ARMBaseInstrInfo.cpp revision 38ae997e63e3e1bb2c8679e01ea74cf8fd0be893
1//===- ARMBaseInstrInfo.cpp - ARM Instruction Information -------*- C++ -*-===// 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// This file contains the Base ARM implementation of the TargetInstrInfo class. 11// 12//===----------------------------------------------------------------------===// 13 14#include "ARMBaseInstrInfo.h" 15#include "ARM.h" 16#include "ARMAddressingModes.h" 17#include "ARMConstantPoolValue.h" 18#include "ARMMachineFunctionInfo.h" 19#include "ARMRegisterInfo.h" 20#include "ARMGenInstrInfo.inc" 21#include "llvm/Constants.h" 22#include "llvm/Function.h" 23#include "llvm/GlobalValue.h" 24#include "llvm/ADT/STLExtras.h" 25#include "llvm/CodeGen/LiveVariables.h" 26#include "llvm/CodeGen/MachineConstantPool.h" 27#include "llvm/CodeGen/MachineFrameInfo.h" 28#include "llvm/CodeGen/MachineInstrBuilder.h" 29#include "llvm/CodeGen/MachineJumpTableInfo.h" 30#include "llvm/CodeGen/MachineMemOperand.h" 31#include "llvm/CodeGen/MachineRegisterInfo.h" 32#include "llvm/CodeGen/PseudoSourceValue.h" 33#include "llvm/MC/MCAsmInfo.h" 34#include "llvm/Support/CommandLine.h" 35#include "llvm/Support/Debug.h" 36#include "llvm/Support/ErrorHandling.h" 37using namespace llvm; 38 39static cl::opt<bool> 40EnableARM3Addr("enable-arm-3-addr-conv", cl::Hidden, 41 cl::desc("Enable ARM 2-addr to 3-addr conv")); 42 43ARMBaseInstrInfo::ARMBaseInstrInfo(const ARMSubtarget& STI) 44 : TargetInstrInfoImpl(ARMInsts, array_lengthof(ARMInsts)), 45 Subtarget(STI) { 46} 47 48MachineInstr * 49ARMBaseInstrInfo::convertToThreeAddress(MachineFunction::iterator &MFI, 50 MachineBasicBlock::iterator &MBBI, 51 LiveVariables *LV) const { 52 // FIXME: Thumb2 support. 53 54 if (!EnableARM3Addr) 55 return NULL; 56 57 MachineInstr *MI = MBBI; 58 MachineFunction &MF = *MI->getParent()->getParent(); 59 uint64_t TSFlags = MI->getDesc().TSFlags; 60 bool isPre = false; 61 switch ((TSFlags & ARMII::IndexModeMask) >> ARMII::IndexModeShift) { 62 default: return NULL; 63 case ARMII::IndexModePre: 64 isPre = true; 65 break; 66 case ARMII::IndexModePost: 67 break; 68 } 69 70 // Try splitting an indexed load/store to an un-indexed one plus an add/sub 71 // operation. 72 unsigned MemOpc = getUnindexedOpcode(MI->getOpcode()); 73 if (MemOpc == 0) 74 return NULL; 75 76 MachineInstr *UpdateMI = NULL; 77 MachineInstr *MemMI = NULL; 78 unsigned AddrMode = (TSFlags & ARMII::AddrModeMask); 79 const TargetInstrDesc &TID = MI->getDesc(); 80 unsigned NumOps = TID.getNumOperands(); 81 bool isLoad = !TID.mayStore(); 82 const MachineOperand &WB = isLoad ? MI->getOperand(1) : MI->getOperand(0); 83 const MachineOperand &Base = MI->getOperand(2); 84 const MachineOperand &Offset = MI->getOperand(NumOps-3); 85 unsigned WBReg = WB.getReg(); 86 unsigned BaseReg = Base.getReg(); 87 unsigned OffReg = Offset.getReg(); 88 unsigned OffImm = MI->getOperand(NumOps-2).getImm(); 89 ARMCC::CondCodes Pred = (ARMCC::CondCodes)MI->getOperand(NumOps-1).getImm(); 90 switch (AddrMode) { 91 default: 92 assert(false && "Unknown indexed op!"); 93 return NULL; 94 case ARMII::AddrMode2: { 95 bool isSub = ARM_AM::getAM2Op(OffImm) == ARM_AM::sub; 96 unsigned Amt = ARM_AM::getAM2Offset(OffImm); 97 if (OffReg == 0) { 98 if (ARM_AM::getSOImmVal(Amt) == -1) 99 // Can't encode it in a so_imm operand. This transformation will 100 // add more than 1 instruction. Abandon! 101 return NULL; 102 UpdateMI = BuildMI(MF, MI->getDebugLoc(), 103 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg) 104 .addReg(BaseReg).addImm(Amt) 105 .addImm(Pred).addReg(0).addReg(0); 106 } else if (Amt != 0) { 107 ARM_AM::ShiftOpc ShOpc = ARM_AM::getAM2ShiftOpc(OffImm); 108 unsigned SOOpc = ARM_AM::getSORegOpc(ShOpc, Amt); 109 UpdateMI = BuildMI(MF, MI->getDebugLoc(), 110 get(isSub ? ARM::SUBrs : ARM::ADDrs), WBReg) 111 .addReg(BaseReg).addReg(OffReg).addReg(0).addImm(SOOpc) 112 .addImm(Pred).addReg(0).addReg(0); 113 } else 114 UpdateMI = BuildMI(MF, MI->getDebugLoc(), 115 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg) 116 .addReg(BaseReg).addReg(OffReg) 117 .addImm(Pred).addReg(0).addReg(0); 118 break; 119 } 120 case ARMII::AddrMode3 : { 121 bool isSub = ARM_AM::getAM3Op(OffImm) == ARM_AM::sub; 122 unsigned Amt = ARM_AM::getAM3Offset(OffImm); 123 if (OffReg == 0) 124 // Immediate is 8-bits. It's guaranteed to fit in a so_imm operand. 125 UpdateMI = BuildMI(MF, MI->getDebugLoc(), 126 get(isSub ? ARM::SUBri : ARM::ADDri), WBReg) 127 .addReg(BaseReg).addImm(Amt) 128 .addImm(Pred).addReg(0).addReg(0); 129 else 130 UpdateMI = BuildMI(MF, MI->getDebugLoc(), 131 get(isSub ? ARM::SUBrr : ARM::ADDrr), WBReg) 132 .addReg(BaseReg).addReg(OffReg) 133 .addImm(Pred).addReg(0).addReg(0); 134 break; 135 } 136 } 137 138 std::vector<MachineInstr*> NewMIs; 139 if (isPre) { 140 if (isLoad) 141 MemMI = BuildMI(MF, MI->getDebugLoc(), 142 get(MemOpc), MI->getOperand(0).getReg()) 143 .addReg(WBReg).addReg(0).addImm(0).addImm(Pred); 144 else 145 MemMI = BuildMI(MF, MI->getDebugLoc(), 146 get(MemOpc)).addReg(MI->getOperand(1).getReg()) 147 .addReg(WBReg).addReg(0).addImm(0).addImm(Pred); 148 NewMIs.push_back(MemMI); 149 NewMIs.push_back(UpdateMI); 150 } else { 151 if (isLoad) 152 MemMI = BuildMI(MF, MI->getDebugLoc(), 153 get(MemOpc), MI->getOperand(0).getReg()) 154 .addReg(BaseReg).addReg(0).addImm(0).addImm(Pred); 155 else 156 MemMI = BuildMI(MF, MI->getDebugLoc(), 157 get(MemOpc)).addReg(MI->getOperand(1).getReg()) 158 .addReg(BaseReg).addReg(0).addImm(0).addImm(Pred); 159 if (WB.isDead()) 160 UpdateMI->getOperand(0).setIsDead(); 161 NewMIs.push_back(UpdateMI); 162 NewMIs.push_back(MemMI); 163 } 164 165 // Transfer LiveVariables states, kill / dead info. 166 if (LV) { 167 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 168 MachineOperand &MO = MI->getOperand(i); 169 if (MO.isReg() && MO.getReg() && 170 TargetRegisterInfo::isVirtualRegister(MO.getReg())) { 171 unsigned Reg = MO.getReg(); 172 173 LiveVariables::VarInfo &VI = LV->getVarInfo(Reg); 174 if (MO.isDef()) { 175 MachineInstr *NewMI = (Reg == WBReg) ? UpdateMI : MemMI; 176 if (MO.isDead()) 177 LV->addVirtualRegisterDead(Reg, NewMI); 178 } 179 if (MO.isUse() && MO.isKill()) { 180 for (unsigned j = 0; j < 2; ++j) { 181 // Look at the two new MI's in reverse order. 