1 //===-- HexagonFrameLowering.cpp - Define frame lowering ------------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 // 9 //===----------------------------------------------------------------------===// 10 11 #define DEBUG_TYPE "hexagon-pei" 12 13 #include "HexagonFrameLowering.h" 14 #include "Hexagon.h" 15 #include "HexagonInstrInfo.h" 16 #include "HexagonMachineFunctionInfo.h" 17 #include "HexagonRegisterInfo.h" 18 #include "HexagonSubtarget.h" 19 #include "HexagonTargetMachine.h" 20 #include "llvm/ADT/BitVector.h" 21 #include "llvm/ADT/PostOrderIterator.h" 22 #include "llvm/ADT/STLExtras.h" 23 #include "llvm/CodeGen/MachineDominators.h" 24 #include "llvm/CodeGen/MachineInstrBuilder.h" 25 #include "llvm/CodeGen/MachineFunction.h" 26 #include "llvm/CodeGen/MachineFunctionPass.h" 27 #include "llvm/CodeGen/MachineInstrBuilder.h" 28 #include "llvm/CodeGen/MachineModuleInfo.h" 29 #include "llvm/CodeGen/MachinePostDominators.h" 30 #include "llvm/CodeGen/MachineRegisterInfo.h" 31 #include "llvm/CodeGen/RegisterScavenging.h" 32 #include "llvm/IR/Function.h" 33 #include "llvm/IR/Type.h" 34 #include "llvm/Support/CommandLine.h" 35 #include "llvm/Support/Debug.h" 36 #include "llvm/Support/raw_ostream.h" 37 #include "llvm/Target/TargetInstrInfo.h" 38 #include "llvm/Target/TargetMachine.h" 39 #include "llvm/Target/TargetOptions.h" 40 41 // Hexagon stack frame layout as defined by the ABI: 42 // 43 // Incoming arguments 44 // passed via stack 45 // | 46 // | 47 // SP during function's FP during function's | 48 // +-- runtime (top of stack) runtime (bottom) --+ | 49 // | | | 50 // --++---------------------+------------------+-----------------++-+------- 51 // | parameter area for | variable-size | fixed-size |LR| arg 52 // | called functions | local objects | local objects |FP| 53 // --+----------------------+------------------+-----------------+--+------- 54 // <- size known -> <- size unknown -> <- size known -> 55 // 56 // Low address High address 57 // 58 // <--- stack growth 59 // 60 // 61 // - In any circumstances, the outgoing function arguments are always accessi- 62 // ble using the SP, and the incoming arguments are accessible using the FP. 63 // - If the local objects are not aligned, they can always be accessed using 64 // the FP. 65 // - If there are no variable-sized objects, the local objects can always be 66 // accessed using the SP, regardless whether they are aligned or not. (The 67 // alignment padding will be at the bottom of the stack (highest address), 68 // and so the offset with respect to the SP will be known at the compile- 69 // -time.) 70 // 71 // The only complication occurs if there are both, local aligned objects, and 72 // dynamically allocated (variable-sized) objects. The alignment pad will be 73 // placed between the FP and the local objects, thus preventing the use of the 74 // FP to access the local objects. At the same time, the variable-sized objects 75 // will be between the SP and the local objects, thus introducing an unknown 76 // distance from the SP to the locals. 77 // 78 // To avoid this problem, a new register is created that holds the aligned 79 // address of the bottom of the stack, referred in the sources as AP (aligned 80 // pointer). The AP will be equal to "FP-p", where "p" is the smallest pad 81 // that aligns AP to the required boundary (a maximum of the alignments of 82 // all stack objects, fixed- and variable-sized). All local objects[1] will 83 // then use AP as the base pointer. 84 // [1] The exception is with "fixed" stack objects. "Fixed" stack objects get 85 // their name from being allocated at fixed locations on the stack, relative 86 // to the FP. In the presence of dynamic allocation and local alignment, such 87 // objects can only be accessed through the FP. 88 // 89 // Illustration of the AP: 90 // FP --+ 91 // | 92 // ---------------+---------------------+-----+-----------------------++-+-- 93 // Rest of the | Local stack objects | Pad | Fixed stack objects |LR| 94 // stack frame | (aligned) | | (CSR, spills, etc.) |FP| 95 // ---------------+---------------------+-----+-----------------+-----+--+-- 96 // |<-- Multiple of the -->| 97 // stack alignment +-- AP 98 // 99 // The AP is set up at the beginning of the function. Since it is not a dedi- 100 // cated (reserved) register, it needs to be kept live throughout the function 101 // to be available as the base register for local object accesses. 102 // Normally, an address of a stack objects is obtained by a pseudo-instruction 103 // TFR_FI. To access local objects with the AP register present, a different 104 // pseudo-instruction needs to be used: TFR_FIA. The TFR_FIA takes one extra 105 // argument compared to TFR_FI: the first input register is the AP register. 106 // This keeps the register live between its definition and its uses. 107 108 // The AP register is originally set up using pseudo-instruction ALIGNA: 109 // AP = ALIGNA A 110 // where 111 // A - required stack alignment 112 // The alignment value must be the maximum of all alignments required by 113 // any stack object. 114 115 // The dynamic allocation uses a pseudo-instruction ALLOCA: 116 // Rd = ALLOCA Rs, A 117 // where 118 // Rd - address of the allocated space 119 // Rs - minimum size (the actual allocated can be larger to accommodate 120 // alignment) 121 // A - required alignment 122 123 124 using namespace llvm; 125 126 static cl::opt<bool> DisableDeallocRet("disable-hexagon-dealloc-ret", 127 cl::Hidden, cl::desc("Disable Dealloc Return for Hexagon target")); 128 129 130 static cl::opt<int> NumberScavengerSlots("number-scavenger-slots", 131 cl::Hidden, cl::desc("Set the number of scavenger slots"), cl::init(2), 132 cl::ZeroOrMore); 133 134 static cl::opt<int> SpillFuncThreshold("spill-func-threshold", 135 cl::Hidden, cl::desc("Specify O2(not Os) spill func threshold"), 136 cl::init(6), cl::ZeroOrMore); 137 138 static cl::opt<int> SpillFuncThresholdOs("spill-func-threshold-Os", 139 cl::Hidden, cl::desc("Specify Os spill func threshold"), 140 cl::init(1), cl::ZeroOrMore); 141 142 static cl::opt<bool> EnableShrinkWrapping("hexagon-shrink-frame", 143 cl::init(true), cl::Hidden, cl::ZeroOrMore, 144 cl::desc("Enable stack frame shrink wrapping")); 145 146 static cl::opt<unsigned> ShrinkLimit("shrink-frame-limit", cl::init(UINT_MAX), 147 cl::Hidden, cl::ZeroOrMore, cl::desc("Max count of stack frame " 148 "shrink-wraps")); 149 150 static cl::opt<bool> UseAllocframe("use-allocframe", cl::init(true), 151 cl::Hidden, cl::desc("Use allocframe more conservatively")); 152 153 154 namespace llvm { 155 void