bool CallAnalyzer::visitBranchInst(BranchInst &BI) {
// We model unconditional branches as essentially free -- they really
// shouldn't exist at all, but handling them makes the behavior of the
// inliner more regular and predictable. Interestingly, conditional branches
// which will fold away are also free.
return BI.isUnconditional() || isa<ConstantInt>(BI.getCondition()) ||
dyn_cast_or_null<ConstantInt>(
SimplifiedValues.lookup(BI.getCondition()));
}
int BranchProbabilities::CheckFloatHeuristic()
{
// Heuristic fails if the last instruction is not a conditional branch
BranchInst *BI = dyn_cast<BranchInst>(_TI);
if ((!BI) || (BI->isUnconditional()))
return -1;
// All float comparisons are done with the fcmp instruction
FCmpInst *fcmp = dyn_cast<FCmpInst>(BI->getCondition());
if (!fcmp)
return -1;
// Choose the prefered branch depending on if this is an eq or neq comp
switch (fcmp->getPredicate())
{
case FCmpInst::FCMP_OEQ:
case FCmpInst::FCMP_UEQ:
return 1;
case FCmpInst::FCMP_ONE:
case FCmpInst::FCMP_UNE:
return 0;
case FCmpInst::FCMP_FALSE:
case FCmpInst::FCMP_TRUE:
assert("true or false predicate should have been folded!");
default:
return -1;
}
}
bool BranchProbabilityInfo::calcFloatingPointHeuristics(BasicBlock *BB) {
BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
if (!BI || !BI->isConditional())
return false;
Value *Cond = BI->getCondition();
FCmpInst *FCmp = dyn_cast<FCmpInst>(Cond);
if (!FCmp)
return false;
bool isProb;
if (FCmp->isEquality()) {
// f1 == f2 -> Unlikely
// f1 != f2 -> Likely
isProb = !FCmp->isTrueWhenEqual();
} else if (FCmp->getPredicate() == FCmpInst::FCMP_ORD) {
// !isnan -> Likely
isProb = true;
} else if (FCmp->getPredicate() == FCmpInst::FCMP_UNO) {
// isnan -> Unlikely
isProb = false;
} else {
return false;
}
unsigned TakenIdx = 0, NonTakenIdx = 1;
if (!isProb)
std::swap(TakenIdx, NonTakenIdx);
setEdgeWeight(BB, TakenIdx, FPH_TAKEN_WEIGHT);
setEdgeWeight(BB, NonTakenIdx, FPH_NONTAKEN_WEIGHT);
return true;
}
/////////////////////////////////////
//findLastSyn() //
//Find the last instruction in BB. Any instructions before this one must be replicated except Synchpoint
/////////////////////////////////////
Instruction* InsDuplica::findLastCond (BasicBlock *BB) {
TerminatorInst *lastIns = BB->getTerminator();
assert(!(isa<SwitchInst>(lastIns)) && "Find a SwitchInst! You need to lower SwitchInst.");
BranchInst * BI = dyn_cast<BranchInst>(lastIns);
if (BI && (BI->isConditional())) {
Value *cond = BI->getCondition();
Instruction* condIns = dyn_cast<Instruction>(cond);
//find condIns. But we have to make sure condIns is the second to the last instruction in BB
assert(condIns && "Branch Condition must not be trivial");
if ((condIns->getNextNode())!=lastIns) {
//assert((condIns->getParent() == BB) && "condIns is not in the same BB as br!");
//if condIns is not in the same BB as br. We have to leave it there
if (condIns->getParent() != BB) return lastIns;
//if condIns is a PHINode, since we can not move, we just return BI
//if condIns has more than one use, better not to reorder
if ((isa<PHINode>(condIns))||!(condIns->hasOneUse())) return lastIns;
condIns->moveBefore(lastIns); //we moved condInst right before br
#ifdef Jing_DEBUG
std::cerr << "adjust order of condIns "<< condIns->getName() <<" in " << BB->getName() <<"\n";
#endif
}
#ifdef Jing_DEBUG
std::cerr << "findLastCond returns " << condIns->getName() <<"\n";
#endif
return condIns;
}
//return the terminator instruction
return lastIns;
}
void TempScopInfo::buildCondition(BasicBlock *BB, BasicBlock *RegionEntry) {
BBCond Cond;
DomTreeNode *BBNode = DT->getNode(BB), *EntryNode = DT->getNode(RegionEntry);
assert(BBNode && EntryNode && "Get null node while building condition!");
// Walk up the dominance tree until reaching the entry node. Add all
// conditions on the path to BB except if BB postdominates the block
// containing the condition.
while (BBNode != EntryNode) {
BasicBlock *CurBB = BBNode->getBlock();
BBNode = BBNode->getIDom();
assert(BBNode && "BBNode should not reach the root node!");
if (PDT->dominates(CurBB, BBNode->getBlock()))
continue;
BranchInst *Br = dyn_cast<BranchInst>(BBNode->getBlock()->getTerminator());
assert(Br && "A Valid Scop should only contain branch instruction");
if (Br->isUnconditional())
continue;
// Is BB on the ELSE side of the branch?
bool inverted = DT->dominates(Br->getSuccessor(1), BB);
Comparison *Cmp;
buildAffineCondition(*(Br->getCondition()), inverted, &Cmp);
Cond.push_back(*Cmp);
}
if (!Cond.empty())
BBConds[BB] = Cond;
}
void MakeDispatcherPass::ConvertCmp(Function& function)
{
typedef std::vector< Instruction * > InstList;
InstList insts;
for (Function::iterator BB = function.begin(), bbE = function.end(); BB != bbE; ++BB)
{
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;)
{
if (isa< CmpInst >(I))
{
insts.push_back(I);
}
if (isa< BranchInst >(I))
{
BasicBlock::iterator save = I;
BranchInst* branchInst = dynamic_cast< BranchInst *>(&*I);
if (branchInst->isConditional() && !insts.empty())
{
Value* valbranch = NULL;
valbranch = branchInst->getCondition();
ShowType(dynamic_cast<CmpInst*>(insts[0]));
CreateInt3(BB, I);
I++;
save->eraseFromParent();
insts.pop_back();
continue;
}
}
I++;
}
}
}
//It receives a BasicBLock and makes table of predicates and its respective gated instructions
void bSSA::makeTable (BasicBlock *BB, Function *F) {
Value *condition;
TerminatorInst *ti = BB->getTerminator();
BranchInst *bi = NULL;
SwitchInst *si=NULL;
PostDominatorTree &PD = getAnalysis<PostDominatorTree>(*F);
ProcessedBB.clear();
if ((bi = dyn_cast<BranchInst>(ti)) && bi->isConditional()) { //If the terminator instruction is a conditional branch
condition = bi->getCondition();
//Including the predicate on the predicatesVector
predicatesVector.push_back(new Pred(condition));
//Make a "Flooding" on each sucessor gated the instruction on Influence Region of the predicate
for (unsigned int i=0; i<bi->getNumSuccessors(); i++) {
findIR (BB, bi->getSuccessor(i),PD);
}
}else if ((si = dyn_cast<SwitchInst>(ti))) {
condition = si->getCondition();
//Including the predicate on the predicatesVector
predicatesVector.push_back(new Pred(condition));
//Make a "Flooding" on each sucessor gated the instruction on Influence Region of the predicate
for (unsigned int i=0; i<si->getNumSuccessors(); i++) {
findIR (BB, si->getSuccessor(i),PD);
}
}
}
bool BranchProbabilityAnalysis::calcZeroHeuristics(BasicBlock *BB) {
BranchInst * BI = dyn_cast<BranchInst>(BB->getTerminator());
if (!BI || !BI->isConditional())
return false;
Value *Cond = BI->getCondition();
ICmpInst *CI = dyn_cast<ICmpInst>(Cond);
if (!CI)
return false;
Value *RHS = CI->getOperand(1);
ConstantInt *CV = dyn_cast<ConstantInt>(RHS);
if (!CV)
return false;
bool isProb;
if (CV->isZero()) {
switch (CI->getPredicate()) {
case CmpInst::ICMP_EQ:
// X == 0 -> Unlikely
isProb = false;
break;
case CmpInst::ICMP_NE:
// X != 0 -> Likely
isProb = true;
break;
case CmpInst::ICMP_SLT:
// X < 0 -> Unlikely
isProb = false;
break;
case CmpInst::ICMP_SGT:
// X > 0 -> Likely
isProb = true;
break;
default:
return false;
}
} else if (CV->isOne() && CI->getPredicate() == CmpInst::ICMP_SLT) {
// InstCombine canonicalizes X <= 0 into X < 1.
// X <= 0 -> Unlikely
isProb = false;
} else if (CV->isAllOnesValue() && CI->getPredicate() == CmpInst::ICMP_SGT) {
// InstCombine canonicalizes X >= 0 into X > -1.
// X >= 0 -> Likely
isProb = true;
} else {
return false;
}
BasicBlock *Taken = BI->getSuccessor(0);
BasicBlock *NonTaken = BI->getSuccessor(1);
if (!isProb)
std::swap(Taken, NonTaken);
BP->setEdgeWeight(BB, Taken, ZH_TAKEN_WEIGHT);
BP->setEdgeWeight(BB, NonTaken, ZH_NONTAKEN_WEIGHT);
return true;
}
// Calculate Edge Weights using "Pointer Heuristics". Predict a comparsion
// between two pointer or pointer and NULL will fail.
bool BranchProbabilityInfo::calcPointerHeuristics(BasicBlock *BB) {
BranchInst * BI = dyn_cast<BranchInst>(BB->getTerminator());
if (!BI || !BI->isConditional())
return false;
Value *Cond = BI->getCondition();
ICmpInst *CI = dyn_cast<ICmpInst>(Cond);
if (!CI || !CI->isEquality())
return false;
Value *LHS = CI->getOperand(0);
if (!LHS->getType()->isPointerTy())
return false;
assert(CI->getOperand(1)->getType()->isPointerTy());
// p != 0 -> isProb = true
// p == 0 -> isProb = false
// p != q -> isProb = true
// p == q -> isProb = false;
unsigned TakenIdx = 0, NonTakenIdx = 1;
bool isProb = CI->getPredicate() == ICmpInst::ICMP_NE;
if (!isProb)
std::swap(TakenIdx, NonTakenIdx);
setEdgeWeight(BB, TakenIdx, PH_TAKEN_WEIGHT);
setEdgeWeight(BB, NonTakenIdx, PH_NONTAKEN_WEIGHT);
return true;
}
// Look for control dependencies on a read.
bool branchesOn(BasicBlock *bb, Value *load,
ICmpInst **icmpOut, int *outIdx) {
// XXX: make this platform configured; on some platforms maybe an
// atomic cmpxchg does /not/ behave like it branches on the old value
if (isa<AtomicCmpXchgInst>(load) || isa<AtomicRMWInst>(load)) {
if (icmpOut) *icmpOut = nullptr;
if (outIdx) *outIdx = 0;
return true;
}
// TODO: we should be able to follow values through phi nodes,
// since we are path dependent anyways.
