1 //===- LowerSwitch.cpp - Eliminate Switch instructions --------------------===//
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 // The LowerSwitch transformation rewrites switch instructions with a sequence
11 // of branches, which allows targets to get away with not implementing the
12 // switch instruction until it is convenient.
13 //
14 //===----------------------------------------------------------------------===//
15
16 #include "llvm/Transforms/Scalar.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/IR/CFG.h"
19 #include "llvm/IR/Constants.h"
20 #include "llvm/IR/Function.h"
21 #include "llvm/IR/Instructions.h"
22 #include "llvm/IR/LLVMContext.h"
23 #include "llvm/Pass.h"
24 #include "llvm/Support/Compiler.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/Support/raw_ostream.h"
27 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
28 #include "llvm/Transforms/Utils/UnifyFunctionExitNodes.h"
29 #include <algorithm>
30 using namespace llvm;
31
32 #define DEBUG_TYPE "lower-switch"
33
34 namespace {
35 struct IntRange {
36 int64_t Low, High;
37 };
38 // Return true iff R is covered by Ranges.
IsInRanges(const IntRange & R,const std::vector<IntRange> & Ranges)39 static bool IsInRanges(const IntRange &R,
40 const std::vector<IntRange> &Ranges) {
41 // Note: Ranges must be sorted, non-overlapping and non-adjacent.
42
43 // Find the first range whose High field is >= R.High,
44 // then check if the Low field is <= R.Low. If so, we
45 // have a Range that covers R.
46 auto I = std::lower_bound(
47 Ranges.begin(), Ranges.end(), R,
48 [](const IntRange &A, const IntRange &B) { return A.High < B.High; });
49 return I != Ranges.end() && I->Low <= R.Low;
50 }
51
52 /// Replace all SwitchInst instructions with chained branch instructions.
53 class LowerSwitch : public FunctionPass {
54 public:
55 static char ID; // Pass identification, replacement for typeid
LowerSwitch()56 LowerSwitch() : FunctionPass(ID) {
57 initializeLowerSwitchPass(*PassRegistry::getPassRegistry());
58 }
59
60 bool runOnFunction(Function &F) override;
61
getAnalysisUsage(AnalysisUsage & AU) const62 void getAnalysisUsage(AnalysisUsage &AU) const override {
63 // This is a cluster of orthogonal Transforms
64 AU.addPreserved<UnifyFunctionExitNodes>();
65 AU.addPreservedID(LowerInvokePassID);
66 }
67
68 struct CaseRange {
69 ConstantInt* Low;
70 ConstantInt* High;
71 BasicBlock* BB;
72
CaseRange__anon1982c9500111::LowerSwitch::CaseRange73 CaseRange(ConstantInt *low, ConstantInt *high, BasicBlock *bb)
74 : Low(low), High(high), BB(bb) {}
75 };
76
77 typedef std::vector<CaseRange> CaseVector;
78 typedef std::vector<CaseRange>::iterator CaseItr;
79 private:
80 void processSwitchInst(SwitchInst *SI, SmallPtrSetImpl<BasicBlock*> &DeleteList);
81
82 BasicBlock *switchConvert(CaseItr Begin, CaseItr End,
83 ConstantInt *LowerBound, ConstantInt *UpperBound,
84 Value *Val, BasicBlock *Predecessor,
85 BasicBlock *OrigBlock, BasicBlock *Default,
86 const std::vector<IntRange> &UnreachableRanges);
87 BasicBlock *newLeafBlock(CaseRange &Leaf, Value *Val, BasicBlock *OrigBlock,
88 BasicBlock *Default);
89 unsigned Clusterify(CaseVector &Cases, SwitchInst *SI);
90 };
91
92 /// The comparison function for sorting the switch case values in the vector.
93 /// WARNING: Case ranges should be disjoint!