182 MachineInstr *NewMI = NewMIs[j]; 183 if (!NewMI->readsRegister(Reg)) 184 continue; 185 LV->addVirtualRegisterKilled(Reg, NewMI); 186 if (VI.removeKill(MI)) 187 VI.Kills.push_back(NewMI); 188 break; 189 } 190 } 191 } 192 } 193 } 194 195 MFI->insert(MBBI, NewMIs[1]); 196 MFI->insert(MBBI, NewMIs[0]); 197 return NewMIs[0]; 198} 199 200bool 201ARMBaseInstrInfo::spillCalleeSavedRegisters(MachineBasicBlock &MBB, 202 MachineBasicBlock::iterator MI, 203 const std::vector<CalleeSavedInfo> &CSI, 204 const TargetRegisterInfo *TRI) const { 205 if (CSI.empty()) 206 return false; 207 208 DebugLoc DL; 209 if (MI != MBB.end()) DL = MI->getDebugLoc(); 210 211 for (unsigned i = 0, e = CSI.size(); i != e; ++i) { 212 unsigned Reg = CSI[i].getReg(); 213 bool isKill = true; 214 215 // Add the callee-saved register as live-in unless it's LR and 216 // @llvm.returnaddress is called. If LR is returned for @llvm.returnaddress 217 // then it's already added to the function and entry block live-in sets. 218 if (Reg == ARM::LR) { 219 MachineFunction &MF = *MBB.getParent(); 220 if (MF.getFrameInfo()->isReturnAddressTaken() && 221 MF.getRegInfo().isLiveIn(Reg)) 222 isKill = false; 223 } 224 225 if (isKill) 226 MBB.addLiveIn(Reg); 227 228 // Insert the spill to the stack frame. The register is killed at the spill 229 // 230 const TargetRegisterClass *RC = TRI->getMinimalPhysRegClass(Reg); 231 storeRegToStackSlot(MBB, MI, Reg, isKill, 232 CSI[i].getFrameIdx(), RC, TRI); 233 } 234 return true; 235} 236 237// Branch analysis. 238bool 239ARMBaseInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB,MachineBasicBlock *&TBB, 240 MachineBasicBlock *&FBB, 241 SmallVectorImpl<MachineOperand> &Cond, 242 bool AllowModify) const { 243 // If the block has no terminators, it just falls into the block after it. 244 MachineBasicBlock::iterator I = MBB.end(); 245 if (I == MBB.begin()) 246 return false; 247 --I; 248 while (I->isDebugValue()) { 249 if (I == MBB.begin()) 250 return false; 251 --I; 252 } 253 if (!isUnpredicatedTerminator(I)) 254 return false; 255 256 // Get the last instruction in the block. 257 MachineInstr *LastInst = I; 258 259 // If there is only one terminator instruction, process it. 260 unsigned LastOpc = LastInst->getOpcode(); 261 if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) { 262 if (isUncondBranchOpcode(LastOpc)) { 263 TBB = LastInst->getOperand(0).getMBB(); 264 return false; 265 } 266 if (isCondBranchOpcode(LastOpc)) { 267 // Block ends with fall-through condbranch. 268 TBB = LastInst->getOperand(0).getMBB(); 269 Cond.push_back(LastInst->getOperand(1)); 270 Cond.push_back(LastInst->getOperand(2)); 271 return false; 272 } 273 return true; // Can't handle indirect branch. 274 } 275 276 // Get the instruction before it if it is a terminator. 277 MachineInstr *SecondLastInst = I; 278 279 // If there are three terminators, we don't know what sort of block this is. 280 if (SecondLastInst && I != MBB.begin() && isUnpredicatedTerminator(--I)) 281 return true; 282 283 // If the block ends with a B and a Bcc, handle it. 284 unsigned SecondLastOpc = SecondLastInst->getOpcode(); 285 if (isCondBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) { 286 TBB = SecondLastInst->getOperand(0).getMBB(); 287 Cond.push_back(SecondLastInst->getOperand(1)); 288 Cond.push_back(SecondLastInst->getOperand(2)); 289 FBB = LastInst->getOperand(0).getMBB(); 290 return false; 291 } 292 293 // If the block ends with two unconditional branches, handle it. The second 294 // one is not executed, so remove it. 295 if (isUncondBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) { 296 TBB = SecondLastInst->getOperand(0).getMBB(); 297 I = LastInst; 298 if (AllowModify) 299 I->eraseFromParent(); 300 return false; 301 } 302 303 // ...likewise if it ends with a branch table followed by an unconditional 304 // branch. The branch folder can create these, and we must get rid of them for 305 // correctness of Thumb constant islands. 306 if ((isJumpTableBranchOpcode(SecondLastOpc) || 307 isIndirectBranchOpcode(SecondLastOpc)) && 308 isUncondBranchOpcode(LastOpc)) { 309 I = LastInst; 310 if (AllowModify) 311 I->eraseFromParent(); 312 return true; 313 } 314 315 // Otherwise, can't handle this. 316 return true; 317} 318 319 320unsigned ARMBaseInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const { 321 MachineBasicBlock::iterator I = MBB.end(); 322 if (I == MBB.begin()) return 0; 323 --I; 324 while (I->isDebugValue()) { 325 if (I == MBB.begin()) 326 return 0; 327 --I; 328 } 329 if (!isUncondBranchOpcode(I->getOpcode()) && 330 !isCondBranchOpcode(I->getOpcode())) 331 return 0; 332 333 // Remove the branch. 334 I->eraseFromParent(); 335 336 I = MBB.end(); 337 338 if (I == MBB.begin()) return 1; 339 --I; 340 if (!isCondBranchOpcode(I->getOpcode())) 341 return 1; 342 343 // Remove the branch. 344 I->eraseFromParent(); 345 return 2; 346} 347 348unsigned 349ARMBaseInstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB, 350 MachineBasicBlock *FBB, 351 const SmallVectorImpl<MachineOperand> &Cond, 352 DebugLoc DL) const { 353 ARMFunctionInfo *AFI = MBB.getParent()->getInfo<ARMFunctionInfo>(); 354 int BOpc = !AFI->isThumbFunction() 355 ? ARM::B : (AFI->isThumb2Function() ? ARM::t2B : ARM::tB); 356 int BccOpc = !AFI->isThumbFunction() 357 ? ARM::Bcc : (AFI->isThumb2Function() ? ARM::t2Bcc : ARM::tBcc); 358 359 // Shouldn't be a fall through. 360 assert(TBB && "InsertBranch must not be told to insert a fallthrough"); 361 assert((Cond.size() == 2 || Cond.size() == 0) && 362 "ARM branch conditions have two components!"); 363 364 if (FBB == 0) { 365 if (Cond.empty()) // Unconditional branch? 366 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB); 367 else 368 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB) 369 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg()); 370 return 1; 371 } 372 373 // Two-way conditional branch. 