lib/Transforms/Utils/BypassSlowDivision.cpp

   1 //===-- BypassSlowDivision.cpp - Bypass slow division ---------------------===//
   2 //
   3 //                     The LLVM Compiler Infrastructure
   4 //
   5 // This file is distributed under the University of Illinois Open Source
   6 // License. See LICENSE.TXT for details.
   7 //
   8 //===----------------------------------------------------------------------===//
   9 //
  10 // This file contains an optimization for div and rem on architectures that
  11 // execute short instructions significantly faster than longer instructions.
  12 // For example, on Intel Atom 32-bit divides are slow enough that during
  13 // runtime it is profitable to check the value of the operands, and if they are
  14 // positive and less than 256 use an unsigned 8-bit divide.
  15 //
  16 //===----------------------------------------------------------------------===//
  17
  18 #define DEBUG_TYPE "bypass-slow-division"
  19 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
  20 #include "llvm/ADT/DenseMap.h"
  21 #include "llvm/IR/Function.h"
  22 #include "llvm/IR/IRBuilder.h"
  23 #include "llvm/IR/Instructions.h"
  24
  25 using namespace llvm;
  26
  27 namespace {
  28   struct DivOpInfo {
  29     bool SignedOp;
  30     Value *Dividend;
  31     Value *Divisor;
  32
  33     DivOpInfo(bool InSignedOp, Value *InDividend, Value *InDivisor)
  34       : SignedOp(InSignedOp), Dividend(InDividend), Divisor(InDivisor) {}
  35   };
  36
  37   struct DivPhiNodes {
  38     PHINode *Quotient;
  39     PHINode *Remainder;
  40
  41     DivPhiNodes(PHINode *InQuotient, PHINode *InRemainder)
  42       : Quotient(InQuotient), Remainder(InRemainder) {}
  43   };
  44 }
  45
  46 namespace llvm {
  47   template<>
  48   struct DenseMapInfo<DivOpInfo> {
  49     static bool isEqual(const DivOpInfo &Val1, const DivOpInfo &Val2) {
  50       return Val1.SignedOp == Val2.SignedOp &&
  51              Val1.Dividend == Val2.Dividend &&
  52              Val1.Divisor == Val2.Divisor;
  53     }
  54
  55     static DivOpInfo getEmptyKey() {
  56       return DivOpInfo(false, 0, 0);
  57     }
  58
  59     static DivOpInfo getTombstoneKey() {
  60       return DivOpInfo(true, 0, 0);
  61     }
  62
  63     static unsigned getHashValue(const DivOpInfo &Val) {
  64       return (unsigned)(reinterpret_cast<uintptr_t>(Val.Dividend) ^
  65                         reinterpret_cast<uintptr_t>(Val.Divisor)) ^
  66                         (unsigned)Val.SignedOp;
  67     }
  68   };
  69
  70   typedef DenseMap<DivOpInfo, DivPhiNodes> DivCacheTy;
  71 }
  72
  73 // insertFastDiv - Substitutes the div/rem instruction with code that checks the
  74 // value of the operands and uses a shorter-faster div/rem instruction when
  75 // possible and the longer-slower div/rem instruction otherwise.
  76 static bool insertFastDiv(Function &F,
  77                           Function::iterator &I,
  78                           BasicBlock::iterator &J,
  79                           IntegerType *BypassType,
  80                           bool UseDivOp,
  81                           bool UseSignedOp,
  82                           DivCacheTy &PerBBDivCache) {
  83   // Get instruction operands
  84   Instruction *Instr = J;
  85   Value *Dividend = Instr->getOperand(0);
  86   Value *Divisor = Instr->getOperand(1);
  87
  88   if (isa<ConstantInt>(Divisor) ||
  89       (isa<ConstantInt>(Dividend) && isa<ConstantInt>(Divisor))) {
  90     // Operations with immediate values should have
  91     // been solved and replaced during compile time.
