1 //===-- Thumb2SizeReduction.cpp - Thumb2 code size reduction pass -*- 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 #include "ARM.h" 11 #include "ARMBaseInstrInfo.h" 12 #include "ARMSubtarget.h" 13 #include "MCTargetDesc/ARMAddressingModes.h" 14 #include "Thumb2InstrInfo.h" 15 #include "llvm/ADT/DenseMap.h" 16 #include "llvm/ADT/PostOrderIterator.h" 17 #include "llvm/ADT/Statistic.h" 18 #include "llvm/CodeGen/MachineFunctionPass.h" 19 #include "llvm/CodeGen/MachineInstr.h" 20 #include "llvm/CodeGen/MachineInstrBuilder.h" 21 #include "llvm/IR/Function.h" // To access Function attributes 22 #include "llvm/Support/CommandLine.h" 23 #include "llvm/Support/Debug.h" 24 #include "llvm/Target/TargetMachine.h" 25 using namespace llvm; 26 27 #define DEBUG_TYPE "t2-reduce-size" 28 29 STATISTIC(NumNarrows, "Number of 32-bit instrs reduced to 16-bit ones"); 30 STATISTIC(Num2Addrs, "Number of 32-bit instrs reduced to 2addr 16-bit ones"); 31 STATISTIC(NumLdSts, "Number of 32-bit load / store reduced to 16-bit ones"); 32 33 static cl::opt<int> ReduceLimit("t2-reduce-limit", 34 cl::init(-1), cl::Hidden); 35 static cl::opt<int> ReduceLimit2Addr("t2-reduce-limit2", 36 cl::init(-1), cl::Hidden); 37 static cl::opt<int> ReduceLimitLdSt("t2-reduce-limit3", 38 cl::init(-1), cl::Hidden); 39 40 namespace { 41 /// ReduceTable - A static table with information on mapping from wide 42 /// opcodes to narrow 43 struct ReduceEntry { 44 uint16_t WideOpc; // Wide opcode 45 uint16_t NarrowOpc1; // Narrow opcode to transform to 46 uint16_t NarrowOpc2; // Narrow opcode when it's two-address 47 uint8_t Imm1Limit; // Limit of immediate field (bits) 48 uint8_t Imm2Limit; // Limit of immediate field when it's two-address 49 unsigned LowRegs1 : 1; // Only possible if low-registers are used 50 unsigned LowRegs2 : 1; // Only possible if low-registers are used (2addr) 51 unsigned PredCC1 : 2; // 0 - If predicated, cc is on and vice versa. 52 // 1 - No cc field. 53 // 2 - Always set CPSR. 54 unsigned PredCC2 : 2; 55 unsigned PartFlag : 1; // 16-bit instruction does partial flag update 56 unsigned Special : 1; // Needs to be dealt with specially 57 unsigned AvoidMovs: 1; // Avoid movs with shifter operand (for Swift) 58 }; 59 60 static const ReduceEntry ReduceTable[] = { 61 // Wide, Narrow1, Narrow2, imm1,imm2, lo1, lo2, P/C,PF,S,AM 62 { ARM::t2ADCrr, 0, ARM::tADC, 0, 0, 0, 1, 0,0, 0,0,0 }, 63 { ARM::t2ADDri, ARM::tADDi3, ARM::tADDi8, 3, 8, 1, 1, 0,0, 0,1,0 }, 64 { ARM::t2ADDrr, ARM::tADDrr, ARM::tADDhirr, 0, 0, 1, 0, 0,1, 0,0,0 }, 65 { ARM::t2ADDSri,ARM::tADDi3, ARM::tADDi8, 3, 8, 1, 1, 2,2, 0,1,0 }, 66 { ARM::t2ADDSrr,ARM::tADDrr, 0, 0, 0, 1, 0, 2,0, 0,1,0 }, 67 { ARM::t2ANDrr, 0, ARM::tAND, 0, 0, 0, 1, 0,0, 1,0,0 }, 68 { ARM::t2ASRri, ARM::tASRri, 0, 5, 0, 1, 0, 0,0, 1,0,1 }, 69 { ARM::t2ASRrr, 0, ARM::tASRrr, 0, 0, 0, 1, 0,0, 1,0,1 }, 70 { ARM::t2BICrr, 0, ARM::tBIC, 0, 0, 0, 1, 0,0, 1,0,0 }, 71 //FIXME: Disable CMN, as CCodes are backwards from compare expectations 72 //{ ARM::t2CMNrr, ARM::tCMN, 0, 0, 0, 1, 0, 2,0, 0,0,0 }, 73 { ARM::t2CMNzrr, ARM::tCMNz, 0, 0, 0, 1, 0, 2,0, 0,0,0 }, 74 { ARM::t2CMPri, ARM::tCMPi8, 0, 8, 0, 1, 0, 2,0, 0,0,0 }, 75 { ARM::t2CMPrr, ARM::tCMPhir, 0, 0, 0, 0, 0, 2,0, 0,1,0 }, 76 { ARM::t2EORrr, 0, ARM::tEOR, 0, 0, 0, 1, 0,0, 1,0,0 }, 77 // FIXME: adr.n immediate offset must be multiple of 4. 78 //{ ARM::t2LEApcrelJT,ARM::tLEApcrelJT, 0, 0, 0, 1, 0, 1,0, 0,0,0 }, 79 { ARM::t2LSLri, ARM::tLSLri, 0, 5, 0, 1, 0, 0,0, 1,0,1 }, 80 { ARM::t2LSLrr, 0, ARM::tLSLrr, 0, 0, 0, 1, 0,0, 1,0,1 }, 81 { ARM::t2LSRri, ARM::tLSRri, 0, 5, 0, 1, 0, 0,0, 1,0,1 }, 82 { ARM::t2LSRrr, 0, ARM::tLSRrr, 0, 0, 0, 1, 0,0, 1,0,1 }, 83 { ARM::t2MOVi, ARM::tMOVi8, 0, 8, 0, 1, 0, 0,0, 1,0,0 }, 84 { ARM::t2MOVi16,ARM::tMOVi8, 0, 8, 0, 1, 0, 0,0, 1,1,0 }, 85 // FIXME: Do we need the 16-bit 'S' variant? 