gem5/cpu/beta_cpu/mem_dep_unit.hh
Kevin Lim 2fb632dbda Check in of various updates to the CPU. Mainly adds in stats, improves
branch prediction, and makes memory dependence work properly.

SConscript:
    Added return address stack, tournament predictor.
cpu/base_cpu.cc:
    Added debug break and print statements.
cpu/base_dyn_inst.cc:
cpu/base_dyn_inst.hh:
    Comment out possibly unneeded variables.
cpu/beta_cpu/2bit_local_pred.cc:
    2bit predictor no longer speculatively updates itself.
cpu/beta_cpu/alpha_dyn_inst.hh:
    Comment formatting.
cpu/beta_cpu/alpha_full_cpu.hh:
    Formatting
cpu/beta_cpu/alpha_full_cpu_builder.cc:
    Added new parameters for branch predictors, and IQ parameters.
cpu/beta_cpu/alpha_full_cpu_impl.hh:
    Register stats.
cpu/beta_cpu/alpha_params.hh:
    Added parameters for IQ, branch predictors, and store sets.
cpu/beta_cpu/bpred_unit.cc:
    Removed one class.
cpu/beta_cpu/bpred_unit.hh:
    Add in RAS, stats.  Changed branch predictor unit functionality
    so that it holds a history of past branches so it can update, and also
    hold a proper history of the RAS so it can be restored on branch
    mispredicts.
cpu/beta_cpu/bpred_unit_impl.hh:
    Added in stats, history of branches, RAS.  Now bpred unit actually
    modifies the instruction's predicted next PC.
cpu/beta_cpu/btb.cc:
    Add in sanity checks.
cpu/beta_cpu/comm.hh:
    Add in communication where needed, remove it where it's not.
cpu/beta_cpu/commit.hh:
cpu/beta_cpu/rename.hh:
cpu/beta_cpu/rename_impl.hh:
    Add in stats.
cpu/beta_cpu/commit_impl.hh:
    Stats, update what is sent back on branch mispredict.
cpu/beta_cpu/cpu_policy.hh:
    Change the bpred unit being used.
cpu/beta_cpu/decode.hh:
cpu/beta_cpu/decode_impl.hh:
    Stats.
cpu/beta_cpu/fetch.hh:
    Stats, change squash so it can handle squashes from decode differently
    than squashes from commit.
cpu/beta_cpu/fetch_impl.hh:
    Add in stats.  Change how a cache line is fetched.  Update to work with
    caches.  Also have separate functions for different behavior if squash
    is coming from decode vs commit.
cpu/beta_cpu/free_list.hh:
    Remove some old comments.
cpu/beta_cpu/full_cpu.cc:
cpu/beta_cpu/full_cpu.hh:
    Added function to remove instructions from back of instruction list
    until a certain sequence number.
cpu/beta_cpu/iew.hh:
    Stats, separate squashing behavior due to branches vs memory.
cpu/beta_cpu/iew_impl.hh:
    Stats, separate squashing behavior for branches vs memory.
cpu/beta_cpu/inst_queue.cc:
    Debug stuff
cpu/beta_cpu/inst_queue.hh:
    Stats, change how mem dep unit works, debug stuff
cpu/beta_cpu/inst_queue_impl.hh:
    Stats, change how mem dep unit works, debug stuff.  Also add in
    parameters that used to be hardcoded.
cpu/beta_cpu/mem_dep_unit.hh:
cpu/beta_cpu/mem_dep_unit_impl.hh:
    Add in stats, change how memory dependence unit works.  It now holds
    the memory instructions that are waiting for their memory dependences
    to resolve.  It provides which instructions are ready directly to the
    IQ.
cpu/beta_cpu/regfile.hh:
    Fix up sanity checks.
cpu/beta_cpu/rename_map.cc:
    Fix loop variable type.
cpu/beta_cpu/rob_impl.hh:
    Remove intermediate DynInstPtr
cpu/beta_cpu/store_set.cc:
    Add in debugging statements.
cpu/beta_cpu/store_set.hh:
    Reorder function arguments to match the rest of the calls.

