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 "BranchFolding.h"
29 #include "llvm/ADT/ArrayRef.h"
30 #include "llvm/ADT/DenseMap.h"
31 #include "llvm/ADT/STLExtras.h"
32 #include "llvm/ADT/SetVector.h"
33 #include "llvm/ADT/SmallPtrSet.h"
34 #include "llvm/ADT/SmallVector.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/Analysis/BlockFrequencyInfoImpl.h"
37 #include "llvm/CodeGen/MachineBasicBlock.h"
38 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
39 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
40 #include "llvm/CodeGen/MachineFunction.h"
41 #include "llvm/CodeGen/MachineFunctionPass.h"
42 #include "llvm/CodeGen/MachineLoopInfo.h"
43 #include "llvm/CodeGen/MachineModuleInfo.h"
44 #include "llvm/CodeGen/MachinePostDominators.h"
45 #include "llvm/CodeGen/TailDuplicator.h"
46 #include "llvm/CodeGen/TargetInstrInfo.h"
47 #include "llvm/CodeGen/TargetLowering.h"
48 #include "llvm/CodeGen/TargetPassConfig.h"
49 #include "llvm/CodeGen/TargetSubtargetInfo.h"
50 #include "llvm/IR/DebugLoc.h"
51 #include "llvm/IR/Function.h"
52 #include "llvm/Pass.h"
53 #include "llvm/Support/Allocator.h"
54 #include "llvm/Support/BlockFrequency.h"
55 #include "llvm/Support/BranchProbability.h"
56 #include "llvm/Support/CodeGen.h"
57 #include "llvm/Support/CommandLine.h"
58 #include "llvm/Support/Compiler.h"
59 #include "llvm/Support/Debug.h"
60 #include "llvm/Support/raw_ostream.h"
61 #include "llvm/Target/TargetMachine.h"
74 #define DEBUG_TYPE "block-placement"
76 STATISTIC(NumCondBranches, "Number of conditional branches");
77 STATISTIC(NumUncondBranches, "Number of unconditional branches");
78 STATISTIC(CondBranchTakenFreq,
79 "Potential frequency of taking conditional branches");
80 STATISTIC(UncondBranchTakenFreq,
81 "Potential frequency of taking unconditional branches");
83 static cl::opt<unsigned> AlignAllBlock("align-all-blocks",
84 cl::desc("Force the alignment of all "
85 "blocks in the function."),
86 cl::init(0), cl::Hidden);
88 static cl::opt<unsigned> AlignAllNonFallThruBlocks(
89 "align-all-nofallthru-blocks",
90 cl::desc("Force the alignment of all "
91 "blocks that have no fall-through predecessors (i.e. don't add "
92 "nops that are executed)."),
93 cl::init(0), cl::Hidden);
95 // FIXME: Find a good default for this flag and remove the flag.
96 static cl::opt<unsigned> ExitBlockBias(
97 "block-placement-exit-block-bias",
98 cl::desc("Block frequency percentage a loop exit block needs "
99 "over the original exit to be considered the new exit."),
100 cl::init(0), cl::Hidden);
103 // - Outlining: placement of a basic block outside the chain or hot path.
105 static cl::opt<unsigned> LoopToColdBlockRatio(
106 "loop-to-cold-block-ratio",
107 cl::desc("Outline loop blocks from loop chain if (frequency of loop) / "
108 "(frequency of block) is greater than this ratio"),
109 cl::init(5), cl::Hidden);
111 static cl::opt<bool> ForceLoopColdBlock(
112 "force-loop-cold-block",
113 cl::desc("Force outlining cold blocks from loops."),
114 cl::init(false), cl::Hidden);
117 PreciseRotationCost("precise-rotation-cost",
118 cl::desc("Model the cost of loop rotation more "
119 "precisely by using profile data."),
120 cl::init(false), cl::Hidden);
123 ForcePreciseRotationCost("force-precise-rotation-cost",
124 cl::desc("Force the use of precise cost "
125 "loop rotation strategy."),
126 cl::init(false), cl::Hidden);
128 static cl::opt<unsigned> MisfetchCost(
130 cl::desc("Cost that models the probabilistic risk of an instruction "
131 "misfetch due to a jump comparing to falling through, whose cost "
133 cl::init(1), cl::Hidden);
135 static cl::opt<unsigned> JumpInstCost("jump-inst-cost",
136 cl::desc("Cost of jump instructions."),
137 cl::init(1), cl::Hidden);
139 TailDupPlacement("tail-dup-placement",
140 cl::desc("Perform tail duplication during placement. "
141 "Creates more fallthrough opportunites in "
142 "outline branches."),
143 cl::init(true), cl::Hidden);
146 BranchFoldPlacement("branch-fold-placement",
147 cl::desc("Perform branch folding during placement. "
148 "Reduces code size."),
149 cl::init(true), cl::Hidden);
151 // Heuristic for tail duplication.
152 static cl::opt<unsigned> TailDupPlacementThreshold(
153 "tail-dup-placement-threshold",
154 cl::desc("Instruction cutoff for tail duplication during layout. "
155 "Tail merging during layout is forced to have a threshold "
156 "that won't conflict."), cl::init(2),
159 // Heuristic for aggressive tail duplication.
160 static cl::opt<unsigned> TailDupPlacementAggressiveThreshold(
161 "tail-dup-placement-aggressive-threshold",
162 cl::desc("Instruction cutoff for aggressive tail duplication during "
163 "layout. Used at -O3. Tail merging during layout is forced to "
164 "have a threshold that won't conflict."), cl::init(4),
167 // Heuristic for tail duplication.
168 static cl::opt<unsigned> TailDupPlacementPenalty(
169 "tail-dup-placement-penalty",
170 cl::desc("Cost penalty for blocks that can avoid breaking CFG by copying. "
171 "Copying can increase fallthrough, but it also increases icache "
172 "pressure. This parameter controls the penalty to account for that. "
173 "Percent as integer."),
177 // Heuristic for triangle chains.
178 static cl::opt<unsigned> TriangleChainCount(
179 "triangle-chain-count",
180 cl::desc("Number of triangle-shaped-CFG's that need to be in a row for the "
181 "triangle tail duplication heuristic to kick in. 0 to disable."),
185 extern cl::opt<unsigned> StaticLikelyProb;
186 extern cl::opt<unsigned> ProfileLikelyProb;
188 // Internal option used to control BFI display only after MBP pass.
189 // Defined in CodeGen/MachineBlockFrequencyInfo.cpp:
190 // -view-block-layout-with-bfi=
191 extern cl::opt<GVDAGType> ViewBlockLayoutWithBFI;
193 // Command line option to specify the name of the function for CFG dump
194 // Defined in Analysis/BlockFrequencyInfo.cpp: -view-bfi-func-name=
195 extern cl::opt<std::string> ViewBlockFreqFuncName;
201 /// \brief Type for our function-wide basic block -> block chain mapping.
202 using BlockToChainMapType = DenseMap<const MachineBasicBlock *, BlockChain *>;
204 /// \brief A chain of blocks which will be laid out contiguously.
206 /// This is the datastructure representing a chain of consecutive blocks that
207 /// are profitable to layout together in order to maximize fallthrough
208 /// probabilities and code locality. We also can use a block chain to represent
209 /// a sequence of basic blocks which have some external (correctness)
210 /// requirement for sequential layout.
212 /// Chains can be built around a single basic block and can be merged to grow
213 /// them. They participate in a block-to-chain mapping, which is updated
214 /// automatically as chains are merged together.
216 /// \brief The sequence of blocks belonging to this chain.
218 /// This is the sequence of blocks for a particular chain. These will be laid
219 /// out in-order within the function.
220 SmallVector<MachineBasicBlock *, 4> Blocks;
222 /// \brief A handle to the function-wide basic block to block chain mapping.
224 /// This is retained in each block chain to simplify the computation of child
225 /// block chains for SCC-formation and iteration. We store the edges to child
226 /// basic blocks, and map them back to their associated chains using this
228 BlockToChainMapType &BlockToChain;
231 /// \brief Construct a new BlockChain.
233 /// This builds a new block chain representing a single basic block in the
234 /// function. It also registers itself as the chain that block participates
235 /// in with the BlockToChain mapping.
236 BlockChain(BlockToChainMapType &BlockToChain, MachineBasicBlock *BB)
237 : Blocks(1, BB), BlockToChain(BlockToChain) {
238 assert(BB && "Cannot create a chain with a null basic block");
239 BlockToChain[BB] = this;
242 /// \brief Iterator over blocks within the chain.
243 using iterator = SmallVectorImpl<MachineBasicBlock *>::iterator;
244 using const_iterator = SmallVectorImpl<MachineBasicBlock *>::const_iterator;
246 /// \brief Beginning of blocks within the chain.
247 iterator begin() { return Blocks.begin(); }
248 const_iterator begin() const { return Blocks.begin(); }
250 /// \brief End of blocks within the chain.
251 iterator end() { return Blocks.end(); }
252 const_iterator end() const { return Blocks.end(); }
254 bool remove(MachineBasicBlock* BB) {
255 for(iterator i = begin(); i != end(); ++i) {
264 /// \brief Merge a block chain into this one.
266 /// This routine merges a block chain into this one. It takes care of forming
267 /// a contiguous sequence of basic blocks, updating the edge list, and
268 /// updating the block -> chain mapping. It does not free or tear down the
269 /// old chain, but the old chain's block list is no longer valid.
270 void merge(MachineBasicBlock *BB, BlockChain *Chain) {
271 assert(BB && "Can't merge a null block.");
272 assert(!Blocks.empty() && "Can't merge into an empty chain.");
274 // Fast path in case we don't have a chain already.
276 assert(!BlockToChain[BB] &&
277 "Passed chain is null, but BB has entry in BlockToChain.");
278 Blocks.push_back(BB);
279 BlockToChain[BB] = this;
283 assert(BB == *Chain->begin() && "Passed BB is not head of Chain.");
284 assert(Chain->begin() != Chain->end());
286 // Update the incoming blocks to point to this chain, and add them to the
288 for (MachineBasicBlock *ChainBB : *Chain) {
289 Blocks.push_back(ChainBB);
290 assert(BlockToChain[ChainBB] == Chain && "Incoming blocks not in chain.");
291 BlockToChain[ChainBB] = this;
296 /// \brief Dump the blocks in this chain.
297 LLVM_DUMP_METHOD void dump() {
298 for (MachineBasicBlock *MBB : *this)
303 /// \brief Count of predecessors of any block within the chain which have not
304 /// yet been scheduled. In general, we will delay scheduling this chain
305 /// until those predecessors are scheduled (or we find a sufficiently good
306 /// reason to override this heuristic.) Note that when forming loop chains,
307 /// blocks outside the loop are ignored and treated as if they were already
310 /// Note: This field is reinitialized multiple times - once for each loop,
311 /// and then once for the function as a whole.
312 unsigned UnscheduledPredecessors = 0;
315 class MachineBlockPlacement : public MachineFunctionPass {
316 /// \brief A type for a block filter set.
317 using BlockFilterSet = SmallSetVector<const MachineBasicBlock *, 16>;
319 /// Pair struct containing basic block and taildup profitiability
320 struct BlockAndTailDupResult {
321 MachineBasicBlock *BB;
325 /// Triple struct containing edge weight and the edge.
326 struct WeightedEdge {
327 BlockFrequency Weight;
328 MachineBasicBlock *Src;
329 MachineBasicBlock *Dest;
332 /// \brief work lists of blocks that are ready to be laid out
333 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
334 SmallVector<MachineBasicBlock *, 16> EHPadWorkList;
336 /// Edges that have already been computed as optimal.
337 DenseMap<const MachineBasicBlock *, BlockAndTailDupResult> ComputedEdges;
339 /// \brief Machine Function
342 /// \brief A handle to the branch probability pass.
343 const MachineBranchProbabilityInfo *MBPI;
345 /// \brief A handle to the function-wide block frequency pass.
346 std::unique_ptr<BranchFolder::MBFIWrapper> MBFI;
348 /// \brief A handle to the loop info.
349 MachineLoopInfo *MLI;
351 /// \brief Preferred loop exit.
352 /// Member variable for convenience. It may be removed by duplication deep
353 /// in the call stack.
354 MachineBasicBlock *PreferredLoopExit;
356 /// \brief A handle to the target's instruction info.
357 const TargetInstrInfo *TII;
359 /// \brief A handle to the target's lowering info.
360 const TargetLoweringBase *TLI;
362 /// \brief A handle to the post dominator tree.
363 MachinePostDominatorTree *MPDT;
365 /// \brief Duplicator used to duplicate tails during placement.
367 /// Placement decisions can open up new tail duplication opportunities, but
368 /// since tail duplication affects placement decisions of later blocks, it
369 /// must be done inline.
370 TailDuplicator TailDup;
372 /// \brief Allocator and owner of BlockChain structures.
374 /// We build BlockChains lazily while processing the loop structure of
375 /// a function. To reduce malloc traffic, we allocate them using this
376 /// slab-like allocator, and destroy them after the pass completes. An
377 /// important guarantee is that this allocator produces stable pointers to
379 SpecificBumpPtrAllocator<BlockChain> ChainAllocator;
381 /// \brief Function wide BasicBlock to BlockChain mapping.
383 /// This mapping allows efficiently moving from any given basic block to the
384 /// BlockChain it participates in, if any. We use it to, among other things,
385 /// allow implicitly defining edges between chains as the existing edges
386 /// between basic blocks.
387 DenseMap<const MachineBasicBlock *, BlockChain *> BlockToChain;
390 /// The set of basic blocks that have terminators that cannot be fully
391 /// analyzed. These basic blocks cannot be re-ordered safely by
392 /// MachineBlockPlacement, and we must preserve physical layout of these
393 /// blocks and their successors through the pass.
394 SmallPtrSet<MachineBasicBlock *, 4> BlocksWithUnanalyzableExits;
397 /// Decrease the UnscheduledPredecessors count for all blocks in chain, and
398 /// if the count goes to 0, add them to the appropriate work list.
399 void markChainSuccessors(
400 const BlockChain &Chain, const MachineBasicBlock *LoopHeaderBB,
401 const BlockFilterSet *BlockFilter = nullptr);
403 /// Decrease the UnscheduledPredecessors count for a single block, and
404 /// if the count goes to 0, add them to the appropriate work list.
