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/MachineFunction.h"
37 #include "llvm/CodeGen/MachineFunctionPass.h"
38 #include "llvm/CodeGen/MachineLoopInfo.h"
39 #include "llvm/CodeGen/MachineModuleInfo.h"
40 #include "llvm/Support/Allocator.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Target/TargetInstrInfo.h"
44 #include "llvm/Target/TargetLowering.h"
45 #include "llvm/Target/TargetSubtargetInfo.h"
49 #define DEBUG_TYPE "block-placement2"
51 STATISTIC(NumCondBranches, "Number of conditional branches");
52 STATISTIC(NumUncondBranches, "Number of uncondittional branches");
53 STATISTIC(CondBranchTakenFreq,
54 "Potential frequency of taking conditional branches");
55 STATISTIC(UncondBranchTakenFreq,
56 "Potential frequency of taking unconditional branches");
58 static cl::opt<unsigned> AlignAllBlock("align-all-blocks",
59 cl::desc("Force the alignment of all "
60 "blocks in the function."),
61 cl::init(0), cl::Hidden);
63 // FIXME: Find a good default for this flag and remove the flag.
64 static cl::opt<unsigned>
65 ExitBlockBias("block-placement-exit-block-bias",
66 cl::desc("Block frequency percentage a loop exit block needs "
67 "over the original exit to be considered the new exit."),
68 cl::init(0), cl::Hidden);
72 /// \brief Type for our function-wide basic block -> block chain mapping.
73 typedef DenseMap<MachineBasicBlock *, BlockChain *> BlockToChainMapType;
77 /// \brief A chain of blocks which will be laid out contiguously.
79 /// This is the datastructure representing a chain of consecutive blocks that
80 /// are profitable to layout together in order to maximize fallthrough
81 /// probabilities and code locality. We also can use a block chain to represent
82 /// a sequence of basic blocks which have some external (correctness)
83 /// requirement for sequential layout.
85 /// Chains can be built around a single basic block and can be merged to grow
86 /// them. They participate in a block-to-chain mapping, which is updated
87 /// automatically as chains are merged together.
89 /// \brief The sequence of blocks belonging to this chain.
91 /// This is the sequence of blocks for a particular chain. These will be laid
92 /// out in-order within the function.
93 SmallVector<MachineBasicBlock *, 4> Blocks;
95 /// \brief A handle to the function-wide basic block to block chain mapping.
97 /// This is retained in each block chain to simplify the computation of child
98 /// block chains for SCC-formation and iteration. We store the edges to child
99 /// basic blocks, and map them back to their associated chains using this
101 BlockToChainMapType &BlockToChain;
104 /// \brief Construct a new BlockChain.
106 /// This builds a new block chain representing a single basic block in the
107 /// function. It also registers itself as the chain that block participates
108 /// in with the BlockToChain mapping.
109 BlockChain(BlockToChainMapType &BlockToChain, MachineBasicBlock *BB)
110 : Blocks(1, BB), BlockToChain(BlockToChain), LoopPredecessors(0) {
111 assert(BB && "Cannot create a chain with a null basic block");
112 BlockToChain[BB] = this;
115 /// \brief Iterator over blocks within the chain.
116 typedef SmallVectorImpl<MachineBasicBlock *>::iterator iterator;
118 /// \brief Beginning of blocks within the chain.
119 iterator begin() { return Blocks.begin(); }
121 /// \brief End of blocks within the chain.
122 iterator end() { return Blocks.end(); }
124 /// \brief Merge a block chain into this one.
126 /// This routine merges a block chain into this one. It takes care of forming
127 /// a contiguous sequence of basic blocks, updating the edge list, and
128 /// updating the block -> chain mapping. It does not free or tear down the
129 /// old chain, but the old chain's block list is no longer valid.
130 void merge(MachineBasicBlock *BB, BlockChain *Chain) {
132 assert(!Blocks.empty());
134 // Fast path in case we don't have a chain already.
136 assert(!BlockToChain[BB]);
137 Blocks.push_back(BB);
138 BlockToChain[BB] = this;
142 assert(BB == *Chain->begin());
143 assert(Chain->begin() != Chain->end());
145 // Update the incoming blocks to point to this chain, and add them to the
147 for (BlockChain::iterator BI = Chain->begin(), BE = Chain->end();
149 Blocks.push_back(*BI);
150 assert(BlockToChain[*BI] == Chain && "Incoming blocks not in chain");
151 BlockToChain[*BI] = this;
156 /// \brief Dump the blocks in this chain.
157 LLVM_DUMP_METHOD void dump() {
158 for (iterator I = begin(), E = end(); I != E; ++I)
163 /// \brief Count of predecessors within the loop currently being processed.
165 /// This count is updated at each loop we process to represent the number of
166 /// in-loop predecessors of this chain.
167 unsigned LoopPredecessors;
172 class MachineBlockPlacement : public MachineFunctionPass {
173 /// \brief A typedef for a block filter set.
174 typedef SmallPtrSet<MachineBasicBlock *, 16> BlockFilterSet;
176 /// \brief A handle to the branch probability pass.
177 const MachineBranchProbabilityInfo *MBPI;
179 /// \brief A handle to the function-wide block frequency pass.
180 const MachineBlockFrequencyInfo *MBFI;
182 /// \brief A handle to the loop info.
183 const MachineLoopInfo *MLI;
185 /// \brief A handle to the target's instruction info.
186 const TargetInstrInfo *TII;
188 /// \brief A handle to the target's lowering info.
189 const TargetLoweringBase *TLI;
191 /// \brief Allocator and owner of BlockChain structures.
