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_LSQ_UNIT_HH__
33#define __CPU_O3_LSQ_UNIT_HH__
34
35#include <algorithm>
36#include <cstring>
37#include <map>
38#include <queue>
39
40#include "arch/faults.hh"
41#include "arch/locked_mem.hh"
42#include "config/full_system.hh"
43#include "config/the_isa.hh"
44#include "base/fast_alloc.hh"
45#include "base/hashmap.hh"
46#include "cpu/inst_seq.hh"
47#include "mem/packet.hh"
48#include "mem/port.hh"
49
50class DerivO3CPUParams;
51
52/**
53 * Class that implements the actual LQ and SQ for each specific
54 * thread. Both are circular queues; load entries are freed upon
55 * committing, while store entries are freed once they writeback. The
56 * LSQUnit tracks if there are memory ordering violations, and also
57 * detects partial load to store forwarding cases (a store only has
58 * part of a load's data) that requires the load to wait until the
59 * store writes back. In the former case it holds onto the instruction
60 * until the dependence unit looks at it, and in the latter it stalls
61 * the LSQ until the store writes back. At that point the load is
62 * replayed.
63 */
64template <class Impl>
65class LSQUnit {
66 public:
67 typedef typename Impl::O3CPU O3CPU;
68 typedef typename Impl::DynInstPtr DynInstPtr;
69 typedef typename Impl::CPUPol::IEW IEW;
70 typedef typename Impl::CPUPol::LSQ LSQ;
71 typedef typename Impl::CPUPol::IssueStruct IssueStruct;
72
73 public:
74 /** Constructs an LSQ unit. init() must be called prior to use. */
75 LSQUnit();
76
77 /** Initializes the LSQ unit with the specified number of entries. */
78 void init(O3CPU *cpu_ptr, IEW *iew_ptr, DerivO3CPUParams *params,
79 LSQ *lsq_ptr, unsigned maxLQEntries, unsigned maxSQEntries,
80 unsigned id);
81
82 /** Returns the name of the LSQ unit. */
83 std::string name() const;
84
85 /** Registers statistics. */
86 void regStats();
87
88 /** Sets the pointer to the dcache port. */
89 void setDcachePort(Port *dcache_port);
90
91 /** Switches out LSQ unit. */
92 void switchOut();
93
94 /** Takes over from another CPU's thread. */
95 void takeOverFrom();
96
97 /** Returns if the LSQ is switched out. */
98 bool isSwitchedOut() { return switchedOut; }
99
100 /** Ticks the LSQ unit, which in this case only resets the number of
101 * used cache ports.
102 * @todo: Move the number of used ports up to the LSQ level so it can
103 * be shared by all LSQ units.
104 */
105 void tick() { usedPorts = 0; }
106
107 /** Inserts an instruction. */
108 void insert(DynInstPtr &inst);
109 /** Inserts a load instruction. */
110 void insertLoad(DynInstPtr &load_inst);
111 /** Inserts a store instruction. */
112 void insertStore(DynInstPtr &store_inst);
113
114 /** Executes a load instruction. */
115 Fault executeLoad(DynInstPtr &inst);
116
117 Fault executeLoad(int lq_idx) { panic("Not implemented"); return NoFault; }
118 /** Executes a store instruction. */
119 Fault executeStore(DynInstPtr &inst);
120
121 /** Commits the head load. */
122 void commitLoad();
123 /** Commits loads older than a specific sequence number. */
124 void commitLoads(InstSeqNum &youngest_inst);
125
126 /** Commits stores older than a specific sequence number. */
127 void commitStores(InstSeqNum &youngest_inst);
128
129 /** Writes back stores. */
130 void writebackStores();
131
132 /** Completes the data access that has been returned from the
133 * memory system. */
134 void completeDataAccess(PacketPtr pkt);
135
136 /** Clears all the entries in the LQ. */
137 void clearLQ();
138
139 /** Clears all the entries in the SQ. */
140 void clearSQ();
141
142 /** Resizes the LQ to a given size. */
143 void resizeLQ(unsigned size);
144
145 /** Resizes the SQ to a given size. */
146 void resizeSQ(unsigned size);
147
148 /** Squashes all instructions younger than a specific sequence number. */
149 void squash(const InstSeqNum &squashed_num);
150
151 /** Returns if there is a memory ordering violation. Value is reset upon
152 * call to getMemDepViolator().
153 */
154 bool violation() { return memDepViolator; }
155
156 /** Returns the memory ordering violator. */
157 DynInstPtr getMemDepViolator();
158
159 /** Returns if a load became blocked due to the memory system. */
160 bool loadBlocked()
161 { return isLoadBlocked; }
162
163 /** Clears the signal that a load became blocked. */
164 void clearLoadBlocked()
165 { isLoadBlocked = false; }
166
167 /** Returns if the blocked load was handled. */
168 bool isLoadBlockedHandled()
169 { return loadBlockedHandled; }
170
171 /** Records the blocked load as being handled. */
172 void setLoadBlockedHandled()
173 { loadBlockedHandled = true; }
174
175 /** Returns the number of free entries (min of free LQ and SQ entries). */
176 unsigned numFreeEntries();
177
178 /** Returns the number of loads ready to execute. */
179 int numLoadsReady();
180
181 /** Returns the number of loads in the LQ. */
182 int numLoads() { return loads; }
183
184 /** Returns the number of stores in the SQ. */
185 int numStores() { return stores; }
186
187 /** Returns if either the LQ or SQ is full. */
188 bool isFull() { return lqFull() || sqFull(); }
189
190 /** Returns if the LQ is full. */
191 bool lqFull() { return loads >= (LQEntries - 1); }
192
193 /** Returns if the SQ is full. */
194 bool sqFull() { return stores >= (SQEntries - 1); }
195
196 /** Returns the number of instructions in the LSQ. */
197 unsigned getCount() { return loads + stores; }
198
199 /** Returns if there are any stores to writeback. */
200 bool hasStoresToWB() { return storesToWB; }
201
202 /** Returns the number of stores to writeback. */
203 int numStoresToWB() { return storesToWB; }
204
205 /** Returns if the LSQ unit will writeback on this cycle. */
206 bool willWB() { return storeQueue[storeWBIdx].canWB &&
207 !storeQueue[storeWBIdx].completed &&
208 !isStoreBlocked; }
209
210 /** Handles doing the retry. */
211 void recvRetry();
212
213 private:
214 /** Writes back the instruction, sending it to IEW. */
215 void writeback(DynInstPtr &inst, PacketPtr pkt);
216
217 /** Writes back a store that couldn't be completed the previous cycle. */
218 void writebackPendingStore();
219
220 /** Handles completing the send of a store to memory. */
221 void storePostSend(PacketPtr pkt);
222
223 /** Completes the store at the specified index. */
224 void completeStore(int store_idx);
225
226 /** Attempts to send a store to the cache. */
227 bool sendStore(PacketPtr data_pkt);
228
229 /** Increments the given store index (circular queue). */
230 inline void incrStIdx(int &store_idx);
231 /** Decrements the given store index (circular queue). */
