1 //===- AArch64LoadStoreOptimizer.cpp - AArch64 load/store opt. pass -------===// 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 // 9 // This file contains a pass that performs load / store related peephole 10 // optimizations. This pass should be run after register allocation. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "AArch64InstrInfo.h" 15 #include "AArch64Subtarget.h" 16 #include "MCTargetDesc/AArch64AddressingModes.h" 17 #include "llvm/ADT/BitVector.h" 18 #include "llvm/ADT/SmallVector.h" 19 #include "llvm/ADT/Statistic.h" 20 #include "llvm/ADT/StringRef.h" 21 #include "llvm/ADT/iterator_range.h" 22 #include "llvm/Analysis/AliasAnalysis.h" 23 #include "llvm/CodeGen/MachineBasicBlock.h" 24 #include "llvm/CodeGen/MachineFunction.h" 25 #include "llvm/CodeGen/MachineFunctionPass.h" 26 #include "llvm/CodeGen/MachineInstr.h" 27 #include "llvm/CodeGen/MachineInstrBuilder.h" 28 #include "llvm/CodeGen/MachineOperand.h" 29 #include "llvm/CodeGen/MachineRegisterInfo.h" 30 #include "llvm/CodeGen/TargetRegisterInfo.h" 31 #include "llvm/IR/DebugLoc.h" 32 #include "llvm/MC/MCRegisterInfo.h" 33 #include "llvm/Pass.h" 34 #include "llvm/Support/CommandLine.h" 35 #include "llvm/Support/Debug.h" 36 #include "llvm/Support/DebugCounter.h" 37 #include "llvm/Support/ErrorHandling.h" 38 #include "llvm/Support/raw_ostream.h" 39 #include <cassert> 40 #include <cstdint> 41 #include <functional> 42 #include <iterator> 43 #include <limits> 44 45 using namespace llvm; 46 47 #define DEBUG_TYPE "aarch64-ldst-opt" 48 49 STATISTIC(NumPairCreated, "Number of load/store pair instructions generated"); 50 STATISTIC(NumPostFolded, "Number of post-index updates folded"); 51 STATISTIC(NumPreFolded, "Number of pre-index updates folded"); 52 STATISTIC(NumUnscaledPairCreated, 53 "Number of load/store from unscaled generated"); 54 STATISTIC(NumZeroStoresPromoted, "Number of narrow zero stores promoted"); 55 STATISTIC(NumLoadsFromStoresPromoted, "Number of loads from stores promoted"); 56 57 DEBUG_COUNTER(RegRenamingCounter, DEBUG_TYPE "-reg-renaming", 58 "Controls which pairs are considered for renaming"); 59 60 // The LdStLimit limits how far we search for load/store pairs. 61 static cl::opt<unsigned> LdStLimit("aarch64-load-store-scan-limit", 62 cl::init(20), cl::Hidden); 63 64 // The UpdateLimit limits how far we search for update instructions when we form 65 // pre-/post-index instructions. 66 static cl::opt<unsigned> UpdateLimit("aarch64-update-scan-limit", cl::init(100), 67 cl::Hidden); 68 69 // Enable register renaming to find additional store pairing opportunities. 70 static cl::opt<bool> EnableRenaming("aarch64-load-store-renaming", 71 cl::init(false), cl::Hidden); 72 73 #define AARCH64_LOAD_STORE_OPT_NAME "AArch64 load / store optimization pass" 74 75 namespace { 76 77 using LdStPairFlags = struct LdStPairFlags { 78 // If a matching instruction is found, MergeForward is set to true if the 79 // merge is to remove the first instruction and replace the second with 80 // a pair-wise insn, and false if the reverse is true. 81 bool MergeForward = false; 82 83 // SExtIdx gives the index of the result of the load pair that must be 84 // extended. The value of SExtIdx assumes that the paired load produces the 85 // value in this order: (I, returned iterator), i.e., -1 means no value has 86 // to be extended, 0 means I, and 1 means the returned iterator. 87 int SExtIdx = -1; 88 89 // If not none, RenameReg can be used to rename the result register of the 90 // first store in a pair. Currently this only works when merging stores 91 // forward. 92 Optional<MCPhysReg> RenameReg = None; 93 94 LdStPairFlags() = default; 95 96 void setMergeForward(bool V = true) { MergeForward = V; } 97 bool getMergeForward() const { return MergeForward; } 98 99 void setSExtIdx(int V) { SExtIdx = V; } 100 int getSExtIdx() const { return SExtIdx; } 101 102 void setRenameReg(MCPhysReg R) { RenameReg = R; } 103 void clearRenameReg() { RenameReg = None; } 104 Optional<MCPhysReg> getRenameReg() const { return RenameReg; } 105 }; 106 107 struct AArch64LoadStoreOpt : public MachineFunctionPass { 108 static char ID; 109 110 AArch64LoadStoreOpt() : MachineFunctionPass(ID) { 111 initializeAArch64LoadStoreOptPass(*PassRegistry::getPassRegistry()); 112 } 113 114 AliasAnalysis *AA; 115 const AArch64InstrInfo *TII; 116 const TargetRegisterInfo *TRI; 117 const AArch64Subtarget *Subtarget; 118 119 // Track which register units have been modified and used. 120 LiveRegUnits ModifiedRegUnits, UsedRegUnits; 121 LiveRegUnits DefinedInBB; 122 123 void getAnalysisUsage(AnalysisUsage &AU) const override { 124 AU.addRequired<AAResultsWrapperPass>(); 125 MachineFunctionPass::getAnalysisUsage(AU); 126 } 127 128 // Scan the instructions looking for a load/store that can be combined 129 // with the current instruction into a load/store pair. 130 // Return the matching instruction if one is found, else MBB->end(). 131 MachineBasicBlock::iterator findMatchingInsn(MachineBasicBlock::iterator I, 132 LdStPairFlags &Flags, 133 unsigned Limit, 134 bool FindNarrowMerge); 135 136 // Scan the instructions looking for a store that writes to the address from 137 // which the current load instruction reads. Return true if one is found. 138 bool findMatchingStore(MachineBasicBlock::iterator I, unsigned Limit, 139 MachineBasicBlock::iterator &StoreI); 140 141 // Merge the two instructions indicated into a wider narrow store instruction. 142 MachineBasicBlock::iterator 143 mergeNarrowZeroStores(MachineBasicBlock::iterator I, 144 MachineBasicBlock::iterator MergeMI, 145 const LdStPairFlags &Flags); 146 147 // Merge the two instructions indicated into a single pair-wise instruction. 148 MachineBasicBlock::iterator 149 mergePairedInsns(MachineBasicBlock::iterator I, 150 MachineBasicBlock::iterator Paired, 151 const LdStPairFlags &Flags); 152 153 // Promote the load that reads directly from the address stored to. 154 MachineBasicBlock::iterator 155 promoteLoadFromStore(MachineBasicBlock::iterator LoadI, 156 MachineBasicBlock::iterator StoreI); 157 158 // Scan the instruction list to find a base register update that can 159 // be combined with the current instruction (a load or store) using 160 // pre or post indexed addressing with writeback. Scan forwards. 161 MachineBasicBlock::iterator 162 findMatchingUpdateInsnForward(MachineBasicBlock::iterator I, 163 int UnscaledOffset, unsigned Limit); 164 165 // Scan the instruction list to find a base register update that can 166 // be combined with the current instruction (a load or store) using 167 // pre or post indexed addressing with writeback. Scan backwards. 168 MachineBasicBlock::iterator 169 findMatchingUpdateInsnBackward(MachineBasicBlock::iterator I, unsigned Limit); 170 171 // Find an instruction that updates the base register of the ld/st 172 // instruction. 173 bool isMatchingUpdateInsn(MachineInstr &MemMI, MachineInstr &MI, 174 unsigned BaseReg, int Offset); 175 176 // Merge a pre- or post-index base register update into a ld/st instruction. 177 MachineBasicBlock::iterator 178 mergeUpdateInsn(MachineBasicBlock::iterator I, 179 MachineBasicBlock::iterator Update, bool IsPreIdx); 180 181 // Find and merge zero store instructions. 182 bool tryToMergeZeroStInst(MachineBasicBlock::iterator &MBBI); 183 184 // Find and pair ldr/str instructions. 185 bool tryToPairLdStInst(MachineBasicBlock::iterator &MBBI); 186 187 // Find and promote load instructions which read directly from store. 188 bool tryToPromoteLoadFromStore(MachineBasicBlock::iterator &MBBI); 189 190 // Find and merge a base register updates before or after a ld/st instruction. 191 bool tryToMergeLdStUpdate(MachineBasicBlock::iterator &MBBI); 192 193 bool optimizeBlock(MachineBasicBlock &MBB, bool EnableNarrowZeroStOpt); 194 195 bool runOnMachineFunction(MachineFunction &Fn) override; 196 197 MachineFunctionProperties getRequiredProperties() const override { 198 return MachineFunctionProperties().set( 199 MachineFunctionProperties::Property::NoVRegs); 200 } 201 202 StringRef getPassName() const override { return AARCH64_LOAD_STORE_OPT_NAME; } 203 }; 204 205 char AArch64LoadStoreOpt::ID = 0; 206 207 } // end anonymous namespace 208 209 INITIALIZE_PASS(AArch64LoadStoreOpt, "aarch64-ldst-opt", 210 AARCH64_LOAD_STORE_OPT_NAME, false, false) 211 212 static bool isNarrowStore(unsigned Opc) { 213 switch (Opc) { 214 default: 215 return false; 216 case AArch64::STRBBui: 217 case AArch64::STURBBi: 218 case AArch64::STRHHui: 219 case AArch64::STURHHi: 220 return true; 221 } 222 } 223 224 // These instruction set memory tag and either keep memory contents unchanged or 225 // set it to zero, ignoring the address part of the source register. 