#include "llvm/Analysis/ScalarEvolutionExpander.h"
#include "llvm/Analysis/LoopInfo.h"
+#include "llvm/IntrinsicInst.h"
#include "llvm/LLVMContext.h"
#include "llvm/Target/TargetData.h"
#include "llvm/ADT/STLExtras.h"
using namespace llvm;
+/// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP,
+/// reusing an existing cast if a suitable one exists, moving an existing
+/// cast if a suitable one exists but isn't in the right place, or
+/// creating a new one.
+Value *SCEVExpander::ReuseOrCreateCast(Value *V, const Type *Ty,
+ Instruction::CastOps Op,
+ BasicBlock::iterator IP) {
+ // Check to see if there is already a cast!
+ for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
+ UI != E; ++UI) {
+ User *U = *UI;
+ if (U->getType() == Ty)
+ if (CastInst *CI = dyn_cast<CastInst>(U))
+ if (CI->getOpcode() == Op) {
+ // If the cast isn't where we want it, fix it.
+ if (BasicBlock::iterator(CI) != IP) {
+ // Create a new cast, and leave the old cast in place in case
+ // it is being used as an insert point. Clear its operand
+ // so that it doesn't hold anything live.
+ Instruction *NewCI = CastInst::Create(Op, V, Ty, "", IP);
+ NewCI->takeName(CI);
+ CI->replaceAllUsesWith(NewCI);
+ CI->setOperand(0, UndefValue::get(V->getType()));
+ rememberInstruction(NewCI);
+ return NewCI;
+ }
+ rememberInstruction(CI);
+ return CI;
+ }
+ }
+
+ // Create a new cast.
+ Instruction *I = CastInst::Create(Op, V, Ty, V->getName(), IP);
+ rememberInstruction(I);
+ return I;
+}
+
/// InsertNoopCastOfTo - Insert a cast of V to the specified type,
/// which must be possible with a noop cast, doing what we can to share
/// the casts.
return CE->getOperand(0);
}
+ // Fold a cast of a constant.
if (Constant *C = dyn_cast<Constant>(V))
return ConstantExpr::getCast(Op, C, Ty);
+ // Cast the argument at the beginning of the entry block, after
+ // any bitcasts of other arguments.
if (Argument *A = dyn_cast<Argument>(V)) {
- // Check to see if there is already a cast!
- for (Value::use_iterator UI = A->use_begin(), E = A->use_end();
- UI != E; ++UI)
- if ((*UI)->getType() == Ty)
- if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI)))
- if (CI->getOpcode() == Op) {
- // If the cast isn't the first instruction of the function, move it.
- if (BasicBlock::iterator(CI) !=
- A->getParent()->getEntryBlock().begin()) {
- // Recreate the cast at the beginning of the entry block.
- // The old cast is left in place in case it is being used
- // as an insert point.
- Instruction *NewCI =
- CastInst::Create(Op, V, Ty, "",
- A->getParent()->getEntryBlock().begin());
- NewCI->takeName(CI);
- CI->replaceAllUsesWith(NewCI);
- return NewCI;
- }
- return CI;
- }
-
- Instruction *I = CastInst::Create(Op, V, Ty, V->getName(),
- A->getParent()->getEntryBlock().begin());
- rememberInstruction(I);
- return I;
+ BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin();
+ while ((isa<BitCastInst>(IP) &&
+ isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) &&
+ cast<BitCastInst>(IP)->getOperand(0) != A) ||
+ isa<DbgInfoIntrinsic>(IP))
+ ++IP;
+ return ReuseOrCreateCast(A, Ty, Op, IP);
}
+ // Cast the instruction immediately after the instruction.
Instruction *I = cast<Instruction>(V);
-
- // Check to see if there is already a cast. If there is, use it.
- for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
- UI != E; ++UI) {
- if ((*UI)->getType() == Ty)
- if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI)))
- if (CI->getOpcode() == Op) {
- BasicBlock::iterator It = I; ++It;
- if (isa<InvokeInst>(I))
- It = cast<InvokeInst>(I)->getNormalDest()->begin();
- while (isa<PHINode>(It)) ++It;
- if (It != BasicBlock::iterator(CI)) {
- // Recreate the cast after the user.
- // The old cast is left in place in case it is being used
- // as an insert point.
- Instruction *NewCI = CastInst::Create(Op, V, Ty, "", It);
- NewCI->takeName(CI);
- CI->replaceAllUsesWith(NewCI);
- rememberInstruction(NewCI);
- return NewCI;
- }
- rememberInstruction(CI);
- return CI;
- }
- }
BasicBlock::iterator IP = I; ++IP;
if (InvokeInst *II = dyn_cast<InvokeInst>(I))
IP = II->getNormalDest()->begin();
- while (isa<PHINode>(IP)) ++IP;
- Instruction *CI = CastInst::Create(Op, V, Ty, V->getName(), IP);
- rememberInstruction(CI);
- return CI;
+ while (isa<PHINode>(IP) || isa<DbgInfoIntrinsic>(IP)) ++IP;
+ return ReuseOrCreateCast(I, Ty, Op, IP);
}
/// InsertBinop - Insert the specified binary operator, doing a small amount
if (IP != BlockBegin) {
--IP;
for (; ScanLimit; --IP, --ScanLimit) {
+ // Don't count dbg.value against the ScanLimit, to avoid perturbing the
+ // generated code.
