1 //===-- MachineBlockPlacement.cpp - Basic Block Code Layout optimization --===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements basic block placement transformations using the CFG
11 // structure and branch probability estimates.
13 // The pass strives to preserve the structure of the CFG (that is, retain
14 // a topological ordering of basic blocks) in the absense of a *strong* signal
15 // to the contrary from probabilities. However, within the CFG structure, it
16 // attempts to choose an ordering which favors placing more likely sequences of
17 // blocks adjacent to each other.
19 // The algorithm works from the inner-most loop within a function outward, and
20 // at each stage walks through the basic blocks, trying to coalesce them into
21 // sequential chains where allowed by the CFG (or demanded by heavy
22 // probabilities). Finally, it walks the blocks in topological order, and the
23 // first time it reaches a chain of basic blocks, it schedules them in the
26 //===----------------------------------------------------------------------===//
28 #define DEBUG_TYPE "block-placement2"
29 #include "llvm/CodeGen/MachineBasicBlock.h"
30 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
31 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineFunctionPass.h"
34 #include "llvm/CodeGen/MachineLoopInfo.h"
35 #include "llvm/CodeGen/MachineModuleInfo.h"
36 #include "llvm/CodeGen/Passes.h"
37 #include "llvm/Support/Allocator.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/ADT/DenseMap.h"
40 #include "llvm/ADT/SmallPtrSet.h"
41 #include "llvm/ADT/SmallVector.h"
42 #include "llvm/ADT/Statistic.h"
43 #include "llvm/Target/TargetInstrInfo.h"
44 #include "llvm/Target/TargetLowering.h"
48 STATISTIC(NumCondBranches, "Number of conditional branches");
49 STATISTIC(NumUncondBranches, "Number of uncondittional branches");
50 STATISTIC(CondBranchTakenFreq,
51 "Potential frequency of taking conditional branches");
52 STATISTIC(UncondBranchTakenFreq,
53 "Potential frequency of taking unconditional branches");
57 /// \brief Type for our function-wide basic block -> block chain mapping.
58 typedef DenseMap<MachineBasicBlock *, BlockChain *> BlockToChainMapType;
62 /// \brief A chain of blocks which will be laid out contiguously.
64 /// This is the datastructure representing a chain of consecutive blocks that
65 /// are profitable to layout together in order to maximize fallthrough
66 /// probabilities. We also can use a block chain to represent a sequence of
67 /// basic blocks which have some external (correctness) requirement for
68 /// sequential layout.
70 /// Eventually, the block chains will form a directed graph over the function.
71 /// We provide an SCC-supporting-iterator in order to quicky build and walk the
72 /// SCCs of block chains within a function.
74 /// The block chains also have support for calculating and caching probability
75 /// information related to the chain itself versus other chains. This is used
76 /// for ranking during the final layout of block chains.
78 /// \brief The sequence of blocks belonging to this chain.
80 /// This is the sequence of blocks for a particular chain. These will be laid
81 /// out in-order within the function.
82 SmallVector<MachineBasicBlock *, 4> Blocks;
84 /// \brief A handle to the function-wide basic block to block chain mapping.
86 /// This is retained in each block chain to simplify the computation of child
87 /// block chains for SCC-formation and iteration. We store the edges to child
88 /// basic blocks, and map them back to their associated chains using this
90 BlockToChainMapType &BlockToChain;
93 /// \brief Construct a new BlockChain.
95 /// This builds a new block chain representing a single basic block in the
96 /// function. It also registers itself as the chain that block participates
97 /// in with the BlockToChain mapping.
98 BlockChain(BlockToChainMapType &BlockToChain, MachineBasicBlock *BB)
99 : Blocks(1, BB), BlockToChain(BlockToChain), LoopPredecessors(0) {
100 assert(BB && "Cannot create a chain with a null basic block");
101 BlockToChain[BB] = this;
104 /// \brief Iterator over blocks within the chain.
105 typedef SmallVectorImpl<MachineBasicBlock *>::const_iterator iterator;
107 /// \brief Beginning of blocks within the chain.