374 BuildMI(&MBB, DL, get(BccOpc)).addMBB(TBB) 375 .addImm(Cond[0].getImm()).addReg(Cond[1].getReg()); 376 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB); 377 return 2; 378} 379 380bool ARMBaseInstrInfo:: 381ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const { 382 ARMCC::CondCodes CC = (ARMCC::CondCodes)(int)Cond[0].getImm(); 383 Cond[0].setImm(ARMCC::getOppositeCondition(CC)); 384 return false; 385} 386 387bool ARMBaseInstrInfo:: 388PredicateInstruction(MachineInstr *MI, 389 const SmallVectorImpl<MachineOperand> &Pred) const { 390 unsigned Opc = MI->getOpcode(); 391 if (isUncondBranchOpcode(Opc)) { 392 MI->setDesc(get(getMatchingCondBranchOpcode(Opc))); 393 MI->addOperand(MachineOperand::CreateImm(Pred[0].getImm())); 394 MI->addOperand(MachineOperand::CreateReg(Pred[1].getReg(), false)); 395 return true; 396 } 397 398 int PIdx = MI->findFirstPredOperandIdx(); 399 if (PIdx != -1) { 400 MachineOperand &PMO = MI->getOperand(PIdx); 401 PMO.setImm(Pred[0].getImm()); 402 MI->getOperand(PIdx+1).setReg(Pred[1].getReg()); 403 return true; 404 } 405 return false; 406} 407 408bool ARMBaseInstrInfo:: 409SubsumesPredicate(const SmallVectorImpl<MachineOperand> &Pred1, 410 const SmallVectorImpl<MachineOperand> &Pred2) const { 411 if (Pred1.size() > 2 || Pred2.size() > 2) 412 return false; 413 414 ARMCC::CondCodes CC1 = (ARMCC::CondCodes)Pred1[0].getImm(); 415 ARMCC::CondCodes CC2 = (ARMCC::CondCodes)Pred2[0].getImm(); 416 if (CC1 == CC2) 417 return true; 418 419 switch (CC1) { 420 default: 421 return false; 422 case ARMCC::AL: 423 return true; 424 case ARMCC::HS: 425 return CC2 == ARMCC::HI; 426 case ARMCC::LS: 427 return CC2 == ARMCC::LO || CC2 == ARMCC::EQ; 428 case ARMCC::GE: 429 return CC2 == ARMCC::GT; 430 case ARMCC::LE: 431 return CC2 == ARMCC::LT; 432 } 433} 434 435bool ARMBaseInstrInfo::DefinesPredicate(MachineInstr *MI, 436 std::vector<MachineOperand> &Pred) const { 437 // FIXME: This confuses implicit_def with optional CPSR def. 438 const TargetInstrDesc &TID = MI->getDesc(); 439 if (!TID.getImplicitDefs() && !TID.hasOptionalDef()) 440 return false; 441 442 bool Found = false; 443 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 444 const MachineOperand &MO = MI->getOperand(i); 445 if (MO.isReg() && MO.getReg() == ARM::CPSR) { 446 Pred.push_back(MO); 447 Found = true; 448 } 449 } 450 451 return Found; 452} 453 454/// isPredicable - Return true if the specified instruction can be predicated. 455/// By default, this returns true for every instruction with a 456/// PredicateOperand. 457bool ARMBaseInstrInfo::isPredicable(MachineInstr *MI) const { 458 const TargetInstrDesc &TID = MI->getDesc(); 459 if (!TID.isPredicable()) 460 return false; 461 462 if ((TID.TSFlags & ARMII::DomainMask) == ARMII::DomainNEON) { 463 ARMFunctionInfo *AFI = 464 MI->getParent()->getParent()->getInfo<ARMFunctionInfo>(); 465 return AFI->isThumb2Function(); 466 } 467 return true; 468} 469 470/// FIXME: Works around a gcc miscompilation with -fstrict-aliasing. 471DISABLE_INLINE 472static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT, 473 unsigned JTI); 474static unsigned getNumJTEntries(const std::vector<MachineJumpTableEntry> &JT, 475 unsigned JTI) { 476 assert(JTI < JT.size()); 477 return JT[JTI].MBBs.size(); 478} 479 480/// GetInstSize - Return the size of the specified MachineInstr. 481/// 482unsigned ARMBaseInstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const { 483 const MachineBasicBlock &MBB = *MI->getParent(); 484 const MachineFunction *MF = MBB.getParent(); 485 const MCAsmInfo *MAI = MF->getTarget().getMCAsmInfo(); 486 487 // Basic size info comes from the TSFlags field. 488 const TargetInstrDesc &TID = MI->getDesc(); 489 uint64_t TSFlags = TID.TSFlags; 490 491 unsigned Opc = MI->getOpcode(); 492 switch ((TSFlags & ARMII::SizeMask) >> ARMII::SizeShift) { 493 default: { 494 // If this machine instr is an inline asm, measure it. 495 if (MI->getOpcode() == ARM::INLINEASM) 496 return getInlineAsmLength(MI->getOperand(0).getSymbolName(), *MAI); 497 if (MI->isLabel()) 498 return 0; 499 switch (Opc) { 500 default: 501 llvm_unreachable("Unknown or unset size field for instr!"); 502 case TargetOpcode::IMPLICIT_DEF: 503 case TargetOpcode::KILL: 504 case TargetOpcode::PROLOG_LABEL: 505 case TargetOpcode::EH_LABEL: 506 case TargetOpcode::DBG_VALUE: 507 return 0; 508 } 509 break; 510 } 511 case ARMII::Size8Bytes: return 8; // ARM instruction x 2. 512 case ARMII::Size4Bytes: return 4; // ARM / Thumb2 instruction. 513 case ARMII::Size2Bytes: return 2; // Thumb1 instruction. 514 case ARMII::SizeSpecial: { 515 switch (Opc) { 516 case ARM::CONSTPOOL_ENTRY: 517 // If this machine instr is a constant pool entry, its size is recorded as 518 // operand #2. 519 return MI->getOperand(2).getImm(); 520 case ARM::Int_eh_sjlj_longjmp: 521 return 16; 522 case ARM::tInt_eh_sjlj_longjmp: 523 return 10; 524 case ARM::Int_eh_sjlj_setjmp: 525 case ARM::Int_eh_sjlj_setjmp_nofp: 526 return 20; 527 case ARM::tInt_eh_sjlj_setjmp: 528 case ARM::t2Int_eh_sjlj_setjmp: 529 case ARM::t2Int_eh_sjlj_setjmp_nofp: 530 return 12; 531 case ARM::BR_JTr: 532 case ARM::BR_JTm: 533 case ARM::BR_JTadd: 534 case ARM::tBR_JTr: 535 case ARM::t2BR_JT: 536 case ARM::t2TBB: 537 case ARM::t2TBH: { 538 // These are jumptable branches, i.e. a branch followed by an inlined 539 // jumptable. The size is 4 + 4 * number of entries. For TBB, each 540 // entry is one byte; TBH two byte each. 541 unsigned EntrySize = (Opc == ARM::t2TBB) 542 ? 1 : ((Opc == ARM::t2TBH) ? 2 : 4); 543 unsigned NumOps = TID.getNumOperands(); 544 MachineOperand JTOP = 545 MI->getOperand(NumOps - (TID.isPredicable() ? 3 : 2)); 546 unsigned JTI = JTOP.getIndex(); 547 const MachineJumpTableInfo *MJTI = MF->getJumpTableInfo(); 548 assert(MJTI != 0); 549 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables(); 550 assert(JTI < JT.size()); 551 // Thumb instructions are 2 byte aligned, but JT entries are 4 byte 552 // 4 aligned. The assembler / linker may add 2 byte padding just before 553 // the JT entries. The size does not include this padding; the 554 // constant islands pass does separate bookkeeping for it. 555 // FIXME: If we know the size of the function is less than (1 << 16) *2 556 // bytes, we can use 16-bit entries instead. Then there won't be an 557 // alignment issue. 