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 absense 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 #define DEBUG_TYPE "block-placement2"
29 #include "llvm/CodeGen/MachineBasicBlock.h"
30 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
31 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineFunctionPass.h"
34 #include "llvm/CodeGen/MachineLoopInfo.h"
35 #include "llvm/CodeGen/MachineModuleInfo.h"
36 #include "llvm/CodeGen/Passes.h"
37 #include "llvm/Support/Allocator.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/ADT/DenseMap.h"
40 #include "llvm/ADT/SmallPtrSet.h"
41 #include "llvm/ADT/SmallVector.h"
42 #include "llvm/ADT/Statistic.h"
43 #include "llvm/Target/TargetInstrInfo.h"
44 #include "llvm/Target/TargetLowering.h"
48 STATISTIC(NumCondBranches, "Number of conditional branches");
49 STATISTIC(NumUncondBranches, "Number of uncondittional branches");
50 STATISTIC(CondBranchTakenFreq,
51 "Potential frequency of taking conditional branches");
52 STATISTIC(UncondBranchTakenFreq,
53 "Potential frequency of taking unconditional branches");
57 /// \brief Type for our function-wide basic block -> block chain mapping.
58 typedef DenseMap<MachineBasicBlock *, BlockChain *> BlockToChainMapType;
62 /// \brief A chain of blocks which will be laid out contiguously.
64 /// This is the datastructure representing a chain of consecutive blocks that
65 /// are profitable to layout together in order to maximize fallthrough
66 /// probabilities. We also can use a block chain to represent a sequence of
67 /// basic blocks which have some external (correctness) requirement for
68 /// sequential layout.
70 /// Eventually, the block chains will form a directed graph over the function.
71 /// We provide an SCC-supporting-iterator in order to quicky build and walk the
72 /// SCCs of block chains within a function.
74 /// The block chains also have support for calculating and caching probability
75 /// information related to the chain itself versus other chains. This is used
76 /// for ranking during the final layout of block chains.
78 /// \brief The sequence of blocks belonging to this chain.
80 /// This is the sequence of blocks for a particular chain. These will be laid
81 /// out in-order within the function.
82 SmallVector<MachineBasicBlock *, 4> Blocks;
84 /// \brief A handle to the function-wide basic block to block chain mapping.
86 /// This is retained in each block chain to simplify the computation of child
87 /// block chains for SCC-formation and iteration. We store the edges to child
88 /// basic blocks, and map them back to their associated chains using this
90 BlockToChainMapType &BlockToChain;
93 /// \brief Construct a new BlockChain.
95 /// This builds a new block chain representing a single basic block in the
96 /// function. It also registers itself as the chain that block participates
97 /// in with the BlockToChain mapping.
98 BlockChain(BlockToChainMapType &BlockToChain, MachineBasicBlock *BB)
99 : Blocks(1, BB), BlockToChain(BlockToChain), LoopPredecessors(0) {
100 assert(BB && "Cannot create a chain with a null basic block");
101 BlockToChain[BB] = this;
104 /// \brief Iterator over blocks within the chain.
105 typedef SmallVectorImpl<MachineBasicBlock *>::const_iterator iterator;
107 /// \brief Beginning of blocks within the chain.
108 iterator begin() const { return Blocks.begin(); }
110 /// \brief End of blocks within the chain.
111 iterator end() const { return Blocks.end(); }
113 /// \brief Merge a block chain into this one.
115 /// This routine merges a block chain into this one. It takes care of forming
116 /// a contiguous sequence of basic blocks, updating the edge list, and
117 /// updating the block -> chain mapping. It does not free or tear down the
118 /// old chain, but the old chain's block list is no longer valid.
119 void merge(MachineBasicBlock *BB, BlockChain *Chain) {
121 assert(!Blocks.empty());
123 // Fast path in case we don't have a chain already.
125 assert(!BlockToChain[BB]);
126 Blocks.push_back(BB);
127 BlockToChain[BB] = this;
131 assert(BB == *Chain->begin());
132 assert(Chain->begin() != Chain->end());
134 // Update the incoming blocks to point to this chain, and add them to the
136 for (BlockChain::iterator BI = Chain->begin(), BE = Chain->end();
138 Blocks.push_back(*BI);
139 assert(BlockToChain[*BI] == Chain && "Incoming blocks not in chain");
140 BlockToChain[*BI] = this;
144 /// \brief Count of predecessors within the loop currently being processed.
146 /// This count is updated at each loop we process to represent the number of
147 /// in-loop predecessors of this chain.