108 iterator begin() const { return Blocks.begin(); }
110 /// \brief End of blocks within the chain.
111 iterator end() const { return Blocks.end(); }
113 /// \brief Merge a block chain into this one.
115 /// This routine merges a block chain into this one. It takes care of forming
116 /// a contiguous sequence of basic blocks, updating the edge list, and
117 /// updating the block -> chain mapping. It does not free or tear down the
118 /// old chain, but the old chain's block list is no longer valid.
119 void merge(MachineBasicBlock *BB, BlockChain *Chain) {
121 assert(!Blocks.empty());
123 // Fast path in case we don't have a chain already.
125 assert(!BlockToChain[BB]);
126 Blocks.push_back(BB);
127 BlockToChain[BB] = this;
131 assert(BB == *Chain->begin());
132 assert(Chain->begin() != Chain->end());
134 // Update the incoming blocks to point to this chain, and add them to the
136 for (BlockChain::iterator BI = Chain->begin(), BE = Chain->end();
138 Blocks.push_back(*BI);
139 assert(BlockToChain[*BI] == Chain && "Incoming blocks not in chain");
140 BlockToChain[*BI] = this;
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 acording 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(MachineFunction &F,
216 const BlockFilterSet &LoopBlockSet);
217 void buildLoopChains(MachineFunction &F, MachineLoop &L);
218 void buildCFGChains(MachineFunction &F);
219 void AlignLoops(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 effeciently 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);
503 MachineBasicBlock *BestSucc = 0;
505 // Look for the best viable successor if there is one to place immediately
507 BestSucc = selectBestSuccessor(BB, Chain, BlockFilter);
509 // If an immediate successor isn't available, look for the best viable
510 // block among those we've identified as not violating the loop's CFG at
511 // this point. This won't be a fallthrough, but it will increase locality.
513 BestSucc = selectBestCandidateBlock(Chain, BlockWorkList, BlockFilter);
516 BestSucc = getFirstUnplacedBlock(F, Chain, PrevUnplacedBlockIt,
521 DEBUG(dbgs() << "Unnatural loop CFG detected, forcibly merging the "
522 "layout successor until the CFG reduces\n");
525 // Place this block, updating the datastructures to reflect its placement.
526 BlockChain &SuccChain = *BlockToChain[BestSucc];
527 // Zero out LoopPredecessors for the successor we're about to merge in case
528 // we selected a successor that didn't fit naturally into the CFG.
529 SuccChain.LoopPredecessors = 0;
530 DEBUG(dbgs() << "Merging from " << getBlockNum(BB)
531 << " to " << getBlockNum(BestSucc) << "\n");
532 markChainSuccessors(SuccChain, LoopHeaderBB, BlockWorkList, BlockFilter);
533 Chain.merge(BestSucc, &SuccChain);
534 BB = *llvm::prior(Chain.end());
537 DEBUG(dbgs() << "Finished forming chain for header block "
538 << getBlockNum(*Chain.begin()) << "\n");
541 /// \brief Find the best loop top block for layout.
543 /// This routine implements the logic to analyze the loop looking for the best
544 /// block to layout at the top of the loop. Typically this is done to maximize
545 /// fallthrough opportunities.
547 MachineBlockPlacement::findBestLoopTop(MachineFunction &F,
549 const BlockFilterSet &LoopBlockSet) {
550 // We don't want to layout the loop linearly in all cases. If the loop header
551 // is just a normal basic block in the loop, we want to look for what block
552 // within the loop is the best one to layout at the top. However, if the loop
553 // header has be pre-merged into a chain due to predecessors not having
554 // analyzable branches, *and* the predecessor it is merged with is *not* part
555 // of the loop, rotating the header into the middle of the loop will create
556 // a non-contiguous range of blocks which is Very Bad. So start with the
557 // header and only rotate if safe.
