/// If there's a single exit block, sink any loop-invariant values that
/// were defined in the preheader but not used inside the loop into the
/// exit block to reduce register pressure in the loop.
void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
BasicBlock *ExitBlock = L->getExitBlock();
if (!ExitBlock) return;
BasicBlock *Preheader = L->getLoopPreheader();
if (!Preheader) return;
Instruction *InsertPt = ExitBlock->getFirstNonPHI();
BasicBlock::iterator I = Preheader->getTerminator();
while (I != Preheader->begin()) {
--I;
// New instructions were inserted at the end of the preheader.
if (isa<PHINode>(I))
break;
// Don't move instructions which might have side effects, since the side
// effects need to complete before instructions inside the loop. Also
// don't move instructions which might read memory, since the loop may
// modify memory. Note that it's okay if the instruction might have
// undefined behavior: LoopSimplify guarantees that the preheader
// dominates the exit block.
if (I->mayHaveSideEffects() || I->mayReadFromMemory())
continue;
// Don't sink static AllocaInsts out of the entry block, which would
// turn them into dynamic allocas!
if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
if (AI->isStaticAlloca())
continue;
// Determine if there is a use in or before the loop (direct or
// otherwise).
bool UsedInLoop = false;
for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
UI != UE; ++UI) {
BasicBlock *UseBB = cast<Instruction>(UI)->getParent();
if (PHINode *P = dyn_cast<PHINode>(UI)) {
unsigned i =
PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
UseBB = P->getIncomingBlock(i);
}
if (UseBB == Preheader || L->contains(UseBB)) {
UsedInLoop = true;
break;
}
}
// If there is, the def must remain in the preheader.
if (UsedInLoop)
continue;
// Otherwise, sink it to the exit block.
Instruction *ToMove = I;
bool Done = false;
if (I != Preheader->begin())
--I;
else
Done = true;
ToMove->moveBefore(InsertPt);
if (Done)
break;
InsertPt = ToMove;
}
}
/// FindAllMemoryUses - Recursively walk all the uses of I until we find a
/// memory use. If we find an obviously non-foldable instruction, return true.
/// Add the ultimately found memory instructions to MemoryUses.
static bool FindAllMemoryUses(Instruction *I,
SmallVectorImpl<std::pair<Instruction*,unsigned> > &MemoryUses,
SmallPtrSet<Instruction*, 16> &ConsideredInsts,
const TargetLowering &TLI) {
// If we already considered this instruction, we're done.
if (!ConsideredInsts.insert(I))
return false;
// If this is an obviously unfoldable instruction, bail out.
if (!MightBeFoldableInst(I))
return true;
// Loop over all the uses, recursively processing them.
for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
UI != E; ++UI) {
User *U = *UI;
if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
MemoryUses.push_back(std::make_pair(LI, UI.getOperandNo()));
continue;
}
if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
unsigned opNo = UI.getOperandNo();
if (opNo == 0) return true; // Storing addr, not into addr.
MemoryUses.push_back(std::make_pair(SI, opNo));
continue;
}
if (CallInst *CI = dyn_cast<CallInst>(U)) {
InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledValue());
if (!IA) return true;
// If this is a memory operand, we're cool, otherwise bail out.
if (!IsOperandAMemoryOperand(CI, IA, I, TLI))
return true;
continue;
}
if (FindAllMemoryUses(cast<Instruction>(U), MemoryUses, ConsideredInsts,
TLI))
return true;
}
return false;
}
void ARM64PromoteConstant::
computeInsertionPoints(Constant *Val,
InsertionPointsPerFunc &InsPtsPerFunc) {
DEBUG(dbgs() << "** Compute insertion points **\n");
for (Value::use_iterator UseIt = Val->use_begin(), EndUseIt = Val->use_end();
UseIt != EndUseIt; ++UseIt) {
// If the user is not an Instruction, we cannot modify it
if (!isa<Instruction>(*UseIt))
continue;
// Filter out uses that should not be converted
if (!shouldConvertUse(Val, cast<Instruction>(*UseIt), UseIt.getOperandNo()))
continue;
DEBUG(dbgs() << "Considered use, opidx " << UseIt.getOperandNo() << ":\n");
DEBUG(UseIt->print(dbgs()));
DEBUG(dbgs() << '\n');
Instruction *InsertionPoint = findInsertionPoint(UseIt);
DEBUG(dbgs() << "Considered insertion point:\n");
DEBUG(InsertionPoint->print(dbgs()));
DEBUG(dbgs() << '\n');
// Check if the current insertion point is useless, i.e., it is dominated
// by another one.
