//===- IndVarSimplify.cpp - Induction Variable Elimination ----------------===//
-//
+//
// The LLVM Compiler Infrastructure
//
-// This file was developed by the LLVM research group and is distributed under
-// the University of Illinois Open Source License. See LICENSE.TXT for details.
-//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
//===----------------------------------------------------------------------===//
//
-// Guarantees that all loops with identifiable, linear, induction variables will
-// be transformed to have a single, canonical, induction variable. After this
-// pass runs, it guarantees the the first PHI node of the header block in the
-// loop is the canonical induction variable if there is one.
+// This transformation analyzes and transforms the induction variables (and
+// 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
+// induction value against the exit value. This turns loops like:
+// 'for (i = 7; i*i < 1000; ++i)' into 'for (i = 0; i != 25; ++i)'
+// 2. Any use outside of the loop of an expression derived from the indvar
+// is changed to compute the derived value outside of the loop, eliminating
+// the dependence on the exit value of the induction variable. If the only
+// 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/Transforms/Scalar.h"
-#include "llvm/Analysis/InductionVariable.h"
-#include "llvm/Analysis/LoopInfo.h"
-#include "llvm/iPHINode.h"
-#include "llvm/iOther.h"
-#include "llvm/Type.h"
+#include "llvm/BasicBlock.h"
#include "llvm/Constants.h"
+#include "llvm/Instructions.h"
+#include "llvm/Type.h"
+#include "llvm/Analysis/ScalarEvolutionExpander.h"
+#include "llvm/Analysis/LoopInfo.h"
+#include "llvm/Analysis/LoopPass.h"
#include "llvm/Support/CFG.h"
-#include "Support/Debug.h"
-#include "Support/Statistic.h"
-#include "Support/STLExtras.h"
-#include <algorithm>
-
-namespace llvm {
+#include "llvm/Support/Compiler.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/GetElementPtrTypeIterator.h"
+#include "llvm/Transforms/Utils/Local.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/Statistic.h"
+using namespace llvm;
+
+STATISTIC(NumRemoved , "Number of aux indvars removed");
+STATISTIC(NumPointer , "Number of pointer indvars promoted");
+STATISTIC(NumInserted, "Number of canonical indvars added");
+STATISTIC(NumReplaced, "Number of exit values replaced");
+STATISTIC(NumLFTR , "Number of loop exit tests replaced");
namespace {
- Statistic<> NumRemoved ("indvars", "Number of aux indvars removed");
- Statistic<> NumInserted("indvars", "Number of canonical indvars added");
+ class VISIBILITY_HIDDEN IndVarSimplify : public LoopPass {
+ LoopInfo *LI;
+ ScalarEvolution *SE;
+ bool Changed;
+ public:
+
+ static char ID; // Pass identification, replacement for typeid
+ IndVarSimplify() : LoopPass(&ID) {}
+
+ bool runOnLoop(Loop *L, LPPassManager &LPM);
+ bool doInitialization(Loop *L, LPPassManager &LPM);
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.addRequired<ScalarEvolution>();
+ AU.addRequiredID(LCSSAID);
+ AU.addRequiredID(LoopSimplifyID);
+ AU.addRequired<LoopInfo>();
+ AU.addPreservedID(LoopSimplifyID);
+ AU.addPreservedID(LCSSAID);
+ AU.setPreservesCFG();
+ }
+
+ private:
+
+ void EliminatePointerRecurrence(PHINode *PN, BasicBlock *Preheader,
+ std::set<Instruction*> &DeadInsts);
+ Instruction *LinearFunctionTestReplace(Loop *L, SCEV *IterationCount,
+ SCEVExpander &RW);
+ void RewriteLoopExitValues(Loop *L, SCEV *IterationCount);
+
+ void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
+
+ void OptimizeCanonicalIVType(Loop *L);
+ };
}
-// InsertCast - Cast Val to Ty, setting a useful name on the cast if Val has a
-// name...
-//
-static Instruction *InsertCast(Value *Val, const Type *Ty,
- Instruction *InsertBefore) {
- return new CastInst(Val, Ty, Val->getName()+"-casted", InsertBefore);
+char IndVarSimplify::ID = 0;
+static RegisterPass<IndVarSimplify>
+X("indvars", "Canonicalize Induction Variables");
+
+LoopPass *llvm::createIndVarSimplifyPass() {
+ return new IndVarSimplify();
}
-static bool TransformLoop(LoopInfo *Loops, Loop *Loop) {
- // Transform all subloops before this loop...