+ if (isa<DbgInfoIntrinsic>(IP))
+ ScanLimit++;
if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
IP->getOperand(1) == RHS)
return IP;
}
}
+ // Save the original insertion point so we can restore it when we're done.
+ BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
+ BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
+
+ // Move the insertion point out of as many loops as we can.
+ while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
+ if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break;
+ BasicBlock *Preheader = L->getLoopPreheader();
+ if (!Preheader) break;
+
+ // Ok, move up a level.
+ Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
+ }
+
// If we haven't found this binop, insert it.
Value *BO = Builder.CreateBinOp(Opcode, LHS, RHS, "tmp");
rememberInstruction(BO);
+
+ // Restore the original insert point.
+ if (SaveInsertBB)
+ restoreInsertPoint(SaveInsertBB, SaveInsertPt);
+
return BO;
}
/// FactorOutConstant - Test if S is divisible by Factor, using signed
/// division. If so, update S with Factor divided out and return true.
-/// S need not be evenly divisble if a reasonable remainder can be
+/// S need not be evenly divisible if a reasonable remainder can be
/// computed.
/// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
/// unnecessary; in its place, just signed-divide Ops[i] by the scale and
// x/x == 1.
if (S == Factor) {
- S = SE.getIntegerSCEV(1, S->getType());
+ S = SE.getConstant(S->getType(), 1);
return true;
}
const SCEVConstant *FC = cast<SCEVConstant>(Factor);
if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
- const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
- SmallVector<const SCEV *, 4> NewMulOps(MOperands.begin(),
- MOperands.end());
+ SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
NewMulOps[0] =
SE.getConstant(C->getValue()->getValue().sdiv(
FC->getValue()->getValue()));
// Mul's operands. If so, we can just remove it.
for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
const SCEV *SOp = M->getOperand(i);
- const SCEV *Remainder = SE.getIntegerSCEV(0, SOp->getType());
+ const SCEV *Remainder = SE.getConstant(SOp->getType(), 0);
if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) &&
Remainder->isZero()) {
- const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
- SmallVector<const SCEV *, 4> NewMulOps(MOperands.begin(),
- MOperands.end());
+ SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end());
NewMulOps[i] = SOp;
S = SE.getMulExpr(NewMulOps);
return true;
// In an AddRec, check if both start and step are divisible.
if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
const SCEV *Step = A->getStepRecurrence(SE);
- const SCEV *StepRem = SE.getIntegerSCEV(0, Step->getType());
+ const SCEV *StepRem = SE.getConstant(Step->getType(), 0);
if (!FactorOutConstant(Step, StepRem, Factor, SE, TD))
return false;
if (!StepRem->isZero())
SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
// Let ScalarEvolution sort and simplify the non-addrecs list.
const SCEV *Sum = NoAddRecs.empty() ?
- SE.getIntegerSCEV(0, Ty) :
+ SE.getConstant(Ty, 0) :
SE.getAddExpr(NoAddRecs);
// If it returned an add, use the operands. Otherwise it simplified
// the sum into a single value, so just use that.
+ Ops.clear();
if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
- Ops = Add->getOperands();
- else {
- Ops.clear();
- if (!Sum->isZero())
- Ops.push_back(Sum);
- }
+ Ops.append(Add->op_begin(), Add->op_end());
+ else if (!Sum->isZero())
+ Ops.push_back(Sum);
// Then append the addrecs.
- Ops.insert(Ops.end(), AddRecs.begin(), AddRecs.end());
+ Ops.append(AddRecs.begin(), AddRecs.end());
}
/// SplitAddRecs - Flatten a list of add operands, moving addrec start values
while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
const SCEV *Start = A->getStart();
if (Start->isZero()) break;
- const SCEV *Zero = SE.getIntegerSCEV(0, Ty);
+ const SCEV *Zero = SE.getConstant(Ty, 0);
AddRecs.push_back(SE.getAddRecExpr(Zero,
A->getStepRecurrence(SE),
A->getLoop()));
if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
Ops[i] = Zero;
- Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
+ Ops.append(Add->op_begin(), Add->op_end());
e += Add->getNumOperands();
} else {
Ops[i] = Start;
}
if (!AddRecs.empty()) {
// Add the addrecs onto the end of the list.
- Ops.insert(Ops.end(), AddRecs.begin(), AddRecs.end());
+ Ops.append(AddRecs.begin(), AddRecs.end());
// Resort the operand list, moving any constants to the front.
SimplifyAddOperands(Ops, Ty, SE);
}
// the indices index into the element or field type selected by the
// preceding index.
for (;;) {
- const SCEV *ElSize = SE.getAllocSizeExpr(ElTy);
// If the scale size is not 0, attempt to factor out a scale for
// array indexing.
