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 // The pass runs after the PrologEpilogInserter where we emit the CFI 13 // instructions. In order to preserve the correctness of the unwind informaiton, 14 // the pass should not change the order of any two instructions, one of which 15 // has the FrameSetup/FrameDestroy flag or, alternatively, apply an add-hoc fix 16 // to unwind information. 17 // 18 //===----------------------------------------------------------------------===// 19 20 #include "AArch64InstrInfo.h" 21 #include "AArch64MachineFunctionInfo.h" 22 #include "AArch64Subtarget.h" 23 #include "MCTargetDesc/AArch64AddressingModes.h" 24 #include "llvm/ADT/BitVector.h" 25 #include "llvm/ADT/SmallVector.h" 26 #include "llvm/ADT/Statistic.h" 27 #include "llvm/ADT/StringRef.h" 28 #include "llvm/ADT/iterator_range.h" 29 #include "llvm/Analysis/AliasAnalysis.h" 30 #include "llvm/CodeGen/MachineBasicBlock.h" 31 #include "llvm/CodeGen/MachineFunction.h" 32 #include "llvm/CodeGen/MachineFunctionPass.h" 33 #include "llvm/CodeGen/MachineInstr.h" 34 #include "llvm/CodeGen/MachineInstrBuilder.h" 35 #include "llvm/CodeGen/MachineOperand.h" 36 #include "llvm/CodeGen/MachineRegisterInfo.h" 37 #include "llvm/CodeGen/TargetRegisterInfo.h" 38 #include "llvm/IR/DebugLoc.h" 39 #include "llvm/MC/MCAsmInfo.h" 40 #include "llvm/MC/MCDwarf.h" 41 #include "llvm/MC/MCRegisterInfo.h" 42 #include "llvm/Pass.h" 43 #include "llvm/Support/CommandLine.h" 44 #include "llvm/Support/Debug.h" 45 #include "llvm/Support/DebugCounter.h" 46 #include "llvm/Support/ErrorHandling.h" 47 #include "llvm/Support/raw_ostream.h" 48 #include <cassert> 49 #include <cstdint> 50 #include <functional> 51 #include <iterator> 52 #include <limits> 53 #include <optional> 54 55 using namespace llvm; 56 57 #define DEBUG_TYPE "aarch64-ldst-opt" 58 59 STATISTIC(NumPairCreated, "Number of load/store pair instructions generated"); 60 STATISTIC(NumPostFolded, "Number of post-index updates folded"); 61 STATISTIC(NumPreFolded, "Number of pre-index updates folded"); 62 STATISTIC(NumUnscaledPairCreated, 63 "Number of load/store from unscaled generated"); 64 STATISTIC(NumZeroStoresPromoted, "Number of narrow zero stores promoted"); 65 STATISTIC(NumLoadsFromStoresPromoted, "Number of loads from stores promoted"); 66 67 DEBUG_COUNTER(RegRenamingCounter, DEBUG_TYPE "-reg-renaming", 68 "Controls which pairs are considered for renaming"); 69 70 // The LdStLimit limits how far we search for load/store pairs. 71 static cl::opt<unsigned> LdStLimit("aarch64-load-store-scan-limit", 72 cl::init(20), cl::Hidden); 73 74 // The UpdateLimit limits how far we search for update instructions when we form 75 // pre-/post-index instructions. 76 static cl::opt<unsigned> UpdateLimit("aarch64-update-scan-limit", cl::init(100), 77 cl::Hidden); 78 79 // Enable register renaming to find additional store pairing opportunities. 80 static cl::opt<bool> EnableRenaming("aarch64-load-store-renaming", 81 cl::init(true), cl::Hidden); 82 83 #define AARCH64_LOAD_STORE_OPT_NAME "AArch64 load / store optimization pass" 84 85 namespace { 86 87 using LdStPairFlags = struct LdStPairFlags { 88 // If a matching instruction is found, MergeForward is set to true if the 89 // merge is to remove the first instruction and replace the second with 90 // a pair-wise insn, and false if the reverse is true. 91 bool MergeForward = false; 92 93 // SExtIdx gives the index of the result of the load pair that must be 94 // extended. The value of SExtIdx assumes that the paired load produces the 95 // value in this order: (I, returned iterator), i.e., -1 means no value has 96 // to be extended, 0 means I, and 1 means the returned iterator. 97 int SExtIdx = -1; 98 99 // If not none, RenameReg can be used to rename the result register of the 100 // first store in a pair. Currently this only works when merging stores 101 // forward. 102 std::optional<MCPhysReg> RenameReg; 103 104 LdStPairFlags() = default; 105 106 void setMergeForward(bool V = true) { MergeForward = V; } 107 bool getMergeForward() const { return MergeForward; } 108 109 void setSExtIdx(int V) { SExtIdx = V; } 110 int getSExtIdx() const { return SExtIdx; } 111 112 void setRenameReg(MCPhysReg R) { RenameReg = R; } 113 void clearRenameReg() { RenameReg = std::nullopt; } 114 std::optional<MCPhysReg> getRenameReg() const { return RenameReg; } 115 }; 116 117 struct AArch64LoadStoreOpt : public MachineFunctionPass { 118 static char ID; 119 120 AArch64LoadStoreOpt() : MachineFunctionPass(ID) { 121 initializeAArch64LoadStoreOptPass(*PassRegistry::getPassRegistry()); 122 } 123 124 AliasAnalysis *AA; 125 const AArch64InstrInfo *TII; 126 const TargetRegisterInfo *TRI; 127 const AArch64Subtarget *Subtarget; 128 129 // Track which register units have been modified and used. 130 LiveRegUnits ModifiedRegUnits, UsedRegUnits; 131 LiveRegUnits DefinedInBB; 132 133 void getAnalysisUsage(AnalysisUsage &AU) const override { 134 AU.addRequired<AAResultsWrapperPass>(); 135 MachineFunctionPass::getAnalysisUsage(AU); 136 } 137 138 // Scan the instructions looking for a load/store that can be combined 139 // with the current instruction into a load/store pair. 140 // Return the matching instruction if one is found, else MBB->end(). 141 MachineBasicBlock::iterator findMatchingInsn(MachineBasicBlock::iterator I, 142 LdStPairFlags &Flags, 143 unsigned Limit, 144 bool FindNarrowMerge); 145 146 // Scan the instructions looking for a store that writes to the address from 147 // which the current load instruction reads. Return true if one is found. 148 bool findMatchingStore(MachineBasicBlock::iterator I, unsigned Limit, 149 MachineBasicBlock::iterator &StoreI); 150 151 // Merge the two instructions indicated into a wider narrow store instruction. 152 MachineBasicBlock::iterator 153 mergeNarrowZeroStores(MachineBasicBlock::iterator I, 154 MachineBasicBlock::iterator MergeMI, 155 const LdStPairFlags &Flags); 156 157 // Merge the two instructions indicated into a single pair-wise instruction. 158 MachineBasicBlock::iterator 159 mergePairedInsns(MachineBasicBlock::iterator I, 160 MachineBasicBlock::iterator Paired, 161 const LdStPairFlags &Flags); 162 163 // Promote the load that reads directly from the address stored to. 164 MachineBasicBlock::iterator 165 promoteLoadFromStore(MachineBasicBlock::iterator LoadI, 166 MachineBasicBlock::iterator StoreI); 167 168 // Scan the instruction list to find a base register update that can 169 // be combined with the current instruction (a load or store) using 170 // pre or post indexed addressing with writeback. Scan forwards. 171 MachineBasicBlock::iterator 172 findMatchingUpdateInsnForward(MachineBasicBlock::iterator I, 173 int UnscaledOffset, unsigned Limit); 174 175 // Scan the instruction list to find a base register update that can 176 // be combined with the current instruction (a load or store) using 177 // pre or post indexed addressing with writeback. Scan backwards. 178 MachineBasicBlock::iterator 179 findMatchingUpdateInsnBackward(MachineBasicBlock::iterator I, unsigned Limit); 180 181 // Find an instruction that updates the base register of the ld/st 182 // instruction. 183 bool isMatchingUpdateInsn(MachineInstr &MemMI, MachineInstr &MI, 184 unsigned BaseReg, int Offset); 185 186 // Merge a pre- or post-index base register update into a ld/st instruction. 187 MachineBasicBlock::iterator 188 mergeUpdateInsn(MachineBasicBlock::iterator I, 189 MachineBasicBlock::iterator Update, bool IsPreIdx); 190 191 // Find and merge zero store instructions. 192 bool tryToMergeZeroStInst(MachineBasicBlock::iterator &MBBI); 193 194 // Find and pair ldr/str instructions. 195 bool tryToPairLdStInst(MachineBasicBlock::iterator &MBBI); 196 197 // Find and promote load instructions which read directly from store. 198 bool tryToPromoteLoadFromStore(MachineBasicBlock::iterator &MBBI); 199 200 // Find and merge a base register updates before or after a ld/st instruction. 201 bool tryToMergeLdStUpdate(MachineBasicBlock::iterator &MBBI); 202 203 bool optimizeBlock(MachineBasicBlock &MBB, bool EnableNarrowZeroStOpt); 204 205 bool runOnMachineFunction(MachineFunction &Fn) override; 206 207 MachineFunctionProperties getRequiredProperties() const override { 208 return MachineFunctionProperties().set( 209 MachineFunctionProperties::Property::NoVRegs); 210 } 211 212 StringRef getPassName() const override { return AARCH64_LOAD_STORE_OPT_NAME; } 213 }; 214 215 char AArch64LoadStoreOpt::ID = 0; 216 217 } // end anonymous namespace 218 219 INITIALIZE_PASS(AArch64LoadStoreOpt, "aarch64-ldst-opt", 220 AARCH64_LOAD_STORE_OPT_NAME, false, false) 221 222 static bool isNarrowStore(unsigned Opc) { 223 switch (Opc) { 224 default: 225 return false; 226 case AArch64::STRBBui: 227 case AArch64::STURBBi: 228 case AArch64::STRHHui: 229 case AArch64::STURHHi: 230 return true; 231 } 232 } 233 234 // These instruction set memory tag and either keep memory contents unchanged or 235 // set it to zero, ignoring the address part of the source register. 236 static bool isTagStore(const MachineInstr &MI) { 237 switch (MI.getOpcode()) { 238 default: 239 return false; 240 case AArch64::STGi: 241 case AArch64::STZGi: 242 case AArch64::ST2Gi: 243 case AArch64::STZ2Gi: 244 return true; 245 } 246 } 247 248 static unsigned getMatchingNonSExtOpcode(unsigned Opc, 249 bool *IsValidLdStrOpc = nullptr) { 250 if (IsValidLdStrOpc) 251 *IsValidLdStrOpc = true; 252 switch (Opc) { 253 default: 254 if (IsValidLdStrOpc) 255 *IsValidLdStrOpc = false; 256 return std::numeric_limits<unsigned>::max(); 257 case AArch64::STRDui: 258 case AArch64::STURDi: 259 case AArch64::STRDpre: 260 case AArch64::STRQui: 261 case AArch64::STURQi: 262 case AArch64::STRQpre: 263 case AArch64::STRBBui: 264 case AArch64::STURBBi: 265 case AArch64::STRHHui: 266 case AArch64::STURHHi: 267 case AArch64::STRWui: 268 case AArch64::STRWpre: 269 case AArch64::STURWi: 270 case AArch64::STRXui: 271 case AArch64::STRXpre: 272 case AArch64::STURXi: 273 case AArch64::LDRDui: 274 case AArch64::LDURDi: 275 case AArch64::LDRDpre: 276 case AArch64::LDRQui: 277 case AArch64::LDURQi: 278 case AArch64::LDRQpre: 279 case AArch64::LDRWui: 280 case AArch64::LDURWi: 281 case AArch64::LDRWpre: 282 case AArch64::LDRXui: 283 case AArch64::LDURXi: 284 case AArch64::LDRXpre: 285 case AArch64::STRSui: 286 case AArch64::STURSi: 287 case AArch64::STRSpre: 288 case AArch64::LDRSui: 289 case AArch64::LDURSi: 290 case AArch64::LDRSpre: 291 return Opc; 292 case AArch64::LDRSWui: 293 return AArch64::LDRWui; 294 case AArch64::LDURSWi: 295 return AArch64::LDURWi; 296 } 297 } 298 299 static unsigned getMatchingWideOpcode(unsigned Opc) { 300 switch (Opc) { 301 default: 302 llvm_unreachable("Opcode has no wide equivalent!"); 303 case AArch64::STRBBui: 