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 and code locality. We also can use a block chain to represent
67 /// a sequence of basic blocks which have some external (correctness)
68 /// requirement for sequential layout.
70 /// Chains can be built around a single basic block and can be merged to grow
71 /// them. They participate in a block-to-chain mapping, which is updated
72 /// automatically as chains are merged together.
74 /// \brief The sequence of blocks belonging to this chain.
76 /// This is the sequence of blocks for a particular chain. These will be laid
77 /// out in-order within the function.
78 SmallVector<MachineBasicBlock *, 4> Blocks;
80 /// \brief A handle to the function-wide basic block to block chain mapping.
82 /// This is retained in each block chain to simplify the computation of child
83 /// block chains for SCC-formation and iteration. We store the edges to child
84 /// basic blocks, and map them back to their associated chains using this
86 BlockToChainMapType &BlockToChain;
89 /// \brief Construct a new BlockChain.
91 /// This builds a new block chain representing a single basic block in the
92 /// function. It also registers itself as the chain that block participates
93 /// in with the BlockToChain mapping.
94 BlockChain(BlockToChainMapType &BlockToChain, MachineBasicBlock *BB)
95 : Blocks(1, BB), BlockToChain(BlockToChain), LoopPredecessors(0) {
96 assert(BB && "Cannot create a chain with a null basic block");
97 BlockToChain[BB] = this;
100 /// \brief Iterator over blocks within the chain.
101 typedef SmallVectorImpl<MachineBasicBlock *>::iterator iterator;
103 /// \brief Beginning of blocks within the chain.
104 iterator begin() { return Blocks.begin(); }
106 /// \brief End of blocks within the chain.
107 iterator end() { return Blocks.end(); }
109 /// \brief Merge a block chain into this one.
111 /// This routine merges a block chain into this one. It takes care of forming
112 /// a contiguous sequence of basic blocks, updating the edge list, and
113 /// updating the block -> chain mapping. It does not free or tear down the
114 /// old chain, but the old chain's block list is no longer valid.
115 void merge(MachineBasicBlock *BB, BlockChain *Chain) {
117 assert(!Blocks.empty());
119 // Fast path in case we don't have a chain already.
121 assert(!BlockToChain[BB]);
122 Blocks.push_back(BB);
123 BlockToChain[BB] = this;
127 assert(BB == *Chain->begin());
128 assert(Chain->begin() != Chain->end());
130 // Update the incoming blocks to point to this chain, and add them to the
132 for (BlockChain::iterator BI = Chain->begin(), BE = Chain->end();
134 Blocks.push_back(*BI);
135 assert(BlockToChain[*BI] == Chain && "Incoming blocks not in chain");
136 BlockToChain[*BI] = this;
141 /// \brief Dump the blocks in this chain.
142 void dump() LLVM_ATTRIBUTE_USED {
143 for (iterator I = begin(), E = end(); I != E; ++I)
148 /// \brief Count of predecessors within the loop currently being processed.
150 /// This count is updated at each loop we process to represent the number of
151 /// in-loop predecessors of this chain.
152 unsigned LoopPredecessors;
157 class MachineBlockPlacement : public MachineFunctionPass {
158 /// \brief A typedef for a block filter set.
159 typedef SmallPtrSet<MachineBasicBlock *, 16> BlockFilterSet;
161 /// \brief A handle to the branch probability pass.
162 const MachineBranchProbabilityInfo *MBPI;
164 /// \brief A handle to the function-wide block frequency pass.
165 const MachineBlockFrequencyInfo *MBFI;
167 /// \brief A handle to the loop info.
168 const MachineLoopInfo *MLI;
170 /// \brief A handle to the target's instruction info.
171 const TargetInstrInfo *TII;
173 /// \brief A handle to the target's lowering info.
174 const TargetLowering *TLI;
176 /// \brief Allocator and owner of BlockChain structures.
178 /// We build BlockChains lazily while processing the loop structure of
179 /// a function. To reduce malloc traffic, we allocate them using this
180 /// slab-like allocator, and destroy them after the pass completes. An
181 /// important guarantee is that this allocator produces stable pointers to
183 SpecificBumpPtrAllocator<BlockChain> ChainAllocator;
185 /// \brief Function wide BasicBlock to BlockChain mapping.
187 /// This mapping allows efficiently moving from any given basic block to the
188 /// BlockChain it participates in, if any. We use it to, among other things,
189 /// allow implicitly defining edges between chains as the existing edges
190 /// between basic blocks.
