1 //===-- LCSSA.cpp - Convert loops into loop-closed SSA form ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass transforms loops by placing phi nodes at the end of the loops for
11 // all values that are live across the loop boundary. For example, it turns
12 // the left into the right code:
14 // for (...) for (...)
19 // X3 = phi(X1, X2) X3 = phi(X1, X2)
20 // ... = X3 + 4 X4 = phi(X3)
23 // This is still valid LLVM; the extra phi nodes are purely redundant, and will
24 // be trivially eliminated by InstCombine. The major benefit of this
25 // transformation is that it makes many other loop optimizations, such as
26 // LoopUnswitching, simpler.
28 //===----------------------------------------------------------------------===//
30 #define DEBUG_TYPE "lcssa"
31 #include "llvm/Transforms/Scalar.h"
32 #include "llvm/Constants.h"
33 #include "llvm/Pass.h"
34 #include "llvm/Function.h"
35 #include "llvm/Instructions.h"
36 #include "llvm/LLVMContext.h"
37 #include "llvm/ADT/SetVector.h"
38 #include "llvm/ADT/Statistic.h"
39 #include "llvm/Analysis/Dominators.h"
40 #include "llvm/Analysis/LoopPass.h"
41 #include "llvm/Analysis/ScalarEvolution.h"
42 #include "llvm/Support/CFG.h"
43 #include "llvm/Support/Compiler.h"
44 #include "llvm/Support/PredIteratorCache.h"
49 STATISTIC(NumLCSSA, "Number of live out of a loop variables");
52 struct VISIBILITY_HIDDEN LCSSA : public LoopPass {
53 static char ID; // Pass identification, replacement for typeid
54 LCSSA() : LoopPass(&ID) {}
56 // Cached analysis information for the current function.
59 std::vector<BasicBlock*> LoopBlocks;
60 PredIteratorCache PredCache;
63 virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
65 void ProcessInstruction(Instruction* Instr,
66 const SmallVector<BasicBlock*, 8>& exitBlocks);
68 /// This transformation requires natural loop information & requires that
69 /// loop preheaders be inserted into the CFG. It maintains both of these,
70 /// as well as the CFG. It also requires dominator information.
72 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
74 AU.addRequiredID(LoopSimplifyID);
75 AU.addPreservedID(LoopSimplifyID);
76 AU.addRequiredTransitive<LoopInfo>();
77 AU.addPreserved<LoopInfo>();
78 AU.addRequiredTransitive<DominatorTree>();
79 AU.addPreserved<ScalarEvolution>();
80 AU.addPreserved<DominatorTree>();
82 // Request DominanceFrontier now, even though LCSSA does
83 // not use it. This allows Pass Manager to schedule Dominance
84 // Frontier early enough such that one LPPassManager can handle
85 // multiple loop transformation passes.
86 AU.addRequired<DominanceFrontier>();
87 AU.addPreserved<DominanceFrontier>();
91 /// verifyAnalysis() - Verify loop nest.
92 virtual void verifyAnalysis() const {
93 // Check the special guarantees that LCSSA makes.
94 assert(L->isLCSSAForm() && "LCSSA form not preserved!");
97 void getLoopValuesUsedOutsideLoop(Loop *L,
98 SetVector<Instruction*> &AffectedValues,
99 const SmallVector<BasicBlock*, 8>& exitBlocks);
101 Value *GetValueForBlock(DomTreeNode *BB, Instruction *OrigInst,
102 DenseMap<DomTreeNode*, Value*> &Phis);
104 /// inLoop - returns true if the given block is within the current loop
105 bool inLoop(BasicBlock* B) {
106 return std::binary_search(LoopBlocks.begin(), LoopBlocks.end(), B);
112 static RegisterPass<LCSSA> X("lcssa", "Loop-Closed SSA Form Pass");
114 Pass *llvm::createLCSSAPass() { return new LCSSA(); }
115 const PassInfo *const llvm::LCSSAID = &X;
117 /// runOnFunction - Process all loops in the function, inner-most out.
118 bool LCSSA::runOnLoop(Loop *l, LPPassManager &LPM) {
122 LI = &LPM.getAnalysis<LoopInfo>();
123 DT = &getAnalysis<DominatorTree>();
125 // Speed up queries by creating a sorted list of blocks
127 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
128 std::sort(LoopBlocks.begin(), LoopBlocks.end());
130 SmallVector<BasicBlock*, 8> exitBlocks;
131 L->getExitBlocks(exitBlocks);
133 SetVector<Instruction*> AffectedValues;
134 getLoopValuesUsedOutsideLoop(L, AffectedValues, exitBlocks);
136 // If no values are affected, we can save a lot of work, since we know that
137 // nothing will be changed.
138 if (AffectedValues.empty())
141 // Iterate over all affected values for this loop and insert Phi nodes
142 // for them in the appropriate exit blocks
144 for (SetVector<Instruction*>::iterator I = AffectedValues.begin(),
145 E = AffectedValues.end(); I != E; ++I)
146 ProcessInstruction(*I, exitBlocks);
148 assert(L->isLCSSAForm());
153 /// processInstruction - Given a live-out instruction, insert LCSSA Phi nodes,
154 /// eliminate all out-of-loop uses.
155 void LCSSA::ProcessInstruction(Instruction *Instr,
156 const SmallVector<BasicBlock*, 8>& exitBlocks) {
157 ++NumLCSSA; // We are applying the transformation
159 // Keep track of the blocks that have the value available already.
