lib/Transforms/Scalar/Float2Int.cpp

   1 //===- Float2Int.cpp - Demote floating point ops to work on integers ------===//
   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 Float2Int pass, which aims to demote floating
  10 // point operations to work on integers, where that is losslessly possible.
  11 //
  12 //===----------------------------------------------------------------------===//
  13
  14 #define DEBUG_TYPE "float2int"
  15
  16 #include "llvm/Transforms/Scalar/Float2Int.h"
  17 #include "llvm/ADT/APInt.h"
  18 #include "llvm/ADT/APSInt.h"
  19 #include "llvm/ADT/SmallVector.h"
  20 #include "llvm/Analysis/AliasAnalysis.h"
  21 #include "llvm/Analysis/GlobalsModRef.h"
  22 #include "llvm/IR/Constants.h"
  23 #include "llvm/IR/IRBuilder.h"
  24 #include "llvm/IR/InstIterator.h"
  25 #include "llvm/IR/Instructions.h"
  26 #include "llvm/IR/Module.h"
  27 #include "llvm/Pass.h"
  28 #include "llvm/Support/Debug.h"
  29 #include "llvm/Support/raw_ostream.h"
  30 #include "llvm/Transforms/Scalar.h"
  31 #include <deque>
  32 #include <functional> // For std::function
  33 using namespace llvm;
  34
  35 // The algorithm is simple. Start at instructions that convert from the
  36 // float to the int domain: fptoui, fptosi and fcmp. Walk up the def-use
  37 // graph, using an equivalence datastructure to unify graphs that interfere.
  38 //
  39 // Mappable instructions are those with an integer corrollary that, given
  40 // integer domain inputs, produce an integer output; fadd, for example.
  41 //
  42 // If a non-mappable instruction is seen, this entire def-use graph is marked
  43 // as non-transformable. If we see an instruction that converts from the
  44 // integer domain to FP domain (uitofp,sitofp), we terminate our walk.
  45
  46 /// The largest integer type worth dealing with.
  47 static cl::opt<unsigned>
  48 MaxIntegerBW("float2int-max-integer-bw", cl::init(64), cl::Hidden,
  49              cl::desc("Max integer bitwidth to consider in float2int"
  50                       "(default=64)"));
  51
  52 namespace {
  53   struct Float2IntLegacyPass : public FunctionPass {
  54     static char ID; // Pass identification, replacement for typeid
  55     Float2IntLegacyPass() : FunctionPass(ID) {
  56       initializeFloat2IntLegacyPassPass(*PassRegistry::getPassRegistry());
  57     }
  58
  59     bool runOnFunction(Function &F) override {
  60       if (skipFunction(F))
  61         return false;
  62
  63       return Impl.runImpl(F);
  64     }
  65
  66     void getAnalysisUsage(AnalysisUsage &AU) const override {
  67       AU.setPreservesCFG();
  68       AU.addPreserved<GlobalsAAWrapperPass>();
  69     }
  70
  71   private:
  72     Float2IntPass Impl;
  73   };
  74 }
  75
  76 char Float2IntLegacyPass::ID = 0;
  77 INITIALIZE_PASS(Float2IntLegacyPass, "float2int", "Float to int", false, false)
  78
  79 // Given a FCmp predicate, return a matching ICmp predicate if one
  80 // exists, otherwise return BAD_ICMP_PREDICATE.
  81 static CmpInst::Predicate mapFCmpPred(CmpInst::Predicate P) {
  82   switch (P) {
  83   case CmpInst::FCMP_OEQ:
  84   case CmpInst::FCMP_UEQ:
  85     return CmpInst::ICMP_EQ;
  86   case CmpInst::FCMP_OGT:
  87   case CmpInst::FCMP_UGT:
  88     return CmpInst::ICMP_SGT;
  89   case CmpInst::FCMP_OGE:
  90   case CmpInst::FCMP_UGE:
  91     return CmpInst::ICMP_SGE;
  92   case CmpInst::FCMP_OLT:
  93   case CmpInst::FCMP_ULT:
  94     return CmpInst::ICMP_SLT;
  95   case CmpInst::FCMP_OLE:
  96   case CmpInst::FCMP_ULE:
  97     return CmpInst::ICMP_SLE;
  98   case CmpInst::FCMP_ONE:
  99   case CmpInst::FCMP_UNE:
 100     return CmpInst::ICMP_NE;
 101   default:
 102     return CmpInst::BAD_ICMP_PREDICATE;
 103   }
 104 }
 105
 106 // Given a floating point binary operator, return the matching
 107 // integer version.