193 /// We build BlockChains lazily while processing the loop structure of
194 /// a function. To reduce malloc traffic, we allocate them using this
195 /// slab-like allocator, and destroy them after the pass completes. An
196 /// important guarantee is that this allocator produces stable pointers to
198 SpecificBumpPtrAllocator<BlockChain> ChainAllocator;
200 /// \brief Function wide BasicBlock to BlockChain mapping.
202 /// This mapping allows efficiently moving from any given basic block to the
203 /// BlockChain it participates in, if any. We use it to, among other things,
204 /// allow implicitly defining edges between chains as the existing edges
205 /// between basic blocks.
206 DenseMap<MachineBasicBlock *, BlockChain *> BlockToChain;
208 void markChainSuccessors(BlockChain &Chain,
209 MachineBasicBlock *LoopHeaderBB,
210 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
211 const BlockFilterSet *BlockFilter = nullptr);
212 MachineBasicBlock *selectBestSuccessor(MachineBasicBlock *BB,
214 const BlockFilterSet *BlockFilter);
215 MachineBasicBlock *selectBestCandidateBlock(
216 BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList,
217 const BlockFilterSet *BlockFilter);
218 MachineBasicBlock *getFirstUnplacedBlock(
220 const BlockChain &PlacedChain,
221 MachineFunction::iterator &PrevUnplacedBlockIt,
222 const BlockFilterSet *BlockFilter);
223 void buildChain(MachineBasicBlock *BB, BlockChain &Chain,
224 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
225 const BlockFilterSet *BlockFilter = nullptr);
226 MachineBasicBlock *findBestLoopTop(MachineLoop &L,
227 const BlockFilterSet &LoopBlockSet);
228 MachineBasicBlock *findBestLoopExit(MachineFunction &F,
230 const BlockFilterSet &LoopBlockSet);
231 void buildLoopChains(MachineFunction &F, MachineLoop &L);
232 void rotateLoop(BlockChain &LoopChain, MachineBasicBlock *ExitingBB,
233 const BlockFilterSet &LoopBlockSet);
234 void buildCFGChains(MachineFunction &F);
237 static char ID; // Pass identification, replacement for typeid
238 MachineBlockPlacement() : MachineFunctionPass(ID) {
239 initializeMachineBlockPlacementPass(*PassRegistry::getPassRegistry());
242 bool runOnMachineFunction(MachineFunction &F) override;
244 void getAnalysisUsage(AnalysisUsage &AU) const override {
245 AU.addRequired<MachineBranchProbabilityInfo>();
246 AU.addRequired<MachineBlockFrequencyInfo>();
247 AU.addRequired<MachineLoopInfo>();
248 MachineFunctionPass::getAnalysisUsage(AU);
253 char MachineBlockPlacement::ID = 0;
254 char &llvm::MachineBlockPlacementID = MachineBlockPlacement::ID;
255 INITIALIZE_PASS_BEGIN(MachineBlockPlacement, "block-placement2",
256 "Branch Probability Basic Block Placement", false, false)
257 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
258 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
259 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
260 INITIALIZE_PASS_END(MachineBlockPlacement, "block-placement2",
261 "Branch Probability Basic Block Placement", false, false)
264 /// \brief Helper to print the name of a MBB.
266 /// Only used by debug logging.
267 static std::string getBlockName(MachineBasicBlock *BB) {
269 raw_string_ostream OS(Result);
270 OS << "BB#" << BB->getNumber()
271 << " (derived from LLVM BB '" << BB->getName() << "')";
276 /// \brief Helper to print the number of a MBB.
278 /// Only used by debug logging.
279 static std::string getBlockNum(MachineBasicBlock *BB) {
281 raw_string_ostream OS(Result);
282 OS << "BB#" << BB->getNumber();
288 /// \brief Mark a chain's successors as having one fewer preds.
290 /// When a chain is being merged into the "placed" chain, this routine will
291 /// quickly walk the successors of each block in the chain and mark them as
292 /// having one fewer active predecessor. It also adds any successors of this
293 /// chain which reach the zero-predecessor state to the worklist passed in.
294 void MachineBlockPlacement::markChainSuccessors(
296 MachineBasicBlock *LoopHeaderBB,
297 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
298 const BlockFilterSet *BlockFilter) {
299 // Walk all the blocks in this chain, marking their successors as having
300 // a predecessor placed.
301 for (BlockChain::iterator CBI = Chain.begin(), CBE = Chain.end();
303 // Add any successors for which this is the only un-placed in-loop
304 // predecessor to the worklist as a viable candidate for CFG-neutral
305 // placement. No subsequent placement of this block will violate the CFG
306 // shape, so we get to use heuristics to choose a favorable placement.
307 for (MachineBasicBlock::succ_iterator SI = (*CBI)->succ_begin(),
308 SE = (*CBI)->succ_end();
310 if (BlockFilter && !BlockFilter->count(*SI))
312 BlockChain &SuccChain = *BlockToChain[*SI];
313 // Disregard edges within a fixed chain, or edges to the loop header.
314 if (&Chain == &SuccChain || *SI == LoopHeaderBB)
317 // This is a cross-chain edge that is within the loop, so decrement the
318 // loop predecessor count of the destination chain.
319 if (SuccChain.LoopPredecessors > 0 && --SuccChain.LoopPredecessors == 0)
320 BlockWorkList.push_back(*SuccChain.begin());
325 /// \brief Select the best successor for a block.
327 /// This looks across all successors of a particular block and attempts to
328 /// select the "best" one to be the layout successor. It only considers direct
329 /// successors which also pass the block filter. It will attempt to avoid
330 /// breaking CFG structure, but cave and break such structures in the case of
331 /// very hot successor edges.
