1/* 2 * Copyright (c) 2014 The University of Wisconsin 3 * 4 * Copyright (c) 2006 INRIA (Institut National de Recherche en 5 * Informatique et en Automatique / French National Research Institute 6 * for Computer Science and Applied Mathematics) 7 * 8 * All rights reserved. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions are 12 * met: redistributions of source code must retain the above copyright 13 * notice, this list of conditions and the following disclaimer; 14 * redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in the 16 * documentation and/or other materials provided with the distribution; 17 * neither the name of the copyright holders nor the names of its 18 * contributors may be used to endorse or promote products derived from 19 * this software without specific prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 32 * 33 * Authors: Vignyan Reddy, Dibakar Gope and Arthur Perais, 34 * from André Seznec's code. 35 */ 36 37/* @file 38 * Implementation of a L-TAGE branch predictor 39 */ 40 41#include "cpu/pred/ltage.hh" 42 43#include "base/intmath.hh" 44#include "base/logging.hh" 45#include "base/random.hh" 46#include "base/trace.hh" 47#include "debug/Fetch.hh" 48#include "debug/LTage.hh" 49 50LTAGE::LTAGE(const LTAGEParams *params) 51 : TAGE(params), 52 logSizeLoopPred(params->logSizeLoopPred), 53 loopTableAgeBits(params->loopTableAgeBits), 54 loopTableConfidenceBits(params->loopTableConfidenceBits), 55 loopTableTagBits(params->loopTableTagBits), 56 loopTableIterBits(params->loopTableIterBits), 57 logLoopTableAssoc(params->logLoopTableAssoc), 58 confidenceThreshold((1 << loopTableConfidenceBits) - 1), 59 loopTagMask((1 << loopTableTagBits) - 1), 60 loopNumIterMask((1 << loopTableIterBits) - 1), 61 loopSetMask((1 << (logSizeLoopPred - logLoopTableAssoc)) - 1), 62 loopUseCounter(0), 63 withLoopBits(params->withLoopBits), 64 useDirectionBit(params->useDirectionBit), 65 useSpeculation(params->useSpeculation), 66 useHashing(params->useHashing) 67{ 68 // we use uint16_t type for these vales, so they cannot be more than 69 // 16 bits 70 assert(loopTableTagBits <= 16); 71 assert(loopTableIterBits <= 16); 72 73 assert(logSizeLoopPred >= logLoopTableAssoc); 74 75 ltable = new LoopEntry[ULL(1) << logSizeLoopPred]; 76} 77 78int 79LTAGE::lindex(Addr pc_in) const 80{ 81 // The loop table is implemented as a linear table 82 // If associativity is N (N being 1 << logLoopTableAssoc), 83 // the first N entries are for set 0, the next N entries are for set 1, 84 // and so on. 85 // Thus, this function calculates the set and then it gets left shifted 86 // by logLoopTableAssoc in order to return the index of the first of the 87 // N entries of the set 88 Addr mask = (ULL(1) << (logSizeLoopPred - logLoopTableAssoc)) - 1; 89 Addr pc = pc_in >> instShiftAmt; 90 if (useHashing) { 91 // copied from TAGE-SC-L 92 // (http://www.jilp.org/cbp2016/code/AndreSeznecLimited.tar.gz) 93 pc ^= (pc_in >> (instShiftAmt + logLoopTableAssoc)); 94 } 95 return ((pc & mask) << logLoopTableAssoc); 96} 97 98int 99LTAGE::finallindex(int index, int lowPcBits, int way) const 100{ 101 // copied from TAGE-SC-L 102 // (http://www.jilp.org/cbp2016/code/AndreSeznecLimited.tar.gz) 103 return (useHashing ? (index ^ ((lowPcBits >> way) << logLoopTableAssoc)) : 104 (index)) 105 + way; 106} 107 108//loop prediction: only used if high confidence 109bool 110LTAGE::getLoop(Addr pc, LTageBranchInfo* bi, bool speculative) const 111{ 112 bi->loopHit = -1; 113 bi->loopPredValid = false; 114 bi->loopIndex = lindex(pc); 115 unsigned pcShift = instShiftAmt + logSizeLoopPred - logLoopTableAssoc; 116 bi->loopTag = ((pc) >> pcShift) & loopTagMask; 117 118 if (useHashing) { 119 bi->loopTag ^= ((pc >> (pcShift + logSizeLoopPred)) & loopTagMask); 120 bi->loopLowPcBits = (pc >> pcShift) & loopSetMask; 121 } 122 123 for (int i = 0; i < (1 << logLoopTableAssoc); i++) { 124 int idx = finallindex(bi->loopIndex, bi->loopLowPcBits, i); 125 if (ltable[idx].tag == bi->loopTag) { 126 bi->loopHit = i; 127 bi->loopPredValid = 128 ltable[idx].confidence == confidenceThreshold; 129 130 uint16_t iter = speculative ? ltable[idx].currentIterSpec 131 : ltable[idx].currentIter; 132 133 if ((iter + 1) == ltable[idx].numIter) { 134 return useDirectionBit ? !(ltable[idx].dir) : false; 135 } else { 136 return useDirectionBit ? (ltable[idx].dir) : true; 137 } 138 } 139 } 140 return false; 141} 142 143void 144LTAGE::specLoopUpdate(bool taken, LTageBranchInfo* bi) 145{ 146 if (bi->loopHit>=0) { 147 int index = finallindex(bi->loopIndex, bi->loopLowPcBits, bi->loopHit); 148 if (taken != ltable[index].dir) { 149 ltable[index].currentIterSpec = 0; 150 } else { 151 ltable[index].currentIterSpec = 152 (ltable[index].currentIterSpec + 1) & loopNumIterMask; 153 } 154 } 155} 156 157void 158LTAGE::loopUpdate(Addr pc, bool taken, LTageBranchInfo* bi) 159{ 160 int idx = finallindex(bi->loopIndex, bi->loopLowPcBits, bi->loopHit); 161 if (bi->loopHit >= 0) { 162 //already a hit 163 if (bi->loopPredValid) { 164 if (taken != bi->loopPred) { 165 // free the entry 166 ltable[idx].numIter = 0; 167 ltable[idx].age = 0; 168 ltable[idx].confidence = 0; 169 ltable[idx].currentIter = 0; 170 return; 171 } else if (bi->loopPred != bi->tagePred) { 172 DPRINTF(LTage, "Loop Prediction success:%lx\n",pc); 173 unsignedCtrUpdate(ltable[idx].age, true, loopTableAgeBits); 174 } 175 } 176 177 ltable[idx].currentIter = 178 (ltable[idx].currentIter + 1) & loopNumIterMask; 179 if (ltable[idx].currentIter > ltable[idx].numIter) { 180 ltable[idx].confidence = 0; 181 if (ltable[idx].numIter != 0) { 182 // free the entry 183 ltable[idx].numIter = 0; 184 ltable[idx].age = 0; 185 ltable[idx].confidence = 0; 186 } 187 } 188 189 if (taken != (useDirectionBit ? ltable[idx].dir : true)) { 190 if (ltable[idx].currentIter == ltable[idx].numIter) { 191 DPRINTF(LTage, "Loop End predicted successfully:%lx\n", pc); 192 193 unsignedCtrUpdate(ltable[idx].confidence, true, 194 loopTableConfidenceBits); 195 //just do not predict when the loop count is 1 or 2 196 if (ltable[idx].numIter < 3) { 197 // free the entry 198 ltable[idx].dir = taken; // ignored if no useDirectionBit 199 ltable[idx].numIter = 0; 200 ltable[idx].age = 0; 201 ltable[idx].confidence = 0; 202 } 203 } else { 204 DPRINTF(LTage, "Loop End predicted incorrectly:%lx\n", pc); 205 if (ltable[idx].numIter == 0) { 206 // first complete nest; 207 ltable[idx].confidence = 0; 208 ltable[idx].numIter = ltable[idx].currentIter; 209 } else { 210 //not the same number of iterations as last time: free the 211 //entry 212 ltable[idx].numIter = 0; 213 ltable[idx].age = 0; 214 ltable[idx].confidence = 0; 215 } 216 } 217 ltable[idx].currentIter = 0; 218 } 219 220 } else if (useDirectionBit ? 