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 #include "llvm/Transforms/Utils/BypassSlowDivision.h"
  19 #include "llvm/ADT/DenseMap.h"
  20 #include "llvm/IR/Function.h"
  21 #include "llvm/IR/IRBuilder.h"
  22 #include "llvm/IR/Instructions.h"
  23
  24 using namespace llvm;
  25
  26 #define DEBUG_TYPE "bypass-slow-division"
  27
  28 namespace {
  29   struct DivOpInfo {
  30     bool SignedOp;
  31     Value *Dividend;
  32     Value *Divisor;
  33
  34     DivOpInfo(bool InSignedOp, Value *InDividend, Value *InDivisor)
  35       : SignedOp(InSignedOp), Dividend(InDividend), Divisor(InDivisor) {}
  36   };
  37
  38   struct DivPhiNodes {
  39     PHINode *Quotient;
  40     PHINode *Remainder;
  41
  42     DivPhiNodes(PHINode *InQuotient, PHINode *InRemainder)
  43       : Quotient(InQuotient), Remainder(InRemainder) {}
  44   };
  45 }
  46
  47 namespace llvm {
  48   template<>
  49   struct DenseMapInfo<DivOpInfo> {
  50     static bool isEqual(const DivOpInfo &Val1, const DivOpInfo &Val2) {
  51       return Val1.SignedOp == Val2.SignedOp &&
  52              Val1.Dividend == Val2.Dividend &&
  53              Val1.Divisor == Val2.Divisor;
  54     }
  55
  56     static DivOpInfo getEmptyKey() {
  57       return DivOpInfo(false, 0, 0);
  58     }
  59
  60     static DivOpInfo getTombstoneKey() {
  61       return DivOpInfo(true, 0, 0);
  62     }
  63
  64     static unsigned getHashValue(const DivOpInfo &Val) {
  65       return (unsigned)(reinterpret_cast<uintptr_t>(Val.Dividend) ^
  66                         reinterpret_cast<uintptr_t>(Val.Divisor)) ^
  67                         (unsigned)Val.SignedOp;
  68     }
  69   };
  70
  71   typedef DenseMap<DivOpInfo, DivPhiNodes> DivCacheTy;
  72 }
  73
  74 // insertFastDiv - Substitutes the div/rem instruction with code that checks the
  75 // value of the operands and uses a shorter-faster div/rem instruction when
  76 // possible and the longer-slower div/rem instruction otherwise.
  77 static bool insertFastDiv(Function &F,
  78                           Function::iterator &I,
  79                           BasicBlock::iterator &J,
  80                           IntegerType *BypassType,
  81                           bool UseDivOp,
  82                           bool UseSignedOp,
  83                           DivCacheTy &PerBBDivCache) {
  84   // Get instruction operands
  85   Instruction *Instr = J;
  86   Value *Dividend = Instr->getOperand(0);
  87   Value *Divisor = Instr->getOperand(1);
  88
  89   if (isa<ConstantInt>(Divisor) ||
  90       (isa<ConstantInt>(Dividend) && isa<ConstantInt>(Divisor))) {
  91     // Operations with immediate values should have
  92     // been solved and replaced during compile time.
