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 #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 *>::iterator iterator;
107 /// \brief Beginning of blocks within the chain.
108 iterator begin() { return Blocks.begin(); }
110 /// \brief End of blocks within the chain.
111 iterator end() { 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;
145 /// \brief Dump the blocks in this chain.
146 void dump() LLVM_ATTRIBUTE_USED {
147 for (iterator I = begin(), E = end(); I != E; ++I)
152 /// \brief Count of predecessors within the loop currently being processed.
154 /// This count is updated at each loop we process to represent the number of
155 /// in-loop predecessors of this chain.
156 unsigned LoopPredecessors;
161 class MachineBlockPlacement : public MachineFunctionPass {
162 /// \brief A typedef for a block filter set.
163 typedef SmallPtrSet<MachineBasicBlock *, 16> BlockFilterSet;
165 /// \brief A handle to the branch probability pass.
166 const MachineBranchProbabilityInfo *MBPI;
168 /// \brief A handle to the function-wide block frequency pass.
169 const MachineBlockFrequencyInfo *MBFI;
171 /// \brief A handle to the loop info.
172 const MachineLoopInfo *MLI;
174 /// \brief A handle to the target's instruction info.
175 const TargetInstrInfo *TII;
177 /// \brief A handle to the target's lowering info.
178 const TargetLowering *TLI;
180 /// \brief Allocator and owner of BlockChain structures.
182 /// We build BlockChains lazily by merging together high probability BB
183 /// sequences according to the "Algo2" in the paper mentioned at the top of
184 /// the file. To reduce malloc traffic, we allocate them using this slab-like
185 /// allocator, and destroy them after the pass completes.
186 SpecificBumpPtrAllocator<BlockChain> ChainAllocator;
188 /// \brief Function wide BasicBlock to BlockChain mapping.
190 /// This mapping allows efficiently moving from any given basic block to the
191 /// BlockChain it participates in, if any. We use it to, among other things,
192 /// allow implicitly defining edges between chains as the existing edges
193 /// between basic blocks.
194 DenseMap<MachineBasicBlock *, BlockChain *> BlockToChain;
196 void markChainSuccessors(BlockChain &Chain,
197 MachineBasicBlock *LoopHeaderBB,
198 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
199 const BlockFilterSet *BlockFilter = 0);
200 MachineBasicBlock *selectBestSuccessor(MachineBasicBlock *BB,
202 const BlockFilterSet *BlockFilter);
203 MachineBasicBlock *selectBestCandidateBlock(
204 BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList,
205 const BlockFilterSet *BlockFilter);
206 MachineBasicBlock *getFirstUnplacedBlock(
208 const BlockChain &PlacedChain,
209 MachineFunction::iterator &PrevUnplacedBlockIt,
210 const BlockFilterSet *BlockFilter);
211 void buildChain(MachineBasicBlock *BB, BlockChain &Chain,
212 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
213 const BlockFilterSet *BlockFilter = 0);
214 MachineBasicBlock *findBestLoopTop(MachineLoop &L,
215 const BlockFilterSet &LoopBlockSet);
216 MachineBasicBlock *findBestLoopExit(MachineFunction &F,
218 const BlockFilterSet &LoopBlockSet);
219 void buildLoopChains(MachineFunction &F, MachineLoop &L);
220 void rotateLoop(BlockChain &LoopChain, MachineBasicBlock *ExitingBB,
221 const BlockFilterSet &LoopBlockSet);
222 void buildCFGChains(MachineFunction &F);
225 static char ID; // Pass identification, replacement for typeid
226 MachineBlockPlacement() : MachineFunctionPass(ID) {
227 initializeMachineBlockPlacementPass(*PassRegistry::getPassRegistry());
230 bool runOnMachineFunction(MachineFunction &F);
232 void getAnalysisUsage(AnalysisUsage &AU) const {
233 AU.addRequired<MachineBranchProbabilityInfo>();
234 AU.addRequired<MachineBlockFrequencyInfo>();
235 AU.addRequired<MachineLoopInfo>();
236 MachineFunctionPass::getAnalysisUsage(AU);
241 char MachineBlockPlacement::ID = 0;
242 char &llvm::MachineBlockPlacementID = MachineBlockPlacement::ID;
243 INITIALIZE_PASS_BEGIN(MachineBlockPlacement, "block-placement2",
244 "Branch Probability Basic Block Placement", false, false)
245 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
246 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
247 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
248 INITIALIZE_PASS_END(MachineBlockPlacement, "block-placement2",
249 "Branch Probability Basic Block Placement", false, false)
252 /// \brief Helper to print the name of a MBB.