94 struct CaseCmp {
operator ()__anon1982c9500111::CaseCmp95 bool operator () (const LowerSwitch::CaseRange& C1,
96 const LowerSwitch::CaseRange& C2) {
97
98 const ConstantInt* CI1 = cast<const ConstantInt>(C1.Low);
99 const ConstantInt* CI2 = cast<const ConstantInt>(C2.High);
100 return CI1->getValue().slt(CI2->getValue());
101 }
102 };
103 }
104
105 char LowerSwitch::ID = 0;
106 INITIALIZE_PASS(LowerSwitch, "lowerswitch",
107 "Lower SwitchInst's to branches", false, false)
108
109 // Publicly exposed interface to pass...
110 char &llvm::LowerSwitchID = LowerSwitch::ID;
111 // createLowerSwitchPass - Interface to this file...
createLowerSwitchPass()112 FunctionPass *llvm::createLowerSwitchPass() {
113 return new LowerSwitch();
114 }
115
runOnFunction(Function & F)116 bool LowerSwitch::runOnFunction(Function &F) {
117 bool Changed = false;
118 SmallPtrSet<BasicBlock*, 8> DeleteList;
119
120 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
121 BasicBlock *Cur = &*I++; // Advance over block so we don't traverse new blocks
122
123 // If the block is a dead Default block that will be deleted later, don't
124 // waste time processing it.
125 if (DeleteList.count(Cur))
126 continue;
127
128 if (SwitchInst *SI = dyn_cast<SwitchInst>(Cur->getTerminator())) {
129 Changed = true;
130 processSwitchInst(SI, DeleteList);
131 }
132 }
133
134 for (BasicBlock* BB: DeleteList) {
135 DeleteDeadBlock(BB);
136 }
137
138 return Changed;
139 }
140
141 /// Used for debugging purposes.
142 static raw_ostream& operator<<(raw_ostream &O,
143 const LowerSwitch::CaseVector &C)
144 LLVM_ATTRIBUTE_USED;
operator <<(raw_ostream & O,const LowerSwitch::CaseVector & C)145 static raw_ostream& operator<<(raw_ostream &O,
146 const LowerSwitch::CaseVector &C) {
147 O << "[";
148
149 for (LowerSwitch::CaseVector::const_iterator B = C.begin(),
150 E = C.end(); B != E; ) {
151 O << *B->Low << " -" << *B->High;
152 if (++B != E) O << ", ";
153 }
154
155 return O << "]";
156 }
157
158 /// \brief Update the first occurrence of the "switch statement" BB in the PHI
159 /// node with the "new" BB. The other occurrences will:
160 ///
161 /// 1) Be updated by subsequent calls to this function. Switch statements may
162 /// have more than one outcoming edge into the same BB if they all have the same
163 /// value. When the switch statement is converted these incoming edges are now
164 /// coming from multiple BBs.
165 /// 2) Removed if subsequent incoming values now share the same case, i.e.,
166 /// multiple outcome edges are condensed into one. This is necessary to keep the
167 /// number of phi values equal to the number of branches to SuccBB.
fixPhis(BasicBlock * SuccBB,BasicBlock * OrigBB,BasicBlock * NewBB,unsigned NumMergedCases)168 static void fixPhis(BasicBlock *SuccBB, BasicBlock *OrigBB, BasicBlock *NewBB,
169 unsigned NumMergedCases) {
170 for (BasicBlock::iterator I = SuccBB->begin(),
171 IE = SuccBB->getFirstNonPHI()->getIterator();
172 I != IE; ++I) {
173 PHINode *PN = cast<PHINode>(I);
174
175 // Only update the first occurrence.
176 unsigned Idx = 0, E = PN->getNumIncomingValues();
177 unsigned LocalNumMergedCases = NumMergedCases;
178 for (; Idx != E; ++Idx) {
179 if (PN->getIncomingBlock(Idx) == OrigBB) {
180 PN->setIncomingBlock(Idx, NewBB);
181 break;
182 }
183 }
184
185 // Remove additional occurrences coming from condensed cases and keep the
186 // number of incoming values equal to the number of branches to SuccBB.