initializeHexagonCallFrameInformationPass(PassRegistry&); 156 FunctionPass *createHexagonCallFrameInformation(); 157 } 158 159 namespace { 160 class HexagonCallFrameInformation : public MachineFunctionPass { 161 public: 162 static char ID; 163 HexagonCallFrameInformation() : MachineFunctionPass(ID) { 164 PassRegistry &PR = *PassRegistry::getPassRegistry(); 165 initializeHexagonCallFrameInformationPass(PR); 166 } 167 bool runOnMachineFunction(MachineFunction &MF) override; 168 }; 169 170 char HexagonCallFrameInformation::ID = 0; 171 } 172 173 bool HexagonCallFrameInformation::runOnMachineFunction(MachineFunction &MF) { 174 auto &HFI = *MF.getSubtarget<HexagonSubtarget>().getFrameLowering(); 175 bool NeedCFI = MF.getMMI().hasDebugInfo() || 176 MF.getFunction()->needsUnwindTableEntry(); 177 178 if (!NeedCFI) 179 return false; 180 HFI.insertCFIInstructions(MF); 181 return true; 182 } 183 184 INITIALIZE_PASS(HexagonCallFrameInformation, "hexagon-cfi", 185 "Hexagon call frame information", false, false) 186 187 FunctionPass *llvm::createHexagonCallFrameInformation() { 188 return new HexagonCallFrameInformation(); 189 } 190 191 192 namespace { 193 /// Map a register pair Reg to the subregister that has the greater "number", 194 /// i.e. D3 (aka R7:6) will be mapped to R7, etc. 195 unsigned getMax32BitSubRegister(unsigned Reg, const TargetRegisterInfo &TRI, 196 bool hireg = true) { 197 if (Reg < Hexagon::D0 || Reg > Hexagon::D15) 198 return Reg; 199 200 unsigned RegNo = 0; 201 for (MCSubRegIterator SubRegs(Reg, &TRI); SubRegs.isValid(); ++SubRegs) { 202 if (hireg) { 203 if (*SubRegs > RegNo) 204 RegNo = *SubRegs; 205 } else { 206 if (!RegNo || *SubRegs < RegNo) 207 RegNo = *SubRegs; 208 } 209 } 210 return RegNo; 211 } 212 213 /// Returns the callee saved register with the largest id in the vector. 214 unsigned getMaxCalleeSavedReg(const std::vector<CalleeSavedInfo> &CSI, 215 const TargetRegisterInfo &TRI) { 216 assert(Hexagon::R1 > 0 && 217 "Assume physical registers are encoded as positive integers"); 218 if (CSI.empty()) 219 return 0; 220 221 unsigned Max = getMax32BitSubRegister(CSI[0].getReg(), TRI); 222 for (unsigned I = 1, E = CSI.size(); I < E; ++I) { 223 unsigned Reg = getMax32BitSubRegister(CSI[I].getReg(), TRI); 224 if (Reg > Max) 225 Max = Reg; 226 } 227 return Max; 228 } 229 230 /// Checks if the basic block contains any instruction that needs a stack 231 /// frame to be already in place. 232 bool needsStackFrame(const MachineBasicBlock &MBB, const BitVector &CSR) { 233 for (auto &I : MBB) { 234 const MachineInstr *MI = &I; 235 if (MI->isCall()) 236 return true; 237 unsigned Opc = MI->getOpcode(); 238 switch (Opc) { 239 case Hexagon::ALLOCA: 240 case Hexagon::ALIGNA: 241 return true; 242 default: 243 break; 244 } 245 // Check individual operands. 246 for (const MachineOperand &MO : MI->operands()) { 247 // While the presence of a frame index does not prove that a stack 248 // frame will be required, all frame indexes should be within alloc- 249 // frame/deallocframe. Otherwise, the code that translates a frame 250 // index into an offset would have to be aware of the placement of 251 // the frame creation/destruction instructions. 252 if (MO.isFI()) 253 return true; 254 if (!MO.isReg()) 255 continue; 256 unsigned R = MO.getReg(); 257 // Virtual registers will need scavenging, which then may require 258 // a stack slot. 259 if (TargetRegisterInfo::isVirtualRegister(R)) 260 return true; 261 if (CSR[R]) 262 return true; 263 } 264 } 265 return false; 266 } 267 268 /// Returns true if MBB has a machine instructions that indicates a tail call 269 /// in the block. 270 bool hasTailCall(const MachineBasicBlock &MBB) { 271 MachineBasicBlock::const_iterator I = MBB.getLastNonDebugInstr(); 272 unsigned RetOpc = I->getOpcode(); 273 return RetOpc == Hexagon::TCRETURNi || RetOpc == Hexagon::TCRETURNr; 274 } 275 276 /// Returns true if MBB contains an instruction that returns. 277 bool hasReturn(const MachineBasicBlock &MBB) { 278 for (auto I = MBB.getFirstTerminator(), E = MBB.end(); I != E; ++I) 279 if (I->isReturn()) 280 return true; 281 return false; 282 } 283 } 284 285 286 /// Implements shrink-wrapping of the stack frame. By default, stack frame 287 /// is created in the function entry block, and is cleaned up in every block 288 /// that returns. This function finds alternate blocks: one for the frame 289 /// setup (prolog) and one for the cleanup (epilog). 290 void HexagonFrameLowering::findShrunkPrologEpilog(MachineFunction &MF, 291 MachineBasicBlock *&PrologB, MachineBasicBlock *&EpilogB) const { 292 static unsigned ShrinkCounter = 0; 293 294 if (ShrinkLimit.getPosition()) { 295 if (ShrinkCounter >= ShrinkLimit) 296 return; 297 ShrinkCounter++; 298 } 299 300 auto &HST = static_cast<const HexagonSubtarget&>(MF.getSubtarget()); 301 auto &HRI = *HST.getRegisterInfo(); 302 303 MachineDominatorTree MDT; 304 MDT.runOnMachineFunction(MF); 305 MachinePostDominatorTree MPT; 306 MPT.runOnMachineFunction(MF); 307 308 typedef DenseMap<unsigned,unsigned> UnsignedMap; 309 UnsignedMap RPO; 310 typedef ReversePostOrderTraversal<const MachineFunction*> RPOTType; 311 RPOTType RPOT(&MF); 312 unsigned RPON = 0; 313 for (RPOTType::rpo_iterator I = RPOT.begin(), E = RPOT.end(); I != E; ++I) 314 RPO[(*I)->getNumber()] = RPON++; 315 316 // Don't process functions that have loops, at least for now. Placement 317 // of prolog and epilog must take loop structure into account. For simpli- 318 // city don't do it right now. 319 for (auto &I : MF) { 320 unsigned BN = RPO[I.getNumber()]; 321 for (auto SI = I.succ_begin(), SE = I.succ_end(); SI != SE; ++SI) { 322 // If found a back-edge, return. 323 if (RPO[(*SI)->getNumber()] <= BN) 324 return; 325 } 326 } 327 328 // Collect the set of blocks that need a stack frame to execute. Scan 329 // each block for uses/defs of callee-saved registers, calls, etc. 330 SmallVector<MachineBasicBlock*,16> SFBlocks; 331 BitVector CSR(Hexagon::NUM_TARGET_REGS); 332 for (const MCPhysReg *P = HRI.getCalleeSavedRegs(&MF); *P; ++P) 333 CSR[*P] = true; 334 335 for (auto &I : MF) 336 if (needsStackFrame(I, CSR)) 337 SFBlocks.push_back(&I); 338 339 DEBUG({ 340 dbgs() << "Blocks needing SF: {"; 341 for (auto &B : SFBlocks) 342 dbgs() << " BB#" << B->getNumber(); 343 dbgs() << " }\n"; 344 }); 345 // No frame needed? 346 if (SFBlocks.empty()) 347 return; 348 349 // Pick a common dominator and a common post-dominator. 