333 /// \returns The best successor block found, or null if none are viable.
334 MachineBasicBlock *MachineBlockPlacement::selectBestSuccessor(
335 MachineBasicBlock *BB, BlockChain &Chain,
336 const BlockFilterSet *BlockFilter) {
337 const BranchProbability HotProb(4, 5); // 80%
339 MachineBasicBlock *BestSucc = nullptr;
340 // FIXME: Due to the performance of the probability and weight routines in
341 // the MBPI analysis, we manually compute probabilities using the edge
342 // weights. This is suboptimal as it means that the somewhat subtle
343 // definition of edge weight semantics is encoded here as well. We should
344 // improve the MBPI interface to efficiently support query patterns such as
346 uint32_t BestWeight = 0;
347 uint32_t WeightScale = 0;
348 uint32_t SumWeight = MBPI->getSumForBlock(BB, WeightScale);
349 DEBUG(dbgs() << "Attempting merge from: " << getBlockName(BB) << "\n");
350 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
353 if (BlockFilter && !BlockFilter->count(*SI))
355 BlockChain &SuccChain = *BlockToChain[*SI];
356 if (&SuccChain == &Chain) {
357 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> Already merged!\n");
360 if (*SI != *SuccChain.begin()) {
361 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> Mid chain!\n");
365 uint32_t SuccWeight = MBPI->getEdgeWeight(BB, *SI);
366 BranchProbability SuccProb(SuccWeight / WeightScale, SumWeight);
368 // Only consider successors which are either "hot", or wouldn't violate
369 // any CFG constraints.
370 if (SuccChain.LoopPredecessors != 0) {
371 if (SuccProb < HotProb) {
372 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> " << SuccProb
373 << " (prob) (CFG conflict)\n");
377 // Make sure that a hot successor doesn't have a globally more important
379 BlockFrequency CandidateEdgeFreq
380 = MBFI->getBlockFreq(BB) * SuccProb * HotProb.getCompl();
381 bool BadCFGConflict = false;
382 for (MachineBasicBlock::pred_iterator PI = (*SI)->pred_begin(),
383 PE = (*SI)->pred_end();
385 if (*PI == *SI || (BlockFilter && !BlockFilter->count(*PI)) ||
386 BlockToChain[*PI] == &Chain)
388 BlockFrequency PredEdgeFreq
389 = MBFI->getBlockFreq(*PI) * MBPI->getEdgeProbability(*PI, *SI);
390 if (PredEdgeFreq >= CandidateEdgeFreq) {
391 BadCFGConflict = true;
395 if (BadCFGConflict) {
396 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> " << SuccProb
397 << " (prob) (non-cold CFG conflict)\n");
402 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> " << SuccProb
404 << (SuccChain.LoopPredecessors != 0 ? " (CFG break)" : "")
406 if (BestSucc && BestWeight >= SuccWeight)
409 BestWeight = SuccWeight;
414 /// \brief Select the best block from a worklist.
416 /// This looks through the provided worklist as a list of candidate basic
417 /// blocks and select the most profitable one to place. The definition of
418 /// profitable only really makes sense in the context of a loop. This returns
419 /// the most frequently visited block in the worklist, which in the case of
420 /// a loop, is the one most desirable to be physically close to the rest of the
421 /// loop body in order to improve icache behavior.
423 /// \returns The best block found, or null if none are viable.
424 MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock(
425 BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList,
426 const BlockFilterSet *BlockFilter) {
427 // Once we need to walk the worklist looking for a candidate, cleanup the
428 // worklist of already placed entries.
429 // FIXME: If this shows up on profiles, it could be folded (at the cost of
430 // some code complexity) into the loop below.
431 WorkList.erase(std::remove_if(WorkList.begin(), WorkList.end(),
432 [&](MachineBasicBlock *BB) {
433 return BlockToChain.lookup(BB) == &Chain;
437 MachineBasicBlock *BestBlock = nullptr;
438 BlockFrequency BestFreq;
439 for (SmallVectorImpl<MachineBasicBlock *>::iterator WBI = WorkList.begin(),
440 WBE = WorkList.end();
442 BlockChain &SuccChain = *BlockToChain[*WBI];
443 if (&SuccChain == &Chain) {
444 DEBUG(dbgs() << " " << getBlockName(*WBI)
445 << " -> Already merged!\n");
448 assert(SuccChain.LoopPredecessors == 0 && "Found CFG-violating block");
450 BlockFrequency CandidateFreq = MBFI->getBlockFreq(*WBI);
451 DEBUG(dbgs() << " " << getBlockName(*WBI) << " -> ";
452 MBFI->printBlockFreq(dbgs(), CandidateFreq) << " (freq)\n");
453 if (BestBlock && BestFreq >= CandidateFreq)
456 BestFreq = CandidateFreq;
461 /// \brief Retrieve the first unplaced basic block.
463 /// This routine is called when we are unable to use the CFG to walk through
464 /// all of the basic blocks and form a chain due to unnatural loops in the CFG.
465 /// We walk through the function's blocks in order, starting from the
466 /// LastUnplacedBlockIt. We update this iterator on each call to avoid
467 /// re-scanning the entire sequence on repeated calls to this routine.
468 MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock(
469 MachineFunction &F, const BlockChain &PlacedChain,
470 MachineFunction::iterator &PrevUnplacedBlockIt,
471 const BlockFilterSet *BlockFilter) {
472 for (MachineFunction::iterator I = PrevUnplacedBlockIt, E = F.end(); I != E;
474 if (BlockFilter && !BlockFilter->count(I))
476 if (BlockToChain[I] != &PlacedChain) {
477 PrevUnplacedBlockIt = I;
478 // Now select the head of the chain to which the unplaced block belongs
479 // as the block to place. This will force the entire chain to be placed,
480 // and satisfies the requirements of merging chains.
