inst_queue_impl.hh revision 10017
111308Santhony.gutierrez@amd.com/* 211308Santhony.gutierrez@amd.com * Copyright (c) 2011-2012 ARM Limited 311308Santhony.gutierrez@amd.com * Copyright (c) 2013 Advanced Micro Devices, Inc. 411308Santhony.gutierrez@amd.com * All rights reserved. 511308Santhony.gutierrez@amd.com * 611308Santhony.gutierrez@amd.com * The license below extends only to copyright in the software and shall 711308Santhony.gutierrez@amd.com * not be construed as granting a license to any other intellectual 811308Santhony.gutierrez@amd.com * property including but not limited to intellectual property relating 911308Santhony.gutierrez@amd.com * to a hardware implementation of the functionality of the software 1011308Santhony.gutierrez@amd.com * licensed hereunder. You may use the software subject to the license 1111308Santhony.gutierrez@amd.com * terms below provided that you ensure that this notice is replicated 1211308Santhony.gutierrez@amd.com * unmodified and in its entirety in all distributions of the software, 1311308Santhony.gutierrez@amd.com * modified or unmodified, in source code or in binary form. 1411308Santhony.gutierrez@amd.com * 1511308Santhony.gutierrez@amd.com * Copyright (c) 2004-2006 The Regents of The University of Michigan 1611308Santhony.gutierrez@amd.com * All rights reserved. 1711308Santhony.gutierrez@amd.com * 1811308Santhony.gutierrez@amd.com * Redistribution and use in source and binary forms, with or without 1911308Santhony.gutierrez@amd.com * modification, are permitted provided that the following conditions are 2011308Santhony.gutierrez@amd.com * met: redistributions of source code must retain the above copyright 2111308Santhony.gutierrez@amd.com * notice, this list of conditions and the following disclaimer; 2211308Santhony.gutierrez@amd.com * redistributions in binary form must reproduce the above copyright 2311308Santhony.gutierrez@amd.com * notice, this list of conditions and the following disclaimer in the 2411308Santhony.gutierrez@amd.com * documentation and/or other materials provided with the distribution; 2511308Santhony.gutierrez@amd.com * neither the name of the copyright holders nor the names of its 2611308Santhony.gutierrez@amd.com * contributors may be used to endorse or promote products derived from 2711308Santhony.gutierrez@amd.com * this software without specific prior written permission. 2811308Santhony.gutierrez@amd.com * 2911308Santhony.gutierrez@amd.com * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 3011308Santhony.gutierrez@amd.com * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 3111308Santhony.gutierrez@amd.com * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 3211308Santhony.gutierrez@amd.com * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 3311308Santhony.gutierrez@amd.com * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 3411308Santhony.gutierrez@amd.com * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 3511308Santhony.gutierrez@amd.com * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 3611308Santhony.gutierrez@amd.com * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 3711308Santhony.gutierrez@amd.com * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 3811308Santhony.gutierrez@amd.com * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 3911308Santhony.gutierrez@amd.com * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 4011308Santhony.gutierrez@amd.com * 4111308Santhony.gutierrez@amd.com * Authors: Kevin Lim 4211308Santhony.gutierrez@amd.com * Korey Sewell 4311308Santhony.gutierrez@amd.com */ 4411308Santhony.gutierrez@amd.com 4511308Santhony.gutierrez@amd.com#ifndef __CPU_O3_INST_QUEUE_IMPL_HH__ 4611308Santhony.gutierrez@amd.com#define __CPU_O3_INST_QUEUE_IMPL_HH__ 4711308Santhony.gutierrez@amd.com 4811308Santhony.gutierrez@amd.com#include <limits> 4911308Santhony.gutierrez@amd.com#include <vector> 5011308Santhony.gutierrez@amd.com 5111308Santhony.gutierrez@amd.com#include "cpu/o3/fu_pool.hh" 5211308Santhony.gutierrez@amd.com#include "cpu/o3/inst_queue.hh" 5311308Santhony.gutierrez@amd.com#include "debug/IQ.hh" 5411308Santhony.gutierrez@amd.com#include "enums/OpClass.hh" 5511308Santhony.gutierrez@amd.com#include "params/DerivO3CPU.hh" 5611308Santhony.gutierrez@amd.com#include "sim/core.hh" 5711308Santhony.gutierrez@amd.com 5811308Santhony.gutierrez@amd.com// clang complains about std::set being overloaded with Packet::set if 5911308Santhony.gutierrez@amd.com// we open up the entire namespace std 6011308Santhony.gutierrez@amd.comusing std::list; 6111308Santhony.gutierrez@amd.com 6211308Santhony.gutierrez@amd.comtemplate <class Impl> 6311308Santhony.gutierrez@amd.comInstructionQueue<Impl>::FUCompletion::FUCompletion(DynInstPtr &_inst, 6411308Santhony.gutierrez@amd.com int fu_idx, InstructionQueue<Impl> *iq_ptr) 6511308Santhony.gutierrez@amd.com : Event(Stat_Event_Pri, AutoDelete), 6611308Santhony.gutierrez@amd.com inst(_inst), fuIdx(fu_idx), iqPtr(iq_ptr), freeFU(false) 6711308Santhony.gutierrez@amd.com{ 6811308Santhony.gutierrez@amd.com} 6911308Santhony.gutierrez@amd.com 7011308Santhony.gutierrez@amd.comtemplate <class Impl> 7111308Santhony.gutierrez@amd.comvoid 7211308Santhony.gutierrez@amd.comInstructionQueue<Impl>::FUCompletion::process() 7311308Santhony.gutierrez@amd.com{ 7411308Santhony.gutierrez@amd.com iqPtr->processFUCompletion(inst, freeFU ? fuIdx : -1); 7511308Santhony.gutierrez@amd.com inst = NULL; 7611308Santhony.gutierrez@amd.com} 7711308Santhony.gutierrez@amd.com 7811308Santhony.gutierrez@amd.com 7911308Santhony.gutierrez@amd.comtemplate <class Impl> 8011308Santhony.gutierrez@amd.comconst char * 8111308Santhony.gutierrez@amd.comInstructionQueue<Impl>::FUCompletion::description() const 8211308Santhony.gutierrez@amd.com{ 8311308Santhony.gutierrez@amd.com return "Functional unit completion"; 8411308Santhony.gutierrez@amd.com} 8511308Santhony.gutierrez@amd.com 8611308Santhony.gutierrez@amd.comtemplate <class Impl> 