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/Utils/LCSSA.h"
31#include "llvm/ADT/STLExtras.h"
32#include "llvm/ADT/Statistic.h"
33#include "llvm/Analysis/AliasAnalysis.h"
34#include "llvm/Analysis/BasicAliasAnalysis.h"
35#include "llvm/Analysis/GlobalsModRef.h"
36#include "llvm/Analysis/LoopPass.h"
37#include "llvm/Analysis/ScalarEvolution.h"
38#include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
39#include "llvm/IR/Constants.h"
40#include "llvm/IR/Dominators.h"
41#include "llvm/IR/Function.h"
42#include "llvm/IR/Instructions.h"
43#include "llvm/IR/PredIteratorCache.h"
44#include "llvm/Pass.h"
45#include "llvm/Transforms/Scalar.h"
46#include "llvm/Transforms/Utils/LoopUtils.h"
47#include "llvm/Transforms/Utils/SSAUpdater.h"
48using namespace llvm;
49
50#define DEBUG_TYPE "lcssa"
51
52STATISTIC(NumLCSSA, "Number of live out of a loop variables");
53
54/// Return true if the specified block is in the list.
55static bool isExitBlock(BasicBlock *BB,
56                        const SmallVectorImpl<BasicBlock *> &ExitBlocks) {
57  return find(ExitBlocks, BB) != ExitBlocks.end();
58}
59
60/// Given an instruction in the loop, check to see if it has any uses that are
61/// outside the current loop.  If so, insert LCSSA PHI nodes and rewrite the
62/// uses.
63static bool processInstruction(Loop &L, Instruction &Inst, DominatorTree &DT,
64                               const SmallVectorImpl<BasicBlock *> &ExitBlocks,
65                               PredIteratorCache &PredCache, LoopInfo *LI) {
66  SmallVector<Use *, 16> UsesToRewrite;
67
68  // Tokens cannot be used in PHI nodes, so we skip over them.
69  // We can run into tokens which are live out of a loop with catchswitch
70  // instructions in Windows EH if the catchswitch has one catchpad which
71  // is inside the loop and another which is not.
72  if (Inst.getType()->isTokenTy())
73    return false;
74
75  BasicBlock *InstBB = Inst.getParent();
76
77  for (Use &U : Inst.uses()) {
78    Instruction *User = cast<Instruction>(U.getUser());
79    BasicBlock *UserBB = User->getParent();
80    if (PHINode *PN = dyn_cast<PHINode>(User))
81      UserBB = PN->getIncomingBlock(U);
82
83    if (InstBB != UserBB && !L.contains(UserBB))
84      UsesToRewrite.push_back(&U);
85  }
86
87  // If there are no uses outside the loop, exit with no change.
88  if (UsesToRewrite.empty())
89    return false;
90
91  ++NumLCSSA; // We are applying the transformation
92
93  // Invoke instructions are special in that their result value is not available
94  // along their unwind edge. The code below tests to see whether DomBB
95  // dominates the value, so adjust DomBB to the normal destination block,
96  // which is effectively where the value is first usable.
97  BasicBlock *DomBB = Inst.getParent();
98  if (InvokeInst *Inv = dyn_cast<InvokeInst>(&Inst))
99    DomBB = Inv->getNormalDest();
100
101  DomTreeNode *DomNode = DT.getNode(DomBB);
102
103  SmallVector<PHINode *, 16> AddedPHIs;
104  SmallVector<PHINode *, 8> PostProcessPHIs;
105
106  SSAUpdater SSAUpdate;
107  SSAUpdate.Initialize(Inst.getType(), Inst.getName());
108
109  // Insert the LCSSA phi's into all of the exit blocks dominated by the
110  // value, and add them to the Phi's map.
111  for (BasicBlock *ExitBB : ExitBlocks) {
112    if (!DT.dominates(DomNode, DT.getNode(ExitBB)))
113      continue;
114
115    // If we already inserted something for this BB, don't reprocess it.
116    if (SSAUpdate.HasValueForBlock(ExitBB))
117      continue;
118
119    PHINode *PN = PHINode::Create(Inst.getType(), PredCache.size(ExitBB),
120                                  Inst.getName() + ".lcssa", &ExitBB->front());
121
122    // Add inputs from inside the loop for this PHI.
123    for (BasicBlock *Pred : PredCache.get(ExitBB)) {
124      PN->addIncoming(&Inst, Pred);
125
126      // If the exit block has a predecessor not within the loop, arrange for
127      // the incoming value use corresponding to that predecessor to be
128      // rewritten in terms of a different LCSSA PHI.
