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 bool allowTailDupPlacement() const {
518 return TailDupPlacement && !F->getTarget().requiresStructuredCFG();
521 void getAnalysisUsage(AnalysisUsage &AU) const override {
522 AU.addRequired<MachineBranchProbabilityInfo>();
523 AU.addRequired<MachineBlockFrequencyInfo>();
524 if (TailDupPlacement)
525 AU.addRequired<MachinePostDominatorTree>();
526 AU.addRequired<MachineLoopInfo>();
527 AU.addRequired<TargetPassConfig>();
528 MachineFunctionPass::getAnalysisUsage(AU);
532 } // end anonymous namespace
534 char MachineBlockPlacement::ID = 0;
536 char &llvm::MachineBlockPlacementID = MachineBlockPlacement::ID;
538 INITIALIZE_PASS_BEGIN(MachineBlockPlacement, DEBUG_TYPE,
539 "Branch Probability Basic Block Placement", false, false)
540 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
541 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
542 INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTree)
543 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
544 INITIALIZE_PASS_END(MachineBlockPlacement, DEBUG_TYPE,
545 "Branch Probability Basic Block Placement", false, false)
548 /// \brief Helper to print the name of a MBB.
550 /// Only used by debug logging.
551 static std::string getBlockName(const MachineBasicBlock *BB) {
553 raw_string_ostream OS(Result);
554 OS << printMBBReference(*BB);
555 OS << " ('" << BB->getName() << "')";
561 /// \brief Mark a chain's successors as having one fewer preds.
563 /// When a chain is being merged into the "placed" chain, this routine will
564 /// quickly walk the successors of each block in the chain and mark them as
565 /// having one fewer active predecessor. It also adds any successors of this
566 /// chain which reach the zero-predecessor state to the appropriate worklist.
567 void MachineBlockPlacement::markChainSuccessors(
568 const BlockChain &Chain, const MachineBasicBlock *LoopHeaderBB,
569 const BlockFilterSet *BlockFilter) {
570 // Walk all the blocks in this chain, marking their successors as having
571 // a predecessor placed.
572 for (MachineBasicBlock *MBB : Chain) {
573 markBlockSuccessors(Chain, MBB, LoopHeaderBB, BlockFilter);
577 /// \brief Mark a single block's successors as having one fewer preds.
579 /// Under normal circumstances, this is only called by markChainSuccessors,
580 /// but if a block that was to be placed is completely tail-duplicated away,
581 /// and was duplicated into the chain end, we need to redo markBlockSuccessors
582 /// for just that block.
583 void MachineBlockPlacement::markBlockSuccessors(
584 const BlockChain &Chain, const MachineBasicBlock *MBB,
585 const MachineBasicBlock *LoopHeaderBB, const BlockFilterSet *BlockFilter) {
586 // Add any successors for which this is the only un-placed in-loop
587 // predecessor to the worklist as a viable candidate for CFG-neutral
588 // placement. No subsequent placement of this block will violate the CFG
589 // shape, so we get to use heuristics to choose a favorable placement.
590 for (MachineBasicBlock *Succ : MBB->successors()) {
591 if (BlockFilter && !BlockFilter->count(Succ))
593 BlockChain &SuccChain = *BlockToChain[Succ];
594 // Disregard edges within a fixed chain, or edges to the loop header.
595 if (&Chain == &SuccChain || Succ == LoopHeaderBB)
598 // This is a cross-chain edge that is within the loop, so decrement the
599 // loop predecessor count of the destination chain.
600 if (SuccChain.UnscheduledPredecessors == 0 ||
601 --SuccChain.UnscheduledPredecessors > 0)
604 auto *NewBB = *SuccChain.begin();
605 if (NewBB->isEHPad())
606 EHPadWorkList.push_back(NewBB);
608 BlockWorkList.push_back(NewBB);
612 /// This helper function collects the set of successors of block
613 /// \p BB that are allowed to be its layout successors, and return
614 /// the total branch probability of edges from \p BB to those
616 BranchProbability MachineBlockPlacement::collectViableSuccessors(
617 const MachineBasicBlock *BB, const BlockChain &Chain,
618 const BlockFilterSet *BlockFilter,
619 SmallVector<MachineBasicBlock *, 4> &Successors) {
620 // Adjust edge probabilities by excluding edges pointing to blocks that is
621 // either not in BlockFilter or is already in the current chain. Consider the
630 // Assume A->C is very hot (>90%), and C->D has a 50% probability, then after
631 // A->C is chosen as a fall-through, D won't be selected as a successor of C
632 // due to CFG constraint (the probability of C->D is not greater than
633 // HotProb to break topo-order). If we exclude E that is not in BlockFilter
634 // when calculating the probability of C->D, D will be selected and we
635 // will get A C D B as the layout of this loop.
636 auto AdjustedSumProb = BranchProbability::getOne();
637 for (MachineBasicBlock *Succ : BB->successors()) {
638 bool SkipSucc = false;
639 if (Succ->isEHPad() || (BlockFilter && !BlockFilter->count(Succ))) {
642 BlockChain *SuccChain = BlockToChain[Succ];
643 if (SuccChain == &Chain) {
645 } else if (Succ != *SuccChain->begin()) {
646 DEBUG(dbgs() << " " << getBlockName(Succ) << " -> Mid chain!\n");
651 AdjustedSumProb -= MBPI->getEdgeProbability(BB, Succ);
653 Successors.push_back(Succ);
656 return AdjustedSumProb;
659 /// The helper function returns the branch probability that is adjusted
660 /// or normalized over the new total \p AdjustedSumProb.
661 static BranchProbability
662 getAdjustedProbability(BranchProbability OrigProb,
663 BranchProbability AdjustedSumProb) {
664 BranchProbability SuccProb;
665 uint32_t SuccProbN = OrigProb.getNumerator();
666 uint32_t SuccProbD = AdjustedSumProb.getNumerator();
667 if (SuccProbN >= SuccProbD)
668 SuccProb = BranchProbability::getOne();
670 SuccProb = BranchProbability(SuccProbN, SuccProbD);
675 /// Check if \p BB has exactly the successors in \p Successors.
677 hasSameSuccessors(MachineBasicBlock &BB,
678 SmallPtrSetImpl<const MachineBasicBlock *> &Successors) {
679 if (BB.succ_size() != Successors.size())
681 // We don't want to count self-loops
682 if (Successors.count(&BB))
684 for (MachineBasicBlock *Succ : BB.successors())
685 if (!Successors.count(Succ))
690 /// Check if a block should be tail duplicated to increase fallthrough
692 /// \p BB Block to check.
693 bool MachineBlockPlacement::shouldTailDuplicate(MachineBasicBlock *BB) {
694 // Blocks with single successors don't create additional fallthrough
695 // opportunities. Don't duplicate them. TODO: When conditional exits are
696 // analyzable, allow them to be duplicated.
697 bool IsSimple = TailDup.isSimpleBB(BB);
699 if (BB->succ_size() == 1)
701 return TailDup.shouldTailDuplicate(IsSimple, *BB);
704 /// Compare 2 BlockFrequency's with a small penalty for \p A.
705 /// In order to be conservative, we apply a X% penalty to account for
706 /// increased icache pressure and static heuristics. For small frequencies
707 /// we use only the numerators to improve accuracy. For simplicity, we assume the
708 /// penalty is less than 100%
709 /// TODO(iteratee): Use 64-bit fixed point edge frequencies everywhere.
710 static bool greaterWithBias(BlockFrequency A, BlockFrequency B,
711 uint64_t EntryFreq) {
712 BranchProbability ThresholdProb(TailDupPlacementPenalty, 100);
713 BlockFrequency Gain = A - B;
714 return (Gain / ThresholdProb).getFrequency() >= EntryFreq;
717 /// Check the edge frequencies to see if tail duplication will increase
718 /// fallthroughs. It only makes sense to call this function when
719 /// \p Succ would not be chosen otherwise. Tail duplication of \p Succ is
720 /// always locally profitable if we would have picked \p Succ without
721 /// considering duplication.
722 bool MachineBlockPlacement::isProfitableToTailDup(
723 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
724 BranchProbability QProb,
725 const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
726 // We need to do a probability calculation to make sure this is profitable.
727 // First: does succ have a successor that post-dominates? This affects the
728 // calculation. The 2 relevant cases are:
743 // '=' : Branch taken for that CFG edge
744 // In the second case, Placing Succ while duplicating it into C prevents the
745 // fallthrough of Succ into either D or PDom, because they now have C as an
746 // unplaced predecessor
748 // Start by figuring out which case we fall into
749 MachineBasicBlock *PDom = nullptr;
750 SmallVector<MachineBasicBlock *, 4> SuccSuccs;
751 // Only scan the relevant successors
752 auto AdjustedSuccSumProb =
753 collectViableSuccessors(Succ, Chain, BlockFilter, SuccSuccs);
754 BranchProbability PProb = MBPI->getEdgeProbability(BB, Succ);
755 auto BBFreq = MBFI->getBlockFreq(BB);
756 auto SuccFreq = MBFI->getBlockFreq(Succ);
757 BlockFrequency P = BBFreq * PProb;
758 BlockFrequency Qout = BBFreq * QProb;
759 uint64_t EntryFreq = MBFI->getEntryFreq();
760 // If there are no more successors, it is profitable to copy, as it strictly
761 // increases fallthrough.
762 if (SuccSuccs.size() == 0)
763 return greaterWithBias(P, Qout, EntryFreq);
765 auto BestSuccSucc = BranchProbability::getZero();
766 // Find the PDom or the best Succ if no PDom exists.
767 for (MachineBasicBlock *SuccSucc : SuccSuccs) {
768 auto Prob = MBPI->getEdgeProbability(Succ, SuccSucc);
769 if (Prob > BestSuccSucc)
772 if (MPDT->dominates(SuccSucc, Succ)) {
777 // For the comparisons, we need to know Succ's best incoming edge that isn't
779 auto SuccBestPred = BlockFrequency(0);
780 for (MachineBasicBlock *SuccPred : Succ->predecessors()) {
781 if (SuccPred == Succ || SuccPred == BB
782 || BlockToChain[SuccPred] == &Chain
783 || (BlockFilter && !BlockFilter->count(SuccPred)))
785 auto Freq = MBFI->getBlockFreq(SuccPred)
786 * MBPI->getEdgeProbability(SuccPred, Succ);
787 if (Freq > SuccBestPred)
790 // Qin is Succ's best unplaced incoming edge that isn't BB
791 BlockFrequency Qin = SuccBestPred;
792 // If it doesn't have a post-dominating successor, here is the calculation:
804 // '=' : Branch taken for that CFG edge
805 // Cost in the first case is: P + V
806 // For this calculation, we always assume P > Qout. If Qout > P
807 // The result of this function will be ignored at the caller.
808 // Let F = SuccFreq - Qin
809 // Cost in the second case is: Qout + min(Qin, F) * U + max(Qin, F) * V
811 if (PDom == nullptr || !Succ->isSuccessor(PDom)) {
812 BranchProbability UProb = BestSuccSucc;
813 BranchProbability VProb = AdjustedSuccSumProb - UProb;
814 BlockFrequency F = SuccFreq - Qin;
815 BlockFrequency V = SuccFreq * VProb;
816 BlockFrequency QinU = std::min(Qin, F) * UProb;
817 BlockFrequency BaseCost = P + V;
818 BlockFrequency DupCost = Qout + QinU + std::max(Qin, F) * VProb;
819 return greaterWithBias(BaseCost, DupCost, EntryFreq);
821 BranchProbability UProb = MBPI->getEdgeProbability(Succ, PDom);
822 BranchProbability VProb = AdjustedSuccSumProb - UProb;
823 BlockFrequency U = SuccFreq * UProb;
824 BlockFrequency V = SuccFreq * VProb;
825 BlockFrequency F = SuccFreq - Qin;
826 // If there is a post-dominating successor, here is the calculation:
828 // | \Qout | \ | \Qout | \
830 // = C' |P C = C' |P C
831 // | /Qin | | | /Qin | |
832 // | / | C' (+Succ) | / | C' (+Succ)
833 // Succ Succ /| Succ Succ /|
834 // | \ V | \/ | | \ V | \/ |
835 // |U \ |U /\ =? |U = |U /\ |
836 // = D = = =?| | D | = =|
841 // '=' : Branch taken for that CFG edge
842 // The cost for taken branches in the first case is P + U
843 // Let F = SuccFreq - Qin
844 // The cost in the second case (assuming independence), given the layout:
845 // BB, Succ, (C+Succ), D, Dom or the layout:
846 // BB, Succ, D, Dom, (C+Succ)
847 // is Qout + max(F, Qin) * U + min(F, Qin)
848 // compare P + U vs Qout + P * U + Qin.
