commit.hh revision 2980:eab855f06b79
1/*
2 * Copyright (c) 2004-2006 The Regents of The University of Michigan
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions are
7 * met: redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer;
9 * redistributions in binary form must reproduce the above copyright
10 * notice, this list of conditions and the following disclaimer in the
11 * documentation and/or other materials provided with the distribution;
12 * neither the name of the copyright holders nor the names of its
13 * contributors may be used to endorse or promote products derived from
14 * this software without specific prior written permission.
15 *
16 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
17 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
18 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
19 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
20 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
21 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
22 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
26 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27 *
28 * Authors: Kevin Lim
29 *          Korey Sewell
30 */
31
32#ifndef __CPU_O3_COMMIT_HH__
33#define __CPU_O3_COMMIT_HH__
34
35#include "base/statistics.hh"
36#include "base/timebuf.hh"
37#include "cpu/exetrace.hh"
38#include "cpu/inst_seq.hh"
39
40template <class>
41class O3ThreadState;
42
43/**
44 * DefaultCommit handles single threaded and SMT commit. Its width is
45 * specified by the parameters; each cycle it tries to commit that
46 * many instructions. The SMT policy decides which thread it tries to
47 * commit instructions from. Non- speculative instructions must reach
48 * the head of the ROB before they are ready to execute; once they
49 * reach the head, commit will broadcast the instruction's sequence
50 * number to the previous stages so that they can issue/ execute the
51 * instruction. Only one non-speculative instruction is handled per
52 * cycle. Commit is responsible for handling all back-end initiated
53 * redirects.  It receives the redirect, and then broadcasts it to all
54 * stages, indicating the sequence number they should squash until,
55 * and any necessary branch misprediction information as well. It
56 * priortizes redirects by instruction's age, only broadcasting a
57 * redirect if it corresponds to an instruction that should currently
58 * be in the ROB. This is done by tracking the sequence number of the
59 * youngest instruction in the ROB, which gets updated to any
60 * squashing instruction's sequence number, and only broadcasting a
61 * redirect if it corresponds to an older instruction. Commit also
62 * supports multiple cycle squashing, to model a ROB that can only
63 * remove a certain number of instructions per cycle.
64 */
65template<class Impl>
66class DefaultCommit
67{
68  public:
69    // Typedefs from the Impl.
70    typedef typename Impl::O3CPU O3CPU;
71    typedef typename Impl::DynInstPtr DynInstPtr;
72    typedef typename Impl::Params Params;
73    typedef typename Impl::CPUPol CPUPol;
74
75    typedef typename CPUPol::RenameMap RenameMap;
76    typedef typename CPUPol::ROB ROB;
77
78    typedef typename CPUPol::TimeStruct TimeStruct;
79    typedef typename CPUPol::FetchStruct FetchStruct;
80    typedef typename CPUPol::IEWStruct IEWStruct;
81    typedef typename CPUPol::RenameStruct RenameStruct;
82
83    typedef typename CPUPol::Fetch Fetch;
84    typedef typename CPUPol::IEW IEW;
85
86    typedef O3ThreadState<Impl> Thread;
87
88    /** Event class used to schedule a squash due to a trap (fault or
89     * interrupt) to happen on a specific cycle.
90     */
91    class TrapEvent : public Event {
92      private:
93        DefaultCommit<Impl> *commit;
94        unsigned tid;
95
96      public:
97        TrapEvent(DefaultCommit<Impl> *_commit, unsigned _tid);
98
99        void process();
100        const char *description();
101    };
102
103    /** Overall commit status. Used to determine if the CPU can deschedule
104     * itself due to a lack of activity.
