//
//===----------------------------------------------------------------------===//
-#define DEBUG_TYPE "indvars"
#include "llvm/Transforms/Scalar.h"
-#include "llvm/BasicBlock.h"
-#include "llvm/Constants.h"
-#include "llvm/Instructions.h"
-#include "llvm/IntrinsicInst.h"
-#include "llvm/LLVMContext.h"
-#include "llvm/Type.h"
-#include "llvm/Analysis/Dominators.h"
-#include "llvm/Analysis/ScalarEvolutionExpander.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/LoopPass.h"
-#include "llvm/Support/CFG.h"
+#include "llvm/Analysis/ScalarEvolutionExpander.h"
+#include "llvm/Analysis/TargetLibraryInfo.h"
+#include "llvm/Analysis/TargetTransformInfo.h"
+#include "llvm/IR/BasicBlock.h"
+#include "llvm/IR/CFG.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/IR/Dominators.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/IntrinsicInst.h"
+#include "llvm/IR/LLVMContext.h"
+#include "llvm/IR/Type.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
-#include "llvm/Transforms/Utils/Local.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
+#include "llvm/Transforms/Utils/Local.h"
#include "llvm/Transforms/Utils/SimplifyIndVar.h"
-#include "llvm/DataLayout.h"
-#include "llvm/Target/TargetLibraryInfo.h"
-#include "llvm/ADT/DenseMap.h"
-#include "llvm/ADT/SmallVector.h"
-#include "llvm/ADT/Statistic.h"
using namespace llvm;
+#define DEBUG_TYPE "indvars"
+
STATISTIC(NumWidened , "Number of indvars widened");
STATISTIC(NumReplaced , "Number of exit values replaced");
STATISTIC(NumLFTR , "Number of loop exit tests replaced");
"verify-indvars", cl::Hidden,
cl::desc("Verify the ScalarEvolution result after running indvars"));
+static cl::opt<bool> ReduceLiveIVs("liv-reduce", cl::Hidden,
+ cl::desc("Reduce live induction variables."));
+
namespace {
class IndVarSimplify : public LoopPass {
- LoopInfo *LI;
- ScalarEvolution *SE;
- DominatorTree *DT;
- DataLayout *TD;
- TargetLibraryInfo *TLI;
+ LoopInfo *LI;
+ ScalarEvolution *SE;
+ DominatorTree *DT;
+ TargetLibraryInfo *TLI;
+ const TargetTransformInfo *TTI;
SmallVector<WeakVH, 16> DeadInsts;
bool Changed;
public:
static char ID; // Pass identification, replacement for typeid
- IndVarSimplify() : LoopPass(ID), LI(0), SE(0), DT(0), TD(0),
- Changed(false) {
+ IndVarSimplify()
+ : LoopPass(ID), LI(nullptr), SE(nullptr), DT(nullptr), Changed(false) {
initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
}
- virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
+ bool runOnLoop(Loop *L, LPPassManager &LPM) override;
- virtual void getAnalysisUsage(AnalysisUsage &AU) const {
- AU.addRequired<DominatorTree>();
- AU.addRequired<LoopInfo>();
+ void getAnalysisUsage(AnalysisUsage &AU) const override {
+ AU.addRequired<DominatorTreeWrapperPass>();
+ AU.addRequired<LoopInfoWrapperPass>();
AU.addRequired<ScalarEvolution>();
AU.addRequiredID(LoopSimplifyID);
AU.addRequiredID(LCSSAID);
}
private:
- virtual void releaseMemory() {
+ void releaseMemory() override {
DeadInsts.clear();
}
char IndVarSimplify::ID = 0;
INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
"Induction Variable Simplification", false, false)
-INITIALIZE_PASS_DEPENDENCY(DominatorTree)
-INITIALIZE_PASS_DEPENDENCY(LoopInfo)
+INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
+INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
INITIALIZE_PASS_DEPENDENCY(LCSSA)
if (!PHI)
return User;
- Instruction *InsertPt = 0;
+ Instruction *InsertPt = nullptr;
for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) {
if (PHI->getIncomingValue(i) != Def)
continue;
// an add or increment value can not be represented by an integer.
BinaryOperator *Incr =
dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge));
- if (Incr == 0 || Incr->getOpcode() != Instruction::FAdd) return;
+ if (Incr == nullptr || Incr->getOpcode() != Instruction::FAdd) return;
// If this is not an add of the PHI with a constantfp, or if the constant fp
// is not an integer, bail out.
ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1));
int64_t IncValue;
- if (IncValueVal == 0 || Incr->getOperand(0) != PN ||
+ if (IncValueVal == nullptr || Incr->getOperand(0) != PN ||
!ConvertToSInt(IncValueVal->getValueAPF(), IncValue))
return;
// Check Incr uses. One user is PN and the other user is an exit condition
// used by the conditional terminator.
- Value::use_iterator IncrUse = Incr->use_begin();
+ Value::user_iterator IncrUse = Incr->user_begin();
Instruction *U1 = cast<Instruction>(*IncrUse++);
- if (IncrUse == Incr->use_end()) return;
+ if (IncrUse == Incr->user_end()) return;
Instruction *U2 = cast<Instruction>(*IncrUse++);
- if (IncrUse != Incr->use_end()) return;
+ if (IncrUse != Incr->user_end()) return;
// Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't
// only used by a branch, we can't transform it.
