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/Support/ErrorHandling.h"
40 #include "llvm/ADT/DenseMap.h"
41 #include "llvm/ADT/PostOrderIterator.h"
42 #include "llvm/ADT/SCCIterator.h"
43 #include "llvm/ADT/SmallPtrSet.h"
44 #include "llvm/ADT/SmallVector.h"
45 #include "llvm/ADT/Statistic.h"
46 #include "llvm/Target/TargetInstrInfo.h"
47 #include "llvm/Target/TargetLowering.h"
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");
59 /// \brief A structure for storing a weighted edge.
61 /// This stores an edge and its weight, computed as the product of the
62 /// frequency that the starting block is entered with the probability of
63 /// a particular exit block.
65 BlockFrequency EdgeFrequency;
66 MachineBasicBlock *From, *To;
68 bool operator<(const WeightedEdge &RHS) const {
69 return EdgeFrequency < RHS.EdgeFrequency;
76 /// \brief Type for our function-wide basic block -> block chain mapping.
77 typedef DenseMap<MachineBasicBlock *, BlockChain *> BlockToChainMapType;
81 /// \brief A chain of blocks which will be laid out contiguously.
83 /// This is the datastructure representing a chain of consecutive blocks that
84 /// are profitable to layout together in order to maximize fallthrough
85 /// probabilities. We also can use a block chain to represent a sequence of
86 /// basic blocks which have some external (correctness) requirement for
87 /// sequential layout.
89 /// Eventually, the block chains will form a directed graph over the function.
90 /// We provide an SCC-supporting-iterator in order to quicky build and walk the
91 /// SCCs of block chains within a function.
93 /// The block chains also have support for calculating and caching probability
94 /// information related to the chain itself versus other chains. This is used
95 /// for ranking during the final layout of block chains.
97 /// \brief The sequence of blocks belonging to this chain.
99 /// This is the sequence of blocks for a particular chain. These will be laid
100 /// out in-order within the function.
101 SmallVector<MachineBasicBlock *, 4> Blocks;
103 /// \brief A handle to the function-wide basic block to block chain mapping.
105 /// This is retained in each block chain to simplify the computation of child
106 /// block chains for SCC-formation and iteration. We store the edges to child
107 /// basic blocks, and map them back to their associated chains using this
109 BlockToChainMapType &BlockToChain;
112 /// \brief Construct a new BlockChain.
114 /// This builds a new block chain representing a single basic block in the
115 /// function. It also registers itself as the chain that block participates
116 /// in with the BlockToChain mapping.
117 BlockChain(BlockToChainMapType &BlockToChain, MachineBasicBlock *BB)
118 : Blocks(1, BB), BlockToChain(BlockToChain), LoopPredecessors(0) {
119 assert(BB && "Cannot create a chain with a null basic block");
120 BlockToChain[BB] = this;
123 /// \brief Iterator over blocks within the chain.
124 typedef SmallVectorImpl<MachineBasicBlock *>::iterator iterator;
125 typedef SmallVectorImpl<MachineBasicBlock *>::reverse_iterator
128 /// \brief Beginning of blocks within the chain.
129 iterator begin() { return Blocks.begin(); }
130 reverse_iterator rbegin() { return Blocks.rbegin(); }
132 /// \brief End of blocks within the chain.
133 iterator end() { return Blocks.end(); }
134 reverse_iterator rend() { return Blocks.rend(); }
136 /// \brief Merge a block chain into this one.
138 /// This routine merges a block chain into this one. It takes care of forming
139 /// a contiguous sequence of basic blocks, updating the edge list, and
140 /// updating the block -> chain mapping. It does not free or tear down the
141 /// old chain, but the old chain's block list is no longer valid.
142 void merge(MachineBasicBlock *BB, BlockChain *Chain) {
144 assert(!Blocks.empty());
146 // Fast path in case we don't have a chain already.
148 assert(!BlockToChain[BB]);
149 Blocks.push_back(BB);
150 BlockToChain[BB] = this;
154 assert(BB == *Chain->begin());
155 assert(Chain->begin() != Chain->end());
157 // Update the incoming blocks to point to this chain, and add them to the
159 for (BlockChain::iterator BI = Chain->begin(), BE = Chain->end();
161 Blocks.push_back(*BI);
162 assert(BlockToChain[*BI] == Chain && "Incoming blocks not in chain");
163 BlockToChain[*BI] = this;
167 /// \brief Count of predecessors within the loop currently being processed.
169 /// This count is updated at each loop we process to represent the number of
170 /// in-loop predecessors of this chain.
