1 //===-- SIOptimizeExecMaskingPreRA.cpp ------------------------------------===// 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 This pass removes redundant S_OR_B64 instructions enabling lanes in 12 /// the exec. If two SI_END_CF (lowered as S_OR_B64) come together without any 13 /// vector instructions between them we can only keep outer SI_END_CF, given 14 /// that CFG is structured and exec bits of the outer end statement are always 15 /// not less than exec bit of the inner one. 16 /// 17 /// This needs to be done before the RA to eliminate saved exec bits registers 18 /// but after register coalescer to have no vector registers copies in between 19 /// of different end cf statements. 20 /// 21 //===----------------------------------------------------------------------===// 22 23 #include "AMDGPU.h" 24 #include "AMDGPUSubtarget.h" 25 #include "SIInstrInfo.h" 26 #include "llvm/CodeGen/LiveIntervalAnalysis.h" 27 #include "llvm/CodeGen/MachineFunctionPass.h" 28 29 using namespace llvm; 30 31 #define DEBUG_TYPE "si-optimize-exec-masking-pre-ra" 32 33 namespace { 34 35 class SIOptimizeExecMaskingPreRA : public MachineFunctionPass { 36 public: 37 static char ID; 38 39 public: 40 SIOptimizeExecMaskingPreRA() : MachineFunctionPass(ID) { 41 initializeSIOptimizeExecMaskingPreRAPass(*PassRegistry::getPassRegistry()); 42 } 43 44 bool runOnMachineFunction(MachineFunction &MF) override; 45 46 StringRef getPassName() const override { 47 return "SI optimize exec mask operations pre-RA"; 48 } 49 50 void getAnalysisUsage(AnalysisUsage &AU) const override { 51 AU.addRequired<LiveIntervals>(); 52 AU.setPreservesAll(); 53 MachineFunctionPass::getAnalysisUsage(AU); 54 } 55 }; 56 57 } // End anonymous namespace. 58 59 INITIALIZE_PASS_BEGIN(SIOptimizeExecMaskingPreRA, DEBUG_TYPE, 60 "SI optimize exec mask operations pre-RA", false, false) 61 INITIALIZE_PASS_DEPENDENCY(LiveIntervals) 62 INITIALIZE_PASS_END(SIOptimizeExecMaskingPreRA, DEBUG_TYPE, 63 "SI optimize exec mask operations pre-RA", false, false) 64 65 char SIOptimizeExecMaskingPreRA::ID = 0; 66 67 char &llvm::SIOptimizeExecMaskingPreRAID = SIOptimizeExecMaskingPreRA::ID; 68 69 FunctionPass *llvm::createSIOptimizeExecMaskingPreRAPass() { 70 return new SIOptimizeExecMaskingPreRA(); 71 } 72 73 static bool isEndCF(const MachineInstr& MI, const SIRegisterInfo* TRI) { 74 return MI.getOpcode() == AMDGPU::S_OR_B64 && 75 MI.modifiesRegister(AMDGPU::EXEC, TRI); 76 } 77 78 static bool isFullExecCopy(const MachineInstr& MI) { 79 return MI.isFullCopy() && MI.getOperand(1).getReg() == AMDGPU::EXEC; 80 } 81 82 static unsigned getOrNonExecReg(const MachineInstr &MI, 83 const SIInstrInfo &TII) { 84 auto Op = TII.getNamedOperand(MI, AMDGPU::OpName::src1); 85 if (Op->isReg() && Op->getReg() != AMDGPU::EXEC) 86 return Op->getReg(); 87 Op = TII.getNamedOperand(MI, AMDGPU::OpName::src0); 88 if (Op->isReg() && Op->getReg() != AMDGPU::EXEC) 89 return Op->getReg(); 90 return AMDGPU::NoRegister; 91 } 92 93 static MachineInstr* getOrExecSource(const MachineInstr &MI, 94 const SIInstrInfo &TII, 95 const MachineRegisterInfo &MRI) { 96 auto SavedExec = getOrNonExecReg(MI, TII); 97 if (SavedExec == AMDGPU::NoRegister) 98 return nullptr; 99 auto SaveExecInst = MRI.getUniqueVRegDef(SavedExec); 100 if (!SaveExecInst || !isFullExecCopy(*SaveExecInst)) 101 return nullptr; 102 return SaveExecInst; 103 } 104 105 bool SIOptimizeExecMaskingPreRA::runOnMachineFunction(MachineFunction &MF) { 106 if (skipFunction(*MF.getFunction())) 107 return false; 108 109 const SISubtarget &ST = MF.getSubtarget<SISubtarget>(); 110 const SIRegisterInfo *TRI = ST.getRegisterInfo(); 111 const SIInstrInfo *TII = ST.getInstrInfo(); 112 MachineRegisterInfo &MRI = MF.getRegInfo(); 113 LiveIntervals *LIS = &getAnalysis<LiveIntervals>(); 114 DenseSet<unsigned> RecalcRegs({AMDGPU::EXEC_LO, AMDGPU::EXEC_HI}); 115 bool Changed = false; 116 117 for (MachineBasicBlock &MBB : MF) { 118 119 // Try to remove unneeded instructions before s_endpgm. 