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 #include "llvm/CodeGen/Passes.h"
29 #include "llvm/ADT/DenseMap.h"
30 #include "llvm/ADT/SmallPtrSet.h"
31 #include "llvm/ADT/SmallVector.h"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/CodeGen/MachineBasicBlock.h"
34 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
35 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
36 #include "llvm/CodeGen/MachineDominators.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/Support/raw_ostream.h"
45 #include "llvm/Target/TargetInstrInfo.h"
46 #include "llvm/Target/TargetLowering.h"
47 #include "llvm/Target/TargetSubtargetInfo.h"
51 #define DEBUG_TYPE "block-placement"
53 STATISTIC(NumCondBranches, "Number of conditional branches");
54 STATISTIC(NumUncondBranches, "Number of unconditional branches");
55 STATISTIC(CondBranchTakenFreq,
56 "Potential frequency of taking conditional branches");
57 STATISTIC(UncondBranchTakenFreq,
58 "Potential frequency of taking unconditional branches");
60 static cl::opt<unsigned> AlignAllBlock("align-all-blocks",
61 cl::desc("Force the alignment of all "
62 "blocks in the function."),
63 cl::init(0), cl::Hidden);
65 // FIXME: Find a good default for this flag and remove the flag.
66 static cl::opt<unsigned> ExitBlockBias(
67 "block-placement-exit-block-bias",
68 cl::desc("Block frequency percentage a loop exit block needs "
69 "over the original exit to be considered the new exit."),
70 cl::init(0), cl::Hidden);
72 static cl::opt<bool> OutlineOptionalBranches(
73 "outline-optional-branches",
74 cl::desc("Put completely optional branches, i.e. branches with a common "
75 "post dominator, out of line."),
76 cl::init(false), cl::Hidden);
78 static cl::opt<unsigned> OutlineOptionalThreshold(
79 "outline-optional-threshold",
80 cl::desc("Don't outline optional branches that are a single block with an "
81 "instruction count below this threshold"),
82 cl::init(4), cl::Hidden);
85 PreciseRotationCost("precise-rotation-cost",
86 cl::desc("Model the cost of loop rotation more "
87 "precisely by using profile data."),
88 cl::init(false), cl::Hidden);
90 static cl::opt<unsigned> MisfetchCost(
92 cl::desc("Cost that models the probablistic risk of an instruction "
93 "misfetch due to a jump comparing to falling through, whose cost "
95 cl::init(1), cl::Hidden);
97 static cl::opt<unsigned> JumpInstCost("jump-inst-cost",
98 cl::desc("Cost of jump instructions."),
99 cl::init(1), cl::Hidden);
103 /// \brief Type for our function-wide basic block -> block chain mapping.
104 typedef DenseMap<MachineBasicBlock *, BlockChain *> BlockToChainMapType;
108 /// \brief A chain of blocks which will be laid out contiguously.
110 /// This is the datastructure representing a chain of consecutive blocks that
111 /// are profitable to layout together in order to maximize fallthrough
112 /// probabilities and code locality. We also can use a block chain to represent
113 /// a sequence of basic blocks which have some external (correctness)
114 /// requirement for sequential layout.
116 /// Chains can be built around a single basic block and can be merged to grow
117 /// them. They participate in a block-to-chain mapping, which is updated
118 /// automatically as chains are merged together.
120 /// \brief The sequence of blocks belonging to this chain.
122 /// This is the sequence of blocks for a particular chain. These will be laid
123 /// out in-order within the function.
124 SmallVector<MachineBasicBlock *, 4> Blocks;
126 /// \brief A handle to the function-wide basic block to block chain mapping.
128 /// This is retained in each block chain to simplify the computation of child
129 /// block chains for SCC-formation and iteration. We store the edges to child
130 /// basic blocks, and map them back to their associated chains using this
132 BlockToChainMapType &BlockToChain;
135 /// \brief Construct a new BlockChain.
137 /// This builds a new block chain representing a single basic block in the
138 /// function. It also registers itself as the chain that block participates
139 /// in with the BlockToChain mapping.
140 BlockChain(BlockToChainMapType &BlockToChain, MachineBasicBlock *BB)
141 : Blocks(1, BB), BlockToChain(BlockToChain), LoopPredecessors(0) {
142 assert(BB && "Cannot create a chain with a null basic block");
143 BlockToChain[BB] = this;
146 /// \brief Iterator over blocks within the chain.
147 typedef SmallVectorImpl<MachineBasicBlock *>::iterator iterator;
149 /// \brief Beginning of blocks within the chain.
150 iterator begin() { return Blocks.begin(); }
152 /// \brief End of blocks within the chain.
153 iterator end() { return Blocks.end(); }
155 /// \brief Merge a block chain into this one.
157 /// This routine merges a block chain into this one. It takes care of forming
158 /// a contiguous sequence of basic blocks, updating the edge list, and
159 /// updating the block -> chain mapping. It does not free or tear down the
160 /// old chain, but the old chain's block list is no longer valid.
161 void merge(MachineBasicBlock *BB, BlockChain *Chain) {
163 assert(!Blocks.empty());
165 // Fast path in case we don't have a chain already.
167 assert(!BlockToChain[BB]);
168 Blocks.push_back(BB);
169 BlockToChain[BB] = this;
173 assert(BB == *Chain->begin());
174 assert(Chain->begin() != Chain->end());
176 // Update the incoming blocks to point to this chain, and add them to the
178 for (MachineBasicBlock *ChainBB : *Chain) {
179 Blocks.push_back(ChainBB);
180 assert(BlockToChain[ChainBB] == Chain && "Incoming blocks not in chain");
181 BlockToChain[ChainBB] = this;
186 /// \brief Dump the blocks in this chain.
187 LLVM_DUMP_METHOD void dump() {
188 for (MachineBasicBlock *MBB : *this)
193 /// \brief Count of predecessors within the loop currently being processed.
195 /// This count is updated at each loop we process to represent the number of
196 /// in-loop predecessors of this chain.
197 unsigned LoopPredecessors;
202 class MachineBlockPlacement : public MachineFunctionPass {
203 /// \brief A typedef for a block filter set.
204 typedef SmallPtrSet<MachineBasicBlock *, 16> BlockFilterSet;
206 /// \brief A handle to the branch probability pass.
207 const MachineBranchProbabilityInfo *MBPI;
209 /// \brief A handle to the function-wide block frequency pass.
210 const MachineBlockFrequencyInfo *MBFI;
212 /// \brief A handle to the loop info.
213 const MachineLoopInfo *MLI;
215 /// \brief A handle to the target's instruction info.
216 const TargetInstrInfo *TII;
218 /// \brief A handle to the target's lowering info.
219 const TargetLoweringBase *TLI;
221 /// \brief A handle to the post dominator tree.
222 MachineDominatorTree *MDT;
224 /// \brief A set of blocks that are unavoidably execute, i.e. they dominate
225 /// all terminators of the MachineFunction.
226 SmallPtrSet<MachineBasicBlock *, 4> UnavoidableBlocks;
228 /// \brief Allocator and owner of BlockChain structures.