129      if (!L.contains(Pred))
130        UsesToRewrite.push_back(
131            &PN->getOperandUse(PN->getOperandNumForIncomingValue(
132                 PN->getNumIncomingValues() - 1)));
133    }
134
135    AddedPHIs.push_back(PN);
136
137    // Remember that this phi makes the value alive in this block.
138    SSAUpdate.AddAvailableValue(ExitBB, PN);
139
140    // LoopSimplify might fail to simplify some loops (e.g. when indirect
141    // branches are involved). In such situations, it might happen that an exit
142    // for Loop L1 is the header of a disjoint Loop L2. Thus, when we create
143    // PHIs in such an exit block, we are also inserting PHIs into L2's header.
144    // This could break LCSSA form for L2 because these inserted PHIs can also
145    // have uses outside of L2. Remember all PHIs in such situation as to
146    // revisit than later on. FIXME: Remove this if indirectbr support into
147    // LoopSimplify gets improved.
148    if (auto *OtherLoop = LI->getLoopFor(ExitBB))
149      if (!L.contains(OtherLoop))
150        PostProcessPHIs.push_back(PN);
151  }
152
153  // Rewrite all uses outside the loop in terms of the new PHIs we just
154  // inserted.
155  for (Use *UseToRewrite : UsesToRewrite) {
156    // If this use is in an exit block, rewrite to use the newly inserted PHI.
157    // This is required for correctness because SSAUpdate doesn't handle uses in
158    // the same block.  It assumes the PHI we inserted is at the end of the
159    // block.
160    Instruction *User = cast<Instruction>(UseToRewrite->getUser());
161    BasicBlock *UserBB = User->getParent();
162    if (PHINode *PN = dyn_cast<PHINode>(User))
163      UserBB = PN->getIncomingBlock(*UseToRewrite);
164
165    if (isa<PHINode>(UserBB->begin()) && isExitBlock(UserBB, ExitBlocks)) {
166      // Tell the VHs that the uses changed. This updates SCEV's caches.
167      if (UseToRewrite->get()->hasValueHandle())
168        ValueHandleBase::ValueIsRAUWd(*UseToRewrite, &UserBB->front());
169      UseToRewrite->set(&UserBB->front());
170      continue;
171    }
172
173    // Otherwise, do full PHI insertion.
174    SSAUpdate.RewriteUse(*UseToRewrite);
175  }
176
177  // Post process PHI instructions that were inserted into another disjoint loop
178  // and update their exits properly.
179  for (auto *I : PostProcessPHIs) {
180    if (I->use_empty())
181      continue;
182
183    BasicBlock *PHIBB = I->getParent();
184    Loop *OtherLoop = LI->getLoopFor(PHIBB);
185    SmallVector<BasicBlock *, 8> EBs;
186    OtherLoop->getExitBlocks(EBs);
187    if (EBs.empty())
188      continue;
189
190    // Recurse and re-process each PHI instruction. FIXME: we should really
191    // convert this entire thing to a worklist approach where we process a
192    // vector of instructions...
193    processInstruction(*OtherLoop, *I, DT, EBs, PredCache, LI);
194  }
195
196  // Remove PHI nodes that did not have any uses rewritten.
197  for (PHINode *PN : AddedPHIs)
198    if (PN->use_empty())
199      PN->eraseFromParent();
200
201  return true;
202}
203
204/// Return true if the specified block dominates at least
205/// one of the blocks in the specified list.
206static bool
207blockDominatesAnExit(BasicBlock *BB,
208                     DominatorTree &DT,
209                     const SmallVectorImpl<BasicBlock *> &ExitBlocks) {
210  DomTreeNode *DomNode = DT.getNode(BB);
211  return llvm::any_of(ExitBlocks, [&](BasicBlock * EB) {
212    return DT.dominates(DomNode, DT.getNode(EB));
213  });
214}
215
216bool llvm::formLCSSA(Loop &L, DominatorTree &DT, LoopInfo *LI,
217                     ScalarEvolution *SE) {
218  bool Changed = false;
219
220  // Get the set of exiting blocks.
221  SmallVector<BasicBlock *, 8> ExitBlocks;
222  L.getExitBlocks(ExitBlocks);
223
224  if (ExitBlocks.empty())
225    return false;
226
227  PredIteratorCache PredCache;
228
229  // Look at all the instructions in the loop, checking to see if they have uses
230  // outside the loop.  If so, rewrite those uses.
231  for (BasicBlock *BB : L.blocks()) {
232    // For large loops, avoid use-scanning by using dominance information:  In
233    // particular, if a block does not dominate any of the loop exits, then none
234    // of the values defined in the block could be used outside the loop.
235    if (!blockDominatesAnExit(BB, DT, ExitBlocks))
236      continue;
237
238    for (Instruction &I : *BB) {
239      // Reject two common cases fast: instructions with no uses (like stores)
240      // and instructions with one use that is in the same block as this.