BranchInst *br = dyn_cast<BranchInst>(bb->getTerminator());
if (!br || !br->isConditional()) return false;
// TODO: We only check one level of things. Check deeper?
// We pretty heavily restrict what operations we handle here.
// Some would just be wrong (like call), but really icmp is
// the main one, so. Probably we should be able to also
// pick through casts and wideness changes.
ICmpInst *icmp = dyn_cast<ICmpInst>(br->getCondition());
if (!icmp) return false;
int idx = 0;
for (auto v : icmp->operand_values()) {
if (getRealValue(v) == load) {
if (icmpOut) *icmpOut = icmp;
if (outIdx) *outIdx = idx;
return true;
}
++idx;
}
return false;
}
void AMDGPUAnnotateUniformValues::visitBranchInst(BranchInst &I) {
if (I.isUnconditional())
return;
Value *Cond = I.getCondition();
if (!DA->isUniform(Cond))
return;
setUniformMetadata(I.getParent()->getTerminator());
}
bool BranchProbabilityAnalysis::calcZeroHeuristics(BasicBlock *BB) {
BranchInst * BI = dyn_cast<BranchInst>(BB->getTerminator());
if (!BI || !BI->isConditional())
return false;
Value *Cond = BI->getCondition();
ICmpInst *CI = dyn_cast<ICmpInst>(Cond);
if (!CI)
return false;
Value *RHS = CI->getOperand(1);
ConstantInt *CV = dyn_cast<ConstantInt>(RHS);
if (!CV || !CV->isZero())
return false;
bool isProb;
switch (CI->getPredicate()) {
case CmpInst::ICMP_EQ:
// Equal to zero is not expected to be taken.
isProb = false;
break;
case CmpInst::ICMP_NE:
// Not equal to zero is expected.
isProb = true;
break;
case CmpInst::ICMP_SLT:
// Less or equal to zero is not expected.
// X < 0 -> Unlikely
isProb = false;
break;
case CmpInst::ICMP_UGT:
case CmpInst::ICMP_SGT:
// Greater or equal to zero is expected.
// X > 0 -> Likely
isProb = true;
break;
default:
return false;
};
BasicBlock *Taken = BI->getSuccessor(0);
BasicBlock *NonTaken = BI->getSuccessor(1);
if (!isProb)
std::swap(Taken, NonTaken);
BP->setEdgeWeight(BB, Taken, ZH_TAKEN_WEIGHT);
BP->setEdgeWeight(BB, NonTaken, ZH_NONTAKEN_WEIGHT);
return true;
}
bool IRTranslator::translateBr(const BranchInst &BrInst) {
unsigned Succ = 0;
if (!BrInst.isUnconditional()) {
// We want a G_BRCOND to the true BB followed by an unconditional branch.
unsigned Tst = getOrCreateVReg(*BrInst.getCondition());
const BasicBlock &TrueTgt = *cast<BasicBlock>(BrInst.getSuccessor(Succ++));
MachineBasicBlock &TrueBB = getOrCreateBB(TrueTgt);
MIRBuilder.buildBrCond(LLT{*BrInst.getCondition()->getType()}, Tst, TrueBB);
}
const BasicBlock &BrTgt = *cast<BasicBlock>(BrInst.getSuccessor(Succ));
MachineBasicBlock &TgtBB = getOrCreateBB(BrTgt);
MIRBuilder.buildBr(TgtBB);
// Link successors.
MachineBasicBlock &CurBB = MIRBuilder.getMBB();
for (const BasicBlock *Succ : BrInst.successors())
CurBB.addSuccessor(&getOrCreateBB(*Succ));
return true;
}
void Interpreter::visitBranchInst(BranchInst &I) {
ExecutionContext &SF = ECStack.back();
BasicBlock *Dest;
Dest = I.getSuccessor(0); // Uncond branches have a fixed dest...
if (!I.isUnconditional()) {
Value *Cond = I.getCondition();
if (getOperandValue(Cond, SF).BoolVal == 0) // If false cond...
Dest = I.getSuccessor(1);
}
SwitchToNewBasicBlock(Dest, SF);
}
void Graph::DFS2_visit(DFSNode * DFS) {
DFS->C = GRAY;
if (DFS->T == UNKNOWN){
Instruction *I = dynamic_cast<Instruction *>(DFS->Bb->getTerminator());
BranchInst *BI = static_cast<BranchInst *>(I);
if (BI->isConditional()){
//branchmap[BI] = dfs->bb;
BranchMap[BI->getCondition()] = DFS->Bb;
if (TypeStack.empty())
DFS->T = IF;
else if (TypeStack.top() != ENDIF)
DFS->T = IF;
//else if (TypeStack.top() == ENDIF && Time == DFS->DTime - 2)
// DFS->T = ELSEIF;
else if (DFS->Bb->getName().substr(0,7) == "if.else")
DFS->T = ELSEIF;
//else if (TypeStack.top() == ELSEIF)
// DFS->T = ELSEIF;
else if(TypeStack.top() == ENDIF && (Time != DFS->DTime - 1))
DFS->T = IF;
else
DEBUG (errs() << "\n\n\n\n\n\n\nError at: " << " Top = " << TypeStack.top() << " Time = " << Time << " DTime = " << DFS->DTime << "\n\n\n\n\n\n\n\n\n");
}
else {
DEBUG (errs() << "Bb: " <<DFS->Bb << " is not conditional\n");
}
}
if (DFS->T != UNKNOWN){
Time = DFS->DTime;
TypeStack.push(DFS->T);
DEBUG (errs() << "Stack: " << TypeStack.top() <<'\n');
}
DEBUG (errs() << "Type of: " << DFS->Bb << " is: " << DFS->T
<< " global time: " << Time << " dfstime: " << DFS->DTime << '\n');
// for each vector adjacent to dfs
std::vector<DFSNode *> Svec = DFS->Vertex;
for (std::vector<DFSNode *>::iterator It = Svec.begin(); It != Svec.end(); ++It)
{
DFSNode *DN = *It;
if (DN->C == WHITE)
DFS2_visit(DN);
}
DFS->C = BLACK;
if (DFS->T != UNKNOWN)
Time = DFS->FTime;
}
// Predict that a comparison in which a register is an operand, the register is
// used before being defined in a successor block, and the successor block
// does not post-dominate will reach the successor block.
int BranchProbabilities::CheckGuardHeuristic()
{
BranchInst *BI = dyn_cast<BranchInst>(_TI);
bool bUses[2] = {false, false};
// If we don't have a conditional branch, abandon
if ((!BI) || (BI->isUnconditional()))
return -1;
// If the condition is not immediately dependent on a comparison, abandon
CmpInst *cmp = dyn_cast<CmpInst>(BI->getCondition());
if (!cmp)
return -1;
for (int i = 0; i < 2; i++)
{
if (_bPostDoms[i])
continue;
// Get the values being compared
Value *v = cmp->getOperand(i);
// For all uses of the first value check if the use post-dominates
for (Value::use_iterator UI = v->use_begin(), UE = v->use_end();
UI != UE; ++UI)
{
// if the use is not an instruction, skip it
Instruction *I = dyn_cast<Instruction>(*UI);
if (!I)
continue;
BasicBlock *UsingBlock = I->getParent();
// Check if the use is in either successor
for (int i = 0; i < 2; i++)
if (UsingBlock == _Succ[i])
bUses[i] = true;
}
}
if (bUses[0] == bUses[1])
return -1;
if (bUses[0])
return 0;
else
return 1;
}
/// Annotate the control flow with intrinsics so the backend can
/// recognize if/then/else and loops.
bool SIAnnotateControlFlow::runOnFunction(Function &F) {
DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
DA = &getAnalysis<LegacyDivergenceAnalysis>();
for (df_iterator<BasicBlock *> I = df_begin(&F.getEntryBlock()),
E = df_end(&F.getEntryBlock()); I != E; ++I) {
BasicBlock *BB = *I;
BranchInst *Term = dyn_cast<BranchInst>(BB->getTerminator());
if (!Term || Term->isUnconditional()) {
if (isTopOfStack(BB))
closeControlFlow(BB);
continue;
}
if (I.nodeVisited(Term->getSuccessor(1))) {
if (isTopOfStack(BB))
closeControlFlow(BB);
handleLoop(Term);
continue;
}
if (isTopOfStack(BB)) {
PHINode *Phi = dyn_cast<PHINode>(Term->getCondition());
if (Phi && Phi->getParent() == BB && isElse(Phi)) {
insertElse(Term);
eraseIfUnused(Phi);
continue;
}
closeControlFlow(BB);
}
openIf(Term);
}
if (!Stack.empty()) {
// CFG was probably not structured.
report_fatal_error("failed to annotate CFG");
}
return true;
}
/// MatchPointerHeuristic - Predict that a comparison of a pointer against
/// null or of two pointers will fail.
/// @returns a Prediction that is a pair in which the first element is the
/// successor taken, and the second the successor not taken.
Prediction BranchHeuristicsInfo::MatchPointerHeuristic(BasicBlock *root) const {
// Last instruction of basic block.