86 { ARM::t2MOVr,ARM::tMOVr, 0, 0, 0, 0, 0, 1,0, 0,0,0 }, 87 { ARM::t2MUL, 0, ARM::tMUL, 0, 0, 0, 1, 0,0, 1,0,0 }, 88 { ARM::t2MVNr, ARM::tMVN, 0, 0, 0, 1, 0, 0,0, 0,0,0 }, 89 { ARM::t2ORRrr, 0, ARM::tORR, 0, 0, 0, 1, 0,0, 1,0,0 }, 90 { ARM::t2REV, ARM::tREV, 0, 0, 0, 1, 0, 1,0, 0,0,0 }, 91 { ARM::t2REV16, ARM::tREV16, 0, 0, 0, 1, 0, 1,0, 0,0,0 }, 92 { ARM::t2REVSH, ARM::tREVSH, 0, 0, 0, 1, 0, 1,0, 0,0,0 }, 93 { ARM::t2RORrr, 0, ARM::tROR, 0, 0, 0, 1, 0,0, 1,0,0 }, 94 { ARM::t2RSBri, ARM::tRSB, 0, 0, 0, 1, 0, 0,0, 0,1,0 }, 95 { ARM::t2RSBSri,ARM::tRSB, 0, 0, 0, 1, 0, 2,0, 0,1,0 }, 96 { ARM::t2SBCrr, 0, ARM::tSBC, 0, 0, 0, 1, 0,0, 0,0,0 }, 97 { ARM::t2SUBri, ARM::tSUBi3, ARM::tSUBi8, 3, 8, 1, 1, 0,0, 0,0,0 }, 98 { ARM::t2SUBrr, ARM::tSUBrr, 0, 0, 0, 1, 0, 0,0, 0,0,0 }, 99 { ARM::t2SUBSri,ARM::tSUBi3, ARM::tSUBi8, 3, 8, 1, 1, 2,2, 0,0,0 }, 100 { ARM::t2SUBSrr,ARM::tSUBrr, 0, 0, 0, 1, 0, 2,0, 0,0,0 }, 101 { ARM::t2SXTB, ARM::tSXTB, 0, 0, 0, 1, 0, 1,0, 0,1,0 }, 102 { ARM::t2SXTH, ARM::tSXTH, 0, 0, 0, 1, 0, 1,0, 0,1,0 }, 103 { ARM::t2TSTrr, ARM::tTST, 0, 0, 0, 1, 0, 2,0, 0,0,0 }, 104 { ARM::t2UXTB, ARM::tUXTB, 0, 0, 0, 1, 0, 1,0, 0,1,0 }, 105 { ARM::t2UXTH, ARM::tUXTH, 0, 0, 0, 1, 0, 1,0, 0,1,0 }, 106 107 // FIXME: Clean this up after splitting each Thumb load / store opcode 108 // into multiple ones. 109 { ARM::t2LDRi12,ARM::tLDRi, ARM::tLDRspi, 5, 8, 1, 0, 0,0, 0,1,0 }, 110 { ARM::t2LDRs, ARM::tLDRr, 0, 0, 0, 1, 0, 0,0, 0,1,0 }, 111 { ARM::t2LDRBi12,ARM::tLDRBi, 0, 5, 0, 1, 0, 0,0, 0,1,0 }, 112 { ARM::t2LDRBs, ARM::tLDRBr, 0, 0, 0, 1, 0, 0,0, 0,1,0 }, 113 { ARM::t2LDRHi12,ARM::tLDRHi, 0, 5, 0, 1, 0, 0,0, 0,1,0 }, 114 { ARM::t2LDRHs, ARM::tLDRHr, 0, 0, 0, 1, 0, 0,0, 0,1,0 }, 115 { ARM::t2LDRSBs,ARM::tLDRSB, 0, 0, 0, 1, 0, 0,0, 0,1,0 }, 116 { ARM::t2LDRSHs,ARM::tLDRSH, 0, 0, 0, 1, 0, 0,0, 0,1,0 }, 117 { ARM::t2STRi12,ARM::tSTRi, ARM::tSTRspi, 5, 8, 1, 0, 0,0, 0,1,0 }, 118 { ARM::t2STRs, ARM::tSTRr, 0, 0, 0, 1, 0, 0,0, 0,1,0 }, 119 { ARM::t2STRBi12,ARM::tSTRBi, 0, 5, 0, 1, 0, 0,0, 0,1,0 }, 120 { ARM::t2STRBs, ARM::tSTRBr, 0, 0, 0, 1, 0, 0,0, 0,1,0 }, 121 { ARM::t2STRHi12,ARM::tSTRHi, 0, 5, 0, 1, 0, 0,0, 0,1,0 }, 122 { ARM::t2STRHs, ARM::tSTRHr, 0, 0, 0, 1, 0, 0,0, 0,1,0 }, 123 124 { ARM::t2LDMIA, ARM::tLDMIA, 0, 0, 0, 1, 1, 1,1, 0,1,0 }, 125 { ARM::t2LDMIA_RET,0, ARM::tPOP_RET, 0, 0, 1, 1, 1,1, 0,1,0 }, 126 { ARM::t2LDMIA_UPD,ARM::tLDMIA_UPD,ARM::tPOP,0, 0, 1, 1, 1,1, 0,1,0 }, 127 // ARM::t2STM (with no basereg writeback) has no Thumb1 equivalent 128 { ARM::t2STMIA_UPD,ARM::tSTMIA_UPD, 0, 0, 0, 1, 1, 1,1, 0,1,0 }, 129 { ARM::t2STMDB_UPD, 0, ARM::tPUSH, 0, 0, 1, 1, 1,1, 0,1,0 } 130 }; 131 132 class Thumb2SizeReduce : public MachineFunctionPass { 133 public: 134 static char ID; 135 Thumb2SizeReduce(); 136 137 const Thumb2InstrInfo *TII; 138 const ARMSubtarget *STI; 139 140 bool runOnMachineFunction(MachineFunction &MF) override; 141 142 const char *getPassName() const override { 143 return "Thumb2 instruction size reduction pass"; 144 } 145 146 private: 147 /// ReduceOpcodeMap - Maps wide opcode to index of entry in ReduceTable. 148 DenseMap<unsigned, unsigned> ReduceOpcodeMap; 149 150 bool canAddPseudoFlagDep(MachineInstr *Use, bool IsSelfLoop); 151 152 bool VerifyPredAndCC(MachineInstr *MI, const ReduceEntry &Entry, 153 bool is2Addr, ARMCC::CondCodes Pred, 154 bool LiveCPSR, bool &HasCC, bool &CCDead); 155 156 bool ReduceLoadStore(MachineBasicBlock &MBB, MachineInstr *MI, 157 const ReduceEntry &Entry); 158 159 bool ReduceSpecial(MachineBasicBlock &MBB, MachineInstr *MI, 160 const ReduceEntry &Entry, bool LiveCPSR, bool IsSelfLoop); 161 162 /// ReduceTo2Addr - Reduce a 32-bit instruction to a 16-bit two-address 163 /// instruction. 164 bool ReduceTo2Addr(MachineBasicBlock &MBB, MachineInstr *MI, 165 const ReduceEntry &Entry, bool LiveCPSR, 166 bool IsSelfLoop); 167 168 /// ReduceToNarrow - Reduce a 32-bit instruction to a 16-bit 169 /// non-two-address instruction. 170 bool ReduceToNarrow(MachineBasicBlock &MBB, MachineInstr *MI, 171 const ReduceEntry &Entry, bool LiveCPSR, 172 bool IsSelfLoop); 173 174 /// ReduceMI - Attempt to reduce MI, return true on success. 175 bool ReduceMI(MachineBasicBlock &MBB, MachineInstr *MI, 176 bool LiveCPSR, bool IsSelfLoop); 177 178 /// ReduceMBB - Reduce width of instructions in the specified basic block. 179 bool ReduceMBB(MachineBasicBlock &MBB); 180 181 bool OptimizeSize; 182 bool MinimizeSize; 183 184 // Last instruction to define CPSR in the current block. 185 MachineInstr *CPSRDef; 186 // Was CPSR last defined by a high latency instruction? 187 // When CPSRDef is null, this refers to CPSR defs in predecessors. 188 bool HighLatencyCPSR; 189 190 struct MBBInfo { 191 // The flags leaving this block have high latency. 