1 //===- ARMConstantIslandPass.cpp - ARM constant islands -------------------===// 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 splits the constant pool up into 'islands' 10 // which are scattered through-out the function. This is required due to the 11 // limited pc-relative displacements that ARM has. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "ARM.h" 16 #include "ARMBaseInstrInfo.h" 17 #include "ARMBasicBlockInfo.h" 18 #include "ARMMachineFunctionInfo.h" 19 #include "ARMSubtarget.h" 20 #include "MCTargetDesc/ARMBaseInfo.h" 21 #include "MVETailPredUtils.h" 22 #include "Thumb2InstrInfo.h" 23 #include "Utils/ARMBaseInfo.h" 24 #include "llvm/ADT/DenseMap.h" 25 #include "llvm/ADT/STLExtras.h" 26 #include "llvm/ADT/SmallSet.h" 27 #include "llvm/ADT/SmallVector.h" 28 #include "llvm/ADT/Statistic.h" 29 #include "llvm/ADT/StringRef.h" 30 #include "llvm/CodeGen/LivePhysRegs.h" 31 #include "llvm/CodeGen/MachineBasicBlock.h" 32 #include "llvm/CodeGen/MachineConstantPool.h" 33 #include "llvm/CodeGen/MachineDominators.h" 34 #include "llvm/CodeGen/MachineFunction.h" 35 #include "llvm/CodeGen/MachineFunctionPass.h" 36 #include "llvm/CodeGen/MachineInstr.h" 37 #include "llvm/CodeGen/MachineJumpTableInfo.h" 38 #include "llvm/CodeGen/MachineOperand.h" 39 #include "llvm/CodeGen/MachineRegisterInfo.h" 40 #include "llvm/Config/llvm-config.h" 41 #include "llvm/IR/DataLayout.h" 42 #include "llvm/IR/DebugLoc.h" 43 #include "llvm/MC/MCInstrDesc.h" 44 #include "llvm/Pass.h" 45 #include "llvm/Support/CommandLine.h" 46 #include "llvm/Support/Compiler.h" 47 #include "llvm/Support/Debug.h" 48 #include "llvm/Support/ErrorHandling.h" 49 #include "llvm/Support/Format.h" 50 #include "llvm/Support/raw_ostream.h" 51 #include <algorithm> 52 #include <cassert> 53 #include <cstdint> 54 #include <iterator> 55 #include <utility> 56 #include <vector> 57 58 using namespace llvm; 59 60 #define DEBUG_TYPE "arm-cp-islands" 61 62 #define ARM_CP_ISLANDS_OPT_NAME \ 63 "ARM constant island placement and branch shortening pass" 64 STATISTIC(NumCPEs, "Number of constpool entries"); 65 STATISTIC(NumSplit, "Number of uncond branches inserted"); 66 STATISTIC(NumCBrFixed, "Number of cond branches fixed"); 67 STATISTIC(NumUBrFixed, "Number of uncond branches fixed"); 68 STATISTIC(NumTBs, "Number of table branches generated"); 69 STATISTIC(NumT2CPShrunk, "Number of Thumb2 constantpool instructions shrunk"); 70 STATISTIC(NumT2BrShrunk, "Number of Thumb2 immediate branches shrunk"); 71 STATISTIC(NumCBZ, "Number of CBZ / CBNZ formed"); 72 STATISTIC(NumJTMoved, "Number of jump table destination blocks moved"); 73 STATISTIC(NumJTInserted, "Number of jump table intermediate blocks inserted"); 74 STATISTIC(NumLEInserted, "Number of LE backwards branches inserted"); 75 76 static cl::opt<bool> 77 AdjustJumpTableBlocks("arm-adjust-jump-tables", cl::Hidden, cl::init(true), 78 cl::desc("Adjust basic block layout to better use TB[BH]")); 79 80 static cl::opt<unsigned> 81 CPMaxIteration("arm-constant-island-max-iteration", cl::Hidden, cl::init(30), 82 cl::desc("The max number of iteration for converge")); 83 84 static cl::opt<bool> SynthesizeThumb1TBB( 85 "arm-synthesize-thumb-1-tbb", cl::Hidden, cl::init(true), 86 cl::desc("Use compressed jump tables in Thumb-1 by synthesizing an " 87 "equivalent to the TBB/TBH instructions")); 88 89 namespace { 90 91 /// ARMConstantIslands - Due to limited PC-relative displacements, ARM 92 /// requires constant pool entries to be scattered among the instructions 93 /// inside a function. To do this, it completely ignores the normal LLVM 94 /// constant pool; instead, it places constants wherever it feels like with 95 /// special instructions. 96 /// 97 /// The terminology used in this pass includes: 98 /// Islands - Clumps of constants placed in the function. 99 /// Water - Potential places where an island could be formed. 100 /// CPE - A constant pool entry that has been placed somewhere, which 101 /// tracks a list of users. 102 class ARMConstantIslands : public MachineFunctionPass { 103 std::unique_ptr<ARMBasicBlockUtils> BBUtils = nullptr; 104 105 /// WaterList - A sorted list of basic blocks where islands could be placed 106 /// (i.e. blocks that don't fall through to the following block, due 107 /// to a return, unreachable, or unconditional branch). 108 std::vector<MachineBasicBlock*> WaterList; 109 110 /// NewWaterList - The subset of WaterList that was created since the 111 /// previous iteration by inserting unconditional branches. 112 SmallSet<MachineBasicBlock*, 4> NewWaterList; 113 114 using water_iterator = std::vector<MachineBasicBlock *>::iterator; 115 116 /// CPUser - One user of a constant pool, keeping the machine instruction 117 /// pointer, the constant pool being referenced, and the max displacement 118 /// allowed from the instruction to the CP. The HighWaterMark records the 119 /// highest basic block where a new CPEntry can be placed. To ensure this 120 /// pass terminates, the CP entries are initially placed at the end of the 121 /// function and then move monotonically to lower addresses. The 122 /// exception to this rule is when the current CP entry for a particular 123 /// CPUser is out of range, but there is another CP entry for the same 124 /// constant value in range. We want to use the existing in-range CP 125 /// entry, but if it later moves out of range, the search for new water 126 /// should resume where it left off. The HighWaterMark is used to record 127 /// that point. 128 struct CPUser { 129 MachineInstr *MI; 130 MachineInstr *CPEMI; 131 MachineBasicBlock *HighWaterMark; 132 unsigned MaxDisp; 133 bool NegOk; 134 bool IsSoImm; 135 bool KnownAlignment = false; 136 137 CPUser(MachineInstr *mi, MachineInstr *cpemi, unsigned maxdisp, 138 bool neg, bool soimm) 139 : MI(mi), CPEMI(cpemi), MaxDisp(maxdisp), NegOk(neg), IsSoImm(soimm) { 140 HighWaterMark = CPEMI->getParent(); 141 } 142 143 /// getMaxDisp - Returns the maximum displacement supported by MI. 144 /// Correct for unknown alignment. 145 /// Conservatively subtract 2 bytes to handle weird alignment effects. 146 unsigned getMaxDisp() const { 147 return (KnownAlignment ? MaxDisp : MaxDisp - 2) - 2; 148 } 149 }; 150 151 /// CPUsers - Keep track of all of the machine instructions that use various 152 /// constant pools and their max displacement. 153 std::vector<CPUser> CPUsers; 154 155 /// CPEntry - One per constant pool entry, keeping the machine instruction 156 /// pointer, the constpool index, and the number of CPUser's which 157 /// reference this entry. 158 struct CPEntry { 159 MachineInstr *CPEMI; 160 unsigned CPI; 161 unsigned RefCount; 162 163 CPEntry(MachineInstr *cpemi, unsigned cpi, unsigned rc = 0) 164 : CPEMI(cpemi), CPI(cpi), RefCount(rc) {} 165 }; 166 167 /// CPEntries - Keep track of all of the constant pool entry machine 168 /// instructions. For each original constpool index (i.e. those that existed 169 /// upon entry to this pass), it keeps a vector of entries. Original 170 /// elements are cloned as we go along; the clones are put in the vector of 171 /// the original element, but have distinct CPIs. 172 /// 173 /// The first half of CPEntries contains generic constants, the second half 174 /// contains jump tables. Use getCombinedIndex on a generic CPEMI to look up 175 /// which vector it will be in here. 176 std::vector<std::vector<CPEntry>> CPEntries; 177 178 /// Maps a JT index to the offset in CPEntries containing copies of that 179 /// table. The equivalent map for a CONSTPOOL_ENTRY is the identity. 180 DenseMap<int, int> JumpTableEntryIndices; 181 182 /// Maps a JT index to the LEA that actually uses the index to calculate its 183 /// base address. 184 DenseMap<int, int> JumpTableUserIndices; 185 186 // Maps a MachineBasicBlock to the number of jump tables entries. 187 DenseMap<const MachineBasicBlock *, int> BlockJumpTableRefCount; 188 189 /// ImmBranch - One per immediate branch, keeping the machine instruction 190 /// pointer, conditional or unconditional, the max displacement, 191 /// and (if isCond is true) the corresponding unconditional branch 192 /// opcode. 193 struct ImmBranch { 194 MachineInstr *MI; 195 unsigned MaxDisp : 31; 196 bool isCond : 1; 197 unsigned UncondBr; 198 199 ImmBranch(MachineInstr *mi, unsigned maxdisp, bool cond, unsigned ubr) 200 : MI(mi), MaxDisp(maxdisp), isCond(cond), UncondBr(ubr) {} 201 }; 202 203 /// ImmBranches - Keep track of all the immediate branch instructions. 204 std::vector<ImmBranch> ImmBranches; 205 206 /// PushPopMIs - Keep track of all the Thumb push / pop instructions. 207 SmallVector<MachineInstr*, 4> PushPopMIs; 208 209 /// T2JumpTables - Keep track of all the Thumb2 jumptable instructions. 210 SmallVector<MachineInstr*, 4> T2JumpTables; 211 212 MachineFunction *MF; 213 MachineConstantPool *MCP; 214 const ARMBaseInstrInfo *TII; 215 const ARMSubtarget *STI; 216 ARMFunctionInfo *AFI; 217 MachineDominatorTree *DT = nullptr; 218 bool isThumb; 219 bool isThumb1; 220 bool isThumb2; 221 bool isPositionIndependentOrROPI; 222 223 public: 224 static char ID; 225 226 ARMConstantIslands() : MachineFunctionPass(ID) {} 227 228 bool runOnMachineFunction(MachineFunction &MF) override; 229 230 void getAnalysisUsage(AnalysisUsage &AU) const override { 231 AU.addRequired<MachineDominatorTreeWrapperPass>(); 232 MachineFunctionPass::getAnalysisUsage(AU); 233 } 234 235 MachineFunctionProperties getRequiredProperties() const override { 236 return MachineFunctionProperties().set( 237 MachineFunctionProperties::Property::NoVRegs); 238 } 239 240 StringRef getPassName() const override { 241 return ARM_CP_ISLANDS_OPT_NAME; 242 } 243 244 private: 245 void doInitialConstPlacement(std::vector<MachineInstr *> &CPEMIs); 246 void doInitialJumpTablePlacement(std::vector<MachineInstr *> &CPEMIs); 247 bool BBHasFallthrough(MachineBasicBlock *MBB); 248 CPEntry *findConstPoolEntry(unsigned CPI, const MachineInstr *CPEMI); 249 Align getCPEAlign(const MachineInstr *CPEMI); 250 void scanFunctionJumpTables(); 251 void initializeFunctionInfo(const std::vector<MachineInstr*> &CPEMIs); 252 MachineBasicBlock *splitBlockBeforeInstr(MachineInstr *MI); 253 void updateForInsertedWaterBlock(MachineBasicBlock *NewBB); 254 bool decrementCPEReferenceCount(unsigned CPI, MachineInstr* CPEMI); 255 unsigned getCombinedIndex(const MachineInstr *CPEMI); 256 int findInRangeCPEntry(CPUser& U, unsigned UserOffset); 257 bool findAvailableWater(CPUser&U, unsigned UserOffset, 258 water_iterator &WaterIter, bool CloserWater); 259 void createNewWater(unsigned CPUserIndex, unsigned UserOffset, 260 MachineBasicBlock *&NewMBB); 261 bool handleConstantPoolUser(unsigned CPUserIndex, bool CloserWater); 262 void removeDeadCPEMI(MachineInstr *CPEMI); 263 bool removeUnusedCPEntries(); 264 bool isCPEntryInRange(MachineInstr *MI, unsigned UserOffset, 265 MachineInstr *CPEMI, unsigned Disp, bool NegOk, 266 bool DoDump = false); 267 bool isWaterInRange(unsigned UserOffset, MachineBasicBlock *Water, 268 CPUser &U, unsigned &Growth); 269 bool fixupImmediateBr(ImmBranch &Br); 270 bool fixupConditionalBr(ImmBranch &Br); 271 bool fixupUnconditionalBr(ImmBranch &Br); 272 bool optimizeThumb2Instructions(); 273 bool optimizeThumb2Branches(); 274 bool reorderThumb2JumpTables(); 275 bool preserveBaseRegister(MachineInstr *JumpMI, MachineInstr *LEAMI, 276 unsigned &DeadSize, bool &CanDeleteLEA, 277 bool &BaseRegKill); 278 bool optimizeThumb2JumpTables(); 279 MachineBasicBlock *adjustJTTargetBlockForward(unsigned JTI, 280 MachineBasicBlock *BB, 281 MachineBasicBlock *JTBB); 282 283 unsigned getUserOffset(CPUser&) const; 284 void dumpBBs(); 285 void verify(); 286 287 bool isOffsetInRange(unsigned UserOffset, unsigned TrialOffset, 288 unsigned Disp, bool NegativeOK, bool IsSoImm = false); 289 bool isOffsetInRange(unsigned UserOffset, unsigned TrialOffset, 290 const CPUser &U) { 291 return isOffsetInRange(UserOffset, TrialOffset, 292 U.getMaxDisp(), U.NegOk, U.IsSoImm); 293 } 294 }; 295 296 } // end anonymous namespace 297 298 char ARMConstantIslands::ID = 0; 299 300 /// verify - check BBOffsets, BBSizes, alignment of islands 301 void ARMConstantIslands::verify() { 302 #ifndef NDEBUG 303 BBInfoVector &BBInfo = BBUtils->getBBInfo(); 304 assert(is_sorted(*MF, [&BBInfo](const MachineBasicBlock &LHS, 305 const MachineBasicBlock &RHS) { 306 return BBInfo[LHS.getNumber()].postOffset() < 307 BBInfo[RHS.getNumber()].postOffset(); 308 })); 309 LLVM_DEBUG(dbgs() << "Verifying " << CPUsers.size() << " CP users.