232 inline void decrStIdx(int &store_idx);
233 /** Increments the given load index (circular queue). */
234 inline void incrLdIdx(int &load_idx);
235 /** Decrements the given load index (circular queue). */
236 inline void decrLdIdx(int &load_idx);
237
238 public:
239 /** Debugging function to dump instructions in the LSQ. */
240 void dumpInsts();
241
242 private:
243 /** Pointer to the CPU. */
244 O3CPU *cpu;
245
246 /** Pointer to the IEW stage. */
247 IEW *iewStage;
248
249 /** Pointer to the LSQ. */
250 LSQ *lsq;
251
252 /** Pointer to the dcache port. Used only for sending. */
253 Port *dcachePort;
254
255 /** Derived class to hold any sender state the LSQ needs. */
256 class LSQSenderState : public Packet::SenderState, public FastAlloc
257 {
258 public:
259 /** Default constructor. */
260 LSQSenderState()
261 : noWB(false), isSplit(false), pktToSend(false), outstanding(1),
262 mainPkt(NULL), pendingPacket(NULL)
263 { }
264
265 /** Instruction who initiated the access to memory. */
266 DynInstPtr inst;
267 /** Whether or not it is a load. */
268 bool isLoad;
269 /** The LQ/SQ index of the instruction. */
270 int idx;
271 /** Whether or not the instruction will need to writeback. */
272 bool noWB;
273 /** Whether or not this access is split in two. */
274 bool isSplit;
275 /** Whether or not there is a packet that needs sending. */
276 bool pktToSend;
277 /** Number of outstanding packets to complete. */
278 int outstanding;
279 /** The main packet from a split load, used during writeback. */
280 PacketPtr mainPkt;
281 /** A second packet from a split store that needs sending. */
282 PacketPtr pendingPacket;
283
284 /** Completes a packet and returns whether the access is finished. */
285 inline bool complete() { return --outstanding == 0; }
286 };
287
288 /** Writeback event, specifically for when stores forward data to loads. */
289 class WritebackEvent : public Event {
290 public:
291 /** Constructs a writeback event. */
292 WritebackEvent(DynInstPtr &_inst, PacketPtr pkt, LSQUnit *lsq_ptr);
293
294 /** Processes the writeback event. */
295 void process();
296
297 /** Returns the description of this event. */
298 const char *description() const;
299
300 private:
301 /** Instruction whose results are being written back. */
302 DynInstPtr inst;
303
304 /** The packet that would have been sent to memory. */
305 PacketPtr pkt;
306
307 /** The pointer to the LSQ unit that issued the store. */
308 LSQUnit<Impl> *lsqPtr;
309 };
310
311 public:
312 struct SQEntry {
313 /** Constructs an empty store queue entry. */
314 SQEntry()
315 : inst(NULL), req(NULL), size(0),
316 canWB(0), committed(0), completed(0)
317 {
318 std::memset(data, 0, sizeof(data));
319 }
320
321 /** Constructs a store queue entry for a given instruction. */
322 SQEntry(DynInstPtr &_inst)
323 : inst(_inst), req(NULL), sreqLow(NULL), sreqHigh(NULL), size(0),
324 isSplit(0), canWB(0), committed(0), completed(0)
325 {
326 std::memset(data, 0, sizeof(data));
327 }
328
329 /** The store instruction. */
330 DynInstPtr inst;
331 /** The request for the store. */
332 RequestPtr req;
333 /** The split requests for the store. */
334 RequestPtr sreqLow;
335 RequestPtr sreqHigh;
336 /** The size of the store. */
337 int size;
338 /** The store data. */
339 char data[16];
340 /** Whether or not the store is split into two requests. */
341 bool isSplit;
342 /** Whether or not the store can writeback. */
343 bool canWB;
344 /** Whether or not the store is committed. */
345 bool committed;
346 /** Whether or not the store is completed. */
347 bool completed;
348 };
349
350 private:
351 /** The LSQUnit thread id. */
352 ThreadID lsqID;
353
354 /** The store queue. */
355 std::vector<SQEntry> storeQueue;
356
357 /** The load queue. */
358 std::vector<DynInstPtr> loadQueue;
359
360 /** The number of LQ entries, plus a sentinel entry (circular queue).
361 * @todo: Consider having var that records the true number of LQ entries.
362 */
363 unsigned LQEntries;
364 /** The number of SQ entries, plus a sentinel entry (circular queue).
365 * @todo: Consider having var that records the true number of SQ entries.
366 */
367 unsigned SQEntries;
368
369 /** The number of load instructions in the LQ. */
370 int loads;
371 /** The number of store instructions in the SQ. */
372 int stores;
373 /** The number of store instructions in the SQ waiting to writeback. */
374 int storesToWB;
375
376 /** The index of the head instruction in the LQ. */
377 int loadHead;
378 /** The index of the tail instruction in the LQ. */
379 int loadTail;
380
381 /** The index of the head instruction in the SQ. */
382 int storeHead;
383 /** The index of the first instruction that may be ready to be
384 * written back, and has not yet been written back.
385 */
386 int storeWBIdx;
387 /** The index of the tail instruction in the SQ. */
388 int storeTail;
389
390 /// @todo Consider moving to a more advanced model with write vs read ports
391 /** The number of cache ports available each cycle. */
392 int cachePorts;
393
394 /** The number of used cache ports in this cycle. */
395 int usedPorts;
396
397 /** Is the LSQ switched out. */
398 bool switchedOut;
399
400 //list<InstSeqNum> mshrSeqNums;
401
402 /** Wire to read information from the issue stage time queue. */
403 typename TimeBuffer<IssueStruct>::wire fromIssue;
404
405 /** Whether or not the LSQ is stalled. */
406 bool stalled;
407 /** The store that causes the stall due to partial store to load
408 * forwarding.
409 */
410 InstSeqNum stallingStoreIsn;
411 /** The index of the above store. */
412 int stallingLoadIdx;
413
414 /** The packet that needs to be retried. */
415 PacketPtr retryPkt;
416
417 /** Whehter or not a store is blocked due to the memory system. */
418 bool isStoreBlocked;
419
420 /** Whether or not a load is blocked due to the memory system. */
421 bool isLoadBlocked;
422
423 /** Has the blocked load been handled. */
424 bool loadBlockedHandled;
425
426 /** The sequence number of the blocked load. */
427 InstSeqNum blockedLoadSeqNum;
428
429 /** The oldest load that caused a memory ordering violation. */
430 DynInstPtr memDepViolator;
431
432 /** Whether or not there is a packet that couldn't be sent because of
433 * a lack of cache ports. */
434 bool hasPendingPkt;
435
436 /** The packet that is pending free cache ports. */
437 PacketPtr pendingPkt;
438
439 // Will also need how many read/write ports the Dcache has. Or keep track
440 // of that in stage that is one level up, and only call executeLoad/Store
441 // the appropriate number of times.
