lib/Transforms/Scalar/CorrelatedValuePropagation.cpp

   1 //===- CorrelatedValuePropagation.cpp - Propagate CFG-derived info --------===//
   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 file implements the Correlated Value Propagation pass.
  10 //
  11 //===----------------------------------------------------------------------===//
  12
  13 #include "llvm/Transforms/Scalar/CorrelatedValuePropagation.h"
  14 #include "llvm/ADT/DepthFirstIterator.h"
  15 #include "llvm/ADT/Optional.h"
  16 #include "llvm/ADT/SmallVector.h"
  17 #include "llvm/ADT/Statistic.h"
  18 #include "llvm/Analysis/DomTreeUpdater.h"
  19 #include "llvm/Analysis/GlobalsModRef.h"
  20 #include "llvm/Analysis/InstructionSimplify.h"
  21 #include "llvm/Analysis/LazyValueInfo.h"
  22 #include "llvm/IR/Attributes.h"
  23 #include "llvm/IR/BasicBlock.h"
  24 #include "llvm/IR/CFG.h"
  25 #include "llvm/IR/CallSite.h"
  26 #include "llvm/IR/Constant.h"
  27 #include "llvm/IR/ConstantRange.h"
  28 #include "llvm/IR/Constants.h"
  29 #include "llvm/IR/DerivedTypes.h"
  30 #include "llvm/IR/Function.h"
  31 #include "llvm/IR/IRBuilder.h"
  32 #include "llvm/IR/InstrTypes.h"
  33 #include "llvm/IR/Instruction.h"
  34 #include "llvm/IR/Instructions.h"
  35 #include "llvm/IR/IntrinsicInst.h"
  36 #include "llvm/IR/Operator.h"
  37 #include "llvm/IR/PassManager.h"
  38 #include "llvm/IR/Type.h"
  39 #include "llvm/IR/Value.h"
  40 #include "llvm/Pass.h"
  41 #include "llvm/Support/Casting.h"
  42 #include "llvm/Support/CommandLine.h"
  43 #include "llvm/Support/Debug.h"
  44 #include "llvm/Support/raw_ostream.h"
  45 #include "llvm/Transforms/Scalar.h"
  46 #include "llvm/Transforms/Utils/Local.h"
  47 #include <cassert>
  48 #include <utility>
  49
  50 using namespace llvm;
  51
  52 #define DEBUG_TYPE "correlated-value-propagation"
  53
  54 STATISTIC(NumPhis,      "Number of phis propagated");
  55 STATISTIC(NumPhiCommon, "Number of phis deleted via common incoming value");
  56 STATISTIC(NumSelects,   "Number of selects propagated");
  57 STATISTIC(NumMemAccess, "Number of memory access targets propagated");
  58 STATISTIC(NumCmps,      "Number of comparisons propagated");
  59 STATISTIC(NumReturns,   "Number of return values propagated");
  60 STATISTIC(NumDeadCases, "Number of switch cases removed");
  61 STATISTIC(NumSDivs,     "Number of sdiv converted to udiv");
  62 STATISTIC(NumUDivs,     "Number of udivs whose width was decreased");
  63 STATISTIC(NumAShrs,     "Number of ashr converted to lshr");
  64 STATISTIC(NumSRems,     "Number of srem converted to urem");
  65 STATISTIC(NumOverflows, "Number of overflow checks removed");
  66 STATISTIC(NumSaturating,
  67     "Number of saturating arithmetics converted to normal arithmetics");
  68
  69 static cl::opt<bool> DontAddNoWrapFlags("cvp-dont-add-nowrap-flags", cl::init(false));
  70
  71 namespace {
  72
  73   class CorrelatedValuePropagation : public FunctionPass {
  74   public:
  75     static char ID;
  76
  77     CorrelatedValuePropagation(): FunctionPass(ID) {
  78      initializeCorrelatedValuePropagationPass(*PassRegistry::getPassRegistry());
  79     }
  80
  81     bool runOnFunction(Function &F) override;
  82
  83     void getAnalysisUsage(AnalysisUsage &AU) const override {
  84       AU.addRequired<DominatorTreeWrapperPass>();
  85       AU.addRequired<LazyValueInfoWrapperPass>();
  86       AU.addPreserved<GlobalsAAWrapperPass>();
  87       AU.addPreserved<DominatorTreeWrapperPass>();
  88       AU.addPreserved<LazyValueInfoWrapperPass>();
  89     }
  90   };
  91
  92 } // end anonymous namespace
  93
  94 char CorrelatedValuePropagation::ID = 0;
  95
  96 INITIALIZE_PASS_BEGIN(CorrelatedValuePropagation, "correlated-propagation",
  97                 "Value Propagation", false, false)
  98 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
  99 INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)
 100 INITIALIZE_PASS_END(CorrelatedValuePropagation, "correlated-propagation",
 101                 "Value Propagation", false, false)
 102
