1 //===- SwitchLoweringUtils.cpp - Switch Lowering --------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file contains switch inst lowering optimizations and utilities for
10 // codegen, so that it can be used for both SelectionDAG and GlobalISel.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/CodeGen/SwitchLoweringUtils.h"
15 #include "llvm/CodeGen/FunctionLoweringInfo.h"
16 #include "llvm/CodeGen/MachineJumpTableInfo.h"
17 #include "llvm/CodeGen/TargetLowering.h"
18 #include "llvm/Target/TargetMachine.h"
19 
20 using namespace llvm;
21 using namespace SwitchCG;
22 
23 uint64_t SwitchCG::getJumpTableRange(const CaseClusterVector &Clusters,
24                                      unsigned First, unsigned Last) {
25   assert(Last >= First);
26   const APInt &LowCase = Clusters[First].Low->getValue();
27   const APInt &HighCase = Clusters[Last].High->getValue();
28   assert(LowCase.getBitWidth() == HighCase.getBitWidth());
29 
30   // FIXME: A range of consecutive cases has 100% density, but only requires one
31   // comparison to lower. We should discriminate against such consecutive ranges
32   // in jump tables.
33   return (HighCase - LowCase).getLimitedValue((UINT64_MAX - 1) / 100) + 1;
34 }
35 
36 uint64_t
37 SwitchCG::getJumpTableNumCases(const SmallVectorImpl<unsigned> &TotalCases,
38                                unsigned First, unsigned Last) {
39   assert(Last >= First);
40   assert(TotalCases[Last] >= TotalCases[First]);
41   uint64_t NumCases =
42       TotalCases[Last] - (First == 0 ? 0 : TotalCases[First - 1]);
43   return NumCases;
44 }
45 
46 void SwitchCG::SwitchLowering::findJumpTables(CaseClusterVector &Clusters,
47                                               const SwitchInst *SI,
48                                               MachineBasicBlock *DefaultMBB,
49                                               ProfileSummaryInfo *PSI,
50                                               BlockFrequencyInfo *BFI) {
51 #ifndef NDEBUG
52   // Clusters must be non-empty, sorted, and only contain Range clusters.
53   assert(!Clusters.empty());
54   for (CaseCluster &C : Clusters)
55     assert(C.Kind == CC_Range);
56   for (unsigned i = 1, e = Clusters.size(); i < e; ++i)
57     assert(Clusters[i - 1].High->getValue().slt(Clusters[i].Low->getValue()));
58 #endif
59 
60   assert(TLI && "TLI not set!");
61   if (!TLI->areJTsAllowed(SI->getParent()->getParent()))
62     return;
63 
64   const unsigned MinJumpTableEntries = TLI->getMinimumJumpTableEntries();
65   const unsigned SmallNumberOfEntries = MinJumpTableEntries / 2;
66 
67   // Bail if not enough cases.
68   const int64_t N = Clusters.size();
69   if (N < 2 || N < MinJumpTableEntries)
70     return;
71 
72   // Accumulated number of cases in each cluster and those prior to it.
73   SmallVector<unsigned, 8> TotalCases(N);
74   for (unsigned i = 0; i < N; ++i) {
75     const APInt &Hi = Clusters[i].High->getValue();
76     const APInt &Lo = Clusters[i].Low->getValue();
77     TotalCases[i] = (Hi - Lo).getLimitedValue() + 1;
78     if (i != 0)
79       TotalCases[i] += TotalCases[i - 1];
80   }
81 
82   uint64_t Range = getJumpTableRange(Clusters,0, N - 1);
83   uint64_t NumCases = getJumpTableNumCases(TotalCases, 0, N - 1);
84   assert(NumCases < UINT64_MAX / 100);
85   assert(Range >= NumCases);
86 
87   // Cheap case: the whole range may be suitable for jump table.
88   if (TLI->isSuitableForJumpTable(SI, NumCases, Range, PSI, BFI)) {
89     CaseCluster JTCluster;
90     if (buildJumpTable(Clusters, 0, N - 1, SI, DefaultMBB, JTCluster)) {
91       Clusters[0] = JTCluster;
92       Clusters.resize(1);
93       return;
94     }
95   }
96 
97   // The algorithm below is not suitable for -O0.
98   if (TM->getOptLevel() == CodeGenOpt::None)
99     return;
100 
101   // Split Clusters into minimum number of dense partitions. The algorithm uses
102   // the same idea as Kannan & Proebsting "Correction to 'Producing Good Code
103   // for the Case Statement'" (1994), but builds the MinPartitions array in
104   // reverse order to make it easier to reconstruct the partitions in ascending
105   // order. In the choice between two optimal partitionings, it picks the one
106   // which yields more jump tables.