254 /// Only used by debug logging.
255 static std::string getBlockName(MachineBasicBlock *BB) {
257 raw_string_ostream OS(Result);
258 OS << "BB#" << BB->getNumber()
259 << " (derived from LLVM BB '" << BB->getName() << "')";
264 /// \brief Helper to print the number of a MBB.
266 /// Only used by debug logging.
267 static std::string getBlockNum(MachineBasicBlock *BB) {
269 raw_string_ostream OS(Result);
270 OS << "BB#" << BB->getNumber();
276 /// \brief Mark a chain's successors as having one fewer preds.
278 /// When a chain is being merged into the "placed" chain, this routine will
279 /// quickly walk the successors of each block in the chain and mark them as
280 /// having one fewer active predecessor. It also adds any successors of this
281 /// chain which reach the zero-predecessor state to the worklist passed in.
282 void MachineBlockPlacement::markChainSuccessors(
284 MachineBasicBlock *LoopHeaderBB,
285 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
286 const BlockFilterSet *BlockFilter) {
287 // Walk all the blocks in this chain, marking their successors as having
288 // a predecessor placed.
289 for (BlockChain::iterator CBI = Chain.begin(), CBE = Chain.end();
291 // Add any successors for which this is the only un-placed in-loop
292 // predecessor to the worklist as a viable candidate for CFG-neutral
293 // placement. No subsequent placement of this block will violate the CFG
294 // shape, so we get to use heuristics to choose a favorable placement.
295 for (MachineBasicBlock::succ_iterator SI = (*CBI)->succ_begin(),
296 SE = (*CBI)->succ_end();
298 if (BlockFilter && !BlockFilter->count(*SI))
300 BlockChain &SuccChain = *BlockToChain[*SI];
301 // Disregard edges within a fixed chain, or edges to the loop header.
302 if (&Chain == &SuccChain || *SI == LoopHeaderBB)
305 // This is a cross-chain edge that is within the loop, so decrement the
306 // loop predecessor count of the destination chain.
307 if (SuccChain.LoopPredecessors > 0 && --SuccChain.LoopPredecessors == 0)
308 BlockWorkList.push_back(*SuccChain.begin());
313 /// \brief Select the best successor for a block.
315 /// This looks across all successors of a particular block and attempts to
316 /// select the "best" one to be the layout successor. It only considers direct
317 /// successors which also pass the block filter. It will attempt to avoid
318 /// breaking CFG structure, but cave and break such structures in the case of
319 /// very hot successor edges.
321 /// \returns The best successor block found, or null if none are viable.
322 MachineBasicBlock *MachineBlockPlacement::selectBestSuccessor(
323 MachineBasicBlock *BB, BlockChain &Chain,
324 const BlockFilterSet *BlockFilter) {
325 const BranchProbability HotProb(4, 5); // 80%
327 MachineBasicBlock *BestSucc = 0;
328 // FIXME: Due to the performance of the probability and weight routines in
329 // the MBPI analysis, we manually compute probabilities using the edge
330 // weights. This is suboptimal as it means that the somewhat subtle
331 // definition of edge weight semantics is encoded here as well. We should
332 // improve the MBPI interface to efficiently support query patterns such as
334 uint32_t BestWeight = 0;
335 uint32_t WeightScale = 0;
336 uint32_t SumWeight = MBPI->getSumForBlock(BB, WeightScale);
337 DEBUG(dbgs() << "Attempting merge from: " << getBlockName(BB) << "\n");
338 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
341 if (BlockFilter && !BlockFilter->count(*SI))
343 BlockChain &SuccChain = *BlockToChain[*SI];
344 if (&SuccChain == &Chain) {
345 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> Already merged!\n");
348 if (*SI != *SuccChain.begin()) {
349 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> Mid chain!\n");
353 uint32_t SuccWeight = MBPI->getEdgeWeight(BB, *SI);
354 BranchProbability SuccProb(SuccWeight / WeightScale, SumWeight);
356 // Only consider successors which are either "hot", or wouldn't violate
357 // any CFG constraints.
