1 //===-- MachineBlockPlacement.cpp - Basic Block Code Layout optimization --===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements basic block placement transformations using the CFG
11 // structure and branch probability estimates.
13 // The pass strives to preserve the structure of the CFG (that is, retain
14 // a topological ordering of basic blocks) in the absence of a *strong* signal
15 // to the contrary from probabilities. However, within the CFG structure, it
16 // attempts to choose an ordering which favors placing more likely sequences of
17 // blocks adjacent to each other.
19 // The algorithm works from the inner-most loop within a function outward, and
20 // at each stage walks through the basic blocks, trying to coalesce them into
21 // sequential chains where allowed by the CFG (or demanded by heavy
22 // probabilities). Finally, it walks the blocks in topological order, and the
23 // first time it reaches a chain of basic blocks, it schedules them in the
26 //===----------------------------------------------------------------------===//
28 #include "llvm/CodeGen/Passes.h"
29 #include "llvm/CodeGen/TargetPassConfig.h"
30 #include "BranchFolding.h"
31 #include "llvm/ADT/DenseMap.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/CodeGen/MachineBasicBlock.h"
37 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
38 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
39 #include "llvm/CodeGen/MachineFunction.h"
40 #include "llvm/CodeGen/MachineFunctionPass.h"
41 #include "llvm/CodeGen/MachineLoopInfo.h"
42 #include "llvm/CodeGen/MachineModuleInfo.h"
43 #include "llvm/CodeGen/MachinePostDominators.h"
44 #include "llvm/CodeGen/TailDuplicator.h"
45 #include "llvm/Support/Allocator.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/raw_ostream.h"
49 #include "llvm/Target/TargetInstrInfo.h"
50 #include "llvm/Target/TargetLowering.h"
51 #include "llvm/Target/TargetSubtargetInfo.h"
57 #define DEBUG_TYPE "block-placement"
59 STATISTIC(NumCondBranches, "Number of conditional branches");
60 STATISTIC(NumUncondBranches, "Number of unconditional branches");
61 STATISTIC(CondBranchTakenFreq,
62 "Potential frequency of taking conditional branches");
63 STATISTIC(UncondBranchTakenFreq,
64 "Potential frequency of taking unconditional branches");
66 static cl::opt<unsigned> AlignAllBlock("align-all-blocks",
67 cl::desc("Force the alignment of all "
68 "blocks in the function."),
69 cl::init(0), cl::Hidden);
71 static cl::opt<unsigned> AlignAllNonFallThruBlocks(
72 "align-all-nofallthru-blocks",
73 cl::desc("Force the alignment of all "
74 "blocks that have no fall-through predecessors (i.e. don't add "
75 "nops that are executed)."),
76 cl::init(0), cl::Hidden);
78 // FIXME: Find a good default for this flag and remove the flag.
79 static cl::opt<unsigned> ExitBlockBias(
80 "block-placement-exit-block-bias",
81 cl::desc("Block frequency percentage a loop exit block needs "
82 "over the original exit to be considered the new exit."),
83 cl::init(0), cl::Hidden);
86 // - Outlining: placement of a basic block outside the chain or hot path.
88 static cl::opt<unsigned> LoopToColdBlockRatio(
89 "loop-to-cold-block-ratio",
90 cl::desc("Outline loop blocks from loop chain if (frequency of loop) / "
91 "(frequency of block) is greater than this ratio"),
92 cl::init(5), cl::Hidden);
95 PreciseRotationCost("precise-rotation-cost",
96 cl::desc("Model the cost of loop rotation more "
97 "precisely by using profile data."),
98 cl::init(false), cl::Hidden);
100 ForcePreciseRotationCost("force-precise-rotation-cost",
101 cl::desc("Force the use of precise cost "
102 "loop rotation strategy."),
103 cl::init(false), cl::Hidden);
105 static cl::opt<unsigned> MisfetchCost(
107 cl::desc("Cost that models the probabilistic risk of an instruction "
108 "misfetch due to a jump comparing to falling through, whose cost "
110 cl::init(1), cl::Hidden);
112 static cl::opt<unsigned> JumpInstCost("jump-inst-cost",
113 cl::desc("Cost of jump instructions."),
114 cl::init(1), cl::Hidden);
116 TailDupPlacement("tail-dup-placement",
117 cl::desc("Perform tail duplication during placement. "
118 "Creates more fallthrough opportunites in "
119 "outline branches."),
120 cl::init(true), cl::Hidden);
123 BranchFoldPlacement("branch-fold-placement",
124 cl::desc("Perform branch folding during placement. "
125 "Reduces code size."),
126 cl::init(true), cl::Hidden);
128 // Heuristic for tail duplication.
129 static cl::opt<unsigned> TailDupPlacementThreshold(
130 "tail-dup-placement-threshold",
131 cl::desc("Instruction cutoff for tail duplication during layout. "
132 "Tail merging during layout is forced to have a threshold "
133 "that won't conflict."), cl::init(2),
136 // Heuristic for aggressive tail duplication.
137 static cl::opt<unsigned> TailDupPlacementAggressiveThreshold(
138 "tail-dup-placement-aggressive-threshold",
139 cl::desc("Instruction cutoff for aggressive tail duplication during "
140 "layout. Used at -O3. Tail merging during layout is forced to "
141 "have a threshold that won't conflict."), cl::init(3),
144 // Heuristic for tail duplication.
145 static cl::opt<unsigned> TailDupPlacementPenalty(
146 "tail-dup-placement-penalty",
147 cl::desc("Cost penalty for blocks that can avoid breaking CFG by copying. "
148 "Copying can increase fallthrough, but it also increases icache "
149 "pressure. This parameter controls the penalty to account for that. "
150 "Percent as integer."),
154 // Heuristic for triangle chains.
155 static cl::opt<unsigned> TriangleChainCount(
156 "triangle-chain-count",
157 cl::desc("Number of triangle-shaped-CFG's that need to be in a row for the "
158 "triangle tail duplication heuristic to kick in. 0 to disable."),
162 extern cl::opt<unsigned> StaticLikelyProb;
163 extern cl::opt<unsigned> ProfileLikelyProb;
165 // Internal option used to control BFI display only after MBP pass.
166 // Defined in CodeGen/MachineBlockFrequencyInfo.cpp:
167 // -view-block-layout-with-bfi=
168 extern cl::opt<GVDAGType> ViewBlockLayoutWithBFI;
170 // Command line option to specify the name of the function for CFG dump
171 // Defined in Analysis/BlockFrequencyInfo.cpp: -view-bfi-func-name=
172 extern cl::opt<std::string> ViewBlockFreqFuncName;
176 /// \brief Type for our function-wide basic block -> block chain mapping.
177 typedef DenseMap<const MachineBasicBlock *, BlockChain *> BlockToChainMapType;
181 /// \brief A chain of blocks which will be laid out contiguously.
183 /// This is the datastructure representing a chain of consecutive blocks that
184 /// are profitable to layout together in order to maximize fallthrough
185 /// probabilities and code locality. We also can use a block chain to represent
186 /// a sequence of basic blocks which have some external (correctness)
187 /// requirement for sequential layout.
189 /// Chains can be built around a single basic block and can be merged to grow
190 /// them. They participate in a block-to-chain mapping, which is updated
191 /// automatically as chains are merged together.
193 /// \brief The sequence of blocks belonging to this chain.
195 /// This is the sequence of blocks for a particular chain. These will be laid
196 /// out in-order within the function.
197 SmallVector<MachineBasicBlock *, 4> Blocks;
199 /// \brief A handle to the function-wide basic block to block chain mapping.
201 /// This is retained in each block chain to simplify the computation of child
202 /// block chains for SCC-formation and iteration. We store the edges to child
203 /// basic blocks, and map them back to their associated chains using this
205 BlockToChainMapType &BlockToChain;
208 /// \brief Construct a new BlockChain.
210 /// This builds a new block chain representing a single basic block in the
211 /// function. It also registers itself as the chain that block participates
212 /// in with the BlockToChain mapping.
213 BlockChain(BlockToChainMapType &BlockToChain, MachineBasicBlock *BB)
214 : Blocks(1, BB), BlockToChain(BlockToChain), UnscheduledPredecessors(0) {
215 assert(BB && "Cannot create a chain with a null basic block");
216 BlockToChain[BB] = this;
219 /// \brief Iterator over blocks within the chain.
220 typedef SmallVectorImpl<MachineBasicBlock *>::iterator iterator;
221 typedef SmallVectorImpl<MachineBasicBlock *>::const_iterator const_iterator;
223 /// \brief Beginning of blocks within the chain.
224 iterator begin() { return Blocks.begin(); }
225 const_iterator begin() const { return Blocks.begin(); }
227 /// \brief End of blocks within the chain.
228 iterator end() { return Blocks.end(); }
229 const_iterator end() const { return Blocks.end(); }
231 bool remove(MachineBasicBlock* BB) {
232 for(iterator i = begin(); i != end(); ++i) {
241 /// \brief Merge a block chain into this one.
243 /// This routine merges a block chain into this one. It takes care of forming
244 /// a contiguous sequence of basic blocks, updating the edge list, and
245 /// updating the block -> chain mapping. It does not free or tear down the
246 /// old chain, but the old chain's block list is no longer valid.
247 void merge(MachineBasicBlock *BB, BlockChain *Chain) {
249 assert(!Blocks.empty());
251 // Fast path in case we don't have a chain already.
253 assert(!BlockToChain[BB]);
254 Blocks.push_back(BB);
255 BlockToChain[BB] = this;
259 assert(BB == *Chain->begin());
260 assert(Chain->begin() != Chain->end());
262 // Update the incoming blocks to point to this chain, and add them to the
264 for (MachineBasicBlock *ChainBB : *Chain) {
265 Blocks.push_back(ChainBB);
266 assert(BlockToChain[ChainBB] == Chain && "Incoming blocks not in chain");
267 BlockToChain[ChainBB] = this;
272 /// \brief Dump the blocks in this chain.
273 LLVM_DUMP_METHOD void dump() {
274 for (MachineBasicBlock *MBB : *this)
279 /// \brief Count of predecessors of any block within the chain which have not
280 /// yet been scheduled. In general, we will delay scheduling this chain
281 /// until those predecessors are scheduled (or we find a sufficiently good
282 /// reason to override this heuristic.) Note that when forming loop chains,
283 /// blocks outside the loop are ignored and treated as if they were already
286 /// Note: This field is reinitialized multiple times - once for each loop,
287 /// and then once for the function as a whole.
288 unsigned UnscheduledPredecessors;
293 class MachineBlockPlacement : public MachineFunctionPass {
294 /// \brief A typedef for a block filter set.
295 typedef SmallSetVector<const MachineBasicBlock *, 16> BlockFilterSet;
297 /// Pair struct containing basic block and taildup profitiability
298 struct BlockAndTailDupResult {
299 MachineBasicBlock *BB;
303 /// Triple struct containing edge weight and the edge.
304 struct WeightedEdge {
305 BlockFrequency Weight;
306 MachineBasicBlock *Src;
307 MachineBasicBlock *Dest;
310 /// \brief work lists of blocks that are ready to be laid out
311 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
312 SmallVector<MachineBasicBlock *, 16> EHPadWorkList;
314 /// Edges that have already been computed as optimal.
315 DenseMap<const MachineBasicBlock *, BlockAndTailDupResult> ComputedEdges;
317 /// \brief Machine Function
320 /// \brief A handle to the branch probability pass.
321 const MachineBranchProbabilityInfo *MBPI;
323 /// \brief A handle to the function-wide block frequency pass.
324 std::unique_ptr<BranchFolder::MBFIWrapper> MBFI;
326 /// \brief A handle to the loop info.
327 MachineLoopInfo *MLI;
329 /// \brief Preferred loop exit.
330 /// Member variable for convenience. It may be removed by duplication deep
331 /// in the call stack.
332 MachineBasicBlock *PreferredLoopExit;
334 /// \brief A handle to the target's instruction info.
335 const TargetInstrInfo *TII;
337 /// \brief A handle to the target's lowering info.
338 const TargetLoweringBase *TLI;
340 /// \brief A handle to the post dominator tree.
341 MachinePostDominatorTree *MPDT;
343 /// \brief Duplicator used to duplicate tails during placement.
345 /// Placement decisions can open up new tail duplication opportunities, but
346 /// since tail duplication affects placement decisions of later blocks, it
347 /// must be done inline.
348 TailDuplicator TailDup;
350 /// \brief Allocator and owner of BlockChain structures.
352 /// We build BlockChains lazily while processing the loop structure of
353 /// a function. To reduce malloc traffic, we allocate them using this
354 /// slab-like allocator, and destroy them after the pass completes. An
355 /// important guarantee is that this allocator produces stable pointers to
357 SpecificBumpPtrAllocator<BlockChain> ChainAllocator;
359 /// \brief Function wide BasicBlock to BlockChain mapping.
361 /// This mapping allows efficiently moving from any given basic block to the
362 /// BlockChain it participates in, if any. We use it to, among other things,
363 /// allow implicitly defining edges between chains as the existing edges
364 /// between basic blocks.
365 DenseMap<const MachineBasicBlock *, BlockChain *> BlockToChain;
368 /// The set of basic blocks that have terminators that cannot be fully
369 /// analyzed. These basic blocks cannot be re-ordered safely by
370 /// MachineBlockPlacement, and we must preserve physical layout of these
371 /// blocks and their successors through the pass.
372 SmallPtrSet<MachineBasicBlock *, 4> BlocksWithUnanalyzableExits;
375 /// Decrease the UnscheduledPredecessors count for all blocks in chain, and
376 /// if the count goes to 0, add them to the appropriate work list.
377 void markChainSuccessors(
378 const BlockChain &Chain, const MachineBasicBlock *LoopHeaderBB,
379 const BlockFilterSet *BlockFilter = nullptr);
381 /// Decrease the UnscheduledPredecessors count for a single block, and
382 /// if the count goes to 0, add them to the appropriate work list.