192 bool HighLatencyCPSR; 193 // Has this block been visited yet? 194 bool Visited; 195 196 MBBInfo() : HighLatencyCPSR(false), Visited(false) {} 197 }; 198 199 SmallVector<MBBInfo, 8> BlockInfo; 200 }; 201 char Thumb2SizeReduce::ID = 0; 202 } 203 204 Thumb2SizeReduce::Thumb2SizeReduce() : MachineFunctionPass(ID) { 205 OptimizeSize = MinimizeSize = false; 206 for (unsigned i = 0, e = array_lengthof(ReduceTable); i != e; ++i) { 207 unsigned FromOpc = ReduceTable[i].WideOpc; 208 if (!ReduceOpcodeMap.insert(std::make_pair(FromOpc, i)).second) 209 assert(false && "Duplicated entries?"); 210 } 211 } 212 213 static bool HasImplicitCPSRDef(const MCInstrDesc &MCID) { 214 for (const uint16_t *Regs = MCID.getImplicitDefs(); *Regs; ++Regs) 215 if (*Regs == ARM::CPSR) 216 return true; 217 return false; 218 } 219 220 // Check for a likely high-latency flag def. 221 static bool isHighLatencyCPSR(MachineInstr *Def) { 222 switch(Def->getOpcode()) { 223 case ARM::FMSTAT: 224 case ARM::tMUL: 225 return true; 226 } 227 return false; 228 } 229 230 /// canAddPseudoFlagDep - For A9 (and other out-of-order) implementations, 231 /// the 's' 16-bit instruction partially update CPSR. Abort the 232 /// transformation to avoid adding false dependency on last CPSR setting 233 /// instruction which hurts the ability for out-of-order execution engine 234 /// to do register renaming magic. 235 /// This function checks if there is a read-of-write dependency between the 236 /// last instruction that defines the CPSR and the current instruction. If there 237 /// is, then there is no harm done since the instruction cannot be retired 238 /// before the CPSR setting instruction anyway. 239 /// Note, we are not doing full dependency analysis here for the sake of compile 240 /// time. We're not looking for cases like: 241 /// r0 = muls ... 242 /// r1 = add.w r0, ... 243 /// ... 244 /// = mul.w r1 245 /// In this case it would have been ok to narrow the mul.w to muls since there 246 /// are indirect RAW dependency between the muls and the mul.w 247 bool 248 Thumb2SizeReduce::canAddPseudoFlagDep(MachineInstr *Use, bool FirstInSelfLoop) { 249 // Disable the check for -Oz (aka OptimizeForSizeHarder). 250 if (MinimizeSize || !STI->avoidCPSRPartialUpdate()) 251 return false; 252 253 if (!CPSRDef) 254 // If this BB loops back to itself, conservatively avoid narrowing the 255 // first instruction that does partial flag update. 256 return HighLatencyCPSR || FirstInSelfLoop; 257 258 SmallSet<unsigned, 2> Defs; 259 for (const MachineOperand &MO : CPSRDef->operands()) { 260 if (!MO.isReg() || MO.isUndef() || MO.isUse()) 261 continue; 262 unsigned Reg = MO.getReg(); 263 if (Reg == 0 || Reg == ARM::CPSR) 264 continue; 265 Defs.insert(Reg); 266 } 267 268 for (const MachineOperand &MO : Use->operands()) { 269 if (!MO.isReg() || MO.isUndef() || MO.isDef()) 270 continue; 271 unsigned Reg = MO.getReg(); 272 if (Defs.count(Reg)) 273 return false; 274 } 275 276 // If the current CPSR has high latency, try to avoid the false dependency. 277 if (HighLatencyCPSR) 278 return true; 279 280 // tMOVi8 usually doesn't start long dependency chains, and there are a lot 281 // of them, so always shrink them when CPSR doesn't have high latency. 282 if (Use->getOpcode() == ARM::t2MOVi || 283 Use->getOpcode() == ARM::t2MOVi16) 284 return false; 285 286 // No read-after-write dependency. The narrowing will add false dependency. 287 return true; 288 } 289 290 bool 291 Thumb2SizeReduce::VerifyPredAndCC(MachineInstr *MI, const ReduceEntry &Entry, 292 bool is2Addr, ARMCC::CondCodes Pred, 293 bool LiveCPSR, bool &HasCC, bool &CCDead) { 294 if ((is2Addr && Entry.PredCC2 == 0) || 295 (!is2Addr && Entry.PredCC1 == 0)) { 296 if (Pred == ARMCC::AL) { 297 // Not predicated, must set CPSR. 298 if (!HasCC) { 299 // Original instruction was not setting CPSR, but CPSR is not 300 // currently live anyway. It's ok to set it. The CPSR def is 301 // dead though. 302 if (!LiveCPSR) { 303 HasCC = true; 304 CCDead = true; 305 return true; 306 } 307 return false; 308 } 309 } else { 310 // Predicated, must not set CPSR. 311 if (HasCC) 312 return false; 313 } 314 } else if ((is2Addr && Entry.PredCC2 == 2) || 315 (!is2Addr && Entry.PredCC1 == 2)) { 316 /// Old opcode has an optional def of CPSR. 317 if (HasCC) 318 return true; 319 // If old opcode does not implicitly define CPSR, then it's not ok since 320 // these new opcodes' CPSR def is not meant to be thrown away. e.g. CMP. 321 if (!HasImplicitCPSRDef(MI->getDesc())) 322 return false; 323 HasCC = true; 324 } else { 325 // 16-bit instruction does not set CPSR. 326 if (HasCC) 