1 //===-- LICM.cpp - Loop Invariant Code Motion Pass ------------------------===//
3 // This pass is a simple loop invariant code motion pass.
5 // Note that this pass does NOT require pre-headers to exist on loops in the
6 // CFG, but if there is not distinct preheader for a loop, the hoisted code will
7 // be *DUPLICATED* in every basic block, outside of the loop, that preceeds the
8 // loop header. Additionally, any use of one of these hoisted expressions
9 // cannot be loop invariant itself, because the expression hoisted gets a PHI
10 // node that is loop variant.
12 // For these reasons, and many more, it makes sense to run a pass before this
13 // that ensures that there are preheaders on all loops. That said, we don't
16 //===----------------------------------------------------------------------===//
18 #include "llvm/Transforms/Scalar.h"
19 #include "llvm/Transforms/Utils/Local.h"
20 #include "llvm/Analysis/LoopInfo.h"
21 #include "llvm/Analysis/AliasAnalysis.h"
22 #include "llvm/iOperators.h"
23 #include "llvm/iPHINode.h"
24 #include "llvm/iMemory.h"
25 #include "llvm/Support/InstVisitor.h"
26 #include "llvm/Support/CFG.h"
27 #include "Support/STLExtras.h"
28 #include "Support/StatisticReporter.h"
32 static Statistic<> NumHoistedNPH("licm\t\t- Number of insts hoisted to multiple"
33 " loop preds (bad, no loop pre-header)");
34 static Statistic<> NumHoistedPH("licm\t\t- Number of insts hoisted to a loop "
36 static Statistic<> NumHoistedLoads("licm\t\t- Number of load insts hoisted");
39 struct LICM : public FunctionPass, public InstVisitor<LICM> {
40 virtual bool runOnFunction(Function &F);
42 // This transformation requires natural loop information...
43 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
45 AU.addRequired<LoopInfo>();
46 AU.addRequired<AliasAnalysis>();
50 // List of predecessor blocks for the current loop - These blocks are where
51 // we hoist loop invariants to for the current loop.
53 std::vector<BasicBlock*> LoopPreds, LoopBackEdges;
55 Loop *CurLoop; // The current loop we are working on...
56 bool Changed; // Set to true when we change anything.
57 AliasAnalysis *AA; // Currently AliasAnalysis information
59 // visitLoop - Hoist expressions out of the specified loop...
60 void visitLoop(Loop *L);
62 // notInCurrentLoop - Little predicate that returns true if the specified
63 // basic block is in a subloop of the current one, not the current one
66 bool notInCurrentLoop(BasicBlock *BB) {
67 for (unsigned i = 0, e = CurLoop->getSubLoops().size(); i != e; ++i)
68 if (CurLoop->getSubLoops()[i]->contains(BB))
69 return true; // A subloop actually contains this block!
73 // hoist - When an instruction is found to only use loop invariant operands
74 // that is safe to hoist, this instruction is called to do the dirty work.
76 void hoist(Instruction &I);
78 // pointerInvalidatedByLoop - Return true if the body of this loop may store
79 // into the memory location pointed to by V.
81 bool pointerInvalidatedByLoop(Value *V);
83 // isLoopInvariant - Return true if the specified value is loop invariant
84 inline bool isLoopInvariant(Value *V) {
85 if (Instruction *I = dyn_cast<Instruction>(V))
86 return !CurLoop->contains(I->getParent());
87 return true; // All non-instructions are loop invariant
90 // visitBasicBlock - Run LICM on a particular block.
91 void visitBasicBlock(BasicBlock *BB);
93 // Instruction visitation handlers... these basically control whether or not
94 // the specified instruction types are hoisted.
96 friend class InstVisitor<LICM>;
97 void visitBinaryOperator(Instruction &I) {
98 if (isLoopInvariant(I.getOperand(0)) && isLoopInvariant(I.getOperand(1)))
101 void visitCastInst(CastInst &CI) {
102 Instruction &I = (Instruction&)CI;
103 if (isLoopInvariant(I.getOperand(0))) hoist(I);
105 void visitShiftInst(ShiftInst &I) { visitBinaryOperator((Instruction&)I); }
107 void visitLoadInst(LoadInst &LI) {
108 if (isLoopInvariant(LI.getOperand(0)) &&
109 !pointerInvalidatedByLoop(LI.getOperand(0)))
113 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
114 Instruction &I = (Instruction&)GEPI;
115 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
116 if (!isLoopInvariant(I.getOperand(i))) return;
121 RegisterOpt<LICM> X("licm", "Loop Invariant Code Motion");
124 Pass *createLICMPass() { return new LICM(); }
126 bool LICM::runOnFunction(Function &) {
127 // get our loop information...
128 const std::vector<Loop*> &TopLevelLoops =
129 getAnalysis<LoopInfo>().getTopLevelLoops();
131 // Get our alias analysis information...
132 AA = &getAnalysis<AliasAnalysis>();
134 // Traverse loops in postorder, hoisting expressions out of the deepest loops
138 std::for_each(TopLevelLoops.begin(), TopLevelLoops.end(),
139 bind_obj(this, &LICM::visitLoop));
143 void LICM::visitLoop(Loop *L) {
144 // Recurse through all subloops before we process this loop...
