//
// 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/SetVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/Dominators.h"
-#include "llvm/Analysis/LoopInfo.h"
+#include "llvm/Analysis/LoopPass.h"
+#include "llvm/Analysis/ScalarEvolution.h"
#include "llvm/Support/CFG.h"
+#include "llvm/Support/Compiler.h"
#include <algorithm>
-#include <cassert>
#include <map>
-#include <vector>
-
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
+ struct VISIBILITY_HIDDEN LCSSA : public LoopPass {
+ static char ID; // Pass identification, replacement for typeid
+ LCSSA() : LoopPass(&ID) {}
+
+ // Cached analysis information for the current function.
+ LoopInfo *LI;
+ DominatorTree *DT;
+ std::vector<BasicBlock*> LoopBlocks;
- virtual bool runOnFunction(Function &F);
- bool visitSubloop(Loop* L);
- void processInstruction(Instruction* Instr,
- const std::vector<BasicBlock*>& LoopBlocks,
- const std::vector<BasicBlock*>& exitBlocks);
+ virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
+
+ void ProcessInstruction(Instruction* Instr,
+ const SmallVector<BasicBlock*, 8>& exitBlocks);
/// This transformation requires natural loop information & requires that
/// loop preheaders be inserted into the CFG. It maintains both of these,
AU.addRequired<LoopInfo>();
AU.addPreserved<LoopInfo>();
AU.addRequired<DominatorTree>();
- AU.addRequired<DominanceFrontier>();
+ 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,
- const std::vector<BasicBlock*>& LoopBlocks);
- Instruction *getValueDominatingBlock(BasicBlock *BB,
- std::map<BasicBlock*, Instruction*> PotDoms);
+ void getLoopValuesUsedOutsideLoop(Loop *L,
+ SetVector<Instruction*> &AffectedValues);
+
+ Value *GetValueForBlock(DomTreeNode *BB, Instruction *OrigInst,
+ DenseMap<DomTreeNode*, Value*> &Phis);
+
+ /// inLoop - returns true if the given block is within the current loop
+ bool inLoop(BasicBlock* B) {
+ 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");
-FunctionPass *llvm::createLCSSAPass() { return new LCSSA(); }
+Pass *llvm::createLCSSAPass() { return new LCSSA(); }
+const PassInfo *const llvm::LCSSAID = &X;
-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 *L, LPPassManager &LPM) {
+
+ LI = &LPM.getAnalysis<LoopInfo>();
DT = &getAnalysis<DominatorTree>();
-
- for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I) {
- changed |= visitSubloop(*I);
- }
-
- return changed;
-}
-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());
+ LoopBlocks.clear();
+ LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
std::sort(LoopBlocks.begin(), LoopBlocks.end());
- std::set<Instruction*> AffectedValues = getLoopValuesUsedOutsideLoop(L,
- LoopBlocks);
+ SetVector<Instruction*> AffectedValues;
+ getLoopValuesUsedOutsideLoop(L, AffectedValues);
// If no values are affected, we can save a lot of work, since we know that
// nothing will be changed.
if (AffectedValues.empty())
return false;
- std::vector<BasicBlock*> exitBlocks;
- L->getExitBlocks(exitBlocks);
-
+ SmallVector<BasicBlock*, 8> exitBlocks;
+ L->getExitBlocks(exitBlocks);
// Iterate over all affected values for this loop and insert Phi nodes
// for them in the appropriate exit blocks
- for (std::set<Instruction*>::iterator I = AffectedValues.begin(),
- E = AffectedValues.end(); I != E; ++I) {
- processInstruction(*I, LoopBlocks, exitBlocks);
- }
+ for (SetVector<Instruction*>::iterator I = AffectedValues.begin(),
+ E = AffectedValues.end(); I != E; ++I)
+ ProcessInstruction(*I, exitBlocks);
+
+ assert(L->isLCSSAForm());
- return true; // FIXME: Should be more intelligent in our return value.
+ return true;
}
-/// processInstruction -
-void LCSSA::processInstruction(Instruction* Instr,
- const std::vector<BasicBlock*>& LoopBlocks,
- const std::vector<BasicBlock*>& exitBlocks)
-{
+/// processInstruction - Given a live-out instruction, insert LCSSA Phi nodes,
+/// eliminate all out-of-loop uses.
+void LCSSA::ProcessInstruction(Instruction *Instr,
+ const SmallVector<BasicBlock*, 8>& exitBlocks) {
++NumLCSSA; // We are applying the transformation
+
+ // Keep track of the blocks that have the value available already.