558 unsigned InstSize = (Opc == ARM::tBR_JTr || Opc == ARM::t2BR_JT) ? 2 : 4; 559 unsigned NumEntries = getNumJTEntries(JT, JTI); 560 if (Opc == ARM::t2TBB && (NumEntries & 1)) 561 // Make sure the instruction that follows TBB is 2-byte aligned. 562 // FIXME: Constant island pass should insert an "ALIGN" instruction 563 // instead. 564 ++NumEntries; 565 return NumEntries * EntrySize + InstSize; 566 } 567 default: 568 // Otherwise, pseudo-instruction sizes are zero. 569 return 0; 570 } 571 } 572 } 573 return 0; // Not reached 574} 575 576unsigned 577ARMBaseInstrInfo::isLoadFromStackSlot(const MachineInstr *MI, 578 int &FrameIndex) const { 579 switch (MI->getOpcode()) { 580 default: break; 581 case ARM::LDR: 582 case ARM::t2LDRs: // FIXME: don't use t2LDRs to access frame. 583 if (MI->getOperand(1).isFI() && 584 MI->getOperand(2).isReg() && 585 MI->getOperand(3).isImm() && 586 MI->getOperand(2).getReg() == 0 && 587 MI->getOperand(3).getImm() == 0) { 588 FrameIndex = MI->getOperand(1).getIndex(); 589 return MI->getOperand(0).getReg(); 590 } 591 break; 592 case ARM::t2LDRi12: 593 case ARM::tRestore: 594 if (MI->getOperand(1).isFI() && 595 MI->getOperand(2).isImm() && 596 MI->getOperand(2).getImm() == 0) { 597 FrameIndex = MI->getOperand(1).getIndex(); 598 return MI->getOperand(0).getReg(); 599 } 600 break; 601 case ARM::VLDRD: 602 case ARM::VLDRS: 603 if (MI->getOperand(1).isFI() && 604 MI->getOperand(2).isImm() && 605 MI->getOperand(2).getImm() == 0) { 606 FrameIndex = MI->getOperand(1).getIndex(); 607 return MI->getOperand(0).getReg(); 608 } 609 break; 610 } 611 612 return 0; 613} 614 615unsigned 616ARMBaseInstrInfo::isStoreToStackSlot(const MachineInstr *MI, 617 int &FrameIndex) const { 618 switch (MI->getOpcode()) { 619 default: break; 620 case ARM::STR: 621 case ARM::t2STRs: // FIXME: don't use t2STRs to access frame. 622 if (MI->getOperand(1).isFI() && 623 MI->getOperand(2).isReg() && 624 MI->getOperand(3).isImm() && 625 MI->getOperand(2).getReg() == 0 && 626 MI->getOperand(3).getImm() == 0) { 627 FrameIndex = MI->getOperand(1).getIndex(); 628 return MI->getOperand(0).getReg(); 629 } 630 break; 631 case ARM::t2STRi12: 632 case ARM::tSpill: 633 if (MI->getOperand(1).isFI() && 634 MI->getOperand(2).isImm() && 635 MI->getOperand(2).getImm() == 0) { 636 FrameIndex = MI->getOperand(1).getIndex(); 637 return MI->getOperand(0).getReg(); 638 } 639 break; 640 case ARM::VSTRD: 641 case ARM::VSTRS: 642 if (MI->getOperand(1).isFI() && 643 MI->getOperand(2).isImm() && 644 MI->getOperand(2).getImm() == 0) { 645 FrameIndex = MI->getOperand(1).getIndex(); 646 return MI->getOperand(0).getReg(); 647 } 648 break; 649 } 650 651 return 0; 652} 653 654void ARMBaseInstrInfo::copyPhysReg(MachineBasicBlock &MBB, 655 MachineBasicBlock::iterator I, DebugLoc DL, 656 unsigned DestReg, unsigned SrcReg, 657 bool KillSrc) const { 658 bool GPRDest = ARM::GPRRegClass.contains(DestReg); 659 bool GPRSrc = ARM::GPRRegClass.contains(SrcReg); 660 661 if (GPRDest && GPRSrc) { 662 AddDefaultCC(AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::MOVr), DestReg) 663 .addReg(SrcReg, getKillRegState(KillSrc)))); 664 return; 665 } 666 667 bool SPRDest = ARM::SPRRegClass.contains(DestReg); 668 bool SPRSrc = ARM::SPRRegClass.contains(SrcReg); 669 670 unsigned Opc; 671 if (SPRDest && SPRSrc) 672 Opc = ARM::VMOVS; 673 else if (GPRDest && SPRSrc) 674 Opc = ARM::VMOVRS; 675 else if (SPRDest && GPRSrc) 676 Opc = ARM::VMOVSR; 677 else if (ARM::DPRRegClass.contains(DestReg, SrcReg)) 678 Opc = ARM::VMOVD; 679 else if (ARM::QPRRegClass.contains(DestReg, SrcReg)) 680 Opc = ARM::VMOVQ; 681 else if (ARM::QQPRRegClass.contains(DestReg, SrcReg)) 682 Opc = ARM::VMOVQQ; 683 else if (ARM::QQQQPRRegClass.contains(DestReg, SrcReg)) 684 Opc = ARM::VMOVQQQQ; 685 else 686 llvm_unreachable("Impossible reg-to-reg copy"); 687 688 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(Opc), DestReg); 689 MIB.addReg(SrcReg, getKillRegState(KillSrc)); 690 if (Opc != ARM::VMOVQQ && Opc != ARM::VMOVQQQQ) 691 AddDefaultPred(MIB); 692} 693 694static const 695MachineInstrBuilder &AddDReg(MachineInstrBuilder &MIB, 696 unsigned Reg, unsigned SubIdx, unsigned State, 697 const TargetRegisterInfo *TRI) { 698 if (!SubIdx) 699 return MIB.addReg(Reg, State); 700 701 if (TargetRegisterInfo::isPhysicalRegister(Reg)) 702 return MIB.addReg(TRI->getSubReg(Reg, SubIdx), State); 703 return MIB.addReg(Reg, State, SubIdx); 704} 705 706void ARMBaseInstrInfo:: 707storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I, 708 unsigned SrcReg, bool isKill, int FI, 709 const TargetRegisterClass *RC, 710 const TargetRegisterInfo *TRI) const { 711 DebugLoc DL; 712 if (I != MBB.end()) DL = I->getDebugLoc(); 713 MachineFunction &MF = *MBB.getParent(); 714 MachineFrameInfo &MFI = *MF.getFrameInfo(); 715 unsigned Align = MFI.getObjectAlignment(FI); 716 717 MachineMemOperand *MMO = 718 MF.getMachineMemOperand(PseudoSourceValue::getFixedStack(FI), 719 MachineMemOperand::MOStore, 0, 720 MFI.getObjectSize(FI), 721 Align); 722 723 // tGPR is used sometimes in ARM instructions that need to avoid using 724 // certain registers. Just treat it as GPR here. Likewise, rGPR. 725 if (RC == ARM::tGPRRegisterClass || RC == ARM::tcGPRRegisterClass 726 || RC == ARM::rGPRRegisterClass) 727 RC = ARM::GPRRegisterClass; 728 729 switch (RC->getID()) { 730 case ARM::GPRRegClassID: 731 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::STR)) 732 .addReg(SrcReg, getKillRegState(isKill)) 733 .addFrameIndex(FI).addReg(0).addImm(0).addMemOperand(MMO)); 734 break; 735 case ARM::SPRRegClassID: 736 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRS)) 737 .addReg(SrcReg, getKillRegState(isKill)) 738 .addFrameIndex(FI).addImm(0).addMemOperand(MMO)); 739 break; 740 case ARM::DPRRegClassID: 