327 return false; 328 } 329 330 return true; 331 } 332 333 static bool VerifyLowRegs(MachineInstr *MI) { 334 unsigned Opc = MI->getOpcode(); 335 bool isPCOk = (Opc == ARM::t2LDMIA_RET || Opc == ARM::t2LDMIA || 336 Opc == ARM::t2LDMDB || Opc == ARM::t2LDMIA_UPD || 337 Opc == ARM::t2LDMDB_UPD); 338 bool isLROk = (Opc == ARM::t2STMIA_UPD || Opc == ARM::t2STMDB_UPD); 339 bool isSPOk = isPCOk || isLROk; 340 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 341 const MachineOperand &MO = MI->getOperand(i); 342 if (!MO.isReg() || MO.isImplicit()) 343 continue; 344 unsigned Reg = MO.getReg(); 345 if (Reg == 0 || Reg == ARM::CPSR) 346 continue; 347 if (isPCOk && Reg == ARM::PC) 348 continue; 349 if (isLROk && Reg == ARM::LR) 350 continue; 351 if (Reg == ARM::SP) { 352 if (isSPOk) 353 continue; 354 if (i == 1 && (Opc == ARM::t2LDRi12 || Opc == ARM::t2STRi12)) 355 // Special case for these ldr / str with sp as base register. 356 continue; 357 } 358 if (!isARMLowRegister(Reg)) 359 return false; 360 } 361 return true; 362 } 363 364 bool 365 Thumb2SizeReduce::ReduceLoadStore(MachineBasicBlock &MBB, MachineInstr *MI, 366 const ReduceEntry &Entry) { 367 if (ReduceLimitLdSt != -1 && ((int)NumLdSts >= ReduceLimitLdSt)) 368 return false; 369 370 unsigned Scale = 1; 371 bool HasImmOffset = false; 372 bool HasShift = false; 373 bool HasOffReg = true; 374 bool isLdStMul = false; 375 unsigned Opc = Entry.NarrowOpc1; 376 unsigned OpNum = 3; // First 'rest' of operands. 377 uint8_t ImmLimit = Entry.Imm1Limit; 378 379 switch (Entry.WideOpc) { 380 default: 381 llvm_unreachable("Unexpected Thumb2 load / store opcode!"); 382 case ARM::t2LDRi12: 383 case ARM::t2STRi12: 384 if (MI->getOperand(1).getReg() == ARM::SP) { 385 Opc = Entry.NarrowOpc2; 386 ImmLimit = Entry.Imm2Limit; 387 HasOffReg = false; 388 } 389 390 Scale = 4; 391 HasImmOffset = true; 392 HasOffReg = false; 393 break; 394 case ARM::t2LDRBi12: 395 case ARM::t2STRBi12: 396 HasImmOffset = true; 397 HasOffReg = false; 398 break; 399 case ARM::t2LDRHi12: 400 case ARM::t2STRHi12: 401 Scale = 2; 402 HasImmOffset = true; 403 HasOffReg = false; 404 break; 405 case ARM::t2LDRs: 406 case ARM::t2LDRBs: 407 case ARM::t2LDRHs: 408 case ARM::t2LDRSBs: 409 case ARM::t2LDRSHs: 410 case ARM::t2STRs: 411 case ARM::t2STRBs: 412 case ARM::t2STRHs: 413 HasShift = true; 414 OpNum = 4; 415 break; 416 case ARM::t2LDMIA: 417 case ARM::t2LDMDB: { 418 unsigned BaseReg = MI->getOperand(0).getReg(); 419 if (!isARMLowRegister(BaseReg) || Entry.WideOpc != ARM::t2LDMIA) 420 return false; 421 422 // For the non-writeback version (this one), the base register must be 423 // one of the registers being loaded. 424 bool isOK = false; 425 for (unsigned i = 4; i < MI->getNumOperands(); ++i) { 426 if (MI->getOperand(i).getReg() == BaseReg) { 427 isOK = true; 428 break; 429 } 430 } 431 432 if (!isOK) 433 return false; 434 435 OpNum = 0; 436 isLdStMul = true; 437 break; 438 } 439 case ARM::t2LDMIA_RET: { 440 unsigned BaseReg = MI->getOperand(1).getReg(); 441 if (BaseReg != ARM::SP) 442 return false; 443 Opc = Entry.NarrowOpc2; // tPOP_RET 444 OpNum = 2; 445 isLdStMul = true; 446 break; 447 } 448 case ARM::t2LDMIA_UPD: 449 case ARM::t2LDMDB_UPD: 450 case ARM::t2STMIA_UPD: 451 case ARM::t2STMDB_UPD: { 452 OpNum = 0; 453 454 unsigned BaseReg = MI->getOperand(1).getReg(); 455 if (BaseReg == ARM::SP && 456 (Entry.WideOpc == ARM::t2LDMIA_UPD || 457 Entry.WideOpc == ARM::t2STMDB_UPD)) { 458 Opc = Entry.NarrowOpc2; // tPOP or tPUSH 459 OpNum = 2; 460 } else if (!isARMLowRegister(BaseReg) || 461 (Entry.WideOpc != ARM::t2LDMIA_UPD && 462 Entry.WideOpc != ARM::t2STMIA_UPD)) { 463 return false; 464 } 465 466 isLdStMul = true; 467 break; 468 } 469 } 470 471 unsigned OffsetReg = 0; 472 bool OffsetKill = false; 473 if (HasShift) { 474 OffsetReg = MI->getOperand(2).getReg(); 475 OffsetKill = MI->getOperand(2).isKill(); 476 477 if (MI->getOperand(3).getImm()) 478 // Thumb1 addressing mode doesn't support shift. 479 return false; 480 } 481 482 unsigned OffsetImm = 0; 483 if (HasImmOffset) { 484 OffsetImm = MI->getOperand(2).getImm(); 485 unsigned MaxOffset = ((1 << ImmLimit) - 1) * Scale; 486 487 if ((OffsetImm & (Scale - 1)) || OffsetImm > MaxOffset) 488 // Make sure the immediate field fits. 489 return false; 490 } 491 492 // Add the 16-bit load / store instruction. 