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) {
248 assert(BB && "Can't merge a null block.");
249 assert(!Blocks.empty() && "Can't merge into an empty chain.");
251 // Fast path in case we don't have a chain already.
253 assert(!BlockToChain[BB] &&
254 "Passed chain is null, but BB has entry in BlockToChain.");
255 Blocks.push_back(BB);
256 BlockToChain[BB] = this;
260 assert(BB == *Chain->begin() && "Passed BB is not head of Chain.");
261 assert(Chain->begin() != Chain->end());
263 // Update the incoming blocks to point to this chain, and add them to the
265 for (MachineBasicBlock *ChainBB : *Chain) {
266 Blocks.push_back(ChainBB);
267 assert(BlockToChain[ChainBB] == Chain && "Incoming blocks not in chain.");
268 BlockToChain[ChainBB] = this;
273 /// \brief Dump the blocks in this chain.
274 LLVM_DUMP_METHOD void dump() {
275 for (MachineBasicBlock *MBB : *this)
280 /// \brief Count of predecessors of any block within the chain which have not
281 /// yet been scheduled. In general, we will delay scheduling this chain
282 /// until those predecessors are scheduled (or we find a sufficiently good
283 /// reason to override this heuristic.) Note that when forming loop chains,
284 /// blocks outside the loop are ignored and treated as if they were already
287 /// Note: This field is reinitialized multiple times - once for each loop,
288 /// and then once for the function as a whole.
289 unsigned UnscheduledPredecessors;
294 class MachineBlockPlacement : public MachineFunctionPass {
295 /// \brief A typedef for a block filter set.
296 typedef SmallSetVector<const MachineBasicBlock *, 16> BlockFilterSet;
298 /// Pair struct containing basic block and taildup profitiability
299 struct BlockAndTailDupResult {
300 MachineBasicBlock *BB;
304 /// Triple struct containing edge weight and the edge.
305 struct WeightedEdge {
306 BlockFrequency Weight;
307 MachineBasicBlock *Src;
308 MachineBasicBlock *Dest;
311 /// \brief work lists of blocks that are ready to be laid out
312 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
313 SmallVector<MachineBasicBlock *, 16> EHPadWorkList;
315 /// Edges that have already been computed as optimal.
316 DenseMap<const MachineBasicBlock *, BlockAndTailDupResult> ComputedEdges;
318 /// \brief Machine Function
321 /// \brief A handle to the branch probability pass.
322 const MachineBranchProbabilityInfo *MBPI;
324 /// \brief A handle to the function-wide block frequency pass.
325 std::unique_ptr<BranchFolder::MBFIWrapper> MBFI;
327 /// \brief A handle to the loop info.
328 MachineLoopInfo *MLI;
330 /// \brief Preferred loop exit.
331 /// Member variable for convenience. It may be removed by duplication deep
332 /// in the call stack.
333 MachineBasicBlock *PreferredLoopExit;
335 /// \brief A handle to the target's instruction info.
336 const TargetInstrInfo *TII;
338 /// \brief A handle to the target's lowering info.
339 const TargetLoweringBase *TLI;
341 /// \brief A handle to the post dominator tree.
342 MachinePostDominatorTree *MPDT;
344 /// \brief Duplicator used to duplicate tails during placement.
346 /// Placement decisions can open up new tail duplication opportunities, but
347 /// since tail duplication affects placement decisions of later blocks, it
348 /// must be done inline.
349 TailDuplicator TailDup;
351 /// \brief Allocator and owner of BlockChain structures.
353 /// We build BlockChains lazily while processing the loop structure of
354 /// a function. To reduce malloc traffic, we allocate them using this
355 /// slab-like allocator, and destroy them after the pass completes. An
356 /// important guarantee is that this allocator produces stable pointers to
358 SpecificBumpPtrAllocator<BlockChain> ChainAllocator;
360 /// \brief Function wide BasicBlock to BlockChain mapping.
362 /// This mapping allows efficiently moving from any given basic block to the
363 /// BlockChain it participates in, if any. We use it to, among other things,
364 /// allow implicitly defining edges between chains as the existing edges
365 /// between basic blocks.
366 DenseMap<const MachineBasicBlock *, BlockChain *> BlockToChain;
369 /// The set of basic blocks that have terminators that cannot be fully
370 /// analyzed. These basic blocks cannot be re-ordered safely by
371 /// MachineBlockPlacement, and we must preserve physical layout of these
372 /// blocks and their successors through the pass.
373 SmallPtrSet<MachineBasicBlock *, 4> BlocksWithUnanalyzableExits;
376 /// Decrease the UnscheduledPredecessors count for all blocks in chain, and
377 /// if the count goes to 0, add them to the appropriate work list.
378 void markChainSuccessors(
379 const BlockChain &Chain, const MachineBasicBlock *LoopHeaderBB,
380 const BlockFilterSet *BlockFilter = nullptr);
382 /// Decrease the UnscheduledPredecessors count for a single block, and
383 /// if the count goes to 0, add them to the appropriate work list.
384 void markBlockSuccessors(
385 const BlockChain &Chain, const MachineBasicBlock *BB,
386 const MachineBasicBlock *LoopHeaderBB,
387 const BlockFilterSet *BlockFilter = nullptr);
390 collectViableSuccessors(
391 const MachineBasicBlock *BB, const BlockChain &Chain,
392 const BlockFilterSet *BlockFilter,
393 SmallVector<MachineBasicBlock *, 4> &Successors);
394 bool shouldPredBlockBeOutlined(
395 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
396 const BlockChain &Chain, const BlockFilterSet *BlockFilter,
397 BranchProbability SuccProb, BranchProbability HotProb);
398 bool repeatedlyTailDuplicateBlock(
399 MachineBasicBlock *BB, MachineBasicBlock *&LPred,
400 const MachineBasicBlock *LoopHeaderBB,
401 BlockChain &Chain, BlockFilterSet *BlockFilter,
402 MachineFunction::iterator &PrevUnplacedBlockIt);
403 bool maybeTailDuplicateBlock(
404 MachineBasicBlock *BB, MachineBasicBlock *LPred,
405 BlockChain &Chain, BlockFilterSet *BlockFilter,
406 MachineFunction::iterator &PrevUnplacedBlockIt,
407 bool &DuplicatedToPred);
408 bool hasBetterLayoutPredecessor(
409 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
410 const BlockChain &SuccChain, BranchProbability SuccProb,
411 BranchProbability RealSuccProb, const BlockChain &Chain,
412 const BlockFilterSet *BlockFilter);
413 BlockAndTailDupResult selectBestSuccessor(
414 const MachineBasicBlock *BB, const BlockChain &Chain,
415 const BlockFilterSet *BlockFilter);
416 MachineBasicBlock *selectBestCandidateBlock(
417 const BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList);
418 MachineBasicBlock *getFirstUnplacedBlock(
419 const BlockChain &PlacedChain,
420 MachineFunction::iterator &PrevUnplacedBlockIt,
421 const BlockFilterSet *BlockFilter);
423 /// \brief Add a basic block to the work list if it is appropriate.
425 /// If the optional parameter BlockFilter is provided, only MBB
426 /// present in the set will be added to the worklist. If nullptr
427 /// is provided, no filtering occurs.
428 void fillWorkLists(const MachineBasicBlock *MBB,
429 SmallPtrSetImpl<BlockChain *> &UpdatedPreds,
430 const BlockFilterSet *BlockFilter);
431 void buildChain(const MachineBasicBlock *BB, BlockChain &Chain,
432 BlockFilterSet *BlockFilter = nullptr);
433 MachineBasicBlock *findBestLoopTop(
434 const MachineLoop &L, const BlockFilterSet &LoopBlockSet);
435 MachineBasicBlock *findBestLoopExit(
436 const MachineLoop &L, const BlockFilterSet &LoopBlockSet);
437 BlockFilterSet collectLoopBlockSet(const MachineLoop &L);
438 void buildLoopChains(const MachineLoop &L);
440 BlockChain &LoopChain, const MachineBasicBlock *ExitingBB,
441 const BlockFilterSet &LoopBlockSet);
442 void rotateLoopWithProfile(
443 BlockChain &LoopChain, const MachineLoop &L,
444 const BlockFilterSet &LoopBlockSet);
445 void buildCFGChains();
446 void optimizeBranches();
448 /// Returns true if a block should be tail-duplicated to increase fallthrough
450 bool shouldTailDuplicate(MachineBasicBlock *BB);
451 /// Check the edge frequencies to see if tail duplication will increase
453 bool isProfitableToTailDup(
454 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
455 BranchProbability AdjustedSumProb,
456 const BlockChain &Chain, const BlockFilterSet *BlockFilter);
457 /// Check for a trellis layout.
458 bool isTrellis(const MachineBasicBlock *BB,
459 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
460 const BlockChain &Chain, const BlockFilterSet *BlockFilter);
461 /// Get the best successor given a trellis layout.
462 BlockAndTailDupResult getBestTrellisSuccessor(
463 const MachineBasicBlock *BB,
464 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
465 BranchProbability AdjustedSumProb, const BlockChain &Chain,
466 const BlockFilterSet *BlockFilter);
467 /// Get the best pair of non-conflicting edges.
468 static std::pair<WeightedEdge, WeightedEdge> getBestNonConflictingEdges(
469 const MachineBasicBlock *BB,
470 MutableArrayRef<SmallVector<WeightedEdge, 8>> Edges);
471 /// Returns true if a block can tail duplicate into all unplaced
472 /// predecessors. Filters based on loop.
473 bool canTailDuplicateUnplacedPreds(
474 const MachineBasicBlock *BB, MachineBasicBlock *Succ,
475 const BlockChain &Chain, const BlockFilterSet *BlockFilter);
476 /// Find chains of triangles to tail-duplicate where a global analysis works,
477 /// but a local analysis would not find them.
478 void precomputeTriangleChains();
481 static char ID; // Pass identification, replacement for typeid
482 MachineBlockPlacement() : MachineFunctionPass(ID) {
483 initializeMachineBlockPlacementPass(*PassRegistry::getPassRegistry());
486 bool runOnMachineFunction(MachineFunction &F) override;
488 void getAnalysisUsage(AnalysisUsage &AU) const override {
489 AU.addRequired<MachineBranchProbabilityInfo>();
490 AU.addRequired<MachineBlockFrequencyInfo>();
491 if (TailDupPlacement)
492 AU.addRequired<MachinePostDominatorTree>();
493 AU.addRequired<MachineLoopInfo>();
494 AU.addRequired<TargetPassConfig>();
495 MachineFunctionPass::getAnalysisUsage(AU);
500 char MachineBlockPlacement::ID = 0;
501 char &llvm::MachineBlockPlacementID = MachineBlockPlacement::ID;
502 INITIALIZE_PASS_BEGIN(MachineBlockPlacement, DEBUG_TYPE,
503 "Branch Probability Basic Block Placement", false, false)
504 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
505 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
506 INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTree)
507 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
508 INITIALIZE_PASS_END(MachineBlockPlacement, DEBUG_TYPE,
509 "Branch Probability Basic Block Placement", false, false)
512 /// \brief Helper to print the name of a MBB.
514 /// Only used by debug logging.
515 static std::string getBlockName(const MachineBasicBlock *BB) {
517 raw_string_ostream OS(Result);
518 OS << "BB#" << BB->getNumber();
519 OS << " ('" << BB->getName() << "')";
525 /// \brief Mark a chain's successors as having one fewer preds.
527 /// When a chain is being merged into the "placed" chain, this routine will
528 /// quickly walk the successors of each block in the chain and mark them as
529 /// having one fewer active predecessor. It also adds any successors of this
530 /// chain which reach the zero-predecessor state to the appropriate worklist.
531 void MachineBlockPlacement::markChainSuccessors(
532 const BlockChain &Chain, const MachineBasicBlock *LoopHeaderBB,
533 const BlockFilterSet *BlockFilter) {
534 // Walk all the blocks in this chain, marking their successors as having
535 // a predecessor placed.
536 for (MachineBasicBlock *MBB : Chain) {
537 markBlockSuccessors(Chain, MBB, LoopHeaderBB, BlockFilter);
541 /// \brief Mark a single block's successors as having one fewer preds.
543 /// Under normal circumstances, this is only called by markChainSuccessors,
544 /// but if a block that was to be placed is completely tail-duplicated away,
545 /// and was duplicated into the chain end, we need to redo markBlockSuccessors
546 /// for just that block.
547 void MachineBlockPlacement::markBlockSuccessors(
548 const BlockChain &Chain, const MachineBasicBlock *MBB,
549 const MachineBasicBlock *LoopHeaderBB, const BlockFilterSet *BlockFilter) {
550 // Add any successors for which this is the only un-placed in-loop
551 // predecessor to the worklist as a viable candidate for CFG-neutral
552 // placement. No subsequent placement of this block will violate the CFG
553 // shape, so we get to use heuristics to choose a favorable placement.
554 for (MachineBasicBlock *Succ : MBB->successors()) {
555 if (BlockFilter && !BlockFilter->count(Succ))
557 BlockChain &SuccChain = *BlockToChain[Succ];
558 // Disregard edges within a fixed chain, or edges to the loop header.
559 if (&Chain == &SuccChain || Succ == LoopHeaderBB)
562 // This is a cross-chain edge that is within the loop, so decrement the
563 // loop predecessor count of the destination chain.
