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/MachineFunction.h"
37 #include "llvm/CodeGen/MachineFunctionPass.h"
38 #include "llvm/CodeGen/MachineLoopInfo.h"
39 #include "llvm/CodeGen/MachineModuleInfo.h"
40 #include "llvm/Support/Allocator.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Target/TargetInstrInfo.h"
44 #include "llvm/Target/TargetLowering.h"
45 #include "llvm/Target/TargetSubtargetInfo.h"
49 #define DEBUG_TYPE "block-placement2"
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");
58 static cl::opt<unsigned> AlignAllBlock("align-all-blocks",
59 cl::desc("Force the alignment of all "
60 "blocks in the function."),
61 cl::init(0), cl::Hidden);
63 static cl::opt<bool> OnlyHotBadCFGConflictCheck(
64 "only-hot-bad-cfg-conflict-check",
65 cl::desc("Only check that a hot successor doesn't have a hot predecessor."),
66 cl::init(false), cl::Hidden);
68 static cl::opt<bool> NoBadCFGConflictCheck(
69 "no-bad-cfg-conflict-check",
70 cl::desc("Don't check whether a hot successor has a more important "
72 cl::init(false), cl::Hidden);
74 // FIXME: Find a good default for this flag and remove the flag.
75 static cl::opt<unsigned>
76 ExitBlockBias("block-placement-exit-block-bias",
77 cl::desc("Block frequency percentage a loop exit block needs "
78 "over the original exit to be considered the new exit."),
79 cl::init(0), cl::Hidden);
83 /// \brief Type for our function-wide basic block -> block chain mapping.
84 typedef DenseMap<MachineBasicBlock *, BlockChain *> BlockToChainMapType;
88 /// \brief A chain of blocks which will be laid out contiguously.
90 /// This is the datastructure representing a chain of consecutive blocks that
91 /// are profitable to layout together in order to maximize fallthrough
92 /// probabilities and code locality. We also can use a block chain to represent
93 /// a sequence of basic blocks which have some external (correctness)
94 /// requirement for sequential layout.
96 /// Chains can be built around a single basic block and can be merged to grow
97 /// them. They participate in a block-to-chain mapping, which is updated
98 /// automatically as chains are merged together.
100 /// \brief The sequence of blocks belonging to this chain.
102 /// This is the sequence of blocks for a particular chain. These will be laid
103 /// out in-order within the function.
104 SmallVector<MachineBasicBlock *, 4> Blocks;
106 /// \brief A handle to the function-wide basic block to block chain mapping.
108 /// This is retained in each block chain to simplify the computation of child
109 /// block chains for SCC-formation and iteration. We store the edges to child
110 /// basic blocks, and map them back to their associated chains using this
112 BlockToChainMapType &BlockToChain;
115 /// \brief Construct a new BlockChain.
117 /// This builds a new block chain representing a single basic block in the
118 /// function. It also registers itself as the chain that block participates
119 /// in with the BlockToChain mapping.
120 BlockChain(BlockToChainMapType &BlockToChain, MachineBasicBlock *BB)
121 : Blocks(1, BB), BlockToChain(BlockToChain), LoopPredecessors(0) {
122 assert(BB && "Cannot create a chain with a null basic block");
123 BlockToChain[BB] = this;
126 /// \brief Iterator over blocks within the chain.
127 typedef SmallVectorImpl<MachineBasicBlock *>::iterator iterator;
129 /// \brief Beginning of blocks within the chain.
130 iterator begin() { return Blocks.begin(); }
132 /// \brief End of blocks within the chain.
133 iterator end() { return Blocks.end(); }
135 /// \brief Merge a block chain into this one.
137 /// This routine merges a block chain into this one. It takes care of forming
138 /// a contiguous sequence of basic blocks, updating the edge list, and
139 /// updating the block -> chain mapping. It does not free or tear down the
140 /// old chain, but the old chain's block list is no longer valid.
141 void merge(MachineBasicBlock *BB, BlockChain *Chain) {
143 assert(!Blocks.empty());
145 // Fast path in case we don't have a chain already.
147 assert(!BlockToChain[BB]);
148 Blocks.push_back(BB);
149 BlockToChain[BB] = this;
153 assert(BB == *Chain->begin());
154 assert(Chain->begin() != Chain->end());
156 // Update the incoming blocks to point to this chain, and add them to the
158 for (BlockChain::iterator BI = Chain->begin(), BE = Chain->end();
160 Blocks.push_back(*BI);
161 assert(BlockToChain[*BI] == Chain && "Incoming blocks not in chain");
162 BlockToChain[*BI] = this;
167 /// \brief Dump the blocks in this chain.
168 LLVM_DUMP_METHOD void dump() {
169 for (iterator I = begin(), E = end(); I != E; ++I)
174 /// \brief Count of predecessors within the loop currently being processed.
176 /// This count is updated at each loop we process to represent the number of
177 /// in-loop predecessors of this chain.
178 unsigned LoopPredecessors;
183 class MachineBlockPlacement : public MachineFunctionPass {
184 /// \brief A typedef for a block filter set.
185 typedef SmallPtrSet<MachineBasicBlock *, 16> BlockFilterSet;
187 /// \brief A handle to the branch probability pass.
188 const MachineBranchProbabilityInfo *MBPI;
190 /// \brief A handle to the function-wide block frequency pass.
191 const MachineBlockFrequencyInfo *MBFI;
193 /// \brief A handle to the loop info.
194 const MachineLoopInfo *MLI;
196 /// \brief A handle to the target's instruction info.
197 const TargetInstrInfo *TII;
199 /// \brief A handle to the target's lowering info.
200 const TargetLoweringBase *TLI;
202 /// \brief Allocator and owner of BlockChain structures.
