1 //===-- MachineBlockPlacement.cpp - Basic Block Code Layout optimization --===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements basic block placement transformations using the CFG
11 // structure and branch probability estimates.
13 // The pass strives to preserve the structure of the CFG (that is, retain
14 // a topological ordering of basic blocks) in the absence of a *strong* signal
15 // to the contrary from probabilities. However, within the CFG structure, it
16 // attempts to choose an ordering which favors placing more likely sequences of
17 // blocks adjacent to each other.
19 // The algorithm works from the inner-most loop within a function outward, and
20 // at each stage walks through the basic blocks, trying to coalesce them into
21 // sequential chains where allowed by the CFG (or demanded by heavy
22 // probabilities). Finally, it walks the blocks in topological order, and the
23 // first time it reaches a chain of basic blocks, it schedules them in the
26 //===----------------------------------------------------------------------===//
28 #include "llvm/CodeGen/Passes.h"
29 #include "llvm/ADT/DenseMap.h"
30 #include "llvm/ADT/SmallPtrSet.h"
31 #include "llvm/ADT/SmallVector.h"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/CodeGen/MachineBasicBlock.h"
34 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
35 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
36 #include "llvm/CodeGen/MachineDominators.h"
37 #include "llvm/CodeGen/MachineFunction.h"
38 #include "llvm/CodeGen/MachineFunctionPass.h"
39 #include "llvm/CodeGen/MachineLoopInfo.h"
40 #include "llvm/CodeGen/MachineModuleInfo.h"
41 #include "llvm/Support/Allocator.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/Support/raw_ostream.h"
45 #include "llvm/Target/TargetInstrInfo.h"
46 #include "llvm/Target/TargetLowering.h"
47 #include "llvm/Target/TargetSubtargetInfo.h"
51 #define DEBUG_TYPE "block-placement"
53 STATISTIC(NumCondBranches, "Number of conditional branches");
54 STATISTIC(NumUncondBranches, "Number of uncondittional branches");
55 STATISTIC(CondBranchTakenFreq,
56 "Potential frequency of taking conditional branches");
57 STATISTIC(UncondBranchTakenFreq,
58 "Potential frequency of taking unconditional branches");
60 static cl::opt<unsigned> AlignAllBlock("align-all-blocks",
61 cl::desc("Force the alignment of all "
62 "blocks in the function."),
63 cl::init(0), cl::Hidden);
65 // FIXME: Find a good default for this flag and remove the flag.
66 static cl::opt<unsigned> ExitBlockBias(
67 "block-placement-exit-block-bias",
68 cl::desc("Block frequency percentage a loop exit block needs "
69 "over the original exit to be considered the new exit."),
70 cl::init(0), cl::Hidden);
72 static cl::opt<bool> OutlineOptionalBranches(
73 "outline-optional-branches",
74 cl::desc("Put completely optional branches, i.e. branches with a common "
75 "post dominator, out of line."),
76 cl::init(false), cl::Hidden);
78 static cl::opt<unsigned> OutlineOptionalThreshold(
79 "outline-optional-threshold",
80 cl::desc("Don't outline optional branches that are a single block with an "
81 "instruction count below this threshold"),
82 cl::init(4), cl::Hidden);
86 /// \brief Type for our function-wide basic block -> block chain mapping.
87 typedef DenseMap<MachineBasicBlock *, BlockChain *> BlockToChainMapType;
91 /// \brief A chain of blocks which will be laid out contiguously.
93 /// This is the datastructure representing a chain of consecutive blocks that
94 /// are profitable to layout together in order to maximize fallthrough
95 /// probabilities and code locality. We also can use a block chain to represent
96 /// a sequence of basic blocks which have some external (correctness)
97 /// requirement for sequential layout.
99 /// Chains can be built around a single basic block and can be merged to grow
100 /// them. They participate in a block-to-chain mapping, which is updated
101 /// automatically as chains are merged together.
103 /// \brief The sequence of blocks belonging to this chain.
105 /// This is the sequence of blocks for a particular chain. These will be laid
106 /// out in-order within the function.
107 SmallVector<MachineBasicBlock *, 4> Blocks;
109 /// \brief A handle to the function-wide basic block to block chain mapping.
111 /// This is retained in each block chain to simplify the computation of child
112 /// block chains for SCC-formation and iteration. We store the edges to child
113 /// basic blocks, and map them back to their associated chains using this
115 BlockToChainMapType &BlockToChain;
118 /// \brief Construct a new BlockChain.
120 /// This builds a new block chain representing a single basic block in the
121 /// function. It also registers itself as the chain that block participates
122 /// in with the BlockToChain mapping.
123 BlockChain(BlockToChainMapType &BlockToChain, MachineBasicBlock *BB)
124 : Blocks(1, BB), BlockToChain(BlockToChain), LoopPredecessors(0) {
125 assert(BB && "Cannot create a chain with a null basic block");
126 BlockToChain[BB] = this;
129 /// \brief Iterator over blocks within the chain.
130 typedef SmallVectorImpl<MachineBasicBlock *>::iterator iterator;
132 /// \brief Beginning of blocks within the chain.
133 iterator begin() { return Blocks.begin(); }
135 /// \brief End of blocks within the chain.
136 iterator end() { return Blocks.end(); }
138 /// \brief Merge a block chain into this one.
140 /// This routine merges a block chain into this one. It takes care of forming
141 /// a contiguous sequence of basic blocks, updating the edge list, and
142 /// updating the block -> chain mapping. It does not free or tear down the
143 /// old chain, but the old chain's block list is no longer valid.
144 void merge(MachineBasicBlock *BB, BlockChain *Chain) {
146 assert(!Blocks.empty());
148 // Fast path in case we don't have a chain already.
150 assert(!BlockToChain[BB]);
151 Blocks.push_back(BB);
152 BlockToChain[BB] = this;
156 assert(BB == *Chain->begin());
157 assert(Chain->begin() != Chain->end());
159 // Update the incoming blocks to point to this chain, and add them to the
161 for (MachineBasicBlock *ChainBB : *Chain) {
162 Blocks.push_back(ChainBB);
163 assert(BlockToChain[ChainBB] == Chain && "Incoming blocks not in chain");
164 BlockToChain[ChainBB] = this;
169 /// \brief Dump the blocks in this chain.