358 if (SuccChain.LoopPredecessors != 0) {
359 if (SuccProb < HotProb) {
360 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> CFG conflict\n");
364 // Make sure that a hot successor doesn't have a globally more important
366 BlockFrequency CandidateEdgeFreq
367 = MBFI->getBlockFreq(BB) * SuccProb * HotProb.getCompl();
368 bool BadCFGConflict = false;
369 for (MachineBasicBlock::pred_iterator PI = (*SI)->pred_begin(),
370 PE = (*SI)->pred_end();
372 if (*PI == *SI || (BlockFilter && !BlockFilter->count(*PI)) ||
373 BlockToChain[*PI] == &Chain)
375 BlockFrequency PredEdgeFreq
376 = MBFI->getBlockFreq(*PI) * MBPI->getEdgeProbability(*PI, *SI);
377 if (PredEdgeFreq >= CandidateEdgeFreq) {
378 BadCFGConflict = true;
382 if (BadCFGConflict) {
383 DEBUG(dbgs() << " " << getBlockName(*SI)
384 << " -> non-cold CFG conflict\n");
389 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> " << SuccProb
391 << (SuccChain.LoopPredecessors != 0 ? " (CFG break)" : "")
393 if (BestSucc && BestWeight >= SuccWeight)
396 BestWeight = SuccWeight;
402 /// \brief Predicate struct to detect blocks already placed.
403 class IsBlockPlaced {
404 const BlockChain &PlacedChain;
405 const BlockToChainMapType &BlockToChain;
408 IsBlockPlaced(const BlockChain &PlacedChain,
409 const BlockToChainMapType &BlockToChain)
410 : PlacedChain(PlacedChain), BlockToChain(BlockToChain) {}
412 bool operator()(MachineBasicBlock *BB) const {
413 return BlockToChain.lookup(BB) == &PlacedChain;
418 /// \brief Select the best block from a worklist.
420 /// This looks through the provided worklist as a list of candidate basic
421 /// blocks and select the most profitable one to place. The definition of
422 /// profitable only really makes sense in the context of a loop. This returns
423 /// the most frequently visited block in the worklist, which in the case of
424 /// a loop, is the one most desirable to be physically close to the rest of the
425 /// loop body in order to improve icache behavior.
427 /// \returns The best block found, or null if none are viable.
428 MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock(
429 BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList,
430 const BlockFilterSet *BlockFilter) {
431 // Once we need to walk the worklist looking for a candidate, cleanup the
432 // worklist of already placed entries.
433 // FIXME: If this shows up on profiles, it could be folded (at the cost of
434 // some code complexity) into the loop below.
435 WorkList.erase(std::remove_if(WorkList.begin(), WorkList.end(),
436 IsBlockPlaced(Chain, BlockToChain)),
439 MachineBasicBlock *BestBlock = 0;
440 BlockFrequency BestFreq;
441 for (SmallVectorImpl<MachineBasicBlock *>::iterator WBI = WorkList.begin(),
442 WBE = WorkList.end();
444 BlockChain &SuccChain = *BlockToChain[*WBI];
445 if (&SuccChain == &Chain) {
446 DEBUG(dbgs() << " " << getBlockName(*WBI)
447 << " -> Already merged!\n");
450 assert(SuccChain.LoopPredecessors == 0 && "Found CFG-violating block");
452 BlockFrequency CandidateFreq = MBFI->getBlockFreq(*WBI);
453 DEBUG(dbgs() << " " << getBlockName(*WBI) << " -> " << CandidateFreq
455 if (BestBlock && BestFreq >= CandidateFreq)
458 BestFreq = CandidateFreq;
463 /// \brief Retrieve the first unplaced basic block.
465 /// This routine is called when we are unable to use the CFG to walk through
466 /// all of the basic blocks and form a chain due to unnatural loops in the CFG.
467 /// We walk through the function's blocks in order, starting from the
468 /// LastUnplacedBlockIt. We update this iterator on each call to avoid
469 /// re-scanning the entire sequence on repeated calls to this routine.