230 /// We build BlockChains lazily while processing the loop structure of
231 /// a function. To reduce malloc traffic, we allocate them using this
232 /// slab-like allocator, and destroy them after the pass completes. An
233 /// important guarantee is that this allocator produces stable pointers to
235 SpecificBumpPtrAllocator<BlockChain> ChainAllocator;
237 /// \brief Function wide BasicBlock to BlockChain mapping.
239 /// This mapping allows efficiently moving from any given basic block to the
240 /// BlockChain it participates in, if any. We use it to, among other things,
241 /// allow implicitly defining edges between chains as the existing edges
242 /// between basic blocks.
243 DenseMap<MachineBasicBlock *, BlockChain *> BlockToChain;
245 void markChainSuccessors(BlockChain &Chain, MachineBasicBlock *LoopHeaderBB,
246 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
247 const BlockFilterSet *BlockFilter = nullptr);
248 MachineBasicBlock *selectBestSuccessor(MachineBasicBlock *BB,
250 const BlockFilterSet *BlockFilter);
252 selectBestCandidateBlock(BlockChain &Chain,
253 SmallVectorImpl<MachineBasicBlock *> &WorkList,
254 const BlockFilterSet *BlockFilter);
256 getFirstUnplacedBlock(MachineFunction &F, const BlockChain &PlacedChain,
257 MachineFunction::iterator &PrevUnplacedBlockIt,
258 const BlockFilterSet *BlockFilter);
259 void buildChain(MachineBasicBlock *BB, BlockChain &Chain,
260 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
261 const BlockFilterSet *BlockFilter = nullptr);
262 MachineBasicBlock *findBestLoopTop(MachineLoop &L,
263 const BlockFilterSet &LoopBlockSet);
264 MachineBasicBlock *findBestLoopExit(MachineFunction &F, MachineLoop &L,
265 const BlockFilterSet &LoopBlockSet);
266 void buildLoopChains(MachineFunction &F, MachineLoop &L);
267 void rotateLoop(BlockChain &LoopChain, MachineBasicBlock *ExitingBB,
268 const BlockFilterSet &LoopBlockSet);
269 void rotateLoopWithProfile(BlockChain &LoopChain, MachineLoop &L,
270 const BlockFilterSet &LoopBlockSet);
271 void buildCFGChains(MachineFunction &F);
274 static char ID; // Pass identification, replacement for typeid
275 MachineBlockPlacement() : MachineFunctionPass(ID) {
276 initializeMachineBlockPlacementPass(*PassRegistry::getPassRegistry());
279 bool runOnMachineFunction(MachineFunction &F) override;
281 void getAnalysisUsage(AnalysisUsage &AU) const override {
282 AU.addRequired<MachineBranchProbabilityInfo>();
283 AU.addRequired<MachineBlockFrequencyInfo>();
284 AU.addRequired<MachineDominatorTree>();
285 AU.addRequired<MachineLoopInfo>();
286 MachineFunctionPass::getAnalysisUsage(AU);
291 char MachineBlockPlacement::ID = 0;
292 char &llvm::MachineBlockPlacementID = MachineBlockPlacement::ID;
293 INITIALIZE_PASS_BEGIN(MachineBlockPlacement, "block-placement",
294 "Branch Probability Basic Block Placement", false, false)
295 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
296 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
297 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
298 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
299 INITIALIZE_PASS_END(MachineBlockPlacement, "block-placement",
300 "Branch Probability Basic Block Placement", false, false)
303 /// \brief Helper to print the name of a MBB.
305 /// Only used by debug logging.
306 static std::string getBlockName(MachineBasicBlock *BB) {
308 raw_string_ostream OS(Result);
309 OS << "BB#" << BB->getNumber();
310 OS << " (derived from LLVM BB '" << BB->getName() << "')";
315 /// \brief Helper to print the number of a MBB.
317 /// Only used by debug logging.
318 static std::string getBlockNum(MachineBasicBlock *BB) {
320 raw_string_ostream OS(Result);
321 OS << "BB#" << BB->getNumber();
327 /// \brief Mark a chain's successors as having one fewer preds.
329 /// When a chain is being merged into the "placed" chain, this routine will
330 /// quickly walk the successors of each block in the chain and mark them as
331 /// having one fewer active predecessor. It also adds any successors of this
332 /// chain which reach the zero-predecessor state to the worklist passed in.
333 void MachineBlockPlacement::markChainSuccessors(
334 BlockChain &Chain, MachineBasicBlock *LoopHeaderBB,
335 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
336 const BlockFilterSet *BlockFilter) {
337 // Walk all the blocks in this chain, marking their successors as having
338 // a predecessor placed.
339 for (MachineBasicBlock *MBB : Chain) {
340 // Add any successors for which this is the only un-placed in-loop
341 // predecessor to the worklist as a viable candidate for CFG-neutral
342 // placement. No subsequent placement of this block will violate the CFG
343 // shape, so we get to use heuristics to choose a favorable placement.
344 for (MachineBasicBlock *Succ : MBB->successors()) {
345 if (BlockFilter && !BlockFilter->count(Succ))
347 BlockChain &SuccChain = *BlockToChain[Succ];
348 // Disregard edges within a fixed chain, or edges to the loop header.
349 if (&Chain == &SuccChain || Succ == LoopHeaderBB)
352 // This is a cross-chain edge that is within the loop, so decrement the
353 // loop predecessor count of the destination chain.
354 if (SuccChain.LoopPredecessors > 0 && --SuccChain.LoopPredecessors == 0)
355 BlockWorkList.push_back(*SuccChain.begin());
360 /// \brief Select the best successor for a block.
362 /// This looks across all successors of a particular block and attempts to
363 /// select the "best" one to be the layout successor. It only considers direct
364 /// successors which also pass the block filter. It will attempt to avoid
365 /// breaking CFG structure, but cave and break such structures in the case of
366 /// very hot successor edges.
368 /// \returns The best successor block found, or null if none are viable.
370 MachineBlockPlacement::selectBestSuccessor(MachineBasicBlock *BB,
372 const BlockFilterSet *BlockFilter) {
373 const BranchProbability HotProb(4, 5); // 80%
375 MachineBasicBlock *BestSucc = nullptr;
376 // FIXME: Due to the performance of the probability and weight routines in
377 // the MBPI analysis, we manually compute probabilities using the edge
378 // weights. This is suboptimal as it means that the somewhat subtle
379 // definition of edge weight semantics is encoded here as well. We should
380 // improve the MBPI interface to efficiently support query patterns such as
382 uint32_t BestWeight = 0;
383 uint32_t WeightScale = 0;
384 uint32_t SumWeight = MBPI->getSumForBlock(BB, WeightScale);
385 DEBUG(dbgs() << "Attempting merge from: " << getBlockName(BB) << "\n");
386 for (MachineBasicBlock *Succ : BB->successors()) {
387 if (BlockFilter && !BlockFilter->count(Succ))
389 BlockChain &SuccChain = *BlockToChain[Succ];
390 if (&SuccChain == &Chain) {
391 DEBUG(dbgs() << " " << getBlockName(Succ) << " -> Already merged!\n");
394 if (Succ != *SuccChain.begin()) {
395 DEBUG(dbgs() << " " << getBlockName(Succ) << " -> Mid chain!\n");
399 uint32_t SuccWeight = MBPI->getEdgeWeight(BB, Succ);
400 BranchProbability SuccProb(SuccWeight / WeightScale, SumWeight);
402 // If we outline optional branches, look whether Succ is unavoidable, i.e.