FCmpInst *Compare = dyn_cast<FCmpInst>(U1);
if (!Compare)
Compare = dyn_cast<FCmpInst>(U2);
- if (Compare == 0 || !Compare->hasOneUse() ||
- !isa<BranchInst>(Compare->use_back()))
+ if (!Compare || !Compare->hasOneUse() ||
+ !isa<BranchInst>(Compare->user_back()))
return;
- BranchInst *TheBr = cast<BranchInst>(Compare->use_back());
+ BranchInst *TheBr = cast<BranchInst>(Compare->user_back());
// We need to verify that the branch actually controls the iteration count
// of the loop. If not, the new IV can overflow and no one will notice.
// transform it.
ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
int64_t ExitValue;
- if (ExitValueVal == 0 ||
+ if (ExitValueVal == nullptr ||
!ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
return;
unsigned NumPreds = PN->getNumIncomingValues();
+ // We would like to be able to RAUW single-incoming value PHI nodes. We
+ // have to be certain this is safe even when this is an LCSSA PHI node.
+ // While the computed exit value is no longer varying in *this* loop, the
+ // exit block may be an exit block for an outer containing loop as well,
+ // the exit value may be varying in the outer loop, and thus it may still
+ // require an LCSSA PHI node. The safe case is when this is
+ // single-predecessor PHI node (LCSSA) and the exit block containing it is
+ // part of the enclosing loop, or this is the outer most loop of the nest.
+ // In either case the exit value could (at most) be varying in the same
+ // loop body as the phi node itself. Thus if it is in turn used outside of
+ // an enclosing loop it will only be via a separate LCSSA node.
+ bool LCSSASafePhiForRAUW =
+ NumPreds == 1 &&
+ (!L->getParentLoop() || L->getParentLoop() == LI->getLoopFor(ExitBB));
+
// Iterate over all of the PHI nodes.
BasicBlock::iterator BBI = ExitBB->begin();
while ((PN = dyn_cast<PHINode>(BBI++))) {
// and varies predictably *inside* the loop. Evaluate the value it
// contains when the loop exits, if possible.
const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
- if (!SE->isLoopInvariant(ExitValue, L))
+ if (!SE->isLoopInvariant(ExitValue, L) ||
+ !isSafeToExpand(ExitValue, *SE))
continue;
+ // Computing the value outside of the loop brings no benefit if :
+ // - it is definitely used inside the loop in a way which can not be
+ // optimized away.
+ // - no use outside of the loop can take advantage of hoisting the
+ // computation out of the loop
+ if (ExitValue->getSCEVType()>=scMulExpr) {
+ unsigned NumHardInternalUses = 0;
+ unsigned NumSoftExternalUses = 0;
+ unsigned NumUses = 0;
+ for (auto IB = Inst->user_begin(), IE = Inst->user_end();
+ IB != IE && NumUses <= 6; ++IB) {
+ Instruction *UseInstr = cast<Instruction>(*IB);
+ unsigned Opc = UseInstr->getOpcode();
+ NumUses++;
+ if (L->contains(UseInstr)) {
+ if (Opc == Instruction::Call || Opc == Instruction::Ret)
+ NumHardInternalUses++;
+ } else {
+ if (Opc == Instruction::PHI) {
+ // Do not count the Phi as a use. LCSSA may have inserted
+ // plenty of trivial ones.
+ NumUses--;
+ for (auto PB = UseInstr->user_begin(),
+ PE = UseInstr->user_end();
+ PB != PE && NumUses <= 6; ++PB, ++NumUses) {
+ unsigned PhiOpc = cast<Instruction>(*PB)->getOpcode();
+ if (PhiOpc != Instruction::Call && PhiOpc != Instruction::Ret)
+ NumSoftExternalUses++;
+ }
+ continue;
+ }
+ if (Opc != Instruction::Call && Opc != Instruction::Ret)
+ NumSoftExternalUses++;
+ }
+ }
+ if (NumUses <= 6 && NumHardInternalUses && !NumSoftExternalUses)
+ continue;
+ }
+
Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
if (isInstructionTriviallyDead(Inst, TLI))
DeadInsts.push_back(Inst);
- if (NumPreds == 1) {
- // Completely replace a single-pred PHI. This is safe, because the
- // NewVal won't be variant in the loop, so we don't need an LCSSA phi
- // node anymore.
+ // If we determined that this PHI is safe to replace even if an LCSSA
+ // PHI, do so.
+ if (LCSSASafePhiForRAUW) {
PN->replaceAllUsesWith(ExitVal);
PN->eraseFromParent();
}
}
- if (NumPreds != 1) {
- // Clone the PHI and delete the original one. This lets IVUsers and
- // any other maps purge the original user from their records.
+
+ // If we were unable to completely replace the PHI node, clone the PHI
+ // and delete the original one. This lets IVUsers and any other maps
+ // purge the original user from their records.
