lsq_unit.hh revision 14083
1/*
2 * Copyright (c) 2012-2014,2017-2018 ARM Limited
3 * All rights reserved
4 *
5 * The license below extends only to copyright in the software and shall
6 * not be construed as granting a license to any other intellectual
7 * property including but not limited to intellectual property relating
8 * to a hardware implementation of the functionality of the software
9 * licensed hereunder.  You may use the software subject to the license
10 * terms below provided that you ensure that this notice is replicated
11 * unmodified and in its entirety in all distributions of the software,
12 * modified or unmodified, in source code or in binary form.
13 *
14 * Copyright (c) 2004-2006 The Regents of The University of Michigan
15 * Copyright (c) 2013 Advanced Micro Devices, Inc.
16 * All rights reserved.
17 *
18 * Redistribution and use in source and binary forms, with or without
19 * modification, are permitted provided that the following conditions are
20 * met: redistributions of source code must retain the above copyright
21 * notice, this list of conditions and the following disclaimer;
22 * redistributions in binary form must reproduce the above copyright
23 * notice, this list of conditions and the following disclaimer in the
24 * documentation and/or other materials provided with the distribution;
25 * neither the name of the copyright holders nor the names of its
26 * contributors may be used to endorse or promote products derived from
27 * this software without specific prior written permission.
28 *
29 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
30 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
31 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
32 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
33 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
34 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
35 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
36 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
37 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
38 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
39 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
40 *
41 * Authors: Kevin Lim
42 *          Korey Sewell
43 */
44
45#ifndef __CPU_O3_LSQ_UNIT_HH__
46#define __CPU_O3_LSQ_UNIT_HH__
47
48#include <algorithm>
49#include <cstring>
50#include <map>
51#include <queue>
52
53#include "arch/generic/debugfaults.hh"
54#include "arch/generic/vec_reg.hh"
55#include "arch/isa_traits.hh"
56#include "arch/locked_mem.hh"
57#include "arch/mmapped_ipr.hh"
58#include "config/the_isa.hh"
59#include "cpu/inst_seq.hh"
60#include "cpu/timebuf.hh"
61#include "debug/LSQUnit.hh"
62#include "mem/packet.hh"
63#include "mem/port.hh"
64
65struct DerivO3CPUParams;
66#include "base/circular_queue.hh"
67
68/**
69 * Class that implements the actual LQ and SQ for each specific
70 * thread.  Both are circular queues; load entries are freed upon
71 * committing, while store entries are freed once they writeback. The
72 * LSQUnit tracks if there are memory ordering violations, and also
73 * detects partial load to store forwarding cases (a store only has
74 * part of a load's data) that requires the load to wait until the
75 * store writes back. In the former case it holds onto the instruction
76 * until the dependence unit looks at it, and in the latter it stalls
77 * the LSQ until the store writes back. At that point the load is
78 * replayed.
