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 "llvm/CodeGen/MachineBlockPlacement.h"
28#include "BranchFolding.h"
29#include "llvm/ADT/ArrayRef.h"
30#include "llvm/ADT/DenseMap.h"
31#include "llvm/ADT/STLExtras.h"
32#include "llvm/ADT/SetVector.h"
33#include "llvm/ADT/SmallPtrSet.h"
34#include "llvm/ADT/SmallVector.h"
35#include "llvm/ADT/Statistic.h"
36#include "llvm/Analysis/BlockFrequencyInfoImpl.h"
37#include "llvm/Analysis/ProfileSummaryInfo.h"
38#include "llvm/CodeGen/MBFIWrapper.h"
39#include "llvm/CodeGen/MachineBasicBlock.h"
40#include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
41#include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
42#include "llvm/CodeGen/MachineFunction.h"
43#include "llvm/CodeGen/MachineFunctionPass.h"
44#include "llvm/CodeGen/MachineLoopInfo.h"
45#include "llvm/CodeGen/MachinePostDominators.h"
46#include "llvm/CodeGen/MachineSizeOpts.h"
47#include "llvm/CodeGen/TailDuplicator.h"
48#include "llvm/CodeGen/TargetInstrInfo.h"
49#include "llvm/CodeGen/TargetLowering.h"
50#include "llvm/CodeGen/TargetPassConfig.h"
51#include "llvm/CodeGen/TargetSubtargetInfo.h"
52#include "llvm/IR/DebugLoc.h"
53#include "llvm/IR/Function.h"
54#include "llvm/IR/PrintPasses.h"
55#include "llvm/InitializePasses.h"
56#include "llvm/Pass.h"
57#include "llvm/Support/Allocator.h"
58#include "llvm/Support/BlockFrequency.h"
59#include "llvm/Support/BranchProbability.h"
60#include "llvm/Support/CodeGen.h"
61#include "llvm/Support/CommandLine.h"
62#include "llvm/Support/Compiler.h"
63#include "llvm/Support/Debug.h"
64#include "llvm/Support/raw_ostream.h"
65#include "llvm/Target/TargetMachine.h"
66#include "llvm/Transforms/Utils/CodeLayout.h"
67#include <algorithm>
68#include <cassert>
69#include <cstdint>
70#include <iterator>
71#include <memory>
72#include <string>
73#include <tuple>
74#include <utility>
75#include <vector>
76
77using namespace llvm;
78
79#define DEBUG_TYPE "block-placement"
80
81STATISTIC(NumCondBranches, "Number of conditional branches");
82STATISTIC(NumUncondBranches, "Number of unconditional branches");
83STATISTIC(CondBranchTakenFreq,
84 "Potential frequency of taking conditional branches");
85STATISTIC(UncondBranchTakenFreq,
86 "Potential frequency of taking unconditional branches");
87
88static cl::opt<unsigned> AlignAllBlock(
89 "align-all-blocks",
90 cl::desc("Force the alignment of all blocks in the function in log2 format "
91 "(e.g 4 means align on 16B boundaries)."),
92 cl::init(Val: 0), cl::Hidden);
93
94static cl::opt<unsigned> AlignAllNonFallThruBlocks(
95 "align-all-nofallthru-blocks",
96 cl::desc("Force the alignment of all blocks that have no fall-through "
97 "predecessors (i.e. don't add nops that are executed). In log2 "
98 "format (e.g 4 means align on 16B boundaries)."),
99 cl::init(Val: 0), cl::Hidden);
100
101static cl::opt<unsigned> MaxBytesForAlignmentOverride(
102 "max-bytes-for-alignment",
103 cl::desc("Forces the maximum bytes allowed to be emitted when padding for "
104 "alignment"),
105 cl::init(Val: 0), cl::Hidden);
106
107static cl::opt<unsigned> PredecessorLimit(
108 "block-placement-predecessor-limit",
109 cl::desc("For blocks with more predecessors, certain layout optimizations"
110 "will be disabled to prevent quadratic compile time."),
111 cl::init(Val: 1000), cl::Hidden);
112
113// FIXME: Find a good default for this flag and remove the flag.
114static cl::opt<unsigned> ExitBlockBias(
115 "block-placement-exit-block-bias",
116 cl::desc("Block frequency percentage a loop exit block needs "
117 "over the original exit to be considered the new exit."),
118 cl::init(Val: 0), cl::Hidden);
119
120// Definition:
121// - Outlining: placement of a basic block outside the chain or hot path.
122
123static cl::opt<unsigned> LoopToColdBlockRatio(
124 "loop-to-cold-block-ratio",
125 cl::desc("Outline loop blocks from loop chain if (frequency of loop) / "
126 "(frequency of block) is greater than this ratio"),
127 cl::init(Val: 5), cl::Hidden);
128
129static cl::opt<bool>
130 ForceLoopColdBlock("force-loop-cold-block",
131 cl::desc("Force outlining cold blocks from loops."),
132 cl::init(Val: false), cl::Hidden);
133
134static cl::opt<bool>
135 PreciseRotationCost("precise-rotation-cost",
136 cl::desc("Model the cost of loop rotation more "
137 "precisely by using profile data."),
138 cl::init(Val: false), cl::Hidden);
139
140static cl::opt<bool>
141 ForcePreciseRotationCost("force-precise-rotation-cost",
142 cl::desc("Force the use of precise cost "
143 "loop rotation strategy."),
144 cl::init(Val: false), cl::Hidden);
145
146static cl::opt<unsigned> MisfetchCost(
147 "misfetch-cost",
148 cl::desc("Cost that models the probabilistic risk of an instruction "
149 "misfetch due to a jump comparing to falling through, whose cost "
150 "is zero."),
151 cl::init(Val: 1), cl::Hidden);
152
153static cl::opt<unsigned> JumpInstCost("jump-inst-cost",
154 cl::desc("Cost of jump instructions."),
155 cl::init(Val: 1), cl::Hidden);
156static cl::opt<bool>
157 TailDupPlacement("tail-dup-placement",
158 cl::desc("Perform tail duplication during placement. "
159 "Creates more fallthrough opportunities in "
160 "outline branches."),
161 cl::init(Val: true), cl::Hidden);
162
163static cl::opt<bool>
164 BranchFoldPlacement("branch-fold-placement",
165 cl::desc("Perform branch folding during placement. "
166 "Reduces code size."),
167 cl::init(Val: true), cl::Hidden);
168
169// Heuristic for tail duplication.
170static cl::opt<unsigned> TailDupPlacementThreshold(
171 "tail-dup-placement-threshold",
172 cl::desc("Instruction cutoff for tail duplication during layout. "
173 "Tail merging during layout is forced to have a threshold "
174 "that won't conflict."),
175 cl::init(Val: 2), cl::Hidden);
176
177// Heuristic for aggressive tail duplication.
178static cl::opt<unsigned> TailDupPlacementAggressiveThreshold(
179 "tail-dup-placement-aggressive-threshold",
180 cl::desc("Instruction cutoff for aggressive tail duplication during "
181 "layout. Used at -O3. Tail merging during layout is forced to "
182 "have a threshold that won't conflict."),
183 cl::init(Val: 4), cl::Hidden);
184
185// Heuristic for tail duplication.
186static cl::opt<unsigned> TailDupPlacementPenalty(
187 "tail-dup-placement-penalty",
188 cl::desc(
189 "Cost penalty for blocks that can avoid breaking CFG by copying. "
190 "Copying can increase fallthrough, but it also increases icache "
191 "pressure. This parameter controls the penalty to account for that. "
192 "Percent as integer."),
193 cl::init(Val: 2), cl::Hidden);
194
195// Heuristic for tail duplication if profile count is used in cost model.
196static cl::opt<unsigned> TailDupProfilePercentThreshold(
197 "tail-dup-profile-percent-threshold",
198 cl::desc("If profile count information is used in tail duplication cost "
199 "model, the gained fall through number from tail duplication "
200 "should be at least this percent of hot count."),
201 cl::init(Val: 50), cl::Hidden);
202
203// Heuristic for triangle chains.
204static cl::opt<unsigned> TriangleChainCount(
205 "triangle-chain-count",
206 cl::desc("Number of triangle-shaped-CFG's that need to be in a row for the "
207 "triangle tail duplication heuristic to kick in. 0 to disable."),
208 cl::init(Val: 2), cl::Hidden);
209
210// Use case: When block layout is visualized after MBP pass, the basic blocks
211// are labeled in layout order; meanwhile blocks could be numbered in a
212// different order. It's hard to map between the graph and pass output.
213// With this option on, the basic blocks are renumbered in function layout
214// order. For debugging only.
215static cl::opt<bool> RenumberBlocksBeforeView(
216 "renumber-blocks-before-view",
217 cl::desc(
218 "If true, basic blocks are re-numbered before MBP layout is printed "
219 "into a dot graph. Only used when a function is being printed."),
220 cl::init(Val: false), cl::Hidden);
221
222static cl::opt<unsigned> ExtTspBlockPlacementMaxBlocks(
223 "ext-tsp-block-placement-max-blocks",
224 cl::desc("Maximum number of basic blocks in a function to run ext-TSP "
225 "block placement."),
226 cl::init(UINT_MAX), cl::Hidden);
227
228// Apply the ext-tsp algorithm minimizing the size of a binary.
229static cl::opt<bool>
230 ApplyExtTspForSize("apply-ext-tsp-for-size", cl::init(Val: false), cl::Hidden,
231 cl::desc("Use ext-tsp for size-aware block placement."));
232
233namespace llvm {
234extern cl::opt<bool> EnableExtTspBlockPlacement;
235extern cl::opt<bool> ApplyExtTspWithoutProfile;
236extern cl::opt<unsigned> StaticLikelyProb;
237extern cl::opt<unsigned> ProfileLikelyProb;
238
239// Internal option used to control BFI display only after MBP pass.
240// Defined in CodeGen/MachineBlockFrequencyInfo.cpp:
241// -view-block-layout-with-bfi=
242extern cl::opt<GVDAGType> ViewBlockLayoutWithBFI;
243
244// Command line option to specify the name of the function for CFG dump
245// Defined in Analysis/BlockFrequencyInfo.cpp: -view-bfi-func-name=
246extern cl::opt<std::string> ViewBlockFreqFuncName;
247} // namespace llvm
248
249namespace {
250
251class BlockChain;
252
253/// Type for our function-wide basic block -> block chain mapping.
254using BlockToChainMapType = DenseMap<const MachineBasicBlock *, BlockChain *>;
255
256/// A chain of blocks which will be laid out contiguously.
257///
258/// This is the datastructure representing a chain of consecutive blocks that
259/// are profitable to layout together in order to maximize fallthrough
260/// probabilities and code locality. We also can use a block chain to represent
261/// a sequence of basic blocks which have some external (correctness)
262/// requirement for sequential layout.
263///
264/// Chains can be built around a single basic block and can be merged to grow
265/// them. They participate in a block-to-chain mapping, which is updated
266/// automatically as chains are merged together.
267class BlockChain {
268 /// The sequence of blocks belonging to this chain.
269 ///
270 /// This is the sequence of blocks for a particular chain. These will be laid
271 /// out in-order within the function.
272 SmallVector<MachineBasicBlock *, 4> Blocks;
273
274 /// A handle to the function-wide basic block to block chain mapping.
275 ///
276 /// This is retained in each block chain to simplify the computation of child
277 /// block chains for SCC-formation and iteration. We store the edges to child
278 /// basic blocks, and map them back to their associated chains using this
279 /// structure.
280 BlockToChainMapType &BlockToChain;
281
282public:
283 /// Construct a new BlockChain.
284 ///
285 /// This builds a new block chain representing a single basic block in the
286 /// function. It also registers itself as the chain that block participates
287 /// in with the BlockToChain mapping.
288 BlockChain(BlockToChainMapType &BlockToChain, MachineBasicBlock *BB)
289 : Blocks(1, BB), BlockToChain(BlockToChain) {
290 assert(BB && "Cannot create a chain with a null basic block");
291 BlockToChain[BB] = this;
292 }
293
294 /// Iterator over blocks within the chain.
295 using iterator = SmallVectorImpl<MachineBasicBlock *>::iterator;
296 using const_iterator = SmallVectorImpl<MachineBasicBlock *>::const_iterator;
297
298 /// Beginning of blocks within the chain.
299 iterator begin() { return Blocks.begin(); }
300 const_iterator begin() const { return Blocks.begin(); }
301
302 /// End of blocks within the chain.
303 iterator end() { return Blocks.end(); }
304 const_iterator end() const { return Blocks.end(); }
305
306 bool remove(MachineBasicBlock *BB) {
307 for (iterator i = begin(); i != end(); ++i) {
308 if (*i == BB) {
309 Blocks.erase(CI: i);
310 return true;
311 }
312 }
313 return false;
314 }
315
316 /// Merge a block chain into this one.
317 ///
318 /// This routine merges a block chain into this one. It takes care of forming
319 /// a contiguous sequence of basic blocks, updating the edge list, and
320 /// updating the block -> chain mapping. It does not free or tear down the
321 /// old chain, but the old chain's block list is no longer valid.
322 void merge(MachineBasicBlock *BB, BlockChain *Chain) {
323 assert(BB && "Can't merge a null block.");
324 assert(!Blocks.empty() && "Can't merge into an empty chain.");
325
326 // Fast path in case we don't have a chain already.
327 if (!Chain) {
328 assert(!BlockToChain[BB] &&
329 "Passed chain is null, but BB has entry in BlockToChain.");
330 Blocks.push_back(Elt: BB);
331 BlockToChain[BB] = this;
332 return;
333 }
334
335 assert(BB == *Chain->begin() && "Passed BB is not head of Chain.");
336 assert(Chain->begin() != Chain->end());
337
338 // Update the incoming blocks to point to this chain, and add them to the
339 // chain structure.
340 for (MachineBasicBlock *ChainBB : *Chain) {
341 Blocks.push_back(Elt: ChainBB);
342 assert(BlockToChain[ChainBB] == Chain && "Incoming blocks not in chain.");
343 BlockToChain[ChainBB] = this;
344 }
345 }
346
347#ifndef NDEBUG
348 /// Dump the blocks in this chain.
349 LLVM_DUMP_METHOD void dump() {
350 for (MachineBasicBlock *MBB : *this)
351 MBB->dump();
352 }
353#endif // NDEBUG
354
355 /// Count of predecessors of any block within the chain which have not
356 /// yet been scheduled. In general, we will delay scheduling this chain
357 /// until those predecessors are scheduled (or we find a sufficiently good
358 /// reason to override this heuristic.) Note that when forming loop chains,
359 /// blocks outside the loop are ignored and treated as if they were already
360 /// scheduled.
361 ///
362 /// Note: This field is reinitialized multiple times - once for each loop,
363 /// and then once for the function as a whole.
364 unsigned UnscheduledPredecessors = 0;
365};
366
367class MachineBlockPlacement {
368 /// A type for a block filter set.
369 using BlockFilterSet = SmallSetVector<const MachineBasicBlock *, 16>;
370
371 /// Pair struct containing basic block and taildup profitability
372 struct BlockAndTailDupResult {
373 MachineBasicBlock *BB = nullptr;
374 bool ShouldTailDup;
375 };
376
377 /// Triple struct containing edge weight and the edge.
378 struct WeightedEdge {
379 BlockFrequency Weight;
380 MachineBasicBlock *Src = nullptr;
381 MachineBasicBlock *Dest = nullptr;
382 };
383
384 /// work lists of blocks that are ready to be laid out
385 SmallVector<MachineBasicBlock *, 16> BlockWorkList;
386 SmallVector<MachineBasicBlock *, 16> EHPadWorkList;
387
388 /// Edges that have already been computed as optimal.
389 DenseMap<const MachineBasicBlock *, BlockAndTailDupResult> ComputedEdges;
390
391 /// Machine Function
392 MachineFunction *F = nullptr;
393
394 /// A handle to the branch probability pass.
395 const MachineBranchProbabilityInfo *MBPI = nullptr;
396
397 /// A handle to the function-wide block frequency pass.
398 std::unique_ptr<MBFIWrapper> MBFI;
399
400 /// A handle to the loop info.
401 MachineLoopInfo *MLI = nullptr;
402
403 /// Preferred loop exit.
404 /// Member variable for convenience. It may be removed by duplication deep
405 /// in the call stack.
406 MachineBasicBlock *PreferredLoopExit = nullptr;
407
408 /// A handle to the target's instruction info.
409 const TargetInstrInfo *TII = nullptr;
410
411 /// A handle to the target's lowering info.
412 const TargetLoweringBase *TLI = nullptr;
413
414 /// A handle to the post dominator tree.
415 MachinePostDominatorTree *MPDT = nullptr;
416
417 ProfileSummaryInfo *PSI = nullptr;
418
419 // Tail merging is also determined based on
420 // whether structured CFG is required.
421 bool AllowTailMerge;
422
423 CodeGenOptLevel OptLevel;
424
425 /// Duplicator used to duplicate tails during placement.
426 ///
427 /// Placement decisions can open up new tail duplication opportunities, but
428 /// since tail duplication affects placement decisions of later blocks, it
429 /// must be done inline.
430 TailDuplicator TailDup;
431
432 /// Partial tail duplication threshold.
433 BlockFrequency DupThreshold;
434
435 unsigned TailDupSize;
436
437 /// True: use block profile count to compute tail duplication cost.
438 /// False: use block frequency to compute tail duplication cost.
439 bool UseProfileCount = false;
440
441 /// Allocator and owner of BlockChain structures.
442 ///
443 /// We build BlockChains lazily while processing the loop structure of
444 /// a function. To reduce malloc traffic, we allocate them using this
445 /// slab-like allocator, and destroy them after the pass completes. An
446 /// important guarantee is that this allocator produces stable pointers to
447 /// the chains.
448 SpecificBumpPtrAllocator<BlockChain> ChainAllocator;
449
450 /// Function wide BasicBlock to BlockChain mapping.
451 ///
452 /// This mapping allows efficiently moving from any given basic block to the
453 /// BlockChain it participates in, if any. We use it to, among other things,
454 /// allow implicitly defining edges between chains as the existing edges
455 /// between basic blocks.
456 DenseMap<const MachineBasicBlock *, BlockChain *> BlockToChain;
457
458#ifndef NDEBUG
459 /// The set of basic blocks that have terminators that cannot be fully
460 /// analyzed. These basic blocks cannot be re-ordered safely by
461 /// MachineBlockPlacement, and we must preserve physical layout of these
462 /// blocks and their successors through the pass.
463 SmallPtrSet<MachineBasicBlock *, 4> BlocksWithUnanalyzableExits;
464#endif
465
466 /// Get block profile count or frequency according to UseProfileCount.
467 /// The return value is used to model tail duplication cost.
468 BlockFrequency getBlockCountOrFrequency(const MachineBasicBlock *BB) {
469 if (UseProfileCount) {
470 auto Count = MBFI->getBlockProfileCount(MBB: BB);
471 if (Count)
472 return BlockFrequency(*Count);
473 else
474 return BlockFrequency(0);
475 } else
476 return MBFI->getBlockFreq(MBB: BB);
477 }
478
479 /// Scale the DupThreshold according to basic block size.
480 BlockFrequency scaleThreshold(MachineBasicBlock *BB);
481 void initTailDupThreshold();
482
483 /// Decrease the UnscheduledPredecessors count for all blocks in chain, and
484 /// if the count goes to 0, add them to the appropriate work list.
485 void markChainSuccessors(const BlockChain &Chain,
486 const MachineBasicBlock *LoopHeaderBB,
487 const BlockFilterSet *BlockFilter = nullptr);
488
489 /// Decrease the UnscheduledPredecessors count for a single block, and
490 /// if the count goes to 0, add them to the appropriate work list.
491 void markBlockSuccessors(const BlockChain &Chain, const MachineBasicBlock *BB,
492 const MachineBasicBlock *LoopHeaderBB,
493 const BlockFilterSet *BlockFilter = nullptr);
494
495 BranchProbability
496 collectViableSuccessors(const MachineBasicBlock *BB, const BlockChain &Chain,
497 const BlockFilterSet *BlockFilter,
498 SmallVector<MachineBasicBlock *, 4> &Successors);
499 bool isBestSuccessor(MachineBasicBlock *BB, MachineBasicBlock *Pred,
500 BlockFilterSet *BlockFilter);
501 void findDuplicateCandidates(SmallVectorImpl<MachineBasicBlock *> &Candidates,
502 MachineBasicBlock *BB,
503 BlockFilterSet *BlockFilter);
504 bool repeatedlyTailDuplicateBlock(
505 MachineBasicBlock *BB, MachineBasicBlock *&LPred,
506 const MachineBasicBlock *LoopHeaderBB, BlockChain &Chain,
507 BlockFilterSet *BlockFilter,
508 MachineFunction::iterator &PrevUnplacedBlockIt,
509 BlockFilterSet::iterator &PrevUnplacedBlockInFilterIt);
510 bool
511 maybeTailDuplicateBlock(MachineBasicBlock *BB, MachineBasicBlock *LPred,
512 BlockChain &Chain, BlockFilterSet *BlockFilter,
513 MachineFunction::iterator &PrevUnplacedBlockIt,
514 BlockFilterSet::iterator &PrevUnplacedBlockInFilterIt,
515 bool &DuplicatedToLPred);
516 bool hasBetterLayoutPredecessor(const MachineBasicBlock *BB,
517 const MachineBasicBlock *Succ,
518 const BlockChain &SuccChain,
519 BranchProbability SuccProb,
520 BranchProbability RealSuccProb,
521 const BlockChain &Chain,
522 const BlockFilterSet *BlockFilter);
523 BlockAndTailDupResult selectBestSuccessor(const MachineBasicBlock *BB,
524 const BlockChain &Chain,
525 const BlockFilterSet *BlockFilter);
526 MachineBasicBlock *
527 selectBestCandidateBlock(const BlockChain &Chain,
528 SmallVectorImpl<MachineBasicBlock *> &WorkList);
529 MachineBasicBlock *
530 getFirstUnplacedBlock(const BlockChain &PlacedChain,
531 MachineFunction::iterator &PrevUnplacedBlockIt);
532 MachineBasicBlock *
533 getFirstUnplacedBlock(const BlockChain &PlacedChain,
534 BlockFilterSet::iterator &PrevUnplacedBlockInFilterIt,
535 const BlockFilterSet *BlockFilter);
536
537 /// Add a basic block to the work list if it is appropriate.
538 ///
539 /// If the optional parameter BlockFilter is provided, only MBB
540 /// present in the set will be added to the worklist. If nullptr
541 /// is provided, no filtering occurs.