226 static bool isTagStore(const MachineInstr &MI) { 227 switch (MI.getOpcode()) { 228 default: 229 return false; 230 case AArch64::STGOffset: 231 case AArch64::STZGOffset: 232 case AArch64::ST2GOffset: 233 case AArch64::STZ2GOffset: 234 return true; 235 } 236 } 237 238 static unsigned getMatchingNonSExtOpcode(unsigned Opc, 239 bool *IsValidLdStrOpc = nullptr) { 240 if (IsValidLdStrOpc) 241 *IsValidLdStrOpc = true; 242 switch (Opc) { 243 default: 244 if (IsValidLdStrOpc) 245 *IsValidLdStrOpc = false; 246 return std::numeric_limits<unsigned>::max(); 247 case AArch64::STRDui: 248 case AArch64::STURDi: 249 case AArch64::STRQui: 250 case AArch64::STURQi: 251 case AArch64::STRBBui: 252 case AArch64::STURBBi: 253 case AArch64::STRHHui: 254 case AArch64::STURHHi: 255 case AArch64::STRWui: 256 case AArch64::STURWi: 257 case AArch64::STRXui: 258 case AArch64::STURXi: 259 case AArch64::LDRDui: 260 case AArch64::LDURDi: 261 case AArch64::LDRQui: 262 case AArch64::LDURQi: 263 case AArch64::LDRWui: 264 case AArch64::LDURWi: 265 case AArch64::LDRXui: 266 case AArch64::LDURXi: 267 case AArch64::STRSui: 268 case AArch64::STURSi: 269 case AArch64::LDRSui: 270 case AArch64::LDURSi: 271 return Opc; 272 case AArch64::LDRSWui: 273 return AArch64::LDRWui; 274 case AArch64::LDURSWi: 275 return AArch64::LDURWi; 276 } 277 } 278 279 static unsigned getMatchingWideOpcode(unsigned Opc) { 280 switch (Opc) { 281 default: 282 llvm_unreachable("Opcode has no wide equivalent!"); 283 case AArch64::STRBBui: 284 return AArch64::STRHHui; 285 case AArch64::STRHHui: 286 return AArch64::STRWui; 287 case AArch64::STURBBi: 288 return AArch64::STURHHi; 289 case AArch64::STURHHi: 290 return AArch64::STURWi; 291 case AArch64::STURWi: 292 return AArch64::STURXi; 293 case AArch64::STRWui: 294 return AArch64::STRXui; 295 } 296 } 297 298 static unsigned getMatchingPairOpcode(unsigned Opc) { 299 switch (Opc) { 300 default: 301 llvm_unreachable("Opcode has no pairwise equivalent!"); 302 case AArch64::STRSui: 303 case AArch64::STURSi: 304 return AArch64::STPSi; 305 case AArch64::STRDui: 306 case AArch64::STURDi: 307 return AArch64::STPDi; 308 case AArch64::STRQui: 309 case AArch64::STURQi: 310 return AArch64::STPQi; 311 case AArch64::STRWui: 312 case AArch64::STURWi: 313 return AArch64::STPWi; 314 case AArch64::STRXui: 315 case AArch64::STURXi: 316 return AArch64::STPXi; 317 case AArch64::LDRSui: 318 case AArch64::LDURSi: 319 return AArch64::LDPSi; 320 case AArch64::LDRDui: 321 case AArch64::LDURDi: 322 return AArch64::LDPDi; 323 case AArch64::LDRQui: 324 case AArch64::LDURQi: 325 return AArch64::LDPQi; 326 case AArch64::LDRWui: 327 case AArch64::LDURWi: 328 return AArch64::LDPWi; 329 case AArch64::LDRXui: 330 case AArch64::LDURXi: 331 return AArch64::LDPXi; 332 case AArch64::LDRSWui: 333 case AArch64::LDURSWi: 334 return AArch64::LDPSWi; 335 } 336 } 337 338 static unsigned isMatchingStore(MachineInstr &LoadInst, 339 MachineInstr &StoreInst) { 340 unsigned LdOpc = LoadInst.getOpcode(); 341 unsigned StOpc = StoreInst.getOpcode(); 342 switch (LdOpc) { 343 default: 344 llvm_unreachable("Unsupported load instruction!"); 345 case AArch64::LDRBBui: 346 return StOpc == AArch64::STRBBui || StOpc == AArch64::STRHHui || 347 StOpc == AArch64::STRWui || StOpc == AArch64::STRXui; 348 case AArch64::LDURBBi: 349 return StOpc == AArch64::STURBBi || StOpc == AArch64::STURHHi || 350 StOpc == AArch64::STURWi || StOpc == AArch64::STURXi; 351 case AArch64::LDRHHui: 352 return StOpc == AArch64::STRHHui || StOpc == AArch64::STRWui || 353 StOpc == AArch64::STRXui; 354 case AArch64::LDURHHi: 355 return StOpc == AArch64::STURHHi || StOpc == AArch64::STURWi || 356 StOpc == AArch64::STURXi; 357 case AArch64::LDRWui: 358 return StOpc == AArch64::STRWui || StOpc == AArch64::STRXui; 359 case AArch64::LDURWi: 360 return StOpc == AArch64::STURWi || StOpc == AArch64::STURXi; 361 case AArch64::LDRXui: 362 return StOpc == AArch64::STRXui; 363 case AArch64::LDURXi: 364 return StOpc == AArch64::STURXi; 365 } 366 } 367 368 static unsigned getPreIndexedOpcode(unsigned Opc) { 369 // FIXME: We don't currently support creating pre-indexed loads/stores when 370 // the load or store is the unscaled version. If we decide to perform such an 371 // optimization in the future the cases for the unscaled loads/stores will 372 // need to be added here. 373 switch (Opc) { 374 default: 375 llvm_unreachable("Opcode has no pre-indexed equivalent!"); 376 case AArch64::STRSui: 377 return AArch64::STRSpre; 378 case AArch64::STRDui: 379 return AArch64::STRDpre; 380 case AArch64::STRQui: 381 return AArch64::STRQpre; 382 case AArch64::STRBBui: 383 return AArch64::STRBBpre; 384 case AArch64::STRHHui: 385 return AArch64::STRHHpre; 386 case AArch64::STRWui: 387 return AArch64::STRWpre; 388 case AArch64::STRXui: 389 return AArch64::STRXpre; 390 case AArch64::LDRSui: 391 return AArch64::LDRSpre; 392 case AArch64::LDRDui: 393 return AArch64::LDRDpre; 394 case AArch64::LDRQui: 395 return AArch64::LDRQpre; 396 case AArch64::LDRBBui: 397 return AArch64::LDRBBpre; 398 case AArch64::LDRHHui: 399 return AArch64::LDRHHpre; 400 case AArch64::LDRWui: 401 return AArch64::LDRWpre; 402 case AArch64::LDRXui: 403 return AArch64::LDRXpre; 404 case AArch64::LDRSWui: 405 return AArch64::LDRSWpre; 406 case AArch64::LDPSi: 407 return AArch64::LDPSpre; 408 case AArch64::LDPSWi: 409 return AArch64::LDPSWpre; 410 case AArch64::LDPDi: 411 return AArch64::LDPDpre; 412 case AArch64::LDPQi: 413 return AArch64::LDPQpre; 414 case AArch64::LDPWi: 415 return AArch64::LDPWpre; 416 case AArch64::LDPXi: 417 return AArch64::LDPXpre; 418 case AArch64::STPSi: 419 return AArch64::STPSpre; 420 case AArch64::STPDi: 421 return AArch64::STPDpre; 422 case AArch64::STPQi: 423 return AArch64::STPQpre; 424 case AArch64::STPWi: 425 return AArch64::STPWpre; 426 case AArch64::STPXi: 427 return AArch64::STPXpre; 428 case AArch64::STGOffset: 429 return AArch64::STGPreIndex; 430 case AArch64::STZGOffset: 431 return AArch64::STZGPreIndex; 432 case AArch64::ST2GOffset: 433 return AArch64::ST2GPreIndex; 434 case AArch64::STZ2GOffset: 435 return AArch64::STZ2GPreIndex; 436 case AArch64::STGPi: 437 return AArch64::STGPpre; 438 } 439 } 440 441 static unsigned getPostIndexedOpcode(unsigned Opc) { 442 switch (Opc) { 443 default: 444 llvm_unreachable("Opcode has no post-indexed wise equivalent!"); 445 case AArch64::STRSui: 446 case AArch64::STURSi: 447 return AArch64::STRSpost; 448 case AArch64::STRDui: 449 case AArch64::STURDi: 450 return AArch64::STRDpost; 451 case AArch64::STRQui: 452 case AArch64::STURQi: 453 return AArch64::STRQpost; 454 case AArch64::STRBBui: 455 return AArch64::STRBBpost; 456 case AArch64::STRHHui: 457 return AArch64::STRHHpost; 458 case AArch64::STRWui: 459 case AArch64::STURWi: 460 return AArch64::STRWpost; 461 case AArch64::STRXui: 462 case AArch64::STURXi: 463 return AArch64::STRXpost; 464 case AArch64::LDRSui: 465 case AArch64::LDURSi: 466 return AArch64::LDRSpost; 467 case AArch64::LDRDui: 468 case AArch64::LDURDi: 469 return AArch64::LDRDpost; 470 case AArch64::LDRQui: 471 case AArch64::LDURQi: 472 return AArch64::LDRQpost; 473 case AArch64::LDRBBui: 474 return AArch64::LDRBBpost; 475 case AArch64::LDRHHui: 476 return AArch64::LDRHHpost; 477 case AArch64::LDRWui: 478 case AArch64::LDURWi: 479 return AArch64::LDRWpost; 480 case AArch64::LDRXui: 481 case AArch64::LDURXi: 482 return AArch64::LDRXpost; 483 case AArch64::LDRSWui: 484 return AArch64::LDRSWpost; 485 case AArch64::LDPSi: 486 return AArch64::LDPSpost; 487 case AArch64::LDPSWi: 488 return AArch64::LDPSWpost; 489 case AArch64::LDPDi: 490 return AArch64::LDPDpost; 491 case AArch64::LDPQi: 492 return AArch64::LDPQpost; 493 case AArch64::LDPWi: 494 return AArch64::LDPWpost; 495 case AArch64::LDPXi: 496 return AArch64::LDPXpost; 497 case AArch64::STPSi: 498 return AArch64::STPSpost; 499 case AArch64::STPDi: 500 return AArch64::STPDpost; 501 case AArch64::STPQi: 502 return AArch64::STPQpost; 503 case AArch64::STPWi: 504 return AArch64::STPWpost; 505 case AArch64::STPXi: 506 return AArch64::STPXpost; 507 case AArch64::STGOffset: 508 return AArch64::STGPostIndex; 509 case AArch64::STZGOffset: 510 return AArch64::STZGPostIndex; 511 case AArch64::ST2GOffset: 512 return AArch64::ST2GPostIndex; 513 case AArch64::STZ2GOffset: 514 return AArch64::STZ2GPostIndex; 515 case AArch64::STGPi: 516 return AArch64::STGPpost; 517 } 518 } 519 520 static bool isPairedLdSt(const MachineInstr &MI) { 521 switch (MI.getOpcode()) { 522 default: 523 return false; 524 case AArch64::LDPSi: 525 case AArch64::LDPSWi: 526 case AArch64::LDPDi: 527 case AArch64::LDPQi: 528 case AArch64::LDPWi: 529 case AArch64::LDPXi: 530 case AArch64::STPSi: 531 case AArch64::STPDi: 532 case AArch64::STPQi: 533 case AArch64::STPWi: 534 case AArch64::STPXi: 535 case AArch64::STGPi: 536 return true; 537 } 538 } 539 540 // Returns the scale and offset range of pre/post indexed variants of MI. 541 static void getPrePostIndexedMemOpInfo(const MachineInstr &MI, int &Scale, 542 int &MinOffset, int &MaxOffset) { 543 bool IsPaired = isPairedLdSt(MI); 544 bool IsTagStore = isTagStore(MI); 545 // ST*G and all paired ldst have the same scale in pre/post-indexed variants 546 // as in the "unsigned offset" variant. 547 // All other pre/post indexed ldst instructions are unscaled. 548 Scale = (IsTagStore || IsPaired) ? AArch64InstrInfo::getMemScale(MI) : 1; 549 550 if (IsPaired) { 551 MinOffset = -64; 552 MaxOffset = 63; 553 } else { 554 MinOffset = -256; 555 MaxOffset = 255; 556 } 557 } 558 559 static MachineOperand &getLdStRegOp(MachineInstr &MI, 560 unsigned PairedRegOp = 0) { 561 assert(PairedRegOp < 2 && "Unexpected register operand idx."); 562 unsigned Idx = isPairedLdSt(MI) ? PairedRegOp : 0; 563 return MI.getOperand(Idx); 564 } 565 566 static const MachineOperand &getLdStBaseOp(const MachineInstr &MI) { 567 unsigned Idx = isPairedLdSt(MI) ? 2 : 1; 568 return MI.getOperand(Idx); 569 } 570 571 static const MachineOperand &getLdStOffsetOp(const MachineInstr &MI) { 572 unsigned Idx = isPairedLdSt(MI) ? 