850 // The 3rd and 4th cases cover when Dom would be chosen to follow Succ.
852 // For the 3rd case, the cost is P + 2 * V
853 // For the 4th case, the cost is Qout + min(Qin, F) * U + max(Qin, F) * V + V
854 // We choose 4 over 3 when (P + V) > Qout + min(Qin, F) * U + max(Qin, F) * V
855 if (UProb > AdjustedSuccSumProb / 2 &&
856 !hasBetterLayoutPredecessor(Succ, PDom, *BlockToChain[PDom], UProb, UProb,
859 return greaterWithBias(
860 (P + V), (Qout + std::max(Qin, F) * VProb + std::min(Qin, F) * UProb),
863 return greaterWithBias((P + U),
864 (Qout + std::min(Qin, F) * AdjustedSuccSumProb +
865 std::max(Qin, F) * UProb),
869 /// Check for a trellis layout. \p BB is the upper part of a trellis if its
870 /// successors form the lower part of a trellis. A successor set S forms the
871 /// lower part of a trellis if all of the predecessors of S are either in S or
872 /// have all of S as successors. We ignore trellises where BB doesn't have 2
873 /// successors because for fewer than 2, it's trivial, and for 3 or greater they
874 /// are very uncommon and complex to compute optimally. Allowing edges within S
875 /// is not strictly a trellis, but the same algorithm works, so we allow it.
876 bool MachineBlockPlacement::isTrellis(
877 const MachineBasicBlock *BB,
878 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
879 const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
880 // Technically BB could form a trellis with branching factor higher than 2.
881 // But that's extremely uncommon.
882 if (BB->succ_size() != 2 || ViableSuccs.size() != 2)
885 SmallPtrSet<const MachineBasicBlock *, 2> Successors(BB->succ_begin(),
887 // To avoid reviewing the same predecessors twice.
888 SmallPtrSet<const MachineBasicBlock *, 8> SeenPreds;
890 for (MachineBasicBlock *Succ : ViableSuccs) {
892 for (auto SuccPred : Succ->predecessors()) {
893 // Allow triangle successors, but don't count them.
894 if (Successors.count(SuccPred)) {
895 // Make sure that it is actually a triangle.
896 for (MachineBasicBlock *CheckSucc : SuccPred->successors())
897 if (!Successors.count(CheckSucc))
901 const BlockChain *PredChain = BlockToChain[SuccPred];
902 if (SuccPred == BB || (BlockFilter && !BlockFilter->count(SuccPred)) ||
903 PredChain == &Chain || PredChain == BlockToChain[Succ])
906 // Perform the successor check only once.
907 if (!SeenPreds.insert(SuccPred).second)
909 if (!hasSameSuccessors(*SuccPred, Successors))
912 // If one of the successors has only BB as a predecessor, it is not a
920 /// Pick the highest total weight pair of edges that can both be laid out.
921 /// The edges in \p Edges[0] are assumed to have a different destination than
922 /// the edges in \p Edges[1]. Simple counting shows that the best pair is either
923 /// the individual highest weight edges to the 2 different destinations, or in
924 /// case of a conflict, one of them should be replaced with a 2nd best edge.
925 std::pair<MachineBlockPlacement::WeightedEdge,
926 MachineBlockPlacement::WeightedEdge>
927 MachineBlockPlacement::getBestNonConflictingEdges(
928 const MachineBasicBlock *BB,
929 MutableArrayRef<SmallVector<MachineBlockPlacement::WeightedEdge, 8>>
931 // Sort the edges, and then for each successor, find the best incoming
932 // predecessor. If the best incoming predecessors aren't the same,
933 // then that is clearly the best layout. If there is a conflict, one of the
934 // successors will have to fallthrough from the second best predecessor. We
935 // compare which combination is better overall.
937 // Sort for highest frequency.
938 auto Cmp = [](WeightedEdge A, WeightedEdge B) { return A.Weight > B.Weight; };
940 std::stable_sort(Edges[0].begin(), Edges[0].end(), Cmp);
941 std::stable_sort(Edges[1].begin(), Edges[1].end(), Cmp);
942 auto BestA = Edges[0].begin();
943 auto BestB = Edges[1].begin();
944 // Arrange for the correct answer to be in BestA and BestB
945 // If the 2 best edges don't conflict, the answer is already there.
946 if (BestA->Src == BestB->Src) {
947 // Compare the total fallthrough of (Best + Second Best) for both pairs
948 auto SecondBestA = std::next(BestA);
949 auto SecondBestB = std::next(BestB);
950 BlockFrequency BestAScore = BestA->Weight + SecondBestB->Weight;
951 BlockFrequency BestBScore = BestB->Weight + SecondBestA->Weight;
952 if (BestAScore < BestBScore)
957 // Arrange for the BB edge to be in BestA if it exists.
958 if (BestB->Src == BB)
959 std::swap(BestA, BestB);
960 return std::make_pair(*BestA, *BestB);
963 /// Get the best successor from \p BB based on \p BB being part of a trellis.
964 /// We only handle trellises with 2 successors, so the algorithm is
965 /// straightforward: Find the best pair of edges that don't conflict. We find
966 /// the best incoming edge for each successor in the trellis. If those conflict,
967 /// we consider which of them should be replaced with the second best.
968 /// Upon return the two best edges will be in \p BestEdges. If one of the edges
969 /// comes from \p BB, it will be in \p BestEdges[0]
970 MachineBlockPlacement::BlockAndTailDupResult
971 MachineBlockPlacement::getBestTrellisSuccessor(
972 const MachineBasicBlock *BB,
973 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
974 BranchProbability AdjustedSumProb, const BlockChain &Chain,
975 const BlockFilterSet *BlockFilter) {
977 BlockAndTailDupResult Result = {nullptr, false};
978 SmallPtrSet<const MachineBasicBlock *, 4> Successors(BB->succ_begin(),
981 // We assume size 2 because it's common. For general n, we would have to do
982 // the Hungarian algorithm, but it's not worth the complexity because more
983 // than 2 successors is fairly uncommon, and a trellis even more so.
984 if (Successors.size() != 2 || ViableSuccs.size() != 2)
987 // Collect the edge frequencies of all edges that form the trellis.
988 SmallVector<WeightedEdge, 8> Edges[2];
990 for (auto Succ : ViableSuccs) {
991 for (MachineBasicBlock *SuccPred : Succ->predecessors()) {
992 // Skip any placed predecessors that are not BB
994 if ((BlockFilter && !BlockFilter->count(SuccPred)) ||
995 BlockToChain[SuccPred] == &Chain ||
996 BlockToChain[SuccPred] == BlockToChain[Succ])
998 BlockFrequency EdgeFreq = MBFI->getBlockFreq(SuccPred) *
999 MBPI->getEdgeProbability(SuccPred, Succ);
1000 Edges[SuccIndex].push_back({EdgeFreq, SuccPred, Succ});
1005 // Pick the best combination of 2 edges from all the edges in the trellis.
1006 WeightedEdge BestA, BestB;
1007 std::tie(BestA, BestB) = getBestNonConflictingEdges(BB, Edges);
1009 if (BestA.Src != BB) {
1010 // If we have a trellis, and BB doesn't have the best fallthrough edges,
1011 // we shouldn't choose any successor. We've already looked and there's a
1012 // better fallthrough edge for all the successors.
1013 DEBUG(dbgs() << "Trellis, but not one of the chosen edges.\n");
1017 // Did we pick the triangle edge? If tail-duplication is profitable, do
1018 // that instead. Otherwise merge the triangle edge now while we know it is
1020 if (BestA.Dest == BestB.Src) {
1021 // The edges are BB->Succ1->Succ2, and we're looking to see if BB->Succ2
1023 MachineBasicBlock *Succ1 = BestA.Dest;
1024 MachineBasicBlock *Succ2 = BestB.Dest;
1025 // Check to see if tail-duplication would be profitable.
1026 if (allowTailDupPlacement() && shouldTailDuplicate(Succ2) &&
1027 canTailDuplicateUnplacedPreds(BB, Succ2, Chain, BlockFilter) &&
1028 isProfitableToTailDup(BB, Succ2, MBPI->getEdgeProbability(BB, Succ1),
1029 Chain, BlockFilter)) {
1030 DEBUG(BranchProbability Succ2Prob = getAdjustedProbability(
1031 MBPI->getEdgeProbability(BB, Succ2), AdjustedSumProb);
1032 dbgs() << " Selected: " << getBlockName(Succ2)
1033 << ", probability: " << Succ2Prob << " (Tail Duplicate)\n");
1035 Result.ShouldTailDup = true;
1039 // We have already computed the optimal edge for the other side of the
1041 ComputedEdges[BestB.Src] = { BestB.Dest, false };
1043 auto TrellisSucc = BestA.Dest;
1044 DEBUG(BranchProbability SuccProb = getAdjustedProbability(
1045 MBPI->getEdgeProbability(BB, TrellisSucc), AdjustedSumProb);
1046 dbgs() << " Selected: " << getBlockName(TrellisSucc)
1047 << ", probability: " << SuccProb << " (Trellis)\n");
1048 Result.BB = TrellisSucc;
1052 /// When the option allowTailDupPlacement() is on, this method checks if the
1053 /// fallthrough candidate block \p Succ (of block \p BB) can be tail-duplicated
1054 /// into all of its unplaced, unfiltered predecessors, that are not BB.
1055 bool MachineBlockPlacement::canTailDuplicateUnplacedPreds(
1056 const MachineBasicBlock *BB, MachineBasicBlock *Succ,
1057 const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
1058 if (!shouldTailDuplicate(Succ))
1061 // For CFG checking.
1062 SmallPtrSet<const MachineBasicBlock *, 4> Successors(BB->succ_begin(),
1064 for (MachineBasicBlock *Pred : Succ->predecessors()) {
1065 // Make sure all unplaced and unfiltered predecessors can be
1066 // tail-duplicated into.
1067 // Skip any blocks that are already placed or not in this loop.
1068 if (Pred == BB || (BlockFilter && !BlockFilter->count(Pred))
1069 || BlockToChain[Pred] == &Chain)
1071 if (!TailDup.canTailDuplicate(Succ, Pred)) {
1072 if (Successors.size() > 1 && hasSameSuccessors(*Pred, Successors))
1073 // This will result in a trellis after tail duplication, so we don't
1074 // need to copy Succ into this predecessor. In the presence
1075 // of a trellis tail duplication can continue to be profitable.
1091 // After BB was duplicated into C, the layout looks like the one on the
1092 // right. BB and C now have the same successors. When considering
1093 // whether Succ can be duplicated into all its unplaced predecessors, we
1095 // We can do this because C already has a profitable fallthrough, namely
1096 // D. TODO(iteratee): ignore sufficiently cold predecessors for
1097 // duplication and for this test.
1099 // This allows trellises to be laid out in 2 separate chains
1100 // (A,B,Succ,...) and later (C,D,...) This is a reasonable heuristic
1101 // because it allows the creation of 2 fallthrough paths with links
1102 // between them, and we correctly identify the best layout for these
1103 // CFGs. We want to extend trellises that the user created in addition
1104 // to trellises created by tail-duplication, so we just look for the
1113 /// Find chains of triangles where we believe it would be profitable to
1114 /// tail-duplicate them all, but a local analysis would not find them.
1115 /// There are 3 ways this can be profitable:
1116 /// 1) The post-dominators marked 50% are actually taken 55% (This shrinks with
1118 /// 2) The chains are statically correlated. Branch probabilities have a very
1119 /// U-shaped distribution.
1120 /// [http://nrs.harvard.edu/urn-3:HUL.InstRepos:24015805]
1121 /// If the branches in a chain are likely to be from the same side of the
1122 /// distribution as their predecessor, but are independent at runtime, this
1123 /// transformation is profitable. (Because the cost of being wrong is a small
1124 /// fixed cost, unlike the standard triangle layout where the cost of being
1125 /// wrong scales with the # of triangles.)
1126 /// 3) The chains are dynamically correlated. If the probability that a previous
1127 /// branch was taken positively influences whether the next branch will be
1129 /// We believe that 2 and 3 are common enough to justify the small margin in 1.