542 void fillWorkLists(const MachineBasicBlock *MBB,
543 SmallPtrSetImpl<BlockChain *> &UpdatedPreds,
544 const BlockFilterSet *BlockFilter);
545
546 void buildChain(const MachineBasicBlock *BB, BlockChain &Chain,
547 BlockFilterSet *BlockFilter = nullptr);
548 bool canMoveBottomBlockToTop(const MachineBasicBlock *BottomBlock,
549 const MachineBasicBlock *OldTop);
550 bool hasViableTopFallthrough(const MachineBasicBlock *Top,
551 const BlockFilterSet &LoopBlockSet);
552 BlockFrequency TopFallThroughFreq(const MachineBasicBlock *Top,
553 const BlockFilterSet &LoopBlockSet);
554 BlockFrequency FallThroughGains(const MachineBasicBlock *NewTop,
555 const MachineBasicBlock *OldTop,
556 const MachineBasicBlock *ExitBB,
557 const BlockFilterSet &LoopBlockSet);
558 MachineBasicBlock *findBestLoopTopHelper(MachineBasicBlock *OldTop,
559 const MachineLoop &L,
560 const BlockFilterSet &LoopBlockSet);
561 MachineBasicBlock *findBestLoopTop(const MachineLoop &L,
562 const BlockFilterSet &LoopBlockSet);
563 MachineBasicBlock *findBestLoopExit(const MachineLoop &L,
564 const BlockFilterSet &LoopBlockSet,
565 BlockFrequency &ExitFreq);
566 BlockFilterSet collectLoopBlockSet(const MachineLoop &L);
567 void buildLoopChains(const MachineLoop &L);
568 void rotateLoop(BlockChain &LoopChain, const MachineBasicBlock *ExitingBB,
569 BlockFrequency ExitFreq, const BlockFilterSet &LoopBlockSet);
570 void rotateLoopWithProfile(BlockChain &LoopChain, const MachineLoop &L,
571 const BlockFilterSet &LoopBlockSet);
572 void buildCFGChains();
573 void optimizeBranches();
574 void alignBlocks();
575 /// Returns true if a block should be tail-duplicated to increase fallthrough
576 /// opportunities.
577 bool shouldTailDuplicate(MachineBasicBlock *BB);
578 /// Check the edge frequencies to see if tail duplication will increase
579 /// fallthroughs.
580 bool isProfitableToTailDup(const MachineBasicBlock *BB,
581 const MachineBasicBlock *Succ,
582 BranchProbability QProb, const BlockChain &Chain,
583 const BlockFilterSet *BlockFilter);
584
585 /// Check for a trellis layout.
586 bool isTrellis(const MachineBasicBlock *BB,
587 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
588 const BlockChain &Chain, const BlockFilterSet *BlockFilter);
589
590 /// Get the best successor given a trellis layout.
591 BlockAndTailDupResult getBestTrellisSuccessor(
592 const MachineBasicBlock *BB,
593 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
594 BranchProbability AdjustedSumProb, const BlockChain &Chain,
595 const BlockFilterSet *BlockFilter);
596
597 /// Get the best pair of non-conflicting edges.
598 static std::pair<WeightedEdge, WeightedEdge> getBestNonConflictingEdges(
599 const MachineBasicBlock *BB,
600 MutableArrayRef<SmallVector<WeightedEdge, 8>> Edges);
601
602 /// Returns true if a block can tail duplicate into all unplaced
603 /// predecessors. Filters based on loop.
604 bool canTailDuplicateUnplacedPreds(const MachineBasicBlock *BB,
605 MachineBasicBlock *Succ,
606 const BlockChain &Chain,
607 const BlockFilterSet *BlockFilter);
608
609 /// Find chains of triangles to tail-duplicate where a global analysis works,
610 /// but a local analysis would not find them.
611 void precomputeTriangleChains();
612
613 /// Apply a post-processing step optimizing block placement.
614 void applyExtTsp(bool OptForSize);
615
616 /// Modify the existing block placement in the function and adjust all jumps.
617 void assignBlockOrder(const std::vector<const MachineBasicBlock *> &NewOrder);
618
619 /// Create a single CFG chain from the current block order.
620 void createCFGChainExtTsp();
621
622public:
623 MachineBlockPlacement(const MachineBranchProbabilityInfo *MBPI,
624 MachineLoopInfo *MLI, ProfileSummaryInfo *PSI,
625 std::unique_ptr<MBFIWrapper> MBFI,
626 MachinePostDominatorTree *MPDT, bool AllowTailMerge)
627 : MBPI(MBPI), MBFI(std::move(MBFI)), MLI(MLI), MPDT(MPDT), PSI(PSI),
628 AllowTailMerge(AllowTailMerge) {};
629
630 bool run(MachineFunction &F);
631
632 static bool allowTailDupPlacement(MachineFunction &MF) {
633 return TailDupPlacement && !MF.getTarget().requiresStructuredCFG();
634 }
635};
636
637class MachineBlockPlacementLegacy : public MachineFunctionPass {
638public:
639 static char ID; // Pass identification, replacement for typeid
640
641 MachineBlockPlacementLegacy() : MachineFunctionPass(ID) {
642 initializeMachineBlockPlacementLegacyPass(*PassRegistry::getPassRegistry());
643 }
644
645 bool runOnMachineFunction(MachineFunction &MF) override {
646 if (skipFunction(F: MF.getFunction()))
647 return false;
648
649 auto *MBPI =
650 &getAnalysis<MachineBranchProbabilityInfoWrapperPass>().getMBPI();
651 auto MBFI = std::make_unique<MBFIWrapper>(
652 args&: getAnalysis<MachineBlockFrequencyInfoWrapperPass>().getMBFI());
653 auto *MLI = &getAnalysis<MachineLoopInfoWrapperPass>().getLI();
654 auto *MPDT = MachineBlockPlacement::allowTailDupPlacement(MF)
655 ? &getAnalysis<MachinePostDominatorTreeWrapperPass>()
656 .getPostDomTree()
657 : nullptr;
658 auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
659 auto *PassConfig = &getAnalysis<TargetPassConfig>();
660 bool AllowTailMerge = PassConfig->getEnableTailMerge();
661 return MachineBlockPlacement(MBPI, MLI, PSI, std::move(MBFI), MPDT,
662 AllowTailMerge)
663 .run(F&: MF);
664 }
665
666 void getAnalysisUsage(AnalysisUsage &AU) const override {
667 AU.addRequired<MachineBranchProbabilityInfoWrapperPass>();
668 AU.addRequired<MachineBlockFrequencyInfoWrapperPass>();
669 if (TailDupPlacement)
670 AU.addRequired<MachinePostDominatorTreeWrapperPass>();
671 AU.addRequired<MachineLoopInfoWrapperPass>();
672 AU.addRequired<ProfileSummaryInfoWrapperPass>();
673 AU.addRequired<TargetPassConfig>();
674 MachineFunctionPass::getAnalysisUsage(AU);
675 }
676};
677
678} // end anonymous namespace
679
680char MachineBlockPlacementLegacy::ID = 0;
681
682char &llvm::MachineBlockPlacementID = MachineBlockPlacementLegacy::ID;
683
684INITIALIZE_PASS_BEGIN(MachineBlockPlacementLegacy, DEBUG_TYPE,
685 "Branch Probability Basic Block Placement", false, false)
686INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfoWrapperPass)
687INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfoWrapperPass)
688INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTreeWrapperPass)
689INITIALIZE_PASS_DEPENDENCY(MachineLoopInfoWrapperPass)
690INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
691INITIALIZE_PASS_END(MachineBlockPlacementLegacy, DEBUG_TYPE,
692 "Branch Probability Basic Block Placement", false, false)
693
694#ifndef NDEBUG
695/// Helper to print the name of a MBB.
696///
697/// Only used by debug logging.
698static std::string getBlockName(const MachineBasicBlock *BB) {
699 std::string Result;
700 raw_string_ostream OS(Result);
701 OS << printMBBReference(*BB);
702 OS << " ('" << BB->getName() << "')";
703 OS.flush();
704 return Result;
705}
706#endif
707
708/// Mark a chain's successors as having one fewer preds.
709///
710/// When a chain is being merged into the "placed" chain, this routine will
711/// quickly walk the successors of each block in the chain and mark them as
712/// having one fewer active predecessor. It also adds any successors of this
713/// chain which reach the zero-predecessor state to the appropriate worklist.
714void MachineBlockPlacement::markChainSuccessors(
715 const BlockChain &Chain, const MachineBasicBlock *LoopHeaderBB,
716 const BlockFilterSet *BlockFilter) {
717 // Walk all the blocks in this chain, marking their successors as having
718 // a predecessor placed.
719 for (MachineBasicBlock *MBB : Chain) {
720 markBlockSuccessors(Chain, BB: MBB, LoopHeaderBB, BlockFilter);
721 }
722}
723
724/// Mark a single block's successors as having one fewer preds.
725///
726/// Under normal circumstances, this is only called by markChainSuccessors,
727/// but if a block that was to be placed is completely tail-duplicated away,
728/// and was duplicated into the chain end, we need to redo markBlockSuccessors
729/// for just that block.
730void MachineBlockPlacement::markBlockSuccessors(
731 const BlockChain &Chain, const MachineBasicBlock *MBB,
732 const MachineBasicBlock *LoopHeaderBB, const BlockFilterSet *BlockFilter) {
733 // Add any successors for which this is the only un-placed in-loop
734 // predecessor to the worklist as a viable candidate for CFG-neutral
735 // placement. No subsequent placement of this block will violate the CFG
736 // shape, so we get to use heuristics to choose a favorable placement.
737 for (MachineBasicBlock *Succ : MBB->successors()) {
738 if (BlockFilter && !BlockFilter->count(key: Succ))
739 continue;
740 BlockChain &SuccChain = *BlockToChain[Succ];
741 // Disregard edges within a fixed chain, or edges to the loop header.
742 if (&Chain == &SuccChain || Succ == LoopHeaderBB)
743 continue;
744
745 // This is a cross-chain edge that is within the loop, so decrement the
746 // loop predecessor count of the destination chain.
747 if (SuccChain.UnscheduledPredecessors == 0 ||
748 --SuccChain.UnscheduledPredecessors > 0)
749 continue;
750
751 auto *NewBB = *SuccChain.begin();
752 if (NewBB->isEHPad())
753 EHPadWorkList.push_back(Elt: NewBB);
754 else
755 BlockWorkList.push_back(Elt: NewBB);
756 }
757}
758
759/// This helper function collects the set of successors of block
760/// \p BB that are allowed to be its layout successors, and return
761/// the total branch probability of edges from \p BB to those
762/// blocks.
763BranchProbability MachineBlockPlacement::collectViableSuccessors(
764 const MachineBasicBlock *BB, const BlockChain &Chain,
765 const BlockFilterSet *BlockFilter,
766 SmallVector<MachineBasicBlock *, 4> &Successors) {
767 // Adjust edge probabilities by excluding edges pointing to blocks that is
768 // either not in BlockFilter or is already in the current chain. Consider the
769 // following CFG:
770 //
771 // --->A
772 // | / \
773 // | B C
774 // | \ / \
775 // ----D E
776 //
777 // Assume A->C is very hot (>90%), and C->D has a 50% probability, then after
778 // A->C is chosen as a fall-through, D won't be selected as a successor of C
779 // due to CFG constraint (the probability of C->D is not greater than
780 // HotProb to break topo-order). If we exclude E that is not in BlockFilter
781 // when calculating the probability of C->D, D will be selected and we
782 // will get A C D B as the layout of this loop.
783 auto AdjustedSumProb = BranchProbability::getOne();
784 for (MachineBasicBlock *Succ : BB->successors()) {
785 bool SkipSucc = false;
786 if (Succ->isEHPad() || (BlockFilter && !BlockFilter->count(key: Succ))) {
787 SkipSucc = true;
788 } else {
789 BlockChain *SuccChain = BlockToChain[Succ];
790 if (SuccChain == &Chain) {
791 SkipSucc = true;
792 } else if (Succ != *SuccChain->begin()) {
793 LLVM_DEBUG(dbgs() << " " << getBlockName(Succ)
794 << " -> Mid chain!\n");
795 continue;
796 }
797 }
798 if (SkipSucc)
799 AdjustedSumProb -= MBPI->getEdgeProbability(Src: BB, Dst: Succ);
800 else
801 Successors.push_back(Elt: Succ);
802 }
803
804 return AdjustedSumProb;
805}
806
807/// The helper function returns the branch probability that is adjusted
808/// or normalized over the new total \p AdjustedSumProb.
809static BranchProbability
810getAdjustedProbability(BranchProbability OrigProb,
811 BranchProbability AdjustedSumProb) {
812 BranchProbability SuccProb;
813 uint32_t SuccProbN = OrigProb.getNumerator();
814 uint32_t SuccProbD = AdjustedSumProb.getNumerator();
815 if (SuccProbN >= SuccProbD)
816 SuccProb = BranchProbability::getOne();
817 else
818 SuccProb = BranchProbability(SuccProbN, SuccProbD);
819
820 return SuccProb;
821}
822
823/// Check if \p BB has exactly the successors in \p Successors.
824static bool
825hasSameSuccessors(MachineBasicBlock &BB,
826 SmallPtrSetImpl<const MachineBasicBlock *> &Successors) {
827 if (BB.succ_size() != Successors.size())
828 return false;
829 // We don't want to count self-loops
830 if (Successors.count(Ptr: &BB))
831 return false;
832 for (MachineBasicBlock *Succ : BB.successors())
833 if (!Successors.count(Ptr: Succ))
834 return false;
835 return true;
836}
837
838/// Check if a block should be tail duplicated to increase fallthrough
839/// opportunities.
840/// \p BB Block to check.
841bool MachineBlockPlacement::shouldTailDuplicate(MachineBasicBlock *BB) {
842 // Blocks with single successors don't create additional fallthrough
843 // opportunities. Don't duplicate them. TODO: When conditional exits are
844 // analyzable, allow them to be duplicated.
845 bool IsSimple = TailDup.isSimpleBB(TailBB: BB);
846
847 if (BB->succ_size() == 1)
848 return false;
849 return TailDup.shouldTailDuplicate(IsSimple, TailBB&: *BB);
850}
851
852/// Compare 2 BlockFrequency's with a small penalty for \p A.
853/// In order to be conservative, we apply a X% penalty to account for
854/// increased icache pressure and static heuristics. For small frequencies
855/// we use only the numerators to improve accuracy. For simplicity, we assume
856/// the penalty is less than 100%
857/// TODO(iteratee): Use 64-bit fixed point edge frequencies everywhere.
858static bool greaterWithBias(BlockFrequency A, BlockFrequency B,
859 BlockFrequency EntryFreq) {
860 BranchProbability ThresholdProb(TailDupPlacementPenalty, 100);
861 BlockFrequency Gain = A - B;
862 return (Gain / ThresholdProb) >= EntryFreq;
863}
864
865/// Check the edge frequencies to see if tail duplication will increase
866/// fallthroughs. It only makes sense to call this function when
867/// \p Succ would not be chosen otherwise. Tail duplication of \p Succ is
868/// always locally profitable if we would have picked \p Succ without
869/// considering duplication.
870bool MachineBlockPlacement::isProfitableToTailDup(
871 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
872 BranchProbability QProb, const BlockChain &Chain,
873 const BlockFilterSet *BlockFilter) {
874 // We need to do a probability calculation to make sure this is profitable.
875 // First: does succ have a successor that post-dominates? This affects the
876 // calculation. The 2 relevant cases are:
877 // BB BB
878 // | \Qout | \Qout
879 // P| C |P C
880 // = C' = C'
881 // | /Qin | /Qin
882 // | / | /
883 // Succ Succ
884 // / \ | \ V
885 // U/ =V |U \
886 // / \ = D
887 // D E | /
888 // | /
889 // |/
890 // PDom
891 // '=' : Branch taken for that CFG edge
892 // In the second case, Placing Succ while duplicating it into C prevents the
893 // fallthrough of Succ into either D or PDom, because they now have C as an
894 // unplaced predecessor
895
896 // Start by figuring out which case we fall into
897 MachineBasicBlock *PDom = nullptr;
898 SmallVector<MachineBasicBlock *, 4> SuccSuccs;
899 // Only scan the relevant successors
900 auto AdjustedSuccSumProb =
901 collectViableSuccessors(BB: Succ, Chain, BlockFilter, Successors&: SuccSuccs);
902 BranchProbability PProb = MBPI->getEdgeProbability(Src: BB, Dst: Succ);
903 auto BBFreq = MBFI->getBlockFreq(MBB: BB);
904 auto SuccFreq = MBFI->getBlockFreq(MBB: Succ);
905 BlockFrequency P = BBFreq * PProb;
906 BlockFrequency Qout = BBFreq * QProb;
907 BlockFrequency EntryFreq = MBFI->getEntryFreq();
908 // If there are no more successors, it is profitable to copy, as it strictly
909 // increases fallthrough.
910 if (SuccSuccs.size() == 0)
911 return greaterWithBias(A: P, B: Qout, EntryFreq);
912
913 auto BestSuccSucc = BranchProbability::getZero();
914 // Find the PDom or the best Succ if no PDom exists.
915 for (MachineBasicBlock *SuccSucc : SuccSuccs) {
916 auto Prob = MBPI->getEdgeProbability(Src: Succ, Dst: SuccSucc);
917 if (Prob > BestSuccSucc)
918 BestSuccSucc = Prob;
919 if (PDom == nullptr)
920 if (MPDT->dominates(A: SuccSucc, B: Succ)) {
921 PDom = SuccSucc;
922 break;
923 }
924 }
925 // For the comparisons, we need to know Succ's best incoming edge that isn't
926 // from BB.
927 auto SuccBestPred = BlockFrequency(0);
928 for (MachineBasicBlock *SuccPred : Succ->predecessors()) {
929 if (SuccPred == Succ || SuccPred == BB ||
930 BlockToChain[SuccPred] == &Chain ||
931 (BlockFilter && !BlockFilter->count(key: SuccPred)))
932 continue;
933 auto Freq =
934 MBFI->getBlockFreq(MBB: SuccPred) * MBPI->getEdgeProbability(Src: SuccPred, Dst: Succ);
935 if (Freq > SuccBestPred)
936 SuccBestPred = Freq;
937 }
938 // Qin is Succ's best unplaced incoming edge that isn't BB
939 BlockFrequency Qin = SuccBestPred;
940 // If it doesn't have a post-dominating successor, here is the calculation:
941 // BB BB
942 // | \Qout | \
943 // P| C | =
944 // = C' | C
945 // | /Qin | |
946 // | / | C' (+Succ)
947 // Succ Succ /|
948 // / \ | \/ |
949 // U/ =V | == |
950 // / \ | / \|
951 // D E D E
952 // '=' : Branch taken for that CFG edge
953 // Cost in the first case is: P + V
954 // For this calculation, we always assume P > Qout. If Qout > P
955 // The result of this function will be ignored at the caller.
956 // Let F = SuccFreq - Qin
957 // Cost in the second case is: Qout + min(Qin, F) * U + max(Qin, F) * V
958
959 if (PDom == nullptr || !Succ->isSuccessor(MBB: PDom)) {
960 BranchProbability UProb = BestSuccSucc;
961 BranchProbability VProb = AdjustedSuccSumProb - UProb;
962 BlockFrequency F = SuccFreq - Qin;
963 BlockFrequency V = SuccFreq * VProb;
964 BlockFrequency QinU = std::min(a: Qin, b: F) * UProb;
965 BlockFrequency BaseCost = P + V;
966 BlockFrequency DupCost = Qout + QinU + std::max(a: Qin, b: F) * VProb;
967 return greaterWithBias(A: BaseCost, B: DupCost, EntryFreq);
968 }
969 BranchProbability UProb = MBPI->getEdgeProbability(Src: Succ, Dst: PDom);
970 BranchProbability VProb = AdjustedSuccSumProb - UProb;
971 BlockFrequency U = SuccFreq * UProb;
972 BlockFrequency V = SuccFreq * VProb;
973 BlockFrequency F = SuccFreq - Qin;
974 // If there is a post-dominating successor, here is the calculation:
975 // BB BB BB BB
976 // | \Qout | \ | \Qout | \
977 // |P C | = |P C | =
978 // = C' |P C = C' |P C
979 // | /Qin | | | /Qin | |
980 // | / | C' (+Succ) | / | C' (+Succ)
981 // Succ Succ /| Succ Succ /|
982 // | \ V | \/ | | \ V | \/ |
983 // |U \ |U /\ =? |U = |U /\ |
984 // = D = = =?| | D | = =|
985 // | / |/ D | / |/ D
986 // | / | / | = | /
987 // |/ | / |/ | =
988 // Dom Dom Dom Dom
989 // '=' : Branch taken for that CFG edge
990 // The cost for taken branches in the first case is P + U
991 // Let F = SuccFreq - Qin
992 // The cost in the second case (assuming independence), given the layout:
993 // BB, Succ, (C+Succ), D, Dom or the layout:
994 // BB, Succ, D, Dom, (C+Succ)
995 // is Qout + max(F, Qin) * U + min(F, Qin)
996 // compare P + U vs Qout + P * U + Qin.
997 //
998 // The 3rd and 4th cases cover when Dom would be chosen to follow Succ.
999 //
1000 // For the 3rd case, the cost is P + 2 * V
1001 // For the 4th case, the cost is Qout + min(Qin, F) * U + max(Qin, F) * V + V
1002 // We choose 4 over 3 when (P + V) > Qout + min(Qin, F) * U + max(Qin, F) * V
1003 if (UProb > AdjustedSuccSumProb / 2 &&
1004 !hasBetterLayoutPredecessor(BB: Succ, Succ: PDom, SuccChain: *BlockToChain[PDom], SuccProb: UProb, RealSuccProb: UProb,
1005 Chain, BlockFilter))
1006 // Cases 3 & 4
1007 return greaterWithBias(
1008 A: (P + V), B: (Qout + std::max(a: Qin, b: F) * VProb + std::min(a: Qin, b: F) * UProb),
1009 EntryFreq);
1010 // Cases 1 & 2
1011 return greaterWithBias(A: (P + U),
1012 B: (Qout + std::min(a: Qin, b: F) * AdjustedSuccSumProb +
1013 std::max(a: Qin, b: F) * UProb),
1014 EntryFreq);
1015}
1016
1017/// Check for a trellis layout. \p BB is the upper part of a trellis if its
1018/// successors form the lower part of a trellis. A successor set S forms the
1019/// lower part of a trellis if all of the predecessors of S are either in S or
1020/// have all of S as successors. We ignore trellises where BB doesn't have 2
1021/// successors because for fewer than 2, it's trivial, and for 3 or greater they
1022/// are very uncommon and complex to compute optimally. Allowing edges within S
1023/// is not strictly a trellis, but the same algorithm works, so we allow it.
1024bool MachineBlockPlacement::isTrellis(
1025 const MachineBasicBlock *BB,
1026 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
1027 const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
1028 // Technically BB could form a trellis with branching factor higher than 2.
1029 // But that's extremely uncommon.
1030 if (BB->succ_size() != 2 || ViableSuccs.size() != 2)
1031 return false;
1032
1033 SmallPtrSet<const MachineBasicBlock *, 2> Successors(llvm::from_range,
1034 BB->successors());
1035 // To avoid reviewing the same predecessors twice.
1036 SmallPtrSet<const MachineBasicBlock *, 8> SeenPreds;
1037
1038 for (MachineBasicBlock *Succ : ViableSuccs) {
1039 // Compile-time optimization: runtime is quadratic in the number of
1040 // predecessors. For such uncommon cases, exit early.
1041 if (Succ->pred_size() > PredecessorLimit)
1042 return false;
1043
1044 int PredCount = 0;
1045 for (auto *SuccPred : Succ->predecessors()) {
1046 // Allow triangle successors, but don't count them.
1047 if (Successors.count(Ptr: SuccPred)) {
1048 // Make sure that it is actually a triangle.
