xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/MachineBlockPlacement.cpp (revision 6580f5c38dd5b01aeeaed16b370f1a12423437f0)
1 //===- MachineBlockPlacement.cpp - Basic Block Code Layout optimization ---===//
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 implements basic block placement transformations using the CFG
10 // structure and branch probability estimates.
11 //
12 // The pass strives to preserve the structure of the CFG (that is, retain
13 // a topological ordering of basic blocks) in the absence of a *strong* signal
14 // to the contrary from probabilities. However, within the CFG structure, it
15 // attempts to choose an ordering which favors placing more likely sequences of
16 // blocks adjacent to each other.
17 //
18 // The algorithm works from the inner-most loop within a function outward, and
19 // at each stage walks through the basic blocks, trying to coalesce them into
20 // sequential chains where allowed by the CFG (or demanded by heavy
21 // probabilities). Finally, it walks the blocks in topological order, and the
22 // first time it reaches a chain of basic blocks, it schedules them in the
23 // function in-order.
24 //
25 //===----------------------------------------------------------------------===//
26 
27 #include "BranchFolding.h"
28 #include "llvm/ADT/ArrayRef.h"
29 #include "llvm/ADT/DenseMap.h"
30 #include "llvm/ADT/STLExtras.h"
31 #include "llvm/ADT/SetVector.h"
32 #include "llvm/ADT/SmallPtrSet.h"
33 #include "llvm/ADT/SmallVector.h"
34 #include "llvm/ADT/Statistic.h"
35 #include "llvm/Analysis/BlockFrequencyInfoImpl.h"
36 #include "llvm/Analysis/ProfileSummaryInfo.h"
37 #include "llvm/CodeGen/MBFIWrapper.h"
38 #include "llvm/CodeGen/MachineBasicBlock.h"
39 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
40 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
41 #include "llvm/CodeGen/MachineFunction.h"
42 #include "llvm/CodeGen/MachineFunctionPass.h"
43 #include "llvm/CodeGen/MachineLoopInfo.h"
44 #include "llvm/CodeGen/MachinePostDominators.h"
45 #include "llvm/CodeGen/MachineSizeOpts.h"
46 #include "llvm/CodeGen/TailDuplicator.h"
47 #include "llvm/CodeGen/TargetInstrInfo.h"
48 #include "llvm/CodeGen/TargetLowering.h"
49 #include "llvm/CodeGen/TargetPassConfig.h"
50 #include "llvm/CodeGen/TargetSubtargetInfo.h"
51 #include "llvm/IR/DebugLoc.h"
52 #include "llvm/IR/Function.h"
53 #include "llvm/IR/PrintPasses.h"
54 #include "llvm/InitializePasses.h"
55 #include "llvm/Pass.h"
56 #include "llvm/Support/Allocator.h"
57 #include "llvm/Support/BlockFrequency.h"
58 #include "llvm/Support/BranchProbability.h"
59 #include "llvm/Support/CodeGen.h"
60 #include "llvm/Support/CommandLine.h"
61 #include "llvm/Support/Compiler.h"
62 #include "llvm/Support/Debug.h"
63 #include "llvm/Support/raw_ostream.h"
64 #include "llvm/Target/TargetMachine.h"
65 #include "llvm/Transforms/Utils/CodeLayout.h"
66 #include <algorithm>
67 #include <cassert>
68 #include <cstdint>
69 #include <iterator>
70 #include <memory>
71 #include <string>
72 #include <tuple>
73 #include <utility>
74 #include <vector>
75 
76 using namespace llvm;
77 
78 #define DEBUG_TYPE "block-placement"
79 
80 STATISTIC(NumCondBranches, "Number of conditional branches");
81 STATISTIC(NumUncondBranches, "Number of unconditional branches");
82 STATISTIC(CondBranchTakenFreq,
83           "Potential frequency of taking conditional branches");
84 STATISTIC(UncondBranchTakenFreq,
85           "Potential frequency of taking unconditional branches");
86 
87 static cl::opt<unsigned> AlignAllBlock(
88     "align-all-blocks",
89     cl::desc("Force the alignment of all blocks in the function in log2 format "
90              "(e.g 4 means align on 16B boundaries)."),
91     cl::init(0), cl::Hidden);
92 
93 static cl::opt<unsigned> AlignAllNonFallThruBlocks(
94     "align-all-nofallthru-blocks",
95     cl::desc("Force the alignment of all blocks that have no fall-through "
96              "predecessors (i.e. don't add nops that are executed). In log2 "
97              "format (e.g 4 means align on 16B boundaries)."),
98     cl::init(0), cl::Hidden);
99 
100 static cl::opt<unsigned> MaxBytesForAlignmentOverride(
101     "max-bytes-for-alignment",
102     cl::desc("Forces the maximum bytes allowed to be emitted when padding for "
103              "alignment"),
104     cl::init(0), cl::Hidden);
105 
106 // FIXME: Find a good default for this flag and remove the flag.
107 static cl::opt<unsigned> ExitBlockBias(
108     "block-placement-exit-block-bias",
109     cl::desc("Block frequency percentage a loop exit block needs "
110              "over the original exit to be considered the new exit."),
111     cl::init(0), cl::Hidden);
112 
113 // Definition:
114 // - Outlining: placement of a basic block outside the chain or hot path.
115 
116 static cl::opt<unsigned> LoopToColdBlockRatio(
117     "loop-to-cold-block-ratio",
118     cl::desc("Outline loop blocks from loop chain if (frequency of loop) / "
119              "(frequency of block) is greater than this ratio"),
120     cl::init(5), cl::Hidden);
121 
122 static cl::opt<bool> ForceLoopColdBlock(
123     "force-loop-cold-block",
124     cl::desc("Force outlining cold blocks from loops."),
125     cl::init(false), cl::Hidden);
126 
127 static cl::opt<bool>
128     PreciseRotationCost("precise-rotation-cost",
129                         cl::desc("Model the cost of loop rotation more "
130                                  "precisely by using profile data."),
131                         cl::init(false), cl::Hidden);
132 
133 static cl::opt<bool>
134     ForcePreciseRotationCost("force-precise-rotation-cost",
135                              cl::desc("Force the use of precise cost "
136                                       "loop rotation strategy."),
137                              cl::init(false), cl::Hidden);
138 
139 static cl::opt<unsigned> MisfetchCost(
140     "misfetch-cost",
141     cl::desc("Cost that models the probabilistic risk of an instruction "
142              "misfetch due to a jump comparing to falling through, whose cost "
143              "is zero."),
144     cl::init(1), cl::Hidden);
145 
146 static cl::opt<unsigned> JumpInstCost("jump-inst-cost",
147                                       cl::desc("Cost of jump instructions."),
148                                       cl::init(1), cl::Hidden);
149 static cl::opt<bool>
150 TailDupPlacement("tail-dup-placement",
151               cl::desc("Perform tail duplication during placement. "
152                        "Creates more fallthrough opportunites in "
153                        "outline branches."),
154               cl::init(true), cl::Hidden);
155 
156 static cl::opt<bool>
157 BranchFoldPlacement("branch-fold-placement",
158               cl::desc("Perform branch folding during placement. "
159                        "Reduces code size."),
160               cl::init(true), cl::Hidden);
161 
162 // Heuristic for tail duplication.
163 static cl::opt<unsigned> TailDupPlacementThreshold(
164     "tail-dup-placement-threshold",
165     cl::desc("Instruction cutoff for tail duplication during layout. "
166              "Tail merging during layout is forced to have a threshold "
167              "that won't conflict."), cl::init(2),
168     cl::Hidden);
169 
170 // Heuristic for aggressive tail duplication.
171 static cl::opt<unsigned> TailDupPlacementAggressiveThreshold(
172     "tail-dup-placement-aggressive-threshold",
173     cl::desc("Instruction cutoff for aggressive tail duplication during "
174              "layout. Used at -O3. Tail merging during layout is forced to "
175              "have a threshold that won't conflict."), cl::init(4),
176     cl::Hidden);
177 
178 // Heuristic for tail duplication.
179 static cl::opt<unsigned> TailDupPlacementPenalty(
180     "tail-dup-placement-penalty",
181     cl::desc("Cost penalty for blocks that can avoid breaking CFG by copying. "
182              "Copying can increase fallthrough, but it also increases icache "
183              "pressure. This parameter controls the penalty to account for that. "
184              "Percent as integer."),
185     cl::init(2),
186     cl::Hidden);
187 
188 // Heuristic for tail duplication if profile count is used in cost model.
189 static cl::opt<unsigned> TailDupProfilePercentThreshold(
190     "tail-dup-profile-percent-threshold",
191     cl::desc("If profile count information is used in tail duplication cost "
192              "model, the gained fall through number from tail duplication "
193              "should be at least this percent of hot count."),
194     cl::init(50), cl::Hidden);
195 
196 // Heuristic for triangle chains.
197 static cl::opt<unsigned> TriangleChainCount(
198     "triangle-chain-count",
199     cl::desc("Number of triangle-shaped-CFG's that need to be in a row for the "
200              "triangle tail duplication heuristic to kick in. 0 to disable."),
201     cl::init(2),
202     cl::Hidden);
203 
204 // Use case: When block layout is visualized after MBP pass, the basic blocks
205 // are labeled in layout order; meanwhile blocks could be numbered in a
206 // different order. It's hard to map between the graph and pass output.
207 // With this option on, the basic blocks are renumbered in function layout
208 // order. For debugging only.
209 static cl::opt<bool> RenumberBlocksBeforeView(
210     "renumber-blocks-before-view",
211     cl::desc(
212         "If true, basic blocks are re-numbered before MBP layout is printed "
213         "into a dot graph. Only used when a function is being printed."),
214     cl::init(false), cl::Hidden);
215 
216 namespace llvm {
217 extern cl::opt<bool> EnableExtTspBlockPlacement;
218 extern cl::opt<bool> ApplyExtTspWithoutProfile;
219 extern cl::opt<unsigned> StaticLikelyProb;
220 extern cl::opt<unsigned> ProfileLikelyProb;
221 
222 // Internal option used to control BFI display only after MBP pass.
223 // Defined in CodeGen/MachineBlockFrequencyInfo.cpp:
224 // -view-block-layout-with-bfi=
225 extern cl::opt<GVDAGType> ViewBlockLayoutWithBFI;
226 
227 // Command line option to specify the name of the function for CFG dump
228 // Defined in Analysis/BlockFrequencyInfo.cpp:  -view-bfi-func-name=
229 extern cl::opt<std::string> ViewBlockFreqFuncName;
230 } // namespace llvm
231 
232 namespace {
233 
234 class BlockChain;
235 
236 /// Type for our function-wide basic block -> block chain mapping.
237 using BlockToChainMapType = DenseMap<const MachineBasicBlock *, BlockChain *>;
238 
239 /// A chain of blocks which will be laid out contiguously.
240 ///
241 /// This is the datastructure representing a chain of consecutive blocks that
242 /// are profitable to layout together in order to maximize fallthrough
243 /// probabilities and code locality. We also can use a block chain to represent
244 /// a sequence of basic blocks which have some external (correctness)
245 /// requirement for sequential layout.
246 ///
247 /// Chains can be built around a single basic block and can be merged to grow
248 /// them. They participate in a block-to-chain mapping, which is updated
249 /// automatically as chains are merged together.
250 class BlockChain {
251   /// The sequence of blocks belonging to this chain.
252   ///
253   /// This is the sequence of blocks for a particular chain. These will be laid
254   /// out in-order within the function.
255   SmallVector<MachineBasicBlock *, 4> Blocks;
256 
257   /// A handle to the function-wide basic block to block chain mapping.
258   ///
259   /// This is retained in each block chain to simplify the computation of child
260   /// block chains for SCC-formation and iteration. We store the edges to child
261   /// basic blocks, and map them back to their associated chains using this
262   /// structure.
263   BlockToChainMapType &BlockToChain;
264 
265 public:
266   /// Construct a new BlockChain.
267   ///
268   /// This builds a new block chain representing a single basic block in the
269   /// function. It also registers itself as the chain that block participates
270   /// in with the BlockToChain mapping.
271   BlockChain(BlockToChainMapType &BlockToChain, MachineBasicBlock *BB)
272       : Blocks(1, BB), BlockToChain(BlockToChain) {
273     assert(BB && "Cannot create a chain with a null basic block");
274     BlockToChain[BB] = this;
275   }
276 
277   /// Iterator over blocks within the chain.
278   using iterator = SmallVectorImpl<MachineBasicBlock *>::iterator;
279   using const_iterator = SmallVectorImpl<MachineBasicBlock *>::const_iterator;
280 
281   /// Beginning of blocks within the chain.
282   iterator begin() { return Blocks.begin(); }
283   const_iterator begin() const { return Blocks.begin(); }
284 
285   /// End of blocks within the chain.
286   iterator end() { return Blocks.end(); }
287   const_iterator end() const { return Blocks.end(); }
288 
289   bool remove(MachineBasicBlock* BB) {
290     for(iterator i = begin(); i != end(); ++i) {
291       if (*i == BB) {
292         Blocks.erase(i);
293         return true;
294       }
295     }
296     return false;
297   }
298 
299   /// Merge a block chain into this one.
300   ///
301   /// This routine merges a block chain into this one. It takes care of forming
302   /// a contiguous sequence of basic blocks, updating the edge list, and
303   /// updating the block -> chain mapping. It does not free or tear down the
304   /// old chain, but the old chain's block list is no longer valid.
305   void merge(MachineBasicBlock *BB, BlockChain *Chain) {
306     assert(BB && "Can't merge a null block.");
307     assert(!Blocks.empty() && "Can't merge into an empty chain.");
308 
309     // Fast path in case we don't have a chain already.
310     if (!Chain) {
311       assert(!BlockToChain[BB] &&
312              "Passed chain is null, but BB has entry in BlockToChain.");
313       Blocks.push_back(BB);
314       BlockToChain[BB] = this;
315       return;
316     }
317 
318     assert(BB == *Chain->begin() && "Passed BB is not head of Chain.");
319     assert(Chain->begin() != Chain->end());
320 
321     // Update the incoming blocks to point to this chain, and add them to the
322     // chain structure.
323     for (MachineBasicBlock *ChainBB : *Chain) {
324       Blocks.push_back(ChainBB);
325       assert(BlockToChain[ChainBB] == Chain && "Incoming blocks not in chain.");
326       BlockToChain[ChainBB] = this;
327     }
328   }
329 
330 #ifndef NDEBUG
331   /// Dump the blocks in this chain.
332   LLVM_DUMP_METHOD void dump() {
333     for (MachineBasicBlock *MBB : *this)
334       MBB->dump();
335   }
336 #endif // NDEBUG
337 
338   /// Count of predecessors of any block within the chain which have not
339   /// yet been scheduled.  In general, we will delay scheduling this chain
340   /// until those predecessors are scheduled (or we find a sufficiently good
341   /// reason to override this heuristic.)  Note that when forming loop chains,
342   /// blocks outside the loop are ignored and treated as if they were already
343   /// scheduled.
344   ///
345   /// Note: This field is reinitialized multiple times - once for each loop,
346   /// and then once for the function as a whole.
347   unsigned UnscheduledPredecessors = 0;
348 };
349 
350 class MachineBlockPlacement : public MachineFunctionPass {
351   /// A type for a block filter set.
352   using BlockFilterSet = SmallSetVector<const MachineBasicBlock *, 16>;
353 
354   /// Pair struct containing basic block and taildup profitability
355   struct BlockAndTailDupResult {
356     MachineBasicBlock *BB = nullptr;
357     bool ShouldTailDup;
358   };
359 
360   /// Triple struct containing edge weight and the edge.
361   struct WeightedEdge {
362     BlockFrequency Weight;
363     MachineBasicBlock *Src = nullptr;
364     MachineBasicBlock *Dest = nullptr;
365   };
366 
367   /// work lists of blocks that are ready to be laid out
368   SmallVector<MachineBasicBlock *, 16> BlockWorkList;
369   SmallVector<MachineBasicBlock *, 16> EHPadWorkList;
370 
371   /// Edges that have already been computed as optimal.
372   DenseMap<const MachineBasicBlock *, BlockAndTailDupResult> ComputedEdges;
373 
374   /// Machine Function
375   MachineFunction *F = nullptr;
376 
377   /// A handle to the branch probability pass.
378   const MachineBranchProbabilityInfo *MBPI = nullptr;
379 
380   /// A handle to the function-wide block frequency pass.
381   std::unique_ptr<MBFIWrapper> MBFI;
382 
383   /// A handle to the loop info.
384   MachineLoopInfo *MLI = nullptr;
385 
386   /// Preferred loop exit.
387   /// Member variable for convenience. It may be removed by duplication deep
388   /// in the call stack.
389   MachineBasicBlock *PreferredLoopExit = nullptr;
390 
391   /// A handle to the target's instruction info.
392   const TargetInstrInfo *TII = nullptr;
393 
394   /// A handle to the target's lowering info.
395   const TargetLoweringBase *TLI = nullptr;
396 
397   /// A handle to the post dominator tree.
398   MachinePostDominatorTree *MPDT = nullptr;
399 
400   ProfileSummaryInfo *PSI = nullptr;
401 
402   /// Duplicator used to duplicate tails during placement.
403   ///
404   /// Placement decisions can open up new tail duplication opportunities, but
405   /// since tail duplication affects placement decisions of later blocks, it
406   /// must be done inline.
407   TailDuplicator TailDup;
408 
409   /// Partial tail duplication threshold.
410   BlockFrequency DupThreshold;
411 
412   /// True:  use block profile count to compute tail duplication cost.
413   /// False: use block frequency to compute tail duplication cost.
414   bool UseProfileCount = false;
415 
416   /// Allocator and owner of BlockChain structures.
417   ///
418   /// We build BlockChains lazily while processing the loop structure of
419   /// a function. To reduce malloc traffic, we allocate them using this
420   /// slab-like allocator, and destroy them after the pass completes. An
421   /// important guarantee is that this allocator produces stable pointers to
422   /// the chains.
423   SpecificBumpPtrAllocator<BlockChain> ChainAllocator;
424 
425   /// Function wide BasicBlock to BlockChain mapping.
426   ///
427   /// This mapping allows efficiently moving from any given basic block to the
428   /// BlockChain it participates in, if any. We use it to, among other things,
429   /// allow implicitly defining edges between chains as the existing edges
430   /// between basic blocks.
431   DenseMap<const MachineBasicBlock *, BlockChain *> BlockToChain;
432 
433 #ifndef NDEBUG
434   /// The set of basic blocks that have terminators that cannot be fully
435   /// analyzed.  These basic blocks cannot be re-ordered safely by
436   /// MachineBlockPlacement, and we must preserve physical layout of these
437   /// blocks and their successors through the pass.
438   SmallPtrSet<MachineBasicBlock *, 4> BlocksWithUnanalyzableExits;
439 #endif
440 
441   /// Get block profile count or frequency according to UseProfileCount.
442   /// The return value is used to model tail duplication cost.
443   BlockFrequency getBlockCountOrFrequency(const MachineBasicBlock *BB) {
444     if (UseProfileCount) {
445       auto Count = MBFI->getBlockProfileCount(BB);
446       if (Count)
447         return BlockFrequency(*Count);
448       else
449         return BlockFrequency(0);
450     } else
451       return MBFI->getBlockFreq(BB);
452   }
453 
454   /// Scale the DupThreshold according to basic block size.
455   BlockFrequency scaleThreshold(MachineBasicBlock *BB);
456   void initDupThreshold();
457 
458   /// Decrease the UnscheduledPredecessors count for all blocks in chain, and
459   /// if the count goes to 0, add them to the appropriate work list.
460   void markChainSuccessors(
461       const BlockChain &Chain, const MachineBasicBlock *LoopHeaderBB,
462       const BlockFilterSet *BlockFilter = nullptr);
463 
464   /// Decrease the UnscheduledPredecessors count for a single block, and
465   /// if the count goes to 0, add them to the appropriate work list.
466   void markBlockSuccessors(
467       const BlockChain &Chain, const MachineBasicBlock *BB,
468       const MachineBasicBlock *LoopHeaderBB,
469       const BlockFilterSet *BlockFilter = nullptr);
470 
471   BranchProbability
472   collectViableSuccessors(
473       const MachineBasicBlock *BB, const BlockChain &Chain,
474       const BlockFilterSet *BlockFilter,
475       SmallVector<MachineBasicBlock *, 4> &Successors);
476   bool isBestSuccessor(MachineBasicBlock *BB, MachineBasicBlock *Pred,
477                        BlockFilterSet *BlockFilter);
478   void findDuplicateCandidates(SmallVectorImpl<MachineBasicBlock *> &Candidates,
479                                MachineBasicBlock *BB,
480                                BlockFilterSet *BlockFilter);
481   bool repeatedlyTailDuplicateBlock(
482       MachineBasicBlock *BB, MachineBasicBlock *&LPred,
483       const MachineBasicBlock *LoopHeaderBB,
484       BlockChain &Chain, BlockFilterSet *BlockFilter,
485       MachineFunction::iterator &PrevUnplacedBlockIt);
486   bool maybeTailDuplicateBlock(
487       MachineBasicBlock *BB, MachineBasicBlock *LPred,
488       BlockChain &Chain, BlockFilterSet *BlockFilter,
489       MachineFunction::iterator &PrevUnplacedBlockIt,
490       bool &DuplicatedToLPred);
491   bool hasBetterLayoutPredecessor(
492       const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
493       const BlockChain &SuccChain, BranchProbability SuccProb,
494       BranchProbability RealSuccProb, const BlockChain &Chain,
495       const BlockFilterSet *BlockFilter);
496   BlockAndTailDupResult selectBestSuccessor(
497       const MachineBasicBlock *BB, const BlockChain &Chain,
498       const BlockFilterSet *BlockFilter);
499   MachineBasicBlock *selectBestCandidateBlock(
500       const BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList);
501   MachineBasicBlock *getFirstUnplacedBlock(
502       const BlockChain &PlacedChain,
503       MachineFunction::iterator &PrevUnplacedBlockIt,
504       const BlockFilterSet *BlockFilter);
505 
506   /// Add a basic block to the work list if it is appropriate.