\n"); 310 for (CPUser &U : CPUsers) { 311 unsigned UserOffset = getUserOffset(U); 312 // Verify offset using the real max displacement without the safety 313 // adjustment. 314 if (isCPEntryInRange(U.MI, UserOffset, U.CPEMI, U.getMaxDisp()+2, U.NegOk, 315 /* DoDump = */ true)) { 316 LLVM_DEBUG(dbgs() << "OK\n"); 317 continue; 318 } 319 LLVM_DEBUG(dbgs() << "Out of range.\n"); 320 dumpBBs(); 321 LLVM_DEBUG(MF->dump()); 322 llvm_unreachable("Constant pool entry out of range!"); 323 } 324 #endif 325 } 326 327 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 328 /// print block size and offset information - debugging 329 LLVM_DUMP_METHOD void ARMConstantIslands::dumpBBs() { 330 LLVM_DEBUG({ 331 BBInfoVector &BBInfo = BBUtils->getBBInfo(); 332 for (unsigned J = 0, E = BBInfo.size(); J !=E; ++J) { 333 const BasicBlockInfo &BBI = BBInfo[J]; 334 dbgs() << format("%08x %bb.%u\t", BBI.Offset, J) 335 << " kb=" << unsigned(BBI.KnownBits) 336 << " ua=" << unsigned(BBI.Unalign) << " pa=" << Log2(BBI.PostAlign) 337 << format(" size=%#x\n", BBInfo[J].Size); 338 } 339 }); 340 } 341 #endif 342 343 // Align blocks where the previous block does not fall through. This may add 344 // extra NOP's but they will not be executed. It uses the PrefLoopAlignment as a 345 // measure of how much to align, and only runs at CodeGenOptLevel::Aggressive. 346 static bool AlignBlocks(MachineFunction *MF, const ARMSubtarget *STI) { 347 if (MF->getTarget().getOptLevel() != CodeGenOptLevel::Aggressive || 348 MF->getFunction().hasOptSize()) 349 return false; 350 351 auto *TLI = STI->getTargetLowering(); 352 const Align Alignment = TLI->getPrefLoopAlignment(); 353 if (Alignment < 4) 354 return false; 355 356 bool Changed = false; 357 bool PrevCanFallthough = true; 358 for (auto &MBB : *MF) { 359 if (!PrevCanFallthough) { 360 Changed = true; 361 MBB.setAlignment(Alignment); 362 } 363 364 PrevCanFallthough = MBB.canFallThrough(); 365 366 // For LOB's, the ARMLowOverheadLoops pass may remove the unconditional 367 // branch later in the pipeline. 368 if (STI->hasLOB()) { 369 for (const auto &MI : reverse(MBB.terminators())) { 370 if (MI.getOpcode() == ARM::t2B && 371 MI.getOperand(0).getMBB() == MBB.getNextNode()) 372 continue; 373 if (isLoopStart(MI) || MI.getOpcode() == ARM::t2LoopEnd || 374 MI.getOpcode() == ARM::t2LoopEndDec) { 375 PrevCanFallthough = true; 376 break; 377 } 378 // Any other terminator - nothing to do 379 break; 380 } 381 } 382 } 383 384 return Changed; 385 } 386 387 bool ARMConstantIslands::runOnMachineFunction(MachineFunction &mf) { 388 MF = &mf; 389 MCP = mf.getConstantPool(); 390 BBUtils = std::make_unique<ARMBasicBlockUtils>(mf); 391 392 LLVM_DEBUG(dbgs() << "***** ARMConstantIslands: " 393 << MCP->getConstants().size() << " CP entries, aligned to " 394 << MCP->getConstantPoolAlign().value() << " bytes *****\n"); 395 396 STI = &MF->getSubtarget<ARMSubtarget>(); 397 TII = STI->getInstrInfo(); 398 isPositionIndependentOrROPI = 399 STI->getTargetLowering()->isPositionIndependent() || STI->isROPI(); 400 AFI = MF->getInfo<ARMFunctionInfo>(); 401 DT = &getAnalysis<MachineDominatorTreeWrapperPass>().getDomTree(); 402 403 isThumb = AFI->isThumbFunction(); 404 isThumb1 = AFI->isThumb1OnlyFunction(); 405 isThumb2 = AFI->isThumb2Function(); 406 407 bool GenerateTBB = isThumb2 || (isThumb1 && SynthesizeThumb1TBB); 408 // TBB generation code in this constant island pass has not been adapted to 409 // deal with speculation barriers. 410 if (STI->hardenSlsRetBr()) 411 GenerateTBB = false; 412 413 // Renumber all of the machine basic blocks in the function, guaranteeing that 414 // the numbers agree with the position of the block in the function. 415 MF->RenumberBlocks(); 416 DT->updateBlockNumbers(); 417 418 // Try to reorder and otherwise adjust the block layout to make good use 419 // of the TB[BH] instructions. 420 bool MadeChange = false; 421 if (GenerateTBB && AdjustJumpTableBlocks) { 422 scanFunctionJumpTables(); 423 MadeChange |= reorderThumb2JumpTables(); 424 // Data is out of date, so clear it. It'll be re-computed later. 425 T2JumpTables.clear(); 426 // Blocks may have shifted around. Keep the numbering up to date. 427 MF->RenumberBlocks(); 428 DT->updateBlockNumbers(); 429 } 430 431 // Align any non-fallthrough blocks 432 MadeChange |= AlignBlocks(MF, STI); 433 434 // Perform the initial placement of the constant pool entries. To start with, 435 // we put them all at the end of the function. 436 std::vector<MachineInstr*> CPEMIs; 437 if (!MCP->isEmpty()) 438 doInitialConstPlacement(CPEMIs); 439 440 if (MF->getJumpTableInfo()) 441 doInitialJumpTablePlacement(CPEMIs); 442 443 /// The next UID to take is the first unused one. 444 AFI->initPICLabelUId(CPEMIs.size()); 445 446 // Do the initial scan of the function, building up information about the 447 // sizes of each block, the location of all the water, and finding all of the 448 // constant pool users. 449 initializeFunctionInfo(CPEMIs); 450 CPEMIs.clear(); 451 LLVM_DEBUG(dumpBBs()); 452 453 // Functions with jump tables need an alignment of 4 because they use the ADR 454 // instruction, which aligns the PC to 4 bytes before adding an offset. 455 if (!T2JumpTables.empty()) 456 MF->ensureAlignment(Align(4)); 457 458 /// Remove dead constant pool entries. 459 MadeChange |= removeUnusedCPEntries(); 460 461 // Iteratively place constant pool entries and fix up branches until there 462 // is no change. 463 unsigned NoCPIters = 0, NoBRIters = 0; 464 while (true) { 465 LLVM_DEBUG(dbgs() << "Beginning CP iteration #" << NoCPIters << '\n'); 466 bool CPChange = false; 467 for (unsigned i = 0, e = CPUsers.size(); i != e; ++i) 468 // For most inputs, it converges in no more than 5 iterations. 469 // If it doesn't end in 10, the input may have huge BB or many CPEs. 470 // In this case, we will try different heuristics. 471 CPChange |= handleConstantPoolUser(i, NoCPIters >= CPMaxIteration / 2); 472 if (CPChange && ++NoCPIters > CPMaxIteration) 473 report_fatal_error("Constant Island pass failed to converge!"); 474 LLVM_DEBUG(dumpBBs()); 475 476 // Clear NewWaterList now. If we split a block for branches, it should 477 // appear as "new water" for the next iteration of constant pool placement. 478 NewWaterList.clear(); 479 480 LLVM_DEBUG(dbgs() << "Beginning BR iteration #" << NoBRIters << '\n'); 481 bool BRChange = false; 482 for (unsigned i = 0, e = ImmBranches.size(); i != e; ++i) 483 BRChange |= fixupImmediateBr(ImmBranches[i]); 484 if (BRChange && ++NoBRIters > 30) 485 report_fatal_error("Branch Fix Up pass failed to converge!"); 486 LLVM_DEBUG(dumpBBs()); 487 488 if (!CPChange && !BRChange) 489 break; 490 MadeChange = true; 491 } 492 493 // Shrink 32-bit Thumb2 load and store instructions. 494 if (isThumb2 && !STI->prefers32BitThumb()) 495 MadeChange |= optimizeThumb2Instructions(); 496 497 // Shrink 32-bit branch instructions. 498 if (isThumb && STI->hasV8MBaselineOps()) 499 MadeChange |= optimizeThumb2Branches(); 500 501 // Optimize jump tables using TBB / TBH. 502 if (GenerateTBB && !STI->genExecuteOnly()) 503 MadeChange |= optimizeThumb2JumpTables(); 504 505 // After a while, this might be made debug-only, but it is not expensive. 506 verify(); 507 508 // Save the mapping between original and cloned constpool entries. 509 for (unsigned i = 0, e = CPEntries.size(); i != e; ++i) { 510 for (unsigned j = 0, je = CPEntries[i].size(); j != je; ++j) { 511 const CPEntry & CPE = CPEntries[i][j]; 512 if (CPE.CPEMI && CPE.CPEMI->getOperand(1).isCPI()) 513 AFI->recordCPEClone(i, CPE.CPI); 514 } 515 } 516 517 LLVM_DEBUG(dbgs() << '\n'; dumpBBs()); 518 519 BBUtils->clear(); 520 WaterList.clear(); 521 CPUsers.clear(); 522 CPEntries.clear(); 523 JumpTableEntryIndices.clear(); 524 JumpTableUserIndices.clear(); 525 BlockJumpTableRefCount.clear(); 526 ImmBranches.clear(); 527 PushPopMIs.clear(); 528 T2JumpTables.clear(); 529 530 return MadeChange; 531 } 532 533 /// Perform the initial placement of the regular constant pool entries. 534 /// To start with, we put them all at the end of the function. 535 void 536 ARMConstantIslands::doInitialConstPlacement(std::vector<MachineInstr*> &CPEMIs) { 537 // Create the basic block to hold the CPE's. 538 MachineBasicBlock *BB = MF->CreateMachineBasicBlock(); 539 MF->push_back(BB); 540 541 // MachineConstantPool measures alignment in bytes. 542 const Align MaxAlign = MCP->getConstantPoolAlign(); 543 const unsigned MaxLogAlign = Log2(MaxAlign); 544 545 // Mark the basic block as required by the const-pool. 546 BB->setAlignment(MaxAlign); 547 548 // The function needs to be as aligned as the basic blocks. The linker may 549 // move functions around based on their alignment. 550 // Special case: halfword literals still need word alignment on the function. 551 Align FuncAlign = MaxAlign; 552 if (MaxAlign == 2) 553 FuncAlign = Align(4); 554 MF->ensureAlignment(FuncAlign); 555 556 // Order the entries in BB by descending alignment. That ensures correct 557 // alignment of all entries as long as BB is sufficiently aligned. Keep 558 // track of the insertion point for each alignment. We are going to bucket 559 // sort the entries as they are created. 560 SmallVector<MachineBasicBlock::iterator, 8> InsPoint(MaxLogAlign + 1, 561 BB->end()); 562 563 // Add all of the constants from the constant pool to the end block, use an 564 // identity mapping of CPI's to CPE's. 565 const std::vector<MachineConstantPoolEntry> &CPs = MCP->getConstants(); 566 567 const DataLayout &TD = MF->getDataLayout(); 568 for (unsigned i = 0, e = CPs.size(); i != e; ++i) { 569 unsigned Size = CPs[i].getSizeInBytes(TD); 570 Align Alignment = CPs[i].getAlign(); 571 // Verify that all constant pool entries are a multiple of their alignment. 572 // If not, we would have to pad them out so that instructions stay aligned. 573 assert(isAligned(Alignment, Size) && "CP Entry not multiple of 4 bytes!"); 574 575 // Insert CONSTPOOL_ENTRY before entries with a smaller alignment. 576 unsigned LogAlign = Log2(Alignment); 577 MachineBasicBlock::iterator InsAt = InsPoint[LogAlign]; 578 MachineInstr *CPEMI = 579 BuildMI(*BB, InsAt, DebugLoc(), TII->get(ARM::CONSTPOOL_ENTRY)) 580 .addImm(i).addConstantPoolIndex(i).addImm(Size); 581 CPEMIs.push_back(CPEMI); 582 583 // Ensure that future entries with higher alignment get inserted before 584 // CPEMI. This is bucket sort with iterators. 585 for (unsigned a = LogAlign + 1; a <= MaxLogAlign; ++a) 586 if (InsPoint[a] == InsAt) 587 InsPoint[a] = CPEMI; 588 589 // Add a new CPEntry, but no corresponding CPUser yet. 590 CPEntries.emplace_back(1, CPEntry(CPEMI, i)); 591 ++NumCPEs; 592 LLVM_DEBUG(dbgs() << "Moved CPI#" << i << " to end of function, size = " 593 << Size << ", align = " << Alignment.value() << '\n'); 594 } 595 LLVM_DEBUG(BB->dump()); 596 } 597 598 /// Do initial placement of the jump tables. Because Thumb2's TBB and TBH 599 /// instructions can be made more efficient if the jump table immediately 600 /// follows the instruction, it's best to place them immediately next to their 601 /// jumps to begin with. In almost all cases they'll never be moved from that 602 /// position. 603 void ARMConstantIslands::doInitialJumpTablePlacement( 604 std::vector<MachineInstr *> &CPEMIs) { 605 unsigned i = CPEntries.size(); 606 auto MJTI = MF->getJumpTableInfo(); 607 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables(); 608 609 // Only inline jump tables are placed in the function. 