170 LLVM_DUMP_METHOD void dump() {
171 for (MachineBasicBlock *MBB : *this)
176 /// \brief Count of predecessors within the loop currently being processed.
178 /// This count is updated at each loop we process to represent the number of
179 /// in-loop predecessors of this chain.
180 unsigned LoopPredecessors;
185 class MachineBlockPlacement : public MachineFunctionPass {
186 /// \brief A typedef for a block filter set.
187 typedef SmallPtrSet<MachineBasicBlock *, 16> BlockFilterSet;
189 /// \brief A handle to the branch probability pass.
190 const MachineBranchProbabilityInfo *MBPI;
192 /// \brief A handle to the function-wide block frequency pass.
193 const MachineBlockFrequencyInfo *MBFI;
195 /// \brief A handle to the loop info.
196 const MachineLoopInfo *MLI;
198 /// \brief A handle to the target's instruction info.
199 const TargetInstrInfo *TII;
201 /// \brief A handle to the target's lowering info.
202 const TargetLoweringBase *TLI;
204 /// \brief A handle to the post dominator tree.
205 MachineDominatorTree *MDT;
207 /// \brief A set of blocks that are unavoidably execute, i.e. they dominate
208 /// all terminators of the MachineFunction.
209 SmallPtrSet<MachineBasicBlock *, 4> UnavoidableBlocks;
211 /// \brief Allocator and owner of BlockChain structures.
213 /// We build BlockChains lazily while processing the loop structure of
214 /// a function. To reduce malloc traffic, we allocate them using this
215 /// slab-like allocator, and destroy them after the pass completes. An
216 /// important guarantee is that this allocator produces stable pointers to
218 SpecificBumpPtrAllocator<BlockChain> ChainAllocator;
220 /// \brief Function wide BasicBlock to BlockChain mapping.
222 /// This mapping allows efficiently moving from any given basic block to the
223 /// BlockChain it participates in, if any. We use it to, among other things,
224 /// allow implicitly defining edges between chains as the existing edges
225 /// between basic blocks.
226 DenseMap<MachineBasicBlock *, BlockChain *> BlockToChain;
228 void markChainSuccessors(BlockChain &Chain, MachineBasicBlock *LoopHeaderBB,
229 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
230 const BlockFilterSet *BlockFilter = nullptr);
231 MachineBasicBlock *selectBestSuccessor(MachineBasicBlock *BB,
233 const BlockFilterSet *BlockFilter);
235 selectBestCandidateBlock(BlockChain &Chain,
236 SmallVectorImpl<MachineBasicBlock *> &WorkList,
237 const BlockFilterSet *BlockFilter);
239 getFirstUnplacedBlock(MachineFunction &F, const BlockChain &PlacedChain,
240 MachineFunction::iterator &PrevUnplacedBlockIt,
241 const BlockFilterSet *BlockFilter);
242 void buildChain(MachineBasicBlock *BB, BlockChain &Chain,
243 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
244 const BlockFilterSet *BlockFilter = nullptr);
245 MachineBasicBlock *findBestLoopTop(MachineLoop &L,
246 const BlockFilterSet &LoopBlockSet);
247 MachineBasicBlock *findBestLoopExit(MachineFunction &F, MachineLoop &L,
248 const BlockFilterSet &LoopBlockSet);
249 void buildLoopChains(MachineFunction &F, MachineLoop &L);
250 void rotateLoop(BlockChain &LoopChain, MachineBasicBlock *ExitingBB,
251 const BlockFilterSet &LoopBlockSet);
252 void buildCFGChains(MachineFunction &F);
255 static char ID; // Pass identification, replacement for typeid
256 MachineBlockPlacement() : MachineFunctionPass(ID) {
257 initializeMachineBlockPlacementPass(*PassRegistry::getPassRegistry());
260 bool runOnMachineFunction(MachineFunction &F) override;
262 void getAnalysisUsage(AnalysisUsage &AU) const override {
263 AU.addRequired<MachineBranchProbabilityInfo>();
264 AU.addRequired<MachineBlockFrequencyInfo>();
265 AU.addRequired<MachineDominatorTree>();
266 AU.addRequired<MachineLoopInfo>();
267 MachineFunctionPass::getAnalysisUsage(AU);
272 char MachineBlockPlacement::ID = 0;
273 char &llvm::MachineBlockPlacementID = MachineBlockPlacement::ID;
274 INITIALIZE_PASS_BEGIN(MachineBlockPlacement, "block-placement",
275 "Branch Probability Basic Block Placement", false, false)
276 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
277 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
278 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
279 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
280 INITIALIZE_PASS_END(MachineBlockPlacement, "block-placement",
281 "Branch Probability Basic Block Placement", false, false)
284 /// \brief Helper to print the name of a MBB.
286 /// Only used by debug logging.
287 static std::string getBlockName(MachineBasicBlock *BB) {
289 raw_string_ostream OS(Result);
290 OS << "BB#" << BB->getNumber();
291 OS << " (derived from LLVM BB '" << BB->getName() << "')";
296 /// \brief Helper to print the number of a MBB.
298 /// Only used by debug logging.
299 static std::string getBlockNum(MachineBasicBlock *BB) {
301 raw_string_ostream OS(Result);
302 OS << "BB#" << BB->getNumber();
308 /// \brief Mark a chain's successors as having one fewer preds.
310 /// When a chain is being merged into the "placed" chain, this routine will
311 /// quickly walk the successors of each block in the chain and mark them as
312 /// having one fewer active predecessor. It also adds any successors of this
313 /// chain which reach the zero-predecessor state to the worklist passed in.
314 void MachineBlockPlacement::markChainSuccessors(
315 BlockChain &Chain, MachineBasicBlock *LoopHeaderBB,
316 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
317 const BlockFilterSet *BlockFilter) {
318 // Walk all the blocks in this chain, marking their successors as having
319 // a predecessor placed.
