//===- 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 transformation analyzes and transforms the induction variables (and
// computations derived from them) into simpler forms suitable for subsequent
// analysis and transformation.
//
-// This transformation make the following changes to each loop with an
+// 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.
#include "llvm/Constants.h"
#include "llvm/Instructions.h"
#include "llvm/Type.h"
-#include "llvm/Analysis/ScalarEvolutionExpressions.h"
+#include "llvm/Analysis/ScalarEvolutionExpander.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Support/CFG.h"
+#include "llvm/Support/GetElementPtrTypeIterator.h"
#include "llvm/Transforms/Utils/Local.h"
-#include "Support/CommandLine.h"
-#include "Support/Statistic.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/ADT/Statistic.h"
using namespace llvm;
namespace {
AU.addRequired<ScalarEvolution>();
AU.addRequired<LoopInfo>();
AU.addPreservedID(LoopSimplifyID);
+ AU.addPreservedID(LCSSAID);
AU.setPreservesCFG();
}
private:
void runOnLoop(Loop *L);
void EliminatePointerRecurrence(PHINode *PN, BasicBlock *Preheader,
std::set<Instruction*> &DeadInsts);
- void LinearFunctionTestReplace(Loop *L, SCEV *IterationCount,
- ScalarEvolutionRewriter &RW);
+ Instruction *LinearFunctionTestReplace(Loop *L, SCEV *IterationCount,
+ SCEVExpander &RW);
void RewriteLoopExitValues(Loop *L);
void DeleteTriviallyDeadInstructions(std::set<Instruction*> &Insts);
};
- RegisterOpt<IndVarSimplify> X("indvars", "Canonicalize Induction Variables");
+ RegisterPass<IndVarSimplify> X("indvars", "Canonicalize Induction Variables");
}
-Pass *llvm::createIndVarSimplifyPass() {
+FunctionPass *llvm::createIndVarSimplifyPass() {
return new IndVarSimplify();
}
-
/// 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.
if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
Insts.insert(U);
SE->deleteInstructionFromRecords(I);
- I->getParent()->getInstList().erase(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,
+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)))
+ dyn_cast<GetElementPtrInst>(PN->getIncomingValue(BackedgeIdx)))
if (GEPI->getOperand(0) == PN) {
- assert(GEPI->getNumOperands() == 2 && "GEP types must mismatch!");
-
+ assert(GEPI->getNumOperands() == 2 && "GEP types must match!");
+
// 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.
// Insert a new integer PHI node into the top of the block.
PHINode *NewPhi = new PHINode(AddedVal->getType(),
PN->getName()+".rec", PN);
- NewPhi->addIncoming(Constant::getNullValue(NewPhi->getType()),
- Preheader);
+ NewPhi->addIncoming(Constant::getNullValue(NewPhi->getType()), Preheader);
+
// Create the new add instruction.
- Value *NewAdd = BinaryOperator::create(Instruction::Add, NewPhi,
- AddedVal,
- GEPI->getName()+".rec", GEPI);
+ 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.
+ std::vector<Value*> CEIdxs(CE->op_begin()+1, CE->op_end()-1);
+ Constant *NCE = ConstantExpr::getGetElementPtr(CE->getOperand(0),
+ CEIdxs);
+ GetElementPtrInst *NGEPI =
+ new GetElementPtrInst(NCE, Constant::getNullValue(Type::IntTy),
+ NewAdd, GEPI->getName(), 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.
/// 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.
-void IndVarSimplify::LinearFunctionTestReplace(Loop *L, SCEV *IterationCount,
- ScalarEvolutionRewriter &RW) {
+///
+/// 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.
- if (L->getExitBlocks().size() != 1) return;
- BasicBlock *ExitBlock = L->getExitBlocks()[0];
+ std::vector<BasicBlock*> 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;
if (ExitingBlock == 0)
ExitingBlock = *PI;
else
- return; // Multiple exits from loop to this block.
