• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===-- MachineBlockPlacement.cpp - Basic Block Code Layout optimization --===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements basic block placement transformations using the CFG
11 // structure and branch probability estimates.
12 //
13 // The pass strives to preserve the structure of the CFG (that is, retain
14 // a topological ordering of basic blocks) in the absence of a *strong* signal
15 // to the contrary from probabilities. However, within the CFG structure, it
16 // attempts to choose an ordering which favors placing more likely sequences of
17 // blocks adjacent to each other.
18 //
19 // The algorithm works from the inner-most loop within a function outward, and
20 // at each stage walks through the basic blocks, trying to coalesce them into
21 // sequential chains where allowed by the CFG (or demanded by heavy
22 // probabilities). Finally, it walks the blocks in topological order, and the
23 // first time it reaches a chain of basic blocks, it schedules them in the
24 // function in-order.
25 //
26 //===----------------------------------------------------------------------===//
27 
28 #include "llvm/CodeGen/Passes.h"
29 #include "llvm/CodeGen/TargetPassConfig.h"
30 #include "BranchFolding.h"
31 #include "llvm/ADT/DenseMap.h"
32 #include "llvm/ADT/SmallPtrSet.h"
33 #include "llvm/ADT/SmallVector.h"
34 #include "llvm/ADT/Statistic.h"
35 #include "llvm/CodeGen/MachineBasicBlock.h"
36 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
37 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
38 #include "llvm/CodeGen/MachineDominators.h"
39 #include "llvm/CodeGen/MachineFunction.h"
40 #include "llvm/CodeGen/MachineFunctionPass.h"
41 #include "llvm/CodeGen/MachineLoopInfo.h"
42 #include "llvm/CodeGen/MachineModuleInfo.h"
43 #include "llvm/Support/Allocator.h"
44 #include "llvm/Support/CommandLine.h"
45 #include "llvm/Support/Debug.h"
46 #include "llvm/Support/raw_ostream.h"
47 #include "llvm/Target/TargetInstrInfo.h"
48 #include "llvm/Target/TargetLowering.h"
49 #include "llvm/Target/TargetSubtargetInfo.h"
50 #include <algorithm>
51 using namespace llvm;
52 
53 #define DEBUG_TYPE "block-placement"
54 
55 STATISTIC(NumCondBranches, "Number of conditional branches");
56 STATISTIC(NumUncondBranches, "Number of unconditional branches");
57 STATISTIC(CondBranchTakenFreq,
58           "Potential frequency of taking conditional branches");
59 STATISTIC(UncondBranchTakenFreq,
60           "Potential frequency of taking unconditional branches");
61 
62 static cl::opt<unsigned> AlignAllBlock("align-all-blocks",
63                                        cl::desc("Force the alignment of all "
64                                                 "blocks in the function."),
65                                        cl::init(0), cl::Hidden);
66 
67 static cl::opt<unsigned> AlignAllNonFallThruBlocks(
68     "align-all-nofallthru-blocks",
69     cl::desc("Force the alignment of all "
70              "blocks that have no fall-through predecessors (i.e. don't add "
71              "nops that are executed)."),
72     cl::init(0), cl::Hidden);
73 
74 // FIXME: Find a good default for this flag and remove the flag.