442 /** Total number of loads forwaded from LSQ stores. */
443 Stats::Scalar lsqForwLoads;
444
445 /** Total number of loads ignored due to invalid addresses. */
446 Stats::Scalar invAddrLoads;
447
448 /** Total number of squashed loads. */
449 Stats::Scalar lsqSquashedLoads;
450
451 /** Total number of responses from the memory system that are
452 * ignored due to the instruction already being squashed. */
453 Stats::Scalar lsqIgnoredResponses;
454
455 /** Tota number of memory ordering violations. */
456 Stats::Scalar lsqMemOrderViolation;
457
458 /** Total number of squashed stores. */
459 Stats::Scalar lsqSquashedStores;
460
461 /** Total number of software prefetches ignored due to invalid addresses. */
462 Stats::Scalar invAddrSwpfs;
463
464 /** Ready loads blocked due to partial store-forwarding. */
465 Stats::Scalar lsqBlockedLoads;
466
467 /** Number of loads that were rescheduled. */
468 Stats::Scalar lsqRescheduledLoads;
469
470 /** Number of times the LSQ is blocked due to the cache. */
471 Stats::Scalar lsqCacheBlocked;
472
473 public:
474 /** Executes the load at the given index. */
475 Fault read(Request *req, Request *sreqLow, Request *sreqHigh,
476 uint8_t *data, int load_idx);
477
478 /** Executes the store at the given index. */
479 Fault write(Request *req, Request *sreqLow, Request *sreqHigh,
480 uint8_t *data, int store_idx);
481
482 /** Returns the index of the head load instruction. */
483 int getLoadHead() { return loadHead; }
484 /** Returns the sequence number of the head load instruction. */
485 InstSeqNum getLoadHeadSeqNum()
486 {
487 if (loadQueue[loadHead]) {
488 return loadQueue[loadHead]->seqNum;
489 } else {
490 return 0;
491 }
492
493 }
494
495 /** Returns the index of the head store instruction. */
496 int getStoreHead() { return storeHead; }
497 /** Returns the sequence number of the head store instruction. */
498 InstSeqNum getStoreHeadSeqNum()
499 {
500 if (storeQueue[storeHead].inst) {
501 return storeQueue[storeHead].inst->seqNum;
502 } else {
503 return 0;
504 }
505
506 }
507
508 /** Returns whether or not the LSQ unit is stalled. */
509 bool isStalled() { return stalled; }
510};
511
512template <class Impl>
513Fault
514LSQUnit<Impl>::read(Request *req, Request *sreqLow, Request *sreqHigh,
515 uint8_t *data, int load_idx)
516{
517 DynInstPtr load_inst = loadQueue[load_idx];
518
519 assert(load_inst);
520
521 assert(!load_inst->isExecuted());
522
523 // Make sure this isn't an uncacheable access
524 // A bit of a hackish way to get uncached accesses to work only if they're
525 // at the head of the LSQ and are ready to commit (at the head of the ROB
526 // too).
527 if (req->isUncacheable() &&
528 (load_idx != loadHead || !load_inst->isAtCommit())) {
529 iewStage->rescheduleMemInst(load_inst);
530 ++lsqRescheduledLoads;
531 DPRINTF(LSQUnit, "Uncachable load [sn:%lli] PC %s\n",
532 load_inst->seqNum, load_inst->pcState());
533
534 // Must delete request now that it wasn't handed off to
535 // memory. This is quite ugly. @todo: Figure out the proper
536 // place to really handle request deletes.
537 delete req;
538 if (TheISA::HasUnalignedMemAcc && sreqLow) {
539 delete sreqLow;
540 delete sreqHigh;
541 }
542 return TheISA::genMachineCheckFault();
543 }
544
545 // Check the SQ for any previous stores that might lead to forwarding
546 int store_idx = load_inst->sqIdx;
547
548 int store_size = 0;
549
550 DPRINTF(LSQUnit, "Read called, load idx: %i, store idx: %i, "
551 "storeHead: %i addr: %#x%s\n",
552 load_idx, store_idx, storeHead, req->getPaddr(),
553 sreqLow ? " split" : "");
554
555 if (req->isLLSC()) {
556 assert(!sreqLow);
557 // Disable recording the result temporarily. Writing to misc
558 // regs normally updates the result, but this is not the
559 // desired behavior when handling store conditionals.
560 load_inst->recordResult = false;
561 TheISA::handleLockedRead(load_inst.get(), req);
562 load_inst->recordResult = true;
563 }
564
565 while (store_idx != -1) {
566 // End once we've reached the top of the LSQ
567 if (store_idx == storeWBIdx) {
568 break;
569 }
570
571 // Move the index to one younger
572 if (--store_idx < 0)
573 store_idx += SQEntries;
574
575 assert(storeQueue[store_idx].inst);
576
577 store_size = storeQueue[store_idx].size;
578
579 if (store_size == 0)
580 continue;
581 else if (storeQueue[store_idx].inst->uncacheable())
582 continue;
583
584 assert(storeQueue[store_idx].inst->effAddrValid);
585
586 // Check if the store data is within the lower and upper bounds of
587 // addresses that the request needs.
588 bool store_has_lower_limit =
589 req->getVaddr() >= storeQueue[store_idx].inst->effAddr;
590 bool store_has_upper_limit =
591 (req->getVaddr() + req->getSize()) <=
592 (storeQueue[store_idx].inst->effAddr + store_size);
593 bool lower_load_has_store_part =
594 req->getVaddr() < (storeQueue[store_idx].inst->effAddr +
595 store_size);
596 bool upper_load_has_store_part =
597 (req->getVaddr() + req->getSize()) >
598 storeQueue[store_idx].inst->effAddr;
599
600 // If the store's data has all of the data needed, we can forward.
601 if ((store_has_lower_limit && store_has_upper_limit)) {
602 // Get shift amount for offset into the store's data.
603 int shift_amt = req->getVaddr() & (store_size - 1);
604
605 memcpy(data, storeQueue[store_idx].data + shift_amt,
606 req->getSize());
607
608 assert(!load_inst->memData);
609 load_inst->memData = new uint8_t[64];
610
611 memcpy(load_inst->memData,
612 storeQueue[store_idx].data + shift_amt, req->getSize());
613
614 DPRINTF(LSQUnit, "Forwarding from store idx %i to load to "
615 "addr %#x, data %#x\n",
616 store_idx, req->getVaddr(), data);
617
618 PacketPtr data_pkt = new Packet(req, MemCmd::ReadReq,
619 Packet::Broadcast);
620 data_pkt->dataStatic(load_inst->memData);
621
622 WritebackEvent *wb = new WritebackEvent(load_inst, data_pkt, this);
623
624 // We'll say this has a 1 cycle load-store forwarding latency
625 // for now.
626 // @todo: Need to make this a parameter.
| 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_LSQ_UNIT_HH__
33#define __CPU_O3_LSQ_UNIT_HH__
34
35#include <algorithm>
36#include <cstring>
37#include <map>
38#include <queue>
39
40#include "arch/faults.hh"
41#include "arch/locked_mem.hh"
42#include "config/full_system.hh"
43#include "config/the_isa.hh"
44#include "base/fast_alloc.hh"
45#include "base/hashmap.hh"
46#include "cpu/inst_seq.hh"
47#include "mem/packet.hh"
48#include "mem/port.hh"
49
50class DerivO3CPUParams;
51
52/**
53 * Class that implements the actual LQ and SQ for each specific
54 * thread. Both are circular queues; load entries are freed upon
55 * committing, while store entries are freed once they writeback. The
56 * LSQUnit tracks if there are memory ordering violations, and also
57 * detects partial load to store forwarding cases (a store only has
58 * part of a load's data) that requires the load to wait until the
59 * store writes back. In the former case it holds onto the instruction
60 * until the dependence unit looks at it, and in the latter it stalls
61 * the LSQ until the store writes back. At that point the load is
62 * replayed.