493 DebugLoc dl = MI->getDebugLoc(); 494 MachineInstrBuilder MIB = BuildMI(MBB, MI, dl, TII->get(Opc)); 495 if (!isLdStMul) { 496 MIB.addOperand(MI->getOperand(0)); 497 MIB.addOperand(MI->getOperand(1)); 498 499 if (HasImmOffset) 500 MIB.addImm(OffsetImm / Scale); 501 502 assert((!HasShift || OffsetReg) && "Invalid so_reg load / store address!"); 503 504 if (HasOffReg) 505 MIB.addReg(OffsetReg, getKillRegState(OffsetKill)); 506 } 507 508 // Transfer the rest of operands. 509 for (unsigned e = MI->getNumOperands(); OpNum != e; ++OpNum) 510 MIB.addOperand(MI->getOperand(OpNum)); 511 512 // Transfer memoperands. 513 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end()); 514 515 // Transfer MI flags. 516 MIB.setMIFlags(MI->getFlags()); 517 518 DEBUG(errs() << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB); 519 520 MBB.erase_instr(MI); 521 ++NumLdSts; 522 return true; 523 } 524 525 bool 526 Thumb2SizeReduce::ReduceSpecial(MachineBasicBlock &MBB, MachineInstr *MI, 527 const ReduceEntry &Entry, 528 bool LiveCPSR, bool IsSelfLoop) { 529 unsigned Opc = MI->getOpcode(); 530 if (Opc == ARM::t2ADDri) { 531 // If the source register is SP, try to reduce to tADDrSPi, otherwise 532 // it's a normal reduce. 533 if (MI->getOperand(1).getReg() != ARM::SP) { 534 if (ReduceTo2Addr(MBB, MI, Entry, LiveCPSR, IsSelfLoop)) 535 return true; 536 return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, IsSelfLoop); 537 } 538 // Try to reduce to tADDrSPi. 539 unsigned Imm = MI->getOperand(2).getImm(); 540 // The immediate must be in range, the destination register must be a low 541 // reg, the predicate must be "always" and the condition flags must not 542 // be being set. 543 if (Imm & 3 || Imm > 1020) 544 return false; 545 if (!isARMLowRegister(MI->getOperand(0).getReg())) 546 return false; 547 if (MI->getOperand(3).getImm() != ARMCC::AL) 548 return false; 549 const MCInstrDesc &MCID = MI->getDesc(); 550 if (MCID.hasOptionalDef() && 551 MI->getOperand(MCID.getNumOperands()-1).getReg() == ARM::CPSR) 552 return false; 553 554 MachineInstrBuilder MIB = BuildMI(MBB, MI, MI->getDebugLoc(), 555 TII->get(ARM::tADDrSPi)) 556 .addOperand(MI->getOperand(0)) 557 .addOperand(MI->getOperand(1)) 558 .addImm(Imm / 4); // The tADDrSPi has an implied scale by four. 559 AddDefaultPred(MIB); 560 561 // Transfer MI flags. 562 MIB.setMIFlags(MI->getFlags()); 563 564 DEBUG(errs() << "Converted 32-bit: " << *MI << " to 16-bit: " <<*MIB); 565 566 MBB.erase_instr(MI); 567 ++NumNarrows; 568 return true; 569 } 570 571 if (Entry.LowRegs1 && !VerifyLowRegs(MI)) 572 return false; 573 574 if (MI->mayLoad() || MI->mayStore()) 575 return ReduceLoadStore(MBB, MI, Entry); 576 577 switch (Opc) { 578 default: break; 579 case ARM::t2ADDSri: 580 case ARM::t2ADDSrr: { 581 unsigned PredReg = 0; 582 if (getInstrPredicate(MI, PredReg) == ARMCC::AL) { 583 switch (Opc) { 584 default: break; 585 case ARM::t2ADDSri: { 586 if (ReduceTo2Addr(MBB, MI, Entry, LiveCPSR, IsSelfLoop)) 587 return true; 588 // fallthrough 589 } 590 case ARM::t2ADDSrr: 591 return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, IsSelfLoop); 592 } 593 } 594 break; 595 } 596 case ARM::t2RSBri: 597 case ARM::t2RSBSri: 598 case ARM::t2SXTB: 599 case ARM::t2SXTH: 600 case ARM::t2UXTB: 601 case ARM::t2UXTH: 602 if (MI->getOperand(2).getImm() == 0) 603 return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, IsSelfLoop); 604 break; 605 case ARM::t2MOVi16: 606 // Can convert only 'pure' immediate operands, not immediates obtained as 607 // globals' addresses. 608 if (MI->getOperand(1).isImm()) 609 return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, IsSelfLoop); 610 break; 611 case ARM::t2CMPrr: { 612 // Try to reduce to the lo-reg only version first. Why there are two 613 // versions of the instruction is a mystery. 614 // It would be nice to just have two entries in the master table that 615 // are prioritized, but the table assumes a unique entry for each 616 // source insn opcode. So for now, we hack a local entry record to use. 617 static const ReduceEntry NarrowEntry = 618 { ARM::t2CMPrr,ARM::tCMPr, 0, 0, 0, 1, 1,2, 0, 0,1,0 }; 619 if (ReduceToNarrow(MBB, MI, NarrowEntry, LiveCPSR, IsSelfLoop)) 620 return true; 621 return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, IsSelfLoop); 622 } 623 } 624 return false; 625 } 626 627 bool 628 Thumb2SizeReduce::ReduceTo2Addr(MachineBasicBlock &MBB, MachineInstr *MI, 629 const ReduceEntry &Entry, 630 bool LiveCPSR, bool IsSelfLoop) { 631 632 if (ReduceLimit2Addr != -1 && ((int)Num2Addrs >= ReduceLimit2Addr)) 633 return false; 634 635 if (!MinimizeSize && !OptimizeSize && Entry.AvoidMovs && 636 STI->avoidMOVsShifterOperand()) 637 // Don't issue movs with shifter operand for some CPUs unless we 638 // are optimizing / minimizing for size. 639 return false; 640 641 unsigned Reg0 = MI->getOperand(0).getReg(); 642 unsigned Reg1 = MI->getOperand(1).getReg(); 643 // t2MUL is "special". The tied source operand is second, not first. 