1049 for (MachineBasicBlock *CheckSucc : SuccPred->successors())
1050 if (!Successors.count(Ptr: CheckSucc))
1051 return false;
1052 continue;
1053 }
1054 const BlockChain *PredChain = BlockToChain[SuccPred];
1055 if (SuccPred == BB || (BlockFilter && !BlockFilter->count(key: SuccPred)) ||
1056 PredChain == &Chain || PredChain == BlockToChain[Succ])
1057 continue;
1058 ++PredCount;
1059 // Perform the successor check only once.
1060 if (!SeenPreds.insert(Ptr: SuccPred).second)
1061 continue;
1062 if (!hasSameSuccessors(BB&: *SuccPred, Successors))
1063 return false;
1064 }
1065 // If one of the successors has only BB as a predecessor, it is not a
1066 // trellis.
1067 if (PredCount < 1)
1068 return false;
1069 }
1070 return true;
1071}
1072
1073/// Pick the highest total weight pair of edges that can both be laid out.
1074/// The edges in \p Edges[0] are assumed to have a different destination than
1075/// the edges in \p Edges[1]. Simple counting shows that the best pair is either
1076/// the individual highest weight edges to the 2 different destinations, or in
1077/// case of a conflict, one of them should be replaced with a 2nd best edge.
1078std::pair<MachineBlockPlacement::WeightedEdge,
1079 MachineBlockPlacement::WeightedEdge>
1080MachineBlockPlacement::getBestNonConflictingEdges(
1081 const MachineBasicBlock *BB,
1082 MutableArrayRef<SmallVector<MachineBlockPlacement::WeightedEdge, 8>>
1083 Edges) {
1084 // Sort the edges, and then for each successor, find the best incoming
1085 // predecessor. If the best incoming predecessors aren't the same,
1086 // then that is clearly the best layout. If there is a conflict, one of the
1087 // successors will have to fallthrough from the second best predecessor. We
1088 // compare which combination is better overall.
1089
1090 // Sort for highest frequency.
1091 auto Cmp = [](WeightedEdge A, WeightedEdge B) { return A.Weight > B.Weight; };
1092
1093 llvm::stable_sort(Range&: Edges[0], C: Cmp);
1094 llvm::stable_sort(Range&: Edges[1], C: Cmp);
1095 auto BestA = Edges[0].begin();
1096 auto BestB = Edges[1].begin();
1097 // Arrange for the correct answer to be in BestA and BestB
1098 // If the 2 best edges don't conflict, the answer is already there.
1099 if (BestA->Src == BestB->Src) {
1100 // Compare the total fallthrough of (Best + Second Best) for both pairs
1101 auto SecondBestA = std::next(x: BestA);
1102 auto SecondBestB = std::next(x: BestB);
1103 BlockFrequency BestAScore = BestA->Weight + SecondBestB->Weight;
1104 BlockFrequency BestBScore = BestB->Weight + SecondBestA->Weight;
1105 if (BestAScore < BestBScore)
1106 BestA = SecondBestA;
1107 else
1108 BestB = SecondBestB;
1109 }
1110 // Arrange for the BB edge to be in BestA if it exists.
1111 if (BestB->Src == BB)
1112 std::swap(a&: BestA, b&: BestB);
1113 return std::make_pair(x&: *BestA, y&: *BestB);
1114}
1115
1116/// Get the best successor from \p BB based on \p BB being part of a trellis.
1117/// We only handle trellises with 2 successors, so the algorithm is
1118/// straightforward: Find the best pair of edges that don't conflict. We find
1119/// the best incoming edge for each successor in the trellis. If those conflict,
1120/// we consider which of them should be replaced with the second best.
1121/// Upon return the two best edges will be in \p BestEdges. If one of the edges
1122/// comes from \p BB, it will be in \p BestEdges[0]
1123MachineBlockPlacement::BlockAndTailDupResult
1124MachineBlockPlacement::getBestTrellisSuccessor(
1125 const MachineBasicBlock *BB,
1126 const SmallVectorImpl<MachineBasicBlock *> &ViableSuccs,
1127 BranchProbability AdjustedSumProb, const BlockChain &Chain,
1128 const BlockFilterSet *BlockFilter) {
1129
1130 BlockAndTailDupResult Result = {.BB: nullptr, .ShouldTailDup: false};
1131 SmallPtrSet<const MachineBasicBlock *, 4> Successors(llvm::from_range,
1132 BB->successors());
1133
1134 // We assume size 2 because it's common. For general n, we would have to do
1135 // the Hungarian algorithm, but it's not worth the complexity because more
1136 // than 2 successors is fairly uncommon, and a trellis even more so.
1137 if (Successors.size() != 2 || ViableSuccs.size() != 2)
1138 return Result;
1139
1140 // Collect the edge frequencies of all edges that form the trellis.
1141 SmallVector<WeightedEdge, 8> Edges[2];
1142 int SuccIndex = 0;
1143 for (auto *Succ : ViableSuccs) {
1144 for (MachineBasicBlock *SuccPred : Succ->predecessors()) {
1145 // Skip any placed predecessors that are not BB
1146 if (SuccPred != BB) {
1147 if (BlockFilter && !BlockFilter->count(key: SuccPred))
1148 continue;
1149 const BlockChain *SuccPredChain = BlockToChain[SuccPred];
1150 if (SuccPredChain == &Chain || SuccPredChain == BlockToChain[Succ])
1151 continue;
1152 }
1153 BlockFrequency EdgeFreq = MBFI->getBlockFreq(MBB: SuccPred) *
1154 MBPI->getEdgeProbability(Src: SuccPred, Dst: Succ);
1155 Edges[SuccIndex].push_back(Elt: {.Weight: EdgeFreq, .Src: SuccPred, .Dest: Succ});
1156 }
1157 ++SuccIndex;
1158 }
1159
1160 // Pick the best combination of 2 edges from all the edges in the trellis.
1161 WeightedEdge BestA, BestB;
1162 std::tie(args&: BestA, args&: BestB) = getBestNonConflictingEdges(BB, Edges);
1163
1164 if (BestA.Src != BB) {
1165 // If we have a trellis, and BB doesn't have the best fallthrough edges,
1166 // we shouldn't choose any successor. We've already looked and there's a
1167 // better fallthrough edge for all the successors.
1168 LLVM_DEBUG(dbgs() << "Trellis, but not one of the chosen edges.\n");
1169 return Result;
1170 }
1171
1172 // Did we pick the triangle edge? If tail-duplication is profitable, do
1173 // that instead. Otherwise merge the triangle edge now while we know it is
1174 // optimal.
1175 if (BestA.Dest == BestB.Src) {
1176 // The edges are BB->Succ1->Succ2, and we're looking to see if BB->Succ2
1177 // would be better.
1178 MachineBasicBlock *Succ1 = BestA.Dest;
1179 MachineBasicBlock *Succ2 = BestB.Dest;
1180 // Check to see if tail-duplication would be profitable.
1181 if (allowTailDupPlacement(MF&: *F) && shouldTailDuplicate(BB: Succ2) &&
1182 canTailDuplicateUnplacedPreds(BB, Succ: Succ2, Chain, BlockFilter) &&
1183 isProfitableToTailDup(BB, Succ: Succ2, QProb: MBPI->getEdgeProbability(Src: BB, Dst: Succ1),
1184 Chain, BlockFilter)) {
1185 LLVM_DEBUG(BranchProbability Succ2Prob = getAdjustedProbability(
1186 MBPI->getEdgeProbability(BB, Succ2), AdjustedSumProb);
1187 dbgs() << " Selected: " << getBlockName(Succ2)
1188 << ", probability: " << Succ2Prob
1189 << " (Tail Duplicate)\n");
1190 Result.BB = Succ2;
1191 Result.ShouldTailDup = true;
1192 return Result;
1193 }
1194 }
1195 // We have already computed the optimal edge for the other side of the
1196 // trellis.
1197 ComputedEdges[BestB.Src] = {.BB: BestB.Dest, .ShouldTailDup: false};
1198
1199 auto TrellisSucc = BestA.Dest;
1200 LLVM_DEBUG(BranchProbability SuccProb = getAdjustedProbability(
1201 MBPI->getEdgeProbability(BB, TrellisSucc), AdjustedSumProb);
1202 dbgs() << " Selected: " << getBlockName(TrellisSucc)
1203 << ", probability: " << SuccProb << " (Trellis)\n");
1204 Result.BB = TrellisSucc;
1205 return Result;
1206}
1207
1208/// When the option allowTailDupPlacement() is on, this method checks if the
1209/// fallthrough candidate block \p Succ (of block \p BB) can be tail-duplicated
1210/// into all of its unplaced, unfiltered predecessors, that are not BB.
1211bool MachineBlockPlacement::canTailDuplicateUnplacedPreds(
1212 const MachineBasicBlock *BB, MachineBasicBlock *Succ,
1213 const BlockChain &Chain, const BlockFilterSet *BlockFilter) {
1214 if (!shouldTailDuplicate(BB: Succ))
1215 return false;
1216
1217 // The result of canTailDuplicate.
1218 bool Duplicate = true;
1219 // Number of possible duplication.
1220 unsigned int NumDup = 0;
1221
1222 // For CFG checking.
1223 SmallPtrSet<const MachineBasicBlock *, 4> Successors(llvm::from_range,
1224 BB->successors());
1225 for (MachineBasicBlock *Pred : Succ->predecessors()) {
1226 // Make sure all unplaced and unfiltered predecessors can be
1227 // tail-duplicated into.
1228 // Skip any blocks that are already placed or not in this loop.
1229 if (Pred == BB || (BlockFilter && !BlockFilter->count(key: Pred)) ||
1230 (BlockToChain[Pred] == &Chain && !Succ->succ_empty()))
1231 continue;
1232 if (!TailDup.canTailDuplicate(TailBB: Succ, PredBB: Pred)) {
1233 if (Successors.size() > 1 && hasSameSuccessors(BB&: *Pred, Successors))
1234 // This will result in a trellis after tail duplication, so we don't
1235 // need to copy Succ into this predecessor. In the presence
1236 // of a trellis tail duplication can continue to be profitable.
1237 // For example:
1238 // A A
1239 // |\ |\
1240 // | \ | \
1241 // | C | C+BB
1242 // | / | |
1243 // |/ | |
1244 // BB => BB |
1245 // |\ |\/|
1246 // | \ |/\|
1247 // | D | D
1248 // | / | /
1249 // |/ |/
1250 // Succ Succ
1251 //
1252 // After BB was duplicated into C, the layout looks like the one on the
1253 // right. BB and C now have the same successors. When considering
1254 // whether Succ can be duplicated into all its unplaced predecessors, we
1255 // ignore C.
1256 // We can do this because C already has a profitable fallthrough, namely
1257 // D. TODO(iteratee): ignore sufficiently cold predecessors for
1258 // duplication and for this test.
1259 //
1260 // This allows trellises to be laid out in 2 separate chains
1261 // (A,B,Succ,...) and later (C,D,...) This is a reasonable heuristic
1262 // because it allows the creation of 2 fallthrough paths with links
1263 // between them, and we correctly identify the best layout for these
1264 // CFGs. We want to extend trellises that the user created in addition
1265 // to trellises created by tail-duplication, so we just look for the
1266 // CFG.
1267 continue;
1268 Duplicate = false;
1269 continue;
1270 }
1271 NumDup++;
1272 }
1273
1274 // No possible duplication in current filter set.
1275 if (NumDup == 0)
1276 return false;
1277
1278 // If profile information is available, findDuplicateCandidates can do more
1279 // precise benefit analysis.
1280 if (F->getFunction().hasProfileData())
1281 return true;
1282
1283 // This is mainly for function exit BB.
1284 // The integrated tail duplication is really designed for increasing
1285 // fallthrough from predecessors from Succ to its successors. We may need
1286 // other machanism to handle different cases.
1287 if (Succ->succ_empty())
1288 return true;
1289
1290 // Plus the already placed predecessor.
1291 NumDup++;
1292
1293 // If the duplication candidate has more unplaced predecessors than
1294 // successors, the extra duplication can't bring more fallthrough.
1295 //
1296 // Pred1 Pred2 Pred3
1297 // \ | /
1298 // \ | /
1299 // \ | /
1300 // Dup
1301 // / \
1302 // / \
1303 // Succ1 Succ2
1304 //
1305 // In this example Dup has 2 successors and 3 predecessors, duplication of Dup
1306 // can increase the fallthrough from Pred1 to Succ1 and from Pred2 to Succ2,
1307 // but the duplication into Pred3 can't increase fallthrough.
1308 //
1309 // A small number of extra duplication may not hurt too much. We need a better
1310 // heuristic to handle it.
1311 if ((NumDup > Succ->succ_size()) || !Duplicate)
1312 return false;
1313
1314 return true;
1315}
1316
1317/// Find chains of triangles where we believe it would be profitable to
1318/// tail-duplicate them all, but a local analysis would not find them.
1319/// There are 3 ways this can be profitable:
1320/// 1) The post-dominators marked 50% are actually taken 55% (This shrinks with
1321/// longer chains)
1322/// 2) The chains are statically correlated. Branch probabilities have a very
1323/// U-shaped distribution.
1324/// [http://nrs.harvard.edu/urn-3:HUL.InstRepos:24015805]
1325/// If the branches in a chain are likely to be from the same side of the
1326/// distribution as their predecessor, but are independent at runtime, this
1327/// transformation is profitable. (Because the cost of being wrong is a small
1328/// fixed cost, unlike the standard triangle layout where the cost of being
1329/// wrong scales with the # of triangles.)
1330/// 3) The chains are dynamically correlated. If the probability that a previous
1331/// branch was taken positively influences whether the next branch will be
1332/// taken
1333/// We believe that 2 and 3 are common enough to justify the small margin in 1.
1334void MachineBlockPlacement::precomputeTriangleChains() {
1335 struct TriangleChain {
1336 std::vector<MachineBasicBlock *> Edges;
1337
1338 TriangleChain(MachineBasicBlock *src, MachineBasicBlock *dst)
1339 : Edges({src, dst}) {}
1340
1341 void append(MachineBasicBlock *dst) {
1342 assert(getKey()->isSuccessor(dst) &&
1343 "Attempting to append a block that is not a successor.");
1344 Edges.push_back(x: dst);
1345 }
1346
1347 unsigned count() const { return Edges.size() - 1; }
1348
1349 MachineBasicBlock *getKey() const { return Edges.back(); }
1350 };
1351
1352 if (TriangleChainCount == 0)
1353 return;
1354
1355 LLVM_DEBUG(dbgs() << "Pre-computing triangle chains.\n");
1356 // Map from last block to the chain that contains it. This allows us to extend
1357 // chains as we find new triangles.
1358 DenseMap<const MachineBasicBlock *, TriangleChain> TriangleChainMap;
1359 for (MachineBasicBlock &BB : *F) {
1360 // If BB doesn't have 2 successors, it doesn't start a triangle.
1361 if (BB.succ_size() != 2)
1362 continue;
1363 MachineBasicBlock *PDom = nullptr;
1364 for (MachineBasicBlock *Succ : BB.successors()) {
1365 if (!MPDT->dominates(A: Succ, B: &BB))
1366 continue;
1367 PDom = Succ;
1368 break;
1369 }
1370 // If BB doesn't have a post-dominating successor, it doesn't form a
1371 // triangle.
1372 if (PDom == nullptr)
1373 continue;
1374 // If PDom has a hint that it is low probability, skip this triangle.
1375 if (MBPI->getEdgeProbability(Src: &BB, Dst: PDom) < BranchProbability(50, 100))
1376 continue;
1377 // If PDom isn't eligible for duplication, this isn't the kind of triangle
1378 // we're looking for.
1379 if (!shouldTailDuplicate(BB: PDom))
1380 continue;
1381 bool CanTailDuplicate = true;
1382 // If PDom can't tail-duplicate into it's non-BB predecessors, then this
1383 // isn't the kind of triangle we're looking for.
1384 for (MachineBasicBlock *Pred : PDom->predecessors()) {
1385 if (Pred == &BB)
1386 continue;
1387 if (!TailDup.canTailDuplicate(TailBB: PDom, PredBB: Pred)) {
1388 CanTailDuplicate = false;
1389 break;
1390 }
1391 }
1392 // If we can't tail-duplicate PDom to its predecessors, then skip this
1393 // triangle.
1394 if (!CanTailDuplicate)
1395 continue;
1396
1397 // Now we have an interesting triangle. Insert it if it's not part of an
1398 // existing chain.
1399 // Note: This cannot be replaced with a call insert() or emplace() because
1400 // the find key is BB, but the insert/emplace key is PDom.
1401 auto Found = TriangleChainMap.find(Val: &BB);
1402 // If it is, remove the chain from the map, grow it, and put it back in the
1403 // map with the end as the new key.
1404 if (Found != TriangleChainMap.end()) {
1405 TriangleChain Chain = std::move(Found->second);
1406 TriangleChainMap.erase(I: Found);
1407 Chain.append(dst: PDom);
1408 TriangleChainMap.insert(KV: std::make_pair(x: Chain.getKey(), y: std::move(Chain)));
1409 } else {
1410 auto InsertResult = TriangleChainMap.try_emplace(Key: PDom, Args: &BB, Args&: PDom);
1411 assert(InsertResult.second && "Block seen twice.");
1412 (void)InsertResult;
1413 }
1414 }
1415
1416 // Iterating over a DenseMap is safe here, because the only thing in the body
1417 // of the loop is inserting into another DenseMap (ComputedEdges).
1418 // ComputedEdges is never iterated, so this doesn't lead to non-determinism.
1419 for (auto &ChainPair : TriangleChainMap) {
1420 TriangleChain &Chain = ChainPair.second;
1421 // Benchmarking has shown that due to branch correlation duplicating 2 or
1422 // more triangles is profitable, despite the calculations assuming
1423 // independence.
1424 if (Chain.count() < TriangleChainCount)
1425 continue;
1426 MachineBasicBlock *dst = Chain.Edges.back();
1427 Chain.Edges.pop_back();
1428 for (MachineBasicBlock *src : reverse(C&: Chain.Edges)) {
1429 LLVM_DEBUG(dbgs() << "Marking edge: " << getBlockName(src) << "->"
1430 << getBlockName(dst)
1431 << " as pre-computed based on triangles.\n");
1432
1433 auto InsertResult = ComputedEdges.insert(KV: {src, {.BB: dst, .ShouldTailDup: true}});
1434 assert(InsertResult.second && "Block seen twice.");
1435 (void)InsertResult;
1436
1437 dst = src;
1438 }
1439 }
1440}
1441
1442// When profile is not present, return the StaticLikelyProb.
1443// When profile is available, we need to handle the triangle-shape CFG.
1444static BranchProbability
1445getLayoutSuccessorProbThreshold(const MachineBasicBlock *BB) {
1446 if (!BB->getParent()->getFunction().hasProfileData())
1447 return BranchProbability(StaticLikelyProb, 100);
1448 if (BB->succ_size() == 2) {
1449 const MachineBasicBlock *Succ1 = *BB->succ_begin();
1450 const MachineBasicBlock *Succ2 = *(BB->succ_begin() + 1);
1451 if (Succ1->isSuccessor(MBB: Succ2) || Succ2->isSuccessor(MBB: Succ1)) {
1452 /* See case 1 below for the cost analysis. For BB->Succ to
1453 * be taken with smaller cost, the following needs to hold:
1454 * Prob(BB->Succ) > 2 * Prob(BB->Pred)
1455 * So the threshold T in the calculation below
1456 * (1-T) * Prob(BB->Succ) > T * Prob(BB->Pred)
1457 * So T / (1 - T) = 2, Yielding T = 2/3
1458 * Also adding user specified branch bias, we have
1459 * T = (2/3)*(ProfileLikelyProb/50)
1460 * = (2*ProfileLikelyProb)/150)
1461 */
1462 return BranchProbability(2 * ProfileLikelyProb, 150);
1463 }
1464 }
1465 return BranchProbability(ProfileLikelyProb, 100);
1466}
1467
1468/// Checks to see if the layout candidate block \p Succ has a better layout
1469/// predecessor than \c BB. If yes, returns true.
1470/// \p SuccProb: The probability adjusted for only remaining blocks.
1471/// Only used for logging
1472/// \p RealSuccProb: The un-adjusted probability.
1473/// \p Chain: The chain that BB belongs to and Succ is being considered for.
1474/// \p BlockFilter: if non-null, the set of blocks that make up the loop being
1475/// considered
1476bool MachineBlockPlacement::hasBetterLayoutPredecessor(
1477 const MachineBasicBlock *BB, const MachineBasicBlock *Succ,
1478 const BlockChain &SuccChain, BranchProbability SuccProb,
1479 BranchProbability RealSuccProb, const BlockChain &Chain,
1480 const BlockFilterSet *BlockFilter) {
1481
1482 // There isn't a better layout when there are no unscheduled predecessors.
1483 if (SuccChain.UnscheduledPredecessors == 0)
1484 return false;
1485
1486 // Compile-time optimization: runtime is quadratic in the number of
1487 // predecessors. For such uncommon cases, exit early.
1488 if (Succ->pred_size() > PredecessorLimit)
1489 return false;
1490
1491 // There are two basic scenarios here:
1492 // -------------------------------------
1493 // Case 1: triangular shape CFG (if-then):
1494 // BB
1495 // | \
1496 // | \
1497 // | Pred
1498 // | /
1499 // Succ
1500 // In this case, we are evaluating whether to select edge -> Succ, e.g.
1501 // set Succ as the layout successor of BB. Picking Succ as BB's
1502 // successor breaks the CFG constraints (FIXME: define these constraints).
1503 // With this layout, Pred BB
1504 // is forced to be outlined, so the overall cost will be cost of the
1505 // branch taken from BB to Pred, plus the cost of back taken branch
1506 // from Pred to Succ, as well as the additional cost associated
1507 // with the needed unconditional jump instruction from Pred To Succ.
1508
1509 // The cost of the topological order layout is the taken branch cost
1510 // from BB to Succ, so to make BB->Succ a viable candidate, the following
1511 // must hold:
1512 // 2 * freq(BB->Pred) * taken_branch_cost + unconditional_jump_cost
1513 // < freq(BB->Succ) * taken_branch_cost.
1514 // Ignoring unconditional jump cost, we get
1515 // freq(BB->Succ) > 2 * freq(BB->Pred), i.e.,
1516 // prob(BB->Succ) > 2 * prob(BB->Pred)
1517 //
1518 // When real profile data is available, we can precisely compute the
1519 // probability threshold that is needed for edge BB->Succ to be considered.
1520 // Without profile data, the heuristic requires the branch bias to be
1521 // a lot larger to make sure the signal is very strong (e.g. 80% default).
1522 // -----------------------------------------------------------------
1523 // Case 2: diamond like CFG (if-then-else):
1524 // S
1525 // / \
1526 // | \
1527 // BB Pred
1528 // \ /
1529 // Succ
1530 // ..
1531 //
1532 // The current block is BB and edge BB->Succ is now being evaluated.
1533 // Note that edge S->BB was previously already selected because
1534 // prob(S->BB) > prob(S->Pred).