191 DenseMap<MachineBasicBlock *, BlockChain *> BlockToChain;
193 void markChainSuccessors(BlockChain &Chain,
194 MachineBasicBlock *LoopHeaderBB,
195 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
196 const BlockFilterSet *BlockFilter = 0);
197 MachineBasicBlock *selectBestSuccessor(MachineBasicBlock *BB,
199 const BlockFilterSet *BlockFilter);
200 MachineBasicBlock *selectBestCandidateBlock(
201 BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList,
202 const BlockFilterSet *BlockFilter);
203 MachineBasicBlock *getFirstUnplacedBlock(
205 const BlockChain &PlacedChain,
206 MachineFunction::iterator &PrevUnplacedBlockIt,
207 const BlockFilterSet *BlockFilter);
208 void buildChain(MachineBasicBlock *BB, BlockChain &Chain,
209 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
210 const BlockFilterSet *BlockFilter = 0);
211 MachineBasicBlock *findBestLoopTop(MachineLoop &L,
212 const BlockFilterSet &LoopBlockSet);
213 MachineBasicBlock *findBestLoopExit(MachineFunction &F,
215 const BlockFilterSet &LoopBlockSet);
216 void buildLoopChains(MachineFunction &F, MachineLoop &L);
217 void rotateLoop(BlockChain &LoopChain, MachineBasicBlock *ExitingBB,
218 const BlockFilterSet &LoopBlockSet);
219 void buildCFGChains(MachineFunction &F);
222 static char ID; // Pass identification, replacement for typeid
223 MachineBlockPlacement() : MachineFunctionPass(ID) {
224 initializeMachineBlockPlacementPass(*PassRegistry::getPassRegistry());
227 bool runOnMachineFunction(MachineFunction &F);
229 void getAnalysisUsage(AnalysisUsage &AU) const {
230 AU.addRequired<MachineBranchProbabilityInfo>();
231 AU.addRequired<MachineBlockFrequencyInfo>();
232 AU.addRequired<MachineLoopInfo>();
233 MachineFunctionPass::getAnalysisUsage(AU);
238 char MachineBlockPlacement::ID = 0;
239 char &llvm::MachineBlockPlacementID = MachineBlockPlacement::ID;
240 INITIALIZE_PASS_BEGIN(MachineBlockPlacement, "block-placement2",
241 "Branch Probability Basic Block Placement", false, false)
242 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
243 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
244 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
245 INITIALIZE_PASS_END(MachineBlockPlacement, "block-placement2",
246 "Branch Probability Basic Block Placement", false, false)
249 /// \brief Helper to print the name of a MBB.
251 /// Only used by debug logging.
252 static std::string getBlockName(MachineBasicBlock *BB) {
254 raw_string_ostream OS(Result);
255 OS << "BB#" << BB->getNumber()
256 << " (derived from LLVM BB '" << BB->getName() << "')";
261 /// \brief Helper to print the number of a MBB.
263 /// Only used by debug logging.
264 static std::string getBlockNum(MachineBasicBlock *BB) {
266 raw_string_ostream OS(Result);
267 OS << "BB#" << BB->getNumber();
273 /// \brief Mark a chain's successors as having one fewer preds.
275 /// When a chain is being merged into the "placed" chain, this routine will
276 /// quickly walk the successors of each block in the chain and mark them as
277 /// having one fewer active predecessor. It also adds any successors of this
278 /// chain which reach the zero-predecessor state to the worklist passed in.
279 void MachineBlockPlacement::markChainSuccessors(
281 MachineBasicBlock *LoopHeaderBB,
282 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
283 const BlockFilterSet *BlockFilter) {
284 // Walk all the blocks in this chain, marking their successors as having
285 // a predecessor placed.
286 for (BlockChain::iterator CBI = Chain.begin(), CBE = Chain.end();
288 // Add any successors for which this is the only un-placed in-loop
289 // predecessor to the worklist as a viable candidate for CFG-neutral
290 // placement. No subsequent placement of this block will violate the CFG
291 // shape, so we get to use heuristics to choose a favorable placement.
292 for (MachineBasicBlock::succ_iterator SI = (*CBI)->succ_begin(),
293 SE = (*CBI)->succ_end();
295 if (BlockFilter && !BlockFilter->count(*SI))
297 BlockChain &SuccChain = *BlockToChain[*SI];
298 // Disregard edges within a fixed chain, or edges to the loop header.
299 if (&Chain == &SuccChain || *SI == LoopHeaderBB)
302 // This is a cross-chain edge that is within the loop, so decrement the
303 // loop predecessor count of the destination chain.
304 if (SuccChain.LoopPredecessors > 0 && --SuccChain.LoopPredecessors == 0)
305 BlockWorkList.push_back(*SuccChain.begin());
310 /// \brief Select the best successor for a block.
312 /// This looks across all successors of a particular block and attempts to
313 /// select the "best" one to be the layout successor. It only considers direct
314 /// successors which also pass the block filter. It will attempt to avoid
315 /// breaking CFG structure, but cave and break such structures in the case of
316 /// very hot successor edges.
318 /// \returns The best successor block found, or null if none are viable.
319 MachineBasicBlock *MachineBlockPlacement::selectBestSuccessor(
320 MachineBasicBlock *BB, BlockChain &Chain,
321 const BlockFilterSet *BlockFilter) {
322 const BranchProbability HotProb(4, 5); // 80%
324 MachineBasicBlock *BestSucc = 0;
325 // FIXME: Due to the performance of the probability and weight routines in
326 // the MBPI analysis, we manually compute probabilities using the edge
327 // weights. This is suboptimal as it means that the somewhat subtle
328 // definition of edge weight semantics is encoded here as well. We should
329 // improve the MBPI interface to efficiently support query patterns such as
331 uint32_t BestWeight = 0;
332 uint32_t WeightScale = 0;
333 uint32_t SumWeight = MBPI->getSumForBlock(BB, WeightScale);
334 DEBUG(dbgs() << "Attempting merge from: " << getBlockName(BB) << "\n");
335 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
338 if (BlockFilter && !BlockFilter->count(*SI))
340 BlockChain &SuccChain = *BlockToChain[*SI];
341 if (&SuccChain == &Chain) {
342 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> Already merged!\n");
345 if (*SI != *SuccChain.begin()) {
346 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> Mid chain!\n");
350 uint32_t SuccWeight = MBPI->getEdgeWeight(BB, *SI);
351 BranchProbability SuccProb(SuccWeight / WeightScale, SumWeight);
353 // Only consider successors which are either "hot", or wouldn't violate
354 // any CFG constraints.