403 // dominates all terminators of the MachineFunction. If it does, other
404 // successors must be optional. Don't do this for cold branches.
405 if (OutlineOptionalBranches && SuccProb > HotProb.getCompl() &&
406 UnavoidableBlocks.count(Succ) > 0) {
407 auto HasShortOptionalBranch = [&]() {
408 for (MachineBasicBlock *Pred : Succ->predecessors()) {
409 // Check whether there is an unplaced optional branch.
410 if (Pred == Succ || (BlockFilter && !BlockFilter->count(Pred)) ||
411 BlockToChain[Pred] == &Chain)
413 // Check whether the optional branch has exactly one BB.
414 if (Pred->pred_size() > 1 || *Pred->pred_begin() != BB)
416 // Check whether the optional branch is small.
417 if (Pred->size() < OutlineOptionalThreshold)
422 if (!HasShortOptionalBranch())
426 // Only consider successors which are either "hot", or wouldn't violate
427 // any CFG constraints.
428 if (SuccChain.LoopPredecessors != 0) {
429 if (SuccProb < HotProb) {
430 DEBUG(dbgs() << " " << getBlockName(Succ) << " -> " << SuccProb
431 << " (prob) (CFG conflict)\n");
435 // Make sure that a hot successor doesn't have a globally more
436 // important predecessor.
437 BlockFrequency CandidateEdgeFreq =
438 MBFI->getBlockFreq(BB) * SuccProb * HotProb.getCompl();
439 bool BadCFGConflict = false;
440 for (MachineBasicBlock *Pred : Succ->predecessors()) {
441 if (Pred == Succ || (BlockFilter && !BlockFilter->count(Pred)) ||
442 BlockToChain[Pred] == &Chain)
444 BlockFrequency PredEdgeFreq =
445 MBFI->getBlockFreq(Pred) * MBPI->getEdgeProbability(Pred, Succ);
446 if (PredEdgeFreq >= CandidateEdgeFreq) {
447 BadCFGConflict = true;
451 if (BadCFGConflict) {
452 DEBUG(dbgs() << " " << getBlockName(Succ) << " -> " << SuccProb
453 << " (prob) (non-cold CFG conflict)\n");
458 DEBUG(dbgs() << " " << getBlockName(Succ) << " -> " << SuccProb
460 << (SuccChain.LoopPredecessors != 0 ? " (CFG break)" : "")
462 if (BestSucc && BestWeight >= SuccWeight)
465 BestWeight = SuccWeight;
470 /// \brief Select the best block from a worklist.
472 /// This looks through the provided worklist as a list of candidate basic
473 /// blocks and select the most profitable one to place. The definition of
474 /// profitable only really makes sense in the context of a loop. This returns
475 /// the most frequently visited block in the worklist, which in the case of
476 /// a loop, is the one most desirable to be physically close to the rest of the
477 /// loop body in order to improve icache behavior.
479 /// \returns The best block found, or null if none are viable.
480 MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock(
481 BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList,
482 const BlockFilterSet *BlockFilter) {
483 // Once we need to walk the worklist looking for a candidate, cleanup the
484 // worklist of already placed entries.
485 // FIXME: If this shows up on profiles, it could be folded (at the cost of
486 // some code complexity) into the loop below.
487 WorkList.erase(std::remove_if(WorkList.begin(), WorkList.end(),
488 [&](MachineBasicBlock *BB) {
489 return BlockToChain.lookup(BB) == &Chain;
493 MachineBasicBlock *BestBlock = nullptr;
494 BlockFrequency BestFreq;
495 for (MachineBasicBlock *MBB : WorkList) {
496 BlockChain &SuccChain = *BlockToChain[MBB];
497 if (&SuccChain == &Chain) {
498 DEBUG(dbgs() << " " << getBlockName(MBB) << " -> Already merged!\n");
501 assert(SuccChain.LoopPredecessors == 0 && "Found CFG-violating block");
503 BlockFrequency CandidateFreq = MBFI->getBlockFreq(MBB);
504 DEBUG(dbgs() << " " << getBlockName(MBB) << " -> ";
505 MBFI->printBlockFreq(dbgs(), CandidateFreq) << " (freq)\n");
506 if (BestBlock && BestFreq >= CandidateFreq)
509 BestFreq = CandidateFreq;
514 /// \brief Retrieve the first unplaced basic block.
516 /// This routine is called when we are unable to use the CFG to walk through
517 /// all of the basic blocks and form a chain due to unnatural loops in the CFG.
518 /// We walk through the function's blocks in order, starting from the
519 /// LastUnplacedBlockIt. We update this iterator on each call to avoid
520 /// re-scanning the entire sequence on repeated calls to this routine.
521 MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock(
522 MachineFunction &F, const BlockChain &PlacedChain,
523 MachineFunction::iterator &PrevUnplacedBlockIt,
524 const BlockFilterSet *BlockFilter) {
525 for (MachineFunction::iterator I = PrevUnplacedBlockIt, E = F.end(); I != E;
527 if (BlockFilter && !BlockFilter->count(&*I))
529 if (BlockToChain[&*I] != &PlacedChain) {
530 PrevUnplacedBlockIt = I;
531 // Now select the head of the chain to which the unplaced block belongs
532 // as the block to place. This will force the entire chain to be placed,
533 // and satisfies the requirements of merging chains.
534 return *BlockToChain[&*I]->begin();
540 void MachineBlockPlacement::buildChain(
541 MachineBasicBlock *BB, BlockChain &Chain,
542 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
543 const BlockFilterSet *BlockFilter) {
545 assert(BlockToChain[BB] == &Chain);
546 MachineFunction &F = *BB->getParent();
547 MachineFunction::iterator PrevUnplacedBlockIt = F.begin();
549 MachineBasicBlock *LoopHeaderBB = BB;
550 markChainSuccessors(Chain, LoopHeaderBB, BlockWorkList, BlockFilter);
551 BB = *std::prev(Chain.end());
554 assert(BlockToChain[BB] == &Chain);
555 assert(*std::prev(Chain.end()) == BB);
557 // Look for the best viable successor if there is one to place immediately
559 MachineBasicBlock *BestSucc = selectBestSuccessor(BB, Chain, BlockFilter);
561 // If an immediate successor isn't available, look for the best viable
562 // block among those we've identified as not violating the loop's CFG at
563 // this point. This won't be a fallthrough, but it will increase locality.
