Home | History | Annotate | Download | only in CodeGen
      1 //==- ScheduleDAGInstrs.h - MachineInstr Scheduling --------------*- C++ -*-==//
      2 //
      3 //                     The LLVM Compiler Infrastructure
      4 //
      5 // This file is distributed under the University of Illinois Open Source
      6 // License. See LICENSE.TXT for details.
      7 //
      8 //===----------------------------------------------------------------------===//
      9 //
     10 // This file implements the ScheduleDAGInstrs class, which implements
     11 // scheduling for a MachineInstr-based dependency graph.
     12 //
     13 //===----------------------------------------------------------------------===//
     14 
     15 #ifndef SCHEDULEDAGINSTRS_H
     16 #define SCHEDULEDAGINSTRS_H
     17 
     18 #include "llvm/CodeGen/MachineDominators.h"
     19 #include "llvm/CodeGen/MachineLoopInfo.h"
     20 #include "llvm/CodeGen/ScheduleDAG.h"
     21 #include "llvm/Support/Compiler.h"
     22 #include "llvm/Target/TargetRegisterInfo.h"
     23 #include "llvm/ADT/SmallSet.h"
     24 #include "llvm/ADT/SparseSet.h"
     25 #include <map>
     26 
     27 namespace llvm {
     28   class MachineLoopInfo;
     29   class MachineDominatorTree;
     30   class LiveIntervals;
     31 
     32   /// LoopDependencies - This class analyzes loop-oriented register
     33   /// dependencies, which are used to guide scheduling decisions.
     34   /// For example, loop induction variable increments should be
     35   /// scheduled as soon as possible after the variable's last use.
     36   ///
     37   class LoopDependencies {
     38     const MachineLoopInfo &MLI;
     39     const MachineDominatorTree &MDT;
     40 
     41   public:
     42     typedef std::map<unsigned, std::pair<const MachineOperand *, unsigned> >
     43       LoopDeps;
     44     LoopDeps Deps;
     45 
     46     LoopDependencies(const MachineLoopInfo &mli,
     47                      const MachineDominatorTree &mdt) :
     48       MLI(mli), MDT(mdt) {}
     49 
     50     /// VisitLoop - Clear out any previous state and analyze the given loop.
     51     ///
     52     void VisitLoop(const MachineLoop *Loop) {
     53       assert(Deps.empty() && "stale loop dependencies");
     54 
     55       MachineBasicBlock *Header = Loop->getHeader();
     56       SmallSet<unsigned, 8> LoopLiveIns;
     57       for (MachineBasicBlock::livein_iterator LI = Header->livein_begin(),
     58            LE = Header->livein_end(); LI != LE; ++LI)
     59         LoopLiveIns.insert(*LI);
     60 
     61       const MachineDomTreeNode *Node = MDT.getNode(Header);
     62       const MachineBasicBlock *MBB = Node->getBlock();
     63       assert(Loop->contains(MBB) &&
     64              "Loop does not contain header!");
     65       VisitRegion(Node, MBB, Loop, LoopLiveIns);
     66     }
     67 
     68   private:
     69     void VisitRegion(const MachineDomTreeNode *Node,
     70                      const MachineBasicBlock *MBB,
     71                      const MachineLoop *Loop,
     72                      const SmallSet<unsigned, 8> &LoopLiveIns) {
     73       unsigned Count = 0;
     74       for (MachineBasicBlock::const_iterator I = MBB->begin(), E = MBB->end();
     75            I != E; ++I) {
     76         const MachineInstr *MI = I;
     77         if (MI->isDebugValue())
     78           continue;
     79         for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
     80           const MachineOperand &MO = MI->getOperand(i);
     81           if (!MO.isReg() || !MO.isUse())
     82             continue;
     83           unsigned MOReg = MO.getReg();
     84           if (LoopLiveIns.count(MOReg))
     85             Deps.insert(std::make_pair(MOReg, std::make_pair(&MO, Count)));
     86         }
     87         ++Count; // Not every iteration due to dbg_value above.