558 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
559 if (!LoopBlockSet.count(*HeaderChain.begin()))
560 return L.getHeader();
562 BlockFrequency BestExitEdgeFreq;
563 MachineBasicBlock *ExitingBB = 0;
564 MachineBasicBlock *LoopingBB = 0;
565 // If there are exits to outer loops, loop rotation can severely limit
566 // fallthrough opportunites unless it selects such an exit. Keep a set of
567 // blocks where rotating to exit with that block will reach an outer loop.
568 SmallPtrSet<MachineBasicBlock *, 4> BlocksExitingToOuterLoop;
570 DEBUG(dbgs() << "Finding best loop exit for: "
571 << getBlockName(L.getHeader()) << "\n");
572 for (MachineLoop::block_iterator I = L.block_begin(),
575 BlockChain &Chain = *BlockToChain[*I];
576 // Ensure that this block is at the end of a chain; otherwise it could be
577 // mid-way through an inner loop or a successor of an analyzable branch.
578 if (*I != *llvm::prior(Chain.end()))
581 // Now walk the successors. We need to establish whether this has a viable
582 // exiting successor and whether it has a viable non-exiting successor.
583 // We store the old exiting state and restore it if a viable looping
584 // successor isn't found.
585 MachineBasicBlock *OldExitingBB = ExitingBB;
586 BlockFrequency OldBestExitEdgeFreq = BestExitEdgeFreq;
587 // We also compute and store the best looping successor for use in layout.
588 MachineBasicBlock *BestLoopSucc = 0;
589 // FIXME: Due to the performance of the probability and weight routines in
590 // the MBPI analysis, we use the internal weights. This is only valid
591 // because it is purely a ranking function, we don't care about anything
592 // but the relative values.
593 uint32_t BestLoopSuccWeight = 0;
594 // FIXME: We also manually compute the probabilities to avoid quadratic
596 uint32_t WeightScale = 0;
597 uint32_t SumWeight = MBPI->getSumForBlock(*I, WeightScale);
598 for (MachineBasicBlock::succ_iterator SI = (*I)->succ_begin(),
599 SE = (*I)->succ_end();
601 if ((*SI)->isLandingPad())
605 BlockChain &SuccChain = *BlockToChain[*SI];
606 // Don't split chains, either this chain or the successor's chain.
607 if (&Chain == &SuccChain || *SI != *SuccChain.begin()) {
608 DEBUG(dbgs() << " " << (LoopBlockSet.count(*SI) ? "looping: "
610 << getBlockName(*I) << " -> "
611 << getBlockName(*SI) << " (chain conflict)\n");
615 uint32_t SuccWeight = MBPI->getEdgeWeight(*I, *SI);
616 if (LoopBlockSet.count(*SI)) {
617 DEBUG(dbgs() << " looping: " << getBlockName(*I) << " -> "
618 << getBlockName(*SI) << " (" << SuccWeight << ")\n");
619 if (BestLoopSucc && BestLoopSuccWeight >= SuccWeight)
623 BestLoopSuccWeight = SuccWeight;
627 BranchProbability SuccProb(SuccWeight / WeightScale, SumWeight);
628 BlockFrequency ExitEdgeFreq = MBFI->getBlockFreq(*I) * SuccProb;
629 DEBUG(dbgs() << " exiting: " << getBlockName(*I) << " -> "
630 << getBlockName(*SI) << " (" << ExitEdgeFreq << ")\n");
631 // Note that we slightly bias this toward an existing layout successor to
632 // retain incoming order in the absence of better information.
633 // FIXME: Should we bias this more strongly? It's pretty weak.
634 if (!ExitingBB || ExitEdgeFreq > BestExitEdgeFreq ||
635 ((*I)->isLayoutSuccessor(*SI) &&
636 !(ExitEdgeFreq < BestExitEdgeFreq))) {
637 BestExitEdgeFreq = ExitEdgeFreq;
641 if (MachineLoop *ExitLoop = MLI->getLoopFor(*SI))
642 if (ExitLoop->contains(&L))
643 BlocksExitingToOuterLoop.insert(*I);
646 // Restore the old exiting state, no viable looping successor was found.
648 ExitingBB = OldExitingBB;
649 BestExitEdgeFreq = OldBestExitEdgeFreq;
653 // If this was best exiting block thus far, also record the looping block.