564 if (SuccChain.UnscheduledPredecessors == 0 ||
565 --SuccChain.UnscheduledPredecessors > 0)
568 auto *NewBB = *SuccChain.begin();
569 if (NewBB->isEHPad())
570 EHPadWorkList.push_back(NewBB);
572 BlockWorkList.push_back(NewBB);
576 /// This helper function collects the set of successors of block
577 /// \p BB that are allowed to be its layout successors, and return
578 /// the total branch probability of edges from \p BB to those
580 BranchProbability MachineBlockPlacement::collectViableSuccessors(
581 const MachineBasicBlock *BB, const BlockChain &Chain,
582 const BlockFilterSet *BlockFilter,
583 SmallVector<MachineBasicBlock *, 4> &Successors) {
584 // Adjust edge probabilities by excluding edges pointing to blocks that is
585 // either not in BlockFilter or is already in the current chain. Consider the
594 // Assume A->C is very hot (>90%), and C->D has a 50% probability, then after
595 // A->C is chosen as a fall-through, D won't be selected as a successor of C
596 // due to CFG constraint (the probability of C->D is not greater than
597 // HotProb to break top-order). If we exclude E that is not in BlockFilter
598 // when calculating the probability of C->D, D will be selected and we
599 // will get A C D B as the layout of this loop.
600 auto AdjustedSumProb = BranchProbability::getOne();
601 for (MachineBasicBlock *Succ : BB->successors()) {
602 bool SkipSucc = false;
603 if (Succ->isEHPad() || (BlockFilter && !BlockFilter->count(Succ))) {
606 BlockChain *SuccChain = BlockToChain[Succ];
607 if (SuccChain == &Chain) {
609 } else if (Succ != *SuccChain->begin()) {
610 DEBUG(dbgs() << " " << getBlockName(Succ) << " -> Mid chain!\n");
615 AdjustedSumProb -= MBPI->getEdgeProbability(BB, Succ);
617 Successors.push_back(Succ);
620 return AdjustedSumProb;
623 /// The helper function returns the branch probability that is adjusted
624 /// or normalized over the new total \p AdjustedSumProb.
625 static BranchProbability
626 getAdjustedProbability(BranchProbability OrigProb,
627 BranchProbability AdjustedSumProb) {
628 BranchProbability SuccProb;
629 uint32_t SuccProbN = OrigProb.getNumerator();
630 uint32_t SuccProbD = AdjustedSumProb.getNumerator();
631 if (SuccProbN >= SuccProbD)
632 SuccProb = BranchProbability::getOne();
634 SuccProb = BranchProbability(SuccProbN, SuccProbD);
639 /// Check if \p BB has exactly the successors in \p Successors.
641 hasSameSuccessors(MachineBasicBlock &BB,
642 SmallPtrSetImpl<const MachineBasicBlock *> &Successors) {
643 if (BB.succ_size() != Successors.size())
645 // We don't want to count self-loops
646 if (Successors.count(&BB))
648 for (MachineBasicBlock *Succ : BB.successors())
649 if (!Successors.count(Succ))
654 /// Check if a block should be tail duplicated to increase fallthrough
656 /// \p BB Block to check.
657 bool MachineBlockPlacement::shouldTailDuplicate(MachineBasicBlock *BB) {
658 // Blocks with single successors don't create additional fallthrough
659 // opportunities. Don't duplicate them. TODO: When conditional exits are
660 // analyzable, allow them to be duplicated.
661 bool IsSimple = TailDup.isSimpleBB(BB);
663 if (BB->succ_size() == 1)
665 return TailDup.shouldTailDuplicate(IsSimple, *BB);
668 /// Compare 2 BlockFrequency's with a small penalty for \p A.
669 /// In order to be conservative, we apply a X% penalty to account for
670 /// increased icache pressure and static heuristics. For small frequencies
671 /// we use only the numerators to improve accuracy. For simplicity, we assume the
672 /// penalty is less than 100%
673 /// TODO(iteratee): Use 64-bit fixed point edge frequencies everywhere.
674 static bool greaterWithBias(BlockFrequency A, BlockFrequency B,
675 uint64_t EntryFreq) {
676 BranchProbability ThresholdProb(TailDupPlacementPenalty, 100);
677 BlockFrequency Gain = A - B;
678 return (Gain / ThresholdProb).getFrequency() >= EntryFreq;
681 /// Check the edge frequencies to see if tail duplication will increase
682 /// fallthroughs. It only makes sense to call this function when
683 /// \p Succ would not be chosen otherwise. Tail duplication of \p Succ is
684 /// always locally profitable if we would have picked \p Succ without
685 /// considering duplication.
686 bool MachineBlockPlacement::isProfitableToTailDup(
687 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
688 BranchProbability QProb,
689 const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
690 // We need to do a probability calculation to make sure this is profitable.
691 // First: does succ have a successor that post-dominates? This affects the
692 // calculation. The 2 relevant cases are:
707 // '=' : Branch taken for that CFG edge
708 // In the second case, Placing Succ while duplicating it into C prevents the
709 // fallthrough of Succ into either D or PDom, because they now have C as an
710 // unplaced predecessor
712 // Start by figuring out which case we fall into
713 MachineBasicBlock *PDom = nullptr;
714 SmallVector<MachineBasicBlock *, 4> SuccSuccs;
715 // Only scan the relevant successors
716 auto AdjustedSuccSumProb =
717 collectViableSuccessors(Succ, Chain, BlockFilter, SuccSuccs);
718 BranchProbability PProb = MBPI->getEdgeProbability(BB, Succ);
719 auto BBFreq = MBFI->getBlockFreq(BB);
720 auto SuccFreq = MBFI->getBlockFreq(Succ);
721 BlockFrequency P = BBFreq * PProb;
722 BlockFrequency Qout = BBFreq * QProb;
723 uint64_t EntryFreq = MBFI->getEntryFreq();
724 // If there are no more successors, it is profitable to copy, as it strictly
725 // increases fallthrough.
726 if (SuccSuccs.size() == 0)
727 return greaterWithBias(P, Qout, EntryFreq);
729 auto BestSuccSucc = BranchProbability::getZero();
730 // Find the PDom or the best Succ if no PDom exists.
731 for (MachineBasicBlock *SuccSucc : SuccSuccs) {
732 auto Prob = MBPI->getEdgeProbability(Succ, SuccSucc);
733 if (Prob > BestSuccSucc)
736 if (MPDT->dominates(SuccSucc, Succ)) {
741 // For the comparisons, we need to know Succ's best incoming edge that isn't
743 auto SuccBestPred = BlockFrequency(0);
744 for (MachineBasicBlock *SuccPred : Succ->predecessors()) {
745 if (SuccPred == Succ || SuccPred == BB
746 || BlockToChain[SuccPred] == &Chain
747 || (BlockFilter && !BlockFilter->count(SuccPred)))
749 auto Freq = MBFI->getBlockFreq(SuccPred)
750 * MBPI->getEdgeProbability(SuccPred, Succ);
751 if (Freq > SuccBestPred)
754 // Qin is Succ's best unplaced incoming edge that isn't BB
755 BlockFrequency Qin = SuccBestPred;
756 // If it doesn't have a post-dominating successor, here is the calculation:
768 // '=' : Branch taken for that CFG edge
769 // Cost in the first case is: P + V
770 // For this calculation, we always assume P > Qout. If Qout > P
771 // The result of this function will be ignored at the caller.
772 // Let F = SuccFreq - Qin
773 // Cost in the second case is: Qout + min(Qin, F) * U + max(Qin, F) * V
775 if (PDom == nullptr || !Succ->isSuccessor(PDom)) {
776 BranchProbability UProb = BestSuccSucc;
777 BranchProbability VProb = AdjustedSuccSumProb - UProb;
778 BlockFrequency F = SuccFreq - Qin;
779 BlockFrequency V = SuccFreq * VProb;
780 BlockFrequency QinU = std::min(Qin, F) * UProb;
781 BlockFrequency BaseCost = P + V;
782 BlockFrequency DupCost = Qout + QinU + std::max(Qin, F) * VProb;
783 return greaterWithBias(BaseCost, DupCost, EntryFreq);
785 BranchProbability UProb = MBPI->getEdgeProbability(Succ, PDom);
786 BranchProbability VProb = AdjustedSuccSumProb - UProb;
787 BlockFrequency U = SuccFreq * UProb;
788 BlockFrequency V = SuccFreq * VProb;
789 BlockFrequency F = SuccFreq - Qin;
790 // If there is a post-dominating successor, here is the calculation:
792 // | \Qout | \ | \Qout | \
794 // = C' |P C = C' |P C
795 // | /Qin | | | /Qin | |
796 // | / | C' (+Succ) | / | C' (+Succ)
797 // Succ Succ /| Succ Succ /|
798 // | \ V | \/ | | \ V | \/ |
799 // |U \ |U /\ =? |U = |U /\ |
800 // = D = = =?| | D | = =|
805 // '=' : Branch taken for that CFG edge
806 // The cost for taken branches in the first case is P + U
807 // Let F = SuccFreq - Qin
808 // The cost in the second case (assuming independence), given the layout:
809 // BB, Succ, (C+Succ), D, Dom or the layout:
810 // BB, Succ, D, Dom, (C+Succ)
811 // is Qout + max(F, Qin) * U + min(F, Qin)
812 // compare P + U vs Qout + P * U + Qin.
814 // The 3rd and 4th cases cover when Dom would be chosen to follow Succ.
816 // For the 3rd case, the cost is P + 2 * V
817 // For the 4th case, the cost is Qout + min(Qin, F) * U + max(Qin, F) * V + V
818 // We choose 4 over 3 when (P + V) > Qout + min(Qin, F) * U + max(Qin, F) * V
819 if (UProb > AdjustedSuccSumProb / 2 &&
820 !hasBetterLayoutPredecessor(Succ, PDom, *BlockToChain[PDom], UProb, UProb,
823 return greaterWithBias(
824 (P + V), (Qout + std::max(Qin, F) * VProb + std::min(Qin, F) * UProb),
827 return greaterWithBias((P + U),
828 (Qout + std::min(Qin, F) * AdjustedSuccSumProb +
829 std::max(Qin, F) * UProb),
833 /// Check for a trellis layout. \p BB is the upper part of a trellis if its
834 /// successors form the lower part of a trellis. A successor set S forms the
835 /// lower part of a trellis if all of the predecessors of S are either in S or
836 /// have all of S as successors. We ignore trellises where BB doesn't have 2
837 /// successors because for fewer than 2, it's trivial, and for 3 or greater they
838 /// are very uncommon and complex to compute optimally. Allowing edges within S
839 /// is not strictly a trellis, but the same algorithm works, so we allow it.
840 bool MachineBlockPlacement::isTrellis(
841 const MachineBasicBlock *BB,
842 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
843 const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
844 // Technically BB could form a trellis with branching factor higher than 2.
845 // But that's extremely uncommon.
846 if (BB->succ_size() != 2 || ViableSuccs.size() != 2)
849 SmallPtrSet<const MachineBasicBlock *, 2> Successors(BB->succ_begin(),
851 // To avoid reviewing the same predecessors twice.
852 SmallPtrSet<const MachineBasicBlock *, 8> SeenPreds;
854 for (MachineBasicBlock *Succ : ViableSuccs) {
856 for (auto SuccPred : Succ->predecessors()) {
857 // Allow triangle successors, but don't count them.
858 if (Successors.count(SuccPred)) {
859 // Make sure that it is actually a triangle.
860 for (MachineBasicBlock *CheckSucc : SuccPred->successors())
861 if (!Successors.count(CheckSucc))
865 const BlockChain *PredChain = BlockToChain[SuccPred];
866 if (SuccPred == BB || (BlockFilter && !BlockFilter->count(SuccPred)) ||
867 PredChain == &Chain || PredChain == BlockToChain[Succ])
870 // Perform the successor check only once.
871 if (!SeenPreds.insert(SuccPred).second)
873 if (!hasSameSuccessors(*SuccPred, Successors))
876 // If one of the successors has only BB as a predecessor, it is not a
884 /// Pick the highest total weight pair of edges that can both be laid out.
885 /// The edges in \p Edges[0] are assumed to have a different destination than
886 /// the edges in \p Edges[1]. Simple counting shows that the best pair is either
887 /// the individual highest weight edges to the 2 different destinations, or in
888 /// case of a conflict, one of them should be replaced with a 2nd best edge.
889 std::pair<MachineBlockPlacement::WeightedEdge,
890 MachineBlockPlacement::WeightedEdge>
891 MachineBlockPlacement::getBestNonConflictingEdges(
892 const MachineBasicBlock *BB,
893 MutableArrayRef<SmallVector<MachineBlockPlacement::WeightedEdge, 8>>
895 // Sort the edges, and then for each successor, find the best incoming
896 // predecessor. If the best incoming predecessors aren't the same,
897 // then that is clearly the best layout. If there is a conflict, one of the
898 // successors will have to fallthrough from the second best predecessor. We
899 // compare which combination is better overall.
901 // Sort for highest frequency.
902 auto Cmp = [](WeightedEdge A, WeightedEdge B) { return A.Weight > B.Weight; };
904 std::stable_sort(Edges[0].begin(), Edges[0].end(), Cmp);
905 std::stable_sort(Edges[1].begin(), Edges[1].end(), Cmp);
906 auto BestA = Edges[0].begin();
907 auto BestB = Edges[1].begin();
908 // Arrange for the correct answer to be in BestA and BestB
909 // If the 2 best edges don't conflict, the answer is already there.