171 unsigned LoopPredecessors;
176 class MachineBlockPlacement : public MachineFunctionPass {
177 /// \brief A typedef for a block filter set.
178 typedef SmallPtrSet<MachineBasicBlock *, 16> BlockFilterSet;
180 /// \brief A handle to the branch probability pass.
181 const MachineBranchProbabilityInfo *MBPI;
183 /// \brief A handle to the function-wide block frequency pass.
184 const MachineBlockFrequencyInfo *MBFI;
186 /// \brief A handle to the loop info.
187 const MachineLoopInfo *MLI;
189 /// \brief A handle to the target's instruction info.
190 const TargetInstrInfo *TII;
192 /// \brief A handle to the target's lowering info.
193 const TargetLowering *TLI;
195 /// \brief Allocator and owner of BlockChain structures.
197 /// We build BlockChains lazily by merging together high probability BB
198 /// sequences acording to the "Algo2" in the paper mentioned at the top of
199 /// the file. To reduce malloc traffic, we allocate them using this slab-like
200 /// allocator, and destroy them after the pass completes.
201 SpecificBumpPtrAllocator<BlockChain> ChainAllocator;
203 /// \brief Function wide BasicBlock to BlockChain mapping.
205 /// This mapping allows efficiently moving from any given basic block to the
206 /// BlockChain it participates in, if any. We use it to, among other things,
207 /// allow implicitly defining edges between chains as the existing edges
208 /// between basic blocks.
209 DenseMap<MachineBasicBlock *, BlockChain *> BlockToChain;
211 void markChainSuccessors(BlockChain &Chain,
212 MachineBasicBlock *LoopHeaderBB,
213 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
214 const BlockFilterSet *BlockFilter = 0);
215 MachineBasicBlock *selectBestSuccessor(MachineBasicBlock *BB,
217 const BlockFilterSet *BlockFilter);
218 MachineBasicBlock *selectBestCandidateBlock(
219 BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList,
220 const BlockFilterSet *BlockFilter);
221 MachineBasicBlock *getFirstUnplacedBlock(
223 const BlockChain &PlacedChain,
224 MachineFunction::iterator &PrevUnplacedBlockIt,
225 const BlockFilterSet *BlockFilter);
226 void buildChain(MachineBasicBlock *BB, BlockChain &Chain,
227 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
228 const BlockFilterSet *BlockFilter = 0);
229 MachineBasicBlock *findBestLoopTop(MachineFunction &F,
231 const BlockFilterSet &LoopBlockSet);
232 void buildLoopChains(MachineFunction &F, MachineLoop &L);
233 void buildCFGChains(MachineFunction &F);
234 void AlignLoops(MachineFunction &F);
237 static char ID; // Pass identification, replacement for typeid
238 MachineBlockPlacement() : MachineFunctionPass(ID) {
239 initializeMachineBlockPlacementPass(*PassRegistry::getPassRegistry());
242 bool runOnMachineFunction(MachineFunction &F);
244 void getAnalysisUsage(AnalysisUsage &AU) const {
245 AU.addRequired<MachineBranchProbabilityInfo>();
246 AU.addRequired<MachineBlockFrequencyInfo>();
247 AU.addRequired<MachineLoopInfo>();
248 MachineFunctionPass::getAnalysisUsage(AU);
251 const char *getPassName() const { return "Block Placement"; }
255 char MachineBlockPlacement::ID = 0;
256 INITIALIZE_PASS_BEGIN(MachineBlockPlacement, "block-placement2",
257 "Branch Probability Basic Block Placement", false, false)
258 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
259 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
260 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
261 INITIALIZE_PASS_END(MachineBlockPlacement, "block-placement2",
262 "Branch Probability Basic Block Placement", false, false)
264 FunctionPass *llvm::createMachineBlockPlacementPass() {
265 return new MachineBlockPlacement();
269 /// \brief Helper to print the name of a MBB.
271 /// Only used by debug logging.
272 static std::string getBlockName(MachineBasicBlock *BB) {
274 raw_string_ostream OS(Result);
275 OS << "BB#" << BB->getNumber()
276 << " (derived from LLVM BB '" << BB->getName() << "')";
281 /// \brief Helper to print the number of a MBB.
283 /// Only used by debug logging.
284 static std::string getBlockNum(MachineBasicBlock *BB) {
286 raw_string_ostream OS(Result);
287 OS << "BB#" << BB->getNumber();
293 /// \brief Mark a chain's successors as having one fewer preds.
