 486 MachineFunctionPass::getAnalysisUsage(AU); 487 } 488 489 bool runOnMachineFunction(MachineFunction &MF) override; 490 491 MachineFunctionProperties getRequiredProperties() const override { 492 return MachineFunctionProperties().setNoVRegs().setTracksLiveness(); 493 } 494 495 StringRef getPassName() const override { 496 return ARM_LOW_OVERHEAD_LOOPS_NAME; 497 } 498 499 private: 500 bool ProcessLoop(MachineLoop *ML); 501 502 bool RevertNonLoops(); 503 504 void RevertWhile(MachineInstr *MI) const; 505 void RevertDo(MachineInstr *MI) const; 506 507 bool RevertLoopDec(MachineInstr *MI) const; 508 509 void RevertLoopEnd(MachineInstr *MI, bool SkipCmp = false) const; 510 511 void RevertLoopEndDec(MachineInstr *MI) const; 512 513 void ConvertVPTBlocks(LowOverheadLoop &LoLoop); 514 515 MachineInstr *ExpandLoopStart(LowOverheadLoop &LoLoop); 516 517 void Expand(LowOverheadLoop &LoLoop); 518 519 void IterationCountDCE(LowOverheadLoop &LoLoop); 520 }; 521 } 522 523 char ARMLowOverheadLoops::ID = 0; 524 525 INITIALIZE_PASS(ARMLowOverheadLoops, DEBUG_TYPE, ARM_LOW_OVERHEAD_LOOPS_NAME, 526 false, false) 527 528 static bool TryRemove(MachineInstr *MI, ReachingDefAnalysis &RDA, 529 InstSet &ToRemove, InstSet &Ignore) { 530 531 // Check that we can remove all of Killed without having to modify any IT 532 // blocks. 533 auto WontCorruptITs = [](InstSet &Killed, ReachingDefAnalysis &RDA) { 534 // Collect the dead code and the MBBs in which they reside. 535 SmallPtrSet<MachineBasicBlock*, 2> BasicBlocks; 536 for (auto *Dead : Killed) 537 BasicBlocks.insert(Dead->getParent()); 538 539 // Collect IT blocks in all affected basic blocks. 540 std::map<MachineInstr *, SmallPtrSet<MachineInstr *, 2>> ITBlocks; 541 for (auto *MBB : BasicBlocks) { 542 for (auto &IT : *MBB) { 543 if (IT.getOpcode() != ARM::t2IT) 544 continue; 545 RDA.getReachingLocalUses(&IT, MCRegister::from(ARM::ITSTATE), 546 ITBlocks[&IT]); 547 } 548 } 549 550 // If we're removing all of the instructions within an IT block, then 551 // also remove the IT instruction. 552 SmallPtrSet<MachineInstr *, 2> ModifiedITs; 553 SmallPtrSet<MachineInstr *, 2> RemoveITs; 554 for (auto *Dead : Killed) { 555 if (MachineOperand *MO = 556 Dead->findRegisterUseOperand(ARM::ITSTATE, /*TRI=*/nullptr)) { 557 MachineInstr *IT = RDA.getMIOperand(Dead, *MO); 558 RemoveITs.insert(IT); 559 auto &CurrentBlock = ITBlocks[IT]; 560 CurrentBlock.erase(Dead); 561 if (CurrentBlock.empty()) 562 ModifiedITs.erase(IT); 563 else 564 ModifiedITs.insert(IT); 565 } 566 } 567 if (!ModifiedITs.empty()) 568 return false; 569 Killed.insert_range(RemoveITs); 570 return true; 571 }; 572 573 SmallPtrSet<MachineInstr *, 2> Uses; 574 if (!RDA.isSafeToRemove(MI, Uses, Ignore)) 575 return false; 576 577 if (WontCorruptITs(Uses, RDA)) { 578 ToRemove.insert_range(Uses); 579 LLVM_DEBUG(dbgs() << "ARM Loops: Able to remove: " << *MI 580 << " - can also remove:\n"; 581 for (auto *Use : Uses) 582 dbgs() << " - " << *Use); 583 584 SmallPtrSet<MachineInstr*, 4> Killed; 585 RDA.collectKilledOperands(MI, Killed); 586 if (WontCorruptITs(Killed, RDA)) { 587 ToRemove.insert_range(Killed); 588 LLVM_DEBUG(for (auto *Dead : Killed) 589 dbgs() << " - " << *Dead); 590 } 591 return true; 592 } 593 return false; 594 } 595 596 bool LowOverheadLoop::ValidateTailPredicate() { 597 if (!IsTailPredicationLegal()) { 598 LLVM_DEBUG(if (VCTPs.empty()) 599 dbgs() << "ARM Loops: Didn't find a VCTP instruction.\n"; 600 dbgs() << "ARM Loops: Tail-predication is not valid.\n"); 601 return false; 602 } 603 604 assert(!VCTPs.empty() && "VCTP instruction expected but is not set"); 605 assert(ML.getBlocks().size() == 1 && 606 "Shouldn't be processing a loop with more than one block"); 607 608 if (DisableTailPredication) { 609 LLVM_DEBUG(dbgs() << "ARM Loops: tail-predication is disabled\n"); 610 return false; 611 } 612 613 if (!VPTstate.isValid(RDA)) { 614 LLVM_DEBUG(dbgs() << "ARM Loops: Invalid VPT state.\n"); 615 return false; 616 } 617 618 if (!ValidateLiveOuts()) { 619 LLVM_DEBUG(dbgs() << "ARM Loops: Invalid live outs.\n"); 620 return false; 621 } 622 623 // For tail predication, we need to provide the number of elements, instead 624 // of the iteration count, to the loop start instruction. The number of 625 // elements is provided to the vctp instruction, so we need to check that 626 // we can use this register at InsertPt. 627 MachineInstr *VCTP = VCTPs.back(); 628 if (Start->getOpcode() == ARM::t2DoLoopStartTP || 629 Start->getOpcode() == ARM::t2WhileLoopStartTP) { 630 TPNumElements = Start->getOperand(2); 631 StartInsertPt = Start; 632 StartInsertBB = Start->getParent(); 633 } else { 634 TPNumElements = VCTP->getOperand(1); 635 MCRegister NumElements = TPNumElements.getReg().asMCReg(); 636 637 // If the register is defined within loop, then we can't perform TP. 638 // TODO: Check whether this is just a mov of a register that would be 639 // available. 640 if (RDA.hasLocalDefBefore(VCTP, NumElements)) { 641 LLVM_DEBUG(dbgs() << "ARM Loops: VCTP operand is defined in the loop.\n"); 642 return false; 643 } 644 645 // The element count register maybe defined after InsertPt, in which case we 646 // need to try to move either InsertPt or the def so that the [w|d]lstp can 647 // use the value. 648 649 if (StartInsertPt != StartInsertBB->end() && 650 !RDA.isReachingDefLiveOut(&*StartInsertPt, NumElements)) { 651 if (auto *ElemDef = 652 RDA.getLocalLiveOutMIDef(StartInsertBB, NumElements)) { 653 if (RDA.isSafeToMoveForwards(ElemDef, &*StartInsertPt)) { 654 ElemDef->removeFromParent(); 655 StartInsertBB->insert(StartInsertPt, ElemDef); 656 LLVM_DEBUG(dbgs() 657 << "ARM Loops: Moved element count def: " << *ElemDef); 658 } else if (RDA.isSafeToMoveBackwards(&*StartInsertPt, ElemDef)) { 659 StartInsertPt->removeFromParent(); 660 StartInsertBB->insertAfter(MachineBasicBlock::iterator(ElemDef), 661 &*StartInsertPt); 662 LLVM_DEBUG(dbgs() << "ARM Loops: Moved start past: " << *ElemDef); 663 } else { 664 // If we fail to move an instruction and the element count is provided 665 // by a mov, use the mov operand if it will have the same value at the 666 // insertion point 667 MachineOperand Operand = ElemDef->getOperand(1); 668 if (isMovRegOpcode(ElemDef->getOpcode()) && 669 RDA.getUniqueReachingMIDef(ElemDef, Operand.getReg().asMCReg()) == 670 RDA.getUniqueReachingMIDef(&*StartInsertPt, 671 Operand.getReg().asMCReg())) { 672 TPNumElements = Operand; 673 NumElements = TPNumElements.getReg(); 674 } else { 675 LLVM_DEBUG(dbgs() 676 << "ARM Loops: Unable to move element count to loop " 677 << "start instruction.