204 /// We build BlockChains lazily while processing the loop structure of
205 /// a function. To reduce malloc traffic, we allocate them using this
206 /// slab-like allocator, and destroy them after the pass completes. An
207 /// important guarantee is that this allocator produces stable pointers to
209 SpecificBumpPtrAllocator<BlockChain> ChainAllocator;
211 /// \brief Function wide BasicBlock to BlockChain mapping.
213 /// This mapping allows efficiently moving from any given basic block to the
214 /// BlockChain it participates in, if any. We use it to, among other things,
215 /// allow implicitly defining edges between chains as the existing edges
216 /// between basic blocks.
217 DenseMap<MachineBasicBlock *, BlockChain *> BlockToChain;
219 void markChainSuccessors(BlockChain &Chain,
220 MachineBasicBlock *LoopHeaderBB,
221 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
222 const BlockFilterSet *BlockFilter = nullptr);
223 MachineBasicBlock *selectBestSuccessor(MachineBasicBlock *BB,
225 const BlockFilterSet *BlockFilter);
226 MachineBasicBlock *selectBestCandidateBlock(
227 BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList,
228 const BlockFilterSet *BlockFilter);
229 MachineBasicBlock *getFirstUnplacedBlock(
231 const BlockChain &PlacedChain,
232 MachineFunction::iterator &PrevUnplacedBlockIt,
233 const BlockFilterSet *BlockFilter);
234 void buildChain(MachineBasicBlock *BB, BlockChain &Chain,
235 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
236 const BlockFilterSet *BlockFilter = nullptr);
237 MachineBasicBlock *findBestLoopTop(MachineLoop &L,
238 const BlockFilterSet &LoopBlockSet);
239 MachineBasicBlock *findBestLoopExit(MachineFunction &F,
241 const BlockFilterSet &LoopBlockSet);
242 void buildLoopChains(MachineFunction &F, MachineLoop &L);
243 void rotateLoop(BlockChain &LoopChain, MachineBasicBlock *ExitingBB,
244 const BlockFilterSet &LoopBlockSet);
245 void buildCFGChains(MachineFunction &F);
248 static char ID; // Pass identification, replacement for typeid
249 MachineBlockPlacement() : MachineFunctionPass(ID) {
250 initializeMachineBlockPlacementPass(*PassRegistry::getPassRegistry());
253 bool runOnMachineFunction(MachineFunction &F) override;
255 void getAnalysisUsage(AnalysisUsage &AU) const override {
256 AU.addRequired<MachineBranchProbabilityInfo>();
257 AU.addRequired<MachineBlockFrequencyInfo>();
258 AU.addRequired<MachineLoopInfo>();
259 MachineFunctionPass::getAnalysisUsage(AU);
264 char MachineBlockPlacement::ID = 0;
265 char &llvm::MachineBlockPlacementID = MachineBlockPlacement::ID;
266 INITIALIZE_PASS_BEGIN(MachineBlockPlacement, "block-placement2",
267 "Branch Probability Basic Block Placement", false, false)
268 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
269 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
270 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
271 INITIALIZE_PASS_END(MachineBlockPlacement, "block-placement2",
272 "Branch Probability Basic Block Placement", false, false)
275 /// \brief Helper to print the name of a MBB.
277 /// Only used by debug logging.
278 static std::string getBlockName(MachineBasicBlock *BB) {
280 raw_string_ostream OS(Result);
281 OS << "BB#" << BB->getNumber()
282 << " (derived from LLVM BB '" << BB->getName() << "')";
287 /// \brief Helper to print the number of a MBB.
289 /// Only used by debug logging.
290 static std::string getBlockNum(MachineBasicBlock *BB) {
292 raw_string_ostream OS(Result);
293 OS << "BB#" << BB->getNumber();
299 /// \brief Mark a chain's successors as having one fewer preds.
301 /// When a chain is being merged into the "placed" chain, this routine will
302 /// quickly walk the successors of each block in the chain and mark them as
303 /// having one fewer active predecessor. It also adds any successors of this
304 /// chain which reach the zero-predecessor state to the worklist passed in.
305 void MachineBlockPlacement::markChainSuccessors(
307 MachineBasicBlock *LoopHeaderBB,
308 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
309 const BlockFilterSet *BlockFilter) {
310 // Walk all the blocks in this chain, marking their successors as having
311 // a predecessor placed.
312 for (BlockChain::iterator CBI = Chain.begin(), CBE = Chain.end();
314 // Add any successors for which this is the only un-placed in-loop
315 // predecessor to the worklist as a viable candidate for CFG-neutral
316 // placement. No subsequent placement of this block will violate the CFG
317 // shape, so we get to use heuristics to choose a favorable placement.
318 for (MachineBasicBlock::succ_iterator SI = (*CBI)->succ_begin(),
319 SE = (*CBI)->succ_end();
321 if (BlockFilter && !BlockFilter->count(*SI))
323 BlockChain &SuccChain = *BlockToChain[*SI];
324 // Disregard edges within a fixed chain, or edges to the loop header.
325 if (&Chain == &SuccChain || *SI == LoopHeaderBB)
328 // This is a cross-chain edge that is within the loop, so decrement the
329 // loop predecessor count of the destination chain.
330 if (SuccChain.LoopPredecessors > 0 && --SuccChain.LoopPredecessors == 0)
331 BlockWorkList.push_back(*SuccChain.begin());
336 /// \brief Select the best successor for a block.
338 /// This looks across all successors of a particular block and attempts to
339 /// select the "best" one to be the layout successor. It only considers direct
340 /// successors which also pass the block filter. It will attempt to avoid
341 /// breaking CFG structure, but cave and break such structures in the case of
342 /// very hot successor edges.