105     */
106    enum CommitStatus{
107        Active,
108        Inactive
109    };
110
111    /** Individual thread status. */
112    enum ThreadStatus {
113        Running,
114        Idle,
115        ROBSquashing,
116        TrapPending,
117        FetchTrapPending
118    };
119
120    /** Commit policy for SMT mode. */
121    enum CommitPolicy {
122        Aggressive,
123        RoundRobin,
124        OldestReady
125    };
126
127  private:
128    /** Overall commit status. */
129    CommitStatus _status;
130    /** Next commit status, to be set at the end of the cycle. */
131    CommitStatus _nextStatus;
132    /** Per-thread status. */
133    ThreadStatus commitStatus[Impl::MaxThreads];
134    /** Commit policy used in SMT mode. */
135    CommitPolicy commitPolicy;
136
137  public:
138    /** Construct a DefaultCommit with the given parameters. */
139    DefaultCommit(Params *params);
140
141    /** Returns the name of the DefaultCommit. */
142    std::string name() const;
143
144    /** Registers statistics. */
145    void regStats();
146
147    /** Sets the CPU pointer. */
148    void setCPU(O3CPU *cpu_ptr);
149
150    /** Sets the list of threads. */
151    void setThreads(std::vector<Thread *> &threads);
152
153    /** Sets the main time buffer pointer, used for backwards communication. */
154    void setTimeBuffer(TimeBuffer<TimeStruct> *tb_ptr);
155
156    void setFetchQueue(TimeBuffer<FetchStruct> *fq_ptr);
157
158    /** Sets the pointer to the queue coming from rename. */
159    void setRenameQueue(TimeBuffer<RenameStruct> *rq_ptr);
160
161    /** Sets the pointer to the queue coming from IEW. */
162    void setIEWQueue(TimeBuffer<IEWStruct> *iq_ptr);
163
164    /** Sets the pointer to the IEW stage. */
165    void setIEWStage(IEW *iew_stage);
166
167    /** Skid buffer between rename and commit. */
168    std::queue<DynInstPtr> skidBuffer;
169
170    /** The pointer to the IEW stage. Used solely to ensure that
171     * various events (traps, interrupts, syscalls) do not occur until
172     * all stores have written back.
173     */
174    IEW *iewStage;
175
176    /** Sets pointer to list of active threads. */
177    void setActiveThreads(std::list<unsigned> *at_ptr);
178
179    /** Sets pointer to the commited state rename map. */
180    void setRenameMap(RenameMap rm_ptr[Impl::MaxThreads]);
181
182    /** Sets pointer to the ROB. */
183    void setROB(ROB *rob_ptr);
184
185    /** Initializes stage by sending back the number of free entries. */
186    void initStage();
187
188    /** Initializes the draining of commit. */
189    bool drain();
190
191    /** Resumes execution after draining. */
192    void resume();
193
194    /** Completes the switch out of commit. */
195    void switchOut();
196
197    /** Takes over from another CPU's thread. */
198    void takeOverFrom();
199
200    /** Ticks the commit stage, which tries to commit instructions. */
201    void tick();
202
203    /** Handles any squashes that are sent from IEW, and adds instructions
204     * to the ROB and tries to commit instructions.
205     */
206    void commit();
207
208    /** Returns the number of free ROB entries for a specific thread. */
209    unsigned numROBFreeEntries(unsigned tid);
210
211    /** Generates an event to schedule a squash due to a trap. */
212    void generateTrapEvent(unsigned tid);
213
214    /** Records that commit needs to initiate a squash due to an
215     * external state update through the TC.
216     */
217    void generateTCEvent(unsigned tid);
218
219  private:
220    /** Updates the overall status of commit with the nextStatus, and
221     * tell the CPU if commit is active/inactive.
222     */
223    void updateStatus();
224
225    /** Sets the next status based on threads' statuses, which becomes the
226     * current status at the end of the cycle.
227     */
228    void setNextStatus();
229
230    /** Checks if the ROB is completed with squashing. This is for the case
231     * where the ROB can take multiple cycles to complete squashing.
232     */
233    bool robDoneSquashing();
234
235    /** Returns if any of the threads have the number of ROB entries changed
236     * on this cycle. Used to determine if the number of free ROB entries needs
237     * to be sent back to previous stages.