910 if (BestA->Src == BestB->Src) {
911 // Compare the total fallthrough of (Best + Second Best) for both pairs
912 auto SecondBestA = std::next(BestA);
913 auto SecondBestB = std::next(BestB);
914 BlockFrequency BestAScore = BestA->Weight + SecondBestB->Weight;
915 BlockFrequency BestBScore = BestB->Weight + SecondBestA->Weight;
916 if (BestAScore < BestBScore)
921 // Arrange for the BB edge to be in BestA if it exists.
922 if (BestB->Src == BB)
923 std::swap(BestA, BestB);
924 return std::make_pair(*BestA, *BestB);
927 /// Get the best successor from \p BB based on \p BB being part of a trellis.
928 /// We only handle trellises with 2 successors, so the algorithm is
929 /// straightforward: Find the best pair of edges that don't conflict. We find
930 /// the best incoming edge for each successor in the trellis. If those conflict,
931 /// we consider which of them should be replaced with the second best.
932 /// Upon return the two best edges will be in \p BestEdges. If one of the edges
933 /// comes from \p BB, it will be in \p BestEdges[0]
934 MachineBlockPlacement::BlockAndTailDupResult
935 MachineBlockPlacement::getBestTrellisSuccessor(
936 const MachineBasicBlock *BB,
937 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
938 BranchProbability AdjustedSumProb, const BlockChain &Chain,
939 const BlockFilterSet *BlockFilter) {
941 BlockAndTailDupResult Result = {nullptr, false};
942 SmallPtrSet<const MachineBasicBlock *, 4> Successors(BB->succ_begin(),
945 // We assume size 2 because it's common. For general n, we would have to do
946 // the Hungarian algorithm, but it's not worth the complexity because more
947 // than 2 successors is fairly uncommon, and a trellis even more so.
948 if (Successors.size() != 2 || ViableSuccs.size() != 2)
951 // Collect the edge frequencies of all edges that form the trellis.
952 SmallVector<WeightedEdge, 8> Edges[2];
954 for (auto Succ : ViableSuccs) {
955 for (MachineBasicBlock *SuccPred : Succ->predecessors()) {
956 // Skip any placed predecessors that are not BB
958 if ((BlockFilter && !BlockFilter->count(SuccPred)) ||
959 BlockToChain[SuccPred] == &Chain ||
960 BlockToChain[SuccPred] == BlockToChain[Succ])
962 BlockFrequency EdgeFreq = MBFI->getBlockFreq(SuccPred) *
963 MBPI->getEdgeProbability(SuccPred, Succ);
964 Edges[SuccIndex].push_back({EdgeFreq, SuccPred, Succ});
969 // Pick the best combination of 2 edges from all the edges in the trellis.
970 WeightedEdge BestA, BestB;
971 std::tie(BestA, BestB) = getBestNonConflictingEdges(BB, Edges);
973 if (BestA.Src != BB) {
974 // If we have a trellis, and BB doesn't have the best fallthrough edges,
975 // we shouldn't choose any successor. We've already looked and there's a
976 // better fallthrough edge for all the successors.
977 DEBUG(dbgs() << "Trellis, but not one of the chosen edges.\n");
981 // Did we pick the triangle edge? If tail-duplication is profitable, do
982 // that instead. Otherwise merge the triangle edge now while we know it is
984 if (BestA.Dest == BestB.Src) {
985 // The edges are BB->Succ1->Succ2, and we're looking to see if BB->Succ2
987 MachineBasicBlock *Succ1 = BestA.Dest;
988 MachineBasicBlock *Succ2 = BestB.Dest;
989 // Check to see if tail-duplication would be profitable.
990 if (TailDupPlacement && shouldTailDuplicate(Succ2) &&
991 canTailDuplicateUnplacedPreds(BB, Succ2, Chain, BlockFilter) &&
992 isProfitableToTailDup(BB, Succ2, MBPI->getEdgeProbability(BB, Succ1),
993 Chain, BlockFilter)) {
994 DEBUG(BranchProbability Succ2Prob = getAdjustedProbability(
995 MBPI->getEdgeProbability(BB, Succ2), AdjustedSumProb);
996 dbgs() << " Selected: " << getBlockName(Succ2)
997 << ", probability: " << Succ2Prob << " (Tail Duplicate)\n");
999 Result.ShouldTailDup = true;
1003 // We have already computed the optimal edge for the other side of the
1005 ComputedEdges[BestB.Src] = { BestB.Dest, false };
1007 auto TrellisSucc = BestA.Dest;
1008 DEBUG(BranchProbability SuccProb = getAdjustedProbability(
1009 MBPI->getEdgeProbability(BB, TrellisSucc), AdjustedSumProb);
1010 dbgs() << " Selected: " << getBlockName(TrellisSucc)
1011 << ", probability: " << SuccProb << " (Trellis)\n");
1012 Result.BB = TrellisSucc;
1016 /// When the option TailDupPlacement is on, this method checks if the
1017 /// fallthrough candidate block \p Succ (of block \p BB) can be tail-duplicated
1018 /// into all of its unplaced, unfiltered predecessors, that are not BB.
1019 bool MachineBlockPlacement::canTailDuplicateUnplacedPreds(
1020 const MachineBasicBlock *BB, MachineBasicBlock *Succ,
1021 const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
1022 if (!shouldTailDuplicate(Succ))
1025 // For CFG checking.
1026 SmallPtrSet<const MachineBasicBlock *, 4> Successors(BB->succ_begin(),
1028 for (MachineBasicBlock *Pred : Succ->predecessors()) {
1029 // Make sure all unplaced and unfiltered predecessors can be
1030 // tail-duplicated into.
1031 // Skip any blocks that are already placed or not in this loop.
1032 if (Pred == BB || (BlockFilter && !BlockFilter->count(Pred))
1033 || BlockToChain[Pred] == &Chain)
1035 if (!TailDup.canTailDuplicate(Succ, Pred)) {
1036 if (Successors.size() > 1 && hasSameSuccessors(*Pred, Successors))
1037 // This will result in a trellis after tail duplication, so we don't
1038 // need to copy Succ into this predecessor. In the presence
1039 // of a trellis tail duplication can continue to be profitable.
1055 // After BB was duplicated into C, the layout looks like the one on the
1056 // right. BB and C now have the same successors. When considering
1057 // whether Succ can be duplicated into all its unplaced predecessors, we
1059 // We can do this because C already has a profitable fallthrough, namely
1060 // D. TODO(iteratee): ignore sufficiently cold predecessors for
1061 // duplication and for this test.
1063 // This allows trellises to be laid out in 2 separate chains
1064 // (A,B,Succ,...) and later (C,D,...) This is a reasonable heuristic
1065 // because it allows the creation of 2 fallthrough paths with links
1066 // between them, and we correctly identify the best layout for these
1067 // CFGs. We want to extend trellises that the user created in addition
1068 // to trellises created by tail-duplication, so we just look for the
1077 /// Find chains of triangles where we believe it would be profitable to
1078 /// tail-duplicate them all, but a local analysis would not find them.
1079 /// There are 3 ways this can be profitable:
1080 /// 1) The post-dominators marked 50% are actually taken 55% (This shrinks with
1082 /// 2) The chains are statically correlated. Branch probabilities have a very
1083 /// U-shaped distribution.
1084 /// [http://nrs.harvard.edu/urn-3:HUL.InstRepos:24015805]
1085 /// If the branches in a chain are likely to be from the same side of the
1086 /// distribution as their predecessor, but are independent at runtime, this
1087 /// transformation is profitable. (Because the cost of being wrong is a small
1088 /// fixed cost, unlike the standard triangle layout where the cost of being
1089 /// wrong scales with the # of triangles.)
1090 /// 3) The chains are dynamically correlated. If the probability that a previous
1091 /// branch was taken positively influences whether the next branch will be
1093 /// We believe that 2 and 3 are common enough to justify the small margin in 1.
1094 void MachineBlockPlacement::precomputeTriangleChains() {
1095 struct TriangleChain {
1096 std::vector<MachineBasicBlock *> Edges;
1097 TriangleChain(MachineBasicBlock *src, MachineBasicBlock *dst)
1098 : Edges({src, dst}) {}
1100 void append(MachineBasicBlock *dst) {
1101 assert(getKey()->isSuccessor(dst) &&
1102 "Attempting to append a block that is not a successor.");
1103 Edges.push_back(dst);
1106 unsigned count() const { return Edges.size() - 1; }
1108 MachineBasicBlock *getKey() const {
1109 return Edges.back();
1113 if (TriangleChainCount == 0)
1116 DEBUG(dbgs() << "Pre-computing triangle chains.\n");
1117 // Map from last block to the chain that contains it. This allows us to extend
1118 // chains as we find new triangles.
1119 DenseMap<const MachineBasicBlock *, TriangleChain> TriangleChainMap;
1120 for (MachineBasicBlock &BB : *F) {
1121 // If BB doesn't have 2 successors, it doesn't start a triangle.
1122 if (BB.succ_size() != 2)
1124 MachineBasicBlock *PDom = nullptr;
1125 for (MachineBasicBlock *Succ : BB.successors()) {
1126 if (!MPDT->dominates(Succ, &BB))
1131 // If BB doesn't have a post-dominating successor, it doesn't form a
1133 if (PDom == nullptr)
1135 // If PDom has a hint that it is low probability, skip this triangle.
1136 if (MBPI->getEdgeProbability(&BB, PDom) < BranchProbability(50, 100))
1138 // If PDom isn't eligible for duplication, this isn't the kind of triangle
1139 // we're looking for.
1140 if (!shouldTailDuplicate(PDom))
1142 bool CanTailDuplicate = true;
1143 // If PDom can't tail-duplicate into it's non-BB predecessors, then this
1144 // isn't the kind of triangle we're looking for.
1145 for (MachineBasicBlock* Pred : PDom->predecessors()) {
1148 if (!TailDup.canTailDuplicate(PDom, Pred)) {
1149 CanTailDuplicate = false;
1153 // If we can't tail-duplicate PDom to its predecessors, then skip this
1155 if (!CanTailDuplicate)
1158 // Now we have an interesting triangle. Insert it if it's not part of an
1160 // Note: This cannot be replaced with a call insert() or emplace() because
1161 // the find key is BB, but the insert/emplace key is PDom.
1162 auto Found = TriangleChainMap.find(&BB);
1163 // If it is, remove the chain from the map, grow it, and put it back in the
1164 // map with the end as the new key.
1165 if (Found != TriangleChainMap.end()) {
1166 TriangleChain Chain = std::move(Found->second);
1167 TriangleChainMap.erase(Found);
1169 TriangleChainMap.insert(std::make_pair(Chain.getKey(), std::move(Chain)));
1171 auto InsertResult = TriangleChainMap.try_emplace(PDom, &BB, PDom);
1172 assert(InsertResult.second && "Block seen twice.");
1177 // Iterating over a DenseMap is safe here, because the only thing in the body
1178 // of the loop is inserting into another DenseMap (ComputedEdges).
1179 // ComputedEdges is never iterated, so this doesn't lead to non-determinism.
1180 for (auto &ChainPair : TriangleChainMap) {
1181 TriangleChain &Chain = ChainPair.second;
1182 // Benchmarking has shown that due to branch correlation duplicating 2 or
1183 // more triangles is profitable, despite the calculations assuming
1185 if (Chain.count() < TriangleChainCount)
1187 MachineBasicBlock *dst = Chain.Edges.back();
1188 Chain.Edges.pop_back();
1189 for (MachineBasicBlock *src : reverse(Chain.Edges)) {
1190 DEBUG(dbgs() << "Marking edge: " << getBlockName(src) << "->" <<
1191 getBlockName(dst) << " as pre-computed based on triangles.\n");
1193 auto InsertResult = ComputedEdges.insert({src, {dst, true}});
1194 assert(InsertResult.second && "Block seen twice.");
1202 // When profile is not present, return the StaticLikelyProb.
1203 // When profile is available, we need to handle the triangle-shape CFG.
1204 static BranchProbability getLayoutSuccessorProbThreshold(
1205 const MachineBasicBlock *BB) {
1206 if (!BB->getParent()->getFunction()->getEntryCount())
1207 return BranchProbability(StaticLikelyProb, 100);
1208 if (BB->succ_size() == 2) {
1209 const MachineBasicBlock *Succ1 = *BB->succ_begin();
1210 const MachineBasicBlock *Succ2 = *(BB->succ_begin() + 1);
1211 if (Succ1->isSuccessor(Succ2) || Succ2->isSuccessor(Succ1)) {
1212 /* See case 1 below for the cost analysis. For BB->Succ to
1213 * be taken with smaller cost, the following needs to hold:
1214 * Prob(BB->Succ) > 2 * Prob(BB->Pred)
1215 * So the threshold T in the calculation below
1216 * (1-T) * Prob(BB->Succ) > T * Prob(BB->Pred)
1217 * So T / (1 - T) = 2, Yielding T = 2/3
1218 * Also adding user specified branch bias, we have
1219 * T = (2/3)*(ProfileLikelyProb/50)
1220 * = (2*ProfileLikelyProb)/150)
1222 return BranchProbability(2 * ProfileLikelyProb, 150);
1225 return BranchProbability(ProfileLikelyProb, 100);
1228 /// Checks to see if the layout candidate block \p Succ has a better layout
1229 /// predecessor than \c BB. If yes, returns true.
1230 /// \p SuccProb: The probability adjusted for only remaining blocks.
