1 //===-- LCSSA.cpp - Convert loops into loop-closed SSA form ---------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
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:
13 //
14 // for (...)                for (...)
15 //   if (c)                   if (c)
16 //     X1 = ...                 X1 = ...
17 //   else                     else
18 //     X2 = ...                 X2 = ...
19 //   X3 = phi(X1, X2)         X3 = phi(X1, X2)
20 // ... = X3 + 4             X4 = phi(X3)
21 //                          ... = X4 + 4
22 //
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.
27 //
28 //===----------------------------------------------------------------------===//
29 
30 #include "llvm/Transforms/Scalar.h"
31 #include "llvm/ADT/STLExtras.h"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/Analysis/AliasAnalysis.h"
34 #include "llvm/Analysis/LoopPass.h"
35 #include "llvm/Analysis/ScalarEvolution.h"
36 #include "llvm/IR/Constants.h"
37 #include "llvm/IR/Dominators.h"
38 #include "llvm/IR/Function.h"
39 #include "llvm/IR/Instructions.h"
40 #include "llvm/IR/PredIteratorCache.h"
41 #include "llvm/Pass.h"
42 #include "llvm/Transforms/Utils/LoopUtils.h"
43 #include "llvm/Transforms/Utils/SSAUpdater.h"
44 using namespace llvm;
45 
46 #define DEBUG_TYPE "lcssa"
47 
48 STATISTIC(NumLCSSA, "Number of live out of a loop variables");
49 
50 /// Return true if the specified block is in the list.
isExitBlock(BasicBlock * BB,const SmallVectorImpl<BasicBlock * > & ExitBlocks)51 static bool isExitBlock(BasicBlock *BB,
52                         const SmallVectorImpl<BasicBlock *> &ExitBlocks) {
53   for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
54     if (ExitBlocks[i] == BB)
55       return true;
56   return false;
57 }
58 
59 /// Given an instruction in the loop, check to see if it has any uses that are
60 /// outside the current loop.  If so, insert LCSSA PHI nodes and rewrite the
61 /// uses.
processInstruction(Loop & L,Instruction & Inst,DominatorTree & DT,const SmallVectorImpl<BasicBlock * > & ExitBlocks,PredIteratorCache & PredCache,LoopInfo * LI)62 static bool processInstruction(Loop &L, Instruction &Inst, DominatorTree &DT,
63                                const SmallVectorImpl<BasicBlock *> &ExitBlocks,
64                                PredIteratorCache &PredCache, LoopInfo *LI) {
65   SmallVector<Use *, 16> UsesToRewrite;
66 
67   BasicBlock *InstBB = Inst.getParent();
68 
69   for (Use &U : Inst.uses()) {
70     Instruction *User = cast<Instruction>(U.getUser());
71     BasicBlock *UserBB = User->getParent();
72     if (PHINode *PN = dyn_cast<PHINode>(User))
73       UserBB = PN->getIncomingBlock(U);
74 
75     if (InstBB != UserBB && !L.contains(UserBB))
76       UsesToRewrite.push_back(&U);
77   }
78 
79   // If there are no uses outside the loop, exit with no change.
80   if (UsesToRewrite.empty())
81     return false;
82 
83   ++NumLCSSA; // We are applying the transformation
84 
85   // Invoke instructions are special in that their result value is not available
86   // along their unwind edge. The code below tests to see whether DomBB
87   // dominates
88   // the value, so adjust DomBB to the normal destination block, which is
89   // effectively where the value is first usable.
90   BasicBlock *DomBB = Inst.getParent();
91   if (InvokeInst *Inv = dyn_cast<InvokeInst>(&Inst))
92     DomBB = Inv->getNormalDest();
93 
94   DomTreeNode *DomNode = DT.getNode(DomBB);
95 
96   SmallVector<PHINode *, 16> AddedPHIs;
97   SmallVector<PHINode *, 8> PostProcessPHIs;
98 
99   SSAUpdater SSAUpdate;
100   SSAUpdate.Initialize(Inst.getType(), Inst.getName());
101 
102   // Insert the LCSSA phi's into all of the exit blocks dominated by the
103   // value, and add them to the Phi's map.
104   for (SmallVectorImpl<BasicBlock *>::const_iterator BBI = ExitBlocks.begin(),
105                                                      BBE = ExitBlocks.end();
106        BBI != BBE; ++BBI) {
107     BasicBlock *ExitBB = *BBI;
108     if (!DT.dominates(DomNode, DT.getNode(ExitBB)))
109       continue;
110 
111     // If we already inserted something for this BB, don't reprocess it.
112     if (SSAUpdate.HasValueForBlock(ExitBB))
113       continue;
114 
115     PHINode *PN = PHINode::Create(Inst.getType(), PredCache.GetNumPreds(ExitBB),
116                                   Inst.getName() + ".lcssa", ExitBB->begin());
117 
118     // Add inputs from inside the loop for this PHI.
