//===- ARMConstantIslandPass.cpp - ARM constant islands -------------------===// // // 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 file contains a pass that splits the constant pool up into 'islands' // which are scattered through-out the function. This is required due to the // limited pc-relative displacements that ARM has. // //===----------------------------------------------------------------------===// #include "ARM.h" #include "ARMBaseInstrInfo.h" #include "ARMBasicBlockInfo.h" #include "ARMMachineFunctionInfo.h" #include "ARMSubtarget.h" #include "MCTargetDesc/ARMBaseInfo.h" #include "MVETailPredUtils.h" #include "Thumb2InstrInfo.h" #include "Utils/ARMBaseInfo.h" #include "llvm/ADT/DenseMap.h" #include "llvm/ADT/STLExtras.h" #include "llvm/ADT/SmallSet.h" #include "llvm/ADT/SmallVector.h" #include "llvm/ADT/Statistic.h" #include "llvm/ADT/StringRef.h" #include "llvm/CodeGen/LivePhysRegs.h" #include "llvm/CodeGen/MachineBasicBlock.h" #include "llvm/CodeGen/MachineConstantPool.h" #include "llvm/CodeGen/MachineDominators.h" #include "llvm/CodeGen/MachineFunction.h" #include "llvm/CodeGen/MachineFunctionPass.h" #include "llvm/CodeGen/MachineInstr.h" #include "llvm/CodeGen/MachineJumpTableInfo.h" #include "llvm/CodeGen/MachineOperand.h" #include "llvm/CodeGen/MachineRegisterInfo.h" #include "llvm/Config/llvm-config.h" #include "llvm/IR/DataLayout.h" #include "llvm/IR/DebugLoc.h" #include "llvm/MC/MCInstrDesc.h" #include "llvm/Pass.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Compiler.h" #include "llvm/Support/Debug.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/Format.h" #include "llvm/Support/MathExtras.h" #include "llvm/Support/raw_ostream.h" #include #include #include #include #include #include using namespace llvm; #define DEBUG_TYPE "arm-cp-islands" #define ARM_CP_ISLANDS_OPT_NAME \ "ARM constant island placement and branch shortening pass" STATISTIC(NumCPEs, "Number of constpool entries"); STATISTIC(NumSplit, "Number of uncond branches inserted"); STATISTIC(NumCBrFixed, "Number of cond branches fixed"); STATISTIC(NumUBrFixed, "Number of uncond branches fixed"); STATISTIC(NumTBs, "Number of table branches generated"); STATISTIC(NumT2CPShrunk, "Number of Thumb2 constantpool instructions shrunk"); STATISTIC(NumT2BrShrunk, "Number of Thumb2 immediate branches shrunk"); STATISTIC(NumCBZ, "Number of CBZ / CBNZ formed"); STATISTIC(NumJTMoved, "Number of jump table destination blocks moved"); STATISTIC(NumJTInserted, "Number of jump table intermediate blocks inserted"); STATISTIC(NumLEInserted, "Number of LE backwards branches inserted"); static cl::opt AdjustJumpTableBlocks("arm-adjust-jump-tables", cl::Hidden, cl::init(true), cl::desc("Adjust basic block layout to better use TB[BH]")); static cl::opt CPMaxIteration("arm-constant-island-max-iteration", cl::Hidden, cl::init(30), cl::desc("The max number of iteration for converge")); static cl::opt SynthesizeThumb1TBB( "arm-synthesize-thumb-1-tbb", cl::Hidden, cl::init(true), cl::desc("Use compressed jump tables in Thumb-1 by synthesizing an " "equivalent to the TBB/TBH instructions")); namespace { /// ARMConstantIslands - Due to limited PC-relative displacements, ARM /// requires constant pool entries to be scattered among the instructions /// inside a function. To do this, it completely ignores the normal LLVM /// constant pool; instead, it places constants wherever it feels like with /// special instructions. /// /// The terminology used in this pass includes: /// Islands - Clumps of constants placed in the function. /// Water - Potential places where an island could be formed. /// CPE - A constant pool entry that has been placed somewhere, which /// tracks a list of users. class ARMConstantIslands : public MachineFunctionPass { std::unique_ptr BBUtils = nullptr; /// WaterList - A sorted list of basic blocks where islands could be placed /// (i.e. blocks that don't fall through to the following block, due /// to a return, unreachable, or unconditional branch). std::vector WaterList; /// NewWaterList - The subset of WaterList that was created since the /// previous iteration by inserting unconditional branches. SmallSet NewWaterList; using water_iterator = std::vector::iterator; /// CPUser - One user of a constant pool, keeping the machine instruction /// pointer, the constant pool being referenced, and the max displacement /// allowed from the instruction to the CP. The HighWaterMark records the /// highest basic block where a new CPEntry can be placed. To ensure this /// pass terminates, the CP entries are initially placed at the end of the /// function and then move monotonically to lower addresses. The /// exception to this rule is when the current CP entry for a particular /// CPUser is out of range, but there is another CP entry for the same /// constant value in range. We want to use the existing in-range CP /// entry, but if it later moves out of range, the search for new water /// should resume where it left off. The HighWaterMark is used to record /// that point. struct CPUser { MachineInstr *MI; MachineInstr *CPEMI; MachineBasicBlock *HighWaterMark; unsigned MaxDisp; bool NegOk; bool IsSoImm; bool KnownAlignment = false; CPUser(MachineInstr *mi, MachineInstr *cpemi, unsigned maxdisp, bool neg, bool soimm) : MI(mi), CPEMI(cpemi), MaxDisp(maxdisp), NegOk(neg), IsSoImm(soimm) { HighWaterMark = CPEMI->getParent(); } /// getMaxDisp - Returns the maximum displacement supported by MI. /// Correct for unknown alignment. /// Conservatively subtract 2 bytes to handle weird alignment effects. unsigned getMaxDisp() const { return (KnownAlignment ? MaxDisp : MaxDisp - 2) - 2; } }; /// CPUsers - Keep track of all of the machine instructions that use various /// constant pools and their max displacement. std::vector CPUsers; /// CPEntry - One per constant pool entry, keeping the machine instruction /// pointer, the constpool index, and the number of CPUser's which /// reference this entry. struct CPEntry { MachineInstr *CPEMI; unsigned CPI; unsigned RefCount; CPEntry(MachineInstr *cpemi, unsigned cpi, unsigned rc = 0) : CPEMI(cpemi), CPI(cpi), RefCount(rc) {} }; /// CPEntries - Keep track of all of the constant pool entry machine /// instructions. For each original constpool index (i.e. those that existed /// upon entry to this pass), it keeps a vector of entries. Original /// elements are cloned as we go along; the clones are put in the vector of /// the original element, but have distinct CPIs. /// /// The first half of CPEntries contains generic constants, the second half /// contains jump tables. Use getCombinedIndex on a generic CPEMI to look up /// which vector it will be in here. std::vector> CPEntries; /// Maps a JT index to the offset in CPEntries containing copies of that /// table. The equivalent map for a CONSTPOOL_ENTRY is the identity. DenseMap JumpTableEntryIndices; /// Maps a JT index to the LEA that actually uses the index to calculate its /// base address. DenseMap JumpTableUserIndices; // Maps a MachineBasicBlock to the number of jump tables entries. DenseMap BlockJumpTableRefCount; /// ImmBranch - One per immediate branch, keeping the machine instruction /// pointer, conditional or unconditional, the max displacement, /// and (if isCond is true) the corresponding unconditional branch /// opcode. struct ImmBranch { MachineInstr *MI; unsigned MaxDisp : 31; bool isCond : 1; unsigned UncondBr; ImmBranch(MachineInstr *mi, unsigned maxdisp, bool cond, unsigned ubr) : MI(mi), MaxDisp(maxdisp), isCond(cond), UncondBr(ubr) {} }; /// ImmBranches - Keep track of all the immediate branch instructions. std::vector ImmBranches; /// PushPopMIs - Keep track of all the Thumb push / pop instructions. SmallVector PushPopMIs; /// T2JumpTables - Keep track of all the Thumb2 jumptable instructions. SmallVector T2JumpTables; MachineFunction *MF; MachineConstantPool *MCP; const ARMBaseInstrInfo *TII; const ARMSubtarget *STI; ARMFunctionInfo *AFI; MachineDominatorTree *DT = nullptr; bool isThumb; bool isThumb1; bool isThumb2; bool isPositionIndependentOrROPI; public: static char ID; ARMConstantIslands() : MachineFunctionPass(ID) {} bool runOnMachineFunction(MachineFunction &MF) override; void getAnalysisUsage(AnalysisUsage &AU) const override { AU.addRequired(); MachineFunctionPass::getAnalysisUsage(AU); } MachineFunctionProperties getRequiredProperties() const override { return MachineFunctionProperties().set( MachineFunctionProperties::Property::NoVRegs); } StringRef getPassName() const override { return ARM_CP_ISLANDS_OPT_NAME; } private: void doInitialConstPlacement(std::vector &CPEMIs); void doInitialJumpTablePlacement(std::vector &CPEMIs); bool BBHasFallthrough(MachineBasicBlock *MBB); CPEntry *findConstPoolEntry(unsigned CPI, const MachineInstr *CPEMI); Align getCPEAlign(const MachineInstr *CPEMI); void scanFunctionJumpTables(); void initializeFunctionInfo(const std::vector &CPEMIs); MachineBasicBlock *splitBlockBeforeInstr(MachineInstr *MI); void updateForInsertedWaterBlock(MachineBasicBlock *NewBB); bool decrementCPEReferenceCount(unsigned CPI, MachineInstr* CPEMI); unsigned getCombinedIndex(const MachineInstr *CPEMI); int findInRangeCPEntry(CPUser& U, unsigned UserOffset); bool findAvailableWater(CPUser&U, unsigned UserOffset, water_iterator &WaterIter, bool CloserWater); void createNewWater(unsigned CPUserIndex, unsigned UserOffset, MachineBasicBlock *&NewMBB); bool handleConstantPoolUser(unsigned CPUserIndex, bool CloserWater); void removeDeadCPEMI(MachineInstr *CPEMI); bool removeUnusedCPEntries(); bool isCPEntryInRange(MachineInstr *MI, unsigned UserOffset, MachineInstr *CPEMI, unsigned Disp, bool NegOk, bool DoDump = false); bool isWaterInRange(unsigned UserOffset, MachineBasicBlock *Water, CPUser &U, unsigned &Growth); bool fixupImmediateBr(ImmBranch &Br); bool fixupConditionalBr(ImmBranch &Br); bool fixupUnconditionalBr(ImmBranch &Br); bool optimizeThumb2Instructions(); bool optimizeThumb2Branches(); bool reorderThumb2JumpTables(); bool preserveBaseRegister(MachineInstr *JumpMI, MachineInstr *LEAMI, unsigned &DeadSize, bool &CanDeleteLEA, bool &BaseRegKill); bool optimizeThumb2JumpTables(); MachineBasicBlock *adjustJTTargetBlockForward(unsigned JTI, MachineBasicBlock *BB, MachineBasicBlock *JTBB); unsigned getUserOffset(CPUser&) const; void dumpBBs(); void verify(); bool isOffsetInRange(unsigned UserOffset, unsigned TrialOffset, unsigned Disp, bool NegativeOK, bool IsSoImm = false); bool isOffsetInRange(unsigned UserOffset, unsigned TrialOffset, const CPUser &U) { return isOffsetInRange(UserOffset, TrialOffset, U.getMaxDisp(), U.NegOk, U.IsSoImm); } }; } // end anonymous namespace char ARMConstantIslands::ID = 0; /// verify - check BBOffsets, BBSizes, alignment of islands void ARMConstantIslands::verify() { #ifndef NDEBUG BBInfoVector &BBInfo = BBUtils->getBBInfo(); assert(is_sorted(*MF, [&BBInfo](const MachineBasicBlock &LHS, const MachineBasicBlock &RHS) { return BBInfo[LHS.getNumber()].postOffset() < BBInfo[RHS.getNumber()].postOffset(); })); LLVM_DEBUG(dbgs() << "Verifying " << CPUsers.size() << " CP users.