1 //===- LowerSwitch.cpp - Eliminate Switch instructions --------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/IR/BasicBlock.h"
21 #include "llvm/IR/CFG.h"
22 #include "llvm/IR/Constants.h"
23 #include "llvm/IR/Function.h"
24 #include "llvm/IR/InstrTypes.h"
25 #include "llvm/IR/Instructions.h"
26 #include "llvm/IR/Value.h"
27 #include "llvm/Pass.h"
28 #include "llvm/Support/Casting.h"
29 #include "llvm/Support/Compiler.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/raw_ostream.h"
32 #include "llvm/Transforms/Utils.h"
33 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
43 #define DEBUG_TYPE "lower-switch"
51 } // end anonymous namespace
53 // Return true iff R is covered by Ranges.
54 static bool IsInRanges(const IntRange &R,
55 const std::vector<IntRange> &Ranges) {
56 // Note: Ranges must be sorted, non-overlapping and non-adjacent.
58 // Find the first range whose High field is >= R.High,
59 // then check if the Low field is <= R.Low. If so, we
60 // have a Range that covers R.
61 auto I = std::lower_bound(
62 Ranges.begin(), Ranges.end(), R,
63 [](const IntRange &A, const IntRange &B) { return A.High < B.High; });
64 return I != Ranges.end() && I->Low <= R.Low;
69 /// Replace all SwitchInst instructions with chained branch instructions.
70 class LowerSwitch : public FunctionPass {
72 // Pass identification, replacement for typeid
75 LowerSwitch() : FunctionPass(ID) {
76 initializeLowerSwitchPass(*PassRegistry::getPassRegistry());
79 bool runOnFunction(Function &F) override;
86 CaseRange(ConstantInt *low, ConstantInt *high, BasicBlock *bb)
87 : Low(low), High(high), BB(bb) {}
90 using CaseVector = std::vector<CaseRange>;
91 using CaseItr = std::vector<CaseRange>::iterator;
94 void processSwitchInst(SwitchInst *SI, SmallPtrSetImpl<BasicBlock*> &DeleteList);
96 BasicBlock *switchConvert(CaseItr Begin, CaseItr End,
97 ConstantInt *LowerBound, ConstantInt *UpperBound,
98 Value *Val, BasicBlock *Predecessor,
99 BasicBlock *OrigBlock, BasicBlock *Default,
100 const std::vector<IntRange> &UnreachableRanges);
101 BasicBlock *newLeafBlock(CaseRange &Leaf, Value *Val, BasicBlock *OrigBlock,
102 BasicBlock *Default);
103 unsigned Clusterify(CaseVector &Cases, SwitchInst *SI);
106 /// The comparison function for sorting the switch case values in the vector.
107 /// WARNING: Case ranges should be disjoint!
109 bool operator()(const LowerSwitch::CaseRange& C1,
110 const LowerSwitch::CaseRange& C2) {
111 const ConstantInt* CI1 = cast<const ConstantInt>(C1.Low);
112 const ConstantInt* CI2 = cast<const ConstantInt>(C2.High);
113 return CI1->getValue().slt(CI2->getValue());
117 } // end anonymous namespace
119 char LowerSwitch::ID = 0;
121 // Publicly exposed interface to pass...
122 char &llvm::LowerSwitchID = LowerSwitch::ID;
124 INITIALIZE_PASS(LowerSwitch, "lowerswitch",
125 "Lower SwitchInst's to branches", false, false)
127 // createLowerSwitchPass - Interface to this file...
128 FunctionPass *llvm::createLowerSwitchPass() {
129 return new LowerSwitch();
132 bool LowerSwitch::runOnFunction(Function &F) {
133 bool Changed = false;
134 SmallPtrSet<BasicBlock*, 8> DeleteList;
136 for (Function::iterator I = F.begin(), E = F.end(); I != E; ) {
137 BasicBlock *Cur = &*I++; // Advance over block so we don't traverse new blocks
139 // If the block is a dead Default block that will be deleted later, don't
140 // waste time processing it.