405 void markBlockSuccessors(
406 const BlockChain &Chain, const MachineBasicBlock *BB,
407 const MachineBasicBlock *LoopHeaderBB,
408 const BlockFilterSet *BlockFilter = nullptr);
411 collectViableSuccessors(
412 const MachineBasicBlock *BB, const BlockChain &Chain,
413 const BlockFilterSet *BlockFilter,
414 SmallVector<MachineBasicBlock *, 4> &Successors);
415 bool shouldPredBlockBeOutlined(
416 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
417 const BlockChain &Chain, const BlockFilterSet *BlockFilter,
418 BranchProbability SuccProb, BranchProbability HotProb);
419 bool repeatedlyTailDuplicateBlock(
420 MachineBasicBlock *BB, MachineBasicBlock *&LPred,
421 const MachineBasicBlock *LoopHeaderBB,
422 BlockChain &Chain, BlockFilterSet *BlockFilter,
423 MachineFunction::iterator &PrevUnplacedBlockIt);
424 bool maybeTailDuplicateBlock(
425 MachineBasicBlock *BB, MachineBasicBlock *LPred,
426 BlockChain &Chain, BlockFilterSet *BlockFilter,
427 MachineFunction::iterator &PrevUnplacedBlockIt,
428 bool &DuplicatedToPred);
429 bool hasBetterLayoutPredecessor(
430 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
431 const BlockChain &SuccChain, BranchProbability SuccProb,
432 BranchProbability RealSuccProb, const BlockChain &Chain,
433 const BlockFilterSet *BlockFilter);
434 BlockAndTailDupResult selectBestSuccessor(
435 const MachineBasicBlock *BB, const BlockChain &Chain,
436 const BlockFilterSet *BlockFilter);
437 MachineBasicBlock *selectBestCandidateBlock(
438 const BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList);
439 MachineBasicBlock *getFirstUnplacedBlock(
440 const BlockChain &PlacedChain,
441 MachineFunction::iterator &PrevUnplacedBlockIt,
442 const BlockFilterSet *BlockFilter);
444 /// \brief Add a basic block to the work list if it is appropriate.
446 /// If the optional parameter BlockFilter is provided, only MBB
447 /// present in the set will be added to the worklist. If nullptr
448 /// is provided, no filtering occurs.
449 void fillWorkLists(const MachineBasicBlock *MBB,
450 SmallPtrSetImpl<BlockChain *> &UpdatedPreds,
451 const BlockFilterSet *BlockFilter);
453 void buildChain(const MachineBasicBlock *BB, BlockChain &Chain,
454 BlockFilterSet *BlockFilter = nullptr);
455 MachineBasicBlock *findBestLoopTop(
456 const MachineLoop &L, const BlockFilterSet &LoopBlockSet);
457 MachineBasicBlock *findBestLoopExit(
458 const MachineLoop &L, const BlockFilterSet &LoopBlockSet);
459 BlockFilterSet collectLoopBlockSet(const MachineLoop &L);
460 void buildLoopChains(const MachineLoop &L);
462 BlockChain &LoopChain, const MachineBasicBlock *ExitingBB,
463 const BlockFilterSet &LoopBlockSet);
464 void rotateLoopWithProfile(
465 BlockChain &LoopChain, const MachineLoop &L,
466 const BlockFilterSet &LoopBlockSet);
467 void buildCFGChains();
468 void optimizeBranches();
470 /// Returns true if a block should be tail-duplicated to increase fallthrough
472 bool shouldTailDuplicate(MachineBasicBlock *BB);
473 /// Check the edge frequencies to see if tail duplication will increase
475 bool isProfitableToTailDup(
476 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
477 BranchProbability AdjustedSumProb,
478 const BlockChain &Chain, const BlockFilterSet *BlockFilter);
480 /// Check for a trellis layout.
481 bool isTrellis(const MachineBasicBlock *BB,
482 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
483 const BlockChain &Chain, const BlockFilterSet *BlockFilter);
485 /// Get the best successor given a trellis layout.
486 BlockAndTailDupResult getBestTrellisSuccessor(
487 const MachineBasicBlock *BB,
488 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
489 BranchProbability AdjustedSumProb, const BlockChain &Chain,
490 const BlockFilterSet *BlockFilter);
492 /// Get the best pair of non-conflicting edges.
493 static std::pair<WeightedEdge, WeightedEdge> getBestNonConflictingEdges(
494 const MachineBasicBlock *BB,
495 MutableArrayRef<SmallVector<WeightedEdge, 8>> Edges);
497 /// Returns true if a block can tail duplicate into all unplaced
498 /// predecessors. Filters based on loop.
499 bool canTailDuplicateUnplacedPreds(
500 const MachineBasicBlock *BB, MachineBasicBlock *Succ,
501 const BlockChain &Chain, const BlockFilterSet *BlockFilter);
503 /// Find chains of triangles to tail-duplicate where a global analysis works,
504 /// but a local analysis would not find them.
505 void precomputeTriangleChains();
508 static char ID; // Pass identification, replacement for typeid
510 MachineBlockPlacement() : MachineFunctionPass(ID) {
511 initializeMachineBlockPlacementPass(*PassRegistry::getPassRegistry());
514 bool runOnMachineFunction(MachineFunction &F) override;
516 void getAnalysisUsage(AnalysisUsage &AU) const override {
517 AU.addRequired<MachineBranchProbabilityInfo>();
518 AU.addRequired<MachineBlockFrequencyInfo>();
519 if (TailDupPlacement)
520 AU.addRequired<MachinePostDominatorTree>();
521 AU.addRequired<MachineLoopInfo>();
522 AU.addRequired<TargetPassConfig>();
523 MachineFunctionPass::getAnalysisUsage(AU);
527 } // end anonymous namespace
529 char MachineBlockPlacement::ID = 0;
531 char &llvm::MachineBlockPlacementID = MachineBlockPlacement::ID;
533 INITIALIZE_PASS_BEGIN(MachineBlockPlacement, DEBUG_TYPE,
534 "Branch Probability Basic Block Placement", false, false)
535 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
536 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
537 INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTree)
538 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
539 INITIALIZE_PASS_END(MachineBlockPlacement, DEBUG_TYPE,
540 "Branch Probability Basic Block Placement", false, false)
543 /// \brief Helper to print the name of a MBB.
545 /// Only used by debug logging.
546 static std::string getBlockName(const MachineBasicBlock *BB) {
548 raw_string_ostream OS(Result);
549 OS << printMBBReference(*BB);
550 OS << " ('" << BB->getName() << "')";
556 /// \brief Mark a chain's successors as having one fewer preds.
558 /// When a chain is being merged into the "placed" chain, this routine will
559 /// quickly walk the successors of each block in the chain and mark them as
560 /// having one fewer active predecessor. It also adds any successors of this
561 /// chain which reach the zero-predecessor state to the appropriate worklist.
562 void MachineBlockPlacement::markChainSuccessors(
563 const BlockChain &Chain, const MachineBasicBlock *LoopHeaderBB,
564 const BlockFilterSet *BlockFilter) {
565 // Walk all the blocks in this chain, marking their successors as having
566 // a predecessor placed.
567 for (MachineBasicBlock *MBB : Chain) {
568 markBlockSuccessors(Chain, MBB, LoopHeaderBB, BlockFilter);
572 /// \brief Mark a single block's successors as having one fewer preds.
574 /// Under normal circumstances, this is only called by markChainSuccessors,
575 /// but if a block that was to be placed is completely tail-duplicated away,
576 /// and was duplicated into the chain end, we need to redo markBlockSuccessors
577 /// for just that block.
578 void MachineBlockPlacement::markBlockSuccessors(
579 const BlockChain &Chain, const MachineBasicBlock *MBB,
580 const MachineBasicBlock *LoopHeaderBB, const BlockFilterSet *BlockFilter) {
581 // Add any successors for which this is the only un-placed in-loop
582 // predecessor to the worklist as a viable candidate for CFG-neutral
583 // placement. No subsequent placement of this block will violate the CFG
584 // shape, so we get to use heuristics to choose a favorable placement.
585 for (MachineBasicBlock *Succ : MBB->successors()) {
586 if (BlockFilter && !BlockFilter->count(Succ))
588 BlockChain &SuccChain = *BlockToChain[Succ];
589 // Disregard edges within a fixed chain, or edges to the loop header.
590 if (&Chain == &SuccChain || Succ == LoopHeaderBB)
593 // This is a cross-chain edge that is within the loop, so decrement the
594 // loop predecessor count of the destination chain.
595 if (SuccChain.UnscheduledPredecessors == 0 ||
596 --SuccChain.UnscheduledPredecessors > 0)
599 auto *NewBB = *SuccChain.begin();
600 if (NewBB->isEHPad())
601 EHPadWorkList.push_back(NewBB);
603 BlockWorkList.push_back(NewBB);
607 /// This helper function collects the set of successors of block
608 /// \p BB that are allowed to be its layout successors, and return
609 /// the total branch probability of edges from \p BB to those
611 BranchProbability MachineBlockPlacement::collectViableSuccessors(
612 const MachineBasicBlock *BB, const BlockChain &Chain,
613 const BlockFilterSet *BlockFilter,
614 SmallVector<MachineBasicBlock *, 4> &Successors) {
615 // Adjust edge probabilities by excluding edges pointing to blocks that is
616 // either not in BlockFilter or is already in the current chain. Consider the
625 // Assume A->C is very hot (>90%), and C->D has a 50% probability, then after
626 // A->C is chosen as a fall-through, D won't be selected as a successor of C
627 // due to CFG constraint (the probability of C->D is not greater than
628 // HotProb to break topo-order). If we exclude E that is not in BlockFilter
629 // when calculating the probability of C->D, D will be selected and we
630 // will get A C D B as the layout of this loop.
631 auto AdjustedSumProb = BranchProbability::getOne();
632 for (MachineBasicBlock *Succ : BB->successors()) {
633 bool SkipSucc = false;
634 if (Succ->isEHPad() || (BlockFilter && !BlockFilter->count(Succ))) {
637 BlockChain *SuccChain = BlockToChain[Succ];
638 if (SuccChain == &Chain) {
640 } else if (Succ != *SuccChain->begin()) {
641 DEBUG(dbgs() << " " << getBlockName(Succ) << " -> Mid chain!\n");
646 AdjustedSumProb -= MBPI->getEdgeProbability(BB, Succ);
648 Successors.push_back(Succ);
651 return AdjustedSumProb;
654 /// The helper function returns the branch probability that is adjusted
655 /// or normalized over the new total \p AdjustedSumProb.
656 static BranchProbability
657 getAdjustedProbability(BranchProbability OrigProb,
658 BranchProbability AdjustedSumProb) {
659 BranchProbability SuccProb;
660 uint32_t SuccProbN = OrigProb.getNumerator();
661 uint32_t SuccProbD = AdjustedSumProb.getNumerator();
662 if (SuccProbN >= SuccProbD)
663 SuccProb = BranchProbability::getOne();
665 SuccProb = BranchProbability(SuccProbN, SuccProbD);
670 /// Check if \p BB has exactly the successors in \p Successors.
672 hasSameSuccessors(MachineBasicBlock &BB,
673 SmallPtrSetImpl<const MachineBasicBlock *> &Successors) {
674 if (BB.succ_size() != Successors.size())
676 // We don't want to count self-loops
677 if (Successors.count(&BB))
679 for (MachineBasicBlock *Succ : BB.successors())
680 if (!Successors.count(Succ))
685 /// Check if a block should be tail duplicated to increase fallthrough
687 /// \p BB Block to check.
688 bool MachineBlockPlacement::shouldTailDuplicate(MachineBasicBlock *BB) {
689 // Blocks with single successors don't create additional fallthrough
690 // opportunities. Don't duplicate them. TODO: When conditional exits are
691 // analyzable, allow them to be duplicated.
692 bool IsSimple = TailDup.isSimpleBB(BB);
694 if (BB->succ_size() == 1)
696 return TailDup.shouldTailDuplicate(IsSimple, *BB);
699 /// Compare 2 BlockFrequency's with a small penalty for \p A.
700 /// In order to be conservative, we apply a X% penalty to account for
701 /// increased icache pressure and static heuristics. For small frequencies
702 /// we use only the numerators to improve accuracy. For simplicity, we assume the
703 /// penalty is less than 100%
704 /// TODO(iteratee): Use 64-bit fixed point edge frequencies everywhere.
705 static bool greaterWithBias(BlockFrequency A, BlockFrequency B,
706 uint64_t EntryFreq) {
707 BranchProbability ThresholdProb(TailDupPlacementPenalty, 100);
708 BlockFrequency Gain = A - B;
709 return (Gain / ThresholdProb).getFrequency() >= EntryFreq;
712 /// Check the edge frequencies to see if tail duplication will increase
713 /// fallthroughs. It only makes sense to call this function when
714 /// \p Succ would not be chosen otherwise. Tail duplication of \p Succ is
715 /// always locally profitable if we would have picked \p Succ without
716 /// considering duplication.
717 bool MachineBlockPlacement::isProfitableToTailDup(
718 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
719 BranchProbability QProb,
720 const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
721 // We need to do a probability calculation to make sure this is profitable.
722 // First: does succ have a successor that post-dominates? This affects the
723 // calculation. The 2 relevant cases are:
738 // '=' : Branch taken for that CFG edge
739 // In the second case, Placing Succ while duplicating it into C prevents the
740 // fallthrough of Succ into either D or PDom, because they now have C as an
741 // unplaced predecessor
743 // Start by figuring out which case we fall into
744 MachineBasicBlock *PDom = nullptr;
745 SmallVector<MachineBasicBlock *, 4> SuccSuccs;
746 // Only scan the relevant successors
747 auto AdjustedSuccSumProb =
748 collectViableSuccessors(Succ, Chain, BlockFilter, SuccSuccs);
749 BranchProbability PProb = MBPI->getEdgeProbability(BB, Succ);
750 auto BBFreq = MBFI->getBlockFreq(BB);
751 auto SuccFreq = MBFI->getBlockFreq(Succ);
752 BlockFrequency P = BBFreq * PProb;
753 BlockFrequency Qout = BBFreq * QProb;
754 uint64_t EntryFreq = MBFI->getEntryFreq();
755 // If there are no more successors, it is profitable to copy, as it strictly
756 // increases fallthrough.
757 if (SuccSuccs.size() == 0)
758 return greaterWithBias(P, Qout, EntryFreq);
760 auto BestSuccSucc = BranchProbability::getZero();
761 // Find the PDom or the best Succ if no PDom exists.
762 for (MachineBasicBlock *SuccSucc : SuccSuccs) {
763 auto Prob = MBPI->getEdgeProbability(Succ, SuccSucc);
764 if (Prob > BestSuccSucc)
767 if (MPDT->dominates(SuccSucc, Succ)) {
772 // For the comparisons, we need to know Succ's best incoming edge that isn't
774 auto SuccBestPred = BlockFrequency(0);
775 for (MachineBasicBlock *SuccPred : Succ->predecessors()) {
776 if (SuccPred == Succ || SuccPred == BB
777 || BlockToChain[SuccPred] == &Chain
778 || (BlockFilter && !BlockFilter->count(SuccPred)))
780 auto Freq = MBFI->getBlockFreq(SuccPred)
781 * MBPI->getEdgeProbability(SuccPred, Succ);
782 if (Freq > SuccBestPred)
785 // Qin is Succ's best unplaced incoming edge that isn't BB
786 BlockFrequency Qin = SuccBestPred;
787 // If it doesn't have a post-dominating successor, here is the calculation:
799 // '=' : Branch taken for that CFG edge
800 // Cost in the first case is: P + V
801 // For this calculation, we always assume P > Qout. If Qout > P
802 // The result of this function will be ignored at the caller.