507   ///
508   /// If the optional parameter BlockFilter is provided, only MBB
509   /// present in the set will be added to the worklist. If nullptr
510   /// is provided, no filtering occurs.
511   void fillWorkLists(const MachineBasicBlock *MBB,
512                      SmallPtrSetImpl<BlockChain *> &UpdatedPreds,
513                      const BlockFilterSet *BlockFilter);
514 
515   void buildChain(const MachineBasicBlock *BB, BlockChain &Chain,
516                   BlockFilterSet *BlockFilter = nullptr);
517   bool canMoveBottomBlockToTop(const MachineBasicBlock *BottomBlock,
518                                const MachineBasicBlock *OldTop);
519   bool hasViableTopFallthrough(const MachineBasicBlock *Top,
520                                const BlockFilterSet &LoopBlockSet);
521   BlockFrequency TopFallThroughFreq(const MachineBasicBlock *Top,
522                                     const BlockFilterSet &LoopBlockSet);
523   BlockFrequency FallThroughGains(const MachineBasicBlock *NewTop,
524                                   const MachineBasicBlock *OldTop,
525                                   const MachineBasicBlock *ExitBB,
526                                   const BlockFilterSet &LoopBlockSet);
527   MachineBasicBlock *findBestLoopTopHelper(MachineBasicBlock *OldTop,
528       const MachineLoop &L, const BlockFilterSet &LoopBlockSet);
529   MachineBasicBlock *findBestLoopTop(
530       const MachineLoop &L, const BlockFilterSet &LoopBlockSet);
531   MachineBasicBlock *findBestLoopExit(
532       const MachineLoop &L, const BlockFilterSet &LoopBlockSet,
533       BlockFrequency &ExitFreq);
534   BlockFilterSet collectLoopBlockSet(const MachineLoop &L);
535   void buildLoopChains(const MachineLoop &L);
536   void rotateLoop(
537       BlockChain &LoopChain, const MachineBasicBlock *ExitingBB,
538       BlockFrequency ExitFreq, const BlockFilterSet &LoopBlockSet);
539   void rotateLoopWithProfile(
540       BlockChain &LoopChain, const MachineLoop &L,
541       const BlockFilterSet &LoopBlockSet);
542   void buildCFGChains();
543   void optimizeBranches();
544   void alignBlocks();
545   /// Returns true if a block should be tail-duplicated to increase fallthrough
546   /// opportunities.
547   bool shouldTailDuplicate(MachineBasicBlock *BB);
548   /// Check the edge frequencies to see if tail duplication will increase
549   /// fallthroughs.
550   bool isProfitableToTailDup(
551     const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
552     BranchProbability QProb,
553     const BlockChain &Chain, const BlockFilterSet *BlockFilter);
554 
555   /// Check for a trellis layout.
556   bool isTrellis(const MachineBasicBlock *BB,
557                  const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
558                  const BlockChain &Chain, const BlockFilterSet *BlockFilter);
559 
560   /// Get the best successor given a trellis layout.
561   BlockAndTailDupResult getBestTrellisSuccessor(
562       const MachineBasicBlock *BB,
563       const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
564       BranchProbability AdjustedSumProb, const BlockChain &Chain,
565       const BlockFilterSet *BlockFilter);
566 
567   /// Get the best pair of non-conflicting edges.
568   static std::pair<WeightedEdge, WeightedEdge> getBestNonConflictingEdges(
569       const MachineBasicBlock *BB,
570       MutableArrayRef<SmallVector<WeightedEdge, 8>> Edges);
571 
572   /// Returns true if a block can tail duplicate into all unplaced
573   /// predecessors. Filters based on loop.
574   bool canTailDuplicateUnplacedPreds(
575       const MachineBasicBlock *BB, MachineBasicBlock *Succ,
576       const BlockChain &Chain, const BlockFilterSet *BlockFilter);
577 
578   /// Find chains of triangles to tail-duplicate where a global analysis works,
579   /// but a local analysis would not find them.
580   void precomputeTriangleChains();
581 
582   /// Apply a post-processing step optimizing block placement.
583   void applyExtTsp();
584 
585   /// Modify the existing block placement in the function and adjust all jumps.
586   void assignBlockOrder(const std::vector<const MachineBasicBlock *> &NewOrder);
587 
588   /// Create a single CFG chain from the current block order.
589   void createCFGChainExtTsp();
590 
591 public:
592   static char ID; // Pass identification, replacement for typeid
593 
594   MachineBlockPlacement() : MachineFunctionPass(ID) {
595     initializeMachineBlockPlacementPass(*PassRegistry::getPassRegistry());
596   }
597 
598   bool runOnMachineFunction(MachineFunction &F) override;
599 
600   bool allowTailDupPlacement() const {
601     assert(F);
602     return TailDupPlacement && !F->getTarget().requiresStructuredCFG();
603   }
604 
605   void getAnalysisUsage(AnalysisUsage &AU) const override {
606     AU.addRequired<MachineBranchProbabilityInfo>();
607     AU.addRequired<MachineBlockFrequencyInfo>();
608     if (TailDupPlacement)
609       AU.addRequired<MachinePostDominatorTree>();
610     AU.addRequired<MachineLoopInfo>();
611     AU.addRequired<ProfileSummaryInfoWrapperPass>();
612     AU.addRequired<TargetPassConfig>();
613     MachineFunctionPass::getAnalysisUsage(AU);
614   }
615 };
616 
617 } // end anonymous namespace
618 
619 char MachineBlockPlacement::ID = 0;
620 
621 char &llvm::MachineBlockPlacementID = MachineBlockPlacement::ID;
622 
623 INITIALIZE_PASS_BEGIN(MachineBlockPlacement, DEBUG_TYPE,
624                       "Branch Probability Basic Block Placement", false, false)
625 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
626 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
627 INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTree)
628 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
629 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
630 INITIALIZE_PASS_END(MachineBlockPlacement, DEBUG_TYPE,
631                     "Branch Probability Basic Block Placement", false, false)
632 
633 #ifndef NDEBUG
634 /// Helper to print the name of a MBB.
635 ///
636 /// Only used by debug logging.
637 static std::string getBlockName(const MachineBasicBlock *BB) {
638   std::string Result;
639   raw_string_ostream OS(Result);
640   OS << printMBBReference(*BB);
641   OS << " ('" << BB->getName() << "')";
642   OS.flush();
643   return Result;
644 }
645 #endif
646 
647 /// Mark a chain's successors as having one fewer preds.
648 ///
649 /// When a chain is being merged into the "placed" chain, this routine will
650 /// quickly walk the successors of each block in the chain and mark them as
651 /// having one fewer active predecessor. It also adds any successors of this
652 /// chain which reach the zero-predecessor state to the appropriate worklist.
653 void MachineBlockPlacement::markChainSuccessors(
654     const BlockChain &Chain, const MachineBasicBlock *LoopHeaderBB,
655     const BlockFilterSet *BlockFilter) {
656   // Walk all the blocks in this chain, marking their successors as having
657   // a predecessor placed.
658   for (MachineBasicBlock *MBB : Chain) {
659     markBlockSuccessors(Chain, MBB, LoopHeaderBB, BlockFilter);
660   }
661 }
662 
663 /// Mark a single block's successors as having one fewer preds.
664 ///
665 /// Under normal circumstances, this is only called by markChainSuccessors,
666 /// but if a block that was to be placed is completely tail-duplicated away,
667 /// and was duplicated into the chain end, we need to redo markBlockSuccessors
668 /// for just that block.
669 void MachineBlockPlacement::markBlockSuccessors(
670     const BlockChain &Chain, const MachineBasicBlock *MBB,
671     const MachineBasicBlock *LoopHeaderBB, const BlockFilterSet *BlockFilter) {
672   // Add any successors for which this is the only un-placed in-loop
673   // predecessor to the worklist as a viable candidate for CFG-neutral
674   // placement. No subsequent placement of this block will violate the CFG
675   // shape, so we get to use heuristics to choose a favorable placement.
676   for (MachineBasicBlock *Succ : MBB->successors()) {
677     if (BlockFilter && !BlockFilter->count(Succ))
678       continue;
679     BlockChain &SuccChain = *BlockToChain[Succ];
680     // Disregard edges within a fixed chain, or edges to the loop header.
681     if (&Chain == &SuccChain || Succ == LoopHeaderBB)
682       continue;
683 
684     // This is a cross-chain edge that is within the loop, so decrement the
685     // loop predecessor count of the destination chain.
686     if (SuccChain.UnscheduledPredecessors == 0 ||
687         --SuccChain.UnscheduledPredecessors > 0)
688       continue;
689 
690     auto *NewBB = *SuccChain.begin();
691     if (NewBB->isEHPad())
692       EHPadWorkList.push_back(NewBB);
693     else
694       BlockWorkList.push_back(NewBB);
695   }
696 }
697 
698 /// This helper function collects the set of successors of block
699 /// \p BB that are allowed to be its layout successors, and return
700 /// the total branch probability of edges from \p BB to those
701 /// blocks.
702 BranchProbability MachineBlockPlacement::collectViableSuccessors(
703     const MachineBasicBlock *BB, const BlockChain &Chain,
704     const BlockFilterSet *BlockFilter,
705     SmallVector<MachineBasicBlock *, 4> &Successors) {
706   // Adjust edge probabilities by excluding edges pointing to blocks that is
707   // either not in BlockFilter or is already in the current chain. Consider the
708   // following CFG:
709   //
710   //     --->A
711   //     |  / \
712   //     | B   C
713   //     |  \ / \
714   //     ----D   E
715   //
716   // Assume A->C is very hot (>90%), and C->D has a 50% probability, then after
717   // A->C is chosen as a fall-through, D won't be selected as a successor of C
718   // due to CFG constraint (the probability of C->D is not greater than
719   // HotProb to break topo-order). If we exclude E that is not in BlockFilter
720   // when calculating the probability of C->D, D will be selected and we
721   // will get A C D B as the layout of this loop.
722   auto AdjustedSumProb = BranchProbability::getOne();
723   for (MachineBasicBlock *Succ : BB->successors()) {
724     bool SkipSucc = false;
725     if (Succ->isEHPad() || (BlockFilter && !BlockFilter->count(Succ))) {
726       SkipSucc = true;
727     } else {
728       BlockChain *SuccChain = BlockToChain[Succ];
729       if (SuccChain == &Chain) {
730         SkipSucc = true;
731       } else if (Succ != *SuccChain->begin()) {
732         LLVM_DEBUG(dbgs() << "    " << getBlockName(Succ)
733                           << " -> Mid chain!\n");
734         continue;
735       }
736     }
737     if (SkipSucc)
738       AdjustedSumProb -= MBPI->getEdgeProbability(BB, Succ);
739     else
740       Successors.push_back(Succ);
741   }
742 
743   return AdjustedSumProb;
744 }
745 
746 /// The helper function returns the branch probability that is adjusted
747 /// or normalized over the new total \p AdjustedSumProb.
748 static BranchProbability
749 getAdjustedProbability(BranchProbability OrigProb,
750                        BranchProbability AdjustedSumProb) {
751   BranchProbability SuccProb;
752   uint32_t SuccProbN = OrigProb.getNumerator();
753   uint32_t SuccProbD = AdjustedSumProb.getNumerator();
754   if (SuccProbN >= SuccProbD)
755     SuccProb = BranchProbability::getOne();
756   else
757     SuccProb = BranchProbability(SuccProbN, SuccProbD);
758 
759   return SuccProb;
760 }
761 
762 /// Check if \p BB has exactly the successors in \p Successors.
763 static bool
764 hasSameSuccessors(MachineBasicBlock &BB,
765                   SmallPtrSetImpl<const MachineBasicBlock *> &Successors) {
766   if (BB.succ_size() != Successors.size())
767     return false;
768   // We don't want to count self-loops
769   if (Successors.count(&BB))
770     return false;
771   for (MachineBasicBlock *Succ : BB.successors())
772     if (!Successors.count(Succ))
773       return false;
774   return true;
775 }
776 
777 /// Check if a block should be tail duplicated to increase fallthrough
778 /// opportunities.
779 /// \p BB Block to check.
780 bool MachineBlockPlacement::shouldTailDuplicate(MachineBasicBlock *BB) {
781   // Blocks with single successors don't create additional fallthrough
782   // opportunities. Don't duplicate them. TODO: When conditional exits are
783   // analyzable, allow them to be duplicated.
784   bool IsSimple = TailDup.isSimpleBB(BB);
785 
786   if (BB->succ_size() == 1)
787     return false;
788   return TailDup.shouldTailDuplicate(IsSimple, *BB);
789 }
790 
791 /// Compare 2 BlockFrequency's with a small penalty for \p A.
792 /// In order to be conservative, we apply a X% penalty to account for
793 /// increased icache pressure and static heuristics. For small frequencies
794 /// we use only the numerators to improve accuracy. For simplicity, we assume the
795 /// penalty is less than 100%
796 /// TODO(iteratee): Use 64-bit fixed point edge frequencies everywhere.
797 static bool greaterWithBias(BlockFrequency A, BlockFrequency B,
798                             BlockFrequency EntryFreq) {
799   BranchProbability ThresholdProb(TailDupPlacementPenalty, 100);
800   BlockFrequency Gain = A - B;
801   return (Gain / ThresholdProb) >= EntryFreq;
802 }
803 
804 /// Check the edge frequencies to see if tail duplication will increase
805 /// fallthroughs. It only makes sense to call this function when
806 /// \p Succ would not be chosen otherwise. Tail duplication of \p Succ is
807 /// always locally profitable if we would have picked \p Succ without
808 /// considering duplication.
809 bool MachineBlockPlacement::isProfitableToTailDup(
810     const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
811     BranchProbability QProb,
812     const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
813   // We need to do a probability calculation to make sure this is profitable.
814   // First: does succ have a successor that post-dominates? This affects the
815   // calculation. The 2 relevant cases are:
816   //    BB         BB
817   //    | \Qout    | \Qout
818   //   P|  C       |P C
819   //    =   C'     =   C'
820   //    |  /Qin    |  /Qin
821   //    | /        | /
822   //    Succ       Succ
823   //    / \        | \  V
824   //  U/   =V      |U \
825   //  /     \      =   D
826   //  D      E     |  /
827   //               | /
828   //               |/
829   //               PDom
830   //  '=' : Branch taken for that CFG edge
831   // In the second case, Placing Succ while duplicating it into C prevents the
832   // fallthrough of Succ into either D or PDom, because they now have C as an
833   // unplaced predecessor
834 
835   // Start by figuring out which case we fall into
836   MachineBasicBlock *PDom = nullptr;
837   SmallVector<MachineBasicBlock *, 4> SuccSuccs;
838   // Only scan the relevant successors
839   auto AdjustedSuccSumProb =
840       collectViableSuccessors(Succ, Chain, BlockFilter, SuccSuccs);
841   BranchProbability PProb = MBPI->getEdgeProbability(BB, Succ);
842   auto BBFreq = MBFI->getBlockFreq(BB);
843   auto SuccFreq = MBFI->getBlockFreq(Succ);
844   BlockFrequency P = BBFreq * PProb;
845   BlockFrequency Qout = BBFreq * QProb;
846   BlockFrequency EntryFreq = MBFI->getEntryFreq();
847   // If there are no more successors, it is profitable to copy, as it strictly
848   // increases fallthrough.
849   if (SuccSuccs.size() == 0)
850     return greaterWithBias(P, Qout, EntryFreq);
851 
852   auto BestSuccSucc = BranchProbability::getZero();
853   // Find the PDom or the best Succ if no PDom exists.
854   for (MachineBasicBlock *SuccSucc : SuccSuccs) {
855     auto Prob = MBPI->getEdgeProbability(Succ, SuccSucc);
856     if (Prob > BestSuccSucc)
857       BestSuccSucc = Prob;
858     if (PDom == nullptr)
859       if (MPDT->dominates(SuccSucc, Succ)) {
860         PDom = SuccSucc;
861         break;
862       }
863   }
864   // For the comparisons, we need to know Succ's best incoming edge that isn't
865   // from BB.
866   auto SuccBestPred = BlockFrequency(0);
867   for (MachineBasicBlock *SuccPred : Succ->predecessors()) {
868     if (SuccPred == Succ || SuccPred == BB
869         || BlockToChain[SuccPred] == &Chain
870         || (BlockFilter && !BlockFilter->count(SuccPred)))
871       continue;
872     auto Freq = MBFI->getBlockFreq(SuccPred)
873         * MBPI->getEdgeProbability(SuccPred, Succ);
874     if (Freq > SuccBestPred)
875       SuccBestPred = Freq;
876   }
877   // Qin is Succ's best unplaced incoming edge that isn't BB
878   BlockFrequency Qin = SuccBestPred;
879   // If it doesn't have a post-dominating successor, here is the calculation:
880   //    BB        BB
881   //    | \Qout   |  \
882   //   P|  C      |   =
883   //    =   C'    |    C
884   //    |  /Qin   |     |
885   //    | /       |     C' (+Succ)
886   //    Succ      Succ /|
887   //    / \       |  \/ |
888   //  U/   =V     |  == |
889   //  /     \     | /  \|
890   //  D      E    D     E
891   //  '=' : Branch taken for that CFG edge
892   //  Cost in the first case is: P + V
893   //  For this calculation, we always assume P > Qout. If Qout > P
894   //  The result of this function will be ignored at the caller.
895   //  Let F = SuccFreq - Qin
896   //  Cost in the second case is: Qout + min(Qin, F) * U + max(Qin, F) * V
897 
898   if (PDom == nullptr || !Succ->isSuccessor(PDom)) {
899     BranchProbability UProb = BestSuccSucc;
900     BranchProbability VProb = AdjustedSuccSumProb - UProb;
901     BlockFrequency F = SuccFreq - Qin;
902     BlockFrequency V = SuccFreq * VProb;
903     BlockFrequency QinU = std::min(Qin, F) * UProb;
904     BlockFrequency BaseCost = P + V;
905     BlockFrequency DupCost = Qout + QinU + std::max(Qin, F) * VProb;
906     return greaterWithBias(BaseCost, DupCost, EntryFreq);
907   }
908   BranchProbability UProb = MBPI->getEdgeProbability(Succ, PDom);
909   BranchProbability VProb = AdjustedSuccSumProb - UProb;
910   BlockFrequency U = SuccFreq * UProb;
911   BlockFrequency V = SuccFreq * VProb;
912   BlockFrequency F = SuccFreq - Qin;
913   // If there is a post-dominating successor, here is the calculation:
914   // BB         BB                 BB          BB
915   // | \Qout    |   \               | \Qout     |  \
916   // |P C       |    =              |P C        |   =
917   // =   C'     |P    C             =   C'      |P   C
918   // |  /Qin    |      |            |  /Qin     |     |
919   // | /        |      C' (+Succ)   | /         |     C' (+Succ)
920   // Succ       Succ  /|            Succ        Succ /|
921   // | \  V     |   \/ |            | \  V      |  \/ |
922   // |U \       |U  /\ =?           |U =        |U /\ |
923   // =   D      = =  =?|            |   D       | =  =|
924   // |  /       |/     D            |  /        |/    D
925   // | /        |     /             | =         |    /
926   // |/         |    /              |/          |   =
927   // Dom         Dom                Dom         Dom
928   //  '=' : Branch taken for that CFG edge
929   // The cost for taken branches in the first case is P + U
930   // Let F = SuccFreq - Qin
931   // The cost in the second case (assuming independence), given the layout:
932   // BB, Succ, (C+Succ), D, Dom or the layout:
933   // BB, Succ, D, Dom, (C+Succ)
934   // is Qout + max(F, Qin) * U + min(F, Qin)
935   // compare P + U vs Qout + P * U + Qin.
936   //
937   // The 3rd and 4th cases cover when Dom would be chosen to follow Succ.
938   //
939   // For the 3rd case, the cost is P + 2 * V
940   // For the 4th case, the cost is Qout + min(Qin, F) * U + max(Qin, F) * V + V
941   // We choose 4 over 3 when (P + V) > Qout + min(Qin, F) * U + max(Qin, F) * V
942   if (UProb > AdjustedSuccSumProb / 2 &&
943       !hasBetterLayoutPredecessor(Succ, PDom, *BlockToChain[PDom], UProb, UProb,
944                                   Chain, BlockFilter))
945     // Cases 3 & 4
946     return greaterWithBias(
947         (P + V), (Qout + std::max(Qin, F) * VProb + std::min(Qin, F) * UProb),
948         EntryFreq);
949   // Cases 1 & 2
950   return greaterWithBias((P + U),
951                          (Qout + std::min(Qin, F) * AdjustedSuccSumProb +
952                           std::max(Qin, F) * UProb),
953                          EntryFreq);
954 }
955 
956 /// Check for a trellis layout. \p BB is the upper part of a trellis if its
957 /// successors form the lower part of a trellis. A successor set S forms the
958 /// lower part of a trellis if all of the predecessors of S are either in S or
959 /// have all of S as successors. We ignore trellises where BB doesn't have 2
960 /// successors because for fewer than 2, it's trivial, and for 3 or greater they
961 /// are very uncommon and complex to compute optimally. Allowing edges within S
962 /// is not strictly a trellis, but the same algorithm works, so we allow it.
963 bool MachineBlockPlacement::isTrellis(
964     const MachineBasicBlock *BB,
965     const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
966     const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
967   // Technically BB could form a trellis with branching factor higher than 2.
968   // But that's extremely uncommon.