355 if (SuccChain.LoopPredecessors != 0) {
356 if (SuccProb < HotProb) {
357 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> CFG conflict\n");
361 // Make sure that a hot successor doesn't have a globally more important
363 BlockFrequency CandidateEdgeFreq
364 = MBFI->getBlockFreq(BB) * SuccProb * HotProb.getCompl();
365 bool BadCFGConflict = false;
366 for (MachineBasicBlock::pred_iterator PI = (*SI)->pred_begin(),
367 PE = (*SI)->pred_end();
369 if (*PI == *SI || (BlockFilter && !BlockFilter->count(*PI)) ||
370 BlockToChain[*PI] == &Chain)
372 BlockFrequency PredEdgeFreq
373 = MBFI->getBlockFreq(*PI) * MBPI->getEdgeProbability(*PI, *SI);
374 if (PredEdgeFreq >= CandidateEdgeFreq) {
375 BadCFGConflict = true;
379 if (BadCFGConflict) {
380 DEBUG(dbgs() << " " << getBlockName(*SI)
381 << " -> non-cold CFG conflict\n");
386 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> " << SuccProb
388 << (SuccChain.LoopPredecessors != 0 ? " (CFG break)" : "")
390 if (BestSucc && BestWeight >= SuccWeight)
393 BestWeight = SuccWeight;
399 /// \brief Predicate struct to detect blocks already placed.
400 class IsBlockPlaced {
401 const BlockChain &PlacedChain;
402 const BlockToChainMapType &BlockToChain;
405 IsBlockPlaced(const BlockChain &PlacedChain,
406 const BlockToChainMapType &BlockToChain)
407 : PlacedChain(PlacedChain), BlockToChain(BlockToChain) {}
409 bool operator()(MachineBasicBlock *BB) const {
410 return BlockToChain.lookup(BB) == &PlacedChain;
415 /// \brief Select the best block from a worklist.
417 /// This looks through the provided worklist as a list of candidate basic
418 /// blocks and select the most profitable one to place. The definition of
419 /// profitable only really makes sense in the context of a loop. This returns
420 /// the most frequently visited block in the worklist, which in the case of
421 /// a loop, is the one most desirable to be physically close to the rest of the
422 /// loop body in order to improve icache behavior.
424 /// \returns The best block found, or null if none are viable.
425 MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock(
426 BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList,
427 const BlockFilterSet *BlockFilter) {
428 // Once we need to walk the worklist looking for a candidate, cleanup the
429 // worklist of already placed entries.
430 // FIXME: If this shows up on profiles, it could be folded (at the cost of
431 // some code complexity) into the loop below.
432 WorkList.erase(std::remove_if(WorkList.begin(), WorkList.end(),
433 IsBlockPlaced(Chain, BlockToChain)),
436 MachineBasicBlock *BestBlock = 0;
437 BlockFrequency BestFreq;
438 for (SmallVectorImpl<MachineBasicBlock *>::iterator WBI = WorkList.begin(),
439 WBE = WorkList.end();
441 BlockChain &SuccChain = *BlockToChain[*WBI];
442 if (&SuccChain == &Chain) {
443 DEBUG(dbgs() << " " << getBlockName(*WBI)
444 << " -> Already merged!\n");
447 assert(SuccChain.LoopPredecessors == 0 && "Found CFG-violating block");
449 BlockFrequency CandidateFreq = MBFI->getBlockFreq(*WBI);
450 DEBUG(dbgs() << " " << getBlockName(*WBI) << " -> " << CandidateFreq
452 if (BestBlock && BestFreq >= CandidateFreq)
455 BestFreq = CandidateFreq;
460 /// \brief Retrieve the first unplaced basic block.
462 /// This routine is called when we are unable to use the CFG to walk through
463 /// all of the basic blocks and form a chain due to unnatural loops in the CFG.
464 /// We walk through the function's blocks in order, starting from the
465 /// LastUnplacedBlockIt. We update this iterator on each call to avoid
466 /// re-scanning the entire sequence on repeated calls to this routine.
467 MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock(
468 MachineFunction &F, const BlockChain &PlacedChain,
469 MachineFunction::iterator &PrevUnplacedBlockIt,
470 const BlockFilterSet *BlockFilter) {
471 for (MachineFunction::iterator I = PrevUnplacedBlockIt, E = F.end(); I != E;
473 if (BlockFilter && !BlockFilter->count(I))
475 if (BlockToChain[I] != &PlacedChain) {
476 PrevUnplacedBlockIt = I;
477 // Now select the head of the chain to which the unplaced block belongs
478 // as the block to place. This will force the entire chain to be placed,
479 // and satisfies the requirements of merging chains.