+ return 0; // Multiple exits from loop to this block.
}
assert(ExitingBlock && "Loop info is broken");
if (!isa<BranchInst>(ExitingBlock->getTerminator()))
- return; // Can't rewrite non-branch yet
+ 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!");
- std::set<Instruction*> InstructionsToDelete;
- if (Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()))
- InstructionsToDelete.insert(Cond);
-
+ 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.
// Expand the code for the iteration count into the preheader of the loop.
BasicBlock *Preheader = L->getLoopPreheader();
- Value *ExitCnt = RW.ExpandCodeFor(TripCount, Preheader->getTerminator(),
+ Value *ExitCnt = RW.expandCodeFor(TripCount, Preheader->getTerminator(),
IndVar->getType());
// Insert a new setne or seteq instruction before the branch.
BI->setCondition(Cond);
++NumLFTR;
Changed = true;
-
- DeleteTriviallyDeadInstructions(InstructionsToDelete);
+ return PotentiallyDeadInst;
}
// Scan all of the instructions in the loop, looking at those that have
// extra-loop users and which are recurrences.
- ScalarEvolutionRewriter Rewriter(*SE, *LI);
+ 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;
- if (L->getExitBlocks().size() == 1)
- BlockToInsertInto = L->getExitBlocks()[0];
+ std::vector<BasicBlock*> ExitBlocks;
+ L->getExitBlocks(ExitBlocks);
+ if (ExitBlocks.size() == 1)
+ BlockToInsertInto = ExitBlocks[0];
else
BlockToInsertInto = Preheader;
BasicBlock::iterator InsertPt = BlockToInsertInto->begin();
while (isa<PHINode>(InsertPt)) ++InsertPt;
+ bool HasConstantItCount = isa<SCEVConstant>(SE->getIterationCount(L));
+
std::set<Instruction*> InstructionsToDelete;
-
+
for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i)
if (LI->getLoopFor(L->getBlocks()[i]) == L) { // Not in a subloop...
BasicBlock *BB = L->getBlocks()[i];
- for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
+ for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E;) {
if (I->getType()->isInteger()) { // Is an integer instruction
SCEVHandle SH = SE->getSCEV(I);
- if (SH->hasComputableLoopEvolution(L)) { // Varies predictably
+ if (SH->hasComputableLoopEvolution(L) || // Varies predictably
+ HasConstantItCount) {
// Find out if this predictably varying value is actually used
// outside of the loop. "extra" as opposed to "intra".
- std::vector<User*> ExtraLoopUsers;
+ std::vector<Instruction*> ExtraLoopUsers;
for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
- UI != E; ++UI)
- if (!L->contains(cast<Instruction>(*UI)->getParent()))
- ExtraLoopUsers.push_back(*UI);
+ UI != E; ++UI) {
+ Instruction *User = cast<Instruction>(*UI);
+ if (!L->contains(User->getParent())) {
+ // If this is a PHI node in the exit block and we're inserting,
+ // into the exit block, it must have a single entry. In this
+ // case, we can't insert the code after the PHI and have the PHI
+ // still use it. Instead, don't insert the the PHI.
+ if (PHINode *PN = dyn_cast<PHINode>(User)) {
+ // FIXME: This is a case where LCSSA pessimizes code, this
+ // should be fixed better.
+ if (PN->getNumOperands() == 2 &&
+ PN->getParent() == BlockToInsertInto)
+ continue;
+ }
+ ExtraLoopUsers.push_back(User);
+ }
+ }
+
if (!ExtraLoopUsers.empty()) {
// Okay, this instruction has a user outside of the current loop
// and varies predictably in this loop. Evaluate the value it
// contains when the loop exits, and insert code for it.