\n"); for (CPUser &U : CPUsers) { unsigned UserOffset = getUserOffset(U); // Verify offset using the real max displacement without the safety // adjustment. if (isCPEntryInRange(U.MI, UserOffset, U.CPEMI, U.getMaxDisp()+2, U.NegOk, /* DoDump = */ true)) { LLVM_DEBUG(dbgs() << "OK\n"); continue; } LLVM_DEBUG(dbgs() << "Out of range.\n"); dumpBBs(); LLVM_DEBUG(MF->dump()); llvm_unreachable("Constant pool entry out of range!"); } #endif } #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) /// print block size and offset information - debugging LLVM_DUMP_METHOD void ARMConstantIslands::dumpBBs() { LLVM_DEBUG({ BBInfoVector &BBInfo = BBUtils->getBBInfo(); for (unsigned J = 0, E = BBInfo.size(); J !=E; ++J) { const BasicBlockInfo &BBI = BBInfo[J]; dbgs() << format("%08x %bb.%u\t", BBI.Offset, J) << " kb=" << unsigned(BBI.KnownBits) << " ua=" << unsigned(BBI.Unalign) << " pa=" << Log2(BBI.PostAlign) << format(" size=%#x\n", BBInfo[J].Size); } }); } #endif // Align blocks where the previous block does not fall through. This may add // extra NOP's but they will not be executed. It uses the PrefLoopAlignment as a // measure of how much to align, and only runs at CodeGenOpt::Aggressive. static bool AlignBlocks(MachineFunction *MF, const ARMSubtarget *STI) { if (MF->getTarget().getOptLevel() != CodeGenOpt::Aggressive || MF->getFunction().hasOptSize()) return false; auto *TLI = STI->getTargetLowering(); const Align Alignment = TLI->getPrefLoopAlignment(); if (Alignment < 4) return false; bool Changed = false; bool PrevCanFallthough = true; for (auto &MBB : *MF) { if (!PrevCanFallthough) { Changed = true; MBB.setAlignment(Alignment); } PrevCanFallthough = MBB.canFallThrough(); // For LOB's, the ARMLowOverheadLoops pass may remove the unconditional // branch later in the pipeline. if (STI->hasLOB()) { for (const auto &MI : reverse(MBB.terminators())) { if (MI.getOpcode() == ARM::t2B && MI.getOperand(0).getMBB() == MBB.getNextNode()) continue; if (isLoopStart(MI) || MI.getOpcode() == ARM::t2LoopEnd || MI.getOpcode() == ARM::t2LoopEndDec) { PrevCanFallthough = true; break; } // Any other terminator - nothing to do break; } } } return Changed; } bool ARMConstantIslands::runOnMachineFunction(MachineFunction &mf) { MF = &mf; MCP = mf.getConstantPool(); BBUtils = std::unique_ptr(new ARMBasicBlockUtils(mf)); LLVM_DEBUG(dbgs() << "***** ARMConstantIslands: " << MCP->getConstants().size() << " CP entries, aligned to " << MCP->getConstantPoolAlign().value() << " bytes *****\n"); STI = &MF->getSubtarget(); TII = STI->getInstrInfo(); isPositionIndependentOrROPI = STI->getTargetLowering()->isPositionIndependent() || STI->isROPI(); AFI = MF->getInfo(); DT = &getAnalysis(); isThumb = AFI->isThumbFunction(); isThumb1 = AFI->isThumb1OnlyFunction(); isThumb2 = AFI->isThumb2Function(); bool GenerateTBB = isThumb2 || (isThumb1 && SynthesizeThumb1TBB); // TBB generation code in this constant island pass has not been adapted to // deal with speculation barriers. if (STI->hardenSlsRetBr()) GenerateTBB = false; // Renumber all of the machine basic blocks in the function, guaranteeing that // the numbers agree with the position of the block in the function. MF->RenumberBlocks(); // Try to reorder and otherwise adjust the block layout to make good use // of the TB[BH] instructions. bool MadeChange = false; if (GenerateTBB && AdjustJumpTableBlocks) { scanFunctionJumpTables(); MadeChange |= reorderThumb2JumpTables(); // Data is out of date, so clear it. It'll be re-computed later. T2JumpTables.clear(); // Blocks may have shifted around. Keep the numbering up to date. MF->RenumberBlocks(); } // Align any non-fallthrough blocks MadeChange |= AlignBlocks(MF, STI); // Perform the initial placement of the constant pool entries. To start with, // we put them all at the end of the function. std::vector CPEMIs; if (!MCP->isEmpty()) doInitialConstPlacement(CPEMIs); if (MF->getJumpTableInfo()) doInitialJumpTablePlacement(CPEMIs); /// The next UID to take is the first unused one. AFI->initPICLabelUId(CPEMIs.size()); // Do the initial scan of the function, building up information about the // sizes of each block, the location of all the water, and finding all of the // constant pool users. initializeFunctionInfo(CPEMIs); CPEMIs.clear(); LLVM_DEBUG(dumpBBs()); // Functions with jump tables need an alignment of 4 because they use the ADR // instruction, which aligns the PC to 4 bytes before adding an offset. if (!T2JumpTables.empty()) MF->ensureAlignment(Align(4)); /// Remove dead constant pool entries. MadeChange |= removeUnusedCPEntries(); // Iteratively place constant pool entries and fix up branches until there // is no change. unsigned NoCPIters = 0, NoBRIters = 0; while (true) { LLVM_DEBUG(dbgs() << "Beginning CP iteration #" << NoCPIters << '\n'); bool CPChange = false; for (unsigned i = 0, e = CPUsers.size(); i != e; ++i) // For most inputs, it converges in no more than 5 iterations. // If it doesn't end in 10, the input may have huge BB or many CPEs. // In this case, we will try different heuristics. CPChange |= handleConstantPoolUser(i, NoCPIters >= CPMaxIteration / 2); if (CPChange && ++NoCPIters > CPMaxIteration) report_fatal_error("Constant Island pass failed to converge!"); LLVM_DEBUG(dumpBBs()); // Clear NewWaterList now. If we split a block for branches, it should // appear as "new water" for the next iteration of constant pool placement. NewWaterList.clear(); LLVM_DEBUG(dbgs() << "Beginning BR iteration #" << NoBRIters << '\n'); bool BRChange = false; for (unsigned i = 0, e = ImmBranches.size(); i != e; ++i) BRChange |= fixupImmediateBr(ImmBranches[i]); if (BRChange && ++NoBRIters > 30) report_fatal_error("Branch Fix Up pass failed to converge!"); LLVM_DEBUG(dumpBBs()); if (!CPChange && !BRChange) break; MadeChange = true; } // Shrink 32-bit Thumb2 load and store instructions. if (isThumb2 && !STI->prefers32BitThumb()) MadeChange |= optimizeThumb2Instructions(); // Shrink 32-bit branch instructions. if (isThumb && STI->hasV8MBaselineOps()) MadeChange |= optimizeThumb2Branches(); // Optimize jump tables using TBB / TBH. if (GenerateTBB && !STI->genExecuteOnly()) MadeChange |= optimizeThumb2JumpTables(); // After a while, this might be made debug-only, but it is not expensive. verify(); // Save the mapping between original and cloned constpool entries. for (unsigned i = 0, e = CPEntries.size(); i != e; ++i) { for (unsigned j = 0, je = CPEntries[i].size(); j != je; ++j) { const CPEntry & CPE = CPEntries[i][j]; if (CPE.CPEMI && CPE.CPEMI->getOperand(1).isCPI()) AFI->recordCPEClone(i, CPE.CPI); } } LLVM_DEBUG(dbgs() << '\n'; dumpBBs()); BBUtils->clear(); WaterList.clear(); CPUsers.clear(); CPEntries.clear(); JumpTableEntryIndices.clear(); JumpTableUserIndices.clear(); BlockJumpTableRefCount.clear(); ImmBranches.clear(); PushPopMIs.clear(); T2JumpTables.clear(); return MadeChange; } /// Perform the initial placement of the regular constant pool entries. /// To start with, we put them all at the end of the function. void ARMConstantIslands::doInitialConstPlacement(std::vector &CPEMIs) { // Create the basic block to hold the CPE's. MachineBasicBlock *BB = MF->CreateMachineBasicBlock(); MF->push_back(BB); // MachineConstantPool measures alignment in bytes. const Align MaxAlign = MCP->getConstantPoolAlign(); const unsigned MaxLogAlign = Log2(MaxAlign); // Mark the basic block as required by the const-pool. BB->setAlignment(MaxAlign); // The function needs to be as aligned as the basic blocks. The linker may // move functions around based on their alignment. // Special case: halfword literals still need word alignment on the function. Align FuncAlign = MaxAlign; if (MaxAlign == 2) FuncAlign = Align(4); MF->ensureAlignment(FuncAlign); // Order the entries in BB by descending alignment. That ensures correct // alignment of all entries as long as BB is sufficiently aligned. Keep // track of the insertion point for each alignment. We are going to bucket // sort the entries as they are created. SmallVector InsPoint(MaxLogAlign + 1, BB->end()); // Add all of the constants from the constant pool to the end block, use an // identity mapping of CPI's to CPE's. const std::vector &CPs = MCP->getConstants(); const DataLayout &TD = MF->getDataLayout(); for (unsigned i = 0, e = CPs.size(); i != e; ++i) { unsigned Size = CPs[i].getSizeInBytes(TD); Align Alignment = CPs[i].getAlign(); // Verify that all constant pool entries are a multiple of their alignment. // If not, we would have to pad them out so that instructions stay aligned. assert(isAligned(Alignment, Size) && "CP Entry not multiple of 4 bytes!"); // Insert CONSTPOOL_ENTRY before entries with a smaller alignment. unsigned LogAlign = Log2(Alignment); MachineBasicBlock::iterator InsAt = InsPoint[LogAlign]; MachineInstr *CPEMI = BuildMI(*BB, InsAt, DebugLoc(), TII->get(ARM::CONSTPOOL_ENTRY)) .addImm(i).addConstantPoolIndex(i).addImm(Size); CPEMIs.push_back(CPEMI); // Ensure that future entries with higher alignment get inserted before // CPEMI. This is bucket sort with iterators. for (unsigned a = LogAlign + 1; a <= MaxLogAlign; ++a) if (InsPoint[a] == InsAt) InsPoint[a] = CPEMI; // Add a new CPEntry, but no corresponding CPUser yet. CPEntries.emplace_back(1, CPEntry(CPEMI, i)); ++NumCPEs; LLVM_DEBUG(dbgs() << "Moved CPI#" << i << " to end of function, size = " << Size << ", align = " << Alignment.value() << '\n'); } LLVM_DEBUG(BB->dump()); } /// Do initial placement of the jump tables. Because Thumb2's TBB and TBH /// instructions can be made more efficient if the jump table immediately /// follows the instruction, it's best to place them immediately next to their /// jumps to begin with. In almost all cases they'll never be moved from that /// position. void ARMConstantIslands::doInitialJumpTablePlacement( std::vector &CPEMIs) { unsigned i = CPEntries.size(); auto MJTI = MF->getJumpTableInfo(); const std::vector &JT = MJTI->getJumpTables(); // Only inline jump tables are placed in the function. if (MJTI->getEntryKind() != MachineJumpTableInfo::EK_Inline) return; MachineBasicBlock *LastCorrectlyNumberedBB = nullptr; for (MachineBasicBlock &MBB : *MF) { auto MI = MBB.getLastNonDebugInstr(); // Look past potential SpeculationBarriers at end of BB. while (MI != MBB.end() && (isSpeculationBarrierEndBBOpcode(MI->getOpcode()) || MI->isDebugInstr())) --MI; if (MI == MBB.end()) continue; unsigned JTOpcode; switch (MI->getOpcode()) { default: continue; case ARM::BR_JTadd: case ARM::BR_JTr: case ARM::tBR_JTr: case ARM::BR_JTm_i12: case ARM::BR_JTm_rs: // These instructions are emitted only in ARM or Thumb1 modes which do not // support PACBTI. Hence we don't add BTI instructions in the destination // blocks. assert(!MF->getInfo()->branchTargetEnforcement() && "Branch protection must not be enabled for Arm or Thumb1 modes"); JTOpcode = ARM::JUMPTABLE_ADDRS; break; case ARM::t2BR_JT: JTOpcode = ARM::JUMPTABLE_INSTS; break; case ARM::tTBB_JT: case ARM::t2TBB_JT: JTOpcode = ARM::JUMPTABLE_TBB; break; case ARM::tTBH_JT: case ARM::t2TBH_JT: JTOpcode = ARM::JUMPTABLE_TBH; break; } unsigned NumOps = MI->getDesc().getNumOperands(); MachineOperand JTOp = MI->getOperand(NumOps - (MI->isPredicable() ? 