304 return AArch64::STRHHui; 305 case AArch64::STRHHui: 306 return AArch64::STRWui; 307 case AArch64::STURBBi: 308 return AArch64::STURHHi; 309 case AArch64::STURHHi: 310 return AArch64::STURWi; 311 case AArch64::STURWi: 312 return AArch64::STURXi; 313 case AArch64::STRWui: 314 return AArch64::STRXui; 315 } 316 } 317 318 static unsigned getMatchingPairOpcode(unsigned Opc) { 319 switch (Opc) { 320 default: 321 llvm_unreachable("Opcode has no pairwise equivalent!"); 322 case AArch64::STRSui: 323 case AArch64::STURSi: 324 return AArch64::STPSi; 325 case AArch64::STRSpre: 326 return AArch64::STPSpre; 327 case AArch64::STRDui: 328 case AArch64::STURDi: 329 return AArch64::STPDi; 330 case AArch64::STRDpre: 331 return AArch64::STPDpre; 332 case AArch64::STRQui: 333 case AArch64::STURQi: 334 return AArch64::STPQi; 335 case AArch64::STRQpre: 336 return AArch64::STPQpre; 337 case AArch64::STRWui: 338 case AArch64::STURWi: 339 return AArch64::STPWi; 340 case AArch64::STRWpre: 341 return AArch64::STPWpre; 342 case AArch64::STRXui: 343 case AArch64::STURXi: 344 return AArch64::STPXi; 345 case AArch64::STRXpre: 346 return AArch64::STPXpre; 347 case AArch64::LDRSui: 348 case AArch64::LDURSi: 349 return AArch64::LDPSi; 350 case AArch64::LDRSpre: 351 return AArch64::LDPSpre; 352 case AArch64::LDRDui: 353 case AArch64::LDURDi: 354 return AArch64::LDPDi; 355 case AArch64::LDRDpre: 356 return AArch64::LDPDpre; 357 case AArch64::LDRQui: 358 case AArch64::LDURQi: 359 return AArch64::LDPQi; 360 case AArch64::LDRQpre: 361 return AArch64::LDPQpre; 362 case AArch64::LDRWui: 363 case AArch64::LDURWi: 364 return AArch64::LDPWi; 365 case AArch64::LDRWpre: 366 return AArch64::LDPWpre; 367 case AArch64::LDRXui: 368 case AArch64::LDURXi: 369 return AArch64::LDPXi; 370 case AArch64::LDRXpre: 371 return AArch64::LDPXpre; 372 case AArch64::LDRSWui: 373 case AArch64::LDURSWi: 374 return AArch64::LDPSWi; 375 } 376 } 377 378 static unsigned isMatchingStore(MachineInstr &LoadInst, 379 MachineInstr &StoreInst) { 380 unsigned LdOpc = LoadInst.getOpcode(); 381 unsigned StOpc = StoreInst.getOpcode(); 382 switch (LdOpc) { 383 default: 384 llvm_unreachable("Unsupported load instruction!"); 385 case AArch64::LDRBBui: 386 return StOpc == AArch64::STRBBui || StOpc == AArch64::STRHHui || 387 StOpc == AArch64::STRWui || StOpc == AArch64::STRXui; 388 case AArch64::LDURBBi: 389 return StOpc == AArch64::STURBBi || StOpc == AArch64::STURHHi || 390 StOpc == AArch64::STURWi || StOpc == AArch64::STURXi; 391 case AArch64::LDRHHui: 392 return StOpc == AArch64::STRHHui || StOpc == AArch64::STRWui || 393 StOpc == AArch64::STRXui; 394 case AArch64::LDURHHi: 395 return StOpc == AArch64::STURHHi || StOpc == AArch64::STURWi || 396 StOpc == AArch64::STURXi; 397 case AArch64::LDRWui: 398 return StOpc == AArch64::STRWui || StOpc == AArch64::STRXui; 399 case AArch64::LDURWi: 400 return StOpc == AArch64::STURWi || StOpc == AArch64::STURXi; 401 case AArch64::LDRXui: 402 return StOpc == AArch64::STRXui; 403 case AArch64::LDURXi: 404 return StOpc == AArch64::STURXi; 405 } 406 } 407 408 static unsigned getPreIndexedOpcode(unsigned Opc) { 409 // FIXME: We don't currently support creating pre-indexed loads/stores when 410 // the load or store is the unscaled version. If we decide to perform such an 411 // optimization in the future the cases for the unscaled loads/stores will 412 // need to be added here. 413 switch (Opc) { 414 default: 415 llvm_unreachable("Opcode has no pre-indexed equivalent!"); 416 case AArch64::STRSui: 417 return AArch64::STRSpre; 418 case AArch64::STRDui: 419 return AArch64::STRDpre; 420 case AArch64::STRQui: 421 return AArch64::STRQpre; 422 case AArch64::STRBBui: 423 return AArch64::STRBBpre; 424 case AArch64::STRHHui: 425 return AArch64::STRHHpre; 426 case AArch64::STRWui: 427 return AArch64::STRWpre; 428 case AArch64::STRXui: 429 return AArch64::STRXpre; 430 case AArch64::LDRSui: 431 return AArch64::LDRSpre; 432 case AArch64::LDRDui: 433 return AArch64::LDRDpre; 434 case AArch64::LDRQui: 435 return AArch64::LDRQpre; 436 case AArch64::LDRBBui: 437 return AArch64::LDRBBpre; 438 case AArch64::LDRHHui: 439 return AArch64::LDRHHpre; 440 case AArch64::LDRWui: 441 return AArch64::LDRWpre; 442 case AArch64::LDRXui: 443 return AArch64::LDRXpre; 444 case AArch64::LDRSWui: 445 return AArch64::LDRSWpre; 446 case AArch64::LDPSi: 447 return AArch64::LDPSpre; 448 case AArch64::LDPSWi: 449 return AArch64::LDPSWpre; 450 case AArch64::LDPDi: 451 return AArch64::LDPDpre; 452 case AArch64::LDPQi: 453 return AArch64::LDPQpre; 454 case AArch64::LDPWi: 455 return AArch64::LDPWpre; 456 case AArch64::LDPXi: 457 return AArch64::LDPXpre; 458 case AArch64::STPSi: 459 return AArch64::STPSpre; 460 case AArch64::STPDi: 461 return AArch64::STPDpre; 462 case AArch64::STPQi: 463 return AArch64::STPQpre; 464 case AArch64::STPWi: 465 return AArch64::STPWpre; 466 case AArch64::STPXi: 467 return AArch64::STPXpre; 468 case AArch64::STGi: 469 return AArch64::STGPreIndex; 470 case AArch64::STZGi: 471 return AArch64::STZGPreIndex; 472 case AArch64::ST2Gi: 473 return AArch64::ST2GPreIndex; 474 case AArch64::STZ2Gi: 475 return AArch64::STZ2GPreIndex; 476 case AArch64::STGPi: 477 return AArch64::STGPpre; 478 } 479 } 480 481 static unsigned getPostIndexedOpcode(unsigned Opc) { 482 switch (Opc) { 483 default: 484 llvm_unreachable("Opcode has no post-indexed wise equivalent!"); 485 case AArch64::STRSui: 486 case AArch64::STURSi: 487 return AArch64::STRSpost; 488 case AArch64::STRDui: 489 case AArch64::STURDi: 490 return AArch64::STRDpost; 491 case AArch64::STRQui: 492 case AArch64::STURQi: 493 return AArch64::STRQpost; 494 case AArch64::STRBBui: 495 return AArch64::STRBBpost; 496 case AArch64::STRHHui: 497 return AArch64::STRHHpost; 498 case AArch64::STRWui: 499 case AArch64::STURWi: 500 return AArch64::STRWpost; 501 case AArch64::STRXui: 502 case AArch64::STURXi: 503 return AArch64::STRXpost; 504 case AArch64::LDRSui: 505 case AArch64::LDURSi: 506 return AArch64::LDRSpost; 507 case AArch64::LDRDui: 508 case AArch64::LDURDi: 509 return AArch64::LDRDpost; 510 case AArch64::LDRQui: 511 case AArch64::LDURQi: 512 return AArch64::LDRQpost; 513 case AArch64::LDRBBui: 514 return AArch64::LDRBBpost; 515 case AArch64::LDRHHui: 516 return AArch64::LDRHHpost; 517 case AArch64::LDRWui: 518 case AArch64::LDURWi: 519 return AArch64::LDRWpost; 520 case AArch64::LDRXui: 521 case AArch64::LDURXi: 522 return AArch64::LDRXpost; 523 case AArch64::LDRSWui: 524 return AArch64::LDRSWpost; 525 case AArch64::LDPSi: 526 return AArch64::LDPSpost; 527 case AArch64::LDPSWi: 528 return AArch64::LDPSWpost; 529 case AArch64::LDPDi: 530 return AArch64::LDPDpost; 531 case AArch64::LDPQi: 532 return AArch64::LDPQpost; 533 case AArch64::LDPWi: 534 return AArch64::LDPWpost; 535 case AArch64::LDPXi: 536 return AArch64::LDPXpost; 537 case AArch64::STPSi: 538 return AArch64::STPSpost; 539 case AArch64::STPDi: 540 return AArch64::STPDpost; 541 case AArch64::STPQi: 542 return AArch64::STPQpost; 543 case AArch64::STPWi: 544 return AArch64::STPWpost; 545 case AArch64::STPXi: 546 return AArch64::STPXpost; 547 case AArch64::STGi: 548 return AArch64::STGPostIndex; 549 case AArch64::STZGi: 550 return AArch64::STZGPostIndex; 551 case AArch64::ST2Gi: 552 return AArch64::ST2GPostIndex; 553 case AArch64::STZ2Gi: 554 return AArch64::STZ2GPostIndex; 555 case AArch64::STGPi: 556 return AArch64::STGPpost; 557 } 558 } 559 560 static bool isPreLdStPairCandidate(MachineInstr &FirstMI, MachineInstr &MI) { 561 562 unsigned OpcA = FirstMI.getOpcode(); 563 unsigned OpcB = MI.getOpcode(); 564 565 switch (OpcA) { 566 default: 567 return false; 568 case AArch64::STRSpre: 569 return (OpcB == AArch64::STRSui) || (OpcB == AArch64::STURSi); 570 case AArch64::STRDpre: 571 return (OpcB == AArch64::STRDui) || (OpcB == AArch64::STURDi); 572 case AArch64::STRQpre: 573 return (OpcB == AArch64::STRQui) || (OpcB == AArch64::STURQi); 574 case AArch64::STRWpre: 575 return (OpcB == AArch64::STRWui) || (OpcB == AArch64::STURWi); 576 case AArch64::STRXpre: 577 return (OpcB == AArch64::STRXui) || (OpcB == AArch64::STURXi); 578 case AArch64::LDRSpre: 579 return (OpcB == AArch64::LDRSui) || (OpcB == AArch64::LDURSi); 580 case AArch64::LDRDpre: 581 return (OpcB == AArch64::LDRDui) || (OpcB == AArch64::LDURDi); 582 case AArch64::LDRQpre: 583 return (OpcB == AArch64::LDRQui) || (OpcB == AArch64::LDURQi); 584 case AArch64::LDRWpre: 585 return (OpcB == AArch64::LDRWui) || (OpcB == AArch64::LDURWi); 586 case AArch64::LDRXpre: 587 return (OpcB == AArch64::LDRXui) || (OpcB == AArch64::LDURXi); 588 } 589 } 590 591 // Returns the scale and offset range of pre/post indexed variants of MI. 592 static void getPrePostIndexedMemOpInfo(const MachineInstr &MI, int &Scale, 593 int &MinOffset, int &MaxOffset) { 594 bool IsPaired = AArch64InstrInfo::isPairedLdSt(MI); 595 bool IsTagStore = isTagStore(MI); 596 // ST*G and all paired ldst have the same scale in pre/post-indexed variants 597 // as in the "unsigned offset" variant. 598 // All other pre/post indexed ldst instructions are unscaled. 599 Scale = (IsTagStore || IsPaired) ? AArch64InstrInfo::getMemScale(MI) : 1; 600 601 if (IsPaired) { 602 MinOffset = -64; 603 MaxOffset = 63; 604 } else { 605 MinOffset = -256; 606 MaxOffset = 255; 607 } 608 } 609 610 static MachineOperand &getLdStRegOp(MachineInstr &MI, 611 unsigned PairedRegOp = 0) { 612 assert(PairedRegOp < 2 && "Unexpected register operand idx."); 613 bool IsPreLdSt = AArch64InstrInfo::isPreLdSt(MI); 614 if (IsPreLdSt) 615 PairedRegOp += 1; 616 unsigned Idx = 617 AArch64InstrInfo::isPairedLdSt(MI) || IsPreLdSt ? PairedRegOp : 0; 618 return MI.getOperand(Idx); 619 } 620 621 static bool isLdOffsetInRangeOfSt(MachineInstr &LoadInst, 622 MachineInstr &StoreInst, 623 const AArch64InstrInfo *TII) { 624 assert(isMatchingStore(LoadInst, StoreInst) && "Expect only matched ld/st."); 625 int LoadSize = TII->getMemScale(LoadInst); 626 int StoreSize = TII->getMemScale(StoreInst); 627 int UnscaledStOffset = 628 TII->hasUnscaledLdStOffset(StoreInst) 629 ? AArch64InstrInfo::getLdStOffsetOp(StoreInst).getImm() 630 : AArch64InstrInfo::getLdStOffsetOp(StoreInst).getImm() * StoreSize; 631 int UnscaledLdOffset = 632 TII->hasUnscaledLdStOffset(LoadInst) 633 ? AArch64InstrInfo::getLdStOffsetOp(LoadInst).getImm() 634 : AArch64InstrInfo::getLdStOffsetOp(LoadInst).getImm() * LoadSize; 635 return (UnscaledStOffset <= UnscaledLdOffset) && 636 (UnscaledLdOffset + LoadSize <= (UnscaledStOffset + StoreSize)); 637 } 638 639 static bool isPromotableZeroStoreInst(MachineInstr &MI) { 640 unsigned Opc = MI.getOpcode(); 641 return (Opc == AArch64::STRWui || Opc == AArch64::STURWi || 642 isNarrowStore(Opc)) && 643 getLdStRegOp(MI).getReg() == AArch64::WZR; 644 } 645 646 static bool isPromotableLoadFromStore(MachineInstr &MI) { 647 switch (MI.getOpcode()) { 648 default: 649 return false; 650 // Scaled instructions. 