- bool Changed = reduce_apply_bool(Loop->getSubLoops().begin(),
- Loop->getSubLoops().end(),
- std::bind1st(std::ptr_fun(TransformLoop), Loops));
- // Get the header node for this loop. All of the phi nodes that could be
- // induction variables must live in this basic block.
- //
- BasicBlock *Header = Loop->getHeader();
-
- // Loop over all of the PHI nodes in the basic block, calculating the
- // induction variables that they represent... stuffing the induction variable
- // info into a vector...
- //
- std::vector<InductionVariable> IndVars; // Induction variables for block
- BasicBlock::iterator AfterPHIIt = Header->begin();
- for (; PHINode *PN = dyn_cast<PHINode>(AfterPHIIt); ++AfterPHIIt)
- IndVars.push_back(InductionVariable(PN, Loops));
- // AfterPHIIt now points to first non-phi instruction...
-
- // If there are no phi nodes in this basic block, there can't be indvars...
- if (IndVars.empty()) return Changed;
-
- // Loop over the induction variables, looking for a canonical induction
- // variable, and checking to make sure they are not all unknown induction
- // variables.
- //
- bool FoundIndVars = false;
- InductionVariable *Canonical = 0;
- for (unsigned i = 0; i < IndVars.size(); ++i) {
- if (IndVars[i].InductionType == InductionVariable::Canonical &&
- !isa<PointerType>(IndVars[i].Phi->getType()))
- Canonical = &IndVars[i];
- if (IndVars[i].InductionType != InductionVariable::Unknown)
- FoundIndVars = true;
+/// DeleteTriviallyDeadInstructions - If any of the instructions is the
+/// specified set are trivially dead, delete them and see if this makes any of
+/// their operands subsequently dead.
+void IndVarSimplify::
+DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts) {
+ while (!Insts.empty()) {
+ Instruction *I = *Insts.begin();
+ Insts.erase(Insts.begin());
+ if (isInstructionTriviallyDead(I)) {
+ for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
+ if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
+ Insts.insert(U);
+ SE->deleteValueFromRecords(I);
+ DOUT << "INDVARS: Deleting: " << *I;
+ I->eraseFromParent();
+ Changed = true;
+ }
}
+}
+
+
+/// EliminatePointerRecurrence - Check to see if this is a trivial GEP pointer
+/// recurrence. If so, change it into an integer recurrence, permitting
+/// analysis by the SCEV routines.
+void IndVarSimplify::EliminatePointerRecurrence(PHINode *PN,
+ BasicBlock *Preheader,
+ std::set<Instruction*> &DeadInsts) {
+ assert(PN->getNumIncomingValues() == 2 && "Noncanonicalized loop!");
+ unsigned PreheaderIdx = PN->getBasicBlockIndex(Preheader);
+ unsigned BackedgeIdx = PreheaderIdx^1;
+ if (GetElementPtrInst *GEPI =
+ dyn_cast<GetElementPtrInst>(PN->getIncomingValue(BackedgeIdx)))
+ if (GEPI->getOperand(0) == PN) {
+ assert(GEPI->getNumOperands() == 2 && "GEP types must match!");
+ DOUT << "INDVARS: Eliminating pointer recurrence: " << *GEPI;
+
+ // Okay, we found a pointer recurrence. Transform this pointer
+ // recurrence into an integer recurrence. Compute the value that gets
+ // added to the pointer at every iteration.
+ Value *AddedVal = GEPI->getOperand(1);
+
+ // Insert a new integer PHI node into the top of the block.
+ PHINode *NewPhi = PHINode::Create(AddedVal->getType(),
+ PN->getName()+".rec", PN);
+ NewPhi->addIncoming(Constant::getNullValue(NewPhi->getType()), Preheader);
+
+ // Create the new add instruction.
+ Value *NewAdd = BinaryOperator::CreateAdd(NewPhi, AddedVal,
+ GEPI->getName()+".rec", GEPI);
+ NewPhi->addIncoming(NewAdd, PN->getIncomingBlock(BackedgeIdx));
+
+ // Update the existing GEP to use the recurrence.
+ GEPI->setOperand(0, PN->getIncomingValue(PreheaderIdx));
+
+ // Update the GEP to use the new recurrence we just inserted.
+ GEPI->setOperand(1, NewAdd);
+
+ // If the incoming value is a constant expr GEP, try peeling out the array
+ // 0 index if possible to make things simpler.
