1 //===-- LoopUnswitch.cpp - Hoist loop-invariant conditionals in loop ------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass transforms loops that contain branches on loop-invariant conditions
11 // to have multiple loops. For example, it turns the left into the right code:
20 // This can increase the size of the code exponentially (doubling it every time
21 // a loop is unswitched) so we only unswitch if the resultant code will be
22 // smaller than a threshold.
24 // This pass expects LICM to be run before it to hoist invariant conditions out
25 // of the loop, to make the unswitching opportunity obvious.
27 //===----------------------------------------------------------------------===//
29 #define DEBUG_TYPE "loop-unswitch"
30 #include "llvm/Transforms/Scalar.h"
31 #include "llvm/Constants.h"
32 #include "llvm/Function.h"
33 #include "llvm/Instructions.h"
34 #include "llvm/Analysis/Dominators.h"
35 #include "llvm/Analysis/LoopInfo.h"
36 #include "llvm/Transforms/Utils/Cloning.h"
37 #include "llvm/Transforms/Utils/Local.h"
38 #include "Support/Debug.h"
39 #include "Support/Statistic.h"
43 Statistic<> NumUnswitched("loop-unswitch", "Number of loops unswitched");
45 class LoopUnswitch : public FunctionPass {
46 LoopInfo *LI; // Loop information
49 virtual bool runOnFunction(Function &F);
50 bool visitLoop(Loop *L);
52 /// This transformation requires natural loop information & requires that
53 /// loop preheaders be inserted into the CFG...
55 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
56 AU.addRequiredID(LoopSimplifyID);
57 //AU.addRequired<DominatorSet>();
58 AU.addRequired<LoopInfo>();
59 AU.addPreserved<LoopInfo>();
63 void VersionLoop(Value *LIC, Loop *L);
64 BasicBlock *SplitBlock(BasicBlock *BB, bool SplitAtTop);
65 void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, bool Val);
67 RegisterOpt<LoopUnswitch> X("loop-unswitch", "Unswitch loops");
70 FunctionPass *createLoopUnswitchPass() { return new LoopUnswitch(); }
72 bool LoopUnswitch::runOnFunction(Function &F) {
74 LI = &getAnalysis<LoopInfo>();
75 DS = 0; //&getAnalysis<DominatorSet>();
77 // Transform all the top-level loops. Copy the loop list so that the child
78 // can update the loop tree if it needs to delete the loop.
79 std::vector<Loop*> SubLoops(LI->begin(), LI->end());
80 for (unsigned i = 0, e = SubLoops.size(); i != e; ++i)
81 Changed |= visitLoop(SubLoops[i]);
86 bool LoopUnswitch::visitLoop(Loop *L) {
89 // Recurse through all subloops before we process this loop. Copy the loop
90 // list so that the child can update the loop tree if it needs to delete the
92 std::vector<Loop*> SubLoops(L->begin(), L->end());
93 for (unsigned i = 0, e = SubLoops.size(); i != e; ++i)
94 Changed |= visitLoop(SubLoops[i]);
96 // Loop over all of the basic blocks in the loop. If we find an interior
97 // block that is branching on a loop-invariant condition, we can unswitch this
99 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
101 TerminatorInst *TI = (*I)->getTerminator();
102 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
103 if (!isa<Constant>(SI) && L->isLoopInvariant(SI->getCondition()))
104 DEBUG(std::cerr << "Can't unswitching 'switch' loop %"
105 << L->getHeader()->getName() << ", cost = "
106 << L->getBlocks().size() << "\n" << **I);
107 } else if (BranchInst *BI = dyn_cast<BranchInst>(TI))
108 if (BI->isConditional() && !isa<Constant>(BI->getCondition()) &&
109 L->isLoopInvariant(BI->getCondition())) {
110 // Check to see if it would be profitable to unswitch this loop.
111 if (L->getBlocks().size() > 10) {
112 DEBUG(std::cerr << "NOT unswitching loop %"
113 << L->getHeader()->getName() << ", cost too high: "
114 << L->getBlocks().size() << "\n");
116 // FIXME: check for profitability.
117 //std::cerr << "BEFORE:\n"; LI->dump();
119 VersionLoop(BI->getCondition(), L);
121 //std::cerr << "AFTER:\n"; LI->dump();
130 /// SplitBlock - Split the specified basic block into two pieces. If SplitAtTop
131 /// is false, this splits the block so the second half only has an unconditional
132 /// branch. If SplitAtTop is true, it makes it so the first half of the block
133 /// only has an unconditional branch in it.