1231 /// Only used for logging
1232 /// \p RealSuccProb: The un-adjusted probability.
1233 /// \p Chain: The chain that BB belongs to and Succ is being considered for.
1234 /// \p BlockFilter: if non-null, the set of blocks that make up the loop being
1236 bool MachineBlockPlacement::hasBetterLayoutPredecessor(
1237 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
1238 const BlockChain &SuccChain, BranchProbability SuccProb,
1239 BranchProbability RealSuccProb, const BlockChain &Chain,
1240 const BlockFilterSet *BlockFilter) {
1242 // There isn't a better layout when there are no unscheduled predecessors.
1243 if (SuccChain.UnscheduledPredecessors == 0)
1246 // There are two basic scenarios here:
1247 // -------------------------------------
1248 // Case 1: triangular shape CFG (if-then):
1255 // In this case, we are evaluating whether to select edge -> Succ, e.g.
1256 // set Succ as the layout successor of BB. Picking Succ as BB's
1257 // successor breaks the CFG constraints (FIXME: define these constraints).
1258 // With this layout, Pred BB
1259 // is forced to be outlined, so the overall cost will be cost of the
1260 // branch taken from BB to Pred, plus the cost of back taken branch
1261 // from Pred to Succ, as well as the additional cost associated
1262 // with the needed unconditional jump instruction from Pred To Succ.
1264 // The cost of the topological order layout is the taken branch cost
1265 // from BB to Succ, so to make BB->Succ a viable candidate, the following
1267 // 2 * freq(BB->Pred) * taken_branch_cost + unconditional_jump_cost
1268 // < freq(BB->Succ) * taken_branch_cost.
1269 // Ignoring unconditional jump cost, we get
1270 // freq(BB->Succ) > 2 * freq(BB->Pred), i.e.,
1271 // prob(BB->Succ) > 2 * prob(BB->Pred)
1273 // When real profile data is available, we can precisely compute the
1274 // probability threshold that is needed for edge BB->Succ to be considered.
1275 // Without profile data, the heuristic requires the branch bias to be
1276 // a lot larger to make sure the signal is very strong (e.g. 80% default).
1277 // -----------------------------------------------------------------
1278 // Case 2: diamond like CFG (if-then-else):
1287 // The current block is BB and edge BB->Succ is now being evaluated.
1288 // Note that edge S->BB was previously already selected because
1289 // prob(S->BB) > prob(S->Pred).
1290 // At this point, 2 blocks can be placed after BB: Pred or Succ. If we
1291 // choose Pred, we will have a topological ordering as shown on the left
1292 // in the picture below. If we choose Succ, we have the solution as shown
1301 // | pred-- | Succ--
1303 // ---succ ---pred--
1305 // cost = freq(S->Pred) + freq(BB->Succ) cost = 2 * freq (S->Pred)
1306 // = freq(S->Pred) + freq(S->BB)
1308 // If we have profile data (i.e, branch probabilities can be trusted), the
1309 // cost (number of taken branches) with layout S->BB->Succ->Pred is 2 *
1310 // freq(S->Pred) while the cost of topo order is freq(S->Pred) + freq(S->BB).
1311 // We know Prob(S->BB) > Prob(S->Pred), so freq(S->BB) > freq(S->Pred), which
1312 // means the cost of topological order is greater.
1313 // When profile data is not available, however, we need to be more
1314 // conservative. If the branch prediction is wrong, breaking the topo-order
1315 // will actually yield a layout with large cost. For this reason, we need
1316 // strong biased branch at block S with Prob(S->BB) in order to select
1317 // BB->Succ. This is equivalent to looking the CFG backward with backward
1318 // edge: Prob(Succ->BB) needs to >= HotProb in order to be selected (without
1320 // --------------------------------------------------------------------------
1321 // Case 3: forked diamond
1333 // The current block is BB and edge BB->S1 is now being evaluated.
1334 // As above S->BB was already selected because
1335 // prob(S->BB) > prob(S->Pred). Assume that prob(BB->S1) >= prob(BB->S2).
1343 // | Pred----| | S1----
1345 // --(S1 or S2) ---Pred--
1349 // topo-cost = freq(S->Pred) + freq(BB->S1) + freq(BB->S2)
1350 // + min(freq(Pred->S1), freq(Pred->S2))
1351 // Non-topo-order cost:
1352 // non-topo-cost = 2 * freq(S->Pred) + freq(BB->S2).
1353 // To be conservative, we can assume that min(freq(Pred->S1), freq(Pred->S2))
1354 // is 0. Then the non topo layout is better when
1355 // freq(S->Pred) < freq(BB->S1).
1356 // This is exactly what is checked below.
1357 // Note there are other shapes that apply (Pred may not be a single block,
1358 // but they all fit this general pattern.)
1359 BranchProbability HotProb = getLayoutSuccessorProbThreshold(BB);
1361 // Make sure that a hot successor doesn't have a globally more
1362 // important predecessor.
1363 BlockFrequency CandidateEdgeFreq = MBFI->getBlockFreq(BB) * RealSuccProb;
1364 bool BadCFGConflict = false;
1366 for (MachineBasicBlock *Pred : Succ->predecessors()) {
1367 if (Pred == Succ || BlockToChain[Pred] == &SuccChain ||
1368 (BlockFilter && !BlockFilter->count(Pred)) ||
1369 BlockToChain[Pred] == &Chain ||
1370 // This check is redundant except for look ahead. This function is
1371 // called for lookahead by isProfitableToTailDup when BB hasn't been
1375 // Do backward checking.
1376 // For all cases above, we need a backward checking to filter out edges that
1377 // are not 'strongly' biased.
1381 // We select edge BB->Succ if
1382 // freq(BB->Succ) > freq(Succ) * HotProb
1383 // i.e. freq(BB->Succ) > freq(BB->Succ) * HotProb + freq(Pred->Succ) *
1385 // i.e. freq((BB->Succ) * (1 - HotProb) > freq(Pred->Succ) * HotProb
1386 // Case 1 is covered too, because the first equation reduces to:
1387 // prob(BB->Succ) > HotProb. (freq(Succ) = freq(BB) for a triangle)
1388 BlockFrequency PredEdgeFreq =
1389 MBFI->getBlockFreq(Pred) * MBPI->getEdgeProbability(Pred, Succ);
1390 if (PredEdgeFreq * HotProb >= CandidateEdgeFreq * HotProb.getCompl()) {
1391 BadCFGConflict = true;
1396 if (BadCFGConflict) {
1397 DEBUG(dbgs() << " Not a candidate: " << getBlockName(Succ) << " -> " << SuccProb
1398 << " (prob) (non-cold CFG conflict)\n");
1405 /// \brief Select the best successor for a block.
1407 /// This looks across all successors of a particular block and attempts to
1408 /// select the "best" one to be the layout successor. It only considers direct
1409 /// successors which also pass the block filter. It will attempt to avoid
1410 /// breaking CFG structure, but cave and break such structures in the case of
1411 /// very hot successor edges.
1413 /// \returns The best successor block found, or null if none are viable, along
1414 /// with a boolean indicating if tail duplication is necessary.
1415 MachineBlockPlacement::BlockAndTailDupResult
1416 MachineBlockPlacement::selectBestSuccessor(
1417 const MachineBasicBlock *BB, const BlockChain &Chain,
1418 const BlockFilterSet *BlockFilter) {
1419 const BranchProbability HotProb(StaticLikelyProb, 100);
1421 BlockAndTailDupResult BestSucc = { nullptr, false };
1422 auto BestProb = BranchProbability::getZero();
1424 SmallVector<MachineBasicBlock *, 4> Successors;
1425 auto AdjustedSumProb =
1426 collectViableSuccessors(BB, Chain, BlockFilter, Successors);
1428 DEBUG(dbgs() << "Selecting best successor for: " << getBlockName(BB) << "\n");
1430 // if we already precomputed the best successor for BB, return that if still
1432 auto FoundEdge = ComputedEdges.find(BB);
1433 if (FoundEdge != ComputedEdges.end()) {
1434 MachineBasicBlock *Succ = FoundEdge->second.BB;
1435 ComputedEdges.erase(FoundEdge);
1436 BlockChain *SuccChain = BlockToChain[Succ];
1437 if (BB->isSuccessor(Succ) && (!BlockFilter || BlockFilter->count(Succ)) &&
1438 SuccChain != &Chain && Succ == *SuccChain->begin())
1439 return FoundEdge->second;
1442 // if BB is part of a trellis, Use the trellis to determine the optimal
1443 // fallthrough edges
1444 if (isTrellis(BB, Successors, Chain, BlockFilter))
1445 return getBestTrellisSuccessor(BB, Successors, AdjustedSumProb, Chain,
1448 // For blocks with CFG violations, we may be able to lay them out anyway with
1449 // tail-duplication. We keep this vector so we can perform the probability
1450 // calculations the minimum number of times.
1451 SmallVector<std::tuple<BranchProbability, MachineBasicBlock *>, 4>
1453 for (MachineBasicBlock *Succ : Successors) {
1454 auto RealSuccProb = MBPI->getEdgeProbability(BB, Succ);
1455 BranchProbability SuccProb =
1456 getAdjustedProbability(RealSuccProb, AdjustedSumProb);
1458 BlockChain &SuccChain = *BlockToChain[Succ];
1459 // Skip the edge \c BB->Succ if block \c Succ has a better layout
1460 // predecessor that yields lower global cost.
1461 if (hasBetterLayoutPredecessor(BB, Succ, SuccChain, SuccProb, RealSuccProb,
1462 Chain, BlockFilter)) {
1463 // If tail duplication would make Succ profitable, place it.
1464 if (TailDupPlacement && shouldTailDuplicate(Succ))
1465 DupCandidates.push_back(std::make_tuple(SuccProb, Succ));
1470 dbgs() << " Candidate: " << getBlockName(Succ) << ", probability: "
1472 << (SuccChain.UnscheduledPredecessors != 0 ? " (CFG break)" : "")
1475 if (BestSucc.BB && BestProb >= SuccProb) {
1476 DEBUG(dbgs() << " Not the best candidate, continuing\n");
1480 DEBUG(dbgs() << " Setting it as best candidate\n");
1482 BestProb = SuccProb;
1484 // Handle the tail duplication candidates in order of decreasing probability.
1485 // Stop at the first one that is profitable. Also stop if they are less
1486 // profitable than BestSucc. Position is important because we preserve it and
1487 // prefer first best match. Here we aren't comparing in order, so we capture
1488 // the position instead.
1489 if (DupCandidates.size() != 0) {
1491 [](const std::tuple<BranchProbability, MachineBasicBlock *> &a,
1492 const std::tuple<BranchProbability, MachineBasicBlock *> &b) {
1493 return std::get<0>(a) > std::get<0>(b);
1495 std::stable_sort(DupCandidates.begin(), DupCandidates.end(), cmp);
1497 for(auto &Tup : DupCandidates) {
1498 BranchProbability DupProb;
1499 MachineBasicBlock *Succ;
1500 std::tie(DupProb, Succ) = Tup;
1501 if (DupProb < BestProb)
1503 if (canTailDuplicateUnplacedPreds(BB, Succ, Chain, BlockFilter)
1504 && (isProfitableToTailDup(BB, Succ, BestProb, Chain, BlockFilter))) {
1506 dbgs() << " Candidate: " << getBlockName(Succ) << ", probability: "
1508 << " (Tail Duplicate)\n");
1510 BestSucc.ShouldTailDup = true;
1516 DEBUG(dbgs() << " Selected: " << getBlockName(BestSucc.BB) << "\n");
1521 /// \brief Select the best block from a worklist.
1523 /// This looks through the provided worklist as a list of candidate basic
1524 /// blocks and select the most profitable one to place. The definition of
1525 /// profitable only really makes sense in the context of a loop. This returns
1526 /// the most frequently visited block in the worklist, which in the case of
1527 /// a loop, is the one most desirable to be physically close to the rest of the
1528 /// loop body in order to improve i-cache behavior.
1530 /// \returns The best block found, or null if none are viable.
1531 MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock(
1532 const BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList) {
1533 // Once we need to walk the worklist looking for a candidate, cleanup the
1534 // worklist of already placed entries.
1535 // FIXME: If this shows up on profiles, it could be folded (at the cost of
1536 // some code complexity) into the loop below.
1537 WorkList.erase(remove_if(WorkList,
1538 [&](MachineBasicBlock *BB) {
1539 return BlockToChain.lookup(BB) == &Chain;
1543 if (WorkList.empty())
1546 bool IsEHPad = WorkList[0]->isEHPad();
1548 MachineBasicBlock *BestBlock = nullptr;
1549 BlockFrequency BestFreq;
1550 for (MachineBasicBlock *MBB : WorkList) {
1551 assert(MBB->isEHPad() == IsEHPad &&
1552 "EHPad mismatch between block and work list.");
1554 BlockChain &SuccChain = *BlockToChain[MBB];
1555 if (&SuccChain == &Chain)
1558 assert(SuccChain.UnscheduledPredecessors == 0 &&
1559 "Found CFG-violating block");
1561 BlockFrequency CandidateFreq = MBFI->getBlockFreq(MBB);
1562 DEBUG(dbgs() << " " << getBlockName(MBB) << " -> ";
1563 MBFI->printBlockFreq(dbgs(), CandidateFreq) << " (freq)\n");
1565 // For ehpad, we layout the least probable first as to avoid jumping back
1566 // from least probable landingpads to more probable ones.
