Line data Source code
1 : //===- MachineBlockPlacement.cpp - Basic Block Code Layout optimization ---===//
2 : //
3 : // The LLVM Compiler Infrastructure
4 : //
5 : // This file is distributed under the University of Illinois Open Source
6 : // License. See LICENSE.TXT for details.
7 : //
8 : //===----------------------------------------------------------------------===//
9 : //
10 : // This file implements basic block placement transformations using the CFG
11 : // structure and branch probability estimates.
12 : //
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.
18 : //
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
24 : // function in-order.
25 : //
26 : //===----------------------------------------------------------------------===//
27 :
28 : #include "BranchFolding.h"
29 : #include "llvm/ADT/ArrayRef.h"
30 : #include "llvm/ADT/DenseMap.h"
31 : #include "llvm/ADT/STLExtras.h"
32 : #include "llvm/ADT/SetVector.h"
33 : #include "llvm/ADT/SmallPtrSet.h"
34 : #include "llvm/ADT/SmallVector.h"
35 : #include "llvm/ADT/Statistic.h"
36 : #include "llvm/Analysis/BlockFrequencyInfoImpl.h"
37 : #include "llvm/CodeGen/MachineBasicBlock.h"
38 : #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
39 : #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
40 : #include "llvm/CodeGen/MachineFunction.h"
41 : #include "llvm/CodeGen/MachineFunctionPass.h"
42 : #include "llvm/CodeGen/MachineLoopInfo.h"
43 : #include "llvm/CodeGen/MachineModuleInfo.h"
44 : #include "llvm/CodeGen/MachinePostDominators.h"
45 : #include "llvm/CodeGen/TailDuplicator.h"
46 : #include "llvm/CodeGen/TargetInstrInfo.h"
47 : #include "llvm/CodeGen/TargetLowering.h"
48 : #include "llvm/CodeGen/TargetPassConfig.h"
49 : #include "llvm/CodeGen/TargetSubtargetInfo.h"
50 : #include "llvm/IR/DebugLoc.h"
51 : #include "llvm/IR/Function.h"
52 : #include "llvm/Pass.h"
53 : #include "llvm/Support/Allocator.h"
54 : #include "llvm/Support/BlockFrequency.h"
55 : #include "llvm/Support/BranchProbability.h"
56 : #include "llvm/Support/CodeGen.h"
57 : #include "llvm/Support/CommandLine.h"
58 : #include "llvm/Support/Compiler.h"
59 : #include "llvm/Support/Debug.h"
60 : #include "llvm/Support/raw_ostream.h"
61 : #include "llvm/Target/TargetMachine.h"
62 : #include <algorithm>
63 : #include <cassert>
64 : #include <cstdint>
65 : #include <iterator>
66 : #include <memory>
67 : #include <string>
68 : #include <tuple>
69 : #include <utility>
70 : #include <vector>
71 :
72 : using namespace llvm;
73 :
74 : #define DEBUG_TYPE "block-placement"
75 :
76 : STATISTIC(NumCondBranches, "Number of conditional branches");
77 : STATISTIC(NumUncondBranches, "Number of unconditional branches");
78 : STATISTIC(CondBranchTakenFreq,
79 : "Potential frequency of taking conditional branches");
80 : STATISTIC(UncondBranchTakenFreq,
81 : "Potential frequency of taking unconditional branches");
82 :
83 : static cl::opt<unsigned> AlignAllBlock("align-all-blocks",
84 : cl::desc("Force the alignment of all "
85 : "blocks in the function."),
86 : cl::init(0), cl::Hidden);
87 :
88 : static cl::opt<unsigned> AlignAllNonFallThruBlocks(
89 : "align-all-nofallthru-blocks",
90 : cl::desc("Force the alignment of all "
91 : "blocks that have no fall-through predecessors (i.e. don't add "
92 : "nops that are executed)."),
93 : cl::init(0), cl::Hidden);
94 :
95 : // FIXME: Find a good default for this flag and remove the flag.
96 : static cl::opt<unsigned> ExitBlockBias(
97 : "block-placement-exit-block-bias",
98 : cl::desc("Block frequency percentage a loop exit block needs "
99 : "over the original exit to be considered the new exit."),
100 : cl::init(0), cl::Hidden);
101 :
102 : // Definition:
103 : // - Outlining: placement of a basic block outside the chain or hot path.
104 :
105 : static cl::opt<unsigned> LoopToColdBlockRatio(
106 : "loop-to-cold-block-ratio",
107 : cl::desc("Outline loop blocks from loop chain if (frequency of loop) / "
108 : "(frequency of block) is greater than this ratio"),
109 : cl::init(5), cl::Hidden);
110 :
111 : static cl::opt<bool> ForceLoopColdBlock(
112 : "force-loop-cold-block",
113 : cl::desc("Force outlining cold blocks from loops."),
114 : cl::init(false), cl::Hidden);
115 :
116 : static cl::opt<bool>
117 : PreciseRotationCost("precise-rotation-cost",
118 : cl::desc("Model the cost of loop rotation more "
119 : "precisely by using profile data."),
120 : cl::init(false), cl::Hidden);
121 :
122 : static cl::opt<bool>
123 : ForcePreciseRotationCost("force-precise-rotation-cost",
124 : cl::desc("Force the use of precise cost "
125 : "loop rotation strategy."),
126 : cl::init(false), cl::Hidden);
127 :
128 : static cl::opt<unsigned> MisfetchCost(
129 : "misfetch-cost",
130 : cl::desc("Cost that models the probabilistic risk of an instruction "
131 : "misfetch due to a jump comparing to falling through, whose cost "
132 : "is zero."),
133 : cl::init(1), cl::Hidden);
134 :
135 : static cl::opt<unsigned> JumpInstCost("jump-inst-cost",
136 : cl::desc("Cost of jump instructions."),
137 : cl::init(1), cl::Hidden);
138 : static cl::opt<bool>
139 : TailDupPlacement("tail-dup-placement",
140 : cl::desc("Perform tail duplication during placement. "
141 : "Creates more fallthrough opportunites in "
142 : "outline branches."),
143 : cl::init(true), cl::Hidden);
144 :
145 : static cl::opt<bool>
146 : BranchFoldPlacement("branch-fold-placement",
147 : cl::desc("Perform branch folding during placement. "
148 : "Reduces code size."),
149 : cl::init(true), cl::Hidden);
150 :
151 : // Heuristic for tail duplication.
152 : static cl::opt<unsigned> TailDupPlacementThreshold(
153 : "tail-dup-placement-threshold",
154 : cl::desc("Instruction cutoff for tail duplication during layout. "
155 : "Tail merging during layout is forced to have a threshold "
156 : "that won't conflict."), cl::init(2),
157 : cl::Hidden);
158 :
159 : // Heuristic for aggressive tail duplication.
160 : static cl::opt<unsigned> TailDupPlacementAggressiveThreshold(
161 : "tail-dup-placement-aggressive-threshold",
162 : cl::desc("Instruction cutoff for aggressive tail duplication during "
163 : "layout. Used at -O3. Tail merging during layout is forced to "
164 : "have a threshold that won't conflict."), cl::init(4),
165 : cl::Hidden);
166 :
167 : // Heuristic for tail duplication.
168 : static cl::opt<unsigned> TailDupPlacementPenalty(
169 : "tail-dup-placement-penalty",
170 : cl::desc("Cost penalty for blocks that can avoid breaking CFG by copying. "
171 : "Copying can increase fallthrough, but it also increases icache "
172 : "pressure. This parameter controls the penalty to account for that. "
173 : "Percent as integer."),
174 : cl::init(2),
175 : cl::Hidden);
176 :
177 : // Heuristic for triangle chains.
178 : static cl::opt<unsigned> TriangleChainCount(
179 : "triangle-chain-count",
180 : cl::desc("Number of triangle-shaped-CFG's that need to be in a row for the "
181 : "triangle tail duplication heuristic to kick in. 0 to disable."),
182 : cl::init(2),
183 : cl::Hidden);
184 :
185 : extern cl::opt<unsigned> StaticLikelyProb;
186 : extern cl::opt<unsigned> ProfileLikelyProb;
187 :
188 : // Internal option used to control BFI display only after MBP pass.
189 : // Defined in CodeGen/MachineBlockFrequencyInfo.cpp:
190 : // -view-block-layout-with-bfi=
191 : extern cl::opt<GVDAGType> ViewBlockLayoutWithBFI;
192 :
193 : // Command line option to specify the name of the function for CFG dump
194 : // Defined in Analysis/BlockFrequencyInfo.cpp: -view-bfi-func-name=
195 : extern cl::opt<std::string> ViewBlockFreqFuncName;
196 :
197 : namespace {
198 :
199 : class BlockChain;
200 :
201 : /// Type for our function-wide basic block -> block chain mapping.
202 : using BlockToChainMapType = DenseMap<const MachineBasicBlock *, BlockChain *>;
203 :
204 : /// A chain of blocks which will be laid out contiguously.
205 : ///
206 : /// This is the datastructure representing a chain of consecutive blocks that
207 : /// are profitable to layout together in order to maximize fallthrough
208 : /// probabilities and code locality. We also can use a block chain to represent
209 : /// a sequence of basic blocks which have some external (correctness)
210 : /// requirement for sequential layout.
211 : ///
212 : /// Chains can be built around a single basic block and can be merged to grow
213 : /// them. They participate in a block-to-chain mapping, which is updated
214 : /// automatically as chains are merged together.
215 347913 : class BlockChain {
216 : /// The sequence of blocks belonging to this chain.
217 : ///
218 : /// This is the sequence of blocks for a particular chain. These will be laid
219 : /// out in-order within the function.
220 : SmallVector<MachineBasicBlock *, 4> Blocks;
221 :
222 : /// A handle to the function-wide basic block to block chain mapping.
223 : ///
224 : /// This is retained in each block chain to simplify the computation of child
225 : /// block chains for SCC-formation and iteration. We store the edges to child
226 : /// basic blocks, and map them back to their associated chains using this
227 : /// structure.
228 : BlockToChainMapType &BlockToChain;
229 :
230 : public:
231 : /// Construct a new BlockChain.
232 : ///
233 : /// This builds a new block chain representing a single basic block in the
234 : /// function. It also registers itself as the chain that block participates
235 : /// in with the BlockToChain mapping.
236 347913 : BlockChain(BlockToChainMapType &BlockToChain, MachineBasicBlock *BB)
237 347913 : : Blocks(1, BB), BlockToChain(BlockToChain) {
238 : assert(BB && "Cannot create a chain with a null basic block");
239 347913 : BlockToChain[BB] = this;
240 347913 : }
241 :
242 : /// Iterator over blocks within the chain.
243 : using iterator = SmallVectorImpl<MachineBasicBlock *>::iterator;
244 : using const_iterator = SmallVectorImpl<MachineBasicBlock *>::const_iterator;
245 :
246 : /// Beginning of blocks within the chain.
247 : iterator begin() { return Blocks.begin(); }
248 : const_iterator begin() const { return Blocks.begin(); }
249 :
250 : /// End of blocks within the chain.
251 : iterator end() { return Blocks.end(); }
252 : const_iterator end() const { return Blocks.end(); }
253 :
254 712 : bool remove(MachineBasicBlock* BB) {
255 712 : for(iterator i = begin(); i != end(); ++i) {
256 712 : if (*i == BB) {
257 712 : Blocks.erase(i);
258 712 : return true;
259 : }
260 : }
261 : return false;
262 : }
263 :
264 : /// Merge a block chain into this one.
265 : ///
266 : /// This routine merges a block chain into this one. It takes care of forming
267 : /// a contiguous sequence of basic blocks, updating the edge list, and
268 : /// updating the block -> chain mapping. It does not free or tear down the
269 : /// old chain, but the old chain's block list is no longer valid.
270 330795 : void merge(MachineBasicBlock *BB, BlockChain *Chain) {
271 : assert(BB && "Can't merge a null block.");
272 : assert(!Blocks.empty() && "Can't merge into an empty chain.");
273 :
274 : // Fast path in case we don't have a chain already.
275 330795 : if (!Chain) {
276 : assert(!BlockToChain[BB] &&
277 : "Passed chain is null, but BB has entry in BlockToChain.");
278 1497 : Blocks.push_back(BB);
279 1497 : BlockToChain[BB] = this;
280 1497 : return;
281 : }
282 :
283 : assert(BB == *Chain->begin() && "Passed BB is not head of Chain.");
284 : assert(Chain->begin() != Chain->end());
285 :
286 : // Update the incoming blocks to point to this chain, and add them to the
287 : // chain structure.
288 709734 : for (MachineBasicBlock *ChainBB : *Chain) {
289 380436 : Blocks.push_back(ChainBB);
290 : assert(BlockToChain[ChainBB] == Chain && "Incoming blocks not in chain.");
291 380436 : BlockToChain[ChainBB] = this;
292 : }
293 : }
294 :
295 : #ifndef NDEBUG
296 : /// Dump the blocks in this chain.
297 : LLVM_DUMP_METHOD void dump() {
298 : for (MachineBasicBlock *MBB : *this)
299 : MBB->dump();
300 : }
301 : #endif // NDEBUG
302 :
303 : /// Count of predecessors of any block within the chain which have not
304 : /// yet been scheduled. In general, we will delay scheduling this chain
305 : /// until those predecessors are scheduled (or we find a sufficiently good
306 : /// reason to override this heuristic.) Note that when forming loop chains,
307 : /// blocks outside the loop are ignored and treated as if they were already
308 : /// scheduled.
309 : ///
310 : /// Note: This field is reinitialized multiple times - once for each loop,
311 : /// and then once for the function as a whole.
312 : unsigned UnscheduledPredecessors = 0;
313 : };
314 :
315 : class MachineBlockPlacement : public MachineFunctionPass {
316 : /// A type for a block filter set.
317 : using BlockFilterSet = SmallSetVector<const MachineBasicBlock *, 16>;
318 :
319 : /// Pair struct containing basic block and taildup profitiability
320 : struct BlockAndTailDupResult {
321 : MachineBasicBlock *BB;
322 : bool ShouldTailDup;
323 : };
324 :
325 : /// Triple struct containing edge weight and the edge.
326 : struct WeightedEdge {
327 : BlockFrequency Weight;
328 : MachineBasicBlock *Src;
329 : MachineBasicBlock *Dest;
330 : };
331 :
332 : /// work lists of blocks that are ready to be laid out
333 : SmallVector<MachineBasicBlock *, 16> BlockWorkList;
334 : SmallVector<MachineBasicBlock *, 16> EHPadWorkList;
335 :
336 : /// Edges that have already been computed as optimal.
337 : DenseMap<const MachineBasicBlock *, BlockAndTailDupResult> ComputedEdges;
338 :
339 : /// Machine Function
340 : MachineFunction *F;
341 :
342 : /// A handle to the branch probability pass.
343 : const MachineBranchProbabilityInfo *MBPI;
344 :
345 : /// A handle to the function-wide block frequency pass.
346 : std::unique_ptr<BranchFolder::MBFIWrapper> MBFI;
347 :
348 : /// A handle to the loop info.
349 : MachineLoopInfo *MLI;
350 :
351 : /// Preferred loop exit.
352 : /// Member variable for convenience. It may be removed by duplication deep
353 : /// in the call stack.
354 : MachineBasicBlock *PreferredLoopExit;
355 :
356 : /// A handle to the target's instruction info.
357 : const TargetInstrInfo *TII;
358 :
359 : /// A handle to the target's lowering info.
360 : const TargetLoweringBase *TLI;
361 :
362 : /// A handle to the post dominator tree.
363 : MachinePostDominatorTree *MPDT;
364 :
365 : /// Duplicator used to duplicate tails during placement.
366 : ///
367 : /// Placement decisions can open up new tail duplication opportunities, but
368 : /// since tail duplication affects placement decisions of later blocks, it
369 : /// must be done inline.
370 : TailDuplicator TailDup;
371 :
372 : /// Allocator and owner of BlockChain structures.
373 : ///
374 : /// We build BlockChains lazily while processing the loop structure of
375 : /// a function. To reduce malloc traffic, we allocate them using this
376 : /// slab-like allocator, and destroy them after the pass completes. An
377 : /// important guarantee is that this allocator produces stable pointers to
378 : /// the chains.
379 : SpecificBumpPtrAllocator<BlockChain> ChainAllocator;
380 :
381 : /// Function wide BasicBlock to BlockChain mapping.
382 : ///
383 : /// This mapping allows efficiently moving from any given basic block to the
384 : /// BlockChain it participates in, if any. We use it to, among other things,
385 : /// allow implicitly defining edges between chains as the existing edges
386 : /// between basic blocks.
387 : DenseMap<const MachineBasicBlock *, BlockChain *> BlockToChain;
388 :
389 : #ifndef NDEBUG
390 : /// The set of basic blocks that have terminators that cannot be fully
391 : /// analyzed. These basic blocks cannot be re-ordered safely by
392 : /// MachineBlockPlacement, and we must preserve physical layout of these
393 : /// blocks and their successors through the pass.
394 : SmallPtrSet<MachineBasicBlock *, 4> BlocksWithUnanalyzableExits;
395 : #endif
396 :
397 : /// Decrease the UnscheduledPredecessors count for all blocks in chain, and
398 : /// if the count goes to 0, add them to the appropriate work list.
399 : void markChainSuccessors(
400 : const BlockChain &Chain, const MachineBasicBlock *LoopHeaderBB,
401 : const BlockFilterSet *BlockFilter = nullptr);
402 :
403 : /// Decrease the UnscheduledPredecessors count for a single block, and
404 : /// if the count goes to 0, add them to the appropriate work list.