383 void markBlockSuccessors(
384 const BlockChain &Chain, const MachineBasicBlock *BB,
385 const MachineBasicBlock *LoopHeaderBB,
386 const BlockFilterSet *BlockFilter = nullptr);
389 collectViableSuccessors(
390 const MachineBasicBlock *BB, const BlockChain &Chain,
391 const BlockFilterSet *BlockFilter,
392 SmallVector<MachineBasicBlock *, 4> &Successors);
393 bool shouldPredBlockBeOutlined(
394 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
395 const BlockChain &Chain, const BlockFilterSet *BlockFilter,
396 BranchProbability SuccProb, BranchProbability HotProb);
397 bool repeatedlyTailDuplicateBlock(
398 MachineBasicBlock *BB, MachineBasicBlock *&LPred,
399 const MachineBasicBlock *LoopHeaderBB,
400 BlockChain &Chain, BlockFilterSet *BlockFilter,
401 MachineFunction::iterator &PrevUnplacedBlockIt);
402 bool maybeTailDuplicateBlock(
403 MachineBasicBlock *BB, MachineBasicBlock *LPred,
404 BlockChain &Chain, BlockFilterSet *BlockFilter,
405 MachineFunction::iterator &PrevUnplacedBlockIt,
406 bool &DuplicatedToPred);
407 bool hasBetterLayoutPredecessor(
408 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
409 const BlockChain &SuccChain, BranchProbability SuccProb,
410 BranchProbability RealSuccProb, const BlockChain &Chain,
411 const BlockFilterSet *BlockFilter);
412 BlockAndTailDupResult selectBestSuccessor(
413 const MachineBasicBlock *BB, const BlockChain &Chain,
414 const BlockFilterSet *BlockFilter);
415 MachineBasicBlock *selectBestCandidateBlock(
416 const BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList);
417 MachineBasicBlock *getFirstUnplacedBlock(
418 const BlockChain &PlacedChain,
419 MachineFunction::iterator &PrevUnplacedBlockIt,
420 const BlockFilterSet *BlockFilter);
422 /// \brief Add a basic block to the work list if it is appropriate.
424 /// If the optional parameter BlockFilter is provided, only MBB
425 /// present in the set will be added to the worklist. If nullptr
426 /// is provided, no filtering occurs.
427 void fillWorkLists(const MachineBasicBlock *MBB,
428 SmallPtrSetImpl<BlockChain *> &UpdatedPreds,
429 const BlockFilterSet *BlockFilter);
430 void buildChain(const MachineBasicBlock *BB, BlockChain &Chain,
431 BlockFilterSet *BlockFilter = nullptr);
432 MachineBasicBlock *findBestLoopTop(
433 const MachineLoop &L, const BlockFilterSet &LoopBlockSet);
434 MachineBasicBlock *findBestLoopExit(
435 const MachineLoop &L, const BlockFilterSet &LoopBlockSet);
436 BlockFilterSet collectLoopBlockSet(const MachineLoop &L);
437 void buildLoopChains(const MachineLoop &L);
439 BlockChain &LoopChain, const MachineBasicBlock *ExitingBB,
440 const BlockFilterSet &LoopBlockSet);
441 void rotateLoopWithProfile(
442 BlockChain &LoopChain, const MachineLoop &L,
443 const BlockFilterSet &LoopBlockSet);
444 void buildCFGChains();
445 void optimizeBranches();
447 /// Returns true if a block should be tail-duplicated to increase fallthrough
449 bool shouldTailDuplicate(MachineBasicBlock *BB);
450 /// Check the edge frequencies to see if tail duplication will increase
452 bool isProfitableToTailDup(
453 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
454 BranchProbability AdjustedSumProb,
455 const BlockChain &Chain, const BlockFilterSet *BlockFilter);
456 /// Check for a trellis layout.
457 bool isTrellis(const MachineBasicBlock *BB,
458 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
459 const BlockChain &Chain, const BlockFilterSet *BlockFilter);
460 /// Get the best successor given a trellis layout.
461 BlockAndTailDupResult getBestTrellisSuccessor(
462 const MachineBasicBlock *BB,
463 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
464 BranchProbability AdjustedSumProb, const BlockChain &Chain,
465 const BlockFilterSet *BlockFilter);
466 /// Get the best pair of non-conflicting edges.
467 static std::pair<WeightedEdge, WeightedEdge> getBestNonConflictingEdges(
468 const MachineBasicBlock *BB,
469 MutableArrayRef<SmallVector<WeightedEdge, 8>> Edges);
470 /// Returns true if a block can tail duplicate into all unplaced
471 /// predecessors. Filters based on loop.
472 bool canTailDuplicateUnplacedPreds(
473 const MachineBasicBlock *BB, MachineBasicBlock *Succ,
474 const BlockChain &Chain, const BlockFilterSet *BlockFilter);
475 /// Find chains of triangles to tail-duplicate where a global analysis works,
476 /// but a local analysis would not find them.
477 void precomputeTriangleChains();
480 static char ID; // Pass identification, replacement for typeid
481 MachineBlockPlacement() : MachineFunctionPass(ID) {
482 initializeMachineBlockPlacementPass(*PassRegistry::getPassRegistry());
485 bool runOnMachineFunction(MachineFunction &F) override;
487 void getAnalysisUsage(AnalysisUsage &AU) const override {
488 AU.addRequired<MachineBranchProbabilityInfo>();
489 AU.addRequired<MachineBlockFrequencyInfo>();
490 if (TailDupPlacement)
491 AU.addRequired<MachinePostDominatorTree>();
492 AU.addRequired<MachineLoopInfo>();
493 AU.addRequired<TargetPassConfig>();
494 MachineFunctionPass::getAnalysisUsage(AU);
499 char MachineBlockPlacement::ID = 0;
500 char &llvm::MachineBlockPlacementID = MachineBlockPlacement::ID;
501 INITIALIZE_PASS_BEGIN(MachineBlockPlacement, "block-placement",
502 "Branch Probability Basic Block Placement", false, false)
503 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
504 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
505 INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTree)
506 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
507 INITIALIZE_PASS_END(MachineBlockPlacement, "block-placement",
508 "Branch Probability Basic Block Placement", false, false)
511 /// \brief Helper to print the name of a MBB.
513 /// Only used by debug logging.
514 static std::string getBlockName(const MachineBasicBlock *BB) {
516 raw_string_ostream OS(Result);
517 OS << "BB#" << BB->getNumber();
518 OS << " ('" << BB->getName() << "')";
524 /// \brief Mark a chain's successors as having one fewer preds.
526 /// When a chain is being merged into the "placed" chain, this routine will
527 /// quickly walk the successors of each block in the chain and mark them as
528 /// having one fewer active predecessor. It also adds any successors of this
529 /// chain which reach the zero-predecessor state to the appropriate worklist.
530 void MachineBlockPlacement::markChainSuccessors(
531 const BlockChain &Chain, const MachineBasicBlock *LoopHeaderBB,
532 const BlockFilterSet *BlockFilter) {
533 // Walk all the blocks in this chain, marking their successors as having
534 // a predecessor placed.
535 for (MachineBasicBlock *MBB : Chain) {
536 markBlockSuccessors(Chain, MBB, LoopHeaderBB, BlockFilter);
540 /// \brief Mark a single block's successors as having one fewer preds.
542 /// Under normal circumstances, this is only called by markChainSuccessors,
543 /// but if a block that was to be placed is completely tail-duplicated away,
544 /// and was duplicated into the chain end, we need to redo markBlockSuccessors
545 /// for just that block.
546 void MachineBlockPlacement::markBlockSuccessors(
547 const BlockChain &Chain, const MachineBasicBlock *MBB,
548 const MachineBasicBlock *LoopHeaderBB, const BlockFilterSet *BlockFilter) {
549 // Add any successors for which this is the only un-placed in-loop
550 // predecessor to the worklist as a viable candidate for CFG-neutral
551 // placement. No subsequent placement of this block will violate the CFG
552 // shape, so we get to use heuristics to choose a favorable placement.
553 for (MachineBasicBlock *Succ : MBB->successors()) {
554 if (BlockFilter && !BlockFilter->count(Succ))
556 BlockChain &SuccChain = *BlockToChain[Succ];
557 // Disregard edges within a fixed chain, or edges to the loop header.
558 if (&Chain == &SuccChain || Succ == LoopHeaderBB)
561 // This is a cross-chain edge that is within the loop, so decrement the
562 // loop predecessor count of the destination chain.
563 if (SuccChain.UnscheduledPredecessors == 0 ||
564 --SuccChain.UnscheduledPredecessors > 0)
567 auto *NewBB = *SuccChain.begin();
568 if (NewBB->isEHPad())
569 EHPadWorkList.push_back(NewBB);
571 BlockWorkList.push_back(NewBB);
575 /// This helper function collects the set of successors of block
576 /// \p BB that are allowed to be its layout successors, and return
577 /// the total branch probability of edges from \p BB to those
579 BranchProbability MachineBlockPlacement::collectViableSuccessors(
580 const MachineBasicBlock *BB, const BlockChain &Chain,
581 const BlockFilterSet *BlockFilter,
582 SmallVector<MachineBasicBlock *, 4> &Successors) {
583 // Adjust edge probabilities by excluding edges pointing to blocks that is
584 // either not in BlockFilter or is already in the current chain. Consider the
593 // Assume A->C is very hot (>90%), and C->D has a 50% probability, then after
594 // A->C is chosen as a fall-through, D won't be selected as a successor of C
595 // due to CFG constraint (the probability of C->D is not greater than
596 // HotProb to break top-order). If we exclude E that is not in BlockFilter
597 // when calculating the probability of C->D, D will be selected and we
598 // will get A C D B as the layout of this loop.
599 auto AdjustedSumProb = BranchProbability::getOne();
600 for (MachineBasicBlock *Succ : BB->successors()) {
601 bool SkipSucc = false;
602 if (Succ->isEHPad() || (BlockFilter && !BlockFilter->count(Succ))) {
605 BlockChain *SuccChain = BlockToChain[Succ];
606 if (SuccChain == &Chain) {
608 } else if (Succ != *SuccChain->begin()) {
609 DEBUG(dbgs() << " " << getBlockName(Succ) << " -> Mid chain!\n");
614 AdjustedSumProb -= MBPI->getEdgeProbability(BB, Succ);
616 Successors.push_back(Succ);
619 return AdjustedSumProb;
622 /// The helper function returns the branch probability that is adjusted
623 /// or normalized over the new total \p AdjustedSumProb.
624 static BranchProbability
625 getAdjustedProbability(BranchProbability OrigProb,
626 BranchProbability AdjustedSumProb) {
627 BranchProbability SuccProb;
628 uint32_t SuccProbN = OrigProb.getNumerator();
629 uint32_t SuccProbD = AdjustedSumProb.getNumerator();
630 if (SuccProbN >= SuccProbD)
631 SuccProb = BranchProbability::getOne();
633 SuccProb = BranchProbability(SuccProbN, SuccProbD);
638 /// Check if \p BB has exactly the successors in \p Successors.
640 hasSameSuccessors(MachineBasicBlock &BB,
641 SmallPtrSetImpl<const MachineBasicBlock *> &Successors) {
642 if (BB.succ_size() != Successors.size())
644 // We don't want to count self-loops
645 if (Successors.count(&BB))
647 for (MachineBasicBlock *Succ : BB.successors())
648 if (!Successors.count(Succ))
653 /// Check if a block should be tail duplicated to increase fallthrough
655 /// \p BB Block to check.
656 bool MachineBlockPlacement::shouldTailDuplicate(MachineBasicBlock *BB) {
657 // Blocks with single successors don't create additional fallthrough
658 // opportunities. Don't duplicate them. TODO: When conditional exits are
659 // analyzable, allow them to be duplicated.
660 bool IsSimple = TailDup.isSimpleBB(BB);
662 if (BB->succ_size() == 1)
664 return TailDup.shouldTailDuplicate(IsSimple, *BB);
667 /// Compare 2 BlockFrequency's with a small penalty for \p A.
668 /// In order to be conservative, we apply a X% penalty to account for
669 /// increased icache pressure and static heuristics. For small frequencies
670 /// we use only the numerators to improve accuracy. For simplicity, we assume the
671 /// penalty is less than 100%
672 /// TODO(iteratee): Use 64-bit fixed point edge frequencies everywhere.
673 static bool greaterWithBias(BlockFrequency A, BlockFrequency B,
674 uint64_t EntryFreq) {
675 BranchProbability ThresholdProb(TailDupPlacementPenalty, 100);
676 BlockFrequency Gain = A - B;
677 return (Gain / ThresholdProb).getFrequency() >= EntryFreq;
680 /// Check the edge frequencies to see if tail duplication will increase
681 /// fallthroughs. It only makes sense to call this function when
682 /// \p Succ would not be chosen otherwise. Tail duplication of \p Succ is
683 /// always locally profitable if we would have picked \p Succ without
684 /// considering duplication.
685 bool MachineBlockPlacement::isProfitableToTailDup(
686 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
687 BranchProbability QProb,
688 const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
689 // We need to do a probability calculation to make sure this is profitable.
690 // First: does succ have a successor that post-dominates? This affects the
691 // calculation. The 2 relevant cases are:
706 // '=' : Branch taken for that CFG edge
707 // In the second case, Placing Succ while duplicating it into C prevents the
708 // fallthrough of Succ into either D or PDom, because they now have C as an
709 // unplaced predecessor
711 // Start by figuring out which case we fall into
712 MachineBasicBlock *PDom = nullptr;
713 SmallVector<MachineBasicBlock *, 4> SuccSuccs;
714 // Only scan the relevant successors
715 auto AdjustedSuccSumProb =
716 collectViableSuccessors(Succ, Chain, BlockFilter, SuccSuccs);
717 BranchProbability PProb = MBPI->getEdgeProbability(BB, Succ);
718 auto BBFreq = MBFI->getBlockFreq(BB);
719 auto SuccFreq = MBFI->getBlockFreq(Succ);
720 BlockFrequency P = BBFreq * PProb;
721 BlockFrequency Qout = BBFreq * QProb;
722 uint64_t EntryFreq = MBFI->getEntryFreq();
723 // If there are no more successors, it is profitable to copy, as it strictly
724 // increases fallthrough.
725 if (SuccSuccs.size() == 0)
726 return greaterWithBias(P, Qout, EntryFreq);
728 auto BestSuccSucc = BranchProbability::getZero();
729 // Find the PDom or the best Succ if no PDom exists.