969   if (BB->succ_size() != 2 || ViableSuccs.size() != 2)
970     return false;
971 
972   SmallPtrSet<const MachineBasicBlock *, 2> Successors(BB->succ_begin(),
973                                                        BB->succ_end());
974   // To avoid reviewing the same predecessors twice.
975   SmallPtrSet<const MachineBasicBlock *, 8> SeenPreds;
976 
977   for (MachineBasicBlock *Succ : ViableSuccs) {
978     int PredCount = 0;
979     for (auto *SuccPred : Succ->predecessors()) {
980       // Allow triangle successors, but don't count them.
981       if (Successors.count(SuccPred)) {
982         // Make sure that it is actually a triangle.
983         for (MachineBasicBlock *CheckSucc : SuccPred->successors())
984           if (!Successors.count(CheckSucc))
985             return false;
986         continue;
987       }
988       const BlockChain *PredChain = BlockToChain[SuccPred];
989       if (SuccPred == BB || (BlockFilter && !BlockFilter->count(SuccPred)) ||
990           PredChain == &Chain || PredChain == BlockToChain[Succ])
991         continue;
992       ++PredCount;
993       // Perform the successor check only once.
994       if (!SeenPreds.insert(SuccPred).second)
995         continue;
996       if (!hasSameSuccessors(*SuccPred, Successors))
997         return false;
998     }
999     // If one of the successors has only BB as a predecessor, it is not a
1000     // trellis.
1001     if (PredCount < 1)
1002       return false;
1003   }
1004   return true;
1005 }
1006 
1007 /// Pick the highest total weight pair of edges that can both be laid out.
1008 /// The edges in \p Edges[0] are assumed to have a different destination than
1009 /// the edges in \p Edges[1]. Simple counting shows that the best pair is either
1010 /// the individual highest weight edges to the 2 different destinations, or in
1011 /// case of a conflict, one of them should be replaced with a 2nd best edge.
1012 std::pair<MachineBlockPlacement::WeightedEdge,
1013           MachineBlockPlacement::WeightedEdge>
1014 MachineBlockPlacement::getBestNonConflictingEdges(
1015     const MachineBasicBlock *BB,
1016     MutableArrayRef<SmallVector<MachineBlockPlacement::WeightedEdge, 8>>
1017         Edges) {
1018   // Sort the edges, and then for each successor, find the best incoming
1019   // predecessor. If the best incoming predecessors aren't the same,
1020   // then that is clearly the best layout. If there is a conflict, one of the
1021   // successors will have to fallthrough from the second best predecessor. We
1022   // compare which combination is better overall.
1023 
1024   // Sort for highest frequency.
1025   auto Cmp = [](WeightedEdge A, WeightedEdge B) { return A.Weight > B.Weight; };
1026 
1027   llvm::stable_sort(Edges[0], Cmp);
1028   llvm::stable_sort(Edges[1], Cmp);
1029   auto BestA = Edges[0].begin();
1030   auto BestB = Edges[1].begin();
1031   // Arrange for the correct answer to be in BestA and BestB
1032   // If the 2 best edges don't conflict, the answer is already there.
1033   if (BestA->Src == BestB->Src) {
1034     // Compare the total fallthrough of (Best + Second Best) for both pairs
1035     auto SecondBestA = std::next(BestA);
1036     auto SecondBestB = std::next(BestB);
1037     BlockFrequency BestAScore = BestA->Weight + SecondBestB->Weight;
1038     BlockFrequency BestBScore = BestB->Weight + SecondBestA->Weight;
1039     if (BestAScore < BestBScore)
1040       BestA = SecondBestA;
1041     else
1042       BestB = SecondBestB;
1043   }
1044   // Arrange for the BB edge to be in BestA if it exists.
1045   if (BestB->Src == BB)
1046     std::swap(BestA, BestB);
1047   return std::make_pair(*BestA, *BestB);
1048 }
1049 
1050 /// Get the best successor from \p BB based on \p BB being part of a trellis.
1051 /// We only handle trellises with 2 successors, so the algorithm is
1052 /// straightforward: Find the best pair of edges that don't conflict. We find
1053 /// the best incoming edge for each successor in the trellis. If those conflict,
1054 /// we consider which of them should be replaced with the second best.
1055 /// Upon return the two best edges will be in \p BestEdges. If one of the edges
1056 /// comes from \p BB, it will be in \p BestEdges[0]
1057 MachineBlockPlacement::BlockAndTailDupResult
1058 MachineBlockPlacement::getBestTrellisSuccessor(
1059     const MachineBasicBlock *BB,
1060     const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
1061     BranchProbability AdjustedSumProb, const BlockChain &Chain,
1062     const BlockFilterSet *BlockFilter) {
1063 
1064   BlockAndTailDupResult Result = {nullptr, false};
1065   SmallPtrSet<const MachineBasicBlock *, 4> Successors(BB->succ_begin(),
1066                                                        BB->succ_end());
1067 
1068   // We assume size 2 because it's common. For general n, we would have to do
1069   // the Hungarian algorithm, but it's not worth the complexity because more
1070   // than 2 successors is fairly uncommon, and a trellis even more so.
1071   if (Successors.size() != 2 || ViableSuccs.size() != 2)
1072     return Result;
1073 
1074   // Collect the edge frequencies of all edges that form the trellis.
1075   SmallVector<WeightedEdge, 8> Edges[2];
1076   int SuccIndex = 0;
1077   for (auto *Succ : ViableSuccs) {
1078     for (MachineBasicBlock *SuccPred : Succ->predecessors()) {
1079       // Skip any placed predecessors that are not BB
1080       if (SuccPred != BB)
1081         if ((BlockFilter && !BlockFilter->count(SuccPred)) ||
1082             BlockToChain[SuccPred] == &Chain ||
1083             BlockToChain[SuccPred] == BlockToChain[Succ])
1084           continue;
1085       BlockFrequency EdgeFreq = MBFI->getBlockFreq(SuccPred) *
1086                                 MBPI->getEdgeProbability(SuccPred, Succ);
1087       Edges[SuccIndex].push_back({EdgeFreq, SuccPred, Succ});
1088     }
1089     ++SuccIndex;
1090   }
1091 
1092   // Pick the best combination of 2 edges from all the edges in the trellis.
1093   WeightedEdge BestA, BestB;
1094   std::tie(BestA, BestB) = getBestNonConflictingEdges(BB, Edges);
1095 
1096   if (BestA.Src != BB) {
1097     // If we have a trellis, and BB doesn't have the best fallthrough edges,
1098     // we shouldn't choose any successor. We've already looked and there's a
1099     // better fallthrough edge for all the successors.
1100     LLVM_DEBUG(dbgs() << "Trellis, but not one of the chosen edges.\n");
1101     return Result;
1102   }
1103 
1104   // Did we pick the triangle edge? If tail-duplication is profitable, do
1105   // that instead. Otherwise merge the triangle edge now while we know it is
1106   // optimal.
1107   if (BestA.Dest == BestB.Src) {
1108     // The edges are BB->Succ1->Succ2, and we're looking to see if BB->Succ2
1109     // would be better.
1110     MachineBasicBlock *Succ1 = BestA.Dest;
1111     MachineBasicBlock *Succ2 = BestB.Dest;
1112     // Check to see if tail-duplication would be profitable.
1113     if (allowTailDupPlacement() && shouldTailDuplicate(Succ2) &&
1114         canTailDuplicateUnplacedPreds(BB, Succ2, Chain, BlockFilter) &&
1115         isProfitableToTailDup(BB, Succ2, MBPI->getEdgeProbability(BB, Succ1),
1116                               Chain, BlockFilter)) {
1117       LLVM_DEBUG(BranchProbability Succ2Prob = getAdjustedProbability(
1118                      MBPI->getEdgeProbability(BB, Succ2), AdjustedSumProb);
1119                  dbgs() << "    Selected: " << getBlockName(Succ2)
1120                         << ", probability: " << Succ2Prob
1121                         << " (Tail Duplicate)\n");
1122       Result.BB = Succ2;
1123       Result.ShouldTailDup = true;
1124       return Result;
1125     }
1126   }
1127   // We have already computed the optimal edge for the other side of the
1128   // trellis.
1129   ComputedEdges[BestB.Src] = { BestB.Dest, false };
1130 
1131   auto TrellisSucc = BestA.Dest;
1132   LLVM_DEBUG(BranchProbability SuccProb = getAdjustedProbability(
1133                  MBPI->getEdgeProbability(BB, TrellisSucc), AdjustedSumProb);
1134              dbgs() << "    Selected: " << getBlockName(TrellisSucc)
1135                     << ", probability: " << SuccProb << " (Trellis)\n");
1136   Result.BB = TrellisSucc;
1137   return Result;
1138 }
1139 
1140 /// When the option allowTailDupPlacement() is on, this method checks if the
1141 /// fallthrough candidate block \p Succ (of block \p BB) can be tail-duplicated
1142 /// into all of its unplaced, unfiltered predecessors, that are not BB.
1143 bool MachineBlockPlacement::canTailDuplicateUnplacedPreds(
1144     const MachineBasicBlock *BB, MachineBasicBlock *Succ,
1145     const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
1146   if (!shouldTailDuplicate(Succ))
1147     return false;
1148 
1149   // The result of canTailDuplicate.
1150   bool Duplicate = true;
1151   // Number of possible duplication.
1152   unsigned int NumDup = 0;
1153 
1154   // For CFG checking.
1155   SmallPtrSet<const MachineBasicBlock *, 4> Successors(BB->succ_begin(),
1156                                                        BB->succ_end());
1157   for (MachineBasicBlock *Pred : Succ->predecessors()) {
1158     // Make sure all unplaced and unfiltered predecessors can be
1159     // tail-duplicated into.
1160     // Skip any blocks that are already placed or not in this loop.
1161     if (Pred == BB || (BlockFilter && !BlockFilter->count(Pred))
1162         || (BlockToChain[Pred] == &Chain && !Succ->succ_empty()))
1163       continue;
1164     if (!TailDup.canTailDuplicate(Succ, Pred)) {
1165       if (Successors.size() > 1 && hasSameSuccessors(*Pred, Successors))
1166         // This will result in a trellis after tail duplication, so we don't
1167         // need to copy Succ into this predecessor. In the presence
1168         // of a trellis tail duplication can continue to be profitable.
1169         // For example:
1170         // A            A
1171         // |\           |\
1172         // | \          | \
1173         // |  C         |  C+BB
1174         // | /          |  |
1175         // |/           |  |
1176         // BB    =>     BB |
1177         // |\           |\/|
1178         // | \          |/\|
1179         // |  D         |  D
1180         // | /          | /
1181         // |/           |/
1182         // Succ         Succ
1183         //
1184         // After BB was duplicated into C, the layout looks like the one on the
1185         // right. BB and C now have the same successors. When considering
1186         // whether Succ can be duplicated into all its unplaced predecessors, we
1187         // ignore C.
1188         // We can do this because C already has a profitable fallthrough, namely
1189         // D. TODO(iteratee): ignore sufficiently cold predecessors for
1190         // duplication and for this test.
1191         //
1192         // This allows trellises to be laid out in 2 separate chains
1193         // (A,B,Succ,...) and later (C,D,...) This is a reasonable heuristic
1194         // because it allows the creation of 2 fallthrough paths with links
1195         // between them, and we correctly identify the best layout for these
1196         // CFGs. We want to extend trellises that the user created in addition
1197         // to trellises created by tail-duplication, so we just look for the
1198         // CFG.
1199         continue;
1200       Duplicate = false;
1201       continue;
1202     }
1203     NumDup++;
1204   }
1205 
1206   // No possible duplication in current filter set.
1207   if (NumDup == 0)
1208     return false;
1209 
1210   // If profile information is available, findDuplicateCandidates can do more
1211   // precise benefit analysis.
1212   if (F->getFunction().hasProfileData())
1213     return true;
1214 
1215   // This is mainly for function exit BB.
1216   // The integrated tail duplication is really designed for increasing
1217   // fallthrough from predecessors from Succ to its successors. We may need
1218   // other machanism to handle different cases.
1219   if (Succ->succ_empty())
1220     return true;
1221 
1222   // Plus the already placed predecessor.
1223   NumDup++;
1224 
1225   // If the duplication candidate has more unplaced predecessors than
1226   // successors, the extra duplication can't bring more fallthrough.
1227   //
1228   //     Pred1 Pred2 Pred3
1229   //         \   |   /
1230   //          \  |  /
1231   //           \ | /
1232   //            Dup
1233   //            / \
1234   //           /   \
1235   //       Succ1  Succ2
1236   //
1237   // In this example Dup has 2 successors and 3 predecessors, duplication of Dup
1238   // can increase the fallthrough from Pred1 to Succ1 and from Pred2 to Succ2,
1239   // but the duplication into Pred3 can't increase fallthrough.
1240   //
1241   // A small number of extra duplication may not hurt too much. We need a better
1242   // heuristic to handle it.
1243   if ((NumDup > Succ->succ_size()) || !Duplicate)
1244     return false;
1245 
1246   return true;
1247 }
1248 
1249 /// Find chains of triangles where we believe it would be profitable to
1250 /// tail-duplicate them all, but a local analysis would not find them.
1251 /// There are 3 ways this can be profitable:
1252 /// 1) The post-dominators marked 50% are actually taken 55% (This shrinks with
1253 ///    longer chains)
1254 /// 2) The chains are statically correlated. Branch probabilities have a very
1255 ///    U-shaped distribution.
1256 ///    [http://nrs.harvard.edu/urn-3:HUL.InstRepos:24015805]
1257 ///    If the branches in a chain are likely to be from the same side of the
1258 ///    distribution as their predecessor, but are independent at runtime, this
1259 ///    transformation is profitable. (Because the cost of being wrong is a small
1260 ///    fixed cost, unlike the standard triangle layout where the cost of being
1261 ///    wrong scales with the # of triangles.)
1262 /// 3) The chains are dynamically correlated. If the probability that a previous
1263 ///    branch was taken positively influences whether the next branch will be
1264 ///    taken
1265 /// We believe that 2 and 3 are common enough to justify the small margin in 1.
1266 void MachineBlockPlacement::precomputeTriangleChains() {
1267   struct TriangleChain {
1268     std::vector<MachineBasicBlock *> Edges;
1269 
1270     TriangleChain(MachineBasicBlock *src, MachineBasicBlock *dst)
1271         : Edges({src, dst}) {}
1272 
1273     void append(MachineBasicBlock *dst) {
1274       assert(getKey()->isSuccessor(dst) &&
1275              "Attempting to append a block that is not a successor.");
1276       Edges.push_back(dst);
1277     }
1278 
1279     unsigned count() const { return Edges.size() - 1; }
1280 
1281     MachineBasicBlock *getKey() const {
1282       return Edges.back();
1283     }
1284   };
1285 
1286   if (TriangleChainCount == 0)
1287     return;
1288 
1289   LLVM_DEBUG(dbgs() << "Pre-computing triangle chains.\n");
1290   // Map from last block to the chain that contains it. This allows us to extend
1291   // chains as we find new triangles.
1292   DenseMap<const MachineBasicBlock *, TriangleChain> TriangleChainMap;
1293   for (MachineBasicBlock &BB : *F) {
1294     // If BB doesn't have 2 successors, it doesn't start a triangle.
1295     if (BB.succ_size() != 2)
1296       continue;
1297     MachineBasicBlock *PDom = nullptr;
1298     for (MachineBasicBlock *Succ : BB.successors()) {
1299       if (!MPDT->dominates(Succ, &BB))
1300         continue;
1301       PDom = Succ;
1302       break;
1303     }
1304     // If BB doesn't have a post-dominating successor, it doesn't form a
1305     // triangle.
1306     if (PDom == nullptr)
1307       continue;
1308     // If PDom has a hint that it is low probability, skip this triangle.
1309     if (MBPI->getEdgeProbability(&BB, PDom) < BranchProbability(50, 100))
1310       continue;
1311     // If PDom isn't eligible for duplication, this isn't the kind of triangle
1312     // we're looking for.
1313     if (!shouldTailDuplicate(PDom))
1314       continue;
1315     bool CanTailDuplicate = true;
1316     // If PDom can't tail-duplicate into it's non-BB predecessors, then this
1317     // isn't the kind of triangle we're looking for.
1318     for (MachineBasicBlock* Pred : PDom->predecessors()) {
1319       if (Pred == &BB)
1320         continue;
1321       if (!TailDup.canTailDuplicate(PDom, Pred)) {
1322         CanTailDuplicate = false;
1323         break;
1324       }
1325     }
1326     // If we can't tail-duplicate PDom to its predecessors, then skip this
1327     // triangle.
1328     if (!CanTailDuplicate)
1329       continue;
1330 
1331     // Now we have an interesting triangle. Insert it if it's not part of an
1332     // existing chain.
1333     // Note: This cannot be replaced with a call insert() or emplace() because
1334     // the find key is BB, but the insert/emplace key is PDom.
1335     auto Found = TriangleChainMap.find(&BB);
1336     // If it is, remove the chain from the map, grow it, and put it back in the
1337     // map with the end as the new key.
1338     if (Found != TriangleChainMap.end()) {
1339       TriangleChain Chain = std::move(Found->second);
1340       TriangleChainMap.erase(Found);
1341       Chain.append(PDom);
1342       TriangleChainMap.insert(std::make_pair(Chain.getKey(), std::move(Chain)));
1343     } else {
1344       auto InsertResult = TriangleChainMap.try_emplace(PDom, &BB, PDom);
1345       assert(InsertResult.second && "Block seen twice.");
1346       (void)InsertResult;
1347     }
1348   }
1349 
1350   // Iterating over a DenseMap is safe here, because the only thing in the body
1351   // of the loop is inserting into another DenseMap (ComputedEdges).
1352   // ComputedEdges is never iterated, so this doesn't lead to non-determinism.
1353   for (auto &ChainPair : TriangleChainMap) {
1354     TriangleChain &Chain = ChainPair.second;
1355     // Benchmarking has shown that due to branch correlation duplicating 2 or
1356     // more triangles is profitable, despite the calculations assuming
1357     // independence.
1358     if (Chain.count() < TriangleChainCount)
1359       continue;
1360     MachineBasicBlock *dst = Chain.Edges.back();
1361     Chain.Edges.pop_back();
1362     for (MachineBasicBlock *src : reverse(Chain.Edges)) {
1363       LLVM_DEBUG(dbgs() << "Marking edge: " << getBlockName(src) << "->"
1364                         << getBlockName(dst)
1365                         << " as pre-computed based on triangles.\n");
1366 
1367       auto InsertResult = ComputedEdges.insert({src, {dst, true}});
1368       assert(InsertResult.second && "Block seen twice.");
1369       (void)InsertResult;
1370 
1371       dst = src;
1372     }
1373   }
1374 }
1375 
1376 // When profile is not present, return the StaticLikelyProb.
1377 // When profile is available, we need to handle the triangle-shape CFG.
1378 static BranchProbability getLayoutSuccessorProbThreshold(
1379       const MachineBasicBlock *BB) {
1380   if (!BB->getParent()->getFunction().hasProfileData())
1381     return BranchProbability(StaticLikelyProb, 100);
1382   if (BB->succ_size() == 2) {
1383     const MachineBasicBlock *Succ1 = *BB->succ_begin();
1384     const MachineBasicBlock *Succ2 = *(BB->succ_begin() + 1);
1385     if (Succ1->isSuccessor(Succ2) || Succ2->isSuccessor(Succ1)) {
1386       /* See case 1 below for the cost analysis. For BB->Succ to
1387        * be taken with smaller cost, the following needs to hold:
1388        *   Prob(BB->Succ) > 2 * Prob(BB->Pred)
1389        *   So the threshold T in the calculation below
1390        *   (1-T) * Prob(BB->Succ) > T * Prob(BB->Pred)
1391        *   So T / (1 - T) = 2, Yielding T = 2/3
1392        * Also adding user specified branch bias, we have
1393        *   T = (2/3)*(ProfileLikelyProb/50)
1394        *     = (2*ProfileLikelyProb)/150)
1395        */
1396       return BranchProbability(2 * ProfileLikelyProb, 150);
1397     }
1398   }
1399   return BranchProbability(ProfileLikelyProb, 100);
1400 }
1401 
1402 /// Checks to see if the layout candidate block \p Succ has a better layout
1403 /// predecessor than \c BB. If yes, returns true.
1404 /// \p SuccProb: The probability adjusted for only remaining blocks.
1405 ///   Only used for logging
1406 /// \p RealSuccProb: The un-adjusted probability.
1407 /// \p Chain: The chain that BB belongs to and Succ is being considered for.
1408 /// \p BlockFilter: if non-null, the set of blocks that make up the loop being
1409 ///    considered
1410 bool MachineBlockPlacement::hasBetterLayoutPredecessor(
1411     const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
1412     const BlockChain &SuccChain, BranchProbability SuccProb,
1413     BranchProbability RealSuccProb, const BlockChain &Chain,
1414     const BlockFilterSet *BlockFilter) {
1415 
1416   // There isn't a better layout when there are no unscheduled predecessors.
1417   if (SuccChain.UnscheduledPredecessors == 0)
1418     return false;
1419 
1420   // There are two basic scenarios here:
1421   // -------------------------------------
1422   // Case 1: triangular shape CFG (if-then):
1423   //     BB
1424   //     | \
1425   //     |  \
1426   //     |   Pred
1427   //     |   /
1428   //     Succ
1429   // In this case, we are evaluating whether to select edge -> Succ, e.g.
1430   // set Succ as the layout successor of BB. Picking Succ as BB's
1431   // successor breaks the CFG constraints (FIXME: define these constraints).
1432   // With this layout, Pred BB
1433   // is forced to be outlined, so the overall cost will be cost of the
1434   // branch taken from BB to Pred, plus the cost of back taken branch
1435   // from Pred to Succ, as well as the additional cost associated
1436   // with the needed unconditional jump instruction from Pred To Succ.