295 /// When a chain is being merged into the "placed" chain, this routine will
296 /// quickly walk the successors of each block in the chain and mark them as
297 /// having one fewer active predecessor. It also adds any successors of this
298 /// chain which reach the zero-predecessor state to the worklist passed in.
299 void MachineBlockPlacement::markChainSuccessors(
301 MachineBasicBlock *LoopHeaderBB,
302 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
303 const BlockFilterSet *BlockFilter) {
304 // Walk all the blocks in this chain, marking their successors as having
305 // a predecessor placed.
306 for (BlockChain::iterator CBI = Chain.begin(), CBE = Chain.end();
308 // Add any successors for which this is the only un-placed in-loop
309 // predecessor to the worklist as a viable candidate for CFG-neutral
310 // placement. No subsequent placement of this block will violate the CFG
311 // shape, so we get to use heuristics to choose a favorable placement.
312 for (MachineBasicBlock::succ_iterator SI = (*CBI)->succ_begin(),
313 SE = (*CBI)->succ_end();
315 if (BlockFilter && !BlockFilter->count(*SI))
317 BlockChain &SuccChain = *BlockToChain[*SI];
318 // Disregard edges within a fixed chain, or edges to the loop header.
319 if (&Chain == &SuccChain || *SI == LoopHeaderBB)
322 // This is a cross-chain edge that is within the loop, so decrement the
323 // loop predecessor count of the destination chain.
324 if (SuccChain.LoopPredecessors > 0 && --SuccChain.LoopPredecessors == 0)
325 BlockWorkList.push_back(*SuccChain.begin());
330 /// \brief Select the best successor for a block.
332 /// This looks across all successors of a particular block and attempts to
333 /// select the "best" one to be the layout successor. It only considers direct
334 /// successors which also pass the block filter. It will attempt to avoid
335 /// breaking CFG structure, but cave and break such structures in the case of
336 /// very hot successor edges.
338 /// \returns The best successor block found, or null if none are viable.
339 MachineBasicBlock *MachineBlockPlacement::selectBestSuccessor(
340 MachineBasicBlock *BB, BlockChain &Chain,
341 const BlockFilterSet *BlockFilter) {
342 const BranchProbability HotProb(4, 5); // 80%
344 MachineBasicBlock *BestSucc = 0;
345 // FIXME: Due to the performance of the probability and weight routines in
346 // the MBPI analysis, we manually compute probabilities using the edge
347 // weights. This is suboptimal as it means that the somewhat subtle
348 // definition of edge weight semantics is encoded here as well. We should
349 // improve the MBPI interface to effeciently support query patterns such as
351 uint32_t BestWeight = 0;
352 uint32_t WeightScale = 0;
353 uint32_t SumWeight = MBPI->getSumForBlock(BB, WeightScale);
354 DEBUG(dbgs() << "Attempting merge from: " << getBlockName(BB) << "\n");
355 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
358 if (BlockFilter && !BlockFilter->count(*SI))
360 BlockChain &SuccChain = *BlockToChain[*SI];
361 if (&SuccChain == &Chain) {
362 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> Already merged!\n");
365 if (*SI != *SuccChain.begin()) {
366 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> Mid chain!\n");
370 uint32_t SuccWeight = MBPI->getEdgeWeight(BB, *SI);
371 BranchProbability SuccProb(SuccWeight / WeightScale, SumWeight);
373 // Only consider successors which are either "hot", or wouldn't violate
374 // any CFG constraints.
375 if (SuccChain.LoopPredecessors != 0) {
376 if (SuccProb < HotProb) {
377 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> CFG conflict\n");
381 // Make sure that a hot successor doesn't have a globally more important
383 BlockFrequency CandidateEdgeFreq
384 = MBFI->getBlockFreq(BB) * SuccProb * HotProb.getCompl();
385 bool BadCFGConflict = false;
386 for (MachineBasicBlock::pred_iterator PI = (*SI)->pred_begin(),
387 PE = (*SI)->pred_end();
389 if (*PI == *SI || (BlockFilter && !BlockFilter->count(*PI)) ||
390 BlockToChain[*PI] == &Chain)
392 BlockFrequency PredEdgeFreq
393 = MBFI->getBlockFreq(*PI) * MBPI->getEdgeProbability(*PI, *SI);
394 if (PredEdgeFreq >= CandidateEdgeFreq) {
395 BadCFGConflict = true;
399 if (BadCFGConflict) {
400 DEBUG(dbgs() << " " << getBlockName(*SI)
401 << " -> non-cold CFG conflict\n");
406 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> " << SuccProb
408 << (SuccChain.LoopPredecessors != 0 ? " (CFG break)" : "")
410 if (BestSucc && BestWeight >= SuccWeight)
413 BestWeight = SuccWeight;
419 /// \brief Predicate struct to detect blocks already placed.
