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/Passes.h"
30 #include "llvm/ADT/DenseMap.h"
31 #include "llvm/ADT/SmallPtrSet.h"
32 #include "llvm/ADT/SmallVector.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/CodeGen/MachineBasicBlock.h"
35 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
36 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
37 #include "llvm/CodeGen/MachineFunction.h"
38 #include "llvm/CodeGen/MachineFunctionPass.h"
39 #include "llvm/CodeGen/MachineLoopInfo.h"
40 #include "llvm/CodeGen/MachineModuleInfo.h"
41 #include "llvm/Support/Allocator.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/Target/TargetInstrInfo.h"
45 #include "llvm/Target/TargetLowering.h"
49 STATISTIC(NumCondBranches, "Number of conditional branches");
50 STATISTIC(NumUncondBranches, "Number of uncondittional branches");
51 STATISTIC(CondBranchTakenFreq,
52 "Potential frequency of taking conditional branches");
53 STATISTIC(UncondBranchTakenFreq,
54 "Potential frequency of taking unconditional branches");
56 static cl::opt<unsigned> AlignAllBlock("align-all-blocks",
57 cl::desc("Force the alignment of all "
58 "blocks in the function."),
59 cl::init(0), cl::Hidden);
61 // FIXME: Find a good default for this flag and remove the flag.
62 static cl::opt<unsigned>
63 ExitBlockBias("block-placement-exit-block-bias",
64 cl::desc("Block frequency percentage a loop exit block needs "
65 "over the original exit to be considered the new exit."),
66 cl::init(0), cl::Hidden);
70 /// \brief Type for our function-wide basic block -> block chain mapping.
71 typedef DenseMap<MachineBasicBlock *, BlockChain *> BlockToChainMapType;
75 /// \brief A chain of blocks which will be laid out contiguously.
77 /// This is the datastructure representing a chain of consecutive blocks that
78 /// are profitable to layout together in order to maximize fallthrough
79 /// probabilities and code locality. We also can use a block chain to represent
80 /// a sequence of basic blocks which have some external (correctness)
81 /// requirement for sequential layout.
83 /// Chains can be built around a single basic block and can be merged to grow
84 /// them. They participate in a block-to-chain mapping, which is updated
85 /// automatically as chains are merged together.
87 /// \brief The sequence of blocks belonging to this chain.
89 /// This is the sequence of blocks for a particular chain. These will be laid
90 /// out in-order within the function.
91 SmallVector<MachineBasicBlock *, 4> Blocks;
93 /// \brief A handle to the function-wide basic block to block chain mapping.
95 /// This is retained in each block chain to simplify the computation of child
96 /// block chains for SCC-formation and iteration. We store the edges to child
97 /// basic blocks, and map them back to their associated chains using this
99 BlockToChainMapType &BlockToChain;
102 /// \brief Construct a new BlockChain.
104 /// This builds a new block chain representing a single basic block in the
105 /// function. It also registers itself as the chain that block participates
106 /// in with the BlockToChain mapping.
107 BlockChain(BlockToChainMapType &BlockToChain, MachineBasicBlock *BB)
108 : Blocks(1, BB), BlockToChain(BlockToChain), LoopPredecessors(0) {
109 assert(BB && "Cannot create a chain with a null basic block");
110 BlockToChain[BB] = this;
113 /// \brief Iterator over blocks within the chain.
114 typedef SmallVectorImpl<MachineBasicBlock *>::iterator iterator;
116 /// \brief Beginning of blocks within the chain.
117 iterator begin() { return Blocks.begin(); }
119 /// \brief End of blocks within the chain.
120 iterator end() { return Blocks.end(); }
122 /// \brief Merge a block chain into this one.
124 /// This routine merges a block chain into this one. It takes care of forming
125 /// a contiguous sequence of basic blocks, updating the edge list, and
126 /// updating the block -> chain mapping. It does not free or tear down the
127 /// old chain, but the old chain's block list is no longer valid.
128 void merge(MachineBasicBlock *BB, BlockChain *Chain) {
130 assert(!Blocks.empty());
132 // Fast path in case we don't have a chain already.
134 assert(!BlockToChain[BB]);
135 Blocks.push_back(BB);
136 BlockToChain[BB] = this;
140 assert(BB == *Chain->begin());
141 assert(Chain->begin() != Chain->end());
143 // Update the incoming blocks to point to this chain, and add them to the
145 for (BlockChain::iterator BI = Chain->begin(), BE = Chain->end();
147 Blocks.push_back(*BI);
148 assert(BlockToChain[*BI] == Chain && "Incoming blocks not in chain");
149 BlockToChain[*BI] = this;
154 /// \brief Dump the blocks in this chain.
155 LLVM_DUMP_METHOD void dump() {
156 for (iterator I = begin(), E = end(); I != E; ++I)
161 /// \brief Count of predecessors within the loop currently being processed.
163 /// This count is updated at each loop we process to represent the number of
164 /// in-loop predecessors of this chain.
165 unsigned LoopPredecessors;
170 class MachineBlockPlacement : public MachineFunctionPass {
171 /// \brief A typedef for a block filter set.
172 typedef SmallPtrSet<MachineBasicBlock *, 16> BlockFilterSet;
174 /// \brief A handle to the branch probability pass.
175 const MachineBranchProbabilityInfo *MBPI;
177 /// \brief A handle to the function-wide block frequency pass.