344 /// \returns The best successor block found, or null if none are viable.
345 MachineBasicBlock *MachineBlockPlacement::selectBestSuccessor(
346 MachineBasicBlock *BB, BlockChain &Chain,
347 const BlockFilterSet *BlockFilter) {
348 const BranchProbability HotProb(4, 5); // 80%
350 MachineBasicBlock *BestSucc = nullptr;
351 // FIXME: Due to the performance of the probability and weight routines in
352 // the MBPI analysis, we manually compute probabilities using the edge
353 // weights. This is suboptimal as it means that the somewhat subtle
354 // definition of edge weight semantics is encoded here as well. We should
355 // improve the MBPI interface to efficiently support query patterns such as
357 uint32_t BestWeight = 0;
358 uint32_t WeightScale = 0;
359 uint32_t SumWeight = MBPI->getSumForBlock(BB, WeightScale);
360 DEBUG(dbgs() << "Attempting merge from: " << getBlockName(BB) << "\n");
361 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
364 if (BlockFilter && !BlockFilter->count(*SI))
366 BlockChain &SuccChain = *BlockToChain[*SI];
367 if (&SuccChain == &Chain) {
368 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> Already merged!\n");
371 if (*SI != *SuccChain.begin()) {
372 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> Mid chain!\n");
376 uint32_t SuccWeight = MBPI->getEdgeWeight(BB, *SI);
377 BranchProbability SuccProb(SuccWeight / WeightScale, SumWeight);
379 // Only consider successors which are either "hot", or wouldn't violate
380 // any CFG constraints.
381 if (SuccChain.LoopPredecessors != 0) {
382 if (SuccProb < HotProb) {
383 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> " << SuccProb
384 << " (prob) (CFG conflict)\n");
388 if (!NoBadCFGConflictCheck) {
389 // Make sure that a hot successor doesn't have a globally more
390 // important predecessor.
391 BlockFrequency CandidateEdgeFreq =
392 OnlyHotBadCFGConflictCheck
393 ? MBFI->getBlockFreq(BB) * SuccProb
394 : MBFI->getBlockFreq(BB) * SuccProb * HotProb.getCompl();
395 bool BadCFGConflict = false;
396 for (MachineBasicBlock::pred_iterator PI = (*SI)->pred_begin(),
397 PE = (*SI)->pred_end();
399 if (*PI == *SI || (BlockFilter && !BlockFilter->count(*PI)) ||
400 BlockToChain[*PI] == &Chain)
402 BlockFrequency PredEdgeFreq =
403 MBFI->getBlockFreq(*PI) * MBPI->getEdgeProbability(*PI, *SI);
404 if (PredEdgeFreq >= CandidateEdgeFreq) {
405 BadCFGConflict = true;
409 if (BadCFGConflict) {
410 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> " << SuccProb
411 << " (prob) (non-cold CFG conflict)\n");
417 DEBUG(dbgs() << " " << getBlockName(*SI) << " -> " << SuccProb
419 << (SuccChain.LoopPredecessors != 0 ? " (CFG break)" : "")
421 if (BestSucc && BestWeight >= SuccWeight)
424 BestWeight = SuccWeight;
429 /// \brief Select the best block from a worklist.
431 /// This looks through the provided worklist as a list of candidate basic
432 /// blocks and select the most profitable one to place. The definition of
433 /// profitable only really makes sense in the context of a loop. This returns
434 /// the most frequently visited block in the worklist, which in the case of
435 /// a loop, is the one most desirable to be physically close to the rest of the
436 /// loop body in order to improve icache behavior.
438 /// \returns The best block found, or null if none are viable.
439 MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock(
440 BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList,
441 const BlockFilterSet *BlockFilter) {
442 // Once we need to walk the worklist looking for a candidate, cleanup the
443 // worklist of already placed entries.
444 // FIXME: If this shows up on profiles, it could be folded (at the cost of
445 // some code complexity) into the loop below.
446 WorkList.erase(std::remove_if(WorkList.begin(), WorkList.end(),
447 [&](MachineBasicBlock *BB) {
448 return BlockToChain.lookup(BB) == &Chain;
452 MachineBasicBlock *BestBlock = nullptr;
453 BlockFrequency BestFreq;
454 for (SmallVectorImpl<MachineBasicBlock *>::iterator WBI = WorkList.begin(),
455 WBE = WorkList.end();
457 BlockChain &SuccChain = *BlockToChain[*WBI];
458 if (&SuccChain == &Chain) {
459 DEBUG(dbgs() << " " << getBlockName(*WBI)
460 << " -> Already merged!\n");
463 assert(SuccChain.LoopPredecessors == 0 && "Found CFG-violating block");
465 BlockFrequency CandidateFreq = MBFI->getBlockFreq(*WBI);
466 DEBUG(dbgs() << " " << getBlockName(*WBI) << " -> ";
467 MBFI->printBlockFreq(dbgs(), CandidateFreq) << " (freq)\n");
468 if (BestBlock && BestFreq >= CandidateFreq)
471 BestFreq = CandidateFreq;
476 /// \brief Retrieve the first unplaced basic block.
478 /// This routine is called when we are unable to use the CFG to walk through
479 /// all of the basic blocks and form a chain due to unnatural loops in the CFG.
480 /// We walk through the function's blocks in order, starting from the
481 /// LastUnplacedBlockIt. We update this iterator on each call to avoid
482 /// re-scanning the entire sequence on repeated calls to this routine.