145 std::for_each(L->getSubLoops().begin(), L->getSubLoops().end(),
146 bind_obj(this, &LICM::visitLoop));
149 // Calculate the set of predecessors for this loop. The predecessors for this
150 // loop are equal to the predecessors for the header node of the loop that are
151 // not themselves in the loop.
153 BasicBlock *Header = L->getHeader();
155 // Calculate the sets of predecessors and backedges of the loop...
156 LoopBackEdges.insert(LoopBackEdges.end(),pred_begin(Header),pred_end(Header));
158 std::vector<BasicBlock*>::iterator LPI =
159 std::partition(LoopBackEdges.begin(), LoopBackEdges.end(),
160 bind_obj(CurLoop, &Loop::contains));
162 // Move all predecessors to the LoopPreds vector...
163 LoopPreds.insert(LoopPreds.end(), LPI, LoopBackEdges.end());
165 // Remove predecessors from backedges list...
166 LoopBackEdges.erase(LPI, LoopBackEdges.end());
169 // The only way that there could be no predecessors to a loop is if the loop
170 // is not reachable. Since we don't care about optimizing dead loops,
171 // summarily ignore them.
173 if (LoopPreds.empty()) return;
175 // We want to visit all of the instructions in this loop... that are not parts
176 // of our subloops (they have already had their invariants hoisted out of
177 // their loop, into this loop, so there is no need to process the BODIES of
180 std::vector<BasicBlock*> BBs(L->getBlocks().begin(), L->getBlocks().end());
182 // Remove blocks that are actually in subloops...
183 BBs.erase(std::remove_if(BBs.begin(), BBs.end(),
184 bind_obj(this, &LICM::notInCurrentLoop)), BBs.end());
186 // Visit all of the basic blocks we have chosen, hoisting out the instructions
187 // as neccesary. This leaves dead copies of the instruction in the loop
190 for_each(BBs.begin(), BBs.end(), bind_obj(this, &LICM::visitBasicBlock));
192 // Clear out loops state information for the next iteration
195 LoopBackEdges.clear();
198 void LICM::visitBasicBlock(BasicBlock *BB) {
199 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
202 if (dceInstruction(I))
210 void LICM::hoist(Instruction &Inst) {
211 if (Inst.use_empty()) return; // Don't (re) hoist dead instructions!
212 //cerr << "Hoisting " << Inst;
214 BasicBlock *Header = CurLoop->getHeader();
216 // Old instruction will be removed, so take it's name...
217 string InstName = Inst.getName();
220 if (isa<LoadInst>(Inst))
223 // The common case is that we have a pre-header. Generate special case code
224 // that is faster if that is the case.
226 if (LoopPreds.size() == 1) {
227 BasicBlock *Pred = LoopPreds[0];
229 // Create a new copy of the instruction, for insertion into Pred.
230 Instruction *New = Inst.clone();
231 New->setName(InstName);
233 // Insert the new node in Pred, before the terminator.
234 Pred->getInstList().insert(--Pred->end(), New);
236 // Kill the old instruction...
237 Inst.replaceAllUsesWith(New);
241 // No loop pre-header, insert a PHI node into header to capture all of the
242 // incoming versions of the value.
244 PHINode *LoopVal = new PHINode(Inst.getType(), InstName+".phi");
246 // Insert the new PHI node into the loop header...
247 Header->getInstList().push_front(LoopVal);
249 // Insert cloned versions of the instruction into all of the loop preds.
250 for (unsigned i = 0, e = LoopPreds.size(); i != e; ++i) {
251 BasicBlock *Pred = LoopPreds[i];
253 // Create a new copy of the instruction, for insertion into Pred.
254 Instruction *New = Inst.clone();
255 New->setName(InstName);
257 // Insert the new node in Pred, before the terminator.
258 Pred->getInstList().insert(--Pred->end(), New);
260 // Add the incoming value to the PHI node.
261 LoopVal->addIncoming(New, Pred);
264 // Add incoming values to the PHI node for all backedges in the loop...
265 for (unsigned i = 0, e = LoopBackEdges.size(); i != e; ++i)
266 LoopVal->addIncoming(LoopVal, LoopBackEdges[i]);
268 // Replace all uses of the old version of the instruction in the loop with
269 // the new version that is out of the loop. We know that this is ok,
270 // because the new definition is in the loop header, which dominates the
271 // entire loop body. The old definition was defined _inside_ of the loop,
272 // so the scope cannot extend outside of the loop, so we're ok.
274 Inst.replaceAllUsesWith(LoopVal);
281 // pointerInvalidatedByLoop - Return true if the body of this loop may store
282 // into the memory location pointed to by V.
284 bool LICM::pointerInvalidatedByLoop(Value *V) {
285 // Check to see if any of the basic blocks in CurLoop invalidate V.
286 for (unsigned i = 0, e = CurLoop->getBlocks().size(); i != e; ++i)
287 if (AA->canBasicBlockModify(*CurLoop->getBlocks()[i], V))