405 : void markBlockSuccessors(
406 : const BlockChain &Chain, const MachineBasicBlock *BB,
407 : const MachineBasicBlock *LoopHeaderBB,
408 : const BlockFilterSet *BlockFilter = nullptr);
409 :
410 : BranchProbability
411 : collectViableSuccessors(
412 : const MachineBasicBlock *BB, const BlockChain &Chain,
413 : const BlockFilterSet *BlockFilter,
414 : SmallVector<MachineBasicBlock *, 4> &Successors);
415 : bool shouldPredBlockBeOutlined(
416 : const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
417 : const BlockChain &Chain, const BlockFilterSet *BlockFilter,
418 : BranchProbability SuccProb, BranchProbability HotProb);
419 : bool repeatedlyTailDuplicateBlock(
420 : MachineBasicBlock *BB, MachineBasicBlock *&LPred,
421 : const MachineBasicBlock *LoopHeaderBB,
422 : BlockChain &Chain, BlockFilterSet *BlockFilter,
423 : MachineFunction::iterator &PrevUnplacedBlockIt);
424 : bool maybeTailDuplicateBlock(
425 : MachineBasicBlock *BB, MachineBasicBlock *LPred,
426 : BlockChain &Chain, BlockFilterSet *BlockFilter,
427 : MachineFunction::iterator &PrevUnplacedBlockIt,
428 : bool &DuplicatedToLPred);
429 : bool hasBetterLayoutPredecessor(
430 : const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
431 : const BlockChain &SuccChain, BranchProbability SuccProb,
432 : BranchProbability RealSuccProb, const BlockChain &Chain,
433 : const BlockFilterSet *BlockFilter);
434 : BlockAndTailDupResult selectBestSuccessor(
435 : const MachineBasicBlock *BB, const BlockChain &Chain,
436 : const BlockFilterSet *BlockFilter);
437 : MachineBasicBlock *selectBestCandidateBlock(
438 : const BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList);
439 : MachineBasicBlock *getFirstUnplacedBlock(
440 : const BlockChain &PlacedChain,
441 : MachineFunction::iterator &PrevUnplacedBlockIt,
442 : const BlockFilterSet *BlockFilter);
443 :
444 : /// Add a basic block to the work list if it is appropriate.
445 : ///
446 : /// If the optional parameter BlockFilter is provided, only MBB
447 : /// present in the set will be added to the worklist. If nullptr
448 : /// is provided, no filtering occurs.
449 : void fillWorkLists(const MachineBasicBlock *MBB,
450 : SmallPtrSetImpl<BlockChain *> &UpdatedPreds,
451 : const BlockFilterSet *BlockFilter);
452 :
453 : void buildChain(const MachineBasicBlock *BB, BlockChain &Chain,
454 : BlockFilterSet *BlockFilter = nullptr);
455 : MachineBasicBlock *findBestLoopTop(
456 : const MachineLoop &L, const BlockFilterSet &LoopBlockSet);
457 : MachineBasicBlock *findBestLoopExit(
458 : const MachineLoop &L, const BlockFilterSet &LoopBlockSet);
459 : BlockFilterSet collectLoopBlockSet(const MachineLoop &L);
460 : void buildLoopChains(const MachineLoop &L);
461 : void rotateLoop(
462 : BlockChain &LoopChain, const MachineBasicBlock *ExitingBB,
463 : const BlockFilterSet &LoopBlockSet);
464 : void rotateLoopWithProfile(
465 : BlockChain &LoopChain, const MachineLoop &L,
466 : const BlockFilterSet &LoopBlockSet);
467 : void buildCFGChains();
468 : void optimizeBranches();
469 : void alignBlocks();
470 : /// Returns true if a block should be tail-duplicated to increase fallthrough
471 : /// opportunities.
472 : bool shouldTailDuplicate(MachineBasicBlock *BB);
473 : /// Check the edge frequencies to see if tail duplication will increase
474 : /// fallthroughs.
475 : bool isProfitableToTailDup(
476 : const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
477 : BranchProbability QProb,
478 : const BlockChain &Chain, const BlockFilterSet *BlockFilter);
479 :
480 : /// Check for a trellis layout.
481 : bool isTrellis(const MachineBasicBlock *BB,
482 : const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
483 : const BlockChain &Chain, const BlockFilterSet *BlockFilter);
484 :
485 : /// Get the best successor given a trellis layout.
486 : BlockAndTailDupResult getBestTrellisSuccessor(
487 : const MachineBasicBlock *BB,
488 : const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
489 : BranchProbability AdjustedSumProb, const BlockChain &Chain,
490 : const BlockFilterSet *BlockFilter);
491 :
492 : /// Get the best pair of non-conflicting edges.
493 : static std::pair<WeightedEdge, WeightedEdge> getBestNonConflictingEdges(
494 : const MachineBasicBlock *BB,
495 : MutableArrayRef<SmallVector<WeightedEdge, 8>> Edges);
496 :
497 : /// Returns true if a block can tail duplicate into all unplaced
498 : /// predecessors. Filters based on loop.
499 : bool canTailDuplicateUnplacedPreds(
500 : const MachineBasicBlock *BB, MachineBasicBlock *Succ,
501 : const BlockChain &Chain, const BlockFilterSet *BlockFilter);
502 :
503 : /// Find chains of triangles to tail-duplicate where a global analysis works,
504 : /// but a local analysis would not find them.
505 : void precomputeTriangleChains();
506 :
507 : public:
508 : static char ID; // Pass identification, replacement for typeid
509 :
510 40376 : MachineBlockPlacement() : MachineFunctionPass(ID) {
511 20188 : initializeMachineBlockPlacementPass(*PassRegistry::getPassRegistry());
512 20188 : }
513 :
514 : bool runOnMachineFunction(MachineFunction &F) override;
515 :
516 0 : bool allowTailDupPlacement() const {
517 : assert(F);
518 810616 : return TailDupPlacement && !F->getTarget().requiresStructuredCFG();
519 : }
520 :
521 20031 : void getAnalysisUsage(AnalysisUsage &AU) const override {
522 : AU.addRequired<MachineBranchProbabilityInfo>();
523 : AU.addRequired<MachineBlockFrequencyInfo>();
524 20031 : if (TailDupPlacement)
525 : AU.addRequired<MachinePostDominatorTree>();
526 : AU.addRequired<MachineLoopInfo>();
527 : AU.addRequired<TargetPassConfig>();
528 20031 : MachineFunctionPass::getAnalysisUsage(AU);
529 20031 : }
530 : };
531 :
532 : } // end anonymous namespace
533 :
534 : char MachineBlockPlacement::ID = 0;
535 :
536 : char &llvm::MachineBlockPlacementID = MachineBlockPlacement::ID;
537 :
538 31780 : INITIALIZE_PASS_BEGIN(MachineBlockPlacement, DEBUG_TYPE,
539 : "Branch Probability Basic Block Placement", false, false)
540 31780 : INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
541 31780 : INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
542 31780 : INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTree)
543 31780 : INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
544 105335 : INITIALIZE_PASS_END(MachineBlockPlacement, DEBUG_TYPE,
545 : "Branch Probability Basic Block Placement", false, false)
546 :
547 : #ifndef NDEBUG
548 : /// Helper to print the name of a MBB.
549 : ///
550 : /// Only used by debug logging.
551 : static std::string getBlockName(const MachineBasicBlock *BB) {
552 : std::string Result;
553 : raw_string_ostream OS(Result);
554 : OS << printMBBReference(*BB);
555 : OS << " ('" << BB->getName() << "')";
556 : OS.flush();
557 : return Result;
558 : }
559 : #endif
560 :
561 : /// Mark a chain's successors as having one fewer preds.
562 : ///
563 : /// When a chain is being merged into the "placed" chain, this routine will
564 : /// quickly walk the successors of each block in the chain and mark them as
565 : /// having one fewer active predecessor. It also adds any successors of this
566 : /// chain which reach the zero-predecessor state to the appropriate worklist.
567 356926 : void MachineBlockPlacement::markChainSuccessors(
568 : const BlockChain &Chain, const MachineBasicBlock *LoopHeaderBB,
569 : const BlockFilterSet *BlockFilter) {
570 : // Walk all the blocks in this chain, marking their successors as having
571 : // a predecessor placed.
572 766329 : for (MachineBasicBlock *MBB : Chain) {
573 409403 : markBlockSuccessors(Chain, MBB, LoopHeaderBB, BlockFilter);
574 : }
575 356926 : }
576 :
577 : /// Mark a single block's successors as having one fewer preds.
578 : ///
579 : /// Under normal circumstances, this is only called by markChainSuccessors,
580 : /// but if a block that was to be placed is completely tail-duplicated away,
581 : /// and was duplicated into the chain end, we need to redo markBlockSuccessors
582 : /// for just that block.
583 410115 : void MachineBlockPlacement::markBlockSuccessors(
584 : const BlockChain &Chain, const MachineBasicBlock *MBB,
585 : const MachineBasicBlock *LoopHeaderBB, const BlockFilterSet *BlockFilter) {
586 : // Add any successors for which this is the only un-placed in-loop
587 : // predecessor to the worklist as a viable candidate for CFG-neutral
588 : // placement. No subsequent placement of this block will violate the CFG
589 : // shape, so we get to use heuristics to choose a favorable placement.
590 982351 : for (MachineBasicBlock *Succ : MBB->successors()) {
591 675921 : if (BlockFilter && !BlockFilter->count(Succ))
592 249624 : continue;
593 546560 : BlockChain &SuccChain = *BlockToChain[Succ];
594 : // Disregard edges within a fixed chain, or edges to the loop header.
595 546560 : if (&Chain == &SuccChain || Succ == LoopHeaderBB)
596 : continue;
597 :
598 : // This is a cross-chain edge that is within the loop, so decrement the
599 : // loop predecessor count of the destination chain.
600 456954 : if (SuccChain.UnscheduledPredecessors == 0 ||
601 445861 : --SuccChain.UnscheduledPredecessors > 0)
602 134342 : continue;
603 :
604 322612 : auto *NewBB = *SuccChain.begin();
605 322612 : if (NewBB->isEHPad())
606 50793 : EHPadWorkList.push_back(NewBB);
607 : else
608 271819 : BlockWorkList.push_back(NewBB);
609 : }
610 410115 : }
611 :
612 : /// This helper function collects the set of successors of block
613 : /// \p BB that are allowed to be its layout successors, and return
614 : /// the total branch probability of edges from \p BB to those
615 : /// blocks.
616 360757 : BranchProbability MachineBlockPlacement::collectViableSuccessors(
617 : const MachineBasicBlock *BB, const BlockChain &Chain,
618 : const BlockFilterSet *BlockFilter,
619 : SmallVector<MachineBasicBlock *, 4> &Successors) {
620 : // Adjust edge probabilities by excluding edges pointing to blocks that is
621 : // either not in BlockFilter or is already in the current chain. Consider the
622 : // following CFG:
623 : //
624 : // --->A
625 : // | / \
626 : // | B C
627 : // | \ / \
628 : // ----D E
629 : //
630 : // Assume A->C is very hot (>90%), and C->D has a 50% probability, then after
631 : // A->C is chosen as a fall-through, D won't be selected as a successor of C
632 : // due to CFG constraint (the probability of C->D is not greater than
633 : // HotProb to break topo-order). If we exclude E that is not in BlockFilter
634 : // when calculating the probability of C->D, D will be selected and we
635 : // will get A C D B as the layout of this loop.
636 : auto AdjustedSumProb = BranchProbability::getOne();
637 850055 : for (MachineBasicBlock *Succ : BB->successors()) {
638 : bool SkipSucc = false;
639 569522 : if (Succ->isEHPad() || (BlockFilter && !BlockFilter->count(Succ))) {
640 : SkipSucc = true;
641 : } else {
642 416609 : BlockChain *SuccChain = BlockToChain[Succ];
643 416609 : if (SuccChain == &Chain) {
644 : SkipSucc = true;
645 385485 : } else if (Succ != *SuccChain->begin()) {
646 : LLVM_DEBUG(dbgs() << " " << getBlockName(Succ)
647 : << " -> Mid chain!\n");
648 : continue;
649 : }
650 : }
651 : if (SkipSucc)
652 103813 : AdjustedSumProb -= MBPI->getEdgeProbability(BB, Succ);
653 : else
654 384176 : Successors.push_back(Succ);
655 : }
656 :
657 360757 : return AdjustedSumProb;
658 : }
659 :
660 : /// The helper function returns the branch probability that is adjusted
661 : /// or normalized over the new total \p AdjustedSumProb.
662 : static BranchProbability
663 : getAdjustedProbability(BranchProbability OrigProb,
664 : BranchProbability AdjustedSumProb) {
665 : BranchProbability SuccProb;
666 : uint32_t SuccProbN = OrigProb.getNumerator();
667 : uint32_t SuccProbD = AdjustedSumProb.getNumerator();
668 378558 : if (SuccProbN >= SuccProbD)
669 : SuccProb = BranchProbability::getOne();
670 : else
671 147549 : SuccProb = BranchProbability(SuccProbN, SuccProbD);
672 :
673 : return SuccProb;
674 : }
675 :
676 : /// Check if \p BB has exactly the successors in \p Successors.
677 : static bool
678 7893 : hasSameSuccessors(MachineBasicBlock &BB,
679 : SmallPtrSetImpl<const MachineBasicBlock *> &Successors) {
680 15786 : if (BB.succ_size() != Successors.size())
681 : return false;
682 : // We don't want to count self-loops
683 5181 : if (Successors.count(&BB))
684 : return false;
685 9789 : for (MachineBasicBlock *Succ : BB.successors())
686 8614 : if (!Successors.count(Succ))
687 : return false;
688 : return true;
689 : }
690 :
691 : /// Check if a block should be tail duplicated to increase fallthrough
692 : /// opportunities.
693 : /// \p BB Block to check.
694 365291 : bool MachineBlockPlacement::shouldTailDuplicate(MachineBasicBlock *BB) {
695 : // Blocks with single successors don't create additional fallthrough
696 : // opportunities. Don't duplicate them. TODO: When conditional exits are
697 : // analyzable, allow them to be duplicated.
698 365291 : bool IsSimple = TailDup.isSimpleBB(BB);
699 :
700 365291 : if (BB->succ_size() == 1)
701 : return false;
702 237596 : return TailDup.shouldTailDuplicate(IsSimple, *BB);
703 : }
704 :
705 : /// Compare 2 BlockFrequency's with a small penalty for \p A.
706 : /// In order to be conservative, we apply a X% penalty to account for
707 : /// increased icache pressure and static heuristics. For small frequencies
708 : /// we use only the numerators to improve accuracy. For simplicity, we assume the
709 : /// penalty is less than 100%
710 : /// TODO(iteratee): Use 64-bit fixed point edge frequencies everywhere.
711 3119 : static bool greaterWithBias(BlockFrequency A, BlockFrequency B,
712 : uint64_t EntryFreq) {
713 3119 : BranchProbability ThresholdProb(TailDupPlacementPenalty, 100);
714 3119 : BlockFrequency Gain = A - B;
715 3119 : return (Gain / ThresholdProb).getFrequency() >= EntryFreq;
716 : }
717 :
718 : /// Check the edge frequencies to see if tail duplication will increase
719 : /// fallthroughs. It only makes sense to call this function when
720 : /// \p Succ would not be chosen otherwise. Tail duplication of \p Succ is
721 : /// always locally profitable if we would have picked \p Succ without
722 : /// considering duplication.
723 3119 : bool MachineBlockPlacement::isProfitableToTailDup(
724 : const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
725 : BranchProbability QProb,
726 : const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
727 : // We need to do a probability calculation to make sure this is profitable.
728 : // First: does succ have a successor that post-dominates? This affects the
729 : // calculation. The 2 relevant cases are:
730 : // BB BB
731 : // | \Qout | \Qout
732 : // P| C |P C
733 : // = C' = C'
734 : // | /Qin | /Qin
735 : // | / | /
736 : // Succ Succ
737 : // / \ | \ V
738 : // U/ =V |U \
739 : // / \ = D
740 : // D E | /
741 : // | /
742 : // |/
743 : // PDom
744 : // '=' : Branch taken for that CFG edge
745 : // In the second case, Placing Succ while duplicating it into C prevents the
746 : // fallthrough of Succ into either D or PDom, because they now have C as an
747 : // unplaced predecessor
748 :
749 : // Start by figuring out which case we fall into
750 : MachineBasicBlock *PDom = nullptr;
751 : SmallVector<MachineBasicBlock *, 4> SuccSuccs;
752 : // Only scan the relevant successors
753 : auto AdjustedSuccSumProb =
754 3119 : collectViableSuccessors(Succ, Chain, BlockFilter, SuccSuccs);
755 3119 : BranchProbability PProb = MBPI->getEdgeProbability(BB, Succ);
756 3119 : auto BBFreq = MBFI->getBlockFreq(BB);
757 3119 : auto SuccFreq = MBFI->getBlockFreq(Succ);
758 3119 : BlockFrequency P = BBFreq * PProb;
759 3119 : BlockFrequency Qout = BBFreq * QProb;
760 3119 : uint64_t EntryFreq = MBFI->getEntryFreq();
761 : // If there are no more successors, it is profitable to copy, as it strictly
762 : // increases fallthrough.
763 6238 : if (SuccSuccs.size() == 0)
764 1701 : return greaterWithBias(P, Qout, EntryFreq);
765 :
766 : auto BestSuccSucc = BranchProbability::getZero();
767 : // Find the PDom or the best Succ if no PDom exists.
768 2995 : for (MachineBasicBlock *SuccSucc : SuccSuccs) {
769 2210 : auto Prob = MBPI->getEdgeProbability(Succ, SuccSucc);
770 2210 : if (Prob > BestSuccSucc)
771 : BestSuccSucc = Prob;
772 : if (PDom == nullptr)
773 2210 : if (MPDT->dominates(SuccSucc, Succ)) {
774 : PDom = SuccSucc;
775 633 : break;
776 : }
777 : }
778 : // For the comparisons, we need to know Succ's best incoming edge that isn't
779 : // from BB.
780 : auto SuccBestPred = BlockFrequency(0);
781 4840 : for (MachineBasicBlock *SuccPred : Succ->predecessors()) {
782 3422 : if (SuccPred == Succ || SuccPred == BB
783 2004 : || BlockToChain[SuccPred] == &Chain
784 5430 : || (BlockFilter && !BlockFilter->count(SuccPred)))
785 1823 : continue;
786 1599 : auto Freq = MBFI->getBlockFreq(SuccPred)
787 3198 : * MBPI->getEdgeProbability(SuccPred, Succ);
788 1599 : if (Freq > SuccBestPred)
789 : SuccBestPred = Freq;
790 : }
791 : // Qin is Succ's best unplaced incoming edge that isn't BB
792 1418 : BlockFrequency Qin = SuccBestPred;
793 : // If it doesn't have a post-dominating successor, here is the calculation:
794 : // BB BB
795 : // | \Qout | \
796 : // P| C | =
797 : // = C' | C
798 : // | /Qin | |
799 : // | / | C' (+Succ)
800 : // Succ Succ /|
801 : // / \ | \/ |
802 : // U/ =V | == |
803 : // / \ | / \|
804 : // D E D E
805 : // '=' : Branch taken for that CFG edge
806 : // Cost in the first case is: P + V
807 : // For this calculation, we always assume P > Qout. If Qout > P
808 : // The result of this function will be ignored at the caller.