2 : 1)); unsigned JTI = JTOp.getIndex(); unsigned Size = JT[JTI].MBBs.size() * sizeof(uint32_t); MachineBasicBlock *JumpTableBB = MF->CreateMachineBasicBlock(); MF->insert(std::next(MachineFunction::iterator(MBB)), JumpTableBB); MachineInstr *CPEMI = BuildMI(*JumpTableBB, JumpTableBB->begin(), DebugLoc(), TII->get(JTOpcode)) .addImm(i++) .addJumpTableIndex(JTI) .addImm(Size); CPEMIs.push_back(CPEMI); CPEntries.emplace_back(1, CPEntry(CPEMI, JTI)); JumpTableEntryIndices.insert(std::make_pair(JTI, CPEntries.size() - 1)); if (!LastCorrectlyNumberedBB) LastCorrectlyNumberedBB = &MBB; } // If we did anything then we need to renumber the subsequent blocks. if (LastCorrectlyNumberedBB) MF->RenumberBlocks(LastCorrectlyNumberedBB); } /// BBHasFallthrough - Return true if the specified basic block can fallthrough /// into the block immediately after it. bool ARMConstantIslands::BBHasFallthrough(MachineBasicBlock *MBB) { // Get the next machine basic block in the function. MachineFunction::iterator MBBI = MBB->getIterator(); // Can't fall off end of function. if (std::next(MBBI) == MBB->getParent()->end()) return false; MachineBasicBlock *NextBB = &*std::next(MBBI); if (!MBB->isSuccessor(NextBB)) return false; // Try to analyze the end of the block. A potential fallthrough may already // have an unconditional branch for whatever reason. MachineBasicBlock *TBB, *FBB; SmallVector Cond; bool TooDifficult = TII->analyzeBranch(*MBB, TBB, FBB, Cond); return TooDifficult || FBB == nullptr; } /// findConstPoolEntry - Given the constpool index and CONSTPOOL_ENTRY MI, /// look up the corresponding CPEntry. ARMConstantIslands::CPEntry * ARMConstantIslands::findConstPoolEntry(unsigned CPI, const MachineInstr *CPEMI) { std::vector &CPEs = CPEntries[CPI]; // Number of entries per constpool index should be small, just do a // linear search. for (CPEntry &CPE : CPEs) if (CPE.CPEMI == CPEMI) return &CPE; return nullptr; } /// getCPEAlign - Returns the required alignment of the constant pool entry /// represented by CPEMI. Align ARMConstantIslands::getCPEAlign(const MachineInstr *CPEMI) { switch (CPEMI->getOpcode()) { case ARM::CONSTPOOL_ENTRY: break; case ARM::JUMPTABLE_TBB: return isThumb1 ? Align(4) : Align(1); case ARM::JUMPTABLE_TBH: return isThumb1 ? Align(4) : Align(2); case ARM::JUMPTABLE_INSTS: return Align(2); case ARM::JUMPTABLE_ADDRS: return Align(4); default: llvm_unreachable("unknown constpool entry kind"); } unsigned CPI = getCombinedIndex(CPEMI); assert(CPI < MCP->getConstants().size() && "Invalid constant pool index."); return MCP->getConstants()[CPI].getAlign(); } // Exception landing pads, blocks that has their adress taken, and function // entry blocks will always be (potential) indirect jump targets, regardless of // whether they are referenced by or not by jump tables. static bool isAlwaysIndirectTarget(const MachineBasicBlock &MBB) { return MBB.isEHPad() || MBB.hasAddressTaken() || &MBB == &MBB.getParent()->front(); } /// scanFunctionJumpTables - Do a scan of the function, building up /// information about the sizes of each block and the locations of all /// the jump tables. void ARMConstantIslands::scanFunctionJumpTables() { for (MachineBasicBlock &MBB : *MF) { for (MachineInstr &I : MBB) if (I.isBranch() && (I.getOpcode() == ARM::t2BR_JT || I.getOpcode() == ARM::tBR_JTr)) T2JumpTables.push_back(&I); } if (!MF->getInfo()->branchTargetEnforcement()) return; if (const MachineJumpTableInfo *JTI = MF->getJumpTableInfo()) for (const MachineJumpTableEntry &JTE : JTI->getJumpTables()) for (const MachineBasicBlock *MBB : JTE.MBBs) { if (isAlwaysIndirectTarget(*MBB)) // Set the reference count essentially to infinity, it will never // reach zero and the BTI Instruction will never be removed. BlockJumpTableRefCount[MBB] = std::numeric_limits::max(); else ++BlockJumpTableRefCount[MBB]; } } /// initializeFunctionInfo - Do the initial scan of the function, building up /// information about the sizes of each block, the location of all the water, /// and finding all of the constant pool users. void ARMConstantIslands:: initializeFunctionInfo(const std::vector &CPEMIs) { BBUtils->computeAllBlockSizes(); BBInfoVector &BBInfo = BBUtils->getBBInfo(); // The known bits of the entry block offset are determined by the function // alignment. BBInfo.front().KnownBits = Log2(MF->getAlignment()); // Compute block offsets and known bits. BBUtils->adjustBBOffsetsAfter(&MF->front()); // We only care about jump table instructions when jump tables are inline. MachineJumpTableInfo *MJTI = MF->getJumpTableInfo(); bool InlineJumpTables = MJTI && MJTI->getEntryKind() == MachineJumpTableInfo::EK_Inline; // Now go back through the instructions and build up our data structures. for (MachineBasicBlock &MBB : *MF) { // If this block doesn't fall through into the next MBB, then this is // 'water' that a constant pool island could be placed. if (!BBHasFallthrough(&MBB)) WaterList.push_back(&MBB); for (MachineInstr &I : MBB) { if (I.isDebugInstr()) continue; unsigned Opc = I.getOpcode(); if (I.isBranch()) { bool isCond = false; unsigned Bits = 0; unsigned Scale = 1; int UOpc = Opc; switch (Opc) { default: continue; // Ignore other JT branches case ARM::t2BR_JT: case ARM::tBR_JTr: if (InlineJumpTables) T2JumpTables.push_back(&I); continue; // Does not get an entry in ImmBranches case ARM::Bcc: isCond = true; UOpc = ARM::B; [[fallthrough]]; case ARM::B: Bits = 24; Scale = 4; break; case ARM::tBcc: isCond = true; UOpc = ARM::tB; Bits = 8; Scale = 2; break; case ARM::tB: Bits = 11; Scale = 2; break; case ARM::t2Bcc: isCond = true; UOpc = ARM::t2B; Bits = 20; Scale = 2; break; case ARM::t2B: Bits = 24; Scale = 2; break; } // Record this immediate branch. unsigned MaxOffs = ((1 << (Bits-1))-1) * Scale; ImmBranches.push_back(ImmBranch(&I, MaxOffs, isCond, UOpc)); } if (Opc == ARM::tPUSH || Opc == ARM::tPOP_RET) PushPopMIs.push_back(&I); if (Opc == ARM::CONSTPOOL_ENTRY || Opc == ARM::JUMPTABLE_ADDRS || Opc == ARM::JUMPTABLE_INSTS || Opc == ARM::JUMPTABLE_TBB || Opc == ARM::JUMPTABLE_TBH) continue; // Scan the instructions for constant pool operands. for (unsigned op = 0, e = I.getNumOperands(); op != e; ++op) if (I.getOperand(op).isCPI() || (I.getOperand(op).isJTI() && InlineJumpTables)) { // We found one. The addressing mode tells us the max displacement // from the PC that this instruction permits. // Basic size info comes from the TSFlags field. unsigned Bits = 0; unsigned Scale = 1; bool NegOk = false; bool IsSoImm = false; switch (Opc) { default: llvm_unreachable("Unknown addressing mode for CP reference!"); // Taking the address of a CP entry. case ARM::LEApcrel: case ARM::LEApcrelJT: { // This takes a SoImm, which is 8 bit immediate rotated. We'll // pretend the maximum offset is 255 * 4. Since each instruction // 4 byte wide, this is always correct. We'll check for other // displacements that fits in a SoImm as well. Bits = 8; NegOk = true; IsSoImm = true; unsigned CPI = I.getOperand(op).getIndex(); assert(CPI < CPEMIs.size()); MachineInstr *CPEMI = CPEMIs[CPI]; const Align CPEAlign = getCPEAlign(CPEMI); const unsigned LogCPEAlign = Log2(CPEAlign); if (LogCPEAlign >= 2) Scale = 4; else // For constants with less than 4-byte alignment, // we'll pretend the maximum offset is 255 * 1. Scale = 1; } break; case ARM::t2LEApcrel: case ARM::t2LEApcrelJT: Bits = 12; NegOk = true; break; case ARM::tLEApcrel: case ARM::tLEApcrelJT: Bits = 8; Scale = 4; break; case ARM::LDRBi12: case ARM::LDRi12: case ARM::LDRcp: case ARM::t2LDRpci: case ARM::t2LDRHpci: case ARM::t2LDRSHpci: case ARM::t2LDRBpci: case ARM::t2LDRSBpci: Bits = 12; // +-offset_12 NegOk = true; break; case ARM::tLDRpci: Bits = 8; Scale = 4; // +(offset_8*4) break; case ARM::VLDRD: case ARM::VLDRS: Bits = 8; Scale = 4; // +-(offset_8*4) NegOk = true; break; case ARM::VLDRH: Bits = 8; Scale = 2; // +-(offset_8*2) NegOk = true; break; } // Remember that this is a user of a CP entry. unsigned CPI = I.getOperand(op).getIndex(); if (I.getOperand(op).isJTI()) { JumpTableUserIndices.insert(std::make_pair(CPI, CPUsers.size())); CPI = JumpTableEntryIndices[CPI]; } MachineInstr *CPEMI = CPEMIs[CPI]; unsigned MaxOffs = ((1 << Bits)-1) * Scale; CPUsers.push_back(CPUser(&I, CPEMI, MaxOffs, NegOk, IsSoImm)); // Increment corresponding CPEntry reference count. CPEntry *CPE = findConstPoolEntry(CPI, CPEMI); assert(CPE && "Cannot find a corresponding CPEntry!"); CPE->RefCount++; // Instructions can only use one CP entry, don't bother scanning the // rest of the operands. break; } } } } /// CompareMBBNumbers - Little predicate function to sort the WaterList by MBB /// ID. static bool CompareMBBNumbers(const MachineBasicBlock *LHS, const MachineBasicBlock *RHS) { return LHS->getNumber() < RHS->getNumber(); } /// updateForInsertedWaterBlock - When a block is newly inserted into the /// machine function, it upsets all of the block numbers. Renumber the blocks /// and update the arrays that parallel this numbering. void ARMConstantIslands::updateForInsertedWaterBlock(MachineBasicBlock *NewBB) { // Renumber the MBB's to keep them consecutive. NewBB->getParent()->RenumberBlocks(NewBB); // Insert an entry into BBInfo to align it properly with the (newly // renumbered) block numbers. BBUtils->insert(NewBB->getNumber(), BasicBlockInfo()); // Next, update WaterList. Specifically, we need to add NewMBB as having // available water after it. water_iterator IP = llvm::lower_bound(WaterList, NewBB, CompareMBBNumbers); WaterList.insert(IP, NewBB); } /// Split the basic block containing MI into two blocks, which are joined by /// an unconditional branch. Update data structures and renumber blocks to /// account for this change and returns the newly created block. MachineBasicBlock *ARMConstantIslands::splitBlockBeforeInstr(MachineInstr *MI) { MachineBasicBlock *OrigBB = MI->getParent(); // Collect liveness information at MI. LivePhysRegs LRs(*MF->getSubtarget().getRegisterInfo()); LRs.addLiveOuts(*OrigBB); auto LivenessEnd = ++MachineBasicBlock::iterator(MI).getReverse(); for (MachineInstr &LiveMI : make_range(OrigBB->rbegin(), LivenessEnd)) LRs.stepBackward(LiveMI); // Create a new MBB for the code after the OrigBB. MachineBasicBlock *NewBB = MF->CreateMachineBasicBlock(OrigBB->getBasicBlock()); MachineFunction::iterator MBBI = ++OrigBB->getIterator(); MF->insert(MBBI, NewBB); // Splice the instructions starting with MI over to NewBB. NewBB->splice(NewBB->end(), OrigBB, MI, OrigBB->end()); // Add an unconditional branch from OrigBB to NewBB. // Note the new unconditional branch is not being recorded. // There doesn't seem to be meaningful DebugInfo available; this doesn't // correspond to anything in the source. unsigned Opc = isThumb ? (isThumb2 ? ARM::t2B : ARM::tB) : ARM::B; if (!isThumb) BuildMI(OrigBB, DebugLoc(), TII->get(Opc)).addMBB(NewBB); else BuildMI(OrigBB, DebugLoc(), TII->get(Opc)) .addMBB(NewBB) .add(predOps(ARMCC::AL)); ++NumSplit; // Update the CFG. All succs of OrigBB are now succs of NewBB. NewBB->transferSuccessors(OrigBB); // OrigBB branches to NewBB. OrigBB->addSuccessor(NewBB); // Update live-in information in the new block. MachineRegisterInfo &MRI = MF->getRegInfo(); for (MCPhysReg L : LRs) if (!MRI.isReserved(L)) NewBB->addLiveIn(L); // Update internal data structures to account for the newly inserted MBB. // This is almost the same as updateForInsertedWaterBlock, except that // the Water goes after OrigBB, not NewBB. MF->RenumberBlocks(NewBB); // Insert an entry into BBInfo to align it properly with the (newly // renumbered) block numbers. BBUtils->insert(NewBB->getNumber(), BasicBlockInfo()); // Next, update WaterList. Specifically, we need to add OrigMBB as having // available water after it (but not if it's already there, which happens // when splitting before a conditional branch that is followed by an // unconditional branch - in that case we want to insert NewBB). water_iterator IP = llvm::lower_bound(WaterList, OrigBB, CompareMBBNumbers); MachineBasicBlock* WaterBB = *IP; if (WaterBB == OrigBB) WaterList.insert(std::next(IP), NewBB); else WaterList.insert(IP, OrigBB); NewWaterList.insert(OrigBB); // Figure out how large the OrigBB is. As the first half of the original // block, it cannot contain a tablejump. The size includes // the new jump we added. (It should be possible to do this without // recounting everything, but it's very confusing, and this is rarely // executed.) BBUtils->computeBlockSize(OrigBB); // Figure out how large the NewMBB is. As the second half of the original // block, it may contain a tablejump. BBUtils->computeBlockSize(NewBB); // All BBOffsets following these blocks must be modified. BBUtils->adjustBBOffsetsAfter(OrigBB); return NewBB; } /// getUserOffset - Compute the offset of U.MI as seen by the hardware /// displacement computation. Update U.KnownAlignment to match its current /// basic block location. unsigned ARMConstantIslands::getUserOffset(CPUser &U) const { unsigned UserOffset = BBUtils->getOffsetOf(U.MI); SmallVectorImpl &BBInfo = BBUtils->getBBInfo(); const BasicBlockInfo &BBI = BBInfo[U.MI->getParent()->getNumber()]; unsigned KnownBits = BBI.internalKnownBits(); // The value read from PC is offset from the actual instruction address. UserOffset += (isThumb ? 4 : 8); // Because of inline assembly, we may not know the alignment (mod 4) of U.MI. // Make sure U.getMaxDisp() returns a constrained range. U.KnownAlignment = (KnownBits >= 2); // On Thumb, offsets==2 mod 4 are rounded down by the hardware for // purposes of the displacement computation; compensate for that here. // For unknown alignments, getMaxDisp() constrains the range instead. if (isThumb && U.KnownAlignment) UserOffset &= ~3u; return UserOffset; } /// isOffsetInRange - Checks whether UserOffset (the location of a constant pool /// reference) is within MaxDisp of TrialOffset (a proposed location of a /// constant pool entry). /// UserOffset is computed by getUserOffset above to include PC adjustments. If /// the mod 4 alignment of UserOffset is not known, the uncertainty must be /// subtracted from MaxDisp instead. CPUser::getMaxDisp() does that. bool ARMConstantIslands::isOffsetInRange(unsigned UserOffset, unsigned TrialOffset, unsigned MaxDisp, bool NegativeOK, bool IsSoImm) { if (UserOffset <= TrialOffset) { // User before the Trial. if (TrialOffset - UserOffset <= MaxDisp) return true; // FIXME: Make use full range of soimm values. } else if (NegativeOK) { if (UserOffset - TrialOffset <= MaxDisp) return true; // FIXME: Make use full range of soimm values. } return false; } /// isWaterInRange - Returns true if a CPE placed after the specified /// Water (a basic block) will be in range for the specific MI. /// /// Compute how much the function will grow by inserting a CPE after Water. bool ARMConstantIslands::isWaterInRange(unsigned UserOffset, MachineBasicBlock* Water, CPUser &U, unsigned &Growth) { BBInfoVector &BBInfo = BBUtils->getBBInfo(); const Align CPEAlign = getCPEAlign(U.CPEMI); const unsigned CPEOffset = BBInfo[Water->getNumber()].postOffset(CPEAlign); unsigned NextBlockOffset; Align NextBlockAlignment; MachineFunction::const_iterator NextBlock = Water->getIterator(); if (++NextBlock == MF->end()) { NextBlockOffset = BBInfo[Water->getNumber()].postOffset(); } else { NextBlockOffset = BBInfo[NextBlock->getNumber()].Offset; NextBlockAlignment = NextBlock->getAlignment(); } unsigned Size = U.CPEMI->getOperand(2).getImm(); unsigned CPEEnd = CPEOffset + Size; // The CPE may be able to hide in the alignment padding before the next // block. It may also cause more padding to be required if it is more aligned // that the next block. if (CPEEnd > NextBlockOffset) { Growth = CPEEnd - NextBlockOffset; // Compute the padding that would go at the end of the CPE to align the next // block. Growth += offsetToAlignment(CPEEnd, NextBlockAlignment); // If the CPE is to be inserted before the instruction, that will raise // the offset of the instruction. Also account for unknown alignment padding // in blocks between CPE and the user. if (CPEOffset < UserOffset) UserOffset += Growth + UnknownPadding(MF->getAlignment(), Log2(CPEAlign)); } else // CPE fits in existing padding. Growth = 0; return isOffsetInRange(UserOffset, CPEOffset, U); } /// isCPEntryInRange - Returns true if the distance between specific MI and /// specific ConstPool entry instruction can fit in MI's displacement field. bool ARMConstantIslands::isCPEntryInRange(MachineInstr *MI, unsigned UserOffset, MachineInstr *CPEMI, unsigned MaxDisp, bool NegOk, bool DoDump) { unsigned CPEOffset = BBUtils->getOffsetOf(CPEMI); if (DoDump) { LLVM_DEBUG({ BBInfoVector &BBInfo = BBUtils->getBBInfo(); unsigned Block = MI->getParent()->getNumber(); const BasicBlockInfo &BBI = BBInfo[Block]; dbgs() << "User of CPE#" << CPEMI->getOperand(0).getImm() << " max delta=" << MaxDisp << format(" insn address=%#x", UserOffset) << " in " << printMBBReference(*MI->getParent()) << ": " << format("%#x-%x\t", BBI.Offset, BBI.postOffset()) << *MI << format("CPE address=%#x offset=%+d: ", CPEOffset, int(CPEOffset - UserOffset)); }); } return isOffsetInRange(UserOffset, CPEOffset, MaxDisp, NegOk); } #ifndef NDEBUG /// BBIsJumpedOver - Return true of the specified basic block's only predecessor /// unconditionally branches to its only successor. static bool BBIsJumpedOver(MachineBasicBlock *MBB) { if (MBB->pred_size() != 1 || MBB->succ_size() != 1) return false; MachineBasicBlock *Succ = *MBB->succ_begin(); MachineBasicBlock *Pred = *MBB->pred_begin(); MachineInstr *PredMI = &Pred->back(); if (PredMI->getOpcode() == ARM::B || PredMI->getOpcode() == ARM::tB || PredMI->getOpcode() == ARM::t2B) return PredMI->getOperand(0).getMBB() == Succ; return false; } #endif // NDEBUG /// decrementCPEReferenceCount - find the constant pool entry with index CPI /// and instruction CPEMI, and decrement its refcount. If the refcount /// becomes 0 remove the entry and instruction. Returns true if we removed /// the entry, false if we didn't. bool ARMConstantIslands::decrementCPEReferenceCount(unsigned CPI, MachineInstr *CPEMI) { // Find the old entry. Eliminate it if it is no longer used. CPEntry *CPE = findConstPoolEntry(CPI, CPEMI); assert(CPE && "Unexpected!"); if (--CPE->RefCount == 0) { removeDeadCPEMI(CPEMI); CPE->CPEMI = nullptr; --NumCPEs; return true; } return false; } unsigned ARMConstantIslands::getCombinedIndex(const MachineInstr *CPEMI) { if (CPEMI->getOperand(1).isCPI()) return CPEMI->getOperand(1).getIndex(); return JumpTableEntryIndices[CPEMI->getOperand(1).getIndex()]; } /// LookForCPEntryInRange - see if the currently referenced CPE is in range; /// if not, see if an in-range clone of the CPE is in range, and if so, /// change the data structures so the user references the clone. Returns: /// 0 = no existing entry found /// 1 = entry found, and there were no code insertions or deletions /// 2 = entry found, and there were code insertions or deletions int ARMConstantIslands::findInRangeCPEntry(CPUser& U, unsigned UserOffset) { MachineInstr *UserMI = U.MI; MachineInstr *CPEMI = U.CPEMI; // Check to see if the CPE is already in-range. if (isCPEntryInRange(UserMI, UserOffset, CPEMI, U.getMaxDisp(), U.NegOk, true)) { LLVM_DEBUG(dbgs() << "In range\n"); return 1; } // No. Look for previously created clones of the CPE that are in range. unsigned CPI = getCombinedIndex(CPEMI); std::vector &CPEs = CPEntries[CPI]; for (CPEntry &CPE : CPEs) { // We already tried this one if (CPE.CPEMI == CPEMI) continue; // Removing CPEs can leave empty entries, skip if (CPE.CPEMI == nullptr) continue; if (isCPEntryInRange(UserMI, UserOffset, CPE.CPEMI, U.getMaxDisp(), U.NegOk)) { LLVM_DEBUG(dbgs() << "Replacing CPE#" << CPI << " with CPE#" << CPE.CPI << "\n"); // Point the CPUser node to the replacement U.CPEMI = CPE.CPEMI; // Change the CPI in the instruction operand to refer to the clone. for (MachineOperand &MO : UserMI->operands()) if (MO.isCPI()) { MO.setIndex(CPE.CPI); break; } // Adjust the refcount of the clone... CPE.RefCount++; // ...and the original. If we didn't remove the old entry, none of the // addresses changed, so we don't need another pass. return decrementCPEReferenceCount(CPI, CPEMI) ? 