178 const MachineBlockFrequencyInfo *MBFI;
180 /// \brief A handle to the loop info.
181 const MachineLoopInfo *MLI;
183 /// \brief A handle to the target's instruction info.
184 const TargetInstrInfo *TII;
186 /// \brief A handle to the target's lowering info.
187 const TargetLoweringBase *TLI;
189 /// \brief Allocator and owner of BlockChain structures.
191 /// We build BlockChains lazily while processing the loop structure of
192 /// a function. To reduce malloc traffic, we allocate them using this
193 /// slab-like allocator, and destroy them after the pass completes. An
194 /// important guarantee is that this allocator produces stable pointers to
196 SpecificBumpPtrAllocator<BlockChain> ChainAllocator;
198 /// \brief Function wide BasicBlock to BlockChain mapping.
200 /// This mapping allows efficiently moving from any given basic block to the
201 /// BlockChain it participates in, if any. We use it to, among other things,
202 /// allow implicitly defining edges between chains as the existing edges
203 /// between basic blocks.
204 DenseMap<MachineBasicBlock *, BlockChain *> BlockToChain;
206 void markChainSuccessors(BlockChain &Chain,
207 MachineBasicBlock *LoopHeaderBB,
208 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
209 const BlockFilterSet *BlockFilter = 0);
210 MachineBasicBlock *selectBestSuccessor(MachineBasicBlock *BB,
212 const BlockFilterSet *BlockFilter);
213 MachineBasicBlock *selectBestCandidateBlock(
214 BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList,
215 const BlockFilterSet *BlockFilter);
216 MachineBasicBlock *getFirstUnplacedBlock(
218 const BlockChain &PlacedChain,
219 MachineFunction::iterator &PrevUnplacedBlockIt,
220 const BlockFilterSet *BlockFilter);
221 void buildChain(MachineBasicBlock *BB, BlockChain &Chain,
222 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
223 const BlockFilterSet *BlockFilter = 0);
224 MachineBasicBlock *findBestLoopTop(MachineLoop &L,
225 const BlockFilterSet &LoopBlockSet);
226 MachineBasicBlock *findBestLoopExit(MachineFunction &F,
228 const BlockFilterSet &LoopBlockSet);
229 void buildLoopChains(MachineFunction &F, MachineLoop &L);
230 void rotateLoop(BlockChain &LoopChain, MachineBasicBlock *ExitingBB,
231 const BlockFilterSet &LoopBlockSet);
232 void buildCFGChains(MachineFunction &F);
235 static char ID; // Pass identification, replacement for typeid
236 MachineBlockPlacement() : MachineFunctionPass(ID) {
237 initializeMachineBlockPlacementPass(*PassRegistry::getPassRegistry());
240 bool runOnMachineFunction(MachineFunction &F);
242 void getAnalysisUsage(AnalysisUsage &AU) const {
243 AU.addRequired<MachineBranchProbabilityInfo>();
244 AU.addRequired<MachineBlockFrequencyInfo>();
245 AU.addRequired<MachineLoopInfo>();
246 MachineFunctionPass::getAnalysisUsage(AU);
251 char MachineBlockPlacement::ID = 0;
252 char &llvm::MachineBlockPlacementID = MachineBlockPlacement::ID;
253 INITIALIZE_PASS_BEGIN(MachineBlockPlacement, "block-placement2",
254 "Branch Probability Basic Block Placement", false, false)
255 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
256 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
257 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
258 INITIALIZE_PASS_END(MachineBlockPlacement, "block-placement2",
259 "Branch Probability Basic Block Placement", false, false)
262 /// \brief Helper to print the name of a MBB.
264 /// Only used by debug logging.
265 static std::string getBlockName(MachineBasicBlock *BB) {
267 raw_string_ostream OS(Result);
268 OS << "BB#" << BB->getNumber()
269 << " (derived from LLVM BB '" << BB->getName() << "')";
274 /// \brief Helper to print the number of a MBB.
276 /// Only used by debug logging.
277 static std::string getBlockNum(MachineBasicBlock *BB) {
279 raw_string_ostream OS(Result);
280 OS << "BB#" << BB->getNumber();
286 /// \brief Mark a chain's successors as having one fewer preds.
288 /// When a chain is being merged into the "placed" chain, this routine will
289 /// quickly walk the successors of each block in the chain and mark them as
290 /// having one fewer active predecessor. It also adds any successors of this
291 /// chain which reach the zero-predecessor state to the worklist passed in.
292 void MachineBlockPlacement::markChainSuccessors(
294 MachineBasicBlock *LoopHeaderBB,
295 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
296 const BlockFilterSet *BlockFilter) {
297 // Walk all the blocks in this chain, marking their successors as having
298 // a predecessor placed.
299 for (BlockChain::iterator CBI = Chain.begin(), CBE = Chain.end();
301 // Add any successors for which this is the only un-placed in-loop
302 // predecessor to the worklist as a viable candidate for CFG-neutral
303 // placement. No subsequent placement of this block will violate the CFG
304 // shape, so we get to use heuristics to choose a favorable placement.
305 for (MachineBasicBlock::succ_iterator SI = (*CBI)->succ_begin(),
306 SE = (*CBI)->succ_end();
308 if (BlockFilter && !BlockFilter->count(*SI))
310 BlockChain &SuccChain = *BlockToChain[*SI];
311 // Disregard edges within a fixed chain, or edges to the loop header.
