//===-- MVEVPTBlockPass.cpp - Insert MVE VPT blocks -----------------------===// // // 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 // //===----------------------------------------------------------------------===// #include "ARM.h" #include "ARMMachineFunctionInfo.h" #include "ARMSubtarget.h" #include "MCTargetDesc/ARMBaseInfo.h" #include "Thumb2InstrInfo.h" #include "llvm/ADT/SmallSet.h" #include "llvm/ADT/SmallVector.h" #include "llvm/ADT/Statistic.h" #include "llvm/ADT/StringRef.h" #include "llvm/CodeGen/MachineBasicBlock.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/MachineOperand.h" #include "llvm/IR/DebugLoc.h" #include "llvm/MC/MCInstrDesc.h" #include "llvm/MC/MCRegisterInfo.h" #include "llvm/Support/Debug.h" #include #include using namespace llvm; #define DEBUG_TYPE "arm-mve-vpt" namespace { class MVEVPTBlock : public MachineFunctionPass { public: static char ID; const Thumb2InstrInfo *TII; const TargetRegisterInfo *TRI; MVEVPTBlock() : MachineFunctionPass(ID) {} bool runOnMachineFunction(MachineFunction &Fn) override; MachineFunctionProperties getRequiredProperties() const override { return MachineFunctionProperties().set( MachineFunctionProperties::Property::NoVRegs); } StringRef getPassName() const override { return "MVE VPT block insertion pass"; } private: bool InsertVPTBlocks(MachineBasicBlock &MBB); }; char MVEVPTBlock::ID = 0; } // end anonymous namespace INITIALIZE_PASS(MVEVPTBlock, DEBUG_TYPE, "ARM MVE VPT block pass", false, false) static MachineInstr *findVCMPToFoldIntoVPST(MachineBasicBlock::iterator MI, const TargetRegisterInfo *TRI, unsigned &NewOpcode) { // Search backwards to the instruction that defines VPR. This may or not // be a VCMP, we check that after this loop. If we find another instruction // that reads cpsr, we return nullptr. MachineBasicBlock::iterator CmpMI = MI; while (CmpMI != MI->getParent()->begin()) { --CmpMI; if (CmpMI->modifiesRegister(ARM::VPR, TRI)) break; if (CmpMI->readsRegister(ARM::VPR, TRI)) break; } if (CmpMI == MI) return nullptr; NewOpcode = VCMPOpcodeToVPT(CmpMI->getOpcode()); if (NewOpcode == 0) return nullptr; // Search forward from CmpMI to MI, checking if either register was def'd if (registerDefinedBetween(CmpMI->getOperand(1).getReg(), std::next(CmpMI), MI, TRI)) return nullptr; if (registerDefinedBetween(CmpMI->getOperand(2).getReg(), std::next(CmpMI), MI, TRI)) return nullptr; return &*CmpMI; } // Advances Iter past a block of predicated instructions. // Returns true if it successfully skipped the whole block of predicated // instructions. Returns false when it stopped early (due to MaxSteps), or if // Iter didn't point to a predicated instruction. static bool StepOverPredicatedInstrs(MachineBasicBlock::instr_iterator &Iter, MachineBasicBlock::instr_iterator EndIter, unsigned MaxSteps, unsigned &NumInstrsSteppedOver) { ARMVCC::VPTCodes NextPred = ARMVCC::None; Register PredReg; NumInstrsSteppedOver = 0; while (Iter != EndIter) { NextPred = getVPTInstrPredicate(*Iter, PredReg); assert(NextPred != ARMVCC::Else && "VPT block pass does not expect Else preds"); if (NextPred == ARMVCC::None || MaxSteps == 0) break; --MaxSteps; ++Iter; ++NumInstrsSteppedOver; }; return NumInstrsSteppedOver != 0 && (NextPred == ARMVCC::None || Iter == EndIter); } // Returns true if at least one instruction in the range [Iter, End) defines // or kills VPR. static bool IsVPRDefinedOrKilledByBlock(MachineBasicBlock::iterator Iter, MachineBasicBlock::iterator End) { for (; Iter != End; ++Iter) if (Iter->definesRegister(ARM::VPR) || Iter->killsRegister(ARM::VPR)) return true; return false; } // Creates a T, TT, TTT or TTTT BlockMask depending on BlockSize. static ARM::PredBlockMask GetInitialBlockMask(unsigned BlockSize) { switch (BlockSize) { case 1: return ARM::PredBlockMask::T; case 2: return ARM::PredBlockMask::TT; case 3: return ARM::PredBlockMask::TTT; case 4: return ARM::PredBlockMask::TTTT; default: llvm_unreachable("Invalid BlockSize!"); } } // Given an iterator (Iter) that points at an instruction with a "Then" // predicate, tries to create the largest block of continuous predicated // instructions possible, and returns the VPT Block Mask of that block. // // This will try to perform some minor optimization in order to maximize the // size of the block. static ARM::PredBlockMask CreateVPTBlock(MachineBasicBlock::instr_iterator &Iter, MachineBasicBlock::instr_iterator EndIter, SmallVectorImpl &DeadInstructions) { MachineBasicBlock::instr_iterator BlockBeg = Iter; (void)BlockBeg; assert(getVPTInstrPredicate(*Iter) == ARMVCC::Then && "Expected a Predicated Instruction"); LLVM_DEBUG(dbgs() << "VPT block created for: "; Iter->dump()); unsigned BlockSize; StepOverPredicatedInstrs(Iter, EndIter, 4, BlockSize); LLVM_DEBUG(for (MachineBasicBlock::instr_iterator AddedInstIter = std::next(BlockBeg); AddedInstIter != Iter; ++AddedInstIter) { dbgs() << " adding: "; AddedInstIter->dump(); }); // Generate the initial BlockMask ARM::PredBlockMask BlockMask = GetInitialBlockMask(BlockSize); // Remove VPNOTs while there's still room in the block, so we can make the // largest block possible. ARMVCC::VPTCodes CurrentPredicate = ARMVCC::Else; while (BlockSize < 4 && Iter != EndIter && Iter->getOpcode() == ARM::MVE_VPNOT) { // Try to skip all of the predicated instructions after the VPNOT, stopping // after (4 - BlockSize). If we can't skip them all, stop. unsigned ElseInstCnt = 0; MachineBasicBlock::instr_iterator VPNOTBlockEndIter = std::next(Iter); if (!StepOverPredicatedInstrs(VPNOTBlockEndIter, EndIter, (4 - BlockSize), ElseInstCnt)) break; // Check if this VPNOT can be removed or not: It can only be removed if at // least one of the predicated instruction that follows it kills or sets // VPR. if (!IsVPRDefinedOrKilledByBlock(Iter, VPNOTBlockEndIter)) break; LLVM_DEBUG(dbgs() << " removing VPNOT: "; Iter->dump();); // Record the new size of the block BlockSize += ElseInstCnt; assert(BlockSize <= 4 && "Block is too large!"); // Record the VPNot to remove it later. DeadInstructions.push_back(&*Iter); ++Iter; // Replace the predicates of the instructions we're adding. // Note that we are using "Iter" to iterate over the block so we can update // it at the same time. for (; Iter != VPNOTBlockEndIter; ++Iter) { // Find the register in which the predicate is int OpIdx = findFirstVPTPredOperandIdx(*Iter); assert(OpIdx != -1); // Change the predicate and update the mask Iter->getOperand(OpIdx).setImm(CurrentPredicate); BlockMask = expandPredBlockMask(BlockMask, CurrentPredicate); LLVM_DEBUG(dbgs() << " adding : "; Iter->dump()); } CurrentPredicate = (CurrentPredicate == ARMVCC::Then ? ARMVCC::Else : ARMVCC::Then); } return BlockMask; } bool MVEVPTBlock::InsertVPTBlocks(MachineBasicBlock &Block) { bool Modified = false; MachineBasicBlock::instr_iterator MBIter = Block.instr_begin(); MachineBasicBlock::instr_iterator EndIter = Block.instr_end(); SmallVector DeadInstructions; while (MBIter != EndIter) { MachineInstr *MI = &*MBIter; Register PredReg; DebugLoc DL = MI->getDebugLoc(); ARMVCC::VPTCodes Pred = getVPTInstrPredicate(*MI, PredReg); // The idea of the predicate is that None, Then and Else are for use when // handling assembly language: they correspond to the three possible // suffixes "", "t" and "e" on the mnemonic. So when instructions are read // from assembly source or disassembled from object code, you expect to // see a mixture whenever there's a long VPT block. But in code // generation, we hope we'll never generate an Else as input to this pass. assert(Pred != ARMVCC::Else && "VPT block pass does not expect Else preds"); if (Pred == ARMVCC::None) { ++MBIter; continue; } ARM::PredBlockMask BlockMask = CreateVPTBlock(MBIter, EndIter, DeadInstructions); // Search back for a VCMP that can be folded to create a VPT, or else // create a VPST directly MachineInstrBuilder MIBuilder; unsigned NewOpcode; LLVM_DEBUG(dbgs() << " final block mask: " << (unsigned)BlockMask << "\n"); if (MachineInstr *VCMP = findVCMPToFoldIntoVPST(MI, TRI, NewOpcode)) { LLVM_DEBUG(dbgs() << " folding VCMP into VPST: "; VCMP->dump()); MIBuilder = BuildMI(Block, MI, DL, TII->get(NewOpcode)); MIBuilder.addImm((uint64_t)BlockMask); MIBuilder.add(VCMP->getOperand(1)); MIBuilder.add(VCMP->getOperand(2)); MIBuilder.add(VCMP->getOperand(3)); // We need to remove any kill flags between the original VCMP and the new // insertion point. for (MachineInstr &MII : make_range(VCMP->getIterator(), MI->getIterator())) { MII.clearRegisterKills(VCMP->getOperand(1).getReg(), TRI); MII.clearRegisterKills(VCMP->getOperand(2).getReg(), TRI); } VCMP->eraseFromParent(); } else { MIBuilder = BuildMI(Block, MI, DL, TII->get(ARM::MVE_VPST)); MIBuilder.addImm((uint64_t)BlockMask); } // Erase all dead instructions (VPNOT's). Do that now so that they do not // mess with the bundle creation. for (MachineInstr *DeadMI : DeadInstructions) DeadMI->eraseFromParent(); DeadInstructions.clear(); finalizeBundle( Block, MachineBasicBlock::instr_iterator(MIBuilder.getInstr()), MBIter); Modified = true; } return Modified; } bool MVEVPTBlock::runOnMachineFunction(MachineFunction &Fn) { const ARMSubtarget &STI = static_cast(Fn.getSubtarget()); if (!STI.isThumb2() || !STI.hasMVEIntegerOps()) return false; TII = static_cast(STI.getInstrInfo()); TRI = STI.getRegisterInfo(); LLVM_DEBUG(dbgs() << "********** ARM MVE VPT BLOCKS **********\n" << "********** Function: " << Fn.getName() << '\n'); bool Modified = false; for (MachineBasicBlock &MBB : Fn) Modified |= InsertVPTBlocks(MBB); LLVM_DEBUG(dbgs() << "**************************************\n"); return Modified; } /// createMVEVPTBlock - Returns an instance of the MVE VPT block /// insertion pass. FunctionPass *llvm::createMVEVPTBlockPass() { return new MVEVPTBlock(); }