135 /// This method updates the LoopInfo for this function to correctly reflect the
136 /// CFG changes made.
137 BasicBlock *LoopUnswitch::SplitBlock(BasicBlock *BB, bool SplitAtTop) {
138 BasicBlock::iterator SplitPoint;
140 SplitPoint = BB->getTerminator();
142 SplitPoint = BB->begin();
143 while (isa<PHINode>(SplitPoint)) ++SplitPoint;
146 BasicBlock *New = BB->splitBasicBlock(SplitPoint, BB->getName()+".tail");
147 // New now lives in whichever loop that BB used to.
148 if (Loop *L = LI->getLoopFor(BB))
149 L->addBasicBlockToLoop(New, *LI);
150 return SplitAtTop ? BB : New;
154 // RemapInstruction - Convert the instruction operands from referencing the
155 // current values into those specified by ValueMap.
157 static inline void RemapInstruction(Instruction *I,
158 std::map<const Value *, Value*> &ValueMap) {
159 for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) {
160 Value *Op = I->getOperand(op);
161 std::map<const Value *, Value*>::iterator It = ValueMap.find(Op);
162 if (It != ValueMap.end()) Op = It->second;
163 I->setOperand(op, Op);
167 /// CloneLoop - Recursively clone the specified loop and all of its children,
168 /// mapping the blocks with the specified map.
169 static Loop *CloneLoop(Loop *L, Loop *PL, std::map<const Value*, Value*> &VM,
171 Loop *New = new Loop();
174 PL->addChildLoop(New);
176 LI->addTopLevelLoop(New);
178 // Add all of the blocks in L to the new loop.
179 for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
181 if (LI->getLoopFor(*I) == L)
182 New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), *LI);
184 // Add all of the subloops to the new loop.
185 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
186 CloneLoop(*I, New, VM, LI);
192 /// InsertPHINodesForUsesOutsideLoop - If this instruction is used outside of
193 /// the specified loop, insert a PHI node in the appropriate exit block to merge
194 /// the values in the two different loop versions.
196 /// Most values are not used outside of the loop they are defined in, so be
197 /// efficient for this case.
200 InsertPHINodesForUsesOutsideLoop(Instruction *OI, Instruction *NI,
201 DominatorSet &DS, Loop *OL, Loop *NL,
202 std::vector<BasicBlock*> &OldExitBlocks,
203 std::map<const Value*, Value*> &ValueMap) {
204 assert(OI->getType() == NI->getType() && OI->getOpcode() == NI->getOpcode() &&
205 "Hrm, should be mapping between identical instructions!");
206 for (Value::use_iterator UI = OI->use_begin(), E = OI->use_end(); UI != E;
208 if (!OL->contains(cast<Instruction>(*UI)->getParent()) &&
209 !NL->contains(cast<Instruction>(*UI)->getParent()))
210 goto UsedOutsideOfLoop;
214 // Okay, this instruction is used outside of the current loop. Insert a PHI
215 // nodes for the instruction merging the values together.
217 // FIXME: For now we just spill the object to the stack, assuming that a
218 // subsequent mem2reg pass will clean up after us. This should be improved in
220 // 1. If there is only one exit block, trivially insert the PHI nodes
221 // 2. Once we update domfrontier, we should do the promotion after everything
223 AllocaInst *Result = DemoteRegToStack(*OI);
225 // Store to the stack location right after the new instruction.
226 BasicBlock::iterator InsertPoint = NI;
227 if (InvokeInst *II = dyn_cast<InvokeInst>(NI))
228 InsertPoint = II->getNormalDest()->begin();
231 while (isa<PHINode>(InsertPoint)) ++InsertPoint;
232 new StoreInst(NI, Result, InsertPoint);
238 /// VersionLoop - We determined that the loop is profitable to unswitch and
239 /// contains a branch on a loop invariant condition. Split it into loop
240 /// versions and test the condition outside of either loop.
241 void LoopUnswitch::VersionLoop(Value *LIC, Loop *L) {
242 Function *F = L->getHeader()->getParent();
244 DEBUG(std::cerr << "loop-unswitch: Unswitching loop %"
245 << L->getHeader()->getName() << " [" << L->getBlocks().size()
246 << " blocks] in Function " << F->getName()
247 << " on cond:" << *LIC << "\n");
249 std::vector<BasicBlock*> LoopBlocks;
251 // First step, split the preheader and exit blocks, and add these blocks to
252 // the LoopBlocks list.
