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                                               std::optional<SDLoc> SL,
49                                               MachineBasicBlock *DefaultMBB,
50                                               ProfileSummaryInfo *PSI,
51                                               BlockFrequencyInfo *BFI) {
52 #ifndef NDEBUG
53   // Clusters must be non-empty, sorted, and only contain Range clusters.
54   assert(!Clusters.empty());
55   for (CaseCluster &C : Clusters)
56     assert(C.Kind == CC_Range);
57   for (unsigned i = 1, e = Clusters.size(); i < e; ++i)
58     assert(Clusters[i - 1].High->getValue().slt(Clusters[i].Low->getValue()));
59 #endif
60 
61   assert(TLI && "TLI not set!");
62   if (!TLI->areJTsAllowed(SI->getParent()->getParent()))
63     return;
64 
65   const unsigned MinJumpTableEntries = TLI->getMinimumJumpTableEntries();
66   const unsigned SmallNumberOfEntries = MinJumpTableEntries / 2;
67 
68   // Bail if not enough cases.
69   const int64_t N = Clusters.size();
70   if (N < 2 || N < MinJumpTableEntries)
71     return;
72 
73   // Accumulated number of cases in each cluster and those prior to it.
74   SmallVector<unsigned, 8> TotalCases(N);
75   for (unsigned i = 0; i < N; ++i) {
76     const APInt &Hi = Clusters[i].High->getValue();
77     const APInt &Lo = Clusters[i].Low->getValue();
78     TotalCases[i] = (Hi - Lo).getLimitedValue() + 1;
79     if (i != 0)
80       TotalCases[i] += TotalCases[i - 1];
81   }
82 
83   uint64_t Range = getJumpTableRange(Clusters,0, N - 1);
84   uint64_t NumCases = getJumpTableNumCases(TotalCases, 0, N - 1);
85   assert(NumCases < UINT64_MAX / 100);
86   assert(Range >= NumCases);
87 
88   // Cheap case: the whole range may be suitable for jump table.
89   if (TLI->isSuitableForJumpTable(SI, NumCases, Range, PSI, BFI)) {
90     CaseCluster JTCluster;
91     if (buildJumpTable(Clusters, 0, N - 1, SI, SL, DefaultMBB, JTCluster)) {
92       Clusters[0] = JTCluster;
93       Clusters.resize(1);
94       return;
95     }
96   }
97 
98   // The algorithm below is not suitable for -O0.
99   if (TM->getOptLevel() == CodeGenOptLevel::None)
100     return;
101 
102   // Split Clusters into minimum number of dense partitions. The algorithm uses
103   // the same idea as Kannan & Proebsting "Correction to 'Producing Good Code
104   // for the Case Statement'" (1994), but builds the MinPartitions array in
105   // reverse order to make it easier to reconstruct the partitions in ascending
106   // order. In the choice between two optimal partitionings, it picks the one
107   // which yields more jump tables.
108 
109   // MinPartitions[i] is the minimum nbr of partitions of Clusters[i..N-1].
110   SmallVector<unsigned, 8> MinPartitions(N);
111   // LastElement[i] is the last element of the partition starting at i.
112   SmallVector<unsigned, 8> LastElement(N);
113   // PartitionsScore[i] is used to break ties when choosing between two
114   // partitionings resulting in the same number of partitions.
115   SmallVector<unsigned, 8> PartitionsScore(N);
116   // For PartitionsScore, a small number of comparisons is considered as good as
117   // a jump table and a single comparison is considered better than a jump
118   // table.
119   enum PartitionScores : unsigned {
120     NoTable = 0,
121     Table = 1,
122     FewCases = 1,
123     SingleCase = 2
124   };
125 
126   // Base case: There is only one way to partition Clusters[N-1].
127   MinPartitions[N - 1] = 1;
128   LastElement[N - 1] = N - 1;
129   PartitionsScore[N - 1] = PartitionScores::SingleCase;
130 
131   // Note: loop indexes are signed to avoid underflow.
132   for (int64_t i = N - 2; i >= 0; i--) {
133     // Find optimal partitioning of Clusters[i..N-1].
