003 File Manager
Current Path:
/usr/src/contrib/llvm-project/llvm/lib/Target/AArch64
usr
/
src
/
contrib
/
llvm-project
/
llvm
/
lib
/
Target
/
AArch64
/
📁
..
📄
AArch64.h
(4.09 KB)
📄
AArch64.td
(49.85 KB)
📄
AArch64A53Fix835769.cpp
(8.3 KB)
📄
AArch64A57FPLoadBalancing.cpp
(25.73 KB)
📄
AArch64AdvSIMDScalarPass.cpp
(16.1 KB)
📄
AArch64AsmPrinter.cpp
(49.26 KB)
📄
AArch64BranchTargets.cpp
(4.91 KB)
📄
AArch64CallingConvention.cpp
(6.63 KB)
📄
AArch64CallingConvention.h
(2.62 KB)
📄
AArch64CallingConvention.td
(23.84 KB)
📄
AArch64CleanupLocalDynamicTLSPass.cpp
(5.53 KB)
📄
AArch64CollectLOH.cpp
(20.07 KB)
📄
AArch64Combine.td
(3.27 KB)
📄
AArch64CompressJumpTables.cpp
(5.06 KB)
📄
AArch64CondBrTuning.cpp
(10.19 KB)
📄
AArch64ConditionOptimizer.cpp
(15.26 KB)
📄
AArch64ConditionalCompares.cpp
(33.26 KB)
📄
AArch64DeadRegisterDefinitionsPass.cpp
(7.76 KB)
📄
AArch64ExpandImm.cpp
(14.25 KB)
📄
AArch64ExpandImm.h
(959 B)
📄
AArch64ExpandPseudoInsts.cpp
(38.08 KB)
📄
AArch64FalkorHWPFFix.cpp
(23.3 KB)
📄
AArch64FastISel.cpp
(171.76 KB)
📄
AArch64FrameLowering.cpp
(124.13 KB)
📄
AArch64FrameLowering.h
(5.54 KB)
📄
AArch64GenRegisterBankInfo.def
(11 KB)
📄
AArch64ISelDAGToDAG.cpp
(180.26 KB)
📄
AArch64ISelLowering.cpp
(578.95 KB)
📄
AArch64ISelLowering.h
(33.88 KB)
📄
AArch64InstrAtomics.td
(20.33 KB)
📄
AArch64InstrFormats.td
(430.93 KB)
📄
AArch64InstrGISel.td
(4.29 KB)
📄
AArch64InstrInfo.cpp
(243.23 KB)
📄
AArch64InstrInfo.h
(19.92 KB)
📄
AArch64InstrInfo.td
(374.84 KB)
📄
AArch64LoadStoreOptimizer.cpp
(77.04 KB)
📄
AArch64MCInstLower.cpp
(11.72 KB)
📄
AArch64MCInstLower.h
(1.69 KB)
📄
AArch64MachineFunctionInfo.cpp
(1.02 KB)
📄
AArch64MachineFunctionInfo.h
(12.9 KB)
📄
AArch64MacroFusion.cpp
(11.47 KB)
📄
AArch64MacroFusion.h
(891 B)
📄
AArch64PBQPRegAlloc.cpp
(11.35 KB)
📄
AArch64PBQPRegAlloc.h
(1.3 KB)
📄
AArch64PerfectShuffle.h
(382.04 KB)
📄
AArch64PfmCounters.td
(713 B)
📄
AArch64PromoteConstant.cpp
(22.43 KB)
📄
AArch64RedundantCopyElimination.cpp
(17.09 KB)
📄
AArch64RegisterBanks.td
(719 B)
📄
AArch64RegisterInfo.cpp
(29.6 KB)
📄
AArch64RegisterInfo.h
(5.5 KB)
📄
AArch64RegisterInfo.td
(51 KB)
📄
AArch64SIMDInstrOpt.cpp
(26.07 KB)
📄
AArch64SLSHardening.cpp
(15.92 KB)
📄
AArch64SVEInstrInfo.td
(169.43 KB)
📄
AArch64SchedA53.td
(15.28 KB)
📄
AArch64SchedA57.td
(34.69 KB)
📄
AArch64SchedA57WriteRes.td
(19.87 KB)
📄
AArch64SchedCyclone.td
(29.82 KB)
📄
AArch64SchedExynosM3.td
(42.57 KB)
📄
AArch64SchedExynosM4.td
(49.81 KB)
📄
AArch64SchedExynosM5.td
(50.74 KB)
📄
AArch64SchedFalkor.td
(5.3 KB)
📄
AArch64SchedFalkorDetails.td
(67.66 KB)
📄
AArch64SchedKryo.td
(6.21 KB)
📄
AArch64SchedKryoDetails.td
(82.63 KB)
📄
AArch64SchedPredExynos.td
(7.5 KB)
📄
AArch64SchedPredicates.td
(27.86 KB)
📄
AArch64SchedThunderX.td
(14.99 KB)
📄
AArch64SchedThunderX2T99.td
(68.58 KB)
📄
AArch64SchedThunderX3T110.td
(68.77 KB)
📄
AArch64Schedule.td
(3.86 KB)
📄
AArch64SelectionDAGInfo.cpp
(5.54 KB)
📄
AArch64SelectionDAGInfo.h
(1.46 KB)
📄
AArch64SpeculationHardening.cpp
(29.6 KB)
📄
AArch64StackOffset.h
(5.01 KB)
📄
AArch64StackTagging.cpp
(24.31 KB)
📄
AArch64StackTaggingPreRA.cpp
(7.25 KB)
📄
AArch64StorePairSuppress.cpp
(6.26 KB)
📄
AArch64Subtarget.cpp
(12.38 KB)
📄
AArch64Subtarget.h
(18.49 KB)
📄
AArch64SystemOperands.td