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEPI->getOperand(0)))
+ if (CE->getOpcode() == Instruction::GetElementPtr) {
+ unsigned NumOps = CE->getNumOperands();
+ assert(NumOps > 1 && "CE folding didn't work!");
+ if (CE->getOperand(NumOps-1)->isNullValue()) {
+ // Check to make sure the last index really is an array index.
+ gep_type_iterator GTI = gep_type_begin(CE);
+ for (unsigned i = 1, e = CE->getNumOperands()-1;
+ i != e; ++i, ++GTI)
+ /*empty*/;
+ if (isa<SequentialType>(*GTI)) {
+ // Pull the last index out of the constant expr GEP.
+ SmallVector<Value*, 8> CEIdxs(CE->op_begin()+1, CE->op_end()-1);
+ Constant *NCE = ConstantExpr::getGetElementPtr(CE->getOperand(0),
+ &CEIdxs[0],
+ CEIdxs.size());
+ Value *Idx[2];
+ Idx[0] = Constant::getNullValue(Type::Int32Ty);
+ Idx[1] = NewAdd;
+ GetElementPtrInst *NGEPI = GetElementPtrInst::Create(
+ NCE, Idx, Idx + 2,
+ GEPI->getName(), GEPI);
+ SE->deleteValueFromRecords(GEPI);
+ GEPI->replaceAllUsesWith(NGEPI);
+ GEPI->eraseFromParent();
+ GEPI = NGEPI;
+ }
+ }
+ }
+
+
+ // Finally, if there are any other users of the PHI node, we must
+ // insert a new GEP instruction that uses the pre-incremented version
+ // of the induction amount.
+ if (!PN->use_empty()) {
+ BasicBlock::iterator InsertPos = PN; ++InsertPos;
+ while (isa<PHINode>(InsertPos)) ++InsertPos;
+ Value *PreInc =
+ GetElementPtrInst::Create(PN->getIncomingValue(PreheaderIdx),
+ NewPhi, "", InsertPos);
+ PreInc->takeName(PN);
+ PN->replaceAllUsesWith(PreInc);
+ }
+
+ // Delete the old PHI for sure, and the GEP if its otherwise unused.
+ DeadInsts.insert(PN);
- // No induction variables, bail early... don't add a canonical indvar
- if (!FoundIndVars) return Changed;
-
- // Okay, we want to convert other induction variables to use a canonical
- // indvar. If we don't have one, add one now...
- if (!Canonical) {
- // Create the PHI node for the new induction variable, and insert the phi
- // node at the start of the PHI nodes...
- PHINode *PN = new PHINode(Type::UIntTy, "cann-indvar", Header->begin());
-
- // Create the increment instruction to add one to the counter...
- Instruction *Add = BinaryOperator::create(Instruction::Add, PN,
- ConstantUInt::get(Type::UIntTy,1),
- "add1-indvar", AfterPHIIt);
-
- // Figure out which block is incoming and which is the backedge for the loop
- BasicBlock *Incoming, *BackEdgeBlock;
- pred_iterator PI = pred_begin(Header);
- assert(PI != pred_end(Header) && "Loop headers should have 2 preds!");
- if (Loop->contains(*PI)) { // First pred is back edge...
- BackEdgeBlock = *PI++;
- Incoming = *PI++;
- } else {
- Incoming = *PI++;
- BackEdgeBlock = *PI++;
+ ++NumPointer;
+ Changed = true;
}
- assert(PI == pred_end(Header) && "Loop headers should have 2 preds!");
-
- // Add incoming values for the PHI node...
- PN->addIncoming(Constant::getNullValue(Type::UIntTy), Incoming);
- PN->addIncoming(Add, BackEdgeBlock);
-
- // Analyze the new induction variable...
- IndVars.push_back(InductionVariable(PN, Loops));
- assert(IndVars.back().InductionType == InductionVariable::Canonical &&
- "Just inserted canonical indvar that is not canonical!");
- Canonical = &IndVars.back();
- ++NumInserted;
- Changed = true;
+}
+
+/// 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.
+///
+/// This method returns a "potentially dead" instruction whose computation chain
+/// should be deleted when convenient.
+Instruction *IndVarSimplify::LinearFunctionTestReplace(Loop *L,
+ SCEV *IterationCount,
+ SCEVExpander &RW) {
+ // Find the exit block for the loop. We can currently only handle loops with
+ // a single exit.
+ SmallVector<BasicBlock*, 8> ExitBlocks;
+ L->getExitBlocks(ExitBlocks);
+ if (ExitBlocks.size() != 1) return 0;
+ BasicBlock *ExitBlock = ExitBlocks[0];
+
+ // Make sure there is only one predecessor block in the loop.