565 BestSucc = selectBestCandidateBlock(Chain, BlockWorkList, BlockFilter);
569 getFirstUnplacedBlock(F, Chain, PrevUnplacedBlockIt, BlockFilter);
573 DEBUG(dbgs() << "Unnatural loop CFG detected, forcibly merging the "
574 "layout successor until the CFG reduces\n");
577 // Place this block, updating the datastructures to reflect its placement.
578 BlockChain &SuccChain = *BlockToChain[BestSucc];
579 // Zero out LoopPredecessors for the successor we're about to merge in case
580 // we selected a successor that didn't fit naturally into the CFG.
581 SuccChain.LoopPredecessors = 0;
582 DEBUG(dbgs() << "Merging from " << getBlockNum(BB) << " to "
583 << getBlockNum(BestSucc) << "\n");
584 markChainSuccessors(SuccChain, LoopHeaderBB, BlockWorkList, BlockFilter);
585 Chain.merge(BestSucc, &SuccChain);
586 BB = *std::prev(Chain.end());
589 DEBUG(dbgs() << "Finished forming chain for header block "
590 << getBlockNum(*Chain.begin()) << "\n");
593 /// \brief Find the best loop top block for layout.
595 /// Look for a block which is strictly better than the loop header for laying
596 /// out at the top of the loop. This looks for one and only one pattern:
597 /// a latch block with no conditional exit. This block will cause a conditional
598 /// jump around it or will be the bottom of the loop if we lay it out in place,
599 /// but if it it doesn't end up at the bottom of the loop for any reason,
600 /// rotation alone won't fix it. Because such a block will always result in an
601 /// unconditional jump (for the backedge) rotating it in front of the loop
602 /// header is always profitable.
604 MachineBlockPlacement::findBestLoopTop(MachineLoop &L,
605 const BlockFilterSet &LoopBlockSet) {
606 // Check that the header hasn't been fused with a preheader block due to
607 // crazy branches. If it has, we need to start with the header at the top to
608 // prevent pulling the preheader into the loop body.
609 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
610 if (!LoopBlockSet.count(*HeaderChain.begin()))
611 return L.getHeader();
613 DEBUG(dbgs() << "Finding best loop top for: " << getBlockName(L.getHeader())
616 BlockFrequency BestPredFreq;
617 MachineBasicBlock *BestPred = nullptr;
618 for (MachineBasicBlock *Pred : L.getHeader()->predecessors()) {
619 if (!LoopBlockSet.count(Pred))
621 DEBUG(dbgs() << " header pred: " << getBlockName(Pred) << ", "
622 << Pred->succ_size() << " successors, ";
623 MBFI->printBlockFreq(dbgs(), Pred) << " freq\n");
624 if (Pred->succ_size() > 1)
627 BlockFrequency PredFreq = MBFI->getBlockFreq(Pred);
628 if (!BestPred || PredFreq > BestPredFreq ||
629 (!(PredFreq < BestPredFreq) &&
630 Pred->isLayoutSuccessor(L.getHeader()))) {
632 BestPredFreq = PredFreq;
636 // If no direct predecessor is fine, just use the loop header.
638 return L.getHeader();
640 // Walk backwards through any straight line of predecessors.
641 while (BestPred->pred_size() == 1 &&
642 (*BestPred->pred_begin())->succ_size() == 1 &&
643 *BestPred->pred_begin() != L.getHeader())
644 BestPred = *BestPred->pred_begin();
646 DEBUG(dbgs() << " final top: " << getBlockName(BestPred) << "\n");
650 /// \brief Find the best loop exiting block for layout.
652 /// This routine implements the logic to analyze the loop looking for the best
653 /// block to layout at the top of the loop. Typically this is done to maximize
654 /// fallthrough opportunities.
656 MachineBlockPlacement::findBestLoopExit(MachineFunction &F, MachineLoop &L,
657 const BlockFilterSet &LoopBlockSet) {
658 // We don't want to layout the loop linearly in all cases. If the loop header
659 // is just a normal basic block in the loop, we want to look for what block
660 // within the loop is the best one to layout at the top. However, if the loop
661 // header has be pre-merged into a chain due to predecessors not having
662 // analyzable branches, *and* the predecessor it is merged with is *not* part
663 // of the loop, rotating the header into the middle of the loop will create
664 // a non-contiguous range of blocks which is Very Bad. So start with the
665 // header and only rotate if safe.
666 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
667 if (!LoopBlockSet.count(*HeaderChain.begin()))
670 BlockFrequency BestExitEdgeFreq;
671 unsigned BestExitLoopDepth = 0;
672 MachineBasicBlock *ExitingBB = nullptr;
673 // If there are exits to outer loops, loop rotation can severely limit
674 // fallthrough opportunites unless it selects such an exit. Keep a set of
675 // blocks where rotating to exit with that block will reach an outer loop.
676 SmallPtrSet<MachineBasicBlock *, 4> BlocksExitingToOuterLoop;
678 DEBUG(dbgs() << "Finding best loop exit for: " << getBlockName(L.getHeader())
680 for (MachineBasicBlock *MBB : L.getBlocks()) {
681 BlockChain &Chain = *BlockToChain[MBB];
682 // Ensure that this block is at the end of a chain; otherwise it could be
683 // mid-way through an inner loop or a successor of an unanalyzable branch.
684 if (MBB != *std::prev(Chain.end()))
687 // Now walk the successors. We need to establish whether this has a viable
688 // exiting successor and whether it has a viable non-exiting successor.
689 // We store the old exiting state and restore it if a viable looping
690 // successor isn't found.
691 MachineBasicBlock *OldExitingBB = ExitingBB;
692 BlockFrequency OldBestExitEdgeFreq = BestExitEdgeFreq;
693 bool HasLoopingSucc = false;
694 // FIXME: Due to the performance of the probability and weight routines in
695 // the MBPI analysis, we use the internal weights and manually compute the
696 // probabilities to avoid quadratic behavior.
697 uint32_t WeightScale = 0;
698 uint32_t SumWeight = MBPI->getSumForBlock(MBB, WeightScale);
699 for (MachineBasicBlock *Succ : MBB->successors()) {
704 BlockChain &SuccChain = *BlockToChain[Succ];
705 // Don't split chains, either this chain or the successor's chain.
