1 //===-- AArch64A57FPLoadBalancing.cpp - Balance FP ops statically on A57---===// 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 // For best-case performance on Cortex-A57, we should try to use a balanced 9 // mix of odd and even D-registers when performing a critical sequence of 10 // independent, non-quadword FP/ASIMD floating-point multiply or 11 // multiply-accumulate operations. 12 // 13 // This pass attempts to detect situations where the register allocation may 14 // adversely affect this load balancing and to change the registers used so as 15 // to better utilize the CPU. 16 // 17 // Ideally we'd just take each multiply or multiply-accumulate in turn and 18 // allocate it alternating even or odd registers. However, multiply-accumulates 19 // are most efficiently performed in the same functional unit as their 20 // accumulation operand. Therefore this pass tries to find maximal sequences 21 // ("Chains") of multiply-accumulates linked via their accumulation operand, 22 // and assign them all the same "color" (oddness/evenness). 23 // 24 // This optimization affects S-register and D-register floating point 25 // multiplies and FMADD/FMAs, as well as vector (floating point only) muls and 26 // FMADD/FMA. Q register instructions (and 128-bit vector instructions) are 27 // not affected. 28 //===----------------------------------------------------------------------===// 29 30 #include "AArch64.h" 31 #include "AArch64InstrInfo.h" 32 #include "AArch64Subtarget.h" 33 #include "llvm/ADT/EquivalenceClasses.h" 34 #include "llvm/CodeGen/MachineFunction.h" 35 #include "llvm/CodeGen/MachineFunctionPass.h" 36 #include "llvm/CodeGen/MachineInstr.h" 37 #include "llvm/CodeGen/MachineInstrBuilder.h" 38 #include "llvm/CodeGen/MachineRegisterInfo.h" 39 #include "llvm/CodeGen/RegisterClassInfo.h" 40 #include "llvm/CodeGen/RegisterScavenging.h" 41 #include "llvm/Support/CommandLine.h" 42 #include "llvm/Support/Debug.h" 43 #include "llvm/Support/raw_ostream.h" 44 using namespace llvm; 45 46 #define DEBUG_TYPE "aarch64-a57-fp-load-balancing" 47 48 // Enforce the algorithm to use the scavenged register even when the original 49 // destination register is the correct color. Used for testing. 50 static cl::opt<bool> 51 TransformAll("aarch64-a57-fp-load-balancing-force-all", 52 cl::desc("Always modify dest registers regardless of color"), 53 cl::init(false), cl::Hidden); 54 55 // Never use the balance information obtained from chains - return a specific 56 // color always. Used for testing. 57 static cl::opt<unsigned> 58 OverrideBalance("aarch64-a57-fp-load-balancing-override", 59 cl::desc("Ignore balance information, always return " 60 "(1: Even, 2: Odd)."), 61 cl::init(0), cl::Hidden); 62 63 //===----------------------------------------------------------------------===// 64 // Helper functions 65 66 // Is the instruction a type of multiply on 64-bit (or 32-bit) FPRs? 67 static bool isMul(MachineInstr *MI) { 68 switch (MI->getOpcode()) { 69 case AArch64::FMULSrr: 70 case AArch64::FNMULSrr: 71 case AArch64::FMULDrr: 72 case AArch64::FNMULDrr: 73 return true; 74 default: 75 return false; 76 } 77 } 78 79 // Is the instruction a type of FP multiply-accumulate on 64-bit (or 32-bit) FPRs? 80 static bool isMla(MachineInstr *MI) { 81 switch (MI->getOpcode()) { 82 case AArch64::FMSUBSrrr: 83 case AArch64::FMADDSrrr: 84 case AArch64::FNMSUBSrrr: 85 case AArch64::FNMADDSrrr: 86 case AArch64::FMSUBDrrr: 87 case AArch64::FMADDDrrr: 88 case AArch64::FNMSUBDrrr: 89 case AArch64::FNMADDDrrr: 90 return true; 91 default: 92 return false; 93 } 94 } 95 96 //===----------------------------------------------------------------------===// 97 98 namespace { 99 /// A "color", which is either even or odd. Yes, these aren't really colors 100 /// but the algorithm is conceptually doing two-color graph coloring. 