lib/Target/Hexagon/HexagonVectorLoopCarriedReuse.cpp

   1 //===- HexagonVectorLoopCarriedReuse.cpp ----------------------------------===//
   2 //
   3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
   4 // See https://llvm.org/LICENSE.txt for license information.
   5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
   6 //
   7 //===----------------------------------------------------------------------===//
   8 //
   9 // This pass removes the computation of provably redundant expressions that have
  10 // been computed earlier in a previous iteration. It relies on the use of PHIs
  11 // to identify loop carried dependences. This is scalar replacement for vector
  12 // types.
  13 //
  14 //-----------------------------------------------------------------------------
  15 // Motivation: Consider the case where we have the following loop structure.
  16 //
  17 // Loop:
  18 //  t0 = a[i];
  19 //  t1 = f(t0);
  20 //  t2 = g(t1);
  21 //  ...
  22 //  t3 = a[i+1];
  23 //  t4 = f(t3);
  24 //  t5 = g(t4);
  25 //  t6 = op(t2, t5)
  26 //  cond_branch <Loop>
  27 //
  28 // This can be converted to
  29 //  t00 = a[0];
  30 //  t10 = f(t00);
  31 //  t20 = g(t10);
  32 // Loop:
  33 //  t2 = t20;
  34 //  t3 = a[i+1];
  35 //  t4 = f(t3);
  36 //  t5 = g(t4);
  37 //  t6 = op(t2, t5)
  38 //  t20 = t5
  39 //  cond_branch <Loop>
  40 //
  41 // SROA does a good job of reusing a[i+1] as a[i] in the next iteration.
  42 // Such a loop comes to this pass in the following form.
  43 //
  44 // LoopPreheader:
  45 //  X0 = a[0];
  46 // Loop:
  47 //  X2 = PHI<(X0, LoopPreheader), (X1, Loop)>
  48 //  t1 = f(X2)   <-- I1
  49 //  t2 = g(t1)
  50 //  ...
  51 //  X1 = a[i+1]
  52 //  t4 = f(X1)   <-- I2
  53 //  t5 = g(t4)
  54 //  t6 = op(t2, t5)
  55 //  cond_branch <Loop>
  56 //
  57 // In this pass, we look for PHIs such as X2 whose incoming values come only
  58 // from the Loop Preheader and over the backedge and additionaly, both these
  59 // values are the results of the same operation in terms of opcode. We call such
  60 // a PHI node a dependence chain or DepChain. In this case, the dependence of X2
  61 // over X1 is carried over only one iteration and so the DepChain is only one
  62 // PHI node long.
  63 //
  64 // Then, we traverse the uses of the PHI (X2) and the uses of the value of the
  65 // PHI coming  over the backedge (X1). We stop at the first pair of such users
  66 // I1 (of X2) and I2 (of X1) that meet the following conditions.
  67 // 1. I1 and I2 are the same operation, but with different operands.
  68 // 2. X2 and X1 are used at the same operand number in the two instructions.
  69 // 3. All other operands Op1 of I1 and Op2 of I2 are also such that there is a
  70 //    a DepChain from Op1 to Op2 of the same length as that between X2 and X1.
  71 //
  72 // We then make the following transformation
  73 // LoopPreheader:
  74 //  X0 = a[0];
  75 //  Y0 = f(X0);
  76 // Loop:
  77 //  X2 = PHI<(X0, LoopPreheader), (X1, Loop)>
  78 //  Y2 = PHI<(Y0, LoopPreheader), (t4, Loop)>
  79 //  t1 = f(X2)   <-- Will be removed by DCE.
  80 //  t2 = g(Y2)
  81 //  ...
  82 //  X1 = a[i+1]
  83 //  t4 = f(X1)
  84 //  t5 = g(t4)
  85 //  t6 = op(t2, t5)
  86 //  cond_branch <Loop>
  87 //
  88 // We proceed until we cannot find any more such instructions I1 and I2.
  89 //
  90 // --- DepChains & Loop carried dependences ---
  91 // Consider a single basic block loop such as
  92 //
  93 // LoopPreheader:
  94 //  X0 = ...
