1 //===- Thumb2InstrInfo.cpp - Thumb-2 Instruction Information --------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file contains the Thumb-2 implementation of the TargetInstrInfo class. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "Thumb2InstrInfo.h" 14 #include "ARMMachineFunctionInfo.h" 15 #include "ARMSubtarget.h" 16 #include "MCTargetDesc/ARMAddressingModes.h" 17 #include "llvm/CodeGen/MachineBasicBlock.h" 18 #include "llvm/CodeGen/MachineFrameInfo.h" 19 #include "llvm/CodeGen/MachineFunction.h" 20 #include "llvm/CodeGen/MachineInstr.h" 21 #include "llvm/CodeGen/MachineInstrBuilder.h" 22 #include "llvm/CodeGen/MachineMemOperand.h" 23 #include "llvm/CodeGen/MachineOperand.h" 24 #include "llvm/CodeGen/MachineRegisterInfo.h" 25 #include "llvm/CodeGen/TargetRegisterInfo.h" 26 #include "llvm/IR/DebugLoc.h" 27 #include "llvm/MC/MCInst.h" 28 #include "llvm/MC/MCInstBuilder.h" 29 #include "llvm/MC/MCInstrDesc.h" 30 #include "llvm/Support/CommandLine.h" 31 #include "llvm/Support/ErrorHandling.h" 32 #include "llvm/Support/MathExtras.h" 33 #include "llvm/Target/TargetMachine.h" 34 #include <cassert> 35 36 using namespace llvm; 37 38 static cl::opt<bool> 39 OldT2IfCvt("old-thumb2-ifcvt", cl::Hidden, 40 cl::desc("Use old-style Thumb2 if-conversion heuristics"), 41 cl::init(false)); 42 43 static cl::opt<bool> 44 PreferNoCSEL("prefer-no-csel", cl::Hidden, 45 cl::desc("Prefer predicated Move to CSEL"), 46 cl::init(false)); 47 48 Thumb2InstrInfo::Thumb2InstrInfo(const ARMSubtarget &STI) 49 : ARMBaseInstrInfo(STI) {} 50 51 /// Return the noop instruction to use for a noop. 52 MCInst Thumb2InstrInfo::getNop() const { 53 return MCInstBuilder(ARM::tHINT).addImm(0).addImm(ARMCC::AL).addReg(0); 54 } 55 56 unsigned Thumb2InstrInfo::getUnindexedOpcode(unsigned Opc) const { 57 // FIXME 58 return 0; 59 } 60 61 void 62 Thumb2InstrInfo::ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail, 63 MachineBasicBlock *NewDest) const { 64 MachineBasicBlock *MBB = Tail->getParent(); 65 ARMFunctionInfo *AFI = MBB->getParent()->getInfo<ARMFunctionInfo>(); 66 if (!AFI->hasITBlocks() || Tail->isBranch()) { 67 TargetInstrInfo::ReplaceTailWithBranchTo(Tail, NewDest); 68 return; 69 } 70 71 // If the first instruction of Tail is predicated, we may have to update 72 // the IT instruction. 73 Register PredReg; 74 ARMCC::CondCodes CC = getInstrPredicate(*Tail, PredReg); 75 MachineBasicBlock::iterator MBBI = Tail; 76 if (CC != ARMCC::AL) 77 // Expecting at least the t2IT instruction before it. 78 --MBBI; 79 80 // Actually replace the tail. 81 TargetInstrInfo::ReplaceTailWithBranchTo(Tail, NewDest); 82 83 // Fix up IT. 84 if (CC != ARMCC::AL) { 85 MachineBasicBlock::iterator E = MBB->begin(); 86 unsigned Count = 4; // At most 4 instructions in an IT block. 87 while (Count && MBBI != E) { 88 if (MBBI->isDebugInstr()) { 89 --MBBI; 90 continue; 91 } 92 if (MBBI->getOpcode() == ARM::t2IT) { 93 unsigned Mask = MBBI->getOperand(1).getImm(); 94 if (Count == 4) 95 MBBI->eraseFromParent(); 96 else { 97 unsigned MaskOn = 1 << Count; 98 unsigned MaskOff = ~(MaskOn - 1); 99 MBBI->getOperand(1).setImm((Mask & MaskOff) | MaskOn); 100 } 101 return; 102 } 103 --MBBI; 104 --Count; 105 } 106 107 // Ctrl flow can reach here if branch folding is run before IT block 108 // formation pass. 109 } 110 } 111 112 bool 113 Thumb2InstrInfo::isLegalToSplitMBBAt(MachineBasicBlock &MBB, 114 MachineBasicBlock::iterator MBBI) const { 115 while (MBBI->isDebugInstr()) { 116 ++MBBI; 117 if (MBBI == MBB.end()) 118 return false; 119 } 120 121 Register PredReg; 122 return getITInstrPredicate(*MBBI, PredReg) == ARMCC::AL; 123 } 124 125 MachineInstr * 126 Thumb2InstrInfo::optimizeSelect(MachineInstr &MI, 127 SmallPtrSetImpl<MachineInstr *> &SeenMIs, 128 bool PreferFalse) const { 129 // Try to use the base optimizeSelect, which uses canFoldIntoMOVCC to fold the 130 // MOVCC into another instruction. If that fails on 8.1-M fall back to using a 131 // CSEL. 132 MachineInstr *RV = ARMBaseInstrInfo::optimizeSelect(MI, SeenMIs, PreferFalse); 133 if (!RV && getSubtarget().hasV8_1MMainlineOps() && !PreferNoCSEL) { 134 Register DestReg = MI.getOperand(0).getReg(); 135 136 if (!DestReg.isVirtual()) 137 return nullptr; 138 139 MachineInstrBuilder NewMI = BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), 140 get(ARM::t2CSEL), DestReg) 141 .add(MI.getOperand(2)) 142 .add(MI.getOperand(1)) 143 .add(MI.getOperand(3)); 144 SeenMIs.insert(NewMI); 145 return NewMI; 146 } 147 return RV; 148 } 149 150 void Thumb2InstrInfo::copyPhysReg(MachineBasicBlock &MBB, 151 MachineBasicBlock::iterator I, 152 const DebugLoc &DL, MCRegister DestReg, 153 MCRegister SrcReg, bool KillSrc) const { 154 // Handle SPR, DPR, and QPR copies. 155 if (!ARM::GPRRegClass.contains(DestReg, SrcReg)) 156 return ARMBaseInstrInfo::copyPhysReg(MBB, I, DL, DestReg, SrcReg, KillSrc); 157 158 BuildMI(MBB, I, DL, get(ARM::tMOVr), DestReg) 159 .addReg(SrcReg, getKillRegState(KillSrc)) 160 .add(predOps(ARMCC::AL)); 161 } 162 163 void Thumb2InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB, 164 MachineBasicBlock::iterator I, 165 Register SrcReg, bool isKill, int FI, 166 const TargetRegisterClass *RC, 167 const TargetRegisterInfo *TRI, 168 Register VReg) const { 169 DebugLoc DL; 170 if (I != MBB.end()) DL = I->getDebugLoc(); 171 172 MachineFunction &MF = *MBB.getParent(); 173 MachineFrameInfo &MFI = MF.getFrameInfo(); 174 MachineMemOperand *MMO = MF.getMachineMemOperand( 175 MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOStore, 176 MFI.getObjectSize(FI), MFI.getObjectAlign(FI)); 177 178 if (ARM::GPRRegClass.hasSubClassEq(RC)) { 179 BuildMI(MBB, I, DL, get(ARM::t2STRi12)) 180 .addReg(SrcReg, getKillRegState(isKill)) 181 .addFrameIndex(FI) 182 .addImm(0) 183 .addMemOperand(MMO) 184 .add(predOps(ARMCC::AL)); 185 return; 186 } 187 188 if (ARM::GPRPairRegClass.hasSubClassEq(RC)) { 189 // Thumb2 STRD expects its dest-registers to be in rGPR. Not a problem for 190 // gsub_0, but needs an extra constraint for gsub_1 (which could be sp 191 // otherwise). 192 if (SrcReg.isVirtual()) { 193 MachineRegisterInfo *MRI = &MF.getRegInfo(); 194 MRI->constrainRegClass(SrcReg, &ARM::GPRPairnospRegClass); 195 } 196 197 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(ARM::t2STRDi8)); 198 AddDReg(MIB, SrcReg, ARM::gsub_0, getKillRegState(isKill), TRI); 199 AddDReg(MIB, SrcReg, ARM::gsub_1, 0, TRI); 200 MIB.addFrameIndex(FI).addImm(0).addMemOperand(MMO).add(predOps(ARMCC::AL)); 201 return; 202 } 203 204 ARMBaseInstrInfo::storeRegToStackSlot(MBB, I, SrcReg, isKill, FI, RC, TRI, 205 Register()); 206 } 207 208 void Thumb2InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB, 209 MachineBasicBlock::iterator I, 210 Register DestReg, int FI, 211 const TargetRegisterClass *RC, 212 const TargetRegisterInfo *TRI, 213 Register VReg) const { 214 MachineFunction &MF = *MBB.getParent(); 215 MachineFrameInfo &MFI = MF.getFrameInfo(); 216 MachineMemOperand *MMO = MF.getMachineMemOperand( 217 MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOLoad, 218 MFI.getObjectSize(FI), MFI.getObjectAlign(FI)); 219 DebugLoc DL; 220 if (I != MBB.end()) DL = I->getDebugLoc(); 221 222 if (ARM::GPRRegClass.hasSubClassEq(RC)) { 223 BuildMI(MBB, I, DL, get(ARM::t2LDRi12), DestReg) 224 .addFrameIndex(FI) 225 .addImm(0) 226 .addMemOperand(MMO) 227 .add(predOps(ARMCC::AL)); 228 return; 229 } 230 231 if (ARM::GPRPairRegClass.hasSubClassEq(RC)) { 232 // Thumb2 LDRD expects its dest-registers to be in rGPR. Not a problem for 233 // gsub_0, but needs an extra constraint for gsub_1 (which could be sp 234 // otherwise). 235 if (DestReg.isVirtual()) { 236 MachineRegisterInfo *MRI = &MF.getRegInfo(); 237 MRI->constrainRegClass(DestReg, &ARM::GPRPairnospRegClass); 238 } 239 240 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(ARM::t2LDRDi8)); 241 AddDReg(MIB, DestReg, ARM::gsub_0, RegState::DefineNoRead, TRI); 242 AddDReg(MIB, DestReg, ARM::gsub_1, RegState::DefineNoRead, TRI); 243 MIB.addFrameIndex(FI).addImm(0).addMemOperand(MMO).add(predOps(ARMCC::AL)); 244 245 if (DestReg.isPhysical()) 246 MIB.addReg(DestReg, RegState::ImplicitDefine); 247 return; 248 } 249 250 ARMBaseInstrInfo::loadRegFromStackSlot(MBB, I, DestReg, FI, RC, TRI, 251 Register()); 252 } 253 254 void Thumb2InstrInfo::expandLoadStackGuard( 255 MachineBasicBlock::iterator MI) const { 256 MachineFunction &MF = *MI->getParent()->getParent(); 257 Module &M = *MF.getFunction().getParent(); 258 259 if (M.getStackProtectorGuard() == "tls") { 260 expandLoadStackGuardBase(MI, ARM::t2MRC, ARM::t2LDRi12); 261 return; 262 } 263 264 const GlobalValue *GV = 265 cast<GlobalValue>((*MI->memoperands_begin())->getValue()); 266 267 if (MF.getSubtarget<ARMSubtarget>().isGVInGOT(GV)) 268 expandLoadStackGuardBase(MI, ARM::t2LDRLIT_ga_pcrel, ARM::t2LDRi12); 269 else if (MF.getTarget().isPositionIndependent()) 270 expandLoadStackGuardBase(MI, ARM::t2MOV_ga_pcrel, ARM::t2LDRi12); 271 else 272 expandLoadStackGuardBase(MI, ARM::t2MOVi32imm, ARM::t2LDRi12); 273 } 274 275 MachineInstr *Thumb2InstrInfo::commuteInstructionImpl(MachineInstr &MI, 276 bool NewMI, 277 unsigned OpIdx1, 278 unsigned OpIdx2) const { 279 switch (MI.getOpcode()) { 280 case ARM::MVE_VMAXNMAf16: 281 case ARM::MVE_VMAXNMAf32: 282 case ARM::MVE_VMINNMAf16: 283 case ARM::MVE_VMINNMAf32: 284 // Don't allow predicated instructions to be commuted. 285 if (getVPTInstrPredicate(MI) != ARMVCC::None) 286 return nullptr; 287 } 288 return ARMBaseInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2); 289 } 290 291 void llvm::emitT2RegPlusImmediate(MachineBasicBlock &MBB, 292 MachineBasicBlock::iterator &MBBI, 293 const DebugLoc &dl, Register DestReg, 294 Register BaseReg, int NumBytes, 295 ARMCC::CondCodes Pred, Register PredReg, 296 const ARMBaseInstrInfo &TII, 297 unsigned MIFlags) { 298 if (NumBytes == 0 && DestReg != BaseReg) { 299 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr), DestReg) 300 .addReg(BaseReg, RegState::Kill) 301 .addImm((unsigned)Pred).addReg(PredReg).setMIFlags(MIFlags); 302 return; 303 } 304 305 bool isSub = NumBytes < 0; 306 if (isSub) NumBytes = -NumBytes; 307 308 // If profitable, use a movw or movt to materialize the offset. 309 // FIXME: Use the scavenger to grab a scratch register. 310 if (DestReg != ARM::SP && DestReg != BaseReg && 311 NumBytes >= 4096 && 312 ARM_AM::getT2SOImmVal(NumBytes) == -1) { 313 bool Fits = false; 314 if (NumBytes < 65536) { 315 // Use a movw to materialize the 16-bit constant. 