809 : // Let F = SuccFreq - Qin
810 : // Cost in the second case is: Qout + min(Qin, F) * U + max(Qin, F) * V
811 :
812 1418 : if (PDom == nullptr || !Succ->isSuccessor(PDom)) {
813 : BranchProbability UProb = BestSuccSucc;
814 785 : BranchProbability VProb = AdjustedSuccSumProb - UProb;
815 785 : BlockFrequency F = SuccFreq - Qin;
816 785 : BlockFrequency V = SuccFreq * VProb;
817 785 : BlockFrequency QinU = std::min(Qin, F) * UProb;
818 785 : BlockFrequency BaseCost = P + V;
819 785 : BlockFrequency DupCost = Qout + QinU + std::max(Qin, F) * VProb;
820 785 : return greaterWithBias(BaseCost, DupCost, EntryFreq);
821 : }
822 633 : BranchProbability UProb = MBPI->getEdgeProbability(Succ, PDom);
823 633 : BranchProbability VProb = AdjustedSuccSumProb - UProb;
824 633 : BlockFrequency U = SuccFreq * UProb;
825 633 : BlockFrequency V = SuccFreq * VProb;
826 633 : BlockFrequency F = SuccFreq - Qin;
827 : // If there is a post-dominating successor, here is the calculation:
828 : // BB BB BB BB
829 : // | \Qout | \ | \Qout | \
830 : // |P C | = |P C | =
831 : // = C' |P C = C' |P C
832 : // | /Qin | | | /Qin | |
833 : // | / | C' (+Succ) | / | C' (+Succ)
834 : // Succ Succ /| Succ Succ /|
835 : // | \ V | \/ | | \ V | \/ |
836 : // |U \ |U /\ =? |U = |U /\ |
837 : // = D = = =?| | D | = =|
838 : // | / |/ D | / |/ D
839 : // | / | / | = | /
840 : // |/ | / |/ | =
841 : // Dom Dom Dom Dom
842 : // '=' : Branch taken for that CFG edge
843 : // The cost for taken branches in the first case is P + U
844 : // Let F = SuccFreq - Qin
845 : // The cost in the second case (assuming independence), given the layout:
846 : // BB, Succ, (C+Succ), D, Dom or the layout:
847 : // BB, Succ, D, Dom, (C+Succ)
848 : // is Qout + max(F, Qin) * U + min(F, Qin)
849 : // compare P + U vs Qout + P * U + Qin.
850 : //
851 : // The 3rd and 4th cases cover when Dom would be chosen to follow Succ.
852 : //
853 : // For the 3rd case, the cost is P + 2 * V
854 : // For the 4th case, the cost is Qout + min(Qin, F) * U + max(Qin, F) * V + V
855 : // We choose 4 over 3 when (P + V) > Qout + min(Qin, F) * U + max(Qin, F) * V
856 637 : if (UProb > AdjustedSuccSumProb / 2 &&
857 637 : !hasBetterLayoutPredecessor(Succ, PDom, *BlockToChain[PDom], UProb, UProb,
858 : Chain, BlockFilter))
859 : // Cases 3 & 4
860 0 : return greaterWithBias(
861 0 : (P + V), (Qout + std::max(Qin, F) * VProb + std::min(Qin, F) * UProb),
862 : EntryFreq);
863 : // Cases 1 & 2
864 1266 : return greaterWithBias((P + U),
865 1266 : (Qout + std::min(Qin, F) * AdjustedSuccSumProb +
866 : std::max(Qin, F) * UProb),
867 : EntryFreq);
868 : }
869 :
870 : /// Check for a trellis layout. \p BB is the upper part of a trellis if its
871 : /// successors form the lower part of a trellis. A successor set S forms the
872 : /// lower part of a trellis if all of the predecessors of S are either in S or
873 : /// have all of S as successors. We ignore trellises where BB doesn't have 2
874 : /// successors because for fewer than 2, it's trivial, and for 3 or greater they
875 : /// are very uncommon and complex to compute optimally. Allowing edges within S
876 : /// is not strictly a trellis, but the same algorithm works, so we allow it.
877 356531 : bool MachineBlockPlacement::isTrellis(
878 : const MachineBasicBlock *BB,
879 : const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
880 : const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
881 : // Technically BB could form a trellis with branching factor higher than 2.
882 : // But that's extremely uncommon.
883 356531 : if (BB->succ_size() != 2 || ViableSuccs.size() != 2)
884 : return false;
885 :
886 : SmallPtrSet<const MachineBasicBlock *, 2> Successors(BB->succ_begin(),
887 73022 : BB->succ_end());
888 : // To avoid reviewing the same predecessors twice.
889 : SmallPtrSet<const MachineBasicBlock *, 8> SeenPreds;
890 :
891 74386 : for (MachineBasicBlock *Succ : ViableSuccs) {
892 : int PredCount = 0;
893 162796 : for (auto SuccPred : Succ->predecessors()) {
894 : // Allow triangle successors, but don't count them.
895 96413 : if (Successors.count(SuccPred)) {
896 : // Make sure that it is actually a triangle.
897 13362 : for (MachineBasicBlock *CheckSucc : SuccPred->successors())
898 8310 : if (!Successors.count(CheckSucc))
899 : return false;
900 : continue;
901 : }
902 89684 : const BlockChain *PredChain = BlockToChain[SuccPred];
903 21839 : if (SuccPred == BB || (BlockFilter && !BlockFilter->count(SuccPred)) ||
904 97165 : PredChain == &Chain || PredChain == BlockToChain[Succ])
905 82347 : continue;
906 7337 : ++PredCount;
907 : // Perform the successor check only once.
908 7337 : if (!SeenPreds.insert(SuccPred).second)
909 : continue;
910 6680 : if (!hasSameSuccessors(*SuccPred, Successors))
911 : return false;
912 : }
913 : // If one of the successors has only BB as a predecessor, it is not a
914 : // trellis.
915 66383 : if (PredCount < 1)
916 : return false;
917 : }
918 : return true;
919 : }
920 :
921 : /// Pick the highest total weight pair of edges that can both be laid out.
922 : /// The edges in \p Edges[0] are assumed to have a different destination than
923 : /// the edges in \p Edges[1]. Simple counting shows that the best pair is either
924 : /// the individual highest weight edges to the 2 different destinations, or in
925 : /// case of a conflict, one of them should be replaced with a 2nd best edge.
926 : std::pair<MachineBlockPlacement::WeightedEdge,
927 : MachineBlockPlacement::WeightedEdge>
928 0 : MachineBlockPlacement::getBestNonConflictingEdges(
929 : const MachineBasicBlock *BB,
930 : MutableArrayRef<SmallVector<MachineBlockPlacement::WeightedEdge, 8>>
931 : Edges) {
932 : // Sort the edges, and then for each successor, find the best incoming
933 : // predecessor. If the best incoming predecessors aren't the same,
934 : // then that is clearly the best layout. If there is a conflict, one of the
935 : // successors will have to fallthrough from the second best predecessor. We
936 : // compare which combination is better overall.
937 :
938 : // Sort for highest frequency.
939 : auto Cmp = [](WeightedEdge A, WeightedEdge B) { return A.Weight > B.Weight; };
940 :
941 : std::stable_sort(Edges[0].begin(), Edges[0].end(), Cmp);
942 : std::stable_sort(Edges[1].begin(), Edges[1].end(), Cmp);
943 : auto BestA = Edges[0].begin();
944 : auto BestB = Edges[1].begin();
945 : // Arrange for the correct answer to be in BestA and BestB
946 : // If the 2 best edges don't conflict, the answer is already there.
947 0 : if (BestA->Src == BestB->Src) {
948 : // Compare the total fallthrough of (Best + Second Best) for both pairs
949 : auto SecondBestA = std::next(BestA);
950 : auto SecondBestB = std::next(BestB);
951 0 : BlockFrequency BestAScore = BestA->Weight + SecondBestB->Weight;
952 0 : BlockFrequency BestBScore = BestB->Weight + SecondBestA->Weight;
953 0 : if (BestAScore < BestBScore)
954 : BestA = SecondBestA;
955 : else
956 : BestB = SecondBestB;
957 : }
958 : // Arrange for the BB edge to be in BestA if it exists.
959 0 : if (BestB->Src == BB)
960 : std::swap(BestA, BestB);
961 0 : return std::make_pair(*BestA, *BestB);
962 : }
963 :
964 : /// Get the best successor from \p BB based on \p BB being part of a trellis.
965 : /// We only handle trellises with 2 successors, so the algorithm is
966 : /// straightforward: Find the best pair of edges that don't conflict. We find
967 : /// the best incoming edge for each successor in the trellis. If those conflict,
968 : /// we consider which of them should be replaced with the second best.
969 : /// Upon return the two best edges will be in \p BestEdges. If one of the edges
970 : /// comes from \p BB, it will be in \p BestEdges[0]
971 : MachineBlockPlacement::BlockAndTailDupResult
972 0 : MachineBlockPlacement::getBestTrellisSuccessor(
973 : const MachineBasicBlock *BB,
974 : const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
975 : BranchProbability AdjustedSumProb, const BlockChain &Chain,
976 : const BlockFilterSet *BlockFilter) {
977 :
978 : BlockAndTailDupResult Result = {nullptr, false};
979 : SmallPtrSet<const MachineBasicBlock *, 4> Successors(BB->succ_begin(),
980 0 : BB->succ_end());
981 :
982 : // We assume size 2 because it's common. For general n, we would have to do
983 : // the Hungarian algorithm, but it's not worth the complexity because more
984 : // than 2 successors is fairly uncommon, and a trellis even more so.
985 0 : if (Successors.size() != 2 || ViableSuccs.size() != 2)
986 0 : return Result;
987 :
988 : // Collect the edge frequencies of all edges that form the trellis.
989 0 : SmallVector<WeightedEdge, 8> Edges[2];
990 : int SuccIndex = 0;
991 0 : for (auto Succ : ViableSuccs) {
992 0 : for (MachineBasicBlock *SuccPred : Succ->predecessors()) {
993 : // Skip any placed predecessors that are not BB
994 0 : if (SuccPred != BB)
995 0 : if ((BlockFilter && !BlockFilter->count(SuccPred)) ||
996 0 : BlockToChain[SuccPred] == &Chain ||
997 0 : BlockToChain[SuccPred] == BlockToChain[Succ])
998 0 : continue;
999 0 : BlockFrequency EdgeFreq = MBFI->getBlockFreq(SuccPred) *
1000 0 : MBPI->getEdgeProbability(SuccPred, Succ);
1001 0 : Edges[SuccIndex].push_back({EdgeFreq, SuccPred, Succ});
1002 : }
1003 0 : ++SuccIndex;
1004 : }
1005 :
1006 : // Pick the best combination of 2 edges from all the edges in the trellis.
1007 : WeightedEdge BestA, BestB;
1008 0 : std::tie(BestA, BestB) = getBestNonConflictingEdges(BB, Edges);
1009 :
1010 0 : if (BestA.Src != BB) {
1011 : // If we have a trellis, and BB doesn't have the best fallthrough edges,
1012 : // we shouldn't choose any successor. We've already looked and there's a
1013 : // better fallthrough edge for all the successors.
1014 : LLVM_DEBUG(dbgs() << "Trellis, but not one of the chosen edges.\n");
1015 0 : return Result;
1016 : }
1017 :
1018 : // Did we pick the triangle edge? If tail-duplication is profitable, do
1019 : // that instead. Otherwise merge the triangle edge now while we know it is
1020 : // optimal.
1021 0 : if (BestA.Dest == BestB.Src) {
1022 : // The edges are BB->Succ1->Succ2, and we're looking to see if BB->Succ2
1023 : // would be better.
1024 : MachineBasicBlock *Succ1 = BestA.Dest;
1025 0 : MachineBasicBlock *Succ2 = BestB.Dest;
1026 : // Check to see if tail-duplication would be profitable.
1027 0 : if (allowTailDupPlacement() && shouldTailDuplicate(Succ2) &&
1028 0 : canTailDuplicateUnplacedPreds(BB, Succ2, Chain, BlockFilter) &&
1029 0 : isProfitableToTailDup(BB, Succ2, MBPI->getEdgeProbability(BB, Succ1),
1030 : Chain, BlockFilter)) {
1031 : LLVM_DEBUG(BranchProbability Succ2Prob = getAdjustedProbability(
1032 : MBPI->getEdgeProbability(BB, Succ2), AdjustedSumProb);
1033 : dbgs() << " Selected: " << getBlockName(Succ2)
1034 : << ", probability: " << Succ2Prob
1035 : << " (Tail Duplicate)\n");
1036 : Result.BB = Succ2;
1037 : Result.ShouldTailDup = true;
1038 0 : return Result;
1039 : }
1040 : }
1041 : // We have already computed the optimal edge for the other side of the
1042 : // trellis.
1043 0 : ComputedEdges[BestB.Src] = { BestB.Dest, false };
1044 :
1045 : auto TrellisSucc = BestA.Dest;
1046 : LLVM_DEBUG(BranchProbability SuccProb = getAdjustedProbability(
1047 : MBPI->getEdgeProbability(BB, TrellisSucc), AdjustedSumProb);
1048 : dbgs() << " Selected: " << getBlockName(TrellisSucc)
1049 : << ", probability: " << SuccProb << " (Trellis)\n");
1050 : Result.BB = TrellisSucc;
1051 0 : return Result;
1052 : }
1053 :
1054 : /// When the option allowTailDupPlacement() is on, this method checks if the
1055 : /// fallthrough candidate block \p Succ (of block \p BB) can be tail-duplicated
1056 : /// into all of its unplaced, unfiltered predecessors, that are not BB.
1057 4371 : bool MachineBlockPlacement::canTailDuplicateUnplacedPreds(
1058 : const MachineBasicBlock *BB, MachineBasicBlock *Succ,
1059 : const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
1060 4371 : if (!shouldTailDuplicate(Succ))
1061 : return false;
1062 :
1063 : // For CFG checking.
1064 : SmallPtrSet<const MachineBasicBlock *, 4> Successors(BB->succ_begin(),
1065 4371 : BB->succ_end());
1066 13303 : for (MachineBasicBlock *Pred : Succ->predecessors()) {
1067 : // Make sure all unplaced and unfiltered predecessors can be
1068 : // tail-duplicated into.
1069 : // Skip any blocks that are already placed or not in this loop.
1070 7332 : if (Pred == BB || (BlockFilter && !BlockFilter->count(Pred))
1071 16526 : || BlockToChain[Pred] == &Chain)
1072 5260 : continue;
1073 4924 : if (!TailDup.canTailDuplicate(Succ, Pred)) {
1074 2980 : if (Successors.size() > 1 && hasSameSuccessors(*Pred, Successors))
1075 : // This will result in a trellis after tail duplication, so we don't
1076 : // need to copy Succ into this predecessor. In the presence
1077 : // of a trellis tail duplication can continue to be profitable.
1078 : // For example:
1079 : // A A
1080 : // |\ |\
1081 : // | \ | \
1082 : // | C | C+BB
1083 : // | / | |
1084 : // |/ | |
1085 : // BB => BB |
1086 : // |\ |\/|
1087 : // | \ |/\|
1088 : // | D | D
1089 : // | / | /
1090 : // |/ |/
1091 : // Succ Succ
1092 : //
1093 : // After BB was duplicated into C, the layout looks like the one on the
1094 : // right. BB and C now have the same successors. When considering
1095 : // whether Succ can be duplicated into all its unplaced predecessors, we
1096 : // ignore C.
1097 : // We can do this because C already has a profitable fallthrough, namely
1098 : // D. TODO(iteratee): ignore sufficiently cold predecessors for
1099 : // duplication and for this test.
1100 : //
1101 : // This allows trellises to be laid out in 2 separate chains
1102 : // (A,B,Succ,...) and later (C,D,...) This is a reasonable heuristic
1103 : // because it allows the creation of 2 fallthrough paths with links
1104 : // between them, and we correctly identify the best layout for these
1105 : // CFGs. We want to extend trellises that the user created in addition
1106 : // to trellises created by tail-duplication, so we just look for the
1107 : // CFG.
1108 : continue;
1109 : return false;
1110 : }
1111 : }
1112 : return true;
1113 : }
1114 :
1115 : /// Find chains of triangles where we believe it would be profitable to
1116 : /// tail-duplicate them all, but a local analysis would not find them.
1117 : /// There are 3 ways this can be profitable:
1118 : /// 1) The post-dominators marked 50% are actually taken 55% (This shrinks with
1119 : /// longer chains)
1120 : /// 2) The chains are statically correlated. Branch probabilities have a very
1121 : /// U-shaped distribution.
1122 : /// [http://nrs.harvard.edu/urn-3:HUL.InstRepos:24015805]
1123 : /// If the branches in a chain are likely to be from the same side of the
1124 : /// distribution as their predecessor, but are independent at runtime, this
1125 : /// transformation is profitable. (Because the cost of being wrong is a small
1126 : /// fixed cost, unlike the standard triangle layout where the cost of being
1127 : /// wrong scales with the # of triangles.)
1128 : /// 3) The chains are dynamically correlated. If the probability that a previous
1129 : /// branch was taken positively influences whether the next branch will be
1130 : /// taken
1131 : /// We believe that 2 and 3 are common enough to justify the small margin in 1.
