/*
- * Copyright (c) 2011 ARM Limited
+ * Copyright (c) 2011-2013 ARM Limited
* All rights reserved
*
* The license below extends only to copyright in the software and shall
#include <vector>
+// Before we do anything else, check if this build is the NULL ISA,
+// and if so stop here
+#include "config/the_isa.hh"
+#if THE_ISA == NULL_ISA
+#include "arch/null/cpu_dummy.hh"
+#else
#include "arch/interrupts.hh"
#include "arch/isa_traits.hh"
#include "arch/microcode_rom.hh"
#include "base/statistics.hh"
-#include "config/the_isa.hh"
#include "mem/mem_object.hh"
#include "sim/eventq.hh"
#include "sim/full_system.hh"
#include "sim/insttracer.hh"
+#include "sim/probe/pmu.hh"
+#include "sim/system.hh"
+#include "debug/Mwait.hh"
-class BaseCPUParams;
-class BranchPred;
+class BaseCPU;
+struct BaseCPUParams;
class CheckerCPU;
class ThreadContext;
-class System;
-class Port;
-namespace TheISA
+struct AddressMonitor
{
- class Predecoder;
-}
+ AddressMonitor();
+ bool doMonitor(PacketPtr pkt);
+
+ bool armed;
+ Addr vAddr;
+ Addr pAddr;
+ uint64_t val;
+ bool waiting; // 0=normal, 1=mwaiting
+ bool gotWakeup;
+};
class CPUProgressEvent : public Event
{
class BaseCPU : public MemObject
{
protected:
- // CPU's clock period in terms of the number of ticks of curTime.
- Tick clock;
- // @todo remove me after debugging with legion done
+
+ /// Instruction count used for SPARC misc register
+ /// @todo unify this with the counters that cpus individually keep
Tick instCnt;
+
// every cpu has an id, put it in the base cpu
// Set at initialization, only time a cpuId might change is during a
// takeover (which should be done from within the BaseCPU anyway,
// therefore no setCpuId() method is provided
int _cpuId;
+ /** Each cpu will have a socket ID that corresponds to its physical location
+ * in the system. This is usually used to bucket cpu cores under single DVFS
+ * domain. This information may also be required by the OS to identify the
+ * cpu core grouping (as in the case of ARM via MPIDR register)
+ */
+ const uint32_t _socketId;
+
/** instruction side request id that must be placed in all requests */
MasterID _instMasterId;
/** data side request id that must be placed in all requests */
MasterID _dataMasterId;
- /**
- * Define a base class for the CPU ports (instruction and data)
- * that is refined in the subclasses. This class handles the
- * common cases, i.e. the functional accesses and the status
- * changes and address range queries. The default behaviour for
- * both atomic and timing access is to panic and the corresponding
- * subclasses have to override these methods.
+ /** An intrenal representation of a task identifier within gem5. This is
+ * used so the CPU can add which taskId (which is an internal representation
+ * of the OS process ID) to each request so components in the memory system
+ * can track which process IDs are ultimately interacting with them
*/
- class CpuPort : public Port
- {
- public:
-
- /**
- * Create a CPU port with a name and a structural owner.
- *
- * @param _name port name including the owner
- * @param _name structural owner of this port
- */
- CpuPort(const std::string& _name, MemObject* _owner) :
- Port(_name, _owner)
- { }
+ uint32_t _taskId;
- protected:
+ /** The current OS process ID that is executing on this processor. This is
+ * used to generate a taskId */
+ uint32_t _pid;
- virtual bool recvTiming(PacketPtr pkt);
+ /** Is the CPU switched out or active? */
+ bool _switchedOut;
- virtual Tick recvAtomic(PacketPtr pkt);
+ /** Cache the cache line size that we get from the system */
+ const unsigned int _cacheLineSize;
- virtual void recvRetry();
-
- void recvFunctional(PacketPtr pkt);
+ public:
- void recvRangeChange();
+ /**
+ * Purely virtual method that returns a reference to the data
+ * port. All subclasses must implement this method.
+ *
+ * @return a reference to the data port
+ */
+ virtual MasterPort &getDataPort() = 0;
- };
+ /**
+ * Purely virtual method that returns a reference to the instruction
+ * port. All subclasses must implement this method.