730 for (MachineBasicBlock *SuccSucc : SuccSuccs) {
731 auto Prob = MBPI->getEdgeProbability(Succ, SuccSucc);
732 if (Prob > BestSuccSucc)
735 if (MPDT->dominates(SuccSucc, Succ)) {
740 // For the comparisons, we need to know Succ's best incoming edge that isn't
742 auto SuccBestPred = BlockFrequency(0);
743 for (MachineBasicBlock *SuccPred : Succ->predecessors()) {
744 if (SuccPred == Succ || SuccPred == BB
745 || BlockToChain[SuccPred] == &Chain
746 || (BlockFilter && !BlockFilter->count(SuccPred)))
748 auto Freq = MBFI->getBlockFreq(SuccPred)
749 * MBPI->getEdgeProbability(SuccPred, Succ);
750 if (Freq > SuccBestPred)
753 // Qin is Succ's best unplaced incoming edge that isn't BB
754 BlockFrequency Qin = SuccBestPred;
755 // If it doesn't have a post-dominating successor, here is the calculation:
767 // '=' : Branch taken for that CFG edge
768 // Cost in the first case is: P + V
769 // For this calculation, we always assume P > Qout. If Qout > P
770 // The result of this function will be ignored at the caller.
771 // Let F = SuccFreq - Qin
772 // Cost in the second case is: Qout + min(Qin, F) * U + max(Qin, F) * V
774 if (PDom == nullptr || !Succ->isSuccessor(PDom)) {
775 BranchProbability UProb = BestSuccSucc;
776 BranchProbability VProb = AdjustedSuccSumProb - UProb;
777 BlockFrequency F = SuccFreq - Qin;
778 BlockFrequency V = SuccFreq * VProb;
779 BlockFrequency QinU = std::min(Qin, F) * UProb;
780 BlockFrequency BaseCost = P + V;
781 BlockFrequency DupCost = Qout + QinU + std::max(Qin, F) * VProb;
782 return greaterWithBias(BaseCost, DupCost, EntryFreq);
784 BranchProbability UProb = MBPI->getEdgeProbability(Succ, PDom);
785 BranchProbability VProb = AdjustedSuccSumProb - UProb;
786 BlockFrequency U = SuccFreq * UProb;
787 BlockFrequency V = SuccFreq * VProb;
788 BlockFrequency F = SuccFreq - Qin;
789 // If there is a post-dominating successor, here is the calculation:
791 // | \Qout | \ | \Qout | \
793 // = C' |P C = C' |P C
794 // | /Qin | | | /Qin | |
795 // | / | C' (+Succ) | / | C' (+Succ)
796 // Succ Succ /| Succ Succ /|
797 // | \ V | \/ | | \ V | \/ |
798 // |U \ |U /\ =? |U = |U /\ |
799 // = D = = =?| | D | = =|
804 // '=' : Branch taken for that CFG edge
805 // The cost for taken branches in the first case is P + U
806 // Let F = SuccFreq - Qin
807 // The cost in the second case (assuming independence), given the layout:
808 // BB, Succ, (C+Succ), D, Dom or the layout:
809 // BB, Succ, D, Dom, (C+Succ)
810 // is Qout + max(F, Qin) * U + min(F, Qin)
811 // compare P + U vs Qout + P * U + Qin.
813 // The 3rd and 4th cases cover when Dom would be chosen to follow Succ.
815 // For the 3rd case, the cost is P + 2 * V
816 // For the 4th case, the cost is Qout + min(Qin, F) * U + max(Qin, F) * V + V
817 // We choose 4 over 3 when (P + V) > Qout + min(Qin, F) * U + max(Qin, F) * V
818 if (UProb > AdjustedSuccSumProb / 2 &&
819 !hasBetterLayoutPredecessor(Succ, PDom, *BlockToChain[PDom], UProb, UProb,
822 return greaterWithBias(
823 (P + V), (Qout + std::max(Qin, F) * VProb + std::min(Qin, F) * UProb),
826 return greaterWithBias((P + U),
827 (Qout + std::min(Qin, F) * AdjustedSuccSumProb +
828 std::max(Qin, F) * UProb),
832 /// Check for a trellis layout. \p BB is the upper part of a trellis if its
833 /// successors form the lower part of a trellis. A successor set S forms the
834 /// lower part of a trellis if all of the predecessors of S are either in S or
835 /// have all of S as successors. We ignore trellises where BB doesn't have 2
836 /// successors because for fewer than 2, it's trivial, and for 3 or greater they
837 /// are very uncommon and complex to compute optimally. Allowing edges within S
838 /// is not strictly a trellis, but the same algorithm works, so we allow it.
839 bool MachineBlockPlacement::isTrellis(
840 const MachineBasicBlock *BB,
841 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
842 const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
843 // Technically BB could form a trellis with branching factor higher than 2.
844 // But that's extremely uncommon.
845 if (BB->succ_size() != 2 || ViableSuccs.size() != 2)
848 SmallPtrSet<const MachineBasicBlock *, 2> Successors(BB->succ_begin(),
850 // To avoid reviewing the same predecessors twice.
851 SmallPtrSet<const MachineBasicBlock *, 8> SeenPreds;
853 for (MachineBasicBlock *Succ : ViableSuccs) {
855 for (auto SuccPred : Succ->predecessors()) {
856 // Allow triangle successors, but don't count them.
857 if (Successors.count(SuccPred)) {
858 // Make sure that it is actually a triangle.
859 for (MachineBasicBlock *CheckSucc : SuccPred->successors())
860 if (!Successors.count(CheckSucc))
864 const BlockChain *PredChain = BlockToChain[SuccPred];
865 if (SuccPred == BB || (BlockFilter && !BlockFilter->count(SuccPred)) ||
866 PredChain == &Chain || PredChain == BlockToChain[Succ])
869 // Perform the successor check only once.
870 if (!SeenPreds.insert(SuccPred).second)
872 if (!hasSameSuccessors(*SuccPred, Successors))
875 // If one of the successors has only BB as a predecessor, it is not a
883 /// Pick the highest total weight pair of edges that can both be laid out.
884 /// The edges in \p Edges[0] are assumed to have a different destination than
885 /// the edges in \p Edges[1]. Simple counting shows that the best pair is either
886 /// the individual highest weight edges to the 2 different destinations, or in
887 /// case of a conflict, one of them should be replaced with a 2nd best edge.
888 std::pair<MachineBlockPlacement::WeightedEdge,
889 MachineBlockPlacement::WeightedEdge>
890 MachineBlockPlacement::getBestNonConflictingEdges(
891 const MachineBasicBlock *BB,
892 MutableArrayRef<SmallVector<MachineBlockPlacement::WeightedEdge, 8>>
894 // Sort the edges, and then for each successor, find the best incoming
895 // predecessor. If the best incoming predecessors aren't the same,
896 // then that is clearly the best layout. If there is a conflict, one of the
897 // successors will have to fallthrough from the second best predecessor. We
898 // compare which combination is better overall.
900 // Sort for highest frequency.
901 auto Cmp = [](WeightedEdge A, WeightedEdge B) { return A.Weight > B.Weight; };
903 std::stable_sort(Edges[0].begin(), Edges[0].end(), Cmp);
904 std::stable_sort(Edges[1].begin(), Edges[1].end(), Cmp);
905 auto BestA = Edges[0].begin();
906 auto BestB = Edges[1].begin();
907 // Arrange for the correct answer to be in BestA and BestB
908 // If the 2 best edges don't conflict, the answer is already there.
909 if (BestA->Src == BestB->Src) {
910 // Compare the total fallthrough of (Best + Second Best) for both pairs
911 auto SecondBestA = std::next(BestA);
912 auto SecondBestB = std::next(BestB);
913 BlockFrequency BestAScore = BestA->Weight + SecondBestB->Weight;
914 BlockFrequency BestBScore = BestB->Weight + SecondBestA->Weight;
915 if (BestAScore < BestBScore)
920 // Arrange for the BB edge to be in BestA if it exists.
921 if (BestB->Src == BB)
922 std::swap(BestA, BestB);
923 return std::make_pair(*BestA, *BestB);
926 /// Get the best successor from \p BB based on \p BB being part of a trellis.
927 /// We only handle trellises with 2 successors, so the algorithm is
928 /// straightforward: Find the best pair of edges that don't conflict. We find
929 /// the best incoming edge for each successor in the trellis. If those conflict,
930 /// we consider which of them should be replaced with the second best.
931 /// Upon return the two best edges will be in \p BestEdges. If one of the edges
932 /// comes from \p BB, it will be in \p BestEdges[0]
933 MachineBlockPlacement::BlockAndTailDupResult
934 MachineBlockPlacement::getBestTrellisSuccessor(
935 const MachineBasicBlock *BB,
936 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
937 BranchProbability AdjustedSumProb, const BlockChain &Chain,
938 const BlockFilterSet *BlockFilter) {
940 BlockAndTailDupResult Result = {nullptr, false};
941 SmallPtrSet<const MachineBasicBlock *, 4> Successors(BB->succ_begin(),
944 // We assume size 2 because it's common. For general n, we would have to do
945 // the Hungarian algorithm, but it's not worth the complexity because more
946 // than 2 successors is fairly uncommon, and a trellis even more so.
947 if (Successors.size() != 2 || ViableSuccs.size() != 2)
950 // Collect the edge frequencies of all edges that form the trellis.
951 SmallVector<WeightedEdge, 8> Edges[2];
953 for (auto Succ : ViableSuccs) {
954 for (MachineBasicBlock *SuccPred : Succ->predecessors()) {
955 // Skip any placed predecessors that are not BB
957 if ((BlockFilter && !BlockFilter->count(SuccPred)) ||
958 BlockToChain[SuccPred] == &Chain ||
959 BlockToChain[SuccPred] == BlockToChain[Succ])
961 BlockFrequency EdgeFreq = MBFI->getBlockFreq(SuccPred) *
962 MBPI->getEdgeProbability(SuccPred, Succ);
963 Edges[SuccIndex].push_back({EdgeFreq, SuccPred, Succ});
968 // Pick the best combination of 2 edges from all the edges in the trellis.
969 WeightedEdge BestA, BestB;
970 std::tie(BestA, BestB) = getBestNonConflictingEdges(BB, Edges);
972 if (BestA.Src != BB) {
973 // If we have a trellis, and BB doesn't have the best fallthrough edges,
974 // we shouldn't choose any successor. We've already looked and there's a
975 // better fallthrough edge for all the successors.
976 DEBUG(dbgs() << "Trellis, but not one of the chosen edges.\n");
980 // Did we pick the triangle edge? If tail-duplication is profitable, do
981 // that instead. Otherwise merge the triangle edge now while we know it is
983 if (BestA.Dest == BestB.Src) {
984 // The edges are BB->Succ1->Succ2, and we're looking to see if BB->Succ2
986 MachineBasicBlock *Succ1 = BestA.Dest;
987 MachineBasicBlock *Succ2 = BestB.Dest;
988 // Check to see if tail-duplication would be profitable.
989 if (TailDupPlacement && shouldTailDuplicate(Succ2) &&
990 canTailDuplicateUnplacedPreds(BB, Succ2, Chain, BlockFilter) &&
991 isProfitableToTailDup(BB, Succ2, MBPI->getEdgeProbability(BB, Succ1),
992 Chain, BlockFilter)) {
993 DEBUG(BranchProbability Succ2Prob = getAdjustedProbability(
994 MBPI->getEdgeProbability(BB, Succ2), AdjustedSumProb);
995 dbgs() << " Selected: " << getBlockName(Succ2)
996 << ", probability: " << Succ2Prob << " (Tail Duplicate)\n");
998 Result.ShouldTailDup = true;
1002 // We have already computed the optimal edge for the other side of the
1004 ComputedEdges[BestB.Src] = { BestB.Dest, false };
1006 auto TrellisSucc = BestA.Dest;
1007 DEBUG(BranchProbability SuccProb = getAdjustedProbability(
1008 MBPI->getEdgeProbability(BB, TrellisSucc), AdjustedSumProb);
1009 dbgs() << " Selected: " << getBlockName(TrellisSucc)
1010 << ", probability: " << SuccProb << " (Trellis)\n");
1011 Result.BB = TrellisSucc;
1015 /// When the option TailDupPlacement is on, this method checks if the
1016 /// fallthrough candidate block \p Succ (of block \p BB) can be tail-duplicated
1017 /// into all of its unplaced, unfiltered predecessors, that are not BB.
1018 bool MachineBlockPlacement::canTailDuplicateUnplacedPreds(
1019 const MachineBasicBlock *BB, MachineBasicBlock *Succ,
1020 const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
1021 if (!shouldTailDuplicate(Succ))
1024 // For CFG checking.
1025 SmallPtrSet<const MachineBasicBlock *, 4> Successors(BB->succ_begin(),
1027 for (MachineBasicBlock *Pred : Succ->predecessors()) {
1028 // Make sure all unplaced and unfiltered predecessors can be
1029 // tail-duplicated into.
1030 // Skip any blocks that are already placed or not in this loop.
1031 if (Pred == BB || (BlockFilter && !BlockFilter->count(Pred))
1032 || BlockToChain[Pred] == &Chain)
1034 if (!TailDup.canTailDuplicate(Succ, Pred)) {
1035 if (Successors.size() > 1 && hasSameSuccessors(*Pred, Successors))
1036 // This will result in a trellis after tail duplication, so we don't
1037 // need to copy Succ into this predecessor. In the presence
1038 // of a trellis tail duplication can continue to be profitable.
1054 // After BB was duplicated into C, the layout looks like the one on the
1055 // right. BB and C now have the same successors. When considering
1056 // whether Succ can be duplicated into all its unplaced predecessors, we
1058 // We can do this because C already has a profitable fallthrough, namely
1059 // D. TODO(iteratee): ignore sufficiently cold predecessors for
1060 // duplication and for this test.
1062 // This allows trellises to be laid out in 2 separate chains
1063 // (A,B,Succ,...) and later (C,D,...) This is a reasonable heuristic
1064 // because it allows the creation of 2 fallthrough paths with links
1065 // between them, and we correctly identify the best layout for these
1066 // CFGs. We want to extend trellises that the user created in addition
1067 // to trellises created by tail-duplication, so we just look for the
1076 /// Find chains of triangles where we believe it would be profitable to
1077 /// tail-duplicate them all, but a local analysis would not find them.
1078 /// There are 3 ways this can be profitable:
1079 /// 1) The post-dominators marked 50% are actually taken 55% (This shrinks with
1081 /// 2) The chains are statically correlated. Branch probabilities have a very
1082 /// U-shaped distribution.
1083 /// [http://nrs.harvard.edu/urn-3:HUL.InstRepos:24015805]
1084 /// If the branches in a chain are likely to be from the same side of the
1085 /// distribution as their predecessor, but are independent at runtime, this
1086 /// transformation is profitable. (Because the cost of being wrong is a small
1087 /// fixed cost, unlike the standard triangle layout where the cost of being
1088 /// wrong scales with the # of triangles.)