--HG--
extra : convert_revision : aabf9b1fecd1d743265dfc3b174d6159937c6f44
2004-10-21 18:02:36 -04:00

141 lines
4.2 KiB
C++

#ifndef __MEM_DEP_UNIT_HH__
#define __MEM_DEP_UNIT_HH__
#include <set>
#include <map>
#include "cpu/inst_seq.hh"
#include "base/statistics.hh"
/**
* Memory dependency unit class. This holds the memory dependence predictor.
* As memory operations are issued to the IQ, they are also issued to this
* unit, which then looks up the prediction as to what they are dependent
* upon. This unit must be checked prior to a memory operation being able
* to issue. Although this is templated, it's somewhat hard to make a generic
* memory dependence unit. This one is mostly for store sets; it will be
* quite limited in what other memory dependence predictions it can also
* utilize. Thus this class should be most likely be rewritten for other
* dependence prediction schemes.
*/
template <class MemDepPred, class Impl>
class MemDepUnit {
public:
typedef typename Impl::Params Params;
typedef typename Impl::DynInstPtr DynInstPtr;
public:
MemDepUnit(Params &params);
void regStats();
void insert(DynInstPtr &inst);
void insertNonSpec(DynInstPtr &inst);
void regsReady(DynInstPtr &inst);
void nonSpecInstReady(DynInstPtr &inst);
void issue(DynInstPtr &inst);
void wakeDependents(DynInstPtr &inst);
void squash(const InstSeqNum &squashed_num);
void violation(DynInstPtr &store_inst, DynInstPtr &violating_load);
// Will want to make this operation relatively fast. Right now it
// kind of sucks.
DynInstPtr &top();
void pop();
inline bool empty()
{ return readyInsts.empty(); }
private:
typedef typename std::set<InstSeqNum>::iterator sn_it_t;
typedef typename std::map<InstSeqNum, DynInstPtr>::iterator dyn_it_t;
// Forward declarations so that the following two typedefs work.
class Dependency;
class ltDependency;
typedef typename std::set<Dependency, ltDependency>::iterator dep_it_t;
typedef typename std::map<InstSeqNum, vector<dep_it_t> >::iterator
sd_it_t;
struct Dependency {
Dependency(const InstSeqNum &_seqNum)
: seqNum(_seqNum), regsReady(0), memDepReady(0)
{ }
Dependency(const InstSeqNum &_seqNum, bool _regsReady,
bool _memDepReady)
: seqNum(_seqNum), regsReady(_regsReady),
memDepReady(_memDepReady)
{ }
InstSeqNum seqNum;
mutable bool regsReady;
mutable bool memDepReady;
mutable sd_it_t storeDep;
};
struct ltDependency {
bool operator() (const Dependency &lhs, const Dependency &rhs)
{
return lhs.seqNum < rhs.seqNum;
}
};
private:
inline void moveToReady(dep_it_t &woken_inst);
private:
/** List of instructions that have passed through rename, yet are still
* waiting on either a memory dependence to resolve or source registers to
* become available before they can issue.
*/
std::set<Dependency, ltDependency> waitingInsts;
/** List of instructions that have all their predicted memory dependences
* resolved and their source registers ready.
*/
std::set<InstSeqNum> readyInsts;
// Change this to hold a vector of iterators, which will point to the
// entry of the waiting instructions.
/** List of stores' sequence numbers, each of which has a vector of
* iterators. The iterators point to the appropriate node within
* waitingInsts that has the depenendent instruction.
*/
std::map<InstSeqNum, vector<dep_it_t> > storeDependents;
// For now will implement this as a map...hash table might not be too
// bad, or could move to something that mimics the current dependency
// graph.
std::map<InstSeqNum, DynInstPtr> memInsts;
// Iterator pointer to the top instruction which has is ready.
// Is set by the top() call.
dyn_it_t topInst;
/** The memory dependence predictor. It is accessed upon new
* instructions being added to the IQ, and responds by telling
* this unit what instruction the newly added instruction is dependent
* upon.
*/
MemDepPred depPred;
Stats::Scalar<> insertedLoads;
Stats::Scalar<> insertedStores;
Stats::Scalar<> conflictingLoads;
Stats::Scalar<> conflictingStores;
};
#endif