79 */
80template <class Impl>
81class LSQUnit
82{
83  public:
84    static constexpr auto MaxDataBytes = MaxVecRegLenInBytes;
85
86    typedef typename Impl::O3CPU O3CPU;
87    typedef typename Impl::DynInstPtr DynInstPtr;
88    typedef typename Impl::CPUPol::IEW IEW;
89    typedef typename Impl::CPUPol::LSQ LSQ;
90    typedef typename Impl::CPUPol::IssueStruct IssueStruct;
91
92    using LSQSenderState = typename LSQ::LSQSenderState;
93    using LSQRequest = typename Impl::CPUPol::LSQ::LSQRequest;
94  private:
95    class LSQEntry
96    {
97      private:
98        /** The instruction. */
99        DynInstPtr inst;
100        /** The request. */
101        LSQRequest* req;
102        /** The size of the operation. */
103        uint32_t _size;
104        /** Valid entry. */
105        bool _valid;
106      public:
107        /** Constructs an empty store queue entry. */
108        LSQEntry()
109            : inst(nullptr), req(nullptr), _size(0), _valid(false)
110        {
111        }
112
113        ~LSQEntry()
114        {
115            inst = nullptr;
116            if (req != nullptr) {
117                req->freeLSQEntry();
118                req = nullptr;
119            }
120        }
121
122        void
123        clear()
124        {
125            inst = nullptr;
126            if (req != nullptr) {
127                req->freeLSQEntry();
128            }
129            req = nullptr;
130            _valid = false;
131            _size = 0;
132        }
133
134        void
135        set(const DynInstPtr& inst)
136        {
137            assert(!_valid);
138            this->inst = inst;
139            _valid = true;
140            _size = 0;
141        }
142        LSQRequest* request() { return req; }
143        void setRequest(LSQRequest* r) { req = r; }
144        bool hasRequest() { return req != nullptr; }
145        /** Member accessors. */
146        /** @{ */
147        bool valid() const { return _valid; }
148        uint32_t& size() { return _size; }
149        const uint32_t& size() const { return _size; }
150        const DynInstPtr& instruction() const { return inst; }
151        /** @} */
152    };
153
154    class SQEntry : public LSQEntry
155    {
156      private:
157        /** The store data. */
158        char _data[MaxDataBytes];
159        /** Whether or not the store can writeback. */
160        bool _canWB;
161        /** Whether or not the store is committed. */
162        bool _committed;
163        /** Whether or not the store is completed. */
164        bool _completed;
165        /** Does this request write all zeros and thus doesn't
166         * have any data attached to it. Used for cache block zero
167         * style instructs (ARM DC ZVA; ALPHA WH64)
168         */
169        bool _isAllZeros;
170      public:
171        static constexpr size_t DataSize = sizeof(_data);
172        /** Constructs an empty store queue entry. */
173        SQEntry()
174            : _canWB(false), _committed(false), _completed(false),
175              _isAllZeros(false)
176        {
177            std::memset(_data, 0, DataSize);
178        }
179
180        ~SQEntry()
181        {
182        }
183
184        void
185        set(const DynInstPtr& inst)
186        {
187            LSQEntry::set(inst);
188        }
189
190        void
191        clear()
192        {
193            LSQEntry::clear();
194            _canWB = _completed = _committed = _isAllZeros = false;
195        }
196        /** Member accessors. */
197        /** @{ */
198        bool& canWB() { return _canWB; }
199        const bool& canWB() const { return _canWB; }
200        bool& completed() { return _completed; }
201        const bool& completed() const { return _completed; }
202        bool& committed() { return _committed; }
203        const bool& committed() const { return _committed; }
204        bool& isAllZeros() { return _isAllZeros; }
205        const bool& isAllZeros() const { return _isAllZeros; }
206        char* data() { return _data; }
207        const char* data() const { return _data; }
208        /** @} */
209    };
210    using LQEntry = LSQEntry;
211
212  public:
213    using LoadQueue = CircularQueue<LQEntry>;
214    using StoreQueue = CircularQueue<SQEntry>;
215
216  public:
217    /** Constructs an LSQ unit. init() must be called prior to use. */
218    LSQUnit(uint32_t lqEntries, uint32_t sqEntries);
219
220    /** We cannot copy LSQUnit because it has stats for which copy
221     * contructor is deleted explicitly. However, STL vector requires
222     * a valid copy constructor for the base type at compile time.