480 return *BlockToChain[I]->begin();
486 void MachineBlockPlacement::buildChain(
487 MachineBasicBlock *BB,
489 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
490 const BlockFilterSet *BlockFilter) {
492 assert(BlockToChain[BB] == &Chain);
493 MachineFunction &F = *BB->getParent();
494 MachineFunction::iterator PrevUnplacedBlockIt = F.begin();
496 MachineBasicBlock *LoopHeaderBB = BB;
497 markChainSuccessors(Chain, LoopHeaderBB, BlockWorkList, BlockFilter);
498 BB = *llvm::prior(Chain.end());
501 assert(BlockToChain[BB] == &Chain);
502 assert(*llvm::prior(Chain.end()) == BB);
504 // Look for the best viable successor if there is one to place immediately
506 MachineBasicBlock *BestSucc = selectBestSuccessor(BB, Chain, BlockFilter);
508 // If an immediate successor isn't available, look for the best viable
509 // block among those we've identified as not violating the loop's CFG at
510 // this point. This won't be a fallthrough, but it will increase locality.
512 BestSucc = selectBestCandidateBlock(Chain, BlockWorkList, BlockFilter);
515 BestSucc = getFirstUnplacedBlock(F, Chain, PrevUnplacedBlockIt,
520 DEBUG(dbgs() << "Unnatural loop CFG detected, forcibly merging the "
521 "layout successor until the CFG reduces\n");
524 // Place this block, updating the datastructures to reflect its placement.
525 BlockChain &SuccChain = *BlockToChain[BestSucc];
526 // Zero out LoopPredecessors for the successor we're about to merge in case
527 // we selected a successor that didn't fit naturally into the CFG.
528 SuccChain.LoopPredecessors = 0;
529 DEBUG(dbgs() << "Merging from " << getBlockNum(BB)
530 << " to " << getBlockNum(BestSucc) << "\n");
531 markChainSuccessors(SuccChain, LoopHeaderBB, BlockWorkList, BlockFilter);
532 Chain.merge(BestSucc, &SuccChain);
533 BB = *llvm::prior(Chain.end());
536 DEBUG(dbgs() << "Finished forming chain for header block "
537 << getBlockNum(*Chain.begin()) << "\n");
540 /// \brief Find the best loop top block for layout.
542 /// Look for a block which is strictly better than the loop header for laying
543 /// out at the top of the loop. This looks for one and only one pattern:
544 /// a latch block with no conditional exit. This block will cause a conditional
545 /// jump around it or will be the bottom of the loop if we lay it out in place,
546 /// but if it it doesn't end up at the bottom of the loop for any reason,
547 /// rotation alone won't fix it. Because such a block will always result in an
548 /// unconditional jump (for the backedge) rotating it in front of the loop
549 /// header is always profitable.
551 MachineBlockPlacement::findBestLoopTop(MachineLoop &L,
552 const BlockFilterSet &LoopBlockSet) {
553 // Check that the header hasn't been fused with a preheader block due to
554 // crazy branches. If it has, we need to start with the header at the top to
555 // prevent pulling the preheader into the loop body.
556 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
557 if (!LoopBlockSet.count(*HeaderChain.begin()))
558 return L.getHeader();
560 DEBUG(dbgs() << "Finding best loop top for: "
561 << getBlockName(L.getHeader()) << "\n");
563 BlockFrequency BestPredFreq;
564 MachineBasicBlock *BestPred = 0;
565 for (MachineBasicBlock::pred_iterator PI = L.getHeader()->pred_begin(),
566 PE = L.getHeader()->pred_end();
568 MachineBasicBlock *Pred = *PI;
569 if (!LoopBlockSet.count(Pred))
571 DEBUG(dbgs() << " header pred: " << getBlockName(Pred) << ", "
572 << Pred->succ_size() << " successors, "
573 << MBFI->getBlockFreq(Pred) << " freq\n");
574 if (Pred->succ_size() > 1)
577 BlockFrequency PredFreq = MBFI->getBlockFreq(Pred);
578 if (!BestPred || PredFreq > BestPredFreq ||
579 (!(PredFreq < BestPredFreq) &&
580 Pred->isLayoutSuccessor(L.getHeader()))) {
582 BestPredFreq = PredFreq;
586 // If no direct predecessor is fine, just use the loop header.
588 return L.getHeader();
590 // Walk backwards through any straight line of predecessors.
591 while (BestPred->pred_size() == 1 &&
592 (*BestPred->pred_begin())->succ_size() == 1 &&
593 *BestPred->pred_begin() != L.getHeader())
594 BestPred = *BestPred->pred_begin();
596 DEBUG(dbgs() << " final top: " << getBlockName(BestPred) << "\n");
601 /// \brief Find the best loop exiting block for layout.
603 /// This routine implements the logic to analyze the loop looking for the best
604 /// block to layout at the top of the loop. Typically this is done to maximize
605 /// fallthrough opportunities.
607 MachineBlockPlacement::findBestLoopExit(MachineFunction &F,
609 const BlockFilterSet &LoopBlockSet) {
610 // We don't want to layout the loop linearly in all cases. If the loop header
611 // is just a normal basic block in the loop, we want to look for what block
612 // within the loop is the best one to layout at the top. However, if the loop
613 // header has be pre-merged into a chain due to predecessors not having
614 // analyzable branches, *and* the predecessor it is merged with is *not* part
615 // of the loop, rotating the header into the middle of the loop will create
616 // a non-contiguous range of blocks which is Very Bad. So start with the
617 // header and only rotate if safe.
