lsq_unit.hh

来自「M5,一个功能强大的多处理器系统模拟器.很多针对处理器架构,性能的研究都使用它作」· HH 代码 · 共 724 行 · 第 1/2 页

/*
 * Copyright (c) 2004, 2005, 2006
 * The Regents of The University of Michigan
 * All Rights Reserved
 *
 * This code is part of the M5 simulator.
 *
 * Permission is granted to use, copy, create derivative works and
 * redistribute this software and such derivative works for any
 * purpose, so long as the copyright notice above, this grant of
 * permission, and the disclaimer below appear in all copies made; and
 * so long as the name of The University of Michigan is not used in
 * any advertising or publicity pertaining to the use or distribution
 * of this software without specific, written prior authorization.
 *
 * THIS SOFTWARE IS PROVIDED AS IS, WITHOUT REPRESENTATION FROM THE
 * UNIVERSITY OF MICHIGAN AS TO ITS FITNESS FOR ANY PURPOSE, AND
 * WITHOUT WARRANTY BY THE UNIVERSITY OF MICHIGAN OF ANY KIND, EITHER
 * EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION THE IMPLIED
 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
 * PURPOSE. THE REGENTS OF THE UNIVERSITY OF MICHIGAN SHALL NOT BE
 * LIABLE FOR ANY DAMAGES, INCLUDING DIRECT, SPECIAL, INDIRECT,
 * INCIDENTAL, OR CONSEQUENTIAL DAMAGES, WITH RESPECT TO ANY CLAIM
 * ARISING OUT OF OR IN CONNECTION WITH THE USE OF THE SOFTWARE, EVEN
 * IF IT HAS BEEN OR IS HEREAFTER ADVISED OF THE POSSIBILITY OF SUCH
 * DAMAGES.
 *
 * Authors: Kevin T. Lim
 *          Korey L. Sewell
 */

#ifndef __CPU_O3_LSQ_UNIT_HH__
#define __CPU_O3_LSQ_UNIT_HH__

#include <algorithm>
#include <cstring>
#include <map>
#include <queue>

#include "arch/faults.hh"
#include "arch/locked_mem.hh"
#include "config/full_system.hh"
#include "base/hashmap.hh"
#include "cpu/inst_seq.hh"
#include "mem/packet.hh"
#include "mem/port.hh"

/**
 * Class that implements the actual LQ and SQ for each specific
 * thread.  Both are circular queues; load entries are freed upon
 * committing, while store entries are freed once they writeback. The
 * LSQUnit tracks if there are memory ordering violations, and also
 * detects partial load to store forwarding cases (a store only has
 * part of a load's data) that requires the load to wait until the
 * store writes back. In the former case it holds onto the instruction
 * until the dependence unit looks at it, and in the latter it stalls
 * the LSQ until the store writes back. At that point the load is
 * replayed.
 */
template <class Impl>
class LSQUnit {
  protected:
    typedef TheISA::IntReg IntReg;
  public:
    typedef typename Impl::Params Params;
    typedef typename Impl::O3CPU O3CPU;
    typedef typename Impl::DynInstPtr DynInstPtr;
    typedef typename Impl::CPUPol::IEW IEW;
    typedef typename Impl::CPUPol::LSQ LSQ;
    typedef typename Impl::CPUPol::IssueStruct IssueStruct;

  public:
    /** Constructs an LSQ unit. init() must be called prior to use. */
    LSQUnit();

    /** Initializes the LSQ unit with the specified number of entries. */
    void init(O3CPU *cpu_ptr, IEW *iew_ptr, Params *params, LSQ *lsq_ptr,
              unsigned maxLQEntries, unsigned maxSQEntries, unsigned id);

    /** Returns the name of the LSQ unit. */
    std::string name() const;

    /** Registers statistics. */
    void regStats();

    /** Sets the pointer to the dcache port. */
    void setDcachePort(Port *dcache_port);

    /** Switches out LSQ unit. */
    void switchOut();

    /** Takes over from another CPU's thread. */
    void takeOverFrom();