420 class IsBlockPlaced {
421 const BlockChain &PlacedChain;
422 const BlockToChainMapType &BlockToChain;
425 IsBlockPlaced(const BlockChain &PlacedChain,
426 const BlockToChainMapType &BlockToChain)
427 : PlacedChain(PlacedChain), BlockToChain(BlockToChain) {}
429 bool operator()(MachineBasicBlock *BB) const {
430 return BlockToChain.lookup(BB) == &PlacedChain;
435 /// \brief Select the best block from a worklist.
437 /// This looks through the provided worklist as a list of candidate basic
438 /// blocks and select the most profitable one to place. The definition of
439 /// profitable only really makes sense in the context of a loop. This returns
440 /// the most frequently visited block in the worklist, which in the case of
441 /// a loop, is the one most desirable to be physically close to the rest of the
442 /// loop body in order to improve icache behavior.
444 /// \returns The best block found, or null if none are viable.
445 MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock(
446 BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList,
447 const BlockFilterSet *BlockFilter) {
448 // Once we need to walk the worklist looking for a candidate, cleanup the
449 // worklist of already placed entries.
450 // FIXME: If this shows up on profiles, it could be folded (at the cost of
451 // some code complexity) into the loop below.
452 WorkList.erase(std::remove_if(WorkList.begin(), WorkList.end(),
453 IsBlockPlaced(Chain, BlockToChain)),
456 MachineBasicBlock *BestBlock = 0;
457 BlockFrequency BestFreq;
458 for (SmallVectorImpl<MachineBasicBlock *>::iterator WBI = WorkList.begin(),
459 WBE = WorkList.end();
461 assert(!BlockFilter || BlockFilter->count(*WBI));
462 BlockChain &SuccChain = *BlockToChain[*WBI];
463 if (&SuccChain == &Chain) {
464 DEBUG(dbgs() << " " << getBlockName(*WBI)
465 << " -> Already merged!\n");
468 assert(SuccChain.LoopPredecessors == 0 && "Found CFG-violating block");
470 BlockFrequency CandidateFreq = MBFI->getBlockFreq(*WBI);
471 DEBUG(dbgs() << " " << getBlockName(*WBI) << " -> " << CandidateFreq
473 if (BestBlock && BestFreq >= CandidateFreq)
476 BestFreq = CandidateFreq;
481 /// \brief Retrieve the first unplaced basic block.
483 /// This routine is called when we are unable to use the CFG to walk through
484 /// all of the basic blocks and form a chain due to unnatural loops in the CFG.
485 /// We walk through the function's blocks in order, starting from the
486 /// LastUnplacedBlockIt. We update this iterator on each call to avoid
487 /// re-scanning the entire sequence on repeated calls to this routine.
488 MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock(
489 MachineFunction &F, const BlockChain &PlacedChain,
490 MachineFunction::iterator &PrevUnplacedBlockIt,
491 const BlockFilterSet *BlockFilter) {
492 for (MachineFunction::iterator I = PrevUnplacedBlockIt, E = F.end(); I != E;
494 if (BlockFilter && !BlockFilter->count(I))
496 if (BlockToChain[I] != &PlacedChain) {
497 PrevUnplacedBlockIt = I;
498 // Now select the head of the chain to which the unplaced block belongs
499 // as the block to place. This will force the entire chain to be placed,
500 // and satisfies the requirements of merging chains.
501 return *BlockToChain[I]->begin();
507 void MachineBlockPlacement::buildChain(
508 MachineBasicBlock *BB,
510 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
511 const BlockFilterSet *BlockFilter) {
513 assert(BlockToChain[BB] == &Chain);
514 MachineFunction &F = *BB->getParent();
515 MachineFunction::iterator PrevUnplacedBlockIt = F.begin();
517 MachineBasicBlock *LoopHeaderBB = BB;
518 markChainSuccessors(Chain, LoopHeaderBB, BlockWorkList, BlockFilter);
519 BB = *llvm::prior(Chain.end());
522 assert(BlockToChain[BB] == &Chain);
523 assert(*llvm::prior(Chain.end()) == BB);
524 MachineBasicBlock *BestSucc = 0;
526 // Look for the best viable successor if there is one to place immediately
528 BestSucc = selectBestSuccessor(BB, Chain, BlockFilter);
530 // If an immediate successor isn't available, look for the best viable
531 // block among those we've identified as not violating the loop's CFG at
532 // this point. This won't be a fallthrough, but it will increase locality.
