1 //===-- SILowerControlFlow.cpp - Use predicates for control flow ----------===// 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 /// \file 11 /// \brief Insert wait instructions for memory reads and writes. 12 /// 13 /// Memory reads and writes are issued asynchronously, so we need to insert 14 /// S_WAITCNT instructions when we want to access any of their results or 15 /// overwrite any register that's used asynchronously. 16 // 17 //===----------------------------------------------------------------------===// 18 19 #include "AMDGPU.h" 20 #include "SIInstrInfo.h" 21 #include "SIMachineFunctionInfo.h" 22 #include "llvm/CodeGen/MachineFunction.h" 23 #include "llvm/CodeGen/MachineFunctionPass.h" 24 #include "llvm/CodeGen/MachineInstrBuilder.h" 25 #include "llvm/CodeGen/MachineRegisterInfo.h" 26 27 using namespace llvm; 28 29 namespace { 30 31 /// \brief One variable for each of the hardware counters 32 typedef union { 33 struct { 34 unsigned VM; 35 unsigned EXP; 36 unsigned LGKM; 37 } Named; 38 unsigned Array[3]; 39 40 } Counters; 41 42 typedef Counters RegCounters[512]; 43 typedef std::pair<unsigned, unsigned> RegInterval; 44 45 class SIInsertWaits : public MachineFunctionPass { 46 47 private: 48 static char ID; 49 const SIInstrInfo *TII; 50 const SIRegisterInfo &TRI; 51 const MachineRegisterInfo *MRI; 52 53 /// \brief Constant hardware limits 54 static const Counters WaitCounts; 55 56 /// \brief Constant zero value 57 static const Counters ZeroCounts; 58 59 /// \brief Counter values we have already waited on. 60 Counters WaitedOn; 61 62 /// \brief Counter values for last instruction issued. 63 Counters LastIssued; 64 65 /// \brief Registers used by async instructions. 66 RegCounters UsedRegs; 67 68 /// \brief Registers defined by async instructions. 69 RegCounters DefinedRegs; 70 71 /// \brief Different export instruction types seen since last wait. 72 unsigned ExpInstrTypesSeen; 73 74 /// \brief Get increment/decrement amount for this instruction. 75 Counters getHwCounts(MachineInstr &MI); 76 77 /// \brief Is operand relevant for async execution? 78 bool isOpRelevant(MachineOperand &Op); 79 80 /// \brief Get register interval an operand affects. 81 RegInterval getRegInterval(MachineOperand &Op); 82 83 /// \brief Handle instructions async components 84 void pushInstruction(MachineInstr &MI); 85 86 /// \brief Insert the actual wait instruction 87 bool insertWait(MachineBasicBlock &MBB, 88 MachineBasicBlock::iterator I, 89 const Counters &Counts); 90 91 /// \brief Do we need def2def checks? 92 bool unorderedDefines(MachineInstr &MI); 93 94 /// \brief Resolve all operand dependencies to counter requirements 95 Counters handleOperands(MachineInstr &MI); 96 97 public: 98 SIInsertWaits(TargetMachine &tm) : 99 MachineFunctionPass(ID), 100 TII(static_cast<const SIInstrInfo*>(tm.getInstrInfo())), 101 TRI(TII->getRegisterInfo()) { } 102 103 virtual bool runOnMachineFunction(MachineFunction &MF); 104 105 const char *getPassName() const { 106 return "SI insert wait instructions"; 107 } 108 109 }; 110 111 } // End anonymous namespace 112 113 char SIInsertWaits::ID = 0; 114 115 const Counters SIInsertWaits::WaitCounts = { { 15, 7, 7 } }; 116 const Counters SIInsertWaits::ZeroCounts = { { 0, 0, 0 } }; 117 118 FunctionPass *llvm::createSIInsertWaits(TargetMachine &tm) { 119 return new SIInsertWaits(tm); 120 } 121 122 Counters SIInsertWaits::getHwCounts(MachineInstr &MI) { 123 124 uint64_t TSFlags = TII->get(MI.getOpcode()).TSFlags; 125 Counters Result; 126 127 Result.Named.VM = !!