TerminatorInst *TI = root->getTerminator();
// Basic block successors. True and False branches.
BasicBlock *trueSuccessor = TI->getSuccessor(0);
BasicBlock *falseSuccessor = TI->getSuccessor(1);
// Is the last instruction a Branch Instruction?
BranchInst *BI = dyn_cast<BranchInst>(TI);
if (!BI || !BI->isConditional())
return empty;
// Conditional instruction.
Value *cond = BI->getCondition();
// Pointer comparisons are integer comparisons.
ICmpInst *II = dyn_cast<ICmpInst>(cond);
if (!II)
return empty;
// An integer comparison has always two operands.
Value *operand1 = II->getOperand(0);
Value *operand2 = II->getOperand(1);
// Obtain the type of comparison.
enum ICmpInst::Predicate signedPred = II->getSignedPredicate();
// The heuristic states that it must be compared against null,
// but in LLVM, null is also a PointerType, so it only requires
// to test if there is a comparison between two pointers.
if (signedPred == ICmpInst::ICMP_EQ &&
isa<PointerType>(operand1->getType()) && // NULL is a pointer type too
isa<PointerType>(operand2->getType())) { // NULL is a pointer type too
return std::make_pair(falseSuccessor, trueSuccessor);
} else if (signedPred != ICmpInst::ICMP_EQ &&
isa<PointerType>(operand1->getType()) &&
isa<PointerType>(operand2->getType())) {
return std::make_pair(trueSuccessor, falseSuccessor);
}
return empty;
}
/// \brief Annotate the control flow with intrinsics so the backend can
/// recognize if/then/else and loops.
bool SIAnnotateControlFlow::runOnFunction(Function &F) {
DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
DA = &getAnalysis<DivergenceAnalysis>();
for (df_iterator<BasicBlock *> I = df_begin(&F.getEntryBlock()),
E = df_end(&F.getEntryBlock()); I != E; ++I) {
BranchInst *Term = dyn_cast<BranchInst>((*I)->getTerminator());
if (!Term || Term->isUnconditional()) {
if (isTopOfStack(*I))
closeControlFlow(*I);
continue;
}
if (I.nodeVisited(Term->getSuccessor(1))) {
if (isTopOfStack(*I))
closeControlFlow(*I);
handleLoop(Term);
continue;
}
if (isTopOfStack(*I)) {
PHINode *Phi = dyn_cast<PHINode>(Term->getCondition());
if (Phi && Phi->getParent() == *I && isElse(Phi)) {
insertElse(Term);
eraseIfUnused(Phi);
continue;
}
closeControlFlow(*I);
}
openIf(Term);
}
assert(Stack.empty());
return true;
}
int BranchProbabilities::CheckPointerHeuristic()
{
// Heuristic fails if the last instruction is not a conditional branch
BranchInst *BI = dyn_cast<BranchInst>(_TI);
if ((!BI) || (BI->isUnconditional()))
return -1;
// All pointer comparisons are done with the icmp instruction
ICmpInst *icmp = dyn_cast<ICmpInst>(BI->getCondition());
if (!icmp)
return -1;
Value *v[2];
v[0] = icmp->getOperand(0);
v[1] = icmp->getOperand(1);
// Make sure we're comparing pointers
if (isa<PointerType>(v[0]->getType()))
{
assert(isa<PointerType>(v[1]->getType()) && "v[1] is not a pointer!");
// Choose the prefered branch depending on if this is an eq or neq comp
switch (icmp->getPredicate())
{
case ICmpInst::ICMP_EQ:
return 1;
case ICmpInst::ICMP_NE:
return 0;
default:
assert("non-equality comparison of pointers");
return -1;
}
}
return -1;
}
bool BranchProbabilityInfo::calcZeroHeuristics(BasicBlock *BB) {
BranchInst * BI = dyn_cast<BranchInst>(BB->getTerminator());
if (!BI || !BI->isConditional())
return false;
Value *Cond = BI->getCondition();
ICmpInst *CI = dyn_cast<ICmpInst>(Cond);
if (!CI)
return false;
Value *RHS = CI->getOperand(1);
ConstantInt *CV = dyn_cast<ConstantInt>(RHS);
if (!CV)
return false;
// If the LHS is the result of AND'ing a value with a single bit bitmask,
// we don't have information about probabilities.
if (Instruction *LHS = dyn_cast<Instruction>(CI->getOperand(0)))
if (LHS->getOpcode() == Instruction::And)
if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(LHS->getOperand(1)))
if (AndRHS->getUniqueInteger().isPowerOf2())
return false;
bool isProb;
if (CV->isZero()) {
switch (CI->getPredicate()) {
case CmpInst::ICMP_EQ:
// X == 0 -> Unlikely
isProb = false;
break;
case CmpInst::ICMP_NE:
// X != 0 -> Likely
isProb = true;
break;
case CmpInst::ICMP_SLT:
// X < 0 -> Unlikely
isProb = false;
break;
case CmpInst::ICMP_SGT:
// X > 0 -> Likely
isProb = true;
break;
default:
return false;
}
} else if (CV->isOne() && CI->getPredicate() == CmpInst::ICMP_SLT) {
// InstCombine canonicalizes X <= 0 into X < 1.
// X <= 0 -> Unlikely
isProb = false;
} else if (CV->isAllOnesValue()) {
switch (CI->getPredicate()) {
case CmpInst::ICMP_EQ:
// X == -1 -> Unlikely
isProb = false;
break;
case CmpInst::ICMP_NE:
// X != -1 -> Likely
isProb = true;
break;
case CmpInst::ICMP_SGT:
// InstCombine canonicalizes X >= 0 into X > -1.
// X >= 0 -> Likely
isProb = true;
break;
default:
return false;
}
} else {
return false;
}
unsigned TakenIdx = 0, NonTakenIdx = 1;
if (!isProb)
std::swap(TakenIdx, NonTakenIdx);
BranchProbability TakenProb(ZH_TAKEN_WEIGHT,
ZH_TAKEN_WEIGHT + ZH_NONTAKEN_WEIGHT);
setEdgeProbability(BB, TakenIdx, TakenProb);
setEdgeProbability(BB, NonTakenIdx, TakenProb.getCompl());
return true;
}
bool ScopDetection::isValidCFG(BasicBlock &BB,
DetectionContext &Context) const {
Region &RefRegion = Context.CurRegion;
TerminatorInst *TI = BB.getTerminator();
// Return instructions are only valid if the region is the top level region.
if (isa<ReturnInst>(TI) && !RefRegion.getExit() && TI->getNumOperands() == 0)
return true;
BranchInst *Br = dyn_cast<BranchInst>(TI);
if (!Br)
return invalid<ReportNonBranchTerminator>(Context, /*Assert=*/true, &BB);
if (Br->isUnconditional())
return true;
Value *Condition = Br->getCondition();
// UndefValue is not allowed as condition.
if (isa<UndefValue>(Condition))
return invalid<ReportUndefCond>(Context, /*Assert=*/true, &BB);
// Only Constant and ICmpInst are allowed as condition.
if (!(isa<Constant>(Condition) || isa<ICmpInst>(Condition)))
return invalid<ReportInvalidCond>(Context, /*Assert=*/true, &BB);
// Allow perfectly nested conditions.
assert(Br->getNumSuccessors() == 2 && "Unexpected number of successors");
if (ICmpInst *ICmp = dyn_cast<ICmpInst>(Condition)) {
// Unsigned comparisons are not allowed. They trigger overflow problems
// in the code generation.
//
// TODO: This is not sufficient and just hides bugs. However it does pretty
// well.
if (ICmp->isUnsigned())
return false;
// Are both operands of the ICmp affine?
if (isa<UndefValue>(ICmp->getOperand(0)) ||
isa<UndefValue>(ICmp->getOperand(1)))
return invalid<ReportUndefOperand>(Context, /*Assert=*/true, &BB);
Loop *L = LI->getLoopFor(ICmp->getParent());
const SCEV *LHS = SE->getSCEVAtScope(ICmp->getOperand(0), L);
const SCEV *RHS = SE->getSCEVAtScope(ICmp->getOperand(1), L);
if (!isAffineExpr(&Context.CurRegion, LHS, *SE) ||
!isAffineExpr(&Context.CurRegion, RHS, *SE))
return invalid<ReportNonAffBranch>(Context, /*Assert=*/true, &BB, LHS,
RHS);
}
// Allow loop exit conditions.
Loop *L = LI->getLoopFor(&BB);
if (L && L->getExitingBlock() == &BB)
return true;
// Allow perfectly nested conditions.
Region *R = RI->getRegionFor(&BB);
if (R->getEntry() != &BB)
return invalid<ReportCondition>(Context, /*Assert=*/true, &BB);
return true;
}
bool LoopIndexSplit::splitLoop() {
SplitCondition = NULL;
if (ExitCondition->getPredicate() == ICmpInst::ICMP_NE
|| ExitCondition->getPredicate() == ICmpInst::ICMP_EQ)
return false;
BasicBlock *Header = L->getHeader();
BasicBlock *Latch = L->getLoopLatch();
BranchInst *SBR = NULL; // Split Condition Branch
BranchInst *EBR = cast<BranchInst>(ExitCondition->getParent()->getTerminator());
// If Exiting block includes loop variant instructions then this
// loop may not be split safely.
BasicBlock *ExitingBlock = ExitCondition->getParent();
if (!cleanBlock(ExitingBlock)) return false;
LLVMContext &Context = Header->getContext();
for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
I != E; ++I) {
BranchInst *BR = dyn_cast<BranchInst>((*I)->getTerminator());
if (!BR || BR->isUnconditional()) continue;
ICmpInst *CI = dyn_cast<ICmpInst>(BR->getCondition());
if (!CI || CI == ExitCondition
|| CI->getPredicate() == ICmpInst::ICMP_NE
|| CI->getPredicate() == ICmpInst::ICMP_EQ)
continue;
// Unable to handle triangle loops at the moment.