1132 17054 : void MachineBlockPlacement::precomputeTriangleChains() {
1133 2566 : struct TriangleChain {
1134 : std::vector<MachineBasicBlock *> Edges;
1135 :
1136 : TriangleChain(MachineBasicBlock *src, MachineBasicBlock *dst)
1137 2112 : : Edges({src, dst}) {}
1138 :
1139 : void append(MachineBasicBlock *dst) {
1140 : assert(getKey()->isSuccessor(dst) &&
1141 : "Attempting to append a block that is not a successor.");
1142 454 : Edges.push_back(dst);
1143 : }
1144 :
1145 4224 : unsigned count() const { return Edges.size() - 1; }
1146 :
1147 : MachineBasicBlock *getKey() const {
1148 454 : return Edges.back();
1149 : }
1150 : };
1151 :
1152 17054 : if (TriangleChainCount == 0)
1153 0 : return;
1154 :
1155 : LLVM_DEBUG(dbgs() << "Pre-computing triangle chains.\n");
1156 : // Map from last block to the chain that contains it. This allows us to extend
1157 : // chains as we find new triangles.
1158 : DenseMap<const MachineBasicBlock *, TriangleChain> TriangleChainMap;
1159 291067 : for (MachineBasicBlock &BB : *F) {
1160 : // If BB doesn't have 2 successors, it doesn't start a triangle.
1161 274013 : if (BB.succ_size() != 2)
1162 271447 : continue;
1163 113948 : MachineBasicBlock *PDom = nullptr;
1164 309060 : for (MachineBasicBlock *Succ : BB.successors()) {
1165 214967 : if (!MPDT->dominates(Succ, &BB))
1166 : continue;
1167 19855 : PDom = Succ;
1168 19855 : break;
1169 : }
1170 : // If BB doesn't have a post-dominating successor, it doesn't form a
1171 : // triangle.
1172 113948 : if (PDom == nullptr)
1173 : continue;
1174 : // If PDom has a hint that it is low probability, skip this triangle.
1175 19855 : if (MBPI->getEdgeProbability(&BB, PDom) < BranchProbability(50, 100))
1176 : continue;
1177 : // If PDom isn't eligible for duplication, this isn't the kind of triangle
1178 : // we're looking for.
1179 11861 : if (!shouldTailDuplicate(PDom))
1180 : continue;
1181 : bool CanTailDuplicate = true;
1182 : // If PDom can't tail-duplicate into it's non-BB predecessors, then this
1183 : // isn't the kind of triangle we're looking for.
1184 9069 : for (MachineBasicBlock* Pred : PDom->predecessors()) {
1185 6503 : if (Pred == &BB)
1186 : continue;
1187 3584 : if (!TailDup.canTailDuplicate(PDom, Pred)) {
1188 : CanTailDuplicate = false;
1189 : break;
1190 : }
1191 : }
1192 : // If we can't tail-duplicate PDom to its predecessors, then skip this
1193 : // triangle.
1194 3405 : if (!CanTailDuplicate)
1195 : continue;
1196 :
1197 : // Now we have an interesting triangle. Insert it if it's not part of an
1198 : // existing chain.
1199 : // Note: This cannot be replaced with a call insert() or emplace() because
1200 : // the find key is BB, but the insert/emplace key is PDom.
1201 2566 : auto Found = TriangleChainMap.find(&BB);
1202 : // If it is, remove the chain from the map, grow it, and put it back in the
1203 : // map with the end as the new key.
1204 2566 : if (Found != TriangleChainMap.end()) {
1205 : TriangleChain Chain = std::move(Found->second);
1206 : TriangleChainMap.erase(Found);
1207 454 : Chain.append(PDom);
1208 454 : TriangleChainMap.insert(std::make_pair(Chain.getKey(), std::move(Chain)));
1209 : } else {
1210 2112 : auto InsertResult = TriangleChainMap.try_emplace(PDom, &BB, PDom);
1211 : assert(InsertResult.second && "Block seen twice.");
1212 : (void)InsertResult;
1213 : }
1214 : }
1215 :
1216 : // Iterating over a DenseMap is safe here, because the only thing in the body
1217 : // of the loop is inserting into another DenseMap (ComputedEdges).
1218 : // ComputedEdges is never iterated, so this doesn't lead to non-determinism.
1219 19166 : for (auto &ChainPair : TriangleChainMap) {
1220 : TriangleChain &Chain = ChainPair.second;
1221 : // Benchmarking has shown that due to branch correlation duplicating 2 or
1222 : // more triangles is profitable, despite the calculations assuming
1223 : // independence.
1224 2112 : if (Chain.count() < TriangleChainCount)
1225 : continue;
1226 184 : MachineBasicBlock *dst = Chain.Edges.back();
1227 : Chain.Edges.pop_back();
1228 822 : for (MachineBasicBlock *src : reverse(Chain.Edges)) {
1229 : LLVM_DEBUG(dbgs() << "Marking edge: " << getBlockName(src) << "->"
1230 : << getBlockName(dst)
1231 : << " as pre-computed based on triangles.\n");
1232 :
1233 638 : auto InsertResult = ComputedEdges.insert({src, {dst, true}});
1234 : assert(InsertResult.second && "Block seen twice.");
1235 : (void)InsertResult;
1236 :
1237 : dst = src;
1238 : }
1239 : }
1240 : }
1241 :
1242 : // When profile is not present, return the StaticLikelyProb.
1243 : // When profile is available, we need to handle the triangle-shape CFG.
1244 111027 : static BranchProbability getLayoutSuccessorProbThreshold(
1245 : const MachineBasicBlock *BB) {
1246 111027 : if (!BB->getParent()->getFunction().hasProfileData())
1247 111011 : return BranchProbability(StaticLikelyProb, 100);
1248 16 : if (BB->succ_size() == 2) {
1249 6 : const MachineBasicBlock *Succ1 = *BB->succ_begin();
1250 6 : const MachineBasicBlock *Succ2 = *(BB->succ_begin() + 1);
1251 6 : if (Succ1->isSuccessor(Succ2) || Succ2->isSuccessor(Succ1)) {
1252 : /* See case 1 below for the cost analysis. For BB->Succ to
1253 : * be taken with smaller cost, the following needs to hold:
1254 : * Prob(BB->Succ) > 2 * Prob(BB->Pred)
1255 : * So the threshold T in the calculation below
1256 : * (1-T) * Prob(BB->Succ) > T * Prob(BB->Pred)
1257 : * So T / (1 - T) = 2, Yielding T = 2/3
1258 : * Also adding user specified branch bias, we have
1259 : * T = (2/3)*(ProfileLikelyProb/50)
1260 : * = (2*ProfileLikelyProb)/150)
1261 : */
1262 4 : return BranchProbability(2 * ProfileLikelyProb, 150);
1263 : }
1264 : }
1265 12 : return BranchProbability(ProfileLikelyProb, 100);
1266 : }
1267 :
1268 : /// Checks to see if the layout candidate block \p Succ has a better layout
1269 : /// predecessor than \c BB. If yes, returns true.
1270 : /// \p SuccProb: The probability adjusted for only remaining blocks.
1271 : /// Only used for logging
1272 : /// \p RealSuccProb: The un-adjusted probability.
1273 : /// \p Chain: The chain that BB belongs to and Succ is being considered for.
1274 : /// \p BlockFilter: if non-null, the set of blocks that make up the loop being
1275 : /// considered
1276 0 : bool MachineBlockPlacement::hasBetterLayoutPredecessor(
1277 : const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
1278 : const BlockChain &SuccChain, BranchProbability SuccProb,
1279 : BranchProbability RealSuccProb, const BlockChain &Chain,
1280 : const BlockFilterSet *BlockFilter) {
1281 :
1282 : // There isn't a better layout when there are no unscheduled predecessors.
1283 0 : if (SuccChain.UnscheduledPredecessors == 0)
1284 0 : return false;
1285 :
1286 : // There are two basic scenarios here:
1287 : // -------------------------------------
1288 : // Case 1: triangular shape CFG (if-then):
1289 : // BB
1290 : // | \
1291 : // | \
1292 : // | Pred
1293 : // | /
1294 : // Succ
1295 : // In this case, we are evaluating whether to select edge -> Succ, e.g.
1296 : // set Succ as the layout successor of BB. Picking Succ as BB's
1297 : // successor breaks the CFG constraints (FIXME: define these constraints).
1298 : // With this layout, Pred BB
1299 : // is forced to be outlined, so the overall cost will be cost of the
1300 : // branch taken from BB to Pred, plus the cost of back taken branch
1301 : // from Pred to Succ, as well as the additional cost associated
1302 : // with the needed unconditional jump instruction from Pred To Succ.
1303 :
1304 : // The cost of the topological order layout is the taken branch cost
1305 : // from BB to Succ, so to make BB->Succ a viable candidate, the following
1306 : // must hold:
1307 : // 2 * freq(BB->Pred) * taken_branch_cost + unconditional_jump_cost
1308 : // < freq(BB->Succ) * taken_branch_cost.
1309 : // Ignoring unconditional jump cost, we get
1310 : // freq(BB->Succ) > 2 * freq(BB->Pred), i.e.,
1311 : // prob(BB->Succ) > 2 * prob(BB->Pred)
1312 : //
1313 : // When real profile data is available, we can precisely compute the
1314 : // probability threshold that is needed for edge BB->Succ to be considered.
1315 : // Without profile data, the heuristic requires the branch bias to be
1316 : // a lot larger to make sure the signal is very strong (e.g. 80% default).
1317 : // -----------------------------------------------------------------
1318 : // Case 2: diamond like CFG (if-then-else):
1319 : // S
1320 : // / \
1321 : // | \
1322 : // BB Pred
1323 : // \ /
1324 : // Succ
1325 : // ..
1326 : //
1327 : // The current block is BB and edge BB->Succ is now being evaluated.
1328 : // Note that edge S->BB was previously already selected because
1329 : // prob(S->BB) > prob(S->Pred).
1330 : // At this point, 2 blocks can be placed after BB: Pred or Succ. If we
1331 : // choose Pred, we will have a topological ordering as shown on the left
1332 : // in the picture below. If we choose Succ, we have the solution as shown
1333 : // on the right:
1334 : //
1335 : // topo-order:
1336 : //
1337 : // S----- ---S
1338 : // | | | |
1339 : // ---BB | | BB
1340 : // | | | |
1341 : // | Pred-- | Succ--
1342 : // | | | |
1343 : // ---Succ ---Pred--
1344 : //
1345 : // cost = freq(S->Pred) + freq(BB->Succ) cost = 2 * freq (S->Pred)
1346 : // = freq(S->Pred) + freq(S->BB)
1347 : //
1348 : // If we have profile data (i.e, branch probabilities can be trusted), the
1349 : // cost (number of taken branches) with layout S->BB->Succ->Pred is 2 *
1350 : // freq(S->Pred) while the cost of topo order is freq(S->Pred) + freq(S->BB).
1351 : // We know Prob(S->BB) > Prob(S->Pred), so freq(S->BB) > freq(S->Pred), which
1352 : // means the cost of topological order is greater.
1353 : // When profile data is not available, however, we need to be more
1354 : // conservative. If the branch prediction is wrong, breaking the topo-order
1355 : // will actually yield a layout with large cost. For this reason, we need
1356 : // strong biased branch at block S with Prob(S->BB) in order to select
1357 : // BB->Succ. This is equivalent to looking the CFG backward with backward
1358 : // edge: Prob(Succ->BB) needs to >= HotProb in order to be selected (without
1359 : // profile data).
1360 : // --------------------------------------------------------------------------
1361 : // Case 3: forked diamond
1362 : // S
1363 : // / \
1364 : // / \
1365 : // BB Pred
1366 : // | \ / |
1367 : // | \ / |
1368 : // | X |
1369 : // | / \ |
1370 : // | / \ |
1371 : // S1 S2
1372 : //
1373 : // The current block is BB and edge BB->S1 is now being evaluated.
1374 : // As above S->BB was already selected because
1375 : // prob(S->BB) > prob(S->Pred). Assume that prob(BB->S1) >= prob(BB->S2).
1376 : //
1377 : // topo-order:
1378 : //
1379 : // S-------| ---S
1380 : // | | | |
1381 : // ---BB | | BB
1382 : // | | | |
1383 : // | Pred----| | S1----
1384 : // | | | |
1385 : // --(S1 or S2) ---Pred--
1386 : // |
1387 : // S2
1388 : //
1389 : // topo-cost = freq(S->Pred) + freq(BB->S1) + freq(BB->S2)
1390 : // + min(freq(Pred->S1), freq(Pred->S2))
1391 : // Non-topo-order cost:
1392 : // non-topo-cost = 2 * freq(S->Pred) + freq(BB->S2).
1393 : // To be conservative, we can assume that min(freq(Pred->S1), freq(Pred->S2))
1394 : // is 0. Then the non topo layout is better when
1395 : // freq(S->Pred) < freq(BB->S1).
1396 : // This is exactly what is checked below.
1397 : // Note there are other shapes that apply (Pred may not be a single block,
1398 : // but they all fit this general pattern.)
1399 0 : BranchProbability HotProb = getLayoutSuccessorProbThreshold(BB);
1400 :
1401 : // Make sure that a hot successor doesn't have a globally more
1402 : // important predecessor.
1403 0 : BlockFrequency CandidateEdgeFreq = MBFI->getBlockFreq(BB) * RealSuccProb;
1404 : bool BadCFGConflict = false;
1405 :
1406 0 : for (MachineBasicBlock *Pred : Succ->predecessors()) {
1407 0 : if (Pred == Succ || BlockToChain[Pred] == &SuccChain ||
1408 0 : (BlockFilter && !BlockFilter->count(Pred)) ||
1409 0 : BlockToChain[Pred] == &Chain ||
1410 : // This check is redundant except for look ahead. This function is
1411 : // called for lookahead by isProfitableToTailDup when BB hasn't been
1412 : // placed yet.
1413 : (Pred == BB))
1414 0 : continue;
1415 : // Do backward checking.
1416 : // For all cases above, we need a backward checking to filter out edges that
1417 : // are not 'strongly' biased.
1418 : // BB Pred
1419 : // \ /
1420 : // Succ
1421 : // We select edge BB->Succ if
1422 : // freq(BB->Succ) > freq(Succ) * HotProb
1423 : // i.e. freq(BB->Succ) > freq(BB->Succ) * HotProb + freq(Pred->Succ) *
1424 : // HotProb
1425 : // i.e. freq((BB->Succ) * (1 - HotProb) > freq(Pred->Succ) * HotProb
1426 : // Case 1 is covered too, because the first equation reduces to:
1427 : // prob(BB->Succ) > HotProb. (freq(Succ) = freq(BB) for a triangle)
1428 : BlockFrequency PredEdgeFreq =
1429 0 : MBFI->getBlockFreq(Pred) * MBPI->getEdgeProbability(Pred, Succ);
1430 0 : if (PredEdgeFreq * HotProb >= CandidateEdgeFreq * HotProb.getCompl()) {
1431 : BadCFGConflict = true;
1432 0 : break;
1433 : }
1434 : }
1435 :
1436 : if (BadCFGConflict) {
1437 : LLVM_DEBUG(dbgs() << " Not a candidate: " << getBlockName(Succ) << " -> "
1438 : << SuccProb << " (prob) (non-cold CFG conflict)\n");
1439 0 : return true;
1440 : }
1441 :
1442 : return false;
1443 : }
1444 :
1445 : /// Select the best successor for a block.
1446 : ///
1447 : /// This looks across all successors of a particular block and attempts to
1448 : /// select the "best" one to be the layout successor. It only considers direct
1449 : /// successors which also pass the block filter. It will attempt to avoid
1450 : /// breaking CFG structure, but cave and break such structures in the case of
1451 : /// very hot successor edges.
1452 : ///
1453 : /// \returns The best successor block found, or null if none are viable, along
1454 : /// with a boolean indicating if tail duplication is necessary.
1455 : MachineBlockPlacement::BlockAndTailDupResult
1456 357638 : MachineBlockPlacement::selectBestSuccessor(
1457 : const MachineBasicBlock *BB, const BlockChain &Chain,
1458 : const BlockFilterSet *BlockFilter) {
1459 357638 : const BranchProbability HotProb(StaticLikelyProb, 100);
1460 :
1461 : BlockAndTailDupResult BestSucc = { nullptr, false };
1462 : auto BestProb = BranchProbability::getZero();
1463 :
1464 : SmallVector<MachineBasicBlock *, 4> Successors;
1465 : auto AdjustedSumProb =
1466 357638 : collectViableSuccessors(BB, Chain, BlockFilter, Successors);
1467 :
1468 : LLVM_DEBUG(dbgs() << "Selecting best successor for: " << getBlockName(BB)
1469 : << "\n");
1470 :
1471 : // if we already precomputed the best successor for BB, return that if still
1472 : // applicable.
1473 357638 : auto FoundEdge = ComputedEdges.find(BB);
1474 357638 : if (FoundEdge != ComputedEdges.end()) {
1475 1114 : MachineBasicBlock *Succ = FoundEdge->second.BB;
1476 : ComputedEdges.erase(FoundEdge);
1477 1114 : BlockChain *SuccChain = BlockToChain[Succ];
1478 2320 : if (BB->isSuccessor(Succ) && (!BlockFilter || BlockFilter->count(Succ)) &&
1479 2221 : SuccChain != &Chain && Succ == *SuccChain->begin())
1480 1107 : return FoundEdge->second;
1481 : }
1482 :
1483 : // if BB is part of a trellis, Use the trellis to determine the optimal
1484 : // fallthrough edges
1485 356531 : if (isTrellis(BB, Successors, Chain, BlockFilter))
1486 : return getBestTrellisSuccessor(BB, Successors, AdjustedSumProb, Chain,
1487 583 : BlockFilter);
1488 :
1489 : // For blocks with CFG violations, we may be able to lay them out anyway with
1490 : // tail-duplication. We keep this vector so we can perform the probability
1491 : // calculations the minimum number of times.
1492 : SmallVector<std::tuple<BranchProbability, MachineBasicBlock *>, 4>
1493 : DupCandidates;
1494 734506 : for (MachineBasicBlock *Succ : Successors) {
1495 378558 : auto RealSuccProb = MBPI->getEdgeProbability(BB, Succ);
1496 : BranchProbability SuccProb =
1497 : getAdjustedProbability(RealSuccProb, AdjustedSumProb);
1498 :
1499 378558 : BlockChain &SuccChain = *BlockToChain[Succ];
1500 : // Skip the edge \c BB->Succ if block \c Succ has a better layout
1501 : // predecessor that yields lower global cost.
1502 378558 : if (hasBetterLayoutPredecessor(BB, Succ, SuccChain, SuccProb, RealSuccProb,
1503 : Chain, BlockFilter)) {
1504 : // If tail duplication would make Succ profitable, place it.