651 case AArch64::LDRBBui: 652 case AArch64::LDRHHui: 653 case AArch64::LDRWui: 654 case AArch64::LDRXui: 655 // Unscaled instructions. 656 case AArch64::LDURBBi: 657 case AArch64::LDURHHi: 658 case AArch64::LDURWi: 659 case AArch64::LDURXi: 660 return true; 661 } 662 } 663 664 static bool isMergeableLdStUpdate(MachineInstr &MI) { 665 unsigned Opc = MI.getOpcode(); 666 switch (Opc) { 667 default: 668 return false; 669 // Scaled instructions. 670 case AArch64::STRSui: 671 case AArch64::STRDui: 672 case AArch64::STRQui: 673 case AArch64::STRXui: 674 case AArch64::STRWui: 675 case AArch64::STRHHui: 676 case AArch64::STRBBui: 677 case AArch64::LDRSui: 678 case AArch64::LDRDui: 679 case AArch64::LDRQui: 680 case AArch64::LDRXui: 681 case AArch64::LDRWui: 682 case AArch64::LDRHHui: 683 case AArch64::LDRBBui: 684 case AArch64::STGi: 685 case AArch64::STZGi: 686 case AArch64::ST2Gi: 687 case AArch64::STZ2Gi: 688 case AArch64::STGPi: 689 // Unscaled instructions. 690 case AArch64::STURSi: 691 case AArch64::STURDi: 692 case AArch64::STURQi: 693 case AArch64::STURWi: 694 case AArch64::STURXi: 695 case AArch64::LDURSi: 696 case AArch64::LDURDi: 697 case AArch64::LDURQi: 698 case AArch64::LDURWi: 699 case AArch64::LDURXi: 700 // Paired instructions. 701 case AArch64::LDPSi: 702 case AArch64::LDPSWi: 703 case AArch64::LDPDi: 704 case AArch64::LDPQi: 705 case AArch64::LDPWi: 706 case AArch64::LDPXi: 707 case AArch64::STPSi: 708 case AArch64::STPDi: 709 case AArch64::STPQi: 710 case AArch64::STPWi: 711 case AArch64::STPXi: 712 // Make sure this is a reg+imm (as opposed to an address reloc). 713 if (!AArch64InstrInfo::getLdStOffsetOp(MI).isImm()) 714 return false; 715 716 return true; 717 } 718 } 719 720 MachineBasicBlock::iterator 721 AArch64LoadStoreOpt::mergeNarrowZeroStores(MachineBasicBlock::iterator I, 722 MachineBasicBlock::iterator MergeMI, 723 const LdStPairFlags &Flags) { 724 assert(isPromotableZeroStoreInst(*I) && isPromotableZeroStoreInst(*MergeMI) && 725 "Expected promotable zero stores."); 726 727 MachineBasicBlock::iterator E = I->getParent()->end(); 728 MachineBasicBlock::iterator NextI = next_nodbg(I, E); 729 // If NextI is the second of the two instructions to be merged, we need 730 // to skip one further. Either way we merge will invalidate the iterator, 731 // and we don't need to scan the new instruction, as it's a pairwise 732 // instruction, which we're not considering for further action anyway. 733 if (NextI == MergeMI) 734 NextI = next_nodbg(NextI, E); 735 736 unsigned Opc = I->getOpcode(); 737 unsigned MergeMIOpc = MergeMI->getOpcode(); 738 bool IsScaled = !TII->hasUnscaledLdStOffset(Opc); 739 bool IsMergedMIScaled = !TII->hasUnscaledLdStOffset(MergeMIOpc); 740 int OffsetStride = IsScaled ? TII->getMemScale(*I) : 1; 741 int MergeMIOffsetStride = IsMergedMIScaled ? TII->getMemScale(*MergeMI) : 1; 742 743 bool MergeForward = Flags.getMergeForward(); 744 // Insert our new paired instruction after whichever of the paired 745 // instructions MergeForward indicates. 746 MachineBasicBlock::iterator InsertionPoint = MergeForward ? MergeMI : I; 747 // Also based on MergeForward is from where we copy the base register operand 748 // so we get the flags compatible with the input code. 749 const MachineOperand &BaseRegOp = 750 MergeForward ? AArch64InstrInfo::getLdStBaseOp(*MergeMI) 751 : AArch64InstrInfo::getLdStBaseOp(*I); 752 753 // Which register is Rt and which is Rt2 depends on the offset order. 754 int64_t IOffsetInBytes = 755 AArch64InstrInfo::getLdStOffsetOp(*I).getImm() * OffsetStride; 756 int64_t MIOffsetInBytes = 757 AArch64InstrInfo::getLdStOffsetOp(*MergeMI).getImm() * 758 MergeMIOffsetStride; 759 // Select final offset based on the offset order. 760 int64_t OffsetImm; 761 if (IOffsetInBytes > MIOffsetInBytes) 762 OffsetImm = MIOffsetInBytes; 763 else 764 OffsetImm = IOffsetInBytes; 765 766 int NewOpcode = getMatchingWideOpcode(Opc); 767 bool FinalIsScaled = !TII->hasUnscaledLdStOffset(NewOpcode); 768 769 // Adjust final offset if the result opcode is a scaled store. 770 if (FinalIsScaled) { 771 int NewOffsetStride = FinalIsScaled ? TII->getMemScale(NewOpcode) : 1; 772 assert(((OffsetImm % NewOffsetStride) == 0) && 773 "Offset should be a multiple of the store memory scale"); 774 OffsetImm = OffsetImm / NewOffsetStride; 775 } 776 777 // Construct the new instruction. 778 DebugLoc DL = I->getDebugLoc(); 779 MachineBasicBlock *MBB = I->getParent(); 780 MachineInstrBuilder MIB; 781 MIB = BuildMI(*MBB, InsertionPoint, DL, TII->get(getMatchingWideOpcode(Opc))) 782 .addReg(isNarrowStore(Opc) ? AArch64::WZR : AArch64::XZR) 783 .add(BaseRegOp) 784 .addImm(OffsetImm) 785 .cloneMergedMemRefs({&*I, &*MergeMI}) 786 .setMIFlags(I->mergeFlagsWith(*MergeMI)); 787 (void)MIB; 788 789 LLVM_DEBUG(dbgs() << "Creating wider store. Replacing instructions:\n "); 790 LLVM_DEBUG(I->print(dbgs())); 791 LLVM_DEBUG(dbgs() << " "); 792 LLVM_DEBUG(MergeMI->print(dbgs())); 793 LLVM_DEBUG(dbgs() << " with instruction:\n "); 794 LLVM_DEBUG(((MachineInstr *)MIB)->print(dbgs())); 795 LLVM_DEBUG(dbgs() << "\n"); 796 797 // Erase the old instructions. 798 I->eraseFromParent(); 799 MergeMI->eraseFromParent(); 800 return NextI; 801 } 802 803 // Apply Fn to all instructions between MI and the beginning of the block, until 804 // a def for DefReg is reached. Returns true, iff Fn returns true for all 805 // visited instructions. Stop after visiting Limit iterations. 806 static bool forAllMIsUntilDef(MachineInstr &MI, MCPhysReg DefReg, 807 const TargetRegisterInfo *TRI, unsigned Limit, 808 std::function<bool(MachineInstr &, bool)> &Fn) { 809 auto MBB = MI.getParent(); 810 for (MachineInstr &I : 811 instructionsWithoutDebug(MI.getReverseIterator(), MBB->instr_rend())) { 812 if (!Limit) 813 return false; 814 --Limit; 815 816 bool isDef = any_of(I.operands(), [DefReg, TRI](MachineOperand &MOP) { 817 return MOP.isReg() && MOP.isDef() && !MOP.isDebug() && MOP.getReg() && 818 TRI->regsOverlap(MOP.getReg(), DefReg); 819 }); 820 if (!Fn(I, isDef)) 821 return false; 822 if (isDef) 823 break; 824 } 825 return true; 826 } 827 828 static void updateDefinedRegisters(MachineInstr &MI, LiveRegUnits &Units, 829 const TargetRegisterInfo *TRI) { 830 831 for (const MachineOperand &MOP : phys_regs_and_masks(MI)) 832 if (MOP.isReg() && MOP.isKill()) 833 Units.removeReg(MOP.getReg()); 834 835 for (const MachineOperand &MOP : phys_regs_and_masks(MI)) 836 if (MOP.isReg() && !MOP.isKill()) 837 Units.addReg(MOP.getReg()); 838 } 839 840 MachineBasicBlock::iterator 841 AArch64LoadStoreOpt::mergePairedInsns(MachineBasicBlock::iterator I, 842 MachineBasicBlock::iterator Paired, 843 const LdStPairFlags &Flags) { 844 MachineBasicBlock::iterator E = I->getParent()->end(); 845 MachineBasicBlock::iterator NextI = next_nodbg(I, E); 846 // If NextI is the second of the two instructions to be merged, we need 847 // to skip one further. Either way we merge will invalidate the iterator, 848 // and we don't need to scan the new instruction, as it's a pairwise 849 // instruction, which we're not considering for further action anyway. 850 if (NextI == Paired) 851 NextI = next_nodbg(NextI, E); 852 853 int SExtIdx = Flags.getSExtIdx(); 854 unsigned Opc = 855 SExtIdx == -1 ? I->getOpcode() : getMatchingNonSExtOpcode(I->getOpcode()); 856 bool IsUnscaled = TII->hasUnscaledLdStOffset(Opc); 857 int OffsetStride = IsUnscaled ? TII->getMemScale(*I) : 1; 858 859 bool MergeForward = Flags.getMergeForward(); 860 861 std::optional<MCPhysReg> RenameReg = Flags.getRenameReg(); 862 if (MergeForward && RenameReg) { 863 MCRegister RegToRename = getLdStRegOp(*I).getReg(); 864 DefinedInBB.addReg(*RenameReg); 865 866 // Return the sub/super register for RenameReg, matching the size of 867 // OriginalReg. 868 auto GetMatchingSubReg = [this, 869 RenameReg](MCPhysReg OriginalReg) -> MCPhysReg { 870 for (MCPhysReg SubOrSuper : TRI->sub_and_superregs_inclusive(*RenameReg)) 871 if (TRI->getMinimalPhysRegClass(OriginalReg) == 872 TRI->getMinimalPhysRegClass(SubOrSuper)) 873 return SubOrSuper; 874 llvm_unreachable("Should have found matching sub or super register!"); 875 }; 876 877 std::function<bool(MachineInstr &, bool)> UpdateMIs = 878 [this, RegToRename, GetMatchingSubReg](MachineInstr &MI, bool IsDef) { 879 if (IsDef) { 880 bool SeenDef = false; 881 for (auto &MOP : MI.operands()) { 882 // Rename the first explicit definition and all implicit 883 // definitions matching RegToRename. 884 if (MOP.isReg() && !MOP.isDebug() && MOP.getReg() && 885 (!SeenDef || (MOP.isDef() && MOP.isImplicit())) && 886 TRI->regsOverlap(MOP.getReg(), RegToRename)) { 887 assert((MOP.isImplicit() || 888 (MOP.isRenamable() && !MOP.isEarlyClobber())) && 889 "Need renamable operands"); 890 MOP.setReg(GetMatchingSubReg(MOP.getReg())); 891 SeenDef = true; 892 } 893 } 894 } else { 895 for (auto &MOP : MI.operands()) { 896 if (MOP.isReg() && !MOP.isDebug() && MOP.getReg() && 897 TRI->regsOverlap(MOP.getReg(), RegToRename)) { 898 assert((MOP.isImplicit() || 899 (MOP.isRenamable() && !MOP.isEarlyClobber())) && 900 "Need renamable operands"); 901 MOP.setReg(GetMatchingSubReg(MOP.getReg())); 902 } 903 } 904 } 905 LLVM_DEBUG(dbgs() << "Renamed " << MI << "\n"); 906 return true; 907 }; 908 forAllMIsUntilDef(*I, RegToRename, TRI, LdStLimit, UpdateMIs); 909 910 #if !defined(NDEBUG) 911 // Make sure the register used for renaming is not used between the paired 912 // instructions. That would trash the content before the new paired 913 // instruction. 914 for (auto &MI : 915 iterator_range<MachineInstrBundleIterator<llvm::MachineInstr>>( 916 std::next(I), std::next(Paired))) 917 assert(all_of(MI.operands(), 918 [this, &RenameReg](const MachineOperand &MOP) { 919 return !MOP.isReg() || MOP.isDebug() || !MOP.getReg() || 920 MOP.isUndef() || 921 !TRI->regsOverlap(MOP.getReg(), *RenameReg); 922 }) && 923 "Rename register used between paired instruction, trashing the " 924 "content"); 925 #endif 926 } 927 928 // Insert our new paired instruction after whichever of the paired 929 // instructions MergeForward indicates. 930 MachineBasicBlock::iterator InsertionPoint = MergeForward ? Paired : I; 931 // Also based on MergeForward is from where we copy the base register operand 932 // so we get the flags compatible with the input code. 933 const MachineOperand &BaseRegOp = 934 MergeForward ? AArch64InstrInfo::getLdStBaseOp(*Paired) 935 : AArch64InstrInfo::getLdStBaseOp(*I); 936 937 int Offset = AArch64InstrInfo::getLdStOffsetOp(*I).getImm(); 938 int PairedOffset = AArch64InstrInfo::getLdStOffsetOp(*Paired).getImm(); 939 bool PairedIsUnscaled = TII->hasUnscaledLdStOffset(Paired->getOpcode()); 940 if (IsUnscaled != PairedIsUnscaled) { 941 // We're trying to pair instructions that differ in how they are scaled. If 942 // I is scaled then scale the offset of Paired accordingly. Otherwise, do 943 // the opposite (i.e., make Paired's offset unscaled). 