\n"); 678 return false; 679 } 680 } 681 } 682 } 683 684 // Especially in the case of while loops, InsertBB may not be the 685 // preheader, so we need to check that the register isn't redefined 686 // before entering the loop. 687 auto CannotProvideElements = [this](MachineBasicBlock *MBB, 688 MCRegister NumElements) { 689 if (MBB->empty()) 690 return false; 691 // NumElements is redefined in this block. 692 if (RDA.hasLocalDefBefore(&MBB->back(), NumElements)) 693 return true; 694 695 // Don't continue searching up through multiple predecessors. 696 if (MBB->pred_size() > 1) 697 return true; 698 699 return false; 700 }; 701 702 // Search backwards for a def, until we get to InsertBB. 703 MachineBasicBlock *MBB = Preheader; 704 while (MBB && MBB != StartInsertBB) { 705 if (CannotProvideElements(MBB, NumElements)) { 706 LLVM_DEBUG(dbgs() << "ARM Loops: Unable to provide element count.\n"); 707 return false; 708 } 709 MBB = *MBB->pred_begin(); 710 } 711 } 712 713 // Could inserting the [W|D]LSTP cause some unintended affects? In a perfect 714 // world the [w|d]lstp instruction would be last instruction in the preheader 715 // and so it would only affect instructions within the loop body. But due to 716 // scheduling, and/or the logic in this pass (above), the insertion point can 717 // be moved earlier. So if the Loop Start isn't the last instruction in the 718 // preheader, and if the initial element count is smaller than the vector 719 // width, the Loop Start instruction will immediately generate one or more 720 // false lane mask which can, incorrectly, affect the proceeding MVE 721 // instructions in the preheader. 722 if (std::any_of(StartInsertPt, StartInsertBB->end(), shouldInspect)) { 723 LLVM_DEBUG(dbgs() << "ARM Loops: Instruction blocks [W|D]LSTP\n"); 724 return false; 725 } 726 727 // For any DoubleWidthResultInstrs we found whilst scanning instructions, they 728 // need to compute an output size that is smaller than the VCTP mask operates 729 // on. The VecSize of the DoubleWidthResult is the larger vector size - the 730 // size it extends into, so any VCTP VecSize <= is valid. 731 unsigned VCTPVecSize = getVecSize(*VCTP); 732 for (MachineInstr *MI : DoubleWidthResultInstrs) { 733 unsigned InstrVecSize = getVecSize(*MI); 734 if (InstrVecSize > VCTPVecSize) { 735 LLVM_DEBUG(dbgs() << "ARM Loops: Double width result larger than VCTP " 736 << "VecSize:\n" << *MI); 737 return false; 738 } 739 } 740 741 // Check that the value change of the element count is what we expect and 742 // that the predication will be equivalent. For this we need: 743 // NumElements = NumElements - VectorWidth. The sub will be a sub immediate 744 // and we can also allow register copies within the chain too. 745 auto IsValidSub = [](MachineInstr *MI, int ExpectedVecWidth) { 746 return -getAddSubImmediate(*MI) == ExpectedVecWidth; 747 }; 748 749 MachineBasicBlock *MBB = VCTP->getParent(); 750 // Remove modifications to the element count since they have no purpose in a 751 // tail predicated loop. Explicitly refer to the vctp operand no matter which 752 // register NumElements has been assigned to, since that is what the 753 // modifications will be using 754 if (auto *Def = RDA.getUniqueReachingMIDef( 755 &MBB->back(), VCTP->getOperand(1).getReg().asMCReg())) { 756 SmallPtrSet<MachineInstr*, 2> ElementChain; 757 SmallPtrSet<MachineInstr*, 2> Ignore; 758 unsigned ExpectedVectorWidth = getTailPredVectorWidth(VCTP->getOpcode()); 759 760 Ignore.insert_range(VCTPs); 761 762 if (TryRemove(Def, RDA, ElementChain, Ignore)) { 763 bool FoundSub = false; 764 765 for (auto *MI : ElementChain) { 766 if (isMovRegOpcode(MI->getOpcode())) 767 continue; 768 769 if (isSubImmOpcode(MI->getOpcode())) { 770 if (FoundSub || !IsValidSub(MI, ExpectedVectorWidth)) { 771 LLVM_DEBUG(dbgs() << "ARM Loops: Unexpected instruction in element" 772 " count: " << *MI); 773 return false; 774 } 775 FoundSub = true; 776 } else { 777 LLVM_DEBUG(dbgs() << "ARM Loops: Unexpected instruction in element" 778 " count: " << *MI); 779 return false; 780 } 781 } 782 ToRemove.insert_range(ElementChain); 783 } 784 } 785 786 // If we converted the LoopStart to a t2DoLoopStartTP/t2WhileLoopStartTP, we 787 // can also remove any extra instructions in the preheader, which often 788 // includes a now unused MOV. 789 if ((Start->getOpcode() == ARM::t2DoLoopStartTP || 790 Start->getOpcode() == ARM::t2WhileLoopStartTP) && 791 Preheader && !Preheader->empty() && 792 !RDA.hasLocalDefBefore(VCTP, VCTP->getOperand(1).getReg())) { 793 if (auto *Def = RDA.getUniqueReachingMIDef( 794 &Preheader->back(), VCTP->getOperand(1).getReg().asMCReg())) { 795 SmallPtrSet<MachineInstr *, 2> Ignore(llvm::from_range, VCTPs); 796 TryRemove(Def, RDA, ToRemove, Ignore); 797 } 798 } 799 800 return true; 801 } 802 803 static bool isRegInClass(const MachineOperand &MO, 804 const TargetRegisterClass *Class) { 805 return MO.isReg() && MO.getReg() && Class->contains(MO.getReg()); 806 } 807 808 // MVE 'narrowing' operate on half a lane, reading from half and writing 809 // to half, which are referred to has the top and bottom half. The other 810 // half retains its previous value. 811 static bool retainsPreviousHalfElement(const MachineInstr &MI) { 812 const MCInstrDesc &MCID = MI.getDesc(); 813 uint64_t Flags = MCID.TSFlags; 814 return (Flags & ARMII::RetainsPreviousHalfElement) != 0; 815 } 816 817 // Some MVE instructions read from the top/bottom halves of their operand(s) 818 // and generate a vector result with result elements that are double the 819 // width of the input. 820 static bool producesDoubleWidthResult(const MachineInstr &MI) { 821 const MCInstrDesc &MCID = MI.getDesc(); 822 uint64_t Flags = MCID.TSFlags; 823 return (Flags & ARMII::DoubleWidthResult) != 0; 824 } 825 826 // Can this instruction generate a non-zero result when given only zeroed 827 // operands? This allows us to know that, given operands with false bytes 828 // zeroed by masked loads, that the result will also contain zeros in those 829 // bytes. 830 static bool canGenerateNonZeros(const MachineInstr &MI) { 831 832 // Check for instructions which can write into a larger element size, 833 // possibly writing into a previous zero'd lane. 834 if (producesDoubleWidthResult(MI)) 835 return true; 836 837 switch (MI.getOpcode()) { 838 default: 839 break; 840 // FIXME: VNEG FP and -0? I think we'll need to handle this once we allow 841 // fp16 -> fp32 vector conversions. 