483 MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock(
484 MachineFunction &F, const BlockChain &PlacedChain,
485 MachineFunction::iterator &PrevUnplacedBlockIt,
486 const BlockFilterSet *BlockFilter) {
487 for (MachineFunction::iterator I = PrevUnplacedBlockIt, E = F.end(); I != E;
489 if (BlockFilter && !BlockFilter->count(I))
491 if (BlockToChain[I] != &PlacedChain) {
492 PrevUnplacedBlockIt = I;
493 // Now select the head of the chain to which the unplaced block belongs
494 // as the block to place. This will force the entire chain to be placed,
495 // and satisfies the requirements of merging chains.
496 return *BlockToChain[I]->begin();
502 void MachineBlockPlacement::buildChain(
503 MachineBasicBlock *BB,
505 SmallVectorImpl<MachineBasicBlock *> &BlockWorkList,
506 const BlockFilterSet *BlockFilter) {
508 assert(BlockToChain[BB] == &Chain);
509 MachineFunction &F = *BB->getParent();
510 MachineFunction::iterator PrevUnplacedBlockIt = F.begin();
512 MachineBasicBlock *LoopHeaderBB = BB;
513 markChainSuccessors(Chain, LoopHeaderBB, BlockWorkList, BlockFilter);
514 BB = *std::prev(Chain.end());
517 assert(BlockToChain[BB] == &Chain);
518 assert(*std::prev(Chain.end()) == BB);
520 // Look for the best viable successor if there is one to place immediately
522 MachineBasicBlock *BestSucc = selectBestSuccessor(BB, Chain, BlockFilter);
524 // If an immediate successor isn't available, look for the best viable
525 // block among those we've identified as not violating the loop's CFG at
526 // this point. This won't be a fallthrough, but it will increase locality.
528 BestSucc = selectBestCandidateBlock(Chain, BlockWorkList, BlockFilter);
531 BestSucc = getFirstUnplacedBlock(F, Chain, PrevUnplacedBlockIt,
536 DEBUG(dbgs() << "Unnatural loop CFG detected, forcibly merging the "
537 "layout successor until the CFG reduces\n");
540 // Place this block, updating the datastructures to reflect its placement.
541 BlockChain &SuccChain = *BlockToChain[BestSucc];
542 // Zero out LoopPredecessors for the successor we're about to merge in case
543 // we selected a successor that didn't fit naturally into the CFG.
544 SuccChain.LoopPredecessors = 0;
545 DEBUG(dbgs() << "Merging from " << getBlockNum(BB)
546 << " to " << getBlockNum(BestSucc) << "\n");
547 markChainSuccessors(SuccChain, LoopHeaderBB, BlockWorkList, BlockFilter);
548 Chain.merge(BestSucc, &SuccChain);
549 BB = *std::prev(Chain.end());
552 DEBUG(dbgs() << "Finished forming chain for header block "
553 << getBlockNum(*Chain.begin()) << "\n");
556 /// \brief Find the best loop top block for layout.
558 /// Look for a block which is strictly better than the loop header for laying
559 /// out at the top of the loop. This looks for one and only one pattern:
560 /// a latch block with no conditional exit. This block will cause a conditional
561 /// jump around it or will be the bottom of the loop if we lay it out in place,
562 /// but if it it doesn't end up at the bottom of the loop for any reason,
563 /// rotation alone won't fix it. Because such a block will always result in an
564 /// unconditional jump (for the backedge) rotating it in front of the loop
565 /// header is always profitable.
567 MachineBlockPlacement::findBestLoopTop(MachineLoop &L,
568 const BlockFilterSet &LoopBlockSet) {
569 // Check that the header hasn't been fused with a preheader block due to
570 // crazy branches. If it has, we need to start with the header at the top to
571 // prevent pulling the preheader into the loop body.
572 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
573 if (!LoopBlockSet.count(*HeaderChain.begin()))
574 return L.getHeader();
576 DEBUG(dbgs() << "Finding best loop top for: "
577 << getBlockName(L.getHeader()) << "\n");
579 BlockFrequency BestPredFreq;
580 MachineBasicBlock *BestPred = nullptr;
581 for (MachineBasicBlock::pred_iterator PI = L.getHeader()->pred_begin(),
582 PE = L.getHeader()->pred_end();
584 MachineBasicBlock *Pred = *PI;
585 if (!LoopBlockSet.count(Pred))
587 DEBUG(dbgs() << " header pred: " << getBlockName(Pred) << ", "
588 << Pred->succ_size() << " successors, ";
589 MBFI->printBlockFreq(dbgs(), Pred) << " freq\n");
590 if (Pred->succ_size() > 1)
593 BlockFrequency PredFreq = MBFI->getBlockFreq(Pred);
594 if (!BestPred || PredFreq > BestPredFreq ||
595 (!(PredFreq < BestPredFreq) &&
596 Pred->isLayoutSuccessor(L.getHeader()))) {
598 BestPredFreq = PredFreq;
602 // If no direct predecessor is fine, just use the loop header.
604 return L.getHeader();
606 // Walk backwards through any straight line of predecessors.
607 while (BestPred->pred_size() == 1 &&
608 (*BestPred->pred_begin())->succ_size() == 1 &&
609 *BestPred->pred_begin() != L.getHeader())
610 BestPred = *BestPred->pred_begin();
612 DEBUG(dbgs() << " final top: " << getBlockName(BestPred) << "\n");
617 /// \brief Find the best loop exiting block for layout.
619 /// This routine implements the logic to analyze the loop looking for the best
620 /// block to layout at the top of the loop. Typically this is done to maximize
621 /// fallthrough opportunities.
623 MachineBlockPlacement::findBestLoopExit(MachineFunction &F,
625 const BlockFilterSet &LoopBlockSet) {
626 // We don't want to layout the loop linearly in all cases. If the loop header
627 // is just a normal basic block in the loop, we want to look for what block
628 // within the loop is the best one to layout at the top. However, if the loop
629 // header has be pre-merged into a chain due to predecessors not having
630 // analyzable branches, *and* the predecessor it is merged with is *not* part
631 // of the loop, rotating the header into the middle of the loop will create
632 // a non-contiguous range of blocks which is Very Bad. So start with the
633 // header and only rotate if safe.
