1 //===-- MachineBlockPlacement.cpp - Basic Block Code Layout optimization --===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements basic block placement transformations using the CFG
11 // structure and branch probability estimates.
13 // The pass strives to preserve the structure of the CFG (that is, retain
14 // a topological ordering of basic blocks) in the absense of a *strong* signal
15 // to the contrary from probabilities. However, within the CFG structure, it
16 // attempts to choose an ordering which favors placing more likely sequences of
17 // blocks adjacent to each other.
19 // The algorithm works from the inner-most loop within a function outward, and
20 // at each stage walks through the basic blocks, trying to coalesce them into
21 // sequential chains where allowed by the CFG (or demanded by heavy
22 // probabilities). Finally, it walks the blocks in topological order, and the
23 // first time it reaches a chain of basic blocks, it schedules them in the
26 //===----------------------------------------------------------------------===//
28 #define DEBUG_TYPE "block-placement2"
29 #include "llvm/CodeGen/MachineBasicBlock.h"
30 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
31 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineFunctionPass.h"
34 #include "llvm/CodeGen/MachineLoopInfo.h"
35 #include "llvm/CodeGen/MachineModuleInfo.h"
36 #include "llvm/CodeGen/Passes.h"
37 #include "llvm/Support/Allocator.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/ErrorHandling.h"
40 #include "llvm/ADT/DenseMap.h"
41 #include "llvm/ADT/PostOrderIterator.h"
42 #include "llvm/ADT/SCCIterator.h"
43 #include "llvm/ADT/SmallPtrSet.h"
44 #include "llvm/ADT/SmallVector.h"
45 #include "llvm/ADT/Statistic.h"
46 #include "llvm/Target/TargetInstrInfo.h"
47 #include "llvm/Target/TargetLowering.h"
51 STATISTIC(NumCondBranches, "Number of conditional branches");
52 STATISTIC(NumUncondBranches, "Number of uncondittional branches");
53 STATISTIC(CondBranchTakenFreq,
54 "Potential frequency of taking conditional branches");
55 STATISTIC(UncondBranchTakenFreq,
56 "Potential frequency of taking unconditional branches");
59 /// \brief A structure for storing a weighted edge.
61 /// This stores an edge and its weight, computed as the product of the
62 /// frequency that the starting block is entered with the probability of
63 /// a particular exit block.
65 BlockFrequency EdgeFrequency;
66 MachineBasicBlock *From, *To;
68 bool operator<(const WeightedEdge &RHS) const {
69 return EdgeFrequency < RHS.EdgeFrequency;
76 /// \brief Type for our function-wide basic block -> block chain mapping.
77 typedef DenseMap<MachineBasicBlock *, BlockChain *> BlockToChainMapType;
81 /// \brief A chain of blocks which will be laid out contiguously.
83 /// This is the datastructure representing a chain of consecutive blocks that
84 /// are profitable to layout together in order to maximize fallthrough
85 /// probabilities. We also can use a block chain to represent a sequence of
86 /// basic blocks which have some external (correctness) requirement for
87 /// sequential layout.
89 /// Eventually, the block chains will form a directed graph over the function.
90 /// We provide an SCC-supporting-iterator in order to quicky build and walk the
91 /// SCCs of block chains within a function.
93 /// The block chains also have support for calculating and caching probability
94 /// information related to the chain itself versus other chains. This is used
95 /// for ranking during the final layout of block chains.
97 /// \brief The sequence of blocks belonging to this chain.
99 /// This is the sequence of blocks for a particular chain. These will be laid
100 /// out in-order within the function.
101 SmallVector<MachineBasicBlock *, 4> Blocks;
103 /// \brief A handle to the function-wide basic block to block chain mapping.
105 /// This is retained in each block chain to simplify the computation of child
106 /// block chains for SCC-formation and iteration. We store the edges to child
107 /// basic blocks, and map them back to their associated chains using this
109 BlockToChainMapType &BlockToChain;
112 /// \brief Construct a new BlockChain.