644 if (MI->getOpcode() == ARM::t2MUL) { 645 unsigned Reg2 = MI->getOperand(2).getReg(); 646 // Early exit if the regs aren't all low regs. 647 if (!isARMLowRegister(Reg0) || !isARMLowRegister(Reg1) 648 || !isARMLowRegister(Reg2)) 649 return false; 650 if (Reg0 != Reg2) { 651 // If the other operand also isn't the same as the destination, we 652 // can't reduce. 653 if (Reg1 != Reg0) 654 return false; 655 // Try to commute the operands to make it a 2-address instruction. 656 MachineInstr *CommutedMI = TII->commuteInstruction(MI); 657 if (!CommutedMI) 658 return false; 659 } 660 } else if (Reg0 != Reg1) { 661 // Try to commute the operands to make it a 2-address instruction. 662 unsigned CommOpIdx1, CommOpIdx2; 663 if (!TII->findCommutedOpIndices(MI, CommOpIdx1, CommOpIdx2) || 664 CommOpIdx1 != 1 || MI->getOperand(CommOpIdx2).getReg() != Reg0) 665 return false; 666 MachineInstr *CommutedMI = TII->commuteInstruction(MI); 667 if (!CommutedMI) 668 return false; 669 } 670 if (Entry.LowRegs2 && !isARMLowRegister(Reg0)) 671 return false; 672 if (Entry.Imm2Limit) { 673 unsigned Imm = MI->getOperand(2).getImm(); 674 unsigned Limit = (1 << Entry.Imm2Limit) - 1; 675 if (Imm > Limit) 676 return false; 677 } else { 678 unsigned Reg2 = MI->getOperand(2).getReg(); 679 if (Entry.LowRegs2 && !isARMLowRegister(Reg2)) 680 return false; 681 } 682 683 // Check if it's possible / necessary to transfer the predicate. 684 const MCInstrDesc &NewMCID = TII->get(Entry.NarrowOpc2); 685 unsigned PredReg = 0; 686 ARMCC::CondCodes Pred = getInstrPredicate(MI, PredReg); 687 bool SkipPred = false; 688 if (Pred != ARMCC::AL) { 689 if (!NewMCID.isPredicable()) 690 // Can't transfer predicate, fail. 691 return false; 692 } else { 693 SkipPred = !NewMCID.isPredicable(); 694 } 695 696 bool HasCC = false; 697 bool CCDead = false; 698 const MCInstrDesc &MCID = MI->getDesc(); 699 if (MCID.hasOptionalDef()) { 700 unsigned NumOps = MCID.getNumOperands(); 701 HasCC = (MI->getOperand(NumOps-1).getReg() == ARM::CPSR); 702 if (HasCC && MI->getOperand(NumOps-1).isDead()) 703 CCDead = true; 704 } 705 if (!VerifyPredAndCC(MI, Entry, true, Pred, LiveCPSR, HasCC, CCDead)) 706 return false; 707 708 // Avoid adding a false dependency on partial flag update by some 16-bit 709 // instructions which has the 's' bit set. 710 if (Entry.PartFlag && NewMCID.hasOptionalDef() && HasCC && 711 canAddPseudoFlagDep(MI, IsSelfLoop)) 712 return false; 713 714 // Add the 16-bit instruction. 715 DebugLoc dl = MI->getDebugLoc(); 716 MachineInstrBuilder MIB = BuildMI(MBB, MI, dl, NewMCID); 717 MIB.addOperand(MI->getOperand(0)); 718 if (NewMCID.hasOptionalDef()) { 719 if (HasCC) 720 AddDefaultT1CC(MIB, CCDead); 721 else 722 AddNoT1CC(MIB); 723 } 724 725 // Transfer the rest of operands. 726 unsigned NumOps = MCID.getNumOperands(); 727 for (unsigned i = 1, e = MI->getNumOperands(); i != e; ++i) { 728 if (i < NumOps && MCID.OpInfo[i].isOptionalDef()) 729 continue; 730 if (SkipPred && MCID.OpInfo[i].isPredicate()) 731 continue; 732 MIB.addOperand(MI->getOperand(i)); 733 } 734 735 // Transfer MI flags. 736 MIB.setMIFlags(MI->getFlags()); 737 738 DEBUG(errs() << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB); 739 740 MBB.erase_instr(MI); 741 ++Num2Addrs; 742 return true; 743 } 744 745 bool 746 Thumb2SizeReduce::ReduceToNarrow(MachineBasicBlock &MBB, MachineInstr *MI, 747 const ReduceEntry &Entry, 748 bool LiveCPSR, bool IsSelfLoop) { 749 if (ReduceLimit != -1 && ((int)NumNarrows >= ReduceLimit)) 750 return false; 751 752 if (!MinimizeSize && !OptimizeSize && Entry.AvoidMovs && 753 STI->avoidMOVsShifterOperand()) 754 // Don't issue movs with shifter operand for some CPUs unless we 755 // are optimizing / minimizing for size. 756 return false; 757 758 unsigned Limit = ~0U; 759 if (Entry.Imm1Limit) 760 Limit = (1 << Entry.Imm1Limit) - 1; 761 762 const MCInstrDesc &MCID = MI->getDesc(); 763 for (unsigned i = 0, e = MCID.getNumOperands(); i != e; ++i) { 764 if (MCID.OpInfo[i].isPredicate()) 765 continue; 766 const MachineOperand &MO = MI->getOperand(i); 767 if (MO.isReg()) { 768 unsigned Reg = MO.getReg(); 769 if (!Reg || Reg == ARM::CPSR) 770 continue; 771 if (Entry.LowRegs1 && !isARMLowRegister(Reg)) 772 return false; 773 } else if (MO.isImm() && 774 !MCID.OpInfo[i].isPredicate()) { 775 if (((unsigned)MO.getImm()) > Limit) 776 return false; 777 } 778 } 779 780 // Check if it's possible / necessary to transfer the predicate. 