610 if (MJTI->getEntryKind() != MachineJumpTableInfo::EK_Inline) 611 return; 612 613 MachineBasicBlock *LastCorrectlyNumberedBB = nullptr; 614 for (MachineBasicBlock &MBB : *MF) { 615 auto MI = MBB.getLastNonDebugInstr(); 616 // Look past potential SpeculationBarriers at end of BB. 617 while (MI != MBB.end() && 618 (isSpeculationBarrierEndBBOpcode(MI->getOpcode()) || 619 MI->isDebugInstr())) 620 --MI; 621 622 if (MI == MBB.end()) 623 continue; 624 625 unsigned JTOpcode; 626 switch (MI->getOpcode()) { 627 default: 628 continue; 629 case ARM::BR_JTadd: 630 case ARM::BR_JTr: 631 case ARM::tBR_JTr: 632 case ARM::BR_JTm_i12: 633 case ARM::BR_JTm_rs: 634 // These instructions are emitted only in ARM or Thumb1 modes which do not 635 // support PACBTI. Hence we don't add BTI instructions in the destination 636 // blocks. 637 assert(!MF->getInfo<ARMFunctionInfo>()->branchTargetEnforcement() && 638 "Branch protection must not be enabled for Arm or Thumb1 modes"); 639 JTOpcode = ARM::JUMPTABLE_ADDRS; 640 break; 641 case ARM::t2BR_JT: 642 JTOpcode = ARM::JUMPTABLE_INSTS; 643 break; 644 case ARM::tTBB_JT: 645 case ARM::t2TBB_JT: 646 JTOpcode = ARM::JUMPTABLE_TBB; 647 break; 648 case ARM::tTBH_JT: 649 case ARM::t2TBH_JT: 650 JTOpcode = ARM::JUMPTABLE_TBH; 651 break; 652 } 653 654 unsigned NumOps = MI->getDesc().getNumOperands(); 655 MachineOperand JTOp = 656 MI->getOperand(NumOps - (MI->isPredicable() ? 2 : 1)); 657 unsigned JTI = JTOp.getIndex(); 658 unsigned Size = JT[JTI].MBBs.size() * sizeof(uint32_t); 659 MachineBasicBlock *JumpTableBB = MF->CreateMachineBasicBlock(); 660 MF->insert(std::next(MachineFunction::iterator(MBB)), JumpTableBB); 661 MachineInstr *CPEMI = BuildMI(*JumpTableBB, JumpTableBB->begin(), 662 DebugLoc(), TII->get(JTOpcode)) 663 .addImm(i++) 664 .addJumpTableIndex(JTI) 665 .addImm(Size); 666 CPEMIs.push_back(CPEMI); 667 CPEntries.emplace_back(1, CPEntry(CPEMI, JTI)); 668 JumpTableEntryIndices.insert(std::make_pair(JTI, CPEntries.size() - 1)); 669 if (!LastCorrectlyNumberedBB) 670 LastCorrectlyNumberedBB = &MBB; 671 } 672 673 // If we did anything then we need to renumber the subsequent blocks. 674 if (LastCorrectlyNumberedBB) { 675 MF->RenumberBlocks(LastCorrectlyNumberedBB); 676 DT->updateBlockNumbers(); 677 } 678 } 679 680 /// BBHasFallthrough - Return true if the specified basic block can fallthrough 681 /// into the block immediately after it. 682 bool ARMConstantIslands::BBHasFallthrough(MachineBasicBlock *MBB) { 683 // Get the next machine basic block in the function. 684 MachineFunction::iterator MBBI = MBB->getIterator(); 685 // Can't fall off end of function. 686 if (std::next(MBBI) == MBB->getParent()->end()) 687 return false; 688 689 MachineBasicBlock *NextBB = &*std::next(MBBI); 690 if (!MBB->isSuccessor(NextBB)) 691 return false; 692 693 // Try to analyze the end of the block. A potential fallthrough may already 694 // have an unconditional branch for whatever reason. 695 MachineBasicBlock *TBB, *FBB; 696 SmallVector<MachineOperand, 4> Cond; 697 bool TooDifficult = TII->analyzeBranch(*MBB, TBB, FBB, Cond); 698 return TooDifficult || FBB == nullptr; 699 } 700 701 /// findConstPoolEntry - Given the constpool index and CONSTPOOL_ENTRY MI, 702 /// look up the corresponding CPEntry. 703 ARMConstantIslands::CPEntry * 704 ARMConstantIslands::findConstPoolEntry(unsigned CPI, 705 const MachineInstr *CPEMI) { 706 std::vector<CPEntry> &CPEs = CPEntries[CPI]; 707 // Number of entries per constpool index should be small, just do a 708 // linear search. 709 for (CPEntry &CPE : CPEs) 710 if (CPE.CPEMI == CPEMI) 711 return &CPE; 712 return nullptr; 713 } 714 715 /// getCPEAlign - Returns the required alignment of the constant pool entry 716 /// represented by CPEMI. 717 Align ARMConstantIslands::getCPEAlign(const MachineInstr *CPEMI) { 718 switch (CPEMI->getOpcode()) { 719 case ARM::CONSTPOOL_ENTRY: 720 break; 721 case ARM::JUMPTABLE_TBB: 722 return isThumb1 ? Align(4) : Align(1); 723 case ARM::JUMPTABLE_TBH: 724 return isThumb1 ? Align(4) : Align(2); 725 case ARM::JUMPTABLE_INSTS: 726 return Align(2); 727 case ARM::JUMPTABLE_ADDRS: 728 return Align(4); 729 default: 730 llvm_unreachable("unknown constpool entry kind"); 731 } 732 733 unsigned CPI = getCombinedIndex(CPEMI); 734 assert(CPI < MCP->getConstants().size() && "Invalid constant pool index."); 735 return MCP->getConstants()[CPI].getAlign(); 736 } 737 738 // Exception landing pads, blocks that has their adress taken, and function 739 // entry blocks will always be (potential) indirect jump targets, regardless of 740 // whether they are referenced by or not by jump tables. 741 static bool isAlwaysIndirectTarget(const MachineBasicBlock &MBB) { 742 return MBB.isEHPad() || MBB.hasAddressTaken() || 743 &MBB == &MBB.getParent()->front(); 744 } 745 746 /// scanFunctionJumpTables - Do a scan of the function, building up 747 /// information about the sizes of each block and the locations of all 748 /// the jump tables. 749 void ARMConstantIslands::scanFunctionJumpTables() { 750 for (MachineBasicBlock &MBB : *MF) { 751 for (MachineInstr &I : MBB) 752 if (I.isBranch() && 753 (I.getOpcode() == ARM::t2BR_JT || I.getOpcode() == ARM::tBR_JTr)) 754 T2JumpTables.push_back(&I); 755 } 756 757 if (!MF->getInfo<ARMFunctionInfo>()->branchTargetEnforcement()) 758 return; 759 760 if (const MachineJumpTableInfo *JTI = MF->getJumpTableInfo()) 761 for (const MachineJumpTableEntry &JTE : JTI->getJumpTables()) 762 for (const MachineBasicBlock *MBB : JTE.MBBs) { 763 if (isAlwaysIndirectTarget(*MBB)) 764 // Set the reference count essentially to infinity, it will never 765 // reach zero and the BTI Instruction will never be removed. 766 BlockJumpTableRefCount[MBB] = std::numeric_limits<int>::max(); 767 else 768 ++BlockJumpTableRefCount[MBB]; 769 } 770 } 771 772 /// initializeFunctionInfo - Do the initial scan of the function, building up 773 /// information about the sizes of each block, the location of all the water, 774 /// and finding all of the constant pool users. 775 void ARMConstantIslands:: 776 initializeFunctionInfo(const std::vector<MachineInstr*> &CPEMIs) { 777 778 BBUtils->computeAllBlockSizes(); 779 BBInfoVector &BBInfo = BBUtils->getBBInfo(); 780 // The known bits of the entry block offset are determined by the function 781 // alignment. 782 BBInfo.front().KnownBits = Log2(MF->getAlignment()); 783 784 // Compute block offsets and known bits. 785 BBUtils->adjustBBOffsetsAfter(&MF->front()); 786 787 // We only care about jump table instructions when jump tables are inline. 788 MachineJumpTableInfo *MJTI = MF->getJumpTableInfo(); 789 bool InlineJumpTables = 790 MJTI && MJTI->getEntryKind() == MachineJumpTableInfo::EK_Inline; 791 792 // Now go back through the instructions and build up our data structures. 793 for (MachineBasicBlock &MBB : *MF) { 794 // If this block doesn't fall through into the next MBB, then this is 795 // 'water' that a constant pool island could be placed. 796 if (!BBHasFallthrough(&MBB)) 797 WaterList.push_back(&MBB); 798 799 for (MachineInstr &I : MBB) { 800 if (I.isDebugInstr()) 801 continue; 802 803 unsigned Opc = I.getOpcode(); 804 if (I.isBranch()) { 805 bool isCond = false; 806 unsigned Bits = 0; 807 unsigned Scale = 1; 808 int UOpc = Opc; 809 switch (Opc) { 810 default: 811 continue; // Ignore other JT branches 812 case ARM::t2BR_JT: 813 case ARM::tBR_JTr: 814 if (InlineJumpTables) 815 T2JumpTables.push_back(&I); 816 continue; // Does not get an entry in ImmBranches 817 case ARM::Bcc: 818 isCond = true; 819 UOpc = ARM::B; 820 [[fallthrough]]; 821 case ARM::B: 822 Bits = 24; 823 Scale = 4; 824 break; 825 case ARM::tBcc: 826 isCond = true; 827 UOpc = ARM::tB; 828 Bits = 8; 829 Scale = 2; 830 break; 831 case ARM::tB: 832 Bits = 11; 833 Scale = 2; 834 break; 835 case ARM::t2Bcc: 836 isCond = true; 837 UOpc = ARM::t2B; 838 Bits = 20; 839 Scale = 2; 840 break; 841 case ARM::t2B: 842 Bits = 24; 843 Scale = 2; 844 break; 845 } 846 847 // Record this immediate branch. 848 unsigned MaxOffs = ((1 << (Bits-1))-1) * Scale; 849 ImmBranches.push_back(ImmBranch(&I, MaxOffs, isCond, UOpc)); 850 } 851 852 if (Opc == ARM::tPUSH || Opc == ARM::tPOP_RET) 853 PushPopMIs.push_back(&I); 854 855 if (Opc == ARM::CONSTPOOL_ENTRY || Opc == ARM::JUMPTABLE_ADDRS || 856 Opc == ARM::JUMPTABLE_INSTS || Opc == ARM::JUMPTABLE_TBB || 857 Opc == ARM::JUMPTABLE_TBH) 858 continue; 859 860 // Scan the instructions for constant pool operands. 861 for (unsigned op = 0, e = I.getNumOperands(); op != e; ++op) 862 if (I.getOperand(op).isCPI() || 863 (I.getOperand(op).isJTI() && InlineJumpTables)) { 864 // We found one. The addressing mode tells us the max displacement 865 // from the PC that this instruction permits. 866 867 // Basic size info comes from the TSFlags field. 868 unsigned Bits = 0; 869 unsigned Scale = 1; 870 bool NegOk = false; 871 bool IsSoImm = false; 872 873 switch (Opc) { 874 default: 875 llvm_unreachable("Unknown addressing mode for CP reference!"); 876 877 // Taking the address of a CP entry. 878 case ARM::LEApcrel: 879 case ARM::LEApcrelJT: { 880 // This takes a SoImm, which is 8 bit immediate rotated. We'll 881 // pretend the maximum offset is 255 * 4. Since each instruction 882 // 4 byte wide, this is always correct. We'll check for other 883 // displacements that fits in a SoImm as well. 884 Bits = 8; 885 NegOk = true; 886 IsSoImm = true; 887 unsigned CPI = I.getOperand(op).getIndex(); 888 assert(CPI < CPEMIs.size()); 889 MachineInstr *CPEMI = CPEMIs[CPI]; 890 const Align CPEAlign = getCPEAlign(CPEMI); 891 const unsigned LogCPEAlign = Log2(CPEAlign); 892 if (LogCPEAlign >= 2) 893 Scale = 4; 894 else 895 // For constants with less than 4-byte alignment, 896 // we'll pretend the maximum offset is 255 * 1. 897 Scale = 1; 898 } 899 break; 900 case ARM::t2LEApcrel: 901 case ARM::t2LEApcrelJT: 902 Bits = 12; 903 NegOk = true; 904 break; 905 case ARM::tLEApcrel: 906 case ARM::tLEApcrelJT: 907 Bits = 8; 908 Scale = 4; 909 break; 910 911 case ARM::LDRBi12: 912 case ARM::LDRi12: 913 case ARM::LDRcp: 914 case ARM::t2LDRpci: 915 case ARM::t2LDRHpci: 916 case ARM::t2LDRSHpci: 917 case ARM::t2LDRBpci: 918 case ARM::t2LDRSBpci: 919 Bits = 12; // +-offset_12 920 NegOk = true; 921 break; 922 923 case ARM::tLDRpci: 924 Bits = 8; 925 Scale = 4; // +(offset_8*4) 926 break; 927 928 case ARM::VLDRD: 929 case ARM::VLDRS: 930 Bits = 8; 931 Scale = 4; // +-(offset_8*4) 932 NegOk = true; 933 break; 934 case ARM::VLDRH: 935 Bits = 8; 936 Scale = 2; // +-(offset_8*2) 937 NegOk = true; 938 break; 939 } 940 941 // Remember that this is a user of a CP entry. 942 unsigned CPI = I.getOperand(op).getIndex(); 943 if (I.getOperand(op).isJTI()) { 944 JumpTableUserIndices.insert(std::make_pair(CPI, CPUsers.size())); 945 CPI = JumpTableEntryIndices[CPI]; 946 } 947 948 MachineInstr *CPEMI = CPEMIs[CPI]; 949 unsigned MaxOffs = ((1 << Bits)-1) * Scale; 950 CPUsers.push_back(CPUser(&I, CPEMI, MaxOffs, NegOk, IsSoImm)); 951 952 // Increment corresponding CPEntry reference count. 953 CPEntry *CPE = findConstPoolEntry(CPI, CPEMI); 954 assert(CPE && "Cannot find a corresponding CPEntry!"); 955 CPE->RefCount++; 956 957 // Instructions can only use one CP entry, don't bother scanning the 958 // rest of the operands. 959 break; 960 } 961 } 962 } 963 } 964 965 /// CompareMBBNumbers - Little predicate function to sort the WaterList by MBB 966 /// ID. 967 static bool CompareMBBNumbers(const MachineBasicBlock *LHS, 968 const MachineBasicBlock *RHS) { 969 return LHS->getNumber() < RHS->getNumber(); 970 } 971 972 /// updateForInsertedWaterBlock - When a block is newly inserted into the 973 /// machine function, it upsets all of the block numbers. Renumber the blocks 974 /// and update the arrays that parallel this numbering. 975 void ARMConstantIslands::updateForInsertedWaterBlock(MachineBasicBlock *NewBB) { 976 // Renumber the MBB's to keep them consecutive. 977 NewBB->getParent()->RenumberBlocks(NewBB); 978 DT->updateBlockNumbers(); 979 980 // Insert an entry into BBInfo to align it properly with the (newly 981 // renumbered) block numbers. 982 BBUtils->insert(NewBB->getNumber(), BasicBlockInfo()); 983 984 // Next, update WaterList. Specifically, we need to add NewMBB as having 985 // available water after it. 986 water_iterator IP = llvm::lower_bound(WaterList, NewBB, CompareMBBNumbers); 987 WaterList.insert(IP, NewBB); 988 } 989 990 /// Split the basic block containing MI into two blocks, which are joined by 991 /// an unconditional branch. Update data structures and renumber blocks to 992 /// account for this change and returns the newly created block. 