GVN.cpp revision f07054d98a5fcd59f3a30853f4b54a74a74986e5
156c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks//===- GVN.cpp - Eliminate redundant values and loads ---------------------===// 256c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks// 356c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks// The LLVM Compiler Infrastructure 456c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks// 556c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks// This file is distributed under the University of Illinois Open Source 656c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks// License. See LICENSE.TXT for details. 756c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks// 856c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks//===----------------------------------------------------------------------===// 956c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks// 107df30109963092559d3760c0661a020f9daf1030The Android Open Source Project// This pass performs global value numbering to eliminate fully redundant 117df30109963092559d3760c0661a020f9daf1030The Android Open Source Project// instructions. It also performs simple dead load elimination. 127df30109963092559d3760c0661a020f9daf1030The Android Open Source Project// 137df30109963092559d3760c0661a020f9daf1030The Android Open Source Project// Note that this pass does the value numbering itself; it does not use the 147df30109963092559d3760c0661a020f9daf1030The Android Open Source Project// ValueNumbering analysis passes. 157df30109963092559d3760c0661a020f9daf1030The Android Open Source Project// 167df30109963092559d3760c0661a020f9daf1030The Android Open Source Project//===----------------------------------------------------------------------===// 177df30109963092559d3760c0661a020f9daf1030The Android Open Source Project 187df30109963092559d3760c0661a020f9daf1030The Android Open Source Project#define DEBUG_TYPE "gvn" 197df30109963092559d3760c0661a020f9daf1030The Android Open Source Project#include "llvm/Transforms/Scalar.h" 207df30109963092559d3760c0661a020f9daf1030The Android Open Source Project#include "llvm/GlobalVariable.h" 217df30109963092559d3760c0661a020f9daf1030The Android Open Source Project#include "llvm/IntrinsicInst.h" 2256c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks#include "llvm/LLVMContext.h" 2356c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks#include "llvm/Analysis/AliasAnalysis.h" 2456c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks#include "llvm/Analysis/ConstantFolding.h" 2556c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks#include "llvm/Analysis/Dominators.h" 2656c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks#include "llvm/Analysis/InstructionSimplify.h" 2756c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks#include "llvm/Analysis/Loads.h" 2856c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks#include "llvm/Analysis/MemoryBuiltins.h" 2956c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks#include "llvm/Analysis/MemoryDependenceAnalysis.h" 3056c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks#include "llvm/Analysis/PHITransAddr.h" 3156c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks#include "llvm/Analysis/ValueTracking.h" 3256c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks#include "llvm/Assembly/Writer.h" 3356c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks#include "llvm/Target/TargetData.h" 3456c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks#include "llvm/Transforms/Utils/BasicBlockUtils.h" 3556c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks#include "llvm/Transforms/Utils/SSAUpdater.h" 3656c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks#include "llvm/ADT/DenseMap.h" 3756c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks#include "llvm/ADT/DepthFirstIterator.h" 3856c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks#include "llvm/ADT/SmallPtrSet.h" 3956c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks#include "llvm/ADT/Statistic.h" 4056c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks#include "llvm/Support/Allocator.h" 4156c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks#include "llvm/Support/CommandLine.h" 4256c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks#include "llvm/Support/Debug.h" 4356c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks#include "llvm/Support/IRBuilder.h" 4456c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparksusing namespace llvm; 4556c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 4656c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave SparksSTATISTIC(NumGVNInstr, "Number of instructions deleted"); 4756c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave SparksSTATISTIC(NumGVNLoad, "Number of loads deleted"); 4856c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave SparksSTATISTIC(NumGVNPRE, "Number of instructions PRE'd"); 4956c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave SparksSTATISTIC(NumGVNBlocks, "Number of blocks merged"); 5056c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave SparksSTATISTIC(NumPRELoad, "Number of loads PRE'd"); 5156c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 5256c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparksstatic cl::opt<bool> EnablePRE("enable-pre", 5356c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks cl::init(true), cl::Hidden); 5456c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparksstatic cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true)); 5556c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 5656c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks//===----------------------------------------------------------------------===// 5756c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks// ValueTable Class 5856c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks//===----------------------------------------------------------------------===// 5956c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 6056c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks/// This class holds the mapping between values and value numbers. It is used 6156c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks/// as an efficient mechanism to determine the expression-wise equivalence of 6256c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks/// two values. 6356c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparksnamespace { 6456c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks struct Expression { 6556c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks uint32_t opcode; 6656c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks const Type* type; 6756c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks SmallVector<uint32_t, 4> varargs; 6856c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 6956c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks Expression() { } 7056c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks Expression(uint32_t o) : opcode(o) { } 7156c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 7256c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks bool operator==(const Expression &other) const { 7356c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks if (opcode != other.opcode) 7456c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks return false; 7556c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks else if (opcode == ~0U || opcode == ~1U) 7656c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks return true; 7756c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks else if (type != other.type) 7856c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks return false; 7956c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks else if (varargs != other.varargs) 8056c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks return false; 8156c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks return true; 8256c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks } 8356c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks }; 8456c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 8556c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks class ValueTable { 8656c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks private: 8756c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks DenseMap<Value*, uint32_t> valueNumbering; 8856c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks DenseMap<Expression, uint32_t> expressionNumbering; 8956c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks AliasAnalysis* AA; 9056c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks MemoryDependenceAnalysis* MD; 9156c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks DominatorTree* DT; 9256c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 9356c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks uint32_t nextValueNumber; 9456c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 9556c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks Expression create_expression(Instruction* I); 9656c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks uint32_t lookup_or_add_call(CallInst* C); 9756c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks public: 9856c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks ValueTable() : nextValueNumber(1) { } 9956c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks uint32_t lookup_or_add(Value *V); 10056c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks uint32_t lookup(Value *V) const; 10156c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks void add(Value *V, uint32_t num); 10256c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks void clear(); 10356c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks void erase(Value *v); 10456c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks void setAliasAnalysis(AliasAnalysis* A) { AA = A; } 10556c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks AliasAnalysis *getAliasAnalysis() const { return AA; } 10656c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks void setMemDep(MemoryDependenceAnalysis* M) { MD = M; } 10756c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks void setDomTree(DominatorTree* D) { DT = D; } 10856c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks uint32_t getNextUnusedValueNumber() { return nextValueNumber; } 10956c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks void verifyRemoved(const Value *) const; 11056c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks }; 11156c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks} 11256c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 11356c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparksnamespace llvm { 11456c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparkstemplate <> struct DenseMapInfo<Expression> { 11556c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks static inline Expression getEmptyKey() { 11656c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks return ~0U; 11756c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks } 11856c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 11956c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks static inline Expression getTombstoneKey() { 12056c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks return ~1U; 12156c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks } 12256c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 12356c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks static unsigned getHashValue(const Expression e) { 12456c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks unsigned hash = e.opcode; 12556c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 12656c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks hash = ((unsigned)((uintptr_t)e.type >> 4) ^ 12756c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks (unsigned)((uintptr_t)e.type >> 9)); 12856c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 12956c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(), 13056c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks E = e.varargs.end(); I != E; ++I) 13156c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks hash = *I + hash * 37; 13256c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 13356c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks return hash; 13456c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks } 13556c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks static bool isEqual(const Expression &LHS, const Expression &RHS) { 13656c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks return LHS == RHS; 13756c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks } 13856c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks}; 13956c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 14056c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks} 14156c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 14256c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks//===----------------------------------------------------------------------===// 14356c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks// ValueTable Internal Functions 14456c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks//===----------------------------------------------------------------------===// 14556c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 14656c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 14756c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave SparksExpression ValueTable::create_expression(Instruction *I) { 14856c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks Expression e; 14956c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks e.type = I->getType(); 15056c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks e.opcode = I->getOpcode(); 15156c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end(); 15256c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks OI != OE; ++OI) 15356c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks e.varargs.push_back(lookup_or_add(*OI)); 15456c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 15556c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks if (CmpInst *C = dyn_cast<CmpInst>(I)) 15656c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks e.opcode = (C->getOpcode() << 8) | C->getPredicate(); 15756c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks else if (ExtractValueInst *E = dyn_cast<ExtractValueInst>(I)) { 15856c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks for (ExtractValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end(); 15956c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks II != IE; ++II) 16056c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks e.varargs.push_back(*II); 16156c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) { 16256c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end(); 16356c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks II != IE; ++II) 16456c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks e.varargs.push_back(*II); 16556c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks } 16656c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 16756c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks return e; 16856c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks} 16956c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 17056c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks//===----------------------------------------------------------------------===// 17156c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks// ValueTable External Functions 17256c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks//===----------------------------------------------------------------------===// 17356c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 17456c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks/// add - Insert a value into the table with a specified value number. 17556c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparksvoid ValueTable::add(Value *V, uint32_t num) { 17656c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks valueNumbering.insert(std::make_pair(V, num)); 17756c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks} 17856c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks 17956c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparksuint32_t ValueTable::lookup_or_add_call(CallInst* C) { 18056c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks if (AA->doesNotAccessMemory(C)) { 18156c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks Expression exp = create_expression(C); 18256c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks uint32_t& e = expressionNumbering[exp]; 18356c99cd2c2c1e6ab038dac5fced5b92ccf11ff6cDave Sparks if (!e) e = nextValueNumber++; 184 valueNumbering[C] = e; 185 return e; 186 } else if (AA->onlyReadsMemory(C)) { 187 Expression exp = create_expression(C); 188 uint32_t& e = expressionNumbering[exp]; 189 if (!e) { 190 e = nextValueNumber++; 191 valueNumbering[C] = e; 192 return e; 193 } 194 if (!MD) { 195 e = nextValueNumber++; 196 valueNumbering[C] = e; 197 return e; 198 } 199 200 MemDepResult local_dep = MD->getDependency(C); 201 202 if (!local_dep.isDef() && !local_dep.isNonLocal()) { 203 valueNumbering[C] = nextValueNumber; 204 return nextValueNumber++; 205 } 206 207 if (local_dep.isDef()) { 208 CallInst* local_cdep = cast<CallInst>(local_dep.getInst()); 209 210 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) { 211 valueNumbering[C] = nextValueNumber; 212 return nextValueNumber++; 213 } 214 215 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 216 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 217 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i)); 218 if (c_vn != cd_vn) { 219 valueNumbering[C] = nextValueNumber; 220 return nextValueNumber++; 221 } 222 } 223 224 uint32_t v = lookup_or_add(local_cdep); 225 valueNumbering[C] = v; 226 return v; 227 } 228 229 // Non-local case. 230 const MemoryDependenceAnalysis::NonLocalDepInfo &deps = 231 MD->getNonLocalCallDependency(CallSite(C)); 232 // FIXME: call/call dependencies for readonly calls should return def, not 233 // clobber! Move the checking logic to MemDep! 234 CallInst* cdep = 0; 235 236 // Check to see if we have a single dominating call instruction that is 237 // identical to C. 238 for (unsigned i = 0, e = deps.size(); i != e; ++i) { 239 const NonLocalDepEntry *I = &deps[i]; 240 // Ignore non-local dependencies. 241 if (I->getResult().isNonLocal()) 242 continue; 243 244 // We don't handle non-depedencies. If we already have a call, reject 245 // instruction dependencies. 246 if (I->getResult().isClobber() || cdep != 0) { 247 cdep = 0; 248 break; 249 } 250 251 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst()); 252 // FIXME: All duplicated with non-local case. 253 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){ 254 cdep = NonLocalDepCall; 255 continue; 256 } 257 258 cdep = 0; 259 break; 260 } 261 262 if (!cdep) { 263 valueNumbering[C] = nextValueNumber; 264 return nextValueNumber++; 265 } 266 267 if (cdep->getNumArgOperands() != C->getNumArgOperands()) { 268 valueNumbering[C] = nextValueNumber; 269 return nextValueNumber++; 270 } 271 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) { 272 uint32_t c_vn = lookup_or_add(C->getArgOperand(i)); 273 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i)); 274 if (c_vn != cd_vn) { 275 valueNumbering[C] = nextValueNumber; 276 return nextValueNumber++; 277 } 278 } 279 280 uint32_t v = lookup_or_add(cdep); 281 valueNumbering[C] = v; 282 return v; 283 284 } else { 285 valueNumbering[C] = nextValueNumber; 286 return nextValueNumber++; 287 } 288} 289 290/// lookup_or_add - Returns the value number for the specified value, assigning 291/// it a new number if it did not have one before. 292uint32_t ValueTable::lookup_or_add(Value *V) { 293 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V); 294 if (VI != valueNumbering.end()) 295 return VI->second; 296 297 if (!isa<Instruction>(V)) { 298 valueNumbering[V] = nextValueNumber; 299 return nextValueNumber++; 300 } 301 302 Instruction* I = cast<Instruction>(V); 303 Expression exp; 304 switch (I->getOpcode()) { 305 case Instruction::Call: 306 return lookup_or_add_call(cast<CallInst>(I)); 307 case Instruction::Add: 308 case Instruction::FAdd: 309 case Instruction::Sub: 310 case Instruction::FSub: 311 case Instruction::Mul: 312 case Instruction::FMul: 313 case Instruction::UDiv: 314 case Instruction::SDiv: 315 case Instruction::FDiv: 316 case Instruction::URem: 317 case Instruction::SRem: 318 case Instruction::FRem: 319 case Instruction::Shl: 320 case Instruction::LShr: 321 case Instruction::AShr: 322 case Instruction::And: 323 case Instruction::Or : 324 case Instruction::Xor: 325 case Instruction::ICmp: 326 case Instruction::FCmp: 327 case Instruction::Trunc: 328 case Instruction::ZExt: 329 case Instruction::SExt: 330 case Instruction::FPToUI: 331 case Instruction::FPToSI: 332 case Instruction::UIToFP: 333 case Instruction::SIToFP: 334 case Instruction::FPTrunc: 335 case Instruction::FPExt: 336 case Instruction::PtrToInt: 337 case Instruction::IntToPtr: 338 case Instruction::BitCast: 339 case Instruction::Select: 340 case Instruction::ExtractElement: 341 case Instruction::InsertElement: 342 case Instruction::ShuffleVector: 343 case Instruction::ExtractValue: 344 case Instruction::InsertValue: 345 case Instruction::GetElementPtr: 346 exp = create_expression(I); 347 break; 348 default: 349 valueNumbering[V] = nextValueNumber; 350 return nextValueNumber++; 351 } 352 353 uint32_t& e = expressionNumbering[exp]; 354 if (!e) e = nextValueNumber++; 355 valueNumbering[V] = e; 356 return e; 357} 358 359/// lookup - Returns the value number of the specified value. Fails if 360/// the value has not yet been numbered. 