1535 // At this point, 2 blocks can be placed after BB: Pred or Succ. If we
1536 // choose Pred, we will have a topological ordering as shown on the left
1537 // in the picture below. If we choose Succ, we have the solution as shown
1538 // on the right:
1539 //
1540 // topo-order:
1541 //
1542 // S----- ---S
1543 // | | | |
1544 // ---BB | | BB
1545 // | | | |
1546 // | Pred-- | Succ--
1547 // | | | |
1548 // ---Succ ---Pred--
1549 //
1550 // cost = freq(S->Pred) + freq(BB->Succ) cost = 2 * freq (S->Pred)
1551 // = freq(S->Pred) + freq(S->BB)
1552 //
1553 // If we have profile data (i.e, branch probabilities can be trusted), the
1554 // cost (number of taken branches) with layout S->BB->Succ->Pred is 2 *
1555 // freq(S->Pred) while the cost of topo order is freq(S->Pred) + freq(S->BB).
1556 // We know Prob(S->BB) > Prob(S->Pred), so freq(S->BB) > freq(S->Pred), which
1557 // means the cost of topological order is greater.
1558 // When profile data is not available, however, we need to be more
1559 // conservative. If the branch prediction is wrong, breaking the topo-order
1560 // will actually yield a layout with large cost. For this reason, we need
1561 // strong biased branch at block S with Prob(S->BB) in order to select
1562 // BB->Succ. This is equivalent to looking the CFG backward with backward
1563 // edge: Prob(Succ->BB) needs to >= HotProb in order to be selected (without
1564 // profile data).
1565 // --------------------------------------------------------------------------
1566 // Case 3: forked diamond
1567 // S
1568 // / \
1569 // / \
1570 // BB Pred
1571 // | \ / |
1572 // | \ / |
1573 // | X |
1574 // | / \ |
1575 // | / \ |
1576 // S1 S2
1577 //
1578 // The current block is BB and edge BB->S1 is now being evaluated.
1579 // As above S->BB was already selected because
1580 // prob(S->BB) > prob(S->Pred). Assume that prob(BB->S1) >= prob(BB->S2).
1581 //
1582 // topo-order:
1583 //
1584 // S-------| ---S
1585 // | | | |
1586 // ---BB | | BB
1587 // | | | |
1588 // | Pred----| | S1----
1589 // | | | |
1590 // --(S1 or S2) ---Pred--
1591 // |
1592 // S2
1593 //
1594 // topo-cost = freq(S->Pred) + freq(BB->S1) + freq(BB->S2)
1595 // + min(freq(Pred->S1), freq(Pred->S2))
1596 // Non-topo-order cost:
1597 // non-topo-cost = 2 * freq(S->Pred) + freq(BB->S2).
1598 // To be conservative, we can assume that min(freq(Pred->S1), freq(Pred->S2))
1599 // is 0. Then the non topo layout is better when
1600 // freq(S->Pred) < freq(BB->S1).
1601 // This is exactly what is checked below.
1602 // Note there are other shapes that apply (Pred may not be a single block,
1603 // but they all fit this general pattern.)
1604 BranchProbability HotProb = getLayoutSuccessorProbThreshold(BB);
1605
1606 // Make sure that a hot successor doesn't have a globally more
1607 // important predecessor.
1608 BlockFrequency CandidateEdgeFreq = MBFI->getBlockFreq(MBB: BB) * RealSuccProb;
1609 bool BadCFGConflict = false;
1610
1611 for (MachineBasicBlock *Pred : Succ->predecessors()) {
1612 BlockChain *PredChain = BlockToChain[Pred];
1613 if (Pred == Succ || PredChain == &SuccChain ||
1614 (BlockFilter && !BlockFilter->count(key: Pred)) || PredChain == &Chain ||
1615 Pred != *std::prev(x: PredChain->end()) ||
1616 // This check is redundant except for look ahead. This function is
1617 // called for lookahead by isProfitableToTailDup when BB hasn't been
1618 // placed yet.
1619 (Pred == BB))
1620 continue;
1621 // Do backward checking.
1622 // For all cases above, we need a backward checking to filter out edges that
1623 // are not 'strongly' biased.
1624 // BB Pred
1625 // \ /
1626 // Succ
1627 // We select edge BB->Succ if
1628 // freq(BB->Succ) > freq(Succ) * HotProb
1629 // i.e. freq(BB->Succ) > freq(BB->Succ) * HotProb + freq(Pred->Succ) *
1630 // HotProb
1631 // i.e. freq((BB->Succ) * (1 - HotProb) > freq(Pred->Succ) * HotProb
1632 // Case 1 is covered too, because the first equation reduces to:
1633 // prob(BB->Succ) > HotProb. (freq(Succ) = freq(BB) for a triangle)
1634 BlockFrequency PredEdgeFreq =
1635 MBFI->getBlockFreq(MBB: Pred) * MBPI->getEdgeProbability(Src: Pred, Dst: Succ);
1636 if (PredEdgeFreq * HotProb >= CandidateEdgeFreq * HotProb.getCompl()) {
1637 BadCFGConflict = true;
1638 break;
1639 }
1640 }
1641
1642 if (BadCFGConflict) {
1643 LLVM_DEBUG(dbgs() << " Not a candidate: " << getBlockName(Succ) << " -> "
1644 << SuccProb << " (prob) (non-cold CFG conflict)\n");
1645 return true;
1646 }
1647
1648 return false;
1649}
1650
1651/// Select the best successor for a block.
1652///
1653/// This looks across all successors of a particular block and attempts to
1654/// select the "best" one to be the layout successor. It only considers direct
1655/// successors which also pass the block filter. It will attempt to avoid
1656/// breaking CFG structure, but cave and break such structures in the case of
1657/// very hot successor edges.
1658///
1659/// \returns The best successor block found, or null if none are viable, along
1660/// with a boolean indicating if tail duplication is necessary.
1661MachineBlockPlacement::BlockAndTailDupResult
1662MachineBlockPlacement::selectBestSuccessor(const MachineBasicBlock *BB,
1663 const BlockChain &Chain,
1664 const BlockFilterSet *BlockFilter) {
1665 const BranchProbability HotProb(StaticLikelyProb, 100);
1666
1667 BlockAndTailDupResult BestSucc = {.BB: nullptr, .ShouldTailDup: false};
1668 auto BestProb = BranchProbability::getZero();
1669
1670 SmallVector<MachineBasicBlock *, 4> Successors;
1671 auto AdjustedSumProb =
1672 collectViableSuccessors(BB, Chain, BlockFilter, Successors);
1673
1674 LLVM_DEBUG(dbgs() << "Selecting best successor for: " << getBlockName(BB)
1675 << "\n");
1676
1677 // if we already precomputed the best successor for BB, return that if still
1678 // applicable.
1679 auto FoundEdge = ComputedEdges.find(Val: BB);
1680 if (FoundEdge != ComputedEdges.end()) {
1681 MachineBasicBlock *Succ = FoundEdge->second.BB;
1682 ComputedEdges.erase(I: FoundEdge);
1683 BlockChain *SuccChain = BlockToChain[Succ];
1684 if (BB->isSuccessor(MBB: Succ) && (!BlockFilter || BlockFilter->count(key: Succ)) &&
1685 SuccChain != &Chain && Succ == *SuccChain->begin())
1686 return FoundEdge->second;
1687 }
1688
1689 // if BB is part of a trellis, Use the trellis to determine the optimal
1690 // fallthrough edges
1691 if (isTrellis(BB, ViableSuccs: Successors, Chain, BlockFilter))
1692 return getBestTrellisSuccessor(BB, ViableSuccs: Successors, AdjustedSumProb, Chain,
1693 BlockFilter);
1694
1695 // For blocks with CFG violations, we may be able to lay them out anyway with
1696 // tail-duplication. We keep this vector so we can perform the probability
1697 // calculations the minimum number of times.
1698 SmallVector<std::pair<BranchProbability, MachineBasicBlock *>, 4>
1699 DupCandidates;
1700 for (MachineBasicBlock *Succ : Successors) {
1701 auto RealSuccProb = MBPI->getEdgeProbability(Src: BB, Dst: Succ);
1702 BranchProbability SuccProb =
1703 getAdjustedProbability(OrigProb: RealSuccProb, AdjustedSumProb);
1704
1705 BlockChain &SuccChain = *BlockToChain[Succ];
1706 // Skip the edge \c BB->Succ if block \c Succ has a better layout
1707 // predecessor that yields lower global cost.
1708 if (hasBetterLayoutPredecessor(BB, Succ, SuccChain, SuccProb, RealSuccProb,
1709 Chain, BlockFilter)) {
1710 // If tail duplication would make Succ profitable, place it.
1711 if (allowTailDupPlacement(MF&: *F) && shouldTailDuplicate(BB: Succ))
1712 DupCandidates.emplace_back(Args&: SuccProb, Args&: Succ);
1713 continue;
1714 }
1715
1716 LLVM_DEBUG(
1717 dbgs() << " Candidate: " << getBlockName(Succ)
1718 << ", probability: " << SuccProb
1719 << (SuccChain.UnscheduledPredecessors != 0 ? " (CFG break)" : "")
1720 << "\n");
1721
1722 if (BestSucc.BB && BestProb >= SuccProb) {
1723 LLVM_DEBUG(dbgs() << " Not the best candidate, continuing\n");
1724 continue;
1725 }
1726
1727 LLVM_DEBUG(dbgs() << " Setting it as best candidate\n");
1728 BestSucc.BB = Succ;
1729 BestProb = SuccProb;
1730 }
1731 // Handle the tail duplication candidates in order of decreasing probability.
1732 // Stop at the first one that is profitable. Also stop if they are less
1733 // profitable than BestSucc. Position is important because we preserve it and
1734 // prefer first best match. Here we aren't comparing in order, so we capture
1735 // the position instead.
1736 llvm::stable_sort(Range&: DupCandidates,
1737 C: [](std::tuple<BranchProbability, MachineBasicBlock *> L,
1738 std::tuple<BranchProbability, MachineBasicBlock *> R) {
1739 return std::get<0>(t&: L) > std::get<0>(t&: R);
1740 });
1741 for (auto &Tup : DupCandidates) {
1742 BranchProbability DupProb;
1743 MachineBasicBlock *Succ;
1744 std::tie(args&: DupProb, args&: Succ) = Tup;
1745 if (DupProb < BestProb)
1746 break;
1747 if (canTailDuplicateUnplacedPreds(BB, Succ, Chain, BlockFilter) &&
1748 (isProfitableToTailDup(BB, Succ, QProb: BestProb, Chain, BlockFilter))) {
1749 LLVM_DEBUG(dbgs() << " Candidate: " << getBlockName(Succ)
1750 << ", probability: " << DupProb
1751 << " (Tail Duplicate)\n");
1752 BestSucc.BB = Succ;
1753 BestSucc.ShouldTailDup = true;
1754 break;
1755 }
1756 }
1757
1758 if (BestSucc.BB)
1759 LLVM_DEBUG(dbgs() << " Selected: " << getBlockName(BestSucc.BB) << "\n");
1760
1761 return BestSucc;
1762}
1763
1764/// Select the best block from a worklist.
1765///
1766/// This looks through the provided worklist as a list of candidate basic
1767/// blocks and select the most profitable one to place. The definition of
1768/// profitable only really makes sense in the context of a loop. This returns
1769/// the most frequently visited block in the worklist, which in the case of
1770/// a loop, is the one most desirable to be physically close to the rest of the
1771/// loop body in order to improve i-cache behavior.
1772///
1773/// \returns The best block found, or null if none are viable.
1774MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock(
1775 const BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList) {
1776 // Once we need to walk the worklist looking for a candidate, cleanup the
1777 // worklist of already placed entries.
1778 // FIXME: If this shows up on profiles, it could be folded (at the cost of
1779 // some code complexity) into the loop below.
1780 llvm::erase_if(C&: WorkList, P: [&](MachineBasicBlock *BB) {
1781 return BlockToChain.lookup(Val: BB) == &Chain;
1782 });
1783
1784 if (WorkList.empty())
1785 return nullptr;
1786
1787 bool IsEHPad = WorkList[0]->isEHPad();
1788
1789 MachineBasicBlock *BestBlock = nullptr;
1790 BlockFrequency BestFreq;
1791 for (MachineBasicBlock *MBB : WorkList) {
1792 assert(MBB->isEHPad() == IsEHPad &&
1793 "EHPad mismatch between block and work list.");
1794
1795 BlockChain &SuccChain = *BlockToChain[MBB];
1796 if (&SuccChain == &Chain)
1797 continue;
1798
1799 assert(SuccChain.UnscheduledPredecessors == 0 &&
1800 "Found CFG-violating block");
1801
1802 BlockFrequency CandidateFreq = MBFI->getBlockFreq(MBB);
1803 LLVM_DEBUG(dbgs() << " " << getBlockName(MBB) << " -> "
1804 << printBlockFreq(MBFI->getMBFI(), CandidateFreq)
1805 << " (freq)\n");
1806
1807 // For ehpad, we layout the least probable first as to avoid jumping back
1808 // from least probable landingpads to more probable ones.
1809 //
1810 // FIXME: Using probability is probably (!) not the best way to achieve
1811 // this. We should probably have a more principled approach to layout
1812 // cleanup code.
1813 //
1814 // The goal is to get:
1815 //
1816 // +--------------------------+
1817 // | V
1818 // InnerLp -> InnerCleanup OuterLp -> OuterCleanup -> Resume
1819 //
1820 // Rather than:
1821 //
1822 // +-------------------------------------+
1823 // V |
1824 // OuterLp -> OuterCleanup -> Resume InnerLp -> InnerCleanup
1825 if (BestBlock && (IsEHPad ^ (BestFreq >= CandidateFreq)))
1826 continue;
1827
1828 BestBlock = MBB;
1829 BestFreq = CandidateFreq;
1830 }
1831
1832 return BestBlock;
1833}
1834
1835/// Retrieve the first unplaced basic block in the entire function.
1836///
1837/// This routine is called when we are unable to use the CFG to walk through
1838/// all of the basic blocks and form a chain due to unnatural loops in the CFG.
1839/// We walk through the function's blocks in order, starting from the
1840/// LastUnplacedBlockIt. We update this iterator on each call to avoid
1841/// re-scanning the entire sequence on repeated calls to this routine.
1842MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock(
1843 const BlockChain &PlacedChain,
1844 MachineFunction::iterator &PrevUnplacedBlockIt) {
1845
1846 for (MachineFunction::iterator I = PrevUnplacedBlockIt, E = F->end(); I != E;
1847 ++I) {
1848 if (BlockChain *Chain = BlockToChain[&*I]; Chain != &PlacedChain) {
1849 PrevUnplacedBlockIt = I;
1850 // Now select the head of the chain to which the unplaced block belongs
1851 // as the block to place. This will force the entire chain to be placed,
1852 // and satisfies the requirements of merging chains.
1853 return *Chain->begin();
1854 }
1855 }
1856 return nullptr;
1857}
1858
1859/// Retrieve the first unplaced basic block among the blocks in BlockFilter.
1860///
1861/// This is similar to getFirstUnplacedBlock for the entire function, but since
1862/// the size of BlockFilter is typically far less than the number of blocks in
1863/// the entire function, iterating through the BlockFilter is more efficient.
1864/// When processing the entire funciton, using the version without BlockFilter
1865/// has a complexity of #(loops in function) * #(blocks in function), while this
1866/// version has a complexity of sum(#(loops in block) foreach block in function)
1867/// which is always smaller. For long function mostly sequential in structure,
1868/// the complexity is amortized to 1 * #(blocks in function).
1869MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock(
1870 const BlockChain &PlacedChain,
1871 BlockFilterSet::iterator &PrevUnplacedBlockInFilterIt,
1872 const BlockFilterSet *BlockFilter) {
1873 assert(BlockFilter);
1874 for (; PrevUnplacedBlockInFilterIt != BlockFilter->end();
1875 ++PrevUnplacedBlockInFilterIt) {
1876 BlockChain *C = BlockToChain[*PrevUnplacedBlockInFilterIt];
1877 if (C != &PlacedChain) {
1878 return *C->begin();
1879 }
1880 }
1881 return nullptr;
1882}
1883
1884void MachineBlockPlacement::fillWorkLists(
1885 const MachineBasicBlock *MBB, SmallPtrSetImpl<BlockChain *> &UpdatedPreds,
1886 const BlockFilterSet *BlockFilter = nullptr) {
1887 BlockChain &Chain = *BlockToChain[MBB];
1888 if (!UpdatedPreds.insert(Ptr: &Chain).second)
1889 return;
1890
1891 assert(
1892 Chain.UnscheduledPredecessors == 0 &&
1893 "Attempting to place block with unscheduled predecessors in worklist.");
1894 for (MachineBasicBlock *ChainBB : Chain) {
1895 assert(BlockToChain[ChainBB] == &Chain &&
1896 "Block in chain doesn't match BlockToChain map.");
1897 for (MachineBasicBlock *Pred : ChainBB->predecessors()) {
1898 if (BlockFilter && !BlockFilter->count(key: Pred))
1899 continue;
1900 if (BlockToChain[Pred] == &Chain)
1901 continue;
1902 ++Chain.UnscheduledPredecessors;
1903 }
1904 }
1905
1906 if (Chain.UnscheduledPredecessors != 0)
1907 return;
1908
1909 MachineBasicBlock *BB = *Chain.begin();
1910 if (BB->isEHPad())
1911 EHPadWorkList.push_back(Elt: BB);
1912 else
1913 BlockWorkList.push_back(Elt: BB);
1914}
1915
1916void MachineBlockPlacement::buildChain(const MachineBasicBlock *HeadBB,
1917 BlockChain &Chain,
1918 BlockFilterSet *BlockFilter) {
1919 assert(HeadBB && "BB must not be null.\n");
1920 assert(BlockToChain[HeadBB] == &Chain && "BlockToChainMap mis-match.\n");
1921 MachineFunction::iterator PrevUnplacedBlockIt = F->begin();
1922 BlockFilterSet::iterator PrevUnplacedBlockInFilterIt;
1923 if (BlockFilter)
1924 PrevUnplacedBlockInFilterIt = BlockFilter->begin();
1925
1926 const MachineBasicBlock *LoopHeaderBB = HeadBB;
1927 markChainSuccessors(Chain, LoopHeaderBB, BlockFilter);
1928 MachineBasicBlock *BB = *std::prev(x: Chain.end());
1929 while (true) {
1930 assert(BB && "null block found at end of chain in loop.");
1931 assert(BlockToChain[BB] == &Chain && "BlockToChainMap mis-match in loop.");
1932 assert(*std::prev(Chain.end()) == BB && "BB Not found at end of chain.");
1933
1934 // Look for the best viable successor if there is one to place immediately
1935 // after this block.
1936 auto Result = selectBestSuccessor(BB, Chain, BlockFilter);
1937 MachineBasicBlock *BestSucc = Result.BB;
1938 bool ShouldTailDup = Result.ShouldTailDup;
1939 if (allowTailDupPlacement(MF&: *F))
1940 ShouldTailDup |= (BestSucc && canTailDuplicateUnplacedPreds(
1941 BB, Succ: BestSucc, Chain, BlockFilter));
1942
1943 // If an immediate successor isn't available, look for the best viable
1944 // block among those we've identified as not violating the loop's CFG at
1945 // this point. This won't be a fallthrough, but it will increase locality.
1946 if (!BestSucc)
1947 BestSucc = selectBestCandidateBlock(Chain, WorkList&: BlockWorkList);
1948 if (!BestSucc)
1949 BestSucc = selectBestCandidateBlock(Chain, WorkList&: EHPadWorkList);
1950
1951 if (!BestSucc) {
1952 if (BlockFilter)
1953 BestSucc = getFirstUnplacedBlock(PlacedChain: Chain, PrevUnplacedBlockInFilterIt,
1954 BlockFilter);
1955 else
1956 BestSucc = getFirstUnplacedBlock(PlacedChain: Chain, PrevUnplacedBlockIt);
1957 if (!BestSucc)
1958 break;
1959
1960 LLVM_DEBUG(dbgs() << "Unnatural loop CFG detected, forcibly merging the "
1961 "layout successor until the CFG reduces\n");
1962 }
1963
1964 // Placement may have changed tail duplication opportunities.
1965 // Check for that now.
1966 if (allowTailDupPlacement(MF&: *F) && BestSucc && ShouldTailDup) {
1967 repeatedlyTailDuplicateBlock(BB: BestSucc, LPred&: BB, LoopHeaderBB, Chain,
1968 BlockFilter, PrevUnplacedBlockIt,
1969 PrevUnplacedBlockInFilterIt);
1970 // If the chosen successor was duplicated into BB, don't bother laying
1971 // it out, just go round the loop again with BB as the chain end.
1972 if (!BB->isSuccessor(MBB: BestSucc))
1973 continue;
1974 }
1975
1976 // Place this block, updating the datastructures to reflect its placement.
1977 BlockChain &SuccChain = *BlockToChain[BestSucc];
1978 // Zero out UnscheduledPredecessors for the successor we're about to merge
1979 // in case we selected a successor that didn't fit naturally into the CFG.
1980 SuccChain.UnscheduledPredecessors = 0;
1981 LLVM_DEBUG(dbgs() << "Merging from " << getBlockName(BB) << " to "
1982 << getBlockName(BestSucc) << "\n");
1983 markChainSuccessors(Chain: SuccChain, LoopHeaderBB, BlockFilter);
1984 Chain.merge(BB: BestSucc, Chain: &SuccChain);
1985 BB = *std::prev(x: Chain.end());
1986 }
1987
1988 LLVM_DEBUG(dbgs() << "Finished forming chain for header block "
1989 << getBlockName(*Chain.begin()) << "\n");
1990}
1991
1992// If bottom of block BB has only one successor OldTop, in most cases it is
1993// profitable to move it before OldTop, except the following case:
1994//
1995// -->OldTop<-
1996// | . |
1997// | . |
1998// | . |
1999// ---Pred |
2000// | |
2001// BB-----
2002//
2003// If BB is moved before OldTop, Pred needs a taken branch to BB, and it can't
2004// layout the other successor below it, so it can't reduce taken branch.
2005// In this case we keep its original layout.
2006bool MachineBlockPlacement::canMoveBottomBlockToTop(
2007 const MachineBasicBlock *BottomBlock, const MachineBasicBlock *OldTop) {
2008 if (BottomBlock->pred_size() != 1)
2009 return true;
2010 MachineBasicBlock *Pred = *BottomBlock->pred_begin();
2011 if (Pred->succ_size() != 2)
2012 return true;
2013
2014 MachineBasicBlock *OtherBB = *Pred->succ_begin();
2015 if (OtherBB == BottomBlock)
2016 OtherBB = *Pred->succ_rbegin();
2017 if (OtherBB == OldTop)
2018 return false;
2019
2020 return true;
2021}
2022
2023// Find out the possible fall through frequence to the top of a loop.
2024BlockFrequency
2025MachineBlockPlacement::TopFallThroughFreq(const MachineBasicBlock *Top,
2026 const BlockFilterSet &LoopBlockSet) {
2027 BlockFrequency MaxFreq = BlockFrequency(0);
2028 for (MachineBasicBlock *Pred : Top->predecessors()) {
2029 BlockChain *PredChain = BlockToChain[Pred];
2030 if (!LoopBlockSet.count(key: Pred) &&
2031 (!PredChain || Pred == *std::prev(x: PredChain->end()))) {
2032 // Found a Pred block can be placed before Top.
2033 // Check if Top is the best successor of Pred.
2034 auto TopProb = MBPI->getEdgeProbability(Src: Pred, Dst: Top);
2035 bool TopOK = true;
2036 for (MachineBasicBlock *Succ : Pred->successors()) {
2037 auto SuccProb = MBPI->getEdgeProbability(Src: Pred, Dst: Succ);
2038 BlockChain *SuccChain = BlockToChain[Succ];
2039 // Check if Succ can be placed after Pred.
