//===-- ARMLowOverheadLoops.cpp - CodeGen Low-overhead Loops ---*- C++ -*-===// // // 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 // //===----------------------------------------------------------------------===// /// \file /// Finalize v8.1-m low-overhead loops by converting the associated pseudo /// instructions into machine operations. /// The expectation is that the loop contains three pseudo instructions: /// - t2*LoopStart - placed in the preheader or pre-preheader. The do-loop /// form should be in the preheader, whereas the while form should be in the /// preheaders only predecessor. /// - t2LoopDec - placed within in the loop body. /// - t2LoopEnd - the loop latch terminator. /// /// In addition to this, we also look for the presence of the VCTP instruction, /// which determines whether we can generated the tail-predicated low-overhead /// loop form. /// /// Assumptions and Dependencies: /// Low-overhead loops are constructed and executed using a setup instruction: /// DLS, WLS, DLSTP or WLSTP and an instruction that loops back: LE or LETP. /// WLS(TP) and LE(TP) are branching instructions with a (large) limited range /// but fixed polarity: WLS can only branch forwards and LE can only branch /// backwards. These restrictions mean that this pass is dependent upon block /// layout and block sizes, which is why it's the last pass to run. The same is /// true for ConstantIslands, but this pass does not increase the size of the /// basic blocks, nor does it change the CFG. Instructions are mainly removed /// during the transform and pseudo instructions are replaced by real ones. In /// some cases, when we have to revert to a 'normal' loop, we have to introduce /// multiple instructions for a single pseudo (see RevertWhile and /// RevertLoopEnd). To handle this situation, t2WhileLoopStart and t2LoopEnd /// are defined to be as large as this maximum sequence of replacement /// instructions. /// /// A note on VPR.P0 (the lane mask): /// VPT, VCMP, VPNOT and VCTP won't overwrite VPR.P0 when they update it in a /// "VPT Active" context (which includes low-overhead loops and vpt blocks). /// They will simply "and" the result of their calculation with the current /// value of VPR.P0. You can think of it like this: /// \verbatim /// if VPT active: ; Between a DLSTP/LETP, or for predicated instrs /// VPR.P0 &= Value /// else /// VPR.P0 = Value /// \endverbatim /// When we're inside the low-overhead loop (between DLSTP and LETP), we always /// fall in the "VPT active" case, so we can consider that all VPR writes by /// one of those instruction is actually a "and". //===----------------------------------------------------------------------===// #include "ARM.h" #include "ARMBaseInstrInfo.h" #include "ARMBaseRegisterInfo.h" #include "ARMBasicBlockInfo.h" #include "ARMSubtarget.h" #include "Thumb2InstrInfo.h" #include "llvm/ADT/SetOperations.h" #include "llvm/ADT/SmallSet.h" #include "llvm/CodeGen/LivePhysRegs.h" #include "llvm/CodeGen/MachineFunctionPass.h" #include "llvm/CodeGen/MachineLoopInfo.h" #include "llvm/CodeGen/MachineLoopUtils.h" #include "llvm/CodeGen/MachineRegisterInfo.h" #include "llvm/CodeGen/Passes.h" #include "llvm/CodeGen/ReachingDefAnalysis.h" #include "llvm/MC/MCInstrDesc.h" using namespace llvm; #define DEBUG_TYPE "arm-low-overhead-loops" #define ARM_LOW_OVERHEAD_LOOPS_NAME "ARM Low Overhead Loops pass" namespace { using InstSet = SmallPtrSetImpl; class PostOrderLoopTraversal { MachineLoop &ML; MachineLoopInfo &MLI; SmallPtrSet Visited; SmallVector Order; public: PostOrderLoopTraversal(MachineLoop &ML, MachineLoopInfo &MLI) : ML(ML), MLI(MLI) { } const SmallVectorImpl &getOrder() const { return Order; } // Visit all the blocks within the loop, as well as exit blocks and any // blocks properly dominating the header. void ProcessLoop() { std::function Search = [this, &Search] (MachineBasicBlock *MBB) -> void { if (Visited.count(MBB)) return; Visited.insert(MBB); for (auto *Succ : MBB->successors()) { if (!ML.contains(Succ)) continue; Search(Succ); } Order.push_back(MBB); }; // Insert exit blocks. SmallVector ExitBlocks; ML.getExitBlocks(ExitBlocks); for (auto *MBB : ExitBlocks) Order.push_back(MBB); // Then add the loop body. Search(ML.getHeader()); // Then try the preheader and its predecessors. std::function GetPredecessor = [this, &GetPredecessor] (MachineBasicBlock *MBB) -> void { Order.push_back(MBB); if (MBB->pred_size() == 1) GetPredecessor(*MBB->pred_begin()); }; if (auto *Preheader = ML.getLoopPreheader()) GetPredecessor(Preheader); else if (auto *Preheader = MLI.findLoopPreheader(&ML, true)) GetPredecessor(Preheader); } }; struct PredicatedMI { MachineInstr *MI = nullptr; SetVector Predicates; public: PredicatedMI(MachineInstr *I, SetVector &Preds) : MI(I) { assert(I && "Instruction must not be null!"); Predicates.insert(Preds.begin(), Preds.end()); } }; // Represent a VPT block, a list of instructions that begins with a VPT/VPST // and has a maximum of four proceeding instructions. All instructions within // the block are predicated upon the vpr and we allow instructions to define // the vpr within in the block too. class VPTBlock { // The predicate then instruction, which is either a VPT, or a VPST // instruction. std::unique_ptr PredicateThen; PredicatedMI *Divergent = nullptr; SmallVector Insts; public: VPTBlock(MachineInstr *MI, SetVector &Preds) { PredicateThen = std::make_unique(MI, Preds); } void addInst(MachineInstr *MI, SetVector &Preds) { LLVM_DEBUG(dbgs() << "ARM Loops: Adding predicated MI: " << *MI); if (!Divergent && !set_difference(Preds, PredicateThen->Predicates).empty()) { Divergent = &Insts.back(); LLVM_DEBUG(dbgs() << " - has divergent predicate: " << *Divergent->MI); } Insts.emplace_back(MI, Preds); assert(Insts.size() <= 4 && "Too many instructions in VPT block!"); } // Have we found an instruction within the block which defines the vpr? If // so, not all the instructions in the block will have the same predicate. bool HasNonUniformPredicate() const { return Divergent != nullptr; } // Is the given instruction part of the predicate set controlling the entry // to the block. bool IsPredicatedOn(MachineInstr *MI) const { return PredicateThen->Predicates.count(MI); } // Returns true if this is a VPT instruction. bool isVPT() const { return !isVPST(); } // Returns true if this is a VPST instruction. bool isVPST() const { return PredicateThen->MI->getOpcode() == ARM::MVE_VPST; } // Is the given instruction the only predicate which controls the entry to // the block. bool IsOnlyPredicatedOn(MachineInstr *MI) const { return IsPredicatedOn(MI) && PredicateThen->Predicates.size() == 1; } unsigned size() const { return Insts.size(); } SmallVectorImpl &getInsts() { return Insts; } MachineInstr *getPredicateThen() const { return PredicateThen->MI; } PredicatedMI *getDivergent() const { return Divergent; } }; struct Reduction { MachineInstr *Init; MachineInstr &Copy; MachineInstr &Reduce; MachineInstr &VPSEL; Reduction(MachineInstr *Init, MachineInstr *Mov, MachineInstr *Add, MachineInstr *Sel) : Init(Init), Copy(*Mov), Reduce(*Add), VPSEL(*Sel) { } }; struct LowOverheadLoop { MachineLoop &ML; MachineBasicBlock *Preheader = nullptr; MachineLoopInfo &MLI; ReachingDefAnalysis &RDA; const TargetRegisterInfo &TRI; const ARMBaseInstrInfo &TII; MachineFunction *MF = nullptr; MachineInstr *InsertPt = nullptr; MachineInstr *Start = nullptr; MachineInstr *Dec = nullptr; MachineInstr *End = nullptr; MachineInstr *VCTP = nullptr; SmallPtrSet SecondaryVCTPs; VPTBlock *CurrentBlock = nullptr; SetVector CurrentPredicate; SmallVector VPTBlocks; SmallPtrSet ToRemove; SmallVector, 1> Reductions; SmallPtrSet BlockMasksToRecompute; bool Revert = false; bool CannotTailPredicate = false; LowOverheadLoop(MachineLoop &ML, MachineLoopInfo &MLI, ReachingDefAnalysis &RDA, const TargetRegisterInfo &TRI, const ARMBaseInstrInfo &TII) : ML(ML), MLI(MLI), RDA(RDA), TRI(TRI), TII(TII) { MF = ML.getHeader()->getParent(); if (auto *MBB = ML.getLoopPreheader()) Preheader = MBB; else if (auto *MBB = MLI.findLoopPreheader(&ML, true)) Preheader = MBB; } // If this is an MVE instruction, check that we know how to use tail // predication with it. Record VPT blocks and return whether the // instruction is valid for tail predication. bool ValidateMVEInst(MachineInstr *MI); void AnalyseMVEInst(MachineInstr *MI) { CannotTailPredicate = !ValidateMVEInst(MI); } bool IsTailPredicationLegal() const { // For now, let's keep things really simple and only support a single // block for tail predication. return !Revert && FoundAllComponents() && VCTP && !CannotTailPredicate && ML.getNumBlocks() == 1; } // Check that the predication in the loop will be equivalent once we // perform the conversion. Also ensure that we can provide the number // of elements to the loop start instruction. bool ValidateTailPredicate(MachineInstr *StartInsertPt); // See whether the live-out instructions are a reduction that we can fixup // later. bool FindValidReduction(InstSet &LiveMIs, InstSet &LiveOutUsers); // Check that any values available outside of the loop will be the same // after tail predication conversion. bool ValidateLiveOuts(); // Is it safe to define LR with DLS/WLS? // LR can be defined if it is the operand to start, because it's the same // value, or if it's going to be equivalent to the operand to Start. MachineInstr *isSafeToDefineLR(); // Check the branch targets are within range and we satisfy our // restrictions. void CheckLegality(ARMBasicBlockUtils *BBUtils); bool FoundAllComponents() const { return Start && Dec && End; } SmallVectorImpl &getVPTBlocks() { return VPTBlocks; } // Return the loop iteration count, or the number of elements if we're tail // predicating. MachineOperand &getCount() { return IsTailPredicationLegal() ? VCTP->getOperand(1) : Start->getOperand(0); } unsigned getStartOpcode() const { bool IsDo = Start->getOpcode() == ARM::t2DoLoopStart; if (!IsTailPredicationLegal()) return IsDo ? ARM::t2DLS : ARM::t2WLS; return VCTPOpcodeToLSTP(VCTP->getOpcode(), IsDo); } void dump() const { if (Start) dbgs() << "ARM Loops: Found Loop Start: " << *Start; if (Dec) dbgs() << "ARM Loops: Found Loop Dec: " << *Dec; if (End) dbgs() << "ARM Loops: Found Loop End: " << *End; if (VCTP) dbgs() << "ARM Loops: Found VCTP: " << *VCTP; if (!FoundAllComponents()) dbgs() << "ARM Loops: Not a low-overhead loop.\n"; else if (!(Start && Dec && End)) dbgs() << "ARM Loops: Failed to find all loop components.\n"; } }; class ARMLowOverheadLoops : public MachineFunctionPass { MachineFunction *MF = nullptr; MachineLoopInfo *MLI = nullptr; ReachingDefAnalysis *RDA = nullptr; const ARMBaseInstrInfo *TII = nullptr; MachineRegisterInfo *MRI = nullptr; const TargetRegisterInfo *TRI = nullptr; std::unique_ptr BBUtils = nullptr; public: static char ID; ARMLowOverheadLoops() : MachineFunctionPass(ID) { } void getAnalysisUsage(AnalysisUsage &AU) const override { AU.setPreservesCFG(); AU.addRequired(); AU.addRequired(); MachineFunctionPass::getAnalysisUsage(AU); } bool runOnMachineFunction(MachineFunction &MF) override; MachineFunctionProperties getRequiredProperties() const override { return MachineFunctionProperties().set( MachineFunctionProperties::Property::NoVRegs).set( MachineFunctionProperties::Property::TracksLiveness); } StringRef getPassName() const override { return ARM_LOW_OVERHEAD_LOOPS_NAME; } private: bool ProcessLoop(MachineLoop *ML); bool RevertNonLoops(); void RevertWhile(MachineInstr *MI) const; bool RevertLoopDec(MachineInstr *MI) const; void RevertLoopEnd(MachineInstr *MI, bool SkipCmp = false) const; void ConvertVPTBlocks(LowOverheadLoop &LoLoop); void FixupReductions(LowOverheadLoop &LoLoop) const; MachineInstr *ExpandLoopStart(LowOverheadLoop &LoLoop); void Expand(LowOverheadLoop &LoLoop); void IterationCountDCE(LowOverheadLoop &LoLoop); }; } char ARMLowOverheadLoops::ID = 0; INITIALIZE_PASS(ARMLowOverheadLoops, DEBUG_TYPE, ARM_LOW_OVERHEAD_LOOPS_NAME, false, false) MachineInstr *LowOverheadLoop::isSafeToDefineLR() { // We can define LR because LR already contains the same value. if (Start->getOperand(0).getReg() == ARM::LR) return Start; unsigned CountReg = Start->getOperand(0).getReg(); auto IsMoveLR = [&CountReg](MachineInstr *MI) { return MI->getOpcode() == ARM::tMOVr && MI->getOperand(0).getReg() == ARM::LR && MI->getOperand(1).getReg() == CountReg && MI->getOperand(2).getImm() == ARMCC::AL; }; MachineBasicBlock *MBB = Start->getParent(); // Find an insertion point: // - Is there a (mov lr, Count) before Start? If so, and nothing else writes // to Count before Start, we can insert at that mov. if (auto *LRDef = RDA.getUniqueReachingMIDef(Start, ARM::LR)) if (IsMoveLR(LRDef) && RDA.hasSameReachingDef(Start, LRDef, CountReg)) return LRDef; // - Is there a (mov lr, Count) after Start? If so, and nothing else writes // to Count after Start, we can insert at that mov. if (auto *LRDef = RDA.getLocalLiveOutMIDef(MBB, ARM::LR)) if (IsMoveLR(LRDef) && RDA.hasSameReachingDef(Start, LRDef, CountReg)) return LRDef; // We've found no suitable LR def and Start doesn't use LR directly. Can we // just define LR anyway? return RDA.isSafeToDefRegAt(Start, ARM::LR) ? Start : nullptr; } bool LowOverheadLoop::ValidateTailPredicate(MachineInstr *StartInsertPt) { assert(VCTP && "VCTP instruction expected but is not set"); // All predication within the loop should be based on vctp. If the block // isn't predicated on entry, check whether the vctp is within the block // and that all other