350 MachineBasicBlock *DomB = SFBlocks[0]; 351 for (unsigned i = 1, n = SFBlocks.size(); i < n; ++i) { 352 DomB = MDT.findNearestCommonDominator(DomB, SFBlocks[i]); 353 if (!DomB) 354 break; 355 } 356 MachineBasicBlock *PDomB = SFBlocks[0]; 357 for (unsigned i = 1, n = SFBlocks.size(); i < n; ++i) { 358 PDomB = MPT.findNearestCommonDominator(PDomB, SFBlocks[i]); 359 if (!PDomB) 360 break; 361 } 362 DEBUG({ 363 dbgs() << "Computed dom block: BB#"; 364 if (DomB) dbgs() << DomB->getNumber(); 365 else dbgs() << "<null>"; 366 dbgs() << ", computed pdom block: BB#"; 367 if (PDomB) dbgs() << PDomB->getNumber(); 368 else dbgs() << "<null>"; 369 dbgs() << "\n"; 370 }); 371 if (!DomB || !PDomB) 372 return; 373 374 // Make sure that DomB dominates PDomB and PDomB post-dominates DomB. 375 if (!MDT.dominates(DomB, PDomB)) { 376 DEBUG(dbgs() << "Dom block does not dominate pdom block\n"); 377 return; 378 } 379 if (!MPT.dominates(PDomB, DomB)) { 380 DEBUG(dbgs() << "PDom block does not post-dominate dom block\n"); 381 return; 382 } 383 384 // Finally, everything seems right. 385 PrologB = DomB; 386 EpilogB = PDomB; 387 } 388 389 /// Perform most of the PEI work here: 390 /// - saving/restoring of the callee-saved registers, 391 /// - stack frame creation and destruction. 392 /// Normally, this work is distributed among various functions, but doing it 393 /// in one place allows shrink-wrapping of the stack frame. 394 void HexagonFrameLowering::emitPrologue(MachineFunction &MF, 395 MachineBasicBlock &MBB) const { 396 auto &HST = static_cast<const HexagonSubtarget&>(MF.getSubtarget()); 397 auto &HRI = *HST.getRegisterInfo(); 398 399 assert(&MF.front() == &MBB && "Shrink-wrapping not yet supported"); 400 MachineFrameInfo *MFI = MF.getFrameInfo(); 401 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo(); 402 403 MachineBasicBlock *PrologB = &MF.front(), *EpilogB = nullptr; 404 if (EnableShrinkWrapping) 405 findShrunkPrologEpilog(MF, PrologB, EpilogB); 406 407 insertCSRSpillsInBlock(*PrologB, CSI, HRI); 408 insertPrologueInBlock(*PrologB); 409 410 if (EpilogB) { 411 insertCSRRestoresInBlock(*EpilogB, CSI, HRI); 412 insertEpilogueInBlock(*EpilogB); 413 } else { 414 for (auto &B : MF) 415 if (B.isReturnBlock()) 416 insertCSRRestoresInBlock(B, CSI, HRI); 417 418 for (auto &B : MF) 419 if (B.isReturnBlock()) 420 insertEpilogueInBlock(B); 421 } 422 } 423 424 425 void HexagonFrameLowering::insertPrologueInBlock(MachineBasicBlock &MBB) const { 426 MachineFunction &MF = *MBB.getParent(); 427 MachineFrameInfo *MFI = MF.getFrameInfo(); 428 auto &HST = MF.getSubtarget<HexagonSubtarget>(); 429 auto &HII = *HST.getInstrInfo(); 430 auto &HRI = *HST.getRegisterInfo(); 431 DebugLoc dl; 432 433 unsigned MaxAlign = std::max(MFI->getMaxAlignment(), getStackAlignment()); 434 435 // Calculate the total stack frame size. 436 // Get the number of bytes to allocate from the FrameInfo. 437 unsigned FrameSize = MFI->getStackSize(); 438 // Round up the max call frame size to the max alignment on the stack. 439 unsigned MaxCFA = RoundUpToAlignment(MFI->getMaxCallFrameSize(), MaxAlign); 440 MFI->setMaxCallFrameSize(MaxCFA); 441 442 FrameSize = MaxCFA + RoundUpToAlignment(FrameSize, MaxAlign); 443 MFI->setStackSize(FrameSize); 444 445 bool AlignStack = (MaxAlign > getStackAlignment()); 446 447 // Get the number of bytes to allocate from the FrameInfo. 448 unsigned NumBytes = MFI->getStackSize(); 449 unsigned SP = HRI.getStackRegister(); 450 unsigned MaxCF = MFI->getMaxCallFrameSize(); 451 MachineBasicBlock::iterator InsertPt = MBB.begin(); 452 453 auto *FuncInfo = MF.getInfo<HexagonMachineFunctionInfo>(); 454 auto &AdjustRegs = FuncInfo->getAllocaAdjustInsts(); 455 456 for (auto MI : AdjustRegs) { 457 assert((MI->getOpcode() == Hexagon::ALLOCA) && "Expected alloca"); 458 expandAlloca(MI, HII, SP, MaxCF); 459 MI->eraseFromParent(); 460 } 461 462 if (!hasFP(MF)) 463 return; 464 465 // Check for overflow. 466 // Hexagon_TODO: Ugh! hardcoding. Is there an API that can be used? 467 const unsigned int ALLOCFRAME_MAX = 16384; 468 469 // Create a dummy memory operand to avoid allocframe from being treated as 470 // a volatile memory reference. 471 MachineMemOperand *MMO = 472 MF.getMachineMemOperand(MachinePointerInfo(), MachineMemOperand::MOStore, 473 4, 4); 474 475 if (NumBytes >= ALLOCFRAME_MAX) { 476 // Emit allocframe(#0). 477 BuildMI(MBB, InsertPt, dl, HII.get(Hexagon::S2_allocframe)) 478 .addImm(0) 479 .addMemOperand(MMO); 480 481 // Subtract offset from frame pointer. 482 // We use a caller-saved non-parameter register for that. 483 unsigned CallerSavedReg = HRI.getFirstCallerSavedNonParamReg(); 484 BuildMI(MBB, InsertPt, dl, HII.get(Hexagon::CONST32_Int_Real), 485 CallerSavedReg).addImm(NumBytes); 486 BuildMI(MBB, InsertPt, dl, HII.get(Hexagon::A2_sub), SP) 487 .addReg(SP) 488 .addReg(CallerSavedReg); 489 } else { 490 BuildMI(MBB, InsertPt, dl, HII.get(Hexagon::S2_allocframe)) 491 .addImm(NumBytes) 492 .addMemOperand(MMO); 493 } 494 495 if (AlignStack) { 496 BuildMI(MBB, InsertPt, dl, HII.get(Hexagon::A2_andir), SP) 497 .addReg(SP) 498 .addImm(-int64_t(MaxAlign)); 499 } 500 } 501 502 void HexagonFrameLowering::insertEpilogueInBlock(MachineBasicBlock &MBB) const { 503 MachineFunction &MF = *MBB.getParent(); 504 if (!hasFP(MF)) 505 return; 506 507 auto &HST = static_cast<const HexagonSubtarget&>(MF.getSubtarget()); 508 auto &HII = *HST.getInstrInfo(); 509 auto &HRI = *HST.getRegisterInfo(); 510 unsigned SP = HRI.getStackRegister(); 511 512 MachineInstr *RetI = nullptr; 513 for (auto &I : MBB) { 514 if (!I.isReturn()) 515 continue; 516 RetI = &I; 517 break; 518 } 519 unsigned RetOpc = RetI ? RetI->getOpcode() : 0; 520 521 MachineBasicBlock::iterator InsertPt = MBB.getFirstTerminator(); 522 DebugLoc DL; 523 if (InsertPt != MBB.end()) 524 DL = InsertPt->getDebugLoc(); 525 else if (!MBB.empty()) 526 DL = std::prev(MBB.end())->getDebugLoc(); 527 528 // Handle EH_RETURN. 529 if (RetOpc == Hexagon::EH_RETURN_JMPR) { 530 BuildMI(MBB, InsertPt, DL, HII.get(Hexagon::L2_deallocframe)); 531 BuildMI(MBB, InsertPt, DL, HII.get(Hexagon::A2_add), SP) 532 .addReg(SP) 533 .addReg(Hexagon::R28); 534 return; 535 } 536 537 // Check for RESTORE_DEALLOC_RET* tail call. Don't emit an extra dealloc- 538 // frame instruction if we encounter it. 539 if (RetOpc == Hexagon::RESTORE_DEALLOC_RET_JMP_V4) { 540 MachineBasicBlock::iterator It = RetI; 541 ++It; 542 // Delete all instructions after the RESTORE (except labels). 543 while (It != MBB.end()) { 544 if (!It->isLabel()) 545 It = MBB.erase(It); 546 else 547 ++It; 548 } 549 return; 550 } 551 552 // It is possible that the restoring code is a call to a library function. 