803 // Let F = SuccFreq - Qin
804 // Cost in the second case is: Qout + min(Qin, F) * U + max(Qin, F) * V
806 if (PDom == nullptr || !Succ->isSuccessor(PDom)) {
807 BranchProbability UProb = BestSuccSucc;
808 BranchProbability VProb = AdjustedSuccSumProb - UProb;
809 BlockFrequency F = SuccFreq - Qin;
810 BlockFrequency V = SuccFreq * VProb;
811 BlockFrequency QinU = std::min(Qin, F) * UProb;
812 BlockFrequency BaseCost = P + V;
813 BlockFrequency DupCost = Qout + QinU + std::max(Qin, F) * VProb;
814 return greaterWithBias(BaseCost, DupCost, EntryFreq);
816 BranchProbability UProb = MBPI->getEdgeProbability(Succ, PDom);
817 BranchProbability VProb = AdjustedSuccSumProb - UProb;
818 BlockFrequency U = SuccFreq * UProb;
819 BlockFrequency V = SuccFreq * VProb;
820 BlockFrequency F = SuccFreq - Qin;
821 // If there is a post-dominating successor, here is the calculation:
823 // | \Qout | \ | \Qout | \
825 // = C' |P C = C' |P C
826 // | /Qin | | | /Qin | |
827 // | / | C' (+Succ) | / | C' (+Succ)
828 // Succ Succ /| Succ Succ /|
829 // | \ V | \/ | | \ V | \/ |
830 // |U \ |U /\ =? |U = |U /\ |
831 // = D = = =?| | D | = =|
836 // '=' : Branch taken for that CFG edge
837 // The cost for taken branches in the first case is P + U
838 // Let F = SuccFreq - Qin
839 // The cost in the second case (assuming independence), given the layout:
840 // BB, Succ, (C+Succ), D, Dom or the layout:
841 // BB, Succ, D, Dom, (C+Succ)
842 // is Qout + max(F, Qin) * U + min(F, Qin)
843 // compare P + U vs Qout + P * U + Qin.
845 // The 3rd and 4th cases cover when Dom would be chosen to follow Succ.
847 // For the 3rd case, the cost is P + 2 * V
848 // For the 4th case, the cost is Qout + min(Qin, F) * U + max(Qin, F) * V + V
849 // We choose 4 over 3 when (P + V) > Qout + min(Qin, F) * U + max(Qin, F) * V
850 if (UProb > AdjustedSuccSumProb / 2 &&
851 !hasBetterLayoutPredecessor(Succ, PDom, *BlockToChain[PDom], UProb, UProb,
854 return greaterWithBias(
855 (P + V), (Qout + std::max(Qin, F) * VProb + std::min(Qin, F) * UProb),
858 return greaterWithBias((P + U),
859 (Qout + std::min(Qin, F) * AdjustedSuccSumProb +
860 std::max(Qin, F) * UProb),
864 /// Check for a trellis layout. \p BB is the upper part of a trellis if its
865 /// successors form the lower part of a trellis. A successor set S forms the
866 /// lower part of a trellis if all of the predecessors of S are either in S or
867 /// have all of S as successors. We ignore trellises where BB doesn't have 2
868 /// successors because for fewer than 2, it's trivial, and for 3 or greater they
869 /// are very uncommon and complex to compute optimally. Allowing edges within S
870 /// is not strictly a trellis, but the same algorithm works, so we allow it.
871 bool MachineBlockPlacement::isTrellis(
872 const MachineBasicBlock *BB,
873 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
874 const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
875 // Technically BB could form a trellis with branching factor higher than 2.
876 // But that's extremely uncommon.
877 if (BB->succ_size() != 2 || ViableSuccs.size() != 2)
880 SmallPtrSet<const MachineBasicBlock *, 2> Successors(BB->succ_begin(),
882 // To avoid reviewing the same predecessors twice.
883 SmallPtrSet<const MachineBasicBlock *, 8> SeenPreds;
885 for (MachineBasicBlock *Succ : ViableSuccs) {
887 for (auto SuccPred : Succ->predecessors()) {
888 // Allow triangle successors, but don't count them.
889 if (Successors.count(SuccPred)) {
890 // Make sure that it is actually a triangle.
891 for (MachineBasicBlock *CheckSucc : SuccPred->successors())
892 if (!Successors.count(CheckSucc))
896 const BlockChain *PredChain = BlockToChain[SuccPred];
897 if (SuccPred == BB || (BlockFilter && !BlockFilter->count(SuccPred)) ||
898 PredChain == &Chain || PredChain == BlockToChain[Succ])
901 // Perform the successor check only once.
902 if (!SeenPreds.insert(SuccPred).second)
904 if (!hasSameSuccessors(*SuccPred, Successors))
907 // If one of the successors has only BB as a predecessor, it is not a
915 /// Pick the highest total weight pair of edges that can both be laid out.
916 /// The edges in \p Edges[0] are assumed to have a different destination than
917 /// the edges in \p Edges[1]. Simple counting shows that the best pair is either
918 /// the individual highest weight edges to the 2 different destinations, or in
919 /// case of a conflict, one of them should be replaced with a 2nd best edge.
920 std::pair<MachineBlockPlacement::WeightedEdge,
921 MachineBlockPlacement::WeightedEdge>
922 MachineBlockPlacement::getBestNonConflictingEdges(
923 const MachineBasicBlock *BB,
924 MutableArrayRef<SmallVector<MachineBlockPlacement::WeightedEdge, 8>>
926 // Sort the edges, and then for each successor, find the best incoming
927 // predecessor. If the best incoming predecessors aren't the same,
928 // then that is clearly the best layout. If there is a conflict, one of the
929 // successors will have to fallthrough from the second best predecessor. We
930 // compare which combination is better overall.
932 // Sort for highest frequency.
933 auto Cmp = [](WeightedEdge A, WeightedEdge B) { return A.Weight > B.Weight; };
935 std::stable_sort(Edges[0].begin(), Edges[0].end(), Cmp);
936 std::stable_sort(Edges[1].begin(), Edges[1].end(), Cmp);
937 auto BestA = Edges[0].begin();
938 auto BestB = Edges[1].begin();
939 // Arrange for the correct answer to be in BestA and BestB
940 // If the 2 best edges don't conflict, the answer is already there.
941 if (BestA->Src == BestB->Src) {
942 // Compare the total fallthrough of (Best + Second Best) for both pairs
943 auto SecondBestA = std::next(BestA);
944 auto SecondBestB = std::next(BestB);
945 BlockFrequency BestAScore = BestA->Weight + SecondBestB->Weight;
946 BlockFrequency BestBScore = BestB->Weight + SecondBestA->Weight;
947 if (BestAScore < BestBScore)
952 // Arrange for the BB edge to be in BestA if it exists.
953 if (BestB->Src == BB)
954 std::swap(BestA, BestB);
955 return std::make_pair(*BestA, *BestB);
958 /// Get the best successor from \p BB based on \p BB being part of a trellis.
959 /// We only handle trellises with 2 successors, so the algorithm is
960 /// straightforward: Find the best pair of edges that don't conflict. We find
961 /// the best incoming edge for each successor in the trellis. If those conflict,
962 /// we consider which of them should be replaced with the second best.
963 /// Upon return the two best edges will be in \p BestEdges. If one of the edges
964 /// comes from \p BB, it will be in \p BestEdges[0]
965 MachineBlockPlacement::BlockAndTailDupResult
966 MachineBlockPlacement::getBestTrellisSuccessor(
967 const MachineBasicBlock *BB,
968 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
969 BranchProbability AdjustedSumProb, const BlockChain &Chain,
970 const BlockFilterSet *BlockFilter) {
972 BlockAndTailDupResult Result = {nullptr, false};
973 SmallPtrSet<const MachineBasicBlock *, 4> Successors(BB->succ_begin(),
976 // We assume size 2 because it's common. For general n, we would have to do
977 // the Hungarian algorithm, but it's not worth the complexity because more
978 // than 2 successors is fairly uncommon, and a trellis even more so.
979 if (Successors.size() != 2 || ViableSuccs.size() != 2)
982 // Collect the edge frequencies of all edges that form the trellis.
983 SmallVector<WeightedEdge, 8> Edges[2];
985 for (auto Succ : ViableSuccs) {
986 for (MachineBasicBlock *SuccPred : Succ->predecessors()) {
987 // Skip any placed predecessors that are not BB
989 if ((BlockFilter && !BlockFilter->count(SuccPred)) ||
990 BlockToChain[SuccPred] == &Chain ||
991 BlockToChain[SuccPred] == BlockToChain[Succ])
993 BlockFrequency EdgeFreq = MBFI->getBlockFreq(SuccPred) *
994 MBPI->getEdgeProbability(SuccPred, Succ);
995 Edges[SuccIndex].push_back({EdgeFreq, SuccPred, Succ});
1000 // Pick the best combination of 2 edges from all the edges in the trellis.
1001 WeightedEdge BestA, BestB;
1002 std::tie(BestA, BestB) = getBestNonConflictingEdges(BB, Edges);
1004 if (BestA.Src != BB) {
1005 // If we have a trellis, and BB doesn't have the best fallthrough edges,
1006 // we shouldn't choose any successor. We've already looked and there's a
1007 // better fallthrough edge for all the successors.
1008 DEBUG(dbgs() << "Trellis, but not one of the chosen edges.\n");
1012 // Did we pick the triangle edge? If tail-duplication is profitable, do
1013 // that instead. Otherwise merge the triangle edge now while we know it is
1015 if (BestA.Dest == BestB.Src) {
1016 // The edges are BB->Succ1->Succ2, and we're looking to see if BB->Succ2
1018 MachineBasicBlock *Succ1 = BestA.Dest;
1019 MachineBasicBlock *Succ2 = BestB.Dest;
1020 // Check to see if tail-duplication would be profitable.
1021 if (TailDupPlacement && shouldTailDuplicate(Succ2) &&
1022 canTailDuplicateUnplacedPreds(BB, Succ2, Chain, BlockFilter) &&
1023 isProfitableToTailDup(BB, Succ2, MBPI->getEdgeProbability(BB, Succ1),
1024 Chain, BlockFilter)) {
1025 DEBUG(BranchProbability Succ2Prob = getAdjustedProbability(
1026 MBPI->getEdgeProbability(BB, Succ2), AdjustedSumProb);
1027 dbgs() << " Selected: " << getBlockName(Succ2)
1028 << ", probability: " << Succ2Prob << " (Tail Duplicate)\n");
1030 Result.ShouldTailDup = true;
1034 // We have already computed the optimal edge for the other side of the
1036 ComputedEdges[BestB.Src] = { BestB.Dest, false };
1038 auto TrellisSucc = BestA.Dest;
1039 DEBUG(BranchProbability SuccProb = getAdjustedProbability(
1040 MBPI->getEdgeProbability(BB, TrellisSucc), AdjustedSumProb);
1041 dbgs() << " Selected: " << getBlockName(TrellisSucc)
1042 << ", probability: " << SuccProb << " (Trellis)\n");
1043 Result.BB = TrellisSucc;
1047 /// When the option TailDupPlacement is on, this method checks if the
1048 /// fallthrough candidate block \p Succ (of block \p BB) can be tail-duplicated
1049 /// into all of its unplaced, unfiltered predecessors, that are not BB.
1050 bool MachineBlockPlacement::canTailDuplicateUnplacedPreds(
1051 const MachineBasicBlock *BB, MachineBasicBlock *Succ,
1052 const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
1053 if (!shouldTailDuplicate(Succ))
1056 // For CFG checking.
1057 SmallPtrSet<const MachineBasicBlock *, 4> Successors(BB->succ_begin(),
1059 for (MachineBasicBlock *Pred : Succ->predecessors()) {
1060 // Make sure all unplaced and unfiltered predecessors can be
1061 // tail-duplicated into.
1062 // Skip any blocks that are already placed or not in this loop.
1063 if (Pred == BB || (BlockFilter && !BlockFilter->count(Pred))
1064 || BlockToChain[Pred] == &Chain)
1066 if (!TailDup.canTailDuplicate(Succ, Pred)) {
1067 if (Successors.size() > 1 && hasSameSuccessors(*Pred, Successors))
1068 // This will result in a trellis after tail duplication, so we don't
1069 // need to copy Succ into this predecessor. In the presence
1070 // of a trellis tail duplication can continue to be profitable.
1086 // After BB was duplicated into C, the layout looks like the one on the
1087 // right. BB and C now have the same successors. When considering
1088 // whether Succ can be duplicated into all its unplaced predecessors, we
1090 // We can do this because C already has a profitable fallthrough, namely
1091 // D. TODO(iteratee): ignore sufficiently cold predecessors for
1092 // duplication and for this test.
1094 // This allows trellises to be laid out in 2 separate chains
1095 // (A,B,Succ,...) and later (C,D,...) This is a reasonable heuristic
1096 // because it allows the creation of 2 fallthrough paths with links
1097 // between them, and we correctly identify the best layout for these
1098 // CFGs. We want to extend trellises that the user created in addition
1099 // to trellises created by tail-duplication, so we just look for the
1108 /// Find chains of triangles where we believe it would be profitable to
1109 /// tail-duplicate them all, but a local analysis would not find them.
1110 /// There are 3 ways this can be profitable:
1111 /// 1) The post-dominators marked 50% are actually taken 55% (This shrinks with
1113 /// 2) The chains are statically correlated. Branch probabilities have a very
1114 /// U-shaped distribution.
1115 /// [http://nrs.harvard.edu/urn-3:HUL.InstRepos:24015805]
1116 /// If the branches in a chain are likely to be from the same side of the
1117 /// distribution as their predecessor, but are independent at runtime, this
1118 /// transformation is profitable. (Because the cost of being wrong is a small
1119 /// fixed cost, unlike the standard triangle layout where the cost of being
1120 /// wrong scales with the # of triangles.)
1121 /// 3) The chains are dynamically correlated. If the probability that a previous
1122 /// branch was taken positively influences whether the next branch will be
1124 /// We believe that 2 and 3 are common enough to justify the small margin in 1.
1125 void MachineBlockPlacement::precomputeTriangleChains() {
1126 struct TriangleChain {
1127 std::vector<MachineBasicBlock *> Edges;
1129 TriangleChain(MachineBasicBlock *src, MachineBasicBlock *dst)
1130 : Edges({src, dst}) {}
1132 void append(MachineBasicBlock *dst) {
1133 assert(getKey()->isSuccessor(dst) &&
1134 "Attempting to append a block that is not a successor.");
1135 Edges.push_back(dst);
1138 unsigned count() const { return Edges.size() - 1; }
1140 MachineBasicBlock *getKey() const {
1141 return Edges.back();
1145 if (TriangleChainCount == 0)
1148 DEBUG(dbgs() << "Pre-computing triangle chains.\n");
1149 // Map from last block to the chain that contains it. This allows us to extend
1150 // chains as we find new triangles.