655 LoopingBB = BestLoopSucc;
657 // Without a candidate exitting block or with only a single block in the
658 // loop, just use the loop header to layout the loop.
659 if (!ExitingBB || L.getNumBlocks() == 1)
660 return L.getHeader();
662 // Also, if we have exit blocks which lead to outer loops but didn't select
663 // one of them as the exiting block we are rotating toward, disable loop
664 // rotation altogether.
665 if (!BlocksExitingToOuterLoop.empty() &&
666 !BlocksExitingToOuterLoop.count(ExitingBB))
667 return L.getHeader();
669 assert(LoopingBB && "All successors of a loop block are exit blocks!");
670 DEBUG(dbgs() << " Best exiting block: " << getBlockName(ExitingBB) << "\n");
671 DEBUG(dbgs() << " Best top block: " << getBlockName(LoopingBB) << "\n");
675 /// \brief Forms basic block chains from the natural loop structures.
677 /// These chains are designed to preserve the existing *structure* of the code
678 /// as much as possible. We can then stitch the chains together in a way which
679 /// both preserves the topological structure and minimizes taken conditional
681 void MachineBlockPlacement::buildLoopChains(MachineFunction &F,
683 // First recurse through any nested loops, building chains for those inner
685 for (MachineLoop::iterator LI = L.begin(), LE = L.end(); LI != LE; ++LI)
686 buildLoopChains(F, **LI);
688 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
689 BlockFilterSet LoopBlockSet(L.block_begin(), L.block_end());
691 MachineBasicBlock *LayoutTop = findBestLoopTop(F, L, LoopBlockSet);
692 BlockChain &LoopChain = *BlockToChain[LayoutTop];
694 // FIXME: This is a really lame way of walking the chains in the loop: we
695 // walk the blocks, and use a set to prevent visiting a particular chain
697 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
698 assert(LoopChain.LoopPredecessors == 0);
699 UpdatedPreds.insert(&LoopChain);
700 for (MachineLoop::block_iterator BI = L.block_begin(),
703 BlockChain &Chain = *BlockToChain[*BI];
704 if (!UpdatedPreds.insert(&Chain))
707 assert(Chain.LoopPredecessors == 0);
708 for (BlockChain::iterator BCI = Chain.begin(), BCE = Chain.end();
710 assert(BlockToChain[*BCI] == &Chain);
711 for (MachineBasicBlock::pred_iterator PI = (*BCI)->pred_begin(),
712 PE = (*BCI)->pred_end();
714 if (BlockToChain[*PI] == &Chain || !LoopBlockSet.count(*PI))
716 ++Chain.LoopPredecessors;
720 if (Chain.LoopPredecessors == 0)
721 BlockWorkList.push_back(*Chain.begin());
724 buildChain(LayoutTop, LoopChain, BlockWorkList, &LoopBlockSet);
727 // Crash at the end so we get all of the debugging output first.
728 bool BadLoop = false;
729 if (LoopChain.LoopPredecessors) {
731 dbgs() << "Loop chain contains a block without its preds placed!\n"
732 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
733 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n";
735 for (BlockChain::iterator BCI = LoopChain.begin(), BCE = LoopChain.end();
737 if (!LoopBlockSet.erase(*BCI)) {
738 // We don't mark the loop as bad here because there are real situations
739 // where this can occur. For example, with an unanalyzable fallthrough
740 // from a loop block to a non-loop block or vice versa.
741 dbgs() << "Loop chain contains a block not contained by the loop!\n"
742 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
743 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
744 << " Bad block: " << getBlockName(*BCI) << "\n";
747 if (!LoopBlockSet.empty()) {
749 for (BlockFilterSet::iterator LBI = LoopBlockSet.begin(),
750 LBE = LoopBlockSet.end();
752 dbgs() << "Loop contains blocks never placed into a chain!\n"
753 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
754 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
755 << " Bad block: " << getBlockName(*LBI) << "\n";
757 assert(!BadLoop && "Detected problems with the placement of this loop.");
761 void MachineBlockPlacement::buildCFGChains(MachineFunction &F) {
762 // Ensure that every BB in the function has an associated chain to simplify
763 // the assumptions of the remaining algorithm.