  95 //  Y0 = ...
  96 // Loop:
  97 //  X2 = PHI<(X0, LoopPreheader), (X1, Loop)>
  98 //  Y2 = PHI<(Y0, LoopPreheader), (X2, Loop)>
  99 //  ...
 100 //  X1 = ...
 101 //  ...
 102 //  cond_branch <Loop>
 103 //
 104 // Then there is a dependence between X2 and X1 that goes back one iteration,
 105 // i.e. X1 is used as X2 in the very next iteration. We represent this as a
 106 // DepChain from X2 to X1 (X2->X1).
 107 // Similarly, there is a dependence between Y2 and X1 that goes back two
 108 // iterations. X1 is used as Y2 two iterations after it is computed. This is
 109 // represented by a DepChain as (Y2->X2->X1).
 110 //
 111 // A DepChain has the following properties.
 112 // 1. Num of edges in DepChain = Number of Instructions in DepChain = Number of
 113 //    iterations of carried dependence + 1.
 114 // 2. All instructions in the DepChain except the last are PHIs.
 115 //
 116 //===----------------------------------------------------------------------===//
 117
 118 #include "llvm/ADT/SetVector.h"
 119 #include "llvm/ADT/SmallVector.h"
 120 #include "llvm/ADT/Statistic.h"
 121 #include "llvm/Analysis/LoopInfo.h"
 122 #include "llvm/Analysis/LoopPass.h"
 123 #include "llvm/IR/BasicBlock.h"
 124 #include "llvm/IR/DerivedTypes.h"
 125 #include "llvm/IR/IRBuilder.h"
 126 #include "llvm/IR/Instruction.h"
 127 #include "llvm/IR/Instructions.h"
 128 #include "llvm/IR/IntrinsicInst.h"
 129 #include "llvm/IR/Intrinsics.h"
 130 #include "llvm/IR/Use.h"
 131 #include "llvm/IR/User.h"
 132 #include "llvm/IR/Value.h"
 133 #include "llvm/Pass.h"
 134 #include "llvm/Support/Casting.h"
 135 #include "llvm/Support/CommandLine.h"
 136 #include "llvm/Support/Compiler.h"
 137 #include "llvm/Support/Debug.h"
 138 #include "llvm/Support/raw_ostream.h"
 139 #include "llvm/Transforms/Scalar.h"
 140 #include "llvm/Transforms/Utils.h"
 141 #include <algorithm>
 142 #include <cassert>
 143 #include <cstddef>
 144 #include <map>
 145 #include <memory>
 146 #include <set>
 147
 148 using namespace llvm;
 149
 150 #define DEBUG_TYPE "hexagon-vlcr"
 151
 152 STATISTIC(HexagonNumVectorLoopCarriedReuse,
 153           "Number of values that were reused from a previous iteration.");
 154
 155 static cl::opt<int> HexagonVLCRIterationLim("hexagon-vlcr-iteration-lim",
 156     cl::Hidden,
 157     cl::desc("Maximum distance of loop carried dependences that are handled"),
 158     cl::init(2), cl::ZeroOrMore);
 159
 160 namespace llvm {
 161
 162 void initializeHexagonVectorLoopCarriedReusePass(PassRegistry&);
 163 Pass *createHexagonVectorLoopCarriedReusePass();
 164
 165 } // end namespace llvm
 166
 167 namespace {
 168
 169   // See info about DepChain in the comments at the top of this file.