944 int MemSize = TII->getMemScale(*Paired); 945 if (PairedIsUnscaled) { 946 // If the unscaled offset isn't a multiple of the MemSize, we can't 947 // pair the operations together. 948 assert(!(PairedOffset % TII->getMemScale(*Paired)) && 949 "Offset should be a multiple of the stride!"); 950 PairedOffset /= MemSize; 951 } else { 952 PairedOffset *= MemSize; 953 } 954 } 955 956 // Which register is Rt and which is Rt2 depends on the offset order. 957 // However, for pre load/stores the Rt should be the one of the pre 958 // load/store. 959 MachineInstr *RtMI, *Rt2MI; 960 if (Offset == PairedOffset + OffsetStride && 961 !AArch64InstrInfo::isPreLdSt(*I)) { 962 RtMI = &*Paired; 963 Rt2MI = &*I; 964 // Here we swapped the assumption made for SExtIdx. 965 // I.e., we turn ldp I, Paired into ldp Paired, I. 966 // Update the index accordingly. 967 if (SExtIdx != -1) 968 SExtIdx = (SExtIdx + 1) % 2; 969 } else { 970 RtMI = &*I; 971 Rt2MI = &*Paired; 972 } 973 int OffsetImm = AArch64InstrInfo::getLdStOffsetOp(*RtMI).getImm(); 974 // Scale the immediate offset, if necessary. 975 if (TII->hasUnscaledLdStOffset(RtMI->getOpcode())) { 976 assert(!(OffsetImm % TII->getMemScale(*RtMI)) && 977 "Unscaled offset cannot be scaled."); 978 OffsetImm /= TII->getMemScale(*RtMI); 979 } 980 981 // Construct the new instruction. 982 MachineInstrBuilder MIB; 983 DebugLoc DL = I->getDebugLoc(); 984 MachineBasicBlock *MBB = I->getParent(); 985 MachineOperand RegOp0 = getLdStRegOp(*RtMI); 986 MachineOperand RegOp1 = getLdStRegOp(*Rt2MI); 987 // Kill flags may become invalid when moving stores for pairing. 988 if (RegOp0.isUse()) { 989 if (!MergeForward) { 990 // Clear kill flags on store if moving upwards. Example: 991 // STRWui %w0, ... 992 // USE %w1 993 // STRWui kill %w1 ; need to clear kill flag when moving STRWui upwards 994 RegOp0.setIsKill(false); 995 RegOp1.setIsKill(false); 996 } else { 997 // Clear kill flags of the first stores register. Example: 998 // STRWui %w1, ... 999 // USE kill %w1 ; need to clear kill flag when moving STRWui downwards 1000 // STRW %w0 1001 Register Reg = getLdStRegOp(*I).getReg(); 1002 for (MachineInstr &MI : make_range(std::next(I), Paired)) 1003 MI.clearRegisterKills(Reg, TRI); 1004 } 1005 } 1006 1007 unsigned int MatchPairOpcode = getMatchingPairOpcode(Opc); 1008 MIB = BuildMI(*MBB, InsertionPoint, DL, TII->get(MatchPairOpcode)); 1009 1010 // Adds the pre-index operand for pre-indexed ld/st pairs. 1011 if (AArch64InstrInfo::isPreLdSt(*RtMI)) 1012 MIB.addReg(BaseRegOp.getReg(), RegState::Define); 1013 1014 MIB.add(RegOp0) 1015 .add(RegOp1) 1016 .add(BaseRegOp) 1017 .addImm(OffsetImm) 1018 .cloneMergedMemRefs({&*I, &*Paired}) 1019 .setMIFlags(I->mergeFlagsWith(*Paired)); 1020 1021 (void)MIB; 1022 1023 LLVM_DEBUG( 1024 dbgs() << "Creating pair load/store. Replacing instructions:\n "); 1025 LLVM_DEBUG(I->print(dbgs())); 1026 LLVM_DEBUG(dbgs() << " "); 1027 LLVM_DEBUG(Paired->print(dbgs())); 1028 LLVM_DEBUG(dbgs() << " with instruction:\n "); 1029 if (SExtIdx != -1) { 1030 // Generate the sign extension for the proper result of the ldp. 1031 // I.e., with X1, that would be: 1032 // %w1 = KILL %w1, implicit-def %x1 1033 // %x1 = SBFMXri killed %x1, 0, 31 1034 MachineOperand &DstMO = MIB->getOperand(SExtIdx); 1035 // Right now, DstMO has the extended register, since it comes from an 1036 // extended opcode. 1037 Register DstRegX = DstMO.getReg(); 1038 // Get the W variant of that register. 1039 Register DstRegW = TRI->getSubReg(DstRegX, AArch64::sub_32); 1040 // Update the result of LDP to use the W instead of the X variant. 1041 DstMO.setReg(DstRegW); 1042 LLVM_DEBUG(((MachineInstr *)MIB)->print(dbgs())); 1043 LLVM_DEBUG(dbgs() << "\n"); 1044 // Make the machine verifier happy by providing a definition for 1045 // the X register. 1046 // Insert this definition right after the generated LDP, i.e., before 1047 // InsertionPoint. 1048 MachineInstrBuilder MIBKill = 1049 BuildMI(*MBB, InsertionPoint, DL, TII->get(TargetOpcode::KILL), DstRegW) 1050 .addReg(DstRegW) 1051 .addReg(DstRegX, RegState::Define); 1052 MIBKill->getOperand(2).setImplicit(); 1053 // Create the sign extension. 1054 MachineInstrBuilder MIBSXTW = 1055 BuildMI(*MBB, InsertionPoint, DL, TII->get(AArch64::SBFMXri), DstRegX) 1056 .addReg(DstRegX) 1057 .addImm(0) 1058 .addImm(31); 1059 (void)MIBSXTW; 1060 LLVM_DEBUG(dbgs() << " Extend operand:\n "); 1061 LLVM_DEBUG(((MachineInstr *)MIBSXTW)->print(dbgs())); 1062 } else { 1063 LLVM_DEBUG(((MachineInstr *)MIB)->print(dbgs())); 1064 } 1065 LLVM_DEBUG(dbgs() << "\n"); 1066 1067 if (MergeForward) 1068 for (const MachineOperand &MOP : phys_regs_and_masks(*I)) 1069 if (MOP.isReg() && MOP.isKill()) 1070 DefinedInBB.addReg(MOP.getReg()); 1071 1072 // Erase the old instructions. 1073 I->eraseFromParent(); 1074 Paired->eraseFromParent(); 1075 1076 return NextI; 1077 } 1078 1079 MachineBasicBlock::iterator 1080 AArch64LoadStoreOpt::promoteLoadFromStore(MachineBasicBlock::iterator LoadI, 1081 MachineBasicBlock::iterator StoreI) { 1082 MachineBasicBlock::iterator NextI = 1083 next_nodbg(LoadI, LoadI->getParent()->end()); 1084 1085 int LoadSize = TII->getMemScale(*LoadI); 1086 int StoreSize = TII->getMemScale(*StoreI); 1087 Register LdRt = getLdStRegOp(*LoadI).getReg(); 1088 const MachineOperand &StMO = getLdStRegOp(*StoreI); 1089 Register StRt = getLdStRegOp(*StoreI).getReg(); 1090 bool IsStoreXReg = TRI->getRegClass(AArch64::GPR64RegClassID)->contains(StRt); 1091 1092 assert((IsStoreXReg || 1093 TRI->getRegClass(AArch64::GPR32RegClassID)->contains(StRt)) && 1094 "Unexpected RegClass"); 1095 1096 MachineInstr *BitExtMI; 1097 if (LoadSize == StoreSize && (LoadSize == 4 || LoadSize == 8)) { 1098 // Remove the load, if the destination register of the loads is the same 1099 // register for stored value. 1100 if (StRt == LdRt && LoadSize == 8) { 1101 for (MachineInstr &MI : make_range(StoreI->getIterator(), 1102 LoadI->getIterator())) { 1103 if (MI.killsRegister(StRt, TRI)) { 1104 MI.clearRegisterKills(StRt, TRI); 1105 break; 1106 } 1107 } 1108 LLVM_DEBUG(dbgs() << "Remove load instruction:\n "); 1109 LLVM_DEBUG(LoadI->print(dbgs())); 1110 LLVM_DEBUG(dbgs() << "\n"); 1111 LoadI->eraseFromParent(); 1112 return NextI; 1113 } 1114 // Replace the load with a mov if the load and store are in the same size. 1115 BitExtMI = 1116 BuildMI(*LoadI->getParent(), LoadI, LoadI->getDebugLoc(), 1117 TII->get(IsStoreXReg ? AArch64::ORRXrs : AArch64::ORRWrs), LdRt) 1118 .addReg(IsStoreXReg ? AArch64::XZR : AArch64::WZR) 1119 .add(StMO) 1120 .addImm(AArch64_AM::getShifterImm(AArch64_AM::LSL, 0)) 1121 .setMIFlags(LoadI->getFlags()); 1122 } else { 1123 // FIXME: Currently we disable this transformation in big-endian targets as 1124 // performance and correctness are verified only in little-endian. 1125 if (!Subtarget->isLittleEndian()) 1126 return NextI; 1127 bool IsUnscaled = TII->hasUnscaledLdStOffset(*LoadI); 1128 assert(IsUnscaled == TII->hasUnscaledLdStOffset(*StoreI) && 1129 "Unsupported ld/st match"); 1130 assert(LoadSize <= StoreSize && "Invalid load size"); 1131 int UnscaledLdOffset = 1132 IsUnscaled 1133 ? AArch64InstrInfo::getLdStOffsetOp(*LoadI).getImm() 1134 : AArch64InstrInfo::getLdStOffsetOp(*LoadI).getImm() * LoadSize; 1135 int UnscaledStOffset = 1136 IsUnscaled 1137 ? AArch64InstrInfo::getLdStOffsetOp(*StoreI).getImm() 1138 : AArch64InstrInfo::getLdStOffsetOp(*StoreI).getImm() * StoreSize; 1139 int Width = LoadSize * 8; 1140 Register DestReg = 1141 IsStoreXReg ? Register(TRI->getMatchingSuperReg( 1142 LdRt, AArch64::sub_32, &AArch64::GPR64RegClass)) 1143 : LdRt; 1144 1145 assert((UnscaledLdOffset >= UnscaledStOffset && 1146 (UnscaledLdOffset + LoadSize) <= UnscaledStOffset + StoreSize) && 1147 "Invalid offset"); 1148 1149 int Immr = 8 * (UnscaledLdOffset - UnscaledStOffset); 1150 int Imms = Immr + Width - 1; 1151 if (UnscaledLdOffset == UnscaledStOffset) { 1152 uint32_t AndMaskEncoded = ((IsStoreXReg ? 1 : 0) << 12) // N 1153 | ((Immr) << 6) // immr 1154 | ((Imms) << 0) // imms 1155 ; 1156 1157 BitExtMI = 1158 BuildMI(*LoadI->getParent(), LoadI, LoadI->getDebugLoc(), 1159 TII->get(IsStoreXReg ? AArch64::ANDXri : AArch64::ANDWri), 1160 DestReg) 1161 .add(StMO) 1162 .addImm(AndMaskEncoded) 1163 .setMIFlags(LoadI->getFlags()); 1164 } else { 1165 BitExtMI = 1166 BuildMI(*LoadI->getParent(), LoadI, LoadI->getDebugLoc(), 1167 TII->get(IsStoreXReg ? AArch64::UBFMXri : AArch64::UBFMWri), 1168 DestReg) 1169 .add(StMO) 1170 .addImm(Immr) 1171 .addImm(Imms) 1172 .setMIFlags(LoadI->getFlags()); 1173 } 1174 } 1175 1176 // Clear kill flags between store and load. 1177 for (MachineInstr &MI : make_range(StoreI->getIterator(), 1178 BitExtMI->getIterator())) 1179 if (MI.killsRegister(StRt, TRI)) { 1180 MI.clearRegisterKills(StRt, TRI); 1181 break; 1182 } 1183 1184 LLVM_DEBUG(dbgs() << "Promoting load by replacing :\n "); 1185 LLVM_DEBUG(StoreI->print(dbgs())); 1186 LLVM_DEBUG(dbgs() << " "); 1187 LLVM_DEBUG(LoadI->print(dbgs())); 1188 LLVM_DEBUG(dbgs() << " with instructions:\n "); 1189 LLVM_DEBUG(StoreI->print(dbgs())); 1190 LLVM_DEBUG(dbgs() << " "); 1191 LLVM_DEBUG((BitExtMI)->print(dbgs())); 1192 LLVM_DEBUG(dbgs() << "\n"); 1193 1194 // Erase the old instructions. 1195 LoadI->eraseFromParent(); 1196 return NextI; 1197 } 1198 1199 static bool inBoundsForPair(bool IsUnscaled, int Offset, int OffsetStride) { 1200 // Convert the byte-offset used by unscaled into an "element" offset used 1201 // by the scaled pair load/store instructions. 1202 if (IsUnscaled) { 1203 // If the byte-offset isn't a multiple of the stride, there's no point 1204 // trying to match it. 1205 if (Offset % OffsetStride) 1206 return false; 1207 Offset /= OffsetStride; 1208 } 1209 return Offset <= 63 && Offset >= -64; 1210 } 1211 1212 // Do alignment, specialized to power of 2 and for signed ints, 1213 // avoiding having to do a C-style cast from uint_64t to int when 1214 // using alignTo from include/llvm/Support/MathExtras.h. 1215 // FIXME: Move this function to include/MathExtras.h? 1216 static int alignTo(int Num, int PowOf2) { 1217 return (Num + PowOf2 - 1) & ~(PowOf2 - 1); 1218 } 1219 1220 static bool mayAlias(MachineInstr &MIa, 1221 SmallVectorImpl<MachineInstr *> &MemInsns, 1222 AliasAnalysis *AA) { 1223 for (MachineInstr *MIb : MemInsns) 1224 if (MIa.mayAlias(AA, *MIb, /*UseTBAA*/ false)) 1225 return true; 1226 1227 return false; 1228 } 1229 1230 bool AArch64LoadStoreOpt::findMatchingStore( 1231 MachineBasicBlock::iterator I, unsigned Limit, 1232 MachineBasicBlock::iterator &StoreI) { 1233 MachineBasicBlock::iterator B = I->getParent()->begin(); 1234 MachineBasicBlock::iterator MBBI = I; 1235 MachineInstr &LoadMI = *I; 1236 Register BaseReg = AArch64InstrInfo::getLdStBaseOp(LoadMI).getReg(); 1237 1238 // If the load is the first instruction in the block, there's obviously 1239 // not any matching store. 