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/CodeGen/MachineInstr.h" 20 #include "llvm/CodeGen/MachineInstrBuilder.h" 21 #include "llvm/CodeGen/MachineFunctionPass.h" 22 #include "llvm/CodeGen/MachineRegisterInfo.h" 23 #include "llvm/Target/TargetRegisterInfo.h" 24 #include "llvm/ADT/SmallPtrSet.h" 25 #include "llvm/ADT/Statistic.h" 26 #include "llvm/Support/CommandLine.h" 27 #include "llvm/Support/Debug.h" 28 #include "llvm/Support/raw_ostream.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 isA9; 55 unsigned MIIdx; 56 MachineInstr* LastMIs[4]; 57 SmallPtrSet<MachineInstr*, 4> IgnoreStall; 58 59 void clearStack(); 60 void pushStack(MachineInstr *MI); 61 MachineInstr *getAccDefMI(MachineInstr *MI) const; 62 unsigned getDefReg(MachineInstr *MI) const; 63 bool hasRAWHazard(unsigned Reg, MachineInstr *MI) const; 64 bool FindMLxHazard(MachineInstr *MI); 65 void ExpandFPMLxInstruction(MachineBasicBlock &MBB, MachineInstr *MI, 66 unsigned MulOpc, unsigned AddSubOpc, 67 bool NegAcc, bool HasLane); 68 bool ExpandFPMLxInstructions(MachineBasicBlock &MBB); 69 }; 70 char MLxExpansion::ID = 0; 71 } 72 73 void MLxExpansion::clearStack() { 74 std::fill(LastMIs, LastMIs + 4, (MachineInstr*)0); 75 MIIdx = 0; 76 } 77 78 void MLxExpansion::pushStack(MachineInstr *MI) { 79 LastMIs[MIIdx] = MI; 80 if (++MIIdx == 4) 81 MIIdx = 0; 82 } 83 84 MachineInstr *MLxExpansion::getAccDefMI(MachineInstr *MI) const { 85 // Look past COPY and INSERT_SUBREG instructions to find the 86 // real definition MI. This is important for _sfp instructions. 87 unsigned Reg = MI->getOperand(1).getReg(); 88 if (TargetRegisterInfo::isPhysicalRegister(Reg)) 89 return 0; 90 91 MachineBasicBlock *MBB = MI->getParent(); 92 MachineInstr *DefMI = MRI->getVRegDef(Reg); 93 while (true) { 94 if (DefMI->getParent() != MBB) 95 break; 96 if (DefMI->isCopyLike()) { 97 Reg = DefMI->getOperand(1).getReg(); 98 if (TargetRegisterInfo::isVirtualRegister(Reg)) { 99 DefMI = MRI->getVRegDef(Reg); 100 continue; 101 } 102 } else if (DefMI->isInsertSubreg()) { 103 Reg = DefMI->getOperand(2).getReg(); 104 if (TargetRegisterInfo::isVirtualRegister(Reg)) { 105 DefMI = MRI->getVRegDef(Reg); 106 continue; 107 } 108 } 109 break; 110 } 111 return DefMI; 112 } 113 114 unsigned MLxExpansion::getDefReg(MachineInstr *MI) const { 115 unsigned Reg = MI->getOperand(0).getReg(); 116 if (TargetRegisterInfo::isPhysicalRegister(Reg) || 117 !MRI->hasOneNonDBGUse(Reg)) 118 return Reg; 119 120 MachineBasicBlock *MBB = MI->getParent(); 121 MachineInstr *UseMI = &*MRI->use_nodbg_begin(Reg); 122 if (UseMI->getParent() != MBB) 123 return Reg; 124 125 while (UseMI->isCopy() || UseMI->isInsertSubreg()) { 126 Reg = UseMI->getOperand(0).getReg(); 127 if (TargetRegisterInfo::isPhysicalRegister(Reg) || 128 !MRI->hasOneNonDBGUse(Reg)) 129 return Reg; 130 UseMI = &*MRI->use_nodbg_begin(Reg); 131 if (UseMI->getParent() != MBB) 132 return Reg; 133 } 134 135 return Reg; 136 } 137 138 bool MLxExpansion::hasRAWHazard(unsigned Reg, MachineInstr *MI) const { 139 // FIXME: Detect integer instructions properly. 