2 : 1; } } return 0; } /// getUnconditionalBrDisp - Returns the maximum displacement that can fit in /// the specific unconditional branch instruction. static inline unsigned getUnconditionalBrDisp(int Opc) { switch (Opc) { case ARM::tB: return ((1<<10)-1)*2; case ARM::t2B: return ((1<<23)-1)*2; default: break; } return ((1<<23)-1)*4; } /// findAvailableWater - Look for an existing entry in the WaterList in which /// we can place the CPE referenced from U so it's within range of U's MI. /// Returns true if found, false if not. If it returns true, WaterIter /// is set to the WaterList entry. For Thumb, prefer water that will not /// introduce padding to water that will. To ensure that this pass /// terminates, the CPE location for a particular CPUser is only allowed to /// move to a lower address, so search backward from the end of the list and /// prefer the first water that is in range. bool ARMConstantIslands::findAvailableWater(CPUser &U, unsigned UserOffset, water_iterator &WaterIter, bool CloserWater) { if (WaterList.empty()) return false; unsigned BestGrowth = ~0u; // The nearest water without splitting the UserBB is right after it. // If the distance is still large (we have a big BB), then we need to split it // if we don't converge after certain iterations. This helps the following // situation to converge: // BB0: // Big BB // BB1: // Constant Pool // When a CP access is out of range, BB0 may be used as water. However, // inserting islands between BB0 and BB1 makes other accesses out of range. MachineBasicBlock *UserBB = U.MI->getParent(); BBInfoVector &BBInfo = BBUtils->getBBInfo(); const Align CPEAlign = getCPEAlign(U.CPEMI); unsigned MinNoSplitDisp = BBInfo[UserBB->getNumber()].postOffset(CPEAlign); if (CloserWater && MinNoSplitDisp > U.getMaxDisp() / 2) return false; for (water_iterator IP = std::prev(WaterList.end()), B = WaterList.begin();; --IP) { MachineBasicBlock* WaterBB = *IP; // Check if water is in range and is either at a lower address than the // current "high water mark" or a new water block that was created since // the previous iteration by inserting an unconditional branch. In the // latter case, we want to allow resetting the high water mark back to // this new water since we haven't seen it before. Inserting branches // should be relatively uncommon and when it does happen, we want to be // sure to take advantage of it for all the CPEs near that block, so that // we don't insert more branches than necessary. // When CloserWater is true, we try to find the lowest address after (or // equal to) user MI's BB no matter of padding growth. unsigned Growth; if (isWaterInRange(UserOffset, WaterBB, U, Growth) && (WaterBB->getNumber() < U.HighWaterMark->getNumber() || NewWaterList.count(WaterBB) || WaterBB == U.MI->getParent()) && Growth < BestGrowth) { // This is the least amount of required padding seen so far. BestGrowth = Growth; WaterIter = IP; LLVM_DEBUG(dbgs() << "Found water after " << printMBBReference(*WaterBB) << " Growth=" << Growth << '\n'); if (CloserWater && WaterBB == U.MI->getParent()) return true; // Keep looking unless it is perfect and we're not looking for the lowest // possible address. if (!CloserWater && BestGrowth == 0) return true; } if (IP == B) break; } return BestGrowth != ~0u; } /// createNewWater - No existing WaterList entry will work for /// CPUsers[CPUserIndex], so create a place to put the CPE. The end of the /// block is used if in range, and the conditional branch munged so control /// flow is correct. Otherwise the block is split to create a hole with an /// unconditional branch around it. In either case NewMBB is set to a /// block following which the new island can be inserted (the WaterList /// is not adjusted). void ARMConstantIslands::createNewWater(unsigned CPUserIndex, unsigned UserOffset, MachineBasicBlock *&NewMBB) { CPUser &U = CPUsers[CPUserIndex]; MachineInstr *UserMI = U.MI; MachineInstr *CPEMI = U.CPEMI; const Align CPEAlign = getCPEAlign(CPEMI); MachineBasicBlock *UserMBB = UserMI->getParent(); BBInfoVector &BBInfo = BBUtils->getBBInfo(); const BasicBlockInfo &UserBBI = BBInfo[UserMBB->getNumber()]; // If the block does not end in an unconditional branch already, and if the // end of the block is within range, make new water there. (The addition // below is for the unconditional branch we will be adding: 4 bytes on ARM + // Thumb2, 2 on Thumb1. if (BBHasFallthrough(UserMBB)) { // Size of branch to insert. unsigned Delta = isThumb1 ? 2 : 4; // Compute the offset where the CPE will begin. unsigned CPEOffset = UserBBI.postOffset(CPEAlign) + Delta; if (isOffsetInRange(UserOffset, CPEOffset, U)) { LLVM_DEBUG(dbgs() << "Split at end of " << printMBBReference(*UserMBB) << format(", expected CPE offset %#x\n", CPEOffset)); NewMBB = &*++UserMBB->getIterator(); // Add an unconditional branch from UserMBB to fallthrough block. Record // it for branch lengthening; this new branch will not get out of range, // but if the preceding conditional branch is out of range, the targets // will be exchanged, and the altered branch may be out of range, so the // machinery has to know about it. int UncondBr = isThumb ? ((isThumb2) ? ARM::t2B : ARM::tB) : ARM::B; if (!isThumb) BuildMI(UserMBB, DebugLoc(), TII->get(UncondBr)).addMBB(NewMBB); else BuildMI(UserMBB, DebugLoc(), TII->get(UncondBr)) .addMBB(NewMBB) .add(predOps(ARMCC::AL)); unsigned MaxDisp = getUnconditionalBrDisp(UncondBr); ImmBranches.push_back(ImmBranch(&UserMBB->back(), MaxDisp, false, UncondBr)); BBUtils->computeBlockSize(UserMBB); BBUtils->adjustBBOffsetsAfter(UserMBB); return; } } // What a big block. Find a place within the block to split it. This is a // little tricky on Thumb1 since instructions are 2 bytes and constant pool // entries are 4 bytes: if instruction I references island CPE, and // instruction I+1 references CPE', it will not work well to put CPE as far // forward as possible, since then CPE' cannot immediately follow it (that // location is 2 bytes farther away from I+1 than CPE was from I) and we'd // need to create a new island. So, we make a first guess, then walk through // the instructions between the one currently being looked at and the // possible insertion point, and make sure any other instructions that // reference CPEs will be able to use the same island area; if not, we back // up the insertion point. // Try to split the block so it's fully aligned. Compute the latest split // point where we can add a 4-byte branch instruction, and then align to // Align which is the largest possible alignment in the function. const Align Align = MF->getAlignment(); assert(Align >= CPEAlign && "Over-aligned constant pool entry"); unsigned KnownBits = UserBBI.internalKnownBits(); unsigned UPad = UnknownPadding(Align, KnownBits); unsigned BaseInsertOffset = UserOffset + U.getMaxDisp() - UPad; LLVM_DEBUG(dbgs() << format("Split in middle of big block before %#x", BaseInsertOffset)); // The 4 in the following is for the unconditional branch we'll be inserting // (allows for long branch on Thumb1). Alignment of the island is handled // inside isOffsetInRange. BaseInsertOffset -= 4; LLVM_DEBUG(dbgs() << format(", adjusted to %#x", BaseInsertOffset) << " la=" << Log2(Align) << " kb=" << KnownBits << " up=" << UPad << '\n'); // This could point off the end of the block if we've already got constant // pool entries following this block; only the last one is in the water list. // Back past any possible branches (allow for a conditional and a maximally // long unconditional). if (BaseInsertOffset + 8 >= UserBBI.postOffset()) { // Ensure BaseInsertOffset is larger than the offset of the instruction // following UserMI so that the loop which searches for the split point // iterates at least once. BaseInsertOffset = std::max(UserBBI.postOffset() - UPad - 8, UserOffset + TII->getInstSizeInBytes(*UserMI) + 1); // If the CP is referenced(ie, UserOffset) is in first four instructions // after IT, this recalculated BaseInsertOffset could be in the middle of // an IT block. If it is, change the BaseInsertOffset to just after the // IT block. This still make the CP Entry is in range becuase of the // following reasons. // 1. The initial BaseseInsertOffset calculated is (UserOffset + // U.getMaxDisp() - UPad). // 2. An IT block is only at most 4 instructions plus the "it" itself (18 // bytes). // 3. All the relevant instructions support much larger Maximum // displacement. MachineBasicBlock::iterator I = UserMI; ++I; Register PredReg; for (unsigned Offset = UserOffset + TII->getInstSizeInBytes(*UserMI); I->getOpcode() != ARM::t2IT && getITInstrPredicate(*I, PredReg) != ARMCC::AL; Offset += TII->getInstSizeInBytes(*I), I = std::next(I)) { BaseInsertOffset = std::max(BaseInsertOffset, Offset + TII->getInstSizeInBytes(*I) + 1); assert(I != UserMBB->end() && "Fell off end of block"); } LLVM_DEBUG(dbgs() << format("Move inside block: %#x\n", BaseInsertOffset)); } unsigned EndInsertOffset = BaseInsertOffset + 4 + UPad + CPEMI->getOperand(2).getImm(); MachineBasicBlock::iterator MI = UserMI; ++MI; unsigned CPUIndex = CPUserIndex+1; unsigned NumCPUsers = CPUsers.size(); MachineInstr *LastIT = nullptr; for (unsigned Offset = UserOffset + TII->getInstSizeInBytes(*UserMI); Offset < BaseInsertOffset; Offset += TII->getInstSizeInBytes(*MI), MI = std::next(MI)) { assert(MI != UserMBB->end() && "Fell off end of block"); if (CPUIndex < NumCPUsers && CPUsers[CPUIndex].MI == &*MI) { CPUser &U = CPUsers[CPUIndex]; if (!isOffsetInRange(Offset, EndInsertOffset, U)) { // Shift intertion point by one unit of alignment so it is within reach. BaseInsertOffset -= Align.value(); EndInsertOffset -= Align.value(); } // This is overly conservative, as we don't account for CPEMIs being // reused within the block, but it doesn't matter much. Also assume CPEs // are added in order with alignment padding. We may eventually be able // to pack the aligned CPEs better. EndInsertOffset += U.CPEMI->getOperand(2).getImm(); CPUIndex++; } // Remember the last IT instruction. if (MI->getOpcode() == ARM::t2IT) LastIT = &*MI; } --MI; // Avoid splitting an IT block. if (LastIT) { Register PredReg; ARMCC::CondCodes CC = getITInstrPredicate(*MI, PredReg); if (CC != ARMCC::AL) MI = LastIT; } // Avoid splitting a MOVW+MOVT pair with a relocation on Windows. // On Windows, this instruction pair is covered by one single // IMAGE_REL_ARM_MOV32T relocation