lib/Transforms/Scalar/ConstantHoisting.cpp

   1 //===- ConstantHoisting.cpp - Prepare code for expensive constants --------===//
   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 identifies expensive constants to hoist and coalesces them to
  10 // better prepare it for SelectionDAG-based code generation. This works around
  11 // the limitations of the basic-block-at-a-time approach.
  12 //
  13 // First it scans all instructions for integer constants and calculates its
  14 // cost. If the constant can be folded into the instruction (the cost is
  15 // TCC_Free) or the cost is just a simple operation (TCC_BASIC), then we don't
  16 // consider it expensive and leave it alone. This is the default behavior and
  17 // the default implementation of getIntImmCost will always return TCC_Free.
  18 //
  19 // If the cost is more than TCC_BASIC, then the integer constant can't be folded
  20 // into the instruction and it might be beneficial to hoist the constant.
  21 // Similar constants are coalesced to reduce register pressure and
  22 // materialization code.
  23 //
  24 // When a constant is hoisted, it is also hidden behind a bitcast to force it to
  25 // be live-out of the basic block. Otherwise the constant would be just
  26 // duplicated and each basic block would have its own copy in the SelectionDAG.
  27 // The SelectionDAG recognizes such constants as opaque and doesn't perform
  28 // certain transformations on them, which would create a new expensive constant.
  29 //
  30 // This optimization is only applied to integer constants in instructions and
  31 // simple (this means not nested) constant cast expressions. For example:
  32 // %0 = load i64* inttoptr (i64 big_constant to i64*)
  33 //===----------------------------------------------------------------------===//
  34
  35 #include "llvm/Transforms/Scalar/ConstantHoisting.h"
  36 #include "llvm/ADT/APInt.h"
  37 #include "llvm/ADT/DenseMap.h"
  38 #include "llvm/ADT/None.h"
  39 #include "llvm/ADT/Optional.h"
  40 #include "llvm/ADT/SmallPtrSet.h"
  41 #include "llvm/ADT/SmallVector.h"
  42 #include "llvm/ADT/Statistic.h"
  43 #include "llvm/Analysis/BlockFrequencyInfo.h"
  44 #include "llvm/Analysis/TargetTransformInfo.h"
  45 #include "llvm/Transforms/Utils/Local.h"
  46 #include "llvm/IR/BasicBlock.h"
  47 #include "llvm/IR/Constants.h"
  48 #include "llvm/IR/DebugInfoMetadata.h"
  49 #include "llvm/IR/Dominators.h"
  50 #include "llvm/IR/Function.h"
  51 #include "llvm/IR/InstrTypes.h"
  52 #include "llvm/IR/Instruction.h"
  53 #include "llvm/IR/Instructions.h"
  54 #include "llvm/IR/IntrinsicInst.h"
  55 #include "llvm/IR/Value.h"
  56 #include "llvm/Pass.h"
  57 #include "llvm/Support/BlockFrequency.h"
  58 #include "llvm/Support/Casting.h"
  59 #include "llvm/Support/CommandLine.h"
  60 #include "llvm/Support/Debug.h"
  61 #include "llvm/Support/raw_ostream.h"
  62 #include "llvm/Transforms/Scalar.h"
  63 #include <algorithm>
  64 #include <cassert>
  65 #include <cstdint>
  66 #include <iterator>
  67 #include <tuple>
  68 #include <utility>
  69
  70 using namespace llvm;
  71 using namespace consthoist;
  72
  73 #define DEBUG_TYPE "consthoist"
  74
  75 STATISTIC(NumConstantsHoisted, "Number of constants hoisted");
  76 STATISTIC(NumConstantsRebased, "Number of constants rebased");
  77
  78 static cl::opt<bool> ConstHoistWithBlockFrequency(
  79     "consthoist-with-block-frequency", cl::init(true), cl::Hidden,
  80     cl::desc("Enable the use of the block frequency analysis to reduce the "
  81              "chance to execute const materialization more frequently than "
  82              "without hoisting."));
  83
  84 static cl::opt<bool> ConstHoistGEP(
  85     "consthoist-gep", cl::init(false), cl::Hidden,
  86     cl::desc("Try hoisting constant gep expressions"));
  87
  88 static cl::opt<unsigned>
  89 MinNumOfDependentToRebase("consthoist-min-num-to-rebase",
  90     cl::desc("Do not rebase if number of dependent constants of a Base is less "
  91              "than this number."),
  92     cl::init(0), cl::Hidden);
  93
  94 namespace {
  95
  96 /// The constant hoisting pass.
  97 class ConstantHoistingLegacyPass : public FunctionPass {
  98 public:
  99   static char ID; // Pass identification, replacement for typeid
 100
 101   ConstantHoistingLegacyPass() : FunctionPass(ID) {
 102     initializeConstantHoistingLegacyPassPass(*PassRegistry::getPassRegistry());
 103   }
 104
 105   bool runOnFunction(Function &Fn) override;
 106
 107   StringRef getPassName() const override { return "Constant Hoisting"; }
 108
 109   void getAnalysisUsage(AnalysisUsage &AU) const override {
 110     AU.setPreservesCFG();
 111     if (ConstHoistWithBlockFrequency)
 112       AU.addRequired<BlockFrequencyInfoWrapperPass>();
 113     AU.addRequired<DominatorTreeWrapperPass>();
 114     AU.addRequired<TargetTransformInfoWrapperPass>();
 115   }
 116
 117   void releaseMemory() override { Impl.releaseMemory(); }
 118
 119 private:
 120   ConstantHoistingPass Impl;
 121 };
 122
 123 } // end anonymous namespace
 124
 125 char ConstantHoistingLegacyPass::ID = 0;
 126
 127 INITIALIZE_PASS_BEGIN(ConstantHoistingLegacyPass, "consthoist",
 128                       "Constant Hoisting", false, false)
 129 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
 130 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
 131 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
 132 INITIALIZE_PASS_END(ConstantHoistingLegacyPass, "consthoist",
 133                     "Constant Hoisting", false, false)
 134
 135 FunctionPass *llvm::createConstantHoistingPass() {
 136   return new ConstantHoistingLegacyPass();
 137 }
 138
 139 /// Perform the constant hoisting optimization for the given function.