320 for (MachineBasicBlock *MBB : Chain) {
321 // Add any successors for which this is the only un-placed in-loop
322 // predecessor to the worklist as a viable candidate for CFG-neutral
323 // placement. No subsequent placement of this block will violate the CFG
324 // shape, so we get to use heuristics to choose a favorable placement.
325 for (MachineBasicBlock *Succ : MBB->successors()) {
326 if (BlockFilter && !BlockFilter->count(Succ))
328 BlockChain &SuccChain = *BlockToChain[Succ];
329 // Disregard edges within a fixed chain, or edges to the loop header.
330 if (&Chain == &SuccChain || Succ == LoopHeaderBB)
333 // This is a cross-chain edge that is within the loop, so decrement the
334 // loop predecessor count of the destination chain.
335 if (SuccChain.LoopPredecessors > 0 && --SuccChain.LoopPredecessors == 0)
336 BlockWorkList.push_back(*SuccChain.begin());
341 /// \brief Select the best successor for a block.
343 /// This looks across all successors of a particular block and attempts to
344 /// select the "best" one to be the layout successor. It only considers direct
345 /// successors which also pass the block filter. It will attempt to avoid
346 /// breaking CFG structure, but cave and break such structures in the case of
347 /// very hot successor edges.
349 /// \returns The best successor block found, or null if none are viable.
351 MachineBlockPlacement::selectBestSuccessor(MachineBasicBlock *BB,
353 const BlockFilterSet *BlockFilter) {
354 const BranchProbability HotProb(4, 5); // 80%
356 MachineBasicBlock *BestSucc = nullptr;
357 // FIXME: Due to the performance of the probability and weight routines in
358 // the MBPI analysis, we manually compute probabilities using the edge
359 // weights. This is suboptimal as it means that the somewhat subtle
360 // definition of edge weight semantics is encoded here as well. We should
361 // improve the MBPI interface to efficiently support query patterns such as
363 uint32_t BestWeight = 0;
364 uint32_t WeightScale = 0;
365 uint32_t SumWeight = MBPI->getSumForBlock(BB, WeightScale);
366 DEBUG(dbgs() << "Attempting merge from: " << getBlockName(BB) << "\n");
367 for (MachineBasicBlock *Succ : BB->successors()) {
368 if (BlockFilter && !BlockFilter->count(Succ))
370 BlockChain &SuccChain = *BlockToChain[Succ];
371 if (&SuccChain == &Chain) {
372 DEBUG(dbgs() << " " << getBlockName(Succ) << " -> Already merged!\n");
375 if (Succ != *SuccChain.begin()) {
376 DEBUG(dbgs() << " " << getBlockName(Succ) << " -> Mid chain!\n");
380 uint32_t SuccWeight = MBPI->getEdgeWeight(BB, Succ);
381 BranchProbability SuccProb(SuccWeight / WeightScale, SumWeight);
383 // If we outline optional branches, look whether Succ is unavoidable, i.e.
384 // dominates all terminators of the MachineFunction. If it does, other
385 // successors must be optional. Don't do this for cold branches.
386 if (OutlineOptionalBranches && SuccProb > HotProb.getCompl() &&
387 UnavoidableBlocks.count(Succ) > 0) {
388 auto HasShortOptionalBranch = [&]() {
389 for (MachineBasicBlock *Pred : Succ->predecessors()) {
390 // Check whether there is an unplaced optional branch.
391 if (Pred == Succ || (BlockFilter && !BlockFilter->count(Pred)) ||
392 BlockToChain[Pred] == &Chain)
394 // Check whether the optional branch has exactly one BB.
395 if (Pred->pred_size() > 1 || *Pred->pred_begin() != BB)
397 // Check whether the optional branch is small.
398 if (Pred->size() < OutlineOptionalThreshold)
403 if (!HasShortOptionalBranch())
407 // Only consider successors which are either "hot", or wouldn't violate
408 // any CFG constraints.
409 if (SuccChain.LoopPredecessors != 0) {
410 if (SuccProb < HotProb) {
411 DEBUG(dbgs() << " " << getBlockName(Succ) << " -> " << SuccProb
412 << " (prob) (CFG conflict)\n");
416 // Make sure that a hot successor doesn't have a globally more
417 // important predecessor.
418 BlockFrequency CandidateEdgeFreq =
419 MBFI->getBlockFreq(BB) * SuccProb * HotProb.getCompl();
420 bool BadCFGConflict = false;
421 for (MachineBasicBlock *Pred : Succ->predecessors()) {
422 if (Pred == Succ || (BlockFilter && !BlockFilter->count(Pred)) ||
423 BlockToChain[Pred] == &Chain)
425 BlockFrequency PredEdgeFreq =
426 MBFI->getBlockFreq(Pred) * MBPI->getEdgeProbability(Pred, Succ);
427 if (PredEdgeFreq >= CandidateEdgeFreq) {
428 BadCFGConflict = true;
432 if (BadCFGConflict) {
433 DEBUG(dbgs() << " " << getBlockName(Succ) << " -> " << SuccProb
434 << " (prob) (non-cold CFG conflict)\n");
439 DEBUG(dbgs() << " " << getBlockName(Succ) << " -> " << SuccProb
441 << (SuccChain.LoopPredecessors != 0 ? " (CFG break)" : "")
443 if (BestSucc && BestWeight >= SuccWeight)
446 BestWeight = SuccWeight;
451 /// \brief Select the best block from a worklist.
453 /// This looks through the provided worklist as a list of candidate basic
454 /// blocks and select the most profitable one to place. The definition of
455 /// profitable only really makes sense in the context of a loop. This returns
456 /// the most frequently visited block in the worklist, which in the case of
457 /// a loop, is the one most desirable to be physically close to the rest of the
458 /// loop body in order to improve icache behavior.
460 /// \returns The best block found, or null if none are viable.
461 MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock(
462 BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList,
463 const BlockFilterSet *BlockFilter) {
464 // Once we need to walk the worklist looking for a candidate, cleanup the
465 // worklist of already placed entries.
466 // FIXME: If this shows up on profiles, it could be folded (at the cost of
467 // some code complexity) into the loop below.