842 // Instructions that perform a NOT will generate 1s from 0s. 843 case ARM::MVE_VMVN: 844 case ARM::MVE_VORN: 845 // Count leading zeros will do just that! 846 case ARM::MVE_VCLZs8: 847 case ARM::MVE_VCLZs16: 848 case ARM::MVE_VCLZs32: 849 return true; 850 } 851 return false; 852 } 853 854 // Look at its register uses to see if it only can only receive zeros 855 // into its false lanes which would then produce zeros. Also check that 856 // the output register is also defined by an FalseLanesZero instruction 857 // so that if tail-predication happens, the lanes that aren't updated will 858 // still be zeros. 859 static bool producesFalseLanesZero(MachineInstr &MI, 860 const TargetRegisterClass *QPRs, 861 const ReachingDefAnalysis &RDA, 862 InstSet &FalseLanesZero) { 863 if (canGenerateNonZeros(MI)) 864 return false; 865 866 bool isPredicated = isVectorPredicated(&MI); 867 // Predicated loads will write zeros to the falsely predicated bytes of the 868 // destination register. 869 if (MI.mayLoad()) 870 return isPredicated; 871 872 auto IsZeroInit = [](MachineInstr *Def) { 873 return !isVectorPredicated(Def) && 874 Def->getOpcode() == ARM::MVE_VMOVimmi32 && 875 Def->getOperand(1).getImm() == 0; 876 }; 877 878 bool AllowScalars = isHorizontalReduction(MI); 879 for (auto &MO : MI.operands()) { 880 if (!MO.isReg() || !MO.getReg()) 881 continue; 882 if (!isRegInClass(MO, QPRs) && AllowScalars) 883 continue; 884 // Skip the lr predicate reg 885 int PIdx = llvm::findFirstVPTPredOperandIdx(MI); 886 if (PIdx != -1 && (int)MO.getOperandNo() == PIdx + 2) 887 continue; 888 889 // Check that this instruction will produce zeros in its false lanes: 890 // - If it only consumes false lanes zero or constant 0 (vmov #0) 891 // - If it's predicated, it only matters that it's def register already has 892 // false lane zeros, so we can ignore the uses. 893 SmallPtrSet<MachineInstr *, 2> Defs; 894 RDA.getGlobalReachingDefs(&MI, MO.getReg(), Defs); 895 if (Defs.empty()) 896 return false; 897 for (auto *Def : Defs) { 898 if (Def == &MI || FalseLanesZero.count(Def) || IsZeroInit(Def)) 899 continue; 900 if (MO.isUse() && isPredicated) 901 continue; 902 return false; 903 } 904 } 905 LLVM_DEBUG(dbgs() << "ARM Loops: Always False Zeros: " << MI); 906 return true; 907 } 908 909 bool LowOverheadLoop::ValidateLiveOuts() { 910 // We want to find out if the tail-predicated version of this loop will 911 // produce the same values as the loop in its original form. For this to 912 // be true, the newly inserted implicit predication must not change the 913 // the (observable) results. 914 // We're doing this because many instructions in the loop will not be 915 // predicated and so the conversion from VPT predication to tail-predication 916 // can result in different values being produced; due to the tail-predication 917 // preventing many instructions from updating their falsely predicated 918 // lanes. This analysis assumes that all the instructions perform lane-wise 919 // operations and don't perform any exchanges. 920 // A masked load, whether through VPT or tail predication, will write zeros 921 // to any of the falsely predicated bytes. So, from the loads, we know that 922 // the false lanes are zeroed and here we're trying to track that those false 923 // lanes remain zero, or where they change, the differences are masked away 924 // by their user(s). 925 // All MVE stores have to be predicated, so we know that any predicate load 926 // operands, or stored results are equivalent already. Other explicitly 927 // predicated instructions will perform the same operation in the original 928 // loop and the tail-predicated form too. Because of this, we can insert 929 // loads, stores and other predicated instructions into our Predicated 930 // set and build from there. 931 const TargetRegisterClass *QPRs = TRI.getRegClass(ARM::MQPRRegClassID); 932 SetVector<MachineInstr *> FalseLanesUnknown; 933 SmallPtrSet<MachineInstr *, 4> FalseLanesZero; 934 SmallPtrSet<MachineInstr *, 4> Predicated; 935 MachineBasicBlock *Header = ML.getHeader(); 936 937 LLVM_DEBUG(dbgs() << "ARM Loops: Validating Live outs\n"); 938 939 for (auto &MI : *Header) { 940 if (!shouldInspect(MI)) 941 continue; 942 943 if (isVCTP(&MI) || isVPTOpcode(MI.getOpcode())) 944 continue; 945 946 bool isPredicated = isVectorPredicated(&MI); 947 bool retainsOrReduces = 948 retainsPreviousHalfElement(MI) || isHorizontalReduction(MI); 949 950 if (isPredicated) 951 Predicated.insert(&MI); 952 if (producesFalseLanesZero(MI, QPRs, RDA, FalseLanesZero)) 953 FalseLanesZero.insert(&MI); 954 else if (MI.getNumDefs() == 0) 955 continue; 956 else if (!isPredicated && retainsOrReduces) { 957 LLVM_DEBUG(dbgs() << " Unpredicated instruction that retainsOrReduces: " << MI); 958 return false; 959 } else if (!isPredicated && MI.getOpcode() != ARM::MQPRCopy) 960 FalseLanesUnknown.insert(&MI); 961 } 962 963 LLVM_DEBUG({ 964 dbgs() << " Predicated:\n"; 965 for (auto *I : Predicated) 966 dbgs() << " " << *I; 967 dbgs() << " FalseLanesZero:\n"; 968 for (auto *I : FalseLanesZero) 969 dbgs() << " " << *I; 970 dbgs() << " FalseLanesUnknown:\n"; 971 for (auto *I : FalseLanesUnknown) 972 dbgs() << " " << *I; 973 }); 974 975 auto HasPredicatedUsers = [this](MachineInstr *MI, const MachineOperand &MO, 976 SmallPtrSetImpl<MachineInstr *> &Predicated) { 977 SmallPtrSet<MachineInstr *, 2> Uses; 978 RDA.getGlobalUses(MI, MO.getReg().asMCReg(), Uses); 979 for (auto *Use : Uses) { 980 if (Use != MI && !Predicated.count(Use)) 981 return false; 982 } 983 return true; 984 }; 985 986 // Visit the unknowns in reverse so that we can start at the values being 987 // stored and then we can work towards the leaves, hopefully adding more 988 // instructions to Predicated. Successfully terminating the loop means that 989 // all the unknown values have to found to be masked by predicated user(s). 990 // For any unpredicated values, we store them in NonPredicated so that we 991 // can later check whether these form a reduction. 992 SmallPtrSet<MachineInstr*, 2> NonPredicated; 993 for (auto *MI : reverse(FalseLanesUnknown)) { 994 for (auto &MO : MI->operands()) { 995 if (!isRegInClass(MO, QPRs) || !MO.isDef()) 996 continue; 997 if (!HasPredicatedUsers(MI, MO, Predicated)) { 998 LLVM_DEBUG(dbgs() << " Found an unknown def of : " 999 << TRI.getRegAsmName(MO.getReg()) << " at " << *MI); 1000 NonPredicated.insert(MI); 1001 break; 1002 } 1003 } 1004 // Any unknown false lanes have been masked away by the user(s). 