101 enum class Color { Even, Odd }; 102 #ifndef NDEBUG 103 static const char *ColorNames[2] = { "Even", "Odd" }; 104 #endif 105 106 class Chain; 107 108 class AArch64A57FPLoadBalancing : public MachineFunctionPass { 109 MachineRegisterInfo *MRI; 110 const TargetRegisterInfo *TRI; 111 RegisterClassInfo RCI; 112 113 public: 114 static char ID; 115 explicit AArch64A57FPLoadBalancing() : MachineFunctionPass(ID) { 116 initializeAArch64A57FPLoadBalancingPass(*PassRegistry::getPassRegistry()); 117 } 118 119 bool runOnMachineFunction(MachineFunction &F) override; 120 121 MachineFunctionProperties getRequiredProperties() const override { 122 return MachineFunctionProperties().set( 123 MachineFunctionProperties::Property::NoVRegs); 124 } 125 126 StringRef getPassName() const override { 127 return "A57 FP Anti-dependency breaker"; 128 } 129 130 void getAnalysisUsage(AnalysisUsage &AU) const override { 131 AU.setPreservesCFG(); 132 MachineFunctionPass::getAnalysisUsage(AU); 133 } 134 135 private: 136 bool runOnBasicBlock(MachineBasicBlock &MBB); 137 bool colorChainSet(std::vector<Chain*> GV, MachineBasicBlock &MBB, 138 int &Balance); 139 bool colorChain(Chain *G, Color C, MachineBasicBlock &MBB); 140 int scavengeRegister(Chain *G, Color C, MachineBasicBlock &MBB); 141 void scanInstruction(MachineInstr *MI, unsigned Idx, 142 std::map<unsigned, Chain*> &Active, 143 std::vector<std::unique_ptr<Chain>> &AllChains); 144 void maybeKillChain(MachineOperand &MO, unsigned Idx, 145 std::map<unsigned, Chain*> &RegChains); 146 Color getColor(unsigned Register); 147 Chain *getAndEraseNext(Color PreferredColor, std::vector<Chain*> &L); 148 }; 149 } 150 151 char AArch64A57FPLoadBalancing::ID = 0; 152 153 INITIALIZE_PASS_BEGIN(AArch64A57FPLoadBalancing, DEBUG_TYPE, 154 "AArch64 A57 FP Load-Balancing", false, false) 155 INITIALIZE_PASS_END(AArch64A57FPLoadBalancing, DEBUG_TYPE, 156 "AArch64 A57 FP Load-Balancing", false, false) 157 158 namespace { 159 /// A Chain is a sequence of instructions that are linked together by 160 /// an accumulation operand. For example: 161 /// 162 /// fmul def d0, ? 163 /// fmla def d1, ?, ?, killed d0 164 /// fmla def d2, ?, ?, killed d1 165 /// 166 /// There may be other instructions interleaved in the sequence that 167 /// do not belong to the chain. These other instructions must not use 168 /// the "chain" register at any point. 169 /// 170 /// We currently only support chains where the "chain" operand is killed 171 /// at each link in the chain for simplicity. 172 /// A chain has three important instructions - Start, Last and Kill. 173 /// * The start instruction is the first instruction in the chain. 174 /// * Last is the final instruction in the chain. 175 /// * Kill may or may not be defined. If defined, Kill is the instruction 176 /// where the outgoing value of the Last instruction is killed. 177 /// This information is important as if we know the outgoing value is 178 /// killed with no intervening uses, we can safely change its register. 179 /// 180 /// Without a kill instruction, we must assume the outgoing value escapes 181 /// beyond our model and either must not change its register or must 182 /// create a fixup FMOV to keep the old register value consistent. 183 /// 184 class Chain { 185 public: 186 /// The important (marker) instructions. 187 MachineInstr *StartInst, *LastInst, *KillInst; 188 /// The index, from the start of the basic block, that each marker 189 /// appears. These are stored so we can do quick interval tests. 190 unsigned StartInstIdx, LastInstIdx, KillInstIdx; 191 /// All instructions in the chain. 