361uint32_t ValueTable::lookup(Value *V) const { 362 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V); 363 assert(VI != valueNumbering.end() && "Value not numbered?"); 364 return VI->second; 365} 366 367/// clear - Remove all entries from the ValueTable 368void ValueTable::clear() { 369 valueNumbering.clear(); 370 expressionNumbering.clear(); 371 nextValueNumber = 1; 372} 373 374/// erase - Remove a value from the value numbering 375void ValueTable::erase(Value *V) { 376 valueNumbering.erase(V); 377} 378 379/// verifyRemoved - Verify that the value is removed from all internal data 380/// structures. 381void ValueTable::verifyRemoved(const Value *V) const { 382 for (DenseMap<Value*, uint32_t>::const_iterator 383 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) { 384 assert(I->first != V && "Inst still occurs in value numbering map!"); 385 } 386} 387 388//===----------------------------------------------------------------------===// 389// GVN Pass 390//===----------------------------------------------------------------------===// 391 392namespace { 393 394 class GVN : public FunctionPass { 395 bool runOnFunction(Function &F); 396 public: 397 static char ID; // Pass identification, replacement for typeid 398 explicit GVN(bool noloads = false) 399 : FunctionPass(ID), NoLoads(noloads), MD(0) { 400 initializeGVNPass(*PassRegistry::getPassRegistry()); 401 } 402 403 private: 404 bool NoLoads; 405 MemoryDependenceAnalysis *MD; 406 DominatorTree *DT; 407 const TargetData* TD; 408 409 ValueTable VN; 410 411 /// LeaderTable - A mapping from value numbers to lists of Value*'s that 412 /// have that value number. Use findLeader to query it. 413 struct LeaderTableEntry { 414 Value *Val; 415 BasicBlock *BB; 416 LeaderTableEntry *Next; 417 }; 418 DenseMap<uint32_t, LeaderTableEntry> LeaderTable; 419 BumpPtrAllocator TableAllocator; 420 421 SmallVector<Instruction*, 8> InstrsToErase; 422 423 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for 424 /// its value number. 425 void addToLeaderTable(uint32_t N, Value *V, BasicBlock *BB) { 426 LeaderTableEntry& Curr = LeaderTable[N]; 427 if (!Curr.Val) { 428 Curr.Val = V; 429 Curr.BB = BB; 430 return; 431 } 432 433 LeaderTableEntry* Node = TableAllocator.Allocate<LeaderTableEntry>(); 434 Node->Val = V; 435 Node->BB = BB; 436 Node->Next = Curr.Next; 437 Curr.Next = Node; 438 } 439 440 /// removeFromLeaderTable - Scan the list of values corresponding to a given 441 /// value number, and remove the given value if encountered. 442 void removeFromLeaderTable(uint32_t N, Value *V, BasicBlock *BB) { 443 LeaderTableEntry* Prev = 0; 444 LeaderTableEntry* Curr = &LeaderTable[N]; 445 446 while (Curr->Val != V || Curr->BB != BB) { 447 Prev = Curr; 448 Curr = Curr->Next; 449 } 450 451 if (Prev) { 452 Prev->Next = Curr->Next; 453 } else { 454 if (!Curr->Next) { 455 Curr->Val = 0; 456 Curr->BB = 0; 457 } else { 458 LeaderTableEntry* Next = Curr->Next; 459 Curr->Val = Next->Val; 460 Curr->BB = Next->BB; 461 Curr->Next = Next->Next; 462 } 463 } 464 } 465 466 // List of critical edges to be split between iterations. 467 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit; 468 469 // This transformation requires dominator postdominator info 470 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 471 AU.addRequired<DominatorTree>(); 472 if (!NoLoads) 473 AU.addRequired<MemoryDependenceAnalysis>(); 474 AU.addRequired<AliasAnalysis>(); 475 476 AU.addPreserved<DominatorTree>(); 477 AU.addPreserved<AliasAnalysis>(); 478 } 479 480 // Helper fuctions 481 // FIXME: eliminate or document these better 482 bool processLoad(LoadInst *L); 483 bool processInstruction(Instruction *I); 484 bool processNonLocalLoad(LoadInst *L); 485 bool processBlock(BasicBlock *BB); 486 void dump(DenseMap<uint32_t, Value*> &d); 487 bool iterateOnFunction(Function &F); 488 bool performPRE(Function &F); 489 Value *findLeader(BasicBlock *BB, uint32_t num); 490 void cleanupGlobalSets(); 491 void verifyRemoved(const Instruction *I) const; 492 bool splitCriticalEdges(); 493 }; 494 495 char GVN::ID = 0; 496} 497 498// createGVNPass - The public interface to this file... 499FunctionPass *llvm::createGVNPass(bool NoLoads) { 500 return new GVN(NoLoads); 501} 502 503INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false) 504INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis) 505INITIALIZE_PASS_DEPENDENCY(DominatorTree) 506INITIALIZE_AG_DEPENDENCY(AliasAnalysis) 507INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false) 508 509void GVN::dump(DenseMap<uint32_t, Value*>& d) { 510 errs() << "{\n"; 511 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(), 512 E = d.end(); I != E; ++I) { 513 errs() << I->first << "\n"; 514 I->second->dump(); 515 } 516 errs() << "}\n"; 517} 518 519/// IsValueFullyAvailableInBlock - Return true if we can prove that the value 520/// we're analyzing is fully available in the specified block. As we go, keep 521/// track of which blocks we know are fully alive in FullyAvailableBlocks. This 522/// map is actually a tri-state map with the following values: 523/// 0) we know the block *is not* fully available. 524/// 1) we know the block *is* fully available. 525/// 2) we do not know whether the block is fully available or not, but we are 526/// currently speculating that it will be. 527/// 3) we are speculating for this block and have used that to speculate for 528/// other blocks. 529static bool IsValueFullyAvailableInBlock(BasicBlock *BB, 530 DenseMap<BasicBlock*, char> &FullyAvailableBlocks) { 531 // Optimistically assume that the block is fully available and check to see 532 // if we already know about this block in one lookup. 533 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV = 534 FullyAvailableBlocks.insert(std::make_pair(BB, 2)); 535 536 // If the entry already existed for this block, return the precomputed value. 537 if (!IV.second) { 538 // If this is a speculative "available" value, mark it as being used for 539 // speculation of other blocks. 540 if (IV.first->second == 2) 541 IV.first->second = 3; 542 return IV.first->second != 0; 543 } 544 545 // Otherwise, see if it is fully available in all predecessors. 546 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 547 548 // If this block has no predecessors, it isn't live-in here. 549 if (PI == PE) 550 goto SpeculationFailure; 551 552 for (; PI != PE; ++PI) 553 // If the value isn't fully available in one of our predecessors, then it 554 // isn't fully available in this block either. Undo our previous 555 // optimistic assumption and bail out. 556 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks)) 557 goto SpeculationFailure; 558 559 return true; 560 561// SpeculationFailure - If we get here, we found out that this is not, after 562// all, a fully-available block. We have a problem if we speculated on this and 563// used the speculation to mark other blocks as available. 564SpeculationFailure: 565 char &BBVal = FullyAvailableBlocks[BB]; 566 567 // If we didn't speculate on this, just return with it set to false. 568 if (BBVal == 2) { 569 BBVal = 0; 570 return false; 571 } 572 573 // If we did speculate on this value, we could have blocks set to 1 that are 574 // incorrect. Walk the (transitive) successors of this block and mark them as 575 // 0 if set to one. 576 SmallVector<BasicBlock*, 32> BBWorklist; 577 BBWorklist.push_back(BB); 578 579 do { 580 BasicBlock *Entry = BBWorklist.pop_back_val(); 581 // Note that this sets blocks to 0 (unavailable) if they happen to not 582 // already be in FullyAvailableBlocks. This is safe. 583 char &EntryVal = FullyAvailableBlocks[Entry]; 584 if (EntryVal == 0) continue; // Already unavailable. 585 586 // Mark as unavailable. 587 EntryVal = 0; 588 589 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I) 590 BBWorklist.push_back(*I); 591 } while (!BBWorklist.empty()); 592 593 return false; 594} 595 596 597/// CanCoerceMustAliasedValueToLoad - Return true if 598/// CoerceAvailableValueToLoadType will succeed. 599static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal, 600 const Type *LoadTy, 601 const TargetData &TD) { 602 // If the loaded or stored value is an first class array or struct, don't try 603 // to transform them. We need to be able to bitcast to integer. 604 if (LoadTy->isStructTy() || LoadTy->isArrayTy() || 605 StoredVal->getType()->isStructTy() || 606 StoredVal->getType()->isArrayTy()) 607 return false; 608 609 // The store has to be at least as big as the load. 610 if (TD.getTypeSizeInBits(StoredVal->getType()) < 611 TD.getTypeSizeInBits(LoadTy)) 612 return false; 613 614 return true; 615} 616 617 618/// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and 619/// then a load from a must-aliased pointer of a different type, try to coerce 620/// the stored value. LoadedTy is the type of the load we want to replace and 621/// InsertPt is the place to insert new instructions. 622/// 623/// If we can't do it, return null. 624static Value *CoerceAvailableValueToLoadType(Value *StoredVal, 625 const Type *LoadedTy, 626 Instruction *InsertPt, 627 const TargetData &TD) { 628 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD)) 629 return 0; 630 631 // If this is already the right type, just return it. 632 const Type *StoredValTy = StoredVal->getType(); 633 634 uint64_t StoreSize = TD.getTypeStoreSizeInBits(StoredValTy); 635 uint64_t LoadSize = TD.getTypeStoreSizeInBits(LoadedTy); 636 637 // If the store and reload are the same size, we can always reuse it. 638 if (StoreSize == LoadSize) { 639 // Pointer to Pointer -> use bitcast. 640 if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy()) 641 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt); 642 643 // Convert source pointers to integers, which can be bitcast. 644 if (StoredValTy->isPointerTy()) { 645 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 646 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 647 } 648 649 const Type *TypeToCastTo = LoadedTy; 650 if (TypeToCastTo->isPointerTy()) 651 TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext()); 652 653 if (StoredValTy != TypeToCastTo) 654 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt); 655 656 // Cast to pointer if the load needs a pointer type. 657 if (LoadedTy->isPointerTy()) 658 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt); 659 660 return StoredVal; 661 } 662 663 // If the loaded value is smaller than the available value, then we can 664 // extract out a piece from it. If the available value is too small, then we 665 // can't do anything. 666 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail"); 667 668 // Convert source pointers to integers, which can be manipulated. 669 if (StoredValTy->isPointerTy()) { 670 StoredValTy = TD.getIntPtrType(StoredValTy->getContext()); 671 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt); 672 } 673 674 // Convert vectors and fp to integer, which can be manipulated. 