1437 
1438   // The cost of the topological order layout is the taken branch cost
1439   // from BB to Succ, so to make BB->Succ a viable candidate, the following
1440   // must hold:
1441   //     2 * freq(BB->Pred) * taken_branch_cost + unconditional_jump_cost
1442   //      < freq(BB->Succ) *  taken_branch_cost.
1443   // Ignoring unconditional jump cost, we get
1444   //    freq(BB->Succ) > 2 * freq(BB->Pred), i.e.,
1445   //    prob(BB->Succ) > 2 * prob(BB->Pred)
1446   //
1447   // When real profile data is available, we can precisely compute the
1448   // probability threshold that is needed for edge BB->Succ to be considered.
1449   // Without profile data, the heuristic requires the branch bias to be
1450   // a lot larger to make sure the signal is very strong (e.g. 80% default).
1451   // -----------------------------------------------------------------
1452   // Case 2: diamond like CFG (if-then-else):
1453   //     S
1454   //    / \
1455   //   |   \
1456   //  BB    Pred
1457   //   \    /
1458   //    Succ
1459   //    ..
1460   //
1461   // The current block is BB and edge BB->Succ is now being evaluated.
1462   // Note that edge S->BB was previously already selected because
1463   // prob(S->BB) > prob(S->Pred).
1464   // At this point, 2 blocks can be placed after BB: Pred or Succ. If we
1465   // choose Pred, we will have a topological ordering as shown on the left
1466   // in the picture below. If we choose Succ, we have the solution as shown
1467   // on the right:
1468   //
1469   //   topo-order:
1470   //
1471   //       S-----                             ---S
1472   //       |    |                             |  |
1473   //    ---BB   |                             |  BB
1474   //    |       |                             |  |
1475   //    |  Pred--                             |  Succ--
1476   //    |  |                                  |       |
1477   //    ---Succ                               ---Pred--
1478   //
1479   // cost = freq(S->Pred) + freq(BB->Succ)    cost = 2 * freq (S->Pred)
1480   //      = freq(S->Pred) + freq(S->BB)
1481   //
1482   // If we have profile data (i.e, branch probabilities can be trusted), the
1483   // cost (number of taken branches) with layout S->BB->Succ->Pred is 2 *
1484   // freq(S->Pred) while the cost of topo order is freq(S->Pred) + freq(S->BB).
1485   // We know Prob(S->BB) > Prob(S->Pred), so freq(S->BB) > freq(S->Pred), which
1486   // means the cost of topological order is greater.
1487   // When profile data is not available, however, we need to be more
1488   // conservative. If the branch prediction is wrong, breaking the topo-order
1489   // will actually yield a layout with large cost. For this reason, we need
1490   // strong biased branch at block S with Prob(S->BB) in order to select
1491   // BB->Succ. This is equivalent to looking the CFG backward with backward
1492   // edge: Prob(Succ->BB) needs to >= HotProb in order to be selected (without
1493   // profile data).
1494   // --------------------------------------------------------------------------
1495   // Case 3: forked diamond
1496   //       S
1497   //      / \
1498   //     /   \
1499   //   BB    Pred
1500   //   | \   / |
1501   //   |  \ /  |
1502   //   |   X   |
1503   //   |  / \  |
1504   //   | /   \ |
1505   //   S1     S2
1506   //
1507   // The current block is BB and edge BB->S1 is now being evaluated.
1508   // As above S->BB was already selected because
1509   // prob(S->BB) > prob(S->Pred). Assume that prob(BB->S1) >= prob(BB->S2).
1510   //
1511   // topo-order:
1512   //
1513   //     S-------|                     ---S
1514   //     |       |                     |  |
1515   //  ---BB      |                     |  BB
1516   //  |          |                     |  |
1517   //  |  Pred----|                     |  S1----
1518   //  |  |                             |       |
1519   //  --(S1 or S2)                     ---Pred--
1520   //                                        |
1521   //                                       S2
1522   //
1523   // topo-cost = freq(S->Pred) + freq(BB->S1) + freq(BB->S2)
1524   //    + min(freq(Pred->S1), freq(Pred->S2))
1525   // Non-topo-order cost:
1526   // non-topo-cost = 2 * freq(S->Pred) + freq(BB->S2).
1527   // To be conservative, we can assume that min(freq(Pred->S1), freq(Pred->S2))
1528   // is 0. Then the non topo layout is better when
1529   // freq(S->Pred) < freq(BB->S1).
1530   // This is exactly what is checked below.
1531   // Note there are other shapes that apply (Pred may not be a single block,
1532   // but they all fit this general pattern.)
1533   BranchProbability HotProb = getLayoutSuccessorProbThreshold(BB);
1534 
1535   // Make sure that a hot successor doesn't have a globally more
1536   // important predecessor.
1537   BlockFrequency CandidateEdgeFreq = MBFI->getBlockFreq(BB) * RealSuccProb;
1538   bool BadCFGConflict = false;
1539 
1540   for (MachineBasicBlock *Pred : Succ->predecessors()) {
1541     BlockChain *PredChain = BlockToChain[Pred];
1542     if (Pred == Succ || PredChain == &SuccChain ||
1543         (BlockFilter && !BlockFilter->count(Pred)) ||
1544         PredChain == &Chain || Pred != *std::prev(PredChain->end()) ||
1545         // This check is redundant except for look ahead. This function is
1546         // called for lookahead by isProfitableToTailDup when BB hasn't been
1547         // placed yet.
1548         (Pred == BB))
1549       continue;
1550     // Do backward checking.
1551     // For all cases above, we need a backward checking to filter out edges that
1552     // are not 'strongly' biased.
1553     // BB  Pred
1554     //  \ /
1555     //  Succ
1556     // We select edge BB->Succ if
1557     //      freq(BB->Succ) > freq(Succ) * HotProb
1558     //      i.e. freq(BB->Succ) > freq(BB->Succ) * HotProb + freq(Pred->Succ) *
1559     //      HotProb
1560     //      i.e. freq((BB->Succ) * (1 - HotProb) > freq(Pred->Succ) * HotProb
1561     // Case 1 is covered too, because the first equation reduces to:
1562     // prob(BB->Succ) > HotProb. (freq(Succ) = freq(BB) for a triangle)
1563     BlockFrequency PredEdgeFreq =
1564         MBFI->getBlockFreq(Pred) * MBPI->getEdgeProbability(Pred, Succ);
1565     if (PredEdgeFreq * HotProb >= CandidateEdgeFreq * HotProb.getCompl()) {
1566       BadCFGConflict = true;
1567       break;
1568     }
1569   }
1570 
1571   if (BadCFGConflict) {
1572     LLVM_DEBUG(dbgs() << "    Not a candidate: " << getBlockName(Succ) << " -> "
1573                       << SuccProb << " (prob) (non-cold CFG conflict)\n");
1574     return true;
1575   }
1576 
1577   return false;
1578 }
1579 
1580 /// Select the best successor for a block.
1581 ///
1582 /// This looks across all successors of a particular block and attempts to
1583 /// select the "best" one to be the layout successor. It only considers direct
1584 /// successors which also pass the block filter. It will attempt to avoid
1585 /// breaking CFG structure, but cave and break such structures in the case of
1586 /// very hot successor edges.
1587 ///
1588 /// \returns The best successor block found, or null if none are viable, along
1589 /// with a boolean indicating if tail duplication is necessary.
1590 MachineBlockPlacement::BlockAndTailDupResult
1591 MachineBlockPlacement::selectBestSuccessor(
1592     const MachineBasicBlock *BB, const BlockChain &Chain,
1593     const BlockFilterSet *BlockFilter) {
1594   const BranchProbability HotProb(StaticLikelyProb, 100);
1595 
1596   BlockAndTailDupResult BestSucc = { nullptr, false };
1597   auto BestProb = BranchProbability::getZero();
1598 
1599   SmallVector<MachineBasicBlock *, 4> Successors;
1600   auto AdjustedSumProb =
1601       collectViableSuccessors(BB, Chain, BlockFilter, Successors);
1602 
1603   LLVM_DEBUG(dbgs() << "Selecting best successor for: " << getBlockName(BB)
1604                     << "\n");
1605 
1606   // if we already precomputed the best successor for BB, return that if still
1607   // applicable.
1608   auto FoundEdge = ComputedEdges.find(BB);
1609   if (FoundEdge != ComputedEdges.end()) {
1610     MachineBasicBlock *Succ = FoundEdge->second.BB;
1611     ComputedEdges.erase(FoundEdge);
1612     BlockChain *SuccChain = BlockToChain[Succ];
1613     if (BB->isSuccessor(Succ) && (!BlockFilter || BlockFilter->count(Succ)) &&
1614         SuccChain != &Chain && Succ == *SuccChain->begin())
1615       return FoundEdge->second;
1616   }
1617 
1618   // if BB is part of a trellis, Use the trellis to determine the optimal
1619   // fallthrough edges
1620   if (isTrellis(BB, Successors, Chain, BlockFilter))
1621     return getBestTrellisSuccessor(BB, Successors, AdjustedSumProb, Chain,
1622                                    BlockFilter);
1623 
1624   // For blocks with CFG violations, we may be able to lay them out anyway with
1625   // tail-duplication. We keep this vector so we can perform the probability
1626   // calculations the minimum number of times.
1627   SmallVector<std::pair<BranchProbability, MachineBasicBlock *>, 4>
1628       DupCandidates;
1629   for (MachineBasicBlock *Succ : Successors) {
1630     auto RealSuccProb = MBPI->getEdgeProbability(BB, Succ);
1631     BranchProbability SuccProb =
1632         getAdjustedProbability(RealSuccProb, AdjustedSumProb);
1633 
1634     BlockChain &SuccChain = *BlockToChain[Succ];
1635     // Skip the edge \c BB->Succ if block \c Succ has a better layout
1636     // predecessor that yields lower global cost.
1637     if (hasBetterLayoutPredecessor(BB, Succ, SuccChain, SuccProb, RealSuccProb,
1638                                    Chain, BlockFilter)) {
1639       // If tail duplication would make Succ profitable, place it.
1640       if (allowTailDupPlacement() && shouldTailDuplicate(Succ))
1641         DupCandidates.emplace_back(SuccProb, Succ);
1642       continue;
1643     }
1644 
1645     LLVM_DEBUG(
1646         dbgs() << "    Candidate: " << getBlockName(Succ)
1647                << ", probability: " << SuccProb
1648                << (SuccChain.UnscheduledPredecessors != 0 ? " (CFG break)" : "")
1649                << "\n");
1650 
1651     if (BestSucc.BB && BestProb >= SuccProb) {
1652       LLVM_DEBUG(dbgs() << "    Not the best candidate, continuing\n");
1653       continue;
1654     }
1655 
1656     LLVM_DEBUG(dbgs() << "    Setting it as best candidate\n");
1657     BestSucc.BB = Succ;
1658     BestProb = SuccProb;
1659   }
1660   // Handle the tail duplication candidates in order of decreasing probability.
1661   // Stop at the first one that is profitable. Also stop if they are less
1662   // profitable than BestSucc. Position is important because we preserve it and
1663   // prefer first best match. Here we aren't comparing in order, so we capture
1664   // the position instead.
1665   llvm::stable_sort(DupCandidates,
1666                     [](std::tuple<BranchProbability, MachineBasicBlock *> L,
1667                        std::tuple<BranchProbability, MachineBasicBlock *> R) {
1668                       return std::get<0>(L) > std::get<0>(R);
1669                     });
1670   for (auto &Tup : DupCandidates) {
1671     BranchProbability DupProb;
1672     MachineBasicBlock *Succ;
1673     std::tie(DupProb, Succ) = Tup;
1674     if (DupProb < BestProb)
1675       break;
1676     if (canTailDuplicateUnplacedPreds(BB, Succ, Chain, BlockFilter)
1677         && (isProfitableToTailDup(BB, Succ, BestProb, Chain, BlockFilter))) {
1678       LLVM_DEBUG(dbgs() << "    Candidate: " << getBlockName(Succ)
1679                         << ", probability: " << DupProb
1680                         << " (Tail Duplicate)\n");
1681       BestSucc.BB = Succ;
1682       BestSucc.ShouldTailDup = true;
1683       break;
1684     }
1685   }
1686 
1687   if (BestSucc.BB)
1688     LLVM_DEBUG(dbgs() << "    Selected: " << getBlockName(BestSucc.BB) << "\n");
1689 
1690   return BestSucc;
1691 }
1692 
1693 /// Select the best block from a worklist.
1694 ///
1695 /// This looks through the provided worklist as a list of candidate basic
1696 /// blocks and select the most profitable one to place. The definition of
1697 /// profitable only really makes sense in the context of a loop. This returns
1698 /// the most frequently visited block in the worklist, which in the case of
1699 /// a loop, is the one most desirable to be physically close to the rest of the
1700 /// loop body in order to improve i-cache behavior.
1701 ///
1702 /// \returns The best block found, or null if none are viable.
1703 MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock(
1704     const BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList) {
1705   // Once we need to walk the worklist looking for a candidate, cleanup the
1706   // worklist of already placed entries.
1707   // FIXME: If this shows up on profiles, it could be folded (at the cost of
1708   // some code complexity) into the loop below.
1709   llvm::erase_if(WorkList, [&](MachineBasicBlock *BB) {
1710     return BlockToChain.lookup(BB) == &Chain;
1711   });
1712 
1713   if (WorkList.empty())
1714     return nullptr;
1715 
1716   bool IsEHPad = WorkList[0]->isEHPad();
1717 
1718   MachineBasicBlock *BestBlock = nullptr;
1719   BlockFrequency BestFreq;
1720   for (MachineBasicBlock *MBB : WorkList) {
1721     assert(MBB->isEHPad() == IsEHPad &&
1722            "EHPad mismatch between block and work list.");
1723 
1724     BlockChain &SuccChain = *BlockToChain[MBB];
1725     if (&SuccChain == &Chain)
1726       continue;
1727 
1728     assert(SuccChain.UnscheduledPredecessors == 0 &&
1729            "Found CFG-violating block");
1730 
1731     BlockFrequency CandidateFreq = MBFI->getBlockFreq(MBB);
1732     LLVM_DEBUG(dbgs() << "    " << getBlockName(MBB) << " -> "
1733                       << printBlockFreq(MBFI->getMBFI(), CandidateFreq)
1734                       << " (freq)\n");
1735 
1736     // For ehpad, we layout the least probable first as to avoid jumping back
1737     // from least probable landingpads to more probable ones.
1738     //
1739     // FIXME: Using probability is probably (!) not the best way to achieve
1740     // this. We should probably have a more principled approach to layout
1741     // cleanup code.
1742     //
1743     // The goal is to get:
1744     //
1745     //                 +--------------------------+
1746     //                 |                          V
1747     // InnerLp -> InnerCleanup    OuterLp -> OuterCleanup -> Resume
1748     //
1749     // Rather than:
1750     //
1751     //                 +-------------------------------------+
1752     //                 V                                     |
1753     // OuterLp -> OuterCleanup -> Resume     InnerLp -> InnerCleanup
1754     if (BestBlock && (IsEHPad ^ (BestFreq >= CandidateFreq)))
1755       continue;
1756 
1757     BestBlock = MBB;
1758     BestFreq = CandidateFreq;
1759   }
1760 
1761   return BestBlock;
1762 }
1763 
1764 /// Retrieve the first unplaced basic block.
1765 ///
1766 /// This routine is called when we are unable to use the CFG to walk through
1767 /// all of the basic blocks and form a chain due to unnatural loops in the CFG.
1768 /// We walk through the function's blocks in order, starting from the
1769 /// LastUnplacedBlockIt. We update this iterator on each call to avoid
1770 /// re-scanning the entire sequence on repeated calls to this routine.
1771 MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock(
1772     const BlockChain &PlacedChain,
1773     MachineFunction::iterator &PrevUnplacedBlockIt,
1774     const BlockFilterSet *BlockFilter) {
1775   for (MachineFunction::iterator I = PrevUnplacedBlockIt, E = F->end(); I != E;
1776        ++I) {
1777     if (BlockFilter && !BlockFilter->count(&*I))
1778       continue;
1779     if (BlockToChain[&*I] != &PlacedChain) {
1780       PrevUnplacedBlockIt = I;
1781       // Now select the head of the chain to which the unplaced block belongs
1782       // as the block to place. This will force the entire chain to be placed,
1783       // and satisfies the requirements of merging chains.
1784       return *BlockToChain[&*I]->begin();
1785     }
1786   }
1787   return nullptr;
1788 }
1789 
1790 void MachineBlockPlacement::fillWorkLists(
1791     const MachineBasicBlock *MBB,
1792     SmallPtrSetImpl<BlockChain *> &UpdatedPreds,
1793     const BlockFilterSet *BlockFilter = nullptr) {
1794   BlockChain &Chain = *BlockToChain[MBB];
1795   if (!UpdatedPreds.insert(&Chain).second)
1796     return;
1797 
1798   assert(
1799       Chain.UnscheduledPredecessors == 0 &&
1800       "Attempting to place block with unscheduled predecessors in worklist.");
1801   for (MachineBasicBlock *ChainBB : Chain) {
1802     assert(BlockToChain[ChainBB] == &Chain &&
1803            "Block in chain doesn't match BlockToChain map.");
1804     for (MachineBasicBlock *Pred : ChainBB->predecessors()) {
1805       if (BlockFilter && !BlockFilter->count(Pred))
1806         continue;
1807       if (BlockToChain[Pred] == &Chain)
1808         continue;
1809       ++Chain.UnscheduledPredecessors;
1810     }
1811   }
1812 
1813   if (Chain.UnscheduledPredecessors != 0)
1814     return;
1815 
1816   MachineBasicBlock *BB = *Chain.begin();
1817   if (BB->isEHPad())
1818     EHPadWorkList.push_back(BB);
1819   else
1820     BlockWorkList.push_back(BB);
1821 }
1822 
1823 void MachineBlockPlacement::buildChain(
1824     const MachineBasicBlock *HeadBB, BlockChain &Chain,
1825     BlockFilterSet *BlockFilter) {
1826   assert(HeadBB && "BB must not be null.\n");
1827   assert(BlockToChain[HeadBB] == &Chain && "BlockToChainMap mis-match.\n");
1828   MachineFunction::iterator PrevUnplacedBlockIt = F->begin();
1829 
1830   const MachineBasicBlock *LoopHeaderBB = HeadBB;
1831   markChainSuccessors(Chain, LoopHeaderBB, BlockFilter);
1832   MachineBasicBlock *BB = *std::prev(Chain.end());
1833   while (true) {
1834     assert(BB && "null block found at end of chain in loop.");
1835     assert(BlockToChain[BB] == &Chain && "BlockToChainMap mis-match in loop.");
1836     assert(*std::prev(Chain.end()) == BB && "BB Not found at end of chain.");
1837 
1838 
1839     // Look for the best viable successor if there is one to place immediately
1840     // after this block.
1841     auto Result = selectBestSuccessor(BB, Chain, BlockFilter);
1842     MachineBasicBlock* BestSucc = Result.BB;
1843     bool ShouldTailDup = Result.ShouldTailDup;
1844     if (allowTailDupPlacement())
1845       ShouldTailDup |= (BestSucc && canTailDuplicateUnplacedPreds(BB, BestSucc,
1846                                                                   Chain,
1847                                                                   BlockFilter));
1848 
1849     // If an immediate successor isn't available, look for the best viable
1850     // block among those we've identified as not violating the loop's CFG at
1851     // this point. This won't be a fallthrough, but it will increase locality.
1852     if (!BestSucc)
1853       BestSucc = selectBestCandidateBlock(Chain, BlockWorkList);
1854     if (!BestSucc)
1855       BestSucc = selectBestCandidateBlock(Chain, EHPadWorkList);
1856 
1857     if (!BestSucc) {
1858       BestSucc = getFirstUnplacedBlock(Chain, PrevUnplacedBlockIt, BlockFilter);
1859       if (!BestSucc)
1860         break;
1861 
1862       LLVM_DEBUG(dbgs() << "Unnatural loop CFG detected, forcibly merging the "
1863                            "layout successor until the CFG reduces\n");
1864     }
1865 
1866     // Placement may have changed tail duplication opportunities.
1867     // Check for that now.
1868     if (allowTailDupPlacement() && BestSucc && ShouldTailDup) {
1869       repeatedlyTailDuplicateBlock(BestSucc, BB, LoopHeaderBB, Chain,
1870                                        BlockFilter, PrevUnplacedBlockIt);
1871       // If the chosen successor was duplicated into BB, don't bother laying
1872       // it out, just go round the loop again with BB as the chain end.
1873       if (!BB->isSuccessor(BestSucc))
1874         continue;
1875     }
1876 
1877     // Place this block, updating the datastructures to reflect its placement.
1878     BlockChain &SuccChain = *BlockToChain[BestSucc];
1879     // Zero out UnscheduledPredecessors for the successor we're about to merge in case
1880     // we selected a successor that didn't fit naturally into the CFG.
1881     SuccChain.UnscheduledPredecessors = 0;
1882     LLVM_DEBUG(dbgs() << "Merging from " << getBlockName(BB) << " to "
1883                       << getBlockName(BestSucc) << "\n");
1884     markChainSuccessors(SuccChain, LoopHeaderBB, BlockFilter);
1885     Chain.merge(BestSucc, &SuccChain);
1886     BB = *std::prev(Chain.end());
1887   }
1888 
1889   LLVM_DEBUG(dbgs() << "Finished forming chain for header block "
1890                     << getBlockName(*Chain.begin()) << "\n");
1891 }
1892 
1893 // If bottom of block BB has only one successor OldTop, in most cases it is
1894 // profitable to move it before OldTop, except the following case:
1895 //
1896 //     -->OldTop<-
1897 //     |    .    |
1898 //     |    .    |
1899 //     |    .    |
1900 //     ---Pred   |
1901 //          |    |
1902 //         BB-----
1903 //
1904 // If BB is moved before OldTop, Pred needs a taken branch to BB, and it can't
1905 // layout the other successor below it, so it can't reduce taken branch.