+ BasicBlock *ExitingBlock = 0;
+ for (pred_iterator PI = pred_begin(ExitBlock), PE = pred_end(ExitBlock);
+ PI != PE; ++PI)
+ if (L->contains(*PI)) {
+ if (ExitingBlock == 0)
+ ExitingBlock = *PI;
+ else
+ return 0; // Multiple exits from loop to this block.
+ }
+ assert(ExitingBlock && "Loop info is broken");
+
+ if (!isa<BranchInst>(ExitingBlock->getTerminator()))
+ return 0; // Can't rewrite non-branch yet
+ BranchInst *BI = cast<BranchInst>(ExitingBlock->getTerminator());
+ assert(BI->isConditional() && "Must be conditional to be part of loop!");
+
+ Instruction *PotentiallyDeadInst = dyn_cast<Instruction>(BI->getCondition());
+
+ // 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.
+ BasicBlock *Header = L->getHeader();
+ pred_iterator HPI = pred_begin(Header);
+ assert(HPI != pred_end(Header) && "Loop with zero preds???");
+ if (!L->contains(*HPI)) ++HPI;
+ assert(HPI != pred_end(Header) && L->contains(*HPI) &&
+ "No backedge in loop?");
+
+ SCEVHandle TripCount = IterationCount;
+ Value *IndVar;
+ if (*HPI == ExitingBlock) {
+ // The IterationCount expression contains the number of times that the
+ // backedge actually branches to the loop header. This is one less than the
+ // number of times the loop executes, so add one to it.
+ ConstantInt *OneC = ConstantInt::get(IterationCount->getType(), 1);
+ TripCount = SE->getAddExpr(IterationCount, SE->getConstant(OneC));
+ IndVar = L->getCanonicalInductionVariableIncrement();
} else {
- // If we have a canonical induction variable, make sure that it is the first
- // one in the basic block.
- if (&Header->front() != Canonical->Phi)
- Header->getInstList().splice(Header->begin(), Header->getInstList(),
- Canonical->Phi);
+ // We have to use the preincremented value...
+ IndVar = L->getCanonicalInductionVariable();
}
+
+ DOUT << "INDVARS: LFTR: TripCount = " << *TripCount
+ << " IndVar = " << *IndVar << "\n";
+
+ // Expand the code for the iteration count into the preheader of the loop.
+ BasicBlock *Preheader = L->getLoopPreheader();
+ Value *ExitCnt = RW.expandCodeFor(TripCount, Preheader->getTerminator());
+
+ // 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;
+
+ Value *Cond = new ICmpInst(Opcode, IndVar, ExitCnt, "exitcond", BI);
+ BI->setCondition(Cond);
+ ++NumLFTR;
+ Changed = true;
+ return PotentiallyDeadInst;
+}
- DEBUG(std::cerr << "Induction variables:\n");
- // Get the current loop iteration count, which is always the value of the
- // canonical phi node...
- //
- PHINode *IterCount = Canonical->Phi;
+/// 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
+/// substitute the exit values from the loop into any instructions outside of
+/// the loop that use the final values of the current expressions.
+void IndVarSimplify::RewriteLoopExitValues(Loop *L, SCEV *IterationCount) {
+ 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, *LI);
+
+ // 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();
+
+ bool HasConstantItCount = isa<SCEVConstant>(IterationCount);
+
+ std::set<Instruction*> InstructionsToDelete;
+ 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
+ // the exit blocks of the loop to find them.
+ for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
+ BasicBlock *ExitBB = ExitBlocks[i];
+
+ // If there are no PHI nodes in this exit block, then no values defined
+ // inside the loop are used on this path, skip it.
+ PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
+ if (!PN) continue;
+
+ unsigned NumPreds = PN->getNumIncomingValues();
+
+ // Iterate over all of the PHI nodes.
+ BasicBlock::iterator BBI = ExitBB->begin();
+ while ((PN = dyn_cast<PHINode>(BBI++))) {
+
+ // 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()))
+ continue;
+
+ // If this pred is for a subloop, not L itself, skip it.
+ if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
+ continue; // The Block is in a subloop, skip it.
+
+ // Check that InVal is defined in the loop.
+ Instruction *Inst = cast<Instruction>(InVal);
+ if (!L->contains(Inst->getParent()))
+ continue;
+
+ // We require that this value either have a computable evolution or that
+ // the loop have a constant iteration count. In the case where the loop
+ // has a constant iteration count, we can sometimes force evaluation of
+ // the exit value through brute force.
