1//===- MachineBlockPlacement.cpp - Basic Block Code Layout optimization ---===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file implements basic block placement transformations using the CFG
10// structure and branch probability estimates.
11//
12// The pass strives to preserve the structure of the CFG (that is, retain
13// a topological ordering of basic blocks) in the absence of a *strong* signal
14// to the contrary from probabilities. However, within the CFG structure, it
15// attempts to choose an ordering which favors placing more likely sequences of
16// blocks adjacent to each other.
17//
18// The algorithm works from the inner-most loop within a function outward, and
19// at each stage walks through the basic blocks, trying to coalesce them into
20// sequential chains where allowed by the CFG (or demanded by heavy
21// probabilities). Finally, it walks the blocks in topological order, and the
22// first time it reaches a chain of basic blocks, it schedules them in the
23// function in-order.
24//
25//===----------------------------------------------------------------------===//
26
27#include "BranchFolding.h"
28#include "llvm/ADT/ArrayRef.h"
29#include "llvm/ADT/DenseMap.h"
30#include "llvm/ADT/STLExtras.h"
31#include "llvm/ADT/SetVector.h"
32#include "llvm/ADT/SmallPtrSet.h"
33#include "llvm/ADT/SmallVector.h"
34#include "llvm/ADT/Statistic.h"
35#include "llvm/Analysis/BlockFrequencyInfoImpl.h"
36#include "llvm/Analysis/ProfileSummaryInfo.h"
37#include "llvm/CodeGen/MBFIWrapper.h"
38#include "llvm/CodeGen/MachineBasicBlock.h"
39#include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
40#include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
41#include "llvm/CodeGen/MachineFunction.h"
42#include "llvm/CodeGen/MachineFunctionPass.h"
43#include "llvm/CodeGen/MachineLoopInfo.h"
44#include "llvm/CodeGen/MachinePostDominators.h"
45#include "llvm/CodeGen/MachineSizeOpts.h"
46#include "llvm/CodeGen/TailDuplicator.h"
47#include "llvm/CodeGen/TargetInstrInfo.h"
48#include "llvm/CodeGen/TargetLowering.h"
49#include "llvm/CodeGen/TargetPassConfig.h"
50#include "llvm/CodeGen/TargetSubtargetInfo.h"
51#include "llvm/IR/DebugLoc.h"
52#include "llvm/IR/Function.h"
53#include "llvm/IR/PrintPasses.h"
54#include "llvm/InitializePasses.h"
55#include "llvm/Pass.h"
56#include "llvm/Support/Allocator.h"
57#include "llvm/Support/BlockFrequency.h"
58#include "llvm/Support/BranchProbability.h"
59#include "llvm/Support/CodeGen.h"
60#include "llvm/Support/CommandLine.h"
61#include "llvm/Support/Compiler.h"
62#include "llvm/Support/Debug.h"
63#include "llvm/Support/raw_ostream.h"
64#include "llvm/Target/TargetMachine.h"
65#include "llvm/Transforms/Utils/CodeLayout.h"
66#include <algorithm>
67#include <cassert>
68#include <cstdint>
69#include <iterator>
70#include <memory>
71#include <string>
72#include <tuple>
73#include <utility>
74#include <vector>
75
76using namespace llvm;
77
78#define DEBUG_TYPE "block-placement"
79
80STATISTIC(NumCondBranches, "Number of conditional branches");
81STATISTIC(NumUncondBranches, "Number of unconditional branches");
82STATISTIC(CondBranchTakenFreq,
83 "Potential frequency of taking conditional branches");
84STATISTIC(UncondBranchTakenFreq,
85 "Potential frequency of taking unconditional branches");
86
87static cl::opt<unsigned> AlignAllBlock(
88 "align-all-blocks",
89 cl::desc("Force the alignment of all blocks in the function in log2 format "
90 "(e.g 4 means align on 16B boundaries)."),
91 cl::init(Val: 0), cl::Hidden);
92
93static cl::opt<unsigned> AlignAllNonFallThruBlocks(
94 "align-all-nofallthru-blocks",
95 cl::desc("Force the alignment of all blocks that have no fall-through "
96 "predecessors (i.e. don't add nops that are executed). In log2 "
97 "format (e.g 4 means align on 16B boundaries)."),
98 cl::init(Val: 0), cl::Hidden);
99
100static cl::opt<unsigned> MaxBytesForAlignmentOverride(
101 "max-bytes-for-alignment",
102 cl::desc("Forces the maximum bytes allowed to be emitted when padding for "
103 "alignment"),
104 cl::init(Val: 0), cl::Hidden);
105
106// FIXME: Find a good default for this flag and remove the flag.
107static cl::opt<unsigned> ExitBlockBias(
108 "block-placement-exit-block-bias",
109 cl::desc("Block frequency percentage a loop exit block needs "
110 "over the original exit to be considered the new exit."),
111 cl::init(Val: 0), cl::Hidden);
112
113// Definition:
114// - Outlining: placement of a basic block outside the chain or hot path.
115
116static cl::opt<unsigned> LoopToColdBlockRatio(
117 "loop-to-cold-block-ratio",
118 cl::desc("Outline loop blocks from loop chain if (frequency of loop) / "
119 "(frequency of block) is greater than this ratio"),
120 cl::init(Val: 5), cl::Hidden);
121
122static cl::opt<bool> ForceLoopColdBlock(
123 "force-loop-cold-block",
124 cl::desc("Force outlining cold blocks from loops."),
125 cl::init(Val: false), cl::Hidden);
126
127static cl::opt<bool>
128 PreciseRotationCost("precise-rotation-cost",
129 cl::desc("Model the cost of loop rotation more "
130 "precisely by using profile data."),
131 cl::init(Val: false), cl::Hidden);
132
133static cl::opt<bool>
134 ForcePreciseRotationCost("force-precise-rotation-cost",
135 cl::desc("Force the use of precise cost "
136 "loop rotation strategy."),
137 cl::init(Val: false), cl::Hidden);
138
139static cl::opt<unsigned> MisfetchCost(
140 "misfetch-cost",
141 cl::desc("Cost that models the probabilistic risk of an instruction "
142 "misfetch due to a jump comparing to falling through, whose cost "
143 "is zero."),
144 cl::init(Val: 1), cl::Hidden);
145
146static cl::opt<unsigned> JumpInstCost("jump-inst-cost",
147 cl::desc("Cost of jump instructions."),
148 cl::init(Val: 1), cl::Hidden);
149static cl::opt<bool>
150TailDupPlacement("tail-dup-placement",
151 cl::desc("Perform tail duplication during placement. "
152 "Creates more fallthrough opportunites in "
153 "outline branches."),
154 cl::init(Val: true), cl::Hidden);
155
156static cl::opt<bool>
157BranchFoldPlacement("branch-fold-placement",
158 cl::desc("Perform branch folding during placement. "
159 "Reduces code size."),
160 cl::init(Val: true), cl::Hidden);
161
162// Heuristic for tail duplication.
163static cl::opt<unsigned> TailDupPlacementThreshold(
164 "tail-dup-placement-threshold",
165 cl::desc("Instruction cutoff for tail duplication during layout. "
166 "Tail merging during layout is forced to have a threshold "
167 "that won't conflict."), cl::init(Val: 2),
168 cl::Hidden);
169
170// Heuristic for aggressive tail duplication.
171static cl::opt<unsigned> TailDupPlacementAggressiveThreshold(
172 "tail-dup-placement-aggressive-threshold",
173 cl::desc("Instruction cutoff for aggressive tail duplication during "
174 "layout. Used at -O3. Tail merging during layout is forced to "
175 "have a threshold that won't conflict."), cl::init(Val: 4),
176 cl::Hidden);
177
178// Heuristic for tail duplication.
179static cl::opt<unsigned> TailDupPlacementPenalty(
180 "tail-dup-placement-penalty",
181 cl::desc("Cost penalty for blocks that can avoid breaking CFG by copying. "
182 "Copying can increase fallthrough, but it also increases icache "
183 "pressure. This parameter controls the penalty to account for that. "
184 "Percent as integer."),
185 cl::init(Val: 2),
186 cl::Hidden);
187
188// Heuristic for tail duplication if profile count is used in cost model.
189static cl::opt<unsigned> TailDupProfilePercentThreshold(
190 "tail-dup-profile-percent-threshold",
191 cl::desc("If profile count information is used in tail duplication cost "
192 "model, the gained fall through number from tail duplication "
193 "should be at least this percent of hot count."),
194 cl::init(Val: 50), cl::Hidden);
195
196// Heuristic for triangle chains.
197static cl::opt<unsigned> TriangleChainCount(
198 "triangle-chain-count",
199 cl::desc("Number of triangle-shaped-CFG's that need to be in a row for the "
200 "triangle tail duplication heuristic to kick in. 0 to disable."),
201 cl::init(Val: 2),
202 cl::Hidden);
203
204// Use case: When block layout is visualized after MBP pass, the basic blocks
205// are labeled in layout order; meanwhile blocks could be numbered in a
206// different order. It's hard to map between the graph and pass output.
207// With this option on, the basic blocks are renumbered in function layout
208// order. For debugging only.
209static cl::opt<bool> RenumberBlocksBeforeView(
210 "renumber-blocks-before-view",
211 cl::desc(
212 "If true, basic blocks are re-numbered before MBP layout is printed "
213 "into a dot graph. Only used when a function is being printed."),
214 cl::init(Val: false), cl::Hidden);
215
216namespace llvm {
217extern cl::opt<bool> EnableExtTspBlockPlacement;
218extern cl::opt<bool> ApplyExtTspWithoutProfile;
219extern cl::opt<unsigned> StaticLikelyProb;
220extern cl::opt<unsigned> ProfileLikelyProb;
221
222// Internal option used to control BFI display only after MBP pass.
223// Defined in CodeGen/MachineBlockFrequencyInfo.cpp:
224// -view-block-layout-with-bfi=
225extern cl::opt<GVDAGType> ViewBlockLayoutWithBFI;
226
227// Command line option to specify the name of the function for CFG dump
228// Defined in Analysis/BlockFrequencyInfo.cpp: -view-bfi-func-name=
229extern cl::opt<std::string> ViewBlockFreqFuncName;
230} // namespace llvm
231
232namespace {
233
234class BlockChain;
235
236/// Type for our function-wide basic block -> block chain mapping.
237using BlockToChainMapType = DenseMap<const MachineBasicBlock *, BlockChain *>;
238
239/// A chain of blocks which will be laid out contiguously.
240///
241/// This is the datastructure representing a chain of consecutive blocks that
242/// are profitable to layout together in order to maximize fallthrough
243/// probabilities and code locality. We also can use a block chain to represent
244/// a sequence of basic blocks which have some external (correctness)
245/// requirement for sequential layout.
246///
247/// Chains can be built around a single basic block and can be merged to grow
248/// them. They participate in a block-to-chain mapping, which is updated
249/// automatically as chains are merged together.
250class BlockChain {
251 /// The sequence of blocks belonging to this chain.
252 ///
253 /// This is the sequence of blocks for a particular chain. These will be laid
254 /// out in-order within the function.
255 SmallVector<MachineBasicBlock *, 4> Blocks;
256
257 /// A handle to the function-wide basic block to block chain mapping.
258 ///
259 /// This is retained in each block chain to simplify the computation of child
260 /// block chains for SCC-formation and iteration. We store the edges to child
261 /// basic blocks, and map them back to their associated chains using this
262 /// structure.
263 BlockToChainMapType &BlockToChain;
264
265public:
266 /// Construct a new BlockChain.
267 ///
268 /// This builds a new block chain representing a single basic block in the
269 /// function. It also registers itself as the chain that block participates
270 /// in with the BlockToChain mapping.
271 BlockChain(BlockToChainMapType &BlockToChain, MachineBasicBlock *BB)
272 : Blocks(1, BB), BlockToChain(BlockToChain) {
273 assert(BB && "Cannot create a chain with a null basic block");
274 BlockToChain[BB] = this;
275 }
276
277 /// Iterator over blocks within the chain.
278 using iterator = SmallVectorImpl<MachineBasicBlock *>::iterator;
279 using const_iterator = SmallVectorImpl<MachineBasicBlock *>::const_iterator;
280
281 /// Beginning of blocks within the chain.
282 iterator begin() { return Blocks.begin(); }
283 const_iterator begin() const { return Blocks.begin(); }
284
285 /// End of blocks within the chain.
286 iterator end() { return Blocks.end(); }
287 const_iterator end() const { return Blocks.end(); }
288
289 bool remove(MachineBasicBlock* BB) {
290 for(iterator i = begin(); i != end(); ++i) {
291 if (*i == BB) {
292 Blocks.erase(CI: i);
293 return true;
294 }
295 }
296 return false;
297 }
298
299 /// Merge a block chain into this one.
300 ///
301 /// This routine merges a block chain into this one. It takes care of forming
302 /// a contiguous sequence of basic blocks, updating the edge list, and
303 /// updating the block -> chain mapping. It does not free or tear down the
304 /// old chain, but the old chain's block list is no longer valid.
305 void merge(MachineBasicBlock *BB, BlockChain *Chain) {
306 assert(BB && "Can't merge a null block.");
307 assert(!Blocks.empty() && "Can't merge into an empty chain.");
308
309 // Fast path in case we don't have a chain already.
310 if (!Chain) {
311 assert(!BlockToChain[BB] &&
312 "Passed chain is null, but BB has entry in BlockToChain.");
313 Blocks.push_back(Elt: BB);
314 BlockToChain[BB] = this;
315 return;
316 }
317
318 assert(BB == *Chain->begin() && "Passed BB is not head of Chain.");
319 assert(Chain->begin() != Chain->end());
320
321 // Update the incoming blocks to point to this chain, and add them to the
322 // chain structure.
323 for (MachineBasicBlock *ChainBB : *Chain) {
324 Blocks.push_back(Elt: ChainBB);
325 assert(BlockToChain[ChainBB] == Chain && "Incoming blocks not in chain.");
326 BlockToChain[ChainBB] = this;
327 }
328 }
329
330#ifndef NDEBUG
331 /// Dump the blocks in this chain.
332 LLVM_DUMP_METHOD void dump() {
333 for (MachineBasicBlock *MBB : *this)
334 MBB->dump();
335 }
336#endif // NDEBUG
337
338 /// Count of predecessors of any block within the chain which have not
339 /// yet been scheduled. In general, we will delay scheduling this chain
340 /// until those predecessors are scheduled (or we find a sufficiently good
341 /// reason to override this heuristic.) Note that when forming loop chains,
342 /// blocks outside the loop are ignored and treated as if they were already
343 /// scheduled.
344 ///
345 /// Note: This field is reinitialized multiple times - once for each loop,
346 /// and then once for the function as a whole.
347 unsigned UnscheduledPredecessors = 0;
348};
349
350class MachineBlockPlacement : public MachineFunctionPass {
351 /// A type for a block filter set.
352 using BlockFilterSet = SmallSetVector<const MachineBasicBlock *, 16>;
353
354 /// Pair struct containing basic block and taildup profitability
355 struct BlockAndTailDupResult {
356 MachineBasicBlock *BB = nullptr;
357 bool ShouldTailDup;
358 };
359
360 /// Triple struct containing edge weight and the edge.
361 struct WeightedEdge {
362 BlockFrequency Weight;
363 MachineBasicBlock *Src = nullptr;
364 MachineBasicBlock *Dest = nullptr;
365 };
366
367 /// work lists of blocks that are ready to be laid out
368 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
369 SmallVector<MachineBasicBlock *, 16> EHPadWorkList;
370
371 /// Edges that have already been computed as optimal.
372 DenseMap<const MachineBasicBlock *, BlockAndTailDupResult> ComputedEdges;
373
374 /// Machine Function
375 MachineFunction *F = nullptr;
376
377 /// A handle to the branch probability pass.
378 const MachineBranchProbabilityInfo *MBPI = nullptr;
379
380 /// A handle to the function-wide block frequency pass.
381 std::unique_ptr<MBFIWrapper> MBFI;
382
383 /// A handle to the loop info.
384 MachineLoopInfo *MLI = nullptr;
385
386 /// Preferred loop exit.
387 /// Member variable for convenience. It may be removed by duplication deep
388 /// in the call stack.
389 MachineBasicBlock *PreferredLoopExit = nullptr;
390
391 /// A handle to the target's instruction info.
392 const TargetInstrInfo *TII = nullptr;
393
394 /// A handle to the target's lowering info.
395 const TargetLoweringBase *TLI = nullptr;
396
397 /// A handle to the post dominator tree.
398 MachinePostDominatorTree *MPDT = nullptr;
399
400 ProfileSummaryInfo *PSI = nullptr;
401
402 /// Duplicator used to duplicate tails during placement.
403 ///
404 /// Placement decisions can open up new tail duplication opportunities, but
405 /// since tail duplication affects placement decisions of later blocks, it
406 /// must be done inline.
407 TailDuplicator TailDup;
408
409 /// Partial tail duplication threshold.
410 BlockFrequency DupThreshold;
411
412 /// True: use block profile count to compute tail duplication cost.
413 /// False: use block frequency to compute tail duplication cost.
414 bool UseProfileCount = false;
415
416 /// Allocator and owner of BlockChain structures.
417 ///
418 /// We build BlockChains lazily while processing the loop structure of
419 /// a function. To reduce malloc traffic, we allocate them using this
420 /// slab-like allocator, and destroy them after the pass completes. An
421 /// important guarantee is that this allocator produces stable pointers to
422 /// the chains.
423 SpecificBumpPtrAllocator<BlockChain> ChainAllocator;
424
425 /// Function wide BasicBlock to BlockChain mapping.
426 ///
427 /// This mapping allows efficiently moving from any given basic block to the
428 /// BlockChain it participates in, if any. We use it to, among other things,
429 /// allow implicitly defining edges between chains as the existing edges
430 /// between basic blocks.
431 DenseMap<const MachineBasicBlock *, BlockChain *> BlockToChain;
432
433#ifndef NDEBUG
434 /// The set of basic blocks that have terminators that cannot be fully
435 /// analyzed. These basic blocks cannot be re-ordered safely by
436 /// MachineBlockPlacement, and we must preserve physical layout of these
437 /// blocks and their successors through the pass.
438 SmallPtrSet<MachineBasicBlock *, 4> BlocksWithUnanalyzableExits;
439#endif
440
441 /// Get block profile count or frequency according to UseProfileCount.
442 /// The return value is used to model tail duplication cost.
443 BlockFrequency getBlockCountOrFrequency(const MachineBasicBlock *BB) {
444 if (UseProfileCount) {
445 auto Count = MBFI->getBlockProfileCount(MBB: BB);
446 if (Count)
447 return BlockFrequency(*Count);
448 else
449 return BlockFrequency(0);
450 } else
451 return MBFI->getBlockFreq(MBB: BB);
452 }
453
454 /// Scale the DupThreshold according to basic block size.
455 BlockFrequency scaleThreshold(MachineBasicBlock *BB);
456 void initDupThreshold();
457
458 /// Decrease the UnscheduledPredecessors count for all blocks in chain, and
459 /// if the count goes to 0, add them to the appropriate work list.
460 void markChainSuccessors(
461 const BlockChain &Chain, const MachineBasicBlock *LoopHeaderBB,
462 const BlockFilterSet *BlockFilter = nullptr);
463
464 /// Decrease the UnscheduledPredecessors count for a single block, and
465 /// if the count goes to 0, add them to the appropriate work list.
466 void markBlockSuccessors(
467 const BlockChain &Chain, const MachineBasicBlock *BB,
468 const MachineBasicBlock *LoopHeaderBB,
469 const BlockFilterSet *BlockFilter = nullptr);
470
471 BranchProbability
472 collectViableSuccessors(
473 const MachineBasicBlock *BB, const BlockChain &Chain,
474 const BlockFilterSet *BlockFilter,
475 SmallVector<MachineBasicBlock *, 4> &Successors);
476 bool isBestSuccessor(MachineBasicBlock *BB, MachineBasicBlock *Pred,
477 BlockFilterSet *BlockFilter);
478 void findDuplicateCandidates(SmallVectorImpl<MachineBasicBlock *> &Candidates,
479 MachineBasicBlock *BB,
480 BlockFilterSet *BlockFilter);
481 bool repeatedlyTailDuplicateBlock(
482 MachineBasicBlock *BB, MachineBasicBlock *&LPred,
483 const MachineBasicBlock *LoopHeaderBB, BlockChain &Chain,
484 BlockFilterSet *BlockFilter,
485 MachineFunction::iterator &PrevUnplacedBlockIt,
486 BlockFilterSet::iterator &PrevUnplacedBlockInFilterIt);
487 bool
488 maybeTailDuplicateBlock(MachineBasicBlock *BB, MachineBasicBlock *LPred,
489 BlockChain &Chain, BlockFilterSet *BlockFilter,
490 MachineFunction::iterator &PrevUnplacedBlockIt,
491 BlockFilterSet::iterator &PrevUnplacedBlockInFilterIt,
492 bool &DuplicatedToLPred);
493 bool hasBetterLayoutPredecessor(
494 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
495 const BlockChain &SuccChain, BranchProbability SuccProb,
496 BranchProbability RealSuccProb, const BlockChain &Chain,
497 const BlockFilterSet *BlockFilter);
498 BlockAndTailDupResult selectBestSuccessor(
499 const MachineBasicBlock *BB, const BlockChain &Chain,
500 const BlockFilterSet *BlockFilter);
501 MachineBasicBlock *selectBestCandidateBlock(
502 const BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList);
503 MachineBasicBlock *
504 getFirstUnplacedBlock(const BlockChain &PlacedChain,
505 MachineFunction::iterator &PrevUnplacedBlockIt);
506 MachineBasicBlock *
507 getFirstUnplacedBlock(const BlockChain &PlacedChain,
508 BlockFilterSet::iterator &PrevUnplacedBlockInFilterIt,
509 const BlockFilterSet *BlockFilter);
510
511 /// Add a basic block to the work list if it is appropriate.
512 ///
513 /// If the optional parameter BlockFilter is provided, only MBB
514 /// present in the set will be added to the worklist. If nullptr
515 /// is provided, no filtering occurs.