75 static cl::opt<unsigned> ExitBlockBias(
76     "block-placement-exit-block-bias",
77     cl::desc("Block frequency percentage a loop exit block needs "
78              "over the original exit to be considered the new exit."),
79     cl::init(0), cl::Hidden);
80 
81 static cl::opt<bool> OutlineOptionalBranches(
82     "outline-optional-branches",
83     cl::desc("Put completely optional branches, i.e. branches with a common "
84              "post dominator, out of line."),
85     cl::init(false), cl::Hidden);
86 
87 static cl::opt<unsigned> OutlineOptionalThreshold(
88     "outline-optional-threshold",
89     cl::desc("Don't outline optional branches that are a single block with an "
90              "instruction count below this threshold"),
91     cl::init(4), cl::Hidden);
92 
93 static cl::opt<unsigned> LoopToColdBlockRatio(
94     "loop-to-cold-block-ratio",
95     cl::desc("Outline loop blocks from loop chain if (frequency of loop) / "
96              "(frequency of block) is greater than this ratio"),
97     cl::init(5), cl::Hidden);
98 
99 static cl::opt<bool>
100     PreciseRotationCost("precise-rotation-cost",
101                         cl::desc("Model the cost of loop rotation more "
102                                  "precisely by using profile data."),
103                         cl::init(false), cl::Hidden);
104 static cl::opt<bool>
105     ForcePreciseRotationCost("force-precise-rotation-cost",
106                              cl::desc("Force the use of precise cost "
107                                       "loop rotation strategy."),
108                              cl::init(false), cl::Hidden);
109 
110 static cl::opt<unsigned> MisfetchCost(
111     "misfetch-cost",
112     cl::desc("Cost that models the probablistic risk of an instruction "
113              "misfetch due to a jump comparing to falling through, whose cost "
114              "is zero."),
115     cl::init(1), cl::Hidden);
116 
117 static cl::opt<unsigned> JumpInstCost("jump-inst-cost",
118                                       cl::desc("Cost of jump instructions."),
119                                       cl::init(1), cl::Hidden);
120 
121 static cl::opt<bool>
122 BranchFoldPlacement("branch-fold-placement",
123               cl::desc("Perform branch folding during placement. "
124                        "Reduces code size."),
125               cl::init(true), cl::Hidden);
126 
127 extern cl::opt<unsigned> StaticLikelyProb;
128 extern cl::opt<unsigned> ProfileLikelyProb;
129 
130 namespace {
131 class BlockChain;
132 /// \brief Type for our function-wide basic block -> block chain mapping.
133 typedef DenseMap<MachineBasicBlock *, BlockChain *> BlockToChainMapType;
134 }
135 
136 namespace {
137 /// \brief A chain of blocks which will be laid out contiguously.
138 ///
139 /// This is the datastructure representing a chain of consecutive blocks that
140 /// are profitable to layout together in order to maximize fallthrough
141 /// probabilities and code locality. We also can use a block chain to represent
142 /// a sequence of basic blocks which have some external (correctness)
143 /// requirement for sequential layout.
144 ///
145 /// Chains can be built around a single basic block and can be merged to grow
146 /// them. They participate in a block-to-chain mapping, which is updated
147 /// automatically as chains are merged together.
148 class BlockChain {
149   /// \brief The sequence of blocks belonging to this chain.
150   ///
151   /// This is the sequence of blocks for a particular chain. These will be laid
152   /// out in-order within the function.
153   SmallVector<MachineBasicBlock *, 4> Blocks;
154 
155   /// \brief A handle to the function-wide basic block to block chain mapping.
156   ///
157   /// This is retained in each block chain to simplify the computation of child
158   /// block chains for SCC-formation and iteration. We store the edges to child
159   /// basic blocks, and map them back to their associated chains using this
160   /// structure.
161   BlockToChainMapType &BlockToChain;
162 
163 public:
164   /// \brief Construct a new BlockChain.
165   ///
166   /// This builds a new block chain representing a single basic block in the
167   /// function. It also registers itself as the chain that block participates
168   /// in with the BlockToChain mapping.
BlockChain(BlockToChainMapType & BlockToChain,MachineBasicBlock * BB)169   BlockChain(BlockToChainMapType &BlockToChain, MachineBasicBlock *BB)
170       : Blocks(1, BB), BlockToChain(BlockToChain), UnscheduledPredecessors(0) {
171     assert(BB && "Cannot create a chain with a null basic block");
172     BlockToChain[BB] = this;
173   }
174 
175   /// \brief Iterator over blocks within the chain.