3 : 2; 573 return MI.getOperand(Idx); 574 } 575 576 static bool isLdOffsetInRangeOfSt(MachineInstr &LoadInst, 577 MachineInstr &StoreInst, 578 const AArch64InstrInfo *TII) { 579 assert(isMatchingStore(LoadInst, StoreInst) && "Expect only matched ld/st."); 580 int LoadSize = TII->getMemScale(LoadInst); 581 int StoreSize = TII->getMemScale(StoreInst); 582 int UnscaledStOffset = TII->isUnscaledLdSt(StoreInst) 583 ? getLdStOffsetOp(StoreInst).getImm() 584 : getLdStOffsetOp(StoreInst).getImm() * StoreSize; 585 int UnscaledLdOffset = TII->isUnscaledLdSt(LoadInst) 586 ? getLdStOffsetOp(LoadInst).getImm() 587 : getLdStOffsetOp(LoadInst).getImm() * LoadSize; 588 return (UnscaledStOffset <= UnscaledLdOffset) && 589 (UnscaledLdOffset + LoadSize <= (UnscaledStOffset + StoreSize)); 590 } 591 592 static bool isPromotableZeroStoreInst(MachineInstr &MI) { 593 unsigned Opc = MI.getOpcode(); 594 return (Opc == AArch64::STRWui || Opc == AArch64::STURWi || 595 isNarrowStore(Opc)) && 596 getLdStRegOp(MI).getReg() == AArch64::WZR; 597 } 598 599 static bool isPromotableLoadFromStore(MachineInstr &MI) { 600 switch (MI.getOpcode()) { 601 default: 602 return false; 603 // Scaled instructions. 604 case AArch64::LDRBBui: 605 case AArch64::LDRHHui: 606 case AArch64::LDRWui: 607 case AArch64::LDRXui: 608 // Unscaled instructions. 609 case AArch64::LDURBBi: 610 case AArch64::LDURHHi: 611 case AArch64::LDURWi: 612 case AArch64::LDURXi: 613 return true; 614 } 615 } 616 617 static bool isMergeableLdStUpdate(MachineInstr &MI) { 618 unsigned Opc = MI.getOpcode(); 619 switch (Opc) { 620 default: 621 return false; 622 // Scaled instructions. 623 case AArch64::STRSui: 624 case AArch64::STRDui: 625 case AArch64::STRQui: 626 case AArch64::STRXui: 627 case AArch64::STRWui: 628 case AArch64::STRHHui: 629 case AArch64::STRBBui: 630 case AArch64::LDRSui: 631 case AArch64::LDRDui: 632 case AArch64::LDRQui: 633 case AArch64::LDRXui: 634 case AArch64::LDRWui: 635 case AArch64::LDRHHui: 636 case AArch64::LDRBBui: 637 case AArch64::STGOffset: 638 case AArch64::STZGOffset: 639 case AArch64::ST2GOffset: 640 case AArch64::STZ2GOffset: 641 case AArch64::STGPi: 642 // Unscaled instructions. 643 case AArch64::STURSi: 644 case AArch64::STURDi: 645 case AArch64::STURQi: 646 case AArch64::STURWi: 647 case AArch64::STURXi: 648 case AArch64::LDURSi: 649 case AArch64::LDURDi: 650 case AArch64::LDURQi: 651 case AArch64::LDURWi: 652 case AArch64::LDURXi: 653 // Paired instructions. 654 case AArch64::LDPSi: 655 case AArch64::LDPSWi: 656 case AArch64::LDPDi: 657 case AArch64::LDPQi: 658 case AArch64::LDPWi: 659 case AArch64::LDPXi: 660 case AArch64::STPSi: 661 case AArch64::STPDi: 662 case AArch64::STPQi: 663 case AArch64::STPWi: 664 case AArch64::STPXi: 665 // Make sure this is a reg+imm (as opposed to an address reloc). 666 if (!getLdStOffsetOp(MI).isImm()) 667 return false; 668 669 return true; 670 } 671 } 672 673 MachineBasicBlock::iterator 674 AArch64LoadStoreOpt::mergeNarrowZeroStores(MachineBasicBlock::iterator I, 675 MachineBasicBlock::iterator MergeMI, 676 const LdStPairFlags &Flags) { 677 assert(isPromotableZeroStoreInst(*I) && isPromotableZeroStoreInst(*MergeMI) && 678 "Expected promotable zero stores."); 679 680 MachineBasicBlock::iterator NextI = I; 681 ++NextI; 682 // If NextI is the second of the two instructions to be merged, we need 683 // to skip one further. Either way we merge will invalidate the iterator, 684 // and we don't need to scan the new instruction, as it's a pairwise 685 // instruction, which we're not considering for further action anyway. 686 if (NextI == MergeMI) 687 ++NextI; 688 689 unsigned Opc = I->getOpcode(); 690 bool IsScaled = !TII->isUnscaledLdSt(Opc); 691 int OffsetStride = IsScaled ? 1 : TII->getMemScale(*I); 692 693 bool MergeForward = Flags.getMergeForward(); 694 // Insert our new paired instruction after whichever of the paired 695 // instructions MergeForward indicates. 696 MachineBasicBlock::iterator InsertionPoint = MergeForward ? MergeMI : I; 697 // Also based on MergeForward is from where we copy the base register operand 698 // so we get the flags compatible with the input code. 699 const MachineOperand &BaseRegOp = 700 MergeForward ? getLdStBaseOp(*MergeMI) : getLdStBaseOp(*I); 701 702 // Which register is Rt and which is Rt2 depends on the offset order. 703 MachineInstr *RtMI; 704 if (getLdStOffsetOp(*I).getImm() == 705 getLdStOffsetOp(*MergeMI).getImm() + OffsetStride) 706 RtMI = &*MergeMI; 707 else 708 RtMI = &*I; 709 710 int OffsetImm = getLdStOffsetOp(*RtMI).getImm(); 711 // Change the scaled offset from small to large type. 712 if (IsScaled) { 713 assert(((OffsetImm & 1) == 0) && "Unexpected offset to merge"); 714 OffsetImm /= 2; 715 } 716 717 // Construct the new instruction. 718 DebugLoc DL = I->getDebugLoc(); 719 MachineBasicBlock *MBB = I->getParent(); 720 MachineInstrBuilder MIB; 721 MIB = BuildMI(*MBB, InsertionPoint, DL, TII->get(getMatchingWideOpcode(Opc))) 722 .addReg(isNarrowStore(Opc) ? AArch64::WZR : AArch64::XZR) 723 .add(BaseRegOp) 724 .addImm(OffsetImm) 725 .cloneMergedMemRefs({&*I, &*MergeMI}) 726 .setMIFlags(I->mergeFlagsWith(*MergeMI)); 727 (void)MIB; 728 729 LLVM_DEBUG(dbgs() << "Creating wider store. Replacing instructions:\n "); 730 LLVM_DEBUG(I->print(dbgs())); 731 LLVM_DEBUG(dbgs() << " "); 732 LLVM_DEBUG(MergeMI->print(dbgs())); 733 LLVM_DEBUG(dbgs() << " with instruction:\n "); 734 LLVM_DEBUG(((MachineInstr *)MIB)->print(dbgs())); 735 LLVM_DEBUG(dbgs() << "\n"); 736 737 // Erase the old instructions. 738 I->eraseFromParent(); 739 MergeMI->eraseFromParent(); 740 return NextI; 741 } 742 743 // Apply Fn to all instructions between MI and the beginning of the block, until 744 // a def for DefReg is reached. Returns true, iff Fn returns true for all 745 // visited instructions. Stop after visiting Limit iterations. 746 static bool forAllMIsUntilDef(MachineInstr &MI, MCPhysReg DefReg, 747 const TargetRegisterInfo *TRI, unsigned Limit, 748 std::function<bool(MachineInstr &, bool)> &Fn) { 749 auto MBB = MI.getParent(); 750 for (MachineBasicBlock::reverse_iterator I = MI.getReverseIterator(), 751 E = MBB->rend(); 752 I != E; I++) { 753 if (!Limit) 754 return false; 755 --Limit; 756 757 bool isDef = any_of(I->operands(), [DefReg, TRI](MachineOperand &MOP) { 758 return MOP.isReg() && MOP.isDef() && !MOP.isDebug() && MOP.getReg() && 759 TRI->regsOverlap(MOP.getReg(), DefReg); 760 }); 761 if (!Fn(*I, isDef)) 762 return false; 763 if (isDef) 764 break; 765 } 766 return true; 767 } 768 769 static void updateDefinedRegisters(MachineInstr &MI, LiveRegUnits &Units, 770 const TargetRegisterInfo *TRI) { 771 772 for (const MachineOperand &MOP : phys_regs_and_masks(MI)) 773 if (MOP.isReg() && MOP.isKill()) 774 Units.removeReg(MOP.getReg()); 775 776 for (const MachineOperand &MOP : phys_regs_and_masks(MI)) 777 if (MOP.isReg() && !MOP.isKill()) 778 Units.addReg(MOP.getReg()); 779 } 780 781 MachineBasicBlock::iterator 782 AArch64LoadStoreOpt::mergePairedInsns(MachineBasicBlock::iterator I, 783 MachineBasicBlock::iterator Paired, 784 const LdStPairFlags &Flags) { 785 MachineBasicBlock::iterator NextI = I; 786 ++NextI; 787 // If NextI is the second of the two instructions to be merged, we need 788 // to skip one further. Either way we merge will invalidate the iterator, 789 // and we don't need to scan the new instruction, as it's a pairwise 790 // instruction, which we're not considering for further action anyway. 791 if (NextI == Paired) 792 ++NextI; 793 794 int SExtIdx = Flags.getSExtIdx(); 795 unsigned Opc = 796 SExtIdx == -1 ? I->getOpcode() : getMatchingNonSExtOpcode(I->getOpcode()); 797 bool IsUnscaled = TII->isUnscaledLdSt(Opc); 798 int OffsetStride = IsUnscaled ? TII->getMemScale(*I) : 1; 799 800 bool MergeForward = Flags.getMergeForward(); 801 802 Optional<MCPhysReg> RenameReg = Flags.getRenameReg(); 803 if (MergeForward && RenameReg) { 804 MCRegister RegToRename = getLdStRegOp(*I).getReg(); 805 DefinedInBB.addReg(*RenameReg); 806 807 // Return the sub/super register for RenameReg, matching the size of 808 // OriginalReg. 809 auto GetMatchingSubReg = [this, 810 RenameReg](MCPhysReg OriginalReg) -> MCPhysReg { 811 for (MCPhysReg SubOrSuper : TRI->sub_and_superregs_inclusive(*RenameReg)) 812 if (TRI->getMinimalPhysRegClass(OriginalReg) == 813 TRI->getMinimalPhysRegClass(SubOrSuper)) 814 return SubOrSuper; 815 llvm_unreachable("Should have found matching sub or super register!"); 816 }; 817 818 std::function<bool(MachineInstr &, bool)> UpdateMIs = 819 [this, RegToRename, GetMatchingSubReg](MachineInstr &MI, bool IsDef) { 820 if (IsDef) { 821 bool SeenDef = false; 822 for (auto &MOP : MI.operands()) { 823 // Rename the first explicit definition and all implicit 824 // definitions matching RegToRename. 825 if (MOP.isReg() && !MOP.isDebug() && MOP.getReg() && 826 (!SeenDef || (MOP.isDef() && MOP.isImplicit())) && 827 TRI->regsOverlap(MOP.getReg(), RegToRename)) { 828 assert((MOP.isImplicit() || 829 (MOP.isRenamable() && !MOP.isEarlyClobber())) && 830 "Need renamable operands"); 831 MOP.setReg(GetMatchingSubReg(MOP.getReg())); 832 SeenDef = true; 833 } 834 } 835 } else { 836 for (auto &MOP : MI.operands()) { 837 if (MOP.isReg() && !MOP.isDebug() && MOP.getReg() && 838 TRI->regsOverlap(MOP.getReg(), RegToRename)) { 839 assert((MOP.isImplicit() || 840 (MOP.isRenamable() && !MOP.isEarlyClobber())) && 841 "Need renamable operands"); 842 MOP.setReg(GetMatchingSubReg(MOP.getReg())); 843 } 844 } 845 } 846 LLVM_DEBUG(dbgs() << "Renamed " << MI << "\n"); 847 return true; 848 }; 849 forAllMIsUntilDef(*I, RegToRename, TRI, LdStLimit, UpdateMIs); 850 851 #if !defined(NDEBUG) 852 // Make sure the register used for renaming is not used between the paired 853 // instructions. That would trash the content before the new paired 854 // instruction. 