1906 // In this case we keep its original layout.
1907 bool
1908 MachineBlockPlacement::canMoveBottomBlockToTop(
1909     const MachineBasicBlock *BottomBlock,
1910     const MachineBasicBlock *OldTop) {
1911   if (BottomBlock->pred_size() != 1)
1912     return true;
1913   MachineBasicBlock *Pred = *BottomBlock->pred_begin();
1914   if (Pred->succ_size() != 2)
1915     return true;
1916 
1917   MachineBasicBlock *OtherBB = *Pred->succ_begin();
1918   if (OtherBB == BottomBlock)
1919     OtherBB = *Pred->succ_rbegin();
1920   if (OtherBB == OldTop)
1921     return false;
1922 
1923   return true;
1924 }
1925 
1926 // Find out the possible fall through frequence to the top of a loop.
1927 BlockFrequency
1928 MachineBlockPlacement::TopFallThroughFreq(
1929     const MachineBasicBlock *Top,
1930     const BlockFilterSet &LoopBlockSet) {
1931   BlockFrequency MaxFreq = BlockFrequency(0);
1932   for (MachineBasicBlock *Pred : Top->predecessors()) {
1933     BlockChain *PredChain = BlockToChain[Pred];
1934     if (!LoopBlockSet.count(Pred) &&
1935         (!PredChain || Pred == *std::prev(PredChain->end()))) {
1936       // Found a Pred block can be placed before Top.
1937       // Check if Top is the best successor of Pred.
1938       auto TopProb = MBPI->getEdgeProbability(Pred, Top);
1939       bool TopOK = true;
1940       for (MachineBasicBlock *Succ : Pred->successors()) {
1941         auto SuccProb = MBPI->getEdgeProbability(Pred, Succ);
1942         BlockChain *SuccChain = BlockToChain[Succ];
1943         // Check if Succ can be placed after Pred.
1944         // Succ should not be in any chain, or it is the head of some chain.
1945         if (!LoopBlockSet.count(Succ) && (SuccProb > TopProb) &&
1946             (!SuccChain || Succ == *SuccChain->begin())) {
1947           TopOK = false;
1948           break;
1949         }
1950       }
1951       if (TopOK) {
1952         BlockFrequency EdgeFreq = MBFI->getBlockFreq(Pred) *
1953                                   MBPI->getEdgeProbability(Pred, Top);
1954         if (EdgeFreq > MaxFreq)
1955           MaxFreq = EdgeFreq;
1956       }
1957     }
1958   }
1959   return MaxFreq;
1960 }
1961 
1962 // Compute the fall through gains when move NewTop before OldTop.
1963 //
1964 // In following diagram, edges marked as "-" are reduced fallthrough, edges
1965 // marked as "+" are increased fallthrough, this function computes
1966 //
1967 //      SUM(increased fallthrough) - SUM(decreased fallthrough)
1968 //
1969 //              |
1970 //              | -
1971 //              V
1972 //        --->OldTop
1973 //        |     .
1974 //        |     .
1975 //       +|     .    +
1976 //        |   Pred --->
1977 //        |     |-
1978 //        |     V
1979 //        --- NewTop <---
1980 //              |-
1981 //              V
1982 //
1983 BlockFrequency
1984 MachineBlockPlacement::FallThroughGains(
1985     const MachineBasicBlock *NewTop,
1986     const MachineBasicBlock *OldTop,
1987     const MachineBasicBlock *ExitBB,
1988     const BlockFilterSet &LoopBlockSet) {
1989   BlockFrequency FallThrough2Top = TopFallThroughFreq(OldTop, LoopBlockSet);
1990   BlockFrequency FallThrough2Exit = BlockFrequency(0);
1991   if (ExitBB)
1992     FallThrough2Exit = MBFI->getBlockFreq(NewTop) *
1993         MBPI->getEdgeProbability(NewTop, ExitBB);
1994   BlockFrequency BackEdgeFreq = MBFI->getBlockFreq(NewTop) *
1995       MBPI->getEdgeProbability(NewTop, OldTop);
1996 
1997   // Find the best Pred of NewTop.
1998   MachineBasicBlock *BestPred = nullptr;
1999   BlockFrequency FallThroughFromPred = BlockFrequency(0);
2000   for (MachineBasicBlock *Pred : NewTop->predecessors()) {
2001     if (!LoopBlockSet.count(Pred))
2002       continue;
2003     BlockChain *PredChain = BlockToChain[Pred];
2004     if (!PredChain || Pred == *std::prev(PredChain->end())) {
2005       BlockFrequency EdgeFreq =
2006           MBFI->getBlockFreq(Pred) * MBPI->getEdgeProbability(Pred, NewTop);
2007       if (EdgeFreq > FallThroughFromPred) {
2008         FallThroughFromPred = EdgeFreq;
2009         BestPred = Pred;
2010       }
2011     }
2012   }
2013 
2014   // If NewTop is not placed after Pred, another successor can be placed
2015   // after Pred.
2016   BlockFrequency NewFreq = BlockFrequency(0);
2017   if (BestPred) {
2018     for (MachineBasicBlock *Succ : BestPred->successors()) {
2019       if ((Succ == NewTop) || (Succ == BestPred) || !LoopBlockSet.count(Succ))
2020         continue;
2021       if (ComputedEdges.contains(Succ))
2022         continue;
2023       BlockChain *SuccChain = BlockToChain[Succ];
2024       if ((SuccChain && (Succ != *SuccChain->begin())) ||
2025           (SuccChain == BlockToChain[BestPred]))
2026         continue;
2027       BlockFrequency EdgeFreq = MBFI->getBlockFreq(BestPred) *
2028                                 MBPI->getEdgeProbability(BestPred, Succ);
2029       if (EdgeFreq > NewFreq)
2030         NewFreq = EdgeFreq;
2031     }
2032     BlockFrequency OrigEdgeFreq = MBFI->getBlockFreq(BestPred) *
2033                                   MBPI->getEdgeProbability(BestPred, NewTop);
2034     if (NewFreq > OrigEdgeFreq) {
2035       // If NewTop is not the best successor of Pred, then Pred doesn't
2036       // fallthrough to NewTop. So there is no FallThroughFromPred and
2037       // NewFreq.
2038       NewFreq = BlockFrequency(0);
2039       FallThroughFromPred = BlockFrequency(0);
2040     }
2041   }
2042 
2043   BlockFrequency Result = BlockFrequency(0);
2044   BlockFrequency Gains = BackEdgeFreq + NewFreq;
2045   BlockFrequency Lost =
2046       FallThrough2Top + FallThrough2Exit + FallThroughFromPred;
2047   if (Gains > Lost)
2048     Result = Gains - Lost;
2049   return Result;
2050 }
2051 
2052 /// Helper function of findBestLoopTop. Find the best loop top block
2053 /// from predecessors of old top.
2054 ///
2055 /// Look for a block which is strictly better than the old top for laying
2056 /// out before the old top of the loop. This looks for only two patterns:
2057 ///
2058 ///     1. a block has only one successor, the old loop top
2059 ///
2060 ///        Because such a block will always result in an unconditional jump,
2061 ///        rotating it in front of the old top is always profitable.
2062 ///
2063 ///     2. a block has two successors, one is old top, another is exit
2064 ///        and it has more than one predecessors
2065 ///
2066 ///        If it is below one of its predecessors P, only P can fall through to
2067 ///        it, all other predecessors need a jump to it, and another conditional
2068 ///        jump to loop header. If it is moved before loop header, all its
2069 ///        predecessors jump to it, then fall through to loop header. So all its
2070 ///        predecessors except P can reduce one taken branch.
2071 ///        At the same time, move it before old top increases the taken branch
2072 ///        to loop exit block, so the reduced taken branch will be compared with
2073 ///        the increased taken branch to the loop exit block.
2074 MachineBasicBlock *
2075 MachineBlockPlacement::findBestLoopTopHelper(
2076     MachineBasicBlock *OldTop,
2077     const MachineLoop &L,
2078     const BlockFilterSet &LoopBlockSet) {
2079   // Check that the header hasn't been fused with a preheader block due to
2080   // crazy branches. If it has, we need to start with the header at the top to
2081   // prevent pulling the preheader into the loop body.
2082   BlockChain &HeaderChain = *BlockToChain[OldTop];
2083   if (!LoopBlockSet.count(*HeaderChain.begin()))
2084     return OldTop;
2085   if (OldTop != *HeaderChain.begin())
2086     return OldTop;
2087 
2088   LLVM_DEBUG(dbgs() << "Finding best loop top for: " << getBlockName(OldTop)
2089                     << "\n");
2090 
2091   BlockFrequency BestGains = BlockFrequency(0);
2092   MachineBasicBlock *BestPred = nullptr;
2093   for (MachineBasicBlock *Pred : OldTop->predecessors()) {
2094     if (!LoopBlockSet.count(Pred))
2095       continue;
2096     if (Pred == L.getHeader())
2097       continue;
2098     LLVM_DEBUG(dbgs() << "   old top pred: " << getBlockName(Pred) << ", has "
2099                       << Pred->succ_size() << " successors, "
2100                       << printBlockFreq(MBFI->getMBFI(), *Pred) << " freq\n");
2101     if (Pred->succ_size() > 2)
2102       continue;
2103 
2104     MachineBasicBlock *OtherBB = nullptr;
2105     if (Pred->succ_size() == 2) {
2106       OtherBB = *Pred->succ_begin();
2107       if (OtherBB == OldTop)
2108         OtherBB = *Pred->succ_rbegin();
2109     }
2110 
2111     if (!canMoveBottomBlockToTop(Pred, OldTop))
2112       continue;
2113 
2114     BlockFrequency Gains = FallThroughGains(Pred, OldTop, OtherBB,
2115                                             LoopBlockSet);
2116     if ((Gains > BlockFrequency(0)) &&
2117         (Gains > BestGains ||
2118          ((Gains == BestGains) && Pred->isLayoutSuccessor(OldTop)))) {
2119       BestPred = Pred;
2120       BestGains = Gains;
2121     }
2122   }
2123 
2124   // If no direct predecessor is fine, just use the loop header.
2125   if (!BestPred) {
2126     LLVM_DEBUG(dbgs() << "    final top unchanged\n");
2127     return OldTop;
2128   }
2129 
2130   // Walk backwards through any straight line of predecessors.
2131   while (BestPred->pred_size() == 1 &&
2132          (*BestPred->pred_begin())->succ_size() == 1 &&
2133          *BestPred->pred_begin() != L.getHeader())
2134     BestPred = *BestPred->pred_begin();
2135 
2136   LLVM_DEBUG(dbgs() << "    final top: " << getBlockName(BestPred) << "\n");
2137   return BestPred;
2138 }
2139 
2140 /// Find the best loop top block for layout.
2141 ///
2142 /// This function iteratively calls findBestLoopTopHelper, until no new better
2143 /// BB can be found.
2144 MachineBasicBlock *
2145 MachineBlockPlacement::findBestLoopTop(const MachineLoop &L,
2146                                        const BlockFilterSet &LoopBlockSet) {
2147   // Placing the latch block before the header may introduce an extra branch
2148   // that skips this block the first time the loop is executed, which we want
2149   // to avoid when optimising for size.
2150   // FIXME: in theory there is a case that does not introduce a new branch,
2151   // i.e. when the layout predecessor does not fallthrough to the loop header.
2152   // In practice this never happens though: there always seems to be a preheader
2153   // that can fallthrough and that is also placed before the header.
2154   bool OptForSize = F->getFunction().hasOptSize() ||
2155                     llvm::shouldOptimizeForSize(L.getHeader(), PSI, MBFI.get());
2156   if (OptForSize)
2157     return L.getHeader();
2158 
2159   MachineBasicBlock *OldTop = nullptr;
2160   MachineBasicBlock *NewTop = L.getHeader();
2161   while (NewTop != OldTop) {
2162     OldTop = NewTop;
2163     NewTop = findBestLoopTopHelper(OldTop, L, LoopBlockSet);
2164     if (NewTop != OldTop)
2165       ComputedEdges[NewTop] = { OldTop, false };
2166   }
2167   return NewTop;
2168 }
2169 
2170 /// Find the best loop exiting block for layout.
2171 ///
2172 /// This routine implements the logic to analyze the loop looking for the best
2173 /// block to layout at the top of the loop. Typically this is done to maximize
2174 /// fallthrough opportunities.
2175 MachineBasicBlock *
2176 MachineBlockPlacement::findBestLoopExit(const MachineLoop &L,
2177                                         const BlockFilterSet &LoopBlockSet,
2178                                         BlockFrequency &ExitFreq) {
2179   // We don't want to layout the loop linearly in all cases. If the loop header
2180   // is just a normal basic block in the loop, we want to look for what block
2181   // within the loop is the best one to layout at the top. However, if the loop
2182   // header has be pre-merged into a chain due to predecessors not having
2183   // analyzable branches, *and* the predecessor it is merged with is *not* part
2184   // of the loop, rotating the header into the middle of the loop will create
2185   // a non-contiguous range of blocks which is Very Bad. So start with the
2186   // header and only rotate if safe.
2187   BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
2188   if (!LoopBlockSet.count(*HeaderChain.begin()))
2189     return nullptr;
2190 
2191   BlockFrequency BestExitEdgeFreq;
2192   unsigned BestExitLoopDepth = 0;
2193   MachineBasicBlock *ExitingBB = nullptr;
2194   // If there are exits to outer loops, loop rotation can severely limit
2195   // fallthrough opportunities unless it selects such an exit. Keep a set of
2196   // blocks where rotating to exit with that block will reach an outer loop.
2197   SmallPtrSet<MachineBasicBlock *, 4> BlocksExitingToOuterLoop;
2198 
2199   LLVM_DEBUG(dbgs() << "Finding best loop exit for: "
2200                     << getBlockName(L.getHeader()) << "\n");
2201   for (MachineBasicBlock *MBB : L.getBlocks()) {
2202     BlockChain &Chain = *BlockToChain[MBB];
2203     // Ensure that this block is at the end of a chain; otherwise it could be
2204     // mid-way through an inner loop or a successor of an unanalyzable branch.
2205     if (MBB != *std::prev(Chain.end()))
2206       continue;
2207 
2208     // Now walk the successors. We need to establish whether this has a viable
2209     // exiting successor and whether it has a viable non-exiting successor.
2210     // We store the old exiting state and restore it if a viable looping
2211     // successor isn't found.
2212     MachineBasicBlock *OldExitingBB = ExitingBB;
2213     BlockFrequency OldBestExitEdgeFreq = BestExitEdgeFreq;
2214     bool HasLoopingSucc = false;
2215     for (MachineBasicBlock *Succ : MBB->successors()) {
2216       if (Succ->isEHPad())
2217         continue;
2218       if (Succ == MBB)
2219         continue;
2220       BlockChain &SuccChain = *BlockToChain[Succ];
2221       // Don't split chains, either this chain or the successor's chain.
2222       if (&Chain == &SuccChain) {
2223         LLVM_DEBUG(dbgs() << "    exiting: " << getBlockName(MBB) << " -> "
2224                           << getBlockName(Succ) << " (chain conflict)\n");
2225         continue;
2226       }
2227 
2228       auto SuccProb = MBPI->getEdgeProbability(MBB, Succ);
2229       if (LoopBlockSet.count(Succ)) {
2230         LLVM_DEBUG(dbgs() << "    looping: " << getBlockName(MBB) << " -> "
2231                           << getBlockName(Succ) << " (" << SuccProb << ")\n");
2232         HasLoopingSucc = true;
2233         continue;
2234       }
2235 
2236       unsigned SuccLoopDepth = 0;
2237       if (MachineLoop *ExitLoop = MLI->getLoopFor(Succ)) {
2238         SuccLoopDepth = ExitLoop->getLoopDepth();
2239         if (ExitLoop->contains(&L))
2240           BlocksExitingToOuterLoop.insert(MBB);
2241       }
2242 
2243       BlockFrequency ExitEdgeFreq = MBFI->getBlockFreq(MBB) * SuccProb;
2244       LLVM_DEBUG(
2245           dbgs() << "    exiting: " << getBlockName(MBB) << " -> "
2246                  << getBlockName(Succ) << " [L:" << SuccLoopDepth << "] ("
2247                  << printBlockFreq(MBFI->getMBFI(), ExitEdgeFreq) << ")\n");
2248       // Note that we bias this toward an existing layout successor to retain
2249       // incoming order in the absence of better information. The exit must have
2250       // a frequency higher than the current exit before we consider breaking
2251       // the layout.
2252       BranchProbability Bias(100 - ExitBlockBias, 100);
2253       if (!ExitingBB || SuccLoopDepth > BestExitLoopDepth ||
2254           ExitEdgeFreq > BestExitEdgeFreq ||
2255           (MBB->isLayoutSuccessor(Succ) &&
2256            !(ExitEdgeFreq < BestExitEdgeFreq * Bias))) {
2257         BestExitEdgeFreq = ExitEdgeFreq;
2258         ExitingBB = MBB;
2259       }
2260     }
2261 
2262     if (!HasLoopingSucc) {
2263       // Restore the old exiting state, no viable looping successor was found.
2264       ExitingBB = OldExitingBB;
2265       BestExitEdgeFreq = OldBestExitEdgeFreq;
2266     }
2267   }
2268   // Without a candidate exiting block or with only a single block in the
2269   // loop, just use the loop header to layout the loop.
2270   if (!ExitingBB) {
2271     LLVM_DEBUG(
2272         dbgs() << "    No other candidate exit blocks, using loop header\n");
2273     return nullptr;
2274   }
2275   if (L.getNumBlocks() == 1) {
2276     LLVM_DEBUG(dbgs() << "    Loop has 1 block, using loop header as exit\n");
2277     return nullptr;
2278   }
2279 
2280   // Also, if we have exit blocks which lead to outer loops but didn't select
2281   // one of them as the exiting block we are rotating toward, disable loop
2282   // rotation altogether.
2283   if (!BlocksExitingToOuterLoop.empty() &&
2284       !BlocksExitingToOuterLoop.count(ExitingBB))
2285     return nullptr;
2286 
2287   LLVM_DEBUG(dbgs() << "  Best exiting block: " << getBlockName(ExitingBB)
2288                     << "\n");
2289   ExitFreq = BestExitEdgeFreq;
2290   return ExitingBB;
2291 }
2292 
2293 /// Check if there is a fallthrough to loop header Top.
2294 ///
2295 ///   1. Look for a Pred that can be layout before Top.
2296 ///   2. Check if Top is the most possible successor of Pred.
2297 bool
2298 MachineBlockPlacement::hasViableTopFallthrough(
2299     const MachineBasicBlock *Top,
2300     const BlockFilterSet &LoopBlockSet) {
2301   for (MachineBasicBlock *Pred : Top->predecessors()) {
2302     BlockChain *PredChain = BlockToChain[Pred];
2303     if (!LoopBlockSet.count(Pred) &&
2304         (!PredChain || Pred == *std::prev(PredChain->end()))) {
2305       // Found a Pred block can be placed before Top.
2306       // Check if Top is the best successor of Pred.
2307       auto TopProb = MBPI->getEdgeProbability(Pred, Top);
2308       bool TopOK = true;
2309       for (MachineBasicBlock *Succ : Pred->successors()) {
2310         auto SuccProb = MBPI->getEdgeProbability(Pred, Succ);
2311         BlockChain *SuccChain = BlockToChain[Succ];
2312         // Check if Succ can be placed after Pred.
2313         // Succ should not be in any chain, or it is the head of some chain.
2314         if ((!SuccChain || Succ == *SuccChain->begin()) && SuccProb > TopProb) {
2315           TopOK = false;
2316           break;
2317         }
2318       }
2319       if (TopOK)
2320         return true;
2321     }
2322   }
2323   return false;
2324 }
2325 
2326 /// Attempt to rotate an exiting block to the bottom of the loop.
2327 ///
2328 /// Once we have built a chain, try to rotate it to line up the hot exit block
2329 /// with fallthrough out of the loop if doing so doesn't introduce unnecessary
2330 /// branches. For example, if the loop has fallthrough into its header and out
2331 /// of its bottom already, don't rotate it.
2332 void MachineBlockPlacement::rotateLoop(BlockChain &LoopChain,
2333                                        const MachineBasicBlock *ExitingBB,
2334                                        BlockFrequency ExitFreq,
2335                                        const BlockFilterSet &LoopBlockSet) {
2336   if (!ExitingBB)
2337     return;
2338 
2339   MachineBasicBlock *Top = *LoopChain.begin();
2340   MachineBasicBlock *Bottom = *std::prev(LoopChain.end());
2341 
2342   // If ExitingBB is already the last one in a chain then nothing to do.
2343   if (Bottom == ExitingBB)
2344     return;
2345 
2346   // The entry block should always be the first BB in a function.
2347   if (Top->isEntryBlock())
2348     return;
2349 
2350   bool ViableTopFallthrough = hasViableTopFallthrough(Top, LoopBlockSet);
2351 
2352   // If the header has viable fallthrough, check whether the current loop
2353   // bottom is a viable exiting block. If so, bail out as rotating will
2354   // introduce an unnecessary branch.
2355   if (ViableTopFallthrough) {
2356     for (MachineBasicBlock *Succ : Bottom->successors()) {
2357       BlockChain *SuccChain = BlockToChain[Succ];
2358       if (!LoopBlockSet.count(Succ) &&
2359           (!SuccChain || Succ == *SuccChain->begin()))
2360         return;
2361     }
2362 
2363     // Rotate will destroy the top fallthrough, we need to ensure the new exit
2364     // frequency is larger than top fallthrough.