1151 DenseMap<const MachineBasicBlock *, TriangleChain> TriangleChainMap;
1152 for (MachineBasicBlock &BB : *F) {
1153 // If BB doesn't have 2 successors, it doesn't start a triangle.
1154 if (BB.succ_size() != 2)
1156 MachineBasicBlock *PDom = nullptr;
1157 for (MachineBasicBlock *Succ : BB.successors()) {
1158 if (!MPDT->dominates(Succ, &BB))
1163 // If BB doesn't have a post-dominating successor, it doesn't form a
1165 if (PDom == nullptr)
1167 // If PDom has a hint that it is low probability, skip this triangle.
1168 if (MBPI->getEdgeProbability(&BB, PDom) < BranchProbability(50, 100))
1170 // If PDom isn't eligible for duplication, this isn't the kind of triangle
1171 // we're looking for.
1172 if (!shouldTailDuplicate(PDom))
1174 bool CanTailDuplicate = true;
1175 // If PDom can't tail-duplicate into it's non-BB predecessors, then this
1176 // isn't the kind of triangle we're looking for.
1177 for (MachineBasicBlock* Pred : PDom->predecessors()) {
1180 if (!TailDup.canTailDuplicate(PDom, Pred)) {
1181 CanTailDuplicate = false;
1185 // If we can't tail-duplicate PDom to its predecessors, then skip this
1187 if (!CanTailDuplicate)
1190 // Now we have an interesting triangle. Insert it if it's not part of an
1192 // Note: This cannot be replaced with a call insert() or emplace() because
1193 // the find key is BB, but the insert/emplace key is PDom.
1194 auto Found = TriangleChainMap.find(&BB);
1195 // If it is, remove the chain from the map, grow it, and put it back in the
1196 // map with the end as the new key.
1197 if (Found != TriangleChainMap.end()) {
1198 TriangleChain Chain = std::move(Found->second);
1199 TriangleChainMap.erase(Found);
1201 TriangleChainMap.insert(std::make_pair(Chain.getKey(), std::move(Chain)));
1203 auto InsertResult = TriangleChainMap.try_emplace(PDom, &BB, PDom);
1204 assert(InsertResult.second && "Block seen twice.");
1209 // Iterating over a DenseMap is safe here, because the only thing in the body
1210 // of the loop is inserting into another DenseMap (ComputedEdges).
1211 // ComputedEdges is never iterated, so this doesn't lead to non-determinism.
1212 for (auto &ChainPair : TriangleChainMap) {
1213 TriangleChain &Chain = ChainPair.second;
1214 // Benchmarking has shown that due to branch correlation duplicating 2 or
1215 // more triangles is profitable, despite the calculations assuming
1217 if (Chain.count() < TriangleChainCount)
1219 MachineBasicBlock *dst = Chain.Edges.back();
1220 Chain.Edges.pop_back();
1221 for (MachineBasicBlock *src : reverse(Chain.Edges)) {
1222 DEBUG(dbgs() << "Marking edge: " << getBlockName(src) << "->" <<
1223 getBlockName(dst) << " as pre-computed based on triangles.\n");
1225 auto InsertResult = ComputedEdges.insert({src, {dst, true}});
1226 assert(InsertResult.second && "Block seen twice.");
1234 // When profile is not present, return the StaticLikelyProb.
1235 // When profile is available, we need to handle the triangle-shape CFG.
1236 static BranchProbability getLayoutSuccessorProbThreshold(
1237 const MachineBasicBlock *BB) {
1238 if (!BB->getParent()->getFunction().getEntryCount())
1239 return BranchProbability(StaticLikelyProb, 100);
1240 if (BB->succ_size() == 2) {
1241 const MachineBasicBlock *Succ1 = *BB->succ_begin();
1242 const MachineBasicBlock *Succ2 = *(BB->succ_begin() + 1);
1243 if (Succ1->isSuccessor(Succ2) || Succ2->isSuccessor(Succ1)) {
1244 /* See case 1 below for the cost analysis. For BB->Succ to
1245 * be taken with smaller cost, the following needs to hold:
1246 * Prob(BB->Succ) > 2 * Prob(BB->Pred)
1247 * So the threshold T in the calculation below
1248 * (1-T) * Prob(BB->Succ) > T * Prob(BB->Pred)
1249 * So T / (1 - T) = 2, Yielding T = 2/3
1250 * Also adding user specified branch bias, we have
1251 * T = (2/3)*(ProfileLikelyProb/50)
1252 * = (2*ProfileLikelyProb)/150)
1254 return BranchProbability(2 * ProfileLikelyProb, 150);
1257 return BranchProbability(ProfileLikelyProb, 100);
1260 /// Checks to see if the layout candidate block \p Succ has a better layout
1261 /// predecessor than \c BB. If yes, returns true.
1262 /// \p SuccProb: The probability adjusted for only remaining blocks.
1263 /// Only used for logging
1264 /// \p RealSuccProb: The un-adjusted probability.
1265 /// \p Chain: The chain that BB belongs to and Succ is being considered for.
1266 /// \p BlockFilter: if non-null, the set of blocks that make up the loop being
1268 bool MachineBlockPlacement::hasBetterLayoutPredecessor(
1269 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
1270 const BlockChain &SuccChain, BranchProbability SuccProb,
1271 BranchProbability RealSuccProb, const BlockChain &Chain,
1272 const BlockFilterSet *BlockFilter) {
1274 // There isn't a better layout when there are no unscheduled predecessors.
1275 if (SuccChain.UnscheduledPredecessors == 0)
1278 // There are two basic scenarios here:
1279 // -------------------------------------
1280 // Case 1: triangular shape CFG (if-then):
1287 // In this case, we are evaluating whether to select edge -> Succ, e.g.
1288 // set Succ as the layout successor of BB. Picking Succ as BB's
1289 // successor breaks the CFG constraints (FIXME: define these constraints).
1290 // With this layout, Pred BB
1291 // is forced to be outlined, so the overall cost will be cost of the
1292 // branch taken from BB to Pred, plus the cost of back taken branch
1293 // from Pred to Succ, as well as the additional cost associated
1294 // with the needed unconditional jump instruction from Pred To Succ.
1296 // The cost of the topological order layout is the taken branch cost
1297 // from BB to Succ, so to make BB->Succ a viable candidate, the following
1299 // 2 * freq(BB->Pred) * taken_branch_cost + unconditional_jump_cost
1300 // < freq(BB->Succ) * taken_branch_cost.
1301 // Ignoring unconditional jump cost, we get
1302 // freq(BB->Succ) > 2 * freq(BB->Pred), i.e.,
1303 // prob(BB->Succ) > 2 * prob(BB->Pred)
1305 // When real profile data is available, we can precisely compute the
1306 // probability threshold that is needed for edge BB->Succ to be considered.
1307 // Without profile data, the heuristic requires the branch bias to be
1308 // a lot larger to make sure the signal is very strong (e.g. 80% default).
1309 // -----------------------------------------------------------------
1310 // Case 2: diamond like CFG (if-then-else):
1319 // The current block is BB and edge BB->Succ is now being evaluated.
1320 // Note that edge S->BB was previously already selected because
1321 // prob(S->BB) > prob(S->Pred).
1322 // At this point, 2 blocks can be placed after BB: Pred or Succ. If we
1323 // choose Pred, we will have a topological ordering as shown on the left
1324 // in the picture below. If we choose Succ, we have the solution as shown
1333 // | Pred-- | Succ--
1335 // ---Succ ---Pred--
1337 // cost = freq(S->Pred) + freq(BB->Succ) cost = 2 * freq (S->Pred)
1338 // = freq(S->Pred) + freq(S->BB)
1340 // If we have profile data (i.e, branch probabilities can be trusted), the
1341 // cost (number of taken branches) with layout S->BB->Succ->Pred is 2 *
1342 // freq(S->Pred) while the cost of topo order is freq(S->Pred) + freq(S->BB).
1343 // We know Prob(S->BB) > Prob(S->Pred), so freq(S->BB) > freq(S->Pred), which
1344 // means the cost of topological order is greater.
1345 // When profile data is not available, however, we need to be more
1346 // conservative. If the branch prediction is wrong, breaking the topo-order
1347 // will actually yield a layout with large cost. For this reason, we need
1348 // strong biased branch at block S with Prob(S->BB) in order to select
1349 // BB->Succ. This is equivalent to looking the CFG backward with backward
1350 // edge: Prob(Succ->BB) needs to >= HotProb in order to be selected (without
1352 // --------------------------------------------------------------------------
1353 // Case 3: forked diamond
1365 // The current block is BB and edge BB->S1 is now being evaluated.
1366 // As above S->BB was already selected because
1367 // prob(S->BB) > prob(S->Pred). Assume that prob(BB->S1) >= prob(BB->S2).
1375 // | Pred----| | S1----
1377 // --(S1 or S2) ---Pred--
1381 // topo-cost = freq(S->Pred) + freq(BB->S1) + freq(BB->S2)
1382 // + min(freq(Pred->S1), freq(Pred->S2))
1383 // Non-topo-order cost:
1384 // non-topo-cost = 2 * freq(S->Pred) + freq(BB->S2).
1385 // To be conservative, we can assume that min(freq(Pred->S1), freq(Pred->S2))
1386 // is 0. Then the non topo layout is better when
1387 // freq(S->Pred) < freq(BB->S1).
1388 // This is exactly what is checked below.
1389 // Note there are other shapes that apply (Pred may not be a single block,
1390 // but they all fit this general pattern.)
1391 BranchProbability HotProb = getLayoutSuccessorProbThreshold(BB);
1393 // Make sure that a hot successor doesn't have a globally more
1394 // important predecessor.
1395 BlockFrequency CandidateEdgeFreq = MBFI->getBlockFreq(BB) * RealSuccProb;
1396 bool BadCFGConflict = false;
1398 for (MachineBasicBlock *Pred : Succ->predecessors()) {
1399 if (Pred == Succ || BlockToChain[Pred] == &SuccChain ||
1400 (BlockFilter && !BlockFilter->count(Pred)) ||
1401 BlockToChain[Pred] == &Chain ||
1402 // This check is redundant except for look ahead. This function is
1403 // called for lookahead by isProfitableToTailDup when BB hasn't been
1407 // Do backward checking.
1408 // For all cases above, we need a backward checking to filter out edges that
1409 // are not 'strongly' biased.
1413 // We select edge BB->Succ if
1414 // freq(BB->Succ) > freq(Succ) * HotProb
1415 // i.e. freq(BB->Succ) > freq(BB->Succ) * HotProb + freq(Pred->Succ) *
1417 // i.e. freq((BB->Succ) * (1 - HotProb) > freq(Pred->Succ) * HotProb
1418 // Case 1 is covered too, because the first equation reduces to:
1419 // prob(BB->Succ) > HotProb. (freq(Succ) = freq(BB) for a triangle)
1420 BlockFrequency PredEdgeFreq =
1421 MBFI->getBlockFreq(Pred) * MBPI->getEdgeProbability(Pred, Succ);
1422 if (PredEdgeFreq * HotProb >= CandidateEdgeFreq * HotProb.getCompl()) {
1423 BadCFGConflict = true;
1428 if (BadCFGConflict) {
1429 DEBUG(dbgs() << " Not a candidate: " << getBlockName(Succ) << " -> " << SuccProb
1430 << " (prob) (non-cold CFG conflict)\n");
1437 /// \brief Select the best successor for a block.
1439 /// This looks across all successors of a particular block and attempts to
1440 /// select the "best" one to be the layout successor. It only considers direct
1441 /// successors which also pass the block filter. It will attempt to avoid
1442 /// breaking CFG structure, but cave and break such structures in the case of
1443 /// very hot successor edges.
1445 /// \returns The best successor block found, or null if none are viable, along
1446 /// with a boolean indicating if tail duplication is necessary.
1447 MachineBlockPlacement::BlockAndTailDupResult
1448 MachineBlockPlacement::selectBestSuccessor(
1449 const MachineBasicBlock *BB, const BlockChain &Chain,
1450 const BlockFilterSet *BlockFilter) {
1451 const BranchProbability HotProb(StaticLikelyProb, 100);
1453 BlockAndTailDupResult BestSucc = { nullptr, false };
1454 auto BestProb = BranchProbability::getZero();
1456 SmallVector<MachineBasicBlock *, 4> Successors;
1457 auto AdjustedSumProb =
1458 collectViableSuccessors(BB, Chain, BlockFilter, Successors);
1460 DEBUG(dbgs() << "Selecting best successor for: " << getBlockName(BB) << "\n");
1462 // if we already precomputed the best successor for BB, return that if still
1464 auto FoundEdge = ComputedEdges.find(BB);
1465 if (FoundEdge != ComputedEdges.end()) {
1466 MachineBasicBlock *Succ = FoundEdge->second.BB;
1467 ComputedEdges.erase(FoundEdge);
1468 BlockChain *SuccChain = BlockToChain[Succ];
1469 if (BB->isSuccessor(Succ) && (!BlockFilter || BlockFilter->count(Succ)) &&
1470 SuccChain != &Chain && Succ == *SuccChain->begin())
1471 return FoundEdge->second;
1474 // if BB is part of a trellis, Use the trellis to determine the optimal
1475 // fallthrough edges
1476 if (isTrellis(BB, Successors, Chain, BlockFilter))
1477 return getBestTrellisSuccessor(BB, Successors, AdjustedSumProb, Chain,
1480 // For blocks with CFG violations, we may be able to lay them out anyway with
1481 // tail-duplication. We keep this vector so we can perform the probability
1482 // calculations the minimum number of times.
1483 SmallVector<std::tuple<BranchProbability, MachineBasicBlock *>, 4>
1485 for (MachineBasicBlock *Succ : Successors) {
1486 auto RealSuccProb = MBPI->getEdgeProbability(BB, Succ);
1487 BranchProbability SuccProb =
1488 getAdjustedProbability(RealSuccProb, AdjustedSumProb);
1490 BlockChain &SuccChain = *BlockToChain[Succ];
1491 // Skip the edge \c BB->Succ if block \c Succ has a better layout
1492 // predecessor that yields lower global cost.
1493 if (hasBetterLayoutPredecessor(BB, Succ, SuccChain, SuccProb, RealSuccProb,
1494 Chain, BlockFilter)) {
1495 // If tail duplication would make Succ profitable, place it.