 170   using ChainOfDependences = SmallVector<Instruction *, 4>;
 171
 172   class DepChain {
 173     ChainOfDependences Chain;
 174
 175   public:
 176     bool isIdentical(DepChain &Other) const {
 177       if (Other.size() != size())
 178         return false;
 179       ChainOfDependences &OtherChain = Other.getChain();
 180       for (int i = 0; i < size(); ++i) {
 181         if (Chain[i] != OtherChain[i])
 182           return false;
 183       }
 184       return true;
 185     }
 186
 187     ChainOfDependences &getChain() {
 188       return Chain;
 189     }
 190
 191     int size() const {
 192       return Chain.size();
 193     }
 194
 195     void clear() {
 196       Chain.clear();
 197     }
 198
 199     void push_back(Instruction *I) {
 200       Chain.push_back(I);
 201     }
 202
 203     int iterations() const {
 204       return size() - 1;
 205     }
 206
 207     Instruction *front() const {
 208       return Chain.front();
 209     }
 210
 211     Instruction *back() const {
 212       return Chain.back();
 213     }
 214
 215     Instruction *&operator[](const int index) {
 216       return Chain[index];
 217     }
 218
 219    friend raw_ostream &operator<< (raw_ostream &OS, const DepChain &D);
 220   };
 221
 222   LLVM_ATTRIBUTE_UNUSED
 223   raw_ostream &operator<<(raw_ostream &OS, const DepChain &D) {
 224     const ChainOfDependences &CD = D.Chain;
 225     int ChainSize = CD.size();
 226     OS << "**DepChain Start::**\n";
 227     for (int i = 0; i < ChainSize -1; ++i) {
 228       OS << *(CD[i]) << " -->\n";
 229     }
 230     OS << *CD[ChainSize-1] << "\n";
 231     return OS;
 232   }
 233
 234   struct ReuseValue {
 235     Instruction *Inst2Replace = nullptr;
 236
 237     // In the new PHI node that we'll construct this is the value that'll be
 238     // used over the backedge. This is teh value that gets reused from a
 239     // previous iteration.
 240     Instruction *BackedgeInst = nullptr;
 241     std::map<Instruction *, DepChain *> DepChains;
 242     int Iterations = -1;
 243
 244     ReuseValue() = default;
 245
 246     void reset() {
 247       Inst2Replace = nullptr;
 248       BackedgeInst = nullptr;
 249       DepChains.clear();
 250       Iterations = -1;
 251     }
 252     bool isDefined() { return Inst2Replace != nullptr; }
 253   };
 254
 255   LLVM_ATTRIBUTE_UNUSED
 256   raw_ostream &operator<<(raw_ostream &OS, const ReuseValue &RU) {
 257     OS << "** ReuseValue ***\n";
 258     OS << "Instruction to Replace: " << *(RU.Inst2Replace) << "\n";
 259     OS << "Backedge Instruction: " << *(RU.BackedgeInst) << "\n";
 260     return OS;
 261   }
 262
 263   class HexagonVectorLoopCarriedReuse : public LoopPass {
 264   public:
 265     static char ID;
 266
 267     explicit HexagonVectorLoopCarriedReuse() : LoopPass(ID) {
 268       PassRegistry *PR = PassRegistry::getPassRegistry();
 269       initializeHexagonVectorLoopCarriedReusePass(*PR);
 270     }
 271
 272     StringRef getPassName() const override {
 273       return "Hexagon-specific loop carried reuse for HVX vectors";
 274     }
 275
 276     void getAnalysisUsage(AnalysisUsage &AU) const override {
 277       AU.addRequired<LoopInfoWrapperPass>();
 278       AU.addRequiredID(LoopSimplifyID);
 279       AU.addRequiredID(LCSSAID);
 280       AU.addPreservedID(LCSSAID);
 281       AU.setPreservesCFG();
 282     }
 283
 284     bool runOnLoop(Loop *L, LPPassManager &LPM) override;
 285
 286   private:
 287     SetVector<DepChain *> Dependences;
 288     std::set<Instruction *> ReplacedInsts;
 289     Loop *CurLoop;
 290     ReuseValue ReuseCandidate;
 291
 292     bool doVLCR();
 293     void findLoopCarriedDeps();
 294     void findValueToReuse();
 295     void findDepChainFromPHI(Instruction *I, DepChain &D);
 296     void reuseValue();
 297     Value *findValueInBlock(Value *Op, BasicBlock *BB);
 298     DepChain *getDepChainBtwn(Instruction *I1, Instruction *I2, int Iters);
 299     bool isEquivalentOperation(Instruction *I1, Instruction *I2);
 300     bool canReplace(Instruction *I);
 301     bool isCallInstCommutative(CallInst *C);
 302   };
 303
 304 } // end anonymous namespace
 305
 306 char HexagonVectorLoopCarriedReuse::ID = 0;
 307
 308 INITIALIZE_PASS_BEGIN(HexagonVectorLoopCarriedReuse, "hexagon-vlcr",
 309     "Hexagon-specific predictive commoning for HVX vectors", false, false)
 310 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
 311 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
 312 INITIALIZE_PASS_DEPENDENCY(LCSSAWrapperPass)
 313 INITIALIZE_PASS_END(HexagonVectorLoopCarriedReuse, "hexagon-vlcr",
 314     "Hexagon-specific predictive commoning for HVX vectors", false, false)
 315
 316 bool HexagonVectorLoopCarriedReuse::runOnLoop(Loop *L, LPPassManager &LPM) {
 317   if (skipLoop(L))
 318     return false;
 319
 320   if (!L->getLoopPreheader())
 321     return false;
 322
 323   // Work only on innermost loops.