63 */
64template <class Impl>
65class LSQUnit {
66 public:
67 typedef typename Impl::O3CPU O3CPU;
68 typedef typename Impl::DynInstPtr DynInstPtr;
69 typedef typename Impl::CPUPol::IEW IEW;
70 typedef typename Impl::CPUPol::LSQ LSQ;
71 typedef typename Impl::CPUPol::IssueStruct IssueStruct;
72
73 public:
74 /** Constructs an LSQ unit. init() must be called prior to use. */
75 LSQUnit();
76
77 /** Initializes the LSQ unit with the specified number of entries. */
78 void init(O3CPU *cpu_ptr, IEW *iew_ptr, DerivO3CPUParams *params,
79 LSQ *lsq_ptr, unsigned maxLQEntries, unsigned maxSQEntries,
80 unsigned id);
81
82 /** Returns the name of the LSQ unit. */
83 std::string name() const;
84
85 /** Registers statistics. */
86 void regStats();
87
88 /** Sets the pointer to the dcache port. */
89 void setDcachePort(Port *dcache_port);
90
91 /** Switches out LSQ unit. */
92 void switchOut();
93
94 /** Takes over from another CPU's thread. */
95 void takeOverFrom();
96
97 /** Returns if the LSQ is switched out. */
98 bool isSwitchedOut() { return switchedOut; }
99
100 /** Ticks the LSQ unit, which in this case only resets the number of
101 * used cache ports.
102 * @todo: Move the number of used ports up to the LSQ level so it can
103 * be shared by all LSQ units.
104 */
105 void tick() { usedPorts = 0; }
106
107 /** Inserts an instruction. */
108 void insert(DynInstPtr &inst);
109 /** Inserts a load instruction. */
110 void insertLoad(DynInstPtr &load_inst);
111 /** Inserts a store instruction. */
112 void insertStore(DynInstPtr &store_inst);
113
114 /** Executes a load instruction. */
115 Fault executeLoad(DynInstPtr &inst);
116
117 Fault executeLoad(int lq_idx) { panic("Not implemented"); return NoFault; }
118 /** Executes a store instruction. */
119 Fault executeStore(DynInstPtr &inst);
120
121 /** Commits the head load. */
122 void commitLoad();
123 /** Commits loads older than a specific sequence number. */
124 void commitLoads(InstSeqNum &youngest_inst);
125
126 /** Commits stores older than a specific sequence number. */
127 void commitStores(InstSeqNum &youngest_inst);
128
129 /** Writes back stores. */
130 void writebackStores();
131
132 /** Completes the data access that has been returned from the
133 * memory system. */
134 void completeDataAccess(PacketPtr pkt);
135
136 /** Clears all the entries in the LQ. */
137 void clearLQ();
138
139 /** Clears all the entries in the SQ. */
140 void clearSQ();
141
142 /** Resizes the LQ to a given size. */
143 void resizeLQ(unsigned size);
144
145 /** Resizes the SQ to a given size. */
146 void resizeSQ(unsigned size);
147
148 /** Squashes all instructions younger than a specific sequence number. */
149 void squash(const InstSeqNum &squashed_num);
150
151 /** Returns if there is a memory ordering violation. Value is reset upon
152 * call to getMemDepViolator().
153 */
154 bool violation() { return memDepViolator; }
155
156 /** Returns the memory ordering violator. */
157 DynInstPtr getMemDepViolator();
158
159 /** Returns if a load became blocked due to the memory system. */
160 bool loadBlocked()
161 { return isLoadBlocked; }
162
163 /** Clears the signal that a load became blocked. */
164 void clearLoadBlocked()
165 { isLoadBlocked = false; }
166
167 /** Returns if the blocked load was handled. */
168 bool isLoadBlockedHandled()
169 { return loadBlockedHandled; }
170
171 /** Records the blocked load as being handled. */
172 void setLoadBlockedHandled()
173 { loadBlockedHandled = true; }
174
175 /** Returns the number of free entries (min of free LQ and SQ entries). */
176 unsigned numFreeEntries();
177
178 /** Returns the number of loads ready to execute. */
179 int numLoadsReady();
180
181 /** Returns the number of loads in the LQ. */
182 int numLoads() { return loads; }
183
184 /** Returns the number of stores in the SQ. */
185 int numStores() { return stores; }
186
187 /** Returns if either the LQ or SQ is full. */
188 bool isFull() { return lqFull() || sqFull(); }
189
190 /** Returns if the LQ is full. */
191 bool lqFull() { return loads >= (LQEntries - 1); }
192
193 /** Returns if the SQ is full. */
194 bool sqFull() { return stores >= (SQEntries - 1); }
195
196 /** Returns the number of instructions in the LSQ. */
197 unsigned getCount() { return loads + stores; }
198
199 /** Returns if there are any stores to writeback. */
200 bool hasStoresToWB() { return storesToWB; }
201
202 /** Returns the number of stores to writeback. */
203 int numStoresToWB() { return storesToWB; }
204
205 /** Returns if the LSQ unit will writeback on this cycle. */
206 bool willWB() { return storeQueue[storeWBIdx].canWB &&
207 !storeQueue[storeWBIdx].completed &&
208 !isStoreBlocked; }
209
210 /** Handles doing the retry. */
211 void recvRetry();
212
213 private:
214 /** Writes back the instruction, sending it to IEW. */
215 void writeback(DynInstPtr &inst, PacketPtr pkt);
216
217 /** Writes back a store that couldn't be completed the previous cycle. */
218 void writebackPendingStore();
219
220 /** Handles completing the send of a store to memory. */
221 void storePostSend(PacketPtr pkt);
222
223 /** Completes the store at the specified index. */
224 void completeStore(int store_idx);
225
226 /** Attempts to send a store to the cache. */
227 bool sendStore(PacketPtr data_pkt);
228
229 /** Increments the given store index (circular queue). */
230 inline void incrStIdx(int &store_idx);
231 /** Decrements the given store index (circular queue). */
232 inline void decrStIdx(int &store_idx);
233 /** Increments the given load index (circular queue). */
234 inline void incrLdIdx(int &load_idx);
235 /** Decrements the given load index (circular queue). */
236 inline void decrLdIdx(int &load_idx);
237
238 public:
239 /** Debugging function to dump instructions in the LSQ. */
240 void dumpInsts();
241
242 private:
243 /** Pointer to the CPU. */
244 O3CPU *cpu;
245
246 /** Pointer to the IEW stage. */
247 IEW *iewStage;
248
249 /** Pointer to the LSQ. */
250 LSQ *lsq;
251