(81.75 KB)
📄
AArch64TargetMachine.cpp
(26.65 KB)
📄
AArch64TargetMachine.h
(3.25 KB)
📄
AArch64TargetObjectFile.cpp
(3.35 KB)
📄
AArch64TargetObjectFile.h
(2.31 KB)
📄
AArch64TargetTransformInfo.cpp
(42.6 KB)
📄
AArch64TargetTransformInfo.h
(8.68 KB)
📁
AsmParser
📁
Disassembler
📁
GISel
📁
MCTargetDesc
📄
SVEInstrFormats.td
(304.03 KB)
📄
SVEIntrinsicOpts.cpp
(8.13 KB)
📁
TargetInfo
📁
Utils
Editing: AArch64StorePairSuppress.cpp
//===--- AArch64StorePairSuppress.cpp --- Suppress store pair formation ---===// // // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. // See https://llvm.org/LICENSE.txt for license information. // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception // //===----------------------------------------------------------------------===// // // This pass identifies floating point stores that should not be combined into // store pairs. Later we may do the same for floating point loads. // ===---------------------------------------------------------------------===// #include "AArch64InstrInfo.h" #include "llvm/CodeGen/MachineFunction.h" #include "llvm/CodeGen/MachineFunctionPass.h" #include "llvm/CodeGen/MachineInstr.h" #include "llvm/CodeGen/MachineTraceMetrics.h" #include "llvm/CodeGen/TargetInstrInfo.h" #include "llvm/CodeGen/TargetSchedule.h" #include "llvm/Support/Debug.h" #include "llvm/Support/raw_ostream.h" using namespace llvm; #define DEBUG_TYPE "aarch64-stp-suppress" #define STPSUPPRESS_PASS_NAME "AArch64 Store Pair Suppression" namespace { class AArch64StorePairSuppress : public MachineFunctionPass { const AArch64InstrInfo *TII; const TargetRegisterInfo *TRI; const MachineRegisterInfo *MRI; TargetSchedModel SchedModel; MachineTraceMetrics *Traces; MachineTraceMetrics::Ensemble *MinInstr; public: static char ID; AArch64StorePairSuppress() : MachineFunctionPass(ID) { initializeAArch64StorePairSuppressPass(*PassRegistry::getPassRegistry()); } StringRef getPassName() const override { return STPSUPPRESS_PASS_NAME; } bool runOnMachineFunction(MachineFunction &F) override; private: bool shouldAddSTPToBlock(const MachineBasicBlock *BB); bool isNarrowFPStore(const MachineInstr &MI); void getAnalysisUsage(AnalysisUsage &AU) const override { AU.setPreservesCFG(); AU.addRequired<MachineTraceMetrics>(); AU.addPreserved<MachineTraceMetrics>(); MachineFunctionPass::getAnalysisUsage(AU); } }; char AArch64StorePairSuppress::ID = 0; } // anonymous INITIALIZE_PASS(AArch64StorePairSuppress, "aarch64-stp-suppress", STPSUPPRESS_PASS_NAME, false, false) FunctionPass *llvm::createAArch64StorePairSuppressPass() { return new AArch64StorePairSuppress(); } /// Return true if an STP can be added to this block without increasing the /// critical resource height. STP is good to form in Ld/St limited blocks and /// bad to form in float-point limited blocks. This is true independent of the /// critical path. If the critical path is longer than the resource height, the /// extra vector ops can limit physreg renaming. Otherwise, it could simply /// oversaturate the vector units. bool AArch64StorePairSuppress::shouldAddSTPToBlock(const MachineBasicBlock *BB) { if (!MinInstr) MinInstr = Traces->getEnsemble(MachineTraceMetrics::TS_MinInstrCount); MachineTraceMetrics::Trace BBTrace = MinInstr->getTrace(BB); unsigned ResLength = BBTrace.getResourceLength(); // Get the machine model's scheduling class for STPQi. // Bypass TargetSchedule's SchedClass resolution since we only have an opcode. unsigned SCIdx = TII->get(AArch64::STPDi).getSchedClass(); const MCSchedClassDesc *SCDesc = SchedModel.getMCSchedModel()->getSchedClassDesc(SCIdx); // If a subtarget does not define resources for STPQi, bail here. if (SCDesc->isValid() && !SCDesc->isVariant()) { unsigned ResLenWithSTP = BBTrace.getResourceLength(None, SCDesc); if (ResLenWithSTP > ResLength) { LLVM_DEBUG(dbgs() << " Suppress STP in BB: " << BB->getNumber() << " resources " << ResLength << " -> " << ResLenWithSTP << "\n"); return false; } } return true; } /// Return true if this is a floating-point store smaller than the V reg. On /// cyclone, these require a vector shuffle before storing a pair. /// Ideally we would call getMatchingPairOpcode() and have the machine model /// tell us if it's profitable with no cpu knowledge here. /// /// FIXME: We plan to develop a decent Target abstraction for simple loads and /// stores. Until then use a nasty switch similar to AArch64LoadStoreOptimizer. bool AArch64StorePairSuppress::isNarrowFPStore(const MachineInstr &MI) { switch (MI.getOpcode()) { default: return false; case AArch64::STRSui: case AArch64::STRDui: case AArch64::STURSi: case AArch64::STURDi: return true; } } bool AArch64StorePairSuppress::runOnMachineFunction(MachineFunction &MF) { if (skipFunction(MF.getFunction())) return false; const TargetSubtargetInfo &ST = MF.getSubtarget(); TII = static_cast<const AArch64InstrInfo *>(ST.getInstrInfo()); TRI = ST.getRegisterInfo(); MRI = &MF.getRegInfo(); SchedModel.init(&ST); Traces = &getAnalysis<MachineTraceMetrics>(); MinInstr = nullptr; LLVM_DEBUG(dbgs() << "*** " << getPassName() << ": " << MF.getName() << '\n'); if (!SchedModel.hasInstrSchedModel()) { LLVM_DEBUG(dbgs() << " Skipping pass: no machine model present.\n"); return false; } // Check for a sequence of stores to the same base address. We don't need to // precisely determine whether a store pair can be formed. But we do want to // filter out most situations where we can't form store pairs to avoid // computing trace metrics in those cases. for (auto &MBB : MF) { bool SuppressSTP = false; unsigned PrevBaseReg = 0; for (auto &MI : MBB) { if (!isNarrowFPStore(MI)) continue; const MachineOperand *BaseOp; int64_t Offset; bool OffsetIsScalable; if (TII->getMemOperandWithOffset(MI, BaseOp, Offset, OffsetIsScalable, TRI) && BaseOp->isReg()) { Register BaseReg = BaseOp->getReg(); if (PrevBaseReg == BaseReg) { // If this block can take STPs, skip ahead to the next block. if (!SuppressSTP && shouldAddSTPToBlock(MI.getParent())) break; // Otherwise, continue unpairing the stores in this block. LLVM_DEBUG(dbgs() << "Unpairing store " << MI << "\n"); SuppressSTP = true; TII->suppressLdStPair(MI); } PrevBaseReg = BaseReg; } else PrevBaseReg = 0; } } // This pass just sets some internal MachineMemOperand flags. It can't really // invalidate anything. return false; }
Upload File
Create Folder