107 
108   // MinPartitions[i] is the minimum nbr of partitions of Clusters[i..N-1].
109   SmallVector<unsigned, 8> MinPartitions(N);
110   // LastElement[i] is the last element of the partition starting at i.
111   SmallVector<unsigned, 8> LastElement(N);
112   // PartitionsScore[i] is used to break ties when choosing between two
113   // partitionings resulting in the same number of partitions.
114   SmallVector<unsigned, 8> PartitionsScore(N);
115   // For PartitionsScore, a small number of comparisons is considered as good as
116   // a jump table and a single comparison is considered better than a jump
117   // table.
118   enum PartitionScores : unsigned {
119     NoTable = 0,
120     Table = 1,
121     FewCases = 1,
122     SingleCase = 2
123   };
124 
125   // Base case: There is only one way to partition Clusters[N-1].
126   MinPartitions[N - 1] = 1;
127   LastElement[N - 1] = N - 1;
128   PartitionsScore[N - 1] = PartitionScores::SingleCase;
129 
130   // Note: loop indexes are signed to avoid underflow.
131   for (int64_t i = N - 2; i >= 0; i--) {
132     // Find optimal partitioning of Clusters[i..N-1].
133     // Baseline: Put Clusters[i] into a partition on its own.
134     MinPartitions[i] = MinPartitions[i + 1] + 1;
135     LastElement[i] = i;
136     PartitionsScore[i] = PartitionsScore[i + 1] + PartitionScores::SingleCase;
137 
138     // Search for a solution that results in fewer partitions.
139     for (int64_t j = N - 1; j > i; j--) {
140       // Try building a partition from Clusters[i..j].
141       Range = getJumpTableRange(Clusters, i, j);
142       NumCases = getJumpTableNumCases(TotalCases, i, j);
143       assert(NumCases < UINT64_MAX / 100);
144       assert(Range >= NumCases);
145 
146       if (TLI->isSuitableForJumpTable(SI, NumCases, Range, PSI, BFI)) {
147         unsigned NumPartitions = 1 + (j == N - 1 ? 0 : MinPartitions[j + 1]);
148         unsigned Score = j == N - 1 ? 0 : PartitionsScore[j + 1];
149         int64_t NumEntries = j - i + 1;
150 
151         if (NumEntries == 1)
152           Score += PartitionScores::SingleCase;
153         else if (NumEntries <= SmallNumberOfEntries)
154           Score += PartitionScores::FewCases;
155         else if (NumEntries >= MinJumpTableEntries)
156           Score += PartitionScores::Table;
157 
158         // If this leads to fewer partitions, or to the same number of
159         // partitions with better score, it is a better partitioning.
160         if (NumPartitions < MinPartitions[i] ||
161             (NumPartitions == MinPartitions[i] && Score > PartitionsScore[i])) {
162           MinPartitions[i] = NumPartitions;
163           LastElement[i] = j;
164           PartitionsScore[i] = Score;
165         }
166       }
167     }
168   }
169 
170   // Iterate over the partitions, replacing some with jump tables in-place.
171   unsigned DstIndex = 0;
172   for (unsigned First = 0, Last; First < N; First = Last + 1) {
173     Last = LastElement[First];
174     assert(Last >= First);
175     assert(DstIndex <= First);
176     unsigned NumClusters = Last - First + 1;
177 
178     CaseCluster JTCluster;
179     if (NumClusters >= MinJumpTableEntries &&
180         buildJumpTable(Clusters, First, Last, SI, DefaultMBB, JTCluster)) {
181       Clusters[DstIndex++] = JTCluster;
182     } else {
183       for (unsigned I = First; I <= Last; ++I)
184         std::memmove(&Clusters[DstIndex++], &Clusters[I], sizeof(Clusters[I]));
185     }
186   }
187   Clusters.resize(DstIndex);
188 }
189 
190 bool SwitchCG::SwitchLowering::buildJumpTable(const CaseClusterVector &Clusters,
191                                               unsigned First, unsigned Last,
192                                               const SwitchInst *SI,
193                                               MachineBasicBlock *DefaultMBB,
194                                               CaseCluster &JTCluster) {
195   assert(First <= Last);
196 
197   auto Prob = BranchProbability::getZero();
198   unsigned NumCmps = 0;
199   std::vector<MachineBasicBlock*> Table;
200   DenseMap<MachineBasicBlock*, BranchProbability> JTProbs;
201 
202   // Initialize probabilities in JTProbs.