187 SmallVector<unsigned, 8> Indices;
188 for (++Idx; LocalNumMergedCases > 0 && Idx < E; ++Idx)
189 if (PN->getIncomingBlock(Idx) == OrigBB) {
190 Indices.push_back(Idx);
191 LocalNumMergedCases--;
192 }
193 // Remove incoming values in the reverse order to prevent invalidating
194 // *successive* index.
195 for (auto III = Indices.rbegin(), IIE = Indices.rend(); III != IIE; ++III)
196 PN->removeIncomingValue(*III);
197 }
198 }
199
200 /// Convert the switch statement into a binary lookup of the case values.
201 /// The function recursively builds this tree. LowerBound and UpperBound are
202 /// used to keep track of the bounds for Val that have already been checked by
203 /// a block emitted by one of the previous calls to switchConvert in the call
204 /// stack.
205 BasicBlock *
switchConvert(CaseItr Begin,CaseItr End,ConstantInt * LowerBound,ConstantInt * UpperBound,Value * Val,BasicBlock * Predecessor,BasicBlock * OrigBlock,BasicBlock * Default,const std::vector<IntRange> & UnreachableRanges)206 LowerSwitch::switchConvert(CaseItr Begin, CaseItr End, ConstantInt *LowerBound,
207 ConstantInt *UpperBound, Value *Val,
208 BasicBlock *Predecessor, BasicBlock *OrigBlock,
209 BasicBlock *Default,
210 const std::vector<IntRange> &UnreachableRanges) {
211 unsigned Size = End - Begin;
212
213 if (Size == 1) {
214 // Check if the Case Range is perfectly squeezed in between
215 // already checked Upper and Lower bounds. If it is then we can avoid
216 // emitting the code that checks if the value actually falls in the range
217 // because the bounds already tell us so.
218 if (Begin->Low == LowerBound && Begin->High == UpperBound) {
219 unsigned NumMergedCases = 0;
220 if (LowerBound && UpperBound)
221 NumMergedCases =
222 UpperBound->getSExtValue() - LowerBound->getSExtValue();
223 fixPhis(Begin->BB, OrigBlock, Predecessor, NumMergedCases);
224 return Begin->BB;
225 }
226 return newLeafBlock(*Begin, Val, OrigBlock, Default);
227 }
228
229 unsigned Mid = Size / 2;
230 std::vector<CaseRange> LHS(Begin, Begin + Mid);
231 DEBUG(dbgs() << "LHS: " << LHS << "\n");
232 std::vector<CaseRange> RHS(Begin + Mid, End);
233 DEBUG(dbgs() << "RHS: " << RHS << "\n");
234
235 CaseRange &Pivot = *(Begin + Mid);
236 DEBUG(dbgs() << "Pivot ==> "
237 << Pivot.Low->getValue()
238 << " -" << Pivot.High->getValue() << "\n");
239
240 // NewLowerBound here should never be the integer minimal value.
241 // This is because it is computed from a case range that is never
242 // the smallest, so there is always a case range that has at least
243 // a smaller value.
244 ConstantInt *NewLowerBound = Pivot.Low;
245
246 // Because NewLowerBound is never the smallest representable integer
247 // it is safe here to subtract one.
248 ConstantInt *NewUpperBound = ConstantInt::get(NewLowerBound->getContext(),
249 NewLowerBound->getValue() - 1);
250
251 if (!UnreachableRanges.empty()) {
252 // Check if the gap between LHS's highest and NewLowerBound is unreachable.