534 BestSucc = selectBestCandidateBlock(Chain, BlockWorkList, BlockFilter);
537 BestSucc = getFirstUnplacedBlock(F, Chain, PrevUnplacedBlockIt,
542 DEBUG(dbgs() << "Unnatural loop CFG detected, forcibly merging the "
543 "layout successor until the CFG reduces\n");
546 // Place this block, updating the datastructures to reflect its placement.
547 BlockChain &SuccChain = *BlockToChain[BestSucc];
548 // Zero out LoopPredecessors for the successor we're about to merge in case
549 // we selected a successor that didn't fit naturally into the CFG.
550 SuccChain.LoopPredecessors = 0;
551 DEBUG(dbgs() << "Merging from " << getBlockNum(BB)
552 << " to " << getBlockNum(BestSucc) << "\n");
553 markChainSuccessors(SuccChain, LoopHeaderBB, BlockWorkList, BlockFilter);
554 Chain.merge(BestSucc, &SuccChain);
555 BB = *llvm::prior(Chain.end());
558 DEBUG(dbgs() << "Finished forming chain for header block "
559 << getBlockNum(*Chain.begin()) << "\n");
562 /// \brief Find the best loop top block for layout.
564 /// This routine implements the logic to analyze the loop looking for the best
565 /// block to layout at the top of the loop. Typically this is done to maximize
566 /// fallthrough opportunities.
568 MachineBlockPlacement::findBestLoopTop(MachineFunction &F,
570 const BlockFilterSet &LoopBlockSet) {
571 BlockFrequency BestExitEdgeFreq;
572 MachineBasicBlock *ExitingBB = 0;
573 MachineBasicBlock *LoopingBB = 0;
574 // If there are exits to outer loops, loop rotation can severely limit
575 // fallthrough opportunites unless it selects such an exit. Keep a set of
576 // blocks where rotating to exit with that block will reach an outer loop.
577 SmallPtrSet<MachineBasicBlock *, 4> BlocksExitingToOuterLoop;
579 DEBUG(dbgs() << "Finding best loop exit for: "
580 << getBlockName(L.getHeader()) << "\n");
581 for (MachineLoop::block_iterator I = L.block_begin(),
584 BlockChain &Chain = *BlockToChain[*I];
585 // Ensure that this block is at the end of a chain; otherwise it could be
586 // mid-way through an inner loop or a successor of an analyzable branch.
587 if (*I != *llvm::prior(Chain.end()))
590 // Now walk the successors. We need to establish whether this has a viable
591 // exiting successor and whether it has a viable non-exiting successor.
592 // We store the old exiting state and restore it if a viable looping
593 // successor isn't found.
594 MachineBasicBlock *OldExitingBB = ExitingBB;
595 BlockFrequency OldBestExitEdgeFreq = BestExitEdgeFreq;
596 // We also compute and store the best looping successor for use in layout.
597 MachineBasicBlock *BestLoopSucc = 0;
598 // FIXME: Due to the performance of the probability and weight routines in
599 // the MBPI analysis, we use the internal weights. This is only valid
600 // because it is purely a ranking function, we don't care about anything
601 // but the relative values.
602 uint32_t BestLoopSuccWeight = 0;
603 // FIXME: We also manually compute the probabilities to avoid quadratic
605 uint32_t WeightScale = 0;
606 uint32_t SumWeight = MBPI->getSumForBlock(*I, WeightScale);
607 for (MachineBasicBlock::succ_iterator SI = (*I)->succ_begin(),
608 SE = (*I)->succ_end();
610 if ((*SI)->isLandingPad())
614 BlockChain &SuccChain = *BlockToChain[*SI];
615 // Don't split chains, either this chain or the successor's chain.
616 if (&Chain == &SuccChain || *SI != *SuccChain.begin()) {
617 DEBUG(dbgs() << " " << (LoopBlockSet.count(*SI) ? "looping: "
619 << getBlockName(*I) << " -> "
620 << getBlockName(*SI) << " (chain conflict)\n");
624 uint32_t SuccWeight = MBPI->getEdgeWeight(*I, *SI);
625 if (LoopBlockSet.count(*SI)) {
626 DEBUG(dbgs() << " looping: " << getBlockName(*I) << " -> "
627 << getBlockName(*SI) << " (" << SuccWeight << ")\n");
628 if (BestLoopSucc && BestLoopSuccWeight >= SuccWeight)
632 BestLoopSuccWeight = SuccWeight;
636 BranchProbability SuccProb(SuccWeight / WeightScale, SumWeight);
637 BlockFrequency ExitEdgeFreq = MBFI->getBlockFreq(*I) * SuccProb;
638 DEBUG(dbgs() << " exiting: " << getBlockName(*I) << " -> "
639 << getBlockName(*SI) << " (" << ExitEdgeFreq << ")\n");
640 // Note that we slightly bias this toward an existing layout successor to
641 // retain incoming order in the absence of better information.