855 for (auto &MI : 856 iterator_range<MachineInstrBundleIterator<llvm::MachineInstr>>( 857 std::next(I), std::next(Paired))) 858 assert(all_of(MI.operands(), 859 [this, &RenameReg](const MachineOperand &MOP) { 860 return !MOP.isReg() || MOP.isDebug() || !MOP.getReg() || 861 !TRI->regsOverlap(MOP.getReg(), *RenameReg); 862 }) && 863 "Rename register used between paired instruction, trashing the " 864 "content"); 865 #endif 866 } 867 868 // Insert our new paired instruction after whichever of the paired 869 // instructions MergeForward indicates. 870 MachineBasicBlock::iterator InsertionPoint = MergeForward ? Paired : I; 871 // Also based on MergeForward is from where we copy the base register operand 872 // so we get the flags compatible with the input code. 873 const MachineOperand &BaseRegOp = 874 MergeForward ? getLdStBaseOp(*Paired) : getLdStBaseOp(*I); 875 876 int Offset = getLdStOffsetOp(*I).getImm(); 877 int PairedOffset = getLdStOffsetOp(*Paired).getImm(); 878 bool PairedIsUnscaled = TII->isUnscaledLdSt(Paired->getOpcode()); 879 if (IsUnscaled != PairedIsUnscaled) { 880 // We're trying to pair instructions that differ in how they are scaled. If 881 // I is scaled then scale the offset of Paired accordingly. Otherwise, do 882 // the opposite (i.e., make Paired's offset unscaled). 883 int MemSize = TII->getMemScale(*Paired); 884 if (PairedIsUnscaled) { 885 // If the unscaled offset isn't a multiple of the MemSize, we can't 886 // pair the operations together. 887 assert(!(PairedOffset % TII->getMemScale(*Paired)) && 888 "Offset should be a multiple of the stride!"); 889 PairedOffset /= MemSize; 890 } else { 891 PairedOffset *= MemSize; 892 } 893 } 894 895 // Which register is Rt and which is Rt2 depends on the offset order. 896 MachineInstr *RtMI, *Rt2MI; 897 if (Offset == PairedOffset + OffsetStride) { 898 RtMI = &*Paired; 899 Rt2MI = &*I; 900 // Here we swapped the assumption made for SExtIdx. 901 // I.e., we turn ldp I, Paired into ldp Paired, I. 902 // Update the index accordingly. 903 if (SExtIdx != -1) 904 SExtIdx = (SExtIdx + 1) % 2; 905 } else { 906 RtMI = &*I; 907 Rt2MI = &*Paired; 908 } 909 int OffsetImm = getLdStOffsetOp(*RtMI).getImm(); 910 // Scale the immediate offset, if necessary. 911 if (TII->isUnscaledLdSt(RtMI->getOpcode())) { 912 assert(!(OffsetImm % TII->getMemScale(*RtMI)) && 913 "Unscaled offset cannot be scaled."); 914 OffsetImm /= TII->getMemScale(*RtMI); 915 } 916 917 // Construct the new instruction. 918 MachineInstrBuilder MIB; 919 DebugLoc DL = I->getDebugLoc(); 920 MachineBasicBlock *MBB = I->getParent(); 921 MachineOperand RegOp0 = getLdStRegOp(*RtMI); 922 MachineOperand RegOp1 = getLdStRegOp(*Rt2MI); 923 // Kill flags may become invalid when moving stores for pairing. 924 if (RegOp0.isUse()) { 925 if (!MergeForward) { 926 // Clear kill flags on store if moving upwards. Example: 927 // STRWui %w0, ... 928 // USE %w1 929 // STRWui kill %w1 ; need to clear kill flag when moving STRWui upwards 930 RegOp0.setIsKill(false); 931 RegOp1.setIsKill(false); 932 } else { 933 // Clear kill flags of the first stores register. Example: 934 // STRWui %w1, ... 935 // USE kill %w1 ; need to clear kill flag when moving STRWui downwards 936 // STRW %w0 937 Register Reg = getLdStRegOp(*I).getReg(); 938 for (MachineInstr &MI : make_range(std::next(I), Paired)) 939 MI.clearRegisterKills(Reg, TRI); 940 } 941 } 942 MIB = BuildMI(*MBB, InsertionPoint, DL, TII->get(getMatchingPairOpcode(Opc))) 943 .add(RegOp0) 944 .add(RegOp1) 945 .add(BaseRegOp) 946 .addImm(OffsetImm) 947 .cloneMergedMemRefs({&*I, &*Paired}) 948 .setMIFlags(I->mergeFlagsWith(*Paired)); 949 950 (void)MIB; 951 952 LLVM_DEBUG( 953 dbgs() << "Creating pair load/store. Replacing instructions:\n "); 954 LLVM_DEBUG(I->print(dbgs())); 955 LLVM_DEBUG(dbgs() << " "); 956 LLVM_DEBUG(Paired->print(dbgs())); 957 LLVM_DEBUG(dbgs() << " with instruction:\n "); 958 if (SExtIdx != -1) { 959 // Generate the sign extension for the proper result of the ldp. 960 // I.e., with X1, that would be: 961 // %w1 = KILL %w1, implicit-def %x1 962 // %x1 = SBFMXri killed %x1, 0, 31 963 MachineOperand &DstMO = MIB->getOperand(SExtIdx); 964 // Right now, DstMO has the extended register, since it comes from an 965 // extended opcode. 966 Register DstRegX = DstMO.getReg(); 967 // Get the W variant of that register. 968 Register DstRegW = TRI->getSubReg(DstRegX, AArch64::sub_32); 969 // Update the result of LDP to use the W instead of the X variant. 970 DstMO.setReg(DstRegW); 971 LLVM_DEBUG(((MachineInstr *)MIB)->print(dbgs())); 972 LLVM_DEBUG(dbgs() << "\n"); 973 // Make the machine verifier happy by providing a definition for 974 // the X register. 975 // Insert this definition right after the generated LDP, i.e., before 976 // InsertionPoint. 977 MachineInstrBuilder MIBKill = 978 BuildMI(*MBB, InsertionPoint, DL, TII->get(TargetOpcode::KILL), DstRegW) 979 .addReg(DstRegW) 980 .addReg(DstRegX, RegState::Define); 981 MIBKill->getOperand(2).setImplicit(); 982 // Create the sign extension. 983 MachineInstrBuilder MIBSXTW = 984 BuildMI(*MBB, InsertionPoint, DL, TII->get(AArch64::SBFMXri), DstRegX) 985 .addReg(DstRegX) 986 .addImm(0) 987 .addImm(31); 988 (void)MIBSXTW; 989 LLVM_DEBUG(dbgs() << " Extend operand:\n "); 990 LLVM_DEBUG(((MachineInstr *)MIBSXTW)->print(dbgs())); 991 } else { 992 LLVM_DEBUG(((MachineInstr *)MIB)->print(dbgs())); 993 } 994 LLVM_DEBUG(dbgs() << "\n"); 995 996 if (MergeForward) 997 for (const MachineOperand &MOP : phys_regs_and_masks(*I)) 998 if (MOP.isReg() && MOP.isKill()) 999 DefinedInBB.addReg(MOP.getReg()); 1000 1001 // Erase the old instructions. 1002 I->eraseFromParent(); 1003 Paired->eraseFromParent(); 1004 1005 return NextI; 1006 } 1007 1008 MachineBasicBlock::iterator 1009 AArch64LoadStoreOpt::promoteLoadFromStore(MachineBasicBlock::iterator LoadI, 1010 MachineBasicBlock::iterator StoreI) { 1011 MachineBasicBlock::iterator NextI = LoadI; 1012 ++NextI; 1013 1014 int LoadSize = TII->getMemScale(*LoadI); 1015 int StoreSize = TII->getMemScale(*StoreI); 1016 Register LdRt = getLdStRegOp(*LoadI).getReg(); 1017 const MachineOperand &StMO = getLdStRegOp(*StoreI); 1018 Register StRt = getLdStRegOp(*StoreI).getReg(); 1019 bool IsStoreXReg = TRI->getRegClass(AArch64::GPR64RegClassID)->contains(StRt); 1020 1021 assert((IsStoreXReg || 1022 TRI->getRegClass(AArch64::GPR32RegClassID)->contains(StRt)) && 1023 "Unexpected RegClass"); 1024 1025 MachineInstr *BitExtMI; 1026 if (LoadSize == StoreSize && (LoadSize == 4 || LoadSize == 8)) { 1027 // Remove the load, if the destination register of the loads is the same 1028 // register for stored value. 1029 if (StRt == LdRt && LoadSize == 8) { 1030 for (MachineInstr &MI : make_range(StoreI->getIterator(), 1031 LoadI->getIterator())) { 1032 if (MI.killsRegister(StRt, TRI)) { 1033 MI.clearRegisterKills(StRt, TRI); 1034 break; 1035 } 1036 } 1037 LLVM_DEBUG(dbgs() << "Remove load instruction:\n "); 1038 LLVM_DEBUG(LoadI->print(dbgs())); 1039 LLVM_DEBUG(dbgs() << "\n"); 1040 LoadI->eraseFromParent(); 1041 return NextI; 1042 } 1043 // Replace the load with a mov if the load and store are in the same size. 1044 BitExtMI = 1045 BuildMI(*LoadI->getParent(), LoadI, LoadI->getDebugLoc(), 1046 TII->get(IsStoreXReg ? AArch64::ORRXrs : AArch64::ORRWrs), LdRt) 1047 .addReg(IsStoreXReg ? AArch64::XZR : AArch64::WZR) 1048 .add(StMO) 1049 .addImm(AArch64_AM::getShifterImm(AArch64_AM::LSL, 0)) 1050 .setMIFlags(LoadI->getFlags()); 1051 } else { 1052 // FIXME: Currently we disable this transformation in big-endian targets as 1053 // performance and correctness are verified only in little-endian. 1054 if (!Subtarget->isLittleEndian()) 1055 return NextI; 1056 bool IsUnscaled = TII->isUnscaledLdSt(*LoadI); 1057 assert(IsUnscaled == TII->isUnscaledLdSt(*StoreI) && 1058 "Unsupported ld/st match"); 1059 assert(LoadSize <= StoreSize && "Invalid load size"); 1060 int UnscaledLdOffset = IsUnscaled 1061 ? getLdStOffsetOp(*LoadI).getImm() 1062 : getLdStOffsetOp(*LoadI).getImm() * LoadSize; 1063 int UnscaledStOffset = IsUnscaled 1064 ? getLdStOffsetOp(*StoreI).getImm() 1065 : getLdStOffsetOp(*StoreI).getImm() * StoreSize; 1066 int Width = LoadSize * 8; 1067 unsigned DestReg = 1068 IsStoreXReg ? Register(TRI->getMatchingSuperReg( 1069 LdRt, AArch64::sub_32, &AArch64::GPR64RegClass)) 1070 : LdRt; 1071 1072 assert((UnscaledLdOffset >= UnscaledStOffset && 1073 (UnscaledLdOffset + LoadSize) <= UnscaledStOffset + StoreSize) && 1074 "Invalid offset"); 1075 1076 int Immr = 8 * (UnscaledLdOffset - UnscaledStOffset); 1077 int Imms = Immr + Width - 1; 1078 if (UnscaledLdOffset == UnscaledStOffset) { 1079 uint32_t AndMaskEncoded = ((IsStoreXReg ? 1 : 0) << 12) // N 1080 | ((Immr) << 6) // immr 1081 | ((Imms) << 0) // imms 1082 ; 1083 1084 BitExtMI = 1085 BuildMI(*LoadI->getParent(), LoadI, LoadI->getDebugLoc(), 1086 TII->get(IsStoreXReg ? AArch64::ANDXri : AArch64::ANDWri), 1087 DestReg) 1088 .add(StMO) 1089 .addImm(AndMaskEncoded) 1090 .setMIFlags(LoadI->getFlags()); 1091 } else { 1092 BitExtMI = 1093 BuildMI(*LoadI->getParent(), LoadI, LoadI->getDebugLoc(), 1094 TII->get(IsStoreXReg ? AArch64::UBFMXri : AArch64::UBFMWri), 1095 DestReg) 1096 .add(StMO) 1097 .addImm(Immr) 1098 .addImm(Imms) 1099 .setMIFlags(LoadI->getFlags()); 1100 } 1101 } 1102 1103 // Clear kill flags between store and load. 1104 for (MachineInstr &MI : make_range(StoreI->getIterator(), 1105 BitExtMI->getIterator())) 1106 if (MI.killsRegister(StRt, TRI)) { 1107 MI.clearRegisterKills(StRt, TRI); 1108 break; 1109 } 1110 1111 LLVM_DEBUG(dbgs() << "Promoting load by replacing :\n "); 1112 LLVM_DEBUG(StoreI->print(dbgs())); 1113 LLVM_DEBUG(dbgs() << " "); 1114 LLVM_DEBUG(LoadI->print(dbgs())); 1115 LLVM_DEBUG(dbgs() << " with instructions:\n "); 1116 LLVM_DEBUG(StoreI->print(dbgs())); 1117 LLVM_DEBUG(dbgs() << " "); 1118 LLVM_DEBUG((BitExtMI)->print(dbgs())); 1119 LLVM_DEBUG(dbgs() << "\n"); 1120 1121 // Erase the old instructions. 