468 WorkList.erase(std::remove_if(WorkList.begin(), WorkList.end(),
469 [&](MachineBasicBlock *BB) {
470 return BlockToChain.lookup(BB) == &Chain;
474 MachineBasicBlock *BestBlock = nullptr;
475 BlockFrequency BestFreq;
476 for (MachineBasicBlock *MBB : WorkList) {
477 BlockChain &SuccChain = *BlockToChain[MBB];
478 if (&SuccChain == &Chain) {
479 DEBUG(dbgs() << " " << getBlockName(MBB) << " -> Already merged!\n");
482 assert(SuccChain.LoopPredecessors == 0 && "Found CFG-violating block");
484 BlockFrequency CandidateFreq = MBFI->getBlockFreq(MBB);
485 DEBUG(dbgs() << " " << getBlockName(MBB) << " -> ";
486 MBFI->printBlockFreq(dbgs(), CandidateFreq) << " (freq)\n");
487 if (BestBlock && BestFreq >= CandidateFreq)
490 BestFreq = CandidateFreq;
495 /// \brief Retrieve the first unplaced basic block.
497 /// This routine is called when we are unable to use the CFG to walk through
498 /// all of the basic blocks and form a chain due to unnatural loops in the CFG.
499 /// We walk through the function's blocks in order, starting from the
500 /// LastUnplacedBlockIt. We update this iterator on each call to avoid
501 /// re-scanning the entire sequence on repeated calls to this routine.
502 MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock(
503 MachineFunction &F, const BlockChain &PlacedChain,
504 MachineFunction::iterator &PrevUnplacedBlockIt,
505 const BlockFilterSet *BlockFilter) {
506 for (MachineFunction::iterator I = PrevUnplacedBlockIt, E = F.end(); I != E;
508 if (BlockFilter && !BlockFilter->count(I))
510 if (BlockToChain[I] != &PlacedChain) {
511 PrevUnplacedBlockIt = I;
512 // Now select the head of the chain to which the unplaced block belongs
513 // as the block to place. This will force the entire chain to be placed,
514 // and satisfies the requirements of merging chains.
515 return *BlockToChain[I]->begin();
521 void MachineBlockPlacement::buildChain(
522 MachineBasicBlock *BB, BlockChain &Chain,
523 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
524 const BlockFilterSet *BlockFilter) {
526 assert(BlockToChain[BB] == &Chain);
527 MachineFunction &F = *BB->getParent();
528 MachineFunction::iterator PrevUnplacedBlockIt = F.begin();
530 MachineBasicBlock *LoopHeaderBB = BB;
531 markChainSuccessors(Chain, LoopHeaderBB, BlockWorkList, BlockFilter);
532 BB = *std::prev(Chain.end());
535 assert(BlockToChain[BB] == &Chain);
536 assert(*std::prev(Chain.end()) == BB);
538 // Look for the best viable successor if there is one to place immediately
540 MachineBasicBlock *BestSucc = selectBestSuccessor(BB, Chain, BlockFilter);
542 // If an immediate successor isn't available, look for the best viable
543 // block among those we've identified as not violating the loop's CFG at
544 // this point. This won't be a fallthrough, but it will increase locality.
546 BestSucc = selectBestCandidateBlock(Chain, BlockWorkList, BlockFilter);
550 getFirstUnplacedBlock(F, Chain, PrevUnplacedBlockIt, BlockFilter);
554 DEBUG(dbgs() << "Unnatural loop CFG detected, forcibly merging the "
555 "layout successor until the CFG reduces\n");
558 // Place this block, updating the datastructures to reflect its placement.
559 BlockChain &SuccChain = *BlockToChain[BestSucc];
560 // Zero out LoopPredecessors for the successor we're about to merge in case
561 // we selected a successor that didn't fit naturally into the CFG.
562 SuccChain.LoopPredecessors = 0;
563 DEBUG(dbgs() << "Merging from " << getBlockNum(BB) << " to "
564 << getBlockNum(BestSucc) << "\n");
565 markChainSuccessors(SuccChain, LoopHeaderBB, BlockWorkList, BlockFilter);
566 Chain.merge(BestSucc, &SuccChain);
567 BB = *std::prev(Chain.end());
570 DEBUG(dbgs() << "Finished forming chain for header block "
571 << getBlockNum(*Chain.begin()) << "\n");
574 /// \brief Find the best loop top block for layout.
576 /// Look for a block which is strictly better than the loop header for laying
577 /// out at the top of the loop. This looks for one and only one pattern:
578 /// a latch block with no conditional exit. This block will cause a conditional
579 /// jump around it or will be the bottom of the loop if we lay it out in place,
580 /// but if it it doesn't end up at the bottom of the loop for any reason,
581 /// rotation alone won't fix it. Because such a block will always result in an
582 /// unconditional jump (for the backedge) rotating it in front of the loop
583 /// header is always profitable.
585 MachineBlockPlacement::findBestLoopTop(MachineLoop &L,
586 const BlockFilterSet &LoopBlockSet) {
587 // Check that the header hasn't been fused with a preheader block due to
588 // crazy branches. If it has, we need to start with the header at the top to
589 // prevent pulling the preheader into the loop body.
590 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
591 if (!LoopBlockSet.count(*HeaderChain.begin()))
592 return L.getHeader();
594 DEBUG(dbgs() << "Finding best loop top for: " << getBlockName(L.getHeader())
597 BlockFrequency BestPredFreq;
598 MachineBasicBlock *BestPred = nullptr;
599 for (MachineBasicBlock *Pred : L.getHeader()->predecessors()) {
600 if (!LoopBlockSet.count(Pred))
602 DEBUG(dbgs() << " header pred: " << getBlockName(Pred) << ", "
603 << Pred->succ_size() << " successors, ";
604 MBFI->printBlockFreq(dbgs(), Pred) << " freq\n");
605 if (Pred->succ_size() > 1)
608 BlockFrequency PredFreq = MBFI->getBlockFreq(Pred);
609 if (!BestPred || PredFreq > BestPredFreq ||
610 (!(PredFreq < BestPredFreq) &&
611 Pred->isLayoutSuccessor(L.getHeader()))) {
613 BestPredFreq = PredFreq;
617 // If no direct predecessor is fine, just use the loop header.