642 // FIXME: Should we bias this more strongly? It's pretty weak.
643 if (!ExitingBB || ExitEdgeFreq > BestExitEdgeFreq ||
644 ((*I)->isLayoutSuccessor(*SI) &&
645 !(ExitEdgeFreq < BestExitEdgeFreq))) {
646 BestExitEdgeFreq = ExitEdgeFreq;
650 if (MachineLoop *ExitLoop = MLI->getLoopFor(*SI))
651 if (ExitLoop->contains(&L))
652 BlocksExitingToOuterLoop.insert(*I);
655 // Restore the old exiting state, no viable looping successor was found.
657 ExitingBB = OldExitingBB;
658 BestExitEdgeFreq = OldBestExitEdgeFreq;
662 // If this was best exiting block thus far, also record the looping block.
664 LoopingBB = BestLoopSucc;
666 // Without a candidate exitting block or with only a single block in the
667 // loop, just use the loop header to layout the loop.
668 if (!ExitingBB || L.getNumBlocks() == 1)
669 return L.getHeader();
671 // Also, if we have exit blocks which lead to outer loops but didn't select
672 // one of them as the exiting block we are rotating toward, disable loop
673 // rotation altogether.
674 if (!BlocksExitingToOuterLoop.empty() &&
675 !BlocksExitingToOuterLoop.count(ExitingBB))
676 return L.getHeader();
678 assert(LoopingBB && "All successors of a loop block are exit blocks!");
679 DEBUG(dbgs() << " Best exiting block: " << getBlockName(ExitingBB) << "\n");
680 DEBUG(dbgs() << " Best top block: " << getBlockName(LoopingBB) << "\n");
684 /// \brief Forms basic block chains from the natural loop structures.
686 /// These chains are designed to preserve the existing *structure* of the code
687 /// as much as possible. We can then stitch the chains together in a way which
688 /// both preserves the topological structure and minimizes taken conditional
690 void MachineBlockPlacement::buildLoopChains(MachineFunction &F,
692 // First recurse through any nested loops, building chains for those inner
694 for (MachineLoop::iterator LI = L.begin(), LE = L.end(); LI != LE; ++LI)
695 buildLoopChains(F, **LI);
697 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
698 BlockFilterSet LoopBlockSet(L.block_begin(), L.block_end());
700 MachineBasicBlock *LayoutTop = findBestLoopTop(F, L, LoopBlockSet);
701 BlockChain &LoopChain = *BlockToChain[LayoutTop];
703 // FIXME: This is a really lame way of walking the chains in the loop: we
704 // walk the blocks, and use a set to prevent visiting a particular chain
706 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
707 assert(BlockToChain[LayoutTop]->LoopPredecessors == 0);
708 UpdatedPreds.insert(BlockToChain[LayoutTop]);
709 for (MachineLoop::block_iterator BI = L.block_begin(),
712 BlockChain &Chain = *BlockToChain[*BI];
713 if (!UpdatedPreds.insert(&Chain))
716 assert(Chain.LoopPredecessors == 0);
717 for (BlockChain::iterator BCI = Chain.begin(), BCE = Chain.end();
719 assert(BlockToChain[*BCI] == &Chain);
720 for (MachineBasicBlock::pred_iterator PI = (*BCI)->pred_begin(),
721 PE = (*BCI)->pred_end();
723 if (BlockToChain[*PI] == &Chain || !LoopBlockSet.count(*PI))
725 ++Chain.LoopPredecessors;
729 if (Chain.LoopPredecessors == 0)
730 BlockWorkList.push_back(*Chain.begin());
733 buildChain(LayoutTop, LoopChain, BlockWorkList, &LoopBlockSet);
736 // Crash at the end so we get all of the debugging output first.