2365     BlockFrequency FallThrough2Top = TopFallThroughFreq(Top, LoopBlockSet);
2366     if (FallThrough2Top >= ExitFreq)
2367       return;
2368   }
2369 
2370   BlockChain::iterator ExitIt = llvm::find(LoopChain, ExitingBB);
2371   if (ExitIt == LoopChain.end())
2372     return;
2373 
2374   // Rotating a loop exit to the bottom when there is a fallthrough to top
2375   // trades the entry fallthrough for an exit fallthrough.
2376   // If there is no bottom->top edge, but the chosen exit block does have
2377   // a fallthrough, we break that fallthrough for nothing in return.
2378 
2379   // Let's consider an example. We have a built chain of basic blocks
2380   // B1, B2, ..., Bn, where Bk is a ExitingBB - chosen exit block.
2381   // By doing a rotation we get
2382   // Bk+1, ..., Bn, B1, ..., Bk
2383   // Break of fallthrough to B1 is compensated by a fallthrough from Bk.
2384   // If we had a fallthrough Bk -> Bk+1 it is broken now.
2385   // It might be compensated by fallthrough Bn -> B1.
2386   // So we have a condition to avoid creation of extra branch by loop rotation.
2387   // All below must be true to avoid loop rotation:
2388   //   If there is a fallthrough to top (B1)
2389   //   There was fallthrough from chosen exit block (Bk) to next one (Bk+1)
2390   //   There is no fallthrough from bottom (Bn) to top (B1).
2391   // Please note that there is no exit fallthrough from Bn because we checked it
2392   // above.
2393   if (ViableTopFallthrough) {
2394     assert(std::next(ExitIt) != LoopChain.end() &&
2395            "Exit should not be last BB");
2396     MachineBasicBlock *NextBlockInChain = *std::next(ExitIt);
2397     if (ExitingBB->isSuccessor(NextBlockInChain))
2398       if (!Bottom->isSuccessor(Top))
2399         return;
2400   }
2401 
2402   LLVM_DEBUG(dbgs() << "Rotating loop to put exit " << getBlockName(ExitingBB)
2403                     << " at bottom\n");
2404   std::rotate(LoopChain.begin(), std::next(ExitIt), LoopChain.end());
2405 }
2406 
2407 /// Attempt to rotate a loop based on profile data to reduce branch cost.
2408 ///
2409 /// With profile data, we can determine the cost in terms of missed fall through
2410 /// opportunities when rotating a loop chain and select the best rotation.
2411 /// Basically, there are three kinds of cost to consider for each rotation:
2412 ///    1. The possibly missed fall through edge (if it exists) from BB out of
2413 ///    the loop to the loop header.
2414 ///    2. The possibly missed fall through edges (if they exist) from the loop
2415 ///    exits to BB out of the loop.
2416 ///    3. The missed fall through edge (if it exists) from the last BB to the
2417 ///    first BB in the loop chain.
2418 ///  Therefore, the cost for a given rotation is the sum of costs listed above.
2419 ///  We select the best rotation with the smallest cost.
2420 void MachineBlockPlacement::rotateLoopWithProfile(
2421     BlockChain &LoopChain, const MachineLoop &L,
2422     const BlockFilterSet &LoopBlockSet) {
2423   auto RotationPos = LoopChain.end();
2424   MachineBasicBlock *ChainHeaderBB = *LoopChain.begin();
2425 
2426   // The entry block should always be the first BB in a function.
2427   if (ChainHeaderBB->isEntryBlock())
2428     return;
2429 
2430   BlockFrequency SmallestRotationCost = BlockFrequency::max();
2431 
2432   // A utility lambda that scales up a block frequency by dividing it by a
2433   // branch probability which is the reciprocal of the scale.
2434   auto ScaleBlockFrequency = [](BlockFrequency Freq,
2435                                 unsigned Scale) -> BlockFrequency {
2436     if (Scale == 0)
2437       return BlockFrequency(0);
2438     // Use operator / between BlockFrequency and BranchProbability to implement
2439     // saturating multiplication.
2440     return Freq / BranchProbability(1, Scale);
2441   };
2442 
2443   // Compute the cost of the missed fall-through edge to the loop header if the
2444   // chain head is not the loop header. As we only consider natural loops with
2445   // single header, this computation can be done only once.
2446   BlockFrequency HeaderFallThroughCost(0);
2447   for (auto *Pred : ChainHeaderBB->predecessors()) {
2448     BlockChain *PredChain = BlockToChain[Pred];
2449     if (!LoopBlockSet.count(Pred) &&
2450         (!PredChain || Pred == *std::prev(PredChain->end()))) {
2451       auto EdgeFreq = MBFI->getBlockFreq(Pred) *
2452           MBPI->getEdgeProbability(Pred, ChainHeaderBB);
2453       auto FallThruCost = ScaleBlockFrequency(EdgeFreq, MisfetchCost);
2454       // If the predecessor has only an unconditional jump to the header, we
2455       // need to consider the cost of this jump.
2456       if (Pred->succ_size() == 1)
2457         FallThruCost += ScaleBlockFrequency(EdgeFreq, JumpInstCost);
2458       HeaderFallThroughCost = std::max(HeaderFallThroughCost, FallThruCost);
2459     }
2460   }
2461 
2462   // Here we collect all exit blocks in the loop, and for each exit we find out
2463   // its hottest exit edge. For each loop rotation, we define the loop exit cost
2464   // as the sum of frequencies of exit edges we collect here, excluding the exit
2465   // edge from the tail of the loop chain.
2466   SmallVector<std::pair<MachineBasicBlock *, BlockFrequency>, 4> ExitsWithFreq;
2467   for (auto *BB : LoopChain) {
2468     auto LargestExitEdgeProb = BranchProbability::getZero();
2469     for (auto *Succ : BB->successors()) {
2470       BlockChain *SuccChain = BlockToChain[Succ];
2471       if (!LoopBlockSet.count(Succ) &&
2472           (!SuccChain || Succ == *SuccChain->begin())) {
2473         auto SuccProb = MBPI->getEdgeProbability(BB, Succ);
2474         LargestExitEdgeProb = std::max(LargestExitEdgeProb, SuccProb);
2475       }
2476     }
2477     if (LargestExitEdgeProb > BranchProbability::getZero()) {
2478       auto ExitFreq = MBFI->getBlockFreq(BB) * LargestExitEdgeProb;
2479       ExitsWithFreq.emplace_back(BB, ExitFreq);
2480     }
2481   }
2482 
2483   // In this loop we iterate every block in the loop chain and calculate the
2484   // cost assuming the block is the head of the loop chain. When the loop ends,
2485   // we should have found the best candidate as the loop chain's head.
2486   for (auto Iter = LoopChain.begin(), TailIter = std::prev(LoopChain.end()),
2487             EndIter = LoopChain.end();
2488        Iter != EndIter; Iter++, TailIter++) {
2489     // TailIter is used to track the tail of the loop chain if the block we are
2490     // checking (pointed by Iter) is the head of the chain.
2491     if (TailIter == LoopChain.end())
2492       TailIter = LoopChain.begin();
2493 
2494     auto TailBB = *TailIter;
2495 
2496     // Calculate the cost by putting this BB to the top.
2497     BlockFrequency Cost = BlockFrequency(0);
2498 
2499     // If the current BB is the loop header, we need to take into account the
2500     // cost of the missed fall through edge from outside of the loop to the
2501     // header.
2502     if (Iter != LoopChain.begin())
2503       Cost += HeaderFallThroughCost;
2504 
2505     // Collect the loop exit cost by summing up frequencies of all exit edges
2506     // except the one from the chain tail.
2507     for (auto &ExitWithFreq : ExitsWithFreq)
2508       if (TailBB != ExitWithFreq.first)
2509         Cost += ExitWithFreq.second;
2510 
2511     // The cost of breaking the once fall-through edge from the tail to the top
2512     // of the loop chain. Here we need to consider three cases:
2513     // 1. If the tail node has only one successor, then we will get an
2514     //    additional jmp instruction. So the cost here is (MisfetchCost +
2515     //    JumpInstCost) * tail node frequency.
2516     // 2. If the tail node has two successors, then we may still get an
2517     //    additional jmp instruction if the layout successor after the loop
2518     //    chain is not its CFG successor. Note that the more frequently executed
2519     //    jmp instruction will be put ahead of the other one. Assume the
2520     //    frequency of those two branches are x and y, where x is the frequency
2521     //    of the edge to the chain head, then the cost will be
2522     //    (x * MisfetechCost + min(x, y) * JumpInstCost) * tail node frequency.
2523     // 3. If the tail node has more than two successors (this rarely happens),
2524     //    we won't consider any additional cost.
2525     if (TailBB->isSuccessor(*Iter)) {
2526       auto TailBBFreq = MBFI->getBlockFreq(TailBB);
2527       if (TailBB->succ_size() == 1)
2528         Cost += ScaleBlockFrequency(TailBBFreq, MisfetchCost + JumpInstCost);
2529       else if (TailBB->succ_size() == 2) {
2530         auto TailToHeadProb = MBPI->getEdgeProbability(TailBB, *Iter);
2531         auto TailToHeadFreq = TailBBFreq * TailToHeadProb;
2532         auto ColderEdgeFreq = TailToHeadProb > BranchProbability(1, 2)
2533                                   ? TailBBFreq * TailToHeadProb.getCompl()
2534                                   : TailToHeadFreq;
2535         Cost += ScaleBlockFrequency(TailToHeadFreq, MisfetchCost) +
2536                 ScaleBlockFrequency(ColderEdgeFreq, JumpInstCost);
2537       }
2538     }
2539 
2540     LLVM_DEBUG(dbgs() << "The cost of loop rotation by making "
2541                       << getBlockName(*Iter) << " to the top: "
2542                       << printBlockFreq(MBFI->getMBFI(), Cost) << "\n");
2543 
2544     if (Cost < SmallestRotationCost) {
2545       SmallestRotationCost = Cost;
2546       RotationPos = Iter;
2547     }
2548   }
2549 
2550   if (RotationPos != LoopChain.end()) {
2551     LLVM_DEBUG(dbgs() << "Rotate loop by making " << getBlockName(*RotationPos)
2552                       << " to the top\n");
2553     std::rotate(LoopChain.begin(), RotationPos, LoopChain.end());
2554   }
2555 }
2556 
2557 /// Collect blocks in the given loop that are to be placed.
2558 ///
2559 /// When profile data is available, exclude cold blocks from the returned set;
2560 /// otherwise, collect all blocks in the loop.
2561 MachineBlockPlacement::BlockFilterSet
2562 MachineBlockPlacement::collectLoopBlockSet(const MachineLoop &L) {
2563   BlockFilterSet LoopBlockSet;
2564 
2565   // Filter cold blocks off from LoopBlockSet when profile data is available.
2566   // Collect the sum of frequencies of incoming edges to the loop header from
2567   // outside. If we treat the loop as a super block, this is the frequency of
2568   // the loop. Then for each block in the loop, we calculate the ratio between
2569   // its frequency and the frequency of the loop block. When it is too small,
2570   // don't add it to the loop chain. If there are outer loops, then this block
2571   // will be merged into the first outer loop chain for which this block is not
2572   // cold anymore. This needs precise profile data and we only do this when
2573   // profile data is available.
2574   if (F->getFunction().hasProfileData() || ForceLoopColdBlock) {
2575     BlockFrequency LoopFreq(0);
2576     for (auto *LoopPred : L.getHeader()->predecessors())
2577       if (!L.contains(LoopPred))
2578         LoopFreq += MBFI->getBlockFreq(LoopPred) *
2579                     MBPI->getEdgeProbability(LoopPred, L.getHeader());
2580 
2581     for (MachineBasicBlock *LoopBB : L.getBlocks()) {
2582       if (LoopBlockSet.count(LoopBB))
2583         continue;
2584       auto Freq = MBFI->getBlockFreq(LoopBB).getFrequency();
2585       if (Freq == 0 || LoopFreq.getFrequency() / Freq > LoopToColdBlockRatio)
2586         continue;
2587       BlockChain *Chain = BlockToChain[LoopBB];
2588       for (MachineBasicBlock *ChainBB : *Chain)
2589         LoopBlockSet.insert(ChainBB);
2590     }
2591   } else
2592     LoopBlockSet.insert(L.block_begin(), L.block_end());
2593 
2594   return LoopBlockSet;
2595 }
2596 
2597 /// Forms basic block chains from the natural loop structures.
2598 ///
2599 /// These chains are designed to preserve the existing *structure* of the code
2600 /// as much as possible. We can then stitch the chains together in a way which
2601 /// both preserves the topological structure and minimizes taken conditional
2602 /// branches.
2603 void MachineBlockPlacement::buildLoopChains(const MachineLoop &L) {
2604   // First recurse through any nested loops, building chains for those inner
2605   // loops.
2606   for (const MachineLoop *InnerLoop : L)
2607     buildLoopChains(*InnerLoop);
2608 
2609   assert(BlockWorkList.empty() &&
2610          "BlockWorkList not empty when starting to build loop chains.");
2611   assert(EHPadWorkList.empty() &&
2612          "EHPadWorkList not empty when starting to build loop chains.");
2613   BlockFilterSet LoopBlockSet = collectLoopBlockSet(L);
2614 
2615   // Check if we have profile data for this function. If yes, we will rotate
2616   // this loop by modeling costs more precisely which requires the profile data
2617   // for better layout.
2618   bool RotateLoopWithProfile =
2619       ForcePreciseRotationCost ||
2620       (PreciseRotationCost && F->getFunction().hasProfileData());
2621 
2622   // First check to see if there is an obviously preferable top block for the
2623   // loop. This will default to the header, but may end up as one of the
2624   // predecessors to the header if there is one which will result in strictly
2625   // fewer branches in the loop body.
2626   MachineBasicBlock *LoopTop = findBestLoopTop(L, LoopBlockSet);
2627 
2628   // If we selected just the header for the loop top, look for a potentially
2629   // profitable exit block in the event that rotating the loop can eliminate
2630   // branches by placing an exit edge at the bottom.
2631   //
2632   // Loops are processed innermost to uttermost, make sure we clear
2633   // PreferredLoopExit before processing a new loop.
2634   PreferredLoopExit = nullptr;
2635   BlockFrequency ExitFreq;
2636   if (!RotateLoopWithProfile && LoopTop == L.getHeader())
2637     PreferredLoopExit = findBestLoopExit(L, LoopBlockSet, ExitFreq);
2638 
2639   BlockChain &LoopChain = *BlockToChain[LoopTop];
2640 
2641   // FIXME: This is a really lame way of walking the chains in the loop: we
2642   // walk the blocks, and use a set to prevent visiting a particular chain
2643   // twice.
2644   SmallPtrSet<BlockChain *, 4> UpdatedPreds;
2645   assert(LoopChain.UnscheduledPredecessors == 0 &&
2646          "LoopChain should not have unscheduled predecessors.");
2647   UpdatedPreds.insert(&LoopChain);
2648 
2649   for (const MachineBasicBlock *LoopBB : LoopBlockSet)
2650     fillWorkLists(LoopBB, UpdatedPreds, &LoopBlockSet);
2651 
2652   buildChain(LoopTop, LoopChain, &LoopBlockSet);
2653 
2654   if (RotateLoopWithProfile)
2655     rotateLoopWithProfile(LoopChain, L, LoopBlockSet);
2656   else
2657     rotateLoop(LoopChain, PreferredLoopExit, ExitFreq, LoopBlockSet);
2658 
2659   LLVM_DEBUG({
2660     // Crash at the end so we get all of the debugging output first.
2661     bool BadLoop = false;
2662     if (LoopChain.UnscheduledPredecessors) {
2663       BadLoop = true;
2664       dbgs() << "Loop chain contains a block without its preds placed!\n"
2665              << "  Loop header:  " << getBlockName(*L.block_begin()) << "\n"
2666              << "  Chain header: " << getBlockName(*LoopChain.begin()) << "\n";
2667     }
2668     for (MachineBasicBlock *ChainBB : LoopChain) {
2669       dbgs() << "          ... " << getBlockName(ChainBB) << "\n";
2670       if (!LoopBlockSet.remove(ChainBB)) {
2671         // We don't mark the loop as bad here because there are real situations
2672         // where this can occur. For example, with an unanalyzable fallthrough
2673         // from a loop block to a non-loop block or vice versa.
2674         dbgs() << "Loop chain contains a block not contained by the loop!\n"
2675                << "  Loop header:  " << getBlockName(*L.block_begin()) << "\n"
2676                << "  Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
2677                << "  Bad block:    " << getBlockName(ChainBB) << "\n";
2678       }
2679     }
2680 
2681     if (!LoopBlockSet.empty()) {
2682       BadLoop = true;
2683       for (const MachineBasicBlock *LoopBB : LoopBlockSet)
2684         dbgs() << "Loop contains blocks never placed into a chain!\n"
2685                << "  Loop header:  " << getBlockName(*L.block_begin()) << "\n"
2686                << "  Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
2687                << "  Bad block:    " << getBlockName(LoopBB) << "\n";
2688     }
2689     assert(!BadLoop && "Detected problems with the placement of this loop.");
2690   });
2691 
2692   BlockWorkList.clear();
2693   EHPadWorkList.clear();
2694 }
2695 
2696 void MachineBlockPlacement::buildCFGChains() {
2697   // Ensure that every BB in the function has an associated chain to simplify
2698   // the assumptions of the remaining algorithm.
2699   SmallVector<MachineOperand, 4> Cond; // For analyzeBranch.
2700   for (MachineFunction::iterator FI = F->begin(), FE = F->end(); FI != FE;
2701        ++FI) {
2702     MachineBasicBlock *BB = &*FI;
2703     BlockChain *Chain =
2704         new (ChainAllocator.Allocate()) BlockChain(BlockToChain, BB);
2705     // Also, merge any blocks which we cannot reason about and must preserve
2706     // the exact fallthrough behavior for.
2707     while (true) {
2708       Cond.clear();
2709       MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For analyzeBranch.
2710       if (!TII->analyzeBranch(*BB, TBB, FBB, Cond) || !FI->canFallThrough())
2711         break;
2712 
2713       MachineFunction::iterator NextFI = std::next(FI);
2714       MachineBasicBlock *NextBB = &*NextFI;
2715       // Ensure that the layout successor is a viable block, as we know that
2716       // fallthrough is a possibility.
2717       assert(NextFI != FE && "Can't fallthrough past the last block.");
2718       LLVM_DEBUG(dbgs() << "Pre-merging due to unanalyzable fallthrough: "
2719                         << getBlockName(BB) << " -> " << getBlockName(NextBB)
2720                         << "\n");
2721       Chain->merge(NextBB, nullptr);
2722 #ifndef NDEBUG
2723       BlocksWithUnanalyzableExits.insert(&*BB);
2724 #endif
2725       FI = NextFI;
2726       BB = NextBB;
2727     }
2728   }
2729 
2730   // Build any loop-based chains.
2731   PreferredLoopExit = nullptr;
2732   for (MachineLoop *L : *MLI)
2733     buildLoopChains(*L);
2734 
2735   assert(BlockWorkList.empty() &&
2736          "BlockWorkList should be empty before building final chain.");
2737   assert(EHPadWorkList.empty() &&
2738          "EHPadWorkList should be empty before building final chain.");
2739 
2740   SmallPtrSet<BlockChain *, 4> UpdatedPreds;
2741   for (MachineBasicBlock &MBB : *F)
2742     fillWorkLists(&MBB, UpdatedPreds);
2743 
2744   BlockChain &FunctionChain = *BlockToChain[&F->front()];
2745   buildChain(&F->front(), FunctionChain);
2746 
2747 #ifndef NDEBUG
2748   using FunctionBlockSetType = SmallPtrSet<MachineBasicBlock *, 16>;
2749 #endif
2750   LLVM_DEBUG({
2751     // Crash at the end so we get all of the debugging output first.
2752     bool BadFunc = false;
2753     FunctionBlockSetType FunctionBlockSet;
2754     for (MachineBasicBlock &MBB : *F)
2755       FunctionBlockSet.insert(&MBB);
2756 
2757     for (MachineBasicBlock *ChainBB : FunctionChain)
2758       if (!FunctionBlockSet.erase(ChainBB)) {
2759         BadFunc = true;
2760         dbgs() << "Function chain contains a block not in the function!\n"
2761                << "  Bad block:    " << getBlockName(ChainBB) << "\n";
2762       }
2763 
2764     if (!FunctionBlockSet.empty()) {
2765       BadFunc = true;
2766       for (MachineBasicBlock *RemainingBB : FunctionBlockSet)
2767         dbgs() << "Function contains blocks never placed into a chain!\n"
2768                << "  Bad block:    " << getBlockName(RemainingBB) << "\n";
2769     }
2770     assert(!BadFunc && "Detected problems with the block placement.");
2771   });
2772 
2773   // Remember original layout ordering, so we can update terminators after
2774   // reordering to point to the original layout successor.
2775   SmallVector<MachineBasicBlock *, 4> OriginalLayoutSuccessors(
2776       F->getNumBlockIDs());
2777   {
2778     MachineBasicBlock *LastMBB = nullptr;
2779     for (auto &MBB : *F) {
2780       if (LastMBB != nullptr)
2781         OriginalLayoutSuccessors[LastMBB->getNumber()] = &MBB;
2782       LastMBB = &MBB;
2783     }
2784     OriginalLayoutSuccessors[F->back().getNumber()] = nullptr;
2785   }
2786 
2787   // Splice the blocks into place.