1568 // FIXME: Using probability is probably (!) not the best way to achieve
1569 // this. We should probably have a more principled approach to layout
1572 // The goal is to get:
1574 // +--------------------------+
1576 // InnerLp -> InnerCleanup OuterLp -> OuterCleanup -> Resume
1580 // +-------------------------------------+
1582 // OuterLp -> OuterCleanup -> Resume InnerLp -> InnerCleanup
1583 if (BestBlock && (IsEHPad ^ (BestFreq >= CandidateFreq)))
1587 BestFreq = CandidateFreq;
1593 /// \brief Retrieve the first unplaced basic block.
1595 /// This routine is called when we are unable to use the CFG to walk through
1596 /// all of the basic blocks and form a chain due to unnatural loops in the CFG.
1597 /// We walk through the function's blocks in order, starting from the
1598 /// LastUnplacedBlockIt. We update this iterator on each call to avoid
1599 /// re-scanning the entire sequence on repeated calls to this routine.
1600 MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock(
1601 const BlockChain &PlacedChain,
1602 MachineFunction::iterator &PrevUnplacedBlockIt,
1603 const BlockFilterSet *BlockFilter) {
1604 for (MachineFunction::iterator I = PrevUnplacedBlockIt, E = F->end(); I != E;
1606 if (BlockFilter && !BlockFilter->count(&*I))
1608 if (BlockToChain[&*I] != &PlacedChain) {
1609 PrevUnplacedBlockIt = I;
1610 // Now select the head of the chain to which the unplaced block belongs
1611 // as the block to place. This will force the entire chain to be placed,
1612 // and satisfies the requirements of merging chains.
1613 return *BlockToChain[&*I]->begin();
1619 void MachineBlockPlacement::fillWorkLists(
1620 const MachineBasicBlock *MBB,
1621 SmallPtrSetImpl<BlockChain *> &UpdatedPreds,
1622 const BlockFilterSet *BlockFilter = nullptr) {
1623 BlockChain &Chain = *BlockToChain[MBB];
1624 if (!UpdatedPreds.insert(&Chain).second)
1628 Chain.UnscheduledPredecessors == 0 &&
1629 "Attempting to place block with unscheduled predecessors in worklist.");
1630 for (MachineBasicBlock *ChainBB : Chain) {
1631 assert(BlockToChain[ChainBB] == &Chain &&
1632 "Block in chain doesn't match BlockToChain map.");
1633 for (MachineBasicBlock *Pred : ChainBB->predecessors()) {
1634 if (BlockFilter && !BlockFilter->count(Pred))
1636 if (BlockToChain[Pred] == &Chain)
1638 ++Chain.UnscheduledPredecessors;
1642 if (Chain.UnscheduledPredecessors != 0)
1645 MachineBasicBlock *BB = *Chain.begin();
1647 EHPadWorkList.push_back(BB);
1649 BlockWorkList.push_back(BB);
1652 void MachineBlockPlacement::buildChain(
1653 const MachineBasicBlock *HeadBB, BlockChain &Chain,
1654 BlockFilterSet *BlockFilter) {
1655 assert(HeadBB && "BB must not be null.\n");
1656 assert(BlockToChain[HeadBB] == &Chain && "BlockToChainMap mis-match.\n");
1657 MachineFunction::iterator PrevUnplacedBlockIt = F->begin();
1659 const MachineBasicBlock *LoopHeaderBB = HeadBB;
1660 markChainSuccessors(Chain, LoopHeaderBB, BlockFilter);
1661 MachineBasicBlock *BB = *std::prev(Chain.end());
1663 assert(BB && "null block found at end of chain in loop.");
1664 assert(BlockToChain[BB] == &Chain && "BlockToChainMap mis-match in loop.");
1665 assert(*std::prev(Chain.end()) == BB && "BB Not found at end of chain.");
1668 // Look for the best viable successor if there is one to place immediately
1669 // after this block.
1670 auto Result = selectBestSuccessor(BB, Chain, BlockFilter);
1671 MachineBasicBlock* BestSucc = Result.BB;
1672 bool ShouldTailDup = Result.ShouldTailDup;
1673 if (TailDupPlacement)
1674 ShouldTailDup |= (BestSucc && shouldTailDuplicate(BestSucc));
1676 // If an immediate successor isn't available, look for the best viable
1677 // block among those we've identified as not violating the loop's CFG at
1678 // this point. This won't be a fallthrough, but it will increase locality.
1680 BestSucc = selectBestCandidateBlock(Chain, BlockWorkList);
1682 BestSucc = selectBestCandidateBlock(Chain, EHPadWorkList);
1685 BestSucc = getFirstUnplacedBlock(Chain, PrevUnplacedBlockIt, BlockFilter);
1689 DEBUG(dbgs() << "Unnatural loop CFG detected, forcibly merging the "
1690 "layout successor until the CFG reduces\n");
1693 // Placement may have changed tail duplication opportunities.
1694 // Check for that now.
1695 if (TailDupPlacement && BestSucc && ShouldTailDup) {
1696 // If the chosen successor was duplicated into all its predecessors,
1697 // don't bother laying it out, just go round the loop again with BB as
1699 if (repeatedlyTailDuplicateBlock(BestSucc, BB, LoopHeaderBB, Chain,
1700 BlockFilter, PrevUnplacedBlockIt))
1704 // Place this block, updating the datastructures to reflect its placement.
1705 BlockChain &SuccChain = *BlockToChain[BestSucc];
1706 // Zero out UnscheduledPredecessors for the successor we're about to merge in case
1707 // we selected a successor that didn't fit naturally into the CFG.
1708 SuccChain.UnscheduledPredecessors = 0;
1709 DEBUG(dbgs() << "Merging from " << getBlockName(BB) << " to "
1710 << getBlockName(BestSucc) << "\n");
1711 markChainSuccessors(SuccChain, LoopHeaderBB, BlockFilter);
1712 Chain.merge(BestSucc, &SuccChain);
1713 BB = *std::prev(Chain.end());
1716 DEBUG(dbgs() << "Finished forming chain for header block "
1717 << getBlockName(*Chain.begin()) << "\n");
1720 /// \brief Find the best loop top block for layout.
1722 /// Look for a block which is strictly better than the loop header for laying
1723 /// out at the top of the loop. This looks for one and only one pattern:
1724 /// a latch block with no conditional exit. This block will cause a conditional
1725 /// jump around it or will be the bottom of the loop if we lay it out in place,
1726 /// but if it it doesn't end up at the bottom of the loop for any reason,
1727 /// rotation alone won't fix it. Because such a block will always result in an
1728 /// unconditional jump (for the backedge) rotating it in front of the loop
1729 /// header is always profitable.
1731 MachineBlockPlacement::findBestLoopTop(const MachineLoop &L,
1732 const BlockFilterSet &LoopBlockSet) {
1733 // Placing the latch block before the header may introduce an extra branch
1734 // that skips this block the first time the loop is executed, which we want
1735 // to avoid when optimising for size.
1736 // FIXME: in theory there is a case that does not introduce a new branch,
1737 // i.e. when the layout predecessor does not fallthrough to the loop header.
1738 // In practice this never happens though: there always seems to be a preheader
1739 // that can fallthrough and that is also placed before the header.
1740 if (F->getFunction()->optForSize())
1741 return L.getHeader();
1743 // Check that the header hasn't been fused with a preheader block due to
1744 // crazy branches. If it has, we need to start with the header at the top to
1745 // prevent pulling the preheader into the loop body.
1746 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
1747 if (!LoopBlockSet.count(*HeaderChain.begin()))
1748 return L.getHeader();
1750 DEBUG(dbgs() << "Finding best loop top for: " << getBlockName(L.getHeader())
1753 BlockFrequency BestPredFreq;
1754 MachineBasicBlock *BestPred = nullptr;
1755 for (MachineBasicBlock *Pred : L.getHeader()->predecessors()) {
1756 if (!LoopBlockSet.count(Pred))
1758 DEBUG(dbgs() << " header pred: " << getBlockName(Pred) << ", has "
1759 << Pred->succ_size() << " successors, ";
1760 MBFI->printBlockFreq(dbgs(), Pred) << " freq\n");
1761 if (Pred->succ_size() > 1)
1764 BlockFrequency PredFreq = MBFI->getBlockFreq(Pred);
1765 if (!BestPred || PredFreq > BestPredFreq ||
1766 (!(PredFreq < BestPredFreq) &&
1767 Pred->isLayoutSuccessor(L.getHeader()))) {
1769 BestPredFreq = PredFreq;
1773 // If no direct predecessor is fine, just use the loop header.
1775 DEBUG(dbgs() << " final top unchanged\n");
1776 return L.getHeader();
1779 // Walk backwards through any straight line of predecessors.
1780 while (BestPred->pred_size() == 1 &&
1781 (*BestPred->pred_begin())->succ_size() == 1 &&
1782 *BestPred->pred_begin() != L.getHeader())
1783 BestPred = *BestPred->pred_begin();
1785 DEBUG(dbgs() << " final top: " << getBlockName(BestPred) << "\n");
1789 /// \brief Find the best loop exiting block for layout.
1791 /// This routine implements the logic to analyze the loop looking for the best
1792 /// block to layout at the top of the loop. Typically this is done to maximize
1793 /// fallthrough opportunities.
1795 MachineBlockPlacement::findBestLoopExit(const MachineLoop &L,
1796 const BlockFilterSet &LoopBlockSet) {
1797 // We don't want to layout the loop linearly in all cases. If the loop header
1798 // is just a normal basic block in the loop, we want to look for what block
1799 // within the loop is the best one to layout at the top. However, if the loop
1800 // header has be pre-merged into a chain due to predecessors not having
1801 // analyzable branches, *and* the predecessor it is merged with is *not* part
1802 // of the loop, rotating the header into the middle of the loop will create
1803 // a non-contiguous range of blocks which is Very Bad. So start with the
1804 // header and only rotate if safe.
1805 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
1806 if (!LoopBlockSet.count(*HeaderChain.begin()))
1809 BlockFrequency BestExitEdgeFreq;
1810 unsigned BestExitLoopDepth = 0;
1811 MachineBasicBlock *ExitingBB = nullptr;
1812 // If there are exits to outer loops, loop rotation can severely limit
1813 // fallthrough opportunities unless it selects such an exit. Keep a set of
1814 // blocks where rotating to exit with that block will reach an outer loop.
1815 SmallPtrSet<MachineBasicBlock *, 4> BlocksExitingToOuterLoop;
1817 DEBUG(dbgs() << "Finding best loop exit for: " << getBlockName(L.getHeader())
1819 for (MachineBasicBlock *MBB : L.getBlocks()) {
1820 BlockChain &Chain = *BlockToChain[MBB];
1821 // Ensure that this block is at the end of a chain; otherwise it could be
1822 // mid-way through an inner loop or a successor of an unanalyzable branch.
1823 if (MBB != *std::prev(Chain.end()))
1826 // Now walk the successors. We need to establish whether this has a viable
1827 // exiting successor and whether it has a viable non-exiting successor.
1828 // We store the old exiting state and restore it if a viable looping
1829 // successor isn't found.
1830 MachineBasicBlock *OldExitingBB = ExitingBB;
1831 BlockFrequency OldBestExitEdgeFreq = BestExitEdgeFreq;
1832 bool HasLoopingSucc = false;
1833 for (MachineBasicBlock *Succ : MBB->successors()) {
1834 if (Succ->isEHPad())
1838 BlockChain &SuccChain = *BlockToChain[Succ];
1839 // Don't split chains, either this chain or the successor's chain.
1840 if (&Chain == &SuccChain) {
1841 DEBUG(dbgs() << " exiting: " << getBlockName(MBB) << " -> "
1842 << getBlockName(Succ) << " (chain conflict)\n");
1846 auto SuccProb = MBPI->getEdgeProbability(MBB, Succ);
1847 if (LoopBlockSet.count(Succ)) {
1848 DEBUG(dbgs() << " looping: " << getBlockName(MBB) << " -> "
1849 << getBlockName(Succ) << " (" << SuccProb << ")\n");
1850 HasLoopingSucc = true;
1854 unsigned SuccLoopDepth = 0;
1855 if (MachineLoop *ExitLoop = MLI->getLoopFor(Succ)) {
1856 SuccLoopDepth = ExitLoop->getLoopDepth();
1857 if (ExitLoop->contains(&L))
1858 BlocksExitingToOuterLoop.insert(MBB);
1861 BlockFrequency ExitEdgeFreq = MBFI->getBlockFreq(MBB) * SuccProb;
1862 DEBUG(dbgs() << " exiting: " << getBlockName(MBB) << " -> "
1863 << getBlockName(Succ) << " [L:" << SuccLoopDepth << "] (";
1864 MBFI->printBlockFreq(dbgs(), ExitEdgeFreq) << ")\n");
1865 // Note that we bias this toward an existing layout successor to retain
1866 // incoming order in the absence of better information. The exit must have
1867 // a frequency higher than the current exit before we consider breaking
1869 BranchProbability Bias(100 - ExitBlockBias, 100);
1870 if (!ExitingBB || SuccLoopDepth > BestExitLoopDepth ||
1871 ExitEdgeFreq > BestExitEdgeFreq ||
1872 (MBB->isLayoutSuccessor(Succ) &&
1873 !(ExitEdgeFreq < BestExitEdgeFreq * Bias))) {
1874 BestExitEdgeFreq = ExitEdgeFreq;
1879 if (!HasLoopingSucc) {
1880 // Restore the old exiting state, no viable looping successor was found.