252 /** Pointer to the dcache port. Used only for sending. */
253 Port *dcachePort;
254
255 /** Derived class to hold any sender state the LSQ needs. */
256 class LSQSenderState : public Packet::SenderState, public FastAlloc
257 {
258 public:
259 /** Default constructor. */
260 LSQSenderState()
261 : noWB(false), isSplit(false), pktToSend(false), outstanding(1),
262 mainPkt(NULL), pendingPacket(NULL)
263 { }
264
265 /** Instruction who initiated the access to memory. */
266 DynInstPtr inst;
267 /** Whether or not it is a load. */
268 bool isLoad;
269 /** The LQ/SQ index of the instruction. */
270 int idx;
271 /** Whether or not the instruction will need to writeback. */
272 bool noWB;
273 /** Whether or not this access is split in two. */
274 bool isSplit;
275 /** Whether or not there is a packet that needs sending. */
276 bool pktToSend;
277 /** Number of outstanding packets to complete. */
278 int outstanding;
279 /** The main packet from a split load, used during writeback. */
280 PacketPtr mainPkt;
281 /** A second packet from a split store that needs sending. */
282 PacketPtr pendingPacket;
283
284 /** Completes a packet and returns whether the access is finished. */
285 inline bool complete() { return --outstanding == 0; }
286 };
287
288 /** Writeback event, specifically for when stores forward data to loads. */
289 class WritebackEvent : public Event {
290 public:
291 /** Constructs a writeback event. */
292 WritebackEvent(DynInstPtr &_inst, PacketPtr pkt, LSQUnit *lsq_ptr);
293
294 /** Processes the writeback event. */
295 void process();
296
297 /** Returns the description of this event. */
298 const char *description() const;
299
300 private:
301 /** Instruction whose results are being written back. */
302 DynInstPtr inst;
303
304 /** The packet that would have been sent to memory. */
305 PacketPtr pkt;
306
307 /** The pointer to the LSQ unit that issued the store. */
308 LSQUnit<Impl> *lsqPtr;
309 };
310
311 public:
312 struct SQEntry {
313 /** Constructs an empty store queue entry. */
314 SQEntry()
315 : inst(NULL), req(NULL), size(0),
316 canWB(0), committed(0), completed(0)
317 {
318 std::memset(data, 0, sizeof(data));
319 }
320
321 /** Constructs a store queue entry for a given instruction. */
322 SQEntry(DynInstPtr &_inst)
323 : inst(_inst), req(NULL), sreqLow(NULL), sreqHigh(NULL), size(0),
324 isSplit(0), canWB(0), committed(0), completed(0)
325 {
326 std::memset(data, 0, sizeof(data));
327 }
328
329 /** The store instruction. */
330 DynInstPtr inst;
331 /** The request for the store. */
332 RequestPtr req;
333 /** The split requests for the store. */
334 RequestPtr sreqLow;
335 RequestPtr sreqHigh;
336 /** The size of the store. */
337 int size;
338 /** The store data. */
339 char data[16];
340 /** Whether or not the store is split into two requests. */
341 bool isSplit;
342 /** Whether or not the store can writeback. */
343 bool canWB;
344 /** Whether or not the store is committed. */
345 bool committed;
346 /** Whether or not the store is completed. */
347 bool completed;
348 };
349
350 private:
351 /** The LSQUnit thread id. */
352 ThreadID lsqID;
353
354 /** The store queue. */
355 std::vector<SQEntry> storeQueue;
356
357 /** The load queue. */
358 std::vector<DynInstPtr> loadQueue;
359
360 /** The number of LQ entries, plus a sentinel entry (circular queue).
361 * @todo: Consider having var that records the true number of LQ entries.
362 */
363 unsigned LQEntries;
364 /** The number of SQ entries, plus a sentinel entry (circular queue).
365 * @todo: Consider having var that records the true number of SQ entries.
366 */
367 unsigned SQEntries;
368
369 /** The number of load instructions in the LQ. */
370 int loads;
371 /** The number of store instructions in the SQ. */
372 int stores;
373 /** The number of store instructions in the SQ waiting to writeback. */
374 int storesToWB;
375
376 /** The index of the head instruction in the LQ. */
377 int loadHead;
378 /** The index of the tail instruction in the LQ. */
379 int loadTail;
380
381 /** The index of the head instruction in the SQ. */
382 int storeHead;
383 /** The index of the first instruction that may be ready to be
384 * written back, and has not yet been written back.
385 */
386 int storeWBIdx;
387 /** The index of the tail instruction in the SQ. */
388 int storeTail;
389
390 /// @todo Consider moving to a more advanced model with write vs read ports
391 /** The number of cache ports available each cycle. */
392 int cachePorts;
393
394 /** The number of used cache ports in this cycle. */
395 int usedPorts;
396
397 /** Is the LSQ switched out. */
398 bool switchedOut;
399
400 //list<InstSeqNum> mshrSeqNums;
401
402 /** Wire to read information from the issue stage time queue. */
403 typename TimeBuffer<IssueStruct>::wire fromIssue;
404
405 /** Whether or not the LSQ is stalled. */
406 bool stalled;
407 /** The store that causes the stall due to partial store to load
408 * forwarding.
409 */
410 InstSeqNum stallingStoreIsn;
411 /** The index of the above store. */
412 int stallingLoadIdx;
413
414 /** The packet that needs to be retried. */
415 PacketPtr retryPkt;
416
417 /** Whehter or not a store is blocked due to the memory system. */
418 bool isStoreBlocked;
419
420 /** Whether or not a load is blocked due to the memory system. */
421 bool isLoadBlocked;
422
423 /** Has the blocked load been handled. */
424 bool loadBlockedHandled;
425
426 /** The sequence number of the blocked load. */
427 InstSeqNum blockedLoadSeqNum;
428
429 /** The oldest load that caused a memory ordering violation. */
430 DynInstPtr memDepViolator;
431
432 /** Whether or not there is a packet that couldn't be sent because of
433 * a lack of cache ports. */
434 bool hasPendingPkt;
435
436 /** The packet that is pending free cache ports. */
437 PacketPtr pendingPkt;
438
439 // Will also need how many read/write ports the Dcache has. Or keep track
440 // of that in stage that is one level up, and only call executeLoad/Store
441 // the appropriate number of times.