634 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
635 if (!LoopBlockSet.count(*HeaderChain.begin()))
638 BlockFrequency BestExitEdgeFreq;
639 unsigned BestExitLoopDepth = 0;
640 MachineBasicBlock *ExitingBB = nullptr;
641 // If there are exits to outer loops, loop rotation can severely limit
642 // fallthrough opportunites unless it selects such an exit. Keep a set of
643 // blocks where rotating to exit with that block will reach an outer loop.
644 SmallPtrSet<MachineBasicBlock *, 4> BlocksExitingToOuterLoop;
646 DEBUG(dbgs() << "Finding best loop exit for: "
647 << getBlockName(L.getHeader()) << "\n");
648 for (MachineLoop::block_iterator I = L.block_begin(),
651 BlockChain &Chain = *BlockToChain[*I];
652 // Ensure that this block is at the end of a chain; otherwise it could be
653 // mid-way through an inner loop or a successor of an analyzable branch.
654 if (*I != *std::prev(Chain.end()))
657 // Now walk the successors. We need to establish whether this has a viable
658 // exiting successor and whether it has a viable non-exiting successor.
659 // We store the old exiting state and restore it if a viable looping
660 // successor isn't found.
661 MachineBasicBlock *OldExitingBB = ExitingBB;
662 BlockFrequency OldBestExitEdgeFreq = BestExitEdgeFreq;
663 bool HasLoopingSucc = false;
664 // FIXME: Due to the performance of the probability and weight routines in
665 // the MBPI analysis, we use the internal weights and manually compute the
666 // probabilities to avoid quadratic behavior.
667 uint32_t WeightScale = 0;
668 uint32_t SumWeight = MBPI->getSumForBlock(*I, WeightScale);
669 for (MachineBasicBlock::succ_iterator SI = (*I)->succ_begin(),
670 SE = (*I)->succ_end();
672 if ((*SI)->isLandingPad())
676 BlockChain &SuccChain = *BlockToChain[*SI];
677 // Don't split chains, either this chain or the successor's chain.
678 if (&Chain == &SuccChain) {
679 DEBUG(dbgs() << " exiting: " << getBlockName(*I) << " -> "
680 << getBlockName(*SI) << " (chain conflict)\n");
684 uint32_t SuccWeight = MBPI->getEdgeWeight(*I, *SI);
685 if (LoopBlockSet.count(*SI)) {
686 DEBUG(dbgs() << " looping: " << getBlockName(*I) << " -> "
687 << getBlockName(*SI) << " (" << SuccWeight << ")\n");
688 HasLoopingSucc = true;
692 unsigned SuccLoopDepth = 0;
693 if (MachineLoop *ExitLoop = MLI->getLoopFor(*SI)) {
694 SuccLoopDepth = ExitLoop->getLoopDepth();
695 if (ExitLoop->contains(&L))
696 BlocksExitingToOuterLoop.insert(*I);
699 BranchProbability SuccProb(SuccWeight / WeightScale, SumWeight);
700 BlockFrequency ExitEdgeFreq = MBFI->getBlockFreq(*I) * SuccProb;
701 DEBUG(dbgs() << " exiting: " << getBlockName(*I) << " -> "
702 << getBlockName(*SI) << " [L:" << SuccLoopDepth
704 MBFI->printBlockFreq(dbgs(), ExitEdgeFreq) << ")\n");
705 // Note that we bias this toward an existing layout successor to retain
706 // incoming order in the absence of better information. The exit must have
707 // a frequency higher than the current exit before we consider breaking
709 BranchProbability Bias(100 - ExitBlockBias, 100);
710 if (!ExitingBB || BestExitLoopDepth < SuccLoopDepth ||
711 ExitEdgeFreq > BestExitEdgeFreq ||
712 ((*I)->isLayoutSuccessor(*SI) &&
713 !(ExitEdgeFreq < BestExitEdgeFreq * Bias))) {
714 BestExitEdgeFreq = ExitEdgeFreq;
719 // Restore the old exiting state, no viable looping successor was found.
720 if (!HasLoopingSucc) {
721 ExitingBB = OldExitingBB;
722 BestExitEdgeFreq = OldBestExitEdgeFreq;
726 // Without a candidate exiting block or with only a single block in the
727 // loop, just use the loop header to layout the loop.
728 if (!ExitingBB || L.getNumBlocks() == 1)
731 // Also, if we have exit blocks which lead to outer loops but didn't select
732 // one of them as the exiting block we are rotating toward, disable loop
733 // rotation altogether.
734 if (!BlocksExitingToOuterLoop.empty() &&
735 !BlocksExitingToOuterLoop.count(ExitingBB))
738 DEBUG(dbgs() << " Best exiting block: " << getBlockName(ExitingBB) << "\n");
742 /// \brief Attempt to rotate an exiting block to the bottom of the loop.
744 /// Once we have built a chain, try to rotate it to line up the hot exit block
745 /// with fallthrough out of the loop if doing so doesn't introduce unnecessary
746 /// branches. For example, if the loop has fallthrough into its header and out
747 /// of its bottom already, don't rotate it.
