1 //===-- MLxExpansionPass.cpp - Expand MLx instrs to avoid hazards ---------===// 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 // Expand VFP / NEON floating point MLA / MLS instructions (each to a pair of 11 // multiple and add / sub instructions) when special VMLx hazards are detected. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #define DEBUG_TYPE "mlx-expansion" 16 #include "ARM.h" 17 #include "ARMBaseInstrInfo.h" 18 #include "ARMSubtarget.h" 19 #include "llvm/ADT/SmallPtrSet.h" 20 #include "llvm/ADT/Statistic.h" 21 #include "llvm/CodeGen/MachineFunctionPass.h" 22 #include "llvm/CodeGen/MachineInstr.h" 23 #include "llvm/CodeGen/MachineInstrBuilder.h" 24 #include "llvm/CodeGen/MachineRegisterInfo.h" 25 #include "llvm/Support/CommandLine.h" 26 #include "llvm/Support/Debug.h" 27 #include "llvm/Support/raw_ostream.h" 28 #include "llvm/Target/TargetRegisterInfo.h" 29 using namespace llvm; 30 31 static cl::opt<bool> 32 ForceExapnd("expand-all-fp-mlx", cl::init(false), cl::Hidden); 33 static cl::opt<unsigned> 34 ExpandLimit("expand-limit", cl::init(~0U), cl::Hidden); 35 36 STATISTIC(NumExpand, "Number of fp MLA / MLS instructions expanded"); 37 38 namespace { 39 struct MLxExpansion : public MachineFunctionPass { 40 static char ID; 41 MLxExpansion() : MachineFunctionPass(ID) {} 42 43 virtual bool runOnMachineFunction(MachineFunction &Fn); 44 45 virtual const char *getPassName() const { 46 return "ARM MLA / MLS expansion pass"; 47 } 48 49 private: 50 const ARMBaseInstrInfo *TII; 51 const TargetRegisterInfo *TRI; 52 MachineRegisterInfo *MRI; 53 54 bool isLikeA9; 55 bool isSwift; 56 unsigned MIIdx; 57 MachineInstr* LastMIs[4]; 58 SmallPtrSet<MachineInstr*, 4> IgnoreStall; 59 60 void clearStack(); 61 void pushStack(MachineInstr *MI); 62 MachineInstr *getAccDefMI(MachineInstr *MI) const; 63 unsigned getDefReg(MachineInstr *MI) const; 64 bool hasLoopHazard(MachineInstr *MI) const; 65 bool hasRAWHazard(unsigned Reg, MachineInstr *MI) const; 66 bool FindMLxHazard(MachineInstr *MI); 67 void ExpandFPMLxInstruction(MachineBasicBlock &MBB, MachineInstr *MI, 68 unsigned MulOpc, unsigned AddSubOpc, 69 bool NegAcc, bool HasLane); 70 bool ExpandFPMLxInstructions(MachineBasicBlock &MBB); 71 }; 72 char MLxExpansion::ID = 0; 73 } 74 75 void MLxExpansion::clearStack() { 76 std::fill(LastMIs, LastMIs + 4, (MachineInstr*)0); 77 MIIdx = 0; 78 } 79 80 void MLxExpansion::pushStack(MachineInstr *MI) { 81 LastMIs[MIIdx] = MI; 82 if (++MIIdx == 4) 83 MIIdx = 0; 84 } 85 86 MachineInstr *MLxExpansion::getAccDefMI(MachineInstr *MI) const { 87 // Look past COPY and INSERT_SUBREG instructions to find the 88 // real definition MI. This is important for _sfp instructions. 