// computations derived from them) into simpler forms suitable for subsequent
// analysis and transformation.
//
-// This transformation makes the following changes to each loop with an
-// identifiable induction variable:
-// 1. All loops are transformed to have a SINGLE canonical induction variable
-// 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.
-//
// If the trip count of a loop is computable, this pass also makes the following
// changes:
// 1. The exit condition for the loop is canonicalized to compare the
// purpose of the loop is to compute the exit value of some derived
// expression, this transformation will make the loop dead.
//
-// This transformation should be followed by strength reduction after all of the
-// desired loop transformations have been performed. Additionally, on targets
-// where it is profitable, the loop could be transformed to count down to zero
-// (the "do loop" optimization).
-//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "indvars"
#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/IVUsers.h"
#include "llvm/Analysis/ScalarEvolutionExpander.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/GetElementPtrTypeIterator.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/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/SetVector.h"
#include "llvm/ADT/Statistic.h"
-#include "llvm/ADT/STLExtras.h"
using namespace llvm;
-STATISTIC(NumRemoved , "Number of aux indvars removed");
-STATISTIC(NumInserted, "Number of canonical indvars added");
-STATISTIC(NumReplaced, "Number of exit values replaced");
-STATISTIC(NumLFTR , "Number of loop exit tests replaced");
+STATISTIC(NumWidened , "Number of indvars widened");
+STATISTIC(NumReplaced , "Number of exit values replaced");
+STATISTIC(NumLFTR , "Number of loop exit tests replaced");
+STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated");
+STATISTIC(NumElimIV , "Number of congruent IVs eliminated");
+
+// Trip count verification can be enabled by default under NDEBUG if we
+// implement a strong expression equivalence checker in SCEV. Until then, we
+// use the verify-indvars flag, which may assert in some cases.
+static cl::opt<bool> VerifyIndvars(
+ "verify-indvars", cl::Hidden,
+ cl::desc("Verify the ScalarEvolution result after running indvars"));
namespace {
- class VISIBILITY_HIDDEN IndVarSimplify : public LoopPass {
- IVUsers *IU;
+ class IndVarSimplify : public LoopPass {
LoopInfo *LI;
ScalarEvolution *SE;
+ DominatorTree *DT;
+ DataLayout *TD;
+ TargetLibraryInfo *TLI;
+
+ SmallVector<WeakVH, 16> DeadInsts;
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), LI(0), SE(0), DT(0), TD(0),
+ Changed(false) {
+ initializeIndVarSimplifyPass(*PassRegistry::getPassRegistry());
+ }
+
+ 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.addPreserved<ScalarEvolution>();
+ AU.addPreservedID(LoopSimplifyID);
+ AU.addPreservedID(LCSSAID);
+ AU.setPreservesCFG();
+ }
private:
+ virtual void releaseMemory() {
+ DeadInsts.clear();
+ }
- void RewriteNonIntegerIVs(Loop *L);
+ bool isValidRewrite(Value *FromVal, Value *ToVal);
- ICmpInst *LinearFunctionTestReplace(Loop *L, SCEVHandle BackedgeTakenCount,
- Value *IndVar,
- BasicBlock *ExitingBlock,
- BranchInst *BI,
- SCEVExpander &Rewriter);
- void RewriteLoopExitValues(Loop *L, const SCEV *BackedgeTakenCount);
+ void HandleFloatingPointIV(Loop *L, PHINode *PH);
+ void RewriteNonIntegerIVs(Loop *L);
- void RewriteIVExpressions(Loop *L, const Type *LargestType,
- SCEVExpander &Rewriter);
+ void SimplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LPPassManager &LPM);
- void SinkUnusedInvariants(Loop *L, SCEVExpander &Rewriter);
+ void RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter);
- void FixUsesBeforeDefs(Loop *L, SCEVExpander &Rewriter);
+ Value *LinearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount,
+ PHINode *IndVar, SCEVExpander &Rewriter);
- void HandleFloatingPointIV(Loop *L, PHINode *PH);
+ void SinkUnusedInvariants(Loop *L);
};
}
char IndVarSimplify::ID = 0;
-static RegisterPass<IndVarSimplify>
-X("indvars", "Canonicalize Induction Variables");
+INITIALIZE_PASS_BEGIN(IndVarSimplify, "indvars",
+ "Induction Variable Simplification", false, false)
+INITIALIZE_PASS_DEPENDENCY(DominatorTree)
+INITIALIZE_PASS_DEPENDENCY(LoopInfo)
+INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
+INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
+INITIALIZE_PASS_DEPENDENCY(LCSSA)
+INITIALIZE_PASS_END(IndVarSimplify, "indvars",
+ "Induction Variable Simplification", false, false)
Pass *llvm::createIndVarSimplifyPass() {
return new IndVarSimplify();
}
-/// LinearFunctionTestReplace - This method rewrites the exit condition of the
-/// loop to be a canonical != comparison against the incremented loop induction
-/// variable. This pass is able to rewrite the exit tests of any loop where the
-/// 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,
- Value *IndVar,
- BasicBlock *ExitingBlock,
- BranchInst *BI,
- SCEVExpander &Rewriter) {
- // If the exiting block is not the same as the backedge block, we must compare
- // against the preincremented value, otherwise we prefer to compare against
- // the post-incremented value.
- Value *CmpIndVar;
- SCEVHandle 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 =
- SE->getAddExpr(BackedgeTakenCount,
- SE->getIntegerSCEV(1, BackedgeTakenCount->getType()));
- if ((isa<SCEVConstant>(N) && !N->isZero()) ||
- SE->isLoopGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
- // No overflow. Cast the sum.
- RHS = SE->getTruncateOrZeroExtend(N, IndVar->getType());
- } else {
- // Potential overflow. Cast before doing the add.
- RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
- IndVar->getType());
- RHS = SE->getAddExpr(RHS,
- SE->getIntegerSCEV(1, IndVar->getType()));
+/// isValidRewrite - Return true if the SCEV expansion generated by the
+/// rewriter can replace the original value. SCEV guarantees that it
+/// produces the same value, but the way it is produced may be illegal IR.
+/// Ideally, this function will only be called for verification.
+bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) {
+ // If an SCEV expression subsumed multiple pointers, its expansion could
+ // reassociate the GEP changing the base pointer. This is illegal because the
+ // final address produced by a GEP chain must be inbounds relative to its
+ // underlying object. Otherwise basic alias analysis, among other things,
+ // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid
+ // producing an expression involving multiple pointers. Until then, we must
+ // bail out here.
+ //
+ // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject
+ // because it understands lcssa phis while SCEV does not.
+ Value *FromPtr = FromVal;
+ Value *ToPtr = ToVal;
+ if (GEPOperator *GEP = dyn_cast<GEPOperator>(FromVal)) {
+ FromPtr = GEP->getPointerOperand();
+ }
+ if (GEPOperator *GEP = dyn_cast<GEPOperator>(ToVal)) {
+ ToPtr = GEP->getPointerOperand();
+ }
+ if (FromPtr != FromVal || ToPtr != ToVal) {
+ // Quickly check the common case
+ if (FromPtr == ToPtr)
+ return true;
+
+ // SCEV may have rewritten an expression that produces the GEP's pointer
+ // operand. That's ok as long as the pointer operand has the same base
+ // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the
+ // base of a recurrence. This handles the case in which SCEV expansion
+ // converts a pointer type recurrence into a nonrecurrent pointer base
+ // indexed by an integer recurrence.
+
+ // If the GEP base pointer is a vector of pointers, abort.
+ if (!FromPtr->getType()->isPointerTy() || !ToPtr->getType()->isPointerTy())
+ return false;
+
+ const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr));
+ const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr));
+ if (FromBase == ToBase)
+ return true;
+
+ DEBUG(dbgs() << "INDVARS: GEP rewrite bail out "
+ << *FromBase << " != " << *ToBase << "\n");
+
+ return false;
+ }
+ return true;
+}
+
+/// Determine the insertion point for this user. By default, insert immediately
+/// before the user. SCEVExpander or LICM will 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.
+static Instruction *getInsertPointForUses(Instruction *User, Value *Def,
+ DominatorTree *DT) {
+ PHINode *PHI = dyn_cast<PHINode>(User);
+ if (!PHI)
+ return User;
+
+ Instruction *InsertPt = 0;
+ for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) {
+ if (PHI->getIncomingValue(i) != Def)
+ continue;
+
+ BasicBlock *InsertBB = PHI->getIncomingBlock(i);
+ if (!InsertPt) {
+ InsertPt = InsertBB->getTerminator();
+ continue;
}
+ InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB);
+ InsertPt = InsertBB->getTerminator();
+ }
+ assert(InsertPt && "Missing phi operand");
+ assert((!isa<Instruction>(Def) ||
+ DT->dominates(cast<Instruction>(Def), InsertPt)) &&
+ "def does not dominate all uses");
+ return InsertPt;
+}
+
+//===----------------------------------------------------------------------===//
+// RewriteNonIntegerIVs and helpers. Prefer integer IVs.