+ if (!LCSSASafePhiForRAUW) {
PHINode *NewPN = cast<PHINode>(PN->clone());
NewPN->takeName(PN);
NewPN->insertBefore(PN);
struct WideIVInfo {
PHINode *NarrowIV;
Type *WidestNativeType; // Widest integer type created [sz]ext
- bool IsSigned; // Was an sext user seen before a zext?
+ bool IsSigned; // Was a sext user seen before a zext?
- WideIVInfo() : NarrowIV(0), WidestNativeType(0), IsSigned(false) {}
- };
-
- class WideIVVisitor : public IVVisitor {
- ScalarEvolution *SE;
- const DataLayout *TD;
-
- public:
- WideIVInfo WI;
-
- WideIVVisitor(PHINode *NarrowIV, ScalarEvolution *SCEV,
- const DataLayout *TData) :
- SE(SCEV), TD(TData) { WI.NarrowIV = NarrowIV; }
-
- // Implement the interface used by simplifyUsersOfIV.
- virtual void visitCast(CastInst *Cast);
+ WideIVInfo() : NarrowIV(nullptr), WidestNativeType(nullptr),
+ IsSigned(false) {}
};
}
/// visitCast - Update information about the induction variable that is
/// extended by this sign or zero extend operation. This is used to determine
/// the final width of the IV before actually widening it.
-void WideIVVisitor::visitCast(CastInst *Cast) {
+static void visitIVCast(CastInst *Cast, WideIVInfo &WI, ScalarEvolution *SE,
+ const TargetTransformInfo *TTI) {
bool IsSigned = Cast->getOpcode() == Instruction::SExt;
if (!IsSigned && Cast->getOpcode() != Instruction::ZExt)
return;
Type *Ty = Cast->getType();
uint64_t Width = SE->getTypeSizeInBits(Ty);
- if (TD && !TD->isLegalInteger(Width))
+ if (!Cast->getModule()->getDataLayout().isLegalInteger(Width))
+ return;
+
+ // Cast is either an sext or zext up to this point.
+ // We should not widen an indvar if arithmetics on the wider indvar are more
+ // expensive than those on the narrower indvar. We check only the cost of ADD
+ // because at least an ADD is required to increment the induction variable. We
+ // could compute more comprehensively the cost of all instructions on the
+ // induction variable when necessary.
+ if (TTI &&
+ TTI->getArithmeticInstrCost(Instruction::Add, Ty) >
+ TTI->getArithmeticInstrCost(Instruction::Add,
+ Cast->getOperand(0)->getType())) {
return;
+ }
if (!WI.WidestNativeType) {
WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
Instruction *NarrowUse;
Instruction *WideDef;
- NarrowIVDefUse(): NarrowDef(0), NarrowUse(0), WideDef(0) {}
+ NarrowIVDefUse(): NarrowDef(nullptr), NarrowUse(nullptr), WideDef(nullptr) {}
NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD):
NarrowDef(ND), NarrowUse(NU), WideDef(WD) {}
L(LI->getLoopFor(OrigPhi->getParent())),
SE(SEv),
DT(DTree),
- WidePhi(0),
- WideInc(0),
- WideIncExpr(0),
+ WidePhi(nullptr),
+ WideInc(nullptr),
+ WideIncExpr(nullptr),
DeadInsts(DI) {
assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
}
const SCEVAddRecExpr* GetExtendedOperandRecurrence(NarrowIVDefUse DU);
+ const SCEV *GetSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
+ unsigned OpCode) const;
+
Instruction *WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter);
+ bool WidenLoopCompare(NarrowIVDefUse DU);
+
void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
};
} // anonymous namespace
unsigned Opcode = DU.NarrowUse->getOpcode();
switch (Opcode) {
default:
- return 0;
+ return nullptr;
case Instruction::Add:
case Instruction::Mul:
case Instruction::UDiv:
}
}
+const SCEV *WidenIV::GetSCEVByOpCode(const SCEV *LHS, const SCEV *RHS,
+ unsigned OpCode) const {
+ if (OpCode == Instruction::Add)
+ return SE->getAddExpr(LHS, RHS);
+ if (OpCode == Instruction::Sub)
+ return SE->getMinusSCEV(LHS, RHS);
+ if (OpCode == Instruction::Mul)
+ return SE->getMulExpr(LHS, RHS);
+
+ llvm_unreachable("Unsupported opcode.");
+}
+
/// No-wrap operations can transfer sign extension of their result to their
/// operands. Generate the SCEV value for the widened operation without
/// actually modifying the IR yet. If the expression after extending the
/// operands is an AddRec for this loop, return it.
const SCEVAddRecExpr* WidenIV::GetExtendedOperandRecurrence(NarrowIVDefUse DU) {
+
// Handle the common case of add<nsw/nuw>
- if (DU.NarrowUse->getOpcode() != Instruction::Add)
- return 0;
+ const unsigned OpCode = DU.NarrowUse->getOpcode();
+ // Only Add/Sub/Mul instructions supported yet.
+ if (OpCode != Instruction::Add && OpCode != Instruction::Sub &&
+ OpCode != Instruction::Mul)
+ return nullptr;
// One operand (NarrowDef) has already been extended to WideDef. Now determine
// if extending the other will lead to a recurrence.