120 if (MBB.succ_empty()) { 121 if (MBB.empty() || MBB.back().getOpcode() != AMDGPU::S_ENDPGM) 122 continue; 123 124 SmallVector<MachineBasicBlock*, 4> Blocks({&MBB}); 125 126 while (!Blocks.empty()) { 127 auto CurBB = Blocks.pop_back_val(); 128 auto I = CurBB->rbegin(), E = CurBB->rend(); 129 if (I != E) { 130 if (I->isUnconditionalBranch() || I->getOpcode() == AMDGPU::S_ENDPGM) 131 ++I; 132 else if (I->isBranch()) 133 continue; 134 } 135 136 while (I != E) { 137 if (I->isDebugValue()) 138 continue; 139 if (I->mayStore() || I->isBarrier() || I->isCall() || 140 I->hasUnmodeledSideEffects() || I->hasOrderedMemoryRef()) 141 break; 142 143 DEBUG(dbgs() << "Removing no effect instruction: " << *I << '\n'); 144 145 for (auto &Op : I->operands()) { 146 if (Op.isReg()) 147 RecalcRegs.insert(Op.getReg()); 148 } 149 150 auto Next = std::next(I); 151 LIS->RemoveMachineInstrFromMaps(*I); 152 I->eraseFromParent(); 153 I = Next; 154 155 Changed = true; 156 } 157 158 if (I != E) 159 continue; 160 161 // Try to ascend predecessors. 162 for (auto *Pred : CurBB->predecessors()) { 163 if (Pred->succ_size() == 1) 164 Blocks.push_back(Pred); 165 } 166 } 167 continue; 168 } 169 170 // Try to collapse adjacent endifs. 171 auto Lead = MBB.begin(), E = MBB.end(); 172 if (MBB.succ_size() != 1 || Lead == E || !isEndCF(*Lead, TRI)) 173 continue; 174 175 const MachineBasicBlock* Succ = *MBB.succ_begin(); 176 if (!MBB.isLayoutSuccessor(Succ)) 177 continue; 178 179 auto I = std::next(Lead); 180 181 for ( ; I != E; ++I) 182 if (!TII->isSALU(*I) || I->readsRegister(AMDGPU::EXEC, TRI)) 183 break; 184 185 if (I != E) 186 continue; 187 188 const auto NextLead = Succ->begin(); 189 if (NextLead == Succ->end() || !isEndCF(*NextLead, TRI) || 190 !getOrExecSource(*NextLead, *TII, MRI)) 191 continue; 192 193 DEBUG(dbgs() << "Redundant EXEC = S_OR_B64 found: " << *Lead << '\n'); 194 195 auto SaveExec = getOrExecSource(*Lead, *TII, MRI); 196 unsigned SaveExecReg = getOrNonExecReg(*Lead, *TII); 197 for (auto &Op : Lead->operands()) { 198 if (Op.isReg()) 199 RecalcRegs.insert(Op.getReg()); 200 } 201 202 LIS->RemoveMachineInstrFromMaps(*Lead); 203 Lead->eraseFromParent(); 204 if (SaveExecReg) { 205 LIS->removeInterval(SaveExecReg); 206 LIS->createAndComputeVirtRegInterval(SaveExecReg); 207 } 208 209 Changed = true; 210 211 // If the only use of saved exec in the removed instruction is S_AND_B64 212 // fold the copy now. 213 if (!SaveExec || !SaveExec->isFullCopy()) 214 continue; 215 216 unsigned SavedExec = SaveExec->getOperand(0).getReg(); 217 bool SafeToReplace = true; 218 for (auto& U : MRI.use_nodbg_instructions(SavedExec)) { 219 if (U.getParent() != SaveExec->getParent()) { 220 SafeToReplace = false; 221 break; 222 } 223 224 DEBUG(dbgs() << "Redundant EXEC COPY: " << *SaveExec << '\n'); 225 } 226 227 if (SafeToReplace) { 228 LIS->RemoveMachineInstrFromMaps(*SaveExec); 229 SaveExec->eraseFromParent(); 230 MRI.replaceRegWith(SavedExec, AMDGPU::EXEC); 231 LIS->removeInterval(SavedExec); 232 } 233 } 234 235 if (Changed) { 236 for (auto Reg : RecalcRegs) { 237 if (TargetRegisterInfo::isVirtualRegister(Reg)) { 238 LIS->removeInterval(Reg); 239 if (!MRI.reg_empty(Reg)) 240 LIS->createAndComputeVirtRegInterval(Reg); 241 } else { 242 for (MCRegUnitIterator U(Reg, TRI); U.isValid(); ++U) 243 LIS->removeRegUnit(*U); 244 } 245 } 246 } 247 248 return Changed; 249 } 250