134     // Baseline: Put Clusters[i] into a partition on its own.
135     MinPartitions[i] = MinPartitions[i + 1] + 1;
136     LastElement[i] = i;
137     PartitionsScore[i] = PartitionsScore[i + 1] + PartitionScores::SingleCase;
138 
139     // Search for a solution that results in fewer partitions.
140     for (int64_t j = N - 1; j > i; j--) {
141       // Try building a partition from Clusters[i..j].
142       Range = getJumpTableRange(Clusters, i, j);
143       NumCases = getJumpTableNumCases(TotalCases, i, j);
144       assert(NumCases < UINT64_MAX / 100);
145       assert(Range >= NumCases);
146 
147       if (TLI->isSuitableForJumpTable(SI, NumCases, Range, PSI, BFI)) {
148         unsigned NumPartitions = 1 + (j == N - 1 ? 0 : MinPartitions[j + 1]);
149         unsigned Score = j == N - 1 ? 0 : PartitionsScore[j + 1];
150         int64_t NumEntries = j - i + 1;
151 
152         if (NumEntries == 1)
153           Score += PartitionScores::SingleCase;
154         else if (NumEntries <= SmallNumberOfEntries)
155           Score += PartitionScores::FewCases;
156         else if (NumEntries >= MinJumpTableEntries)
157           Score += PartitionScores::Table;
158 
159         // If this leads to fewer partitions, or to the same number of
160         // partitions with better score, it is a better partitioning.
161         if (NumPartitions < MinPartitions[i] ||
162             (NumPartitions == MinPartitions[i] && Score > PartitionsScore[i])) {
163           MinPartitions[i] = NumPartitions;
164           LastElement[i] = j;
165           PartitionsScore[i] = Score;
166         }
167       }
168     }
169   }
170 
171   // Iterate over the partitions, replacing some with jump tables in-place.
172   unsigned DstIndex = 0;
173   for (unsigned First = 0, Last; First < N; First = Last + 1) {
174     Last = LastElement[First];
175     assert(Last >= First);
176     assert(DstIndex <= First);
177     unsigned NumClusters = Last - First + 1;
178 
179     CaseCluster JTCluster;
180     if (NumClusters >= MinJumpTableEntries &&
181         buildJumpTable(Clusters, First, Last, SI, SL, DefaultMBB, JTCluster)) {
182       Clusters[DstIndex++] = JTCluster;
183     } else {
184       for (unsigned I = First; I <= Last; ++I)
185         std::memmove(&Clusters[DstIndex++], &Clusters[I], sizeof(Clusters[I]));
186     }
187   }
188   Clusters.resize(DstIndex);
189 }
190 
191 bool SwitchCG::SwitchLowering::buildJumpTable(const CaseClusterVector &Clusters,
192                                               unsigned First, unsigned Last,
193                                               const SwitchInst *SI,
194                                               const std::optional<SDLoc> &SL,
195                                               MachineBasicBlock *DefaultMBB,
196                                               CaseCluster &JTCluster) {
197   assert(First <= Last);
198 
199   auto Prob = BranchProbability::getZero();
200   unsigned NumCmps = 0;
201   std::vector<MachineBasicBlock*> Table;
202   DenseMap<MachineBasicBlock*, BranchProbability> JTProbs;
203 
204   // Initialize probabilities in JTProbs.
205   for (unsigned I = First; I <= Last; ++I)
206     JTProbs[Clusters[I].MBB] = BranchProbability::getZero();
207 
208   for (unsigned I = First; I <= Last; ++I) {
209     assert(Clusters[I].Kind == CC_Range);
210     Prob += Clusters[I].Prob;
211     const APInt &Low = Clusters[I].Low->getValue();
212     const APInt &High = Clusters[I].High->getValue();
213     NumCmps += (Low == High) ? 1 : 2;
214     if (I != First) {
215       // Fill the gap between this and the previous cluster.