140 const MCInstrDesc &MCID = MI->getDesc(); 141 unsigned Domain = MCID.TSFlags & ARMII::DomainMask; 142 if (MCID.mayStore()) 143 return false; 144 unsigned Opcode = MCID.getOpcode(); 145 if (Opcode == ARM::VMOVRS || Opcode == ARM::VMOVRRD) 146 return false; 147 if ((Domain & ARMII::DomainVFP) || (Domain & ARMII::DomainNEON)) 148 return MI->readsRegister(Reg, TRI); 149 return false; 150 } 151 152 153 bool MLxExpansion::FindMLxHazard(MachineInstr *MI) { 154 if (NumExpand >= ExpandLimit) 155 return false; 156 157 if (ForceExapnd) 158 return true; 159 160 MachineInstr *DefMI = getAccDefMI(MI); 161 if (TII->isFpMLxInstruction(DefMI->getOpcode())) { 162 // r0 = vmla 163 // r3 = vmla r0, r1, r2 164 // takes 16 - 17 cycles 165 // 166 // r0 = vmla 167 // r4 = vmul r1, r2 168 // r3 = vadd r0, r4 169 // takes about 14 - 15 cycles even with vmul stalling for 4 cycles. 170 IgnoreStall.insert(DefMI); 171 return true; 172 } 173 174 if (IgnoreStall.count(MI)) 175 return false; 176 177 // If a VMLA.F is followed by an VADD.F or VMUL.F with no RAW hazard, the 178 // VADD.F or VMUL.F will stall 4 cycles before issue. The 4 cycle stall 179 // preserves the in-order retirement of the instructions. 180 // Look at the next few instructions, if *most* of them can cause hazards, 181 // then the scheduler can't *fix* this, we'd better break up the VMLA. 182 unsigned Limit1 = isA9 ? 1 : 4; 183 unsigned Limit2 = isA9 ? 1 : 4; 184 for (unsigned i = 1; i <= 4; ++i) { 185 int Idx = ((int)MIIdx - i + 4) % 4; 186 MachineInstr *NextMI = LastMIs[Idx]; 187 if (!NextMI) 188 continue; 189 190 if (TII->canCauseFpMLxStall(NextMI->getOpcode())) { 191 if (i <= Limit1) 192 return true; 193 } 194 195 // Look for VMLx RAW hazard. 196 if (i <= Limit2 && hasRAWHazard(getDefReg(MI), NextMI)) 197 return true; 198 } 199 200 return false; 201 } 202 203 /// ExpandFPMLxInstructions - Expand a MLA / MLS instruction into a pair 204 /// of MUL + ADD / SUB instructions. 205 void 206 MLxExpansion::ExpandFPMLxInstruction(MachineBasicBlock &MBB, MachineInstr *MI, 207 unsigned MulOpc, unsigned AddSubOpc, 208 bool NegAcc, bool HasLane) { 209 unsigned DstReg = MI->getOperand(0).getReg(); 210 bool DstDead = MI->getOperand(0).isDead(); 211 unsigned AccReg = MI->getOperand(1).getReg(); 212 unsigned Src1Reg = MI->getOperand(2).getReg(); 213 unsigned Src2Reg = MI->getOperand(3).getReg(); 214 bool Src1Kill = MI->getOperand(2).isKill(); 215 bool Src2Kill = MI->getOperand(3).isKill(); 216 unsigned LaneImm = HasLane ? MI->getOperand(4).getImm() : 0; 217 unsigned NextOp = HasLane ? 