instructions are then predicated on it. for (auto &Block : VPTBlocks) { if (Block.IsPredicatedOn(VCTP)) continue; if (Block.HasNonUniformPredicate() && !isVCTP(Block.getDivergent()->MI)) { LLVM_DEBUG(dbgs() << "ARM Loops: Found unsupported diverging predicate: " << *Block.getDivergent()->MI); return false; } SmallVectorImpl &Insts = Block.getInsts(); for (auto &PredMI : Insts) { // Check the instructions in the block and only allow: // - VCTPs // - Instructions predicated on the main VCTP // - Any VCMP // - VCMPs just "and" their result with VPR.P0. Whether they are // located before/after the VCTP is irrelevant - the end result will // be the same in both cases, so there's no point in requiring them // to be located after the VCTP! if (PredMI.Predicates.count(VCTP) || isVCTP(PredMI.MI) || VCMPOpcodeToVPT(PredMI.MI->getOpcode()) != 0) continue; LLVM_DEBUG(dbgs() << "ARM Loops: Can't convert: " << *PredMI.MI << " - which is predicated on:\n"; for (auto *MI : PredMI.Predicates) dbgs() << " - " << *MI); return false; } } if (!ValidateLiveOuts()) return false; // For tail predication, we need to provide the number of elements, instead // of the iteration count, to the loop start instruction. The number of // elements is provided to the vctp instruction, so we need to check that // we can use this register at InsertPt. Register NumElements = VCTP->getOperand(1).getReg(); // If the register is defined within loop, then we can't perform TP. // TODO: Check whether this is just a mov of a register that would be // available. if (RDA.hasLocalDefBefore(VCTP, NumElements)) { LLVM_DEBUG(dbgs() << "ARM Loops: VCTP operand is defined in the loop.\n"); return false; } // The element count register maybe defined after InsertPt, in which case we // need to try to move either InsertPt or the def so that the [w|d]lstp can // use the value. // TODO: On failing to move an instruction, check if the count is provided by // a mov and whether we can use the mov operand directly. MachineBasicBlock *InsertBB = StartInsertPt->getParent(); if (!RDA.isReachingDefLiveOut(StartInsertPt, NumElements)) { if (auto *ElemDef = RDA.getLocalLiveOutMIDef(InsertBB, NumElements)) { if (RDA.isSafeToMoveForwards(ElemDef, StartInsertPt)) { ElemDef->removeFromParent(); InsertBB->insert(MachineBasicBlock::iterator(StartInsertPt), ElemDef); LLVM_DEBUG(dbgs() << "ARM Loops: Moved element count def: " << *ElemDef); } else if (RDA.isSafeToMoveBackwards(StartInsertPt, ElemDef)) { StartInsertPt->removeFromParent(); InsertBB->insertAfter(MachineBasicBlock::iterator(ElemDef), StartInsertPt); LLVM_DEBUG(dbgs() << "ARM Loops: Moved start past: " << *ElemDef); } else { LLVM_DEBUG(dbgs() << "ARM Loops: Unable to move element count to loop " << "start instruction.\n"); return false; } } } // Especially in the case of while loops, InsertBB may not be the // preheader, so we need to check that the register isn't redefined // before entering the loop. auto CannotProvideElements = [this](MachineBasicBlock *MBB, Register NumElements) { // NumElements is redefined in this block. if (RDA.hasLocalDefBefore(&MBB->back(), NumElements)) return true; // Don't continue searching up through multiple predecessors. if (MBB->pred_size() > 1) return true; return false; }; // First, find the block that looks like the preheader. MachineBasicBlock *MBB = Preheader; if (!MBB) { LLVM_DEBUG(dbgs() << "ARM Loops: Didn't find preheader.\n"); return false; } // Then search backwards for a def, until we get to InsertBB. while (MBB != InsertBB) { if (CannotProvideElements(MBB, NumElements)) { LLVM_DEBUG(dbgs() << "ARM Loops: Unable to provide element count.\n"); return false; } MBB = *MBB->pred_begin(); } // Check that the value change of the element count is what we expect and // that the predication will be equivalent. For this we need: // NumElements = NumElements - VectorWidth. The sub will be a sub immediate // and we can also allow register copies within the chain too. auto IsValidSub = [](MachineInstr *MI, int ExpectedVecWidth) { return -getAddSubImmediate(*MI) == ExpectedVecWidth; }; MBB = VCTP->getParent(); if (auto *Def = RDA.getUniqueReachingMIDef(&MBB->back(), NumElements)) { SmallPtrSet ElementChain; SmallPtrSet Ignore = { VCTP }; unsigned ExpectedVectorWidth = getTailPredVectorWidth(VCTP->getOpcode()); Ignore.insert(SecondaryVCTPs.begin(), SecondaryVCTPs.end()); if (RDA.isSafeToRemove(Def, ElementChain, Ignore)) { bool FoundSub = false; for (auto *MI : ElementChain) { if (isMovRegOpcode(MI->getOpcode())) continue; if (isSubImmOpcode(MI->getOpcode())) { if (FoundSub || !IsValidSub(MI, ExpectedVectorWidth)) return false; FoundSub = true; } else return false; } LLVM_DEBUG(dbgs() << "ARM Loops: Will remove element count chain:\n"; for (auto *MI : ElementChain) dbgs() << " - " << *MI); ToRemove.insert(ElementChain.begin(), ElementChain.end()); } } return true; } static bool isVectorPredicated(MachineInstr *MI) { int PIdx = llvm::findFirstVPTPredOperandIdx(*MI); return PIdx != -1 && MI->getOperand(PIdx + 1).getReg() == ARM::VPR; } static bool isRegInClass(const MachineOperand &MO, const TargetRegisterClass *Class) { return MO.isReg() && MO.getReg() && Class->contains(MO.getReg()); } // MVE 'narrowing' operate on half a lane, reading from half and writing // to half, which are referred to has the top and bottom half. The other // half retains its previous value. static bool retainsPreviousHalfElement(const MachineInstr &MI) { const MCInstrDesc &MCID = MI.getDesc(); uint64_t Flags = MCID.TSFlags; return (Flags & ARMII::RetainsPreviousHalfElement) != 0; } // Some MVE instructions read from the top/bottom halves of their operand(s) // and generate a vector result with result elements that are double the // width of the input. static bool producesDoubleWidthResult(const MachineInstr &MI) { const MCInstrDesc &MCID = MI.getDesc(); uint64_t Flags = MCID.TSFlags; return (Flags & ARMII::DoubleWidthResult) != 0; } static bool isHorizontalReduction(const MachineInstr &MI) { const MCInstrDesc &MCID = MI.getDesc(); uint64_t Flags = MCID.TSFlags; return (Flags & ARMII::HorizontalReduction) != 0; } // Can this instruction generate a non-zero result when given only zeroed // operands? This allows us to know that, given operands with false bytes // zeroed by masked loads, that the result will also contain zeros in those // bytes. static bool canGenerateNonZeros(const MachineInstr &MI) { // Check for instructions which can write into a larger element size, // possibly writing into a previous zero'd lane. if (producesDoubleWidthResult(MI)) return true; switch (MI.getOpcode()) { default: break; // FIXME: VNEG FP and -0? I think we'll need to handle this once we allow // fp16 -> fp32 vector conversions. // Instructions that perform a NOT