// which starts at zero and steps by one.
// 2. The canonical induction variable is guaranteed to be the first PHI node
// in the loop header block.
-// 3. Any pointer arithmetic recurrences are raised to use array subscripts.
+// 3. The canonical induction variable is guaranteed to be in a wide enough
+// type so that IV expressions need not be (directly) zero-extended or
+// sign-extended.
+// 4. Any pointer arithmetic recurrences are raised to use array subscripts.
//
// If the trip count of a loop is computable, this pass also makes the following
// changes:
#include "llvm/BasicBlock.h"
#include "llvm/Constants.h"
#include "llvm/Instructions.h"
+#include "llvm/LLVMContext.h"
#include "llvm/Type.h"
#include "llvm/Analysis/Dominators.h"
#include "llvm/Analysis/IVUsers.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/LoopPass.h"
#include "llvm/Support/CFG.h"
-#include "llvm/Support/Compiler.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/Support/CommandLine.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/STLExtras.h"
STATISTIC(NumLFTR , "Number of loop exit tests replaced");
namespace {
- class VISIBILITY_HIDDEN IndVarSimplify : public LoopPass {
+ class IndVarSimplify : public LoopPass {
IVUsers *IU;
LoopInfo *LI;
ScalarEvolution *SE;
+ DominatorTree *DT;
bool Changed;
public:
- static char ID; // Pass identification, replacement for typeid
- IndVarSimplify() : LoopPass(&ID) {}
-
- virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
-
- virtual void getAnalysisUsage(AnalysisUsage &AU) const {
- AU.addRequired<DominatorTree>();
- AU.addRequired<ScalarEvolution>();
- AU.addRequiredID(LCSSAID);
- AU.addRequiredID(LoopSimplifyID);
- AU.addRequired<LoopInfo>();
- AU.addRequired<IVUsers>();
- AU.addPreserved<ScalarEvolution>();
- AU.addPreservedID(LoopSimplifyID);
- AU.addPreserved<IVUsers>();
- AU.addPreservedID(LCSSAID);
- AU.setPreservesCFG();
- }
+ static char ID; // Pass identification, replacement for typeid
+ IndVarSimplify() : LoopPass(&ID) {}
+
+ virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.addRequired<DominatorTree>();
+ AU.addRequired<LoopInfo>();
+ AU.addRequired<ScalarEvolution>();
+ AU.addRequiredID(LoopSimplifyID);
+ AU.addRequiredID(LCSSAID);
+ AU.addRequired<IVUsers>();
+ AU.addPreserved<ScalarEvolution>();
+ AU.addPreservedID(LoopSimplifyID);
+ AU.addPreservedID(LCSSAID);
+ AU.addPreserved<IVUsers>();
+ AU.setPreservesCFG();
+ }
private:
void RewriteNonIntegerIVs(Loop *L);
- ICmpInst *LinearFunctionTestReplace(Loop *L, SCEVHandle BackedgeTakenCount,
+ ICmpInst *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
Value *IndVar,
BasicBlock *ExitingBlock,
BranchInst *BI,
SCEVExpander &Rewriter);
- void RewriteLoopExitValues(Loop *L, const SCEV *BackedgeTakenCount);
+ void RewriteLoopExitValues(Loop *L, const SCEV *BackedgeTakenCount,
+ SCEVExpander &Rewriter);
void RewriteIVExpressions(Loop *L, const Type *LargestType,
SCEVExpander &Rewriter);
- void SinkUnusedInvariants(Loop *L, SCEVExpander &Rewriter);
-
- void FixUsesBeforeDefs(Loop *L, SCEVExpander &Rewriter);
+ void SinkUnusedInvariants(Loop *L);
void HandleFloatingPointIV(Loop *L, PHINode *PH);
};
/// SCEV analysis can determine a loop-invariant trip count of the loop, which
/// is actually a much broader range than just linear tests.
ICmpInst *IndVarSimplify::LinearFunctionTestReplace(Loop *L,
- SCEVHandle BackedgeTakenCount,
+ const SCEV *BackedgeTakenCount,
Value *IndVar,
BasicBlock *ExitingBlock,
BranchInst *BI,
// against the preincremented value, otherwise we prefer to compare against
// the post-incremented value.
