//===- InlineSpiller.cpp - Insert spills and restores inline --------------===// // // 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 // //===----------------------------------------------------------------------===// // // The inline spiller modifies the machine function directly instead of // inserting spills and restores in VirtRegMap. // //===----------------------------------------------------------------------===// #include "SplitKit.h" #include "llvm/ADT/ArrayRef.h" #include "llvm/ADT/DenseMap.h" #include "llvm/ADT/MapVector.h" #include "llvm/ADT/STLExtras.h" #include "llvm/ADT/SetVector.h" #include "llvm/ADT/SmallPtrSet.h" #include "llvm/ADT/SmallVector.h" #include "llvm/ADT/Statistic.h" #include "llvm/Analysis/AliasAnalysis.h" #include "llvm/CodeGen/LiveInterval.h" #include "llvm/CodeGen/LiveIntervals.h" #include "llvm/CodeGen/LiveRangeEdit.h" #include "llvm/CodeGen/LiveStacks.h" #include "llvm/CodeGen/MachineBasicBlock.h" #include "llvm/CodeGen/MachineBlockFrequencyInfo.h" #include "llvm/CodeGen/MachineDominators.h" #include "llvm/CodeGen/MachineFunction.h" #include "llvm/CodeGen/MachineFunctionPass.h" #include "llvm/CodeGen/MachineInstr.h" #include "llvm/CodeGen/MachineInstrBuilder.h" #include "llvm/CodeGen/MachineInstrBundle.h" #include "llvm/CodeGen/MachineLoopInfo.h" #include "llvm/CodeGen/MachineOperand.h" #include "llvm/CodeGen/MachineRegisterInfo.h" #include "llvm/CodeGen/SlotIndexes.h" #include "llvm/CodeGen/Spiller.h" #include "llvm/CodeGen/StackMaps.h" #include "llvm/CodeGen/TargetInstrInfo.h" #include "llvm/CodeGen/TargetOpcodes.h" #include "llvm/CodeGen/TargetRegisterInfo.h" #include "llvm/CodeGen/TargetSubtargetInfo.h" #include "llvm/CodeGen/VirtRegMap.h" #include "llvm/Config/llvm-config.h" #include "llvm/Support/BlockFrequency.h" #include "llvm/Support/BranchProbability.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Compiler.h" #include "llvm/Support/Debug.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/raw_ostream.h" #include #include #include #include #include using namespace llvm; #define DEBUG_TYPE "regalloc" STATISTIC(NumSpilledRanges, "Number of spilled live ranges"); STATISTIC(NumSnippets, "Number of spilled snippets"); STATISTIC(NumSpills, "Number of spills inserted"); STATISTIC(NumSpillsRemoved, "Number of spills removed"); STATISTIC(NumReloads, "Number of reloads inserted"); STATISTIC(NumReloadsRemoved, "Number of reloads removed"); STATISTIC(NumFolded, "Number of folded stack accesses"); STATISTIC(NumFoldedLoads, "Number of folded loads"); STATISTIC(NumRemats, "Number of rematerialized defs for spilling"); static cl::opt DisableHoisting("disable-spill-hoist", cl::Hidden, cl::desc("Disable inline spill hoisting")); static cl::opt RestrictStatepointRemat("restrict-statepoint-remat", cl::init(false), cl::Hidden, cl::desc("Restrict remat for statepoint operands")); namespace { class HoistSpillHelper : private LiveRangeEdit::Delegate { MachineFunction &MF; LiveIntervals &LIS; LiveStacks &LSS; MachineDominatorTree &MDT; MachineLoopInfo &Loops; VirtRegMap &VRM; MachineRegisterInfo &MRI; const TargetInstrInfo &TII; const TargetRegisterInfo &TRI; const MachineBlockFrequencyInfo &MBFI; InsertPointAnalysis IPA; // Map from StackSlot to the LiveInterval of the original register. // Note the LiveInterval of the original register may have been deleted // after it is spilled. We keep a copy here to track the range where // spills can be moved. DenseMap> StackSlotToOrigLI; // Map from pair of (StackSlot and Original VNI) to a set of spills which // have the same stackslot and have equal values defined by Original VNI. // These spills are mergeable and are hoist candidates. using MergeableSpillsMap = MapVector, SmallPtrSet>; MergeableSpillsMap MergeableSpills; /// This is the map from original register to a set containing all its /// siblings. To hoist a spill to another BB, we need to find out a live /// sibling there and use it as the source of the new spill. DenseMap> Virt2SiblingsMap; bool isSpillCandBB(LiveInterval &OrigLI, VNInfo &OrigVNI, MachineBasicBlock &BB, Register &LiveReg); void rmRedundantSpills( SmallPtrSet &Spills, SmallVectorImpl &SpillsToRm, DenseMap &SpillBBToSpill); void getVisitOrders( MachineBasicBlock *Root, SmallPtrSet &Spills, SmallVectorImpl &Orders, SmallVectorImpl &SpillsToRm, DenseMap &SpillsToKeep, DenseMap &SpillBBToSpill); void runHoistSpills(LiveInterval &OrigLI, VNInfo &OrigVNI, SmallPtrSet &Spills, SmallVectorImpl &SpillsToRm, DenseMap &SpillsToIns); public: HoistSpillHelper(MachineFunctionPass &pass, MachineFunction &mf, VirtRegMap &vrm) : MF(mf), LIS(pass.getAnalysis()), LSS(pass.getAnalysis()), MDT(pass.getAnalysis()), Loops(pass.getAnalysis()), VRM(vrm), MRI(mf.getRegInfo()), TII(*mf.getSubtarget().getInstrInfo()), TRI(*mf.getSubtarget().getRegisterInfo()), MBFI(pass.getAnalysis()), IPA(LIS, mf.getNumBlockIDs()) {} void addToMergeableSpills(MachineInstr &Spill, int StackSlot, unsigned Original); bool rmFromMergeableSpills(MachineInstr &Spill, int StackSlot); void hoistAllSpills(); void LRE_DidCloneVirtReg(Register, Register) override; }; class InlineSpiller : public Spiller { MachineFunction &MF; LiveIntervals &LIS; LiveStacks &LSS; MachineDominatorTree &MDT; MachineLoopInfo &Loops; VirtRegMap &VRM; MachineRegisterInfo &MRI; const TargetInstrInfo &TII; const TargetRegisterInfo &TRI; const MachineBlockFrequencyInfo &MBFI; // Variables that are valid during spill(), but used by multiple methods. LiveRangeEdit *Edit = nullptr; LiveInterval *StackInt = nullptr; int StackSlot; Register Original; // All registers to spill to StackSlot, including the main register. SmallVector RegsToSpill; // All COPY instructions to/from snippets. // They are ignored since both operands refer to the same stack slot. // For bundled copies, this will only include the first header copy. SmallPtrSet SnippetCopies; // Values that failed to remat at some point. SmallPtrSet UsedValues; // Dead defs generated during spilling. SmallVector DeadDefs; // Object records spills information and does the hoisting. HoistSpillHelper HSpiller; // Live range weight calculator. VirtRegAuxInfo &VRAI; ~InlineSpiller() override = default; public: InlineSpiller(MachineFunctionPass &Pass, MachineFunction &MF, VirtRegMap &VRM, VirtRegAuxInfo &VRAI) : MF(MF), LIS(Pass.getAnalysis()), LSS(Pass.getAnalysis()), MDT(Pass.getAnalysis()), Loops(Pass.getAnalysis()), VRM(VRM), MRI(MF.getRegInfo()), TII(*MF.getSubtarget().getInstrInfo()), TRI(*MF.getSubtarget().getRegisterInfo()), MBFI(Pass.getAnalysis()), HSpiller(Pass, MF, VRM), VRAI(VRAI) {} void spill(LiveRangeEdit &) override; void postOptimization() override; private: bool isSnippet(const LiveInterval &SnipLI); void collectRegsToSpill(); bool isRegToSpill(Register Reg) { return is_contained(RegsToSpill, Reg); } bool isSibling(Register Reg); bool hoistSpillInsideBB(LiveInterval &SpillLI, MachineInstr &CopyMI); void eliminateRedundantSpills(LiveInterval &LI, VNInfo *VNI); void markValueUsed(LiveInterval*, VNInfo*); bool canGuaranteeAssignmentAfterRemat(Register VReg, MachineInstr &MI); bool reMaterializeFor(LiveInterval &, MachineInstr &MI); void reMaterializeAll(); bool coalesceStackAccess(MachineInstr *MI, Register Reg); bool foldMemoryOperand(ArrayRef>, MachineInstr *LoadMI = nullptr); void insertReload(Register VReg, SlotIndex, MachineBasicBlock::iterator MI); void insertSpill(Register VReg, bool isKill, MachineBasicBlock::iterator MI); void spillAroundUses(Register Reg); void spillAll(); }; } // end anonymous namespace Spiller::~Spiller() = default; void Spiller::anchor() {} Spiller *llvm::createInlineSpiller(MachineFunctionPass &Pass, MachineFunction &MF, VirtRegMap &VRM, VirtRegAuxInfo &VRAI) { return new InlineSpiller(Pass, MF, VRM, VRAI); } //===----------------------------------------------------------------------===// // Snippets //===----------------------------------------------------------------------===// // When spilling a virtual register, we also spill any snippets it is connected // to. The snippets are small live ranges that only have a single real use, // leftovers from live range splitting. Spilling them enables memory operand // folding or tightens the live range around the single use. // // This minimizes register pressure and maximizes the store-to-load distance for // spill slots which can be important in tight loops. /// If MI is a COPY to or from Reg, return the other register, otherwise return /// 0. static Register isCopyOf(const MachineInstr &MI, Register Reg) { assert(!MI.isBundled()); if (!MI.isCopy()) return Register(); const MachineOperand &DstOp = MI.getOperand(0); const MachineOperand &SrcOp = MI.getOperand(1); // TODO: Probably only worth allowing subreg copies with undef dests. if (DstOp.getSubReg() != SrcOp.getSubReg()) return Register(); if (DstOp.getReg() == Reg) return SrcOp.getReg(); if (SrcOp.getReg() == Reg) return DstOp.getReg(); return Register(); } /// Check for a copy bundle as formed by SplitKit. static Register isCopyOfBundle(const MachineInstr &FirstMI, Register Reg) { if (!FirstMI.isBundled()) return isCopyOf(FirstMI, Reg); assert(!FirstMI.isBundledWithPred() && FirstMI.isBundledWithSucc() && "expected to see first instruction in bundle"); Register SnipReg; MachineBasicBlock::const_instr_iterator I = FirstMI.getIterator(); while (I->isBundledWithSucc()) { const MachineInstr &MI = *I; if (!MI.isCopy()) return Register(); const MachineOperand &DstOp = MI.getOperand(0); const MachineOperand &SrcOp = MI.getOperand(1); if (DstOp.getReg() == Reg) { if (!SnipReg) SnipReg = SrcOp.getReg(); else if (SnipReg != SrcOp.getReg()) return Register(); } else if (SrcOp.getReg() == Reg) { if (!SnipReg) SnipReg = DstOp.getReg(); else if (SnipReg != DstOp.getReg()) return Register(); } ++I; } return Register(); } static void getVDefInterval(const MachineInstr &MI, LiveIntervals &LIS) { for (const MachineOperand &MO : MI.all_defs()) if (MO.getReg().isVirtual()) LIS.getInterval(MO.getReg()); } /// isSnippet - Identify if a live interval is a snippet that should be spilled. /// It is assumed that SnipLI is a virtual register with the same original as /// Edit->getReg(). bool InlineSpiller::isSnippet(const LiveInterval &SnipLI) { Register Reg = Edit->getReg(); // A snippet is a tiny live range with only a single instruction using it // besides copies to/from Reg or spills/fills. // Exception is done for statepoint instructions which will fold fills // into their operands. // We accept: // // %snip = COPY %Reg / FILL fi# // %snip = USE %snip // %snip = STATEPOINT %snip in var arg area // %Reg = COPY %snip / SPILL %snip, fi# // if (!LIS.intervalIsInOneMBB(SnipLI)) return false; // Number of defs should not exceed 2 not accounting defs coming from // statepoint instructions. unsigned NumValNums = SnipLI.getNumValNums(); for (auto *VNI : SnipLI.vnis()) { MachineInstr *MI = LIS.getInstructionFromIndex(VNI->def); if (MI->getOpcode() == TargetOpcode::STATEPOINT) --NumValNums; } if (NumValNums > 2) return false; MachineInstr *UseMI = nullptr; // Check that all uses satisfy our criteria. for (MachineRegisterInfo::reg_bundle_nodbg_iterator RI = MRI.reg_bundle_nodbg_begin(SnipLI.reg()), E = MRI.reg_bundle_nodbg_end(); RI != E;) { MachineInstr &MI = *RI++; // Allow copies to/from Reg. if (isCopyOfBundle(MI, Reg)) continue; // Allow stack slot loads. int FI; if (SnipLI.reg() == TII.isLoadFromStackSlot(MI, FI) && FI == StackSlot) continue; // Allow stack slot stores. if (SnipLI.reg() == TII.isStoreToStackSlot(MI, FI) && FI == StackSlot) continue; if (StatepointOpers::isFoldableReg(&MI, SnipLI.reg())) continue; // Allow a single additional instruction. if (UseMI && &MI != UseMI) return false; UseMI = &MI; } return true; } /// collectRegsToSpill - Collect live range snippets that only have a single /// real use. void InlineSpiller::collectRegsToSpill() { Register Reg = Edit->getReg(); // Main register always spills. RegsToSpill.assign(1, Reg); SnippetCopies.clear(); // Snippets all have the same original, so there can't be any for an original // register. if (Original == Reg) return; for (MachineInstr &MI : llvm::make_early_inc_range(MRI.reg_bundles(Reg))) { Register SnipReg = isCopyOfBundle(MI, Reg); if (!isSibling(SnipReg)) continue; LiveInterval &SnipLI = LIS.getInterval(SnipReg); if (!isSnippet(SnipLI)) continue; SnippetCopies.insert(&MI); if (isRegToSpill(SnipReg)) continue; RegsToSpill.push_back(SnipReg); LLVM_DEBUG(dbgs() << "\talso spill snippet " << SnipLI << '\n'); ++NumSnippets; } } bool InlineSpiller::isSibling(Register Reg) { return Reg.isVirtual() && VRM.getOriginal(Reg) == Original; } /// It is beneficial to spill to earlier place in the same BB in case /// as follows: /// There is an alternative def earlier in the same MBB. /// Hoist the spill as far as possible in SpillMBB. This can ease /// register pressure: /// /// x = def /// y = use x /// s = copy x /// /// Hoisting the spill of s to immediately after the def removes the /// interference between x and y: /// /// x = def /// spill x /// y = use killed x /// /// This hoist only helps when the copy kills its source. /// bool InlineSpiller::hoistSpillInsideBB(LiveInterval &SpillLI, MachineInstr &CopyMI) { SlotIndex Idx = LIS.getInstructionIndex(CopyMI); #ifndef NDEBUG VNInfo *VNI = SpillLI.getVNInfoAt(Idx.getRegSlot()); assert(VNI && VNI->def == Idx.getRegSlot() && "Not defined by copy"); #endif Register SrcReg = CopyMI.getOperand(1).getReg(); LiveInterval &SrcLI = LIS.getInterval(SrcReg); VNInfo *SrcVNI = SrcLI.getVNInfoAt(Idx); LiveQueryResult SrcQ = SrcLI.Query(Idx); MachineBasicBlock *DefMBB = LIS.getMBBFromIndex(SrcVNI->def); if (DefMBB != CopyMI.getParent() || !SrcQ.isKill()) return false; // Conservatively extend the stack slot range to the range of the original // value. We may be able to do better with stack slot coloring by being more // careful here. assert(StackInt && "No stack slot assigned yet."); LiveInterval &OrigLI = LIS.getInterval(Original); VNInfo *OrigVNI = OrigLI.getVNInfoAt(Idx); StackInt->MergeValueInAsValue(OrigLI, OrigVNI, StackInt->getValNumInfo(0)); LLVM_DEBUG(dbgs() << "\tmerged orig valno " << OrigVNI->id << ": " << *StackInt << '\n'); // We are going to spill SrcVNI immediately after its def, so clear out // any later spills of the same value. eliminateRedundantSpills(SrcLI, SrcVNI); MachineBasicBlock *MBB = LIS.getMBBFromIndex(SrcVNI->def); MachineBasicBlock::iterator MII; if (SrcVNI->isPHIDef()) MII = MBB->SkipPHIsLabelsAndDebug(MBB->begin()); else { MachineInstr *DefMI = LIS.getInstructionFromIndex(SrcVNI->def); assert(DefMI && "Defining instruction disappeared"); MII = DefMI; ++MII; } MachineInstrSpan MIS(MII, MBB); // Insert spill without kill flag immediately after def. TII.storeRegToStackSlot(*MBB, MII, SrcReg, false, StackSlot, MRI.getRegClass(SrcReg), &TRI, Register()); LIS.InsertMachineInstrRangeInMaps(MIS.begin(), MII); for (const MachineInstr &MI : make_range(MIS.begin(), MII)) getVDefInterval(MI, LIS); --MII; // Point to store instruction. LLVM_DEBUG(dbgs() << "\thoisted: " << SrcVNI->def << '\t' << *MII); // If there is only 1 store instruction is required for spill, add it // to mergeable list. In X86 AMX, 2 intructions are required to store. // We disable the merge for this case. if (MIS.begin() == MII) HSpiller.addToMergeableSpills(*MII, StackSlot, Original); ++NumSpills; return true; } /// eliminateRedundantSpills - SLI:VNI is known to be on the stack. Remove any /// redundant spills of this value in SLI.reg and sibling copies. void InlineSpiller::eliminateRedundantSpills(LiveInterval &SLI, VNInfo *VNI) { assert(VNI && "Missing value"); SmallVector, 8> WorkList; WorkList.push_back(std::make_pair(&SLI, VNI)); assert(StackInt && "No stack slot assigned yet."); do { LiveInterval *LI; std::tie(LI, VNI) = WorkList.pop_back_val(); Register Reg = LI->reg(); LLVM_DEBUG(dbgs() << "Checking redundant spills for " << VNI->id << '@' << VNI->def << " in " << *LI << '\n'); // Regs to spill are taken care of. if (isRegToSpill(Reg)) continue; // Add all of VNI's live range to StackInt. StackInt->MergeValueInAsValue(*LI, VNI, StackInt->getValNumInfo(0)); LLVM_DEBUG(dbgs() << "Merged to stack int: " << *StackInt << '\n'); // Find all spills and copies of VNI. for (MachineInstr &MI : llvm::make_early_inc_range(MRI.use_nodbg_bundles(Reg))) { if (!MI.isCopy() && !MI.mayStore()) continue; SlotIndex Idx = LIS.getInstructionIndex(MI); if (LI->getVNInfoAt(Idx) != VNI) continue; // Follow sibling copies down the dominator tree. if (Register DstReg = isCopyOfBundle(MI, Reg)) { if (isSibling(DstReg)) { LiveInterval &DstLI = LIS.getInterval(DstReg); VNInfo *DstVNI = DstLI.getVNInfoAt(Idx.getRegSlot()); assert(DstVNI && "Missing defined value"); assert(DstVNI->def == Idx.getRegSlot() && "Wrong copy def slot"); WorkList.push_back(std::make_pair(&DstLI, DstVNI)); } continue; } // Erase spills. int FI; if (Reg == TII.isStoreToStackSlot(MI, FI) && FI == StackSlot) { LLVM_DEBUG(dbgs() << "Redundant spill " << Idx << '\t' << MI); // eliminateDeadDefs won't normally remove stores, so switch opcode. MI.setDesc(TII.get(TargetOpcode::KILL)); DeadDefs.push_back(&MI); ++NumSpillsRemoved; if (HSpiller.rmFromMergeableSpills(MI, StackSlot)) --NumSpills; } } } while (!WorkList.empty()); } //===----------------------------------------------------------------------===// // Rematerialization //===----------------------------------------------------------------------===// /// markValueUsed - Remember that VNI failed to rematerialize, so its defining /// instruction cannot be eliminated. See through snippet copies void InlineSpiller::markValueUsed(LiveInterval *LI, VNInfo *VNI) { SmallVector, 8> WorkList; WorkList.push_back(std::make_pair(LI, VNI)); do { std::tie(LI, VNI) = WorkList.pop_back_val(); if (!UsedValues.insert(VNI).second) continue; if (VNI->isPHIDef()) { MachineBasicBlock *MBB = LIS.getMBBFromIndex(VNI->def); for (MachineBasicBlock *P : MBB->predecessors()) { VNInfo *PVNI = LI->getVNInfoBefore(LIS.getMBBEndIdx(P)); if (PVNI) WorkList.push_back(std::make_pair(LI, PVNI)); } continue; } // Follow snippet copies. MachineInstr *MI = LIS.getInstructionFromIndex(VNI->def); if (!SnippetCopies.count(MI)) continue; LiveInterval &SnipLI = LIS.getInterval(MI->getOperand(1).getReg()); assert(isRegToSpill(SnipLI.reg()) && "Unexpected register in copy"); VNInfo *SnipVNI = SnipLI.getVNInfoAt(VNI->def.getRegSlot(true)); assert(SnipVNI && "Snippet undefined before copy"); WorkList.push_back(std::make_pair(&SnipLI, SnipVNI)); } while (!WorkList.empty()); } bool InlineSpiller::canGuaranteeAssignmentAfterRemat(Register VReg, MachineInstr &MI) { if (!RestrictStatepointRemat) return true; // Here's a quick explanation of the problem we're trying to handle here: // * There are some pseudo instructions with more vreg uses than there are // physical registers on the machine. // * This is normally handled by spilling the vreg, and folding the reload // into the user instruction. (Thus decreasing the number of used vregs // until the remainder can be assigned to physregs.) // * However, since we may try to spill vregs in any order, we can end up // trying to spill each operand to the instruction, and then rematting it // instead. When that happens, the new live intervals (for the remats) are // expected to be trivially assignable (i.e. RS_Done). However, since we // may have more remats than physregs, we're guaranteed to fail to assign // one. // At the moment, we only handle this for STATEPOINTs since they're the only // pseudo op where we've seen this. If we start seeing other instructions // with the same problem, we need to revisit this. if (MI.getOpcode() != TargetOpcode::STATEPOINT) return true; // For STATEPOINTs we allow re-materialization for fixed arguments only hoping // that number of physical registers is enough to cover all fixed arguments. // If it is not true we need to revisit it. for (unsigned Idx = StatepointOpers(&MI).getVarIdx(), EndIdx = MI.getNumOperands(); Idx < EndIdx; ++Idx) { MachineOperand &MO = MI.getOperand(Idx); if (MO.isReg() && MO.getReg() == VReg) return false; } return true; } /// reMaterializeFor - Attempt to rematerialize before MI instead of reloading. bool InlineSpiller::reMaterializeFor(LiveInterval &VirtReg, MachineInstr &MI) { // Analyze instruction SmallVector, 8> Ops; VirtRegInfo RI = AnalyzeVirtRegInBundle(MI, VirtReg.reg(), &Ops); if (!RI.Reads) return false; SlotIndex UseIdx = LIS.getInstructionIndex(MI).getRegSlot(true); VNInfo *ParentVNI = VirtReg.getVNInfoAt(UseIdx.getBaseIndex()); if (!ParentVNI) { LLVM_DEBUG(dbgs() << "\tadding flags: "); for (MachineOperand &MO : MI.all_uses()) if (MO.getReg() == VirtReg.reg()) MO.setIsUndef(); LLVM_DEBUG(dbgs() << UseIdx << '\t' << MI); return true; } if (SnippetCopies.count(&MI)) return false; LiveInterval &OrigLI = LIS.getInterval(Original); VNInfo *OrigVNI = OrigLI.getVNInfoAt(UseIdx); LiveRangeEdit::Remat RM(ParentVNI); RM.OrigMI = LIS.getInstructionFromIndex(OrigVNI->def); if (!Edit->canRematerializeAt(RM, OrigVNI, UseIdx, false)) { markValueUsed(&VirtReg, ParentVNI); LLVM_DEBUG(dbgs() << "\tcannot remat for " << UseIdx << '\t' << MI); return false; } // If the instruction also writes VirtReg.reg, it had better not require the // same