//===----- HexagonNewValueJump.cpp - Hexagon Backend New Value Jump -------===// // // The LLVM Compiler Infrastructure // // This file is distributed under the University of Illinois Open Source // License. See LICENSE.TXT for details. // //===----------------------------------------------------------------------===// // // This implements NewValueJump pass in Hexagon. // Ideally, we should merge this as a Peephole pass prior to register // allocation, but because we have a spill in between the feeder and new value // jump instructions, we are forced to write after register allocation. // Having said that, we should re-attempt to pull this earlier at some point // in future. // The basic approach looks for sequence of predicated jump, compare instruciton // that genereates the predicate and, the feeder to the predicate. Once it finds // all, it collapses compare and jump instruction into a new valu jump // intstructions. // // //===----------------------------------------------------------------------===// #define DEBUG_TYPE "hexagon-nvj" #include "Hexagon.h" #include "HexagonInstrInfo.h" #include "HexagonMachineFunctionInfo.h" #include "HexagonRegisterInfo.h" #include "HexagonSubtarget.h" #include "HexagonTargetMachine.h" #include "llvm/ADT/DenseMap.h" #include "llvm/ADT/Statistic.h" #include "llvm/CodeGen/LiveVariables.h" #include "llvm/CodeGen/MachineFunctionAnalysis.h" #include "llvm/CodeGen/MachineFunctionPass.h" #include "llvm/CodeGen/MachineInstrBuilder.h" #include "llvm/CodeGen/MachineRegisterInfo.h" #include "llvm/CodeGen/Passes.h" #include "llvm/CodeGen/ScheduleDAGInstrs.h" #include "llvm/PassSupport.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Compiler.h" #include "llvm/Support/Debug.h" #include "llvm/Target/TargetInstrInfo.h" #include "llvm/Target/TargetMachine.h" #include "llvm/Target/TargetRegisterInfo.h" #include using namespace llvm; STATISTIC(NumNVJGenerated, "Number of New Value Jump Instructions created"); static cl::opt DbgNVJCount("nvj-count", cl::init(-1), cl::Hidden, cl::desc( "Maximum number of predicated jumps to be converted to New Value Jump")); static cl::opt DisableNewValueJumps("disable-nvjump", cl::Hidden, cl::ZeroOrMore, cl::init(false), cl::desc("Disable New Value Jumps")); namespace { struct HexagonNewValueJump : public MachineFunctionPass { const HexagonInstrInfo *QII; const HexagonRegisterInfo *QRI; public: static char ID; HexagonNewValueJump() : MachineFunctionPass(ID) { } virtual void getAnalysisUsage(AnalysisUsage &AU) const { MachineFunctionPass::getAnalysisUsage(AU); } const char *getPassName() const { return "Hexagon NewValueJump"; } virtual bool runOnMachineFunction(MachineFunction &Fn); private: }; } // end of anonymous namespace char HexagonNewValueJump::ID = 0; // We have identified this II could be feeder to NVJ, // verify that it can be. static bool canBeFeederToNewValueJump(const HexagonInstrInfo *QII, const TargetRegisterInfo *TRI, MachineBasicBlock::iterator II, MachineBasicBlock::iterator end, MachineBasicBlock::iterator skip, MachineFunction &MF) { // Predicated instruction can not be feeder to NVJ. if (QII->isPredicated(II)) return false; // Bail out if feederReg is a paired register (double regs in // our case). One would think that we can check to see if a given // register cmpReg1 or cmpReg2 is a sub register of feederReg // using -- if (QRI->isSubRegister(feederReg, cmpReg1) logic // before the callsite of this function // But we can not as it comes in the following fashion. // %D0 = Hexagon_S2_lsr_r_p %D0, %R2 // %R0 = KILL %R0, %D0 // %P0 = CMPEQri %R0, 0 // Hence, we need to check if it's a KILL instruction. if (II->getOpcode() == TargetOpcode::KILL) return