LLVM  7.0.0svn
DivRemPairs.cpp
Go to the documentation of this file.
1 //===- DivRemPairs.cpp - Hoist/decompose division and remainder -*- C++ -*-===//
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 pass hoists and/or decomposes integer division and remainder
11 // instructions to enable CFG improvements and better codegen.
12 //
13 //===----------------------------------------------------------------------===//
14 
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/MapVector.h"
18 #include "llvm/ADT/Statistic.h"
21 #include "llvm/IR/Dominators.h"
22 #include "llvm/IR/Function.h"
23 #include "llvm/Pass.h"
24 #include "llvm/Transforms/Scalar.h"
26 using namespace llvm;
27 
28 #define DEBUG_TYPE "div-rem-pairs"
29 STATISTIC(NumPairs, "Number of div/rem pairs");
30 STATISTIC(NumHoisted, "Number of instructions hoisted");
31 STATISTIC(NumDecomposed, "Number of instructions decomposed");
32 
33 /// Find matching pairs of integer div/rem ops (they have the same numerator,
34 /// denominator, and signedness). If they exist in different basic blocks, bring
35 /// them together by hoisting or replace the common division operation that is
36 /// implicit in the remainder:
37 /// X % Y <--> X - ((X / Y) * Y).
38 ///
39 /// We can largely ignore the normal safety and cost constraints on speculation
40 /// of these ops when we find a matching pair. This is because we are already
41 /// guaranteed that any exceptions and most cost are already incurred by the
42 /// first member of the pair.
43 ///
44 /// Note: This transform could be an oddball enhancement to EarlyCSE, GVN, or
45 /// SimplifyCFG, but it's split off on its own because it's different enough
46 /// that it doesn't quite match the stated objectives of those passes.
47 static bool optimizeDivRem(Function &F, const TargetTransformInfo &TTI,
48  const DominatorTree &DT) {
49  bool Changed = false;
50 
51  // Insert all divide and remainder instructions into maps keyed by their
52  // operands and opcode (signed or unsigned).
54  // Use a MapVector for RemMap so that instructions are moved/inserted in a
55  // deterministic order.
57  for (auto &BB : F) {
58  for (auto &I : BB) {
59  if (I.getOpcode() == Instruction::SDiv)
60  DivMap[DivRemMapKey(true, I.getOperand(0), I.getOperand(1))] = &I;
61  else if (I.getOpcode() == Instruction::UDiv)
62  DivMap[DivRemMapKey(false, I.getOperand(0), I.getOperand(1))] = &I;
63  else if (I.getOpcode() == Instruction::SRem)
64  RemMap[DivRemMapKey(true, I.getOperand(0), I.getOperand(1))] = &I;
65  else if (I.getOpcode() == Instruction::URem)
66  RemMap[DivRemMapKey(false, I.getOperand(0), I.getOperand(1))] = &I;
67  }
68  }
69 
70  // We can iterate over either map because we are only looking for matched
71  // pairs. Choose remainders for efficiency because they are usually even more
72  // rare than division.
73  for (auto &RemPair : RemMap) {
74  // Find the matching division instruction from the division map.
75  Instruction *DivInst = DivMap[RemPair.first];
76  if (!DivInst)
77  continue;
78 
79  // We have a matching pair of div/rem instructions. If one dominates the
80  // other, hoist and/or replace one.
81  NumPairs++;
82  Instruction *RemInst = RemPair.second;
83  bool IsSigned = DivInst->getOpcode() == Instruction::SDiv;
84  bool HasDivRemOp = TTI.hasDivRemOp(DivInst->getType(), IsSigned);
85 
86  // If the target supports div+rem and the instructions are in the same block
87  // already, there's nothing to do. The backend should handle this. If the
88  // target does not support div+rem, then we will decompose the rem.
89  if (HasDivRemOp && RemInst->getParent() == DivInst->getParent())
90  continue;
91 
92  bool DivDominates = DT.dominates(DivInst, RemInst);
93  if (!DivDominates && !DT.dominates(RemInst, DivInst))
94  continue;
95 
96  if (HasDivRemOp) {
97  // The target has a single div/rem operation. Hoist the lower instruction
98  // to make the matched pair visible to the backend.
99  if (DivDominates)
100  RemInst->moveAfter(DivInst);
101  else
102  DivInst->moveAfter(RemInst);
103  NumHoisted++;
104  } else {
105  // The target does not have a single div/rem operation. Decompose the
106  // remainder calculation as:
107  // X % Y --> X - ((X / Y) * Y).
108  Value *X = RemInst->getOperand(0);
109  Value *Y = RemInst->getOperand(1);
110  Instruction *Mul = BinaryOperator::CreateMul(DivInst, Y);
111  Instruction *Sub = BinaryOperator::CreateSub(X, Mul);
112 
113  // If the remainder dominates, then hoist the division up to that block:
114  //
115  // bb1:
116  // %rem = srem %x, %y
117  // bb2:
118  // %div = sdiv %x, %y
119  // -->
120  // bb1:
121  // %div = sdiv %x, %y
122  // %mul = mul %div, %y
123  // %rem = sub %x, %mul
124  //
125  // If the division dominates, it's already in the right place. The mul+sub
126  // will be in a different block because we don't assume that they are
127  // cheap to speculatively execute:
128  //
129  // bb1:
130  // %div = sdiv %x, %y
131  // bb2:
132  // %rem = srem %x, %y
133  // -->
134  // bb1:
135  // %div = sdiv %x, %y
136  // bb2:
137  // %mul = mul %div, %y
138  // %rem = sub %x, %mul
139  //
140  // If the div and rem are in the same block, we do the same transform,
141  // but any code movement would be within the same block.