253 BasicBlock *OrigPreheader = L->getLoopPreheader();
254 LoopBlocks.push_back(SplitBlock(OrigPreheader, false));
256 // We want the loop to come after the preheader, but before the exit blocks.
257 LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
259 std::vector<BasicBlock*> ExitBlocks;
260 L->getExitBlocks(ExitBlocks);
261 std::sort(ExitBlocks.begin(), ExitBlocks.end());
262 ExitBlocks.erase(std::unique(ExitBlocks.begin(), ExitBlocks.end()),
264 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
265 LoopBlocks.push_back(ExitBlocks[i] = SplitBlock(ExitBlocks[i], true));
267 // Next step, clone all of the basic blocks that make up the loop (including
268 // the loop preheader and exit blocks), keeping track of the mapping between
269 // the instructions and blocks.
270 std::vector<BasicBlock*> NewBlocks;
271 NewBlocks.reserve(LoopBlocks.size());
272 std::map<const Value*, Value*> ValueMap;
273 for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
274 NewBlocks.push_back(CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F));
275 ValueMap[LoopBlocks[i]] = NewBlocks.back(); // Keep the BB mapping.
278 // Splice the newly inserted blocks into the function right before the
279 // original preheader.
280 F->getBasicBlockList().splice(LoopBlocks[0], F->getBasicBlockList(),
281 NewBlocks[0], F->end());
283 // Now we create the new Loop object for the versioned loop.
284 Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI);
285 if (Loop *Parent = L->getParentLoop()) {
286 // Make sure to add the cloned preheader and exit blocks to the parent loop
288 Parent->addBasicBlockToLoop(NewBlocks[0], *LI);
289 for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
290 Parent->addBasicBlockToLoop(cast<BasicBlock>(ValueMap[ExitBlocks[i]]),
294 // Rewrite the code to refer to itself.
295 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
296 for (BasicBlock::iterator I = NewBlocks[i]->begin(),
297 E = NewBlocks[i]->end(); I != E; ++I)
298 RemapInstruction(I, ValueMap);
300 // If the instructions are used outside of the loop, insert a PHI node in any
301 // exit blocks dominated by the instruction.
302 for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
303 for (BasicBlock::iterator OI = LoopBlocks[i]->begin(),
304 E = LoopBlocks[i]->end(); OI != E; ++OI)
305 if (!OI->use_empty()) {
306 std::map<const Value*,Value*>::iterator OII = ValueMap.find(OI);
307 // The PHINode rewriting stuff can insert stores that are not in the
308 // mapping. Don't mess around with them.
309 if (OII != ValueMap.end()) {
310 Instruction *NI = cast<Instruction>(OII->second);
311 InsertPHINodesForUsesOutsideLoop(OI, NI, *DS, L, NewLoop,
312 ExitBlocks, ValueMap);
316 // Rewrite the original preheader to select between versions of the loop.
317 assert(isa<BranchInst>(OrigPreheader->getTerminator()) &&
318 cast<BranchInst>(OrigPreheader->getTerminator())->isUnconditional() &&
319 OrigPreheader->getTerminator()->getSuccessor(0) == LoopBlocks[0] &&
320 "Preheader splitting did not work correctly!");
321 // Remove the unconditional branch to LoopBlocks[0].
322 OrigPreheader->getInstList().pop_back();
324 // Insert a conditional branch on LIC to the two preheaders. The original
325 // code is the true version and the new code is the false version.
326 new BranchInst(LoopBlocks[0], NewBlocks[0], LIC, OrigPreheader);
328 // Now we rewrite the original code to know that the condition is true and the
329 // new code to know that the condition is false.
330 RewriteLoopBodyWithConditionConstant(L, LIC, true);
331 RewriteLoopBodyWithConditionConstant(NewLoop, LIC, false);
334 // Try to unswitch each of our new loops now!
339 // RewriteLoopBodyWithConditionConstant - We know that the boolean value LIC has
340 // the value specified by Val in the specified loop. Rewrite any uses of LIC or
341 // of properties correlated to it.
342 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
344 // FIXME: Support correlated properties, like:
350 ConstantBool *BoolVal = ConstantBool::get(Val);
352 std::vector<User*> Users(LIC->use_begin(), LIC->use_end());
353 for (unsigned i = 0, e = Users.size(); i != e; ++i)
354 if (Instruction *U = dyn_cast<Instruction>(Users[i]))
355 if (L->contains(U->getParent()))
356 U->replaceUsesOfWith(LIC, BoolVal);