141 if (DeleteList.count(Cur))
144 if (SwitchInst *SI = dyn_cast<SwitchInst>(Cur->getTerminator())) {
146 processSwitchInst(SI, DeleteList);
150 for (BasicBlock* BB: DeleteList) {
157 /// Used for debugging purposes.
158 static raw_ostream& operator<<(raw_ostream &O,
159 const LowerSwitch::CaseVector &C)
162 static raw_ostream& operator<<(raw_ostream &O,
163 const LowerSwitch::CaseVector &C) {
166 for (LowerSwitch::CaseVector::const_iterator B = C.begin(),
167 E = C.end(); B != E; ) {
168 O << *B->Low << " -" << *B->High;
169 if (++B != E) O << ", ";
175 /// Update the first occurrence of the "switch statement" BB in the PHI
176 /// node with the "new" BB. The other occurrences will:
178 /// 1) Be updated by subsequent calls to this function. Switch statements may
179 /// have more than one outcoming edge into the same BB if they all have the same
180 /// value. When the switch statement is converted these incoming edges are now
181 /// coming from multiple BBs.
182 /// 2) Removed if subsequent incoming values now share the same case, i.e.,
183 /// multiple outcome edges are condensed into one. This is necessary to keep the
184 /// number of phi values equal to the number of branches to SuccBB.
185 static void fixPhis(BasicBlock *SuccBB, BasicBlock *OrigBB, BasicBlock *NewBB,
186 unsigned NumMergedCases) {
187 for (BasicBlock::iterator I = SuccBB->begin(),
188 IE = SuccBB->getFirstNonPHI()->getIterator();
190 PHINode *PN = cast<PHINode>(I);
192 // Only update the first occurrence.
193 unsigned Idx = 0, E = PN->getNumIncomingValues();
194 unsigned LocalNumMergedCases = NumMergedCases;
195 for (; Idx != E; ++Idx) {
196 if (PN->getIncomingBlock(Idx) == OrigBB) {
197 PN->setIncomingBlock(Idx, NewBB);
202 // Remove additional occurrences coming from condensed cases and keep the
203 // number of incoming values equal to the number of branches to SuccBB.
204 SmallVector<unsigned, 8> Indices;
205 for (++Idx; LocalNumMergedCases > 0 && Idx < E; ++Idx)
206 if (PN->getIncomingBlock(Idx) == OrigBB) {
207 Indices.push_back(Idx);
208 LocalNumMergedCases--;
210 // Remove incoming values in the reverse order to prevent invalidating
211 // *successive* index.
212 for (unsigned III : llvm::reverse(Indices))
213 PN->removeIncomingValue(III);
217 /// Convert the switch statement into a binary lookup of the case values.
218 /// The function recursively builds this tree. LowerBound and UpperBound are
219 /// used to keep track of the bounds for Val that have already been checked by
220 /// a block emitted by one of the previous calls to switchConvert in the call
223 LowerSwitch::switchConvert(CaseItr Begin, CaseItr End, ConstantInt *LowerBound,
224 ConstantInt *UpperBound, Value *Val,
225 BasicBlock *Predecessor, BasicBlock *OrigBlock,
227 const std::vector<IntRange> &UnreachableRanges) {
228 unsigned Size = End - Begin;
231 // Check if the Case Range is perfectly squeezed in between
232 // already checked Upper and Lower bounds. If it is then we can avoid
233 // emitting the code that checks if the value actually falls in the range
234 // because the bounds already tell us so.
235 if (Begin->Low == LowerBound && Begin->High == UpperBound) {
236 unsigned NumMergedCases = 0;
237 if (LowerBound && UpperBound)
239 UpperBound->getSExtValue() - LowerBound->getSExtValue();
240 fixPhis(Begin->BB, OrigBlock, Predecessor, NumMergedCases);
243 return newLeafBlock(*Begin, Val, OrigBlock, Default);
246 unsigned Mid = Size / 2;
247 std::vector<CaseRange> LHS(Begin, Begin + Mid);
248 DEBUG(dbgs() << "LHS: " << LHS << "\n");
249 std::vector<CaseRange> RHS(Begin + Mid, End);
250 DEBUG(dbgs() << "RHS: " << RHS << "\n");
252 CaseRange &Pivot = *(Begin + Mid);
253 DEBUG(dbgs() << "Pivot ==> "
254 << Pivot.Low->getValue()
255 << " -" << Pivot.High->getValue() << "\n");
257 // NewLowerBound here should never be the integer minimal value.