470 MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock(
471 MachineFunction &F, const BlockChain &PlacedChain,
472 MachineFunction::iterator &PrevUnplacedBlockIt,
473 const BlockFilterSet *BlockFilter) {
474 for (MachineFunction::iterator I = PrevUnplacedBlockIt, E = F.end(); I != E;
476 if (BlockFilter && !BlockFilter->count(I))
478 if (BlockToChain[I] != &PlacedChain) {
479 PrevUnplacedBlockIt = I;
480 // Now select the head of the chain to which the unplaced block belongs
481 // as the block to place. This will force the entire chain to be placed,
482 // and satisfies the requirements of merging chains.
483 return *BlockToChain[I]->begin();
489 void MachineBlockPlacement::buildChain(
490 MachineBasicBlock *BB,
492 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
493 const BlockFilterSet *BlockFilter) {
495 assert(BlockToChain[BB] == &Chain);
496 MachineFunction &F = *BB->getParent();
497 MachineFunction::iterator PrevUnplacedBlockIt = F.begin();
499 MachineBasicBlock *LoopHeaderBB = BB;
500 markChainSuccessors(Chain, LoopHeaderBB, BlockWorkList, BlockFilter);
501 BB = *llvm::prior(Chain.end());
504 assert(BlockToChain[BB] == &Chain);
505 assert(*llvm::prior(Chain.end()) == BB);
506 MachineBasicBlock *BestSucc = 0;
508 // Look for the best viable successor if there is one to place immediately
510 BestSucc = selectBestSuccessor(BB, Chain, BlockFilter);
512 // If an immediate successor isn't available, look for the best viable
513 // block among those we've identified as not violating the loop's CFG at
514 // this point. This won't be a fallthrough, but it will increase locality.
516 BestSucc = selectBestCandidateBlock(Chain, BlockWorkList, BlockFilter);
519 BestSucc = getFirstUnplacedBlock(F, Chain, PrevUnplacedBlockIt,
524 DEBUG(dbgs() << "Unnatural loop CFG detected, forcibly merging the "
525 "layout successor until the CFG reduces\n");
528 // Place this block, updating the datastructures to reflect its placement.
529 BlockChain &SuccChain = *BlockToChain[BestSucc];
530 // Zero out LoopPredecessors for the successor we're about to merge in case
531 // we selected a successor that didn't fit naturally into the CFG.
532 SuccChain.LoopPredecessors = 0;
533 DEBUG(dbgs() << "Merging from " << getBlockNum(BB)
534 << " to " << getBlockNum(BestSucc) << "\n");
535 markChainSuccessors(SuccChain, LoopHeaderBB, BlockWorkList, BlockFilter);
536 Chain.merge(BestSucc, &SuccChain);
537 BB = *llvm::prior(Chain.end());
540 DEBUG(dbgs() << "Finished forming chain for header block "
541 << getBlockNum(*Chain.begin()) << "\n");
544 /// \brief Find the best loop top block for layout.
546 /// Look for a block which is strictly better than the loop header for laying
547 /// out at the top of the loop. This looks for one and only one pattern:
548 /// a latch block with no conditional exit. This block will cause a conditional
549 /// jump around it or will be the bottom of the loop if we lay it out in place,
550 /// but if it it doesn't end up at the bottom of the loop for any reason,
551 /// rotation alone won't fix it. Because such a block will always result in an
552 /// unconditional jump (for the backedge) rotating it in front of the loop
553 /// header is always profitable.
555 MachineBlockPlacement::findBestLoopTop(MachineLoop &L,
556 const BlockFilterSet &LoopBlockSet) {
557 // Check that the header hasn't been fused with a preheader block due to
558 // crazy branches. If it has, we need to start with the header at the top to
559 // prevent pulling the preheader into the loop body.