// In triangle loop, split condition is in header and one of the
// the split destination is loop latch. If split condition is EQ
// then such loops are already handle in processOneIterationLoop().
if (Header == (*I)
&& (Latch == BR->getSuccessor(0) || Latch == BR->getSuccessor(1)))
continue;
// If the block does not dominate the latch then this is not a diamond.
// Such loop may not benefit from index split.
if (!DT->dominates((*I), Latch))
continue;
// If split condition branches heads do not have single predecessor,
// SplitCondBlock, then is not possible to remove inactive branch.
if (!BR->getSuccessor(0)->getSinglePredecessor()
|| !BR->getSuccessor(1)->getSinglePredecessor())
return false;
// If the merge point for BR is not loop latch then skip this condition.
if (BR->getSuccessor(0) != Latch) {
DominanceFrontier::iterator DF0 = DF->find(BR->getSuccessor(0));
assert (DF0 != DF->end() && "Unable to find dominance frontier");
if (!DF0->second.count(Latch))
continue;
}
if (BR->getSuccessor(1) != Latch) {
DominanceFrontier::iterator DF1 = DF->find(BR->getSuccessor(1));
assert (DF1 != DF->end() && "Unable to find dominance frontier");
if (!DF1->second.count(Latch))
continue;
}
SplitCondition = CI;
SBR = BR;
break;
}
if (!SplitCondition)
return false;
// If the predicate sign does not match then skip.
if (ExitCondition->isSigned() != SplitCondition->isSigned())
return false;
unsigned EVOpNum = (ExitCondition->getOperand(1) == IVExitValue);
unsigned SVOpNum = IVBasedValues.count(SplitCondition->getOperand(0));
Value *SplitValue = SplitCondition->getOperand(SVOpNum);
if (!L->isLoopInvariant(SplitValue))
return false;
if (!IVBasedValues.count(SplitCondition->getOperand(!SVOpNum)))
return false;
// Normalize loop conditions so that it is easier to calculate new loop
// bounds.
if (IVisGT(*ExitCondition) || IVisGE(*ExitCondition)) {
ExitCondition->setPredicate(ExitCondition->getInversePredicate());
BasicBlock *T = EBR->getSuccessor(0);
EBR->setSuccessor(0, EBR->getSuccessor(1));
EBR->setSuccessor(1, T);
}
if (IVisGT(*SplitCondition) || IVisGE(*SplitCondition)) {
SplitCondition->setPredicate(SplitCondition->getInversePredicate());
BasicBlock *T = SBR->getSuccessor(0);
SBR->setSuccessor(0, SBR->getSuccessor(1));
SBR->setSuccessor(1, T);
}
//[*] Calculate new loop bounds.
Value *AEV = SplitValue;
Value *BSV = SplitValue;
bool Sign = SplitCondition->isSigned();
Instruction *PHTerm = L->getLoopPreheader()->getTerminator();
if (IVisLT(*ExitCondition)) {
if (IVisLT(*SplitCondition)) {
/* Do nothing */
}
else if (IVisLE(*SplitCondition)) {
AEV = getPlusOne(SplitValue, Sign, PHTerm, Context);
BSV = getPlusOne(SplitValue, Sign, PHTerm, Context);
} else {
assert (0 && "Unexpected split condition!");
}
}
else if (IVisLE(*ExitCondition)) {
if (IVisLT(*SplitCondition)) {
AEV = getMinusOne(SplitValue, Sign, PHTerm, Context);
}
else if (IVisLE(*SplitCondition)) {
BSV = getPlusOne(SplitValue, Sign, PHTerm, Context);
} else {
assert (0 && "Unexpected split condition!");
}
} else {
assert (0 && "Unexpected exit condition!");
}
AEV = getMin(AEV, IVExitValue, Sign, PHTerm);
BSV = getMax(BSV, IVStartValue, Sign, PHTerm);
// [*] Clone Loop
DenseMap<const Value *, Value *> ValueMap;
Loop *BLoop = CloneLoop(L, LPM, LI, ValueMap, this);
Loop *ALoop = L;
// [*] ALoop's exiting edge enters BLoop's header.
// ALoop's original exit block becomes BLoop's exit block.
PHINode *B_IndVar = cast<PHINode>(ValueMap[IndVar]);
BasicBlock *A_ExitingBlock = ExitCondition->getParent();
BranchInst *A_ExitInsn =
dyn_cast<BranchInst>(A_ExitingBlock->getTerminator());
assert (A_ExitInsn && "Unable to find suitable loop exit branch");
BasicBlock *B_ExitBlock = A_ExitInsn->getSuccessor(1);
BasicBlock *B_Header = BLoop->getHeader();
if (ALoop->contains(B_ExitBlock)) {
B_ExitBlock = A_ExitInsn->getSuccessor(0);
A_ExitInsn->setSuccessor(0, B_Header);
} else
A_ExitInsn->setSuccessor(1, B_Header);
// [*] Update ALoop's exit value using new exit value.
ExitCondition->setOperand(EVOpNum, AEV);
// [*] Update BLoop's header phi nodes. Remove incoming PHINode's from
// original loop's preheader. Add incoming PHINode values from
// ALoop's exiting block. Update BLoop header's domiantor info.
// Collect inverse map of Header PHINodes.
DenseMap<Value *, Value *> InverseMap;
for (BasicBlock::iterator BI = ALoop->getHeader()->begin(),
BE = ALoop->getHeader()->end(); BI != BE; ++BI) {
if (PHINode *PN = dyn_cast<PHINode>(BI)) {
PHINode *PNClone = cast<PHINode>(ValueMap[PN]);
InverseMap[PNClone] = PN;
} else
break;
}
BasicBlock *A_Preheader = ALoop->getLoopPreheader();
for (BasicBlock::iterator BI = B_Header->begin(), BE = B_Header->end();
BI != BE; ++BI) {
if (PHINode *PN = dyn_cast<PHINode>(BI)) {
// Remove incoming value from original preheader.
PN->removeIncomingValue(A_Preheader);
// Add incoming value from A_ExitingBlock.
if (PN == B_IndVar)
PN->addIncoming(BSV, A_ExitingBlock);
else {
PHINode *OrigPN = cast<PHINode>(InverseMap[PN]);
Value *V2 = NULL;
// If loop header is also loop exiting block then
// OrigPN is incoming value for B loop header.
if (A_ExitingBlock == ALoop->getHeader())
V2 = OrigPN;
else
V2 = OrigPN->getIncomingValueForBlock(A_ExitingBlock);
PN->addIncoming(V2, A_ExitingBlock);
}
} else
break;
}
DT->changeImmediateDominator(B_Header, A_ExitingBlock);
DF->changeImmediateDominator(B_Header, A_ExitingBlock, DT);
// [*] Update BLoop's exit block. Its new predecessor is BLoop's exit
// block. Remove incoming PHINode values from ALoop's exiting block.
// Add new incoming values from BLoop's incoming exiting value.
// Update BLoop exit block's dominator info..
BasicBlock *B_ExitingBlock = cast<BasicBlock>(ValueMap[A_ExitingBlock]);
for (BasicBlock::iterator BI = B_ExitBlock->begin(), BE = B_ExitBlock->end();
BI != BE; ++BI) {
if (PHINode *PN = dyn_cast<PHINode>(BI)) {
PN->addIncoming(ValueMap[PN->getIncomingValueForBlock(A_ExitingBlock)],
B_ExitingBlock);
PN->removeIncomingValue(A_ExitingBlock);
} else
break;
}
DT->changeImmediateDominator(B_ExitBlock, B_ExitingBlock);
DF->changeImmediateDominator(B_ExitBlock, B_ExitingBlock, DT);
//[*] Split ALoop's exit edge. This creates a new block which
// serves two purposes. First one is to hold PHINode defnitions
// to ensure that ALoop's LCSSA form. Second use it to act
// as a preheader for BLoop.
BasicBlock *A_ExitBlock = SplitEdge(A_ExitingBlock, B_Header, this);
//[*] Preserve ALoop's LCSSA form. Create new forwarding PHINodes
// in A_ExitBlock to redefine outgoing PHI definitions from ALoop.
for(BasicBlock::iterator BI = B_Header->begin(), BE = B_Header->end();
BI != BE; ++BI) {
if (PHINode *PN = dyn_cast<PHINode>(BI)) {
Value *V1 = PN->getIncomingValueForBlock(A_ExitBlock);
PHINode *newPHI = PHINode::Create(PN->getType(), PN->getName());
newPHI->addIncoming(V1, A_ExitingBlock);
A_ExitBlock->getInstList().push_front(newPHI);
PN->removeIncomingValue(A_ExitBlock);
PN->addIncoming(newPHI, A_ExitBlock);
} else
break;
}
//[*] Eliminate split condition's inactive branch from ALoop.
BasicBlock *A_SplitCondBlock = SplitCondition->getParent();
BranchInst *A_BR = cast<BranchInst>(A_SplitCondBlock->getTerminator());
BasicBlock *A_InactiveBranch = NULL;
BasicBlock *A_ActiveBranch = NULL;
A_ActiveBranch = A_BR->getSuccessor(0);
A_InactiveBranch = A_BR->getSuccessor(1);
A_BR->setUnconditionalDest(A_ActiveBranch);
removeBlocks(A_InactiveBranch, L, A_ActiveBranch);
//[*] Eliminate split condition's inactive branch in from BLoop.
BasicBlock *B_SplitCondBlock = cast<BasicBlock>(ValueMap[A_SplitCondBlock]);
BranchInst *B_BR = cast<BranchInst>(B_SplitCondBlock->getTerminator());
BasicBlock *B_InactiveBranch = NULL;
BasicBlock *B_ActiveBranch = NULL;
B_ActiveBranch = B_BR->getSuccessor(1);
B_InactiveBranch = B_BR->getSuccessor(0);
B_BR->setUnconditionalDest(B_ActiveBranch);
removeBlocks(B_InactiveBranch, BLoop, B_ActiveBranch);
BasicBlock *A_Header = ALoop->getHeader();
if (A_ExitingBlock == A_Header)
return true;
//[*] Move exit condition into split condition block to avoid
// executing dead loop iteration.
