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 /// 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/LiveIntervals.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->isDebugInstr()) { 138 I = std::next(I); 139 continue; 140 } 141 142 if (I->mayStore() || I->isBarrier() || I->isCall() || 143 I->hasUnmodeledSideEffects() || I->hasOrderedMemoryRef()) 144 break; 145 146 LLVM_DEBUG(dbgs() 147 << "Removing no effect instruction: " << *I << '\n'); 148 149 for (auto &Op : I->operands()) { 150 if (Op.isReg()) 151 RecalcRegs.insert(Op.getReg()); 152 } 153 154 auto Next = std::next(I); 155 LIS->RemoveMachineInstrFromMaps(*I); 156 I->eraseFromParent(); 157 I = Next; 158 159 Changed = true; 160 } 161 162 if (I != E) 163 continue; 164 165 // Try to ascend predecessors. 166 for (auto *Pred : CurBB->predecessors()) { 167 if (Pred->succ_size() == 1) 168 Blocks.push_back(Pred); 169 } 170 } 171 continue; 172 } 173 174 // Try to collapse adjacent endifs. 175 auto Lead = MBB.begin(), E = MBB.end(); 176 if (MBB.succ_size() != 1 || Lead == E || !isEndCF(*Lead, TRI)) 177 continue; 178 179 const MachineBasicBlock* Succ = *MBB.succ_begin(); 180 if (!MBB.isLayoutSuccessor(Succ)) 181 continue; 182 183 auto I = std::next(Lead); 184 185 for ( ; I != E; ++I) 186 if (!TII->isSALU(*I) || I->readsRegister(AMDGPU::EXEC, TRI)) 187 break; 188 189 if (I != E) 190 continue; 191 192 const auto NextLead = Succ->begin(); 193 if (NextLead == Succ->end() || !isEndCF(*NextLead, TRI) || 194 !getOrExecSource(*NextLead, *TII, MRI)) 195 continue; 196 197 LLVM_DEBUG(dbgs() << "Redundant EXEC = S_OR_B64 found: " << *Lead << '\n'); 198 199 auto SaveExec = getOrExecSource(*Lead, *TII, MRI); 200 unsigned SaveExecReg = getOrNonExecReg(*Lead, *TII); 201 for (auto &Op : Lead->operands()) { 202 if (Op.isReg()) 203 RecalcRegs.insert(Op.getReg()); 204 } 205 206 LIS->RemoveMachineInstrFromMaps(*Lead); 207 Lead->eraseFromParent(); 208 if (SaveExecReg) { 209 LIS->removeInterval(SaveExecReg); 210 LIS->createAndComputeVirtRegInterval(SaveExecReg); 211 } 212 213 Changed = true; 214 215 // If the only use of saved exec in the removed instruction is S_AND_B64 216 // fold the copy now. 217 if (!SaveExec || !SaveExec->isFullCopy()) 218 continue; 219 220 unsigned SavedExec = SaveExec->getOperand(0).getReg(); 221 bool SafeToReplace = true; 222 for (auto& U : MRI.use_nodbg_instructions(SavedExec)) { 223 if (U.getParent() != SaveExec->getParent()) { 224 SafeToReplace = false; 225 break; 226 } 227 228 LLVM_DEBUG(dbgs() << "Redundant EXEC COPY: " << *SaveExec << '\n'); 229 } 230 231 if (SafeToReplace) { 232 LIS->RemoveMachineInstrFromMaps(*SaveExec); 233 SaveExec->eraseFromParent(); 234 MRI.replaceRegWith(SavedExec, AMDGPU::EXEC); 235 LIS->removeInterval(SavedExec); 236 } 237 } 238 239 if (Changed) { 240 for (auto Reg : RecalcRegs) { 241 if (TargetRegisterInfo::isVirtualRegister(Reg)) { 242 LIS->removeInterval(Reg); 243 if (!MRI.reg_empty(Reg)) 244 LIS->createAndComputeVirtRegInterval(Reg); 245 } else { 246 for (MCRegUnitIterator U(Reg, TRI); U.isValid(); ++U) 247 LIS->removeRegUnit(*U); 248 } 249 } 250 } 251 252 return Changed; 253 } 254