SmallVector<const SCEV *, 8> ScaledOps;
- if (ElTy->isSized() && !ElSize->isZero()) {
- SmallVector<const SCEV *, 8> NewOps;
- for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
- const SCEV *Op = Ops[i];
- const SCEV *Remainder = SE.getIntegerSCEV(0, Ty);
- if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) {
- // Op now has ElSize factored out.
- ScaledOps.push_back(Op);
- if (!Remainder->isZero())
- NewOps.push_back(Remainder);
- AnyNonZeroIndices = true;
- } else {
- // The operand was not divisible, so add it to the list of operands
- // we'll scan next iteration.
- NewOps.push_back(Ops[i]);
+ if (ElTy->isSized()) {
+ const SCEV *ElSize = SE.getSizeOfExpr(ElTy);
+ if (!ElSize->isZero()) {
+ SmallVector<const SCEV *, 8> NewOps;
+ for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
+ const SCEV *Op = Ops[i];
+ const SCEV *Remainder = SE.getConstant(Ty, 0);
+ if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) {
+ // Op now has ElSize factored out.
+ ScaledOps.push_back(Op);
+ if (!Remainder->isZero())
+ NewOps.push_back(Remainder);
+ AnyNonZeroIndices = true;
+ } else {
+ // The operand was not divisible, so add it to the list of operands
+ // we'll scan next iteration.
+ NewOps.push_back(Ops[i]);
+ }
+ }
+ // If we made any changes, update Ops.
+ if (!ScaledOps.empty()) {
+ Ops = NewOps;
+ SimplifyAddOperands(Ops, Ty, SE);
}
- }
- // If we made any changes, update Ops.
- if (!ScaledOps.empty()) {
- Ops = NewOps;
- SimplifyAddOperands(Ops, Ty, SE);
}
}
}
}
} else {
- // Without TargetData, just check for a SCEVFieldOffsetExpr of the
+ // Without TargetData, just check for an offsetof expression of the
// appropriate struct type.
for (unsigned i = 0, e = Ops.size(); i != e; ++i)
- if (const SCEVFieldOffsetExpr *FO =
- dyn_cast<SCEVFieldOffsetExpr>(Ops[i]))
- if (FO->getStructType() == STy) {
- unsigned FieldNo = FO->getFieldNo();
- GepIndices.push_back(
- ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
- FieldNo));
- ElTy = STy->getTypeAtIndex(FieldNo);
+ if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) {
+ const Type *CTy;
+ Constant *FieldNo;
+ if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) {
+ GepIndices.push_back(FieldNo);
+ ElTy =
+ STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue());
Ops[i] = SE.getConstant(Ty, 0);
AnyNonZeroIndices = true;
FoundFieldNo = true;
break;
}
+ }
}
// If no struct field offsets were found, tentatively assume that
// field zero was selected (since the zero offset would obviously
break;
}
- // If none of the operands were convertable to proper GEP indices, cast
+ // If none of the operands were convertible to proper GEP indices, cast
// the base to i8* and do an ugly getelementptr with that. It's still
// better than ptrtoint+arithmetic+inttoptr at least.
if (!AnyNonZeroIndices) {
if (IP != BlockBegin) {
--IP;
for (; ScanLimit; --IP, --ScanLimit) {
+ // Don't count dbg.value against the ScanLimit, to avoid perturbing the
+ // generated code.
+ if (isa<DbgInfoIntrinsic>(IP))
+ ScanLimit++;
if (IP->getOpcode() == Instruction::GetElementPtr &&
IP->getOperand(0) == V && IP->getOperand(1) == Idx)
return IP;
}
}
+ // Save the original insertion point so we can restore it when we're done.
+ BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
+ BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
+
+ // Move the insertion point out of as many loops as we can.
+ while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
+ if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break;
+ BasicBlock *Preheader = L->getLoopPreheader();
+ if (!Preheader) break;
+
+ // Ok, move up a level.
+ Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
+ }
+
// Emit a GEP.
Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
rememberInstruction(GEP);
+
+ // Restore the original insert point.
+ if (SaveInsertBB)
+ restoreInsertPoint(SaveInsertBB, SaveInsertPt);
+
return GEP;
}
+ // Save the original insertion point so we can restore it when we're done.
+ BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
+ BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
+
+ // Move the insertion point out of as many loops as we can.
+ while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) {
+ if (!L->isLoopInvariant(V)) break;
+
+ bool AnyIndexNotLoopInvariant = false;
+ for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(),
+ E = GepIndices.end(); I != E; ++I)
+ if (!L->isLoopInvariant(*I)) {
+ AnyIndexNotLoopInvariant = true;
+ break;
+ }
+ if (AnyIndexNotLoopInvariant)
+ break;
+
+ BasicBlock *Preheader = L->getLoopPreheader();
+ if (!Preheader) break;
+
+ // Ok, move up a level.
+ Builder.SetInsertPoint(Preheader, Preheader->getTerminator());
+ }
+
// Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
// because ScalarEvolution may have changed the address arithmetic to
// compute a value which is beyond the end of the allocated object.
"scevgep");
Ops.push_back(SE.getUnknown(GEP));
rememberInstruction(GEP);
+
+ // Restore the original insert point.