 140 bool ConstantHoistingLegacyPass::runOnFunction(Function &Fn) {
 141   if (skipFunction(Fn))
 142     return false;
 143
 144   LLVM_DEBUG(dbgs() << "********** Begin Constant Hoisting **********\n");
 145   LLVM_DEBUG(dbgs() << "********** Function: " << Fn.getName() << '\n');
 146
 147   bool MadeChange =
 148       Impl.runImpl(Fn, getAnalysis<TargetTransformInfoWrapperPass>().getTTI(Fn),
 149                    getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
 150                    ConstHoistWithBlockFrequency
 151                        ? &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI()
 152                        : nullptr,
 153                    Fn.getEntryBlock());
 154
 155   if (MadeChange) {
 156     LLVM_DEBUG(dbgs() << "********** Function after Constant Hoisting: "
 157                       << Fn.getName() << '\n');
 158     LLVM_DEBUG(dbgs() << Fn);
 159   }
 160   LLVM_DEBUG(dbgs() << "********** End Constant Hoisting **********\n");
 161
 162   return MadeChange;
 163 }
 164
 165 /// Find the constant materialization insertion point.
 166 Instruction *ConstantHoistingPass::findMatInsertPt(Instruction *Inst,
 167                                                    unsigned Idx) const {
 168   // If the operand is a cast instruction, then we have to materialize the
 169   // constant before the cast instruction.
 170   if (Idx != ~0U) {
 171     Value *Opnd = Inst->getOperand(Idx);
 172     if (auto CastInst = dyn_cast<Instruction>(Opnd))
 173       if (CastInst->isCast())
 174         return CastInst;
 175   }
 176
 177   // The simple and common case. This also includes constant expressions.
 178   if (!isa<PHINode>(Inst) && !Inst->isEHPad())
 179     return Inst;
 180
 181   // We can't insert directly before a phi node or an eh pad. Insert before
 182   // the terminator of the incoming or dominating block.
 183   assert(Entry != Inst->getParent() && "PHI or landing pad in entry block!");
 184   if (Idx != ~0U && isa<PHINode>(Inst))
 185     return cast<PHINode>(Inst)->getIncomingBlock(Idx)->getTerminator();
 186
 187   // This must be an EH pad. Iterate over immediate dominators until we find a
 188   // non-EH pad. We need to skip over catchswitch blocks, which are both EH pads
 189   // and terminators.
 190   auto IDom = DT->getNode(Inst->getParent())->getIDom();
 191   while (IDom->getBlock()->isEHPad()) {
 192     assert(Entry != IDom->getBlock() && "eh pad in entry block");
 193     IDom = IDom->getIDom();
 194   }
 195
 196   return IDom->getBlock()->getTerminator();
 197 }
 198
 199 /// Given \p BBs as input, find another set of BBs which collectively
 200 /// dominates \p BBs and have the minimal sum of frequencies. Return the BB
 201 /// set found in \p BBs.
 202 static void findBestInsertionSet(DominatorTree &DT, BlockFrequencyInfo &BFI,
 203                                  BasicBlock *Entry,
 204                                  SmallPtrSet<BasicBlock *, 8> &BBs) {
 205   assert(!BBs.count(Entry) && "Assume Entry is not in BBs");
 206   // Nodes on the current path to the root.
 207   SmallPtrSet<BasicBlock *, 8> Path;
 208   // Candidates includes any block 'BB' in set 'BBs' that is not strictly
 209   // dominated by any other blocks in set 'BBs', and all nodes in the path
 210   // in the dominator tree from Entry to 'BB'.
 211   SmallPtrSet<BasicBlock *, 16> Candidates;
 212   for (auto BB : BBs) {
 213     Path.clear();
 214     // Walk up the dominator tree until Entry or another BB in BBs
 215     // is reached. Insert the nodes on the way to the Path.
 216     BasicBlock *Node = BB;
 217     // The "Path" is a candidate path to be added into Candidates set.
 218     bool isCandidate = false;
 219     do {
 220       Path.insert(Node);
 221       if (Node == Entry || Candidates.count(Node)) {
 222         isCandidate = true;
 223         break;
 224       }
 225       assert(DT.getNode(Node)->getIDom() &&
 226              "Entry doens't dominate current Node");
 227       Node = DT.getNode(Node)->getIDom()->getBlock();
 228     } while (!BBs.count(Node));
 229
 230     // If isCandidate is false, Node is another Block in BBs dominating
 231     // current 'BB'. Drop the nodes on the Path.
 232     if (!isCandidate)
 233       continue;
 234
 235     // Add nodes on the Path into Candidates.
 236     Candidates.insert(Path.begin(), Path.end());
 237   }
 238
 239   // Sort the nodes in Candidates in top-down order and save the nodes
 240   // in Orders.
 241   unsigned Idx = 0;
 242   SmallVector<BasicBlock *, 16> Orders;
 243   Orders.push_back(Entry);
 244   while (Idx != Orders.size()) {
 245     BasicBlock *Node = Orders[Idx++];
 246     for (auto ChildDomNode : DT.getNode(Node)->getChildren()) {
 247       if (Candidates.count(ChildDomNode->getBlock()))
 248         Orders.push_back(ChildDomNode->getBlock());
 249     }
 250   }
 251
 252   // Visit Orders in bottom-up order.
 253   using InsertPtsCostPair =
 254       std::pair<SmallPtrSet<BasicBlock *, 16>, BlockFrequency>;
 255
 256   // InsertPtsMap is a map from a BB to the best insertion points for the
 257   // subtree of BB (subtree not including the BB itself).