481 return *BlockToChain[I]->begin();
487 void MachineBlockPlacement::buildChain(
488 MachineBasicBlock *BB,
490 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
491 const BlockFilterSet *BlockFilter) {
493 assert(BlockToChain[BB] == &Chain);
494 MachineFunction &F = *BB->getParent();
495 MachineFunction::iterator PrevUnplacedBlockIt = F.begin();
497 MachineBasicBlock *LoopHeaderBB = BB;
498 markChainSuccessors(Chain, LoopHeaderBB, BlockWorkList, BlockFilter);
499 BB = *std::prev(Chain.end());
502 assert(BlockToChain[BB] == &Chain);
503 assert(*std::prev(Chain.end()) == BB);
505 // Look for the best viable successor if there is one to place immediately
507 MachineBasicBlock *BestSucc = selectBestSuccessor(BB, Chain, BlockFilter);
509 // If an immediate successor isn't available, look for the best viable
510 // block among those we've identified as not violating the loop's CFG at
511 // this point. This won't be a fallthrough, but it will increase locality.
513 BestSucc = selectBestCandidateBlock(Chain, BlockWorkList, BlockFilter);
516 BestSucc = getFirstUnplacedBlock(F, Chain, PrevUnplacedBlockIt,
521 DEBUG(dbgs() << "Unnatural loop CFG detected, forcibly merging the "
522 "layout successor until the CFG reduces\n");
525 // Place this block, updating the datastructures to reflect its placement.
526 BlockChain &SuccChain = *BlockToChain[BestSucc];
527 // Zero out LoopPredecessors for the successor we're about to merge in case
528 // we selected a successor that didn't fit naturally into the CFG.
529 SuccChain.LoopPredecessors = 0;
530 DEBUG(dbgs() << "Merging from " << getBlockNum(BB)
531 << " to " << getBlockNum(BestSucc) << "\n");
532 markChainSuccessors(SuccChain, LoopHeaderBB, BlockWorkList, BlockFilter);
533 Chain.merge(BestSucc, &SuccChain);
534 BB = *std::prev(Chain.end());
537 DEBUG(dbgs() << "Finished forming chain for header block "
538 << getBlockNum(*Chain.begin()) << "\n");
541 /// \brief Find the best loop top block for layout.
543 /// Look for a block which is strictly better than the loop header for laying
544 /// out at the top of the loop. This looks for one and only one pattern:
545 /// a latch block with no conditional exit. This block will cause a conditional
546 /// jump around it or will be the bottom of the loop if we lay it out in place,
547 /// but if it it doesn't end up at the bottom of the loop for any reason,
548 /// rotation alone won't fix it. Because such a block will always result in an
549 /// unconditional jump (for the backedge) rotating it in front of the loop
550 /// header is always profitable.
552 MachineBlockPlacement::findBestLoopTop(MachineLoop &L,
553 const BlockFilterSet &LoopBlockSet) {
554 // Check that the header hasn't been fused with a preheader block due to
555 // crazy branches. If it has, we need to start with the header at the top to
556 // prevent pulling the preheader into the loop body.
557 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
558 if (!LoopBlockSet.count(*HeaderChain.begin()))
559 return L.getHeader();
561 DEBUG(dbgs() << "Finding best loop top for: "
562 << getBlockName(L.getHeader()) << "\n");
564 BlockFrequency BestPredFreq;
565 MachineBasicBlock *BestPred = nullptr;
566 for (MachineBasicBlock::pred_iterator PI = L.getHeader()->pred_begin(),
567 PE = L.getHeader()->pred_end();
569 MachineBasicBlock *Pred = *PI;
570 if (!LoopBlockSet.count(Pred))
572 DEBUG(dbgs() << " header pred: " << getBlockName(Pred) << ", "
573 << Pred->succ_size() << " successors, ";
574 MBFI->printBlockFreq(dbgs(), Pred) << " freq\n");
575 if (Pred->succ_size() > 1)
578 BlockFrequency PredFreq = MBFI->getBlockFreq(Pred);
579 if (!BestPred || PredFreq > BestPredFreq ||
580 (!(PredFreq < BestPredFreq) &&
581 Pred->isLayoutSuccessor(L.getHeader()))) {
583 BestPredFreq = PredFreq;
587 // If no direct predecessor is fine, just use the loop header.
589 return L.getHeader();
591 // Walk backwards through any straight line of predecessors.
592 while (BestPred->pred_size() == 1 &&
593 (*BestPred->pred_begin())->succ_size() == 1 &&
594 *BestPred->pred_begin() != L.getHeader())
595 BestPred = *BestPred->pred_begin();
597 DEBUG(dbgs() << " final top: " << getBlockName(BestPred) << "\n");
602 /// \brief Find the best loop exiting block for layout.
604 /// This routine implements the logic to analyze the loop looking for the best
605 /// block to layout at the top of the loop. Typically this is done to maximize
606 /// fallthrough opportunities.
608 MachineBlockPlacement::findBestLoopExit(MachineFunction &F,
610 const BlockFilterSet &LoopBlockSet) {
611 // We don't want to layout the loop linearly in all cases. If the loop header
612 // is just a normal basic block in the loop, we want to look for what block
613 // within the loop is the best one to layout at the top. However, if the loop
614 // header has be pre-merged into a chain due to predecessors not having
615 // analyzable branches, *and* the predecessor it is merged with is *not* part
616 // of the loop, rotating the header into the middle of the loop will create
617 // a non-contiguous range of blocks which is Very Bad. So start with the
618 // header and only rotate if safe.