+//===----------------------------------------------------------------------===//
+
+/// ConvertToSInt - Convert APF to an integer, if possible.
+static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) {
+ bool isExact = false;
+ // See if we can convert this to an int64_t
+ uint64_t UIntVal;
+ if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero,
+ &isExact) != APFloat::opOK || !isExact)
+ return false;
+ IntVal = UIntVal;
+ return true;
+}
+
+/// HandleFloatingPointIV - If the loop has floating induction variable
+/// then insert corresponding integer induction variable if possible.
+/// For example,
+/// for(double i = 0; i < 10000; ++i)
+/// bar(i)
+/// is converted into
+/// for(int i = 0; i < 10000; ++i)
+/// bar((double)i);
+///
+void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PN) {
+ unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
+ unsigned BackEdge = IncomingEdge^1;
+
+ // Check incoming value.
+ ConstantFP *InitValueVal =
+ dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge));
+
+ int64_t InitValue;
+ if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue))
+ return;
+
+ // Check IV increment. Reject this PN if increment operation is not
+ // 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 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 ||
+ !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();
+ Instruction *U1 = cast<Instruction>(*IncrUse++);
+ if (IncrUse == Incr->use_end()) return;
+ Instruction *U2 = cast<Instruction>(*IncrUse++);
+ if (IncrUse != Incr->use_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()))
+ return;
+
+ BranchInst *TheBr = cast<BranchInst>(Compare->use_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.
+ // The branch block must be in the loop and one of the successors must be out
+ // of the loop.
+ assert(TheBr->isConditional() && "Can't use fcmp if not conditional");
+ if (!L->contains(TheBr->getParent()) ||
+ (L->contains(TheBr->getSuccessor(0)) &&
+ L->contains(TheBr->getSuccessor(1))))
+ return;
+
+
+ // If it isn't a comparison with an integer-as-fp (the exit value), we can't
+ // transform it.
+ ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1));
+ int64_t ExitValue;
+ if (ExitValueVal == 0 ||
+ !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue))
+ return;
+
+ // Find new predicate for integer comparison.
+ CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
+ switch (Compare->getPredicate()) {
+ default: return; // Unknown comparison.
+ case CmpInst::FCMP_OEQ:
+ case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break;
+ case CmpInst::FCMP_ONE:
+ case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break;
+ case CmpInst::FCMP_OGT:
+ case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break;
+ case CmpInst::FCMP_OGE:
+ case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break;
+ case CmpInst::FCMP_OLT:
+ case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break;
+ case CmpInst::FCMP_OLE:
+ case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break;
+ }
+
+ // We convert the floating point induction variable to a signed i32 value if
+ // we can. This is only safe if the comparison will not overflow in a way
+ // that won't be trapped by the integer equivalent operations. Check for this
+ // now.
+ // TODO: We could use i64 if it is native and the range requires it.
+
+ // The start/stride/exit values must all fit in signed i32.
+ if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue))
+ return;
+
+ // If not actually striding (add x, 0.0), avoid touching the code.
+ if (IncValue == 0)
+ return;
+
+ // Positive and negative strides have different safety conditions.
+ if (IncValue > 0) {
+ // If we have a positive stride, we require the init to be less than the
+ // exit value.
+ if (InitValue >= ExitValue)
+ return;
+
+ uint32_t Range = uint32_t(ExitValue-InitValue);
+ // Check for infinite loop, either:
+ // while (i <= Exit) or until (i > Exit)
+ if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) {
+ if (++Range == 0) return; // Range overflows.
+ }
+
+ unsigned Leftover = Range % uint32_t(IncValue);
+
+ // If this is an equality comparison, we require that the strided value
+ // exactly land on the exit value, otherwise the IV condition will wrap
+ // around and do things the fp IV wouldn't.
+ if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
+ Leftover != 0)
+ return;
+
+ // If the stride would wrap around the i32 before exiting, we can't
+ // transform the IV.
+ if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue)
+ return;
- // 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 = L->getCanonicalInductionVariableIncrement();
} else {
- // We have to use the preincremented value...
- RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
- IndVar->getType());
- CmpIndVar = IndVar;
+ // If we have a negative stride, we require the init to be greater than the
+ // exit value.
+ if (InitValue <= ExitValue)
+ return;
+
+ uint32_t Range = uint32_t(InitValue-ExitValue);
+ // Check for infinite loop, either:
+ // while (i >= Exit) or until (i < Exit)
+ if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) {
+ if (++Range == 0) return; // Range overflows.
+ }
+
+ unsigned Leftover = Range % uint32_t(-IncValue);
+
+ // If this is an equality comparison, we require that the strided value
+ // exactly land on the exit value, otherwise the IV condition will wrap
+ // around and do things the fp IV wouldn't.
+ if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) &&
+ Leftover != 0)
+ return;
+
+ // If the stride would wrap around the i32 before exiting, we can't
+ // transform the IV.
+ if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue)
+ return;
}
- // 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());
+ IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext());
- // Insert a new icmp_ne or icmp_eq instruction before the branch.
- ICmpInst::Predicate Opcode;
- if (L->contains(BI->getSuccessor(0)))
- Opcode = ICmpInst::ICMP_NE;
- else
- Opcode = ICmpInst::ICMP_EQ;
+ // Insert new integer induction variable.
+ PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN);
+ NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue),
+ PN->getIncomingBlock(IncomingEdge));
+
+ Value *NewAdd =
+ BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue),
+ Incr->getName()+".int", Incr);
+ NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge));
- DOUT << "INDVARS: Rewriting loop exit condition to:\n"
- << " LHS:" << *CmpIndVar // includes a newline
- << " op:\t"
- << (Opcode == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
- << " RHS:\t" << *RHS << "\n";
+ ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd,
+ ConstantInt::get(Int32Ty, ExitValue),
+ Compare->getName());
- ICmpInst *Cond = new ICmpInst(Opcode, CmpIndVar, ExitCnt, "exitcond", BI);
+ // In the following deletions, PN may become dead and may be deleted.
+ // Use a WeakVH to observe whether this happens.
+ WeakVH WeakPH = PN;
- Instruction *OrigCond = cast<Instruction>(BI->getCondition());
- OrigCond->replaceAllUsesWith(Cond);
- RecursivelyDeleteTriviallyDeadInstructions(OrigCond);
+ // Delete the old floating point exit comparison. The branch starts using the
+ // new comparison.
+ NewCompare->takeName(Compare);
+ Compare->replaceAllUsesWith(NewCompare);
+ RecursivelyDeleteTriviallyDeadInstructions(Compare, TLI);
- ++NumLFTR;
+ // Delete the old floating point increment.
+ Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
+ RecursivelyDeleteTriviallyDeadInstructions(Incr, TLI);
+
+ // If the FP induction variable still has uses, this is because something else
+ // in the loop uses its value. In order to canonicalize the induction
+ // variable, we chose to eliminate the IV and rewrite it in terms of an
+ // int->fp cast.
+ //
+ // We give preference to sitofp over uitofp because it is faster on most
+ // platforms.
+ if (WeakPH) {
+ Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv",
+ PN->getParent()->getFirstInsertionPt());
+ PN->replaceAllUsesWith(Conv);
+ RecursivelyDeleteTriviallyDeadInstructions(PN, TLI);
+ }
Changed = true;
- return Cond;
}
+void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
+ // First step. Check to see if there are any floating-point recurrences.
+ // If there are, change them into integer recurrences, permitting analysis by
+ // the SCEV routines.
+ //
+ BasicBlock *Header = L->getHeader();
+
+ SmallVector<WeakVH, 8> PHIs;
+ for (BasicBlock::iterator I = Header->begin();
+ PHINode *PN = dyn_cast<PHINode>(I); ++I)
+ PHIs.push_back(PN);
+
+ for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
+ if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i]))
+ HandleFloatingPointIV(L, PN);
+
+ // If the loop previously had floating-point IV, ScalarEvolution
+ // 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->forgetLoop(L);
+}
+
+//===----------------------------------------------------------------------===//
+// RewriteLoopExitValues - Optimize IV users outside the loop.
+// As a side effect, reduces the amount of IV processing within the loop.
+//===----------------------------------------------------------------------===//
+
/// RewriteLoopExitValues - 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 that are recurrent in the loop, and
/// happen later, except that it's more powerful in some cases, because it's
/// 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) {
+void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) {
// Verify the input to the pass in already in LCSSA form.
- assert(L->isLCSSAForm());
+ assert(L->isLCSSAForm(*DT));
- 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
+
+ // SCEV only supports integer expressions for now.