- SCEVHandle ExitValue = SE->getSCEVAtScope(I,L->getParentLoop());
+ SCEVHandle ExitValue = SE->getSCEVAtScope(I, L->getParentLoop());
if (!isa<SCEVCouldNotCompute>(ExitValue)) {
Changed = true;
++NumReplaced;
- Value *NewVal = Rewriter.ExpandCodeFor(ExitValue, InsertPt,
+ // Remember the next instruction. The rewriter can move code
+ // around in some cases.
+ BasicBlock::iterator NextI = I; ++NextI;
+
+ Value *NewVal = Rewriter.expandCodeFor(ExitValue, InsertPt,
I->getType());
// Rewrite any users of the computed value outside of the loop
// with the newly computed value.
- for (unsigned i = 0, e = ExtraLoopUsers.size(); i != e; ++i)
- ExtraLoopUsers[i]->replaceUsesOfWith(I, NewVal);
+ for (unsigned i = 0, e = ExtraLoopUsers.size(); i != e; ++i) {
+ PHINode* PN = dyn_cast<PHINode>(ExtraLoopUsers[i]);
+ if (PN && PN->getNumOperands() == 2 &&
+ !L->contains(PN->getParent())) {
+ // We're dealing with an LCSSA Phi. Handle it specially.
+ Instruction* LCSSAInsertPt = BlockToInsertInto->begin();
+
+ Instruction* NewInstr = dyn_cast<Instruction>(NewVal);
+ if (NewInstr && !isa<PHINode>(NewInstr) &&
+ !L->contains(NewInstr->getParent()))
+ for (unsigned j = 0; j < NewInstr->getNumOperands(); ++j){
+ Instruction* PredI =
+ dyn_cast<Instruction>(NewInstr->getOperand(j));
+ if (PredI && L->contains(PredI->getParent())) {
+ PHINode* NewLCSSA = new PHINode(PredI->getType(),
+ PredI->getName() + ".lcssa",
+ LCSSAInsertPt);
+ NewLCSSA->addIncoming(PredI,
+ BlockToInsertInto->getSinglePredecessor());
+
+ NewInstr->replaceUsesOfWith(PredI, NewLCSSA);
+ }
+ }
+
+ PN->replaceAllUsesWith(NewVal);
+ PN->eraseFromParent();
+ } else {
+ ExtraLoopUsers[i]->replaceUsesOfWith(I, NewVal);
+ }
+ }
// If this instruction is dead now, schedule it to be removed.
if (I->use_empty())
InstructionsToDelete.insert(I);
+ I = NextI;
+ continue; // Skip the ++I
}
}
}
}
+
+ // Next instruction. Continue instruction skips this.
+ ++I;
+ }
}
DeleteTriviallyDeadInstructions(InstructionsToDelete);
//
BasicBlock *Header = L->getHeader();
BasicBlock *Preheader = L->getLoopPreheader();
-
+
std::set<Instruction*> DeadInsts;
- for (BasicBlock::iterator I = Header->begin();
- PHINode *PN = dyn_cast<PHINode>(I); ++I)
+ for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
+ PHINode *PN = cast<PHINode>(I);
if (isa<PointerType>(PN->getType()))
EliminatePointerRecurrence(PN, Preheader, DeadInsts);
+ }
if (!DeadInsts.empty())
DeleteTriviallyDeadInstructions(DeadInsts);
// auxillary induction variables.
std::vector<std::pair<PHINode*, SCEVHandle> > IndVars;
- for (BasicBlock::iterator I = Header->begin();
- PHINode *PN = dyn_cast<PHINode>(I); ++I)
+ 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))
- if (SE->shouldSubstituteIndVar(SCEV)) // HACK!
- IndVars.push_back(std::make_pair(PN, SCEV));
+ // 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 there are no induction variables in the loop, there is nothing more to
// do.