89 unsigned Reg = MI->getOperand(1).getReg(); 90 if (TargetRegisterInfo::isPhysicalRegister(Reg)) 91 return 0; 92 93 MachineBasicBlock *MBB = MI->getParent(); 94 MachineInstr *DefMI = MRI->getVRegDef(Reg); 95 while (true) { 96 if (DefMI->getParent() != MBB) 97 break; 98 if (DefMI->isCopyLike()) { 99 Reg = DefMI->getOperand(1).getReg(); 100 if (TargetRegisterInfo::isVirtualRegister(Reg)) { 101 DefMI = MRI->getVRegDef(Reg); 102 continue; 103 } 104 } else if (DefMI->isInsertSubreg()) { 105 Reg = DefMI->getOperand(2).getReg(); 106 if (TargetRegisterInfo::isVirtualRegister(Reg)) { 107 DefMI = MRI->getVRegDef(Reg); 108 continue; 109 } 110 } 111 break; 112 } 113 return DefMI; 114 } 115 116 unsigned MLxExpansion::getDefReg(MachineInstr *MI) const { 117 unsigned Reg = MI->getOperand(0).getReg(); 118 if (TargetRegisterInfo::isPhysicalRegister(Reg) || 119 !MRI->hasOneNonDBGUse(Reg)) 120 return Reg; 121 122 MachineBasicBlock *MBB = MI->getParent(); 123 MachineInstr *UseMI = &*MRI->use_nodbg_begin(Reg); 124 if (UseMI->getParent() != MBB) 125 return Reg; 126 127 while (UseMI->isCopy() || UseMI->isInsertSubreg()) { 128 Reg = UseMI->getOperand(0).getReg(); 129 if (TargetRegisterInfo::isPhysicalRegister(Reg) || 130 !MRI->hasOneNonDBGUse(Reg)) 131 return Reg; 132 UseMI = &*MRI->use_nodbg_begin(Reg); 133 if (UseMI->getParent() != MBB) 134 return Reg; 135 } 136 137 return Reg; 138 } 139 140 /// hasLoopHazard - Check whether an MLx instruction is chained to itself across 141 /// a single-MBB loop. 142 bool MLxExpansion::hasLoopHazard(MachineInstr *MI) const { 143 unsigned Reg = MI->getOperand(1).getReg(); 144 if (TargetRegisterInfo::isPhysicalRegister(Reg)) 145 return false; 146 147 MachineBasicBlock *MBB = MI->getParent(); 148 MachineInstr *DefMI = MRI->getVRegDef(Reg); 149 while (true) { 150 outer_continue: 151 if (DefMI->getParent() != MBB) 152 break; 153 154 if (DefMI->isPHI()) { 155 for (unsigned i = 1, e = DefMI->getNumOperands(); i < e; i += 2) { 156 if (DefMI->getOperand(i + 1).getMBB() == MBB) { 157 unsigned SrcReg = DefMI->getOperand(i).getReg(); 158 if (TargetRegisterInfo::isVirtualRegister(SrcReg)) { 159 DefMI = MRI->getVRegDef(SrcReg); 160 goto outer_continue; 161 } 162 } 163 } 164 } else if (DefMI->isCopyLike()) { 165 Reg = DefMI->getOperand(1).getReg(); 166 if (TargetRegisterInfo::isVirtualRegister(Reg)) { 167 DefMI = MRI->getVRegDef(Reg); 168 continue; 169 } 170 } else if (DefMI->isInsertSubreg()) { 171 Reg = DefMI->getOperand(2).getReg(); 172 if (TargetRegisterInfo::isVirtualRegister(Reg)) { 173 DefMI = MRI->getVRegDef(Reg); 174 continue; 175 } 176 } 177 178 break; 179 } 180 181 return DefMI == MI; 182 } 183 184 bool MLxExpansion::hasRAWHazard(unsigned Reg, MachineInstr *MI) const { 185 // FIXME: Detect integer instructions properly. 186 const MCInstrDesc &MCID = MI->getDesc(); 187 unsigned Domain = MCID.TSFlags & ARMII::DomainMask; 188 if (MI->mayStore()) 189 return false; 190 unsigned Opcode = MCID.getOpcode(); 191 if (Opcode == ARM::VMOVRS || Opcode == ARM::VMOVRRD) 192 return false; 193 if ((Domain & ARMII::DomainVFP) || (Domain & ARMII::DomainNEON)) 194 return MI->readsRegister(Reg, TRI); 195 return false; 196 } 197 198 static bool isFpMulInstruction(unsigned Opcode) { 199 switch (Opcode) { 200 case