675 if (!StoredValTy->isIntegerTy()) { 676 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize); 677 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt); 678 } 679 680 // If this is a big-endian system, we need to shift the value down to the low 681 // bits so that a truncate will work. 682 if (TD.isBigEndian()) { 683 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize); 684 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt); 685 } 686 687 // Truncate the integer to the right size now. 688 const Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize); 689 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt); 690 691 if (LoadedTy == NewIntTy) 692 return StoredVal; 693 694 // If the result is a pointer, inttoptr. 695 if (LoadedTy->isPointerTy()) 696 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt); 697 698 // Otherwise, bitcast. 699 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt); 700} 701 702/// AnalyzeLoadFromClobberingWrite - This function is called when we have a 703/// memdep query of a load that ends up being a clobbering memory write (store, 704/// memset, memcpy, memmove). This means that the write *may* provide bits used 705/// by the load but we can't be sure because the pointers don't mustalias. 706/// 707/// Check this case to see if there is anything more we can do before we give 708/// up. This returns -1 if we have to give up, or a byte number in the stored 709/// value of the piece that feeds the load. 710static int AnalyzeLoadFromClobberingWrite(const Type *LoadTy, Value *LoadPtr, 711 Value *WritePtr, 712 uint64_t WriteSizeInBits, 713 const TargetData &TD) { 714 // If the loaded or stored value is an first class array or struct, don't try 715 // to transform them. We need to be able to bitcast to integer. 716 if (LoadTy->isStructTy() || LoadTy->isArrayTy()) 717 return -1; 718 719 int64_t StoreOffset = 0, LoadOffset = 0; 720 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr, StoreOffset,TD); 721 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, TD); 722 if (StoreBase != LoadBase) 723 return -1; 724 725 // If the load and store are to the exact same address, they should have been 726 // a must alias. AA must have gotten confused. 727 // FIXME: Study to see if/when this happens. One case is forwarding a memset 728 // to a load from the base of the memset. 729#if 0 730 if (LoadOffset == StoreOffset) { 731 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n" 732 << "Base = " << *StoreBase << "\n" 733 << "Store Ptr = " << *WritePtr << "\n" 734 << "Store Offs = " << StoreOffset << "\n" 735 << "Load Ptr = " << *LoadPtr << "\n"; 736 abort(); 737 } 738#endif 739 740 // If the load and store don't overlap at all, the store doesn't provide 741 // anything to the load. In this case, they really don't alias at all, AA 742 // must have gotten confused. 743 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy); 744 745 if ((WriteSizeInBits & 7) | (LoadSize & 7)) 746 return -1; 747 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes. 748 LoadSize >>= 3; 749 750 751 bool isAAFailure = false; 752 if (StoreOffset < LoadOffset) 753 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset; 754 else 755 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset; 756 757 if (isAAFailure) { 758#if 0 759 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n" 760 << "Base = " << *StoreBase << "\n" 761 << "Store Ptr = " << *WritePtr << "\n" 762 << "Store Offs = " << StoreOffset << "\n" 763 << "Load Ptr = " << *LoadPtr << "\n"; 764 abort(); 765#endif 766 return -1; 767 } 768 769 // If the Load isn't completely contained within the stored bits, we don't 770 // have all the bits to feed it. We could do something crazy in the future 771 // (issue a smaller load then merge the bits in) but this seems unlikely to be 772 // valuable. 773 if (StoreOffset > LoadOffset || 774 StoreOffset+StoreSize < LoadOffset+LoadSize) 775 return -1; 776 777 // Okay, we can do this transformation. Return the number of bytes into the 778 // store that the load is. 779 return LoadOffset-StoreOffset; 780} 781 782/// AnalyzeLoadFromClobberingStore - This function is called when we have a 783/// memdep query of a load that ends up being a clobbering store. 784static int AnalyzeLoadFromClobberingStore(const Type *LoadTy, Value *LoadPtr, 785 StoreInst *DepSI, 786 const TargetData &TD) { 787 // Cannot handle reading from store of first-class aggregate yet. 788 if (DepSI->getValueOperand()->getType()->isStructTy() || 789 DepSI->getValueOperand()->getType()->isArrayTy()) 790 return -1; 791 792 Value *StorePtr = DepSI->getPointerOperand(); 793 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType()); 794 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 795 StorePtr, StoreSize, TD); 796} 797 798/// AnalyzeLoadFromClobberingLoad - This function is called when we have a 799/// memdep query of a load that ends up being clobbered by another load. See if 800/// the other load can feed into the second load. 801static int AnalyzeLoadFromClobberingLoad(const Type *LoadTy, Value *LoadPtr, 802 LoadInst *DepLI, const TargetData &TD){ 803 // Cannot handle reading from store of first-class aggregate yet. 804 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy()) 805 return -1; 806 807 Value *DepPtr = DepLI->getPointerOperand(); 808 uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType()); 809 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD); 810 if (R != -1) return R; 811 812 // If we have a load/load clobber an DepLI can be widened to cover this load, 813 // then we should widen it! 814 int64_t LoadOffs = 0; 815 const Value *LoadBase = 816 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, TD); 817 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 818 819 unsigned Size = MemoryDependenceAnalysis:: 820 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, TD); 821 if (Size == 0) return -1; 822 823 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, TD); 824} 825 826 827 828static int AnalyzeLoadFromClobberingMemInst(const Type *LoadTy, Value *LoadPtr, 829 MemIntrinsic *MI, 830 const TargetData &TD) { 831 // If the mem operation is a non-constant size, we can't handle it. 832 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength()); 833 if (SizeCst == 0) return -1; 834 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8; 835 836 // If this is memset, we just need to see if the offset is valid in the size 837 // of the memset.. 838 if (MI->getIntrinsicID() == Intrinsic::memset) 839 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(), 840 MemSizeInBits, TD); 841 842 // If we have a memcpy/memmove, the only case we can handle is if this is a 843 // copy from constant memory. In that case, we can read directly from the 844 // constant memory. 845 MemTransferInst *MTI = cast<MemTransferInst>(MI); 846 847 Constant *Src = dyn_cast<Constant>(MTI->getSource()); 848 if (Src == 0) return -1; 849 850 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD)); 851 if (GV == 0 || !GV->isConstant()) return -1; 852 853 // See if the access is within the bounds of the transfer. 854 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, 855 MI->getDest(), MemSizeInBits, TD); 856 if (Offset == -1) 857 return Offset; 858 859 // Otherwise, see if we can constant fold a load from the constant with the 860 // offset applied as appropriate. 861 Src = ConstantExpr::getBitCast(Src, 862 llvm::Type::getInt8PtrTy(Src->getContext())); 863 Constant *OffsetCst = 864 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 865 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1); 866 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 867 if (ConstantFoldLoadFromConstPtr(Src, &TD)) 868 return Offset; 869 return -1; 870} 871 872 873/// GetStoreValueForLoad - This function is called when we have a 874/// memdep query of a load that ends up being a clobbering store. This means 875/// that the store provides bits used by the load but we the pointers don't 876/// mustalias. Check this case to see if there is anything more we can do 877/// before we give up. 878static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset, 879 const Type *LoadTy, 880 Instruction *InsertPt, const TargetData &TD){ 881 LLVMContext &Ctx = SrcVal->getType()->getContext(); 882 883 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8; 884 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8; 885 886 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 887 888 // Compute which bits of the stored value are being used by the load. Convert 889 // to an integer type to start with. 890 if (SrcVal->getType()->isPointerTy()) 891 SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx), "tmp"); 892 if (!SrcVal->getType()->isIntegerTy()) 893 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8), 894 "tmp"); 895 896 // Shift the bits to the least significant depending on endianness. 897 unsigned ShiftAmt; 898 if (TD.isLittleEndian()) 899 ShiftAmt = Offset*8; 900 else 901 ShiftAmt = (StoreSize-LoadSize-Offset)*8; 902 903 if (ShiftAmt) 904 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt, "tmp"); 905 906 if (LoadSize != StoreSize) 907 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8), 908 "tmp"); 909 910 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD); 911} 912 913/// GetStoreValueForLoad - This function is called when we have a 914/// memdep query of a load that ends up being a clobbering load. This means 915/// that the load *may* provide bits used by the load but we can't be sure 916/// because the pointers don't mustalias. Check this case to see if there is 917/// anything more we can do before we give up. 918static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset, 919 const Type *LoadTy, 920 Instruction *InsertPt, const TargetData &TD, 921 MemoryDependenceAnalysis &MD) { 922 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to 923 // widen SrcVal out to a larger load. 924 unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType()); 925 unsigned LoadSize = TD.getTypeStoreSize(LoadTy); 926 if (Offset+LoadSize > SrcValSize) { 927 assert(!SrcVal->isVolatile() && "Cannot widen volatile load!"); 928 assert(isa<IntegerType>(SrcVal->getType())&&"Can't widen non-integer load"); 929 // If we have a load/load clobber an DepLI can be widened to cover this 930 // load, then we should widen it to the next power of 2 size big enough! 