which covers both instructions. If a // constant island is injected inbetween them, the relocation will clobber // the instruction and fail to update the MOVT instruction. // (These instructions are bundled up until right before the ConstantIslands // pass.) if (STI->isTargetWindows() && isThumb && MI->getOpcode() == ARM::t2MOVTi16 && (MI->getOperand(2).getTargetFlags() & ARMII::MO_OPTION_MASK) == ARMII::MO_HI16) { --MI; assert(MI->getOpcode() == ARM::t2MOVi16 && (MI->getOperand(1).getTargetFlags() & ARMII::MO_OPTION_MASK) == ARMII::MO_LO16); } // We really must not split an IT block. #ifndef NDEBUG Register PredReg; assert(!isThumb || getITInstrPredicate(*MI, PredReg) == ARMCC::AL); #endif NewMBB = splitBlockBeforeInstr(&*MI); } /// handleConstantPoolUser - Analyze the specified user, checking to see if it /// is out-of-range. If so, pick up the constant pool value and move it some /// place in-range. Return true if we changed any addresses (thus must run /// another pass of branch lengthening), false otherwise. bool ARMConstantIslands::handleConstantPoolUser(unsigned CPUserIndex, bool CloserWater) { CPUser &U = CPUsers[CPUserIndex]; MachineInstr *UserMI = U.MI; MachineInstr *CPEMI = U.CPEMI; unsigned CPI = getCombinedIndex(CPEMI); unsigned Size = CPEMI->getOperand(2).getImm(); // Compute this only once, it's expensive. unsigned UserOffset = getUserOffset(U); // See if the current entry is within range, or there is a clone of it // in range. int result = findInRangeCPEntry(U, UserOffset); if (result==1) return false; else if (result==2) return true; // No existing clone of this CPE is within range. // We will be generating a new clone. Get a UID for it. unsigned ID = AFI->createPICLabelUId(); // Look for water where we can place this CPE. MachineBasicBlock *NewIsland = MF->CreateMachineBasicBlock(); MachineBasicBlock *NewMBB; water_iterator IP; if (findAvailableWater(U, UserOffset, IP, CloserWater)) { LLVM_DEBUG(dbgs() << "Found water in range\n"); MachineBasicBlock *WaterBB = *IP; // If the original WaterList entry was "new water" on this iteration, // propagate that to the new island. This is just keeping NewWaterList // updated to match the WaterList, which will be updated below. if (NewWaterList.erase(WaterBB)) NewWaterList.insert(NewIsland); // The new CPE goes before the following block (NewMBB). NewMBB = &*++WaterBB->getIterator(); } else { // No water found. LLVM_DEBUG(dbgs() << "No water found\n"); createNewWater(CPUserIndex, UserOffset, NewMBB); // splitBlockBeforeInstr adds to WaterList, which is important when it is // called while handling branches so that the water will be seen on the // next iteration for constant pools, but in this context, we don't want // it. Check for this so it will be removed from the WaterList. // Also remove any entry from NewWaterList. MachineBasicBlock *WaterBB = &*--NewMBB->getIterator(); IP = find(WaterList, WaterBB); if (IP != WaterList.end()) NewWaterList.erase(WaterBB); // We are adding new water. Update NewWaterList. NewWaterList.insert(NewIsland); } // Always align the new block because CP entries can be smaller than 4 // bytes. Be careful not to decrease the existing alignment, e.g. NewMBB may // be an already aligned constant pool block. const Align Alignment = isThumb ? Align(2) : Align(4); if (NewMBB->getAlignment() < Alignment) NewMBB->setAlignment(Alignment); // Remove the original WaterList entry; we want subsequent insertions in // this vicinity to go after the one we're about to insert. This // considerably reduces the number of times we have to move the same CPE // more than once and is also important to ensure the algorithm terminates. if (IP != WaterList.end()) WaterList.erase(IP); // Okay, we know we can put an island before NewMBB now, do it! MF->insert(NewMBB->getIterator(), NewIsland); // Update internal data structures to account for the newly inserted MBB. updateForInsertedWaterBlock(NewIsland); // Now that we have an island to add the CPE to, clone the original CPE and // add it to the island. U.HighWaterMark = NewIsland; U.CPEMI = BuildMI(NewIsland, DebugLoc(), CPEMI->getDesc()) .addImm(ID) .add(CPEMI->getOperand(1)) .addImm(Size); CPEntries[CPI].push_back(CPEntry(U.CPEMI, ID, 1)); ++NumCPEs; // Decrement the old entry, and remove it if refcount becomes 0. decrementCPEReferenceCount(CPI, CPEMI); // Mark the basic block as aligned as required by the const-pool entry. NewIsland->setAlignment(getCPEAlign(U.CPEMI)); // Increase the size of the island block to account for the new entry. BBUtils->adjustBBSize(NewIsland, Size); BBUtils->adjustBBOffsetsAfter(&*--NewIsland->getIterator()); // Finally, change the CPI in the instruction operand to be ID. for (MachineOperand &MO : UserMI->operands()) if (MO.isCPI()) { MO.setIndex(ID); break; } LLVM_DEBUG( dbgs() << " Moved CPE to #" << ID << " CPI=" << CPI << format(" offset=%#x\n", BBUtils->getBBInfo()[NewIsland->getNumber()].Offset)); return true; } /// removeDeadCPEMI - Remove a dead constant pool entry instruction. Update /// sizes and offsets of impacted basic blocks. void ARMConstantIslands::removeDeadCPEMI(MachineInstr *CPEMI) { MachineBasicBlock *CPEBB = CPEMI->getParent(); unsigned Size = CPEMI->getOperand(2).getImm(); CPEMI->eraseFromParent(); BBInfoVector &BBInfo = BBUtils->getBBInfo(); BBUtils->adjustBBSize(CPEBB, -Size); // All succeeding offsets have the current size value added in, fix this. if (CPEBB->empty()) { BBInfo[CPEBB->getNumber()].Size = 0; // This block no longer needs to be aligned. CPEBB->setAlignment(Align(1)); } else { // Entries are sorted by descending alignment, so realign from the front. CPEBB->setAlignment(getCPEAlign(&*CPEBB->begin())); } BBUtils->adjustBBOffsetsAfter(CPEBB); // An island has only one predecessor BB and one successor BB. Check if // this BB's predecessor jumps directly to this BB's successor. This // shouldn't happen currently. assert(!BBIsJumpedOver(CPEBB) && "How did this happen?"); // FIXME: remove the empty blocks after all the work is done? } /// removeUnusedCPEntries - Remove constant pool entries whose refcounts /// are zero. bool ARMConstantIslands::removeUnusedCPEntries() { unsigned MadeChange = false; for (std::vector &CPEs : CPEntries) { for (CPEntry &CPE : CPEs) { if (CPE.RefCount == 0 && CPE.CPEMI) { removeDeadCPEMI(CPE.CPEMI); CPE.CPEMI = nullptr; MadeChange = true; } } } return MadeChange; } /// fixupImmediateBr - Fix up an immediate branch whose destination is too far /// away to fit in its displacement field. bool ARMConstantIslands::fixupImmediateBr(ImmBranch &Br) { MachineInstr *MI = Br.MI; MachineBasicBlock *DestBB = MI->getOperand(0).getMBB(); // Check to see if the DestBB is already in-range. if (BBUtils->isBBInRange(MI, DestBB, Br.MaxDisp)) return false; if (!Br.isCond) return fixupUnconditionalBr(Br); return fixupConditionalBr(Br); } /// fixupUnconditionalBr - Fix up an unconditional branch whose destination is /// too far away to fit in its displacement field. If the LR register has been /// spilled in the epilogue, then we can use BL to implement a far jump. /// Otherwise, add an intermediate branch instruction to a branch. bool ARMConstantIslands::fixupUnconditionalBr(ImmBranch &Br) { MachineInstr *MI = Br.MI; MachineBasicBlock *MBB = MI->getParent(); if (!isThumb1) llvm_unreachable("fixupUnconditionalBr is Thumb1 only!"); if (!AFI->isLRSpilled()) report_fatal_error("underestimated function size"); // Use BL to implement far jump. Br.MaxDisp = (1 << 21) * 2; MI->setDesc(TII->get(ARM::tBfar)); BBInfoVector &BBInfo = BBUtils->getBBInfo(); BBInfo[MBB->getNumber()].Size += 2; BBUtils->adjustBBOffsetsAfter(MBB); ++NumUBrFixed; LLVM_DEBUG(dbgs() << " Changed B to long jump " << *MI); return true; } /// fixupConditionalBr - Fix up a conditional branch whose destination is too /// far away to fit in its displacement field. It is converted to an inverse /// conditional branch + an unconditional branch to the destination. bool ARMConstantIslands::fixupConditionalBr(ImmBranch &Br) { MachineInstr *MI = Br.MI; MachineBasicBlock *DestBB = MI->getOperand(0).getMBB(); // Add an unconditional branch to the destination and invert the branch // condition to jump over it: // blt L1 // => // bge L2 // b L1 // L2: ARMCC::CondCodes CC = (ARMCC::CondCodes)MI->getOperand(1).getImm(); CC = ARMCC::getOppositeCondition(CC); Register CCReg = MI->getOperand(2).getReg(); // If the branch is at the end of its MBB and that has a fall-through block, // direct the updated conditional branch to the fall-through block. Otherwise, // split the MBB before the next instruction. MachineBasicBlock *MBB = MI->getParent(); MachineInstr *BMI = &MBB->back(); bool NeedSplit = (BMI != MI) || !BBHasFallthrough(MBB); ++NumCBrFixed; if (BMI != MI) { if (std::next(MachineBasicBlock::iterator(MI)) == std::prev(MBB->end()) && BMI->getOpcode() == Br.UncondBr) { // Last MI in the BB is an unconditional branch. Can we simply invert the // condition and swap destinations: // beq L1 // b L2 // => // bne L2 // b L1 MachineBasicBlock *NewDest = BMI->getOperand(0).getMBB(); if (BBUtils->isBBInRange(MI, NewDest, Br.MaxDisp)) { LLVM_DEBUG( dbgs() << " Invert Bcc condition and swap its destination with " << *BMI); BMI->getOperand(0).setMBB(DestBB); MI->getOperand(0).setMBB(NewDest); MI->getOperand(1).setImm(CC); return true; } } } if (NeedSplit) { splitBlockBeforeInstr(MI); // No need for the branch to the next block. We're adding an unconditional // branch to the destination. int delta = TII->getInstSizeInBytes(MBB->back()); BBUtils->adjustBBSize(MBB, -delta); MBB->back().eraseFromParent(); // The conditional successor will be swapped between the BBs after this, so // update CFG. MBB->addSuccessor(DestBB); std::next(MBB->getIterator())->removeSuccessor(DestBB); // BBInfo[SplitBB].Offset is wrong temporarily, fixed below } MachineBasicBlock *NextBB = &*++MBB->getIterator(); LLVM_DEBUG(dbgs() << " Insert B to " << printMBBReference(*DestBB) << " also invert condition and change dest. to " << printMBBReference(*NextBB) << "\n"); // Insert a new conditional branch and a new unconditional branch. // Also update the ImmBranch as well as adding a new entry for the new branch. BuildMI(MBB, DebugLoc(), TII->get(MI->getOpcode())) .addMBB(NextBB).addImm(CC).addReg(CCReg); Br.MI = &MBB->back(); BBUtils->adjustBBSize(MBB, TII->getInstSizeInBytes(MBB->back())); if (isThumb) BuildMI(MBB, DebugLoc(), TII->get(Br.UncondBr)) .addMBB(DestBB) .add(predOps(ARMCC::AL)); else BuildMI(MBB, DebugLoc(), TII->get(Br.UncondBr)).addMBB(DestBB); BBUtils->adjustBBSize(MBB, TII->getInstSizeInBytes(MBB->back())); unsigned