+ if (SaveInsertBB)
+ restoreInsertPoint(SaveInsertBB, SaveInsertPt);
+
return expand(SE.getAddExpr(Ops));
}
return SC->getValue()->getValue().isNegative();
}
-Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
- int NumOperands = S->getNumOperands();
- const Type *Ty = SE.getEffectiveSCEVType(S->getType());
+/// PickMostRelevantLoop - Given two loops pick the one that's most relevant for
+/// SCEV expansion. If they are nested, this is the most nested. If they are
+/// neighboring, pick the later.
+static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B,
+ DominatorTree &DT) {
+ if (!A) return B;
+ if (!B) return A;
+ if (A->contains(B)) return B;
+ if (B->contains(A)) return A;
+ if (DT.dominates(A->getHeader(), B->getHeader())) return B;
+ if (DT.dominates(B->getHeader(), A->getHeader())) return A;
+ return A; // Arbitrarily break the tie.
+}
- // Find the index of an operand to start with. Choose the operand with
- // pointer type, if there is one, or the last operand otherwise.
- int PIdx = 0;
- for (; PIdx != NumOperands - 1; ++PIdx)
- if (isa<PointerType>(S->getOperand(PIdx)->getType())) break;
-
- // Expand code for the operand that we chose.
- Value *V = expand(S->getOperand(PIdx));
-
- // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
- // comments on expandAddToGEP for details.
- if (const PointerType *PTy = dyn_cast<PointerType>(V->getType())) {
- // Take the operand at PIdx out of the list.
- const SmallVectorImpl<const SCEV *> &Ops = S->getOperands();
- SmallVector<const SCEV *, 8> NewOps;
- NewOps.insert(NewOps.end(), Ops.begin(), Ops.begin() + PIdx);
- NewOps.insert(NewOps.end(), Ops.begin() + PIdx + 1, Ops.end());
- // Make a GEP.
- return expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, V);
+/// GetRelevantLoop - Get the most relevant loop associated with the given
+/// expression, according to PickMostRelevantLoop.
+static const Loop *GetRelevantLoop(const SCEV *S, LoopInfo &LI,
+ DominatorTree &DT) {
+ if (isa<SCEVConstant>(S))
+ return 0;
+ if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) {
+ if (const Instruction *I = dyn_cast<Instruction>(U->getValue()))
+ return LI.getLoopFor(I->getParent());
+ return 0;
}
+ if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) {
+ const Loop *L = 0;
+ if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
+ L = AR->getLoop();
+ for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end();
+ I != E; ++I)
+ L = PickMostRelevantLoop(L, GetRelevantLoop(*I, LI, DT), DT);
+ return L;
+ }
+ if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
+ return GetRelevantLoop(C->getOperand(), LI, DT);
+ if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S))
+ return PickMostRelevantLoop(GetRelevantLoop(D->getLHS(), LI, DT),
+ GetRelevantLoop(D->getRHS(), LI, DT),
+ DT);
+ llvm_unreachable("Unexpected SCEV type!");
+}
+
+namespace {
+
+/// LoopCompare - Compare loops by PickMostRelevantLoop.
+class LoopCompare {
+ DominatorTree &DT;
+public:
+ explicit LoopCompare(DominatorTree &dt) : DT(dt) {}
+
+ bool operator()(std::pair<const Loop *, const SCEV *> LHS,
+ std::pair<const Loop *, const SCEV *> RHS) const {
+ // Keep pointer operands sorted at the end.
+ if (LHS.second->getType()->isPointerTy() !=
+ RHS.second->getType()->isPointerTy())
+ return LHS.second->getType()->isPointerTy();
+
+ // Compare loops with PickMostRelevantLoop.
+ if (LHS.first != RHS.first)
+ return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first;
+
+ // If one operand is a non-constant negative and the other is not,
+ // put the non-constant negative on the right so that a sub can
+ // be used instead of a negate and add.
+ if (isNonConstantNegative(LHS.second)) {
+ if (!isNonConstantNegative(RHS.second))
+ return false;
+ } else if (isNonConstantNegative(RHS.second))
+ return true;
- // Otherwise, we'll expand the rest of the SCEVAddExpr as plain integer
- // arithmetic.
- V = InsertNoopCastOfTo(V, Ty);
+ // Otherwise they are equivalent according to this comparison.
+ return false;
+ }
+};
+
+}
- // Emit a bunch of add instructions
- for (int i = NumOperands-1; i >= 0; --i) {
- if (i == PIdx) continue;
- const SCEV *Op = S->getOperand(i);
- if (isNonConstantNegative(Op)) {
+Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
+ const Type *Ty = SE.getEffectiveSCEVType(S->getType());
+
+ // Collect all the add operands in a loop, along with their associated loops.
+ // Iterate in reverse so that constants are emitted last, all else equal, and
+ // so that pointer operands are inserted first, which the code below relies on
+ // to form more involved GEPs.
+ SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
+ for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()),
+ E(S->op_begin()); I != E; ++I)
+ OpsAndLoops.push_back(std::make_pair(GetRelevantLoop(*I, *SE.LI, *SE.DT),
+ *I));
+
+ // Sort by loop. Use a stable sort so that constants follow non-constants and
+ // pointer operands precede non-pointer operands.
+ std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
+
+ // Emit instructions to add all the operands. Hoist as much as possible
+ // out of loops, and form meaningful getelementptrs where possible.
+ Value *Sum = 0;
+ for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
+ I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
+ const Loop *CurLoop = I->first;
+ const SCEV *Op = I->second;
+ if (!Sum) {
+ // This is the first operand. Just expand it.
+ Sum = expand(Op);
+ ++I;
+ } else if (const PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) {
+ // The running sum expression is a pointer. Try to form a getelementptr
+ // at this level with that as the base.
+ SmallVector<const SCEV *, 4> NewOps;
+ for (; I != E && I->first == CurLoop; ++I) {
+ // If the operand is SCEVUnknown and not instructions, peek through
+ // it, to enable more of it to be folded into the GEP.
+ const SCEV *X = I->second;
+ if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X))
+ if (!isa<Instruction>(U->getValue()))
+ X = SE.getSCEV(U->getValue());
+ NewOps.push_back(X);
+ }
+ Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum);
+ } else if (const PointerType *PTy = dyn_cast<PointerType>(Op->getType())) {
+ // The running sum is an integer, and there's a pointer at this level.
+ // Try to form a getelementptr. If the running sum is instructions,
+ // use a SCEVUnknown to avoid re-analyzing them.
+ SmallVector<const SCEV *, 4> NewOps;
+ NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) :
+ SE.getSCEV(Sum));
+ for (++I; I != E && I->first == CurLoop; ++I)
+ NewOps.push_back(I->second);
+ Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op));
+ } else if (isNonConstantNegative(Op)) {
+ // Instead of doing a negate and add, just do a subtract.
Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty);
- V = InsertBinop(Instruction::Sub, V, W);
+ Sum = InsertNoopCastOfTo(Sum, Ty);
+ Sum = InsertBinop(Instruction::Sub, Sum, W);
+ ++I;
} else {
+ // A simple add.
Value *W = expandCodeFor(Op, Ty);
- V = InsertBinop(Instruction::Add, V, W);
+ Sum = InsertNoopCastOfTo(Sum, Ty);
+ // Canonicalize a constant to the RHS.
+ if (isa<Constant>(Sum)) std::swap(Sum, W);
+ Sum = InsertBinop(Instruction::Add, Sum, W);
+ ++I;
}
}
- return V;
+
+ return Sum;
}
Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
const Type *Ty = SE.getEffectiveSCEVType(S->getType());
- int FirstOp = 0; // Set if we should emit a subtract.
- if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getOperand(0)))
- if (SC->getValue()->isAllOnesValue())
- FirstOp = 1;
-
- int i = S->getNumOperands()-2;
- Value *V = expandCodeFor(S->getOperand(i+1), Ty);
-
- // Emit a bunch of multiply instructions
- for (; i >= FirstOp; --i) {
- Value *W = expandCodeFor(S->getOperand(i), Ty);
- V = InsertBinop(Instruction::Mul, V, W);
+
+ // Collect all the mul operands in a loop, along with their associated loops.
+ // Iterate in reverse so that constants are emitted last, all else equal.
+ SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops;
+ for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()),
+ E(S->op_begin()); I != E; ++I)
+ OpsAndLoops.push_back(std::make_pair(GetRelevantLoop(*I, *SE.LI, *SE.DT),
+ *I));
+
+ // Sort by loop. Use a stable sort so that constants follow non-constants.
+ std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT));
+
+ // Emit instructions to mul all the operands. Hoist as much as possible
+ // out of loops.
+ Value *Prod = 0;
+ for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator
+ I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) {
+ const SCEV *Op = I->second;
+ if (!Prod) {
+ // This is the first operand. Just expand it.
+ Prod = expand(Op);
+ ++I;
+ } else if (Op->isAllOnesValue()) {
+ // Instead of doing a multiply by negative one, just do a negate.
+ Prod = InsertNoopCastOfTo(Prod, Ty);
+ Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod);
+ ++I;
+ } else {
+ // A simple mul.
+ Value *W = expandCodeFor(Op, Ty);
+ Prod = InsertNoopCastOfTo(Prod, Ty);
+ // Canonicalize a constant to the RHS.
+ if (isa<Constant>(Prod)) std::swap(Prod, W);
+ Prod = InsertBinop(Instruction::Mul, Prod, W);
+ ++I;
+ }
}
- // -1 * ... ---> 0 - ...