114 /// This builds a new block chain representing a single basic block in the
115 /// function. It also registers itself as the chain that block participates
116 /// in with the BlockToChain mapping.
117 BlockChain(BlockToChainMapType &BlockToChain, MachineBasicBlock *BB)
118 : Blocks(1, BB), BlockToChain(BlockToChain), LoopPredecessors(0) {
119 assert(BB && "Cannot create a chain with a null basic block");
120 BlockToChain[BB] = this;
123 /// \brief Iterator over blocks within the chain.
124 typedef SmallVectorImpl<MachineBasicBlock *>::const_iterator iterator;
126 /// \brief Beginning of blocks within the chain.
127 iterator begin() const { return Blocks.begin(); }
129 /// \brief End of blocks within the chain.
130 iterator end() const { return Blocks.end(); }
132 /// \brief Merge a block chain into this one.
134 /// This routine merges a block chain into this one. It takes care of forming
135 /// a contiguous sequence of basic blocks, updating the edge list, and
136 /// updating the block -> chain mapping. It does not free or tear down the
137 /// old chain, but the old chain's block list is no longer valid.
138 void merge(MachineBasicBlock *BB, BlockChain *Chain) {
140 assert(!Blocks.empty());
142 // Fast path in case we don't have a chain already.
144 assert(!BlockToChain[BB]);
145 Blocks.push_back(BB);
146 BlockToChain[BB] = this;
150 assert(BB == *Chain->begin());
151 assert(Chain->begin() != Chain->end());
153 // Update the incoming blocks to point to this chain, and add them to the
155 for (BlockChain::iterator BI = Chain->begin(), BE = Chain->end();
157 Blocks.push_back(*BI);
158 assert(BlockToChain[*BI] == Chain && "Incoming blocks not in chain");
159 BlockToChain[*BI] = this;
163 /// \brief Count of predecessors within the loop currently being processed.
165 /// This count is updated at each loop we process to represent the number of
166 /// in-loop predecessors of this chain.
167 unsigned LoopPredecessors;
172 class MachineBlockPlacement : public MachineFunctionPass {
173 /// \brief A typedef for a block filter set.
174 typedef SmallPtrSet<MachineBasicBlock *, 16> BlockFilterSet;
176 /// \brief A handle to the branch probability pass.
177 const MachineBranchProbabilityInfo *MBPI;
179 /// \brief A handle to the function-wide block frequency pass.
180 const MachineBlockFrequencyInfo *MBFI;
182 /// \brief A handle to the loop info.
183 const MachineLoopInfo *MLI;
185 /// \brief A handle to the target's instruction info.
186 const TargetInstrInfo *TII;
188 /// \brief A handle to the target's lowering info.
189 const TargetLowering *TLI;
191 /// \brief Allocator and owner of BlockChain structures.
193 /// We build BlockChains lazily by merging together high probability BB
194 /// sequences acording to the "Algo2" in the paper mentioned at the top of
195 /// the file. To reduce malloc traffic, we allocate them using this slab-like
196 /// allocator, and destroy them after the pass completes.
197 SpecificBumpPtrAllocator<BlockChain> ChainAllocator;
199 /// \brief Function wide BasicBlock to BlockChain mapping.
201 /// This mapping allows efficiently moving from any given basic block to the
202 /// BlockChain it participates in, if any. We use it to, among other things,
203 /// allow implicitly defining edges between chains as the existing edges
204 /// between basic blocks.