553 // All of the restore* functions include "deallocframe", so we need to make 554 // sure that we don't add an extra one. 555 bool NeedsDeallocframe = true; 556 if (!MBB.empty() && InsertPt != MBB.begin()) { 557 MachineBasicBlock::iterator PrevIt = std::prev(InsertPt); 558 unsigned COpc = PrevIt->getOpcode(); 559 if (COpc == Hexagon::RESTORE_DEALLOC_BEFORE_TAILCALL_V4) 560 NeedsDeallocframe = false; 561 } 562 563 if (!NeedsDeallocframe) 564 return; 565 // If the returning instruction is JMPret, replace it with dealloc_return, 566 // otherwise just add deallocframe. The function could be returning via a 567 // tail call. 568 if (RetOpc != Hexagon::JMPret || DisableDeallocRet) { 569 BuildMI(MBB, InsertPt, DL, HII.get(Hexagon::L2_deallocframe)); 570 return; 571 } 572 unsigned NewOpc = Hexagon::L4_return; 573 MachineInstr *NewI = BuildMI(MBB, RetI, DL, HII.get(NewOpc)); 574 // Transfer the function live-out registers. 575 NewI->copyImplicitOps(MF, RetI); 576 MBB.erase(RetI); 577 } 578 579 580 namespace { 581 bool IsAllocFrame(MachineBasicBlock::const_iterator It) { 582 if (!It->isBundle()) 583 return It->getOpcode() == Hexagon::S2_allocframe; 584 auto End = It->getParent()->instr_end(); 585 MachineBasicBlock::const_instr_iterator I = It.getInstrIterator(); 586 while (++I != End && I->isBundled()) 587 if (I->getOpcode() == Hexagon::S2_allocframe) 588 return true; 589 return false; 590 } 591 592 MachineBasicBlock::iterator FindAllocFrame(MachineBasicBlock &B) { 593 for (auto &I : B) 594 if (IsAllocFrame(I)) 595 return I; 596 return B.end(); 597 } 598 } 599 600 601 void HexagonFrameLowering::insertCFIInstructions(MachineFunction &MF) const { 602 for (auto &B : MF) { 603 auto AF = FindAllocFrame(B); 604 if (AF == B.end()) 605 continue; 606 insertCFIInstructionsAt(B, ++AF); 607 } 608 } 609 610 611 void HexagonFrameLowering::insertCFIInstructionsAt(MachineBasicBlock &MBB, 612 MachineBasicBlock::iterator At) const { 613 MachineFunction &MF = *MBB.getParent(); 614 MachineFrameInfo *MFI = MF.getFrameInfo(); 615 MachineModuleInfo &MMI = MF.getMMI(); 616 auto &HST = MF.getSubtarget<HexagonSubtarget>(); 617 auto &HII = *HST.getInstrInfo(); 618 auto &HRI = *HST.getRegisterInfo(); 619 620 // If CFI instructions have debug information attached, something goes 621 // wrong with the final assembly generation: the prolog_end is placed 622 // in a wrong location. 623 DebugLoc DL; 624 const MCInstrDesc &CFID = HII.get(TargetOpcode::CFI_INSTRUCTION); 625 626 MCSymbol *FrameLabel = MMI.getContext().createTempSymbol(); 627 628 if (hasFP(MF)) { 629 unsigned DwFPReg = HRI.getDwarfRegNum(HRI.getFrameRegister(), true); 630 unsigned DwRAReg = HRI.getDwarfRegNum(HRI.getRARegister(), true); 631 632 // Define CFA via an offset from the value of FP. 633 // 634 // -8 -4 0 (SP) 635 // --+----+----+--------------------- 636 // | FP | LR | increasing addresses --> 637 // --+----+----+--------------------- 638 // | +-- Old SP (before allocframe) 639 // +-- New FP (after allocframe) 640 // 641 // MCCFIInstruction::createDefCfa subtracts the offset from the register. 642 // MCCFIInstruction::createOffset takes the offset without sign change. 643 auto DefCfa = MCCFIInstruction::createDefCfa(FrameLabel, DwFPReg, -8); 644 BuildMI(MBB, At, DL, CFID) 645 .addCFIIndex(MMI.addFrameInst(DefCfa)); 646 // R31 (return addr) = CFA - 4 647 auto OffR31 = MCCFIInstruction::createOffset(FrameLabel, DwRAReg, -4); 648 BuildMI(MBB, At, DL, CFID) 649 .addCFIIndex(MMI.addFrameInst(OffR31)); 650 // R30 (frame ptr) = CFA - 8 651 auto OffR30 = MCCFIInstruction::createOffset(FrameLabel, DwFPReg, -8); 652 BuildMI(MBB, At, DL, CFID) 653 .addCFIIndex(MMI.addFrameInst(OffR30)); 654 } 655 656 static unsigned int RegsToMove[] = { 657 Hexagon::R1, Hexagon::R0, Hexagon::R3, Hexagon::R2, 658 Hexagon::R17, Hexagon::R16, Hexagon::R19, Hexagon::R18, 659 Hexagon::R21, Hexagon::R20, Hexagon::R23, Hexagon::R22, 660 Hexagon::R25, Hexagon::R24, Hexagon::R27, Hexagon::R26, 661 Hexagon::D0, Hexagon::D1, Hexagon::D8, Hexagon::D9, 662 Hexagon::D10, Hexagon::D11, Hexagon::D12, Hexagon::D13, 663 Hexagon::NoRegister 664 }; 665 666 const std::vector<CalleeSavedInfo> &CSI = MFI->getCalleeSavedInfo(); 667 668 for (unsigned i = 0; RegsToMove[i] != Hexagon::NoRegister; ++i) { 669 unsigned Reg = RegsToMove[i]; 670 auto IfR = [Reg] (const CalleeSavedInfo &C) -> bool { 671 return C.getReg() == Reg; 672 }; 673 auto F = std::find_if(CSI.begin(), CSI.end(), IfR); 674 if (F == CSI.end()) 675 continue; 676 677 // Subtract 8 to make room for R30 and R31, which are added above. 678 unsigned FrameReg; 679 int64_t Offset = getFrameIndexReference(MF, F->getFrameIdx(), FrameReg) - 8; 680 681 if (Reg < Hexagon::D0 || Reg > Hexagon::D15) { 682 unsigned DwarfReg = HRI.getDwarfRegNum(Reg, true); 683 auto OffReg = MCCFIInstruction::createOffset(FrameLabel, DwarfReg, 684 Offset); 685 BuildMI(MBB, At, DL, CFID) 686 .addCFIIndex(MMI.addFrameInst(OffReg)); 687 } else { 688 // Split the double regs into subregs, and generate appropriate 689 // cfi_offsets. 690 // The only reason, we are split double regs is, llvm-mc does not 691 // understand paired registers for cfi_offset. 692 // Eg .cfi_offset r1:0, -64 693 694 unsigned HiReg = HRI.getSubReg(Reg, Hexagon::subreg_hireg); 695 unsigned LoReg = HRI.getSubReg(Reg, Hexagon::subreg_loreg); 696 unsigned HiDwarfReg = HRI.getDwarfRegNum(HiReg, true); 697 unsigned LoDwarfReg = HRI.getDwarfRegNum(LoReg, true); 698 auto OffHi = MCCFIInstruction::createOffset(FrameLabel, HiDwarfReg, 699 Offset+4); 700 BuildMI(MBB, At, DL, CFID) 701 .addCFIIndex(MMI.addFrameInst(OffHi)); 702 auto OffLo = MCCFIInstruction::createOffset(FrameLabel, LoDwarfReg, 703 Offset); 704 BuildMI(MBB, At, DL, CFID) 705 .addCFIIndex(MMI.addFrameInst(OffLo)); 706 } 707 } 708 } 709 710 711 bool HexagonFrameLowering::hasFP(const MachineFunction &MF) const { 712 auto &MFI = *MF.getFrameInfo(); 713 auto &HRI = *MF.getSubtarget<HexagonSubtarget>().getRegisterInfo(); 714 715 bool HasFixed = MFI.getNumFixedObjects(); 716 bool HasPrealloc = const_cast<MachineFrameInfo&>(MFI) 717 .getLocalFrameObjectCount(); 718 bool HasExtraAlign = HRI.needsStackRealignment(MF); 719 bool HasAlloca = MFI.hasVarSizedObjects(); 720 721 // Insert ALLOCFRAME if we need to or at -O0 for the debugger. Think 722 // that this shouldn't be required, but doing so now because gcc does and 723 // gdb can't break at the start of the function without it. Will remove if 724 // this turns out to be a gdb bug. 725 // 726 if (MF.getTarget().getOptLevel() == CodeGenOpt::None) 727 return true; 728 729 // By default we want to use SP (since it's always there). FP requires 730 // some setup (i.e. ALLOCFRAME). 731 // Fixed and preallocated objects need FP if the distance from them to 732 // the SP is unknown (as is with alloca or aligna). 