1005 if (!NonPredicated.contains(MI)) 1006 Predicated.insert(MI); 1007 } 1008 1009 SmallPtrSet<MachineInstr *, 2> LiveOutMIs; 1010 SmallVector<MachineBasicBlock *, 2> ExitBlocks; 1011 ML.getExitBlocks(ExitBlocks); 1012 assert(ML.getNumBlocks() == 1 && "Expected single block loop!"); 1013 assert(ExitBlocks.size() == 1 && "Expected a single exit block"); 1014 MachineBasicBlock *ExitBB = ExitBlocks.front(); 1015 for (const MachineBasicBlock::RegisterMaskPair &RegMask : ExitBB->liveins()) { 1016 // TODO: Instead of blocking predication, we could move the vctp to the exit 1017 // block and calculate it's operand there in or the preheader. 1018 if (RegMask.PhysReg == ARM::VPR) { 1019 LLVM_DEBUG(dbgs() << " VPR is live in to the exit block."); 1020 return false; 1021 } 1022 // Check Q-regs that are live in the exit blocks. We don't collect scalars 1023 // because they won't be affected by lane predication. 1024 if (QPRs->contains(RegMask.PhysReg)) 1025 if (auto *MI = RDA.getLocalLiveOutMIDef(Header, RegMask.PhysReg)) 1026 LiveOutMIs.insert(MI); 1027 } 1028 1029 // We've already validated that any VPT predication within the loop will be 1030 // equivalent when we perform the predication transformation; so we know that 1031 // any VPT predicated instruction is predicated upon VCTP. Any live-out 1032 // instruction needs to be predicated, so check this here. The instructions 1033 // in NonPredicated have been found to be a reduction that we can ensure its 1034 // legality. Any MQPRCopy found will need to validate its input as if it was 1035 // live out. 1036 SmallVector<MachineInstr *> Worklist(LiveOutMIs.begin(), LiveOutMIs.end()); 1037 while (!Worklist.empty()) { 1038 MachineInstr *MI = Worklist.pop_back_val(); 1039 if (MI->getOpcode() == ARM::MQPRCopy) { 1040 VMOVCopies.insert(MI); 1041 MachineInstr *CopySrc = 1042 RDA.getUniqueReachingMIDef(MI, MI->getOperand(1).getReg()); 1043 if (CopySrc) 1044 Worklist.push_back(CopySrc); 1045 } else if (NonPredicated.count(MI) && FalseLanesUnknown.contains(MI)) { 1046 LLVM_DEBUG(dbgs() << " Unable to handle live out: " << *MI); 1047 VMOVCopies.clear(); 1048 return false; 1049 } 1050 } 1051 1052 return true; 1053 } 1054 1055 void LowOverheadLoop::Validate(ARMBasicBlockUtils *BBUtils) { 1056 if (Revert) 1057 return; 1058 1059 // Check branch target ranges: WLS[TP] can only branch forwards and LE[TP] 1060 // can only jump back. 1061 auto ValidateRanges = [](MachineInstr *Start, MachineInstr *End, 1062 ARMBasicBlockUtils *BBUtils, MachineLoop &ML) { 1063 MachineBasicBlock *TgtBB = End->getOpcode() == ARM::t2LoopEnd 1064 ? End->getOperand(1).getMBB() 1065 : End->getOperand(2).getMBB(); 1066 // TODO Maybe there's cases where the target doesn't have to be the header, 1067 // but for now be safe and revert. 1068 if (TgtBB != ML.getHeader()) { 1069 LLVM_DEBUG(dbgs() << "ARM Loops: LoopEnd is not targeting header.\n"); 1070 return false; 1071 } 1072 1073 // The WLS and LE instructions have 12-bits for the label offset. WLS 1074 // requires a positive offset, while LE uses negative. 1075 if (BBUtils->getOffsetOf(End) < BBUtils->getOffsetOf(ML.getHeader()) || 1076 !BBUtils->isBBInRange(End, ML.getHeader(), 4094)) { 1077 LLVM_DEBUG(dbgs() << "ARM Loops: LE offset is out-of-range\n"); 1078 return false; 1079 } 1080 1081 if (isWhileLoopStart(*Start)) { 1082 MachineBasicBlock *TargetBB = getWhileLoopStartTargetBB(*Start); 1083 if (BBUtils->getOffsetOf(Start) > BBUtils->getOffsetOf(TargetBB) || 1084 !BBUtils->isBBInRange(Start, TargetBB, 4094)) { 1085 LLVM_DEBUG(dbgs() << "ARM Loops: WLS offset is out-of-range!\n"); 1086 return false; 1087 } 1088 } 1089 return true; 1090 }; 1091 1092 StartInsertPt = MachineBasicBlock::iterator(Start); 1093 StartInsertBB = Start->getParent(); 1094 LLVM_DEBUG(dbgs() << "ARM Loops: Will insert LoopStart at " 1095 << *StartInsertPt); 1096 1097 Revert = !ValidateRanges(Start, End, BBUtils, ML); 1098 CannotTailPredicate = !ValidateTailPredicate(); 1099 } 1100 1101 bool LowOverheadLoop::AddVCTP(MachineInstr *MI) { 1102 LLVM_DEBUG(dbgs() << "ARM Loops: Adding VCTP: " << *MI); 1103 if (VCTPs.empty()) { 1104 VCTPs.push_back(MI); 1105 return true; 1106 } 1107 1108 // If we find another VCTP, check whether it uses the same value as the main VCTP. 1109 // If it does, store it in the VCTPs set, else refuse it. 1110 MachineInstr *Prev = VCTPs.back(); 1111 if (!Prev->getOperand(1).isIdenticalTo(MI->getOperand(1)) || 1112 !RDA.hasSameReachingDef(Prev, MI, MI->getOperand(1).getReg().asMCReg())) { 1113 LLVM_DEBUG(dbgs() << "ARM Loops: Found VCTP with a different reaching " 1114 "definition from the main VCTP"); 1115 return false; 1116 } 1117 VCTPs.push_back(MI); 1118 return true; 1119 } 1120 1121 static bool ValidateMVEStore(MachineInstr *MI, MachineLoop *ML) { 1122 1123 auto GetFrameIndex = [](MachineMemOperand *Operand) { 1124 const PseudoSourceValue *PseudoValue = Operand->getPseudoValue(); 1125 if (PseudoValue && PseudoValue->kind() == PseudoSourceValue::FixedStack) { 1126 if (const auto *FS = dyn_cast<FixedStackPseudoSourceValue>(PseudoValue)) { 1127 return FS->getFrameIndex(); 1128 } 1129 } 1130 return -1; 1131 }; 1132 1133 auto IsStackOp = [GetFrameIndex](MachineInstr *I) { 1134 switch (I->getOpcode()) { 1135 case ARM::MVE_VSTRWU32: 1136 case ARM::MVE_VLDRWU32: { 1137 return I->getOperand(1).getReg() == ARM::SP && 1138 I->memoperands().size() == 1 && 1139 GetFrameIndex(I->memoperands().front()) >= 0; 1140 } 1141 default: 1142 return false; 1143 } 1144 }; 1145 1146 // An unpredicated vector register spill is allowed if all of the uses of the 1147 // stack slot are within the loop 1148 if (MI->getOpcode() != ARM::MVE_VSTRWU32 || !IsStackOp(MI)) 1149 return false; 1150 1151 // Search all blocks after the loop for accesses to the same stack slot. 1152 // ReachingDefAnalysis doesn't work for sp as it relies on registers being 1153 // live-out (which sp never is) to know what blocks to look in 1154 if (MI->memoperands().size() == 0) 1155 return false; 1156 int FI = GetFrameIndex(MI->memoperands().front()); 1157 1158 auto &FrameInfo = MI->getParent()->getParent()->getFrameInfo(); 1159 if (FI == -1 || !FrameInfo.isSpillSlotObjectIndex(FI)) 1160 return false; 1161 1162 SmallVector<MachineBasicBlock *> Frontier; 1163 ML->getExitBlocks(Frontier); 1164 SmallPtrSet<MachineBasicBlock *, 4> Visited{MI->getParent()}; 1165 unsigned Idx = 0; 1166 while (Idx < Frontier.size()) { 1167 MachineBasicBlock *BB = Frontier[Idx]; 1168 bool LookAtSuccessors = true; 1169 for (auto &I : *BB) { 1170 if (!IsStackOp(&I) || I.memoperands().size() == 0) 1171 continue; 1172 if (GetFrameIndex(I.memoperands().front()) != FI) 1173 continue; 1174 // If this block has a store to the stack slot before any loads then we 1175 // can ignore the block 1176 if (I.getOpcode() == ARM::MVE_VSTRWU32) { 1177 LookAtSuccessors = false; 1178 break; 1179 } 1180 // If the store and the load are using the same stack slot then the 1181 // store isn't valid for tail predication 1182 if (I.getOpcode() == ARM::MVE_VLDRWU32) 1183 return false; 1184 } 1185 1186 if (LookAtSuccessors) { 1187 for (auto *Succ : BB->successors()) { 1188 if (!Visited.contains(Succ) && !is_contained(Frontier, Succ)) 1189 Frontier.push_back(Succ); 1190 } 1191 } 1192 Visited.insert(BB); 1193 Idx++; 1194 } 1195 1196 return true; 1197 } 1198 1199 bool LowOverheadLoop::ValidateMVEInst(MachineInstr *MI) { 1200 if (CannotTailPredicate) 1201 return false; 1202 1203 if (!shouldInspect(*MI)) 1204 return true; 1205 1206 if (MI->getOpcode() == ARM::MVE_VPSEL || 1207 MI->getOpcode() == ARM::MVE_VPNOT) { 1208 // TODO: Allow VPSEL and VPNOT, we currently cannot because: 1209 // 1) It will use the VPR as a predicate operand, but doesn't have to be 1210 // instead a VPT block, which means we can assert while building up 1211 // the VPT block because we don't find another VPT or VPST to being a new 1212 // one. 1213 // 2) VPSEL still requires a VPR operand even after tail predicating, 1214 // which means we can't remove it unless there is another 1215 // instruction, such as vcmp, that can provide the VPR def. 1216 return false; 1217 } 1218 1219 // Record all VCTPs and check that they're equivalent to one another. 1220 if (isVCTP(MI) && !AddVCTP(MI)) 1221 return false; 1222 1223 // Inspect uses first so that any instructions that alter the VPR don't 1224 // alter the predicate upon themselves. 1225 const MCInstrDesc &MCID = MI->getDesc(); 1226 bool IsUse = false; 1227 unsigned LastOpIdx = MI->getNumOperands() - 1; 1228 for (const auto &Op : enumerate(reverse(MCID.operands()))) { 1229 const MachineOperand &MO = MI->getOperand(LastOpIdx - Op.index()); 1230 if (!MO.isReg() || !MO.isUse() || MO.getReg() != ARM::VPR) 1231 continue; 1232 1233 if (ARM::isVpred(Op.value().OperandType)) { 1234 VPTstate.addInst(MI); 1235 IsUse = true; 1236 } else if (MI->getOpcode() != ARM::MVE_VPST) { 1237 LLVM_DEBUG(dbgs() << "ARM Loops: Found instruction using vpr: " << *MI); 1238 return false; 1239 } 1240 } 1241 1242 // If we find an instruction that has been marked as not valid for tail 1243 // predication, only allow the instruction if it's contained within a valid 1244 // VPT block. 1245 bool RequiresExplicitPredication = 1246 (MCID.TSFlags & ARMII::ValidForTailPredication) == 0; 1247 if (isDomainMVE(MI) && RequiresExplicitPredication) { 1248 if (MI->getOpcode() == ARM::MQPRCopy) 1249 return true; 1250 if (!IsUse && producesDoubleWidthResult(*MI)) { 1251 DoubleWidthResultInstrs.insert(MI); 1252 return true; 1253 } 1254 1255 LLVM_DEBUG(if (!IsUse) dbgs() 1256 << "ARM Loops: Can't tail predicate: " << *MI); 1257 return IsUse; 1258 } 1259 1260 // If the instruction is already explicitly predicated, then the conversion 1261 // will be fine, but ensure that all store operations are predicated. 1262 if (MI->mayStore() && !ValidateMVEStore(MI, &ML)) 1263 return IsUse; 1264 1265 // If this instruction defines the VPR, update the predicate for the 1266 // proceeding instructions. 1267 if (isVectorPredicate(MI)) { 1268 // Clear the existing predicate when we're not in VPT Active state, 1269 // otherwise we add to it. 1270 if (!isVectorPredicated(MI)) 1271 VPTstate.resetPredicate(MI); 1272 else 1273 VPTstate.addPredicate(MI); 1274 } 1275 1276 // Finally once the predicate has been modified, we can start a new VPT 1277 // block if necessary. 1278 if (isVPTOpcode(MI->getOpcode())) 1279 VPTstate.CreateVPTBlock(MI); 1280 1281 return true; 1282 } 1283 1284 bool ARMLowOverheadLoops::runOnMachineFunction(MachineFunction &mf) { 1285 const ARMSubtarget &ST = mf.getSubtarget<ARMSubtarget>(); 1286 if (!ST.hasLOB()) 1287 return false; 1288 1289 MF = &mf; 1290 LLVM_DEBUG(dbgs() << "ARM Loops on " << MF->getName() << " ------------- \n"); 1291 1292 MLI = &getAnalysis<MachineLoopInfoWrapperPass>().getLI(); 1293 RDA = &getAnalysis<ReachingDefAnalysis>(); 1294 MF->getProperties().setTracksLiveness(); 1295 MRI = &MF->getRegInfo(); 1296 TII = ST.getInstrInfo(); 1297 TRI = ST.getRegisterInfo(); 1298 BBUtils = std::make_unique<ARMBasicBlockUtils>(*MF); 1299 BBUtils->computeAllBlockSizes(); 1300 BBUtils->adjustBBOffsetsAfter(&MF->front()); 1301 1302 bool Changed = false; 1303 for (auto *ML : *MLI) { 1304 if (ML->isOutermost()) 1305 Changed |= ProcessLoop(ML); 1306 } 1307 Changed |= RevertNonLoops(); 1308 return Changed; 1309 } 1310 1311 bool ARMLowOverheadLoops::ProcessLoop(MachineLoop *ML) { 1312 bool Changed = false; 1313 1314 // Process inner loops first. 1315 for (MachineLoop *L : *ML) 1316 Changed |= ProcessLoop(L); 1317 1318 LLVM_DEBUG({ 1319 dbgs() << "ARM Loops: Processing loop containing:\n"; 1320 if (auto *Preheader = ML->getLoopPreheader()) 1321 dbgs() << " - Preheader: " << printMBBReference(*Preheader) << "\n"; 1322 else if (auto *Preheader = MLI->findLoopPreheader(ML, true, true)) 1323 dbgs() << " - Preheader: " << printMBBReference(*Preheader) << "\n"; 1324 for (auto *MBB : ML->getBlocks()) 1325 dbgs() << " - Block: " << printMBBReference(*MBB) << "\n"; 1326 }); 1327 1328 // Search the given block for a loop start instruction. If one isn't found, 1329 // and there's only one predecessor block, search that one too. 1330 std::function<MachineInstr*(MachineBasicBlock*)> SearchForStart = 1331 [&SearchForStart](MachineBasicBlock *MBB) -> MachineInstr* { 1332 for (auto &MI : *MBB) { 1333 if (isLoopStart(MI)) 1334 return &MI; 1335 } 1336 if (MBB->pred_size() == 1) 1337 return SearchForStart(*MBB->pred_begin()); 1338 return nullptr; 1339 }; 1340 1341 LowOverheadLoop LoLoop(*ML, *MLI, *RDA, *TRI, *TII); 1342 // Search the preheader for the start intrinsic. 1343 // FIXME: I don't see why we shouldn't be supporting multiple predecessors 1344 // with potentially multiple set.loop.iterations, so we need to enable this. 1345 if (LoLoop.Preheader) 1346 LoLoop.Start = SearchForStart(LoLoop.Preheader); 1347 else 1348 return Changed; 1349 1350 // Find the low-overhead loop components and decide whether or not to fall 1351 // back to a normal loop. Also look for a vctp instructions and decide 1352 // whether we can convert that predicate using tail predication. 