741 case ARM::DPR_VFP2RegClassID: 742 case ARM::DPR_8RegClassID: 743 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTRD)) 744 .addReg(SrcReg, getKillRegState(isKill)) 745 .addFrameIndex(FI).addImm(0).addMemOperand(MMO)); 746 break; 747 case ARM::QPRRegClassID: 748 case ARM::QPR_VFP2RegClassID: 749 case ARM::QPR_8RegClassID: 750 // FIXME: Neon instructions should support predicates 751 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) { 752 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1q)) 753 .addFrameIndex(FI).addImm(16) 754 .addReg(SrcReg, getKillRegState(isKill)) 755 .addMemOperand(MMO)); 756 } else { 757 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMQ)) 758 .addReg(SrcReg, getKillRegState(isKill)) 759 .addFrameIndex(FI) 760 .addImm(ARM_AM::getAM5Opc(ARM_AM::ia, 4)) 761 .addMemOperand(MMO)); 762 } 763 break; 764 case ARM::QQPRRegClassID: 765 case ARM::QQPR_VFP2RegClassID: 766 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) { 767 // FIXME: It's possible to only store part of the QQ register if the 768 // spilled def has a sub-register index. 769 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(ARM::VST1d64Q)) 770 .addFrameIndex(FI).addImm(16); 771 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI); 772 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI); 773 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI); 774 MIB = AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI); 775 AddDefaultPred(MIB.addMemOperand(MMO)); 776 } else { 777 MachineInstrBuilder MIB = 778 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMD)) 779 .addFrameIndex(FI) 780 .addImm(ARM_AM::getAM5Opc(ARM_AM::ia, 4))) 781 .addMemOperand(MMO); 782 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI); 783 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI); 784 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI); 785 AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI); 786 } 787 break; 788 case ARM::QQQQPRRegClassID: { 789 MachineInstrBuilder MIB = 790 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMD)) 791 .addFrameIndex(FI) 792 .addImm(ARM_AM::getAM5Opc(ARM_AM::ia, 4))) 793 .addMemOperand(MMO); 794 MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI); 795 MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI); 796 MIB = AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI); 797 MIB = AddDReg(MIB, SrcReg, ARM::dsub_3, 0, TRI); 798 MIB = AddDReg(MIB, SrcReg, ARM::dsub_4, 0, TRI); 799 MIB = AddDReg(MIB, SrcReg, ARM::dsub_5, 0, TRI); 800 MIB = AddDReg(MIB, SrcReg, ARM::dsub_6, 0, TRI); 801 AddDReg(MIB, SrcReg, ARM::dsub_7, 0, TRI); 802 break; 803 } 804 default: 805 llvm_unreachable("Unknown regclass!"); 806 } 807} 808 809void ARMBaseInstrInfo:: 810loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I, 811 unsigned DestReg, int FI, 812 const TargetRegisterClass *RC, 813 const TargetRegisterInfo *TRI) const { 814 DebugLoc DL; 815 if (I != MBB.end()) DL = I->getDebugLoc(); 816 MachineFunction &MF = *MBB.getParent(); 817 MachineFrameInfo &MFI = *MF.getFrameInfo(); 818 unsigned Align = MFI.getObjectAlignment(FI); 819 MachineMemOperand *MMO = 820 MF.getMachineMemOperand(PseudoSourceValue::getFixedStack(FI), 821 MachineMemOperand::MOLoad, 0, 822 MFI.getObjectSize(FI), 823 Align); 824 825 // tGPR is used sometimes in ARM instructions that need to avoid using 826 // certain registers. Just treat it as GPR here. 827 if (RC == ARM::tGPRRegisterClass || RC == ARM::tcGPRRegisterClass 828 || RC == ARM::rGPRRegisterClass) 829 RC = ARM::GPRRegisterClass; 830 831 switch (RC->getID()) { 832 case ARM::GPRRegClassID: 833 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::LDR), DestReg) 834 .addFrameIndex(FI).addReg(0).addImm(0).addMemOperand(MMO)); 835 break; 836 case ARM::SPRRegClassID: 837 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRS), DestReg) 838 .addFrameIndex(FI).addImm(0).addMemOperand(MMO)); 839 break; 840 case ARM::DPRRegClassID: 841 case ARM::DPR_VFP2RegClassID: 842 case ARM::DPR_8RegClassID: 843 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDRD), DestReg) 844 .addFrameIndex(FI).addImm(0).addMemOperand(MMO)); 845 break; 846 case ARM::QPRRegClassID: 847 case ARM::QPR_VFP2RegClassID: 848 case ARM::QPR_8RegClassID: 849 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) { 850 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1q), DestReg) 851 .addFrameIndex(FI).addImm(16) 852 .addMemOperand(MMO)); 853 } else { 854 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMQ), DestReg) 855 .addFrameIndex(FI) 856 .addImm(ARM_AM::getAM5Opc(ARM_AM::ia, 4)) 857 .addMemOperand(MMO)); 858 } 859 break; 860 case ARM::QQPRRegClassID: 861 case ARM::QQPR_VFP2RegClassID: 862 if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) { 863 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(ARM::VLD1d64Q)); 864 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::Define, TRI); 865 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::Define, TRI); 866 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::Define, TRI); 867 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::Define, TRI); 868 AddDefaultPred(MIB.addFrameIndex(FI).addImm(16).addMemOperand(MMO)); 869 } else { 870 MachineInstrBuilder MIB = 871 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMD)) 872 .addFrameIndex(FI) 873 .addImm(ARM_AM::getAM5Opc(ARM_AM::ia, 4))) 874 .addMemOperand(MMO); 875 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::Define, TRI); 876 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::Define, TRI); 877 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::Define, TRI); 878 AddDReg(MIB, DestReg, ARM::dsub_3, RegState::Define, TRI); 879 } 880 break; 881 case ARM::QQQQPRRegClassID: { 882 MachineInstrBuilder MIB = 883 AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMD)) 884 .addFrameIndex(FI) 885 .addImm(ARM_AM::getAM5Opc(ARM_AM::ia, 4))) 886 .addMemOperand(MMO); 887 MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::Define, TRI); 888 MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::Define, TRI); 