205 DenseMap<MachineBasicBlock *, BlockChain *> BlockToChain;
207 void markChainSuccessors(BlockChain &Chain,
208 MachineBasicBlock *LoopHeaderBB,
209 SmallVectorImpl<MachineBasicBlock *> &Blocks,
210 const BlockFilterSet *BlockFilter = 0);
211 MachineBasicBlock *selectBestSuccessor(MachineBasicBlock *BB,
213 const BlockFilterSet *BlockFilter);
214 MachineBasicBlock *selectBestCandidateBlock(
215 BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList,
216 const BlockFilterSet *BlockFilter);
217 void buildChain(MachineBasicBlock *BB, BlockChain &Chain,
218 SmallVectorImpl<MachineBasicBlock *> &Blocks,
219 const BlockFilterSet *BlockFilter = 0);
220 void buildLoopChains(MachineFunction &F, MachineLoop &L);
221 void buildCFGChains(MachineFunction &F);
222 void AlignLoops(MachineFunction &F);
225 static char ID; // Pass identification, replacement for typeid
226 MachineBlockPlacement() : MachineFunctionPass(ID) {
227 initializeMachineBlockPlacementPass(*PassRegistry::getPassRegistry());
230 bool runOnMachineFunction(MachineFunction &F);
232 void getAnalysisUsage(AnalysisUsage &AU) const {
233 AU.addRequired<MachineBranchProbabilityInfo>();
234 AU.addRequired<MachineBlockFrequencyInfo>();
235 AU.addRequired<MachineLoopInfo>();
236 MachineFunctionPass::getAnalysisUsage(AU);
239 const char *getPassName() const { return "Block Placement"; }
243 char MachineBlockPlacement::ID = 0;
244 INITIALIZE_PASS_BEGIN(MachineBlockPlacement, "block-placement2",
245 "Branch Probability Basic Block Placement", false, false)
246 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
247 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
248 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
249 INITIALIZE_PASS_END(MachineBlockPlacement, "block-placement2",
250 "Branch Probability Basic Block Placement", false, false)
252 FunctionPass *llvm::createMachineBlockPlacementPass() {
253 return new MachineBlockPlacement();
257 /// \brief Helper to print the name of a MBB.
259 /// Only used by debug logging.
260 static std::string getBlockName(MachineBasicBlock *BB) {
262 raw_string_ostream OS(Result);
263 OS << "BB#" << BB->getNumber()
264 << " (derived from LLVM BB '" << BB->getName() << "')";
269 /// \brief Helper to print the number of a MBB.
271 /// Only used by debug logging.
272 static std::string getBlockNum(MachineBasicBlock *BB) {
274 raw_string_ostream OS(Result);
275 OS << "BB#" << BB->getNumber();
281 /// \brief Mark a chain's successors as having one fewer preds.
283 /// When a chain is being merged into the "placed" chain, this routine will
284 /// quickly walk the successors of each block in the chain and mark them as
285 /// having one fewer active predecessor. It also adds any successors of this
286 /// chain which reach the zero-predecessor state to the worklist passed in.
287 void MachineBlockPlacement::markChainSuccessors(
289 MachineBasicBlock *LoopHeaderBB,
290 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
291 const BlockFilterSet *BlockFilter) {
292 // Walk all the blocks in this chain, marking their successors as having
293 // a predecessor placed.
294 for (BlockChain::iterator CBI = Chain.begin(), CBE = Chain.end();
296 // Add any successors for which this is the only un-placed in-loop
297 // predecessor to the worklist as a viable candidate for CFG-neutral
298 // placement. No subsequent placement of this block will violate the CFG
299 // shape, so we get to use heuristics to choose a favorable placement.
300 for (MachineBasicBlock::succ_iterator SI = (*CBI)->succ_begin(),
301 SE = (*CBI)->succ_end();
303 if (BlockFilter && !BlockFilter->count(*SI))
305 BlockChain &SuccChain = *BlockToChain[*SI];
306 // Disregard edges within a fixed chain, or edges to the loop header.
307 if (&Chain == &SuccChain || *SI == LoopHeaderBB)
310 // This is a cross-chain edge that is within the loop, so decrement the
311 // loop predecessor count of the destination chain.