258 // This is because it is computed from a case range that is never
259 // the smallest, so there is always a case range that has at least
261 ConstantInt *NewLowerBound = Pivot.Low;
263 // Because NewLowerBound is never the smallest representable integer
264 // it is safe here to subtract one.
265 ConstantInt *NewUpperBound = ConstantInt::get(NewLowerBound->getContext(),
266 NewLowerBound->getValue() - 1);
268 if (!UnreachableRanges.empty()) {
269 // Check if the gap between LHS's highest and NewLowerBound is unreachable.
270 int64_t GapLow = LHS.back().High->getSExtValue() + 1;
271 int64_t GapHigh = NewLowerBound->getSExtValue() - 1;
272 IntRange Gap = { GapLow, GapHigh };
273 if (GapHigh >= GapLow && IsInRanges(Gap, UnreachableRanges))
274 NewUpperBound = LHS.back().High;
277 DEBUG(dbgs() << "LHS Bounds ==> ";
279 dbgs() << LowerBound->getSExtValue();
283 dbgs() << " - " << NewUpperBound->getSExtValue() << "\n";
284 dbgs() << "RHS Bounds ==> ";
285 dbgs() << NewLowerBound->getSExtValue() << " - ";
287 dbgs() << UpperBound->getSExtValue() << "\n";
292 // Create a new node that checks if the value is < pivot. Go to the
293 // left branch if it is and right branch if not.
294 Function* F = OrigBlock->getParent();
295 BasicBlock* NewNode = BasicBlock::Create(Val->getContext(), "NodeBlock");
297 ICmpInst* Comp = new ICmpInst(ICmpInst::ICMP_SLT,
298 Val, Pivot.Low, "Pivot");
300 BasicBlock *LBranch = switchConvert(LHS.begin(), LHS.end(), LowerBound,
301 NewUpperBound, Val, NewNode, OrigBlock,
302 Default, UnreachableRanges);
303 BasicBlock *RBranch = switchConvert(RHS.begin(), RHS.end(), NewLowerBound,
304 UpperBound, Val, NewNode, OrigBlock,
305 Default, UnreachableRanges);
307 F->getBasicBlockList().insert(++OrigBlock->getIterator(), NewNode);
308 NewNode->getInstList().push_back(Comp);
310 BranchInst::Create(LBranch, RBranch, Comp, NewNode);
314 /// Create a new leaf block for the binary lookup tree. It checks if the
315 /// switch's value == the case's value. If not, then it jumps to the default
316 /// branch. At this point in the tree, the value can't be another valid case
317 /// value, so the jump to the "default" branch is warranted.
318 BasicBlock* LowerSwitch::newLeafBlock(CaseRange& Leaf, Value* Val,
319 BasicBlock* OrigBlock,
320 BasicBlock* Default) {
321 Function* F = OrigBlock->getParent();
322 BasicBlock* NewLeaf = BasicBlock::Create(Val->getContext(), "LeafBlock");
323 F->getBasicBlockList().insert(++OrigBlock->getIterator(), NewLeaf);
326 ICmpInst* Comp = nullptr;
327 if (Leaf.Low == Leaf.High) {
328 // Make the seteq instruction...