 108 static Instruction::BinaryOps mapBinOpcode(unsigned Opcode) {
 109   switch (Opcode) {
 110   default: llvm_unreachable("Unhandled opcode!");
 111   case Instruction::FAdd: return Instruction::Add;
 112   case Instruction::FSub: return Instruction::Sub;
 113   case Instruction::FMul: return Instruction::Mul;
 114   }
 115 }
 116
 117 // Find the roots - instructions that convert from the FP domain to
 118 // integer domain.
 119 void Float2IntPass::findRoots(Function &F, SmallPtrSet<Instruction*,8> &Roots) {
 120   for (auto &I : instructions(F)) {
 121     if (isa<VectorType>(I.getType()))
 122       continue;
 123     switch (I.getOpcode()) {
 124     default: break;
 125     case Instruction::FPToUI:
 126     case Instruction::FPToSI:
 127       Roots.insert(&I);
 128       break;
 129     case Instruction::FCmp:
 130       if (mapFCmpPred(cast<CmpInst>(&I)->getPredicate()) !=
 131           CmpInst::BAD_ICMP_PREDICATE)
 132         Roots.insert(&I);
 133       break;
 134     }
 135   }
 136 }
 137
 138 // Helper - mark I as having been traversed, having range R.
 139 void Float2IntPass::seen(Instruction *I, ConstantRange R) {
 140   LLVM_DEBUG(dbgs() << "F2I: " << *I << ":" << R << "\n");
 141   auto IT = SeenInsts.find(I);
 142   if (IT != SeenInsts.end())
 143     IT->second = std::move(R);
 144   else
 145     SeenInsts.insert(std::make_pair(I, std::move(R)));
 146 }
 147
 148 // Helper - get a range representing a poison value.
 149 ConstantRange Float2IntPass::badRange() {
 150   return ConstantRange(MaxIntegerBW + 1, true);
 151 }
 152 ConstantRange Float2IntPass::unknownRange() {
 153   return ConstantRange(MaxIntegerBW + 1, false);
 154 }
 155 ConstantRange Float2IntPass::validateRange(ConstantRange R) {
 156   if (R.getBitWidth() > MaxIntegerBW + 1)
 157     return badRange();
 158   return R;
 159 }
 160
 161 // The most obvious way to structure the search is a depth-first, eager
 162 // search from each root. However, that require direct recursion and so
 163 // can only handle small instruction sequences. Instead, we split the search
 164 // up into two phases:
 165 //   - walkBackwards:  A breadth-first walk of the use-def graph starting from
 166 //                     the roots. Populate "SeenInsts" with interesting
 167 //                     instructions and poison values if they're obvious and
 168 //                     cheap to compute. Calculate the equivalance set structure
 169 //                     while we're here too.
 170 //   - walkForwards:  Iterate over SeenInsts in reverse order, so we visit
 171 //                     defs before their uses. Calculate the real range info.
 172
 173 // Breadth-first walk of the use-def graph; determine the set of nodes
 174 // we care about and eagerly determine if some of them are poisonous.
 175 void Float2IntPass::walkBackwards(const SmallPtrSetImpl<Instruction*> &Roots) {
 176   std::deque<Instruction*> Worklist(Roots.begin(), Roots.end());
 177   while (!Worklist.empty()) {
 178     Instruction *I = Worklist.back();
 179     Worklist.pop_back();
 180
 181     if (SeenInsts.find(I) != SeenInsts.end())
 182       // Seen already.
 183       continue;
 184
 185     switch (I->getOpcode()) {
 186       // FIXME: Handle select and phi nodes.
 187     default:
 188       // Path terminated uncleanly.
 189       seen(I, badRange());
 190       break;
 191
 192     case Instruction::UIToFP:
 193     case Instruction::SIToFP: {
 194       // Path terminated cleanly - use the type of the integer input to seed
 195       // the analysis.
 196       unsigned BW = I->getOperand(0)->getType()->getPrimitiveSizeInBits();
 197       auto Input = ConstantRange(BW, true);
 198       auto CastOp = (Instruction::CastOps)I->getOpcode();
 199       seen(I, validateRange(Input.castOp(CastOp, MaxIntegerBW+1)));
 200       continue;
 201     }
 202
 203     case Instruction::FAdd:
 204     case Instruction::FSub:
 205     case Instruction::FMul:
 206     case Instruction::FPToUI:
 207     case Instruction::FPToSI:
 208     case Instruction::FCmp:
 209       seen(I, unknownRange());
 210       break;
 211     }
 212
 213     for (Value *O : I->operands()) {
 214       if (Instruction *OI = dyn_cast<Instruction>(O)) {
 215         // Unify def-use chains if they interfere.