1496 if (TailDupPlacement && shouldTailDuplicate(Succ))
1497 DupCandidates.push_back(std::make_tuple(SuccProb, Succ));
1502 dbgs() << " Candidate: " << getBlockName(Succ) << ", probability: "
1504 << (SuccChain.UnscheduledPredecessors != 0 ? " (CFG break)" : "")
1507 if (BestSucc.BB && BestProb >= SuccProb) {
1508 DEBUG(dbgs() << " Not the best candidate, continuing\n");
1512 DEBUG(dbgs() << " Setting it as best candidate\n");
1514 BestProb = SuccProb;
1516 // Handle the tail duplication candidates in order of decreasing probability.
1517 // Stop at the first one that is profitable. Also stop if they are less
1518 // profitable than BestSucc. Position is important because we preserve it and
1519 // prefer first best match. Here we aren't comparing in order, so we capture
1520 // the position instead.
1521 if (DupCandidates.size() != 0) {
1523 [](const std::tuple<BranchProbability, MachineBasicBlock *> &a,
1524 const std::tuple<BranchProbability, MachineBasicBlock *> &b) {
1525 return std::get<0>(a) > std::get<0>(b);
1527 std::stable_sort(DupCandidates.begin(), DupCandidates.end(), cmp);
1529 for(auto &Tup : DupCandidates) {
1530 BranchProbability DupProb;
1531 MachineBasicBlock *Succ;
1532 std::tie(DupProb, Succ) = Tup;
1533 if (DupProb < BestProb)
1535 if (canTailDuplicateUnplacedPreds(BB, Succ, Chain, BlockFilter)
1536 && (isProfitableToTailDup(BB, Succ, BestProb, Chain, BlockFilter))) {
1538 dbgs() << " Candidate: " << getBlockName(Succ) << ", probability: "
1540 << " (Tail Duplicate)\n");
1542 BestSucc.ShouldTailDup = true;
1548 DEBUG(dbgs() << " Selected: " << getBlockName(BestSucc.BB) << "\n");
1553 /// \brief Select the best block from a worklist.
1555 /// This looks through the provided worklist as a list of candidate basic
1556 /// blocks and select the most profitable one to place. The definition of
1557 /// profitable only really makes sense in the context of a loop. This returns
1558 /// the most frequently visited block in the worklist, which in the case of
1559 /// a loop, is the one most desirable to be physically close to the rest of the
1560 /// loop body in order to improve i-cache behavior.
1562 /// \returns The best block found, or null if none are viable.
1563 MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock(
1564 const BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList) {
1565 // Once we need to walk the worklist looking for a candidate, cleanup the
1566 // worklist of already placed entries.
1567 // FIXME: If this shows up on profiles, it could be folded (at the cost of
1568 // some code complexity) into the loop below.
1569 WorkList.erase(llvm::remove_if(WorkList,
1570 [&](MachineBasicBlock *BB) {
1571 return BlockToChain.lookup(BB) == &Chain;
1575 if (WorkList.empty())
1578 bool IsEHPad = WorkList[0]->isEHPad();
1580 MachineBasicBlock *BestBlock = nullptr;
1581 BlockFrequency BestFreq;
1582 for (MachineBasicBlock *MBB : WorkList) {
1583 assert(MBB->isEHPad() == IsEHPad &&
1584 "EHPad mismatch between block and work list.");
1586 BlockChain &SuccChain = *BlockToChain[MBB];
1587 if (&SuccChain == &Chain)
1590 assert(SuccChain.UnscheduledPredecessors == 0 &&
1591 "Found CFG-violating block");
1593 BlockFrequency CandidateFreq = MBFI->getBlockFreq(MBB);
1594 DEBUG(dbgs() << " " << getBlockName(MBB) << " -> ";
1595 MBFI->printBlockFreq(dbgs(), CandidateFreq) << " (freq)\n");
1597 // For ehpad, we layout the least probable first as to avoid jumping back
1598 // from least probable landingpads to more probable ones.
1600 // FIXME: Using probability is probably (!) not the best way to achieve
1601 // this. We should probably have a more principled approach to layout
1604 // The goal is to get:
1606 // +--------------------------+
1608 // InnerLp -> InnerCleanup OuterLp -> OuterCleanup -> Resume
1612 // +-------------------------------------+
1614 // OuterLp -> OuterCleanup -> Resume InnerLp -> InnerCleanup
1615 if (BestBlock && (IsEHPad ^ (BestFreq >= CandidateFreq)))
1619 BestFreq = CandidateFreq;
1625 /// \brief Retrieve the first unplaced basic block.
1627 /// This routine is called when we are unable to use the CFG to walk through
1628 /// all of the basic blocks and form a chain due to unnatural loops in the CFG.
1629 /// We walk through the function's blocks in order, starting from the
1630 /// LastUnplacedBlockIt. We update this iterator on each call to avoid
1631 /// re-scanning the entire sequence on repeated calls to this routine.
1632 MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock(
1633 const BlockChain &PlacedChain,
1634 MachineFunction::iterator &PrevUnplacedBlockIt,
1635 const BlockFilterSet *BlockFilter) {
1636 for (MachineFunction::iterator I = PrevUnplacedBlockIt, E = F->end(); I != E;
1638 if (BlockFilter && !BlockFilter->count(&*I))
1640 if (BlockToChain[&*I] != &PlacedChain) {
1641 PrevUnplacedBlockIt = I;
1642 // Now select the head of the chain to which the unplaced block belongs
1643 // as the block to place. This will force the entire chain to be placed,
1644 // and satisfies the requirements of merging chains.
1645 return *BlockToChain[&*I]->begin();
1651 void MachineBlockPlacement::fillWorkLists(
1652 const MachineBasicBlock *MBB,
1653 SmallPtrSetImpl<BlockChain *> &UpdatedPreds,
1654 const BlockFilterSet *BlockFilter = nullptr) {
1655 BlockChain &Chain = *BlockToChain[MBB];
1656 if (!UpdatedPreds.insert(&Chain).second)
1660 Chain.UnscheduledPredecessors == 0 &&
1661 "Attempting to place block with unscheduled predecessors in worklist.");
1662 for (MachineBasicBlock *ChainBB : Chain) {
1663 assert(BlockToChain[ChainBB] == &Chain &&
1664 "Block in chain doesn't match BlockToChain map.");
1665 for (MachineBasicBlock *Pred : ChainBB->predecessors()) {
1666 if (BlockFilter && !BlockFilter->count(Pred))
1668 if (BlockToChain[Pred] == &Chain)
1670 ++Chain.UnscheduledPredecessors;
1674 if (Chain.UnscheduledPredecessors != 0)
1677 MachineBasicBlock *BB = *Chain.begin();
1679 EHPadWorkList.push_back(BB);
1681 BlockWorkList.push_back(BB);
1684 void MachineBlockPlacement::buildChain(
1685 const MachineBasicBlock *HeadBB, BlockChain &Chain,
1686 BlockFilterSet *BlockFilter) {
1687 assert(HeadBB && "BB must not be null.\n");
1688 assert(BlockToChain[HeadBB] == &Chain && "BlockToChainMap mis-match.\n");
1689 MachineFunction::iterator PrevUnplacedBlockIt = F->begin();
1691 const MachineBasicBlock *LoopHeaderBB = HeadBB;
1692 markChainSuccessors(Chain, LoopHeaderBB, BlockFilter);
1693 MachineBasicBlock *BB = *std::prev(Chain.end());
1695 assert(BB && "null block found at end of chain in loop.");
1696 assert(BlockToChain[BB] == &Chain && "BlockToChainMap mis-match in loop.");
1697 assert(*std::prev(Chain.end()) == BB && "BB Not found at end of chain.");
1700 // Look for the best viable successor if there is one to place immediately
1701 // after this block.
1702 auto Result = selectBestSuccessor(BB, Chain, BlockFilter);
1703 MachineBasicBlock* BestSucc = Result.BB;
1704 bool ShouldTailDup = Result.ShouldTailDup;
1705 if (TailDupPlacement)
1706 ShouldTailDup |= (BestSucc && shouldTailDuplicate(BestSucc));
1708 // If an immediate successor isn't available, look for the best viable
1709 // block among those we've identified as not violating the loop's CFG at
1710 // this point. This won't be a fallthrough, but it will increase locality.
1712 BestSucc = selectBestCandidateBlock(Chain, BlockWorkList);
1714 BestSucc = selectBestCandidateBlock(Chain, EHPadWorkList);
1717 BestSucc = getFirstUnplacedBlock(Chain, PrevUnplacedBlockIt, BlockFilter);
1721 DEBUG(dbgs() << "Unnatural loop CFG detected, forcibly merging the "
1722 "layout successor until the CFG reduces\n");
1725 // Placement may have changed tail duplication opportunities.
1726 // Check for that now.
1727 if (TailDupPlacement && BestSucc && ShouldTailDup) {
1728 // If the chosen successor was duplicated into all its predecessors,
1729 // don't bother laying it out, just go round the loop again with BB as
1731 if (repeatedlyTailDuplicateBlock(BestSucc, BB, LoopHeaderBB, Chain,
1732 BlockFilter, PrevUnplacedBlockIt))
1736 // Place this block, updating the datastructures to reflect its placement.
1737 BlockChain &SuccChain = *BlockToChain[BestSucc];
1738 // Zero out UnscheduledPredecessors for the successor we're about to merge in case
1739 // we selected a successor that didn't fit naturally into the CFG.
1740 SuccChain.UnscheduledPredecessors = 0;
1741 DEBUG(dbgs() << "Merging from " << getBlockName(BB) << " to "
1742 << getBlockName(BestSucc) << "\n");
1743 markChainSuccessors(SuccChain, LoopHeaderBB, BlockFilter);
1744 Chain.merge(BestSucc, &SuccChain);
1745 BB = *std::prev(Chain.end());
1748 DEBUG(dbgs() << "Finished forming chain for header block "
1749 << getBlockName(*Chain.begin()) << "\n");
1752 /// \brief Find the best loop top block for layout.
1754 /// Look for a block which is strictly better than the loop header for laying
1755 /// out at the top of the loop. This looks for one and only one pattern:
1756 /// a latch block with no conditional exit. This block will cause a conditional
1757 /// jump around it or will be the bottom of the loop if we lay it out in place,
1758 /// but if it it doesn't end up at the bottom of the loop for any reason,
1759 /// rotation alone won't fix it. Because such a block will always result in an
1760 /// unconditional jump (for the backedge) rotating it in front of the loop
1761 /// header is always profitable.
1763 MachineBlockPlacement::findBestLoopTop(const MachineLoop &L,
1764 const BlockFilterSet &LoopBlockSet) {
1765 // Placing the latch block before the header may introduce an extra branch
1766 // that skips this block the first time the loop is executed, which we want
1767 // to avoid when optimising for size.
1768 // FIXME: in theory there is a case that does not introduce a new branch,
1769 // i.e. when the layout predecessor does not fallthrough to the loop header.
1770 // In practice this never happens though: there always seems to be a preheader
1771 // that can fallthrough and that is also placed before the header.
1772 if (F->getFunction().optForSize())
1773 return L.getHeader();
1775 // Check that the header hasn't been fused with a preheader block due to
1776 // crazy branches. If it has, we need to start with the header at the top to
1777 // prevent pulling the preheader into the loop body.
1778 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
1779 if (!LoopBlockSet.count(*HeaderChain.begin()))
1780 return L.getHeader();
1782 DEBUG(dbgs() << "Finding best loop top for: " << getBlockName(L.getHeader())
1785 BlockFrequency BestPredFreq;
1786 MachineBasicBlock *BestPred = nullptr;
1787 for (MachineBasicBlock *Pred : L.getHeader()->predecessors()) {
1788 if (!LoopBlockSet.count(Pred))
1790 DEBUG(dbgs() << " header pred: " << getBlockName(Pred) << ", has "
1791 << Pred->succ_size() << " successors, ";
1792 MBFI->printBlockFreq(dbgs(), Pred) << " freq\n");
1793 if (Pred->succ_size() > 1)
1796 BlockFrequency PredFreq = MBFI->getBlockFreq(Pred);
1797 if (!BestPred || PredFreq > BestPredFreq ||
1798 (!(PredFreq < BestPredFreq) &&
1799 Pred->isLayoutSuccessor(L.getHeader()))) {
1801 BestPredFreq = PredFreq;
1805 // If no direct predecessor is fine, just use the loop header.
1807 DEBUG(dbgs() << " final top unchanged\n");
1808 return L.getHeader();
1811 // Walk backwards through any straight line of predecessors.
1812 while (BestPred->pred_size() == 1 &&
1813 (*BestPred->pred_begin())->succ_size() == 1 &&
1814 *BestPred->pred_begin() != L.getHeader())
1815 BestPred = *BestPred->pred_begin();
1817 DEBUG(dbgs() << " final top: " << getBlockName(BestPred) << "\n");
1821 /// \brief Find the best loop exiting block for layout.
1823 /// This routine implements the logic to analyze the loop looking for the best
1824 /// block to layout at the top of the loop. Typically this is done to maximize
1825 /// fallthrough opportunities.
1827 MachineBlockPlacement::findBestLoopExit(const MachineLoop &L,
1828 const BlockFilterSet &LoopBlockSet) {
1829 // We don't want to layout the loop linearly in all cases. If the loop header
1830 // is just a normal basic block in the loop, we want to look for what block
1831 // within the loop is the best one to layout at the top. However, if the loop
1832 // header has be pre-merged into a chain due to predecessors not having
1833 // analyzable branches, *and* the predecessor it is merged with is *not* part
1834 // of the loop, rotating the header into the middle of the loop will create
1835 // a non-contiguous range of blocks which is Very Bad. So start with the
1836 // header and only rotate if safe.
1837 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
1838 if (!LoopBlockSet.count(*HeaderChain.begin()))
1841 BlockFrequency BestExitEdgeFreq;
1842 unsigned BestExitLoopDepth = 0;
1843 MachineBasicBlock *ExitingBB = nullptr;
1844 // If there are exits to outer loops, loop rotation can severely limit
1845 // fallthrough opportunities unless it selects such an exit. Keep a set of
1846 // blocks where rotating to exit with that block will reach an outer loop.
1847 SmallPtrSet<MachineBasicBlock *, 4> BlocksExitingToOuterLoop;
1849 DEBUG(dbgs() << "Finding best loop exit for: " << getBlockName(L.getHeader())
1851 for (MachineBasicBlock *MBB : L.getBlocks()) {
1852 BlockChain &Chain = *BlockToChain[MBB];
1853 // Ensure that this block is at the end of a chain; otherwise it could be
1854 // mid-way through an inner loop or a successor of an unanalyzable branch.
1855 if (MBB != *std::prev(Chain.end()))
1858 // Now walk the successors. We need to establish whether this has a viable
1859 // exiting successor and whether it has a viable non-exiting successor.
1860 // We store the old exiting state and restore it if a viable looping
1861 // successor isn't found.