 258   DenseMap<BasicBlock *, InsertPtsCostPair> InsertPtsMap;
 259   InsertPtsMap.reserve(Orders.size() + 1);
 260   for (auto RIt = Orders.rbegin(); RIt != Orders.rend(); RIt++) {
 261     BasicBlock *Node = *RIt;
 262     bool NodeInBBs = BBs.count(Node);
 263     SmallPtrSet<BasicBlock *, 16> &InsertPts = InsertPtsMap[Node].first;
 264     BlockFrequency &InsertPtsFreq = InsertPtsMap[Node].second;
 265
 266     // Return the optimal insert points in BBs.
 267     if (Node == Entry) {
 268       BBs.clear();
 269       if (InsertPtsFreq > BFI.getBlockFreq(Node) ||
 270           (InsertPtsFreq == BFI.getBlockFreq(Node) && InsertPts.size() > 1))
 271         BBs.insert(Entry);
 272       else
 273         BBs.insert(InsertPts.begin(), InsertPts.end());
 274       break;
 275     }
 276
 277     BasicBlock *Parent = DT.getNode(Node)->getIDom()->getBlock();
 278     // Initially, ParentInsertPts is empty and ParentPtsFreq is 0. Every child
 279     // will update its parent's ParentInsertPts and ParentPtsFreq.
 280     SmallPtrSet<BasicBlock *, 16> &ParentInsertPts = InsertPtsMap[Parent].first;
 281     BlockFrequency &ParentPtsFreq = InsertPtsMap[Parent].second;
 282     // Choose to insert in Node or in subtree of Node.
 283     // Don't hoist to EHPad because we may not find a proper place to insert
 284     // in EHPad.
 285     // If the total frequency of InsertPts is the same as the frequency of the
 286     // target Node, and InsertPts contains more than one nodes, choose hoisting
 287     // to reduce code size.
 288     if (NodeInBBs ||
 289         (!Node->isEHPad() &&
 290          (InsertPtsFreq > BFI.getBlockFreq(Node) ||
 291           (InsertPtsFreq == BFI.getBlockFreq(Node) && InsertPts.size() > 1)))) {
 292       ParentInsertPts.insert(Node);
 293       ParentPtsFreq += BFI.getBlockFreq(Node);
 294     } else {
 295       ParentInsertPts.insert(InsertPts.begin(), InsertPts.end());
 296       ParentPtsFreq += InsertPtsFreq;
 297     }
 298   }
 299 }
 300
 301 /// Find an insertion point that dominates all uses.
 302 SmallPtrSet<Instruction *, 8> ConstantHoistingPass::findConstantInsertionPoint(
 303     const ConstantInfo &ConstInfo) const {
 304   assert(!ConstInfo.RebasedConstants.empty() && "Invalid constant info entry.");
 305   // Collect all basic blocks.
 306   SmallPtrSet<BasicBlock *, 8> BBs;
 307   SmallPtrSet<Instruction *, 8> InsertPts;
 308   for (auto const &RCI : ConstInfo.RebasedConstants)
 309     for (auto const &U : RCI.Uses)
 310       BBs.insert(findMatInsertPt(U.Inst, U.OpndIdx)->getParent());
 311
 312   if (BBs.count(Entry)) {
 313     InsertPts.insert(&Entry->front());
 314     return InsertPts;
 315   }
 316
 317   if (BFI) {
 318     findBestInsertionSet(*DT, *BFI, Entry, BBs);
 319     for (auto BB : BBs) {
 320       BasicBlock::iterator InsertPt = BB->begin();
 321       for (; isa<PHINode>(InsertPt) || InsertPt->isEHPad(); ++InsertPt)
 322         ;
 323       InsertPts.insert(&*InsertPt);
 324     }
 325     return InsertPts;
 326   }
 327
 328   while (BBs.size() >= 2) {
 329     BasicBlock *BB, *BB1, *BB2;
 330     BB1 = *BBs.begin();
 331     BB2 = *std::next(BBs.begin());
 332     BB = DT->findNearestCommonDominator(BB1, BB2);
 333     if (BB == Entry) {
 334       InsertPts.insert(&Entry->front());
 335       return InsertPts;
 336     }
 337     BBs.erase(BB1);
 338     BBs.erase(BB2);
 339     BBs.insert(BB);
 340   }
 341   assert((BBs.size() == 1) && "Expected only one element.");
 342   Instruction &FirstInst = (*BBs.begin())->front();
 343   InsertPts.insert(findMatInsertPt(&FirstInst));
 344   return InsertPts;
 345 }
 346
 347 /// Record constant integer ConstInt for instruction Inst at operand
 348 /// index Idx.
 349 ///
 350 /// The operand at index Idx is not necessarily the constant integer itself. It
 351 /// could also be a cast instruction or a constant expression that uses the
 352 /// constant integer.
 353 void ConstantHoistingPass::collectConstantCandidates(
 354     ConstCandMapType &ConstCandMap, Instruction *Inst, unsigned Idx,
 355     ConstantInt *ConstInt) {
 356   unsigned Cost;
 357   // Ask the target about the cost of materializing the constant for the given
 358   // instruction and operand index.
 359   if (auto IntrInst = dyn_cast<IntrinsicInst>(Inst))
 360     Cost = TTI->getIntImmCost(IntrInst->getIntrinsicID(), Idx,
 361                               ConstInt->getValue(), ConstInt->getType());
 362   else
 363     Cost = TTI->getIntImmCost(Inst->getOpcode(), Idx, ConstInt->getValue(),
 364                               ConstInt->getType());
 365
 366   // Ignore cheap integer constants.
 367   if (Cost > TargetTransformInfo::TCC_Basic) {
 368     ConstCandMapType::iterator Itr;
 369     bool Inserted;
 370     ConstPtrUnionType Cand = ConstInt;
 371     std::tie(Itr, Inserted) = ConstCandMap.insert(std::make_pair(Cand, 0));
 372     if (Inserted) {
 373       ConstIntCandVec.push_back(ConstantCandidate(ConstInt));
 374       Itr->second = ConstIntCandVec.size() - 1;
 375     }
 376     ConstIntCandVec[Itr->second].addUser(Inst, Idx, Cost);
 377     LLVM_DEBUG(if (isa<ConstantInt>(Inst->getOperand(Idx))) dbgs()
 378                    << "Collect constant " << *ConstInt << " from " << *Inst
 379                    << " with cost " << Cost << '\n';
 380                else dbgs() << "Collect constant " << *ConstInt
 381                            << " indirectly from " << *Inst << " via "
 382                            << *Inst->getOperand(Idx) << " with cost " << Cost
 383                            << '\n';);
 384   }
 385 }
 386
 387 /// Record constant GEP expression for instruction Inst at operand index Idx.