MaxDisp = getUnconditionalBrDisp(Br.UncondBr); ImmBranches.push_back(ImmBranch(&MBB->back(), MaxDisp, false, Br.UncondBr)); // Remove the old conditional branch. It may or may not still be in MBB. BBUtils->adjustBBSize(MI->getParent(), -TII->getInstSizeInBytes(*MI)); MI->eraseFromParent(); BBUtils->adjustBBOffsetsAfter(MBB); return true; } bool ARMConstantIslands::optimizeThumb2Instructions() { bool MadeChange = false; // Shrink ADR and LDR from constantpool. for (CPUser &U : CPUsers) { unsigned Opcode = U.MI->getOpcode(); unsigned NewOpc = 0; unsigned Scale = 1; unsigned Bits = 0; switch (Opcode) { default: break; case ARM::t2LEApcrel: if (isARMLowRegister(U.MI->getOperand(0).getReg())) { NewOpc = ARM::tLEApcrel; Bits = 8; Scale = 4; } break; case ARM::t2LDRpci: if (isARMLowRegister(U.MI->getOperand(0).getReg())) { NewOpc = ARM::tLDRpci; Bits = 8; Scale = 4; } break; } if (!NewOpc) continue; unsigned UserOffset = getUserOffset(U); unsigned MaxOffs = ((1 << Bits) - 1) * Scale; // Be conservative with inline asm. if (!U.KnownAlignment) MaxOffs -= 2; // FIXME: Check if offset is multiple of scale if scale is not 4. if (isCPEntryInRange(U.MI, UserOffset, U.CPEMI, MaxOffs, false, true)) { LLVM_DEBUG(dbgs() << "Shrink: " << *U.MI); U.MI->setDesc(TII->get(NewOpc)); MachineBasicBlock *MBB = U.MI->getParent(); BBUtils->adjustBBSize(MBB, -2); BBUtils->adjustBBOffsetsAfter(MBB); ++NumT2CPShrunk; MadeChange = true; } } return MadeChange; } bool ARMConstantIslands::optimizeThumb2Branches() { auto TryShrinkBranch = [this](ImmBranch &Br) { unsigned Opcode = Br.MI->getOpcode(); unsigned NewOpc = 0; unsigned Scale = 1; unsigned Bits = 0; switch (Opcode) { default: break; case ARM::t2B: NewOpc = ARM::tB; Bits = 11; Scale = 2; break; case ARM::t2Bcc: NewOpc = ARM::tBcc; Bits = 8; Scale = 2; break; } if (NewOpc) { unsigned MaxOffs = ((1 << (Bits-1))-1) * Scale; MachineBasicBlock *DestBB = Br.MI->getOperand(0).getMBB(); if (BBUtils->isBBInRange(Br.MI, DestBB, MaxOffs)) { LLVM_DEBUG(dbgs() << "Shrink branch: " << *Br.MI); Br.MI->setDesc(TII->get(NewOpc)); MachineBasicBlock *MBB = Br.MI->getParent(); BBUtils->adjustBBSize(MBB, -2); BBUtils->adjustBBOffsetsAfter(MBB); ++NumT2BrShrunk; return true; } } return false; }; struct ImmCompare { MachineInstr* MI = nullptr; unsigned NewOpc = 0; }; auto FindCmpForCBZ = [this](ImmBranch &Br, ImmCompare &ImmCmp, MachineBasicBlock *DestBB) { ImmCmp.MI = nullptr; ImmCmp.NewOpc = 0; // If the conditional branch doesn't kill CPSR, then CPSR can be liveout // so this transformation is not safe. if (!Br.MI->killsRegister(ARM::CPSR)) return false; Register PredReg; unsigned NewOpc = 0; ARMCC::CondCodes Pred = getInstrPredicate(*Br.MI, PredReg); if (Pred == ARMCC::EQ) NewOpc = ARM::tCBZ; else if (Pred == ARMCC::NE) NewOpc = ARM::tCBNZ; else return false; // Check if the distance is within 126. Subtract starting offset by 2 // because the cmp will be eliminated. unsigned BrOffset = BBUtils->getOffsetOf(Br.MI) + 4 - 2; BBInfoVector &BBInfo = BBUtils->getBBInfo(); unsigned DestOffset = BBInfo[DestBB->getNumber()].Offset; if (BrOffset >= DestOffset || (DestOffset - BrOffset) > 126) return false; // Search backwards to find a tCMPi8 auto *TRI = STI->getRegisterInfo(); MachineInstr *CmpMI = findCMPToFoldIntoCBZ(Br.MI, TRI); if (!CmpMI || CmpMI->getOpcode() != ARM::tCMPi8) return false; ImmCmp.MI = CmpMI; ImmCmp.NewOpc = NewOpc; return true; }; auto TryConvertToLE = [this](ImmBranch &Br, ImmCompare &Cmp) { if (Br.MI->getOpcode() != ARM::t2Bcc || !STI->hasLOB() || STI->hasMinSize()) return false; MachineBasicBlock *MBB = Br.MI->getParent(); MachineBasicBlock *DestBB = Br.MI->getOperand(0).getMBB(); if (BBUtils->getOffsetOf(MBB) < BBUtils->getOffsetOf(DestBB) || !BBUtils->isBBInRange(Br.MI, DestBB, 4094)) return false; if (!DT->dominates(DestBB, MBB)) return false; // We queried for the CBN?Z opcode based upon the 'ExitBB', the opposite // target of Br. So now we need to reverse the condition. Cmp.NewOpc = Cmp.NewOpc == ARM::tCBZ ? ARM::tCBNZ : ARM::tCBZ; MachineInstrBuilder MIB = BuildMI(*MBB, Br.MI, Br.MI->getDebugLoc(), TII->get(ARM::t2LE)); // Swapped a t2Bcc for a t2LE, so no need to update the size of the block. MIB.add(Br.MI->getOperand(0)); Br.MI->eraseFromParent(); Br.MI = MIB; ++NumLEInserted; return true; }; bool MadeChange = false; // The order in which branches appear in ImmBranches is approximately their // order within the function body. By visiting later branches first, we reduce // the distance between earlier forward branches and their targets, making it // more likely that the cbn?z optimization, which can only apply to forward // branches, will succeed. for (ImmBranch &Br : reverse(ImmBranches)) { MachineBasicBlock *DestBB = Br.MI->getOperand(0).getMBB(); MachineBasicBlock *MBB = Br.MI->getParent(); MachineBasicBlock *ExitBB = &MBB->back() == Br.MI ? MBB->getFallThrough() : MBB->back().getOperand(0).getMBB(); ImmCompare Cmp; if (FindCmpForCBZ(Br, Cmp, ExitBB) && TryConvertToLE(Br, Cmp)) { DestBB = ExitBB; MadeChange = true; } else { FindCmpForCBZ(Br, Cmp, DestBB); MadeChange |= TryShrinkBranch(Br); } unsigned Opcode = Br.MI->getOpcode(); if ((Opcode != ARM::tBcc && Opcode != ARM::t2LE) || !Cmp.NewOpc) continue; Register Reg = Cmp.MI->getOperand(0).getReg(); // Check for Kill flags on Reg. If they are present remove them and set kill // on the new CBZ. auto *TRI = STI->getRegisterInfo(); MachineBasicBlock::iterator KillMI = Br.MI; bool RegKilled = false; do { --KillMI; if (KillMI->killsRegister(Reg, TRI)) { KillMI->clearRegisterKills(Reg, TRI); RegKilled = true; break; } } while (KillMI != Cmp.MI); // Create the new CBZ/CBNZ LLVM_DEBUG(dbgs() << "Fold: " << *Cmp.MI << " and: " << *Br.MI); MachineInstr *NewBR = BuildMI(*MBB, Br.MI, Br.MI->getDebugLoc(), TII->get(Cmp.NewOpc)) .addReg(Reg, getKillRegState(RegKilled) | getRegState(Cmp.MI->getOperand(0))) .addMBB(DestBB, Br.MI->getOperand(0).getTargetFlags()); Cmp.MI->eraseFromParent(); if (Br.MI->getOpcode() == ARM::tBcc) { Br.MI->eraseFromParent(); Br.MI = NewBR; BBUtils->adjustBBSize(MBB, -2); } else if (MBB->back().getOpcode() != ARM::t2LE) { // An LE has been generated, but it's not the terminator - that is an // unconditional branch. However, the logic has now been reversed with the // CBN?Z being the conditional branch and the LE being the unconditional // branch. So this means we can remove the redundant unconditional branch // at the end of the block. MachineInstr *LastMI = &MBB->back(); BBUtils->adjustBBSize(MBB, -LastMI->getDesc().getSize()); LastMI->eraseFromParent(); } BBUtils->adjustBBOffsetsAfter(MBB); ++NumCBZ; MadeChange = true; } return MadeChange; } static bool isSimpleIndexCalc(MachineInstr &I, unsigned EntryReg, unsigned BaseReg) { if (I.getOpcode() != ARM::t2ADDrs) return false; if (I.getOperand(0).getReg() != EntryReg) return false; if (I.getOperand(1).getReg() != BaseReg) return false; // FIXME: what about CC and IdxReg? return true; } /// While trying to form a TBB/TBH instruction, we may (if the table /// doesn't immediately follow the BR_JT) need access to the start of the /// jump-table. We know one instruction that produces such a register; this /// function works out whether that definition can be preserved to the BR_JT, /// possibly by removing an intervening addition (which is usually needed to /// calculate the actual entry to jump to). bool ARMConstantIslands::preserveBaseRegister(MachineInstr *JumpMI, MachineInstr *LEAMI, unsigned &DeadSize, bool &CanDeleteLEA, bool &BaseRegKill) { if (JumpMI->getParent() != LEAMI->getParent()) return false; // Now we hope that we have at least these instructions in the basic block: // BaseReg = t2LEA ... // [...] // EntryReg = t2ADDrs BaseReg, ... // [...] // t2BR_JT EntryReg // // We have to be very conservative about what we recognise here though. The // main perturbing factors to watch out for are: // + Spills at any point in the chain: not direct problems but we would // expect a blocking Def of the spilled register so in practice what we // can do is limited. // + EntryReg == BaseReg: this is the one situation we should allow a Def // of BaseReg, but only if the t2ADDrs can be removed. // + Some instruction other than t2ADDrs computing the entry. Not seen in // the wild, but we should be careful. Register EntryReg = JumpMI->getOperand(0).getReg(); Register BaseReg = LEAMI->getOperand(0).getReg(); CanDeleteLEA = true; BaseRegKill = false; MachineInstr *RemovableAdd = nullptr; MachineBasicBlock::iterator I(LEAMI); for (++I; &*I != JumpMI; ++I) { if (isSimpleIndexCalc(*I, EntryReg, BaseReg)) { RemovableAdd = &*I; break; } for (const MachineOperand &MO : I->operands()) { if (!MO.isReg() || !MO.getReg()) continue; if (MO.isDef() && MO.getReg() == BaseReg) return false; if (MO.isUse() && MO.getReg() == BaseReg) { BaseRegKill = BaseRegKill || MO.isKill(); CanDeleteLEA = false; } } } if (!RemovableAdd) return true; // Check the add really is removable, and that nothing else in the block // clobbers BaseReg. for (++I; &*I != JumpMI; ++I) { for (const MachineOperand &MO : I->operands()) { if (!MO.isReg() || !MO.getReg()) continue; if (MO.isDef() && MO.getReg() == BaseReg) return false; if (MO.isUse() && MO.getReg() == EntryReg) RemovableAdd = nullptr; } } if (RemovableAdd) { RemovableAdd->eraseFromParent(); DeadSize += isThumb2 ? 