223     */
224    LSQUnit(const LSQUnit &l) { panic("LSQUnit is not copy-able"); }
225
226    /** Initializes the LSQ unit with the specified number of entries. */
227    void init(O3CPU *cpu_ptr, IEW *iew_ptr, DerivO3CPUParams *params,
228            LSQ *lsq_ptr, unsigned id);
229
230    /** Returns the name of the LSQ unit. */
231    std::string name() const;
232
233    /** Registers statistics. */
234    void regStats();
235
236    /** Sets the pointer to the dcache port. */
237    void setDcachePort(MasterPort *dcache_port);
238
239    /** Perform sanity checks after a drain. */
240    void drainSanityCheck() const;
241
242    /** Takes over from another CPU's thread. */
243    void takeOverFrom();
244
245    /** Inserts an instruction. */
246    void insert(const DynInstPtr &inst);
247    /** Inserts a load instruction. */
248    void insertLoad(const DynInstPtr &load_inst);
249    /** Inserts a store instruction. */
250    void insertStore(const DynInstPtr &store_inst);
251
252    /** Check for ordering violations in the LSQ. For a store squash if we
253     * ever find a conflicting load. For a load, only squash if we
254     * an external snoop invalidate has been seen for that load address
255     * @param load_idx index to start checking at
256     * @param inst the instruction to check
257     */
258    Fault checkViolations(typename LoadQueue::iterator& loadIt,
259            const DynInstPtr& inst);
260
261    /** Check if an incoming invalidate hits in the lsq on a load
262     * that might have issued out of order wrt another load beacuse
263     * of the intermediate invalidate.
264     */
265    void checkSnoop(PacketPtr pkt);
266
267    /** Executes a load instruction. */
268    Fault executeLoad(const DynInstPtr &inst);
269
270    Fault executeLoad(int lq_idx) { panic("Not implemented"); return NoFault; }
271    /** Executes a store instruction. */
272    Fault executeStore(const DynInstPtr &inst);
273
274    /** Commits the head load. */
275    void commitLoad();
276    /** Commits loads older than a specific sequence number. */
277    void commitLoads(InstSeqNum &youngest_inst);
278
279    /** Commits stores older than a specific sequence number. */
280    void commitStores(InstSeqNum &youngest_inst);
281
282    /** Writes back stores. */
283    void writebackStores();
284
285    /** Completes the data access that has been returned from the
286     * memory system. */
287    void completeDataAccess(PacketPtr pkt);
288
289    /** Squashes all instructions younger than a specific sequence number. */
290    void squash(const InstSeqNum &squashed_num);
291
292    /** Returns if there is a memory ordering violation. Value is reset upon
293     * call to getMemDepViolator().
294     */
295    bool violation() { return memDepViolator; }
296
297    /** Returns the memory ordering violator. */
298    DynInstPtr getMemDepViolator();
299
300    /** Returns the number of free LQ entries. */
301    unsigned numFreeLoadEntries();
302
303    /** Returns the number of free SQ entries. */
304    unsigned numFreeStoreEntries();
305
306    /** Returns the number of loads in the LQ. */
307    int numLoads() { return loads; }
308
309    /** Returns the number of stores in the SQ. */
310    int numStores() { return stores; }
311
312    /** Returns if either the LQ or SQ is full. */
313    bool isFull() { return lqFull() || sqFull(); }
314
315    /** Returns if both the LQ and SQ are empty. */
316    bool isEmpty() const { return lqEmpty() && sqEmpty(); }
317
318    /** Returns if the LQ is full. */
319    bool lqFull() { return loadQueue.full(); }
320
321    /** Returns if the SQ is full. */
322    bool sqFull() { return storeQueue.full(); }
323
324    /** Returns if the LQ is empty. */
325    bool lqEmpty() const { return loads == 0; }
326
327    /** Returns if the SQ is empty. */
328    bool sqEmpty() const { return stores == 0; }
329
330    /** Returns the number of instructions in the LSQ. */
331    unsigned getCount() { return loads + stores; }
332
333    /** Returns if there are any stores to writeback. */
334    bool hasStoresToWB() { return storesToWB; }
335
336    /** Returns the number of stores to writeback. */
337    int numStoresToWB() { return storesToWB; }
338
339    /** Returns if the LSQ unit will writeback on this cycle. */
340    bool
341    willWB()
342    {
343        return storeWBIt.dereferenceable() &&
344                        storeWBIt->valid() &&
345                        storeWBIt->canWB() &&
346                        !storeWBIt->completed() &&
347                        !isStoreBlocked;
348    }
349
350    /** Handles doing the retry. */
351    void recvRetry();
352
353    unsigned int cacheLineSize();
354  private:
355    /** Reset the LSQ state */
356    void resetState();
357
358    /** Writes back the instruction, sending it to IEW. */
359    void writeback(const DynInstPtr &inst, PacketPtr pkt);
360
361    /** Try to finish a previously blocked write back attempt */
362    void writebackBlockedStore();
363
364    /** Completes the store at the specified index. */
365    void completeStore(typename StoreQueue::iterator store_idx);
366
367    /** Handles completing the send of a store to memory. */
368    void storePostSend();
369
370  public:
371    /** Attempts to send a packet to the cache.