 388 void ConstantHoistingPass::collectConstantCandidates(
 389     ConstCandMapType &ConstCandMap, Instruction *Inst, unsigned Idx,
 390     ConstantExpr *ConstExpr) {
 391   // TODO: Handle vector GEPs
 392   if (ConstExpr->getType()->isVectorTy())
 393     return;
 394
 395   GlobalVariable *BaseGV = dyn_cast<GlobalVariable>(ConstExpr->getOperand(0));
 396   if (!BaseGV)
 397     return;
 398
 399   // Get offset from the base GV.
 400   PointerType *GVPtrTy = dyn_cast<PointerType>(BaseGV->getType());
 401   IntegerType *PtrIntTy = DL->getIntPtrType(*Ctx, GVPtrTy->getAddressSpace());
 402   APInt Offset(DL->getTypeSizeInBits(PtrIntTy), /*val*/0, /*isSigned*/true);
 403   auto *GEPO = cast<GEPOperator>(ConstExpr);
 404   if (!GEPO->accumulateConstantOffset(*DL, Offset))
 405     return;
 406
 407   if (!Offset.isIntN(32))
 408     return;
 409
 410   // A constant GEP expression that has a GlobalVariable as base pointer is
 411   // usually lowered to a load from constant pool. Such operation is unlikely
 412   // to be cheaper than compute it by <Base + Offset>, which can be lowered to
 413   // an ADD instruction or folded into Load/Store instruction.
 414   int Cost = TTI->getIntImmCost(Instruction::Add, 1, Offset, PtrIntTy);
 415   ConstCandVecType &ExprCandVec = ConstGEPCandMap[BaseGV];
 416   ConstCandMapType::iterator Itr;
 417   bool Inserted;
 418   ConstPtrUnionType Cand = ConstExpr;
 419   std::tie(Itr, Inserted) = ConstCandMap.insert(std::make_pair(Cand, 0));
 420   if (Inserted) {
 421     ExprCandVec.push_back(ConstantCandidate(
 422         ConstantInt::get(Type::getInt32Ty(*Ctx), Offset.getLimitedValue()),
 423         ConstExpr));
 424     Itr->second = ExprCandVec.size() - 1;
 425   }
 426   ExprCandVec[Itr->second].addUser(Inst, Idx, Cost);
 427 }
 428
 429 /// Check the operand for instruction Inst at index Idx.
 430 void ConstantHoistingPass::collectConstantCandidates(
 431     ConstCandMapType &ConstCandMap, Instruction *Inst, unsigned Idx) {
 432   Value *Opnd = Inst->getOperand(Idx);
 433
 434   // Visit constant integers.
 435   if (auto ConstInt = dyn_cast<ConstantInt>(Opnd)) {
 436     collectConstantCandidates(ConstCandMap, Inst, Idx, ConstInt);
 437     return;
 438   }
 439
 440   // Visit cast instructions that have constant integers.
 441   if (auto CastInst = dyn_cast<Instruction>(Opnd)) {
 442     // Only visit cast instructions, which have been skipped. All other
 443     // instructions should have already been visited.
 444     if (!CastInst->isCast())
 445       return;
 446
 447     if (auto *ConstInt = dyn_cast<ConstantInt>(CastInst->getOperand(0))) {
 448       // Pretend the constant is directly used by the instruction and ignore
 449       // the cast instruction.
 450       collectConstantCandidates(ConstCandMap, Inst, Idx, ConstInt);
 451       return;
 452     }
 453   }
 454
 455   // Visit constant expressions that have constant integers.
 456   if (auto ConstExpr = dyn_cast<ConstantExpr>(Opnd)) {
 457     // Handle constant gep expressions.
 458     if (ConstHoistGEP && ConstExpr->isGEPWithNoNotionalOverIndexing())
 459       collectConstantCandidates(ConstCandMap, Inst, Idx, ConstExpr);
 460
 461     // Only visit constant cast expressions.
 462     if (!ConstExpr->isCast())
 463       return;
 464
 465     if (auto ConstInt = dyn_cast<ConstantInt>(ConstExpr->getOperand(0))) {
 466       // Pretend the constant is directly used by the instruction and ignore
 467       // the constant expression.
 468       collectConstantCandidates(ConstCandMap, Inst, Idx, ConstInt);
 469       return;
 470     }
 471   }
 472 }
 473
 474 /// Scan the instruction for expensive integer constants and record them
 475 /// in the constant candidate vector.
 476 void ConstantHoistingPass::collectConstantCandidates(
 477     ConstCandMapType &ConstCandMap, Instruction *Inst) {
 478   // Skip all cast instructions. They are visited indirectly later on.
 479   if (Inst->isCast())
 480     return;
 481
 482   // Scan all operands.
 483   for (unsigned Idx = 0, E = Inst->getNumOperands(); Idx != E; ++Idx) {
 484     // The cost of materializing the constants (defined in
 485     // `TargetTransformInfo::getIntImmCost`) for instructions which only take
 486     // constant variables is lower than `TargetTransformInfo::TCC_Basic`. So
 487     // it's safe for us to collect constant candidates from all IntrinsicInsts.
 488     if (canReplaceOperandWithVariable(Inst, Idx) || isa<IntrinsicInst>(Inst)) {
 489       collectConstantCandidates(ConstCandMap, Inst, Idx);
 490     }
 491   } // end of for all operands
 492 }
 493
 494 /// Collect all integer constants in the function that cannot be folded
 495 /// into an instruction itself.