 324   if (!L->getSubLoops().empty())
 325     return false;
 326
 327   // Work only on single basic blocks loops.
 328   if (L->getNumBlocks() != 1)
 329     return false;
 330
 331   CurLoop = L;
 332
 333   return doVLCR();
 334 }
 335
 336 bool HexagonVectorLoopCarriedReuse::isCallInstCommutative(CallInst *C) {
 337   switch (C->getCalledFunction()->getIntrinsicID()) {
 338     case Intrinsic::hexagon_V6_vaddb:
 339     case Intrinsic::hexagon_V6_vaddb_128B:
 340     case Intrinsic::hexagon_V6_vaddh:
 341     case Intrinsic::hexagon_V6_vaddh_128B:
 342     case Intrinsic::hexagon_V6_vaddw:
 343     case Intrinsic::hexagon_V6_vaddw_128B:
 344     case Intrinsic::hexagon_V6_vaddubh:
 345     case Intrinsic::hexagon_V6_vaddubh_128B:
 346     case Intrinsic::hexagon_V6_vadduhw:
 347     case Intrinsic::hexagon_V6_vadduhw_128B:
 348     case Intrinsic::hexagon_V6_vaddhw:
 349     case Intrinsic::hexagon_V6_vaddhw_128B:
 350     case Intrinsic::hexagon_V6_vmaxb:
 351     case Intrinsic::hexagon_V6_vmaxb_128B:
 352     case Intrinsic::hexagon_V6_vmaxh:
 353     case Intrinsic::hexagon_V6_vmaxh_128B:
 354     case Intrinsic::hexagon_V6_vmaxw:
 355     case Intrinsic::hexagon_V6_vmaxw_128B:
 356     case Intrinsic::hexagon_V6_vmaxub:
 357     case Intrinsic::hexagon_V6_vmaxub_128B:
 358     case Intrinsic::hexagon_V6_vmaxuh:
 359     case Intrinsic::hexagon_V6_vmaxuh_128B:
 360     case Intrinsic::hexagon_V6_vminub:
 361     case Intrinsic::hexagon_V6_vminub_128B:
 362     case Intrinsic::hexagon_V6_vminuh:
 363     case Intrinsic::hexagon_V6_vminuh_128B:
 364     case Intrinsic::hexagon_V6_vminb:
 365     case Intrinsic::hexagon_V6_vminb_128B:
 366     case Intrinsic::hexagon_V6_vminh:
 367     case Intrinsic::hexagon_V6_vminh_128B:
 368     case Intrinsic::hexagon_V6_vminw:
 369     case Intrinsic::hexagon_V6_vminw_128B:
 370     case Intrinsic::hexagon_V6_vmpyub:
 371     case Intrinsic::hexagon_V6_vmpyub_128B:
 372     case Intrinsic::hexagon_V6_vmpyuh:
 373     case Intrinsic::hexagon_V6_vmpyuh_128B:
 374     case Intrinsic::hexagon_V6_vavgub:
 375     case Intrinsic::hexagon_V6_vavgub_128B:
 376     case Intrinsic::hexagon_V6_vavgh:
 377     case Intrinsic::hexagon_V6_vavgh_128B:
 378     case Intrinsic::hexagon_V6_vavguh:
 379     case Intrinsic::hexagon_V6_vavguh_128B:
 380     case Intrinsic::hexagon_V6_vavgw:
 381     case Intrinsic::hexagon_V6_vavgw_128B:
 382     case Intrinsic::hexagon_V6_vavgb:
 383     case Intrinsic::hexagon_V6_vavgb_128B:
 384     case Intrinsic::hexagon_V6_vavguw:
 385     case Intrinsic::hexagon_V6_vavguw_128B:
 386     case Intrinsic::hexagon_V6_vabsdiffh:
 387     case Intrinsic::hexagon_V6_vabsdiffh_128B:
 388     case Intrinsic::hexagon_V6_vabsdiffub:
 389     case Intrinsic::hexagon_V6_vabsdiffub_128B:
 390     case Intrinsic::hexagon_V6_vabsdiffuh:
 391     case Intrinsic::hexagon_V6_vabsdiffuh_128B:
 392     case Intrinsic::hexagon_V6_vabsdiffw:
 393     case Intrinsic::hexagon_V6_vabsdiffw_128B:
 394       return true;
 395     default:
 396       return false;
 397   }
 398 }
 399
 400 bool HexagonVectorLoopCarriedReuse::isEquivalentOperation(Instruction *I1,
 401                                                           Instruction *I2) {
 402   if (!I1->isSameOperationAs(I2))
 403     return false;
 404   // This check is in place specifically for intrinsics. isSameOperationAs will
 405   // return two for any two hexagon intrinsics because they are essentially the