(TSFlags & SIInstrFlags::VM_CNT); 128 129 // Only consider stores or EXP for EXP_CNT 130 Result.Named.EXP = !!(TSFlags & SIInstrFlags::EXP_CNT && 131 (MI.getOpcode() == AMDGPU::EXP || MI.getDesc().mayStore())); 132 133 // LGKM may uses larger values 134 if (TSFlags & SIInstrFlags::LGKM_CNT) { 135 136 MachineOperand &Op = MI.getOperand(0); 137 assert(Op.isReg() && "First LGKM operand must be a register!"); 138 139 unsigned Reg = Op.getReg(); 140 unsigned Size = TRI.getMinimalPhysRegClass(Reg)->getSize(); 141 Result.Named.LGKM = Size > 4 ? 2 : 1; 142 143 } else { 144 Result.Named.LGKM = 0; 145 } 146 147 return Result; 148 } 149 150 bool SIInsertWaits::isOpRelevant(MachineOperand &Op) { 151 152 // Constants are always irrelevant 153 if (!Op.isReg()) 154 return false; 155 156 // Defines are always relevant 157 if (Op.isDef()) 158 return true; 159 160 // For exports all registers are relevant 161 MachineInstr &MI = *Op.getParent(); 162 if (MI.getOpcode() == AMDGPU::EXP) 163 return true; 164 165 // For stores the stored value is also relevant 166 if (!MI.getDesc().mayStore()) 167 return false; 168 169 for (MachineInstr::mop_iterator I = MI.operands_begin(), 170 E = MI.operands_end(); I != E; ++I) { 171 172 if (I->isReg() && I->isUse()) 173 return Op.isIdenticalTo(*I); 174 } 175 176 return false; 177 } 178 179 RegInterval SIInsertWaits::getRegInterval(MachineOperand &Op) { 180 181 if (!Op.isReg()) 182 return std::make_pair(0, 0); 183 184 unsigned Reg = Op.getReg(); 185 unsigned Size = TRI.getMinimalPhysRegClass(Reg)->getSize(); 186 187 assert(Size >= 4); 188 189 RegInterval Result; 190 Result.first = TRI.getEncodingValue(Reg); 191 Result.second = Result.first + Size / 4; 192 193 return Result; 194 } 195 196 void SIInsertWaits::pushInstruction(MachineInstr &MI) { 197 198 // Get the hardware counter increments and sum them up 199 Counters Increment = getHwCounts(MI); 200 unsigned Sum = 0; 201 202 for (unsigned i = 0; i < 3; ++i) { 203 LastIssued.Array[i] += Increment.Array[i]; 204 Sum += Increment.Array[i]; 205 } 206 207 // If we don't increase anything then that's it 208 if (Sum == 0) 209 return; 210 211 // Remember which export instructions we have seen 212 if (Increment.Named.EXP) { 213 ExpInstrTypesSeen |= MI.getOpcode() == AMDGPU::EXP ? 1 : 2; 214 } 215 216 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { 217 218 MachineOperand &Op = MI.getOperand(i); 219 if (!isOpRelevant(Op)) 220 continue; 221 222 RegInterval Interval = getRegInterval(Op); 223 for (unsigned j = Interval.first; j < Interval.second; ++j) { 224 225 // Remember which registers we define 226 if (Op.isDef()) 227 DefinedRegs[j] = LastIssued; 228 229 // and which one we are using 230 if (Op.isUse()) 231 UsedRegs[j] = LastIssued; 232 } 233 } 234 } 235 236 bool SIInsertWaits::insertWait(MachineBasicBlock &MBB, 237 MachineBasicBlock::iterator I, 238 const