2788   MachineFunction::iterator InsertPos = F->begin();
2789   LLVM_DEBUG(dbgs() << "[MBP] Function: " << F->getName() << "\n");
2790   for (MachineBasicBlock *ChainBB : FunctionChain) {
2791     LLVM_DEBUG(dbgs() << (ChainBB == *FunctionChain.begin() ? "Placing chain "
2792                                                             : "          ... ")
2793                       << getBlockName(ChainBB) << "\n");
2794     if (InsertPos != MachineFunction::iterator(ChainBB))
2795       F->splice(InsertPos, ChainBB);
2796     else
2797       ++InsertPos;
2798 
2799     // Update the terminator of the previous block.
2800     if (ChainBB == *FunctionChain.begin())
2801       continue;
2802     MachineBasicBlock *PrevBB = &*std::prev(MachineFunction::iterator(ChainBB));
2803 
2804     // FIXME: It would be awesome of updateTerminator would just return rather
2805     // than assert when the branch cannot be analyzed in order to remove this
2806     // boiler plate.
2807     Cond.clear();
2808     MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For analyzeBranch.
2809 
2810 #ifndef NDEBUG
2811     if (!BlocksWithUnanalyzableExits.count(PrevBB)) {
2812       // Given the exact block placement we chose, we may actually not _need_ to
2813       // be able to edit PrevBB's terminator sequence, but not being _able_ to
2814       // do that at this point is a bug.
2815       assert((!TII->analyzeBranch(*PrevBB, TBB, FBB, Cond) ||
2816               !PrevBB->canFallThrough()) &&
2817              "Unexpected block with un-analyzable fallthrough!");
2818       Cond.clear();
2819       TBB = FBB = nullptr;
2820     }
2821 #endif
2822 
2823     // The "PrevBB" is not yet updated to reflect current code layout, so,
2824     //   o. it may fall-through to a block without explicit "goto" instruction
2825     //      before layout, and no longer fall-through it after layout; or
2826     //   o. just opposite.
2827     //
2828     // analyzeBranch() may return erroneous value for FBB when these two
2829     // situations take place. For the first scenario FBB is mistakenly set NULL;
2830     // for the 2nd scenario, the FBB, which is expected to be NULL, is
2831     // mistakenly pointing to "*BI".
2832     // Thus, if the future change needs to use FBB before the layout is set, it
2833     // has to correct FBB first by using the code similar to the following:
2834     //
2835     // if (!Cond.empty() && (!FBB || FBB == ChainBB)) {
2836     //   PrevBB->updateTerminator();
2837     //   Cond.clear();
2838     //   TBB = FBB = nullptr;
2839     //   if (TII->analyzeBranch(*PrevBB, TBB, FBB, Cond)) {
2840     //     // FIXME: This should never take place.
2841     //     TBB = FBB = nullptr;
2842     //   }
2843     // }
2844     if (!TII->analyzeBranch(*PrevBB, TBB, FBB, Cond)) {
2845       PrevBB->updateTerminator(OriginalLayoutSuccessors[PrevBB->getNumber()]);
2846     }
2847   }
2848 
2849   // Fixup the last block.
2850   Cond.clear();
2851   MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For analyzeBranch.
2852   if (!TII->analyzeBranch(F->back(), TBB, FBB, Cond)) {
2853     MachineBasicBlock *PrevBB = &F->back();
2854     PrevBB->updateTerminator(OriginalLayoutSuccessors[PrevBB->getNumber()]);
2855   }
2856 
2857   BlockWorkList.clear();
2858   EHPadWorkList.clear();
2859 }
2860 
2861 void MachineBlockPlacement::optimizeBranches() {
2862   BlockChain &FunctionChain = *BlockToChain[&F->front()];
2863   SmallVector<MachineOperand, 4> Cond; // For analyzeBranch.
2864 
2865   // Now that all the basic blocks in the chain have the proper layout,
2866   // make a final call to analyzeBranch with AllowModify set.
2867   // Indeed, the target may be able to optimize the branches in a way we
2868   // cannot because all branches may not be analyzable.
2869   // E.g., the target may be able to remove an unconditional branch to
2870   // a fallthrough when it occurs after predicated terminators.
2871   for (MachineBasicBlock *ChainBB : FunctionChain) {
2872     Cond.clear();
2873     MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For analyzeBranch.
2874     if (!TII->analyzeBranch(*ChainBB, TBB, FBB, Cond, /*AllowModify*/ true)) {
2875       // If PrevBB has a two-way branch, try to re-order the branches
2876       // such that we branch to the successor with higher probability first.
2877       if (TBB && !Cond.empty() && FBB &&
2878           MBPI->getEdgeProbability(ChainBB, FBB) >
2879               MBPI->getEdgeProbability(ChainBB, TBB) &&
2880           !TII->reverseBranchCondition(Cond)) {
2881         LLVM_DEBUG(dbgs() << "Reverse order of the two branches: "
2882                           << getBlockName(ChainBB) << "\n");
2883         LLVM_DEBUG(dbgs() << "    Edge probability: "
2884                           << MBPI->getEdgeProbability(ChainBB, FBB) << " vs "
2885                           << MBPI->getEdgeProbability(ChainBB, TBB) << "\n");
2886         DebugLoc dl; // FIXME: this is nowhere
2887         TII->removeBranch(*ChainBB);
2888         TII->insertBranch(*ChainBB, FBB, TBB, Cond, dl);
2889       }
2890     }
2891   }
2892 }
2893 
2894 void MachineBlockPlacement::alignBlocks() {
2895   // Walk through the backedges of the function now that we have fully laid out
2896   // the basic blocks and align the destination of each backedge. We don't rely
2897   // exclusively on the loop info here so that we can align backedges in
2898   // unnatural CFGs and backedges that were introduced purely because of the
2899   // loop rotations done during this layout pass.
2900   if (F->getFunction().hasMinSize() ||
2901       (F->getFunction().hasOptSize() && !TLI->alignLoopsWithOptSize()))
2902     return;
2903   BlockChain &FunctionChain = *BlockToChain[&F->front()];
2904   if (FunctionChain.begin() == FunctionChain.end())
2905     return; // Empty chain.
2906 
2907   const BranchProbability ColdProb(1, 5); // 20%
2908   BlockFrequency EntryFreq = MBFI->getBlockFreq(&F->front());
2909   BlockFrequency WeightedEntryFreq = EntryFreq * ColdProb;
2910   for (MachineBasicBlock *ChainBB : FunctionChain) {
2911     if (ChainBB == *FunctionChain.begin())
2912       continue;
2913 
2914     // Don't align non-looping basic blocks. These are unlikely to execute
2915     // enough times to matter in practice. Note that we'll still handle
2916     // unnatural CFGs inside of a natural outer loop (the common case) and
2917     // rotated loops.
2918     MachineLoop *L = MLI->getLoopFor(ChainBB);
2919     if (!L)
2920       continue;
2921 
2922     const Align TLIAlign = TLI->getPrefLoopAlignment(L);
2923     unsigned MDAlign = 1;
2924     MDNode *LoopID = L->getLoopID();
2925     if (LoopID) {
2926       for (unsigned I = 1, E = LoopID->getNumOperands(); I < E; ++I) {
2927         MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(I));
2928         if (MD == nullptr)
2929           continue;
2930         MDString *S = dyn_cast<MDString>(MD->getOperand(0));
2931         if (S == nullptr)
2932           continue;
2933         if (S->getString() == "llvm.loop.align") {
2934           assert(MD->getNumOperands() == 2 &&
2935                  "per-loop align metadata should have two operands.");
2936           MDAlign =
2937               mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue();
2938           assert(MDAlign >= 1 && "per-loop align value must be positive.");
2939         }
2940       }
2941     }
2942 
2943     // Use max of the TLIAlign and MDAlign
2944     const Align LoopAlign = std::max(TLIAlign, Align(MDAlign));
2945     if (LoopAlign == 1)
2946       continue; // Don't care about loop alignment.
2947 
2948     // If the block is cold relative to the function entry don't waste space
2949     // aligning it.
2950     BlockFrequency Freq = MBFI->getBlockFreq(ChainBB);
2951     if (Freq < WeightedEntryFreq)
2952       continue;
2953 
2954     // If the block is cold relative to its loop header, don't align it
2955     // regardless of what edges into the block exist.
2956     MachineBasicBlock *LoopHeader = L->getHeader();
2957     BlockFrequency LoopHeaderFreq = MBFI->getBlockFreq(LoopHeader);
2958     if (Freq < (LoopHeaderFreq * ColdProb))
2959       continue;
2960 
2961     // If the global profiles indicates so, don't align it.
2962     if (llvm::shouldOptimizeForSize(ChainBB, PSI, MBFI.get()) &&
2963         !TLI->alignLoopsWithOptSize())
2964       continue;
2965 
2966     // Check for the existence of a non-layout predecessor which would benefit
2967     // from aligning this block.
2968     MachineBasicBlock *LayoutPred =
2969         &*std::prev(MachineFunction::iterator(ChainBB));
2970 
2971     auto DetermineMaxAlignmentPadding = [&]() {
2972       // Set the maximum bytes allowed to be emitted for alignment.
2973       unsigned MaxBytes;
2974       if (MaxBytesForAlignmentOverride.getNumOccurrences() > 0)
2975         MaxBytes = MaxBytesForAlignmentOverride;
2976       else
2977         MaxBytes = TLI->getMaxPermittedBytesForAlignment(ChainBB);
2978       ChainBB->setMaxBytesForAlignment(MaxBytes);
2979     };
2980 
2981     // Force alignment if all the predecessors are jumps. We already checked
2982     // that the block isn't cold above.
2983     if (!LayoutPred->isSuccessor(ChainBB)) {
2984       ChainBB->setAlignment(LoopAlign);
2985       DetermineMaxAlignmentPadding();
2986       continue;
2987     }
2988 
2989     // Align this block if the layout predecessor's edge into this block is
2990     // cold relative to the block. When this is true, other predecessors make up
2991     // all of the hot entries into the block and thus alignment is likely to be
2992     // important.
2993     BranchProbability LayoutProb =
2994         MBPI->getEdgeProbability(LayoutPred, ChainBB);
2995     BlockFrequency LayoutEdgeFreq = MBFI->getBlockFreq(LayoutPred) * LayoutProb;
2996     if (LayoutEdgeFreq <= (Freq * ColdProb)) {
2997       ChainBB->setAlignment(LoopAlign);
2998       DetermineMaxAlignmentPadding();
2999     }
3000   }
3001 }
3002 
3003 /// Tail duplicate \p BB into (some) predecessors if profitable, repeating if
3004 /// it was duplicated into its chain predecessor and removed.
3005 /// \p BB    - Basic block that may be duplicated.
3006 ///
3007 /// \p LPred - Chosen layout predecessor of \p BB.
3008 ///            Updated to be the chain end if LPred is removed.
3009 /// \p Chain - Chain to which \p LPred belongs, and \p BB will belong.
3010 /// \p BlockFilter - Set of blocks that belong to the loop being laid out.
3011 ///                  Used to identify which blocks to update predecessor
3012 ///                  counts.
3013 /// \p PrevUnplacedBlockIt - Iterator pointing to the last block that was
3014 ///                          chosen in the given order due to unnatural CFG
3015 ///                          only needed if \p BB is removed and
3016 ///                          \p PrevUnplacedBlockIt pointed to \p BB.
3017 /// @return true if \p BB was removed.
3018 bool MachineBlockPlacement::repeatedlyTailDuplicateBlock(
3019     MachineBasicBlock *BB, MachineBasicBlock *&LPred,
3020     const MachineBasicBlock *LoopHeaderBB,
3021     BlockChain &Chain, BlockFilterSet *BlockFilter,
3022     MachineFunction::iterator &PrevUnplacedBlockIt) {
3023   bool Removed, DuplicatedToLPred;
3024   bool DuplicatedToOriginalLPred;
3025   Removed = maybeTailDuplicateBlock(BB, LPred, Chain, BlockFilter,
3026                                     PrevUnplacedBlockIt,
3027                                     DuplicatedToLPred);
3028   if (!Removed)
3029     return false;
3030   DuplicatedToOriginalLPred = DuplicatedToLPred;
3031   // Iteratively try to duplicate again. It can happen that a block that is
3032   // duplicated into is still small enough to be duplicated again.
3033   // No need to call markBlockSuccessors in this case, as the blocks being
3034   // duplicated from here on are already scheduled.
3035   while (DuplicatedToLPred && Removed) {
3036     MachineBasicBlock *DupBB, *DupPred;
3037     // The removal callback causes Chain.end() to be updated when a block is
3038     // removed. On the first pass through the loop, the chain end should be the
3039     // same as it was on function entry. On subsequent passes, because we are
3040     // duplicating the block at the end of the chain, if it is removed the
3041     // chain will have shrunk by one block.
3042     BlockChain::iterator ChainEnd = Chain.end();
3043     DupBB = *(--ChainEnd);
3044     // Now try to duplicate again.
3045     if (ChainEnd == Chain.begin())
3046       break;
3047     DupPred = *std::prev(ChainEnd);
3048     Removed = maybeTailDuplicateBlock(DupBB, DupPred, Chain, BlockFilter,
3049                                       PrevUnplacedBlockIt,
3050                                       DuplicatedToLPred);
3051   }
3052   // If BB was duplicated into LPred, it is now scheduled. But because it was
3053   // removed, markChainSuccessors won't be called for its chain. Instead we
3054   // call markBlockSuccessors for LPred to achieve the same effect. This must go
3055   // at the end because repeating the tail duplication can increase the number
3056   // of unscheduled predecessors.
3057   LPred = *std::prev(Chain.end());
3058   if (DuplicatedToOriginalLPred)
3059     markBlockSuccessors(Chain, LPred, LoopHeaderBB, BlockFilter);
3060   return true;
3061 }
3062 
3063 /// Tail duplicate \p BB into (some) predecessors if profitable.
3064 /// \p BB    - Basic block that may be duplicated
3065 /// \p LPred - Chosen layout predecessor of \p BB
3066 /// \p Chain - Chain to which \p LPred belongs, and \p BB will belong.
3067 /// \p BlockFilter - Set of blocks that belong to the loop being laid out.
3068 ///                  Used to identify which blocks to update predecessor
3069 ///                  counts.
3070 /// \p PrevUnplacedBlockIt - Iterator pointing to the last block that was
3071 ///                          chosen in the given order due to unnatural CFG
3072 ///                          only needed if \p BB is removed and
3073 ///                          \p PrevUnplacedBlockIt pointed to \p BB.
3074 /// \p DuplicatedToLPred - True if the block was duplicated into LPred.
3075 /// \return  - True if the block was duplicated into all preds and removed.
3076 bool MachineBlockPlacement::maybeTailDuplicateBlock(
3077     MachineBasicBlock *BB, MachineBasicBlock *LPred,
3078     BlockChain &Chain, BlockFilterSet *BlockFilter,
3079     MachineFunction::iterator &PrevUnplacedBlockIt,
3080     bool &DuplicatedToLPred) {
3081   DuplicatedToLPred = false;
3082   if (!shouldTailDuplicate(BB))
3083     return false;
3084 
3085   LLVM_DEBUG(dbgs() << "Redoing tail duplication for Succ#" << BB->getNumber()
3086                     << "\n");
3087 
3088   // This has to be a callback because none of it can be done after
3089   // BB is deleted.
3090   bool Removed = false;
3091   auto RemovalCallback =
3092       [&](MachineBasicBlock *RemBB) {
3093         // Signal to outer function
3094         Removed = true;
3095 
3096         // Conservative default.
3097         bool InWorkList = true;
3098         // Remove from the Chain and Chain Map
3099         if (BlockToChain.count(RemBB)) {
3100           BlockChain *Chain = BlockToChain[RemBB];
3101           InWorkList = Chain->UnscheduledPredecessors == 0;
3102           Chain->remove(RemBB);
3103           BlockToChain.erase(RemBB);
3104         }
3105 
3106         // Handle the unplaced block iterator
3107         if (&(*PrevUnplacedBlockIt) == RemBB) {
3108           PrevUnplacedBlockIt++;
3109         }
3110 
3111         // Handle the Work Lists
3112         if (InWorkList) {
3113           SmallVectorImpl<MachineBasicBlock *> &RemoveList = BlockWorkList;
3114           if (RemBB->isEHPad())
3115             RemoveList = EHPadWorkList;
3116           llvm::erase(RemoveList, RemBB);
3117         }
3118 
3119         // Handle the filter set
3120         if (BlockFilter) {
3121           BlockFilter->remove(RemBB);
3122         }
3123 
3124         // Remove the block from loop info.
3125         MLI->removeBlock(RemBB);
3126         if (RemBB == PreferredLoopExit)
3127           PreferredLoopExit = nullptr;
3128 
3129         LLVM_DEBUG(dbgs() << "TailDuplicator deleted block: "
3130                           << getBlockName(RemBB) << "\n");
3131       };
3132   auto RemovalCallbackRef =
3133       function_ref<void(MachineBasicBlock*)>(RemovalCallback);
3134 
3135   SmallVector<MachineBasicBlock *, 8> DuplicatedPreds;
3136   bool IsSimple = TailDup.isSimpleBB(BB);
3137   SmallVector<MachineBasicBlock *, 8> CandidatePreds;
3138   SmallVectorImpl<MachineBasicBlock *> *CandidatePtr = nullptr;
3139   if (F->getFunction().hasProfileData()) {
3140     // We can do partial duplication with precise profile information.
3141     findDuplicateCandidates(CandidatePreds, BB, BlockFilter);
3142     if (CandidatePreds.size() == 0)
3143       return false;
3144     if (CandidatePreds.size() < BB->pred_size())
3145       CandidatePtr = &CandidatePreds;
3146   }
3147   TailDup.tailDuplicateAndUpdate(IsSimple, BB, LPred, &DuplicatedPreds,
3148                                  &RemovalCallbackRef, CandidatePtr);
3149 
3150   // Update UnscheduledPredecessors to reflect tail-duplication.
3151   DuplicatedToLPred = false;
3152   for (MachineBasicBlock *Pred : DuplicatedPreds) {
3153     // We're only looking for unscheduled predecessors that match the filter.
3154     BlockChain* PredChain = BlockToChain[Pred];
3155     if (Pred == LPred)
3156       DuplicatedToLPred = true;
3157     if (Pred == LPred || (BlockFilter && !BlockFilter->count(Pred))
3158         || PredChain == &Chain)
3159       continue;
3160     for (MachineBasicBlock *NewSucc : Pred->successors()) {
3161       if (BlockFilter && !BlockFilter->count(NewSucc))
3162         continue;
3163       BlockChain *NewChain = BlockToChain[NewSucc];
3164       if (NewChain != &Chain && NewChain != PredChain)
3165         NewChain->UnscheduledPredecessors++;
3166     }
3167   }
3168   return Removed;
3169 }
3170 
3171 // Count the number of actual machine instructions.
3172 static uint64_t countMBBInstruction(MachineBasicBlock *MBB) {
3173   uint64_t InstrCount = 0;
3174   for (MachineInstr &MI : *MBB) {
3175     if (!MI.isPHI() && !MI.isMetaInstruction())
3176       InstrCount += 1;
3177   }
3178   return InstrCount;
3179 }
3180 
3181 // The size cost of duplication is the instruction size of the duplicated block.
3182 // So we should scale the threshold accordingly. But the instruction size is not
3183 // available on all targets, so we use the number of instructions instead.
3184 BlockFrequency MachineBlockPlacement::scaleThreshold(MachineBasicBlock *BB) {
3185   return BlockFrequency(DupThreshold.getFrequency() * countMBBInstruction(BB));
3186 }
3187 
3188 // Returns true if BB is Pred's best successor.
3189 bool MachineBlockPlacement::isBestSuccessor(MachineBasicBlock *BB,
3190                                             MachineBasicBlock *Pred,
3191                                             BlockFilterSet *BlockFilter) {
3192   if (BB == Pred)
3193     return false;
3194   if (BlockFilter && !BlockFilter->count(Pred))
3195     return false;
3196   BlockChain *PredChain = BlockToChain[Pred];
3197   if (PredChain && (Pred != *std::prev(PredChain->end())))
3198     return false;
3199 
3200   // Find the successor with largest probability excluding BB.
3201   BranchProbability BestProb = BranchProbability::getZero();
3202   for (MachineBasicBlock *Succ : Pred->successors())
3203     if (Succ != BB) {
3204       if (BlockFilter && !BlockFilter->count(Succ))
3205         continue;
3206       BlockChain *SuccChain = BlockToChain[Succ];
3207       if (SuccChain && (Succ != *SuccChain->begin()))
3208         continue;
3209       BranchProbability SuccProb = MBPI->getEdgeProbability(Pred, Succ);
3210       if (SuccProb > BestProb)
3211         BestProb = SuccProb;
3212     }
3213 
3214   BranchProbability BBProb = MBPI->getEdgeProbability(Pred, BB);
3215   if (BBProb <= BestProb)
3216     return false;
3217 
3218   // Compute the number of reduced taken branches if Pred falls through to BB
3219   // instead of another successor. Then compare it with threshold.
3220   BlockFrequency PredFreq = getBlockCountOrFrequency(Pred);
3221   BlockFrequency Gain = PredFreq * (BBProb - BestProb);
3222   return Gain > scaleThreshold(BB);
3223 }
3224 
3225 // Find out the predecessors of BB and BB can be beneficially duplicated into
3226 // them.
3227 void MachineBlockPlacement::findDuplicateCandidates(
3228     SmallVectorImpl<MachineBasicBlock *> &Candidates,
3229     MachineBasicBlock *BB,
3230     BlockFilterSet *BlockFilter) {
3231   MachineBasicBlock *Fallthrough = nullptr;
3232   BranchProbability DefaultBranchProb = BranchProbability::getZero();
3233   BlockFrequency BBDupThreshold(scaleThreshold(BB));
3234   SmallVector<MachineBasicBlock *, 8> Preds(BB->predecessors());
3235   SmallVector<MachineBasicBlock *, 8> Succs(BB->successors());
3236 
3237   // Sort for highest frequency.