993 MachineBasicBlock *ARMConstantIslands::splitBlockBeforeInstr(MachineInstr *MI) { 994 MachineBasicBlock *OrigBB = MI->getParent(); 995 996 // Collect liveness information at MI. 997 LivePhysRegs LRs(*MF->getSubtarget().getRegisterInfo()); 998 LRs.addLiveOuts(*OrigBB); 999 auto LivenessEnd = ++MachineBasicBlock::iterator(MI).getReverse(); 1000 for (MachineInstr &LiveMI : make_range(OrigBB->rbegin(), LivenessEnd)) 1001 LRs.stepBackward(LiveMI); 1002 1003 // Create a new MBB for the code after the OrigBB. 1004 MachineBasicBlock *NewBB = 1005 MF->CreateMachineBasicBlock(OrigBB->getBasicBlock()); 1006 MachineFunction::iterator MBBI = ++OrigBB->getIterator(); 1007 MF->insert(MBBI, NewBB); 1008 1009 // Splice the instructions starting with MI over to NewBB. 1010 NewBB->splice(NewBB->end(), OrigBB, MI, OrigBB->end()); 1011 1012 // Add an unconditional branch from OrigBB to NewBB. 1013 // Note the new unconditional branch is not being recorded. 1014 // There doesn't seem to be meaningful DebugInfo available; this doesn't 1015 // correspond to anything in the source. 1016 unsigned Opc = isThumb ? (isThumb2 ? ARM::t2B : ARM::tB) : ARM::B; 1017 if (!isThumb) 1018 BuildMI(OrigBB, DebugLoc(), TII->get(Opc)).addMBB(NewBB); 1019 else 1020 BuildMI(OrigBB, DebugLoc(), TII->get(Opc)) 1021 .addMBB(NewBB) 1022 .add(predOps(ARMCC::AL)); 1023 ++NumSplit; 1024 1025 // Update the CFG. All succs of OrigBB are now succs of NewBB. 1026 NewBB->transferSuccessors(OrigBB); 1027 1028 // OrigBB branches to NewBB. 1029 OrigBB->addSuccessor(NewBB); 1030 1031 // Update live-in information in the new block. 1032 MachineRegisterInfo &MRI = MF->getRegInfo(); 1033 for (MCPhysReg L : LRs) 1034 if (!MRI.isReserved(L)) 1035 NewBB->addLiveIn(L); 1036 1037 // Update internal data structures to account for the newly inserted MBB. 1038 // This is almost the same as updateForInsertedWaterBlock, except that 1039 // the Water goes after OrigBB, not NewBB. 1040 MF->RenumberBlocks(NewBB); 1041 DT->updateBlockNumbers(); 1042 1043 // Insert an entry into BBInfo to align it properly with the (newly 1044 // renumbered) block numbers. 1045 BBUtils->insert(NewBB->getNumber(), BasicBlockInfo()); 1046 1047 // Next, update WaterList. Specifically, we need to add OrigMBB as having 1048 // available water after it (but not if it's already there, which happens 1049 // when splitting before a conditional branch that is followed by an 1050 // unconditional branch - in that case we want to insert NewBB). 1051 water_iterator IP = llvm::lower_bound(WaterList, OrigBB, CompareMBBNumbers); 1052 MachineBasicBlock* WaterBB = *IP; 1053 if (WaterBB == OrigBB) 1054 WaterList.insert(std::next(IP), NewBB); 1055 else 1056 WaterList.insert(IP, OrigBB); 1057 NewWaterList.insert(OrigBB); 1058 1059 // Figure out how large the OrigBB is. As the first half of the original 1060 // block, it cannot contain a tablejump. The size includes 1061 // the new jump we added. (It should be possible to do this without 1062 // recounting everything, but it's very confusing, and this is rarely 1063 // executed.) 1064 BBUtils->computeBlockSize(OrigBB); 1065 1066 // Figure out how large the NewMBB is. As the second half of the original 1067 // block, it may contain a tablejump. 1068 BBUtils->computeBlockSize(NewBB); 1069 1070 // All BBOffsets following these blocks must be modified. 1071 BBUtils->adjustBBOffsetsAfter(OrigBB); 1072 1073 return NewBB; 1074 } 1075 1076 /// getUserOffset - Compute the offset of U.MI as seen by the hardware 1077 /// displacement computation. Update U.KnownAlignment to match its current 1078 /// basic block location. 1079 unsigned ARMConstantIslands::getUserOffset(CPUser &U) const { 1080 unsigned UserOffset = BBUtils->getOffsetOf(U.MI); 1081 1082 SmallVectorImpl<BasicBlockInfo> &BBInfo = BBUtils->getBBInfo(); 1083 const BasicBlockInfo &BBI = BBInfo[U.MI->getParent()->getNumber()]; 1084 unsigned KnownBits = BBI.internalKnownBits(); 1085 1086 // The value read from PC is offset from the actual instruction address. 1087 UserOffset += (isThumb ? 4 : 8); 1088 1089 // Because of inline assembly, we may not know the alignment (mod 4) of U.MI. 1090 // Make sure U.getMaxDisp() returns a constrained range. 1091 U.KnownAlignment = (KnownBits >= 2); 1092 1093 // On Thumb, offsets==2 mod 4 are rounded down by the hardware for 1094 // purposes of the displacement computation; compensate for that here. 1095 // For unknown alignments, getMaxDisp() constrains the range instead. 1096 if (isThumb && U.KnownAlignment) 1097 UserOffset &= ~3u; 1098 1099 return UserOffset; 1100 } 1101 1102 /// isOffsetInRange - Checks whether UserOffset (the location of a constant pool 1103 /// reference) is within MaxDisp of TrialOffset (a proposed location of a 1104 /// constant pool entry). 1105 /// UserOffset is computed by getUserOffset above to include PC adjustments. If 1106 /// the mod 4 alignment of UserOffset is not known, the uncertainty must be 1107 /// subtracted from MaxDisp instead. CPUser::getMaxDisp() does that. 1108 bool ARMConstantIslands::isOffsetInRange(unsigned UserOffset, 1109 unsigned TrialOffset, unsigned MaxDisp, 1110 bool NegativeOK, bool IsSoImm) { 1111 if (UserOffset <= TrialOffset) { 1112 // User before the Trial. 1113 if (TrialOffset - UserOffset <= MaxDisp) 1114 return true; 1115 // FIXME: Make use full range of soimm values. 1116 } else if (NegativeOK) { 1117 if (UserOffset - TrialOffset <= MaxDisp) 1118 return true; 1119 // FIXME: Make use full range of soimm values. 1120 } 1121 return false; 1122 } 1123 1124 /// isWaterInRange - Returns true if a CPE placed after the specified 1125 /// Water (a basic block) will be in range for the specific MI. 1126 /// 1127 /// Compute how much the function will grow by inserting a CPE after Water. 1128 bool ARMConstantIslands::isWaterInRange(unsigned UserOffset, 1129 MachineBasicBlock* Water, CPUser &U, 1130 unsigned &Growth) { 1131 BBInfoVector &BBInfo = BBUtils->getBBInfo(); 1132 const Align CPEAlign = getCPEAlign(U.CPEMI); 1133 const unsigned CPEOffset = BBInfo[Water->getNumber()].postOffset(CPEAlign); 1134 unsigned NextBlockOffset; 1135 Align NextBlockAlignment; 1136 MachineFunction::const_iterator NextBlock = Water->getIterator(); 1137 if (++NextBlock == MF->end()) { 1138 NextBlockOffset = BBInfo[Water->getNumber()].postOffset(); 1139 } else { 1140 NextBlockOffset = BBInfo[NextBlock->getNumber()].Offset; 1141 NextBlockAlignment = NextBlock->getAlignment(); 1142 } 1143 unsigned Size = U.CPEMI->getOperand(2).getImm(); 1144 unsigned CPEEnd = CPEOffset + Size; 1145 1146 // The CPE may be able to hide in the alignment padding before the next 1147 // block. It may also cause more padding to be required if it is more aligned 1148 // that the next block. 1149 if (CPEEnd > NextBlockOffset) { 1150 Growth = CPEEnd - NextBlockOffset; 1151 // Compute the padding that would go at the end of the CPE to align the next 1152 // block. 1153 Growth += offsetToAlignment(CPEEnd, NextBlockAlignment); 1154 1155 // If the CPE is to be inserted before the instruction, that will raise 1156 // the offset of the instruction. Also account for unknown alignment padding 1157 // in blocks between CPE and the user. 1158 if (CPEOffset < UserOffset) 1159 UserOffset += Growth + UnknownPadding(MF->getAlignment(), Log2(CPEAlign)); 1160 } else 1161 // CPE fits in existing padding. 1162 Growth = 0; 1163 1164 return isOffsetInRange(UserOffset, CPEOffset, U); 1165 } 1166 1167 /// isCPEntryInRange - Returns true if the distance between specific MI and 1168 /// specific ConstPool entry instruction can fit in MI's displacement field. 1169 bool ARMConstantIslands::isCPEntryInRange(MachineInstr *MI, unsigned UserOffset, 1170 MachineInstr *CPEMI, unsigned MaxDisp, 1171 bool NegOk, bool DoDump) { 1172 unsigned CPEOffset = BBUtils->getOffsetOf(CPEMI); 1173 1174 if (DoDump) { 1175 LLVM_DEBUG({ 1176 BBInfoVector &BBInfo = BBUtils->getBBInfo(); 1177 unsigned Block = MI->getParent()->getNumber(); 1178 const BasicBlockInfo &BBI = BBInfo[Block]; 1179 dbgs() << "User of CPE#" << CPEMI->getOperand(0).getImm() 1180 << " max delta=" << MaxDisp 1181 << format(" insn address=%#x", UserOffset) << " in " 1182 << printMBBReference(*MI->getParent()) << ": " 1183 << format("%#x-%x\t", BBI.Offset, BBI.postOffset()) << *MI 1184 << format("CPE address=%#x offset=%+d: ", CPEOffset, 1185 int(CPEOffset - UserOffset)); 1186 }); 1187 } 1188 1189 return isOffsetInRange(UserOffset, CPEOffset, MaxDisp, NegOk); 1190 } 1191 1192 #ifndef NDEBUG 1193 /// BBIsJumpedOver - Return true of the specified basic block's only predecessor 1194 /// unconditionally branches to its only successor. 1195 static bool BBIsJumpedOver(MachineBasicBlock *MBB) { 1196 if (MBB->pred_size() != 1 || MBB->succ_size() != 1) 1197 return false; 1198 1199 MachineBasicBlock *Succ = *MBB->succ_begin(); 1200 MachineBasicBlock *Pred = *MBB->pred_begin(); 1201 MachineInstr *PredMI = &Pred->back(); 1202 if (PredMI->getOpcode() == ARM::B || PredMI->getOpcode() == ARM::tB 1203 || PredMI->getOpcode() == ARM::t2B) 1204 return PredMI->getOperand(0).getMBB() == Succ; 1205 return false; 1206 } 1207 #endif // NDEBUG 1208 1209 /// decrementCPEReferenceCount - find the constant pool entry with index CPI 1210 /// and instruction CPEMI, and decrement its refcount. If the refcount 1211 /// becomes 0 remove the entry and instruction. Returns true if we removed 1212 /// the entry, false if we didn't. 1213 bool ARMConstantIslands::decrementCPEReferenceCount(unsigned CPI, 1214 MachineInstr *CPEMI) { 1215 // Find the old entry. Eliminate it if it is no longer used. 1216 CPEntry *CPE = findConstPoolEntry(CPI, CPEMI); 1217 assert(CPE && "Unexpected!"); 1218 if (--CPE->RefCount == 0) { 1219 removeDeadCPEMI(CPEMI); 1220 CPE->CPEMI = nullptr; 1221 --NumCPEs; 1222 return true; 1223 } 1224 return false; 1225 } 1226 1227 unsigned ARMConstantIslands::getCombinedIndex(const MachineInstr *CPEMI) { 1228 if (CPEMI->getOperand(1).isCPI()) 1229 return CPEMI->getOperand(1).getIndex(); 1230 1231 return JumpTableEntryIndices[CPEMI->getOperand(1).getIndex()]; 1232 } 1233 1234 /// LookForCPEntryInRange - see if the currently referenced CPE is in range; 1235 /// if not, see if an in-range clone of the CPE is in range, and if so, 1236 /// change the data structures so the user references the clone. Returns: 1237 /// 0 = no existing entry found 1238 /// 1 = entry found, and there were no code insertions or deletions 1239 /// 2 = entry found, and there were code insertions or deletions 1240 int ARMConstantIslands::findInRangeCPEntry(CPUser& U, unsigned UserOffset) { 1241 MachineInstr *UserMI = U.MI; 1242 MachineInstr *CPEMI = U.CPEMI; 1243 1244 // Check to see if the CPE is already in-range. 1245 if (isCPEntryInRange(UserMI, UserOffset, CPEMI, U.getMaxDisp(), U.NegOk, 1246 true)) { 1247 LLVM_DEBUG(dbgs() << "In range\n"); 1248 return 1; 1249 } 1250 1251 // No. Look for previously created clones of the CPE that are in range. 1252 unsigned CPI = getCombinedIndex(CPEMI); 1253 std::vector<CPEntry> &CPEs = CPEntries[CPI]; 1254 for (CPEntry &CPE : CPEs) { 1255 // We already tried this one 1256 if (CPE.CPEMI == CPEMI) 1257 continue; 1258 // Removing CPEs can leave empty entries, skip 1259 if (CPE.CPEMI == nullptr) 1260 continue; 1261 if (isCPEntryInRange(UserMI, UserOffset, CPE.CPEMI, U.getMaxDisp(), 1262 U.NegOk)) { 1263 LLVM_DEBUG(dbgs() << "Replacing CPE#" << CPI << " with CPE#" << CPE.CPI 1264 << "\n"); 1265 // Point the CPUser node to the replacement 1266 U.CPEMI = CPE.CPEMI; 1267 // Change the CPI in the instruction operand to refer to the clone. 1268 for (MachineOperand &MO : UserMI->operands()) 1269 if (MO.isCPI()) { 1270 MO.setIndex(CPE.CPI); 1271 break; 1272 } 1273 // Adjust the refcount of the clone... 1274 CPE.RefCount++; 1275 // ...and the original. If we didn't remove the old entry, none of the 1276 // addresses changed, so we don't need another pass. 1277 return decrementCPEReferenceCount(CPI, CPEMI) ? 