316 BuildMI(MBB, MBBI, dl, TII.get(ARM::t2MOVi16), DestReg) 317 .addImm(NumBytes) 318 .addImm((unsigned)Pred).addReg(PredReg).setMIFlags(MIFlags); 319 Fits = true; 320 } else if ((NumBytes & 0xffff) == 0) { 321 // Use a movt to materialize the 32-bit constant. 322 BuildMI(MBB, MBBI, dl, TII.get(ARM::t2MOVTi16), DestReg) 323 .addReg(DestReg) 324 .addImm(NumBytes >> 16) 325 .addImm((unsigned)Pred).addReg(PredReg).setMIFlags(MIFlags); 326 Fits = true; 327 } 328 329 if (Fits) { 330 if (isSub) { 331 BuildMI(MBB, MBBI, dl, TII.get(ARM::t2SUBrr), DestReg) 332 .addReg(BaseReg) 333 .addReg(DestReg, RegState::Kill) 334 .add(predOps(Pred, PredReg)) 335 .add(condCodeOp()) 336 .setMIFlags(MIFlags); 337 } else { 338 // Here we know that DestReg is not SP but we do not 339 // know anything about BaseReg. t2ADDrr is an invalid 340 // instruction is SP is used as the second argument, but 341 // is fine if SP is the first argument. To be sure we 342 // do not generate invalid encoding, put BaseReg first. 343 BuildMI(MBB, MBBI, dl, TII.get(ARM::t2ADDrr), DestReg) 344 .addReg(BaseReg) 345 .addReg(DestReg, RegState::Kill) 346 .add(predOps(Pred, PredReg)) 347 .add(condCodeOp()) 348 .setMIFlags(MIFlags); 349 } 350 return; 351 } 352 } 353 354 while (NumBytes) { 355 unsigned ThisVal = NumBytes; 356 unsigned Opc = 0; 357 if (DestReg == ARM::SP && BaseReg != ARM::SP) { 358 // mov sp, rn. Note t2MOVr cannot be used. 359 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr), DestReg) 360 .addReg(BaseReg) 361 .setMIFlags(MIFlags) 362 .add(predOps(ARMCC::AL)); 363 BaseReg = ARM::SP; 364 continue; 365 } 366 367 assert((DestReg != ARM::SP || BaseReg == ARM::SP) && 368 "Writing to SP, from other register."); 369 370 // Try to use T1, as it smaller 371 if ((DestReg == ARM::SP) && (ThisVal < ((1 << 7) - 1) * 4)) { 372 assert((ThisVal & 3) == 0 && "Stack update is not multiple of 4?"); 373 Opc = isSub ? ARM::tSUBspi : ARM::tADDspi; 374 BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg) 375 .addReg(BaseReg) 376 .addImm(ThisVal / 4) 377 .setMIFlags(MIFlags) 378 .add(predOps(ARMCC::AL)); 379 break; 380 } 381 bool HasCCOut = true; 382 int ImmIsT2SO = ARM_AM::getT2SOImmVal(ThisVal); 383 bool ToSP = DestReg == ARM::SP; 384 unsigned t2SUB = ToSP ? ARM::t2SUBspImm : ARM::t2SUBri; 385 unsigned t2ADD = ToSP ? ARM::t2ADDspImm : ARM::t2ADDri; 386 unsigned t2SUBi12 = ToSP ? ARM::t2SUBspImm12 : ARM::t2SUBri12; 387 unsigned t2ADDi12 = ToSP ? ARM::t2ADDspImm12 : ARM::t2ADDri12; 388 Opc = isSub ? t2SUB : t2ADD; 389 // Prefer T2: sub rd, rn, so_imm | sub sp, sp, so_imm 390 if (ImmIsT2SO != -1) { 391 NumBytes = 0; 392 } else if (ThisVal < 4096) { 393 // Prefer T3 if can make it in a single go: subw rd, rn, imm12 | subw sp, 394 // sp, imm12 395 Opc = isSub ? t2SUBi12 : t2ADDi12; 396 HasCCOut = false; 397 NumBytes = 0; 398 } else { 399 // Use one T2 instruction to reduce NumBytes 400 // FIXME: Move this to ARMAddressingModes.h? 401 unsigned RotAmt = llvm::countl_zero(ThisVal); 402 ThisVal = ThisVal & llvm::rotr<uint32_t>(0xff000000U, RotAmt); 403 NumBytes &= ~ThisVal; 404 assert(ARM_AM::getT2SOImmVal(ThisVal) != -1 && 405 "Bit extraction didn't work?"); 406 } 407 408 // Build the new ADD / SUB. 409 MachineInstrBuilder MIB = BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg) 410 .addReg(BaseReg, RegState::Kill) 411 .addImm(ThisVal) 412 .add(predOps(ARMCC::AL)) 413 .setMIFlags(MIFlags); 414 if (HasCCOut) 415 MIB.add(condCodeOp()); 416 417 BaseReg = DestReg; 418 } 419 } 420 421 static unsigned 422 