619 return L.getHeader();
621 // Walk backwards through any straight line of predecessors.
622 while (BestPred->pred_size() == 1 &&
623 (*BestPred->pred_begin())->succ_size() == 1 &&
624 *BestPred->pred_begin() != L.getHeader())
625 BestPred = *BestPred->pred_begin();
627 DEBUG(dbgs() << " final top: " << getBlockName(BestPred) << "\n");
631 /// \brief Find the best loop exiting block for layout.
633 /// This routine implements the logic to analyze the loop looking for the best
634 /// block to layout at the top of the loop. Typically this is done to maximize
635 /// fallthrough opportunities.
637 MachineBlockPlacement::findBestLoopExit(MachineFunction &F, MachineLoop &L,
638 const BlockFilterSet &LoopBlockSet) {
639 // We don't want to layout the loop linearly in all cases. If the loop header
640 // is just a normal basic block in the loop, we want to look for what block
641 // within the loop is the best one to layout at the top. However, if the loop
642 // header has be pre-merged into a chain due to predecessors not having
643 // analyzable branches, *and* the predecessor it is merged with is *not* part
644 // of the loop, rotating the header into the middle of the loop will create
645 // a non-contiguous range of blocks which is Very Bad. So start with the
646 // header and only rotate if safe.
647 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
648 if (!LoopBlockSet.count(*HeaderChain.begin()))
651 BlockFrequency BestExitEdgeFreq;
652 unsigned BestExitLoopDepth = 0;
653 MachineBasicBlock *ExitingBB = nullptr;
654 // If there are exits to outer loops, loop rotation can severely limit
655 // fallthrough opportunites unless it selects such an exit. Keep a set of
656 // blocks where rotating to exit with that block will reach an outer loop.
657 SmallPtrSet<MachineBasicBlock *, 4> BlocksExitingToOuterLoop;
659 DEBUG(dbgs() << "Finding best loop exit for: " << getBlockName(L.getHeader())
661 for (MachineBasicBlock *MBB : L.getBlocks()) {
662 BlockChain &Chain = *BlockToChain[MBB];
663 // Ensure that this block is at the end of a chain; otherwise it could be
664 // mid-way through an inner loop or a successor of an unanalyzable branch.
665 if (MBB != *std::prev(Chain.end()))
668 // Now walk the successors. We need to establish whether this has a viable
669 // exiting successor and whether it has a viable non-exiting successor.
670 // We store the old exiting state and restore it if a viable looping
671 // successor isn't found.
672 MachineBasicBlock *OldExitingBB = ExitingBB;
673 BlockFrequency OldBestExitEdgeFreq = BestExitEdgeFreq;
674 bool HasLoopingSucc = false;
675 // FIXME: Due to the performance of the probability and weight routines in
676 // the MBPI analysis, we use the internal weights and manually compute the
677 // probabilities to avoid quadratic behavior.
678 uint32_t WeightScale = 0;
679 uint32_t SumWeight = MBPI->getSumForBlock(MBB, WeightScale);
680 for (MachineBasicBlock *Succ : MBB->successors()) {
681 if (Succ->isLandingPad())
685 BlockChain &SuccChain = *BlockToChain[Succ];
686 // Don't split chains, either this chain or the successor's chain.
687 if (&Chain == &SuccChain) {
688 DEBUG(dbgs() << " exiting: " << getBlockName(MBB) << " -> "
689 << getBlockName(Succ) << " (chain conflict)\n");
693 uint32_t SuccWeight = MBPI->getEdgeWeight(MBB, Succ);
694 if (LoopBlockSet.count(Succ)) {
695 DEBUG(dbgs() << " looping: " << getBlockName(MBB) << " -> "
696 << getBlockName(Succ) << " (" << SuccWeight << ")\n");
697 HasLoopingSucc = true;
701 unsigned SuccLoopDepth = 0;
702 if (MachineLoop *ExitLoop = MLI->getLoopFor(Succ)) {
703 SuccLoopDepth = ExitLoop->getLoopDepth();
704 if (ExitLoop->contains(&L))
705 BlocksExitingToOuterLoop.insert(MBB);
708 BranchProbability SuccProb(SuccWeight / WeightScale, SumWeight);
709 BlockFrequency ExitEdgeFreq = MBFI->getBlockFreq(MBB) * SuccProb;
710 DEBUG(dbgs() << " exiting: " << getBlockName(MBB) << " -> "
711 << getBlockName(Succ) << " [L:" << SuccLoopDepth << "] (";
712 MBFI->printBlockFreq(dbgs(), ExitEdgeFreq) << ")\n");
713 // Note that we bias this toward an existing layout successor to retain
714 // incoming order in the absence of better information. The exit must have
715 // a frequency higher than the current exit before we consider breaking
717 BranchProbability Bias(100 - ExitBlockBias, 100);
718 if (!ExitingBB || SuccLoopDepth > BestExitLoopDepth ||
719 ExitEdgeFreq > BestExitEdgeFreq ||
720 (MBB->isLayoutSuccessor(Succ) &&
721 !(ExitEdgeFreq < BestExitEdgeFreq * Bias))) {
722 BestExitEdgeFreq = ExitEdgeFreq;
727 if (!HasLoopingSucc) {
728 // Restore the old exiting state, no viable looping successor was found.
729 ExitingBB = OldExitingBB;
730 BestExitEdgeFreq = OldBestExitEdgeFreq;
734 // Without a candidate exiting block or with only a single block in the
735 // loop, just use the loop header to layout the loop.
736 if (!ExitingBB || L.getNumBlocks() == 1)
739 // Also, if we have exit blocks which lead to outer loops but didn't select
740 // one of them as the exiting block we are rotating toward, disable loop
741 // rotation altogether.
742 if (!BlocksExitingToOuterLoop.empty() &&
743 !BlocksExitingToOuterLoop.count(ExitingBB))
746 DEBUG(dbgs() << " Best exiting block: " << getBlockName(ExitingBB) << "\n");
750 /// \brief Attempt to rotate an exiting block to the bottom of the loop.