Counters &Required) { 239 240 // End of program? No need to wait on anything 241 if (I != MBB.end() && I->getOpcode() == AMDGPU::S_ENDPGM) 242 return false; 243 244 // Figure out if the async instructions execute in order 245 bool Ordered[3]; 246 247 // VM_CNT is always ordered 248 Ordered[0] = true; 249 250 // EXP_CNT is unordered if we have both EXP & VM-writes 251 Ordered[1] = ExpInstrTypesSeen == 3; 252 253 // LGKM_CNT is handled as always unordered. TODO: Handle LDS and GDS 254 Ordered[2] = false; 255 256 // The values we are going to put into the S_WAITCNT instruction 257 Counters Counts = WaitCounts; 258 259 // Do we really need to wait? 260 bool NeedWait = false; 261 262 for (unsigned i = 0; i < 3; ++i) { 263 264 if (Required.Array[i] <= WaitedOn.Array[i]) 265 continue; 266 267 NeedWait = true; 268 269 if (Ordered[i]) { 270 unsigned Value = LastIssued.Array[i] - Required.Array[i]; 271 272 // adjust the value to the real hardware posibilities 273 Counts.Array[i] = std::min(Value, WaitCounts.Array[i]); 274 275 } else 276 Counts.Array[i] = 0; 277 278 // Remember on what we have waited on 279 WaitedOn.Array[i] = LastIssued.Array[i] - Counts.Array[i]; 280 } 281 282 if (!NeedWait) 283 return false; 284 285 // Reset EXP_CNT instruction types 286 if (Counts.Named.EXP == 0) 287 ExpInstrTypesSeen = 0; 288 289 // Build the wait instruction 290 BuildMI(MBB, I, DebugLoc(), TII->get(AMDGPU::S_WAITCNT)) 291 .addImm((Counts.Named.VM & 0xF) | 292 ((Counts.Named.EXP & 0x7) << 4) | 293 ((Counts.Named.LGKM & 0x7) << 8)); 294 295 return true; 296 } 297 298 /// \brief helper function for handleOperands 299 static void increaseCounters(Counters &Dst, const Counters &Src) { 300 301 for (unsigned i = 0; i < 3; ++i) 302 Dst.Array[i] = std::max(Dst.Array[i], Src.Array[i]); 303 } 304 305 Counters SIInsertWaits::handleOperands(MachineInstr &MI) { 306 307 Counters Result = ZeroCounts; 308 309 // For each register affected by this 310 // instruction increase the result sequence 311 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { 312 313 MachineOperand &Op = MI.getOperand(i); 314 RegInterval Interval = getRegInterval(Op); 315 for (unsigned j = Interval.first; j < Interval.second; ++j) { 316 317 if (Op.isDef()) { 318 increaseCounters(Result, UsedRegs[j]); 319 increaseCounters(Result, DefinedRegs[j]); 320 } 321 322 if (Op.isUse()) 323 increaseCounters(Result, DefinedRegs[j]); 324 } 325 } 326 327 return Result; 328 } 329 330 bool SIInsertWaits::runOnMachineFunction(MachineFunction &MF) { 331 332 bool Changes = false; 333 334 MRI = &MF.getRegInfo(); 335 336 WaitedOn = ZeroCounts; 337 LastIssued = ZeroCounts; 338 339 memset(&UsedRegs, 0, sizeof(UsedRegs)); 340 memset(&DefinedRegs, 0, sizeof(DefinedRegs)); 341 342 for (MachineFunction::iterator BI = MF.begin(), BE = MF.end(); 343 BI != BE; ++BI) { 344 345 MachineBasicBlock &MBB = *BI; 346 for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end(); 347 I != E; ++I) { 348 349 Changes |= insertWait(MBB, I, handleOperands(*I)); 350 pushInstruction(*I); 351 } 352 353 // Wait for everything at the end of the MBB 354 Changes |= insertWait(MBB, MBB.getFirstTerminator(), LastIssued); 355 } 356 357 return Changes; 358 } 359