+ *
+ * @return a reference to the instruction port
+ */
+ virtual MasterPort &getInstPort() = 0;
- public:
/** Reads this CPU's ID. */
- int cpuId() { return _cpuId; }
+ int cpuId() const { return _cpuId; }
+
+ /** Reads this CPU's Socket ID. */
+ uint32_t socketId() const { return _socketId; }
/** Reads this CPU's unique data requestor ID */
MasterID dataMasterId() { return _dataMasterId; }
/** Reads this CPU's unique instruction requestor ID */
MasterID instMasterId() { return _instMasterId; }
-// Tick currentTick;
- inline Tick frequency() const { return SimClock::Frequency / clock; }
- inline Tick ticks(int numCycles) const { return clock * numCycles; }
- inline Tick curCycle() const { return curTick() / clock; }
- inline Tick tickToCycles(Tick val) const { return val / clock; }
+ /**
+ * Get a master port on this CPU. All CPUs have a data and
+ * instruction port, and this method uses getDataPort and
+ * getInstPort of the subclasses to resolve the two ports.
+ *
+ * @param if_name the port name
+ * @param idx ignored index
+ *
+ * @return a reference to the port with the given name
+ */
+ BaseMasterPort &getMasterPort(const std::string &if_name,
+ PortID idx = InvalidPortID) override;
+
+ /** Get cpu task id */
+ uint32_t taskId() const { return _taskId; }
+ /** Set cpu task id */
+ void taskId(uint32_t id) { _taskId = id; }
+
+ uint32_t getPid() const { return _pid; }
+ void setPid(uint32_t pid) { _pid = pid; }
+
inline void workItemBegin() { numWorkItemsStarted++; }
inline void workItemEnd() { numWorkItemsCompleted++; }
// @todo remove me after debugging with legion done
Tick instCount() { return instCnt; }
- /** The next cycle the CPU should be scheduled, given a cache
- * access or quiesce event returning on this cycle. This function
- * may return curTick() if the CPU should run on the current cycle.
- */
- Tick nextCycle();
-
- /** The next cycle the CPU should be scheduled, given a cache
- * access or quiesce event returning on the given Tick. This
- * function may return curTick() if the CPU should run on the
- * current cycle.
- * @param begin_tick The tick that the event is completing on.
- */
- Tick nextCycle(Tick begin_tick);
-
TheISA::MicrocodeRom microcodeRom;
protected:
- TheISA::Interrupts *interrupts;
+ std::vector<TheISA::Interrupts*> interrupts;
public:
TheISA::Interrupts *
- getInterruptController()
+ getInterruptController(ThreadID tid)
{
- return interrupts;
+ if (interrupts.empty())
+ return NULL;
+
+ assert(interrupts.size() > tid);
+ return interrupts[tid];
}
- virtual void wakeup() = 0;
+ virtual void wakeup(ThreadID tid) = 0;
void
- postInterrupt(int int_num, int index)
+ postInterrupt(ThreadID tid, int int_num, int index)
{
- interrupts->post(int_num, index);
+ interrupts[tid]->post(int_num, index);
if (FullSystem)
- wakeup();
+ wakeup(tid);
}
void
- clearInterrupt(int int_num, int index)
+ clearInterrupt(ThreadID tid, int int_num, int index)
{
- interrupts->clear(int_num, index);
+ interrupts[tid]->clear(int_num, index);
}
void
- clearInterrupts()
+ clearInterrupts(ThreadID tid)
{
- interrupts->clearAll();
+ interrupts[tid]->clearAll();
}
bool
checkInterrupts(ThreadContext *tc) const
{
- return FullSystem && interrupts->checkInterrupts(tc);
+ return FullSystem && interrupts[tc->threadId()]->checkInterrupts(tc);
}
class ProfileEvent : public Event
protected:
std::vector<ThreadContext *> threadContexts;
- std::vector<TheISA::Predecoder *> predecoders;
Trace::InstTracer * tracer;
public:
+
+ /** Invalid or unknown Pid. Possible when operating system is not present
+ * or has not assigned a pid yet */
+ static const uint32_t invldPid = std::numeric_limits<uint32_t>::max();
+
// Mask to align PCs to MachInst sized boundaries
static const Addr PCMask = ~((Addr)sizeof(TheISA::MachInst) - 1);
/// Provide access to the tracer pointer
Trace::InstTracer * getTracer() { return tracer; }
- /// Notify the CPU that the indicated context is now active. The
- /// delay parameter indicates the number of ticks to wait before
- /// executing (typically 0 or 1).
- virtual void activateContext(ThreadID thread_num, int delay) {}
+ /// Notify the CPU that the indicated context is now active.