 103 // Public interface to the Value Propagation pass
 104 Pass *llvm::createCorrelatedValuePropagationPass() {
 105   return new CorrelatedValuePropagation();
 106 }
 107
 108 static bool processSelect(SelectInst *S, LazyValueInfo *LVI) {
 109   if (S->getType()->isVectorTy()) return false;
 110   if (isa<Constant>(S->getOperand(0))) return false;
 111
 112   Constant *C = LVI->getConstant(S->getCondition(), S->getParent(), S);
 113   if (!C) return false;
 114
 115   ConstantInt *CI = dyn_cast<ConstantInt>(C);
 116   if (!CI) return false;
 117
 118   Value *ReplaceWith = S->getTrueValue();
 119   Value *Other = S->getFalseValue();
 120   if (!CI->isOne()) std::swap(ReplaceWith, Other);
 121   if (ReplaceWith == S) ReplaceWith = UndefValue::get(S->getType());
 122
 123   S->replaceAllUsesWith(ReplaceWith);
 124   S->eraseFromParent();
 125
 126   ++NumSelects;
 127
 128   return true;
 129 }
 130
 131 /// Try to simplify a phi with constant incoming values that match the edge
 132 /// values of a non-constant value on all other edges:
 133 /// bb0:
 134 ///   %isnull = icmp eq i8* %x, null
 135 ///   br i1 %isnull, label %bb2, label %bb1
 136 /// bb1:
 137 ///   br label %bb2
 138 /// bb2:
 139 ///   %r = phi i8* [ %x, %bb1 ], [ null, %bb0 ]
 140 /// -->
 141 ///   %r = %x
 142 static bool simplifyCommonValuePhi(PHINode *P, LazyValueInfo *LVI,
 143                                    DominatorTree *DT) {
 144   // Collect incoming constants and initialize possible common value.
 145   SmallVector<std::pair<Constant *, unsigned>, 4> IncomingConstants;
 146   Value *CommonValue = nullptr;
 147   for (unsigned i = 0, e = P->getNumIncomingValues(); i != e; ++i) {
 148     Value *Incoming = P->getIncomingValue(i);
 149     if (auto *IncomingConstant = dyn_cast<Constant>(Incoming)) {
 150       IncomingConstants.push_back(std::make_pair(IncomingConstant, i));
 151     } else if (!CommonValue) {
 152       // The potential common value is initialized to the first non-constant.
 153       CommonValue = Incoming;
 154     } else if (Incoming != CommonValue) {
 155       // There can be only one non-constant common value.
 156       return false;
 157     }
 158   }
 159
 160   if (!CommonValue || IncomingConstants.empty())
 161     return false;
 162
 163   // The common value must be valid in all incoming blocks.
 164   BasicBlock *ToBB = P->getParent();
 165   if (auto *CommonInst = dyn_cast<Instruction>(CommonValue))
 166     if (!DT->dominates(CommonInst, ToBB))
 167       return false;
 168
 169   // We have a phi with exactly 1 variable incoming value and 1 or more constant
 170   // incoming values. See if all constant incoming values can be mapped back to
 171   // the same incoming variable value.
 172   for (auto &IncomingConstant : IncomingConstants) {
 173     Constant *C = IncomingConstant.first;
 174     BasicBlock *IncomingBB = P->getIncomingBlock(IncomingConstant.second);
 175     if (C != LVI->getConstantOnEdge(CommonValue, IncomingBB, ToBB, P))
 176       return false;
 177   }
 178
 179   // All constant incoming values map to the same variable along the incoming
 180   // edges of the phi. The phi is unnecessary.
 181   P->replaceAllUsesWith(CommonValue);
 182   P->eraseFromParent();
 183   ++NumPhiCommon;
 184   return true;
 185 }
 186
 187 static bool processPHI(PHINode *P, LazyValueInfo *LVI, DominatorTree *DT,
 188                        const SimplifyQuery &SQ) {
 189   bool Changed = false;
 190
 191   BasicBlock *BB = P->getParent();
 192   for (unsigned i = 0, e = P->getNumIncomingValues(); i < e; ++i) {
 193     Value *Incoming = P->getIncomingValue(i);
 194     if (isa<Constant>(Incoming)) continue;
 195
 196     Value *V = LVI->getConstantOnEdge(Incoming, P->getIncomingBlock(i), BB, P);
 197
 198     // Look if the incoming value is a select with a scalar condition for which
 199     // LVI can tells us the value. In that case replace the incoming value with
 200     // the appropriate value of the select. This often allows us to remove the
 201     // select later.