618 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
619 if (!LoopBlockSet.count(*HeaderChain.begin()))
622 BlockFrequency BestExitEdgeFreq;
623 unsigned BestExitLoopDepth = 0;
624 MachineBasicBlock *ExitingBB = 0;
625 // If there are exits to outer loops, loop rotation can severely limit
626 // fallthrough opportunites unless it selects such an exit. Keep a set of
627 // blocks where rotating to exit with that block will reach an outer loop.
628 SmallPtrSet<MachineBasicBlock *, 4> BlocksExitingToOuterLoop;
630 DEBUG(dbgs() << "Finding best loop exit for: "
631 << getBlockName(L.getHeader()) << "\n");
632 for (MachineLoop::block_iterator I = L.block_begin(),
635 BlockChain &Chain = *BlockToChain[*I];
636 // Ensure that this block is at the end of a chain; otherwise it could be
637 // mid-way through an inner loop or a successor of an analyzable branch.
638 if (*I != *llvm::prior(Chain.end()))
641 // Now walk the successors. We need to establish whether this has a viable
642 // exiting successor and whether it has a viable non-exiting successor.
643 // We store the old exiting state and restore it if a viable looping
644 // successor isn't found.
645 MachineBasicBlock *OldExitingBB = ExitingBB;
646 BlockFrequency OldBestExitEdgeFreq = BestExitEdgeFreq;
647 bool HasLoopingSucc = false;
648 // FIXME: Due to the performance of the probability and weight routines in
649 // the MBPI analysis, we use the internal weights and manually compute the
650 // probabilities to avoid quadratic behavior.
651 uint32_t WeightScale = 0;
652 uint32_t SumWeight = MBPI->getSumForBlock(*I, WeightScale);
653 for (MachineBasicBlock::succ_iterator SI = (*I)->succ_begin(),
654 SE = (*I)->succ_end();
656 if ((*SI)->isLandingPad())
660 BlockChain &SuccChain = *BlockToChain[*SI];
661 // Don't split chains, either this chain or the successor's chain.
662 if (&Chain == &SuccChain) {
663 DEBUG(dbgs() << " exiting: " << getBlockName(*I) << " -> "
664 << getBlockName(*SI) << " (chain conflict)\n");
668 uint32_t SuccWeight = MBPI->getEdgeWeight(*I, *SI);
669 if (LoopBlockSet.count(*SI)) {
670 DEBUG(dbgs() << " looping: " << getBlockName(*I) << " -> "
671 << getBlockName(*SI) << " (" << SuccWeight << ")\n");
672 HasLoopingSucc = true;
676 unsigned SuccLoopDepth = 0;
677 if (MachineLoop *ExitLoop = MLI->getLoopFor(*SI)) {
678 SuccLoopDepth = ExitLoop->getLoopDepth();
679 if (ExitLoop->contains(&L))
680 BlocksExitingToOuterLoop.insert(*I);
683 BranchProbability SuccProb(SuccWeight / WeightScale, SumWeight);
684 BlockFrequency ExitEdgeFreq = MBFI->getBlockFreq(*I) * SuccProb;
685 DEBUG(dbgs() << " exiting: " << getBlockName(*I) << " -> "
686 << getBlockName(*SI) << " [L:" << SuccLoopDepth
687 << "] (" << ExitEdgeFreq << ")\n");
688 // Note that we slightly bias this toward an existing layout successor to
689 // retain incoming order in the absence of better information.
690 // FIXME: Should we bias this more strongly? It's pretty weak.
691 if (!ExitingBB || BestExitLoopDepth < SuccLoopDepth ||
692 ExitEdgeFreq > BestExitEdgeFreq ||
693 ((*I)->isLayoutSuccessor(*SI) &&
694 !(ExitEdgeFreq < BestExitEdgeFreq))) {
695 BestExitEdgeFreq = ExitEdgeFreq;
700 // Restore the old exiting state, no viable looping successor was found.
701 if (!HasLoopingSucc) {
702 ExitingBB = OldExitingBB;
703 BestExitEdgeFreq = OldBestExitEdgeFreq;
707 // Without a candidate exiting block or with only a single block in the
708 // loop, just use the loop header to layout the loop.
709 if (!ExitingBB || L.getNumBlocks() == 1)
712 // Also, if we have exit blocks which lead to outer loops but didn't select
713 // one of them as the exiting block we are rotating toward, disable loop
714 // rotation altogether.
715 if (!BlocksExitingToOuterLoop.empty() &&
716 !BlocksExitingToOuterLoop.count(ExitingBB))
719 DEBUG(dbgs() << " Best exiting block: " << getBlockName(ExitingBB) << "\n");
723 /// \brief Attempt to rotate an exiting block to the bottom of the loop.
725 /// Once we have built a chain, try to rotate it to line up the hot exit block
726 /// with fallthrough out of the loop if doing so doesn't introduce unnecessary
727 /// branches. For example, if the loop has fallthrough into its header and out
728 /// of its bottom already, don't rotate it.