1089 /// 3) The chains are dynamically correlated. If the probability that a previous
1090 /// branch was taken positively influences whether the next branch will be
1092 /// We believe that 2 and 3 are common enough to justify the small margin in 1.
1093 void MachineBlockPlacement::precomputeTriangleChains() {
1094 struct TriangleChain {
1095 std::vector<MachineBasicBlock *> Edges;
1096 TriangleChain(MachineBasicBlock *src, MachineBasicBlock *dst)
1097 : Edges({src, dst}) {}
1099 void append(MachineBasicBlock *dst) {
1100 assert(getKey()->isSuccessor(dst) &&
1101 "Attempting to append a block that is not a successor.");
1102 Edges.push_back(dst);
1105 unsigned count() const { return Edges.size() - 1; }
1107 MachineBasicBlock *getKey() const {
1108 return Edges.back();
1112 if (TriangleChainCount == 0)
1115 DEBUG(dbgs() << "Pre-computing triangle chains.\n");
1116 // Map from last block to the chain that contains it. This allows us to extend
1117 // chains as we find new triangles.
1118 DenseMap<const MachineBasicBlock *, TriangleChain> TriangleChainMap;
1119 for (MachineBasicBlock &BB : *F) {
1120 // If BB doesn't have 2 successors, it doesn't start a triangle.
1121 if (BB.succ_size() != 2)
1123 MachineBasicBlock *PDom = nullptr;
1124 for (MachineBasicBlock *Succ : BB.successors()) {
1125 if (!MPDT->dominates(Succ, &BB))
1130 // If BB doesn't have a post-dominating successor, it doesn't form a
1132 if (PDom == nullptr)
1134 // If PDom has a hint that it is low probability, skip this triangle.
1135 if (MBPI->getEdgeProbability(&BB, PDom) < BranchProbability(50, 100))
1137 // If PDom isn't eligible for duplication, this isn't the kind of triangle
1138 // we're looking for.
1139 if (!shouldTailDuplicate(PDom))
1141 bool CanTailDuplicate = true;
1142 // If PDom can't tail-duplicate into it's non-BB predecessors, then this
1143 // isn't the kind of triangle we're looking for.
1144 for (MachineBasicBlock* Pred : PDom->predecessors()) {
1147 if (!TailDup.canTailDuplicate(PDom, Pred)) {
1148 CanTailDuplicate = false;
1152 // If we can't tail-duplicate PDom to its predecessors, then skip this
1154 if (!CanTailDuplicate)
1157 // Now we have an interesting triangle. Insert it if it's not part of an
1159 // Note: This cannot be replaced with a call insert() or emplace() because
1160 // the find key is BB, but the insert/emplace key is PDom.
1161 auto Found = TriangleChainMap.find(&BB);
1162 // If it is, remove the chain from the map, grow it, and put it back in the
1163 // map with the end as the new key.
1164 if (Found != TriangleChainMap.end()) {
1165 TriangleChain Chain = std::move(Found->second);
1166 TriangleChainMap.erase(Found);
1168 TriangleChainMap.insert(std::make_pair(Chain.getKey(), std::move(Chain)));
1170 auto InsertResult = TriangleChainMap.try_emplace(PDom, &BB, PDom);
1171 assert(InsertResult.second && "Block seen twice.");
1176 // Iterating over a DenseMap is safe here, because the only thing in the body
1177 // of the loop is inserting into another DenseMap (ComputedEdges).
1178 // ComputedEdges is never iterated, so this doesn't lead to non-determinism.
1179 for (auto &ChainPair : TriangleChainMap) {
1180 TriangleChain &Chain = ChainPair.second;
1181 // Benchmarking has shown that due to branch correlation duplicating 2 or
1182 // more triangles is profitable, despite the calculations assuming
1184 if (Chain.count() < TriangleChainCount)
1186 MachineBasicBlock *dst = Chain.Edges.back();
1187 Chain.Edges.pop_back();
1188 for (MachineBasicBlock *src : reverse(Chain.Edges)) {
1189 DEBUG(dbgs() << "Marking edge: " << getBlockName(src) << "->" <<
1190 getBlockName(dst) << " as pre-computed based on triangles.\n");
1192 auto InsertResult = ComputedEdges.insert({src, {dst, true}});
1193 assert(InsertResult.second && "Block seen twice.");
1201 // When profile is not present, return the StaticLikelyProb.
1202 // When profile is available, we need to handle the triangle-shape CFG.
1203 static BranchProbability getLayoutSuccessorProbThreshold(
1204 const MachineBasicBlock *BB) {
1205 if (!BB->getParent()->getFunction()->getEntryCount())
1206 return BranchProbability(StaticLikelyProb, 100);
1207 if (BB->succ_size() == 2) {
1208 const MachineBasicBlock *Succ1 = *BB->succ_begin();
1209 const MachineBasicBlock *Succ2 = *(BB->succ_begin() + 1);
1210 if (Succ1->isSuccessor(Succ2) || Succ2->isSuccessor(Succ1)) {
1211 /* See case 1 below for the cost analysis. For BB->Succ to
1212 * be taken with smaller cost, the following needs to hold:
1213 * Prob(BB->Succ) > 2 * Prob(BB->Pred)
1214 * So the threshold T in the calculation below
1215 * (1-T) * Prob(BB->Succ) > T * Prob(BB->Pred)
1216 * So T / (1 - T) = 2, Yielding T = 2/3
1217 * Also adding user specified branch bias, we have
1218 * T = (2/3)*(ProfileLikelyProb/50)
1219 * = (2*ProfileLikelyProb)/150)
1221 return BranchProbability(2 * ProfileLikelyProb, 150);
1224 return BranchProbability(ProfileLikelyProb, 100);
1227 /// Checks to see if the layout candidate block \p Succ has a better layout
1228 /// predecessor than \c BB. If yes, returns true.
1229 /// \p SuccProb: The probability adjusted for only remaining blocks.
1230 /// Only used for logging
1231 /// \p RealSuccProb: The un-adjusted probability.
1232 /// \p Chain: The chain that BB belongs to and Succ is being considered for.
1233 /// \p BlockFilter: if non-null, the set of blocks that make up the loop being
1235 bool MachineBlockPlacement::hasBetterLayoutPredecessor(
1236 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
1237 const BlockChain &SuccChain, BranchProbability SuccProb,
1238 BranchProbability RealSuccProb, const BlockChain &Chain,
1239 const BlockFilterSet *BlockFilter) {
1241 // There isn't a better layout when there are no unscheduled predecessors.
1242 if (SuccChain.UnscheduledPredecessors == 0)
1245 // There are two basic scenarios here:
1246 // -------------------------------------
1247 // Case 1: triangular shape CFG (if-then):
1254 // In this case, we are evaluating whether to select edge -> Succ, e.g.
1255 // set Succ as the layout successor of BB. Picking Succ as BB's
1256 // successor breaks the CFG constraints (FIXME: define these constraints).
1257 // With this layout, Pred BB
1258 // is forced to be outlined, so the overall cost will be cost of the
1259 // branch taken from BB to Pred, plus the cost of back taken branch
1260 // from Pred to Succ, as well as the additional cost associated
1261 // with the needed unconditional jump instruction from Pred To Succ.
1263 // The cost of the topological order layout is the taken branch cost
1264 // from BB to Succ, so to make BB->Succ a viable candidate, the following
1266 // 2 * freq(BB->Pred) * taken_branch_cost + unconditional_jump_cost
1267 // < freq(BB->Succ) * taken_branch_cost.
1268 // Ignoring unconditional jump cost, we get
1269 // freq(BB->Succ) > 2 * freq(BB->Pred), i.e.,
1270 // prob(BB->Succ) > 2 * prob(BB->Pred)
1272 // When real profile data is available, we can precisely compute the
1273 // probability threshold that is needed for edge BB->Succ to be considered.
1274 // Without profile data, the heuristic requires the branch bias to be
1275 // a lot larger to make sure the signal is very strong (e.g. 80% default).
1276 // -----------------------------------------------------------------
1277 // Case 2: diamond like CFG (if-then-else):
1286 // The current block is BB and edge BB->Succ is now being evaluated.
1287 // Note that edge S->BB was previously already selected because
1288 // prob(S->BB) > prob(S->Pred).
1289 // At this point, 2 blocks can be placed after BB: Pred or Succ. If we
1290 // choose Pred, we will have a topological ordering as shown on the left
1291 // in the picture below. If we choose Succ, we have the solution as shown
1300 // | pred-- | Succ--
1302 // ---succ ---pred--
1304 // cost = freq(S->Pred) + freq(BB->Succ) cost = 2 * freq (S->Pred)
1305 // = freq(S->Pred) + freq(S->BB)
1307 // If we have profile data (i.e, branch probabilities can be trusted), the
1308 // cost (number of taken branches) with layout S->BB->Succ->Pred is 2 *
1309 // freq(S->Pred) while the cost of topo order is freq(S->Pred) + freq(S->BB).
1310 // We know Prob(S->BB) > Prob(S->Pred), so freq(S->BB) > freq(S->Pred), which
1311 // means the cost of topological order is greater.
1312 // When profile data is not available, however, we need to be more
1313 // conservative. If the branch prediction is wrong, breaking the topo-order
1314 // will actually yield a layout with large cost. For this reason, we need
1315 // strong biased branch at block S with Prob(S->BB) in order to select
1316 // BB->Succ. This is equivalent to looking the CFG backward with backward
1317 // edge: Prob(Succ->BB) needs to >= HotProb in order to be selected (without
1319 // --------------------------------------------------------------------------
1320 // Case 3: forked diamond
1332 // The current block is BB and edge BB->S1 is now being evaluated.
1333 // As above S->BB was already selected because
1334 // prob(S->BB) > prob(S->Pred). Assume that prob(BB->S1) >= prob(BB->S2).
1342 // | Pred----| | S1----
1344 // --(S1 or S2) ---Pred--
1348 // topo-cost = freq(S->Pred) + freq(BB->S1) + freq(BB->S2)
1349 // + min(freq(Pred->S1), freq(Pred->S2))
1350 // Non-topo-order cost:
1351 // non-topo-cost = 2 * freq(S->Pred) + freq(BB->S2).
1352 // To be conservative, we can assume that min(freq(Pred->S1), freq(Pred->S2))
1353 // is 0. Then the non topo layout is better when
1354 // freq(S->Pred) < freq(BB->S1).
1355 // This is exactly what is checked below.
1356 // Note there are other shapes that apply (Pred may not be a single block,
1357 // but they all fit this general pattern.)
1358 BranchProbability HotProb = getLayoutSuccessorProbThreshold(BB);
1360 // Make sure that a hot successor doesn't have a globally more
1361 // important predecessor.
1362 BlockFrequency CandidateEdgeFreq = MBFI->getBlockFreq(BB) * RealSuccProb;
1363 bool BadCFGConflict = false;
1365 for (MachineBasicBlock *Pred : Succ->predecessors()) {
1366 if (Pred == Succ || BlockToChain[Pred] == &SuccChain ||
1367 (BlockFilter && !BlockFilter->count(Pred)) ||
1368 BlockToChain[Pred] == &Chain ||
1369 // This check is redundant except for look ahead. This function is
1370 // called for lookahead by isProfitableToTailDup when BB hasn't been
1374 // Do backward checking.
1375 // For all cases above, we need a backward checking to filter out edges that
1376 // are not 'strongly' biased.
1380 // We select edge BB->Succ if
1381 // freq(BB->Succ) > freq(Succ) * HotProb
1382 // i.e. freq(BB->Succ) > freq(BB->Succ) * HotProb + freq(Pred->Succ) *
1384 // i.e. freq((BB->Succ) * (1 - HotProb) > freq(Pred->Succ) * HotProb
1385 // Case 1 is covered too, because the first equation reduces to:
1386 // prob(BB->Succ) > HotProb. (freq(Succ) = freq(BB) for a triangle)
1387 BlockFrequency PredEdgeFreq =
1388 MBFI->getBlockFreq(Pred) * MBPI->getEdgeProbability(Pred, Succ);
1389 if (PredEdgeFreq * HotProb >= CandidateEdgeFreq * HotProb.getCompl()) {
1390 BadCFGConflict = true;
1395 if (BadCFGConflict) {
1396 DEBUG(dbgs() << " Not a candidate: " << getBlockName(Succ) << " -> " << SuccProb
1397 << " (prob) (non-cold CFG conflict)\n");
1404 /// \brief Select the best successor for a block.
1406 /// This looks across all successors of a particular block and attempts to
1407 /// select the "best" one to be the layout successor. It only considers direct
1408 /// successors which also pass the block filter. It will attempt to avoid
1409 /// breaking CFG structure, but cave and break such structures in the case of
1410 /// very hot successor edges.
1412 /// \returns The best successor block found, or null if none are viable, along
1413 /// with a boolean indicating if tail duplication is necessary.
1414 MachineBlockPlacement::BlockAndTailDupResult
1415 MachineBlockPlacement::selectBestSuccessor(
1416 const MachineBasicBlock *BB, const BlockChain &Chain,
1417 const BlockFilterSet *BlockFilter) {
1418 const BranchProbability HotProb(StaticLikelyProb, 100);
1420 BlockAndTailDupResult BestSucc = { nullptr, false };
1421 auto BestProb = BranchProbability::getZero();
1423 SmallVector<MachineBasicBlock *, 4> Successors;
1424 auto AdjustedSumProb =
1425 collectViableSuccessors(BB, Chain, BlockFilter, Successors);
1427 DEBUG(dbgs() << "Selecting best successor for: " << getBlockName(BB) << "\n");
1429 // if we already precomputed the best successor for BB, return that if still
1431 auto FoundEdge = ComputedEdges.find(BB);
1432 if (FoundEdge != ComputedEdges.end()) {
1433 MachineBasicBlock *Succ = FoundEdge->second.BB;
1434 ComputedEdges.erase(FoundEdge);
1435 BlockChain *SuccChain = BlockToChain[Succ];
1436 if (BB->isSuccessor(Succ) && (!BlockFilter || BlockFilter->count(Succ)) &&
1437 SuccChain != &Chain && Succ == *SuccChain->begin())
1438 return FoundEdge->second;
1441 // if BB is part of a trellis, Use the trellis to determine the optimal
1442 // fallthrough edges
1443 if (isTrellis(BB, Successors, Chain, BlockFilter))
1444 return getBestTrellisSuccessor(BB, Successors, AdjustedSumProb, Chain,
1447 // For blocks with CFG violations, we may be able to lay them out anyway with
1448 // tail-duplication. We keep this vector so we can perform the probability
1449 // calculations the minimum number of times.