 202     if (!V) {
 203       SelectInst *SI = dyn_cast<SelectInst>(Incoming);
 204       if (!SI) continue;
 205
 206       Value *Condition = SI->getCondition();
 207       if (!Condition->getType()->isVectorTy()) {
 208         if (Constant *C = LVI->getConstantOnEdge(
 209                 Condition, P->getIncomingBlock(i), BB, P)) {
 210           if (C->isOneValue()) {
 211             V = SI->getTrueValue();
 212           } else if (C->isZeroValue()) {
 213             V = SI->getFalseValue();
 214           }
 215           // Once LVI learns to handle vector types, we could also add support
 216           // for vector type constants that are not all zeroes or all ones.
 217         }
 218       }
 219
 220       // Look if the select has a constant but LVI tells us that the incoming
 221       // value can never be that constant. In that case replace the incoming
 222       // value with the other value of the select. This often allows us to
 223       // remove the select later.
 224       if (!V) {
 225         Constant *C = dyn_cast<Constant>(SI->getFalseValue());
 226         if (!C) continue;
 227
 228         if (LVI->getPredicateOnEdge(ICmpInst::ICMP_EQ, SI, C,
 229               P->getIncomingBlock(i), BB, P) !=
 230             LazyValueInfo::False)
 231           continue;
 232         V = SI->getTrueValue();
 233       }
 234
 235       LLVM_DEBUG(dbgs() << "CVP: Threading PHI over " << *SI << '\n');
 236     }
 237
 238     P->setIncomingValue(i, V);
 239     Changed = true;
 240   }
 241
 242   if (Value *V = SimplifyInstruction(P, SQ)) {
 243     P->replaceAllUsesWith(V);
 244     P->eraseFromParent();
 245     Changed = true;
 246   }
 247
 248   if (!Changed)
 249     Changed = simplifyCommonValuePhi(P, LVI, DT);
 250
 251   if (Changed)
 252     ++NumPhis;
 253
 254   return Changed;
 255 }
 256
 257 static bool processMemAccess(Instruction *I, LazyValueInfo *LVI) {
 258   Value *Pointer = nullptr;
 259   if (LoadInst *L = dyn_cast<LoadInst>(I))
 260     Pointer = L->getPointerOperand();
 261   else
 262     Pointer = cast<StoreInst>(I)->getPointerOperand();
 263
 264   if (isa<Constant>(Pointer)) return false;
 265
 266   Constant *C = LVI->getConstant(Pointer, I->getParent(), I);
 267   if (!C) return false;
 268
 269   ++NumMemAccess;
 270   I->replaceUsesOfWith(Pointer, C);
 271   return true;
 272 }
 273
 274 /// See if LazyValueInfo's ability to exploit edge conditions or range
 275 /// information is sufficient to prove this comparison. Even for local
 276 /// conditions, this can sometimes prove conditions instcombine can't by
 277 /// exploiting range information.
 278 static bool processCmp(CmpInst *Cmp, LazyValueInfo *LVI) {
 279   Value *Op0 = Cmp->getOperand(0);
 280   auto *C = dyn_cast<Constant>(Cmp->getOperand(1));
 281   if (!C)
 282     return false;
 283
 284   // As a policy choice, we choose not to waste compile time on anything where
 285   // the comparison is testing local values.  While LVI can sometimes reason
 286   // about such cases, it's not its primary purpose.  We do make sure to do
 287   // the block local query for uses from terminator instructions, but that's
 288   // handled in the code for each terminator.
 289   auto *I = dyn_cast<Instruction>(Op0);
 290   if (I && I->getParent() == Cmp->getParent())
 291     return false;
 292
 293   LazyValueInfo::Tristate Result =
 294       LVI->getPredicateAt(Cmp->getPredicate(), Op0, C, Cmp);
 295   if (Result == LazyValueInfo::Unknown)
 296     return false;
 297
 298   ++NumCmps;
 299   Constant *TorF = ConstantInt::get(Type::getInt1Ty(Cmp->getContext()), Result);
 300   Cmp->replaceAllUsesWith(TorF);
 301   Cmp->eraseFromParent();
 302   return true;
 303 }
 304
 305 /// Simplify a switch instruction by removing cases which can never fire. If the
 306 /// uselessness of a case could be determined locally then constant propagation
 307 /// would already have figured it out. Instead, walk the predecessors and
 308 /// statically evaluate cases based on information available on that edge. Cases
 309 /// that cannot fire no matter what the incoming edge can safely be removed. If
 310 /// a case fires on every incoming edge then the entire switch can be removed
 311 /// and replaced with a branch to the case destination.