InsertionPoints &InsertPts =
InsPtsPerFunc[InsertionPoint->getParent()->getParent()];
if (isDominated(InsertionPoint, UseIt, InsertPts))
continue;
// This insertion point is useful, check if we can merge some insertion
// point in a common dominator or if NewPt dominates an existing one.
if (tryAndMerge(InsertionPoint, UseIt, InsertPts))
continue;
DEBUG(dbgs() << "Keep considered insertion point\n");
// It is definitely useful by its own
InsertPts[InsertionPoint].push_back(UseIt);
}
}
/// AllUsesDominatedByBlock - Return true if all uses of the specified value
/// occur in blocks dominated by the specified block.
bool Sinking::AllUsesDominatedByBlock(Instruction *Inst,
BasicBlock *BB) const {
// Ignoring debug uses is necessary so debug info doesn't affect the code.
// This may leave a referencing dbg_value in the original block, before
// the definition of the vreg. Dwarf generator handles this although the
// user might not get the right info at runtime.
for (Value::use_iterator I = Inst->use_begin(),
E = Inst->use_end(); I != E; ++I) {
// Determine the block of the use.
Instruction *UseInst = cast<Instruction>(*I);
BasicBlock *UseBlock = UseInst->getParent();
if (PHINode *PN = dyn_cast<PHINode>(UseInst)) {
// PHI nodes use the operand in the predecessor block, not the block with
// the PHI.
unsigned Num = PHINode::getIncomingValueNumForOperand(I.getOperandNo());
UseBlock = PN->getIncomingBlock(Num);
}
// Check that it dominates.
if (!DT->dominates(BB, UseBlock))
return false;
}
return true;
}
/// SplitCriticalEdge - If this edge is a critical edge, insert a new node to
/// split the critical edge. This will update DominatorTree and
/// DominatorFrontier information if it is available, thus calling this pass
/// will not invalidate either of them. This returns the new block if the edge
/// was split, null otherwise.
///
/// If MergeIdenticalEdges is true (not the default), *all* edges from TI to the
/// specified successor will be merged into the same critical edge block.
/// This is most commonly interesting with switch instructions, which may
/// have many edges to any one destination. This ensures that all edges to that
/// dest go to one block instead of each going to a different block, but isn't
/// the standard definition of a "critical edge".
///
/// It is invalid to call this function on a critical edge that starts at an
/// IndirectBrInst. Splitting these edges will almost always create an invalid
/// program because the address of the new block won't be the one that is jumped
/// to.
///
BasicBlock *llvm::SplitCriticalEdge(TerminatorInst *TI, unsigned SuccNum,
Pass *P, bool MergeIdenticalEdges) {
if (!isCriticalEdge(TI, SuccNum, MergeIdenticalEdges)) return 0;
assert(!isa<IndirectBrInst>(TI) &&
"Cannot split critical edge from IndirectBrInst");
BasicBlock *TIBB = TI->getParent();
BasicBlock *DestBB = TI->getSuccessor(SuccNum);
// Create a new basic block, linking it into the CFG.
BasicBlock *NewBB = BasicBlock::Create(TI->getContext(),
TIBB->getName() + "." + DestBB->getName() + "_crit_edge");
// Create our unconditional branch.
BranchInst::Create(DestBB, NewBB);
// Branch to the new block, breaking the edge.
TI->setSuccessor(SuccNum, NewBB);
// Insert the block into the function... right after the block TI lives in.