764 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
765 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
766 MachineBasicBlock *BB = FI;
768 = new (ChainAllocator.Allocate()) BlockChain(BlockToChain, BB);
769 // Also, merge any blocks which we cannot reason about and must preserve
770 // the exact fallthrough behavior for.
773 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch.
774 if (!TII->AnalyzeBranch(*BB, TBB, FBB, Cond) || !FI->canFallThrough())
777 MachineFunction::iterator NextFI(llvm::next(FI));
778 MachineBasicBlock *NextBB = NextFI;
779 // Ensure that the layout successor is a viable block, as we know that
780 // fallthrough is a possibility.
781 assert(NextFI != FE && "Can't fallthrough past the last block.");
782 DEBUG(dbgs() << "Pre-merging due to unanalyzable fallthrough: "
783 << getBlockName(BB) << " -> " << getBlockName(NextBB)
785 Chain->merge(NextBB, 0);
791 // Build any loop-based chains.
792 for (MachineLoopInfo::iterator LI = MLI->begin(), LE = MLI->end(); LI != LE;
794 buildLoopChains(F, **LI);
796 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
798 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
799 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
800 MachineBasicBlock *BB = &*FI;
801 BlockChain &Chain = *BlockToChain[BB];
802 if (!UpdatedPreds.insert(&Chain))
805 assert(Chain.LoopPredecessors == 0);
806 for (BlockChain::iterator BCI = Chain.begin(), BCE = Chain.end();
808 assert(BlockToChain[*BCI] == &Chain);
809 for (MachineBasicBlock::pred_iterator PI = (*BCI)->pred_begin(),
810 PE = (*BCI)->pred_end();
812 if (BlockToChain[*PI] == &Chain)
814 ++Chain.LoopPredecessors;
818 if (Chain.LoopPredecessors == 0)
819 BlockWorkList.push_back(*Chain.begin());
822 BlockChain &FunctionChain = *BlockToChain[&F.front()];
823 buildChain(&F.front(), FunctionChain, BlockWorkList);
825 typedef SmallPtrSet<MachineBasicBlock *, 16> FunctionBlockSetType;
827 // Crash at the end so we get all of the debugging output first.
828 bool BadFunc = false;
829 FunctionBlockSetType FunctionBlockSet;
830 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
831 FunctionBlockSet.insert(FI);
833 for (BlockChain::iterator BCI = FunctionChain.begin(),
834 BCE = FunctionChain.end();
836 if (!FunctionBlockSet.erase(*BCI)) {
838 dbgs() << "Function chain contains a block not in the function!\n"
839 << " Bad block: " << getBlockName(*BCI) << "\n";
842 if (!FunctionBlockSet.empty()) {
844 for (FunctionBlockSetType::iterator FBI = FunctionBlockSet.begin(),
845 FBE = FunctionBlockSet.end();
847 dbgs() << "Function contains blocks never placed into a chain!\n"
848 << " Bad block: " << getBlockName(*FBI) << "\n";
850 assert(!BadFunc && "Detected problems with the block placement.");
853 // Splice the blocks into place.
854 MachineFunction::iterator InsertPos = F.begin();
855 for (BlockChain::iterator BI = FunctionChain.begin(),
856 BE = FunctionChain.end();
858 DEBUG(dbgs() << (BI == FunctionChain.begin() ? "Placing chain "
860 << getBlockName(*BI) << "\n");
861 if (InsertPos != MachineFunction::iterator(*BI))
862 F.splice(InsertPos, *BI);
866 // Update the terminator of the previous block.
867 if (BI == FunctionChain.begin())
869 MachineBasicBlock *PrevBB = llvm::prior(MachineFunction::iterator(*BI));
871 // FIXME: It would be awesome of updateTerminator would just return rather
872 // than assert when the branch cannot be analyzed in order to remove this
875 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch.