1122 LoadI->eraseFromParent(); 1123 return NextI; 1124 } 1125 1126 static bool inBoundsForPair(bool IsUnscaled, int Offset, int OffsetStride) { 1127 // Convert the byte-offset used by unscaled into an "element" offset used 1128 // by the scaled pair load/store instructions. 1129 if (IsUnscaled) { 1130 // If the byte-offset isn't a multiple of the stride, there's no point 1131 // trying to match it. 1132 if (Offset % OffsetStride) 1133 return false; 1134 Offset /= OffsetStride; 1135 } 1136 return Offset <= 63 && Offset >= -64; 1137 } 1138 1139 // Do alignment, specialized to power of 2 and for signed ints, 1140 // avoiding having to do a C-style cast from uint_64t to int when 1141 // using alignTo from include/llvm/Support/MathExtras.h. 1142 // FIXME: Move this function to include/MathExtras.h? 1143 static int alignTo(int Num, int PowOf2) { 1144 return (Num + PowOf2 - 1) & ~(PowOf2 - 1); 1145 } 1146 1147 static bool mayAlias(MachineInstr &MIa, MachineInstr &MIb, 1148 AliasAnalysis *AA) { 1149 // One of the instructions must modify memory. 1150 if (!MIa.mayStore() && !MIb.mayStore()) 1151 return false; 1152 1153 // Both instructions must be memory operations. 1154 if (!MIa.mayLoadOrStore() && !MIb.mayLoadOrStore()) 1155 return false; 1156 1157 return MIa.mayAlias(AA, MIb, /*UseTBAA*/false); 1158 } 1159 1160 static bool mayAlias(MachineInstr &MIa, 1161 SmallVectorImpl<MachineInstr *> &MemInsns, 1162 AliasAnalysis *AA) { 1163 for (MachineInstr *MIb : MemInsns) 1164 if (mayAlias(MIa, *MIb, AA)) 1165 return true; 1166 1167 return false; 1168 } 1169 1170 bool AArch64LoadStoreOpt::findMatchingStore( 1171 MachineBasicBlock::iterator I, unsigned Limit, 1172 MachineBasicBlock::iterator &StoreI) { 1173 MachineBasicBlock::iterator B = I->getParent()->begin(); 1174 MachineBasicBlock::iterator MBBI = I; 1175 MachineInstr &LoadMI = *I; 1176 Register BaseReg = getLdStBaseOp(LoadMI).getReg(); 1177 1178 // If the load is the first instruction in the block, there's obviously 1179 // not any matching store. 1180 if (MBBI == B) 1181 return false; 1182 1183 // Track which register units have been modified and used between the first 1184 // insn and the second insn. 1185 ModifiedRegUnits.clear(); 1186 UsedRegUnits.clear(); 1187 1188 unsigned Count = 0; 1189 do { 1190 --MBBI; 1191 MachineInstr &MI = *MBBI; 1192 1193 // Don't count transient instructions towards the search limit since there 1194 // may be different numbers of them if e.g. debug information is present. 1195 if (!MI.isTransient()) 1196 ++Count; 1197 1198 // If the load instruction reads directly from the address to which the 1199 // store instruction writes and the stored value is not modified, we can 1200 // promote the load. Since we do not handle stores with pre-/post-index, 1201 // it's unnecessary to check if BaseReg is modified by the store itself. 1202 if (MI.mayStore() && isMatchingStore(LoadMI, MI) && 1203 BaseReg == getLdStBaseOp(MI).getReg() && 1204 isLdOffsetInRangeOfSt(LoadMI, MI, TII) && 1205 ModifiedRegUnits.available(getLdStRegOp(MI).getReg())) { 1206 StoreI = MBBI; 1207 return true; 1208 } 1209 1210 if (MI.isCall()) 1211 return false; 1212 1213 // Update modified / uses register units. 1214 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits, TRI); 1215 1216 // Otherwise, if the base register is modified, we have no match, so 1217 // return early. 1218 if (!ModifiedRegUnits.available(BaseReg)) 1219 return false; 1220 1221 // If we encounter a store aliased with the load, return early. 1222 if (MI.mayStore() && mayAlias(LoadMI, MI, AA)) 1223 return false; 1224 } while (MBBI != B && Count < Limit); 1225 return false; 1226 } 1227 1228 // Returns true if FirstMI and MI are candidates for merging or pairing. 1229 // Otherwise, returns false. 1230 static bool areCandidatesToMergeOrPair(MachineInstr &FirstMI, MachineInstr &MI, 1231 LdStPairFlags &Flags, 1232 const AArch64InstrInfo *TII) { 1233 // If this is volatile or if pairing is suppressed, not a candidate. 1234 if (MI.hasOrderedMemoryRef() || TII->isLdStPairSuppressed(MI)) 1235 return false; 1236 1237 // We should have already checked FirstMI for pair suppression and volatility. 1238 assert(!FirstMI.hasOrderedMemoryRef() && 1239 !TII->isLdStPairSuppressed(FirstMI) && 1240 "FirstMI shouldn't get here if either of these checks are true."); 1241 1242 unsigned OpcA = FirstMI.getOpcode(); 1243 unsigned OpcB = MI.getOpcode(); 1244 1245 // Opcodes match: nothing more to check. 1246 if (OpcA == OpcB) 1247 return true; 1248 1249 // Try to match a sign-extended load/store with a zero-extended load/store. 1250 bool IsValidLdStrOpc, PairIsValidLdStrOpc; 1251 unsigned NonSExtOpc = getMatchingNonSExtOpcode(OpcA, &IsValidLdStrOpc); 1252 assert(IsValidLdStrOpc && 1253 "Given Opc should be a Load or Store with an immediate"); 1254 // OpcA will be the first instruction in the pair. 1255 if (NonSExtOpc == getMatchingNonSExtOpcode(OpcB, &PairIsValidLdStrOpc)) { 1256 Flags.setSExtIdx(NonSExtOpc == (unsigned)OpcA ? 1 : 0); 1257 return true; 1258 } 1259 1260 // If the second instruction isn't even a mergable/pairable load/store, bail 1261 // out. 1262 if (!PairIsValidLdStrOpc) 1263 return false; 1264 1265 // FIXME: We don't support merging narrow stores with mixed scaled/unscaled 1266 // offsets. 1267 if (isNarrowStore(OpcA) || isNarrowStore(OpcB)) 1268 return false; 1269 1270 // Try to match an unscaled load/store with a scaled load/store. 1271 return TII->isUnscaledLdSt(OpcA) != TII->isUnscaledLdSt(OpcB) && 1272 getMatchingPairOpcode(OpcA) == getMatchingPairOpcode(OpcB); 1273 1274 // FIXME: Can we also match a mixed sext/zext unscaled/scaled pair? 1275 } 1276 1277 static bool 1278 canRenameUpToDef(MachineInstr &FirstMI, LiveRegUnits &UsedInBetween, 1279 SmallPtrSetImpl<const TargetRegisterClass *> &RequiredClasses, 1280 const TargetRegisterInfo *TRI) { 1281 if (!FirstMI.mayStore()) 1282 return false; 1283 1284 // Check if we can find an unused register which we can use to rename 1285 // the register used by the first load/store. 1286 auto *RegClass = TRI->getMinimalPhysRegClass(getLdStRegOp(FirstMI).getReg()); 1287 MachineFunction &MF = *FirstMI.getParent()->getParent(); 1288 if (!RegClass || !MF.getRegInfo().tracksLiveness()) 1289 return false; 1290 1291 auto RegToRename = getLdStRegOp(FirstMI).getReg(); 1292 // For now, we only rename if the store operand gets killed at the store. 1293 if (!getLdStRegOp(FirstMI).isKill() && 1294 !any_of(FirstMI.operands(), 1295 [TRI, RegToRename](const MachineOperand &MOP) { 1296 return MOP.isReg() && !MOP.isDebug() && MOP.getReg() && 1297 MOP.isImplicit() && MOP.isKill() && 1298 TRI->regsOverlap(RegToRename, MOP.getReg()); 1299 })) { 1300 LLVM_DEBUG(dbgs() << " Operand not killed at " << FirstMI << "\n"); 1301 return false; 1302 } 1303 auto canRenameMOP = [](const MachineOperand &MOP) { 1304 return MOP.isImplicit() || 1305 (MOP.isRenamable() && !MOP.isEarlyClobber() && !MOP.isTied()); 1306 }; 1307 1308 bool FoundDef = false; 1309 1310 // For each instruction between FirstMI and the previous def for RegToRename, 1311 // we 1312 // * check if we can rename RegToRename in this instruction 1313 // * collect the registers used and required register classes for RegToRename. 1314 std::function<bool(MachineInstr &, bool)> CheckMIs = [&](MachineInstr &MI, 1315 bool IsDef) { 1316 LLVM_DEBUG(dbgs() << "Checking " << MI << "\n"); 1317 // Currently we do not try to rename across frame-setup instructions. 1318 if (MI.getFlag(MachineInstr::FrameSetup)) { 1319 LLVM_DEBUG(dbgs() << " Cannot rename framesetup instructions currently (" 1320 << MI << ")\n"); 1321 return false; 1322 } 1323 1324 UsedInBetween.accumulate(MI); 1325 1326 // For a definition, check that we can rename the definition and exit the 1327 // loop. 1328 FoundDef = IsDef; 1329 1330 // For defs, check if we can rename the first def of RegToRename. 1331 if (FoundDef) { 1332 // For some pseudo instructions, we might not generate code in the end 1333 // (e.g. KILL) and we would end up without a correct def for the rename 1334 // register. 1335 // TODO: This might be overly conservative and we could handle those cases 1336 // in multiple ways: 1337 // 1. Insert an extra copy, to materialize the def. 1338 // 2. Skip pseudo-defs until we find an non-pseudo def. 1339 if (MI.isPseudo()) { 1340 LLVM_DEBUG(dbgs() << " Cannot rename pseudo instruction " << MI 1341 << "\n"); 1342 return false; 1343 } 1344 1345 for (auto &MOP : MI.operands()) { 1346 if (!MOP.isReg() || !MOP.isDef() || MOP.isDebug() || !MOP.getReg() || 1347 !TRI->regsOverlap(MOP.getReg(), RegToRename)) 1348 continue; 1349 if (!canRenameMOP(MOP)) { 1350 LLVM_DEBUG(dbgs() 1351 << " Cannot rename " << MOP << " in " << MI << "\n"); 1352 return false; 1353 } 1354 RequiredClasses.insert(TRI->getMinimalPhysRegClass(MOP.getReg())); 1355 } 1356 return true; 1357 } else { 1358 for (auto &MOP : MI.operands()) { 1359 if (!MOP.isReg() || MOP.isDebug() || !MOP.getReg() || 1360 !TRI->regsOverlap(MOP.getReg(), RegToRename)) 1361 continue; 1362 1363 if (!canRenameMOP(MOP)) { 1364 LLVM_DEBUG(dbgs() 1365 << " Cannot rename " << MOP << " in " << MI << "\n"); 1366 return false; 1367 } 1368 RequiredClasses.insert(TRI->getMinimalPhysRegClass(MOP.getReg())); 1369 } 1370 } 1371 return true; 1372 }; 1373 1374 if (!forAllMIsUntilDef(FirstMI, RegToRename, TRI, LdStLimit, CheckMIs)) 1375 return false; 1376 1377 if (!FoundDef) { 1378 LLVM_DEBUG(dbgs() << " Did not find definition for register in BB\n"); 1379 return false; 1380 } 1381 return true; 1382 } 1383 1384 // Check if we can find a physical register for renaming. This register must: 1385 // * not be defined up to FirstMI (checking DefinedInBB) 1386 // * not used between the MI and the defining instruction of the register to 1387 // rename (checked using UsedInBetween). 