192 std::set<MachineInstr*> Insts; 193 /// True if KillInst cannot be modified. If this is true, 194 /// we cannot change LastInst's outgoing register. 195 /// This will be true for tied values and regmasks. 196 bool KillIsImmutable; 197 /// The "color" of LastInst. This will be the preferred chain color, 198 /// as changing intermediate nodes is easy but changing the last 199 /// instruction can be more tricky. 200 Color LastColor; 201 202 Chain(MachineInstr *MI, unsigned Idx, Color C) 203 : StartInst(MI), LastInst(MI), KillInst(nullptr), 204 StartInstIdx(Idx), LastInstIdx(Idx), KillInstIdx(0), 205 LastColor(C) { 206 Insts.insert(MI); 207 } 208 209 /// Add a new instruction into the chain. The instruction's dest operand 210 /// has the given color. 211 void add(MachineInstr *MI, unsigned Idx, Color C) { 212 LastInst = MI; 213 LastInstIdx = Idx; 214 LastColor = C; 215 assert((KillInstIdx == 0 || LastInstIdx < KillInstIdx) && 216 "Chain: broken invariant. A Chain can only be killed after its last " 217 "def"); 218 219 Insts.insert(MI); 220 } 221 222 /// Return true if MI is a member of the chain. 223 bool contains(MachineInstr &MI) { return Insts.count(&MI) > 0; } 224 225 /// Return the number of instructions in the chain. 226 unsigned size() const { 227 return Insts.size(); 228 } 229 230 /// Inform the chain that its last active register (the dest register of 231 /// LastInst) is killed by MI with no intervening uses or defs. 232 void setKill(MachineInstr *MI, unsigned Idx, bool Immutable) { 233 KillInst = MI; 234 KillInstIdx = Idx; 235 KillIsImmutable = Immutable; 236 assert((KillInstIdx == 0 || LastInstIdx < KillInstIdx) && 237 "Chain: broken invariant. A Chain can only be killed after its last " 238 "def"); 239 } 240 241 /// Return the first instruction in the chain. 242 MachineInstr *getStart() const { return StartInst; } 243 /// Return the last instruction in the chain. 244 MachineInstr *getLast() const { return LastInst; } 245 /// Return the "kill" instruction (as set with setKill()) or NULL. 246 MachineInstr *getKill() const { return KillInst; } 247 /// Return an instruction that can be used as an iterator for the end 248 /// of the chain. This is the maximum of KillInst (if set) and LastInst. 249 MachineBasicBlock::iterator end() const { 250 return ++MachineBasicBlock::iterator(KillInst ? KillInst : LastInst); 251 } 252 MachineBasicBlock::iterator begin() const { return getStart(); } 253 254 /// Can the Kill instruction (assuming one exists) be modified? 255 bool isKillImmutable() const { return KillIsImmutable; } 256 257 /// Return the preferred color of this chain. 258 Color getPreferredColor() { 259 if (OverrideBalance != 0) 260 return OverrideBalance == 1 ? Color::Even : Color::Odd; 261 return LastColor; 262 } 263 264 /// Return true if this chain (StartInst..KillInst) overlaps with Other. 265 bool rangeOverlapsWith(const Chain &Other) const { 266 unsigned End = KillInst ? KillInstIdx : LastInstIdx; 267 unsigned OtherEnd = Other.KillInst ? 268 Other.KillInstIdx : Other.LastInstIdx; 269 270 return StartInstIdx <= OtherEnd && Other.StartInstIdx <= End; 271 } 272 273 /// Return true if this chain starts before Other. 274 bool startsBefore(const Chain *Other) const { 275 return StartInstIdx < Other->StartInstIdx; 276 } 277 278 /// Return true if the group will require a fixup MOV at the end. 279 bool requiresFixup() const { 280 return (getKill() && isKillImmutable()) || !getKill(); 281 } 282 283 /// Return a simple string representation of the chain. 