148 unsigned LoopPredecessors;
153 class MachineBlockPlacement : public MachineFunctionPass {
154 /// \brief A typedef for a block filter set.
155 typedef SmallPtrSet<MachineBasicBlock *, 16> BlockFilterSet;
157 /// \brief A handle to the branch probability pass.
158 const MachineBranchProbabilityInfo *MBPI;
160 /// \brief A handle to the function-wide block frequency pass.
161 const MachineBlockFrequencyInfo *MBFI;
163 /// \brief A handle to the loop info.
164 const MachineLoopInfo *MLI;
166 /// \brief A handle to the target's instruction info.
167 const TargetInstrInfo *TII;
169 /// \brief A handle to the target's lowering info.
170 const TargetLowering *TLI;
172 /// \brief Allocator and owner of BlockChain structures.
174 /// We build BlockChains lazily by merging together high probability BB
175 /// sequences acording to the "Algo2" in the paper mentioned at the top of
176 /// the file. To reduce malloc traffic, we allocate them using this slab-like
177 /// allocator, and destroy them after the pass completes.
178 SpecificBumpPtrAllocator<BlockChain> ChainAllocator;
180 /// \brief Function wide BasicBlock to BlockChain mapping.
182 /// This mapping allows efficiently moving from any given basic block to the
183 /// BlockChain it participates in, if any. We use it to, among other things,
184 /// allow implicitly defining edges between chains as the existing edges
185 /// between basic blocks.
186 DenseMap<MachineBasicBlock *, BlockChain *> BlockToChain;
188 void markChainSuccessors(BlockChain &Chain,
189 MachineBasicBlock *LoopHeaderBB,
190 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
191 const BlockFilterSet *BlockFilter = 0);
192 MachineBasicBlock *selectBestSuccessor(MachineBasicBlock *BB,
194 const BlockFilterSet *BlockFilter);
195 MachineBasicBlock *selectBestCandidateBlock(
196 BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList,
197 const BlockFilterSet *BlockFilter);
198 MachineBasicBlock *getFirstUnplacedBlock(
200 const BlockChain &PlacedChain,
201 MachineFunction::iterator &PrevUnplacedBlockIt,
202 const BlockFilterSet *BlockFilter);
203 void buildChain(MachineBasicBlock *BB, BlockChain &Chain,
204 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
205 const BlockFilterSet *BlockFilter = 0);
206 MachineBasicBlock *findBestLoopTop(MachineFunction &F,
208 const BlockFilterSet &LoopBlockSet);
209 void buildLoopChains(MachineFunction &F, MachineLoop &L);
210 void buildCFGChains(MachineFunction &F);
211 void AlignLoops(MachineFunction &F);
214 static char ID; // Pass identification, replacement for typeid
215 MachineBlockPlacement() : MachineFunctionPass(ID) {
216 initializeMachineBlockPlacementPass(*PassRegistry::getPassRegistry());
219 bool runOnMachineFunction(MachineFunction &F);
221 void getAnalysisUsage(AnalysisUsage &AU) const {
222 AU.addRequired<MachineBranchProbabilityInfo>();
223 AU.addRequired<MachineBlockFrequencyInfo>();
224 AU.addRequired<MachineLoopInfo>();
225 MachineFunctionPass::getAnalysisUsage(AU);
228 const char *getPassName() const { return "Block Placement"; }
232 char MachineBlockPlacement::ID = 0;
233 INITIALIZE_PASS_BEGIN(MachineBlockPlacement, "block-placement2",
234 "Branch Probability Basic Block Placement", false, false)
235 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
236 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
237 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
238 INITIALIZE_PASS_END(MachineBlockPlacement, "block-placement2",
239 "Branch Probability Basic Block Placement", false, false)
241 FunctionPass *llvm::createMachineBlockPlacementPass() {
242 return new MachineBlockPlacement();
246 /// \brief Helper to print the name of a MBB.
248 /// Only used by debug logging.
249 static std::string getBlockName(MachineBasicBlock *BB) {
251 raw_string_ostream OS(Result);
252 OS << "BB#" << BB->getNumber()
253 << " (derived from LLVM BB '" << BB->getName() << "')";
258 /// \brief Helper to print the number of a MBB.
260 /// Only used by debug logging.
261 static std::string getBlockNum(MachineBasicBlock *BB) {
263 raw_string_ostream OS(Result);
264 OS << "BB#" << BB->getNumber();
270 /// \brief Mark a chain's successors as having one fewer preds.