203   for (unsigned I = First; I <= Last; ++I)
204     JTProbs[Clusters[I].MBB] = BranchProbability::getZero();
205 
206   for (unsigned I = First; I <= Last; ++I) {
207     assert(Clusters[I].Kind == CC_Range);
208     Prob += Clusters[I].Prob;
209     const APInt &Low = Clusters[I].Low->getValue();
210     const APInt &High = Clusters[I].High->getValue();
211     NumCmps += (Low == High) ? 1 : 2;
212     if (I != First) {
213       // Fill the gap between this and the previous cluster.
214       const APInt &PreviousHigh = Clusters[I - 1].High->getValue();
215       assert(PreviousHigh.slt(Low));
216       uint64_t Gap = (Low - PreviousHigh).getLimitedValue() - 1;
217       for (uint64_t J = 0; J < Gap; J++)
218         Table.push_back(DefaultMBB);
219     }
220     uint64_t ClusterSize = (High - Low).getLimitedValue() + 1;
221     for (uint64_t J = 0; J < ClusterSize; ++J)
222       Table.push_back(Clusters[I].MBB);
223     JTProbs[Clusters[I].MBB] += Clusters[I].Prob;
224   }
225 
226   unsigned NumDests = JTProbs.size();
227   if (TLI->isSuitableForBitTests(NumDests, NumCmps,
228                                  Clusters[First].Low->getValue(),
229                                  Clusters[Last].High->getValue(), *DL)) {
230     // Clusters[First..Last] should be lowered as bit tests instead.
231     return false;
232   }
233 
234   // Create the MBB that will load from and jump through the table.
235   // Note: We create it here, but it's not inserted into the function yet.
236   MachineFunction *CurMF = FuncInfo.MF;
237   MachineBasicBlock *JumpTableMBB =
238       CurMF->CreateMachineBasicBlock(SI->getParent());
239 
240   // Add successors. Note: use table order for determinism.
241   SmallPtrSet<MachineBasicBlock *, 8> Done;
242   for (MachineBasicBlock *Succ : Table) {
243     if (Done.count(Succ))
244       continue;
245     addSuccessorWithProb(JumpTableMBB, Succ, JTProbs[Succ]);
246     Done.insert(Succ);
247   }
248   JumpTableMBB->normalizeSuccProbs();
249 
250   unsigned JTI = CurMF->getOrCreateJumpTableInfo(TLI->getJumpTableEncoding())
251                      ->createJumpTableIndex(Table);
252 
253   // Set up the jump table info.
254   JumpTable JT(-1U, JTI, JumpTableMBB, nullptr);
255   JumpTableHeader JTH(Clusters[First].Low->getValue(),
256                       Clusters[Last].High->getValue(), SI->getCondition(),
257                       nullptr, false);
258   JTCases.emplace_back(std::move(JTH), std::move(JT));
259 
260   JTCluster = CaseCluster::jumpTable(Clusters[First].Low, Clusters[Last].High,
261                                      JTCases.size() - 1, Prob);
262   return true;
263 }
264 
265 void SwitchCG::SwitchLowering::findBitTestClusters(CaseClusterVector &Clusters,
266                                                    const SwitchInst *SI) {
267   // Partition Clusters into as few subsets as possible, where each subset has a
268   // range that fits in a machine word and has <= 3 unique destinations.
269 
270 #ifndef NDEBUG
271   // Clusters must be sorted and contain Range or JumpTable clusters.
272   assert(!Clusters.empty());
273   assert(Clusters[0].Kind == CC_Range || Clusters[0].Kind == CC_JumpTable);
274   for (const CaseCluster &C : Clusters)
275     assert(C.Kind == CC_Range || C.Kind == CC_JumpTable);
276   for (unsigned i = 1; i < Clusters.size(); ++i)
277     assert(Clusters[i-1].High->getValue().slt(Clusters[i].Low->getValue()));
278 #endif
279 
280   // The algorithm below is not suitable for -O0.
281   if (TM->getOptLevel() == CodeGenOpt::None)
282     return;
283 
284   // If target does not have legal shift left, do not emit bit tests at all.
285   EVT PTy = TLI->getPointerTy(*DL);
286   if (!TLI->isOperationLegal(ISD::SHL, PTy))
287     return;
288 
289   int BitWidth = PTy.getSizeInBits();
290   const int64_t N = Clusters.size();
291 
292   // MinPartitions[i] is the minimum nbr of partitions of Clusters[i..N-1].