 216         ECs.unionSets(I, OI);
 217         if (SeenInsts.find(I)->second != badRange())
 218           Worklist.push_back(OI);
 219       } else if (!isa<ConstantFP>(O)) {
 220         // Not an instruction or ConstantFP? we can't do anything.
 221         seen(I, badRange());
 222       }
 223     }
 224   }
 225 }
 226
 227 // Walk forwards down the list of seen instructions, so we visit defs before
 228 // uses.
 229 void Float2IntPass::walkForwards() {
 230   for (auto &It : reverse(SeenInsts)) {
 231     if (It.second != unknownRange())
 232       continue;
 233
 234     Instruction *I = It.first;
 235     std::function<ConstantRange(ArrayRef<ConstantRange>)> Op;
 236     switch (I->getOpcode()) {
 237       // FIXME: Handle select and phi nodes.
 238     default:
 239     case Instruction::UIToFP:
 240     case Instruction::SIToFP:
 241       llvm_unreachable("Should have been handled in walkForwards!");
 242
 243     case Instruction::FAdd:
 244     case Instruction::FSub:
 245     case Instruction::FMul:
 246       Op = [I](ArrayRef<ConstantRange> Ops) {
 247         assert(Ops.size() == 2 && "its a binary operator!");
 248         auto BinOp = (Instruction::BinaryOps) I->getOpcode();
 249         return Ops[0].binaryOp(BinOp, Ops[1]);
 250       };
 251       break;
 252
 253     //
 254     // Root-only instructions - we'll only see these if they're the
 255     //                          first node in a walk.
 256     //
 257     case Instruction::FPToUI:
 258     case Instruction::FPToSI:
 259       Op = [I](ArrayRef<ConstantRange> Ops) {
 260         assert(Ops.size() == 1 && "FPTo[US]I is a unary operator!");
 261         // Note: We're ignoring the casts output size here as that's what the
 262         // caller expects.
 263         auto CastOp = (Instruction::CastOps)I->getOpcode();
 264         return Ops[0].castOp(CastOp, MaxIntegerBW+1);
 265       };
 266       break;
 267
 268     case Instruction::FCmp:
 269       Op = [](ArrayRef<ConstantRange> Ops) {
 270         assert(Ops.size() == 2 && "FCmp is a binary operator!");
 271         return Ops[0].unionWith(Ops[1]);
 272       };
 273       break;
 274     }
 275
 276     bool Abort = false;
 277     SmallVector<ConstantRange,4> OpRanges;
 278     for (Value *O : I->operands()) {
 279       if (Instruction *OI = dyn_cast<Instruction>(O)) {
 280         assert(SeenInsts.find(OI) != SeenInsts.end() &&
 281                "def not seen before use!");
 282         OpRanges.push_back(SeenInsts.find(OI)->second);
 283       } else if (ConstantFP *CF = dyn_cast<ConstantFP>(O)) {
 284         // Work out if the floating point number can be losslessly represented
 285         // as an integer.
 286         // APFloat::convertToInteger(&Exact) purports to do what we want, but
 287         // the exactness can be too precise. For example, negative zero can
 288         // never be exactly converted to an integer.
 289         //
 290         // Instead, we ask APFloat to round itself to an integral value - this
 291         // preserves sign-of-zero - then compare the result with the original.
 292         //
 293         const APFloat &F = CF->getValueAPF();
 294
 295         // First, weed out obviously incorrect values. Non-finite numbers
 296         // can't be represented and neither can negative zero, unless
 297         // we're in fast math mode.
 298         if (!F.isFinite() ||
 299             (F.isZero() && F.isNegative() && isa<FPMathOperator>(I) &&
 300              !I->hasNoSignedZeros())) {
 301           seen(I, badRange());
 302           Abort = true;
 303           break;
 304         }
 305
 306         APFloat NewF = F;
 307         auto Res = NewF.roundToIntegral(APFloat::rmNearestTiesToEven);
 308         if (Res != APFloat::opOK || NewF.compare(F) != APFloat::cmpEqual) {
 309           seen(I, badRange());
 310           Abort = true;
 311           break;
 312         }
 313         // OK, it's representable. Now get it.