+ virtual void activateContext(ThreadID thread_num);
/// Notify the CPU that the indicated context is now suspended.
- virtual void suspendContext(ThreadID thread_num) {}
-
- /// Notify the CPU that the indicated context is now deallocated.
- virtual void deallocateContext(ThreadID thread_num) {}
+ /// Check if possible to enter a lower power state
+ virtual void suspendContext(ThreadID thread_num);
/// Notify the CPU that the indicated context is now halted.
virtual void haltContext(ThreadID thread_num) {}
int findContext(ThreadContext *tc);
/// Given a thread num get tho thread context for it
- ThreadContext *getContext(int tn) { return threadContexts[tn]; }
+ virtual ThreadContext *getContext(int tn) { return threadContexts[tn]; }
+
+ /// Get the number of thread contexts available
+ unsigned numContexts() { return threadContexts.size(); }
+
+ /// Convert ContextID to threadID
+ ThreadID contextToThread(ContextID cid)
+ { return static_cast<ThreadID>(cid - threadContexts[0]->contextId()); }
public:
typedef BaseCPUParams Params;
const Params *params() const
{ return reinterpret_cast<const Params *>(_params); }
- BaseCPU(Params *params);
+ BaseCPU(Params *params, bool is_checker = false);
virtual ~BaseCPU();
- virtual void init();
- virtual void startup();
- virtual void regStats();
+ void init() override;
+ void startup() override;
+ void regStats() override;
- virtual void activateWhenReady(ThreadID tid) {};
+ void regProbePoints() override;
void registerThreadContexts();
- /// Prepare for another CPU to take over execution. When it is
- /// is ready (drained pipe) it signals the sampler.
+ /**
+ * Prepare for another CPU to take over execution.
+ *
+ * When this method exits, all internal state should have been
+ * flushed. After the method returns, the simulator calls
+ * takeOverFrom() on the new CPU with this CPU as its parameter.
+ */
virtual void switchOut();
- /// Take over execution from the given CPU. Used for warm-up and
- /// sampling.
- virtual void takeOverFrom(BaseCPU *);
+ /**
+ * Load the state of a CPU from the previous CPU object, invoked
+ * on all new CPUs that are about to be switched in.
+ *
+ * A CPU model implementing this method is expected to initialize
+ * its state from the old CPU and connect its memory (unless they
+ * are already connected) to the memories connected to the old
+ * CPU.
+ *
+ * @param cpu CPU to initialize read state from.
+ */
+ virtual void takeOverFrom(BaseCPU *cpu);
+
+ /**
+ * Flush all TLBs in the CPU.
+ *
+ * This method is mainly used to flush stale translations when
+ * switching CPUs. It is also exported to the Python world to
+ * allow it to request a TLB flush after draining the CPU to make
+ * it easier to compare traces when debugging
+ * handover/checkpointing.
+ */
+ void flushTLBs();
+
+ /**
+ * Determine if the CPU is switched out.
+ *
+ * @return True if the CPU is switched out, false otherwise.
+ */
+ bool switchedOut() const { return _switchedOut; }
+
+ /**
+ * Verify that the system is in a memory mode supported by the
+ * CPU.
+ *
+ * Implementations are expected to query the system for the
+ * current memory mode and ensure that it is what the CPU model
+ * expects. If the check fails, the implementation should
+ * terminate the simulation using fatal().
+ */
+ virtual void verifyMemoryMode() const { };
/**
* Number of threads we're actually simulating (<= SMT_MAX_THREADS).
System *system;
- Tick phase;
+ /**
+ * Get the cache line size of the system.
+ */
+ inline unsigned int cacheLineSize() const { return _cacheLineSize; }
/**
* Serialize this object to the given output stream.
+ *
+ * @note CPU models should normally overload the serializeThread()
+ * method instead of the serialize() method as this provides a
+ * uniform data format for all CPU models and promotes better code
+ * reuse.
+ *
* @param os The stream to serialize to.
*/
- virtual void serialize(std::ostream &os);
+ void serialize(CheckpointOut &cp) const override;
/**
* Reconstruct the state of this object from a checkpoint.
+ *
+ * @note CPU models should normally overload the
+ * unserializeThread() method instead of the unserialize() method
+ * as this provides a uniform data format for all CPU models and
+ * promotes better code reuse.
+
* @param cp The checkpoint use.
- * @param section The section name of this object
+ * @param section The section name of this object.