false; // Make sure there there is no 'def' or 'use' of any of the uses of // feeder insn between it's definition, this MI and jump, jmpInst // skipping compare, cmpInst. // Here's the example. // r21=memub(r22+r24<<#0) // p0 = cmp.eq(r21, #0) // r4=memub(r3+r21<<#0) // if (p0.new) jump:t .LBB29_45 // Without this check, it will be converted into // r4=memub(r3+r21<<#0) // r21=memub(r22+r24<<#0) // p0 = cmp.eq(r21, #0) // if (p0.new) jump:t .LBB29_45 // and result WAR hazards if converted to New Value Jump. for (unsigned i = 0; i < II->getNumOperands(); ++i) { if (II->getOperand(i).isReg() && (II->getOperand(i).isUse() || II->getOperand(i).isDef())) { MachineBasicBlock::iterator localII = II; ++localII; unsigned Reg = II->getOperand(i).getReg(); for (MachineBasicBlock::iterator localBegin = localII; localBegin != end; ++localBegin) { if (localBegin == skip ) continue; // Check for Subregisters too. if (localBegin->modifiesRegister(Reg, TRI) || localBegin->readsRegister(Reg, TRI)) return false; } } } return true; } // These are the common checks that need to performed // to determine if // 1. compare instruction can be moved before jump. // 2. feeder to the compare instruction can be moved before jump. static bool commonChecksToProhibitNewValueJump(bool afterRA, MachineBasicBlock::iterator MII) { // If store in path, bail out. if (MII->getDesc().mayStore()) return false; // if call in path, bail out. if (MII->getOpcode() == Hexagon::CALLv3) return false; // if NVJ is running prior to RA, do the following checks. if (!afterRA) { // The following Target Opcode instructions are spurious // to new value jump. If they are in the path, bail out. // KILL sets kill flag on the opcode. It also sets up a // single register, out of pair. // %D0 = Hexagon_S2_lsr_r_p %D0, %R2 // %R0 = KILL %R0, %D0 // %P0 = CMPEQri %R0, 0 // PHI can be anything after RA. // COPY can remateriaze things in between feeder, compare and nvj. if (MII->getOpcode() == TargetOpcode::KILL || MII->getOpcode() == TargetOpcode::PHI || MII->getOpcode() == TargetOpcode::COPY) return false; // The following pseudo Hexagon instructions sets "use" and "def" // of registers by individual passes in the backend. At this time, // we don't know the scope of usage and definitions of these // instructions. if (MII->getOpcode() == Hexagon::TFR_condset_rr || MII->getOpcode() == Hexagon::TFR_condset_ii || MII->getOpcode() == Hexagon::TFR_condset_ri || MII->getOpcode() == Hexagon::TFR_condset_ir || MII->getOpcode() == Hexagon::LDriw_pred || MII->getOpcode() == Hexagon::STriw_pred) return false; } return true; } static bool canCompareBeNewValueJump(const HexagonInstrInfo *QII, const TargetRegisterInfo *TRI, MachineBasicBlock::iterator II, unsigned pReg, bool secondReg, bool optLocation, MachineBasicBlock::iterator end, MachineFunction &MF) { MachineInstr *MI = II; // If the second operand of the compare is an imm, make sure it's in the // range specified by the arch. if (!secondReg) { int64_t v = MI->getOperand(2).getImm(); if (MI->getOpcode() == Hexagon::CMPGEri || (MI->getOpcode() == Hexagon::CMPGEUri && v > 0)) --v; if (!