2 : 1; 1278 } 1279 } 1280 return 0; 1281 } 1282 1283 /// getUnconditionalBrDisp - Returns the maximum displacement that can fit in 1284 /// the specific unconditional branch instruction. 1285 static inline unsigned getUnconditionalBrDisp(int Opc) { 1286 switch (Opc) { 1287 case ARM::tB: 1288 return ((1<<10)-1)*2; 1289 case ARM::t2B: 1290 return ((1<<23)-1)*2; 1291 default: 1292 break; 1293 } 1294 1295 return ((1<<23)-1)*4; 1296 } 1297 1298 /// findAvailableWater - Look for an existing entry in the WaterList in which 1299 /// we can place the CPE referenced from U so it's within range of U's MI. 1300 /// Returns true if found, false if not. If it returns true, WaterIter 1301 /// is set to the WaterList entry. For Thumb, prefer water that will not 1302 /// introduce padding to water that will. To ensure that this pass 1303 /// terminates, the CPE location for a particular CPUser is only allowed to 1304 /// move to a lower address, so search backward from the end of the list and 1305 /// prefer the first water that is in range. 1306 bool ARMConstantIslands::findAvailableWater(CPUser &U, unsigned UserOffset, 1307 water_iterator &WaterIter, 1308 bool CloserWater) { 1309 if (WaterList.empty()) 1310 return false; 1311 1312 unsigned BestGrowth = ~0u; 1313 // The nearest water without splitting the UserBB is right after it. 1314 // If the distance is still large (we have a big BB), then we need to split it 1315 // if we don't converge after certain iterations. This helps the following 1316 // situation to converge: 1317 // BB0: 1318 // Big BB 1319 // BB1: 1320 // Constant Pool 1321 // When a CP access is out of range, BB0 may be used as water. However, 1322 // inserting islands between BB0 and BB1 makes other accesses out of range. 1323 MachineBasicBlock *UserBB = U.MI->getParent(); 1324 BBInfoVector &BBInfo = BBUtils->getBBInfo(); 1325 const Align CPEAlign = getCPEAlign(U.CPEMI); 1326 unsigned MinNoSplitDisp = 1327 BBInfo[UserBB->getNumber()].postOffset(CPEAlign) - UserOffset; 1328 if (CloserWater && MinNoSplitDisp > U.getMaxDisp() / 2) 1329 return false; 1330 for (water_iterator IP = std::prev(WaterList.end()), B = WaterList.begin();; 1331 --IP) { 1332 MachineBasicBlock* WaterBB = *IP; 1333 // Check if water is in range and is either at a lower address than the 1334 // current "high water mark" or a new water block that was created since 1335 // the previous iteration by inserting an unconditional branch. In the 1336 // latter case, we want to allow resetting the high water mark back to 1337 // this new water since we haven't seen it before. Inserting branches 1338 // should be relatively uncommon and when it does happen, we want to be 1339 // sure to take advantage of it for all the CPEs near that block, so that 1340 // we don't insert more branches than necessary. 1341 // When CloserWater is true, we try to find the lowest address after (or 1342 // equal to) user MI's BB no matter of padding growth. 1343 unsigned Growth; 1344 if (isWaterInRange(UserOffset, WaterBB, U, Growth) && 1345 (WaterBB->getNumber() < U.HighWaterMark->getNumber() || 1346 NewWaterList.count(WaterBB) || WaterBB == U.MI->getParent()) && 1347 Growth < BestGrowth) { 1348 // This is the least amount of required padding seen so far. 1349 BestGrowth = Growth; 1350 WaterIter = IP; 1351 LLVM_DEBUG(dbgs() << "Found water after " << printMBBReference(*WaterBB) 1352 << " Growth=" << Growth << '\n'); 1353 1354 if (CloserWater && WaterBB == U.MI->getParent()) 1355 return true; 1356 // Keep looking unless it is perfect and we're not looking for the lowest 1357 // possible address. 1358 if (!CloserWater && BestGrowth == 0) 1359 return true; 1360 } 1361 if (IP == B) 1362 break; 1363 } 1364 return BestGrowth != ~0u; 1365 } 1366 1367 /// createNewWater - No existing WaterList entry will work for 1368 /// CPUsers[CPUserIndex], so create a place to put the CPE. The end of the 1369 /// block is used if in range, and the conditional branch munged so control 1370 /// flow is correct. Otherwise the block is split to create a hole with an 1371 /// unconditional branch around it. In either case NewMBB is set to a 1372 /// block following which the new island can be inserted (the WaterList 1373 /// is not adjusted). 1374 void ARMConstantIslands::createNewWater(unsigned CPUserIndex, 1375 unsigned UserOffset, 1376 MachineBasicBlock *&NewMBB) { 1377 CPUser &U = CPUsers[CPUserIndex]; 1378 MachineInstr *UserMI = U.MI; 1379 MachineInstr *CPEMI = U.CPEMI; 1380 const Align CPEAlign = getCPEAlign(CPEMI); 1381 MachineBasicBlock *UserMBB = UserMI->getParent(); 1382 BBInfoVector &BBInfo = BBUtils->getBBInfo(); 1383 const BasicBlockInfo &UserBBI = BBInfo[UserMBB->getNumber()]; 1384 1385 // If the block does not end in an unconditional branch already, and if the 1386 // end of the block is within range, make new water there. (The addition 1387 // below is for the unconditional branch we will be adding: 4 bytes on ARM + 1388 // Thumb2, 2 on Thumb1. 1389 if (BBHasFallthrough(UserMBB)) { 1390 // Size of branch to insert. 1391 unsigned Delta = isThumb1 ? 2 : 4; 1392 // Compute the offset where the CPE will begin. 1393 unsigned CPEOffset = UserBBI.postOffset(CPEAlign) + Delta; 1394 1395 if (isOffsetInRange(UserOffset, CPEOffset, U)) { 1396 LLVM_DEBUG(dbgs() << "Split at end of " << printMBBReference(*UserMBB) 1397 << format(", expected CPE offset %#x\n", CPEOffset)); 1398 NewMBB = &*++UserMBB->getIterator(); 1399 // Add an unconditional branch from UserMBB to fallthrough block. Record 1400 // it for branch lengthening; this new branch will not get out of range, 1401 // but if the preceding conditional branch is out of range, the targets 1402 // will be exchanged, and the altered branch may be out of range, so the 1403 // machinery has to know about it. 1404 int UncondBr = isThumb ? ((isThumb2) ? ARM::t2B : ARM::tB) : ARM::B; 1405 if (!isThumb) 1406 BuildMI(UserMBB, DebugLoc(), TII->get(UncondBr)).addMBB(NewMBB); 1407 else 1408 BuildMI(UserMBB, DebugLoc(), TII->get(UncondBr)) 1409 .addMBB(NewMBB) 1410 .add(predOps(ARMCC::AL)); 1411 unsigned MaxDisp = getUnconditionalBrDisp(UncondBr); 1412 ImmBranches.push_back(ImmBranch(&UserMBB->back(), 1413 MaxDisp, false, UncondBr)); 1414 BBUtils->computeBlockSize(UserMBB); 1415 BBUtils->adjustBBOffsetsAfter(UserMBB); 1416 return; 1417 } 1418 } 1419 1420 // What a big block. Find a place within the block to split it. This is a 1421 // little tricky on Thumb1 since instructions are 2 bytes and constant pool 1422 // entries are 4 bytes: if instruction I references island CPE, and 1423 // instruction I+1 references CPE', it will not work well to put CPE as far 1424 // forward as possible, since then CPE' cannot immediately follow it (that 1425 // location is 2 bytes farther away from I+1 than CPE was from I) and we'd 1426 // need to create a new island. So, we make a first guess, then walk through 1427 // the instructions between the one currently being looked at and the 1428 // possible insertion point, and make sure any other instructions that 1429 // reference CPEs will be able to use the same island area; if not, we back 1430 // up the insertion point. 1431 1432 // Try to split the block so it's fully aligned. Compute the latest split 1433 // point where we can add a 4-byte branch instruction, and then align to 1434 // Align which is the largest possible alignment in the function. 1435 const Align Align = MF->getAlignment(); 1436 assert(Align >= CPEAlign && "Over-aligned constant pool entry"); 1437 unsigned KnownBits = UserBBI.internalKnownBits(); 1438 unsigned UPad = UnknownPadding(Align, KnownBits); 1439 unsigned BaseInsertOffset = UserOffset + U.getMaxDisp() - UPad; 1440 LLVM_DEBUG(dbgs() << format("Split in middle of big block before %#x", 1441 BaseInsertOffset)); 1442 1443 // The 4 in the following is for the unconditional branch we'll be inserting 1444 // (allows for long branch on Thumb1). Alignment of the island is handled 1445 // inside isOffsetInRange. 1446 BaseInsertOffset -= 4; 1447 1448 LLVM_DEBUG(dbgs() << format(", adjusted to %#x", BaseInsertOffset) 1449 << " la=" << Log2(Align) << " kb=" << KnownBits 1450 << " up=" << UPad << '\n'); 1451 1452 // This could point off the end of the block if we've already got constant 1453 // pool entries following this block; only the last one is in the water list. 1454 // Back past any possible branches (allow for a conditional and a maximally 1455 // long unconditional). 1456 if (BaseInsertOffset + 8 >= UserBBI.postOffset()) { 1457 // Ensure BaseInsertOffset is larger than the offset of the instruction 1458 // following UserMI so that the loop which searches for the split point 1459 // iterates at least once. 1460 BaseInsertOffset = 1461 std::max(UserBBI.postOffset() - UPad - 8, 1462 UserOffset + TII->getInstSizeInBytes(*UserMI) + 1); 1463 // If the CP is referenced(ie, UserOffset) is in first four instructions 1464 // after IT, this recalculated BaseInsertOffset could be in the middle of 1465 // an IT block. If it is, change the BaseInsertOffset to just after the 1466 // IT block. This still make the CP Entry is in range becuase of the 1467 // following reasons. 1468 // 1. The initial BaseseInsertOffset calculated is (UserOffset + 1469 // U.getMaxDisp() - UPad). 1470 // 2. An IT block is only at most 4 instructions plus the "it" itself (18 1471 // bytes). 1472 // 3. All the relevant instructions support much larger Maximum 1473 // displacement. 1474 MachineBasicBlock::iterator I = UserMI; 1475 ++I; 1476 Register PredReg; 1477 for (unsigned Offset = UserOffset + TII->getInstSizeInBytes(*UserMI); 1478 I->getOpcode() != ARM::t2IT && 1479 getITInstrPredicate(*I, PredReg) != ARMCC::AL; 1480 Offset += TII->getInstSizeInBytes(*I), I = std::next(I)) { 1481 BaseInsertOffset = 1482 std::max(BaseInsertOffset, Offset + TII->getInstSizeInBytes(*I) + 1); 1483 assert(I != UserMBB->end() && "Fell off end of block"); 1484 } 1485 LLVM_DEBUG(dbgs() << format("Move inside block: %#x\n", BaseInsertOffset)); 1486 } 1487 unsigned EndInsertOffset = BaseInsertOffset + 4 + UPad + 1488 CPEMI->getOperand(2).getImm(); 1489 MachineBasicBlock::iterator MI = UserMI; 1490 ++MI; 1491 unsigned CPUIndex = CPUserIndex+1; 1492 unsigned NumCPUsers = CPUsers.size(); 1493 MachineInstr *LastIT = nullptr; 1494 for (unsigned Offset = UserOffset + TII->getInstSizeInBytes(*UserMI); 1495 Offset < BaseInsertOffset; 1496 Offset += TII->getInstSizeInBytes(*MI), MI = std::next(MI)) { 1497 assert(MI != UserMBB->end() && "Fell off end of block"); 1498 if (CPUIndex < NumCPUsers && CPUsers[CPUIndex].MI == &*MI) { 1499 CPUser &U = CPUsers[CPUIndex]; 1500 if (!isOffsetInRange(Offset, EndInsertOffset, U)) { 1501 // Shift intertion point by one unit of alignment so it is within reach. 1502 BaseInsertOffset -= Align.value(); 1503 EndInsertOffset -= Align.value(); 1504 } 1505 // This is overly conservative, as we don't account for CPEMIs being 1506 // reused within the block, but it doesn't matter much. Also assume CPEs 1507 // are added in order with alignment padding. We may eventually be able 1508 // to pack the aligned CPEs better. 1509 EndInsertOffset += U.CPEMI->getOperand(2).getImm(); 1510 CPUIndex++; 1511 } 1512 1513 // Remember the last IT instruction. 1514 if (MI->getOpcode() == ARM::t2IT) 1515 LastIT = &*MI; 1516 } 1517 1518 --MI; 1519 1520 // Avoid splitting an IT block. 1521 if (LastIT) { 1522 Register PredReg; 1523 ARMCC::CondCodes CC = getITInstrPredicate(*MI, PredReg); 1524 if (CC != ARMCC::AL) 1525 MI = LastIT; 1526 } 1527 1528 // Avoid splitting a MOVW+MOVT pair with a relocation on Windows. 1529 // On Windows, this instruction pair is covered by one single 1530 // IMAGE_REL_ARM_MOV32T relocation which covers both instructions. If a 1531 // constant island is injected inbetween them, the relocation will clobber 1532 // the instruction and fail to update the MOVT instruction. 1533 // (These instructions are bundled up until right before the ConstantIslands 1534 // pass.) 1535 if (STI->isTargetWindows() && isThumb && MI->getOpcode() == ARM::t2MOVTi16 && 1536 (MI->getOperand(2).getTargetFlags() & ARMII::MO_OPTION_MASK) == 1537 ARMII::MO_HI16) { 1538 --MI; 1539 assert(MI->getOpcode() == ARM::t2MOVi16 && 1540 (MI->getOperand(1).getTargetFlags() & ARMII::MO_OPTION_MASK) == 1541 ARMII::MO_LO16); 1542 } 1543 1544 // We really must not split an IT block. 