 312 static bool processSwitch(SwitchInst *I, LazyValueInfo *LVI,
 313                           DominatorTree *DT) {
 314   DomTreeUpdater DTU(*DT, DomTreeUpdater::UpdateStrategy::Lazy);
 315   Value *Cond = I->getCondition();
 316   BasicBlock *BB = I->getParent();
 317
 318   // If the condition was defined in same block as the switch then LazyValueInfo
 319   // currently won't say anything useful about it, though in theory it could.
 320   if (isa<Instruction>(Cond) && cast<Instruction>(Cond)->getParent() == BB)
 321     return false;
 322
 323   // If the switch is unreachable then trying to improve it is a waste of time.
 324   pred_iterator PB = pred_begin(BB), PE = pred_end(BB);
 325   if (PB == PE) return false;
 326
 327   // Analyse each switch case in turn.
 328   bool Changed = false;
 329   DenseMap<BasicBlock*, int> SuccessorsCount;
 330   for (auto *Succ : successors(BB))
 331     SuccessorsCount[Succ]++;
 332
 333   { // Scope for SwitchInstProfUpdateWrapper. It must not live during
 334     // ConstantFoldTerminator() as the underlying SwitchInst can be changed.
 335     SwitchInstProfUpdateWrapper SI(*I);
 336
 337     for (auto CI = SI->case_begin(), CE = SI->case_end(); CI != CE;) {
 338       ConstantInt *Case = CI->getCaseValue();
 339
 340       // Check to see if the switch condition is equal to/not equal to the case
 341       // value on every incoming edge, equal/not equal being the same each time.
 342       LazyValueInfo::Tristate State = LazyValueInfo::Unknown;
 343       for (pred_iterator PI = PB; PI != PE; ++PI) {
 344         // Is the switch condition equal to the case value?
 345         LazyValueInfo::Tristate Value = LVI->getPredicateOnEdge(CmpInst::ICMP_EQ,
 346                                                                 Cond, Case, *PI,
 347                                                                 BB, SI);
 348         // Give up on this case if nothing is known.
 349         if (Value == LazyValueInfo::Unknown) {
 350           State = LazyValueInfo::Unknown;
 351           break;
 352         }
 353
 354         // If this was the first edge to be visited, record that all other edges
 355         // need to give the same result.
 356         if (PI == PB) {
 357           State = Value;
 358           continue;
 359         }
 360
 361         // If this case is known to fire for some edges and known not to fire for
 362         // others then there is nothing we can do - give up.
 363         if (Value != State) {
 364           State = LazyValueInfo::Unknown;
 365           break;
 366         }
 367       }
 368
 369       if (State == LazyValueInfo::False) {
 370         // This case never fires - remove it.
 371         BasicBlock *Succ = CI->getCaseSuccessor();
 372         Succ->removePredecessor(BB);
 373         CI = SI.removeCase(CI);
 374         CE = SI->case_end();
 375
 376         // The condition can be modified by removePredecessor's PHI simplification
 377         // logic.
 378         Cond = SI->getCondition();
 379
 380         ++NumDeadCases;
 381         Changed = true;
 382         if (--SuccessorsCount[Succ] == 0)
 383           DTU.applyUpdatesPermissive({{DominatorTree::Delete, BB, Succ}});
 384         continue;
 385       }
 386       if (State == LazyValueInfo::True) {
 387         // This case always fires.  Arrange for the switch to be turned into an
 388         // unconditional branch by replacing the switch condition with the case
 389         // value.
 390         SI->setCondition(Case);
 391         NumDeadCases += SI->getNumCases();
 392         Changed = true;
 393         break;
 394       }
 395
 396       // Increment the case iterator since we didn't delete it.
 397       ++CI;
 398     }
 399   }
 400
 401   if (Changed)
 402     // If the switch has been simplified to the point where it can be replaced
 403     // by a branch then do so now.
 404     ConstantFoldTerminator(BB, /*DeleteDeadConditions = */ false,
 405                            /*TLI = */ nullptr, &DTU);
 406   return Changed;
 407 }
 408
 409 // See if we can prove that the given binary op intrinsic will not overflow.