293   SmallVector<unsigned, 8> MinPartitions(N);
294   // LastElement[i] is the last element of the partition starting at i.
295   SmallVector<unsigned, 8> LastElement(N);
296 
297   // FIXME: This might not be the best algorithm for finding bit test clusters.
298 
299   // Base case: There is only one way to partition Clusters[N-1].
300   MinPartitions[N - 1] = 1;
301   LastElement[N - 1] = N - 1;
302 
303   // Note: loop indexes are signed to avoid underflow.
304   for (int64_t i = N - 2; i >= 0; --i) {
305     // Find optimal partitioning of Clusters[i..N-1].
306     // Baseline: Put Clusters[i] into a partition on its own.
307     MinPartitions[i] = MinPartitions[i + 1] + 1;
308     LastElement[i] = i;
309 
310     // Search for a solution that results in fewer partitions.
311     // Note: the search is limited by BitWidth, reducing time complexity.
312     for (int64_t j = std::min(N - 1, i + BitWidth - 1); j > i; --j) {
313       // Try building a partition from Clusters[i..j].
314 
315       // Check the range.
316       if (!TLI->rangeFitsInWord(Clusters[i].Low->getValue(),
317                                 Clusters[j].High->getValue(), *DL))
318         continue;
319 
320       // Check nbr of destinations and cluster types.
321       // FIXME: This works, but doesn't seem very efficient.
322       bool RangesOnly = true;
323       BitVector Dests(FuncInfo.MF->getNumBlockIDs());
324       for (int64_t k = i; k <= j; k++) {
325         if (Clusters[k].Kind != CC_Range) {
326           RangesOnly = false;
327           break;
328         }
329         Dests.set(Clusters[k].MBB->getNumber());
330       }
331       if (!RangesOnly || Dests.count() > 3)
332         break;
333 
334       // Check if it's a better partition.
335       unsigned NumPartitions = 1 + (j == N - 1 ? 0 : MinPartitions[j + 1]);
336       if (NumPartitions < MinPartitions[i]) {
337         // Found a better partition.
338         MinPartitions[i] = NumPartitions;
339         LastElement[i] = j;
340       }
341     }
342   }
343 
344   // Iterate over the partitions, replacing with bit-test clusters in-place.
345   unsigned DstIndex = 0;
346   for (unsigned First = 0, Last; First < N; First = Last + 1) {
347     Last = LastElement[First];
348     assert(First <= Last);
349     assert(DstIndex <= First);
350 
351     CaseCluster BitTestCluster;
352     if (buildBitTests(Clusters, First, Last, SI, BitTestCluster)) {
353       Clusters[DstIndex++] = BitTestCluster;
354     } else {
355       size_t NumClusters = Last - First + 1;
356       std::memmove(&Clusters[DstIndex], &Clusters[First],
357                    sizeof(Clusters[0]) * NumClusters);
358       DstIndex += NumClusters;
359     }
360   }
361   Clusters.resize(DstIndex);
362 }
363 
364 bool SwitchCG::SwitchLowering::buildBitTests(CaseClusterVector &Clusters,
365                                              unsigned First, unsigned Last,
366                                              const SwitchInst *SI,
367                                              CaseCluster &BTCluster) {
368   assert(First <= Last);
369   if (First == Last)
370     return false;
371 
372   BitVector Dests(FuncInfo.MF->getNumBlockIDs());
373   unsigned NumCmps = 0;
374   for (int64_t I = First; I <= Last; ++I) {
375     assert(Clusters[I].Kind == CC_Range);
376     Dests.set(Clusters[I].MBB->getNumber());
377     NumCmps += (Clusters[I].Low == Clusters[I].High) ? 1 : 2;
378   }
379   unsigned NumDests = Dests.count();
380 
381   APInt Low = Clusters[First].Low->getValue();
382   APInt High = Clusters[Last].High->getValue();
383   assert(Low.slt(High));
384 
385   if (!TLI->isSuitableForBitTests(NumDests, NumCmps, Low, High, *DL))
386     return false;
387 
388   APInt LowBound;
389   APInt CmpRange;
390 
391   const int BitWidth = TLI->getPointerTy(*DL).getSizeInBits();
392   assert(TLI->rangeFitsInWord(Low, High, *DL) &&
393          "Case range must fit in bit mask!");
394 
395   // Check if the clusters cover a contiguous range such that no value in the
396   // range will jump to the default statement.