238     */
239    bool changedROBEntries();
240
241    /** Squashes all in flight instructions. */
242    void squashAll(unsigned tid);
243
244    /** Handles squashing due to a trap. */
245    void squashFromTrap(unsigned tid);
246
247    /** Handles squashing due to an TC write. */
248    void squashFromTC(unsigned tid);
249
250    /** Commits as many instructions as possible. */
251    void commitInsts();
252
253    /** Tries to commit the head ROB instruction passed in.
254     * @param head_inst The instruction to be committed.
255     */
256    bool commitHead(DynInstPtr &head_inst, unsigned inst_num);
257
258    /** Gets instructions from rename and inserts them into the ROB. */
259    void getInsts();
260
261    /** Insert all instructions from rename into skidBuffer */
262    void skidInsert();
263
264    /** Marks completed instructions using information sent from IEW. */
265    void markCompletedInsts();
266
267    /** Gets the thread to commit, based on the SMT policy. */
268    int getCommittingThread();
269
270    /** Returns the thread ID to use based on a round robin policy. */
271    int roundRobin();
272
273    /** Returns the thread ID to use based on an oldest instruction policy. */
274    int oldestReady();
275
276  public:
277    /** Returns the PC of the head instruction of the ROB.
278     * @todo: Probably remove this function as it returns only thread 0.
279     */
280    uint64_t readPC() { return PC[0]; }
281
282    /** Returns the PC of a specific thread. */
283    uint64_t readPC(unsigned tid) { return PC[tid]; }
284
285    /** Sets the PC of a specific thread. */
286    void setPC(uint64_t val, unsigned tid) { PC[tid] = val; }
287
288    /** Reads the next PC of a specific thread. */
289    uint64_t readNextPC(unsigned tid) { return nextPC[tid]; }
290
291    /** Sets the next PC of a specific thread. */
292    void setNextPC(uint64_t val, unsigned tid) { nextPC[tid] = val; }
293
294    /** Reads the next NPC of a specific thread. */
295    uint64_t readNextNPC(unsigned tid) { return nextNPC[tid]; }
296
297    /** Sets the next NPC of a specific thread. */
298    void setNextNPC(uint64_t val, unsigned tid) { nextNPC[tid] = val; }
299
300  private:
301    /** Time buffer interface. */
302    TimeBuffer<TimeStruct> *timeBuffer;
303
304    /** Wire to write information heading to previous stages. */
305    typename TimeBuffer<TimeStruct>::wire toIEW;
306
307    /** Wire to read information from IEW (for ROB). */
308    typename TimeBuffer<TimeStruct>::wire robInfoFromIEW;
309
310    TimeBuffer<FetchStruct> *fetchQueue;
311
312    typename TimeBuffer<FetchStruct>::wire fromFetch;
313
314    /** IEW instruction queue interface. */
315    TimeBuffer<IEWStruct> *iewQueue;
316
317    /** Wire to read information from IEW queue. */
318    typename TimeBuffer<IEWStruct>::wire fromIEW;
319
320    /** Rename instruction queue interface, for ROB. */
321    TimeBuffer<RenameStruct> *renameQueue;
322
323    /** Wire to read information from rename queue. */
324    typename TimeBuffer<RenameStruct>::wire fromRename;
325
326  public:
327    /** ROB interface. */
328    ROB *rob;
329
330  private:
331    /** Pointer to O3CPU. */
332    O3CPU *cpu;
333
334    /** Vector of all of the threads. */
335    std::vector<Thread *> thread;
336
337    /** Records that commit has written to the time buffer this cycle. Used for
338     * the CPU to determine if it can deschedule itself if there is no activity.
339     */
340    bool wroteToTimeBuffer;
341
342    /** Records if the number of ROB entries has changed this cycle. If it has,
343     * then the number of free entries must be re-broadcast.