+ DenseMap<DomTreeNode*, Value*> Phis;
+
+ DomTreeNode *InstrNode = DT->getNode(Instr->getParent());
+
+ // Insert the LCSSA phi's into the exit blocks (dominated by the value), and
+ // add them to the Phi's map.
+ for (SmallVector<BasicBlock*, 8>::const_iterator BBI = exitBlocks.begin(),
+ BBE = exitBlocks.end(); BBI != BBE; ++BBI) {
+ BasicBlock *BB = *BBI;
+ DomTreeNode *ExitBBNode = DT->getNode(BB);
+ Value *&Phi = Phis[ExitBBNode];
+ if (!Phi && DT->dominates(InstrNode, ExitBBNode)) {
+ PHINode *PN = PHINode::Create(Instr->getType(), Instr->getName()+".lcssa",
+ BB->begin());
+ PN->reserveOperandSpace(std::distance(pred_begin(BB), pred_end(BB)));
+
+ // Remember that this phi makes the value alive in this block.
+ Phi = PN;
+
+ // Add inputs from inside the loop for this PHI.
+ for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
+ PN->addIncoming(Instr, *PI);
+ }
+ }
- std::map<BasicBlock*, Instruction*> Phis;
- Phis[Instr->getParent()] = Instr;
-
- // Phi nodes that need to be IDF-processed
- std::vector<PHINode*> workList;
- for (std::vector<BasicBlock*>::const_iterator BBI = exitBlocks.begin(),
- BBE = exitBlocks.end(); BBI != BBE; ++BBI)
- if (DT->getNode(Instr->getParent())->dominates(DT->getNode(*BBI))) {
- PHINode *phi = new PHINode(Instr->getType(), "lcssa", (*BBI)->begin());
- workList.push_back(phi);
- Phis[*BBI] = phi;
+ // Record all uses of Instr outside the loop. We need to rewrite these. The
+ // LCSSA phis won't be included because they use the value in the loop.
+ for (Value::use_iterator UI = Instr->use_begin(), E = Instr->use_end();
+ UI != E;) {
+ BasicBlock *UserBB = cast<Instruction>(*UI)->getParent();
+ if (PHINode *P = dyn_cast<PHINode>(*UI)) {
+ UserBB = P->getIncomingBlock(UI);
}
-
- // Calculate the IDF of these LCSSA Phi nodes, inserting new Phi's where
- // necessary. Keep track of these new Phi's in Phis.
- while (!workList.empty()) {
- PHINode *CurPHI = workList.back();
- workList.pop_back();
- // Get the current Phi's DF, and insert Phi nodes. Add these new
- // nodes to our worklist.
- DominanceFrontier::const_iterator it = DF->find(CurPHI->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 (Phis[*P] == 0) {
- // Still doesn't have operands...
- PHINode *phi = new PHINode(Instr->getType(), "lcssa", (*P)->begin());
- Phis[*P] = phi;
-
- workList.push_back(phi);
- }
- }
+ // If the user is in the loop, don't rewrite it!
+ if (UserBB == Instr->getParent() || inLoop(UserBB)) {
+ ++UI;
+ 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(CurPHI->getParent()),
- E = pred_end(CurPHI->getParent()); PI != E; ++PI)
- CurPHI->addIncoming(getValueDominatingBlock(*PI, Phis),
- *PI);
- }
-
- // Find all uses of the affected value, and replace them with the
- // appropriate Phi.
- std::vector<Instruction*> Uses;
- for (Instruction::use_iterator UI = Instr->use_begin(), UE = Instr->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()) &&
- !(std::binary_search(exitBlocks.begin(), exitBlocks.end(),
- use->getParent()) && isa<PHINode>(use)))
- Uses.push_back(use);
- }
-
- // Deliberately remove the initial instruction from Phis set.
- Phis.erase(Instr->getParent());
-
- for (std::vector<Instruction*>::iterator II = Uses.begin(), IE = Uses.end();
- II != IE; ++II) {
- if (PHINode* phi = dyn_cast<PHINode>(*II)) {
- for (unsigned int i = 0; i < phi->getNumIncomingValues(); ++i) {
- // FIXME: Replace a Phi entry if and only if the corresponding
- // predecessor is dominated.