  93     return false;
  94   }
  95
  96   // Basic Block is split before divide
  97   BasicBlock *MainBB = I;
  98   BasicBlock *SuccessorBB = I->splitBasicBlock(J);
  99   ++I; //advance iterator I to successorBB
 100
 101   // Add new basic block for slow divide operation
 102   BasicBlock *SlowBB = BasicBlock::Create(F.getContext(), "",
 103                                           MainBB->getParent(), SuccessorBB);
 104   SlowBB->moveBefore(SuccessorBB);
 105   IRBuilder<> SlowBuilder(SlowBB, SlowBB->begin());
 106   Value *SlowQuotientV;
 107   Value *SlowRemainderV;
 108   if (UseSignedOp) {
 109     SlowQuotientV = SlowBuilder.CreateSDiv(Dividend, Divisor);
 110     SlowRemainderV = SlowBuilder.CreateSRem(Dividend, Divisor);
 111   } else {
 112     SlowQuotientV = SlowBuilder.CreateUDiv(Dividend, Divisor);
 113     SlowRemainderV = SlowBuilder.CreateURem(Dividend, Divisor);
 114   }
 115   SlowBuilder.CreateBr(SuccessorBB);
 116
 117   // Add new basic block for fast divide operation
 118   BasicBlock *FastBB = BasicBlock::Create(F.getContext(), "",
 119                                           MainBB->getParent(), SuccessorBB);
 120   FastBB->moveBefore(SlowBB);
 121   IRBuilder<> FastBuilder(FastBB, FastBB->begin());
 122   Value *ShortDivisorV = FastBuilder.CreateCast(Instruction::Trunc, Divisor,
 123                                                 BypassType);
 124   Value *ShortDividendV = FastBuilder.CreateCast(Instruction::Trunc, Dividend,
 125                                                  BypassType);
 126
 127   // udiv/urem because optimization only handles positive numbers
 128   Value *ShortQuotientV = FastBuilder.CreateExactUDiv(ShortDividendV,
 129                                                       ShortDivisorV);
 130   Value *ShortRemainderV = FastBuilder.CreateURem(ShortDividendV,
 131                                                   ShortDivisorV);
 132   Value *FastQuotientV = FastBuilder.CreateCast(Instruction::ZExt,
 133                                                 ShortQuotientV,
 134                                                 Dividend->getType());
 135   Value *FastRemainderV = FastBuilder.CreateCast(Instruction::ZExt,
 136                                                  ShortRemainderV,
 137                                                  Dividend->getType());
 138   FastBuilder.CreateBr(SuccessorBB);
 139
 140   // Phi nodes for result of div and rem
 141   IRBuilder<> SuccessorBuilder(SuccessorBB, SuccessorBB->begin());
 142   PHINode *QuoPhi = SuccessorBuilder.CreatePHI(Instr->getType(), 2);
 143   QuoPhi->addIncoming(SlowQuotientV, SlowBB);
 144   QuoPhi->addIncoming(FastQuotientV, FastBB);
 145   PHINode *RemPhi = SuccessorBuilder.CreatePHI(Instr->getType(), 2);
 146   RemPhi->addIncoming(SlowRemainderV, SlowBB);
 147   RemPhi->addIncoming(FastRemainderV, FastBB);
 148
 149   // Replace Instr with appropriate phi node
 150   if (UseDivOp)
 151     Instr->replaceAllUsesWith(QuoPhi);
 152   else
 153     Instr->replaceAllUsesWith(RemPhi);
 154   Instr->eraseFromParent();
 155
 156   // Combine operands into a single value with OR for value testing below
 157   MainBB->getInstList().back().eraseFromParent();
 158   IRBuilder<> MainBuilder(MainBB, MainBB->end());
 159   Value *OrV = MainBuilder.CreateOr(Dividend, Divisor);
 160
 161   // BitMask is inverted to check if the operands are
 162   // larger than the bypass type
 163   uint64_t BitMask = ~BypassType->getBitMask();
 164   Value *AndV = MainBuilder.CreateAnd(OrV, BitMask);
 165
 166   // Compare operand values and branch
 167   Value *ZeroV = ConstantInt::getSigned(Dividend->getType(), 0);
 168   Value *CmpV = MainBuilder.CreateICmpEQ(AndV, ZeroV);
 169   MainBuilder.CreateCondBr(CmpV, FastBB, SlowBB);
 170
 171   // point iterator J at first instruction of successorBB
 172   J = I->begin();
 173
 174   // Cache phi nodes to be used later in place of other instances
 175   // of div or rem with the same sign, dividend, and divisor
 176   DivOpInfo Key(UseSignedOp, Dividend, Divisor);
 177   DivPhiNodes Value(QuoPhi, RemPhi);
 178   PerBBDivCache.insert(std::pair<DivOpInfo, DivPhiNodes>(Key, Value));
 179   return true;
 180 }
 181
 182 // reuseOrInsertFastDiv - Reuses previously computed dividend or remainder if
 183 // operands and operation are identical. Otherwise call insertFastDiv to perform
 184 // the optimization and cache the resulting dividend and remainder.