1130 void MachineBlockPlacement::precomputeTriangleChains() {
1131 struct TriangleChain {
1132 std::vector<MachineBasicBlock *> Edges;
1134 TriangleChain(MachineBasicBlock *src, MachineBasicBlock *dst)
1135 : Edges({src, dst}) {}
1137 void append(MachineBasicBlock *dst) {
1138 assert(getKey()->isSuccessor(dst) &&
1139 "Attempting to append a block that is not a successor.");
1140 Edges.push_back(dst);
1143 unsigned count() const { return Edges.size() - 1; }
1145 MachineBasicBlock *getKey() const {
1146 return Edges.back();
1150 if (TriangleChainCount == 0)
1153 DEBUG(dbgs() << "Pre-computing triangle chains.\n");
1154 // Map from last block to the chain that contains it. This allows us to extend
1155 // chains as we find new triangles.
1156 DenseMap<const MachineBasicBlock *, TriangleChain> TriangleChainMap;
1157 for (MachineBasicBlock &BB : *F) {
1158 // If BB doesn't have 2 successors, it doesn't start a triangle.
1159 if (BB.succ_size() != 2)
1161 MachineBasicBlock *PDom = nullptr;
1162 for (MachineBasicBlock *Succ : BB.successors()) {
1163 if (!MPDT->dominates(Succ, &BB))
1168 // If BB doesn't have a post-dominating successor, it doesn't form a
1170 if (PDom == nullptr)
1172 // If PDom has a hint that it is low probability, skip this triangle.
1173 if (MBPI->getEdgeProbability(&BB, PDom) < BranchProbability(50, 100))
1175 // If PDom isn't eligible for duplication, this isn't the kind of triangle
1176 // we're looking for.
1177 if (!shouldTailDuplicate(PDom))
1179 bool CanTailDuplicate = true;
1180 // If PDom can't tail-duplicate into it's non-BB predecessors, then this
1181 // isn't the kind of triangle we're looking for.
1182 for (MachineBasicBlock* Pred : PDom->predecessors()) {
1185 if (!TailDup.canTailDuplicate(PDom, Pred)) {
1186 CanTailDuplicate = false;
1190 // If we can't tail-duplicate PDom to its predecessors, then skip this
1192 if (!CanTailDuplicate)
1195 // Now we have an interesting triangle. Insert it if it's not part of an
1197 // Note: This cannot be replaced with a call insert() or emplace() because
1198 // the find key is BB, but the insert/emplace key is PDom.
1199 auto Found = TriangleChainMap.find(&BB);
1200 // If it is, remove the chain from the map, grow it, and put it back in the
1201 // map with the end as the new key.
1202 if (Found != TriangleChainMap.end()) {
1203 TriangleChain Chain = std::move(Found->second);
1204 TriangleChainMap.erase(Found);
1206 TriangleChainMap.insert(std::make_pair(Chain.getKey(), std::move(Chain)));
1208 auto InsertResult = TriangleChainMap.try_emplace(PDom, &BB, PDom);
1209 assert(InsertResult.second && "Block seen twice.");
1214 // Iterating over a DenseMap is safe here, because the only thing in the body
1215 // of the loop is inserting into another DenseMap (ComputedEdges).
1216 // ComputedEdges is never iterated, so this doesn't lead to non-determinism.
1217 for (auto &ChainPair : TriangleChainMap) {
1218 TriangleChain &Chain = ChainPair.second;
1219 // Benchmarking has shown that due to branch correlation duplicating 2 or
1220 // more triangles is profitable, despite the calculations assuming
1222 if (Chain.count() < TriangleChainCount)
1224 MachineBasicBlock *dst = Chain.Edges.back();
1225 Chain.Edges.pop_back();
1226 for (MachineBasicBlock *src : reverse(Chain.Edges)) {
1227 DEBUG(dbgs() << "Marking edge: " << getBlockName(src) << "->" <<
1228 getBlockName(dst) << " as pre-computed based on triangles.\n");
1230 auto InsertResult = ComputedEdges.insert({src, {dst, true}});
1231 assert(InsertResult.second && "Block seen twice.");
1239 // When profile is not present, return the StaticLikelyProb.
1240 // When profile is available, we need to handle the triangle-shape CFG.
1241 static BranchProbability getLayoutSuccessorProbThreshold(
1242 const MachineBasicBlock *BB) {
1243 if (!BB->getParent()->getFunction().hasProfileData())
1244 return BranchProbability(StaticLikelyProb, 100);
1245 if (BB->succ_size() == 2) {
1246 const MachineBasicBlock *Succ1 = *BB->succ_begin();
1247 const MachineBasicBlock *Succ2 = *(BB->succ_begin() + 1);
1248 if (Succ1->isSuccessor(Succ2) || Succ2->isSuccessor(Succ1)) {
1249 /* See case 1 below for the cost analysis. For BB->Succ to
1250 * be taken with smaller cost, the following needs to hold:
1251 * Prob(BB->Succ) > 2 * Prob(BB->Pred)
1252 * So the threshold T in the calculation below
1253 * (1-T) * Prob(BB->Succ) > T * Prob(BB->Pred)
1254 * So T / (1 - T) = 2, Yielding T = 2/3
1255 * Also adding user specified branch bias, we have
1256 * T = (2/3)*(ProfileLikelyProb/50)
1257 * = (2*ProfileLikelyProb)/150)
1259 return BranchProbability(2 * ProfileLikelyProb, 150);
1262 return BranchProbability(ProfileLikelyProb, 100);
1265 /// Checks to see if the layout candidate block \p Succ has a better layout
1266 /// predecessor than \c BB. If yes, returns true.
1267 /// \p SuccProb: The probability adjusted for only remaining blocks.
1268 /// Only used for logging
1269 /// \p RealSuccProb: The un-adjusted probability.
1270 /// \p Chain: The chain that BB belongs to and Succ is being considered for.
1271 /// \p BlockFilter: if non-null, the set of blocks that make up the loop being
1273 bool MachineBlockPlacement::hasBetterLayoutPredecessor(
1274 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
1275 const BlockChain &SuccChain, BranchProbability SuccProb,
1276 BranchProbability RealSuccProb, const BlockChain &Chain,
1277 const BlockFilterSet *BlockFilter) {
1279 // There isn't a better layout when there are no unscheduled predecessors.
1280 if (SuccChain.UnscheduledPredecessors == 0)
1283 // There are two basic scenarios here:
1284 // -------------------------------------
1285 // Case 1: triangular shape CFG (if-then):
1292 // In this case, we are evaluating whether to select edge -> Succ, e.g.
1293 // set Succ as the layout successor of BB. Picking Succ as BB's
1294 // successor breaks the CFG constraints (FIXME: define these constraints).
1295 // With this layout, Pred BB
1296 // is forced to be outlined, so the overall cost will be cost of the
1297 // branch taken from BB to Pred, plus the cost of back taken branch
1298 // from Pred to Succ, as well as the additional cost associated
1299 // with the needed unconditional jump instruction from Pred To Succ.
1301 // The cost of the topological order layout is the taken branch cost
1302 // from BB to Succ, so to make BB->Succ a viable candidate, the following
1304 // 2 * freq(BB->Pred) * taken_branch_cost + unconditional_jump_cost
1305 // < freq(BB->Succ) * taken_branch_cost.
1306 // Ignoring unconditional jump cost, we get
1307 // freq(BB->Succ) > 2 * freq(BB->Pred), i.e.,
1308 // prob(BB->Succ) > 2 * prob(BB->Pred)
1310 // When real profile data is available, we can precisely compute the
1311 // probability threshold that is needed for edge BB->Succ to be considered.
1312 // Without profile data, the heuristic requires the branch bias to be
1313 // a lot larger to make sure the signal is very strong (e.g. 80% default).
1314 // -----------------------------------------------------------------
1315 // Case 2: diamond like CFG (if-then-else):
1324 // The current block is BB and edge BB->Succ is now being evaluated.
1325 // Note that edge S->BB was previously already selected because
1326 // prob(S->BB) > prob(S->Pred).
1327 // At this point, 2 blocks can be placed after BB: Pred or Succ. If we
1328 // choose Pred, we will have a topological ordering as shown on the left
1329 // in the picture below. If we choose Succ, we have the solution as shown
1338 // | Pred-- | Succ--
1340 // ---Succ ---Pred--
1342 // cost = freq(S->Pred) + freq(BB->Succ) cost = 2 * freq (S->Pred)
1343 // = freq(S->Pred) + freq(S->BB)
1345 // If we have profile data (i.e, branch probabilities can be trusted), the
1346 // cost (number of taken branches) with layout S->BB->Succ->Pred is 2 *
1347 // freq(S->Pred) while the cost of topo order is freq(S->Pred) + freq(S->BB).
1348 // We know Prob(S->BB) > Prob(S->Pred), so freq(S->BB) > freq(S->Pred), which
1349 // means the cost of topological order is greater.
1350 // When profile data is not available, however, we need to be more
1351 // conservative. If the branch prediction is wrong, breaking the topo-order
1352 // will actually yield a layout with large cost. For this reason, we need
1353 // strong biased branch at block S with Prob(S->BB) in order to select
1354 // BB->Succ. This is equivalent to looking the CFG backward with backward
1355 // edge: Prob(Succ->BB) needs to >= HotProb in order to be selected (without
1357 // --------------------------------------------------------------------------
1358 // Case 3: forked diamond
1370 // The current block is BB and edge BB->S1 is now being evaluated.
1371 // As above S->BB was already selected because
1372 // prob(S->BB) > prob(S->Pred). Assume that prob(BB->S1) >= prob(BB->S2).
1380 // | Pred----| | S1----
1382 // --(S1 or S2) ---Pred--
1386 // topo-cost = freq(S->Pred) + freq(BB->S1) + freq(BB->S2)
1387 // + min(freq(Pred->S1), freq(Pred->S2))
1388 // Non-topo-order cost:
1389 // non-topo-cost = 2 * freq(S->Pred) + freq(BB->S2).
1390 // To be conservative, we can assume that min(freq(Pred->S1), freq(Pred->S2))
1391 // is 0. Then the non topo layout is better when
1392 // freq(S->Pred) < freq(BB->S1).
1393 // This is exactly what is checked below.
1394 // Note there are other shapes that apply (Pred may not be a single block,
1395 // but they all fit this general pattern.)
1396 BranchProbability HotProb = getLayoutSuccessorProbThreshold(BB);
1398 // Make sure that a hot successor doesn't have a globally more
1399 // important predecessor.
1400 BlockFrequency CandidateEdgeFreq = MBFI->getBlockFreq(BB) * RealSuccProb;
1401 bool BadCFGConflict = false;
1403 for (MachineBasicBlock *Pred : Succ->predecessors()) {
1404 if (Pred == Succ || BlockToChain[Pred] == &SuccChain ||
1405 (BlockFilter && !BlockFilter->count(Pred)) ||
1406 BlockToChain[Pred] == &Chain ||
1407 // This check is redundant except for look ahead. This function is
1408 // called for lookahead by isProfitableToTailDup when BB hasn't been
1412 // Do backward checking.
1413 // For all cases above, we need a backward checking to filter out edges that
1414 // are not 'strongly' biased.
1418 // We select edge BB->Succ if
1419 // freq(BB->Succ) > freq(Succ) * HotProb
1420 // i.e. freq(BB->Succ) > freq(BB->Succ) * HotProb + freq(Pred->Succ) *
1422 // i.e. freq((BB->Succ) * (1 - HotProb) > freq(Pred->Succ) * HotProb
1423 // Case 1 is covered too, because the first equation reduces to:
1424 // prob(BB->Succ) > HotProb. (freq(Succ) = freq(BB) for a triangle)
1425 BlockFrequency PredEdgeFreq =
1426 MBFI->getBlockFreq(Pred) * MBPI->getEdgeProbability(Pred, Succ);
1427 if (PredEdgeFreq * HotProb >= CandidateEdgeFreq * HotProb.getCompl()) {
1428 BadCFGConflict = true;
1433 if (BadCFGConflict) {
1434 DEBUG(dbgs() << " Not a candidate: " << getBlockName(Succ) << " -> " << SuccProb
1435 << " (prob) (non-cold CFG conflict)\n");
1442 /// \brief Select the best successor for a block.
1444 /// This looks across all successors of a particular block and attempts to
1445 /// select the "best" one to be the layout successor. It only considers direct
1446 /// successors which also pass the block filter. It will attempt to avoid
1447 /// breaking CFG structure, but cave and break such structures in the case of
1448 /// very hot successor edges.