+ if (!PN->getType()->isIntegerTy() && !PN->getType()->isPointerTy())
+ continue;
+
+ // It's necessary to tell ScalarEvolution about this explicitly so that
+ // it can walk the def-use list and forget all SCEVs, as it may not be
+ // watching the PHI itself. Once the new exit value is in place, there
+ // may not be a def-use connection between the loop and every instruction
+ // which got a SCEVAddRecExpr for that loop.
+ SE->forgetValue(PN);
// 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
// in the loop, skip it.
Value *InVal = PN->getIncomingValue(i);
- if (!isa<Instruction>(InVal) ||
- // SCEV only supports integer expressions for now.
- (!isa<IntegerType>(InVal->getType()) &&
- !isa<PointerType>(InVal->getType())))
+ if (!isa<Instruction>(InVal))
continue;
// If this pred is for a subloop, not L itself, skip it.
// 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 (!SE->isLoopInvariant(ExitValue, L))
continue;
- Changed = true;
- ++NumReplaced;
+ Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst);
- // 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);
+ DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n'
+ << " LoopVal = " << *Inst << "\n");
- DOUT << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal
- << " LoopVal = " << *Inst << "\n";
+ if (!isValidRewrite(Inst, ExitVal)) {
+ DeadInsts.push_back(ExitVal);
+ continue;
+ }
+ Changed = true;
+ ++NumReplaced;
PN->setIncomingValue(i, ExitVal);
- // If this instruction is dead now, delete it.
- RecursivelyDeleteTriviallyDeadInstructions(Inst);
+ // If this instruction is dead now, delete it. Don't do it now to avoid
+ // invalidating iterators.
+ if (isInstructionTriviallyDead(Inst, TLI))
+ DeadInsts.push_back(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);
- RecursivelyDeleteTriviallyDeadInstructions(PN);
- break;
+ 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.
+ PHINode *NewPN = cast<PHINode>(PN->clone());
+ NewPN->takeName(PN);
+ NewPN->insertBefore(PN);
+ PN->replaceAllUsesWith(NewPN);
+ PN->eraseFromParent();
+ }
}
}
+
+ // The insertion point instruction may have been deleted; clear it out
+ // so that the rewriter doesn't trip over it later.
+ Rewriter.clearInsertPoint();
}
-void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
- // First step. Check to see if there are any floating-point recurrences.
- // If there are, change them into integer recurrences, permitting analysis by
- // the SCEV routines.
- //
- BasicBlock *Header = L->getHeader();
+//===----------------------------------------------------------------------===//
+// IV Widening - Extend the width of an IV to cover its widest uses.
+//===----------------------------------------------------------------------===//
- SmallVector<WeakVH, 8> PHIs;
- for (BasicBlock::iterator I = Header->begin();
- PHINode *PN = dyn_cast<PHINode>(I); ++I)
- PHIs.push_back(PN);
+namespace {
+ // Collect information about induction variables that are used by sign/zero
+ // extend operations. This information is recorded by CollectExtend and
+ // provides the input to WidenIV.
+ struct WideIVInfo {
+ PHINode *NarrowIV;
+ Type *WidestNativeType; // Widest integer type created [sz]ext
+ bool IsSigned; // Was an sext user seen before a zext?
+
+ WideIVInfo() : NarrowIV(0), WidestNativeType(0), IsSigned(false) {}
+ };
- for (unsigned i = 0, e = PHIs.size(); i != e; ++i)
- if (PHINode *PN = dyn_cast_or_null<PHINode>(PHIs[i]))
- HandleFloatingPointIV(L, PN);
+ class WideIVVisitor : public IVVisitor {
+ ScalarEvolution *SE;
+ const DataLayout *TD;
- // If the loop previously had floating-point IV, ScalarEvolution
- // 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);
-}
+ public:
+ WideIVInfo WI;
-bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
- IU = &getAnalysis<IVUsers>();
- LI = &getAnalysis<LoopInfo>();
- SE = &getAnalysis<ScalarEvolution>();
- Changed = false;
+ WideIVVisitor(PHINode *NarrowIV, ScalarEvolution *SCEV,
+ const DataLayout *TData) :
+ SE(SCEV), TD(TData) { WI.NarrowIV = NarrowIV; }
- // If there are any floating-point recurrences, attempt to
- // transform them to use integer recurrences.
- RewriteNonIntegerIVs(L);
+ // Implement the interface used by simplifyUsersOfIV.
+ virtual void visitCast(CastInst *Cast);
+ };
+}
- BasicBlock *Header = L->getHeader();
- BasicBlock *ExitingBlock = L->getExitingBlock(); // may be null
- SCEVHandle BackedgeTakenCount = SE->getBackedgeTakenCount(L);
+/// 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) {
+ bool IsSigned = Cast->getOpcode() == Instruction::SExt;
+ if (!IsSigned && Cast->getOpcode() != Instruction::ZExt)
+ return;
- // 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
- // that are recurrent in the loop, and substitute the exit values from the
- // loop into any instructions outside of the loop that use the final values of
- // the current expressions.
- //
- if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
- RewriteLoopExitValues(L, BackedgeTakenCount);
-
- // Compute the type of the largest recurrence expression, and decide whether
- // a canonical induction variable should be inserted.
- const Type *LargestType = 0;
- bool NeedCannIV = false;
- if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
- LargestType = BackedgeTakenCount->getType();
- LargestType = SE->getEffectiveSCEVType(LargestType);
- // If we have a known trip count and a single exit block, we'll be
- // rewriting the loop exit test condition below, which requires a
- // canonical induction variable.
- if (ExitingBlock)
- NeedCannIV = true;
- }
- for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) {
- SCEVHandle 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 =
- IU->IVUsesByStride.find(IU->StrideOrder[i]);
- assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!");
-
- if (!SI->second->Users.empty())
- NeedCannIV = true;
+ Type *Ty = Cast->getType();
+ uint64_t Width = SE->getTypeSizeInBits(Ty);
+ if (TD && !TD->isLegalInteger(Width))
+ return;
+
+ if (!WI.WidestNativeType) {
+ WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
+ WI.IsSigned = IsSigned;
+ return;
}
- // Create a rewriter object which we'll use to transform the code with.
- SCEVExpander Rewriter(*SE);
+ // We extend the IV to satisfy the sign of its first user, arbitrarily.
+ if (WI.IsSigned != IsSigned)
+ return;
- // 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);
- ++NumInserted;
- Changed = true;
- DOUT << "INDVARS: New CanIV: " << *IndVar;
- }
+ if (Width > SE->getTypeSizeInBits(WI.WidestNativeType))
+ WI.WidestNativeType = SE->getEffectiveSCEVType(Ty);
+}
- // 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.
- ICmpInst *NewICmp = 0;
- if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount) && ExitingBlock) {
- assert(NeedCannIV &&
- "LinearFunctionTestReplace requires a canonical induction variable");
- // Can't rewrite non-branch yet.
- if (BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator()))
- NewICmp = LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
- ExitingBlock, BI, Rewriter);
- }
+namespace {
- Rewriter.setInsertionPoint(Header->getFirstNonPHI());
+/// NarrowIVDefUse - Record a link in the Narrow IV def-use chain along with the
+/// WideIV that computes the same value as the Narrow IV def. This avoids
+/// caching Use* pointers.
+struct NarrowIVDefUse {
+ Instruction *NarrowDef;
+ Instruction *NarrowUse;
+ Instruction *WideDef;
- // Rewrite IV-derived expressions.
- RewriteIVExpressions(L, LargestType, Rewriter);
+ NarrowIVDefUse(): NarrowDef(0), NarrowUse(0), WideDef(0) {}
- // 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);
+ NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD):
+ NarrowDef(ND), NarrowUse(NU), WideDef(WD) {}
+};
- // Reorder instructions to avoid use-before-def conditions.
- FixUsesBeforeDefs(L, Rewriter);
+/// WidenIV - The goal of this transform is to remove sign and zero extends
+/// without creating any new induction variables. To do this, it creates a new
+/// phi of the wider type and redirects all users, either removing extends or
+/// inserting truncs whenever we stop propagating the type.
+///
+class WidenIV {
+ // Parameters
+ PHINode *OrigPhi;
+ Type *WideType;
+ bool IsSigned;
+
+ // Context
+ LoopInfo *LI;
+ Loop *L;
+ ScalarEvolution *SE;
+ DominatorTree *DT;
+
+ // Result
+ PHINode *WidePhi;
+ Instruction *WideInc;
+ const SCEV *WideIncExpr;
+ SmallVectorImpl<WeakVH> &DeadInsts;
+
+ SmallPtrSet<Instruction*,16> Widened;
+ SmallVector<NarrowIVDefUse, 8> NarrowIVUsers;
+
+public:
+ WidenIV(const WideIVInfo &WI, LoopInfo *LInfo,
+ ScalarEvolution *SEv, DominatorTree *DTree,
+ SmallVectorImpl<WeakVH> &DI) :
+ OrigPhi(WI.NarrowIV),
+ WideType(WI.WidestNativeType),
+ IsSigned(WI.IsSigned),
+ LI(LInfo),
+ L(LI->getLoopFor(OrigPhi->getParent())),
+ SE(SEv),
+ DT(DTree),
+ WidePhi(0),
+ WideInc(0),
+ WideIncExpr(0),
+ DeadInsts(DI) {
+ assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV");
+ }
- // For completeness, inform IVUsers of the IV use in the newly-created
- // loop exit test instruction.