negativeOffsetOpcode(unsigned opcode) 423 { 424 switch (opcode) { 425 case ARM::t2LDRi12: return ARM::t2LDRi8; 426 case ARM::t2LDRHi12: return ARM::t2LDRHi8; 427 case ARM::t2LDRBi12: return ARM::t2LDRBi8; 428 case ARM::t2LDRSHi12: return ARM::t2LDRSHi8; 429 case ARM::t2LDRSBi12: return ARM::t2LDRSBi8; 430 case ARM::t2STRi12: return ARM::t2STRi8; 431 case ARM::t2STRBi12: return ARM::t2STRBi8; 432 case ARM::t2STRHi12: return ARM::t2STRHi8; 433 case ARM::t2PLDi12: return ARM::t2PLDi8; 434 case ARM::t2PLDWi12: return ARM::t2PLDWi8; 435 case ARM::t2PLIi12: return ARM::t2PLIi8; 436 437 case ARM::t2LDRi8: 438 case ARM::t2LDRHi8: 439 case ARM::t2LDRBi8: 440 case ARM::t2LDRSHi8: 441 case ARM::t2LDRSBi8: 442 case ARM::t2STRi8: 443 case ARM::t2STRBi8: 444 case ARM::t2STRHi8: 445 case ARM::t2PLDi8: 446 case ARM::t2PLDWi8: 447 case ARM::t2PLIi8: 448 return opcode; 449 450 default: 451 llvm_unreachable("unknown thumb2 opcode."); 452 } 453 } 454 455 static unsigned 456 positiveOffsetOpcode(unsigned opcode) 457 { 458 switch (opcode) { 459 case ARM::t2LDRi8: return ARM::t2LDRi12; 460 case ARM::t2LDRHi8: return ARM::t2LDRHi12; 461 case ARM::t2LDRBi8: return ARM::t2LDRBi12; 462 case ARM::t2LDRSHi8: return ARM::t2LDRSHi12; 463 case ARM::t2LDRSBi8: return ARM::t2LDRSBi12; 464 case ARM::t2STRi8: return ARM::t2STRi12; 465 case ARM::t2STRBi8: return ARM::t2STRBi12; 466 case ARM::t2STRHi8: return ARM::t2STRHi12; 467 case ARM::t2PLDi8: return ARM::t2PLDi12; 468 case ARM::t2PLDWi8: return ARM::t2PLDWi12; 469 case ARM::t2PLIi8: return ARM::t2PLIi12; 470 471 case ARM::t2LDRi12: 472 case ARM::t2LDRHi12: 473 case ARM::t2LDRBi12: 474 case ARM::t2LDRSHi12: 475 case ARM::t2LDRSBi12: 476 case ARM::t2STRi12: 477 case ARM::t2STRBi12: 478 case ARM::t2STRHi12: 479 case ARM::t2PLDi12: 480 case ARM::t2PLDWi12: 481 case ARM::t2PLIi12: 482 return opcode; 483 484 default: 485 llvm_unreachable("unknown thumb2 opcode."); 486 } 487 } 488 489 static unsigned 490 immediateOffsetOpcode(unsigned opcode) 491 { 492 switch (opcode) { 493 case ARM::t2LDRs: return ARM::t2LDRi12; 494 case ARM::t2LDRHs: return ARM::t2LDRHi12; 495 case ARM::t2LDRBs: return ARM::t2LDRBi12; 496 case ARM::t2LDRSHs: return ARM::t2LDRSHi12; 497 case ARM::t2LDRSBs: return ARM::t2LDRSBi12; 498 case ARM::t2STRs: return ARM::t2STRi12; 499 case ARM::t2STRBs: return ARM::t2STRBi12; 500 case ARM::t2STRHs: return ARM::t2STRHi12; 501 case ARM::t2PLDs: return ARM::t2PLDi12; 502 case ARM::t2PLDWs: return ARM::t2PLDWi12; 503 case ARM::t2PLIs: return ARM::t2PLIi12; 504 505 case ARM::t2LDRi12: 506 case ARM::t2LDRHi12: 507 case ARM::t2LDRBi12: 508 case ARM::t2LDRSHi12: 509 case ARM::t2LDRSBi12: 510 case ARM::t2STRi12: 511 case ARM::t2STRBi12: 512 case ARM::t2STRHi12: 513 case ARM::t2PLDi12: 514 case ARM::t2PLDWi12: 515 case ARM::t2PLIi12: 516 case ARM::t2LDRi8: 517 case ARM::t2LDRHi8: 518 case ARM::t2LDRBi8: 519 case ARM::t2LDRSHi8: 520 case ARM::t2LDRSBi8: 521 case ARM::t2STRi8: 522 case ARM::t2STRBi8: 523 case ARM::t2STRHi8: 524 case ARM::t2PLDi8: 525 case ARM::t2PLDWi8: 526 case ARM::t2PLIi8: 527 return opcode; 528 529 default: 530 llvm_unreachable("unknown thumb2 opcode."); 531 } 532 } 533 534 bool llvm::rewriteT2FrameIndex(MachineInstr &MI, unsigned FrameRegIdx, 535 Register FrameReg, int &Offset, 536 const ARMBaseInstrInfo &TII, 537 const TargetRegisterInfo *TRI) { 538 unsigned Opcode = MI.getOpcode(); 539 const MCInstrDesc &Desc = MI.getDesc(); 540 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask); 541 bool isSub = false; 542 543 MachineFunction &MF = *MI.getParent()->getParent(); 544 const TargetRegisterClass *RegClass = 545 TII.getRegClass(Desc, FrameRegIdx, TRI, MF); 546 547 // Memory operands in inline assembly always use AddrModeT2_i12. 