     88       }
     89 
     90       const std::vector<MachineDomTreeNode*> &Children = Node->getChildren();
     91       for (std::vector<MachineDomTreeNode*>::const_iterator I =
     92            Children.begin(), E = Children.end(); I != E; ++I) {
     93         const MachineDomTreeNode *ChildNode = *I;
     94         MachineBasicBlock *ChildBlock = ChildNode->getBlock();
     95         if (Loop->contains(ChildBlock))
     96           VisitRegion(ChildNode, ChildBlock, Loop, LoopLiveIns);
     97       }
     98     }
     99   };
    100 
    101   /// An individual mapping from virtual register number to SUnit.
    102   struct VReg2SUnit {
    103     unsigned VirtReg;
    104     SUnit *SU;
    105 
    106     VReg2SUnit(unsigned reg, SUnit *su): VirtReg(reg), SU(su) {}
    107 
    108     unsigned getSparseSetKey() const {
    109       return TargetRegisterInfo::virtReg2Index(VirtReg);
    110     }
    111   };
    112 
    113   /// Combine a SparseSet with a 1x1 vector to track physical registers.
    114   /// The SparseSet allows iterating over the (few) live registers for quickly
    115   /// comparing against a regmask or clearing the set.
    116   ///
    117   /// Storage for the map is allocated once for the pass. The map can be
    118   /// cleared between scheduling regions without freeing unused entries.
    119   class Reg2SUnitsMap {
    120     SparseSet<unsigned> PhysRegSet;
    121     std::vector<std::vector<SUnit*> > SUnits;
    122   public:
    123     typedef SparseSet<unsigned>::const_iterator const_iterator;
    124 
    125     // Allow iteration over register numbers (keys) in the map. If needed, we
    126     // can provide an iterator over SUnits (values) as well.
    127     const_iterator reg_begin() const { return PhysRegSet.begin(); }
    128     const_iterator reg_end() const { return PhysRegSet.end(); }
    129 
    130     /// Initialize the map with the number of registers.
    131     /// If the map is already large enough, no allocation occurs.
    132     /// For simplicity we expect the map to be empty().
    133     void setRegLimit(unsigned Limit);
    134 
    135     /// Returns true if the map is empty.
    136     bool empty() const { return PhysRegSet.empty(); }
    137 
    138     /// Clear the map without deallocating storage.
    139     void clear();
    140 
    141     bool contains(unsigned Reg) const { return PhysRegSet.count(Reg); }
    142 
    143     /// If this register is mapped, return its existing SUnits vector.
    144     /// Otherwise map the register and return an empty SUnits vector.
    145     std::vector<SUnit *> &operator[](unsigned Reg) {
    146       bool New = PhysRegSet.insert(Reg).second;
    147       assert((!New || SUnits[Reg].empty()) && "stale SUnits vector");
    148       (void)New;
    149       return SUnits[Reg];
    150     }
    151 
    152     /// Erase an existing element without freeing memory.
    153     void erase(unsigned Reg) {
    154       PhysRegSet.erase(Reg);
    155       SUnits[Reg].clear();
    156     }
    157   };
    158 
    159   /// Use SparseSet as a SparseMap by relying on the fact that it never
    160   /// compares ValueT's, only unsigned keys. This allows the set to be cleared
    161   /// between scheduling regions in constant time as long as ValueT does not
    162   /// require a destructor.
    163   typedef SparseSet<VReg2SUnit> VReg2SUnitMap;
    164 
    165   /// ScheduleDAGInstrs - A ScheduleDAG subclass for scheduling lists of
    166   /// MachineInstrs.
    167   class ScheduleDAGInstrs : public ScheduleDAG {
    168   protected:
    169     const MachineLoopInfo &MLI;
    170     const MachineDominatorTree &MDT;
    171     const MachineFrameInfo *MFI;
    172     const InstrItineraryData *InstrItins;
    173 
    174     /// Live Intervals provides reaching defs in preRA scheduling.