- unsigned ExtendOperIdx = DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0;
+ const unsigned ExtendOperIdx =
+ DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0;
assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU");
- const SCEV *ExtendOperExpr = 0;
+ const SCEV *ExtendOperExpr = nullptr;
const OverflowingBinaryOperator *OBO =
cast<OverflowingBinaryOperator>(DU.NarrowUse);
if (IsSigned && OBO->hasNoSignedWrap())
ExtendOperExpr = SE->getZeroExtendExpr(
SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
else
- return 0;
+ return nullptr;
- // When creating this AddExpr, don't apply the current operations NSW or NUW
+ // When creating this SCEV expr, don't apply the current operations NSW or NUW
// flags. This instruction may be guarded by control flow that the no-wrap
// behavior depends on. Non-control-equivalent instructions can be mapped to
// the same SCEV expression, and it would be incorrect to transfer NSW/NUW
// semantics to those operations.
- const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(
- SE->getAddExpr(SE->getSCEV(DU.WideDef), ExtendOperExpr));
+ const SCEV *lhs = SE->getSCEV(DU.WideDef);
+ const SCEV *rhs = ExtendOperExpr;
+
+ // Let's swap operands to the initial order for the case of non-commutative
+ // operations, like SUB. See PR21014.
+ if (ExtendOperIdx == 0)
+ std::swap(lhs, rhs);
+ const SCEVAddRecExpr *AddRec =
+ dyn_cast<SCEVAddRecExpr>(GetSCEVByOpCode(lhs, rhs, OpCode));
if (!AddRec || AddRec->getLoop() != L)
- return 0;
+ return nullptr;
return AddRec;
}
/// recurrence. Otherwise return NULL.
const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) {
if (!SE->isSCEVable(NarrowUse->getType()))
- return 0;
+ return nullptr;
const SCEV *NarrowExpr = SE->getSCEV(NarrowUse);
if (SE->getTypeSizeInBits(NarrowExpr->getType())
>= SE->getTypeSizeInBits(WideType)) {
// NarrowUse implicitly widens its operand. e.g. a gep with a narrow
// index. So don't follow this use.
- return 0;
+ return nullptr;
}
const SCEV *WideExpr = IsSigned ?
SE->getZeroExtendExpr(NarrowExpr, WideType);
const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
if (!AddRec || AddRec->getLoop() != L)
- return 0;
+ return nullptr;
return AddRec;
}
+/// This IV user cannot be widen. Replace this use of the original narrow IV
+/// with a truncation of the new wide IV to isolate and eliminate the narrow IV.
+static void truncateIVUse(NarrowIVDefUse DU, DominatorTree *DT) {
+ DEBUG(dbgs() << "INDVARS: Truncate IV " << *DU.WideDef
+ << " for user " << *DU.NarrowUse << "\n");
+ IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
+ Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType());
+ DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc);
+}
+
+/// If the narrow use is a compare instruction, then widen the compare
+// (and possibly the other operand). The extend operation is hoisted into the
+// loop preheader as far as possible.
+bool WidenIV::WidenLoopCompare(NarrowIVDefUse DU) {
+ ICmpInst *Cmp = dyn_cast<ICmpInst>(DU.NarrowUse);
+ if (!Cmp)
+ return false;
+
+ // Sign of IV user and compare must match.
+ if (IsSigned != CmpInst::isSigned(Cmp->getPredicate()))
+ return false;
+
+ Value *Op = Cmp->getOperand(Cmp->getOperand(0) == DU.NarrowDef ? 1 : 0);
+ unsigned CastWidth = SE->getTypeSizeInBits(Op->getType());
+ unsigned IVWidth = SE->getTypeSizeInBits(WideType);
+ assert (CastWidth <= IVWidth && "Unexpected width while widening compare.");
+
+ // Widen the compare instruction.
+ IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
+ DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
+
+ // Widen the other operand of the compare, if necessary.
+ if (CastWidth < IVWidth) {
+ Value *ExtOp = getExtend(Op, WideType, IsSigned, Cmp);
+ DU.NarrowUse->replaceUsesOfWith(Op, ExtOp);
+ }
+ return true;
+}
+
/// WidenIVUse - Determine whether an individual user of the narrow IV can be
/// widened. If so, return the wide clone of the user.
Instruction *WidenIV::WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter) {
// Stop traversing the def-use chain at inner-loop phis or post-loop phis.
- if (isa<PHINode>(DU.NarrowUse) &&
- LI->getLoopFor(DU.NarrowUse->getParent()) != L)
- return 0;
-
+ if (PHINode *UsePhi = dyn_cast<PHINode>(DU.NarrowUse)) {
+ if (LI->getLoopFor(UsePhi->getParent()) != L) {
+ // For LCSSA phis, sink the truncate outside the loop.
+ // After SimplifyCFG most loop exit targets have a single predecessor.
+ // Otherwise fall back to a truncate within the loop.