733 if ((HasFixed || HasPrealloc) && (HasAlloca || HasExtraAlign)) 734 return true; 735 736 if (MFI.getStackSize() > 0) { 737 if (UseAllocframe) 738 return true; 739 } 740 741 if (MFI.hasCalls() || 742 MF.getInfo<HexagonMachineFunctionInfo>()->hasClobberLR()) 743 return true; 744 745 return false; 746 } 747 748 749 enum SpillKind { 750 SK_ToMem, 751 SK_FromMem, 752 SK_FromMemTailcall 753 }; 754 755 static const char * 756 getSpillFunctionFor(unsigned MaxReg, SpillKind SpillType) { 757 const char * V4SpillToMemoryFunctions[] = { 758 "__save_r16_through_r17", 759 "__save_r16_through_r19", 760 "__save_r16_through_r21", 761 "__save_r16_through_r23", 762 "__save_r16_through_r25", 763 "__save_r16_through_r27" }; 764 765 const char * V4SpillFromMemoryFunctions[] = { 766 "__restore_r16_through_r17_and_deallocframe", 767 "__restore_r16_through_r19_and_deallocframe", 768 "__restore_r16_through_r21_and_deallocframe", 769 "__restore_r16_through_r23_and_deallocframe", 770 "__restore_r16_through_r25_and_deallocframe", 771 "__restore_r16_through_r27_and_deallocframe" }; 772 773 const char * V4SpillFromMemoryTailcallFunctions[] = { 774 "__restore_r16_through_r17_and_deallocframe_before_tailcall", 775 "__restore_r16_through_r19_and_deallocframe_before_tailcall", 776 "__restore_r16_through_r21_and_deallocframe_before_tailcall", 777 "__restore_r16_through_r23_and_deallocframe_before_tailcall", 778 "__restore_r16_through_r25_and_deallocframe_before_tailcall", 779 "__restore_r16_through_r27_and_deallocframe_before_tailcall" 780 }; 781 782 const char **SpillFunc = nullptr; 783 784 switch(SpillType) { 785 case SK_ToMem: 786 SpillFunc = V4SpillToMemoryFunctions; 787 break; 788 case SK_FromMem: 789 SpillFunc = V4SpillFromMemoryFunctions; 790 break; 791 case SK_FromMemTailcall: 792 SpillFunc = V4SpillFromMemoryTailcallFunctions; 793 break; 794 } 795 assert(SpillFunc && "Unknown spill kind"); 796 797 // Spill all callee-saved registers up to the highest register used. 798 switch (MaxReg) { 799 case Hexagon::R17: 800 return SpillFunc[0]; 801 case Hexagon::R19: 802 return SpillFunc[1]; 803 case Hexagon::R21: 804 return SpillFunc[2]; 805 case Hexagon::R23: 806 return SpillFunc[3]; 807 case Hexagon::R25: 808 return SpillFunc[4]; 809 case Hexagon::R27: 810 return SpillFunc[5]; 811 default: 812 llvm_unreachable("Unhandled maximum callee save register"); 813 } 814 return 0; 815 } 816 817 /// Adds all callee-saved registers up to MaxReg to the instruction. 818 static void addCalleeSaveRegistersAsImpOperand(MachineInstr *Inst, 819 unsigned MaxReg, bool IsDef) { 820 // Add the callee-saved registers as implicit uses. 821 for (unsigned R = Hexagon::R16; R <= MaxReg; ++R) { 822 MachineOperand ImpUse = MachineOperand::CreateReg(R, IsDef, true); 823 Inst->addOperand(ImpUse); 824 } 825 } 826 827 828 int HexagonFrameLowering::getFrameIndexReference(const MachineFunction &MF, 829 int FI, unsigned &FrameReg) const { 830 auto &MFI = *MF.getFrameInfo(); 831 auto &HRI = *MF.getSubtarget<HexagonSubtarget>().getRegisterInfo(); 832 833 // Large parts of this code are shared with HRI::eliminateFrameIndex. 834 int Offset = MFI.getObjectOffset(FI); 835 bool HasAlloca = MFI.hasVarSizedObjects(); 836 bool HasExtraAlign = HRI.needsStackRealignment(MF); 837 bool NoOpt = MF.getTarget().getOptLevel() == CodeGenOpt::None; 838 839 unsigned SP = HRI.getStackRegister(), FP = HRI.getFrameRegister(); 840 unsigned AP = 0; 841 if (const MachineInstr *AI = getAlignaInstr(MF)) 842 AP = AI->getOperand(0).getReg(); 843 unsigned FrameSize = MFI.getStackSize(); 844 845 bool UseFP = false, UseAP = false; // Default: use SP (except at -O0). 846 // Use FP at -O0, except when there are objects with extra alignment. 847 // That additional alignment requirement may cause a pad to be inserted, 848 // which will make it impossible to use FP to access objects located 849 // past the pad. 850 if (NoOpt && !HasExtraAlign) 851 UseFP = true; 852 if (MFI.isFixedObjectIndex(FI) || MFI.isObjectPreAllocated(FI)) { 853 // Fixed and preallocated objects will be located before any padding 854 // so FP must be used to access them. 855 UseFP |= (HasAlloca || HasExtraAlign); 856 } else { 857 if (HasAlloca) { 858 if (HasExtraAlign) 859 UseAP = true; 860 else 861 UseFP = true; 862 } 863 } 864 865 // If FP was picked, then there had better be FP. 866 bool HasFP = hasFP(MF); 867 assert((HasFP || !UseFP) && "This function must have frame pointer"); 868 869 // Having FP implies allocframe. Allocframe will store extra 8 bytes: 870 // FP/LR. If the base register is used to access an object across these 871 // 8 bytes, then the offset will need to be adjusted by 8. 872 // 873 // After allocframe: 874 // HexagonISelLowering adds 8 to ---+ 875 // the offsets of all stack-based | 876 // arguments (*) | 877 // | 878 // getObjectOffset < 0 0 8 getObjectOffset >= 8 879 // ------------------------+-----+------------------------> increasing 880 // <local objects> |FP/LR| <input arguments> addresses 881 // -----------------+------+-----+------------------------> 882 // | | 883 // SP/AP point --+ +-- FP points here (**) 884 // somewhere on 885 // this side of FP/LR 886 // 887 // (*) See LowerFormalArguments. The FP/LR is assumed to be present. 888 // (**) *FP == old-FP. FP+0..7 are the bytes of FP/LR. 889 890 // The lowering assumes that FP/LR is present, and so the offsets of 891 // the formal arguments start at 8. If FP/LR is not there we need to 892 // reduce the offset by 8. 893 if (Offset > 0 && !HasFP) 894 Offset -= 8; 895 896 if (UseFP) 897 FrameReg = FP; 898 else if (UseAP) 899 FrameReg = AP; 900 else 901 FrameReg = SP; 902 903 // Calculate the actual offset in the instruction. If there is no FP 904 // (in other words, no allocframe), then SP will not be adjusted (i.e. 905 // there will be no SP -= FrameSize), so the frame size should not be 906 // added to the calculated offset. 907 int RealOffset = Offset; 908 if (!UseFP && !UseAP && HasFP) 909 RealOffset = FrameSize+Offset; 910 return RealOffset; 911 } 912 913 914 bool HexagonFrameLowering::insertCSRSpillsInBlock(MachineBasicBlock &MBB, 915 const CSIVect &CSI, const HexagonRegisterInfo &HRI) const { 916 if (CSI.empty()) 917 return true; 918 919 MachineBasicBlock::iterator MI = MBB.begin(); 920 MachineFunction &MF = *MBB.getParent(); 921 auto &HII = *MF.getSubtarget<HexagonSubtarget>().getInstrInfo(); 922 923 if (useSpillFunction(MF, CSI)) { 924 unsigned MaxReg = getMaxCalleeSavedReg(CSI, HRI); 925 const char *SpillFun = getSpillFunctionFor(MaxReg, SK_ToMem); 926 // Call spill function. 