312 if (SuccChain.LoopPredecessors > 0 && --SuccChain.LoopPredecessors == 0)
313 BlockWorkList.push_back(*SI);
318 /// \brief Select the best successor for a block.
320 /// This looks across all successors of a particular block and attempts to
321 /// select the "best" one to be the layout successor. It only considers direct
322 /// successors which also pass the block filter. It will attempt to avoid
323 /// breaking CFG structure, but cave and break such structures in the case of
324 /// very hot successor edges.
326 /// \returns The best successor block found, or null if none are viable.
327 MachineBasicBlock *MachineBlockPlacement::selectBestSuccessor(
328 MachineBasicBlock *BB, BlockChain &Chain,
329 const BlockFilterSet *BlockFilter) {
330 const BranchProbability HotProb(4, 5); // 80%
332 MachineBasicBlock *BestSucc = 0;
333 BranchProbability BestProb = BranchProbability::getZero();
334 DEBUG(dbgs() << "Attempting merge from: " << getBlockName(BB) << "\n");
335 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
338 if (BlockFilter && !BlockFilter->count(*SI))
340 BlockChain &SuccChain = *BlockToChain[*SI];
341 if (&SuccChain == &Chain) {
342 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> Already merged!\n");
346 BranchProbability SuccProb = MBPI->getEdgeProbability(BB, *SI);
348 // Only consider successors which are either "hot", or wouldn't violate
349 // any CFG constraints.
350 if (SuccChain.LoopPredecessors != 0 && SuccProb < HotProb) {
351 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> CFG conflict\n");
355 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> " << SuccProb
357 << (SuccChain.LoopPredecessors != 0 ? " (CFG break)" : "")
359 if (BestSucc && BestProb >= SuccProb)
367 /// \brief Select the best block from a worklist.
369 /// This looks through the provided worklist as a list of candidate basic
370 /// blocks and select the most profitable one to place. The definition of
371 /// profitable only really makes sense in the context of a loop. This returns
372 /// the most frequently visited block in the worklist, which in the case of
373 /// a loop, is the one most desirable to be physically close to the rest of the
374 /// loop body in order to improve icache behavior.
376 /// \returns The best block found, or null if none are viable.
377 MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock(
378 BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList,
379 const BlockFilterSet *BlockFilter) {
380 MachineBasicBlock *BestBlock = 0;
381 BlockFrequency BestFreq;
382 for (SmallVectorImpl<MachineBasicBlock *>::iterator WBI = WorkList.begin(),
383 WBE = WorkList.end();
385 if (BlockFilter && !BlockFilter->count(*WBI))
387 BlockChain &SuccChain = *BlockToChain[*WBI];
388 if (&SuccChain == &Chain) {
389 DEBUG(dbgs() << " " << getBlockName(*WBI)
390 << " -> Already merged!\n");
393 assert(SuccChain.LoopPredecessors == 0 && "Found CFG-violating block");
395 BlockFrequency CandidateFreq = MBFI->getBlockFreq(*WBI);
396 DEBUG(dbgs() << " " << getBlockName(*WBI) << " -> " << CandidateFreq
398 if (BestBlock && BestFreq >= CandidateFreq)
401 BestFreq = CandidateFreq;
406 void MachineBlockPlacement::buildChain(
407 MachineBasicBlock *BB,
409 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
410 const BlockFilterSet *BlockFilter) {
412 assert(BlockToChain[BB] == &Chain);
413 assert(*Chain.begin() == BB);
414 MachineBasicBlock *LoopHeaderBB = BB;
415 markChainSuccessors(Chain, LoopHeaderBB, BlockWorkList, BlockFilter);
416 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
417 BB = *llvm::prior(Chain.end());
420 assert(BlockToChain[BB] == &Chain);
421 assert(*llvm::prior(Chain.end()) == BB);
422 MachineBasicBlock *BestSucc = 0;
424 // Check for unreasonable branches, and forcibly merge the existing layout
425 // successor for them. We can handle cases that AnalyzeBranch can't: jump
426 // tables etc are fine. The case we want to handle specially is when there
427 // is potential fallthrough, but the branch cannot be analyzed. This
428 // includes blocks without terminators as well as other cases.