1881 ExitingBB = OldExitingBB;
1882 BestExitEdgeFreq = OldBestExitEdgeFreq;
1885 // Without a candidate exiting block or with only a single block in the
1886 // loop, just use the loop header to layout the loop.
1888 DEBUG(dbgs() << " No other candidate exit blocks, using loop header\n");
1891 if (L.getNumBlocks() == 1) {
1892 DEBUG(dbgs() << " Loop has 1 block, using loop header as exit\n");
1896 // Also, if we have exit blocks which lead to outer loops but didn't select
1897 // one of them as the exiting block we are rotating toward, disable loop
1898 // rotation altogether.
1899 if (!BlocksExitingToOuterLoop.empty() &&
1900 !BlocksExitingToOuterLoop.count(ExitingBB))
1903 DEBUG(dbgs() << " Best exiting block: " << getBlockName(ExitingBB) << "\n");
1907 /// \brief Attempt to rotate an exiting block to the bottom of the loop.
1909 /// Once we have built a chain, try to rotate it to line up the hot exit block
1910 /// with fallthrough out of the loop if doing so doesn't introduce unnecessary
1911 /// branches. For example, if the loop has fallthrough into its header and out
1912 /// of its bottom already, don't rotate it.
1913 void MachineBlockPlacement::rotateLoop(BlockChain &LoopChain,
1914 const MachineBasicBlock *ExitingBB,
1915 const BlockFilterSet &LoopBlockSet) {
1919 MachineBasicBlock *Top = *LoopChain.begin();
1920 bool ViableTopFallthrough = false;
1921 for (MachineBasicBlock *Pred : Top->predecessors()) {
1922 BlockChain *PredChain = BlockToChain[Pred];
1923 if (!LoopBlockSet.count(Pred) &&
1924 (!PredChain || Pred == *std::prev(PredChain->end()))) {
1925 ViableTopFallthrough = true;
1930 // If the header has viable fallthrough, check whether the current loop
1931 // bottom is a viable exiting block. If so, bail out as rotating will
1932 // introduce an unnecessary branch.
1933 if (ViableTopFallthrough) {
1934 MachineBasicBlock *Bottom = *std::prev(LoopChain.end());
1935 for (MachineBasicBlock *Succ : Bottom->successors()) {
1936 BlockChain *SuccChain = BlockToChain[Succ];
1937 if (!LoopBlockSet.count(Succ) &&
1938 (!SuccChain || Succ == *SuccChain->begin()))
1943 BlockChain::iterator ExitIt = find(LoopChain, ExitingBB);
1944 if (ExitIt == LoopChain.end())
1947 std::rotate(LoopChain.begin(), std::next(ExitIt), LoopChain.end());
1950 /// \brief Attempt to rotate a loop based on profile data to reduce branch cost.
1952 /// With profile data, we can determine the cost in terms of missed fall through
1953 /// opportunities when rotating a loop chain and select the best rotation.
1954 /// Basically, there are three kinds of cost to consider for each rotation:
1955 /// 1. The possibly missed fall through edge (if it exists) from BB out of
1956 /// the loop to the loop header.
1957 /// 2. The possibly missed fall through edges (if they exist) from the loop
1958 /// exits to BB out of the loop.
1959 /// 3. The missed fall through edge (if it exists) from the last BB to the
1960 /// first BB in the loop chain.
1961 /// Therefore, the cost for a given rotation is the sum of costs listed above.
1962 /// We select the best rotation with the smallest cost.
1963 void MachineBlockPlacement::rotateLoopWithProfile(
1964 BlockChain &LoopChain, const MachineLoop &L,
1965 const BlockFilterSet &LoopBlockSet) {
1966 auto HeaderBB = L.getHeader();
1967 auto HeaderIter = find(LoopChain, HeaderBB);
1968 auto RotationPos = LoopChain.end();
1970 BlockFrequency SmallestRotationCost = BlockFrequency::getMaxFrequency();
1972 // A utility lambda that scales up a block frequency by dividing it by a
1973 // branch probability which is the reciprocal of the scale.
1974 auto ScaleBlockFrequency = [](BlockFrequency Freq,
1975 unsigned Scale) -> BlockFrequency {
1978 // Use operator / between BlockFrequency and BranchProbability to implement
1979 // saturating multiplication.
1980 return Freq / BranchProbability(1, Scale);
1983 // Compute the cost of the missed fall-through edge to the loop header if the
1984 // chain head is not the loop header. As we only consider natural loops with
1985 // single header, this computation can be done only once.
1986 BlockFrequency HeaderFallThroughCost(0);
1987 for (auto *Pred : HeaderBB->predecessors()) {
1988 BlockChain *PredChain = BlockToChain[Pred];
1989 if (!LoopBlockSet.count(Pred) &&
1990 (!PredChain || Pred == *std::prev(PredChain->end()))) {
1992 MBFI->getBlockFreq(Pred) * MBPI->getEdgeProbability(Pred, HeaderBB);
1993 auto FallThruCost = ScaleBlockFrequency(EdgeFreq, MisfetchCost);
1994 // If the predecessor has only an unconditional jump to the header, we
1995 // need to consider the cost of this jump.
1996 if (Pred->succ_size() == 1)
1997 FallThruCost += ScaleBlockFrequency(EdgeFreq, JumpInstCost);
1998 HeaderFallThroughCost = std::max(HeaderFallThroughCost, FallThruCost);
2002 // Here we collect all exit blocks in the loop, and for each exit we find out
2003 // its hottest exit edge. For each loop rotation, we define the loop exit cost
2004 // as the sum of frequencies of exit edges we collect here, excluding the exit
2005 // edge from the tail of the loop chain.
2006 SmallVector<std::pair<MachineBasicBlock *, BlockFrequency>, 4> ExitsWithFreq;
2007 for (auto BB : LoopChain) {
2008 auto LargestExitEdgeProb = BranchProbability::getZero();
2009 for (auto *Succ : BB->successors()) {
2010 BlockChain *SuccChain = BlockToChain[Succ];
2011 if (!LoopBlockSet.count(Succ) &&
2012 (!SuccChain || Succ == *SuccChain->begin())) {
2013 auto SuccProb = MBPI->getEdgeProbability(BB, Succ);
2014 LargestExitEdgeProb = std::max(LargestExitEdgeProb, SuccProb);
2017 if (LargestExitEdgeProb > BranchProbability::getZero()) {
2018 auto ExitFreq = MBFI->getBlockFreq(BB) * LargestExitEdgeProb;
2019 ExitsWithFreq.emplace_back(BB, ExitFreq);
2023 // In this loop we iterate every block in the loop chain and calculate the
2024 // cost assuming the block is the head of the loop chain. When the loop ends,
2025 // we should have found the best candidate as the loop chain's head.
2026 for (auto Iter = LoopChain.begin(), TailIter = std::prev(LoopChain.end()),
2027 EndIter = LoopChain.end();
2028 Iter != EndIter; Iter++, TailIter++) {
2029 // TailIter is used to track the tail of the loop chain if the block we are
2030 // checking (pointed by Iter) is the head of the chain.
2031 if (TailIter == LoopChain.end())
2032 TailIter = LoopChain.begin();
2034 auto TailBB = *TailIter;
2036 // Calculate the cost by putting this BB to the top.
2037 BlockFrequency Cost = 0;
2039 // If the current BB is the loop header, we need to take into account the
2040 // cost of the missed fall through edge from outside of the loop to the
2042 if (Iter != HeaderIter)
2043 Cost += HeaderFallThroughCost;
2045 // Collect the loop exit cost by summing up frequencies of all exit edges
2046 // except the one from the chain tail.
2047 for (auto &ExitWithFreq : ExitsWithFreq)
2048 if (TailBB != ExitWithFreq.first)
2049 Cost += ExitWithFreq.second;
2051 // The cost of breaking the once fall-through edge from the tail to the top
2052 // of the loop chain. Here we need to consider three cases:
2053 // 1. If the tail node has only one successor, then we will get an
2054 // additional jmp instruction. So the cost here is (MisfetchCost +
2055 // JumpInstCost) * tail node frequency.
2056 // 2. If the tail node has two successors, then we may still get an
2057 // additional jmp instruction if the layout successor after the loop
2058 // chain is not its CFG successor. Note that the more frequently executed
2059 // jmp instruction will be put ahead of the other one. Assume the
2060 // frequency of those two branches are x and y, where x is the frequency
2061 // of the edge to the chain head, then the cost will be
2062 // (x * MisfetechCost + min(x, y) * JumpInstCost) * tail node frequency.
2063 // 3. If the tail node has more than two successors (this rarely happens),
2064 // we won't consider any additional cost.
2065 if (TailBB->isSuccessor(*Iter)) {
2066 auto TailBBFreq = MBFI->getBlockFreq(TailBB);
2067 if (TailBB->succ_size() == 1)
2068 Cost += ScaleBlockFrequency(TailBBFreq.getFrequency(),
2069 MisfetchCost + JumpInstCost);
2070 else if (TailBB->succ_size() == 2) {
2071 auto TailToHeadProb = MBPI->getEdgeProbability(TailBB, *Iter);
2072 auto TailToHeadFreq = TailBBFreq * TailToHeadProb;
2073 auto ColderEdgeFreq = TailToHeadProb > BranchProbability(1, 2)
2074 ? TailBBFreq * TailToHeadProb.getCompl()
2076 Cost += ScaleBlockFrequency(TailToHeadFreq, MisfetchCost) +
2077 ScaleBlockFrequency(ColderEdgeFreq, JumpInstCost);
2081 DEBUG(dbgs() << "The cost of loop rotation by making " << getBlockName(*Iter)
2082 << " to the top: " << Cost.getFrequency() << "\n");
2084 if (Cost < SmallestRotationCost) {
2085 SmallestRotationCost = Cost;
2090 if (RotationPos != LoopChain.end()) {
2091 DEBUG(dbgs() << "Rotate loop by making " << getBlockName(*RotationPos)
2092 << " to the top\n");
2093 std::rotate(LoopChain.begin(), RotationPos, LoopChain.end());
2097 /// \brief Collect blocks in the given loop that are to be placed.
2099 /// When profile data is available, exclude cold blocks from the returned set;
2100 /// otherwise, collect all blocks in the loop.
2101 MachineBlockPlacement::BlockFilterSet
2102 MachineBlockPlacement::collectLoopBlockSet(const MachineLoop &L) {
2103 BlockFilterSet LoopBlockSet;
2105 // Filter cold blocks off from LoopBlockSet when profile data is available.
2106 // Collect the sum of frequencies of incoming edges to the loop header from
2107 // outside. If we treat the loop as a super block, this is the frequency of
2108 // the loop. Then for each block in the loop, we calculate the ratio between
2109 // its frequency and the frequency of the loop block. When it is too small,
2110 // don't add it to the loop chain. If there are outer loops, then this block
2111 // will be merged into the first outer loop chain for which this block is not
2112 // cold anymore. This needs precise profile data and we only do this when
2113 // profile data is available.
2114 if (F->getFunction()->getEntryCount()) {
2115 BlockFrequency LoopFreq(0);
2116 for (auto LoopPred : L.getHeader()->predecessors())
2117 if (!L.contains(LoopPred))
2118 LoopFreq += MBFI->getBlockFreq(LoopPred) *
2119 MBPI->getEdgeProbability(LoopPred, L.getHeader());
2121 for (MachineBasicBlock *LoopBB : L.getBlocks()) {
2122 auto Freq = MBFI->getBlockFreq(LoopBB).getFrequency();
2123 if (Freq == 0 || LoopFreq.getFrequency() / Freq > LoopToColdBlockRatio)
2125 LoopBlockSet.insert(LoopBB);
2128 LoopBlockSet.insert(L.block_begin(), L.block_end());
2130 return LoopBlockSet;
2133 /// \brief Forms basic block chains from the natural loop structures.
2135 /// These chains are designed to preserve the existing *structure* of the code
2136 /// as much as possible. We can then stitch the chains together in a way which
2137 /// both preserves the topological structure and minimizes taken conditional
2139 void MachineBlockPlacement::buildLoopChains(const MachineLoop &L) {
2140 // First recurse through any nested loops, building chains for those inner
2142 for (const MachineLoop *InnerLoop : L)
2143 buildLoopChains(*InnerLoop);
2145 assert(BlockWorkList.empty() &&
2146 "BlockWorkList not empty when starting to build loop chains.");
2147 assert(EHPadWorkList.empty() &&
2148 "EHPadWorkList not empty when starting to build loop chains.");
2149 BlockFilterSet LoopBlockSet = collectLoopBlockSet(L);
2151 // Check if we have profile data for this function. If yes, we will rotate
2152 // this loop by modeling costs more precisely which requires the profile data
2153 // for better layout.
2154 bool RotateLoopWithProfile =
2155 ForcePreciseRotationCost ||
2156 (PreciseRotationCost && F->getFunction()->getEntryCount());
2158 // First check to see if there is an obviously preferable top block for the
2159 // loop. This will default to the header, but may end up as one of the
2160 // predecessors to the header if there is one which will result in strictly
2161 // fewer branches in the loop body.
2162 // When we use profile data to rotate the loop, this is unnecessary.
2163 MachineBasicBlock *LoopTop =
2164 RotateLoopWithProfile ? L.getHeader() : findBestLoopTop(L, LoopBlockSet);
2166 // If we selected just the header for the loop top, look for a potentially
2167 // profitable exit block in the event that rotating the loop can eliminate
2168 // branches by placing an exit edge at the bottom.
