//
// The LLVM Compiler Infrastructure
//
-// This file was developed by Owen Anderson 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.
//
//===----------------------------------------------------------------------===//
//
// the left into the right code:
//
// for (...) for (...)
-// if (c) if(c)
+// if (c) if (c)
// X1 = ... X1 = ...
// else else
// X2 = ... X2 = ...
// X3 = phi(X1, X2) X3 = phi(X1, X2)
-// ... = X3 + 4 X4 = phi(X3)
-// ... = X4 + 4
+// ... = X3 + 4 X4 = phi(X3)
+// ... = X4 + 4
//
// This is still valid LLVM; the extra phi nodes are purely redundant, and will
// be trivially eliminated by InstCombine. The major benefit of this
//
//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "lcssa"
#include "llvm/Transforms/Scalar.h"
+#include "llvm/Constants.h"
#include "llvm/Pass.h"
#include "llvm/Function.h"
#include "llvm/Instructions.h"
-#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/Dominators.h"
-#include "llvm/Analysis/LoopInfo.h"
-#include "llvm/Support/CFG.h"
-#include <algorithm>
-#include <map>
-#include <vector>
-
+#include "llvm/Analysis/LoopPass.h"
+#include "llvm/Analysis/ScalarEvolution.h"
+#include "llvm/Transforms/Utils/SSAUpdater.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/ADT/STLExtras.h"
+#include "llvm/Support/PredIteratorCache.h"
using namespace llvm;
+STATISTIC(NumLCSSA, "Number of live out of a loop variables");
+
namespace {
- static Statistic<> NumLCSSA("lcssa",
- "Number of live out of a loop variables");
-
- class LCSSA : public FunctionPass {
- public:
-
-
- LoopInfo *LI; // Loop information
- DominatorTree *DT; // Dominator Tree for the current Loop...
- DominanceFrontier *DF; // Current Dominance Frontier
-
- virtual bool runOnFunction(Function &F);
- bool visitSubloop(Loop* L);
+ struct LCSSA : public LoopPass {
+ static char ID; // Pass identification, replacement for typeid
+ LCSSA() : LoopPass(&ID) {}
+
+ // Cached analysis information for the current function.
+ DominatorTree *DT;
+ std::vector<BasicBlock*> LoopBlocks;
+ PredIteratorCache PredCache;
+ Loop *L;
+ virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
+
/// This transformation requires natural loop information & requires that
/// loop preheaders be inserted into the CFG. It maintains both of these,
/// as well as the CFG. It also requires dominator information.
///
virtual void getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesCFG();
+
+ // LCSSA doesn't actually require LoopSimplify, but the PassManager
+ // doesn't know how to schedule LoopSimplify by itself.
AU.addRequiredID(LoopSimplifyID);
AU.addPreservedID(LoopSimplifyID);
- AU.addRequired<LoopInfo>();
+ AU.addRequiredTransitive<LoopInfo>();
AU.addPreserved<LoopInfo>();
- AU.addRequired<DominatorTree>();
- AU.addRequired<DominanceFrontier>();
+ AU.addRequiredTransitive<DominatorTree>();
+ AU.addPreserved<ScalarEvolution>();
+ AU.addPreserved<DominatorTree>();
+
+ // Request DominanceFrontier now, even though LCSSA does
+ // not use it. This allows Pass Manager to schedule Dominance
+ // Frontier early enough such that one LPPassManager can handle
+ // multiple loop transformation passes.
+ AU.addRequired<DominanceFrontier>();
+ AU.addPreserved<DominanceFrontier>();
}
private:
- std::set<Instruction*> getLoopValuesUsedOutsideLoop(Loop *L,
- std::vector<BasicBlock*> LoopBlocks);
+ bool ProcessInstruction(Instruction *Inst,
+ const SmallVectorImpl<BasicBlock*> &ExitBlocks);
+
+ /// verifyAnalysis() - Verify loop nest.
+ virtual void verifyAnalysis() const {
+ // Check the special guarantees that LCSSA makes.
+ assert(L->isLCSSAForm(*DT) && "LCSSA form not preserved!");
+ }
+
+ /// inLoop - returns true if the given block is within the current loop
+ bool inLoop(BasicBlock *B) const {
+ return std::binary_search(LoopBlocks.begin(), LoopBlocks.end(), B);
+ }
};
+}
- RegisterOpt<LCSSA> X("lcssa", "Loop-Closed SSA Form Pass");
+char LCSSA::ID = 0;
+static RegisterPass<LCSSA> X("lcssa", "Loop-Closed SSA Form Pass");
+
+Pass *llvm::createLCSSAPass() { return new LCSSA(); }
+const PassInfo *const llvm::LCSSAID = &X;
+
+
+/// BlockDominatesAnExit - Return true if the specified block dominates at least
+/// one of the blocks in the specified list.