931 unsigned NewLoadSize = Offset+LoadSize; 932 if (!isPowerOf2_32(NewLoadSize)) 933 NewLoadSize = NextPowerOf2(NewLoadSize); 934 935 Value *PtrVal = SrcVal->getPointerOperand(); 936 937 IRBuilder<> Builder(SrcVal->getParent(), SrcVal); 938 const Type *DestPTy = 939 IntegerType::get(LoadTy->getContext(), NewLoadSize*8); 940 DestPTy = PointerType::get(DestPTy, 941 cast<PointerType>(PtrVal->getType())->getAddressSpace()); 942 943 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy); 944 LoadInst *NewLoad = Builder.CreateLoad(PtrVal); 945 NewLoad->takeName(SrcVal); 946 NewLoad->setAlignment(SrcVal->getAlignment()); 947 948 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n"); 949 DEBUG(dbgs() << "TO: " << *NewLoad << "\n"); 950 951 // Replace uses of the original load with the wider load. On a big endian 952 // system, we need to shift down to get the relevant bits. 953 Value *RV = NewLoad; 954 if (TD.isBigEndian()) 955 RV = Builder.CreateLShr(RV, 956 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits()); 957 RV = Builder.CreateTrunc(RV, SrcVal->getType()); 958 SrcVal->replaceAllUsesWith(RV); 959 MD.removeInstruction(SrcVal); 960 SrcVal = NewLoad; 961 } 962 963 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, TD); 964} 965 966 967/// GetMemInstValueForLoad - This function is called when we have a 968/// memdep query of a load that ends up being a clobbering mem intrinsic. 969static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset, 970 const Type *LoadTy, Instruction *InsertPt, 971 const TargetData &TD){ 972 LLVMContext &Ctx = LoadTy->getContext(); 973 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8; 974 975 IRBuilder<> Builder(InsertPt->getParent(), InsertPt); 976 977 // We know that this method is only called when the mem transfer fully 978 // provides the bits for the load. 979 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) { 980 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and 981 // independently of what the offset is. 982 Value *Val = MSI->getValue(); 983 if (LoadSize != 1) 984 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8)); 985 986 Value *OneElt = Val; 987 988 // Splat the value out to the right number of bits. 989 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) { 990 // If we can double the number of bytes set, do it. 991 if (NumBytesSet*2 <= LoadSize) { 992 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8); 993 Val = Builder.CreateOr(Val, ShVal); 994 NumBytesSet <<= 1; 995 continue; 996 } 997 998 // Otherwise insert one byte at a time. 999 Value *ShVal = Builder.CreateShl(Val, 1*8); 1000 Val = Builder.CreateOr(OneElt, ShVal); 1001 ++NumBytesSet; 1002 } 1003 1004 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD); 1005 } 1006 1007 // Otherwise, this is a memcpy/memmove from a constant global. 1008 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst); 1009 Constant *Src = cast<Constant>(MTI->getSource()); 1010 1011 // Otherwise, see if we can constant fold a load from the constant with the 1012 // offset applied as appropriate. 1013 Src = ConstantExpr::getBitCast(Src, 1014 llvm::Type::getInt8PtrTy(Src->getContext())); 1015 Constant *OffsetCst = 1016 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset); 1017 Src = ConstantExpr::getGetElementPtr(Src, &OffsetCst, 1); 1018 Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy)); 1019 return ConstantFoldLoadFromConstPtr(Src, &TD); 1020} 1021 1022namespace { 1023 1024struct AvailableValueInBlock { 1025 /// BB - The basic block in question. 1026 BasicBlock *BB; 1027 enum ValType { 1028 SimpleVal, // A simple offsetted value that is accessed. 1029 LoadVal, // A value produced by a load. 1030 MemIntrin // A memory intrinsic which is loaded from. 1031 }; 1032 1033 /// V - The value that is live out of the block. 1034 PointerIntPair<Value *, 2, ValType> Val; 1035 1036 /// Offset - The byte offset in Val that is interesting for the load query. 1037 unsigned Offset; 1038 1039 static AvailableValueInBlock get(BasicBlock *BB, Value *V, 1040 unsigned Offset = 0) { 1041 AvailableValueInBlock Res; 1042 Res.BB = BB; 1043 Res.Val.setPointer(V); 1044 Res.Val.setInt(SimpleVal); 1045 Res.Offset = Offset; 1046 return Res; 1047 } 1048 1049 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI, 1050 unsigned Offset = 0) { 1051 AvailableValueInBlock Res; 1052 Res.BB = BB; 1053 Res.Val.setPointer(MI); 1054 Res.Val.setInt(MemIntrin); 1055 Res.Offset = Offset; 1056 return Res; 1057 } 1058 1059 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI, 1060 unsigned Offset = 0) { 1061 AvailableValueInBlock Res; 1062 Res.BB = BB; 1063 Res.Val.setPointer(LI); 1064 Res.Val.setInt(LoadVal); 1065 Res.Offset = Offset; 1066 return Res; 1067 } 1068 1069 bool isSimpleValue() const { return Val.getInt() == SimpleVal; } 1070 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; } 1071 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; } 1072 1073 Value *getSimpleValue() const { 1074 assert(isSimpleValue() && "Wrong accessor"); 1075 return Val.getPointer(); 1076 } 1077 1078 LoadInst *getCoercedLoadValue() const { 1079 assert(isCoercedLoadValue() && "Wrong accessor"); 1080 return cast<LoadInst>(Val.getPointer()); 1081 } 1082 1083 MemIntrinsic *getMemIntrinValue() const { 1084 assert(isMemIntrinValue() && "Wrong accessor"); 1085 return cast<MemIntrinsic>(Val.getPointer()); 1086 } 1087 1088 /// MaterializeAdjustedValue - Emit code into this block to adjust the value 1089 /// defined here to the specified type. This handles various coercion cases. 1090 Value *MaterializeAdjustedValue(const Type *LoadTy, 1091 const TargetData *TD, 1092 MemoryDependenceAnalysis &MD) const { 1093 Value *Res; 1094 if (isSimpleValue()) { 1095 Res = getSimpleValue(); 1096 if (Res->getType() != LoadTy) { 1097 assert(TD && "Need target data to handle type mismatch case"); 1098 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), 1099 *TD); 1100 1101 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " " 1102 << *getSimpleValue() << '\n' 1103 << *Res << '\n' << "\n\n\n"); 1104 } 1105 } else if (isCoercedLoadValue()) { 1106 LoadInst *Load = getCoercedLoadValue(); 1107 if (Load->getType() == LoadTy && Offset == 0) { 1108 Res = Load; 1109 } else { 1110 assert(TD && "Need target data to handle type mismatch case"); 1111 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(), 1112 *TD, MD); 1113 1114 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " " 1115 << *getCoercedLoadValue() << '\n' 1116 << *Res << '\n' << "\n\n\n"); 1117 } 1118 } else { 1119 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, 1120 LoadTy, BB->getTerminator(), *TD); 1121 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset 1122 << " " << *getMemIntrinValue() << '\n' 1123 << *Res << '\n' << "\n\n\n"); 1124 } 1125 return Res; 1126 } 1127}; 1128 1129} // end anonymous namespace 1130 1131/// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock, 1132/// construct SSA form, allowing us to eliminate LI. This returns the value 1133/// that should be used at LI's definition site. 1134static Value *ConstructSSAForLoadSet(LoadInst *LI, 1135 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock, 1136 const TargetData *TD, 1137 const DominatorTree &DT, 1138 AliasAnalysis *AA, 1139 MemoryDependenceAnalysis &MD) { 1140 // Check for the fully redundant, dominating load case. In this case, we can 1141 // just use the dominating value directly. 1142 if (ValuesPerBlock.size() == 1 && 1143 DT.properlyDominates(ValuesPerBlock[0].BB, LI->getParent())) 1144 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), TD, MD); 1145 1146 // Otherwise, we have to construct SSA form. 1147 SmallVector<PHINode*, 8> NewPHIs; 1148 SSAUpdater SSAUpdate(&NewPHIs); 1149 SSAUpdate.Initialize(LI->getType(), LI->getName()); 1150 1151 const Type *LoadTy = LI->getType(); 1152 1153 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1154 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1155 BasicBlock *BB = AV.BB; 1156 1157 if (SSAUpdate.HasValueForBlock(BB)) 1158 continue; 1159 1160 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, TD,MD)); 1161 } 1162 1163 // Perform PHI construction. 1164 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent()); 1165 1166 // If new PHI nodes were created, notify alias analysis. 1167 if (V->getType()->isPointerTy()) 1168 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) 1169 AA->copyValue(LI, NewPHIs[i]); 1170 1171 // Now that we've copied information to the new PHIs, scan through 1172 // them again and inform alias analysis that we've added potentially 1173 // escaping uses to any values that are operands to these PHIs. 1174 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) { 1175 PHINode *P = NewPHIs[i]; 1176 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) 1177 AA->addEscapingUse(P->getOperandUse(2*ii)); 1178 } 1179 1180 return V; 1181} 1182 1183static bool isLifetimeStart(const Instruction *Inst) { 1184 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst)) 1185 return II->getIntrinsicID() == Intrinsic::lifetime_start; 1186 return false; 1187} 1188 1189/// processNonLocalLoad - Attempt to eliminate a load whose dependencies are 1190/// non-local by performing PHI construction. 1191bool GVN::processNonLocalLoad(LoadInst *LI) { 1192 // Find the non-local dependencies of the load. 1193 SmallVector<NonLocalDepResult, 64> Deps; 1194 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI); 1195 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps); 1196 //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: " 1197 // << Deps.size() << *LI << '\n'); 1198 1199 // If we had to process more than one hundred blocks to find the 1200 // dependencies, this load isn't worth worrying about. Optimizing 1201 // it will be too expensive. 1202 if (Deps.size() > 100) 1203 return false; 1204 1205 // If we had a phi translation failure, we'll have a single entry which is a 1206 // clobber in the current block. Reject this early. 1207 if (Deps.size() == 1 && Deps[0].getResult().isClobber() && 1208 Deps[0].getResult().getInst()->getParent() == LI->getParent()) { 1209 DEBUG( 1210 dbgs() << "GVN: non-local load "; 1211 WriteAsOperand(dbgs(), LI); 1212 dbgs() << " is clobbered by " << *Deps[0].getResult().getInst() << '\n'; 1213 ); 1214 return false; 1215 } 1216 1217 // Filter out useless results (non-locals, etc). Keep track of the blocks 1218 // where we have a value available in repl, also keep track of whether we see 1219 // dependencies that produce an unknown value for the load (such as a call 1220 // that could potentially clobber the load). 1221 SmallVector<AvailableValueInBlock, 16> ValuesPerBlock; 1222 SmallVector<BasicBlock*, 16> UnavailableBlocks; 1223 1224 for (unsigned i = 0, e = Deps.size(); i != e; ++i) { 1225 BasicBlock *DepBB = Deps[i].getBB(); 1226 MemDepResult DepInfo = Deps[i].getResult(); 1227 1228 if (DepInfo.isClobber()) { 1229 // The address being loaded in this non-local block may not be the same as 1230 // the pointer operand of the load if PHI translation occurs. Make sure 1231 // to consider the right address. 1232 Value *Address = Deps[i].getAddress(); 1233 1234 // If the dependence is to a store that writes to a superset of the bits 1235 // read by the load, we can extract the bits we need for the load from the 1236 // stored value. 