430 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch.
431 if (TII->AnalyzeBranch(*BB, TBB, FBB, Cond) && BB->canFallThrough()) {
432 MachineFunction::iterator I(BB);
433 assert(llvm::next(I) != BB->getParent()->end() &&
434 "The final block in the function can fallthrough!");
435 BestSucc = llvm::next(I);
438 // Otherwise, look for the best viable successor if there is one to place
439 // immediately after this block.
441 BestSucc = selectBestSuccessor(BB, Chain, BlockFilter);
443 // If an immediate successor isn't available, look for the best viable
444 // block among those we've identified as not violating the loop's CFG at
445 // this point. This won't be a fallthrough, but it will increase locality.
447 BestSucc = selectBestCandidateBlock(Chain, BlockWorkList, BlockFilter);
450 DEBUG(dbgs() << "Finished forming chain for header block "
451 << getBlockNum(*Chain.begin()) << "\n");
455 // Place this block, updating the datastructures to reflect its placement.
456 BlockChain &SuccChain = *BlockToChain[BestSucc];
457 DEBUG(dbgs() << "Merging from " << getBlockNum(BB)
458 << " to " << getBlockNum(BestSucc) << "\n");
459 markChainSuccessors(SuccChain, LoopHeaderBB, BlockWorkList, BlockFilter);
460 Chain.merge(BestSucc, &SuccChain);
461 BB = *llvm::prior(Chain.end());
465 /// \brief Forms basic block chains from the natural loop structures.
467 /// These chains are designed to preserve the existing *structure* of the code
468 /// as much as possible. We can then stitch the chains together in a way which
469 /// both preserves the topological structure and minimizes taken conditional
471 void MachineBlockPlacement::buildLoopChains(MachineFunction &F,
473 // First recurse through any nested loops, building chains for those inner
475 for (MachineLoop::iterator LI = L.begin(), LE = L.end(); LI != LE; ++LI)
476 buildLoopChains(F, **LI);
478 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
479 BlockFilterSet LoopBlockSet(L.block_begin(), L.block_end());
481 // FIXME: This is a really lame way of walking the chains in the loop: we
482 // walk the blocks, and use a set to prevent visiting a particular chain
484 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
485 for (MachineLoop::block_iterator BI = L.block_begin(),
488 BlockChain &Chain = *BlockToChain[*BI];
489 if (!UpdatedPreds.insert(&Chain) || BI == L.block_begin())
492 assert(Chain.LoopPredecessors == 0);
493 for (BlockChain::iterator BCI = Chain.begin(), BCE = Chain.end();
495 assert(BlockToChain[*BCI] == &Chain);
496 for (MachineBasicBlock::pred_iterator PI = (*BCI)->pred_begin(),
497 PE = (*BCI)->pred_end();
499 if (BlockToChain[*PI] == &Chain || !LoopBlockSet.count(*PI))
501 ++Chain.LoopPredecessors;
505 if (Chain.LoopPredecessors == 0)
506 BlockWorkList.push_back(*BI);
509 BlockChain &LoopChain = *BlockToChain[L.getHeader()];
510 buildChain(*L.block_begin(), LoopChain, BlockWorkList, &LoopBlockSet);
513 if (LoopChain.LoopPredecessors)
514 dbgs() << "Loop chain contains a block without its preds placed!\n"
515 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
516 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n";
517 for (BlockChain::iterator BCI = LoopChain.begin(), BCE = LoopChain.end();
519 if (!LoopBlockSet.erase(*BCI))
520 dbgs() << "Loop chain contains a block not contained by the loop!\n"
521 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
522 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
523 << " Bad block: " << getBlockName(*BCI) << "\n";
525 if (!LoopBlockSet.empty())
526 for (SmallPtrSet<MachineBasicBlock *, 16>::iterator LBI = LoopBlockSet.begin(), LBE = LoopBlockSet.end();
528 dbgs() << "Loop contains blocks never placed into a chain!\n"
529 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
530 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
531 << " Bad block: " << getBlockName(*LBI) << "\n";
535 void MachineBlockPlacement::buildCFGChains(MachineFunction &F) {
536 // Ensure that every BB in the function has an associated chain to simplify
537 // the assumptions of the remaining algorithm.