372     * Check if there are ports available. Return true if
373     * there are, false if there are not.
374     */
375    bool trySendPacket(bool isLoad, PacketPtr data_pkt);
376
377
378    /** Debugging function to dump instructions in the LSQ. */
379    void dumpInsts() const;
380
381    /** Schedule event for the cpu. */
382    void schedule(Event& ev, Tick when) { cpu->schedule(ev, when); }
383
384    BaseTLB* dTLB() { return cpu->dtb; }
385
386  private:
387    /** Pointer to the CPU. */
388    O3CPU *cpu;
389
390    /** Pointer to the IEW stage. */
391    IEW *iewStage;
392
393    /** Pointer to the LSQ. */
394    LSQ *lsq;
395
396    /** Pointer to the dcache port.  Used only for sending. */
397    MasterPort *dcachePort;
398
399    /** Particularisation of the LSQSenderState to the LQ. */
400    class LQSenderState : public LSQSenderState
401    {
402        using LSQSenderState::alive;
403      public:
404        LQSenderState(typename LoadQueue::iterator idx_)
405            : LSQSenderState(idx_->request(), true), idx(idx_) { }
406
407        /** The LQ index of the instruction. */
408        typename LoadQueue::iterator idx;
409        //virtual LSQRequest* request() { return idx->request(); }
410        virtual void
411        complete()
412        {
413            //if (alive())
414            //  idx->request()->senderState(nullptr);
415        }
416    };
417
418    /** Particularisation of the LSQSenderState to the SQ. */
419    class SQSenderState : public LSQSenderState
420    {
421        using LSQSenderState::alive;
422      public:
423        SQSenderState(typename StoreQueue::iterator idx_)
424            : LSQSenderState(idx_->request(), false), idx(idx_) { }
425        /** The SQ index of the instruction. */
426        typename StoreQueue::iterator idx;
427        //virtual LSQRequest* request() { return idx->request(); }
428        virtual void
429        complete()
430        {
431            //if (alive())
432            //   idx->request()->senderState(nullptr);
433        }
434    };
435
436    /** Writeback event, specifically for when stores forward data to loads. */
437    class WritebackEvent : public Event
438    {
439      public:
440        /** Constructs a writeback event. */
441        WritebackEvent(const DynInstPtr &_inst, PacketPtr pkt,
442                LSQUnit *lsq_ptr);
443
444        /** Processes the writeback event. */
445        void process();
446
447        /** Returns the description of this event. */
448        const char *description() const;
449
450      private:
451        /** Instruction whose results are being written back. */
452        DynInstPtr inst;
453
454        /** The packet that would have been sent to memory. */
455        PacketPtr pkt;
456
457        /** The pointer to the LSQ unit that issued the store. */
458        LSQUnit<Impl> *lsqPtr;
459    };
460
461  public:
462    /**
463     * Handles writing back and completing the load or store that has
464     * returned from memory.