442 /** Total number of loads forwaded from LSQ stores. */
443 Stats::Scalar lsqForwLoads;
444
445 /** Total number of loads ignored due to invalid addresses. */
446 Stats::Scalar invAddrLoads;
447
448 /** Total number of squashed loads. */
449 Stats::Scalar lsqSquashedLoads;
450
451 /** Total number of responses from the memory system that are
452 * ignored due to the instruction already being squashed. */
453 Stats::Scalar lsqIgnoredResponses;
454
455 /** Tota number of memory ordering violations. */
456 Stats::Scalar lsqMemOrderViolation;
457
458 /** Total number of squashed stores. */
459 Stats::Scalar lsqSquashedStores;
460
461 /** Total number of software prefetches ignored due to invalid addresses. */
462 Stats::Scalar invAddrSwpfs;
463
464 /** Ready loads blocked due to partial store-forwarding. */
465 Stats::Scalar lsqBlockedLoads;
466
467 /** Number of loads that were rescheduled. */
468 Stats::Scalar lsqRescheduledLoads;
469
470 /** Number of times the LSQ is blocked due to the cache. */
471 Stats::Scalar lsqCacheBlocked;
472
473 public:
474 /** Executes the load at the given index. */
475 Fault read(Request *req, Request *sreqLow, Request *sreqHigh,
476 uint8_t *data, int load_idx);
477
478 /** Executes the store at the given index. */
479 Fault write(Request *req, Request *sreqLow, Request *sreqHigh,
480 uint8_t *data, int store_idx);
481
482 /** Returns the index of the head load instruction. */
483 int getLoadHead() { return loadHead; }
484 /** Returns the sequence number of the head load instruction. */
485 InstSeqNum getLoadHeadSeqNum()
486 {
487 if (loadQueue[loadHead]) {
488 return loadQueue[loadHead]->seqNum;
489 } else {
490 return 0;
491 }
492
493 }
494
495 /** Returns the index of the head store instruction. */
496 int getStoreHead() { return storeHead; }
497 /** Returns the sequence number of the head store instruction. */
498 InstSeqNum getStoreHeadSeqNum()
499 {
500 if (storeQueue[storeHead].inst) {
501 return storeQueue[storeHead].inst->seqNum;
502 } else {
503 return 0;
504 }
505
506 }
507
508 /** Returns whether or not the LSQ unit is stalled. */
509 bool isStalled() { return stalled; }
510};
511
512template <class Impl>
513Fault
514LSQUnit<Impl>::read(Request *req, Request *sreqLow, Request *sreqHigh,
515 uint8_t *data, int load_idx)
516{
517 DynInstPtr load_inst = loadQueue[load_idx];
518
519 assert(load_inst);
520
521 assert(!load_inst->isExecuted());
522
523 // Make sure this isn't an uncacheable access
524 // A bit of a hackish way to get uncached accesses to work only if they're
525 // at the head of the LSQ and are ready to commit (at the head of the ROB
526 // too).
527 if (req->isUncacheable() &&
528 (load_idx != loadHead || !load_inst->isAtCommit())) {
529 iewStage->rescheduleMemInst(load_inst);
530 ++lsqRescheduledLoads;
531 DPRINTF(LSQUnit, "Uncachable load [sn:%lli] PC %s\n",
532 load_inst->seqNum, load_inst->pcState());
533
534 // Must delete request now that it wasn't handed off to
535 // memory. This is quite ugly. @todo: Figure out the proper
536 // place to really handle request deletes.
537 delete req;
538 if (TheISA::HasUnalignedMemAcc && sreqLow) {
539 delete sreqLow;
540 delete sreqHigh;
541 }
542 return TheISA::genMachineCheckFault();
543 }
544
545 // Check the SQ for any previous stores that might lead to forwarding
546 int store_idx = load_inst->sqIdx;
547
548 int store_size = 0;
549
550 DPRINTF(LSQUnit, "Read called, load idx: %i, store idx: %i, "
551 "storeHead: %i addr: %#x%s\n",
552 load_idx, store_idx, storeHead, req->getPaddr(),
553 sreqLow ? " split" : "");
554
555 if (req->isLLSC()) {
556 assert(!sreqLow);
557 // Disable recording the result temporarily. Writing to misc
558 // regs normally updates the result, but this is not the
559 // desired behavior when handling store conditionals.
560 load_inst->recordResult = false;
561 TheISA::handleLockedRead(load_inst.get(), req);
562 load_inst->recordResult = true;
563 }
564
565 while (store_idx != -1) {
566 // End once we've reached the top of the LSQ
567 if (store_idx == storeWBIdx) {
568 break;
569 }
570
571 // Move the index to one younger
572 if (--store_idx < 0)
573 store_idx += SQEntries;
574
575 assert(storeQueue[store_idx].inst);
576
577 store_size = storeQueue[store_idx].size;
578
579 if (store_size == 0)
580 continue;
581 else if (storeQueue[store_idx].inst->uncacheable())
582 continue;
583
584 assert(storeQueue[store_idx].inst->effAddrValid);
585
586 // Check if the store data is within the lower and upper bounds of
587 // addresses that the request needs.
588 bool store_has_lower_limit =
589 req->getVaddr() >= storeQueue[store_idx].inst->effAddr;
590 bool store_has_upper_limit =
591 (req->getVaddr() + req->getSize()) <=
592 (storeQueue[store_idx].inst->effAddr + store_size);
593 bool lower_load_has_store_part =
594 req->getVaddr() < (storeQueue[store_idx].inst->effAddr +
595 store_size);
596 bool upper_load_has_store_part =
597 (req->getVaddr() + req->getSize()) >
598 storeQueue[store_idx].inst->effAddr;
599
600 // If the store's data has all of the data needed, we can forward.
601 if ((store_has_lower_limit && store_has_upper_limit)) {
602 // Get shift amount for offset into the store's data.
603 int shift_amt = req->getVaddr() & (store_size - 1);
604
605 memcpy(data, storeQueue[store_idx].data + shift_amt,
606 req->getSize());
607
608 assert(!load_inst->memData);
609 load_inst->memData = new uint8_t[64];
610
611 memcpy(load_inst->memData,
612 storeQueue[store_idx].data + shift_amt, req->getSize());
613
614 DPRINTF(LSQUnit, "Forwarding from store idx %i to load to "
615 "addr %#x, data %#x\n",
616 store_idx, req->getVaddr(), data);
617
618 PacketPtr data_pkt = new Packet(req, MemCmd::ReadReq,
619 Packet::Broadcast);
620 data_pkt->dataStatic(load_inst->memData);
621
622 WritebackEvent *wb = new WritebackEvent(load_inst, data_pkt, this);
623
624 // We'll say this has a 1 cycle load-store forwarding latency
625 // for now.
626 // @todo: Need to make this a parameter.
|
628
629 // Don't need to do anything special for split loads.
630 if (TheISA::HasUnalignedMemAcc && sreqLow) {
631 delete sreqLow;
632 delete sreqHigh;
633 }
634
635 ++lsqForwLoads;
636 return NoFault;
637 } else if ((store_has_lower_limit && lower_load_has_store_part) ||
638 (store_has_upper_limit && upper_load_has_store_part) ||
639 (lower_load_has_store_part && upper_load_has_store_part)) {
640 // This is the partial store-load forwarding case where a store
641 // has only part of the load's data.
642
643 // If it's already been written back, then don't worry about
644 // stalling on it.
645 if (storeQueue[store_idx].completed) {
646 panic("Should not check one of these");
647 continue;
648 }
649
650 // Must stall load and force it to retry, so long as it's the oldest
651 // load that needs to do so.