Value *CmpIndVar;
- SCEVHandle RHS = BackedgeTakenCount;
+ const SCEV *RHS = BackedgeTakenCount;
if (ExitingBlock == 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.
- SCEVHandle Zero = SE->getIntegerSCEV(0, BackedgeTakenCount->getType());
- SCEVHandle N =
+ const SCEV *Zero = SE->getIntegerSCEV(0, BackedgeTakenCount->getType());
+ const SCEV *N =
SE->getAddExpr(BackedgeTakenCount,
SE->getIntegerSCEV(1, BackedgeTakenCount->getType()));
if ((isa<SCEVConstant>(N) && !N->isZero()) ||
CmpIndVar = IndVar;
}
- // Expand the code for the iteration count into the preheader of the loop.
- BasicBlock *Preheader = L->getLoopPreheader();
- Value *ExitCnt = Rewriter.expandCodeFor(RHS, CmpIndVar->getType(),
- Preheader->getTerminator());
+ // Expand the code for the iteration count.
+ assert(RHS->isLoopInvariant(L) &&
+ "Computed iteration count is not loop invariant!");
+ Value *ExitCnt = Rewriter.expandCodeFor(RHS, IndVar->getType(), BI);
// Insert a new icmp_ne or icmp_eq instruction before the branch.
ICmpInst::Predicate Opcode;
else
Opcode = ICmpInst::ICMP_EQ;
- DOUT << "INDVARS: Rewriting loop exit condition to:\n"
- << " LHS:" << *CmpIndVar // includes a newline
- << " op:\t"
- << (Opcode == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
- << " RHS:\t" << *RHS << "\n";
+ DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
+ << " LHS:" << *CmpIndVar << '\n'
+ << " op:\t"
+ << (Opcode == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
+ << " RHS:\t" << *RHS << "\n");
- ICmpInst *Cond = new ICmpInst(Opcode, CmpIndVar, ExitCnt, "exitcond", BI);
+ ICmpInst *Cond = new ICmpInst(BI, Opcode, CmpIndVar, ExitCnt, "exitcond");
Instruction *OrigCond = cast<Instruction>(BI->getCondition());
- OrigCond->replaceAllUsesWith(Cond);
+ // It's tempting to use replaceAllUsesWith here to fully replace the old
+ // comparison, but that's not immediately safe, since users of the old
+ // comparison may not be dominated by the new comparison. Instead, just
+ // update the branch to use the new comparison; in the common case this
+ // will make old comparison dead.
+ BI->setCondition(Cond);
RecursivelyDeleteTriviallyDeadInstructions(OrigCond);
++NumLFTR;
/// able to brute-force evaluate arbitrary instructions as long as they have
/// constant operands at the beginning of the loop.
void IndVarSimplify::RewriteLoopExitValues(Loop *L,
- const SCEV *BackedgeTakenCount) {
+ const SCEV *BackedgeTakenCount,
+ SCEVExpander &Rewriter) {
// Verify the input to the pass in already in LCSSA form.
assert(L->isLCSSAForm());
- BasicBlock *Preheader = L->getLoopPreheader();
-
- // Scan all of the instructions in the loop, looking at those that have
- // extra-loop users and which are recurrences.
- SCEVExpander Rewriter(*SE);
-
- // We insert the code into the preheader of the loop if the loop contains
- // multiple exit blocks, or in the exit block if there is exactly one.
- BasicBlock *BlockToInsertInto;
SmallVector<BasicBlock*, 8> ExitBlocks;
L->getUniqueExitBlocks(ExitBlocks);
- if (ExitBlocks.size() == 1)
- BlockToInsertInto = ExitBlocks[0];
- else
- BlockToInsertInto = Preheader;
- BasicBlock::iterator InsertPt = BlockToInsertInto->getFirstNonPHI();
-
- std::map<Instruction*, Value*> ExitValues;
// Find all values that are computed inside the loop, but used outside of it.
// Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan
// Iterate over all of the PHI nodes.