 410 static bool willNotOverflow(BinaryOpIntrinsic *BO, LazyValueInfo *LVI) {
 411   ConstantRange LRange = LVI->getConstantRange(
 412       BO->getLHS(), BO->getParent(), BO);
 413   ConstantRange RRange = LVI->getConstantRange(
 414       BO->getRHS(), BO->getParent(), BO);
 415   ConstantRange NWRegion = ConstantRange::makeGuaranteedNoWrapRegion(
 416       BO->getBinaryOp(), RRange, BO->getNoWrapKind());
 417   return NWRegion.contains(LRange);
 418 }
 419
 420 // Rewrite this with.overflow intrinsic as non-overflowing.
 421 static void processOverflowIntrinsic(WithOverflowInst *WO) {
 422   IRBuilder<> B(WO);
 423   Value *NewOp = B.CreateBinOp(
 424       WO->getBinaryOp(), WO->getLHS(), WO->getRHS(), WO->getName());
 425   // Constant-folding could have happened.
 426   if (auto *Inst = dyn_cast<Instruction>(NewOp)) {
 427     if (WO->isSigned())
 428       Inst->setHasNoSignedWrap();
 429     else
 430       Inst->setHasNoUnsignedWrap();
 431   }
 432
 433   StructType *ST = cast<StructType>(WO->getType());
 434   Constant *Struct = ConstantStruct::get(ST,
 435       { UndefValue::get(ST->getElementType(0)),
 436         ConstantInt::getFalse(ST->getElementType(1)) });
 437   Value *NewI = B.CreateInsertValue(Struct, NewOp, 0);
 438   WO->replaceAllUsesWith(NewI);
 439   WO->eraseFromParent();
 440   ++NumOverflows;
 441 }
 442
 443 static void processSaturatingInst(SaturatingInst *SI) {
 444   BinaryOperator *BinOp = BinaryOperator::Create(
 445       SI->getBinaryOp(), SI->getLHS(), SI->getRHS(), SI->getName(), SI);
 446   BinOp->setDebugLoc(SI->getDebugLoc());
 447   if (SI->isSigned())
 448     BinOp->setHasNoSignedWrap();
 449   else
 450     BinOp->setHasNoUnsignedWrap();
 451
 452   SI->replaceAllUsesWith(BinOp);
 453   SI->eraseFromParent();
 454   ++NumSaturating;
 455 }
 456
 457 /// Infer nonnull attributes for the arguments at the specified callsite.
 458 static bool processCallSite(CallSite CS, LazyValueInfo *LVI) {
 459   SmallVector<unsigned, 4> ArgNos;
 460   unsigned ArgNo = 0;
 461
 462   if (auto *WO = dyn_cast<WithOverflowInst>(CS.getInstruction())) {
 463     if (WO->getLHS()->getType()->isIntegerTy() && willNotOverflow(WO, LVI)) {
 464       processOverflowIntrinsic(WO);
 465       return true;
 466     }
 467   }
 468
 469   if (auto *SI = dyn_cast<SaturatingInst>(CS.getInstruction())) {
 470     if (SI->getType()->isIntegerTy() && willNotOverflow(SI, LVI)) {
 471       processSaturatingInst(SI);
 472       return true;
 473     }
 474   }
 475
 476   // Deopt bundle operands are intended to capture state with minimal
 477   // perturbance of the code otherwise.  If we can find a constant value for
 478   // any such operand and remove a use of the original value, that's
 479   // desireable since it may allow further optimization of that value (e.g. via
 480   // single use rules in instcombine).  Since deopt uses tend to,
 481   // idiomatically, appear along rare conditional paths, it's reasonable likely
 482   // we may have a conditional fact with which LVI can fold.
 483   if (auto DeoptBundle = CS.getOperandBundle(LLVMContext::OB_deopt)) {
 484     bool Progress = false;
 485     for (const Use &ConstU : DeoptBundle->Inputs) {
 486       Use &U = const_cast<Use&>(ConstU);
 487       Value *V = U.get();
 488       if (V->getType()->isVectorTy()) continue;
 489       if (isa<Constant>(V)) continue;
 490
 491       Constant *C = LVI->getConstant(V, CS.getParent(), CS.getInstruction());
 492       if (!C) continue;
 493       U.set(C);
 494       Progress = true;
 495     }
 496     if (Progress)
 497       return true;
 498   }
 499
 500   for (Value *V : CS.args()) {
 501     PointerType *Type = dyn_cast<PointerType>(V->getType());
 502     // Try to mark pointer typed parameters as non-null.  We skip the
 503     // relatively expensive analysis for constants which are obviously either
 504     // null or non-null to start with.