619 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
620 if (!LoopBlockSet.count(*HeaderChain.begin()))
623 BlockFrequency BestExitEdgeFreq;
624 unsigned BestExitLoopDepth = 0;
625 MachineBasicBlock *ExitingBB = nullptr;
626 // If there are exits to outer loops, loop rotation can severely limit
627 // fallthrough opportunites unless it selects such an exit. Keep a set of
628 // blocks where rotating to exit with that block will reach an outer loop.
629 SmallPtrSet<MachineBasicBlock *, 4> BlocksExitingToOuterLoop;
631 DEBUG(dbgs() << "Finding best loop exit for: "
632 << getBlockName(L.getHeader()) << "\n");
633 for (MachineLoop::block_iterator I = L.block_begin(),
636 BlockChain &Chain = *BlockToChain[*I];
637 // Ensure that this block is at the end of a chain; otherwise it could be
638 // mid-way through an inner loop or a successor of an analyzable branch.
639 if (*I != *std::prev(Chain.end()))
642 // Now walk the successors. We need to establish whether this has a viable
643 // exiting successor and whether it has a viable non-exiting successor.
644 // We store the old exiting state and restore it if a viable looping
645 // successor isn't found.
646 MachineBasicBlock *OldExitingBB = ExitingBB;
647 BlockFrequency OldBestExitEdgeFreq = BestExitEdgeFreq;
648 bool HasLoopingSucc = false;
649 // FIXME: Due to the performance of the probability and weight routines in
650 // the MBPI analysis, we use the internal weights and manually compute the
651 // probabilities to avoid quadratic behavior.
652 uint32_t WeightScale = 0;
653 uint32_t SumWeight = MBPI->getSumForBlock(*I, WeightScale);
654 for (MachineBasicBlock::succ_iterator SI = (*I)->succ_begin(),
655 SE = (*I)->succ_end();
657 if ((*SI)->isLandingPad())
661 BlockChain &SuccChain = *BlockToChain[*SI];
662 // Don't split chains, either this chain or the successor's chain.
663 if (&Chain == &SuccChain) {
664 DEBUG(dbgs() << " exiting: " << getBlockName(*I) << " -> "
665 << getBlockName(*SI) << " (chain conflict)\n");
669 uint32_t SuccWeight = MBPI->getEdgeWeight(*I, *SI);
670 if (LoopBlockSet.count(*SI)) {
671 DEBUG(dbgs() << " looping: " << getBlockName(*I) << " -> "
672 << getBlockName(*SI) << " (" << SuccWeight << ")\n");
673 HasLoopingSucc = true;
677 unsigned SuccLoopDepth = 0;
678 if (MachineLoop *ExitLoop = MLI->getLoopFor(*SI)) {
679 SuccLoopDepth = ExitLoop->getLoopDepth();
680 if (ExitLoop->contains(&L))
681 BlocksExitingToOuterLoop.insert(*I);
684 BranchProbability SuccProb(SuccWeight / WeightScale, SumWeight);
685 BlockFrequency ExitEdgeFreq = MBFI->getBlockFreq(*I) * SuccProb;
686 DEBUG(dbgs() << " exiting: " << getBlockName(*I) << " -> "
687 << getBlockName(*SI) << " [L:" << SuccLoopDepth
689 MBFI->printBlockFreq(dbgs(), ExitEdgeFreq) << ")\n");
690 // Note that we bias this toward an existing layout successor to retain
691 // incoming order in the absence of better information. The exit must have
692 // a frequency higher than the current exit before we consider breaking
694 BranchProbability Bias(100 - ExitBlockBias, 100);
695 if (!ExitingBB || BestExitLoopDepth < SuccLoopDepth ||
696 ExitEdgeFreq > BestExitEdgeFreq ||
697 ((*I)->isLayoutSuccessor(*SI) &&
698 !(ExitEdgeFreq < BestExitEdgeFreq * Bias))) {
699 BestExitEdgeFreq = ExitEdgeFreq;
704 // Restore the old exiting state, no viable looping successor was found.
705 if (!HasLoopingSucc) {
706 ExitingBB = OldExitingBB;
707 BestExitEdgeFreq = OldBestExitEdgeFreq;
711 // Without a candidate exiting block or with only a single block in the
712 // loop, just use the loop header to layout the loop.
713 if (!ExitingBB || L.getNumBlocks() == 1)
716 // Also, if we have exit blocks which lead to outer loops but didn't select
717 // one of them as the exiting block we are rotating toward, disable loop
718 // rotation altogether.
719 if (!BlocksExitingToOuterLoop.empty() &&
720 !BlocksExitingToOuterLoop.count(ExitingBB))
723 DEBUG(dbgs() << " Best exiting block: " << getBlockName(ExitingBB) << "\n");
727 /// \brief Attempt to rotate an exiting block to the bottom of the loop.
729 /// Once we have built a chain, try to rotate it to line up the hot exit block
730 /// with fallthrough out of the loop if doing so doesn't introduce unnecessary
731 /// branches. For example, if the loop has fallthrough into its header and out
732 /// of its bottom already, don't rotate it.
733 void MachineBlockPlacement::rotateLoop(BlockChain &LoopChain,
734 MachineBasicBlock *ExitingBB,
735 const BlockFilterSet &LoopBlockSet) {
739 MachineBasicBlock *Top = *LoopChain.begin();
740 bool ViableTopFallthrough = false;
741 for (MachineBasicBlock::pred_iterator PI = Top->pred_begin(),
742 PE = Top->pred_end();
744 BlockChain *PredChain = BlockToChain[*PI];
745 if (!LoopBlockSet.count(*PI) &&
746 (!PredChain || *PI == *std::prev(PredChain->end()))) {
747 ViableTopFallthrough = true;
752 // If the header has viable fallthrough, check whether the current loop
753 // bottom is a viable exiting block. If so, bail out as rotating will
754 // introduce an unnecessary branch.