+static bool BlockDominatesAnExit(BasicBlock *BB,
+ const SmallVectorImpl<BasicBlock*> &ExitBlocks,
+ DominatorTree *DT) {
+ DomTreeNode *DomNode = DT->getNode(BB);
+ for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
+ if (DT->dominates(DomNode, DT->getNode(ExitBlocks[i])))
+ return true;
+
+ return false;
}
-FunctionPass *llvm::createLCSSAPass() { return new LCSSA(); }
-bool LCSSA::runOnFunction(Function &F) {
- bool changed = false;
- LI = &getAnalysis<LoopInfo>();
- DF = &getAnalysis<DominanceFrontier>();
+/// runOnFunction - Process all loops in the function, inner-most out.
+bool LCSSA::runOnLoop(Loop *TheLoop, LPPassManager &LPM) {
+ L = TheLoop;
+
DT = &getAnalysis<DominatorTree>();
+
+ // Get the set of exiting blocks.
+ SmallVector<BasicBlock*, 8> ExitBlocks;
+ L->getExitBlocks(ExitBlocks);
+
+ if (ExitBlocks.empty())
+ return false;
+
+ // Speed up queries by creating a sorted vector of blocks.
+ LoopBlocks.clear();
+ LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
+ array_pod_sort(LoopBlocks.begin(), LoopBlocks.end());
+
+ // Look at all the instructions in the loop, checking to see if they have uses
+ // outside the loop. If so, rewrite those uses.
+ bool MadeChange = false;
+
+ for (Loop::block_iterator BBI = L->block_begin(), E = L->block_end();
+ BBI != E; ++BBI) {
+ BasicBlock *BB = *BBI;
- for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I) {
- changed |= visitSubloop(*I);
- }
+ // For large loops, avoid use-scanning by using dominance information: In
+ // particular, if a block does not dominate any of the loop exits, then none
+ // of the values defined in the block could be used outside the loop.
+ if (!BlockDominatesAnExit(BB, ExitBlocks, DT))
+ continue;
+
+ for (BasicBlock::iterator I = BB->begin(), E = BB->end();
+ I != E; ++I) {
+ // Reject two common cases fast: instructions with no uses (like stores)
+ // and instructions with one use that is in the same block as this.
+ if (I->use_empty() ||
+ (I->hasOneUse() && I->use_back()->getParent() == BB &&
+ !isa<PHINode>(I->use_back())))
+ continue;
- return changed;
+ MadeChange |= ProcessInstruction(I, ExitBlocks);
+ }
+ }
+
+ assert(L->isLCSSAForm(*DT));
+ PredCache.clear();
+
+ return MadeChange;
}
-bool LCSSA::visitSubloop(Loop* L) {
- for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
- visitSubloop(*I);
-
- // Speed up queries by creating a sorted list of blocks
- std::vector<BasicBlock*> LoopBlocks(L->block_begin(), L->block_end());
- std::sort(LoopBlocks.begin(), LoopBlocks.end());
+/// isExitBlock - Return true if the specified block is in the list.
+static bool isExitBlock(BasicBlock *BB,
+ const SmallVectorImpl<BasicBlock*> &ExitBlocks) {
+ for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
+ if (ExitBlocks[i] == BB)
+ return true;
+ return false;
+}
+
+/// ProcessInstruction - Given an instruction in the loop, check to see if it
+/// has any uses that are outside the current loop. If so, insert LCSSA PHI
+/// nodes and rewrite the uses.
+bool LCSSA::ProcessInstruction(Instruction *Inst,
+ const SmallVectorImpl<BasicBlock*> &ExitBlocks) {
+ SmallVector<Use*, 16> UsesToRewrite;
- std::set<Instruction*> AffectedValues = getLoopValuesUsedOutsideLoop(L,
- LoopBlocks);
+ BasicBlock *InstBB = Inst->getParent();
- std::vector<BasicBlock*> exitBlocks;
- L->getExitBlocks(exitBlocks);
+ for (Value::use_iterator UI = Inst->use_begin(), E = Inst->use_end();
+ UI != E; ++UI) {
+ BasicBlock *UserBB = cast<Instruction>(*UI)->getParent();
+ if (PHINode *PN = dyn_cast<PHINode>(*UI))
+ UserBB = PN->getIncomingBlock(UI);
+
+ if (InstBB != UserBB && !inLoop(UserBB))
+ UsesToRewrite.push_back(&UI.getUse());
+ }
- // Phi nodes that need to be IDF-processed
- std::vector<PHINode*> workList;
+ // If there are no uses outside the loop, exit with no change.