 314         APSInt Int(MaxIntegerBW+1, false);
 315         bool Exact;
 316         CF->getValueAPF().convertToInteger(Int,
 317                                            APFloat::rmNearestTiesToEven,
 318                                            &Exact);
 319         OpRanges.push_back(ConstantRange(Int));
 320       } else {
 321         llvm_unreachable("Should have already marked this as badRange!");
 322       }
 323     }
 324
 325     // Reduce the operands' ranges to a single range and return.
 326     if (!Abort)
 327       seen(I, Op(OpRanges));
 328   }
 329 }
 330
 331 // If there is a valid transform to be done, do it.
 332 bool Float2IntPass::validateAndTransform() {
 333   bool MadeChange = false;
 334
 335   // Iterate over every disjoint partition of the def-use graph.
 336   for (auto It = ECs.begin(), E = ECs.end(); It != E; ++It) {
 337     ConstantRange R(MaxIntegerBW + 1, false);
 338     bool Fail = false;
 339     Type *ConvertedToTy = nullptr;
 340
 341     // For every member of the partition, union all the ranges together.
 342     for (auto MI = ECs.member_begin(It), ME = ECs.member_end();
 343          MI != ME; ++MI) {
 344       Instruction *I = *MI;
 345       auto SeenI = SeenInsts.find(I);
 346       if (SeenI == SeenInsts.end())
 347         continue;
 348
 349       R = R.unionWith(SeenI->second);
 350       // We need to ensure I has no users that have not been seen.
 351       // If it does, transformation would be illegal.
 352       //
 353       // Don't count the roots, as they terminate the graphs.
 354       if (Roots.count(I) == 0) {
 355         // Set the type of the conversion while we're here.
 356         if (!ConvertedToTy)
 357           ConvertedToTy = I->getType();
 358         for (User *U : I->users()) {
 359           Instruction *UI = dyn_cast<Instruction>(U);
 360           if (!UI || SeenInsts.find(UI) == SeenInsts.end()) {
 361             LLVM_DEBUG(dbgs() << "F2I: Failing because of " << *U << "\n");
 362             Fail = true;
 363             break;
 364           }
 365         }
 366       }
 367       if (Fail)
 368         break;
 369     }
 370
 371     // If the set was empty, or we failed, or the range is poisonous,
 372     // bail out.
 373     if (ECs.member_begin(It) == ECs.member_end() || Fail ||
 374         R.isFullSet() || R.isSignWrappedSet())
 375       continue;
 376     assert(ConvertedToTy && "Must have set the convertedtoty by this point!");
 377
 378     // The number of bits required is the maximum of the upper and
 379     // lower limits, plus one so it can be signed.
 380     unsigned MinBW = std::max(R.getLower().getMinSignedBits(),
 381                               R.getUpper().getMinSignedBits()) + 1;
 382     LLVM_DEBUG(dbgs() << "F2I: MinBitwidth=" << MinBW << ", R: " << R << "\n");
 383
 384     // If we've run off the realms of the exactly representable integers,
 385     // the floating point result will differ from an integer approximation.
 386
 387     // Do we need more bits than are in the mantissa of the type we converted
 388     // to? semanticsPrecision returns the number of mantissa bits plus one
 389     // for the sign bit.
 390     unsigned MaxRepresentableBits
 391       = APFloat::semanticsPrecision(ConvertedToTy->getFltSemantics()) - 1;
 392     if (MinBW > MaxRepresentableBits) {
 393       LLVM_DEBUG(dbgs() << "F2I: Value not guaranteed to be representable!\n");
 394       continue;
 395     }
 396     if (MinBW > 64) {
 397       LLVM_DEBUG(
 398           dbgs() << "F2I: Value requires more than 64 bits to represent!\n");
 399       continue;
 400     }
 401
 402     // OK, R is known to be representable. Now pick a type for it.
 403     // FIXME: Pick the smallest legal type that will fit.
 404     Type *Ty = (MinBW > 32) ? Type::getInt64Ty(*Ctx) : Type::getInt32Ty(*Ctx);
 405
 406     for (auto MI = ECs.member_begin(It), ME = ECs.member_end();
 407          MI != ME; ++MI)
 408       convert(*MI, Ty);
 409     MadeChange = true;
 410   }
 411
 412   return MadeChange;
 413 }
 414
 415 Value *Float2IntPass::convert(Instruction *I, Type *ToTy) {
 416   if (ConvertedInsts.find(I) != ConvertedInsts.end())
 417     // Already converted this instruction.