+ if (UsePhi->getNumOperands() != 1)
+ truncateIVUse(DU, DT);
+ else {
+ PHINode *WidePhi =
+ PHINode::Create(DU.WideDef->getType(), 1, UsePhi->getName() + ".wide",
+ UsePhi);
+ WidePhi->addIncoming(DU.WideDef, UsePhi->getIncomingBlock(0));
+ IRBuilder<> Builder(WidePhi->getParent()->getFirstInsertionPt());
+ Value *Trunc = Builder.CreateTrunc(WidePhi, DU.NarrowDef->getType());
+ UsePhi->replaceAllUsesWith(Trunc);
+ DeadInsts.push_back(UsePhi);
+ DEBUG(dbgs() << "INDVARS: Widen lcssa phi " << *UsePhi
+ << " to " << *WidePhi << "\n");
+ }
+ return nullptr;
+ }
+ }
// Our raison d'etre! Eliminate sign and zero extension.
if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) {
Value *NewDef = DU.WideDef;
// push the uses of WideDef here.
// No further widening is needed. The deceased [sz]ext had done it for us.
- return 0;
+ return nullptr;
}
// Does this user itself evaluate to a recurrence after widening?
const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse);
+ if (!WideAddRec)
+ WideAddRec = GetExtendedOperandRecurrence(DU);
+
if (!WideAddRec) {
- WideAddRec = GetExtendedOperandRecurrence(DU);
- }
- if (!WideAddRec) {
+ // If use is a loop condition, try to promote the condition instead of
+ // truncating the IV first.
+ if (WidenLoopCompare(DU))
+ return nullptr;
+
// This user does not evaluate to a recurence after widening, so don't
// follow it. Instead insert a Trunc to kill off the original use,
// eventually isolating the original narrow IV so it can be removed.
- IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT));
- Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType());
- DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc);
- return 0;
+ truncateIVUse(DU, DT);
+ return nullptr;
}
// Assume block terminators cannot evaluate to a recurrence. We can't to
// insert a Trunc after a terminator if there happens to be a critical edge.
// Reuse the IV increment that SCEVExpander created as long as it dominates
// NarrowUse.
- Instruction *WideUse = 0;
+ Instruction *WideUse = nullptr;
if (WideAddRec == WideIncExpr
&& Rewriter.hoistIVInc(WideInc, DU.NarrowUse))
WideUse = WideInc;
else {
WideUse = CloneIVUser(DU);
if (!WideUse)
- return 0;
+ return nullptr;
}
// Evaluation of WideAddRec ensured that the narrow expression could be
// extended outside the loop without overflow. This suggests that the wide use
DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
<< ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
DeadInsts.push_back(WideUse);
- return 0;
+ return nullptr;
}
// Returning WideUse pushes it on the worklist.
/// pushNarrowIVUsers - Add eligible users of NarrowDef to NarrowIVUsers.
///
void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) {
- for (Value::use_iterator UI = NarrowDef->use_begin(),
- UE = NarrowDef->use_end(); UI != UE; ++UI) {
- Instruction *NarrowUse = cast<Instruction>(*UI);
+ for (User *U : NarrowDef->users()) {
+ Instruction *NarrowUser = cast<Instruction>(U);
// Handle data flow merges and bizarre phi cycles.
- if (!Widened.insert(NarrowUse))
+ if (!Widened.insert(NarrowUser).second)
continue;
- NarrowIVUsers.push_back(NarrowIVDefUse(NarrowDef, NarrowUse, WideDef));
+ NarrowIVUsers.push_back(NarrowIVDefUse(NarrowDef, NarrowUser, WideDef));
}
}
// Is this phi an induction variable?
const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
if (!AddRec)
- return NULL;
+ return nullptr;
// Widen the induction variable expression.
const SCEV *WideIVExpr = IsSigned ?
// Can the IV be extended outside the loop without overflow?
AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
if (!AddRec || AddRec->getLoop() != L)
- return NULL;
+ return nullptr;
// An AddRec must have loop-invariant operands. Since this AddRec is
// materialized by a loop header phi, the expression cannot have any post-loop
return WidePhi;
}
+//===----------------------------------------------------------------------===//
+// Live IV Reduction - Minimize IVs live across the loop.
+//===----------------------------------------------------------------------===//
+
+
//===----------------------------------------------------------------------===//
// Simplification of IV users based on SCEV evaluation.
//===----------------------------------------------------------------------===//
+namespace {
+ class IndVarSimplifyVisitor : public IVVisitor {
+ ScalarEvolution *SE;
+ const TargetTransformInfo *TTI;
+ PHINode *IVPhi;
+
+ public:
+ WideIVInfo WI;
+
+ IndVarSimplifyVisitor(PHINode *IV, ScalarEvolution *SCEV,
+ const TargetTransformInfo *TTI,
+ const DominatorTree *DTree)
+ : SE(SCEV), TTI(TTI), IVPhi(IV) {
+ DT = DTree;
+ WI.NarrowIV = IVPhi;
+ if (ReduceLiveIVs)
+ setSplitOverflowIntrinsics();
+ }
+
+ // Implement the interface used by simplifyUsersOfIV.