516 void fillWorkLists(const MachineBasicBlock *MBB,
517 SmallPtrSetImpl<BlockChain *> &UpdatedPreds,
518 const BlockFilterSet *BlockFilter);
519
520 void buildChain(const MachineBasicBlock *BB, BlockChain &Chain,
521 BlockFilterSet *BlockFilter = nullptr);
522 bool canMoveBottomBlockToTop(const MachineBasicBlock *BottomBlock,
523 const MachineBasicBlock *OldTop);
524 bool hasViableTopFallthrough(const MachineBasicBlock *Top,
525 const BlockFilterSet &LoopBlockSet);
526 BlockFrequency TopFallThroughFreq(const MachineBasicBlock *Top,
527 const BlockFilterSet &LoopBlockSet);
528 BlockFrequency FallThroughGains(const MachineBasicBlock *NewTop,
529 const MachineBasicBlock *OldTop,
530 const MachineBasicBlock *ExitBB,
531 const BlockFilterSet &LoopBlockSet);
532 MachineBasicBlock *findBestLoopTopHelper(MachineBasicBlock *OldTop,
533 const MachineLoop &L, const BlockFilterSet &LoopBlockSet);
534 MachineBasicBlock *findBestLoopTop(
535 const MachineLoop &L, const BlockFilterSet &LoopBlockSet);
536 MachineBasicBlock *findBestLoopExit(
537 const MachineLoop &L, const BlockFilterSet &LoopBlockSet,
538 BlockFrequency &ExitFreq);
539 BlockFilterSet collectLoopBlockSet(const MachineLoop &L);
540 void buildLoopChains(const MachineLoop &L);
541 void rotateLoop(
542 BlockChain &LoopChain, const MachineBasicBlock *ExitingBB,
543 BlockFrequency ExitFreq, const BlockFilterSet &LoopBlockSet);
544 void rotateLoopWithProfile(
545 BlockChain &LoopChain, const MachineLoop &L,
546 const BlockFilterSet &LoopBlockSet);
547 void buildCFGChains();
548 void optimizeBranches();
549 void alignBlocks();
550 /// Returns true if a block should be tail-duplicated to increase fallthrough
551 /// opportunities.
552 bool shouldTailDuplicate(MachineBasicBlock *BB);
553 /// Check the edge frequencies to see if tail duplication will increase
554 /// fallthroughs.
555 bool isProfitableToTailDup(
556 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
557 BranchProbability QProb,
558 const BlockChain &Chain, const BlockFilterSet *BlockFilter);
559
560 /// Check for a trellis layout.
561 bool isTrellis(const MachineBasicBlock *BB,
562 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
563 const BlockChain &Chain, const BlockFilterSet *BlockFilter);
564
565 /// Get the best successor given a trellis layout.
566 BlockAndTailDupResult getBestTrellisSuccessor(
567 const MachineBasicBlock *BB,
568 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
569 BranchProbability AdjustedSumProb, const BlockChain &Chain,
570 const BlockFilterSet *BlockFilter);
571
572 /// Get the best pair of non-conflicting edges.
573 static std::pair<WeightedEdge, WeightedEdge> getBestNonConflictingEdges(
574 const MachineBasicBlock *BB,
575 MutableArrayRef<SmallVector<WeightedEdge, 8>> Edges);
576
577 /// Returns true if a block can tail duplicate into all unplaced
578 /// predecessors. Filters based on loop.
579 bool canTailDuplicateUnplacedPreds(
580 const MachineBasicBlock *BB, MachineBasicBlock *Succ,
581 const BlockChain &Chain, const BlockFilterSet *BlockFilter);
582
583 /// Find chains of triangles to tail-duplicate where a global analysis works,
584 /// but a local analysis would not find them.
585 void precomputeTriangleChains();
586
587 /// Apply a post-processing step optimizing block placement.
588 void applyExtTsp();
589
590 /// Modify the existing block placement in the function and adjust all jumps.
591 void assignBlockOrder(const std::vector<const MachineBasicBlock *> &NewOrder);
592
593 /// Create a single CFG chain from the current block order.
594 void createCFGChainExtTsp();
595
596public:
597 static char ID; // Pass identification, replacement for typeid
598
599 MachineBlockPlacement() : MachineFunctionPass(ID) {
600 initializeMachineBlockPlacementPass(*PassRegistry::getPassRegistry());
601 }
602
603 bool runOnMachineFunction(MachineFunction &F) override;
604
605 bool allowTailDupPlacement() const {
606 assert(F);
607 return TailDupPlacement && !F->getTarget().requiresStructuredCFG();
608 }
609
610 void getAnalysisUsage(AnalysisUsage &AU) const override {
611 AU.addRequired<MachineBranchProbabilityInfoWrapperPass>();
612 AU.addRequired<MachineBlockFrequencyInfoWrapperPass>();
613 if (TailDupPlacement)
614 AU.addRequired<MachinePostDominatorTreeWrapperPass>();
615 AU.addRequired<MachineLoopInfoWrapperPass>();
616 AU.addRequired<ProfileSummaryInfoWrapperPass>();
617 AU.addRequired<TargetPassConfig>();
618 MachineFunctionPass::getAnalysisUsage(AU);
619 }
620};
621
622} // end anonymous namespace
623
624char MachineBlockPlacement::ID = 0;
625
626char &llvm::MachineBlockPlacementID = MachineBlockPlacement::ID;
627
628INITIALIZE_PASS_BEGIN(MachineBlockPlacement, DEBUG_TYPE,
629 "Branch Probability Basic Block Placement", false, false)
630INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfoWrapperPass)
631INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfoWrapperPass)
632INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTreeWrapperPass)
633INITIALIZE_PASS_DEPENDENCY(MachineLoopInfoWrapperPass)
634INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
635INITIALIZE_PASS_END(MachineBlockPlacement, DEBUG_TYPE,
636 "Branch Probability Basic Block Placement", false, false)
637
638#ifndef NDEBUG
639/// Helper to print the name of a MBB.
640///
641/// Only used by debug logging.
642static std::string getBlockName(const MachineBasicBlock *BB) {
643 std::string Result;
644 raw_string_ostream OS(Result);
645 OS << printMBBReference(*BB);
646 OS << " ('" << BB->getName() << "')";
647 OS.flush();
648 return Result;
649}
650#endif
651
652/// Mark a chain's successors as having one fewer preds.
653///
654/// When a chain is being merged into the "placed" chain, this routine will
655/// quickly walk the successors of each block in the chain and mark them as
656/// having one fewer active predecessor. It also adds any successors of this
657/// chain which reach the zero-predecessor state to the appropriate worklist.
658void MachineBlockPlacement::markChainSuccessors(
659 const BlockChain &Chain, const MachineBasicBlock *LoopHeaderBB,
660 const BlockFilterSet *BlockFilter) {
661 // Walk all the blocks in this chain, marking their successors as having
662 // a predecessor placed.
663 for (MachineBasicBlock *MBB : Chain) {
664 markBlockSuccessors(Chain, BB: MBB, LoopHeaderBB, BlockFilter);
665 }
666}
667
668/// Mark a single block's successors as having one fewer preds.
669///
670/// Under normal circumstances, this is only called by markChainSuccessors,
671/// but if a block that was to be placed is completely tail-duplicated away,
672/// and was duplicated into the chain end, we need to redo markBlockSuccessors
673/// for just that block.
674void MachineBlockPlacement::markBlockSuccessors(
675 const BlockChain &Chain, const MachineBasicBlock *MBB,
676 const MachineBasicBlock *LoopHeaderBB, const BlockFilterSet *BlockFilter) {
677 // Add any successors for which this is the only un-placed in-loop
678 // predecessor to the worklist as a viable candidate for CFG-neutral
679 // placement. No subsequent placement of this block will violate the CFG
680 // shape, so we get to use heuristics to choose a favorable placement.
681 for (MachineBasicBlock *Succ : MBB->successors()) {
682 if (BlockFilter && !BlockFilter->count(key: Succ))
683 continue;
684 BlockChain &SuccChain = *BlockToChain[Succ];
685 // Disregard edges within a fixed chain, or edges to the loop header.
686 if (&Chain == &SuccChain || Succ == LoopHeaderBB)
687 continue;
688
689 // This is a cross-chain edge that is within the loop, so decrement the
690 // loop predecessor count of the destination chain.
691 if (SuccChain.UnscheduledPredecessors == 0 ||
692 --SuccChain.UnscheduledPredecessors > 0)
693 continue;
694
695 auto *NewBB = *SuccChain.begin();
696 if (NewBB->isEHPad())
697 EHPadWorkList.push_back(Elt: NewBB);
698 else
699 BlockWorkList.push_back(Elt: NewBB);
700 }
701}
702
703/// This helper function collects the set of successors of block
704/// \p BB that are allowed to be its layout successors, and return
705/// the total branch probability of edges from \p BB to those
706/// blocks.
707BranchProbability MachineBlockPlacement::collectViableSuccessors(
708 const MachineBasicBlock *BB, const BlockChain &Chain,
709 const BlockFilterSet *BlockFilter,
710 SmallVector<MachineBasicBlock *, 4> &Successors) {
711 // Adjust edge probabilities by excluding edges pointing to blocks that is
712 // either not in BlockFilter or is already in the current chain. Consider the
713 // following CFG:
714 //
715 // --->A
716 // | / \
717 // | B C
718 // | \ / \
719 // ----D E
720 //
721 // Assume A->C is very hot (>90%), and C->D has a 50% probability, then after
722 // A->C is chosen as a fall-through, D won't be selected as a successor of C
723 // due to CFG constraint (the probability of C->D is not greater than
724 // HotProb to break topo-order). If we exclude E that is not in BlockFilter
725 // when calculating the probability of C->D, D will be selected and we
726 // will get A C D B as the layout of this loop.
727 auto AdjustedSumProb = BranchProbability::getOne();
728 for (MachineBasicBlock *Succ : BB->successors()) {
729 bool SkipSucc = false;
730 if (Succ->isEHPad() || (BlockFilter && !BlockFilter->count(key: Succ))) {
731 SkipSucc = true;
732 } else {
733 BlockChain *SuccChain = BlockToChain[Succ];
734 if (SuccChain == &Chain) {
735 SkipSucc = true;
736 } else if (Succ != *SuccChain->begin()) {
737 LLVM_DEBUG(dbgs() << " " << getBlockName(Succ)
738 << " -> Mid chain!\n");
739 continue;
740 }
741 }
742 if (SkipSucc)
743 AdjustedSumProb -= MBPI->getEdgeProbability(Src: BB, Dst: Succ);
744 else
745 Successors.push_back(Elt: Succ);
746 }
747
748 return AdjustedSumProb;
749}
750
751/// The helper function returns the branch probability that is adjusted
752/// or normalized over the new total \p AdjustedSumProb.
753static BranchProbability
754getAdjustedProbability(BranchProbability OrigProb,
755 BranchProbability AdjustedSumProb) {
756 BranchProbability SuccProb;
757 uint32_t SuccProbN = OrigProb.getNumerator();
758 uint32_t SuccProbD = AdjustedSumProb.getNumerator();
759 if (SuccProbN >= SuccProbD)
760 SuccProb = BranchProbability::getOne();
761 else
762 SuccProb = BranchProbability(SuccProbN, SuccProbD);
763
764 return SuccProb;
765}
766
767/// Check if \p BB has exactly the successors in \p Successors.
768static bool
769hasSameSuccessors(MachineBasicBlock &BB,
770 SmallPtrSetImpl<const MachineBasicBlock *> &Successors) {
771 if (BB.succ_size() != Successors.size())
772 return false;
773 // We don't want to count self-loops
774 if (Successors.count(Ptr: &BB))
775 return false;
776 for (MachineBasicBlock *Succ : BB.successors())
777 if (!Successors.count(Ptr: Succ))
778 return false;
779 return true;
780}
781
782/// Check if a block should be tail duplicated to increase fallthrough
783/// opportunities.
784/// \p BB Block to check.
785bool MachineBlockPlacement::shouldTailDuplicate(MachineBasicBlock *BB) {
786 // Blocks with single successors don't create additional fallthrough
787 // opportunities. Don't duplicate them. TODO: When conditional exits are
788 // analyzable, allow them to be duplicated.
789 bool IsSimple = TailDup.isSimpleBB(TailBB: BB);
790
791 if (BB->succ_size() == 1)
792 return false;
793 return TailDup.shouldTailDuplicate(IsSimple, TailBB&: *BB);
794}
795
796/// Compare 2 BlockFrequency's with a small penalty for \p A.
797/// In order to be conservative, we apply a X% penalty to account for
798/// increased icache pressure and static heuristics. For small frequencies
799/// we use only the numerators to improve accuracy. For simplicity, we assume the
800/// penalty is less than 100%
801/// TODO(iteratee): Use 64-bit fixed point edge frequencies everywhere.
802static bool greaterWithBias(BlockFrequency A, BlockFrequency B,
803 BlockFrequency EntryFreq) {
804 BranchProbability ThresholdProb(TailDupPlacementPenalty, 100);
805 BlockFrequency Gain = A - B;
806 return (Gain / ThresholdProb) >= EntryFreq;
807}
808
809/// Check the edge frequencies to see if tail duplication will increase
810/// fallthroughs. It only makes sense to call this function when
811/// \p Succ would not be chosen otherwise. Tail duplication of \p Succ is
812/// always locally profitable if we would have picked \p Succ without
813/// considering duplication.
814bool MachineBlockPlacement::isProfitableToTailDup(
815 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
816 BranchProbability QProb,
817 const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
818 // We need to do a probability calculation to make sure this is profitable.
819 // First: does succ have a successor that post-dominates? This affects the
820 // calculation. The 2 relevant cases are:
821 // BB BB
822 // | \Qout | \Qout
823 // P| C |P C
824 // = C' = C'
825 // | /Qin | /Qin
826 // | / | /
827 // Succ Succ
828 // / \ | \ V
829 // U/ =V |U \
830 // / \ = D
831 // D E | /
832 // | /
833 // |/
834 // PDom
835 // '=' : Branch taken for that CFG edge
836 // In the second case, Placing Succ while duplicating it into C prevents the
837 // fallthrough of Succ into either D or PDom, because they now have C as an
838 // unplaced predecessor
839
840 // Start by figuring out which case we fall into
841 MachineBasicBlock *PDom = nullptr;
842 SmallVector<MachineBasicBlock *, 4> SuccSuccs;
843 // Only scan the relevant successors
844 auto AdjustedSuccSumProb =
845 collectViableSuccessors(BB: Succ, Chain, BlockFilter, Successors&: SuccSuccs);
846 BranchProbability PProb = MBPI->getEdgeProbability(Src: BB, Dst: Succ);
847 auto BBFreq = MBFI->getBlockFreq(MBB: BB);
848 auto SuccFreq = MBFI->getBlockFreq(MBB: Succ);
849 BlockFrequency P = BBFreq * PProb;
850 BlockFrequency Qout = BBFreq * QProb;
851 BlockFrequency EntryFreq = MBFI->getEntryFreq();
852 // If there are no more successors, it is profitable to copy, as it strictly
853 // increases fallthrough.
854 if (SuccSuccs.size() == 0)
855 return greaterWithBias(A: P, B: Qout, EntryFreq);
856
857 auto BestSuccSucc = BranchProbability::getZero();
858 // Find the PDom or the best Succ if no PDom exists.
859 for (MachineBasicBlock *SuccSucc : SuccSuccs) {
860 auto Prob = MBPI->getEdgeProbability(Src: Succ, Dst: SuccSucc);
861 if (Prob > BestSuccSucc)
862 BestSuccSucc = Prob;
863 if (PDom == nullptr)
864 if (MPDT->dominates(A: SuccSucc, B: Succ)) {
865 PDom = SuccSucc;
866 break;
867 }
868 }
869 // For the comparisons, we need to know Succ's best incoming edge that isn't
870 // from BB.
871 auto SuccBestPred = BlockFrequency(0);
872 for (MachineBasicBlock *SuccPred : Succ->predecessors()) {
873 if (SuccPred == Succ || SuccPred == BB
874 || BlockToChain[SuccPred] == &Chain
875 || (BlockFilter && !BlockFilter->count(key: SuccPred)))
876 continue;
877 auto Freq = MBFI->getBlockFreq(MBB: SuccPred)
878 * MBPI->getEdgeProbability(Src: SuccPred, Dst: Succ);
879 if (Freq > SuccBestPred)
880 SuccBestPred = Freq;
881 }
882 // Qin is Succ's best unplaced incoming edge that isn't BB
883 BlockFrequency Qin = SuccBestPred;
884 // If it doesn't have a post-dominating successor, here is the calculation:
885 // BB BB
886 // | \Qout | \
887 // P| C | =
888 // = C' | C
889 // | /Qin | |
890 // | / | C' (+Succ)
891 // Succ Succ /|
892 // / \ | \/ |
893 // U/ =V | == |
894 // / \ | / \|
895 // D E D E
896 // '=' : Branch taken for that CFG edge
897 // Cost in the first case is: P + V
898 // For this calculation, we always assume P > Qout. If Qout > P
899 // The result of this function will be ignored at the caller.
900 // Let F = SuccFreq - Qin
901 // Cost in the second case is: Qout + min(Qin, F) * U + max(Qin, F) * V
902
903 if (PDom == nullptr || !Succ->isSuccessor(MBB: PDom)) {
904 BranchProbability UProb = BestSuccSucc;
905 BranchProbability VProb = AdjustedSuccSumProb - UProb;
906 BlockFrequency F = SuccFreq - Qin;
907 BlockFrequency V = SuccFreq * VProb;
908 BlockFrequency QinU = std::min(a: Qin, b: F) * UProb;
909 BlockFrequency BaseCost = P + V;
910 BlockFrequency DupCost = Qout + QinU + std::max(a: Qin, b: F) * VProb;
911 return greaterWithBias(A: BaseCost, B: DupCost, EntryFreq);
912 }
913 BranchProbability UProb = MBPI->getEdgeProbability(Src: Succ, Dst: PDom);
914 BranchProbability VProb = AdjustedSuccSumProb - UProb;
915 BlockFrequency U = SuccFreq * UProb;
916 BlockFrequency V = SuccFreq * VProb;
917 BlockFrequency F = SuccFreq - Qin;
918 // If there is a post-dominating successor, here is the calculation:
919 // BB BB BB BB
920 // | \Qout | \ | \Qout | \
921 // |P C | = |P C | =
922 // = C' |P C = C' |P C
923 // | /Qin | | | /Qin | |
924 // | / | C' (+Succ) | / | C' (+Succ)
925 // Succ Succ /| Succ Succ /|
926 // | \ V | \/ | | \ V | \/ |
927 // |U \ |U /\ =? |U = |U /\ |
928 // = D = = =?| | D | = =|
929 // | / |/ D | / |/ D
930 // | / | / | = | /
931 // |/ | / |/ | =
932 // Dom Dom Dom Dom
933 // '=' : Branch taken for that CFG edge
934 // The cost for taken branches in the first case is P + U
935 // Let F = SuccFreq - Qin
936 // The cost in the second case (assuming independence), given the layout:
937 // BB, Succ, (C+Succ), D, Dom or the layout:
938 // BB, Succ, D, Dom, (C+Succ)
939 // is Qout + max(F, Qin) * U + min(F, Qin)
940 // compare P + U vs Qout + P * U + Qin.
941 //
942 // The 3rd and 4th cases cover when Dom would be chosen to follow Succ.
943 //
944 // For the 3rd case, the cost is P + 2 * V
945 // For the 4th case, the cost is Qout + min(Qin, F) * U + max(Qin, F) * V + V
946 // We choose 4 over 3 when (P + V) > Qout + min(Qin, F) * U + max(Qin, F) * V
947 if (UProb > AdjustedSuccSumProb / 2 &&
948 !hasBetterLayoutPredecessor(BB: Succ, Succ: PDom, SuccChain: *BlockToChain[PDom], SuccProb: UProb, RealSuccProb: UProb,
949 Chain, BlockFilter))
950 // Cases 3 & 4
951 return greaterWithBias(
952 A: (P + V), B: (Qout + std::max(a: Qin, b: F) * VProb + std::min(a: Qin, b: F) * UProb),
953 EntryFreq);
954 // Cases 1 & 2
955 return greaterWithBias(A: (P + U),
956 B: (Qout + std::min(a: Qin, b: F) * AdjustedSuccSumProb +
957 std::max(a: Qin, b: F) * UProb),
958 EntryFreq);
959}
960
961/// Check for a trellis layout. \p BB is the upper part of a trellis if its
962/// successors form the lower part of a trellis. A successor set S forms the
963/// lower part of a trellis if all of the predecessors of S are either in S or
964/// have all of S as successors. We ignore trellises where BB doesn't have 2
965/// successors because for fewer than 2, it's trivial, and for 3 or greater they
966/// are very uncommon and complex to compute optimally. Allowing edges within S
967/// is not strictly a trellis, but the same algorithm works, so we allow it.
968bool MachineBlockPlacement::isTrellis(
969 const MachineBasicBlock *BB,
970 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
971 const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
972 // Technically BB could form a trellis with branching factor higher than 2.
973 // But that's extremely uncommon.
974 if (BB->succ_size() != 2 || ViableSuccs.size() != 2)
975 return false;
976
977 SmallPtrSet<const MachineBasicBlock *, 2> Successors(BB->succ_begin(),
978 BB->succ_end());
979 // To avoid reviewing the same predecessors twice.
980 SmallPtrSet<const MachineBasicBlock *, 8> SeenPreds;
981
982 for (MachineBasicBlock *Succ : ViableSuccs) {
983 int PredCount = 0;
984 for (auto *SuccPred : Succ->predecessors()) {
985 // Allow triangle successors, but don't count them.
986 if (Successors.count(Ptr: SuccPred)) {
987 // Make sure that it is actually a triangle.
988 for (MachineBasicBlock *CheckSucc : SuccPred->successors())
989 if (!Successors.count(Ptr: CheckSucc))
990 return false;
991 continue;
992 }
993 const BlockChain *PredChain = BlockToChain[SuccPred];
994 if (SuccPred == BB || (BlockFilter && !BlockFilter->count(key: SuccPred)) ||
995 PredChain == &Chain || PredChain == BlockToChain[Succ])
996 continue;
997 ++PredCount;
998 // Perform the successor check only once.