927 DebugLoc DL = MI != MBB.end() ? MI->getDebugLoc() : DebugLoc(); 928 MachineInstr *SaveRegsCall = 929 BuildMI(MBB, MI, DL, HII.get(Hexagon::SAVE_REGISTERS_CALL_V4)) 930 .addExternalSymbol(SpillFun); 931 // Add callee-saved registers as use. 932 addCalleeSaveRegistersAsImpOperand(SaveRegsCall, MaxReg, false); 933 // Add live in registers. 934 for (unsigned I = 0; I < CSI.size(); ++I) 935 MBB.addLiveIn(CSI[I].getReg()); 936 return true; 937 } 938 939 for (unsigned i = 0, n = CSI.size(); i < n; ++i) { 940 unsigned Reg = CSI[i].getReg(); 941 // Add live in registers. We treat eh_return callee saved register r0 - r3 942 // specially. They are not really callee saved registers as they are not 943 // supposed to be killed. 944 bool IsKill = !HRI.isEHReturnCalleeSaveReg(Reg); 945 int FI = CSI[i].getFrameIdx(); 946 const TargetRegisterClass *RC = HRI.getMinimalPhysRegClass(Reg); 947 HII.storeRegToStackSlot(MBB, MI, Reg, IsKill, FI, RC, &HRI); 948 if (IsKill) 949 MBB.addLiveIn(Reg); 950 } 951 return true; 952 } 953 954 955 bool HexagonFrameLowering::insertCSRRestoresInBlock(MachineBasicBlock &MBB, 956 const CSIVect &CSI, const HexagonRegisterInfo &HRI) const { 957 if (CSI.empty()) 958 return false; 959 960 MachineBasicBlock::iterator MI = MBB.getFirstTerminator(); 961 MachineFunction &MF = *MBB.getParent(); 962 auto &HII = *MF.getSubtarget<HexagonSubtarget>().getInstrInfo(); 963 964 if (useRestoreFunction(MF, CSI)) { 965 bool HasTC = hasTailCall(MBB) || !hasReturn(MBB); 966 unsigned MaxR = getMaxCalleeSavedReg(CSI, HRI); 967 SpillKind Kind = HasTC ? SK_FromMemTailcall : SK_FromMem; 968 const char *RestoreFn = getSpillFunctionFor(MaxR, Kind); 969 970 // Call spill function. 971 DebugLoc DL = MI != MBB.end() ? MI->getDebugLoc() 972 : MBB.getLastNonDebugInstr()->getDebugLoc(); 973 MachineInstr *DeallocCall = nullptr; 974 975 if (HasTC) { 976 unsigned ROpc = Hexagon::RESTORE_DEALLOC_BEFORE_TAILCALL_V4; 977 DeallocCall = BuildMI(MBB, MI, DL, HII.get(ROpc)) 978 .addExternalSymbol(RestoreFn); 979 } else { 980 // The block has a return. 981 MachineBasicBlock::iterator It = MBB.getFirstTerminator(); 982 assert(It->isReturn() && std::next(It) == MBB.end()); 983 unsigned ROpc = Hexagon::RESTORE_DEALLOC_RET_JMP_V4; 984 DeallocCall = BuildMI(MBB, It, DL, HII.get(ROpc)) 985 .addExternalSymbol(RestoreFn); 986 // Transfer the function live-out registers. 987 DeallocCall->copyImplicitOps(MF, It); 988 } 989 addCalleeSaveRegistersAsImpOperand(DeallocCall, MaxR, true); 990 return true; 991 } 992 993 for (unsigned i = 0; i < CSI.size(); ++i) { 994 unsigned Reg = CSI[i].getReg(); 995 const TargetRegisterClass *RC = HRI.getMinimalPhysRegClass(Reg); 996 int FI = CSI[i].getFrameIdx(); 997 HII.loadRegFromStackSlot(MBB, MI, Reg, FI, RC, &HRI); 998 } 999 return true; 1000 } 1001 1002 1003 void HexagonFrameLowering::eliminateCallFramePseudoInstr(MachineFunction &MF, 1004 MachineBasicBlock &MBB, MachineBasicBlock::iterator I) const { 1005 MachineInstr &MI = *I; 1006 unsigned Opc = MI.getOpcode(); 1007 (void)Opc; // Silence compiler warning. 1008 assert((Opc == Hexagon::ADJCALLSTACKDOWN || Opc == Hexagon::ADJCALLSTACKUP) && 1009 "Cannot handle this call frame pseudo instruction"); 1010 MBB.erase(I); 1011 } 1012 1013 1014 void HexagonFrameLowering::processFunctionBeforeFrameFinalized( 1015 MachineFunction &MF, RegScavenger *RS) const { 1016 // If this function has uses aligned stack and also has variable sized stack 1017 // objects, then we need to map all spill slots to fixed positions, so that 1018 // they can be accessed through FP. Otherwise they would have to be accessed 1019 // via AP, which may not be available at the particular place in the program. 1020 MachineFrameInfo *MFI = MF.getFrameInfo(); 1021 bool HasAlloca = MFI->hasVarSizedObjects(); 1022 bool NeedsAlign = (MFI->getMaxAlignment() > getStackAlignment()); 1023 1024 if (!HasAlloca || !NeedsAlign) 1025 return; 1026 1027 unsigned LFS = MFI->getLocalFrameSize(); 1028 int Offset = -LFS; 1029 for (int i = 0, e = MFI->getObjectIndexEnd(); i != e; ++i) { 1030 if (!MFI->isSpillSlotObjectIndex(i) || MFI->isDeadObjectIndex(i)) 1031 continue; 1032 int S = MFI->getObjectSize(i); 1033 LFS += S; 1034 Offset -= S; 1035 MFI->mapLocalFrameObject(i, Offset); 1036 } 1037 1038 MFI->setLocalFrameSize(LFS); 1039 unsigned A = MFI->getLocalFrameMaxAlign(); 1040 assert(A <= 8 && "Unexpected local frame alignment"); 1041 if (A == 0) 1042 MFI->setLocalFrameMaxAlign(8); 1043 MFI->setUseLocalStackAllocationBlock(true); 1044 } 1045 1046 /// Returns true if there is no caller saved registers available. 1047 static bool needToReserveScavengingSpillSlots(MachineFunction &MF, 1048 const HexagonRegisterInfo &HRI) { 1049 MachineRegisterInfo &MRI = MF.getRegInfo(); 1050 const MCPhysReg *CallerSavedRegs = HRI.getCallerSavedRegs(&MF); 1051 // Check for an unused caller-saved register. 1052 for ( ; *CallerSavedRegs; ++CallerSavedRegs) { 1053 MCPhysReg FreeReg = *CallerSavedRegs; 1054 if (!MRI.reg_nodbg_empty(FreeReg)) 1055 continue; 1056 1057 // Check aliased register usage. 1058 bool IsCurrentRegUsed = false; 1059 for (MCRegAliasIterator AI(FreeReg, &HRI, false); AI.isValid(); ++AI) 1060 if (!MRI.reg_nodbg_empty(*AI)) { 1061 IsCurrentRegUsed = true; 1062 break; 1063 } 1064 if (IsCurrentRegUsed) 1065 continue; 1066 1067 // Neither directly used nor used through an aliased register. 1068 return false; 1069 } 1070 // All caller-saved registers are used. 1071 return true; 1072 } 1073 1074 1075 /// Replaces the predicate spill code pseudo instructions by valid instructions. 1076 bool HexagonFrameLowering::replacePredRegPseudoSpillCode(MachineFunction &MF) 1077 const { 1078 auto &HST = static_cast<const HexagonSubtarget&>(MF.getSubtarget()); 1079 auto &HII = *HST.getInstrInfo(); 1080 MachineRegisterInfo &MRI = MF.getRegInfo(); 1081 bool HasReplacedPseudoInst = false; 1082 // Replace predicate spill pseudo instructions by real code. 1083 // Loop over all of the basic blocks. 1084 for (MachineFunction::iterator MBBb = MF.begin(), MBBe = MF.end(); 1085 MBBb != MBBe; ++MBBb) { 1086 MachineBasicBlock *MBB = &*MBBb; 1087 // Traverse the basic block. 1088 MachineBasicBlock::iterator NextII; 1089 for (MachineBasicBlock::iterator MII = MBB->begin(); MII != MBB->end(); 1090 MII = NextII) { 1091 MachineInstr *MI = MII; 1092 NextII = std::next(MII); 1093 int Opc = MI->getOpcode(); 1094 if (Opc == Hexagon::STriw_pred) { 1095 HasReplacedPseudoInst = true; 1096 // STriw_pred FI, 0, SrcReg; 1097 unsigned VirtReg = MRI.createVirtualRegister(&Hexagon::IntRegsRegClass); 1098 unsigned SrcReg = MI->getOperand(2).getReg(); 1099 bool IsOrigSrcRegKilled = MI->getOperand(2).isKill(); 1100 1101 assert(MI->getOperand(0).isFI() && "Expect a frame index"); 1102 assert(Hexagon::PredRegsRegClass.contains(SrcReg) && 1103 "Not a predicate register"); 1104 1105 // Insert transfer to general purpose register. 