1388 // * is available in all used register classes (checked using RequiredClasses). 1389 static Optional<MCPhysReg> tryToFindRegisterToRename( 1390 MachineInstr &FirstMI, MachineInstr &MI, LiveRegUnits &DefinedInBB, 1391 LiveRegUnits &UsedInBetween, 1392 SmallPtrSetImpl<const TargetRegisterClass *> &RequiredClasses, 1393 const TargetRegisterInfo *TRI) { 1394 auto &MF = *FirstMI.getParent()->getParent(); 1395 MachineRegisterInfo &RegInfo = MF.getRegInfo(); 1396 1397 // Checks if any sub- or super-register of PR is callee saved. 1398 auto AnySubOrSuperRegCalleePreserved = [&MF, TRI](MCPhysReg PR) { 1399 return any_of(TRI->sub_and_superregs_inclusive(PR), 1400 [&MF, TRI](MCPhysReg SubOrSuper) { 1401 return TRI->isCalleeSavedPhysReg(SubOrSuper, MF); 1402 }); 1403 }; 1404 1405 // Check if PR or one of its sub- or super-registers can be used for all 1406 // required register classes. 1407 auto CanBeUsedForAllClasses = [&RequiredClasses, TRI](MCPhysReg PR) { 1408 return all_of(RequiredClasses, [PR, TRI](const TargetRegisterClass *C) { 1409 return any_of(TRI->sub_and_superregs_inclusive(PR), 1410 [C, TRI](MCPhysReg SubOrSuper) { 1411 return C == TRI->getMinimalPhysRegClass(SubOrSuper); 1412 }); 1413 }); 1414 }; 1415 1416 auto *RegClass = TRI->getMinimalPhysRegClass(getLdStRegOp(FirstMI).getReg()); 1417 for (const MCPhysReg &PR : *RegClass) { 1418 if (DefinedInBB.available(PR) && UsedInBetween.available(PR) && 1419 !RegInfo.isReserved(PR) && !AnySubOrSuperRegCalleePreserved(PR) && 1420 CanBeUsedForAllClasses(PR)) { 1421 DefinedInBB.addReg(PR); 1422 LLVM_DEBUG(dbgs() << "Found rename register " << printReg(PR, TRI) 1423 << "\n"); 1424 return {PR}; 1425 } 1426 } 1427 LLVM_DEBUG(dbgs() << "No rename register found from " 1428 << TRI->getRegClassName(RegClass) << "\n"); 1429 return None; 1430 } 1431 1432 /// Scan the instructions looking for a load/store that can be combined with the 1433 /// current instruction into a wider equivalent or a load/store pair. 1434 MachineBasicBlock::iterator 1435 AArch64LoadStoreOpt::findMatchingInsn(MachineBasicBlock::iterator I, 1436 LdStPairFlags &Flags, unsigned Limit, 1437 bool FindNarrowMerge) { 1438 MachineBasicBlock::iterator E = I->getParent()->end(); 1439 MachineBasicBlock::iterator MBBI = I; 1440 MachineBasicBlock::iterator MBBIWithRenameReg; 1441 MachineInstr &FirstMI = *I; 1442 ++MBBI; 1443 1444 bool MayLoad = FirstMI.mayLoad(); 1445 bool IsUnscaled = TII->isUnscaledLdSt(FirstMI); 1446 Register Reg = getLdStRegOp(FirstMI).getReg(); 1447 Register BaseReg = getLdStBaseOp(FirstMI).getReg(); 1448 int Offset = getLdStOffsetOp(FirstMI).getImm(); 1449 int OffsetStride = IsUnscaled ? TII->getMemScale(FirstMI) : 1; 1450 bool IsPromotableZeroStore = isPromotableZeroStoreInst(FirstMI); 1451 1452 Optional<bool> MaybeCanRename = None; 1453 if (!EnableRenaming) 1454 MaybeCanRename = {false}; 1455 1456 SmallPtrSet<const TargetRegisterClass *, 5> RequiredClasses; 1457 LiveRegUnits UsedInBetween; 1458 UsedInBetween.init(*TRI); 1459 1460 Flags.clearRenameReg(); 1461 1462 // Track which register units have been modified and used between the first 1463 // insn (inclusive) and the second insn. 1464 ModifiedRegUnits.clear(); 1465 UsedRegUnits.clear(); 1466 1467 // Remember any instructions that read/write memory between FirstMI and MI. 1468 SmallVector<MachineInstr *, 4> MemInsns; 1469 1470 for (unsigned Count = 0; MBBI != E && Count < Limit; ++MBBI) { 1471 MachineInstr &MI = *MBBI; 1472 1473 UsedInBetween.accumulate(MI); 1474 1475 // Don't count transient instructions towards the search limit since there 1476 // may be different numbers of them if e.g. debug information is present. 1477 if (!MI.isTransient()) 1478 ++Count; 1479 1480 Flags.setSExtIdx(-1); 1481 if (areCandidatesToMergeOrPair(FirstMI, MI, Flags, TII) && 1482 getLdStOffsetOp(MI).isImm()) { 1483 assert(MI.mayLoadOrStore() && "Expected memory operation."); 1484 // If we've found another instruction with the same opcode, check to see 1485 // if the base and offset are compatible with our starting instruction. 1486 // These instructions all have scaled immediate operands, so we just 1487 // check for +1/-1. Make sure to check the new instruction offset is 1488 // actually an immediate and not a symbolic reference destined for 1489 // a relocation. 1490 Register MIBaseReg = getLdStBaseOp(MI).getReg(); 1491 int MIOffset = getLdStOffsetOp(MI).getImm(); 1492 bool MIIsUnscaled = TII->isUnscaledLdSt(MI); 1493 if (IsUnscaled != MIIsUnscaled) { 1494 // We're trying to pair instructions that differ in how they are scaled. 1495 // If FirstMI is scaled then scale the offset of MI accordingly. 1496 // Otherwise, do the opposite (i.e., make MI's offset unscaled). 1497 int MemSize = TII->getMemScale(MI); 1498 if (MIIsUnscaled) { 1499 // If the unscaled offset isn't a multiple of the MemSize, we can't 1500 // pair the operations together: bail and keep looking. 1501 if (MIOffset % MemSize) { 1502 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, 1503 UsedRegUnits, TRI); 1504 MemInsns.push_back(&MI); 1505 continue; 1506 } 1507 MIOffset /= MemSize; 1508 } else { 1509 MIOffset *= MemSize; 1510 } 1511 } 1512 1513 if (BaseReg == MIBaseReg && ((Offset == MIOffset + OffsetStride) || 1514 (Offset + OffsetStride == MIOffset))) { 1515 int MinOffset = Offset < MIOffset ? Offset : MIOffset; 1516 if (FindNarrowMerge) { 1517 // If the alignment requirements of the scaled wide load/store 1518 // instruction can't express the offset of the scaled narrow input, 1519 // bail and keep looking. For promotable zero stores, allow only when 1520 // the stored value is the same (i.e., WZR). 1521 if ((!IsUnscaled && alignTo(MinOffset, 2) != MinOffset) || 1522 (IsPromotableZeroStore && Reg != getLdStRegOp(MI).getReg())) { 1523 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, 1524 UsedRegUnits, TRI); 1525 MemInsns.push_back(&MI); 1526 continue; 1527 } 1528 } else { 1529 // Pairwise instructions have a 7-bit signed offset field. Single 1530 // insns have a 12-bit unsigned offset field. If the resultant 1531 // immediate offset of merging these instructions is out of range for 1532 // a pairwise instruction, bail and keep looking. 1533 if (!inBoundsForPair(IsUnscaled, MinOffset, OffsetStride)) { 1534 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, 1535 UsedRegUnits, TRI); 1536 MemInsns.push_back(&MI); 1537 continue; 1538 } 1539 // If the alignment requirements of the paired (scaled) instruction 1540 // can't express the offset of the unscaled input, bail and keep 1541 // looking. 1542 if (IsUnscaled && (alignTo(MinOffset, OffsetStride) != MinOffset)) { 1543 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, 1544 UsedRegUnits, TRI); 1545 MemInsns.push_back(&MI); 1546 continue; 1547 } 1548 } 1549 // If the destination register of the loads is the same register, bail 1550 // and keep looking. A load-pair instruction with both destination 1551 // registers the same is UNPREDICTABLE and will result in an exception. 1552 if (MayLoad && Reg == getLdStRegOp(MI).getReg()) { 1553 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits, 1554 TRI); 1555 MemInsns.push_back(&MI); 1556 continue; 1557 } 1558 1559 // If the Rt of the second instruction was not modified or used between 1560 // the two instructions and none of the instructions between the second 1561 // and first alias with the second, we can combine the second into the 1562 // first. 1563 if (ModifiedRegUnits.available(getLdStRegOp(MI).getReg()) && 1564 !(MI.mayLoad() && 1565 !UsedRegUnits.available(getLdStRegOp(MI).getReg())) && 1566 !mayAlias(MI, MemInsns, AA)) { 1567 1568 Flags.setMergeForward(false); 1569 Flags.clearRenameReg(); 1570 return MBBI; 1571 } 1572 1573 // Likewise, if the Rt of the first instruction is not modified or used 1574 // between the two instructions and none of the instructions between the 1575 // first and the second alias with the first, we can combine the first 1576 // into the second. 1577 if (!(MayLoad && 1578 !UsedRegUnits.available(getLdStRegOp(FirstMI).getReg())) && 1579 !mayAlias(FirstMI, MemInsns, AA)) { 1580 1581 if (ModifiedRegUnits.available(getLdStRegOp(FirstMI).getReg())) { 1582 Flags.setMergeForward(true); 1583 Flags.clearRenameReg(); 1584 return MBBI; 1585 } 1586 1587 if (DebugCounter::shouldExecute(RegRenamingCounter)) { 1588 if (!MaybeCanRename) 1589 MaybeCanRename = {canRenameUpToDef(FirstMI, UsedInBetween, 1590 RequiredClasses, TRI)}; 1591 1592 if (*MaybeCanRename) { 1593 Optional<MCPhysReg> MaybeRenameReg = tryToFindRegisterToRename( 1594 FirstMI, MI, DefinedInBB, UsedInBetween, RequiredClasses, 1595 TRI); 1596 if (MaybeRenameReg) { 1597 Flags.setRenameReg(*MaybeRenameReg); 1598 Flags.setMergeForward(true); 1599 MBBIWithRenameReg = MBBI; 1600 } 1601 } 1602 } 1603 } 1604 // Unable to combine these instructions due to interference in between. 