 406   // same instruciton (CallInst). We need to scratch the surface to see if they
 407   // are calls to the same function.
 408   if (CallInst *C1 = dyn_cast<CallInst>(I1)) {
 409     if (CallInst *C2 = dyn_cast<CallInst>(I2)) {
 410       if (C1->getCalledFunction() != C2->getCalledFunction())
 411         return false;
 412     }
 413   }
 414
 415   // If both the Instructions are of Vector Type and any of the element
 416   // is integer constant, check their values too for equivalence.
 417   if (I1->getType()->isVectorTy() && I2->getType()->isVectorTy()) {
 418     unsigned NumOperands = I1->getNumOperands();
 419     for (unsigned i = 0; i < NumOperands; ++i) {
 420       ConstantInt *C1 = dyn_cast<ConstantInt>(I1->getOperand(i));
 421       ConstantInt *C2 = dyn_cast<ConstantInt>(I2->getOperand(i));
 422       if(!C1) continue;
 423       assert(C2);
 424       if (C1->getSExtValue() != C2->getSExtValue())
 425         return false;
 426     }
 427   }
 428
 429   return true;
 430 }
 431
 432 bool HexagonVectorLoopCarriedReuse::canReplace(Instruction *I) {
 433   const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
 434   if (!II)
 435     return true;
 436
 437   switch (II->getIntrinsicID()) {
 438   case Intrinsic::hexagon_V6_hi:
 439   case Intrinsic::hexagon_V6_lo:
 440   case Intrinsic::hexagon_V6_hi_128B:
 441   case Intrinsic::hexagon_V6_lo_128B:
 442     LLVM_DEBUG(dbgs() << "Not considering for reuse: " << *II << "\n");
 443     return false;
 444   default:
 445     return true;
 446   }
 447 }
 448 void HexagonVectorLoopCarriedReuse::findValueToReuse() {
 449   for (auto *D : Dependences) {
 450     LLVM_DEBUG(dbgs() << "Processing dependence " << *(D->front()) << "\n");
 451     if (D->iterations() > HexagonVLCRIterationLim) {
 452       LLVM_DEBUG(
 453           dbgs()
 454           << ".. Skipping because number of iterations > than the limit\n");
 455       continue;
 456     }
 457
 458     PHINode *PN = cast<PHINode>(D->front());
 459     Instruction *BEInst = D->back();
 460     int Iters = D->iterations();
 461     BasicBlock *BB = PN->getParent();
 462     LLVM_DEBUG(dbgs() << "Checking if any uses of " << *PN
 463                       << " can be reused\n");
 464
 465     SmallVector<Instruction *, 4> PNUsers;
 466     for (auto UI = PN->use_begin(), E = PN->use_end(); UI != E; ++UI) {
 467       Use &U = *UI;
 468       Instruction *User = cast<Instruction>(U.getUser());
 469
 470       if (User->getParent() != BB)
 471         continue;
 472       if (ReplacedInsts.count(User)) {
 473         LLVM_DEBUG(dbgs() << *User
 474                           << " has already been replaced. Skipping...\n");
 475         continue;
 476       }
 477       if (isa<PHINode>(User))
 478         continue;
 479       if (User->mayHaveSideEffects())
 480         continue;
 481       if (!canReplace(User))
 482         continue;
 483
 484       PNUsers.push_back(User);
 485     }
 486     LLVM_DEBUG(dbgs() << PNUsers.size() << " use(s) of the PHI in the block\n");
 487
 488     // For each interesting use I of PN, find an Instruction BEUser that
 489     // performs the same operation as I on BEInst and whose other operands,
 490     // if any, can also be rematerialized in OtherBB. We stop when we find the
 491     // first such Instruction BEUser. This is because once BEUser is
 492     // rematerialized in OtherBB, we may find more such "fixup" opportunities
 493     // in this block. So, we'll start over again.