 496 void ConstantHoistingPass::collectConstantCandidates(Function &Fn) {
 497   ConstCandMapType ConstCandMap;
 498   for (BasicBlock &BB : Fn)
 499     for (Instruction &Inst : BB)
 500       collectConstantCandidates(ConstCandMap, &Inst);
 501 }
 502
 503 // This helper function is necessary to deal with values that have different
 504 // bit widths (APInt Operator- does not like that). If the value cannot be
 505 // represented in uint64 we return an "empty" APInt. This is then interpreted
 506 // as the value is not in range.
 507 static Optional<APInt> calculateOffsetDiff(const APInt &V1, const APInt &V2) {
 508   Optional<APInt> Res = None;
 509   unsigned BW = V1.getBitWidth() > V2.getBitWidth() ?
 510                 V1.getBitWidth() : V2.getBitWidth();
 511   uint64_t LimVal1 = V1.getLimitedValue();
 512   uint64_t LimVal2 = V2.getLimitedValue();
 513
 514   if (LimVal1 == ~0ULL || LimVal2 == ~0ULL)
 515     return Res;
 516
 517   uint64_t Diff = LimVal1 - LimVal2;
 518   return APInt(BW, Diff, true);
 519 }
 520
 521 // From a list of constants, one needs to picked as the base and the other
 522 // constants will be transformed into an offset from that base constant. The
 523 // question is which we can pick best? For example, consider these constants
 524 // and their number of uses:
 525 //
 526 //  Constants| 2 | 4 | 12 | 42 |
 527 //  NumUses  | 3 | 2 |  8 |  7 |
 528 //
 529 // Selecting constant 12 because it has the most uses will generate negative
 530 // offsets for constants 2 and 4 (i.e. -10 and -8 respectively). If negative
 531 // offsets lead to less optimal code generation, then there might be better
 532 // solutions. Suppose immediates in the range of 0..35 are most optimally
 533 // supported by the architecture, then selecting constant 2 is most optimal
 534 // because this will generate offsets: 0, 2, 10, 40. Offsets 0, 2 and 10 are in
 535 // range 0..35, and thus 3 + 2 + 8 = 13 uses are in range. Selecting 12 would
 536 // have only 8 uses in range, so choosing 2 as a base is more optimal. Thus, in
 537 // selecting the base constant the range of the offsets is a very important
 538 // factor too that we take into account here. This algorithm calculates a total
 539 // costs for selecting a constant as the base and substract the costs if
 540 // immediates are out of range. It has quadratic complexity, so we call this
 541 // function only when we're optimising for size and there are less than 100
 542 // constants, we fall back to the straightforward algorithm otherwise
 543 // which does not do all the offset calculations.
 544 unsigned
 545 ConstantHoistingPass::maximizeConstantsInRange(ConstCandVecType::iterator S,
 546                                            ConstCandVecType::iterator E,
 547                                            ConstCandVecType::iterator &MaxCostItr) {
 548   unsigned NumUses = 0;
 549
 550   if(!Entry->getParent()->optForSize() || std::distance(S,E) > 100) {
 551     for (auto ConstCand = S; ConstCand != E; ++ConstCand) {
 552       NumUses += ConstCand->Uses.size();
 553       if (ConstCand->CumulativeCost > MaxCostItr->CumulativeCost)
 554         MaxCostItr = ConstCand;
 555     }
 556     return NumUses;
 557   }
 558
 559   LLVM_DEBUG(dbgs() << "== Maximize constants in range ==\n");
 560   int MaxCost = -1;
 561   for (auto ConstCand = S; ConstCand != E; ++ConstCand) {
 562     auto Value = ConstCand->ConstInt->getValue();
 563     Type *Ty = ConstCand->ConstInt->getType();
 564     int Cost = 0;
 565     NumUses += ConstCand->Uses.size();
 566     LLVM_DEBUG(dbgs() << "= Constant: " << ConstCand->ConstInt->getValue()
 567                       << "\n");
 568
 569     for (auto User : ConstCand->Uses) {
 570       unsigned Opcode = User.Inst->getOpcode();
 571       unsigned OpndIdx = User.OpndIdx;
 572       Cost += TTI->getIntImmCost(Opcode, OpndIdx, Value, Ty);
 573       LLVM_DEBUG(dbgs() << "Cost: " << Cost << "\n");
 574
 575       for (auto C2 = S; C2 != E; ++C2) {
 576         Optional<APInt> Diff = calculateOffsetDiff(
 577                                    C2->ConstInt->getValue(),
 578                                    ConstCand->ConstInt->getValue());
 579         if (Diff) {
 580           const int ImmCosts =
 581             TTI->getIntImmCodeSizeCost(Opcode, OpndIdx, Diff.getValue(), Ty);
 582           Cost -= ImmCosts;
 583           LLVM_DEBUG(dbgs() << "Offset " << Diff.getValue() << " "
 584                             << "has penalty: " << ImmCosts << "\n"
 585                             << "Adjusted cost: " << Cost << "\n");
 586         }
 587       }
 588     }
 589     LLVM_DEBUG(dbgs() << "Cumulative cost: " << Cost << "\n");
 590     if (Cost > MaxCost) {
 591       MaxCost = Cost;
 592       MaxCostItr = ConstCand;
 593       LLVM_DEBUG(dbgs() << "New candidate: " << MaxCostItr->ConstInt->getValue()
 594                         << "\n");
 595     }
 596   }
 597   return NumUses;
 598 }
 599
 600 /// Find the base constant within the given range and rebase all other
 601 /// constants with respect to the base constant.
 602 void ConstantHoistingPass::findAndMakeBaseConstant(
 603     ConstCandVecType::iterator S, ConstCandVecType::iterator E,
 604     SmallVectorImpl<consthoist::ConstantInfo> &ConstInfoVec) {
 605   auto MaxCostItr = S;
 606   unsigned NumUses = maximizeConstantsInRange(S, E, MaxCostItr);
 607
 608   // Don't hoist constants that have only one use.