1862 MachineBasicBlock *OldExitingBB = ExitingBB;
1863 BlockFrequency OldBestExitEdgeFreq = BestExitEdgeFreq;
1864 bool HasLoopingSucc = false;
1865 for (MachineBasicBlock *Succ : MBB->successors()) {
1866 if (Succ->isEHPad())
1870 BlockChain &SuccChain = *BlockToChain[Succ];
1871 // Don't split chains, either this chain or the successor's chain.
1872 if (&Chain == &SuccChain) {
1873 DEBUG(dbgs() << " exiting: " << getBlockName(MBB) << " -> "
1874 << getBlockName(Succ) << " (chain conflict)\n");
1878 auto SuccProb = MBPI->getEdgeProbability(MBB, Succ);
1879 if (LoopBlockSet.count(Succ)) {
1880 DEBUG(dbgs() << " looping: " << getBlockName(MBB) << " -> "
1881 << getBlockName(Succ) << " (" << SuccProb << ")\n");
1882 HasLoopingSucc = true;
1886 unsigned SuccLoopDepth = 0;
1887 if (MachineLoop *ExitLoop = MLI->getLoopFor(Succ)) {
1888 SuccLoopDepth = ExitLoop->getLoopDepth();
1889 if (ExitLoop->contains(&L))
1890 BlocksExitingToOuterLoop.insert(MBB);
1893 BlockFrequency ExitEdgeFreq = MBFI->getBlockFreq(MBB) * SuccProb;
1894 DEBUG(dbgs() << " exiting: " << getBlockName(MBB) << " -> "
1895 << getBlockName(Succ) << " [L:" << SuccLoopDepth << "] (";
1896 MBFI->printBlockFreq(dbgs(), ExitEdgeFreq) << ")\n");
1897 // Note that we bias this toward an existing layout successor to retain
1898 // incoming order in the absence of better information. The exit must have
1899 // a frequency higher than the current exit before we consider breaking
1901 BranchProbability Bias(100 - ExitBlockBias, 100);
1902 if (!ExitingBB || SuccLoopDepth > BestExitLoopDepth ||
1903 ExitEdgeFreq > BestExitEdgeFreq ||
1904 (MBB->isLayoutSuccessor(Succ) &&
1905 !(ExitEdgeFreq < BestExitEdgeFreq * Bias))) {
1906 BestExitEdgeFreq = ExitEdgeFreq;
1911 if (!HasLoopingSucc) {
1912 // Restore the old exiting state, no viable looping successor was found.
1913 ExitingBB = OldExitingBB;
1914 BestExitEdgeFreq = OldBestExitEdgeFreq;
1917 // Without a candidate exiting block or with only a single block in the
1918 // loop, just use the loop header to layout the loop.
1920 DEBUG(dbgs() << " No other candidate exit blocks, using loop header\n");
1923 if (L.getNumBlocks() == 1) {
1924 DEBUG(dbgs() << " Loop has 1 block, using loop header as exit\n");
1928 // Also, if we have exit blocks which lead to outer loops but didn't select
1929 // one of them as the exiting block we are rotating toward, disable loop
1930 // rotation altogether.
1931 if (!BlocksExitingToOuterLoop.empty() &&
1932 !BlocksExitingToOuterLoop.count(ExitingBB))
1935 DEBUG(dbgs() << " Best exiting block: " << getBlockName(ExitingBB) << "\n");
1939 /// \brief Attempt to rotate an exiting block to the bottom of the loop.
1941 /// Once we have built a chain, try to rotate it to line up the hot exit block
1942 /// with fallthrough out of the loop if doing so doesn't introduce unnecessary
1943 /// branches. For example, if the loop has fallthrough into its header and out
1944 /// of its bottom already, don't rotate it.
1945 void MachineBlockPlacement::rotateLoop(BlockChain &LoopChain,
1946 const MachineBasicBlock *ExitingBB,
1947 const BlockFilterSet &LoopBlockSet) {
1951 MachineBasicBlock *Top = *LoopChain.begin();
1952 MachineBasicBlock *Bottom = *std::prev(LoopChain.end());
1954 // If ExitingBB is already the last one in a chain then nothing to do.
1955 if (Bottom == ExitingBB)
1958 bool ViableTopFallthrough = false;
1959 for (MachineBasicBlock *Pred : Top->predecessors()) {
1960 BlockChain *PredChain = BlockToChain[Pred];
1961 if (!LoopBlockSet.count(Pred) &&
1962 (!PredChain || Pred == *std::prev(PredChain->end()))) {
1963 ViableTopFallthrough = true;
1968 // If the header has viable fallthrough, check whether the current loop
1969 // bottom is a viable exiting block. If so, bail out as rotating will
1970 // introduce an unnecessary branch.
1971 if (ViableTopFallthrough) {
1972 for (MachineBasicBlock *Succ : Bottom->successors()) {
1973 BlockChain *SuccChain = BlockToChain[Succ];
1974 if (!LoopBlockSet.count(Succ) &&
1975 (!SuccChain || Succ == *SuccChain->begin()))
1980 BlockChain::iterator ExitIt = llvm::find(LoopChain, ExitingBB);
1981 if (ExitIt == LoopChain.end())
1984 // Rotating a loop exit to the bottom when there is a fallthrough to top
1985 // trades the entry fallthrough for an exit fallthrough.
1986 // If there is no bottom->top edge, but the chosen exit block does have
1987 // a fallthrough, we break that fallthrough for nothing in return.
1989 // Let's consider an example. We have a built chain of basic blocks
1990 // B1, B2, ..., Bn, where Bk is a ExitingBB - chosen exit block.
1991 // By doing a rotation we get
1992 // Bk+1, ..., Bn, B1, ..., Bk
1993 // Break of fallthrough to B1 is compensated by a fallthrough from Bk.
1994 // If we had a fallthrough Bk -> Bk+1 it is broken now.
1995 // It might be compensated by fallthrough Bn -> B1.
1996 // So we have a condition to avoid creation of extra branch by loop rotation.
1997 // All below must be true to avoid loop rotation:
1998 // If there is a fallthrough to top (B1)
1999 // There was fallthrough from chosen exit block (Bk) to next one (Bk+1)
2000 // There is no fallthrough from bottom (Bn) to top (B1).
2001 // Please note that there is no exit fallthrough from Bn because we checked it
2003 if (ViableTopFallthrough) {
2004 assert(std::next(ExitIt) != LoopChain.end() &&
2005 "Exit should not be last BB");
2006 MachineBasicBlock *NextBlockInChain = *std::next(ExitIt);
2007 if (ExitingBB->isSuccessor(NextBlockInChain))
2008 if (!Bottom->isSuccessor(Top))
2012 DEBUG(dbgs() << "Rotating loop to put exit " << getBlockName(ExitingBB)
2014 std::rotate(LoopChain.begin(), std::next(ExitIt), LoopChain.end());
2017 /// \brief Attempt to rotate a loop based on profile data to reduce branch cost.
2019 /// With profile data, we can determine the cost in terms of missed fall through
2020 /// opportunities when rotating a loop chain and select the best rotation.
2021 /// Basically, there are three kinds of cost to consider for each rotation:
2022 /// 1. The possibly missed fall through edge (if it exists) from BB out of
2023 /// the loop to the loop header.
2024 /// 2. The possibly missed fall through edges (if they exist) from the loop
2025 /// exits to BB out of the loop.
2026 /// 3. The missed fall through edge (if it exists) from the last BB to the
2027 /// first BB in the loop chain.
2028 /// Therefore, the cost for a given rotation is the sum of costs listed above.
2029 /// We select the best rotation with the smallest cost.
2030 void MachineBlockPlacement::rotateLoopWithProfile(
2031 BlockChain &LoopChain, const MachineLoop &L,
2032 const BlockFilterSet &LoopBlockSet) {
2033 auto HeaderBB = L.getHeader();
2034 auto HeaderIter = llvm::find(LoopChain, HeaderBB);
2035 auto RotationPos = LoopChain.end();
2037 BlockFrequency SmallestRotationCost = BlockFrequency::getMaxFrequency();
2039 // A utility lambda that scales up a block frequency by dividing it by a
2040 // branch probability which is the reciprocal of the scale.
2041 auto ScaleBlockFrequency = [](BlockFrequency Freq,
2042 unsigned Scale) -> BlockFrequency {
2045 // Use operator / between BlockFrequency and BranchProbability to implement
2046 // saturating multiplication.
2047 return Freq / BranchProbability(1, Scale);
2050 // Compute the cost of the missed fall-through edge to the loop header if the
2051 // chain head is not the loop header. As we only consider natural loops with
2052 // single header, this computation can be done only once.
2053 BlockFrequency HeaderFallThroughCost(0);
2054 for (auto *Pred : HeaderBB->predecessors()) {
2055 BlockChain *PredChain = BlockToChain[Pred];
2056 if (!LoopBlockSet.count(Pred) &&
2057 (!PredChain || Pred == *std::prev(PredChain->end()))) {
2059 MBFI->getBlockFreq(Pred) * MBPI->getEdgeProbability(Pred, HeaderBB);
2060 auto FallThruCost = ScaleBlockFrequency(EdgeFreq, MisfetchCost);
2061 // If the predecessor has only an unconditional jump to the header, we
2062 // need to consider the cost of this jump.
2063 if (Pred->succ_size() == 1)
2064 FallThruCost += ScaleBlockFrequency(EdgeFreq, JumpInstCost);
2065 HeaderFallThroughCost = std::max(HeaderFallThroughCost, FallThruCost);
2069 // Here we collect all exit blocks in the loop, and for each exit we find out
2070 // its hottest exit edge. For each loop rotation, we define the loop exit cost
2071 // as the sum of frequencies of exit edges we collect here, excluding the exit
2072 // edge from the tail of the loop chain.
2073 SmallVector<std::pair<MachineBasicBlock *, BlockFrequency>, 4> ExitsWithFreq;
2074 for (auto BB : LoopChain) {
2075 auto LargestExitEdgeProb = BranchProbability::getZero();
2076 for (auto *Succ : BB->successors()) {
2077 BlockChain *SuccChain = BlockToChain[Succ];
2078 if (!LoopBlockSet.count(Succ) &&
2079 (!SuccChain || Succ == *SuccChain->begin())) {
2080 auto SuccProb = MBPI->getEdgeProbability(BB, Succ);
2081 LargestExitEdgeProb = std::max(LargestExitEdgeProb, SuccProb);
2084 if (LargestExitEdgeProb > BranchProbability::getZero()) {
2085 auto ExitFreq = MBFI->getBlockFreq(BB) * LargestExitEdgeProb;
2086 ExitsWithFreq.emplace_back(BB, ExitFreq);
2090 // In this loop we iterate every block in the loop chain and calculate the
2091 // cost assuming the block is the head of the loop chain. When the loop ends,
2092 // we should have found the best candidate as the loop chain's head.
2093 for (auto Iter = LoopChain.begin(), TailIter = std::prev(LoopChain.end()),
2094 EndIter = LoopChain.end();
2095 Iter != EndIter; Iter++, TailIter++) {
2096 // TailIter is used to track the tail of the loop chain if the block we are
2097 // checking (pointed by Iter) is the head of the chain.
2098 if (TailIter == LoopChain.end())
2099 TailIter = LoopChain.begin();
2101 auto TailBB = *TailIter;
2103 // Calculate the cost by putting this BB to the top.
2104 BlockFrequency Cost = 0;
2106 // If the current BB is the loop header, we need to take into account the
2107 // cost of the missed fall through edge from outside of the loop to the
2109 if (Iter != HeaderIter)
2110 Cost += HeaderFallThroughCost;
2112 // Collect the loop exit cost by summing up frequencies of all exit edges
2113 // except the one from the chain tail.
2114 for (auto &ExitWithFreq : ExitsWithFreq)
2115 if (TailBB != ExitWithFreq.first)
2116 Cost += ExitWithFreq.second;
2118 // The cost of breaking the once fall-through edge from the tail to the top
2119 // of the loop chain. Here we need to consider three cases:
2120 // 1. If the tail node has only one successor, then we will get an
2121 // additional jmp instruction. So the cost here is (MisfetchCost +
2122 // JumpInstCost) * tail node frequency.
2123 // 2. If the tail node has two successors, then we may still get an
2124 // additional jmp instruction if the layout successor after the loop
2125 // chain is not its CFG successor. Note that the more frequently executed
2126 // jmp instruction will be put ahead of the other one. Assume the
2127 // frequency of those two branches are x and y, where x is the frequency
2128 // of the edge to the chain head, then the cost will be
2129 // (x * MisfetechCost + min(x, y) * JumpInstCost) * tail node frequency.
2130 // 3. If the tail node has more than two successors (this rarely happens),
2131 // we won't consider any additional cost.
2132 if (TailBB->isSuccessor(*Iter)) {
2133 auto TailBBFreq = MBFI->getBlockFreq(TailBB);
2134 if (TailBB->succ_size() == 1)
2135 Cost += ScaleBlockFrequency(TailBBFreq.getFrequency(),
2136 MisfetchCost + JumpInstCost);
2137 else if (TailBB->succ_size() == 2) {
2138 auto TailToHeadProb = MBPI->getEdgeProbability(TailBB, *Iter);
2139 auto TailToHeadFreq = TailBBFreq * TailToHeadProb;
2140 auto ColderEdgeFreq = TailToHeadProb > BranchProbability(1, 2)
2141 ? TailBBFreq * TailToHeadProb.getCompl()
2143 Cost += ScaleBlockFrequency(TailToHeadFreq, MisfetchCost) +
2144 ScaleBlockFrequency(ColderEdgeFreq, JumpInstCost);
2148 DEBUG(dbgs() << "The cost of loop rotation by making " << getBlockName(*Iter)
2149 << " to the top: " << Cost.getFrequency() << "\n");
2151 if (Cost < SmallestRotationCost) {
2152 SmallestRotationCost = Cost;
2157 if (RotationPos != LoopChain.end()) {
2158 DEBUG(dbgs() << "Rotate loop by making " << getBlockName(*RotationPos)
2159 << " to the top\n");
2160 std::rotate(LoopChain.begin(), RotationPos, LoopChain.end());
2164 /// \brief Collect blocks in the given loop that are to be placed.
2166 /// When profile data is available, exclude cold blocks from the returned set;
2167 /// otherwise, collect all blocks in the loop.
2168 MachineBlockPlacement::BlockFilterSet
2169 MachineBlockPlacement::collectLoopBlockSet(const MachineLoop &L) {
2170 BlockFilterSet LoopBlockSet;
2172 // Filter cold blocks off from LoopBlockSet when profile data is available.
2173 // Collect the sum of frequencies of incoming edges to the loop header from
2174 // outside. If we treat the loop as a super block, this is the frequency of
2175 // the loop. Then for each block in the loop, we calculate the ratio between
2176 // its frequency and the frequency of the loop block. When it is too small,
2177 // don't add it to the loop chain. If there are outer loops, then this block
2178 // will be merged into the first outer loop chain for which this block is not
2179 // cold anymore. This needs precise profile data and we only do this when
2180 // profile data is available.