272 /// When a chain is being merged into the "placed" chain, this routine will
273 /// quickly walk the successors of each block in the chain and mark them as
274 /// having one fewer active predecessor. It also adds any successors of this
275 /// chain which reach the zero-predecessor state to the worklist passed in.
276 void MachineBlockPlacement::markChainSuccessors(
278 MachineBasicBlock *LoopHeaderBB,
279 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
280 const BlockFilterSet *BlockFilter) {
281 // Walk all the blocks in this chain, marking their successors as having
282 // a predecessor placed.
283 for (BlockChain::iterator CBI = Chain.begin(), CBE = Chain.end();
285 // Add any successors for which this is the only un-placed in-loop
286 // predecessor to the worklist as a viable candidate for CFG-neutral
287 // placement. No subsequent placement of this block will violate the CFG
288 // shape, so we get to use heuristics to choose a favorable placement.
289 for (MachineBasicBlock::succ_iterator SI = (*CBI)->succ_begin(),
290 SE = (*CBI)->succ_end();
292 if (BlockFilter && !BlockFilter->count(*SI))
294 BlockChain &SuccChain = *BlockToChain[*SI];
295 // Disregard edges within a fixed chain, or edges to the loop header.
296 if (&Chain == &SuccChain || *SI == LoopHeaderBB)
299 // This is a cross-chain edge that is within the loop, so decrement the
300 // loop predecessor count of the destination chain.
301 if (SuccChain.LoopPredecessors > 0 && --SuccChain.LoopPredecessors == 0)
302 BlockWorkList.push_back(*SuccChain.begin());
307 /// \brief Select the best successor for a block.
309 /// This looks across all successors of a particular block and attempts to
310 /// select the "best" one to be the layout successor. It only considers direct
311 /// successors which also pass the block filter. It will attempt to avoid
312 /// breaking CFG structure, but cave and break such structures in the case of
313 /// very hot successor edges.
315 /// \returns The best successor block found, or null if none are viable.
316 MachineBasicBlock *MachineBlockPlacement::selectBestSuccessor(
317 MachineBasicBlock *BB, BlockChain &Chain,
318 const BlockFilterSet *BlockFilter) {
319 const BranchProbability HotProb(4, 5); // 80%
321 MachineBasicBlock *BestSucc = 0;
322 // FIXME: Due to the performance of the probability and weight routines in
323 // the MBPI analysis, we manually compute probabilities using the edge
324 // weights. This is suboptimal as it means that the somewhat subtle
325 // definition of edge weight semantics is encoded here as well. We should
326 // improve the MBPI interface to effeciently support query patterns such as
328 uint32_t BestWeight = 0;
329 uint32_t WeightScale = 0;
330 uint32_t SumWeight = MBPI->getSumForBlock(BB, WeightScale);
331 DEBUG(dbgs() << "Attempting merge from: " << getBlockName(BB) << "\n");
332 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
335 if (BlockFilter && !BlockFilter->count(*SI))
337 BlockChain &SuccChain = *BlockToChain[*SI];
338 if (&SuccChain == &Chain) {
339 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> Already merged!\n");
342 if (*SI != *SuccChain.begin()) {
343 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> Mid chain!\n");
347 uint32_t SuccWeight = MBPI->getEdgeWeight(BB, *SI);
348 BranchProbability SuccProb(SuccWeight / WeightScale, SumWeight);
350 // Only consider successors which are either "hot", or wouldn't violate
351 // any CFG constraints.
352 if (SuccChain.LoopPredecessors != 0) {
353 if (SuccProb < HotProb) {
354 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> CFG conflict\n");
358 // Make sure that a hot successor doesn't have a globally more important
360 BlockFrequency CandidateEdgeFreq
361 = MBFI->getBlockFreq(BB) * SuccProb * HotProb.getCompl();
362 bool BadCFGConflict = false;
363 for (MachineBasicBlock::pred_iterator PI = (*SI)->pred_begin(),
364 PE = (*SI)->pred_end();
366 if (*PI == *SI || (BlockFilter && !BlockFilter->count(*PI)) ||
367 BlockToChain[*PI] == &Chain)
369 BlockFrequency PredEdgeFreq
370 = MBFI->getBlockFreq(*PI) * MBPI->getEdgeProbability(*PI, *SI);
371 if (PredEdgeFreq >= CandidateEdgeFreq) {
372 BadCFGConflict = true;
376 if (BadCFGConflict) {
377 DEBUG(dbgs() << " " << getBlockName(*SI)
378 << " -> non-cold CFG conflict\n");
383 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> " << SuccProb
385 << (SuccChain.LoopPredecessors != 0 ? " (CFG break)" : "")
387 if (BestSucc && BestWeight >= SuccWeight)
390 BestWeight = SuccWeight;
396 /// \brief Predicate struct to detect blocks already placed.