5 : 4; 218 ARMCC::CondCodes Pred = (ARMCC::CondCodes)MI->getOperand(NextOp).getImm(); 219 unsigned PredReg = MI->getOperand(++NextOp).getReg(); 220 221 const MCInstrDesc &MCID1 = TII->get(MulOpc); 222 const MCInstrDesc &MCID2 = TII->get(AddSubOpc); 223 unsigned TmpReg = MRI->createVirtualRegister(TII->getRegClass(MCID1, 0, TRI)); 224 225 MachineInstrBuilder MIB = BuildMI(MBB, *MI, MI->getDebugLoc(), MCID1, TmpReg) 226 .addReg(Src1Reg, getKillRegState(Src1Kill)) 227 .addReg(Src2Reg, getKillRegState(Src2Kill)); 228 if (HasLane) 229 MIB.addImm(LaneImm); 230 MIB.addImm(Pred).addReg(PredReg); 231 232 MIB = BuildMI(MBB, *MI, MI->getDebugLoc(), MCID2) 233 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstDead)); 234 235 if (NegAcc) { 236 bool AccKill = MRI->hasOneNonDBGUse(AccReg); 237 MIB.addReg(TmpReg, getKillRegState(true)) 238 .addReg(AccReg, getKillRegState(AccKill)); 239 } else { 240 MIB.addReg(AccReg).addReg(TmpReg, getKillRegState(true)); 241 } 242 MIB.addImm(Pred).addReg(PredReg); 243 244 DEBUG({ 245 dbgs() << "Expanding: " << *MI; 246 dbgs() << " to:\n"; 247 MachineBasicBlock::iterator MII = MI; 248 MII = llvm::prior(MII); 249 MachineInstr &MI2 = *MII; 250 MII = llvm::prior(MII); 251 MachineInstr &MI1 = *MII; 252 dbgs() << " " << MI1; 253 dbgs() << " " << MI2; 254 }); 255 256 MI->eraseFromParent(); 257 ++NumExpand; 258 } 259 260 bool MLxExpansion::ExpandFPMLxInstructions(MachineBasicBlock &MBB) { 261 bool Changed = false; 262 263 clearStack(); 264 IgnoreStall.clear(); 265 266 unsigned Skip = 0; 267 MachineBasicBlock::reverse_iterator MII = MBB.rbegin(), E = MBB.rend(); 268 while (MII != E) { 269 MachineInstr *MI = &*MII; 270 271 if (MI->isLabel() || MI->isImplicitDef() || MI->isCopy()) { 272 ++MII; 273 continue; 274 } 275 276 const MCInstrDesc &MCID = MI->getDesc(); 277 if (MCID.isBarrier()) { 278 clearStack(); 279 Skip = 0; 280 ++MII; 281 continue; 282 } 283 284 unsigned Domain = MCID.TSFlags & ARMII::DomainMask; 285 if (Domain == ARMII::DomainGeneral) { 286 if (++Skip == 2) 287 // Assume dual issues of non-VFP / NEON instructions. 288 pushStack(0); 289 } else { 290 Skip = 0; 291 292 unsigned MulOpc, AddSubOpc; 293 bool NegAcc, HasLane; 294 if (!TII->isFpMLxInstruction(MCID.getOpcode(), 295 MulOpc, AddSubOpc, NegAcc, HasLane) || 296 !FindMLxHazard(MI)) 297 pushStack(MI); 298 else { 299 ExpandFPMLxInstruction(MBB, MI, MulOpc, AddSubOpc, NegAcc, HasLane); 300 E = MBB.rend(); // May have changed if MI was the 1st instruction. 301 Changed = true; 302 continue; 303 } 304 } 305 306 ++MII; 307 } 308 309 return Changed; 310 } 311 312 bool MLxExpansion::runOnMachineFunction(MachineFunction &Fn) { 313 TII = static_cast<const ARMBaseInstrInfo*>(Fn.getTarget().getInstrInfo()); 314 TRI = Fn.getTarget().getRegisterInfo(); 315 MRI = &Fn.getRegInfo(); 316 const ARMSubtarget *STI = &Fn.getTarget().getSubtarget<ARMSubtarget>(); 317 isA9 = STI->isCortexA9(); 318 319 bool Modified = false; 320 for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E; 321 ++MFI) { 322 MachineBasicBlock &MBB = *MFI; 323 Modified |= ExpandFPMLxInstructions(MBB); 324 } 325 326 return Modified; 327 } 328 329 FunctionPass *llvm::createMLxExpansionPass() { 330 return new MLxExpansion(); 331 } 332