2181 if (F->getFunction().getEntryCount() || ForceLoopColdBlock) {
2182 BlockFrequency LoopFreq(0);
2183 for (auto LoopPred : L.getHeader()->predecessors())
2184 if (!L.contains(LoopPred))
2185 LoopFreq += MBFI->getBlockFreq(LoopPred) *
2186 MBPI->getEdgeProbability(LoopPred, L.getHeader());
2188 for (MachineBasicBlock *LoopBB : L.getBlocks()) {
2189 auto Freq = MBFI->getBlockFreq(LoopBB).getFrequency();
2190 if (Freq == 0 || LoopFreq.getFrequency() / Freq > LoopToColdBlockRatio)
2192 LoopBlockSet.insert(LoopBB);
2195 LoopBlockSet.insert(L.block_begin(), L.block_end());
2197 return LoopBlockSet;
2200 /// \brief Forms basic block chains from the natural loop structures.
2202 /// These chains are designed to preserve the existing *structure* of the code
2203 /// as much as possible. We can then stitch the chains together in a way which
2204 /// both preserves the topological structure and minimizes taken conditional
2206 void MachineBlockPlacement::buildLoopChains(const MachineLoop &L) {
2207 // First recurse through any nested loops, building chains for those inner
2209 for (const MachineLoop *InnerLoop : L)
2210 buildLoopChains(*InnerLoop);
2212 assert(BlockWorkList.empty() &&
2213 "BlockWorkList not empty when starting to build loop chains.");
2214 assert(EHPadWorkList.empty() &&
2215 "EHPadWorkList not empty when starting to build loop chains.");
2216 BlockFilterSet LoopBlockSet = collectLoopBlockSet(L);
2218 // Check if we have profile data for this function. If yes, we will rotate
2219 // this loop by modeling costs more precisely which requires the profile data
2220 // for better layout.
2221 bool RotateLoopWithProfile =
2222 ForcePreciseRotationCost ||
2223 (PreciseRotationCost && F->getFunction().getEntryCount());
2225 // First check to see if there is an obviously preferable top block for the
2226 // loop. This will default to the header, but may end up as one of the
2227 // predecessors to the header if there is one which will result in strictly
2228 // fewer branches in the loop body.
2229 // When we use profile data to rotate the loop, this is unnecessary.
2230 MachineBasicBlock *LoopTop =
2231 RotateLoopWithProfile ? L.getHeader() : findBestLoopTop(L, LoopBlockSet);
2233 // If we selected just the header for the loop top, look for a potentially
2234 // profitable exit block in the event that rotating the loop can eliminate
2235 // branches by placing an exit edge at the bottom.
2237 // Loops are processed innermost to uttermost, make sure we clear
2238 // PreferredLoopExit before processing a new loop.
2239 PreferredLoopExit = nullptr;
2240 if (!RotateLoopWithProfile && LoopTop == L.getHeader())
2241 PreferredLoopExit = findBestLoopExit(L, LoopBlockSet);
2243 BlockChain &LoopChain = *BlockToChain[LoopTop];
2245 // FIXME: This is a really lame way of walking the chains in the loop: we
2246 // walk the blocks, and use a set to prevent visiting a particular chain
2248 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
2249 assert(LoopChain.UnscheduledPredecessors == 0 &&
2250 "LoopChain should not have unscheduled predecessors.");
2251 UpdatedPreds.insert(&LoopChain);
2253 for (const MachineBasicBlock *LoopBB : LoopBlockSet)
2254 fillWorkLists(LoopBB, UpdatedPreds, &LoopBlockSet);
2256 buildChain(LoopTop, LoopChain, &LoopBlockSet);
2258 if (RotateLoopWithProfile)
2259 rotateLoopWithProfile(LoopChain, L, LoopBlockSet);
2261 rotateLoop(LoopChain, PreferredLoopExit, LoopBlockSet);
2264 // Crash at the end so we get all of the debugging output first.
2265 bool BadLoop = false;
2266 if (LoopChain.UnscheduledPredecessors) {
2268 dbgs() << "Loop chain contains a block without its preds placed!\n"
2269 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
2270 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n";
2272 for (MachineBasicBlock *ChainBB : LoopChain) {
2273 dbgs() << " ... " << getBlockName(ChainBB) << "\n";
2274 if (!LoopBlockSet.remove(ChainBB)) {
2275 // We don't mark the loop as bad here because there are real situations
2276 // where this can occur. For example, with an unanalyzable fallthrough
2277 // from a loop block to a non-loop block or vice versa.
2278 dbgs() << "Loop chain contains a block not contained by the loop!\n"
2279 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
2280 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
2281 << " Bad block: " << getBlockName(ChainBB) << "\n";
2285 if (!LoopBlockSet.empty()) {
2287 for (const MachineBasicBlock *LoopBB : LoopBlockSet)
2288 dbgs() << "Loop contains blocks never placed into a chain!\n"
2289 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
2290 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
2291 << " Bad block: " << getBlockName(LoopBB) << "\n";
2293 assert(!BadLoop && "Detected problems with the placement of this loop.");
2296 BlockWorkList.clear();
2297 EHPadWorkList.clear();
2300 void MachineBlockPlacement::buildCFGChains() {
2301 // Ensure that every BB in the function has an associated chain to simplify
2302 // the assumptions of the remaining algorithm.
2303 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
2304 for (MachineFunction::iterator FI = F->begin(), FE = F->end(); FI != FE;
2306 MachineBasicBlock *BB = &*FI;
2308 new (ChainAllocator.Allocate()) BlockChain(BlockToChain, BB);
2309 // Also, merge any blocks which we cannot reason about and must preserve
2310 // the exact fallthrough behavior for.
2313 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
2314 if (!TII->analyzeBranch(*BB, TBB, FBB, Cond) || !FI->canFallThrough())
2317 MachineFunction::iterator NextFI = std::next(FI);
2318 MachineBasicBlock *NextBB = &*NextFI;
2319 // Ensure that the layout successor is a viable block, as we know that
2320 // fallthrough is a possibility.
2321 assert(NextFI != FE && "Can't fallthrough past the last block.");
2322 DEBUG(dbgs() << "Pre-merging due to unanalyzable fallthrough: "
2323 << getBlockName(BB) << " -> " << getBlockName(NextBB)
2325 Chain->merge(NextBB, nullptr);
2327 BlocksWithUnanalyzableExits.insert(&*BB);
2334 // Build any loop-based chains.
2335 PreferredLoopExit = nullptr;
2336 for (MachineLoop *L : *MLI)
2337 buildLoopChains(*L);
2339 assert(BlockWorkList.empty() &&
2340 "BlockWorkList should be empty before building final chain.");
2341 assert(EHPadWorkList.empty() &&
2342 "EHPadWorkList should be empty before building final chain.");
2344 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
2345 for (MachineBasicBlock &MBB : *F)
2346 fillWorkLists(&MBB, UpdatedPreds);
2348 BlockChain &FunctionChain = *BlockToChain[&F->front()];
2349 buildChain(&F->front(), FunctionChain);
2352 using FunctionBlockSetType = SmallPtrSet<MachineBasicBlock *, 16>;
2355 // Crash at the end so we get all of the debugging output first.
2356 bool BadFunc = false;
2357 FunctionBlockSetType FunctionBlockSet;
2358 for (MachineBasicBlock &MBB : *F)
2359 FunctionBlockSet.insert(&MBB);
2361 for (MachineBasicBlock *ChainBB : FunctionChain)
2362 if (!FunctionBlockSet.erase(ChainBB)) {
2364 dbgs() << "Function chain contains a block not in the function!\n"
2365 << " Bad block: " << getBlockName(ChainBB) << "\n";
2368 if (!FunctionBlockSet.empty()) {
2370 for (MachineBasicBlock *RemainingBB : FunctionBlockSet)
2371 dbgs() << "Function contains blocks never placed into a chain!\n"
2372 << " Bad block: " << getBlockName(RemainingBB) << "\n";
2374 assert(!BadFunc && "Detected problems with the block placement.");
2377 // Splice the blocks into place.
2378 MachineFunction::iterator InsertPos = F->begin();
2379 DEBUG(dbgs() << "[MBP] Function: "<< F->getName() << "\n");
2380 for (MachineBasicBlock *ChainBB : FunctionChain) {
2381 DEBUG(dbgs() << (ChainBB == *FunctionChain.begin() ? "Placing chain "
2383 << getBlockName(ChainBB) << "\n");
2384 if (InsertPos != MachineFunction::iterator(ChainBB))
2385 F->splice(InsertPos, ChainBB);
2389 // Update the terminator of the previous block.
2390 if (ChainBB == *FunctionChain.begin())
2392 MachineBasicBlock *PrevBB = &*std::prev(MachineFunction::iterator(ChainBB));
2394 // FIXME: It would be awesome of updateTerminator would just return rather
2395 // than assert when the branch cannot be analyzed in order to remove this
2398 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
2401 if (!BlocksWithUnanalyzableExits.count(PrevBB)) {
2402 // Given the exact block placement we chose, we may actually not _need_ to
2403 // be able to edit PrevBB's terminator sequence, but not being _able_ to
2404 // do that at this point is a bug.
2405 assert((!TII->analyzeBranch(*PrevBB, TBB, FBB, Cond) ||
2406 !PrevBB->canFallThrough()) &&
2407 "Unexpected block with un-analyzable fallthrough!");
2409 TBB = FBB = nullptr;
2413 // The "PrevBB" is not yet updated to reflect current code layout, so,
2414 // o. it may fall-through to a block without explicit "goto" instruction
2415 // before layout, and no longer fall-through it after layout; or
2416 // o. just opposite.
2418 // analyzeBranch() may return erroneous value for FBB when these two
2419 // situations take place. For the first scenario FBB is mistakenly set NULL;
2420 // for the 2nd scenario, the FBB, which is expected to be NULL, is
2421 // mistakenly pointing to "*BI".
2422 // Thus, if the future change needs to use FBB before the layout is set, it
2423 // has to correct FBB first by using the code similar to the following:
2425 // if (!Cond.empty() && (!FBB || FBB == ChainBB)) {
2426 // PrevBB->updateTerminator();
2428 // TBB = FBB = nullptr;
2429 // if (TII->analyzeBranch(*PrevBB, TBB, FBB, Cond)) {
2430 // // FIXME: This should never take place.
2431 // TBB = FBB = nullptr;
2434 if (!TII->analyzeBranch(*PrevBB, TBB, FBB, Cond))
2435 PrevBB->updateTerminator();
2438 // Fixup the last block.
2440 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
2441 if (!TII->analyzeBranch(F->back(), TBB, FBB, Cond))
2442 F->back().updateTerminator();
2444 BlockWorkList.clear();
2445 EHPadWorkList.clear();
2448 void MachineBlockPlacement::optimizeBranches() {
2449 BlockChain &FunctionChain = *BlockToChain[&F->front()];
2450 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
2452 // Now that all the basic blocks in the chain have the proper layout,
2453 // make a final call to AnalyzeBranch with AllowModify set.
2454 // Indeed, the target may be able to optimize the branches in a way we
2455 // cannot because all branches may not be analyzable.
2456 // E.g., the target may be able to remove an unconditional branch to
2457 // a fallthrough when it occurs after predicated terminators.
2458 for (MachineBasicBlock *ChainBB : FunctionChain) {
2460 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
2461 if (!TII->analyzeBranch(*ChainBB, TBB, FBB, Cond, /*AllowModify*/ true)) {
2462 // If PrevBB has a two-way branch, try to re-order the branches
2463 // such that we branch to the successor with higher probability first.
2464 if (TBB && !Cond.empty() && FBB &&
2465 MBPI->getEdgeProbability(ChainBB, FBB) >
2466 MBPI->getEdgeProbability(ChainBB, TBB) &&
2467 !TII->reverseBranchCondition(Cond)) {
2468 DEBUG(dbgs() << "Reverse order of the two branches: "
2469 << getBlockName(ChainBB) << "\n");
2470 DEBUG(dbgs() << " Edge probability: "
2471 << MBPI->getEdgeProbability(ChainBB, FBB) << " vs "
2472 << MBPI->getEdgeProbability(ChainBB, TBB) << "\n");
2473 DebugLoc dl; // FIXME: this is nowhere
2474 TII->removeBranch(*ChainBB);
2475 TII->insertBranch(*ChainBB, FBB, TBB, Cond, dl);
2476 ChainBB->updateTerminator();
2482 void MachineBlockPlacement::alignBlocks() {
2483 // Walk through the backedges of the function now that we have fully laid out
2484 // the basic blocks and align the destination of each backedge. We don't rely
2485 // exclusively on the loop info here so that we can align backedges in
2486 // unnatural CFGs and backedges that were introduced purely because of the
2487 // loop rotations done during this layout pass.
2488 if (F->getFunction().optForSize())
2490 BlockChain &FunctionChain = *BlockToChain[&F->front()];
2491 if (FunctionChain.begin() == FunctionChain.end())
2492 return; // Empty chain.
2494 const BranchProbability ColdProb(1, 5); // 20%
2495 BlockFrequency EntryFreq = MBFI->getBlockFreq(&F->front());
2496 BlockFrequency WeightedEntryFreq = EntryFreq * ColdProb;
2497 for (MachineBasicBlock *ChainBB : FunctionChain) {
2498 if (ChainBB == *FunctionChain.begin())
2501 // Don't align non-looping basic blocks. These are unlikely to execute
2502 // enough times to matter in practice. Note that we'll still handle
2503 // unnatural CFGs inside of a natural outer loop (the common case) and
2505 MachineLoop *L = MLI->getLoopFor(ChainBB);
2509 unsigned Align = TLI->getPrefLoopAlignment(L);
2511 continue; // Don't care about loop alignment.
2513 // If the block is cold relative to the function entry don't waste space
2515 BlockFrequency Freq = MBFI->getBlockFreq(ChainBB);
2516 if (Freq < WeightedEntryFreq)
2519 // If the block is cold relative to its loop header, don't align it
2520 // regardless of what edges into the block exist.
2521 MachineBasicBlock *LoopHeader = L->getHeader();
2522 BlockFrequency LoopHeaderFreq = MBFI->getBlockFreq(LoopHeader);
2523 if (Freq < (LoopHeaderFreq * ColdProb))
2526 // Check for the existence of a non-layout predecessor which would benefit
2527 // from aligning this block.
2528 MachineBasicBlock *LayoutPred =
2529 &*std::prev(MachineFunction::iterator(ChainBB));
2531 // Force alignment if all the predecessors are jumps. We already checked
2532 // that the block isn't cold above.