(isUInt<5>(v) || ((MI->getOpcode() == Hexagon::CMPEQri || MI->getOpcode() == Hexagon::CMPGTri || MI->getOpcode() == Hexagon::CMPGEri) && (v == -1)))) return false; } unsigned cmpReg1, cmpOp2 = 0; // cmpOp2 assignment silences compiler warning. cmpReg1 = MI->getOperand(1).getReg(); if (secondReg) { cmpOp2 = MI->getOperand(2).getReg(); // Make sure that that second register is not from COPY // At machine code level, we don't need this, but if we decide // to move new value jump prior to RA, we would be needing this. MachineRegisterInfo &MRI = MF.getRegInfo(); if (secondReg && !TargetRegisterInfo::isPhysicalRegister(cmpOp2)) { MachineInstr *def = MRI.getVRegDef(cmpOp2); if (def->getOpcode() == TargetOpcode::COPY) return false; } } // Walk the instructions after the compare (predicate def) to the jump, // and satisfy the following conditions. ++II ; for (MachineBasicBlock::iterator localII = II; localII != end; ++localII) { // Check 1. // If "common" checks fail, bail out. if (!commonChecksToProhibitNewValueJump(optLocation, localII)) return false; // Check 2. // If there is a def or use of predicate (result of compare), bail out. if (localII->modifiesRegister(pReg, TRI) || localII->readsRegister(pReg, TRI)) return false; // Check 3. // If there is a def of any of the use of the compare (operands of compare), // bail out. // Eg. // p0 = cmp.eq(r2, r0) // r2 = r4 // if (p0.new) jump:t .LBB28_3 if (localII->modifiesRegister(cmpReg1, TRI) || (secondReg && localII->modifiesRegister(cmpOp2, TRI))) return false; } return true; } // Given a compare operator, return a matching New Value Jump // compare operator. Make sure that MI here is included in // HexagonInstrInfo.cpp::isNewValueJumpCandidate static unsigned getNewValueJumpOpcode(const MachineInstr *MI, int reg, bool secondRegNewified) { switch (MI->getOpcode()) { case Hexagon::CMPEQrr: return Hexagon::JMP_EQrrPt_nv_V4; case Hexagon::CMPEQri: { if (reg >= 0) return Hexagon::JMP_EQriPt_nv_V4; else return Hexagon::JMP_EQriPtneg_nv_V4; } case Hexagon::CMPLTrr: case Hexagon::CMPGTrr: { if (secondRegNewified) return Hexagon::JMP_GTrrdnPt_nv_V4; else return Hexagon::JMP_GTrrPt_nv_V4; } case Hexagon::CMPGEri: { if (reg >= 1) return Hexagon::JMP_GTriPt_nv_V4; else return Hexagon::JMP_GTriPtneg_nv_V4; } case Hexagon::CMPGTri: { if (reg >= 0) return Hexagon::JMP_GTriPt_nv_V4; else return Hexagon::JMP_GTriPtneg_nv_V4; } case Hexagon::CMPLTUrr: case Hexagon::CMPGTUrr: { if (secondRegNewified) return Hexagon::JMP_GTUrrdnPt_nv_V4; else return Hexagon::JMP_GTUrrPt_nv_V4; } case Hexagon::CMPGTUri: return Hexagon::JMP_GTUriPt_nv_V4; case Hexagon::CMPGEUri: { if (reg == 0) return Hexagon::JMP_EQrrPt_nv_V4; else return Hexagon::JMP_GTUriPt_nv_V4; } default: llvm_unreachable("Could not find matching New Value Jump instruction."); } // return *some value* to avoid compiler warning return 0; } bool HexagonNewValueJump::runOnMachineFunction(MachineFunction &MF) { DEBUG(dbgs() << "********** Hexagon New Value Jump **********\n" << "********** Function: " << MF.getName() << "\n"); #if 0 // for now disable this, if we move NewValueJump before register // allocation we need this information. LiveVariables &LVs = getAnalysis(); #endif QII = static_cast(MF.getTarget().getInstrInfo()); QRI = static_cast(MF.getTarget().getRegisterInfo()); if (!QRI->Subtarget.hasV4TOps() || DisableNewValueJumps) { return false; } int nvjCount = DbgNVJCount; int nvjGenerated = 0; // Loop through all the bb's of the function for (MachineFunction::iterator MBBb = MF.begin(), MBBe = MF.end(); MBBb != MBBe; ++MBBb) { MachineBasicBlock* MBB = MBBb; DEBUG(dbgs() << "** dumping bb ** " << MBB->getNumber() << "\n"); DEBUG(MBB->dump()); DEBUG(dbgs() << "\n" << "********** dumping instr bottom up **********\n"); bool foundJump = false; bool foundCompare = false; bool invertPredicate = false; unsigned predReg = 0; // predicate reg of the jump. unsigned cmpReg1 = 0; int cmpOp2 = 0; bool MO1IsKill = false; bool MO2IsKill = false; MachineBasicBlock::iterator jmpPos; MachineBasicBlock::iterator cmpPos; MachineInstr *cmpInstr = NULL, *jmpInstr = NULL; MachineBasicBlock *jmpTarget = NULL; bool afterRA = false; bool isSecondOpReg = false; bool isSecondOpNewified = false; // Traverse the basic block - bottom up for (MachineBasicBlock::iterator MII = MBB->end(), E = MBB->begin(); MII != E;) { MachineInstr *MI = --MII; if (MI->isDebugValue()) { continue; } if ((nvjCount == 0) || (nvjCount > -1 && nvjCount <= nvjGenerated)) break; DEBUG(dbgs() << "Instr: "; MI->dump(); dbgs() << "\n"); if (!foundJump && (MI->getOpcode() == Hexagon::JMP_c || MI->getOpcode() == Hexagon::JMP_cNot || MI->getOpcode() == Hexagon::JMP_cdnPt || MI->getOpcode() == Hexagon::JMP_cdnPnt || MI->getOpcode() == Hexagon::JMP_cdnNotPt || MI->getOpcode() == Hexagon::JMP_cdnNotPnt)) { // This is where you would insert your compare and // instr that feeds compare jmpPos = MII; jmpInstr = MI; predReg = MI->getOperand(0).getReg(); afterRA = TargetRegisterInfo::isPhysicalRegister(predReg); // If ifconverter had not messed up with the kill flags of the // operands, the following check on the kill flag would suffice. // if(!jmpInstr->getOperand(0).isKill()) break; // This predicate register is live out out of BB // this would only work if we can actually use Live // variable analysis on phy regs - but LLVM does not // provide LV analysis on phys regs. //if(LVs.isLiveOut(predReg, *MBB)) break; // Get all the successors of this block - which will always // be 2. Check if the predicate register is live in in those // successor. If yes, we can not delete the predicate - // I am doing this only because LLVM does not provide LiveOut // at the BB level. bool predLive = false; for (MachineBasicBlock::const_succ_iterator SI = MBB->succ_begin(), SIE = MBB->succ_end(); SI != SIE; ++SI) { MachineBasicBlock* succMBB = *SI; if (succMBB->isLiveIn(predReg)) { predLive = true; } } if (predLive) break; jmpTarget = MI->getOperand(1).getMBB(); foundJump = true; if (MI->getOpcode() == Hexagon::JMP_cNot || MI->getOpcode() == Hexagon::JMP_cdnNotPt || MI->getOpcode() == Hexagon::JMP_cdnNotPnt) { invertPredicate = true; } continue; } // No new value jump if there is a barrier. A barrier has to be in its // own packet. A barrier has zero operands. We conservatively bail out // here if we see any instruction with zero operands. if (foundJump && MI->getNumOperands() == 0) break; if (foundJump && !foundCompare && MI->getOperand(0).isReg() && MI->getOperand(0).getReg() == predReg) { // Not all compares can be new value compare. Arch Spec: 7.6.1.1 if (QII->isNewValueJumpCandidate(MI)) { assert((MI->getDesc().isCompare()) && "Only compare instruction can be collapsed into New Value Jump"); isSecondOpReg = MI->getOperand(2).isReg(); if (!canCompareBeNewValueJump(QII, QRI, MII, predReg, isSecondOpReg, afterRA, jmpPos, MF)) break; cmpInstr = MI; cmpPos = MII; foundCompare = true; // We need cmpReg1 and cmpOp2(imm or reg) while building // new value jump instruction. cmpReg1 = MI->getOperand(1).getReg(); if (MI->getOperand(1).isKill()) MO1IsKill = true; if (isSecondOpReg) { cmpOp2 = MI->getOperand(2).getReg(); if (MI->getOperand(2).isKill()) MO2IsKill = true; } else cmpOp2 = MI->getOperand(2).getImm(); continue; } } if (foundCompare && foundJump) { // If "common" checks fail, bail out on this BB. if (!commonChecksToProhibitNewValueJump(afterRA, MII)) break; bool foundFeeder = false; MachineBasicBlock::iterator feederPos = MII; if (MI->getOperand(0).isReg() && MI->getOperand(0).isDef() && (MI->getOperand(0).getReg() == cmpReg1 || (isSecondOpReg && MI->getOperand(0).getReg() == (unsigned) cmpOp2))) { unsigned feederReg = MI->getOperand(0).getReg(); // First try to see if we can get the feeder from the first operand // of the compare. If we can not, and if secondOpReg is true // (second operand of the compare is also register), try that one. // TODO: Try to come up with some heuristic to figure out which // feeder would benefit. if (feederReg == cmpReg1) { if (!canBeFeederToNewValueJump(QII, QRI, MII, jmpPos, cmpPos, MF)) { if (!isSecondOpReg) break; else continue; } else foundFeeder = true; } if (!foundFeeder && isSecondOpReg && feederReg == (unsigned) cmpOp2) if (!canBeFeederToNewValueJump(QII, QRI, MII, jmpPos, cmpPos, MF)) break; if (isSecondOpReg) { // In case of CMPLT, or CMPLTU, or EQ with the second register // to newify, swap the operands. if (cmpInstr->getOpcode() == Hexagon::CMPLTrr || cmpInstr->getOpcode() == Hexagon::CMPLTUrr || (cmpInstr->getOpcode() == Hexagon::CMPEQrr && feederReg == (unsigned) cmpOp2)) { unsigned tmp = cmpReg1; bool tmpIsKill = MO1IsKill; cmpReg1 = cmpOp2; MO1IsKill = MO2IsKill; cmpOp2 = tmp; MO2IsKill = tmpIsKill; } // Now we have swapped the operands, all we need to check is, // if the second operand (after swap) is the feeder. // And if it is, make a note. if (feederReg == (unsigned)cmpOp2) isSecondOpNewified = true; } // Now that we are moving feeder close the jump, // make sure we are respecting the kill values of // the operands of the feeder. bool updatedIsKill = false; for (unsigned i = 0; i < MI->getNumOperands(); i++) { MachineOperand &MO = MI->getOperand(i); if (MO.isReg() && MO.isUse()) { unsigned feederReg = MO.getReg(); for (MachineBasicBlock::iterator localII = feederPos, end = jmpPos; localII != end; localII++) { MachineInstr *localMI = localII; for (unsigned j = 0; j < localMI->getNumOperands(); j++) { MachineOperand &localMO = localMI->getOperand(j); if (localMO.isReg() && localMO.isUse() && localMO.isKill() && feederReg == localMO.getReg()) { // We found that there is kill of a use register // Set up a kill flag on the register localMO.setIsKill(false); MO.setIsKill(); updatedIsKill = true; break; } } if (updatedIsKill) break; } } if (updatedIsKill) break; } MBB->splice(jmpPos, MI->getParent(), MI); MBB->splice(jmpPos, MI->getParent(), cmpInstr); DebugLoc dl = MI->getDebugLoc(); MachineInstr *NewMI; assert((QII->isNewValueJumpCandidate(cmpInstr)) && "This compare is not a New Value Jump candidate."); unsigned opc = getNewValueJumpOpcode(cmpInstr, cmpOp2, isSecondOpNewified); if (invertPredicate) opc = QII->getInvertedPredicatedOpcode(opc); // Manage the conversions from CMPGEUri to either CMPEQrr // or CMPGTUri properly. See Arch spec for CMPGEUri instructions. // This has to be after the getNewValueJumpOpcode function call as // second operand of the compare could be modified in this logic. if (cmpInstr->getOpcode() == Hexagon::CMPGEUri) { if (cmpOp2 == 0) { cmpOp2 = cmpReg1; MO2IsKill = MO1IsKill; isSecondOpReg = true; } else --cmpOp2; } // Manage the conversions from CMPGEri to CMPGTUri properly. // See Arch spec for CMPGEri instructions. if (cmpInstr->getOpcode() == Hexagon::CMPGEri) --cmpOp2; if (isSecondOpReg) { NewMI = BuildMI(*MBB, jmpPos, dl, QII->get(opc)) .addReg(cmpReg1, getKillRegState(MO1IsKill)) .addReg(cmpOp2, getKillRegState(MO2IsKill)) .addMBB(jmpTarget); } else { NewMI = BuildMI(*MBB, jmpPos, dl, QII->get(opc)) .addReg(cmpReg1, getKillRegState(MO1IsKill)) .addImm(cmpOp2) .addMBB(jmpTarget); } assert(NewMI && "New Value Jump Instruction Not created!"); if (cmpInstr->getOperand(0).isReg() && cmpInstr->getOperand(0).isKill()) cmpInstr->getOperand(0).setIsKill(false); if (cmpInstr->getOperand(1).isReg() && cmpInstr->getOperand(1).isKill()) cmpInstr->getOperand(1).setIsKill(false); cmpInstr->eraseFromParent(); jmpInstr->eraseFromParent(); ++nvjGenerated; ++NumNVJGenerated; break; } } } } return true; } FunctionPass *llvm::createHexagonNewValueJump() { return new HexagonNewValueJump(); }