ICmpInst *B_ExitCondition = cast<ICmpInst>(ValueMap[ExitCondition]);
Instruction *B_IndVarIncrement = cast<Instruction>(ValueMap[IVIncrement]);
ICmpInst *B_SplitCondition = cast<ICmpInst>(ValueMap[SplitCondition]);
moveExitCondition(A_SplitCondBlock, A_ActiveBranch, A_ExitBlock, ExitCondition,
cast<ICmpInst>(SplitCondition), IndVar, IVIncrement,
ALoop, EVOpNum);
moveExitCondition(B_SplitCondBlock, B_ActiveBranch,
B_ExitBlock, B_ExitCondition,
B_SplitCondition, B_IndVar, B_IndVarIncrement,
BLoop, EVOpNum);
NumIndexSplit++;
return true;
}
/// updateLoopIterationSpace -- Update loop's iteration space if loop
/// body is executed for certain IV range only. For example,
///
/// for (i = 0; i < N; ++i) {
/// if ( i > A && i < B) {
/// ...
/// }
/// }
/// is transformed to iterators from A to B, if A > 0 and B < N.
///
bool LoopIndexSplit::updateLoopIterationSpace() {
SplitCondition = NULL;
if (ExitCondition->getPredicate() == ICmpInst::ICMP_NE
|| ExitCondition->getPredicate() == ICmpInst::ICMP_EQ)
return false;
BasicBlock *Latch = L->getLoopLatch();
BasicBlock *Header = L->getHeader();
BranchInst *BR = dyn_cast<BranchInst>(Header->getTerminator());
if (!BR) return false;
if (!isa<BranchInst>(Latch->getTerminator())) return false;
if (BR->isUnconditional()) return false;
BinaryOperator *AND = dyn_cast<BinaryOperator>(BR->getCondition());
if (!AND) return false;
if (AND->getOpcode() != Instruction::And) return false;
ICmpInst *Op0 = dyn_cast<ICmpInst>(AND->getOperand(0));
ICmpInst *Op1 = dyn_cast<ICmpInst>(AND->getOperand(1));
if (!Op0 || !Op1)
return false;
IVBasedValues.insert(AND);
IVBasedValues.insert(Op0);
IVBasedValues.insert(Op1);
if (!cleanBlock(Header)) return false;
BasicBlock *ExitingBlock = ExitCondition->getParent();
if (!cleanBlock(ExitingBlock)) return false;
// If the merge point for BR is not loop latch then skip this loop.
if (BR->getSuccessor(0) != Latch) {
DominanceFrontier::iterator DF0 = DF->find(BR->getSuccessor(0));
assert (DF0 != DF->end() && "Unable to find dominance frontier");
if (!DF0->second.count(Latch))
return false;
}
if (BR->getSuccessor(1) != Latch) {
DominanceFrontier::iterator DF1 = DF->find(BR->getSuccessor(1));
assert (DF1 != DF->end() && "Unable to find dominance frontier");
if (!DF1->second.count(Latch))
return false;
}
// Verify that loop exiting block has only two predecessor, where one pred
// is split condition block. The other predecessor will become exiting block's
// dominator after CFG is updated. TODO : Handle CFG's where exiting block has
// more then two predecessors. This requires extra work in updating dominator
// information.
BasicBlock *ExitingBBPred = NULL;
for (pred_iterator PI = pred_begin(ExitingBlock), PE = pred_end(ExitingBlock);
PI != PE; ++PI) {
BasicBlock *BB = *PI;
if (Header == BB)
continue;
if (ExitingBBPred)
return false;
else
ExitingBBPred = BB;
}
if (!restrictLoopBound(*Op0))
return false;
if (!restrictLoopBound(*Op1))
return false;
// Update CFG.
if (BR->getSuccessor(0) == ExitingBlock)
BR->setUnconditionalDest(BR->getSuccessor(1));
else
BR->setUnconditionalDest(BR->getSuccessor(0));
AND->eraseFromParent();
if (Op0->use_empty())
Op0->eraseFromParent();
if (Op1->use_empty())
Op1->eraseFromParent();
// Update domiantor info. Now, ExitingBlock has only one predecessor,
// ExitingBBPred, and it is ExitingBlock's immediate domiantor.
DT->changeImmediateDominator(ExitingBlock, ExitingBBPred);
BasicBlock *ExitBlock = ExitingBlock->getTerminator()->getSuccessor(1);
if (L->contains(ExitBlock))
ExitBlock = ExitingBlock->getTerminator()->getSuccessor(0);
// If ExitingBlock is a member of the loop basic blocks' DF list then
// replace ExitingBlock with header and exit block in the DF list
DominanceFrontier::iterator ExitingBlockDF = DF->find(ExitingBlock);
for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
I != E; ++I) {
BasicBlock *BB = *I;
if (BB == Header || BB == ExitingBlock)
continue;
DominanceFrontier::iterator BBDF = DF->find(BB);
DominanceFrontier::DomSetType::iterator DomSetI = BBDF->second.begin();
DominanceFrontier::DomSetType::iterator DomSetE = BBDF->second.end();
while (DomSetI != DomSetE) {
DominanceFrontier::DomSetType::iterator CurrentItr = DomSetI;
++DomSetI;
BasicBlock *DFBB = *CurrentItr;
if (DFBB == ExitingBlock) {
BBDF->second.erase(DFBB);
for (DominanceFrontier::DomSetType::iterator
EBI = ExitingBlockDF->second.begin(),
EBE = ExitingBlockDF->second.end(); EBI != EBE; ++EBI)
BBDF->second.insert(*EBI);
}
}
}
NumRestrictBounds++;
return true;
}
/// processOneIterationLoop -- Eliminate loop if loop body is executed
/// only once. For example,
/// for (i = 0; i < N; ++i) {
/// if ( i == X) {
/// ...
/// }
/// }
///
bool LoopIndexSplit::processOneIterationLoop() {
SplitCondition = NULL;
BasicBlock *Latch = L->getLoopLatch();
BasicBlock *Header = L->getHeader();
BranchInst *BR = dyn_cast<BranchInst>(Header->getTerminator());
if (!BR) return false;
if (!isa<BranchInst>(Latch->getTerminator())) return false;
if (BR->isUnconditional()) return false;
SplitCondition = dyn_cast<ICmpInst>(BR->getCondition());
if (!SplitCondition) return false;
if (SplitCondition == ExitCondition) return false;
if (SplitCondition->getPredicate() != ICmpInst::ICMP_EQ) return false;
if (BR->getOperand(1) != Latch) return false;
if (!IVBasedValues.count(SplitCondition->getOperand(0))
&& !IVBasedValues.count(SplitCondition->getOperand(1)))
return false;
// If IV is used outside the loop then this loop traversal is required.
// FIXME: Calculate and use last IV value.
if (isUsedOutsideLoop(IVIncrement, L))
return false;
// If BR operands are not IV or not loop invariants then skip this loop.
Value *OPV = SplitCondition->getOperand(0);
Value *SplitValue = SplitCondition->getOperand(1);
if (!L->isLoopInvariant(SplitValue))
std::swap(OPV, SplitValue);
if (!L->isLoopInvariant(SplitValue))
return false;
Instruction *OPI = dyn_cast<Instruction>(OPV);
if (!OPI)
return false;
if (OPI->getParent() != Header || isUsedOutsideLoop(OPI, L))
return false;
Value *StartValue = IVStartValue;
Value *ExitValue = IVExitValue;;
if (OPV != IndVar) {
// If BR operand is IV based then use this operand to calculate
// effective conditions for loop body.
BinaryOperator *BOPV = dyn_cast<BinaryOperator>(OPV);
if (!BOPV)
return false;
if (BOPV->getOpcode() != Instruction::Add)
return false;
StartValue = BinaryOperator::CreateAdd(OPV, StartValue, "" , BR);
ExitValue = BinaryOperator::CreateAdd(OPV, ExitValue, "" , BR);
}
if (!cleanBlock(Header))
return false;
if (!cleanBlock(Latch))
return false;
// If the merge point for BR is not loop latch then skip this loop.
if (BR->getSuccessor(0) != Latch) {
DominanceFrontier::iterator DF0 = DF->find(BR->getSuccessor(0));
assert (DF0 != DF->end() && "Unable to find dominance frontier");
if (!DF0->second.count(Latch))
return false;
}
if (BR->getSuccessor(1) != Latch) {
DominanceFrontier::iterator DF1 = DF->find(BR->getSuccessor(1));
assert (DF1 != DF->end() && "Unable to find dominance frontier");
if (!DF1->second.count(Latch))
return false;
}
// Now, Current loop L contains compare instruction
// that compares induction variable, IndVar, against loop invariant. And
// entire (i.e. meaningful) loop body is dominated by this compare
// instruction. In such case eliminate
// loop structure surrounding this loop body. For example,
// for (int i = start; i < end; ++i) {
// if ( i == somevalue) {
// loop_body
// }
// }
// can be transformed into
// if (somevalue >= start && somevalue < end) {
// i = somevalue;
// loop_body
// }
// Replace index variable with split value in loop body. Loop body is executed
// only when index variable is equal to split value.
IndVar->replaceAllUsesWith(SplitValue);
// Replace split condition in header.
// Transform
// SplitCondition : icmp eq i32 IndVar, SplitValue
// into
// c1 = icmp uge i32 SplitValue, StartValue
// c2 = icmp ult i32 SplitValue, ExitValue
// and i32 c1, c2
Instruction *C1 = new ICmpInst(BR, ExitCondition->isSigned() ?
ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE,
SplitValue, StartValue, "lisplit");
CmpInst::Predicate C2P = ExitCondition->getPredicate();
BranchInst *LatchBR = cast<BranchInst>(Latch->getTerminator());
if (LatchBR->getOperand(1) != Header)
C2P = CmpInst::getInversePredicate(C2P);
Instruction *C2 = new ICmpInst(BR, C2P, SplitValue, ExitValue, "lisplit");
Instruction *NSplitCond = BinaryOperator::CreateAnd(C1, C2, "lisplit", BR);
SplitCondition->replaceAllUsesWith(NSplitCond);
SplitCondition->eraseFromParent();
// Remove Latch to Header edge.