1545 #ifndef NDEBUG 1546 Register PredReg; 1547 assert(!isThumb || getITInstrPredicate(*MI, PredReg) == ARMCC::AL); 1548 #endif 1549 NewMBB = splitBlockBeforeInstr(&*MI); 1550 } 1551 1552 /// handleConstantPoolUser - Analyze the specified user, checking to see if it 1553 /// is out-of-range. If so, pick up the constant pool value and move it some 1554 /// place in-range. Return true if we changed any addresses (thus must run 1555 /// another pass of branch lengthening), false otherwise. 1556 bool ARMConstantIslands::handleConstantPoolUser(unsigned CPUserIndex, 1557 bool CloserWater) { 1558 CPUser &U = CPUsers[CPUserIndex]; 1559 MachineInstr *UserMI = U.MI; 1560 MachineInstr *CPEMI = U.CPEMI; 1561 unsigned CPI = getCombinedIndex(CPEMI); 1562 unsigned Size = CPEMI->getOperand(2).getImm(); 1563 // Compute this only once, it's expensive. 1564 unsigned UserOffset = getUserOffset(U); 1565 1566 // See if the current entry is within range, or there is a clone of it 1567 // in range. 1568 int result = findInRangeCPEntry(U, UserOffset); 1569 if (result==1) return false; 1570 else if (result==2) return true; 1571 1572 // No existing clone of this CPE is within range. 1573 // We will be generating a new clone. Get a UID for it. 1574 unsigned ID = AFI->createPICLabelUId(); 1575 1576 // Look for water where we can place this CPE. 1577 MachineBasicBlock *NewIsland = MF->CreateMachineBasicBlock(); 1578 MachineBasicBlock *NewMBB; 1579 water_iterator IP; 1580 if (findAvailableWater(U, UserOffset, IP, CloserWater)) { 1581 LLVM_DEBUG(dbgs() << "Found water in range\n"); 1582 MachineBasicBlock *WaterBB = *IP; 1583 1584 // If the original WaterList entry was "new water" on this iteration, 1585 // propagate that to the new island. This is just keeping NewWaterList 1586 // updated to match the WaterList, which will be updated below. 1587 if (NewWaterList.erase(WaterBB)) 1588 NewWaterList.insert(NewIsland); 1589 1590 // The new CPE goes before the following block (NewMBB). 1591 NewMBB = &*++WaterBB->getIterator(); 1592 } else { 1593 // No water found. 1594 LLVM_DEBUG(dbgs() << "No water found\n"); 1595 createNewWater(CPUserIndex, UserOffset, NewMBB); 1596 1597 // splitBlockBeforeInstr adds to WaterList, which is important when it is 1598 // called while handling branches so that the water will be seen on the 1599 // next iteration for constant pools, but in this context, we don't want 1600 // it. Check for this so it will be removed from the WaterList. 1601 // Also remove any entry from NewWaterList. 1602 MachineBasicBlock *WaterBB = &*--NewMBB->getIterator(); 1603 IP = find(WaterList, WaterBB); 1604 if (IP != WaterList.end()) 1605 NewWaterList.erase(WaterBB); 1606 1607 // We are adding new water. Update NewWaterList. 1608 NewWaterList.insert(NewIsland); 1609 } 1610 // Always align the new block because CP entries can be smaller than 4 1611 // bytes. Be careful not to decrease the existing alignment, e.g. NewMBB may 1612 // be an already aligned constant pool block. 1613 const Align Alignment = isThumb ? Align(2) : Align(4); 1614 if (NewMBB->getAlignment() < Alignment) 1615 NewMBB->setAlignment(Alignment); 1616 1617 // Remove the original WaterList entry; we want subsequent insertions in 1618 // this vicinity to go after the one we're about to insert. This 1619 // considerably reduces the number of times we have to move the same CPE 1620 // more than once and is also important to ensure the algorithm terminates. 1621 if (IP != WaterList.end()) 1622 WaterList.erase(IP); 1623 1624 // Okay, we know we can put an island before NewMBB now, do it! 1625 MF->insert(NewMBB->getIterator(), NewIsland); 1626 1627 // Update internal data structures to account for the newly inserted MBB. 1628 updateForInsertedWaterBlock(NewIsland); 1629 1630 // Now that we have an island to add the CPE to, clone the original CPE and 1631 // add it to the island. 1632 U.HighWaterMark = NewIsland; 1633 U.CPEMI = BuildMI(NewIsland, DebugLoc(), CPEMI->getDesc()) 1634 .addImm(ID) 1635 .add(CPEMI->getOperand(1)) 1636 .addImm(Size); 1637 CPEntries[CPI].push_back(CPEntry(U.CPEMI, ID, 1)); 1638 ++NumCPEs; 1639 1640 // Decrement the old entry, and remove it if refcount becomes 0. 1641 decrementCPEReferenceCount(CPI, CPEMI); 1642 1643 // Mark the basic block as aligned as required by the const-pool entry. 1644 NewIsland->setAlignment(getCPEAlign(U.CPEMI)); 1645 1646 // Increase the size of the island block to account for the new entry. 1647 BBUtils->adjustBBSize(NewIsland, Size); 1648 BBUtils->adjustBBOffsetsAfter(&*--NewIsland->getIterator()); 1649 1650 // Finally, change the CPI in the instruction operand to be ID. 1651 for (MachineOperand &MO : UserMI->operands()) 1652 if (MO.isCPI()) { 1653 MO.setIndex(ID); 1654 break; 1655 } 1656 1657 LLVM_DEBUG( 1658 dbgs() << " Moved CPE to #" << ID << " CPI=" << CPI 1659 << format(" offset=%#x\n", 1660 BBUtils->getBBInfo()[NewIsland->getNumber()].Offset)); 1661 1662 return true; 1663 } 1664 1665 /// removeDeadCPEMI - Remove a dead constant pool entry instruction. Update 1666 /// sizes and offsets of impacted basic blocks. 1667 void ARMConstantIslands::removeDeadCPEMI(MachineInstr *CPEMI) { 1668 MachineBasicBlock *CPEBB = CPEMI->getParent(); 1669 unsigned Size = CPEMI->getOperand(2).getImm(); 1670 CPEMI->eraseFromParent(); 1671 BBInfoVector &BBInfo = BBUtils->getBBInfo(); 1672 BBUtils->adjustBBSize(CPEBB, -Size); 1673 // All succeeding offsets have the current size value added in, fix this. 1674 if (CPEBB->empty()) { 1675 BBInfo[CPEBB->getNumber()].Size = 0; 1676 1677 // This block no longer needs to be aligned. 1678 CPEBB->setAlignment(Align(1)); 1679 } else { 1680 // Entries are sorted by descending alignment, so realign from the front. 1681 CPEBB->setAlignment(getCPEAlign(&*CPEBB->begin())); 1682 } 1683 1684 BBUtils->adjustBBOffsetsAfter(CPEBB); 1685 // An island has only one predecessor BB and one successor BB. Check if 1686 // this BB's predecessor jumps directly to this BB's successor. This 1687 // shouldn't happen currently. 1688 assert(!BBIsJumpedOver(CPEBB) && "How did this happen?"); 1689 // FIXME: remove the empty blocks after all the work is done? 1690 } 1691 1692 /// removeUnusedCPEntries - Remove constant pool entries whose refcounts 1693 /// are zero. 1694 bool ARMConstantIslands::removeUnusedCPEntries() { 1695 unsigned MadeChange = false; 1696 for (std::vector<CPEntry> &CPEs : CPEntries) { 1697 for (CPEntry &CPE : CPEs) { 1698 if (CPE.RefCount == 0 && CPE.CPEMI) { 1699 removeDeadCPEMI(CPE.CPEMI); 1700 CPE.CPEMI = nullptr; 1701 MadeChange = true; 1702 } 1703 } 1704 } 1705 return MadeChange; 1706 } 1707 1708 1709 /// fixupImmediateBr - Fix up an immediate branch whose destination is too far 1710 /// away to fit in its displacement field. 1711 bool ARMConstantIslands::fixupImmediateBr(ImmBranch &Br) { 1712 MachineInstr *MI = Br.MI; 1713 MachineBasicBlock *DestBB = MI->getOperand(0).getMBB(); 1714 1715 // Check to see if the DestBB is already in-range. 1716 if (BBUtils->isBBInRange(MI, DestBB, Br.MaxDisp)) 1717 return false; 1718 1719 if (!Br.isCond) 1720 return fixupUnconditionalBr(Br); 1721 return fixupConditionalBr(Br); 1722 } 1723 1724 /// fixupUnconditionalBr - Fix up an unconditional branch whose destination is 1725 /// too far away to fit in its displacement field. If the LR register has been 1726 /// spilled in the epilogue, then we can use BL to implement a far jump. 1727 /// Otherwise, add an intermediate branch instruction to a branch. 1728 bool 1729 ARMConstantIslands::fixupUnconditionalBr(ImmBranch &Br) { 1730 MachineInstr *MI = Br.MI; 1731 MachineBasicBlock *MBB = MI->getParent(); 1732 if (!isThumb1) 1733 llvm_unreachable("fixupUnconditionalBr is Thumb1 only!"); 1734 1735 if (!AFI->isLRSpilled()) 1736 report_fatal_error("underestimated function size"); 1737 1738 // Use BL to implement far jump. 1739 Br.MaxDisp = (1 << 21) * 2; 1740 MI->setDesc(TII->get(ARM::tBfar)); 1741 BBInfoVector &BBInfo = BBUtils->getBBInfo(); 1742 BBInfo[MBB->getNumber()].Size += 2; 1743 BBUtils->adjustBBOffsetsAfter(MBB); 1744 ++NumUBrFixed; 1745 1746 LLVM_DEBUG(dbgs() << " Changed B to long jump " << *MI); 1747 1748 return true; 1749 } 1750 1751 /// fixupConditionalBr - Fix up a conditional branch whose destination is too 1752 /// far away to fit in its displacement field. It is converted to an inverse 1753 /// conditional branch + an unconditional branch to the destination. 1754 bool 1755 ARMConstantIslands::fixupConditionalBr(ImmBranch &Br) { 1756 MachineInstr *MI = Br.MI; 1757 MachineBasicBlock *DestBB = MI->getOperand(0).getMBB(); 1758 1759 // Add an unconditional branch to the destination and invert the branch 1760 // condition to jump over it: 1761 // blt L1 1762 // => 1763 // bge L2 1764 // b L1 1765 // L2: 1766 ARMCC::CondCodes CC = (ARMCC::CondCodes)MI->getOperand(1).getImm(); 1767 CC = ARMCC::getOppositeCondition(CC); 1768 Register CCReg = MI->getOperand(2).getReg(); 1769 1770 // If the branch is at the end of its MBB and that has a fall-through block, 1771 // direct the updated conditional branch to the fall-through block. Otherwise, 1772 // split the MBB before the next instruction. 1773 MachineBasicBlock *MBB = MI->getParent(); 1774 MachineInstr *BMI = &MBB->back(); 1775 bool NeedSplit = (BMI != MI) || !BBHasFallthrough(MBB); 1776 1777 ++NumCBrFixed; 1778 if (BMI != MI) { 1779 if (std::next(MachineBasicBlock::iterator(MI)) == std::prev(MBB->end()) && 1780 BMI->getOpcode() == Br.UncondBr) { 1781 // Last MI in the BB is an unconditional branch. Can we simply invert the 1782 // condition and swap destinations: 1783 // beq L1 1784 // b L2 1785 // => 1786 // bne L2 1787 // b L1 1788 MachineBasicBlock *NewDest = BMI->getOperand(0).getMBB(); 1789 if (BBUtils->isBBInRange(MI, NewDest, Br.MaxDisp)) { 1790 LLVM_DEBUG( 1791 dbgs() << " Invert Bcc condition and swap its destination with " 1792 << *BMI); 1793 BMI->getOperand(0).setMBB(DestBB); 1794 MI->getOperand(0).setMBB(NewDest); 1795 MI->getOperand(1).setImm(CC); 1796 return true; 1797 } 1798 } 1799 } 1800 1801 if (NeedSplit) { 1802 splitBlockBeforeInstr(MI); 1803 // No need for the branch to the next block. We're adding an unconditional 1804 // branch to the destination. 1805 int delta = TII->getInstSizeInBytes(MBB->back()); 1806 BBUtils->adjustBBSize(MBB, -delta); 1807 MBB->back().eraseFromParent(); 1808 1809 // The conditional successor will be swapped between the BBs after this, so 1810 // update CFG. 1811 MBB->addSuccessor(DestBB); 1812 std::next(MBB->getIterator())->removeSuccessor(DestBB); 1813 1814 // BBInfo[SplitBB].Offset is wrong temporarily, fixed below 1815 } 1816 MachineBasicBlock *NextBB = &*++MBB->getIterator(); 1817 1818 LLVM_DEBUG(dbgs() << " Insert B to " << printMBBReference(*DestBB) 1819 << " also invert condition and change dest. to " 1820 << printMBBReference(*NextBB) << "\n"); 1821 1822 // Insert a new conditional branch and a new unconditional branch. 1823 // Also update the ImmBranch as well as adding a new entry for the new branch. 1824 BuildMI(MBB, DebugLoc(), TII->get(MI->getOpcode())) 1825 .addMBB(NextBB).addImm(CC).addReg(CCReg); 1826 Br.MI = &MBB->back(); 1827 BBUtils->adjustBBSize(MBB, TII->getInstSizeInBytes(MBB->back())); 1828 if (isThumb) 1829 BuildMI(MBB, DebugLoc(), TII->get(Br.UncondBr)) 1830 .addMBB(DestBB) 1831 .add(predOps(ARMCC::AL)); 1832 else 1833 BuildMI(MBB, DebugLoc(), TII->get(Br.UncondBr)).addMBB(DestBB); 1834 BBUtils->adjustBBSize(MBB, TII->getInstSizeInBytes(MBB->back())); 1835 unsigned MaxDisp = getUnconditionalBrDisp(Br.UncondBr); 1836 ImmBranches.push_back(ImmBranch(&MBB->back(), MaxDisp, false, Br.UncondBr)); 1837 1838 // Remove the old conditional branch. It may or may not still be in MBB. 1839 BBUtils->adjustBBSize(MI->getParent(), -TII->getInstSizeInBytes(*MI)); 1840 MI->eraseFromParent(); 1841 BBUtils->adjustBBOffsetsAfter(MBB); 1842 return true; 1843 } 1844 1845 bool ARMConstantIslands::optimizeThumb2Instructions() { 1846 bool MadeChange = false; 1847 1848 // Shrink ADR and LDR from constantpool. 