748 void MachineBlockPlacement::rotateLoop(BlockChain &LoopChain,
749 MachineBasicBlock *ExitingBB,
750 const BlockFilterSet &LoopBlockSet) {
754 MachineBasicBlock *Top = *LoopChain.begin();
755 bool ViableTopFallthrough = false;
756 for (MachineBasicBlock::pred_iterator PI = Top->pred_begin(),
757 PE = Top->pred_end();
759 BlockChain *PredChain = BlockToChain[*PI];
760 if (!LoopBlockSet.count(*PI) &&
761 (!PredChain || *PI == *std::prev(PredChain->end()))) {
762 ViableTopFallthrough = true;
767 // If the header has viable fallthrough, check whether the current loop
768 // bottom is a viable exiting block. If so, bail out as rotating will
769 // introduce an unnecessary branch.
770 if (ViableTopFallthrough) {
771 MachineBasicBlock *Bottom = *std::prev(LoopChain.end());
772 for (MachineBasicBlock::succ_iterator SI = Bottom->succ_begin(),
773 SE = Bottom->succ_end();
775 BlockChain *SuccChain = BlockToChain[*SI];
776 if (!LoopBlockSet.count(*SI) &&
777 (!SuccChain || *SI == *SuccChain->begin()))
782 BlockChain::iterator ExitIt = std::find(LoopChain.begin(), LoopChain.end(),
784 if (ExitIt == LoopChain.end())
787 std::rotate(LoopChain.begin(), std::next(ExitIt), LoopChain.end());
790 /// \brief Forms basic block chains from the natural loop structures.
792 /// These chains are designed to preserve the existing *structure* of the code
793 /// as much as possible. We can then stitch the chains together in a way which
794 /// both preserves the topological structure and minimizes taken conditional
796 void MachineBlockPlacement::buildLoopChains(MachineFunction &F,
798 // First recurse through any nested loops, building chains for those inner
800 for (MachineLoop::iterator LI = L.begin(), LE = L.end(); LI != LE; ++LI)
801 buildLoopChains(F, **LI);
803 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
804 BlockFilterSet LoopBlockSet(L.block_begin(), L.block_end());
806 // First check to see if there is an obviously preferable top block for the
807 // loop. This will default to the header, but may end up as one of the
808 // predecessors to the header if there is one which will result in strictly
809 // fewer branches in the loop body.
810 MachineBasicBlock *LoopTop = findBestLoopTop(L, LoopBlockSet);
812 // If we selected just the header for the loop top, look for a potentially
813 // profitable exit block in the event that rotating the loop can eliminate
814 // branches by placing an exit edge at the bottom.
815 MachineBasicBlock *ExitingBB = nullptr;
816 if (LoopTop == L.getHeader())
817 ExitingBB = findBestLoopExit(F, L, LoopBlockSet);
819 BlockChain &LoopChain = *BlockToChain[LoopTop];
821 // FIXME: This is a really lame way of walking the chains in the loop: we
822 // walk the blocks, and use a set to prevent visiting a particular chain
824 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
825 assert(LoopChain.LoopPredecessors == 0);
826 UpdatedPreds.insert(&LoopChain);
827 for (MachineLoop::block_iterator BI = L.block_begin(),
830 BlockChain &Chain = *BlockToChain[*BI];
831 if (!UpdatedPreds.insert(&Chain).second)
834 assert(Chain.LoopPredecessors == 0);
835 for (BlockChain::iterator BCI = Chain.begin(), BCE = Chain.end();
837 assert(BlockToChain[*BCI] == &Chain);
838 for (MachineBasicBlock::pred_iterator PI = (*BCI)->pred_begin(),
839 PE = (*BCI)->pred_end();
841 if (BlockToChain[*PI] == &Chain || !LoopBlockSet.count(*PI))
843 ++Chain.LoopPredecessors;
847 if (Chain.LoopPredecessors == 0)
848 BlockWorkList.push_back(*Chain.begin());
851 buildChain(LoopTop, LoopChain, BlockWorkList, &LoopBlockSet);
852 rotateLoop(LoopChain, ExitingBB, LoopBlockSet);
855 // Crash at the end so we get all of the debugging output first.
856 bool BadLoop = false;
857 if (LoopChain.LoopPredecessors) {
859 dbgs() << "Loop chain contains a block without its preds placed!\n"
860 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
861 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n";
863 for (BlockChain::iterator BCI = LoopChain.begin(), BCE = LoopChain.end();
865 dbgs() << " ... " << getBlockName(*BCI) << "\n";
866 if (!LoopBlockSet.erase(*BCI)) {
867 // We don't mark the loop as bad here because there are real situations
868 // where this can occur. For example, with an unanalyzable fallthrough
869 // from a loop block to a non-loop block or vice versa.
870 dbgs() << "Loop chain contains a block not contained by the loop!\n"
871 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
872 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
873 << " Bad block: " << getBlockName(*BCI) << "\n";
877 if (!LoopBlockSet.empty()) {
879 for (BlockFilterSet::iterator LBI = LoopBlockSet.begin(),
880 LBE = LoopBlockSet.end();
882 dbgs() << "Loop contains blocks never placed into a chain!\n"
883 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
884 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
885 << " Bad block: " << getBlockName(*LBI) << "\n";
887 assert(!BadLoop && "Detected problems with the placement of this loop.");
891 void MachineBlockPlacement::buildCFGChains(MachineFunction &F) {
892 // Ensure that every BB in the function has an associated chain to simplify
893 // the assumptions of the remaining algorithm.
894 SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
895 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
896 MachineBasicBlock *BB = FI;
898 = new (ChainAllocator.Allocate()) BlockChain(BlockToChain, BB);
899 // Also, merge any blocks which we cannot reason about and must preserve
900 // the exact fallthrough behavior for.
903 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
904 if (!TII->AnalyzeBranch(*BB, TBB, FBB, Cond) || !FI->canFallThrough())
907 MachineFunction::iterator NextFI(std::next(FI));
908 MachineBasicBlock *NextBB = NextFI;
909 // Ensure that the layout successor is a viable block, as we know that
910 // fallthrough is a possibility.