Function &F = *TIBB->getParent();
Function::iterator FBBI = TIBB;
F.getBasicBlockList().insert(++FBBI, NewBB);
// If there are any PHI nodes in DestBB, we need to update them so that they
// merge incoming values from NewBB instead of from TIBB.
if (PHINode *APHI = dyn_cast<PHINode>(DestBB->begin())) {
// This conceptually does:
// foreach (PHINode *PN in DestBB)
// PN->setIncomingBlock(PN->getIncomingBlock(TIBB), NewBB);
// but is optimized for two cases.
if (APHI->getNumIncomingValues() <= 8) { // Small # preds case.
unsigned BBIdx = 0;
for (BasicBlock::iterator I = DestBB->begin(); isa<PHINode>(I); ++I) {
// We no longer enter through TIBB, now we come in through NewBB.
// Revector exactly one entry in the PHI node that used to come from
// TIBB to come from NewBB.
PHINode *PN = cast<PHINode>(I);
// Reuse the previous value of BBIdx if it lines up. In cases where we
// have multiple phi nodes with *lots* of predecessors, this is a speed
// win because we don't have to scan the PHI looking for TIBB. This
// happens because the BB list of PHI nodes are usually in the same
// order.
if (PN->getIncomingBlock(BBIdx) != TIBB)
BBIdx = PN->getBasicBlockIndex(TIBB);
PN->setIncomingBlock(BBIdx, NewBB);
}
} else {
// However, the foreach loop is slow for blocks with lots of predecessors
// because PHINode::getIncomingBlock is O(n) in # preds. Instead, walk
// the user list of TIBB to find the PHI nodes.
SmallPtrSet<PHINode*, 16> UpdatedPHIs;
for (Value::use_iterator UI = TIBB->use_begin(), E = TIBB->use_end();
UI != E; ) {
Value::use_iterator Use = UI++;
if (PHINode *PN = dyn_cast<PHINode>(Use)) {
// Remove one entry from each PHI.
if (PN->getParent() == DestBB && UpdatedPHIs.insert(PN))
PN->setOperand(Use.getOperandNo(), NewBB);
}
}
}
}
// If there are any other edges from TIBB to DestBB, update those to go
// through the split block, making those edges non-critical as well (and
// reducing the number of phi entries in the DestBB if relevant).
if (MergeIdenticalEdges) {
for (unsigned i = SuccNum+1, e = TI->getNumSuccessors(); i != e; ++i) {
if (TI->getSuccessor(i) != DestBB) continue;
// Remove an entry for TIBB from DestBB phi nodes.
DestBB->removePredecessor(TIBB);
// We found another edge to DestBB, go to NewBB instead.
TI->setSuccessor(i, NewBB);
}
}
// If we don't have a pass object, we can't update anything...
if (P == 0) return NewBB;
DominatorTree *DT = P->getAnalysisIfAvailable<DominatorTree>();
DominanceFrontier *DF = P->getAnalysisIfAvailable<DominanceFrontier>();
LoopInfo *LI = P->getAnalysisIfAvailable<LoopInfo>();
ProfileInfo *PI = P->getAnalysisIfAvailable<ProfileInfo>();
// If we have nothing to update, just return.
if (DT == 0 && DF == 0 && LI == 0 && PI == 0)
return NewBB;
// Now update analysis information. Since the only predecessor of NewBB is
// the TIBB, TIBB clearly dominates NewBB. TIBB usually doesn't dominate
// anything, as there are other successors of DestBB. However, if all other
// predecessors of DestBB are already dominated by DestBB (e.g. DestBB is a
// loop header) then NewBB dominates DestBB.
SmallVector<BasicBlock*, 8> OtherPreds;
// If there is a PHI in the block, loop over predecessors with it, which is
// faster than iterating pred_begin/end.
if (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) {
for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
if (PN->getIncomingBlock(i) != NewBB)
OtherPreds.push_back(PN->getIncomingBlock(i));
} else {
for (pred_iterator I = pred_begin(DestBB), E = pred_end(DestBB);
I != E; ++I)
if (*I != NewBB)
OtherPreds.push_back(*I);
}
bool NewBBDominatesDestBB = true;
// Should we update DominatorTree information?
if (DT) {
DomTreeNode *TINode = DT->getNode(TIBB);
// The new block is not the immediate dominator for any other nodes, but
// TINode is the immediate dominator for the new node.