- if (NewICmp)
- IU->AddUsersIfInteresting(cast<Instruction>(NewICmp->getOperand(0)));
+ PHINode *CreateWideIV(SCEVExpander &Rewriter);
- // Clean up dead instructions.
- DeleteDeadPHIs(L->getHeader());
- // Check a post-condition.
- assert(L->isLCSSAForm() && "Indvars did not leave the loop in lcssa form!");
- return Changed;
-}
+protected:
+ Value *getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
+ Instruction *Use);
-void IndVarSimplify::RewriteIVExpressions(Loop *L, const Type *LargestType,
- SCEVExpander &Rewriter) {
- SmallVector<WeakVH, 16> DeadInsts;
+ Instruction *CloneIVUser(NarrowIVDefUse DU);
- // Rewrite all induction variable expressions in terms of the canonical
- // induction variable.
- //
- // If there were induction variables of other sizes or offsets, manually
- // add the offsets to the primary induction variable and cast, avoiding
- // 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];
-
- std::map<SCEVHandle, 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();
- Instruction *User = UI->getUser();
- bool isSigned = UI->isSigned();
-
- // Compute the final addrec to expand into code.
- SCEVHandle 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
- // polynomial evaluations inside of the loop, and the str reduction pass
- // 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()))
- continue;
+ const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse);
- 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);
- else
- PromotedStride = SE->getZeroExtendExpr(PromotedStride,
- LargestType);
- }
- }
+ const SCEVAddRecExpr* GetExtendedOperandRecurrence(NarrowIVDefUse DU);
- // 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);
- }
- }
+ Instruction *WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter);
- DOUT << "INDVARS: Made offset-and-trunc IV for offset "
- << *IVTy << " " << *Offset << ": ";
- DEBUG(WriteAsOperand(*DOUT, V, false));
- DOUT << "\n";
+ void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef);
+};
+} // anonymous namespace
- // Now expand it into actual Instructions and patch it into place.
- NewVal = Rewriter.expandCodeFor(AR, UseTy);
- }
+/// isLoopInvariant - Perform a quick domtree based check for loop invariance
+/// assuming that V is used within the loop. LoopInfo::isLoopInvariant() seems
+/// gratuitous for this purpose.
+static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) {
+ Instruction *Inst = dyn_cast<Instruction>(V);
+ if (!Inst)
+ return true;
+
+ return DT->properlyDominates(Inst->getParent(), L->getHeader());
+}
- // 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";
- ++NumRemoved;
- Changed = true;
-
- // The old value may be dead now.
- DeadInsts.push_back(Op);
+Value *WidenIV::getExtend(Value *NarrowOper, Type *WideType, bool IsSigned,
+ Instruction *Use) {
+ // Set the debug location and conservative insertion point.
+ IRBuilder<> Builder(Use);
+ // Hoist the insertion point into loop preheaders as far as possible.
+ for (const Loop *L = LI->getLoopFor(Use->getParent());
+ L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT);
+ L = L->getParentLoop())
+ Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator());
+
+ return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) :
+ Builder.CreateZExt(NarrowOper, WideType);
+}
+
+/// CloneIVUser - Instantiate a wide operation to replace a narrow
+/// operation. This only needs to handle operations that can evaluation to
+/// SCEVAddRec. It can safely return 0 for any operation we decide not to clone.
+Instruction *WidenIV::CloneIVUser(NarrowIVDefUse DU) {
+ unsigned Opcode = DU.NarrowUse->getOpcode();
+ switch (Opcode) {
+ default:
+ return 0;
+ case Instruction::Add:
+ case Instruction::Mul:
+ case Instruction::UDiv:
+ case Instruction::Sub:
+ case Instruction::And:
+ case Instruction::Or:
+ case Instruction::Xor:
+ case Instruction::Shl:
+ case Instruction::LShr:
+ case Instruction::AShr:
+ DEBUG(dbgs() << "Cloning IVUser: " << *DU.NarrowUse << "\n");
+
+ // Replace NarrowDef operands with WideDef. Otherwise, we don't know
+ // anything about the narrow operand yet so must insert a [sz]ext. It is
+ // probably loop invariant and will be folded or hoisted. If it actually
+ // comes from a widened IV, it should be removed during a future call to
+ // WidenIVUse.
+ Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef) ? DU.WideDef :
+ getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned, DU.NarrowUse);
+ Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef) ? DU.WideDef :
+ getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned, DU.NarrowUse);
+
+ BinaryOperator *NarrowBO = cast<BinaryOperator>(DU.NarrowUse);
+ BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(),
+ LHS, RHS,
+ NarrowBO->getName());
+ IRBuilder<> Builder(DU.NarrowUse);
+ Builder.Insert(WideBO);
+ if (const OverflowingBinaryOperator *OBO =
+ dyn_cast<OverflowingBinaryOperator>(NarrowBO)) {
+ if (OBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap();
+ if (OBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap();
}
+ return WideBO;
}
+}
- // Now that we're done iterating through lists, clean up any instructions
- // which are now dead.
- while (!DeadInsts.empty()) {
- Instruction *Inst = dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val());
- if (Inst)
- RecursivelyDeleteTriviallyDeadInstructions(Inst);
- }
+/// 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;
+
+ // 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;
+ assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU");
+
+ const SCEV *ExtendOperExpr = 0;
+ const OverflowingBinaryOperator *OBO =
+ cast<OverflowingBinaryOperator>(DU.NarrowUse);
+ if (IsSigned && OBO->hasNoSignedWrap())
+ ExtendOperExpr = SE->getSignExtendExpr(
+ SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
+ else if(!IsSigned && OBO->hasNoUnsignedWrap())
+ ExtendOperExpr = SE->getZeroExtendExpr(
+ SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType);
+ else
+ return 0;
+
+ // When creating this AddExpr, 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));
+
+ if (!AddRec || AddRec->getLoop() != L)
+ return 0;
+ return AddRec;
}
-/// 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) {
- BasicBlock *ExitBlock = L->getExitBlock();
- if (!ExitBlock) return;
+/// GetWideRecurrence - Is this instruction potentially interesting from
+/// IVUsers' perspective after widening it's type? In other words, can the
+/// extend be safely hoisted out of the loop with SCEV reducing the value to a
+/// recurrence on the same loop. If so, return the sign or zero extended
+/// recurrence. Otherwise return NULL.
+const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) {
+ if (!SE->isSCEVable(NarrowUse->getType()))
+ return 0;
+
+ 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;
+ }
- Instruction *NonPHI = ExitBlock->getFirstNonPHI();
- BasicBlock *Preheader = L->getLoopPreheader();
- 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))
- break;
- // Determine if there is a use in or before the loop (direct or
- // otherwise).
- bool UsedInLoop = false;
- for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
- UI != UE; ++UI) {
- BasicBlock *UseBB = cast<Instruction>(UI)->getParent();
- if (PHINode *P = dyn_cast<PHINode>(UI)) {
- unsigned i =
- PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
- UseBB = P->getIncomingBlock(i);
+ const SCEV *WideExpr = IsSigned ?
+ SE->getSignExtendExpr(NarrowExpr, WideType) :
+ SE->getZeroExtendExpr(NarrowExpr, WideType);
+ const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr);
+ if (!AddRec || AddRec->getLoop() != L)
+ return 0;
+ return AddRec;
+}
+
+/// 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;
+
+ // Our raison d'etre! Eliminate sign and zero extension.
+ if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) {
+ Value *NewDef = DU.WideDef;
+ if (DU.NarrowUse->getType() != WideType) {
+ unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType());
+ unsigned IVWidth = SE->getTypeSizeInBits(WideType);
+ if (CastWidth < IVWidth) {
+ // The cast isn't as wide as the IV, so insert a Trunc.
+ IRBuilder<> Builder(DU.NarrowUse);
+ NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType());
}
- if (UseBB == Preheader || L->contains(UseBB)) {
- UsedInLoop = true;
- break;
+ else {
+ // A wider extend was hidden behind a narrower one. This may induce
+ // another round of IV widening in which the intermediate IV becomes
+ // dead. It should be very rare.
+ DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi
+ << " not wide enough to subsume " << *DU.NarrowUse << "\n");
+ DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef);
+ NewDef = DU.NarrowUse;
}
}
- // If there is, the def must remain in the preheader.