397 class IsBlockPlaced {
398 const BlockChain &PlacedChain;
399 const BlockToChainMapType &BlockToChain;
402 IsBlockPlaced(const BlockChain &PlacedChain,
403 const BlockToChainMapType &BlockToChain)
404 : PlacedChain(PlacedChain), BlockToChain(BlockToChain) {}
406 bool operator()(MachineBasicBlock *BB) const {
407 return BlockToChain.lookup(BB) == &PlacedChain;
412 /// \brief Select the best block from a worklist.
414 /// This looks through the provided worklist as a list of candidate basic
415 /// blocks and select the most profitable one to place. The definition of
416 /// profitable only really makes sense in the context of a loop. This returns
417 /// the most frequently visited block in the worklist, which in the case of
418 /// a loop, is the one most desirable to be physically close to the rest of the
419 /// loop body in order to improve icache behavior.
421 /// \returns The best block found, or null if none are viable.
422 MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock(
423 BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList,
424 const BlockFilterSet *BlockFilter) {
425 // Once we need to walk the worklist looking for a candidate, cleanup the
426 // worklist of already placed entries.
427 // FIXME: If this shows up on profiles, it could be folded (at the cost of
428 // some code complexity) into the loop below.
429 WorkList.erase(std::remove_if(WorkList.begin(), WorkList.end(),
430 IsBlockPlaced(Chain, BlockToChain)),
433 MachineBasicBlock *BestBlock = 0;
434 BlockFrequency BestFreq;
435 for (SmallVectorImpl<MachineBasicBlock *>::iterator WBI = WorkList.begin(),
436 WBE = WorkList.end();
438 assert(!BlockFilter || BlockFilter->count(*WBI));
439 BlockChain &SuccChain = *BlockToChain[*WBI];
440 if (&SuccChain == &Chain) {
441 DEBUG(dbgs() << " " << getBlockName(*WBI)
442 << " -> Already merged!\n");
445 assert(SuccChain.LoopPredecessors == 0 && "Found CFG-violating block");
447 BlockFrequency CandidateFreq = MBFI->getBlockFreq(*WBI);
448 DEBUG(dbgs() << " " << getBlockName(*WBI) << " -> " << CandidateFreq
450 if (BestBlock && BestFreq >= CandidateFreq)
453 BestFreq = CandidateFreq;
458 /// \brief Retrieve the first unplaced basic block.
460 /// This routine is called when we are unable to use the CFG to walk through
461 /// all of the basic blocks and form a chain due to unnatural loops in the CFG.
462 /// We walk through the function's blocks in order, starting from the
463 /// LastUnplacedBlockIt. We update this iterator on each call to avoid
464 /// re-scanning the entire sequence on repeated calls to this routine.
465 MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock(
466 MachineFunction &F, const BlockChain &PlacedChain,
467 MachineFunction::iterator &PrevUnplacedBlockIt,
468 const BlockFilterSet *BlockFilter) {
469 for (MachineFunction::iterator I = PrevUnplacedBlockIt, E = F.end(); I != E;
471 if (BlockFilter && !BlockFilter->count(I))
473 if (BlockToChain[I] != &PlacedChain) {
474 PrevUnplacedBlockIt = I;
475 // Now select the head of the chain to which the unplaced block belongs
476 // as the block to place. This will force the entire chain to be placed,
477 // and satisfies the requirements of merging chains.
478 return *BlockToChain[I]->begin();
484 void MachineBlockPlacement::buildChain(
485 MachineBasicBlock *BB,
487 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
488 const BlockFilterSet *BlockFilter) {
490 assert(BlockToChain[BB] == &Chain);
491 MachineFunction &F = *BB->getParent();
492 MachineFunction::iterator PrevUnplacedBlockIt = F.begin();
494 MachineBasicBlock *LoopHeaderBB = BB;
495 markChainSuccessors(Chain, LoopHeaderBB, BlockWorkList, BlockFilter);
496 BB = *llvm::prior(Chain.end());
499 assert(BlockToChain[BB] == &Chain);
500 assert(*llvm::prior(Chain.end()) == BB);
501 MachineBasicBlock *BestSucc = 0;
503 // Look for the best viable successor if there is one to place immediately
505 BestSucc = selectBestSuccessor(BB, Chain, BlockFilter);
507 // If an immediate successor isn't available, look for the best viable
508 // block among those we've identified as not violating the loop's CFG at
509 // this point. This won't be a fallthrough, but it will increase locality.