1450 SmallVector<std::tuple<BranchProbability, MachineBasicBlock *>, 4>
1452 for (MachineBasicBlock *Succ : Successors) {
1453 auto RealSuccProb = MBPI->getEdgeProbability(BB, Succ);
1454 BranchProbability SuccProb =
1455 getAdjustedProbability(RealSuccProb, AdjustedSumProb);
1457 BlockChain &SuccChain = *BlockToChain[Succ];
1458 // Skip the edge \c BB->Succ if block \c Succ has a better layout
1459 // predecessor that yields lower global cost.
1460 if (hasBetterLayoutPredecessor(BB, Succ, SuccChain, SuccProb, RealSuccProb,
1461 Chain, BlockFilter)) {
1462 // If tail duplication would make Succ profitable, place it.
1463 if (TailDupPlacement && shouldTailDuplicate(Succ))
1464 DupCandidates.push_back(std::make_tuple(SuccProb, Succ));
1469 dbgs() << " Candidate: " << getBlockName(Succ) << ", probability: "
1471 << (SuccChain.UnscheduledPredecessors != 0 ? " (CFG break)" : "")
1474 if (BestSucc.BB && BestProb >= SuccProb) {
1475 DEBUG(dbgs() << " Not the best candidate, continuing\n");
1479 DEBUG(dbgs() << " Setting it as best candidate\n");
1481 BestProb = SuccProb;
1483 // Handle the tail duplication candidates in order of decreasing probability.
1484 // Stop at the first one that is profitable. Also stop if they are less
1485 // profitable than BestSucc. Position is important because we preserve it and
1486 // prefer first best match. Here we aren't comparing in order, so we capture
1487 // the position instead.
1488 if (DupCandidates.size() != 0) {
1490 [](const std::tuple<BranchProbability, MachineBasicBlock *> &a,
1491 const std::tuple<BranchProbability, MachineBasicBlock *> &b) {
1492 return std::get<0>(a) > std::get<0>(b);
1494 std::stable_sort(DupCandidates.begin(), DupCandidates.end(), cmp);
1496 for(auto &Tup : DupCandidates) {
1497 BranchProbability DupProb;
1498 MachineBasicBlock *Succ;
1499 std::tie(DupProb, Succ) = Tup;
1500 if (DupProb < BestProb)
1502 if (canTailDuplicateUnplacedPreds(BB, Succ, Chain, BlockFilter)
1503 && (isProfitableToTailDup(BB, Succ, BestProb, Chain, BlockFilter))) {
1505 dbgs() << " Candidate: " << getBlockName(Succ) << ", probability: "
1507 << " (Tail Duplicate)\n");
1509 BestSucc.ShouldTailDup = true;
1515 DEBUG(dbgs() << " Selected: " << getBlockName(BestSucc.BB) << "\n");
1520 /// \brief Select the best block from a worklist.
1522 /// This looks through the provided worklist as a list of candidate basic
1523 /// blocks and select the most profitable one to place. The definition of
1524 /// profitable only really makes sense in the context of a loop. This returns
1525 /// the most frequently visited block in the worklist, which in the case of
1526 /// a loop, is the one most desirable to be physically close to the rest of the
1527 /// loop body in order to improve i-cache behavior.
1529 /// \returns The best block found, or null if none are viable.
1530 MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock(
1531 const BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList) {
1532 // Once we need to walk the worklist looking for a candidate, cleanup the
1533 // worklist of already placed entries.
1534 // FIXME: If this shows up on profiles, it could be folded (at the cost of
1535 // some code complexity) into the loop below.
1536 WorkList.erase(remove_if(WorkList,
1537 [&](MachineBasicBlock *BB) {
1538 return BlockToChain.lookup(BB) == &Chain;
1542 if (WorkList.empty())
1545 bool IsEHPad = WorkList[0]->isEHPad();
1547 MachineBasicBlock *BestBlock = nullptr;
1548 BlockFrequency BestFreq;
1549 for (MachineBasicBlock *MBB : WorkList) {
1550 assert(MBB->isEHPad() == IsEHPad);
1552 BlockChain &SuccChain = *BlockToChain[MBB];
1553 if (&SuccChain == &Chain)
1556 assert(SuccChain.UnscheduledPredecessors == 0 && "Found CFG-violating block");
1558 BlockFrequency CandidateFreq = MBFI->getBlockFreq(MBB);
1559 DEBUG(dbgs() << " " << getBlockName(MBB) << " -> ";
1560 MBFI->printBlockFreq(dbgs(), CandidateFreq) << " (freq)\n");
1562 // For ehpad, we layout the least probable first as to avoid jumping back
1563 // from least probable landingpads to more probable ones.
1565 // FIXME: Using probability is probably (!) not the best way to achieve
1566 // this. We should probably have a more principled approach to layout
1569 // The goal is to get:
1571 // +--------------------------+
1573 // InnerLp -> InnerCleanup OuterLp -> OuterCleanup -> Resume
1577 // +-------------------------------------+
1579 // OuterLp -> OuterCleanup -> Resume InnerLp -> InnerCleanup
1580 if (BestBlock && (IsEHPad ^ (BestFreq >= CandidateFreq)))
1584 BestFreq = CandidateFreq;
1590 /// \brief Retrieve the first unplaced basic block.
1592 /// This routine is called when we are unable to use the CFG to walk through
1593 /// all of the basic blocks and form a chain due to unnatural loops in the CFG.
1594 /// We walk through the function's blocks in order, starting from the
1595 /// LastUnplacedBlockIt. We update this iterator on each call to avoid
1596 /// re-scanning the entire sequence on repeated calls to this routine.
1597 MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock(
1598 const BlockChain &PlacedChain,
1599 MachineFunction::iterator &PrevUnplacedBlockIt,
1600 const BlockFilterSet *BlockFilter) {
1601 for (MachineFunction::iterator I = PrevUnplacedBlockIt, E = F->end(); I != E;
1603 if (BlockFilter && !BlockFilter->count(&*I))
1605 if (BlockToChain[&*I] != &PlacedChain) {
1606 PrevUnplacedBlockIt = I;
1607 // Now select the head of the chain to which the unplaced block belongs
1608 // as the block to place. This will force the entire chain to be placed,
1609 // and satisfies the requirements of merging chains.
1610 return *BlockToChain[&*I]->begin();
1616 void MachineBlockPlacement::fillWorkLists(
1617 const MachineBasicBlock *MBB,
1618 SmallPtrSetImpl<BlockChain *> &UpdatedPreds,
1619 const BlockFilterSet *BlockFilter = nullptr) {
1620 BlockChain &Chain = *BlockToChain[MBB];
1621 if (!UpdatedPreds.insert(&Chain).second)
1624 assert(Chain.UnscheduledPredecessors == 0);
1625 for (MachineBasicBlock *ChainBB : Chain) {
1626 assert(BlockToChain[ChainBB] == &Chain);
1627 for (MachineBasicBlock *Pred : ChainBB->predecessors()) {
1628 if (BlockFilter && !BlockFilter->count(Pred))
1630 if (BlockToChain[Pred] == &Chain)
1632 ++Chain.UnscheduledPredecessors;
1636 if (Chain.UnscheduledPredecessors != 0)
1639 MachineBasicBlock *BB = *Chain.begin();
1641 EHPadWorkList.push_back(BB);
1643 BlockWorkList.push_back(BB);
1646 void MachineBlockPlacement::buildChain(
1647 const MachineBasicBlock *HeadBB, BlockChain &Chain,
1648 BlockFilterSet *BlockFilter) {
1649 assert(HeadBB && "BB must not be null.\n");
1650 assert(BlockToChain[HeadBB] == &Chain && "BlockToChainMap mis-match.\n");
1651 MachineFunction::iterator PrevUnplacedBlockIt = F->begin();
1653 const MachineBasicBlock *LoopHeaderBB = HeadBB;
1654 markChainSuccessors(Chain, LoopHeaderBB, BlockFilter);
1655 MachineBasicBlock *BB = *std::prev(Chain.end());
1657 assert(BB && "null block found at end of chain in loop.");
1658 assert(BlockToChain[BB] == &Chain && "BlockToChainMap mis-match in loop.");
1659 assert(*std::prev(Chain.end()) == BB && "BB Not found at end of chain.");
1662 // Look for the best viable successor if there is one to place immediately
1663 // after this block.
1664 auto Result = selectBestSuccessor(BB, Chain, BlockFilter);
1665 MachineBasicBlock* BestSucc = Result.BB;
1666 bool ShouldTailDup = Result.ShouldTailDup;
1667 if (TailDupPlacement)
1668 ShouldTailDup |= (BestSucc && shouldTailDuplicate(BestSucc));
1670 // If an immediate successor isn't available, look for the best viable
1671 // block among those we've identified as not violating the loop's CFG at
1672 // this point. This won't be a fallthrough, but it will increase locality.
1674 BestSucc = selectBestCandidateBlock(Chain, BlockWorkList);
1676 BestSucc = selectBestCandidateBlock(Chain, EHPadWorkList);
1679 BestSucc = getFirstUnplacedBlock(Chain, PrevUnplacedBlockIt, BlockFilter);
1683 DEBUG(dbgs() << "Unnatural loop CFG detected, forcibly merging the "
1684 "layout successor until the CFG reduces\n");
1687 // Placement may have changed tail duplication opportunities.
1688 // Check for that now.
1689 if (TailDupPlacement && BestSucc && ShouldTailDup) {
1690 // If the chosen successor was duplicated into all its predecessors,
1691 // don't bother laying it out, just go round the loop again with BB as
1693 if (repeatedlyTailDuplicateBlock(BestSucc, BB, LoopHeaderBB, Chain,
1694 BlockFilter, PrevUnplacedBlockIt))
1698 // Place this block, updating the datastructures to reflect its placement.
1699 BlockChain &SuccChain = *BlockToChain[BestSucc];
1700 // Zero out UnscheduledPredecessors for the successor we're about to merge in case
1701 // we selected a successor that didn't fit naturally into the CFG.
1702 SuccChain.UnscheduledPredecessors = 0;
1703 DEBUG(dbgs() << "Merging from " << getBlockName(BB) << " to "
1704 << getBlockName(BestSucc) << "\n");
1705 markChainSuccessors(SuccChain, LoopHeaderBB, BlockFilter);
1706 Chain.merge(BestSucc, &SuccChain);
1707 BB = *std::prev(Chain.end());
1710 DEBUG(dbgs() << "Finished forming chain for header block "
1711 << getBlockName(*Chain.begin()) << "\n");
1714 /// \brief Find the best loop top block for layout.
1716 /// Look for a block which is strictly better than the loop header for laying
1717 /// out at the top of the loop. This looks for one and only one pattern:
1718 /// a latch block with no conditional exit. This block will cause a conditional
1719 /// jump around it or will be the bottom of the loop if we lay it out in place,
1720 /// but if it it doesn't end up at the bottom of the loop for any reason,
1721 /// rotation alone won't fix it. Because such a block will always result in an
1722 /// unconditional jump (for the backedge) rotating it in front of the loop
1723 /// header is always profitable.
1725 MachineBlockPlacement::findBestLoopTop(const MachineLoop &L,
1726 const BlockFilterSet &LoopBlockSet) {
1727 // Placing the latch block before the header may introduce an extra branch
1728 // that skips this block the first time the loop is executed, which we want
1729 // to avoid when optimising for size.
1730 // FIXME: in theory there is a case that does not introduce a new branch,
1731 // i.e. when the layout predecessor does not fallthrough to the loop header.
1732 // In practice this never happens though: there always seems to be a preheader
1733 // that can fallthrough and that is also placed before the header.
1734 if (F->getFunction()->optForSize())
1735 return L.getHeader();
1737 // Check that the header hasn't been fused with a preheader block due to
1738 // crazy branches. If it has, we need to start with the header at the top to
1739 // prevent pulling the preheader into the loop body.
1740 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
1741 if (!LoopBlockSet.count(*HeaderChain.begin()))
1742 return L.getHeader();
1744 DEBUG(dbgs() << "Finding best loop top for: " << getBlockName(L.getHeader())
1747 BlockFrequency BestPredFreq;
1748 MachineBasicBlock *BestPred = nullptr;
1749 for (MachineBasicBlock *Pred : L.getHeader()->predecessors()) {
1750 if (!LoopBlockSet.count(Pred))
1752 DEBUG(dbgs() << " header pred: " << getBlockName(Pred) << ", has "
1753 << Pred->succ_size() << " successors, ";
1754 MBFI->printBlockFreq(dbgs(), Pred) << " freq\n");
1755 if (Pred->succ_size() > 1)
1758 BlockFrequency PredFreq = MBFI->getBlockFreq(Pred);
1759 if (!BestPred || PredFreq > BestPredFreq ||
1760 (!(PredFreq < BestPredFreq) &&
1761 Pred->isLayoutSuccessor(L.getHeader()))) {
1763 BestPredFreq = PredFreq;
1767 // If no direct predecessor is fine, just use the loop header.
1769 DEBUG(dbgs() << " final top unchanged\n");
1770 return L.getHeader();
1773 // Walk backwards through any straight line of predecessors.
1774 while (BestPred->pred_size() == 1 &&
1775 (*BestPred->pred_begin())->succ_size() == 1 &&
1776 *BestPred->pred_begin() != L.getHeader())
1777 BestPred = *BestPred->pred_begin();
1779 DEBUG(dbgs() << " final top: " << getBlockName(BestPred) << "\n");
1783 /// \brief Find the best loop exiting block for layout.
1785 /// This routine implements the logic to analyze the loop looking for the best
1786 /// block to layout at the top of the loop. Typically this is done to maximize
1787 /// fallthrough opportunities.
1789 MachineBlockPlacement::findBestLoopExit(const MachineLoop &L,
1790 const BlockFilterSet &LoopBlockSet) {
1791 // We don't want to layout the loop linearly in all cases. If the loop header
1792 // is just a normal basic block in the loop, we want to look for what block
1793 // within the loop is the best one to layout at the top. However, if the loop
1794 // header has be pre-merged into a chain due to predecessors not having
1795 // analyzable branches, *and* the predecessor it is merged with is *not* part
1796 // of the loop, rotating the header into the middle of the loop will create
1797 // a non-contiguous range of blocks which is Very Bad. So start with the
1798 // header and only rotate if safe.