284 std::string str() const { 285 std::string S; 286 raw_string_ostream OS(S); 287 288 OS << "{"; 289 StartInst->print(OS, /* SkipOpers= */true); 290 OS << " -> "; 291 LastInst->print(OS, /* SkipOpers= */true); 292 if (KillInst) { 293 OS << " (kill @ "; 294 KillInst->print(OS, /* SkipOpers= */true); 295 OS << ")"; 296 } 297 OS << "}"; 298 299 return OS.str(); 300 } 301 302 }; 303 304 } // end anonymous namespace 305 306 //===----------------------------------------------------------------------===// 307 308 bool AArch64A57FPLoadBalancing::runOnMachineFunction(MachineFunction &F) { 309 if (skipFunction(F.getFunction())) 310 return false; 311 312 if (!F.getSubtarget<AArch64Subtarget>().balanceFPOps()) 313 return false; 314 315 bool Changed = false; 316 LLVM_DEBUG(dbgs() << "***** AArch64A57FPLoadBalancing *****\n"); 317 318 MRI = &F.getRegInfo(); 319 TRI = F.getRegInfo().getTargetRegisterInfo(); 320 RCI.runOnMachineFunction(F); 321 322 for (auto &MBB : F) { 323 Changed |= runOnBasicBlock(MBB); 324 } 325 326 return Changed; 327 } 328 329 bool AArch64A57FPLoadBalancing::runOnBasicBlock(MachineBasicBlock &MBB) { 330 bool Changed = false; 331 LLVM_DEBUG(dbgs() << "Running on MBB: " << MBB 332 << " - scanning instructions...\n"); 333 334 // First, scan the basic block producing a set of chains. 335 336 // The currently "active" chains - chains that can be added to and haven't 337 // been killed yet. This is keyed by register - all chains can only have one 338 // "link" register between each inst in the chain. 339 std::map<unsigned, Chain*> ActiveChains; 340 std::vector<std::unique_ptr<Chain>> AllChains; 341 unsigned Idx = 0; 342 for (auto &MI : MBB) 343 scanInstruction(&MI, Idx++, ActiveChains, AllChains); 344 345 LLVM_DEBUG(dbgs() << "Scan complete, " << AllChains.size() 346 << " chains created.\n"); 347 348 // Group the chains into disjoint sets based on their liveness range. This is 349 // a poor-man's version of graph coloring. Ideally we'd create an interference 350 // graph and perform full-on graph coloring on that, but; 351 // (a) That's rather heavyweight for only two colors. 352 // (b) We expect multiple disjoint interference regions - in practice the live 353 // range of chains is quite small and they are clustered between loads 354 // and stores. 355 EquivalenceClasses<Chain*> EC; 356 for (auto &I : AllChains) 357 EC.insert(I.get()); 358 359 for (auto &I : AllChains) 360 for (auto &J : AllChains) 361 if (I != J && I->rangeOverlapsWith(*J)) 362 EC.unionSets(I.get(), J.get()); 363 LLVM_DEBUG(dbgs() << "Created " << EC.getNumClasses() << " disjoint sets.\n"); 364 365 // Now we assume that every member of an equivalence class interferes 366 // with every other member of that class, and with no members of other classes. 367 368 // Convert the EquivalenceClasses to a simpler set of sets. 369 std::vector<std::vector<Chain*> > V; 370 for (auto I = EC.begin(), E = EC.end(); I != E; ++I) { 371 std::vector<Chain*> Cs(EC.member_begin(I), EC.member_end()); 372 if (Cs.empty()) continue; 373 V.push_back(std::move(Cs)); 374 } 375 376 // Now we have a set of sets, order them by start address so 377 // we can iterate over them sequentially. 378 llvm::sort(V, 379 [](const std::vector<Chain *> &A, const std::vector<Chain *> &B) { 380 return A.front()->startsBefore(B.front()); 381 }); 382 383 // As we only have two colors, we can track the global (BB-level) balance of 384 // odds versus evens. We aim to keep this near zero to keep both execution 385 // units fed. 386 // Positive means we're even-heavy, negative we're odd-heavy. 