2040 // Succ should not be in any chain, or it is the head of some chain.
2041 if (!LoopBlockSet.count(key: Succ) && (SuccProb > TopProb) &&
2042 (!SuccChain || Succ == *SuccChain->begin())) {
2043 TopOK = false;
2044 break;
2045 }
2046 }
2047 if (TopOK) {
2048 BlockFrequency EdgeFreq =
2049 MBFI->getBlockFreq(MBB: Pred) * MBPI->getEdgeProbability(Src: Pred, Dst: Top);
2050 if (EdgeFreq > MaxFreq)
2051 MaxFreq = EdgeFreq;
2052 }
2053 }
2054 }
2055 return MaxFreq;
2056}
2057
2058// Compute the fall through gains when move NewTop before OldTop.
2059//
2060// In following diagram, edges marked as "-" are reduced fallthrough, edges
2061// marked as "+" are increased fallthrough, this function computes
2062//
2063// SUM(increased fallthrough) - SUM(decreased fallthrough)
2064//
2065// |
2066// | -
2067// V
2068// --->OldTop
2069// | .
2070// | .
2071// +| . +
2072// | Pred --->
2073// | |-
2074// | V
2075// --- NewTop <---
2076// |-
2077// V
2078//
2079BlockFrequency MachineBlockPlacement::FallThroughGains(
2080 const MachineBasicBlock *NewTop, const MachineBasicBlock *OldTop,
2081 const MachineBasicBlock *ExitBB, const BlockFilterSet &LoopBlockSet) {
2082 BlockFrequency FallThrough2Top = TopFallThroughFreq(Top: OldTop, LoopBlockSet);
2083 BlockFrequency FallThrough2Exit = BlockFrequency(0);
2084 if (ExitBB)
2085 FallThrough2Exit =
2086 MBFI->getBlockFreq(MBB: NewTop) * MBPI->getEdgeProbability(Src: NewTop, Dst: ExitBB);
2087 BlockFrequency BackEdgeFreq =
2088 MBFI->getBlockFreq(MBB: NewTop) * MBPI->getEdgeProbability(Src: NewTop, Dst: OldTop);
2089
2090 // Find the best Pred of NewTop.
2091 MachineBasicBlock *BestPred = nullptr;
2092 BlockFrequency FallThroughFromPred = BlockFrequency(0);
2093 for (MachineBasicBlock *Pred : NewTop->predecessors()) {
2094 if (!LoopBlockSet.count(key: Pred))
2095 continue;
2096 BlockChain *PredChain = BlockToChain[Pred];
2097 if (!PredChain || Pred == *std::prev(x: PredChain->end())) {
2098 BlockFrequency EdgeFreq =
2099 MBFI->getBlockFreq(MBB: Pred) * MBPI->getEdgeProbability(Src: Pred, Dst: NewTop);
2100 if (EdgeFreq > FallThroughFromPred) {
2101 FallThroughFromPred = EdgeFreq;
2102 BestPred = Pred;
2103 }
2104 }
2105 }
2106
2107 // If NewTop is not placed after Pred, another successor can be placed
2108 // after Pred.
2109 BlockFrequency NewFreq = BlockFrequency(0);
2110 if (BestPred) {
2111 for (MachineBasicBlock *Succ : BestPred->successors()) {
2112 if ((Succ == NewTop) || (Succ == BestPred) || !LoopBlockSet.count(key: Succ))
2113 continue;
2114 if (ComputedEdges.contains(Val: Succ))
2115 continue;
2116 BlockChain *SuccChain = BlockToChain[Succ];
2117 if ((SuccChain && (Succ != *SuccChain->begin())) ||
2118 (SuccChain == BlockToChain[BestPred]))
2119 continue;
2120 BlockFrequency EdgeFreq = MBFI->getBlockFreq(MBB: BestPred) *
2121 MBPI->getEdgeProbability(Src: BestPred, Dst: Succ);
2122 if (EdgeFreq > NewFreq)
2123 NewFreq = EdgeFreq;
2124 }
2125 BlockFrequency OrigEdgeFreq = MBFI->getBlockFreq(MBB: BestPred) *
2126 MBPI->getEdgeProbability(Src: BestPred, Dst: NewTop);
2127 if (NewFreq > OrigEdgeFreq) {
2128 // If NewTop is not the best successor of Pred, then Pred doesn't
2129 // fallthrough to NewTop. So there is no FallThroughFromPred and
2130 // NewFreq.
2131 NewFreq = BlockFrequency(0);
2132 FallThroughFromPred = BlockFrequency(0);
2133 }
2134 }
2135
2136 BlockFrequency Result = BlockFrequency(0);
2137 BlockFrequency Gains = BackEdgeFreq + NewFreq;
2138 BlockFrequency Lost =
2139 FallThrough2Top + FallThrough2Exit + FallThroughFromPred;
2140 if (Gains > Lost)
2141 Result = Gains - Lost;
2142 return Result;
2143}
2144
2145/// Helper function of findBestLoopTop. Find the best loop top block
2146/// from predecessors of old top.
2147///
2148/// Look for a block which is strictly better than the old top for laying
2149/// out before the old top of the loop. This looks for only two patterns:
2150///
2151/// 1. a block has only one successor, the old loop top
2152///
2153/// Because such a block will always result in an unconditional jump,
2154/// rotating it in front of the old top is always profitable.
2155///
2156/// 2. a block has two successors, one is old top, another is exit
2157/// and it has more than one predecessors
2158///
2159/// If it is below one of its predecessors P, only P can fall through to
2160/// it, all other predecessors need a jump to it, and another conditional
2161/// jump to loop header. If it is moved before loop header, all its
2162/// predecessors jump to it, then fall through to loop header. So all its
2163/// predecessors except P can reduce one taken branch.
2164/// At the same time, move it before old top increases the taken branch
2165/// to loop exit block, so the reduced taken branch will be compared with
2166/// the increased taken branch to the loop exit block.
2167MachineBasicBlock *MachineBlockPlacement::findBestLoopTopHelper(
2168 MachineBasicBlock *OldTop, const MachineLoop &L,
2169 const BlockFilterSet &LoopBlockSet) {
2170 // Check that the header hasn't been fused with a preheader block due to
2171 // crazy branches. If it has, we need to start with the header at the top to
2172 // prevent pulling the preheader into the loop body.
2173 BlockChain &HeaderChain = *BlockToChain[OldTop];
2174 if (!LoopBlockSet.count(key: *HeaderChain.begin()))
2175 return OldTop;
2176 if (OldTop != *HeaderChain.begin())
2177 return OldTop;
2178
2179 LLVM_DEBUG(dbgs() << "Finding best loop top for: " << getBlockName(OldTop)
2180 << "\n");
2181
2182 BlockFrequency BestGains = BlockFrequency(0);
2183 MachineBasicBlock *BestPred = nullptr;
2184 for (MachineBasicBlock *Pred : OldTop->predecessors()) {
2185 if (!LoopBlockSet.count(key: Pred))
2186 continue;
2187 if (Pred == L.getHeader())
2188 continue;
2189 LLVM_DEBUG(dbgs() << " old top pred: " << getBlockName(Pred) << ", has "
2190 << Pred->succ_size() << " successors, "
2191 << printBlockFreq(MBFI->getMBFI(), *Pred) << " freq\n");
2192 if (Pred->succ_size() > 2)
2193 continue;
2194
2195 MachineBasicBlock *OtherBB = nullptr;
2196 if (Pred->succ_size() == 2) {
2197 OtherBB = *Pred->succ_begin();
2198 if (OtherBB == OldTop)
2199 OtherBB = *Pred->succ_rbegin();
2200 }
2201
2202 if (!canMoveBottomBlockToTop(BottomBlock: Pred, OldTop))
2203 continue;
2204
2205 BlockFrequency Gains =
2206 FallThroughGains(NewTop: Pred, OldTop, ExitBB: OtherBB, LoopBlockSet);
2207 if ((Gains > BlockFrequency(0)) &&
2208 (Gains > BestGains ||
2209 ((Gains == BestGains) && Pred->isLayoutSuccessor(MBB: OldTop)))) {
2210 BestPred = Pred;
2211 BestGains = Gains;
2212 }
2213 }
2214
2215 // If no direct predecessor is fine, just use the loop header.
2216 if (!BestPred) {
2217 LLVM_DEBUG(dbgs() << " final top unchanged\n");
2218 return OldTop;
2219 }
2220
2221 // Walk backwards through any straight line of predecessors.
2222 while (BestPred->pred_size() == 1 &&
2223 (*BestPred->pred_begin())->succ_size() == 1 &&
2224 *BestPred->pred_begin() != L.getHeader())
2225 BestPred = *BestPred->pred_begin();
2226
2227 LLVM_DEBUG(dbgs() << " final top: " << getBlockName(BestPred) << "\n");
2228 return BestPred;
2229}
2230
2231/// Find the best loop top block for layout.
2232///
2233/// This function iteratively calls findBestLoopTopHelper, until no new better
2234/// BB can be found.
2235MachineBasicBlock *
2236MachineBlockPlacement::findBestLoopTop(const MachineLoop &L,
2237 const BlockFilterSet &LoopBlockSet) {
2238 // Placing the latch block before the header may introduce an extra branch
2239 // that skips this block the first time the loop is executed, which we want
2240 // to avoid when optimising for size.
2241 // FIXME: in theory there is a case that does not introduce a new branch,
2242 // i.e. when the layout predecessor does not fallthrough to the loop header.
2243 // In practice this never happens though: there always seems to be a preheader
2244 // that can fallthrough and that is also placed before the header.
2245 if (llvm::shouldOptimizeForSize(MBB: L.getHeader(), PSI, MBFIWrapper: MBFI.get()))
2246 return L.getHeader();
2247
2248 MachineBasicBlock *OldTop = nullptr;
2249 MachineBasicBlock *NewTop = L.getHeader();
2250 while (NewTop != OldTop) {
2251 OldTop = NewTop;
2252 NewTop = findBestLoopTopHelper(OldTop, L, LoopBlockSet);
2253 if (NewTop != OldTop)
2254 ComputedEdges[NewTop] = {.BB: OldTop, .ShouldTailDup: false};
2255 }
2256 return NewTop;
2257}
2258
2259/// Find the best loop exiting block for layout.
2260///
2261/// This routine implements the logic to analyze the loop looking for the best
2262/// block to layout at the top of the loop. Typically this is done to maximize
2263/// fallthrough opportunities.
2264MachineBasicBlock *
2265MachineBlockPlacement::findBestLoopExit(const MachineLoop &L,
2266 const BlockFilterSet &LoopBlockSet,
2267 BlockFrequency &ExitFreq) {
2268 // We don't want to layout the loop linearly in all cases. If the loop header
2269 // is just a normal basic block in the loop, we want to look for what block
2270 // within the loop is the best one to layout at the top. However, if the loop
2271 // header has be pre-merged into a chain due to predecessors not having
2272 // analyzable branches, *and* the predecessor it is merged with is *not* part
2273 // of the loop, rotating the header into the middle of the loop will create
2274 // a non-contiguous range of blocks which is Very Bad. So start with the
2275 // header and only rotate if safe.
2276 BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
2277 if (!LoopBlockSet.count(key: *HeaderChain.begin()))
2278 return nullptr;
2279
2280 BlockFrequency BestExitEdgeFreq;
2281 unsigned BestExitLoopDepth = 0;
2282 MachineBasicBlock *ExitingBB = nullptr;
2283 // If there are exits to outer loops, loop rotation can severely limit
2284 // fallthrough opportunities unless it selects such an exit. Keep a set of
2285 // blocks where rotating to exit with that block will reach an outer loop.
2286 SmallPtrSet<MachineBasicBlock *, 4> BlocksExitingToOuterLoop;
2287
2288 LLVM_DEBUG(dbgs() << "Finding best loop exit for: "
2289 << getBlockName(L.getHeader()) << "\n");
2290 for (MachineBasicBlock *MBB : L.getBlocks()) {
2291 BlockChain &Chain = *BlockToChain[MBB];
2292 // Ensure that this block is at the end of a chain; otherwise it could be
2293 // mid-way through an inner loop or a successor of an unanalyzable branch.
2294 if (MBB != *std::prev(x: Chain.end()))
2295 continue;
2296
2297 // Now walk the successors. We need to establish whether this has a viable
2298 // exiting successor and whether it has a viable non-exiting successor.
2299 // We store the old exiting state and restore it if a viable looping
2300 // successor isn't found.
2301 MachineBasicBlock *OldExitingBB = ExitingBB;
2302 BlockFrequency OldBestExitEdgeFreq = BestExitEdgeFreq;
2303 bool HasLoopingSucc = false;
2304 for (MachineBasicBlock *Succ : MBB->successors()) {
2305 if (Succ->isEHPad())
2306 continue;
2307 if (Succ == MBB)
2308 continue;
2309 BlockChain &SuccChain = *BlockToChain[Succ];
2310 // Don't split chains, either this chain or the successor's chain.
2311 if (&Chain == &SuccChain) {
2312 LLVM_DEBUG(dbgs() << " exiting: " << getBlockName(MBB) << " -> "
2313 << getBlockName(Succ) << " (chain conflict)\n");
2314 continue;
2315 }
2316
2317 auto SuccProb = MBPI->getEdgeProbability(Src: MBB, Dst: Succ);
2318 if (LoopBlockSet.count(key: Succ)) {
2319 LLVM_DEBUG(dbgs() << " looping: " << getBlockName(MBB) << " -> "
2320 << getBlockName(Succ) << " (" << SuccProb << ")\n");
2321 HasLoopingSucc = true;
2322 continue;
2323 }
2324
2325 unsigned SuccLoopDepth = 0;
2326 if (MachineLoop *ExitLoop = MLI->getLoopFor(BB: Succ)) {
2327 SuccLoopDepth = ExitLoop->getLoopDepth();
2328 if (ExitLoop->contains(L: &L))
2329 BlocksExitingToOuterLoop.insert(Ptr: MBB);
2330 }
2331
2332 BlockFrequency ExitEdgeFreq = MBFI->getBlockFreq(MBB) * SuccProb;
2333 LLVM_DEBUG(
2334 dbgs() << " exiting: " << getBlockName(MBB) << " -> "
2335 << getBlockName(Succ) << " [L:" << SuccLoopDepth << "] ("
2336 << printBlockFreq(MBFI->getMBFI(), ExitEdgeFreq) << ")\n");
2337 // Note that we bias this toward an existing layout successor to retain
2338 // incoming order in the absence of better information. The exit must have
2339 // a frequency higher than the current exit before we consider breaking
2340 // the layout.
2341 BranchProbability Bias(100 - ExitBlockBias, 100);
2342 if (!ExitingBB || SuccLoopDepth > BestExitLoopDepth ||
2343 ExitEdgeFreq > BestExitEdgeFreq ||
2344 (MBB->isLayoutSuccessor(MBB: Succ) &&
2345 !(ExitEdgeFreq < BestExitEdgeFreq * Bias))) {
2346 BestExitEdgeFreq = ExitEdgeFreq;
2347 ExitingBB = MBB;
2348 }
2349 }
2350
2351 if (!HasLoopingSucc) {
2352 // Restore the old exiting state, no viable looping successor was found.
2353 ExitingBB = OldExitingBB;
2354 BestExitEdgeFreq = OldBestExitEdgeFreq;
2355 }
2356 }
2357 // Without a candidate exiting block or with only a single block in the
2358 // loop, just use the loop header to layout the loop.
2359 if (!ExitingBB) {
2360 LLVM_DEBUG(
2361 dbgs() << " No other candidate exit blocks, using loop header\n");
2362 return nullptr;
2363 }
2364 if (L.getNumBlocks() == 1) {
2365 LLVM_DEBUG(dbgs() << " Loop has 1 block, using loop header as exit\n");
2366 return nullptr;
2367 }
2368
2369 // Also, if we have exit blocks which lead to outer loops but didn't select
2370 // one of them as the exiting block we are rotating toward, disable loop
2371 // rotation altogether.
2372 if (!BlocksExitingToOuterLoop.empty() &&
2373 !BlocksExitingToOuterLoop.count(Ptr: ExitingBB))
2374 return nullptr;
2375
2376 LLVM_DEBUG(dbgs() << " Best exiting block: " << getBlockName(ExitingBB)
2377 << "\n");
2378 ExitFreq = BestExitEdgeFreq;
2379 return ExitingBB;
2380}
2381
2382/// Check if there is a fallthrough to loop header Top.
2383///
2384/// 1. Look for a Pred that can be layout before Top.
2385/// 2. Check if Top is the most possible successor of Pred.
2386bool MachineBlockPlacement::hasViableTopFallthrough(
2387 const MachineBasicBlock *Top, const BlockFilterSet &LoopBlockSet) {
2388 for (MachineBasicBlock *Pred : Top->predecessors()) {
2389 BlockChain *PredChain = BlockToChain[Pred];
2390 if (!LoopBlockSet.count(key: Pred) &&
2391 (!PredChain || Pred == *std::prev(x: PredChain->end()))) {
2392 // Found a Pred block can be placed before Top.
2393 // Check if Top is the best successor of Pred.
2394 auto TopProb = MBPI->getEdgeProbability(Src: Pred, Dst: Top);
2395 bool TopOK = true;
2396 for (MachineBasicBlock *Succ : Pred->successors()) {
2397 auto SuccProb = MBPI->getEdgeProbability(Src: Pred, Dst: Succ);
2398 BlockChain *SuccChain = BlockToChain[Succ];
2399 // Check if Succ can be placed after Pred.
2400 // Succ should not be in any chain, or it is the head of some chain.
2401 if ((!SuccChain || Succ == *SuccChain->begin()) && SuccProb > TopProb) {
2402 TopOK = false;
2403 break;
2404 }
2405 }
2406 if (TopOK)
2407 return true;
2408 }
2409 }
2410 return false;
2411}
2412
2413/// Attempt to rotate an exiting block to the bottom of the loop.
2414///
2415/// Once we have built a chain, try to rotate it to line up the hot exit block
2416/// with fallthrough out of the loop if doing so doesn't introduce unnecessary
2417/// branches. For example, if the loop has fallthrough into its header and out
2418/// of its bottom already, don't rotate it.
2419void MachineBlockPlacement::rotateLoop(BlockChain &LoopChain,
2420 const MachineBasicBlock *ExitingBB,
2421 BlockFrequency ExitFreq,
2422 const BlockFilterSet &LoopBlockSet) {
2423 if (!ExitingBB)
2424 return;
2425
2426 MachineBasicBlock *Top = *LoopChain.begin();
2427 MachineBasicBlock *Bottom = *std::prev(x: LoopChain.end());
2428
2429 // If ExitingBB is already the last one in a chain then nothing to do.
2430 if (Bottom == ExitingBB)
2431 return;
2432
2433 // The entry block should always be the first BB in a function.
2434 if (Top->isEntryBlock())
2435 return;
2436
2437 bool ViableTopFallthrough = hasViableTopFallthrough(Top, LoopBlockSet);
2438
2439 // If the header has viable fallthrough, check whether the current loop
2440 // bottom is a viable exiting block. If so, bail out as rotating will
2441 // introduce an unnecessary branch.
2442 if (ViableTopFallthrough) {
2443 for (MachineBasicBlock *Succ : Bottom->successors()) {
2444 BlockChain *SuccChain = BlockToChain[Succ];
2445 if (!LoopBlockSet.count(key: Succ) &&
2446 (!SuccChain || Succ == *SuccChain->begin()))
2447 return;
2448 }
2449
2450 // Rotate will destroy the top fallthrough, we need to ensure the new exit
2451 // frequency is larger than top fallthrough.
2452 BlockFrequency FallThrough2Top = TopFallThroughFreq(Top, LoopBlockSet);
2453 if (FallThrough2Top >= ExitFreq)
2454 return;
2455 }
2456
2457 BlockChain::iterator ExitIt = llvm::find(Range&: LoopChain, Val: ExitingBB);
2458 if (ExitIt == LoopChain.end())
2459 return;
2460
2461 // Rotating a loop exit to the bottom when there is a fallthrough to top
2462 // trades the entry fallthrough for an exit fallthrough.
2463 // If there is no bottom->top edge, but the chosen exit block does have
2464 // a fallthrough, we break that fallthrough for nothing in return.
2465
2466 // Let's consider an example. We have a built chain of basic blocks
2467 // B1, B2, ..., Bn, where Bk is a ExitingBB - chosen exit block.
2468 // By doing a rotation we get
2469 // Bk+1, ..., Bn, B1, ..., Bk
2470 // Break of fallthrough to B1 is compensated by a fallthrough from Bk.
2471 // If we had a fallthrough Bk -> Bk+1 it is broken now.
2472 // It might be compensated by fallthrough Bn -> B1.
2473 // So we have a condition to avoid creation of extra branch by loop rotation.
2474 // All below must be true to avoid loop rotation:
2475 // If there is a fallthrough to top (B1)
2476 // There was fallthrough from chosen exit block (Bk) to next one (Bk+1)
2477 // There is no fallthrough from bottom (Bn) to top (B1).
2478 // Please note that there is no exit fallthrough from Bn because we checked it
2479 // above.
2480 if (ViableTopFallthrough) {
2481 assert(std::next(ExitIt) != LoopChain.end() &&
2482 "Exit should not be last BB");
2483 MachineBasicBlock *NextBlockInChain = *std::next(x: ExitIt);
2484 if (ExitingBB->isSuccessor(MBB: NextBlockInChain))
2485 if (!Bottom->isSuccessor(MBB: Top))
2486 return;
2487 }
2488
2489 LLVM_DEBUG(dbgs() << "Rotating loop to put exit " << getBlockName(ExitingBB)
2490 << " at bottom\n");
2491 std::rotate(first: LoopChain.begin(), middle: std::next(x: ExitIt), last: LoopChain.end());
2492}
2493
2494/// Attempt to rotate a loop based on profile data to reduce branch cost.
2495///
2496/// With profile data, we can determine the cost in terms of missed fall through
2497/// opportunities when rotating a loop chain and select the best rotation.
2498/// Basically, there are three kinds of cost to consider for each rotation:
2499/// 1. The possibly missed fall through edge (if it exists) from BB out of
2500/// the loop to the loop header.
2501/// 2. The possibly missed fall through edges (if they exist) from the loop
2502/// exits to BB out of the loop.
2503/// 3. The missed fall through edge (if it exists) from the last BB to the
2504/// first BB in the loop chain.
2505/// Therefore, the cost for a given rotation is the sum of costs listed above.
2506/// We select the best rotation with the smallest cost.
2507void MachineBlockPlacement::rotateLoopWithProfile(
2508 BlockChain &LoopChain, const MachineLoop &L,
2509 const BlockFilterSet &LoopBlockSet) {
2510 auto RotationPos = LoopChain.end();
2511 MachineBasicBlock *ChainHeaderBB = *LoopChain.begin();
2512
2513 // The entry block should always be the first BB in a function.
2514 if (ChainHeaderBB->isEntryBlock())
2515 return;
2516
2517 BlockFrequency SmallestRotationCost = BlockFrequency::max();
2518
2519 // A utility lambda that scales up a block frequency by dividing it by a
2520 // branch probability which is the reciprocal of the scale.