    /** Returns if the LSQ is switched out. */
    bool isSwitchedOut() { return switchedOut; }

    /** Ticks the LSQ unit, which in this case only resets the number of
     * used cache ports.
     * @todo: Move the number of used ports up to the LSQ level so it can
     * be shared by all LSQ units.
     */
    void tick() { usedPorts = 0; }

    /** Inserts an instruction. */
    void insert(DynInstPtr &inst);
    /** Inserts a load instruction. */
    void insertLoad(DynInstPtr &load_inst);
    /** Inserts a store instruction. */
    void insertStore(DynInstPtr &store_inst);

    /** Executes a load instruction. */
    Fault executeLoad(DynInstPtr &inst);

    Fault executeLoad(int lq_idx) { panic("Not implemented"); return NoFault; }
    /** Executes a store instruction. */
    Fault executeStore(DynInstPtr &inst);

    /** Commits the head load. */
    void commitLoad();
    /** Commits loads older than a specific sequence number. */
    void commitLoads(InstSeqNum &youngest_inst);

    /** Commits stores older than a specific sequence number. */
    void commitStores(InstSeqNum &youngest_inst);

    /** Writes back stores. */
    void writebackStores();

    /** Completes the data access that has been returned from the
     * memory system. */
    void completeDataAccess(PacketPtr pkt);

    /** Clears all the entries in the LQ. */
    void clearLQ();

    /** Clears all the entries in the SQ. */
    void clearSQ();

    /** Resizes the LQ to a given size. */
    void resizeLQ(unsigned size);

    /** Resizes the SQ to a given size. */
    void resizeSQ(unsigned size);

    /** Squashes all instructions younger than a specific sequence number. */
    void squash(const InstSeqNum &squashed_num);

    /** Returns if there is a memory ordering violation. Value is reset upon
     * call to getMemDepViolator().
     */
    bool violation() { return memDepViolator; }

    /** Returns the memory ordering violator. */
    DynInstPtr getMemDepViolator();

    /** Returns if a load became blocked due to the memory system. */
    bool loadBlocked()
    { return isLoadBlocked; }

    /** Clears the signal that a load became blocked. */
    void clearLoadBlocked()
    { isLoadBlocked = false; }

    /** Returns if the blocked load was handled. */
    bool isLoadBlockedHandled()
    { return loadBlockedHandled; }

    /** Records the blocked load as being handled. */
    void setLoadBlockedHandled()
    { loadBlockedHandled = true; }

    /** Returns the number of free entries (min of free LQ and SQ entries). */
    unsigned numFreeEntries();

    /** Returns the number of loads ready to execute. */
    int numLoadsReady();

    /** Returns the number of loads in the LQ. */
    int numLoads() { return loads; }

    /** Returns the number of stores in the SQ. */
    int numStores() { return stores; }

    /** Returns if either the LQ or SQ is full. */
    bool isFull() { return lqFull() || sqFull(); }

    /** Returns if the LQ is full. */
    bool lqFull() { return loads >= (LQEntries - 1); }

    /** Returns if the SQ is full. */
    bool sqFull() { return stores >= (SQEntries - 1); }

    /** Returns the number of instructions in the LSQ. */
    unsigned getCount() { return loads + stores; }

    /** Returns if there are any stores to writeback. */
    bool hasStoresToWB() { return storesToWB; }

    /** Returns the number of stores to writeback. */
    int numStoresToWB() { return storesToWB; }

    /** Returns if the LSQ unit will writeback on this cycle. */
    bool willWB() { return storeQueue[storeWBIdx].canWB &&
                        !storeQueue[storeWBIdx].completed &&
                        !isStoreBlocked; }

    /** Handles doing the retry. */
    void recvRetry();

  private:
    /** Writes back the instruction, sending it to IEW. */
    void writeback(DynInstPtr &inst, PacketPtr pkt);

    /** Handles completing the send of a store to memory. */
    void storePostSend(PacketPtr pkt);

    /** Completes the store at the specified index. */
    void completeStore(int store_idx);