 609   if (NumUses <= 1)
 610     return;
 611
 612   ConstantInt *ConstInt = MaxCostItr->ConstInt;
 613   ConstantExpr *ConstExpr = MaxCostItr->ConstExpr;
 614   ConstantInfo ConstInfo;
 615   ConstInfo.BaseInt = ConstInt;
 616   ConstInfo.BaseExpr = ConstExpr;
 617   Type *Ty = ConstInt->getType();
 618
 619   // Rebase the constants with respect to the base constant.
 620   for (auto ConstCand = S; ConstCand != E; ++ConstCand) {
 621     APInt Diff = ConstCand->ConstInt->getValue() - ConstInt->getValue();
 622     Constant *Offset = Diff == 0 ? nullptr : ConstantInt::get(Ty, Diff);
 623     Type *ConstTy =
 624         ConstCand->ConstExpr ? ConstCand->ConstExpr->getType() : nullptr;
 625     ConstInfo.RebasedConstants.push_back(
 626       RebasedConstantInfo(std::move(ConstCand->Uses), Offset, ConstTy));
 627   }
 628   ConstInfoVec.push_back(std::move(ConstInfo));
 629 }
 630
 631 /// Finds and combines constant candidates that can be easily
 632 /// rematerialized with an add from a common base constant.
 633 void ConstantHoistingPass::findBaseConstants(GlobalVariable *BaseGV) {
 634   // If BaseGV is nullptr, find base among candidate constant integers;
 635   // Otherwise find base among constant GEPs that share the same BaseGV.
 636   ConstCandVecType &ConstCandVec = BaseGV ?
 637       ConstGEPCandMap[BaseGV] : ConstIntCandVec;
 638   ConstInfoVecType &ConstInfoVec = BaseGV ?
 639       ConstGEPInfoMap[BaseGV] : ConstIntInfoVec;
 640
 641   // Sort the constants by value and type. This invalidates the mapping!
 642   std::stable_sort(ConstCandVec.begin(), ConstCandVec.end(),
 643              [](const ConstantCandidate &LHS, const ConstantCandidate &RHS) {
 644     if (LHS.ConstInt->getType() != RHS.ConstInt->getType())
 645       return LHS.ConstInt->getType()->getBitWidth() <
 646              RHS.ConstInt->getType()->getBitWidth();
 647     return LHS.ConstInt->getValue().ult(RHS.ConstInt->getValue());
 648   });
 649
 650   // Simple linear scan through the sorted constant candidate vector for viable
 651   // merge candidates.
 652   auto MinValItr = ConstCandVec.begin();
 653   for (auto CC = std::next(ConstCandVec.begin()), E = ConstCandVec.end();
 654        CC != E; ++CC) {
 655     if (MinValItr->ConstInt->getType() == CC->ConstInt->getType()) {
 656       Type *MemUseValTy = nullptr;
 657       for (auto &U : CC->Uses) {
 658         auto *UI = U.Inst;
 659         if (LoadInst *LI = dyn_cast<LoadInst>(UI)) {
 660           MemUseValTy = LI->getType();
 661           break;
 662         } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
 663           // Make sure the constant is used as pointer operand of the StoreInst.
 664           if (SI->getPointerOperand() == SI->getOperand(U.OpndIdx)) {
 665             MemUseValTy = SI->getValueOperand()->getType();
 666             break;
 667           }
 668         }
 669       }
 670
 671       // Check if the constant is in range of an add with immediate.
 672       APInt Diff = CC->ConstInt->getValue() - MinValItr->ConstInt->getValue();
 673       if ((Diff.getBitWidth() <= 64) &&
 674           TTI->isLegalAddImmediate(Diff.getSExtValue()) &&
 675           // Check if Diff can be used as offset in addressing mode of the user
 676           // memory instruction.
 677           (!MemUseValTy || TTI->isLegalAddressingMode(MemUseValTy,
 678            /*BaseGV*/nullptr, /*BaseOffset*/Diff.getSExtValue(),
 679            /*HasBaseReg*/true, /*Scale*/0)))
 680         continue;
 681     }
 682     // We either have now a different constant type or the constant is not in
 683     // range of an add with immediate anymore.
 684     findAndMakeBaseConstant(MinValItr, CC, ConstInfoVec);
 685     // Start a new base constant search.
 686     MinValItr = CC;
 687   }
 688   // Finalize the last base constant search.
 689   findAndMakeBaseConstant(MinValItr, ConstCandVec.end(), ConstInfoVec);
 690 }
 691
 692 /// Updates the operand at Idx in instruction Inst with the result of
 693 ///        instruction Mat. If the instruction is a PHI node then special
 694 ///        handling for duplicate values form the same incoming basic block is
 695 ///        required.
 696 /// \return The update will always succeed, but the return value indicated if
 697 ///         Mat was used for the update or not.
 698 static bool updateOperand(Instruction *Inst, unsigned Idx, Instruction *Mat) {
 699   if (auto PHI = dyn_cast<PHINode>(Inst)) {
 700     // Check if any previous operand of the PHI node has the same incoming basic
 701     // block. This is a very odd case that happens when the incoming basic block
 702     // has a switch statement. In this case use the same value as the previous
 703     // operand(s), otherwise we will fail verification due to different values.
 704     // The values are actually the same, but the variable names are different
 705     // and the verifier doesn't like that.
 706     BasicBlock *IncomingBB = PHI->getIncomingBlock(Idx);
 707     for (unsigned i = 0; i < Idx; ++i) {
 708       if (PHI->getIncomingBlock(i) == IncomingBB) {
 709         Value *IncomingVal = PHI->getIncomingValue(i);
 710         Inst->setOperand(Idx, IncomingVal);
 711         return false;
 712       }
 713     }
 714   }
 715
 716   Inst->setOperand(Idx, Mat);
 717   return true;
 718 }
 719
 720 /// Emit materialization code for all rebased constants and update their
 721 /// users.
 722 void ConstantHoistingPass::emitBaseConstants(Instruction *Base,
 723                                              Constant *Offset,
 724                                              Type *Ty,
 725                                              const ConstantUser &ConstUser) {
 726   Instruction *Mat = Base;
 727
 728   // The same offset can be dereferenced to different types in nested struct.