999 if (!SeenPreds.insert(Ptr: SuccPred).second)
1000 continue;
1001 if (!hasSameSuccessors(BB&: *SuccPred, Successors))
1002 return false;
1003 }
1004 // If one of the successors has only BB as a predecessor, it is not a
1005 // trellis.
1006 if (PredCount < 1)
1007 return false;
1008 }
1009 return true;
1010}
1011
1012/// Pick the highest total weight pair of edges that can both be laid out.
1013/// The edges in \p Edges[0] are assumed to have a different destination than
1014/// the edges in \p Edges[1]. Simple counting shows that the best pair is either
1015/// the individual highest weight edges to the 2 different destinations, or in
1016/// case of a conflict, one of them should be replaced with a 2nd best edge.
1017std::pair<MachineBlockPlacement::WeightedEdge,
1018 MachineBlockPlacement::WeightedEdge>
1019MachineBlockPlacement::getBestNonConflictingEdges(
1020 const MachineBasicBlock *BB,
1021 MutableArrayRef<SmallVector<MachineBlockPlacement::WeightedEdge, 8>>
1022 Edges) {
1023 // Sort the edges, and then for each successor, find the best incoming
1024 // predecessor. If the best incoming predecessors aren't the same,
1025 // then that is clearly the best layout. If there is a conflict, one of the
1026 // successors will have to fallthrough from the second best predecessor. We
1027 // compare which combination is better overall.
1028
1029 // Sort for highest frequency.
1030 auto Cmp = [](WeightedEdge A, WeightedEdge B) { return A.Weight > B.Weight; };
1031
1032 llvm::stable_sort(Range&: Edges[0], C: Cmp);
1033 llvm::stable_sort(Range&: Edges[1], C: Cmp);
1034 auto BestA = Edges[0].begin();
1035 auto BestB = Edges[1].begin();
1036 // Arrange for the correct answer to be in BestA and BestB
1037 // If the 2 best edges don't conflict, the answer is already there.
1038 if (BestA->Src == BestB->Src) {
1039 // Compare the total fallthrough of (Best + Second Best) for both pairs
1040 auto SecondBestA = std::next(x: BestA);
1041 auto SecondBestB = std::next(x: BestB);
1042 BlockFrequency BestAScore = BestA->Weight + SecondBestB->Weight;
1043 BlockFrequency BestBScore = BestB->Weight + SecondBestA->Weight;
1044 if (BestAScore < BestBScore)
1045 BestA = SecondBestA;
1046 else
1047 BestB = SecondBestB;
1048 }
1049 // Arrange for the BB edge to be in BestA if it exists.
1050 if (BestB->Src == BB)
1051 std::swap(a&: BestA, b&: BestB);
1052 return std::make_pair(x&: *BestA, y&: *BestB);
1053}
1054
1055/// Get the best successor from \p BB based on \p BB being part of a trellis.
1056/// We only handle trellises with 2 successors, so the algorithm is
1057/// straightforward: Find the best pair of edges that don't conflict. We find
1058/// the best incoming edge for each successor in the trellis. If those conflict,
1059/// we consider which of them should be replaced with the second best.
1060/// Upon return the two best edges will be in \p BestEdges. If one of the edges
1061/// comes from \p BB, it will be in \p BestEdges[0]
1062MachineBlockPlacement::BlockAndTailDupResult
1063MachineBlockPlacement::getBestTrellisSuccessor(
1064 const MachineBasicBlock *BB,
1065 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
1066 BranchProbability AdjustedSumProb, const BlockChain &Chain,
1067 const BlockFilterSet *BlockFilter) {
1068
1069 BlockAndTailDupResult Result = {.BB: nullptr, .ShouldTailDup: false};
1070 SmallPtrSet<const MachineBasicBlock *, 4> Successors(BB->succ_begin(),
1071 BB->succ_end());
1072
1073 // We assume size 2 because it's common. For general n, we would have to do
1074 // the Hungarian algorithm, but it's not worth the complexity because more
1075 // than 2 successors is fairly uncommon, and a trellis even more so.
1076 if (Successors.size() != 2 || ViableSuccs.size() != 2)
1077 return Result;
1078
1079 // Collect the edge frequencies of all edges that form the trellis.
1080 SmallVector<WeightedEdge, 8> Edges[2];
1081 int SuccIndex = 0;
1082 for (auto *Succ : ViableSuccs) {
1083 for (MachineBasicBlock *SuccPred : Succ->predecessors()) {
1084 // Skip any placed predecessors that are not BB
1085 if (SuccPred != BB)
1086 if ((BlockFilter && !BlockFilter->count(key: SuccPred)) ||
1087 BlockToChain[SuccPred] == &Chain ||
1088 BlockToChain[SuccPred] == BlockToChain[Succ])
1089 continue;
1090 BlockFrequency EdgeFreq = MBFI->getBlockFreq(MBB: SuccPred) *
1091 MBPI->getEdgeProbability(Src: SuccPred, Dst: Succ);
1092 Edges[SuccIndex].push_back(Elt: {.Weight: EdgeFreq, .Src: SuccPred, .Dest: Succ});
1093 }
1094 ++SuccIndex;
1095 }
1096
1097 // Pick the best combination of 2 edges from all the edges in the trellis.
1098 WeightedEdge BestA, BestB;
1099 std::tie(args&: BestA, args&: BestB) = getBestNonConflictingEdges(BB, Edges);
1100
1101 if (BestA.Src != BB) {
1102 // If we have a trellis, and BB doesn't have the best fallthrough edges,
1103 // we shouldn't choose any successor. We've already looked and there's a
1104 // better fallthrough edge for all the successors.
1105 LLVM_DEBUG(dbgs() << "Trellis, but not one of the chosen edges.\n");
1106 return Result;
1107 }
1108
1109 // Did we pick the triangle edge? If tail-duplication is profitable, do
1110 // that instead. Otherwise merge the triangle edge now while we know it is
1111 // optimal.
1112 if (BestA.Dest == BestB.Src) {
1113 // The edges are BB->Succ1->Succ2, and we're looking to see if BB->Succ2
1114 // would be better.
1115 MachineBasicBlock *Succ1 = BestA.Dest;
1116 MachineBasicBlock *Succ2 = BestB.Dest;
1117 // Check to see if tail-duplication would be profitable.
1118 if (allowTailDupPlacement() && shouldTailDuplicate(BB: Succ2) &&
1119 canTailDuplicateUnplacedPreds(BB, Succ: Succ2, Chain, BlockFilter) &&
1120 isProfitableToTailDup(BB, Succ: Succ2, QProb: MBPI->getEdgeProbability(Src: BB, Dst: Succ1),
1121 Chain, BlockFilter)) {
1122 LLVM_DEBUG(BranchProbability Succ2Prob = getAdjustedProbability(
1123 MBPI->getEdgeProbability(BB, Succ2), AdjustedSumProb);
1124 dbgs() << " Selected: " << getBlockName(Succ2)
1125 << ", probability: " << Succ2Prob
1126 << " (Tail Duplicate)\n");
1127 Result.BB = Succ2;
1128 Result.ShouldTailDup = true;
1129 return Result;
1130 }
1131 }
1132 // We have already computed the optimal edge for the other side of the
1133 // trellis.
1134 ComputedEdges[BestB.Src] = { .BB: BestB.Dest, .ShouldTailDup: false };
1135
1136 auto TrellisSucc = BestA.Dest;
1137 LLVM_DEBUG(BranchProbability SuccProb = getAdjustedProbability(
1138 MBPI->getEdgeProbability(BB, TrellisSucc), AdjustedSumProb);
1139 dbgs() << " Selected: " << getBlockName(TrellisSucc)
1140 << ", probability: " << SuccProb << " (Trellis)\n");
1141 Result.BB = TrellisSucc;
1142 return Result;
1143}
1144
1145/// When the option allowTailDupPlacement() is on, this method checks if the
1146/// fallthrough candidate block \p Succ (of block \p BB) can be tail-duplicated
1147/// into all of its unplaced, unfiltered predecessors, that are not BB.
1148bool MachineBlockPlacement::canTailDuplicateUnplacedPreds(
1149 const MachineBasicBlock *BB, MachineBasicBlock *Succ,
1150 const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
1151 if (!shouldTailDuplicate(BB: Succ))
1152 return false;
1153
1154 // The result of canTailDuplicate.
1155 bool Duplicate = true;
1156 // Number of possible duplication.
1157 unsigned int NumDup = 0;
1158
1159 // For CFG checking.
1160 SmallPtrSet<const MachineBasicBlock *, 4> Successors(BB->succ_begin(),
1161 BB->succ_end());
1162 for (MachineBasicBlock *Pred : Succ->predecessors()) {
1163 // Make sure all unplaced and unfiltered predecessors can be
1164 // tail-duplicated into.
1165 // Skip any blocks that are already placed or not in this loop.
1166 if (Pred == BB || (BlockFilter && !BlockFilter->count(key: Pred))
1167 || (BlockToChain[Pred] == &Chain && !Succ->succ_empty()))
1168 continue;
1169 if (!TailDup.canTailDuplicate(TailBB: Succ, PredBB: Pred)) {
1170 if (Successors.size() > 1 && hasSameSuccessors(BB&: *Pred, Successors))
1171 // This will result in a trellis after tail duplication, so we don't
1172 // need to copy Succ into this predecessor. In the presence
1173 // of a trellis tail duplication can continue to be profitable.
1174 // For example:
1175 // A A
1176 // |\ |\
1177 // | \ | \
1178 // | C | C+BB
1179 // | / | |
1180 // |/ | |
1181 // BB => BB |
1182 // |\ |\/|
1183 // | \ |/\|
1184 // | D | D
1185 // | / | /
1186 // |/ |/
1187 // Succ Succ
1188 //
1189 // After BB was duplicated into C, the layout looks like the one on the
1190 // right. BB and C now have the same successors. When considering
1191 // whether Succ can be duplicated into all its unplaced predecessors, we
1192 // ignore C.
1193 // We can do this because C already has a profitable fallthrough, namely
1194 // D. TODO(iteratee): ignore sufficiently cold predecessors for
1195 // duplication and for this test.
1196 //
1197 // This allows trellises to be laid out in 2 separate chains
1198 // (A,B,Succ,...) and later (C,D,...) This is a reasonable heuristic
1199 // because it allows the creation of 2 fallthrough paths with links
1200 // between them, and we correctly identify the best layout for these
1201 // CFGs. We want to extend trellises that the user created in addition
1202 // to trellises created by tail-duplication, so we just look for the
1203 // CFG.
1204 continue;
1205 Duplicate = false;
1206 continue;
1207 }
1208 NumDup++;
1209 }
1210
1211 // No possible duplication in current filter set.
1212 if (NumDup == 0)
1213 return false;
1214
1215 // If profile information is available, findDuplicateCandidates can do more
1216 // precise benefit analysis.
1217 if (F->getFunction().hasProfileData())
1218 return true;
1219
1220 // This is mainly for function exit BB.
1221 // The integrated tail duplication is really designed for increasing
1222 // fallthrough from predecessors from Succ to its successors. We may need
1223 // other machanism to handle different cases.
1224 if (Succ->succ_empty())
1225 return true;
1226
1227 // Plus the already placed predecessor.
1228 NumDup++;
1229
1230 // If the duplication candidate has more unplaced predecessors than
1231 // successors, the extra duplication can't bring more fallthrough.
1232 //
1233 // Pred1 Pred2 Pred3
1234 // \ | /
1235 // \ | /
1236 // \ | /
1237 // Dup
1238 // / \
1239 // / \
1240 // Succ1 Succ2
1241 //
1242 // In this example Dup has 2 successors and 3 predecessors, duplication of Dup
1243 // can increase the fallthrough from Pred1 to Succ1 and from Pred2 to Succ2,
1244 // but the duplication into Pred3 can't increase fallthrough.
1245 //
1246 // A small number of extra duplication may not hurt too much. We need a better
1247 // heuristic to handle it.
1248 if ((NumDup > Succ->succ_size()) || !Duplicate)
1249 return false;
1250
1251 return true;
1252}
1253
1254/// Find chains of triangles where we believe it would be profitable to
1255/// tail-duplicate them all, but a local analysis would not find them.
1256/// There are 3 ways this can be profitable:
1257/// 1) The post-dominators marked 50% are actually taken 55% (This shrinks with
1258/// longer chains)
1259/// 2) The chains are statically correlated. Branch probabilities have a very
1260/// U-shaped distribution.
1261/// [http://nrs.harvard.edu/urn-3:HUL.InstRepos:24015805]
1262/// If the branches in a chain are likely to be from the same side of the
1263/// distribution as their predecessor, but are independent at runtime, this
1264/// transformation is profitable. (Because the cost of being wrong is a small
1265/// fixed cost, unlike the standard triangle layout where the cost of being
1266/// wrong scales with the # of triangles.)
1267/// 3) The chains are dynamically correlated. If the probability that a previous
1268/// branch was taken positively influences whether the next branch will be
1269/// taken
1270/// We believe that 2 and 3 are common enough to justify the small margin in 1.
1271void MachineBlockPlacement::precomputeTriangleChains() {
1272 struct TriangleChain {
1273 std::vector<MachineBasicBlock *> Edges;
1274
1275 TriangleChain(MachineBasicBlock *src, MachineBasicBlock *dst)
1276 : Edges({src, dst}) {}
1277
1278 void append(MachineBasicBlock *dst) {
1279 assert(getKey()->isSuccessor(dst) &&
1280 "Attempting to append a block that is not a successor.");
1281 Edges.push_back(x: dst);
1282 }
1283
1284 unsigned count() const { return Edges.size() - 1; }
1285
1286 MachineBasicBlock *getKey() const {
1287 return Edges.back();
1288 }
1289 };
1290
1291 if (TriangleChainCount == 0)
1292 return;
1293
1294 LLVM_DEBUG(dbgs() << "Pre-computing triangle chains.\n");
1295 // Map from last block to the chain that contains it. This allows us to extend
1296 // chains as we find new triangles.
1297 DenseMap<const MachineBasicBlock *, TriangleChain> TriangleChainMap;
1298 for (MachineBasicBlock &BB : *F) {
1299 // If BB doesn't have 2 successors, it doesn't start a triangle.
1300 if (BB.succ_size() != 2)
1301 continue;
1302 MachineBasicBlock *PDom = nullptr;
1303 for (MachineBasicBlock *Succ : BB.successors()) {
1304 if (!MPDT->dominates(A: Succ, B: &BB))
1305 continue;
1306 PDom = Succ;
1307 break;
1308 }
1309 // If BB doesn't have a post-dominating successor, it doesn't form a
1310 // triangle.
1311 if (PDom == nullptr)
1312 continue;
1313 // If PDom has a hint that it is low probability, skip this triangle.
1314 if (MBPI->getEdgeProbability(Src: &BB, Dst: PDom) < BranchProbability(50, 100))
1315 continue;
1316 // If PDom isn't eligible for duplication, this isn't the kind of triangle
1317 // we're looking for.
1318 if (!shouldTailDuplicate(BB: PDom))
1319 continue;
1320 bool CanTailDuplicate = true;
1321 // If PDom can't tail-duplicate into it's non-BB predecessors, then this
1322 // isn't the kind of triangle we're looking for.
1323 for (MachineBasicBlock* Pred : PDom->predecessors()) {
1324 if (Pred == &BB)
1325 continue;
1326 if (!TailDup.canTailDuplicate(TailBB: PDom, PredBB: Pred)) {
1327 CanTailDuplicate = false;
1328 break;
1329 }
1330 }
1331 // If we can't tail-duplicate PDom to its predecessors, then skip this
1332 // triangle.
1333 if (!CanTailDuplicate)
1334 continue;
1335
1336 // Now we have an interesting triangle. Insert it if it's not part of an
1337 // existing chain.
1338 // Note: This cannot be replaced with a call insert() or emplace() because
1339 // the find key is BB, but the insert/emplace key is PDom.
1340 auto Found = TriangleChainMap.find(Val: &BB);
1341 // If it is, remove the chain from the map, grow it, and put it back in the
1342 // map with the end as the new key.
1343 if (Found != TriangleChainMap.end()) {
1344 TriangleChain Chain = std::move(Found->second);
1345 TriangleChainMap.erase(I: Found);
1346 Chain.append(dst: PDom);
1347 TriangleChainMap.insert(KV: std::make_pair(x: Chain.getKey(), y: std::move(Chain)));
1348 } else {
1349 auto InsertResult = TriangleChainMap.try_emplace(Key: PDom, Args: &BB, Args&: PDom);
1350 assert(InsertResult.second && "Block seen twice.");
1351 (void)InsertResult;
1352 }
1353 }
1354
1355 // Iterating over a DenseMap is safe here, because the only thing in the body
1356 // of the loop is inserting into another DenseMap (ComputedEdges).
1357 // ComputedEdges is never iterated, so this doesn't lead to non-determinism.
1358 for (auto &ChainPair : TriangleChainMap) {
1359 TriangleChain &Chain = ChainPair.second;
1360 // Benchmarking has shown that due to branch correlation duplicating 2 or
1361 // more triangles is profitable, despite the calculations assuming
1362 // independence.
1363 if (Chain.count() < TriangleChainCount)
1364 continue;
1365 MachineBasicBlock *dst = Chain.Edges.back();
1366 Chain.Edges.pop_back();
1367 for (MachineBasicBlock *src : reverse(C&: Chain.Edges)) {
1368 LLVM_DEBUG(dbgs() << "Marking edge: " << getBlockName(src) << "->"
1369 << getBlockName(dst)
1370 << " as pre-computed based on triangles.\n");
1371
1372 auto InsertResult = ComputedEdges.insert(KV: {src, {.BB: dst, .ShouldTailDup: true}});
1373 assert(InsertResult.second && "Block seen twice.");
1374 (void)InsertResult;
1375
1376 dst = src;
1377 }
1378 }
1379}
1380
1381// When profile is not present, return the StaticLikelyProb.
1382// When profile is available, we need to handle the triangle-shape CFG.
1383static BranchProbability getLayoutSuccessorProbThreshold(
1384 const MachineBasicBlock *BB) {
1385 if (!BB->getParent()->getFunction().hasProfileData())
1386 return BranchProbability(StaticLikelyProb, 100);
1387 if (BB->succ_size() == 2) {
1388 const MachineBasicBlock *Succ1 = *BB->succ_begin();
1389 const MachineBasicBlock *Succ2 = *(BB->succ_begin() + 1);
1390 if (Succ1->isSuccessor(MBB: Succ2) || Succ2->isSuccessor(MBB: Succ1)) {
1391 /* See case 1 below for the cost analysis. For BB->Succ to
1392 * be taken with smaller cost, the following needs to hold:
1393 * Prob(BB->Succ) > 2 * Prob(BB->Pred)
1394 * So the threshold T in the calculation below
1395 * (1-T) * Prob(BB->Succ) > T * Prob(BB->Pred)
1396 * So T / (1 - T) = 2, Yielding T = 2/3
1397 * Also adding user specified branch bias, we have
1398 * T = (2/3)*(ProfileLikelyProb/50)
1399 * = (2*ProfileLikelyProb)/150)
1400 */
1401 return BranchProbability(2 * ProfileLikelyProb, 150);
1402 }
1403 }
1404 return BranchProbability(ProfileLikelyProb, 100);
1405}
1406
1407/// Checks to see if the layout candidate block \p Succ has a better layout
1408/// predecessor than \c BB. If yes, returns true.
1409/// \p SuccProb: The probability adjusted for only remaining blocks.
1410/// Only used for logging
1411/// \p RealSuccProb: The un-adjusted probability.
1412/// \p Chain: The chain that BB belongs to and Succ is being considered for.
1413/// \p BlockFilter: if non-null, the set of blocks that make up the loop being
1414/// considered
1415bool MachineBlockPlacement::hasBetterLayoutPredecessor(
1416 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
1417 const BlockChain &SuccChain, BranchProbability SuccProb,
1418 BranchProbability RealSuccProb, const BlockChain &Chain,
1419 const BlockFilterSet *BlockFilter) {
1420
1421 // There isn't a better layout when there are no unscheduled predecessors.
1422 if (SuccChain.UnscheduledPredecessors == 0)
1423 return false;
1424
1425 // There are two basic scenarios here:
1426 // -------------------------------------
1427 // Case 1: triangular shape CFG (if-then):
1428 // BB
1429 // | \
1430 // | \
1431 // | Pred
1432 // | /
1433 // Succ
1434 // In this case, we are evaluating whether to select edge -> Succ, e.g.
1435 // set Succ as the layout successor of BB. Picking Succ as BB's
1436 // successor breaks the CFG constraints (FIXME: define these constraints).
1437 // With this layout, Pred BB
1438 // is forced to be outlined, so the overall cost will be cost of the
1439 // branch taken from BB to Pred, plus the cost of back taken branch
1440 // from Pred to Succ, as well as the additional cost associated
1441 // with the needed unconditional jump instruction from Pred To Succ.
1442
1443 // The cost of the topological order layout is the taken branch cost
1444 // from BB to Succ, so to make BB->Succ a viable candidate, the following
1445 // must hold:
1446 // 2 * freq(BB->Pred) * taken_branch_cost + unconditional_jump_cost
1447 // < freq(BB->Succ) * taken_branch_cost.
1448 // Ignoring unconditional jump cost, we get
1449 // freq(BB->Succ) > 2 * freq(BB->Pred), i.e.,
1450 // prob(BB->Succ) > 2 * prob(BB->Pred)
1451 //
1452 // When real profile data is available, we can precisely compute the
1453 // probability threshold that is needed for edge BB->Succ to be considered.
1454 // Without profile data, the heuristic requires the branch bias to be
1455 // a lot larger to make sure the signal is very strong (e.g. 80% default).
1456 // -----------------------------------------------------------------
1457 // Case 2: diamond like CFG (if-then-else):
1458 // S
1459 // / \
1460 // | \
1461 // BB Pred
1462 // \ /
1463 // Succ
1464 // ..
1465 //
1466 // The current block is BB and edge BB->Succ is now being evaluated.
1467 // Note that edge S->BB was previously already selected because
1468 // prob(S->BB) > prob(S->Pred).