BasicBlock::iterator BBI = ExitBB->begin();
while ((PN = dyn_cast<PHINode>(BBI++))) {
-
+ if (PN->use_empty())
+ continue; // dead use, don't replace it
// Iterate over all of the values in all the PHI nodes.
for (unsigned i = 0; i != NumPreds; ++i) {
// If the value being merged in is not integer or is not defined
// Check that InVal is defined in the loop.
Instruction *Inst = cast<Instruction>(InVal);
- if (!L->contains(Inst->getParent()))
+ if (!L->contains(Inst))
continue;
// Okay, this instruction has a user outside of the current loop
// and varies predictably *inside* the loop. Evaluate the value it
// contains when the loop exits, if possible.
- SCEVHandle SH = SE->getSCEV(Inst);
- SCEVHandle ExitValue = SE->getSCEVAtScope(SH, L->getParentLoop());
- if (isa<SCEVCouldNotCompute>(ExitValue) ||
- !ExitValue->isLoopInvariant(L))
+ const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
+ if (!ExitValue->isLoopInvariant(L))
continue;
Changed = true;
++NumReplaced;
- // See if we already computed the exit value for the instruction, if so,
- // just reuse it.
- Value *&ExitVal = ExitValues[Inst];
- if (!ExitVal)
- ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), InsertPt);
+ Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
- DOUT << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal
- << " LoopVal = " << *Inst << "\n";
+ DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
+ << " LoopVal = " << *Inst << "\n");
PN->setIncomingValue(i, ExitVal);
// If this instruction is dead now, delete it.
RecursivelyDeleteTriviallyDeadInstructions(Inst);
- // See if this is a single-entry LCSSA PHI node. If so, we can (and
- // have to) remove
- // the PHI entirely. This is safe, because the NewVal won't be variant
- // in the loop, so we don't need an LCSSA phi node anymore.
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.
PN->replaceAllUsesWith(ExitVal);
- Rewriter.clear();
RecursivelyDeleteTriviallyDeadInstructions(PN);
- break;
}
}
+ 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.
+ PHINode *NewPN = cast<PHINode>(PN->clone());
+ NewPN->takeName(PN);
+ NewPN->insertBefore(PN);
+ PN->replaceAllUsesWith(NewPN);
+ PN->eraseFromParent();
+ }
}
}
}
// may not have been able to compute a trip count. Now that we've done some
// re-writing, the trip count may be computable.
if (Changed)
- SE->forgetLoopBackedgeTakenCount(L);
+ SE->forgetLoop(L);
}
bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
IU = &getAnalysis<IVUsers>();
LI = &getAnalysis<LoopInfo>();
SE = &getAnalysis<ScalarEvolution>();
+ DT = &getAnalysis<DominatorTree>();
Changed = false;
// If there are any floating-point recurrences, attempt to
// transform them to use integer recurrences.
RewriteNonIntegerIVs(L);
- BasicBlock *Header = L->getHeader();
BasicBlock *ExitingBlock = L->getExitingBlock(); // may be null
- SCEVHandle BackedgeTakenCount = SE->getBackedgeTakenCount(L);
+ const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
+
+ // Create a rewriter object which we'll use to transform the code with.
+ SCEVExpander Rewriter(*SE);
// Check to see if this loop has a computable loop-invariant execution count.
// If so, this means that we can compute the final value of any expressions
// the current expressions.
//
if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
- RewriteLoopExitValues(L, BackedgeTakenCount);
+ RewriteLoopExitValues(L, BackedgeTakenCount, Rewriter);
// Compute the type of the largest recurrence expression, and decide whether
// a canonical induction variable should be inserted.
NeedCannIV = true;
}
for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
- SCEVHandle Stride = IU->StrideOrder[i];
+ const SCEV *Stride = IU->StrideOrder[i];
const Type *Ty = SE->getEffectiveSCEVType(Stride->getType());
if (!LargestType ||
SE->getTypeSizeInBits(Ty) >
SE->getTypeSizeInBits(LargestType))
LargestType = Ty;
- std::map<SCEVHandle, IVUsersOfOneStride *>::iterator SI =
+ std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
IU->IVUsesByStride.find(IU->StrideOrder[i]);
assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
NeedCannIV = true;
}
- // Create a rewriter object which we'll use to transform the code with.