560 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
561 if (!LoopBlockSet.count(*HeaderChain.begin()))
562 return L.getHeader();
564 DEBUG(dbgs() << "Finding best loop top for: "
565 << getBlockName(L.getHeader()) << "\n");
567 BlockFrequency BestPredFreq;
568 MachineBasicBlock *BestPred = 0;
569 for (MachineBasicBlock::pred_iterator PI = L.getHeader()->pred_begin(),
570 PE = L.getHeader()->pred_end();
572 MachineBasicBlock *Pred = *PI;
573 if (!LoopBlockSet.count(Pred))
575 DEBUG(dbgs() << " header pred: " << getBlockName(Pred) << ", "
576 << Pred->succ_size() << " successors, "
577 << MBFI->getBlockFreq(Pred) << " freq\n");
578 if (Pred->succ_size() > 1)
581 BlockFrequency PredFreq = MBFI->getBlockFreq(Pred);
582 if (!BestPred || PredFreq > BestPredFreq ||
583 (!(PredFreq < BestPredFreq) &&
584 Pred->isLayoutSuccessor(L.getHeader()))) {
586 BestPredFreq = PredFreq;
590 // If no direct predecessor is fine, just use the loop header.
592 return L.getHeader();
594 // Walk backwards through any straight line of predecessors.
595 while (BestPred->pred_size() == 1 &&
596 (*BestPred->pred_begin())->succ_size() == 1 &&
597 *BestPred->pred_begin() != L.getHeader())
598 BestPred = *BestPred->pred_begin();
600 DEBUG(dbgs() << " final top: " << getBlockName(BestPred) << "\n");
605 /// \brief Find the best loop exiting block for layout.
607 /// This routine implements the logic to analyze the loop looking for the best
608 /// block to layout at the top of the loop. Typically this is done to maximize
609 /// fallthrough opportunities.
611 MachineBlockPlacement::findBestLoopExit(MachineFunction &F,
613 const BlockFilterSet &LoopBlockSet) {
614 // We don't want to layout the loop linearly in all cases. If the loop header
615 // is just a normal basic block in the loop, we want to look for what block
616 // within the loop is the best one to layout at the top. However, if the loop
617 // header has be pre-merged into a chain due to predecessors not having
618 // analyzable branches, *and* the predecessor it is merged with is *not* part
619 // of the loop, rotating the header into the middle of the loop will create
620 // a non-contiguous range of blocks which is Very Bad. So start with the
621 // header and only rotate if safe.
622 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
623 if (!LoopBlockSet.count(*HeaderChain.begin()))
626 BlockFrequency BestExitEdgeFreq;
627 unsigned BestExitLoopDepth = 0;
628 MachineBasicBlock *ExitingBB = 0;
629 // If there are exits to outer loops, loop rotation can severely limit
630 // fallthrough opportunites unless it selects such an exit. Keep a set of
631 // blocks where rotating to exit with that block will reach an outer loop.
632 SmallPtrSet<MachineBasicBlock *, 4> BlocksExitingToOuterLoop;
634 DEBUG(dbgs() << "Finding best loop exit for: "
635 << getBlockName(L.getHeader()) << "\n");
636 for (MachineLoop::block_iterator I = L.block_begin(),
639 BlockChain &Chain = *BlockToChain[*I];
640 // Ensure that this block is at the end of a chain; otherwise it could be
641 // mid-way through an inner loop or a successor of an analyzable branch.
642 if (*I != *llvm::prior(Chain.end()))
645 // Now walk the successors. We need to establish whether this has a viable
646 // exiting successor and whether it has a viable non-exiting successor.
647 // We store the old exiting state and restore it if a viable looping
648 // successor isn't found.
649 MachineBasicBlock *OldExitingBB = ExitingBB;
650 BlockFrequency OldBestExitEdgeFreq = BestExitEdgeFreq;
651 bool HasLoopingSucc = false;
652 // FIXME: Due to the performance of the probability and weight routines in
653 // the MBPI analysis, we use the internal weights and manually compute the
654 // probabilities to avoid quadratic behavior.
655 uint32_t WeightScale = 0;
656 uint32_t SumWeight = MBPI->getSumForBlock(*I, WeightScale);
657 for (MachineBasicBlock::succ_iterator SI = (*I)->succ_begin(),
658 SE = (*I)->succ_end();
660 if ((*SI)->isLandingPad())
664 BlockChain &SuccChain = *BlockToChain[*SI];
665 // Don't split chains, either this chain or the successor's chain.