344     */
345    bool changedROBNumEntries[Impl::MaxThreads];
346
347    /** A counter of how many threads are currently squashing. */
348    int squashCounter;
349
350    /** Records if a thread has to squash this cycle due to a trap. */
351    bool trapSquash[Impl::MaxThreads];
352
353    /** Records if a thread has to squash this cycle due to an XC write. */
354    bool tcSquash[Impl::MaxThreads];
355
356    /** Priority List used for Commit Policy */
357    std::list<unsigned> priority_list;
358
359    /** IEW to Commit delay, in ticks. */
360    unsigned iewToCommitDelay;
361
362    /** Commit to IEW delay, in ticks. */
363    unsigned commitToIEWDelay;
364
365    /** Rename to ROB delay, in ticks. */
366    unsigned renameToROBDelay;
367
368    unsigned fetchToCommitDelay;
369
370    /** Rename width, in instructions.  Used so ROB knows how many
371     *  instructions to get from the rename instruction queue.
372     */
373    unsigned renameWidth;
374
375    /** Commit width, in instructions. */
376    unsigned commitWidth;
377
378    /** Number of Reorder Buffers */
379    unsigned numRobs;
380
381    /** Number of Active Threads */
382    unsigned numThreads;
383
384    /** Is a drain pending. */
385    bool drainPending;
386
387    /** Is commit switched out. */
388    bool switchedOut;
389
390    /** The latency to handle a trap.  Used when scheduling trap
391     * squash event.
392     */
393    Tick trapLatency;
394
395    /** The commit PC of each thread.  Refers to the instruction that
396     * is currently being processed/committed.
397     */
398    Addr PC[Impl::MaxThreads];
399
400    /** The next PC of each thread. */
401    Addr nextPC[Impl::MaxThreads];
402
403    /** The next NPC of each thread. */
404    Addr nextNPC[Impl::MaxThreads];
405
406    /** The sequence number of the youngest valid instruction in the ROB. */
407    InstSeqNum youngestSeqNum[Impl::MaxThreads];
408
409    /** Pointer to the list of active threads. */
410    std::list<unsigned> *activeThreads;
411
412    /** Rename map interface. */
413    RenameMap *renameMap[Impl::MaxThreads];
414
415    /** Updates commit stats based on this instruction. */
416    void updateComInstStats(DynInstPtr &inst);
417
418    /** Stat for the total number of committed instructions. */
419    Stats::Scalar<> commitCommittedInsts;
420    /** Stat for the total number of squashed instructions discarded by commit.
421     */
422    Stats::Scalar<> commitSquashedInsts;
423    /** Stat for the total number of times commit is told to squash.
424     * @todo: Actually increment this stat.
425     */
426    Stats::Scalar<> commitSquashEvents;
427    /** Stat for the total number of times commit has had to stall due to a non-
428     * speculative instruction reaching the head of the ROB.
429     */
430    Stats::Scalar<> commitNonSpecStalls;
431    /** Stat for the total number of branch mispredicts that caused a squash. */
432    Stats::Scalar<> branchMispredicts;
433    /** Distribution of the number of committed instructions each cycle. */
434    Stats::Distribution<> numCommittedDist;
435
436    /** Total number of instructions committed. */
437    Stats::Vector<> statComInst;
438    /** Total number of software prefetches committed. */
439    Stats::Vector<> statComSwp;
440    /** Stat for the total number of committed memory references. */
441    Stats::Vector<> statComRefs;
442    /** Stat for the total number of committed loads. */
443    Stats::Vector<> statComLoads;
444    /** Total number of committed memory barriers. */
445    Stats::Vector<> statComMembars;
446    /** Total number of committed branches. */
447    Stats::Vector<> statComBranches;
448
449    /** Number of cycles where the commit bandwidth limit is reached. */
450    Stats::Scalar<> commitEligibleSamples;
451    /** Number of instructions not committed due to bandwidth limits. */
452    Stats::Vector<> commitEligible;
453};
454
455#endif // __CPU_O3_COMMIT_HH__
456