889 MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::Define, TRI); 890 MIB = AddDReg(MIB, DestReg, ARM::dsub_3, RegState::Define, TRI); 891 MIB = AddDReg(MIB, DestReg, ARM::dsub_4, RegState::Define, TRI); 892 MIB = AddDReg(MIB, DestReg, ARM::dsub_5, RegState::Define, TRI); 893 MIB = AddDReg(MIB, DestReg, ARM::dsub_6, RegState::Define, TRI); 894 AddDReg(MIB, DestReg, ARM::dsub_7, RegState::Define, TRI); 895 break; 896 } 897 default: 898 llvm_unreachable("Unknown regclass!"); 899 } 900} 901 902MachineInstr* 903ARMBaseInstrInfo::emitFrameIndexDebugValue(MachineFunction &MF, 904 int FrameIx, uint64_t Offset, 905 const MDNode *MDPtr, 906 DebugLoc DL) const { 907 MachineInstrBuilder MIB = BuildMI(MF, DL, get(ARM::DBG_VALUE)) 908 .addFrameIndex(FrameIx).addImm(0).addImm(Offset).addMetadata(MDPtr); 909 return &*MIB; 910} 911 912/// Create a copy of a const pool value. Update CPI to the new index and return 913/// the label UID. 914static unsigned duplicateCPV(MachineFunction &MF, unsigned &CPI) { 915 MachineConstantPool *MCP = MF.getConstantPool(); 916 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 917 918 const MachineConstantPoolEntry &MCPE = MCP->getConstants()[CPI]; 919 assert(MCPE.isMachineConstantPoolEntry() && 920 "Expecting a machine constantpool entry!"); 921 ARMConstantPoolValue *ACPV = 922 static_cast<ARMConstantPoolValue*>(MCPE.Val.MachineCPVal); 923 924 unsigned PCLabelId = AFI->createConstPoolEntryUId(); 925 ARMConstantPoolValue *NewCPV = 0; 926 if (ACPV->isGlobalValue()) 927 NewCPV = new ARMConstantPoolValue(ACPV->getGV(), PCLabelId, 928 ARMCP::CPValue, 4); 929 else if (ACPV->isExtSymbol()) 930 NewCPV = new ARMConstantPoolValue(MF.getFunction()->getContext(), 931 ACPV->getSymbol(), PCLabelId, 4); 932 else if (ACPV->isBlockAddress()) 933 NewCPV = new ARMConstantPoolValue(ACPV->getBlockAddress(), PCLabelId, 934 ARMCP::CPBlockAddress, 4); 935 else 936 llvm_unreachable("Unexpected ARM constantpool value type!!"); 937 CPI = MCP->getConstantPoolIndex(NewCPV, MCPE.getAlignment()); 938 return PCLabelId; 939} 940 941void ARMBaseInstrInfo:: 942reMaterialize(MachineBasicBlock &MBB, 943 MachineBasicBlock::iterator I, 944 unsigned DestReg, unsigned SubIdx, 945 const MachineInstr *Orig, 946 const TargetRegisterInfo &TRI) const { 947 unsigned Opcode = Orig->getOpcode(); 948 switch (Opcode) { 949 default: { 950 MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig); 951 MI->substituteRegister(Orig->getOperand(0).getReg(), DestReg, SubIdx, TRI); 952 MBB.insert(I, MI); 953 break; 954 } 955 case ARM::tLDRpci_pic: 956 case ARM::t2LDRpci_pic: { 957 MachineFunction &MF = *MBB.getParent(); 958 unsigned CPI = Orig->getOperand(1).getIndex(); 959 unsigned PCLabelId = duplicateCPV(MF, CPI); 960 MachineInstrBuilder MIB = BuildMI(MBB, I, Orig->getDebugLoc(), get(Opcode), 961 DestReg) 962 .addConstantPoolIndex(CPI).addImm(PCLabelId); 963 (*MIB).setMemRefs(Orig->memoperands_begin(), Orig->memoperands_end()); 964 break; 965 } 966 } 967} 968 969MachineInstr * 970ARMBaseInstrInfo::duplicate(MachineInstr *Orig, MachineFunction &MF) const { 971 MachineInstr *MI = TargetInstrInfoImpl::duplicate(Orig, MF); 972 switch(Orig->getOpcode()) { 973 case ARM::tLDRpci_pic: 974 case ARM::t2LDRpci_pic: { 975 unsigned CPI = Orig->getOperand(1).getIndex(); 976 unsigned PCLabelId = duplicateCPV(MF, CPI); 977 Orig->getOperand(1).setIndex(CPI); 978 Orig->getOperand(2).setImm(PCLabelId); 979 break; 980 } 981 } 982 return MI; 983} 984 985bool ARMBaseInstrInfo::produceSameValue(const MachineInstr *MI0, 986 const MachineInstr *MI1) const { 987 int Opcode = MI0->getOpcode(); 988 if (Opcode == ARM::t2LDRpci || 989 Opcode == ARM::t2LDRpci_pic || 990 Opcode == ARM::tLDRpci || 991 Opcode == ARM::tLDRpci_pic) { 992 if (MI1->getOpcode() != Opcode) 993 return false; 994 if (MI0->getNumOperands() != MI1->getNumOperands()) 995 return false; 996 997 const MachineOperand &MO0 = MI0->getOperand(1); 998 const MachineOperand &MO1 = MI1->getOperand(1); 999 if (MO0.getOffset() != MO1.getOffset()) 1000 return false; 1001 1002 const MachineFunction *MF = MI0->getParent()->getParent(); 1003 const MachineConstantPool *MCP = MF->getConstantPool(); 1004 int CPI0 = MO0.getIndex(); 1005 int CPI1 = MO1.getIndex(); 1006 const MachineConstantPoolEntry &MCPE0 = MCP->getConstants()[CPI0]; 1007 const MachineConstantPoolEntry &MCPE1 = MCP->getConstants()[CPI1]; 1008 ARMConstantPoolValue *ACPV0 = 1009 static_cast<ARMConstantPoolValue*>(MCPE0.Val.MachineCPVal); 1010 ARMConstantPoolValue *ACPV1 = 1011 static_cast<ARMConstantPoolValue*>(MCPE1.Val.MachineCPVal); 1012 return ACPV0->hasSameValue(ACPV1); 1013 } 1014 1015 return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs); 1016} 1017 1018/// areLoadsFromSameBasePtr - This is used by the pre-regalloc scheduler to 1019/// determine if two loads are loading from the same base address. It should 1020/// only return true if the base pointers are the same and the only differences 1021/// between the two addresses is the offset. It also returns the offsets by 1022/// reference. 1023bool ARMBaseInstrInfo::areLoadsFromSameBasePtr(SDNode *Load1, SDNode *Load2, 1024 int64_t &Offset1, 1025 int64_t &Offset2) const { 1026 // Don't worry about Thumb: just ARM and Thumb2. 1027 if (Subtarget.isThumb1Only()) return false; 1028 1029 if (!Load1->isMachineOpcode() || !Load2->isMachineOpcode()) 1030 return false; 1031 1032 switch (Load1->getMachineOpcode()) { 1033 default: 1034 return false; 1035 case ARM::LDR: 1036 case ARM::LDRB: 1037 case ARM::LDRD: 1038 case ARM::LDRH: 1039 case ARM::LDRSB: 1040 case ARM::LDRSH: 1041 case ARM::VLDRD: 1042 case ARM::VLDRS: 1043 case ARM::t2LDRi8: 1044 case ARM::t2LDRDi8: 1045 case ARM::t2LDRSHi8: 1046 case ARM::t2LDRi12: 1047 case ARM::t2LDRSHi12: 1048 break; 1049 } 1050 1051 switch (Load2->getMachineOpcode()) { 1052 default: 1053 return false; 1054 case ARM::LDR: 1055 case ARM::LDRB: 1056 case ARM::LDRD: 1057 case ARM::LDRH: 1058 case ARM::LDRSB: 1059 case ARM::LDRSH: 1060 case ARM::VLDRD: 1061 case ARM::VLDRS: 1062 case ARM::t2LDRi8: 1063 case ARM::t2LDRDi8: 1064 case ARM::t2LDRSHi8: 1065 case ARM::t2LDRi12: 1066 case ARM::t2LDRSHi12: 1067 break; 1068 } 1069 1070 // Check if base addresses and chain operands match. 