8711308Santhony.gutierrez@amd.comInstructionQueue<Impl>::InstructionQueue(O3CPU *cpu_ptr, IEW *iew_ptr, 8811308Santhony.gutierrez@amd.com DerivO3CPUParams *params) 8911308Santhony.gutierrez@amd.com : cpu(cpu_ptr), 9011693Santhony.gutierrez@amd.com iewStage(iew_ptr), 9111693Santhony.gutierrez@amd.com fuPool(params->fuPool), 9211693Santhony.gutierrez@amd.com numEntries(params->numIQEntries), 9311693Santhony.gutierrez@amd.com totalWidth(params->issueWidth), 9411693Santhony.gutierrez@amd.com commitToIEWDelay(params->commitToIEWDelay) 9511693Santhony.gutierrez@amd.com{ 9611308Santhony.gutierrez@amd.com assert(fuPool); 9711308Santhony.gutierrez@amd.com 9811308Santhony.gutierrez@amd.com numThreads = params->numThreads; 9911308Santhony.gutierrez@amd.com 10011308Santhony.gutierrez@amd.com // Set the number of total physical registers 10111308Santhony.gutierrez@amd.com numPhysRegs = params->numPhysIntRegs + params->numPhysFloatRegs + 10211308Santhony.gutierrez@amd.com params->numPhysCCRegs; 10311308Santhony.gutierrez@amd.com 10411308Santhony.gutierrez@amd.com //Create an entry for each physical register within the 10511308Santhony.gutierrez@amd.com //dependency graph. 10611308Santhony.gutierrez@amd.com dependGraph.resize(numPhysRegs); 10711308Santhony.gutierrez@amd.com 10811308Santhony.gutierrez@amd.com // Resize the register scoreboard. 10911308Santhony.gutierrez@amd.com regScoreboard.resize(numPhysRegs); 11011308Santhony.gutierrez@amd.com 11111308Santhony.gutierrez@amd.com //Initialize Mem Dependence Units 11211308Santhony.gutierrez@amd.com for (ThreadID tid = 0; tid < numThreads; tid++) { 11311308Santhony.gutierrez@amd.com memDepUnit[tid].init(params, tid); 11411308Santhony.gutierrez@amd.com memDepUnit[tid].setIQ(this); 11511308Santhony.gutierrez@amd.com } 11611308Santhony.gutierrez@amd.com 11711308Santhony.gutierrez@amd.com resetState(); 11811308Santhony.gutierrez@amd.com 11911308Santhony.gutierrez@amd.com std::string policy = params->smtIQPolicy; 12011308Santhony.gutierrez@amd.com 12111308Santhony.gutierrez@amd.com //Convert string to lowercase 12211308Santhony.gutierrez@amd.com std::transform(policy.begin(), policy.end(), policy.begin(), 12311308Santhony.gutierrez@amd.com (int(*)(int)) tolower); 12411308Santhony.gutierrez@amd.com 12511308Santhony.gutierrez@amd.com //Figure out resource sharing policy 12611308Santhony.gutierrez@amd.com if (policy == "dynamic") { 12711308Santhony.gutierrez@amd.com iqPolicy = Dynamic; 128 129 //Set Max Entries to Total ROB Capacity 130 for (ThreadID tid = 0; tid < numThreads; tid++) { 131 maxEntries[tid] = numEntries; 132 } 133 134 } else if (policy == "partitioned") { 135 iqPolicy = Partitioned; 136 137 //@todo:make work if part_amt doesnt divide evenly. 138 int part_amt = numEntries / numThreads; 139 140 //Divide ROB up evenly 141 for (ThreadID tid = 0; tid < numThreads; tid++) { 142 maxEntries[tid] = part_amt; 143 } 144 145 DPRINTF(IQ, "IQ sharing policy set to Partitioned:" 146 "%i entries per thread.\n",part_amt); 147 } else if (policy == "threshold") { 148 iqPolicy = Threshold; 149 150 double threshold = (double)params->smtIQThreshold / 100; 151 152 int thresholdIQ = (int)((double)threshold * numEntries); 153 154 //Divide up by threshold amount 155 for (ThreadID tid = 0; tid < numThreads; tid++) { 156 maxEntries[tid] = thresholdIQ; 157 } 158 159 DPRINTF(IQ, "IQ sharing policy set to Threshold:" 160 "%i entries per thread.\n",thresholdIQ); 161 } else { 162 assert(0 && "Invalid IQ Sharing Policy.Options Are:{Dynamic," 163 "Partitioned, Threshold}"); 164 } 165} 166 167template <class Impl> 168InstructionQueue<Impl>::~InstructionQueue() 169{ 170 dependGraph.reset(); 171#ifdef DEBUG 172 cprintf("Nodes traversed: %i, removed: %i\n", 173 dependGraph.nodesTraversed, dependGraph.nodesRemoved); 174#endif 175} 176 177template <class Impl> 178std::string 179InstructionQueue<Impl>::name() const 180{ 181 return cpu->name() + ".iq"; 182} 183 184template <class Impl> 185void 186InstructionQueue<Impl>::regStats() 187{ 188 using namespace Stats; 189 iqInstsAdded 190 .name(name() + ".iqInstsAdded") 191 .desc("Number of instructions added to the IQ (excludes non-spec)") 192 .prereq(iqInstsAdded); 193 194 iqNonSpecInstsAdded 195 .name(name() + ".iqNonSpecInstsAdded") 196 .desc("Number of non-speculative instructions added to the IQ") 197 .prereq(iqNonSpecInstsAdded); 198 199 iqInstsIssued 200 .name(name() + ".iqInstsIssued") 201 .desc("Number of instructions issued") 202 .prereq(iqInstsIssued); 203 204 iqIntInstsIssued 205 .name(name() + ".iqIntInstsIssued") 206 .desc("Number of integer instructions issued") 207 .prereq(iqIntInstsIssued); 208 209 iqFloatInstsIssued 210 .name(name() + ".iqFloatInstsIssued") 211 .desc("Number of float instructions issued") 212 .prereq(iqFloatInstsIssued); 213 214 iqBranchInstsIssued 215 .name(name() + ".iqBranchInstsIssued") 216 .desc("Number of branch instructions issued") 217 .prereq(iqBranchInstsIssued); 218 219 iqMemInstsIssued 220 .name(name() + ".iqMemInstsIssued") 221 .desc("Number of memory instructions issued") 222 .prereq(iqMemInstsIssued); 223 224 iqMiscInstsIssued 225 .name(name() + ".iqMiscInstsIssued") 226 .desc("Number of miscellaneous instructions issued") 227 .prereq(iqMiscInstsIssued); 228 229 iqSquashedInstsIssued 230 .name(name() + ".iqSquashedInstsIssued") 231 .desc("Number of squashed instructions issued") 232 .prereq(iqSquashedInstsIssued); 233 234 iqSquashedInstsExamined 235 .name(name() + ".iqSquashedInstsExamined") 236 .desc("Number of squashed instructions iterated over during squash;" 237 " mainly for profiling") 238 .prereq(iqSquashedInstsExamined); 239 240 iqSquashedOperandsExamined 241 .name(name() + ".iqSquashedOperandsExamined") 242 .desc("Number of squashed operands that are examined and possibly " 243 "removed from graph") 244 .prereq(iqSquashedOperandsExamined); 245 246 iqSquashedNonSpecRemoved 247 .name(name() + ".iqSquashedNonSpecRemoved") 248 .desc("Number of squashed non-spec instructions that were removed") 249 .prereq(iqSquashedNonSpecRemoved); 250/* 251 queueResDist 252 .init(Num_OpClasses, 0, 99, 2) 253 .name(name() + ".IQ:residence:") 254 .desc("cycles from dispatch to issue") 255 .flags(total | pdf | cdf ) 256 ; 257 for (int i = 0; i < Num_OpClasses; ++i) { 258 queueResDist.subname(i, opClassStrings[i]); 259 } 260*/ 261 numIssuedDist 262 .init(0,totalWidth,1) 263 .name(name() + ".issued_per_cycle") 264 .desc("Number of insts issued each cycle") 265 .flags(pdf) 266 ; 267/* 268 dist_unissued 269 .init(Num_OpClasses+2) 270 .name(name() + ".unissued_cause") 271 .desc("Reason ready instruction not issued") 272 .flags(pdf | dist) 273 ; 274 for (int i=0; i < (Num_OpClasses + 2); ++i) { 275 dist_unissued.subname(i, unissued_names[i]); 276 } 277*/ 278 statIssuedInstType 279 .init(numThreads,Enums::Num_OpClass) 280 .name(name() + ".FU_type") 281 .desc("Type of FU issued") 282 .flags(total | pdf | dist) 283 ; 284 statIssuedInstType.ysubnames(Enums::OpClassStrings); 285 286 // 287 // How long did instructions for a particular FU type wait prior to issue 288 // 289/* 290 issueDelayDist 291 .init(Num_OpClasses,0,99,2) 292 .name(name() + ".") 293 .desc("cycles from operands ready to issue") 294 .flags(pdf | cdf) 295 ; 296 297 for (int i=0; i<Num_OpClasses; ++i) { 298 std::stringstream subname; 299 subname << opClassStrings[i] << "_delay"; 300 issueDelayDist.subname(i, subname.str()); 301 } 302*/ 303 issueRate 304 .name(name() + ".rate") 305 .desc("Inst issue rate") 306 .flags(total) 307 ; 308 issueRate = iqInstsIssued / cpu->numCycles; 309 310 statFuBusy 311 .init(Num_OpClasses) 312 .name(name() + ".fu_full") 313 .desc("attempts to use FU when none available") 314 .flags(pdf | dist) 315 ; 316 for (int i=0; i < Num_OpClasses; ++i) { 317 statFuBusy.subname(i, Enums::OpClassStrings[i]); 318 } 319 320 fuBusy 321 .init(numThreads) 322 .name(name() + ".fu_busy_cnt") 323 .desc("FU busy when requested") 324 .flags(total) 325 ; 326 327 fuBusyRate 328 .name(name() + ".fu_busy_rate") 329 .desc("FU busy rate (busy events/executed inst)") 330 .flags(total) 331 ; 332 fuBusyRate = fuBusy / iqInstsIssued; 333 334 for (ThreadID tid = 0; tid < numThreads; tid++) { 335 // Tell mem dependence unit to reg stats as well. 336 memDepUnit[tid].regStats(); 337 } 338 339 intInstQueueReads 340 .name(name() + ".int_inst_queue_reads") 341 .desc("Number of integer instruction queue reads") 342 .flags(total); 343 344 intInstQueueWrites 345 .name(name() + ".int_inst_queue_writes") 346 .desc("Number of integer instruction queue writes") 347 .flags(total); 348 349 intInstQueueWakeupAccesses 350 .name(name() + ".int_inst_queue_wakeup_accesses") 351 .desc("Number of integer instruction queue wakeup accesses") 352 .flags(total); 353 354 fpInstQueueReads 355 .name(name() + ".fp_inst_queue_reads") 356 .desc("Number of floating instruction queue reads") 357 .flags(total); 358 359 fpInstQueueWrites 360 .name(name() + ".fp_inst_queue_writes") 361 .desc("Number of floating instruction queue writes") 362 .flags(total); 363 364 fpInstQueueWakeupQccesses 365 .name(name() + ".fp_inst_queue_wakeup_accesses") 366 .desc("Number of floating instruction queue wakeup accesses") 367 .flags(total); 368 369 intAluAccesses 370 .name(name() + ".int_alu_accesses") 371 .desc("Number of integer alu accesses") 372 .flags(total); 373 374 fpAluAccesses 375 .name(name() + ".fp_alu_accesses") 376 .desc("Number of floating point alu accesses") 377 .flags(total); 378 379} 380 381template <class Impl> 382void 383InstructionQueue<Impl>::resetState() 384{ 385 //Initialize thread IQ counts 386 for (ThreadID tid = 0; tid <numThreads; tid++) { 387 count[tid] = 0; 388 instList[tid].clear(); 389 } 390 391 // Initialize the number of free IQ entries. 392 freeEntries = numEntries; 393 394 // Note that in actuality, the registers corresponding to the logical 395 // registers start off as ready. However this doesn't matter for the 396 // IQ as the instruction should have been correctly told if those 397 // registers are ready in rename. Thus it can all be initialized as 398 // unready. 399 for (int i = 0; i < numPhysRegs; ++i) { 400 regScoreboard[i] = false; 401 } 402 403 for (ThreadID tid = 0; tid < numThreads; ++tid) { 404 squashedSeqNum[tid] = 0; 405 } 406 407 for (int i = 0; i < Num_OpClasses; ++i) { 408 while (!readyInsts[i].empty()) 409 readyInsts[i].pop(); 410 queueOnList[i] = false; 411 readyIt[i] = listOrder.end(); 412 } 413 nonSpecInsts.clear(); 414 listOrder.clear(); 415 deferredMemInsts.clear(); 416} 417 418template <class Impl> 419void 420InstructionQueue<Impl>::setActiveThreads(list<ThreadID> *at_ptr) 421{ 422 activeThreads = at_ptr; 423} 424 425template <class Impl> 426void 427InstructionQueue<Impl>::setIssueToExecuteQueue(TimeBuffer<IssueStruct> *i2e_ptr) 428{ 429 issueToExecuteQueue = i2e_ptr; 430} 431 432template <class Impl> 433void 434InstructionQueue<Impl>::setTimeBuffer(TimeBuffer<TimeStruct> *tb_ptr) 435{ 436 timeBuffer = tb_ptr; 437 438 fromCommit = timeBuffer->getWire(-commitToIEWDelay); 439} 440 441template <class Impl> 442void 443InstructionQueue<Impl>::drainSanityCheck() const 444{ 445 assert(dependGraph.empty()); 446 assert(instsToExecute.empty()); 447 for (ThreadID tid = 0; tid < numThreads; ++tid) 448 memDepUnit[tid].drainSanityCheck(); 449} 450 451template <class Impl> 452void 453InstructionQueue<Impl>::takeOverFrom() 454{ 455 resetState(); 456} 457 458template <class Impl> 459int 460InstructionQueue<Impl>::entryAmount(ThreadID num_threads) 461{ 462 if (iqPolicy == Partitioned) { 463 return numEntries / num_threads; 464 } else { 465 return 0; 466 } 467} 468 469 470template <class Impl> 471void 472InstructionQueue<Impl>::resetEntries() 473{ 474 if (iqPolicy != Dynamic || numThreads > 1) { 475 int active_threads = activeThreads->size(); 476 477 list<ThreadID>::iterator threads = activeThreads->begin(); 478 list<ThreadID>::iterator end = activeThreads->end(); 479 480 while (threads != end) { 481 ThreadID tid = *threads++; 482 483 if (iqPolicy == Partitioned) { 484 maxEntries[tid] = numEntries / active_threads; 485 } else if(iqPolicy == Threshold && active_threads == 1) { 486 maxEntries[tid] = numEntries; 487 } 488 } 489 } 490} 491 492template <class Impl> 493unsigned 494InstructionQueue<Impl>::numFreeEntries() 495{ 496 return freeEntries; 497} 498 499template <class Impl> 500unsigned 501InstructionQueue<Impl>::numFreeEntries(ThreadID tid) 502{ 503 return maxEntries[tid] - count[tid]; 504} 505 506// Might want to do something more complex if it knows how many instructions 507// will be issued this cycle. 