1469 // At this point, 2 blocks can be placed after BB: Pred or Succ. If we
1470 // choose Pred, we will have a topological ordering as shown on the left
1471 // in the picture below. If we choose Succ, we have the solution as shown
1472 // on the right:
1473 //
1474 // topo-order:
1475 //
1476 // S----- ---S
1477 // | | | |
1478 // ---BB | | BB
1479 // | | | |
1480 // | Pred-- | Succ--
1481 // | | | |
1482 // ---Succ ---Pred--
1483 //
1484 // cost = freq(S->Pred) + freq(BB->Succ) cost = 2 * freq (S->Pred)
1485 // = freq(S->Pred) + freq(S->BB)
1486 //
1487 // If we have profile data (i.e, branch probabilities can be trusted), the
1488 // cost (number of taken branches) with layout S->BB->Succ->Pred is 2 *
1489 // freq(S->Pred) while the cost of topo order is freq(S->Pred) + freq(S->BB).
1490 // We know Prob(S->BB) > Prob(S->Pred), so freq(S->BB) > freq(S->Pred), which
1491 // means the cost of topological order is greater.
1492 // When profile data is not available, however, we need to be more
1493 // conservative. If the branch prediction is wrong, breaking the topo-order
1494 // will actually yield a layout with large cost. For this reason, we need
1495 // strong biased branch at block S with Prob(S->BB) in order to select
1496 // BB->Succ. This is equivalent to looking the CFG backward with backward
1497 // edge: Prob(Succ->BB) needs to >= HotProb in order to be selected (without
1498 // profile data).
1499 // --------------------------------------------------------------------------
1500 // Case 3: forked diamond
1501 // S
1502 // / \
1503 // / \
1504 // BB Pred
1505 // | \ / |
1506 // | \ / |
1507 // | X |
1508 // | / \ |
1509 // | / \ |
1510 // S1 S2
1511 //
1512 // The current block is BB and edge BB->S1 is now being evaluated.
1513 // As above S->BB was already selected because
1514 // prob(S->BB) > prob(S->Pred). Assume that prob(BB->S1) >= prob(BB->S2).
1515 //
1516 // topo-order:
1517 //
1518 // S-------| ---S
1519 // | | | |
1520 // ---BB | | BB
1521 // | | | |
1522 // | Pred----| | S1----
1523 // | | | |
1524 // --(S1 or S2) ---Pred--
1525 // |
1526 // S2
1527 //
1528 // topo-cost = freq(S->Pred) + freq(BB->S1) + freq(BB->S2)
1529 // + min(freq(Pred->S1), freq(Pred->S2))
1530 // Non-topo-order cost:
1531 // non-topo-cost = 2 * freq(S->Pred) + freq(BB->S2).
1532 // To be conservative, we can assume that min(freq(Pred->S1), freq(Pred->S2))
1533 // is 0. Then the non topo layout is better when
1534 // freq(S->Pred) < freq(BB->S1).
1535 // This is exactly what is checked below.
1536 // Note there are other shapes that apply (Pred may not be a single block,
1537 // but they all fit this general pattern.)
1538 BranchProbability HotProb = getLayoutSuccessorProbThreshold(BB);
1539
1540 // Make sure that a hot successor doesn't have a globally more
1541 // important predecessor.
1542 BlockFrequency CandidateEdgeFreq = MBFI->getBlockFreq(MBB: BB) * RealSuccProb;
1543 bool BadCFGConflict = false;
1544
1545 for (MachineBasicBlock *Pred : Succ->predecessors()) {
1546 BlockChain *PredChain = BlockToChain[Pred];
1547 if (Pred == Succ || PredChain == &SuccChain ||
1548 (BlockFilter && !BlockFilter->count(key: Pred)) ||
1549 PredChain == &Chain || Pred != *std::prev(x: PredChain->end()) ||
1550 // This check is redundant except for look ahead. This function is
1551 // called for lookahead by isProfitableToTailDup when BB hasn't been
1552 // placed yet.
1553 (Pred == BB))
1554 continue;
1555 // Do backward checking.
1556 // For all cases above, we need a backward checking to filter out edges that
1557 // are not 'strongly' biased.
1558 // BB Pred
1559 // \ /
1560 // Succ
1561 // We select edge BB->Succ if
1562 // freq(BB->Succ) > freq(Succ) * HotProb
1563 // i.e. freq(BB->Succ) > freq(BB->Succ) * HotProb + freq(Pred->Succ) *
1564 // HotProb
1565 // i.e. freq((BB->Succ) * (1 - HotProb) > freq(Pred->Succ) * HotProb
1566 // Case 1 is covered too, because the first equation reduces to:
1567 // prob(BB->Succ) > HotProb. (freq(Succ) = freq(BB) for a triangle)
1568 BlockFrequency PredEdgeFreq =
1569 MBFI->getBlockFreq(MBB: Pred) * MBPI->getEdgeProbability(Src: Pred, Dst: Succ);
1570 if (PredEdgeFreq * HotProb >= CandidateEdgeFreq * HotProb.getCompl()) {
1571 BadCFGConflict = true;
1572 break;
1573 }
1574 }
1575
1576 if (BadCFGConflict) {
1577 LLVM_DEBUG(dbgs() << " Not a candidate: " << getBlockName(Succ) << " -> "
1578 << SuccProb << " (prob) (non-cold CFG conflict)\n");
1579 return true;
1580 }
1581
1582 return false;
1583}
1584
1585/// Select the best successor for a block.
1586///
1587/// This looks across all successors of a particular block and attempts to
1588/// select the "best" one to be the layout successor. It only considers direct
1589/// successors which also pass the block filter. It will attempt to avoid
1590/// breaking CFG structure, but cave and break such structures in the case of
1591/// very hot successor edges.
1592///
1593/// \returns The best successor block found, or null if none are viable, along
1594/// with a boolean indicating if tail duplication is necessary.
1595MachineBlockPlacement::BlockAndTailDupResult
1596MachineBlockPlacement::selectBestSuccessor(
1597 const MachineBasicBlock *BB, const BlockChain &Chain,
1598 const BlockFilterSet *BlockFilter) {
1599 const BranchProbability HotProb(StaticLikelyProb, 100);
1600
1601 BlockAndTailDupResult BestSucc = { .BB: nullptr, .ShouldTailDup: false };
1602 auto BestProb = BranchProbability::getZero();
1603
1604 SmallVector<MachineBasicBlock *, 4> Successors;
1605 auto AdjustedSumProb =
1606 collectViableSuccessors(BB, Chain, BlockFilter, Successors);
1607
1608 LLVM_DEBUG(dbgs() << "Selecting best successor for: " << getBlockName(BB)
1609 << "\n");
1610
1611 // if we already precomputed the best successor for BB, return that if still
1612 // applicable.
1613 auto FoundEdge = ComputedEdges.find(Val: BB);
1614 if (FoundEdge != ComputedEdges.end()) {
1615 MachineBasicBlock *Succ = FoundEdge->second.BB;
1616 ComputedEdges.erase(I: FoundEdge);
1617 BlockChain *SuccChain = BlockToChain[Succ];
1618 if (BB->isSuccessor(MBB: Succ) && (!BlockFilter || BlockFilter->count(key: Succ)) &&
1619 SuccChain != &Chain && Succ == *SuccChain->begin())
1620 return FoundEdge->second;
1621 }
1622
1623 // if BB is part of a trellis, Use the trellis to determine the optimal
1624 // fallthrough edges
1625 if (isTrellis(BB, ViableSuccs: Successors, Chain, BlockFilter))
1626 return getBestTrellisSuccessor(BB, ViableSuccs: Successors, AdjustedSumProb, Chain,
1627 BlockFilter);
1628
1629 // For blocks with CFG violations, we may be able to lay them out anyway with
1630 // tail-duplication. We keep this vector so we can perform the probability
1631 // calculations the minimum number of times.
1632 SmallVector<std::pair<BranchProbability, MachineBasicBlock *>, 4>
1633 DupCandidates;
1634 for (MachineBasicBlock *Succ : Successors) {
1635 auto RealSuccProb = MBPI->getEdgeProbability(Src: BB, Dst: Succ);
1636 BranchProbability SuccProb =
1637 getAdjustedProbability(OrigProb: RealSuccProb, AdjustedSumProb);
1638
1639 BlockChain &SuccChain = *BlockToChain[Succ];
1640 // Skip the edge \c BB->Succ if block \c Succ has a better layout
1641 // predecessor that yields lower global cost.
1642 if (hasBetterLayoutPredecessor(BB, Succ, SuccChain, SuccProb, RealSuccProb,
1643 Chain, BlockFilter)) {
1644 // If tail duplication would make Succ profitable, place it.
1645 if (allowTailDupPlacement() && shouldTailDuplicate(BB: Succ))
1646 DupCandidates.emplace_back(Args&: SuccProb, Args&: Succ);
1647 continue;
1648 }
1649
1650 LLVM_DEBUG(
1651 dbgs() << " Candidate: " << getBlockName(Succ)
1652 << ", probability: " << SuccProb
1653 << (SuccChain.UnscheduledPredecessors != 0 ? " (CFG break)" : "")
1654 << "\n");
1655
1656 if (BestSucc.BB && BestProb >= SuccProb) {
1657 LLVM_DEBUG(dbgs() << " Not the best candidate, continuing\n");
1658 continue;
1659 }
1660
1661 LLVM_DEBUG(dbgs() << " Setting it as best candidate\n");
1662 BestSucc.BB = Succ;
1663 BestProb = SuccProb;
1664 }
1665 // Handle the tail duplication candidates in order of decreasing probability.
1666 // Stop at the first one that is profitable. Also stop if they are less
1667 // profitable than BestSucc. Position is important because we preserve it and
1668 // prefer first best match. Here we aren't comparing in order, so we capture
1669 // the position instead.
1670 llvm::stable_sort(Range&: DupCandidates,
1671 C: [](std::tuple<BranchProbability, MachineBasicBlock *> L,
1672 std::tuple<BranchProbability, MachineBasicBlock *> R) {
1673 return std::get<0>(t&: L) > std::get<0>(t&: R);
1674 });
1675 for (auto &Tup : DupCandidates) {
1676 BranchProbability DupProb;
1677 MachineBasicBlock *Succ;
1678 std::tie(args&: DupProb, args&: Succ) = Tup;
1679 if (DupProb < BestProb)
1680 break;
1681 if (canTailDuplicateUnplacedPreds(BB, Succ, Chain, BlockFilter)
1682 && (isProfitableToTailDup(BB, Succ, QProb: BestProb, Chain, BlockFilter))) {
1683 LLVM_DEBUG(dbgs() << " Candidate: " << getBlockName(Succ)
1684 << ", probability: " << DupProb
1685 << " (Tail Duplicate)\n");
1686 BestSucc.BB = Succ;
1687 BestSucc.ShouldTailDup = true;
1688 break;
1689 }
1690 }
1691
1692 if (BestSucc.BB)
1693 LLVM_DEBUG(dbgs() << " Selected: " << getBlockName(BestSucc.BB) << "\n");
1694
1695 return BestSucc;
1696}
1697
1698/// Select the best block from a worklist.
1699///
1700/// This looks through the provided worklist as a list of candidate basic
1701/// blocks and select the most profitable one to place. The definition of
1702/// profitable only really makes sense in the context of a loop. This returns
1703/// the most frequently visited block in the worklist, which in the case of
1704/// a loop, is the one most desirable to be physically close to the rest of the
1705/// loop body in order to improve i-cache behavior.
1706///
1707/// \returns The best block found, or null if none are viable.
1708MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock(
1709 const BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList) {
1710 // Once we need to walk the worklist looking for a candidate, cleanup the
1711 // worklist of already placed entries.
1712 // FIXME: If this shows up on profiles, it could be folded (at the cost of
1713 // some code complexity) into the loop below.
1714 llvm::erase_if(C&: WorkList, P: [&](MachineBasicBlock *BB) {
1715 return BlockToChain.lookup(Val: BB) == &Chain;
1716 });
1717
1718 if (WorkList.empty())
1719 return nullptr;
1720
1721 bool IsEHPad = WorkList[0]->isEHPad();
1722
1723 MachineBasicBlock *BestBlock = nullptr;
1724 BlockFrequency BestFreq;
1725 for (MachineBasicBlock *MBB : WorkList) {
1726 assert(MBB->isEHPad() == IsEHPad &&
1727 "EHPad mismatch between block and work list.");
1728
1729 BlockChain &SuccChain = *BlockToChain[MBB];
1730 if (&SuccChain == &Chain)
1731 continue;
1732
1733 assert(SuccChain.UnscheduledPredecessors == 0 &&
1734 "Found CFG-violating block");
1735
1736 BlockFrequency CandidateFreq = MBFI->getBlockFreq(MBB);
1737 LLVM_DEBUG(dbgs() << " " << getBlockName(MBB) << " -> "
1738 << printBlockFreq(MBFI->getMBFI(), CandidateFreq)
1739 << " (freq)\n");
1740
1741 // For ehpad, we layout the least probable first as to avoid jumping back
1742 // from least probable landingpads to more probable ones.
1743 //
1744 // FIXME: Using probability is probably (!) not the best way to achieve
1745 // this. We should probably have a more principled approach to layout
1746 // cleanup code.
1747 //
1748 // The goal is to get:
1749 //
1750 // +--------------------------+
1751 // | V
1752 // InnerLp -> InnerCleanup OuterLp -> OuterCleanup -> Resume
1753 //
1754 // Rather than:
1755 //
1756 // +-------------------------------------+
1757 // V |
1758 // OuterLp -> OuterCleanup -> Resume InnerLp -> InnerCleanup
1759 if (BestBlock && (IsEHPad ^ (BestFreq >= CandidateFreq)))
1760 continue;
1761
1762 BestBlock = MBB;
1763 BestFreq = CandidateFreq;
1764 }
1765
1766 return BestBlock;
1767}
1768
1769/// Retrieve the first unplaced basic block in the entire function.
1770///
1771/// This routine is called when we are unable to use the CFG to walk through
1772/// all of the basic blocks and form a chain due to unnatural loops in the CFG.
1773/// We walk through the function's blocks in order, starting from the
1774/// LastUnplacedBlockIt. We update this iterator on each call to avoid
1775/// re-scanning the entire sequence on repeated calls to this routine.
1776MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock(
1777 const BlockChain &PlacedChain,
1778 MachineFunction::iterator &PrevUnplacedBlockIt) {
1779
1780 for (MachineFunction::iterator I = PrevUnplacedBlockIt, E = F->end(); I != E;
1781 ++I) {
1782 if (BlockToChain[&*I] != &PlacedChain) {
1783 PrevUnplacedBlockIt = I;
1784 // Now select the head of the chain to which the unplaced block belongs
1785 // as the block to place. This will force the entire chain to be placed,
1786 // and satisfies the requirements of merging chains.
1787 return *BlockToChain[&*I]->begin();
1788 }
1789 }
1790 return nullptr;
1791}
1792
1793/// Retrieve the first unplaced basic block among the blocks in BlockFilter.
1794///
1795/// This is similar to getFirstUnplacedBlock for the entire function, but since
1796/// the size of BlockFilter is typically far less than the number of blocks in
1797/// the entire function, iterating through the BlockFilter is more efficient.
1798/// When processing the entire funciton, using the version without BlockFilter
1799/// has a complexity of #(loops in function) * #(blocks in function), while this
1800/// version has a complexity of sum(#(loops in block) foreach block in function)
1801/// which is always smaller. For long function mostly sequential in structure,
1802/// the complexity is amortized to 1 * #(blocks in function).
1803MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock(
1804 const BlockChain &PlacedChain,
1805 BlockFilterSet::iterator &PrevUnplacedBlockInFilterIt,
1806 const BlockFilterSet *BlockFilter) {
1807 assert(BlockFilter);
1808 for (; PrevUnplacedBlockInFilterIt != BlockFilter->end();
1809 ++PrevUnplacedBlockInFilterIt) {
1810 BlockChain *C = BlockToChain[*PrevUnplacedBlockInFilterIt];
1811 if (C != &PlacedChain) {
1812 return *C->begin();
1813 }
1814 }
1815 return nullptr;
1816}
1817
1818void MachineBlockPlacement::fillWorkLists(
1819 const MachineBasicBlock *MBB,
1820 SmallPtrSetImpl<BlockChain *> &UpdatedPreds,
1821 const BlockFilterSet *BlockFilter = nullptr) {
1822 BlockChain &Chain = *BlockToChain[MBB];
1823 if (!UpdatedPreds.insert(Ptr: &Chain).second)
1824 return;
1825
1826 assert(
1827 Chain.UnscheduledPredecessors == 0 &&
1828 "Attempting to place block with unscheduled predecessors in worklist.");
1829 for (MachineBasicBlock *ChainBB : Chain) {
1830 assert(BlockToChain[ChainBB] == &Chain &&
1831 "Block in chain doesn't match BlockToChain map.");
1832 for (MachineBasicBlock *Pred : ChainBB->predecessors()) {
1833 if (BlockFilter && !BlockFilter->count(key: Pred))
1834 continue;
1835 if (BlockToChain[Pred] == &Chain)
1836 continue;
1837 ++Chain.UnscheduledPredecessors;
1838 }
1839 }
1840
1841 if (Chain.UnscheduledPredecessors != 0)
1842 return;
1843
1844 MachineBasicBlock *BB = *Chain.begin();
1845 if (BB->isEHPad())
1846 EHPadWorkList.push_back(Elt: BB);
1847 else
1848 BlockWorkList.push_back(Elt: BB);
1849}
1850
1851void MachineBlockPlacement::buildChain(
1852 const MachineBasicBlock *HeadBB, BlockChain &Chain,
1853 BlockFilterSet *BlockFilter) {
1854 assert(HeadBB && "BB must not be null.\n");
1855 assert(BlockToChain[HeadBB] == &Chain && "BlockToChainMap mis-match.\n");
1856 MachineFunction::iterator PrevUnplacedBlockIt = F->begin();
1857 BlockFilterSet::iterator PrevUnplacedBlockInFilterIt;
1858 if (BlockFilter)
1859 PrevUnplacedBlockInFilterIt = BlockFilter->begin();
1860
1861 const MachineBasicBlock *LoopHeaderBB = HeadBB;
1862 markChainSuccessors(Chain, LoopHeaderBB, BlockFilter);
1863 MachineBasicBlock *BB = *std::prev(x: Chain.end());
1864 while (true) {
1865 assert(BB && "null block found at end of chain in loop.");
1866 assert(BlockToChain[BB] == &Chain && "BlockToChainMap mis-match in loop.");
1867 assert(*std::prev(Chain.end()) == BB && "BB Not found at end of chain.");
1868
1869
1870 // Look for the best viable successor if there is one to place immediately
1871 // after this block.
1872 auto Result = selectBestSuccessor(BB, Chain, BlockFilter);
1873 MachineBasicBlock* BestSucc = Result.BB;
1874 bool ShouldTailDup = Result.ShouldTailDup;
1875 if (allowTailDupPlacement())
1876 ShouldTailDup |= (BestSucc && canTailDuplicateUnplacedPreds(BB, Succ: BestSucc,
1877 Chain,
1878 BlockFilter));
1879
1880 // If an immediate successor isn't available, look for the best viable
1881 // block among those we've identified as not violating the loop's CFG at
1882 // this point. This won't be a fallthrough, but it will increase locality.
1883 if (!BestSucc)
1884 BestSucc = selectBestCandidateBlock(Chain, WorkList&: BlockWorkList);
1885 if (!BestSucc)
1886 BestSucc = selectBestCandidateBlock(Chain, WorkList&: EHPadWorkList);
1887
1888 if (!BestSucc) {
1889 if (BlockFilter)
1890 BestSucc = getFirstUnplacedBlock(PlacedChain: Chain, PrevUnplacedBlockInFilterIt,
1891 BlockFilter);
1892 else
1893 BestSucc = getFirstUnplacedBlock(PlacedChain: Chain, PrevUnplacedBlockIt);
1894 if (!BestSucc)
1895 break;
1896
1897 LLVM_DEBUG(dbgs() << "Unnatural loop CFG detected, forcibly merging the "
1898 "layout successor until the CFG reduces\n");
1899 }
1900
1901 // Placement may have changed tail duplication opportunities.
1902 // Check for that now.
1903 if (allowTailDupPlacement() && BestSucc && ShouldTailDup) {
1904 repeatedlyTailDuplicateBlock(BB: BestSucc, LPred&: BB, LoopHeaderBB, Chain,
1905 BlockFilter, PrevUnplacedBlockIt,
1906 PrevUnplacedBlockInFilterIt);
1907 // If the chosen successor was duplicated into BB, don't bother laying
1908 // it out, just go round the loop again with BB as the chain end.
1909 if (!BB->isSuccessor(MBB: BestSucc))
1910 continue;
1911 }
1912
1913 // Place this block, updating the datastructures to reflect its placement.
1914 BlockChain &SuccChain = *BlockToChain[BestSucc];
1915 // Zero out UnscheduledPredecessors for the successor we're about to merge in case
1916 // we selected a successor that didn't fit naturally into the CFG.
1917 SuccChain.UnscheduledPredecessors = 0;
1918 LLVM_DEBUG(dbgs() << "Merging from " << getBlockName(BB) << " to "
1919 << getBlockName(BestSucc) << "\n");
1920 markChainSuccessors(Chain: SuccChain, LoopHeaderBB, BlockFilter);
1921 Chain.merge(BB: BestSucc, Chain: &SuccChain);
1922 BB = *std::prev(x: Chain.end());
1923 }
1924
1925 LLVM_DEBUG(dbgs() << "Finished forming chain for header block "
1926 << getBlockName(*Chain.begin()) << "\n");
1927}
1928
1929// If bottom of block BB has only one successor OldTop, in most cases it is
1930// profitable to move it before OldTop, except the following case:
1931//
1932// -->OldTop<-
1933// | . |
1934// | . |
1935// | . |
1936// ---Pred |
1937// | |
1938// BB-----
1939//
1940// If BB is moved before OldTop, Pred needs a taken branch to BB, and it can't
1941// layout the other successor below it, so it can't reduce taken branch.
1942// In this case we keep its original layout.