2521 auto ScaleBlockFrequency = [](BlockFrequency Freq,
2522 unsigned Scale) -> BlockFrequency {
2523 if (Scale == 0)
2524 return BlockFrequency(0);
2525 // Use operator / between BlockFrequency and BranchProbability to implement
2526 // saturating multiplication.
2527 return Freq / BranchProbability(1, Scale);
2528 };
2529
2530 // Compute the cost of the missed fall-through edge to the loop header if the
2531 // chain head is not the loop header. As we only consider natural loops with
2532 // single header, this computation can be done only once.
2533 BlockFrequency HeaderFallThroughCost(0);
2534 for (auto *Pred : ChainHeaderBB->predecessors()) {
2535 BlockChain *PredChain = BlockToChain[Pred];
2536 if (!LoopBlockSet.count(key: Pred) &&
2537 (!PredChain || Pred == *std::prev(x: PredChain->end()))) {
2538 auto EdgeFreq = MBFI->getBlockFreq(MBB: Pred) *
2539 MBPI->getEdgeProbability(Src: Pred, Dst: ChainHeaderBB);
2540 auto FallThruCost = ScaleBlockFrequency(EdgeFreq, MisfetchCost);
2541 // If the predecessor has only an unconditional jump to the header, we
2542 // need to consider the cost of this jump.
2543 if (Pred->succ_size() == 1)
2544 FallThruCost += ScaleBlockFrequency(EdgeFreq, JumpInstCost);
2545 HeaderFallThroughCost = std::max(a: HeaderFallThroughCost, b: FallThruCost);
2546 }
2547 }
2548
2549 // Here we collect all exit blocks in the loop, and for each exit we find out
2550 // its hottest exit edge. For each loop rotation, we define the loop exit cost
2551 // as the sum of frequencies of exit edges we collect here, excluding the exit
2552 // edge from the tail of the loop chain.
2553 SmallVector<std::pair<MachineBasicBlock *, BlockFrequency>, 4> ExitsWithFreq;
2554 for (auto *BB : LoopChain) {
2555 auto LargestExitEdgeProb = BranchProbability::getZero();
2556 for (auto *Succ : BB->successors()) {
2557 BlockChain *SuccChain = BlockToChain[Succ];
2558 if (!LoopBlockSet.count(key: Succ) &&
2559 (!SuccChain || Succ == *SuccChain->begin())) {
2560 auto SuccProb = MBPI->getEdgeProbability(Src: BB, Dst: Succ);
2561 LargestExitEdgeProb = std::max(a: LargestExitEdgeProb, b: SuccProb);
2562 }
2563 }
2564 if (LargestExitEdgeProb > BranchProbability::getZero()) {
2565 auto ExitFreq = MBFI->getBlockFreq(MBB: BB) * LargestExitEdgeProb;
2566 ExitsWithFreq.emplace_back(Args&: BB, Args&: ExitFreq);
2567 }
2568 }
2569
2570 // In this loop we iterate every block in the loop chain and calculate the
2571 // cost assuming the block is the head of the loop chain. When the loop ends,
2572 // we should have found the best candidate as the loop chain's head.
2573 for (auto Iter = LoopChain.begin(), TailIter = std::prev(x: LoopChain.end()),
2574 EndIter = LoopChain.end();
2575 Iter != EndIter; Iter++, TailIter++) {
2576 // TailIter is used to track the tail of the loop chain if the block we are
2577 // checking (pointed by Iter) is the head of the chain.
2578 if (TailIter == LoopChain.end())
2579 TailIter = LoopChain.begin();
2580
2581 auto TailBB = *TailIter;
2582
2583 // Calculate the cost by putting this BB to the top.
2584 BlockFrequency Cost = BlockFrequency(0);
2585
2586 // If the current BB is the loop header, we need to take into account the
2587 // cost of the missed fall through edge from outside of the loop to the
2588 // header.
2589 if (Iter != LoopChain.begin())
2590 Cost += HeaderFallThroughCost;
2591
2592 // Collect the loop exit cost by summing up frequencies of all exit edges
2593 // except the one from the chain tail.
2594 for (auto &ExitWithFreq : ExitsWithFreq)
2595 if (TailBB != ExitWithFreq.first)
2596 Cost += ExitWithFreq.second;
2597
2598 // The cost of breaking the once fall-through edge from the tail to the top
2599 // of the loop chain. Here we need to consider three cases:
2600 // 1. If the tail node has only one successor, then we will get an
2601 // additional jmp instruction. So the cost here is (MisfetchCost +
2602 // JumpInstCost) * tail node frequency.
2603 // 2. If the tail node has two successors, then we may still get an
2604 // additional jmp instruction if the layout successor after the loop
2605 // chain is not its CFG successor. Note that the more frequently executed
2606 // jmp instruction will be put ahead of the other one. Assume the
2607 // frequency of those two branches are x and y, where x is the frequency
2608 // of the edge to the chain head, then the cost will be
2609 // (x * MisfetechCost + min(x, y) * JumpInstCost) * tail node frequency.
2610 // 3. If the tail node has more than two successors (this rarely happens),
2611 // we won't consider any additional cost.
2612 if (TailBB->isSuccessor(MBB: *Iter)) {
2613 auto TailBBFreq = MBFI->getBlockFreq(MBB: TailBB);
2614 if (TailBB->succ_size() == 1)
2615 Cost += ScaleBlockFrequency(TailBBFreq, MisfetchCost + JumpInstCost);
2616 else if (TailBB->succ_size() == 2) {
2617 auto TailToHeadProb = MBPI->getEdgeProbability(Src: TailBB, Dst: *Iter);
2618 auto TailToHeadFreq = TailBBFreq * TailToHeadProb;
2619 auto ColderEdgeFreq = TailToHeadProb > BranchProbability(1, 2)
2620 ? TailBBFreq * TailToHeadProb.getCompl()
2621 : TailToHeadFreq;
2622 Cost += ScaleBlockFrequency(TailToHeadFreq, MisfetchCost) +
2623 ScaleBlockFrequency(ColderEdgeFreq, JumpInstCost);
2624 }
2625 }
2626
2627 LLVM_DEBUG(dbgs() << "The cost of loop rotation by making "
2628 << getBlockName(*Iter) << " to the top: "
2629 << printBlockFreq(MBFI->getMBFI(), Cost) << "\n");
2630
2631 if (Cost < SmallestRotationCost) {
2632 SmallestRotationCost = Cost;
2633 RotationPos = Iter;
2634 }
2635 }
2636
2637 if (RotationPos != LoopChain.end()) {
2638 LLVM_DEBUG(dbgs() << "Rotate loop by making " << getBlockName(*RotationPos)
2639 << " to the top\n");
2640 std::rotate(first: LoopChain.begin(), middle: RotationPos, last: LoopChain.end());
2641 }
2642}
2643
2644/// Collect blocks in the given loop that are to be placed.
2645///
2646/// When profile data is available, exclude cold blocks from the returned set;
2647/// otherwise, collect all blocks in the loop.
2648MachineBlockPlacement::BlockFilterSet
2649MachineBlockPlacement::collectLoopBlockSet(const MachineLoop &L) {
2650 // Collect the blocks in a set ordered by block number, as this gives the same
2651 // order as they appear in the function.
2652 struct MBBCompare {
2653 bool operator()(const MachineBasicBlock *X,
2654 const MachineBasicBlock *Y) const {
2655 return X->getNumber() < Y->getNumber();
2656 }
2657 };
2658 std::set<const MachineBasicBlock *, MBBCompare> LoopBlockSet;
2659
2660 // Filter cold blocks off from LoopBlockSet when profile data is available.
2661 // Collect the sum of frequencies of incoming edges to the loop header from
2662 // outside. If we treat the loop as a super block, this is the frequency of
2663 // the loop. Then for each block in the loop, we calculate the ratio between
2664 // its frequency and the frequency of the loop block. When it is too small,
2665 // don't add it to the loop chain. If there are outer loops, then this block
2666 // will be merged into the first outer loop chain for which this block is not
2667 // cold anymore. This needs precise profile data and we only do this when
2668 // profile data is available.
2669 if (F->getFunction().hasProfileData() || ForceLoopColdBlock) {
2670 BlockFrequency LoopFreq(0);
2671 for (auto *LoopPred : L.getHeader()->predecessors())
2672 if (!L.contains(BB: LoopPred))
2673 LoopFreq += MBFI->getBlockFreq(MBB: LoopPred) *
2674 MBPI->getEdgeProbability(Src: LoopPred, Dst: L.getHeader());
2675
2676 for (MachineBasicBlock *LoopBB : L.getBlocks()) {
2677 if (LoopBlockSet.count(x: LoopBB))
2678 continue;
2679 auto Freq = MBFI->getBlockFreq(MBB: LoopBB).getFrequency();
2680 if (Freq == 0 || LoopFreq.getFrequency() / Freq > LoopToColdBlockRatio)
2681 continue;
2682 BlockChain *Chain = BlockToChain[LoopBB];
2683 for (MachineBasicBlock *ChainBB : *Chain)
2684 LoopBlockSet.insert(x: ChainBB);
2685 }
2686 } else
2687 LoopBlockSet.insert(first: L.block_begin(), last: L.block_end());
2688
2689 // Copy the blocks into a BlockFilterSet, as iterating it is faster than
2690 // std::set. We will only remove blocks and never insert them, which will
2691 // preserve the ordering.
2692 BlockFilterSet Ret(LoopBlockSet.begin(), LoopBlockSet.end());
2693 return Ret;
2694}
2695
2696/// Forms basic block chains from the natural loop structures.
2697///
2698/// These chains are designed to preserve the existing *structure* of the code
2699/// as much as possible. We can then stitch the chains together in a way which
2700/// both preserves the topological structure and minimizes taken conditional
2701/// branches.
2702void MachineBlockPlacement::buildLoopChains(const MachineLoop &L) {
2703 // First recurse through any nested loops, building chains for those inner
2704 // loops.
2705 for (const MachineLoop *InnerLoop : L)
2706 buildLoopChains(L: *InnerLoop);
2707
2708 assert(BlockWorkList.empty() &&
2709 "BlockWorkList not empty when starting to build loop chains.");
2710 assert(EHPadWorkList.empty() &&
2711 "EHPadWorkList not empty when starting to build loop chains.");
2712 BlockFilterSet LoopBlockSet = collectLoopBlockSet(L);
2713
2714 // Check if we have profile data for this function. If yes, we will rotate
2715 // this loop by modeling costs more precisely which requires the profile data
2716 // for better layout.
2717 bool RotateLoopWithProfile =
2718 ForcePreciseRotationCost ||
2719 (PreciseRotationCost && F->getFunction().hasProfileData());
2720
2721 // First check to see if there is an obviously preferable top block for the
2722 // loop. This will default to the header, but may end up as one of the
2723 // predecessors to the header if there is one which will result in strictly
2724 // fewer branches in the loop body.
2725 MachineBasicBlock *LoopTop = findBestLoopTop(L, LoopBlockSet);
2726
2727 // If we selected just the header for the loop top, look for a potentially
2728 // profitable exit block in the event that rotating the loop can eliminate
2729 // branches by placing an exit edge at the bottom.
2730 //
2731 // Loops are processed innermost to uttermost, make sure we clear
2732 // PreferredLoopExit before processing a new loop.
2733 PreferredLoopExit = nullptr;
2734 BlockFrequency ExitFreq;
2735 if (!RotateLoopWithProfile && LoopTop == L.getHeader())
2736 PreferredLoopExit = findBestLoopExit(L, LoopBlockSet, ExitFreq);
2737
2738 BlockChain &LoopChain = *BlockToChain[LoopTop];
2739
2740 // FIXME: This is a really lame way of walking the chains in the loop: we
2741 // walk the blocks, and use a set to prevent visiting a particular chain
2742 // twice.
2743 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
2744 assert(LoopChain.UnscheduledPredecessors == 0 &&
2745 "LoopChain should not have unscheduled predecessors.");
2746 UpdatedPreds.insert(Ptr: &LoopChain);
2747
2748 for (const MachineBasicBlock *LoopBB : LoopBlockSet)
2749 fillWorkLists(MBB: LoopBB, UpdatedPreds, BlockFilter: &LoopBlockSet);
2750
2751 buildChain(HeadBB: LoopTop, Chain&: LoopChain, BlockFilter: &LoopBlockSet);
2752
2753 if (RotateLoopWithProfile)
2754 rotateLoopWithProfile(LoopChain, L, LoopBlockSet);
2755 else
2756 rotateLoop(LoopChain, ExitingBB: PreferredLoopExit, ExitFreq, LoopBlockSet);
2757
2758 LLVM_DEBUG({
2759 // Crash at the end so we get all of the debugging output first.
2760 bool BadLoop = false;
2761 if (LoopChain.UnscheduledPredecessors) {
2762 BadLoop = true;
2763 dbgs() << "Loop chain contains a block without its preds placed!\n"
2764 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
2765 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n";
2766 }
2767 for (MachineBasicBlock *ChainBB : LoopChain) {
2768 dbgs() << " ... " << getBlockName(ChainBB) << "\n";
2769 if (!LoopBlockSet.remove(ChainBB)) {
2770 // We don't mark the loop as bad here because there are real situations
2771 // where this can occur. For example, with an unanalyzable fallthrough
2772 // from a loop block to a non-loop block or vice versa.
2773 dbgs() << "Loop chain contains a block not contained by the loop!\n"
2774 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
2775 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
2776 << " Bad block: " << getBlockName(ChainBB) << "\n";
2777 }
2778 }
2779
2780 if (!LoopBlockSet.empty()) {
2781 BadLoop = true;
2782 for (const MachineBasicBlock *LoopBB : LoopBlockSet)
2783 dbgs() << "Loop contains blocks never placed into a chain!\n"
2784 << " Loop header: " << getBlockName(*L.block_begin()) << "\n"
2785 << " Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
2786 << " Bad block: " << getBlockName(LoopBB) << "\n";
2787 }
2788 assert(!BadLoop && "Detected problems with the placement of this loop.");
2789 });
2790
2791 BlockWorkList.clear();
2792 EHPadWorkList.clear();
2793}
2794
2795void MachineBlockPlacement::buildCFGChains() {
2796 // Ensure that every BB in the function has an associated chain to simplify
2797 // the assumptions of the remaining algorithm.
2798 SmallVector<MachineOperand, 4> Cond; // For analyzeBranch.
2799 for (MachineFunction::iterator FI = F->begin(), FE = F->end(); FI != FE;
2800 ++FI) {
2801 MachineBasicBlock *BB = &*FI;
2802 BlockChain *Chain =
2803 new (ChainAllocator.Allocate()) BlockChain(BlockToChain, BB);
2804 // Also, merge any blocks which we cannot reason about and must preserve
2805 // the exact fallthrough behavior for.
2806 while (true) {
2807 Cond.clear();
2808 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For analyzeBranch.
2809 if (!TII->analyzeBranch(MBB&: *BB, TBB, FBB, Cond) || !FI->canFallThrough())
2810 break;
2811
2812 MachineFunction::iterator NextFI = std::next(x: FI);
2813 MachineBasicBlock *NextBB = &*NextFI;
2814 // Ensure that the layout successor is a viable block, as we know that
2815 // fallthrough is a possibility.
2816 assert(NextFI != FE && "Can't fallthrough past the last block.");
2817 LLVM_DEBUG(dbgs() << "Pre-merging due to unanalyzable fallthrough: "
2818 << getBlockName(BB) << " -> " << getBlockName(NextBB)
2819 << "\n");
2820 Chain->merge(BB: NextBB, Chain: nullptr);
2821#ifndef NDEBUG
2822 BlocksWithUnanalyzableExits.insert(&*BB);
2823#endif
2824 FI = NextFI;
2825 BB = NextBB;
2826 }
2827 }
2828
2829 // Build any loop-based chains.
2830 PreferredLoopExit = nullptr;
2831 for (MachineLoop *L : *MLI)
2832 buildLoopChains(L: *L);
2833
2834 assert(BlockWorkList.empty() &&
2835 "BlockWorkList should be empty before building final chain.");
2836 assert(EHPadWorkList.empty() &&
2837 "EHPadWorkList should be empty before building final chain.");
2838
2839 SmallPtrSet<BlockChain *, 4> UpdatedPreds;
2840 for (MachineBasicBlock &MBB : *F)
2841 fillWorkLists(MBB: &MBB, UpdatedPreds);
2842
2843 BlockChain &FunctionChain = *BlockToChain[&F->front()];
2844 buildChain(HeadBB: &F->front(), Chain&: FunctionChain);
2845
2846#ifndef NDEBUG
2847 using FunctionBlockSetType = SmallPtrSet<MachineBasicBlock *, 16>;
2848#endif
2849 LLVM_DEBUG({
2850 // Crash at the end so we get all of the debugging output first.
2851 bool BadFunc = false;
2852 FunctionBlockSetType FunctionBlockSet;
2853 for (MachineBasicBlock &MBB : *F)
2854 FunctionBlockSet.insert(&MBB);
2855
2856 for (MachineBasicBlock *ChainBB : FunctionChain)
2857 if (!FunctionBlockSet.erase(ChainBB)) {
2858 BadFunc = true;
2859 dbgs() << "Function chain contains a block not in the function!\n"
2860 << " Bad block: " << getBlockName(ChainBB) << "\n";
2861 }
2862
2863 if (!FunctionBlockSet.empty()) {
2864 BadFunc = true;
2865 for (MachineBasicBlock *RemainingBB : FunctionBlockSet)
2866 dbgs() << "Function contains blocks never placed into a chain!\n"
2867 << " Bad block: " << getBlockName(RemainingBB) << "\n";
2868 }
2869 assert(!BadFunc && "Detected problems with the block placement.");
2870 });
2871
2872 // Remember original layout ordering, so we can update terminators after
2873 // reordering to point to the original layout successor.
2874 SmallVector<MachineBasicBlock *, 4> OriginalLayoutSuccessors(
2875 F->getNumBlockIDs());
2876 {
2877 MachineBasicBlock *LastMBB = nullptr;
2878 for (auto &MBB : *F) {
2879 if (LastMBB != nullptr)
2880 OriginalLayoutSuccessors[LastMBB->getNumber()] = &MBB;
2881 LastMBB = &MBB;
2882 }
2883 OriginalLayoutSuccessors[F->back().getNumber()] = nullptr;
2884 }
2885
2886 // Splice the blocks into place.
2887 MachineFunction::iterator InsertPos = F->begin();
2888 LLVM_DEBUG(dbgs() << "[MBP] Function: " << F->getName() << "\n");
2889 for (MachineBasicBlock *ChainBB : FunctionChain) {
2890 LLVM_DEBUG(dbgs() << (ChainBB == *FunctionChain.begin() ? "Placing chain "
2891 : " ... ")
2892 << getBlockName(ChainBB) << "\n");
2893 if (InsertPos != MachineFunction::iterator(ChainBB))
2894 F->splice(InsertPt: InsertPos, MBB: ChainBB);
2895 else
2896 ++InsertPos;
2897
2898 // Update the terminator of the previous block.
2899 if (ChainBB == *FunctionChain.begin())
2900 continue;
2901 MachineBasicBlock *PrevBB = &*std::prev(x: MachineFunction::iterator(ChainBB));
2902
2903 // FIXME: It would be awesome of updateTerminator would just return rather
2904 // than assert when the branch cannot be analyzed in order to remove this
2905 // boiler plate.
2906 Cond.clear();
2907 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For analyzeBranch.
2908
2909#ifndef NDEBUG
2910 if (!BlocksWithUnanalyzableExits.count(PrevBB)) {
2911 // Given the exact block placement we chose, we may actually not _need_ to
2912 // be able to edit PrevBB's terminator sequence, but not being _able_ to
2913 // do that at this point is a bug.
2914 assert((!TII->analyzeBranch(*PrevBB, TBB, FBB, Cond) ||
2915 !PrevBB->canFallThrough()) &&
2916 "Unexpected block with un-analyzable fallthrough!");
2917 Cond.clear();
2918 TBB = FBB = nullptr;
2919 }
2920#endif
2921
2922 // The "PrevBB" is not yet updated to reflect current code layout, so,
2923 // o. it may fall-through to a block without explicit "goto" instruction
2924 // before layout, and no longer fall-through it after layout; or
2925 // o. just opposite.
2926 //
2927 // analyzeBranch() may return erroneous value for FBB when these two
2928 // situations take place. For the first scenario FBB is mistakenly set NULL;
2929 // for the 2nd scenario, the FBB, which is expected to be NULL, is
2930 // mistakenly pointing to "*BI".
2931 // Thus, if the future change needs to use FBB before the layout is set, it
2932 // has to correct FBB first by using the code similar to the following:
2933 //
2934 // if (!Cond.empty() && (!FBB || FBB == ChainBB)) {
2935 // PrevBB->updateTerminator();
2936 // Cond.clear();
2937 // TBB = FBB = nullptr;
2938 // if (TII->analyzeBranch(*PrevBB, TBB, FBB, Cond)) {
2939 // // FIXME: This should never take place.
2940 // TBB = FBB = nullptr;
2941 // }
2942 // }
2943 if (!TII->analyzeBranch(MBB&: *PrevBB, TBB, FBB, Cond)) {
2944 PrevBB->updateTerminator(PreviousLayoutSuccessor: OriginalLayoutSuccessors[PrevBB->getNumber()]);
2945 }
2946 }
2947
2948 // Fixup the last block.
2949 Cond.clear();
2950 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For analyzeBranch.
2951 if (!TII->analyzeBranch(MBB&: F->back(), TBB, FBB, Cond)) {
2952 MachineBasicBlock *PrevBB = &F->back();
2953 PrevBB->updateTerminator(PreviousLayoutSuccessor: OriginalLayoutSuccessors[PrevBB->getNumber()]);
2954 }
2955
2956 BlockWorkList.clear();
2957 EHPadWorkList.clear();
2958}
2959
2960void MachineBlockPlacement::optimizeBranches() {
2961 BlockChain &FunctionChain = *BlockToChain[&F->front()];
2962 SmallVector<MachineOperand, 4> Cond;
2963
2964 // Now that all the basic blocks in the chain have the proper layout,
2965 // make a final call to analyzeBranch with AllowModify set.
2966 // Indeed, the target may be able to optimize the branches in a way we
2967 // cannot because all branches may not be analyzable.
2968 // E.g., the target may be able to remove an unconditional branch to
2969 // a fallthrough when it occurs after predicated terminators.
2970 for (MachineBasicBlock *ChainBB : FunctionChain) {
2971 Cond.clear();
2972 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
2973 if (TII->analyzeBranch(MBB&: *ChainBB, TBB, FBB, Cond, /*AllowModify*/ true))
2974 continue;
2975 if (!TBB || !FBB || Cond.empty())
2976 continue;
2977 // If we are optimizing for size we do not consider the runtime performance.
2978 // Instead, we retain the original branch condition so we have more uniform
2979 // instructions which will benefit ICF.
2980 if (llvm::shouldOptimizeForSize(MBB: ChainBB, PSI, MBFIWrapper: MBFI.get()))
2981 continue;
2982 // If ChainBB has a two-way branch, try to re-order the branches
2983 // such that we branch to the successor with higher probability first.