2169 if (!RotateLoopWithProfile && LoopTop == L.getHeader())
2170 PreferredLoopExit = findBestLoopExit(L, LoopBlockSet);
2172 BlockChain &LoopChain = *BlockToChain[LoopTop];
2174 // FIXME: This is a really lame way of walking the chains in the loop: we
2175 // walk the blocks, and use a set to prevent visiting a particular chain
2177 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
2178 assert(LoopChain.UnscheduledPredecessors == 0 &&
2179 "LoopChain should not have unscheduled predecessors.");
2180 UpdatedPreds.insert(&LoopChain);
2182 for (const MachineBasicBlock *LoopBB : LoopBlockSet)
2183 fillWorkLists(LoopBB, UpdatedPreds, &LoopBlockSet);
2185 buildChain(LoopTop, LoopChain, &LoopBlockSet);
2187 if (RotateLoopWithProfile)
2188 rotateLoopWithProfile(LoopChain, L, LoopBlockSet);
2190 rotateLoop(LoopChain, PreferredLoopExit, LoopBlockSet);
2193 // Crash at the end so we get all of the debugging output first.
2194 bool BadLoop = false;
2195 if (LoopChain.UnscheduledPredecessors) {
2197 dbgs() << "Loop chain contains a block without its preds placed!\n"
2198 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
2199 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n";
2201 for (MachineBasicBlock *ChainBB : LoopChain) {
2202 dbgs() << " ... " << getBlockName(ChainBB) << "\n";
2203 if (!LoopBlockSet.remove(ChainBB)) {
2204 // We don't mark the loop as bad here because there are real situations
2205 // where this can occur. For example, with an unanalyzable fallthrough
2206 // from a loop block to a non-loop block or vice versa.
2207 dbgs() << "Loop chain contains a block not contained by the loop!\n"
2208 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
2209 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
2210 << " Bad block: " << getBlockName(ChainBB) << "\n";
2214 if (!LoopBlockSet.empty()) {
2216 for (const MachineBasicBlock *LoopBB : LoopBlockSet)
2217 dbgs() << "Loop contains blocks never placed into a chain!\n"
2218 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
2219 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
2220 << " Bad block: " << getBlockName(LoopBB) << "\n";
2222 assert(!BadLoop && "Detected problems with the placement of this loop.");
2225 BlockWorkList.clear();
2226 EHPadWorkList.clear();
2229 void MachineBlockPlacement::buildCFGChains() {
2230 // Ensure that every BB in the function has an associated chain to simplify
2231 // the assumptions of the remaining algorithm.
2232 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
2233 for (MachineFunction::iterator FI = F->begin(), FE = F->end(); FI != FE;
2235 MachineBasicBlock *BB = &*FI;
2237 new (ChainAllocator.Allocate()) BlockChain(BlockToChain, BB);
2238 // Also, merge any blocks which we cannot reason about and must preserve
2239 // the exact fallthrough behavior for.
2242 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
2243 if (!TII->analyzeBranch(*BB, TBB, FBB, Cond) || !FI->canFallThrough())
2246 MachineFunction::iterator NextFI = std::next(FI);
2247 MachineBasicBlock *NextBB = &*NextFI;
2248 // Ensure that the layout successor is a viable block, as we know that
2249 // fallthrough is a possibility.
2250 assert(NextFI != FE && "Can't fallthrough past the last block.");
2251 DEBUG(dbgs() << "Pre-merging due to unanalyzable fallthrough: "
2252 << getBlockName(BB) << " -> " << getBlockName(NextBB)
2254 Chain->merge(NextBB, nullptr);
2256 BlocksWithUnanalyzableExits.insert(&*BB);
2263 // Build any loop-based chains.
2264 PreferredLoopExit = nullptr;
2265 for (MachineLoop *L : *MLI)
2266 buildLoopChains(*L);
2268 assert(BlockWorkList.empty() &&
2269 "BlockWorkList should be empty before building final chain.");
2270 assert(EHPadWorkList.empty() &&
2271 "EHPadWorkList should be empty before building final chain.");
2273 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
2274 for (MachineBasicBlock &MBB : *F)
2275 fillWorkLists(&MBB, UpdatedPreds);
2277 BlockChain &FunctionChain = *BlockToChain[&F->front()];
2278 buildChain(&F->front(), FunctionChain);
2281 typedef SmallPtrSet<MachineBasicBlock *, 16> FunctionBlockSetType;
2284 // Crash at the end so we get all of the debugging output first.
2285 bool BadFunc = false;
2286 FunctionBlockSetType FunctionBlockSet;
2287 for (MachineBasicBlock &MBB : *F)
2288 FunctionBlockSet.insert(&MBB);
2290 for (MachineBasicBlock *ChainBB : FunctionChain)
2291 if (!FunctionBlockSet.erase(ChainBB)) {
2293 dbgs() << "Function chain contains a block not in the function!\n"
2294 << " Bad block: " << getBlockName(ChainBB) << "\n";
2297 if (!FunctionBlockSet.empty()) {
2299 for (MachineBasicBlock *RemainingBB : FunctionBlockSet)
2300 dbgs() << "Function contains blocks never placed into a chain!\n"
2301 << " Bad block: " << getBlockName(RemainingBB) << "\n";
2303 assert(!BadFunc && "Detected problems with the block placement.");
2306 // Splice the blocks into place.
2307 MachineFunction::iterator InsertPos = F->begin();
2308 DEBUG(dbgs() << "[MBP] Function: "<< F->getName() << "\n");
2309 for (MachineBasicBlock *ChainBB : FunctionChain) {
2310 DEBUG(dbgs() << (ChainBB == *FunctionChain.begin() ? "Placing chain "
2312 << getBlockName(ChainBB) << "\n");
2313 if (InsertPos != MachineFunction::iterator(ChainBB))
2314 F->splice(InsertPos, ChainBB);
2318 // Update the terminator of the previous block.
2319 if (ChainBB == *FunctionChain.begin())
2321 MachineBasicBlock *PrevBB = &*std::prev(MachineFunction::iterator(ChainBB));
2323 // FIXME: It would be awesome of updateTerminator would just return rather
2324 // than assert when the branch cannot be analyzed in order to remove this
2327 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
2330 if (!BlocksWithUnanalyzableExits.count(PrevBB)) {
2331 // Given the exact block placement we chose, we may actually not _need_ to
2332 // be able to edit PrevBB's terminator sequence, but not being _able_ to
2333 // do that at this point is a bug.
2334 assert((!TII->analyzeBranch(*PrevBB, TBB, FBB, Cond) ||
2335 !PrevBB->canFallThrough()) &&
2336 "Unexpected block with un-analyzable fallthrough!");
2338 TBB = FBB = nullptr;
2342 // The "PrevBB" is not yet updated to reflect current code layout, so,
2343 // o. it may fall-through to a block without explicit "goto" instruction
2344 // before layout, and no longer fall-through it after layout; or
2345 // o. just opposite.
2347 // analyzeBranch() may return erroneous value for FBB when these two
2348 // situations take place. For the first scenario FBB is mistakenly set NULL;
2349 // for the 2nd scenario, the FBB, which is expected to be NULL, is
2350 // mistakenly pointing to "*BI".
2351 // Thus, if the future change needs to use FBB before the layout is set, it
2352 // has to correct FBB first by using the code similar to the following:
2354 // if (!Cond.empty() && (!FBB || FBB == ChainBB)) {
2355 // PrevBB->updateTerminator();
2357 // TBB = FBB = nullptr;
2358 // if (TII->analyzeBranch(*PrevBB, TBB, FBB, Cond)) {
2359 // // FIXME: This should never take place.
2360 // TBB = FBB = nullptr;
2363 if (!TII->analyzeBranch(*PrevBB, TBB, FBB, Cond))
2364 PrevBB->updateTerminator();
2367 // Fixup the last block.
2369 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
2370 if (!TII->analyzeBranch(F->back(), TBB, FBB, Cond))
2371 F->back().updateTerminator();
2373 BlockWorkList.clear();
2374 EHPadWorkList.clear();
2377 void MachineBlockPlacement::optimizeBranches() {
2378 BlockChain &FunctionChain = *BlockToChain[&F->front()];
2379 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
2381 // Now that all the basic blocks in the chain have the proper layout,
2382 // make a final call to AnalyzeBranch with AllowModify set.
2383 // Indeed, the target may be able to optimize the branches in a way we
2384 // cannot because all branches may not be analyzable.
2385 // E.g., the target may be able to remove an unconditional branch to
2386 // a fallthrough when it occurs after predicated terminators.
2387 for (MachineBasicBlock *ChainBB : FunctionChain) {
2389 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
2390 if (!TII->analyzeBranch(*ChainBB, TBB, FBB, Cond, /*AllowModify*/ true)) {
2391 // If PrevBB has a two-way branch, try to re-order the branches
2392 // such that we branch to the successor with higher probability first.
2393 if (TBB && !Cond.empty() && FBB &&
2394 MBPI->getEdgeProbability(ChainBB, FBB) >
2395 MBPI->getEdgeProbability(ChainBB, TBB) &&
2396 !TII->reverseBranchCondition(Cond)) {
2397 DEBUG(dbgs() << "Reverse order of the two branches: "
2398 << getBlockName(ChainBB) << "\n");
2399 DEBUG(dbgs() << " Edge probability: "
2400 << MBPI->getEdgeProbability(ChainBB, FBB) << " vs "
2401 << MBPI->getEdgeProbability(ChainBB, TBB) << "\n");
2402 DebugLoc dl; // FIXME: this is nowhere
2403 TII->removeBranch(*ChainBB);
2404 TII->insertBranch(*ChainBB, FBB, TBB, Cond, dl);
2405 ChainBB->updateTerminator();
2411 void MachineBlockPlacement::alignBlocks() {
2412 // Walk through the backedges of the function now that we have fully laid out
2413 // the basic blocks and align the destination of each backedge. We don't rely
2414 // exclusively on the loop info here so that we can align backedges in
2415 // unnatural CFGs and backedges that were introduced purely because of the
2416 // loop rotations done during this layout pass.
2417 if (F->getFunction()->optForSize())
2419 BlockChain &FunctionChain = *BlockToChain[&F->front()];
2420 if (FunctionChain.begin() == FunctionChain.end())
2421 return; // Empty chain.
2423 const BranchProbability ColdProb(1, 5); // 20%
2424 BlockFrequency EntryFreq = MBFI->getBlockFreq(&F->front());
2425 BlockFrequency WeightedEntryFreq = EntryFreq * ColdProb;
2426 for (MachineBasicBlock *ChainBB : FunctionChain) {
2427 if (ChainBB == *FunctionChain.begin())
2430 // Don't align non-looping basic blocks. These are unlikely to execute
2431 // enough times to matter in practice. Note that we'll still handle
2432 // unnatural CFGs inside of a natural outer loop (the common case) and
2434 MachineLoop *L = MLI->getLoopFor(ChainBB);
2438 unsigned Align = TLI->getPrefLoopAlignment(L);
2440 continue; // Don't care about loop alignment.
2442 // If the block is cold relative to the function entry don't waste space
2444 BlockFrequency Freq = MBFI->getBlockFreq(ChainBB);
2445 if (Freq < WeightedEntryFreq)
2448 // If the block is cold relative to its loop header, don't align it
2449 // regardless of what edges into the block exist.
2450 MachineBasicBlock *LoopHeader = L->getHeader();
2451 BlockFrequency LoopHeaderFreq = MBFI->getBlockFreq(LoopHeader);
2452 if (Freq < (LoopHeaderFreq * ColdProb))
2455 // Check for the existence of a non-layout predecessor which would benefit
2456 // from aligning this block.
2457 MachineBasicBlock *LayoutPred =
2458 &*std::prev(MachineFunction::iterator(ChainBB));
2460 // Force alignment if all the predecessors are jumps. We already checked
2461 // that the block isn't cold above.
2462 if (!LayoutPred->isSuccessor(ChainBB)) {
2463 ChainBB->setAlignment(Align);
2467 // Align this block if the layout predecessor's edge into this block is
2468 // cold relative to the block. When this is true, other predecessors make up
2469 // all of the hot entries into the block and thus alignment is likely to be
2471 BranchProbability LayoutProb =
2472 MBPI->getEdgeProbability(LayoutPred, ChainBB);
2473 BlockFrequency LayoutEdgeFreq = MBFI->getBlockFreq(LayoutPred) * LayoutProb;
2474 if (LayoutEdgeFreq <= (Freq * ColdProb))
2475 ChainBB->setAlignment(Align);
2479 /// Tail duplicate \p BB into (some) predecessors if profitable, repeating if
2480 /// it was duplicated into its chain predecessor and removed.
2481 /// \p BB - Basic block that may be duplicated.
2483 /// \p LPred - Chosen layout predecessor of \p BB.
2484 /// Updated to be the chain end if LPred is removed.
2485 /// \p Chain - Chain to which \p LPred belongs, and \p BB will belong.
2486 /// \p BlockFilter - Set of blocks that belong to the loop being laid out.
2487 /// Used to identify which blocks to update predecessor
2489 /// \p PrevUnplacedBlockIt - Iterator pointing to the last block that was
2490 /// chosen in the given order due to unnatural CFG
2491 /// only needed if \p BB is removed and
2492 /// \p PrevUnplacedBlockIt pointed to \p BB.