1505 105636 : if (allowTailDupPlacement() && shouldTailDuplicate(Succ))
1506 6379 : DupCandidates.push_back(std::make_tuple(SuccProb, Succ));
1507 139354 : continue;
1508 : }
1509 :
1510 : LLVM_DEBUG(
1511 : dbgs() << " Candidate: " << getBlockName(Succ)
1512 : << ", probability: " << SuccProb
1513 : << (SuccChain.UnscheduledPredecessors != 0 ? " (CFG break)" : "")
1514 : << "\n");
1515 :
1516 272786 : if (BestSucc.BB && BestProb >= SuccProb) {
1517 : LLVM_DEBUG(dbgs() << " Not the best candidate, continuing\n");
1518 : continue;
1519 : }
1520 :
1521 : LLVM_DEBUG(dbgs() << " Setting it as best candidate\n");
1522 : BestSucc.BB = Succ;
1523 : BestProb = SuccProb;
1524 : }
1525 : // Handle the tail duplication candidates in order of decreasing probability.
1526 : // Stop at the first one that is profitable. Also stop if they are less
1527 : // profitable than BestSucc. Position is important because we preserve it and
1528 : // prefer first best match. Here we aren't comparing in order, so we capture
1529 : // the position instead.
1530 711896 : if (DupCandidates.size() != 0) {
1531 : auto cmp =
1532 : [](const std::tuple<BranchProbability, MachineBasicBlock *> &a,
1533 : const std::tuple<BranchProbability, MachineBasicBlock *> &b) {
1534 : return std::get<0>(a) > std::get<0>(b);
1535 : };
1536 : std::stable_sort(DupCandidates.begin(), DupCandidates.end(), cmp);
1537 : }
1538 359421 : for(auto &Tup : DupCandidates) {
1539 : BranchProbability DupProb;
1540 : MachineBasicBlock *Succ;
1541 : std::tie(DupProb, Succ) = Tup;
1542 6348 : if (DupProb < BestProb)
1543 : break;
1544 4345 : if (canTailDuplicateUnplacedPreds(BB, Succ, Chain, BlockFilter)
1545 4345 : && (isProfitableToTailDup(BB, Succ, BestProb, Chain, BlockFilter))) {
1546 : LLVM_DEBUG(dbgs() << " Candidate: " << getBlockName(Succ)
1547 : << ", probability: " << DupProb
1548 : << " (Tail Duplicate)\n");
1549 : BestSucc.BB = Succ;
1550 : BestSucc.ShouldTailDup = true;
1551 : break;
1552 : }
1553 : }
1554 :
1555 : if (BestSucc.BB)
1556 : LLVM_DEBUG(dbgs() << " Selected: " << getBlockName(BestSucc.BB) << "\n");
1557 :
1558 355948 : return BestSucc;
1559 : }
1560 :
1561 : /// Select the best block from a worklist.
1562 : ///
1563 : /// This looks through the provided worklist as a list of candidate basic
1564 : /// blocks and select the most profitable one to place. The definition of
1565 : /// profitable only really makes sense in the context of a loop. This returns
1566 : /// the most frequently visited block in the worklist, which in the case of
1567 : /// a loop, is the one most desirable to be physically close to the rest of the
1568 : /// loop body in order to improve i-cache behavior.
1569 : ///
1570 : /// \returns The best block found, or null if none are viable.
1571 215579 : MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock(
1572 : const BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList) {
1573 : // Once we need to walk the worklist looking for a candidate, cleanup the
1574 : // worklist of already placed entries.
1575 : // FIXME: If this shows up on profiles, it could be folded (at the cost of
1576 : // some code complexity) into the loop below.
1577 215579 : WorkList.erase(llvm::remove_if(WorkList,
1578 : [&](MachineBasicBlock *BB) {
1579 0 : return BlockToChain.lookup(BB) == &Chain;
1580 : }),
1581 : WorkList.end());
1582 :
1583 215579 : if (WorkList.empty())
1584 : return nullptr;
1585 :
1586 109058 : bool IsEHPad = WorkList[0]->isEHPad();
1587 :
1588 : MachineBasicBlock *BestBlock = nullptr;
1589 : BlockFrequency BestFreq;
1590 2255383 : for (MachineBasicBlock *MBB : WorkList) {
1591 : assert(MBB->isEHPad() == IsEHPad &&
1592 : "EHPad mismatch between block and work list.");
1593 :
1594 2146325 : BlockChain &SuccChain = *BlockToChain[MBB];
1595 2146325 : if (&SuccChain == &Chain)
1596 1417905 : continue;
1597 :
1598 : assert(SuccChain.UnscheduledPredecessors == 0 &&
1599 : "Found CFG-violating block");
1600 :
1601 2146325 : BlockFrequency CandidateFreq = MBFI->getBlockFreq(MBB);
1602 : LLVM_DEBUG(dbgs() << " " << getBlockName(MBB) << " -> ";
1603 : MBFI->printBlockFreq(dbgs(), CandidateFreq) << " (freq)\n");
1604 :
1605 : // For ehpad, we layout the least probable first as to avoid jumping back
1606 : // from least probable landingpads to more probable ones.
1607 : //
1608 : // FIXME: Using probability is probably (!) not the best way to achieve
1609 : // this. We should probably have a more principled approach to layout
1610 : // cleanup code.
1611 : //
1612 : // The goal is to get:
1613 : //
1614 : // +--------------------------+
1615 : // | V
1616 : // InnerLp -> InnerCleanup OuterLp -> OuterCleanup -> Resume
1617 : //
1618 : // Rather than:
1619 : //
1620 : // +-------------------------------------+
1621 : // V |
1622 : // OuterLp -> OuterCleanup -> Resume InnerLp -> InnerCleanup
1623 2146325 : if (BestBlock && (IsEHPad ^ (BestFreq >= CandidateFreq)))
1624 : continue;
1625 :
1626 : BestBlock = MBB;
1627 728420 : BestFreq = CandidateFreq;
1628 : }
1629 :
1630 : return BestBlock;
1631 : }
1632 :
1633 : /// Retrieve the first unplaced basic block.
1634 : ///
1635 : /// This routine is called when we are unable to use the CFG to walk through
1636 : /// all of the basic blocks and form a chain due to unnatural loops in the CFG.
1637 : /// We walk through the function's blocks in order, starting from the
1638 : /// LastUnplacedBlockIt. We update this iterator on each call to avoid
1639 : /// re-scanning the entire sequence on repeated calls to this routine.
1640 27864 : MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock(
1641 : const BlockChain &PlacedChain,
1642 : MachineFunction::iterator &PrevUnplacedBlockIt,
1643 : const BlockFilterSet *BlockFilter) {
1644 988861 : for (MachineFunction::iterator I = PrevUnplacedBlockIt, E = F->end(); I != E;
1645 : ++I) {
1646 1573698 : if (BlockFilter && !BlockFilter->count(&*I))
1647 : continue;
1648 409621 : if (BlockToChain[&*I] != &PlacedChain) {
1649 236 : PrevUnplacedBlockIt = I;
1650 : // Now select the head of the chain to which the unplaced block belongs
1651 : // as the block to place. This will force the entire chain to be placed,
1652 : // and satisfies the requirements of merging chains.
1653 236 : return *BlockToChain[&*I]->begin();
1654 : }
1655 : }
1656 : return nullptr;
1657 : }
1658 :
1659 410097 : void MachineBlockPlacement::fillWorkLists(
1660 : const MachineBasicBlock *MBB,
1661 : SmallPtrSetImpl<BlockChain *> &UpdatedPreds,
1662 : const BlockFilterSet *BlockFilter = nullptr) {
1663 410097 : BlockChain &Chain = *BlockToChain[MBB];
1664 410097 : if (!UpdatedPreds.insert(&Chain).second)
1665 392193 : return;
1666 :
1667 : assert(
1668 : Chain.UnscheduledPredecessors == 0 &&
1669 : "Attempting to place block with unscheduled predecessors in worklist.");
1670 748247 : for (MachineBasicBlock *ChainBB : Chain) {
1671 : assert(BlockToChain[ChainBB] == &Chain &&
1672 : "Block in chain doesn't match BlockToChain map.");
1673 937014 : for (MachineBasicBlock *Pred : ChainBB->predecessors()) {
1674 605454 : if (BlockFilter && !BlockFilter->count(Pred))
1675 : continue;
1676 535465 : if (BlockToChain[Pred] == &Chain)
1677 : continue;
1678 456203 : ++Chain.UnscheduledPredecessors;
1679 : }
1680 : }
1681 :
1682 347913 : if (Chain.UnscheduledPredecessors != 0)
1683 : return;
1684 :
1685 17904 : MachineBasicBlock *BB = *Chain.begin();
1686 17904 : if (BB->isEHPad())
1687 0 : EHPadWorkList.push_back(BB);
1688 : else
1689 17904 : BlockWorkList.push_back(BB);
1690 : }
1691 :
1692 27628 : void MachineBlockPlacement::buildChain(
1693 : const MachineBasicBlock *HeadBB, BlockChain &Chain,
1694 : BlockFilterSet *BlockFilter) {
1695 : assert(HeadBB && "BB must not be null.\n");
1696 : assert(BlockToChain[HeadBB] == &Chain && "BlockToChainMap mis-match.\n");
1697 27628 : MachineFunction::iterator PrevUnplacedBlockIt = F->begin();
1698 :
1699 : const MachineBasicBlock *LoopHeaderBB = HeadBB;
1700 27628 : markChainSuccessors(Chain, LoopHeaderBB, BlockFilter);
1701 27628 : MachineBasicBlock *BB = *std::prev(Chain.end());
1702 : while (true) {
1703 : assert(BB && "null block found at end of chain in loop.");
1704 : assert(BlockToChain[BB] == &Chain && "BlockToChainMap mis-match in loop.");
1705 : assert(*std::prev(Chain.end()) == BB && "BB Not found at end of chain.");
1706 :
1707 :
1708 : // Look for the best viable successor if there is one to place immediately
1709 : // after this block.
1710 357638 : auto Result = selectBestSuccessor(BB, Chain, BlockFilter);
1711 357638 : MachineBasicBlock* BestSucc = Result.BB;
1712 357638 : bool ShouldTailDup = Result.ShouldTailDup;
1713 : if (allowTailDupPlacement())
1714 554819 : ShouldTailDup |= (BestSucc && shouldTailDuplicate(BestSucc));
1715 :
1716 : // If an immediate successor isn't available, look for the best viable
1717 : // block among those we've identified as not violating the loop's CFG at
1718 : // this point. This won't be a fallthrough, but it will increase locality.
1719 357638 : if (!BestSucc)
1720 136922 : BestSucc = selectBestCandidateBlock(Chain, BlockWorkList);
1721 357638 : if (!BestSucc)
1722 78657 : BestSucc = selectBestCandidateBlock(Chain, EHPadWorkList);
1723 :
1724 357638 : if (!BestSucc) {
1725 27864 : BestSucc = getFirstUnplacedBlock(Chain, PrevUnplacedBlockIt, BlockFilter);
1726 27864 : if (!BestSucc)
1727 : break;
1728 :
1729 : LLVM_DEBUG(dbgs() << "Unnatural loop CFG detected, forcibly merging the "
1730 : "layout successor until the CFG reduces\n");
1731 : }
1732 :
1733 : // Placement may have changed tail duplication opportunities.
1734 : // Check for that now.
1735 329520 : if (allowTailDupPlacement() && BestSucc && ShouldTailDup) {
1736 : // If the chosen successor was duplicated into all its predecessors,
1737 : // don't bother laying it out, just go round the loop again with BB as
1738 : // the chain end.
1739 22424 : if (repeatedlyTailDuplicateBlock(BestSucc, BB, LoopHeaderBB, Chain,
1740 : BlockFilter, PrevUnplacedBlockIt))
1741 712 : continue;
1742 : }
1743 :
1744 : // Place this block, updating the datastructures to reflect its placement.
1745 329298 : BlockChain &SuccChain = *BlockToChain[BestSucc];
1746 : // Zero out UnscheduledPredecessors for the successor we're about to merge in case
1747 : // we selected a successor that didn't fit naturally into the CFG.
1748 329298 : SuccChain.UnscheduledPredecessors = 0;
1749 : LLVM_DEBUG(dbgs() << "Merging from " << getBlockName(BB) << " to "
1750 : << getBlockName(BestSucc) << "\n");
1751 329298 : markChainSuccessors(SuccChain, LoopHeaderBB, BlockFilter);
1752 329298 : Chain.merge(BestSucc, &SuccChain);
1753 329298 : BB = *std::prev(Chain.end());
1754 : }
1755 :
1756 : LLVM_DEBUG(dbgs() << "Finished forming chain for header block "
1757 : << getBlockName(*Chain.begin()) << "\n");
1758 27628 : }
1759 :
1760 : /// Find the best loop top block for layout.
1761 : ///
1762 : /// Look for a block which is strictly better than the loop header for laying
1763 : /// out at the top of the loop. This looks for one and only one pattern:
1764 : /// a latch block with no conditional exit. This block will cause a conditional
1765 : /// jump around it or will be the bottom of the loop if we lay it out in place,
1766 : /// but if it it doesn't end up at the bottom of the loop for any reason,
1767 : /// rotation alone won't fix it. Because such a block will always result in an
1768 : /// unconditional jump (for the backedge) rotating it in front of the loop
1769 : /// header is always profitable.
1770 : MachineBasicBlock *
1771 9721 : MachineBlockPlacement::findBestLoopTop(const MachineLoop &L,
1772 : const BlockFilterSet &LoopBlockSet) {
1773 : // Placing the latch block before the header may introduce an extra branch
1774 : // that skips this block the first time the loop is executed, which we want
1775 : // to avoid when optimising for size.
1776 : // FIXME: in theory there is a case that does not introduce a new branch,
1777 : // i.e. when the layout predecessor does not fallthrough to the loop header.
1778 : // In practice this never happens though: there always seems to be a preheader
1779 : // that can fallthrough and that is also placed before the header.
1780 9721 : if (F->getFunction().optForSize())
1781 103 : return L.getHeader();
1782 :
1783 : // Check that the header hasn't been fused with a preheader block due to
1784 : // crazy branches. If it has, we need to start with the header at the top to
1785 : // prevent pulling the preheader into the loop body.
1786 9618 : BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
1787 19236 : if (!LoopBlockSet.count(*HeaderChain.begin()))
1788 10 : return L.getHeader();
1789 :
1790 : LLVM_DEBUG(dbgs() << "Finding best loop top for: "
1791 : << getBlockName(L.getHeader()) << "\n");
1792 :
1793 : BlockFrequency BestPredFreq;
1794 : MachineBasicBlock *BestPred = nullptr;
1795 29947 : for (MachineBasicBlock *Pred : L.getHeader()->predecessors()) {
1796 20339 : if (!LoopBlockSet.count(Pred))
1797 : continue;
1798 : LLVM_DEBUG(dbgs() << " header pred: " << getBlockName(Pred) << ", has "
1799 : << Pred->succ_size() << " successors, ";
1800 : MBFI->printBlockFreq(dbgs(), Pred) << " freq\n");
1801 10120 : if (Pred->succ_size() > 1)
1802 : continue;
1803 :
1804 1566 : BlockFrequency PredFreq = MBFI->getBlockFreq(Pred);
1805 1566 : if (!BestPred || PredFreq > BestPredFreq ||
1806 55 : (!(PredFreq < BestPredFreq) &&
1807 55 : Pred->isLayoutSuccessor(L.getHeader()))) {
1808 : BestPred = Pred;
1809 : BestPredFreq = PredFreq;
1810 : }
1811 : }
1812 :
1813 : // If no direct predecessor is fine, just use the loop header.
1814 9608 : if (!BestPred) {
1815 : LLVM_DEBUG(dbgs() << " final top unchanged\n");
1816 8151 : return L.getHeader();
1817 : }
1818 :
1819 : // Walk backwards through any straight line of predecessors.
1820 970 : while (BestPred->pred_size() == 1 &&
1821 1459 : (*BestPred->pred_begin())->succ_size() == 1 &&
1822 : *BestPred->pred_begin() != L.getHeader())
1823 : BestPred = *BestPred->pred_begin();
1824 :
1825 : LLVM_DEBUG(dbgs() << " final top: " << getBlockName(BestPred) << "\n");
1826 : return BestPred;
1827 : }
1828 :
1829 : /// Find the best loop exiting block for layout.
1830 : ///
1831 : /// This routine implements the logic to analyze the loop looking for the best
1832 : /// block to layout at the top of the loop. Typically this is done to maximize
1833 : /// fallthrough opportunities.
1834 : MachineBasicBlock *
1835 8541 : MachineBlockPlacement::findBestLoopExit(const MachineLoop &L,
1836 : const BlockFilterSet &LoopBlockSet) {
1837 : // We don't want to layout the loop linearly in all cases. If the loop header
1838 : // is just a normal basic block in the loop, we want to look for what block
1839 : // within the loop is the best one to layout at the top. However, if the loop
1840 : // header has be pre-merged into a chain due to predecessors not having
1841 : // analyzable branches, *and* the predecessor it is merged with is *not* part
1842 : // of the loop, rotating the header into the middle of the loop will create
1843 : // a non-contiguous range of blocks which is Very Bad. So start with the
1844 : // header and only rotate if safe.
1845 8541 : BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
1846 17082 : if (!LoopBlockSet.count(*HeaderChain.begin()))
1847 : return nullptr;
1848 :
1849 : BlockFrequency BestExitEdgeFreq;
1850 : unsigned BestExitLoopDepth = 0;
1851 : MachineBasicBlock *ExitingBB = nullptr;
1852 : // If there are exits to outer loops, loop rotation can severely limit
1853 : // fallthrough opportunities unless it selects such an exit. Keep a set of
1854 : // blocks where rotating to exit with that block will reach an outer loop.
1855 : SmallPtrSet<MachineBasicBlock *, 4> BlocksExitingToOuterLoop;
1856 :
1857 : LLVM_DEBUG(dbgs() << "Finding best loop exit for: "
1858 : << getBlockName(L.getHeader()) << "\n");
1859 54335 : for (MachineBasicBlock *MBB : L.getBlocks()) {
1860 45805 : BlockChain &Chain = *BlockToChain[MBB];
1861 : // Ensure that this block is at the end of a chain; otherwise it could be
1862 : // mid-way through an inner loop or a successor of an unanalyzable branch.