1943bool
1944MachineBlockPlacement::canMoveBottomBlockToTop(
1945 const MachineBasicBlock *BottomBlock,
1946 const MachineBasicBlock *OldTop) {
1947 if (BottomBlock->pred_size() != 1)
1948 return true;
1949 MachineBasicBlock *Pred = *BottomBlock->pred_begin();
1950 if (Pred->succ_size() != 2)
1951 return true;
1952
1953 MachineBasicBlock *OtherBB = *Pred->succ_begin();
1954 if (OtherBB == BottomBlock)
1955 OtherBB = *Pred->succ_rbegin();
1956 if (OtherBB == OldTop)
1957 return false;
1958
1959 return true;
1960}
1961
1962// Find out the possible fall through frequence to the top of a loop.
1963BlockFrequency
1964MachineBlockPlacement::TopFallThroughFreq(
1965 const MachineBasicBlock *Top,
1966 const BlockFilterSet &LoopBlockSet) {
1967 BlockFrequency MaxFreq = BlockFrequency(0);
1968 for (MachineBasicBlock *Pred : Top->predecessors()) {
1969 BlockChain *PredChain = BlockToChain[Pred];
1970 if (!LoopBlockSet.count(key: Pred) &&
1971 (!PredChain || Pred == *std::prev(x: PredChain->end()))) {
1972 // Found a Pred block can be placed before Top.
1973 // Check if Top is the best successor of Pred.
1974 auto TopProb = MBPI->getEdgeProbability(Src: Pred, Dst: Top);
1975 bool TopOK = true;
1976 for (MachineBasicBlock *Succ : Pred->successors()) {
1977 auto SuccProb = MBPI->getEdgeProbability(Src: Pred, Dst: Succ);
1978 BlockChain *SuccChain = BlockToChain[Succ];
1979 // Check if Succ can be placed after Pred.
1980 // Succ should not be in any chain, or it is the head of some chain.
1981 if (!LoopBlockSet.count(key: Succ) && (SuccProb > TopProb) &&
1982 (!SuccChain || Succ == *SuccChain->begin())) {
1983 TopOK = false;
1984 break;
1985 }
1986 }
1987 if (TopOK) {
1988 BlockFrequency EdgeFreq = MBFI->getBlockFreq(MBB: Pred) *
1989 MBPI->getEdgeProbability(Src: Pred, Dst: Top);
1990 if (EdgeFreq > MaxFreq)
1991 MaxFreq = EdgeFreq;
1992 }
1993 }
1994 }
1995 return MaxFreq;
1996}
1997
1998// Compute the fall through gains when move NewTop before OldTop.
1999//
2000// In following diagram, edges marked as "-" are reduced fallthrough, edges
2001// marked as "+" are increased fallthrough, this function computes
2002//
2003// SUM(increased fallthrough) - SUM(decreased fallthrough)
2004//
2005// |
2006// | -
2007// V
2008// --->OldTop
2009// | .
2010// | .
2011// +| . +
2012// | Pred --->
2013// | |-
2014// | V
2015// --- NewTop <---
2016// |-
2017// V
2018//
2019BlockFrequency
2020MachineBlockPlacement::FallThroughGains(
2021 const MachineBasicBlock *NewTop,
2022 const MachineBasicBlock *OldTop,
2023 const MachineBasicBlock *ExitBB,
2024 const BlockFilterSet &LoopBlockSet) {
2025 BlockFrequency FallThrough2Top = TopFallThroughFreq(Top: OldTop, LoopBlockSet);
2026 BlockFrequency FallThrough2Exit = BlockFrequency(0);
2027 if (ExitBB)
2028 FallThrough2Exit = MBFI->getBlockFreq(MBB: NewTop) *
2029 MBPI->getEdgeProbability(Src: NewTop, Dst: ExitBB);
2030 BlockFrequency BackEdgeFreq = MBFI->getBlockFreq(MBB: NewTop) *
2031 MBPI->getEdgeProbability(Src: NewTop, Dst: OldTop);
2032
2033 // Find the best Pred of NewTop.
2034 MachineBasicBlock *BestPred = nullptr;
2035 BlockFrequency FallThroughFromPred = BlockFrequency(0);
2036 for (MachineBasicBlock *Pred : NewTop->predecessors()) {
2037 if (!LoopBlockSet.count(key: Pred))
2038 continue;
2039 BlockChain *PredChain = BlockToChain[Pred];
2040 if (!PredChain || Pred == *std::prev(x: PredChain->end())) {
2041 BlockFrequency EdgeFreq =
2042 MBFI->getBlockFreq(MBB: Pred) * MBPI->getEdgeProbability(Src: Pred, Dst: NewTop);
2043 if (EdgeFreq > FallThroughFromPred) {
2044 FallThroughFromPred = EdgeFreq;
2045 BestPred = Pred;
2046 }
2047 }
2048 }
2049
2050 // If NewTop is not placed after Pred, another successor can be placed
2051 // after Pred.
2052 BlockFrequency NewFreq = BlockFrequency(0);
2053 if (BestPred) {
2054 for (MachineBasicBlock *Succ : BestPred->successors()) {
2055 if ((Succ == NewTop) || (Succ == BestPred) || !LoopBlockSet.count(key: Succ))
2056 continue;
2057 if (ComputedEdges.contains(Val: Succ))
2058 continue;
2059 BlockChain *SuccChain = BlockToChain[Succ];
2060 if ((SuccChain && (Succ != *SuccChain->begin())) ||
2061 (SuccChain == BlockToChain[BestPred]))
2062 continue;
2063 BlockFrequency EdgeFreq = MBFI->getBlockFreq(MBB: BestPred) *
2064 MBPI->getEdgeProbability(Src: BestPred, Dst: Succ);
2065 if (EdgeFreq > NewFreq)
2066 NewFreq = EdgeFreq;
2067 }
2068 BlockFrequency OrigEdgeFreq = MBFI->getBlockFreq(MBB: BestPred) *
2069 MBPI->getEdgeProbability(Src: BestPred, Dst: NewTop);
2070 if (NewFreq > OrigEdgeFreq) {
2071 // If NewTop is not the best successor of Pred, then Pred doesn't
2072 // fallthrough to NewTop. So there is no FallThroughFromPred and
2073 // NewFreq.
2074 NewFreq = BlockFrequency(0);
2075 FallThroughFromPred = BlockFrequency(0);
2076 }
2077 }
2078
2079 BlockFrequency Result = BlockFrequency(0);
2080 BlockFrequency Gains = BackEdgeFreq + NewFreq;
2081 BlockFrequency Lost =
2082 FallThrough2Top + FallThrough2Exit + FallThroughFromPred;
2083 if (Gains > Lost)
2084 Result = Gains - Lost;
2085 return Result;
2086}
2087
2088/// Helper function of findBestLoopTop. Find the best loop top block
2089/// from predecessors of old top.
2090///
2091/// Look for a block which is strictly better than the old top for laying
2092/// out before the old top of the loop. This looks for only two patterns:
2093///
2094/// 1. a block has only one successor, the old loop top
2095///
2096/// Because such a block will always result in an unconditional jump,
2097/// rotating it in front of the old top is always profitable.
2098///
2099/// 2. a block has two successors, one is old top, another is exit
2100/// and it has more than one predecessors
2101///
2102/// If it is below one of its predecessors P, only P can fall through to
2103/// it, all other predecessors need a jump to it, and another conditional
2104/// jump to loop header. If it is moved before loop header, all its
2105/// predecessors jump to it, then fall through to loop header. So all its
2106/// predecessors except P can reduce one taken branch.
2107/// At the same time, move it before old top increases the taken branch
2108/// to loop exit block, so the reduced taken branch will be compared with
2109/// the increased taken branch to the loop exit block.
2110MachineBasicBlock *
2111MachineBlockPlacement::findBestLoopTopHelper(
2112 MachineBasicBlock *OldTop,
2113 const MachineLoop &L,
2114 const BlockFilterSet &LoopBlockSet) {
2115 // Check that the header hasn't been fused with a preheader block due to
2116 // crazy branches. If it has, we need to start with the header at the top to
2117 // prevent pulling the preheader into the loop body.
2118 BlockChain &HeaderChain = *BlockToChain[OldTop];
2119 if (!LoopBlockSet.count(key: *HeaderChain.begin()))
2120 return OldTop;
2121 if (OldTop != *HeaderChain.begin())
2122 return OldTop;
2123
2124 LLVM_DEBUG(dbgs() << "Finding best loop top for: " << getBlockName(OldTop)
2125 << "\n");
2126
2127 BlockFrequency BestGains = BlockFrequency(0);
2128 MachineBasicBlock *BestPred = nullptr;
2129 for (MachineBasicBlock *Pred : OldTop->predecessors()) {
2130 if (!LoopBlockSet.count(key: Pred))
2131 continue;
2132 if (Pred == L.getHeader())
2133 continue;
2134 LLVM_DEBUG(dbgs() << " old top pred: " << getBlockName(Pred) << ", has "
2135 << Pred->succ_size() << " successors, "
2136 << printBlockFreq(MBFI->getMBFI(), *Pred) << " freq\n");
2137 if (Pred->succ_size() > 2)
2138 continue;
2139
2140 MachineBasicBlock *OtherBB = nullptr;
2141 if (Pred->succ_size() == 2) {
2142 OtherBB = *Pred->succ_begin();
2143 if (OtherBB == OldTop)
2144 OtherBB = *Pred->succ_rbegin();
2145 }
2146
2147 if (!canMoveBottomBlockToTop(BottomBlock: Pred, OldTop))
2148 continue;
2149
2150 BlockFrequency Gains = FallThroughGains(NewTop: Pred, OldTop, ExitBB: OtherBB,
2151 LoopBlockSet);
2152 if ((Gains > BlockFrequency(0)) &&
2153 (Gains > BestGains ||
2154 ((Gains == BestGains) && Pred->isLayoutSuccessor(MBB: OldTop)))) {
2155 BestPred = Pred;
2156 BestGains = Gains;
2157 }
2158 }
2159
2160 // If no direct predecessor is fine, just use the loop header.
2161 if (!BestPred) {
2162 LLVM_DEBUG(dbgs() << " final top unchanged\n");
2163 return OldTop;
2164 }
2165
2166 // Walk backwards through any straight line of predecessors.
2167 while (BestPred->pred_size() == 1 &&
2168 (*BestPred->pred_begin())->succ_size() == 1 &&
2169 *BestPred->pred_begin() != L.getHeader())
2170 BestPred = *BestPred->pred_begin();
2171
2172 LLVM_DEBUG(dbgs() << " final top: " << getBlockName(BestPred) << "\n");
2173 return BestPred;
2174}
2175
2176/// Find the best loop top block for layout.
2177///
2178/// This function iteratively calls findBestLoopTopHelper, until no new better
2179/// BB can be found.
2180MachineBasicBlock *
2181MachineBlockPlacement::findBestLoopTop(const MachineLoop &L,
2182 const BlockFilterSet &LoopBlockSet) {
2183 // Placing the latch block before the header may introduce an extra branch
2184 // that skips this block the first time the loop is executed, which we want
2185 // to avoid when optimising for size.
2186 // FIXME: in theory there is a case that does not introduce a new branch,
2187 // i.e. when the layout predecessor does not fallthrough to the loop header.
2188 // In practice this never happens though: there always seems to be a preheader
2189 // that can fallthrough and that is also placed before the header.
2190 bool OptForSize = F->getFunction().hasOptSize() ||
2191 llvm::shouldOptimizeForSize(MBB: L.getHeader(), PSI, MBFIWrapper: MBFI.get());
2192 if (OptForSize)
2193 return L.getHeader();
2194
2195 MachineBasicBlock *OldTop = nullptr;
2196 MachineBasicBlock *NewTop = L.getHeader();
2197 while (NewTop != OldTop) {
2198 OldTop = NewTop;
2199 NewTop = findBestLoopTopHelper(OldTop, L, LoopBlockSet);
2200 if (NewTop != OldTop)
2201 ComputedEdges[NewTop] = { .BB: OldTop, .ShouldTailDup: false };
2202 }
2203 return NewTop;
2204}
2205
2206/// Find the best loop exiting block for layout.
2207///
2208/// This routine implements the logic to analyze the loop looking for the best
2209/// block to layout at the top of the loop. Typically this is done to maximize
2210/// fallthrough opportunities.
2211MachineBasicBlock *
2212MachineBlockPlacement::findBestLoopExit(const MachineLoop &L,
2213 const BlockFilterSet &LoopBlockSet,
2214 BlockFrequency &ExitFreq) {
2215 // We don't want to layout the loop linearly in all cases. If the loop header
2216 // is just a normal basic block in the loop, we want to look for what block
2217 // within the loop is the best one to layout at the top. However, if the loop
2218 // header has be pre-merged into a chain due to predecessors not having
2219 // analyzable branches, *and* the predecessor it is merged with is *not* part
2220 // of the loop, rotating the header into the middle of the loop will create
2221 // a non-contiguous range of blocks which is Very Bad. So start with the
2222 // header and only rotate if safe.
2223 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
2224 if (!LoopBlockSet.count(key: *HeaderChain.begin()))
2225 return nullptr;
2226
2227 BlockFrequency BestExitEdgeFreq;
2228 unsigned BestExitLoopDepth = 0;
2229 MachineBasicBlock *ExitingBB = nullptr;
2230 // If there are exits to outer loops, loop rotation can severely limit
2231 // fallthrough opportunities unless it selects such an exit. Keep a set of
2232 // blocks where rotating to exit with that block will reach an outer loop.
2233 SmallPtrSet<MachineBasicBlock *, 4> BlocksExitingToOuterLoop;
2234
2235 LLVM_DEBUG(dbgs() << "Finding best loop exit for: "
2236 << getBlockName(L.getHeader()) << "\n");
2237 for (MachineBasicBlock *MBB : L.getBlocks()) {
2238 BlockChain &Chain = *BlockToChain[MBB];
2239 // Ensure that this block is at the end of a chain; otherwise it could be
2240 // mid-way through an inner loop or a successor of an unanalyzable branch.
2241 if (MBB != *std::prev(x: Chain.end()))
2242 continue;
2243
2244 // Now walk the successors. We need to establish whether this has a viable
2245 // exiting successor and whether it has a viable non-exiting successor.
2246 // We store the old exiting state and restore it if a viable looping
2247 // successor isn't found.
2248 MachineBasicBlock *OldExitingBB = ExitingBB;
2249 BlockFrequency OldBestExitEdgeFreq = BestExitEdgeFreq;
2250 bool HasLoopingSucc = false;
2251 for (MachineBasicBlock *Succ : MBB->successors()) {
2252 if (Succ->isEHPad())
2253 continue;
2254 if (Succ == MBB)
2255 continue;
2256 BlockChain &SuccChain = *BlockToChain[Succ];
2257 // Don't split chains, either this chain or the successor's chain.
2258 if (&Chain == &SuccChain) {
2259 LLVM_DEBUG(dbgs() << " exiting: " << getBlockName(MBB) << " -> "
2260 << getBlockName(Succ) << " (chain conflict)\n");
2261 continue;
2262 }
2263
2264 auto SuccProb = MBPI->getEdgeProbability(Src: MBB, Dst: Succ);
2265 if (LoopBlockSet.count(key: Succ)) {
2266 LLVM_DEBUG(dbgs() << " looping: " << getBlockName(MBB) << " -> "
2267 << getBlockName(Succ) << " (" << SuccProb << ")\n");
2268 HasLoopingSucc = true;
2269 continue;
2270 }
2271
2272 unsigned SuccLoopDepth = 0;
2273 if (MachineLoop *ExitLoop = MLI->getLoopFor(BB: Succ)) {
2274 SuccLoopDepth = ExitLoop->getLoopDepth();
2275 if (ExitLoop->contains(L: &L))
2276 BlocksExitingToOuterLoop.insert(Ptr: MBB);
2277 }
2278
2279 BlockFrequency ExitEdgeFreq = MBFI->getBlockFreq(MBB) * SuccProb;
2280 LLVM_DEBUG(
2281 dbgs() << " exiting: " << getBlockName(MBB) << " -> "
2282 << getBlockName(Succ) << " [L:" << SuccLoopDepth << "] ("
2283 << printBlockFreq(MBFI->getMBFI(), ExitEdgeFreq) << ")\n");
2284 // Note that we bias this toward an existing layout successor to retain
2285 // incoming order in the absence of better information. The exit must have
2286 // a frequency higher than the current exit before we consider breaking
2287 // the layout.
2288 BranchProbability Bias(100 - ExitBlockBias, 100);
2289 if (!ExitingBB || SuccLoopDepth > BestExitLoopDepth ||
2290 ExitEdgeFreq > BestExitEdgeFreq ||
2291 (MBB->isLayoutSuccessor(MBB: Succ) &&
2292 !(ExitEdgeFreq < BestExitEdgeFreq * Bias))) {
2293 BestExitEdgeFreq = ExitEdgeFreq;
2294 ExitingBB = MBB;
2295 }
2296 }
2297
2298 if (!HasLoopingSucc) {
2299 // Restore the old exiting state, no viable looping successor was found.
2300 ExitingBB = OldExitingBB;
2301 BestExitEdgeFreq = OldBestExitEdgeFreq;
2302 }
2303 }
2304 // Without a candidate exiting block or with only a single block in the
2305 // loop, just use the loop header to layout the loop.
2306 if (!ExitingBB) {
2307 LLVM_DEBUG(
2308 dbgs() << " No other candidate exit blocks, using loop header\n");
2309 return nullptr;
2310 }
2311 if (L.getNumBlocks() == 1) {
2312 LLVM_DEBUG(dbgs() << " Loop has 1 block, using loop header as exit\n");
2313 return nullptr;
2314 }
2315
2316 // Also, if we have exit blocks which lead to outer loops but didn't select
2317 // one of them as the exiting block we are rotating toward, disable loop
2318 // rotation altogether.
2319 if (!BlocksExitingToOuterLoop.empty() &&
2320 !BlocksExitingToOuterLoop.count(Ptr: ExitingBB))
2321 return nullptr;
2322
2323 LLVM_DEBUG(dbgs() << " Best exiting block: " << getBlockName(ExitingBB)
2324 << "\n");
2325 ExitFreq = BestExitEdgeFreq;
2326 return ExitingBB;
2327}
2328
2329/// Check if there is a fallthrough to loop header Top.
2330///
2331/// 1. Look for a Pred that can be layout before Top.
2332/// 2. Check if Top is the most possible successor of Pred.
2333bool
2334MachineBlockPlacement::hasViableTopFallthrough(
2335 const MachineBasicBlock *Top,
2336 const BlockFilterSet &LoopBlockSet) {
2337 for (MachineBasicBlock *Pred : Top->predecessors()) {
2338 BlockChain *PredChain = BlockToChain[Pred];
2339 if (!LoopBlockSet.count(key: Pred) &&
2340 (!PredChain || Pred == *std::prev(x: PredChain->end()))) {
2341 // Found a Pred block can be placed before Top.
2342 // Check if Top is the best successor of Pred.
2343 auto TopProb = MBPI->getEdgeProbability(Src: Pred, Dst: Top);
2344 bool TopOK = true;
2345 for (MachineBasicBlock *Succ : Pred->successors()) {
2346 auto SuccProb = MBPI->getEdgeProbability(Src: Pred, Dst: Succ);
2347 BlockChain *SuccChain = BlockToChain[Succ];
2348 // Check if Succ can be placed after Pred.
2349 // Succ should not be in any chain, or it is the head of some chain.
2350 if ((!SuccChain || Succ == *SuccChain->begin()) && SuccProb > TopProb) {
2351 TopOK = false;
2352 break;
2353 }
2354 }
2355 if (TopOK)
2356 return true;
2357 }
2358 }
2359 return false;
2360}
2361
2362/// Attempt to rotate an exiting block to the bottom of the loop.
2363///
2364/// Once we have built a chain, try to rotate it to line up the hot exit block
2365/// with fallthrough out of the loop if doing so doesn't introduce unnecessary
2366/// branches. For example, if the loop has fallthrough into its header and out
2367/// of its bottom already, don't rotate it.
2368void MachineBlockPlacement::rotateLoop(BlockChain &LoopChain,
2369 const MachineBasicBlock *ExitingBB,
2370 BlockFrequency ExitFreq,
2371 const BlockFilterSet &LoopBlockSet) {
2372 if (!ExitingBB)
2373 return;
2374
2375 MachineBasicBlock *Top = *LoopChain.begin();
2376 MachineBasicBlock *Bottom = *std::prev(x: LoopChain.end());
2377
2378 // If ExitingBB is already the last one in a chain then nothing to do.
2379 if (Bottom == ExitingBB)
2380 return;
2381
2382 // The entry block should always be the first BB in a function.
2383 if (Top->isEntryBlock())
2384 return;
2385
2386 bool ViableTopFallthrough = hasViableTopFallthrough(Top, LoopBlockSet);
2387
2388 // If the header has viable fallthrough, check whether the current loop
2389 // bottom is a viable exiting block. If so, bail out as rotating will
2390 // introduce an unnecessary branch.
2391 if (ViableTopFallthrough) {
2392 for (MachineBasicBlock *Succ : Bottom->successors()) {
2393 BlockChain *SuccChain = BlockToChain[Succ];
2394 if (!LoopBlockSet.count(key: Succ) &&
2395 (!SuccChain || Succ == *SuccChain->begin()))
2396 return;
2397 }
2398
2399 // Rotate will destroy the top fallthrough, we need to ensure the new exit
2400 // frequency is larger than top fallthrough.
2401 BlockFrequency FallThrough2Top = TopFallThroughFreq(Top, LoopBlockSet);
2402 if (FallThrough2Top >= ExitFreq)
2403 return;
2404 }
2405
2406 BlockChain::iterator ExitIt = llvm::find(Range&: LoopChain, Val: ExitingBB);
2407 if (ExitIt == LoopChain.end())
2408 return;
2409
2410 // Rotating a loop exit to the bottom when there is a fallthrough to top
2411 // trades the entry fallthrough for an exit fallthrough.
2412 // If there is no bottom->top edge, but the chosen exit block does have
2413 // a fallthrough, we break that fallthrough for nothing in return.
2414
2415 // Let's consider an example. We have a built chain of basic blocks
2416 // B1, B2, ..., Bn, where Bk is a ExitingBB - chosen exit block.