 729   if (!Offset && Ty && Ty != Base->getType())
 730     Offset = ConstantInt::get(Type::getInt32Ty(*Ctx), 0);
 731
 732   if (Offset) {
 733     Instruction *InsertionPt = findMatInsertPt(ConstUser.Inst,
 734                                                ConstUser.OpndIdx);
 735     if (Ty) {
 736       // Constant being rebased is a ConstantExpr.
 737       PointerType *Int8PtrTy = Type::getInt8PtrTy(*Ctx,
 738           cast<PointerType>(Ty)->getAddressSpace());
 739       Base = new BitCastInst(Base, Int8PtrTy, "base_bitcast", InsertionPt);
 740       Mat = GetElementPtrInst::Create(Int8PtrTy->getElementType(), Base,
 741           Offset, "mat_gep", InsertionPt);
 742       Mat = new BitCastInst(Mat, Ty, "mat_bitcast", InsertionPt);
 743     } else
 744       // Constant being rebased is a ConstantInt.
 745       Mat = BinaryOperator::Create(Instruction::Add, Base, Offset,
 746                                  "const_mat", InsertionPt);
 747
 748     LLVM_DEBUG(dbgs() << "Materialize constant (" << *Base->getOperand(0)
 749                       << " + " << *Offset << ") in BB "
 750                       << Mat->getParent()->getName() << '\n'
 751                       << *Mat << '\n');
 752     Mat->setDebugLoc(ConstUser.Inst->getDebugLoc());
 753   }
 754   Value *Opnd = ConstUser.Inst->getOperand(ConstUser.OpndIdx);
 755
 756   // Visit constant integer.
 757   if (isa<ConstantInt>(Opnd)) {
 758     LLVM_DEBUG(dbgs() << "Update: " << *ConstUser.Inst << '\n');
 759     if (!updateOperand(ConstUser.Inst, ConstUser.OpndIdx, Mat) && Offset)
 760       Mat->eraseFromParent();
 761     LLVM_DEBUG(dbgs() << "To    : " << *ConstUser.Inst << '\n');
 762     return;
 763   }
 764
 765   // Visit cast instruction.
 766   if (auto CastInst = dyn_cast<Instruction>(Opnd)) {
 767     assert(CastInst->isCast() && "Expected an cast instruction!");
 768     // Check if we already have visited this cast instruction before to avoid
 769     // unnecessary cloning.
 770     Instruction *&ClonedCastInst = ClonedCastMap[CastInst];
 771     if (!ClonedCastInst) {
 772       ClonedCastInst = CastInst->clone();
 773       ClonedCastInst->setOperand(0, Mat);
 774       ClonedCastInst->insertAfter(CastInst);
 775       // Use the same debug location as the original cast instruction.
 776       ClonedCastInst->setDebugLoc(CastInst->getDebugLoc());
 777       LLVM_DEBUG(dbgs() << "Clone instruction: " << *CastInst << '\n'
 778                         << "To               : " << *ClonedCastInst << '\n');
 779     }
 780
 781     LLVM_DEBUG(dbgs() << "Update: " << *ConstUser.Inst << '\n');
 782     updateOperand(ConstUser.Inst, ConstUser.OpndIdx, ClonedCastInst);
 783     LLVM_DEBUG(dbgs() << "To    : " << *ConstUser.Inst << '\n');
 784     return;
 785   }
 786
 787   // Visit constant expression.
 788   if (auto ConstExpr = dyn_cast<ConstantExpr>(Opnd)) {
 789     if (ConstExpr->isGEPWithNoNotionalOverIndexing()) {
 790       // Operand is a ConstantGEP, replace it.
 791       updateOperand(ConstUser.Inst, ConstUser.OpndIdx, Mat);
 792       return;
 793     }
 794
 795     // Aside from constant GEPs, only constant cast expressions are collected.
 796     assert(ConstExpr->isCast() && "ConstExpr should be a cast");
 797     Instruction *ConstExprInst = ConstExpr->getAsInstruction();
 798     ConstExprInst->setOperand(0, Mat);
 799     ConstExprInst->insertBefore(findMatInsertPt(ConstUser.Inst,
 800                                                 ConstUser.OpndIdx));
 801
 802     // Use the same debug location as the instruction we are about to update.
 803     ConstExprInst->setDebugLoc(ConstUser.Inst->getDebugLoc());
 804
 805     LLVM_DEBUG(dbgs() << "Create instruction: " << *ConstExprInst << '\n'
 806                       << "From              : " << *ConstExpr << '\n');
 807     LLVM_DEBUG(dbgs() << "Update: " << *ConstUser.Inst << '\n');
 808     if (!updateOperand(ConstUser.Inst, ConstUser.OpndIdx, ConstExprInst)) {
 809       ConstExprInst->eraseFromParent();
 810       if (Offset)
 811         Mat->eraseFromParent();
 812     }
 813     LLVM_DEBUG(dbgs() << "To    : " << *ConstUser.Inst << '\n');
 814     return;
 815   }
 816 }
 817
 818 /// Hoist and hide the base constant behind a bitcast and emit
 819 /// materialization code for derived constants.
 820 bool ConstantHoistingPass::emitBaseConstants(GlobalVariable *BaseGV) {
 821   bool MadeChange = false;
 822   SmallVectorImpl<consthoist::ConstantInfo> &ConstInfoVec =
 823       BaseGV ? ConstGEPInfoMap[BaseGV] : ConstIntInfoVec;
 824   for (auto const &ConstInfo : ConstInfoVec) {
 825     SmallPtrSet<Instruction *, 8> IPSet = findConstantInsertionPoint(ConstInfo);
 826     assert(!IPSet.empty() && "IPSet is empty");
 827
 828     unsigned UsesNum = 0;
 829     unsigned ReBasesNum = 0;
 830     unsigned NotRebasedNum = 0;
 831     for (Instruction *IP : IPSet) {
 832       // First, collect constants depending on this IP of the base.