4 : 2; } else if (BaseReg == EntryReg) { // The add wasn't removable, but clobbered the base for the TBB. So we can't // preserve it. return false; } // We reached the end of the block without seeing another definition of // BaseReg (except, possibly the t2ADDrs, which was removed). BaseReg can be // used in the TBB/TBH if necessary. return true; } /// Returns whether CPEMI is the first instruction in the block /// immediately following JTMI (assumed to be a TBB or TBH terminator). If so, /// we can switch the first register to PC and usually remove the address /// calculation that preceded it. static bool jumpTableFollowsTB(MachineInstr *JTMI, MachineInstr *CPEMI) { MachineFunction::iterator MBB = JTMI->getParent()->getIterator(); MachineFunction *MF = MBB->getParent(); ++MBB; return MBB != MF->end() && !MBB->empty() && &*MBB->begin() == CPEMI; } static void RemoveDeadAddBetweenLEAAndJT(MachineInstr *LEAMI, MachineInstr *JumpMI, unsigned &DeadSize) { // Remove a dead add between the LEA and JT, which used to compute EntryReg, // but the JT now uses PC. Finds the last ADD (if any) that def's EntryReg // and is not clobbered / used. MachineInstr *RemovableAdd = nullptr; Register EntryReg = JumpMI->getOperand(0).getReg(); // Find the last ADD to set EntryReg MachineBasicBlock::iterator I(LEAMI); for (++I; &*I != JumpMI; ++I) { if (I->getOpcode() == ARM::t2ADDrs && I->getOperand(0).getReg() == EntryReg) RemovableAdd = &*I; } if (!RemovableAdd) return; // Ensure EntryReg is not clobbered or used. MachineBasicBlock::iterator J(RemovableAdd); for (++J; &*J != JumpMI; ++J) { for (const MachineOperand &MO : J->operands()) { if (!MO.isReg() || !MO.getReg()) continue; if (MO.isDef() && MO.getReg() == EntryReg) return; if (MO.isUse() && MO.getReg() == EntryReg) return; } } LLVM_DEBUG(dbgs() << "Removing Dead Add: " << *RemovableAdd); RemovableAdd->eraseFromParent(); DeadSize += 4; } /// optimizeThumb2JumpTables - Use tbb / tbh instructions to generate smaller /// jumptables when it's possible. bool ARMConstantIslands::optimizeThumb2JumpTables() { bool MadeChange = false; // FIXME: After the tables are shrunk, can we get rid some of the // constantpool tables? MachineJumpTableInfo *MJTI = MF->getJumpTableInfo(); if (!MJTI) return false; const std::vector &JT = MJTI->getJumpTables(); for (unsigned i = 0, e = T2JumpTables.size(); i != e; ++i) { MachineInstr *MI = T2JumpTables[i]; const MCInstrDesc &MCID = MI->getDesc(); unsigned NumOps = MCID.getNumOperands(); unsigned JTOpIdx = NumOps - (MI->isPredicable() ? 2 : 1); MachineOperand JTOP = MI->getOperand(JTOpIdx); unsigned JTI = JTOP.getIndex(); assert(JTI < JT.size()); bool ByteOk = true; bool HalfWordOk = true; unsigned JTOffset = BBUtils->getOffsetOf(MI) + 4; const std::vector &JTBBs = JT[JTI].MBBs; BBInfoVector &BBInfo = BBUtils->getBBInfo(); for (MachineBasicBlock *MBB : JTBBs) { unsigned DstOffset = BBInfo[MBB->getNumber()].Offset; // Negative offset is not ok. FIXME: We should change BB layout to make // sure all the branches are forward. if (ByteOk && (DstOffset - JTOffset) > ((1<<8)-1)*2) ByteOk = false; unsigned TBHLimit = ((1<<16)-1)*2; if (HalfWordOk && (DstOffset - JTOffset) > TBHLimit) HalfWordOk = false; if (!ByteOk && !HalfWordOk) break; } if (!ByteOk && !HalfWordOk) continue; CPUser &User = CPUsers[JumpTableUserIndices[JTI]]; MachineBasicBlock *MBB = MI->getParent(); if (!MI->getOperand(0).isKill()) // FIXME: needed now? continue; unsigned DeadSize = 0; bool CanDeleteLEA = false; bool BaseRegKill = false; unsigned IdxReg = ~0U; bool IdxRegKill = true; if (isThumb2) { IdxReg = MI->getOperand(1).getReg(); IdxRegKill = MI->getOperand(1).isKill(); bool PreservedBaseReg = preserveBaseRegister(MI, User.MI, DeadSize, CanDeleteLEA, BaseRegKill); if (!jumpTableFollowsTB(MI, User.CPEMI) && !PreservedBaseReg) continue; } else { // We're in thumb-1 mode, so we must have something like: // %idx = tLSLri %idx, 2 // %base = tLEApcrelJT // %t = tLDRr %base, %idx Register BaseReg = User.MI->getOperand(0).getReg(); MachineBasicBlock *UserMBB = User.MI->getParent(); MachineBasicBlock::iterator Shift = User.MI->getIterator(); if (Shift == UserMBB->begin()) continue; Shift = prev_nodbg(Shift, UserMBB->begin()); if (Shift->getOpcode() != ARM::tLSLri || Shift->getOperand(3).getImm() != 2 || !Shift->getOperand(2).isKill()) continue; IdxReg = Shift->getOperand(2).getReg(); Register ShiftedIdxReg = Shift->getOperand(0).getReg(); // It's important that IdxReg is live until the actual TBB/TBH. Most of // the range is checked later, but the LEA might still clobber it and not // actually get removed. if (BaseReg == IdxReg && !jumpTableFollowsTB(MI, User.CPEMI)) continue; MachineInstr *Load = User.MI->getNextNode(); if (Load->getOpcode() != ARM::tLDRr) continue; if (Load->getOperand(1).getReg() != BaseReg || Load->getOperand(2).getReg() != ShiftedIdxReg || !Load->getOperand(2).isKill()) continue; // If we're in PIC mode, there should be another ADD following. auto *TRI = STI->getRegisterInfo(); // %base cannot be redefined after the load as it will appear before // TBB/TBH like: // %base = // %base = // tBB %base, %idx if (registerDefinedBetween(BaseReg, Load->getNextNode(), MBB->end(), TRI)) continue; if (isPositionIndependentOrROPI) { MachineInstr *Add = Load->getNextNode(); if (Add->getOpcode() != ARM::tADDrr || Add->getOperand(2).getReg() != BaseReg || Add->getOperand(3).getReg() != Load->getOperand(0).getReg() || !Add->getOperand(3).isKill()) continue; if (Add->getOperand(0).getReg() != MI->getOperand(0).getReg()) continue; if (registerDefinedBetween(IdxReg, Add->getNextNode(), MI, TRI)) // IdxReg gets redefined in the middle of the sequence. continue; Add->eraseFromParent(); DeadSize += 2; } else { if (Load->getOperand(0).getReg() != MI->getOperand(0).getReg()) continue; if (registerDefinedBetween(IdxReg, Load->getNextNode(), MI, TRI)) // IdxReg gets redefined in the middle of the sequence. continue; } // Now safe to delete the load and lsl. The LEA will be removed later. CanDeleteLEA = true; Shift->eraseFromParent(); Load->eraseFromParent(); DeadSize += 4; } LLVM_DEBUG(dbgs() << "Shrink JT: " << *MI); MachineInstr *CPEMI = User.CPEMI; unsigned Opc = ByteOk ? ARM::t2TBB_JT : ARM::t2TBH_JT; if (!isThumb2) Opc = ByteOk ? ARM::tTBB_JT : ARM::tTBH_JT; MachineBasicBlock::iterator MI_JT = MI; MachineInstr *NewJTMI = BuildMI(*MBB, MI_JT, MI->getDebugLoc(), TII->get(Opc)) .addReg(User.MI->getOperand(0).getReg(), getKillRegState(BaseRegKill)) .addReg(IdxReg, getKillRegState(IdxRegKill)) .addJumpTableIndex(JTI, JTOP.getTargetFlags()) .addImm(CPEMI->getOperand(0).getImm()); LLVM_DEBUG(dbgs() << printMBBReference(*MBB) << ": " << *NewJTMI); unsigned JTOpc = ByteOk ? ARM::JUMPTABLE_TBB : ARM::JUMPTABLE_TBH; CPEMI->setDesc(TII->get(JTOpc)); if (jumpTableFollowsTB(MI, User.CPEMI)) { NewJTMI->getOperand(0).setReg(ARM::PC); NewJTMI->getOperand(0).setIsKill(false); if (CanDeleteLEA) { if (isThumb2) RemoveDeadAddBetweenLEAAndJT(User.MI, MI, DeadSize); User.MI->eraseFromParent(); DeadSize += isThumb2 ? 4 : 2; // The LEA was eliminated, the TBB instruction becomes the only new user // of the jump table. User.MI = NewJTMI; User.MaxDisp = 4; User.NegOk = false; User.IsSoImm = false; User.KnownAlignment = false; } else { // The LEA couldn't be eliminated, so we must add another CPUser to // record the TBB or TBH use. int CPEntryIdx = JumpTableEntryIndices[JTI]; auto &CPEs = CPEntries[CPEntryIdx]; auto Entry = find_if(CPEs, [&](CPEntry &E) { return E.CPEMI == User.CPEMI; }); ++Entry->RefCount; CPUsers.emplace_back(CPUser(NewJTMI, User.CPEMI, 4, false, false)); } } unsigned NewSize = TII->getInstSizeInBytes(*NewJTMI); unsigned OrigSize = TII->getInstSizeInBytes(*MI); MI->eraseFromParent(); int Delta = OrigSize - NewSize + DeadSize; BBInfo[MBB->getNumber()].Size -= Delta; BBUtils->adjustBBOffsetsAfter(MBB); ++NumTBs; MadeChange = true; } return MadeChange; } /// reorderThumb2JumpTables - Adjust the function's block layout to ensure that /// jump tables always branch forwards, since that's what tbb and tbh need. bool ARMConstantIslands::reorderThumb2JumpTables() { bool MadeChange = false; MachineJumpTableInfo *MJTI = MF->getJumpTableInfo(); if (!MJTI) return false; const std::vector &JT = MJTI->getJumpTables(); for (unsigned i = 0, e = T2JumpTables.size(); i != e; ++i) { MachineInstr *MI = T2JumpTables[i]; const MCInstrDesc &MCID = MI->getDesc(); unsigned NumOps = MCID.getNumOperands(); unsigned JTOpIdx = NumOps - (MI->isPredicable() ? 2 : 1); MachineOperand JTOP = MI->getOperand(JTOpIdx); unsigned JTI = JTOP.getIndex(); assert(JTI < JT.size()); // We prefer if target blocks for the jump table come after the jump // instruction so we can use TB[BH]. Loop through the target blocks // and try to adjust them such that that's true. int JTNumber = MI->getParent()->getNumber(); const std::vector &JTBBs = JT[JTI].MBBs; for (MachineBasicBlock *MBB : JTBBs) { int DTNumber = MBB->getNumber(); if (DTNumber < JTNumber) { // The destination precedes the switch. Try to move the block forward // so we have a positive offset. MachineBasicBlock *NewBB = adjustJTTargetBlockForward(JTI, MBB, MI->getParent()); if (NewBB) MJTI->ReplaceMBBInJumpTable(JTI, MBB, NewBB); MadeChange = true; } } } return MadeChange; } MachineBasicBlock *ARMConstantIslands::adjustJTTargetBlockForward( unsigned JTI, MachineBasicBlock *BB, MachineBasicBlock *JTBB) { // If the destination block is terminated by an unconditional branch, // try to move it; otherwise, create a new block following the jump // table that branches back to the actual target. This is a very simple // heuristic. FIXME: We can definitely improve it. MachineBasicBlock *TBB = nullptr, *FBB = nullptr; SmallVector Cond; SmallVector CondPrior; MachineFunction::iterator BBi = BB->getIterator(); MachineFunction::iterator OldPrior = std::prev(BBi); MachineFunction::iterator OldNext = std::next(BBi); // If the block terminator isn't analyzable, don't try to move the block bool B = TII->analyzeBranch(*BB, TBB, FBB, Cond); // If the block ends in an unconditional branch, move it. The prior block // has to have an analyzable terminator for us to move this one. Be paranoid // and make sure we're not trying to move the entry block of the function. if (!B && Cond.empty() && BB != &MF->front() && !TII->analyzeBranch(*OldPrior, TBB, FBB, CondPrior)) { BB->moveAfter(JTBB); OldPrior->updateTerminator(BB); BB->updateTerminator(OldNext != MF->end() ? &*OldNext : nullptr); // Update numbering to account for the block being moved. MF->RenumberBlocks(); ++NumJTMoved; return nullptr; } // Create a new MBB for the code after the jump BB. MachineBasicBlock *NewBB = MF->CreateMachineBasicBlock(JTBB->getBasicBlock()); MachineFunction::iterator MBBI = ++JTBB->getIterator(); MF->insert(MBBI, NewBB); // Copy live-in information to new block. for (const MachineBasicBlock::RegisterMaskPair &RegMaskPair : BB->liveins()) NewBB->addLiveIn(RegMaskPair); // Add an unconditional branch from NewBB to BB. // There doesn't seem to be meaningful DebugInfo available; this doesn't // correspond directly to anything in the source. if (isThumb2) BuildMI(NewBB, DebugLoc(), TII->get(ARM::t2B)) .addMBB(BB) .add(predOps(ARMCC::AL)); else BuildMI(NewBB, DebugLoc(), TII->get(ARM::tB)) .addMBB(BB) .add(predOps(ARMCC::AL)); // Update internal data structures to account for the newly inserted MBB. MF->RenumberBlocks(NewBB); // Update the CFG. NewBB->addSuccessor(BB); JTBB->replaceSuccessor(BB, NewBB); ++NumJTInserted; return NewBB; } /// createARMConstantIslandPass - returns an instance of the constpool /// island pass. FunctionPass *llvm::createARMConstantIslandPass() { return new ARMConstantIslands(); } INITIALIZE_PASS(ARMConstantIslands, "arm-cp-islands", ARM_CP_ISLANDS_OPT_NAME, false, false)