  92     return false;
  93   }
  94
  95   // Basic Block is split before divide
  96   BasicBlock *MainBB = I;
  97   BasicBlock *SuccessorBB = I->splitBasicBlock(J);
  98   ++I; //advance iterator I to successorBB
  99
 100   // Add new basic block for slow divide operation
 101   BasicBlock *SlowBB = BasicBlock::Create(F.getContext(), "",
 102                                           MainBB->getParent(), SuccessorBB);
 103   SlowBB->moveBefore(SuccessorBB);
 104   IRBuilder<> SlowBuilder(SlowBB, SlowBB->begin());
 105   Value *SlowQuotientV;
 106   Value *SlowRemainderV;
 107   if (UseSignedOp) {
 108     SlowQuotientV = SlowBuilder.CreateSDiv(Dividend, Divisor);
 109     SlowRemainderV = SlowBuilder.CreateSRem(Dividend, Divisor);
 110   } else {
 111     SlowQuotientV = SlowBuilder.CreateUDiv(Dividend, Divisor);
 112     SlowRemainderV = SlowBuilder.CreateURem(Dividend, Divisor);
 113   }
 114   SlowBuilder.CreateBr(SuccessorBB);
 115
 116   // Add new basic block for fast divide operation
 117   BasicBlock *FastBB = BasicBlock::Create(F.getContext(), "",
 118                                           MainBB->getParent(), SuccessorBB);
 119   FastBB->moveBefore(SlowBB);
 120   IRBuilder<> FastBuilder(FastBB, FastBB->begin());
 121   Value *ShortDivisorV = FastBuilder.CreateCast(Instruction::Trunc, Divisor,
 122                                                 BypassType);
 123   Value *ShortDividendV = FastBuilder.CreateCast(Instruction::Trunc, Dividend,
 124                                                  BypassType);
 125
 126   // udiv/urem because optimization only handles positive numbers
 127   Value *ShortQuotientV = FastBuilder.CreateExactUDiv(ShortDividendV,
 128                                                       ShortDivisorV);
 129   Value *ShortRemainderV = FastBuilder.CreateURem(ShortDividendV,
 130                                                   ShortDivisorV);
 131   Value *FastQuotientV = FastBuilder.CreateCast(Instruction::ZExt,
 132                                                 ShortQuotientV,
 133                                                 Dividend->getType());
 134   Value *FastRemainderV = FastBuilder.CreateCast(Instruction::ZExt,
 135                                                  ShortRemainderV,
 136                                                  Dividend->getType());
 137   FastBuilder.CreateBr(SuccessorBB);
 138
 139   // Phi nodes for result of div and rem
 140   IRBuilder<> SuccessorBuilder(SuccessorBB, SuccessorBB->begin());
 141   PHINode *QuoPhi = SuccessorBuilder.CreatePHI(Instr->getType(), 2);
 142   QuoPhi->addIncoming(SlowQuotientV, SlowBB);
 143   QuoPhi->addIncoming(FastQuotientV, FastBB);
 144   PHINode *RemPhi = SuccessorBuilder.CreatePHI(Instr->getType(), 2);
 145   RemPhi->addIncoming(SlowRemainderV, SlowBB);
 146   RemPhi->addIncoming(FastRemainderV, FastBB);
 147
 148   // Replace Instr with appropriate phi node
 149   if (UseDivOp)
 150     Instr->replaceAllUsesWith(QuoPhi);
 151   else
 152     Instr->replaceAllUsesWith(RemPhi);
 153   Instr->eraseFromParent();
 154
 155   // Combine operands into a single value with OR for value testing below
 156   MainBB->getInstList().back().eraseFromParent();
 157   IRBuilder<> MainBuilder(MainBB, MainBB->end());
 158   Value *OrV = MainBuilder.CreateOr(Dividend, Divisor);
 159
 160   // BitMask is inverted to check if the operands are
 161   // larger than the bypass type
 162   uint64_t BitMask = ~BypassType->getBitMask();
 163   Value *AndV = MainBuilder.CreateAnd(OrV, BitMask);
 164
 165   // Compare operand values and branch
 166   Value *ZeroV = ConstantInt::getSigned(Dividend->getType(), 0);
 167   Value *CmpV = MainBuilder.CreateICmpEQ(AndV, ZeroV);
 168   MainBuilder.CreateCondBr(CmpV, FastBB, SlowBB);
 169
 170   // point iterator J at first instruction of successorBB
 171   J = I->begin();
 172
 173   // Cache phi nodes to be used later in place of other instances
 174   // of div or rem with the same sign, dividend, and divisor
 175   DivOpInfo Key(UseSignedOp, Dividend, Divisor);
 176   DivPhiNodes Value(QuoPhi, RemPhi);
 177   PerBBDivCache.insert(std::pair<DivOpInfo, DivPhiNodes>(Key, Value));
 178   return true;
 179 }
 180
 181 // reuseOrInsertFastDiv - Reuses previously computed dividend or remainder if
 182 // operands and operation are identical. Otherwise call insertFastDiv to perform
 183 // the optimization and cache the resulting dividend and remainder.