 418     return ConvertedInsts[I];
 419
 420   SmallVector<Value*,4> NewOperands;
 421   for (Value *V : I->operands()) {
 422     // Don't recurse if we're an instruction that terminates the path.
 423     if (I->getOpcode() == Instruction::UIToFP ||
 424         I->getOpcode() == Instruction::SIToFP) {
 425       NewOperands.push_back(V);
 426     } else if (Instruction *VI = dyn_cast<Instruction>(V)) {
 427       NewOperands.push_back(convert(VI, ToTy));
 428     } else if (ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
 429       APSInt Val(ToTy->getPrimitiveSizeInBits(), /*IsUnsigned=*/false);
 430       bool Exact;
 431       CF->getValueAPF().convertToInteger(Val,
 432                                          APFloat::rmNearestTiesToEven,
 433                                          &Exact);
 434       NewOperands.push_back(ConstantInt::get(ToTy, Val));
 435     } else {
 436       llvm_unreachable("Unhandled operand type?");
 437     }
 438   }
 439
 440   // Now create a new instruction.
 441   IRBuilder<> IRB(I);
 442   Value *NewV = nullptr;
 443   switch (I->getOpcode()) {
 444   default: llvm_unreachable("Unhandled instruction!");
 445
 446   case Instruction::FPToUI:
 447     NewV = IRB.CreateZExtOrTrunc(NewOperands[0], I->getType());
 448     break;
 449
 450   case Instruction::FPToSI:
 451     NewV = IRB.CreateSExtOrTrunc(NewOperands[0], I->getType());
 452     break;
 453
 454   case Instruction::FCmp: {
 455     CmpInst::Predicate P = mapFCmpPred(cast<CmpInst>(I)->getPredicate());
 456     assert(P != CmpInst::BAD_ICMP_PREDICATE && "Unhandled predicate!");
 457     NewV = IRB.CreateICmp(P, NewOperands[0], NewOperands[1], I->getName());
 458     break;
 459   }
 460
 461   case Instruction::UIToFP:
 462     NewV = IRB.CreateZExtOrTrunc(NewOperands[0], ToTy);
 463     break;
 464
 465   case Instruction::SIToFP:
 466     NewV = IRB.CreateSExtOrTrunc(NewOperands[0], ToTy);
 467     break;
 468
 469   case Instruction::FAdd:
 470   case Instruction::FSub:
 471   case Instruction::FMul:
 472     NewV = IRB.CreateBinOp(mapBinOpcode(I->getOpcode()),
 473                            NewOperands[0], NewOperands[1],
 474                            I->getName());
 475     break;
 476   }
 477
 478   // If we're a root instruction, RAUW.
 479   if (Roots.count(I))
 480     I->replaceAllUsesWith(NewV);
 481
 482   ConvertedInsts[I] = NewV;
 483   return NewV;
 484 }
 485
 486 // Perform dead code elimination on the instructions we just modified.
 487 void Float2IntPass::cleanup() {
 488   for (auto &I : reverse(ConvertedInsts))
 489     I.first->eraseFromParent();
 490 }
 491
 492 bool Float2IntPass::runImpl(Function &F) {
 493   LLVM_DEBUG(dbgs() << "F2I: Looking at function " << F.getName() << "\n");
 494   // Clear out all state.
 495   ECs = EquivalenceClasses<Instruction*>();
 496   SeenInsts.clear();
 497   ConvertedInsts.clear();
 498   Roots.clear();
 499
 500   Ctx = &F.getParent()->getContext();
 501
 502   findRoots(F, Roots);
 503
 504   walkBackwards(Roots);
 505   walkForwards();
 506
 507   bool Modified = validateAndTransform();
 508   if (Modified)
 509     cleanup();
 510   return Modified;
 511 }
 512
 513 namespace llvm {
 514 FunctionPass *createFloat2IntPass() { return new Float2IntLegacyPass(); }
 515
 516 PreservedAnalyses Float2IntPass::run(Function &F, FunctionAnalysisManager &) {
 517   if (!runImpl(F))
 518     return PreservedAnalyses::all();
 519
 520   PreservedAnalyses PA;
 521   PA.preserveSet<CFGAnalyses>();
 522   PA.preserve<GlobalsAA>();
 523   return PA;
 524 }
 525 } // End namespace llvm