548 if (Opcode == ARM::INLINEASM || Opcode == ARM::INLINEASM_BR) 549 AddrMode = ARMII::AddrModeT2_i12; // FIXME. mode for thumb2? 550 551 const bool IsSP = Opcode == ARM::t2ADDspImm12 || Opcode == ARM::t2ADDspImm; 552 if (IsSP || Opcode == ARM::t2ADDri || Opcode == ARM::t2ADDri12) { 553 Offset += MI.getOperand(FrameRegIdx+1).getImm(); 554 555 Register PredReg; 556 if (Offset == 0 && getInstrPredicate(MI, PredReg) == ARMCC::AL && 557 !MI.definesRegister(ARM::CPSR)) { 558 // Turn it into a move. 559 MI.setDesc(TII.get(ARM::tMOVr)); 560 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false); 561 // Remove offset and remaining explicit predicate operands. 562 do MI.removeOperand(FrameRegIdx+1); 563 while (MI.getNumOperands() > FrameRegIdx+1); 564 MachineInstrBuilder MIB(*MI.getParent()->getParent(), &MI); 565 MIB.add(predOps(ARMCC::AL)); 566 return true; 567 } 568 569 bool HasCCOut = (Opcode != ARM::t2ADDspImm12 && Opcode != ARM::t2ADDri12); 570 571 if (Offset < 0) { 572 Offset = -Offset; 573 isSub = true; 574 MI.setDesc(IsSP ? TII.get(ARM::t2SUBspImm) : TII.get(ARM::t2SUBri)); 575 } else { 576 MI.setDesc(IsSP ? TII.get(ARM::t2ADDspImm) : TII.get(ARM::t2ADDri)); 577 } 578 579 // Common case: small offset, fits into instruction. 580 if (ARM_AM::getT2SOImmVal(Offset) != -1) { 581 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false); 582 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(Offset); 583 // Add cc_out operand if the original instruction did not have one. 584 if (!HasCCOut) 585 MI.addOperand(MachineOperand::CreateReg(0, false)); 586 Offset = 0; 587 return true; 588 } 589 // Another common case: imm12. 590 if (Offset < 4096 && 591 (!HasCCOut || MI.getOperand(MI.getNumOperands()-1).getReg() == 0)) { 592 unsigned NewOpc = isSub ? IsSP ? ARM::t2SUBspImm12 : ARM::t2SUBri12 593 : IsSP ? ARM::t2ADDspImm12 : ARM::t2ADDri12; 594 MI.setDesc(TII.get(NewOpc)); 595 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false); 596 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(Offset); 597 // Remove the cc_out operand. 598 if (HasCCOut) 599 MI.removeOperand(MI.getNumOperands()-1); 600 Offset = 0; 601 return true; 602 } 603 604 // Otherwise, extract 8 adjacent bits from the immediate into this 605 // t2ADDri/t2SUBri. 606 unsigned RotAmt = llvm::countl_zero<unsigned>(Offset); 607 unsigned ThisImmVal = Offset & llvm::rotr<uint32_t>(0xff000000U, RotAmt); 608 609 // We will handle these bits from offset, clear them. 610 Offset &= ~ThisImmVal; 611 612 assert(ARM_AM::getT2SOImmVal(ThisImmVal) != -1 && 613 "Bit extraction didn't work?"); 614 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(ThisImmVal); 615 // Add cc_out operand if the original instruction did not have one. 616 if (!HasCCOut) 617 MI.addOperand(MachineOperand::CreateReg(0, false)); 618 } else { 619 // AddrMode4 and AddrMode6 cannot handle any offset. 620 if (AddrMode == ARMII::AddrMode4 || AddrMode == ARMII::AddrMode6) 621 return false; 622 623 // AddrModeT2_so cannot handle any offset. If there is no offset 624 // register then we change to an immediate version. 