- if (UsedInLoop)
+ if (NewDef != DU.NarrowUse) {
+ DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse
+ << " replaced by " << *DU.WideDef << "\n");
+ ++NumElimExt;
+ DU.NarrowUse->replaceAllUsesWith(NewDef);
+ DeadInsts.push_back(DU.NarrowUse);
+ }
+ // Now that the extend is gone, we want to expose it's uses for potential
+ // further simplification. We don't need to directly inform SimplifyIVUsers
+ // of the new users, because their parent IV will be processed later as a
+ // new loop phi. If we preserved IVUsers analysis, we would also want to
+ // push the uses of WideDef here.
+
+ // No further widening is needed. The deceased [sz]ext had done it for us.
+ return 0;
+ }
+
+ // Does this user itself evaluate to a recurrence after widening?
+ const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse);
+ if (!WideAddRec) {
+ WideAddRec = GetExtendedOperandRecurrence(DU);
+ }
+ if (!WideAddRec) {
+ // 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;
+ }
+ // 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.
+ assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() &&
+ "SCEV is not expected to evaluate a block terminator");
+
+ // Reuse the IV increment that SCEVExpander created as long as it dominates
+ // NarrowUse.
+ Instruction *WideUse = 0;
+ if (WideAddRec == WideIncExpr
+ && Rewriter.hoistIVInc(WideInc, DU.NarrowUse))
+ WideUse = WideInc;
+ else {
+ WideUse = CloneIVUser(DU);
+ if (!WideUse)
+ return 0;
+ }
+ // Evaluation of WideAddRec ensured that the narrow expression could be
+ // extended outside the loop without overflow. This suggests that the wide use
+ // evaluates to the same expression as the extended narrow use, but doesn't
+ // absolutely guarantee it. Hence the following failsafe check. In rare cases
+ // where it fails, we simply throw away the newly created wide use.
+ if (WideAddRec != SE->getSCEV(WideUse)) {
+ DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse
+ << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n");
+ DeadInsts.push_back(WideUse);
+ return 0;
+ }
+
+ // Returning WideUse pushes it on the worklist.
+ return WideUse;
+}
+
+/// 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);
+
+ // Handle data flow merges and bizarre phi cycles.
+ if (!Widened.insert(NarrowUse))
continue;
- // Otherwise, sink it to the exit block.
- Instruction *ToMove = I;
- bool Done = false;
- if (I != Preheader->begin())
- --I;
- else
- Done = true;
- ToMove->moveBefore(NonPHI);
- if (Done)
- break;
+
+ NarrowIVUsers.push_back(NarrowIVDefUse(NarrowDef, NarrowUse, WideDef));
}
}
-/// 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 = next(I);
- 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());
- }
+/// CreateWideIV - Process a single induction variable. First use the
+/// SCEVExpander to create a wide induction variable that evaluates to the same
+/// recurrence as the original narrow IV. Then use a worklist to forward
+/// traverse the narrow IV's def-use chain. After WidenIVUse has processed all
+/// interesting IV users, the narrow IV will be isolated for removal by
+/// DeleteDeadPHIs.
+///
+/// It would be simpler to delete uses as they are processed, but we must avoid
+/// invalidating SCEV expressions.
+///
+PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) {
+ // Is this phi an induction variable?
+ const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi));
+ if (!AddRec)
+ return NULL;
+
+ // Widen the induction variable expression.
+ const SCEV *WideIVExpr = IsSigned ?
+ SE->getSignExtendExpr(AddRec, WideType) :
+ SE->getZeroExtendExpr(AddRec, WideType);
+
+ assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType &&
+ "Expect the new IV expression to preserve its type");
+
+ // Can the IV be extended outside the loop without overflow?
+ AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr);
+ if (!AddRec || AddRec->getLoop() != L)
+ return NULL;
+
+ // An AddRec must have loop-invariant operands. Since this AddRec is
+ // materialized by a loop header phi, the expression cannot have any post-loop
+ // operands, so they must dominate the loop header.
+ assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) &&
+ SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader())
+ && "Loop header phi recurrence inputs do not dominate the loop");
+
+ // The rewriter provides a value for the desired IV expression. This may
+ // either find an existing phi or materialize a new one. Either way, we
+ // expect a well-formed cyclic phi-with-increments. i.e. any operand not part
+ // of the phi-SCC dominates the loop entry.
+ Instruction *InsertPt = L->getHeader()->begin();
+ WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt));
+
+ // Remembering the WideIV increment generated by SCEVExpander allows
+ // WidenIVUse to reuse it when widening the narrow IV's increment. We don't
+ // employ a general reuse mechanism because the call above is the only call to
+ // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses.
+ if (BasicBlock *LatchBlock = L->getLoopLatch()) {
+ WideInc =
+ cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock));
+ WideIncExpr = SE->getSCEV(WideInc);
+ }
+
+ DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n");
+ ++NumWidened;
+
+ // Traverse the def-use chain using a worklist starting at the original IV.
+ assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" );
+
+ Widened.insert(OrigPhi);
+ pushNarrowIVUsers(OrigPhi, WidePhi);
+
+ while (!NarrowIVUsers.empty()) {
+ NarrowIVDefUse DU = NarrowIVUsers.pop_back_val();
+
+ // Process a def-use edge. This may replace the use, so don't hold a
+ // use_iterator across it.
+ Instruction *WideUse = WidenIVUse(DU, Rewriter);
+
+ // Follow all def-use edges from the previous narrow use.
+ if (WideUse)
+ pushNarrowIVUsers(DU.NarrowUse, WideUse);
+
+ // WidenIVUse may have removed the def-use edge.
+ if (DU.NarrowDef->use_empty())
+ DeadInsts.push_back(DU.NarrowDef);
+ }
+ return WidePhi;
+}
+
+//===----------------------------------------------------------------------===//
+// Simplification of IV users based on SCEV evaluation.
+//===----------------------------------------------------------------------===//
+
+
+/// SimplifyAndExtend - Iteratively perform simplification on a worklist of IV
+/// users. Each successive simplification may push more users which may
+/// themselves be candidates for simplification.
+///
+/// Sign/Zero extend elimination is interleaved with IV simplification.
+///
+void IndVarSimplify::SimplifyAndExtend(Loop *L,
+ SCEVExpander &Rewriter,
+ LPPassManager &LPM) {
+ SmallVector<WideIVInfo, 8> WideIVs;
+
+ SmallVector<PHINode*, 8> LoopPhis;
+ for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
+ LoopPhis.push_back(cast<PHINode>(I));
+ }
+ // Each round of simplification iterates through the SimplifyIVUsers worklist
+ // for all current phis, then determines whether any IVs can be
+ // widened. Widening adds new phis to LoopPhis, inducing another round of
+ // simplification on the wide IVs.
+ while (!LoopPhis.empty()) {
+ // Evaluate as many IV expressions as possible before widening any IVs. This
+ // forces SCEV to set no-wrap flags before evaluating sign/zero
+ // extension. The first time SCEV attempts to normalize sign/zero extension,
+ // the result becomes final. So for the most predictable results, we delay
+ // evaluation of sign/zero extend evaluation until needed, and avoid running
+ // other SCEV based analysis prior to SimplifyAndExtend.
+ do {
+ PHINode *CurrIV = LoopPhis.pop_back_val();
+
+ // Information about sign/zero extensions of CurrIV.
+ WideIVVisitor WIV(CurrIV, SE, TD);
+
+ Changed |= simplifyUsersOfIV(CurrIV, SE, &LPM, DeadInsts, &WIV);
+
+ if (WIV.WI.WidestNativeType) {
+ WideIVs.push_back(WIV.WI);
+ }
+ } while(!LoopPhis.empty());
+
+ for (; !WideIVs.empty(); WideIVs.pop_back()) {
+ WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts);
+ if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) {
+ Changed = true;
+ LoopPhis.push_back(WidePhi);
}
}
- } 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,
- uint64_t intIV, uint64_t intEV) {
+//===----------------------------------------------------------------------===//
+// LinearFunctionTestReplace and its kin. Rewrite the loop exit condition.
+//===----------------------------------------------------------------------===//
- if (InitV.getValueAPF().isNegative() || ExitV.getValueAPF().isNegative())
- return true;
+/// Check for expressions that ScalarEvolution generates to compute
+/// 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,
+ ScalarEvolution *SE) {
+ if (!Processed.insert(S))
+ return false;
+
+ // If the backedge-taken count is a UDiv, it's very likely a UDiv that
+ // ScalarEvolution's HowFarToZero or HowManyLessThans produced to compute a
+ // precise expression, rather than a UDiv from the user's code. If we can't
+ // find a UDiv in the code with some simple searching, assume the former and
+ // forego rewriting the loop.