+ SCEVHandle SH = SE->getSCEV(Inst);
+ if (!SH->hasComputableLoopEvolution(L) && !HasConstantItCount)
+ continue; // Cannot get exit evolution for the loop value.
+
+ // 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 ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
+ if (isa<SCEVCouldNotCompute>(ExitValue) ||
+ !ExitValue->isLoopInvariant(L))
+ continue;
+
+ Changed = true;
+ ++NumReplaced;
+
+ // See if we already computed the exit value for the instruction, if so,
+ // just reuse it.
+ Value *&ExitVal = ExitValues[Inst];
+ if (!ExitVal)
+ ExitVal = Rewriter.expandCodeFor(ExitValue, InsertPt);
+
+ DOUT << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal
+ << " LoopVal = " << *Inst << "\n";
+
+ PN->setIncomingValue(i, ExitVal);
+
+ // If this instruction is dead now, schedule it to be removed.
+ if (Inst->use_empty())
+ InstructionsToDelete.insert(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) {
+ SE->deleteValueFromRecords(PN);
+ PN->replaceAllUsesWith(ExitVal);
+ PN->eraseFromParent();
+ break;
+ }
+ }
+ }
+ }
+
+ DeleteTriviallyDeadInstructions(InstructionsToDelete);
+}
+
+bool IndVarSimplify::doInitialization(Loop *L, LPPassManager &LPM) {
- // Loop through and replace all of the auxiliary induction variables with
- // references to the canonical induction variable...
+ Changed = false;
+ // First step. Check to see if there are any trivial GEP pointer recurrences.
+ // If there are, change them into integer recurrences, permitting analysis by
+ // the SCEV routines.
//
- for (unsigned i = 0; i < IndVars.size(); ++i) {
- InductionVariable *IV = &IndVars[i];
+ BasicBlock *Header = L->getHeader();
+ BasicBlock *Preheader = L->getLoopPreheader();
+ SE = &LPM.getAnalysis<ScalarEvolution>();
+
+ std::set<Instruction*> DeadInsts;
+ for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
+ PHINode *PN = cast<PHINode>(I);
+ if (isa<PointerType>(PN->getType()))
+ EliminatePointerRecurrence(PN, Preheader, DeadInsts);
+ }
- DEBUG(IV->print(std::cerr));
+ if (!DeadInsts.empty())
+ DeleteTriviallyDeadInstructions(DeadInsts);
- // Don't do math with pointers...
- const Type *IVTy = IV->Phi->getType();
- if (isa<PointerType>(IVTy)) IVTy = Type::ULongTy;
+ return Changed;
+}
- // Don't modify the canonical indvar or unrecognized indvars...
- if (IV != Canonical && IV->InductionType != InductionVariable::Unknown) {
- Instruction *Val = IterCount;
- if (!isa<ConstantInt>(IV->Step) || // If the step != 1
- !cast<ConstantInt>(IV->Step)->equalsInt(1)) {
+bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
- // If the types are not compatible, insert a cast now...
- if (Val->getType() != IVTy)
- Val = InsertCast(Val, IVTy, AfterPHIIt);
- if (IV->Step->getType() != IVTy)
- IV->Step = InsertCast(IV->Step, IVTy, AfterPHIIt);
- Val = BinaryOperator::create(Instruction::Mul, Val, IV->Step,
- IV->Phi->getName()+"-scale", AfterPHIIt);
- }
+ LI = &getAnalysis<LoopInfo>();
+ SE = &getAnalysis<ScalarEvolution>();
+
+ Changed = false;
+ BasicBlock *Header = L->getHeader();
+ std::set<Instruction*> DeadInsts;
+
+ // Verify the input to the pass in already in LCSSA form.
+ assert(L->isLCSSAForm());
+
+ // 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.
+ //
+ SCEVHandle IterationCount = SE->getIterationCount(L);
+ if (!isa<SCEVCouldNotCompute>(IterationCount))
+ RewriteLoopExitValues(L, IterationCount);
+
+ // Next, analyze all of the induction variables in the loop, canonicalizing
+ // auxillary induction variables.
+ std::vector<std::pair<PHINode*, SCEVHandle> > IndVars;
+
+ for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
+ PHINode *PN = cast<PHINode>(I);
+ if (PN->getType()->isInteger()) { // FIXME: when we have fast-math, enable!
+ SCEVHandle SCEV = SE->getSCEV(PN);
+ if (SCEV->hasComputableLoopEvolution(L))
+ // 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.