ARM::VMULS: 201 case ARM::VMULfd: 202 case ARM::VMULfq: 203 case ARM::VMULD: 204 case ARM::VMULslfd: 205 case ARM::VMULslfq: 206 return true; 207 default: 208 return false; 209 } 210 } 211 212 bool MLxExpansion::FindMLxHazard(MachineInstr *MI) { 213 if (NumExpand >= ExpandLimit) 214 return false; 215 216 if (ForceExapnd) 217 return true; 218 219 MachineInstr *DefMI = getAccDefMI(MI); 220 if (TII->isFpMLxInstruction(DefMI->getOpcode())) { 221 // r0 = vmla 222 // r3 = vmla r0, r1, r2 223 // takes 16 - 17 cycles 224 // 225 // r0 = vmla 226 // r4 = vmul r1, r2 227 // r3 = vadd r0, r4 228 // takes about 14 - 15 cycles even with vmul stalling for 4 cycles. 229 IgnoreStall.insert(DefMI); 230 return true; 231 } 232 233 // On Swift, we mostly care about hazards from multiplication instructions 234 // writing the accumulator and the pipelining of loop iterations by out-of- 235 // order execution. 236 if (isSwift) 237 return isFpMulInstruction(DefMI->getOpcode()) || hasLoopHazard(MI); 238 239 if (IgnoreStall.count(MI)) 240 return false; 241 242 // If a VMLA.F is followed by an VADD.F or VMUL.F with no RAW hazard, the 243 // VADD.F or VMUL.F will stall 4 cycles before issue. The 4 cycle stall 244 // preserves the in-order retirement of the instructions. 245 // Look at the next few instructions, if *most* of them can cause hazards, 246 // then the scheduler can't *fix* this, we'd better break up the VMLA. 247 unsigned Limit1 = isLikeA9 ? 1 : 4; 248 unsigned Limit2 = isLikeA9 ? 1 : 4; 249 for (unsigned i = 1; i <= 4; ++i) { 250 int Idx = ((int)MIIdx - i + 4) % 4; 251 MachineInstr *NextMI = LastMIs[Idx]; 252 if (!NextMI) 253 continue; 254 255 if (TII->canCauseFpMLxStall(NextMI->getOpcode())) { 256 if (i <= Limit1) 257 return true; 258 } 259 260 // Look for VMLx RAW hazard. 261 if (i <= Limit2 && hasRAWHazard(getDefReg(MI), NextMI)) 262 return true; 263 } 264 265 return false; 266 } 267 268 /// ExpandFPMLxInstructions - Expand a MLA / MLS instruction into a pair 269 /// of MUL + ADD / SUB instructions. 270 void 271 MLxExpansion::ExpandFPMLxInstruction(MachineBasicBlock &MBB, MachineInstr *MI, 272 unsigned MulOpc, unsigned AddSubOpc, 273 bool NegAcc, bool HasLane) { 274 unsigned DstReg = MI->getOperand(0).getReg(); 275 bool DstDead = MI->getOperand(0).isDead(); 276 unsigned AccReg = MI->getOperand(1).getReg(); 277 unsigned Src1Reg = MI->getOperand(2).getReg(); 278 unsigned Src2Reg = MI->getOperand(3).getReg(); 279 bool Src1Kill = MI->getOperand(2).isKill(); 280 bool Src2Kill = MI->getOperand(3).isKill(); 281 unsigned LaneImm = HasLane ? MI->getOperand(4).getImm() : 0; 282 unsigned NextOp = HasLane ? 