1450 /// \returns The best successor block found, or null if none are viable, along
1451 /// with a boolean indicating if tail duplication is necessary.
1452 MachineBlockPlacement::BlockAndTailDupResult
1453 MachineBlockPlacement::selectBestSuccessor(
1454 const MachineBasicBlock *BB, const BlockChain &Chain,
1455 const BlockFilterSet *BlockFilter) {
1456 const BranchProbability HotProb(StaticLikelyProb, 100);
1458 BlockAndTailDupResult BestSucc = { nullptr, false };
1459 auto BestProb = BranchProbability::getZero();
1461 SmallVector<MachineBasicBlock *, 4> Successors;
1462 auto AdjustedSumProb =
1463 collectViableSuccessors(BB, Chain, BlockFilter, Successors);
1465 DEBUG(dbgs() << "Selecting best successor for: " << getBlockName(BB) << "\n");
1467 // if we already precomputed the best successor for BB, return that if still
1469 auto FoundEdge = ComputedEdges.find(BB);
1470 if (FoundEdge != ComputedEdges.end()) {
1471 MachineBasicBlock *Succ = FoundEdge->second.BB;
1472 ComputedEdges.erase(FoundEdge);
1473 BlockChain *SuccChain = BlockToChain[Succ];
1474 if (BB->isSuccessor(Succ) && (!BlockFilter || BlockFilter->count(Succ)) &&
1475 SuccChain != &Chain && Succ == *SuccChain->begin())
1476 return FoundEdge->second;
1479 // if BB is part of a trellis, Use the trellis to determine the optimal
1480 // fallthrough edges
1481 if (isTrellis(BB, Successors, Chain, BlockFilter))
1482 return getBestTrellisSuccessor(BB, Successors, AdjustedSumProb, Chain,
1485 // For blocks with CFG violations, we may be able to lay them out anyway with
1486 // tail-duplication. We keep this vector so we can perform the probability
1487 // calculations the minimum number of times.
1488 SmallVector<std::tuple<BranchProbability, MachineBasicBlock *>, 4>
1490 for (MachineBasicBlock *Succ : Successors) {
1491 auto RealSuccProb = MBPI->getEdgeProbability(BB, Succ);
1492 BranchProbability SuccProb =
1493 getAdjustedProbability(RealSuccProb, AdjustedSumProb);
1495 BlockChain &SuccChain = *BlockToChain[Succ];
1496 // Skip the edge \c BB->Succ if block \c Succ has a better layout
1497 // predecessor that yields lower global cost.
1498 if (hasBetterLayoutPredecessor(BB, Succ, SuccChain, SuccProb, RealSuccProb,
1499 Chain, BlockFilter)) {
1500 // If tail duplication would make Succ profitable, place it.
1501 if (allowTailDupPlacement() && shouldTailDuplicate(Succ))
1502 DupCandidates.push_back(std::make_tuple(SuccProb, Succ));
1507 dbgs() << " Candidate: " << getBlockName(Succ) << ", probability: "
1509 << (SuccChain.UnscheduledPredecessors != 0 ? " (CFG break)" : "")
1512 if (BestSucc.BB && BestProb >= SuccProb) {
1513 DEBUG(dbgs() << " Not the best candidate, continuing\n");
1517 DEBUG(dbgs() << " Setting it as best candidate\n");
1519 BestProb = SuccProb;
1521 // Handle the tail duplication candidates in order of decreasing probability.
1522 // Stop at the first one that is profitable. Also stop if they are less
1523 // profitable than BestSucc. Position is important because we preserve it and
1524 // prefer first best match. Here we aren't comparing in order, so we capture
1525 // the position instead.
1526 if (DupCandidates.size() != 0) {
1528 [](const std::tuple<BranchProbability, MachineBasicBlock *> &a,
1529 const std::tuple<BranchProbability, MachineBasicBlock *> &b) {
1530 return std::get<0>(a) > std::get<0>(b);
1532 std::stable_sort(DupCandidates.begin(), DupCandidates.end(), cmp);
1534 for(auto &Tup : DupCandidates) {
1535 BranchProbability DupProb;
1536 MachineBasicBlock *Succ;
1537 std::tie(DupProb, Succ) = Tup;
1538 if (DupProb < BestProb)
1540 if (canTailDuplicateUnplacedPreds(BB, Succ, Chain, BlockFilter)
1541 && (isProfitableToTailDup(BB, Succ, BestProb, Chain, BlockFilter))) {
1543 dbgs() << " Candidate: " << getBlockName(Succ) << ", probability: "
1545 << " (Tail Duplicate)\n");
1547 BestSucc.ShouldTailDup = true;
1553 DEBUG(dbgs() << " Selected: " << getBlockName(BestSucc.BB) << "\n");
1558 /// \brief Select the best block from a worklist.
1560 /// This looks through the provided worklist as a list of candidate basic
1561 /// blocks and select the most profitable one to place. The definition of
1562 /// profitable only really makes sense in the context of a loop. This returns
1563 /// the most frequently visited block in the worklist, which in the case of
1564 /// a loop, is the one most desirable to be physically close to the rest of the
1565 /// loop body in order to improve i-cache behavior.
1567 /// \returns The best block found, or null if none are viable.
1568 MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock(
1569 const BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList) {
1570 // Once we need to walk the worklist looking for a candidate, cleanup the
1571 // worklist of already placed entries.
1572 // FIXME: If this shows up on profiles, it could be folded (at the cost of
1573 // some code complexity) into the loop below.
1574 WorkList.erase(llvm::remove_if(WorkList,
1575 [&](MachineBasicBlock *BB) {
1576 return BlockToChain.lookup(BB) == &Chain;
1580 if (WorkList.empty())
1583 bool IsEHPad = WorkList[0]->isEHPad();
1585 MachineBasicBlock *BestBlock = nullptr;
1586 BlockFrequency BestFreq;
1587 for (MachineBasicBlock *MBB : WorkList) {
1588 assert(MBB->isEHPad() == IsEHPad &&
1589 "EHPad mismatch between block and work list.");
1591 BlockChain &SuccChain = *BlockToChain[MBB];
1592 if (&SuccChain == &Chain)
1595 assert(SuccChain.UnscheduledPredecessors == 0 &&
1596 "Found CFG-violating block");
1598 BlockFrequency CandidateFreq = MBFI->getBlockFreq(MBB);
1599 DEBUG(dbgs() << " " << getBlockName(MBB) << " -> ";
1600 MBFI->printBlockFreq(dbgs(), CandidateFreq) << " (freq)\n");
1602 // For ehpad, we layout the least probable first as to avoid jumping back
1603 // from least probable landingpads to more probable ones.
1605 // FIXME: Using probability is probably (!) not the best way to achieve
1606 // this. We should probably have a more principled approach to layout
1609 // The goal is to get:
1611 // +--------------------------+
1613 // InnerLp -> InnerCleanup OuterLp -> OuterCleanup -> Resume
1617 // +-------------------------------------+
1619 // OuterLp -> OuterCleanup -> Resume InnerLp -> InnerCleanup
1620 if (BestBlock && (IsEHPad ^ (BestFreq >= CandidateFreq)))
1624 BestFreq = CandidateFreq;
1630 /// \brief Retrieve the first unplaced basic block.
1632 /// This routine is called when we are unable to use the CFG to walk through
1633 /// all of the basic blocks and form a chain due to unnatural loops in the CFG.
1634 /// We walk through the function's blocks in order, starting from the
1635 /// LastUnplacedBlockIt. We update this iterator on each call to avoid
1636 /// re-scanning the entire sequence on repeated calls to this routine.
1637 MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock(
1638 const BlockChain &PlacedChain,
1639 MachineFunction::iterator &PrevUnplacedBlockIt,
1640 const BlockFilterSet *BlockFilter) {
1641 for (MachineFunction::iterator I = PrevUnplacedBlockIt, E = F->end(); I != E;
1643 if (BlockFilter && !BlockFilter->count(&*I))
1645 if (BlockToChain[&*I] != &PlacedChain) {
1646 PrevUnplacedBlockIt = I;
1647 // Now select the head of the chain to which the unplaced block belongs
1648 // as the block to place. This will force the entire chain to be placed,
1649 // and satisfies the requirements of merging chains.
1650 return *BlockToChain[&*I]->begin();
1656 void MachineBlockPlacement::fillWorkLists(
1657 const MachineBasicBlock *MBB,
1658 SmallPtrSetImpl<BlockChain *> &UpdatedPreds,
1659 const BlockFilterSet *BlockFilter = nullptr) {
1660 BlockChain &Chain = *BlockToChain[MBB];
1661 if (!UpdatedPreds.insert(&Chain).second)
1665 Chain.UnscheduledPredecessors == 0 &&
1666 "Attempting to place block with unscheduled predecessors in worklist.");
1667 for (MachineBasicBlock *ChainBB : Chain) {
1668 assert(BlockToChain[ChainBB] == &Chain &&
1669 "Block in chain doesn't match BlockToChain map.");
1670 for (MachineBasicBlock *Pred : ChainBB->predecessors()) {
1671 if (BlockFilter && !BlockFilter->count(Pred))
1673 if (BlockToChain[Pred] == &Chain)
1675 ++Chain.UnscheduledPredecessors;
1679 if (Chain.UnscheduledPredecessors != 0)
1682 MachineBasicBlock *BB = *Chain.begin();
1684 EHPadWorkList.push_back(BB);
1686 BlockWorkList.push_back(BB);
1689 void MachineBlockPlacement::buildChain(
1690 const MachineBasicBlock *HeadBB, BlockChain &Chain,
1691 BlockFilterSet *BlockFilter) {
1692 assert(HeadBB && "BB must not be null.\n");
1693 assert(BlockToChain[HeadBB] == &Chain && "BlockToChainMap mis-match.\n");
1694 MachineFunction::iterator PrevUnplacedBlockIt = F->begin();
1696 const MachineBasicBlock *LoopHeaderBB = HeadBB;
1697 markChainSuccessors(Chain, LoopHeaderBB, BlockFilter);
1698 MachineBasicBlock *BB = *std::prev(Chain.end());
1700 assert(BB && "null block found at end of chain in loop.");
1701 assert(BlockToChain[BB] == &Chain && "BlockToChainMap mis-match in loop.");
1702 assert(*std::prev(Chain.end()) == BB && "BB Not found at end of chain.");
1705 // Look for the best viable successor if there is one to place immediately
1706 // after this block.
1707 auto Result = selectBestSuccessor(BB, Chain, BlockFilter);
1708 MachineBasicBlock* BestSucc = Result.BB;
1709 bool ShouldTailDup = Result.ShouldTailDup;
1710 if (allowTailDupPlacement())
1711 ShouldTailDup |= (BestSucc && shouldTailDuplicate(BestSucc));
1713 // If an immediate successor isn't available, look for the best viable
1714 // block among those we've identified as not violating the loop's CFG at
1715 // this point. This won't be a fallthrough, but it will increase locality.
1717 BestSucc = selectBestCandidateBlock(Chain, BlockWorkList);
1719 BestSucc = selectBestCandidateBlock(Chain, EHPadWorkList);
1722 BestSucc = getFirstUnplacedBlock(Chain, PrevUnplacedBlockIt, BlockFilter);
1726 DEBUG(dbgs() << "Unnatural loop CFG detected, forcibly merging the "
1727 "layout successor until the CFG reduces\n");
1730 // Placement may have changed tail duplication opportunities.
1731 // Check for that now.
1732 if (allowTailDupPlacement() && BestSucc && ShouldTailDup) {
1733 // If the chosen successor was duplicated into all its predecessors,
1734 // don't bother laying it out, just go round the loop again with BB as
1736 if (repeatedlyTailDuplicateBlock(BestSucc, BB, LoopHeaderBB, Chain,
1737 BlockFilter, PrevUnplacedBlockIt))
1741 // Place this block, updating the datastructures to reflect its placement.
1742 BlockChain &SuccChain = *BlockToChain[BestSucc];
1743 // Zero out UnscheduledPredecessors for the successor we're about to merge in case
1744 // we selected a successor that didn't fit naturally into the CFG.
1745 SuccChain.UnscheduledPredecessors = 0;
1746 DEBUG(dbgs() << "Merging from " << getBlockName(BB) << " to "
1747 << getBlockName(BestSucc) << "\n");
1748 markChainSuccessors(SuccChain, LoopHeaderBB, BlockFilter);
1749 Chain.merge(BestSucc, &SuccChain);
1750 BB = *std::prev(Chain.end());
1753 DEBUG(dbgs() << "Finished forming chain for header block "
1754 << getBlockName(*Chain.begin()) << "\n");
1757 /// \brief Find the best loop top block for layout.