1071 if (Load1->getOperand(0) != Load2->getOperand(0) || 1072 Load1->getOperand(4) != Load2->getOperand(4)) 1073 return false; 1074 1075 // Index should be Reg0. 1076 if (Load1->getOperand(3) != Load2->getOperand(3)) 1077 return false; 1078 1079 // Determine the offsets. 1080 if (isa<ConstantSDNode>(Load1->getOperand(1)) && 1081 isa<ConstantSDNode>(Load2->getOperand(1))) { 1082 Offset1 = cast<ConstantSDNode>(Load1->getOperand(1))->getSExtValue(); 1083 Offset2 = cast<ConstantSDNode>(Load2->getOperand(1))->getSExtValue(); 1084 return true; 1085 } 1086 1087 return false; 1088} 1089 1090/// shouldScheduleLoadsNear - This is a used by the pre-regalloc scheduler to 1091/// determine (in conjuction with areLoadsFromSameBasePtr) if two loads should 1092/// be scheduled togther. On some targets if two loads are loading from 1093/// addresses in the same cache line, it's better if they are scheduled 1094/// together. This function takes two integers that represent the load offsets 1095/// from the common base address. It returns true if it decides it's desirable 1096/// to schedule the two loads together. "NumLoads" is the number of loads that 1097/// have already been scheduled after Load1. 1098bool ARMBaseInstrInfo::shouldScheduleLoadsNear(SDNode *Load1, SDNode *Load2, 1099 int64_t Offset1, int64_t Offset2, 1100 unsigned NumLoads) const { 1101 // Don't worry about Thumb: just ARM and Thumb2. 1102 if (Subtarget.isThumb1Only()) return false; 1103 1104 assert(Offset2 > Offset1); 1105 1106 if ((Offset2 - Offset1) / 8 > 64) 1107 return false; 1108 1109 if (Load1->getMachineOpcode() != Load2->getMachineOpcode()) 1110 return false; // FIXME: overly conservative? 1111 1112 // Four loads in a row should be sufficient. 1113 if (NumLoads >= 3) 1114 return false; 1115 1116 return true; 1117} 1118 1119bool ARMBaseInstrInfo::isSchedulingBoundary(const MachineInstr *MI, 1120 const MachineBasicBlock *MBB, 1121 const MachineFunction &MF) const { 1122 // Debug info is never a scheduling boundary. It's necessary to be explicit 1123 // due to the special treatment of IT instructions below, otherwise a 1124 // dbg_value followed by an IT will result in the IT instruction being 1125 // considered a scheduling hazard, which is wrong. It should be the actual 1126 // instruction preceding the dbg_value instruction(s), just like it is 1127 // when debug info is not present. 1128 if (MI->isDebugValue()) 1129 return false; 1130 1131 // Terminators and labels can't be scheduled around. 1132 if (MI->getDesc().isTerminator() || MI->isLabel()) 1133 return true; 1134 1135 // Treat the start of the IT block as a scheduling boundary, but schedule 1136 // t2IT along with all instructions following it. 1137 // FIXME: This is a big hammer. But the alternative is to add all potential 1138 // true and anti dependencies to IT block instructions as implicit operands 1139 // to the t2IT instruction. The added compile time and complexity does not 1140 // seem worth it. 1141 MachineBasicBlock::const_iterator I = MI; 1142 // Make sure to skip any dbg_value instructions 1143 while (++I != MBB->end() && I->isDebugValue()) 1144 ; 1145 if (I != MBB->end() && I->getOpcode() == ARM::t2IT) 1146 return true; 1147 1148 // Don't attempt to schedule around any instruction that defines 1149 // a stack-oriented pointer, as it's unlikely to be profitable. This 1150 // saves compile time, because it doesn't require every single 1151 // stack slot reference to depend on the instruction that does the 1152 // modification. 1153 if (MI->definesRegister(ARM::SP)) 1154 return true; 1155 1156 return false; 1157} 1158 1159bool ARMBaseInstrInfo:: 1160isProfitableToIfCvt(MachineBasicBlock &MBB, unsigned NumInstrs) const { 1161 if (!NumInstrs) 1162 return false; 1163 if (Subtarget.getCPUString() == "generic") 1164 // Generic (and overly aggressive) if-conversion limits for testing. 1165 return NumInstrs <= 10; 1166 else if (Subtarget.hasV7Ops()) 1167 return NumInstrs <= 3; 1168 return NumInstrs <= 2; 1169} 1170 1171bool ARMBaseInstrInfo:: 1172isProfitableToIfCvt(MachineBasicBlock &TMBB, unsigned NumT, 1173 MachineBasicBlock &FMBB, unsigned NumF) const { 1174 return NumT && NumF && NumT <= 2 && NumF <= 2; 1175} 1176 1177/// getInstrPredicate - If instruction is predicated, returns its predicate 1178/// condition, otherwise returns AL. It also returns the condition code 1179/// register by reference. 1180ARMCC::CondCodes 1181llvm::getInstrPredicate(const MachineInstr *MI, unsigned &PredReg) { 1182 int PIdx = MI->findFirstPredOperandIdx(); 1183 if (PIdx == -1) { 1184 PredReg = 0; 1185 return ARMCC::AL; 1186 } 1187 1188 PredReg = MI->getOperand(PIdx+1).getReg(); 1189 return (ARMCC::CondCodes)MI->getOperand(PIdx).getImm(); 1190} 1191 1192 1193int llvm::getMatchingCondBranchOpcode(int Opc) { 1194 if (Opc == ARM::B) 1195 return ARM::Bcc; 1196 else if (Opc == ARM::tB) 1197 return ARM::tBcc; 1198 else if (Opc == ARM::t2B) 1199 return ARM::t2Bcc; 1200 1201 llvm_unreachable("Unknown unconditional branch opcode!"); 1202 return 0; 1203} 1204 1205 1206void llvm::emitARMRegPlusImmediate(MachineBasicBlock &MBB, 1207 MachineBasicBlock::iterator &MBBI, DebugLoc dl, 1208 unsigned DestReg, unsigned BaseReg, int NumBytes, 1209 ARMCC::CondCodes Pred, unsigned PredReg, 1210 const ARMBaseInstrInfo &TII) { 1211 bool isSub = NumBytes < 0; 1212 if (isSub) NumBytes = -NumBytes; 1213 1214 while (NumBytes) { 1215 unsigned RotAmt = ARM_AM::getSOImmValRotate(NumBytes); 1216 unsigned ThisVal = NumBytes & ARM_AM::rotr32(0xFF, RotAmt); 1217 assert(ThisVal && "Didn't extract field correctly"); 1218 1219 // We will handle these bits from offset, clear them. 1220 NumBytes &= ~ThisVal; 1221 1222 assert(ARM_AM::getSOImmVal(ThisVal) != -1 && "Bit extraction didn't work?"); 1223 1224 // Build the new ADD / SUB. 1225 unsigned Opc = isSub ? ARM::SUBri : ARM::ADDri; 1226 BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg) 1227 .addReg(BaseReg, RegState::Kill).addImm(ThisVal) 1228 .addImm((unsigned)Pred).addReg(PredReg).addReg(0); 1229 BaseReg = DestReg; 1230 } 1231} 1232 1233bool llvm::rewriteARMFrameIndex(MachineInstr &MI, unsigned FrameRegIdx, 1234 unsigned FrameReg, int &Offset, 1235 const ARMBaseInstrInfo &TII) { 1236 unsigned Opcode = MI.getOpcode(); 1237 const TargetInstrDesc &Desc = MI.getDesc(); 1238 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask); 1239 bool isSub = false; 1240 1241 // Memory operands in inline assembly always use AddrMode2. 