1605 // Keep looking. 1606 } 1607 } 1608 1609 if (Flags.getRenameReg()) 1610 return MBBIWithRenameReg; 1611 1612 // If the instruction wasn't a matching load or store. Stop searching if we 1613 // encounter a call instruction that might modify memory. 1614 if (MI.isCall()) 1615 return E; 1616 1617 // Update modified / uses register units. 1618 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits, TRI); 1619 1620 // Otherwise, if the base register is modified, we have no match, so 1621 // return early. 1622 if (!ModifiedRegUnits.available(BaseReg)) 1623 return E; 1624 1625 // Update list of instructions that read/write memory. 1626 if (MI.mayLoadOrStore()) 1627 MemInsns.push_back(&MI); 1628 } 1629 return E; 1630 } 1631 1632 MachineBasicBlock::iterator 1633 AArch64LoadStoreOpt::mergeUpdateInsn(MachineBasicBlock::iterator I, 1634 MachineBasicBlock::iterator Update, 1635 bool IsPreIdx) { 1636 assert((Update->getOpcode() == AArch64::ADDXri || 1637 Update->getOpcode() == AArch64::SUBXri) && 1638 "Unexpected base register update instruction to merge!"); 1639 MachineBasicBlock::iterator NextI = I; 1640 // Return the instruction following the merged instruction, which is 1641 // the instruction following our unmerged load. Unless that's the add/sub 1642 // instruction we're merging, in which case it's the one after that. 1643 if (++NextI == Update) 1644 ++NextI; 1645 1646 int Value = Update->getOperand(2).getImm(); 1647 assert(AArch64_AM::getShiftValue(Update->getOperand(3).getImm()) == 0 && 1648 "Can't merge 1 << 12 offset into pre-/post-indexed load / store"); 1649 if (Update->getOpcode() == AArch64::SUBXri) 1650 Value = -Value; 1651 1652 unsigned NewOpc = IsPreIdx ? getPreIndexedOpcode(I->getOpcode()) 1653 : getPostIndexedOpcode(I->getOpcode()); 1654 MachineInstrBuilder MIB; 1655 int Scale, MinOffset, MaxOffset; 1656 getPrePostIndexedMemOpInfo(*I, Scale, MinOffset, MaxOffset); 1657 if (!isPairedLdSt(*I)) { 1658 // Non-paired instruction. 1659 MIB = BuildMI(*I->getParent(), I, I->getDebugLoc(), TII->get(NewOpc)) 1660 .add(getLdStRegOp(*Update)) 1661 .add(getLdStRegOp(*I)) 1662 .add(getLdStBaseOp(*I)) 1663 .addImm(Value / Scale) 1664 .setMemRefs(I->memoperands()) 1665 .setMIFlags(I->mergeFlagsWith(*Update)); 1666 } else { 1667 // Paired instruction. 1668 MIB = BuildMI(*I->getParent(), I, I->getDebugLoc(), TII->get(NewOpc)) 1669 .add(getLdStRegOp(*Update)) 1670 .add(getLdStRegOp(*I, 0)) 1671 .add(getLdStRegOp(*I, 1)) 1672 .add(getLdStBaseOp(*I)) 1673 .addImm(Value / Scale) 1674 .setMemRefs(I->memoperands()) 1675 .setMIFlags(I->mergeFlagsWith(*Update)); 1676 } 1677 (void)MIB; 1678 1679 if (IsPreIdx) { 1680 ++NumPreFolded; 1681 LLVM_DEBUG(dbgs() << "Creating pre-indexed load/store."); 1682 } else { 1683 ++NumPostFolded; 1684 LLVM_DEBUG(dbgs() << "Creating post-indexed load/store."); 1685 } 1686 LLVM_DEBUG(dbgs() << " Replacing instructions:\n "); 1687 LLVM_DEBUG(I->print(dbgs())); 1688 LLVM_DEBUG(dbgs() << " "); 1689 LLVM_DEBUG(Update->print(dbgs())); 1690 LLVM_DEBUG(dbgs() << " with instruction:\n "); 1691 LLVM_DEBUG(((MachineInstr *)MIB)->print(dbgs())); 1692 LLVM_DEBUG(dbgs() << "\n"); 1693 1694 // Erase the old instructions for the block. 1695 I->eraseFromParent(); 1696 Update->eraseFromParent(); 1697 1698 return NextI; 1699 } 1700 1701 bool AArch64LoadStoreOpt::isMatchingUpdateInsn(MachineInstr &MemMI, 1702 MachineInstr &MI, 1703 unsigned BaseReg, int Offset) { 1704 switch (MI.getOpcode()) { 1705 default: 1706 break; 1707 case AArch64::SUBXri: 1708 case AArch64::ADDXri: 1709 // Make sure it's a vanilla immediate operand, not a relocation or 1710 // anything else we can't handle. 1711 if (!MI.getOperand(2).isImm()) 1712 break; 1713 // Watch out for 1 << 12 shifted value. 1714 if (AArch64_AM::getShiftValue(MI.getOperand(3).getImm())) 1715 break; 1716 1717 // The update instruction source and destination register must be the 1718 // same as the load/store base register. 1719 if (MI.getOperand(0).getReg() != BaseReg || 1720 MI.getOperand(1).getReg() != BaseReg) 1721 break; 1722 1723 int UpdateOffset = MI.getOperand(2).getImm(); 1724 if (MI.getOpcode() == AArch64::SUBXri) 1725 UpdateOffset = -UpdateOffset; 1726 1727 // The immediate must be a multiple of the scaling factor of the pre/post 1728 // indexed instruction. 1729 int Scale, MinOffset, MaxOffset; 1730 getPrePostIndexedMemOpInfo(MemMI, Scale, MinOffset, MaxOffset); 1731 if (UpdateOffset % Scale != 0) 1732 break; 1733 1734 // Scaled offset must fit in the instruction immediate. 1735 int ScaledOffset = UpdateOffset / Scale; 1736 if (ScaledOffset > MaxOffset || ScaledOffset < MinOffset) 1737 break; 1738 1739 // If we have a non-zero Offset, we check that it matches the amount 1740 // we're adding to the register. 1741 if (!Offset || Offset == UpdateOffset) 1742 return true; 1743 break; 1744 } 1745 return false; 1746 } 1747 1748 MachineBasicBlock::iterator AArch64LoadStoreOpt::findMatchingUpdateInsnForward( 1749 MachineBasicBlock::iterator I, int UnscaledOffset, unsigned Limit) { 1750 MachineBasicBlock::iterator E = I->getParent()->end(); 1751 MachineInstr &MemMI = *I; 1752 MachineBasicBlock::iterator MBBI = I; 1753 1754 Register BaseReg = getLdStBaseOp(MemMI).getReg(); 1755 int MIUnscaledOffset = getLdStOffsetOp(MemMI).getImm() * TII->getMemScale(MemMI); 1756 1757 // Scan forward looking for post-index opportunities. Updating instructions 1758 // can't be formed if the memory instruction doesn't have the offset we're 1759 // looking for. 1760 if (MIUnscaledOffset != UnscaledOffset) 1761 return E; 1762 1763 // If the base register overlaps a source/destination register, we can't 1764 // merge the update. This does not apply to tag store instructions which 1765 // ignore the address part of the source register. 1766 // This does not apply to STGPi as well, which does not have unpredictable 1767 // behavior in this case unlike normal stores, and always performs writeback 1768 // after reading the source register value. 1769 if (!isTagStore(MemMI) && MemMI.getOpcode() != AArch64::STGPi) { 1770 bool IsPairedInsn = isPairedLdSt(MemMI); 1771 for (unsigned i = 0, e = IsPairedInsn ? 2 : 1; i != e; ++i) { 1772 Register DestReg = getLdStRegOp(MemMI, i).getReg(); 1773 if (DestReg == BaseReg || TRI->isSubRegister(BaseReg, DestReg)) 1774 return E; 1775 } 1776 } 1777 1778 // Track which register units have been modified and used between the first 1779 // insn (inclusive) and the second insn. 1780 ModifiedRegUnits.clear(); 1781 UsedRegUnits.clear(); 1782 ++MBBI; 1783 for (unsigned Count = 0; MBBI != E && Count < Limit; ++MBBI) { 1784 MachineInstr &MI = *MBBI; 1785 1786 // Don't count transient instructions towards the search limit since there 1787 // may be different numbers of them if e.g. debug information is present. 1788 if (!MI.isTransient()) 1789 ++Count; 1790 1791 // If we found a match, return it. 1792 if (isMatchingUpdateInsn(*I, MI, BaseReg, UnscaledOffset)) 1793 return MBBI; 1794 1795 // Update the status of what the instruction clobbered and used. 1796 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits, TRI); 1797 1798 // Otherwise, if the base register is used or modified, we have no match, so 1799 // return early. 1800 if (!ModifiedRegUnits.available(BaseReg) || 1801 !UsedRegUnits.available(BaseReg)) 1802 return E; 1803 } 1804 return E; 1805 } 1806 1807 MachineBasicBlock::iterator AArch64LoadStoreOpt::findMatchingUpdateInsnBackward( 1808 MachineBasicBlock::iterator I, unsigned Limit) { 1809 MachineBasicBlock::iterator B = I->getParent()->begin(); 1810 MachineBasicBlock::iterator E = I->getParent()->end(); 1811 MachineInstr &MemMI = *I; 1812 MachineBasicBlock::iterator MBBI = I; 1813 1814 Register BaseReg = getLdStBaseOp(MemMI).getReg(); 1815 int Offset = getLdStOffsetOp(MemMI).getImm(); 1816 1817 // If the load/store is the first instruction in the block, there's obviously 1818 // not any matching update. Ditto if the memory offset isn't zero. 1819 if (MBBI == B || Offset != 0) 1820 return E; 1821 // If the base register overlaps a destination register, we can't 1822 // merge the update. 1823 if (!isTagStore(MemMI)) { 1824 bool IsPairedInsn = isPairedLdSt(MemMI); 1825 for (unsigned i = 0, e = IsPairedInsn ? 2 : 1; i != e; ++i) { 1826 Register DestReg = getLdStRegOp(MemMI, i).getReg(); 1827 if (DestReg == BaseReg || TRI->isSubRegister(BaseReg, DestReg)) 1828 return E; 1829 } 1830 } 1831 1832 // Track which register units have been modified and used between the first 1833 // insn (inclusive) and the second insn. 1834 ModifiedRegUnits.clear(); 1835 UsedRegUnits.clear(); 1836 unsigned Count = 0; 1837 do { 1838 --MBBI; 1839 MachineInstr &MI = *MBBI; 1840 1841 // Don't count transient instructions towards the search limit since there 1842 // may be different numbers of them if e.g. debug information is present. 1843 if (!MI.isTransient()) 1844 ++Count; 1845 1846 // If we found a match, return it. 1847 if (isMatchingUpdateInsn(*I, MI, BaseReg, Offset)) 1848 return MBBI; 1849 1850 // Update the status of what the instruction clobbered and used. 1851 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits, TRI); 1852 1853 // Otherwise, if the base register is used or modified, we have no match, so 1854 // return early. 1855 if (!ModifiedRegUnits.available(BaseReg) || 1856 !UsedRegUnits.available(BaseReg)) 1857 return E; 1858 } while (MBBI != B && Count < Limit); 1859 return E; 1860 } 1861 1862 bool AArch64LoadStoreOpt::tryToPromoteLoadFromStore( 1863 MachineBasicBlock::iterator &MBBI) { 1864 MachineInstr &MI = *MBBI; 1865 // If this is a volatile load, don't mess with it. 1866 if (MI.hasOrderedMemoryRef()) 1867 return false; 1868 1869 // Make sure this is a reg+imm. 1870 // FIXME: It is possible to extend it to handle reg+reg cases. 