    175     LiveIntervals *LIS;
    176 
    177     /// isPostRA flag indicates vregs cannot be present.
    178     bool IsPostRA;
    179 
    180     /// UnitLatencies (misnamed) flag avoids computing def-use latencies, using
    181     /// the def-side latency only.
    182     bool UnitLatencies;
    183 
    184     /// The standard DAG builder does not normally include terminators as DAG
    185     /// nodes because it does not create the necessary dependencies to prevent
    186     /// reordering. A specialized scheduler can overide
    187     /// TargetInstrInfo::isSchedulingBoundary then enable this flag to indicate
    188     /// it has taken responsibility for scheduling the terminator correctly.
    189     bool CanHandleTerminators;
    190 
    191     /// State specific to the current scheduling region.
    192     /// ------------------------------------------------
    193 
    194     /// The block in which to insert instructions
    195     MachineBasicBlock *BB;
    196 
    197     /// The beginning of the range to be scheduled.
    198     MachineBasicBlock::iterator RegionBegin;
    199 
    200     /// The end of the range to be scheduled.
    201     MachineBasicBlock::iterator RegionEnd;
    202 
    203     /// The index in BB of RegionEnd.
    204     unsigned EndIndex;
    205 
    206     /// After calling BuildSchedGraph, each machine instruction in the current
    207     /// scheduling region is mapped to an SUnit.
    208     DenseMap<MachineInstr*, SUnit*> MISUnitMap;
    209 
    210     /// State internal to DAG building.
    211     /// -------------------------------
    212 
    213     /// Defs, Uses - Remember where defs and uses of each register are as we
    214     /// iterate upward through the instructions. This is allocated here instead
    215     /// of inside BuildSchedGraph to avoid the need for it to be initialized and
    216     /// destructed for each block.
    217     Reg2SUnitsMap Defs;
    218     Reg2SUnitsMap Uses;
    219 
    220     /// Track the last instructon in this region defining each virtual register.
    221     VReg2SUnitMap VRegDefs;
    222 
    223     /// PendingLoads - Remember where unknown loads are after the most recent
    224     /// unknown store, as we iterate. As with Defs and Uses, this is here
    225     /// to minimize construction/destruction.
    226     std::vector<SUnit *> PendingLoads;
    227 
    228     /// LoopRegs - Track which registers are used for loop-carried dependencies.
    229     ///
    230     LoopDependencies LoopRegs;
    231 
    232     /// DbgValues - Remember instruction that preceeds DBG_VALUE.
    233     /// These are generated by buildSchedGraph but persist so they can be
    234     /// referenced when emitting the final schedule.
    235     typedef std::vector<std::pair<MachineInstr *, MachineInstr *> >
    236       DbgValueVector;
    237     DbgValueVector DbgValues;
    238     MachineInstr *FirstDbgValue;
    239 
    240   public:
    241     explicit ScheduleDAGInstrs(MachineFunction &mf,
    242                                const MachineLoopInfo &mli,
    243                                const MachineDominatorTree &mdt,
    244                                bool IsPostRAFlag,
    245                                LiveIntervals *LIS = 0);
    246 
    247     virtual ~ScheduleDAGInstrs() {}
    248 
    249     /// begin - Return an iterator to the top of the current scheduling region.
    250     MachineBasicBlock::iterator begin() const { return RegionBegin; }
    251 
    252     /// end - Return an iterator to the bottom of the current scheduling region.
    253     MachineBasicBlock::iterator end() const { return RegionEnd; }
    254 
    255     /// newSUnit - Creates a new SUnit and return a ptr to it.
    256     SUnit *newSUnit(MachineInstr *MI);
    257 
    258     /// getSUnit - Return an existing SUnit for this MI, or NULL.
    259     SUnit *getSUnit(MachineInstr *MI) const;
    260 
    261     /// startBlock - Prepare to perform scheduling in the given block.
    262     virtual void startBlock(MachineBasicBlock *BB);
    263 
    264     /// finishBlock - Clean up after scheduling in the given block.