+ void visitCast(CastInst *Cast) override { visitIVCast(Cast, WI, SE, TTI); }
+ };
+}
/// SimplifyAndExtend - Iteratively perform simplification on a worklist of IV
/// users. Each successive simplification may push more users which may
PHINode *CurrIV = LoopPhis.pop_back_val();
// Information about sign/zero extensions of CurrIV.
- WideIVVisitor WIV(CurrIV, SE, TD);
+ IndVarSimplifyVisitor Visitor(CurrIV, SE, TTI, DT);
- Changed |= simplifyUsersOfIV(CurrIV, SE, &LPM, DeadInsts, &WIV);
+ Changed |= simplifyUsersOfIV(CurrIV, SE, &LPM, DeadInsts, &Visitor);
- if (WIV.WI.WidestNativeType) {
- WideIVs.push_back(WIV.WI);
+ if (Visitor.WI.WidestNativeType) {
+ WideIVs.push_back(Visitor.WI);
}
} while(!LoopPhis.empty());
/// BackedgeTakenInfo. If these expressions have not been reduced, then
/// expanding them may incur additional cost (albeit in the loop preheader).
static bool isHighCostExpansion(const SCEV *S, BranchInst *BI,
- SmallPtrSet<const SCEV*, 8> &Processed,
+ SmallPtrSetImpl<const SCEV*> &Processed,
ScalarEvolution *SE) {
- if (!Processed.insert(S))
+ if (!Processed.insert(S).second)
return false;
// If the backedge-taken count is a UDiv, it's very likely a UDiv that
static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) {
Instruction *IncI = dyn_cast<Instruction>(IncV);
if (!IncI)
- return 0;
+ return nullptr;
switch (IncI->getOpcode()) {
case Instruction::Add:
if (IncI->getNumOperands() == 2)
break;
default:
- return 0;
+ return nullptr;
}
PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0));
if (Phi && Phi->getParent() == L->getHeader()) {
if (isLoopInvariant(IncI->getOperand(1), L, DT))
return Phi;
- return 0;
+ return nullptr;
}
if (IncI->getOpcode() == Instruction::GetElementPtr)
- return 0;
+ return nullptr;
// Allow add/sub to be commuted.
Phi = dyn_cast<PHINode>(IncI->getOperand(1));
if (isLoopInvariant(IncI->getOperand(0), L, DT))
return Phi;
}
- return 0;
+ return nullptr;
}
/// Return the compare guarding the loop latch, or NULL for unrecognized tests.
BasicBlock *LatchBlock = L->getLoopLatch();
// Don't bother with LFTR if the loop is not properly simplified.
if (!LatchBlock)
- return 0;
+ return nullptr;
BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
assert(BI && "expected exit branch");
/// Recursive helper for hasConcreteDef(). Unfortunately, this currently boils
/// down to checking that all operands are constant and listing instructions
/// that may hide undef.
-static bool hasConcreteDefImpl(Value *V, SmallPtrSet<Value*, 8> &Visited,
+static bool hasConcreteDefImpl(Value *V, SmallPtrSetImpl<Value*> &Visited,
unsigned Depth) {
if (isa<Constant>(V))
return !isa<UndefValue>(V);
// Optimistically handle other instructions.
for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
- if (!Visited.insert(*OI))
+ if (!Visited.insert(*OI).second)
continue;
if (!hasConcreteDefImpl(*OI, Visited, Depth+1))
return false;
int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
Value *IncV = Phi->getIncomingValue(LatchIdx);
- for (Value::use_iterator UI = Phi->use_begin(), UE = Phi->use_end();
- UI != UE; ++UI) {
- if (*UI != Cond && *UI != IncV) return false;
- }
+ for (User *U : Phi->users())
+ if (U != Cond && U != IncV) return false;
- for (Value::use_iterator UI = IncV->use_begin(), UE = IncV->use_end();
- UI != UE; ++UI) {
- if (*UI != Cond && *UI != Phi) return false;
- }
+ for (User *U : IncV->users())
+ if (U != Cond && U != Phi) return false;
return true;
}
/// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride.
/// This is difficult in general for SCEV because of potential overflow. But we
/// could at least handle constant BECounts.
-static PHINode *
-FindLoopCounter(Loop *L, const SCEV *BECount,
- ScalarEvolution *SE, DominatorTree *DT, const DataLayout *TD) {
+static PHINode *FindLoopCounter(Loop *L, const SCEV *BECount,
+ ScalarEvolution *SE, DominatorTree *DT) {
uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType());
Value *Cond =
cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition();
// Loop over all of the PHI nodes, looking for a simple counter.
- PHINode *BestPhi = 0;
- const SCEV *BestInit = 0;
+ PHINode *BestPhi = nullptr;
+ const SCEV *BestInit = nullptr;
BasicBlock *LatchBlock = L->getLoopLatch();
assert(LatchBlock && "needsLFTR should guarantee a loop latch");
// AR may be wider than BECount. With eq/ne tests overflow is immaterial.