will generate 1s from 0s. case ARM::MVE_VMVN: case ARM::MVE_VORN: // Count leading zeros will do just that! case ARM::MVE_VCLZs8: case ARM::MVE_VCLZs16: case ARM::MVE_VCLZs32: return true; } return false; } // Look at its register uses to see if it only can only receive zeros // into its false lanes which would then produce zeros. Also check that // the output register is also defined by an FalseLanesZero instruction // so that if tail-predication happens, the lanes that aren't updated will // still be zeros. static bool producesFalseLanesZero(MachineInstr &MI, const TargetRegisterClass *QPRs, const ReachingDefAnalysis &RDA, InstSet &FalseLanesZero) { if (canGenerateNonZeros(MI)) return false; bool AllowScalars = isHorizontalReduction(MI); for (auto &MO : MI.operands()) { if (!MO.isReg() || !MO.getReg()) continue; if (!isRegInClass(MO, QPRs) && AllowScalars) continue; if (auto *OpDef = RDA.getMIOperand(&MI, MO)) if (FalseLanesZero.count(OpDef)) continue; return false; } LLVM_DEBUG(dbgs() << "ARM Loops: Always False Zeros: " << MI); return true; } bool LowOverheadLoop::FindValidReduction(InstSet &LiveMIs, InstSet &LiveOutUsers) { // Also check for reductions where the operation needs to be merging values // from the last and previous loop iterations. This means an instruction // producing a value and a vmov storing the value calculated in the previous // iteration. So we can have two live-out regs, one produced by a vmov and // both being consumed by a vpsel. LLVM_DEBUG(dbgs() << "ARM Loops: Looking for reduction live-outs:\n"; for (auto *MI : LiveMIs) dbgs() << " - " << *MI); if (!Preheader) return false; // Expect a vmov, a vadd and a single vpsel user. // TODO: This means we can't currently support multiple reductions in the // loop. if (LiveMIs.size() != 2 || LiveOutUsers.size() != 1) return false; MachineInstr *VPSEL = *LiveOutUsers.begin(); if (VPSEL->getOpcode() != ARM::MVE_VPSEL) return false; unsigned VPRIdx = llvm::findFirstVPTPredOperandIdx(*VPSEL) + 1; MachineInstr *Pred = RDA.getMIOperand(VPSEL, VPRIdx); if (!Pred || Pred != VCTP) { LLVM_DEBUG(dbgs() << "ARM Loops: Not using equivalent predicate.\n"); return false; } MachineInstr *Reduce = RDA.getMIOperand(VPSEL, 1); if (!Reduce) return false; assert(LiveMIs.count(Reduce) && "Expected MI to be live-out"); // TODO: Support more operations than VADD. switch (VCTP->getOpcode()) { default: return false; case ARM::MVE_VCTP8: if (Reduce->getOpcode() != ARM::MVE_VADDi8) return false; break; case ARM::MVE_VCTP16: if (Reduce->getOpcode() != ARM::MVE_VADDi16) return false; break; case ARM::MVE_VCTP32: if (Reduce->getOpcode() != ARM::MVE_VADDi32) return false; break; } // Test that the reduce op is overwriting ones of its operands. if (Reduce->getOperand(0).getReg() != Reduce->getOperand(1).getReg() && Reduce->getOperand(0).getReg() != Reduce->getOperand(2).getReg()) { LLVM_DEBUG(dbgs() << "ARM Loops: Reducing op isn't overwriting itself.\n"); return false; } // Check that the VORR is actually a VMOV. MachineInstr *Copy = RDA.getMIOperand(VPSEL, 2); if (!Copy || Copy->getOpcode() != ARM::MVE_VORR || !Copy->getOperand(1).isReg() || !Copy->getOperand(2).isReg() || Copy->getOperand(1).getReg() != Copy->getOperand(2).getReg()) return false; assert(LiveMIs.count(Copy) && "Expected MI to be live-out"); // Check that the vadd and vmov are only used by each other and the vpsel. SmallPtrSet CopyUsers; RDA.getGlobalUses(Copy, Copy->getOperand(0).getReg(), CopyUsers); if (CopyUsers.size() > 2 || !CopyUsers.count(Reduce)) { LLVM_DEBUG(dbgs() << "ARM Loops: Copy users unsupported.\n"); return false; } SmallPtrSet ReduceUsers; RDA.getGlobalUses(Reduce, Reduce->getOperand(0).getReg(), ReduceUsers); if (ReduceUsers.size() > 2 || !ReduceUsers.count(Copy)) { LLVM_DEBUG(dbgs() << "ARM Loops: Reduce users unsupported.\n"); return false; } // Then find whether there's an instruction initialising the register that // is storing the reduction. SmallPtrSet Incoming; RDA.getLiveOuts(Preheader, Copy->getOperand(1).getReg(), Incoming); if (Incoming.size() > 1) return false; MachineInstr *Init = Incoming.empty() ? nullptr : *Incoming.begin(); LLVM_DEBUG(dbgs() << "ARM Loops: Found a reduction:\n" << " - " << *Copy << " - " << *Reduce << " - " << *VPSEL); Reductions.push_back(std::make_unique(Init, Copy, Reduce, VPSEL)); return true; } bool LowOverheadLoop::ValidateLiveOuts() { // We want to find out if the tail-predicated version of this loop will // produce the same values as the loop in its original form. For this to // be true, the newly inserted implicit predication must not change the // the (observable) results. // We're doing this because many instructions in the loop will not be // predicated and so the conversion from VPT predication to tail-predication // can result in different values being produced; due to the tail-predication // preventing many instructions from updating their falsely predicated // lanes. This analysis assumes that all the instructions perform lane-wise // operations and don't perform any exchanges. // A masked load, whether through VPT or tail predication, will write zeros // to any of the falsely predicated bytes. So, from the loads, we know that // the false lanes are zeroed and here we're trying to track that those false // lanes remain zero, or where they change, the differences are masked away // by their user(s). // All MVE loads and stores have to be predicated, so we know that any load // operands, or stored results are equivalent already. Other explicitly // predicated instructions will perform the same operation in the original // loop and the tail-predicated form too. Because of this, we can insert // loads, stores and other predicated instructions into our Predicated // set and build from there. const TargetRegisterClass *QPRs = TRI.getRegClass(ARM::MQPRRegClassID); SetVector FalseLanesUnknown; SmallPtrSet FalseLanesZero; SmallPtrSet Predicated; MachineBasicBlock *Header = ML.getHeader(); for (auto &MI : *Header) { const MCInstrDesc &MCID = MI.getDesc(); uint64_t Flags = MCID.TSFlags; if ((Flags & ARMII::DomainMask) != ARMII::DomainMVE) continue; if (isVCTP(&MI) || isVPTOpcode(MI.getOpcode())) continue; // Predicated loads will write zeros to the falsely predicated bytes of the // destination register. if (isVectorPredicated(&MI)) { if (MI.mayLoad()) FalseLanesZero.insert(&MI); Predicated.insert(&MI); continue; } if (MI.getNumDefs() == 0) continue; if (!producesFalseLanesZero(MI, QPRs, RDA, FalseLanesZero)) { // We require retaining and horizontal operations to operate upon zero'd // false lanes to ensure the conversion doesn't change the output. if (retainsPreviousHalfElement(MI) || isHorizontalReduction(MI)) return false; // Otherwise we need to evaluate this