+ */
+ void unserialize(CheckpointIn &cp) override;
+
+ /**
+ * Serialize a single thread.
+ *
+ * @param os The stream to serialize to.
+ * @param tid ID of the current thread.
*/
- virtual void unserialize(Checkpoint *cp, const std::string §ion);
+ virtual void serializeThread(CheckpointOut &cp, ThreadID tid) const {};
/**
- * Return pointer to CPU's branch predictor (NULL if none).
- * @return Branch predictor pointer.
+ * Unserialize one thread.
+ *
+ * @param cp The checkpoint use.
+ * @param section The section name of this thread.
+ * @param tid ID of the current thread.
*/
- virtual BranchPred *getBranchPred() { return NULL; };
+ virtual void unserializeThread(CheckpointIn &cp, ThreadID tid) {};
virtual Counter totalInsts() const = 0;
virtual Counter totalOps() const = 0;
+ /**
+ * Schedule an event that exits the simulation loops after a
+ * predefined number of instructions.
+ *
+ * This method is usually called from the configuration script to
+ * get an exit event some time in the future. It is typically used
+ * when the script wants to simulate for a specific number of
+ * instructions rather than ticks.
+ *
+ * @param tid Thread monitor.
+ * @param insts Number of instructions into the future.
+ * @param cause Cause to signal in the exit event.
+ */
+ void scheduleInstStop(ThreadID tid, Counter insts, const char *cause);
+
+ /**
+ * Schedule an event that exits the simulation loops after a
+ * predefined number of load operations.
+ *
+ * This method is usually called from the configuration script to
+ * get an exit event some time in the future. It is typically used
+ * when the script wants to simulate for a specific number of
+ * loads rather than ticks.
+ *
+ * @param tid Thread monitor.
+ * @param loads Number of load instructions into the future.
+ * @param cause Cause to signal in the exit event.
+ */
+ void scheduleLoadStop(ThreadID tid, Counter loads, const char *cause);
+
+ /**
+ * Get the number of instructions executed by the specified thread
+ * on this CPU. Used by Python to control simulation.
+ *
+ * @param tid Thread monitor
+ * @return Number of instructions executed
+ */
+ uint64_t getCurrentInstCount(ThreadID tid);
+
+ public:
+ /**
+ * @{
+ * @name PMU Probe points.
+ */
+
+ /**
+ * Helper method to trigger PMU probes for a committed
+ * instruction.
+ *
+ * @param inst Instruction that just committed
+ */
+ virtual void probeInstCommit(const StaticInstPtr &inst);
+
+ /**
+ * Helper method to instantiate probe points belonging to this
+ * object.
+ *
+ * @param name Name of the probe point.
+ * @return A unique_ptr to the new probe point.
+ */
+ ProbePoints::PMUUPtr pmuProbePoint(const char *name);
+
+ /** CPU cycle counter */
+ ProbePoints::PMUUPtr ppCycles;
+
+ /**
+ * Instruction commit probe point.
+ *
+ * This probe point is triggered whenever one or more instructions
+ * are committed. It is normally triggered once for every
+ * instruction. However, CPU models committing bundles of
+ * instructions may call notify once for the entire bundle.
+ */
+ ProbePoints::PMUUPtr ppRetiredInsts;
+
+ /** Retired load instructions */
+ ProbePoints::PMUUPtr ppRetiredLoads;
+ /** Retired store instructions */
+ ProbePoints::PMUUPtr ppRetiredStores;
+
+ /** Retired branches (any type) */
+ ProbePoints::PMUUPtr ppRetiredBranches;
+
+ /** @} */
+
+
+
// Function tracing
private:
bool functionTracingEnabled;
Stats::Scalar numCycles;
Stats::Scalar numWorkItemsStarted;
Stats::Scalar numWorkItemsCompleted;
+
+ private:
+ std::vector<AddressMonitor> addressMonitor;
+
+ public:
+ void armMonitor(ThreadID tid, Addr address);
+ bool mwait(ThreadID tid, PacketPtr pkt);
+ void mwaitAtomic(ThreadID tid, ThreadContext *tc, TheISA::TLB *dtb);
+ AddressMonitor *getCpuAddrMonitor(ThreadID tid)
+ {
+ assert(tid < numThreads);
+ return &addressMonitor[tid];
+ }
+
+ Cycles syscallRetryLatency;
};
+#endif // THE_ISA == NULL_ISA
+
#endif // __CPU_BASE_HH__