2493 /// @return true if \p BB was removed.
2494 bool MachineBlockPlacement::repeatedlyTailDuplicateBlock(
2495 MachineBasicBlock *BB, MachineBasicBlock *&LPred,
2496 const MachineBasicBlock *LoopHeaderBB,
2497 BlockChain &Chain, BlockFilterSet *BlockFilter,
2498 MachineFunction::iterator &PrevUnplacedBlockIt) {
2499 bool Removed, DuplicatedToLPred;
2500 bool DuplicatedToOriginalLPred;
2501 Removed = maybeTailDuplicateBlock(BB, LPred, Chain, BlockFilter,
2502 PrevUnplacedBlockIt,
2506 DuplicatedToOriginalLPred = DuplicatedToLPred;
2507 // Iteratively try to duplicate again. It can happen that a block that is
2508 // duplicated into is still small enough to be duplicated again.
2509 // No need to call markBlockSuccessors in this case, as the blocks being
2510 // duplicated from here on are already scheduled.
2511 // Note that DuplicatedToLPred always implies Removed.
2512 while (DuplicatedToLPred) {
2513 assert (Removed && "Block must have been removed to be duplicated into its "
2514 "layout predecessor.");
2515 MachineBasicBlock *DupBB, *DupPred;
2516 // The removal callback causes Chain.end() to be updated when a block is
2517 // removed. On the first pass through the loop, the chain end should be the
2518 // same as it was on function entry. On subsequent passes, because we are
2519 // duplicating the block at the end of the chain, if it is removed the
2520 // chain will have shrunk by one block.
2521 BlockChain::iterator ChainEnd = Chain.end();
2522 DupBB = *(--ChainEnd);
2523 // Now try to duplicate again.
2524 if (ChainEnd == Chain.begin())
2526 DupPred = *std::prev(ChainEnd);
2527 Removed = maybeTailDuplicateBlock(DupBB, DupPred, Chain, BlockFilter,
2528 PrevUnplacedBlockIt,
2531 // If BB was duplicated into LPred, it is now scheduled. But because it was
2532 // removed, markChainSuccessors won't be called for its chain. Instead we
2533 // call markBlockSuccessors for LPred to achieve the same effect. This must go
2534 // at the end because repeating the tail duplication can increase the number
2535 // of unscheduled predecessors.
2536 LPred = *std::prev(Chain.end());
2537 if (DuplicatedToOriginalLPred)
2538 markBlockSuccessors(Chain, LPred, LoopHeaderBB, BlockFilter);
2542 /// Tail duplicate \p BB into (some) predecessors if profitable.
2543 /// \p BB - Basic block that may be duplicated
2544 /// \p LPred - Chosen layout predecessor of \p BB
2545 /// \p Chain - Chain to which \p LPred belongs, and \p BB will belong.
2546 /// \p BlockFilter - Set of blocks that belong to the loop being laid out.
2547 /// Used to identify which blocks to update predecessor
2549 /// \p PrevUnplacedBlockIt - Iterator pointing to the last block that was
2550 /// chosen in the given order due to unnatural CFG
2551 /// only needed if \p BB is removed and
2552 /// \p PrevUnplacedBlockIt pointed to \p BB.
2553 /// \p DuplicatedToLPred - True if the block was duplicated into LPred. Will
2554 /// only be true if the block was removed.
2555 /// \return - True if the block was duplicated into all preds and removed.
2556 bool MachineBlockPlacement::maybeTailDuplicateBlock(
2557 MachineBasicBlock *BB, MachineBasicBlock *LPred,
2558 BlockChain &Chain, BlockFilterSet *BlockFilter,
2559 MachineFunction::iterator &PrevUnplacedBlockIt,
2560 bool &DuplicatedToLPred) {
2561 DuplicatedToLPred = false;
2562 if (!shouldTailDuplicate(BB))
2565 DEBUG(dbgs() << "Redoing tail duplication for Succ#"
2566 << BB->getNumber() << "\n");
2568 // This has to be a callback because none of it can be done after
2570 bool Removed = false;
2571 auto RemovalCallback =
2572 [&](MachineBasicBlock *RemBB) {
2573 // Signal to outer function
2576 // Conservative default.
2577 bool InWorkList = true;
2578 // Remove from the Chain and Chain Map
2579 if (BlockToChain.count(RemBB)) {
2580 BlockChain *Chain = BlockToChain[RemBB];
2581 InWorkList = Chain->UnscheduledPredecessors == 0;
2582 Chain->remove(RemBB);
2583 BlockToChain.erase(RemBB);
2586 // Handle the unplaced block iterator
2587 if (&(*PrevUnplacedBlockIt) == RemBB) {
2588 PrevUnplacedBlockIt++;
2591 // Handle the Work Lists
2593 SmallVectorImpl<MachineBasicBlock *> &RemoveList = BlockWorkList;
2594 if (RemBB->isEHPad())
2595 RemoveList = EHPadWorkList;
2597 remove_if(RemoveList,
2598 [RemBB](MachineBasicBlock *BB) {return BB == RemBB;}),
2602 // Handle the filter set
2604 BlockFilter->remove(RemBB);
2607 // Remove the block from loop info.
2608 MLI->removeBlock(RemBB);
2609 if (RemBB == PreferredLoopExit)
2610 PreferredLoopExit = nullptr;
2612 DEBUG(dbgs() << "TailDuplicator deleted block: "
2613 << getBlockName(RemBB) << "\n");
2615 auto RemovalCallbackRef =
2616 llvm::function_ref<void(MachineBasicBlock*)>(RemovalCallback);
2618 SmallVector<MachineBasicBlock *, 8> DuplicatedPreds;
2619 bool IsSimple = TailDup.isSimpleBB(BB);
2620 TailDup.tailDuplicateAndUpdate(IsSimple, BB, LPred,
2621 &DuplicatedPreds, &RemovalCallbackRef);
2623 // Update UnscheduledPredecessors to reflect tail-duplication.
2624 DuplicatedToLPred = false;
2625 for (MachineBasicBlock *Pred : DuplicatedPreds) {
2626 // We're only looking for unscheduled predecessors that match the filter.
2627 BlockChain* PredChain = BlockToChain[Pred];
2629 DuplicatedToLPred = true;
2630 if (Pred == LPred || (BlockFilter && !BlockFilter->count(Pred))
2631 || PredChain == &Chain)
2633 for (MachineBasicBlock *NewSucc : Pred->successors()) {
2634 if (BlockFilter && !BlockFilter->count(NewSucc))
2636 BlockChain *NewChain = BlockToChain[NewSucc];
2637 if (NewChain != &Chain && NewChain != PredChain)
2638 NewChain->UnscheduledPredecessors++;
2644 bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &MF) {
2645 if (skipFunction(*MF.getFunction()))
2648 // Check for single-block functions and skip them.
2649 if (std::next(MF.begin()) == MF.end())
2653 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
2654 MBFI = llvm::make_unique<BranchFolder::MBFIWrapper>(
2655 getAnalysis<MachineBlockFrequencyInfo>());
2656 MLI = &getAnalysis<MachineLoopInfo>();
2657 TII = MF.getSubtarget().getInstrInfo();
2658 TLI = MF.getSubtarget().getTargetLowering();
2661 // Initialize PreferredLoopExit to nullptr here since it may never be set if
2662 // there are no MachineLoops.
2663 PreferredLoopExit = nullptr;
2665 assert(BlockToChain.empty() &&
2666 "BlockToChain map should be empty before starting placement.");
2667 assert(ComputedEdges.empty() &&
2668 "Computed Edge map should be empty before starting placement.");
2670 unsigned TailDupSize = TailDupPlacementThreshold;
2671 // If only the aggressive threshold is explicitly set, use it.
2672 if (TailDupPlacementAggressiveThreshold.getNumOccurrences() != 0 &&
2673 TailDupPlacementThreshold.getNumOccurrences() == 0)
2674 TailDupSize = TailDupPlacementAggressiveThreshold;
2676 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
2677 // For agressive optimization, we can adjust some thresholds to be less
2679 if (PassConfig->getOptLevel() >= CodeGenOpt::Aggressive) {
2680 // At O3 we should be more willing to copy blocks for tail duplication. This
2681 // increases size pressure, so we only do it at O3
2682 // Do this unless only the regular threshold is explicitly set.
2683 if (TailDupPlacementThreshold.getNumOccurrences() == 0 ||
2684 TailDupPlacementAggressiveThreshold.getNumOccurrences() != 0)
2685 TailDupSize = TailDupPlacementAggressiveThreshold;
2688 if (TailDupPlacement) {
2689 MPDT = &getAnalysis<MachinePostDominatorTree>();
2690 if (MF.getFunction()->optForSize())
2692 TailDup.initMF(MF, MBPI, /* LayoutMode */ true, TailDupSize);
2693 precomputeTriangleChains();
2698 // Changing the layout can create new tail merging opportunities.
2699 // TailMerge can create jump into if branches that make CFG irreducible for
2700 // HW that requires structured CFG.
2701 bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
2702 PassConfig->getEnableTailMerge() &&
2703 BranchFoldPlacement;
2704 // No tail merging opportunities if the block number is less than four.
2705 if (MF.size() > 3 && EnableTailMerge) {
2706 unsigned TailMergeSize = TailDupSize + 1;
2707 BranchFolder BF(/*EnableTailMerge=*/true, /*CommonHoist=*/false, *MBFI,
2708 *MBPI, TailMergeSize);
2710 if (BF.OptimizeFunction(MF, TII, MF.getSubtarget().getRegisterInfo(),
2711 getAnalysisIfAvailable<MachineModuleInfo>(), MLI,
2712 /*AfterBlockPlacement=*/true)) {
2713 // Redo the layout if tail merging creates/removes/moves blocks.
2714 BlockToChain.clear();
2715 ComputedEdges.clear();
2716 // Must redo the post-dominator tree if blocks were changed.
2718 MPDT->runOnMachineFunction(MF);
2719 ChainAllocator.DestroyAll();
2727 BlockToChain.clear();
2728 ComputedEdges.clear();
2729 ChainAllocator.DestroyAll();
2732 // Align all of the blocks in the function to a specific alignment.
2733 for (MachineBasicBlock &MBB : MF)
2734 MBB.setAlignment(AlignAllBlock);
2735 else if (AlignAllNonFallThruBlocks) {
2736 // Align all of the blocks that have no fall-through predecessors to a
2737 // specific alignment.
2738 for (auto MBI = std::next(MF.begin()), MBE = MF.end(); MBI != MBE; ++MBI) {
2739 auto LayoutPred = std::prev(MBI);
2740 if (!LayoutPred->isSuccessor(&*MBI))
2741 MBI->setAlignment(AlignAllNonFallThruBlocks);
2744 if (ViewBlockLayoutWithBFI != GVDT_None &&
2745 (ViewBlockFreqFuncName.empty() ||
2746 F->getFunction()->getName().equals(ViewBlockFreqFuncName))) {
2747 MBFI->view("MBP." + MF.getName(), false);
2751 // We always return true as we have no way to track whether the final order
2752 // differs from the original order.
2757 /// \brief A pass to compute block placement statistics.
2759 /// A separate pass to compute interesting statistics for evaluating block
2760 /// placement. This is separate from the actual placement pass so that they can
2761 /// be computed in the absence of any placement transformations or when using
2762 /// alternative placement strategies.
2763 class MachineBlockPlacementStats : public MachineFunctionPass {
2764 /// \brief A handle to the branch probability pass.
2765 const MachineBranchProbabilityInfo *MBPI;
2767 /// \brief A handle to the function-wide block frequency pass.
2768 const MachineBlockFrequencyInfo *MBFI;
2771 static char ID; // Pass identification, replacement for typeid
2772 MachineBlockPlacementStats() : MachineFunctionPass(ID) {
2773 initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry());
2776 bool runOnMachineFunction(MachineFunction &F) override;
2778 void getAnalysisUsage(AnalysisUsage &AU) const override {
2779 AU.addRequired<MachineBranchProbabilityInfo>();
2780 AU.addRequired<MachineBlockFrequencyInfo>();
2781 AU.setPreservesAll();
2782 MachineFunctionPass::getAnalysisUsage(AU);
2787 char MachineBlockPlacementStats::ID = 0;
2788 char &llvm::MachineBlockPlacementStatsID = MachineBlockPlacementStats::ID;
2789 INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats",
2790 "Basic Block Placement Stats", false, false)
2791 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
2792 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
2793 INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats",
2794 "Basic Block Placement Stats", false, false)
2796 bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) {
2797 // Check for single-block functions and skip them.
2798 if (std::next(F.begin()) == F.end())
2801 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
2802 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
2804 for (MachineBasicBlock &MBB : F) {
2805 BlockFrequency BlockFreq = MBFI->getBlockFreq(&MBB);
2806 Statistic &NumBranches =
2807 (MBB.succ_size() > 1) ? NumCondBranches : NumUncondBranches;
2808 Statistic &BranchTakenFreq =
2809 (MBB.succ_size() > 1) ? CondBranchTakenFreq : UncondBranchTakenFreq;
2810 for (MachineBasicBlock *Succ : MBB.successors()) {
2811 // Skip if this successor is a fallthrough.
2812 if (MBB.isLayoutSuccessor(Succ))
2815 BlockFrequency EdgeFreq =
2816 BlockFreq * MBPI->getEdgeProbability(&MBB, Succ);
2818 BranchTakenFreq += EdgeFreq.getFrequency();