1849 for (CPUser &U : CPUsers) { 1850 unsigned Opcode = U.MI->getOpcode(); 1851 unsigned NewOpc = 0; 1852 unsigned Scale = 1; 1853 unsigned Bits = 0; 1854 switch (Opcode) { 1855 default: break; 1856 case ARM::t2LEApcrel: 1857 if (isARMLowRegister(U.MI->getOperand(0).getReg())) { 1858 NewOpc = ARM::tLEApcrel; 1859 Bits = 8; 1860 Scale = 4; 1861 } 1862 break; 1863 case ARM::t2LDRpci: 1864 if (isARMLowRegister(U.MI->getOperand(0).getReg())) { 1865 NewOpc = ARM::tLDRpci; 1866 Bits = 8; 1867 Scale = 4; 1868 } 1869 break; 1870 } 1871 1872 if (!NewOpc) 1873 continue; 1874 1875 unsigned UserOffset = getUserOffset(U); 1876 unsigned MaxOffs = ((1 << Bits) - 1) * Scale; 1877 1878 // Be conservative with inline asm. 1879 if (!U.KnownAlignment) 1880 MaxOffs -= 2; 1881 1882 // FIXME: Check if offset is multiple of scale if scale is not 4. 1883 if (isCPEntryInRange(U.MI, UserOffset, U.CPEMI, MaxOffs, false, true)) { 1884 LLVM_DEBUG(dbgs() << "Shrink: " << *U.MI); 1885 U.MI->setDesc(TII->get(NewOpc)); 1886 MachineBasicBlock *MBB = U.MI->getParent(); 1887 BBUtils->adjustBBSize(MBB, -2); 1888 BBUtils->adjustBBOffsetsAfter(MBB); 1889 ++NumT2CPShrunk; 1890 MadeChange = true; 1891 } 1892 } 1893 1894 return MadeChange; 1895 } 1896 1897 1898 bool ARMConstantIslands::optimizeThumb2Branches() { 1899 1900 auto TryShrinkBranch = [this](ImmBranch &Br) { 1901 unsigned Opcode = Br.MI->getOpcode(); 1902 unsigned NewOpc = 0; 1903 unsigned Scale = 1; 1904 unsigned Bits = 0; 1905 switch (Opcode) { 1906 default: break; 1907 case ARM::t2B: 1908 NewOpc = ARM::tB; 1909 Bits = 11; 1910 Scale = 2; 1911 break; 1912 case ARM::t2Bcc: 1913 NewOpc = ARM::tBcc; 1914 Bits = 8; 1915 Scale = 2; 1916 break; 1917 } 1918 if (NewOpc) { 1919 unsigned MaxOffs = ((1 << (Bits-1))-1) * Scale; 1920 MachineBasicBlock *DestBB = Br.MI->getOperand(0).getMBB(); 1921 if (BBUtils->isBBInRange(Br.MI, DestBB, MaxOffs)) { 1922 LLVM_DEBUG(dbgs() << "Shrink branch: " << *Br.MI); 1923 Br.MI->setDesc(TII->get(NewOpc)); 1924 MachineBasicBlock *MBB = Br.MI->getParent(); 1925 BBUtils->adjustBBSize(MBB, -2); 1926 BBUtils->adjustBBOffsetsAfter(MBB); 1927 ++NumT2BrShrunk; 1928 return true; 1929 } 1930 } 1931 return false; 1932 }; 1933 1934 struct ImmCompare { 1935 MachineInstr* MI = nullptr; 1936 unsigned NewOpc = 0; 1937 }; 1938 1939 auto FindCmpForCBZ = [this](ImmBranch &Br, ImmCompare &ImmCmp, 1940 MachineBasicBlock *DestBB) { 1941 ImmCmp.MI = nullptr; 1942 ImmCmp.NewOpc = 0; 1943 1944 // If the conditional branch doesn't kill CPSR, then CPSR can be liveout 1945 // so this transformation is not safe. 1946 if (!Br.MI->killsRegister(ARM::CPSR, /*TRI=*/nullptr)) 1947 return false; 1948 1949 Register PredReg; 1950 unsigned NewOpc = 0; 1951 ARMCC::CondCodes Pred = getInstrPredicate(*Br.MI, PredReg); 1952 if (Pred == ARMCC::EQ) 1953 NewOpc = ARM::tCBZ; 1954 else if (Pred == ARMCC::NE) 1955 NewOpc = ARM::tCBNZ; 1956 else 1957 return false; 1958 1959 // Check if the distance is within 126. Subtract starting offset by 2 1960 // because the cmp will be eliminated. 1961 unsigned BrOffset = BBUtils->getOffsetOf(Br.MI) + 4 - 2; 1962 BBInfoVector &BBInfo = BBUtils->getBBInfo(); 1963 unsigned DestOffset = BBInfo[DestBB->getNumber()].Offset; 1964 if (BrOffset >= DestOffset || (DestOffset - BrOffset) > 126) 1965 return false; 1966 1967 // Search backwards to find a tCMPi8 1968 auto *TRI = STI->getRegisterInfo(); 1969 MachineInstr *CmpMI = findCMPToFoldIntoCBZ(Br.MI, TRI); 1970 if (!CmpMI || CmpMI->getOpcode() != ARM::tCMPi8) 1971 return false; 1972 1973 ImmCmp.MI = CmpMI; 1974 ImmCmp.NewOpc = NewOpc; 1975 return true; 1976 }; 1977 1978 auto TryConvertToLE = [this](ImmBranch &Br, ImmCompare &Cmp) { 1979 if (Br.MI->getOpcode() != ARM::t2Bcc || !STI->hasLOB() || 1980 STI->hasMinSize()) 1981 return false; 1982 1983 MachineBasicBlock *MBB = Br.MI->getParent(); 1984 MachineBasicBlock *DestBB = Br.MI->getOperand(0).getMBB(); 1985 if (BBUtils->getOffsetOf(MBB) < BBUtils->getOffsetOf(DestBB) || 1986 !BBUtils->isBBInRange(Br.MI, DestBB, 4094)) 1987 return false; 1988 1989 if (!DT->dominates(DestBB, MBB)) 1990 return false; 1991 1992 // We queried for the CBN?Z opcode based upon the 'ExitBB', the opposite 1993 // target of Br. So now we need to reverse the condition. 1994 Cmp.NewOpc = Cmp.NewOpc == ARM::tCBZ ? ARM::tCBNZ : ARM::tCBZ; 1995 1996 MachineInstrBuilder MIB = BuildMI(*MBB, Br.MI, Br.MI->getDebugLoc(), 1997 TII->get(ARM::t2LE)); 1998 // Swapped a t2Bcc for a t2LE, so no need to update the size of the block. 1999 MIB.add(Br.MI->getOperand(0)); 2000 Br.MI->eraseFromParent(); 2001 Br.MI = MIB; 2002 ++NumLEInserted; 2003 return true; 2004 }; 2005 2006 bool MadeChange = false; 2007 2008 // The order in which branches appear in ImmBranches is approximately their 2009 // order within the function body. By visiting later branches first, we reduce 2010 // the distance between earlier forward branches and their targets, making it 2011 // more likely that the cbn?z optimization, which can only apply to forward 2012 // branches, will succeed. 2013 for (ImmBranch &Br : reverse(ImmBranches)) { 2014 MachineBasicBlock *DestBB = Br.MI->getOperand(0).getMBB(); 2015 MachineBasicBlock *MBB = Br.MI->getParent(); 2016 MachineBasicBlock *ExitBB = &MBB->back() == Br.MI ? 2017 MBB->getFallThrough() : 2018 MBB->back().getOperand(0).getMBB(); 2019 2020 ImmCompare Cmp; 2021 if (FindCmpForCBZ(Br, Cmp, ExitBB) && TryConvertToLE(Br, Cmp)) { 2022 DestBB = ExitBB; 2023 MadeChange = true; 2024 } else { 2025 FindCmpForCBZ(Br, Cmp, DestBB); 2026 MadeChange |= TryShrinkBranch(Br); 2027 } 2028 2029 unsigned Opcode = Br.MI->getOpcode(); 2030 if ((Opcode != ARM::tBcc && Opcode != ARM::t2LE) || !Cmp.NewOpc) 2031 continue; 2032 2033 Register Reg = Cmp.MI->getOperand(0).getReg(); 2034 2035 // Check for Kill flags on Reg. If they are present remove them and set kill 2036 // on the new CBZ. 2037 auto *TRI = STI->getRegisterInfo(); 2038 MachineBasicBlock::iterator KillMI = Br.MI; 2039 bool RegKilled = false; 2040 do { 2041 --KillMI; 2042 if (KillMI->killsRegister(Reg, TRI)) { 2043 KillMI->clearRegisterKills(Reg, TRI); 2044 RegKilled = true; 2045 break; 2046 } 2047 } while (KillMI != Cmp.MI); 2048 2049 // Create the new CBZ/CBNZ 2050 LLVM_DEBUG(dbgs() << "Fold: " << *Cmp.MI << " and: " << *Br.MI); 2051 MachineInstr *NewBR = 2052 BuildMI(*MBB, Br.MI, Br.MI->getDebugLoc(), TII->get(Cmp.NewOpc)) 2053 .addReg(Reg, getKillRegState(RegKilled) | 2054 getRegState(Cmp.MI->getOperand(0))) 2055 .addMBB(DestBB, Br.MI->getOperand(0).getTargetFlags()); 2056 2057 Cmp.MI->eraseFromParent(); 2058 2059 if (Br.MI->getOpcode() == ARM::tBcc) { 2060 Br.MI->eraseFromParent(); 2061 Br.MI = NewBR; 2062 BBUtils->adjustBBSize(MBB, -2); 2063 } else if (MBB->back().getOpcode() != ARM::t2LE) { 2064 // An LE has been generated, but it's not the terminator - that is an 2065 // unconditional branch. However, the logic has now been reversed with the 2066 // CBN?Z being the conditional branch and the LE being the unconditional 2067 // branch. So this means we can remove the redundant unconditional branch 2068 // at the end of the block. 2069 MachineInstr *LastMI = &MBB->back(); 2070 BBUtils->adjustBBSize(MBB, -LastMI->getDesc().getSize()); 2071 LastMI->eraseFromParent(); 2072 } 2073 BBUtils->adjustBBOffsetsAfter(MBB); 2074 ++NumCBZ; 2075 MadeChange = true; 2076 } 2077 2078 return MadeChange; 2079 } 2080 2081 static bool isSimpleIndexCalc(MachineInstr &I, unsigned EntryReg, 2082 unsigned BaseReg) { 2083 if (I.getOpcode() != ARM::t2ADDrs) 2084 return false; 2085 2086 if (I.getOperand(0).getReg() != EntryReg) 2087 return false; 2088 2089 if (I.getOperand(1).getReg() != BaseReg) 2090 return false; 2091 2092 // FIXME: what about CC and IdxReg? 2093 return true; 2094 } 2095 2096 /// While trying to form a TBB/TBH instruction, we may (if the table 2097 /// doesn't immediately follow the BR_JT) need access to the start of the 2098 /// jump-table. We know one instruction that produces such a register; this 2099 /// function works out whether that definition can be preserved to the BR_JT, 2100 /// possibly by removing an intervening addition (which is usually needed to 2101 /// calculate the actual entry to jump to). 2102 bool ARMConstantIslands::preserveBaseRegister(MachineInstr *JumpMI, 2103 MachineInstr *LEAMI, 2104 unsigned &DeadSize, 2105 bool &CanDeleteLEA, 2106 bool &BaseRegKill) { 2107 if (JumpMI->getParent() != LEAMI->getParent()) 2108 return false; 2109 2110 // Now we hope that we have at least these instructions in the basic block: 2111 // BaseReg = t2LEA ... 2112 // [...] 2113 // EntryReg = t2ADDrs BaseReg, ... 2114 // [...] 2115 // t2BR_JT EntryReg 2116 // 2117 // We have to be very conservative about what we recognise here though. The 2118 // main perturbing factors to watch out for are: 2119 // + Spills at any point in the chain: not direct problems but we would 2120 // expect a blocking Def of the spilled register so in practice what we 2121 // can do is limited. 2122 // + EntryReg == BaseReg: this is the one situation we should allow a Def 2123 // of BaseReg, but only if the t2ADDrs can be removed. 2124 // + Some instruction other than t2ADDrs computing the entry. Not seen in 2125 // the wild, but we should be careful. 2126 Register EntryReg = JumpMI->getOperand(0).getReg(); 2127 Register BaseReg = LEAMI->getOperand(0).getReg(); 2128 2129 CanDeleteLEA = true; 2130 BaseRegKill = false; 2131 MachineInstr *RemovableAdd = nullptr; 2132 MachineBasicBlock::iterator I(LEAMI); 2133 for (++I; &*I != JumpMI; ++I) { 2134 if (isSimpleIndexCalc(*I, EntryReg, BaseReg)) { 2135 RemovableAdd = &*I; 2136 break; 2137 } 2138 2139 for (const MachineOperand &MO : I->operands()) { 2140 if (!MO.isReg() || !MO.getReg()) 2141 continue; 2142 if (MO.isDef() && MO.getReg() == BaseReg) 2143 return false; 2144 if (MO.isUse() && MO.getReg() == BaseReg) { 2145 BaseRegKill = BaseRegKill || MO.isKill(); 2146 CanDeleteLEA = false; 2147 } 2148 } 2149 } 2150 2151 if (!RemovableAdd) 2152 return true; 2153 2154 // Check the add really is removable, and that nothing else in the block 2155 // clobbers BaseReg. 2156 for (++I; &*I != JumpMI; ++I) { 2157 for (const MachineOperand &MO : I->operands()) { 2158 if (!MO.isReg() || !MO.getReg()) 2159 continue; 2160 if (MO.isDef() && MO.getReg() == BaseReg) 2161 return false; 2162 if (MO.isUse() && MO.getReg() == EntryReg) 2163 RemovableAdd = nullptr; 2164 } 2165 } 2166 2167 if (RemovableAdd) { 2168 RemovableAdd->eraseFromParent(); 2169 DeadSize += isThumb2 ? 4 : 2; 2170 } else if (BaseReg == EntryReg) { 2171 // The add wasn't removable, but clobbered the base for the TBB. So we can't 2172 // preserve it. 2173 return false; 2174 } 2175 2176 // We reached the end of the block without seeing another definition of 2177 // BaseReg (except, possibly the t2ADDrs, which was removed). BaseReg can be 2178 // used in the TBB/TBH if necessary. 2179 return true; 2180 } 2181 2182 /// Returns whether CPEMI is the first instruction in the block 2183 /// immediately following JTMI (assumed to be a TBB or TBH terminator). If so, 2184 /// we can switch the first register to PC and usually remove the address 2185 /// calculation that preceded it. 2186 static bool jumpTableFollowsTB(MachineInstr *JTMI, MachineInstr *CPEMI) { 2187 MachineFunction::iterator MBB = JTMI->getParent()->getIterator(); 2188 MachineFunction *MF = MBB->getParent(); 2189 ++MBB; 2190 2191 return MBB != MF->end() && !MBB->empty() && &*MBB->begin() == CPEMI; 2192 } 2193 2194 static void RemoveDeadAddBetweenLEAAndJT(MachineInstr *LEAMI, 2195 MachineInstr *JumpMI, 2196 unsigned &DeadSize) { 2197 // Remove a dead add between the LEA and JT, which used to compute EntryReg, 2198 // but the JT now uses PC. Finds the last ADD (if any) that def's EntryReg 2199 // and is not clobbered / used. 2200 MachineInstr *RemovableAdd = nullptr; 2201 Register EntryReg = JumpMI->getOperand(0).getReg(); 2202 2203 // Find the last ADD to set EntryReg 2204 MachineBasicBlock::iterator I(LEAMI); 2205 for (++I; &*I != JumpMI; ++I) { 2206 if (I->getOpcode() == ARM::t2ADDrs && I->getOperand(0).getReg() == EntryReg) 2207 RemovableAdd = &*I; 2208 } 2209 2210 if (!RemovableAdd) 2211 return; 2212 2213 // Ensure EntryReg is not clobbered or used. 2214 MachineBasicBlock::iterator J(RemovableAdd); 2215 for (++J; &*J != JumpMI; ++J) { 2216 for (const MachineOperand &MO : J->operands()) { 2217 if (!MO.isReg() || !MO.getReg()) 2218 continue; 2219 if (MO.isDef() && MO.getReg() == EntryReg) 2220 return; 2221 if (MO.isUse() && MO.getReg() == EntryReg) 2222 return; 2223 } 2224 } 2225 2226 LLVM_DEBUG(dbgs() << "Removing Dead Add: " << *RemovableAdd); 2227 RemovableAdd->eraseFromParent(); 2228 DeadSize += 4; 2229 } 2230 2231 /// optimizeThumb2JumpTables - Use tbb / tbh instructions to generate smaller 2232 /// jumptables when it's possible. 