465     *
466     * @param pkt Response packet from the memory sub-system
467     */
468    bool recvTimingResp(PacketPtr pkt);
469
470  private:
471    /** The LSQUnit thread id. */
472    ThreadID lsqID;
473  public:
474    /** The store queue. */
475    CircularQueue<SQEntry> storeQueue;
476
477    /** The load queue. */
478    LoadQueue loadQueue;
479
480  private:
481    /** The number of places to shift addresses in the LSQ before checking
482     * for dependency violations
483     */
484    unsigned depCheckShift;
485
486    /** Should loads be checked for dependency issues */
487    bool checkLoads;
488
489    /** The number of load instructions in the LQ. */
490    int loads;
491    /** The number of store instructions in the SQ. */
492    int stores;
493    /** The number of store instructions in the SQ waiting to writeback. */
494    int storesToWB;
495
496    /** The index of the first instruction that may be ready to be
497     * written back, and has not yet been written back.
498     */
499    typename StoreQueue::iterator storeWBIt;
500
501    /** Address Mask for a cache block (e.g. ~(cache_block_size-1)) */
502    Addr cacheBlockMask;
503
504    /** Wire to read information from the issue stage time queue. */
505    typename TimeBuffer<IssueStruct>::wire fromIssue;
506
507    /** Whether or not the LSQ is stalled. */
508    bool stalled;
509    /** The store that causes the stall due to partial store to load
510     * forwarding.
511     */
512    InstSeqNum stallingStoreIsn;
513    /** The index of the above store. */
514    int stallingLoadIdx;
515
516    /** The packet that needs to be retried. */
517    PacketPtr retryPkt;
518
519    /** Whehter or not a store is blocked due to the memory system. */
520    bool isStoreBlocked;
521
522    /** Whether or not a store is in flight. */
523    bool storeInFlight;
524
525    /** The oldest load that caused a memory ordering violation. */
526    DynInstPtr memDepViolator;
527
528    /** Whether or not there is a packet that couldn't be sent because of
529     * a lack of cache ports. */
530    bool hasPendingRequest;
531
532    /** The packet that is pending free cache ports. */
533    LSQRequest* pendingRequest;
534
535    /** Flag for memory model. */
536    bool needsTSO;
537
538    // Will also need how many read/write ports the Dcache has.  Or keep track
539    // of that in stage that is one level up, and only call executeLoad/Store
540    // the appropriate number of times.
541    /** Total number of loads forwaded from LSQ stores. */
542    Stats::Scalar lsqForwLoads;
543
544    /** Total number of loads ignored due to invalid addresses. */
545    Stats::Scalar invAddrLoads;
546
547    /** Total number of squashed loads. */
548    Stats::Scalar lsqSquashedLoads;
549
550    /** Total number of responses from the memory system that are
551     * ignored due to the instruction already being squashed. */
552    Stats::Scalar lsqIgnoredResponses;
553
554    /** Tota number of memory ordering violations. */
555    Stats::Scalar lsqMemOrderViolation;
556
557    /** Total number of squashed stores. */
558    Stats::Scalar lsqSquashedStores;
559
560    /** Total number of software prefetches ignored due to invalid addresses. */
561    Stats::Scalar invAddrSwpfs;
562
563    /** Ready loads blocked due to partial store-forwarding. */
564    Stats::Scalar lsqBlockedLoads;
565
566    /** Number of loads that were rescheduled. */
567    Stats::Scalar lsqRescheduledLoads;
568
569    /** Number of times the LSQ is blocked due to the cache. */
570    Stats::Scalar lsqCacheBlocked;
571
572  public:
573    /** Executes the load at the given index. */
574    Fault read(LSQRequest *req, int load_idx);
575
576    /** Executes the store at the given index. */
577    Fault write(LSQRequest *req, uint8_t *data, int store_idx);
578
579    /** Returns the index of the head load instruction. */
580    int getLoadHead() { return loadQueue.head(); }
581
582    /** Returns the sequence number of the head load instruction. */
583    InstSeqNum
584    getLoadHeadSeqNum()
585    {
586        return loadQueue.front().valid()
587            ? loadQueue.front().instruction()->seqNum
588            : 0;
589    }
590
591    /** Returns the index of the head store instruction. */
592    int getStoreHead() { return storeQueue.head(); }
593    /** Returns the sequence number of the head store instruction. */
594    InstSeqNum