397   bool ContiguousRange = true;
398   for (int64_t I = First + 1; I <= Last; ++I) {
399     if (Clusters[I].Low->getValue() != Clusters[I - 1].High->getValue() + 1) {
400       ContiguousRange = false;
401       break;
402     }
403   }
404 
405   if (Low.isStrictlyPositive() && High.slt(BitWidth)) {
406     // Optimize the case where all the case values fit in a word without having
407     // to subtract minValue. In this case, we can optimize away the subtraction.
408     LowBound = APInt::getZero(Low.getBitWidth());
409     CmpRange = High;
410     ContiguousRange = false;
411   } else {
412     LowBound = Low;
413     CmpRange = High - Low;
414   }
415 
416   CaseBitsVector CBV;
417   auto TotalProb = BranchProbability::getZero();
418   for (unsigned i = First; i <= Last; ++i) {
419     // Find the CaseBits for this destination.
420     unsigned j;
421     for (j = 0; j < CBV.size(); ++j)
422       if (CBV[j].BB == Clusters[i].MBB)
423         break;
424     if (j == CBV.size())
425       CBV.push_back(
426           CaseBits(0, Clusters[i].MBB, 0, BranchProbability::getZero()));
427     CaseBits *CB = &CBV[j];
428 
429     // Update Mask, Bits and ExtraProb.
430     uint64_t Lo = (Clusters[i].Low->getValue() - LowBound).getZExtValue();
431     uint64_t Hi = (Clusters[i].High->getValue() - LowBound).getZExtValue();
432     assert(Hi >= Lo && Hi < 64 && "Invalid bit case!");
433     CB->Mask |= (-1ULL >> (63 - (Hi - Lo))) << Lo;
434     CB->Bits += Hi - Lo + 1;
435     CB->ExtraProb += Clusters[i].Prob;
436     TotalProb += Clusters[i].Prob;
437   }
438 
439   BitTestInfo BTI;
440   llvm::sort(CBV, [](const CaseBits &a, const CaseBits &b) {
441     // Sort by probability first, number of bits second, bit mask third.
442     if (a.ExtraProb != b.ExtraProb)
443       return a.ExtraProb > b.ExtraProb;
444     if (a.Bits != b.Bits)
445       return a.Bits > b.Bits;
446     return a.Mask < b.Mask;
447   });
448 
449   for (auto &CB : CBV) {
450     MachineBasicBlock *BitTestBB =
451         FuncInfo.MF->CreateMachineBasicBlock(SI->getParent());
452     BTI.push_back(BitTestCase(CB.Mask, BitTestBB, CB.BB, CB.ExtraProb));
453   }
454   BitTestCases.emplace_back(std::move(LowBound), std::move(CmpRange),
455                             SI->getCondition(), -1U, MVT::Other, false,
456                             ContiguousRange, nullptr, nullptr, std::move(BTI),
457                             TotalProb);
458 
459   BTCluster = CaseCluster::bitTests(Clusters[First].Low, Clusters[Last].High,
460                                     BitTestCases.size() - 1, TotalProb);
461   return true;
462 }
463 
464 void SwitchCG::sortAndRangeify(CaseClusterVector &Clusters) {
465 #ifndef NDEBUG
466   for (const CaseCluster &CC : Clusters)
467     assert(CC.Low == CC.High && "Input clusters must be single-case");
468 #endif
469 
470   llvm::sort(Clusters, [](const CaseCluster &a, const CaseCluster &b) {
471     return a.Low->getValue().slt(b.Low->getValue());
472   });
473 
474   // Merge adjacent clusters with the same destination.
475   const unsigned N = Clusters.size();
476   unsigned DstIndex = 0;
477   for (unsigned SrcIndex = 0; SrcIndex < N; ++SrcIndex) {
478     CaseCluster &CC = Clusters[SrcIndex];
479     const ConstantInt *CaseVal = CC.Low;
480     MachineBasicBlock *Succ = CC.MBB;
481 
482     if (DstIndex != 0 && Clusters[DstIndex - 1].MBB == Succ &&
483         (CaseVal->getValue() - Clusters[DstIndex - 1].High->getValue()) == 1) {
484       // If this case has the same successor and is a neighbour, merge it into
485       // the previous cluster.
486       Clusters[DstIndex - 1].High = CaseVal;
487       Clusters[DstIndex - 1].Prob += CC.Prob;
488     } else {
489       std::memmove(&Clusters[DstIndex++], &Clusters[SrcIndex],
490                    sizeof(Clusters[SrcIndex]));
491     }
492   }
493   Clusters.resize(DstIndex);
494 }
495