508template <class Impl> 509bool 510InstructionQueue<Impl>::isFull() 511{ 512 if (freeEntries == 0) { 513 return(true); 514 } else { 515 return(false); 516 } 517} 518 519template <class Impl> 520bool 521InstructionQueue<Impl>::isFull(ThreadID tid) 522{ 523 if (numFreeEntries(tid) == 0) { 524 return(true); 525 } else { 526 return(false); 527 } 528} 529 530template <class Impl> 531bool 532InstructionQueue<Impl>::hasReadyInsts() 533{ 534 if (!listOrder.empty()) { 535 return true; 536 } 537 538 for (int i = 0; i < Num_OpClasses; ++i) { 539 if (!readyInsts[i].empty()) { 540 return true; 541 } 542 } 543 544 return false; 545} 546 547template <class Impl> 548void 549InstructionQueue<Impl>::insert(DynInstPtr &new_inst) 550{ 551 new_inst->isFloating() ? fpInstQueueWrites++ : intInstQueueWrites++; 552 // Make sure the instruction is valid 553 assert(new_inst); 554 555 DPRINTF(IQ, "Adding instruction [sn:%lli] PC %s to the IQ.\n", 556 new_inst->seqNum, new_inst->pcState()); 557 558 assert(freeEntries != 0); 559 560 instList[new_inst->threadNumber].push_back(new_inst); 561 562 --freeEntries; 563 564 new_inst->setInIQ(); 565 566 // Look through its source registers (physical regs), and mark any 567 // dependencies. 568 addToDependents(new_inst); 569 570 // Have this instruction set itself as the producer of its destination 571 // register(s). 572 addToProducers(new_inst); 573 574 if (new_inst->isMemRef()) { 575 memDepUnit[new_inst->threadNumber].insert(new_inst); 576 } else { 577 addIfReady(new_inst); 578 } 579 580 ++iqInstsAdded; 581 582 count[new_inst->threadNumber]++; 583 584 assert(freeEntries == (numEntries - countInsts())); 585} 586 587template <class Impl> 588void 589InstructionQueue<Impl>::insertNonSpec(DynInstPtr &new_inst) 590{ 591 // @todo: Clean up this code; can do it by setting inst as unable 592 // to issue, then calling normal insert on the inst. 593 new_inst->isFloating() ? fpInstQueueWrites++ : intInstQueueWrites++; 594 595 assert(new_inst); 596 597 nonSpecInsts[new_inst->seqNum] = new_inst; 598 599 DPRINTF(IQ, "Adding non-speculative instruction [sn:%lli] PC %s " 600 "to the IQ.\n", 601 new_inst->seqNum, new_inst->pcState()); 602 603 assert(freeEntries != 0); 604 605 instList[new_inst->threadNumber].push_back(new_inst); 606 607 --freeEntries; 608 609 new_inst->setInIQ(); 610 611 // Have this instruction set itself as the producer of its destination 612 // register(s). 613 addToProducers(new_inst); 614 615 // If it's a memory instruction, add it to the memory dependency 616 // unit. 617 if (new_inst->isMemRef()) { 618 memDepUnit[new_inst->threadNumber].insertNonSpec(new_inst); 619 } 620 621 ++iqNonSpecInstsAdded; 622 623 count[new_inst->threadNumber]++; 624 625 assert(freeEntries == (numEntries - countInsts())); 626} 627 628template <class Impl> 629void 630InstructionQueue<Impl>::insertBarrier(DynInstPtr &barr_inst) 631{ 632 memDepUnit[barr_inst->threadNumber].insertBarrier(barr_inst); 633 634 insertNonSpec(barr_inst); 635} 636 637template <class Impl> 638typename Impl::DynInstPtr 639InstructionQueue<Impl>::getInstToExecute() 640{ 641 assert(!instsToExecute.empty()); 642 DynInstPtr inst = instsToExecute.front(); 643 instsToExecute.pop_front(); 644 if (inst->isFloating()){ 645 fpInstQueueReads++; 646 } else { 647 intInstQueueReads++; 648 } 649 return inst; 650} 651 652template <class Impl> 653void 654InstructionQueue<Impl>::addToOrderList(OpClass op_class) 655{ 656 assert(!readyInsts[op_class].empty()); 657 658 ListOrderEntry queue_entry; 659 660 queue_entry.queueType = op_class; 661 662 queue_entry.oldestInst = readyInsts[op_class].top()->seqNum; 663 664 ListOrderIt list_it = listOrder.begin(); 665 ListOrderIt list_end_it = listOrder.end(); 666 667 while (list_it != list_end_it) { 668 if ((*list_it).oldestInst > queue_entry.oldestInst) { 669 break; 670 } 671 672 list_it++; 673 } 674 675 readyIt[op_class] = listOrder.insert(list_it, queue_entry); 676 queueOnList[op_class] = true; 677} 678 679template <class Impl> 680void 681InstructionQueue<Impl>::moveToYoungerInst(ListOrderIt list_order_it) 682{ 683 // Get iterator of next item on the list 684 // Delete the original iterator 685 // Determine if the next item is either the end of the list or younger 686 // than the new instruction. If so, then add in a new iterator right here. 687 // If not, then move along. 688 ListOrderEntry queue_entry; 689 OpClass op_class = (*list_order_it).queueType; 690 ListOrderIt next_it = list_order_it; 691 692 ++next_it; 693 694 queue_entry.queueType = op_class; 695 queue_entry.oldestInst = readyInsts[op_class].top()->seqNum; 696 697 while (next_it != listOrder.end() && 698 (*next_it).oldestInst < queue_entry.oldestInst) { 699 ++next_it; 700 } 701 702 readyIt[op_class] = listOrder.insert(next_it, queue_entry); 703} 704 705template <class Impl> 706void 707InstructionQueue<Impl>::processFUCompletion(DynInstPtr &inst, int fu_idx) 708{ 709 DPRINTF(IQ, "Processing FU completion [sn:%lli]\n", inst->seqNum); 710 assert(!cpu->switchedOut()); 711 // The CPU could have been sleeping until this op completed (*extremely* 712 // long latency op). Wake it if it was. This may be overkill. 713 iewStage->wakeCPU(); 714 715 if (fu_idx > -1) 716 fuPool->freeUnitNextCycle(fu_idx); 717 718 // @todo: Ensure that these FU Completions happen at the beginning 719 // of a cycle, otherwise they could add too many instructions to 720 // the queue. 