312 if (&Chain == &SuccChain || *SI == LoopHeaderBB)
315 // This is a cross-chain edge that is within the loop, so decrement the
316 // loop predecessor count of the destination chain.
317 if (SuccChain.LoopPredecessors > 0 && --SuccChain.LoopPredecessors == 0)
318 BlockWorkList.push_back(*SuccChain.begin());
323 /// \brief Select the best successor for a block.
325 /// This looks across all successors of a particular block and attempts to
326 /// select the "best" one to be the layout successor. It only considers direct
327 /// successors which also pass the block filter. It will attempt to avoid
328 /// breaking CFG structure, but cave and break such structures in the case of
329 /// very hot successor edges.
331 /// \returns The best successor block found, or null if none are viable.
332 MachineBasicBlock *MachineBlockPlacement::selectBestSuccessor(
333 MachineBasicBlock *BB, BlockChain &Chain,
334 const BlockFilterSet *BlockFilter) {
335 const BranchProbability HotProb(4, 5); // 80%
337 MachineBasicBlock *BestSucc = 0;
338 // FIXME: Due to the performance of the probability and weight routines in
339 // the MBPI analysis, we manually compute probabilities using the edge
340 // weights. This is suboptimal as it means that the somewhat subtle
341 // definition of edge weight semantics is encoded here as well. We should
342 // improve the MBPI interface to efficiently support query patterns such as
344 uint32_t BestWeight = 0;
345 uint32_t WeightScale = 0;
346 uint32_t SumWeight = MBPI->getSumForBlock(BB, WeightScale);
347 DEBUG(dbgs() << "Attempting merge from: " << getBlockName(BB) << "\n");
348 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
351 if (BlockFilter && !BlockFilter->count(*SI))
353 BlockChain &SuccChain = *BlockToChain[*SI];
354 if (&SuccChain == &Chain) {
355 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> Already merged!\n");
358 if (*SI != *SuccChain.begin()) {
359 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> Mid chain!\n");
363 uint32_t SuccWeight = MBPI->getEdgeWeight(BB, *SI);
364 BranchProbability SuccProb(SuccWeight / WeightScale, SumWeight);
366 // Only consider successors which are either "hot", or wouldn't violate
367 // any CFG constraints.
368 if (SuccChain.LoopPredecessors != 0) {
369 if (SuccProb < HotProb) {
370 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> " << SuccProb
371 << " (prob) (CFG conflict)\n");
375 // Make sure that a hot successor doesn't have a globally more important
377 BlockFrequency CandidateEdgeFreq
378 = MBFI->getBlockFreq(BB) * SuccProb * HotProb.getCompl();
379 bool BadCFGConflict = false;
380 for (MachineBasicBlock::pred_iterator PI = (*SI)->pred_begin(),
381 PE = (*SI)->pred_end();
383 if (*PI == *SI || (BlockFilter && !BlockFilter->count(*PI)) ||
384 BlockToChain[*PI] == &Chain)
386 BlockFrequency PredEdgeFreq
387 = MBFI->getBlockFreq(*PI) * MBPI->getEdgeProbability(*PI, *SI);
388 if (PredEdgeFreq >= CandidateEdgeFreq) {
389 BadCFGConflict = true;
393 if (BadCFGConflict) {
394 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> " << SuccProb
395 << " (prob) (non-cold CFG conflict)\n");
400 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> " << SuccProb
402 << (SuccChain.LoopPredecessors != 0 ? " (CFG break)" : "")
404 if (BestSucc && BestWeight >= SuccWeight)
407 BestWeight = SuccWeight;
412 /// \brief Select the best block from a worklist.
414 /// This looks through the provided worklist as a list of candidate basic
415 /// blocks and select the most profitable one to place. The definition of
416 /// profitable only really makes sense in the context of a loop. This returns
417 /// the most frequently visited block in the worklist, which in the case of
418 /// a loop, is the one most desirable to be physically close to the rest of the
419 /// loop body in order to improve icache behavior.
421 /// \returns The best block found, or null if none are viable.
422 MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock(
423 BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList,
424 const BlockFilterSet *BlockFilter) {
425 // Once we need to walk the worklist looking for a candidate, cleanup the
426 // worklist of already placed entries.
427 // FIXME: If this shows up on profiles, it could be folded (at the cost of
428 // some code complexity) into the loop below.
429 WorkList.erase(std::remove_if(WorkList.begin(), WorkList.end(),
430 [&](MachineBasicBlock *BB) {
431 return BlockToChain.lookup(BB) == &Chain;
435 MachineBasicBlock *BestBlock = 0;
436 BlockFrequency BestFreq;
437 for (SmallVectorImpl<MachineBasicBlock *>::iterator WBI = WorkList.begin(),
438 WBE = WorkList.end();
440 BlockChain &SuccChain = *BlockToChain[*WBI];
441 if (&SuccChain == &Chain) {
442 DEBUG(dbgs() << " " << getBlockName(*WBI)
443 << " -> Already merged!\n");
446 assert(SuccChain.LoopPredecessors == 0 && "Found CFG-violating block");
448 BlockFrequency CandidateFreq = MBFI->getBlockFreq(*WBI);
449 DEBUG(dbgs() << " " << getBlockName(*WBI) << " -> ";
450 MBFI->printBlockFreq(dbgs(), CandidateFreq) << " (freq)\n");
451 if (BestBlock && BestFreq >= CandidateFreq)
454 BestFreq = CandidateFreq;
459 /// \brief Retrieve the first unplaced basic block.