1353 for (auto *MBB : reverse(ML->getBlocks())) { 1354 for (auto &MI : *MBB) { 1355 if (MI.isDebugValue()) 1356 continue; 1357 else if (MI.getOpcode() == ARM::t2LoopDec) 1358 LoLoop.Dec = &MI; 1359 else if (MI.getOpcode() == ARM::t2LoopEnd) 1360 LoLoop.End = &MI; 1361 else if (MI.getOpcode() == ARM::t2LoopEndDec) 1362 LoLoop.End = LoLoop.Dec = &MI; 1363 else if (isLoopStart(MI)) 1364 LoLoop.Start = &MI; 1365 else if (MI.getDesc().isCall()) { 1366 // TODO: Though the call will require LE to execute again, does this 1367 // mean we should revert? Always executing LE hopefully should be 1368 // faster than performing a sub,cmp,br or even subs,br. 1369 LoLoop.Revert = true; 1370 LLVM_DEBUG(dbgs() << "ARM Loops: Found call.\n"); 1371 } else { 1372 // Record VPR defs and build up their corresponding vpt blocks. 1373 // Check we know how to tail predicate any mve instructions. 1374 LoLoop.AnalyseMVEInst(&MI); 1375 } 1376 } 1377 } 1378 1379 LLVM_DEBUG(LoLoop.dump()); 1380 if (!LoLoop.FoundAllComponents()) { 1381 LLVM_DEBUG(dbgs() << "ARM Loops: Didn't find loop start, update, end\n"); 1382 return Changed; 1383 } 1384 1385 assert(LoLoop.Start->getOpcode() != ARM::t2WhileLoopStart && 1386 "Expected t2WhileLoopStart to be removed before regalloc!"); 1387 1388 // Check that the only instruction using LoopDec is LoopEnd. This can only 1389 // happen when the Dec and End are separate, not a single t2LoopEndDec. 1390 // TODO: Check for copy chains that really have no effect. 1391 if (LoLoop.Dec != LoLoop.End) { 1392 SmallPtrSet<MachineInstr *, 2> Uses; 1393 RDA->getReachingLocalUses(LoLoop.Dec, MCRegister::from(ARM::LR), Uses); 1394 if (Uses.size() > 1 || !Uses.count(LoLoop.End)) { 1395 LLVM_DEBUG(dbgs() << "ARM Loops: Unable to remove LoopDec.\n"); 1396 LoLoop.Revert = true; 1397 } 1398 } 1399 LoLoop.Validate(BBUtils.get()); 1400 Expand(LoLoop); 1401 return true; 1402 } 1403 1404 // WhileLoopStart holds the exit block, so produce a cmp lr, 0 and then a 1405 // beq that branches to the exit branch. 1406 // TODO: We could also try to generate a cbz if the value in LR is also in 1407 // another low register. 1408 void ARMLowOverheadLoops::RevertWhile(MachineInstr *MI) const { 1409 LLVM_DEBUG(dbgs() << "ARM Loops: Reverting to cmp: " << *MI); 1410 MachineBasicBlock *DestBB = getWhileLoopStartTargetBB(*MI); 1411 unsigned BrOpc = BBUtils->isBBInRange(MI, DestBB, 254) ? 1412 ARM::tBcc : ARM::t2Bcc; 1413 1414 RevertWhileLoopStartLR(MI, TII, BrOpc); 1415 } 1416 1417 void ARMLowOverheadLoops::RevertDo(MachineInstr *MI) const { 1418 LLVM_DEBUG(dbgs() << "ARM Loops: Reverting to mov: " << *MI); 1419 RevertDoLoopStart(MI, TII); 1420 } 1421 1422 bool ARMLowOverheadLoops::RevertLoopDec(MachineInstr *MI) const { 1423 LLVM_DEBUG(dbgs() << "ARM Loops: Reverting to sub: " << *MI); 1424 MachineBasicBlock *MBB = MI->getParent(); 1425 SmallPtrSet<MachineInstr*, 1> Ignore; 1426 for (auto I = MachineBasicBlock::iterator(MI), E = MBB->end(); I != E; ++I) { 1427 if (I->getOpcode() == ARM::t2LoopEnd) { 1428 Ignore.insert(&*I); 1429 break; 1430 } 1431 } 1432 1433 // If nothing defines CPSR between LoopDec and LoopEnd, use a t2SUBS. 1434 bool SetFlags = 1435 RDA->isSafeToDefRegAt(MI, MCRegister::from(ARM::CPSR), Ignore); 1436 1437 llvm::RevertLoopDec(MI, TII, SetFlags); 1438 return SetFlags; 1439 } 1440 1441 // Generate a subs, or sub and cmp, and a branch instead of an LE. 1442 void ARMLowOverheadLoops::RevertLoopEnd(MachineInstr *MI, bool SkipCmp) const { 1443 LLVM_DEBUG(dbgs() << "ARM Loops: Reverting to cmp, br: " << *MI); 1444 1445 MachineBasicBlock *DestBB = MI->getOperand(1).getMBB(); 1446 unsigned BrOpc = BBUtils->isBBInRange(MI, DestBB, 254) ? 1447 ARM::tBcc : ARM::t2Bcc; 1448 1449 llvm::RevertLoopEnd(MI, TII, BrOpc, SkipCmp); 1450 } 1451 1452 // Generate a subs, or sub and cmp, and a branch instead of an LE. 1453 void ARMLowOverheadLoops::RevertLoopEndDec(MachineInstr *MI) const { 1454 LLVM_DEBUG(dbgs() << "ARM Loops: Reverting to subs, br: " << *MI); 1455 assert(MI->getOpcode() == ARM::t2LoopEndDec && "Expected a t2LoopEndDec!"); 1456 MachineBasicBlock *MBB = MI->getParent(); 1457 1458 MachineInstrBuilder MIB = 1459 BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(ARM::t2SUBri)); 1460 MIB.addDef(ARM::LR); 1461 MIB.add(MI->getOperand(1)); 1462 MIB.addImm(1); 1463 MIB.addImm(ARMCC::AL); 1464 MIB.addReg(ARM::NoRegister); 1465 MIB.addReg(ARM::CPSR); 1466 MIB->getOperand(5).setIsDef(true); 1467 1468 MachineBasicBlock *DestBB = MI->getOperand(2).getMBB(); 1469 unsigned BrOpc = 1470 BBUtils->isBBInRange(MI, DestBB, 254) ? ARM::tBcc : ARM::t2Bcc; 1471 1472 // Create bne 1473 MIB = BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(BrOpc)); 1474 MIB.add(MI->getOperand(2)); // branch target 1475 MIB.addImm(ARMCC::NE); // condition code 1476 MIB.addReg(ARM::CPSR); 1477 1478 MI->eraseFromParent(); 1479 } 1480 1481 // Perform dead code elimation on the loop iteration count setup expression. 1482 // If we are tail-predicating, the number of elements to be processed is the 1483 // operand of the VCTP instruction in the vector body, see getCount(), which is 1484 // register $r3 in this example: 1485 // 1486 // $lr = big-itercount-expression 1487 // .. 1488 // $lr = t2DoLoopStart renamable $lr 1489 // vector.body: 1490 // .. 1491 // $vpr = MVE_VCTP32 renamable $r3 1492 // renamable $lr = t2LoopDec killed renamable $lr, 1 1493 // t2LoopEnd renamable $lr, %vector.body 1494 // tB %end 1495 // 1496 // What we would like achieve here is to replace the do-loop start pseudo 1497 // instruction t2DoLoopStart with: 1498 // 1499 // $lr = MVE_DLSTP_32 killed renamable $r3 1500 // 1501 // Thus, $r3 which defines the number of elements, is written to $lr, 1502 // and then we want to delete the whole chain that used to define $lr, 1503 // see the comment below how this chain could look like. 1504 // 1505 void ARMLowOverheadLoops::IterationCountDCE(LowOverheadLoop &LoLoop) { 1506 if (!LoLoop.IsTailPredicationLegal()) 1507 return; 1508 1509 LLVM_DEBUG(dbgs() << "ARM Loops: Trying DCE on loop iteration count.\n"); 1510 1511 MachineInstr *Def = RDA->getMIOperand(LoLoop.Start, 1); 1512 if (!Def) { 1513 LLVM_DEBUG(dbgs() << "ARM Loops: Couldn't find iteration count.\n"); 1514 return; 1515 } 1516 1517 // Collect and remove the users of iteration count. 1518 SmallPtrSet<MachineInstr*, 4> Killed = { LoLoop.Start, LoLoop.Dec, 1519 LoLoop.End }; 1520 if (!TryRemove(Def, *RDA, LoLoop.ToRemove, Killed)) 1521 LLVM_DEBUG(dbgs() << "ARM Loops: Unsafe to remove loop iteration count.