666 if (&Chain == &SuccChain) {
667 DEBUG(dbgs() << " exiting: " << getBlockName(*I) << " -> "
668 << getBlockName(*SI) << " (chain conflict)\n");
672 uint32_t SuccWeight = MBPI->getEdgeWeight(*I, *SI);
673 if (LoopBlockSet.count(*SI)) {
674 DEBUG(dbgs() << " looping: " << getBlockName(*I) << " -> "
675 << getBlockName(*SI) << " (" << SuccWeight << ")\n");
676 HasLoopingSucc = true;
680 unsigned SuccLoopDepth = 0;
681 if (MachineLoop *ExitLoop = MLI->getLoopFor(*SI)) {
682 SuccLoopDepth = ExitLoop->getLoopDepth();
683 if (ExitLoop->contains(&L))
684 BlocksExitingToOuterLoop.insert(*I);
687 BranchProbability SuccProb(SuccWeight / WeightScale, SumWeight);
688 BlockFrequency ExitEdgeFreq = MBFI->getBlockFreq(*I) * SuccProb;
689 DEBUG(dbgs() << " exiting: " << getBlockName(*I) << " -> "
690 << getBlockName(*SI) << " [L:" << SuccLoopDepth
691 << "] (" << ExitEdgeFreq << ")\n");
692 // Note that we slightly bias this toward an existing layout successor to
693 // retain incoming order in the absence of better information.
694 // FIXME: Should we bias this more strongly? It's pretty weak.
695 if (!ExitingBB || BestExitLoopDepth < SuccLoopDepth ||
696 ExitEdgeFreq > BestExitEdgeFreq ||
697 ((*I)->isLayoutSuccessor(*SI) &&
698 !(ExitEdgeFreq < BestExitEdgeFreq))) {
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 == *llvm::prior(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 = *llvm::prior(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(), llvm::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 = 0;
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 = 0, *FBB = 0; // For AnalyzeBranch.
889 if (!TII->AnalyzeBranch(*BB, TBB, FBB, Cond) || !FI->canFallThrough())
892 MachineFunction::iterator NextFI(llvm::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, 0);
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);
940 typedef SmallPtrSet<MachineBasicBlock *, 16> FunctionBlockSetType;
942 // Crash at the end so we get all of the debugging output first.
943 bool BadFunc = false;
944 FunctionBlockSetType FunctionBlockSet;
945 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
946 FunctionBlockSet.insert(FI);
948 for (BlockChain::iterator BCI = FunctionChain.begin(),
949 BCE = FunctionChain.end();
951 if (!FunctionBlockSet.erase(*BCI)) {
953 dbgs() << "Function chain contains a block not in the function!\n"
954 << " Bad block: " << getBlockName(*BCI) << "\n";
957 if (!FunctionBlockSet.empty()) {
959 for (FunctionBlockSetType::iterator FBI = FunctionBlockSet.begin(),
960 FBE = FunctionBlockSet.end();
962 dbgs() << "Function contains blocks never placed into a chain!\n"
963 << " Bad block: " << getBlockName(*FBI) << "\n";
965 assert(!BadFunc && "Detected problems with the block placement.");
968 // Splice the blocks into place.
969 MachineFunction::iterator InsertPos = F.begin();
970 for (BlockChain::iterator BI = FunctionChain.begin(),
971 BE = FunctionChain.end();
973 DEBUG(dbgs() << (BI == FunctionChain.begin() ? "Placing chain "
975 << getBlockName(*BI) << "\n");
976 if (InsertPos != MachineFunction::iterator(*BI))
977 F.splice(InsertPos, *BI);
981 // Update the terminator of the previous block.
982 if (BI == FunctionChain.begin())
984 MachineBasicBlock *PrevBB = llvm::prior(MachineFunction::iterator(*BI));
986 // FIXME: It would be awesome of updateTerminator would just return rather
987 // than assert when the branch cannot be analyzed in order to remove this
990 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch.
991 if (!TII->AnalyzeBranch(*PrevBB, TBB, FBB, Cond))
992 PrevBB->updateTerminator();
995 // Fixup the last block.
997 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch.