721 issueToExecuteQueue->access(-1)->size++; 722 instsToExecute.push_back(inst); 723} 724 725// @todo: Figure out a better way to remove the squashed items from the 726// lists. Checking the top item of each list to see if it's squashed 727// wastes time and forces jumps. 728template <class Impl> 729void 730InstructionQueue<Impl>::scheduleReadyInsts() 731{ 732 DPRINTF(IQ, "Attempting to schedule ready instructions from " 733 "the IQ.\n"); 734 735 IssueStruct *i2e_info = issueToExecuteQueue->access(0); 736 737 DynInstPtr deferred_mem_inst; 738 int total_deferred_mem_issued = 0; 739 while (total_deferred_mem_issued < totalWidth && 740 (deferred_mem_inst = getDeferredMemInstToExecute()) != 0) { 741 issueToExecuteQueue->access(0)->size++; 742 instsToExecute.push_back(deferred_mem_inst); 743 total_deferred_mem_issued++; 744 } 745 746 // Have iterator to head of the list 747 // While I haven't exceeded bandwidth or reached the end of the list, 748 // Try to get a FU that can do what this op needs. 749 // If successful, change the oldestInst to the new top of the list, put 750 // the queue in the proper place in the list. 751 // Increment the iterator. 752 // This will avoid trying to schedule a certain op class if there are no 753 // FUs that handle it. 754 ListOrderIt order_it = listOrder.begin(); 755 ListOrderIt order_end_it = listOrder.end(); 756 int total_issued = 0; 757 758 while (total_issued < (totalWidth - total_deferred_mem_issued) && 759 iewStage->canIssue() && 760 order_it != order_end_it) { 761 OpClass op_class = (*order_it).queueType; 762 763 assert(!readyInsts[op_class].empty()); 764 765 DynInstPtr issuing_inst = readyInsts[op_class].top(); 766 767 issuing_inst->isFloating() ? fpInstQueueReads++ : intInstQueueReads++; 768 769 assert(issuing_inst->seqNum == (*order_it).oldestInst); 770 771 if (issuing_inst->isSquashed()) { 772 readyInsts[op_class].pop(); 773 774 if (!readyInsts[op_class].empty()) { 775 moveToYoungerInst(order_it); 776 } else { 777 readyIt[op_class] = listOrder.end(); 778 queueOnList[op_class] = false; 779 } 780 781 listOrder.erase(order_it++); 782 783 ++iqSquashedInstsIssued; 784 785 continue; 786 } 787 788 int idx = -2; 789 Cycles op_latency = Cycles(1); 790 ThreadID tid = issuing_inst->threadNumber; 791 792 if (op_class != No_OpClass) { 793 idx = fuPool->getUnit(op_class); 794 issuing_inst->isFloating() ? fpAluAccesses++ : intAluAccesses++; 795 if (idx > -1) { 796 op_latency = fuPool->getOpLatency(op_class); 797 } 798 } 799 800 // If we have an instruction that doesn't require a FU, or a 801 // valid FU, then schedule for execution. 802 if (idx == -2 || idx != -1) { 803 if (op_latency == Cycles(1)) { 804 i2e_info->size++; 805 instsToExecute.push_back(issuing_inst); 806 807 // Add the FU onto the list of FU's to be freed next 808 // cycle if we used one. 809 if (idx >= 0) 810 fuPool->freeUnitNextCycle(idx); 811 } else { 812 Cycles issue_latency = fuPool->getIssueLatency(op_class); 813 // Generate completion event for the FU 814 FUCompletion *execution = new FUCompletion(issuing_inst, 815 idx, this); 816 817 cpu->schedule(execution, 818 cpu->clockEdge(Cycles(op_latency - 1))); 819 820 // @todo: Enforce that issue_latency == 1 or op_latency 821 if (issue_latency > Cycles(1)) { 822 // If FU isn't pipelined, then it must be freed 823 // upon the execution completing. 824 execution->setFreeFU(); 825 } else { 826 // Add the FU onto the list of FU's to be freed next cycle. 827 fuPool->freeUnitNextCycle(idx); 828 } 829 } 830 831 DPRINTF(IQ, "Thread %i: Issuing instruction PC %s " 832 "[sn:%lli]\n", 833 tid, issuing_inst->pcState(), 834 issuing_inst->seqNum); 835 836 readyInsts[op_class].pop(); 837 838 if (!readyInsts[op_class].empty()) { 839 moveToYoungerInst(order_it); 840 } else { 841 readyIt[op_class] = listOrder.end(); 842 queueOnList[op_class] = false; 843 } 844 845 issuing_inst->setIssued(); 846 ++total_issued; 847 848#if TRACING_ON 849 issuing_inst->issueTick = curTick() - issuing_inst->fetchTick; 850#endif 851 852 if (!issuing_inst->isMemRef()) { 853 // Memory instructions can not be freed from the IQ until they 854 // complete. 855 ++freeEntries; 856 count[tid]--; 857 issuing_inst->clearInIQ(); 858 } else { 859 memDepUnit[tid].issue(issuing_inst); 860 } 861 862 listOrder.erase(order_it++); 863 statIssuedInstType[tid][op_class]++; 864 iewStage->incrWb(issuing_inst->seqNum); 865 } else { 866 statFuBusy[op_class]++; 867 fuBusy[tid]++; 868 ++order_it; 869 } 870 } 871 872 numIssuedDist.sample(total_issued); 873 iqInstsIssued+= total_issued; 874 875 // If we issued any instructions, tell the CPU we had activity. 876 // @todo If the way deferred memory instructions are handeled due to 877 // translation changes then the deferredMemInsts condition should be removed 878 // from the code below. 879 if (total_issued || total_deferred_mem_issued || deferredMemInsts.size()) { 880 cpu->activityThisCycle(); 881 } else { 882 DPRINTF(IQ, "Not able to schedule any instructions.\n"); 883 } 884} 885 886template <class Impl> 887void 888InstructionQueue<Impl>::scheduleNonSpec(const InstSeqNum &inst) 889{ 890 DPRINTF(IQ, "Marking nonspeculative instruction [sn:%lli] as ready " 891 "to execute.\n", inst); 892 893 NonSpecMapIt inst_it = nonSpecInsts.find(inst); 894 895 assert(inst_it != nonSpecInsts.end()); 896 897 ThreadID tid = (*inst_it).second->threadNumber; 898 899 (*inst_it).second->setAtCommit(); 900 901 (*inst_it).second->setCanIssue(); 902 903 if (!(*inst_it).second->isMemRef()) { 904 addIfReady((*inst_it).second); 905 } else { 906 memDepUnit[tid].nonSpecInstReady((*inst_it).second); 907 } 908 909 (*inst_it).second = NULL; 910 911 nonSpecInsts.erase(inst_it); 912} 913 914template <class Impl> 915void 916InstructionQueue<Impl>::commit(const InstSeqNum &inst, ThreadID tid) 917{ 918 DPRINTF(IQ, "[tid:%i]: Committing instructions older than [sn:%i]\n", 919 tid,inst); 920 921 ListIt iq_it = instList[tid].begin(); 922 923 while (iq_it != instList[tid].end() && 924 (*iq_it)->seqNum <= inst) { 925 ++iq_it; 926 instList[tid].pop_front(); 927 } 928 929 assert(freeEntries == (numEntries - countInsts())); 930} 931 932template <class Impl> 933int 934InstructionQueue<Impl>::wakeDependents(DynInstPtr &completed_inst) 935{ 936 int dependents = 0; 937 938 // The instruction queue here takes care of both floating and int ops 939 if (completed_inst->isFloating()) { 940 fpInstQueueWakeupQccesses++; 941 } else { 942 intInstQueueWakeupAccesses++; 943 } 944 945 DPRINTF(IQ, "Waking dependents of completed instruction.