729 void MachineBlockPlacement::rotateLoop(BlockChain &LoopChain,
730 MachineBasicBlock *ExitingBB,
731 const BlockFilterSet &LoopBlockSet) {
735 MachineBasicBlock *Top = *LoopChain.begin();
736 bool ViableTopFallthrough = false;
737 for (MachineBasicBlock::pred_iterator PI = Top->pred_begin(),
738 PE = Top->pred_end();
740 BlockChain *PredChain = BlockToChain[*PI];
741 if (!LoopBlockSet.count(*PI) &&
742 (!PredChain || *PI == *llvm::prior(PredChain->end()))) {
743 ViableTopFallthrough = true;
748 // If the header has viable fallthrough, check whether the current loop
749 // bottom is a viable exiting block. If so, bail out as rotating will
750 // introduce an unnecessary branch.
751 if (ViableTopFallthrough) {
752 MachineBasicBlock *Bottom = *llvm::prior(LoopChain.end());
753 for (MachineBasicBlock::succ_iterator SI = Bottom->succ_begin(),
754 SE = Bottom->succ_end();
756 BlockChain *SuccChain = BlockToChain[*SI];
757 if (!LoopBlockSet.count(*SI) &&
758 (!SuccChain || *SI == *SuccChain->begin()))
763 BlockChain::iterator ExitIt = std::find(LoopChain.begin(), LoopChain.end(),
765 if (ExitIt == LoopChain.end())
768 std::rotate(LoopChain.begin(), llvm::next(ExitIt), LoopChain.end());
771 /// \brief Forms basic block chains from the natural loop structures.
773 /// These chains are designed to preserve the existing *structure* of the code
774 /// as much as possible. We can then stitch the chains together in a way which
775 /// both preserves the topological structure and minimizes taken conditional
777 void MachineBlockPlacement::buildLoopChains(MachineFunction &F,
779 // First recurse through any nested loops, building chains for those inner
781 for (MachineLoop::iterator LI = L.begin(), LE = L.end(); LI != LE; ++LI)
782 buildLoopChains(F, **LI);
784 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
785 BlockFilterSet LoopBlockSet(L.block_begin(), L.block_end());
787 // First check to see if there is an obviously preferable top block for the
788 // loop. This will default to the header, but may end up as one of the
789 // predecessors to the header if there is one which will result in strictly
790 // fewer branches in the loop body.
791 MachineBasicBlock *LoopTop = findBestLoopTop(L, LoopBlockSet);
793 // If we selected just the header for the loop top, look for a potentially
794 // profitable exit block in the event that rotating the loop can eliminate
795 // branches by placing an exit edge at the bottom.
796 MachineBasicBlock *ExitingBB = 0;
797 if (LoopTop == L.getHeader())
798 ExitingBB = findBestLoopExit(F, L, LoopBlockSet);
800 BlockChain &LoopChain = *BlockToChain[LoopTop];
802 // FIXME: This is a really lame way of walking the chains in the loop: we
803 // walk the blocks, and use a set to prevent visiting a particular chain
805 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
806 assert(LoopChain.LoopPredecessors == 0);
807 UpdatedPreds.insert(&LoopChain);
808 for (MachineLoop::block_iterator BI = L.block_begin(),
811 BlockChain &Chain = *BlockToChain[*BI];
812 if (!UpdatedPreds.insert(&Chain))
815 assert(Chain.LoopPredecessors == 0);
816 for (BlockChain::iterator BCI = Chain.begin(), BCE = Chain.end();
818 assert(BlockToChain[*BCI] == &Chain);
819 for (MachineBasicBlock::pred_iterator PI = (*BCI)->pred_begin(),
820 PE = (*BCI)->pred_end();
822 if (BlockToChain[*PI] == &Chain || !LoopBlockSet.count(*PI))
824 ++Chain.LoopPredecessors;
828 if (Chain.LoopPredecessors == 0)
829 BlockWorkList.push_back(*Chain.begin());
832 buildChain(LoopTop, LoopChain, BlockWorkList, &LoopBlockSet);
833 rotateLoop(LoopChain, ExitingBB, LoopBlockSet);
836 // Crash at the end so we get all of the debugging output first.
837 bool BadLoop = false;
838 if (LoopChain.LoopPredecessors) {
840 dbgs() << "Loop chain contains a block without its preds placed!\n"
841 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
842 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n";
844 for (BlockChain::iterator BCI = LoopChain.begin(), BCE = LoopChain.end();
846 dbgs() << " ... " << getBlockName(*BCI) << "\n";
847 if (!LoopBlockSet.erase(*BCI)) {
848 // We don't mark the loop as bad here because there are real situations
849 // where this can occur. For example, with an unanalyzable fallthrough
850 // from a loop block to a non-loop block or vice versa.
851 dbgs() << "Loop chain contains a block not contained by the loop!\n"
852 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
853 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
854 << " Bad block: " << getBlockName(*BCI) << "\n";
858 if (!LoopBlockSet.empty()) {
860 for (BlockFilterSet::iterator LBI = LoopBlockSet.begin(),
861 LBE = LoopBlockSet.end();
863 dbgs() << "Loop contains blocks never placed into a chain!\n"
864 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
865 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
866 << " Bad block: " << getBlockName(*LBI) << "\n";
868 assert(!BadLoop && "Detected problems with the placement of this loop.");
872 void MachineBlockPlacement::buildCFGChains(MachineFunction &F) {
873 // Ensure that every BB in the function has an associated chain to simplify
874 // the assumptions of the remaining algorithm.
875 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
876 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
877 MachineBasicBlock *BB = FI;
879 = new (ChainAllocator.Allocate()) BlockChain(BlockToChain, BB);
880 // Also, merge any blocks which we cannot reason about and must preserve
881 // the exact fallthrough behavior for.
