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/iOperators.h"
22 #include "llvm/iPHINode.h"
23 #include "llvm/Support/InstVisitor.h"
24 #include "llvm/Support/CFG.h"
25 #include "Support/STLExtras.h"
26 #include "Support/StatisticReporter.h"
30 static Statistic<> NumHoistedNPH("licm\t\t- Number of insts hoisted to multiple"
31 " loop preds (bad, no loop pre-header)");
32 static Statistic<> NumHoistedPH("licm\t\t- Number of insts hoisted to a loop "
36 struct LICM : public FunctionPass, public InstVisitor<LICM> {
37 virtual bool runOnFunction(Function &F);
39 // This transformation requires natural loop information...
40 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
42 AU.addRequired<LoopInfo>();
46 // List of predecessor blocks for the current loop - These blocks are where
47 // we hoist loop invariants to for the current loop.
49 std::vector<BasicBlock*> LoopPreds, LoopBackEdges;
51 Loop *CurLoop; // The current loop we are working on...
52 bool Changed; // Set to true when we change anything.
54 // visitLoop - Hoist expressions out of the specified loop...
55 void visitLoop(Loop *L);
57 // notInCurrentLoop - Little predicate that returns true if the specified
58 // basic block is in a subloop of the current one, not the current one
61 bool notInCurrentLoop(BasicBlock *BB) {
62 for (unsigned i = 0, e = CurLoop->getSubLoops().size(); i != e; ++i)
63 if (CurLoop->getSubLoops()[i]->contains(BB))
64 return true; // A subloop actually contains this block!
68 // hoist - When an instruction is found to only use loop invariant operands
69 // that is safe to hoist, this instruction is called to do the dirty work.
71 void hoist(Instruction &I);
73 // isLoopInvariant - Return true if the specified value is loop invariant
74 inline bool isLoopInvariant(Value *V) {
75 if (Instruction *I = dyn_cast<Instruction>(V))
76 return !CurLoop->contains(I->getParent());
77 return true; // All non-instructions are loop invariant
80 // visitBasicBlock - Run LICM on a particular block.
81 void visitBasicBlock(BasicBlock *BB);
83 // Instruction visitation handlers... these basically control whether or not
84 // the specified instruction types are hoisted.
86 friend class InstVisitor<LICM>;
87 void visitBinaryOperator(Instruction &I) {
88 if (isLoopInvariant(I.getOperand(0)) && isLoopInvariant(I.getOperand(1)))
91 void visitCastInst(CastInst &CI) {
92 Instruction &I = (Instruction&)CI;
93 if (isLoopInvariant(I.getOperand(0))) hoist(I);
95 void visitShiftInst(ShiftInst &I) { visitBinaryOperator((Instruction&)I); }
97 void visitGetElementPtrInst(GetElementPtrInst &GEPI) {
98 Instruction &I = (Instruction&)GEPI;
99 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
100 if (!isLoopInvariant(I.getOperand(i))) return;
105 RegisterOpt<LICM> X("licm", "Loop Invariant Code Motion");
108 Pass *createLICMPass() { return new LICM(); }
110 bool LICM::runOnFunction(Function &) {
111 // get our loop information...
112 const std::vector<Loop*> &TopLevelLoops =
113 getAnalysis<LoopInfo>().getTopLevelLoops();
115 // Traverse loops in postorder, hoisting expressions out of the deepest loops
119 std::for_each(TopLevelLoops.begin(), TopLevelLoops.end(),
120 bind_obj(this, &LICM::visitLoop));
124 void LICM::visitLoop(Loop *L) {
125 // Recurse through all subloops before we process this loop...
126 std::for_each(L->getSubLoops().begin(), L->getSubLoops().end(),
127 bind_obj(this, &LICM::visitLoop));
130 // Calculate the set of predecessors for this loop. The predecessors for this
131 // loop are equal to the predecessors for the header node of the loop that are
132 // not themselves in the loop.
134 BasicBlock *Header = L->getHeader();
136 // Calculate the sets of predecessors and backedges of the loop...