755 if (ViableTopFallthrough) {
756 MachineBasicBlock *Bottom = *std::prev(LoopChain.end());
757 for (MachineBasicBlock::succ_iterator SI = Bottom->succ_begin(),
758 SE = Bottom->succ_end();
760 BlockChain *SuccChain = BlockToChain[*SI];
761 if (!LoopBlockSet.count(*SI) &&
762 (!SuccChain || *SI == *SuccChain->begin()))
767 BlockChain::iterator ExitIt = std::find(LoopChain.begin(), LoopChain.end(),
769 if (ExitIt == LoopChain.end())
772 std::rotate(LoopChain.begin(), std::next(ExitIt), LoopChain.end());
775 /// \brief Forms basic block chains from the natural loop structures.
777 /// These chains are designed to preserve the existing *structure* of the code
778 /// as much as possible. We can then stitch the chains together in a way which
779 /// both preserves the topological structure and minimizes taken conditional
781 void MachineBlockPlacement::buildLoopChains(MachineFunction &F,
783 // First recurse through any nested loops, building chains for those inner
785 for (MachineLoop::iterator LI = L.begin(), LE = L.end(); LI != LE; ++LI)
786 buildLoopChains(F, **LI);
788 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
789 BlockFilterSet LoopBlockSet(L.block_begin(), L.block_end());
791 // First check to see if there is an obviously preferable top block for the
792 // loop. This will default to the header, but may end up as one of the
793 // predecessors to the header if there is one which will result in strictly
794 // fewer branches in the loop body.
795 MachineBasicBlock *LoopTop = findBestLoopTop(L, LoopBlockSet);
797 // If we selected just the header for the loop top, look for a potentially
798 // profitable exit block in the event that rotating the loop can eliminate
799 // branches by placing an exit edge at the bottom.
800 MachineBasicBlock *ExitingBB = nullptr;
801 if (LoopTop == L.getHeader())
802 ExitingBB = findBestLoopExit(F, L, LoopBlockSet);
804 BlockChain &LoopChain = *BlockToChain[LoopTop];
806 // FIXME: This is a really lame way of walking the chains in the loop: we
807 // walk the blocks, and use a set to prevent visiting a particular chain
809 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
810 assert(LoopChain.LoopPredecessors == 0);
811 UpdatedPreds.insert(&LoopChain);
812 for (MachineLoop::block_iterator BI = L.block_begin(),
815 BlockChain &Chain = *BlockToChain[*BI];
816 if (!UpdatedPreds.insert(&Chain))
819 assert(Chain.LoopPredecessors == 0);
820 for (BlockChain::iterator BCI = Chain.begin(), BCE = Chain.end();
822 assert(BlockToChain[*BCI] == &Chain);
823 for (MachineBasicBlock::pred_iterator PI = (*BCI)->pred_begin(),
824 PE = (*BCI)->pred_end();
826 if (BlockToChain[*PI] == &Chain || !LoopBlockSet.count(*PI))
828 ++Chain.LoopPredecessors;
832 if (Chain.LoopPredecessors == 0)
833 BlockWorkList.push_back(*Chain.begin());
836 buildChain(LoopTop, LoopChain, BlockWorkList, &LoopBlockSet);
837 rotateLoop(LoopChain, ExitingBB, LoopBlockSet);
840 // Crash at the end so we get all of the debugging output first.
841 bool BadLoop = false;
842 if (LoopChain.LoopPredecessors) {
844 dbgs() << "Loop chain contains a block without its preds placed!\n"
845 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
846 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n";
848 for (BlockChain::iterator BCI = LoopChain.begin(), BCE = LoopChain.end();
850 dbgs() << " ... " << getBlockName(*BCI) << "\n";
851 if (!LoopBlockSet.erase(*BCI)) {
852 // We don't mark the loop as bad here because there are real situations
853 // where this can occur. For example, with an unanalyzable fallthrough
854 // from a loop block to a non-loop block or vice versa.
855 dbgs() << "Loop chain contains a block not contained by the loop!\n"
856 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
857 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
858 << " Bad block: " << getBlockName(*BCI) << "\n";
862 if (!LoopBlockSet.empty()) {
864 for (BlockFilterSet::iterator LBI = LoopBlockSet.begin(),
865 LBE = LoopBlockSet.end();
867 dbgs() << "Loop contains blocks never placed into a chain!\n"
868 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
869 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
870 << " Bad block: " << getBlockName(*LBI) << "\n";
872 assert(!BadLoop && "Detected problems with the placement of this loop.");
876 void MachineBlockPlacement::buildCFGChains(MachineFunction &F) {
877 // Ensure that every BB in the function has an associated chain to simplify
878 // the assumptions of the remaining algorithm.
879 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
880 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
881 MachineBasicBlock *BB = FI;
883 = new (ChainAllocator.Allocate()) BlockChain(BlockToChain, BB);
884 // Also, merge any blocks which we cannot reason about and must preserve
885 // the exact fallthrough behavior for.
888 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
889 if (!TII->AnalyzeBranch(*BB, TBB, FBB, Cond) || !FI->canFallThrough())
892 MachineFunction::iterator NextFI(std::next(FI));
893 MachineBasicBlock *NextBB = NextFI;
894 // Ensure that the layout successor is a viable block, as we know that
895 // fallthrough is a possibility.