2233 bool ARMConstantIslands::optimizeThumb2JumpTables() { 2234 bool MadeChange = false; 2235 2236 // FIXME: After the tables are shrunk, can we get rid some of the 2237 // constantpool tables? 2238 MachineJumpTableInfo *MJTI = MF->getJumpTableInfo(); 2239 if (!MJTI) return false; 2240 2241 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables(); 2242 for (MachineInstr *MI : T2JumpTables) { 2243 const MCInstrDesc &MCID = MI->getDesc(); 2244 unsigned NumOps = MCID.getNumOperands(); 2245 unsigned JTOpIdx = NumOps - (MI->isPredicable() ? 2 : 1); 2246 MachineOperand JTOP = MI->getOperand(JTOpIdx); 2247 unsigned JTI = JTOP.getIndex(); 2248 assert(JTI < JT.size()); 2249 2250 bool ByteOk = true; 2251 bool HalfWordOk = true; 2252 unsigned JTOffset = BBUtils->getOffsetOf(MI) + 4; 2253 const std::vector<MachineBasicBlock*> &JTBBs = JT[JTI].MBBs; 2254 BBInfoVector &BBInfo = BBUtils->getBBInfo(); 2255 for (MachineBasicBlock *MBB : JTBBs) { 2256 unsigned DstOffset = BBInfo[MBB->getNumber()].Offset; 2257 // Negative offset is not ok. FIXME: We should change BB layout to make 2258 // sure all the branches are forward. 2259 if (ByteOk && (DstOffset - JTOffset) > ((1<<8)-1)*2) 2260 ByteOk = false; 2261 unsigned TBHLimit = ((1<<16)-1)*2; 2262 if (HalfWordOk && (DstOffset - JTOffset) > TBHLimit) 2263 HalfWordOk = false; 2264 if (!ByteOk && !HalfWordOk) 2265 break; 2266 } 2267 2268 if (!ByteOk && !HalfWordOk) 2269 continue; 2270 2271 CPUser &User = CPUsers[JumpTableUserIndices[JTI]]; 2272 MachineBasicBlock *MBB = MI->getParent(); 2273 if (!MI->getOperand(0).isKill()) // FIXME: needed now? 2274 continue; 2275 2276 unsigned DeadSize = 0; 2277 bool CanDeleteLEA = false; 2278 bool BaseRegKill = false; 2279 2280 unsigned IdxReg = ~0U; 2281 bool IdxRegKill = true; 2282 if (isThumb2) { 2283 IdxReg = MI->getOperand(1).getReg(); 2284 IdxRegKill = MI->getOperand(1).isKill(); 2285 2286 bool PreservedBaseReg = 2287 preserveBaseRegister(MI, User.MI, DeadSize, CanDeleteLEA, BaseRegKill); 2288 if (!jumpTableFollowsTB(MI, User.CPEMI) && !PreservedBaseReg) 2289 continue; 2290 } else { 2291 // We're in thumb-1 mode, so we must have something like: 2292 // %idx = tLSLri %idx, 2 2293 // %base = tLEApcrelJT 2294 // %t = tLDRr %base, %idx 2295 Register BaseReg = User.MI->getOperand(0).getReg(); 2296 2297 MachineBasicBlock *UserMBB = User.MI->getParent(); 2298 MachineBasicBlock::iterator Shift = User.MI->getIterator(); 2299 if (Shift == UserMBB->begin()) 2300 continue; 2301 2302 Shift = prev_nodbg(Shift, UserMBB->begin()); 2303 if (Shift->getOpcode() != ARM::tLSLri || 2304 Shift->getOperand(3).getImm() != 2 || 2305 !Shift->getOperand(2).isKill()) 2306 continue; 2307 IdxReg = Shift->getOperand(2).getReg(); 2308 Register ShiftedIdxReg = Shift->getOperand(0).getReg(); 2309 2310 // It's important that IdxReg is live until the actual TBB/TBH. Most of 2311 // the range is checked later, but the LEA might still clobber it and not 2312 // actually get removed. 2313 if (BaseReg == IdxReg && !jumpTableFollowsTB(MI, User.CPEMI)) 2314 continue; 2315 2316 MachineInstr *Load = User.MI->getNextNode(); 2317 if (Load->getOpcode() != ARM::tLDRr) 2318 continue; 2319 if (Load->getOperand(1).getReg() != BaseReg || 2320 Load->getOperand(2).getReg() != ShiftedIdxReg || 2321 !Load->getOperand(2).isKill()) 2322 continue; 2323 2324 // If we're in PIC mode, there should be another ADD following. 2325 auto *TRI = STI->getRegisterInfo(); 2326 2327 // %base cannot be redefined after the load as it will appear before 2328 // TBB/TBH like: 2329 // %base = 2330 // %base = 2331 // tBB %base, %idx 2332 if (registerDefinedBetween(BaseReg, Load->getNextNode(), MBB->end(), TRI)) 2333 continue; 2334 2335 if (isPositionIndependentOrROPI) { 2336 MachineInstr *Add = Load->getNextNode(); 2337 if (Add->getOpcode() != ARM::tADDrr || 2338 Add->getOperand(2).getReg() != BaseReg || 2339 Add->getOperand(3).getReg() != Load->getOperand(0).getReg() || 2340 !Add->getOperand(3).isKill()) 2341 continue; 2342 if (Add->getOperand(0).getReg() != MI->getOperand(0).getReg()) 2343 continue; 2344 if (registerDefinedBetween(IdxReg, Add->getNextNode(), MI, TRI)) 2345 // IdxReg gets redefined in the middle of the sequence. 2346 continue; 2347 Add->eraseFromParent(); 2348 DeadSize += 2; 2349 } else { 2350 if (Load->getOperand(0).getReg() != MI->getOperand(0).getReg()) 2351 continue; 2352 if (registerDefinedBetween(IdxReg, Load->getNextNode(), MI, TRI)) 2353 // IdxReg gets redefined in the middle of the sequence. 2354 continue; 2355 } 2356 2357 // Now safe to delete the load and lsl. The LEA will be removed later. 2358 CanDeleteLEA = true; 2359 Shift->eraseFromParent(); 2360 Load->eraseFromParent(); 2361 DeadSize += 4; 2362 } 2363 2364 LLVM_DEBUG(dbgs() << "Shrink JT: " << *MI); 2365 MachineInstr *CPEMI = User.CPEMI; 2366 unsigned Opc = ByteOk ? ARM::t2TBB_JT : ARM::t2TBH_JT; 2367 if (!isThumb2) 2368 Opc = ByteOk ? ARM::tTBB_JT : ARM::tTBH_JT; 2369 2370 MachineBasicBlock::iterator MI_JT = MI; 2371 MachineInstr *NewJTMI = 2372 BuildMI(*MBB, MI_JT, MI->getDebugLoc(), TII->get(Opc)) 2373 .addReg(User.MI->getOperand(0).getReg(), 2374 getKillRegState(BaseRegKill)) 2375 .addReg(IdxReg, getKillRegState(IdxRegKill)) 2376 .addJumpTableIndex(JTI, JTOP.getTargetFlags()) 2377 .addImm(CPEMI->getOperand(0).getImm()); 2378 LLVM_DEBUG(dbgs() << printMBBReference(*MBB) << ": " << *NewJTMI); 2379 2380 unsigned JTOpc = ByteOk ? ARM::JUMPTABLE_TBB : ARM::JUMPTABLE_TBH; 2381 CPEMI->setDesc(TII->get(JTOpc)); 2382 2383 if (jumpTableFollowsTB(MI, User.CPEMI)) { 2384 NewJTMI->getOperand(0).setReg(ARM::PC); 2385 NewJTMI->getOperand(0).setIsKill(false); 2386 2387 if (CanDeleteLEA) { 2388 if (isThumb2) 2389 RemoveDeadAddBetweenLEAAndJT(User.MI, MI, DeadSize); 2390 2391 User.MI->eraseFromParent(); 2392 DeadSize += isThumb2 ? 4 : 2; 2393 2394 // The LEA was eliminated, the TBB instruction becomes the only new user 2395 // of the jump table. 2396 User.MI = NewJTMI; 2397 User.MaxDisp = 4; 2398 User.NegOk = false; 2399 User.IsSoImm = false; 2400 User.KnownAlignment = false; 2401 } else { 2402 // The LEA couldn't be eliminated, so we must add another CPUser to 2403 // record the TBB or TBH use. 2404 int CPEntryIdx = JumpTableEntryIndices[JTI]; 2405 auto &CPEs = CPEntries[CPEntryIdx]; 2406 auto Entry = 2407 find_if(CPEs, [&](CPEntry &E) { return E.CPEMI == User.CPEMI; }); 2408 ++Entry->RefCount; 2409 CPUsers.emplace_back(CPUser(NewJTMI, User.CPEMI, 4, false, false)); 2410 } 2411 } 2412 2413 unsigned NewSize = TII->getInstSizeInBytes(*NewJTMI); 2414 unsigned OrigSize = TII->getInstSizeInBytes(*MI); 2415 MI->eraseFromParent(); 2416 2417 int Delta = OrigSize - NewSize + DeadSize; 2418 BBInfo[MBB->getNumber()].Size -= Delta; 2419 BBUtils->adjustBBOffsetsAfter(MBB); 2420 2421 ++NumTBs; 2422 MadeChange = true; 2423 } 2424 2425 return MadeChange; 2426 } 2427 2428 /// reorderThumb2JumpTables - Adjust the function's block layout to ensure that 2429 /// jump tables always branch forwards, since that's what tbb and tbh need. 2430 bool ARMConstantIslands::reorderThumb2JumpTables() { 2431 bool MadeChange = false; 2432 2433 MachineJumpTableInfo *MJTI = MF->getJumpTableInfo(); 2434 if (!MJTI) return false; 2435 2436 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables(); 2437 for (MachineInstr *MI : T2JumpTables) { 2438 const MCInstrDesc &MCID = MI->getDesc(); 2439 unsigned NumOps = MCID.getNumOperands(); 2440 unsigned JTOpIdx = NumOps - (MI->isPredicable() ? 2 : 1); 2441 MachineOperand JTOP = MI->getOperand(JTOpIdx); 2442 unsigned JTI = JTOP.getIndex(); 2443 assert(JTI < JT.size()); 2444 2445 // We prefer if target blocks for the jump table come after the jump 2446 // instruction so we can use TB[BH]. Loop through the target blocks 2447 // and try to adjust them such that that's true. 2448 int JTNumber = MI->getParent()->getNumber(); 2449 const std::vector<MachineBasicBlock*> &JTBBs = JT[JTI].MBBs; 2450 for (MachineBasicBlock *MBB : JTBBs) { 2451 int DTNumber = MBB->getNumber(); 2452 2453 if (DTNumber < JTNumber) { 2454 // The destination precedes the switch. Try to move the block forward 2455 // so we have a positive offset. 2456 MachineBasicBlock *NewBB = 2457 adjustJTTargetBlockForward(JTI, MBB, MI->getParent()); 2458 if (NewBB) 2459 MJTI->ReplaceMBBInJumpTable(JTI, MBB, NewBB); 2460 MadeChange = true; 2461 } 2462 } 2463 } 2464 2465 return MadeChange; 2466 } 2467 2468 MachineBasicBlock *ARMConstantIslands::adjustJTTargetBlockForward( 2469 unsigned JTI, MachineBasicBlock *BB, MachineBasicBlock *JTBB) { 2470 // If the destination block is terminated by an unconditional branch, 2471 // try to move it; otherwise, create a new block following the jump 2472 // table that branches back to the actual target. This is a very simple 2473 // heuristic. FIXME: We can definitely improve it. 2474 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; 2475 SmallVector<MachineOperand, 4> Cond; 2476 SmallVector<MachineOperand, 4> CondPrior; 2477 MachineFunction::iterator BBi = BB->getIterator(); 2478 MachineFunction::iterator OldPrior = std::prev(BBi); 2479 MachineFunction::iterator OldNext = std::next(BBi); 2480 2481 // If the block terminator isn't analyzable, don't try to move the block 2482 bool B = TII->analyzeBranch(*BB, TBB, FBB, Cond); 2483 2484 // If the block ends in an unconditional branch, move it. The prior block 2485 // has to have an analyzable terminator for us to move this one. Be paranoid 2486 // and make sure we're not trying to move the entry block of the function. 2487 if (!B && Cond.empty() && BB != &MF->front() && 2488 !TII->analyzeBranch(*OldPrior, TBB, FBB, CondPrior)) { 2489 BB->moveAfter(JTBB); 2490 OldPrior->updateTerminator(BB); 2491 BB->updateTerminator(OldNext != MF->end() ? &*OldNext : nullptr); 2492 // Update numbering to account for the block being moved. 2493 MF->RenumberBlocks(); 2494 DT->updateBlockNumbers(); 2495 ++NumJTMoved; 2496 return nullptr; 2497 } 2498 2499 // Create a new MBB for the code after the jump BB. 2500 MachineBasicBlock *NewBB = 2501 MF->CreateMachineBasicBlock(JTBB->getBasicBlock()); 2502 MachineFunction::iterator MBBI = ++JTBB->getIterator(); 2503 MF->insert(MBBI, NewBB); 2504 2505 // Copy live-in information to new block. 2506 for (const MachineBasicBlock::RegisterMaskPair &RegMaskPair : BB->liveins()) 2507 NewBB->addLiveIn(RegMaskPair); 2508 2509 // Add an unconditional branch from NewBB to BB. 2510 // There doesn't seem to be meaningful DebugInfo available; this doesn't 2511 // correspond directly to anything in the source. 2512 if (isThumb2) 2513 BuildMI(NewBB, DebugLoc(), TII->get(ARM::t2B)) 2514 .addMBB(BB) 2515 .add(predOps(ARMCC::AL)); 2516 else 2517 BuildMI(NewBB, DebugLoc(), TII->get(ARM::tB)) 2518 .addMBB(BB) 2519 .add(predOps(ARMCC::AL)); 2520 2521 // Update internal data structures to account for the newly inserted MBB. 2522 MF->RenumberBlocks(NewBB); 2523 DT->updateBlockNumbers(); 2524 2525 // Update the CFG. 2526 NewBB->addSuccessor(BB); 2527 JTBB->replaceSuccessor(BB, NewBB); 2528 2529 ++NumJTInserted; 2530 return NewBB; 2531 } 2532 2533 /// createARMConstantIslandPass - returns an instance of the constpool 2534 /// island pass. 2535 FunctionPass *llvm::createARMConstantIslandPass() { 2536 return new ARMConstantIslands(); 2537 } 2538 2539 INITIALIZE_PASS(ARMConstantIslands, "arm-cp-islands", ARM_CP_ISLANDS_OPT_NAME, 2540 false, false) 2541