\n"); 946 947 assert(!completed_inst->isSquashed()); 948 949 // Tell the memory dependence unit to wake any dependents on this 950 // instruction if it is a memory instruction. Also complete the memory 951 // instruction at this point since we know it executed without issues. 952 // @todo: Might want to rename "completeMemInst" to something that 953 // indicates that it won't need to be replayed, and call this 954 // earlier. Might not be a big deal. 955 if (completed_inst->isMemRef()) { 956 memDepUnit[completed_inst->threadNumber].wakeDependents(completed_inst); 957 completeMemInst(completed_inst); 958 } else if (completed_inst->isMemBarrier() || 959 completed_inst->isWriteBarrier()) { 960 memDepUnit[completed_inst->threadNumber].completeBarrier(completed_inst); 961 } 962 963 for (int dest_reg_idx = 0; 964 dest_reg_idx < completed_inst->numDestRegs(); 965 dest_reg_idx++) 966 { 967 PhysRegIndex dest_reg = 968 completed_inst->renamedDestRegIdx(dest_reg_idx); 969 970 // Special case of uniq or control registers. They are not 971 // handled by the IQ and thus have no dependency graph entry. 972 // @todo Figure out a cleaner way to handle this. 973 if (dest_reg >= numPhysRegs) { 974 DPRINTF(IQ, "dest_reg :%d, numPhysRegs: %d\n", dest_reg, 975 numPhysRegs); 976 continue; 977 } 978 979 DPRINTF(IQ, "Waking any dependents on register %i.\n", 980 (int) dest_reg); 981 982 //Go through the dependency chain, marking the registers as 983 //ready within the waiting instructions. 984 DynInstPtr dep_inst = dependGraph.pop(dest_reg); 985 986 while (dep_inst) { 987 DPRINTF(IQ, "Waking up a dependent instruction, [sn:%lli] " 988 "PC %s.\n", dep_inst->seqNum, dep_inst->pcState()); 989 990 // Might want to give more information to the instruction 991 // so that it knows which of its source registers is 992 // ready. However that would mean that the dependency 993 // graph entries would need to hold the src_reg_idx. 994 dep_inst->markSrcRegReady(); 995 996 addIfReady(dep_inst); 997 998 dep_inst = dependGraph.pop(dest_reg); 999 1000 ++dependents; 1001 } 1002 1003 // Reset the head node now that all of its dependents have 1004 // been woken up. 1005 assert(dependGraph.empty(dest_reg)); 1006 dependGraph.clearInst(dest_reg); 1007 1008 // Mark the scoreboard as having that register ready. 1009 regScoreboard[dest_reg] = true; 1010 } 1011 return dependents; 1012} 1013 1014template <class Impl> 1015void 1016InstructionQueue<Impl>::addReadyMemInst(DynInstPtr &ready_inst) 1017{ 1018 OpClass op_class = ready_inst->opClass(); 1019 1020 readyInsts[op_class].push(ready_inst); 1021 1022 // Will need to reorder the list if either a queue is not on the list, 1023 // or it has an older instruction than last time. 1024 if (!queueOnList[op_class]) { 1025 addToOrderList(op_class); 1026 } else if (readyInsts[op_class].top()->seqNum < 1027 (*readyIt[op_class]).oldestInst) { 1028 listOrder.erase(readyIt[op_class]); 1029 addToOrderList(op_class); 1030 } 1031 1032 DPRINTF(IQ, "Instruction is ready to issue, putting it onto " 1033 "the ready list, PC %s opclass:%i [sn:%lli].\n", 1034 ready_inst->pcState(), op_class, ready_inst->seqNum); 1035} 1036 1037template <class Impl> 1038void 1039InstructionQueue<Impl>::rescheduleMemInst(DynInstPtr &resched_inst) 1040{ 1041 DPRINTF(IQ, "Rescheduling mem inst [sn:%lli]\n", resched_inst->seqNum); 1042 1043 // Reset DTB translation state 1044 resched_inst->translationStarted(false); 1045 resched_inst->translationCompleted(false); 1046 1047 resched_inst->clearCanIssue(); 1048 memDepUnit[resched_inst->threadNumber].reschedule(resched_inst); 1049} 1050 1051template <class Impl> 1052void 1053InstructionQueue<Impl>::replayMemInst(DynInstPtr &replay_inst) 1054{ 1055 memDepUnit[replay_inst->threadNumber].replay(replay_inst); 1056} 1057 1058template <class Impl> 1059void 1060InstructionQueue<Impl>::completeMemInst(DynInstPtr &completed_inst) 1061{ 1062 ThreadID tid = completed_inst->threadNumber; 1063 1064 DPRINTF(IQ, "Completing mem instruction PC: %s [sn:%lli]\n", 1065 completed_inst->pcState(), completed_inst->seqNum); 1066 1067 ++freeEntries; 1068 1069 completed_inst->memOpDone(true); 1070 1071 memDepUnit[tid].completed(completed_inst); 1072 count[tid]--; 1073} 1074 1075template <class Impl> 1076void 1077InstructionQueue<Impl>::deferMemInst(DynInstPtr &deferred_inst) 1078{ 1079 deferredMemInsts.push_back(deferred_inst); 1080} 1081 1082template <class Impl> 1083typename Impl::DynInstPtr 1084InstructionQueue<Impl>::getDeferredMemInstToExecute() 1085{ 1086 for (ListIt it = deferredMemInsts.begin(); it != deferredMemInsts.end(); 1087 ++it) { 1088 if ((*it)->translationCompleted() || (*it)->isSquashed()) { 1089 DynInstPtr ret = *it; 1090 deferredMemInsts.erase(it); 1091 return ret; 1092 } 1093 } 1094 return NULL; 1095} 1096 1097template <class Impl> 1098void 1099InstructionQueue<Impl>::violation(DynInstPtr &store, 1100 DynInstPtr &faulting_load) 1101{ 1102 intInstQueueWrites++; 1103 memDepUnit[store->threadNumber].violation(store, faulting_load); 1104} 1105 1106template <class Impl> 1107void 1108InstructionQueue<Impl>::squash(ThreadID tid) 1109{ 1110 DPRINTF(IQ, "[tid:%i]: Starting to squash instructions in " 1111 "the IQ.\n", tid); 1112 1113 // Read instruction sequence number of last instruction out of the 1114 // time buffer. 1115 squashedSeqNum[tid] = fromCommit->commitInfo[tid].doneSeqNum; 1116 1117 // Call doSquash if there are insts in the IQ 1118 if (count[tid] > 0) { 1119 doSquash(tid); 1120 } 1121 1122 // Also tell the memory dependence unit to squash. 