5 : 4; 283 ARMCC::CondCodes Pred = (ARMCC::CondCodes)MI->getOperand(NextOp).getImm(); 284 unsigned PredReg = MI->getOperand(++NextOp).getReg(); 285 286 const MCInstrDesc &MCID1 = TII->get(MulOpc); 287 const MCInstrDesc &MCID2 = TII->get(AddSubOpc); 288 const MachineFunction &MF = *MI->getParent()->getParent(); 289 unsigned TmpReg = MRI->createVirtualRegister( 290 TII->getRegClass(MCID1, 0, TRI, MF)); 291 292 MachineInstrBuilder MIB = BuildMI(MBB, MI, MI->getDebugLoc(), MCID1, TmpReg) 293 .addReg(Src1Reg, getKillRegState(Src1Kill)) 294 .addReg(Src2Reg, getKillRegState(Src2Kill)); 295 if (HasLane) 296 MIB.addImm(LaneImm); 297 MIB.addImm(Pred).addReg(PredReg); 298 299 MIB = BuildMI(MBB, MI, MI->getDebugLoc(), MCID2) 300 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstDead)); 301 302 if (NegAcc) { 303 bool AccKill = MRI->hasOneNonDBGUse(AccReg); 304 MIB.addReg(TmpReg, getKillRegState(true)) 305 .addReg(AccReg, getKillRegState(AccKill)); 306 } else { 307 MIB.addReg(AccReg).addReg(TmpReg, getKillRegState(true)); 308 } 309 MIB.addImm(Pred).addReg(PredReg); 310 311 DEBUG({ 312 dbgs() << "Expanding: " << *MI; 313 dbgs() << " to:\n"; 314 MachineBasicBlock::iterator MII = MI; 315 MII = llvm::prior(MII); 316 MachineInstr &MI2 = *MII; 317 MII = llvm::prior(MII); 318 MachineInstr &MI1 = *MII; 319 dbgs() << " " << MI1; 320 dbgs() << " " << MI2; 321 }); 322 323 MI->eraseFromParent(); 324 ++NumExpand; 325 } 326 327 bool MLxExpansion::ExpandFPMLxInstructions(MachineBasicBlock &MBB) { 328 bool Changed = false; 329 330 clearStack(); 331 IgnoreStall.clear(); 332 333 unsigned Skip = 0; 334 MachineBasicBlock::reverse_iterator MII = MBB.rbegin(), E = MBB.rend(); 335 while (MII != E) { 336 MachineInstr *MI = &*MII; 337 338 if (MI->isLabel() || MI->isImplicitDef() || MI->isCopy()) { 339 ++MII; 340 continue; 341 } 342 343 const MCInstrDesc &MCID = MI->getDesc(); 344 if (MI->isBarrier()) { 345 clearStack(); 346 Skip = 0; 347 ++MII; 348 continue; 349 } 350 351 unsigned Domain = MCID.TSFlags & ARMII::DomainMask; 352 if (Domain == ARMII::DomainGeneral) { 353 if (++Skip == 2) 354 // Assume dual issues of non-VFP / NEON instructions. 355 pushStack(0); 356 } else { 357 Skip = 0; 358 359 unsigned MulOpc, AddSubOpc; 360 bool NegAcc, HasLane; 361 if (!TII->isFpMLxInstruction(MCID.getOpcode(), 362 MulOpc, AddSubOpc, NegAcc, HasLane) || 363 !FindMLxHazard(MI)) 364 pushStack(MI); 365 else { 366 ExpandFPMLxInstruction(MBB, MI, MulOpc, AddSubOpc, NegAcc, HasLane); 367 E = MBB.rend(); // May have changed if MI was the 1st instruction. 368 Changed = true; 369 continue; 370 } 371 } 372 373 ++MII; 374 } 375 376 return Changed; 377 } 378 379 bool MLxExpansion::runOnMachineFunction(MachineFunction &Fn) { 380 TII = static_cast<const ARMBaseInstrInfo*>(Fn.getTarget().getInstrInfo()); 381 TRI = Fn.getTarget().getRegisterInfo(); 382 MRI = &Fn.getRegInfo(); 383 const ARMSubtarget *STI = &Fn.getTarget().getSubtarget<ARMSubtarget>(); 384 isLikeA9 = STI->isLikeA9() || STI->isSwift(); 385 isSwift = STI->isSwift(); 386 387 bool Modified = false; 388 for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E; 389 ++MFI) { 390 MachineBasicBlock &MBB = *MFI; 391 Modified |= ExpandFPMLxInstructions(MBB); 392 } 393 394 return Modified; 395 } 396 397 FunctionPass *llvm::createMLxExpansionPass() { 398 return new MLxExpansion(); 399 } 400