    /** Increments the given store index (circular queue). */
    inline void incrStIdx(int &store_idx);
    /** Decrements the given store index (circular queue). */
    inline void decrStIdx(int &store_idx);
    /** Increments the given load index (circular queue). */
    inline void incrLdIdx(int &load_idx);
    /** Decrements the given load index (circular queue). */
    inline void decrLdIdx(int &load_idx);

  public:
    /** Debugging function to dump instructions in the LSQ. */
    void dumpInsts();

  private:
    /** Pointer to the CPU. */
    O3CPU *cpu;

    /** Pointer to the IEW stage. */
    IEW *iewStage;

    /** Pointer to the LSQ. */
    LSQ *lsq;

    /** Pointer to the dcache port.  Used only for sending. */
    Port *dcachePort;

    /** Derived class to hold any sender state the LSQ needs. */
    class LSQSenderState : public Packet::SenderState
    {
      public:
        /** Default constructor. */
        LSQSenderState()
            : noWB(false)
        { }

        /** Instruction who initiated the access to memory. */
        DynInstPtr inst;
        /** Whether or not it is a load. */
        bool isLoad;
        /** The LQ/SQ index of the instruction. */
        int idx;
        /** Whether or not the instruction will need to writeback. */
        bool noWB;
    };

    /** Writeback event, specifically for when stores forward data to loads. */
    class WritebackEvent : public Event {
      public:
        /** Constructs a writeback event. */
        WritebackEvent(DynInstPtr &_inst, PacketPtr pkt, LSQUnit *lsq_ptr);

        /** Processes the writeback event. */
        void process();

        /** Returns the description of this event. */
        const char *description() const;

      private:
        /** Instruction whose results are being written back. */
        DynInstPtr inst;

        /** The packet that would have been sent to memory. */
        PacketPtr pkt;

        /** The pointer to the LSQ unit that issued the store. */
        LSQUnit<Impl> *lsqPtr;
    };

  public:
    struct SQEntry {
        /** Constructs an empty store queue entry. */
        SQEntry()
            : inst(NULL), req(NULL), size(0),
              canWB(0), committed(0), completed(0)
        {
            std::memset(data, 0, sizeof(data));
        }

        /** Constructs a store queue entry for a given instruction. */
        SQEntry(DynInstPtr &_inst)
            : inst(_inst), req(NULL), size(0),
              canWB(0), committed(0), completed(0)
        {
            std::memset(data, 0, sizeof(data));
        }

        /** The store instruction. */
        DynInstPtr inst;
        /** The request for the store. */
        RequestPtr req;
        /** The size of the store. */
        int size;
        /** The store data. */
        char data[sizeof(IntReg)];
        /** Whether or not the store can writeback. */
        bool canWB;
        /** Whether or not the store is committed. */
        bool committed;
        /** Whether or not the store is completed. */
        bool completed;
    };

  private:
    /** The LSQUnit thread id. */
    unsigned lsqID;

    /** The store queue. */
    std::vector<SQEntry> storeQueue;

    /** The load queue. */
    std::vector<DynInstPtr> loadQueue;

    /** The number of LQ entries, plus a sentinel entry (circular queue).
     *  @todo: Consider having var that records the true number of LQ entries.
     */
    unsigned LQEntries;
    /** The number of SQ entries, plus a sentinel entry (circular queue).
     *  @todo: Consider having var that records the true number of SQ entries.
     */
    unsigned SQEntries;

    /** The number of load instructions in the LQ. */
    int loads;
    /** The number of store instructions in the SQ. */
    int stores;
    /** The number of store instructions in the SQ waiting to writeback. */
    int storesToWB;

    /** The index of the head instruction in the LQ. */
    int loadHead;
    /** The index of the tail instruction in the LQ. */
    int loadTail;

    /** The index of the head instruction in the SQ. */
    int storeHead;
    /** The index of the first instruction that may be ready to be
     * written back, and has not yet been written back.
     */
    int storeWBIdx;
    /** The index of the tail instruction in the SQ. */
    int storeTail;