737 bool BadLoop = false;
738 if (LoopChain.LoopPredecessors) {
740 dbgs() << "Loop chain contains a block without its preds placed!\n"
741 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
742 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n";
744 for (BlockChain::iterator BCI = LoopChain.begin(), BCE = LoopChain.end();
746 if (!LoopBlockSet.erase(*BCI)) {
747 // We don't mark the loop as bad here because there are real situations
748 // where this can occur. For example, with an unanalyzable fallthrough
749 // from a loop block to a non-loop block or vice versa.
750 dbgs() << "Loop chain contains a block not contained by the loop!\n"
751 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
752 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
753 << " Bad block: " << getBlockName(*BCI) << "\n";
756 if (!LoopBlockSet.empty()) {
758 for (BlockFilterSet::iterator LBI = LoopBlockSet.begin(),
759 LBE = LoopBlockSet.end();
761 dbgs() << "Loop contains blocks never placed into a chain!\n"
762 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
763 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
764 << " Bad block: " << getBlockName(*LBI) << "\n";
766 assert(!BadLoop && "Detected problems with the placement of this loop.");
770 void MachineBlockPlacement::buildCFGChains(MachineFunction &F) {
771 // Ensure that every BB in the function has an associated chain to simplify
772 // the assumptions of the remaining algorithm.
773 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
774 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
775 MachineBasicBlock *BB = FI;
777 = new (ChainAllocator.Allocate()) BlockChain(BlockToChain, BB);
778 // Also, merge any blocks which we cannot reason about and must preserve
779 // the exact fallthrough behavior for.
782 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch.
783 if (!TII->AnalyzeBranch(*BB, TBB, FBB, Cond) || !FI->canFallThrough())
786 MachineFunction::iterator NextFI(llvm::next(FI));
787 MachineBasicBlock *NextBB = NextFI;
788 // Ensure that the layout successor is a viable block, as we know that
789 // fallthrough is a possibility.
790 assert(NextFI != FE && "Can't fallthrough past the last block.");
791 DEBUG(dbgs() << "Pre-merging due to unanalyzable fallthrough: "
792 << getBlockName(BB) << " -> " << getBlockName(NextBB)
794 Chain->merge(NextBB, 0);
800 // Build any loop-based chains.
801 for (MachineLoopInfo::iterator LI = MLI->begin(), LE = MLI->end(); LI != LE;
803 buildLoopChains(F, **LI);
805 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
807 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
808 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
809 MachineBasicBlock *BB = &*FI;
810 BlockChain &Chain = *BlockToChain[BB];
811 if (!UpdatedPreds.insert(&Chain))
814 assert(Chain.LoopPredecessors == 0);
815 for (BlockChain::iterator BCI = Chain.begin(), BCE = Chain.end();
817 assert(BlockToChain[*BCI] == &Chain);
818 for (MachineBasicBlock::pred_iterator PI = (*BCI)->pred_begin(),
819 PE = (*BCI)->pred_end();
821 if (BlockToChain[*PI] == &Chain)
823 ++Chain.LoopPredecessors;
827 if (Chain.LoopPredecessors == 0)
828 BlockWorkList.push_back(*Chain.begin());
831 BlockChain &FunctionChain = *BlockToChain[&F.front()];
832 buildChain(&F.front(), FunctionChain, BlockWorkList);
834 typedef SmallPtrSet<MachineBasicBlock *, 16> FunctionBlockSetType;
836 // Crash at the end so we get all of the debugging output first.
837 bool BadFunc = false;
838 FunctionBlockSetType FunctionBlockSet;
839 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
840 FunctionBlockSet.insert(FI);
842 for (BlockChain::iterator BCI = FunctionChain.begin(),
843 BCE = FunctionChain.end();
845 if (!FunctionBlockSet.erase(*BCI)) {
847 dbgs() << "Function chain contains a block not in the function!\n"
848 << " Bad block: " << getBlockName(*BCI) << "\n";
851 if (!FunctionBlockSet.empty()) {
853 for (FunctionBlockSetType::iterator FBI = FunctionBlockSet.begin(),
854 FBE = FunctionBlockSet.end();
856 dbgs() << "Function contains blocks never placed into a chain!\n"
857 << " Bad block: " << getBlockName(*FBI) << "\n";
859 assert(!BadFunc && "Detected problems with the block placement.");
862 // Splice the blocks into place.
863 MachineFunction::iterator InsertPos = F.begin();
864 for (BlockChain::iterator BI = FunctionChain.begin(),
865 BE = FunctionChain.end();
867 DEBUG(dbgs() << (BI == FunctionChain.begin() ? "Placing chain "
869 << getBlockName(*BI) << "\n");
870 if (InsertPos != MachineFunction::iterator(*BI))
871 F.splice(InsertPos, *BI);
875 // Update the terminator of the previous block.