2417 // By doing a rotation we get
2418 // Bk+1, ..., Bn, B1, ..., Bk
2419 // Break of fallthrough to B1 is compensated by a fallthrough from Bk.
2420 // If we had a fallthrough Bk -> Bk+1 it is broken now.
2421 // It might be compensated by fallthrough Bn -> B1.
2422 // So we have a condition to avoid creation of extra branch by loop rotation.
2423 // All below must be true to avoid loop rotation:
2424 // If there is a fallthrough to top (B1)
2425 // There was fallthrough from chosen exit block (Bk) to next one (Bk+1)
2426 // There is no fallthrough from bottom (Bn) to top (B1).
2427 // Please note that there is no exit fallthrough from Bn because we checked it
2428 // above.
2429 if (ViableTopFallthrough) {
2430 assert(std::next(ExitIt) != LoopChain.end() &&
2431 "Exit should not be last BB");
2432 MachineBasicBlock *NextBlockInChain = *std::next(x: ExitIt);
2433 if (ExitingBB->isSuccessor(MBB: NextBlockInChain))
2434 if (!Bottom->isSuccessor(MBB: Top))
2435 return;
2436 }
2437
2438 LLVM_DEBUG(dbgs() << "Rotating loop to put exit " << getBlockName(ExitingBB)
2439 << " at bottom\n");
2440 std::rotate(first: LoopChain.begin(), middle: std::next(x: ExitIt), last: LoopChain.end());
2441}
2442
2443/// Attempt to rotate a loop based on profile data to reduce branch cost.
2444///
2445/// With profile data, we can determine the cost in terms of missed fall through
2446/// opportunities when rotating a loop chain and select the best rotation.
2447/// Basically, there are three kinds of cost to consider for each rotation:
2448/// 1. The possibly missed fall through edge (if it exists) from BB out of
2449/// the loop to the loop header.
2450/// 2. The possibly missed fall through edges (if they exist) from the loop
2451/// exits to BB out of the loop.
2452/// 3. The missed fall through edge (if it exists) from the last BB to the
2453/// first BB in the loop chain.
2454/// Therefore, the cost for a given rotation is the sum of costs listed above.
2455/// We select the best rotation with the smallest cost.
2456void MachineBlockPlacement::rotateLoopWithProfile(
2457 BlockChain &LoopChain, const MachineLoop &L,
2458 const BlockFilterSet &LoopBlockSet) {
2459 auto RotationPos = LoopChain.end();
2460 MachineBasicBlock *ChainHeaderBB = *LoopChain.begin();
2461
2462 // The entry block should always be the first BB in a function.
2463 if (ChainHeaderBB->isEntryBlock())
2464 return;
2465
2466 BlockFrequency SmallestRotationCost = BlockFrequency::max();
2467
2468 // A utility lambda that scales up a block frequency by dividing it by a
2469 // branch probability which is the reciprocal of the scale.
2470 auto ScaleBlockFrequency = [](BlockFrequency Freq,
2471 unsigned Scale) -> BlockFrequency {
2472 if (Scale == 0)
2473 return BlockFrequency(0);
2474 // Use operator / between BlockFrequency and BranchProbability to implement
2475 // saturating multiplication.
2476 return Freq / BranchProbability(1, Scale);
2477 };
2478
2479 // Compute the cost of the missed fall-through edge to the loop header if the
2480 // chain head is not the loop header. As we only consider natural loops with
2481 // single header, this computation can be done only once.
2482 BlockFrequency HeaderFallThroughCost(0);
2483 for (auto *Pred : ChainHeaderBB->predecessors()) {
2484 BlockChain *PredChain = BlockToChain[Pred];
2485 if (!LoopBlockSet.count(key: Pred) &&
2486 (!PredChain || Pred == *std::prev(x: PredChain->end()))) {
2487 auto EdgeFreq = MBFI->getBlockFreq(MBB: Pred) *
2488 MBPI->getEdgeProbability(Src: Pred, Dst: ChainHeaderBB);
2489 auto FallThruCost = ScaleBlockFrequency(EdgeFreq, MisfetchCost);
2490 // If the predecessor has only an unconditional jump to the header, we
2491 // need to consider the cost of this jump.
2492 if (Pred->succ_size() == 1)
2493 FallThruCost += ScaleBlockFrequency(EdgeFreq, JumpInstCost);
2494 HeaderFallThroughCost = std::max(a: HeaderFallThroughCost, b: FallThruCost);
2495 }
2496 }
2497
2498 // Here we collect all exit blocks in the loop, and for each exit we find out
2499 // its hottest exit edge. For each loop rotation, we define the loop exit cost
2500 // as the sum of frequencies of exit edges we collect here, excluding the exit
2501 // edge from the tail of the loop chain.
2502 SmallVector<std::pair<MachineBasicBlock *, BlockFrequency>, 4> ExitsWithFreq;
2503 for (auto *BB : LoopChain) {
2504 auto LargestExitEdgeProb = BranchProbability::getZero();
2505 for (auto *Succ : BB->successors()) {
2506 BlockChain *SuccChain = BlockToChain[Succ];
2507 if (!LoopBlockSet.count(key: Succ) &&
2508 (!SuccChain || Succ == *SuccChain->begin())) {
2509 auto SuccProb = MBPI->getEdgeProbability(Src: BB, Dst: Succ);
2510 LargestExitEdgeProb = std::max(a: LargestExitEdgeProb, b: SuccProb);
2511 }
2512 }
2513 if (LargestExitEdgeProb > BranchProbability::getZero()) {
2514 auto ExitFreq = MBFI->getBlockFreq(MBB: BB) * LargestExitEdgeProb;
2515 ExitsWithFreq.emplace_back(Args&: BB, Args&: ExitFreq);
2516 }
2517 }
2518
2519 // In this loop we iterate every block in the loop chain and calculate the
2520 // cost assuming the block is the head of the loop chain. When the loop ends,
2521 // we should have found the best candidate as the loop chain's head.
2522 for (auto Iter = LoopChain.begin(), TailIter = std::prev(x: LoopChain.end()),
2523 EndIter = LoopChain.end();
2524 Iter != EndIter; Iter++, TailIter++) {
2525 // TailIter is used to track the tail of the loop chain if the block we are
2526 // checking (pointed by Iter) is the head of the chain.
2527 if (TailIter == LoopChain.end())
2528 TailIter = LoopChain.begin();
2529
2530 auto TailBB = *TailIter;
2531
2532 // Calculate the cost by putting this BB to the top.
2533 BlockFrequency Cost = BlockFrequency(0);
2534
2535 // If the current BB is the loop header, we need to take into account the
2536 // cost of the missed fall through edge from outside of the loop to the
2537 // header.
2538 if (Iter != LoopChain.begin())
2539 Cost += HeaderFallThroughCost;
2540
2541 // Collect the loop exit cost by summing up frequencies of all exit edges
2542 // except the one from the chain tail.
2543 for (auto &ExitWithFreq : ExitsWithFreq)
2544 if (TailBB != ExitWithFreq.first)
2545 Cost += ExitWithFreq.second;
2546
2547 // The cost of breaking the once fall-through edge from the tail to the top
2548 // of the loop chain. Here we need to consider three cases:
2549 // 1. If the tail node has only one successor, then we will get an
2550 // additional jmp instruction. So the cost here is (MisfetchCost +
2551 // JumpInstCost) * tail node frequency.
2552 // 2. If the tail node has two successors, then we may still get an
2553 // additional jmp instruction if the layout successor after the loop
2554 // chain is not its CFG successor. Note that the more frequently executed
2555 // jmp instruction will be put ahead of the other one. Assume the
2556 // frequency of those two branches are x and y, where x is the frequency
2557 // of the edge to the chain head, then the cost will be
2558 // (x * MisfetechCost + min(x, y) * JumpInstCost) * tail node frequency.
2559 // 3. If the tail node has more than two successors (this rarely happens),
2560 // we won't consider any additional cost.
2561 if (TailBB->isSuccessor(MBB: *Iter)) {
2562 auto TailBBFreq = MBFI->getBlockFreq(MBB: TailBB);
2563 if (TailBB->succ_size() == 1)
2564 Cost += ScaleBlockFrequency(TailBBFreq, MisfetchCost + JumpInstCost);
2565 else if (TailBB->succ_size() == 2) {
2566 auto TailToHeadProb = MBPI->getEdgeProbability(Src: TailBB, Dst: *Iter);
2567 auto TailToHeadFreq = TailBBFreq * TailToHeadProb;
2568 auto ColderEdgeFreq = TailToHeadProb > BranchProbability(1, 2)
2569 ? TailBBFreq * TailToHeadProb.getCompl()
2570 : TailToHeadFreq;
2571 Cost += ScaleBlockFrequency(TailToHeadFreq, MisfetchCost) +
2572 ScaleBlockFrequency(ColderEdgeFreq, JumpInstCost);
2573 }
2574 }
2575
2576 LLVM_DEBUG(dbgs() << "The cost of loop rotation by making "
2577 << getBlockName(*Iter) << " to the top: "
2578 << printBlockFreq(MBFI->getMBFI(), Cost) << "\n");
2579
2580 if (Cost < SmallestRotationCost) {
2581 SmallestRotationCost = Cost;
2582 RotationPos = Iter;
2583 }
2584 }
2585
2586 if (RotationPos != LoopChain.end()) {
2587 LLVM_DEBUG(dbgs() << "Rotate loop by making " << getBlockName(*RotationPos)
2588 << " to the top\n");
2589 std::rotate(first: LoopChain.begin(), middle: RotationPos, last: LoopChain.end());
2590 }
2591}
2592
2593/// Collect blocks in the given loop that are to be placed.
2594///
2595/// When profile data is available, exclude cold blocks from the returned set;
2596/// otherwise, collect all blocks in the loop.
2597MachineBlockPlacement::BlockFilterSet
2598MachineBlockPlacement::collectLoopBlockSet(const MachineLoop &L) {
2599 BlockFilterSet LoopBlockSet;
2600
2601 // Filter cold blocks off from LoopBlockSet when profile data is available.
2602 // Collect the sum of frequencies of incoming edges to the loop header from
2603 // outside. If we treat the loop as a super block, this is the frequency of
2604 // the loop. Then for each block in the loop, we calculate the ratio between
2605 // its frequency and the frequency of the loop block. When it is too small,
2606 // don't add it to the loop chain. If there are outer loops, then this block
2607 // will be merged into the first outer loop chain for which this block is not
2608 // cold anymore. This needs precise profile data and we only do this when
2609 // profile data is available.
2610 if (F->getFunction().hasProfileData() || ForceLoopColdBlock) {
2611 BlockFrequency LoopFreq(0);
2612 for (auto *LoopPred : L.getHeader()->predecessors())
2613 if (!L.contains(BB: LoopPred))
2614 LoopFreq += MBFI->getBlockFreq(MBB: LoopPred) *
2615 MBPI->getEdgeProbability(Src: LoopPred, Dst: L.getHeader());
2616
2617 for (MachineBasicBlock *LoopBB : L.getBlocks()) {
2618 if (LoopBlockSet.count(key: LoopBB))
2619 continue;
2620 auto Freq = MBFI->getBlockFreq(MBB: LoopBB).getFrequency();
2621 if (Freq == 0 || LoopFreq.getFrequency() / Freq > LoopToColdBlockRatio)
2622 continue;
2623 BlockChain *Chain = BlockToChain[LoopBB];
2624 for (MachineBasicBlock *ChainBB : *Chain)
2625 LoopBlockSet.insert(X: ChainBB);
2626 }
2627 } else
2628 LoopBlockSet.insert(Start: L.block_begin(), End: L.block_end());
2629
2630 return LoopBlockSet;
2631}
2632
2633/// Forms basic block chains from the natural loop structures.
2634///
2635/// These chains are designed to preserve the existing *structure* of the code
2636/// as much as possible. We can then stitch the chains together in a way which
2637/// both preserves the topological structure and minimizes taken conditional
2638/// branches.
2639void MachineBlockPlacement::buildLoopChains(const MachineLoop &L) {
2640 // First recurse through any nested loops, building chains for those inner
2641 // loops.
2642 for (const MachineLoop *InnerLoop : L)
2643 buildLoopChains(L: *InnerLoop);
2644
2645 assert(BlockWorkList.empty() &&
2646 "BlockWorkList not empty when starting to build loop chains.");
2647 assert(EHPadWorkList.empty() &&
2648 "EHPadWorkList not empty when starting to build loop chains.");
2649 BlockFilterSet LoopBlockSet = collectLoopBlockSet(L);
2650
2651 // Check if we have profile data for this function. If yes, we will rotate
2652 // this loop by modeling costs more precisely which requires the profile data
2653 // for better layout.
2654 bool RotateLoopWithProfile =
2655 ForcePreciseRotationCost ||
2656 (PreciseRotationCost && F->getFunction().hasProfileData());
2657
2658 // First check to see if there is an obviously preferable top block for the
2659 // loop. This will default to the header, but may end up as one of the
2660 // predecessors to the header if there is one which will result in strictly
2661 // fewer branches in the loop body.
2662 MachineBasicBlock *LoopTop = findBestLoopTop(L, LoopBlockSet);
2663
2664 // If we selected just the header for the loop top, look for a potentially
2665 // profitable exit block in the event that rotating the loop can eliminate
2666 // branches by placing an exit edge at the bottom.
2667 //
2668 // Loops are processed innermost to uttermost, make sure we clear
2669 // PreferredLoopExit before processing a new loop.
2670 PreferredLoopExit = nullptr;
2671 BlockFrequency ExitFreq;
2672 if (!RotateLoopWithProfile && LoopTop == L.getHeader())
2673 PreferredLoopExit = findBestLoopExit(L, LoopBlockSet, ExitFreq);
2674
2675 BlockChain &LoopChain = *BlockToChain[LoopTop];
2676
2677 // FIXME: This is a really lame way of walking the chains in the loop: we
2678 // walk the blocks, and use a set to prevent visiting a particular chain
2679 // twice.
2680 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
2681 assert(LoopChain.UnscheduledPredecessors == 0 &&
2682 "LoopChain should not have unscheduled predecessors.");
2683 UpdatedPreds.insert(Ptr: &LoopChain);
2684
2685 for (const MachineBasicBlock *LoopBB : LoopBlockSet)
2686 fillWorkLists(MBB: LoopBB, UpdatedPreds, BlockFilter: &LoopBlockSet);
2687
2688 buildChain(HeadBB: LoopTop, Chain&: LoopChain, BlockFilter: &LoopBlockSet);
2689
2690 if (RotateLoopWithProfile)
2691 rotateLoopWithProfile(LoopChain, L, LoopBlockSet);
2692 else
2693 rotateLoop(LoopChain, ExitingBB: PreferredLoopExit, ExitFreq, LoopBlockSet);
2694
2695 LLVM_DEBUG({
2696 // Crash at the end so we get all of the debugging output first.
2697 bool BadLoop = false;
2698 if (LoopChain.UnscheduledPredecessors) {
2699 BadLoop = true;
2700 dbgs() << "Loop chain contains a block without its preds placed!\n"
2701 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
2702 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n";
2703 }
2704 for (MachineBasicBlock *ChainBB : LoopChain) {
2705 dbgs() << " ... " << getBlockName(ChainBB) << "\n";
2706 if (!LoopBlockSet.remove(ChainBB)) {
2707 // We don't mark the loop as bad here because there are real situations
2708 // where this can occur. For example, with an unanalyzable fallthrough
2709 // from a loop block to a non-loop block or vice versa.
2710 dbgs() << "Loop chain contains a block not contained by the loop!\n"
2711 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
2712 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
2713 << " Bad block: " << getBlockName(ChainBB) << "\n";
2714 }
2715 }
2716
2717 if (!LoopBlockSet.empty()) {
2718 BadLoop = true;
2719 for (const MachineBasicBlock *LoopBB : LoopBlockSet)
2720 dbgs() << "Loop contains blocks never placed into a chain!\n"
2721 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
2722 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
2723 << " Bad block: " << getBlockName(LoopBB) << "\n";
2724 }
2725 assert(!BadLoop && "Detected problems with the placement of this loop.");
2726 });
2727
2728 BlockWorkList.clear();
2729 EHPadWorkList.clear();
2730}
2731
2732void MachineBlockPlacement::buildCFGChains() {
2733 // Ensure that every BB in the function has an associated chain to simplify
2734 // the assumptions of the remaining algorithm.
2735 SmallVector<MachineOperand, 4> Cond; // For analyzeBranch.
2736 for (MachineFunction::iterator FI = F->begin(), FE = F->end(); FI != FE;
2737 ++FI) {
2738 MachineBasicBlock *BB = &*FI;
2739 BlockChain *Chain =
2740 new (ChainAllocator.Allocate()) BlockChain(BlockToChain, BB);
2741 // Also, merge any blocks which we cannot reason about and must preserve
2742 // the exact fallthrough behavior for.
2743 while (true) {
2744 Cond.clear();
2745 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For analyzeBranch.
2746 if (!TII->analyzeBranch(MBB&: *BB, TBB, FBB, Cond) || !FI->canFallThrough())
2747 break;
2748
2749 MachineFunction::iterator NextFI = std::next(x: FI);
2750 MachineBasicBlock *NextBB = &*NextFI;
2751 // Ensure that the layout successor is a viable block, as we know that
2752 // fallthrough is a possibility.
2753 assert(NextFI != FE && "Can't fallthrough past the last block.");
2754 LLVM_DEBUG(dbgs() << "Pre-merging due to unanalyzable fallthrough: "
2755 << getBlockName(BB) << " -> " << getBlockName(NextBB)
2756 << "\n");
2757 Chain->merge(BB: NextBB, Chain: nullptr);
2758#ifndef NDEBUG
2759 BlocksWithUnanalyzableExits.insert(&*BB);
2760#endif
2761 FI = NextFI;
2762 BB = NextBB;
2763 }
2764 }
2765
2766 // Build any loop-based chains.
2767 PreferredLoopExit = nullptr;
2768 for (MachineLoop *L : *MLI)
2769 buildLoopChains(L: *L);
2770
2771 assert(BlockWorkList.empty() &&
2772 "BlockWorkList should be empty before building final chain.");
2773 assert(EHPadWorkList.empty() &&
2774 "EHPadWorkList should be empty before building final chain.");
2775
2776 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
2777 for (MachineBasicBlock &MBB : *F)
2778 fillWorkLists(MBB: &MBB, UpdatedPreds);
2779
2780 BlockChain &FunctionChain = *BlockToChain[&F->front()];
2781 buildChain(HeadBB: &F->front(), Chain&: FunctionChain);
2782
2783#ifndef NDEBUG
2784 using FunctionBlockSetType = SmallPtrSet<MachineBasicBlock *, 16>;
2785#endif
2786 LLVM_DEBUG({
2787 // Crash at the end so we get all of the debugging output first.
2788 bool BadFunc = false;
2789 FunctionBlockSetType FunctionBlockSet;
2790 for (MachineBasicBlock &MBB : *F)
2791 FunctionBlockSet.insert(&MBB);
2792
2793 for (MachineBasicBlock *ChainBB : FunctionChain)
2794 if (!FunctionBlockSet.erase(ChainBB)) {
2795 BadFunc = true;
2796 dbgs() << "Function chain contains a block not in the function!\n"
2797 << " Bad block: " << getBlockName(ChainBB) << "\n";
2798 }
2799
2800 if (!FunctionBlockSet.empty()) {
2801 BadFunc = true;
2802 for (MachineBasicBlock *RemainingBB : FunctionBlockSet)
2803 dbgs() << "Function contains blocks never placed into a chain!\n"
2804 << " Bad block: " << getBlockName(RemainingBB) << "\n";
2805 }
2806 assert(!BadFunc && "Detected problems with the block placement.");
2807 });
2808
2809 // Remember original layout ordering, so we can update terminators after
2810 // reordering to point to the original layout successor.
2811 SmallVector<MachineBasicBlock *, 4> OriginalLayoutSuccessors(
2812 F->getNumBlockIDs());
2813 {
2814 MachineBasicBlock *LastMBB = nullptr;
2815 for (auto &MBB : *F) {
2816 if (LastMBB != nullptr)
2817 OriginalLayoutSuccessors[LastMBB->getNumber()] = &MBB;
2818 LastMBB = &MBB;
2819 }
2820 OriginalLayoutSuccessors[F->back().getNumber()] = nullptr;
2821 }
2822
2823 // Splice the blocks into place.
2824 MachineFunction::iterator InsertPos = F->begin();
2825 LLVM_DEBUG(dbgs() << "[MBP] Function: " << F->getName() << "\n");
2826 for (MachineBasicBlock *ChainBB : FunctionChain) {
2827 LLVM_DEBUG(dbgs() << (ChainBB == *FunctionChain.begin() ? "Placing chain "
2828 : " ... ")
2829 << getBlockName(ChainBB) << "\n");
2830 if (InsertPos != MachineFunction::iterator(ChainBB))
2831 F->splice(InsertPt: InsertPos, MBB: ChainBB);
2832 else
2833 ++InsertPos;
2834
2835 // Update the terminator of the previous block.
2836 if (ChainBB == *FunctionChain.begin())
2837 continue;
2838 MachineBasicBlock *PrevBB = &*std::prev(x: MachineFunction::iterator(ChainBB));
2839
2840 // FIXME: It would be awesome of updateTerminator would just return rather
2841 // than assert when the branch cannot be analyzed in order to remove this
2842 // boiler plate.
2843 Cond.clear();
2844 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For analyzeBranch.
2845
2846#ifndef NDEBUG
2847 if (!BlocksWithUnanalyzableExits.count(PrevBB)) {
2848 // Given the exact block placement we chose, we may actually not _need_ to
2849 // be able to edit PrevBB's terminator sequence, but not being _able_ to
2850 // do that at this point is a bug.
2851 assert((!TII->analyzeBranch(*PrevBB, TBB, FBB, Cond) ||
2852 !PrevBB->canFallThrough()) &&
2853 "Unexpected block with un-analyzable fallthrough!");
2854 Cond.clear();
2855 TBB = FBB = nullptr;
2856 }
2857#endif
2858
2859 // The "PrevBB" is not yet updated to reflect current code layout, so,
2860 // o. it may fall-through to a block without explicit "goto" instruction
2861 // before layout, and no longer fall-through it after layout; or
2862 // o. just opposite.