- SCEVExpander Rewriter(*SE);
-
// Now that we know the largest of of the induction variable expressions
// in this loop, insert a canonical induction variable of the largest size.
Value *IndVar = 0;
if (NeedCannIV) {
- IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L,LargestType);
+ // Check to see if the loop already has a canonical-looking induction
+ // variable. If one is present and it's wider than the planned canonical
+ // induction variable, temporarily remove it, so that the Rewriter
+ // doesn't attempt to reuse it.
+ PHINode *OldCannIV = L->getCanonicalInductionVariable();
+ if (OldCannIV) {
+ if (SE->getTypeSizeInBits(OldCannIV->getType()) >
+ SE->getTypeSizeInBits(LargestType))
+ OldCannIV->removeFromParent();
+ else
+ OldCannIV = 0;
+ }
+
+ IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L, LargestType);
+
++NumInserted;
Changed = true;
- DOUT << "INDVARS: New CanIV: " << *IndVar;
+ DEBUG(dbgs() << "INDVARS: New CanIV: " << *IndVar << '\n');
+
+ // Now that the official induction variable is established, reinsert
+ // the old canonical-looking variable after it so that the IR remains
+ // consistent. It will be deleted as part of the dead-PHI deletion at
+ // the end of the pass.
+ if (OldCannIV)
+ OldCannIV->insertAfter(cast<Instruction>(IndVar));
}
// If we have a trip count expression, rewrite the loop's exit condition
ExitingBlock, BI, Rewriter);
}
- Rewriter.setInsertionPoint(Header->getFirstNonPHI());
-
- // Rewrite IV-derived expressions.
+ // Rewrite IV-derived expressions. Clears the rewriter cache.
RewriteIVExpressions(L, LargestType, Rewriter);
+ // The Rewriter may not be used from this point on.
+
// Loop-invariant instructions in the preheader that aren't used in the
// loop may be sunk below the loop to reduce register pressure.
- SinkUnusedInvariants(L, Rewriter);
+ SinkUnusedInvariants(L);
- // Reorder instructions to avoid use-before-def conditions.
- FixUsesBeforeDefs(L, Rewriter);
-
- Rewriter.clear();
// For completeness, inform IVUsers of the IV use in the newly-created
// loop exit test instruction.
if (NewICmp)
// the need for the code evaluation methods to insert induction variables
// of different sizes.
for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
- SCEVHandle Stride = IU->StrideOrder[i];
+ const SCEV *Stride = IU->StrideOrder[i];
- std::map<SCEVHandle, IVUsersOfOneStride *>::iterator SI =
+ std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI =
IU->IVUsesByStride.find(IU->StrideOrder[i]);
assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
ilist<IVStrideUse> &List = SI->second->Users;
for (ilist<IVStrideUse>::iterator UI = List.begin(),
E = List.end(); UI != E; ++UI) {
- SCEVHandle Offset = UI->getOffset();
Value *Op = UI->getOperandValToReplace();
+ const Type *UseTy = Op->getType();
Instruction *User = UI->getUser();
- bool isSigned = UI->isSigned();
// Compute the final addrec to expand into code.
- SCEVHandle AR = IU->getReplacementExpr(*UI);
+ const SCEV *AR = IU->getReplacementExpr(*UI);
// FIXME: It is an extremely bad idea to indvar substitute anything more
// complex than affine induction variables. Doing so will put expensive
// currently can only reduce affine polynomials. For now just disable
// indvar subst on anything more complex than an affine addrec, unless
// it can be expanded to a trivial value.
- if (!Stride->isLoopInvariant(L) &&
- !isa<SCEVConstant>(AR) &&
- L->contains(User->getParent()))
+ if (!AR->isLoopInvariant(L) && !Stride->isLoopInvariant(L))
continue;
- Value *NewVal = 0;
- if (AR->isLoopInvariant(L)) {
- BasicBlock::iterator I = Rewriter.getInsertionPoint();
- // Expand loop-invariant values in the loop preheader. They will
- // be sunk to the exit block later, if possible.
- NewVal =
- Rewriter.expandCodeFor(AR, LargestType,
- L->getLoopPreheader()->getTerminator());
- Rewriter.setInsertionPoint(I);
- ++NumReplaced;
- } else {
- const Type *IVTy = Offset->getType();
- const Type *UseTy = Op->getType();
-
- // Promote the Offset and Stride up to the canonical induction
- // variable's bit width.