329 Comp = new ICmpInst(*NewLeaf, ICmpInst::ICMP_EQ, Val,
330 Leaf.Low, "SwitchLeaf");
332 // Make range comparison
333 if (Leaf.Low->isMinValue(true /*isSigned*/)) {
334 // Val >= Min && Val <= Hi --> Val <= Hi
335 Comp = new ICmpInst(*NewLeaf, ICmpInst::ICMP_SLE, Val, Leaf.High,
337 } else if (Leaf.Low->isZero()) {
338 // Val >= 0 && Val <= Hi --> Val <=u Hi
339 Comp = new ICmpInst(*NewLeaf, ICmpInst::ICMP_ULE, Val, Leaf.High,
342 // Emit V-Lo <=u Hi-Lo
343 Constant* NegLo = ConstantExpr::getNeg(Leaf.Low);
344 Instruction* Add = BinaryOperator::CreateAdd(Val, NegLo,
345 Val->getName()+".off",
347 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Leaf.High);
348 Comp = new ICmpInst(*NewLeaf, ICmpInst::ICMP_ULE, Add, UpperBound,
353 // Make the conditional branch...
354 BasicBlock* Succ = Leaf.BB;
355 BranchInst::Create(Succ, Default, Comp, NewLeaf);
357 // If there were any PHI nodes in this successor, rewrite one entry
358 // from OrigBlock to come from NewLeaf.
359 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) {
360 PHINode* PN = cast<PHINode>(I);
361 // Remove all but one incoming entries from the cluster
362 uint64_t Range = Leaf.High->getSExtValue() -
363 Leaf.Low->getSExtValue();
364 for (uint64_t j = 0; j < Range; ++j) {
365 PN->removeIncomingValue(OrigBlock);
368 int BlockIdx = PN->getBasicBlockIndex(OrigBlock);
369 assert(BlockIdx != -1 && "Switch didn't go to this successor??");
370 PN->setIncomingBlock((unsigned)BlockIdx, NewLeaf);
376 /// Transform simple list of Cases into list of CaseRange's.
377 unsigned LowerSwitch::Clusterify(CaseVector& Cases, SwitchInst *SI) {
378 unsigned numCmps = 0;
380 // Start with "simple" cases
381 for (auto Case : SI->cases())
382 Cases.push_back(CaseRange(Case.getCaseValue(), Case.getCaseValue(),
383 Case.getCaseSuccessor()));
385 llvm::sort(Cases.begin(), Cases.end(), CaseCmp());
387 // Merge case into clusters
388 if (Cases.size() >= 2) {
389 CaseItr I = Cases.begin();
390 for (CaseItr J = std::next(I), E = Cases.end(); J != E; ++J) {
391 int64_t nextValue = J->Low->getSExtValue();
392 int64_t currentValue = I->High->getSExtValue();
393 BasicBlock* nextBB = J->BB;
394 BasicBlock* currentBB = I->BB;
396 // If the two neighboring cases go to the same destination, merge them
397 // into a single case.
398 assert(nextValue > currentValue && "Cases should be strictly ascending");
399 if ((nextValue == currentValue + 1) && (currentBB == nextBB)) {
401 // FIXME: Combine branch weights.
402 } else if (++I != J) {
406 Cases.erase(std::next(I), Cases.end());
409 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) {
410 if (I->Low != I->High)
411 // A range counts double, since it requires two compares.
418 /// Replace the specified switch instruction with a sequence of chained if-then
419 /// insts in a balanced binary search.
420 void LowerSwitch::processSwitchInst(SwitchInst *SI,
421 SmallPtrSetImpl<BasicBlock*> &DeleteList) {
422 BasicBlock *CurBlock = SI->getParent();
423 BasicBlock *OrigBlock = CurBlock;
424 Function *F = CurBlock->getParent();
425 Value *Val = SI->getCondition(); // The value we are switching on...
426 BasicBlock* Default = SI->getDefaultDest();
428 // Don't handle unreachable blocks. If there are successors with phis, this
429 // would leave them behind with missing predecessors.
430 if ((CurBlock != &F->getEntryBlock() && pred_empty(CurBlock)) ||
431 CurBlock->getSinglePredecessor() == CurBlock) {
432 DeleteList.insert(CurBlock);
436 // If there is only the default destination, just branch.
437 if (!SI->getNumCases()) {
438 BranchInst::Create(Default, CurBlock);
439 SI->eraseFromParent();
443 // Prepare cases vector.