+ if (isa<SCEVUDivExpr>(S)) {
+ ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition());
+ if (!OrigCond) return true;
+ const SCEV *R = SE->getSCEV(OrigCond->getOperand(1));
+ R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1));
+ if (R != S) {
+ const SCEV *L = SE->getSCEV(OrigCond->getOperand(0));
+ L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1));
+ if (L != S)
+ return true;
+ }
+ }
+
+ // Recurse past add expressions, which commonly occur in the
+ // BackedgeTakenCount. They may already exist in program code, and if not,
+ // they are not too expensive rematerialize.
+ if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
+ for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
+ I != E; ++I) {
+ if (isHighCostExpansion(*I, BI, Processed, SE))
+ return true;
+ }
+ return false;
+ }
- // If the iteration range can be handled by SIToFPInst then use it.
- APInt Max = APInt::getSignedMaxValue(32);
- if (Max.getZExtValue() > static_cast<uint64_t>(abs64(intEV - intIV)))
+ // HowManyLessThans uses a Max expression whenever the loop is not guarded by
+ // the exit condition.
+ if (isa<SCEVSMaxExpr>(S) || isa<SCEVUMaxExpr>(S))
return true;
+ // If we haven't recognized an expensive SCEV pattern, assume it's an
+ // expression produced by program code.
return false;
}
-/// convertToInt - Convert APF to an integer, if possible.
-static bool convertToInt(const APFloat &APF, uint64_t *intVal) {
+/// canExpandBackedgeTakenCount - Return true if this loop's backedge taken
+/// count expression can be safely and cheaply expanded into an instruction
+/// sequence that can be used by LinearFunctionTestReplace.
+///
+/// TODO: This fails for pointer-type loop counters with greater than one byte
+/// strides, consequently preventing LFTR from running. For the purpose of LFTR
+/// we could skip this check in the case that the LFTR loop counter (chosen by
+/// FindLoopCounter) is also pointer type. Instead, we could directly convert
+/// the loop test to an inequality test by checking the target data's alignment
+/// of element types (given that the initial pointer value originates from or is
+/// used by ABI constrained operation, as opposed to inttoptr/ptrtoint).
+/// However, we don't yet have a strong motivation for converting loop tests
+/// into inequality tests.
+static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE) {
+ const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
+ if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
+ BackedgeTakenCount->isZero())
+ return false;
- bool isExact = false;
- if (&APF.getSemantics() == &APFloat::PPCDoubleDouble)
+ if (!L->getExitingBlock())
return false;
- if (APF.convertToInteger(intVal, 32, APF.isNegative(),
- APFloat::rmTowardZero, &isExact)
- != APFloat::opOK)
+
+ // Can't rewrite non-branch yet.
+ BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
+ if (!BI)
return false;
- if (!isExact)
+
+ SmallPtrSet<const SCEV*, 8> Processed;
+ if (isHighCostExpansion(BackedgeTakenCount, BI, Processed, SE))
return false;
+
return true;
+}
+
+/// getLoopPhiForCounter - Return the loop header phi IFF IncV adds a loop
+/// invariant value to the phi.
+static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) {
+ Instruction *IncI = dyn_cast<Instruction>(IncV);
+ if (!IncI)
+ return 0;
+
+ switch (IncI->getOpcode()) {
+ case Instruction::Add:
+ case Instruction::Sub:
+ break;
+ case Instruction::GetElementPtr:
+ // An IV counter must preserve its type.
+ if (IncI->getNumOperands() == 2)
+ break;
+ default:
+ return 0;
+ }
+ 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;
+ }
+ if (IncI->getOpcode() == Instruction::GetElementPtr)
+ return 0;
+
+ // Allow add/sub to be commuted.
+ Phi = dyn_cast<PHINode>(IncI->getOperand(1));
+ if (Phi && Phi->getParent() == L->getHeader()) {
+ if (isLoopInvariant(IncI->getOperand(0), L, DT))
+ return Phi;
+ }
+ return 0;
}
-/// HandleFloatingPointIV - If the loop has floating induction variable
-/// then insert corresponding integer induction variable if possible.
-/// For example,
-/// for(double i = 0; i < 10000; ++i)
-/// bar(i)
-/// is converted into
-/// for(int i = 0; i < 10000; ++i)
-/// bar((double)i);
+/// Return the compare guarding the loop latch, or NULL for unrecognized tests.
+static ICmpInst *getLoopTest(Loop *L) {
+ assert(L->getExitingBlock() && "expected loop exit");
+
+ BasicBlock *LatchBlock = L->getLoopLatch();
+ // Don't bother with LFTR if the loop is not properly simplified.
+ if (!LatchBlock)
+ return 0;
+
+ BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator());
+ assert(BI && "expected exit branch");
+
+ return dyn_cast<ICmpInst>(BI->getCondition());
+}
+
+/// needsLFTR - LinearFunctionTestReplace policy. Return true unless we can show
+/// that the current exit test is already sufficiently canonical.
+static bool needsLFTR(Loop *L, DominatorTree *DT) {
+ // Do LFTR to simplify the exit condition to an ICMP.
+ ICmpInst *Cond = getLoopTest(L);
+ if (!Cond)
+ return true;
+
+ // Do LFTR to simplify the exit ICMP to EQ/NE
+ ICmpInst::Predicate Pred = Cond->getPredicate();
+ if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ)
+ return true;
+
+ // Look for a loop invariant RHS
+ Value *LHS = Cond->getOperand(0);
+ Value *RHS = Cond->getOperand(1);
+ if (!isLoopInvariant(RHS, L, DT)) {
+ if (!isLoopInvariant(LHS, L, DT))
+ return true;
+ std::swap(LHS, RHS);
+ }
+ // Look for a simple IV counter LHS
+ PHINode *Phi = dyn_cast<PHINode>(LHS);
+ if (!Phi)
+ Phi = getLoopPhiForCounter(LHS, L, DT);
+
+ if (!Phi)
+ return true;
+
+ // Do LFTR if PHI node is defined in the loop, but is *not* a counter.
+ int Idx = Phi->getBasicBlockIndex(L->getLoopLatch());
+ if (Idx < 0)
+ return true;
+
+ // Do LFTR if the exit condition's IV is *not* a simple counter.
+ Value *IncV = Phi->getIncomingValue(Idx);
+ return Phi != getLoopPhiForCounter(IncV, L, DT);
+}
+
+/// 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,
+ unsigned Depth) {
+ if (isa<Constant>(V))
+ return !isa<UndefValue>(V);
+
+ if (Depth >= 6)
+ return false;
+
+ // Conservatively handle non-constant non-instructions. For example, Arguments
+ // may be undef.
+ Instruction *I = dyn_cast<Instruction>(V);
+ if (!I)
+ return false;
+
+ // Load and return values may be undef.
+ if(I->mayReadFromMemory() || isa<CallInst>(I) || isa<InvokeInst>(I))
+ return false;
+
+ // Optimistically handle other instructions.
+ for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) {
+ if (!Visited.insert(*OI))
+ continue;
+ if (!hasConcreteDefImpl(*OI, Visited, Depth+1))
+ return false;
+ }
+ return true;
+}
+
+/// Return true if the given value is concrete. We must prove that undef can
+/// never reach it.
///
-void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PH) {
+/// TODO: If we decide that this is a good approach to checking for undef, we
+/// may factor it into a common location.
+static bool hasConcreteDef(Value *V) {
+ SmallPtrSet<Value*, 8> Visited;
+ Visited.insert(V);
+ return hasConcreteDefImpl(V, Visited, 0);
+}
- unsigned IncomingEdge = L->contains(PH->getIncomingBlock(0));
- unsigned BackEdge = IncomingEdge^1;
+/// AlmostDeadIV - Return true if this IV has any uses other than the (soon to
+/// be rewritten) loop exit test.
+static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) {
+ int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
+ Value *IncV = Phi->getIncomingValue(LatchIdx);
- // Check incoming value.
- ConstantFP *InitValue = dyn_cast<ConstantFP>(PH->getIncomingValue(IncomingEdge));
- if (!InitValue) return;
- uint64_t newInitValue = Type::Int32Ty->getPrimitiveSizeInBits();
- if (!convertToInt(InitValue->getValueAPF(), &newInitValue))
- return;
+ for (Value::use_iterator UI = Phi->use_begin(), UE = Phi->use_end();
+ UI != UE; ++UI) {
+ if (*UI != Cond && *UI != IncV) return false;
+ }
- // Check IV increment. Reject this PH if increement operation is not
- // an add or increment value can not be represented by an integer.