+ if (SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SCEV))
+ if (AR->isAffine())
+ IndVars.push_back(std::make_pair(PN, SCEV));
+ }
+ }
- // If the start != 0
- if (IV->Start != Constant::getNullValue(IV->Start->getType())) {
- // If the types are not compatible, insert a cast now...
- if (Val->getType() != IVTy)
- Val = InsertCast(Val, IVTy, AfterPHIIt);
- if (IV->Start->getType() != IVTy)
- IV->Start = InsertCast(IV->Start, IVTy, AfterPHIIt);
-
- // Insert the instruction after the phi nodes...
- Val = BinaryOperator::create(Instruction::Add, Val, IV->Start,
- IV->Phi->getName()+"-offset", AfterPHIIt);
+ // If there are no induction variables in the loop, there is nothing more to
+ // do.
+ if (IndVars.empty()) {
+ // Actually, if we know how many times the loop iterates, lets insert a
+ // canonical induction variable to help subsequent passes.
+ if (!isa<SCEVCouldNotCompute>(IterationCount)) {
+ SCEVExpander Rewriter(*SE, *LI);
+ Rewriter.getOrInsertCanonicalInductionVariable(L,
+ IterationCount->getType());
+ if (Instruction *I = LinearFunctionTestReplace(L, IterationCount,
+ Rewriter)) {
+ std::set<Instruction*> InstructionsToDelete;
+ InstructionsToDelete.insert(I);
+ DeleteTriviallyDeadInstructions(InstructionsToDelete);
}
+ }
+ return Changed;
+ }
- // If the PHI node has a different type than val is, insert a cast now...
- if (Val->getType() != IV->Phi->getType())
- Val = InsertCast(Val, IV->Phi->getType(), AfterPHIIt);
-
- // Replace all uses of the old PHI node with the new computed value...
- IV->Phi->replaceAllUsesWith(Val);
-
- // Move the PHI name to it's new equivalent value...
- std::string OldName = IV->Phi->getName();
- IV->Phi->setName("");
- Val->setName(OldName);
-
- // Delete the old, now unused, phi node...
- Header->getInstList().erase(IV->Phi);
-
- // If the PHI is the last user of any instructions for computing PHI nodes
- // that are irrelevant now, delete those instructions.
- while (!PHIOps.empty()) {
- Instruction *MaybeDead = dyn_cast<Instruction>(PHIOps.back());
- PHIOps.pop_back();
-
- if (MaybeDead && isInstructionTriviallyDead(MaybeDead)) {
- PHIOps.insert(PHIOps.end(), MaybeDead->op_begin(),
- MaybeDead->op_end());
- MaybeDead->getParent()->getInstList().erase(MaybeDead);
-
- // Erase any duplicates entries in the PHIOps list.
- std::vector<Value*>::iterator It =
- std::find(PHIOps.begin(), PHIOps.end(), MaybeDead);
- while (It != PHIOps.end()) {
- PHIOps.erase(It);
- It = std::find(PHIOps.begin(), PHIOps.end(), MaybeDead);
- }
+ // Compute the type of the largest recurrence expression.
+ //
+ const Type *LargestType = IndVars[0].first->getType();
+ bool DifferingSizes = false;
+ for (unsigned i = 1, e = IndVars.size(); i != e; ++i) {
+ const Type *Ty = IndVars[i].first->getType();
+ DifferingSizes |=
+ Ty->getPrimitiveSizeInBits() != LargestType->getPrimitiveSizeInBits();
+ if (Ty->getPrimitiveSizeInBits() > LargestType->getPrimitiveSizeInBits())
+ LargestType = Ty;
+ }
- // Erasing the instruction could invalidate the AfterPHI iterator!
- AfterPHIIt = Header->begin();
- }
+ // Create a rewriter object which we'll use to transform the code with.
+ SCEVExpander Rewriter(*SE, *LI);
+
+ // Now that we know the largest of of the induction variables in this loop,
+ // insert a canonical induction variable of the largest size.
+ Value *IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L,LargestType);
+ ++NumInserted;
+ Changed = true;
+ DOUT << "INDVARS: New CanIV: " << *IndVar;
+
+ if (!isa<SCEVCouldNotCompute>(IterationCount)) {
+ IterationCount = SE->getTruncateOrZeroExtend(IterationCount, LargestType);
+ if (Instruction *DI = LinearFunctionTestReplace(L, IterationCount,Rewriter))
+ DeadInsts.insert(DI);
+ }
+
+ // Now that we have a canonical induction variable, we can rewrite any
+ // recurrences in terms of the induction variable. Start with the auxillary
+ // induction variables, and recursively rewrite any of their uses.