595    getStoreHeadSeqNum()
596    {
597        return storeQueue.front().valid()
598            ? storeQueue.front().instruction()->seqNum
599            : 0;
600    }
601
602    /** Returns whether or not the LSQ unit is stalled. */
603    bool isStalled()  { return stalled; }
604  public:
605    typedef typename CircularQueue<LQEntry>::iterator LQIterator;
606    typedef typename CircularQueue<SQEntry>::iterator SQIterator;
607    typedef CircularQueue<LQEntry> LQueue;
608    typedef CircularQueue<SQEntry> SQueue;
609};
610
611template <class Impl>
612Fault
613LSQUnit<Impl>::read(LSQRequest *req, int load_idx)
614{
615    LQEntry& load_req = loadQueue[load_idx];
616    const DynInstPtr& load_inst = load_req.instruction();
617
618    load_req.setRequest(req);
619    assert(load_inst);
620
621    assert(!load_inst->isExecuted());
622
623    // Make sure this isn't a strictly ordered load
624    // A bit of a hackish way to get strictly ordered accesses to work
625    // only if they're at the head of the LSQ and are ready to commit
626    // (at the head of the ROB too).
627
628    if (req->mainRequest()->isStrictlyOrdered() &&
629        (load_idx != loadQueue.head() || !load_inst->isAtCommit())) {
630        // Tell IQ/mem dep unit that this instruction will need to be
631        // rescheduled eventually
632        iewStage->rescheduleMemInst(load_inst);
633        load_inst->clearIssued();
634        load_inst->effAddrValid(false);
635        ++lsqRescheduledLoads;
636        DPRINTF(LSQUnit, "Strictly ordered load [sn:%lli] PC %s\n",
637                load_inst->seqNum, load_inst->pcState());
638
639        // Must delete request now that it wasn't handed off to
640        // memory.  This is quite ugly.  @todo: Figure out the proper
641        // place to really handle request deletes.
642        load_req.setRequest(nullptr);
643        req->discard();
644        return std::make_shared<GenericISA::M5PanicFault>(
645            "Strictly ordered load [sn:%llx] PC %s\n",
646            load_inst->seqNum, load_inst->pcState());
647    }
648
649    DPRINTF(LSQUnit, "Read called, load idx: %i, store idx: %i, "
650            "storeHead: %i addr: %#x%s\n",
651            load_idx - 1, load_inst->sqIt._idx, storeQueue.head() - 1,
652            req->mainRequest()->getPaddr(), req->isSplit() ? " split" : "");
653
654    if (req->mainRequest()->isLLSC()) {
655        // Disable recording the result temporarily.  Writing to misc
656        // regs normally updates the result, but this is not the
657        // desired behavior when handling store conditionals.
658        load_inst->recordResult(false);
659        TheISA::handleLockedRead(load_inst.get(), req->mainRequest());
660        load_inst->recordResult(true);
661    }
662
663    if (req->mainRequest()->isMmappedIpr()) {
664        assert(!load_inst->memData);
665        load_inst->memData = new uint8_t[MaxDataBytes];
666
667        ThreadContext *thread = cpu->tcBase(lsqID);
668        PacketPtr main_pkt = new Packet(req->mainRequest(), MemCmd::ReadReq);
669
670        main_pkt->dataStatic(load_inst->memData);
671
672        Cycles delay = req->handleIprRead(thread, main_pkt);
673
674        WritebackEvent *wb = new WritebackEvent(load_inst, main_pkt, this);
675        cpu->schedule(wb, cpu->clockEdge(delay));
676        return NoFault;
677    }
678
679    // Check the SQ for any previous stores that might lead to forwarding
680    auto store_it = load_inst->sqIt;
681    assert (store_it >= storeWBIt);
682    // End once we've reached the top of the LSQ
683    while (store_it != storeWBIt) {
684        // Move the index to one younger
685        store_it--;
686        assert(store_it->valid());
687        assert(store_it->instruction()->seqNum < load_inst->seqNum);
688        int store_size = store_it->size();
689
690        // Cache maintenance instructions go down via the store
691        // path but they carry no data and they shouldn't be
692        // considered for forwarding
693        if (store_size != 0 && !store_it->instruction()->strictlyOrdered() &&
694            !(store_it->request()->mainRequest() &&
695              store_it->request()->mainRequest()->isCacheMaintenance())) {
696            assert(store_it->instruction()->effAddrValid());
697
698            // Check if the store data is within the lower and upper bounds of
699            // addresses that the request needs.