176   typedef SmallVectorImpl<MachineBasicBlock *>::iterator iterator;
177 
178   /// \brief Beginning of blocks within the chain.
begin()179   iterator begin() { return Blocks.begin(); }
180 
181   /// \brief End of blocks within the chain.
end()182   iterator end() { return Blocks.end(); }
183 
184   /// \brief Merge a block chain into this one.
185   ///
186   /// This routine merges a block chain into this one. It takes care of forming
187   /// a contiguous sequence of basic blocks, updating the edge list, and
188   /// updating the block -> chain mapping. It does not free or tear down the
189   /// old chain, but the old chain's block list is no longer valid.
merge(MachineBasicBlock * BB,BlockChain * Chain)190   void merge(MachineBasicBlock *BB, BlockChain *Chain) {
191     assert(BB);
192     assert(!Blocks.empty());
193 
194     // Fast path in case we don't have a chain already.
195     if (!Chain) {
196       assert(!BlockToChain[BB]);
197       Blocks.push_back(BB);
198       BlockToChain[BB] = this;
199       return;
200     }
201 
202     assert(BB == *Chain->begin());
203     assert(Chain->begin() != Chain->end());
204 
205     // Update the incoming blocks to point to this chain, and add them to the
206     // chain structure.
207     for (MachineBasicBlock *ChainBB : *Chain) {
208       Blocks.push_back(ChainBB);
209       assert(BlockToChain[ChainBB] == Chain && "Incoming blocks not in chain");
210       BlockToChain[ChainBB] = this;
211     }
212   }
213 
214 #ifndef NDEBUG
215   /// \brief Dump the blocks in this chain.
dump()216   LLVM_DUMP_METHOD void dump() {
217     for (MachineBasicBlock *MBB : *this)
218       MBB->dump();
219   }
220 #endif // NDEBUG
221 
222   /// \brief Count of predecessors of any block within the chain which have not
223   /// yet been scheduled.  In general, we will delay scheduling this chain
224   /// until those predecessors are scheduled (or we find a sufficiently good
225   /// reason to override this heuristic.)  Note that when forming loop chains,
226   /// blocks outside the loop are ignored and treated as if they were already
227   /// scheduled.
228   ///
229   /// Note: This field is reinitialized multiple times - once for each loop,
230   /// and then once for the function as a whole.
231   unsigned UnscheduledPredecessors;
232 };
233 }
234 
235 namespace {
236 class MachineBlockPlacement : public MachineFunctionPass {
237   /// \brief A typedef for a block filter set.
238   typedef SmallPtrSet<MachineBasicBlock *, 16> BlockFilterSet;
239 
240   /// \brief work lists of blocks that are ready to be laid out
241   SmallVector<MachineBasicBlock *, 16> BlockWorkList;
242   SmallVector<MachineBasicBlock *, 16> EHPadWorkList;
243 
244   /// \brief Machine Function
245   MachineFunction *F;
246 
247   /// \brief A handle to the branch probability pass.
248   const MachineBranchProbabilityInfo *MBPI;
249 
250   /// \brief A handle to the function-wide block frequency pass.
251   std::unique_ptr<BranchFolder::MBFIWrapper> MBFI;
252 
253   /// \brief A handle to the loop info.
254   MachineLoopInfo *MLI;
255 
256   /// \brief A handle to the target's instruction info.
257   const TargetInstrInfo *TII;
258 
259   /// \brief A handle to the target's lowering info.
260   const TargetLoweringBase *TLI;
261 
262   /// \brief A handle to the post dominator tree.
263   MachineDominatorTree *MDT;
264 
265   /// \brief A set of blocks that are unavoidably execute, i.e. they dominate
266   /// all terminators of the MachineFunction.
267   SmallPtrSet<MachineBasicBlock *, 4> UnavoidableBlocks;
268 
269   /// \brief Allocator and owner of BlockChain structures.