625 unsigned NewOpc = Opcode; 626 if (AddrMode == ARMII::AddrModeT2_so) { 627 Register OffsetReg = MI.getOperand(FrameRegIdx + 1).getReg(); 628 if (OffsetReg != 0) { 629 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false); 630 return Offset == 0; 631 } 632 633 MI.removeOperand(FrameRegIdx+1); 634 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(0); 635 NewOpc = immediateOffsetOpcode(Opcode); 636 AddrMode = ARMII::AddrModeT2_i12; 637 } 638 639 unsigned NumBits = 0; 640 unsigned Scale = 1; 641 if (AddrMode == ARMII::AddrModeT2_i8neg || 642 AddrMode == ARMII::AddrModeT2_i12) { 643 // i8 supports only negative, and i12 supports only positive, so 644 // based on Offset sign convert Opcode to the appropriate 645 // instruction 646 Offset += MI.getOperand(FrameRegIdx+1).getImm(); 647 if (Offset < 0) { 648 NewOpc = negativeOffsetOpcode(Opcode); 649 NumBits = 8; 650 isSub = true; 651 Offset = -Offset; 652 } else { 653 NewOpc = positiveOffsetOpcode(Opcode); 654 NumBits = 12; 655 } 656 } else if (AddrMode == ARMII::AddrMode5) { 657 // VFP address mode. 658 const MachineOperand &OffOp = MI.getOperand(FrameRegIdx+1); 659 int InstrOffs = ARM_AM::getAM5Offset(OffOp.getImm()); 660 if (ARM_AM::getAM5Op(OffOp.getImm()) == ARM_AM::sub) 661 InstrOffs *= -1; 662 NumBits = 8; 663 Scale = 4; 664 Offset += InstrOffs * 4; 665 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!"); 666 if (Offset < 0) { 667 Offset = -Offset; 668 isSub = true; 669 } 670 } else if (AddrMode == ARMII::AddrMode5FP16) { 671 // VFP address mode. 672 const MachineOperand &OffOp = MI.getOperand(FrameRegIdx+1); 673 int InstrOffs = ARM_AM::getAM5FP16Offset(OffOp.getImm()); 674 if (ARM_AM::getAM5FP16Op(OffOp.getImm()) == ARM_AM::sub) 675 InstrOffs *= -1; 676 NumBits = 8; 677 Scale = 2; 678 Offset += InstrOffs * 2; 679 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!"); 680 if (Offset < 0) { 681 Offset = -Offset; 682 isSub = true; 683 } 684 } else if (AddrMode == ARMII::AddrModeT2_i7s4 || 685 AddrMode == ARMII::AddrModeT2_i7s2 || 686 AddrMode == ARMII::AddrModeT2_i7) { 687 Offset += MI.getOperand(FrameRegIdx + 1).getImm(); 688 unsigned OffsetMask; 689 switch (AddrMode) { 690 case ARMII::AddrModeT2_i7s4: NumBits = 9; OffsetMask = 0x3; break; 691 case ARMII::AddrModeT2_i7s2: NumBits = 8; OffsetMask = 0x1; break; 692 default: NumBits = 7; OffsetMask = 0x0; break; 693 } 694 // MCInst operand expects already scaled value. 695 Scale = 1; 696 assert((Offset & OffsetMask) == 0 && "Can't encode this offset!"); 697 (void)OffsetMask; // squash unused-variable warning at -NDEBUG 698 } else if (AddrMode == ARMII::AddrModeT2_i8s4) { 699 Offset += MI.getOperand(FrameRegIdx + 1).getImm(); 700 NumBits = 8 + 2; 701 // MCInst operand expects already scaled value. 702 Scale = 1; 703 assert((Offset & 3) == 0 && "Can't encode this offset!"); 704 } else if (AddrMode == ARMII::AddrModeT2_ldrex) { 705 Offset += MI.getOperand(FrameRegIdx + 1).getImm() * 4; 706 NumBits = 8; // 8 bits scaled by 4 707 Scale = 4; 708 assert((Offset & 3) == 0 && "Can't encode this offset!"); 709 } else { 710 llvm_unreachable("Unsupported addressing mode!"); 711 } 712 713 if (NewOpc != Opcode) 714 MI.setDesc(TII.get(NewOpc)); 715 716 MachineOperand &ImmOp = MI.getOperand(FrameRegIdx+1); 717 718 // Attempt to fold address computation 719 // Common case: small offset, fits into instruction. We need to make sure 720 // the register class is correct too, for instructions like the MVE 721 // VLDRH.32, which only accepts low tGPR registers. 