1799 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
1800 if (!LoopBlockSet.count(*HeaderChain.begin()))
1803 BlockFrequency BestExitEdgeFreq;
1804 unsigned BestExitLoopDepth = 0;
1805 MachineBasicBlock *ExitingBB = nullptr;
1806 // If there are exits to outer loops, loop rotation can severely limit
1807 // fallthrough opportunities unless it selects such an exit. Keep a set of
1808 // blocks where rotating to exit with that block will reach an outer loop.
1809 SmallPtrSet<MachineBasicBlock *, 4> BlocksExitingToOuterLoop;
1811 DEBUG(dbgs() << "Finding best loop exit for: " << getBlockName(L.getHeader())
1813 for (MachineBasicBlock *MBB : L.getBlocks()) {
1814 BlockChain &Chain = *BlockToChain[MBB];
1815 // Ensure that this block is at the end of a chain; otherwise it could be
1816 // mid-way through an inner loop or a successor of an unanalyzable branch.
1817 if (MBB != *std::prev(Chain.end()))
1820 // Now walk the successors. We need to establish whether this has a viable
1821 // exiting successor and whether it has a viable non-exiting successor.
1822 // We store the old exiting state and restore it if a viable looping
1823 // successor isn't found.
1824 MachineBasicBlock *OldExitingBB = ExitingBB;
1825 BlockFrequency OldBestExitEdgeFreq = BestExitEdgeFreq;
1826 bool HasLoopingSucc = false;
1827 for (MachineBasicBlock *Succ : MBB->successors()) {
1828 if (Succ->isEHPad())
1832 BlockChain &SuccChain = *BlockToChain[Succ];
1833 // Don't split chains, either this chain or the successor's chain.
1834 if (&Chain == &SuccChain) {
1835 DEBUG(dbgs() << " exiting: " << getBlockName(MBB) << " -> "
1836 << getBlockName(Succ) << " (chain conflict)\n");
1840 auto SuccProb = MBPI->getEdgeProbability(MBB, Succ);
1841 if (LoopBlockSet.count(Succ)) {
1842 DEBUG(dbgs() << " looping: " << getBlockName(MBB) << " -> "
1843 << getBlockName(Succ) << " (" << SuccProb << ")\n");
1844 HasLoopingSucc = true;
1848 unsigned SuccLoopDepth = 0;
1849 if (MachineLoop *ExitLoop = MLI->getLoopFor(Succ)) {
1850 SuccLoopDepth = ExitLoop->getLoopDepth();
1851 if (ExitLoop->contains(&L))
1852 BlocksExitingToOuterLoop.insert(MBB);
1855 BlockFrequency ExitEdgeFreq = MBFI->getBlockFreq(MBB) * SuccProb;
1856 DEBUG(dbgs() << " exiting: " << getBlockName(MBB) << " -> "
1857 << getBlockName(Succ) << " [L:" << SuccLoopDepth << "] (";
1858 MBFI->printBlockFreq(dbgs(), ExitEdgeFreq) << ")\n");
1859 // Note that we bias this toward an existing layout successor to retain
1860 // incoming order in the absence of better information. The exit must have
1861 // a frequency higher than the current exit before we consider breaking
1863 BranchProbability Bias(100 - ExitBlockBias, 100);
1864 if (!ExitingBB || SuccLoopDepth > BestExitLoopDepth ||
1865 ExitEdgeFreq > BestExitEdgeFreq ||
1866 (MBB->isLayoutSuccessor(Succ) &&
1867 !(ExitEdgeFreq < BestExitEdgeFreq * Bias))) {
1868 BestExitEdgeFreq = ExitEdgeFreq;
1873 if (!HasLoopingSucc) {
1874 // Restore the old exiting state, no viable looping successor was found.
1875 ExitingBB = OldExitingBB;
1876 BestExitEdgeFreq = OldBestExitEdgeFreq;
1879 // Without a candidate exiting block or with only a single block in the
1880 // loop, just use the loop header to layout the loop.
1882 DEBUG(dbgs() << " No other candidate exit blocks, using loop header\n");
1885 if (L.getNumBlocks() == 1) {
1886 DEBUG(dbgs() << " Loop has 1 block, using loop header as exit\n");
1890 // Also, if we have exit blocks which lead to outer loops but didn't select
1891 // one of them as the exiting block we are rotating toward, disable loop
1892 // rotation altogether.
1893 if (!BlocksExitingToOuterLoop.empty() &&
1894 !BlocksExitingToOuterLoop.count(ExitingBB))
1897 DEBUG(dbgs() << " Best exiting block: " << getBlockName(ExitingBB) << "\n");
1901 /// \brief Attempt to rotate an exiting block to the bottom of the loop.
1903 /// Once we have built a chain, try to rotate it to line up the hot exit block
1904 /// with fallthrough out of the loop if doing so doesn't introduce unnecessary
1905 /// branches. For example, if the loop has fallthrough into its header and out
1906 /// of its bottom already, don't rotate it.
1907 void MachineBlockPlacement::rotateLoop(BlockChain &LoopChain,
1908 const MachineBasicBlock *ExitingBB,
1909 const BlockFilterSet &LoopBlockSet) {
1913 MachineBasicBlock *Top = *LoopChain.begin();
1914 bool ViableTopFallthrough = false;
1915 for (MachineBasicBlock *Pred : Top->predecessors()) {
1916 BlockChain *PredChain = BlockToChain[Pred];
1917 if (!LoopBlockSet.count(Pred) &&
1918 (!PredChain || Pred == *std::prev(PredChain->end()))) {
1919 ViableTopFallthrough = true;
1924 // If the header has viable fallthrough, check whether the current loop
1925 // bottom is a viable exiting block. If so, bail out as rotating will
1926 // introduce an unnecessary branch.
1927 if (ViableTopFallthrough) {
1928 MachineBasicBlock *Bottom = *std::prev(LoopChain.end());
1929 for (MachineBasicBlock *Succ : Bottom->successors()) {
1930 BlockChain *SuccChain = BlockToChain[Succ];
1931 if (!LoopBlockSet.count(Succ) &&
1932 (!SuccChain || Succ == *SuccChain->begin()))
1937 BlockChain::iterator ExitIt = find(LoopChain, ExitingBB);
1938 if (ExitIt == LoopChain.end())
1941 std::rotate(LoopChain.begin(), std::next(ExitIt), LoopChain.end());
1944 /// \brief Attempt to rotate a loop based on profile data to reduce branch cost.
1946 /// With profile data, we can determine the cost in terms of missed fall through
1947 /// opportunities when rotating a loop chain and select the best rotation.
1948 /// Basically, there are three kinds of cost to consider for each rotation:
1949 /// 1. The possibly missed fall through edge (if it exists) from BB out of
1950 /// the loop to the loop header.
1951 /// 2. The possibly missed fall through edges (if they exist) from the loop
1952 /// exits to BB out of the loop.
1953 /// 3. The missed fall through edge (if it exists) from the last BB to the
1954 /// first BB in the loop chain.
1955 /// Therefore, the cost for a given rotation is the sum of costs listed above.
1956 /// We select the best rotation with the smallest cost.
1957 void MachineBlockPlacement::rotateLoopWithProfile(
1958 BlockChain &LoopChain, const MachineLoop &L,
1959 const BlockFilterSet &LoopBlockSet) {
1960 auto HeaderBB = L.getHeader();
1961 auto HeaderIter = find(LoopChain, HeaderBB);
1962 auto RotationPos = LoopChain.end();
1964 BlockFrequency SmallestRotationCost = BlockFrequency::getMaxFrequency();
1966 // A utility lambda that scales up a block frequency by dividing it by a
1967 // branch probability which is the reciprocal of the scale.
1968 auto ScaleBlockFrequency = [](BlockFrequency Freq,
1969 unsigned Scale) -> BlockFrequency {
1972 // Use operator / between BlockFrequency and BranchProbability to implement
1973 // saturating multiplication.
1974 return Freq / BranchProbability(1, Scale);
1977 // Compute the cost of the missed fall-through edge to the loop header if the
1978 // chain head is not the loop header. As we only consider natural loops with
1979 // single header, this computation can be done only once.
1980 BlockFrequency HeaderFallThroughCost(0);
1981 for (auto *Pred : HeaderBB->predecessors()) {
1982 BlockChain *PredChain = BlockToChain[Pred];
1983 if (!LoopBlockSet.count(Pred) &&
1984 (!PredChain || Pred == *std::prev(PredChain->end()))) {
1986 MBFI->getBlockFreq(Pred) * MBPI->getEdgeProbability(Pred, HeaderBB);
1987 auto FallThruCost = ScaleBlockFrequency(EdgeFreq, MisfetchCost);
1988 // If the predecessor has only an unconditional jump to the header, we
1989 // need to consider the cost of this jump.
1990 if (Pred->succ_size() == 1)
1991 FallThruCost += ScaleBlockFrequency(EdgeFreq, JumpInstCost);
1992 HeaderFallThroughCost = std::max(HeaderFallThroughCost, FallThruCost);
1996 // Here we collect all exit blocks in the loop, and for each exit we find out
1997 // its hottest exit edge. For each loop rotation, we define the loop exit cost
1998 // as the sum of frequencies of exit edges we collect here, excluding the exit
1999 // edge from the tail of the loop chain.
2000 SmallVector<std::pair<MachineBasicBlock *, BlockFrequency>, 4> ExitsWithFreq;
2001 for (auto BB : LoopChain) {
2002 auto LargestExitEdgeProb = BranchProbability::getZero();
2003 for (auto *Succ : BB->successors()) {
2004 BlockChain *SuccChain = BlockToChain[Succ];
2005 if (!LoopBlockSet.count(Succ) &&
2006 (!SuccChain || Succ == *SuccChain->begin())) {
2007 auto SuccProb = MBPI->getEdgeProbability(BB, Succ);
2008 LargestExitEdgeProb = std::max(LargestExitEdgeProb, SuccProb);
2011 if (LargestExitEdgeProb > BranchProbability::getZero()) {
2012 auto ExitFreq = MBFI->getBlockFreq(BB) * LargestExitEdgeProb;
2013 ExitsWithFreq.emplace_back(BB, ExitFreq);
2017 // In this loop we iterate every block in the loop chain and calculate the
2018 // cost assuming the block is the head of the loop chain. When the loop ends,
2019 // we should have found the best candidate as the loop chain's head.
2020 for (auto Iter = LoopChain.begin(), TailIter = std::prev(LoopChain.end()),
2021 EndIter = LoopChain.end();
2022 Iter != EndIter; Iter++, TailIter++) {
2023 // TailIter is used to track the tail of the loop chain if the block we are
2024 // checking (pointed by Iter) is the head of the chain.
2025 if (TailIter == LoopChain.end())
2026 TailIter = LoopChain.begin();
2028 auto TailBB = *TailIter;
2030 // Calculate the cost by putting this BB to the top.
2031 BlockFrequency Cost = 0;
2033 // If the current BB is the loop header, we need to take into account the
2034 // cost of the missed fall through edge from outside of the loop to the
2036 if (Iter != HeaderIter)
2037 Cost += HeaderFallThroughCost;
2039 // Collect the loop exit cost by summing up frequencies of all exit edges
2040 // except the one from the chain tail.
2041 for (auto &ExitWithFreq : ExitsWithFreq)
2042 if (TailBB != ExitWithFreq.first)
2043 Cost += ExitWithFreq.second;
2045 // The cost of breaking the once fall-through edge from the tail to the top
2046 // of the loop chain. Here we need to consider three cases:
2047 // 1. If the tail node has only one successor, then we will get an
2048 // additional jmp instruction. So the cost here is (MisfetchCost +
2049 // JumpInstCost) * tail node frequency.
2050 // 2. If the tail node has two successors, then we may still get an
2051 // additional jmp instruction if the layout successor after the loop
2052 // chain is not its CFG successor. Note that the more frequently executed
2053 // jmp instruction will be put ahead of the other one. Assume the
2054 // frequency of those two branches are x and y, where x is the frequency
2055 // of the edge to the chain head, then the cost will be
2056 // (x * MisfetechCost + min(x, y) * JumpInstCost) * tail node frequency.
2057 // 3. If the tail node has more than two successors (this rarely happens),
2058 // we won't consider any additional cost.
2059 if (TailBB->isSuccessor(*Iter)) {
2060 auto TailBBFreq = MBFI->getBlockFreq(TailBB);
2061 if (TailBB->succ_size() == 1)
2062 Cost += ScaleBlockFrequency(TailBBFreq.getFrequency(),
2063 MisfetchCost + JumpInstCost);
2064 else if (TailBB->succ_size() == 2) {
2065 auto TailToHeadProb = MBPI->getEdgeProbability(TailBB, *Iter);
2066 auto TailToHeadFreq = TailBBFreq * TailToHeadProb;
2067 auto ColderEdgeFreq = TailToHeadProb > BranchProbability(1, 2)
2068 ? TailBBFreq * TailToHeadProb.getCompl()
2070 Cost += ScaleBlockFrequency(TailToHeadFreq, MisfetchCost) +
2071 ScaleBlockFrequency(ColderEdgeFreq, JumpInstCost);
2075 DEBUG(dbgs() << "The cost of loop rotation by making " << getBlockName(*Iter)
2076 << " to the top: " << Cost.getFrequency() << "\n");
2078 if (Cost < SmallestRotationCost) {
2079 SmallestRotationCost = Cost;
2084 if (RotationPos != LoopChain.end()) {
2085 DEBUG(dbgs() << "Rotate loop by making " << getBlockName(*RotationPos)
2086 << " to the top\n");
2087 std::rotate(LoopChain.begin(), RotationPos, LoopChain.end());
2091 /// \brief Collect blocks in the given loop that are to be placed.
2093 /// When profile data is available, exclude cold blocks from the returned set;
2094 /// otherwise, collect all blocks in the loop.
2095 MachineBlockPlacement::BlockFilterSet
2096 MachineBlockPlacement::collectLoopBlockSet(const MachineLoop &L) {
2097 BlockFilterSet LoopBlockSet;
2099 // Filter cold blocks off from LoopBlockSet when profile data is available.
2100 // Collect the sum of frequencies of incoming edges to the loop header from
2101 // outside. If we treat the loop as a super block, this is the frequency of
2102 // the loop. Then for each block in the loop, we calculate the ratio between
2103 // its frequency and the frequency of the loop block. When it is too small,
2104 // don't add it to the loop chain. If there are outer loops, then this block
2105 // will be merged into the first outer loop chain for which this block is not
2106 // cold anymore. This needs precise profile data and we only do this when
2107 // profile data is available.