221 ((bi->loopPredValid ? bi->loopPred : bi->tagePred) != taken) : 222 taken) { 223 //try to allocate an entry on taken branch 224 int nrand = random_mt.random<int>(); 225 for (int i = 0; i < (1 << logLoopTableAssoc); i++) { 226 int loop_hit = (nrand + i) & ((1 << logLoopTableAssoc) - 1); 227 idx = bi->loopIndex + loop_hit; 228 if (ltable[idx].age == 0) { 229 DPRINTF(LTage, "Allocating loop pred entry for branch %lx\n", 230 pc); 231 ltable[idx].dir = !taken; // ignored if no useDirectionBit 232 ltable[idx].tag = bi->loopTag; 233 ltable[idx].numIter = 0; 234 ltable[idx].age = (1 << loopTableAgeBits) - 1; 235 ltable[idx].confidence = 0; 236 ltable[idx].currentIter = 1; 237 break; 238 239 } 240 else 241 ltable[idx].age--; 242 } 243 } 244 245} 246 247//prediction 248bool 249LTAGE::predict(ThreadID tid, Addr branch_pc, bool cond_branch, void* &b) 250{ 251 LTageBranchInfo *bi = new LTageBranchInfo(nHistoryTables+1); 252 b = (void*)(bi); 253 254 bool pred_taken = tagePredict(tid, branch_pc, cond_branch, bi); 255 256 if (cond_branch) { 257 // loop prediction 258 bi->loopPred = getLoop(branch_pc, bi, useSpeculation); 259 260 if ((loopUseCounter >= 0) && bi->loopPredValid) { 261 pred_taken = bi->loopPred; 262 bi->provider = LOOP; 263 } 264 DPRINTF(LTage, "Predict for %lx: taken?:%d, loopTaken?:%d, " 265 "loopValid?:%d, loopUseCounter:%d, tagePred:%d, altPred:%d\n", 266 branch_pc, pred_taken, bi->loopPred, bi->loopPredValid, 267 loopUseCounter, bi->tagePred, bi->altTaken); 268 269 if (useSpeculation) { 270 specLoopUpdate(pred_taken, bi); 271 } 272 } 273 274 return pred_taken; 275} 276 277void 278LTAGE::condBranchUpdate(Addr branch_pc, bool taken, 279 TageBranchInfo* tage_bi, int nrand) 280{ 281 LTageBranchInfo* bi = static_cast<LTageBranchInfo*>(tage_bi); 282 283 if (useSpeculation) { 284 // recalculate loop prediction without speculation 285 // It is ok to overwrite the loop prediction fields in bi 286 // as the stats have already been updated with the previous 287 // values 288 bi->loopPred = getLoop(branch_pc, bi, false); 289 } 290 291 if (bi->loopPredValid) { 292 if (bi->tagePred != bi->loopPred) { 293 ctrUpdate(loopUseCounter, 294 (bi->loopPred == taken), 295 withLoopBits); 296 } 297 } 298 299 loopUpdate(branch_pc, taken, bi); 300 301 TAGE::condBranchUpdate(branch_pc, taken, bi, nrand); 302} 303 304void 305LTAGE::squash(ThreadID tid, bool taken, void *bp_history) 306{ 307 TAGE::squash(tid, taken, bp_history); 308 309 LTageBranchInfo* bi = (LTageBranchInfo*)(bp_history); 310 311 if (bi->condBranch) { 312 if (bi->loopHit >= 0) { 313 int idx = finallindex(bi->loopIndex, 314 bi->loopLowPcBits, 315 bi->loopHit); 316 ltable[idx].currentIterSpec = bi->currentIter; 317 } 318 } 319} 320 321void 322LTAGE::squash(ThreadID tid, void *bp_history) 323{ 324 LTageBranchInfo* bi = (LTageBranchInfo*)(bp_history); 325 if (bi->condBranch) { 326 if (bi->loopHit >= 0) { 327 int idx = finallindex(bi->loopIndex, 328 bi->loopLowPcBits, 329 bi->loopHit); 330 ltable[idx].currentIterSpec = bi->currentIter; 331 } 332 } 333 334 TAGE::squash(tid, bp_history); 335} 336 337 338void 339LTAGE::updateStats(bool taken, TageBranchInfo* bi) 340{ 341 TAGE::updateStats(taken, bi); 342 343 LTageBranchInfo * ltage_bi = static_cast<LTageBranchInfo *>(bi); 344 345 if (ltage_bi->provider == LOOP) { 346 if (taken == ltage_bi->loopPred) { 347 loopPredictorCorrect++; 348 } else { 349 loopPredictorWrong++; 350 } 351 } 352} 353 354 355 356void 357LTAGE::regStats() 358{ 359 TAGE::regStats(); 360 361 loopPredictorCorrect 362 .name(name() + ".loopPredictorCorrect") 363 .desc("Number of times the loop predictor is the provider and " 364 "the prediction is correct"); 365 366 loopPredictorWrong 367 .name(name() + ".loopPredictorWrong")
|