\n"); 1522 } 1523 1524 MachineInstr* ARMLowOverheadLoops::ExpandLoopStart(LowOverheadLoop &LoLoop) { 1525 LLVM_DEBUG(dbgs() << "ARM Loops: Expanding LoopStart.\n"); 1526 // When using tail-predication, try to delete the dead code that was used to 1527 // calculate the number of loop iterations. 1528 IterationCountDCE(LoLoop); 1529 1530 MachineBasicBlock::iterator InsertPt = LoLoop.StartInsertPt; 1531 MachineInstr *Start = LoLoop.Start; 1532 MachineBasicBlock *MBB = LoLoop.StartInsertBB; 1533 unsigned Opc = LoLoop.getStartOpcode(); 1534 MachineOperand &Count = LoLoop.getLoopStartOperand(); 1535 1536 // A DLS lr, lr we needn't emit 1537 MachineInstr* NewStart; 1538 if (!DisableOmitDLS && Opc == ARM::t2DLS && Count.isReg() && 1539 Count.getReg() == ARM::LR) { 1540 LLVM_DEBUG(dbgs() << "ARM Loops: Didn't insert start: DLS lr, lr"); 1541 NewStart = nullptr; 1542 } else { 1543 MachineInstrBuilder MIB = 1544 BuildMI(*MBB, InsertPt, Start->getDebugLoc(), TII->get(Opc)); 1545 1546 MIB.addDef(ARM::LR); 1547 MIB.add(Count); 1548 if (isWhileLoopStart(*Start)) 1549 MIB.addMBB(getWhileLoopStartTargetBB(*Start)); 1550 1551 LLVM_DEBUG(dbgs() << "ARM Loops: Inserted start: " << *MIB); 1552 NewStart = &*MIB; 1553 } 1554 1555 LoLoop.ToRemove.insert(Start); 1556 return NewStart; 1557 } 1558 1559 void ARMLowOverheadLoops::ConvertVPTBlocks(LowOverheadLoop &LoLoop) { 1560 auto RemovePredicate = [](MachineInstr *MI) { 1561 if (MI->isDebugInstr()) 1562 return; 1563 LLVM_DEBUG(dbgs() << "ARM Loops: Removing predicate from: " << *MI); 1564 int PIdx = llvm::findFirstVPTPredOperandIdx(*MI); 1565 assert(PIdx >= 1 && "Trying to unpredicate a non-predicated instruction"); 1566 assert(MI->getOperand(PIdx).getImm() == ARMVCC::Then && 1567 "Expected Then predicate!"); 1568 MI->getOperand(PIdx).setImm(ARMVCC::None); 1569 MI->getOperand(PIdx + 1).setReg(0); 1570 }; 1571 1572 for (auto &Block : LoLoop.getVPTBlocks()) { 1573 SmallVectorImpl<MachineInstr *> &Insts = Block.getInsts(); 1574 1575 auto ReplaceVCMPWithVPT = [&](MachineInstr *&TheVCMP, MachineInstr *At) { 1576 assert(TheVCMP && "Replacing a removed or non-existent VCMP"); 1577 // Replace the VCMP with a VPT 1578 MachineInstrBuilder MIB = 1579 BuildMI(*At->getParent(), At, At->getDebugLoc(), 1580 TII->get(VCMPOpcodeToVPT(TheVCMP->getOpcode()))); 1581 MIB.addImm(ARMVCC::Then); 1582 // Register one 1583 MIB.add(TheVCMP->getOperand(1)); 1584 // Register two 1585 MIB.add(TheVCMP->getOperand(2)); 1586 // The comparison code, e.g. ge, eq, lt 1587 MIB.add(TheVCMP->getOperand(3)); 1588 LLVM_DEBUG(dbgs() << "ARM Loops: Combining with VCMP to VPT: " << *MIB); 1589 LoLoop.BlockMasksToRecompute.insert(MIB.getInstr()); 1590 LoLoop.ToRemove.insert(TheVCMP); 1591 TheVCMP = nullptr; 1592 }; 1593 1594 if (LoLoop.VPTstate.isEntryPredicatedOnVCTP(Block, /*exclusive*/ true)) { 1595 MachineInstr *VPST = Insts.front(); 1596 if (Block.hasUniformPredicate()) { 1597 // A vpt block starting with VPST, is only predicated upon vctp and has no 1598 // internal vpr defs: 1599 // - Remove vpst. 1600 // - Unpredicate the remaining instructions. 1601 LLVM_DEBUG(dbgs() << "ARM Loops: Removing VPST: " << *VPST); 1602 for (unsigned i = 1; i < Insts.size(); ++i) 1603 RemovePredicate(Insts[i]); 1604 } else { 1605 // The VPT block has a non-uniform predicate but it uses a vpst and its 1606 // entry is guarded only by a vctp, which means we: 1607 // - Need to remove the original vpst. 1608 // - Then need to unpredicate any following instructions, until 1609 // we come across the divergent vpr def. 1610 // - Insert a new vpst to predicate the instruction(s) that following 1611 // the divergent vpr def. 1612 MachineInstr *Divergent = Block.getDivergent(); 1613 MachineBasicBlock *MBB = Divergent->getParent(); 1614 auto DivergentNext = ++MachineBasicBlock::iterator(Divergent); 1615 while (DivergentNext != MBB->end() && DivergentNext->isDebugInstr()) 1616 ++DivergentNext; 1617 1618 bool DivergentNextIsPredicated = 1619 DivergentNext != MBB->end() && 1620 getVPTInstrPredicate(*DivergentNext) != ARMVCC::None; 1621 1622 for (auto I = ++MachineBasicBlock::iterator(VPST), E = DivergentNext; 1623 I != E; ++I) 1624 RemovePredicate(&*I); 1625 1626 // Check if the instruction defining vpr is a vcmp so it can be combined 1627 // with the VPST This should be the divergent instruction 1628 MachineInstr *VCMP = 1629 VCMPOpcodeToVPT(Divergent->getOpcode()) != 0 ? Divergent : nullptr; 1630 1631 if (DivergentNextIsPredicated) { 1632 // Insert a VPST at the divergent only if the next instruction 1633 // would actually use it. A VCMP following a VPST can be 1634 // merged into a VPT so do that instead if the VCMP exists. 1635 if (!VCMP) { 1636 // Create a VPST (with a null mask for now, we'll recompute it 1637 // later) 1638 MachineInstrBuilder MIB = 1639 BuildMI(*Divergent->getParent(), Divergent, 1640 Divergent->getDebugLoc(), TII->get(ARM::MVE_VPST)); 1641 MIB.addImm(0); 1642 LLVM_DEBUG(dbgs() << "ARM Loops: Created VPST: " << *MIB); 1643 LoLoop.BlockMasksToRecompute.insert(MIB.getInstr()); 1644 } else { 1645 // No RDA checks are necessary here since the VPST would have been 1646 // directly after the VCMP 1647 ReplaceVCMPWithVPT(VCMP, VCMP); 1648 } 1649 } 1650 } 1651 LLVM_DEBUG(dbgs() << "ARM Loops: Removing VPST: " << *VPST); 1652 LoLoop.ToRemove.insert(VPST); 1653 } else if (Block.containsVCTP()) { 1654 // The vctp will be removed, so either the entire block will be dead or 1655 // the block mask of the vp(s)t will need to be recomputed. 1656 MachineInstr *VPST = Insts.front(); 1657 if (Block.size() == 2) { 1658 assert(VPST->getOpcode() == ARM::MVE_VPST && 1659 "Found a VPST in an otherwise empty vpt block"); 1660 LoLoop.ToRemove.insert(VPST); 1661 } else 1662 LoLoop.BlockMasksToRecompute.insert(VPST); 1663 } else if (Insts.front()->getOpcode() == ARM::MVE_VPST) { 1664 // If this block starts with a VPST then attempt to merge it with the 1665 // preceeding un-merged VCMP into a VPT. This VCMP comes from a VPT 1666 // block that no longer exists 1667 MachineInstr *VPST = Insts.front(); 1668 auto Next = ++MachineBasicBlock::iterator(VPST); 1669 assert(getVPTInstrPredicate(*Next) != ARMVCC::None && 1670 "The instruction after a VPST must be predicated"); 1671 (void)Next; 1672 MachineInstr *VprDef = RDA->getUniqueReachingMIDef(VPST, ARM::VPR); 1673 if (VprDef && VCMPOpcodeToVPT(VprDef->getOpcode()) && 1674 !LoLoop.ToRemove.contains(VprDef)) { 1675 MachineInstr *VCMP = VprDef; 1676 // The VCMP and VPST can only be merged if the VCMP's operands will have 1677 // the same values at the VPST. 1678 // If any of the instructions between the VCMP and VPST are predicated 1679 // then a different code path is expected to have merged the VCMP and 1680 // VPST already. 