1237 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) { 1238 if (TD && Address) { 1239 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address, 1240 DepSI, *TD); 1241 if (Offset != -1) { 1242 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1243 DepSI->getValueOperand(), 1244 Offset)); 1245 continue; 1246 } 1247 } 1248 } 1249 1250 // Check to see if we have something like this: 1251 // load i32* P 1252 // load i8* (P+1) 1253 // if we have this, replace the later with an extraction from the former. 1254 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) { 1255 // If this is a clobber and L is the first instruction in its block, then 1256 // we have the first instruction in the entry block. 1257 if (DepLI != LI && Address && TD) { 1258 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(), 1259 LI->getPointerOperand(), 1260 DepLI, *TD); 1261 1262 if (Offset != -1) { 1263 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI, 1264 Offset)); 1265 continue; 1266 } 1267 } 1268 } 1269 1270 // If the clobbering value is a memset/memcpy/memmove, see if we can 1271 // forward a value on from it. 1272 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) { 1273 if (TD && Address) { 1274 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address, 1275 DepMI, *TD); 1276 if (Offset != -1) { 1277 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI, 1278 Offset)); 1279 continue; 1280 } 1281 } 1282 } 1283 1284 UnavailableBlocks.push_back(DepBB); 1285 continue; 1286 } 1287 1288 Instruction *DepInst = DepInfo.getInst(); 1289 1290 // Loading the allocation -> undef. 1291 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) || 1292 // Loading immediately after lifetime begin -> undef. 1293 isLifetimeStart(DepInst)) { 1294 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1295 UndefValue::get(LI->getType()))); 1296 continue; 1297 } 1298 1299 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) { 1300 // Reject loads and stores that are to the same address but are of 1301 // different types if we have to. 1302 if (S->getValueOperand()->getType() != LI->getType()) { 1303 // If the stored value is larger or equal to the loaded value, we can 1304 // reuse it. 1305 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(), 1306 LI->getType(), *TD)) { 1307 UnavailableBlocks.push_back(DepBB); 1308 continue; 1309 } 1310 } 1311 1312 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB, 1313 S->getValueOperand())); 1314 continue; 1315 } 1316 1317 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) { 1318 // If the types mismatch and we can't handle it, reject reuse of the load. 1319 if (LD->getType() != LI->getType()) { 1320 // If the stored value is larger or equal to the loaded value, we can 1321 // reuse it. 1322 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){ 1323 UnavailableBlocks.push_back(DepBB); 1324 continue; 1325 } 1326 } 1327 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD)); 1328 continue; 1329 } 1330 1331 UnavailableBlocks.push_back(DepBB); 1332 continue; 1333 } 1334 1335 // If we have no predecessors that produce a known value for this load, exit 1336 // early. 1337 if (ValuesPerBlock.empty()) return false; 1338 1339 // If all of the instructions we depend on produce a known value for this 1340 // load, then it is fully redundant and we can use PHI insertion to compute 1341 // its value. Insert PHIs and remove the fully redundant value now. 1342 if (UnavailableBlocks.empty()) { 1343 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n'); 1344 1345 // Perform PHI construction. 1346 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT, 1347 VN.getAliasAnalysis(), *MD); 1348 LI->replaceAllUsesWith(V); 1349 1350 if (isa<PHINode>(V)) 1351 V->takeName(LI); 1352 if (V->getType()->isPointerTy()) 1353 MD->invalidateCachedPointerInfo(V); 1354 VN.erase(LI); 1355 InstrsToErase.push_back(LI); 1356 ++NumGVNLoad; 1357 return true; 1358 } 1359 1360 if (!EnablePRE || !EnableLoadPRE) 1361 return false; 1362 1363 // Okay, we have *some* definitions of the value. This means that the value 1364 // is available in some of our (transitive) predecessors. Lets think about 1365 // doing PRE of this load. This will involve inserting a new load into the 1366 // predecessor when it's not available. We could do this in general, but 1367 // prefer to not increase code size. As such, we only do this when we know 1368 // that we only have to insert *one* load (which means we're basically moving 1369 // the load, not inserting a new one). 1370 1371 SmallPtrSet<BasicBlock *, 4> Blockers; 1372 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1373 Blockers.insert(UnavailableBlocks[i]); 1374 1375 // Lets find first basic block with more than one predecessor. Walk backwards 1376 // through predecessors if needed. 1377 BasicBlock *LoadBB = LI->getParent(); 1378 BasicBlock *TmpBB = LoadBB; 1379 1380 bool isSinglePred = false; 1381 bool allSingleSucc = true; 1382 while (TmpBB->getSinglePredecessor()) { 1383 isSinglePred = true; 1384 TmpBB = TmpBB->getSinglePredecessor(); 1385 if (TmpBB == LoadBB) // Infinite (unreachable) loop. 1386 return false; 1387 if (Blockers.count(TmpBB)) 1388 return false; 1389 1390 // If any of these blocks has more than one successor (i.e. if the edge we 1391 // just traversed was critical), then there are other paths through this 1392 // block along which the load may not be anticipated. Hoisting the load 1393 // above this block would be adding the load to execution paths along 1394 // which it was not previously executed. 1395 if (TmpBB->getTerminator()->getNumSuccessors() != 1) 1396 return false; 1397 } 1398 1399 assert(TmpBB); 1400 LoadBB = TmpBB; 1401 1402 // FIXME: It is extremely unclear what this loop is doing, other than 1403 // artificially restricting loadpre. 1404 if (isSinglePred) { 1405 bool isHot = false; 1406 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) { 1407 const AvailableValueInBlock &AV = ValuesPerBlock[i]; 1408 if (AV.isSimpleValue()) 1409 // "Hot" Instruction is in some loop (because it dominates its dep. 1410 // instruction). 1411 if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue())) 1412 if (DT->dominates(LI, I)) { 1413 isHot = true; 1414 break; 1415 } 1416 } 1417 1418 // We are interested only in "hot" instructions. We don't want to do any 1419 // mis-optimizations here. 1420 if (!isHot) 1421 return false; 1422 } 1423 1424 // Check to see how many predecessors have the loaded value fully 1425 // available. 1426 DenseMap<BasicBlock*, Value*> PredLoads; 1427 DenseMap<BasicBlock*, char> FullyAvailableBlocks; 1428 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) 1429 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true; 1430 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i) 1431 FullyAvailableBlocks[UnavailableBlocks[i]] = false; 1432 1433 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit; 1434 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); 1435 PI != E; ++PI) { 1436 BasicBlock *Pred = *PI; 1437 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) { 1438 continue; 1439 } 1440 PredLoads[Pred] = 0; 1441 1442 if (Pred->getTerminator()->getNumSuccessors() != 1) { 1443 if (isa<IndirectBrInst>(Pred->getTerminator())) { 1444 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '" 1445 << Pred->getName() << "': " << *LI << '\n'); 1446 return false; 1447 } 1448 unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB); 1449 NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum)); 1450 } 1451 } 1452 if (!NeedToSplit.empty()) { 1453 toSplit.append(NeedToSplit.begin(), NeedToSplit.end()); 1454 return false; 1455 } 1456 1457 // Decide whether PRE is profitable for this load. 1458 unsigned NumUnavailablePreds = PredLoads.size(); 1459 assert(NumUnavailablePreds != 0 && 1460 "Fully available value should be eliminated above!"); 1461 1462 // If this load is unavailable in multiple predecessors, reject it. 1463 // FIXME: If we could restructure the CFG, we could make a common pred with 1464 // all the preds that don't have an available LI and insert a new load into 1465 // that one block. 1466 if (NumUnavailablePreds != 1) 1467 return false; 1468 1469 // Check if the load can safely be moved to all the unavailable predecessors. 1470 bool CanDoPRE = true; 1471 SmallVector<Instruction*, 8> NewInsts; 1472 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1473 E = PredLoads.end(); I != E; ++I) { 1474 BasicBlock *UnavailablePred = I->first; 1475 1476 // Do PHI translation to get its value in the predecessor if necessary. The 1477 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred. 1478 1479 // If all preds have a single successor, then we know it is safe to insert 1480 // the load on the pred (?!?), so we can insert code to materialize the 1481 // pointer if it is not available. 1482 PHITransAddr Address(LI->getPointerOperand(), TD); 1483 Value *LoadPtr = 0; 1484 if (allSingleSucc) { 1485 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred, 1486 *DT, NewInsts); 1487 } else { 1488 Address.PHITranslateValue(LoadBB, UnavailablePred, DT); 1489 LoadPtr = Address.getAddr(); 1490 } 1491 1492 // If we couldn't find or insert a computation of this phi translated value, 1493 // we fail PRE. 1494 if (LoadPtr == 0) { 1495 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: " 1496 << *LI->getPointerOperand() << "\n"); 1497 CanDoPRE = false; 1498 break; 1499 } 1500 1501 // Make sure it is valid to move this load here. We have to watch out for: 1502 // @1 = getelementptr (i8* p, ... 1503 // test p and branch if == 0 1504 // load @1 1505 // It is valid to have the getelementptr before the test, even if p can 1506 // be 0, as getelementptr only does address arithmetic. 1507 // If we are not pushing the value through any multiple-successor blocks 1508 // we do not have this case. Otherwise, check that the load is safe to 1509 // put anywhere; this can be improved, but should be conservatively safe. 1510 if (!allSingleSucc && 1511 // FIXME: REEVALUTE THIS. 1512 !isSafeToLoadUnconditionally(LoadPtr, 1513 UnavailablePred->getTerminator(), 1514 LI->getAlignment(), TD)) { 1515 CanDoPRE = false; 1516 break; 1517 } 1518 1519 I->second = LoadPtr; 1520 } 1521 1522 if (!CanDoPRE) { 1523 while (!NewInsts.empty()) { 1524 Instruction *I = NewInsts.pop_back_val(); 1525 if (MD) MD->removeInstruction(I); 1526 I->eraseFromParent(); 1527 } 1528 return false; 1529 } 1530 1531 // Okay, we can eliminate this load by inserting a reload in the predecessor 1532 // and using PHI construction to get the value in the other predecessors, do 1533 // it. 1534 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n'); 1535 DEBUG(if (!NewInsts.empty()) 1536 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: " 1537 << *NewInsts.back() << '\n'); 1538 1539 // Assign value numbers to the new instructions. 