// Actually, if we know how many times the loop iterates, lets insert a
// canonical induction variable to help subsequent passes.
if (!isa<SCEVCouldNotCompute>(IterationCount)) {
- ScalarEvolutionRewriter Rewriter(*SE, *LI);
- Rewriter.GetOrInsertCanonicalInductionVariable(L,
+ SCEVExpander Rewriter(*SE, *LI);
+ Rewriter.getOrInsertCanonicalInductionVariable(L,
IterationCount->getType());
- LinearFunctionTestReplace(L, IterationCount, Rewriter);
+ if (Instruction *I = LinearFunctionTestReplace(L, IterationCount,
+ Rewriter)) {
+ std::set<Instruction*> InstructionsToDelete;
+ InstructionsToDelete.insert(I);
+ DeleteTriviallyDeadInstructions(InstructionsToDelete);
+ }
}
return;
}
}
// Create a rewriter object which we'll use to transform the code with.
- ScalarEvolutionRewriter Rewriter(*SE, *LI);
+ 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.
LargestType = LargestType->getUnsignedVersion();
- Value *IndVar = Rewriter.GetOrInsertCanonicalInductionVariable(L,LargestType);
+ Value *IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L,LargestType);
++NumInserted;
Changed = true;
if (!isa<SCEVCouldNotCompute>(IterationCount))
- LinearFunctionTestReplace(L, IterationCount, Rewriter);
-
-#if 0
- // 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.
- // FIXME!
- if (DifferingSizes) {
- std::map<unsigned, Value*> InsertedSizes;
- for (unsigned i = 0, e = IndVars.size(); i != e; ++i) {
- }
- }
-#endif
+ 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
BasicBlock::iterator InsertPt = Header->begin();
while (isa<PHINode>(InsertPt)) ++InsertPt;
+ // 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) {
+ bool InsertedSizes[17] = { false };
+ InsertedSizes[LargestType->getPrimitiveSize()] = true;
+ for (unsigned i = 0, e = IndVars.size(); i != e; ++i)
+ if (!InsertedSizes[IndVars[i].first->getType()->getPrimitiveSize()]) {
+ PHINode *PN = IndVars[i].first;
+ InsertedSizes[PN->getType()->getPrimitiveSize()] = true;
+ Instruction *New = new CastInst(IndVar,
+ PN->getType()->getUnsignedVersion(),
+ "indvar", InsertPt);
+ Rewriter.addInsertedValue(New, SE->getSCEV(New));
+ }
+ }
+
+ // 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.
+ std::map<unsigned, Value*> InsertedSizes;
while (!IndVars.empty()) {
PHINode *PN = IndVars.back().first;
- Value *NewVal = Rewriter.ExpandCodeFor(IndVars.back().second, InsertPt,
+ Value *NewVal = Rewriter.expandCodeFor(IndVars.back().second, InsertPt,
PN->getType());
+ std::string Name = PN->getName();
+ PN->setName("");
+ NewVal->setName(Name);
+
// Replace the old PHI Node with the inserted computation.
PN->replaceAllUsesWith(NewVal);
DeadInsts.insert(PN);
Changed = true;
}
- DeleteTriviallyDeadInstructions(DeadInsts);
-
- // TODO: In the future we could replace all instructions in the loop body with
- // simpler expressions. It's not clear how useful this would be though or if
- // the code expansion cost would be worth it! We probably shouldn't do this
- // until we have a way to reuse expressions already in the code.
#if 0
+ // Now replace all derived expressions in the loop body with simpler
+ // expressions.
for (unsigned i = 0, e = L->getBlocks().size(); i != e; ++i)
if (LI->getLoopFor(L->getBlocks()[i]) == L) { // Not in a subloop...
BasicBlock *BB = L->getBlocks()[i];
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)) {
+ std::string Name = I->getName();
+ I->setName("");
+ V->setName(Name);
+ }
+ I->replaceAllUsesWith(V);
+ DeadInsts.insert(I);
+ ++NumRemoved;
+ Changed = true;
+ }
}
}
#endif
+
+ DeleteTriviallyDeadInstructions(DeadInsts);
+
+ if (mustPreserveAnalysisID(LCSSAID)) assert(L->isLCSSAForm());
}