511 BestSucc = selectBestCandidateBlock(Chain, BlockWorkList, BlockFilter);
514 BestSucc = getFirstUnplacedBlock(F, Chain, PrevUnplacedBlockIt,
519 DEBUG(dbgs() << "Unnatural loop CFG detected, forcibly merging the "
520 "layout successor until the CFG reduces\n");
523 // Place this block, updating the datastructures to reflect its placement.
524 BlockChain &SuccChain = *BlockToChain[BestSucc];
525 // Zero out LoopPredecessors for the successor we're about to merge in case
526 // we selected a successor that didn't fit naturally into the CFG.
527 SuccChain.LoopPredecessors = 0;
528 DEBUG(dbgs() << "Merging from " << getBlockNum(BB)
529 << " to " << getBlockNum(BestSucc) << "\n");
530 markChainSuccessors(SuccChain, LoopHeaderBB, BlockWorkList, BlockFilter);
531 Chain.merge(BestSucc, &SuccChain);
532 BB = *llvm::prior(Chain.end());
535 DEBUG(dbgs() << "Finished forming chain for header block "
536 << getBlockNum(*Chain.begin()) << "\n");
539 /// \brief Find the best loop top block for layout.
541 /// This routine implements the logic to analyze the loop looking for the best
542 /// block to layout at the top of the loop. Typically this is done to maximize
543 /// fallthrough opportunities.
545 MachineBlockPlacement::findBestLoopTop(MachineFunction &F,
547 const BlockFilterSet &LoopBlockSet) {
548 BlockFrequency BestExitEdgeFreq;
549 MachineBasicBlock *ExitingBB = 0;
550 MachineBasicBlock *LoopingBB = 0;
551 // If there are exits to outer loops, loop rotation can severely limit
552 // fallthrough opportunites unless it selects such an exit. Keep a set of
553 // blocks where rotating to exit with that block will reach an outer loop.
554 SmallPtrSet<MachineBasicBlock *, 4> BlocksExitingToOuterLoop;
556 DEBUG(dbgs() << "Finding best loop exit for: "
557 << getBlockName(L.getHeader()) << "\n");
558 for (MachineLoop::block_iterator I = L.block_begin(),
561 BlockChain &Chain = *BlockToChain[*I];
562 // Ensure that this block is at the end of a chain; otherwise it could be
563 // mid-way through an inner loop or a successor of an analyzable branch.
564 if (*I != *llvm::prior(Chain.end()))
567 // Now walk the successors. We need to establish whether this has a viable
568 // exiting successor and whether it has a viable non-exiting successor.
569 // We store the old exiting state and restore it if a viable looping
570 // successor isn't found.
571 MachineBasicBlock *OldExitingBB = ExitingBB;
572 BlockFrequency OldBestExitEdgeFreq = BestExitEdgeFreq;
573 // We also compute and store the best looping successor for use in layout.
574 MachineBasicBlock *BestLoopSucc = 0;
575 // FIXME: Due to the performance of the probability and weight routines in
576 // the MBPI analysis, we use the internal weights. This is only valid
577 // because it is purely a ranking function, we don't care about anything
578 // but the relative values.
579 uint32_t BestLoopSuccWeight = 0;
580 // FIXME: We also manually compute the probabilities to avoid quadratic
582 uint32_t WeightScale = 0;
583 uint32_t SumWeight = MBPI->getSumForBlock(*I, WeightScale);
584 for (MachineBasicBlock::succ_iterator SI = (*I)->succ_begin(),
585 SE = (*I)->succ_end();
587 if ((*SI)->isLandingPad())
591 BlockChain &SuccChain = *BlockToChain[*SI];
592 // Don't split chains, either this chain or the successor's chain.