2108 if (F->getFunction()->getEntryCount()) {
2109 BlockFrequency LoopFreq(0);
2110 for (auto LoopPred : L.getHeader()->predecessors())
2111 if (!L.contains(LoopPred))
2112 LoopFreq += MBFI->getBlockFreq(LoopPred) *
2113 MBPI->getEdgeProbability(LoopPred, L.getHeader());
2115 for (MachineBasicBlock *LoopBB : L.getBlocks()) {
2116 auto Freq = MBFI->getBlockFreq(LoopBB).getFrequency();
2117 if (Freq == 0 || LoopFreq.getFrequency() / Freq > LoopToColdBlockRatio)
2119 LoopBlockSet.insert(LoopBB);
2122 LoopBlockSet.insert(L.block_begin(), L.block_end());
2124 return LoopBlockSet;
2127 /// \brief Forms basic block chains from the natural loop structures.
2129 /// These chains are designed to preserve the existing *structure* of the code
2130 /// as much as possible. We can then stitch the chains together in a way which
2131 /// both preserves the topological structure and minimizes taken conditional
2133 void MachineBlockPlacement::buildLoopChains(const MachineLoop &L) {
2134 // First recurse through any nested loops, building chains for those inner
2136 for (const MachineLoop *InnerLoop : L)
2137 buildLoopChains(*InnerLoop);
2139 assert(BlockWorkList.empty());
2140 assert(EHPadWorkList.empty());
2141 BlockFilterSet LoopBlockSet = collectLoopBlockSet(L);
2143 // Check if we have profile data for this function. If yes, we will rotate
2144 // this loop by modeling costs more precisely which requires the profile data
2145 // for better layout.
2146 bool RotateLoopWithProfile =
2147 ForcePreciseRotationCost ||
2148 (PreciseRotationCost && F->getFunction()->getEntryCount());
2150 // First check to see if there is an obviously preferable top block for the
2151 // loop. This will default to the header, but may end up as one of the
2152 // predecessors to the header if there is one which will result in strictly
2153 // fewer branches in the loop body.
2154 // When we use profile data to rotate the loop, this is unnecessary.
2155 MachineBasicBlock *LoopTop =
2156 RotateLoopWithProfile ? L.getHeader() : findBestLoopTop(L, LoopBlockSet);
2158 // If we selected just the header for the loop top, look for a potentially
2159 // profitable exit block in the event that rotating the loop can eliminate
2160 // branches by placing an exit edge at the bottom.
2161 if (!RotateLoopWithProfile && LoopTop == L.getHeader())
2162 PreferredLoopExit = findBestLoopExit(L, LoopBlockSet);
2164 BlockChain &LoopChain = *BlockToChain[LoopTop];
2166 // FIXME: This is a really lame way of walking the chains in the loop: we
2167 // walk the blocks, and use a set to prevent visiting a particular chain
2169 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
2170 assert(LoopChain.UnscheduledPredecessors == 0);
2171 UpdatedPreds.insert(&LoopChain);
2173 for (const MachineBasicBlock *LoopBB : LoopBlockSet)
2174 fillWorkLists(LoopBB, UpdatedPreds, &LoopBlockSet);
2176 buildChain(LoopTop, LoopChain, &LoopBlockSet);
2178 if (RotateLoopWithProfile)
2179 rotateLoopWithProfile(LoopChain, L, LoopBlockSet);
2181 rotateLoop(LoopChain, PreferredLoopExit, LoopBlockSet);
2184 // Crash at the end so we get all of the debugging output first.
2185 bool BadLoop = false;
2186 if (LoopChain.UnscheduledPredecessors) {
2188 dbgs() << "Loop chain contains a block without its preds placed!\n"
2189 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
2190 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n";
2192 for (MachineBasicBlock *ChainBB : LoopChain) {
2193 dbgs() << " ... " << getBlockName(ChainBB) << "\n";
2194 if (!LoopBlockSet.remove(ChainBB)) {
2195 // We don't mark the loop as bad here because there are real situations
2196 // where this can occur. For example, with an unanalyzable fallthrough
2197 // from a loop block to a non-loop block or vice versa.
2198 dbgs() << "Loop chain contains a block not contained by the loop!\n"
2199 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
2200 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
2201 << " Bad block: " << getBlockName(ChainBB) << "\n";
2205 if (!LoopBlockSet.empty()) {
2207 for (const MachineBasicBlock *LoopBB : LoopBlockSet)
2208 dbgs() << "Loop contains blocks never placed into a chain!\n"
2209 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
2210 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
2211 << " Bad block: " << getBlockName(LoopBB) << "\n";
2213 assert(!BadLoop && "Detected problems with the placement of this loop.");
2216 BlockWorkList.clear();
2217 EHPadWorkList.clear();
2220 void MachineBlockPlacement::buildCFGChains() {
2221 // Ensure that every BB in the function has an associated chain to simplify
2222 // the assumptions of the remaining algorithm.
2223 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
2224 for (MachineFunction::iterator FI = F->begin(), FE = F->end(); FI != FE;
2226 MachineBasicBlock *BB = &*FI;
2228 new (ChainAllocator.Allocate()) BlockChain(BlockToChain, BB);
2229 // Also, merge any blocks which we cannot reason about and must preserve
2230 // the exact fallthrough behavior for.
2233 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
2234 if (!TII->analyzeBranch(*BB, TBB, FBB, Cond) || !FI->canFallThrough())
2237 MachineFunction::iterator NextFI = std::next(FI);
2238 MachineBasicBlock *NextBB = &*NextFI;
2239 // Ensure that the layout successor is a viable block, as we know that
2240 // fallthrough is a possibility.
2241 assert(NextFI != FE && "Can't fallthrough past the last block.");
2242 DEBUG(dbgs() << "Pre-merging due to unanalyzable fallthrough: "
2243 << getBlockName(BB) << " -> " << getBlockName(NextBB)
2245 Chain->merge(NextBB, nullptr);
2247 BlocksWithUnanalyzableExits.insert(&*BB);
2254 // Build any loop-based chains.
2255 PreferredLoopExit = nullptr;
2256 for (MachineLoop *L : *MLI)
2257 buildLoopChains(*L);
2259 assert(BlockWorkList.empty());
2260 assert(EHPadWorkList.empty());
2262 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
2263 for (MachineBasicBlock &MBB : *F)
2264 fillWorkLists(&MBB, UpdatedPreds);
2266 BlockChain &FunctionChain = *BlockToChain[&F->front()];
2267 buildChain(&F->front(), FunctionChain);
2270 typedef SmallPtrSet<MachineBasicBlock *, 16> FunctionBlockSetType;
2273 // Crash at the end so we get all of the debugging output first.
2274 bool BadFunc = false;
2275 FunctionBlockSetType FunctionBlockSet;
2276 for (MachineBasicBlock &MBB : *F)
2277 FunctionBlockSet.insert(&MBB);
2279 for (MachineBasicBlock *ChainBB : FunctionChain)
2280 if (!FunctionBlockSet.erase(ChainBB)) {
2282 dbgs() << "Function chain contains a block not in the function!\n"
2283 << " Bad block: " << getBlockName(ChainBB) << "\n";
2286 if (!FunctionBlockSet.empty()) {
2288 for (MachineBasicBlock *RemainingBB : FunctionBlockSet)
2289 dbgs() << "Function contains blocks never placed into a chain!\n"
2290 << " Bad block: " << getBlockName(RemainingBB) << "\n";
2292 assert(!BadFunc && "Detected problems with the block placement.");
2295 // Splice the blocks into place.
2296 MachineFunction::iterator InsertPos = F->begin();
2297 DEBUG(dbgs() << "[MBP] Function: "<< F->getName() << "\n");
2298 for (MachineBasicBlock *ChainBB : FunctionChain) {
2299 DEBUG(dbgs() << (ChainBB == *FunctionChain.begin() ? "Placing chain "
2301 << getBlockName(ChainBB) << "\n");
2302 if (InsertPos != MachineFunction::iterator(ChainBB))
2303 F->splice(InsertPos, ChainBB);
2307 // Update the terminator of the previous block.
2308 if (ChainBB == *FunctionChain.begin())
2310 MachineBasicBlock *PrevBB = &*std::prev(MachineFunction::iterator(ChainBB));
2312 // FIXME: It would be awesome of updateTerminator would just return rather
2313 // than assert when the branch cannot be analyzed in order to remove this
2316 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
2319 if (!BlocksWithUnanalyzableExits.count(PrevBB)) {
2320 // Given the exact block placement we chose, we may actually not _need_ to
2321 // be able to edit PrevBB's terminator sequence, but not being _able_ to
2322 // do that at this point is a bug.
2323 assert((!TII->analyzeBranch(*PrevBB, TBB, FBB, Cond) ||
2324 !PrevBB->canFallThrough()) &&
2325 "Unexpected block with un-analyzable fallthrough!");
2327 TBB = FBB = nullptr;
2331 // The "PrevBB" is not yet updated to reflect current code layout, so,
2332 // o. it may fall-through to a block without explicit "goto" instruction
2333 // before layout, and no longer fall-through it after layout; or
2334 // o. just opposite.
2336 // analyzeBranch() may return erroneous value for FBB when these two
2337 // situations take place. For the first scenario FBB is mistakenly set NULL;
2338 // for the 2nd scenario, the FBB, which is expected to be NULL, is
2339 // mistakenly pointing to "*BI".
2340 // Thus, if the future change needs to use FBB before the layout is set, it
2341 // has to correct FBB first by using the code similar to the following:
2343 // if (!Cond.empty() && (!FBB || FBB == ChainBB)) {
2344 // PrevBB->updateTerminator();
2346 // TBB = FBB = nullptr;
2347 // if (TII->analyzeBranch(*PrevBB, TBB, FBB, Cond)) {
2348 // // FIXME: This should never take place.
2349 // TBB = FBB = nullptr;
2352 if (!TII->analyzeBranch(*PrevBB, TBB, FBB, Cond))
2353 PrevBB->updateTerminator();
2356 // Fixup the last block.
2358 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
2359 if (!TII->analyzeBranch(F->back(), TBB, FBB, Cond))
2360 F->back().updateTerminator();
2362 BlockWorkList.clear();
2363 EHPadWorkList.clear();
2366 void MachineBlockPlacement::optimizeBranches() {
2367 BlockChain &FunctionChain = *BlockToChain[&F->front()];
2368 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
2370 // Now that all the basic blocks in the chain have the proper layout,
2371 // make a final call to AnalyzeBranch with AllowModify set.
2372 // Indeed, the target may be able to optimize the branches in a way we
2373 // cannot because all branches may not be analyzable.
2374 // E.g., the target may be able to remove an unconditional branch to
2375 // a fallthrough when it occurs after predicated terminators.
2376 for (MachineBasicBlock *ChainBB : FunctionChain) {
2378 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
2379 if (!TII->analyzeBranch(*ChainBB, TBB, FBB, Cond, /*AllowModify*/ true)) {
2380 // If PrevBB has a two-way branch, try to re-order the branches
2381 // such that we branch to the successor with higher probability first.
2382 if (TBB && !Cond.empty() && FBB &&
2383 MBPI->getEdgeProbability(ChainBB, FBB) >
2384 MBPI->getEdgeProbability(ChainBB, TBB) &&
2385 !TII->reverseBranchCondition(Cond)) {
2386 DEBUG(dbgs() << "Reverse order of the two branches: "
2387 << getBlockName(ChainBB) << "\n");
2388 DEBUG(dbgs() << " Edge probability: "
2389 << MBPI->getEdgeProbability(ChainBB, FBB) << " vs "
2390 << MBPI->getEdgeProbability(ChainBB, TBB) << "\n");
2391 DebugLoc dl; // FIXME: this is nowhere
2392 TII->removeBranch(*ChainBB);
2393 TII->insertBranch(*ChainBB, FBB, TBB, Cond, dl);
2394 ChainBB->updateTerminator();
2400 void MachineBlockPlacement::alignBlocks() {
2401 // Walk through the backedges of the function now that we have fully laid out
2402 // the basic blocks and align the destination of each backedge. We don't rely
2403 // exclusively on the loop info here so that we can align backedges in
2404 // unnatural CFGs and backedges that were introduced purely because of the
2405 // loop rotations done during this layout pass.
2406 if (F->getFunction()->optForSize())
2408 BlockChain &FunctionChain = *BlockToChain[&F->front()];
2409 if (FunctionChain.begin() == FunctionChain.end())
2410 return; // Empty chain.
2412 const BranchProbability ColdProb(1, 5); // 20%
2413 BlockFrequency EntryFreq = MBFI->getBlockFreq(&F->front());
2414 BlockFrequency WeightedEntryFreq = EntryFreq * ColdProb;
2415 for (MachineBasicBlock *ChainBB : FunctionChain) {
2416 if (ChainBB == *FunctionChain.begin())
2419 // Don't align non-looping basic blocks. These are unlikely to execute
2420 // enough times to matter in practice. Note that we'll still handle
2421 // unnatural CFGs inside of a natural outer loop (the common case) and
2423 MachineLoop *L = MLI->getLoopFor(ChainBB);
2427 unsigned Align = TLI->getPrefLoopAlignment(L);
2429 continue; // Don't care about loop alignment.
2431 // If the block is cold relative to the function entry don't waste space
2433 BlockFrequency Freq = MBFI->getBlockFreq(ChainBB);
2434 if (Freq < WeightedEntryFreq)
2437 // If the block is cold relative to its loop header, don't align it
2438 // regardless of what edges into the block exist.
2439 MachineBasicBlock *LoopHeader = L->getHeader();
2440 BlockFrequency LoopHeaderFreq = MBFI->getBlockFreq(LoopHeader);
2441 if (Freq < (LoopHeaderFreq * ColdProb))
2444 // Check for the existence of a non-layout predecessor which would benefit
2445 // from aligning this block.
2446 MachineBasicBlock *LayoutPred =
2447 &*std::prev(MachineFunction::iterator(ChainBB));
2449 // Force alignment if all the predecessors are jumps. We already checked
2450 // that the block isn't cold above.