1759 /// Look for a block which is strictly better than the loop header for laying
1760 /// out at the top of the loop. This looks for one and only one pattern:
1761 /// a latch block with no conditional exit. This block will cause a conditional
1762 /// jump around it or will be the bottom of the loop if we lay it out in place,
1763 /// but if it it doesn't end up at the bottom of the loop for any reason,
1764 /// rotation alone won't fix it. Because such a block will always result in an
1765 /// unconditional jump (for the backedge) rotating it in front of the loop
1766 /// header is always profitable.
1768 MachineBlockPlacement::findBestLoopTop(const MachineLoop &L,
1769 const BlockFilterSet &LoopBlockSet) {
1770 // Placing the latch block before the header may introduce an extra branch
1771 // that skips this block the first time the loop is executed, which we want
1772 // to avoid when optimising for size.
1773 // FIXME: in theory there is a case that does not introduce a new branch,
1774 // i.e. when the layout predecessor does not fallthrough to the loop header.
1775 // In practice this never happens though: there always seems to be a preheader
1776 // that can fallthrough and that is also placed before the header.
1777 if (F->getFunction().optForSize())
1778 return L.getHeader();
1780 // Check that the header hasn't been fused with a preheader block due to
1781 // crazy branches. If it has, we need to start with the header at the top to
1782 // prevent pulling the preheader into the loop body.
1783 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
1784 if (!LoopBlockSet.count(*HeaderChain.begin()))
1785 return L.getHeader();
1787 DEBUG(dbgs() << "Finding best loop top for: " << getBlockName(L.getHeader())
1790 BlockFrequency BestPredFreq;
1791 MachineBasicBlock *BestPred = nullptr;
1792 for (MachineBasicBlock *Pred : L.getHeader()->predecessors()) {
1793 if (!LoopBlockSet.count(Pred))
1795 DEBUG(dbgs() << " header pred: " << getBlockName(Pred) << ", has "
1796 << Pred->succ_size() << " successors, ";
1797 MBFI->printBlockFreq(dbgs(), Pred) << " freq\n");
1798 if (Pred->succ_size() > 1)
1801 BlockFrequency PredFreq = MBFI->getBlockFreq(Pred);
1802 if (!BestPred || PredFreq > BestPredFreq ||
1803 (!(PredFreq < BestPredFreq) &&
1804 Pred->isLayoutSuccessor(L.getHeader()))) {
1806 BestPredFreq = PredFreq;
1810 // If no direct predecessor is fine, just use the loop header.
1812 DEBUG(dbgs() << " final top unchanged\n");
1813 return L.getHeader();
1816 // Walk backwards through any straight line of predecessors.
1817 while (BestPred->pred_size() == 1 &&
1818 (*BestPred->pred_begin())->succ_size() == 1 &&
1819 *BestPred->pred_begin() != L.getHeader())
1820 BestPred = *BestPred->pred_begin();
1822 DEBUG(dbgs() << " final top: " << getBlockName(BestPred) << "\n");
1826 /// \brief Find the best loop exiting block for layout.
1828 /// This routine implements the logic to analyze the loop looking for the best
1829 /// block to layout at the top of the loop. Typically this is done to maximize
1830 /// fallthrough opportunities.
1832 MachineBlockPlacement::findBestLoopExit(const MachineLoop &L,
1833 const BlockFilterSet &LoopBlockSet) {
1834 // We don't want to layout the loop linearly in all cases. If the loop header
1835 // is just a normal basic block in the loop, we want to look for what block
1836 // within the loop is the best one to layout at the top. However, if the loop
1837 // header has be pre-merged into a chain due to predecessors not having
1838 // analyzable branches, *and* the predecessor it is merged with is *not* part
1839 // of the loop, rotating the header into the middle of the loop will create
1840 // a non-contiguous range of blocks which is Very Bad. So start with the
1841 // header and only rotate if safe.
1842 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
1843 if (!LoopBlockSet.count(*HeaderChain.begin()))
1846 BlockFrequency BestExitEdgeFreq;
1847 unsigned BestExitLoopDepth = 0;
1848 MachineBasicBlock *ExitingBB = nullptr;
1849 // If there are exits to outer loops, loop rotation can severely limit
1850 // fallthrough opportunities unless it selects such an exit. Keep a set of
1851 // blocks where rotating to exit with that block will reach an outer loop.
1852 SmallPtrSet<MachineBasicBlock *, 4> BlocksExitingToOuterLoop;
1854 DEBUG(dbgs() << "Finding best loop exit for: " << getBlockName(L.getHeader())
1856 for (MachineBasicBlock *MBB : L.getBlocks()) {
1857 BlockChain &Chain = *BlockToChain[MBB];
1858 // Ensure that this block is at the end of a chain; otherwise it could be
1859 // mid-way through an inner loop or a successor of an unanalyzable branch.
1860 if (MBB != *std::prev(Chain.end()))
1863 // Now walk the successors. We need to establish whether this has a viable
1864 // exiting successor and whether it has a viable non-exiting successor.
1865 // We store the old exiting state and restore it if a viable looping
1866 // successor isn't found.
1867 MachineBasicBlock *OldExitingBB = ExitingBB;
1868 BlockFrequency OldBestExitEdgeFreq = BestExitEdgeFreq;
1869 bool HasLoopingSucc = false;
1870 for (MachineBasicBlock *Succ : MBB->successors()) {
1871 if (Succ->isEHPad())
1875 BlockChain &SuccChain = *BlockToChain[Succ];
1876 // Don't split chains, either this chain or the successor's chain.
1877 if (&Chain == &SuccChain) {
1878 DEBUG(dbgs() << " exiting: " << getBlockName(MBB) << " -> "
1879 << getBlockName(Succ) << " (chain conflict)\n");
1883 auto SuccProb = MBPI->getEdgeProbability(MBB, Succ);
1884 if (LoopBlockSet.count(Succ)) {
1885 DEBUG(dbgs() << " looping: " << getBlockName(MBB) << " -> "
1886 << getBlockName(Succ) << " (" << SuccProb << ")\n");
1887 HasLoopingSucc = true;
1891 unsigned SuccLoopDepth = 0;
1892 if (MachineLoop *ExitLoop = MLI->getLoopFor(Succ)) {
1893 SuccLoopDepth = ExitLoop->getLoopDepth();
1894 if (ExitLoop->contains(&L))
1895 BlocksExitingToOuterLoop.insert(MBB);
1898 BlockFrequency ExitEdgeFreq = MBFI->getBlockFreq(MBB) * SuccProb;
1899 DEBUG(dbgs() << " exiting: " << getBlockName(MBB) << " -> "
1900 << getBlockName(Succ) << " [L:" << SuccLoopDepth << "] (";
1901 MBFI->printBlockFreq(dbgs(), ExitEdgeFreq) << ")\n");
1902 // Note that we bias this toward an existing layout successor to retain
1903 // incoming order in the absence of better information. The exit must have
1904 // a frequency higher than the current exit before we consider breaking
1906 BranchProbability Bias(100 - ExitBlockBias, 100);
1907 if (!ExitingBB || SuccLoopDepth > BestExitLoopDepth ||
1908 ExitEdgeFreq > BestExitEdgeFreq ||
1909 (MBB->isLayoutSuccessor(Succ) &&
1910 !(ExitEdgeFreq < BestExitEdgeFreq * Bias))) {
1911 BestExitEdgeFreq = ExitEdgeFreq;
1916 if (!HasLoopingSucc) {
1917 // Restore the old exiting state, no viable looping successor was found.
1918 ExitingBB = OldExitingBB;
1919 BestExitEdgeFreq = OldBestExitEdgeFreq;
1922 // Without a candidate exiting block or with only a single block in the
1923 // loop, just use the loop header to layout the loop.
1925 DEBUG(dbgs() << " No other candidate exit blocks, using loop header\n");
1928 if (L.getNumBlocks() == 1) {
1929 DEBUG(dbgs() << " Loop has 1 block, using loop header as exit\n");
1933 // Also, if we have exit blocks which lead to outer loops but didn't select
1934 // one of them as the exiting block we are rotating toward, disable loop
1935 // rotation altogether.
1936 if (!BlocksExitingToOuterLoop.empty() &&
1937 !BlocksExitingToOuterLoop.count(ExitingBB))
1940 DEBUG(dbgs() << " Best exiting block: " << getBlockName(ExitingBB) << "\n");
1944 /// \brief Attempt to rotate an exiting block to the bottom of the loop.
1946 /// Once we have built a chain, try to rotate it to line up the hot exit block
1947 /// with fallthrough out of the loop if doing so doesn't introduce unnecessary
1948 /// branches. For example, if the loop has fallthrough into its header and out
1949 /// of its bottom already, don't rotate it.
1950 void MachineBlockPlacement::rotateLoop(BlockChain &LoopChain,
1951 const MachineBasicBlock *ExitingBB,
1952 const BlockFilterSet &LoopBlockSet) {
1956 MachineBasicBlock *Top = *LoopChain.begin();
1957 MachineBasicBlock *Bottom = *std::prev(LoopChain.end());
1959 // If ExitingBB is already the last one in a chain then nothing to do.
1960 if (Bottom == ExitingBB)
1963 bool ViableTopFallthrough = false;
1964 for (MachineBasicBlock *Pred : Top->predecessors()) {
1965 BlockChain *PredChain = BlockToChain[Pred];
1966 if (!LoopBlockSet.count(Pred) &&
1967 (!PredChain || Pred == *std::prev(PredChain->end()))) {
1968 ViableTopFallthrough = true;
1973 // If the header has viable fallthrough, check whether the current loop
1974 // bottom is a viable exiting block. If so, bail out as rotating will
1975 // introduce an unnecessary branch.
1976 if (ViableTopFallthrough) {
1977 for (MachineBasicBlock *Succ : Bottom->successors()) {
1978 BlockChain *SuccChain = BlockToChain[Succ];
1979 if (!LoopBlockSet.count(Succ) &&
1980 (!SuccChain || Succ == *SuccChain->begin()))
1985 BlockChain::iterator ExitIt = llvm::find(LoopChain, ExitingBB);
1986 if (ExitIt == LoopChain.end())
1989 // Rotating a loop exit to the bottom when there is a fallthrough to top
1990 // trades the entry fallthrough for an exit fallthrough.
1991 // If there is no bottom->top edge, but the chosen exit block does have
1992 // a fallthrough, we break that fallthrough for nothing in return.
1994 // Let's consider an example. We have a built chain of basic blocks
1995 // B1, B2, ..., Bn, where Bk is a ExitingBB - chosen exit block.
1996 // By doing a rotation we get
1997 // Bk+1, ..., Bn, B1, ..., Bk
1998 // Break of fallthrough to B1 is compensated by a fallthrough from Bk.
1999 // If we had a fallthrough Bk -> Bk+1 it is broken now.
2000 // It might be compensated by fallthrough Bn -> B1.
2001 // So we have a condition to avoid creation of extra branch by loop rotation.
2002 // All below must be true to avoid loop rotation:
2003 // If there is a fallthrough to top (B1)
2004 // There was fallthrough from chosen exit block (Bk) to next one (Bk+1)
2005 // There is no fallthrough from bottom (Bn) to top (B1).
2006 // Please note that there is no exit fallthrough from Bn because we checked it
2008 if (ViableTopFallthrough) {
2009 assert(std::next(ExitIt) != LoopChain.end() &&
2010 "Exit should not be last BB");
2011 MachineBasicBlock *NextBlockInChain = *std::next(ExitIt);
2012 if (ExitingBB->isSuccessor(NextBlockInChain))
2013 if (!Bottom->isSuccessor(Top))
2017 DEBUG(dbgs() << "Rotating loop to put exit " << getBlockName(ExitingBB)
2019 std::rotate(LoopChain.begin(), std::next(ExitIt), LoopChain.end());
2022 /// \brief Attempt to rotate a loop based on profile data to reduce branch cost.
2024 /// With profile data, we can determine the cost in terms of missed fall through
2025 /// opportunities when rotating a loop chain and select the best rotation.
2026 /// Basically, there are three kinds of cost to consider for each rotation:
2027 /// 1. The possibly missed fall through edge (if it exists) from BB out of
2028 /// the loop to the loop header.
2029 /// 2. The possibly missed fall through edges (if they exist) from the loop
2030 /// exits to BB out of the loop.