- if (FirstOp == 1)
- V = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), V);
- return V;
+ return Prod;
}
Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
Base = A->getStart();
Rest = SE.getAddExpr(Rest,
- SE.getAddRecExpr(SE.getIntegerSCEV(0, A->getType()),
+ SE.getAddRecExpr(SE.getConstant(A->getType(), 0),
A->getStepRecurrence(SE),
A->getLoop()));
}
// Reuse a previously-inserted PHI, if present.
for (BasicBlock::iterator I = L->getHeader()->begin();
PHINode *PN = dyn_cast<PHINode>(I); ++I)
- if (isInsertedInstruction(PN) && SE.getSCEV(PN) == Normalized)
- return PN;
+ if (SE.isSCEVable(PN->getType()) &&
+ (SE.getEffectiveSCEVType(PN->getType()) ==
+ SE.getEffectiveSCEVType(Normalized->getType())) &&
+ SE.getSCEV(PN) == Normalized)
+ if (BasicBlock *LatchBlock = L->getLoopLatch()) {
+ Instruction *IncV =
+ cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
+
+ // Determine if this is a well-behaved chain of instructions leading
+ // back to the PHI. It probably will be, if we're scanning an inner
+ // loop already visited by LSR for example, but it wouldn't have
+ // to be.
+ do {
+ if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV)) {
+ IncV = 0;
+ break;
+ }
+ // If any of the operands don't dominate the insert position, bail.
+ // Addrec operands are always loop-invariant, so this can only happen
+ // if there are instructions which haven't been hoisted.
+ for (User::op_iterator OI = IncV->op_begin()+1,
+ OE = IncV->op_end(); OI != OE; ++OI)
+ if (Instruction *OInst = dyn_cast<Instruction>(OI))
+ if (!SE.DT->dominates(OInst, IVIncInsertPos)) {
+ IncV = 0;
+ break;
+ }
+ if (!IncV)
+ break;
+ // Advance to the next instruction.
+ IncV = dyn_cast<Instruction>(IncV->getOperand(0));
+ if (!IncV)
+ break;
+ if (IncV->mayHaveSideEffects()) {
+ IncV = 0;
+ break;
+ }
+ } while (IncV != PN);
+
+ if (IncV) {
+ // Ok, the add recurrence looks usable.
+ // Remember this PHI, even in post-inc mode.
+ InsertedValues.insert(PN);
+ // Remember the increment.
+ IncV = cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock));
+ rememberInstruction(IncV);
+ if (L == IVIncInsertLoop)
+ do {
+ if (SE.DT->dominates(IncV, IVIncInsertPos))
+ break;
+ // Make sure the increment is where we want it. But don't move it
+ // down past a potential existing post-inc user.
+ IncV->moveBefore(IVIncInsertPos);
+ IVIncInsertPos = IncV;
+ IncV = cast<Instruction>(IncV->getOperand(0));
+ } while (IncV != PN);
+ return PN;
+ }
+ }
// Save the original insertion point so we can restore it when we're done.
BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
// negative, insert a sub instead of an add for the increment (unless it's a
// constant, because subtracts of constants are canonicalized to adds).
const SCEV *Step = Normalized->getStepRecurrence(SE);
- bool isPointer = isa<PointerType>(ExpandTy);
+ bool isPointer = ExpandTy->isPointerTy();
bool isNegative = !isPointer && isNonConstantNegative(Step);
if (isNegative)
Step = SE.getNegativeSCEV(Step);
// Restore the original insert point.
if (SaveInsertBB)
- Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
+ restoreInsertPoint(SaveInsertBB, SaveInsertPt);
// Remember this PHI, even in post-inc mode.
InsertedValues.insert(PN);
// Determine a normalized form of this expression, which is the expression
// before any post-inc adjustment is made.
const SCEVAddRecExpr *Normalized = S;
- if (L == PostIncLoop) {
- const SCEV *Step = S->getStepRecurrence(SE);
- Normalized = cast<SCEVAddRecExpr>(SE.getMinusSCEV(S, Step));
+ if (PostIncLoops.count(L)) {
+ PostIncLoopSet Loops;
+ Loops.insert(L);
+ Normalized =
+ cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, 0, 0,
+ Loops, SE, *SE.DT));
}
// Strip off any non-loop-dominating component from the addrec start.
const SCEV *PostLoopOffset = 0;
if (!Start->properlyDominates(L->getHeader(), SE.DT)) {
PostLoopOffset = Start;
- Start = SE.getIntegerSCEV(0, Normalized->getType());
+ Start = SE.getConstant(Normalized->getType(), 0);
Normalized =
cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start,
Normalized->getStepRecurrence(SE),
// Strip off any non-loop-dominating component from the addrec step.
const SCEV *Step = Normalized->getStepRecurrence(SE);
const SCEV *PostLoopScale = 0;
- if (!Step->hasComputableLoopEvolution(L) &&
- !Step->dominates(L->getHeader(), SE.DT)) {
+ if (!Step->dominates(L->getHeader(), SE.DT)) {
PostLoopScale = Step;
- Step = SE.getIntegerSCEV(1, Normalized->getType());
+ Step = SE.getConstant(Normalized->getType(), 1);
Normalized =
cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start, Step,
Normalized->getLoop()));
const Type *ExpandTy = PostLoopScale ? IntTy : STy;
PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy);
- // Accomodate post-inc mode, if necessary.
+ // Accommodate post-inc mode, if necessary.
Value *Result;
- if (L != PostIncLoop)
+ if (!PostIncLoops.count(L))
Result = PN;
else {
// In PostInc mode, use the post-incremented value.