 833       unsigned Uses = 0;
 834       using RebasedUse = std::tuple<Constant *, Type *, ConstantUser>;
 835       SmallVector<RebasedUse, 4> ToBeRebased;
 836       for (auto const &RCI : ConstInfo.RebasedConstants) {
 837         for (auto const &U : RCI.Uses) {
 838           Uses++;
 839           BasicBlock *OrigMatInsertBB =
 840               findMatInsertPt(U.Inst, U.OpndIdx)->getParent();
 841           // If Base constant is to be inserted in multiple places,
 842           // generate rebase for U using the Base dominating U.
 843           if (IPSet.size() == 1 ||
 844               DT->dominates(IP->getParent(), OrigMatInsertBB))
 845             ToBeRebased.push_back(RebasedUse(RCI.Offset, RCI.Ty, U));
 846         }
 847       }
 848       UsesNum = Uses;
 849
 850       // If only few constants depend on this IP of base, skip rebasing,
 851       // assuming the base and the rebased have the same materialization cost.
 852       if (ToBeRebased.size() < MinNumOfDependentToRebase) {
 853         NotRebasedNum += ToBeRebased.size();
 854         continue;
 855       }
 856
 857       // Emit an instance of the base at this IP.
 858       Instruction *Base = nullptr;
 859       // Hoist and hide the base constant behind a bitcast.
 860       if (ConstInfo.BaseExpr) {
 861         assert(BaseGV && "A base constant expression must have an base GV");
 862         Type *Ty = ConstInfo.BaseExpr->getType();
 863         Base = new BitCastInst(ConstInfo.BaseExpr, Ty, "const", IP);
 864       } else {
 865         IntegerType *Ty = ConstInfo.BaseInt->getType();
 866         Base = new BitCastInst(ConstInfo.BaseInt, Ty, "const", IP);
 867       }
 868
 869       Base->setDebugLoc(IP->getDebugLoc());
 870
 871       LLVM_DEBUG(dbgs() << "Hoist constant (" << *ConstInfo.BaseInt
 872                         << ") to BB " << IP->getParent()->getName() << '\n'
 873                         << *Base << '\n');
 874
 875       // Emit materialization code for rebased constants depending on this IP.
 876       for (auto const &R : ToBeRebased) {
 877         Constant *Off = std::get<0>(R);
 878         Type *Ty = std::get<1>(R);
 879         ConstantUser U = std::get<2>(R);
 880         emitBaseConstants(Base, Off, Ty, U);
 881         ReBasesNum++;
 882         // Use the same debug location as the last user of the constant.
 883         Base->setDebugLoc(DILocation::getMergedLocation(
 884             Base->getDebugLoc(), U.Inst->getDebugLoc()));
 885       }
 886       assert(!Base->use_empty() && "The use list is empty!?");
 887       assert(isa<Instruction>(Base->user_back()) &&
 888              "All uses should be instructions.");
 889     }
 890     (void)UsesNum;
 891     (void)ReBasesNum;
 892     (void)NotRebasedNum;
 893     // Expect all uses are rebased after rebase is done.
 894     assert(UsesNum == (ReBasesNum + NotRebasedNum) &&
 895            "Not all uses are rebased");
 896
 897     NumConstantsHoisted++;
 898
 899     // Base constant is also included in ConstInfo.RebasedConstants, so
 900     // deduct 1 from ConstInfo.RebasedConstants.size().
 901     NumConstantsRebased += ConstInfo.RebasedConstants.size() - 1;
 902
 903     MadeChange = true;
 904   }
 905   return MadeChange;
 906 }
 907
 908 /// Check all cast instructions we made a copy of and remove them if they
 909 /// have no more users.
 910 void ConstantHoistingPass::deleteDeadCastInst() const {
 911   for (auto const &I : ClonedCastMap)
 912     if (I.first->use_empty())
 913       I.first->eraseFromParent();
 914 }
 915
 916 /// Optimize expensive integer constants in the given function.
 917 bool ConstantHoistingPass::runImpl(Function &Fn, TargetTransformInfo &TTI,
 918                                    DominatorTree &DT, BlockFrequencyInfo *BFI,
 919                                    BasicBlock &Entry) {
 920   this->TTI = &TTI;
 921   this->DT = &DT;
 922   this->BFI = BFI;
 923   this->DL = &Fn.getParent()->getDataLayout();
 924   this->Ctx = &Fn.getContext();
 925   this->Entry = &Entry;
 926   // Collect all constant candidates.
 927   collectConstantCandidates(Fn);
 928
 929   // Combine constants that can be easily materialized with an add from a common
 930   // base constant.
 931   if (!ConstIntCandVec.empty())
 932     findBaseConstants(nullptr);
 933   for (auto &MapEntry : ConstGEPCandMap)
 934     if (!MapEntry.second.empty())
 935       findBaseConstants(MapEntry.first);
 936
 937   // Finally hoist the base constant and emit materialization code for dependent
 938   // constants.
 939   bool MadeChange = false;
 940   if (!ConstIntInfoVec.empty())
 941     MadeChange = emitBaseConstants(nullptr);
 942   for (auto MapEntry : ConstGEPInfoMap)
 943     if (!MapEntry.second.empty())
 944       MadeChange |= emitBaseConstants(MapEntry.first);
 945
 946
 947   // Cleanup dead instructions.
 948   deleteDeadCastInst();
 949
 950   return MadeChange;
 951 }
 952
 953 PreservedAnalyses ConstantHoistingPass::run(Function &F,
 954                                             FunctionAnalysisManager &AM) {
 955   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
 956   auto &TTI = AM.getResult<TargetIRAnalysis>(F);
 957   auto BFI = ConstHoistWithBlockFrequency
 958                  ? &AM.getResult<BlockFrequencyAnalysis>(F)
 959                  : nullptr;
 960   if (!runImpl(F, TTI, DT, BFI, F.getEntryBlock()))
 961     return PreservedAnalyses::all();
 962
 963   PreservedAnalyses PA;
 964   PA.preserveSet<CFGAnalyses>();
 965   return PA;
 966 }