876 if (BI == FunctionChain.begin())
878 MachineBasicBlock *PrevBB = llvm::prior(MachineFunction::iterator(*BI));
880 // FIXME: It would be awesome of updateTerminator would just return rather
881 // than assert when the branch cannot be analyzed in order to remove this
884 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch.
885 if (!TII->AnalyzeBranch(*PrevBB, TBB, FBB, Cond))
886 PrevBB->updateTerminator();
889 // Fixup the last block.
891 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch.
892 if (!TII->AnalyzeBranch(F.back(), TBB, FBB, Cond))
893 F.back().updateTerminator();
896 /// \brief Recursive helper to align a loop and any nested loops.
897 static void AlignLoop(MachineFunction &F, MachineLoop *L, unsigned Align) {
898 // Recurse through nested loops.
899 for (MachineLoop::iterator I = L->begin(), E = L->end(); I != E; ++I)
900 AlignLoop(F, *I, Align);
902 L->getTopBlock()->setAlignment(Align);
905 /// \brief Align loop headers to target preferred alignments.
906 void MachineBlockPlacement::AlignLoops(MachineFunction &F) {
907 if (F.getFunction()->hasFnAttr(Attribute::OptimizeForSize))
910 unsigned Align = TLI->getPrefLoopAlignment();
912 return; // Don't care about loop alignment.
914 for (MachineLoopInfo::iterator I = MLI->begin(), E = MLI->end(); I != E; ++I)
915 AlignLoop(F, *I, Align);
918 bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &F) {
919 // Check for single-block functions and skip them.
920 if (llvm::next(F.begin()) == F.end())
923 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
924 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
925 MLI = &getAnalysis<MachineLoopInfo>();
926 TII = F.getTarget().getInstrInfo();
927 TLI = F.getTarget().getTargetLowering();
928 assert(BlockToChain.empty());
933 BlockToChain.clear();
934 ChainAllocator.DestroyAll();
936 // We always return true as we have no way to track whether the final order
937 // differs from the original order.
942 /// \brief A pass to compute block placement statistics.
944 /// A separate pass to compute interesting statistics for evaluating block
945 /// placement. This is separate from the actual placement pass so that they can
946 /// be computed in the absense of any placement transformations or when using
947 /// alternative placement strategies.
948 class MachineBlockPlacementStats : public MachineFunctionPass {
949 /// \brief A handle to the branch probability pass.
950 const MachineBranchProbabilityInfo *MBPI;
952 /// \brief A handle to the function-wide block frequency pass.
953 const MachineBlockFrequencyInfo *MBFI;
956 static char ID; // Pass identification, replacement for typeid
957 MachineBlockPlacementStats() : MachineFunctionPass(ID) {
958 initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry());
961 bool runOnMachineFunction(MachineFunction &F);
963 void getAnalysisUsage(AnalysisUsage &AU) const {
964 AU.addRequired<MachineBranchProbabilityInfo>();
965 AU.addRequired<MachineBlockFrequencyInfo>();
966 AU.setPreservesAll();
967 MachineFunctionPass::getAnalysisUsage(AU);
970 const char *getPassName() const { return "Block Placement Stats"; }
974 char MachineBlockPlacementStats::ID = 0;
975 INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats",
976 "Basic Block Placement Stats", false, false)
977 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
978 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
979 INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats",
980 "Basic Block Placement Stats", false, false)
982 FunctionPass *llvm::createMachineBlockPlacementStatsPass() {
983 return new MachineBlockPlacementStats();
986 bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) {
987 // Check for single-block functions and skip them.
988 if (llvm::next(F.begin()) == F.end())
991 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
992 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
994 for (MachineFunction::iterator I = F.begin(), E = F.end(); I != E; ++I) {
995 BlockFrequency BlockFreq = MBFI->getBlockFreq(I);
996 Statistic &NumBranches = (I->succ_size() > 1) ? NumCondBranches
998 Statistic &BranchTakenFreq = (I->succ_size() > 1) ? CondBranchTakenFreq
999 : UncondBranchTakenFreq;
1000 for (MachineBasicBlock::succ_iterator SI = I->succ_begin(),
1003 // Skip if this successor is a fallthrough.
1004 if (I->isLayoutSuccessor(*SI))
1007 BlockFrequency EdgeFreq = BlockFreq * MBPI->getEdgeProbability(I, *SI);
1009 BranchTakenFreq += EdgeFreq.getFrequency();