instruction later to see whether // unknown false lanes will get masked away by their user(s). FalseLanesUnknown.insert(&MI); } else if (!isHorizontalReduction(MI)) FalseLanesZero.insert(&MI); } auto HasPredicatedUsers = [this](MachineInstr *MI, const MachineOperand &MO, SmallPtrSetImpl &Predicated) { SmallPtrSet Uses; RDA.getGlobalUses(MI, MO.getReg(), Uses); for (auto *Use : Uses) { if (Use != MI && !Predicated.count(Use)) return false; } return true; }; // Visit the unknowns in reverse so that we can start at the values being // stored and then we can work towards the leaves, hopefully adding more // instructions to Predicated. Successfully terminating the loop means that // all the unknown values have to found to be masked by predicated user(s). // For any unpredicated values, we store them in NonPredicated so that we // can later check whether these form a reduction. SmallPtrSet NonPredicated; for (auto *MI : reverse(FalseLanesUnknown)) { for (auto &MO : MI->operands()) { if (!isRegInClass(MO, QPRs) || !MO.isDef()) continue; if (!HasPredicatedUsers(MI, MO, Predicated)) { LLVM_DEBUG(dbgs() << "ARM Loops: Found an unknown def of : " << TRI.getRegAsmName(MO.getReg()) << " at " << *MI); NonPredicated.insert(MI); continue; } } // Any unknown false lanes have been masked away by the user(s). Predicated.insert(MI); } SmallPtrSet LiveOutMIs; SmallPtrSet LiveOutUsers; SmallVector ExitBlocks; ML.getExitBlocks(ExitBlocks); assert(ML.getNumBlocks() == 1 && "Expected single block loop!"); assert(ExitBlocks.size() == 1 && "Expected a single exit block"); MachineBasicBlock *ExitBB = ExitBlocks.front(); for (const MachineBasicBlock::RegisterMaskPair &RegMask : ExitBB->liveins()) { // Check Q-regs that are live in the exit blocks. We don't collect scalars // because they won't be affected by lane predication. if (QPRs->contains(RegMask.PhysReg)) { if (auto *MI = RDA.getLocalLiveOutMIDef(Header, RegMask.PhysReg)) LiveOutMIs.insert(MI); RDA.getLiveInUses(ExitBB, RegMask.PhysReg, LiveOutUsers); } } // If we have any non-predicated live-outs, they need to be part of a // reduction that we can fixup later. The reduction that the form of an // operation that uses its previous values through a vmov and then a vpsel // resides in the exit blocks to select the final bytes from n and n-1 // iterations. if (!NonPredicated.empty() && !FindValidReduction(NonPredicated, LiveOutUsers)) return false; // We've already validated that any VPT predication within the loop will be // equivalent when we perform the predication transformation; so we know that // any VPT predicated instruction is predicated upon VCTP. Any live-out // instruction needs to be predicated, so check this here. The instructions // in NonPredicated have been found to be a reduction that we can ensure its // legality. for (auto *MI : LiveOutMIs) if (!isVectorPredicated(MI) && !NonPredicated.count(MI)) return false; return true; } void LowOverheadLoop::CheckLegality(ARMBasicBlockUtils *BBUtils) { if (Revert) return; if (!End->getOperand(1).isMBB()) report_fatal_error("Expected LoopEnd to target basic block"); // TODO Maybe there's cases where the target doesn't have to be the header, // but for now be safe and revert. if (End->getOperand(1).getMBB() != ML.getHeader()) { LLVM_DEBUG(dbgs() << "ARM Loops: LoopEnd is not targetting header.\n"); Revert = true; return; } // The WLS and LE instructions have 12-bits for the label offset. WLS // requires a positive offset, while LE uses negative. if (BBUtils->getOffsetOf(End) < BBUtils->getOffsetOf(ML.getHeader()) || !BBUtils->isBBInRange(End, ML.getHeader(), 4094)) { LLVM_DEBUG(dbgs() << "ARM Loops: LE offset is out-of-range\n"); Revert = true; return; } if (Start->getOpcode() == ARM::t2WhileLoopStart && (BBUtils->getOffsetOf(Start) > BBUtils->getOffsetOf(Start->getOperand(1).getMBB()) || !BBUtils->isBBInRange(Start, Start->getOperand(1).getMBB(), 4094))) { LLVM_DEBUG(dbgs() << "ARM Loops: WLS offset is out-of-range!\n"); Revert = true; return; } InsertPt = Revert ? nullptr : isSafeToDefineLR(); if (!InsertPt) { LLVM_DEBUG(dbgs() << "ARM Loops: Unable to find safe insertion point.\n"); Revert = true; return; } else LLVM_DEBUG(dbgs() << "ARM Loops: Start insertion point: " << *InsertPt); if (!IsTailPredicationLegal()) { LLVM_DEBUG(if (!VCTP) dbgs() << "ARM Loops: Didn't find a VCTP instruction.\n"; dbgs() << "ARM Loops: Tail-predication is not valid.\n"); return; } assert(ML.getBlocks().size() == 1 && "Shouldn't be processing a loop with more than one block"); CannotTailPredicate = !ValidateTailPredicate(InsertPt); LLVM_DEBUG(if (CannotTailPredicate) dbgs() << "ARM Loops: Couldn't validate tail predicate.\n"); } bool LowOverheadLoop::ValidateMVEInst(MachineInstr* MI) { if (CannotTailPredicate) return false; if (isVCTP(MI)) { // If we find another VCTP, check whether it uses the same value as the main VCTP. // If it does, store it in the SecondaryVCTPs set, else refuse it. if (VCTP) { if (!VCTP->getOperand(1).isIdenticalTo(MI->getOperand(1)) || !RDA.hasSameReachingDef(VCTP, MI, MI->getOperand(1).getReg())) { LLVM_DEBUG(dbgs() << "ARM Loops: Found VCTP with a different reaching " "definition from the main VCTP"); return false; } LLVM_DEBUG(dbgs() << "ARM Loops: Found secondary VCTP: " << *MI); SecondaryVCTPs.insert(MI); } else { LLVM_DEBUG(dbgs() << "ARM Loops: Found 'main' VCTP: " << *MI); VCTP = MI; } } else if (isVPTOpcode(MI->getOpcode())) { if (MI->getOpcode() != ARM::MVE_VPST) { assert(MI->findRegisterDefOperandIdx(ARM::VPR) != -1 && "VPT does not implicitly define VPR?!"); CurrentPredicate.insert(MI); } VPTBlocks.emplace_back(MI, CurrentPredicate); CurrentBlock = &VPTBlocks.back(); return true; } else if (MI->getOpcode() == ARM::MVE_VPSEL || MI->getOpcode() == ARM::MVE_VPNOT) { // TODO: Allow VPSEL and VPNOT, we currently cannot because: // 1) It will use the VPR as a predicate operand, but doesn't have to be // instead a VPT block, which means we can assert while building up // the VPT block because we don't find another VPT or VPST to being a new // one. // 2) VPSEL still requires a VPR operand even after tail predicating, // which means we can't remove it unless there is another // instruction, such as vcmp, that can provide the VPR def. return false; } bool IsUse = false; bool IsDef = false; const MCInstrDesc &MCID = MI->getDesc(); for (int i = MI->getNumOperands() - 1; i >= 0; --i) { const MachineOperand &MO = MI->getOperand(i); if (!MO.isReg() || MO.getReg() != ARM::VPR) continue; if (MO.isDef()) { CurrentPredicate.insert(MI); IsDef = true; } else if (ARM::isVpred(MCID.OpInfo[i].OperandType)) { CurrentBlock->addInst(MI, CurrentPredicate); IsUse = true; } else { LLVM_DEBUG(dbgs() << "ARM Loops: Found instruction using vpr: " << *MI); return