- SCEVHandle PromotedOffset = Offset;
- SCEVHandle PromotedStride = Stride;
- if (SE->getTypeSizeInBits(IVTy) != SE->getTypeSizeInBits(LargestType)) {
- // It doesn't matter for correctness whether zero or sign extension
- // is used here, since the value is truncated away below, but if the
- // value is signed, sign extension is more likely to be folded.
- if (isSigned) {
- PromotedOffset = SE->getSignExtendExpr(PromotedOffset, LargestType);
- PromotedStride = SE->getSignExtendExpr(PromotedStride, LargestType);
- } else {
- PromotedOffset = SE->getZeroExtendExpr(PromotedOffset, LargestType);
- // If the stride is obviously negative, use sign extension to
- // produce things like x-1 instead of x+255.
- if (isa<SCEVConstant>(PromotedStride) &&
- cast<SCEVConstant>(PromotedStride)
- ->getValue()->getValue().isNegative())
- PromotedStride = SE->getSignExtendExpr(PromotedStride,
- LargestType);
+ // Determine the insertion point for this user. By default, insert
+ // immediately before the user. The SCEVExpander class will automatically
+ // hoist loop invariants out of the loop. For PHI nodes, there may be
+ // multiple uses, so compute the nearest common dominator for the
+ // incoming blocks.
+ Instruction *InsertPt = User;
+ if (PHINode *PHI = dyn_cast<PHINode>(InsertPt))
+ for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i)
+ if (PHI->getIncomingValue(i) == Op) {
+ if (InsertPt == User)
+ InsertPt = PHI->getIncomingBlock(i)->getTerminator();
else
- PromotedStride = SE->getZeroExtendExpr(PromotedStride,
- LargestType);
+ InsertPt =
+ DT->findNearestCommonDominator(InsertPt->getParent(),
+ PHI->getIncomingBlock(i))
+ ->getTerminator();
}
- }
- // Create the SCEV representing the offset from the canonical
- // induction variable, still in the canonical induction variable's
- // type, so that all expanded arithmetic is done in the same type.
- SCEVHandle NewAR = SE->getAddRecExpr(SE->getIntegerSCEV(0, LargestType),
- PromotedStride, L);
- // Add the PromotedOffset as a separate step, because it may not be
- // loop-invariant.
- NewAR = SE->getAddExpr(NewAR, PromotedOffset);
-
- // Expand the addrec into instructions.
- Value *V = Rewriter.expandCodeFor(NewAR);
-
- // Insert an explicit cast if necessary to truncate the value
- // down to the original stride type. This is done outside of
- // SCEVExpander because in SCEV expressions, a truncate of an
- // addrec is always folded.
- if (LargestType != IVTy) {
- if (SE->getTypeSizeInBits(IVTy) != SE->getTypeSizeInBits(LargestType))
- NewAR = SE->getTruncateExpr(NewAR, IVTy);
- if (Rewriter.isInsertedExpression(NewAR))
- V = Rewriter.expandCodeFor(NewAR);
- else {
- V = Rewriter.InsertCastOfTo(CastInst::getCastOpcode(V, false,
- IVTy, false),
- V, IVTy);
- assert(!isa<SExtInst>(V) && !isa<ZExtInst>(V) &&
- "LargestType wasn't actually the largest type!");
- // Force the rewriter to use this trunc whenever this addrec
- // appears so that it doesn't insert new phi nodes or
- // arithmetic in a different type.
- Rewriter.addInsertedValue(V, NewAR);
- }
- }
-
- DOUT << "INDVARS: Made offset-and-trunc IV for offset "
- << *IVTy << " " << *Offset << ": ";
- DEBUG(WriteAsOperand(*DOUT, V, false));
- DOUT << "\n";
-
- // Now expand it into actual Instructions and patch it into place.
- NewVal = Rewriter.expandCodeFor(AR, UseTy);
- }
+ // Now expand it into actual Instructions and patch it into place.