538 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
540 new (ChainAllocator.Allocate()) BlockChain(BlockToChain, &*FI);
542 // Build any loop-based chains.
543 for (MachineLoopInfo::iterator LI = MLI->begin(), LE = MLI->end(); LI != LE;
545 buildLoopChains(F, **LI);
547 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
549 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
550 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
551 MachineBasicBlock *BB = &*FI;
552 BlockChain &Chain = *BlockToChain[BB];
553 if (!UpdatedPreds.insert(&Chain))
556 assert(Chain.LoopPredecessors == 0);
557 for (BlockChain::iterator BCI = Chain.begin(), BCE = Chain.end();
559 assert(BlockToChain[*BCI] == &Chain);
560 for (MachineBasicBlock::pred_iterator PI = (*BCI)->pred_begin(),
561 PE = (*BCI)->pred_end();
563 if (BlockToChain[*PI] == &Chain)
565 ++Chain.LoopPredecessors;
569 if (Chain.LoopPredecessors == 0)
570 BlockWorkList.push_back(BB);
573 BlockChain &FunctionChain = *BlockToChain[&F.front()];
574 buildChain(&F.front(), FunctionChain, BlockWorkList);
576 typedef SmallPtrSet<MachineBasicBlock *, 16> FunctionBlockSetType;
578 FunctionBlockSetType FunctionBlockSet;
579 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
580 FunctionBlockSet.insert(FI);
582 for (BlockChain::iterator BCI = FunctionChain.begin(), BCE = FunctionChain.end();
584 if (!FunctionBlockSet.erase(*BCI))
585 dbgs() << "Function chain contains a block not in the function!\n"
586 << " Bad block: " << getBlockName(*BCI) << "\n";
588 if (!FunctionBlockSet.empty())
589 for (SmallPtrSet<MachineBasicBlock *, 16>::iterator FBI = FunctionBlockSet.begin(),
590 FBE = FunctionBlockSet.end(); FBI != FBE; ++FBI)
591 dbgs() << "Function contains blocks never placed into a chain!\n"
592 << " Bad block: " << getBlockName(*FBI) << "\n";
595 // Splice the blocks into place.
596 MachineFunction::iterator InsertPos = F.begin();
597 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
598 for (BlockChain::iterator BI = FunctionChain.begin(), BE = FunctionChain.end();
600 DEBUG(dbgs() << (BI == FunctionChain.begin() ? "Placing chain "
602 << getBlockName(*BI) << "\n");
603 if (InsertPos != MachineFunction::iterator(*BI))
604 F.splice(InsertPos, *BI);
608 // Update the terminator of the previous block.
609 if (BI == FunctionChain.begin())
611 MachineBasicBlock *PrevBB = llvm::prior(MachineFunction::iterator(*BI));
613 // FIXME: It would be awesome of updateTerminator would just return rather
614 // than assert when the branch cannot be analyzed in order to remove this
617 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch.
618 if (!TII->AnalyzeBranch(*PrevBB, TBB, FBB, Cond))
619 PrevBB->updateTerminator();
622 // Fixup the last block.