1863 45805 : if (MBB != *std::prev(Chain.end()))
1864 : continue;
1865 :
1866 : // Now walk the successors. We need to establish whether this has a viable
1867 : // exiting successor and whether it has a viable non-exiting successor.
1868 : // We store the old exiting state and restore it if a viable looping
1869 : // successor isn't found.
1870 : MachineBasicBlock *OldExitingBB = ExitingBB;
1871 39411 : BlockFrequency OldBestExitEdgeFreq = BestExitEdgeFreq;
1872 : bool HasLoopingSucc = false;
1873 107387 : for (MachineBasicBlock *Succ : MBB->successors()) {
1874 67976 : if (Succ->isEHPad())
1875 56313 : continue;
1876 61070 : if (Succ == MBB)
1877 : continue;
1878 55391 : BlockChain &SuccChain = *BlockToChain[Succ];
1879 : // Don't split chains, either this chain or the successor's chain.
1880 55391 : if (&Chain == &SuccChain) {
1881 : LLVM_DEBUG(dbgs() << " exiting: " << getBlockName(MBB) << " -> "
1882 : << getBlockName(Succ) << " (chain conflict)\n");
1883 : continue;
1884 : }
1885 :
1886 54726 : auto SuccProb = MBPI->getEdgeProbability(MBB, Succ);
1887 54726 : if (LoopBlockSet.count(Succ)) {
1888 : LLVM_DEBUG(dbgs() << " looping: " << getBlockName(MBB) << " -> "
1889 : << getBlockName(Succ) << " (" << SuccProb << ")\n");
1890 : HasLoopingSucc = true;
1891 : continue;
1892 : }
1893 :
1894 : unsigned SuccLoopDepth = 0;
1895 13922 : if (MachineLoop *ExitLoop = MLI->getLoopFor(Succ)) {
1896 2259 : SuccLoopDepth = ExitLoop->getLoopDepth();
1897 4518 : if (ExitLoop->contains(&L))
1898 1607 : BlocksExitingToOuterLoop.insert(MBB);
1899 : }
1900 :
1901 11663 : BlockFrequency ExitEdgeFreq = MBFI->getBlockFreq(MBB) * SuccProb;
1902 : LLVM_DEBUG(dbgs() << " exiting: " << getBlockName(MBB) << " -> "
1903 : << getBlockName(Succ) << " [L:" << SuccLoopDepth
1904 : << "] (";
1905 : MBFI->printBlockFreq(dbgs(), ExitEdgeFreq) << ")\n");
1906 : // Note that we bias this toward an existing layout successor to retain
1907 : // incoming order in the absence of better information. The exit must have
1908 : // a frequency higher than the current exit before we consider breaking
1909 : // the layout.
1910 11663 : BranchProbability Bias(100 - ExitBlockBias, 100);
1911 2470 : if (!ExitingBB || SuccLoopDepth > BestExitLoopDepth ||
1912 13731 : ExitEdgeFreq > BestExitEdgeFreq ||
1913 3110 : (MBB->isLayoutSuccessor(Succ) &&
1914 1042 : !(ExitEdgeFreq < BestExitEdgeFreq * Bias))) {
1915 9700 : BestExitEdgeFreq = ExitEdgeFreq;
1916 : ExitingBB = MBB;
1917 : }
1918 : }
1919 :
1920 39411 : if (!HasLoopingSucc) {
1921 : // Restore the old exiting state, no viable looping successor was found.
1922 : ExitingBB = OldExitingBB;
1923 5302 : BestExitEdgeFreq = OldBestExitEdgeFreq;
1924 : }
1925 : }
1926 : // Without a candidate exiting block or with only a single block in the
1927 : // loop, just use the loop header to layout the loop.
1928 8530 : if (!ExitingBB) {
1929 : LLVM_DEBUG(
1930 : dbgs() << " No other candidate exit blocks, using loop header\n");
1931 : return nullptr;
1932 : }
1933 3292 : if (L.getNumBlocks() == 1) {
1934 : LLVM_DEBUG(dbgs() << " Loop has 1 block, using loop header as exit\n");
1935 : return nullptr;
1936 : }
1937 :
1938 : // Also, if we have exit blocks which lead to outer loops but didn't select
1939 : // one of them as the exiting block we are rotating toward, disable loop
1940 : // rotation altogether.
1941 4001 : if (!BlocksExitingToOuterLoop.empty() &&
1942 709 : !BlocksExitingToOuterLoop.count(ExitingBB))
1943 7 : return nullptr;
1944 :
1945 : LLVM_DEBUG(dbgs() << " Best exiting block: " << getBlockName(ExitingBB)
1946 : << "\n");
1947 : return ExitingBB;
1948 : }
1949 :
1950 : /// Attempt to rotate an exiting block to the bottom of the loop.
1951 : ///
1952 : /// Once we have built a chain, try to rotate it to line up the hot exit block
1953 : /// with fallthrough out of the loop if doing so doesn't introduce unnecessary
1954 : /// branches. For example, if the loop has fallthrough into its header and out
1955 : /// of its bottom already, don't rotate it.
1956 9721 : void MachineBlockPlacement::rotateLoop(BlockChain &LoopChain,
1957 : const MachineBasicBlock *ExitingBB,
1958 : const BlockFilterSet &LoopBlockSet) {
1959 9721 : if (!ExitingBB)
1960 : return;
1961 :
1962 3276 : MachineBasicBlock *Top = *LoopChain.begin();
1963 3276 : MachineBasicBlock *Bottom = *std::prev(LoopChain.end());
1964 :
1965 : // If ExitingBB is already the last one in a chain then nothing to do.
1966 3276 : if (Bottom == ExitingBB)
1967 : return;
1968 :
1969 : bool ViableTopFallthrough = false;
1970 1178 : for (MachineBasicBlock *Pred : Top->predecessors()) {
1971 1178 : BlockChain *PredChain = BlockToChain[Pred];
1972 1178 : if (!LoopBlockSet.count(Pred) &&
1973 1084 : (!PredChain || Pred == *std::prev(PredChain->end()))) {
1974 : ViableTopFallthrough = true;
1975 : break;
1976 : }
1977 : }
1978 :
1979 : // If the header has viable fallthrough, check whether the current loop
1980 : // bottom is a viable exiting block. If so, bail out as rotating will
1981 : // introduce an unnecessary branch.
1982 1084 : if (ViableTopFallthrough) {
1983 1907 : for (MachineBasicBlock *Succ : Bottom->successors()) {
1984 1776 : BlockChain *SuccChain = BlockToChain[Succ];
1985 1776 : if (!LoopBlockSet.count(Succ) &&
1986 953 : (!SuccChain || Succ == *SuccChain->begin()))
1987 : return;
1988 : }
1989 : }
1990 :
1991 : BlockChain::iterator ExitIt = llvm::find(LoopChain, ExitingBB);
1992 131 : if (ExitIt == LoopChain.end())
1993 : return;
1994 :
1995 : // Rotating a loop exit to the bottom when there is a fallthrough to top
1996 : // trades the entry fallthrough for an exit fallthrough.
1997 : // If there is no bottom->top edge, but the chosen exit block does have
1998 : // a fallthrough, we break that fallthrough for nothing in return.
1999 :
2000 : // Let's consider an example. We have a built chain of basic blocks
2001 : // B1, B2, ..., Bn, where Bk is a ExitingBB - chosen exit block.
2002 : // By doing a rotation we get
2003 : // Bk+1, ..., Bn, B1, ..., Bk
2004 : // Break of fallthrough to B1 is compensated by a fallthrough from Bk.
2005 : // If we had a fallthrough Bk -> Bk+1 it is broken now.
2006 : // It might be compensated by fallthrough Bn -> B1.
2007 : // So we have a condition to avoid creation of extra branch by loop rotation.
2008 : // All below must be true to avoid loop rotation:
2009 : // If there is a fallthrough to top (B1)
2010 : // There was fallthrough from chosen exit block (Bk) to next one (Bk+1)
2011 : // There is no fallthrough from bottom (Bn) to top (B1).
2012 : // Please note that there is no exit fallthrough from Bn because we checked it
2013 : // above.
2014 131 : if (ViableTopFallthrough) {
2015 : assert(std::next(ExitIt) != LoopChain.end() &&
2016 : "Exit should not be last BB");
2017 131 : MachineBasicBlock *NextBlockInChain = *std::next(ExitIt);
2018 131 : if (ExitingBB->isSuccessor(NextBlockInChain))
2019 44 : if (!Bottom->isSuccessor(Top))
2020 : return;
2021 : }
2022 :
2023 : LLVM_DEBUG(dbgs() << "Rotating loop to put exit " << getBlockName(ExitingBB)
2024 : << " at bottom\n");
2025 : std::rotate(LoopChain.begin(), std::next(ExitIt), LoopChain.end());
2026 : }
2027 :
2028 : /// Attempt to rotate a loop based on profile data to reduce branch cost.
2029 : ///
2030 : /// With profile data, we can determine the cost in terms of missed fall through
2031 : /// opportunities when rotating a loop chain and select the best rotation.
2032 : /// Basically, there are three kinds of cost to consider for each rotation:
2033 : /// 1. The possibly missed fall through edge (if it exists) from BB out of
2034 : /// the loop to the loop header.
2035 : /// 2. The possibly missed fall through edges (if they exist) from the loop
2036 : /// exits to BB out of the loop.
2037 : /// 3. The missed fall through edge (if it exists) from the last BB to the
2038 : /// first BB in the loop chain.
2039 : /// Therefore, the cost for a given rotation is the sum of costs listed above.
2040 : /// We select the best rotation with the smallest cost.
2041 4 : void MachineBlockPlacement::rotateLoopWithProfile(
2042 : BlockChain &LoopChain, const MachineLoop &L,
2043 : const BlockFilterSet &LoopBlockSet) {
2044 4 : auto HeaderBB = L.getHeader();
2045 : auto HeaderIter = llvm::find(LoopChain, HeaderBB);
2046 : auto RotationPos = LoopChain.end();
2047 :
2048 : BlockFrequency SmallestRotationCost = BlockFrequency::getMaxFrequency();
2049 :
2050 : // A utility lambda that scales up a block frequency by dividing it by a
2051 : // branch probability which is the reciprocal of the scale.
2052 : auto ScaleBlockFrequency = [](BlockFrequency Freq,
2053 : unsigned Scale) -> BlockFrequency {
2054 : if (Scale == 0)
2055 : return 0;
2056 : // Use operator / between BlockFrequency and BranchProbability to implement
2057 : // saturating multiplication.
2058 : return Freq / BranchProbability(1, Scale);
2059 : };
2060 :
2061 : // Compute the cost of the missed fall-through edge to the loop header if the
2062 : // chain head is not the loop header. As we only consider natural loops with
2063 : // single header, this computation can be done only once.
2064 : BlockFrequency HeaderFallThroughCost(0);
2065 12 : for (auto *Pred : HeaderBB->predecessors()) {
2066 8 : BlockChain *PredChain = BlockToChain[Pred];
2067 8 : if (!LoopBlockSet.count(Pred) &&
2068 4 : (!PredChain || Pred == *std::prev(PredChain->end()))) {
2069 : auto EdgeFreq =
2070 4 : MBFI->getBlockFreq(Pred) * MBPI->getEdgeProbability(Pred, HeaderBB);
2071 4 : auto FallThruCost = ScaleBlockFrequency(EdgeFreq, MisfetchCost);
2072 : // If the predecessor has only an unconditional jump to the header, we
2073 : // need to consider the cost of this jump.
2074 4 : if (Pred->succ_size() == 1)
2075 3 : FallThruCost += ScaleBlockFrequency(EdgeFreq, JumpInstCost);
2076 4 : HeaderFallThroughCost = std::max(HeaderFallThroughCost, FallThruCost);
2077 : }
2078 : }
2079 :
2080 : // Here we collect all exit blocks in the loop, and for each exit we find out
2081 : // its hottest exit edge. For each loop rotation, we define the loop exit cost
2082 : // as the sum of frequencies of exit edges we collect here, excluding the exit
2083 : // edge from the tail of the loop chain.
2084 : SmallVector<std::pair<MachineBasicBlock *, BlockFrequency>, 4> ExitsWithFreq;
2085 22 : for (auto BB : LoopChain) {
2086 18 : auto LargestExitEdgeProb = BranchProbability::getZero();
2087 47 : for (auto *Succ : BB->successors()) {
2088 29 : BlockChain *SuccChain = BlockToChain[Succ];
2089 29 : if (!LoopBlockSet.count(Succ) &&
2090 5 : (!SuccChain || Succ == *SuccChain->begin())) {
2091 5 : auto SuccProb = MBPI->getEdgeProbability(BB, Succ);
2092 5 : LargestExitEdgeProb = std::max(LargestExitEdgeProb, SuccProb);
2093 : }
2094 : }
2095 18 : if (LargestExitEdgeProb > BranchProbability::getZero()) {
2096 5 : auto ExitFreq = MBFI->getBlockFreq(BB) * LargestExitEdgeProb;
2097 5 : ExitsWithFreq.emplace_back(BB, ExitFreq);
2098 : }
2099 : }
2100 :
2101 : // In this loop we iterate every block in the loop chain and calculate the
2102 : // cost assuming the block is the head of the loop chain. When the loop ends,
2103 : // we should have found the best candidate as the loop chain's head.
2104 18 : for (auto Iter = LoopChain.begin(), TailIter = std::prev(LoopChain.end()),
2105 : EndIter = LoopChain.end();
2106 22 : Iter != EndIter; Iter++, TailIter++) {
2107 : // TailIter is used to track the tail of the loop chain if the block we are
2108 : // checking (pointed by Iter) is the head of the chain.
2109 18 : if (TailIter == LoopChain.end())
2110 : TailIter = LoopChain.begin();
2111 :
2112 18 : auto TailBB = *TailIter;
2113 :
2114 : // Calculate the cost by putting this BB to the top.
2115 : BlockFrequency Cost = 0;
2116 :
2117 : // If the current BB is the loop header, we need to take into account the
2118 : // cost of the missed fall through edge from outside of the loop to the
2119 : // header.
2120 18 : if (Iter != HeaderIter)
2121 14 : Cost += HeaderFallThroughCost;
2122 :
2123 : // Collect the loop exit cost by summing up frequencies of all exit edges
2124 : // except the one from the chain tail.
2125 39 : for (auto &ExitWithFreq : ExitsWithFreq)
2126 21 : if (TailBB != ExitWithFreq.first)
2127 16 : Cost += ExitWithFreq.second;
2128 :
2129 : // The cost of breaking the once fall-through edge from the tail to the top
2130 : // of the loop chain. Here we need to consider three cases:
2131 : // 1. If the tail node has only one successor, then we will get an
2132 : // additional jmp instruction. So the cost here is (MisfetchCost +
2133 : // JumpInstCost) * tail node frequency.
2134 : // 2. If the tail node has two successors, then we may still get an
2135 : // additional jmp instruction if the layout successor after the loop
2136 : // chain is not its CFG successor. Note that the more frequently executed
2137 : // jmp instruction will be put ahead of the other one. Assume the
2138 : // frequency of those two branches are x and y, where x is the frequency
2139 : // of the edge to the chain head, then the cost will be
2140 : // (x * MisfetechCost + min(x, y) * JumpInstCost) * tail node frequency.
2141 : // 3. If the tail node has more than two successors (this rarely happens),
2142 : // we won't consider any additional cost.
2143 18 : if (TailBB->isSuccessor(*Iter)) {
2144 14 : auto TailBBFreq = MBFI->getBlockFreq(TailBB);
2145 14 : if (TailBB->succ_size() == 1)
2146 : Cost += ScaleBlockFrequency(TailBBFreq.getFrequency(),
2147 6 : MisfetchCost + JumpInstCost);
2148 11 : else if (TailBB->succ_size() == 2) {
2149 11 : auto TailToHeadProb = MBPI->getEdgeProbability(TailBB, *Iter);
2150 11 : auto TailToHeadFreq = TailBBFreq * TailToHeadProb;
2151 22 : auto ColderEdgeFreq = TailToHeadProb > BranchProbability(1, 2)
2152 6 : ? TailBBFreq * TailToHeadProb.getCompl()
2153 11 : : TailToHeadFreq;
2154 22 : Cost += ScaleBlockFrequency(TailToHeadFreq, MisfetchCost) +
2155 22 : ScaleBlockFrequency(ColderEdgeFreq, JumpInstCost);
2156 : }
2157 : }
2158 :
2159 : LLVM_DEBUG(dbgs() << "The cost of loop rotation by making "
2160 : << getBlockName(*Iter)
2161 : << " to the top: " << Cost.getFrequency() << "\n");
2162 :
2163 18 : if (Cost < SmallestRotationCost) {
2164 : SmallestRotationCost = Cost;
2165 : RotationPos = Iter;
2166 : }
2167 : }
2168 :
2169 4 : if (RotationPos != LoopChain.end()) {
2170 : LLVM_DEBUG(dbgs() << "Rotate loop by making " << getBlockName(*RotationPos)
2171 : << " to the top\n");
2172 : std::rotate(LoopChain.begin(), RotationPos, LoopChain.end());
2173 : }
2174 4 : }
2175 :
2176 : /// Collect blocks in the given loop that are to be placed.
2177 : ///
2178 : /// When profile data is available, exclude cold blocks from the returned set;
2179 : /// otherwise, collect all blocks in the loop.
2180 : MachineBlockPlacement::BlockFilterSet
2181 9725 : MachineBlockPlacement::collectLoopBlockSet(const MachineLoop &L) {
2182 : BlockFilterSet LoopBlockSet;
2183 :
2184 : // Filter cold blocks off from LoopBlockSet when profile data is available.
2185 : // Collect the sum of frequencies of incoming edges to the loop header from
2186 : // outside. If we treat the loop as a super block, this is the frequency of
2187 : // the loop. Then for each block in the loop, we calculate the ratio between
2188 : // its frequency and the frequency of the loop block. When it is too small,
2189 : // don't add it to the loop chain. If there are outer loops, then this block
2190 : // will be merged into the first outer loop chain for which this block is not
2191 : // cold anymore. This needs precise profile data and we only do this when
2192 : // profile data is available.