 494     for (Instruction *I : PNUsers) {
 495       for (auto UI = BEInst->use_begin(), E = BEInst->use_end(); UI != E;
 496            ++UI) {
 497         Use &U = *UI;
 498         Instruction *BEUser = cast<Instruction>(U.getUser());
 499
 500         if (BEUser->getParent() != BB)
 501           continue;
 502         if (!isEquivalentOperation(I, BEUser))
 503           continue;
 504
 505         int NumOperands = I->getNumOperands();
 506
 507         // Take operands of each PNUser one by one and try to find DepChain
 508         // with every operand of the BEUser. If any of the operands of BEUser
 509         // has DepChain with current operand of the PNUser, break the matcher
 510         // loop. Keep doing this for Every PNUser operand. If PNUser operand
 511         // does not have DepChain with any of the BEUser operand, break the
 512         // outer matcher loop, mark the BEUser as null and reset the ReuseCandidate.
 513         // This ensures that DepChain exist for all the PNUser operand with
 514         // BEUser operand. This also ensures that DepChains are independent of
 515         // the positions in PNUser and BEUser.
 516         std::map<Instruction *, DepChain *> DepChains;
 517         CallInst *C1 = dyn_cast<CallInst>(I);
 518         if ((I && I->isCommutative()) || (C1 && isCallInstCommutative(C1))) {
 519           bool Found = false;
 520           for (int OpNo = 0; OpNo < NumOperands; ++OpNo) {
 521             Value *Op = I->getOperand(OpNo);
 522             Instruction *OpInst = dyn_cast<Instruction>(Op);
 523             Found = false;
 524             for (int T = 0; T < NumOperands; ++T) {
 525               Value *BEOp = BEUser->getOperand(T);
 526               Instruction *BEOpInst = dyn_cast<Instruction>(BEOp);
 527               if (!OpInst && !BEOpInst) {
 528                 if (Op == BEOp) {
 529                   Found = true;
 530                   break;
 531                 }
 532               }
 533
 534               if ((OpInst && !BEOpInst) || (!OpInst && BEOpInst))
 535                 continue;
 536
 537               DepChain *D = getDepChainBtwn(OpInst, BEOpInst, Iters);
 538
 539               if (D) {
 540                 Found = true;
 541                 DepChains[OpInst] = D;
 542                 break;
 543               }
 544             }
 545             if (!Found) {
 546               BEUser = nullptr;
 547               break;
 548             }
 549           }
 550         } else {
 551
 552           for (int OpNo = 0; OpNo < NumOperands; ++OpNo) {
 553             Value *Op = I->getOperand(OpNo);
 554             Value *BEOp = BEUser->getOperand(OpNo);
 555
 556             Instruction *OpInst = dyn_cast<Instruction>(Op);
 557             if (!OpInst) {
 558               if (Op == BEOp)
 559                 continue;
 560               // Do not allow reuse to occur when the operands may be different
 561               // values.