896 assert(NextFI != FE && "Can't fallthrough past the last block.");
897 DEBUG(dbgs() << "Pre-merging due to unanalyzable fallthrough: "
898 << getBlockName(BB) << " -> " << getBlockName(NextBB)
900 Chain->merge(NextBB, nullptr);
906 // Build any loop-based chains.
907 for (MachineLoopInfo::iterator LI = MLI->begin(), LE = MLI->end(); LI != LE;
909 buildLoopChains(F, **LI);
911 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
913 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
914 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
915 MachineBasicBlock *BB = &*FI;
916 BlockChain &Chain = *BlockToChain[BB];
917 if (!UpdatedPreds.insert(&Chain))
920 assert(Chain.LoopPredecessors == 0);
921 for (BlockChain::iterator BCI = Chain.begin(), BCE = Chain.end();
923 assert(BlockToChain[*BCI] == &Chain);
924 for (MachineBasicBlock::pred_iterator PI = (*BCI)->pred_begin(),
925 PE = (*BCI)->pred_end();
927 if (BlockToChain[*PI] == &Chain)
929 ++Chain.LoopPredecessors;
933 if (Chain.LoopPredecessors == 0)
934 BlockWorkList.push_back(*Chain.begin());
937 BlockChain &FunctionChain = *BlockToChain[&F.front()];
938 buildChain(&F.front(), FunctionChain, BlockWorkList);
941 typedef SmallPtrSet<MachineBasicBlock *, 16> FunctionBlockSetType;
944 // Crash at the end so we get all of the debugging output first.
945 bool BadFunc = false;
946 FunctionBlockSetType FunctionBlockSet;
947 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
948 FunctionBlockSet.insert(FI);
950 for (BlockChain::iterator BCI = FunctionChain.begin(),
951 BCE = FunctionChain.end();
953 if (!FunctionBlockSet.erase(*BCI)) {
955 dbgs() << "Function chain contains a block not in the function!\n"
956 << " Bad block: " << getBlockName(*BCI) << "\n";
959 if (!FunctionBlockSet.empty()) {
961 for (FunctionBlockSetType::iterator FBI = FunctionBlockSet.begin(),
962 FBE = FunctionBlockSet.end();
964 dbgs() << "Function contains blocks never placed into a chain!\n"
965 << " Bad block: " << getBlockName(*FBI) << "\n";
967 assert(!BadFunc && "Detected problems with the block placement.");
970 // Splice the blocks into place.
971 MachineFunction::iterator InsertPos = F.begin();
972 for (BlockChain::iterator BI = FunctionChain.begin(),
973 BE = FunctionChain.end();
975 DEBUG(dbgs() << (BI == FunctionChain.begin() ? "Placing chain "
977 << getBlockName(*BI) << "\n");
978 if (InsertPos != MachineFunction::iterator(*BI))
979 F.splice(InsertPos, *BI);
983 // Update the terminator of the previous block.
984 if (BI == FunctionChain.begin())
986 MachineBasicBlock *PrevBB = std::prev(MachineFunction::iterator(*BI));
988 // FIXME: It would be awesome of updateTerminator would just return rather
989 // than assert when the branch cannot be analyzed in order to remove this
992 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
993 if (!TII->AnalyzeBranch(*PrevBB, TBB, FBB, Cond)) {
994 // The "PrevBB" is not yet updated to reflect current code layout, so,
995 // o. it may fall-through to a block without explict "goto" instruction
996 // before layout, and no longer fall-through it after layout; or
999 // AnalyzeBranch() may return erroneous value for FBB when these two
1000 // situations take place. For the first scenario FBB is mistakenly set
1001 // NULL; for the 2nd scenario, the FBB, which is expected to be NULL,
1002 // is mistakenly pointing to "*BI".
1004 bool needUpdateBr = true;
1005 if (!Cond.empty() && (!FBB || FBB == *BI)) {
1006 PrevBB->updateTerminator();
1007 needUpdateBr = false;
1009 TBB = FBB = nullptr;
1010 if (TII->AnalyzeBranch(*PrevBB, TBB, FBB, Cond)) {
1011 // FIXME: This should never take place.
1012 TBB = FBB = nullptr;
1016 // If PrevBB has a two-way branch, try to re-order the branches
1017 // such that we branch to the successor with higher weight first.
1018 if (TBB && !Cond.empty() && FBB &&
1019 MBPI->getEdgeWeight(PrevBB, FBB) > MBPI->getEdgeWeight(PrevBB, TBB) &&
1020 !TII->ReverseBranchCondition(Cond)) {
1021 DEBUG(dbgs() << "Reverse order of the two branches: "
1022 << getBlockName(PrevBB) << "\n");
1023 DEBUG(dbgs() << " Edge weight: " << MBPI->getEdgeWeight(PrevBB, FBB)
1024 << " vs " << MBPI->getEdgeWeight(PrevBB, TBB) << "\n");
1025 DebugLoc dl; // FIXME: this is nowhere
1026 TII->RemoveBranch(*PrevBB);
1027 TII->InsertBranch(*PrevBB, FBB, TBB, Cond, dl);
1028 needUpdateBr = true;
1031 PrevBB->updateTerminator();
1035 // Fixup the last block.
1037 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
1038 if (!TII->AnalyzeBranch(F.back(), TBB, FBB, Cond))
1039 F.back().updateTerminator();
1041 // Walk through the backedges of the function now that we have fully laid out
1042 // the basic blocks and align the destination of each backedge. We don't rely
1043 // exclusively on the loop info here so that we can align backedges in
1044 // unnatural CFGs and backedges that were introduced purely because of the
1045 // loop rotations done during this layout pass.