2863 //
2864 // analyzeBranch() may return erroneous value for FBB when these two
2865 // situations take place. For the first scenario FBB is mistakenly set NULL;
2866 // for the 2nd scenario, the FBB, which is expected to be NULL, is
2867 // mistakenly pointing to "*BI".
2868 // Thus, if the future change needs to use FBB before the layout is set, it
2869 // has to correct FBB first by using the code similar to the following:
2870 //
2871 // if (!Cond.empty() && (!FBB || FBB == ChainBB)) {
2872 // PrevBB->updateTerminator();
2873 // Cond.clear();
2874 // TBB = FBB = nullptr;
2875 // if (TII->analyzeBranch(*PrevBB, TBB, FBB, Cond)) {
2876 // // FIXME: This should never take place.
2877 // TBB = FBB = nullptr;
2878 // }
2879 // }
2880 if (!TII->analyzeBranch(MBB&: *PrevBB, TBB, FBB, Cond)) {
2881 PrevBB->updateTerminator(PreviousLayoutSuccessor: OriginalLayoutSuccessors[PrevBB->getNumber()]);
2882 }
2883 }
2884
2885 // Fixup the last block.
2886 Cond.clear();
2887 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For analyzeBranch.
2888 if (!TII->analyzeBranch(MBB&: F->back(), TBB, FBB, Cond)) {
2889 MachineBasicBlock *PrevBB = &F->back();
2890 PrevBB->updateTerminator(PreviousLayoutSuccessor: OriginalLayoutSuccessors[PrevBB->getNumber()]);
2891 }
2892
2893 BlockWorkList.clear();
2894 EHPadWorkList.clear();
2895}
2896
2897void MachineBlockPlacement::optimizeBranches() {
2898 BlockChain &FunctionChain = *BlockToChain[&F->front()];
2899 SmallVector<MachineOperand, 4> Cond; // For analyzeBranch.
2900
2901 // Now that all the basic blocks in the chain have the proper layout,
2902 // make a final call to analyzeBranch with AllowModify set.
2903 // Indeed, the target may be able to optimize the branches in a way we
2904 // cannot because all branches may not be analyzable.
2905 // E.g., the target may be able to remove an unconditional branch to
2906 // a fallthrough when it occurs after predicated terminators.
2907 for (MachineBasicBlock *ChainBB : FunctionChain) {
2908 Cond.clear();
2909 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For analyzeBranch.
2910 if (!TII->analyzeBranch(MBB&: *ChainBB, TBB, FBB, Cond, /*AllowModify*/ true)) {
2911 // If PrevBB has a two-way branch, try to re-order the branches
2912 // such that we branch to the successor with higher probability first.
2913 if (TBB && !Cond.empty() && FBB &&
2914 MBPI->getEdgeProbability(Src: ChainBB, Dst: FBB) >
2915 MBPI->getEdgeProbability(Src: ChainBB, Dst: TBB) &&
2916 !TII->reverseBranchCondition(Cond)) {
2917 LLVM_DEBUG(dbgs() << "Reverse order of the two branches: "
2918 << getBlockName(ChainBB) << "\n");
2919 LLVM_DEBUG(dbgs() << " Edge probability: "
2920 << MBPI->getEdgeProbability(ChainBB, FBB) << " vs "
2921 << MBPI->getEdgeProbability(ChainBB, TBB) << "\n");
2922 DebugLoc dl; // FIXME: this is nowhere
2923 TII->removeBranch(MBB&: *ChainBB);
2924 TII->insertBranch(MBB&: *ChainBB, TBB: FBB, FBB: TBB, Cond, DL: dl);
2925 }
2926 }
2927 }
2928}
2929
2930void MachineBlockPlacement::alignBlocks() {
2931 // Walk through the backedges of the function now that we have fully laid out
2932 // the basic blocks and align the destination of each backedge. We don't rely
2933 // exclusively on the loop info here so that we can align backedges in
2934 // unnatural CFGs and backedges that were introduced purely because of the
2935 // loop rotations done during this layout pass.
2936 if (F->getFunction().hasMinSize() ||
2937 (F->getFunction().hasOptSize() && !TLI->alignLoopsWithOptSize()))
2938 return;
2939 BlockChain &FunctionChain = *BlockToChain[&F->front()];
2940 if (FunctionChain.begin() == FunctionChain.end())
2941 return; // Empty chain.
2942
2943 const BranchProbability ColdProb(1, 5); // 20%
2944 BlockFrequency EntryFreq = MBFI->getBlockFreq(MBB: &F->front());
2945 BlockFrequency WeightedEntryFreq = EntryFreq * ColdProb;
2946 for (MachineBasicBlock *ChainBB : FunctionChain) {
2947 if (ChainBB == *FunctionChain.begin())
2948 continue;
2949
2950 // Don't align non-looping basic blocks. These are unlikely to execute
2951 // enough times to matter in practice. Note that we'll still handle
2952 // unnatural CFGs inside of a natural outer loop (the common case) and
2953 // rotated loops.
2954 MachineLoop *L = MLI->getLoopFor(BB: ChainBB);
2955 if (!L)
2956 continue;
2957
2958 const Align TLIAlign = TLI->getPrefLoopAlignment(ML: L);
2959 unsigned MDAlign = 1;
2960 MDNode *LoopID = L->getLoopID();
2961 if (LoopID) {
2962 for (const MDOperand &MDO : llvm::drop_begin(RangeOrContainer: LoopID->operands())) {
2963 MDNode *MD = dyn_cast<MDNode>(Val: MDO);
2964 if (MD == nullptr)
2965 continue;
2966 MDString *S = dyn_cast<MDString>(Val: MD->getOperand(I: 0));
2967 if (S == nullptr)
2968 continue;
2969 if (S->getString() == "llvm.loop.align") {
2970 assert(MD->getNumOperands() == 2 &&
2971 "per-loop align metadata should have two operands.");
2972 MDAlign =
2973 mdconst::extract<ConstantInt>(MD: MD->getOperand(I: 1))->getZExtValue();
2974 assert(MDAlign >= 1 && "per-loop align value must be positive.");
2975 }
2976 }
2977 }
2978
2979 // Use max of the TLIAlign and MDAlign
2980 const Align LoopAlign = std::max(a: TLIAlign, b: Align(MDAlign));
2981 if (LoopAlign == 1)
2982 continue; // Don't care about loop alignment.
2983
2984 // If the block is cold relative to the function entry don't waste space
2985 // aligning it.
2986 BlockFrequency Freq = MBFI->getBlockFreq(MBB: ChainBB);
2987 if (Freq < WeightedEntryFreq)
2988 continue;
2989
2990 // If the block is cold relative to its loop header, don't align it
2991 // regardless of what edges into the block exist.
2992 MachineBasicBlock *LoopHeader = L->getHeader();
2993 BlockFrequency LoopHeaderFreq = MBFI->getBlockFreq(MBB: LoopHeader);
2994 if (Freq < (LoopHeaderFreq * ColdProb))
2995 continue;
2996
2997 // If the global profiles indicates so, don't align it.
2998 if (llvm::shouldOptimizeForSize(MBB: ChainBB, PSI, MBFIWrapper: MBFI.get()) &&
2999 !TLI->alignLoopsWithOptSize())
3000 continue;
3001
3002 // Check for the existence of a non-layout predecessor which would benefit
3003 // from aligning this block.
3004 MachineBasicBlock *LayoutPred =
3005 &*std::prev(x: MachineFunction::iterator(ChainBB));
3006
3007 auto DetermineMaxAlignmentPadding = [&]() {
3008 // Set the maximum bytes allowed to be emitted for alignment.
3009 unsigned MaxBytes;
3010 if (MaxBytesForAlignmentOverride.getNumOccurrences() > 0)
3011 MaxBytes = MaxBytesForAlignmentOverride;
3012 else
3013 MaxBytes = TLI->getMaxPermittedBytesForAlignment(MBB: ChainBB);
3014 ChainBB->setMaxBytesForAlignment(MaxBytes);
3015 };
3016
3017 // Force alignment if all the predecessors are jumps. We already checked
3018 // that the block isn't cold above.
3019 if (!LayoutPred->isSuccessor(MBB: ChainBB)) {
3020 ChainBB->setAlignment(LoopAlign);
3021 DetermineMaxAlignmentPadding();
3022 continue;
3023 }
3024
3025 // Align this block if the layout predecessor's edge into this block is
3026 // cold relative to the block. When this is true, other predecessors make up
3027 // all of the hot entries into the block and thus alignment is likely to be
3028 // important.
3029 BranchProbability LayoutProb =
3030 MBPI->getEdgeProbability(Src: LayoutPred, Dst: ChainBB);
3031 BlockFrequency LayoutEdgeFreq = MBFI->getBlockFreq(MBB: LayoutPred) * LayoutProb;
3032 if (LayoutEdgeFreq <= (Freq * ColdProb)) {
3033 ChainBB->setAlignment(LoopAlign);
3034 DetermineMaxAlignmentPadding();
3035 }
3036 }
3037}
3038
3039/// Tail duplicate \p BB into (some) predecessors if profitable, repeating if
3040/// it was duplicated into its chain predecessor and removed.
3041/// \p BB - Basic block that may be duplicated.
3042///
3043/// \p LPred - Chosen layout predecessor of \p BB.
3044/// Updated to be the chain end if LPred is removed.
3045/// \p Chain - Chain to which \p LPred belongs, and \p BB will belong.
3046/// \p BlockFilter - Set of blocks that belong to the loop being laid out.
3047/// Used to identify which blocks to update predecessor
3048/// counts.
3049/// \p PrevUnplacedBlockIt - Iterator pointing to the last block that was
3050/// chosen in the given order due to unnatural CFG
3051/// only needed if \p BB is removed and
3052/// \p PrevUnplacedBlockIt pointed to \p BB.
3053/// @return true if \p BB was removed.
3054bool MachineBlockPlacement::repeatedlyTailDuplicateBlock(
3055 MachineBasicBlock *BB, MachineBasicBlock *&LPred,
3056 const MachineBasicBlock *LoopHeaderBB, BlockChain &Chain,
3057 BlockFilterSet *BlockFilter, MachineFunction::iterator &PrevUnplacedBlockIt,
3058 BlockFilterSet::iterator &PrevUnplacedBlockInFilterIt) {
3059 bool Removed, DuplicatedToLPred;
3060 bool DuplicatedToOriginalLPred;
3061 Removed = maybeTailDuplicateBlock(
3062 BB, LPred, Chain, BlockFilter, PrevUnplacedBlockIt,
3063 PrevUnplacedBlockInFilterIt, DuplicatedToLPred);
3064 if (!Removed)
3065 return false;
3066 DuplicatedToOriginalLPred = DuplicatedToLPred;
3067 // Iteratively try to duplicate again. It can happen that a block that is
3068 // duplicated into is still small enough to be duplicated again.
3069 // No need to call markBlockSuccessors in this case, as the blocks being
3070 // duplicated from here on are already scheduled.
3071 while (DuplicatedToLPred && Removed) {
3072 MachineBasicBlock *DupBB, *DupPred;
3073 // The removal callback causes Chain.end() to be updated when a block is
3074 // removed. On the first pass through the loop, the chain end should be the
3075 // same as it was on function entry. On subsequent passes, because we are
3076 // duplicating the block at the end of the chain, if it is removed the
3077 // chain will have shrunk by one block.
3078 BlockChain::iterator ChainEnd = Chain.end();
3079 DupBB = *(--ChainEnd);
3080 // Now try to duplicate again.
3081 if (ChainEnd == Chain.begin())
3082 break;
3083 DupPred = *std::prev(x: ChainEnd);
3084 Removed = maybeTailDuplicateBlock(
3085 BB: DupBB, LPred: DupPred, Chain, BlockFilter, PrevUnplacedBlockIt,
3086 PrevUnplacedBlockInFilterIt, DuplicatedToLPred);
3087 }
3088 // If BB was duplicated into LPred, it is now scheduled. But because it was
3089 // removed, markChainSuccessors won't be called for its chain. Instead we
3090 // call markBlockSuccessors for LPred to achieve the same effect. This must go
3091 // at the end because repeating the tail duplication can increase the number
3092 // of unscheduled predecessors.
3093 LPred = *std::prev(x: Chain.end());
3094 if (DuplicatedToOriginalLPred)
3095 markBlockSuccessors(Chain, MBB: LPred, LoopHeaderBB, BlockFilter);
3096 return true;
3097}
3098
3099/// Tail duplicate \p BB into (some) predecessors if profitable.
3100/// \p BB - Basic block that may be duplicated
3101/// \p LPred - Chosen layout predecessor of \p BB
3102/// \p Chain - Chain to which \p LPred belongs, and \p BB will belong.
3103/// \p BlockFilter - Set of blocks that belong to the loop being laid out.
3104/// Used to identify which blocks to update predecessor
3105/// counts.
3106/// \p PrevUnplacedBlockIt - Iterator pointing to the last block that was
3107/// chosen in the given order due to unnatural CFG
3108/// only needed if \p BB is removed and
3109/// \p PrevUnplacedBlockIt pointed to \p BB.
3110/// \p DuplicatedToLPred - True if the block was duplicated into LPred.
3111/// \return - True if the block was duplicated into all preds and removed.
3112bool MachineBlockPlacement::maybeTailDuplicateBlock(
3113 MachineBasicBlock *BB, MachineBasicBlock *LPred, BlockChain &Chain,
3114 BlockFilterSet *BlockFilter, MachineFunction::iterator &PrevUnplacedBlockIt,
3115 BlockFilterSet::iterator &PrevUnplacedBlockInFilterIt,
3116 bool &DuplicatedToLPred) {
3117 DuplicatedToLPred = false;
3118 if (!shouldTailDuplicate(BB))
3119 return false;
3120
3121 LLVM_DEBUG(dbgs() << "Redoing tail duplication for Succ#" << BB->getNumber()
3122 << "\n");
3123
3124 // This has to be a callback because none of it can be done after
3125 // BB is deleted.
3126 bool Removed = false;
3127 auto RemovalCallback =
3128 [&](MachineBasicBlock *RemBB) {
3129 // Signal to outer function
3130 Removed = true;
3131
3132 // Conservative default.
3133 bool InWorkList = true;
3134 // Remove from the Chain and Chain Map
3135 if (BlockToChain.count(Val: RemBB)) {
3136 BlockChain *Chain = BlockToChain[RemBB];
3137 InWorkList = Chain->UnscheduledPredecessors == 0;
3138 Chain->remove(BB: RemBB);
3139 BlockToChain.erase(Val: RemBB);
3140 }
3141
3142 // Handle the unplaced block iterator
3143 if (&(*PrevUnplacedBlockIt) == RemBB) {
3144 PrevUnplacedBlockIt++;
3145 }
3146
3147 // Handle the Work Lists
3148 if (InWorkList) {
3149 SmallVectorImpl<MachineBasicBlock *> &RemoveList = BlockWorkList;
3150 if (RemBB->isEHPad())
3151 RemoveList = EHPadWorkList;
3152 llvm::erase(C&: RemoveList, V: RemBB);
3153 }
3154
3155 // Handle the filter set
3156 if (BlockFilter) {
3157 auto It = llvm::find(Range&: *BlockFilter, Val: RemBB);
3158 // Erase RemBB from BlockFilter, and keep PrevUnplacedBlockInFilterIt
3159 // pointing to the same element as before.
3160 if (It != BlockFilter->end()) {
3161 if (It < PrevUnplacedBlockInFilterIt) {
3162 const MachineBasicBlock *PrevBB = *PrevUnplacedBlockInFilterIt;
3163 // BlockFilter is a SmallVector so all elements after RemBB are
3164 // shifted to the front by 1 after its deletion.
3165 auto Distance = PrevUnplacedBlockInFilterIt - It - 1;
3166 PrevUnplacedBlockInFilterIt = BlockFilter->erase(I: It) + Distance;
3167 assert(*PrevUnplacedBlockInFilterIt == PrevBB);
3168 (void)PrevBB;
3169 } else if (It == PrevUnplacedBlockInFilterIt)
3170 // The block pointed by PrevUnplacedBlockInFilterIt is erased, we
3171 // have to set it to the next element.
3172 PrevUnplacedBlockInFilterIt = BlockFilter->erase(I: It);
3173 else
3174 BlockFilter->erase(I: It);
3175 }
3176 }
3177
3178 // Remove the block from loop info.
3179 MLI->removeBlock(BB: RemBB);
3180 if (RemBB == PreferredLoopExit)
3181 PreferredLoopExit = nullptr;
3182
3183 LLVM_DEBUG(dbgs() << "TailDuplicator deleted block: "
3184 << getBlockName(RemBB) << "\n");
3185 };
3186 auto RemovalCallbackRef =
3187 function_ref<void(MachineBasicBlock*)>(RemovalCallback);
3188
3189 SmallVector<MachineBasicBlock *, 8> DuplicatedPreds;
3190 bool IsSimple = TailDup.isSimpleBB(TailBB: BB);
3191 SmallVector<MachineBasicBlock *, 8> CandidatePreds;
3192 SmallVectorImpl<MachineBasicBlock *> *CandidatePtr = nullptr;
3193 if (F->getFunction().hasProfileData()) {
3194 // We can do partial duplication with precise profile information.
3195 findDuplicateCandidates(Candidates&: CandidatePreds, BB, BlockFilter);
3196 if (CandidatePreds.size() == 0)
3197 return false;
3198 if (CandidatePreds.size() < BB->pred_size())
3199 CandidatePtr = &CandidatePreds;
3200 }
3201 TailDup.tailDuplicateAndUpdate(IsSimple, MBB: BB, ForcedLayoutPred: LPred, DuplicatedPreds: &DuplicatedPreds,
3202 RemovalCallback: &RemovalCallbackRef, CandidatePtr);
3203
3204 // Update UnscheduledPredecessors to reflect tail-duplication.
3205 DuplicatedToLPred = false;
3206 for (MachineBasicBlock *Pred : DuplicatedPreds) {
3207 // We're only looking for unscheduled predecessors that match the filter.
3208 BlockChain* PredChain = BlockToChain[Pred];
3209 if (Pred == LPred)
3210 DuplicatedToLPred = true;
3211 if (Pred == LPred || (BlockFilter && !BlockFilter->count(key: Pred))
3212 || PredChain == &Chain)
3213 continue;
3214 for (MachineBasicBlock *NewSucc : Pred->successors()) {
3215 if (BlockFilter && !BlockFilter->count(key: NewSucc))
3216 continue;
3217 BlockChain *NewChain = BlockToChain[NewSucc];
3218 if (NewChain != &Chain && NewChain != PredChain)
3219 NewChain->UnscheduledPredecessors++;
3220 }
3221 }
3222 return Removed;
3223}
3224
3225// Count the number of actual machine instructions.
3226static uint64_t countMBBInstruction(MachineBasicBlock *MBB) {
3227 uint64_t InstrCount = 0;
3228 for (MachineInstr &MI : *MBB) {
3229 if (!MI.isPHI() && !MI.isMetaInstruction())
3230 InstrCount += 1;
3231 }
3232 return InstrCount;
3233}
3234
3235// The size cost of duplication is the instruction size of the duplicated block.
3236// So we should scale the threshold accordingly. But the instruction size is not
3237// available on all targets, so we use the number of instructions instead.
3238BlockFrequency MachineBlockPlacement::scaleThreshold(MachineBasicBlock *BB) {
3239 return BlockFrequency(DupThreshold.getFrequency() * countMBBInstruction(MBB: BB));
3240}
3241
3242// Returns true if BB is Pred's best successor.
3243bool MachineBlockPlacement::isBestSuccessor(MachineBasicBlock *BB,
3244 MachineBasicBlock *Pred,
3245 BlockFilterSet *BlockFilter) {
3246 if (BB == Pred)
3247 return false;
3248 if (BlockFilter && !BlockFilter->count(key: Pred))
3249 return false;
3250 BlockChain *PredChain = BlockToChain[Pred];
3251 if (PredChain && (Pred != *std::prev(x: PredChain->end())))
3252 return false;
3253
3254 // Find the successor with largest probability excluding BB.
3255 BranchProbability BestProb = BranchProbability::getZero();
3256 for (MachineBasicBlock *Succ : Pred->successors())
3257 if (Succ != BB) {
3258 if (BlockFilter && !BlockFilter->count(key: Succ))
3259 continue;
3260 BlockChain *SuccChain = BlockToChain[Succ];
3261 if (SuccChain && (Succ != *SuccChain->begin()))
3262 continue;
3263 BranchProbability SuccProb = MBPI->getEdgeProbability(Src: Pred, Dst: Succ);
3264 if (SuccProb > BestProb)
3265 BestProb = SuccProb;
3266 }
3267
3268 BranchProbability BBProb = MBPI->getEdgeProbability(Src: Pred, Dst: BB);
3269 if (BBProb <= BestProb)
3270 return false;
3271
3272 // Compute the number of reduced taken branches if Pred falls through to BB
3273 // instead of another successor. Then compare it with threshold.
3274 BlockFrequency PredFreq = getBlockCountOrFrequency(BB: Pred);
3275 BlockFrequency Gain = PredFreq * (BBProb - BestProb);
3276 return Gain > scaleThreshold(BB);
3277}
3278
3279// Find out the predecessors of BB and BB can be beneficially duplicated into
3280// them.
3281void MachineBlockPlacement::findDuplicateCandidates(
3282 SmallVectorImpl<MachineBasicBlock *> &Candidates,
3283 MachineBasicBlock *BB,
3284 BlockFilterSet *BlockFilter) {
3285 MachineBasicBlock *Fallthrough = nullptr;
3286 BranchProbability DefaultBranchProb = BranchProbability::getZero();
3287 BlockFrequency BBDupThreshold(scaleThreshold(BB));
3288 SmallVector<MachineBasicBlock *, 8> Preds(BB->predecessors());
3289 SmallVector<MachineBasicBlock *, 8> Succs(BB->successors());
3290
3291 // Sort for highest frequency.