137 LoopBackEdges.insert(LoopBackEdges.end(),pred_begin(Header),pred_end(Header));
139 std::vector<BasicBlock*>::iterator LPI =
140 std::partition(LoopBackEdges.begin(), LoopBackEdges.end(),
141 bind_obj(CurLoop, &Loop::contains));
143 // Move all predecessors to the LoopPreds vector...
144 LoopPreds.insert(LoopPreds.end(), LPI, LoopBackEdges.end());
146 // Remove predecessors from backedges list...
147 LoopBackEdges.erase(LPI, LoopBackEdges.end());
150 // The only way that there could be no predecessors to a loop is if the loop
151 // is not reachable. Since we don't care about optimizing dead loops,
152 // summarily ignore them.
154 if (LoopPreds.empty()) return;
156 // We want to visit all of the instructions in this loop... that are not parts
157 // of our subloops (they have already had their invariants hoisted out of
158 // their loop, into this loop, so there is no need to process the BODIES of
161 std::vector<BasicBlock*> BBs(L->getBlocks().begin(), L->getBlocks().end());
163 // Remove blocks that are actually in subloops...
164 BBs.erase(std::remove_if(BBs.begin(), BBs.end(),
165 bind_obj(this, &LICM::notInCurrentLoop)), BBs.end());
167 // Visit all of the basic blocks we have chosen, hoisting out the instructions
168 // as neccesary. This leaves dead copies of the instruction in the loop
171 for_each(BBs.begin(), BBs.end(), bind_obj(this, &LICM::visitBasicBlock));
173 // Clear out loops state information for the next iteration
176 LoopBackEdges.clear();
179 void LICM::visitBasicBlock(BasicBlock *BB) {
180 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ) {
183 if (dceInstruction(I))
191 void LICM::hoist(Instruction &Inst) {
192 if (Inst.use_empty()) return; // Don't (re) hoist dead instructions!
193 //cerr << "Hoisting " << Inst;
195 BasicBlock *Header = CurLoop->getHeader();
197 // Old instruction will be removed, so take it's name...
198 string InstName = Inst.getName();
201 // The common case is that we have a pre-header. Generate special case code
202 // that is faster if that is the case.
204 if (LoopPreds.size() == 1) {
205 BasicBlock *Pred = LoopPreds[0];
207 // Create a new copy of the instruction, for insertion into Pred.
208 Instruction *New = Inst.clone();
209 New->setName(InstName);
211 // Insert the new node in Pred, before the terminator.
212 Pred->getInstList().insert(--Pred->end(), New);
214 // Kill the old instruction...
215 Inst.replaceAllUsesWith(New);
219 // No loop pre-header, insert a PHI node into header to capture all of the
220 // incoming versions of the value.
222 PHINode *LoopVal = new PHINode(Inst.getType(), InstName+".phi");
224 // Insert the new PHI node into the loop header...
225 Header->getInstList().push_front(LoopVal);
227 // Insert cloned versions of the instruction into all of the loop preds.
228 for (unsigned i = 0, e = LoopPreds.size(); i != e; ++i) {
229 BasicBlock *Pred = LoopPreds[i];
231 // Create a new copy of the instruction, for insertion into Pred.
232 Instruction *New = Inst.clone();
233 New->setName(InstName);
235 // Insert the new node in Pred, before the terminator.
236 Pred->getInstList().insert(--Pred->end(), New);
238 // Add the incoming value to the PHI node.
239 LoopVal->addIncoming(New, Pred);
242 // Add incoming values to the PHI node for all backedges in the loop...
243 for (unsigned i = 0, e = LoopBackEdges.size(); i != e; ++i)
244 LoopVal->addIncoming(LoopVal, LoopBackEdges[i]);
246 // Replace all uses of the old version of the instruction in the loop with
247 // the new version that is out of the loop. We know that this is ok,
248 // because the new definition is in the loop header, which dominates the
249 // entire loop body. The old definition was defined _inside_ of the loop,
250 // so the scope cannot extend outside of the loop, so we're ok.
252 Inst.replaceAllUsesWith(LoopVal);