160 DenseMap<DomTreeNode*, Value*> Phis;
162 BasicBlock *DomBB = Instr->getParent();
164 // Invoke instructions are special in that their result value is not available
165 // along their unwind edge. The code below tests to see whether DomBB dominates
166 // the value, so adjust DomBB to the normal destination block, which is
167 // effectively where the value is first usable.
168 if (InvokeInst *Inv = dyn_cast<InvokeInst>(Instr))
169 DomBB = Inv->getNormalDest();
171 DomTreeNode *DomNode = DT->getNode(DomBB);
173 // Insert the LCSSA phi's into the exit blocks (dominated by the value), and
174 // add them to the Phi's map.
175 for (SmallVector<BasicBlock*, 8>::const_iterator BBI = exitBlocks.begin(),
176 BBE = exitBlocks.end(); BBI != BBE; ++BBI) {
177 BasicBlock *BB = *BBI;
178 DomTreeNode *ExitBBNode = DT->getNode(BB);
179 Value *&Phi = Phis[ExitBBNode];
180 if (!Phi && DT->dominates(DomNode, ExitBBNode)) {
181 PHINode *PN = PHINode::Create(Instr->getType(), Instr->getName()+".lcssa",
183 PN->reserveOperandSpace(PredCache.GetNumPreds(BB));
185 // Remember that this phi makes the value alive in this block.
188 // Add inputs from inside the loop for this PHI.
189 for (BasicBlock** PI = PredCache.GetPreds(BB); *PI; ++PI)
190 PN->addIncoming(Instr, *PI);
195 // Record all uses of Instr outside the loop. We need to rewrite these. The
196 // LCSSA phis won't be included because they use the value in the loop.
197 for (Value::use_iterator UI = Instr->use_begin(), E = Instr->use_end();
199 BasicBlock *UserBB = cast<Instruction>(*UI)->getParent();
200 if (PHINode *P = dyn_cast<PHINode>(*UI))
201 UserBB = P->getIncomingBlock(UI);
203 // If the user is in the loop, don't rewrite it!
204 if (UserBB == Instr->getParent() || inLoop(UserBB)) {
209 // Otherwise, patch up uses of the value with the appropriate LCSSA Phi,
210 // inserting PHI nodes into join points where needed.
211 Value *Val = GetValueForBlock(DT->getNode(UserBB), Instr, Phis);
213 // Preincrement the iterator to avoid invalidating it when we change the
215 Use &U = UI.getUse();
221 /// getLoopValuesUsedOutsideLoop - Return any values defined in the loop that
222 /// are used by instructions outside of it.
223 void LCSSA::getLoopValuesUsedOutsideLoop(Loop *L,
224 SetVector<Instruction*> &AffectedValues,
225 const SmallVector<BasicBlock*, 8>& exitBlocks) {
226 // FIXME: For large loops, we may be able to avoid a lot of use-scanning
227 // by using dominance information. In particular, if a block does not
228 // dominate any of the loop exits, then none of the values defined in the
229 // block could be used outside the loop.
230 for (Loop::block_iterator BB = L->block_begin(), BE = L->block_end();
232 for (BasicBlock::iterator I = (*BB)->begin(), E = (*BB)->end(); I != E; ++I)
233 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
235 BasicBlock *UserBB = cast<Instruction>(*UI)->getParent();
236 if (PHINode *p = dyn_cast<PHINode>(*UI))
237 UserBB = p->getIncomingBlock(UI);
239 if (*BB != UserBB && !inLoop(UserBB)) {
240 AffectedValues.insert(I);
247 /// GetValueForBlock - Get the value to use within the specified basic block.
248 /// available values are in Phis.
249 Value *LCSSA::GetValueForBlock(DomTreeNode *BB, Instruction *OrigInst,
250 DenseMap<DomTreeNode*, Value*> &Phis) {
251 // If there is no dominator info for this BB, it is unreachable.
253 return UndefValue::get(OrigInst->getType());
255 // If we have already computed this value, return the previously computed val.
256 if (Phis.count(BB)) return Phis[BB];
258 DomTreeNode *IDom = BB->getIDom();
260 // Otherwise, there are two cases: we either have to insert a PHI node or we
261 // don't. We need to insert a PHI node if this block is not dominated by one
262 // of the exit nodes from the loop (the loop could have multiple exits, and
263 // though the value defined *inside* the loop dominated all its uses, each
264 // exit by itself may not dominate all the uses).
266 // The simplest way to check for this condition is by checking to see if the
267 // idom is in the loop. If so, we *know* that none of the exit blocks
268 // dominate this block. Note that we *know* that the block defining the
269 // original instruction is in the idom chain, because if it weren't, then the
270 // original value didn't dominate this use.
271 if (!inLoop(IDom->getBlock())) {
272 // Idom is not in the loop, we must still be "below" the exit block and must
273 // be fully dominated by the value live in the idom.
274 Value* val = GetValueForBlock(IDom, OrigInst, Phis);
275 Phis.insert(std::make_pair(BB, val));
279 BasicBlock *BBN = BB->getBlock();
281 // Otherwise, the idom is the loop, so we need to insert a PHI node. Do so
282 // now, then get values to fill in the incoming values for the PHI.
283 PHINode *PN = PHINode::Create(OrigInst->getType(),
284 OrigInst->getName() + ".lcssa", BBN->begin());
285 PN->reserveOperandSpace(PredCache.GetNumPreds(BBN));
286 Phis.insert(std::make_pair(BB, PN));
288 // Fill in the incoming values for the block.
289 for (BasicBlock **PI = PredCache.GetPreds(BBN); *PI; ++PI)
290 PN->addIncoming(GetValueForBlock(DT->getNode(*PI), OrigInst, Phis), *PI);