register for uses and defs. if (RI.Tied) { markValueUsed(&VirtReg, ParentVNI); LLVM_DEBUG(dbgs() << "\tcannot remat tied reg: " << UseIdx << '\t' << MI); return false; } // Before rematerializing into a register for a single instruction, try to // fold a load into the instruction. That avoids allocating a new register. if (RM.OrigMI->canFoldAsLoad() && foldMemoryOperand(Ops, RM.OrigMI)) { Edit->markRematerialized(RM.ParentVNI); ++NumFoldedLoads; return true; } // If we can't guarantee that we'll be able to actually assign the new vreg, // we can't remat. if (!canGuaranteeAssignmentAfterRemat(VirtReg.reg(), MI)) { markValueUsed(&VirtReg, ParentVNI); LLVM_DEBUG(dbgs() << "\tcannot remat for " << UseIdx << '\t' << MI); return false; } // Allocate a new register for the remat. Register NewVReg = Edit->createFrom(Original); // Finally we can rematerialize OrigMI before MI. SlotIndex DefIdx = Edit->rematerializeAt(*MI.getParent(), MI, NewVReg, RM, TRI); // We take the DebugLoc from MI, since OrigMI may be attributed to a // different source location. auto *NewMI = LIS.getInstructionFromIndex(DefIdx); NewMI->setDebugLoc(MI.getDebugLoc()); (void)DefIdx; LLVM_DEBUG(dbgs() << "\tremat: " << DefIdx << '\t' << *LIS.getInstructionFromIndex(DefIdx)); // Replace operands for (const auto &OpPair : Ops) { MachineOperand &MO = OpPair.first->getOperand(OpPair.second); if (MO.isReg() && MO.isUse() && MO.getReg() == VirtReg.reg()) { MO.setReg(NewVReg); MO.setIsKill(); } } LLVM_DEBUG(dbgs() << "\t " << UseIdx << '\t' << MI << '\n'); ++NumRemats; return true; } /// reMaterializeAll - Try to rematerialize as many uses as possible, /// and trim the live ranges after. void InlineSpiller::reMaterializeAll() { if (!Edit->anyRematerializable()) return; UsedValues.clear(); // Try to remat before all uses of snippets. bool anyRemat = false; for (Register Reg : RegsToSpill) { LiveInterval &LI = LIS.getInterval(Reg); for (MachineInstr &MI : llvm::make_early_inc_range(MRI.reg_bundles(Reg))) { // Debug values are not allowed to affect codegen. if (MI.isDebugValue()) continue; assert(!MI.isDebugInstr() && "Did not expect to find a use in debug " "instruction that isn't a DBG_VALUE"); anyRemat |= reMaterializeFor(LI, MI); } } if (!anyRemat) return; // Remove any values that were completely rematted. for (Register Reg : RegsToSpill) { LiveInterval &LI = LIS.getInterval(Reg); for (VNInfo *VNI : LI.vnis()) { if (VNI->isUnused() || VNI->isPHIDef() || UsedValues.count(VNI)) continue; MachineInstr *MI = LIS.getInstructionFromIndex(VNI->def); MI->addRegisterDead(Reg, &TRI); if (!MI->allDefsAreDead()) continue; LLVM_DEBUG(dbgs() << "All defs dead: " << *MI); DeadDefs.push_back(MI); } } // Eliminate dead code after remat. Note that some snippet copies may be // deleted here. if (DeadDefs.empty()) return; LLVM_DEBUG(dbgs() << "Remat created " << DeadDefs.size() << " dead defs.\n"); Edit->eliminateDeadDefs(DeadDefs, RegsToSpill); // LiveRangeEdit::eliminateDeadDef is used to remove dead define instructions // after rematerialization. To remove a VNI for a vreg from its LiveInterval, // LiveIntervals::removeVRegDefAt is used. However, after non-PHI VNIs are all // removed, PHI VNI are still left in the LiveInterval. // So to get rid of unused reg, we need to check whether it has non-dbg // reference instead of whether it has non-empty interval. unsigned ResultPos = 0; for (Register Reg : RegsToSpill) { if (MRI.reg_nodbg_empty(Reg)) { Edit->eraseVirtReg(Reg); continue; } assert(LIS.hasInterval(Reg) && (!LIS.getInterval(Reg).empty() || !MRI.reg_nodbg_empty(Reg)) && "Empty and not used live-range?!"); RegsToSpill[ResultPos++] = Reg; } RegsToSpill.erase(RegsToSpill.begin() + ResultPos, RegsToSpill.end()); LLVM_DEBUG(dbgs() << RegsToSpill.size() << " registers to spill after remat.\n"); } //===----------------------------------------------------------------------===// // Spilling //===----------------------------------------------------------------------===// /// If MI is a load or store of StackSlot, it can be removed. bool InlineSpiller::coalesceStackAccess(MachineInstr *MI, Register Reg) { int FI = 0; Register InstrReg = TII.isLoadFromStackSlot(*MI, FI); bool IsLoad = InstrReg; if (!IsLoad) InstrReg = TII.isStoreToStackSlot(*MI, FI); // We have a stack access. Is it the right register and slot? if (InstrReg != Reg || FI != StackSlot) return false; if (!IsLoad) HSpiller.rmFromMergeableSpills(*MI, StackSlot); LLVM_DEBUG(dbgs() << "Coalescing stack access: " << *MI); LIS.RemoveMachineInstrFromMaps(*MI); MI->eraseFromParent(); if (IsLoad) { ++NumReloadsRemoved; --NumReloads; } else { ++NumSpillsRemoved; --NumSpills; } return true; } #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) LLVM_DUMP_METHOD // Dump the range of instructions from B to E with their slot indexes. static void dumpMachineInstrRangeWithSlotIndex(MachineBasicBlock::iterator B, MachineBasicBlock::iterator E, LiveIntervals const &LIS, const char *const header, Register VReg = Register()) { char NextLine = '\n'; char SlotIndent = '\t'; if (std::next(B) == E) { NextLine = ' '; SlotIndent = ' '; } dbgs() << '\t' << header << ": " << NextLine; for (MachineBasicBlock::iterator I = B; I != E; ++I) { SlotIndex Idx = LIS.getInstructionIndex(*I).getRegSlot(); // If a register was passed in and this instruction has it as a // destination that is marked as an early clobber, print the // early-clobber slot index. if (VReg) { MachineOperand *MO = I->findRegisterDefOperand(VReg); if (MO && MO->isEarlyClobber()) Idx = Idx.getRegSlot(true); } dbgs() << SlotIndent << Idx << '\t' << *I; } } #endif /// foldMemoryOperand - Try folding stack slot references in Ops into their /// instructions. /// /// @param Ops Operand indices from AnalyzeVirtRegInBundle(). /// @param LoadMI Load instruction to use instead of stack slot when non-null. /// @return True on success. bool InlineSpiller:: foldMemoryOperand(ArrayRef> Ops, MachineInstr *LoadMI) { if (Ops.empty()) return false; // Don't attempt folding in bundles. MachineInstr *MI = Ops.front().first; if (Ops.back().first != MI || MI->isBundled()) return false; bool WasCopy = MI->isCopy(); Register ImpReg; // TII::foldMemoryOperand will do what we need here for statepoint // (fold load into use and remove corresponding def). We will replace // uses of removed def with loads (spillAroundUses). // For that to work we need to untie def and use to pass it through // foldMemoryOperand and signal foldPatchpoint that it is allowed to // fold them. bool UntieRegs = MI->getOpcode() == TargetOpcode::STATEPOINT; // Spill subregs if the target allows it. // We always want to spill subregs for stackmap/patchpoint pseudos. bool SpillSubRegs = TII.isSubregFoldable() || MI->getOpcode() == TargetOpcode::STATEPOINT || MI->getOpcode() == TargetOpcode::PATCHPOINT || MI->getOpcode() == TargetOpcode::STACKMAP; // TargetInstrInfo::foldMemoryOperand only expects explicit, non-tied // operands. SmallVector FoldOps; for (const auto &OpPair : Ops) { unsigned Idx = OpPair.second; assert(MI == OpPair.first && "Instruction conflict during operand