461 /// This routine is called when we are unable to use the CFG to walk through
462 /// all of the basic blocks and form a chain due to unnatural loops in the CFG.
463 /// We walk through the function's blocks in order, starting from the
464 /// LastUnplacedBlockIt. We update this iterator on each call to avoid
465 /// re-scanning the entire sequence on repeated calls to this routine.
466 MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock(
467 MachineFunction &F, const BlockChain &PlacedChain,
468 MachineFunction::iterator &PrevUnplacedBlockIt,
469 const BlockFilterSet *BlockFilter) {
470 for (MachineFunction::iterator I = PrevUnplacedBlockIt, E = F.end(); I != E;
472 if (BlockFilter && !BlockFilter->count(I))
474 if (BlockToChain[I] != &PlacedChain) {
475 PrevUnplacedBlockIt = I;
476 // Now select the head of the chain to which the unplaced block belongs
477 // as the block to place. This will force the entire chain to be placed,
478 // and satisfies the requirements of merging chains.
479 return *BlockToChain[I]->begin();
485 void MachineBlockPlacement::buildChain(
486 MachineBasicBlock *BB,
488 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
489 const BlockFilterSet *BlockFilter) {
491 assert(BlockToChain[BB] == &Chain);
492 MachineFunction &F = *BB->getParent();
493 MachineFunction::iterator PrevUnplacedBlockIt = F.begin();
495 MachineBasicBlock *LoopHeaderBB = BB;
496 markChainSuccessors(Chain, LoopHeaderBB, BlockWorkList, BlockFilter);
497 BB = *std::prev(Chain.end());
500 assert(BlockToChain[BB] == &Chain);
501 assert(*std::prev(Chain.end()) == BB);
503 // Look for the best viable successor if there is one to place immediately
505 MachineBasicBlock *BestSucc = selectBestSuccessor(BB, Chain, BlockFilter);
507 // If an immediate successor isn't available, look for the best viable
508 // block among those we've identified as not violating the loop's CFG at
509 // this point. This won't be a fallthrough, but it will increase locality.
511 BestSucc = selectBestCandidateBlock(Chain, BlockWorkList, BlockFilter);
514 BestSucc = getFirstUnplacedBlock(F, Chain, PrevUnplacedBlockIt,
519 DEBUG(dbgs() << "Unnatural loop CFG detected, forcibly merging the "
520 "layout successor until the CFG reduces\n");
523 // Place this block, updating the datastructures to reflect its placement.
524 BlockChain &SuccChain = *BlockToChain[BestSucc];
525 // Zero out LoopPredecessors for the successor we're about to merge in case
526 // we selected a successor that didn't fit naturally into the CFG.
527 SuccChain.LoopPredecessors = 0;
528 DEBUG(dbgs() << "Merging from " << getBlockNum(BB)
529 << " to " << getBlockNum(BestSucc) << "\n");
530 markChainSuccessors(SuccChain, LoopHeaderBB, BlockWorkList, BlockFilter);
531 Chain.merge(BestSucc, &SuccChain);
532 BB = *std::prev(Chain.end());
535 DEBUG(dbgs() << "Finished forming chain for header block "
536 << getBlockNum(*Chain.begin()) << "\n");
539 /// \brief Find the best loop top block for layout.
541 /// Look for a block which is strictly better than the loop header for laying
542 /// out at the top of the loop. This looks for one and only one pattern:
543 /// a latch block with no conditional exit. This block will cause a conditional
544 /// jump around it or will be the bottom of the loop if we lay it out in place,
545 /// but if it it doesn't end up at the bottom of the loop for any reason,
546 /// rotation alone won't fix it. Because such a block will always result in an
547 /// unconditional jump (for the backedge) rotating it in front of the loop
548 /// header is always profitable.
550 MachineBlockPlacement::findBestLoopTop(MachineLoop &L,
551 const BlockFilterSet &LoopBlockSet) {
552 // Check that the header hasn't been fused with a preheader block due to
553 // crazy branches. If it has, we need to start with the header at the top to
554 // prevent pulling the preheader into the loop body.
555 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
556 if (!LoopBlockSet.count(*HeaderChain.begin()))
557 return L.getHeader();
559 DEBUG(dbgs() << "Finding best loop top for: "
560 << getBlockName(L.getHeader()) << "\n");
562 BlockFrequency BestPredFreq;
563 MachineBasicBlock *BestPred = 0;
564 for (MachineBasicBlock::pred_iterator PI = L.getHeader()->pred_begin(),
565 PE = L.getHeader()->pred_end();
567 MachineBasicBlock *Pred = *PI;
568 if (!LoopBlockSet.count(Pred))
570 DEBUG(dbgs() << " header pred: " << getBlockName(Pred) << ", "
571 << Pred->succ_size() << " successors, ";
572 MBFI->printBlockFreq(dbgs(), Pred) << " freq\n");
573 if (Pred->succ_size() > 1)
576 BlockFrequency PredFreq = MBFI->getBlockFreq(Pred);
577 if (!BestPred || PredFreq > BestPredFreq ||
578 (!(PredFreq < BestPredFreq) &&
579 Pred->isLayoutSuccessor(L.getHeader()))) {
581 BestPredFreq = PredFreq;
585 // If no direct predecessor is fine, just use the loop header.
587 return L.getHeader();
589 // Walk backwards through any straight line of predecessors.