876 if (!TII->AnalyzeBranch(*PrevBB, TBB, FBB, Cond))
877 PrevBB->updateTerminator();
880 // Fixup the last block.
882 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch.
883 if (!TII->AnalyzeBranch(F.back(), TBB, FBB, Cond))
884 F.back().updateTerminator();
887 /// \brief Recursive helper to align a loop and any nested loops.
888 static void AlignLoop(MachineFunction &F, MachineLoop *L, unsigned Align) {
889 // Recurse through nested loops.
890 for (MachineLoop::iterator I = L->begin(), E = L->end(); I != E; ++I)
891 AlignLoop(F, *I, Align);
893 L->getTopBlock()->setAlignment(Align);
896 /// \brief Align loop headers to target preferred alignments.
897 void MachineBlockPlacement::AlignLoops(MachineFunction &F) {
898 if (F.getFunction()->hasFnAttr(Attribute::OptimizeForSize))
901 unsigned Align = TLI->getPrefLoopAlignment();
903 return; // Don't care about loop alignment.
905 for (MachineLoopInfo::iterator I = MLI->begin(), E = MLI->end(); I != E; ++I)
906 AlignLoop(F, *I, Align);
909 bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &F) {
910 // Check for single-block functions and skip them.
911 if (llvm::next(F.begin()) == F.end())
914 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
915 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
916 MLI = &getAnalysis<MachineLoopInfo>();
917 TII = F.getTarget().getInstrInfo();
918 TLI = F.getTarget().getTargetLowering();
919 assert(BlockToChain.empty());
924 BlockToChain.clear();
925 ChainAllocator.DestroyAll();
927 // We always return true as we have no way to track whether the final order
928 // differs from the original order.
933 /// \brief A pass to compute block placement statistics.
935 /// A separate pass to compute interesting statistics for evaluating block
936 /// placement. This is separate from the actual placement pass so that they can
937 /// be computed in the absense of any placement transformations or when using
938 /// alternative placement strategies.
939 class MachineBlockPlacementStats : public MachineFunctionPass {
940 /// \brief A handle to the branch probability pass.
941 const MachineBranchProbabilityInfo *MBPI;
943 /// \brief A handle to the function-wide block frequency pass.
944 const MachineBlockFrequencyInfo *MBFI;
947 static char ID; // Pass identification, replacement for typeid
948 MachineBlockPlacementStats() : MachineFunctionPass(ID) {
949 initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry());
952 bool runOnMachineFunction(MachineFunction &F);
954 void getAnalysisUsage(AnalysisUsage &AU) const {
955 AU.addRequired<MachineBranchProbabilityInfo>();
956 AU.addRequired<MachineBlockFrequencyInfo>();
957 AU.setPreservesAll();
958 MachineFunctionPass::getAnalysisUsage(AU);
963 char MachineBlockPlacementStats::ID = 0;
964 char &llvm::MachineBlockPlacementStatsID = MachineBlockPlacementStats::ID;
965 INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats",
966 "Basic Block Placement Stats", false, false)
967 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
968 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
969 INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats",
970 "Basic Block Placement Stats", false, false)
972 bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) {
973 // Check for single-block functions and skip them.
974 if (llvm::next(F.begin()) == F.end())
977 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
978 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
980 for (MachineFunction::iterator I = F.begin(), E = F.end(); I != E; ++I) {
981 BlockFrequency BlockFreq = MBFI->getBlockFreq(I);
982 Statistic &NumBranches = (I->succ_size() > 1) ? NumCondBranches
984 Statistic &BranchTakenFreq = (I->succ_size() > 1) ? CondBranchTakenFreq
985 : UncondBranchTakenFreq;
986 for (MachineBasicBlock::succ_iterator SI = I->succ_begin(),
989 // Skip if this successor is a fallthrough.
990 if (I->isLayoutSuccessor(*SI))
993 BlockFrequency EdgeFreq = BlockFreq * MBPI->getEdgeProbability(I, *SI);
995 BranchTakenFreq += EdgeFreq.getFrequency();