- BinaryOperator *Incr =
- dyn_cast<BinaryOperator>(PH->getIncomingValue(BackEdge));
- if (!Incr) return;
- if (Incr->getOpcode() != Instruction::Add) 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();
- if (!convertToInt(IncrValue->getValueAPF(), &newIncrValue))
- return;
+ for (Value::use_iterator UI = IncV->use_begin(), UE = IncV->use_end();
+ UI != UE; ++UI) {
+ if (*UI != Cond && *UI != Phi) return false;
+ }
+ return true;
+}
- // Check Incr uses. One user is PH and the other users is exit condition used
- // by the conditional terminator.
- Value::use_iterator IncrUse = Incr->use_begin();
- Instruction *U1 = cast<Instruction>(IncrUse++);
- if (IncrUse == Incr->use_end()) return;
- Instruction *U2 = cast<Instruction>(IncrUse++);
- if (IncrUse != Incr->use_end()) return;
+/// FindLoopCounter - Find an affine IV in canonical form.
+///
+/// BECount may be an i8* pointer type. The pointer difference is already
+/// valid count without scaling the address stride, so it remains a pointer
+/// expression as far as SCEV is concerned.
+///
+/// Currently only valid for LFTR. See the comments on hasConcreteDef below.
+///
+/// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount
+///
+/// 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) {
+ 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;
+ BasicBlock *LatchBlock = L->getLoopLatch();
+ assert(LatchBlock && "needsLFTR should guarantee a loop latch");
+
+ for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) {
+ PHINode *Phi = cast<PHINode>(I);
+ if (!SE->isSCEVable(Phi->getType()))
+ continue;
- // Find exit condition.
- FCmpInst *EC = dyn_cast<FCmpInst>(U1);
- if (!EC)
- EC = dyn_cast<FCmpInst>(U2);
- if (!EC) return;
-
- if (BranchInst *BI = dyn_cast<BranchInst>(EC->getParent()->getTerminator())) {
- if (!BI->isConditional()) return;
- if (BI->getCondition() != EC) return;
- }
-
- // Find exit value. If exit value can not be represented as an interger then
- // do not handle this floating point PH.
- ConstantFP *EV = NULL;
- unsigned EVIndex = 1;
- if (EC->getOperand(1) == Incr)
- EVIndex = 0;
- EV = dyn_cast<ConstantFP>(EC->getOperand(EVIndex));
- if (!EV) return;
- uint64_t intEV = Type::Int32Ty->getPrimitiveSizeInBits();
- if (!convertToInt(EV->getValueAPF(), &intEV))
- return;
+ // Avoid comparing an integer IV against a pointer Limit.
+ if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy())
+ continue;
- // Find new predicate for integer comparison.
- CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
- switch (EC->getPredicate()) {
- case CmpInst::FCMP_OEQ:
- case CmpInst::FCMP_UEQ:
- NewPred = CmpInst::ICMP_EQ;
- break;
- case CmpInst::FCMP_OGT:
- case CmpInst::FCMP_UGT:
- NewPred = CmpInst::ICMP_UGT;
- break;
- case CmpInst::FCMP_OGE:
- case CmpInst::FCMP_UGE:
- NewPred = CmpInst::ICMP_UGE;
- break;
- case CmpInst::FCMP_OLT:
- case CmpInst::FCMP_ULT:
- NewPred = CmpInst::ICMP_ULT;
- break;
- case CmpInst::FCMP_OLE:
- case CmpInst::FCMP_ULE:
- NewPred = CmpInst::ICMP_ULE;
- break;
- default:
- break;
+ const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi));
+ if (!AR || AR->getLoop() != L || !AR->isAffine())
+ continue;
+
+ // AR may be a pointer type, while BECount is an integer type.
+ // 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)))
+ continue;
+
+ const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
+ if (!Step || !Step->isOne())
+ continue;
+
+ int LatchIdx = Phi->getBasicBlockIndex(LatchBlock);
+ Value *IncV = Phi->getIncomingValue(LatchIdx);
+ if (getLoopPhiForCounter(IncV, L, DT) != Phi)
+ continue;
+
+ // Avoid reusing a potentially undef value to compute other values that may
+ // have originally had a concrete definition.
+ if (!hasConcreteDef(Phi)) {
+ // We explicitly allow unknown phis as long as they are already used by
+ // the loop test. In this case we assume that performing LFTR could not
+ // increase the number of undef users.
+ if (ICmpInst *Cond = getLoopTest(L)) {
+ if (Phi != getLoopPhiForCounter(Cond->getOperand(0), L, DT)
+ && Phi != getLoopPhiForCounter(Cond->getOperand(1), L, DT)) {
+ continue;
+ }
+ }
+ }
+ const SCEV *Init = AR->getStart();
+
+ if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) {
+ // Don't force a live loop counter if another IV can be used.
+ if (AlmostDeadIV(Phi, LatchBlock, Cond))
+ continue;
+
+ // Prefer to count-from-zero. This is a more "canonical" counter form. It
+ // also prefers integer to pointer IVs.
+ if (BestInit->isZero() != Init->isZero()) {
+ if (BestInit->isZero())
+ continue;
+ }
+ // If two IVs both count from zero or both count from nonzero then the
+ // narrower is likely a dead phi that has been widened. Use the wider phi
+ // to allow the other to be eliminated.
+ else if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType()))
+ continue;
+ }
+ BestPhi = Phi;
+ BestInit = Init;
}
- if (NewPred == CmpInst::BAD_ICMP_PREDICATE) return;
+ return BestPhi;
+}
- // Insert new integer induction variable.
- PHINode *NewPHI = PHINode::Create(Type::Int32Ty,
- PH->getName()+".int", PH);
- NewPHI->addIncoming(ConstantInt::get(Type::Int32Ty, newInitValue),
- PH->getIncomingBlock(IncomingEdge));
-
- Value *NewAdd = BinaryOperator::CreateAdd(NewPHI,
- ConstantInt::get(Type::Int32Ty,
- 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);
- 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());
-
- // In the following deltions, PH may become dead and may be deleted.
- // Use a WeakVH to observe whether this happens.
- WeakVH WeakPH = PH;
+/// 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) {
+ const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar));
+ assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter");
+ const SCEV *IVInit = AR->getStart();
+
+ // IVInit may be a pointer while IVCount is an integer when FindLoopCounter
+ // finds a valid pointer IV. Sign extend BECount in order to materialize a
+ // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing
+ // the existing GEPs whenever possible.
+ if (IndVar->getType()->isPointerTy()
+ && !IVCount->getType()->isPointerTy()) {
+
+ Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType());
+ const SCEV *IVOffset = SE->getTruncateOrSignExtend(IVCount, OfsTy);
+
+ // Expand the code for the iteration count.
+ assert(SE->isLoopInvariant(IVOffset, L) &&
+ "Computed iteration count is not loop invariant!");
+ BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
+ Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI);
+
+ Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader());
+ 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())->isOne()
+ && "unit stride pointer IV must be i8*");
+
+ IRBuilder<> Builder(L->getLoopPreheader()->getTerminator());
+ return Builder.CreateGEP(GEPBase, GEPOffset, "lftr.limit");
+ }
+ else {
+ // In any other case, convert both IVInit and IVCount to integers before
+ // comparing. This may result in SCEV expension of pointers, but in practice
+ // SCEV will fold the pointer arithmetic away as such:
+ // 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.
+
+ const SCEV *IVLimit = 0;
+ // For unit stride, IVCount = Start + BECount with 2's complement overflow.
+ // For non-zero Start, compute IVCount here.
+ if (AR->getStart()->isZero())
+ IVLimit = IVCount;
+ else {
+ assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride");
+ const SCEV *IVInit = AR->getStart();
+
+ // For integer IVs, truncate the IV before computing IVInit + BECount.
+ if (SE->getTypeSizeInBits(IVInit->getType())
+ > SE->getTypeSizeInBits(IVCount->getType()))
+ IVInit = SE->getTruncateExpr(IVInit, IVCount->getType());
+
+ IVLimit = SE->getAddExpr(IVInit, IVCount);
+ }
+ // Expand the code for the iteration count.
+ BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
+ IRBuilder<> Builder(BI);
+ assert(SE->isLoopInvariant(IVLimit, L) &&
+ "Computed iteration count is not loop invariant!");
+ // Ensure that we generate the same type as IndVar, or a smaller integer
+ // type. In the presence of null pointer values, we have an integer type
+ // SCEV expression (IVInit) for a pointer type IV value (IndVar).
+ Type *LimitTy = IVCount->getType()->isPointerTy() ?
+ IndVar->getType() : IVCount->getType();
+ return Rewriter.expandCodeFor(IVLimit, LimitTy, BI);
+ }
+}
- // Delete old, floating point, exit comparision instruction.
- EC->replaceAllUsesWith(NewEC);
- RecursivelyDeleteTriviallyDeadInstructions(EC);
+/// LinearFunctionTestReplace - This method rewrites the exit condition of the
+/// loop to be a canonical != comparison against the incremented loop induction
+/// variable. This pass is able to rewrite the exit tests of any loop where the
+/// SCEV analysis can determine a loop-invariant trip count of the loop, which
+/// is actually a much broader range than just linear tests.