1123 memDepUnit[tid].squash(squashedSeqNum[tid], tid); 1124} 1125 1126template <class Impl> 1127void 1128InstructionQueue<Impl>::doSquash(ThreadID tid) 1129{ 1130 // Start at the tail. 1131 ListIt squash_it = instList[tid].end(); 1132 --squash_it; 1133 1134 DPRINTF(IQ, "[tid:%i]: Squashing until sequence number %i!\n", 1135 tid, squashedSeqNum[tid]); 1136 1137 // Squash any instructions younger than the squashed sequence number 1138 // given. 1139 while (squash_it != instList[tid].end() && 1140 (*squash_it)->seqNum > squashedSeqNum[tid]) { 1141 1142 DynInstPtr squashed_inst = (*squash_it); 1143 squashed_inst->isFloating() ? fpInstQueueWrites++ : intInstQueueWrites++; 1144 1145 // Only handle the instruction if it actually is in the IQ and 1146 // hasn't already been squashed in the IQ. 1147 if (squashed_inst->threadNumber != tid || 1148 squashed_inst->isSquashedInIQ()) { 1149 --squash_it; 1150 continue; 1151 } 1152 1153 if (!squashed_inst->isIssued() || 1154 (squashed_inst->isMemRef() && 1155 !squashed_inst->memOpDone())) { 1156 1157 DPRINTF(IQ, "[tid:%i]: Instruction [sn:%lli] PC %s squashed.\n", 1158 tid, squashed_inst->seqNum, squashed_inst->pcState()); 1159 1160 // Remove the instruction from the dependency list. 1161 if (!squashed_inst->isNonSpeculative() && 1162 !squashed_inst->isStoreConditional() && 1163 !squashed_inst->isMemBarrier() && 1164 !squashed_inst->isWriteBarrier()) { 1165 1166 for (int src_reg_idx = 0; 1167 src_reg_idx < squashed_inst->numSrcRegs(); 1168 src_reg_idx++) 1169 { 1170 PhysRegIndex src_reg = 1171 squashed_inst->renamedSrcRegIdx(src_reg_idx); 1172 1173 // Only remove it from the dependency graph if it 1174 // was placed there in the first place. 1175 1176 // Instead of doing a linked list traversal, we 1177 // can just remove these squashed instructions 1178 // either at issue time, or when the register is 1179 // overwritten. The only downside to this is it 1180 // leaves more room for error. 1181 1182 if (!squashed_inst->isReadySrcRegIdx(src_reg_idx) && 1183 src_reg < numPhysRegs) { 1184 dependGraph.remove(src_reg, squashed_inst); 1185 } 1186 1187 1188 ++iqSquashedOperandsExamined; 1189 } 1190 } else if (!squashed_inst->isStoreConditional() || 1191 !squashed_inst->isCompleted()) { 1192 NonSpecMapIt ns_inst_it = 1193 nonSpecInsts.find(squashed_inst->seqNum); 1194 1195 // we remove non-speculative instructions from 1196 // nonSpecInsts already when they are ready, and so we 1197 // cannot always expect to find them 1198 if (ns_inst_it == nonSpecInsts.end()) { 1199 // loads that became ready but stalled on a 1200 // blocked cache are alreayd removed from 1201 // nonSpecInsts, and have not faulted 1202 assert(squashed_inst->getFault() != NoFault || 1203 squashed_inst->isMemRef()); 1204 } else { 1205 1206 (*ns_inst_it).second = NULL; 1207 1208 nonSpecInsts.erase(ns_inst_it); 1209 1210 ++iqSquashedNonSpecRemoved; 1211 } 1212 } 1213 1214 // Might want to also clear out the head of the dependency graph. 1215 1216 // Mark it as squashed within the IQ. 1217 squashed_inst->setSquashedInIQ(); 1218 1219 // @todo: Remove this hack where several statuses are set so the 1220 // inst will flow through the rest of the pipeline. 1221 squashed_inst->setIssued(); 1222 squashed_inst->setCanCommit(); 1223 squashed_inst->clearInIQ(); 1224 1225 //Update Thread IQ Count 1226 count[squashed_inst->threadNumber]--; 1227 1228 ++freeEntries; 1229 } 1230 1231 instList[tid].erase(squash_it--); 1232 ++iqSquashedInstsExamined; 1233 } 1234} 1235 1236template <class Impl> 1237bool 1238InstructionQueue<Impl>::addToDependents(DynInstPtr &new_inst) 1239{ 1240 // Loop through the instruction's source registers, adding 1241 // them to the dependency list if they are not ready. 1242 int8_t total_src_regs = new_inst->numSrcRegs(); 1243 bool return_val = false; 1244 1245 for (int src_reg_idx = 0; 1246 src_reg_idx < total_src_regs; 1247 src_reg_idx++) 1248 { 1249 // Only add it to the dependency graph if it's not ready. 1250 if (!new_inst->isReadySrcRegIdx(src_reg_idx)) { 1251 PhysRegIndex src_reg = new_inst->renamedSrcRegIdx(src_reg_idx); 1252 1253 // Check the IQ's scoreboard to make sure the register 1254 // hasn't become ready while the instruction was in flight 1255 // between stages. Only if it really isn't ready should 1256 // it be added to the dependency graph. 1257 if (src_reg >= numPhysRegs) { 1258 continue; 1259 } else if (regScoreboard[src_reg] == false) { 1260 DPRINTF(IQ, "Instruction PC %s has src reg %i that " 1261 "is being added to the dependency chain.\n", 1262 new_inst->pcState(), src_reg); 1263 1264 dependGraph.insert(src_reg, new_inst); 1265 1266 // Change the return value to indicate that something 1267 // was added to the dependency graph. 1268 return_val = true; 1269 } else { 1270 DPRINTF(IQ, "Instruction PC %s has src reg %i that " 1271 "became ready before it reached the IQ.\n", 1272 new_inst->pcState(), src_reg); 1273 // Mark a register ready within the instruction. 1274 new_inst->markSrcRegReady(src_reg_idx); 1275 } 1276 } 1277 } 1278 1279 return return_val; 1280} 1281 1282template <class Impl> 1283void 1284InstructionQueue<Impl>::addToProducers(DynInstPtr &new_inst) 1285{ 1286 // Nothing really needs to be marked when an instruction becomes 1287 // the producer of a register's value, but for convenience a ptr 1288 // to the producing instruction will be placed in the head node of 1289 // the dependency links. 1290 int8_t total_dest_regs = new_inst->numDestRegs(); 1291 1292 for (int dest_reg_idx = 0; 1293 dest_reg_idx < total_dest_regs; 1294 dest_reg_idx++) 1295 { 1296 PhysRegIndex dest_reg = new_inst->renamedDestRegIdx(dest_reg_idx); 1297 1298 // Instructions that use the misc regs will have a reg number 1299 // higher than the normal physical registers. In this case these 1300 // registers are not renamed, and there is no need to track 1301 // dependencies as these instructions must be executed at commit. 