216       const APInt &PreviousHigh = Clusters[I - 1].High->getValue();
217       assert(PreviousHigh.slt(Low));
218       uint64_t Gap = (Low - PreviousHigh).getLimitedValue() - 1;
219       for (uint64_t J = 0; J < Gap; J++)
220         Table.push_back(DefaultMBB);
221     }
222     uint64_t ClusterSize = (High - Low).getLimitedValue() + 1;
223     for (uint64_t J = 0; J < ClusterSize; ++J)
224       Table.push_back(Clusters[I].MBB);
225     JTProbs[Clusters[I].MBB] += Clusters[I].Prob;
226   }
227 
228   unsigned NumDests = JTProbs.size();
229   if (TLI->isSuitableForBitTests(NumDests, NumCmps,
230                                  Clusters[First].Low->getValue(),
231                                  Clusters[Last].High->getValue(), *DL)) {
232     // Clusters[First..Last] should be lowered as bit tests instead.
233     return false;
234   }
235 
236   // Create the MBB that will load from and jump through the table.
237   // Note: We create it here, but it's not inserted into the function yet.
238   MachineFunction *CurMF = FuncInfo.MF;
239   MachineBasicBlock *JumpTableMBB =
240       CurMF->CreateMachineBasicBlock(SI->getParent());
241 
242   // Add successors. Note: use table order for determinism.
243   SmallPtrSet<MachineBasicBlock *, 8> Done;
244   for (MachineBasicBlock *Succ : Table) {
245     if (Done.count(Succ))
246       continue;
247     addSuccessorWithProb(JumpTableMBB, Succ, JTProbs[Succ]);
248     Done.insert(Succ);
249   }
250   JumpTableMBB->normalizeSuccProbs();
251 
252   unsigned JTI = CurMF->getOrCreateJumpTableInfo(TLI->getJumpTableEncoding())
253                      ->createJumpTableIndex(Table);
254 
255   // Set up the jump table info.
256   JumpTable JT(-1U, JTI, JumpTableMBB, nullptr, SL);
257   JumpTableHeader JTH(Clusters[First].Low->getValue(),
258                       Clusters[Last].High->getValue(), SI->getCondition(),
259                       nullptr, false);
260   JTCases.emplace_back(std::move(JTH), std::move(JT));
261 
262   JTCluster = CaseCluster::jumpTable(Clusters[First].Low, Clusters[Last].High,
263                                      JTCases.size() - 1, Prob);
264   return true;
265 }
266 
267 void SwitchCG::SwitchLowering::findBitTestClusters(CaseClusterVector &Clusters,
268                                                    const SwitchInst *SI) {
269   // Partition Clusters into as few subsets as possible, where each subset has a
270   // range that fits in a machine word and has <= 3 unique destinations.
271 
272 #ifndef NDEBUG
273   // Clusters must be sorted and contain Range or JumpTable clusters.
274   assert(!Clusters.empty());
275   assert(Clusters[0].Kind == CC_Range || Clusters[0].Kind == CC_JumpTable);
276   for (const CaseCluster &C : Clusters)
277     assert(C.Kind == CC_Range || C.Kind == CC_JumpTable);
278   for (unsigned i = 1; i < Clusters.size(); ++i)
279     assert(Clusters[i-1].High->getValue().slt(Clusters[i].Low->getValue()));
280 #endif
281 
282   // The algorithm below is not suitable for -O0.
283   if (TM->getOptLevel() == CodeGenOptLevel::None)
284     return;
285 
286   // If target does not have legal shift left, do not emit bit tests at all.
287   EVT PTy = TLI->getPointerTy(*DL);
288   if (!TLI->isOperationLegal(ISD::SHL, PTy))
289     return;
290 
291   int BitWidth = PTy.getSizeInBits();
292   const int64_t N = Clusters.size();
293 
294   // MinPartitions[i] is the minimum nbr of partitions of Clusters[i..N-1].
295   SmallVector<unsigned, 8> MinPartitions(N);
296   // LastElement[i] is the last element of the partition starting at i.
297   SmallVector<unsigned, 8> LastElement(N);
298 
299   // FIXME: This might not be the best algorithm for finding bit test clusters.
300 
301   // Base case: There is only one way to partition Clusters[N-1].