+ Value *NewVal = Rewriter.expandCodeFor(AR, UseTy, InsertPt);
// Patch the new value into place.
if (Op->hasName())
NewVal->takeName(Op);
User->replaceUsesOfWith(Op, NewVal);
UI->setOperandValToReplace(NewVal);
- DOUT << "INDVARS: Rewrote IV '" << *AR << "' " << *Op
- << " into = " << *NewVal << "\n";
+ DEBUG(dbgs() << "INDVARS: Rewrote IV '" << *AR << "' " << *Op << '\n'
+ << " into = " << *NewVal << "\n");
++NumRemoved;
Changed = true;
}
}
+ // Clear the rewriter cache, because values that are in the rewriter's cache
+ // can be deleted in the loop below, causing the AssertingVH in the cache to
+ // trigger.
+ Rewriter.clear();
// Now that we're done iterating through lists, clean up any instructions
// which are now dead.
while (!DeadInsts.empty()) {
/// 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, SCEVExpander &Rewriter) {
+void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
BasicBlock *ExitBlock = L->getExitBlock();
if (!ExitBlock) return;
- Instruction *NonPHI = ExitBlock->getFirstNonPHI();
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. Only
- // consider those new instructions.
- if (!Rewriter.isInsertedInstruction(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;
--I;
else
Done = true;
- ToMove->moveBefore(NonPHI);
+ ToMove->moveBefore(InsertPt);
if (Done)
break;
+ InsertPt = ToMove;
}
}
-/// Re-schedule the inserted instructions to put defs before uses. This
-/// fixes problems that arrise when SCEV expressions contain loop-variant
-/// values unrelated to the induction variable which are defined inside the
-/// loop. FIXME: It would be better to insert instructions in the right
-/// place so that this step isn't needed.
-void IndVarSimplify::FixUsesBeforeDefs(Loop *L, SCEVExpander &Rewriter) {
- // Visit all the blocks in the loop in pre-order dom-tree dfs order.
- DominatorTree *DT = &getAnalysis<DominatorTree>();
- std::map<Instruction *, unsigned> NumPredsLeft;
- SmallVector<DomTreeNode *, 16> Worklist;
- Worklist.push_back(DT->getNode(L->getHeader()));
- do {
- DomTreeNode *Node = Worklist.pop_back_val();
- for (DomTreeNode::iterator I = Node->begin(), E = Node->end(); I != E; ++I)
- if (L->contains((*I)->getBlock()))
- Worklist.push_back(*I);
- BasicBlock *BB = Node->getBlock();
- // Visit all the instructions in the block top down.
- for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
- // Count the number of operands that aren't properly dominating.
- unsigned NumPreds = 0;
- if (Rewriter.isInsertedInstruction(I) && !isa<PHINode>(I))
- for (User::op_iterator OI = I->op_begin(), OE = I->op_end();
- OI != OE; ++OI)
- if (Instruction *Inst = dyn_cast<Instruction>(OI))
- if (L->contains(Inst->getParent()) && !NumPredsLeft.count(Inst))
- ++NumPreds;
- NumPredsLeft[I] = NumPreds;
- // Notify uses of the position of this instruction, and move the
- // users (and their dependents, recursively) into place after this
- // instruction if it is their last outstanding operand.
- for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
- UI != UE; ++UI) {
- Instruction *Inst = cast<Instruction>(UI);
- std::map<Instruction *, unsigned>::iterator Z = NumPredsLeft.find(Inst);
- if (Z != NumPredsLeft.end() && Z->second != 0 && --Z->second == 0) {
- SmallVector<Instruction *, 4> UseWorkList;
- UseWorkList.push_back(Inst);
- BasicBlock::iterator InsertPt = I;
- if (InvokeInst *II = dyn_cast<InvokeInst>(InsertPt))
- InsertPt = II->getNormalDest()->begin();
- else
- ++InsertPt;
- while (isa<PHINode>(InsertPt)) ++InsertPt;
- do {
- Instruction *Use = UseWorkList.pop_back_val();
- Use->moveBefore(InsertPt);
- NumPredsLeft.erase(Use);
- for (Value::use_iterator IUI = Use->use_begin(),
- IUE = Use->use_end(); IUI != IUE; ++IUI) {
- Instruction *IUIInst = cast<Instruction>(IUI);
- if (L->contains(IUIInst->getParent()) &&
- Rewriter.isInsertedInstruction(IUIInst) &&
- !isa<PHINode>(IUIInst))
- UseWorkList.push_back(IUIInst);
- }
- } while (!UseWorkList.empty());
- }
- }
- }
- } while (!Worklist.empty());
-}
-
/// Return true if it is OK to use SIToFPInst for an inducation variable
/// with given inital and exit values.
static bool useSIToFPInst(ConstantFP &InitV, ConstantFP &ExitV,
// Check incoming value.