706 if (&Chain == &SuccChain) {
707 DEBUG(dbgs() << " exiting: " << getBlockName(MBB) << " -> "
708 << getBlockName(Succ) << " (chain conflict)\n");
712 uint32_t SuccWeight = MBPI->getEdgeWeight(MBB, Succ);
713 if (LoopBlockSet.count(Succ)) {
714 DEBUG(dbgs() << " looping: " << getBlockName(MBB) << " -> "
715 << getBlockName(Succ) << " (" << SuccWeight << ")\n");
716 HasLoopingSucc = true;
720 unsigned SuccLoopDepth = 0;
721 if (MachineLoop *ExitLoop = MLI->getLoopFor(Succ)) {
722 SuccLoopDepth = ExitLoop->getLoopDepth();
723 if (ExitLoop->contains(&L))
724 BlocksExitingToOuterLoop.insert(MBB);
727 BranchProbability SuccProb(SuccWeight / WeightScale, SumWeight);
728 BlockFrequency ExitEdgeFreq = MBFI->getBlockFreq(MBB) * SuccProb;
729 DEBUG(dbgs() << " exiting: " << getBlockName(MBB) << " -> "
730 << getBlockName(Succ) << " [L:" << SuccLoopDepth << "] (";
731 MBFI->printBlockFreq(dbgs(), ExitEdgeFreq) << ")\n");
732 // Note that we bias this toward an existing layout successor to retain
733 // incoming order in the absence of better information. The exit must have
734 // a frequency higher than the current exit before we consider breaking
736 BranchProbability Bias(100 - ExitBlockBias, 100);
737 if (!ExitingBB || SuccLoopDepth > BestExitLoopDepth ||
738 ExitEdgeFreq > BestExitEdgeFreq ||
739 (MBB->isLayoutSuccessor(Succ) &&
740 !(ExitEdgeFreq < BestExitEdgeFreq * Bias))) {
741 BestExitEdgeFreq = ExitEdgeFreq;
746 if (!HasLoopingSucc) {
747 // Restore the old exiting state, no viable looping successor was found.
748 ExitingBB = OldExitingBB;
749 BestExitEdgeFreq = OldBestExitEdgeFreq;
753 // Without a candidate exiting block or with only a single block in the
754 // loop, just use the loop header to layout the loop.
755 if (!ExitingBB || L.getNumBlocks() == 1)
758 // Also, if we have exit blocks which lead to outer loops but didn't select
759 // one of them as the exiting block we are rotating toward, disable loop
760 // rotation altogether.
761 if (!BlocksExitingToOuterLoop.empty() &&
762 !BlocksExitingToOuterLoop.count(ExitingBB))
765 DEBUG(dbgs() << " Best exiting block: " << getBlockName(ExitingBB) << "\n");
769 /// \brief Attempt to rotate an exiting block to the bottom of the loop.
771 /// Once we have built a chain, try to rotate it to line up the hot exit block
772 /// with fallthrough out of the loop if doing so doesn't introduce unnecessary
773 /// branches. For example, if the loop has fallthrough into its header and out
774 /// of its bottom already, don't rotate it.
775 void MachineBlockPlacement::rotateLoop(BlockChain &LoopChain,
776 MachineBasicBlock *ExitingBB,
777 const BlockFilterSet &LoopBlockSet) {
781 MachineBasicBlock *Top = *LoopChain.begin();
782 bool ViableTopFallthrough = false;
783 for (MachineBasicBlock *Pred : Top->predecessors()) {
784 BlockChain *PredChain = BlockToChain[Pred];
785 if (!LoopBlockSet.count(Pred) &&
786 (!PredChain || Pred == *std::prev(PredChain->end()))) {
787 ViableTopFallthrough = true;
792 // If the header has viable fallthrough, check whether the current loop
793 // bottom is a viable exiting block. If so, bail out as rotating will
794 // introduce an unnecessary branch.
795 if (ViableTopFallthrough) {
796 MachineBasicBlock *Bottom = *std::prev(LoopChain.end());
797 for (MachineBasicBlock *Succ : Bottom->successors()) {
798 BlockChain *SuccChain = BlockToChain[Succ];
799 if (!LoopBlockSet.count(Succ) &&
800 (!SuccChain || Succ == *SuccChain->begin()))
805 BlockChain::iterator ExitIt =
806 std::find(LoopChain.begin(), LoopChain.end(), ExitingBB);
807 if (ExitIt == LoopChain.end())
810 std::rotate(LoopChain.begin(), std::next(ExitIt), LoopChain.end());
813 /// \brief Attempt to rotate a loop based on profile data to reduce branch cost.
815 /// With profile data, we can determine the cost in terms of missed fall through
816 /// opportunities when rotating a loop chain and select the best rotation.
817 /// Basically, there are three kinds of cost to consider for each rotation:
818 /// 1. The possibly missed fall through edge (if it exists) from BB out of
819 /// the loop to the loop header.
820 /// 2. The possibly missed fall through edges (if they exist) from the loop
821 /// exits to BB out of the loop.
822 /// 3. The missed fall through edge (if it exists) from the last BB to the
823 /// first BB in the loop chain.
824 /// Therefore, the cost for a given rotation is the sum of costs listed above.
825 /// We select the best rotation with the smallest cost.
826 void MachineBlockPlacement::rotateLoopWithProfile(
827 BlockChain &LoopChain, MachineLoop &L, const BlockFilterSet &LoopBlockSet) {
828 auto HeaderBB = L.getHeader();
829 auto HeaderIter = std::find(LoopChain.begin(), LoopChain.end(), HeaderBB);
830 auto RotationPos = LoopChain.end();
832 BlockFrequency SmallestRotationCost = BlockFrequency::getMaxFrequency();
834 // A utility lambda that scales up a block frequency by dividing it by a
835 // branch probability which is the reciprocal of the scale.
836 auto ScaleBlockFrequency = [](BlockFrequency Freq,
837 unsigned Scale) -> BlockFrequency {
840 // Use operator / between BlockFrequency and BranchProbability to implement
841 // saturating multiplication.
842 return Freq / BranchProbability(1, Scale);
845 // Compute the cost of the missed fall-through edge to the loop header if the
846 // chain head is not the loop header. As we only consider natural loops with
847 // single header, this computation can be done only once.
848 BlockFrequency HeaderFallThroughCost(0);
849 for (auto *Pred : HeaderBB->predecessors()) {
850 BlockChain *PredChain = BlockToChain[Pred];
851 if (!LoopBlockSet.count(Pred) &&
852 (!PredChain || Pred == *std::prev(PredChain->end()))) {
854 MBFI->getBlockFreq(Pred) * MBPI->getEdgeProbability(Pred, HeaderBB);
855 auto FallThruCost = ScaleBlockFrequency(EdgeFreq, MisfetchCost);
856 // If the predecessor has only an unconditional jump to the header, we
857 // need to consider the cost of this jump.
858 if (Pred->succ_size() == 1)
859 FallThruCost += ScaleBlockFrequency(EdgeFreq, JumpInstCost);
860 HeaderFallThroughCost = std::max(HeaderFallThroughCost, FallThruCost);
864 // Here we collect all exit blocks in the loop, and for each exit we find out
865 // its hottest exit edge. For each loop rotation, we define the loop exit cost
866 // as the sum of frequencies of exit edges we collect here, excluding the exit
867 // edge from the tail of the loop chain.