884 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch.
885 if (!TII->AnalyzeBranch(*BB, TBB, FBB, Cond) || !FI->canFallThrough())
888 MachineFunction::iterator NextFI(llvm::next(FI));
889 MachineBasicBlock *NextBB = NextFI;
890 // Ensure that the layout successor is a viable block, as we know that
891 // fallthrough is a possibility.
892 assert(NextFI != FE && "Can't fallthrough past the last block.");
893 DEBUG(dbgs() << "Pre-merging due to unanalyzable fallthrough: "
894 << getBlockName(BB) << " -> " << getBlockName(NextBB)
896 Chain->merge(NextBB, 0);
902 // Build any loop-based chains.
903 for (MachineLoopInfo::iterator LI = MLI->begin(), LE = MLI->end(); LI != LE;
905 buildLoopChains(F, **LI);
907 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
909 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
910 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
911 MachineBasicBlock *BB = &*FI;
912 BlockChain &Chain = *BlockToChain[BB];
913 if (!UpdatedPreds.insert(&Chain))
916 assert(Chain.LoopPredecessors == 0);
917 for (BlockChain::iterator BCI = Chain.begin(), BCE = Chain.end();
919 assert(BlockToChain[*BCI] == &Chain);
920 for (MachineBasicBlock::pred_iterator PI = (*BCI)->pred_begin(),
921 PE = (*BCI)->pred_end();
923 if (BlockToChain[*PI] == &Chain)
925 ++Chain.LoopPredecessors;
929 if (Chain.LoopPredecessors == 0)
930 BlockWorkList.push_back(*Chain.begin());
933 BlockChain &FunctionChain = *BlockToChain[&F.front()];
934 buildChain(&F.front(), FunctionChain, BlockWorkList);
936 typedef SmallPtrSet<MachineBasicBlock *, 16> FunctionBlockSetType;
938 // Crash at the end so we get all of the debugging output first.
939 bool BadFunc = false;
940 FunctionBlockSetType FunctionBlockSet;
941 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
942 FunctionBlockSet.insert(FI);
944 for (BlockChain::iterator BCI = FunctionChain.begin(),
945 BCE = FunctionChain.end();
947 if (!FunctionBlockSet.erase(*BCI)) {
949 dbgs() << "Function chain contains a block not in the function!\n"
950 << " Bad block: " << getBlockName(*BCI) << "\n";
953 if (!FunctionBlockSet.empty()) {
955 for (FunctionBlockSetType::iterator FBI = FunctionBlockSet.begin(),
956 FBE = FunctionBlockSet.end();
958 dbgs() << "Function contains blocks never placed into a chain!\n"
959 << " Bad block: " << getBlockName(*FBI) << "\n";
961 assert(!BadFunc && "Detected problems with the block placement.");
964 // Splice the blocks into place.
965 MachineFunction::iterator InsertPos = F.begin();
966 for (BlockChain::iterator BI = FunctionChain.begin(),
967 BE = FunctionChain.end();
969 DEBUG(dbgs() << (BI == FunctionChain.begin() ? "Placing chain "
971 << getBlockName(*BI) << "\n");
972 if (InsertPos != MachineFunction::iterator(*BI))
973 F.splice(InsertPos, *BI);
977 // Update the terminator of the previous block.
978 if (BI == FunctionChain.begin())
980 MachineBasicBlock *PrevBB = llvm::prior(MachineFunction::iterator(*BI));
982 // FIXME: It would be awesome of updateTerminator would just return rather
983 // than assert when the branch cannot be analyzed in order to remove this
986 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch.
987 if (!TII->AnalyzeBranch(*PrevBB, TBB, FBB, Cond)) {
988 // If PrevBB has a two-way branch, try to re-order the branches
989 // such that we branch to the successor with higher weight first.
990 if (TBB && !Cond.empty() && FBB &&
991 MBPI->getEdgeWeight(PrevBB, FBB) > MBPI->getEdgeWeight(PrevBB, TBB) &&
992 !TII->ReverseBranchCondition(Cond)) {
993 DEBUG(dbgs() << "Reverse order of the two branches: "
994 << getBlockName(PrevBB) << "\n");
995 DEBUG(dbgs() << " Edge weight: " << MBPI->getEdgeWeight(PrevBB, FBB)
996 << " vs " << MBPI->getEdgeWeight(PrevBB, TBB) << "\n");
997 DebugLoc dl; // FIXME: this is nowhere
998 TII->RemoveBranch(*PrevBB);
999 TII->InsertBranch(*PrevBB, FBB, TBB, Cond, dl);
1001 PrevBB->updateTerminator();
1005 // Fixup the last block.
1007 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch.
1008 if (!TII->AnalyzeBranch(F.back(), TBB, FBB, Cond))
1009 F.back().updateTerminator();
1011 // Walk through the backedges of the function now that we have fully laid out
1012 // the basic blocks and align the destination of each backedge. We don't rely
1013 // exclusively on the loop info here so that we can align backedges in
1014 // unnatural CFGs and backedges that were introduced purely because of the
1015 // loop rotations done during this layout pass.
1016 if (F.getFunction()->getFnAttributes().