752 /// Once we have built a chain, try to rotate it to line up the hot exit block
753 /// with fallthrough out of the loop if doing so doesn't introduce unnecessary
754 /// branches. For example, if the loop has fallthrough into its header and out
755 /// of its bottom already, don't rotate it.
756 void MachineBlockPlacement::rotateLoop(BlockChain &LoopChain,
757 MachineBasicBlock *ExitingBB,
758 const BlockFilterSet &LoopBlockSet) {
762 MachineBasicBlock *Top = *LoopChain.begin();
763 bool ViableTopFallthrough = false;
764 for (MachineBasicBlock *Pred : Top->predecessors()) {
765 BlockChain *PredChain = BlockToChain[Pred];
766 if (!LoopBlockSet.count(Pred) &&
767 (!PredChain || Pred == *std::prev(PredChain->end()))) {
768 ViableTopFallthrough = true;
773 // If the header has viable fallthrough, check whether the current loop
774 // bottom is a viable exiting block. If so, bail out as rotating will
775 // introduce an unnecessary branch.
776 if (ViableTopFallthrough) {
777 MachineBasicBlock *Bottom = *std::prev(LoopChain.end());
778 for (MachineBasicBlock *Succ : Bottom->successors()) {
779 BlockChain *SuccChain = BlockToChain[Succ];
780 if (!LoopBlockSet.count(Succ) &&
781 (!SuccChain || Succ == *SuccChain->begin()))
786 BlockChain::iterator ExitIt =
787 std::find(LoopChain.begin(), LoopChain.end(), ExitingBB);
788 if (ExitIt == LoopChain.end())
791 std::rotate(LoopChain.begin(), std::next(ExitIt), LoopChain.end());
794 /// \brief Forms basic block chains from the natural loop structures.
796 /// These chains are designed to preserve the existing *structure* of the code
797 /// as much as possible. We can then stitch the chains together in a way which
798 /// both preserves the topological structure and minimizes taken conditional
800 void MachineBlockPlacement::buildLoopChains(MachineFunction &F,
802 // First recurse through any nested loops, building chains for those inner
804 for (MachineLoop *InnerLoop : L)
805 buildLoopChains(F, *InnerLoop);
807 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
808 BlockFilterSet LoopBlockSet(L.block_begin(), L.block_end());
810 // First check to see if there is an obviously preferable top block for the
811 // loop. This will default to the header, but may end up as one of the
812 // predecessors to the header if there is one which will result in strictly
813 // fewer branches in the loop body.
814 MachineBasicBlock *LoopTop = findBestLoopTop(L, LoopBlockSet);
816 // If we selected just the header for the loop top, look for a potentially
817 // profitable exit block in the event that rotating the loop can eliminate
818 // branches by placing an exit edge at the bottom.
819 MachineBasicBlock *ExitingBB = nullptr;
820 if (LoopTop == L.getHeader())
821 ExitingBB = findBestLoopExit(F, L, LoopBlockSet);
823 BlockChain &LoopChain = *BlockToChain[LoopTop];
825 // FIXME: This is a really lame way of walking the chains in the loop: we
826 // walk the blocks, and use a set to prevent visiting a particular chain
828 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
829 assert(LoopChain.LoopPredecessors == 0);
830 UpdatedPreds.insert(&LoopChain);
831 for (MachineBasicBlock *LoopBB : L.getBlocks()) {
832 BlockChain &Chain = *BlockToChain[LoopBB];
833 if (!UpdatedPreds.insert(&Chain).second)
836 assert(Chain.LoopPredecessors == 0);
837 for (MachineBasicBlock *ChainBB : Chain) {
838 assert(BlockToChain[ChainBB] == &Chain);
839 for (MachineBasicBlock *Pred : ChainBB->predecessors()) {
840 if (BlockToChain[Pred] == &Chain || !LoopBlockSet.count(Pred))
842 ++Chain.LoopPredecessors;
846 if (Chain.LoopPredecessors == 0)
847 BlockWorkList.push_back(*Chain.begin());
850 buildChain(LoopTop, LoopChain, BlockWorkList, &LoopBlockSet);
851 rotateLoop(LoopChain, ExitingBB, LoopBlockSet);
854 // Crash at the end so we get all of the debugging output first.
855 bool BadLoop = false;
856 if (LoopChain.LoopPredecessors) {
858 dbgs() << "Loop chain contains a block without its preds placed!\n"
859 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
860 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n";
862 for (MachineBasicBlock *ChainBB : LoopChain) {
863 dbgs() << " ... " << getBlockName(ChainBB) << "\n";
864 if (!LoopBlockSet.erase(ChainBB)) {
865 // We don't mark the loop as bad here because there are real situations
866 // where this can occur. For example, with an unanalyzable fallthrough
867 // from a loop block to a non-loop block or vice versa.
868 dbgs() << "Loop chain contains a block not contained by the loop!\n"
869 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
870 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
871 << " Bad block: " << getBlockName(ChainBB) << "\n";
875 if (!LoopBlockSet.empty()) {
877 for (MachineBasicBlock *LoopBB : LoopBlockSet)
878 dbgs() << "Loop contains blocks never placed into a chain!\n"
879 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
880 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
881 << " Bad block: " << getBlockName(LoopBB) << "\n";
883 assert(!BadLoop && "Detected problems with the placement of this loop.");
887 void MachineBlockPlacement::buildCFGChains(MachineFunction &F) {
888 // Ensure that every BB in the function has an associated chain to simplify
889 // the assumptions of the remaining algorithm.
890 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
891 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
892 MachineBasicBlock *BB = FI;
894 new (ChainAllocator.Allocate()) BlockChain(BlockToChain, BB);
895 // Also, merge any blocks which we cannot reason about and must preserve
896 // the exact fallthrough behavior for.
899 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
900 if (!TII->AnalyzeBranch(*BB, TBB, FBB, Cond) || !FI->canFallThrough())
903 MachineFunction::iterator NextFI(std::next(FI));
904 MachineBasicBlock *NextBB = NextFI;
905 // Ensure that the layout successor is a viable block, as we know that
906 // fallthrough is a possibility.