1240 if (MBBI == B) 1241 return false; 1242 1243 // Track which register units have been modified and used between the first 1244 // insn and the second insn. 1245 ModifiedRegUnits.clear(); 1246 UsedRegUnits.clear(); 1247 1248 unsigned Count = 0; 1249 do { 1250 MBBI = prev_nodbg(MBBI, B); 1251 MachineInstr &MI = *MBBI; 1252 1253 // Don't count transient instructions towards the search limit since there 1254 // may be different numbers of them if e.g. debug information is present. 1255 if (!MI.isTransient()) 1256 ++Count; 1257 1258 // If the load instruction reads directly from the address to which the 1259 // store instruction writes and the stored value is not modified, we can 1260 // promote the load. Since we do not handle stores with pre-/post-index, 1261 // it's unnecessary to check if BaseReg is modified by the store itself. 1262 // Also we can't handle stores without an immediate offset operand, 1263 // while the operand might be the address for a global variable. 1264 if (MI.mayStore() && isMatchingStore(LoadMI, MI) && 1265 BaseReg == AArch64InstrInfo::getLdStBaseOp(MI).getReg() && 1266 AArch64InstrInfo::getLdStOffsetOp(MI).isImm() && 1267 isLdOffsetInRangeOfSt(LoadMI, MI, TII) && 1268 ModifiedRegUnits.available(getLdStRegOp(MI).getReg())) { 1269 StoreI = MBBI; 1270 return true; 1271 } 1272 1273 if (MI.isCall()) 1274 return false; 1275 1276 // Update modified / uses register units. 1277 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits, TRI); 1278 1279 // Otherwise, if the base register is modified, we have no match, so 1280 // return early. 1281 if (!ModifiedRegUnits.available(BaseReg)) 1282 return false; 1283 1284 // If we encounter a store aliased with the load, return early. 1285 if (MI.mayStore() && LoadMI.mayAlias(AA, MI, /*UseTBAA*/ false)) 1286 return false; 1287 } while (MBBI != B && Count < Limit); 1288 return false; 1289 } 1290 1291 static bool needsWinCFI(const MachineFunction *MF) { 1292 return MF->getTarget().getMCAsmInfo()->usesWindowsCFI() && 1293 MF->getFunction().needsUnwindTableEntry(); 1294 } 1295 1296 // Returns true if FirstMI and MI are candidates for merging or pairing. 1297 // Otherwise, returns false. 1298 static bool areCandidatesToMergeOrPair(MachineInstr &FirstMI, MachineInstr &MI, 1299 LdStPairFlags &Flags, 1300 const AArch64InstrInfo *TII) { 1301 // If this is volatile or if pairing is suppressed, not a candidate. 1302 if (MI.hasOrderedMemoryRef() || TII->isLdStPairSuppressed(MI)) 1303 return false; 1304 1305 // We should have already checked FirstMI for pair suppression and volatility. 1306 assert(!FirstMI.hasOrderedMemoryRef() && 1307 !TII->isLdStPairSuppressed(FirstMI) && 1308 "FirstMI shouldn't get here if either of these checks are true."); 1309 1310 if (needsWinCFI(MI.getMF()) && (MI.getFlag(MachineInstr::FrameSetup) || 1311 MI.getFlag(MachineInstr::FrameDestroy))) 1312 return false; 1313 1314 unsigned OpcA = FirstMI.getOpcode(); 1315 unsigned OpcB = MI.getOpcode(); 1316 1317 // Opcodes match: If the opcodes are pre ld/st there is nothing more to check. 1318 if (OpcA == OpcB) 1319 return !AArch64InstrInfo::isPreLdSt(FirstMI); 1320 1321 // Try to match a sign-extended load/store with a zero-extended load/store. 1322 bool IsValidLdStrOpc, PairIsValidLdStrOpc; 1323 unsigned NonSExtOpc = getMatchingNonSExtOpcode(OpcA, &IsValidLdStrOpc); 1324 assert(IsValidLdStrOpc && 1325 "Given Opc should be a Load or Store with an immediate"); 1326 // OpcA will be the first instruction in the pair. 1327 if (NonSExtOpc == getMatchingNonSExtOpcode(OpcB, &PairIsValidLdStrOpc)) { 1328 Flags.setSExtIdx(NonSExtOpc == (unsigned)OpcA ? 1 : 0); 1329 return true; 1330 } 1331 1332 // If the second instruction isn't even a mergable/pairable load/store, bail 1333 // out. 1334 if (!PairIsValidLdStrOpc) 1335 return false; 1336 1337 // FIXME: We don't support merging narrow stores with mixed scaled/unscaled 1338 // offsets. 1339 if (isNarrowStore(OpcA) || isNarrowStore(OpcB)) 1340 return false; 1341 1342 // The STR<S,D,Q,W,X>pre - STR<S,D,Q,W,X>ui and 1343 // LDR<S,D,Q,W,X>pre-LDR<S,D,Q,W,X>ui 1344 // are candidate pairs that can be merged. 1345 if (isPreLdStPairCandidate(FirstMI, MI)) 1346 return true; 1347 1348 // Try to match an unscaled load/store with a scaled load/store. 1349 return TII->hasUnscaledLdStOffset(OpcA) != TII->hasUnscaledLdStOffset(OpcB) && 1350 getMatchingPairOpcode(OpcA) == getMatchingPairOpcode(OpcB); 1351 1352 // FIXME: Can we also match a mixed sext/zext unscaled/scaled pair? 1353 } 1354 1355 static bool 1356 canRenameUpToDef(MachineInstr &FirstMI, LiveRegUnits &UsedInBetween, 1357 SmallPtrSetImpl<const TargetRegisterClass *> &RequiredClasses, 1358 const TargetRegisterInfo *TRI) { 1359 if (!FirstMI.mayStore()) 1360 return false; 1361 1362 // Check if we can find an unused register which we can use to rename 1363 // the register used by the first load/store. 1364 auto *RegClass = TRI->getMinimalPhysRegClass(getLdStRegOp(FirstMI).getReg()); 1365 MachineFunction &MF = *FirstMI.getParent()->getParent(); 1366 if (!RegClass || !MF.getRegInfo().tracksLiveness()) 1367 return false; 1368 1369 auto RegToRename = getLdStRegOp(FirstMI).getReg(); 1370 // For now, we only rename if the store operand gets killed at the store. 1371 if (!getLdStRegOp(FirstMI).isKill() && 1372 !any_of(FirstMI.operands(), 1373 [TRI, RegToRename](const MachineOperand &MOP) { 1374 return MOP.isReg() && !MOP.isDebug() && MOP.getReg() && 1375 MOP.isImplicit() && MOP.isKill() && 1376 TRI->regsOverlap(RegToRename, MOP.getReg()); 1377 })) { 1378 LLVM_DEBUG(dbgs() << " Operand not killed at " << FirstMI << "\n"); 1379 return false; 1380 } 1381 auto canRenameMOP = [TRI](const MachineOperand &MOP) { 1382 if (MOP.isReg()) { 1383 auto *RegClass = TRI->getMinimalPhysRegClass(MOP.getReg()); 1384 // Renaming registers with multiple disjunct sub-registers (e.g. the 1385 // result of a LD3) means that all sub-registers are renamed, potentially 1386 // impacting other instructions we did not check. Bail out. 1387 // Note that this relies on the structure of the AArch64 register file. In 1388 // particular, a subregister cannot be written without overwriting the 1389 // whole register. 1390 if (RegClass->HasDisjunctSubRegs) { 1391 LLVM_DEBUG( 1392 dbgs() 1393 << " Cannot rename operands with multiple disjunct subregisters (" 1394 << MOP << ")\n"); 1395 return false; 1396 } 1397 } 1398 return MOP.isImplicit() || 1399 (MOP.isRenamable() && !MOP.isEarlyClobber() && !MOP.isTied()); 1400 }; 1401 1402 bool FoundDef = false; 1403 1404 // For each instruction between FirstMI and the previous def for RegToRename, 1405 // we 1406 // * check if we can rename RegToRename in this instruction 1407 // * collect the registers used and required register classes for RegToRename. 1408 std::function<bool(MachineInstr &, bool)> CheckMIs = [&](MachineInstr &MI, 1409 bool IsDef) { 1410 LLVM_DEBUG(dbgs() << "Checking " << MI << "\n"); 1411 // Currently we do not try to rename across frame-setup instructions. 1412 if (MI.getFlag(MachineInstr::FrameSetup)) { 1413 LLVM_DEBUG(dbgs() << " Cannot rename framesetup instructions currently (" 1414 << MI << ")\n"); 1415 return false; 1416 } 1417 1418 UsedInBetween.accumulate(MI); 1419 1420 // For a definition, check that we can rename the definition and exit the 1421 // loop. 1422 FoundDef = IsDef; 1423 1424 // For defs, check if we can rename the first def of RegToRename. 1425 if (FoundDef) { 1426 // For some pseudo instructions, we might not generate code in the end 1427 // (e.g. KILL) and we would end up without a correct def for the rename 1428 // register. 1429 // TODO: This might be overly conservative and we could handle those cases 1430 // in multiple ways: 1431 // 1. Insert an extra copy, to materialize the def. 1432 // 2. Skip pseudo-defs until we find an non-pseudo def. 1433 if (MI.isPseudo()) { 1434 LLVM_DEBUG(dbgs() << " Cannot rename pseudo instruction " << MI 1435 << "\n"); 1436 return false; 1437 } 1438 1439 for (auto &MOP : MI.operands()) { 1440 if (!MOP.isReg() || !MOP.isDef() || MOP.isDebug() || !MOP.getReg() || 1441 !TRI->regsOverlap(MOP.getReg(), RegToRename)) 1442 continue; 1443 if (!canRenameMOP(MOP)) { 1444 LLVM_DEBUG(dbgs() 1445 << " Cannot rename " << MOP << " in " << MI << "\n"); 1446 return false; 1447 } 1448 RequiredClasses.insert(TRI->getMinimalPhysRegClass(MOP.getReg())); 1449 } 1450 return true; 1451 } else { 1452 for (auto &MOP : MI.operands()) { 1453 if (!MOP.isReg() || MOP.isDebug() || !MOP.getReg() || 1454 !TRI->regsOverlap(MOP.getReg(), RegToRename)) 1455 continue; 1456 1457 if (!canRenameMOP(MOP)) { 1458 LLVM_DEBUG(dbgs() 1459 << " Cannot rename " << MOP << " in " << MI << "\n"); 1460 return false; 1461 } 1462 RequiredClasses.insert(TRI->getMinimalPhysRegClass(MOP.getReg())); 1463 } 1464 } 1465 return true; 1466 }; 1467 1468 if (!forAllMIsUntilDef(FirstMI, RegToRename, TRI, LdStLimit, CheckMIs)) 1469 return false; 1470 1471 if (!FoundDef) { 1472 LLVM_DEBUG(dbgs() << " Did not find definition for register in BB\n"); 1473 return false; 1474 } 1475 return true; 1476 } 1477 1478 // Check if we can find a physical register for renaming \p Reg. This register 1479 // must: 1480 // * not be defined already in \p DefinedInBB; DefinedInBB must contain all 1481 // defined registers up to the point where the renamed register will be used, 1482 // * not used in \p UsedInBetween; UsedInBetween must contain all accessed 1483 // registers in the range the rename register will be used, 1484 // * is available in all used register classes (checked using RequiredClasses). 1485 static std::optional<MCPhysReg> tryToFindRegisterToRename( 1486 const MachineFunction &MF, Register Reg, LiveRegUnits &DefinedInBB, 1487 LiveRegUnits &UsedInBetween, 1488 SmallPtrSetImpl<const TargetRegisterClass *> &RequiredClasses, 1489 const TargetRegisterInfo *TRI) { 1490 const MachineRegisterInfo &RegInfo = MF.getRegInfo(); 1491 1492 // Checks if any sub- or super-register of PR is callee saved. 1493 auto AnySubOrSuperRegCalleePreserved = [&MF, TRI](MCPhysReg PR) { 1494 return any_of(TRI->sub_and_superregs_inclusive(PR), 1495 [&MF, TRI](MCPhysReg SubOrSuper) { 1496 return TRI->isCalleeSavedPhysReg(SubOrSuper, MF); 1497 }); 1498 }; 1499 1500 // Check if PR or one of its sub- or super-registers can be used for all 1501 // required register classes. 