1871 if (!getLdStOffsetOp(MI).isImm()) 1872 return false; 1873 1874 // Look backward up to LdStLimit instructions. 1875 MachineBasicBlock::iterator StoreI; 1876 if (findMatchingStore(MBBI, LdStLimit, StoreI)) { 1877 ++NumLoadsFromStoresPromoted; 1878 // Promote the load. Keeping the iterator straight is a 1879 // pain, so we let the merge routine tell us what the next instruction 1880 // is after it's done mucking about. 1881 MBBI = promoteLoadFromStore(MBBI, StoreI); 1882 return true; 1883 } 1884 return false; 1885 } 1886 1887 // Merge adjacent zero stores into a wider store. 1888 bool AArch64LoadStoreOpt::tryToMergeZeroStInst( 1889 MachineBasicBlock::iterator &MBBI) { 1890 assert(isPromotableZeroStoreInst(*MBBI) && "Expected narrow store."); 1891 MachineInstr &MI = *MBBI; 1892 MachineBasicBlock::iterator E = MI.getParent()->end(); 1893 1894 if (!TII->isCandidateToMergeOrPair(MI)) 1895 return false; 1896 1897 // Look ahead up to LdStLimit instructions for a mergable instruction. 1898 LdStPairFlags Flags; 1899 MachineBasicBlock::iterator MergeMI = 1900 findMatchingInsn(MBBI, Flags, LdStLimit, /* FindNarrowMerge = */ true); 1901 if (MergeMI != E) { 1902 ++NumZeroStoresPromoted; 1903 1904 // Keeping the iterator straight is a pain, so we let the merge routine tell 1905 // us what the next instruction is after it's done mucking about. 1906 MBBI = mergeNarrowZeroStores(MBBI, MergeMI, Flags); 1907 return true; 1908 } 1909 return false; 1910 } 1911 1912 // Find loads and stores that can be merged into a single load or store pair 1913 // instruction. 1914 bool AArch64LoadStoreOpt::tryToPairLdStInst(MachineBasicBlock::iterator &MBBI) { 1915 MachineInstr &MI = *MBBI; 1916 MachineBasicBlock::iterator E = MI.getParent()->end(); 1917 1918 if (!TII->isCandidateToMergeOrPair(MI)) 1919 return false; 1920 1921 // Early exit if the offset is not possible to match. (6 bits of positive 1922 // range, plus allow an extra one in case we find a later insn that matches 1923 // with Offset-1) 1924 bool IsUnscaled = TII->isUnscaledLdSt(MI); 1925 int Offset = getLdStOffsetOp(MI).getImm(); 1926 int OffsetStride = IsUnscaled ? TII->getMemScale(MI) : 1; 1927 // Allow one more for offset. 1928 if (Offset > 0) 1929 Offset -= OffsetStride; 1930 if (!inBoundsForPair(IsUnscaled, Offset, OffsetStride)) 1931 return false; 1932 1933 // Look ahead up to LdStLimit instructions for a pairable instruction. 1934 LdStPairFlags Flags; 1935 MachineBasicBlock::iterator Paired = 1936 findMatchingInsn(MBBI, Flags, LdStLimit, /* FindNarrowMerge = */ false); 1937 if (Paired != E) { 1938 ++NumPairCreated; 1939 if (TII->isUnscaledLdSt(MI)) 1940 ++NumUnscaledPairCreated; 1941 // Keeping the iterator straight is a pain, so we let the merge routine tell 1942 // us what the next instruction is after it's done mucking about. 1943 auto Prev = std::prev(MBBI); 1944 MBBI = mergePairedInsns(MBBI, Paired, Flags); 1945 // Collect liveness info for instructions between Prev and the new position 1946 // MBBI. 1947 for (auto I = std::next(Prev); I != MBBI; I++) 1948 updateDefinedRegisters(*I, DefinedInBB, TRI); 1949 1950 return true; 1951 } 1952 return false; 1953 } 1954 1955 bool AArch64LoadStoreOpt::tryToMergeLdStUpdate 1956 (MachineBasicBlock::iterator &MBBI) { 1957 MachineInstr &MI = *MBBI; 1958 MachineBasicBlock::iterator E = MI.getParent()->end(); 1959 MachineBasicBlock::iterator Update; 1960 1961 // Look forward to try to form a post-index instruction. For example, 1962 // ldr x0, [x20] 1963 // add x20, x20, #32 1964 // merged into: 1965 // ldr x0, [x20], #32 1966 Update = findMatchingUpdateInsnForward(MBBI, 0, UpdateLimit); 1967 if (Update != E) { 1968 // Merge the update into the ld/st. 1969 MBBI = mergeUpdateInsn(MBBI, Update, /*IsPreIdx=*/false); 1970 return true; 1971 } 1972 1973 // Don't know how to handle unscaled pre/post-index versions below, so bail. 1974 if (TII->isUnscaledLdSt(MI.getOpcode())) 1975 return false; 1976 1977 // Look back to try to find a pre-index instruction. For example, 1978 // add x0, x0, #8 1979 // ldr x1, [x0] 1980 // merged into: 1981 // ldr x1, [x0, #8]! 1982 Update = findMatchingUpdateInsnBackward(MBBI, UpdateLimit); 1983 if (Update != E) { 1984 // Merge the update into the ld/st. 1985 MBBI = mergeUpdateInsn(MBBI, Update, /*IsPreIdx=*/true); 1986 return true; 1987 } 1988 1989 // The immediate in the load/store is scaled by the size of the memory 1990 // operation. The immediate in the add we're looking for, 1991 // however, is not, so adjust here. 1992 int UnscaledOffset = getLdStOffsetOp(MI).getImm() * TII->getMemScale(MI); 1993 1994 // Look forward to try to find a pre-index instruction. For example, 1995 // ldr x1, [x0, #64] 1996 // add x0, x0, #64 1997 // merged into: 1998 // ldr x1, [x0, #64]! 1999 Update = findMatchingUpdateInsnForward(MBBI, UnscaledOffset, UpdateLimit); 2000 if (Update != E) { 2001 // Merge the update into the ld/st. 2002 MBBI = mergeUpdateInsn(MBBI, Update, /*IsPreIdx=*/true); 2003 return true; 2004 } 2005 2006 return false; 2007 } 2008 2009 bool AArch64LoadStoreOpt::optimizeBlock(MachineBasicBlock &MBB, 2010 bool EnableNarrowZeroStOpt) { 2011 2012 bool Modified = false; 2013 // Four tranformations to do here: 2014 // 1) Find loads that directly read from stores and promote them by 2015 // replacing with mov instructions. If the store is wider than the load, 2016 // the load will be replaced with a bitfield extract. 2017 // e.g., 2018 // str w1, [x0, #4] 2019 // ldrh w2, [x0, #6] 2020 // ; becomes 2021 // str w1, [x0, #4] 2022 // lsr w2, w1, #16 2023 for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end(); 2024 MBBI != E;) { 2025 if (isPromotableLoadFromStore(*MBBI) && tryToPromoteLoadFromStore(MBBI)) 2026 Modified = true; 2027 else 2028 ++MBBI; 2029 } 2030 // 2) Merge adjacent zero stores into a wider store. 2031 // e.g., 2032 // strh wzr, [x0] 2033 // strh wzr, [x0, #2] 2034 // ; becomes 2035 // str wzr, [x0] 2036 // e.g., 2037 // str wzr, [x0] 2038 // str wzr, [x0, #4] 2039 // ; becomes 2040 // str xzr, [x0] 2041 if (EnableNarrowZeroStOpt) 2042 for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end(); 2043 MBBI != E;) { 2044 if (isPromotableZeroStoreInst(*MBBI) && tryToMergeZeroStInst(MBBI)) 2045 Modified = true; 2046 else 2047 ++MBBI; 2048 } 2049 // 3) Find loads and stores that can be merged into a single load or store 2050 // pair instruction. 2051 // e.g., 2052 // ldr x0, [x2] 2053 // ldr x1, [x2, #8] 2054 // ; becomes 2055 // ldp x0, x1, [x2] 2056 2057 if (MBB.getParent()->getRegInfo().tracksLiveness()) { 2058 DefinedInBB.clear(); 2059 DefinedInBB.addLiveIns(MBB); 2060 } 2061 2062 for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end(); 2063 MBBI != E;) { 2064 // Track currently live registers up to this point, to help with 2065 // searching for a rename register on demand. 2066 updateDefinedRegisters(*MBBI, DefinedInBB, TRI); 2067 if (TII->isPairableLdStInst(*MBBI) && tryToPairLdStInst(MBBI)) 2068 Modified = true; 2069 else 2070 ++MBBI; 2071 } 2072 // 4) Find base register updates that can be merged into the load or store 2073 // as a base-reg writeback. 2074 // e.g., 2075 // ldr x0, [x2] 2076 // add x2, x2, #4 2077 // ; becomes 2078 // ldr x0, [x2], #4 2079 for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end(); 2080 MBBI != E;) { 2081 if (isMergeableLdStUpdate(*MBBI) && tryToMergeLdStUpdate(MBBI)) 2082 Modified = true; 2083 else 2084 ++MBBI; 2085 } 2086 2087 return Modified; 2088 } 2089 2090 bool AArch64LoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) { 2091 if (skipFunction(Fn.getFunction())) 2092 return false; 2093 2094 Subtarget = &static_cast<const AArch64Subtarget &>(Fn.getSubtarget()); 2095 TII = static_cast<const AArch64InstrInfo *>(Subtarget->getInstrInfo()); 2096 TRI = Subtarget->getRegisterInfo(); 2097 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); 2098 2099 // Resize the modified and used register unit trackers. We do this once 2100 // per function and then clear the register units each time we optimize a load 2101 // or store. 2102 ModifiedRegUnits.init(*TRI); 2103 UsedRegUnits.init(*TRI); 2104 DefinedInBB.init(*TRI); 2105 2106 bool Modified = false; 2107 bool enableNarrowZeroStOpt = !Subtarget->requiresStrictAlign(); 2108 for (auto &MBB : Fn) { 2109 auto M = optimizeBlock(MBB, enableNarrowZeroStOpt); 2110 Modified |= M; 2111 } 2112 2113 return Modified; 2114 } 2115 2116 // FIXME: Do we need/want a pre-alloc pass like ARM has to try to keep loads and 2117 // stores near one another? Note: The pre-RA instruction scheduler already has 2118 // hooks to try and schedule pairable loads/stores together to improve pairing 2119 // opportunities. Thus, pre-RA pairing pass may not be worth the effort. 2120 2121 // FIXME: When pairing store instructions it's very possible for this pass to 2122 // hoist a store with a KILL marker above another use (without a KILL marker). 2123 // The resulting IR is invalid, but nothing uses the KILL markers after this 2124 // pass, so it's never caused a problem in practice. 2125 2126 /// createAArch64LoadStoreOptimizationPass - returns an instance of the 2127 /// load / store optimization pass. 2128 FunctionPass *llvm::createAArch64LoadStoreOptimizationPass() { 2129 return new AArch64LoadStoreOpt(); 2130 } 2131