593 if (&Chain == &SuccChain || *SI != *SuccChain.begin()) {
594 DEBUG(dbgs() << " " << (LoopBlockSet.count(*SI) ? "looping: "
596 << getBlockName(*I) << " -> "
597 << getBlockName(*SI) << " (chain conflict)\n");
601 uint32_t SuccWeight = MBPI->getEdgeWeight(*I, *SI);
602 if (LoopBlockSet.count(*SI)) {
603 DEBUG(dbgs() << " looping: " << getBlockName(*I) << " -> "
604 << getBlockName(*SI) << " (" << SuccWeight << ")\n");
605 if (BestLoopSucc && BestLoopSuccWeight >= SuccWeight)
609 BestLoopSuccWeight = SuccWeight;
613 BranchProbability SuccProb(SuccWeight / WeightScale, SumWeight);
614 BlockFrequency ExitEdgeFreq = MBFI->getBlockFreq(*I) * SuccProb;
615 DEBUG(dbgs() << " exiting: " << getBlockName(*I) << " -> "
616 << getBlockName(*SI) << " (" << ExitEdgeFreq << ")\n");
617 // Note that we slightly bias this toward an existing layout successor to
618 // retain incoming order in the absence of better information.
619 // FIXME: Should we bias this more strongly? It's pretty weak.
620 if (!ExitingBB || ExitEdgeFreq > BestExitEdgeFreq ||
621 ((*I)->isLayoutSuccessor(*SI) &&
622 !(ExitEdgeFreq < BestExitEdgeFreq))) {
623 BestExitEdgeFreq = ExitEdgeFreq;
627 if (MachineLoop *ExitLoop = MLI->getLoopFor(*SI))
628 if (ExitLoop->contains(&L))
629 BlocksExitingToOuterLoop.insert(*I);
632 // Restore the old exiting state, no viable looping successor was found.
634 ExitingBB = OldExitingBB;
635 BestExitEdgeFreq = OldBestExitEdgeFreq;
639 // If this was best exiting block thus far, also record the looping block.
641 LoopingBB = BestLoopSucc;
643 // Without a candidate exitting block or with only a single block in the
644 // loop, just use the loop header to layout the loop.
645 if (!ExitingBB || L.getNumBlocks() == 1)
646 return L.getHeader();
648 // Also, if we have exit blocks which lead to outer loops but didn't select
649 // one of them as the exiting block we are rotating toward, disable loop
650 // rotation altogether.
651 if (!BlocksExitingToOuterLoop.empty() &&
652 !BlocksExitingToOuterLoop.count(ExitingBB))
653 return L.getHeader();
655 assert(LoopingBB && "All successors of a loop block are exit blocks!");
656 DEBUG(dbgs() << " Best exiting block: " << getBlockName(ExitingBB) << "\n");
657 DEBUG(dbgs() << " Best top block: " << getBlockName(LoopingBB) << "\n");
661 /// \brief Forms basic block chains from the natural loop structures.
663 /// These chains are designed to preserve the existing *structure* of the code
664 /// as much as possible. We can then stitch the chains together in a way which
665 /// both preserves the topological structure and minimizes taken conditional
667 void MachineBlockPlacement::buildLoopChains(MachineFunction &F,
669 // First recurse through any nested loops, building chains for those inner
671 for (MachineLoop::iterator LI = L.begin(), LE = L.end(); LI != LE; ++LI)
672 buildLoopChains(F, **LI);
674 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
675 BlockFilterSet LoopBlockSet(L.block_begin(), L.block_end());
677 MachineBasicBlock *LayoutTop = findBestLoopTop(F, L, LoopBlockSet);
678 BlockChain &LoopChain = *BlockToChain[LayoutTop];
680 // FIXME: This is a really lame way of walking the chains in the loop: we
681 // walk the blocks, and use a set to prevent visiting a particular chain
683 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
684 assert(LoopChain.LoopPredecessors == 0);
685 UpdatedPreds.insert(&LoopChain);
686 for (MachineLoop::block_iterator BI = L.block_begin(),
689 BlockChain &Chain = *BlockToChain[*BI];
690 if (!UpdatedPreds.insert(&Chain))
693 assert(Chain.LoopPredecessors == 0);
694 for (BlockChain::iterator BCI = Chain.begin(), BCE = Chain.end();
696 assert(BlockToChain[*BCI] == &Chain);
697 for (MachineBasicBlock::pred_iterator PI = (*BCI)->pred_begin(),
698 PE = (*BCI)->pred_end();
700 if (BlockToChain[*PI] == &Chain || !LoopBlockSet.count(*PI))
702 ++Chain.LoopPredecessors;
706 if (Chain.LoopPredecessors == 0)
707 BlockWorkList.push_back(*Chain.begin());
710 buildChain(LayoutTop, LoopChain, BlockWorkList, &LoopBlockSet);
713 // Crash at the end so we get all of the debugging output first.