253 int64_t GapLow = LHS.back().High->getSExtValue() + 1;
254 int64_t GapHigh = NewLowerBound->getSExtValue() - 1;
255 IntRange Gap = { GapLow, GapHigh };
256 if (GapHigh >= GapLow && IsInRanges(Gap, UnreachableRanges))
257 NewUpperBound = LHS.back().High;
258 }
259
260 DEBUG(dbgs() << "LHS Bounds ==> ";
261 if (LowerBound) {
262 dbgs() << LowerBound->getSExtValue();
263 } else {
264 dbgs() << "NONE";
265 }
266 dbgs() << " - " << NewUpperBound->getSExtValue() << "\n";
267 dbgs() << "RHS Bounds ==> ";
268 dbgs() << NewLowerBound->getSExtValue() << " - ";
269 if (UpperBound) {
270 dbgs() << UpperBound->getSExtValue() << "\n";
271 } else {
272 dbgs() << "NONE\n";
273 });
274
275 // Create a new node that checks if the value is < pivot. Go to the
276 // left branch if it is and right branch if not.
277 Function* F = OrigBlock->getParent();
278 BasicBlock* NewNode = BasicBlock::Create(Val->getContext(), "NodeBlock");
279
280 ICmpInst* Comp = new ICmpInst(ICmpInst::ICMP_SLT,
281 Val, Pivot.Low, "Pivot");
282
283 BasicBlock *LBranch = switchConvert(LHS.begin(), LHS.end(), LowerBound,
284 NewUpperBound, Val, NewNode, OrigBlock,
285 Default, UnreachableRanges);
286 BasicBlock *RBranch = switchConvert(RHS.begin(), RHS.end(), NewLowerBound,
287 UpperBound, Val, NewNode, OrigBlock,
288 Default, UnreachableRanges);
289
290 F->getBasicBlockList().insert(++OrigBlock->getIterator(), NewNode);
291 NewNode->getInstList().push_back(Comp);
292
293 BranchInst::Create(LBranch, RBranch, Comp, NewNode);
294 return NewNode;
295 }
296
297 /// Create a new leaf block for the binary lookup tree. It checks if the
298 /// switch's value == the case's value. If not, then it jumps to the default
299 /// branch. At this point in the tree, the value can't be another valid case
300 /// value, so the jump to the "default" branch is warranted.
newLeafBlock(CaseRange & Leaf,Value * Val,BasicBlock * OrigBlock,BasicBlock * Default)301 BasicBlock* LowerSwitch::newLeafBlock(CaseRange& Leaf, Value* Val,
302 BasicBlock* OrigBlock,
303 BasicBlock* Default)
304 {
305 Function* F = OrigBlock->getParent();
306 BasicBlock* NewLeaf = BasicBlock::Create(Val->getContext(), "LeafBlock");
307 F->getBasicBlockList().insert(++OrigBlock->getIterator(), NewLeaf);
308
309 // Emit comparison
310 ICmpInst* Comp = nullptr;
311 if (Leaf.Low == Leaf.High) {
312 // Make the seteq instruction...
313 Comp = new ICmpInst(*NewLeaf, ICmpInst::ICMP_EQ, Val,
314 Leaf.Low, "SwitchLeaf");
315 } else {
316 // Make range comparison
317 if (Leaf.Low->isMinValue(true /*isSigned*/)) {
318 // Val >= Min && Val <= Hi --> Val <= Hi
319 Comp = new ICmpInst(*NewLeaf, ICmpInst::ICMP_SLE, Val, Leaf.High,
320 "SwitchLeaf");
321 } else if (Leaf.Low->isZero()) {
322 // Val >= 0 && Val <= Hi --> Val <=u Hi
323 Comp = new ICmpInst(*NewLeaf, ICmpInst::ICMP_ULE, Val, Leaf.High,
324 "SwitchLeaf");
325 } else {
326 // Emit V-Lo <=u Hi-Lo
327 Constant* NegLo = ConstantExpr::getNeg(Leaf.Low);
328 Instruction* Add = BinaryOperator::CreateAdd(Val, NegLo,
329 Val->getName()+".off",
330 NewLeaf);
331 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Leaf.High);
332 Comp = new ICmpInst(*NewLeaf, ICmpInst::ICMP_ULE, Add, UpperBound,
333 "SwitchLeaf");
334 }
335 }
336
337 // Make the conditional branch...