3292 auto CmpSucc = [&](MachineBasicBlock *A, MachineBasicBlock *B) {
3293 return MBPI->getEdgeProbability(Src: BB, Dst: A) > MBPI->getEdgeProbability(Src: BB, Dst: B);
3294 };
3295 auto CmpPred = [&](MachineBasicBlock *A, MachineBasicBlock *B) {
3296 return MBFI->getBlockFreq(MBB: A) > MBFI->getBlockFreq(MBB: B);
3297 };
3298 llvm::stable_sort(Range&: Succs, C: CmpSucc);
3299 llvm::stable_sort(Range&: Preds, C: CmpPred);
3300
3301 auto SuccIt = Succs.begin();
3302 if (SuccIt != Succs.end()) {
3303 DefaultBranchProb = MBPI->getEdgeProbability(Src: BB, Dst: *SuccIt).getCompl();
3304 }
3305
3306 // For each predecessors of BB, compute the benefit of duplicating BB,
3307 // if it is larger than the threshold, add it into Candidates.
3308 //
3309 // If we have following control flow.
3310 //
3311 // PB1 PB2 PB3 PB4
3312 // \ | / /\
3313 // \ | / / \
3314 // \ |/ / \
3315 // BB----/ OB
3316 // /\
3317 // / \
3318 // SB1 SB2
3319 //
3320 // And it can be partially duplicated as
3321 //
3322 // PB2+BB
3323 // | PB1 PB3 PB4
3324 // | | / /\
3325 // | | / / \
3326 // | |/ / \
3327 // | BB----/ OB
3328 // |\ /|
3329 // | X |
3330 // |/ \|
3331 // SB2 SB1
3332 //
3333 // The benefit of duplicating into a predecessor is defined as
3334 // Orig_taken_branch - Duplicated_taken_branch
3335 //
3336 // The Orig_taken_branch is computed with the assumption that predecessor
3337 // jumps to BB and the most possible successor is laid out after BB.
3338 //
3339 // The Duplicated_taken_branch is computed with the assumption that BB is
3340 // duplicated into PB, and one successor is layout after it (SB1 for PB1 and
3341 // SB2 for PB2 in our case). If there is no available successor, the combined
3342 // block jumps to all BB's successor, like PB3 in this example.
3343 //
3344 // If a predecessor has multiple successors, so BB can't be duplicated into
3345 // it. But it can beneficially fall through to BB, and duplicate BB into other
3346 // predecessors.
3347 for (MachineBasicBlock *Pred : Preds) {
3348 BlockFrequency PredFreq = getBlockCountOrFrequency(BB: Pred);
3349
3350 if (!TailDup.canTailDuplicate(TailBB: BB, PredBB: Pred)) {
3351 // BB can't be duplicated into Pred, but it is possible to be layout
3352 // below Pred.
3353 if (!Fallthrough && isBestSuccessor(BB, Pred, BlockFilter)) {
3354 Fallthrough = Pred;
3355 if (SuccIt != Succs.end())
3356 SuccIt++;
3357 }
3358 continue;
3359 }
3360
3361 BlockFrequency OrigCost = PredFreq + PredFreq * DefaultBranchProb;
3362 BlockFrequency DupCost;
3363 if (SuccIt == Succs.end()) {
3364 // Jump to all successors;
3365 if (Succs.size() > 0)
3366 DupCost += PredFreq;
3367 } else {
3368 // Fallthrough to *SuccIt, jump to all other successors;
3369 DupCost += PredFreq;
3370 DupCost -= PredFreq * MBPI->getEdgeProbability(Src: BB, Dst: *SuccIt);
3371 }
3372
3373 assert(OrigCost >= DupCost);
3374 OrigCost -= DupCost;
3375 if (OrigCost > BBDupThreshold) {
3376 Candidates.push_back(Elt: Pred);
3377 if (SuccIt != Succs.end())
3378 SuccIt++;
3379 }
3380 }
3381
3382 // No predecessors can optimally fallthrough to BB.
3383 // So we can change one duplication into fallthrough.
3384 if (!Fallthrough) {
3385 if ((Candidates.size() < Preds.size()) && (Candidates.size() > 0)) {
3386 Candidates[0] = Candidates.back();
3387 Candidates.pop_back();
3388 }
3389 }
3390}
3391
3392void MachineBlockPlacement::initDupThreshold() {
3393 DupThreshold = BlockFrequency(0);
3394 if (!F->getFunction().hasProfileData())
3395 return;
3396
3397 // We prefer to use prifile count.
3398 uint64_t HotThreshold = PSI->getOrCompHotCountThreshold();
3399 if (HotThreshold != UINT64_MAX) {
3400 UseProfileCount = true;
3401 DupThreshold =
3402 BlockFrequency(HotThreshold * TailDupProfilePercentThreshold / 100);
3403 return;
3404 }
3405
3406 // Profile count is not available, we can use block frequency instead.
3407 BlockFrequency MaxFreq = BlockFrequency(0);
3408 for (MachineBasicBlock &MBB : *F) {
3409 BlockFrequency Freq = MBFI->getBlockFreq(MBB: &MBB);
3410 if (Freq > MaxFreq)
3411 MaxFreq = Freq;
3412 }
3413
3414 BranchProbability ThresholdProb(TailDupPlacementPenalty, 100);
3415 DupThreshold = BlockFrequency(MaxFreq * ThresholdProb);
3416 UseProfileCount = false;
3417}
3418
3419bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &MF) {
3420 if (skipFunction(F: MF.getFunction()))
3421 return false;
3422
3423 // Check for single-block functions and skip them.
3424 if (std::next(x: MF.begin()) == MF.end())
3425 return false;
3426
3427 F = &MF;
3428 MBPI = &getAnalysis<MachineBranchProbabilityInfoWrapperPass>().getMBPI();
3429 MBFI = std::make_unique<MBFIWrapper>(
3430 args&: getAnalysis<MachineBlockFrequencyInfoWrapperPass>().getMBFI());
3431 MLI = &getAnalysis<MachineLoopInfoWrapperPass>().getLI();
3432 TII = MF.getSubtarget().getInstrInfo();
3433 TLI = MF.getSubtarget().getTargetLowering();
3434 MPDT = nullptr;
3435 PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
3436
3437 initDupThreshold();
3438
3439 // Initialize PreferredLoopExit to nullptr here since it may never be set if
3440 // there are no MachineLoops.
3441 PreferredLoopExit = nullptr;
3442
3443 assert(BlockToChain.empty() &&
3444 "BlockToChain map should be empty before starting placement.");
3445 assert(ComputedEdges.empty() &&
3446 "Computed Edge map should be empty before starting placement.");
3447
3448 unsigned TailDupSize = TailDupPlacementThreshold;
3449 // If only the aggressive threshold is explicitly set, use it.
3450 if (TailDupPlacementAggressiveThreshold.getNumOccurrences() != 0 &&
3451 TailDupPlacementThreshold.getNumOccurrences() == 0)
3452 TailDupSize = TailDupPlacementAggressiveThreshold;
3453
3454 TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
3455 // For aggressive optimization, we can adjust some thresholds to be less
3456 // conservative.
3457 if (PassConfig->getOptLevel() >= CodeGenOptLevel::Aggressive) {
3458 // At O3 we should be more willing to copy blocks for tail duplication. This
3459 // increases size pressure, so we only do it at O3
3460 // Do this unless only the regular threshold is explicitly set.
3461 if (TailDupPlacementThreshold.getNumOccurrences() == 0 ||
3462 TailDupPlacementAggressiveThreshold.getNumOccurrences() != 0)
3463 TailDupSize = TailDupPlacementAggressiveThreshold;
3464 }
3465
3466 // If there's no threshold provided through options, query the target
3467 // information for a threshold instead.
3468 if (TailDupPlacementThreshold.getNumOccurrences() == 0 &&
3469 (PassConfig->getOptLevel() < CodeGenOptLevel::Aggressive ||
3470 TailDupPlacementAggressiveThreshold.getNumOccurrences() == 0))
3471 TailDupSize = TII->getTailDuplicateSize(OptLevel: PassConfig->getOptLevel());
3472
3473 if (allowTailDupPlacement()) {
3474 MPDT = &getAnalysis<MachinePostDominatorTreeWrapperPass>().getPostDomTree();
3475 bool OptForSize = MF.getFunction().hasOptSize() ||
3476 llvm::shouldOptimizeForSize(MF: &MF, PSI, BFI: &MBFI->getMBFI());
3477 if (OptForSize)
3478 TailDupSize = 1;
3479 bool PreRegAlloc = false;
3480 TailDup.initMF(MF, PreRegAlloc, MBPI, MBFI: MBFI.get(), PSI,
3481 /* LayoutMode */ true, TailDupSize);
3482 precomputeTriangleChains();
3483 }
3484
3485 buildCFGChains();
3486
3487 // Changing the layout can create new tail merging opportunities.
3488 // TailMerge can create jump into if branches that make CFG irreducible for
3489 // HW that requires structured CFG.
3490 bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
3491 PassConfig->getEnableTailMerge() &&
3492 BranchFoldPlacement;
3493 // No tail merging opportunities if the block number is less than four.
3494 if (MF.size() > 3 && EnableTailMerge) {
3495 unsigned TailMergeSize = TailDupSize + 1;
3496 BranchFolder BF(/*DefaultEnableTailMerge=*/true, /*CommonHoist=*/false,
3497 *MBFI, *MBPI, PSI, TailMergeSize);
3498
3499 if (BF.OptimizeFunction(MF, tii: TII, tri: MF.getSubtarget().getRegisterInfo(), mli: MLI,
3500 /*AfterPlacement=*/true)) {
3501 // Redo the layout if tail merging creates/removes/moves blocks.
3502 BlockToChain.clear();
3503 ComputedEdges.clear();
3504 // Must redo the post-dominator tree if blocks were changed.
3505 if (MPDT)
3506 MPDT->recalculate(Func&: MF);
3507 ChainAllocator.DestroyAll();
3508 buildCFGChains();
3509 }
3510 }
3511
3512 // Apply a post-processing optimizing block placement.
3513 if (MF.size() >= 3 && EnableExtTspBlockPlacement &&
3514 (ApplyExtTspWithoutProfile || MF.getFunction().hasProfileData())) {
3515 // Find a new placement and modify the layout of the blocks in the function.
3516 applyExtTsp();
3517
3518 // Re-create CFG chain so that we can optimizeBranches and alignBlocks.
3519 createCFGChainExtTsp();
3520 }
3521
3522 optimizeBranches();
3523 alignBlocks();
3524
3525 BlockToChain.clear();
3526 ComputedEdges.clear();
3527 ChainAllocator.DestroyAll();
3528
3529 bool HasMaxBytesOverride =
3530 MaxBytesForAlignmentOverride.getNumOccurrences() > 0;
3531
3532 if (AlignAllBlock)
3533 // Align all of the blocks in the function to a specific alignment.
3534 for (MachineBasicBlock &MBB : MF) {
3535 if (HasMaxBytesOverride)
3536 MBB.setAlignment(A: Align(1ULL << AlignAllBlock),
3537 MaxBytes: MaxBytesForAlignmentOverride);
3538 else
3539 MBB.setAlignment(Align(1ULL << AlignAllBlock));
3540 }
3541 else if (AlignAllNonFallThruBlocks) {
3542 // Align all of the blocks that have no fall-through predecessors to a
3543 // specific alignment.
3544 for (auto MBI = std::next(x: MF.begin()), MBE = MF.end(); MBI != MBE; ++MBI) {
3545 auto LayoutPred = std::prev(x: MBI);
3546 if (!LayoutPred->isSuccessor(MBB: &*MBI)) {
3547 if (HasMaxBytesOverride)
3548 MBI->setAlignment(A: Align(1ULL << AlignAllNonFallThruBlocks),
3549 MaxBytes: MaxBytesForAlignmentOverride);
3550 else
3551 MBI->setAlignment(Align(1ULL << AlignAllNonFallThruBlocks));
3552 }
3553 }
3554 }
3555 if (ViewBlockLayoutWithBFI != GVDT_None &&
3556 (ViewBlockFreqFuncName.empty() ||
3557 F->getFunction().getName() == ViewBlockFreqFuncName)) {
3558 if (RenumberBlocksBeforeView)
3559 MF.RenumberBlocks();
3560 MBFI->view(Name: "MBP." + MF.getName(), isSimple: false);
3561 }
3562
3563 // We always return true as we have no way to track whether the final order
3564 // differs from the original order.
3565 return true;
3566}
3567
3568void MachineBlockPlacement::applyExtTsp() {
3569 // Prepare data; blocks are indexed by their index in the current ordering.
3570 DenseMap<const MachineBasicBlock *, uint64_t> BlockIndex;
3571 BlockIndex.reserve(NumEntries: F->size());
3572 std::vector<const MachineBasicBlock *> CurrentBlockOrder;
3573 CurrentBlockOrder.reserve(n: F->size());
3574 size_t NumBlocks = 0;
3575 for (const MachineBasicBlock &MBB : *F) {
3576 BlockIndex[&MBB] = NumBlocks++;
3577 CurrentBlockOrder.push_back(x: &MBB);
3578 }
3579
3580 auto BlockSizes = std::vector<uint64_t>(F->size());
3581 auto BlockCounts = std::vector<uint64_t>(F->size());
3582 std::vector<codelayout::EdgeCount> JumpCounts;
3583 for (MachineBasicBlock &MBB : *F) {
3584 // Getting the block frequency.
3585 BlockFrequency BlockFreq = MBFI->getBlockFreq(MBB: &MBB);
3586 BlockCounts[BlockIndex[&MBB]] = BlockFreq.getFrequency();
3587 // Getting the block size:
3588 // - approximate the size of an instruction by 4 bytes, and
3589 // - ignore debug instructions.
3590 // Note: getting the exact size of each block is target-dependent and can be
3591 // done by extending the interface of MCCodeEmitter. Experimentally we do
3592 // not see a perf improvement with the exact block sizes.
3593 auto NonDbgInsts =
3594 instructionsWithoutDebug(It: MBB.instr_begin(), End: MBB.instr_end());
3595 int NumInsts = std::distance(first: NonDbgInsts.begin(), last: NonDbgInsts.end());
3596 BlockSizes[BlockIndex[&MBB]] = 4 * NumInsts;
3597 // Getting jump frequencies.
3598 for (MachineBasicBlock *Succ : MBB.successors()) {
3599 auto EP = MBPI->getEdgeProbability(Src: &MBB, Dst: Succ);
3600 BlockFrequency JumpFreq = BlockFreq * EP;
3601 JumpCounts.push_back(
3602 x: {.src: BlockIndex[&MBB], .dst: BlockIndex[Succ], .count: JumpFreq.getFrequency()});
3603 }
3604 }
3605
3606 LLVM_DEBUG(dbgs() << "Applying ext-tsp layout for |V| = " << F->size()
3607 << " with profile = " << F->getFunction().hasProfileData()
3608 << " (" << F->getName().str() << ")"
3609 << "\n");
3610 LLVM_DEBUG(
3611 dbgs() << format(" original layout score: %0.2f\n",
3612 calcExtTspScore(BlockSizes, BlockCounts, JumpCounts)));
3613
3614 // Run the layout algorithm.
3615 auto NewOrder = computeExtTspLayout(NodeSizes: BlockSizes, NodeCounts: BlockCounts, EdgeCounts: JumpCounts);
3616 std::vector<const MachineBasicBlock *> NewBlockOrder;
3617 NewBlockOrder.reserve(n: F->size());
3618 for (uint64_t Node : NewOrder) {
3619 NewBlockOrder.push_back(x: CurrentBlockOrder[Node]);
3620 }
3621 LLVM_DEBUG(dbgs() << format(" optimized layout score: %0.2f\n",
3622 calcExtTspScore(NewOrder, BlockSizes, BlockCounts,
3623 JumpCounts)));
3624
3625 // Assign new block order.
3626 assignBlockOrder(NewOrder: NewBlockOrder);
3627}
3628
3629void MachineBlockPlacement::assignBlockOrder(
3630 const std::vector<const MachineBasicBlock *> &NewBlockOrder) {
3631 assert(F->size() == NewBlockOrder.size() && "Incorrect size of block order");
3632 F->RenumberBlocks();
3633
3634 bool HasChanges = false;
3635 for (size_t I = 0; I < NewBlockOrder.size(); I++) {
3636 if (NewBlockOrder[I] != F->getBlockNumbered(N: I)) {
3637 HasChanges = true;
3638 break;
3639 }
3640 }
3641 // Stop early if the new block order is identical to the existing one.
3642 if (!HasChanges)
3643 return;
3644
3645 SmallVector<MachineBasicBlock *, 4> PrevFallThroughs(F->getNumBlockIDs());
3646 for (auto &MBB : *F) {
3647 PrevFallThroughs[MBB.getNumber()] = MBB.getFallThrough();
3648 }
3649
3650 // Sort basic blocks in the function according to the computed order.
3651 DenseMap<const MachineBasicBlock *, size_t> NewIndex;
3652 for (const MachineBasicBlock *MBB : NewBlockOrder) {
3653 NewIndex[MBB] = NewIndex.size();
3654 }
3655 F->sort(comp: [&](MachineBasicBlock &L, MachineBasicBlock &R) {
3656 return NewIndex[&L] < NewIndex[&R];
3657 });
3658
3659 // Update basic block branches by inserting explicit fallthrough branches
3660 // when required and re-optimize branches when possible.
3661 const TargetInstrInfo *TII = F->getSubtarget().getInstrInfo();
3662 SmallVector<MachineOperand, 4> Cond;
3663 for (auto &MBB : *F) {
3664 MachineFunction::iterator NextMBB = std::next(x: MBB.getIterator());
3665 MachineFunction::iterator EndIt = MBB.getParent()->end();
3666 auto *FTMBB = PrevFallThroughs[MBB.getNumber()];
3667 // If this block had a fallthrough before we need an explicit unconditional
3668 // branch to that block if the fallthrough block is not adjacent to the
3669 // block in the new order.
3670 if (FTMBB && (NextMBB == EndIt || &*NextMBB != FTMBB)) {
3671 TII->insertUnconditionalBranch(MBB, DestBB: FTMBB, DL: MBB.findBranchDebugLoc());
3672 }
3673
3674 // It might be possible to optimize branches by flipping the condition.
3675 Cond.clear();
3676 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
3677 if (TII->analyzeBranch(MBB, TBB, FBB, Cond))
3678 continue;
3679 MBB.updateTerminator(PreviousLayoutSuccessor: FTMBB);
3680 }
3681
3682#ifndef NDEBUG
3683 // Make sure we correctly constructed all branches.
3684 F->verify(this, "After optimized block reordering");
3685#endif
3686}
3687
3688void MachineBlockPlacement::createCFGChainExtTsp() {
3689 BlockToChain.clear();
3690 ComputedEdges.clear();
3691 ChainAllocator.DestroyAll();
3692
3693 MachineBasicBlock *HeadBB = &F->front();
3694 BlockChain *FunctionChain =
3695 new (ChainAllocator.Allocate()) BlockChain(BlockToChain, HeadBB);
3696
3697 for (MachineBasicBlock &MBB : *F) {
3698 if (HeadBB == &MBB)
3699 continue; // Ignore head of the chain
3700 FunctionChain->merge(BB: &MBB, Chain: nullptr);
3701 }
3702}
3703
3704namespace {
3705
3706/// A pass to compute block placement statistics.
3707///
3708/// A separate pass to compute interesting statistics for evaluating block
3709/// placement. This is separate from the actual placement pass so that they can
3710/// be computed in the absence of any placement transformations or when using
3711/// alternative placement strategies.
3712class MachineBlockPlacementStats : public MachineFunctionPass {
3713 /// A handle to the branch probability pass.
3714 const MachineBranchProbabilityInfo *MBPI;
3715
3716 /// A handle to the function-wide block frequency pass.
3717 const MachineBlockFrequencyInfo *MBFI;
3718
3719public:
3720 static char ID; // Pass identification, replacement for typeid
3721
3722 MachineBlockPlacementStats() : MachineFunctionPass(ID) {
3723 initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry());
3724 }
3725
3726 bool runOnMachineFunction(MachineFunction &F) override;
3727
3728 void getAnalysisUsage(AnalysisUsage &AU) const override {
3729 AU.addRequired<MachineBranchProbabilityInfoWrapperPass>();
3730 AU.addRequired<MachineBlockFrequencyInfoWrapperPass>();
3731 AU.setPreservesAll();
3732 MachineFunctionPass::getAnalysisUsage(AU);
3733 }
3734};
3735
3736} // end anonymous namespace
3737
3738char MachineBlockPlacementStats::ID = 0;
3739
3740char &llvm::MachineBlockPlacementStatsID = MachineBlockPlacementStats::ID;
3741
3742INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats",
3743 "Basic Block Placement Stats", false, false)
3744INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfoWrapperPass)
3745INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfoWrapperPass)
3746INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats",
3747 "Basic Block Placement Stats", false, false)
3748
3749bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) {
3750 // Check for single-block functions and skip them.
3751 if (std::next(x: F.begin()) == F.end())
3752 return false;
3753
3754 if (!isFunctionInPrintList(FunctionName: F.getName()))
3755 return false;
3756
3757 MBPI = &getAnalysis<MachineBranchProbabilityInfoWrapperPass>().getMBPI();
3758 MBFI = &getAnalysis<MachineBlockFrequencyInfoWrapperPass>().getMBFI();
3759
3760 for (MachineBasicBlock &MBB : F) {
3761 BlockFrequency BlockFreq = MBFI->getBlockFreq(MBB: &MBB);
3762 Statistic &NumBranches =
3763 (MBB.succ_size() > 1) ? NumCondBranches : NumUncondBranches;
3764 Statistic &BranchTakenFreq =
3765 (MBB.succ_size() > 1) ? CondBranchTakenFreq : UncondBranchTakenFreq;
3766 for (MachineBasicBlock *Succ : MBB.successors()) {
3767 // Skip if this successor is a fallthrough.
3768 if (MBB.isLayoutSuccessor(MBB: Succ))
3769 continue;
3770
3771 BlockFrequency EdgeFreq =
3772 BlockFreq * MBPI->getEdgeProbability(Src: &MBB, Dst: Succ);
3773 ++NumBranches;
3774 BranchTakenFreq += EdgeFreq.getFrequency();
3775 }
3776 }
3777
3778 return false;
3779}
3780