911 assert(NextFI != FE && "Can't fallthrough past the last block.");
912 DEBUG(dbgs() << "Pre-merging due to unanalyzable fallthrough: "
913 << getBlockName(BB) << " -> " << getBlockName(NextBB)
915 Chain->merge(NextBB, nullptr);
921 // Build any loop-based chains.
922 for (MachineLoopInfo::iterator LI = MLI->begin(), LE = MLI->end(); LI != LE;
924 buildLoopChains(F, **LI);
926 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
928 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
929 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) {
930 MachineBasicBlock *BB = &*FI;
931 BlockChain &Chain = *BlockToChain[BB];
932 if (!UpdatedPreds.insert(&Chain).second)
935 assert(Chain.LoopPredecessors == 0);
936 for (BlockChain::iterator BCI = Chain.begin(), BCE = Chain.end();
938 assert(BlockToChain[*BCI] == &Chain);
939 for (MachineBasicBlock::pred_iterator PI = (*BCI)->pred_begin(),
940 PE = (*BCI)->pred_end();
942 if (BlockToChain[*PI] == &Chain)
944 ++Chain.LoopPredecessors;
948 if (Chain.LoopPredecessors == 0)
949 BlockWorkList.push_back(*Chain.begin());
952 BlockChain &FunctionChain = *BlockToChain[&F.front()];
953 buildChain(&F.front(), FunctionChain, BlockWorkList);
956 typedef SmallPtrSet<MachineBasicBlock *, 16> FunctionBlockSetType;
959 // Crash at the end so we get all of the debugging output first.
960 bool BadFunc = false;
961 FunctionBlockSetType FunctionBlockSet;
962 for (MachineFunction::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI)
963 FunctionBlockSet.insert(FI);
965 for (BlockChain::iterator BCI = FunctionChain.begin(),
966 BCE = FunctionChain.end();
968 if (!FunctionBlockSet.erase(*BCI)) {
970 dbgs() << "Function chain contains a block not in the function!\n"
971 << " Bad block: " << getBlockName(*BCI) << "\n";
974 if (!FunctionBlockSet.empty()) {
976 for (FunctionBlockSetType::iterator FBI = FunctionBlockSet.begin(),
977 FBE = FunctionBlockSet.end();
979 dbgs() << "Function contains blocks never placed into a chain!\n"
980 << " Bad block: " << getBlockName(*FBI) << "\n";
982 assert(!BadFunc && "Detected problems with the block placement.");
985 // Splice the blocks into place.
986 MachineFunction::iterator InsertPos = F.begin();
987 for (BlockChain::iterator BI = FunctionChain.begin(),
988 BE = FunctionChain.end();
990 DEBUG(dbgs() << (BI == FunctionChain.begin() ? "Placing chain "
992 << getBlockName(*BI) << "\n");
993 if (InsertPos != MachineFunction::iterator(*BI))
994 F.splice(InsertPos, *BI);
998 // Update the terminator of the previous block.
999 if (BI == FunctionChain.begin())
1001 MachineBasicBlock *PrevBB = std::prev(MachineFunction::iterator(*BI));
1003 // FIXME: It would be awesome of updateTerminator would just return rather
1004 // than assert when the branch cannot be analyzed in order to remove this
1007 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
1008 if (!TII->AnalyzeBranch(*PrevBB, TBB, FBB, Cond)) {
1009 // The "PrevBB" is not yet updated to reflect current code layout, so,
1010 // o. it may fall-through to a block without explict "goto" instruction
1011 // before layout, and no longer fall-through it after layout; or
1012 // o. just opposite.
1014 // AnalyzeBranch() may return erroneous value for FBB when these two
1015 // situations take place. For the first scenario FBB is mistakenly set
1016 // NULL; for the 2nd scenario, the FBB, which is expected to be NULL,
1017 // is mistakenly pointing to "*BI".
1019 bool needUpdateBr = true;
1020 if (!Cond.empty() && (!FBB || FBB == *BI)) {
1021 PrevBB->updateTerminator();
1022 needUpdateBr = false;
1024 TBB = FBB = nullptr;
1025 if (TII->AnalyzeBranch(*PrevBB, TBB, FBB, Cond)) {
1026 // FIXME: This should never take place.
1027 TBB = FBB = nullptr;
1031 // If PrevBB has a two-way branch, try to re-order the branches
1032 // such that we branch to the successor with higher weight first.
1033 if (TBB && !Cond.empty() && FBB &&
1034 MBPI->getEdgeWeight(PrevBB, FBB) > MBPI->getEdgeWeight(PrevBB, TBB) &&
1035 !TII->ReverseBranchCondition(Cond)) {
1036 DEBUG(dbgs() << "Reverse order of the two branches: "
1037 << getBlockName(PrevBB) << "\n");
1038 DEBUG(dbgs() << " Edge weight: " << MBPI->getEdgeWeight(PrevBB, FBB)
1039 << " vs " << MBPI->getEdgeWeight(PrevBB, TBB) << "\n");
1040 DebugLoc dl; // FIXME: this is nowhere
1041 TII->RemoveBranch(*PrevBB);
1042 TII->InsertBranch(*PrevBB, FBB, TBB, Cond, dl);
1043 needUpdateBr = true;
1046 PrevBB->updateTerminator();
1050 // Fixup the last block.
1052 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
1053 if (!TII->AnalyzeBranch(F.back(), TBB, FBB, Cond))
1054 F.back().updateTerminator();
1056 // Walk through the backedges of the function now that we have fully laid out
1057 // the basic blocks and align the destination of each backedge. We don't rely
1058 // exclusively on the loop info here so that we can align backedges in
1059 // unnatural CFGs and backedges that were introduced purely because of the
1060 // loop rotations done during this layout pass.