+ BasicBlock::iterator InsertPt = Header->getFirstNonPHI();
+
+ // If there were induction variables of other sizes, cast the primary
+ // induction variable to the right size for them, avoiding the need for the
+ // code evaluation methods to insert induction variables of different sizes.
+ if (DifferingSizes) {
+ SmallVector<unsigned,4> InsertedSizes;
+ InsertedSizes.push_back(LargestType->getPrimitiveSizeInBits());
+ for (unsigned i = 0, e = IndVars.size(); i != e; ++i) {
+ unsigned ithSize = IndVars[i].first->getType()->getPrimitiveSizeInBits();
+ if (std::find(InsertedSizes.begin(), InsertedSizes.end(), ithSize)
+ == InsertedSizes.end()) {
+ PHINode *PN = IndVars[i].first;
+ InsertedSizes.push_back(ithSize);
+ Instruction *New = new TruncInst(IndVar, PN->getType(), "indvar",
+ InsertPt);
+ Rewriter.addInsertedValue(New, SE->getSCEV(New));
+ DOUT << "INDVARS: Made trunc IV for " << *PN
+ << " NewVal = " << *New << "\n";
}
+ }
+ }
- Changed = true;
- ++NumRemoved;
+ // Rewrite all induction variables in terms of the canonical induction
+ // variable.
+ while (!IndVars.empty()) {
+ PHINode *PN = IndVars.back().first;
+ Value *NewVal = Rewriter.expandCodeFor(IndVars.back().second, InsertPt);
+ DOUT << "INDVARS: Rewrote IV '" << *IndVars.back().second << "' " << *PN
+ << " into = " << *NewVal << "\n";
+ NewVal->takeName(PN);
+
+ // Replace the old PHI Node with the inserted computation.
+ PN->replaceAllUsesWith(NewVal);
+ DeadInsts.insert(PN);
+ IndVars.pop_back();
+ ++NumRemoved;
+ Changed = true;
+ }
+
+#if 0
+ // Now replace all derived expressions in the loop body with simpler
+ // expressions.
+ for (LoopInfo::block_iterator I = L->block_begin(), E = L->block_end();
+ I != E; ++I) {
+ BasicBlock *BB = *I;
+ if (LI->getLoopFor(BB) == L) { // Not in a subloop...
+ for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
+ if (I->getType()->isInteger() && // Is an integer instruction
+ !I->use_empty() &&
+ !Rewriter.isInsertedInstruction(I)) {
+ SCEVHandle SH = SE->getSCEV(I);
+ Value *V = Rewriter.expandCodeFor(SH, I, I->getType());
+ if (V != I) {
+ if (isa<Instruction>(V))
+ V->takeName(I);
+ I->replaceAllUsesWith(V);
+ DeadInsts.insert(I);
+ ++NumRemoved;
+ Changed = true;
+ }
+ }
}
}
+#endif
+ DeleteTriviallyDeadInstructions(DeadInsts);
+ OptimizeCanonicalIVType(L);
+ assert(L->isLCSSAForm());
return Changed;
}
-namespace {
- struct InductionVariableSimplify : public FunctionPass {
- virtual bool runOnFunction(Function &) {
- LoopInfo &LI = getAnalysis<LoopInfo>();
-
- // Induction Variables live in the header nodes of loops
- return reduce_apply_bool(LI.getTopLevelLoops().begin(),
- LI.getTopLevelLoops().end(),
- std::bind1st(std::ptr_fun(TransformLoop), &LI));
+/// OptimizeCanonicalIVType - If loop induction variable is always
+/// sign or zero extended then extend the type of the induction
+/// variable.
+void IndVarSimplify::OptimizeCanonicalIVType(Loop *L) {
+ PHINode *PH = L->getCanonicalInductionVariable();
+ if (!PH) return;
+
+ // Check loop iteration count.
+ SCEVHandle IC = SE->getIterationCount(L);
+ if (isa<SCEVCouldNotCompute>(IC)) return;
+ SCEVConstant *IterationCount = dyn_cast<SCEVConstant>(IC);
+ if (!IterationCount) return;
+
+ unsigned IncomingEdge = L->contains(PH->getIncomingBlock(0));
+ unsigned BackEdge = IncomingEdge^1;
+
+ // Check IV uses. If all IV uses are either SEXT or ZEXT (except
+ // IV increment instruction) then this IV is suitable for this
+ // transformation.
