1 //===-- AArch64A53Fix835769.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 // This pass changes code to work around Cortex-A53 erratum 835769. 10 // It works around it by inserting a nop instruction in code sequences that 11 // in some circumstances may trigger the erratum. 12 // It inserts a nop instruction between a sequence of the following 2 classes 13 // of instructions: 14 // instr 1: mem-instr (including loads, stores and prefetches). 15 // instr 2: non-SIMD integer multiply-accumulate writing 64-bit X registers. 16 //===----------------------------------------------------------------------===// 17 18 #include "AArch64.h" 19 #include "llvm/ADT/Statistic.h" 20 #include "llvm/CodeGen/MachineFunction.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/Debug.h" 26 #include "llvm/Support/raw_ostream.h" 27 #include "llvm/Target/TargetInstrInfo.h" 28 29 using namespace llvm; 30 31 #define DEBUG_TYPE "aarch64-fix-cortex-a53-835769" 32 33 STATISTIC(NumNopsAdded, "Number of Nops added to work around erratum 835769"); 34 35 //===----------------------------------------------------------------------===// 36 // Helper functions 37 38 // Is the instruction a match for the instruction that comes first in the 39 // sequence of instructions that can trigger the erratum? 40 static bool isFirstInstructionInSequence(MachineInstr *MI) { 41 // Must return true if this instruction is a load, a store or a prefetch. 42 switch (MI->getOpcode()) { 43 case AArch64::PRFMl: 44 case AArch64::PRFMroW: 45 case AArch64::PRFMroX: 46 case AArch64::PRFMui: 47 case AArch64::PRFUMi: 48 return true; 49 default: 50 return MI->mayLoadOrStore(); 51 } 52 } 53 54 // Is the instruction a match for the instruction that comes second in the 55 // sequence that can trigger the erratum? 56 static bool isSecondInstructionInSequence(MachineInstr *MI) { 57 // Must return true for non-SIMD integer multiply-accumulates, writing 58 // to a 64-bit register. 59 switch (MI->getOpcode()) { 60 // Erratum cannot be triggered when the destination register is 32 bits, 61 // therefore only include the following. 62 case AArch64::MSUBXrrr: 63 case AArch64::MADDXrrr: 64 case AArch64::SMADDLrrr: 65 case AArch64::SMSUBLrrr: 66 case AArch64::UMADDLrrr: 67 case AArch64::UMSUBLrrr: 68 // Erratum can only be triggered by multiply-adds, not by regular 69 // non-accumulating multiplies, i.e. when Ra=XZR='11111' 70 return MI->getOperand(3).getReg() != AArch64::XZR; 71 default: 72 return false; 73 } 74 } 75 76 77 //===----------------------------------------------------------------------===// 78 79 namespace { 80 class AArch64A53Fix835769 : public MachineFunctionPass { 81 const TargetInstrInfo *TII; 82 83 public: 84 static char ID; 85 explicit AArch64A53Fix835769() : MachineFunctionPass(ID) {} 86 87 bool runOnMachineFunction(MachineFunction &F) override; 88 89 MachineFunctionProperties getRequiredProperties() const override { 90 return MachineFunctionProperties().set( 91 MachineFunctionProperties::Property::AllVRegsAllocated); 92 } 93 94 const char *getPassName() const override { 95 return "Workaround A53 erratum 835769 pass"; 96 } 97 98 void getAnalysisUsage(AnalysisUsage &AU) const override { 99 AU.setPreservesCFG(); 100 MachineFunctionPass::getAnalysisUsage(AU); 101 } 102 103 private: 104 bool runOnBasicBlock(MachineBasicBlock &MBB); 105 }; 106 char AArch64A53Fix835769::ID = 0; 107 108 } // end anonymous namespace 109 110 //===----------------------------------------------------------------------===// 111 112 bool 113 AArch64A53Fix835769::runOnMachineFunction(MachineFunction &F) { 114 DEBUG(dbgs() << "***** AArch64A53Fix835769 *****\n"); 115 bool Changed = false; 116 TII = F.getSubtarget().getInstrInfo(); 117 118 for (auto &MBB : F) { 119 Changed |= runOnBasicBlock(MBB); 120 } 121 return Changed; 122 } 123 124 // Return the block that was fallen through to get to MBB, if any, 125 // otherwise nullptr. 