624 MachineBasicBlock *TBB = 0, *FBB = 0; // For AnalyzeBranch.
625 if (!TII->AnalyzeBranch(F.back(), TBB, FBB, Cond))
626 F.back().updateTerminator();
629 /// \brief Recursive helper to align a loop and any nested loops.
630 static void AlignLoop(MachineFunction &F, MachineLoop *L, unsigned Align) {
631 // Recurse through nested loops.
632 for (MachineLoop::iterator I = L->begin(), E = L->end(); I != E; ++I)
633 AlignLoop(F, *I, Align);
635 L->getTopBlock()->setAlignment(Align);
638 /// \brief Align loop headers to target preferred alignments.
639 void MachineBlockPlacement::AlignLoops(MachineFunction &F) {
640 if (F.getFunction()->hasFnAttr(Attribute::OptimizeForSize))
643 unsigned Align = TLI->getPrefLoopAlignment();
645 return; // Don't care about loop alignment.
647 for (MachineLoopInfo::iterator I = MLI->begin(), E = MLI->end(); I != E; ++I)
648 AlignLoop(F, *I, Align);
651 bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &F) {
652 // Check for single-block functions and skip them.
653 if (llvm::next(F.begin()) == F.end())
656 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
657 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
658 MLI = &getAnalysis<MachineLoopInfo>();
659 TII = F.getTarget().getInstrInfo();
660 TLI = F.getTarget().getTargetLowering();
661 assert(BlockToChain.empty());
666 BlockToChain.clear();
668 // We always return true as we have no way to track whether the final order
669 // differs from the original order.
674 /// \brief A pass to compute block placement statistics.
676 /// A separate pass to compute interesting statistics for evaluating block
677 /// placement. This is separate from the actual placement pass so that they can
678 /// be computed in the absense of any placement transformations or when using
679 /// alternative placement strategies.
680 class MachineBlockPlacementStats : public MachineFunctionPass {
681 /// \brief A handle to the branch probability pass.
682 const MachineBranchProbabilityInfo *MBPI;
684 /// \brief A handle to the function-wide block frequency pass.
685 const MachineBlockFrequencyInfo *MBFI;
688 static char ID; // Pass identification, replacement for typeid
689 MachineBlockPlacementStats() : MachineFunctionPass(ID) {
690 initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry());
693 bool runOnMachineFunction(MachineFunction &F);
695 void getAnalysisUsage(AnalysisUsage &AU) const {
696 AU.addRequired<MachineBranchProbabilityInfo>();
697 AU.addRequired<MachineBlockFrequencyInfo>();
698 AU.setPreservesAll();
699 MachineFunctionPass::getAnalysisUsage(AU);
702 const char *getPassName() const { return "Block Placement Stats"; }
706 char MachineBlockPlacementStats::ID = 0;
707 INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats",
708 "Basic Block Placement Stats", false, false)
709 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
710 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
711 INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats",
712 "Basic Block Placement Stats", false, false)
714 FunctionPass *llvm::createMachineBlockPlacementStatsPass() {
715 return new MachineBlockPlacementStats();
718 bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) {
719 // Check for single-block functions and skip them.
720 if (llvm::next(F.begin()) == F.end())
723 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
724 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
726 for (MachineFunction::iterator I = F.begin(), E = F.end(); I != E; ++I) {
727 BlockFrequency BlockFreq = MBFI->getBlockFreq(I);
728 Statistic &NumBranches = (I->succ_size() > 1) ? NumCondBranches
730 Statistic &BranchTakenFreq = (I->succ_size() > 1) ? CondBranchTakenFreq
731 : UncondBranchTakenFreq;
732 for (MachineBasicBlock::succ_iterator SI = I->succ_begin(),
735 // Skip if this successor is a fallthrough.
736 if (I->isLayoutSuccessor(*SI))
739 BlockFrequency EdgeFreq = BlockFreq * MBPI->getEdgeProbability(I, *SI);
741 BranchTakenFreq += EdgeFreq.getFrequency();