ConstantFP *InitValue = dyn_cast<ConstantFP>(PH->getIncomingValue(IncomingEdge));
if (!InitValue) return;
- uint64_t newInitValue = Type::Int32Ty->getPrimitiveSizeInBits();
+ uint64_t newInitValue =
+ Type::getInt32Ty(PH->getContext())->getPrimitiveSizeInBits();
if (!convertToInt(InitValue->getValueAPF(), &newInitValue))
return;
BinaryOperator *Incr =
dyn_cast<BinaryOperator>(PH->getIncomingValue(BackEdge));
if (!Incr) return;
- if (Incr->getOpcode() != Instruction::Add) return;
+ if (Incr->getOpcode() != Instruction::FAdd) return;
ConstantFP *IncrValue = NULL;
unsigned IncrVIndex = 1;
if (Incr->getOperand(1) == PH)
IncrVIndex = 0;
IncrValue = dyn_cast<ConstantFP>(Incr->getOperand(IncrVIndex));
if (!IncrValue) return;
- uint64_t newIncrValue = Type::Int32Ty->getPrimitiveSizeInBits();
+ uint64_t newIncrValue =
+ Type::getInt32Ty(PH->getContext())->getPrimitiveSizeInBits();
if (!convertToInt(IncrValue->getValueAPF(), &newIncrValue))
return;
EVIndex = 0;
EV = dyn_cast<ConstantFP>(EC->getOperand(EVIndex));
if (!EV) return;
- uint64_t intEV = Type::Int32Ty->getPrimitiveSizeInBits();
+ uint64_t intEV = Type::getInt32Ty(PH->getContext())->getPrimitiveSizeInBits();
if (!convertToInt(EV->getValueAPF(), &intEV))
return;
if (NewPred == CmpInst::BAD_ICMP_PREDICATE) return;
// Insert new integer induction variable.
- PHINode *NewPHI = PHINode::Create(Type::Int32Ty,
+ PHINode *NewPHI = PHINode::Create(Type::getInt32Ty(PH->getContext()),
PH->getName()+".int", PH);
- NewPHI->addIncoming(ConstantInt::get(Type::Int32Ty, newInitValue),
+ NewPHI->addIncoming(ConstantInt::get(Type::getInt32Ty(PH->getContext()),
+ newInitValue),
PH->getIncomingBlock(IncomingEdge));
Value *NewAdd = BinaryOperator::CreateAdd(NewPHI,
- ConstantInt::get(Type::Int32Ty,
+ ConstantInt::get(Type::getInt32Ty(PH->getContext()),
newIncrValue),
Incr->getName()+".int", Incr);
NewPHI->addIncoming(NewAdd, PH->getIncomingBlock(BackEdge));
// The back edge is edge 1 of newPHI, whatever it may have been in the
// original PHI.
- ConstantInt *NewEV = ConstantInt::get(Type::Int32Ty, intEV);
+ ConstantInt *NewEV = ConstantInt::get(Type::getInt32Ty(PH->getContext()),
+ intEV);
Value *LHS = (EVIndex == 1 ? NewPHI->getIncomingValue(1) : NewEV);
Value *RHS = (EVIndex == 1 ? NewEV : NewPHI->getIncomingValue(1));
- ICmpInst *NewEC = new ICmpInst(NewPred, LHS, RHS, EC->getNameStart(),
- EC->getParent()->getTerminator());
+ ICmpInst *NewEC = new ICmpInst(EC->getParent()->getTerminator(),
+ NewPred, LHS, RHS, EC->getName());
// In the following deltions, PH may become dead and may be deleted.
// Use a WeakVH to observe whether this happens.