1502 auto CanBeUsedForAllClasses = [&RequiredClasses, TRI](MCPhysReg PR) { 1503 return all_of(RequiredClasses, [PR, TRI](const TargetRegisterClass *C) { 1504 return any_of(TRI->sub_and_superregs_inclusive(PR), 1505 [C, TRI](MCPhysReg SubOrSuper) { 1506 return C == TRI->getMinimalPhysRegClass(SubOrSuper); 1507 }); 1508 }); 1509 }; 1510 1511 auto *RegClass = TRI->getMinimalPhysRegClass(Reg); 1512 for (const MCPhysReg &PR : *RegClass) { 1513 if (DefinedInBB.available(PR) && UsedInBetween.available(PR) && 1514 !RegInfo.isReserved(PR) && !AnySubOrSuperRegCalleePreserved(PR) && 1515 CanBeUsedForAllClasses(PR)) { 1516 DefinedInBB.addReg(PR); 1517 LLVM_DEBUG(dbgs() << "Found rename register " << printReg(PR, TRI) 1518 << "\n"); 1519 return {PR}; 1520 } 1521 } 1522 LLVM_DEBUG(dbgs() << "No rename register found from " 1523 << TRI->getRegClassName(RegClass) << "\n"); 1524 return std::nullopt; 1525 } 1526 1527 /// Scan the instructions looking for a load/store that can be combined with the 1528 /// current instruction into a wider equivalent or a load/store pair. 1529 MachineBasicBlock::iterator 1530 AArch64LoadStoreOpt::findMatchingInsn(MachineBasicBlock::iterator I, 1531 LdStPairFlags &Flags, unsigned Limit, 1532 bool FindNarrowMerge) { 1533 MachineBasicBlock::iterator E = I->getParent()->end(); 1534 MachineBasicBlock::iterator MBBI = I; 1535 MachineBasicBlock::iterator MBBIWithRenameReg; 1536 MachineInstr &FirstMI = *I; 1537 MBBI = next_nodbg(MBBI, E); 1538 1539 bool MayLoad = FirstMI.mayLoad(); 1540 bool IsUnscaled = TII->hasUnscaledLdStOffset(FirstMI); 1541 Register Reg = getLdStRegOp(FirstMI).getReg(); 1542 Register BaseReg = AArch64InstrInfo::getLdStBaseOp(FirstMI).getReg(); 1543 int Offset = AArch64InstrInfo::getLdStOffsetOp(FirstMI).getImm(); 1544 int OffsetStride = IsUnscaled ? TII->getMemScale(FirstMI) : 1; 1545 bool IsPromotableZeroStore = isPromotableZeroStoreInst(FirstMI); 1546 1547 std::optional<bool> MaybeCanRename; 1548 if (!EnableRenaming) 1549 MaybeCanRename = {false}; 1550 1551 SmallPtrSet<const TargetRegisterClass *, 5> RequiredClasses; 1552 LiveRegUnits UsedInBetween; 1553 UsedInBetween.init(*TRI); 1554 1555 Flags.clearRenameReg(); 1556 1557 // Track which register units have been modified and used between the first 1558 // insn (inclusive) and the second insn. 1559 ModifiedRegUnits.clear(); 1560 UsedRegUnits.clear(); 1561 1562 // Remember any instructions that read/write memory between FirstMI and MI. 1563 SmallVector<MachineInstr *, 4> MemInsns; 1564 1565 for (unsigned Count = 0; MBBI != E && Count < Limit; 1566 MBBI = next_nodbg(MBBI, E)) { 1567 MachineInstr &MI = *MBBI; 1568 1569 UsedInBetween.accumulate(MI); 1570 1571 // Don't count transient instructions towards the search limit since there 1572 // may be different numbers of them if e.g. debug information is present. 1573 if (!MI.isTransient()) 1574 ++Count; 1575 1576 Flags.setSExtIdx(-1); 1577 if (areCandidatesToMergeOrPair(FirstMI, MI, Flags, TII) && 1578 AArch64InstrInfo::getLdStOffsetOp(MI).isImm()) { 1579 assert(MI.mayLoadOrStore() && "Expected memory operation."); 1580 // If we've found another instruction with the same opcode, check to see 1581 // if the base and offset are compatible with our starting instruction. 1582 // These instructions all have scaled immediate operands, so we just 1583 // check for +1/-1. Make sure to check the new instruction offset is 1584 // actually an immediate and not a symbolic reference destined for 1585 // a relocation. 1586 Register MIBaseReg = AArch64InstrInfo::getLdStBaseOp(MI).getReg(); 1587 int MIOffset = AArch64InstrInfo::getLdStOffsetOp(MI).getImm(); 1588 bool MIIsUnscaled = TII->hasUnscaledLdStOffset(MI); 1589 if (IsUnscaled != MIIsUnscaled) { 1590 // We're trying to pair instructions that differ in how they are scaled. 1591 // If FirstMI is scaled then scale the offset of MI accordingly. 1592 // Otherwise, do the opposite (i.e., make MI's offset unscaled). 1593 int MemSize = TII->getMemScale(MI); 1594 if (MIIsUnscaled) { 1595 // If the unscaled offset isn't a multiple of the MemSize, we can't 1596 // pair the operations together: bail and keep looking. 1597 if (MIOffset % MemSize) { 1598 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, 1599 UsedRegUnits, TRI); 1600 MemInsns.push_back(&MI); 1601 continue; 1602 } 1603 MIOffset /= MemSize; 1604 } else { 1605 MIOffset *= MemSize; 1606 } 1607 } 1608 1609 bool IsPreLdSt = isPreLdStPairCandidate(FirstMI, MI); 1610 1611 if (BaseReg == MIBaseReg) { 1612 // If the offset of the second ld/st is not equal to the size of the 1613 // destination register it can’t be paired with a pre-index ld/st 1614 // pair. Additionally if the base reg is used or modified the operations 1615 // can't be paired: bail and keep looking. 1616 if (IsPreLdSt) { 1617 bool IsOutOfBounds = MIOffset != TII->getMemScale(MI); 1618 bool IsBaseRegUsed = !UsedRegUnits.available( 1619 AArch64InstrInfo::getLdStBaseOp(MI).getReg()); 1620 bool IsBaseRegModified = !ModifiedRegUnits.available( 1621 AArch64InstrInfo::getLdStBaseOp(MI).getReg()); 1622 // If the stored value and the address of the second instruction is 1623 // the same, it needs to be using the updated register and therefore 1624 // it must not be folded. 1625 bool IsMIRegTheSame = 1626 TRI->regsOverlap(getLdStRegOp(MI).getReg(), 1627 AArch64InstrInfo::getLdStBaseOp(MI).getReg()); 1628 if (IsOutOfBounds || IsBaseRegUsed || IsBaseRegModified || 1629 IsMIRegTheSame) { 1630 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, 1631 UsedRegUnits, TRI); 1632 MemInsns.push_back(&MI); 1633 continue; 1634 } 1635 } else { 1636 if ((Offset != MIOffset + OffsetStride) && 1637 (Offset + OffsetStride != MIOffset)) { 1638 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, 1639 UsedRegUnits, TRI); 1640 MemInsns.push_back(&MI); 1641 continue; 1642 } 1643 } 1644 1645 int MinOffset = Offset < MIOffset ? Offset : MIOffset; 1646 if (FindNarrowMerge) { 1647 // If the alignment requirements of the scaled wide load/store 1648 // instruction can't express the offset of the scaled narrow input, 1649 // bail and keep looking. For promotable zero stores, allow only when 1650 // the stored value is the same (i.e., WZR). 1651 if ((!IsUnscaled && alignTo(MinOffset, 2) != MinOffset) || 1652 (IsPromotableZeroStore && Reg != getLdStRegOp(MI).getReg())) { 1653 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, 1654 UsedRegUnits, TRI); 1655 MemInsns.push_back(&MI); 1656 continue; 1657 } 1658 } else { 1659 // Pairwise instructions have a 7-bit signed offset field. Single 1660 // insns have a 12-bit unsigned offset field. If the resultant 1661 // immediate offset of merging these instructions is out of range for 1662 // a pairwise instruction, bail and keep looking. 1663 if (!inBoundsForPair(IsUnscaled, MinOffset, OffsetStride)) { 1664 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, 1665 UsedRegUnits, TRI); 1666 MemInsns.push_back(&MI); 1667 continue; 1668 } 1669 // If the alignment requirements of the paired (scaled) instruction 1670 // can't express the offset of the unscaled input, bail and keep 1671 // looking. 1672 if (IsUnscaled && (alignTo(MinOffset, OffsetStride) != MinOffset)) { 1673 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, 1674 UsedRegUnits, TRI); 1675 MemInsns.push_back(&MI); 1676 continue; 1677 } 1678 } 1679 // If the destination register of one load is the same register or a 1680 // sub/super register of the other load, bail and keep looking. A 1681 // load-pair instruction with both destination registers the same is 1682 // UNPREDICTABLE and will result in an exception. 1683 if (MayLoad && 1684 TRI->isSuperOrSubRegisterEq(Reg, getLdStRegOp(MI).getReg())) { 1685 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits, 1686 TRI); 1687 MemInsns.push_back(&MI); 1688 continue; 1689 } 1690 1691 // If the BaseReg has been modified, then we cannot do the optimization. 1692 // For example, in the following pattern 1693 // ldr x1 [x2] 1694 // ldr x2 [x3] 1695 // ldr x4 [x2, #8], 1696 // the first and third ldr cannot be converted to ldp x1, x4, [x2] 1697 if (!ModifiedRegUnits.available(BaseReg)) 1698 return E; 1699 1700 // If the Rt of the second instruction was not modified or used between 1701 // the two instructions and none of the instructions between the second 1702 // and first alias with the second, we can combine the second into the 1703 // first. 1704 if (ModifiedRegUnits.available(getLdStRegOp(MI).getReg()) && 1705 !(MI.mayLoad() && 1706 !UsedRegUnits.available(getLdStRegOp(MI).getReg())) && 1707 !mayAlias(MI, MemInsns, AA)) { 1708 1709 Flags.setMergeForward(false); 1710 Flags.clearRenameReg(); 1711 return MBBI; 1712 } 1713 1714 // Likewise, if the Rt of the first instruction is not modified or used 1715 // between the two instructions and none of the instructions between the 1716 // first and the second alias with the first, we can combine the first 1717 // into the second. 1718 if (!(MayLoad && 1719 !UsedRegUnits.available(getLdStRegOp(FirstMI).getReg())) && 1720 !mayAlias(FirstMI, MemInsns, AA)) { 1721 1722 if (ModifiedRegUnits.available(getLdStRegOp(FirstMI).getReg())) { 1723 Flags.setMergeForward(true); 1724 Flags.clearRenameReg(); 1725 return MBBI; 1726 } 1727 1728 if (DebugCounter::shouldExecute(RegRenamingCounter)) { 1729 if (!MaybeCanRename) 1730 MaybeCanRename = {canRenameUpToDef(FirstMI, UsedInBetween, 1731 RequiredClasses, TRI)}; 1732 1733 if (*MaybeCanRename) { 1734 std::optional<MCPhysReg> MaybeRenameReg = 1735 tryToFindRegisterToRename(*FirstMI.getParent()->getParent(), 1736 Reg, DefinedInBB, UsedInBetween, 1737 RequiredClasses, TRI); 1738 if (MaybeRenameReg) { 1739 Flags.setRenameReg(*MaybeRenameReg); 1740 Flags.setMergeForward(true); 1741 MBBIWithRenameReg = MBBI; 1742 } 1743 } 1744 } 1745 } 1746 // Unable to combine these instructions due to interference in between. 1747 // Keep looking. 1748 } 1749 } 1750 1751 if (Flags.getRenameReg()) 1752 return MBBIWithRenameReg; 1753 1754 // If the instruction wasn't a matching load or store. Stop searching if we 1755 // encounter a call instruction that might modify memory. 1756 if (MI.isCall()) 1757 return E; 1758 1759 // Update modified / uses register units. 1760 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits, TRI); 1761 1762 // Otherwise, if the base register is modified, we have no match, so 1763 // return early. 1764 if (!ModifiedRegUnits.available(BaseReg)) 1765 return E; 1766 1767 // Update list of instructions that read/write memory. 1768 if (MI.mayLoadOrStore()) 1769 MemInsns.push_back(&MI); 1770 } 1771 return E; 1772 } 1773 1774 static MachineBasicBlock::iterator 1775 maybeMoveCFI(MachineInstr &MI, MachineBasicBlock::iterator MaybeCFI) { 1776 auto End = MI.getParent()->end(); 1777 if (MaybeCFI == End || 1778 MaybeCFI->getOpcode() != TargetOpcode::CFI_INSTRUCTION || 1779 !