// AR may not be a narrower type, or we may never exit.
uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType());
- if (PhiWidth < BCWidth || (TD && !TD->isLegalInteger(PhiWidth)))
+ if (PhiWidth < BCWidth ||
+ !L->getHeader()->getModule()->getDataLayout().isLegalInteger(PhiWidth))
continue;
const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
/// genLoopLimit - Help LinearFunctionTestReplace by generating a value that
/// holds the RHS of the new loop test.
static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L,
- SCEVExpander &Rewriter, ScalarEvolution *SE,
- Type *IntPtrTy) {
+ SCEVExpander &Rewriter, ScalarEvolution *SE) {
const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter");
const SCEV *IVInit = AR->getStart();
if (IndVar->getType()->isPointerTy()
&& !IVCount->getType()->isPointerTy()) {
+ // IVOffset will be the new GEP offset that is interpreted by GEP as a
+ // signed value. IVCount on the other hand represents the loop trip count,
+ // which is an unsigned value. FindLoopCounter only allows induction
+ // variables that have a positive unit stride of one. This means we don't
+ // have to handle the case of negative offsets (yet) and just need to zero
+ // extend IVCount.
Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType());
- const SCEV *IVOffset = SE->getTruncateOrSignExtend(IVCount, OfsTy);
+ const SCEV *IVOffset = SE->getTruncateOrZeroExtend(IVCount, OfsTy);
// Expand the code for the iteration count.
assert(SE->isLoopInvariant(IVOffset, L) &&
assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter");
// We could handle pointer IVs other than i8*, but we need to compensate for
// gep index scaling. See canExpandBackedgeTakenCount comments.
- assert(SE->getSizeOfExpr(
- cast<PointerType>(GEPBase->getType())->getElementType(),
- IntPtrTy)->isOne()
+ assert(SE->getSizeOfExpr(IntegerType::getInt64Ty(IndVar->getContext()),
+ cast<PointerType>(GEPBase->getType())->getElementType())->isOne()
&& "unit stride pointer IV must be i8*");
IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
- return Builder.CreateGEP(GEPBase, GEPOffset, "lftr.limit");
+ return Builder.CreateGEP(nullptr, GEPBase, GEPOffset, "lftr.limit");
}
else {
// In any other case, convert both IVInit and IVCount to integers before
// BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc).
//
// Valid Cases: (1) both integers is most common; (2) both may be pointers
- // for simple memset-style loops; (3) IVInit is an integer and IVCount is a
- // pointer may occur when enable-iv-rewrite generates a canonical IV on top
- // of case #2.
+ // for simple memset-style loops.
+ //
+ // IVInit integer and IVCount pointer would only occur if a canonical IV
+ // were generated on top of case #2, which is not expected.
- const SCEV *IVLimit = 0;
+ const SCEV *IVLimit = nullptr;
// For unit stride, IVCount = Start + BECount with 2's complement overflow.
// For non-zero Start, compute IVCount here.
if (AR->getStart()->isZero())
SCEVExpander &Rewriter) {
assert(canExpandBackedgeTakenCount(L, SE) && "precondition");
- // LFTR can ignore IV overflow and truncate to the width of
- // BECount. This avoids materializing the add(zext(add)) expression.
- Type *CntTy = BackedgeTakenCount->getType();
-
+ // Initialize CmpIndVar and IVCount to their preincremented values.
+ Value *CmpIndVar = IndVar;
const SCEV *IVCount = BackedgeTakenCount;
// If the exiting block is the same as the backedge block, we prefer to
// compare against the post-incremented value, otherwise we must compare
// against the preincremented value.
- Value *CmpIndVar;
if (L->getExitingBlock() == L->getLoopLatch()) {
// Add one to the "backedge-taken" count to get the trip count.
- // If this addition may overflow, we have to be more pessimistic and
- // cast the induction variable before doing the add.
- const SCEV *N =
- SE->getAddExpr(IVCount, SE->getConstant(IVCount->getType(), 1));
- if (CntTy == IVCount->getType())
- IVCount = N;
- else {
- const SCEV *Zero = SE->getConstant(IVCount->getType(), 0);
- if ((isa<SCEVConstant>(N) && !N->isZero()) ||
- SE->isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
- // No overflow. Cast the sum.
- IVCount = SE->getTruncateOrZeroExtend(N, CntTy);
- } else {
- // Potential overflow. Cast before doing the add.
- IVCount = SE->getTruncateOrZeroExtend(IVCount, CntTy);
- IVCount = SE->getAddExpr(IVCount, SE->getConstant(CntTy, 1));
- }
- }
+ // This addition may overflow, which is valid as long as the comparison is
+ // truncated to BackedgeTakenCount->getType().
+ IVCount = SE->getAddExpr(BackedgeTakenCount,
+ SE->getConstant(BackedgeTakenCount->getType(), 1));
// The BackedgeTaken expression contains the number of times that the
// backedge branches to the loop header. This is one less than the
// number of times the loop executes, so use the incremented indvar.
CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock());
- } else {
- // We must use the preincremented value...