1061 if (F.getFunction()->getAttributes().
1062 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize))
1064 if (FunctionChain.begin() == FunctionChain.end())
1065 return; // Empty chain.
1067 const BranchProbability ColdProb(1, 5); // 20%
1068 BlockFrequency EntryFreq = MBFI->getBlockFreq(F.begin());
1069 BlockFrequency WeightedEntryFreq = EntryFreq * ColdProb;
1070 for (BlockChain::iterator BI = std::next(FunctionChain.begin()),
1071 BE = FunctionChain.end();
1073 // Don't align non-looping basic blocks. These are unlikely to execute
1074 // enough times to matter in practice. Note that we'll still handle
1075 // unnatural CFGs inside of a natural outer loop (the common case) and
1077 MachineLoop *L = MLI->getLoopFor(*BI);
1081 unsigned Align = TLI->getPrefLoopAlignment(L);
1083 continue; // Don't care about loop alignment.
1085 // If the block is cold relative to the function entry don't waste space
1087 BlockFrequency Freq = MBFI->getBlockFreq(*BI);
1088 if (Freq < WeightedEntryFreq)
1091 // If the block is cold relative to its loop header, don't align it
1092 // regardless of what edges into the block exist.
1093 MachineBasicBlock *LoopHeader = L->getHeader();
1094 BlockFrequency LoopHeaderFreq = MBFI->getBlockFreq(LoopHeader);
1095 if (Freq < (LoopHeaderFreq * ColdProb))
1098 // Check for the existence of a non-layout predecessor which would benefit
1099 // from aligning this block.
1100 MachineBasicBlock *LayoutPred = *std::prev(BI);
1102 // Force alignment if all the predecessors are jumps. We already checked
1103 // that the block isn't cold above.
1104 if (!LayoutPred->isSuccessor(*BI)) {
1105 (*BI)->setAlignment(Align);
1109 // Align this block if the layout predecessor's edge into this block is
1110 // cold relative to the block. When this is true, other predecessors make up
1111 // all of the hot entries into the block and thus alignment is likely to be
1113 BranchProbability LayoutProb = MBPI->getEdgeProbability(LayoutPred, *BI);
1114 BlockFrequency LayoutEdgeFreq = MBFI->getBlockFreq(LayoutPred) * LayoutProb;
1115 if (LayoutEdgeFreq <= (Freq * ColdProb))
1116 (*BI)->setAlignment(Align);
1120 bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &F) {
1121 // Check for single-block functions and skip them.
1122 if (std::next(F.begin()) == F.end())
1125 if (skipOptnoneFunction(*F.getFunction()))
1128 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
1129 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
1130 MLI = &getAnalysis<MachineLoopInfo>();
1131 TII = F.getSubtarget().getInstrInfo();
1132 TLI = F.getSubtarget().getTargetLowering();
1133 assert(BlockToChain.empty());
1137 BlockToChain.clear();
1138 ChainAllocator.DestroyAll();
1141 // Align all of the blocks in the function to a specific alignment.
1142 for (MachineFunction::iterator FI = F.begin(), FE = F.end();
1144 FI->setAlignment(AlignAllBlock);
1146 // We always return true as we have no way to track whether the final order
1147 // differs from the original order.
1152 /// \brief A pass to compute block placement statistics.
1154 /// A separate pass to compute interesting statistics for evaluating block
1155 /// placement. This is separate from the actual placement pass so that they can
1156 /// be computed in the absence of any placement transformations or when using
1157 /// alternative placement strategies.
1158 class MachineBlockPlacementStats : public MachineFunctionPass {
1159 /// \brief A handle to the branch probability pass.
1160 const MachineBranchProbabilityInfo *MBPI;
1162 /// \brief A handle to the function-wide block frequency pass.
1163 const MachineBlockFrequencyInfo *MBFI;
1166 static char ID; // Pass identification, replacement for typeid
1167 MachineBlockPlacementStats() : MachineFunctionPass(ID) {
1168 initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry());
1171 bool runOnMachineFunction(MachineFunction &F) override;
1173 void getAnalysisUsage(AnalysisUsage &AU) const override {
1174 AU.addRequired<MachineBranchProbabilityInfo>();
1175 AU.addRequired<MachineBlockFrequencyInfo>();
1176 AU.setPreservesAll();
1177 MachineFunctionPass::getAnalysisUsage(AU);
1182 char MachineBlockPlacementStats::ID = 0;
1183 char &llvm::MachineBlockPlacementStatsID = MachineBlockPlacementStats::ID;
1184 INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats",
1185 "Basic Block Placement Stats", false, false)
1186 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
1187 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
1188 INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats",
1189 "Basic Block Placement Stats", false, false)
1191 bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) {
1192 // Check for single-block functions and skip them.
1193 if (std::next(F.begin()) == F.end())
1196 MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
1197 MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
1199 for (MachineFunction::iterator I = F.begin(), E = F.end(); I != E; ++I) {
1200 BlockFrequency BlockFreq = MBFI->getBlockFreq(I);
1201 Statistic &NumBranches = (I->succ_size() > 1) ? NumCondBranches
1202 : NumUncondBranches;
1203 Statistic &BranchTakenFreq = (I->succ_size() > 1) ? CondBranchTakenFreq
1204 : UncondBranchTakenFreq;
1205 for (MachineBasicBlock::succ_iterator SI = I->succ_begin(),
1208 // Skip if this successor is a fallthrough.
1209 if (I->isLayoutSuccessor(*SI))
1212 BlockFrequency EdgeFreq = BlockFreq * MBPI->getEdgeProbability(I, *SI);
1214 BranchTakenFreq += EdgeFreq.getFrequency();