119     for (BasicBlock **PI = PredCache.GetPreds(ExitBB); *PI; ++PI) {
120       PN->addIncoming(&Inst, *PI);
121 
122       // If the exit block has a predecessor not within the loop, arrange for
123       // the incoming value use corresponding to that predecessor to be
124       // rewritten in terms of a different LCSSA PHI.
125       if (!L.contains(*PI))
126         UsesToRewrite.push_back(
127             &PN->getOperandUse(PN->getOperandNumForIncomingValue(
128                  PN->getNumIncomingValues() - 1)));
129     }
130 
131     AddedPHIs.push_back(PN);
132 
133     // Remember that this phi makes the value alive in this block.
134     SSAUpdate.AddAvailableValue(ExitBB, PN);
135 
136     // LoopSimplify might fail to simplify some loops (e.g. when indirect
137     // branches are involved). In such situations, it might happen that an exit
138     // for Loop L1 is the header of a disjoint Loop L2. Thus, when we create
139     // PHIs in such an exit block, we are also inserting PHIs into L2's header.
140     // This could break LCSSA form for L2 because these inserted PHIs can also
141     // have uses outside of L2. Remember all PHIs in such situation as to
142     // revisit than later on. FIXME: Remove this if indirectbr support into
143     // LoopSimplify gets improved.
144     if (auto *OtherLoop = LI->getLoopFor(ExitBB))
145       if (!L.contains(OtherLoop))
146         PostProcessPHIs.push_back(PN);
147   }
148 
149   // Rewrite all uses outside the loop in terms of the new PHIs we just
150   // inserted.
151   for (unsigned i = 0, e = UsesToRewrite.size(); i != e; ++i) {
152     // If this use is in an exit block, rewrite to use the newly inserted PHI.
153     // This is required for correctness because SSAUpdate doesn't handle uses in
154     // the same block.  It assumes the PHI we inserted is at the end of the
155     // block.
156     Instruction *User = cast<Instruction>(UsesToRewrite[i]->getUser());
157     BasicBlock *UserBB = User->getParent();
158     if (PHINode *PN = dyn_cast<PHINode>(User))
159       UserBB = PN->getIncomingBlock(*UsesToRewrite[i]);
160 
161     if (isa<PHINode>(UserBB->begin()) && isExitBlock(UserBB, ExitBlocks)) {
162       // Tell the VHs that the uses changed. This updates SCEV's caches.
163       if (UsesToRewrite[i]->get()->hasValueHandle())
164         ValueHandleBase::ValueIsRAUWd(*UsesToRewrite[i], UserBB->begin());
165       UsesToRewrite[i]->set(UserBB->begin());
166       continue;
167     }
168 
169     // Otherwise, do full PHI insertion.
170     SSAUpdate.RewriteUse(*UsesToRewrite[i]);
171   }
172 
173   // Post process PHI instructions that were inserted into another disjoint loop
174   // and update their exits properly.
175   for (auto *I : PostProcessPHIs) {
176     if (I->use_empty())
177       continue;
178 
179     BasicBlock *PHIBB = I->getParent();
180     Loop *OtherLoop = LI->getLoopFor(PHIBB);
181     SmallVector<BasicBlock *, 8> EBs;
182     OtherLoop->getExitBlocks(EBs);
183     if (EBs.empty())
184       continue;
185 
186     // Recurse and re-process each PHI instruction. FIXME: we should really
187     // convert this entire thing to a worklist approach where we process a
188     // vector of instructions...
189     processInstruction(*OtherLoop, *I, DT, EBs, PredCache, LI);
190   }
191 
192   // Remove PHI nodes that did not have any uses rewritten.
193   for (unsigned i = 0, e = AddedPHIs.size(); i != e; ++i) {
194     if (AddedPHIs[i]->use_empty())
195       AddedPHIs[i]->eraseFromParent();
196   }
197 
198   return true;
199 }
200 
201 /// Return true if the specified block dominates at least
202 /// one of the blocks in the specified list.
203 static bool
blockDominatesAnExit(BasicBlock * BB,DominatorTree & DT,const SmallVectorImpl<BasicBlock * > & ExitBlocks)204 blockDominatesAnExit(BasicBlock *BB,
205                      DominatorTree &DT,
206                      const SmallVectorImpl<BasicBlock *> &ExitBlocks) {
207   DomTreeNode *DomNode = DT.getNode(BB);
208   for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i)
209     if (DT.dominates(DomNode, DT.getNode(ExitBlocks[i])))
210       return true;
211 
212   return false;
213 }
214 
formLCSSA(Loop & L,DominatorTree & DT,LoopInfo * LI,ScalarEvolution * SE)215 bool llvm::formLCSSA(Loop &L, DominatorTree &DT, LoopInfo *LI,
216                      ScalarEvolution *SE) {
217   bool Changed = false;
218 
219   // Get the set of exiting blocks.
220   SmallVector<BasicBlock *, 8> ExitBlocks;
221   L.getExitBlocks(ExitBlocks);
222 
223   if (ExitBlocks.empty())
224     return false;
225 
226   PredIteratorCache PredCache;
227 
228   // Look at all the instructions in the loop, checking to see if they have uses
229   // outside the loop.  If so, rewrite those uses.