 505     if (Type && !CS.paramHasAttr(ArgNo, Attribute::NonNull) &&
 506         !isa<Constant>(V) &&
 507         LVI->getPredicateAt(ICmpInst::ICMP_EQ, V,
 508                             ConstantPointerNull::get(Type),
 509                             CS.getInstruction()) == LazyValueInfo::False)
 510       ArgNos.push_back(ArgNo);
 511     ArgNo++;
 512   }
 513
 514   assert(ArgNo == CS.arg_size() && "sanity check");
 515
 516   if (ArgNos.empty())
 517     return false;
 518
 519   AttributeList AS = CS.getAttributes();
 520   LLVMContext &Ctx = CS.getInstruction()->getContext();
 521   AS = AS.addParamAttribute(Ctx, ArgNos,
 522                             Attribute::get(Ctx, Attribute::NonNull));
 523   CS.setAttributes(AS);
 524
 525   return true;
 526 }
 527
 528 static bool hasPositiveOperands(BinaryOperator *SDI, LazyValueInfo *LVI) {
 529   Constant *Zero = ConstantInt::get(SDI->getType(), 0);
 530   for (Value *O : SDI->operands()) {
 531     auto Result = LVI->getPredicateAt(ICmpInst::ICMP_SGE, O, Zero, SDI);
 532     if (Result != LazyValueInfo::True)
 533       return false;
 534   }
 535   return true;
 536 }
 537
 538 /// Try to shrink a udiv/urem's width down to the smallest power of two that's
 539 /// sufficient to contain its operands.
 540 static bool processUDivOrURem(BinaryOperator *Instr, LazyValueInfo *LVI) {
 541   assert(Instr->getOpcode() == Instruction::UDiv ||
 542          Instr->getOpcode() == Instruction::URem);
 543   if (Instr->getType()->isVectorTy())
 544     return false;
 545
 546   // Find the smallest power of two bitwidth that's sufficient to hold Instr's
 547   // operands.
 548   auto OrigWidth = Instr->getType()->getIntegerBitWidth();
 549   ConstantRange OperandRange(OrigWidth, /*isFullSet=*/false);
 550   for (Value *Operand : Instr->operands()) {
 551     OperandRange = OperandRange.unionWith(
 552         LVI->getConstantRange(Operand, Instr->getParent()));
 553   }
 554   // Don't shrink below 8 bits wide.
 555   unsigned NewWidth = std::max<unsigned>(
 556       PowerOf2Ceil(OperandRange.getUnsignedMax().getActiveBits()), 8);
 557   // NewWidth might be greater than OrigWidth if OrigWidth is not a power of
 558   // two.
 559   if (NewWidth >= OrigWidth)
 560     return false;
 561
 562   ++NumUDivs;
 563   IRBuilder<> B{Instr};
 564   auto *TruncTy = Type::getIntNTy(Instr->getContext(), NewWidth);
 565   auto *LHS = B.CreateTruncOrBitCast(Instr->getOperand(0), TruncTy,
 566                                      Instr->getName() + ".lhs.trunc");
 567   auto *RHS = B.CreateTruncOrBitCast(Instr->getOperand(1), TruncTy,
 568                                      Instr->getName() + ".rhs.trunc");
 569   auto *BO = B.CreateBinOp(Instr->getOpcode(), LHS, RHS, Instr->getName());
 570   auto *Zext = B.CreateZExt(BO, Instr->getType(), Instr->getName() + ".zext");
 571   if (auto *BinOp = dyn_cast<BinaryOperator>(BO))
 572     if (BinOp->getOpcode() == Instruction::UDiv)
 573       BinOp->setIsExact(Instr->isExact());
 574
 575   Instr->replaceAllUsesWith(Zext);
 576   Instr->eraseFromParent();
 577   return true;
 578 }
 579
 580 static bool processSRem(BinaryOperator *SDI, LazyValueInfo *LVI) {
 581   if (SDI->getType()->isVectorTy() || !hasPositiveOperands(SDI, LVI))
 582     return false;
 583
 584   ++NumSRems;
 585   auto *BO = BinaryOperator::CreateURem(SDI->getOperand(0), SDI->getOperand(1),
 586                                         SDI->getName(), SDI);
 587   BO->setDebugLoc(SDI->getDebugLoc());
 588   SDI->replaceAllUsesWith(BO);
 589   SDI->eraseFromParent();
 590
 591   // Try to process our new urem.
 592   processUDivOrURem(BO, LVI);
 593
 594   return true;
 595 }
 596
 597 /// See if LazyValueInfo's ability to exploit edge conditions or range
 598 /// information is sufficient to prove the both operands of this SDiv are
 599 /// positive.  If this is the case, replace the SDiv with a UDiv. Even for local
 600 /// conditions, this can sometimes prove conditions instcombine can't by
 601 /// exploiting range information.