781 const MCInstrDesc &NewMCID = TII->get(Entry.NarrowOpc1); 782 unsigned PredReg = 0; 783 ARMCC::CondCodes Pred = getInstrPredicate(MI, PredReg); 784 bool SkipPred = false; 785 if (Pred != ARMCC::AL) { 786 if (!NewMCID.isPredicable()) 787 // Can't transfer predicate, fail. 788 return false; 789 } else { 790 SkipPred = !NewMCID.isPredicable(); 791 } 792 793 bool HasCC = false; 794 bool CCDead = false; 795 if (MCID.hasOptionalDef()) { 796 unsigned NumOps = MCID.getNumOperands(); 797 HasCC = (MI->getOperand(NumOps-1).getReg() == ARM::CPSR); 798 if (HasCC && MI->getOperand(NumOps-1).isDead()) 799 CCDead = true; 800 } 801 if (!VerifyPredAndCC(MI, Entry, false, Pred, LiveCPSR, HasCC, CCDead)) 802 return false; 803 804 // Avoid adding a false dependency on partial flag update by some 16-bit 805 // instructions which has the 's' bit set. 806 if (Entry.PartFlag && NewMCID.hasOptionalDef() && HasCC && 807 canAddPseudoFlagDep(MI, IsSelfLoop)) 808 return false; 809 810 // Add the 16-bit instruction. 811 DebugLoc dl = MI->getDebugLoc(); 812 MachineInstrBuilder MIB = BuildMI(MBB, MI, dl, NewMCID); 813 MIB.addOperand(MI->getOperand(0)); 814 if (NewMCID.hasOptionalDef()) { 815 if (HasCC) 816 AddDefaultT1CC(MIB, CCDead); 817 else 818 AddNoT1CC(MIB); 819 } 820 821 // Transfer the rest of operands. 822 unsigned NumOps = MCID.getNumOperands(); 823 for (unsigned i = 1, e = MI->getNumOperands(); i != e; ++i) { 824 if (i < NumOps && MCID.OpInfo[i].isOptionalDef()) 825 continue; 826 if ((MCID.getOpcode() == ARM::t2RSBSri || 827 MCID.getOpcode() == ARM::t2RSBri || 828 MCID.getOpcode() == ARM::t2SXTB || 829 MCID.getOpcode() == ARM::t2SXTH || 830 MCID.getOpcode() == ARM::t2UXTB || 831 MCID.getOpcode() == ARM::t2UXTH) && i == 2) 832 // Skip the zero immediate operand, it's now implicit. 833 continue; 834 bool isPred = (i < NumOps && MCID.OpInfo[i].isPredicate()); 835 if (SkipPred && isPred) 836 continue; 837 const MachineOperand &MO = MI->getOperand(i); 838 if (MO.isReg() && MO.isImplicit() && MO.getReg() == ARM::CPSR) 839 // Skip implicit def of CPSR. Either it's modeled as an optional 840 // def now or it's already an implicit def on the new instruction. 841 continue; 842 MIB.addOperand(MO); 843 } 844 if (!MCID.isPredicable() && NewMCID.isPredicable()) 845 AddDefaultPred(MIB); 846 847 // Transfer MI flags. 848 MIB.setMIFlags(MI->getFlags()); 849 850 DEBUG(errs() << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB); 851 852 MBB.erase_instr(MI); 853 ++NumNarrows; 854 return true; 855 } 856 857 static bool UpdateCPSRDef(MachineInstr &MI, bool LiveCPSR, bool &DefCPSR) { 858 bool HasDef = false; 859 for (const MachineOperand &MO : MI.operands()) { 860 if (!MO.isReg() || MO.isUndef() || MO.isUse()) 861 continue; 862 if (MO.getReg() != ARM::CPSR) 863 continue; 864 865 DefCPSR = true; 866 if (!MO.isDead()) 867 HasDef = true; 868 } 869 870 return HasDef || LiveCPSR; 871 } 872 873 static bool UpdateCPSRUse(MachineInstr &MI, bool LiveCPSR) { 874 for (const MachineOperand &MO : MI.operands()) { 875 if (!MO.isReg() || MO.isUndef() || MO.isDef()) 876 continue; 877 if (MO.getReg() != ARM::CPSR) 878 continue; 879 assert(LiveCPSR && "CPSR liveness tracking is wrong!"); 880 if (MO.isKill()) { 881 LiveCPSR = false; 882 break; 883 } 884 } 885 886 return LiveCPSR; 887 } 888 889 bool Thumb2SizeReduce::ReduceMI(MachineBasicBlock &MBB, MachineInstr *MI, 890 bool LiveCPSR, bool IsSelfLoop) { 891 unsigned Opcode = MI->getOpcode(); 892 DenseMap<unsigned, unsigned>::iterator OPI = ReduceOpcodeMap.find(Opcode); 893 if (OPI == ReduceOpcodeMap.end()) 894 return false; 895 const ReduceEntry &Entry = ReduceTable[OPI->second]; 896 897 // Don't attempt normal reductions on "special" cases for now. 898 if (Entry.Special) 899 return ReduceSpecial(MBB, MI, Entry, LiveCPSR, IsSelfLoop); 900 901 // Try to transform to a 16-bit two-address instruction. 