- Instruction* dominator =
- getValueDominatingBlock(phi->getIncomingBlock(i), Phis);
-
- if (phi->getIncomingValue(i) == Instr)
- phi->setIncomingValue(i, dominator);
- }
- } else {
- (*II)->replaceUsesOfWith(Instr,
- getValueDominatingBlock((*II)->getParent(),
- Phis));
- }
+ // Otherwise, patch up uses of the value with the appropriate LCSSA Phi,
+ // inserting PHI nodes into join points where needed.
+ Value *Val = GetValueForBlock(DT->getNode(UserBB), Instr, Phis);
+
+ // Preincrement the iterator to avoid invalidating it when we change the
+ // value.
+ Use &U = UI.getUse();
+ ++UI;
+ U.set(Val);
}
}
/// getLoopValuesUsedOutsideLoop - Return any values defined in the loop that
/// are used by instructions outside of it.
-std::set<Instruction*> LCSSA::getLoopValuesUsedOutsideLoop(Loop *L,
- const std::vector<BasicBlock*>& LoopBlocks) {
-
+void LCSSA::getLoopValuesUsedOutsideLoop(Loop *L,
+ SetVector<Instruction*> &AffectedValues) {
// FIXME: For large loops, we may be able to avoid a lot of 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.
-
- 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)) {
+ if (PHINode* p = dyn_cast<PHINode>(*UI)) {
+ UserBB = p->getIncomingBlock(UI);
+ }
+
+ if (*BB != UserBB && !inLoop(UserBB)) {
AffectedValues.insert(I);
break;
}
}
}
- return AffectedValues;
}
-Instruction *LCSSA::getValueDominatingBlock(BasicBlock *BB,
- std::map<BasicBlock*, Instruction*> PotDoms) {
- DominatorTree::Node* bbNode = DT->getNode(BB);
- while (bbNode != 0) {
- std::map<BasicBlock*, Instruction*>::iterator I =
- PotDoms.find(bbNode->getBlock());
- if (I != PotDoms.end()) {
- return (*I).second;
- }
- bbNode = bbNode->getIDom();
+/// GetValueForBlock - Get the value to use within the specified basic block.
+/// available values are in Phis.
+Value *LCSSA::GetValueForBlock(DomTreeNode *BB, Instruction *OrigInst,
+ DenseMap<DomTreeNode*, Value*> &Phis) {
+ // If there is no dominator info for this BB, it is unreachable.
+ if (BB == 0)
+ return UndefValue::get(OrigInst->getType());
+
+ // If we have already computed this value, return the previously computed val.
+ if (Phis.count(BB)) return Phis[BB];
+
+ DomTreeNode *IDom = BB->getIDom();
+
+ // Otherwise, there are two cases: we either have to insert a PHI node or we
+ // don't. We need to insert a PHI node if this block is not dominated by one
+ // of the exit nodes from the loop (the loop could have multiple exits, and
+ // though the value defined *inside* the loop dominated all its uses, each
+ // exit by itself may not dominate all the uses).
+ //
+ // The simplest way to check for this condition is by checking to see if the
+ // idom is in the loop. If so, we *know* that none of the exit blocks
+ // dominate this block. Note that we *know* that the block defining the
+ // original instruction is in the idom chain, because if it weren't, then the
+ // original value didn't dominate this use.
+ if (!inLoop(IDom->getBlock())) {
+ // Idom is not in the loop, we must still be "below" the exit block and must
+ // be fully dominated by the value live in the idom.
+ Value* val = GetValueForBlock(IDom, OrigInst, Phis);
+ Phis.insert(std::make_pair(BB, val));
+ return val;
}
- assert(0 && "No dominating value found.");
+ BasicBlock *BBN = BB->getBlock();
- return 0;
+ // Otherwise, the idom is the loop, so we need to insert a PHI node. Do so
+ // now, then get values to fill in the incoming values for the PHI.
+ PHINode *PN = PHINode::Create(OrigInst->getType(),
+ OrigInst->getName() + ".lcssa", BBN->begin());
+ PN->reserveOperandSpace(std::distance(pred_begin(BBN), pred_end(BBN)));
+ Phis.insert(std::make_pair(BB, PN));
+
+ // Fill in the incoming values for the block.
+ for (pred_iterator PI = pred_begin(BBN), E = pred_end(BBN); PI != E; ++PI)
+ PN->addIncoming(GetValueForBlock(DT->getNode(*PI), OrigInst, Phis), *PI);
+ return PN;
}
+
projects / oota-llvm.git / blobdiff