590 while (BestPred->pred_size() == 1 &&
591 (*BestPred->pred_begin())->succ_size() == 1 &&
592 *BestPred->pred_begin() != L.getHeader())
593 BestPred = *BestPred->pred_begin();
595 DEBUG(dbgs() << " final top: " << getBlockName(BestPred) << "\n");
600 /// \brief Find the best loop exiting block for layout.
602 /// This routine implements the logic to analyze the loop looking for the best
603 /// block to layout at the top of the loop. Typically this is done to maximize
604 /// fallthrough opportunities.
606 MachineBlockPlacement::findBestLoopExit(MachineFunction &F,
608 const BlockFilterSet &LoopBlockSet) {
609 // We don't want to layout the loop linearly in all cases. If the loop header
610 // is just a normal basic block in the loop, we want to look for what block
611 // within the loop is the best one to layout at the top. However, if the loop
612 // header has be pre-merged into a chain due to predecessors not having
613 // analyzable branches, *and* the predecessor it is merged with is *not* part
614 // of the loop, rotating the header into the middle of the loop will create
615 // a non-contiguous range of blocks which is Very Bad. So start with the
616 // header and only rotate if safe.
617 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
618 if (!LoopBlockSet.count(*HeaderChain.begin()))
621 BlockFrequency BestExitEdgeFreq;
622 unsigned BestExitLoopDepth = 0;
623 MachineBasicBlock *ExitingBB = 0;
624 // If there are exits to outer loops, loop rotation can severely limit
625 // fallthrough opportunites unless it selects such an exit. Keep a set of
626 // blocks where rotating to exit with that block will reach an outer loop.
627 SmallPtrSet<MachineBasicBlock *, 4> BlocksExitingToOuterLoop;
629 DEBUG(dbgs() << "Finding best loop exit for: "
630 << getBlockName(L.getHeader()) << "\n");
631 for (MachineLoop::block_iterator I = L.block_begin(),
634 BlockChain &Chain = *BlockToChain[*I];
635 // Ensure that this block is at the end of a chain; otherwise it could be
636 // mid-way through an inner loop or a successor of an analyzable branch.
637 if (*I != *std::prev(Chain.end()))
640 // Now walk the successors. We need to establish whether this has a viable
641 // exiting successor and whether it has a viable non-exiting successor.
642 // We store the old exiting state and restore it if a viable looping
643 // successor isn't found.
644 MachineBasicBlock *OldExitingBB = ExitingBB;
645 BlockFrequency OldBestExitEdgeFreq = BestExitEdgeFreq;
646 bool HasLoopingSucc = false;
647 // FIXME: Due to the performance of the probability and weight routines in
648 // the MBPI analysis, we use the internal weights and manually compute the
649 // probabilities to avoid quadratic behavior.
650 uint32_t WeightScale = 0;
651 uint32_t SumWeight = MBPI->getSumForBlock(*I, WeightScale);
652 for (MachineBasicBlock::succ_iterator SI = (*I)->succ_begin(),
653 SE = (*I)->succ_end();
655 if ((*SI)->isLandingPad())
659 BlockChain &SuccChain = *BlockToChain[*SI];
660 // Don't split chains, either this chain or the successor's chain.
661 if (&Chain == &SuccChain) {
662 DEBUG(dbgs() << " exiting: " << getBlockName(*I) << " -> "
663 << getBlockName(*SI) << " (chain conflict)\n");
667 uint32_t SuccWeight = MBPI->getEdgeWeight(*I, *SI);
668 if (LoopBlockSet.count(*SI)) {
669 DEBUG(dbgs() << " looping: " << getBlockName(*I) << " -> "
670 << getBlockName(*SI) << " (" << SuccWeight << ")\n");
671 HasLoopingSucc = true;
675 unsigned SuccLoopDepth = 0;
676 if (MachineLoop *ExitLoop = MLI->getLoopFor(*SI)) {
677 SuccLoopDepth = ExitLoop->getLoopDepth();
678 if (ExitLoop->contains(&L))
679 BlocksExitingToOuterLoop.insert(*I);
682 BranchProbability SuccProb(SuccWeight / WeightScale, SumWeight);
683 BlockFrequency ExitEdgeFreq = MBFI->getBlockFreq(*I) * SuccProb;
684 DEBUG(dbgs() << " exiting: " << getBlockName(*I) << " -> "
685 << getBlockName(*SI) << " [L:" << SuccLoopDepth
687 MBFI->printBlockFreq(dbgs(), ExitEdgeFreq) << ")\n");
688 // Note that we bias this toward an existing layout successor to retain
689 // incoming order in the absence of better information. The exit must have
690 // a frequency higher than the current exit before we consider breaking
692 BranchProbability Bias(100 - ExitBlockBias, 100);
693 if (!ExitingBB || BestExitLoopDepth < SuccLoopDepth ||
694 ExitEdgeFreq > BestExitEdgeFreq ||
695 ((*I)->isLayoutSuccessor(*SI) &&
696 !(ExitEdgeFreq < BestExitEdgeFreq * Bias))) {
697 BestExitEdgeFreq = ExitEdgeFreq;
702 // Restore the old exiting state, no viable looping successor was found.
703 if (!HasLoopingSucc) {
704 ExitingBB = OldExitingBB;
705 BestExitEdgeFreq = OldBestExitEdgeFreq;
709 // Without a candidate exiting block or with only a single block in the
710 // loop, just use the loop header to layout the loop.