302   MinPartitions[N - 1] = 1;
303   LastElement[N - 1] = N - 1;
304 
305   // Note: loop indexes are signed to avoid underflow.
306   for (int64_t i = N - 2; i >= 0; --i) {
307     // Find optimal partitioning of Clusters[i..N-1].
308     // Baseline: Put Clusters[i] into a partition on its own.
309     MinPartitions[i] = MinPartitions[i + 1] + 1;
310     LastElement[i] = i;
311 
312     // Search for a solution that results in fewer partitions.
313     // Note: the search is limited by BitWidth, reducing time complexity.
314     for (int64_t j = std::min(N - 1, i + BitWidth - 1); j > i; --j) {
315       // Try building a partition from Clusters[i..j].
316 
317       // Check the range.
318       if (!TLI->rangeFitsInWord(Clusters[i].Low->getValue(),
319                                 Clusters[j].High->getValue(), *DL))
320         continue;
321 
322       // Check nbr of destinations and cluster types.
323       // FIXME: This works, but doesn't seem very efficient.
324       bool RangesOnly = true;
325       BitVector Dests(FuncInfo.MF->getNumBlockIDs());
326       for (int64_t k = i; k <= j; k++) {
327         if (Clusters[k].Kind != CC_Range) {
328           RangesOnly = false;
329           break;
330         }
331         Dests.set(Clusters[k].MBB->getNumber());
332       }
333       if (!RangesOnly || Dests.count() > 3)
334         break;
335 
336       // Check if it's a better partition.
337       unsigned NumPartitions = 1 + (j == N - 1 ? 0 : MinPartitions[j + 1]);
338       if (NumPartitions < MinPartitions[i]) {
339         // Found a better partition.
340         MinPartitions[i] = NumPartitions;
341         LastElement[i] = j;
342       }
343     }
344   }
345 
346   // Iterate over the partitions, replacing with bit-test clusters in-place.
347   unsigned DstIndex = 0;
348   for (unsigned First = 0, Last; First < N; First = Last + 1) {
349     Last = LastElement[First];
350     assert(First <= Last);
351     assert(DstIndex <= First);
352 
353     CaseCluster BitTestCluster;
354     if (buildBitTests(Clusters, First, Last, SI, BitTestCluster)) {
355       Clusters[DstIndex++] = BitTestCluster;
356     } else {
357       size_t NumClusters = Last - First + 1;
358       std::memmove(&Clusters[DstIndex], &Clusters[First],
359                    sizeof(Clusters[0]) * NumClusters);
360       DstIndex += NumClusters;
361     }
362   }
363   Clusters.resize(DstIndex);
364 }
365 
366 bool SwitchCG::SwitchLowering::buildBitTests(CaseClusterVector &Clusters,
367                                              unsigned First, unsigned Last,
368                                              const SwitchInst *SI,
369                                              CaseCluster &BTCluster) {
370   assert(First <= Last);
371   if (First == Last)
372     return false;
373 
374   BitVector Dests(FuncInfo.MF->getNumBlockIDs());
375   unsigned NumCmps = 0;
376   for (int64_t I = First; I <= Last; ++I) {
377     assert(Clusters[I].Kind == CC_Range);
378     Dests.set(Clusters[I].MBB->getNumber());
379     NumCmps += (Clusters[I].Low == Clusters[I].High) ? 1 : 2;
380   }
381   unsigned NumDests = Dests.count();
382 
383   APInt Low = Clusters[First].Low->getValue();
384   APInt High = Clusters[Last].High->getValue();
385   assert(Low.slt(High));
386 
387   if (!TLI->isSuitableForBitTests(NumDests, NumCmps, Low, High, *DL))
388     return false;
389 
390   APInt LowBound;
391   APInt CmpRange;
392 
393   const int BitWidth = TLI->getPointerTy(*DL).getSizeInBits();
394   assert(TLI->rangeFitsInWord(Low, High, *DL) &&
395          "Case range must fit in bit mask!");
396 
397   // Check if the clusters cover a contiguous range such that no value in the
398   // range will jump to the default statement.