2984 if (MBPI->getEdgeProbability(Src: ChainBB, Dst: TBB) >=
2985 MBPI->getEdgeProbability(Src: ChainBB, Dst: FBB))
2986 continue;
2987 if (TII->reverseBranchCondition(Cond))
2988 continue;
2989 LLVM_DEBUG(dbgs() << "Reverse order of the two branches: "
2990 << getBlockName(ChainBB) << "\n");
2991 LLVM_DEBUG(dbgs() << " " << getBlockName(TBB) << " < " << getBlockName(FBB)
2992 << "\n");
2993 auto Dl = ChainBB->findBranchDebugLoc();
2994 TII->removeBranch(MBB&: *ChainBB);
2995 TII->insertBranch(MBB&: *ChainBB, TBB: FBB, FBB: TBB, Cond, DL: Dl);
2996 }
2997}
2998
2999void MachineBlockPlacement::alignBlocks() {
3000 // Walk through the backedges of the function now that we have fully laid out
3001 // the basic blocks and align the destination of each backedge. We don't rely
3002 // exclusively on the loop info here so that we can align backedges in
3003 // unnatural CFGs and backedges that were introduced purely because of the
3004 // loop rotations done during this layout pass.
3005 if (!AlignAllBlock && !AlignAllNonFallThruBlocks) {
3006 if (F->getFunction().hasMinSize() ||
3007 (F->getFunction().hasOptSize() && !TLI->alignLoopsWithOptSize()))
3008 return;
3009 }
3010
3011 BlockChain &FunctionChain = *BlockToChain[&F->front()];
3012 // Empty chain.
3013 if (FunctionChain.begin() == FunctionChain.end())
3014 return;
3015
3016 const BranchProbability ColdProb(1, 5); // 20%
3017 BlockFrequency EntryFreq = MBFI->getBlockFreq(MBB: &F->front());
3018 BlockFrequency WeightedEntryFreq = EntryFreq * ColdProb;
3019 for (MachineBasicBlock *ChainBB : FunctionChain) {
3020 if (ChainBB == *FunctionChain.begin())
3021 continue;
3022
3023 // Don't align non-looping basic blocks. These are unlikely to execute
3024 // enough times to matter in practice. Note that we'll still handle
3025 // unnatural CFGs inside of a natural outer loop (the common case) and
3026 // rotated loops.
3027 MachineLoop *L = MLI->getLoopFor(BB: ChainBB);
3028 if (!L)
3029 continue;
3030
3031 const Align TLIAlign = TLI->getPrefLoopAlignment(ML: L);
3032 unsigned MDAlign = 1;
3033 MDNode *LoopID = L->getLoopID();
3034 if (LoopID) {
3035 for (const MDOperand &MDO : llvm::drop_begin(RangeOrContainer: LoopID->operands())) {
3036 MDNode *MD = dyn_cast<MDNode>(Val: MDO);
3037 if (MD == nullptr)
3038 continue;
3039 MDString *S = dyn_cast<MDString>(Val: MD->getOperand(I: 0));
3040 if (S == nullptr)
3041 continue;
3042 if (S->getString() == "llvm.loop.align") {
3043 assert(MD->getNumOperands() == 2 &&
3044 "per-loop align metadata should have two operands.");
3045 MDAlign =
3046 mdconst::extract<ConstantInt>(MD: MD->getOperand(I: 1))->getZExtValue();
3047 assert(MDAlign >= 1 && "per-loop align value must be positive.");
3048 }
3049 }
3050 }
3051
3052 // Use max of the TLIAlign and MDAlign
3053 const Align LoopAlign = std::max(a: TLIAlign, b: Align(MDAlign));
3054 if (LoopAlign == 1)
3055 continue; // Don't care about loop alignment.
3056
3057 // If the block is cold relative to the function entry don't waste space
3058 // aligning it.
3059 BlockFrequency Freq = MBFI->getBlockFreq(MBB: ChainBB);
3060 if (Freq < WeightedEntryFreq)
3061 continue;
3062
3063 // If the block is cold relative to its loop header, don't align it
3064 // regardless of what edges into the block exist.
3065 MachineBasicBlock *LoopHeader = L->getHeader();
3066 BlockFrequency LoopHeaderFreq = MBFI->getBlockFreq(MBB: LoopHeader);
3067 if (Freq < (LoopHeaderFreq * ColdProb))
3068 continue;
3069
3070 // If the global profiles indicates so, don't align it.
3071 if (llvm::shouldOptimizeForSize(MBB: ChainBB, PSI, MBFIWrapper: MBFI.get()) &&
3072 !TLI->alignLoopsWithOptSize())
3073 continue;
3074
3075 // Check for the existence of a non-layout predecessor which would benefit
3076 // from aligning this block.
3077 MachineBasicBlock *LayoutPred =
3078 &*std::prev(x: MachineFunction::iterator(ChainBB));
3079
3080 auto DetermineMaxAlignmentPadding = [&]() {
3081 // Set the maximum bytes allowed to be emitted for alignment.
3082 unsigned MaxBytes;
3083 if (MaxBytesForAlignmentOverride.getNumOccurrences() > 0)
3084 MaxBytes = MaxBytesForAlignmentOverride;
3085 else
3086 MaxBytes = TLI->getMaxPermittedBytesForAlignment(MBB: ChainBB);
3087 ChainBB->setMaxBytesForAlignment(MaxBytes);
3088 };
3089
3090 // Force alignment if all the predecessors are jumps. We already checked
3091 // that the block isn't cold above.
3092 if (!LayoutPred->isSuccessor(MBB: ChainBB)) {
3093 ChainBB->setAlignment(LoopAlign);
3094 DetermineMaxAlignmentPadding();
3095 continue;
3096 }
3097
3098 // Align this block if the layout predecessor's edge into this block is
3099 // cold relative to the block. When this is true, other predecessors make up
3100 // all of the hot entries into the block and thus alignment is likely to be
3101 // important.
3102 BranchProbability LayoutProb =
3103 MBPI->getEdgeProbability(Src: LayoutPred, Dst: ChainBB);
3104 BlockFrequency LayoutEdgeFreq = MBFI->getBlockFreq(MBB: LayoutPred) * LayoutProb;
3105 if (LayoutEdgeFreq <= (Freq * ColdProb)) {
3106 ChainBB->setAlignment(LoopAlign);
3107 DetermineMaxAlignmentPadding();
3108 }
3109 }
3110
3111 const bool HasMaxBytesOverride =
3112 MaxBytesForAlignmentOverride.getNumOccurrences() > 0;
3113
3114 if (AlignAllBlock)
3115 // Align all of the blocks in the function to a specific alignment.
3116 for (MachineBasicBlock &MBB : *F) {
3117 if (HasMaxBytesOverride)
3118 MBB.setAlignment(A: Align(1ULL << AlignAllBlock),
3119 MaxBytes: MaxBytesForAlignmentOverride);
3120 else
3121 MBB.setAlignment(Align(1ULL << AlignAllBlock));
3122 }
3123 else if (AlignAllNonFallThruBlocks) {
3124 // Align all of the blocks that have no fall-through predecessors to a
3125 // specific alignment.
3126 for (auto MBI = std::next(x: F->begin()), MBE = F->end(); MBI != MBE; ++MBI) {
3127 auto LayoutPred = std::prev(x: MBI);
3128 if (!LayoutPred->isSuccessor(MBB: &*MBI)) {
3129 if (HasMaxBytesOverride)
3130 MBI->setAlignment(A: Align(1ULL << AlignAllNonFallThruBlocks),
3131 MaxBytes: MaxBytesForAlignmentOverride);
3132 else
3133 MBI->setAlignment(Align(1ULL << AlignAllNonFallThruBlocks));
3134 }
3135 }
3136 }
3137}
3138
3139/// Tail duplicate \p BB into (some) predecessors if profitable, repeating if
3140/// it was duplicated into its chain predecessor and removed.
3141/// \p BB - Basic block that may be duplicated.
3142///
3143/// \p LPred - Chosen layout predecessor of \p BB.
3144/// Updated to be the chain end if LPred is removed.
3145/// \p Chain - Chain to which \p LPred belongs, and \p BB will belong.
3146/// \p BlockFilter - Set of blocks that belong to the loop being laid out.
3147/// Used to identify which blocks to update predecessor
3148/// counts.
3149/// \p PrevUnplacedBlockIt - Iterator pointing to the last block that was
3150/// chosen in the given order due to unnatural CFG
3151/// only needed if \p BB is removed and
3152/// \p PrevUnplacedBlockIt pointed to \p BB.
3153/// @return true if \p BB was removed.
3154bool MachineBlockPlacement::repeatedlyTailDuplicateBlock(
3155 MachineBasicBlock *BB, MachineBasicBlock *&LPred,
3156 const MachineBasicBlock *LoopHeaderBB, BlockChain &Chain,
3157 BlockFilterSet *BlockFilter, MachineFunction::iterator &PrevUnplacedBlockIt,
3158 BlockFilterSet::iterator &PrevUnplacedBlockInFilterIt) {
3159 bool Removed, DuplicatedToLPred;
3160 bool DuplicatedToOriginalLPred;
3161 Removed = maybeTailDuplicateBlock(
3162 BB, LPred, Chain, BlockFilter, PrevUnplacedBlockIt,
3163 PrevUnplacedBlockInFilterIt, DuplicatedToLPred);
3164 if (!Removed)
3165 return false;
3166 DuplicatedToOriginalLPred = DuplicatedToLPred;
3167 // Iteratively try to duplicate again. It can happen that a block that is
3168 // duplicated into is still small enough to be duplicated again.
3169 // No need to call markBlockSuccessors in this case, as the blocks being
3170 // duplicated from here on are already scheduled.
3171 while (DuplicatedToLPred && Removed) {
3172 MachineBasicBlock *DupBB, *DupPred;
3173 // The removal callback causes Chain.end() to be updated when a block is
3174 // removed. On the first pass through the loop, the chain end should be the
3175 // same as it was on function entry. On subsequent passes, because we are
3176 // duplicating the block at the end of the chain, if it is removed the
3177 // chain will have shrunk by one block.
3178 BlockChain::iterator ChainEnd = Chain.end();
3179 DupBB = *(--ChainEnd);
3180 // Now try to duplicate again.
3181 if (ChainEnd == Chain.begin())
3182 break;
3183 DupPred = *std::prev(x: ChainEnd);
3184 Removed = maybeTailDuplicateBlock(
3185 BB: DupBB, LPred: DupPred, Chain, BlockFilter, PrevUnplacedBlockIt,
3186 PrevUnplacedBlockInFilterIt, DuplicatedToLPred);
3187 }
3188 // If BB was duplicated into LPred, it is now scheduled. But because it was
3189 // removed, markChainSuccessors won't be called for its chain. Instead we
3190 // call markBlockSuccessors for LPred to achieve the same effect. This must go
3191 // at the end because repeating the tail duplication can increase the number
3192 // of unscheduled predecessors.
3193 LPred = *std::prev(x: Chain.end());
3194 if (DuplicatedToOriginalLPred)
3195 markBlockSuccessors(Chain, MBB: LPred, LoopHeaderBB, BlockFilter);
3196 return true;
3197}
3198
3199/// Tail duplicate \p BB into (some) predecessors if profitable.
3200/// \p BB - Basic block that may be duplicated
3201/// \p LPred - Chosen layout predecessor of \p BB
3202/// \p Chain - Chain to which \p LPred belongs, and \p BB will belong.
3203/// \p BlockFilter - Set of blocks that belong to the loop being laid out.
3204/// Used to identify which blocks to update predecessor
3205/// counts.
3206/// \p PrevUnplacedBlockIt - Iterator pointing to the last block that was
3207/// chosen in the given order due to unnatural CFG
3208/// only needed if \p BB is removed and
3209/// \p PrevUnplacedBlockIt pointed to \p BB.
3210/// \p DuplicatedToLPred - True if the block was duplicated into LPred.
3211/// \return - True if the block was duplicated into all preds and removed.
3212bool MachineBlockPlacement::maybeTailDuplicateBlock(
3213 MachineBasicBlock *BB, MachineBasicBlock *LPred, BlockChain &Chain,
3214 BlockFilterSet *BlockFilter, MachineFunction::iterator &PrevUnplacedBlockIt,
3215 BlockFilterSet::iterator &PrevUnplacedBlockInFilterIt,
3216 bool &DuplicatedToLPred) {
3217 DuplicatedToLPred = false;
3218 if (!shouldTailDuplicate(BB))
3219 return false;
3220
3221 LLVM_DEBUG(dbgs() << "Redoing tail duplication for Succ#" << BB->getNumber()
3222 << "\n");
3223
3224 // This has to be a callback because none of it can be done after
3225 // BB is deleted.
3226 bool Removed = false;
3227 auto RemovalCallback = [&](MachineBasicBlock *RemBB) {
3228 // Signal to outer function
3229 Removed = true;
3230
3231 // Remove from the Chain and Chain Map
3232 if (auto It = BlockToChain.find(Val: RemBB); It != BlockToChain.end()) {
3233 It->second->remove(BB: RemBB);
3234 BlockToChain.erase(I: It);
3235 }
3236
3237 // Handle the unplaced block iterator
3238 if (&(*PrevUnplacedBlockIt) == RemBB) {
3239 PrevUnplacedBlockIt++;
3240 }
3241
3242 // Handle the Work Lists
3243 if (RemBB->isEHPad()) {
3244 llvm::erase(C&: EHPadWorkList, V: RemBB);
3245 } else {
3246 llvm::erase(C&: BlockWorkList, V: RemBB);
3247 }
3248
3249 // Handle the filter set
3250 if (BlockFilter) {
3251 auto It = llvm::find(Range&: *BlockFilter, Val: RemBB);
3252 // Erase RemBB from BlockFilter, and keep PrevUnplacedBlockInFilterIt
3253 // pointing to the same element as before.
3254 if (It != BlockFilter->end()) {
3255 if (It < PrevUnplacedBlockInFilterIt) {
3256 const MachineBasicBlock *PrevBB = *PrevUnplacedBlockInFilterIt;
3257 // BlockFilter is a SmallVector so all elements after RemBB are
3258 // shifted to the front by 1 after its deletion.
3259 auto Distance = PrevUnplacedBlockInFilterIt - It - 1;
3260 PrevUnplacedBlockInFilterIt = BlockFilter->erase(I: It) + Distance;
3261 assert(*PrevUnplacedBlockInFilterIt == PrevBB);
3262 (void)PrevBB;
3263 } else if (It == PrevUnplacedBlockInFilterIt)
3264 // The block pointed by PrevUnplacedBlockInFilterIt is erased, we
3265 // have to set it to the next element.
3266 PrevUnplacedBlockInFilterIt = BlockFilter->erase(I: It);
3267 else
3268 BlockFilter->erase(I: It);
3269 }
3270 }
3271
3272 // Remove the block from loop info.
3273 MLI->removeBlock(BB: RemBB);
3274 if (RemBB == PreferredLoopExit)
3275 PreferredLoopExit = nullptr;
3276
3277 LLVM_DEBUG(dbgs() << "TailDuplicator deleted block: " << getBlockName(RemBB)
3278 << "\n");
3279 };
3280 auto RemovalCallbackRef =
3281 function_ref<void(MachineBasicBlock *)>(RemovalCallback);
3282
3283 SmallVector<MachineBasicBlock *, 8> DuplicatedPreds;
3284 bool IsSimple = TailDup.isSimpleBB(TailBB: BB);
3285 SmallVector<MachineBasicBlock *, 8> CandidatePreds;
3286 SmallVectorImpl<MachineBasicBlock *> *CandidatePtr = nullptr;
3287 if (F->getFunction().hasProfileData()) {
3288 // We can do partial duplication with precise profile information.
3289 findDuplicateCandidates(Candidates&: CandidatePreds, BB, BlockFilter);
3290 if (CandidatePreds.size() == 0)
3291 return false;
3292 if (CandidatePreds.size() < BB->pred_size())
3293 CandidatePtr = &CandidatePreds;
3294 }
3295 TailDup.tailDuplicateAndUpdate(IsSimple, MBB: BB, ForcedLayoutPred: LPred, DuplicatedPreds: &DuplicatedPreds,
3296 RemovalCallback: &RemovalCallbackRef, CandidatePtr);
3297
3298 // Update UnscheduledPredecessors to reflect tail-duplication.
3299 DuplicatedToLPred = false;
3300 for (MachineBasicBlock *Pred : DuplicatedPreds) {
3301 // We're only looking for unscheduled predecessors that match the filter.
3302 BlockChain *PredChain = BlockToChain[Pred];
3303 if (Pred == LPred)
3304 DuplicatedToLPred = true;
3305 if (Pred == LPred || (BlockFilter && !BlockFilter->count(key: Pred)) ||
3306 PredChain == &Chain)
3307 continue;
3308 for (MachineBasicBlock *NewSucc : Pred->successors()) {
3309 if (BlockFilter && !BlockFilter->count(key: NewSucc))
3310 continue;
3311 BlockChain *NewChain = BlockToChain[NewSucc];
3312 if (NewChain != &Chain && NewChain != PredChain)
3313 NewChain->UnscheduledPredecessors++;
3314 }
3315 }
3316 return Removed;
3317}
3318
3319// Count the number of actual machine instructions.
3320static uint64_t countMBBInstruction(MachineBasicBlock *MBB) {
3321 uint64_t InstrCount = 0;
3322 for (MachineInstr &MI : *MBB) {
3323 if (!MI.isPHI() && !MI.isMetaInstruction())
3324 InstrCount += 1;
3325 }
3326 return InstrCount;
3327}
3328
3329// The size cost of duplication is the instruction size of the duplicated block.
3330// So we should scale the threshold accordingly. But the instruction size is not
3331// available on all targets, so we use the number of instructions instead.
3332BlockFrequency MachineBlockPlacement::scaleThreshold(MachineBasicBlock *BB) {
3333 return BlockFrequency(DupThreshold.getFrequency() * countMBBInstruction(MBB: BB));
3334}
3335
3336// Returns true if BB is Pred's best successor.
3337bool MachineBlockPlacement::isBestSuccessor(MachineBasicBlock *BB,
3338 MachineBasicBlock *Pred,
3339 BlockFilterSet *BlockFilter) {
3340 if (BB == Pred)
3341 return false;
3342 if (BlockFilter && !BlockFilter->count(key: Pred))
3343 return false;
3344 BlockChain *PredChain = BlockToChain[Pred];
3345 if (PredChain && (Pred != *std::prev(x: PredChain->end())))
3346 return false;
3347
3348 // Find the successor with largest probability excluding BB.
3349 BranchProbability BestProb = BranchProbability::getZero();
3350 for (MachineBasicBlock *Succ : Pred->successors())
3351 if (Succ != BB) {
3352 if (BlockFilter && !BlockFilter->count(key: Succ))
3353 continue;
3354 BlockChain *SuccChain = BlockToChain[Succ];
3355 if (SuccChain && (Succ != *SuccChain->begin()))
3356 continue;
3357 BranchProbability SuccProb = MBPI->getEdgeProbability(Src: Pred, Dst: Succ);
3358 if (SuccProb > BestProb)
3359 BestProb = SuccProb;
3360 }
3361
3362 BranchProbability BBProb = MBPI->getEdgeProbability(Src: Pred, Dst: BB);
3363 if (BBProb <= BestProb)
3364 return false;
3365
3366 // Compute the number of reduced taken branches if Pred falls through to BB
3367 // instead of another successor. Then compare it with threshold.
3368 BlockFrequency PredFreq = getBlockCountOrFrequency(BB: Pred);
3369 BlockFrequency Gain = PredFreq * (BBProb - BestProb);
3370 return Gain > scaleThreshold(BB);
3371}
3372
3373// Find out the predecessors of BB and BB can be beneficially duplicated into
3374// them.
3375void MachineBlockPlacement::findDuplicateCandidates(
3376 SmallVectorImpl<MachineBasicBlock *> &Candidates, MachineBasicBlock *BB,
3377 BlockFilterSet *BlockFilter) {
3378 MachineBasicBlock *Fallthrough = nullptr;
3379 BranchProbability DefaultBranchProb = BranchProbability::getZero();
3380 BlockFrequency BBDupThreshold(scaleThreshold(BB));
3381 SmallVector<MachineBasicBlock *, 8> Preds(BB->predecessors());
3382 SmallVector<MachineBasicBlock *, 8> Succs(BB->successors());
3383
3384 // Sort for highest frequency.
3385 auto CmpSucc = [&](MachineBasicBlock *A, MachineBasicBlock *B) {
3386 return MBPI->getEdgeProbability(Src: BB, Dst: A) > MBPI->getEdgeProbability(Src: BB, Dst: B);
3387 };
3388 auto CmpPred = [&](MachineBasicBlock *A, MachineBasicBlock *B) {
3389 return MBFI->getBlockFreq(MBB: A) > MBFI->getBlockFreq(MBB: B);
3390 };
3391 llvm::stable_sort(Range&: Succs, C: CmpSucc);
3392 llvm::stable_sort(Range&: Preds, C: CmpPred);
3393
3394 auto SuccIt = Succs.begin();
3395 if (SuccIt != Succs.end()) {
3396 DefaultBranchProb = MBPI->getEdgeProbability(Src: BB, Dst: *SuccIt).getCompl();
3397 }
3398
3399 // For each predecessors of BB, compute the benefit of duplicating BB,
3400 // if it is larger than the threshold, add it into Candidates.
3401 //
3402 // If we have following control flow.
3403 //
3404 // PB1 PB2 PB3 PB4
3405 // \ | / /\
3406 // \ | / / \
3407 // \ |/ / \
3408 // BB----/ OB
3409 // /\
3410 // / \
3411 // SB1 SB2
3412 //
3413 // And it can be partially duplicated as
3414 //
3415 // PB2+BB
3416 // | PB1 PB3 PB4
3417 // | | / /\
3418 // | | / / \
3419 // | |/ / \
3420 // | BB----/ OB
3421 // |\ /|
3422 // | X |
3423 // |/ \|
3424 // SB2 SB1
3425 //
3426 // The benefit of duplicating into a predecessor is defined as
3427 // Orig_taken_branch - Duplicated_taken_branch
3428 //
3429 // The Orig_taken_branch is computed with the assumption that predecessor
3430 // jumps to BB and the most possible successor is laid out after BB.
3431 //
3432 // The Duplicated_taken_branch is computed with the assumption that BB is
3433 // duplicated into PB, and one successor is layout after it (SB1 for PB1 and
3434 // SB2 for PB2 in our case). If there is no available successor, the combined
3435 // block jumps to all BB's successor, like PB3 in this example.
3436 //
3437 // If a predecessor has multiple successors, so BB can't be duplicated into
3438 // it. But it can beneficially fall through to BB, and duplicate BB into other
3439 // predecessors.
3440 for (MachineBasicBlock *Pred : Preds) {
3441 BlockFrequency PredFreq = getBlockCountOrFrequency(BB: Pred);
3442
3443 if (!TailDup.canTailDuplicate(TailBB: BB, PredBB: Pred)) {
3444 // BB can't be duplicated into Pred, but it is possible to be layout
3445 // below Pred.