false; } } // If we find a vpr def that is not already predicated on the vctp, we've // got disjoint predicates that may not be equivalent when we do the // conversion. if (IsDef && !IsUse && VCTP && !isVCTP(MI)) { LLVM_DEBUG(dbgs() << "ARM Loops: Found disjoint vpr def: " << *MI); return false; } uint64_t Flags = MCID.TSFlags; if ((Flags & ARMII::DomainMask) != ARMII::DomainMVE) return true; // If we find an instruction that has been marked as not valid for tail // predication, only allow the instruction if it's contained within a valid // VPT block. if ((Flags & ARMII::ValidForTailPredication) == 0 && !IsUse) { LLVM_DEBUG(dbgs() << "ARM Loops: Can't tail predicate: " << *MI); return false; } // If the instruction is already explicitly predicated, then the conversion // will be fine, but ensure that all memory operations are predicated. return !IsUse && MI->mayLoadOrStore() ? false : true; } bool ARMLowOverheadLoops::runOnMachineFunction(MachineFunction &mf) { const ARMSubtarget &ST = static_cast(mf.getSubtarget()); if (!ST.hasLOB()) return false; MF = &mf; LLVM_DEBUG(dbgs() << "ARM Loops on " << MF->getName() << " ------------- \n"); MLI = &getAnalysis(); RDA = &getAnalysis(); MF->getProperties().set(MachineFunctionProperties::Property::TracksLiveness); MRI = &MF->getRegInfo(); TII = static_cast(ST.getInstrInfo()); TRI = ST.getRegisterInfo(); BBUtils = std::unique_ptr(new ARMBasicBlockUtils(*MF)); BBUtils->computeAllBlockSizes(); BBUtils->adjustBBOffsetsAfter(&MF->front()); bool Changed = false; for (auto ML : *MLI) { if (!ML->getParentLoop()) Changed |= ProcessLoop(ML); } Changed |= RevertNonLoops(); return Changed; } bool ARMLowOverheadLoops::ProcessLoop(MachineLoop *ML) { bool Changed = false; // Process inner loops first. for (auto I = ML->begin(), E = ML->end(); I != E; ++I) Changed |= ProcessLoop(*I); LLVM_DEBUG(dbgs() << "ARM Loops: Processing loop containing:\n"; if (auto *Preheader = ML->getLoopPreheader()) dbgs() << " - " << Preheader->getName() << "\n"; else if (auto *Preheader = MLI->findLoopPreheader(ML)) dbgs() << " - " << Preheader->getName() << "\n"; else if (auto *Preheader = MLI->findLoopPreheader(ML, true)) dbgs() << " - " << Preheader->getName() << "\n"; for (auto *MBB : ML->getBlocks()) dbgs() << " - " << MBB->getName() << "\n"; ); // Search the given block for a loop start instruction. If one isn't found, // and there's only one predecessor block, search that one too. std::function SearchForStart = [&SearchForStart](MachineBasicBlock *MBB) -> MachineInstr* { for (auto &MI : *MBB) { if (isLoopStart(MI)) return &MI; } if (MBB->pred_size() == 1) return SearchForStart(*MBB->pred_begin()); return nullptr; }; LowOverheadLoop LoLoop(*ML, *MLI, *RDA, *TRI, *TII); // Search the preheader for the start intrinsic. // FIXME: I don't see why we shouldn't be supporting multiple predecessors // with potentially multiple set.loop.iterations, so we need to enable this. if (LoLoop.Preheader) LoLoop.Start = SearchForStart(LoLoop.Preheader); else return false; // Find the low-overhead loop components and decide whether or not to fall // back to a normal loop. Also look for a vctp instructions and decide // whether we can convert that predicate using tail predication. for (auto *MBB : reverse(ML->getBlocks())) { for (auto &MI : *MBB) { if (MI.isDebugValue()) continue; else if (MI.getOpcode() == ARM::t2LoopDec) LoLoop.Dec = &MI; else if (MI.getOpcode() == ARM::t2LoopEnd) LoLoop.End = &MI; else if (isLoopStart(MI)) LoLoop.Start = &MI; else if (MI.getDesc().isCall()) { // TODO: Though the call will require LE to execute again, does this // mean we should revert? Always executing LE hopefully should be // faster than performing a sub,cmp,br or even subs,br. LoLoop.Revert = true; LLVM_DEBUG(dbgs() << "ARM Loops: Found call.\n"); } else { // Record VPR defs and build up their corresponding vpt blocks. // Check we know how to tail predicate any mve instructions. LoLoop.AnalyseMVEInst(&MI); } } } LLVM_DEBUG(LoLoop.dump()); if (!LoLoop.FoundAllComponents()) { LLVM_DEBUG(dbgs() << "ARM Loops: Didn't find loop start, update, end\n"); return false; } // Check that the only instruction using LoopDec is LoopEnd. // TODO: Check for copy chains that really have no effect. SmallPtrSet Uses; RDA->getReachingLocalUses(LoLoop.Dec, ARM::LR, Uses); if (Uses.size() > 1 || !Uses.count(LoLoop.End)) { LLVM_DEBUG(dbgs() << "ARM Loops: Unable to remove LoopDec.\n"); LoLoop.Revert = true; } LoLoop.CheckLegality(BBUtils.get()); Expand(LoLoop); return true; } // WhileLoopStart holds the exit block, so produce a cmp lr, 0 and then a // beq that branches to the exit branch. // TODO: We could also try to generate a cbz if the value in LR is also in // another low register. void ARMLowOverheadLoops::RevertWhile(MachineInstr *MI) const { LLVM_DEBUG(dbgs() << "ARM Loops: Reverting to cmp: " << *MI); MachineBasicBlock *MBB = MI->getParent(); MachineInstrBuilder MIB = BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(ARM::t2CMPri)); MIB.add(MI->getOperand(0)); MIB.addImm(0); MIB.addImm(ARMCC::AL); MIB.addReg(ARM::NoRegister); MachineBasicBlock *DestBB = MI->getOperand(1).getMBB(); unsigned BrOpc = BBUtils->isBBInRange(MI, DestBB, 254) ? ARM::tBcc : ARM::t2Bcc; MIB = BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(BrOpc)); MIB.add(MI->getOperand(1)); // branch target MIB.addImm(ARMCC::EQ); // condition code MIB.addReg(ARM::CPSR); MI->eraseFromParent(); } bool ARMLowOverheadLoops::RevertLoopDec(MachineInstr *MI) const { LLVM_DEBUG(dbgs() << "ARM Loops: Reverting to sub: " << *MI); MachineBasicBlock *MBB = MI->getParent(); SmallPtrSet Ignore; for (auto I = MachineBasicBlock::iterator(MI), E = MBB->end(); I != E; ++I) { if (I->getOpcode() == ARM::t2LoopEnd) { Ignore.insert(&*I); break; } } // If nothing defines CPSR between LoopDec and LoopEnd, use a t2SUBS. bool SetFlags = RDA->isSafeToDefRegAt(MI, ARM::CPSR, Ignore); MachineInstrBuilder MIB = BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(ARM::t2SUBri)); MIB.addDef(ARM::LR); MIB.add(MI->getOperand(1)); MIB.add(MI->getOperand(2)); MIB.addImm(ARMCC::AL); MIB.addReg(0); if (SetFlags) { MIB.addReg(ARM::CPSR); MIB->getOperand(5).setIsDef(true); } else MIB.addReg(0); MI->eraseFromParent(); return SetFlags; } // Generate a subs, or sub and cmp, and a branch instead of an LE. void ARMLowOverheadLoops::RevertLoopEnd(MachineInstr *MI, bool SkipCmp) const { LLVM_DEBUG(dbgs() << "ARM Loops: Reverting to cmp, br: " << *MI); MachineBasicBlock *MBB = MI->getParent(); // Create cmp if (!SkipCmp) { MachineInstrBuilder MIB = BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(ARM::t2CMPri)); MIB.addReg(ARM::LR); MIB.addImm(0); MIB.addImm(ARMCC::AL); MIB.addReg(ARM::NoRegister); } MachineBasicBlock *DestBB = MI->getOperand(1).getMBB(); unsigned BrOpc = BBUtils->isBBInRange(MI, DestBB, 254) ? ARM::tBcc : ARM::t2Bcc; // Create bne MachineInstrBuilder MIB = BuildMI(*MBB, MI, MI->getDebugLoc(), TII->get(BrOpc)); MIB.add(MI->getOperand(1)); // branch target MIB.addImm(ARMCC::NE); // condition code MIB.addReg(ARM::CPSR); MI->eraseFromParent(); } // Perform dead code elimation on the loop iteration count setup expression. // If we are tail-predicating, the number of elements to be processed is the // operand of the VCTP instruction in the vector body, see getCount(), which is // register $r3 in this example: // // $lr = big-itercount-expression // .. // t2DoLoopStart renamable $lr // vector.body: // .. // $vpr = MVE_VCTP32 renamable $r3 // renamable $lr = t2LoopDec killed renamable $lr, 1 // t2LoopEnd renamable $lr, %vector.body // tB %end // // What we would like achieve here is to replace the do-loop start pseudo // instruction t2DoLoopStart with: // // $lr = MVE_DLSTP_32 killed renamable $r3 // // Thus, $r3 which defines the number of elements, is written to $lr, // and then we want to delete the whole chain that used to define $lr, // see the comment below how this chain could look like. // void ARMLowOverheadLoops::IterationCountDCE(LowOverheadLoop &LoLoop) { if (!LoLoop.IsTailPredicationLegal()) return; LLVM_DEBUG(dbgs() << "ARM Loops: Trying DCE on loop iteration count.\n"); MachineInstr *Def = RDA->getMIOperand(LoLoop.Start, 0); if (!Def) { LLVM_DEBUG(dbgs() << "ARM Loops: Couldn't find iteration count.\n"); return; } // Collect and remove the users of iteration count. SmallPtrSet Killed = { LoLoop.Start, LoLoop.Dec, LoLoop.End, LoLoop.InsertPt }; SmallPtrSet Remove; if (RDA->isSafeToRemove(Def, Remove, Killed)) LoLoop.ToRemove.insert(Remove.begin(), Remove.end()); else { LLVM_DEBUG(dbgs() << "ARM Loops: Unsafe to remove loop iteration count.\n"); return; } // Collect the dead code and the MBBs in which they reside. RDA->collectKilledOperands(Def, Killed); SmallPtrSet BasicBlocks; for (auto *MI : Killed) BasicBlocks.insert(MI->getParent()); // Collect IT blocks in all affected basic blocks. std::map> ITBlocks; for (auto *MBB : BasicBlocks) { for (auto &MI : *MBB) { if (MI.getOpcode() != ARM::t2IT) continue; RDA->getReachingLocalUses(&MI, ARM::ITSTATE, ITBlocks[&MI]); } } // If we're removing all of the instructions within an IT block, then // also remove the IT instruction. SmallPtrSet ModifiedITs; for (auto *MI : Killed) { if (MachineOperand *MO = MI->findRegisterUseOperand(ARM::ITSTATE)) { MachineInstr *IT = RDA->getMIOperand(MI, *MO); auto &CurrentBlock = ITBlocks[IT]; CurrentBlock.erase(MI); if (CurrentBlock.empty()) ModifiedITs.erase(IT); else ModifiedITs.insert(IT); } } // Delete the killed instructions only if we don't have any IT blocks that // need to be modified because we need to fixup the mask. // TODO: Handle cases where IT blocks are modified. if (ModifiedITs.empty()) { LLVM_DEBUG(dbgs() << "ARM Loops: Will remove iteration count:\n"; for (auto *MI : Killed) dbgs() << " - " << *MI); LoLoop.ToRemove.insert(Killed.begin(), Killed.end()); } else LLVM_DEBUG(dbgs() << "ARM Loops: Would need to modify IT block(s).\n"); } MachineInstr* ARMLowOverheadLoops::ExpandLoopStart(LowOverheadLoop &LoLoop) { LLVM_DEBUG(dbgs() << "ARM Loops: Expanding LoopStart.\n"); // When using tail-predication, try to delete the dead code that was used to // calculate the number of loop iterations. IterationCountDCE(LoLoop); MachineInstr *InsertPt = LoLoop.InsertPt; MachineInstr *Start = LoLoop.Start; MachineBasicBlock *MBB = InsertPt->getParent(); bool IsDo = Start->getOpcode() == ARM::t2DoLoopStart; unsigned Opc = LoLoop.getStartOpcode(); MachineOperand &Count = LoLoop.getCount(); MachineInstrBuilder MIB = BuildMI(*MBB, InsertPt, InsertPt->getDebugLoc(), TII->get(Opc)); MIB.addDef(ARM::LR); MIB.add(Count); if (!IsDo) MIB.add(Start->getOperand(1)); // If we're inserting at a mov lr, then remove it as it's redundant. if (InsertPt != Start) LoLoop.ToRemove.insert(InsertPt); LoLoop.ToRemove.insert(Start); LLVM_DEBUG(dbgs() << "ARM Loops: Inserted start: " << *MIB); return &*MIB; } void ARMLowOverheadLoops::FixupReductions(LowOverheadLoop &LoLoop) const { LLVM_DEBUG(dbgs() << "ARM Loops: Fixing up reduction(s).\n"); auto BuildMov = [this](MachineInstr &InsertPt, Register To, Register From) { MachineBasicBlock *MBB = InsertPt.getParent(); MachineInstrBuilder MIB = BuildMI(*MBB, &InsertPt, InsertPt.getDebugLoc(), TII->get(ARM::MVE_VORR)); MIB.addDef(To); MIB.addReg(From); MIB.addReg(From); MIB.addImm(0); MIB.addReg(0); MIB.addReg(To); LLVM_DEBUG(dbgs() << "ARM Loops: Inserted VMOV: " << *MIB); }; for (auto &Reduction : LoLoop.Reductions) { MachineInstr &Copy = Reduction->Copy; MachineInstr &Reduce = Reduction->Reduce; Register DestReg = Copy.getOperand(0).getReg(); // Change the initialiser if present if (Reduction->Init) { MachineInstr *Init = Reduction->Init; for (unsigned i = 0; i < Init->getNumOperands(); ++i) { MachineOperand &MO = Init->getOperand(i); if (MO.isReg() && MO.isUse() && MO.isTied() && Init->findTiedOperandIdx(i) == 0) Init->getOperand(i).setReg(DestReg); } Init->getOperand(0).setReg(DestReg); LLVM_DEBUG(dbgs() << "ARM Loops: Changed init regs: " << *Init); } else BuildMov(LoLoop.Preheader->instr_back(), DestReg, Copy.getOperand(1).getReg()); // Change the reducing op to write to the register that is used to copy // its value on the next iteration. Also update the tied-def operand. Reduce.getOperand(0).setReg(DestReg); Reduce.getOperand(5).setReg(DestReg); LLVM_DEBUG(dbgs() << "ARM Loops: Changed reduction regs: " << Reduce); // Instead of a vpsel, just copy the register into the necessary one. MachineInstr &VPSEL = Reduction->VPSEL; if (VPSEL.getOperand(0).getReg() != DestReg) BuildMov(VPSEL, VPSEL.getOperand(0).getReg(), DestReg); // Remove the unnecessary instructions. LLVM_DEBUG(dbgs() << "ARM Loops: Removing:\n" << " - " << Copy << " - " << VPSEL << "\n"); Copy.eraseFromParent(); VPSEL.eraseFromParent(); } } void ARMLowOverheadLoops::ConvertVPTBlocks(LowOverheadLoop &LoLoop) { auto RemovePredicate = [](MachineInstr *MI) { LLVM_DEBUG(dbgs() << "ARM Loops: Removing predicate from: " << *MI); if (int PIdx = llvm::findFirstVPTPredOperandIdx(*MI)) { assert(MI->getOperand(PIdx).getImm() == ARMVCC::Then && "Expected Then predicate!"); MI->getOperand(PIdx).setImm(ARMVCC::None); MI->getOperand(PIdx+1).setReg(0); } else llvm_unreachable("trying to unpredicate a non-predicated instruction"); }; // There are a few scenarios which we have to fix up: // 1. VPT Blocks with non-uniform predicates: // - a. When the divergent instruction is a vctp // - b. When the block uses a vpst, and is only predicated on the vctp // - c. When the block uses a vpt and (optionally) contains one or more // vctp. // 2. VPT Blocks with uniform predicates: // - a. The block uses a vpst, and is only predicated on the vctp for (auto &Block : LoLoop.getVPTBlocks()) { SmallVectorImpl &Insts = Block.getInsts(); if (Block.HasNonUniformPredicate()) { PredicatedMI *Divergent = Block.getDivergent(); if (isVCTP(Divergent->MI)) { // The vctp will be removed, so the block mask of the vp(s)t will need // to be recomputed. LoLoop.BlockMasksToRecompute.insert(Block.getPredicateThen()); } else if (Block.isVPST() && Block.IsOnlyPredicatedOn(LoLoop.VCTP)) { // The VPT block has a non-uniform predicate but it uses a vpst and its // entry is guarded only by a vctp, which means we: // - Need to remove the original vpst. // - Then need to unpredicate any following instructions, until // we come across the divergent vpr def. // - Insert a new vpst to predicate the instruction(s) that following // the divergent vpr def. // TODO: We could be producing more VPT blocks than necessary and could // fold the newly created one into a proceeding one. for (auto I = ++MachineBasicBlock::iterator(Block.getPredicateThen()), E = ++MachineBasicBlock::iterator(Divergent->MI); I != E; ++I) RemovePredicate(&*I); unsigned Size = 0; auto E = MachineBasicBlock::reverse_iterator(Divergent->MI); auto I = MachineBasicBlock::reverse_iterator(Insts.back().MI); MachineInstr *InsertAt = nullptr; while (I != E) { InsertAt = &*I; ++Size; ++I; } // Create a VPST (with a null mask for now, we'll recompute it later). MachineInstrBuilder MIB = BuildMI(*InsertAt->getParent(), InsertAt, InsertAt->getDebugLoc(), TII->get(ARM::MVE_VPST)); MIB.addImm(0); LLVM_DEBUG(dbgs() << "ARM Loops: Removing VPST: " << *Block.getPredicateThen()); LLVM_DEBUG(dbgs() << "ARM Loops: Created VPST: " << *MIB); LoLoop.ToRemove.insert(Block.getPredicateThen()); LoLoop.BlockMasksToRecompute.insert(MIB.getInstr()); } // Else, if the block uses a vpt, iterate over the block, removing the // extra VCTPs it may contain. else if (Block.isVPT()) { bool RemovedVCTP = false; for (PredicatedMI &Elt : Block.getInsts()) { MachineInstr *MI = Elt.MI; if (isVCTP(MI)) { LLVM_DEBUG(dbgs() << "ARM Loops: Removing VCTP: " << *MI); LoLoop.ToRemove.insert(MI); RemovedVCTP = true; continue; } } if (RemovedVCTP) LoLoop.BlockMasksToRecompute.insert(Block.getPredicateThen()); } } else if (Block.IsOnlyPredicatedOn(LoLoop.VCTP) && Block.isVPST()) { // A vpt block starting with VPST, is only predicated upon vctp and has no // internal vpr defs: // - Remove vpst. // - Unpredicate the remaining instructions. LLVM_DEBUG(dbgs() << "ARM Loops: Removing VPST: " << *Block.getPredicateThen()); LoLoop.ToRemove.insert(Block.getPredicateThen()); for (auto &PredMI : Insts) RemovePredicate(PredMI.MI); } } LLVM_DEBUG(dbgs() << "ARM Loops: Removing remaining VCTPs...\n"); // Remove the "main" VCTP LoLoop.ToRemove.insert(LoLoop.VCTP); LLVM_DEBUG(dbgs() << " " << *LoLoop.VCTP); // Remove remaining secondary VCTPs for (MachineInstr *VCTP : LoLoop.SecondaryVCTPs) { // All VCTPs that aren't marked for removal yet should be unpredicated ones. // The predicated ones should have already been marked for removal when // visiting the VPT blocks. if (LoLoop.ToRemove.insert(VCTP).second) { assert(getVPTInstrPredicate(*VCTP) == ARMVCC::None && "Removing Predicated VCTP without updating the block mask!"); LLVM_DEBUG(dbgs() << " " << *VCTP); } } } void ARMLowOverheadLoops::Expand(LowOverheadLoop &LoLoop) { // Combine the LoopDec and LoopEnd instructions into LE(TP). auto ExpandLoopEnd = [this](LowOverheadLoop &LoLoop) { MachineInstr *End = LoLoop.End; MachineBasicBlock *MBB = End->getParent(); unsigned Opc = LoLoop.IsTailPredicationLegal() ? ARM::MVE_LETP : ARM::t2LEUpdate; MachineInstrBuilder MIB = BuildMI(*MBB, End, End->getDebugLoc(), TII->get(Opc)); MIB.addDef(ARM::LR); MIB.add(End->getOperand(0)); MIB.add(End->getOperand(1)); LLVM_DEBUG(dbgs() << "ARM Loops: Inserted LE: " << *MIB); LoLoop.ToRemove.insert(LoLoop.Dec); LoLoop.ToRemove.insert(End); return &*MIB; }; // TODO: We should be able to automatically remove these branches before we // get here - probably by teaching analyzeBranch about the pseudo // instructions. // If there is an unconditional branch, after I, that just branches to the // next block, remove it. auto RemoveDeadBranch = [](MachineInstr *I) { MachineBasicBlock *BB = I->getParent(); MachineInstr *Terminator = &BB->instr_back(); if (Terminator->isUnconditionalBranch() && I != Terminator) { MachineBasicBlock *Succ = Terminator->getOperand(0).getMBB(); if (BB->isLayoutSuccessor(Succ)) { LLVM_DEBUG(dbgs() << "ARM Loops: Removing branch: " << *Terminator); Terminator->eraseFromParent(); } } }; if (LoLoop.Revert) { if (LoLoop.Start->getOpcode() == ARM::t2WhileLoopStart) RevertWhile(LoLoop.Start); else LoLoop.Start->eraseFromParent(); bool FlagsAlreadySet = RevertLoopDec(LoLoop.Dec); RevertLoopEnd(LoLoop.End, FlagsAlreadySet); } else { LoLoop.Start = ExpandLoopStart(LoLoop); RemoveDeadBranch(LoLoop.Start); LoLoop.End = ExpandLoopEnd(LoLoop); RemoveDeadBranch(LoLoop.End); if (LoLoop.IsTailPredicationLegal()) { ConvertVPTBlocks(LoLoop); FixupReductions(LoLoop); } for (auto *I : LoLoop.ToRemove) { LLVM_DEBUG(dbgs() << "ARM Loops: Erasing " << *I); I->eraseFromParent(); } for (auto *I : LoLoop.BlockMasksToRecompute) { LLVM_DEBUG(dbgs() << "ARM Loops: Recomputing VPT/VPST Block Mask: " << *I); recomputeVPTBlockMask(*I); LLVM_DEBUG(dbgs() << " ... done: " << *I); } } PostOrderLoopTraversal DFS(LoLoop.ML, *MLI); DFS.ProcessLoop(); const SmallVectorImpl &PostOrder = DFS.getOrder(); for (auto *MBB : PostOrder) { recomputeLiveIns(*MBB); // FIXME: For some reason, the live-in print order is non-deterministic for // our tests and I can't out why... So just sort them. MBB->sortUniqueLiveIns(); } for (auto *MBB : reverse(PostOrder)) recomputeLivenessFlags(*MBB); // We've moved, removed and inserted new instructions, so update RDA. RDA->reset(); } bool ARMLowOverheadLoops::RevertNonLoops() { LLVM_DEBUG(dbgs() << "ARM Loops: Reverting any remaining pseudos...\n"); bool Changed = false; for (auto &MBB : *MF) { SmallVector Starts; SmallVector Decs; SmallVector Ends; for (auto &I : MBB) { if (isLoopStart(I)) Starts.push_back(&I); else if (I.getOpcode() == ARM::t2LoopDec) Decs.push_back(&I); else if (I.getOpcode() == ARM::t2LoopEnd) Ends.push_back(&I); } if (Starts.empty() && Decs.empty() && Ends.empty()) continue; Changed = true; for (auto *Start : Starts) { if (Start->getOpcode() == ARM::t2WhileLoopStart) RevertWhile(Start); else Start->eraseFromParent(); } for (auto *Dec : Decs) RevertLoopDec(Dec); for (auto *End : Ends) RevertLoopEnd(End); } return Changed; } FunctionPass *llvm::createARMLowOverheadLoopsPass() { return new ARMLowOverheadLoops(); }