399   bool ContiguousRange = true;
400   for (int64_t I = First + 1; I <= Last; ++I) {
401     if (Clusters[I].Low->getValue() != Clusters[I - 1].High->getValue() + 1) {
402       ContiguousRange = false;
403       break;
404     }
405   }
406 
407   if (Low.isStrictlyPositive() && High.slt(BitWidth)) {
408     // Optimize the case where all the case values fit in a word without having
409     // to subtract minValue. In this case, we can optimize away the subtraction.
410     LowBound = APInt::getZero(Low.getBitWidth());
411     CmpRange = High;
412     ContiguousRange = false;
413   } else {
414     LowBound = Low;
415     CmpRange = High - Low;
416   }
417 
418   CaseBitsVector CBV;
419   auto TotalProb = BranchProbability::getZero();
420   for (unsigned i = First; i <= Last; ++i) {
421     // Find the CaseBits for this destination.
422     unsigned j;
423     for (j = 0; j < CBV.size(); ++j)
424       if (CBV[j].BB == Clusters[i].MBB)
425         break;
426     if (j == CBV.size())
427       CBV.push_back(
428           CaseBits(0, Clusters[i].MBB, 0, BranchProbability::getZero()));
429     CaseBits *CB = &CBV[j];
430 
431     // Update Mask, Bits and ExtraProb.
432     uint64_t Lo = (Clusters[i].Low->getValue() - LowBound).getZExtValue();
433     uint64_t Hi = (Clusters[i].High->getValue() - LowBound).getZExtValue();
434     assert(Hi >= Lo && Hi < 64 && "Invalid bit case!");
435     CB->Mask |= (-1ULL >> (63 - (Hi - Lo))) << Lo;
436     CB->Bits += Hi - Lo + 1;
437     CB->ExtraProb += Clusters[i].Prob;
438     TotalProb += Clusters[i].Prob;
439   }
440 
441   BitTestInfo BTI;
442   llvm::sort(CBV, [](const CaseBits &a, const CaseBits &b) {
443     // Sort by probability first, number of bits second, bit mask third.
444     if (a.ExtraProb != b.ExtraProb)
445       return a.ExtraProb > b.ExtraProb;
446     if (a.Bits != b.Bits)
447       return a.Bits > b.Bits;
448     return a.Mask < b.Mask;
449   });
450 
451   for (auto &CB : CBV) {
452     MachineBasicBlock *BitTestBB =
453         FuncInfo.MF->CreateMachineBasicBlock(SI->getParent());
454     BTI.push_back(BitTestCase(CB.Mask, BitTestBB, CB.BB, CB.ExtraProb));
455   }
456   BitTestCases.emplace_back(std::move(LowBound), std::move(CmpRange),
457                             SI->getCondition(), -1U, MVT::Other, false,
458                             ContiguousRange, nullptr, nullptr, std::move(BTI),
459                             TotalProb);
460 
461   BTCluster = CaseCluster::bitTests(Clusters[First].Low, Clusters[Last].High,
462                                     BitTestCases.size() - 1, TotalProb);
463   return true;
464 }
465 
466 void SwitchCG::sortAndRangeify(CaseClusterVector &Clusters) {
467 #ifndef NDEBUG
468   for (const CaseCluster &CC : Clusters)
469     assert(CC.Low == CC.High && "Input clusters must be single-case");
470 #endif
471 
472   llvm::sort(Clusters, [](const CaseCluster &a, const CaseCluster &b) {
473     return a.Low->getValue().slt(b.Low->getValue());
474   });
475 
476   // Merge adjacent clusters with the same destination.
477   const unsigned N = Clusters.size();
478   unsigned DstIndex = 0;
479   for (unsigned SrcIndex = 0; SrcIndex < N; ++SrcIndex) {
480     CaseCluster &CC = Clusters[SrcIndex];
481     const ConstantInt *CaseVal = CC.Low;
482     MachineBasicBlock *Succ = CC.MBB;
483 
484     if (DstIndex != 0 && Clusters[DstIndex - 1].MBB == Succ &&
485         (CaseVal->getValue() - Clusters[DstIndex - 1].High->getValue()) == 1) {
486       // If this case has the same successor and is a neighbour, merge it into
487       // the previous cluster.