1017 hasAttribute(Attributes::OptimizeForSize))
1019 unsigned Align = TLI->getPrefLoopAlignment();
1021 return; // Don't care about loop alignment.
1022 if (FunctionChain.begin() == FunctionChain.end())
1023 return; // Empty chain.
1025 const BranchProbability ColdProb(1, 5); // 20%
1026 BlockFrequency EntryFreq = MBFI->getBlockFreq(F.begin());
1027 BlockFrequency WeightedEntryFreq = EntryFreq * ColdProb;
1028 for (BlockChain::iterator BI = llvm::next(FunctionChain.begin()),
1029 BE = FunctionChain.end();
1031 // Don't align non-looping basic blocks. These are unlikely to execute
1032 // enough times to matter in practice. Note that we'll still handle
1033 // unnatural CFGs inside of a natural outer loop (the common case) and
1035 MachineLoop *L = MLI->getLoopFor(*BI);
1039 // If the block is cold relative to the function entry don't waste space
1041 BlockFrequency Freq = MBFI->getBlockFreq(*BI);
1042 if (Freq < WeightedEntryFreq)
1045 // If the block is cold relative to its loop header, don't align it
1046 // regardless of what edges into the block exist.
1047 MachineBasicBlock *LoopHeader = L->getHeader();
1048 BlockFrequency LoopHeaderFreq = MBFI->getBlockFreq(LoopHeader);
1049 if (Freq < (LoopHeaderFreq * ColdProb))
1052 // Check for the existence of a non-layout predecessor which would benefit
1053 // from aligning this block.
1054 MachineBasicBlock *LayoutPred = *llvm::prior(BI);
1056 // Force alignment if all the predecessors are jumps. We already checked
1057 // that the block isn't cold above.
1058 if (!LayoutPred->isSuccessor(*BI)) {
1059 (*BI)->setAlignment(Align);
1063 // Align this block if the layout predecessor's edge into this block is
1064 // cold relative to the block. When this is true, othe predecessors make up
1065 // all of the hot entries into the block and thus alignment is likely to be
1067 BranchProbability LayoutProb = MBPI->getEdgeProbability(LayoutPred, *BI);
1068 BlockFrequency LayoutEdgeFreq = MBFI->getBlockFreq(LayoutPred) * LayoutProb;
1069 if (LayoutEdgeFreq <= (Freq * ColdProb))
1070 (*BI)->setAlignment(Align);
1074 bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &F) {
1075 // Check for single-block functions and skip them.
1076 if (llvm::next(F.begin()) == F.end())
1079 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
1080 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
1081 MLI = &getAnalysis<MachineLoopInfo>();
1082 TII = F.getTarget().getInstrInfo();
1083 TLI = F.getTarget().getTargetLowering();
1084 assert(BlockToChain.empty());
1088 BlockToChain.clear();
1089 ChainAllocator.DestroyAll();
1091 // We always return true as we have no way to track whether the final order
1092 // differs from the original order.
1097 /// \brief A pass to compute block placement statistics.
1099 /// A separate pass to compute interesting statistics for evaluating block
1100 /// placement. This is separate from the actual placement pass so that they can
1101 /// be computed in the absence of any placement transformations or when using
1102 /// alternative placement strategies.
1103 class MachineBlockPlacementStats : public MachineFunctionPass {
1104 /// \brief A handle to the branch probability pass.
1105 const MachineBranchProbabilityInfo *MBPI;
1107 /// \brief A handle to the function-wide block frequency pass.
1108 const MachineBlockFrequencyInfo *MBFI;
1111 static char ID; // Pass identification, replacement for typeid
1112 MachineBlockPlacementStats() : MachineFunctionPass(ID) {
1113 initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry());
1116 bool runOnMachineFunction(MachineFunction &F);
1118 void getAnalysisUsage(AnalysisUsage &AU) const {
1119 AU.addRequired<MachineBranchProbabilityInfo>();
1120 AU.addRequired<MachineBlockFrequencyInfo>();
1121 AU.setPreservesAll();
1122 MachineFunctionPass::getAnalysisUsage(AU);
1127 char MachineBlockPlacementStats::ID = 0;
1128 char &llvm::MachineBlockPlacementStatsID = MachineBlockPlacementStats::ID;
1129 INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats",
1130 "Basic Block Placement Stats", false, false)
1131 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
1132 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
1133 INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats",
1134 "Basic Block Placement Stats", false, false)
1136 bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) {
1137 // Check for single-block functions and skip them.
1138 if (llvm::next(F.begin()) == F.end())
1141 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
1142 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
1144 for (MachineFunction::iterator I = F.begin(), E = F.end(); I != E; ++I) {
1145 BlockFrequency BlockFreq = MBFI->getBlockFreq(I);
1146 Statistic &NumBranches = (I->succ_size() > 1) ? NumCondBranches
1147 : NumUncondBranches;
1148 Statistic &BranchTakenFreq = (I->succ_size() > 1) ? CondBranchTakenFreq
1149 : UncondBranchTakenFreq;
1150 for (MachineBasicBlock::succ_iterator SI = I->succ_begin(),
1153 // Skip if this successor is a fallthrough.
1154 if (I->isLayoutSuccessor(*SI))
1157 BlockFrequency EdgeFreq = BlockFreq * MBPI->getEdgeProbability(I, *SI);
1159 BranchTakenFreq += EdgeFreq.getFrequency();