// Re-apply any non-loop-dominating scale.
if (PostLoopScale) {
+ Result = InsertNoopCastOfTo(Result, IntTy);
Result = Builder.CreateMul(Result,
expandCodeFor(PostLoopScale, IntTy));
rememberInstruction(Result);
const SCEV *const OffsetArray[1] = { PostLoopOffset };
Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result);
} else {
+ Result = InsertNoopCastOfTo(Result, IntTy);
Result = Builder.CreateAdd(Result,
expandCodeFor(PostLoopOffset, IntTy));
rememberInstruction(Result);
// First check for an existing canonical IV in a suitable type.
PHINode *CanonicalIV = 0;
if (PHINode *PN = L->getCanonicalInductionVariable())
- if (SE.isSCEVable(PN->getType()) &&
- isa<IntegerType>(SE.getEffectiveSCEVType(PN->getType())) &&
- SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
+ if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
CanonicalIV = PN;
// Rewrite an AddRec in terms of the canonical induction variable, if
if (CanonicalIV &&
SE.getTypeSizeInBits(CanonicalIV->getType()) >
SE.getTypeSizeInBits(Ty)) {
- const SmallVectorImpl<const SCEV *> &Ops = S->getOperands();
- SmallVector<const SCEV *, 4> NewOps(Ops.size());
- for (unsigned i = 0, e = Ops.size(); i != e; ++i)
- NewOps[i] = SE.getAnyExtendExpr(Ops[i], CanonicalIV->getType());
+ SmallVector<const SCEV *, 4> NewOps(S->getNumOperands());
+ for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i)
+ NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType());
Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop()));
BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
BasicBlock::iterator NewInsertPt =
llvm::next(BasicBlock::iterator(cast<Instruction>(V)));
- while (isa<PHINode>(NewInsertPt)) ++NewInsertPt;
+ while (isa<PHINode>(NewInsertPt) || isa<DbgInfoIntrinsic>(NewInsertPt))
+ ++NewInsertPt;
V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
NewInsertPt);
- Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
+ restoreInsertPoint(SaveInsertBB, SaveInsertPt);
return V;
}
// {X,+,F} --> X + {0,+,F}
if (!S->getStart()->isZero()) {
- const SmallVectorImpl<const SCEV *> &SOperands = S->getOperands();
- SmallVector<const SCEV *, 4> NewOps(SOperands.begin(), SOperands.end());
- NewOps[0] = SE.getIntegerSCEV(0, Ty);
+ SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end());
+ NewOps[0] = SE.getConstant(Ty, 0);
const SCEV *Rest = SE.getAddRecExpr(NewOps, L);
// Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
SE.getUnknown(expand(Rest))));
}
- // {0,+,1} --> Insert a canonical induction variable into the loop!
- if (S->isAffine() &&
- S->getOperand(1) == SE.getIntegerSCEV(1, Ty)) {
- // If there's a canonical IV, just use it.
- if (CanonicalIV) {
- assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
- "IVs with types different from the canonical IV should "
- "already have been handled!");
- return CanonicalIV;
- }
-
+ // If we don't yet have a canonical IV, create one.
+ if (!CanonicalIV) {
// Create and insert the PHI node for the induction variable in the
// specified loop.
BasicBlock *Header = L->getHeader();
- PHINode *PN = PHINode::Create(Ty, "indvar", Header->begin());
- rememberInstruction(PN);
+ CanonicalIV = PHINode::Create(Ty, "indvar", Header->begin());
+ rememberInstruction(CanonicalIV);
Constant *One = ConstantInt::get(Ty, 1);
for (pred_iterator HPI = pred_begin(Header), HPE = pred_end(Header);
- HPI != HPE; ++HPI)
- if (L->contains(*HPI)) {
+ HPI != HPE; ++HPI) {
+ BasicBlock *HP = *HPI;
+ if (L->contains(HP)) {
// Insert a unit add instruction right before the terminator
// corresponding to the back-edge.
- Instruction *Add = BinaryOperator::CreateAdd(PN, One, "indvar.next",
- (*HPI)->getTerminator());
+ Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One,
+ "indvar.next",
+ HP->getTerminator());
rememberInstruction(Add);
- PN->addIncoming(Add, *HPI);
+ CanonicalIV->addIncoming(Add, HP);
} else {
- PN->addIncoming(Constant::getNullValue(Ty), *HPI);
+ CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP);
}
+ }
+ }
+
+ // {0,+,1} --> Insert a canonical induction variable into the loop!
+ if (S->isAffine() && S->getOperand(1)->isOne()) {
+ assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
+ "IVs with types different from the canonical IV should "
+ "already have been handled!");
+ return CanonicalIV;
}
// {0,+,F} --> {0,+,1} * F
- // Get the canonical induction variable I for this loop.
- Value *I = CanonicalIV ?