445 unsigned numCmps = Clusterify(Cases, SI);
446 DEBUG(dbgs() << "Clusterify finished. Total clusters: " << Cases.size()
447 << ". Total compares: " << numCmps << "\n");
448 DEBUG(dbgs() << "Cases: " << Cases << "\n");
451 ConstantInt *LowerBound = nullptr;
452 ConstantInt *UpperBound = nullptr;
453 std::vector<IntRange> UnreachableRanges;
455 if (isa<UnreachableInst>(Default->getFirstNonPHIOrDbg())) {
456 // Make the bounds tightly fitted around the case value range, because we
457 // know that the value passed to the switch must be exactly one of the case
459 assert(!Cases.empty());
460 LowerBound = Cases.front().Low;
461 UpperBound = Cases.back().High;
463 DenseMap<BasicBlock *, unsigned> Popularity;
465 BasicBlock *PopSucc = nullptr;
467 IntRange R = {std::numeric_limits<int64_t>::min(),
468 std::numeric_limits<int64_t>::max()};
469 UnreachableRanges.push_back(R);
470 for (const auto &I : Cases) {
471 int64_t Low = I.Low->getSExtValue();
472 int64_t High = I.High->getSExtValue();
474 IntRange &LastRange = UnreachableRanges.back();
475 if (LastRange.Low == Low) {
476 // There is nothing left of the previous range.
477 UnreachableRanges.pop_back();
479 // Terminate the previous range.
480 assert(Low > LastRange.Low);
481 LastRange.High = Low - 1;
483 if (High != std::numeric_limits<int64_t>::max()) {
484 IntRange R = { High + 1, std::numeric_limits<int64_t>::max() };
485 UnreachableRanges.push_back(R);
489 int64_t N = High - Low + 1;
490 unsigned &Pop = Popularity[I.BB];
491 if ((Pop += N) > MaxPop) {
497 /* UnreachableRanges should be sorted and the ranges non-adjacent. */
498 for (auto I = UnreachableRanges.begin(), E = UnreachableRanges.end();
500 assert(I->Low <= I->High);
503 assert(Next->Low > I->High);
508 // Use the most popular block as the new default, reducing the number of
510 assert(MaxPop > 0 && PopSucc);
514 Cases, [PopSucc](const CaseRange &R) { return R.BB == PopSucc; }),
517 // If there are no cases left, just branch.
519 BranchInst::Create(Default, CurBlock);
520 SI->eraseFromParent();
525 // Create a new, empty default block so that the new hierarchy of
526 // if-then statements go to this and the PHI nodes are happy.
527 BasicBlock *NewDefault = BasicBlock::Create(SI->getContext(), "NewDefault");
528 F->getBasicBlockList().insert(Default->getIterator(), NewDefault);
529 BranchInst::Create(Default, NewDefault);
531 // If there is an entry in any PHI nodes for the default edge, make sure
532 // to update them as well.
533 for (BasicBlock::iterator I = Default->begin(); isa<PHINode>(I); ++I) {
534 PHINode *PN = cast<PHINode>(I);
535 int BlockIdx = PN->getBasicBlockIndex(OrigBlock);
536 assert(BlockIdx != -1 && "Switch didn't go to this successor??");
537 PN->setIncomingBlock((unsigned)BlockIdx, NewDefault);
540 BasicBlock *SwitchBlock =
541 switchConvert(Cases.begin(), Cases.end(), LowerBound, UpperBound, Val,
542 OrigBlock, OrigBlock, NewDefault, UnreachableRanges);
544 // Branch to our shiny new if-then stuff...
545 BranchInst::Create(SwitchBlock, OrigBlock);
547 // We are now done with the switch instruction, delete it.
548 BasicBlock *OldDefault = SI->getDefaultDest();
549 CurBlock->getInstList().erase(SI);
551 // If the Default block has no more predecessors just add it to DeleteList.
552 if (pred_begin(OldDefault) == pred_end(OldDefault))
553 DeleteList.insert(OldDefault);