868 SmallVector<std::pair<MachineBasicBlock *, BlockFrequency>, 4> ExitsWithFreq;
869 for (auto BB : LoopChain) {
870 uint32_t LargestExitEdgeWeight = 0;
871 for (auto *Succ : BB->successors()) {
872 BlockChain *SuccChain = BlockToChain[Succ];
873 if (!LoopBlockSet.count(Succ) &&
874 (!SuccChain || Succ == *SuccChain->begin())) {
875 uint32_t SuccWeight = MBPI->getEdgeWeight(BB, Succ);
876 LargestExitEdgeWeight = std::max(LargestExitEdgeWeight, SuccWeight);
879 if (LargestExitEdgeWeight > 0) {
880 uint32_t WeightScale = 0;
881 uint32_t SumWeight = MBPI->getSumForBlock(BB, WeightScale);
883 MBFI->getBlockFreq(BB) *
884 BranchProbability(LargestExitEdgeWeight / WeightScale, SumWeight);
885 ExitsWithFreq.emplace_back(BB, ExitFreq);
889 // In this loop we iterate every block in the loop chain and calculate the
890 // cost assuming the block is the head of the loop chain. When the loop ends,
891 // we should have found the best candidate as the loop chain's head.
892 for (auto Iter = LoopChain.begin(), TailIter = std::prev(LoopChain.end()),
893 EndIter = LoopChain.end();
894 Iter != EndIter; Iter++, TailIter++) {
895 // TailIter is used to track the tail of the loop chain if the block we are
896 // checking (pointed by Iter) is the head of the chain.
897 if (TailIter == LoopChain.end())
898 TailIter = LoopChain.begin();
900 auto TailBB = *TailIter;
902 // Calculate the cost by putting this BB to the top.
903 BlockFrequency Cost = 0;
905 // If the current BB is the loop header, we need to take into account the
906 // cost of the missed fall through edge from outside of the loop to the
908 if (Iter != HeaderIter)
909 Cost += HeaderFallThroughCost;
911 // Collect the loop exit cost by summing up frequencies of all exit edges
912 // except the one from the chain tail.
913 for (auto &ExitWithFreq : ExitsWithFreq)
914 if (TailBB != ExitWithFreq.first)
915 Cost += ExitWithFreq.second;
917 // The cost of breaking the once fall-through edge from the tail to the top
918 // of the loop chain. Here we need to consider three cases:
919 // 1. If the tail node has only one successor, then we will get an
920 // additional jmp instruction. So the cost here is (MisfetchCost +
921 // JumpInstCost) * tail node frequency.
922 // 2. If the tail node has two successors, then we may still get an
923 // additional jmp instruction if the layout successor after the loop
924 // chain is not its CFG successor. Note that the more frequently executed
925 // jmp instruction will be put ahead of the other one. Assume the
926 // frequency of those two branches are x and y, where x is the frequency
927 // of the edge to the chain head, then the cost will be
928 // (x * MisfetechCost + min(x, y) * JumpInstCost) * tail node frequency.
929 // 3. If the tail node has more than two successors (this rarely happens),
930 // we won't consider any additional cost.
931 if (TailBB->isSuccessor(*Iter)) {
932 auto TailBBFreq = MBFI->getBlockFreq(TailBB);
933 if (TailBB->succ_size() == 1)
934 Cost += ScaleBlockFrequency(TailBBFreq.getFrequency(),
935 MisfetchCost + JumpInstCost);
936 else if (TailBB->succ_size() == 2) {
937 auto TailToHeadProb = MBPI->getEdgeProbability(TailBB, *Iter);
938 auto TailToHeadFreq = TailBBFreq * TailToHeadProb;
939 auto ColderEdgeFreq = TailToHeadProb > BranchProbability(1, 2)
940 ? TailBBFreq * TailToHeadProb.getCompl()
942 Cost += ScaleBlockFrequency(TailToHeadFreq, MisfetchCost) +
943 ScaleBlockFrequency(ColderEdgeFreq, JumpInstCost);
947 DEBUG(dbgs() << "The cost of loop rotation by making " << getBlockNum(*Iter)
948 << " to the top: " << Cost.getFrequency() << "\n");
950 if (Cost < SmallestRotationCost) {
951 SmallestRotationCost = Cost;
956 if (RotationPos != LoopChain.end()) {
957 DEBUG(dbgs() << "Rotate loop by making " << getBlockNum(*RotationPos)
959 std::rotate(LoopChain.begin(), RotationPos, LoopChain.end());
963 /// \brief Forms basic block chains from the natural loop structures.
965 /// These chains are designed to preserve the existing *structure* of the code
966 /// as much as possible. We can then stitch the chains together in a way which
967 /// both preserves the topological structure and minimizes taken conditional
969 void MachineBlockPlacement::buildLoopChains(MachineFunction &F,
971 // First recurse through any nested loops, building chains for those inner
973 for (MachineLoop *InnerLoop : L)
974 buildLoopChains(F, *InnerLoop);
976 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
977 BlockFilterSet LoopBlockSet(L.block_begin(), L.block_end());
979 // Check if we have profile data for this function. If yes, we will rotate
980 // this loop by modeling costs more precisely which requires the profile data
981 // for better layout.
982 bool RotateLoopWithProfile =
983 PreciseRotationCost && F.getFunction()->getEntryCount();
985 // First check to see if there is an obviously preferable top block for the
986 // loop. This will default to the header, but may end up as one of the
987 // predecessors to the header if there is one which will result in strictly
988 // fewer branches in the loop body.
989 // When we use profile data to rotate the loop, this is unnecessary.
990 MachineBasicBlock *LoopTop =
991 RotateLoopWithProfile ? L.getHeader() : findBestLoopTop(L, LoopBlockSet);
993 // If we selected just the header for the loop top, look for a potentially
994 // profitable exit block in the event that rotating the loop can eliminate
995 // branches by placing an exit edge at the bottom.
996 MachineBasicBlock *ExitingBB = nullptr;
997 if (!RotateLoopWithProfile && LoopTop == L.getHeader())
998 ExitingBB = findBestLoopExit(F, L, LoopBlockSet);
1000 BlockChain &LoopChain = *BlockToChain[LoopTop];
1002 // FIXME: This is a really lame way of walking the chains in the loop: we
1003 // walk the blocks, and use a set to prevent visiting a particular chain
1005 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
1006 assert(LoopChain.LoopPredecessors == 0);
1007 UpdatedPreds.insert(&LoopChain);
1008 for (MachineBasicBlock *LoopBB : L.getBlocks()) {
1009 BlockChain &Chain = *BlockToChain[LoopBB];
1010 if (!UpdatedPreds.insert(&Chain).second)
1013 assert(Chain.LoopPredecessors == 0);
1014 for (MachineBasicBlock *ChainBB : Chain) {
1015 assert(BlockToChain[ChainBB] == &Chain);
1016 for (MachineBasicBlock *Pred : ChainBB->predecessors()) {
1017 if (BlockToChain[Pred] == &Chain || !LoopBlockSet.count(Pred))
1019 ++Chain.LoopPredecessors;
1023 if (Chain.LoopPredecessors == 0)
1024 BlockWorkList.push_back(*Chain.begin());
1027 buildChain(LoopTop, LoopChain, BlockWorkList, &LoopBlockSet);
1029 if (RotateLoopWithProfile)
1030 rotateLoopWithProfile(LoopChain, L, LoopBlockSet);
1032 rotateLoop(LoopChain, ExitingBB, LoopBlockSet);
1035 // Crash at the end so we get all of the debugging output first.