BasicBlock *LatchSucc = NULL;
Header->removePredecessor(Latch);
for (succ_iterator SI = succ_begin(Latch), E = succ_end(Latch);
SI != E; ++SI) {
if (Header != *SI)
LatchSucc = *SI;
}
// Clean up latch block.
Value *LatchBRCond = LatchBR->getCondition();
LatchBR->setUnconditionalDest(LatchSucc);
RecursivelyDeleteTriviallyDeadInstructions(LatchBRCond);
LPM->deleteLoopFromQueue(L);
// Update Dominator Info.
// Only CFG change done is to remove Latch to Header edge. This
// does not change dominator tree because Latch did not dominate
// Header.
if (DF) {
DominanceFrontier::iterator HeaderDF = DF->find(Header);
if (HeaderDF != DF->end())
DF->removeFromFrontier(HeaderDF, Header);
DominanceFrontier::iterator LatchDF = DF->find(Latch);
if (LatchDF != DF->end())
DF->removeFromFrontier(LatchDF, Header);
}
++NumIndexSplitRemoved;
return true;
}
// Index split Loop L. Return true if loop is split.
bool LoopIndexSplit::runOnLoop(Loop *IncomingLoop, LPPassManager &LPM_Ref) {
L = IncomingLoop;
LPM = &LPM_Ref;
// If LoopSimplify form is not available, stay out of trouble.
if (!L->isLoopSimplifyForm())
return false;
// FIXME - Nested loops make dominator info updates tricky.
if (!L->getSubLoops().empty())
return false;
DT = &getAnalysis<DominatorTree>();
LI = &getAnalysis<LoopInfo>();
DF = &getAnalysis<DominanceFrontier>();
// Initialize loop data.
IndVar = L->getCanonicalInductionVariable();
if (!IndVar) return false;
bool P1InLoop = L->contains(IndVar->getIncomingBlock(1));
IVStartValue = IndVar->getIncomingValue(!P1InLoop);
IVIncrement = dyn_cast<Instruction>(IndVar->getIncomingValue(P1InLoop));
if (!IVIncrement) return false;
IVBasedValues.clear();
IVBasedValues.insert(IndVar);
IVBasedValues.insert(IVIncrement);
for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
I != E; ++I)
for(BasicBlock::iterator BI = (*I)->begin(), BE = (*I)->end();
BI != BE; ++BI) {
if (BinaryOperator *BO = dyn_cast<BinaryOperator>(BI))
if (BO != IVIncrement
&& (BO->getOpcode() == Instruction::Add
|| BO->getOpcode() == Instruction::Sub))
if (IVBasedValues.count(BO->getOperand(0))
&& L->isLoopInvariant(BO->getOperand(1)))
IVBasedValues.insert(BO);
}
// Reject loop if loop exit condition is not suitable.
BasicBlock *ExitingBlock = L->getExitingBlock();
if (!ExitingBlock)
return false;
BranchInst *EBR = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
if (!EBR) return false;
ExitCondition = dyn_cast<ICmpInst>(EBR->getCondition());
if (!ExitCondition) return false;
if (ExitingBlock != L->getLoopLatch()) return false;
IVExitValue = ExitCondition->getOperand(1);
if (!L->isLoopInvariant(IVExitValue))
IVExitValue = ExitCondition->getOperand(0);
if (!L->isLoopInvariant(IVExitValue))
return false;
if (!IVBasedValues.count(
ExitCondition->getOperand(IVExitValue == ExitCondition->getOperand(0))))
return false;
// If start value is more then exit value where induction variable
// increments by 1 then we are potentially dealing with an infinite loop.
// Do not index split this loop.
if (ConstantInt *SV = dyn_cast<ConstantInt>(IVStartValue))
if (ConstantInt *EV = dyn_cast<ConstantInt>(IVExitValue))
if (SV->getSExtValue() > EV->getSExtValue())
return false;
if (processOneIterationLoop())
return true;
if (updateLoopIterationSpace())
return true;
if (splitLoop())
return true;
return false;
}
bool RangedAddressSanitizer::runOnFunction(Function &F) {
if (getenv("FASAN_DISABLE")) {
SPM_DEBUG( dbgs() << "FASan : disabled\n" );
return false;
}
DL_ = &getAnalysis<DataLayout>();
DT_ = &getAnalysis<DominatorTree>();
LI_ = &getAnalysis<LoopInfo>();
RI_ = &getAnalysis<ReduceIndexation>();
#ifdef ENABLE_REUSE
RE_ = &getAnalysis<RelativeExecutions>();
#endif
RMM_ = &getAnalysis<RelativeMinMax>();
Module_ = F.getParent();
Context_ = &Module_->getContext();
Type *VoidTy = Type::getVoidTy(*Context_);
IntegerType *IntTy = IntegerType::getInt64Ty(*Context_);
IntegerType *BoolTy = IntegerType::getInt1Ty(*Context_);
PointerType *IntPtrTy = PointerType::getUnqual(IntTy);
PointerType *VoidPtrTy = PointerType::getInt8PtrTy(*Context_);
SPM_DEBUG( F.dump() );
outs() << "[IterationInfo]\n";
for (Loop * loop : *LI_) {
ii_visitLoop(loop);
}
outs() << "[EndOfIterationInfo]\n";
#if 0 // disabled initialization,shutdown sequence for FASan
if (F.getName() == "main") {
SPM_DEBUG(dbgs() << "RangedAddressSanitizer: inserting hwloc calls into "
"main function\n");
FunctionType *FnType = FunctionType::get(VoidTy, ArrayRef<Type*>(), false);
IRBuilder<> IRB(&(*F.getEntryBlock().begin()));
Constant *Init = Module_->getOrInsertFunction("__spm_init", FnType);
IRB.CreateCall(Init);
Constant *End = Module_->getOrInsertFunction("__spm_end", FnType);
for (auto &BB : F) {
TerminatorInst *TI = BB.getTerminator();
if (isa<ReturnInst>(TI)) {
IRB.SetInsertPoint(TI);
IRB.CreateCall(End);
}
}
}
#endif
if (!ClFunc.empty() && F.getName() != ClFunc) {
SPM_DEBUG(dbgs() << "RangedAddressSanitizer: skipping function "
<< F.getName() << "\n");
return false;
}
Calls_.clear();
SPM_DEBUG(dbgs() << "RangedAddressSanitizer: processing function "
<< F.getName() << "\n");
std::vector<Type*> ReuseFnFormals = { VoidPtrTy, IntTy, IntTy, IntTy };
FunctionType *ReuseFnType = FunctionType::get(BoolTy, ReuseFnFormals, false);
ReuseFn_ =
F.getParent()->getOrInsertFunction("__fasan_check", ReuseFnType);
ReuseFnDestroy_ =
F.getParent()->getOrInsertFunction("__spm_give", ReuseFnType);
// Visit all loops in bottom-up order (innter-most loops first)
std::set<BasicBlock*> Processed;
auto Entry = DT_->getRootNode();
for (auto ET = po_begin(Entry), EE = po_end(Entry); ET != EE; ++ET) {
BasicBlock *Header = (*ET)->getBlock();
if (LI_->isLoopHeader(Header)) {
SPM_DEBUG(dbgs() << "RangedAddressSanitizer: processing loop at "
<< Header->getName() << "\n");
Loop *L = LI_->getLoopFor(Header);
if (L->getNumBackEdges() != 1 ||
std::distance(pred_begin(Header), pred_end(Header)) != 2) {
SPM_DEBUG(dbgs() << "RangedAddressSanitizer: loop has multiple "
<< "backedges or multiple incoming outer blocks\n");
continue;
}
SPM_DEBUG(dbgs() << "RangedAddressSanitizer: processing loop at "
<< Header->getName() << "\n");
// visit all memory acccesses in this loop
for (auto BB = L->block_begin(), BE = L->block_end(); BB != BE; ++BB) {
if (!Processed.count(*BB)) {
Processed.insert(*BB);
for (auto &I : *(*BB))
generateCallFor(L, &I);
}
}
}
}
// FAsan logic goes here
std::map<const BasicBlock*,BasicBlock*> clonedBlockMap; // keeps track of cloned regions to avoid redundant cloning
std::vector<CallInst*> ToInline;
for (auto &CI : Calls_) {
BasicBlock * Preheader = CI.Preheader;
// TODO decide whether it is worthwhile to optimize for this case
// insert range check
IRBuilder<> IRB(Preheader->getTerminator());
Value *VoidArray = IRB.CreateBitCast(CI.Array, VoidPtrTy);
std::vector<Value*> Args = { VoidArray, CI.Min, CI.Max, CI.Reuse };
CallInst *CR = IRB.CreateCall(ReuseFn_, Args);
ToInline.push_back(CR);
// verify if this loop was already instrumented
TerminatorInst * preHeaderTerm = CR->getParent()->getTerminator();
BranchInst * preHeaderBranch = dyn_cast<BranchInst>(preHeaderTerm);
if (preHeaderBranch && preHeaderBranch->isConditional()) {
// discover the structure of the instrumented code (safe and default region)
// abort, if this does not look like instrumented code
BasicBlock * firstTarget = preHeaderBranch->getSuccessor(0);
BasicBlock * secondTarget = preHeaderBranch->getSuccessor(1);
BasicBlock * safeHeader, * defHeader;
if (clonedBlockMap.count(firstTarget)) {
defHeader = firstTarget;
safeHeader = clonedBlockMap[firstTarget];
assert(safeHeader == secondTarget);
} else {
assert(clonedBlockMap.count(secondTarget));
defHeader = secondTarget;
safeHeader = clonedBlockMap[secondTarget];
assert(safeHeader == firstTarget);
}
SPM_DEBUG( dbgs() << "FASan: (Unsupported) second array in safe region controlled by " << * preHeaderBranch << "\n" );
Loop * defLoop = LI_->getLoopFor(defHeader);
assert(defLoop && "default region is not a loop!");
Loop::block_iterator itBodyBlock,S,E;
S = defLoop->block_begin();
E = defLoop->block_end();
// mark accesses in cloned region as safe
for (itBodyBlock = S;itBodyBlock != E; ++itBodyBlock) {
BasicBlock * defBodyBlock = *itBodyBlock;
BasicBlock * safeBodyBlock = clonedBlockMap[defBodyBlock];