488       Clusters[DstIndex - 1].High = CaseVal;
489       Clusters[DstIndex - 1].Prob += CC.Prob;
490     } else {
491       std::memmove(&Clusters[DstIndex++], &Clusters[SrcIndex],
492                    sizeof(Clusters[SrcIndex]));
493     }
494   }
495   Clusters.resize(DstIndex);
496 }
497 
498 unsigned SwitchCG::SwitchLowering::caseClusterRank(const CaseCluster &CC,
499                                                    CaseClusterIt First,
500                                                    CaseClusterIt Last) {
501   return std::count_if(First, Last + 1, [&](const CaseCluster &X) {
502     if (X.Prob != CC.Prob)
503       return X.Prob > CC.Prob;
504 
505     // Ties are broken by comparing the case value.
506     return X.Low->getValue().slt(CC.Low->getValue());
507   });
508 }
509 
510 llvm::SwitchCG::SwitchLowering::SplitWorkItemInfo
511 SwitchCG::SwitchLowering::computeSplitWorkItemInfo(
512     const SwitchWorkListItem &W) {
513   CaseClusterIt LastLeft = W.FirstCluster;
514   CaseClusterIt FirstRight = W.LastCluster;
515   auto LeftProb = LastLeft->Prob + W.DefaultProb / 2;
516   auto RightProb = FirstRight->Prob + W.DefaultProb / 2;
517 
518   // Move LastLeft and FirstRight towards each other from opposite directions to
519   // find a partitioning of the clusters which balances the probability on both
520   // sides. If LeftProb and RightProb are equal, alternate which side is
521   // taken to ensure 0-probability nodes are distributed evenly.
522   unsigned I = 0;
523   while (LastLeft + 1 < FirstRight) {
524     if (LeftProb < RightProb || (LeftProb == RightProb && (I & 1)))
525       LeftProb += (++LastLeft)->Prob;
526     else
527       RightProb += (--FirstRight)->Prob;
528     I++;
529   }
530 
531   while (true) {
532     // Our binary search tree differs from a typical BST in that ours can have
533     // up to three values in each leaf. The pivot selection above doesn't take
534     // that into account, which means the tree might require more nodes and be
535     // less efficient. We compensate for this here.
536 
537     unsigned NumLeft = LastLeft - W.FirstCluster + 1;
538     unsigned NumRight = W.LastCluster - FirstRight + 1;
539 
540     if (std::min(NumLeft, NumRight) < 3 && std::max(NumLeft, NumRight) > 3) {
541       // If one side has less than 3 clusters, and the other has more than 3,
542       // consider taking a cluster from the other side.
543 
544       if (NumLeft < NumRight) {
545         // Consider moving the first cluster on the right to the left side.
546         CaseCluster &CC = *FirstRight;
547         unsigned RightSideRank = caseClusterRank(CC, FirstRight, W.LastCluster);
548         unsigned LeftSideRank = caseClusterRank(CC, W.FirstCluster, LastLeft);
549         if (LeftSideRank <= RightSideRank) {
550           // Moving the cluster to the left does not demote it.
551           ++LastLeft;
552           ++FirstRight;
553           continue;
554         }
555       } else {
556         assert(NumRight < NumLeft);
557         // Consider moving the last element on the left to the right side.
558         CaseCluster &CC = *LastLeft;
559         unsigned LeftSideRank = caseClusterRank(CC, W.FirstCluster, LastLeft);
560         unsigned RightSideRank = caseClusterRank(CC, FirstRight, W.LastCluster);
561         if (RightSideRank <= LeftSideRank) {
562           // Moving the cluster to the right does not demot it.
563           --LastLeft;
564           --FirstRight;
565           continue;
566         }
567       }
568     }
569     break;
570   }
571 
572   assert(LastLeft + 1 == FirstRight);
573   assert(LastLeft >= W.FirstCluster);
574   assert(FirstRight <= W.LastCluster);
575 
576   return SplitWorkItemInfo{LastLeft, FirstRight, LeftProb, RightProb};
577 }