2193 9725 : if (F->getFunction().hasProfileData() || ForceLoopColdBlock) {
2194 : BlockFrequency LoopFreq(0);
2195 98 : for (auto LoopPred : L.getHeader()->predecessors())
2196 66 : if (!L.contains(LoopPred))
2197 64 : LoopFreq += MBFI->getBlockFreq(LoopPred) *
2198 96 : MBPI->getEdgeProbability(LoopPred, L.getHeader());
2199 :
2200 113 : for (MachineBasicBlock *LoopBB : L.getBlocks()) {
2201 81 : auto Freq = MBFI->getBlockFreq(LoopBB).getFrequency();
2202 81 : if (Freq == 0 || LoopFreq.getFrequency() / Freq > LoopToColdBlockRatio)
2203 : continue;
2204 76 : LoopBlockSet.insert(LoopBB);
2205 : }
2206 : } else
2207 19386 : LoopBlockSet.insert(L.block_begin(), L.block_end());
2208 :
2209 9725 : return LoopBlockSet;
2210 : }
2211 :
2212 : /// Forms basic block chains from the natural loop structures.
2213 : ///
2214 : /// These chains are designed to preserve the existing *structure* of the code
2215 : /// as much as possible. We can then stitch the chains together in a way which
2216 : /// both preserves the topological structure and minimizes taken conditional
2217 : /// branches.
2218 9725 : void MachineBlockPlacement::buildLoopChains(const MachineLoop &L) {
2219 : // First recurse through any nested loops, building chains for those inner
2220 : // loops.
2221 11227 : for (const MachineLoop *InnerLoop : L)
2222 1502 : buildLoopChains(*InnerLoop);
2223 :
2224 : assert(BlockWorkList.empty() &&
2225 : "BlockWorkList not empty when starting to build loop chains.");
2226 : assert(EHPadWorkList.empty() &&
2227 : "EHPadWorkList not empty when starting to build loop chains.");
2228 9725 : BlockFilterSet LoopBlockSet = collectLoopBlockSet(L);
2229 :
2230 : // Check if we have profile data for this function. If yes, we will rotate
2231 : // this loop by modeling costs more precisely which requires the profile data
2232 : // for better layout.
2233 : bool RotateLoopWithProfile =
2234 9725 : ForcePreciseRotationCost ||
2235 6 : (PreciseRotationCost && F->getFunction().hasProfileData());
2236 :
2237 : // First check to see if there is an obviously preferable top block for the
2238 : // loop. This will default to the header, but may end up as one of the
2239 : // predecessors to the header if there is one which will result in strictly
2240 : // fewer branches in the loop body.
2241 : // When we use profile data to rotate the loop, this is unnecessary.
2242 : MachineBasicBlock *LoopTop =
2243 9721 : RotateLoopWithProfile ? L.getHeader() : findBestLoopTop(L, LoopBlockSet);
2244 :
2245 : // If we selected just the header for the loop top, look for a potentially
2246 : // profitable exit block in the event that rotating the loop can eliminate
2247 : // branches by placing an exit edge at the bottom.
2248 : //
2249 : // Loops are processed innermost to uttermost, make sure we clear
2250 : // PreferredLoopExit before processing a new loop.
2251 9725 : PreferredLoopExit = nullptr;
2252 9725 : if (!RotateLoopWithProfile && LoopTop == L.getHeader())
2253 8541 : PreferredLoopExit = findBestLoopExit(L, LoopBlockSet);
2254 :
2255 9725 : BlockChain &LoopChain = *BlockToChain[LoopTop];
2256 :
2257 : // FIXME: This is a really lame way of walking the chains in the loop: we
2258 : // walk the blocks, and use a set to prevent visiting a particular chain
2259 : // twice.
2260 : SmallPtrSet<BlockChain *, 4> UpdatedPreds;
2261 : assert(LoopChain.UnscheduledPredecessors == 0 &&
2262 : "LoopChain should not have unscheduled predecessors.");
2263 9725 : UpdatedPreds.insert(&LoopChain);
2264 :
2265 70695 : for (const MachineBasicBlock *LoopBB : LoopBlockSet)
2266 60970 : fillWorkLists(LoopBB, UpdatedPreds, &LoopBlockSet);
2267 :
2268 9725 : buildChain(LoopTop, LoopChain, &LoopBlockSet);
2269 :
2270 9725 : if (RotateLoopWithProfile)
2271 4 : rotateLoopWithProfile(LoopChain, L, LoopBlockSet);
2272 : else
2273 9721 : rotateLoop(LoopChain, PreferredLoopExit, LoopBlockSet);
2274 :
2275 : LLVM_DEBUG({
2276 : // Crash at the end so we get all of the debugging output first.
2277 : bool BadLoop = false;
2278 : if (LoopChain.UnscheduledPredecessors) {
2279 : BadLoop = true;
2280 : dbgs() << "Loop chain contains a block without its preds placed!\n"
2281 : << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
2282 : << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n";
2283 : }
2284 : for (MachineBasicBlock *ChainBB : LoopChain) {
2285 : dbgs() << " ... " << getBlockName(ChainBB) << "\n";
2286 : if (!LoopBlockSet.remove(ChainBB)) {
2287 : // We don't mark the loop as bad here because there are real situations
2288 : // where this can occur. For example, with an unanalyzable fallthrough
2289 : // from a loop block to a non-loop block or vice versa.
2290 : dbgs() << "Loop chain contains a block not contained by the loop!\n"
2291 : << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
2292 : << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
2293 : << " Bad block: " << getBlockName(ChainBB) << "\n";
2294 : }
2295 : }
2296 :
2297 : if (!LoopBlockSet.empty()) {
2298 : BadLoop = true;
2299 : for (const MachineBasicBlock *LoopBB : LoopBlockSet)
2300 : dbgs() << "Loop contains blocks never placed into a chain!\n"
2301 : << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
2302 : << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
2303 : << " Bad block: " << getBlockName(LoopBB) << "\n";
2304 : }
2305 : assert(!BadLoop && "Detected problems with the placement of this loop.");
2306 : });
2307 :
2308 : BlockWorkList.clear();
2309 : EHPadWorkList.clear();
2310 9725 : }
2311 :
2312 17903 : void MachineBlockPlacement::buildCFGChains() {
2313 : // Ensure that every BB in the function has an associated chain to simplify
2314 : // the assumptions of the remaining algorithm.
2315 : SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
2316 365816 : for (MachineFunction::iterator FI = F->begin(), FE = F->end(); FI != FE;
2317 : ++FI) {
2318 : MachineBasicBlock *BB = &*FI;
2319 : BlockChain *Chain =
2320 347913 : new (ChainAllocator.Allocate()) BlockChain(BlockToChain, BB);
2321 : // Also, merge any blocks which we cannot reason about and must preserve
2322 : // the exact fallthrough behavior for.
2323 : while (true) {
2324 : Cond.clear();
2325 349410 : MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
2326 349410 : if (!TII->analyzeBranch(*BB, TBB, FBB, Cond) || !FI->canFallThrough())
2327 : break;
2328 :
2329 : MachineFunction::iterator NextFI = std::next(FI);
2330 : MachineBasicBlock *NextBB = &*NextFI;
2331 : // Ensure that the layout successor is a viable block, as we know that
2332 : // fallthrough is a possibility.
2333 : assert(NextFI != FE && "Can't fallthrough past the last block.");
2334 : LLVM_DEBUG(dbgs() << "Pre-merging due to unanalyzable fallthrough: "
2335 : << getBlockName(BB) << " -> " << getBlockName(NextBB)
2336 : << "\n");
2337 1497 : Chain->merge(NextBB, nullptr);
2338 : #ifndef NDEBUG
2339 : BlocksWithUnanalyzableExits.insert(&*BB);
2340 : #endif
2341 : FI = NextFI;
2342 : BB = NextBB;
2343 1497 : }
2344 : }
2345 :
2346 : // Build any loop-based chains.
2347 17903 : PreferredLoopExit = nullptr;
2348 26126 : for (MachineLoop *L : *MLI)
2349 8223 : buildLoopChains(*L);
2350 :
2351 : assert(BlockWorkList.empty() &&
2352 : "BlockWorkList should be empty before building final chain.");
2353 : assert(EHPadWorkList.empty() &&
2354 : "EHPadWorkList should be empty before building final chain.");
2355 :
2356 : SmallPtrSet<BlockChain *, 4> UpdatedPreds;
2357 367030 : for (MachineBasicBlock &MBB : *F)
2358 349127 : fillWorkLists(&MBB, UpdatedPreds);
2359 :
2360 35806 : BlockChain &FunctionChain = *BlockToChain[&F->front()];
2361 35806 : buildChain(&F->front(), FunctionChain);
2362 :
2363 : #ifndef NDEBUG
2364 : using FunctionBlockSetType = SmallPtrSet<MachineBasicBlock *, 16>;
2365 : #endif
2366 : LLVM_DEBUG({
2367 : // Crash at the end so we get all of the debugging output first.
2368 : bool BadFunc = false;
2369 : FunctionBlockSetType FunctionBlockSet;
2370 : for (MachineBasicBlock &MBB : *F)
2371 : FunctionBlockSet.insert(&MBB);
2372 :
2373 : for (MachineBasicBlock *ChainBB : FunctionChain)
2374 : if (!FunctionBlockSet.erase(ChainBB)) {
2375 : BadFunc = true;
2376 : dbgs() << "Function chain contains a block not in the function!\n"
2377 : << " Bad block: " << getBlockName(ChainBB) << "\n";
2378 : }
2379 :
2380 : if (!FunctionBlockSet.empty()) {
2381 : BadFunc = true;
2382 : for (MachineBasicBlock *RemainingBB : FunctionBlockSet)
2383 : dbgs() << "Function contains blocks never placed into a chain!\n"
2384 : << " Bad block: " << getBlockName(RemainingBB) << "\n";
2385 : }
2386 : assert(!BadFunc && "Detected problems with the block placement.");
2387 : });
2388 :
2389 : // Splice the blocks into place.
2390 17903 : MachineFunction::iterator InsertPos = F->begin();
2391 : LLVM_DEBUG(dbgs() << "[MBP] Function: " << F->getName() << "\n");
2392 366601 : for (MachineBasicBlock *ChainBB : FunctionChain) {
2393 : LLVM_DEBUG(dbgs() << (ChainBB == *FunctionChain.begin() ? "Placing chain "
2394 : : " ... ")
2395 : << getBlockName(ChainBB) << "\n");
2396 348698 : if (InsertPos != MachineFunction::iterator(ChainBB))
2397 179251 : F->splice(InsertPos, ChainBB);
2398 : else
2399 : ++InsertPos;
2400 :
2401 : // Update the terminator of the previous block.
2402 348698 : if (ChainBB == *FunctionChain.begin())
2403 17903 : continue;
2404 : MachineBasicBlock *PrevBB = &*std::prev(MachineFunction::iterator(ChainBB));
2405 :
2406 : // FIXME: It would be awesome of updateTerminator would just return rather
2407 : // than assert when the branch cannot be analyzed in order to remove this
2408 : // boiler plate.
2409 : Cond.clear();
2410 330795 : MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
2411 :
2412 : #ifndef NDEBUG
2413 : if (!BlocksWithUnanalyzableExits.count(PrevBB)) {
2414 : // Given the exact block placement we chose, we may actually not _need_ to
2415 : // be able to edit PrevBB's terminator sequence, but not being _able_ to
2416 : // do that at this point is a bug.
2417 : assert((!TII->analyzeBranch(*PrevBB, TBB, FBB, Cond) ||
2418 : !PrevBB->canFallThrough()) &&
2419 : "Unexpected block with un-analyzable fallthrough!");
2420 : Cond.clear();
2421 : TBB = FBB = nullptr;
2422 : }
2423 : #endif
2424 :
2425 : // The "PrevBB" is not yet updated to reflect current code layout, so,
2426 : // o. it may fall-through to a block without explicit "goto" instruction
2427 : // before layout, and no longer fall-through it after layout; or
2428 : // o. just opposite.
2429 : //
2430 : // analyzeBranch() may return erroneous value for FBB when these two
2431 : // situations take place. For the first scenario FBB is mistakenly set NULL;
2432 : // for the 2nd scenario, the FBB, which is expected to be NULL, is
2433 : // mistakenly pointing to "*BI".
2434 : // Thus, if the future change needs to use FBB before the layout is set, it
2435 : // has to correct FBB first by using the code similar to the following:
2436 : //
2437 : // if (!Cond.empty() && (!FBB || FBB == ChainBB)) {
2438 : // PrevBB->updateTerminator();
2439 : // Cond.clear();
2440 : // TBB = FBB = nullptr;
2441 : // if (TII->analyzeBranch(*PrevBB, TBB, FBB, Cond)) {
2442 : // // FIXME: This should never take place.
2443 : // TBB = FBB = nullptr;
2444 : // }
2445 : // }
2446 330795 : if (!TII->analyzeBranch(*PrevBB, TBB, FBB, Cond))
2447 318923 : PrevBB->updateTerminator();
2448 : }
2449 :
2450 : // Fixup the last block.
2451 : Cond.clear();
2452 17903 : MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
2453 35806 : if (!TII->analyzeBranch(F->back(), TBB, FBB, Cond))
2454 12806 : F->back().updateTerminator();
2455 :
2456 : BlockWorkList.clear();
2457 : EHPadWorkList.clear();
2458 17903 : }
2459 :
2460 17196 : void MachineBlockPlacement::optimizeBranches() {
2461 34392 : BlockChain &FunctionChain = *BlockToChain[&F->front()];
2462 : SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
2463 :
2464 : // Now that all the basic blocks in the chain have the proper layout,
2465 : // make a final call to AnalyzeBranch with AllowModify set.
2466 : // Indeed, the target may be able to optimize the branches in a way we
2467 : // cannot because all branches may not be analyzable.
2468 : // E.g., the target may be able to remove an unconditional branch to
2469 : // a fallthrough when it occurs after predicated terminators.
2470 290596 : for (MachineBasicBlock *ChainBB : FunctionChain) {
2471 : Cond.clear();
2472 273400 : MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
2473 273400 : if (!TII->analyzeBranch(*ChainBB, TBB, FBB, Cond, /*AllowModify*/ true)) {
2474 : // If PrevBB has a two-way branch, try to re-order the branches
2475 : // such that we branch to the successor with higher probability first.
2476 141834 : if (TBB && !Cond.empty() && FBB &&
2477 252403 : MBPI->getEdgeProbability(ChainBB, FBB) >
2478 253251 : MBPI->getEdgeProbability(ChainBB, TBB) &&
2479 424 : !TII->reverseBranchCondition(Cond)) {
2480 : LLVM_DEBUG(dbgs() << "Reverse order of the two branches: "
2481 : << getBlockName(ChainBB) << "\n");
2482 : LLVM_DEBUG(dbgs() << " Edge probability: "
2483 : << MBPI->getEdgeProbability(ChainBB, FBB) << " vs "
2484 : << MBPI->getEdgeProbability(ChainBB, TBB) << "\n");
2485 424 : DebugLoc dl; // FIXME: this is nowhere
2486 424 : TII->removeBranch(*ChainBB);
2487 848 : TII->insertBranch(*ChainBB, FBB, TBB, Cond, dl);
2488 424 : ChainBB->updateTerminator();
2489 : }
2490 : }
2491 : }
2492 17196 : }
2493 :
2494 17196 : void MachineBlockPlacement::alignBlocks() {
2495 : // Walk through the backedges of the function now that we have fully laid out
2496 : // the basic blocks and align the destination of each backedge. We don't rely
2497 : // exclusively on the loop info here so that we can align backedges in
2498 : // unnatural CFGs and backedges that were introduced purely because of the
2499 : // loop rotations done during this layout pass.
2500 34321 : if (F->getFunction().optForMinSize() ||
2501 17319 : (F->getFunction().optForSize() && !TLI->alignLoopsWithOptSize()))
2502 260 : return;
2503 33872 : BlockChain &FunctionChain = *BlockToChain[&F->front()];
2504 16936 : if (FunctionChain.begin() == FunctionChain.end())
2505 : return; // Empty chain.
2506 :
2507 16936 : const BranchProbability ColdProb(1, 5); // 20%
2508 33872 : BlockFrequency EntryFreq = MBFI->getBlockFreq(&F->front());
2509 16936 : BlockFrequency WeightedEntryFreq = EntryFreq * ColdProb;
2510 289218 : for (MachineBasicBlock *ChainBB : FunctionChain) {
2511 272282 : if (ChainBB == *FunctionChain.begin())
2512 247440 : continue;
2513 :
2514 : // Don't align non-looping basic blocks. These are unlikely to execute
2515 : // enough times to matter in practice. Note that we'll still handle
2516 : // unnatural CFGs inside of a natural outer loop (the common case) and
2517 : // rotated loops.
2518 255346 : MachineLoop *L = MLI->getLoopFor(ChainBB);
2519 38205 : if (!L)
2520 217141 : continue;
2521 :
2522 38205 : unsigned Align = TLI->getPrefLoopAlignment(L);
2523 38205 : if (!Align)
2524 : continue; // Don't care about loop alignment.
2525 :
2526 : // If the block is cold relative to the function entry don't waste space
2527 : // aligning it.
2528 34519 : BlockFrequency Freq = MBFI->getBlockFreq(ChainBB);
2529 34519 : if (Freq < WeightedEntryFreq)
2530 : continue;
2531 :
2532 : // If the block is cold relative to its loop header, don't align it
2533 : // regardless of what edges into the block exist.
2534 : MachineBasicBlock *LoopHeader = L->getHeader();
2535 33031 : BlockFrequency LoopHeaderFreq = MBFI->getBlockFreq(LoopHeader);
2536 33031 : if (Freq < (LoopHeaderFreq * ColdProb))
2537 : continue;
2538 :
2539 : // Check for the existence of a non-layout predecessor which would benefit
2540 : // from aligning this block.
2541 : MachineBasicBlock *LayoutPred =
2542 : &*std::prev(MachineFunction::iterator(ChainBB));
2543 :
2544 : // Force alignment if all the predecessors are jumps. We already checked
2545 : // that the block isn't cold above.