2031 /// 3. The missed fall through edge (if it exists) from the last BB to the
2032 /// first BB in the loop chain.
2033 /// Therefore, the cost for a given rotation is the sum of costs listed above.
2034 /// We select the best rotation with the smallest cost.
2035 void MachineBlockPlacement::rotateLoopWithProfile(
2036 BlockChain &LoopChain, const MachineLoop &L,
2037 const BlockFilterSet &LoopBlockSet) {
2038 auto HeaderBB = L.getHeader();
2039 auto HeaderIter = llvm::find(LoopChain, HeaderBB);
2040 auto RotationPos = LoopChain.end();
2042 BlockFrequency SmallestRotationCost = BlockFrequency::getMaxFrequency();
2044 // A utility lambda that scales up a block frequency by dividing it by a
2045 // branch probability which is the reciprocal of the scale.
2046 auto ScaleBlockFrequency = [](BlockFrequency Freq,
2047 unsigned Scale) -> BlockFrequency {
2050 // Use operator / between BlockFrequency and BranchProbability to implement
2051 // saturating multiplication.
2052 return Freq / BranchProbability(1, Scale);
2055 // Compute the cost of the missed fall-through edge to the loop header if the
2056 // chain head is not the loop header. As we only consider natural loops with
2057 // single header, this computation can be done only once.
2058 BlockFrequency HeaderFallThroughCost(0);
2059 for (auto *Pred : HeaderBB->predecessors()) {
2060 BlockChain *PredChain = BlockToChain[Pred];
2061 if (!LoopBlockSet.count(Pred) &&
2062 (!PredChain || Pred == *std::prev(PredChain->end()))) {
2064 MBFI->getBlockFreq(Pred) * MBPI->getEdgeProbability(Pred, HeaderBB);
2065 auto FallThruCost = ScaleBlockFrequency(EdgeFreq, MisfetchCost);
2066 // If the predecessor has only an unconditional jump to the header, we
2067 // need to consider the cost of this jump.
2068 if (Pred->succ_size() == 1)
2069 FallThruCost += ScaleBlockFrequency(EdgeFreq, JumpInstCost);
2070 HeaderFallThroughCost = std::max(HeaderFallThroughCost, FallThruCost);
2074 // Here we collect all exit blocks in the loop, and for each exit we find out
2075 // its hottest exit edge. For each loop rotation, we define the loop exit cost
2076 // as the sum of frequencies of exit edges we collect here, excluding the exit
2077 // edge from the tail of the loop chain.
2078 SmallVector<std::pair<MachineBasicBlock *, BlockFrequency>, 4> ExitsWithFreq;
2079 for (auto BB : LoopChain) {
2080 auto LargestExitEdgeProb = BranchProbability::getZero();
2081 for (auto *Succ : BB->successors()) {
2082 BlockChain *SuccChain = BlockToChain[Succ];
2083 if (!LoopBlockSet.count(Succ) &&
2084 (!SuccChain || Succ == *SuccChain->begin())) {
2085 auto SuccProb = MBPI->getEdgeProbability(BB, Succ);
2086 LargestExitEdgeProb = std::max(LargestExitEdgeProb, SuccProb);
2089 if (LargestExitEdgeProb > BranchProbability::getZero()) {
2090 auto ExitFreq = MBFI->getBlockFreq(BB) * LargestExitEdgeProb;
2091 ExitsWithFreq.emplace_back(BB, ExitFreq);
2095 // In this loop we iterate every block in the loop chain and calculate the
2096 // cost assuming the block is the head of the loop chain. When the loop ends,
2097 // we should have found the best candidate as the loop chain's head.
2098 for (auto Iter = LoopChain.begin(), TailIter = std::prev(LoopChain.end()),
2099 EndIter = LoopChain.end();
2100 Iter != EndIter; Iter++, TailIter++) {
2101 // TailIter is used to track the tail of the loop chain if the block we are
2102 // checking (pointed by Iter) is the head of the chain.
2103 if (TailIter == LoopChain.end())
2104 TailIter = LoopChain.begin();
2106 auto TailBB = *TailIter;
2108 // Calculate the cost by putting this BB to the top.
2109 BlockFrequency Cost = 0;
2111 // If the current BB is the loop header, we need to take into account the
2112 // cost of the missed fall through edge from outside of the loop to the
2114 if (Iter != HeaderIter)
2115 Cost += HeaderFallThroughCost;
2117 // Collect the loop exit cost by summing up frequencies of all exit edges
2118 // except the one from the chain tail.
2119 for (auto &ExitWithFreq : ExitsWithFreq)
2120 if (TailBB != ExitWithFreq.first)
2121 Cost += ExitWithFreq.second;
2123 // The cost of breaking the once fall-through edge from the tail to the top
2124 // of the loop chain. Here we need to consider three cases:
2125 // 1. If the tail node has only one successor, then we will get an
2126 // additional jmp instruction. So the cost here is (MisfetchCost +
2127 // JumpInstCost) * tail node frequency.
2128 // 2. If the tail node has two successors, then we may still get an
2129 // additional jmp instruction if the layout successor after the loop
2130 // chain is not its CFG successor. Note that the more frequently executed
2131 // jmp instruction will be put ahead of the other one. Assume the
2132 // frequency of those two branches are x and y, where x is the frequency
2133 // of the edge to the chain head, then the cost will be
2134 // (x * MisfetechCost + min(x, y) * JumpInstCost) * tail node frequency.
2135 // 3. If the tail node has more than two successors (this rarely happens),
2136 // we won't consider any additional cost.
2137 if (TailBB->isSuccessor(*Iter)) {
2138 auto TailBBFreq = MBFI->getBlockFreq(TailBB);
2139 if (TailBB->succ_size() == 1)
2140 Cost += ScaleBlockFrequency(TailBBFreq.getFrequency(),
2141 MisfetchCost + JumpInstCost);
2142 else if (TailBB->succ_size() == 2) {
2143 auto TailToHeadProb = MBPI->getEdgeProbability(TailBB, *Iter);
2144 auto TailToHeadFreq = TailBBFreq * TailToHeadProb;
2145 auto ColderEdgeFreq = TailToHeadProb > BranchProbability(1, 2)
2146 ? TailBBFreq * TailToHeadProb.getCompl()
2148 Cost += ScaleBlockFrequency(TailToHeadFreq, MisfetchCost) +
2149 ScaleBlockFrequency(ColderEdgeFreq, JumpInstCost);
2153 DEBUG(dbgs() << "The cost of loop rotation by making " << getBlockName(*Iter)
2154 << " to the top: " << Cost.getFrequency() << "\n");
2156 if (Cost < SmallestRotationCost) {
2157 SmallestRotationCost = Cost;
2162 if (RotationPos != LoopChain.end()) {
2163 DEBUG(dbgs() << "Rotate loop by making " << getBlockName(*RotationPos)
2164 << " to the top\n");
2165 std::rotate(LoopChain.begin(), RotationPos, LoopChain.end());
2169 /// \brief Collect blocks in the given loop that are to be placed.
2171 /// When profile data is available, exclude cold blocks from the returned set;
2172 /// otherwise, collect all blocks in the loop.
2173 MachineBlockPlacement::BlockFilterSet
2174 MachineBlockPlacement::collectLoopBlockSet(const MachineLoop &L) {
2175 BlockFilterSet LoopBlockSet;
2177 // Filter cold blocks off from LoopBlockSet when profile data is available.
2178 // Collect the sum of frequencies of incoming edges to the loop header from
2179 // outside. If we treat the loop as a super block, this is the frequency of
2180 // the loop. Then for each block in the loop, we calculate the ratio between
2181 // its frequency and the frequency of the loop block. When it is too small,
2182 // don't add it to the loop chain. If there are outer loops, then this block
2183 // will be merged into the first outer loop chain for which this block is not
2184 // cold anymore. This needs precise profile data and we only do this when
2185 // profile data is available.
2186 if (F->getFunction().hasProfileData() || ForceLoopColdBlock) {
2187 BlockFrequency LoopFreq(0);
2188 for (auto LoopPred : L.getHeader()->predecessors())
2189 if (!L.contains(LoopPred))
2190 LoopFreq += MBFI->getBlockFreq(LoopPred) *
2191 MBPI->getEdgeProbability(LoopPred, L.getHeader());
2193 for (MachineBasicBlock *LoopBB : L.getBlocks()) {
2194 auto Freq = MBFI->getBlockFreq(LoopBB).getFrequency();
2195 if (Freq == 0 || LoopFreq.getFrequency() / Freq > LoopToColdBlockRatio)
2197 LoopBlockSet.insert(LoopBB);
2200 LoopBlockSet.insert(L.block_begin(), L.block_end());
2202 return LoopBlockSet;
2205 /// \brief Forms basic block chains from the natural loop structures.
2207 /// These chains are designed to preserve the existing *structure* of the code
2208 /// as much as possible. We can then stitch the chains together in a way which
2209 /// both preserves the topological structure and minimizes taken conditional
2211 void MachineBlockPlacement::buildLoopChains(const MachineLoop &L) {
2212 // First recurse through any nested loops, building chains for those inner
2214 for (const MachineLoop *InnerLoop : L)
2215 buildLoopChains(*InnerLoop);
2217 assert(BlockWorkList.empty() &&
2218 "BlockWorkList not empty when starting to build loop chains.");
2219 assert(EHPadWorkList.empty() &&
2220 "EHPadWorkList not empty when starting to build loop chains.");
2221 BlockFilterSet LoopBlockSet = collectLoopBlockSet(L);
2223 // Check if we have profile data for this function. If yes, we will rotate
2224 // this loop by modeling costs more precisely which requires the profile data
2225 // for better layout.
2226 bool RotateLoopWithProfile =
2227 ForcePreciseRotationCost ||
2228 (PreciseRotationCost && F->getFunction().hasProfileData());
2230 // First check to see if there is an obviously preferable top block for the
2231 // loop. This will default to the header, but may end up as one of the
2232 // predecessors to the header if there is one which will result in strictly
2233 // fewer branches in the loop body.
2234 // When we use profile data to rotate the loop, this is unnecessary.
2235 MachineBasicBlock *LoopTop =
2236 RotateLoopWithProfile ? L.getHeader() : findBestLoopTop(L, LoopBlockSet);
2238 // If we selected just the header for the loop top, look for a potentially
2239 // profitable exit block in the event that rotating the loop can eliminate
2240 // branches by placing an exit edge at the bottom.
2242 // Loops are processed innermost to uttermost, make sure we clear
2243 // PreferredLoopExit before processing a new loop.
2244 PreferredLoopExit = nullptr;
2245 if (!RotateLoopWithProfile && LoopTop == L.getHeader())
2246 PreferredLoopExit = findBestLoopExit(L, LoopBlockSet);
2248 BlockChain &LoopChain = *BlockToChain[LoopTop];
2250 // FIXME: This is a really lame way of walking the chains in the loop: we
2251 // walk the blocks, and use a set to prevent visiting a particular chain
2253 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
2254 assert(LoopChain.UnscheduledPredecessors == 0 &&
2255 "LoopChain should not have unscheduled predecessors.");
2256 UpdatedPreds.insert(&LoopChain);
2258 for (const MachineBasicBlock *LoopBB : LoopBlockSet)
2259 fillWorkLists(LoopBB, UpdatedPreds, &LoopBlockSet);
2261 buildChain(LoopTop, LoopChain, &LoopBlockSet);
2263 if (RotateLoopWithProfile)
2264 rotateLoopWithProfile(LoopChain, L, LoopBlockSet);
2266 rotateLoop(LoopChain, PreferredLoopExit, LoopBlockSet);
2269 // Crash at the end so we get all of the debugging output first.
2270 bool BadLoop = false;
2271 if (LoopChain.UnscheduledPredecessors) {
2273 dbgs() << "Loop chain contains a block without its preds placed!\n"
2274 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
2275 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n";
2277 for (MachineBasicBlock *ChainBB : LoopChain) {
2278 dbgs() << " ... " << getBlockName(ChainBB) << "\n";
2279 if (!LoopBlockSet.remove(ChainBB)) {
2280 // We don't mark the loop as bad here because there are real situations
2281 // where this can occur. For example, with an unanalyzable fallthrough
2282 // from a loop block to a non-loop block or vice versa.