folding"); MachineOperand &MO = MI->getOperand(Idx); // No point restoring an undef read, and we'll produce an invalid live // interval. // TODO: Is this really the correct way to handle undef tied uses? if (MO.isUse() && !MO.readsReg() && !MO.isTied()) continue; if (MO.isImplicit()) { ImpReg = MO.getReg(); continue; } if (!SpillSubRegs && MO.getSubReg()) return false; // We cannot fold a load instruction into a def. if (LoadMI && MO.isDef()) return false; // Tied use operands should not be passed to foldMemoryOperand. if (UntieRegs || !MI->isRegTiedToDefOperand(Idx)) FoldOps.push_back(Idx); } // If we only have implicit uses, we won't be able to fold that. // Moreover, TargetInstrInfo::foldMemoryOperand will assert if we try! if (FoldOps.empty()) return false; MachineInstrSpan MIS(MI, MI->getParent()); SmallVector > TiedOps; if (UntieRegs) for (unsigned Idx : FoldOps) { MachineOperand &MO = MI->getOperand(Idx); if (!MO.isTied()) continue; unsigned Tied = MI->findTiedOperandIdx(Idx); if (MO.isUse()) TiedOps.emplace_back(Tied, Idx); else { assert(MO.isDef() && "Tied to not use and def?"); TiedOps.emplace_back(Idx, Tied); } MI->untieRegOperand(Idx); } MachineInstr *FoldMI = LoadMI ? TII.foldMemoryOperand(*MI, FoldOps, *LoadMI, &LIS) : TII.foldMemoryOperand(*MI, FoldOps, StackSlot, &LIS, &VRM); if (!FoldMI) { // Re-tie operands. for (auto Tied : TiedOps) MI->tieOperands(Tied.first, Tied.second); return false; } // Remove LIS for any dead defs in the original MI not in FoldMI. for (MIBundleOperands MO(*MI); MO.isValid(); ++MO) { if (!MO->isReg()) continue; Register Reg = MO->getReg(); if (!Reg || Reg.isVirtual() || MRI.isReserved(Reg)) { continue; } // Skip non-Defs, including undef uses and internal reads. if (MO->isUse()) continue; PhysRegInfo RI = AnalyzePhysRegInBundle(*FoldMI, Reg, &TRI); if (RI.FullyDefined) continue; // FoldMI does not define this physreg. Remove the LI segment. assert(MO->isDead() && "Cannot fold physreg def"); SlotIndex Idx = LIS.getInstructionIndex(*MI).getRegSlot(); LIS.removePhysRegDefAt(Reg.asMCReg(), Idx); } int FI; if (TII.isStoreToStackSlot(*MI, FI) && HSpiller.rmFromMergeableSpills(*MI, FI)) --NumSpills; LIS.ReplaceMachineInstrInMaps(*MI, *FoldMI); // Update the call site info. if (MI->isCandidateForCallSiteEntry()) MI->getMF()->moveCallSiteInfo(MI, FoldMI); // If we've folded a store into an instruction labelled with debug-info, // record a substitution from the old operand to the memory operand. Handle // the simple common case where operand 0 is the one being folded, plus when // the destination operand is also a tied def. More values could be // substituted / preserved with more analysis. if (MI->peekDebugInstrNum() && Ops[0].second == 0) { // Helper lambda. auto MakeSubstitution = [this,FoldMI,MI,&Ops]() { // Substitute old operand zero to the new instructions memory operand. unsigned OldOperandNum = Ops[0].second; unsigned NewNum = FoldMI->getDebugInstrNum(); unsigned OldNum = MI->getDebugInstrNum(); MF.makeDebugValueSubstitution({OldNum, OldOperandNum}, {NewNum, MachineFunction::DebugOperandMemNumber}); }; const MachineOperand &Op0 = MI->getOperand(Ops[0].second); if (Ops.size() == 1 && Op0.isDef()) { MakeSubstitution(); } else if (Ops.size() == 2 && Op0.isDef() && MI->getOperand(1).isTied() && Op0.getReg() == MI->getOperand(1).getReg()) { MakeSubstitution(); } } else if (MI->peekDebugInstrNum()) { // This is a debug-labelled instruction, but the operand being folded isn't // at operand zero. Most likely this means it's a load being folded in. // Substitute any register defs from operand zero up to the one being // folded -- past that point, we don't know what the new operand indexes // will be. MF.substituteDebugValuesForInst(*MI, *FoldMI, Ops[0].second); } MI->eraseFromParent(); // Insert any new instructions other than FoldMI into the LIS maps. assert(!MIS.empty() && "Unexpected empty span of instructions!"); for (MachineInstr &MI : MIS) if (&MI != FoldMI) LIS.InsertMachineInstrInMaps(MI); // TII.foldMemoryOperand may have left some implicit operands on the // instruction. Strip them. if (ImpReg) for (unsigned i = FoldMI->getNumOperands(); i; --i) { MachineOperand &MO = FoldMI->getOperand(i - 1); if (!MO.isReg() || !MO.isImplicit()) break; if (MO.getReg() == ImpReg) FoldMI->removeOperand(i - 1); } LLVM_DEBUG(dumpMachineInstrRangeWithSlotIndex(MIS.begin(), MIS.end(), LIS, "folded")); if (!WasCopy) ++NumFolded; else if (Ops.front().second == 0) { ++NumSpills; // If there is only 1 store instruction is required for spill, add it // to mergeable list. In X86 AMX, 2 intructions are required to store. // We disable the merge for this case. if (std::distance(MIS.begin(), MIS.end()) <= 1) HSpiller.addToMergeableSpills(*FoldMI, StackSlot, Original); } else ++NumReloads; return true; } void InlineSpiller::insertReload(Register NewVReg, SlotIndex Idx, MachineBasicBlock::iterator MI) { MachineBasicBlock &MBB = *MI->getParent(); MachineInstrSpan MIS(MI, &MBB); TII.loadRegFromStackSlot(MBB, MI, NewVReg, StackSlot, MRI.getRegClass(NewVReg), &TRI, Register()); LIS.InsertMachineInstrRangeInMaps(MIS.begin(), MI); LLVM_DEBUG(dumpMachineInstrRangeWithSlotIndex(MIS.begin(), MI, LIS, "reload", NewVReg)); ++NumReloads; } /// Check if \p Def fully defines a VReg with an undefined value. /// If that's the case, that means the value of VReg is actually /// not relevant. static bool isRealSpill(const MachineInstr &Def) { if (!Def.isImplicitDef()) return true; assert(Def.getNumOperands() == 1 && "Implicit def with more than one definition"); // We can say that the VReg defined by Def is undef, only if it is // fully defined by Def. Otherwise, some of the lanes may not be // undef and the value of the VReg matters. return Def.getOperand(0).getSubReg(); } /// insertSpill - Insert a spill of NewVReg after MI. void InlineSpiller::insertSpill(Register NewVReg, bool isKill, MachineBasicBlock::iterator MI) { // Spill are not terminators, so inserting spills after terminators will // violate invariants in MachineVerifier. assert(!MI->isTerminator() && "Inserting a spill after a terminator"); MachineBasicBlock &MBB = *MI->getParent(); MachineInstrSpan MIS(MI, &MBB); MachineBasicBlock::iterator SpillBefore = std::next(MI); bool IsRealSpill = isRealSpill(*MI); if (IsRealSpill) TII.storeRegToStackSlot(MBB, SpillBefore, NewVReg, isKill, StackSlot, MRI.getRegClass(NewVReg), &TRI, Register()); else // Don't spill undef value. // Anything works for undef, in particular keeping the memory // uninitialized is a viable option and it saves code size and // run time. BuildMI(MBB, SpillBefore, MI->getDebugLoc(), TII.get(TargetOpcode::KILL)) .addReg(NewVReg, getKillRegState(isKill)); MachineBasicBlock::iterator Spill = std::next(MI); LIS.InsertMachineInstrRangeInMaps(Spill, MIS.end()); for (const MachineInstr &MI : make_range(Spill, MIS.end())) getVDefInterval(MI, LIS); LLVM_DEBUG( dumpMachineInstrRangeWithSlotIndex(Spill, MIS.end(), LIS, "spill")); ++NumSpills; // If there is only 1 store instruction is required for spill, add it // to mergeable list. In X86 AMX, 2 