722 int ImmedOffset = Offset / Scale; 723 unsigned Mask = (1 << NumBits) - 1; 724 if ((unsigned)Offset <= Mask * Scale && 725 (FrameReg.isVirtual() || RegClass->contains(FrameReg))) { 726 if (FrameReg.isVirtual()) { 727 // Make sure the register class for the virtual register is correct 728 MachineRegisterInfo *MRI = &MF.getRegInfo(); 729 if (!MRI->constrainRegClass(FrameReg, RegClass)) 730 llvm_unreachable("Unable to constrain virtual register class."); 731 } 732 733 // Replace the FrameIndex with fp/sp 734 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false); 735 if (isSub) { 736 if (AddrMode == ARMII::AddrMode5 || AddrMode == ARMII::AddrMode5FP16) 737 // FIXME: Not consistent. 738 ImmedOffset |= 1 << NumBits; 739 else 740 ImmedOffset = -ImmedOffset; 741 } 742 ImmOp.ChangeToImmediate(ImmedOffset); 743 Offset = 0; 744 return true; 745 } 746 747 // Otherwise, offset doesn't fit. Pull in what we can to simplify 748 ImmedOffset = ImmedOffset & Mask; 749 if (isSub) { 750 if (AddrMode == ARMII::AddrMode5 || AddrMode == ARMII::AddrMode5FP16) 751 // FIXME: Not consistent. 752 ImmedOffset |= 1 << NumBits; 753 else { 754 ImmedOffset = -ImmedOffset; 755 if (ImmedOffset == 0) 756 // Change the opcode back if the encoded offset is zero. 757 MI.setDesc(TII.get(positiveOffsetOpcode(NewOpc))); 758 } 759 } 760 ImmOp.ChangeToImmediate(ImmedOffset); 761 Offset &= ~(Mask*Scale); 762 } 763 764 Offset = (isSub) ? -Offset : Offset; 765 return Offset == 0 && (FrameReg.isVirtual() || RegClass->contains(FrameReg)); 766 } 767 768 ARMCC::CondCodes llvm::getITInstrPredicate(const MachineInstr &MI, 769 Register &PredReg) { 770 unsigned Opc = MI.getOpcode(); 771 if (Opc == ARM::tBcc || Opc == ARM::t2Bcc) 772 return ARMCC::AL; 773 return getInstrPredicate(MI, PredReg); 774 } 775 776 int llvm::findFirstVPTPredOperandIdx(const MachineInstr &MI) { 777 const MCInstrDesc &MCID = MI.getDesc(); 778 779 for (unsigned i = 0, e = MCID.getNumOperands(); i != e; ++i) 780 if (ARM::isVpred(MCID.operands()[i].OperandType)) 781 return i; 782 783 return -1; 784 } 785 786 ARMVCC::VPTCodes llvm::getVPTInstrPredicate(const MachineInstr &MI, 787 Register &PredReg) { 788 int PIdx = findFirstVPTPredOperandIdx(MI); 789 if (PIdx == -1) { 790 PredReg = 0; 791 return ARMVCC::None; 792 } 793 794 PredReg = MI.getOperand(PIdx+1).getReg(); 795 return (ARMVCC::VPTCodes)MI.getOperand(PIdx).getImm(); 796 } 797 798 void llvm::recomputeVPTBlockMask(MachineInstr &Instr) { 799 assert(isVPTOpcode(Instr.getOpcode()) && "Not a VPST or VPT Instruction!"); 800 801 MachineOperand &MaskOp = Instr.getOperand(0); 802 assert(MaskOp.isImm() && "Operand 0 is not the block mask of the VPT/VPST?!"); 803 804 MachineBasicBlock::iterator Iter = ++Instr.getIterator(), 805 End = Instr.getParent()->end(); 806 807 while (Iter != End && Iter->isDebugInstr()) 808 ++Iter; 809 810 // Verify that the instruction after the VPT/VPST is predicated (it should 811 // be), and skip it. 812 assert(Iter != End && "Expected some instructions in any VPT block"); 813 assert( 814 getVPTInstrPredicate(*Iter) == ARMVCC::Then && 815 "VPT/VPST should be followed by an instruction with a 'then' predicate!"); 816 ++Iter; 817 818 // Iterate over the predicated instructions, updating the BlockMask as we go. 819 ARM::PredBlockMask BlockMask = ARM::PredBlockMask::T; 820 while (Iter != End) { 821 if (Iter->isDebugInstr()) { 822 ++Iter; 823 continue; 824 } 825 ARMVCC::VPTCodes Pred = getVPTInstrPredicate(*Iter); 826 if (Pred == ARMVCC::None) 827 break; 828 BlockMask = expandPredBlockMask(BlockMask, Pred); 829 ++Iter; 830 } 831 832 // Rewrite the BlockMask. 833 MaskOp.setImm((int64_t)(BlockMask)); 834 } 835