902 if (Entry.NarrowOpc2 && 903 ReduceTo2Addr(MBB, MI, Entry, LiveCPSR, IsSelfLoop)) 904 return true; 905 906 // Try to transform to a 16-bit non-two-address instruction. 907 if (Entry.NarrowOpc1 && 908 ReduceToNarrow(MBB, MI, Entry, LiveCPSR, IsSelfLoop)) 909 return true; 910 911 return false; 912 } 913 914 bool Thumb2SizeReduce::ReduceMBB(MachineBasicBlock &MBB) { 915 bool Modified = false; 916 917 // Yes, CPSR could be livein. 918 bool LiveCPSR = MBB.isLiveIn(ARM::CPSR); 919 MachineInstr *BundleMI = nullptr; 920 921 CPSRDef = nullptr; 922 HighLatencyCPSR = false; 923 924 // Check predecessors for the latest CPSRDef. 925 for (auto *Pred : MBB.predecessors()) { 926 const MBBInfo &PInfo = BlockInfo[Pred->getNumber()]; 927 if (!PInfo.Visited) { 928 // Since blocks are visited in RPO, this must be a back-edge. 929 continue; 930 } 931 if (PInfo.HighLatencyCPSR) { 932 HighLatencyCPSR = true; 933 break; 934 } 935 } 936 937 // If this BB loops back to itself, conservatively avoid narrowing the 938 // first instruction that does partial flag update. 939 bool IsSelfLoop = MBB.isSuccessor(&MBB); 940 MachineBasicBlock::instr_iterator MII = MBB.instr_begin(),E = MBB.instr_end(); 941 MachineBasicBlock::instr_iterator NextMII; 942 for (; MII != E; MII = NextMII) { 943 NextMII = std::next(MII); 944 945 MachineInstr *MI = &*MII; 946 if (MI->isBundle()) { 947 BundleMI = MI; 948 continue; 949 } 950 if (MI->isDebugValue()) 951 continue; 952 953 LiveCPSR = UpdateCPSRUse(*MI, LiveCPSR); 954 955 // Does NextMII belong to the same bundle as MI? 956 bool NextInSameBundle = NextMII != E && NextMII->isBundledWithPred(); 957 958 if (ReduceMI(MBB, MI, LiveCPSR, IsSelfLoop)) { 959 Modified = true; 960 MachineBasicBlock::instr_iterator I = std::prev(NextMII); 961 MI = &*I; 962 // Removing and reinserting the first instruction in a bundle will break 963 // up the bundle. Fix the bundling if it was broken. 964 if (NextInSameBundle && !NextMII->isBundledWithPred()) 965 NextMII->bundleWithPred(); 966 } 967 968 if (!NextInSameBundle && MI->isInsideBundle()) { 969 // FIXME: Since post-ra scheduler operates on bundles, the CPSR kill 970 // marker is only on the BUNDLE instruction. Process the BUNDLE 971 // instruction as we finish with the bundled instruction to work around 972 // the inconsistency. 973 if (BundleMI->killsRegister(ARM::CPSR)) 974 LiveCPSR = false; 975 MachineOperand *MO = BundleMI->findRegisterDefOperand(ARM::CPSR); 976 if (MO && !MO->isDead()) 977 LiveCPSR = true; 978 MO = BundleMI->findRegisterUseOperand(ARM::CPSR); 979 if (MO && !MO->isKill()) 980 LiveCPSR = true; 981 } 982 983 bool DefCPSR = false; 984 LiveCPSR = UpdateCPSRDef(*MI, LiveCPSR, DefCPSR); 985 if (MI->isCall()) { 986 // Calls don't really set CPSR. 987 CPSRDef = nullptr; 988 HighLatencyCPSR = false; 989 IsSelfLoop = false; 990 } else if (DefCPSR) { 991 // This is the last CPSR defining instruction. 992 CPSRDef = MI; 993 HighLatencyCPSR = isHighLatencyCPSR(CPSRDef); 994 IsSelfLoop = false; 995 } 996 } 997 998 MBBInfo &Info = BlockInfo[MBB.getNumber()]; 999 Info.HighLatencyCPSR = HighLatencyCPSR; 1000 Info.Visited = true; 1001 return Modified; 1002 } 1003 1004 bool Thumb2SizeReduce::runOnMachineFunction(MachineFunction &MF) { 1005 const TargetMachine &TM = MF.getTarget(); 1006 TII = static_cast<const Thumb2InstrInfo*>(TM.getInstrInfo()); 1007 STI = &TM.getSubtarget<ARMSubtarget>(); 1008 1009 // Optimizing / minimizing size? 1010 AttributeSet FnAttrs = MF.getFunction()->getAttributes(); 1011 OptimizeSize = FnAttrs.hasAttribute(AttributeSet::FunctionIndex, 1012 Attribute::OptimizeForSize); 1013 MinimizeSize = 1014 FnAttrs.hasAttribute(AttributeSet::FunctionIndex, Attribute::MinSize); 1015 1016 BlockInfo.clear(); 1017 BlockInfo.resize(MF.getNumBlockIDs()); 1018 1019 // Visit blocks in reverse post-order so LastCPSRDef is known for all 1020 // predecessors. 1021 ReversePostOrderTraversal<MachineFunction*> RPOT(&MF); 1022 bool Modified = false; 1023 for (ReversePostOrderTraversal<MachineFunction*>::rpo_iterator 1024 I = RPOT.begin(), E = RPOT.end(); I != E; ++I) 1025 Modified |= ReduceMBB(**I); 1026 return Modified; 1027 } 1028 1029 /// createThumb2SizeReductionPass - Returns an instance of the Thumb2 size 1030 /// reduction pass. 1031 FunctionPass *llvm::createThumb2SizeReductionPass() { 1032 return new Thumb2SizeReduce(); 1033 } 1034