387 // 388 // FIXME: If chains have interdependencies, for example: 389 // mul r0, r1, r2 390 // mul r3, r0, r1 391 // We do not model this and may color each one differently, assuming we'll 392 // get ILP when we obviously can't. This hasn't been seen to be a problem 393 // in practice so far, so we simplify the algorithm by ignoring it. 394 int Parity = 0; 395 396 for (auto &I : V) 397 Changed |= colorChainSet(std::move(I), MBB, Parity); 398 399 return Changed; 400 } 401 402 Chain *AArch64A57FPLoadBalancing::getAndEraseNext(Color PreferredColor, 403 std::vector<Chain*> &L) { 404 if (L.empty()) 405 return nullptr; 406 407 // We try and get the best candidate from L to color next, given that our 408 // preferred color is "PreferredColor". L is ordered from larger to smaller 409 // chains. It is beneficial to color the large chains before the small chains, 410 // but if we can't find a chain of the maximum length with the preferred color, 411 // we fuzz the size and look for slightly smaller chains before giving up and 412 // returning a chain that must be recolored. 413 414 // FIXME: Does this need to be configurable? 415 const unsigned SizeFuzz = 1; 416 unsigned MinSize = L.front()->size() - SizeFuzz; 417 for (auto I = L.begin(), E = L.end(); I != E; ++I) { 418 if ((*I)->size() <= MinSize) { 419 // We've gone past the size limit. Return the previous item. 420 Chain *Ch = *--I; 421 L.erase(I); 422 return Ch; 423 } 424 425 if ((*I)->getPreferredColor() == PreferredColor) { 426 Chain *Ch = *I; 427 L.erase(I); 428 return Ch; 429 } 430 } 431 432 // Bailout case - just return the first item. 433 Chain *Ch = L.front(); 434 L.erase(L.begin()); 435 return Ch; 436 } 437 438 bool AArch64A57FPLoadBalancing::colorChainSet(std::vector<Chain*> GV, 439 MachineBasicBlock &MBB, 440 int &Parity) { 441 bool Changed = false; 442 LLVM_DEBUG(dbgs() << "colorChainSet(): #sets=" << GV.size() << "\n"); 443 444 // Sort by descending size order so that we allocate the most important 445 // sets first. 446 // Tie-break equivalent sizes by sorting chains requiring fixups before 447 // those without fixups. The logic here is that we should look at the 448 // chains that we cannot change before we look at those we can, 449 // so the parity counter is updated and we know what color we should 450 // change them to! 451 // Final tie-break with instruction order so pass output is stable (i.e. not 452 // dependent on malloc'd pointer values). 453 llvm::sort(GV, [](const Chain *G1, const Chain *G2) { 454 if (G1->size() != G2->size()) 455 return G1->size() > G2->size(); 456 if (G1->requiresFixup() != G2->requiresFixup()) 457 return G1->requiresFixup() > G2->requiresFixup(); 458 // Make sure startsBefore() produces a stable final order. 459 assert((G1 == G2 || (G1->startsBefore(G2) ^ G2->startsBefore(G1))) && 460 "Starts before not total order!"); 461 return G1->startsBefore(G2); 462 }); 463 464 Color PreferredColor = Parity < 0 ? Color::Even : Color::Odd; 465 while (Chain *G = getAndEraseNext(PreferredColor, GV)) { 466 // Start off by assuming we'll color to our own preferred color. 467 Color C = PreferredColor; 468 if (Parity == 0) 469 // But if we really don't care, use the chain's preferred color. 470 C = G->getPreferredColor(); 471 472 LLVM_DEBUG(dbgs() << " - Parity=" << Parity 473 << ", Color=" << ColorNames[(int)C] << "\n"); 474 475 // If we'll need a fixup FMOV, don't bother. Testing has shown that this 476 // happens infrequently and when it does it has at least a 50% chance of 477 // slowing code down instead of speeding it up. 478 if (G->requiresFixup() && C != G->getPreferredColor()) { 479 C = G->getPreferredColor(); 480 LLVM_DEBUG(dbgs() << " - " << G->str() 481 << " - not worthwhile changing; " 482 "color remains " 483 << ColorNames[(int)C] << "\n"); 484 } 485 486 Changed |= colorChain(G, C, MBB); 487 488 Parity += (C == Color::Even) ? G->size() : -G->size(); 489 PreferredColor = Parity < 0 ? Color::Even : Color::Odd; 490 } 491 492 return Changed; 493 } 494 495 int AArch64A57FPLoadBalancing::scavengeRegister(Chain *G, Color C, 496 MachineBasicBlock &MBB) { 497 // Can we find an appropriate register that is available throughout the life 498 // of the chain? Simulate liveness backwards until the end of the chain. 