2451 if (!LayoutPred->isSuccessor(ChainBB)) {
2452 ChainBB->setAlignment(Align);
2456 // Align this block if the layout predecessor's edge into this block is
2457 // cold relative to the block. When this is true, other predecessors make up
2458 // all of the hot entries into the block and thus alignment is likely to be
2460 BranchProbability LayoutProb =
2461 MBPI->getEdgeProbability(LayoutPred, ChainBB);
2462 BlockFrequency LayoutEdgeFreq = MBFI->getBlockFreq(LayoutPred) * LayoutProb;
2463 if (LayoutEdgeFreq <= (Freq * ColdProb))
2464 ChainBB->setAlignment(Align);
2468 /// Tail duplicate \p BB into (some) predecessors if profitable, repeating if
2469 /// it was duplicated into its chain predecessor and removed.
2470 /// \p BB - Basic block that may be duplicated.
2472 /// \p LPred - Chosen layout predecessor of \p BB.
2473 /// Updated to be the chain end if LPred is removed.
2474 /// \p Chain - Chain to which \p LPred belongs, and \p BB will belong.
2475 /// \p BlockFilter - Set of blocks that belong to the loop being laid out.
2476 /// Used to identify which blocks to update predecessor
2478 /// \p PrevUnplacedBlockIt - Iterator pointing to the last block that was
2479 /// chosen in the given order due to unnatural CFG
2480 /// only needed if \p BB is removed and
2481 /// \p PrevUnplacedBlockIt pointed to \p BB.
2482 /// @return true if \p BB was removed.
2483 bool MachineBlockPlacement::repeatedlyTailDuplicateBlock(
2484 MachineBasicBlock *BB, MachineBasicBlock *&LPred,
2485 const MachineBasicBlock *LoopHeaderBB,
2486 BlockChain &Chain, BlockFilterSet *BlockFilter,
2487 MachineFunction::iterator &PrevUnplacedBlockIt) {
2488 bool Removed, DuplicatedToLPred;
2489 bool DuplicatedToOriginalLPred;
2490 Removed = maybeTailDuplicateBlock(BB, LPred, Chain, BlockFilter,
2491 PrevUnplacedBlockIt,
2495 DuplicatedToOriginalLPred = DuplicatedToLPred;
2496 // Iteratively try to duplicate again. It can happen that a block that is
2497 // duplicated into is still small enough to be duplicated again.
2498 // No need to call markBlockSuccessors in this case, as the blocks being
2499 // duplicated from here on are already scheduled.
2500 // Note that DuplicatedToLPred always implies Removed.
2501 while (DuplicatedToLPred) {
2502 assert (Removed && "Block must have been removed to be duplicated into its "
2503 "layout predecessor.");
2504 MachineBasicBlock *DupBB, *DupPred;
2505 // The removal callback causes Chain.end() to be updated when a block is
2506 // removed. On the first pass through the loop, the chain end should be the
2507 // same as it was on function entry. On subsequent passes, because we are
2508 // duplicating the block at the end of the chain, if it is removed the
2509 // chain will have shrunk by one block.
2510 BlockChain::iterator ChainEnd = Chain.end();
2511 DupBB = *(--ChainEnd);
2512 // Now try to duplicate again.
2513 if (ChainEnd == Chain.begin())
2515 DupPred = *std::prev(ChainEnd);
2516 Removed = maybeTailDuplicateBlock(DupBB, DupPred, Chain, BlockFilter,
2517 PrevUnplacedBlockIt,
2520 // If BB was duplicated into LPred, it is now scheduled. But because it was
2521 // removed, markChainSuccessors won't be called for its chain. Instead we
2522 // call markBlockSuccessors for LPred to achieve the same effect. This must go
2523 // at the end because repeating the tail duplication can increase the number
2524 // of unscheduled predecessors.
2525 LPred = *std::prev(Chain.end());
2526 if (DuplicatedToOriginalLPred)
2527 markBlockSuccessors(Chain, LPred, LoopHeaderBB, BlockFilter);
2531 /// Tail duplicate \p BB into (some) predecessors if profitable.
2532 /// \p BB - Basic block that may be duplicated
2533 /// \p LPred - Chosen layout predecessor of \p BB
2534 /// \p Chain - Chain to which \p LPred belongs, and \p BB will belong.
2535 /// \p BlockFilter - Set of blocks that belong to the loop being laid out.
2536 /// Used to identify which blocks to update predecessor
2538 /// \p PrevUnplacedBlockIt - Iterator pointing to the last block that was
2539 /// chosen in the given order due to unnatural CFG
2540 /// only needed if \p BB is removed and
2541 /// \p PrevUnplacedBlockIt pointed to \p BB.
2542 /// \p DuplicatedToLPred - True if the block was duplicated into LPred. Will
2543 /// only be true if the block was removed.
2544 /// \return - True if the block was duplicated into all preds and removed.
2545 bool MachineBlockPlacement::maybeTailDuplicateBlock(
2546 MachineBasicBlock *BB, MachineBasicBlock *LPred,
2547 BlockChain &Chain, BlockFilterSet *BlockFilter,
2548 MachineFunction::iterator &PrevUnplacedBlockIt,
2549 bool &DuplicatedToLPred) {
2550 DuplicatedToLPred = false;
2551 if (!shouldTailDuplicate(BB))
2554 DEBUG(dbgs() << "Redoing tail duplication for Succ#"
2555 << BB->getNumber() << "\n");
2557 // This has to be a callback because none of it can be done after
2559 bool Removed = false;
2560 auto RemovalCallback =
2561 [&](MachineBasicBlock *RemBB) {
2562 // Signal to outer function
2565 // Conservative default.
2566 bool InWorkList = true;
2567 // Remove from the Chain and Chain Map
2568 if (BlockToChain.count(RemBB)) {
2569 BlockChain *Chain = BlockToChain[RemBB];
2570 InWorkList = Chain->UnscheduledPredecessors == 0;
2571 Chain->remove(RemBB);
2572 BlockToChain.erase(RemBB);
2575 // Handle the unplaced block iterator
2576 if (&(*PrevUnplacedBlockIt) == RemBB) {
2577 PrevUnplacedBlockIt++;
2580 // Handle the Work Lists
2582 SmallVectorImpl<MachineBasicBlock *> &RemoveList = BlockWorkList;
2583 if (RemBB->isEHPad())
2584 RemoveList = EHPadWorkList;
2586 remove_if(RemoveList,
2587 [RemBB](MachineBasicBlock *BB) {return BB == RemBB;}),
2591 // Handle the filter set
2593 BlockFilter->remove(RemBB);
2596 // Remove the block from loop info.
2597 MLI->removeBlock(RemBB);
2598 if (RemBB == PreferredLoopExit)
2599 PreferredLoopExit = nullptr;
2601 DEBUG(dbgs() << "TailDuplicator deleted block: "
2602 << getBlockName(RemBB) << "\n");
2604 auto RemovalCallbackRef =
2605 llvm::function_ref<void(MachineBasicBlock*)>(RemovalCallback);
2607 SmallVector<MachineBasicBlock *, 8> DuplicatedPreds;
2608 bool IsSimple = TailDup.isSimpleBB(BB);
2609 TailDup.tailDuplicateAndUpdate(IsSimple, BB, LPred,
2610 &DuplicatedPreds, &RemovalCallbackRef);
2612 // Update UnscheduledPredecessors to reflect tail-duplication.
2613 DuplicatedToLPred = false;
2614 for (MachineBasicBlock *Pred : DuplicatedPreds) {
2615 // We're only looking for unscheduled predecessors that match the filter.
2616 BlockChain* PredChain = BlockToChain[Pred];
2618 DuplicatedToLPred = true;
2619 if (Pred == LPred || (BlockFilter && !BlockFilter->count(Pred))
2620 || PredChain == &Chain)
2622 for (MachineBasicBlock *NewSucc : Pred->successors()) {
2623 if (BlockFilter && !BlockFilter->count(NewSucc))
2625 BlockChain *NewChain = BlockToChain[NewSucc];
2626 if (NewChain != &Chain && NewChain != PredChain)
2627 NewChain->UnscheduledPredecessors++;
2633 bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &MF) {
2634 if (skipFunction(*MF.getFunction()))
2637 // Check for single-block functions and skip them.
2638 if (std::next(MF.begin()) == MF.end())
2642 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
2643 MBFI = llvm::make_unique<BranchFolder::MBFIWrapper>(
2644 getAnalysis<MachineBlockFrequencyInfo>());
2645 MLI = &getAnalysis<MachineLoopInfo>();
2646 TII = MF.getSubtarget().getInstrInfo();
2647 TLI = MF.getSubtarget().getTargetLowering();
2650 // Initialize PreferredLoopExit to nullptr here since it may never be set if
2651 // there are no MachineLoops.
2652 PreferredLoopExit = nullptr;
2654 assert(BlockToChain.empty());
2655 assert(ComputedEdges.empty());
2657 unsigned TailDupSize = TailDupPlacementThreshold;
2658 // If only the aggressive threshold is explicitly set, use it.
2659 if (TailDupPlacementAggressiveThreshold.getNumOccurrences() != 0 &&
2660 TailDupPlacementThreshold.getNumOccurrences() == 0)
2661 TailDupSize = TailDupPlacementAggressiveThreshold;
2663 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
2664 // For agressive optimization, we can adjust some thresholds to be less
2666 if (PassConfig->getOptLevel() >= CodeGenOpt::Aggressive) {
2667 // At O3 we should be more willing to copy blocks for tail duplication. This
2668 // increases size pressure, so we only do it at O3
2669 // Do this unless only the regular threshold is explicitly set.
2670 if (TailDupPlacementThreshold.getNumOccurrences() == 0 ||
2671 TailDupPlacementAggressiveThreshold.getNumOccurrences() != 0)
2672 TailDupSize = TailDupPlacementAggressiveThreshold;
2675 if (TailDupPlacement) {
2676 MPDT = &getAnalysis<MachinePostDominatorTree>();
2677 if (MF.getFunction()->optForSize())
2679 TailDup.initMF(MF, MBPI, /* LayoutMode */ true, TailDupSize);
2680 precomputeTriangleChains();
2685 // Changing the layout can create new tail merging opportunities.
2686 // TailMerge can create jump into if branches that make CFG irreducible for
2687 // HW that requires structured CFG.
2688 bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
2689 PassConfig->getEnableTailMerge() &&
2690 BranchFoldPlacement;
2691 // No tail merging opportunities if the block number is less than four.
2692 if (MF.size() > 3 && EnableTailMerge) {
2693 unsigned TailMergeSize = TailDupSize + 1;
2694 BranchFolder BF(/*EnableTailMerge=*/true, /*CommonHoist=*/false, *MBFI,
2695 *MBPI, TailMergeSize);
2697 if (BF.OptimizeFunction(MF, TII, MF.getSubtarget().getRegisterInfo(),
2698 getAnalysisIfAvailable<MachineModuleInfo>(), MLI,
2699 /*AfterBlockPlacement=*/true)) {
2700 // Redo the layout if tail merging creates/removes/moves blocks.
2701 BlockToChain.clear();
2702 ComputedEdges.clear();
2703 // Must redo the post-dominator tree if blocks were changed.
2705 MPDT->runOnMachineFunction(MF);
2706 ChainAllocator.DestroyAll();
2714 BlockToChain.clear();
2715 ComputedEdges.clear();
2716 ChainAllocator.DestroyAll();
2719 // Align all of the blocks in the function to a specific alignment.
2720 for (MachineBasicBlock &MBB : MF)
2721 MBB.setAlignment(AlignAllBlock);
2722 else if (AlignAllNonFallThruBlocks) {
2723 // Align all of the blocks that have no fall-through predecessors to a
2724 // specific alignment.
2725 for (auto MBI = std::next(MF.begin()), MBE = MF.end(); MBI != MBE; ++MBI) {
2726 auto LayoutPred = std::prev(MBI);
2727 if (!LayoutPred->isSuccessor(&*MBI))
2728 MBI->setAlignment(AlignAllNonFallThruBlocks);
2731 if (ViewBlockLayoutWithBFI != GVDT_None &&
2732 (ViewBlockFreqFuncName.empty() ||
2733 F->getFunction()->getName().equals(ViewBlockFreqFuncName))) {
2734 MBFI->view("MBP." + MF.getName(), false);
2738 // We always return true as we have no way to track whether the final order
2739 // differs from the original order.
2744 /// \brief A pass to compute block placement statistics.
2746 /// A separate pass to compute interesting statistics for evaluating block
2747 /// placement. This is separate from the actual placement pass so that they can
2748 /// be computed in the absence of any placement transformations or when using
2749 /// alternative placement strategies.
2750 class MachineBlockPlacementStats : public MachineFunctionPass {
2751 /// \brief A handle to the branch probability pass.
2752 const MachineBranchProbabilityInfo *MBPI;
2754 /// \brief A handle to the function-wide block frequency pass.
2755 const MachineBlockFrequencyInfo *MBFI;
2758 static char ID; // Pass identification, replacement for typeid
2759 MachineBlockPlacementStats() : MachineFunctionPass(ID) {
2760 initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry());
2763 bool runOnMachineFunction(MachineFunction &F) override;
2765 void getAnalysisUsage(AnalysisUsage &AU) const override {
2766 AU.addRequired<MachineBranchProbabilityInfo>();
2767 AU.addRequired<MachineBlockFrequencyInfo>();
2768 AU.setPreservesAll();
2769 MachineFunctionPass::getAnalysisUsage(AU);
2774 char MachineBlockPlacementStats::ID = 0;
2775 char &llvm::MachineBlockPlacementStatsID = MachineBlockPlacementStats::ID;
2776 INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats",
2777 "Basic Block Placement Stats", false, false)
2778 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
2779 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
2780 INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats",
2781 "Basic Block Placement Stats", false, false)
2783 bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) {
2784 // Check for single-block functions and skip them.
2785 if (std::next(F.begin()) == F.end())
2788 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
2789 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
2791 for (MachineBasicBlock &MBB : F) {
2792 BlockFrequency BlockFreq = MBFI->getBlockFreq(&MBB);
2793 Statistic &NumBranches =
2794 (MBB.succ_size() > 1) ? NumCondBranches : NumUncondBranches;
2795 Statistic &BranchTakenFreq =
2796 (MBB.succ_size() > 1) ? CondBranchTakenFreq : UncondBranchTakenFreq;
2797 for (MachineBasicBlock *Succ : MBB.successors()) {
2798 // Skip if this successor is a fallthrough.
2799 if (MBB.isLayoutSuccessor(Succ))
2802 BlockFrequency EdgeFreq =
2803 BlockFreq * MBPI->getEdgeProbability(&MBB, Succ);
2805 BranchTakenFreq += EdgeFreq.getFrequency();