652 if (!stalled ||
653 (stalled &&
654 load_inst->seqNum <
655 loadQueue[stallingLoadIdx]->seqNum)) {
656 stalled = true;
657 stallingStoreIsn = storeQueue[store_idx].inst->seqNum;
658 stallingLoadIdx = load_idx;
659 }
660
661 // Tell IQ/mem dep unit that this instruction will need to be
662 // rescheduled eventually
663 iewStage->rescheduleMemInst(load_inst);
664 iewStage->decrWb(load_inst->seqNum);
665 load_inst->clearIssued();
666 ++lsqRescheduledLoads;
667
668 // Do not generate a writeback event as this instruction is not
669 // complete.
670 DPRINTF(LSQUnit, "Load-store forwarding mis-match. "
671 "Store idx %i to load addr %#x\n",
672 store_idx, req->getVaddr());
673
674 // Must delete request now that it wasn't handed off to
675 // memory. This is quite ugly. @todo: Figure out the
676 // proper place to really handle request deletes.
677 delete req;
678 if (TheISA::HasUnalignedMemAcc && sreqLow) {
679 delete sreqLow;
680 delete sreqHigh;
681 }
682
683 return NoFault;
684 }
685 }
686
687 // If there's no forwarding case, then go access memory
688 DPRINTF(LSQUnit, "Doing memory access for inst [sn:%lli] PC %s\n",
689 load_inst->seqNum, load_inst->pcState());
690
691 assert(!load_inst->memData);
692 load_inst->memData = new uint8_t[64];
693
694 ++usedPorts;
695
696 // if we the cache is not blocked, do cache access
697 bool completedFirst = false;
698 if (!lsq->cacheBlocked()) {
699 MemCmd command =
700 req->isLLSC() ? MemCmd::LoadLockedReq : MemCmd::ReadReq;
701 PacketPtr data_pkt = new Packet(req, command, Packet::Broadcast);
702 PacketPtr fst_data_pkt = NULL;
703 PacketPtr snd_data_pkt = NULL;
704
705 data_pkt->dataStatic(load_inst->memData);
706
707 LSQSenderState *state = new LSQSenderState;
708 state->isLoad = true;
709 state->idx = load_idx;
710 state->inst = load_inst;
711 data_pkt->senderState = state;
712
713 if (!TheISA::HasUnalignedMemAcc || !sreqLow) {
714
715 // Point the first packet at the main data packet.
716 fst_data_pkt = data_pkt;
717 } else {
718
719 // Create the split packets.
720 fst_data_pkt = new Packet(sreqLow, command, Packet::Broadcast);
721 snd_data_pkt = new Packet(sreqHigh, command, Packet::Broadcast);
722
723 fst_data_pkt->dataStatic(load_inst->memData);
724 snd_data_pkt->dataStatic(load_inst->memData + sreqLow->getSize());
725
726 fst_data_pkt->senderState = state;
727 snd_data_pkt->senderState = state;
728
729 state->isSplit = true;
730 state->outstanding = 2;
731 state->mainPkt = data_pkt;
732 }
733
734 if (!dcachePort->sendTiming(fst_data_pkt)) {
735 // Delete state and data packet because a load retry
736 // initiates a pipeline restart; it does not retry.
737 delete state;
738 delete data_pkt->req;
739 delete data_pkt;
740 if (TheISA::HasUnalignedMemAcc && sreqLow) {
741 delete fst_data_pkt->req;
742 delete fst_data_pkt;
743 delete snd_data_pkt->req;
744 delete snd_data_pkt;
745 sreqLow = NULL;
746 sreqHigh = NULL;
747 }
748
749 req = NULL;
750
751 // If the access didn't succeed, tell the LSQ by setting
752 // the retry thread id.
753 lsq->setRetryTid(lsqID);
754 } else if (TheISA::HasUnalignedMemAcc && sreqLow) {
755 completedFirst = true;
756
757 // The first packet was sent without problems, so send this one
758 // too. If there is a problem with this packet then the whole
759 // load will be squashed, so indicate this to the state object.
760 // The first packet will return in completeDataAccess and be
761 // handled there.
762 ++usedPorts;
763 if (!dcachePort->sendTiming(snd_data_pkt)) {
764
765 // The main packet will be deleted in completeDataAccess.
766 delete snd_data_pkt->req;
767 delete snd_data_pkt;
768
769 state->complete();
770
771 req = NULL;
772 sreqHigh = NULL;
773
774 lsq->setRetryTid(lsqID);
775 }
776 }
777 }
778
779 // If the cache was blocked, or has become blocked due to the access,
780 // handle it.
781 if (lsq->cacheBlocked()) {
782 if (req)
783 delete req;
784 if (TheISA::HasUnalignedMemAcc && sreqLow && !completedFirst) {
785 delete sreqLow;
786 delete sreqHigh;
787 }
788
789 ++lsqCacheBlocked;
790
791 iewStage->decrWb(load_inst->seqNum);
792 // There's an older load that's already going to squash.
793 if (isLoadBlocked && blockedLoadSeqNum < load_inst->seqNum)
794 return NoFault;
795
796 // Record that the load was blocked due to memory. This
797 // load will squash all instructions after it, be
798 // refetched, and re-executed.
799 isLoadBlocked = true;
800 loadBlockedHandled = false;
801 blockedLoadSeqNum = load_inst->seqNum;
802 // No fault occurred, even though the interface is blocked.
803 return NoFault;
804 }
805
806 return NoFault;
807}
808
809template <class Impl>
810Fault
811LSQUnit<Impl>::write(Request *req, Request *sreqLow, Request *sreqHigh,
812 uint8_t *data, int store_idx)
813{
814 assert(storeQueue[store_idx].inst);
815
816 DPRINTF(LSQUnit, "Doing write to store idx %i, addr %#x data %#x"
817 " | storeHead:%i [sn:%i]\n",
818 store_idx, req->getPaddr(), data, storeHead,
819 storeQueue[store_idx].inst->seqNum);
820
821 storeQueue[store_idx].req = req;
822 storeQueue[store_idx].sreqLow = sreqLow;
823 storeQueue[store_idx].sreqHigh = sreqHigh;
824 unsigned size = req->getSize();
825 storeQueue[store_idx].size = size;
826 assert(size <= sizeof(storeQueue[store_idx].data));
827
828 // Split stores can only occur in ISAs with unaligned memory accesses. If
829 // a store request has been split, sreqLow and sreqHigh will be non-null.
830 if (TheISA::HasUnalignedMemAcc && sreqLow) {
831 storeQueue[store_idx].isSplit = true;
832 }
833
834 memcpy(storeQueue[store_idx].data, data, size);
835
836 // This function only writes the data to the store queue, so no fault
837 // can happen here.
838 return NoFault;
839}
840
841#endif // __CPU_O3_LSQ_UNIT_HH__
| 628
629 // Don't need to do anything special for split loads.
630 if (TheISA::HasUnalignedMemAcc && sreqLow) {
631 delete sreqLow;
632 delete sreqHigh;
633 }
634
635 ++lsqForwLoads;
636 return NoFault;
637 } else if ((store_has_lower_limit && lower_load_has_store_part) ||
638 (store_has_upper_limit && upper_load_has_store_part) ||
639 (lower_load_has_store_part && upper_load_has_store_part)) {
640 // This is the partial store-load forwarding case where a store
641 // has only part of the load's data.