499 LiveRegUnits Units(*TRI); 500 Units.addLiveOuts(MBB); 501 MachineBasicBlock::iterator I = MBB.end(); 502 MachineBasicBlock::iterator ChainEnd = G->end(); 503 while (I != ChainEnd) { 504 --I; 505 Units.stepBackward(*I); 506 } 507 508 // Check which register units are alive throughout the chain. 509 MachineBasicBlock::iterator ChainBegin = G->begin(); 510 assert(ChainBegin != ChainEnd && "Chain should contain instructions"); 511 do { 512 --I; 513 Units.accumulate(*I); 514 } while (I != ChainBegin); 515 516 // Make sure we allocate in-order, to get the cheapest registers first. 517 unsigned RegClassID = ChainBegin->getDesc().operands()[0].RegClass; 518 auto Ord = RCI.getOrder(TRI->getRegClass(RegClassID)); 519 for (auto Reg : Ord) { 520 if (!Units.available(Reg)) 521 continue; 522 if (C == getColor(Reg)) 523 return Reg; 524 } 525 526 return -1; 527 } 528 529 bool AArch64A57FPLoadBalancing::colorChain(Chain *G, Color C, 530 MachineBasicBlock &MBB) { 531 bool Changed = false; 532 LLVM_DEBUG(dbgs() << " - colorChain(" << G->str() << ", " 533 << ColorNames[(int)C] << ")\n"); 534 535 // Try and obtain a free register of the right class. Without a register 536 // to play with we cannot continue. 537 int Reg = scavengeRegister(G, C, MBB); 538 if (Reg == -1) { 539 LLVM_DEBUG(dbgs() << "Scavenging (thus coloring) failed!\n"); 540 return false; 541 } 542 LLVM_DEBUG(dbgs() << " - Scavenged register: " << printReg(Reg, TRI) << "\n"); 543 544 std::map<unsigned, unsigned> Substs; 545 for (MachineInstr &I : *G) { 546 if (!G->contains(I) && (&I != G->getKill() || G->isKillImmutable())) 547 continue; 548 549 // I is a member of G, or I is a mutable instruction that kills G. 550 551 std::vector<unsigned> ToErase; 552 for (auto &U : I.operands()) { 553 if (U.isReg() && U.isUse() && Substs.find(U.getReg()) != Substs.end()) { 554 Register OrigReg = U.getReg(); 555 U.setReg(Substs[OrigReg]); 556 if (U.isKill()) 557 // Don't erase straight away, because there may be other operands 558 // that also reference this substitution! 559 ToErase.push_back(OrigReg); 560 } else if (U.isRegMask()) { 561 for (auto J : Substs) { 562 if (U.clobbersPhysReg(J.first)) 563 ToErase.push_back(J.first); 564 } 565 } 566 } 567 // Now it's safe to remove the substs identified earlier. 568 for (auto J : ToErase) 569 Substs.erase(J); 570 571 // Only change the def if this isn't the last instruction. 572 if (&I != G->getKill()) { 573 MachineOperand &MO = I.getOperand(0); 574 575 bool Change = TransformAll || getColor(MO.getReg()) != C; 576 if (G->requiresFixup() && &I == G->getLast()) 577 Change = false; 578 579 if (Change) { 580 Substs[MO.getReg()] = Reg; 581 MO.setReg(Reg); 582 583 Changed = true; 584 } 585 } 586 } 587 assert(Substs.size() == 0 && "No substitutions should be left active!"); 588 589 if (G->getKill()) { 590 LLVM_DEBUG(dbgs() << " - Kill instruction seen.\n"); 591 } else { 592 // We didn't have a kill instruction, but we didn't seem to need to change 593 // the destination register anyway. 594 LLVM_DEBUG(dbgs() << " - Destination register not changed.\n"); 595 } 596 return Changed; 597 } 598 599 void AArch64A57FPLoadBalancing::scanInstruction( 600 MachineInstr *MI, unsigned Idx, std::map<unsigned, Chain *> &ActiveChains, 601 std::vector<std::unique_ptr<Chain>> &AllChains) { 602 // Inspect "MI", updating ActiveChains and AllChains. 603 604 if (isMul(MI)) { 605 606 for (auto &I : MI->uses()) 607 maybeKillChain(I, Idx, ActiveChains); 608 for (auto &I : MI->defs()) 609 maybeKillChain(I, Idx, ActiveChains); 610 611 // Create a new chain. Multiplies don't require forwarding so can go on any 612 // unit. 613 Register DestReg = MI->getOperand(0).getReg(); 614 615 LLVM_DEBUG(dbgs() << "New chain started for register " 616 << printReg(DestReg, TRI) << " at " << *MI); 617 618 auto G = std::make_unique<Chain>(MI, Idx, getColor(DestReg)); 619 ActiveChains[DestReg] = G.get(); 620 AllChains.push_back(std::move(G)); 621 622 } else if (isMla(MI)) { 623 624 // It is beneficial to keep MLAs on the same functional unit as their 625 // accumulator operand. 