for(auto & inst : *safeBodyBlock) {
markSafeArrayUse(&inst, CI.Array);
}
}
// add conjunctive test
Value * oldCond = preHeaderBranch->getCondition();
Value * joinedCond = IRB.CreateAnd(oldCond, CR, "allsafe");
preHeaderBranch->setCondition(joinedCond);
} else {
// get loop
Loop* finalLoop = CI.FinalLoop;
Loop::block_iterator itBodyBlock,S,E;
S = finalLoop->block_begin();
E = finalLoop->block_end();
// clone loop body (cloned loop will run unchecked)
ValueToValueMapTy cloneMap;
BasicBlock * clonedHeader = 0;
std::vector<BasicBlock*> clonedBlocks;
for (itBodyBlock = S;itBodyBlock != E; ++itBodyBlock) {
const BasicBlock * bodyBlock = *itBodyBlock;
BasicBlock * clonedBlock = CloneBasicBlock(bodyBlock, cloneMap, "_checked", &F, 0);
cloneMap[bodyBlock] = clonedBlock;
clonedBlockMap[bodyBlock] = clonedBlock;
clonedBlocks.push_back(clonedBlock);
if (bodyBlock == finalLoop->getHeader()) {
clonedHeader = clonedBlock;
SPM_DEBUG( dbgs() << "FASan: loop header case at " << bodyBlock->getName() << "\n" );
} else {
SPM_DEBUG( dbgs() << "FASan: non-header block at " << bodyBlock->getName() << "\n" );
}
}
if (!clonedHeader) {
// TODO run clean-up code
SPM_DEBUG( dbgs() << "FASan: could not find header!\n");
abort();
}
// Remap uses inside cloned region (mark pointers in the region as unguarded)
for (BasicBlock * block : clonedBlocks) {
for(auto & inst : *block) {
RemapInstruction(&inst, cloneMap, RF_IgnoreMissingEntries);
markSafeArrayUse(&inst, CI.Array);
}
}
// TODO fix PHI-nodes in exit blocks
// Rewire terminator of the range check to branch to the cloned region
TerminatorInst * checkTermInst = CR->getParent()->getTerminator();
if (BranchInst * checkBranchInst = dyn_cast<BranchInst>(checkTermInst)) {
if (checkBranchInst->isUnconditional()) {
BasicBlock * defTarget = checkBranchInst->getSuccessor(0);
BranchInst * modifiedBranchInst = BranchInst::Create(clonedHeader, defTarget, CR, checkBranchInst);
checkBranchInst->replaceAllUsesWith(modifiedBranchInst);
checkBranchInst->eraseFromParent();
} else {
SPM_DEBUG( dbgs() << "FASan: Unexpected conditional branch (preheader should branch unconditional, other array checks will introduce conditional branches) " << * checkTermInst << "\n" );
abort();
}
} else {
SPM_DEBUG( dbgs() << "FASan: unsupported terminator type " << * checkTermInst << "\n" );
abort();
}
}
#if 0
IRB.SetInsertPoint(&(*CI.Final->begin()));
IRB.CreateCall(ReuseFnDestroy_, Args);
#endif
SPM_DEBUG(dbgs() << "RangedAddressSanitizer: call instruction: " << *CR
<< "\n");
}
// inline calls
#ifdef FASAN_INLINE_RUNTIME
for (CallInst * call : ToInline) {
assert(call);
InlineFunctionInfo IFI;
InlineFunction(call, IFI, false);
}
#endif
SPM_DEBUG( F.dump() );
return true;
}
/// \brief Simplify one loop and queue further loops for simplification.
///
/// FIXME: Currently this accepts both lots of analyses that it uses and a raw
/// Pass pointer. The Pass pointer is used by numerous utilities to update
/// specific analyses. Rather than a pass it would be much cleaner and more
/// explicit if they accepted the analysis directly and then updated it.
static bool simplifyOneLoop(Loop *L, SmallVectorImpl<Loop *> &Worklist,
DominatorTree *DT, LoopInfo *LI,
ScalarEvolution *SE, Pass *PP,
AssumptionCache *AC) {
bool Changed = false;
ReprocessLoop:
// Check to see that no blocks (other than the header) in this loop have
// predecessors that are not in the loop. This is not valid for natural
// loops, but can occur if the blocks are unreachable. Since they are
// unreachable we can just shamelessly delete those CFG edges!
for (Loop::block_iterator BB = L->block_begin(), E = L->block_end();
BB != E; ++BB) {
if (*BB == L->getHeader()) continue;
SmallPtrSet<BasicBlock*, 4> BadPreds;
for (pred_iterator PI = pred_begin(*BB),
PE = pred_end(*BB); PI != PE; ++PI) {
BasicBlock *P = *PI;
if (!L->contains(P))
BadPreds.insert(P);
}
// Delete each unique out-of-loop (and thus dead) predecessor.
for (BasicBlock *P : BadPreds) {
DEBUG(dbgs() << "LoopSimplify: Deleting edge from dead predecessor "
<< P->getName() << "\n");
// Inform each successor of each dead pred.
for (succ_iterator SI = succ_begin(P), SE = succ_end(P); SI != SE; ++SI)
(*SI)->removePredecessor(P);
// Zap the dead pred's terminator and replace it with unreachable.
TerminatorInst *TI = P->getTerminator();
TI->replaceAllUsesWith(UndefValue::get(TI->getType()));
P->getTerminator()->eraseFromParent();
new UnreachableInst(P->getContext(), P);
Changed = true;
}
}
// If there are exiting blocks with branches on undef, resolve the undef in
// the direction which will exit the loop. This will help simplify loop
// trip count computations.
SmallVector<BasicBlock*, 8> ExitingBlocks;
L->getExitingBlocks(ExitingBlocks);
for (SmallVectorImpl<BasicBlock *>::iterator I = ExitingBlocks.begin(),
E = ExitingBlocks.end(); I != E; ++I)
if (BranchInst *BI = dyn_cast<BranchInst>((*I)->getTerminator()))
if (BI->isConditional()) {
if (UndefValue *Cond = dyn_cast<UndefValue>(BI->getCondition())) {
DEBUG(dbgs() << "LoopSimplify: Resolving \"br i1 undef\" to exit in "
<< (*I)->getName() << "\n");
BI->setCondition(ConstantInt::get(Cond->getType(),
!L->contains(BI->getSuccessor(0))));
// This may make the loop analyzable, force SCEV recomputation.
if (SE)
SE->forgetLoop(L);
Changed = true;
}
}
// Does the loop already have a preheader? If so, don't insert one.
BasicBlock *Preheader = L->getLoopPreheader();
if (!Preheader) {
Preheader = InsertPreheaderForLoop(L, PP);
if (Preheader) {
++NumInserted;
Changed = true;
}
}
// Next, check to make sure that all exit nodes of the loop only have
// predecessors that are inside of the loop. This check guarantees that the
// loop preheader/header will dominate the exit blocks. If the exit block has
// predecessors from outside of the loop, split the edge now.
SmallVector<BasicBlock*, 8> ExitBlocks;
L->getExitBlocks(ExitBlocks);
SmallSetVector<BasicBlock *, 8> ExitBlockSet(ExitBlocks.begin(),
ExitBlocks.end());
for (SmallSetVector<BasicBlock *, 8>::iterator I = ExitBlockSet.begin(),
E = ExitBlockSet.end(); I != E; ++I) {
BasicBlock *ExitBlock = *I;
for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock);
PI != PE; ++PI)
// Must be exactly this loop: no subloops, parent loops, or non-loop preds
// allowed.
if (!L->contains(*PI)) {
if (rewriteLoopExitBlock(L, ExitBlock, DT, LI, PP)) {
++NumInserted;
Changed = true;
}
break;
}
}
// If the header has more than two predecessors at this point (from the
// preheader and from multiple backedges), we must adjust the loop.
BasicBlock *LoopLatch = L->getLoopLatch();
if (!LoopLatch) {
// If this is really a nested loop, rip it out into a child loop. Don't do
// this for loops with a giant number of backedges, just factor them into a
// common backedge instead.
if (L->getNumBackEdges() < 8) {
if (Loop *OuterL = separateNestedLoop(L, Preheader, DT, LI, SE, PP, AC)) {
++NumNested;
// Enqueue the outer loop as it should be processed next in our
// depth-first nest walk.
Worklist.push_back(OuterL);
// This is a big restructuring change, reprocess the whole loop.
Changed = true;
// GCC doesn't tail recursion eliminate this.
// FIXME: It isn't clear we can't rely on LLVM to TRE this.
goto ReprocessLoop;
}
}
// If we either couldn't, or didn't want to, identify nesting of the loops,
// insert a new block that all backedges target, then make it jump to the
// loop header.
LoopLatch = insertUniqueBackedgeBlock(L, Preheader, DT, LI);
if (LoopLatch) {
++NumInserted;
Changed = true;
}
}
const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
// Scan over the PHI nodes in the loop header. Since they now have only two
// incoming values (the loop is canonicalized), we may have simplified the PHI
// down to 'X = phi [X, Y]', which should be replaced with 'Y'.
PHINode *PN;
for (BasicBlock::iterator I = L->getHeader()->begin();
(PN = dyn_cast<PHINode>(I++)); )
if (Value *V = SimplifyInstruction(PN, DL, nullptr, DT, AC)) {
if (SE) SE->forgetValue(PN);
PN->replaceAllUsesWith(V);
PN->eraseFromParent();
}
// If this loop has multiple exits and the exits all go to the same
// block, attempt to merge the exits. This helps several passes, such
// as LoopRotation, which do not support loops with multiple exits.