(MI.getFlag(MachineInstr::FrameSetup) || 1780 MI.getFlag(MachineInstr::FrameDestroy)) || 1781 AArch64InstrInfo::getLdStBaseOp(MI).getReg() != AArch64::SP) 1782 return End; 1783 1784 const MachineFunction &MF = *MI.getParent()->getParent(); 1785 unsigned CFIIndex = MaybeCFI->getOperand(0).getCFIIndex(); 1786 const MCCFIInstruction &CFI = MF.getFrameInstructions()[CFIIndex]; 1787 switch (CFI.getOperation()) { 1788 case MCCFIInstruction::OpDefCfa: 1789 case MCCFIInstruction::OpDefCfaOffset: 1790 return MaybeCFI; 1791 default: 1792 return End; 1793 } 1794 } 1795 1796 MachineBasicBlock::iterator 1797 AArch64LoadStoreOpt::mergeUpdateInsn(MachineBasicBlock::iterator I, 1798 MachineBasicBlock::iterator Update, 1799 bool IsPreIdx) { 1800 assert((Update->getOpcode() == AArch64::ADDXri || 1801 Update->getOpcode() == AArch64::SUBXri) && 1802 "Unexpected base register update instruction to merge!"); 1803 MachineBasicBlock::iterator E = I->getParent()->end(); 1804 MachineBasicBlock::iterator NextI = next_nodbg(I, E); 1805 1806 // If updating the SP and the following instruction is CFA offset related CFI 1807 // instruction move it after the merged instruction. 1808 MachineBasicBlock::iterator CFI = 1809 IsPreIdx ? maybeMoveCFI(*Update, next_nodbg(Update, E)) : E; 1810 1811 // Return the instruction following the merged instruction, which is 1812 // the instruction following our unmerged load. Unless that's the add/sub 1813 // instruction we're merging, in which case it's the one after that. 1814 if (NextI == Update) 1815 NextI = next_nodbg(NextI, E); 1816 1817 int Value = Update->getOperand(2).getImm(); 1818 assert(AArch64_AM::getShiftValue(Update->getOperand(3).getImm()) == 0 && 1819 "Can't merge 1 << 12 offset into pre-/post-indexed load / store"); 1820 if (Update->getOpcode() == AArch64::SUBXri) 1821 Value = -Value; 1822 1823 unsigned NewOpc = IsPreIdx ? getPreIndexedOpcode(I->getOpcode()) 1824 : getPostIndexedOpcode(I->getOpcode()); 1825 MachineInstrBuilder MIB; 1826 int Scale, MinOffset, MaxOffset; 1827 getPrePostIndexedMemOpInfo(*I, Scale, MinOffset, MaxOffset); 1828 if (!AArch64InstrInfo::isPairedLdSt(*I)) { 1829 // Non-paired instruction. 1830 MIB = BuildMI(*I->getParent(), I, I->getDebugLoc(), TII->get(NewOpc)) 1831 .add(getLdStRegOp(*Update)) 1832 .add(getLdStRegOp(*I)) 1833 .add(AArch64InstrInfo::getLdStBaseOp(*I)) 1834 .addImm(Value / Scale) 1835 .setMemRefs(I->memoperands()) 1836 .setMIFlags(I->mergeFlagsWith(*Update)); 1837 } else { 1838 // Paired instruction. 1839 MIB = BuildMI(*I->getParent(), I, I->getDebugLoc(), TII->get(NewOpc)) 1840 .add(getLdStRegOp(*Update)) 1841 .add(getLdStRegOp(*I, 0)) 1842 .add(getLdStRegOp(*I, 1)) 1843 .add(AArch64InstrInfo::getLdStBaseOp(*I)) 1844 .addImm(Value / Scale) 1845 .setMemRefs(I->memoperands()) 1846 .setMIFlags(I->mergeFlagsWith(*Update)); 1847 } 1848 if (CFI != E) { 1849 MachineBasicBlock *MBB = I->getParent(); 1850 MBB->splice(std::next(MIB.getInstr()->getIterator()), MBB, CFI); 1851 } 1852 1853 if (IsPreIdx) { 1854 ++NumPreFolded; 1855 LLVM_DEBUG(dbgs() << "Creating pre-indexed load/store."); 1856 } else { 1857 ++NumPostFolded; 1858 LLVM_DEBUG(dbgs() << "Creating post-indexed load/store."); 1859 } 1860 LLVM_DEBUG(dbgs() << " Replacing instructions:\n "); 1861 LLVM_DEBUG(I->print(dbgs())); 1862 LLVM_DEBUG(dbgs() << " "); 1863 LLVM_DEBUG(Update->print(dbgs())); 1864 LLVM_DEBUG(dbgs() << " with instruction:\n "); 1865 LLVM_DEBUG(((MachineInstr *)MIB)->print(dbgs())); 1866 LLVM_DEBUG(dbgs() << "\n"); 1867 1868 // Erase the old instructions for the block. 1869 I->eraseFromParent(); 1870 Update->eraseFromParent(); 1871 1872 return NextI; 1873 } 1874 1875 bool AArch64LoadStoreOpt::isMatchingUpdateInsn(MachineInstr &MemMI, 1876 MachineInstr &MI, 1877 unsigned BaseReg, int Offset) { 1878 switch (MI.getOpcode()) { 1879 default: 1880 break; 1881 case AArch64::SUBXri: 1882 case AArch64::ADDXri: 1883 // Make sure it's a vanilla immediate operand, not a relocation or 1884 // anything else we can't handle. 1885 if (!MI.getOperand(2).isImm()) 1886 break; 1887 // Watch out for 1 << 12 shifted value. 1888 if (AArch64_AM::getShiftValue(MI.getOperand(3).getImm())) 1889 break; 1890 1891 // The update instruction source and destination register must be the 1892 // same as the load/store base register. 1893 if (MI.getOperand(0).getReg() != BaseReg || 1894 MI.getOperand(1).getReg() != BaseReg) 1895 break; 1896 1897 int UpdateOffset = MI.getOperand(2).getImm(); 1898 if (MI.getOpcode() == AArch64::SUBXri) 1899 UpdateOffset = -UpdateOffset; 1900 1901 // The immediate must be a multiple of the scaling factor of the pre/post 1902 // indexed instruction. 1903 int Scale, MinOffset, MaxOffset; 1904 getPrePostIndexedMemOpInfo(MemMI, Scale, MinOffset, MaxOffset); 1905 if (UpdateOffset % Scale != 0) 1906 break; 1907 1908 // Scaled offset must fit in the instruction immediate. 1909 int ScaledOffset = UpdateOffset / Scale; 1910 if (ScaledOffset > MaxOffset || ScaledOffset < MinOffset) 1911 break; 1912 1913 // If we have a non-zero Offset, we check that it matches the amount 1914 // we're adding to the register. 1915 if (!Offset || Offset == UpdateOffset) 1916 return true; 1917 break; 1918 } 1919 return false; 1920 } 1921 1922 MachineBasicBlock::iterator AArch64LoadStoreOpt::findMatchingUpdateInsnForward( 1923 MachineBasicBlock::iterator I, int UnscaledOffset, unsigned Limit) { 1924 MachineBasicBlock::iterator E = I->getParent()->end(); 1925 MachineInstr &MemMI = *I; 1926 MachineBasicBlock::iterator MBBI = I; 1927 1928 Register BaseReg = AArch64InstrInfo::getLdStBaseOp(MemMI).getReg(); 1929 int MIUnscaledOffset = AArch64InstrInfo::getLdStOffsetOp(MemMI).getImm() * 1930 TII->getMemScale(MemMI); 1931 1932 // Scan forward looking for post-index opportunities. Updating instructions 1933 // can't be formed if the memory instruction doesn't have the offset we're 1934 // looking for. 1935 if (MIUnscaledOffset != UnscaledOffset) 1936 return E; 1937 1938 // If the base register overlaps a source/destination register, we can't 1939 // merge the update. This does not apply to tag store instructions which 1940 // ignore the address part of the source register. 1941 // This does not apply to STGPi as well, which does not have unpredictable 1942 // behavior in this case unlike normal stores, and always performs writeback 1943 // after reading the source register value. 1944 if (!isTagStore(MemMI) && MemMI.getOpcode() != AArch64::STGPi) { 1945 bool IsPairedInsn = AArch64InstrInfo::isPairedLdSt(MemMI); 1946 for (unsigned i = 0, e = IsPairedInsn ? 2 : 1; i != e; ++i) { 1947 Register DestReg = getLdStRegOp(MemMI, i).getReg(); 1948 if (DestReg == BaseReg || TRI->isSubRegister(BaseReg, DestReg)) 1949 return E; 1950 } 1951 } 1952 1953 // Track which register units have been modified and used between the first 1954 // insn (inclusive) and the second insn. 1955 ModifiedRegUnits.clear(); 1956 UsedRegUnits.clear(); 1957 MBBI = next_nodbg(MBBI, E); 1958 1959 // We can't post-increment the stack pointer if any instruction between 1960 // the memory access (I) and the increment (MBBI) can access the memory 1961 // region defined by [SP, MBBI]. 1962 const bool BaseRegSP = BaseReg == AArch64::SP; 1963 if (BaseRegSP && needsWinCFI(I->getMF())) { 1964 // FIXME: For now, we always block the optimization over SP in windows 1965 // targets as it requires to adjust the unwind/debug info, messing up 1966 // the unwind info can actually cause a miscompile. 1967 return E; 1968 } 1969 1970 for (unsigned Count = 0; MBBI != E && Count < Limit; 1971 MBBI = next_nodbg(MBBI, E)) { 1972 MachineInstr &MI = *MBBI; 1973 1974 // Don't count transient instructions towards the search limit since there 1975 // may be different numbers of them if e.g. debug information is present. 1976 if (!MI.isTransient()) 1977 ++Count; 1978 1979 // If we found a match, return it. 1980 if (isMatchingUpdateInsn(*I, MI, BaseReg, UnscaledOffset)) 1981 return MBBI; 1982 1983 // Update the status of what the instruction clobbered and used. 1984 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits, TRI); 1985 1986 // Otherwise, if the base register is used or modified, we have no match, so 1987 // return early. 1988 // If we are optimizing SP, do not allow instructions that may load or store 1989 // in between the load and the optimized value update. 1990 if (!ModifiedRegUnits.available(BaseReg) || 1991 !UsedRegUnits.available(BaseReg) || 1992 (BaseRegSP && MBBI->mayLoadOrStore())) 1993 return E; 1994 } 1995 return E; 1996 } 1997 1998 MachineBasicBlock::iterator AArch64LoadStoreOpt::findMatchingUpdateInsnBackward( 1999 MachineBasicBlock::iterator I, unsigned Limit) { 2000 MachineBasicBlock::iterator B = I->getParent()->begin(); 2001 MachineBasicBlock::iterator E = I->getParent()->end(); 2002 MachineInstr &MemMI = *I; 2003 MachineBasicBlock::iterator MBBI = I; 2004 MachineFunction &MF = *MemMI.getMF(); 2005 2006 Register BaseReg = AArch64InstrInfo::getLdStBaseOp(MemMI).getReg(); 2007 int Offset = AArch64InstrInfo::getLdStOffsetOp(MemMI).getImm(); 2008 2009 // If the load/store is the first instruction in the block, there's obviously 2010 // not any matching update. Ditto if the memory offset isn't zero. 2011 if (MBBI == B || Offset != 0) 2012 return E; 2013 // If the base register overlaps a destination register, we can't 2014 // merge the update. 2015 if (!isTagStore(MemMI)) { 2016 bool IsPairedInsn = AArch64InstrInfo::isPairedLdSt(MemMI); 2017 for (unsigned i = 0, e = IsPairedInsn ? 2 : 1; i != e; ++i) { 2018 Register DestReg = getLdStRegOp(MemMI, i).getReg(); 2019 if (DestReg == BaseReg || TRI->isSubRegister(BaseReg, DestReg)) 2020 return E; 2021 } 2022 } 2023 2024 const bool BaseRegSP = BaseReg == AArch64::SP; 2025 if (BaseRegSP && needsWinCFI(I->getMF())) { 2026 // FIXME: For now, we always block the optimization over SP in windows 2027 // targets as it requires to adjust the unwind/debug info, messing up 2028 // the unwind info can actually cause a miscompile. 2029 return E; 2030 } 2031 2032 const AArch64Subtarget &Subtarget = MF.getSubtarget<AArch64Subtarget>(); 2033 unsigned RedZoneSize = 2034 Subtarget.getTargetLowering()->getRedZoneSize(MF.getFunction()); 2035 2036 // Track which register units have been modified and used between the first 2037 // insn (inclusive) and the second insn. 2038 ModifiedRegUnits.clear(); 2039 UsedRegUnits.clear(); 2040 unsigned Count = 0; 2041 bool MemAcessBeforeSPPreInc = false; 2042 do { 2043 MBBI = prev_nodbg(MBBI, B); 2044 MachineInstr &MI = *MBBI; 2045 2046 // Don't count transient instructions towards the search limit since there 2047 // may be different numbers of them if e.g. debug information is present. 