711 if (!ExitingBB || L.getNumBlocks() == 1)
714 // Also, if we have exit blocks which lead to outer loops but didn't select
715 // one of them as the exiting block we are rotating toward, disable loop
716 // rotation altogether.
717 if (!BlocksExitingToOuterLoop.empty() &&
718 !BlocksExitingToOuterLoop.count(ExitingBB))
721 DEBUG(dbgs() << " Best exiting block: " << getBlockName(ExitingBB) << "\n");
725 /// \brief Attempt to rotate an exiting block to the bottom of the loop.
727 /// Once we have built a chain, try to rotate it to line up the hot exit block
728 /// with fallthrough out of the loop if doing so doesn't introduce unnecessary
729 /// branches. For example, if the loop has fallthrough into its header and out
730 /// of its bottom already, don't rotate it.
731 void MachineBlockPlacement::rotateLoop(BlockChain &LoopChain,
732 MachineBasicBlock *ExitingBB,
733 const BlockFilterSet &LoopBlockSet) {
737 MachineBasicBlock *Top = *LoopChain.begin();
738 bool ViableTopFallthrough = false;
739 for (MachineBasicBlock::pred_iterator PI = Top->pred_begin(),
740 PE = Top->pred_end();
742 BlockChain *PredChain = BlockToChain[*PI];
743 if (!LoopBlockSet.count(*PI) &&
744 (!PredChain || *PI == *std::prev(PredChain->end()))) {
745 ViableTopFallthrough = true;
750 // If the header has viable fallthrough, check whether the current loop
751 // bottom is a viable exiting block. If so, bail out as rotating will
752 // introduce an unnecessary branch.
753 if (ViableTopFallthrough) {
754 MachineBasicBlock *Bottom = *std::prev(LoopChain.end());
755 for (MachineBasicBlock::succ_iterator SI = Bottom->succ_begin(),
756 SE = Bottom->succ_end();
758 BlockChain *SuccChain = BlockToChain[*SI];
759 if (!LoopBlockSet.count(*SI) &&
760 (!SuccChain || *SI == *SuccChain->begin()))
765 BlockChain::iterator ExitIt = std::find(LoopChain.begin(), LoopChain.end(),
767 if (ExitIt == LoopChain.end())
770 std::rotate(LoopChain.begin(), std::next(ExitIt), LoopChain.end());
773 /// \brief Forms basic block chains from the natural loop structures.
775 /// These chains are designed to preserve the existing *structure* of the code
776 /// as much as possible. We can then stitch the chains together in a way which
777 /// both preserves the topological structure and minimizes taken conditional
779 void MachineBlockPlacement::buildLoopChains(MachineFunction &F,
781 // First recurse through any nested loops, building chains for those inner
783 for (MachineLoop::iterator LI = L.begin(), LE = L.end(); LI != LE; ++LI)
784 buildLoopChains(F, **LI);
786 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
787 BlockFilterSet LoopBlockSet(L.block_begin(), L.block_end());
789 // First check to see if there is an obviously preferable top block for the
790 // loop. This will default to the header, but may end up as one of the
791 // predecessors to the header if there is one which will result in strictly
792 // fewer branches in the loop body.
793 MachineBasicBlock *LoopTop = findBestLoopTop(L, LoopBlockSet);
795 // If we selected just the header for the loop top, look for a potentially
796 // profitable exit block in the event that rotating the loop can eliminate
797 // branches by placing an exit edge at the bottom.
798 MachineBasicBlock *ExitingBB = 0;
799 if (LoopTop == L.getHeader())
800 ExitingBB = findBestLoopExit(F, L, LoopBlockSet);
802 BlockChain &LoopChain = *BlockToChain[LoopTop];
804 // FIXME: This is a really lame way of walking the chains in the loop: we
805 // walk the blocks, and use a set to prevent visiting a particular chain
807 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
808 assert(LoopChain.LoopPredecessors == 0);
809 UpdatedPreds.insert(&LoopChain);
810 for (MachineLoop::block_iterator BI = L.block_begin(),
813 BlockChain &Chain = *BlockToChain[*BI];
814 if (!UpdatedPreds.insert(&Chain))
817 assert(Chain.LoopPredecessors == 0);
818 for (BlockChain::iterator BCI = Chain.begin(), BCE = Chain.end();
820 assert(BlockToChain[*BCI] == &Chain);
821 for (MachineBasicBlock::pred_iterator PI = (*BCI)->pred_begin(),
822 PE = (*BCI)->pred_end();
824 if (BlockToChain[*PI] == &Chain || !LoopBlockSet.count(*PI))
826 ++Chain.LoopPredecessors;
830 if (Chain.LoopPredecessors == 0)
831 BlockWorkList.push_back(*Chain.begin());
834 buildChain(LoopTop, LoopChain, BlockWorkList, &LoopBlockSet);
835 rotateLoop(LoopChain, ExitingBB, LoopBlockSet);
838 // Crash at the end so we get all of the debugging output first.
839 bool BadLoop = false;
840 if (LoopChain.LoopPredecessors) {
842 dbgs() << "Loop chain contains a block without its preds placed!\n"
843 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
844 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n";
846 for (BlockChain::iterator BCI = LoopChain.begin(), BCE = LoopChain.end();
848 dbgs() << " ... " << getBlockName(*BCI) << "\n";
849 if (!LoopBlockSet.erase(*BCI)) {
850 // We don't mark the loop as bad here because there are real situations
851 // where this can occur. For example, with an unanalyzable fallthrough
852 // from a loop block to a non-loop block or vice versa.