- IVCount = SE->getTruncateOrZeroExtend(IVCount, CntTy);
- CmpIndVar = IndVar;
}
- Type *IntPtrTy = TD ? TD->getIntPtrType(IndVar->getType()) :
- IntegerType::getInt64Ty(IndVar->getContext());
- Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE, IntPtrTy);
+ Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE);
assert(ExitCnt->getType()->isPointerTy() == IndVar->getType()->isPointerTy()
&& "genLoopLimit missed a cast");
<< " IVCount:\t" << *IVCount << "\n");
IRBuilder<> Builder(BI);
- if (SE->getTypeSizeInBits(CmpIndVar->getType())
- > SE->getTypeSizeInBits(ExitCnt->getType())) {
- CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(),
- "lftr.wideiv");
- }
+ // LFTR can ignore IV overflow and truncate to the width of
+ // BECount. This avoids materializing the add(zext(add)) expression.
+ unsigned CmpIndVarSize = SE->getTypeSizeInBits(CmpIndVar->getType());
+ unsigned ExitCntSize = SE->getTypeSizeInBits(ExitCnt->getType());
+ if (CmpIndVarSize > ExitCntSize) {
+ const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
+ const SCEV *ARStart = AR->getStart();
+ const SCEV *ARStep = AR->getStepRecurrence(*SE);
+ // For constant IVCount, avoid truncation.
+ if (isa<SCEVConstant>(ARStart) && isa<SCEVConstant>(IVCount)) {
+ const APInt &Start = cast<SCEVConstant>(ARStart)->getValue()->getValue();
+ APInt Count = cast<SCEVConstant>(IVCount)->getValue()->getValue();
+ // Note that the post-inc value of BackedgeTakenCount may have overflowed
+ // above such that IVCount is now zero.
+ if (IVCount != BackedgeTakenCount && Count == 0) {
+ Count = APInt::getMaxValue(Count.getBitWidth()).zext(CmpIndVarSize);
+ ++Count;
+ }
+ else
+ Count = Count.zext(CmpIndVarSize);
+ APInt NewLimit;
+ if (cast<SCEVConstant>(ARStep)->getValue()->isNegative())
+ NewLimit = Start - Count;
+ else
+ NewLimit = Start + Count;
+ ExitCnt = ConstantInt::get(CmpIndVar->getType(), NewLimit);
+
+ DEBUG(dbgs() << " Widen RHS:\t" << *ExitCnt << "\n");
+ } else {
+ CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(),
+ "lftr.wideiv");
+ }
+ }
Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond");
Value *OrigCond = BI->getCondition();
// It's tempting to use replaceAllUsesWith here to fully replace the old
// 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) {
- User *U = *UI;
- BasicBlock *UseBB = cast<Instruction>(U)->getParent();
- if (PHINode *P = dyn_cast<PHINode>(U)) {
+ for (Use &U : I->uses()) {
+ Instruction *User = cast<Instruction>(U.getUser());
+ BasicBlock *UseBB = User->getParent();
+ if (PHINode *P = dyn_cast<PHINode>(User)) {
unsigned i =
- PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
+ PHINode::getIncomingValueNumForOperand(U.getOperandNo());
UseBB = P->getIncomingBlock(i);
}
if (UseBB == Preheader || L->contains(UseBB)) {
//===----------------------------------------------------------------------===//
bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
+ if (skipOptnoneFunction(L))
+ return false;
+
// If LoopSimplify form is not available, stay out of trouble. Some notes:
// - LSR currently only supports LoopSimplify-form loops. Indvars'
// canonicalization can be a pessimization without LSR to "clean up"
if (!L->isLoopSimplifyForm())
return false;
- LI = &getAnalysis<LoopInfo>();
+ LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
SE = &getAnalysis<ScalarEvolution>();
- DT = &getAnalysis<DominatorTree>();
- TD = getAnalysisIfAvailable<DataLayout>();
- TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
+ DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
+ auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>();
+ TLI = TLIP ? &TLIP->getTLI() : nullptr;
+ auto *TTIP = getAnalysisIfAvailable<TargetTransformInfoWrapperPass>();
+ TTI = TTIP ? &TTIP->getTTI(*L->getHeader()->getParent()) : nullptr;
+ const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
DeadInsts.clear();
Changed = false;
const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
// Create a rewriter object which we'll use to transform the code with.
- SCEVExpander Rewriter(*SE, "indvars");
+ SCEVExpander Rewriter(*SE, DL, "indvars");
#ifndef NDEBUG
Rewriter.setDebugType(DEBUG_TYPE);
#endif
// If we have a trip count expression, rewrite the loop's exit condition
// using it. We can currently only handle loops with a single exit.
if (canExpandBackedgeTakenCount(L, SE) && needsLFTR(L, DT)) {
- PHINode *IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT, TD);
+ PHINode *IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT);
if (IndVar) {
// Check preconditions for proper SCEVExpander operation. SCEV does not
// express SCEVExpander's dependencies, such as LoopSimplify. Instead any
// pass that uses the SCEVExpander must do it. This does not work well for
- // loop passes because SCEVExpander makes assumptions about all loops, while
- // LoopPassManager only forces the current loop to be simplified.
+ // loop passes because SCEVExpander makes assumptions about all loops,
+ // while LoopPassManager only forces the current loop to be simplified.
//
// FIXME: SCEV expansion has no way to bail out, so the caller must
// explicitly check any assumptions made by SCEV. Brittle.