1681 if (std::none_of(++MachineBasicBlock::iterator(VCMP), 1682 MachineBasicBlock::iterator(VPST), hasVPRUse) && 1683 RDA->hasSameReachingDef(VCMP, VPST, VCMP->getOperand(1).getReg()) && 1684 RDA->hasSameReachingDef(VCMP, VPST, VCMP->getOperand(2).getReg())) { 1685 ReplaceVCMPWithVPT(VCMP, VPST); 1686 LLVM_DEBUG(dbgs() << "ARM Loops: Removing VPST: " << *VPST); 1687 LoLoop.ToRemove.insert(VPST); 1688 } 1689 } 1690 } 1691 } 1692 1693 LoLoop.ToRemove.insert_range(LoLoop.VCTPs); 1694 } 1695 1696 void ARMLowOverheadLoops::Expand(LowOverheadLoop &LoLoop) { 1697 1698 // Combine the LoopDec and LoopEnd instructions into LE(TP). 1699 auto ExpandLoopEnd = [this](LowOverheadLoop &LoLoop) { 1700 MachineInstr *End = LoLoop.End; 1701 MachineBasicBlock *MBB = End->getParent(); 1702 unsigned Opc = LoLoop.IsTailPredicationLegal() ? 1703 ARM::MVE_LETP : ARM::t2LEUpdate; 1704 MachineInstrBuilder MIB = BuildMI(*MBB, End, End->getDebugLoc(), 1705 TII->get(Opc)); 1706 MIB.addDef(ARM::LR); 1707 unsigned Off = LoLoop.Dec == LoLoop.End ? 1 : 0; 1708 MIB.add(End->getOperand(Off + 0)); 1709 MIB.add(End->getOperand(Off + 1)); 1710 LLVM_DEBUG(dbgs() << "ARM Loops: Inserted LE: " << *MIB); 1711 LoLoop.ToRemove.insert(LoLoop.Dec); 1712 LoLoop.ToRemove.insert(End); 1713 return &*MIB; 1714 }; 1715 1716 // TODO: We should be able to automatically remove these branches before we 1717 // get here - probably by teaching analyzeBranch about the pseudo 1718 // instructions. 1719 // If there is an unconditional branch, after I, that just branches to the 1720 // next block, remove it. 1721 auto RemoveDeadBranch = [](MachineInstr *I) { 1722 MachineBasicBlock *BB = I->getParent(); 1723 MachineInstr *Terminator = &BB->instr_back(); 1724 if (Terminator->isUnconditionalBranch() && I != Terminator) { 1725 MachineBasicBlock *Succ = Terminator->getOperand(0).getMBB(); 1726 if (BB->isLayoutSuccessor(Succ)) { 1727 LLVM_DEBUG(dbgs() << "ARM Loops: Removing branch: " << *Terminator); 1728 Terminator->eraseFromParent(); 1729 } 1730 } 1731 }; 1732 1733 // And VMOVCopies need to become 2xVMOVD for tail predication to be valid. 1734 // Anything other MQPRCopy can be converted to MVE_VORR later on. 1735 auto ExpandVMOVCopies = [this](SmallPtrSet<MachineInstr *, 4> &VMOVCopies) { 1736 for (auto *MI : VMOVCopies) { 1737 LLVM_DEBUG(dbgs() << "Converting copy to VMOVD: " << *MI); 1738 assert(MI->getOpcode() == ARM::MQPRCopy && "Only expected MQPRCOPY!"); 1739 MachineBasicBlock *MBB = MI->getParent(); 1740 Register Dst = MI->getOperand(0).getReg(); 1741 Register Src = MI->getOperand(1).getReg(); 1742 auto MIB1 = BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(ARM::VMOVD), 1743 ARM::D0 + (Dst - ARM::Q0) * 2) 1744 .addReg(ARM::D0 + (Src - ARM::Q0) * 2) 1745 .add(predOps(ARMCC::AL)); 1746 (void)MIB1; 1747 LLVM_DEBUG(dbgs() << " into " << *MIB1); 1748 auto MIB2 = BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(ARM::VMOVD), 1749 ARM::D0 + (Dst - ARM::Q0) * 2 + 1) 1750 .addReg(ARM::D0 + (Src - ARM::Q0) * 2 + 1) 1751 .add(predOps(ARMCC::AL)); 1752 LLVM_DEBUG(dbgs() << " and " << *MIB2); 1753 (void)MIB2; 1754 MI->eraseFromParent(); 1755 } 1756 }; 1757 1758 if (LoLoop.Revert) { 1759 if (isWhileLoopStart(*LoLoop.Start)) 1760 RevertWhile(LoLoop.Start); 1761 else 1762 RevertDo(LoLoop.Start); 1763 if (LoLoop.Dec == LoLoop.End) 1764 RevertLoopEndDec(LoLoop.End); 1765 else 1766 RevertLoopEnd(LoLoop.End, RevertLoopDec(LoLoop.Dec)); 1767 } else { 1768 ExpandVMOVCopies(LoLoop.VMOVCopies); 1769 LoLoop.Start = ExpandLoopStart(LoLoop); 1770 if (LoLoop.Start) 1771 RemoveDeadBranch(LoLoop.Start); 1772 LoLoop.End = ExpandLoopEnd(LoLoop); 1773 RemoveDeadBranch(LoLoop.End); 1774 if (LoLoop.IsTailPredicationLegal()) 1775 ConvertVPTBlocks(LoLoop); 1776 for (auto *I : LoLoop.ToRemove) { 1777 LLVM_DEBUG(dbgs() << "ARM Loops: Erasing " << *I); 1778 I->eraseFromParent(); 1779 } 1780 for (auto *I : LoLoop.BlockMasksToRecompute) { 1781 LLVM_DEBUG(dbgs() << "ARM Loops: Recomputing VPT/VPST Block Mask: " << *I); 1782 recomputeVPTBlockMask(*I); 1783 LLVM_DEBUG(dbgs() << " ... done: " << *I); 1784 } 1785 } 1786 1787 PostOrderLoopTraversal DFS(LoLoop.ML, *MLI); 1788 DFS.ProcessLoop(); 1789 const SmallVectorImpl<MachineBasicBlock*> &PostOrder = DFS.getOrder(); 1790 fullyRecomputeLiveIns(PostOrder); 1791 1792 for (auto *MBB : reverse(PostOrder)) 1793 recomputeLivenessFlags(*MBB); 1794 1795 // We've moved, removed and inserted new instructions, so update RDA. 1796 RDA->reset(); 1797 } 1798 1799 bool ARMLowOverheadLoops::RevertNonLoops() { 1800 LLVM_DEBUG(dbgs() << "ARM Loops: Reverting any remaining pseudos...\n"); 1801 bool Changed = false; 1802 1803 for (auto &MBB : *MF) { 1804 SmallVector<MachineInstr*, 4> Starts; 1805 SmallVector<MachineInstr*, 4> Decs; 1806 SmallVector<MachineInstr*, 4> Ends; 1807 SmallVector<MachineInstr *, 4> EndDecs; 1808 SmallVector<MachineInstr *, 4> MQPRCopies; 1809 1810 for (auto &I : MBB) { 1811 if (isLoopStart(I)) 1812 Starts.push_back(&I); 1813 else if (I.getOpcode() == ARM::t2LoopDec) 1814 Decs.push_back(&I); 1815 else if (I.getOpcode() == ARM::t2LoopEnd) 1816 Ends.push_back(&I); 1817 else if (I.getOpcode() == ARM::t2LoopEndDec) 1818 EndDecs.push_back(&I); 1819 else if (I.getOpcode() == ARM::MQPRCopy) 1820 MQPRCopies.push_back(&I); 1821 } 1822 1823 if (Starts.empty() && Decs.empty() && Ends.empty() && EndDecs.empty() && 1824 MQPRCopies.empty()) 1825 continue; 1826 1827 Changed = true; 1828 1829 for (auto *Start : Starts) { 1830 if (isWhileLoopStart(*Start)) 1831 RevertWhile(Start); 1832 else 1833 RevertDo(Start); 1834 } 1835 for (auto *Dec : Decs) 1836 RevertLoopDec(Dec); 1837 1838 for (auto *End : Ends) 1839 RevertLoopEnd(End); 1840 for (auto *End : EndDecs) 1841 RevertLoopEndDec(End); 1842 for (auto *MI : MQPRCopies) { 1843 LLVM_DEBUG(dbgs() << "Converting copy to VORR: " << *MI); 1844 assert(MI->getOpcode() == ARM::MQPRCopy && "Only expected MQPRCOPY!"); 1845 MachineBasicBlock *MBB = MI->getParent(); 1846 auto MIB = BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(ARM::MVE_VORR), 1847 MI->getOperand(0).getReg()) 1848 .add(MI->getOperand(1)) 1849 .add(MI->getOperand(1)); 1850 addUnpredicatedMveVpredROp(MIB, MI->getOperand(0).getReg()); 1851 MI->eraseFromParent(); 1852 } 1853 } 1854 return Changed; 1855 } 1856 1857 FunctionPass *llvm::createARMLowOverheadLoopsPass() { 1858 return new ARMLowOverheadLoops(); 1859 } 1860