853 dbgs() << "Loop chain contains a block not contained by the loop!\n"
854 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
855 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
856 << " Bad block: " << getBlockName(*BCI) << "\n";
860 if (!LoopBlockSet.empty()) {
862 for (BlockFilterSet::iterator LBI = LoopBlockSet.begin(),
863 LBE = LoopBlockSet.end();
865 dbgs() << "Loop contains blocks never placed into a chain!\n"
866 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
867 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
868 << " Bad block: " << getBlockName(*LBI) << "\n";
870 assert(!BadLoop && "Detected problems with the placement of this loop.");
874 void MachineBlockPlacement::buildCFGChains(MachineFunction &F) {
875 // Ensure that every BB in the function has an associated chain to simplify
876 // the assumptions of the remaining algorithm.
877 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
878 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
879 MachineBasicBlock *BB = FI;
881 = new (ChainAllocator.Allocate()) BlockChain(BlockToChain, BB);
882 // Also, merge any blocks which we cannot reason about and must preserve
883 // the exact fallthrough behavior for.
886 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch.
887 if (!TII->AnalyzeBranch(*BB, TBB, FBB, Cond) || !FI->canFallThrough())
890 MachineFunction::iterator NextFI(std::next(FI));
891 MachineBasicBlock *NextBB = NextFI;
892 // Ensure that the layout successor is a viable block, as we know that
893 // fallthrough is a possibility.
894 assert(NextFI != FE && "Can't fallthrough past the last block.");
895 DEBUG(dbgs() << "Pre-merging due to unanalyzable fallthrough: "
896 << getBlockName(BB) << " -> " << getBlockName(NextBB)
898 Chain->merge(NextBB, 0);
904 // Build any loop-based chains.
905 for (MachineLoopInfo::iterator LI = MLI->begin(), LE = MLI->end(); LI != LE;
907 buildLoopChains(F, **LI);
909 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
911 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
912 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
913 MachineBasicBlock *BB = &*FI;
914 BlockChain &Chain = *BlockToChain[BB];
915 if (!UpdatedPreds.insert(&Chain))
918 assert(Chain.LoopPredecessors == 0);
919 for (BlockChain::iterator BCI = Chain.begin(), BCE = Chain.end();
921 assert(BlockToChain[*BCI] == &Chain);
922 for (MachineBasicBlock::pred_iterator PI = (*BCI)->pred_begin(),
923 PE = (*BCI)->pred_end();
925 if (BlockToChain[*PI] == &Chain)
927 ++Chain.LoopPredecessors;
931 if (Chain.LoopPredecessors == 0)
932 BlockWorkList.push_back(*Chain.begin());
935 BlockChain &FunctionChain = *BlockToChain[&F.front()];
936 buildChain(&F.front(), FunctionChain, BlockWorkList);
939 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 = std::prev(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 // The "PrevBB" is not yet updated to reflect current code layout, so,
993 // o. it may fall-through to a block without explict "goto" instruction
994 // before layout, and no longer fall-through it after layout; or
997 // AnalyzeBranch() may return erroneous value for FBB when these two
998 // situations take place. For the first scenario FBB is mistakenly set
999 // NULL; for the 2nd scenario, the FBB, which is expected to be NULL,
1000 // is mistakenly pointing to "*BI".
1002 bool needUpdateBr = true;
1003 if (!Cond.empty() && (!FBB || FBB == *BI)) {
1004 PrevBB->updateTerminator();
1005 needUpdateBr = false;
1008 if (TII->AnalyzeBranch(*PrevBB, TBB, FBB, Cond)) {
1009 // FIXME: This should never take place.
1014 // If PrevBB has a two-way branch, try to re-order the branches
1015 // such that we branch to the successor with higher weight first.
1016 if (TBB && !Cond.empty() && FBB &&
1017 MBPI->getEdgeWeight(PrevBB, FBB) > MBPI->getEdgeWeight(PrevBB, TBB) &&
1018 !TII->ReverseBranchCondition(Cond)) {
1019 DEBUG(dbgs() << "Reverse order of the two branches: "
1020 << getBlockName(PrevBB) << "\n");
1021 DEBUG(dbgs() << " Edge weight: " << MBPI->getEdgeWeight(PrevBB, FBB)
1022 << " vs " << MBPI->getEdgeWeight(PrevBB, TBB) << "\n");
1023 DebugLoc dl; // FIXME: this is nowhere
1024 TII->RemoveBranch(*PrevBB);
1025 TII->InsertBranch(*PrevBB, FBB, TBB, Cond, dl);
1026 needUpdateBr = true;
1029 PrevBB->updateTerminator();
1033 // Fixup the last block.
1035 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch.
1036 if (!TII->AnalyzeBranch(F.back(), TBB, FBB, Cond))
1037 F.back().updateTerminator();
1039 // Walk through the backedges of the function now that we have fully laid out
1040 // the basic blocks and align the destination of each backedge. We don't rely
1041 // exclusively on the loop info here so that we can align backedges in
1042 // unnatural CFGs and backedges that were introduced purely because of the
1043 // loop rotations done during this layout pass.
1044 if (F.getFunction()->getAttributes().
1045 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize))
1047 unsigned Align = TLI->getPrefLoopAlignment();
1049 return; // Don't care about loop alignment.