1302 if (dest_reg >= numPhysRegs) { 1303 continue; 1304 } 1305 1306 if (!dependGraph.empty(dest_reg)) { 1307 dependGraph.dump(); 1308 panic("Dependency graph %i not empty!", dest_reg); 1309 } 1310 1311 dependGraph.setInst(dest_reg, new_inst); 1312 1313 // Mark the scoreboard to say it's not yet ready. 1314 regScoreboard[dest_reg] = false; 1315 } 1316} 1317 1318template <class Impl> 1319void 1320InstructionQueue<Impl>::addIfReady(DynInstPtr &inst) 1321{ 1322 // If the instruction now has all of its source registers 1323 // available, then add it to the list of ready instructions. 1324 if (inst->readyToIssue()) { 1325 1326 //Add the instruction to the proper ready list. 1327 if (inst->isMemRef()) { 1328 1329 DPRINTF(IQ, "Checking if memory instruction can issue.\n"); 1330 1331 // Message to the mem dependence unit that this instruction has 1332 // its registers ready. 1333 memDepUnit[inst->threadNumber].regsReady(inst); 1334 1335 return; 1336 } 1337 1338 OpClass op_class = inst->opClass(); 1339 1340 DPRINTF(IQ, "Instruction is ready to issue, putting it onto " 1341 "the ready list, PC %s opclass:%i [sn:%lli].\n", 1342 inst->pcState(), op_class, inst->seqNum); 1343 1344 readyInsts[op_class].push(inst); 1345 1346 // Will need to reorder the list if either a queue is not on the list, 1347 // or it has an older instruction than last time. 1348 if (!queueOnList[op_class]) { 1349 addToOrderList(op_class); 1350 } else if (readyInsts[op_class].top()->seqNum < 1351 (*readyIt[op_class]).oldestInst) { 1352 listOrder.erase(readyIt[op_class]); 1353 addToOrderList(op_class); 1354 } 1355 } 1356} 1357 1358template <class Impl> 1359int 1360InstructionQueue<Impl>::countInsts() 1361{ 1362#if 0 1363 //ksewell:This works but definitely could use a cleaner write 1364 //with a more intuitive way of counting. Right now it's 1365 //just brute force .... 1366 // Change the #if if you want to use this method. 1367 int total_insts = 0; 1368 1369 for (ThreadID tid = 0; tid < numThreads; ++tid) { 1370 ListIt count_it = instList[tid].begin(); 1371 1372 while (count_it != instList[tid].end()) { 1373 if (!(*count_it)->isSquashed() && !(*count_it)->isSquashedInIQ()) { 1374 if (!(*count_it)->isIssued()) { 1375 ++total_insts; 1376 } else if ((*count_it)->isMemRef() && 1377 !(*count_it)->memOpDone) { 1378 // Loads that have not been marked as executed still count 1379 // towards the total instructions. 1380 ++total_insts; 1381 } 1382 } 1383 1384 ++count_it; 1385 } 1386 } 1387 1388 return total_insts; 1389#else 1390 return numEntries - freeEntries; 1391#endif 1392} 1393 1394template <class Impl> 1395void 1396InstructionQueue<Impl>::dumpLists() 1397{ 1398 for (int i = 0; i < Num_OpClasses; ++i) { 1399 cprintf("Ready list %i size: %i\n", i, readyInsts[i].size()); 1400 1401 cprintf("\n"); 1402 } 1403 1404 cprintf("Non speculative list size: %i\n", nonSpecInsts.size()); 1405 1406 NonSpecMapIt non_spec_it = nonSpecInsts.begin(); 1407 NonSpecMapIt non_spec_end_it = nonSpecInsts.end(); 1408 1409 cprintf("Non speculative list: "); 1410 1411 while (non_spec_it != non_spec_end_it) { 1412 cprintf("%s [sn:%lli]", (*non_spec_it).second->pcState(), 1413 (*non_spec_it).second->seqNum); 1414 ++non_spec_it; 1415 } 1416 1417 cprintf("\n"); 1418 1419 ListOrderIt list_order_it = listOrder.begin(); 1420 ListOrderIt list_order_end_it = listOrder.end(); 1421 int i = 1; 1422 1423 cprintf("List order: "); 1424 1425 while (list_order_it != list_order_end_it) { 1426 cprintf("%i OpClass:%i [sn:%lli] ", i, (*list_order_it).queueType, 1427 (*list_order_it).oldestInst); 1428 1429 ++list_order_it; 1430 ++i; 1431 } 1432 1433 cprintf("\n"); 1434} 1435 1436 1437template <class Impl> 1438void 1439InstructionQueue<Impl>::dumpInsts() 1440{ 1441 for (ThreadID tid = 0; tid < numThreads; ++tid) { 1442 int num = 0; 1443 int valid_num = 0; 1444 ListIt inst_list_it = instList[tid].begin(); 1445 1446 while (inst_list_it != instList[tid].end()) { 1447 cprintf("Instruction:%i\n", num); 1448 if (!(*inst_list_it)->isSquashed()) { 1449 if (!(*inst_list_it)->isIssued()) { 1450 ++valid_num; 1451 cprintf("Count:%i\n", valid_num); 1452 } else if ((*inst_list_it)->isMemRef() && 1453 !(*inst_list_it)->memOpDone()) { 1454 // Loads that have not been marked as executed 1455 // still count towards the total instructions. 1456 ++valid_num; 1457 cprintf("Count:%i\n", valid_num); 1458 } 1459 } 1460 1461 cprintf("PC: %s\n[sn:%lli]\n[tid:%i]\n" 1462 "Issued:%i\nSquashed:%i\n", 1463 (*inst_list_it)->pcState(), 1464 (*inst_list_it)->seqNum, 1465 (*inst_list_it)->threadNumber, 1466 (*inst_list_it)->isIssued(), 1467 (*inst_list_it)->isSquashed()); 1468 1469 if ((*inst_list_it)->isMemRef()) { 1470 cprintf("MemOpDone:%i\n", (*inst_list_it)->memOpDone()); 1471 } 1472 1473 cprintf("\n"); 1474 1475 inst_list_it++; 1476 ++num; 1477 } 1478 } 1479 1480 cprintf("Insts to Execute list:\n"); 1481 1482 int num = 0; 1483 int valid_num = 0; 1484 ListIt inst_list_it = instsToExecute.begin(); 1485 1486 while (inst_list_it != instsToExecute.end()) 1487 { 1488 cprintf("Instruction:%i\n", 1489 num); 1490 if (!(*inst_list_it)->isSquashed()) { 1491 if (!(*inst_list_it)->isIssued()) { 1492 ++valid_num; 1493 cprintf("Count:%i\n", valid_num); 1494 } else if ((*inst_list_it)->isMemRef() && 1495 !(*inst_list_it)->memOpDone()) { 1496 // Loads that have not been marked as executed 1497 // still count towards the total instructions. 1498 ++valid_num; 1499 cprintf("Count:%i\n", valid_num); 1500 } 1501 } 1502 1503 cprintf("PC: %s\n[sn:%lli]\n[tid:%i]\n" 1504 "Issued:%i\nSquashed:%i\n", 1505 (*inst_list_it)->pcState(), 1506 (*inst_list_it)->seqNum, 1507 (*inst_list_it)->threadNumber, 1508 (*inst_list_it)->isIssued(), 1509 (*inst_list_it)->isSquashed()); 1510 1511 if ((*inst_list_it)->isMemRef()) { 1512 cprintf("MemOpDone:%i\n", (*inst_list_it)->memOpDone()); 1513 } 1514 1515 cprintf("\n"); 1516 1517 inst_list_it++; 1518 ++num; 1519 } 1520} 1521 1522#endif//__CPU_O3_INST_QUEUE_IMPL_HH__ 1523