338 BasicBlock* Succ = Leaf.BB;
339 BranchInst::Create(Succ, Default, Comp, NewLeaf);
340
341 // If there were any PHI nodes in this successor, rewrite one entry
342 // from OrigBlock to come from NewLeaf.
343 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) {
344 PHINode* PN = cast<PHINode>(I);
345 // Remove all but one incoming entries from the cluster
346 uint64_t Range = Leaf.High->getSExtValue() -
347 Leaf.Low->getSExtValue();
348 for (uint64_t j = 0; j < Range; ++j) {
349 PN->removeIncomingValue(OrigBlock);
350 }
351
352 int BlockIdx = PN->getBasicBlockIndex(OrigBlock);
353 assert(BlockIdx != -1 && "Switch didn't go to this successor??");
354 PN->setIncomingBlock((unsigned)BlockIdx, NewLeaf);
355 }
356
357 return NewLeaf;
358 }
359
360 /// Transform simple list of Cases into list of CaseRange's.
Clusterify(CaseVector & Cases,SwitchInst * SI)361 unsigned LowerSwitch::Clusterify(CaseVector& Cases, SwitchInst *SI) {
362 unsigned numCmps = 0;
363
364 // Start with "simple" cases
365 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); i != e; ++i)
366 Cases.push_back(CaseRange(i.getCaseValue(), i.getCaseValue(),
367 i.getCaseSuccessor()));
368
369 std::sort(Cases.begin(), Cases.end(), CaseCmp());
370
371 // Merge case into clusters
372 if (Cases.size() >= 2) {
373 CaseItr I = Cases.begin();
374 for (CaseItr J = std::next(I), E = Cases.end(); J != E; ++J) {
375 int64_t nextValue = J->Low->getSExtValue();
376 int64_t currentValue = I->High->getSExtValue();
377 BasicBlock* nextBB = J->BB;
378 BasicBlock* currentBB = I->BB;
379
380 // If the two neighboring cases go to the same destination, merge them
381 // into a single case.
382 assert(nextValue > currentValue && "Cases should be strictly ascending");
383 if ((nextValue == currentValue + 1) && (currentBB == nextBB)) {
384 I->High = J->High;
385 // FIXME: Combine branch weights.
386 } else if (++I != J) {
387 *I = *J;
388 }
389 }
390 Cases.erase(std::next(I), Cases.end());
391 }
392
393 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) {
394 if (I->Low != I->High)
395 // A range counts double, since it requires two compares.
396 ++numCmps;
397 }
398
399 return numCmps;
400 }
401
402 /// Replace the specified switch instruction with a sequence of chained if-then
403 /// insts in a balanced binary search.
processSwitchInst(SwitchInst * SI,SmallPtrSetImpl<BasicBlock * > & DeleteList)404 void LowerSwitch::processSwitchInst(SwitchInst *SI,
405 SmallPtrSetImpl<BasicBlock*> &DeleteList) {
406 BasicBlock *CurBlock = SI->getParent();
407 BasicBlock *OrigBlock = CurBlock;
408 Function *F = CurBlock->getParent();
409 Value *Val = SI->getCondition(); // The value we are switching on...
410 BasicBlock* Default = SI->getDefaultDest();
411
412 // If there is only the default destination, just branch.
413 if (!SI->getNumCases()) {
414 BranchInst::Create(Default, CurBlock);
415 SI->eraseFromParent();
416 return;
417 }
418
419 // Prepare cases vector.
420 CaseVector Cases;
421 unsigned numCmps = Clusterify(Cases, SI);
422 DEBUG(dbgs() << "Clusterify finished. Total clusters: " << Cases.size()
423 << ". Total compares: " << numCmps << "\n");
424 DEBUG(dbgs() << "Cases: " << Cases << "\n");
425 (void)numCmps;
426
427 ConstantInt *LowerBound = nullptr;
428 ConstantInt *UpperBound = nullptr;
429 std::vector<IntRange> UnreachableRanges;
430
431 if (isa<UnreachableInst>(Default->getFirstNonPHIOrDbg())) {
432 // Make the bounds tightly fitted around the case value range, because we
433 // know that the value passed to the switch must be exactly one of the case
434 // values.