 185 static bool reuseOrInsertFastDiv(Function &F,
 186                                  Function::iterator &I,
 187                                  BasicBlock::iterator &J,
 188                                  IntegerType *BypassType,
 189                                  bool UseDivOp,
 190                                  bool UseSignedOp,
 191                                  DivCacheTy &PerBBDivCache) {
 192   // Get instruction operands
 193   Instruction *Instr = J;
 194   DivOpInfo Key(UseSignedOp, Instr->getOperand(0), Instr->getOperand(1));
 195   DivCacheTy::iterator CacheI = PerBBDivCache.find(Key);
 196
 197   if (CacheI == PerBBDivCache.end()) {
 198     // If previous instance does not exist, insert fast div
 199     return insertFastDiv(F, I, J, BypassType, UseDivOp, UseSignedOp,
 200                          PerBBDivCache);
 201   }
 202
 203   // Replace operation value with previously generated phi node
 204   DivPhiNodes &Value = CacheI->second;
 205   if (UseDivOp) {
 206     // Replace all uses of div instruction with quotient phi node
 207     J->replaceAllUsesWith(Value.Quotient);
 208   } else {
 209     // Replace all uses of rem instruction with remainder phi node
 210     J->replaceAllUsesWith(Value.Remainder);
 211   }
 212
 213   // Advance to next operation
 214   ++J;
 215
 216   // Remove redundant operation
 217   Instr->eraseFromParent();
 218   return true;
 219 }
 220
 221 // bypassSlowDivision - This optimization identifies DIV instructions that can
 222 // be profitably bypassed and carried out with a shorter, faster divide.
 223 bool llvm::bypassSlowDivision(Function &F,
 224                               Function::iterator &I,
 225                               const DenseMap<unsigned int, unsigned int> &BypassWidths) {
 226   DivCacheTy DivCache;
 227
 228   bool MadeChange = false;
 229   for (BasicBlock::iterator J = I->begin(); J != I->end(); J++) {
 230
 231     // Get instruction details
 232     unsigned Opcode = J->getOpcode();
 233     bool UseDivOp = Opcode == Instruction::SDiv || Opcode == Instruction::UDiv;
 234     bool UseRemOp = Opcode == Instruction::SRem || Opcode == Instruction::URem;
 235     bool UseSignedOp = Opcode == Instruction::SDiv ||
 236                        Opcode == Instruction::SRem;
 237
 238     // Only optimize div or rem ops
 239     if (!UseDivOp && !UseRemOp)
 240       continue;
 241
 242     // Skip division on vector types, only optimize integer instructions
 243     if (!J->getType()->isIntegerTy())
 244       continue;
 245
 246     // Get bitwidth of div/rem instruction
 247     IntegerType *T = cast<IntegerType>(J->getType());
 248     unsigned int bitwidth = T->getBitWidth();
 249
 250     // Continue if bitwidth is not bypassed
 251     DenseMap<unsigned int, unsigned int>::const_iterator BI = BypassWidths.find(bitwidth);
 252     if (BI == BypassWidths.end())
 253       continue;
 254
 255     // Get type for div/rem instruction with bypass bitwidth
 256     IntegerType *BT = IntegerType::get(J->getContext(), BI->second);
 257
 258     MadeChange |= reuseOrInsertFastDiv(F, I, J, BT, UseDivOp,
 259                                        UseSignedOp, DivCache);
 260   }
 261
 262   return MadeChange;
 263 }