907 assert(NextFI != FE && "Can't fallthrough past the last block.");
908 DEBUG(dbgs() << "Pre-merging due to unanalyzable fallthrough: "
909 << getBlockName(BB) << " -> " << getBlockName(NextBB)
911 Chain->merge(NextBB, nullptr);
917 if (OutlineOptionalBranches) {
918 // Find the nearest common dominator of all of F's terminators.
919 MachineBasicBlock *Terminator = nullptr;
920 for (MachineBasicBlock &MBB : F) {
921 if (MBB.succ_size() == 0) {
922 if (Terminator == nullptr)
925 Terminator = MDT->findNearestCommonDominator(Terminator, &MBB);
929 // MBBs dominating this common dominator are unavoidable.
930 UnavoidableBlocks.clear();
931 for (MachineBasicBlock &MBB : F) {
932 if (MDT->dominates(&MBB, Terminator)) {
933 UnavoidableBlocks.insert(&MBB);
938 // Build any loop-based chains.
939 for (MachineLoop *L : *MLI)
940 buildLoopChains(F, *L);
942 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
944 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
945 for (MachineBasicBlock &MBB : F) {
946 BlockChain &Chain = *BlockToChain[&MBB];
947 if (!UpdatedPreds.insert(&Chain).second)
950 assert(Chain.LoopPredecessors == 0);
951 for (MachineBasicBlock *ChainBB : Chain) {
952 assert(BlockToChain[ChainBB] == &Chain);
953 for (MachineBasicBlock *Pred : ChainBB->predecessors()) {
954 if (BlockToChain[Pred] == &Chain)
956 ++Chain.LoopPredecessors;
960 if (Chain.LoopPredecessors == 0)
961 BlockWorkList.push_back(*Chain.begin());
964 BlockChain &FunctionChain = *BlockToChain[&F.front()];
965 buildChain(&F.front(), FunctionChain, BlockWorkList);
968 typedef SmallPtrSet<MachineBasicBlock *, 16> FunctionBlockSetType;
971 // Crash at the end so we get all of the debugging output first.
972 bool BadFunc = false;
973 FunctionBlockSetType FunctionBlockSet;
974 for (MachineBasicBlock &MBB : F)
975 FunctionBlockSet.insert(&MBB);
977 for (MachineBasicBlock *ChainBB : FunctionChain)
978 if (!FunctionBlockSet.erase(ChainBB)) {
980 dbgs() << "Function chain contains a block not in the function!\n"
981 << " Bad block: " << getBlockName(ChainBB) << "\n";
984 if (!FunctionBlockSet.empty()) {
986 for (MachineBasicBlock *RemainingBB : FunctionBlockSet)
987 dbgs() << "Function contains blocks never placed into a chain!\n"
988 << " Bad block: " << getBlockName(RemainingBB) << "\n";
990 assert(!BadFunc && "Detected problems with the block placement.");
993 // Splice the blocks into place.
994 MachineFunction::iterator InsertPos = F.begin();
995 for (MachineBasicBlock *ChainBB : FunctionChain) {
996 DEBUG(dbgs() << (ChainBB == *FunctionChain.begin() ? "Placing chain "
998 << getBlockName(ChainBB) << "\n");
999 if (InsertPos != MachineFunction::iterator(ChainBB))
1000 F.splice(InsertPos, ChainBB);
1004 // Update the terminator of the previous block.
1005 if (ChainBB == *FunctionChain.begin())
1007 MachineBasicBlock *PrevBB = std::prev(MachineFunction::iterator(ChainBB));
1009 // FIXME: It would be awesome of updateTerminator would just return rather
1010 // than assert when the branch cannot be analyzed in order to remove this
1013 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
1014 if (!TII->AnalyzeBranch(*PrevBB, TBB, FBB, Cond)) {
1015 // The "PrevBB" is not yet updated to reflect current code layout, so,
1016 // o. it may fall-through to a block without explict "goto" instruction
1017 // before layout, and no longer fall-through it after layout; or
1018 // o. just opposite.
1020 // AnalyzeBranch() may return erroneous value for FBB when these two
1021 // situations take place. For the first scenario FBB is mistakenly set
1022 // NULL; for the 2nd scenario, the FBB, which is expected to be NULL,
1023 // is mistakenly pointing to "*BI".
1025 bool needUpdateBr = true;
1026 if (!Cond.empty() && (!FBB || FBB == ChainBB)) {
1027 PrevBB->updateTerminator();
1028 needUpdateBr = false;
1030 TBB = FBB = nullptr;
1031 if (TII->AnalyzeBranch(*PrevBB, TBB, FBB, Cond)) {
1032 // FIXME: This should never take place.
1033 TBB = FBB = nullptr;
1037 // If PrevBB has a two-way branch, try to re-order the branches
1038 // such that we branch to the successor with higher weight first.
1039 if (TBB && !Cond.empty() && FBB &&
1040 MBPI->getEdgeWeight(PrevBB, FBB) > MBPI->getEdgeWeight(PrevBB, TBB) &&
1041 !TII->ReverseBranchCondition(Cond)) {
1042 DEBUG(dbgs() << "Reverse order of the two branches: "
1043 << getBlockName(PrevBB) << "\n");
1044 DEBUG(dbgs() << " Edge weight: " << MBPI->getEdgeWeight(PrevBB, FBB)
1045 << " vs " << MBPI->getEdgeWeight(PrevBB, TBB) << "\n");
1046 DebugLoc dl; // FIXME: this is nowhere
1047 TII->RemoveBranch(*PrevBB);
1048 TII->InsertBranch(*PrevBB, FBB, TBB, Cond, dl);
1049 needUpdateBr = true;
1052 PrevBB->updateTerminator();
1056 // Fixup the last block.