 562               BEUser = nullptr;
 563               break;
 564             }
 565
 566             Instruction *BEOpInst = dyn_cast<Instruction>(BEOp);
 567             DepChain *D = getDepChainBtwn(OpInst, BEOpInst, Iters);
 568
 569             if (D) {
 570               DepChains[OpInst] = D;
 571             } else {
 572               BEUser = nullptr;
 573               break;
 574             }
 575           }
 576         }
 577         if (BEUser) {
 578           LLVM_DEBUG(dbgs() << "Found Value for reuse.\n");
 579           ReuseCandidate.Inst2Replace = I;
 580           ReuseCandidate.BackedgeInst = BEUser;
 581           ReuseCandidate.DepChains = DepChains;
 582           ReuseCandidate.Iterations = Iters;
 583           return;
 584         }
 585         ReuseCandidate.reset();
 586       }
 587     }
 588   }
 589   ReuseCandidate.reset();
 590 }
 591
 592 Value *HexagonVectorLoopCarriedReuse::findValueInBlock(Value *Op,
 593                                                        BasicBlock *BB) {
 594   PHINode *PN = dyn_cast<PHINode>(Op);
 595   assert(PN);
 596   Value *ValueInBlock = PN->getIncomingValueForBlock(BB);
 597   return ValueInBlock;
 598 }
 599
 600 void HexagonVectorLoopCarriedReuse::reuseValue() {
 601   LLVM_DEBUG(dbgs() << ReuseCandidate);
 602   Instruction *Inst2Replace = ReuseCandidate.Inst2Replace;
 603   Instruction *BEInst = ReuseCandidate.BackedgeInst;
 604   int NumOperands = Inst2Replace->getNumOperands();
 605   std::map<Instruction *, DepChain *> &DepChains = ReuseCandidate.DepChains;
 606   int Iterations = ReuseCandidate.Iterations;
 607   BasicBlock *LoopPH = CurLoop->getLoopPreheader();
 608   assert(!DepChains.empty() && "No DepChains");
 609   LLVM_DEBUG(dbgs() << "reuseValue is making the following changes\n");
 610
 611   SmallVector<Instruction *, 4> InstsInPreheader;
 612   for (int i = 0; i < Iterations; ++i) {
 613     Instruction *InstInPreheader = Inst2Replace->clone();
 614     SmallVector<Value *, 4> Ops;
 615     for (int j = 0; j < NumOperands; ++j) {
 616       Instruction *I = dyn_cast<Instruction>(Inst2Replace->getOperand(j));
 617       if (!I)
 618         continue;
 619       // Get the DepChain corresponding to this operand.
 620       DepChain &D = *DepChains[I];
 621       // Get the PHI for the iteration number and find
 622       // the incoming value from the Loop Preheader for
 623       // that PHI.
 624       Value *ValInPreheader = findValueInBlock(D[i], LoopPH);
 625       InstInPreheader->setOperand(j, ValInPreheader);
 626     }
 627     InstsInPreheader.push_back(InstInPreheader);
 628     InstInPreheader->setName(Inst2Replace->getName() + ".hexagon.vlcr");
 629     InstInPreheader->insertBefore(LoopPH->getTerminator());
 630     LLVM_DEBUG(dbgs() << "Added " << *InstInPreheader << " to "
 631                       << LoopPH->getName() << "\n");
 632   }
 633   BasicBlock *BB = BEInst->getParent();
 634   IRBuilder<> IRB(BB);
 635   IRB.SetInsertPoint(BB->getFirstNonPHI());
 636   Value *BEVal = BEInst;
 637   PHINode *NewPhi;
 638   for (int i = Iterations-1; i >=0 ; --i) {
 639     Instruction *InstInPreheader = InstsInPreheader[i];
 640     NewPhi = IRB.CreatePHI(InstInPreheader->getType(), 2);
 641     NewPhi->addIncoming(InstInPreheader, LoopPH);
 642     NewPhi->addIncoming(BEVal, BB);
 643     LLVM_DEBUG(dbgs() << "Adding " << *NewPhi << " to " << BB->getName()
 644                       << "\n");
 645     BEVal = NewPhi;
 646   }
 647   // We are in LCSSA form. So, a value defined inside the Loop is used only
 648   // inside the loop. So, the following is safe.