2546 27807 : if (!LayoutPred->isSuccessor(ChainBB)) {
2547 : ChainBB->setAlignment(Align);
2548 2965 : continue;
2549 : }
2550 :
2551 : // Align this block if the layout predecessor's edge into this block is
2552 : // cold relative to the block. When this is true, other predecessors make up
2553 : // all of the hot entries into the block and thus alignment is likely to be
2554 : // important.
2555 : BranchProbability LayoutProb =
2556 24842 : MBPI->getEdgeProbability(LayoutPred, ChainBB);
2557 24842 : BlockFrequency LayoutEdgeFreq = MBFI->getBlockFreq(LayoutPred) * LayoutProb;
2558 24842 : if (LayoutEdgeFreq <= (Freq * ColdProb))
2559 : ChainBB->setAlignment(Align);
2560 : }
2561 : }
2562 :
2563 : /// Tail duplicate \p BB into (some) predecessors if profitable, repeating if
2564 : /// it was duplicated into its chain predecessor and removed.
2565 : /// \p BB - Basic block that may be duplicated.
2566 : ///
2567 : /// \p LPred - Chosen layout predecessor of \p BB.
2568 : /// Updated to be the chain end if LPred is removed.
2569 : /// \p Chain - Chain to which \p LPred belongs, and \p BB will belong.
2570 : /// \p BlockFilter - Set of blocks that belong to the loop being laid out.
2571 : /// Used to identify which blocks to update predecessor
2572 : /// counts.
2573 : /// \p PrevUnplacedBlockIt - Iterator pointing to the last block that was
2574 : /// chosen in the given order due to unnatural CFG
2575 : /// only needed if \p BB is removed and
2576 : /// \p PrevUnplacedBlockIt pointed to \p BB.
2577 : /// @return true if \p BB was removed.
2578 22424 : bool MachineBlockPlacement::repeatedlyTailDuplicateBlock(
2579 : MachineBasicBlock *BB, MachineBasicBlock *&LPred,
2580 : const MachineBasicBlock *LoopHeaderBB,
2581 : BlockChain &Chain, BlockFilterSet *BlockFilter,
2582 : MachineFunction::iterator &PrevUnplacedBlockIt) {
2583 : bool Removed, DuplicatedToLPred;
2584 : bool DuplicatedToOriginalLPred;
2585 22424 : Removed = maybeTailDuplicateBlock(BB, LPred, Chain, BlockFilter,
2586 : PrevUnplacedBlockIt,
2587 : DuplicatedToLPred);
2588 22424 : if (!Removed)
2589 : return false;
2590 712 : DuplicatedToOriginalLPred = DuplicatedToLPred;
2591 : // Iteratively try to duplicate again. It can happen that a block that is
2592 : // duplicated into is still small enough to be duplicated again.
2593 : // No need to call markBlockSuccessors in this case, as the blocks being
2594 : // duplicated from here on are already scheduled.
2595 : // Note that DuplicatedToLPred always implies Removed.
2596 1251 : while (DuplicatedToLPred) {
2597 : assert(Removed && "Block must have been removed to be duplicated into its "
2598 : "layout predecessor.");
2599 : MachineBasicBlock *DupBB, *DupPred;
2600 : // The removal callback causes Chain.end() to be updated when a block is
2601 : // removed. On the first pass through the loop, the chain end should be the
2602 : // same as it was on function entry. On subsequent passes, because we are
2603 : // duplicating the block at the end of the chain, if it is removed the
2604 : // chain will have shrunk by one block.
2605 : BlockChain::iterator ChainEnd = Chain.end();
2606 712 : DupBB = *(--ChainEnd);
2607 : // Now try to duplicate again.
2608 712 : if (ChainEnd == Chain.begin())
2609 : break;
2610 539 : DupPred = *std::prev(ChainEnd);
2611 539 : Removed = maybeTailDuplicateBlock(DupBB, DupPred, Chain, BlockFilter,
2612 : PrevUnplacedBlockIt,
2613 : DuplicatedToLPred);
2614 : }
2615 : // If BB was duplicated into LPred, it is now scheduled. But because it was
2616 : // removed, markChainSuccessors won't be called for its chain. Instead we
2617 : // call markBlockSuccessors for LPred to achieve the same effect. This must go
2618 : // at the end because repeating the tail duplication can increase the number
2619 : // of unscheduled predecessors.
2620 712 : LPred = *std::prev(Chain.end());
2621 712 : if (DuplicatedToOriginalLPred)
2622 712 : markBlockSuccessors(Chain, LPred, LoopHeaderBB, BlockFilter);
2623 : return true;
2624 : }
2625 :
2626 : /// Tail duplicate \p BB into (some) predecessors if profitable.
2627 : /// \p BB - Basic block that may be duplicated
2628 : /// \p LPred - Chosen layout predecessor of \p BB
2629 : /// \p Chain - Chain to which \p LPred belongs, and \p BB will belong.
2630 : /// \p BlockFilter - Set of blocks that belong to the loop being laid out.
2631 : /// Used to identify which blocks to update predecessor
2632 : /// counts.
2633 : /// \p PrevUnplacedBlockIt - Iterator pointing to the last block that was
2634 : /// chosen in the given order due to unnatural CFG
2635 : /// only needed if \p BB is removed and
2636 : /// \p PrevUnplacedBlockIt pointed to \p BB.
2637 : /// \p DuplicatedToLPred - True if the block was duplicated into LPred. Will
2638 : /// only be true if the block was removed.
2639 : /// \return - True if the block was duplicated into all preds and removed.
2640 22963 : bool MachineBlockPlacement::maybeTailDuplicateBlock(
2641 : MachineBasicBlock *BB, MachineBasicBlock *LPred,
2642 : BlockChain &Chain, BlockFilterSet *BlockFilter,
2643 : MachineFunction::iterator &PrevUnplacedBlockIt,
2644 : bool &DuplicatedToLPred) {
2645 22963 : DuplicatedToLPred = false;
2646 22963 : if (!shouldTailDuplicate(BB))
2647 : return false;
2648 :
2649 : LLVM_DEBUG(dbgs() << "Redoing tail duplication for Succ#" << BB->getNumber()
2650 : << "\n");
2651 :
2652 : // This has to be a callback because none of it can be done after
2653 : // BB is deleted.
2654 22451 : bool Removed = false;
2655 : auto RemovalCallback =
2656 : [&](MachineBasicBlock *RemBB) {
2657 : // Signal to outer function
2658 : Removed = true;
2659 :
2660 : // Conservative default.
2661 : bool InWorkList = true;
2662 : // Remove from the Chain and Chain Map
2663 : if (BlockToChain.count(RemBB)) {
2664 : BlockChain *Chain = BlockToChain[RemBB];
2665 : InWorkList = Chain->UnscheduledPredecessors == 0;
2666 : Chain->remove(RemBB);
2667 : BlockToChain.erase(RemBB);
2668 : }
2669 :
2670 : // Handle the unplaced block iterator
2671 : if (&(*PrevUnplacedBlockIt) == RemBB) {
2672 : PrevUnplacedBlockIt++;
2673 : }
2674 :
2675 : // Handle the Work Lists
2676 : if (InWorkList) {
2677 : SmallVectorImpl<MachineBasicBlock *> &RemoveList = BlockWorkList;
2678 : if (RemBB->isEHPad())
2679 : RemoveList = EHPadWorkList;
2680 : RemoveList.erase(
2681 : llvm::remove_if(RemoveList,
2682 : [RemBB](MachineBasicBlock *BB) {
2683 0 : return BB == RemBB;
2684 : }),
2685 : RemoveList.end());
2686 : }
2687 :
2688 : // Handle the filter set
2689 : if (BlockFilter) {
2690 : BlockFilter->remove(RemBB);
2691 : }
2692 :
2693 : // Remove the block from loop info.
2694 : MLI->removeBlock(RemBB);
2695 : if (RemBB == PreferredLoopExit)
2696 : PreferredLoopExit = nullptr;
2697 :
2698 : LLVM_DEBUG(dbgs() << "TailDuplicator deleted block: "
2699 : << getBlockName(RemBB) << "\n");
2700 22451 : };
2701 : auto RemovalCallbackRef =
2702 : function_ref<void(MachineBasicBlock*)>(RemovalCallback);
2703 :
2704 : SmallVector<MachineBasicBlock *, 8> DuplicatedPreds;
2705 22451 : bool IsSimple = TailDup.isSimpleBB(BB);
2706 22451 : TailDup.tailDuplicateAndUpdate(IsSimple, BB, LPred,
2707 : &DuplicatedPreds, &RemovalCallbackRef);
2708 :
2709 : // Update UnscheduledPredecessors to reflect tail-duplication.
2710 22451 : DuplicatedToLPred = false;
2711 25335 : for (MachineBasicBlock *Pred : DuplicatedPreds) {
2712 : // We're only looking for unscheduled predecessors that match the filter.
2713 2884 : BlockChain* PredChain = BlockToChain[Pred];
2714 2884 : if (Pred == LPred)
2715 712 : DuplicatedToLPred = true;
2716 2654 : if (Pred == LPred || (BlockFilter && !BlockFilter->count(Pred))
2717 4872 : || PredChain == &Chain)
2718 1287 : continue;
2719 4057 : for (MachineBasicBlock *NewSucc : Pred->successors()) {
2720 2896 : if (BlockFilter && !BlockFilter->count(NewSucc))
2721 : continue;
2722 2384 : BlockChain *NewChain = BlockToChain[NewSucc];
2723 2384 : if (NewChain != &Chain && NewChain != PredChain)
2724 2348 : NewChain->UnscheduledPredecessors++;
2725 : }
2726 : }
2727 22451 : return Removed;
2728 : }
2729 :
2730 197814 : bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &MF) {
2731 197814 : if (skipFunction(MF.getFunction()))
2732 : return false;
2733 :
2734 : // Check for single-block functions and skip them.
2735 197629 : if (std::next(MF.begin()) == MF.end())
2736 : return false;
2737 :
2738 17196 : F = &MF;
2739 17196 : MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
2740 17196 : MBFI = llvm::make_unique<BranchFolder::MBFIWrapper>(
2741 : getAnalysis<MachineBlockFrequencyInfo>());
2742 17196 : MLI = &getAnalysis<MachineLoopInfo>();
2743 17196 : TII = MF.getSubtarget().getInstrInfo();
2744 17196 : TLI = MF.getSubtarget().getTargetLowering();
2745 17196 : MPDT = nullptr;
2746 :
2747 : // Initialize PreferredLoopExit to nullptr here since it may never be set if
2748 : // there are no MachineLoops.
2749 17196 : PreferredLoopExit = nullptr;
2750 :
2751 : assert(BlockToChain.empty() &&
2752 : "BlockToChain map should be empty before starting placement.");
2753 : assert(ComputedEdges.empty() &&
2754 : "Computed Edge map should be empty before starting placement.");
2755 :
2756 : unsigned TailDupSize = TailDupPlacementThreshold;
2757 : // If only the aggressive threshold is explicitly set, use it.
2758 17196 : if (TailDupPlacementAggressiveThreshold.getNumOccurrences() != 0 &&
2759 0 : TailDupPlacementThreshold.getNumOccurrences() == 0)
2760 : TailDupSize = TailDupPlacementAggressiveThreshold;
2761 :
2762 17196 : TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
2763 : // For aggressive optimization, we can adjust some thresholds to be less
2764 : // conservative.
2765 17196 : if (PassConfig->getOptLevel() >= CodeGenOpt::Aggressive) {
2766 : // At O3 we should be more willing to copy blocks for tail duplication. This
2767 : // increases size pressure, so we only do it at O3
2768 : // Do this unless only the regular threshold is explicitly set.
2769 266 : if (TailDupPlacementThreshold.getNumOccurrences() == 0 ||
2770 0 : TailDupPlacementAggressiveThreshold.getNumOccurrences() != 0)
2771 : TailDupSize = TailDupPlacementAggressiveThreshold;
2772 : }
2773 :
2774 : if (allowTailDupPlacement()) {
2775 17054 : MPDT = &getAnalysis<MachinePostDominatorTree>();
2776 17054 : if (MF.getFunction().optForSize())
2777 : TailDupSize = 1;
2778 : bool PreRegAlloc = false;
2779 17054 : TailDup.initMF(MF, PreRegAlloc, MBPI, /* LayoutMode */ true, TailDupSize);
2780 17054 : precomputeTriangleChains();
2781 : }
2782 :
2783 17196 : buildCFGChains();
2784 :
2785 : // Changing the layout can create new tail merging opportunities.
2786 : // TailMerge can create jump into if branches that make CFG irreducible for
2787 : // HW that requires structured CFG.
2788 17196 : bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
2789 17196 : PassConfig->getEnableTailMerge() &&
2790 : BranchFoldPlacement;
2791 : // No tail merging opportunities if the block number is less than four.
2792 17196 : if (MF.size() > 3 && EnableTailMerge) {
2793 8461 : unsigned TailMergeSize = TailDupSize + 1;
2794 : BranchFolder BF(/*EnableTailMerge=*/true, /*CommonHoist=*/false, *MBFI,
2795 25383 : *MBPI, TailMergeSize);
2796 :
2797 8461 : if (BF.OptimizeFunction(MF, TII, MF.getSubtarget().getRegisterInfo(),
2798 : getAnalysisIfAvailable<MachineModuleInfo>(), MLI,
2799 : /*AfterBlockPlacement=*/true)) {
2800 : // Redo the layout if tail merging creates/removes/moves blocks.
2801 707 : BlockToChain.clear();
2802 707 : ComputedEdges.clear();
2803 : // Must redo the post-dominator tree if blocks were changed.
2804 707 : if (MPDT)
2805 707 : MPDT->runOnMachineFunction(MF);
2806 707 : ChainAllocator.DestroyAll();
2807 707 : buildCFGChains();
2808 : }
2809 : }
2810 :
2811 17196 : optimizeBranches();
2812 17196 : alignBlocks();
2813 :
2814 17196 : BlockToChain.clear();
2815 17196 : ComputedEdges.clear();
2816 17196 : ChainAllocator.DestroyAll();
2817 :
2818 17196 : if (AlignAllBlock)
2819 : // Align all of the blocks in the function to a specific alignment.
2820 4 : for (MachineBasicBlock &MBB : MF)
2821 : MBB.setAlignment(AlignAllBlock);
2822 17195 : else if (AlignAllNonFallThruBlocks) {
2823 : // Align all of the blocks that have no fall-through predecessors to a
2824 : // specific alignment.
2825 3 : for (auto MBI = std::next(MF.begin()), MBE = MF.end(); MBI != MBE; ++MBI) {
2826 : auto LayoutPred = std::prev(MBI);
2827 2 : if (!LayoutPred->isSuccessor(&*MBI))
2828 : MBI->setAlignment(AlignAllNonFallThruBlocks);
2829 : }
2830 : }
2831 17196 : if (ViewBlockLayoutWithBFI != GVDT_None &&
2832 : (ViewBlockFreqFuncName.empty() ||
2833 17196 : F->getFunction().getName().equals(ViewBlockFreqFuncName))) {
2834 0 : MBFI->view("MBP." + MF.getName(), false);
2835 : }
2836 :
2837 :
2838 : // We always return true as we have no way to track whether the final order
2839 : // differs from the original order.
2840 : return true;
2841 : }
2842 :
2843 : namespace {
2844 :
2845 : /// A pass to compute block placement statistics.
2846 : ///
2847 : /// A separate pass to compute interesting statistics for evaluating block
2848 : /// placement. This is separate from the actual placement pass so that they can
2849 : /// be computed in the absence of any placement transformations or when using
2850 : /// alternative placement strategies.
2851 : class MachineBlockPlacementStats : public MachineFunctionPass {
2852 : /// A handle to the branch probability pass.
2853 : const MachineBranchProbabilityInfo *MBPI;
2854 :
2855 : /// A handle to the function-wide block frequency pass.
2856 : const MachineBlockFrequencyInfo *MBFI;
2857 :
2858 : public:
2859 : static char ID; // Pass identification, replacement for typeid
2860 :
2861 0 : MachineBlockPlacementStats() : MachineFunctionPass(ID) {
2862 0 : initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry());
2863 0 : }
2864 :
2865 : bool runOnMachineFunction(MachineFunction &F) override;
2866 :
2867 0 : void getAnalysisUsage(AnalysisUsage &AU) const override {
2868 : AU.addRequired<MachineBranchProbabilityInfo>();
2869 : AU.addRequired<MachineBlockFrequencyInfo>();
2870 : AU.setPreservesAll();
2871 0 : MachineFunctionPass::getAnalysisUsage(AU);
2872 0 : }
2873 : };
2874 :
2875 : } // end anonymous namespace
2876 :
2877 : char MachineBlockPlacementStats::ID = 0;
2878 :
2879 : char &llvm::MachineBlockPlacementStatsID = MachineBlockPlacementStats::ID;
2880 :
2881 31780 : INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats",
2882 : "Basic Block Placement Stats", false, false)
2883 31780 : INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
2884 31780 : INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
2885 85147 : INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats",
2886 : "Basic Block Placement Stats", false, false)
2887 :
2888 0 : bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) {
2889 : // Check for single-block functions and skip them.
2890 0 : if (std::next(F.begin()) == F.end())
2891 : return false;
2892 :
2893 0 : MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
2894 0 : MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
2895 :
2896 0 : for (MachineBasicBlock &MBB : F) {
2897 0 : BlockFrequency BlockFreq = MBFI->getBlockFreq(&MBB);
2898 : Statistic &NumBranches =
2899 : (MBB.succ_size() > 1) ? NumCondBranches : NumUncondBranches;
2900 : Statistic &BranchTakenFreq =
2901 : (MBB.succ_size() > 1) ? CondBranchTakenFreq : UncondBranchTakenFreq;
2902 0 : for (MachineBasicBlock *Succ : MBB.successors()) {
2903 : // Skip if this successor is a fallthrough.
2904 0 : if (MBB.isLayoutSuccessor(Succ))
2905 0 : continue;
2906 :
2907 : BlockFrequency EdgeFreq =
2908 0 : BlockFreq * MBPI->getEdgeProbability(&MBB, Succ);
2909 : ++NumBranches;
2910 : BranchTakenFreq += EdgeFreq.getFrequency();
2911 : }
2912 : }
2913 :
2914 : return false;
2915 : }
|