714 bool BadLoop = false;
715 if (LoopChain.LoopPredecessors) {
717 dbgs() << "Loop chain contains a block without its preds placed!\n"
718 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
719 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n";
721 for (BlockChain::iterator BCI = LoopChain.begin(), BCE = LoopChain.end();
723 if (!LoopBlockSet.erase(*BCI)) {
724 // We don't mark the loop as bad here because there are real situations
725 // where this can occur. For example, with an unanalyzable fallthrough
726 // from a loop block to a non-loop block or vice versa.
727 dbgs() << "Loop chain contains a block not contained by the loop!\n"
728 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
729 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
730 << " Bad block: " << getBlockName(*BCI) << "\n";
733 if (!LoopBlockSet.empty()) {
735 for (BlockFilterSet::iterator LBI = LoopBlockSet.begin(),
736 LBE = LoopBlockSet.end();
738 dbgs() << "Loop contains blocks never placed into a chain!\n"
739 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
740 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
741 << " Bad block: " << getBlockName(*LBI) << "\n";
743 assert(!BadLoop && "Detected problems with the placement of this loop.");
747 void MachineBlockPlacement::buildCFGChains(MachineFunction &F) {
748 // Ensure that every BB in the function has an associated chain to simplify
749 // the assumptions of the remaining algorithm.
750 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
751 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
752 MachineBasicBlock *BB = FI;
754 = new (ChainAllocator.Allocate()) BlockChain(BlockToChain, BB);
755 // Also, merge any blocks which we cannot reason about and must preserve
756 // the exact fallthrough behavior for.
759 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch.
760 if (!TII->AnalyzeBranch(*BB, TBB, FBB, Cond) || !FI->canFallThrough())
763 MachineFunction::iterator NextFI(llvm::next(FI));
764 MachineBasicBlock *NextBB = NextFI;
765 // Ensure that the layout successor is a viable block, as we know that
766 // fallthrough is a possibility.
767 assert(NextFI != FE && "Can't fallthrough past the last block.");
768 DEBUG(dbgs() << "Pre-merging due to unanalyzable fallthrough: "
769 << getBlockName(BB) << " -> " << getBlockName(NextBB)
771 Chain->merge(NextBB, 0);
777 // Build any loop-based chains.
778 for (MachineLoopInfo::iterator LI = MLI->begin(), LE = MLI->end(); LI != LE;
780 buildLoopChains(F, **LI);
782 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
784 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
785 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
786 MachineBasicBlock *BB = &*FI;
787 BlockChain &Chain = *BlockToChain[BB];
788 if (!UpdatedPreds.insert(&Chain))
791 assert(Chain.LoopPredecessors == 0);
792 for (BlockChain::iterator BCI = Chain.begin(), BCE = Chain.end();
794 assert(BlockToChain[*BCI] == &Chain);
795 for (MachineBasicBlock::pred_iterator PI = (*BCI)->pred_begin(),
796 PE = (*BCI)->pred_end();
798 if (BlockToChain[*PI] == &Chain)
800 ++Chain.LoopPredecessors;
804 if (Chain.LoopPredecessors == 0)
805 BlockWorkList.push_back(*Chain.begin());
808 BlockChain &FunctionChain = *BlockToChain[&F.front()];
809 buildChain(&F.front(), FunctionChain, BlockWorkList);
811 typedef SmallPtrSet<MachineBasicBlock *, 16> FunctionBlockSetType;
813 // Crash at the end so we get all of the debugging output first.
814 bool BadFunc = false;
815 FunctionBlockSetType FunctionBlockSet;
816 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
817 FunctionBlockSet.insert(FI);
819 for (BlockChain::iterator BCI = FunctionChain.begin(),
820 BCE = FunctionChain.end();
822 if (!FunctionBlockSet.erase(*BCI)) {
824 dbgs() << "Function chain contains a block not in the function!\n"
825 << " Bad block: " << getBlockName(*BCI) << "\n";
828 if (!FunctionBlockSet.empty()) {
830 for (FunctionBlockSetType::iterator FBI = FunctionBlockSet.begin(),
831 FBE = FunctionBlockSet.end();
833 dbgs() << "Function contains blocks never placed into a chain!\n"
834 << " Bad block: " << getBlockName(*FBI) << "\n";
836 assert(!BadFunc && "Detected problems with the block placement.");
839 // Splice the blocks into place.
840 MachineFunction::iterator InsertPos = F.begin();
841 for (BlockChain::iterator BI = FunctionChain.begin(),
842 BE = FunctionChain.end();
844 DEBUG(dbgs() << (BI == FunctionChain.begin() ? "Placing chain "
846 << getBlockName(*BI) << "\n");
847 if (InsertPos != MachineFunction::iterator(*BI))
848 F.splice(InsertPos, *BI);
852 // Update the terminator of the previous block.