142 
143  if (!DivDominates)
144  DivInst->moveBefore(RemInst);
145  Mul->insertAfter(RemInst);
146  Sub->insertAfter(Mul);
147 
148  // Now kill the explicit remainder. We have replaced it with:
149  // (sub X, (mul (div X, Y), Y)
150  RemInst->replaceAllUsesWith(Sub);
151  RemInst->eraseFromParent();
152  NumDecomposed++;
153  }
154  Changed = true;
155  }
156 
157  return Changed;
158 }
159 
160 // Pass manager boilerplate below here.
161 
162 namespace {
163 struct DivRemPairsLegacyPass : public FunctionPass {
164  static char ID;
165  DivRemPairsLegacyPass() : FunctionPass(ID) {
167  }
168 
169  void getAnalysisUsage(AnalysisUsage &AU) const override {
172  AU.setPreservesCFG();
176  }
177 
178  bool runOnFunction(Function &F) override {
179  if (skipFunction(F))
180  return false;
181  auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
182  auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
183  return optimizeDivRem(F, TTI, DT);
184  }
185 };
186 }
187 
189 INITIALIZE_PASS_BEGIN(DivRemPairsLegacyPass, "div-rem-pairs",
190  "Hoist/decompose integer division and remainder", false,
191  false)
193 INITIALIZE_PASS_END(DivRemPairsLegacyPass, "div-rem-pairs",
194  "Hoist/decompose integer division and remainder", false,
195  false)
197  return new DivRemPairsLegacyPass();
198 }
199 
204  if (!optimizeDivRem(F, TTI, DT))
205  return PreservedAnalyses::all();
206  // TODO: This pass just hoists/replaces math ops - all analyses are preserved?
208  PA.preserveSet<CFGAnalyses>();
209  PA.preserve<GlobalsAA>();
210  return PA;
211 }
Legacy wrapper pass to provide the GlobalsAAResult object.
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks &#39;this&#39; from the containing basic block and deletes it.
Definition: Instruction.cpp:68
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
AnalysisUsage & addPreserved()
Add the specified Pass class to the set of analyses preserved by this pass.
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
static bool optimizeDivRem(Function &F, const TargetTransformInfo &TTI, const DominatorTree &DT)
Find matching pairs of integer div/rem ops (they have the same numerator, denominator, and signedness).
Definition: DivRemPairs.cpp:47
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:687
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
This is the interface for a simple mod/ref and alias analysis over globals.
Analysis pass providing the TargetTransformInfo.
void initializeDivRemPairsLegacyPassPass(PassRegistry &)
This class implements a map that also provides access to all stored values in a deterministic order...
Definition: MapVector.h:38
STATISTIC(NumFunctions, "Total number of functions")
Analysis pass which computes a DominatorTree.
Definition: Dominators.h:225
F(f)
INITIALIZE_PASS_BEGIN(DivRemPairsLegacyPass, "div-rem-pairs", "Hoist/decompose integer division and remainder", false, false) INITIALIZE_PASS_END(DivRemPairsLegacyPass
static GCMetadataPrinterRegistry::Add< OcamlGCMetadataPrinter > Y("ocaml", "ocaml 3.10-compatible collector")
AnalysisUsage & addRequired()
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:51
div rem Hoist decompose integer division and remainder
virtual void getAnalysisUsage(AnalysisUsage &) const
getAnalysisUsage - This function should be overriden by passes that need analysis information to do t...
Definition: Pass.cpp:92
div rem pairs
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:245
bool hasDivRemOp(Type *DataType, bool IsSigned) const
Return true if the target has a unified operation to calculate division and remainder.
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:126
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:439
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:142
Value * getOperand(unsigned i) const
Definition: User.h:170
static bool runOnFunction(Function &F, bool PostInlining)
Wrapper pass for TargetTransformInfo.
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:153
FunctionPass * createDivRemPairsPass()
Represent the analysis usage information of a pass.
Analysis pass providing a never-invalidated alias analysis result.
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:285
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:159
INITIALIZE_PASS_END(RegBankSelect, DEBUG_TYPE, "Assign register bank of generic virtual registers", false, false) RegBankSelect
This pass provides access to the codegen interfaces that are needed for IR-level transformations.
bool dominates(const Instruction *Def, const Use &U) const
Return true if Def dominates a use in User.
Definition: Dominators.cpp:244
void setPreservesCFG()
This function should be called by the pass, iff they do not:
Definition: Pass.cpp:286
Represents analyses that only rely on functions&#39; control flow.
Definition: PassManager.h:114
PreservedAnalyses run(Function &F, FunctionAnalysisManager &)
void preserveSet()
Mark an analysis set as preserved.
Definition: PassManager.h:189
void insertAfter(Instruction *InsertPos)
Insert an unlinked instruction into a basic block immediately after the specified instruction...
Definition: Instruction.cpp:80
#define I(x, y, z)
Definition: MD5.cpp:58
void preserve()
Mark an analysis as preserved.
Definition: PassManager.h:174
void moveAfter(Instruction *MovePos)
Unlink this instruction from its current basic block and insert it into the basic block that MovePos ...
Definition: Instruction.cpp:91
LLVM Value Representation.
Definition: Value.h:73
void moveBefore(Instruction *MovePos)
Unlink this instruction from its current basic block and insert it into the basic block that MovePos ...
Definition: Instruction.cpp:87
A container for analyses that lazily runs them and caches their results.
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:254
This pass exposes codegen information to IR-level passes.
static BinaryOperator * CreateMul(Value *S1, Value *S2, const Twine &Name, Instruction *InsertBefore, Value *FlagsOp)
const BasicBlock * getParent() const
Definition: Instruction.h:67