1106 // VirtReg = C2_tfrpr SrcPredReg 1107 BuildMI(*MBB, MII, MI->getDebugLoc(), HII.get(Hexagon::C2_tfrpr), 1108 VirtReg).addReg(SrcReg, getKillRegState(IsOrigSrcRegKilled)); 1109 1110 // Change instruction to S2_storeri_io. 1111 // S2_storeri_io FI, 0, VirtReg 1112 MI->setDesc(HII.get(Hexagon::S2_storeri_io)); 1113 MI->getOperand(2).setReg(VirtReg); 1114 MI->getOperand(2).setIsKill(); 1115 1116 } else if (Opc == Hexagon::LDriw_pred) { 1117 // DstReg = LDriw_pred FI, 0 1118 MachineOperand &M0 = MI->getOperand(0); 1119 if (M0.isDead()) { 1120 MBB->erase(MII); 1121 continue; 1122 } 1123 1124 unsigned VirtReg = MRI.createVirtualRegister(&Hexagon::IntRegsRegClass); 1125 unsigned DestReg = MI->getOperand(0).getReg(); 1126 1127 assert(MI->getOperand(1).isFI() && "Expect a frame index"); 1128 assert(Hexagon::PredRegsRegClass.contains(DestReg) && 1129 "Not a predicate register"); 1130 1131 // Change instruction to L2_loadri_io. 1132 // VirtReg = L2_loadri_io FI, 0 1133 MI->setDesc(HII.get(Hexagon::L2_loadri_io)); 1134 MI->getOperand(0).setReg(VirtReg); 1135 1136 // Insert transfer to general purpose register. 1137 // DestReg = C2_tfrrp VirtReg 1138 const MCInstrDesc &D = HII.get(Hexagon::C2_tfrrp); 1139 BuildMI(*MBB, std::next(MII), MI->getDebugLoc(), D, DestReg) 1140 .addReg(VirtReg, getKillRegState(true)); 1141 HasReplacedPseudoInst = true; 1142 } 1143 } 1144 } 1145 return HasReplacedPseudoInst; 1146 } 1147 1148 1149 void HexagonFrameLowering::determineCalleeSaves(MachineFunction &MF, 1150 BitVector &SavedRegs, 1151 RegScavenger *RS) const { 1152 TargetFrameLowering::determineCalleeSaves(MF, SavedRegs, RS); 1153 1154 auto &HST = static_cast<const HexagonSubtarget&>(MF.getSubtarget()); 1155 auto &HRI = *HST.getRegisterInfo(); 1156 1157 bool HasEHReturn = MF.getInfo<HexagonMachineFunctionInfo>()->hasEHReturn(); 1158 1159 // If we have a function containing __builtin_eh_return we want to spill and 1160 // restore all callee saved registers. Pretend that they are used. 1161 if (HasEHReturn) { 1162 for (const MCPhysReg *CSRegs = HRI.getCalleeSavedRegs(&MF); *CSRegs; 1163 ++CSRegs) 1164 SavedRegs.set(*CSRegs); 1165 } 1166 1167 const TargetRegisterClass &RC = Hexagon::IntRegsRegClass; 1168 1169 // Replace predicate register pseudo spill code. 1170 bool HasReplacedPseudoInst = replacePredRegPseudoSpillCode(MF); 1171 1172 // We need to reserve a a spill slot if scavenging could potentially require 1173 // spilling a scavenged register. 1174 if (HasReplacedPseudoInst && needToReserveScavengingSpillSlots(MF, HRI)) { 1175 MachineFrameInfo *MFI = MF.getFrameInfo(); 1176 for (int i=0; i < NumberScavengerSlots; i++) 1177 RS->addScavengingFrameIndex( 1178 MFI->CreateSpillStackObject(RC.getSize(), RC.getAlignment())); 1179 } 1180 } 1181 1182 1183 #ifndef NDEBUG 1184 static void dump_registers(BitVector &Regs, const TargetRegisterInfo &TRI) { 1185 dbgs() << '{'; 1186 for (int x = Regs.find_first(); x >= 0; x = Regs.find_next(x)) { 1187 unsigned R = x; 1188 dbgs() << ' ' << PrintReg(R, &TRI); 1189 } 1190 dbgs() << " }"; 1191 } 1192 #endif 1193 1194 1195 bool HexagonFrameLowering::assignCalleeSavedSpillSlots(MachineFunction &MF, 1196 const TargetRegisterInfo *TRI, std::vector<CalleeSavedInfo> &CSI) const { 1197 DEBUG(dbgs() << LLVM_FUNCTION_NAME << " on " 1198 << MF.getFunction()->getName() << '\n'); 1199 MachineFrameInfo *MFI = MF.getFrameInfo(); 1200 BitVector SRegs(Hexagon::NUM_TARGET_REGS); 1201 1202 // Generate a set of unique, callee-saved registers (SRegs), where each 1203 // register in the set is maximal in terms of sub-/super-register relation, 1204 // i.e. for each R in SRegs, no proper super-register of R is also in SRegs. 1205 1206 // (1) For each callee-saved register, add that register and all of its 1207 // sub-registers to SRegs. 1208 DEBUG(dbgs() << "Initial CS registers: {"); 1209 for (unsigned i = 0, n = CSI.size(); i < n; ++i) { 1210 unsigned R = CSI[i].getReg(); 1211 DEBUG(dbgs() << ' ' << PrintReg(R, TRI)); 1212 for (MCSubRegIterator SR(R, TRI, true); SR.isValid(); ++SR) 1213 SRegs[*SR] = true; 1214 } 1215 DEBUG(dbgs() << " }\n"); 1216 DEBUG(dbgs() << "SRegs.1: "; dump_registers(SRegs, *TRI); dbgs() << "\n"); 1217 1218 // (2) For each reserved register, remove that register and all of its 1219 // sub- and super-registers from SRegs. 1220 BitVector Reserved = TRI->getReservedRegs(MF); 1221 for (int x = Reserved.find_first(); x >= 0; x = Reserved.find_next(x)) { 1222 unsigned R = x; 1223 for (MCSuperRegIterator SR(R, TRI, true); SR.isValid(); ++SR) 1224 SRegs[*SR] = false; 1225 } 1226 DEBUG(dbgs() << "Res: "; dump_registers(Reserved, *TRI); dbgs() << "\n"); 1227 DEBUG(dbgs() << "SRegs.2: "; dump_registers(SRegs, *TRI); dbgs() << "\n"); 1228 1229 // (3) Collect all registers that have at least one sub-register in SRegs, 1230 // and also have no sub-registers that are reserved. These will be the can- 1231 // didates for saving as a whole instead of their individual sub-registers. 1232 // (Saving R17:16 instead of R16 is fine, but only if R17 was not reserved.) 1233 BitVector TmpSup(Hexagon::NUM_TARGET_REGS); 1234 for (int x = SRegs.find_first(); x >= 0; x = SRegs.find_next(x)) { 1235 unsigned R = x; 1236 for (MCSuperRegIterator SR(R, TRI); SR.isValid(); ++SR) 1237 TmpSup[*SR] = true; 1238 } 1239 for (int x = TmpSup.find_first(); x >= 0; x = TmpSup.find_next(x)) { 1240 unsigned R = x; 1241 for (MCSubRegIterator SR(R, TRI, true); SR.isValid(); ++SR) { 1242 if (!Reserved[*SR]) 1243 continue; 1244 TmpSup[R] = false; 1245 break; 1246 } 1247 } 1248 DEBUG(dbgs() << "TmpSup: "; dump_registers(TmpSup, *TRI); dbgs() << "\n"); 1249 1250 // (4) Include all super-registers found in (3) into SRegs. 1251 SRegs |= TmpSup; 1252 DEBUG(dbgs() << "SRegs.4: "; dump_registers(SRegs, *TRI); dbgs() << "\n"); 1253 1254 // (5) For each register R in SRegs, if any super-register of R is in SRegs, 1255 // remove R from SRegs. 1256 for (int x = SRegs.find_first(); x >= 0; x = SRegs.find_next(x)) { 1257 unsigned R = x; 1258 for (MCSuperRegIterator SR(R, TRI); SR.isValid(); ++SR) { 1259 if (!SRegs[*SR]) 1260 continue; 1261 SRegs[R] = false; 1262 break; 1263 } 1264 } 1265 DEBUG(dbgs() << "SRegs.5: "; dump_registers(SRegs, *TRI); dbgs() << "\n"); 1266 1267 // Now, for each register that has a fixed stack slot, create the stack 1268 // object for it. 1269 CSI.clear(); 1270 1271 typedef TargetFrameLowering::SpillSlot SpillSlot; 1272 unsigned NumFixed; 1273 int MinOffset = 0; // CS offsets are negative. 