- CanonicalIV :
- getOrInsertCanonicalInductionVariable(L, Ty);
// If this is a simple linear addrec, emit it now as a special case.
if (S->isAffine()) // {0,+,F} --> i*F
return
expand(SE.getTruncateOrNoop(
- SE.getMulExpr(SE.getUnknown(I),
+ SE.getMulExpr(SE.getUnknown(CanonicalIV),
SE.getNoopOrAnyExtend(S->getOperand(1),
- I->getType())),
+ CanonicalIV->getType())),
Ty));
// If this is a chain of recurrences, turn it into a closed form, using the
// folders, then expandCodeFor the closed form. This allows the folders to
// simplify the expression without having to build a bunch of special code
// into this folder.
- const SCEV *IH = SE.getUnknown(I); // Get I as a "symbolic" SCEV.
+ const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV.
// Promote S up to the canonical IV type, if the cast is foldable.
const SCEV *NewS = S;
- const SCEV *Ext = SE.getNoopOrAnyExtend(S, I->getType());
+ const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType());
if (isa<SCEVAddRecExpr>(Ext))
NewS = Ext;
return LHS;
}
-Value *SCEVExpander::visitFieldOffsetExpr(const SCEVFieldOffsetExpr *S) {
- return ConstantExpr::getOffsetOf(S->getStructType(), S->getFieldNo());
-}
-
-Value *SCEVExpander::visitAllocSizeExpr(const SCEVAllocSizeExpr *S) {
- return ConstantExpr::getSizeOf(S->getAllocType());
+Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty,
+ Instruction *I) {
+ BasicBlock::iterator IP = I;
+ while (isInsertedInstruction(IP) || isa<DbgInfoIntrinsic>(IP))
+ ++IP;
+ Builder.SetInsertPoint(IP->getParent(), IP);
+ return expandCodeFor(SH, Ty);
}
Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty) {
Instruction *InsertPt = Builder.GetInsertPoint();
for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
L = L->getParentLoop())
- if (S->isLoopInvariant(L)) {
+ if (SE.isLoopInvariant(S, L)) {
if (!L) break;
if (BasicBlock *Preheader = L->getLoopPreheader())
InsertPt = Preheader->getTerminator();
// If the SCEV is computable at this level, insert it into the header
// after the PHIs (and after any other instructions that we've inserted
// there) so that it is guaranteed to dominate any user inside the loop.
- if (L && S->hasComputableLoopEvolution(L))
+ if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L))
InsertPt = L->getHeader()->getFirstNonPHI();
- while (isInsertedInstruction(InsertPt))
+ while (isInsertedInstruction(InsertPt) || isa<DbgInfoIntrinsic>(InsertPt))
InsertPt = llvm::next(BasicBlock::iterator(InsertPt));
break;
}
Value *V = visit(S);
// Remember the expanded value for this SCEV at this location.
- if (!PostIncLoop)
+ if (PostIncLoops.empty())
InsertedExpressions[std::make_pair(S, InsertPt)] = V;
- Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
+ restoreInsertPoint(SaveInsertBB, SaveInsertPt);
return V;
}
+void SCEVExpander::rememberInstruction(Value *I) {
+ if (!PostIncLoops.empty())
+ InsertedPostIncValues.insert(I);
+ else
+ InsertedValues.insert(I);
+
+ // If we just claimed an existing instruction and that instruction had
+ // been the insert point, adjust the insert point forward so that
+ // subsequently inserted code will be dominated.
+ if (Builder.GetInsertPoint() == I) {
+ BasicBlock::iterator It = cast<Instruction>(I);
+ do { ++It; } while (isInsertedInstruction(It) ||
+ isa<DbgInfoIntrinsic>(It));
+ Builder.SetInsertPoint(Builder.GetInsertBlock(), It);
+ }
+}
+
+void SCEVExpander::restoreInsertPoint(BasicBlock *BB, BasicBlock::iterator I) {
+ // If we acquired more instructions since the old insert point was saved,
+ // advance past them.
+ while (isInsertedInstruction(I) || isa<DbgInfoIntrinsic>(I)) ++I;
+
+ Builder.SetInsertPoint(BB, I);
+}
+
/// getOrInsertCanonicalInductionVariable - This method returns the
/// canonical induction variable of the specified type for the specified
/// loop (inserting one if there is none). A canonical induction variable
/// starts at zero and steps by one on each iteration.
-Value *
+PHINode *
SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
const Type *Ty) {
- assert(Ty->isInteger() && "Can only insert integer induction variables!");
- const SCEV *H = SE.getAddRecExpr(SE.getIntegerSCEV(0, Ty),
- SE.getIntegerSCEV(1, Ty), L);
+ assert(Ty->isIntegerTy() && "Can only insert integer induction variables!");
+
+ // Build a SCEV for {0,+,1}<L>.
+ const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0),
+ SE.getConstant(Ty, 1), L);
+
+ // Emit code for it.
BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
- Value *V = expandCodeFor(H, 0, L->getHeader()->begin());
+ PHINode *V = cast<PHINode>(expandCodeFor(H, 0, L->getHeader()->begin()));
if (SaveInsertBB)
- Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
+ restoreInsertPoint(SaveInsertBB, SaveInsertPt);
+
return V;
}
projects / oota-llvm.git / blobdiff