1540 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) { 1541 // FIXME: We really _ought_ to insert these value numbers into their 1542 // parent's availability map. However, in doing so, we risk getting into 1543 // ordering issues. If a block hasn't been processed yet, we would be 1544 // marking a value as AVAIL-IN, which isn't what we intend. 1545 VN.lookup_or_add(NewInsts[i]); 1546 } 1547 1548 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(), 1549 E = PredLoads.end(); I != E; ++I) { 1550 BasicBlock *UnavailablePred = I->first; 1551 Value *LoadPtr = I->second; 1552 1553 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false, 1554 LI->getAlignment(), 1555 UnavailablePred->getTerminator()); 1556 1557 // Transfer the old load's TBAA tag to the new load. 1558 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa)) 1559 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag); 1560 1561 // Add the newly created load. 1562 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred, 1563 NewLoad)); 1564 MD->invalidateCachedPointerInfo(LoadPtr); 1565 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n'); 1566 } 1567 1568 // Perform PHI construction. 1569 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, TD, *DT, 1570 VN.getAliasAnalysis(), *MD); 1571 LI->replaceAllUsesWith(V); 1572 if (isa<PHINode>(V)) 1573 V->takeName(LI); 1574 if (V->getType()->isPointerTy()) 1575 MD->invalidateCachedPointerInfo(V); 1576 VN.erase(LI); 1577 InstrsToErase.push_back(LI); 1578 ++NumPRELoad; 1579 return true; 1580} 1581 1582/// processLoad - Attempt to eliminate a load, first by eliminating it 1583/// locally, and then attempting non-local elimination if that fails. 1584bool GVN::processLoad(LoadInst *L) { 1585 if (!MD) 1586 return false; 1587 1588 if (L->isVolatile()) 1589 return false; 1590 1591 // ... to a pointer that has been loaded from before... 1592 MemDepResult Dep = MD->getDependency(L); 1593 1594 // If we have a clobber and target data is around, see if this is a clobber 1595 // that we can fix up through code synthesis. 1596 if (Dep.isClobber() && TD) { 1597 // Check to see if we have something like this: 1598 // store i32 123, i32* %P 1599 // %A = bitcast i32* %P to i8* 1600 // %B = gep i8* %A, i32 1 1601 // %C = load i8* %B 1602 // 1603 // We could do that by recognizing if the clobber instructions are obviously 1604 // a common base + constant offset, and if the previous store (or memset) 1605 // completely covers this load. This sort of thing can happen in bitfield 1606 // access code. 1607 Value *AvailVal = 0; 1608 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) { 1609 int Offset = AnalyzeLoadFromClobberingStore(L->getType(), 1610 L->getPointerOperand(), 1611 DepSI, *TD); 1612 if (Offset != -1) 1613 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset, 1614 L->getType(), L, *TD); 1615 } 1616 1617 // Check to see if we have something like this: 1618 // load i32* P 1619 // load i8* (P+1) 1620 // if we have this, replace the later with an extraction from the former. 1621 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) { 1622 // If this is a clobber and L is the first instruction in its block, then 1623 // we have the first instruction in the entry block. 1624 if (DepLI == L) 1625 return false; 1626 1627 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(), 1628 L->getPointerOperand(), 1629 DepLI, *TD); 1630 if (Offset != -1) 1631 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *TD, *MD); 1632 } 1633 1634 // If the clobbering value is a memset/memcpy/memmove, see if we can forward 1635 // a value on from it. 1636 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) { 1637 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(), 1638 L->getPointerOperand(), 1639 DepMI, *TD); 1640 if (Offset != -1) 1641 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD); 1642 } 1643 1644 if (AvailVal) { 1645 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n' 1646 << *AvailVal << '\n' << *L << "\n\n\n"); 1647 1648 // Replace the load! 1649 L->replaceAllUsesWith(AvailVal); 1650 if (AvailVal->getType()->isPointerTy()) 1651 MD->invalidateCachedPointerInfo(AvailVal); 1652 VN.erase(L); 1653 InstrsToErase.push_back(L); 1654 ++NumGVNLoad; 1655 return true; 1656 } 1657 } 1658 1659 // If the value isn't available, don't do anything! 1660 if (Dep.isClobber()) { 1661 DEBUG( 1662 // fast print dep, using operator<< on instruction is too slow. 1663 dbgs() << "GVN: load "; 1664 WriteAsOperand(dbgs(), L); 1665 Instruction *I = Dep.getInst(); 1666 dbgs() << " is clobbered by " << *I << '\n'; 1667 ); 1668 return false; 1669 } 1670 1671 // If it is defined in another block, try harder. 1672 if (Dep.isNonLocal()) 1673 return processNonLocalLoad(L); 1674 1675 Instruction *DepInst = Dep.getInst(); 1676 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) { 1677 Value *StoredVal = DepSI->getValueOperand(); 1678 1679 // The store and load are to a must-aliased pointer, but they may not 1680 // actually have the same type. See if we know how to reuse the stored 1681 // value (depending on its type). 1682 if (StoredVal->getType() != L->getType()) { 1683 if (TD) { 1684 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(), 1685 L, *TD); 1686 if (StoredVal == 0) 1687 return false; 1688 1689 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal 1690 << '\n' << *L << "\n\n\n"); 1691 } 1692 else 1693 return false; 1694 } 1695 1696 // Remove it! 1697 L->replaceAllUsesWith(StoredVal); 1698 if (StoredVal->getType()->isPointerTy()) 1699 MD->invalidateCachedPointerInfo(StoredVal); 1700 VN.erase(L); 1701 InstrsToErase.push_back(L); 1702 ++NumGVNLoad; 1703 return true; 1704 } 1705 1706 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) { 1707 Value *AvailableVal = DepLI; 1708 1709 // The loads are of a must-aliased pointer, but they may not actually have 1710 // the same type. See if we know how to reuse the previously loaded value 1711 // (depending on its type). 1712 if (DepLI->getType() != L->getType()) { 1713 if (TD) { 1714 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), 1715 L, *TD); 1716 if (AvailableVal == 0) 1717 return false; 1718 1719 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal 1720 << "\n" << *L << "\n\n\n"); 1721 } 1722 else 1723 return false; 1724 } 1725 1726 // Remove it! 1727 L->replaceAllUsesWith(AvailableVal); 1728 if (DepLI->getType()->isPointerTy()) 1729 MD->invalidateCachedPointerInfo(DepLI); 1730 VN.erase(L); 1731 InstrsToErase.push_back(L); 1732 ++NumGVNLoad; 1733 return true; 1734 } 1735 1736 // If this load really doesn't depend on anything, then we must be loading an 1737 // undef value. This can happen when loading for a fresh allocation with no 1738 // intervening stores, for example. 1739 if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) { 1740 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1741 VN.erase(L); 1742 InstrsToErase.push_back(L); 1743 ++NumGVNLoad; 1744 return true; 1745 } 1746 1747 // If this load occurs either right after a lifetime begin, 1748 // then the loaded value is undefined. 1749 if (IntrinsicInst* II = dyn_cast<IntrinsicInst>(DepInst)) { 1750 if (II->getIntrinsicID() == Intrinsic::lifetime_start) { 1751 L->replaceAllUsesWith(UndefValue::get(L->getType())); 1752 VN.erase(L); 1753 InstrsToErase.push_back(L); 1754 ++NumGVNLoad; 1755 return true; 1756 } 1757 } 1758 1759 return false; 1760} 1761 1762// findLeader - In order to find a leader for a given value number at a 1763// specific basic block, we first obtain the list of all Values for that number, 1764// and then scan the list to find one whose block dominates the block in 1765// question. This is fast because dominator tree queries consist of only 1766// a few comparisons of DFS numbers. 1767Value *GVN::findLeader(BasicBlock *BB, uint32_t num) { 1768 LeaderTableEntry Vals = LeaderTable[num]; 1769 if (!Vals.Val) return 0; 1770 1771 Value *Val = 0; 1772 if (DT->dominates(Vals.BB, BB)) { 1773 Val = Vals.Val; 1774 if (isa<Constant>(Val)) return Val; 1775 } 1776 1777 LeaderTableEntry* Next = Vals.Next; 1778 while (Next) { 1779 if (DT->dominates(Next->BB, BB)) { 1780 if (isa<Constant>(Next->Val)) return Next->Val; 1781 if (!Val) Val = Next->Val; 1782 } 1783 1784 Next = Next->Next; 1785 } 1786 1787 return Val; 1788} 1789 1790 1791/// processInstruction - When calculating availability, handle an instruction 1792/// by inserting it into the appropriate sets 1793bool GVN::processInstruction(Instruction *I) { 1794 // Ignore dbg info intrinsics. 1795 if (isa<DbgInfoIntrinsic>(I)) 1796 return false; 1797 1798 // If the instruction can be easily simplified then do so now in preference 1799 // to value numbering it. Value numbering often exposes redundancies, for 1800 // example if it determines that %y is equal to %x then the instruction 1801 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify. 1802 if (Value *V = SimplifyInstruction(I, TD, DT)) { 1803 I->replaceAllUsesWith(V); 1804 if (MD && V->getType()->isPointerTy()) 1805 MD->invalidateCachedPointerInfo(V); 1806 VN.erase(I); 1807 InstrsToErase.push_back(I); 1808 return true; 1809 } 1810 1811 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 1812 if (processLoad(LI)) 1813 return true; 1814 1815 unsigned Num = VN.lookup_or_add(LI); 1816 addToLeaderTable(Num, LI, LI->getParent()); 1817 return false; 1818 } 1819 1820 // For conditions branches, we can perform simple conditional propagation on 1821 // the condition value itself. 1822 if (BranchInst *BI = dyn_cast<BranchInst>(I)) { 1823 if (!BI->isConditional() || isa<Constant>(BI->getCondition())) 1824 return false; 1825 1826 Value *BranchCond = BI->getCondition(); 1827 uint32_t CondVN = VN.lookup_or_add(BranchCond); 1828 1829 BasicBlock *TrueSucc = BI->getSuccessor(0); 1830 BasicBlock *FalseSucc = BI->getSuccessor(1); 1831 1832 if (TrueSucc->getSinglePredecessor()) 1833 addToLeaderTable(CondVN, 1834 ConstantInt::getTrue(TrueSucc->getContext()), 1835 TrueSucc); 1836 if (FalseSucc->getSinglePredecessor()) 1837 addToLeaderTable(CondVN, 1838 ConstantInt::getFalse(TrueSucc->getContext()), 1839 FalseSucc); 1840 1841 return false; 1842 } 1843 1844 // Instructions with void type don't return a value, so there's 1845 // no point in trying to find redudancies in them. 1846 if (I->getType()->isVoidTy()) return false; 1847 1848 uint32_t NextNum = VN.getNextUnusedValueNumber(); 1849 unsigned Num = VN.lookup_or_add(I); 1850 1851 // Allocations are always uniquely numbered, so we can save time and memory 1852 // by fast failing them. 1853 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) { 1854 addToLeaderTable(Num, I, I->getParent()); 1855 return false; 1856 } 1857 1858 // If the number we were assigned was a brand new VN, then we don't 1859 // need to do a lookup to see if the number already exists 1860 // somewhere in the domtree: it can't! 1861 if (Num == NextNum) { 1862 addToLeaderTable(Num, I, I->getParent()); 1863 return false; 1864 } 1865 1866 // Perform fast-path value-number based elimination of values inherited from 1867 // dominators. 