2533 if (!LayoutPred->isSuccessor(ChainBB)) {
2534 ChainBB->setAlignment(Align);
2538 // Align this block if the layout predecessor's edge into this block is
2539 // cold relative to the block. When this is true, other predecessors make up
2540 // all of the hot entries into the block and thus alignment is likely to be
2542 BranchProbability LayoutProb =
2543 MBPI->getEdgeProbability(LayoutPred, ChainBB);
2544 BlockFrequency LayoutEdgeFreq = MBFI->getBlockFreq(LayoutPred) * LayoutProb;
2545 if (LayoutEdgeFreq <= (Freq * ColdProb))
2546 ChainBB->setAlignment(Align);
2550 /// Tail duplicate \p BB into (some) predecessors if profitable, repeating if
2551 /// it was duplicated into its chain predecessor and removed.
2552 /// \p BB - Basic block that may be duplicated.
2554 /// \p LPred - Chosen layout predecessor of \p BB.
2555 /// Updated to be the chain end if LPred is removed.
2556 /// \p Chain - Chain to which \p LPred belongs, and \p BB will belong.
2557 /// \p BlockFilter - Set of blocks that belong to the loop being laid out.
2558 /// Used to identify which blocks to update predecessor
2560 /// \p PrevUnplacedBlockIt - Iterator pointing to the last block that was
2561 /// chosen in the given order due to unnatural CFG
2562 /// only needed if \p BB is removed and
2563 /// \p PrevUnplacedBlockIt pointed to \p BB.
2564 /// @return true if \p BB was removed.
2565 bool MachineBlockPlacement::repeatedlyTailDuplicateBlock(
2566 MachineBasicBlock *BB, MachineBasicBlock *&LPred,
2567 const MachineBasicBlock *LoopHeaderBB,
2568 BlockChain &Chain, BlockFilterSet *BlockFilter,
2569 MachineFunction::iterator &PrevUnplacedBlockIt) {
2570 bool Removed, DuplicatedToLPred;
2571 bool DuplicatedToOriginalLPred;
2572 Removed = maybeTailDuplicateBlock(BB, LPred, Chain, BlockFilter,
2573 PrevUnplacedBlockIt,
2577 DuplicatedToOriginalLPred = DuplicatedToLPred;
2578 // Iteratively try to duplicate again. It can happen that a block that is
2579 // duplicated into is still small enough to be duplicated again.
2580 // No need to call markBlockSuccessors in this case, as the blocks being
2581 // duplicated from here on are already scheduled.
2582 // Note that DuplicatedToLPred always implies Removed.
2583 while (DuplicatedToLPred) {
2584 assert(Removed && "Block must have been removed to be duplicated into its "
2585 "layout predecessor.");
2586 MachineBasicBlock *DupBB, *DupPred;
2587 // The removal callback causes Chain.end() to be updated when a block is
2588 // removed. On the first pass through the loop, the chain end should be the
2589 // same as it was on function entry. On subsequent passes, because we are
2590 // duplicating the block at the end of the chain, if it is removed the
2591 // chain will have shrunk by one block.
2592 BlockChain::iterator ChainEnd = Chain.end();
2593 DupBB = *(--ChainEnd);
2594 // Now try to duplicate again.
2595 if (ChainEnd == Chain.begin())
2597 DupPred = *std::prev(ChainEnd);
2598 Removed = maybeTailDuplicateBlock(DupBB, DupPred, Chain, BlockFilter,
2599 PrevUnplacedBlockIt,
2602 // If BB was duplicated into LPred, it is now scheduled. But because it was
2603 // removed, markChainSuccessors won't be called for its chain. Instead we
2604 // call markBlockSuccessors for LPred to achieve the same effect. This must go
2605 // at the end because repeating the tail duplication can increase the number
2606 // of unscheduled predecessors.
2607 LPred = *std::prev(Chain.end());
2608 if (DuplicatedToOriginalLPred)
2609 markBlockSuccessors(Chain, LPred, LoopHeaderBB, BlockFilter);
2613 /// Tail duplicate \p BB into (some) predecessors if profitable.
2614 /// \p BB - Basic block that may be duplicated
2615 /// \p LPred - Chosen layout predecessor of \p BB
2616 /// \p Chain - Chain to which \p LPred belongs, and \p BB will belong.
2617 /// \p BlockFilter - Set of blocks that belong to the loop being laid out.
2618 /// Used to identify which blocks to update predecessor
2620 /// \p PrevUnplacedBlockIt - Iterator pointing to the last block that was
2621 /// chosen in the given order due to unnatural CFG
2622 /// only needed if \p BB is removed and
2623 /// \p PrevUnplacedBlockIt pointed to \p BB.
2624 /// \p DuplicatedToLPred - True if the block was duplicated into LPred. Will
2625 /// only be true if the block was removed.
2626 /// \return - True if the block was duplicated into all preds and removed.
2627 bool MachineBlockPlacement::maybeTailDuplicateBlock(
2628 MachineBasicBlock *BB, MachineBasicBlock *LPred,
2629 BlockChain &Chain, BlockFilterSet *BlockFilter,
2630 MachineFunction::iterator &PrevUnplacedBlockIt,
2631 bool &DuplicatedToLPred) {
2632 DuplicatedToLPred = false;
2633 if (!shouldTailDuplicate(BB))
2636 DEBUG(dbgs() << "Redoing tail duplication for Succ#"
2637 << BB->getNumber() << "\n");
2639 // This has to be a callback because none of it can be done after
2641 bool Removed = false;
2642 auto RemovalCallback =
2643 [&](MachineBasicBlock *RemBB) {
2644 // Signal to outer function
2647 // Conservative default.
2648 bool InWorkList = true;
2649 // Remove from the Chain and Chain Map
2650 if (BlockToChain.count(RemBB)) {
2651 BlockChain *Chain = BlockToChain[RemBB];
2652 InWorkList = Chain->UnscheduledPredecessors == 0;
2653 Chain->remove(RemBB);
2654 BlockToChain.erase(RemBB);
2657 // Handle the unplaced block iterator
2658 if (&(*PrevUnplacedBlockIt) == RemBB) {
2659 PrevUnplacedBlockIt++;
2662 // Handle the Work Lists
2664 SmallVectorImpl<MachineBasicBlock *> &RemoveList = BlockWorkList;
2665 if (RemBB->isEHPad())
2666 RemoveList = EHPadWorkList;
2668 llvm::remove_if(RemoveList,
2669 [RemBB](MachineBasicBlock *BB) {
2675 // Handle the filter set
2677 BlockFilter->remove(RemBB);
2680 // Remove the block from loop info.
2681 MLI->removeBlock(RemBB);
2682 if (RemBB == PreferredLoopExit)
2683 PreferredLoopExit = nullptr;
2685 DEBUG(dbgs() << "TailDuplicator deleted block: "
2686 << getBlockName(RemBB) << "\n");
2688 auto RemovalCallbackRef =
2689 function_ref<void(MachineBasicBlock*)>(RemovalCallback);
2691 SmallVector<MachineBasicBlock *, 8> DuplicatedPreds;
2692 bool IsSimple = TailDup.isSimpleBB(BB);
2693 TailDup.tailDuplicateAndUpdate(IsSimple, BB, LPred,
2694 &DuplicatedPreds, &RemovalCallbackRef);
2696 // Update UnscheduledPredecessors to reflect tail-duplication.
2697 DuplicatedToLPred = false;
2698 for (MachineBasicBlock *Pred : DuplicatedPreds) {
2699 // We're only looking for unscheduled predecessors that match the filter.
2700 BlockChain* PredChain = BlockToChain[Pred];
2702 DuplicatedToLPred = true;
2703 if (Pred == LPred || (BlockFilter && !BlockFilter->count(Pred))
2704 || PredChain == &Chain)
2706 for (MachineBasicBlock *NewSucc : Pred->successors()) {
2707 if (BlockFilter && !BlockFilter->count(NewSucc))
2709 BlockChain *NewChain = BlockToChain[NewSucc];
2710 if (NewChain != &Chain && NewChain != PredChain)
2711 NewChain->UnscheduledPredecessors++;
2717 bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &MF) {
2718 if (skipFunction(MF.getFunction()))
2721 // Check for single-block functions and skip them.
2722 if (std::next(MF.begin()) == MF.end())
2726 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
2727 MBFI = llvm::make_unique<BranchFolder::MBFIWrapper>(
2728 getAnalysis<MachineBlockFrequencyInfo>());
2729 MLI = &getAnalysis<MachineLoopInfo>();
2730 TII = MF.getSubtarget().getInstrInfo();
2731 TLI = MF.getSubtarget().getTargetLowering();
2734 // Initialize PreferredLoopExit to nullptr here since it may never be set if
2735 // there are no MachineLoops.
2736 PreferredLoopExit = nullptr;
2738 assert(BlockToChain.empty() &&
2739 "BlockToChain map should be empty before starting placement.");
2740 assert(ComputedEdges.empty() &&
2741 "Computed Edge map should be empty before starting placement.");
2743 unsigned TailDupSize = TailDupPlacementThreshold;
2744 // If only the aggressive threshold is explicitly set, use it.
2745 if (TailDupPlacementAggressiveThreshold.getNumOccurrences() != 0 &&
2746 TailDupPlacementThreshold.getNumOccurrences() == 0)
2747 TailDupSize = TailDupPlacementAggressiveThreshold;
2749 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
2750 // For agressive optimization, we can adjust some thresholds to be less
2752 if (PassConfig->getOptLevel() >= CodeGenOpt::Aggressive) {
2753 // At O3 we should be more willing to copy blocks for tail duplication. This
2754 // increases size pressure, so we only do it at O3
2755 // Do this unless only the regular threshold is explicitly set.
2756 if (TailDupPlacementThreshold.getNumOccurrences() == 0 ||
2757 TailDupPlacementAggressiveThreshold.getNumOccurrences() != 0)
2758 TailDupSize = TailDupPlacementAggressiveThreshold;
2761 if (TailDupPlacement) {
2762 MPDT = &getAnalysis<MachinePostDominatorTree>();
2763 if (MF.getFunction().optForSize())
2765 bool PreRegAlloc = false;
2766 TailDup.initMF(MF, PreRegAlloc, MBPI, /* LayoutMode */ true, TailDupSize);
2767 precomputeTriangleChains();
2772 // Changing the layout can create new tail merging opportunities.
2773 // TailMerge can create jump into if branches that make CFG irreducible for
2774 // HW that requires structured CFG.
2775 bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
2776 PassConfig->getEnableTailMerge() &&
2777 BranchFoldPlacement;
2778 // No tail merging opportunities if the block number is less than four.
2779 if (MF.size() > 3 && EnableTailMerge) {
2780 unsigned TailMergeSize = TailDupSize + 1;
2781 BranchFolder BF(/*EnableTailMerge=*/true, /*CommonHoist=*/false, *MBFI,
2782 *MBPI, TailMergeSize);
2784 if (BF.OptimizeFunction(MF, TII, MF.getSubtarget().getRegisterInfo(),
2785 getAnalysisIfAvailable<MachineModuleInfo>(), MLI,
2786 /*AfterBlockPlacement=*/true)) {
2787 // Redo the layout if tail merging creates/removes/moves blocks.
2788 BlockToChain.clear();
2789 ComputedEdges.clear();
2790 // Must redo the post-dominator tree if blocks were changed.
2792 MPDT->runOnMachineFunction(MF);
2793 ChainAllocator.DestroyAll();
2801 BlockToChain.clear();
2802 ComputedEdges.clear();
2803 ChainAllocator.DestroyAll();
2806 // Align all of the blocks in the function to a specific alignment.
2807 for (MachineBasicBlock &MBB : MF)
2808 MBB.setAlignment(AlignAllBlock);
2809 else if (AlignAllNonFallThruBlocks) {
2810 // Align all of the blocks that have no fall-through predecessors to a
2811 // specific alignment.
2812 for (auto MBI = std::next(MF.begin()), MBE = MF.end(); MBI != MBE; ++MBI) {
2813 auto LayoutPred = std::prev(MBI);
2814 if (!LayoutPred->isSuccessor(&*MBI))
2815 MBI->setAlignment(AlignAllNonFallThruBlocks);
2818 if (ViewBlockLayoutWithBFI != GVDT_None &&
2819 (ViewBlockFreqFuncName.empty() ||
2820 F->getFunction().getName().equals(ViewBlockFreqFuncName))) {
2821 MBFI->view("MBP." + MF.getName(), false);
2825 // We always return true as we have no way to track whether the final order
2826 // differs from the original order.
2832 /// \brief A pass to compute block placement statistics.
2834 /// A separate pass to compute interesting statistics for evaluating block
2835 /// placement. This is separate from the actual placement pass so that they can
2836 /// be computed in the absence of any placement transformations or when using
2837 /// alternative placement strategies.
2838 class MachineBlockPlacementStats : public MachineFunctionPass {
2839 /// \brief A handle to the branch probability pass.
2840 const MachineBranchProbabilityInfo *MBPI;
2842 /// \brief A handle to the function-wide block frequency pass.
2843 const MachineBlockFrequencyInfo *MBFI;
2846 static char ID; // Pass identification, replacement for typeid
2848 MachineBlockPlacementStats() : MachineFunctionPass(ID) {
2849 initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry());
2852 bool runOnMachineFunction(MachineFunction &F) override;
2854 void getAnalysisUsage(AnalysisUsage &AU) const override {
2855 AU.addRequired<MachineBranchProbabilityInfo>();
2856 AU.addRequired<MachineBlockFrequencyInfo>();
2857 AU.setPreservesAll();
2858 MachineFunctionPass::getAnalysisUsage(AU);
2862 } // end anonymous namespace
2864 char MachineBlockPlacementStats::ID = 0;
2866 char &llvm::MachineBlockPlacementStatsID = MachineBlockPlacementStats::ID;
2868 INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats",
2869 "Basic Block Placement Stats", false, false)
2870 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
2871 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
2872 INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats",
2873 "Basic Block Placement Stats", false, false)
2875 bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) {
2876 // Check for single-block functions and skip them.
2877 if (std::next(F.begin()) == F.end())
2880 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
2881 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
2883 for (MachineBasicBlock &MBB : F) {
2884 BlockFrequency BlockFreq = MBFI->getBlockFreq(&MBB);
2885 Statistic &NumBranches =
2886 (MBB.succ_size() > 1) ? NumCondBranches : NumUncondBranches;
2887 Statistic &BranchTakenFreq =
2888 (MBB.succ_size() > 1) ? CondBranchTakenFreq : UncondBranchTakenFreq;
2889 for (MachineBasicBlock *Succ : MBB.successors()) {
2890 // Skip if this successor is a fallthrough.
2891 if (MBB.isLayoutSuccessor(Succ))
2894 BlockFrequency EdgeFreq =
2895 BlockFreq * MBPI->getEdgeProbability(&MBB, Succ);
2897 BranchTakenFreq += EdgeFreq.getFrequency();