241      if (I.use_empty() ||
242          (I.hasOneUse() && I.user_back()->getParent() == BB &&
243           !isa<PHINode>(I.user_back())))
244        continue;
245
246      Changed |= processInstruction(L, I, DT, ExitBlocks, PredCache, LI);
247    }
248  }
249
250  // If we modified the code, remove any caches about the loop from SCEV to
251  // avoid dangling entries.
252  // FIXME: This is a big hammer, can we clear the cache more selectively?
253  if (SE && Changed)
254    SE->forgetLoop(&L);
255
256  assert(L.isLCSSAForm(DT));
257
258  return Changed;
259}
260
261/// Process a loop nest depth first.
262bool llvm::formLCSSARecursively(Loop &L, DominatorTree &DT, LoopInfo *LI,
263                                ScalarEvolution *SE) {
264  bool Changed = false;
265
266  // Recurse depth-first through inner loops.
267  for (Loop *SubLoop : L.getSubLoops())
268    Changed |= formLCSSARecursively(*SubLoop, DT, LI, SE);
269
270  Changed |= formLCSSA(L, DT, LI, SE);
271  return Changed;
272}
273
274/// Process all loops in the function, inner-most out.
275static bool formLCSSAOnAllLoops(LoopInfo *LI, DominatorTree &DT,
276                                ScalarEvolution *SE) {
277  bool Changed = false;
278  for (auto &L : *LI)
279    Changed |= formLCSSARecursively(*L, DT, LI, SE);
280  return Changed;
281}
282
283namespace {
284struct LCSSAWrapperPass : public FunctionPass {
285  static char ID; // Pass identification, replacement for typeid
286  LCSSAWrapperPass() : FunctionPass(ID) {
287    initializeLCSSAWrapperPassPass(*PassRegistry::getPassRegistry());
288  }
289
290  // Cached analysis information for the current function.
291  DominatorTree *DT;
292  LoopInfo *LI;
293  ScalarEvolution *SE;
294
295  bool runOnFunction(Function &F) override;
296
297  /// This transformation requires natural loop information & requires that
298  /// loop preheaders be inserted into the CFG.  It maintains both of these,
299  /// as well as the CFG.  It also requires dominator information.
300  void getAnalysisUsage(AnalysisUsage &AU) const override {
301    AU.setPreservesCFG();
302
303    AU.addRequired<DominatorTreeWrapperPass>();
304    AU.addRequired<LoopInfoWrapperPass>();
305    AU.addPreservedID(LoopSimplifyID);
306    AU.addPreserved<AAResultsWrapperPass>();
307    AU.addPreserved<BasicAAWrapperPass>();
308    AU.addPreserved<GlobalsAAWrapperPass>();
309    AU.addPreserved<ScalarEvolutionWrapperPass>();
310    AU.addPreserved<SCEVAAWrapperPass>();
311  }
312};
313}
314
315char LCSSAWrapperPass::ID = 0;
316INITIALIZE_PASS_BEGIN(LCSSAWrapperPass, "lcssa", "Loop-Closed SSA Form Pass",
317                      false, false)
318INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
319INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
320INITIALIZE_PASS_END(LCSSAWrapperPass, "lcssa", "Loop-Closed SSA Form Pass",
321                    false, false)
322
323Pass *llvm::createLCSSAPass() { return new LCSSAWrapperPass(); }
324char &llvm::LCSSAID = LCSSAWrapperPass::ID;
325
326/// Transform \p F into loop-closed SSA form.
327bool LCSSAWrapperPass::runOnFunction(Function &F) {
328  LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
329  DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
330  auto *SEWP = getAnalysisIfAvailable<ScalarEvolutionWrapperPass>();
331  SE = SEWP ? &SEWP->getSE() : nullptr;
332
333  return formLCSSAOnAllLoops(LI, *DT, SE);
334}
335
336PreservedAnalyses LCSSAPass::run(Function &F, AnalysisManager<Function> &AM) {
337  auto &LI = AM.getResult<LoopAnalysis>(F);
338  auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
339  auto *SE = AM.getCachedResult<ScalarEvolutionAnalysis>(F);
340  if (!formLCSSAOnAllLoops(&LI, DT, SE))
341    return PreservedAnalyses::all();
342
343  // FIXME: This should also 'preserve the CFG'.
344  PreservedAnalyses PA;
345  PA.preserve<BasicAA>();
346  PA.preserve<GlobalsAA>();
347  PA.preserve<SCEVAA>();
348  PA.preserve<ScalarEvolutionAnalysis>();
349  return PA;
350}
351