+Value *IndVarSimplify::
+LinearFunctionTestReplace(Loop *L,
+ const SCEV *BackedgeTakenCount,
+ PHINode *IndVar,
+ 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();
+
+ 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));
+ }
+ }
+ // 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;
+ }
- // Delete old, floating point, increment instruction.
- Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
- RecursivelyDeleteTriviallyDeadInstructions(Incr);
-
- // Replace floating induction variable, if it isn't already deleted.
- // Give SIToFPInst preference over UIToFPInst because it is faster on
- // platforms that are widely used.
- if (WeakPH && !PH->use_empty()) {
- if (useSIToFPInst(*InitValue, *EV, newInitValue, intEV)) {
- SIToFPInst *Conv = new SIToFPInst(NewPHI, PH->getType(), "indvar.conv",
- PH->getParent()->getFirstNonPHI());
- PH->replaceAllUsesWith(Conv);
+ Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE);
+ assert(ExitCnt->getType()->isPointerTy() == IndVar->getType()->isPointerTy()
+ && "genLoopLimit missed a cast");
+
+ // Insert a new icmp_ne or icmp_eq instruction before the branch.
+ BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator());
+ ICmpInst::Predicate P;
+ if (L->contains(BI->getSuccessor(0)))
+ P = ICmpInst::ICMP_NE;
+ else
+ P = ICmpInst::ICMP_EQ;
+
+ DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n"
+ << " LHS:" << *CmpIndVar << '\n'
+ << " op:\t"
+ << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
+ << " RHS:\t" << *ExitCnt << "\n"
+ << " IVCount:\t" << *IVCount << "\n");
+
+ IRBuilder<> Builder(BI);
+ if (SE->getTypeSizeInBits(CmpIndVar->getType())
+ > SE->getTypeSizeInBits(ExitCnt->getType())) {
+ 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
+ // 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);
+ DeadInsts.push_back(OrigCond);
+
+ ++NumLFTR;
+ Changed = true;
+ return Cond;
+}
+
+//===----------------------------------------------------------------------===//
+// SinkUnusedInvariants. A late subpass to cleanup loop preheaders.
+//===----------------------------------------------------------------------===//
+
+/// If there's a single exit block, sink any loop-invariant values that
+/// were defined in the preheader but not used inside the loop into the
+/// exit block to reduce register pressure in the loop.
+void IndVarSimplify::SinkUnusedInvariants(Loop *L) {
+ BasicBlock *ExitBlock = L->getExitBlock();
+ if (!ExitBlock) return;
+
+ BasicBlock *Preheader = L->getLoopPreheader();
+ if (!Preheader) return;
+
+ Instruction *InsertPt = ExitBlock->getFirstInsertionPt();
+ BasicBlock::iterator I = Preheader->getTerminator();
+ while (I != Preheader->begin()) {
+ --I;
+ // New instructions were inserted at the end of the preheader.
+ if (isa<PHINode>(I))
+ break;
+
+ // Don't move instructions which might have side effects, since the side
+ // effects need to complete before instructions inside the loop. Also don't
+ // move instructions which might read memory, since the loop may modify
+ // memory. Note that it's okay if the instruction might have undefined
+ // behavior: LoopSimplify guarantees that the preheader dominates the exit
+ // block.
+ if (I->mayHaveSideEffects() || I->mayReadFromMemory())
+ continue;
+
+ // Skip debug info intrinsics.
+ if (isa<DbgInfoIntrinsic>(I))
+ continue;
+
+ // Skip landingpad instructions.
+ if (isa<LandingPadInst>(I))
+ continue;
+
+ // Don't sink alloca: we never want to sink static alloca's out of the
+ // entry block, and correctly sinking dynamic alloca's requires
+ // checks for stacksave/stackrestore intrinsics.
+ // FIXME: Refactor this check somehow?
+ if (isa<AllocaInst>(I))
+ continue;
+
+ // Determine if there is a use in or before the loop (direct or
+ // otherwise).
+ bool UsedInLoop = false;
+ for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
+ UI != UE; ++UI) {
+ User *U = *UI;
+ BasicBlock *UseBB = cast<Instruction>(U)->getParent();
+ if (PHINode *P = dyn_cast<PHINode>(U)) {
+ unsigned i =
+ PHINode::getIncomingValueNumForOperand(UI.getOperandNo());
+ UseBB = P->getIncomingBlock(i);
+ }
+ if (UseBB == Preheader || L->contains(UseBB)) {
+ UsedInLoop = true;
+ break;
+ }
+ }
+
+ // If there is, the def must remain in the preheader.
+ if (UsedInLoop)
+ continue;
+
+ // Otherwise, sink it to the exit block.
+ Instruction *ToMove = I;
+ bool Done = false;
+
+ if (I != Preheader->begin()) {
+ // Skip debug info intrinsics.
+ do {
+ --I;
+ } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin());
+
+ if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin())
+ Done = true;
} else {
- UIToFPInst *Conv = new UIToFPInst(NewPHI, PH->getType(), "indvar.conv",
- PH->getParent()->getFirstNonPHI());
- PH->replaceAllUsesWith(Conv);
+ Done = true;
+ }
+
+ ToMove->moveBefore(InsertPt);
+ if (Done) break;
+ InsertPt = ToMove;
+ }
+}
+
+//===----------------------------------------------------------------------===//
+// IndVarSimplify driver. Manage several subpasses of IV simplification.
+//===----------------------------------------------------------------------===//
+
+bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
+ // 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"
+ // afterwards.
+ // - We depend on having a preheader; in particular,
+ // Loop::getCanonicalInductionVariable only supports loops with preheaders,
+ // and we're in trouble if we can't find the induction variable even when
+ // we've manually inserted one.
+ if (!L->isLoopSimplifyForm())
+ return false;
+
+ LI = &getAnalysis<LoopInfo>();
+ SE = &getAnalysis<ScalarEvolution>();
+ DT = &getAnalysis<DominatorTree>();
+ TD = getAnalysisIfAvailable<DataLayout>();
+ TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
+
+ DeadInsts.clear();
+ Changed = false;
+
+ // If there are any floating-point recurrences, attempt to
+ // transform them to use integer recurrences.
+ RewriteNonIntegerIVs(L);
+
+ const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L);
+
+ // Create a rewriter object which we'll use to transform the code with.
+ SCEVExpander Rewriter(*SE, "indvars");
+#ifndef NDEBUG
+ Rewriter.setDebugType(DEBUG_TYPE);
+#endif
+
+ // Eliminate redundant IV users.
+ //
+ // Simplification works best when run before other consumers of SCEV. We
+ // attempt to avoid evaluating SCEVs for sign/zero extend operations until
+ // other expressions involving loop IVs have been evaluated. This helps SCEV
+ // set no-wrap flags before normalizing sign/zero extension.
+ Rewriter.disableCanonicalMode();
+ SimplifyAndExtend(L, Rewriter, LPM);
+
+ // 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
+ // that are recurrent in the loop, and substitute the exit values from the
+ // loop into any instructions outside of the loop that use the final values of
+ // the current expressions.
+ //
+ if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
+ RewriteLoopExitValues(L, Rewriter);
+
+ // Eliminate redundant IV cycles.
+ NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts);
+
+ // 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);
+ 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.
+ //
+ // FIXME: SCEV expansion has no way to bail out, so the caller must
+ // explicitly check any assumptions made by SCEV. Brittle.
+ const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount);
+ if (!AR || AR->getLoop()->getLoopPreheader())
+ (void)LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
+ Rewriter);
}
- RecursivelyDeleteTriviallyDeadInstructions(PH);
}
+ // 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 (Instruction *Inst =
+ dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()))
+ RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI);
+
+ // 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);
+
+ // Clean up dead instructions.
+ Changed |= DeleteDeadPHIs(L->getHeader(), TLI);
+ // Check a post-condition.
+ assert(L->isLCSSAForm(*DT) &&
+ "Indvars did not leave the loop in lcssa form!");
+
+ // Verify that LFTR, and any other change have not interfered with SCEV's
+ // ability to compute trip count.
+#ifndef NDEBUG
+ if (VerifyIndvars && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
+ SE->forgetLoop(L);
+ const SCEV *NewBECount = SE->getBackedgeTakenCount(L);
+ if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) <
+ SE->getTypeSizeInBits(NewBECount->getType()))
+ NewBECount = SE->getTruncateOrNoop(NewBECount,
+ BackedgeTakenCount->getType());
+ else
+ BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount,
+ NewBECount->getType());
+ assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV");
+ }
+#endif
- // Add a new IVUsers entry for the newly-created integer PHI.
- IU->AddUsersIfInteresting(NewPHI);
+ return Changed;
}