1036 bool BadLoop = false;
1037 if (LoopChain.LoopPredecessors) {
1039 dbgs() << "Loop chain contains a block without its preds placed!\n"
1040 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
1041 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n";
1043 for (MachineBasicBlock *ChainBB : LoopChain) {
1044 dbgs() << " ... " << getBlockName(ChainBB) << "\n";
1045 if (!LoopBlockSet.erase(ChainBB)) {
1046 // We don't mark the loop as bad here because there are real situations
1047 // where this can occur. For example, with an unanalyzable fallthrough
1048 // from a loop block to a non-loop block or vice versa.
1049 dbgs() << "Loop chain contains a block not contained by the loop!\n"
1050 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
1051 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
1052 << " Bad block: " << getBlockName(ChainBB) << "\n";
1056 if (!LoopBlockSet.empty()) {
1058 for (MachineBasicBlock *LoopBB : LoopBlockSet)
1059 dbgs() << "Loop contains blocks never placed into a chain!\n"
1060 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
1061 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
1062 << " Bad block: " << getBlockName(LoopBB) << "\n";
1064 assert(!BadLoop && "Detected problems with the placement of this loop.");
1068 void MachineBlockPlacement::buildCFGChains(MachineFunction &F) {
1069 // Ensure that every BB in the function has an associated chain to simplify
1070 // the assumptions of the remaining algorithm.
1071 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
1072 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
1073 MachineBasicBlock *BB = &*FI;
1075 new (ChainAllocator.Allocate()) BlockChain(BlockToChain, BB);
1076 // Also, merge any blocks which we cannot reason about and must preserve
1077 // the exact fallthrough behavior for.
1080 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
1081 if (!TII->AnalyzeBranch(*BB, TBB, FBB, Cond) || !FI->canFallThrough())
1084 MachineFunction::iterator NextFI = std::next(FI);
1085 MachineBasicBlock *NextBB = &*NextFI;
1086 // Ensure that the layout successor is a viable block, as we know that
1087 // fallthrough is a possibility.
1088 assert(NextFI != FE && "Can't fallthrough past the last block.");
1089 DEBUG(dbgs() << "Pre-merging due to unanalyzable fallthrough: "
1090 << getBlockName(BB) << " -> " << getBlockName(NextBB)
1092 Chain->merge(NextBB, nullptr);
1098 if (OutlineOptionalBranches) {
1099 // Find the nearest common dominator of all of F's terminators.
1100 MachineBasicBlock *Terminator = nullptr;
1101 for (MachineBasicBlock &MBB : F) {
1102 if (MBB.succ_size() == 0) {
1103 if (Terminator == nullptr)
1106 Terminator = MDT->findNearestCommonDominator(Terminator, &MBB);
1110 // MBBs dominating this common dominator are unavoidable.
1111 UnavoidableBlocks.clear();
1112 for (MachineBasicBlock &MBB : F) {
1113 if (MDT->dominates(&MBB, Terminator)) {
1114 UnavoidableBlocks.insert(&MBB);
1119 // Build any loop-based chains.
1120 for (MachineLoop *L : *MLI)
1121 buildLoopChains(F, *L);
1123 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
1125 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
1126 for (MachineBasicBlock &MBB : F) {
1127 BlockChain &Chain = *BlockToChain[&MBB];
1128 if (!UpdatedPreds.insert(&Chain).second)
1131 assert(Chain.LoopPredecessors == 0);
1132 for (MachineBasicBlock *ChainBB : Chain) {
1133 assert(BlockToChain[ChainBB] == &Chain);
1134 for (MachineBasicBlock *Pred : ChainBB->predecessors()) {
1135 if (BlockToChain[Pred] == &Chain)
1137 ++Chain.LoopPredecessors;
1141 if (Chain.LoopPredecessors == 0)
1142 BlockWorkList.push_back(*Chain.begin());
1145 BlockChain &FunctionChain = *BlockToChain[&F.front()];
1146 buildChain(&F.front(), FunctionChain, BlockWorkList);
1149 typedef SmallPtrSet<MachineBasicBlock *, 16> FunctionBlockSetType;
1152 // Crash at the end so we get all of the debugging output first.
1153 bool BadFunc = false;
1154 FunctionBlockSetType FunctionBlockSet;
1155 for (MachineBasicBlock &MBB : F)
1156 FunctionBlockSet.insert(&MBB);
1158 for (MachineBasicBlock *ChainBB : FunctionChain)
1159 if (!FunctionBlockSet.erase(ChainBB)) {
1161 dbgs() << "Function chain contains a block not in the function!\n"
1162 << " Bad block: " << getBlockName(ChainBB) << "\n";
1165 if (!FunctionBlockSet.empty()) {
1167 for (MachineBasicBlock *RemainingBB : FunctionBlockSet)
1168 dbgs() << "Function contains blocks never placed into a chain!\n"
1169 << " Bad block: " << getBlockName(RemainingBB) << "\n";
1171 assert(!BadFunc && "Detected problems with the block placement.");
1174 // Splice the blocks into place.
1175 MachineFunction::iterator InsertPos = F.begin();
1176 for (MachineBasicBlock *ChainBB : FunctionChain) {
1177 DEBUG(dbgs() << (ChainBB == *FunctionChain.begin() ? "Placing chain "
1179 << getBlockName(ChainBB) << "\n");
1180 if (InsertPos != MachineFunction::iterator(ChainBB))
1181 F.splice(InsertPos, ChainBB);
1185 // Update the terminator of the previous block.
1186 if (ChainBB == *FunctionChain.begin())
1188 MachineBasicBlock *PrevBB = &*std::prev(MachineFunction::iterator(ChainBB));
1190 // FIXME: It would be awesome of updateTerminator would just return rather
1191 // than assert when the branch cannot be analyzed in order to remove this
1194 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
1195 if (!TII->AnalyzeBranch(*PrevBB, TBB, FBB, Cond)) {
1196 // The "PrevBB" is not yet updated to reflect current code layout, so,
1197 // o. it may fall-through to a block without explict "goto" instruction
1198 // before layout, and no longer fall-through it after layout; or
1199 // o. just opposite.
1201 // AnalyzeBranch() may return erroneous value for FBB when these two
1202 // situations take place. For the first scenario FBB is mistakenly set
1203 // NULL; for the 2nd scenario, the FBB, which is expected to be NULL,
1204 // is mistakenly pointing to "*BI".
1206 bool needUpdateBr = true;
1207 if (!Cond.empty() && (!FBB || FBB == ChainBB)) {
1208 PrevBB->updateTerminator();
1209 needUpdateBr = false;
1211 TBB = FBB = nullptr;
1212 if (TII->AnalyzeBranch(*PrevBB, TBB, FBB, Cond)) {
1213 // FIXME: This should never take place.
1214 TBB = FBB = nullptr;
1218 // If PrevBB has a two-way branch, try to re-order the branches
1219 // such that we branch to the successor with higher weight first.