998 if (!TII->AnalyzeBranch(F.back(), TBB, FBB, Cond))
999 F.back().updateTerminator();
1001 // Walk through the backedges of the function now that we have fully laid out
1002 // the basic blocks and align the destination of each backedge. We don't rely
1003 // on the loop info here so that we can align backedges in unnatural CFGs and
1004 // backedges that were introduced purely because of the loop rotations done
1005 // during this layout pass.
1006 // FIXME: This isn't quite right, we shouldn't align backedges that result
1007 // from blocks being sunken below the exit block for the function.
1008 if (F.getFunction()->hasFnAttr(Attribute::OptimizeForSize))
1010 unsigned Align = TLI->getPrefLoopAlignment();
1012 return; // Don't care about loop alignment.
1014 SmallPtrSet<MachineBasicBlock *, 16> PreviousBlocks;
1015 for (BlockChain::iterator BI = FunctionChain.begin(),
1016 BE = FunctionChain.end();
1018 PreviousBlocks.insert(*BI);
1019 // Set alignment on the destination of all the back edges in the new
1021 for (MachineBasicBlock::succ_iterator SI = (*BI)->succ_begin(),
1022 SE = (*BI)->succ_end();
1024 if (PreviousBlocks.count(*SI))
1025 (*SI)->setAlignment(Align);
1029 bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &F) {
1030 // Check for single-block functions and skip them.
1031 if (llvm::next(F.begin()) == F.end())
1034 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
1035 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
1036 MLI = &getAnalysis<MachineLoopInfo>();
1037 TII = F.getTarget().getInstrInfo();
1038 TLI = F.getTarget().getTargetLowering();
1039 assert(BlockToChain.empty());
1043 BlockToChain.clear();
1044 ChainAllocator.DestroyAll();
1046 // We always return true as we have no way to track whether the final order
1047 // differs from the original order.
1052 /// \brief A pass to compute block placement statistics.
1054 /// A separate pass to compute interesting statistics for evaluating block
1055 /// placement. This is separate from the actual placement pass so that they can
1056 /// be computed in the absence of any placement transformations or when using
1057 /// alternative placement strategies.
1058 class MachineBlockPlacementStats : public MachineFunctionPass {
1059 /// \brief A handle to the branch probability pass.
1060 const MachineBranchProbabilityInfo *MBPI;
1062 /// \brief A handle to the function-wide block frequency pass.
1063 const MachineBlockFrequencyInfo *MBFI;
1066 static char ID; // Pass identification, replacement for typeid
1067 MachineBlockPlacementStats() : MachineFunctionPass(ID) {
1068 initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry());
1071 bool runOnMachineFunction(MachineFunction &F);
1073 void getAnalysisUsage(AnalysisUsage &AU) const {
1074 AU.addRequired<MachineBranchProbabilityInfo>();
1075 AU.addRequired<MachineBlockFrequencyInfo>();
1076 AU.setPreservesAll();
1077 MachineFunctionPass::getAnalysisUsage(AU);
1082 char MachineBlockPlacementStats::ID = 0;
1083 char &llvm::MachineBlockPlacementStatsID = MachineBlockPlacementStats::ID;
1084 INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats",
1085 "Basic Block Placement Stats", false, false)
1086 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
1087 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
1088 INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats",
1089 "Basic Block Placement Stats", false, false)
1091 bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) {
1092 // Check for single-block functions and skip them.
1093 if (llvm::next(F.begin()) == F.end())
1096 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
1097 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
1099 for (MachineFunction::iterator I = F.begin(), E = F.end(); I != E; ++I) {
1100 BlockFrequency BlockFreq = MBFI->getBlockFreq(I);
1101 Statistic &NumBranches = (I->succ_size() > 1) ? NumCondBranches
1102 : NumUncondBranches;
1103 Statistic &BranchTakenFreq = (I->succ_size() > 1) ? CondBranchTakenFreq
1104 : UncondBranchTakenFreq;
1105 for (MachineBasicBlock::succ_iterator SI = I->succ_begin(),
1108 // Skip if this successor is a fallthrough.
1109 if (I->isLayoutSuccessor(*SI))
1112 BlockFrequency EdgeFreq = BlockFreq * MBPI->getEdgeProbability(I, *SI);
1114 BranchTakenFreq += EdgeFreq.getFrequency();