700            auto req_s = req->mainRequest()->getVaddr();
701            auto req_e = req_s + req->mainRequest()->getSize();
702            auto st_s = store_it->instruction()->effAddr;
703            auto st_e = st_s + store_size;
704
705            bool store_has_lower_limit = req_s >= st_s;
706            bool store_has_upper_limit = req_e <= st_e;
707            bool lower_load_has_store_part = req_s < st_e;
708            bool upper_load_has_store_part = req_e > st_s;
709
710            // If the store entry is not atomic (atomic does not have valid
711            // data), the store has all of the data needed, and
712            // the load is not LLSC, then
713            // we can forward data from the store to the load
714            if (!store_it->instruction()->isAtomic() &&
715                store_has_lower_limit && store_has_upper_limit &&
716                !req->mainRequest()->isLLSC()) {
717
718                // Get shift amount for offset into the store's data.
719                int shift_amt = req->mainRequest()->getVaddr() -
720                    store_it->instruction()->effAddr;
721
722                // Allocate memory if this is the first time a load is issued.
723                if (!load_inst->memData) {
724                    load_inst->memData =
725                        new uint8_t[req->mainRequest()->getSize()];
726                }
727                if (store_it->isAllZeros())
728                    memset(load_inst->memData, 0,
729                            req->mainRequest()->getSize());
730                else
731                    memcpy(load_inst->memData,
732                        store_it->data() + shift_amt,
733                        req->mainRequest()->getSize());
734
735                DPRINTF(LSQUnit, "Forwarding from store idx %i to load to "
736                        "addr %#x\n", store_it._idx,
737                        req->mainRequest()->getVaddr());
738
739                PacketPtr data_pkt = new Packet(req->mainRequest(),
740                        MemCmd::ReadReq);
741                data_pkt->dataStatic(load_inst->memData);
742
743                if (req->isAnyOutstandingRequest()) {
744                    assert(req->_numOutstandingPackets > 0);
745                    // There are memory requests packets in flight already.
746                    // This may happen if the store was not complete the
747                    // first time this load got executed. Signal the senderSate
748                    // that response packets should be discarded.
749                    req->discardSenderState();
750                }
751
752                WritebackEvent *wb = new WritebackEvent(load_inst, data_pkt,
753                        this);
754
755                // We'll say this has a 1 cycle load-store forwarding latency
756                // for now.
757                // @todo: Need to make this a parameter.
758                cpu->schedule(wb, curTick());
759
760                // Don't need to do anything special for split loads.
761                ++lsqForwLoads;
762
763                return NoFault;
764            } else if (
765                // This is the partial store-load forwarding case where a store
766                // has only part of the load's data and the load isn't LLSC
767                (!req->mainRequest()->isLLSC() &&
768                 ((store_has_lower_limit && lower_load_has_store_part) ||
769                  (store_has_upper_limit && upper_load_has_store_part) ||
770                  (lower_load_has_store_part && upper_load_has_store_part))) ||
771                // The load is LLSC, and the store has all or part of the
772                // load's data
773                (req->mainRequest()->isLLSC() &&
774                 ((store_has_lower_limit || upper_load_has_store_part) &&
775                  (store_has_upper_limit || lower_load_has_store_part))) ||
776                // The store entry is atomic and has all or part of the load's
777                // data
778                (store_it->instruction()->isAtomic() &&
779                 ((store_has_lower_limit || upper_load_has_store_part) &&
780                  (store_has_upper_limit || lower_load_has_store_part)))) {
781
782                // If it's already been written back, then don't worry about
783                // stalling on it.