1220 if (TBB && !Cond.empty() && FBB &&
1221 MBPI->getEdgeWeight(PrevBB, FBB) > MBPI->getEdgeWeight(PrevBB, TBB) &&
1222 !TII->ReverseBranchCondition(Cond)) {
1223 DEBUG(dbgs() << "Reverse order of the two branches: "
1224 << getBlockName(PrevBB) << "\n");
1225 DEBUG(dbgs() << " Edge weight: " << MBPI->getEdgeWeight(PrevBB, FBB)
1226 << " vs " << MBPI->getEdgeWeight(PrevBB, TBB) << "\n");
1227 DebugLoc dl; // FIXME: this is nowhere
1228 TII->RemoveBranch(*PrevBB);
1229 TII->InsertBranch(*PrevBB, FBB, TBB, Cond, dl);
1230 needUpdateBr = true;
1233 PrevBB->updateTerminator();
1237 // Fixup the last block.
1239 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
1240 if (!TII->AnalyzeBranch(F.back(), TBB, FBB, Cond))
1241 F.back().updateTerminator();
1243 // Walk through the backedges of the function now that we have fully laid out
1244 // the basic blocks and align the destination of each backedge. We don't rely
1245 // exclusively on the loop info here so that we can align backedges in
1246 // unnatural CFGs and backedges that were introduced purely because of the
1247 // loop rotations done during this layout pass.
1248 // FIXME: Use Function::optForSize().
1249 if (F.getFunction()->hasFnAttribute(Attribute::OptimizeForSize))
1251 if (FunctionChain.begin() == FunctionChain.end())
1252 return; // Empty chain.
1254 const BranchProbability ColdProb(1, 5); // 20%
1255 BlockFrequency EntryFreq = MBFI->getBlockFreq(&F.front());
1256 BlockFrequency WeightedEntryFreq = EntryFreq * ColdProb;
1257 for (MachineBasicBlock *ChainBB : FunctionChain) {
1258 if (ChainBB == *FunctionChain.begin())
1261 // Don't align non-looping basic blocks. These are unlikely to execute
1262 // enough times to matter in practice. Note that we'll still handle
1263 // unnatural CFGs inside of a natural outer loop (the common case) and
1265 MachineLoop *L = MLI->getLoopFor(ChainBB);
1269 unsigned Align = TLI->getPrefLoopAlignment(L);
1271 continue; // Don't care about loop alignment.
1273 // If the block is cold relative to the function entry don't waste space
1275 BlockFrequency Freq = MBFI->getBlockFreq(ChainBB);
1276 if (Freq < WeightedEntryFreq)
1279 // If the block is cold relative to its loop header, don't align it
1280 // regardless of what edges into the block exist.
1281 MachineBasicBlock *LoopHeader = L->getHeader();
1282 BlockFrequency LoopHeaderFreq = MBFI->getBlockFreq(LoopHeader);
1283 if (Freq < (LoopHeaderFreq * ColdProb))
1286 // Check for the existence of a non-layout predecessor which would benefit
1287 // from aligning this block.
1288 MachineBasicBlock *LayoutPred =
1289 &*std::prev(MachineFunction::iterator(ChainBB));
1291 // Force alignment if all the predecessors are jumps. We already checked
1292 // that the block isn't cold above.
1293 if (!LayoutPred->isSuccessor(ChainBB)) {
1294 ChainBB->setAlignment(Align);
1298 // Align this block if the layout predecessor's edge into this block is
1299 // cold relative to the block. When this is true, other predecessors make up
1300 // all of the hot entries into the block and thus alignment is likely to be
1302 BranchProbability LayoutProb =
1303 MBPI->getEdgeProbability(LayoutPred, ChainBB);
1304 BlockFrequency LayoutEdgeFreq = MBFI->getBlockFreq(LayoutPred) * LayoutProb;
1305 if (LayoutEdgeFreq <= (Freq * ColdProb))
1306 ChainBB->setAlignment(Align);
1310 bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &F) {
1311 // Check for single-block functions and skip them.
1312 if (std::next(F.begin()) == F.end())
1315 if (skipOptnoneFunction(*F.getFunction()))
1318 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
1319 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
1320 MLI = &getAnalysis<MachineLoopInfo>();
1321 TII = F.getSubtarget().getInstrInfo();
1322 TLI = F.getSubtarget().getTargetLowering();
1323 MDT = &getAnalysis<MachineDominatorTree>();
1324 assert(BlockToChain.empty());
1328 BlockToChain.clear();
1329 ChainAllocator.DestroyAll();
1332 // Align all of the blocks in the function to a specific alignment.
1333 for (MachineBasicBlock &MBB : F)
1334 MBB.setAlignment(AlignAllBlock);
1336 // We always return true as we have no way to track whether the final order
1337 // differs from the original order.
1342 /// \brief A pass to compute block placement statistics.
1344 /// A separate pass to compute interesting statistics for evaluating block
1345 /// placement. This is separate from the actual placement pass so that they can
1346 /// be computed in the absence of any placement transformations or when using
1347 /// alternative placement strategies.
1348 class MachineBlockPlacementStats : public MachineFunctionPass {
1349 /// \brief A handle to the branch probability pass.
1350 const MachineBranchProbabilityInfo *MBPI;
1352 /// \brief A handle to the function-wide block frequency pass.
1353 const MachineBlockFrequencyInfo *MBFI;
1356 static char ID; // Pass identification, replacement for typeid
1357 MachineBlockPlacementStats() : MachineFunctionPass(ID) {
1358 initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry());
1361 bool runOnMachineFunction(MachineFunction &F) override;
1363 void getAnalysisUsage(AnalysisUsage &AU) const override {
1364 AU.addRequired<MachineBranchProbabilityInfo>();
1365 AU.addRequired<MachineBlockFrequencyInfo>();
1366 AU.setPreservesAll();
1367 MachineFunctionPass::getAnalysisUsage(AU);
1372 char MachineBlockPlacementStats::ID = 0;
1373 char &llvm::MachineBlockPlacementStatsID = MachineBlockPlacementStats::ID;
1374 INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats",
1375 "Basic Block Placement Stats", false, false)
1376 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
1377 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
1378 INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats",
1379 "Basic Block Placement Stats", false, false)
1381 bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) {
1382 // Check for single-block functions and skip them.
1383 if (std::next(F.begin()) == F.end())
1386 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
1387 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
1389 for (MachineBasicBlock &MBB : F) {
1390 BlockFrequency BlockFreq = MBFI->getBlockFreq(&MBB);
1391 Statistic &NumBranches =
1392 (MBB.succ_size() > 1) ? NumCondBranches : NumUncondBranches;
1393 Statistic &BranchTakenFreq =
1394 (MBB.succ_size() > 1) ? CondBranchTakenFreq : UncondBranchTakenFreq;
1395 for (MachineBasicBlock *Succ : MBB.successors()) {
1396 // Skip if this successor is a fallthrough.
1397 if (MBB.isLayoutSuccessor(Succ))
1400 BlockFrequency EdgeFreq =
1401 BlockFreq * MBPI->getEdgeProbability(&MBB, Succ);
1403 BranchTakenFreq += EdgeFreq.getFrequency();