+ if (UsesToRewrite.empty()) return false;
- // Iterate over all affected values for this loop and insert Phi nodes
- // for them in the appropriate exit blocks
- std::map<BasicBlock*, PHINode*> ExitPhis;
- for (std::set<Instruction*>::iterator I = AffectedValues.begin(),
- E = AffectedValues.end(); I != E; ++I) {
- ++NumLCSSA; // We are applying the transformation
- for (std::vector<BasicBlock*>::iterator BBI = exitBlocks.begin(),
- BBE = exitBlocks.end(); BBI != BBE; ++BBI) {
- PHINode *phi = new PHINode((*I)->getType(), "lcssa");
- (*BBI)->getInstList().insert((*BBI)->front(), phi);
- workList.push_back(phi);
- ExitPhis[*BBI] = phi;
-
- // Since LoopSimplify has been run, we know that all of these predecessors
- // are in the loop, so just hook them up in the obvious manner.
- for (pred_iterator PI = pred_begin(*BBI), PE = pred_end(*BBI); PI != PE;
- ++PI)
- phi->addIncoming(*I, *PI);
- }
+ ++NumLCSSA; // We are applying the transformation
+
+ // Invoke instructions are special in that their result value is not available
+ // along their unwind edge. The code below tests to see whether DomBB dominates
+ // the value, so adjust DomBB to the normal destination block, which is
+ // effectively where the value is first usable.
+ BasicBlock *DomBB = Inst->getParent();
+ if (InvokeInst *Inv = dyn_cast<InvokeInst>(Inst))
+ DomBB = Inv->getNormalDest();
+
+ DomTreeNode *DomNode = DT->getNode(DomBB);
+
+ SSAUpdater SSAUpdate;
+ SSAUpdate.Initialize(Inst);
- // Calculate the IDF of these LCSSA Phi nodes, inserting new Phi's where
- // necessary. Keep track of these new Phi's in DFPhis.
- std::map<BasicBlock*, PHINode*> DFPhis;
- for (std::vector<PHINode*>::iterator DFI = workList.begin(),
- E = workList.end(); DFI != E; ++DFI) {
+ // Insert the LCSSA phi's into all of the exit blocks dominated by the
+ // value, and add them to the Phi's map.
+ for (SmallVectorImpl<BasicBlock*>::const_iterator BBI = ExitBlocks.begin(),
+ BBE = ExitBlocks.end(); BBI != BBE; ++BBI) {
+ BasicBlock *ExitBB = *BBI;
+ if (!DT->dominates(DomNode, DT->getNode(ExitBB))) continue;
- // Get the current Phi's DF, and insert Phi nodes. Add these new
- // nodes to our worklist.
- DominanceFrontier::const_iterator it = DF->find((*DFI)->getParent());
- if (it != DF->end()) {
- const DominanceFrontier::DomSetType &S = it->second;
- for (DominanceFrontier::DomSetType::const_iterator P = S.begin(),
- PE = S.end(); P != PE; ++P) {
- if (DFPhis[*P] == 0) {
- // Still doesn't have operands...
- PHINode *phi = new PHINode((*DFI)->getType(), "lcssa");
- (*P)->getInstList().insert((*P)->front(), phi);
- DFPhis[*P] = phi;
-
- workList.push_back(phi);
- }
- }
- }
+ // If we already inserted something for this BB, don't reprocess it.
+ if (SSAUpdate.HasValueForBlock(ExitBB)) continue;
- // Get the predecessor blocks of the current Phi, and use them to hook up
- // the operands of the current Phi to any members of DFPhis that dominate
- // it. This is a nop for the Phis inserted directly in the exit blocks,
- // since they are not dominated by any members of DFPhis.