1274 const SpillSlot *FixedSlots = getCalleeSavedSpillSlots(NumFixed); 1275 for (const SpillSlot *S = FixedSlots; S != FixedSlots+NumFixed; ++S) { 1276 if (!SRegs[S->Reg]) 1277 continue; 1278 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(S->Reg); 1279 int FI = MFI->CreateFixedSpillStackObject(RC->getSize(), S->Offset); 1280 MinOffset = std::min(MinOffset, S->Offset); 1281 CSI.push_back(CalleeSavedInfo(S->Reg, FI)); 1282 SRegs[S->Reg] = false; 1283 } 1284 1285 // There can be some registers that don't have fixed slots. For example, 1286 // we need to store R0-R3 in functions with exception handling. For each 1287 // such register, create a non-fixed stack object. 1288 for (int x = SRegs.find_first(); x >= 0; x = SRegs.find_next(x)) { 1289 unsigned R = x; 1290 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(R); 1291 int Off = MinOffset - RC->getSize(); 1292 unsigned Align = std::min(RC->getAlignment(), getStackAlignment()); 1293 assert(isPowerOf2_32(Align)); 1294 Off &= -Align; 1295 int FI = MFI->CreateFixedSpillStackObject(RC->getSize(), Off); 1296 MinOffset = std::min(MinOffset, Off); 1297 CSI.push_back(CalleeSavedInfo(R, FI)); 1298 SRegs[R] = false; 1299 } 1300 1301 DEBUG({ 1302 dbgs() << "CS information: {"; 1303 for (unsigned i = 0, n = CSI.size(); i < n; ++i) { 1304 int FI = CSI[i].getFrameIdx(); 1305 int Off = MFI->getObjectOffset(FI); 1306 dbgs() << ' ' << PrintReg(CSI[i].getReg(), TRI) << ":fi#" << FI << ":sp"; 1307 if (Off >= 0) 1308 dbgs() << '+'; 1309 dbgs() << Off; 1310 } 1311 dbgs() << " }\n"; 1312 }); 1313 1314 #ifndef NDEBUG 1315 // Verify that all registers were handled. 1316 bool MissedReg = false; 1317 for (int x = SRegs.find_first(); x >= 0; x = SRegs.find_next(x)) { 1318 unsigned R = x; 1319 dbgs() << PrintReg(R, TRI) << ' '; 1320 MissedReg = true; 1321 } 1322 if (MissedReg) 1323 llvm_unreachable("...there are unhandled callee-saved registers!"); 1324 #endif 1325 1326 return true; 1327 } 1328 1329 1330 void HexagonFrameLowering::expandAlloca(MachineInstr *AI, 1331 const HexagonInstrInfo &HII, unsigned SP, unsigned CF) const { 1332 MachineBasicBlock &MB = *AI->getParent(); 1333 DebugLoc DL = AI->getDebugLoc(); 1334 unsigned A = AI->getOperand(2).getImm(); 1335 1336 // Have 1337 // Rd = alloca Rs, #A 1338 // 1339 // If Rs and Rd are different registers, use this sequence: 1340 // Rd = sub(r29, Rs) 1341 // r29 = sub(r29, Rs) 1342 // Rd = and(Rd, #-A) ; if necessary 1343 // r29 = and(r29, #-A) ; if necessary 1344 // Rd = add(Rd, #CF) ; CF size aligned to at most A 1345 // otherwise, do 1346 // Rd = sub(r29, Rs) 1347 // Rd = and(Rd, #-A) ; if necessary 1348 // r29 = Rd 1349 // Rd = add(Rd, #CF) ; CF size aligned to at most A 1350 1351 MachineOperand &RdOp = AI->getOperand(0); 1352 MachineOperand &RsOp = AI->getOperand(1); 1353 unsigned Rd = RdOp.getReg(), Rs = RsOp.getReg(); 1354 1355 // Rd = sub(r29, Rs) 1356 BuildMI(MB, AI, DL, HII.get(Hexagon::A2_sub), Rd) 1357 .addReg(SP) 1358 .addReg(Rs); 1359 if (Rs != Rd) { 1360 // r29 = sub(r29, Rs) 1361 BuildMI(MB, AI, DL, HII.get(Hexagon::A2_sub), SP) 1362 .addReg(SP) 1363 .addReg(Rs); 1364 } 1365 if (A > 8) { 1366 // Rd = and(Rd, #-A) 1367 BuildMI(MB, AI, DL, HII.get(Hexagon::A2_andir), Rd) 1368 .addReg(Rd) 1369 .addImm(-int64_t(A)); 1370 if (Rs != Rd) 1371 BuildMI(MB, AI, DL, HII.get(Hexagon::A2_andir), SP) 1372 .addReg(SP) 1373 .addImm(-int64_t(A)); 1374 } 1375 if (Rs == Rd) { 1376 // r29 = Rd 1377 BuildMI(MB, AI, DL, HII.get(TargetOpcode::COPY), SP) 1378 .addReg(Rd); 1379 } 1380 if (CF > 0) { 1381 // Rd = add(Rd, #CF) 1382 BuildMI(MB, AI, DL, HII.get(Hexagon::A2_addi), Rd) 1383 .addReg(Rd) 1384 .addImm(CF); 1385 } 1386 } 1387 1388 1389 bool HexagonFrameLowering::needsAligna(const MachineFunction &MF) const { 1390 const MachineFrameInfo *MFI = MF.getFrameInfo(); 1391 if (!MFI->hasVarSizedObjects()) 1392 return false; 1393 unsigned MaxA = MFI->getMaxAlignment(); 1394 if (MaxA <= getStackAlignment()) 1395 return false; 1396 return true; 1397 } 1398 1399 1400 const MachineInstr *HexagonFrameLowering::getAlignaInstr( 1401 const MachineFunction &MF) const { 1402 for (auto &B : MF) 1403 for (auto &I : B) 1404 if (I.getOpcode() == Hexagon::ALIGNA) 1405 return &I; 1406 return nullptr; 1407 } 1408 1409 1410 // FIXME: Use Function::optForSize(). 1411 inline static bool isOptSize(const MachineFunction &MF) { 1412 AttributeSet AF = MF.getFunction()->getAttributes(); 1413 return AF.hasAttribute(AttributeSet::FunctionIndex, 1414 Attribute::OptimizeForSize); 1415 } 1416 1417 inline static bool isMinSize(const MachineFunction &MF) { 1418 return MF.getFunction()->optForMinSize(); 1419 } 1420 1421 1422 /// Determine whether the callee-saved register saves and restores should 1423 /// be generated via inline code. If this function returns "true", inline 1424 /// code will be generated. If this function returns "false", additional 1425 /// checks are performed, which may still lead to the inline code. 1426 bool HexagonFrameLowering::shouldInlineCSR(MachineFunction &MF, 1427 const CSIVect &CSI) const { 1428 if (MF.getInfo<HexagonMachineFunctionInfo>()->hasEHReturn()) 1429 return true; 1430 if (!isOptSize(MF) && !isMinSize(MF)) 1431 if (MF.getTarget().getOptLevel() > CodeGenOpt::Default) 1432 return true; 1433 1434 // Check if CSI only has double registers, and if the registers form 1435 // a contiguous block starting from D8. 1436 BitVector Regs(Hexagon::NUM_TARGET_REGS); 1437 for (unsigned i = 0, n = CSI.size(); i < n; ++i) { 1438 unsigned R = CSI[i].getReg(); 1439 if (!Hexagon::DoubleRegsRegClass.contains(R)) 1440 return true; 1441 Regs[R] = true; 1442 } 1443 int F = Regs.find_first(); 1444 if (F != Hexagon::D8) 1445 return true; 1446 while (F >= 0) { 1447 int N = Regs.find_next(F); 1448 if (N >= 0 && N != F+1) 1449 return true; 1450 F = N; 1451 } 1452 1453 return false; 1454 } 1455 1456 1457 bool HexagonFrameLowering::useSpillFunction(MachineFunction &MF, 1458 const CSIVect &CSI) const { 1459 if (shouldInlineCSR(MF, CSI)) 1460 return false; 1461 unsigned NumCSI = CSI.size(); 1462 if (NumCSI <= 1) 1463 return false; 1464 1465 unsigned Threshold = isOptSize(MF) ? SpillFuncThresholdOs 1466 : SpillFuncThreshold; 1467 return Threshold < NumCSI; 1468 } 1469 1470 1471 bool HexagonFrameLowering::useRestoreFunction(MachineFunction &MF, 1472 const CSIVect &CSI) const { 1473 if (shouldInlineCSR(MF, CSI)) 1474 return false; 1475 unsigned NumCSI = CSI.size(); 1476 unsigned Threshold = isOptSize(MF) ? SpillFuncThresholdOs-1 1477 : SpillFuncThreshold; 1478 return Threshold < NumCSI; 1479 } 1480