3446 if (!Fallthrough && isBestSuccessor(BB, Pred, BlockFilter)) {
3447 Fallthrough = Pred;
3448 if (SuccIt != Succs.end())
3449 SuccIt++;
3450 }
3451 continue;
3452 }
3453
3454 BlockFrequency OrigCost = PredFreq + PredFreq * DefaultBranchProb;
3455 BlockFrequency DupCost;
3456 if (SuccIt == Succs.end()) {
3457 // Jump to all successors;
3458 if (Succs.size() > 0)
3459 DupCost += PredFreq;
3460 } else {
3461 // Fallthrough to *SuccIt, jump to all other successors;
3462 DupCost += PredFreq;
3463 DupCost -= PredFreq * MBPI->getEdgeProbability(Src: BB, Dst: *SuccIt);
3464 }
3465
3466 assert(OrigCost >= DupCost);
3467 OrigCost -= DupCost;
3468 if (OrigCost > BBDupThreshold) {
3469 Candidates.push_back(Elt: Pred);
3470 if (SuccIt != Succs.end())
3471 SuccIt++;
3472 }
3473 }
3474
3475 // No predecessors can optimally fallthrough to BB.
3476 // So we can change one duplication into fallthrough.
3477 if (!Fallthrough) {
3478 if ((Candidates.size() < Preds.size()) && (Candidates.size() > 0)) {
3479 Candidates[0] = Candidates.back();
3480 Candidates.pop_back();
3481 }
3482 }
3483}
3484
3485void MachineBlockPlacement::initTailDupThreshold() {
3486 DupThreshold = BlockFrequency(0);
3487 if (F->getFunction().hasProfileData()) {
3488 // We prefer to use prifile count.
3489 uint64_t HotThreshold = PSI->getOrCompHotCountThreshold();
3490 if (HotThreshold != UINT64_MAX) {
3491 UseProfileCount = true;
3492 DupThreshold =
3493 BlockFrequency(HotThreshold * TailDupProfilePercentThreshold / 100);
3494 } else {
3495 // Profile count is not available, we can use block frequency instead.
3496 BlockFrequency MaxFreq = BlockFrequency(0);
3497 for (MachineBasicBlock &MBB : *F) {
3498 BlockFrequency Freq = MBFI->getBlockFreq(MBB: &MBB);
3499 if (Freq > MaxFreq)
3500 MaxFreq = Freq;
3501 }
3502
3503 BranchProbability ThresholdProb(TailDupPlacementPenalty, 100);
3504 DupThreshold = BlockFrequency(MaxFreq * ThresholdProb);
3505 UseProfileCount = false;
3506 }
3507 }
3508
3509 TailDupSize = TailDupPlacementThreshold;
3510 // If only the aggressive threshold is explicitly set, use it.
3511 if (TailDupPlacementAggressiveThreshold.getNumOccurrences() != 0 &&
3512 TailDupPlacementThreshold.getNumOccurrences() == 0)
3513 TailDupSize = TailDupPlacementAggressiveThreshold;
3514
3515 // For aggressive optimization, we can adjust some thresholds to be less
3516 // conservative.
3517 if (OptLevel >= CodeGenOptLevel::Aggressive) {
3518 // At O3 we should be more willing to copy blocks for tail duplication. This
3519 // increases size pressure, so we only do it at O3
3520 // Do this unless only the regular threshold is explicitly set.
3521 if (TailDupPlacementThreshold.getNumOccurrences() == 0 ||
3522 TailDupPlacementAggressiveThreshold.getNumOccurrences() != 0)
3523 TailDupSize = TailDupPlacementAggressiveThreshold;
3524 }
3525
3526 // If there's no threshold provided through options, query the target
3527 // information for a threshold instead.
3528 if (TailDupPlacementThreshold.getNumOccurrences() == 0 &&
3529 (OptLevel < CodeGenOptLevel::Aggressive ||
3530 TailDupPlacementAggressiveThreshold.getNumOccurrences() == 0))
3531 TailDupSize = TII->getTailDuplicateSize(OptLevel);
3532}
3533
3534PreservedAnalyses
3535MachineBlockPlacementPass::run(MachineFunction &MF,
3536 MachineFunctionAnalysisManager &MFAM) {
3537 auto *MBPI = &MFAM.getResult<MachineBranchProbabilityAnalysis>(IR&: MF);
3538 auto MBFI = std::make_unique<MBFIWrapper>(
3539 args&: MFAM.getResult<MachineBlockFrequencyAnalysis>(IR&: MF));
3540 auto *MLI = &MFAM.getResult<MachineLoopAnalysis>(IR&: MF);
3541 auto *MPDT = MachineBlockPlacement::allowTailDupPlacement(MF)
3542 ? &MFAM.getResult<MachinePostDominatorTreeAnalysis>(IR&: MF)
3543 : nullptr;
3544 auto *PSI = MFAM.getResult<ModuleAnalysisManagerMachineFunctionProxy>(IR&: MF)
3545 .getCachedResult<ProfileSummaryAnalysis>(
3546 IR&: *MF.getFunction().getParent());
3547 if (!PSI)
3548 report_fatal_error(reason: "MachineBlockPlacement requires ProfileSummaryAnalysis",
3549 gen_crash_diag: false);
3550
3551 MachineBlockPlacement MBP(MBPI, MLI, PSI, std::move(MBFI), MPDT,
3552 AllowTailMerge);
3553
3554 if (!MBP.run(F&: MF))
3555 return PreservedAnalyses::all();
3556
3557 return getMachineFunctionPassPreservedAnalyses();
3558}
3559
3560void MachineBlockPlacementPass::printPipeline(
3561 raw_ostream &OS,
3562 function_ref<StringRef(StringRef)> MapClassName2PassName) const {
3563 OS << MapClassName2PassName(name());
3564 if (!AllowTailMerge)
3565 OS << "<no-tail-merge>";
3566}
3567
3568bool MachineBlockPlacement::run(MachineFunction &MF) {
3569
3570 // Check for single-block functions and skip them.
3571 if (std::next(x: MF.begin()) == MF.end())
3572 return false;
3573
3574 F = &MF;
3575 OptLevel = F->getTarget().getOptLevel();
3576
3577 TII = MF.getSubtarget().getInstrInfo();
3578 TLI = MF.getSubtarget().getTargetLowering();
3579
3580 // Initialize PreferredLoopExit to nullptr here since it may never be set if
3581 // there are no MachineLoops.
3582 PreferredLoopExit = nullptr;
3583
3584 assert(BlockToChain.empty() &&
3585 "BlockToChain map should be empty before starting placement.");
3586 assert(ComputedEdges.empty() &&
3587 "Computed Edge map should be empty before starting placement.");
3588
3589 // Initialize tail duplication thresholds.
3590 initTailDupThreshold();
3591
3592 const bool OptForSize =
3593 llvm::shouldOptimizeForSize(MF: &MF, PSI, BFI: &MBFI->getMBFI());
3594 // Determine whether to use ext-tsp for perf/size optimization. The method
3595 // is beneficial only for instances with at least 3 basic blocks and it can be
3596 // disabled for huge functions (exceeding a certain size).
3597 bool UseExtTspForPerf = false;
3598 bool UseExtTspForSize = false;
3599 if (3 <= MF.size() && MF.size() <= ExtTspBlockPlacementMaxBlocks) {
3600 UseExtTspForPerf =
3601 EnableExtTspBlockPlacement &&
3602 (ApplyExtTspWithoutProfile || MF.getFunction().hasProfileData());
3603 UseExtTspForSize = OptForSize && ApplyExtTspForSize;
3604 }
3605
3606 // Apply tail duplication.
3607 if (allowTailDupPlacement(MF&: *F)) {
3608 if (OptForSize)
3609 TailDupSize = 1;
3610 const bool PreRegAlloc = false;
3611 TailDup.initMF(MF, PreRegAlloc, MBPI, MBFI: MBFI.get(), PSI,
3612 /* LayoutMode */ true, TailDupSize);
3613 if (!UseExtTspForSize)
3614 precomputeTriangleChains();
3615 }
3616
3617 // Run the main block placement.
3618 if (!UseExtTspForSize)
3619 buildCFGChains();
3620
3621 // Changing the layout can create new tail merging opportunities.
3622 // TailMerge can create jump into if branches that make CFG irreducible for
3623 // HW that requires structured CFG.
3624 const bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
3625 AllowTailMerge && BranchFoldPlacement &&
3626 MF.size() > 3;
3627 // No tail merging opportunities if the block number is less than four.
3628 if (EnableTailMerge) {
3629 const unsigned TailMergeSize = TailDupSize + 1;
3630 BranchFolder BF(/*DefaultEnableTailMerge=*/true, /*CommonHoist=*/false,
3631 *MBFI, *MBPI, PSI, TailMergeSize);
3632
3633 if (BF.OptimizeFunction(MF, tii: TII, tri: MF.getSubtarget().getRegisterInfo(), mli: MLI,
3634 /*AfterPlacement=*/true)) {
3635 // Must redo the post-dominator tree if blocks were changed.
3636 if (MPDT)
3637 MPDT->recalculate(Func&: MF);
3638 if (!UseExtTspForSize) {
3639 // Redo the layout if tail merging creates/removes/moves blocks.
3640 BlockToChain.clear();
3641 ComputedEdges.clear();
3642 ChainAllocator.DestroyAll();
3643 buildCFGChains();
3644 }
3645 }
3646 }
3647
3648 // Apply a post-processing optimizing block placement:
3649 // - find a new placement and modify the layout of the blocks in the function;
3650 // - re-create CFG chains so that we can optimizeBranches and alignBlocks.
3651 if (UseExtTspForPerf || UseExtTspForSize) {
3652 assert(
3653 !(UseExtTspForPerf && UseExtTspForSize) &&
3654 "UseExtTspForPerf and UseExtTspForSize can not be set simultaneously");
3655 applyExtTsp(/*OptForSize=*/UseExtTspForSize);
3656 createCFGChainExtTsp();
3657 }
3658
3659 optimizeBranches();
3660 alignBlocks();
3661
3662 BlockToChain.clear();
3663 ComputedEdges.clear();
3664 ChainAllocator.DestroyAll();
3665
3666 // View the function.
3667 if (ViewBlockLayoutWithBFI != GVDT_None &&
3668 (ViewBlockFreqFuncName.empty() ||
3669 F->getFunction().getName() == ViewBlockFreqFuncName)) {
3670 if (RenumberBlocksBeforeView)
3671 MF.RenumberBlocks();
3672 MBFI->view(Name: "MBP." + MF.getName(), isSimple: false);
3673 }
3674
3675 // We always return true as we have no way to track whether the final order
3676 // differs from the original order.
3677 return true;
3678}
3679
3680void MachineBlockPlacement::applyExtTsp(bool OptForSize) {
3681 // Prepare data; blocks are indexed by their index in the current ordering.
3682 DenseMap<const MachineBasicBlock *, uint64_t> BlockIndex;
3683 BlockIndex.reserve(NumEntries: F->size());
3684 std::vector<const MachineBasicBlock *> CurrentBlockOrder;
3685 CurrentBlockOrder.reserve(n: F->size());
3686 size_t NumBlocks = 0;
3687 for (const MachineBasicBlock &MBB : *F) {
3688 BlockIndex[&MBB] = NumBlocks++;
3689 CurrentBlockOrder.push_back(x: &MBB);
3690 }
3691
3692 SmallVector<uint64_t, 0> BlockCounts(F->size());
3693 SmallVector<uint64_t, 0> BlockSizes(F->size());
3694 SmallVector<codelayout::EdgeCount, 0> JumpCounts;
3695 SmallVector<MachineOperand, 4> Cond; // For analyzeBranch.
3696 SmallVector<const MachineBasicBlock *, 4> Succs;
3697 for (MachineBasicBlock &MBB : *F) {
3698 // Getting the block frequency.
3699 BlockFrequency BlockFreq = MBFI->getBlockFreq(MBB: &MBB);
3700 BlockCounts[BlockIndex[&MBB]] = OptForSize ? 1 : BlockFreq.getFrequency();
3701 // Getting the block size:
3702 // - approximate the size of an instruction by 4 bytes, and
3703 // - ignore debug instructions.
3704 // Note: getting the exact size of each block is target-dependent and can be
3705 // done by extending the interface of MCCodeEmitter. Experimentally we do
3706 // not see a perf improvement with the exact block sizes.
3707 auto NonDbgInsts =
3708 instructionsWithoutDebug(It: MBB.instr_begin(), End: MBB.instr_end());
3709 size_t NumInsts = std::distance(first: NonDbgInsts.begin(), last: NonDbgInsts.end());
3710 BlockSizes[BlockIndex[&MBB]] = 4 * NumInsts;
3711
3712 // Getting jump frequencies.
3713 if (OptForSize) {
3714 Cond.clear();
3715 MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For analyzeBranch.
3716 if (TII->analyzeBranch(MBB, TBB, FBB, Cond))
3717 continue;
3718
3719 const MachineBasicBlock *FTB = MBB.getFallThrough();
3720 // Succs is a collection of distinct destinations of the block reachable
3721 // from MBB via a jump instruction; initialize the list using the three
3722 // (non-necessarily distinct) blocks, FTB, TBB, and FBB.
3723 Succs.clear();
3724 if (TBB && TBB != FTB)
3725 Succs.push_back(Elt: TBB);
3726 if (FBB && FBB != FTB)
3727 Succs.push_back(Elt: FBB);
3728 if (FTB)
3729 Succs.push_back(Elt: FTB);
3730 // Absolute magnitude of non-zero counts does not matter for the
3731 // optimization; prioritize slightly jumps with a single successor, since
3732 // the corresponding jump instruction will be removed from the binary.
3733 const uint64_t Freq = Succs.size() == 1 ? 110 : 100;
3734 for (const MachineBasicBlock *Succ : Succs)
3735 JumpCounts.push_back(Elt: {.src: BlockIndex[&MBB], .dst: BlockIndex[Succ], .count: Freq});
3736 } else {
3737 for (MachineBasicBlock *Succ : MBB.successors()) {
3738 auto EP = MBPI->getEdgeProbability(Src: &MBB, Dst: Succ);
3739 BlockFrequency JumpFreq = BlockFreq * EP;
3740 JumpCounts.push_back(
3741 Elt: {.src: BlockIndex[&MBB], .dst: BlockIndex[Succ], .count: JumpFreq.getFrequency()});
3742 }
3743 }
3744 }
3745
3746 LLVM_DEBUG(dbgs() << "Applying ext-tsp layout for |V| = " << F->size()
3747 << " with profile = " << F->getFunction().hasProfileData()
3748 << " (" << F->getName() << ")" << "\n");
3749
3750 const double OrgScore = calcExtTspScore(NodeSizes: BlockSizes, EdgeCounts: JumpCounts);
3751 LLVM_DEBUG(dbgs() << format(" original layout score: %0.2f\n", OrgScore));
3752
3753 // Run the layout algorithm.
3754 auto NewOrder = computeExtTspLayout(NodeSizes: BlockSizes, NodeCounts: BlockCounts, EdgeCounts: JumpCounts);
3755 std::vector<const MachineBasicBlock *> NewBlockOrder;
3756 NewBlockOrder.reserve(n: F->size());
3757 for (uint64_t Node : NewOrder) {
3758 NewBlockOrder.push_back(x: CurrentBlockOrder[Node]);
3759 }
3760 const double OptScore = calcExtTspScore(Order: NewOrder, NodeSizes: BlockSizes, EdgeCounts: JumpCounts);
3761 LLVM_DEBUG(dbgs() << format(" optimized layout score: %0.2f\n", OptScore));
3762
3763 // If the optimization is unsuccessful, fall back to the original block order.
3764 if (OptForSize && OrgScore > OptScore)
3765 assignBlockOrder(NewOrder: CurrentBlockOrder);
3766 else
3767 assignBlockOrder(NewOrder: NewBlockOrder);
3768}
3769
3770void MachineBlockPlacement::assignBlockOrder(
3771 const std::vector<const MachineBasicBlock *> &NewBlockOrder) {
3772 assert(F->size() == NewBlockOrder.size() && "Incorrect size of block order");
3773 F->RenumberBlocks();
3774 // At this point, we possibly removed blocks from the function, so we can't
3775 // renumber the domtree. At this point, we don't need it anymore, though.
3776 // TODO: move this to the point where the dominator tree is actually
3777 // invalidated (i.e., where blocks are removed without updating the domtree).
3778 MPDT = nullptr;
3779
3780 bool HasChanges = false;
3781 for (size_t I = 0; I < NewBlockOrder.size(); I++) {
3782 if (NewBlockOrder[I] != F->getBlockNumbered(N: I)) {
3783 HasChanges = true;
3784 break;
3785 }
3786 }
3787 // Stop early if the new block order is identical to the existing one.
3788 if (!HasChanges)
3789 return;
3790
3791 SmallVector<MachineBasicBlock *, 4> PrevFallThroughs(F->getNumBlockIDs());
3792 for (auto &MBB : *F) {
3793 PrevFallThroughs[MBB.getNumber()] = MBB.getFallThrough();
3794 }
3795
3796 // Sort basic blocks in the function according to the computed order.
3797 DenseMap<const MachineBasicBlock *, size_t> NewIndex;
3798 for (const MachineBasicBlock *MBB : NewBlockOrder) {
3799 NewIndex[MBB] = NewIndex.size();
3800 }
3801 F->sort(comp: [&](MachineBasicBlock &L, MachineBasicBlock &R) {
3802 return NewIndex[&L] < NewIndex[&R];
3803 });
3804
3805 // Update basic block branches by inserting explicit fallthrough branches
3806 // when required and re-optimize branches when possible.
3807 const TargetInstrInfo *TII = F->getSubtarget().getInstrInfo();
3808 SmallVector<MachineOperand, 4> Cond;
3809 for (auto &MBB : *F) {
3810 MachineFunction::iterator NextMBB = std::next(x: MBB.getIterator());
3811 MachineFunction::iterator EndIt = MBB.getParent()->end();
3812 auto *FTMBB = PrevFallThroughs[MBB.getNumber()];
3813 // If this block had a fallthrough before we need an explicit unconditional
3814 // branch to that block if the fallthrough block is not adjacent to the
3815 // block in the new order.
3816 if (FTMBB && (NextMBB == EndIt || &*NextMBB != FTMBB)) {
3817 TII->insertUnconditionalBranch(MBB, DestBB: FTMBB, DL: MBB.findBranchDebugLoc());
3818 }
3819
3820 // It might be possible to optimize branches by flipping the condition.
3821 Cond.clear();
3822 MachineBasicBlock *TBB = nullptr, *FBB = nullptr;
3823 if (TII->analyzeBranch(MBB, TBB, FBB, Cond))
3824 continue;
3825 MBB.updateTerminator(PreviousLayoutSuccessor: FTMBB);
3826 }
3827}
3828
3829void MachineBlockPlacement::createCFGChainExtTsp() {
3830 BlockToChain.clear();
3831 ComputedEdges.clear();
3832 ChainAllocator.DestroyAll();
3833
3834 MachineBasicBlock *HeadBB = &F->front();
3835 BlockChain *FunctionChain =
3836 new (ChainAllocator.Allocate()) BlockChain(BlockToChain, HeadBB);
3837
3838 for (MachineBasicBlock &MBB : *F) {
3839 if (HeadBB == &MBB)
3840 continue; // Ignore head of the chain
3841 FunctionChain->merge(BB: &MBB, Chain: nullptr);
3842 }
3843}
3844
3845namespace {
3846
3847/// A pass to compute block placement statistics.
3848///
3849/// A separate pass to compute interesting statistics for evaluating block
3850/// placement. This is separate from the actual placement pass so that they can
3851/// be computed in the absence of any placement transformations or when using
3852/// alternative placement strategies.
3853class MachineBlockPlacementStats {
3854 /// A handle to the branch probability pass.
3855 const MachineBranchProbabilityInfo *MBPI;
3856
3857 /// A handle to the function-wide block frequency pass.
3858 const MachineBlockFrequencyInfo *MBFI;
3859
3860public:
3861 MachineBlockPlacementStats(const MachineBranchProbabilityInfo *MBPI,
3862 const MachineBlockFrequencyInfo *MBFI)
3863 : MBPI(MBPI), MBFI(MBFI) {}
3864 bool run(MachineFunction &MF);
3865};
3866
3867class MachineBlockPlacementStatsLegacy : public MachineFunctionPass {
3868public:
3869 static char ID; // Pass identification, replacement for typeid
3870
3871 MachineBlockPlacementStatsLegacy() : MachineFunctionPass(ID) {
3872 initializeMachineBlockPlacementStatsLegacyPass(
3873 *PassRegistry::getPassRegistry());
3874 }
3875
3876 bool runOnMachineFunction(MachineFunction &F) override {
3877 auto *MBPI =
3878 &getAnalysis<MachineBranchProbabilityInfoWrapperPass>().getMBPI();
3879 auto *MBFI = &getAnalysis<MachineBlockFrequencyInfoWrapperPass>().getMBFI();
3880 return MachineBlockPlacementStats(MBPI, MBFI).run(MF&: F);
3881 }
3882
3883 void getAnalysisUsage(AnalysisUsage &AU) const override {
3884 AU.addRequired<MachineBranchProbabilityInfoWrapperPass>();
3885 AU.addRequired<MachineBlockFrequencyInfoWrapperPass>();
3886 AU.setPreservesAll();
3887 MachineFunctionPass::getAnalysisUsage(AU);
3888 }
3889};
3890
3891} // end anonymous namespace
3892
3893char MachineBlockPlacementStatsLegacy::ID = 0;
3894
3895char &llvm::MachineBlockPlacementStatsID = MachineBlockPlacementStatsLegacy::ID;
3896
3897INITIALIZE_PASS_BEGIN(MachineBlockPlacementStatsLegacy, "block-placement-stats",
3898 "Basic Block Placement Stats", false, false)
3899INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfoWrapperPass)
3900INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfoWrapperPass)
3901INITIALIZE_PASS_END(MachineBlockPlacementStatsLegacy, "block-placement-stats",
3902 "Basic Block Placement Stats", false, false)
3903
3904PreservedAnalyses
3905MachineBlockPlacementStatsPass::run(MachineFunction &MF,
3906 MachineFunctionAnalysisManager &MFAM) {
3907 auto &MBPI = MFAM.getResult<MachineBranchProbabilityAnalysis>(IR&: MF);
3908 auto &MBFI = MFAM.getResult<MachineBlockFrequencyAnalysis>(IR&: MF);
3909
3910 MachineBlockPlacementStats(&MBPI, &MBFI).run(MF);
3911 return PreservedAnalyses::all();
3912}
3913
3914bool MachineBlockPlacementStats::run(MachineFunction &F) {
3915 // Check for single-block functions and skip them.
3916 if (std::next(x: F.begin()) == F.end())
3917 return false;
3918
3919 if (!isFunctionInPrintList(FunctionName: F.getName()))
3920 return false;
3921
3922 for (MachineBasicBlock &MBB : F) {
3923 BlockFrequency BlockFreq = MBFI->getBlockFreq(MBB: &MBB);
3924 Statistic &NumBranches =
3925 (MBB.succ_size() > 1) ? NumCondBranches : NumUncondBranches;
3926 Statistic &BranchTakenFreq =
3927 (MBB.succ_size() > 1) ? CondBranchTakenFreq : UncondBranchTakenFreq;
3928 for (MachineBasicBlock *Succ : MBB.successors()) {
3929 // Skip if this successor is a fallthrough.
3930 if (MBB.isLayoutSuccessor(MBB: Succ))
3931 continue;
3932
3933 BlockFrequency EdgeFreq =
3934 BlockFreq * MBPI->getEdgeProbability(Src: &MBB, Dst: Succ);
3935 ++NumBranches;
3936 BranchTakenFreq += EdgeFreq.getFrequency();
3937 }
3938 }
3939
3940 return false;
3941}
3942