intructions are required to store. // We disable the merge for this case. if (IsRealSpill && std::distance(Spill, MIS.end()) <= 1) HSpiller.addToMergeableSpills(*Spill, StackSlot, Original); } /// spillAroundUses - insert spill code around each use of Reg. void InlineSpiller::spillAroundUses(Register Reg) { LLVM_DEBUG(dbgs() << "spillAroundUses " << printReg(Reg) << '\n'); LiveInterval &OldLI = LIS.getInterval(Reg); // Iterate over instructions using Reg. for (MachineInstr &MI : llvm::make_early_inc_range(MRI.reg_bundles(Reg))) { // Debug values are not allowed to affect codegen. if (MI.isDebugValue()) { // Modify DBG_VALUE now that the value is in a spill slot. MachineBasicBlock *MBB = MI.getParent(); LLVM_DEBUG(dbgs() << "Modifying debug info due to spill:\t" << MI); buildDbgValueForSpill(*MBB, &MI, MI, StackSlot, Reg); MBB->erase(MI); continue; } assert(!MI.isDebugInstr() && "Did not expect to find a use in debug " "instruction that isn't a DBG_VALUE"); // Ignore copies to/from snippets. We'll delete them. if (SnippetCopies.count(&MI)) continue; // Stack slot accesses may coalesce away. if (coalesceStackAccess(&MI, Reg)) continue; // Analyze instruction. SmallVector, 8> Ops; VirtRegInfo RI = AnalyzeVirtRegInBundle(MI, Reg, &Ops); // Find the slot index where this instruction reads and writes OldLI. // This is usually the def slot, except for tied early clobbers. SlotIndex Idx = LIS.getInstructionIndex(MI).getRegSlot(); if (VNInfo *VNI = OldLI.getVNInfoAt(Idx.getRegSlot(true))) if (SlotIndex::isSameInstr(Idx, VNI->def)) Idx = VNI->def; // Check for a sibling copy. Register SibReg = isCopyOfBundle(MI, Reg); if (SibReg && isSibling(SibReg)) { // This may actually be a copy between snippets. if (isRegToSpill(SibReg)) { LLVM_DEBUG(dbgs() << "Found new snippet copy: " << MI); SnippetCopies.insert(&MI); continue; } if (RI.Writes) { if (hoistSpillInsideBB(OldLI, MI)) { // This COPY is now dead, the value is already in the stack slot. MI.getOperand(0).setIsDead(); DeadDefs.push_back(&MI); continue; } } else { // This is a reload for a sib-reg copy. Drop spills downstream. LiveInterval &SibLI = LIS.getInterval(SibReg); eliminateRedundantSpills(SibLI, SibLI.getVNInfoAt(Idx)); // The COPY will fold to a reload below. } } // Attempt to fold memory ops. if (foldMemoryOperand(Ops)) continue; // Create a new virtual register for spill/fill. // FIXME: Infer regclass from instruction alone. Register NewVReg = Edit->createFrom(Reg); if (RI.Reads) insertReload(NewVReg, Idx, &MI); // Rewrite instruction operands. bool hasLiveDef = false; for (const auto &OpPair : Ops) { MachineOperand &MO = OpPair.first->getOperand(OpPair.second); MO.setReg(NewVReg); if (MO.isUse()) { if (!OpPair.first->isRegTiedToDefOperand(OpPair.second)) MO.setIsKill(); } else { if (!MO.isDead()) hasLiveDef = true; } } LLVM_DEBUG(dbgs() << "\trewrite: " << Idx << '\t' << MI << '\n'); // FIXME: Use a second vreg if instruction has no tied ops. if (RI.Writes) if (hasLiveDef) insertSpill(NewVReg, true, &MI); } } /// spillAll - Spill all registers remaining after rematerialization. void InlineSpiller::spillAll() { // Update LiveStacks now that we are committed to spilling. if (StackSlot == VirtRegMap::NO_STACK_SLOT) { StackSlot = VRM.assignVirt2StackSlot(Original); StackInt = &LSS.getOrCreateInterval(StackSlot, MRI.getRegClass(Original)); StackInt->getNextValue(SlotIndex(), LSS.getVNInfoAllocator()); } else StackInt = &LSS.getInterval(StackSlot); if (Original != Edit->getReg()) VRM.assignVirt2StackSlot(Edit->getReg(), StackSlot); assert(StackInt->getNumValNums() == 1 && "Bad stack interval values"); for (Register Reg : RegsToSpill) StackInt->MergeSegmentsInAsValue(LIS.getInterval(Reg), StackInt->getValNumInfo(0)); LLVM_DEBUG(dbgs() << "Merged spilled regs: " << *StackInt << '\n'); // Spill around uses of all RegsToSpill. for (Register Reg : RegsToSpill) spillAroundUses(Reg); // Hoisted spills may cause dead code. if (!DeadDefs.empty()) { LLVM_DEBUG(dbgs() << "Eliminating " << DeadDefs.size() << " dead defs\n"); Edit->eliminateDeadDefs(DeadDefs, RegsToSpill); } // Finally delete the SnippetCopies. for (Register Reg : RegsToSpill) { for (MachineInstr &MI : llvm::make_early_inc_range(MRI.reg_instructions(Reg))) { assert(SnippetCopies.count(&MI) && "Remaining use wasn't a snippet copy"); // FIXME: Do this with a LiveRangeEdit callback. LIS.getSlotIndexes()->removeSingleMachineInstrFromMaps(MI); MI.eraseFromBundle(); } } // Delete all spilled registers. for (Register Reg : RegsToSpill) Edit->eraseVirtReg(Reg); } void InlineSpiller::spill(LiveRangeEdit &edit) { ++NumSpilledRanges; Edit = &edit; assert(!Register::isStackSlot(edit.getReg()) && "Trying to spill a stack slot."); // Share a stack slot among all descendants of Original. Original = VRM.getOriginal(edit.getReg()); StackSlot = VRM.getStackSlot(Original); StackInt = nullptr; LLVM_DEBUG(dbgs() << "Inline spilling " << TRI.getRegClassName(MRI.getRegClass(edit.getReg())) << ':' << edit.getParent() << "\nFrom original " << printReg(Original) << '\n'); assert(edit.getParent().isSpillable() && "Attempting to spill already spilled value."); assert(DeadDefs.empty() && "Previous spill didn't remove dead defs"); collectRegsToSpill(); reMaterializeAll(); // Remat may handle everything. if (!RegsToSpill.empty()) spillAll(); Edit->calculateRegClassAndHint(MF, VRAI); } /// Optimizations after all the reg selections and spills are done. void InlineSpiller::postOptimization() { HSpiller.hoistAllSpills(); } /// When a spill is inserted, add the spill to MergeableSpills map. void HoistSpillHelper::addToMergeableSpills(MachineInstr &Spill, int StackSlot, unsigned Original) { BumpPtrAllocator &Allocator = LIS.getVNInfoAllocator(); LiveInterval &OrigLI = LIS.getInterval(Original); // save a copy of LiveInterval in StackSlotToOrigLI because the original // LiveInterval may be cleared after all its references are spilled. if (!StackSlotToOrigLI.contains(StackSlot)) { auto LI = std::make_unique(OrigLI.reg(), OrigLI.weight()); LI->assign(OrigLI, Allocator); StackSlotToOrigLI[StackSlot] = std::move(LI); } SlotIndex Idx = LIS.getInstructionIndex(Spill); VNInfo *OrigVNI = StackSlotToOrigLI[StackSlot]->getVNInfoAt(Idx.getRegSlot()); std::pair MIdx = std::make_pair(StackSlot, OrigVNI); MergeableSpills[MIdx].insert(&Spill); } /// When a spill is removed, remove the spill from MergeableSpills map. /// Return true if the spill is removed successfully. bool HoistSpillHelper::rmFromMergeableSpills(MachineInstr &Spill, int StackSlot) { auto It = StackSlotToOrigLI.find(StackSlot); if (It == StackSlotToOrigLI.end()) return false; SlotIndex Idx = LIS.getInstructionIndex(Spill); VNInfo *OrigVNI = It->second->getVNInfoAt(Idx.getRegSlot()); std::pair MIdx = std::make_pair(StackSlot, OrigVNI); return MergeableSpills[MIdx].erase(&Spill); } /// Check BB to see if it is a possible target BB to place a hoisted spill, /// i.e., there should be a living sibling of OrigReg at the insert point. bool HoistSpillHelper::isSpillCandBB(LiveInterval &OrigLI, VNInfo &OrigVNI, MachineBasicBlock &BB, Register &LiveReg) { SlotIndex Idx = IPA.getLastInsertPoint(OrigLI, BB); // The original def could be after the last insert point in the root block, // we can't hoist to here. if (Idx < OrigVNI.def) { // TODO: We could be better here. If LI is not alive in landing pad // we could hoist spill after LIP. LLVM_DEBUG(dbgs() << "can't spill in root block - def after LIP\n"); return false; } Register OrigReg = OrigLI.reg(); SmallSetVector &Siblings = Virt2SiblingsMap[OrigReg]; assert(OrigLI.getVNInfoAt(Idx) == &OrigVNI && "Unexpected VNI"); for (const Register &SibReg : Siblings) { LiveInterval &LI = LIS.getInterval(SibReg); VNInfo *VNI = LI.getVNInfoAt(Idx); if (VNI) { LiveReg = SibReg; return true; } } return false; } /// Remove redundant spills in the same BB. Save those redundant spills in /// SpillsToRm, and save the spill to keep and its BB in SpillBBToSpill map. void HoistSpillHelper::rmRedundantSpills( SmallPtrSet &Spills, SmallVectorImpl &SpillsToRm, DenseMap &SpillBBToSpill) { // For each spill saw, check SpillBBToSpill[] and see if its BB already has // another spill inside. If a BB contains more than one spill, only keep the // earlier spill with smaller SlotIndex. for (auto *const CurrentSpill : Spills) { MachineBasicBlock *Block = CurrentSpill->getParent(); MachineDomTreeNode *Node = MDT.getBase().getNode(Block); MachineInstr *PrevSpill = SpillBBToSpill[Node]; if (PrevSpill) { SlotIndex PIdx = LIS.getInstructionIndex(*PrevSpill); SlotIndex CIdx = LIS.getInstructionIndex(*CurrentSpill); MachineInstr *SpillToRm = (CIdx > PIdx) ? CurrentSpill : PrevSpill; MachineInstr *SpillToKeep = (CIdx > PIdx) ? PrevSpill : CurrentSpill; SpillsToRm.push_back(SpillToRm); SpillBBToSpill[MDT.getBase().getNode(Block)] = SpillToKeep; } else { SpillBBToSpill[MDT.getBase().getNode(Block)] = CurrentSpill; } } for (auto *const SpillToRm : SpillsToRm) Spills.erase(SpillToRm); } /// Starting from \p Root find a top-down traversal order of the dominator /// tree to visit all basic blocks containing the elements of \p Spills. /// Redundant spills will be found and put into \p SpillsToRm at the same /// time. \p SpillBBToSpill will be populated as part of the process and /// maps a basic block to the first store occurring in the basic block. /// \post SpillsToRm.union(Spills\@post) == Spills\@pre void HoistSpillHelper::getVisitOrders( MachineBasicBlock *Root, SmallPtrSet &Spills, SmallVectorImpl &Orders, SmallVectorImpl &SpillsToRm, DenseMap &SpillsToKeep, DenseMap &SpillBBToSpill) { // The set contains all the possible BB nodes to which we may hoist // original spills. SmallPtrSet WorkSet; // Save the BB nodes on the path from the first BB node containing // non-redundant spill to the Root node. SmallPtrSet NodesOnPath; // All the spills to be hoisted must originate from a single def instruction // to the OrigReg. It means the def instruction should dominate all the spills // to be hoisted. We choose the BB where the def instruction is located as // the Root. MachineDomTreeNode *RootIDomNode = MDT[Root]->getIDom(); // For every node on the dominator tree with spill, walk up on the dominator // tree towards the Root node until it is reached. If there is other node // containing spill in the middle of the path, the previous spill saw will // be redundant and the node containing it will be removed. All the nodes on // the path starting from the first node with non-redundant spill to the Root // node will be added to the WorkSet, which will contain all the possible // locations where spills may be hoisted to after the loop below is done. for (auto *const Spill : Spills) { MachineBasicBlock *Block = Spill->getParent(); MachineDomTreeNode *Node = MDT[Block]; MachineInstr *SpillToRm = nullptr; while (Node != RootIDomNode) { // If Node dominates Block, and it already contains a spill, the spill in // Block will be redundant. if (Node != MDT[Block] && SpillBBToSpill[Node]) { SpillToRm = SpillBBToSpill[MDT[Block]]; break; /// If we see the Node already in WorkSet, the path from the Node to /// the Root node must already be traversed by another spill. /// Then no need to repeat. } else if (WorkSet.count(Node)) { break; } else { NodesOnPath.insert(Node); } Node = Node->getIDom(); } if (SpillToRm) { SpillsToRm.push_back(SpillToRm); } else { // Add a BB containing the original spills to SpillsToKeep -- i.e., // set the initial status before hoisting start. The value of BBs // containing original spills is set to 0, in order to descriminate // with BBs containing hoisted spills which will be inserted to // SpillsToKeep later during hoisting. SpillsToKeep[MDT[Block]] = 0; WorkSet.insert(NodesOnPath.begin(), NodesOnPath.end()); } NodesOnPath.clear(); } // Sort the nodes in WorkSet in top-down order and save the nodes // in Orders. Orders will be used for hoisting in runHoistSpills. unsigned idx = 0; Orders.push_back(MDT.getBase().getNode(Root)); do { MachineDomTreeNode *Node = Orders[idx++]; for (MachineDomTreeNode *Child : Node->children()) { if (WorkSet.count(Child)) Orders.push_back(Child); } } while (idx != Orders.size()); assert(Orders.size() == WorkSet.size() && "Orders have different size with WorkSet"); #ifndef NDEBUG LLVM_DEBUG(dbgs() << "Orders size is " << Orders.size() << "\n"); SmallVector::reverse_iterator RIt = Orders.rbegin(); for (; RIt != Orders.rend(); RIt++) LLVM_DEBUG(dbgs() << "BB" << (*RIt)->getBlock()->getNumber() << ","); LLVM_DEBUG(dbgs() << "\n"); #endif } /// Try to hoist spills according to BB hotness. The spills to removed will /// be saved in \p SpillsToRm. The spills to be inserted will be saved in /// \p SpillsToIns. void HoistSpillHelper::runHoistSpills( LiveInterval &OrigLI, VNInfo &OrigVNI, SmallPtrSet &Spills, SmallVectorImpl &SpillsToRm, DenseMap &SpillsToIns) { // Visit order of dominator tree nodes. SmallVector Orders; // SpillsToKeep contains all the nodes where spills are to be inserted // during hoisting. If the spill to be inserted is an original spill // (not a hoisted one), the value of the map entry is 0. If the spill // is a hoisted spill, the value of the map entry is the VReg to be used // as the source of the spill. DenseMap SpillsToKeep; // Map from BB to the first spill inside of it. DenseMap SpillBBToSpill; rmRedundantSpills(Spills, SpillsToRm, SpillBBToSpill); MachineBasicBlock *Root = LIS.getMBBFromIndex(OrigVNI.def); getVisitOrders(Root, Spills, Orders, SpillsToRm, SpillsToKeep, SpillBBToSpill); // SpillsInSubTreeMap keeps the map from a dom tree node to a pair of // nodes set and the cost of all the spills inside those nodes. // The nodes set are the locations where spills are to be inserted // in the subtree of current node. using NodesCostPair = std::pair, BlockFrequency>; DenseMap SpillsInSubTreeMap; // Iterate Orders set in reverse order, which will be a bottom-up order // in the dominator tree. Once we visit a dom tree node, we know its // children have already been visited and the spill locations in the // subtrees of all the children have been determined. SmallVector::reverse_iterator RIt = Orders.rbegin(); for (; RIt != Orders.rend(); RIt++) { MachineBasicBlock *Block = (*RIt)->getBlock(); // If Block contains an original spill, simply continue. if (SpillsToKeep.contains(*RIt) && !SpillsToKeep[*RIt]) { SpillsInSubTreeMap[*RIt].first.insert(*RIt); // SpillsInSubTreeMap[*RIt].second contains the cost of spill. SpillsInSubTreeMap[*RIt].second = MBFI.getBlockFreq(Block); continue; } // Collect spills in subtree of current node (*RIt) to // SpillsInSubTreeMap[*RIt].first. for (MachineDomTreeNode *Child : (*RIt)->children()) { if (!SpillsInSubTreeMap.contains(Child)) continue; // The stmt "SpillsInSubTree = SpillsInSubTreeMap[*RIt].first" below // should be placed before getting the begin and end iterators of // SpillsInSubTreeMap[Child].first, or else the iterators may be // invalidated when SpillsInSubTreeMap[*RIt] is seen the first time // and the map grows and then the original buckets in the map are moved. SmallPtrSet &SpillsInSubTree = SpillsInSubTreeMap[*RIt].first; BlockFrequency &SubTreeCost = SpillsInSubTreeMap[*RIt].second; SubTreeCost += SpillsInSubTreeMap[Child].second; auto BI = SpillsInSubTreeMap[Child].first.begin(); auto EI = SpillsInSubTreeMap[Child].first.end(); SpillsInSubTree.insert(BI, EI); SpillsInSubTreeMap.erase(Child); } SmallPtrSet &SpillsInSubTree = SpillsInSubTreeMap[*RIt].first; BlockFrequency &SubTreeCost = SpillsInSubTreeMap[*RIt].second; // No spills in subtree, simply continue. if (SpillsInSubTree.empty()) continue; // Check whether Block is a possible candidate to insert spill. Register LiveReg; if (!isSpillCandBB(OrigLI, OrigVNI, *Block, LiveReg)) continue; // If there are multiple spills that could be merged, bias a little // to hoist the spill. BranchProbability MarginProb = (SpillsInSubTree.size() > 1) ? BranchProbability(9, 10) : BranchProbability(1, 1); if (SubTreeCost > MBFI.getBlockFreq(Block) * MarginProb) { // Hoist: Move spills to current Block. for (auto *const SpillBB : SpillsInSubTree) { // When SpillBB is a BB contains original spill, insert the spill // to SpillsToRm. if (SpillsToKeep.contains(SpillBB) && !SpillsToKeep[SpillBB]) { MachineInstr *SpillToRm = SpillBBToSpill[SpillBB]; SpillsToRm.push_back(SpillToRm); } // SpillBB will not contain spill anymore, remove it from SpillsToKeep. SpillsToKeep.erase(SpillBB); } // Current Block is the BB containing the new hoisted spill. Add it to // SpillsToKeep. LiveReg is the source of the new spill. SpillsToKeep[*RIt] = LiveReg; LLVM_DEBUG({ dbgs() << "spills in BB: "; for (const auto Rspill : SpillsInSubTree) dbgs() << Rspill->getBlock()->getNumber() << " "; dbgs() << "were promoted to BB" << (*RIt)->getBlock()->getNumber() << "\n"; }); SpillsInSubTree.clear(); SpillsInSubTree.insert(*RIt); SubTreeCost = MBFI.getBlockFreq(Block); } } // For spills in SpillsToKeep with LiveReg set (i.e., not original spill), // save them to SpillsToIns. for (const auto &Ent : SpillsToKeep) { if (Ent.second) SpillsToIns[Ent.first->getBlock()] = Ent.second; } } /// For spills with equal values, remove redundant spills and hoist those left /// to less hot spots. /// /// Spills with equal values will be collected into the same set in /// MergeableSpills when spill is inserted. These equal spills are originated /// from the same defining instruction and are dominated by the instruction. /// Before hoisting all the equal spills, redundant spills inside in the same /// BB are first marked to be deleted. Then starting from the spills left, walk /// up on the dominator tree towards the Root node where the define instruction /// is located, mark the dominated spills to be deleted along the way and /// collect the BB nodes on the path from non-dominated spills to the define /// instruction into a WorkSet. The nodes in WorkSet are the candidate places /// where we are considering to hoist the spills. We iterate the WorkSet in /// bottom-up order, and for each node, we will decide whether to hoist spills /// inside its subtree to that node. In this way, we can get benefit locally /// even if hoisting all the equal spills to one cold place is impossible. void HoistSpillHelper::hoistAllSpills() { SmallVector NewVRegs; LiveRangeEdit Edit(nullptr, NewVRegs, MF, LIS, &VRM, this); for (unsigned i = 0, e = MRI.getNumVirtRegs(); i != e; ++i) { Register Reg = Register::index2VirtReg(i); Register Original = VRM.getPreSplitReg(Reg); if (!MRI.def_empty(Reg)) Virt2SiblingsMap[Original].insert(Reg); } // Each entry in MergeableSpills contains a spill set with equal values. for (auto &Ent : MergeableSpills) { int Slot = Ent.first.first; LiveInterval &OrigLI = *StackSlotToOrigLI[Slot]; VNInfo *OrigVNI = Ent.first.second; SmallPtrSet &EqValSpills = Ent.second; if (Ent.second.empty()) continue; LLVM_DEBUG({ dbgs() << "\nFor Slot" << Slot << " and VN" << OrigVNI->id << ":\n" << "Equal spills in BB: "; for (const auto spill : EqValSpills) dbgs() << spill->getParent()->getNumber() << " "; dbgs() << "\n"; }); // SpillsToRm is the spill set to be removed from EqValSpills. SmallVector SpillsToRm; // SpillsToIns is the spill set to be newly inserted after hoisting. DenseMap SpillsToIns; runHoistSpills(OrigLI, *OrigVNI, EqValSpills, SpillsToRm, SpillsToIns); LLVM_DEBUG({ dbgs() << "Finally inserted spills in BB: "; for (const auto &Ispill : SpillsToIns) dbgs() << Ispill.first->getNumber() << " "; dbgs() << "\nFinally removed spills in BB: "; for (const auto Rspill : SpillsToRm) dbgs() << Rspill->getParent()->getNumber() << " "; dbgs() << "\n"; }); // Stack live range update. LiveInterval &StackIntvl = LSS.getInterval(Slot); if (!SpillsToIns.empty() || !SpillsToRm.empty()) StackIntvl.MergeValueInAsValue(OrigLI, OrigVNI, StackIntvl.getValNumInfo(0)); // Insert hoisted spills. for (auto const &Insert : SpillsToIns) { MachineBasicBlock *BB = Insert.first; Register LiveReg = Insert.second; MachineBasicBlock::iterator MII = IPA.getLastInsertPointIter(OrigLI, *BB); MachineInstrSpan MIS(MII, BB); TII.storeRegToStackSlot(*BB, MII, LiveReg, false, Slot, MRI.getRegClass(LiveReg), &TRI, Register()); LIS.InsertMachineInstrRangeInMaps(MIS.begin(), MII); for (const MachineInstr &MI : make_range(MIS.begin(), MII)) getVDefInterval(MI, LIS); ++NumSpills; } // Remove redundant spills or change them to dead instructions. NumSpills -= SpillsToRm.size(); for (auto *const RMEnt : SpillsToRm) { RMEnt->setDesc(TII.get(TargetOpcode::KILL)); for (unsigned i = RMEnt->getNumOperands(); i; --i) { MachineOperand &MO = RMEnt->getOperand(i - 1); if (MO.isReg() && MO.isImplicit() && MO.isDef() && !MO.isDead()) RMEnt->removeOperand(i - 1); } } Edit.eliminateDeadDefs(SpillsToRm, std::nullopt); } } /// For VirtReg clone, the \p New register should have the same physreg or /// stackslot as the \p old register. void HoistSpillHelper::LRE_DidCloneVirtReg(Register New, Register Old) { if (VRM.hasPhys(Old)) VRM.assignVirt2Phys(New, VRM.getPhys(Old)); else if (VRM.getStackSlot(Old) != VirtRegMap::NO_STACK_SLOT) VRM.assignVirt2StackSlot(New, VRM.getStackSlot(Old)); else llvm_unreachable("VReg should be assigned either physreg or stackslot"); if (VRM.hasShape(Old)) VRM.assignVirt2Shape(New, VRM.getShape(Old)); }