2283 dbgs() << "Loop chain contains a block not contained by the loop!\n"
2284 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
2285 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
2286 << " Bad block: " << getBlockName(ChainBB) << "\n";
2290 if (!LoopBlockSet.empty()) {
2292 for (const MachineBasicBlock *LoopBB : LoopBlockSet)
2293 dbgs() << "Loop contains blocks never placed into a chain!\n"
2294 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
2295 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
2296 << " Bad block: " << getBlockName(LoopBB) << "\n";
2298 assert(!BadLoop && "Detected problems with the placement of this loop.");
2301 BlockWorkList.clear();
2302 EHPadWorkList.clear();
2305 void MachineBlockPlacement::buildCFGChains() {
2306 // Ensure that every BB in the function has an associated chain to simplify
2307 // the assumptions of the remaining algorithm.
2308 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
2309 for (MachineFunction::iterator FI = F->begin(), FE = F->end(); FI != FE;
2311 MachineBasicBlock *BB = &*FI;
2313 new (ChainAllocator.Allocate()) BlockChain(BlockToChain, BB);
2314 // Also, merge any blocks which we cannot reason about and must preserve
2315 // the exact fallthrough behavior for.
2318 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
2319 if (!TII->analyzeBranch(*BB, TBB, FBB, Cond) || !FI->canFallThrough())
2322 MachineFunction::iterator NextFI = std::next(FI);
2323 MachineBasicBlock *NextBB = &*NextFI;
2324 // Ensure that the layout successor is a viable block, as we know that
2325 // fallthrough is a possibility.
2326 assert(NextFI != FE && "Can't fallthrough past the last block.");
2327 DEBUG(dbgs() << "Pre-merging due to unanalyzable fallthrough: "
2328 << getBlockName(BB) << " -> " << getBlockName(NextBB)
2330 Chain->merge(NextBB, nullptr);
2332 BlocksWithUnanalyzableExits.insert(&*BB);
2339 // Build any loop-based chains.
2340 PreferredLoopExit = nullptr;
2341 for (MachineLoop *L : *MLI)
2342 buildLoopChains(*L);
2344 assert(BlockWorkList.empty() &&
2345 "BlockWorkList should be empty before building final chain.");
2346 assert(EHPadWorkList.empty() &&
2347 "EHPadWorkList should be empty before building final chain.");
2349 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
2350 for (MachineBasicBlock &MBB : *F)
2351 fillWorkLists(&MBB, UpdatedPreds);
2353 BlockChain &FunctionChain = *BlockToChain[&F->front()];
2354 buildChain(&F->front(), FunctionChain);
2357 using FunctionBlockSetType = SmallPtrSet<MachineBasicBlock *, 16>;
2360 // Crash at the end so we get all of the debugging output first.
2361 bool BadFunc = false;
2362 FunctionBlockSetType FunctionBlockSet;
2363 for (MachineBasicBlock &MBB : *F)
2364 FunctionBlockSet.insert(&MBB);
2366 for (MachineBasicBlock *ChainBB : FunctionChain)
2367 if (!FunctionBlockSet.erase(ChainBB)) {
2369 dbgs() << "Function chain contains a block not in the function!\n"
2370 << " Bad block: " << getBlockName(ChainBB) << "\n";
2373 if (!FunctionBlockSet.empty()) {
2375 for (MachineBasicBlock *RemainingBB : FunctionBlockSet)
2376 dbgs() << "Function contains blocks never placed into a chain!\n"
2377 << " Bad block: " << getBlockName(RemainingBB) << "\n";
2379 assert(!BadFunc && "Detected problems with the block placement.");
2382 // Splice the blocks into place.
2383 MachineFunction::iterator InsertPos = F->begin();
2384 DEBUG(dbgs() << "[MBP] Function: "<< F->getName() << "\n");
2385 for (MachineBasicBlock *ChainBB : FunctionChain) {
2386 DEBUG(dbgs() << (ChainBB == *FunctionChain.begin() ? "Placing chain "
2388 << getBlockName(ChainBB) << "\n");
2389 if (InsertPos != MachineFunction::iterator(ChainBB))
2390 F->splice(InsertPos, ChainBB);
2394 // Update the terminator of the previous block.
2395 if (ChainBB == *FunctionChain.begin())
2397 MachineBasicBlock *PrevBB = &*std::prev(MachineFunction::iterator(ChainBB));
2399 // FIXME: It would be awesome of updateTerminator would just return rather
2400 // than assert when the branch cannot be analyzed in order to remove this
2403 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
2406 if (!BlocksWithUnanalyzableExits.count(PrevBB)) {
2407 // Given the exact block placement we chose, we may actually not _need_ to
2408 // be able to edit PrevBB's terminator sequence, but not being _able_ to
2409 // do that at this point is a bug.
2410 assert((!TII->analyzeBranch(*PrevBB, TBB, FBB, Cond) ||
2411 !PrevBB->canFallThrough()) &&
2412 "Unexpected block with un-analyzable fallthrough!");
2414 TBB = FBB = nullptr;
2418 // The "PrevBB" is not yet updated to reflect current code layout, so,
2419 // o. it may fall-through to a block without explicit "goto" instruction
2420 // before layout, and no longer fall-through it after layout; or
2421 // o. just opposite.
2423 // analyzeBranch() may return erroneous value for FBB when these two
2424 // situations take place. For the first scenario FBB is mistakenly set NULL;
2425 // for the 2nd scenario, the FBB, which is expected to be NULL, is
2426 // mistakenly pointing to "*BI".
2427 // Thus, if the future change needs to use FBB before the layout is set, it
2428 // has to correct FBB first by using the code similar to the following:
2430 // if (!Cond.empty() && (!FBB || FBB == ChainBB)) {
2431 // PrevBB->updateTerminator();
2433 // TBB = FBB = nullptr;
2434 // if (TII->analyzeBranch(*PrevBB, TBB, FBB, Cond)) {
2435 // // FIXME: This should never take place.
2436 // TBB = FBB = nullptr;
2439 if (!TII->analyzeBranch(*PrevBB, TBB, FBB, Cond))
2440 PrevBB->updateTerminator();
2443 // Fixup the last block.
2445 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
2446 if (!TII->analyzeBranch(F->back(), TBB, FBB, Cond))
2447 F->back().updateTerminator();
2449 BlockWorkList.clear();
2450 EHPadWorkList.clear();
2453 void MachineBlockPlacement::optimizeBranches() {
2454 BlockChain &FunctionChain = *BlockToChain[&F->front()];
2455 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
2457 // Now that all the basic blocks in the chain have the proper layout,
2458 // make a final call to AnalyzeBranch with AllowModify set.
2459 // Indeed, the target may be able to optimize the branches in a way we
2460 // cannot because all branches may not be analyzable.
2461 // E.g., the target may be able to remove an unconditional branch to
2462 // a fallthrough when it occurs after predicated terminators.
2463 for (MachineBasicBlock *ChainBB : FunctionChain) {
2465 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
2466 if (!TII->analyzeBranch(*ChainBB, TBB, FBB, Cond, /*AllowModify*/ true)) {
2467 // If PrevBB has a two-way branch, try to re-order the branches
2468 // such that we branch to the successor with higher probability first.
2469 if (TBB && !Cond.empty() && FBB &&
2470 MBPI->getEdgeProbability(ChainBB, FBB) >
2471 MBPI->getEdgeProbability(ChainBB, TBB) &&
2472 !TII->reverseBranchCondition(Cond)) {
2473 DEBUG(dbgs() << "Reverse order of the two branches: "
2474 << getBlockName(ChainBB) << "\n");
2475 DEBUG(dbgs() << " Edge probability: "
2476 << MBPI->getEdgeProbability(ChainBB, FBB) << " vs "
2477 << MBPI->getEdgeProbability(ChainBB, TBB) << "\n");
2478 DebugLoc dl; // FIXME: this is nowhere
2479 TII->removeBranch(*ChainBB);
2480 TII->insertBranch(*ChainBB, FBB, TBB, Cond, dl);
2481 ChainBB->updateTerminator();
2487 void MachineBlockPlacement::alignBlocks() {
2488 // Walk through the backedges of the function now that we have fully laid out
2489 // the basic blocks and align the destination of each backedge. We don't rely
2490 // exclusively on the loop info here so that we can align backedges in
2491 // unnatural CFGs and backedges that were introduced purely because of the
2492 // loop rotations done during this layout pass.
2493 if (F->getFunction().optForSize())
2495 BlockChain &FunctionChain = *BlockToChain[&F->front()];
2496 if (FunctionChain.begin() == FunctionChain.end())
2497 return; // Empty chain.
2499 const BranchProbability ColdProb(1, 5); // 20%
2500 BlockFrequency EntryFreq = MBFI->getBlockFreq(&F->front());
2501 BlockFrequency WeightedEntryFreq = EntryFreq * ColdProb;
2502 for (MachineBasicBlock *ChainBB : FunctionChain) {
2503 if (ChainBB == *FunctionChain.begin())
2506 // Don't align non-looping basic blocks. These are unlikely to execute
2507 // enough times to matter in practice. Note that we'll still handle
2508 // unnatural CFGs inside of a natural outer loop (the common case) and
2510 MachineLoop *L = MLI->getLoopFor(ChainBB);
2514 unsigned Align = TLI->getPrefLoopAlignment(L);
2516 continue; // Don't care about loop alignment.
2518 // If the block is cold relative to the function entry don't waste space
2520 BlockFrequency Freq = MBFI->getBlockFreq(ChainBB);
2521 if (Freq < WeightedEntryFreq)
2524 // If the block is cold relative to its loop header, don't align it
2525 // regardless of what edges into the block exist.
2526 MachineBasicBlock *LoopHeader = L->getHeader();
2527 BlockFrequency LoopHeaderFreq = MBFI->getBlockFreq(LoopHeader);
2528 if (Freq < (LoopHeaderFreq * ColdProb))
2531 // Check for the existence of a non-layout predecessor which would benefit
2532 // from aligning this block.
2533 MachineBasicBlock *LayoutPred =
2534 &*std::prev(MachineFunction::iterator(ChainBB));
2536 // Force alignment if all the predecessors are jumps. We already checked
2537 // that the block isn't cold above.
2538 if (!LayoutPred->isSuccessor(ChainBB)) {
2539 ChainBB->setAlignment(Align);
2543 // Align this block if the layout predecessor's edge into this block is
2544 // cold relative to the block. When this is true, other predecessors make up
2545 // all of the hot entries into the block and thus alignment is likely to be
2547 BranchProbability LayoutProb =
2548 MBPI->getEdgeProbability(LayoutPred, ChainBB);
2549 BlockFrequency LayoutEdgeFreq = MBFI->getBlockFreq(LayoutPred) * LayoutProb;
2550 if (LayoutEdgeFreq <= (Freq * ColdProb))
2551 ChainBB->setAlignment(Align);
2555 /// Tail duplicate \p BB into (some) predecessors if profitable, repeating if
2556 /// it was duplicated into its chain predecessor and removed.
2557 /// \p BB - Basic block that may be duplicated.
2559 /// \p LPred - Chosen layout predecessor of \p BB.
2560 /// Updated to be the chain end if LPred is removed.
2561 /// \p Chain - Chain to which \p LPred belongs, and \p BB will belong.
2562 /// \p BlockFilter - Set of blocks that belong to the loop being laid out.
2563 /// Used to identify which blocks to update predecessor
2565 /// \p PrevUnplacedBlockIt - Iterator pointing to the last block that was
2566 /// chosen in the given order due to unnatural CFG
2567 /// only needed if \p BB is removed and
2568 /// \p PrevUnplacedBlockIt pointed to \p BB.
2569 /// @return true if \p BB was removed.
2570 bool MachineBlockPlacement::repeatedlyTailDuplicateBlock(
2571 MachineBasicBlock *BB, MachineBasicBlock *&LPred,
2572 const MachineBasicBlock *LoopHeaderBB,
2573 BlockChain &Chain, BlockFilterSet *BlockFilter,
2574 MachineFunction::iterator &PrevUnplacedBlockIt) {
2575 bool Removed, DuplicatedToLPred;
2576 bool DuplicatedToOriginalLPred;
2577 Removed = maybeTailDuplicateBlock(BB, LPred, Chain, BlockFilter,
2578 PrevUnplacedBlockIt,
2582 DuplicatedToOriginalLPred = DuplicatedToLPred;
2583 // Iteratively try to duplicate again. It can happen that a block that is
2584 // duplicated into is still small enough to be duplicated again.
2585 // No need to call markBlockSuccessors in this case, as the blocks being
2586 // duplicated from here on are already scheduled.
2587 // Note that DuplicatedToLPred always implies Removed.