1050 if (FunctionChain.begin() == FunctionChain.end())
1051 return; // Empty chain.
1053 const BranchProbability ColdProb(1, 5); // 20%
1054 BlockFrequency EntryFreq = MBFI->getBlockFreq(F.begin());
1055 BlockFrequency WeightedEntryFreq = EntryFreq * ColdProb;
1056 for (BlockChain::iterator BI = std::next(FunctionChain.begin()),
1057 BE = FunctionChain.end();
1059 // Don't align non-looping basic blocks. These are unlikely to execute
1060 // enough times to matter in practice. Note that we'll still handle
1061 // unnatural CFGs inside of a natural outer loop (the common case) and
1063 MachineLoop *L = MLI->getLoopFor(*BI);
1067 // If the block is cold relative to the function entry don't waste space
1069 BlockFrequency Freq = MBFI->getBlockFreq(*BI);
1070 if (Freq < WeightedEntryFreq)
1073 // If the block is cold relative to its loop header, don't align it
1074 // regardless of what edges into the block exist.
1075 MachineBasicBlock *LoopHeader = L->getHeader();
1076 BlockFrequency LoopHeaderFreq = MBFI->getBlockFreq(LoopHeader);
1077 if (Freq < (LoopHeaderFreq * ColdProb))
1080 // Check for the existence of a non-layout predecessor which would benefit
1081 // from aligning this block.
1082 MachineBasicBlock *LayoutPred = *std::prev(BI);
1084 // Force alignment if all the predecessors are jumps. We already checked
1085 // that the block isn't cold above.
1086 if (!LayoutPred->isSuccessor(*BI)) {
1087 (*BI)->setAlignment(Align);
1091 // Align this block if the layout predecessor's edge into this block is
1092 // cold relative to the block. When this is true, other predecessors make up
1093 // all of the hot entries into the block and thus alignment is likely to be
1095 BranchProbability LayoutProb = MBPI->getEdgeProbability(LayoutPred, *BI);
1096 BlockFrequency LayoutEdgeFreq = MBFI->getBlockFreq(LayoutPred) * LayoutProb;
1097 if (LayoutEdgeFreq <= (Freq * ColdProb))
1098 (*BI)->setAlignment(Align);
1102 bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &F) {
1103 // Check for single-block functions and skip them.
1104 if (std::next(F.begin()) == F.end())
1107 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
1108 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
1109 MLI = &getAnalysis<MachineLoopInfo>();
1110 TII = F.getTarget().getInstrInfo();
1111 TLI = F.getTarget().getTargetLowering();
1112 assert(BlockToChain.empty());
1116 BlockToChain.clear();
1117 ChainAllocator.DestroyAll();
1120 // Align all of the blocks in the function to a specific alignment.
1121 for (MachineFunction::iterator FI = F.begin(), FE = F.end();
1123 FI->setAlignment(AlignAllBlock);
1125 // We always return true as we have no way to track whether the final order
1126 // differs from the original order.
1131 /// \brief A pass to compute block placement statistics.
1133 /// A separate pass to compute interesting statistics for evaluating block
1134 /// placement. This is separate from the actual placement pass so that they can
1135 /// be computed in the absence of any placement transformations or when using
1136 /// alternative placement strategies.
1137 class MachineBlockPlacementStats : public MachineFunctionPass {
1138 /// \brief A handle to the branch probability pass.
1139 const MachineBranchProbabilityInfo *MBPI;
1141 /// \brief A handle to the function-wide block frequency pass.
1142 const MachineBlockFrequencyInfo *MBFI;
1145 static char ID; // Pass identification, replacement for typeid
1146 MachineBlockPlacementStats() : MachineFunctionPass(ID) {
1147 initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry());
1150 bool runOnMachineFunction(MachineFunction &F);
1152 void getAnalysisUsage(AnalysisUsage &AU) const {
1153 AU.addRequired<MachineBranchProbabilityInfo>();
1154 AU.addRequired<MachineBlockFrequencyInfo>();
1155 AU.setPreservesAll();
1156 MachineFunctionPass::getAnalysisUsage(AU);
1161 char MachineBlockPlacementStats::ID = 0;
1162 char &llvm::MachineBlockPlacementStatsID = MachineBlockPlacementStats::ID;
1163 INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats",
1164 "Basic Block Placement Stats", false, false)
1165 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
1166 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
1167 INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats",
1168 "Basic Block Placement Stats", false, false)
1170 bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) {
1171 // Check for single-block functions and skip them.
1172 if (std::next(F.begin()) == F.end())
1175 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
1176 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
1178 for (MachineFunction::iterator I = F.begin(), E = F.end(); I != E; ++I) {
1179 BlockFrequency BlockFreq = MBFI->getBlockFreq(I);
1180 Statistic &NumBranches = (I->succ_size() > 1) ? NumCondBranches
1181 : NumUncondBranches;
1182 Statistic &BranchTakenFreq = (I->succ_size() > 1) ? CondBranchTakenFreq
1183 : UncondBranchTakenFreq;
1184 for (MachineBasicBlock::succ_iterator SI = I->succ_begin(),
1187 // Skip if this successor is a fallthrough.
1188 if (I->isLayoutSuccessor(*SI))
1191 BlockFrequency EdgeFreq = BlockFreq * MBPI->getEdgeProbability(I, *SI);
1193 BranchTakenFreq += EdgeFreq.getFrequency();