1242 if (Opcode == ARM::INLINEASM) 1243 AddrMode = ARMII::AddrMode2; 1244 1245 if (Opcode == ARM::ADDri) { 1246 Offset += MI.getOperand(FrameRegIdx+1).getImm(); 1247 if (Offset == 0) { 1248 // Turn it into a move. 1249 MI.setDesc(TII.get(ARM::MOVr)); 1250 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false); 1251 MI.RemoveOperand(FrameRegIdx+1); 1252 Offset = 0; 1253 return true; 1254 } else if (Offset < 0) { 1255 Offset = -Offset; 1256 isSub = true; 1257 MI.setDesc(TII.get(ARM::SUBri)); 1258 } 1259 1260 // Common case: small offset, fits into instruction. 1261 if (ARM_AM::getSOImmVal(Offset) != -1) { 1262 // Replace the FrameIndex with sp / fp 1263 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false); 1264 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(Offset); 1265 Offset = 0; 1266 return true; 1267 } 1268 1269 // Otherwise, pull as much of the immedidate into this ADDri/SUBri 1270 // as possible. 1271 unsigned RotAmt = ARM_AM::getSOImmValRotate(Offset); 1272 unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xFF, RotAmt); 1273 1274 // We will handle these bits from offset, clear them. 1275 Offset &= ~ThisImmVal; 1276 1277 // Get the properly encoded SOImmVal field. 1278 assert(ARM_AM::getSOImmVal(ThisImmVal) != -1 && 1279 "Bit extraction didn't work?"); 1280 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(ThisImmVal); 1281 } else { 1282 unsigned ImmIdx = 0; 1283 int InstrOffs = 0; 1284 unsigned NumBits = 0; 1285 unsigned Scale = 1; 1286 switch (AddrMode) { 1287 case ARMII::AddrMode2: { 1288 ImmIdx = FrameRegIdx+2; 1289 InstrOffs = ARM_AM::getAM2Offset(MI.getOperand(ImmIdx).getImm()); 1290 if (ARM_AM::getAM2Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub) 1291 InstrOffs *= -1; 1292 NumBits = 12; 1293 break; 1294 } 1295 case ARMII::AddrMode3: { 1296 ImmIdx = FrameRegIdx+2; 1297 InstrOffs = ARM_AM::getAM3Offset(MI.getOperand(ImmIdx).getImm()); 1298 if (ARM_AM::getAM3Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub) 1299 InstrOffs *= -1; 1300 NumBits = 8; 1301 break; 1302 } 1303 case ARMII::AddrMode4: 1304 case ARMII::AddrMode6: 1305 // Can't fold any offset even if it's zero. 1306 return false; 1307 case ARMII::AddrMode5: { 1308 ImmIdx = FrameRegIdx+1; 1309 InstrOffs = ARM_AM::getAM5Offset(MI.getOperand(ImmIdx).getImm()); 1310 if (ARM_AM::getAM5Op(MI.getOperand(ImmIdx).getImm()) == ARM_AM::sub) 1311 InstrOffs *= -1; 1312 NumBits = 8; 1313 Scale = 4; 1314 break; 1315 } 1316 default: 1317 llvm_unreachable("Unsupported addressing mode!"); 1318 break; 1319 } 1320 1321 Offset += InstrOffs * Scale; 1322 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!"); 1323 if (Offset < 0) { 1324 Offset = -Offset; 1325 isSub = true; 1326 } 1327 1328 // Attempt to fold address comp. if opcode has offset bits 1329 if (NumBits > 0) { 1330 // Common case: small offset, fits into instruction. 1331 MachineOperand &ImmOp = MI.getOperand(ImmIdx); 1332 int ImmedOffset = Offset / Scale; 1333 unsigned Mask = (1 << NumBits) - 1; 1334 if ((unsigned)Offset <= Mask * Scale) { 1335 // Replace the FrameIndex with sp 1336 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false); 1337 if (isSub) 1338 ImmedOffset |= 1 << NumBits; 1339 ImmOp.ChangeToImmediate(ImmedOffset); 1340 Offset = 0; 1341 return true; 1342 } 1343 1344 // Otherwise, it didn't fit. Pull in what we can to simplify the immed. 1345 ImmedOffset = ImmedOffset & Mask; 1346 if (isSub) 1347 ImmedOffset |= 1 << NumBits; 1348 ImmOp.ChangeToImmediate(ImmedOffset); 1349 Offset &= ~(Mask*Scale); 1350 } 1351 } 1352 1353 Offset = (isSub) ? -Offset : Offset; 1354 return Offset == 0; 1355} 1356 1357bool ARMBaseInstrInfo:: 1358AnalyzeCompare(const MachineInstr *MI, unsigned &SrcReg, int &CmpValue) const { 1359 switch (MI->getOpcode()) { 1360 default: break; 1361 case ARM::CMPri: 1362 case ARM::CMPzri: 1363 case ARM::t2CMPri: 1364 case ARM::t2CMPzri: 1365 SrcReg = MI->getOperand(0).getReg(); 1366 CmpValue = MI->getOperand(1).getImm(); 1367 return true; 1368 } 1369 1370 return false; 1371} 1372 1373/// ConvertToSetZeroFlag - Convert the instruction to set the "zero" flag so 1374/// that we can remove a "comparison with zero". 1375bool ARMBaseInstrInfo:: 1376ConvertToSetZeroFlag(MachineInstr *MI, MachineInstr *CmpInstr) const { 1377 // Conservatively refuse to convert an instruction which isn't in the same BB 1378 // as the comparison. 1379 if (MI->getParent() != CmpInstr->getParent()) 1380 return false; 1381 1382 // Check that CPSR isn't set between the comparison instruction and the one we 1383 // want to change. 1384 MachineBasicBlock::const_iterator I = CmpInstr, E = MI; 1385 --I; 1386 for (; I != E; --I) { 1387 const MachineInstr &Instr = *I; 1388 1389 for (unsigned IO = 0, EO = Instr.getNumOperands(); IO != EO; ++IO) { 1390 const MachineOperand &MO = Instr.getOperand(IO); 1391 if (!MO.isReg() || !MO.isDef()) continue; 1392 1393 // This instruction modifies CPSR before the one we want to change. We 1394 // can't do this transformation. 1395 if (MO.getReg() == ARM::CPSR) 1396 return false; 1397 } 1398 } 1399 1400 // Set the "zero" bit in CPSR. 1401 switch (MI->getOpcode()) { 1402 default: break; 1403 case ARM::ADDri: 1404 case ARM::SUBri: 1405 case ARM::t2ADDri: 1406 case ARM::t2SUBri: { 1407 MI->RemoveOperand(5); 1408 MachineInstrBuilder MB(MI); 1409 MB.addReg(ARM::CPSR, RegState::Define | RegState::Implicit); 1410 CmpInstr->eraseFromParent(); 1411 return true; 1412 } 1413 } 1414 1415 return false; 1416} 1417