 184 static bool reuseOrInsertFastDiv(Function &F,
 185                                  Function::iterator &I,
 186                                  BasicBlock::iterator &J,
 187                                  IntegerType *BypassType,
 188                                  bool UseDivOp,
 189                                  bool UseSignedOp,
 190                                  DivCacheTy &PerBBDivCache) {
 191   // Get instruction operands
 192   Instruction *Instr = J;
 193   DivOpInfo Key(UseSignedOp, Instr->getOperand(0), Instr->getOperand(1));
 194   DivCacheTy::iterator CacheI = PerBBDivCache.find(Key);
 195
 196   if (CacheI == PerBBDivCache.end()) {
 197     // If previous instance does not exist, insert fast div
 198     return insertFastDiv(F, I, J, BypassType, UseDivOp, UseSignedOp,
 199                          PerBBDivCache);
 200   }
 201
 202   // Replace operation value with previously generated phi node
 203   DivPhiNodes &Value = CacheI->second;
 204   if (UseDivOp) {
 205     // Replace all uses of div instruction with quotient phi node
 206     J->replaceAllUsesWith(Value.Quotient);
 207   } else {
 208     // Replace all uses of rem instruction with remainder phi node
 209     J->replaceAllUsesWith(Value.Remainder);
 210   }
 211
 212   // Advance to next operation
 213   ++J;
 214
 215   // Remove redundant operation
 216   Instr->eraseFromParent();
 217   return true;
 218 }
 219
 220 // bypassSlowDivision - This optimization identifies DIV instructions that can
 221 // be profitably bypassed and carried out with a shorter, faster divide.
 222 bool llvm::bypassSlowDivision(Function &F,
 223                               Function::iterator &I,
 224                               const DenseMap<unsigned int, unsigned int> &BypassWidths) {
 225   DivCacheTy DivCache;
 226
 227   bool MadeChange = false;
 228   for (BasicBlock::iterator J = I->begin(); J != I->end(); J++) {
 229
 230     // Get instruction details
 231     unsigned Opcode = J->getOpcode();
 232     bool UseDivOp = Opcode == Instruction::SDiv || Opcode == Instruction::UDiv;
 233     bool UseRemOp = Opcode == Instruction::SRem || Opcode == Instruction::URem;
 234     bool UseSignedOp = Opcode == Instruction::SDiv ||
 235                        Opcode == Instruction::SRem;
 236
 237     // Only optimize div or rem ops
 238     if (!UseDivOp && !UseRemOp)
 239       continue;
 240
 241     // Skip division on vector types, only optimize integer instructions
 242     if (!J->getType()->isIntegerTy())
 243       continue;
 244
 245     // Get bitwidth of div/rem instruction
 246     IntegerType *T = cast<IntegerType>(J->getType());
 247     unsigned int bitwidth = T->getBitWidth();
 248
 249     // Continue if bitwidth is not bypassed
 250     DenseMap<unsigned int, unsigned int>::const_iterator BI = BypassWidths.find(bitwidth);
 251     if (BI == BypassWidths.end())
 252       continue;
 253
 254     // Get type for div/rem instruction with bypass bitwidth
 255     IntegerType *BT = IntegerType::get(J->getContext(), BI->second);
 256
 257     MadeChange |= reuseOrInsertFastDiv(F, I, J, BT, UseDivOp,
 258                                        UseSignedOp, DivCache);
 259   }
 260
 261   return MadeChange;
 262 }