 649   Inst2Replace->replaceAllUsesWith(NewPhi);
 650   ReplacedInsts.insert(Inst2Replace);
 651   ++HexagonNumVectorLoopCarriedReuse;
 652 }
 653
 654 bool HexagonVectorLoopCarriedReuse::doVLCR() {
 655   assert(CurLoop->getSubLoops().empty() &&
 656          "Can do VLCR on the innermost loop only");
 657   assert((CurLoop->getNumBlocks() == 1) &&
 658          "Can do VLCR only on single block loops");
 659
 660   bool Changed = false;
 661   bool Continue;
 662
 663   LLVM_DEBUG(dbgs() << "Working on Loop: " << *CurLoop->getHeader() << "\n");
 664   do {
 665     // Reset datastructures.
 666     Dependences.clear();
 667     Continue = false;
 668
 669     findLoopCarriedDeps();
 670     findValueToReuse();
 671     if (ReuseCandidate.isDefined()) {
 672       reuseValue();
 673       Changed = true;
 674       Continue = true;
 675     }
 676     llvm::for_each(Dependences, std::default_delete<DepChain>());
 677   } while (Continue);
 678   return Changed;
 679 }
 680
 681 void HexagonVectorLoopCarriedReuse::findDepChainFromPHI(Instruction *I,
 682                                                         DepChain &D) {
 683   PHINode *PN = dyn_cast<PHINode>(I);
 684   if (!PN) {
 685     D.push_back(I);
 686     return;
 687   } else {
 688     auto NumIncomingValues = PN->getNumIncomingValues();
 689     if (NumIncomingValues != 2) {
 690       D.clear();
 691       return;
 692     }
 693
 694     BasicBlock *BB = PN->getParent();
 695     if (BB != CurLoop->getHeader()) {
 696       D.clear();
 697       return;
 698     }
 699
 700     Value *BEVal = PN->getIncomingValueForBlock(BB);
 701     Instruction *BEInst = dyn_cast<Instruction>(BEVal);
 702     // This is a single block loop with a preheader, so at least
 703     // one value should come over the backedge.
 704     assert(BEInst && "There should be a value over the backedge");
 705
 706     Value *PreHdrVal =
 707       PN->getIncomingValueForBlock(CurLoop->getLoopPreheader());
 708     if(!PreHdrVal || !isa<Instruction>(PreHdrVal)) {
 709       D.clear();
 710       return;
 711     }
 712     D.push_back(PN);
 713     findDepChainFromPHI(BEInst, D);
 714   }
 715 }
 716
 717 DepChain *HexagonVectorLoopCarriedReuse::getDepChainBtwn(Instruction *I1,
 718                                                          Instruction *I2,
 719                                                          int Iters) {
 720   for (auto *D : Dependences) {
 721     if (D->front() == I1 && D->back() == I2 && D->iterations() == Iters)
 722       return D;
 723   }
 724   return nullptr;
 725 }
 726
 727 void HexagonVectorLoopCarriedReuse::findLoopCarriedDeps() {
 728   BasicBlock *BB = CurLoop->getHeader();
 729   for (auto I = BB->begin(), E = BB->end(); I != E && isa<PHINode>(I); ++I) {
 730     auto *PN = cast<PHINode>(I);
 731     if (!isa<VectorType>(PN->getType()))
 732       continue;
 733
 734     DepChain *D = new DepChain();
 735     findDepChainFromPHI(PN, *D);
 736     if (D->size() != 0)
 737       Dependences.insert(D);
 738     else
 739       delete D;
 740   }
 741   LLVM_DEBUG(dbgs() << "Found " << Dependences.size() << " dependences\n");
 742   LLVM_DEBUG(for (size_t i = 0; i < Dependences.size();
 743                   ++i) { dbgs() << *Dependences[i] << "\n"; });
 744 }
 745
 746 Pass *llvm::createHexagonVectorLoopCarriedReusePass() {
 747   return new HexagonVectorLoopCarriedReuse();
 748 }