1046 if (F.getFunction()->getAttributes().
1047 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize))
1049 unsigned Align = TLI->getPrefLoopAlignment();
1051 return; // Don't care about loop alignment.
1052 if (FunctionChain.begin() == FunctionChain.end())
1053 return; // Empty chain.
1055 const BranchProbability ColdProb(1, 5); // 20%
1056 BlockFrequency EntryFreq = MBFI->getBlockFreq(F.begin());
1057 BlockFrequency WeightedEntryFreq = EntryFreq * ColdProb;
1058 for (BlockChain::iterator BI = std::next(FunctionChain.begin()),
1059 BE = FunctionChain.end();
1061 // Don't align non-looping basic blocks. These are unlikely to execute
1062 // enough times to matter in practice. Note that we'll still handle
1063 // unnatural CFGs inside of a natural outer loop (the common case) and
1065 MachineLoop *L = MLI->getLoopFor(*BI);
1069 // If the block is cold relative to the function entry don't waste space
1071 BlockFrequency Freq = MBFI->getBlockFreq(*BI);
1072 if (Freq < WeightedEntryFreq)
1075 // If the block is cold relative to its loop header, don't align it
1076 // regardless of what edges into the block exist.
1077 MachineBasicBlock *LoopHeader = L->getHeader();
1078 BlockFrequency LoopHeaderFreq = MBFI->getBlockFreq(LoopHeader);
1079 if (Freq < (LoopHeaderFreq * ColdProb))
1082 // Check for the existence of a non-layout predecessor which would benefit
1083 // from aligning this block.
1084 MachineBasicBlock *LayoutPred = *std::prev(BI);
1086 // Force alignment if all the predecessors are jumps. We already checked
1087 // that the block isn't cold above.
1088 if (!LayoutPred->isSuccessor(*BI)) {
1089 (*BI)->setAlignment(Align);
1093 // Align this block if the layout predecessor's edge into this block is
1094 // cold relative to the block. When this is true, other predecessors make up
1095 // all of the hot entries into the block and thus alignment is likely to be
1097 BranchProbability LayoutProb = MBPI->getEdgeProbability(LayoutPred, *BI);
1098 BlockFrequency LayoutEdgeFreq = MBFI->getBlockFreq(LayoutPred) * LayoutProb;
1099 if (LayoutEdgeFreq <= (Freq * ColdProb))
1100 (*BI)->setAlignment(Align);
1104 bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &F) {
1105 // Check for single-block functions and skip them.
1106 if (std::next(F.begin()) == F.end())
1109 if (skipOptnoneFunction(*F.getFunction()))
1112 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
1113 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
1114 MLI = &getAnalysis<MachineLoopInfo>();
1115 TII = F.getSubtarget().getInstrInfo();
1116 TLI = F.getSubtarget().getTargetLowering();
1117 assert(BlockToChain.empty());
1121 BlockToChain.clear();
1122 ChainAllocator.DestroyAll();
1125 // Align all of the blocks in the function to a specific alignment.
1126 for (MachineFunction::iterator FI = F.begin(), FE = F.end();
1128 FI->setAlignment(AlignAllBlock);
1130 // We always return true as we have no way to track whether the final order
1131 // differs from the original order.
1136 /// \brief A pass to compute block placement statistics.
1138 /// A separate pass to compute interesting statistics for evaluating block
1139 /// placement. This is separate from the actual placement pass so that they can
1140 /// be computed in the absence of any placement transformations or when using
1141 /// alternative placement strategies.
1142 class MachineBlockPlacementStats : public MachineFunctionPass {
1143 /// \brief A handle to the branch probability pass.
1144 const MachineBranchProbabilityInfo *MBPI;
1146 /// \brief A handle to the function-wide block frequency pass.
1147 const MachineBlockFrequencyInfo *MBFI;
1150 static char ID; // Pass identification, replacement for typeid
1151 MachineBlockPlacementStats() : MachineFunctionPass(ID) {
1152 initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry());
1155 bool runOnMachineFunction(MachineFunction &F) override;
1157 void getAnalysisUsage(AnalysisUsage &AU) const override {
1158 AU.addRequired<MachineBranchProbabilityInfo>();
1159 AU.addRequired<MachineBlockFrequencyInfo>();
1160 AU.setPreservesAll();
1161 MachineFunctionPass::getAnalysisUsage(AU);
1166 char MachineBlockPlacementStats::ID = 0;
1167 char &llvm::MachineBlockPlacementStatsID = MachineBlockPlacementStats::ID;
1168 INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats",
1169 "Basic Block Placement Stats", false, false)
1170 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
1171 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
1172 INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats",
1173 "Basic Block Placement Stats", false, false)
1175 bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) {
1176 // Check for single-block functions and skip them.
1177 if (std::next(F.begin()) == F.end())
1180 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
1181 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
1183 for (MachineFunction::iterator I = F.begin(), E = F.end(); I != E; ++I) {
1184 BlockFrequency BlockFreq = MBFI->getBlockFreq(I);
1185 Statistic &NumBranches = (I->succ_size() > 1) ? NumCondBranches
1186 : NumUncondBranches;
1187 Statistic &BranchTakenFreq = (I->succ_size() > 1) ? CondBranchTakenFreq
1188 : UncondBranchTakenFreq;
1189 for (MachineBasicBlock::succ_iterator SI = I->succ_begin(),
1192 // Skip if this successor is a fallthrough.
1193 if (I->isLayoutSuccessor(*SI))
1196 BlockFrequency EdgeFreq = BlockFreq * MBPI->getEdgeProbability(I, *SI);
1198 BranchTakenFreq += EdgeFreq.getFrequency();