- for (pred_iterator PI = pred_begin((*DFI)->getParent()),
- E = pred_end((*DFI)->getParent()); PI != E; ++PI)
- for (std::map<BasicBlock*, PHINode*>::iterator MI = DFPhis.begin(),
- ME = DFPhis.end(); MI != ME; ++MI)
- if (DT->getNode((*MI).first)->dominates(DT->getNode(*PI))) {
- (*DFI)->addIncoming((*MI).second, *PI);
-
- // Since dominate() is not cheap, don't do it more than we have to.
- break;
- }
+ PHINode *PN = PHINode::Create(Inst->getType(), Inst->getName()+".lcssa",
+ ExitBB->begin());
+ PN->reserveOperandSpace(PredCache.GetNumPreds(ExitBB));
+
+ // Add inputs from inside the loop for this PHI.
+ for (BasicBlock **PI = PredCache.GetPreds(ExitBB); *PI; ++PI) {
+ PN->addIncoming(Inst, *PI);
+
+ // If the exit block has a predecessor not within the loop, arrange for
+ // the incoming value use corresponding to that predecessor to be
+ // rewritten in terms of a different LCSSA PHI.
+ if (!inLoop(*PI))
+ UsesToRewrite.push_back(
+ &PN->getOperandUse(
+ PN->getOperandNumForIncomingValue(PN->getNumIncomingValues()-1)));
}
-
-
- // Find all uses of the affected value, and replace them with the
- // appropriate Phi.
- for (Instruction::use_iterator UI = (*I)->use_begin(), UE=(*I)->use_end();
- UI != UE; ++UI) {
- Instruction* use = cast<Instruction>(*UI);
-
- // Don't need to update uses within the loop body
- if (!std::binary_search(LoopBlocks.begin(), LoopBlocks.end(),
- use->getParent())) {
-
- for (std::map<BasicBlock*, PHINode*>::iterator DI = ExitPhis.begin(),
- DE = ExitPhis.end(); DI != DE; ++DI) {
- if (DT->getNode((*DI).first)->dominates( \
- DT->getNode(use->getParent())) && use != (*DI).second) {
- use->replaceUsesOfWith(*I, (*DI).second);
- break;
- }
- }
-
- for (std::map<BasicBlock*, PHINode*>::iterator DI = DFPhis.begin(),
- DE = DFPhis.end(); DI != DE; ++DI) {
- if (DT->getNode((*DI).first)->dominates( \
- DT->getNode(use->getParent()))) {
- use->replaceUsesOfWith(*I, (*DI).second);
- break;
- }
- }
- }
- }
+ // Remember that this phi makes the value alive in this block.
+ SSAUpdate.AddAvailableValue(ExitBB, PN);
}
- return true; // FIXME: Should be more intelligent in our return value.
-}
+ // Rewrite all uses outside the loop in terms of the new PHIs we just
+ // inserted.
+ for (unsigned i = 0, e = UsesToRewrite.size(); i != e; ++i) {
+ // If this use is in an exit block, rewrite to use the newly inserted PHI.
+ // This is required for correctness because SSAUpdate doesn't handle uses in
+ // the same block. It assumes the PHI we inserted is at the end of the
+ // block.
+ Instruction *User = cast<Instruction>(UsesToRewrite[i]->getUser());
+ BasicBlock *UserBB = User->getParent();
+ if (PHINode *PN = dyn_cast<PHINode>(User))
+ UserBB = PN->getIncomingBlock(*UsesToRewrite[i]);
-/// getLoopValuesUsedOutsideLoop - Return any values defined in the loop that
-/// are used by instructions outside of it.
-std::set<Instruction*> LCSSA::getLoopValuesUsedOutsideLoop(Loop *L,
- std::vector<BasicBlock*> LoopBlocks) {
-
- std::set<Instruction*> AffectedValues;
- for (Loop::block_iterator BB = L->block_begin(), E = L->block_end();
- BB != E; ++BB) {
- for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end(); I != E; ++I)
- for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
- ++UI) {
- BasicBlock *UserBB = cast<Instruction>(*UI)->getParent();
- if (!std::binary_search(LoopBlocks.begin(), LoopBlocks.end(), UserBB))
- AffectedValues.insert(I);
- }
+ if (isa<PHINode>(UserBB->begin()) &&
+ isExitBlock(UserBB, ExitBlocks)) {
+ UsesToRewrite[i]->set(UserBB->begin());
+ continue;
+ }
+
+ // Otherwise, do full PHI insertion.
+ SSAUpdate.RewriteUse(*UsesToRewrite[i]);
}
- return AffectedValues;
+
+ return true;
}
+