//
if (TINode) { // Don't break unreachable code!
DomTreeNode *NewBBNode = DT->addNewBlock(NewBB, TIBB);
DomTreeNode *DestBBNode = 0;
// If NewBBDominatesDestBB hasn't been computed yet, do so with DT.
if (!OtherPreds.empty()) {
DestBBNode = DT->getNode(DestBB);
while (!OtherPreds.empty() && NewBBDominatesDestBB) {
if (DomTreeNode *OPNode = DT->getNode(OtherPreds.back()))
NewBBDominatesDestBB = DT->dominates(DestBBNode, OPNode);
OtherPreds.pop_back();
}
OtherPreds.clear();
}
// If NewBBDominatesDestBB, then NewBB dominates DestBB, otherwise it
// doesn't dominate anything.
if (NewBBDominatesDestBB) {
if (!DestBBNode) DestBBNode = DT->getNode(DestBB);
DT->changeImmediateDominator(DestBBNode, NewBBNode);
}
}
}
// Should we update DominanceFrontier information?
if (DF) {
// If NewBBDominatesDestBB hasn't been computed yet, do so with DF.
if (!OtherPreds.empty()) {
// FIXME: IMPLEMENT THIS!
llvm_unreachable("Requiring domfrontiers but not idom/domtree/domset."
" not implemented yet!");
}
// Since the new block is dominated by its only predecessor TIBB,
// it cannot be in any block's dominance frontier. If NewBB dominates
// DestBB, its dominance frontier is the same as DestBB's, otherwise it is
// just {DestBB}.
DominanceFrontier::DomSetType NewDFSet;
if (NewBBDominatesDestBB) {
DominanceFrontier::iterator I = DF->find(DestBB);
if (I != DF->end()) {
DF->addBasicBlock(NewBB, I->second);
if (I->second.count(DestBB)) {
// However NewBB's frontier does not include DestBB.
DominanceFrontier::iterator NF = DF->find(NewBB);
DF->removeFromFrontier(NF, DestBB);
}
}
else
DF->addBasicBlock(NewBB, DominanceFrontier::DomSetType());
} else {
DominanceFrontier::DomSetType NewDFSet;
NewDFSet.insert(DestBB);
DF->addBasicBlock(NewBB, NewDFSet);
}
}
// Update LoopInfo if it is around.
if (LI) {
if (Loop *TIL = LI->getLoopFor(TIBB)) {
// If one or the other blocks were not in a loop, the new block is not
// either, and thus LI doesn't need to be updated.
if (Loop *DestLoop = LI->getLoopFor(DestBB)) {
if (TIL == DestLoop) {
// Both in the same loop, the NewBB joins loop.
DestLoop->addBasicBlockToLoop(NewBB, LI->getBase());
} else if (TIL->contains(DestLoop)) {
// Edge from an outer loop to an inner loop. Add to the outer loop.
TIL->addBasicBlockToLoop(NewBB, LI->getBase());
} else if (DestLoop->contains(TIL)) {
// Edge from an inner loop to an outer loop. Add to the outer loop.
DestLoop->addBasicBlockToLoop(NewBB, LI->getBase());
} else {
// Edge from two loops with no containment relation. Because these
// are natural loops, we know that the destination block must be the
// header of its loop (adding a branch into a loop elsewhere would
// create an irreducible loop).
assert(DestLoop->getHeader() == DestBB &&
"Should not create irreducible loops!");
if (Loop *P = DestLoop->getParentLoop())
P->addBasicBlockToLoop(NewBB, LI->getBase());
}
}
// If TIBB is in a loop and DestBB is outside of that loop, split the
// other exit blocks of the loop that also have predecessors outside
// the loop, to maintain a LoopSimplify guarantee.
if (!TIL->contains(DestBB) &&
P->mustPreserveAnalysisID(LoopSimplifyID)) {
assert(!TIL->contains(NewBB) &&
"Split point for loop exit is contained in loop!");
// Update LCSSA form in the newly created exit block.
if (P->mustPreserveAnalysisID(LCSSAID)) {
SmallVector<BasicBlock *, 1> OrigPred;
OrigPred.push_back(TIBB);
CreatePHIsForSplitLoopExit(OrigPred, NewBB, DestBB);
}
// For each unique exit block...