1868 Value *repl = findLeader(I->getParent(), Num); 1869 if (repl == 0) { 1870 // Failure, just remember this instance for future use. 1871 addToLeaderTable(Num, I, I->getParent()); 1872 return false; 1873 } 1874 1875 // Remove it! 1876 VN.erase(I); 1877 I->replaceAllUsesWith(repl); 1878 if (MD && repl->getType()->isPointerTy()) 1879 MD->invalidateCachedPointerInfo(repl); 1880 InstrsToErase.push_back(I); 1881 return true; 1882} 1883 1884/// runOnFunction - This is the main transformation entry point for a function. 1885bool GVN::runOnFunction(Function& F) { 1886 if (!NoLoads) 1887 MD = &getAnalysis<MemoryDependenceAnalysis>(); 1888 DT = &getAnalysis<DominatorTree>(); 1889 TD = getAnalysisIfAvailable<TargetData>(); 1890 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>()); 1891 VN.setMemDep(MD); 1892 VN.setDomTree(DT); 1893 1894 bool Changed = false; 1895 bool ShouldContinue = true; 1896 1897 // Merge unconditional branches, allowing PRE to catch more 1898 // optimization opportunities. 1899 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) { 1900 BasicBlock *BB = FI++; 1901 1902 bool removedBlock = MergeBlockIntoPredecessor(BB, this); 1903 if (removedBlock) ++NumGVNBlocks; 1904 1905 Changed |= removedBlock; 1906 } 1907 1908 unsigned Iteration = 0; 1909 while (ShouldContinue) { 1910 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n"); 1911 ShouldContinue = iterateOnFunction(F); 1912 if (splitCriticalEdges()) 1913 ShouldContinue = true; 1914 Changed |= ShouldContinue; 1915 ++Iteration; 1916 } 1917 1918 if (EnablePRE) { 1919 bool PREChanged = true; 1920 while (PREChanged) { 1921 PREChanged = performPRE(F); 1922 Changed |= PREChanged; 1923 } 1924 } 1925 // FIXME: Should perform GVN again after PRE does something. PRE can move 1926 // computations into blocks where they become fully redundant. Note that 1927 // we can't do this until PRE's critical edge splitting updates memdep. 1928 // Actually, when this happens, we should just fully integrate PRE into GVN. 1929 1930 cleanupGlobalSets(); 1931 1932 return Changed; 1933} 1934 1935 1936bool GVN::processBlock(BasicBlock *BB) { 1937 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function 1938 // (and incrementing BI before processing an instruction). 1939 assert(InstrsToErase.empty() && 1940 "We expect InstrsToErase to be empty across iterations"); 1941 bool ChangedFunction = false; 1942 1943 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); 1944 BI != BE;) { 1945 ChangedFunction |= processInstruction(BI); 1946 if (InstrsToErase.empty()) { 1947 ++BI; 1948 continue; 1949 } 1950 1951 // If we need some instructions deleted, do it now. 1952 NumGVNInstr += InstrsToErase.size(); 1953 1954 // Avoid iterator invalidation. 1955 bool AtStart = BI == BB->begin(); 1956 if (!AtStart) 1957 --BI; 1958 1959 for (SmallVector<Instruction*, 4>::iterator I = InstrsToErase.begin(), 1960 E = InstrsToErase.end(); I != E; ++I) { 1961 DEBUG(dbgs() << "GVN removed: " << **I << '\n'); 1962 if (MD) MD->removeInstruction(*I); 1963 (*I)->eraseFromParent(); 1964 DEBUG(verifyRemoved(*I)); 1965 } 1966 InstrsToErase.clear(); 1967 1968 if (AtStart) 1969 BI = BB->begin(); 1970 else 1971 ++BI; 1972 } 1973 1974 return ChangedFunction; 1975} 1976 1977/// performPRE - Perform a purely local form of PRE that looks for diamond 1978/// control flow patterns and attempts to perform simple PRE at the join point. 1979bool GVN::performPRE(Function &F) { 1980 bool Changed = false; 1981 DenseMap<BasicBlock*, Value*> predMap; 1982 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()), 1983 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) { 1984 BasicBlock *CurrentBlock = *DI; 1985 1986 // Nothing to PRE in the entry block. 1987 if (CurrentBlock == &F.getEntryBlock()) continue; 1988 1989 for (BasicBlock::iterator BI = CurrentBlock->begin(), 1990 BE = CurrentBlock->end(); BI != BE; ) { 1991 Instruction *CurInst = BI++; 1992 1993 if (isa<AllocaInst>(CurInst) || 1994 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) || 1995 CurInst->getType()->isVoidTy() || 1996 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() || 1997 isa<DbgInfoIntrinsic>(CurInst)) 1998 continue; 1999 2000 // We don't currently value number ANY inline asm calls. 2001 if (CallInst *CallI = dyn_cast<CallInst>(CurInst)) 2002 if (CallI->isInlineAsm()) 2003 continue; 2004 2005 uint32_t ValNo = VN.lookup(CurInst); 2006 2007 // Look for the predecessors for PRE opportunities. We're 2008 // only trying to solve the basic diamond case, where 2009 // a value is computed in the successor and one predecessor, 2010 // but not the other. We also explicitly disallow cases 2011 // where the successor is its own predecessor, because they're 2012 // more complicated to get right. 2013 unsigned NumWith = 0; 2014 unsigned NumWithout = 0; 2015 BasicBlock *PREPred = 0; 2016 predMap.clear(); 2017 2018 for (pred_iterator PI = pred_begin(CurrentBlock), 2019 PE = pred_end(CurrentBlock); PI != PE; ++PI) { 2020 BasicBlock *P = *PI; 2021 // We're not interested in PRE where the block is its 2022 // own predecessor, or in blocks with predecessors 2023 // that are not reachable. 2024 if (P == CurrentBlock) { 2025 NumWithout = 2; 2026 break; 2027 } else if (!DT->dominates(&F.getEntryBlock(), P)) { 2028 NumWithout = 2; 2029 break; 2030 } 2031 2032 Value* predV = findLeader(P, ValNo); 2033 if (predV == 0) { 2034 PREPred = P; 2035 ++NumWithout; 2036 } else if (predV == CurInst) { 2037 NumWithout = 2; 2038 } else { 2039 predMap[P] = predV; 2040 ++NumWith; 2041 } 2042 } 2043 2044 // Don't do PRE when it might increase code size, i.e. when 2045 // we would need to insert instructions in more than one pred. 2046 if (NumWithout != 1 || NumWith == 0) 2047 continue; 2048 2049 // Don't do PRE across indirect branch. 2050 if (isa<IndirectBrInst>(PREPred->getTerminator())) 2051 continue; 2052 2053 // We can't do PRE safely on a critical edge, so instead we schedule 2054 // the edge to be split and perform the PRE the next time we iterate 2055 // on the function. 2056 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock); 2057 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) { 2058 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum)); 2059 continue; 2060 } 2061 2062 // Instantiate the expression in the predecessor that lacked it. 2063 // Because we are going top-down through the block, all value numbers 2064 // will be available in the predecessor by the time we need them. Any 2065 // that weren't originally present will have been instantiated earlier 2066 // in this loop. 2067 Instruction *PREInstr = CurInst->clone(); 2068 bool success = true; 2069 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) { 2070 Value *Op = PREInstr->getOperand(i); 2071 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op)) 2072 continue; 2073 2074 if (Value *V = findLeader(PREPred, VN.lookup(Op))) { 2075 PREInstr->setOperand(i, V); 2076 } else { 2077 success = false; 2078 break; 2079 } 2080 } 2081 2082 // Fail out if we encounter an operand that is not available in 2083 // the PRE predecessor. This is typically because of loads which 2084 // are not value numbered precisely. 2085 if (!success) { 2086 delete PREInstr; 2087 DEBUG(verifyRemoved(PREInstr)); 2088 continue; 2089 } 2090 2091 PREInstr->insertBefore(PREPred->getTerminator()); 2092 PREInstr->setName(CurInst->getName() + ".pre"); 2093 predMap[PREPred] = PREInstr; 2094 VN.add(PREInstr, ValNo); 2095 ++NumGVNPRE; 2096 2097 // Update the availability map to include the new instruction. 2098 addToLeaderTable(ValNo, PREInstr, PREPred); 2099 2100 // Create a PHI to make the value available in this block. 2101 pred_iterator PB = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock); 2102 PHINode* Phi = PHINode::Create(CurInst->getType(), std::distance(PB, PE), 2103 CurInst->getName() + ".pre-phi", 2104 CurrentBlock->begin()); 2105 for (pred_iterator PI = PB; PI != PE; ++PI) { 2106 BasicBlock *P = *PI; 2107 Phi->addIncoming(predMap[P], P); 2108 } 2109 2110 VN.add(Phi, ValNo); 2111 addToLeaderTable(ValNo, Phi, CurrentBlock); 2112 2113 CurInst->replaceAllUsesWith(Phi); 2114 if (Phi->getType()->isPointerTy()) { 2115 // Because we have added a PHI-use of the pointer value, it has now 2116 // "escaped" from alias analysis' perspective. We need to inform 2117 // AA of this. 2118 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; ++ii) 2119 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(2*ii)); 2120 2121 if (MD) 2122 MD->invalidateCachedPointerInfo(Phi); 2123 } 2124 VN.erase(CurInst); 2125 removeFromLeaderTable(ValNo, CurInst, CurrentBlock); 2126 2127 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n'); 2128 if (MD) MD->removeInstruction(CurInst); 2129 CurInst->eraseFromParent(); 2130 DEBUG(verifyRemoved(CurInst)); 2131 Changed = true; 2132 } 2133 } 2134 2135 if (splitCriticalEdges()) 2136 Changed = true; 2137 2138 return Changed; 2139} 2140 2141/// splitCriticalEdges - Split critical edges found during the previous 2142/// iteration that may enable further optimization. 2143bool GVN::splitCriticalEdges() { 2144 if (toSplit.empty()) 2145 return false; 2146 do { 2147 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val(); 2148 SplitCriticalEdge(Edge.first, Edge.second, this); 2149 } while (!toSplit.empty()); 2150 if (MD) MD->invalidateCachedPredecessors(); 2151 return true; 2152} 2153 2154/// iterateOnFunction - Executes one iteration of GVN 2155bool GVN::iterateOnFunction(Function &F) { 2156 cleanupGlobalSets(); 2157 2158 // Top-down walk of the dominator tree 2159 bool Changed = false; 2160#if 0 2161 // Needed for value numbering with phi construction to work. 2162 ReversePostOrderTraversal<Function*> RPOT(&F); 2163 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(), 2164 RE = RPOT.end(); RI != RE; ++RI) 2165 Changed |= processBlock(*RI); 2166#else 2167 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()), 2168 DE = df_end(DT->getRootNode()); DI != DE; ++DI) 2169 Changed |= processBlock(DI->getBlock()); 2170#endif 2171 2172 return Changed; 2173} 2174 2175void GVN::cleanupGlobalSets() { 2176 VN.clear(); 2177 LeaderTable.clear(); 2178 TableAllocator.Reset(); 2179} 2180 2181/// verifyRemoved - Verify that the specified instruction does not occur in our 2182/// internal data structures. 2183void GVN::verifyRemoved(const Instruction *Inst) const { 2184 VN.verifyRemoved(Inst); 2185 2186 // Walk through the value number scope to make sure the instruction isn't 2187 // ferreted away in it. 2188 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator 2189 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) { 2190 const LeaderTableEntry *Node = &I->second; 2191 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2192 2193 while (Node->Next) { 2194 Node = Node->Next; 2195 assert(Node->Val != Inst && "Inst still in value numbering scope!"); 2196 } 2197 } 2198} 2199