// SimplifyCFG also does this (and this code uses the same utility
// function), however this code is loop-aware, where SimplifyCFG is
// not. That gives it the advantage of being able to hoist
// loop-invariant instructions out of the way to open up more
// opportunities, and the disadvantage of having the responsibility
// to preserve dominator information.
bool UniqueExit = true;
if (!ExitBlocks.empty())
for (unsigned i = 1, e = ExitBlocks.size(); i != e; ++i)
if (ExitBlocks[i] != ExitBlocks[0]) {
UniqueExit = false;
break;
}
if (UniqueExit) {
for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) {
BasicBlock *ExitingBlock = ExitingBlocks[i];
if (!ExitingBlock->getSinglePredecessor()) continue;
BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator());
if (!BI || !BI->isConditional()) continue;
CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition());
if (!CI || CI->getParent() != ExitingBlock) continue;
// Attempt to hoist out all instructions except for the
// comparison and the branch.
bool AllInvariant = true;
bool AnyInvariant = false;
for (BasicBlock::iterator I = ExitingBlock->begin(); &*I != BI; ) {
Instruction *Inst = I++;
// Skip debug info intrinsics.
if (isa<DbgInfoIntrinsic>(Inst))
continue;
if (Inst == CI)
continue;
if (!L->makeLoopInvariant(Inst, AnyInvariant,
Preheader ? Preheader->getTerminator()
: nullptr)) {
AllInvariant = false;
break;
}
}
if (AnyInvariant) {
Changed = true;
// The loop disposition of all SCEV expressions that depend on any
// hoisted values have also changed.
if (SE)
SE->forgetLoopDispositions(L);
}
if (!AllInvariant) continue;
// The block has now been cleared of all instructions except for
// a comparison and a conditional branch. SimplifyCFG may be able
// to fold it now.
if (!FoldBranchToCommonDest(BI))
continue;
// Success. The block is now dead, so remove it from the loop,
// update the dominator tree and delete it.
DEBUG(dbgs() << "LoopSimplify: Eliminating exiting block "
<< ExitingBlock->getName() << "\n");
// Notify ScalarEvolution before deleting this block. Currently assume the
// parent loop doesn't change (spliting edges doesn't count). If blocks,
// CFG edges, or other values in the parent loop change, then we need call
// to forgetLoop() for the parent instead.
if (SE)
SE->forgetLoop(L);
assert(pred_begin(ExitingBlock) == pred_end(ExitingBlock));
Changed = true;
LI->removeBlock(ExitingBlock);
DomTreeNode *Node = DT->getNode(ExitingBlock);
const std::vector<DomTreeNodeBase<BasicBlock> *> &Children =
Node->getChildren();
while (!Children.empty()) {
DomTreeNode *Child = Children.front();
DT->changeImmediateDominator(Child, Node->getIDom());
}
DT->eraseNode(ExitingBlock);
BI->getSuccessor(0)->removePredecessor(ExitingBlock);
BI->getSuccessor(1)->removePredecessor(ExitingBlock);
ExitingBlock->eraseFromParent();
}
}
return Changed;
}
/// GetIfCondition - Given a basic block (BB) with two predecessors,
/// check to see if the merge at this block is due
/// to an "if condition". If so, return the boolean condition that determines
/// which entry into BB will be taken. Also, return by references the block
/// that will be entered from if the condition is true, and the block that will
/// be entered if the condition is false.
///
/// This does no checking to see if the true/false blocks have large or unsavory
/// instructions in them.
Value *llvm::GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue,
BasicBlock *&IfFalse) {
PHINode *SomePHI = dyn_cast<PHINode>(BB->begin());
BasicBlock *Pred1 = NULL;
BasicBlock *Pred2 = NULL;
if (SomePHI) {
if (SomePHI->getNumIncomingValues() != 2)
return NULL;
Pred1 = SomePHI->getIncomingBlock(0);
Pred2 = SomePHI->getIncomingBlock(1);
} else {
pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
if (PI == PE) // No predecessor
return NULL;
Pred1 = *PI++;
if (PI == PE) // Only one predecessor
return NULL;
Pred2 = *PI++;
if (PI != PE) // More than two predecessors
return NULL;
}
// We can only handle branches. Other control flow will be lowered to
// branches if possible anyway.
BranchInst *Pred1Br = dyn_cast<BranchInst>(Pred1->getTerminator());
BranchInst *Pred2Br = dyn_cast<BranchInst>(Pred2->getTerminator());
if (Pred1Br == 0 || Pred2Br == 0)
return 0;
// Eliminate code duplication by ensuring that Pred1Br is conditional if
// either are.
if (Pred2Br->isConditional()) {
// If both branches are conditional, we don't have an "if statement". In
// reality, we could transform this case, but since the condition will be
// required anyway, we stand no chance of eliminating it, so the xform is
// probably not profitable.
if (Pred1Br->isConditional())
return 0;
std::swap(Pred1, Pred2);
std::swap(Pred1Br, Pred2Br);
}
if (Pred1Br->isConditional()) {
// The only thing we have to watch out for here is to make sure that Pred2
// doesn't have incoming edges from other blocks. If it does, the condition
// doesn't dominate BB.
if (Pred2->getSinglePredecessor() == 0)
return 0;
// If we found a conditional branch predecessor, make sure that it branches
// to BB and Pred2Br. If it doesn't, this isn't an "if statement".
if (Pred1Br->getSuccessor(0) == BB &&
Pred1Br->getSuccessor(1) == Pred2) {
IfTrue = Pred1;
IfFalse = Pred2;
} else if (Pred1Br->getSuccessor(0) == Pred2 &&
Pred1Br->getSuccessor(1) == BB) {
IfTrue = Pred2;
IfFalse = Pred1;
} else {
// We know that one arm of the conditional goes to BB, so the other must
// go somewhere unrelated, and this must not be an "if statement".
return 0;
}
return Pred1Br->getCondition();
}
// Ok, if we got here, both predecessors end with an unconditional branch to
// BB. Don't panic! If both blocks only have a single (identical)
// predecessor, and THAT is a conditional branch, then we're all ok!
BasicBlock *CommonPred = Pred1->getSinglePredecessor();
if (CommonPred == 0 || CommonPred != Pred2->getSinglePredecessor())
return 0;
// Otherwise, if this is a conditional branch, then we can use it!
BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator());
if (BI == 0) return 0;
assert(BI->isConditional() && "Two successors but not conditional?");
if (BI->getSuccessor(0) == Pred1) {
IfTrue = Pred1;
IfFalse = Pred2;
} else {
IfTrue = Pred2;
IfFalse = Pred1;
}
return BI->getCondition();
}
void CPFlowFunction::visitBranchInst(BranchInst &BI) {
CPLatticePoint* result = new CPLatticePoint(*(info_in_casted.back()));
info_in_casted.pop_back();
BranchInst* current = &BI;
if (BI.isConditional()) {
Value* cond = BI.getCondition();
if (isa<ICmpInst>(cond)) {
std::pair<Use*, Use *> branches = helper::getOps(BI);
Use* true_branch = branches.first;
Use* false_branch = branches.second;
ICmpInst* cmp = dyn_cast<ICmpInst>(cond);
std::pair<Use*, Use *> operands = helper::getOps(*cmp);
Use* rhs = operands.second;
Use* lhs = operands.first;
ConstantInt* rhs_const = NULL;
ConstantInt* lhs_const = NULL;
// get the rhs/lhs as a constant int
if (isa<ConstantInt>(rhs)) {
rhs_const = dyn_cast<ConstantInt>(rhs);
} else if (result->representation.count(rhs->get()) > 0) {
rhs_const = result->representation[rhs->get()];
} else {
rhs_const = ConstantInt::get(context, llvm::APInt(32, 0, true));
}
if (isa<ConstantInt>(lhs)) {
lhs_const = dyn_cast<ConstantInt>(lhs->get());
} else if (result->representation.count(lhs->get()) > 0) {
lhs_const = result->representation[lhs->get()];
} else {
lhs_const = ConstantInt::get(context, llvm::APInt(32, 0, true));
}
// Create successors
CPLatticePoint* true_branchCLP = new CPLatticePoint(false, false, std::map<Value*,ConstantInt*>(result->representation));
CPLatticePoint* false_branchCLP = new CPLatticePoint(false, false, std::map<Value*,ConstantInt*>(result->representation));
// get the predicate
int predicate = 0;
predicate = cmp->isSigned() ? cmp->getSignedPredicate() : cmp->getUnsignedPredicate();
if (predicate == CmpInst::ICMP_EQ) {
if (isa<ConstantInt>(lhs)) {
true_branchCLP->representation[rhs->get()] = lhs_const;
} else if (isa<ConstantInt>(rhs)) {
true_branchCLP->representation[lhs->get()] = rhs_const;
}
out_map[true_branch->get()] = true_branchCLP;
out_map[false_branch->get()] = false_branchCLP;
} else if (predicate == CmpInst::ICMP_NE) {
if (isa<ConstantInt>(lhs)) {
false_branchCLP->representation[rhs->get()] = lhs_const;
} else if (isa<ConstantInt>(rhs)) {
false_branchCLP->representation[lhs->get()] = rhs_const;
}
out_map[true_branch->get()] = true_branchCLP;
out_map[false_branch->get()] = false_branchCLP;
} else {
for (std::map<Value *, LatticePoint *>::iterator it=out_map.begin(); it != out_map.end(); ++it){
Value* elm = it->first;
out_map[elm] = new CPLatticePoint(*result);
}
}
} else {
for (std::map<Value *, LatticePoint *>::iterator it=out_map.begin(); it != out_map.end(); ++it){
Value* elm = it->first;
out_map[elm] = new CPLatticePoint(*result);
}
}
} else {
for (std::map<Value *, LatticePoint *>::iterator it=out_map.begin(); it != out_map.end(); ++it){
Value* elm = it->first;
out_map[elm] = new CPLatticePoint(*result);
}
}
}