626 Register DestReg = MI->getOperand(0).getReg(); 627 Register AccumReg = MI->getOperand(3).getReg(); 628 629 maybeKillChain(MI->getOperand(1), Idx, ActiveChains); 630 maybeKillChain(MI->getOperand(2), Idx, ActiveChains); 631 if (DestReg != AccumReg) 632 maybeKillChain(MI->getOperand(0), Idx, ActiveChains); 633 634 if (ActiveChains.find(AccumReg) != ActiveChains.end()) { 635 LLVM_DEBUG(dbgs() << "Chain found for accumulator register " 636 << printReg(AccumReg, TRI) << " in MI " << *MI); 637 638 // For simplicity we only chain together sequences of MULs/MLAs where the 639 // accumulator register is killed on each instruction. This means we don't 640 // need to track other uses of the registers we want to rewrite. 641 // 642 // FIXME: We could extend to handle the non-kill cases for more coverage. 643 if (MI->getOperand(3).isKill()) { 644 // Add to chain. 645 LLVM_DEBUG(dbgs() << "Instruction was successfully added to chain.\n"); 646 ActiveChains[AccumReg]->add(MI, Idx, getColor(DestReg)); 647 // Handle cases where the destination is not the same as the accumulator. 648 if (DestReg != AccumReg) { 649 ActiveChains[DestReg] = ActiveChains[AccumReg]; 650 ActiveChains.erase(AccumReg); 651 } 652 return; 653 } 654 655 LLVM_DEBUG( 656 dbgs() << "Cannot add to chain because accumulator operand wasn't " 657 << "marked <kill>!\n"); 658 maybeKillChain(MI->getOperand(3), Idx, ActiveChains); 659 } 660 661 LLVM_DEBUG(dbgs() << "Creating new chain for dest register " 662 << printReg(DestReg, TRI) << "\n"); 663 auto G = std::make_unique<Chain>(MI, Idx, getColor(DestReg)); 664 ActiveChains[DestReg] = G.get(); 665 AllChains.push_back(std::move(G)); 666 667 } else { 668 669 // Non-MUL or MLA instruction. Invalidate any chain in the uses or defs 670 // lists. 671 for (auto &I : MI->uses()) 672 maybeKillChain(I, Idx, ActiveChains); 673 for (auto &I : MI->defs()) 674 maybeKillChain(I, Idx, ActiveChains); 675 676 } 677 } 678 679 void AArch64A57FPLoadBalancing:: 680 maybeKillChain(MachineOperand &MO, unsigned Idx, 681 std::map<unsigned, Chain*> &ActiveChains) { 682 // Given an operand and the set of active chains (keyed by register), 683 // determine if a chain should be ended and remove from ActiveChains. 684 MachineInstr *MI = MO.getParent(); 685 686 if (MO.isReg()) { 687 688 // If this is a KILL of a current chain, record it. 689 if (MO.isKill() && ActiveChains.find(MO.getReg()) != ActiveChains.end()) { 690 LLVM_DEBUG(dbgs() << "Kill seen for chain " << printReg(MO.getReg(), TRI) 691 << "\n"); 692 ActiveChains[MO.getReg()]->setKill(MI, Idx, /*Immutable=*/MO.isTied()); 693 } 694 ActiveChains.erase(MO.getReg()); 695 696 } else if (MO.isRegMask()) { 697 698 for (auto I = ActiveChains.begin(), E = ActiveChains.end(); 699 I != E;) { 700 if (MO.clobbersPhysReg(I->first)) { 701 LLVM_DEBUG(dbgs() << "Kill (regmask) seen for chain " 702 << printReg(I->first, TRI) << "\n"); 703 I->second->setKill(MI, Idx, /*Immutable=*/true); 704 ActiveChains.erase(I++); 705 } else 706 ++I; 707 } 708 709 } 710 } 711 712 Color AArch64A57FPLoadBalancing::getColor(unsigned Reg) { 713 if ((TRI->getEncodingValue(Reg) % 2) == 0) 714 return Color::Even; 715 else 716 return Color::Odd; 717 } 718 719 // Factory function used by AArch64TargetMachine to add the pass to the passmanager. 720 FunctionPass *llvm::createAArch64A57FPLoadBalancing() { 721 return new AArch64A57FPLoadBalancing(); 722 } 723