1058 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
1059 if (!TII->AnalyzeBranch(F.back(), TBB, FBB, Cond))
1060 F.back().updateTerminator();
1062 // Walk through the backedges of the function now that we have fully laid out
1063 // the basic blocks and align the destination of each backedge. We don't rely
1064 // exclusively on the loop info here so that we can align backedges in
1065 // unnatural CFGs and backedges that were introduced purely because of the
1066 // loop rotations done during this layout pass.
1067 // FIXME: Use Function::optForSize().
1068 if (F.getFunction()->hasFnAttribute(Attribute::OptimizeForSize))
1070 if (FunctionChain.begin() == FunctionChain.end())
1071 return; // Empty chain.
1073 const BranchProbability ColdProb(1, 5); // 20%
1074 BlockFrequency EntryFreq = MBFI->getBlockFreq(F.begin());
1075 BlockFrequency WeightedEntryFreq = EntryFreq * ColdProb;
1076 for (MachineBasicBlock *ChainBB : FunctionChain) {
1077 if (ChainBB == *FunctionChain.begin())
1080 // Don't align non-looping basic blocks. These are unlikely to execute
1081 // enough times to matter in practice. Note that we'll still handle
1082 // unnatural CFGs inside of a natural outer loop (the common case) and
1084 MachineLoop *L = MLI->getLoopFor(ChainBB);
1088 unsigned Align = TLI->getPrefLoopAlignment(L);
1090 continue; // Don't care about loop alignment.
1092 // If the block is cold relative to the function entry don't waste space
1094 BlockFrequency Freq = MBFI->getBlockFreq(ChainBB);
1095 if (Freq < WeightedEntryFreq)
1098 // If the block is cold relative to its loop header, don't align it
1099 // regardless of what edges into the block exist.
1100 MachineBasicBlock *LoopHeader = L->getHeader();
1101 BlockFrequency LoopHeaderFreq = MBFI->getBlockFreq(LoopHeader);
1102 if (Freq < (LoopHeaderFreq * ColdProb))
1105 // Check for the existence of a non-layout predecessor which would benefit
1106 // from aligning this block.
1107 MachineBasicBlock *LayoutPred =
1108 &*std::prev(MachineFunction::iterator(ChainBB));
1110 // Force alignment if all the predecessors are jumps. We already checked
1111 // that the block isn't cold above.
1112 if (!LayoutPred->isSuccessor(ChainBB)) {
1113 ChainBB->setAlignment(Align);
1117 // Align this block if the layout predecessor's edge into this block is
1118 // cold relative to the block. When this is true, other predecessors make up
1119 // all of the hot entries into the block and thus alignment is likely to be
1121 BranchProbability LayoutProb =
1122 MBPI->getEdgeProbability(LayoutPred, ChainBB);
1123 BlockFrequency LayoutEdgeFreq = MBFI->getBlockFreq(LayoutPred) * LayoutProb;
1124 if (LayoutEdgeFreq <= (Freq * ColdProb))
1125 ChainBB->setAlignment(Align);
1129 bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &F) {
1130 // Check for single-block functions and skip them.
1131 if (std::next(F.begin()) == F.end())
1134 if (skipOptnoneFunction(*F.getFunction()))
1137 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
1138 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
1139 MLI = &getAnalysis<MachineLoopInfo>();
1140 TII = F.getSubtarget().getInstrInfo();
1141 TLI = F.getSubtarget().getTargetLowering();
1142 MDT = &getAnalysis<MachineDominatorTree>();
1143 assert(BlockToChain.empty());
1147 BlockToChain.clear();
1148 ChainAllocator.DestroyAll();
1151 // Align all of the blocks in the function to a specific alignment.
1152 for (MachineBasicBlock &MBB : F)
1153 MBB.setAlignment(AlignAllBlock);
1155 // We always return true as we have no way to track whether the final order
1156 // differs from the original order.
1161 /// \brief A pass to compute block placement statistics.
1163 /// A separate pass to compute interesting statistics for evaluating block
1164 /// placement. This is separate from the actual placement pass so that they can
1165 /// be computed in the absence of any placement transformations or when using
1166 /// alternative placement strategies.
1167 class MachineBlockPlacementStats : public MachineFunctionPass {
1168 /// \brief A handle to the branch probability pass.
1169 const MachineBranchProbabilityInfo *MBPI;
1171 /// \brief A handle to the function-wide block frequency pass.
1172 const MachineBlockFrequencyInfo *MBFI;
1175 static char ID; // Pass identification, replacement for typeid
1176 MachineBlockPlacementStats() : MachineFunctionPass(ID) {
1177 initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry());
1180 bool runOnMachineFunction(MachineFunction &F) override;
1182 void getAnalysisUsage(AnalysisUsage &AU) const override {
1183 AU.addRequired<MachineBranchProbabilityInfo>();
1184 AU.addRequired<MachineBlockFrequencyInfo>();
1185 AU.setPreservesAll();
1186 MachineFunctionPass::getAnalysisUsage(AU);
1191 char MachineBlockPlacementStats::ID = 0;
1192 char &llvm::MachineBlockPlacementStatsID = MachineBlockPlacementStats::ID;
1193 INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats",
1194 "Basic Block Placement Stats", false, false)
1195 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
1196 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
1197 INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats",
1198 "Basic Block Placement Stats", false, false)
1200 bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) {
1201 // Check for single-block functions and skip them.
1202 if (std::next(F.begin()) == F.end())
1205 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
1206 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
1208 for (MachineBasicBlock &MBB : F) {
1209 BlockFrequency BlockFreq = MBFI->getBlockFreq(&MBB);
1210 Statistic &NumBranches =
1211 (MBB.succ_size() > 1) ? NumCondBranches : NumUncondBranches;
1212 Statistic &BranchTakenFreq =
1213 (MBB.succ_size() > 1) ? CondBranchTakenFreq : UncondBranchTakenFreq;
1214 for (MachineBasicBlock *Succ : MBB.successors()) {
1215 // Skip if this successor is a fallthrough.
1216 if (MBB.isLayoutSuccessor(Succ))
1219 BlockFrequency EdgeFreq =
1220 BlockFreq * MBPI->getEdgeProbability(&MBB, Succ);
1222 BranchTakenFreq += EdgeFreq.getFrequency();