3238   auto CmpSucc = [&](MachineBasicBlock *A, MachineBasicBlock *B) {
3239     return MBPI->getEdgeProbability(BB, A) > MBPI->getEdgeProbability(BB, B);
3240   };
3241   auto CmpPred = [&](MachineBasicBlock *A, MachineBasicBlock *B) {
3242     return MBFI->getBlockFreq(A) > MBFI->getBlockFreq(B);
3243   };
3244   llvm::stable_sort(Succs, CmpSucc);
3245   llvm::stable_sort(Preds, CmpPred);
3246 
3247   auto SuccIt = Succs.begin();
3248   if (SuccIt != Succs.end()) {
3249     DefaultBranchProb = MBPI->getEdgeProbability(BB, *SuccIt).getCompl();
3250   }
3251 
3252   // For each predecessors of BB, compute the benefit of duplicating BB,
3253   // if it is larger than the threshold, add it into Candidates.
3254   //
3255   // If we have following control flow.
3256   //
3257   //     PB1 PB2 PB3 PB4
3258   //      \   |  /    /\
3259   //       \  | /    /  \
3260   //        \ |/    /    \
3261   //         BB----/     OB
3262   //         /\
3263   //        /  \
3264   //      SB1 SB2
3265   //
3266   // And it can be partially duplicated as
3267   //
3268   //   PB2+BB
3269   //      |  PB1 PB3 PB4
3270   //      |   |  /    /\
3271   //      |   | /    /  \
3272   //      |   |/    /    \
3273   //      |  BB----/     OB
3274   //      |\ /|
3275   //      | X |
3276   //      |/ \|
3277   //     SB2 SB1
3278   //
3279   // The benefit of duplicating into a predecessor is defined as
3280   //         Orig_taken_branch - Duplicated_taken_branch
3281   //
3282   // The Orig_taken_branch is computed with the assumption that predecessor
3283   // jumps to BB and the most possible successor is laid out after BB.
3284   //
3285   // The Duplicated_taken_branch is computed with the assumption that BB is
3286   // duplicated into PB, and one successor is layout after it (SB1 for PB1 and
3287   // SB2 for PB2 in our case). If there is no available successor, the combined
3288   // block jumps to all BB's successor, like PB3 in this example.
3289   //
3290   // If a predecessor has multiple successors, so BB can't be duplicated into
3291   // it. But it can beneficially fall through to BB, and duplicate BB into other
3292   // predecessors.
3293   for (MachineBasicBlock *Pred : Preds) {
3294     BlockFrequency PredFreq = getBlockCountOrFrequency(Pred);
3295 
3296     if (!TailDup.canTailDuplicate(BB, Pred)) {
3297       // BB can't be duplicated into Pred, but it is possible to be layout
3298       // below Pred.
3299       if (!Fallthrough && isBestSuccessor(BB, Pred, BlockFilter)) {
3300         Fallthrough = Pred;
3301         if (SuccIt != Succs.end())
3302           SuccIt++;
3303       }
3304       continue;
3305     }
3306 
3307     BlockFrequency OrigCost = PredFreq + PredFreq * DefaultBranchProb;
3308     BlockFrequency DupCost;
3309     if (SuccIt == Succs.end()) {
3310       // Jump to all successors;
3311       if (Succs.size() > 0)
3312         DupCost += PredFreq;
3313     } else {
3314       // Fallthrough to *SuccIt, jump to all other successors;
3315       DupCost += PredFreq;
3316       DupCost -= PredFreq * MBPI->getEdgeProbability(BB, *SuccIt);
3317     }
3318 
3319     assert(OrigCost >= DupCost);
3320     OrigCost -= DupCost;
3321     if (OrigCost > BBDupThreshold) {
3322       Candidates.push_back(Pred);
3323       if (SuccIt != Succs.end())
3324         SuccIt++;
3325     }
3326   }
3327 
3328   // No predecessors can optimally fallthrough to BB.
3329   // So we can change one duplication into fallthrough.
3330   if (!Fallthrough) {
3331     if ((Candidates.size() < Preds.size()) && (Candidates.size() > 0)) {
3332       Candidates[0] = Candidates.back();
3333       Candidates.pop_back();
3334     }
3335   }
3336 }
3337 
3338 void MachineBlockPlacement::initDupThreshold() {
3339   DupThreshold = BlockFrequency(0);
3340   if (!F->getFunction().hasProfileData())
3341     return;
3342 
3343   // We prefer to use prifile count.
3344   uint64_t HotThreshold = PSI->getOrCompHotCountThreshold();
3345   if (HotThreshold != UINT64_MAX) {
3346     UseProfileCount = true;
3347     DupThreshold =
3348         BlockFrequency(HotThreshold * TailDupProfilePercentThreshold / 100);
3349     return;
3350   }
3351 
3352   // Profile count is not available, we can use block frequency instead.
3353   BlockFrequency MaxFreq = BlockFrequency(0);
3354   for (MachineBasicBlock &MBB : *F) {
3355     BlockFrequency Freq = MBFI->getBlockFreq(&MBB);
3356     if (Freq > MaxFreq)
3357       MaxFreq = Freq;
3358   }
3359 
3360   BranchProbability ThresholdProb(TailDupPlacementPenalty, 100);
3361   DupThreshold = BlockFrequency(MaxFreq * ThresholdProb);
3362   UseProfileCount = false;
3363 }
3364 
3365 bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &MF) {
3366   if (skipFunction(MF.getFunction()))
3367     return false;
3368 
3369   // Check for single-block functions and skip them.
3370   if (std::next(MF.begin()) == MF.end())
3371     return false;
3372 
3373   F = &MF;
3374   MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
3375   MBFI = std::make_unique<MBFIWrapper>(
3376       getAnalysis<MachineBlockFrequencyInfo>());
3377   MLI = &getAnalysis<MachineLoopInfo>();
3378   TII = MF.getSubtarget().getInstrInfo();
3379   TLI = MF.getSubtarget().getTargetLowering();
3380   MPDT = nullptr;
3381   PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
3382 
3383   initDupThreshold();
3384 
3385   // Initialize PreferredLoopExit to nullptr here since it may never be set if
3386   // there are no MachineLoops.
3387   PreferredLoopExit = nullptr;
3388 
3389   assert(BlockToChain.empty() &&
3390          "BlockToChain map should be empty before starting placement.");
3391   assert(ComputedEdges.empty() &&
3392          "Computed Edge map should be empty before starting placement.");
3393 
3394   unsigned TailDupSize = TailDupPlacementThreshold;
3395   // If only the aggressive threshold is explicitly set, use it.
3396   if (TailDupPlacementAggressiveThreshold.getNumOccurrences() != 0 &&
3397       TailDupPlacementThreshold.getNumOccurrences() == 0)
3398     TailDupSize = TailDupPlacementAggressiveThreshold;
3399 
3400   TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
3401   // For aggressive optimization, we can adjust some thresholds to be less
3402   // conservative.
3403   if (PassConfig->getOptLevel() >= CodeGenOptLevel::Aggressive) {
3404     // At O3 we should be more willing to copy blocks for tail duplication. This
3405     // increases size pressure, so we only do it at O3
3406     // Do this unless only the regular threshold is explicitly set.
3407     if (TailDupPlacementThreshold.getNumOccurrences() == 0 ||
3408         TailDupPlacementAggressiveThreshold.getNumOccurrences() != 0)
3409       TailDupSize = TailDupPlacementAggressiveThreshold;
3410   }
3411 
3412   // If there's no threshold provided through options, query the target
3413   // information for a threshold instead.
3414   if (TailDupPlacementThreshold.getNumOccurrences() == 0 &&
3415       (PassConfig->getOptLevel() < CodeGenOptLevel::Aggressive ||
3416        TailDupPlacementAggressiveThreshold.getNumOccurrences() == 0))
3417     TailDupSize = TII->getTailDuplicateSize(PassConfig->getOptLevel());
3418 
3419   if (allowTailDupPlacement()) {
3420     MPDT = &getAnalysis<MachinePostDominatorTree>();
3421     bool OptForSize = MF.getFunction().hasOptSize() ||
3422                       llvm::shouldOptimizeForSize(&MF, PSI, &MBFI->getMBFI());
3423     if (OptForSize)
3424       TailDupSize = 1;
3425     bool PreRegAlloc = false;
3426     TailDup.initMF(MF, PreRegAlloc, MBPI, MBFI.get(), PSI,
3427                    /* LayoutMode */ true, TailDupSize);
3428     precomputeTriangleChains();
3429   }
3430 
3431   buildCFGChains();
3432 
3433   // Changing the layout can create new tail merging opportunities.
3434   // TailMerge can create jump into if branches that make CFG irreducible for
3435   // HW that requires structured CFG.
3436   bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
3437                          PassConfig->getEnableTailMerge() &&
3438                          BranchFoldPlacement;
3439   // No tail merging opportunities if the block number is less than four.
3440   if (MF.size() > 3 && EnableTailMerge) {
3441     unsigned TailMergeSize = TailDupSize + 1;
3442     BranchFolder BF(/*DefaultEnableTailMerge=*/true, /*CommonHoist=*/false,
3443                     *MBFI, *MBPI, PSI, TailMergeSize);
3444 
3445     if (BF.OptimizeFunction(MF, TII, MF.getSubtarget().getRegisterInfo(), MLI,
3446                             /*AfterPlacement=*/true)) {
3447       // Redo the layout if tail merging creates/removes/moves blocks.
3448       BlockToChain.clear();
3449       ComputedEdges.clear();
3450       // Must redo the post-dominator tree if blocks were changed.
3451       if (MPDT)
3452         MPDT->runOnMachineFunction(MF);
3453       ChainAllocator.DestroyAll();
3454       buildCFGChains();
3455     }
3456   }
3457 
3458   // Apply a post-processing optimizing block placement.
3459   if (MF.size() >= 3 && EnableExtTspBlockPlacement &&
3460       (ApplyExtTspWithoutProfile || MF.getFunction().hasProfileData())) {
3461     // Find a new placement and modify the layout of the blocks in the function.
3462     applyExtTsp();
3463 
3464     // Re-create CFG chain so that we can optimizeBranches and alignBlocks.
3465     createCFGChainExtTsp();
3466   }
3467 
3468   optimizeBranches();
3469   alignBlocks();
3470 
3471   BlockToChain.clear();
3472   ComputedEdges.clear();
3473   ChainAllocator.DestroyAll();
3474 
3475   bool HasMaxBytesOverride =
3476       MaxBytesForAlignmentOverride.getNumOccurrences() > 0;
3477 
3478   if (AlignAllBlock)
3479     // Align all of the blocks in the function to a specific alignment.
3480     for (MachineBasicBlock &MBB : MF) {
3481       if (HasMaxBytesOverride)
3482         MBB.setAlignment(Align(1ULL << AlignAllBlock),
3483                          MaxBytesForAlignmentOverride);
3484       else
3485         MBB.setAlignment(Align(1ULL << AlignAllBlock));
3486     }
3487   else if (AlignAllNonFallThruBlocks) {
3488     // Align all of the blocks that have no fall-through predecessors to a
3489     // specific alignment.
3490     for (auto MBI = std::next(MF.begin()), MBE = MF.end(); MBI != MBE; ++MBI) {
3491       auto LayoutPred = std::prev(MBI);
3492       if (!LayoutPred->isSuccessor(&*MBI)) {
3493         if (HasMaxBytesOverride)
3494           MBI->setAlignment(Align(1ULL << AlignAllNonFallThruBlocks),
3495                             MaxBytesForAlignmentOverride);
3496         else
3497           MBI->setAlignment(Align(1ULL << AlignAllNonFallThruBlocks));
3498       }
3499     }
3500   }
3501   if (ViewBlockLayoutWithBFI != GVDT_None &&
3502       (ViewBlockFreqFuncName.empty() ||
3503        F->getFunction().getName().equals(ViewBlockFreqFuncName))) {
3504     if (RenumberBlocksBeforeView)
3505       MF.RenumberBlocks();
3506     MBFI->view("MBP." + MF.getName(), false);
3507   }
3508 
3509   // We always return true as we have no way to track whether the final order
3510   // differs from the original order.
3511   return true;
3512 }
3513 
3514 void MachineBlockPlacement::applyExtTsp() {
3515   // Prepare data; blocks are indexed by their index in the current ordering.
3516   DenseMap<const MachineBasicBlock *, uint64_t> BlockIndex;
3517   BlockIndex.reserve(F->size());
3518   std::vector<const MachineBasicBlock *> CurrentBlockOrder;
3519   CurrentBlockOrder.reserve(F->size());
3520   size_t NumBlocks = 0;
3521   for (const MachineBasicBlock &MBB : *F) {
3522     BlockIndex[&MBB] = NumBlocks++;
3523     CurrentBlockOrder.push_back(&MBB);
3524   }
3525 
3526   auto BlockSizes = std::vector<uint64_t>(F->size());
3527   auto BlockCounts = std::vector<uint64_t>(F->size());
3528   std::vector<codelayout::EdgeCount> JumpCounts;
3529   for (MachineBasicBlock &MBB : *F) {
3530     // Getting the block frequency.
3531     BlockFrequency BlockFreq = MBFI->getBlockFreq(&MBB);
3532     BlockCounts[BlockIndex[&MBB]] = BlockFreq.getFrequency();
3533     // Getting the block size:
3534     // - approximate the size of an instruction by 4 bytes, and
3535     // - ignore debug instructions.
3536     // Note: getting the exact size of each block is target-dependent and can be
3537     // done by extending the interface of MCCodeEmitter. Experimentally we do
3538     // not see a perf improvement with the exact block sizes.
3539     auto NonDbgInsts =
3540         instructionsWithoutDebug(MBB.instr_begin(), MBB.instr_end());
3541     int NumInsts = std::distance(NonDbgInsts.begin(), NonDbgInsts.end());
3542     BlockSizes[BlockIndex[&MBB]] = 4 * NumInsts;
3543     // Getting jump frequencies.
3544     for (MachineBasicBlock *Succ : MBB.successors()) {
3545       auto EP = MBPI->getEdgeProbability(&MBB, Succ);
3546       BlockFrequency JumpFreq = BlockFreq * EP;
3547       JumpCounts.push_back(
3548           {BlockIndex[&MBB], BlockIndex[Succ], JumpFreq.getFrequency()});
3549     }
3550   }
3551 
3552   LLVM_DEBUG(dbgs() << "Applying ext-tsp layout for |V| = " << F->size()
3553                     << " with profile = " << F->getFunction().hasProfileData()
3554                     << " (" << F->getName().str() << ")"
3555                     << "\n");
3556   LLVM_DEBUG(
3557       dbgs() << format("  original  layout score: %0.2f\n",
3558                        calcExtTspScore(BlockSizes, BlockCounts, JumpCounts)));
3559 
3560   // Run the layout algorithm.
3561   auto NewOrder = computeExtTspLayout(BlockSizes, BlockCounts, JumpCounts);
3562   std::vector<const MachineBasicBlock *> NewBlockOrder;
3563   NewBlockOrder.reserve(F->size());
3564   for (uint64_t Node : NewOrder) {
3565     NewBlockOrder.push_back(CurrentBlockOrder[Node]);
3566   }
3567   LLVM_DEBUG(dbgs() << format("  optimized layout score: %0.2f\n",
3568                               calcExtTspScore(NewOrder, BlockSizes, BlockCounts,
3569                                               JumpCounts)));
3570 
3571   // Assign new block order.
3572   assignBlockOrder(NewBlockOrder);
3573 }
3574 
3575 void MachineBlockPlacement::assignBlockOrder(
3576     const std::vector<const MachineBasicBlock *> &NewBlockOrder) {
3577   assert(F->size() == NewBlockOrder.size() && "Incorrect size of block order");
3578   F->RenumberBlocks();
3579 
3580   bool HasChanges = false;
3581   for (size_t I = 0; I < NewBlockOrder.size(); I++) {
3582     if (NewBlockOrder[I] != F->getBlockNumbered(I)) {
3583       HasChanges = true;
3584       break;
3585     }
3586   }
3587   // Stop early if the new block order is identical to the existing one.
3588   if (!HasChanges)
3589     return;
3590 
3591   SmallVector<MachineBasicBlock *, 4> PrevFallThroughs(F->getNumBlockIDs());
3592   for (auto &MBB : *F) {
3593     PrevFallThroughs[MBB.getNumber()] = MBB.getFallThrough();
3594   }
3595 
3596   // Sort basic blocks in the function according to the computed order.
3597   DenseMap<const MachineBasicBlock *, size_t> NewIndex;
3598   for (const MachineBasicBlock *MBB : NewBlockOrder) {
3599     NewIndex[MBB] = NewIndex.size();
3600   }
3601   F->sort([&](MachineBasicBlock &L, MachineBasicBlock &R) {
3602     return NewIndex[&L] < NewIndex[&R];
3603   });
3604 
3605   // Update basic block branches by inserting explicit fallthrough branches
3606   // when required and re-optimize branches when possible.
3607   const TargetInstrInfo *TII = F->getSubtarget().getInstrInfo();
3608   SmallVector<MachineOperand, 4> Cond;
3609   for (auto &MBB : *F) {
3610     MachineFunction::iterator NextMBB = std::next(MBB.getIterator());
3611     MachineFunction::iterator EndIt = MBB.getParent()->end();
3612     auto *FTMBB = PrevFallThroughs[MBB.getNumber()];
3613     // If this block had a fallthrough before we need an explicit unconditional
3614     // branch to that block if the fallthrough block is not adjacent to the
3615     // block in the new order.
3616     if (FTMBB && (NextMBB == EndIt || &*NextMBB != FTMBB)) {
3617       TII->insertUnconditionalBranch(MBB, FTMBB, MBB.findBranchDebugLoc());
3618     }
3619 
3620     // It might be possible to optimize branches by flipping the condition.
3621     Cond.clear();
3622     MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
3623     if (TII->analyzeBranch(MBB, TBB, FBB, Cond))
3624       continue;
3625     MBB.updateTerminator(FTMBB);
3626   }
3627 
3628 #ifndef NDEBUG
3629   // Make sure we correctly constructed all branches.
3630   F->verify(this, "After optimized block reordering");
3631 #endif
3632 }
3633 
3634 void MachineBlockPlacement::createCFGChainExtTsp() {
3635   BlockToChain.clear();
3636   ComputedEdges.clear();
3637   ChainAllocator.DestroyAll();
3638 
3639   MachineBasicBlock *HeadBB = &F->front();
3640   BlockChain *FunctionChain =
3641       new (ChainAllocator.Allocate()) BlockChain(BlockToChain, HeadBB);
3642 
3643   for (MachineBasicBlock &MBB : *F) {
3644     if (HeadBB == &MBB)
3645       continue; // Ignore head of the chain
3646     FunctionChain->merge(&MBB, nullptr);
3647   }
3648 }
3649 
3650 namespace {
3651 
3652 /// A pass to compute block placement statistics.
3653 ///
3654 /// A separate pass to compute interesting statistics for evaluating block
3655 /// placement. This is separate from the actual placement pass so that they can
3656 /// be computed in the absence of any placement transformations or when using
3657 /// alternative placement strategies.
3658 class MachineBlockPlacementStats : public MachineFunctionPass {
3659   /// A handle to the branch probability pass.
3660   const MachineBranchProbabilityInfo *MBPI;
3661 
3662   /// A handle to the function-wide block frequency pass.
3663   const MachineBlockFrequencyInfo *MBFI;
3664 
3665 public:
3666   static char ID; // Pass identification, replacement for typeid
3667 
3668   MachineBlockPlacementStats() : MachineFunctionPass(ID) {
3669     initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry());
3670   }
3671 
3672   bool runOnMachineFunction(MachineFunction &F) override;
3673 
3674   void getAnalysisUsage(AnalysisUsage &AU) const override {
3675     AU.addRequired<MachineBranchProbabilityInfo>();
3676     AU.addRequired<MachineBlockFrequencyInfo>();
3677     AU.setPreservesAll();
3678     MachineFunctionPass::getAnalysisUsage(AU);
3679   }
3680 };
3681 
3682 } // end anonymous namespace
3683 
3684 char MachineBlockPlacementStats::ID = 0;
3685 
3686 char &llvm::MachineBlockPlacementStatsID = MachineBlockPlacementStats::ID;
3687 
3688 INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats",
3689                       "Basic Block Placement Stats", false, false)
3690 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
3691 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
3692 INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats",
3693                     "Basic Block Placement Stats", false, false)
3694 
3695 bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) {
3696   // Check for single-block functions and skip them.
3697   if (std::next(F.begin()) == F.end())
3698     return false;
3699 
3700   if (!isFunctionInPrintList(F.getName()))
3701     return false;
3702 
3703   MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
3704   MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
3705 
3706   for (MachineBasicBlock &MBB : F) {
3707     BlockFrequency BlockFreq = MBFI->getBlockFreq(&MBB);
3708     Statistic &NumBranches =
3709         (MBB.succ_size() > 1) ? NumCondBranches : NumUncondBranches;
3710     Statistic &BranchTakenFreq =
3711         (MBB.succ_size() > 1) ? CondBranchTakenFreq : UncondBranchTakenFreq;
3712     for (MachineBasicBlock *Succ : MBB.successors()) {
3713       // Skip if this successor is a fallthrough.
3714       if (MBB.isLayoutSuccessor(Succ))
3715         continue;
3716 
3717       BlockFrequency EdgeFreq =
3718           BlockFreq * MBPI->getEdgeProbability(&MBB, Succ);
3719       ++NumBranches;
3720       BranchTakenFreq += EdgeFreq.getFrequency();
3721     }
3722   }
3723 
3724   return false;
3725 }
3726