2588 while (DuplicatedToLPred) {
2589 assert(Removed && "Block must have been removed to be duplicated into its "
2590 "layout predecessor.");
2591 MachineBasicBlock *DupBB, *DupPred;
2592 // The removal callback causes Chain.end() to be updated when a block is
2593 // removed. On the first pass through the loop, the chain end should be the
2594 // same as it was on function entry. On subsequent passes, because we are
2595 // duplicating the block at the end of the chain, if it is removed the
2596 // chain will have shrunk by one block.
2597 BlockChain::iterator ChainEnd = Chain.end();
2598 DupBB = *(--ChainEnd);
2599 // Now try to duplicate again.
2600 if (ChainEnd == Chain.begin())
2602 DupPred = *std::prev(ChainEnd);
2603 Removed = maybeTailDuplicateBlock(DupBB, DupPred, Chain, BlockFilter,
2604 PrevUnplacedBlockIt,
2607 // If BB was duplicated into LPred, it is now scheduled. But because it was
2608 // removed, markChainSuccessors won't be called for its chain. Instead we
2609 // call markBlockSuccessors for LPred to achieve the same effect. This must go
2610 // at the end because repeating the tail duplication can increase the number
2611 // of unscheduled predecessors.
2612 LPred = *std::prev(Chain.end());
2613 if (DuplicatedToOriginalLPred)
2614 markBlockSuccessors(Chain, LPred, LoopHeaderBB, BlockFilter);
2618 /// Tail duplicate \p BB into (some) predecessors if profitable.
2619 /// \p BB - Basic block that may be duplicated
2620 /// \p LPred - Chosen layout predecessor of \p BB
2621 /// \p Chain - Chain to which \p LPred belongs, and \p BB will belong.
2622 /// \p BlockFilter - Set of blocks that belong to the loop being laid out.
2623 /// Used to identify which blocks to update predecessor
2625 /// \p PrevUnplacedBlockIt - Iterator pointing to the last block that was
2626 /// chosen in the given order due to unnatural CFG
2627 /// only needed if \p BB is removed and
2628 /// \p PrevUnplacedBlockIt pointed to \p BB.
2629 /// \p DuplicatedToLPred - True if the block was duplicated into LPred. Will
2630 /// only be true if the block was removed.
2631 /// \return - True if the block was duplicated into all preds and removed.
2632 bool MachineBlockPlacement::maybeTailDuplicateBlock(
2633 MachineBasicBlock *BB, MachineBasicBlock *LPred,
2634 BlockChain &Chain, BlockFilterSet *BlockFilter,
2635 MachineFunction::iterator &PrevUnplacedBlockIt,
2636 bool &DuplicatedToLPred) {
2637 DuplicatedToLPred = false;
2638 if (!shouldTailDuplicate(BB))
2641 DEBUG(dbgs() << "Redoing tail duplication for Succ#"
2642 << BB->getNumber() << "\n");
2644 // This has to be a callback because none of it can be done after
2646 bool Removed = false;
2647 auto RemovalCallback =
2648 [&](MachineBasicBlock *RemBB) {
2649 // Signal to outer function
2652 // Conservative default.
2653 bool InWorkList = true;
2654 // Remove from the Chain and Chain Map
2655 if (BlockToChain.count(RemBB)) {
2656 BlockChain *Chain = BlockToChain[RemBB];
2657 InWorkList = Chain->UnscheduledPredecessors == 0;
2658 Chain->remove(RemBB);
2659 BlockToChain.erase(RemBB);
2662 // Handle the unplaced block iterator
2663 if (&(*PrevUnplacedBlockIt) == RemBB) {
2664 PrevUnplacedBlockIt++;
2667 // Handle the Work Lists
2669 SmallVectorImpl<MachineBasicBlock *> &RemoveList = BlockWorkList;
2670 if (RemBB->isEHPad())
2671 RemoveList = EHPadWorkList;
2673 llvm::remove_if(RemoveList,
2674 [RemBB](MachineBasicBlock *BB) {
2680 // Handle the filter set
2682 BlockFilter->remove(RemBB);
2685 // Remove the block from loop info.
2686 MLI->removeBlock(RemBB);
2687 if (RemBB == PreferredLoopExit)
2688 PreferredLoopExit = nullptr;
2690 DEBUG(dbgs() << "TailDuplicator deleted block: "
2691 << getBlockName(RemBB) << "\n");
2693 auto RemovalCallbackRef =
2694 function_ref<void(MachineBasicBlock*)>(RemovalCallback);
2696 SmallVector<MachineBasicBlock *, 8> DuplicatedPreds;
2697 bool IsSimple = TailDup.isSimpleBB(BB);
2698 TailDup.tailDuplicateAndUpdate(IsSimple, BB, LPred,
2699 &DuplicatedPreds, &RemovalCallbackRef);
2701 // Update UnscheduledPredecessors to reflect tail-duplication.
2702 DuplicatedToLPred = false;
2703 for (MachineBasicBlock *Pred : DuplicatedPreds) {
2704 // We're only looking for unscheduled predecessors that match the filter.
2705 BlockChain* PredChain = BlockToChain[Pred];
2707 DuplicatedToLPred = true;
2708 if (Pred == LPred || (BlockFilter && !BlockFilter->count(Pred))
2709 || PredChain == &Chain)
2711 for (MachineBasicBlock *NewSucc : Pred->successors()) {
2712 if (BlockFilter && !BlockFilter->count(NewSucc))
2714 BlockChain *NewChain = BlockToChain[NewSucc];
2715 if (NewChain != &Chain && NewChain != PredChain)
2716 NewChain->UnscheduledPredecessors++;
2722 bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &MF) {
2723 if (skipFunction(MF.getFunction()))
2726 // Check for single-block functions and skip them.
2727 if (std::next(MF.begin()) == MF.end())
2731 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
2732 MBFI = llvm::make_unique<BranchFolder::MBFIWrapper>(
2733 getAnalysis<MachineBlockFrequencyInfo>());
2734 MLI = &getAnalysis<MachineLoopInfo>();
2735 TII = MF.getSubtarget().getInstrInfo();
2736 TLI = MF.getSubtarget().getTargetLowering();
2739 // Initialize PreferredLoopExit to nullptr here since it may never be set if
2740 // there are no MachineLoops.
2741 PreferredLoopExit = nullptr;
2743 assert(BlockToChain.empty() &&
2744 "BlockToChain map should be empty before starting placement.");
2745 assert(ComputedEdges.empty() &&
2746 "Computed Edge map should be empty before starting placement.");
2748 unsigned TailDupSize = TailDupPlacementThreshold;
2749 // If only the aggressive threshold is explicitly set, use it.
2750 if (TailDupPlacementAggressiveThreshold.getNumOccurrences() != 0 &&
2751 TailDupPlacementThreshold.getNumOccurrences() == 0)
2752 TailDupSize = TailDupPlacementAggressiveThreshold;
2754 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
2755 // For agressive optimization, we can adjust some thresholds to be less
2757 if (PassConfig->getOptLevel() >= CodeGenOpt::Aggressive) {
2758 // At O3 we should be more willing to copy blocks for tail duplication. This
2759 // increases size pressure, so we only do it at O3
2760 // Do this unless only the regular threshold is explicitly set.
2761 if (TailDupPlacementThreshold.getNumOccurrences() == 0 ||
2762 TailDupPlacementAggressiveThreshold.getNumOccurrences() != 0)
2763 TailDupSize = TailDupPlacementAggressiveThreshold;
2766 if (allowTailDupPlacement()) {
2767 MPDT = &getAnalysis<MachinePostDominatorTree>();
2768 if (MF.getFunction().optForSize())
2770 bool PreRegAlloc = false;
2771 TailDup.initMF(MF, PreRegAlloc, MBPI, /* LayoutMode */ true, TailDupSize);
2772 precomputeTriangleChains();
2777 // Changing the layout can create new tail merging opportunities.
2778 // TailMerge can create jump into if branches that make CFG irreducible for
2779 // HW that requires structured CFG.
2780 bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
2781 PassConfig->getEnableTailMerge() &&
2782 BranchFoldPlacement;
2783 // No tail merging opportunities if the block number is less than four.
2784 if (MF.size() > 3 && EnableTailMerge) {
2785 unsigned TailMergeSize = TailDupSize + 1;
2786 BranchFolder BF(/*EnableTailMerge=*/true, /*CommonHoist=*/false, *MBFI,
2787 *MBPI, TailMergeSize);
2789 if (BF.OptimizeFunction(MF, TII, MF.getSubtarget().getRegisterInfo(),
2790 getAnalysisIfAvailable<MachineModuleInfo>(), MLI,
2791 /*AfterBlockPlacement=*/true)) {
2792 // Redo the layout if tail merging creates/removes/moves blocks.
2793 BlockToChain.clear();
2794 ComputedEdges.clear();
2795 // Must redo the post-dominator tree if blocks were changed.
2797 MPDT->runOnMachineFunction(MF);
2798 ChainAllocator.DestroyAll();
2806 BlockToChain.clear();
2807 ComputedEdges.clear();
2808 ChainAllocator.DestroyAll();
2811 // Align all of the blocks in the function to a specific alignment.
2812 for (MachineBasicBlock &MBB : MF)
2813 MBB.setAlignment(AlignAllBlock);
2814 else if (AlignAllNonFallThruBlocks) {
2815 // Align all of the blocks that have no fall-through predecessors to a
2816 // specific alignment.
2817 for (auto MBI = std::next(MF.begin()), MBE = MF.end(); MBI != MBE; ++MBI) {
2818 auto LayoutPred = std::prev(MBI);
2819 if (!LayoutPred->isSuccessor(&*MBI))
2820 MBI->setAlignment(AlignAllNonFallThruBlocks);
2823 if (ViewBlockLayoutWithBFI != GVDT_None &&
2824 (ViewBlockFreqFuncName.empty() ||
2825 F->getFunction().getName().equals(ViewBlockFreqFuncName))) {
2826 MBFI->view("MBP." + MF.getName(), false);
2830 // We always return true as we have no way to track whether the final order
2831 // differs from the original order.
2837 /// \brief A pass to compute block placement statistics.
2839 /// A separate pass to compute interesting statistics for evaluating block
2840 /// placement. This is separate from the actual placement pass so that they can
2841 /// be computed in the absence of any placement transformations or when using
2842 /// alternative placement strategies.
2843 class MachineBlockPlacementStats : public MachineFunctionPass {
2844 /// \brief A handle to the branch probability pass.
2845 const MachineBranchProbabilityInfo *MBPI;
2847 /// \brief A handle to the function-wide block frequency pass.
2848 const MachineBlockFrequencyInfo *MBFI;
2851 static char ID; // Pass identification, replacement for typeid
2853 MachineBlockPlacementStats() : MachineFunctionPass(ID) {
2854 initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry());
2857 bool runOnMachineFunction(MachineFunction &F) override;
2859 void getAnalysisUsage(AnalysisUsage &AU) const override {
2860 AU.addRequired<MachineBranchProbabilityInfo>();
2861 AU.addRequired<MachineBlockFrequencyInfo>();
2862 AU.setPreservesAll();
2863 MachineFunctionPass::getAnalysisUsage(AU);
2867 } // end anonymous namespace
2869 char MachineBlockPlacementStats::ID = 0;
2871 char &llvm::MachineBlockPlacementStatsID = MachineBlockPlacementStats::ID;
2873 INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats",
2874 "Basic Block Placement Stats", false, false)
2875 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
2876 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
2877 INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats",
2878 "Basic Block Placement Stats", false, false)
2880 bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) {
2881 // Check for single-block functions and skip them.
2882 if (std::next(F.begin()) == F.end())
2885 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
2886 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
2888 for (MachineBasicBlock &MBB : F) {
2889 BlockFrequency BlockFreq = MBFI->getBlockFreq(&MBB);
2890 Statistic &NumBranches =
2891 (MBB.succ_size() > 1) ? NumCondBranches : NumUncondBranches;
2892 Statistic &BranchTakenFreq =
2893 (MBB.succ_size() > 1) ? CondBranchTakenFreq : UncondBranchTakenFreq;
2894 for (MachineBasicBlock *Succ : MBB.successors()) {
2895 // Skip if this successor is a fallthrough.
2896 if (MBB.isLayoutSuccessor(Succ))
2899 BlockFrequency EdgeFreq =
2900 BlockFreq * MBPI->getEdgeProbability(&MBB, Succ);
2902 BranchTakenFreq += EdgeFreq.getFrequency();