230   for (Loop::block_iterator BBI = L.block_begin(), BBE = L.block_end();
231        BBI != BBE; ++BBI) {
232     BasicBlock *BB = *BBI;
233 
234     // For large loops, avoid use-scanning by using dominance information:  In
235     // particular, if a block does not dominate any of the loop exits, then none
236     // of the values defined in the block could be used outside the loop.
237     if (!blockDominatesAnExit(BB, DT, ExitBlocks))
238       continue;
239 
240     for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
241       // Reject two common cases fast: instructions with no uses (like stores)
242       // and instructions with one use that is in the same block as this.
243       if (I->use_empty() ||
244           (I->hasOneUse() && I->user_back()->getParent() == BB &&
245            !isa<PHINode>(I->user_back())))
246         continue;
247 
248       Changed |= processInstruction(L, *I, DT, ExitBlocks, PredCache, LI);
249     }
250   }
251 
252   // If we modified the code, remove any caches about the loop from SCEV to
253   // avoid dangling entries.
254   // FIXME: This is a big hammer, can we clear the cache more selectively?
255   if (SE && Changed)
256     SE->forgetLoop(&L);
257 
258   assert(L.isLCSSAForm(DT));
259 
260   return Changed;
261 }
262 
263 /// Process a loop nest depth first.
formLCSSARecursively(Loop & L,DominatorTree & DT,LoopInfo * LI,ScalarEvolution * SE)264 bool llvm::formLCSSARecursively(Loop &L, DominatorTree &DT, LoopInfo *LI,
265                                 ScalarEvolution *SE) {
266   bool Changed = false;
267 
268   // Recurse depth-first through inner loops.
269   for (Loop::iterator I = L.begin(), E = L.end(); I != E; ++I)
270     Changed |= formLCSSARecursively(**I, DT, LI, SE);
271 
272   Changed |= formLCSSA(L, DT, LI, SE);
273   return Changed;
274 }
275 
276 namespace {
277 struct LCSSA : public FunctionPass {
278   static char ID; // Pass identification, replacement for typeid
LCSSA__anon74a42f660111::LCSSA279   LCSSA() : FunctionPass(ID) {
280     initializeLCSSAPass(*PassRegistry::getPassRegistry());
281   }
282 
283   // Cached analysis information for the current function.
284   DominatorTree *DT;
285   LoopInfo *LI;
286   ScalarEvolution *SE;
287 
288   bool runOnFunction(Function &F) override;
289 
290   /// This transformation requires natural loop information & requires that
291   /// loop preheaders be inserted into the CFG.  It maintains both of these,
292   /// as well as the CFG.  It also requires dominator information.
getAnalysisUsage__anon74a42f660111::LCSSA293   void getAnalysisUsage(AnalysisUsage &AU) const override {
294     AU.setPreservesCFG();
295 
296     AU.addRequired<DominatorTreeWrapperPass>();
297     AU.addRequired<LoopInfoWrapperPass>();
298     AU.addPreservedID(LoopSimplifyID);
299     AU.addPreserved<AliasAnalysis>();
300     AU.addPreserved<ScalarEvolution>();
301   }
302 
303 private:
304   void verifyAnalysis() const override;
305 };
306 }
307 
308 char LCSSA::ID = 0;
309 INITIALIZE_PASS_BEGIN(LCSSA, "lcssa", "Loop-Closed SSA Form Pass", false, false)
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)310 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
311 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
312 INITIALIZE_PASS_END(LCSSA, "lcssa", "Loop-Closed SSA Form Pass", false, false)
313 
314 Pass *llvm::createLCSSAPass() { return new LCSSA(); }
315 char &llvm::LCSSAID = LCSSA::ID;
316 
317 
318 /// Process all loops in the function, inner-most out.
runOnFunction(Function & F)319 bool LCSSA::runOnFunction(Function &F) {
320   bool Changed = false;
321   LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
322   DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
323   SE = getAnalysisIfAvailable<ScalarEvolution>();
324 
325   // Simplify each loop nest in the function.
326   for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
327     Changed |= formLCSSARecursively(**I, *DT, LI, SE);
328 
329   return Changed;
330 }
331 
verifyLoop(Loop & L,DominatorTree & DT)332 static void verifyLoop(Loop &L, DominatorTree &DT) {
333   // Recurse depth-first through inner loops.
334   for (Loop::iterator LI = L.begin(), LE = L.end(); LI != LE; ++LI)
335     verifyLoop(**LI, DT);
336 
337   // Check the special guarantees that LCSSA makes.
338   //assert(L.isLCSSAForm(DT) && "LCSSA form not preserved!");
339 }
340 
verifyAnalysis() const341 void LCSSA::verifyAnalysis() const {
342   // Verify each loop nest in the function, assuming LI still points at that
343   // function's loop info.
344   for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I)
345     verifyLoop(**I, *DT);
346 }
347