SmallVector<BasicBlock *, 4> ExitBlocks;
TIL->getExitBlocks(ExitBlocks);
for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
// Collect all the preds that are inside the loop, and note
// whether there are any preds outside the loop.
SmallVector<BasicBlock *, 4> Preds;
bool HasPredOutsideOfLoop = false;
BasicBlock *Exit = ExitBlocks[i];
for (pred_iterator I = pred_begin(Exit), E = pred_end(Exit);
I != E; ++I)
if (TIL->contains(*I))
Preds.push_back(*I);
else
HasPredOutsideOfLoop = true;
// If there are any preds not in the loop, we'll need to split
// the edges. The Preds.empty() check is needed because a block
// may appear multiple times in the list. We can't use
// getUniqueExitBlocks above because that depends on LoopSimplify
// form, which we're in the process of restoring!
if (!Preds.empty() && HasPredOutsideOfLoop) {
BasicBlock *NewExitBB =
SplitBlockPredecessors(Exit, Preds.data(), Preds.size(),
"split", P);
if (P->mustPreserveAnalysisID(LCSSAID))
CreatePHIsForSplitLoopExit(Preds, NewExitBB, Exit);
}
}
}
// LCSSA form was updated above for the case where LoopSimplify is
// available, which means that all predecessors of loop exit blocks
// are within the loop. Without LoopSimplify form, it would be
// necessary to insert a new phi.
assert((!P->mustPreserveAnalysisID(LCSSAID) ||
P->mustPreserveAnalysisID(LoopSimplifyID)) &&
"SplitCriticalEdge doesn't know how to update LCCSA form "
"without LoopSimplify!");
}
}
// Update ProfileInfo if it is around.
if (PI)
PI->splitEdge(TIBB, DestBB, NewBB, MergeIdenticalEdges);
return NewBB;
}
/// SurveyUse - This looks at a single use of an argument or return value
/// and determines if it should be alive or not. Adds this use to MaybeLiveUses
/// if it causes the used value to become MaybeAlive.
///
/// RetValNum is the return value number to use when this use is used in a
/// return instruction. This is used in the recursion, you should always leave
/// it at 0.
DAE::Liveness DAE::SurveyUse(Value::use_iterator U, UseVector &MaybeLiveUses,
unsigned RetValNum) {
Value *V = *U;
if (ReturnInst *RI = dyn_cast<ReturnInst>(V)) {
// The value is returned from a function. It's only live when the
// function's return value is live. We use RetValNum here, for the case
// that U is really a use of an insertvalue instruction that uses the
// orginal Use.
RetOrArg Use = CreateRet(RI->getParent()->getParent(), RetValNum);
// We might be live, depending on the liveness of Use.
return MarkIfNotLive(Use, MaybeLiveUses);
}
if (InsertValueInst *IV = dyn_cast<InsertValueInst>(V)) {
if (U.getOperandNo() != InsertValueInst::getAggregateOperandIndex()
&& IV->hasIndices())
// The use we are examining is inserted into an aggregate. Our liveness
// depends on all uses of that aggregate, but if it is used as a return
// value, only index at which we were inserted counts.
RetValNum = *IV->idx_begin();
// Note that if we are used as the aggregate operand to the insertvalue,
// we don't change RetValNum, but do survey all our uses.
Liveness Result = MaybeLive;
for (Value::use_iterator I = IV->use_begin(),
E = V->use_end(); I != E; ++I) {
Result = SurveyUse(I, MaybeLiveUses, RetValNum);
if (Result == Live)
break;
}
return Result;
}
CallSite CS = CallSite::get(V);
if (CS.getInstruction()) {
Function *F = CS.getCalledFunction();
if (F) {
// Used in a direct call.