 602 static bool processSDiv(BinaryOperator *SDI, LazyValueInfo *LVI) {
 603   if (SDI->getType()->isVectorTy() || !hasPositiveOperands(SDI, LVI))
 604     return false;
 605
 606   ++NumSDivs;
 607   auto *BO = BinaryOperator::CreateUDiv(SDI->getOperand(0), SDI->getOperand(1),
 608                                         SDI->getName(), SDI);
 609   BO->setDebugLoc(SDI->getDebugLoc());
 610   BO->setIsExact(SDI->isExact());
 611   SDI->replaceAllUsesWith(BO);
 612   SDI->eraseFromParent();
 613
 614   // Try to simplify our new udiv.
 615   processUDivOrURem(BO, LVI);
 616
 617   return true;
 618 }
 619
 620 static bool processAShr(BinaryOperator *SDI, LazyValueInfo *LVI) {
 621   if (SDI->getType()->isVectorTy())
 622     return false;
 623
 624   Constant *Zero = ConstantInt::get(SDI->getType(), 0);
 625   if (LVI->getPredicateAt(ICmpInst::ICMP_SGE, SDI->getOperand(0), Zero, SDI) !=
 626       LazyValueInfo::True)
 627     return false;
 628
 629   ++NumAShrs;
 630   auto *BO = BinaryOperator::CreateLShr(SDI->getOperand(0), SDI->getOperand(1),
 631                                         SDI->getName(), SDI);
 632   BO->setDebugLoc(SDI->getDebugLoc());
 633   BO->setIsExact(SDI->isExact());
 634   SDI->replaceAllUsesWith(BO);
 635   SDI->eraseFromParent();
 636
 637   return true;
 638 }
 639
 640 static bool processBinOp(BinaryOperator *BinOp, LazyValueInfo *LVI) {
 641   using OBO = OverflowingBinaryOperator;
 642
 643   if (DontAddNoWrapFlags)
 644     return false;
 645
 646   if (BinOp->getType()->isVectorTy())
 647     return false;
 648
 649   bool NSW = BinOp->hasNoSignedWrap();
 650   bool NUW = BinOp->hasNoUnsignedWrap();
 651   if (NSW && NUW)
 652     return false;
 653
 654   BasicBlock *BB = BinOp->getParent();
 655
 656   Value *LHS = BinOp->getOperand(0);
 657   Value *RHS = BinOp->getOperand(1);
 658
 659   ConstantRange LRange = LVI->getConstantRange(LHS, BB, BinOp);
 660   ConstantRange RRange = LVI->getConstantRange(RHS, BB, BinOp);
 661
 662   bool Changed = false;
 663   if (!NUW) {
 664     ConstantRange NUWRange = ConstantRange::makeGuaranteedNoWrapRegion(
 665         BinOp->getOpcode(), RRange, OBO::NoUnsignedWrap);
 666     bool NewNUW = NUWRange.contains(LRange);
 667     BinOp->setHasNoUnsignedWrap(NewNUW);
 668     Changed |= NewNUW;
 669   }
 670   if (!NSW) {
 671     ConstantRange NSWRange = ConstantRange::makeGuaranteedNoWrapRegion(
 672         BinOp->getOpcode(), RRange, OBO::NoSignedWrap);
 673     bool NewNSW = NSWRange.contains(LRange);
 674     BinOp->setHasNoSignedWrap(NewNSW);
 675     Changed |= NewNSW;
 676   }
 677
 678   return Changed;
 679 }
 680
 681 static Constant *getConstantAt(Value *V, Instruction *At, LazyValueInfo *LVI) {
 682   if (Constant *C = LVI->getConstant(V, At->getParent(), At))
 683     return C;
 684
 685   // TODO: The following really should be sunk inside LVI's core algorithm, or
 686   // at least the outer shims around such.
 687   auto *C = dyn_cast<CmpInst>(V);
 688   if (!C) return nullptr;
 689
 690   Value *Op0 = C->getOperand(0);
 691   Constant *Op1 = dyn_cast<Constant>(C->getOperand(1));
 692   if (!Op1) return nullptr;
 693
 694   LazyValueInfo::Tristate Result =
 695     LVI->getPredicateAt(C->getPredicate(), Op0, Op1, At);
 696   if (Result == LazyValueInfo::Unknown)
 697     return nullptr;
 698
 699   return (Result == LazyValueInfo::True) ?