853 if (BI == FunctionChain.begin())
855 MachineBasicBlock *PrevBB = llvm::prior(MachineFunction::iterator(*BI));
857 // FIXME: It would be awesome of updateTerminator would just return rather
858 // than assert when the branch cannot be analyzed in order to remove this
861 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch.
862 if (!TII->AnalyzeBranch(*PrevBB, TBB, FBB, Cond))
863 PrevBB->updateTerminator();
866 // Fixup the last block.
868 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch.
869 if (!TII->AnalyzeBranch(F.back(), TBB, FBB, Cond))
870 F.back().updateTerminator();
873 /// \brief Recursive helper to align a loop and any nested loops.
874 static void AlignLoop(MachineFunction &F, MachineLoop *L, unsigned Align) {
875 // Recurse through nested loops.
876 for (MachineLoop::iterator I = L->begin(), E = L->end(); I != E; ++I)
877 AlignLoop(F, *I, Align);
879 L->getTopBlock()->setAlignment(Align);
882 /// \brief Align loop headers to target preferred alignments.
883 void MachineBlockPlacement::AlignLoops(MachineFunction &F) {
884 if (F.getFunction()->hasFnAttr(Attribute::OptimizeForSize))
887 unsigned Align = TLI->getPrefLoopAlignment();
889 return; // Don't care about loop alignment.
891 for (MachineLoopInfo::iterator I = MLI->begin(), E = MLI->end(); I != E; ++I)
892 AlignLoop(F, *I, Align);
895 bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &F) {
896 // Check for single-block functions and skip them.
897 if (llvm::next(F.begin()) == F.end())
900 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
901 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
902 MLI = &getAnalysis<MachineLoopInfo>();
903 TII = F.getTarget().getInstrInfo();
904 TLI = F.getTarget().getTargetLowering();
905 assert(BlockToChain.empty());
910 BlockToChain.clear();
911 ChainAllocator.DestroyAll();
913 // We always return true as we have no way to track whether the final order
914 // differs from the original order.
919 /// \brief A pass to compute block placement statistics.
921 /// A separate pass to compute interesting statistics for evaluating block
922 /// placement. This is separate from the actual placement pass so that they can
923 /// be computed in the absense of any placement transformations or when using
924 /// alternative placement strategies.
925 class MachineBlockPlacementStats : public MachineFunctionPass {
926 /// \brief A handle to the branch probability pass.
927 const MachineBranchProbabilityInfo *MBPI;
929 /// \brief A handle to the function-wide block frequency pass.
930 const MachineBlockFrequencyInfo *MBFI;
933 static char ID; // Pass identification, replacement for typeid
934 MachineBlockPlacementStats() : MachineFunctionPass(ID) {
935 initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry());
938 bool runOnMachineFunction(MachineFunction &F);
940 void getAnalysisUsage(AnalysisUsage &AU) const {
941 AU.addRequired<MachineBranchProbabilityInfo>();
942 AU.addRequired<MachineBlockFrequencyInfo>();
943 AU.setPreservesAll();
944 MachineFunctionPass::getAnalysisUsage(AU);
947 const char *getPassName() const { return "Block Placement Stats"; }
951 char MachineBlockPlacementStats::ID = 0;
952 INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats",
953 "Basic Block Placement Stats", false, false)
954 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
955 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
956 INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats",
957 "Basic Block Placement Stats", false, false)
959 FunctionPass *llvm::createMachineBlockPlacementStatsPass() {
960 return new MachineBlockPlacementStats();
963 bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) {
964 // Check for single-block functions and skip them.
965 if (llvm::next(F.begin()) == F.end())
968 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
969 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
971 for (MachineFunction::iterator I = F.begin(), E = F.end(); I != E; ++I) {
972 BlockFrequency BlockFreq = MBFI->getBlockFreq(I);
973 Statistic &NumBranches = (I->succ_size() > 1) ? NumCondBranches
975 Statistic &BranchTakenFreq = (I->succ_size() > 1) ? CondBranchTakenFreq
976 : UncondBranchTakenFreq;
977 for (MachineBasicBlock::succ_iterator SI = I->succ_begin(),
980 // Skip if this successor is a fallthrough.
981 if (I->isLayoutSuccessor(*SI))
984 BlockFrequency EdgeFreq = BlockFreq * MBPI->getEdgeProbability(I, *SI);
986 BranchTakenFreq += EdgeFreq.getFrequency();