435 assert(!Cases.empty());
436 LowerBound = Cases.front().Low;
437 UpperBound = Cases.back().High;
438
439 DenseMap<BasicBlock *, unsigned> Popularity;
440 unsigned MaxPop = 0;
441 BasicBlock *PopSucc = nullptr;
442
443 IntRange R = { INT64_MIN, INT64_MAX };
444 UnreachableRanges.push_back(R);
445 for (const auto &I : Cases) {
446 int64_t Low = I.Low->getSExtValue();
447 int64_t High = I.High->getSExtValue();
448
449 IntRange &LastRange = UnreachableRanges.back();
450 if (LastRange.Low == Low) {
451 // There is nothing left of the previous range.
452 UnreachableRanges.pop_back();
453 } else {
454 // Terminate the previous range.
455 assert(Low > LastRange.Low);
456 LastRange.High = Low - 1;
457 }
458 if (High != INT64_MAX) {
459 IntRange R = { High + 1, INT64_MAX };
460 UnreachableRanges.push_back(R);
461 }
462
463 // Count popularity.
464 int64_t N = High - Low + 1;
465 unsigned &Pop = Popularity[I.BB];
466 if ((Pop += N) > MaxPop) {
467 MaxPop = Pop;
468 PopSucc = I.BB;
469 }
470 }
471 #ifndef NDEBUG
472 /* UnreachableRanges should be sorted and the ranges non-adjacent. */
473 for (auto I = UnreachableRanges.begin(), E = UnreachableRanges.end();
474 I != E; ++I) {
475 assert(I->Low <= I->High);
476 auto Next = I + 1;
477 if (Next != E) {
478 assert(Next->Low > I->High);
479 }
480 }
481 #endif
482
483 // Use the most popular block as the new default, reducing the number of
484 // cases.
485 assert(MaxPop > 0 && PopSucc);
486 Default = PopSucc;
487 Cases.erase(std::remove_if(
488 Cases.begin(), Cases.end(),
489 [PopSucc](const CaseRange &R) { return R.BB == PopSucc; }),
490 Cases.end());
491
492 // If there are no cases left, just branch.
493 if (Cases.empty()) {
494 BranchInst::Create(Default, CurBlock);
495 SI->eraseFromParent();
496 return;
497 }
498 }
499
500 // Create a new, empty default block so that the new hierarchy of
501 // if-then statements go to this and the PHI nodes are happy.
502 BasicBlock *NewDefault = BasicBlock::Create(SI->getContext(), "NewDefault");
503 F->getBasicBlockList().insert(Default->getIterator(), NewDefault);
504 BranchInst::Create(Default, NewDefault);
505
506 // If there is an entry in any PHI nodes for the default edge, make sure
507 // to update them as well.
508 for (BasicBlock::iterator I = Default->begin(); isa<PHINode>(I); ++I) {
509 PHINode *PN = cast<PHINode>(I);
510 int BlockIdx = PN->getBasicBlockIndex(OrigBlock);
511 assert(BlockIdx != -1 && "Switch didn't go to this successor??");
512 PN->setIncomingBlock((unsigned)BlockIdx, NewDefault);
513 }
514
515 BasicBlock *SwitchBlock =
516 switchConvert(Cases.begin(), Cases.end(), LowerBound, UpperBound, Val,
517 OrigBlock, OrigBlock, NewDefault, UnreachableRanges);
518
519 // Branch to our shiny new if-then stuff...
520 BranchInst::Create(SwitchBlock, OrigBlock);
521
522 // We are now done with the switch instruction, delete it.
523 BasicBlock *OldDefault = SI->getDefaultDest();
524 CurBlock->getInstList().erase(SI);
525
526 // If the Default block has no more predecessors just add it to DeleteList.
527 if (pred_begin(OldDefault) == pred_end(OldDefault))
528 DeleteList.insert(OldDefault);
529 }
530