126 static MachineBasicBlock *getBBFallenThrough(MachineBasicBlock *MBB, 127 const TargetInstrInfo *TII) { 128 // Get the previous machine basic block in the function. 129 MachineFunction::iterator MBBI(MBB); 130 131 // Can't go off top of function. 132 if (MBBI == MBB->getParent()->begin()) 133 return nullptr; 134 135 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; 136 SmallVector<MachineOperand, 2> Cond; 137 138 MachineBasicBlock *PrevBB = &*std::prev(MBBI); 139 for (MachineBasicBlock *S : MBB->predecessors()) 140 if (S == PrevBB && !TII->analyzeBranch(*PrevBB, TBB, FBB, Cond) && !TBB && 141 !FBB) 142 return S; 143 144 return nullptr; 145 } 146 147 // Iterate through fallen through blocks trying to find a previous non-pseudo if 148 // there is one, otherwise return nullptr. Only look for instructions in 149 // previous blocks, not the current block, since we only use this to look at 150 // previous blocks. 151 static MachineInstr *getLastNonPseudo(MachineBasicBlock &MBB, 152 const TargetInstrInfo *TII) { 153 MachineBasicBlock *FMBB = &MBB; 154 155 // If there is no non-pseudo in the current block, loop back around and try 156 // the previous block (if there is one). 157 while ((FMBB = getBBFallenThrough(FMBB, TII))) { 158 for (MachineInstr &I : make_range(FMBB->rbegin(), FMBB->rend())) 159 if (!I.isPseudo()) 160 return &I; 161 } 162 163 // There was no previous non-pseudo in the fallen through blocks 164 return nullptr; 165 } 166 167 static void insertNopBeforeInstruction(MachineBasicBlock &MBB, MachineInstr* MI, 168 const TargetInstrInfo *TII) { 169 // If we are the first instruction of the block, put the NOP at the end of 170 // the previous fallthrough block 171 if (MI == &MBB.front()) { 172 MachineInstr *I = getLastNonPseudo(MBB, TII); 173 assert(I && "Expected instruction"); 174 DebugLoc DL = I->getDebugLoc(); 175 BuildMI(I->getParent(), DL, TII->get(AArch64::HINT)).addImm(0); 176 } 177 else { 178 DebugLoc DL = MI->getDebugLoc(); 179 BuildMI(MBB, MI, DL, TII->get(AArch64::HINT)).addImm(0); 180 } 181 182 ++NumNopsAdded; 183 } 184 185 bool 186 AArch64A53Fix835769::runOnBasicBlock(MachineBasicBlock &MBB) { 187 bool Changed = false; 188 DEBUG(dbgs() << "Running on MBB: " << MBB << " - scanning instructions...\n"); 189 190 // First, scan the basic block, looking for a sequence of 2 instructions 191 // that match the conditions under which the erratum may trigger. 192 193 // List of terminating instructions in matching sequences 194 std::vector<MachineInstr*> Sequences; 195 unsigned Idx = 0; 196 MachineInstr *PrevInstr = nullptr; 197 198 // Try and find the last non-pseudo instruction in any fallen through blocks, 199 // if there isn't one, then we use nullptr to represent that. 200 PrevInstr = getLastNonPseudo(MBB, TII); 201 202 for (auto &MI : MBB) { 203 MachineInstr *CurrInstr = &MI; 204 DEBUG(dbgs() << " Examining: " << MI); 205 if (PrevInstr) { 206 DEBUG(dbgs() << " PrevInstr: " << *PrevInstr 207 << " CurrInstr: " << *CurrInstr 208 << " isFirstInstructionInSequence(PrevInstr): " 209 << isFirstInstructionInSequence(PrevInstr) << "\n" 210 << " isSecondInstructionInSequence(CurrInstr): " 211 << isSecondInstructionInSequence(CurrInstr) << "\n"); 212 if (isFirstInstructionInSequence(PrevInstr) && 213 isSecondInstructionInSequence(CurrInstr)) { 214 DEBUG(dbgs() << " ** pattern found at Idx " << Idx << "!\n"); 215 Sequences.push_back(CurrInstr); 216 } 217 } 218 if (!CurrInstr->isPseudo()) 219 PrevInstr = CurrInstr; 220 ++Idx; 221 } 222 223 DEBUG(dbgs() << "Scan complete, " << Sequences.size() 224 << " occurrences of pattern found.\n"); 225 226 // Then update the basic block, inserting nops between the detected sequences. 227 for (auto &MI : Sequences) { 228 Changed = true; 229 insertNopBeforeInstruction(MBB, MI, TII); 230 } 231 232 return Changed; 233 } 234 235 // Factory function used by AArch64TargetMachine to add the pass to 236 // the passmanager. 237 FunctionPass *llvm::createAArch64A53Fix835769() { 238 return new AArch64A53Fix835769(); 239 } 240