784                if (store_it->completed()) {
785                    panic("Should not check one of these");
786                    continue;
787                }
788
789                // Must stall load and force it to retry, so long as it's the
790                // oldest load that needs to do so.
791                if (!stalled ||
792                    (stalled &&
793                     load_inst->seqNum <
794                     loadQueue[stallingLoadIdx].instruction()->seqNum)) {
795                    stalled = true;
796                    stallingStoreIsn = store_it->instruction()->seqNum;
797                    stallingLoadIdx = load_idx;
798                }
799
800                // Tell IQ/mem dep unit that this instruction will need to be
801                // rescheduled eventually
802                iewStage->rescheduleMemInst(load_inst);
803                load_inst->clearIssued();
804                load_inst->effAddrValid(false);
805                ++lsqRescheduledLoads;
806
807                // Do not generate a writeback event as this instruction is not
808                // complete.
809                DPRINTF(LSQUnit, "Load-store forwarding mis-match. "
810                        "Store idx %i to load addr %#x\n",
811                        store_it._idx, req->mainRequest()->getVaddr());
812
813                // Must discard the request.
814                req->discard();
815                load_req.setRequest(nullptr);
816                return NoFault;
817            }
818        }
819    }
820
821    // If there's no forwarding case, then go access memory
822    DPRINTF(LSQUnit, "Doing memory access for inst [sn:%lli] PC %s\n",
823            load_inst->seqNum, load_inst->pcState());
824
825    // Allocate memory if this is the first time a load is issued.
826    if (!load_inst->memData) {
827        load_inst->memData = new uint8_t[req->mainRequest()->getSize()];
828    }
829
830    // For now, load throughput is constrained by the number of
831    // load FUs only, and loads do not consume a cache port (only
832    // stores do).
833    // @todo We should account for cache port contention
834    // and arbitrate between loads and stores.
835
836    // if we the cache is not blocked, do cache access
837    if (req->senderState() == nullptr) {
838        LQSenderState *state = new LQSenderState(
839                loadQueue.getIterator(load_idx));
840        state->isLoad = true;
841        state->inst = load_inst;
842        state->isSplit = req->isSplit();
843        req->senderState(state);
844    }
845    req->buildPackets();
846    req->sendPacketToCache();
847    if (!req->isSent())
848        iewStage->blockMemInst(load_inst);
849
850    return NoFault;
851}
852
853template <class Impl>
854Fault
855LSQUnit<Impl>::write(LSQRequest *req, uint8_t *data, int store_idx)
856{
857    assert(storeQueue[store_idx].valid());
858
859    DPRINTF(LSQUnit, "Doing write to store idx %i, addr %#x | storeHead:%i "
860            "[sn:%llu]\n",
861            store_idx - 1, req->request()->getPaddr(), storeQueue.head() - 1,
862            storeQueue[store_idx].instruction()->seqNum);
863
864    storeQueue[store_idx].setRequest(req);
865    unsigned size = req->_size;
866    storeQueue[store_idx].size() = size;
867    bool store_no_data =
868        req->mainRequest()->getFlags() & Request::STORE_NO_DATA;
869    storeQueue[store_idx].isAllZeros() = store_no_data;
870    assert(size <= SQEntry::DataSize || store_no_data);
871
872    // copy data into the storeQueue only if the store request has valid data
873    if (!(req->request()->getFlags() & Request::CACHE_BLOCK_ZERO) &&
874        !req->request()->isCacheMaintenance() &&
875        !req->request()->isAtomic())
876        memcpy(storeQueue[store_idx].data(), data, size);
877
878    // This function only writes the data to the store queue, so no fault
879    // can happen here.
880    return NoFault;
881}
882
883#endif // __CPU_O3_LSQ_UNIT_HH__
884