// Find the argument number. We know for sure that this use is an
// argument, since if it was the function argument this would be an
// indirect call and the we know can't be looking at a value of the
// label type (for the invoke instruction).
unsigned ArgNo = CS.getArgumentNo(U.getOperandNo());
if (ArgNo >= F->getFunctionType()->getNumParams())
// The value is passed in through a vararg! Must be live.
return Live;
assert(CS.getArgument(ArgNo)
== CS.getInstruction()->getOperand(U.getOperandNo())
&& "Argument is not where we expected it");
// Value passed to a normal call. It's only live when the corresponding
// argument to the called function turns out live.
RetOrArg Use = CreateArg(F, ArgNo);
return MarkIfNotLive(Use, MaybeLiveUses);
}
}
// Used in any other way? Value must be live.
return Live;
}
/// isOnlyCopiedFromConstantGlobal - Recursively walk the uses of a (derived)
/// pointer to an alloca. Ignore any reads of the pointer, return false if we
/// see any stores or other unknown uses. If we see pointer arithmetic, keep
/// track of whether it moves the pointer (with IsOffset) but otherwise traverse
/// the uses. If we see a memcpy/memmove that targets an unoffseted pointer to
/// the alloca, and if the source pointer is a pointer to a constant global, we
/// can optimize this.
static bool
isOnlyCopiedFromConstantGlobal(Value *V, MemTransferInst *&TheCopy,
SmallVectorImpl<Instruction *> &ToDelete,
bool IsOffset = false) {
// We track lifetime intrinsics as we encounter them. If we decide to go
// ahead and replace the value with the global, this lets the caller quickly
// eliminate the markers.
for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI!=E; ++UI) {
User *U = cast<Instruction>(*UI);
if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
// Ignore non-volatile loads, they are always ok.
if (!LI->isSimple()) return false;
continue;
}
if (BitCastInst *BCI = dyn_cast<BitCastInst>(U)) {
// If uses of the bitcast are ok, we are ok.
if (!isOnlyCopiedFromConstantGlobal(BCI, TheCopy, ToDelete, IsOffset))
return false;
continue;
}
if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
// If the GEP has all zero indices, it doesn't offset the pointer. If it
// doesn't, it does.
if (!isOnlyCopiedFromConstantGlobal(GEP, TheCopy, ToDelete,
IsOffset || !GEP->hasAllZeroIndices()))
return false;
continue;
}
if (CallSite CS = U) {
// If this is the function being called then we treat it like a load and
// ignore it.
if (CS.isCallee(UI))
continue;
// If this is a readonly/readnone call site, then we know it is just a
// load (but one that potentially returns the value itself), so we can
// ignore it if we know that the value isn't captured.
unsigned ArgNo = CS.getArgumentNo(UI);
if (CS.onlyReadsMemory() &&
(CS.getInstruction()->use_empty() || CS.doesNotCapture(ArgNo)))
continue;
// If this is being passed as a byval argument, the caller is making a
// copy, so it is only a read of the alloca.
if (CS.isByValArgument(ArgNo))
continue;
}
// Lifetime intrinsics can be handled by the caller.
if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(U)) {
if (II->getIntrinsicID() == Intrinsic::lifetime_start ||
II->getIntrinsicID() == Intrinsic::lifetime_end) {
assert(II->use_empty() && "Lifetime markers have no result to use!");
ToDelete.push_back(II);
continue;
}
}
// If this is isn't our memcpy/memmove, reject it as something we can't
// handle.
MemTransferInst *MI = dyn_cast<MemTransferInst>(U);
if (MI == 0)
return false;
// If the transfer is using the alloca as a source of the transfer, then
// ignore it since it is a load (unless the transfer is volatile).
if (UI.getOperandNo() == 1) {
if (MI->isVolatile()) return false;
continue;
}
// If we already have seen a copy, reject the second one.
if (TheCopy) return false;
// If the pointer has been offset from the start of the alloca, we can't
// safely handle this.
if (IsOffset) return false;
// If the memintrinsic isn't using the alloca as the dest, reject it.
if (UI.getOperandNo() != 0) return false;
// If the source of the memcpy/move is not a constant global, reject it.
if (!pointsToConstantGlobal(MI->getSource()))
return false;
// Otherwise, the transform is safe. Remember the copy instruction.
TheCopy = MI;
}
return true;
}