642
643 // If it's already been written back, then don't worry about
644 // stalling on it.
645 if (storeQueue[store_idx].completed) {
646 panic("Should not check one of these");
647 continue;
648 }
649
650 // Must stall load and force it to retry, so long as it's the oldest
651 // load that needs to do so.
652 if (!stalled ||
653 (stalled &&
654 load_inst->seqNum <
655 loadQueue[stallingLoadIdx]->seqNum)) {
656 stalled = true;
657 stallingStoreIsn = storeQueue[store_idx].inst->seqNum;
658 stallingLoadIdx = load_idx;
659 }
660
661 // Tell IQ/mem dep unit that this instruction will need to be
662 // rescheduled eventually
663 iewStage->rescheduleMemInst(load_inst);
664 iewStage->decrWb(load_inst->seqNum);
665 load_inst->clearIssued();
666 ++lsqRescheduledLoads;
667
668 // Do not generate a writeback event as this instruction is not
669 // complete.
670 DPRINTF(LSQUnit, "Load-store forwarding mis-match. "
671 "Store idx %i to load addr %#x\n",
672 store_idx, req->getVaddr());
673
674 // Must delete request now that it wasn't handed off to
675 // memory. This is quite ugly. @todo: Figure out the
676 // proper place to really handle request deletes.
677 delete req;
678 if (TheISA::HasUnalignedMemAcc && sreqLow) {
679 delete sreqLow;
680 delete sreqHigh;
681 }
682
683 return NoFault;
684 }
685 }
686
687 // If there's no forwarding case, then go access memory
688 DPRINTF(LSQUnit, "Doing memory access for inst [sn:%lli] PC %s\n",
689 load_inst->seqNum, load_inst->pcState());
690
691 assert(!load_inst->memData);
692 load_inst->memData = new uint8_t[64];
693
694 ++usedPorts;
695
696 // if we the cache is not blocked, do cache access
697 bool completedFirst = false;
698 if (!lsq->cacheBlocked()) {
699 MemCmd command =
700 req->isLLSC() ? MemCmd::LoadLockedReq : MemCmd::ReadReq;
701 PacketPtr data_pkt = new Packet(req, command, Packet::Broadcast);
702 PacketPtr fst_data_pkt = NULL;
703 PacketPtr snd_data_pkt = NULL;
704
705 data_pkt->dataStatic(load_inst->memData);
706
707 LSQSenderState *state = new LSQSenderState;
708 state->isLoad = true;
709 state->idx = load_idx;
710 state->inst = load_inst;
711 data_pkt->senderState = state;
712
713 if (!TheISA::HasUnalignedMemAcc || !sreqLow) {
714
715 // Point the first packet at the main data packet.
716 fst_data_pkt = data_pkt;
717 } else {
718
719 // Create the split packets.
720 fst_data_pkt = new Packet(sreqLow, command, Packet::Broadcast);
721 snd_data_pkt = new Packet(sreqHigh, command, Packet::Broadcast);
722
723 fst_data_pkt->dataStatic(load_inst->memData);
724 snd_data_pkt->dataStatic(load_inst->memData + sreqLow->getSize());
725
726 fst_data_pkt->senderState = state;
727 snd_data_pkt->senderState = state;
728
729 state->isSplit = true;
730 state->outstanding = 2;
731 state->mainPkt = data_pkt;
732 }
733
734 if (!dcachePort->sendTiming(fst_data_pkt)) {
735 // Delete state and data packet because a load retry
736 // initiates a pipeline restart; it does not retry.
737 delete state;
738 delete data_pkt->req;
739 delete data_pkt;
740 if (TheISA::HasUnalignedMemAcc && sreqLow) {
741 delete fst_data_pkt->req;
742 delete fst_data_pkt;
743 delete snd_data_pkt->req;
744 delete snd_data_pkt;
745 sreqLow = NULL;
746 sreqHigh = NULL;
747 }
748
749 req = NULL;
750
751 // If the access didn't succeed, tell the LSQ by setting
752 // the retry thread id.
753 lsq->setRetryTid(lsqID);
754 } else if (TheISA::HasUnalignedMemAcc && sreqLow) {
755 completedFirst = true;
756
757 // The first packet was sent without problems, so send this one
758 // too. If there is a problem with this packet then the whole
759 // load will be squashed, so indicate this to the state object.
760 // The first packet will return in completeDataAccess and be
761 // handled there.
762 ++usedPorts;
763 if (!dcachePort->sendTiming(snd_data_pkt)) {
764
765 // The main packet will be deleted in completeDataAccess.
766 delete snd_data_pkt->req;
767 delete snd_data_pkt;
768
769 state->complete();
770
771 req = NULL;
772 sreqHigh = NULL;
773
774 lsq->setRetryTid(lsqID);
775 }
776 }
777 }
778
779 // If the cache was blocked, or has become blocked due to the access,
780 // handle it.
781 if (lsq->cacheBlocked()) {
782 if (req)
783 delete req;
784 if (TheISA::HasUnalignedMemAcc && sreqLow && !completedFirst) {
785 delete sreqLow;
786 delete sreqHigh;
787 }
788
789 ++lsqCacheBlocked;
790
791 iewStage->decrWb(load_inst->seqNum);
792 // There's an older load that's already going to squash.
793 if (isLoadBlocked && blockedLoadSeqNum < load_inst->seqNum)
794 return NoFault;
795
796 // Record that the load was blocked due to memory. This
797 // load will squash all instructions after it, be
798 // refetched, and re-executed.
799 isLoadBlocked = true;
800 loadBlockedHandled = false;
801 blockedLoadSeqNum = load_inst->seqNum;
802 // No fault occurred, even though the interface is blocked.
803 return NoFault;
804 }
805
806 return NoFault;
807}
808
809template <class Impl>
810Fault
811LSQUnit<Impl>::write(Request *req, Request *sreqLow, Request *sreqHigh,
812 uint8_t *data, int store_idx)
813{
814 assert(storeQueue[store_idx].inst);
815
816 DPRINTF(LSQUnit, "Doing write to store idx %i, addr %#x data %#x"
817 " | storeHead:%i [sn:%i]\n",
818 store_idx, req->getPaddr(), data, storeHead,
819 storeQueue[store_idx].inst->seqNum);
820
821 storeQueue[store_idx].req = req;
822 storeQueue[store_idx].sreqLow = sreqLow;
823 storeQueue[store_idx].sreqHigh = sreqHigh;
824 unsigned size = req->getSize();
825 storeQueue[store_idx].size = size;
826 assert(size <= sizeof(storeQueue[store_idx].data));
827
828 // Split stores can only occur in ISAs with unaligned memory accesses. If
829 // a store request has been split, sreqLow and sreqHigh will be non-null.
830 if (TheISA::HasUnalignedMemAcc && sreqLow) {
831 storeQueue[store_idx].isSplit = true;
832 }
833
834 memcpy(storeQueue[store_idx].data, data, size);
835
836 // This function only writes the data to the store queue, so no fault
837 // can happen here.
838 return NoFault;
839}
840
841#endif // __CPU_O3_LSQ_UNIT_HH__
|