    265     virtual void finishBlock();
    266 
    267     /// Initialize the scheduler state for the next scheduling region.
    268     virtual void enterRegion(MachineBasicBlock *bb,
    269                              MachineBasicBlock::iterator begin,
    270                              MachineBasicBlock::iterator end,
    271                              unsigned endcount);
    272 
    273     /// Notify that the scheduler has finished scheduling the current region.
    274     virtual void exitRegion();
    275 
    276     /// buildSchedGraph - Build SUnits from the MachineBasicBlock that we are
    277     /// input.
    278     void buildSchedGraph(AliasAnalysis *AA);
    279 
    280     /// addSchedBarrierDeps - Add dependencies from instructions in the current
    281     /// list of instructions being scheduled to scheduling barrier. We want to
    282     /// make sure instructions which define registers that are either used by
    283     /// the terminator or are live-out are properly scheduled. This is
    284     /// especially important when the definition latency of the return value(s)
    285     /// are too high to be hidden by the branch or when the liveout registers
    286     /// used by instructions in the fallthrough block.
    287     void addSchedBarrierDeps();
    288 
    289     /// computeLatency - Compute node latency.
    290     ///
    291     virtual void computeLatency(SUnit *SU);
    292 
    293     /// computeOperandLatency - Override dependence edge latency using
    294     /// operand use/def information
    295     ///
    296     virtual void computeOperandLatency(SUnit *Def, SUnit *Use,
    297                                        SDep& dep) const;
    298 
    299     /// schedule - Order nodes according to selected style, filling
    300     /// in the Sequence member.
    301     ///
    302     /// Typically, a scheduling algorithm will implement schedule() without
    303     /// overriding enterRegion() or exitRegion().
    304     virtual void schedule() = 0;
    305 
    306     /// finalizeSchedule - Allow targets to perform final scheduling actions at
    307     /// the level of the whole MachineFunction. By default does nothing.
    308     virtual void finalizeSchedule() {}
    309 
    310     virtual void dumpNode(const SUnit *SU) const;
    311 
    312     /// Return a label for a DAG node that points to an instruction.
    313     virtual std::string getGraphNodeLabel(const SUnit *SU) const;
    314 
    315     /// Return a label for the region of code covered by the DAG.
    316     virtual std::string getDAGName() const;
    317 
    318   protected:
    319     void initSUnits();
    320     void addPhysRegDataDeps(SUnit *SU, const MachineOperand &MO);
    321     void addPhysRegDeps(SUnit *SU, unsigned OperIdx);
    322     void addVRegDefDeps(SUnit *SU, unsigned OperIdx);
    323     void addVRegUseDeps(SUnit *SU, unsigned OperIdx);
    324 
    325     VReg2SUnitMap::iterator findVRegDef(unsigned VirtReg) {
    326       return VRegDefs.find(TargetRegisterInfo::virtReg2Index(VirtReg));
    327     }
    328   };
    329 
    330   /// newSUnit - Creates a new SUnit and return a ptr to it.
    331   inline SUnit *ScheduleDAGInstrs::newSUnit(MachineInstr *MI) {
    332 #ifndef NDEBUG
    333     const SUnit *Addr = SUnits.empty() ? 0 : &SUnits[0];
    334 #endif
    335     SUnits.push_back(SUnit(MI, (unsigned)SUnits.size()));
    336     assert((Addr == 0 || Addr == &SUnits[0]) &&
    337            "SUnits std::vector reallocated on the fly!");
    338     SUnits.back().OrigNode = &SUnits.back();
    339     return &SUnits.back();
    340   }
    341 
    342   /// getSUnit - Return an existing SUnit for this MI, or NULL.
    343   inline SUnit *ScheduleDAGInstrs::getSUnit(MachineInstr *MI) const {
    344     DenseMap<MachineInstr*, SUnit*>::const_iterator I = MISUnitMap.find(MI);
    345     if (I == MISUnitMap.end())
    346       return 0;
    347     return I->second;
    348   }
    349 } // namespace llvm
    350 
    351 #endif
    352