2048 if (!MI.isTransient()) 2049 ++Count; 2050 2051 // If we found a match, return it. 2052 if (isMatchingUpdateInsn(*I, MI, BaseReg, Offset)) { 2053 // Check that the update value is within our red zone limit (which may be 2054 // zero). 2055 if (MemAcessBeforeSPPreInc && MBBI->getOperand(2).getImm() > RedZoneSize) 2056 return E; 2057 return MBBI; 2058 } 2059 2060 // Update the status of what the instruction clobbered and used. 2061 LiveRegUnits::accumulateUsedDefed(MI, ModifiedRegUnits, UsedRegUnits, TRI); 2062 2063 // Otherwise, if the base register is used or modified, we have no match, so 2064 // return early. 2065 if (!ModifiedRegUnits.available(BaseReg) || 2066 !UsedRegUnits.available(BaseReg)) 2067 return E; 2068 // Keep track if we have a memory access before an SP pre-increment, in this 2069 // case we need to validate later that the update amount respects the red 2070 // zone. 2071 if (BaseRegSP && MBBI->mayLoadOrStore()) 2072 MemAcessBeforeSPPreInc = true; 2073 } while (MBBI != B && Count < Limit); 2074 return E; 2075 } 2076 2077 bool AArch64LoadStoreOpt::tryToPromoteLoadFromStore( 2078 MachineBasicBlock::iterator &MBBI) { 2079 MachineInstr &MI = *MBBI; 2080 // If this is a volatile load, don't mess with it. 2081 if (MI.hasOrderedMemoryRef()) 2082 return false; 2083 2084 if (needsWinCFI(MI.getMF()) && MI.getFlag(MachineInstr::FrameDestroy)) 2085 return false; 2086 2087 // Make sure this is a reg+imm. 2088 // FIXME: It is possible to extend it to handle reg+reg cases. 2089 if (!AArch64InstrInfo::getLdStOffsetOp(MI).isImm()) 2090 return false; 2091 2092 // Look backward up to LdStLimit instructions. 2093 MachineBasicBlock::iterator StoreI; 2094 if (findMatchingStore(MBBI, LdStLimit, StoreI)) { 2095 ++NumLoadsFromStoresPromoted; 2096 // Promote the load. Keeping the iterator straight is a 2097 // pain, so we let the merge routine tell us what the next instruction 2098 // is after it's done mucking about. 2099 MBBI = promoteLoadFromStore(MBBI, StoreI); 2100 return true; 2101 } 2102 return false; 2103 } 2104 2105 // Merge adjacent zero stores into a wider store. 2106 bool AArch64LoadStoreOpt::tryToMergeZeroStInst( 2107 MachineBasicBlock::iterator &MBBI) { 2108 assert(isPromotableZeroStoreInst(*MBBI) && "Expected narrow store."); 2109 MachineInstr &MI = *MBBI; 2110 MachineBasicBlock::iterator E = MI.getParent()->end(); 2111 2112 if (!TII->isCandidateToMergeOrPair(MI)) 2113 return false; 2114 2115 // Look ahead up to LdStLimit instructions for a mergable instruction. 2116 LdStPairFlags Flags; 2117 MachineBasicBlock::iterator MergeMI = 2118 findMatchingInsn(MBBI, Flags, LdStLimit, /* FindNarrowMerge = */ true); 2119 if (MergeMI != E) { 2120 ++NumZeroStoresPromoted; 2121 2122 // Keeping the iterator straight is a pain, so we let the merge routine tell 2123 // us what the next instruction is after it's done mucking about. 2124 MBBI = mergeNarrowZeroStores(MBBI, MergeMI, Flags); 2125 return true; 2126 } 2127 return false; 2128 } 2129 2130 // Find loads and stores that can be merged into a single load or store pair 2131 // instruction. 2132 bool AArch64LoadStoreOpt::tryToPairLdStInst(MachineBasicBlock::iterator &MBBI) { 2133 MachineInstr &MI = *MBBI; 2134 MachineBasicBlock::iterator E = MI.getParent()->end(); 2135 2136 if (!TII->isCandidateToMergeOrPair(MI)) 2137 return false; 2138 2139 // Early exit if the offset is not possible to match. (6 bits of positive 2140 // range, plus allow an extra one in case we find a later insn that matches 2141 // with Offset-1) 2142 bool IsUnscaled = TII->hasUnscaledLdStOffset(MI); 2143 int Offset = AArch64InstrInfo::getLdStOffsetOp(MI).getImm(); 2144 int OffsetStride = IsUnscaled ? TII->getMemScale(MI) : 1; 2145 // Allow one more for offset. 2146 if (Offset > 0) 2147 Offset -= OffsetStride; 2148 if (!inBoundsForPair(IsUnscaled, Offset, OffsetStride)) 2149 return false; 2150 2151 // Look ahead up to LdStLimit instructions for a pairable instruction. 2152 LdStPairFlags Flags; 2153 MachineBasicBlock::iterator Paired = 2154 findMatchingInsn(MBBI, Flags, LdStLimit, /* FindNarrowMerge = */ false); 2155 if (Paired != E) { 2156 ++NumPairCreated; 2157 if (TII->hasUnscaledLdStOffset(MI)) 2158 ++NumUnscaledPairCreated; 2159 // Keeping the iterator straight is a pain, so we let the merge routine tell 2160 // us what the next instruction is after it's done mucking about. 2161 auto Prev = std::prev(MBBI); 2162 MBBI = mergePairedInsns(MBBI, Paired, Flags); 2163 // Collect liveness info for instructions between Prev and the new position 2164 // MBBI. 2165 for (auto I = std::next(Prev); I != MBBI; I++) 2166 updateDefinedRegisters(*I, DefinedInBB, TRI); 2167 2168 return true; 2169 } 2170 return false; 2171 } 2172 2173 bool AArch64LoadStoreOpt::tryToMergeLdStUpdate 2174 (MachineBasicBlock::iterator &MBBI) { 2175 MachineInstr &MI = *MBBI; 2176 MachineBasicBlock::iterator E = MI.getParent()->end(); 2177 MachineBasicBlock::iterator Update; 2178 2179 // Look forward to try to form a post-index instruction. For example, 2180 // ldr x0, [x20] 2181 // add x20, x20, #32 2182 // merged into: 2183 // ldr x0, [x20], #32 2184 Update = findMatchingUpdateInsnForward(MBBI, 0, UpdateLimit); 2185 if (Update != E) { 2186 // Merge the update into the ld/st. 2187 MBBI = mergeUpdateInsn(MBBI, Update, /*IsPreIdx=*/false); 2188 return true; 2189 } 2190 2191 // Don't know how to handle unscaled pre/post-index versions below, so bail. 2192 if (TII->hasUnscaledLdStOffset(MI.getOpcode())) 2193 return false; 2194 2195 // Look back to try to find a pre-index instruction. For example, 2196 // add x0, x0, #8 2197 // ldr x1, [x0] 2198 // merged into: 2199 // ldr x1, [x0, #8]! 2200 Update = findMatchingUpdateInsnBackward(MBBI, UpdateLimit); 2201 if (Update != E) { 2202 // Merge the update into the ld/st. 2203 MBBI = mergeUpdateInsn(MBBI, Update, /*IsPreIdx=*/true); 2204 return true; 2205 } 2206 2207 // The immediate in the load/store is scaled by the size of the memory 2208 // operation. The immediate in the add we're looking for, 2209 // however, is not, so adjust here. 2210 int UnscaledOffset = 2211 AArch64InstrInfo::getLdStOffsetOp(MI).getImm() * TII->getMemScale(MI); 2212 2213 // Look forward to try to find a pre-index instruction. For example, 2214 // ldr x1, [x0, #64] 2215 // add x0, x0, #64 2216 // merged into: 2217 // ldr x1, [x0, #64]! 2218 Update = findMatchingUpdateInsnForward(MBBI, UnscaledOffset, UpdateLimit); 2219 if (Update != E) { 2220 // Merge the update into the ld/st. 2221 MBBI = mergeUpdateInsn(MBBI, Update, /*IsPreIdx=*/true); 2222 return true; 2223 } 2224 2225 return false; 2226 } 2227 2228 bool AArch64LoadStoreOpt::optimizeBlock(MachineBasicBlock &MBB, 2229 bool EnableNarrowZeroStOpt) { 2230 2231 bool Modified = false; 2232 // Four tranformations to do here: 2233 // 1) Find loads that directly read from stores and promote them by 2234 // replacing with mov instructions. If the store is wider than the load, 2235 // the load will be replaced with a bitfield extract. 2236 // e.g., 2237 // str w1, [x0, #4] 2238 // ldrh w2, [x0, #6] 2239 // ; becomes 2240 // str w1, [x0, #4] 2241 // lsr w2, w1, #16 2242 for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end(); 2243 MBBI != E;) { 2244 if (isPromotableLoadFromStore(*MBBI) && tryToPromoteLoadFromStore(MBBI)) 2245 Modified = true; 2246 else 2247 ++MBBI; 2248 } 2249 // 2) Merge adjacent zero stores into a wider store. 2250 // e.g., 2251 // strh wzr, [x0] 2252 // strh wzr, [x0, #2] 2253 // ; becomes 2254 // str wzr, [x0] 2255 // e.g., 2256 // str wzr, [x0] 2257 // str wzr, [x0, #4] 2258 // ; becomes 2259 // str xzr, [x0] 2260 if (EnableNarrowZeroStOpt) 2261 for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end(); 2262 MBBI != E;) { 2263 if (isPromotableZeroStoreInst(*MBBI) && tryToMergeZeroStInst(MBBI)) 2264 Modified = true; 2265 else 2266 ++MBBI; 2267 } 2268 // 3) Find loads and stores that can be merged into a single load or store 2269 // pair instruction. 2270 // e.g., 2271 // ldr x0, [x2] 2272 // ldr x1, [x2, #8] 2273 // ; becomes 2274 // ldp x0, x1, [x2] 2275 2276 if (MBB.getParent()->getRegInfo().tracksLiveness()) { 2277 DefinedInBB.clear(); 2278 DefinedInBB.addLiveIns(MBB); 2279 } 2280 2281 for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end(); 2282 MBBI != E;) { 2283 // Track currently live registers up to this point, to help with 2284 // searching for a rename register on demand. 2285 updateDefinedRegisters(*MBBI, DefinedInBB, TRI); 2286 if (TII->isPairableLdStInst(*MBBI) && tryToPairLdStInst(MBBI)) 2287 Modified = true; 2288 else 2289 ++MBBI; 2290 } 2291 // 4) Find base register updates that can be merged into the load or store 2292 // as a base-reg writeback. 2293 // e.g., 2294 // ldr x0, [x2] 2295 // add x2, x2, #4 2296 // ; becomes 2297 // ldr x0, [x2], #4 2298 for (MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end(); 2299 MBBI != E;) { 2300 if (isMergeableLdStUpdate(*MBBI) && tryToMergeLdStUpdate(MBBI)) 2301 Modified = true; 2302 else 2303 ++MBBI; 2304 } 2305 2306 return Modified; 2307 } 2308 2309 bool AArch64LoadStoreOpt::runOnMachineFunction(MachineFunction &Fn) { 2310 if (skipFunction(Fn.getFunction())) 2311 return false; 2312 2313 Subtarget = &Fn.getSubtarget<AArch64Subtarget>(); 2314 TII = static_cast<const AArch64InstrInfo *>(Subtarget->getInstrInfo()); 2315 TRI = Subtarget->getRegisterInfo(); 2316 AA = &getAnalysis<AAResultsWrapperPass>().getAAResults(); 2317 2318 // Resize the modified and used register unit trackers. We do this once 2319 // per function and then clear the register units each time we optimize a load 2320 // or store. 2321 ModifiedRegUnits.init(*TRI); 2322 UsedRegUnits.init(*TRI); 2323 DefinedInBB.init(*TRI); 2324 2325 bool Modified = false; 2326 bool enableNarrowZeroStOpt = !Subtarget->requiresStrictAlign(); 2327 for (auto &MBB : Fn) { 2328 auto M = optimizeBlock(MBB, enableNarrowZeroStOpt); 2329 Modified |= M; 2330 } 2331 2332 return Modified; 2333 } 2334 2335 // FIXME: Do we need/want a pre-alloc pass like ARM has to try to keep loads and 2336 // stores near one another? Note: The pre-RA instruction scheduler already has 2337 // hooks to try and schedule pairable loads/stores together to improve pairing 2338 // opportunities. Thus, pre-RA pairing pass may not be worth the effort. 2339 2340 // FIXME: When pairing store instructions it's very possible for this pass to 2341 // hoist a store with a KILL marker above another use (without a KILL marker). 2342 // The resulting IR is invalid, but nothing uses the KILL markers after this 2343 // pass, so it's never caused a problem in practice. 2344 2345 /// createAArch64LoadStoreOptimizationPass - returns an instance of the 2346 /// load / store optimization pass. 2347 FunctionPass *llvm::createAArch64LoadStoreOptimizationPass() { 2348 return new AArch64LoadStoreOpt(); 2349 } 2350