+ bool isSEXT = false;
+ BinaryOperator *Incr = NULL;
+ const Type *NewType = NULL;
+ for(Value::use_iterator UI = PH->use_begin(), UE = PH->use_end();
+ UI != UE; ++UI) {
+ const Type *CandidateType = NULL;
+ if (ZExtInst *ZI = dyn_cast<ZExtInst>(UI))
+ CandidateType = ZI->getDestTy();
+ else if (SExtInst *SI = dyn_cast<SExtInst>(UI)) {
+ CandidateType = SI->getDestTy();
+ isSEXT = true;
}
-
- virtual void getAnalysisUsage(AnalysisUsage &AU) const {
- AU.addRequired<LoopInfo>();
- AU.addRequiredID(LoopSimplifyID);
- AU.setPreservesCFG();
+ else if ((Incr = dyn_cast<BinaryOperator>(UI))) {
+ // Validate IV increment instruction.
+ if (PH->getIncomingValue(BackEdge) == Incr)
+ continue;
}
- };
- RegisterOpt<InductionVariableSimplify> X("indvars",
- "Canonicalize Induction Variables");
-}
+ if (!CandidateType) {
+ NewType = NULL;
+ break;
+ }
+ if (!NewType)
+ NewType = CandidateType;
+ else if (NewType != CandidateType) {
+ NewType = NULL;
+ break;
+ }
+ }
-Pass *createIndVarSimplifyPass() {
- return new InductionVariableSimplify();
+ // IV uses are not suitable then avoid this transformation.
+ if (!NewType || !Incr)
+ return;
+
+ // IV increment instruction has two uses, one is loop exit condition
+ // and second is the IV (phi node) itself.
+ ICmpInst *Exit = NULL;
+ for(Value::use_iterator II = Incr->use_begin(), IE = Incr->use_end();
+ II != IE; ++II) {
+ if (PH == *II) continue;
+ Exit = dyn_cast<ICmpInst>(*II);
+ break;
+ }
+ if (!Exit) return;
+ ConstantInt *EV = dyn_cast<ConstantInt>(Exit->getOperand(0));
+ if (!EV)
+ EV = dyn_cast<ConstantInt>(Exit->getOperand(1));
+ if (!EV) return;
+
+ // Check iteration count max value to avoid loops that wrap around IV.
+ APInt ICount = IterationCount->getValue()->getValue();
+ if (ICount.isNegative()) return;
+ uint32_t BW = PH->getType()->getPrimitiveSizeInBits();
+ APInt Max = (isSEXT ? APInt::getSignedMaxValue(BW) : APInt::getMaxValue(BW));
+ if (ICount.getZExtValue() > Max.getZExtValue()) return;
+
+ // Extend IV type.
+
+ SCEVExpander Rewriter(*SE, *LI);
+ Value *NewIV = Rewriter.getOrInsertCanonicalInductionVariable(L,NewType);
+ PHINode *NewPH = cast<PHINode>(NewIV);
+ Instruction *NewIncr = cast<Instruction>(NewPH->getIncomingValue(BackEdge));
+
+ // Replace all SEXT or ZEXT uses.
+ SmallVector<Instruction *, 4> PHUses;
+ for(Value::use_iterator UI = PH->use_begin(), UE = PH->use_end();
+ UI != UE; ++UI) {
+ Instruction *I = cast<Instruction>(UI);
+ PHUses.push_back(I);
+ }
+ while (!PHUses.empty()){
+ Instruction *Use = PHUses.back(); PHUses.pop_back();
+ if (Incr == Use) continue;
+
+ SE->deleteValueFromRecords(Use);
+ Use->replaceAllUsesWith(NewIV);
+ Use->eraseFromParent();
+ }
+
+ // Replace exit condition.
+ ConstantInt *NEV = ConstantInt::get(NewType, EV->getZExtValue());
+ Instruction *NE = new ICmpInst(Exit->getPredicate(),
+ NewIncr, NEV, "new.exit",
+ Exit->getParent()->getTerminator());
+ SE->deleteValueFromRecords(Exit);
+ Exit->replaceAllUsesWith(NE);
+ Exit->eraseFromParent();
+
+ // Remove old IV and increment instructions.
+ SE->deleteValueFromRecords(PH);
+ PH->removeIncomingValue((unsigned)0);
+ PH->removeIncomingValue((unsigned)0);
+ SE->deleteValueFromRecords(Incr);
+ Incr->eraseFromParent();
}
-} // End llvm namespace