 700     ConstantInt::getTrue(C->getContext()) :
 701     ConstantInt::getFalse(C->getContext());
 702 }
 703
 704 static bool runImpl(Function &F, LazyValueInfo *LVI, DominatorTree *DT,
 705                     const SimplifyQuery &SQ) {
 706   bool FnChanged = false;
 707   // Visiting in a pre-order depth-first traversal causes us to simplify early
 708   // blocks before querying later blocks (which require us to analyze early
 709   // blocks).  Eagerly simplifying shallow blocks means there is strictly less
 710   // work to do for deep blocks.  This also means we don't visit unreachable
 711   // blocks.
 712   for (BasicBlock *BB : depth_first(&F.getEntryBlock())) {
 713     bool BBChanged = false;
 714     for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) {
 715       Instruction *II = &*BI++;
 716       switch (II->getOpcode()) {
 717       case Instruction::Select:
 718         BBChanged |= processSelect(cast<SelectInst>(II), LVI);
 719         break;
 720       case Instruction::PHI:
 721         BBChanged |= processPHI(cast<PHINode>(II), LVI, DT, SQ);
 722         break;
 723       case Instruction::ICmp:
 724       case Instruction::FCmp:
 725         BBChanged |= processCmp(cast<CmpInst>(II), LVI);
 726         break;
 727       case Instruction::Load:
 728       case Instruction::Store:
 729         BBChanged |= processMemAccess(II, LVI);
 730         break;
 731       case Instruction::Call:
 732       case Instruction::Invoke:
 733         BBChanged |= processCallSite(CallSite(II), LVI);
 734         break;
 735       case Instruction::SRem:
 736         BBChanged |= processSRem(cast<BinaryOperator>(II), LVI);
 737         break;
 738       case Instruction::SDiv:
 739         BBChanged |= processSDiv(cast<BinaryOperator>(II), LVI);
 740         break;
 741       case Instruction::UDiv:
 742       case Instruction::URem:
 743         BBChanged |= processUDivOrURem(cast<BinaryOperator>(II), LVI);
 744         break;
 745       case Instruction::AShr:
 746         BBChanged |= processAShr(cast<BinaryOperator>(II), LVI);
 747         break;
 748       case Instruction::Add:
 749       case Instruction::Sub:
 750         BBChanged |= processBinOp(cast<BinaryOperator>(II), LVI);
 751         break;
 752       }
 753     }
 754
 755     Instruction *Term = BB->getTerminator();
 756     switch (Term->getOpcode()) {
 757     case Instruction::Switch:
 758       BBChanged |= processSwitch(cast<SwitchInst>(Term), LVI, DT);
 759       break;
 760     case Instruction::Ret: {
 761       auto *RI = cast<ReturnInst>(Term);
 762       // Try to determine the return value if we can.  This is mainly here to
 763       // simplify the writing of unit tests, but also helps to enable IPO by
 764       // constant folding the return values of callees.
 765       auto *RetVal = RI->getReturnValue();
 766       if (!RetVal) break; // handle "ret void"
 767       if (isa<Constant>(RetVal)) break; // nothing to do
 768       if (auto *C = getConstantAt(RetVal, RI, LVI)) {
 769         ++NumReturns;
 770         RI->replaceUsesOfWith(RetVal, C);
 771         BBChanged = true;
 772       }
 773     }
 774     }
 775
 776     FnChanged |= BBChanged;
 777   }
 778
 779   return FnChanged;
 780 }
 781
 782 bool CorrelatedValuePropagation::runOnFunction(Function &F) {
 783   if (skipFunction(F))
 784     return false;
 785
 786   LazyValueInfo *LVI = &getAnalysis<LazyValueInfoWrapperPass>().getLVI();
 787   DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
 788
 789   return runImpl(F, LVI, DT, getBestSimplifyQuery(*this, F));
 790 }
 791
 792 PreservedAnalyses
 793 CorrelatedValuePropagationPass::run(Function &F, FunctionAnalysisManager &AM) {
 794   LazyValueInfo *LVI = &AM.getResult<LazyValueAnalysis>(F);
 795   DominatorTree *DT = &AM.getResult<DominatorTreeAnalysis>(F);
 796
 797   bool Changed = runImpl(F, LVI, DT, getBestSimplifyQuery(AM, F));
 798
 799   if (!Changed)
 800     return PreservedAnalyses::all();
 801   PreservedAnalyses PA;
 802   PA.preserve<GlobalsAA>();
 803   PA.preserve<DominatorTreeAnalysis>();
 804   PA.preserve<LazyValueAnalysis>();
 805   return PA;
 806 }