270   ///
271   /// We build BlockChains lazily while processing the loop structure of
272   /// a function. To reduce malloc traffic, we allocate them using this
273   /// slab-like allocator, and destroy them after the pass completes. An
274   /// important guarantee is that this allocator produces stable pointers to
275   /// the chains.
276   SpecificBumpPtrAllocator<BlockChain> ChainAllocator;
277 
278   /// \brief Function wide BasicBlock to BlockChain mapping.
279   ///
280   /// This mapping allows efficiently moving from any given basic block to the
281   /// BlockChain it participates in, if any. We use it to, among other things,
282   /// allow implicitly defining edges between chains as the existing edges
283   /// between basic blocks.
284   DenseMap<MachineBasicBlock *, BlockChain *> BlockToChain;
285 
286   void markChainSuccessors(BlockChain &Chain, MachineBasicBlock *LoopHeaderBB,
287                            const BlockFilterSet *BlockFilter = nullptr);
288   BranchProbability
289   collectViableSuccessors(MachineBasicBlock *BB, BlockChain &Chain,
290                           const BlockFilterSet *BlockFilter,
291                           SmallVector<MachineBasicBlock *, 4> &Successors);
292   bool shouldPredBlockBeOutlined(MachineBasicBlock *BB, MachineBasicBlock *Succ,
293                                  BlockChain &Chain,
294                                  const BlockFilterSet *BlockFilter,
295                                  BranchProbability SuccProb,
296                                  BranchProbability HotProb);
297   bool
298   hasBetterLayoutPredecessor(MachineBasicBlock *BB, MachineBasicBlock *Succ,
299                              BlockChain &SuccChain, BranchProbability SuccProb,
300                              BranchProbability RealSuccProb, BlockChain &Chain,
301                              const BlockFilterSet *BlockFilter);
302   MachineBasicBlock *selectBestSuccessor(MachineBasicBlock *BB,
303                                          BlockChain &Chain,
304                                          const BlockFilterSet *BlockFilter);
305   MachineBasicBlock *
306   selectBestCandidateBlock(BlockChain &Chain,
307                            SmallVectorImpl<MachineBasicBlock *> &WorkList);
308   MachineBasicBlock *
309   getFirstUnplacedBlock(const BlockChain &PlacedChain,
310                         MachineFunction::iterator &PrevUnplacedBlockIt,
311                         const BlockFilterSet *BlockFilter);
312 
313   /// \brief Add a basic block to the work list if it is apropriate.
314   ///
315   /// If the optional parameter BlockFilter is provided, only MBB
316   /// present in the set will be added to the worklist. If nullptr
317   /// is provided, no filtering occurs.
318   void fillWorkLists(MachineBasicBlock *MBB,
319                      SmallPtrSetImpl<BlockChain *> &UpdatedPreds,
320                      const BlockFilterSet *BlockFilter);
321   void buildChain(MachineBasicBlock *BB, BlockChain &Chain,
322                   const BlockFilterSet *BlockFilter = nullptr);
323   MachineBasicBlock *findBestLoopTop(MachineLoop &L,
324                                      const BlockFilterSet &LoopBlockSet);
325   MachineBasicBlock *findBestLoopExit(MachineLoop &L,
326                                       const BlockFilterSet &LoopBlockSet);
327   BlockFilterSet collectLoopBlockSet(MachineLoop &L);
328   void buildLoopChains(MachineLoop &L);
329   void rotateLoop(BlockChain &LoopChain, MachineBasicBlock *ExitingBB,
330                   const BlockFilterSet &LoopBlockSet);
331   void rotateLoopWithProfile(BlockChain &LoopChain, MachineLoop &L,
332                              const BlockFilterSet &LoopBlockSet);
333   void collectMustExecuteBBs();
334   void buildCFGChains();
335   void optimizeBranches();
336   void alignBlocks();
337 
338 public:
339   static char ID; // Pass identification, replacement for typeid
MachineBlockPlacement()340   MachineBlockPlacement() : MachineFunctionPass(ID) {
341     initializeMachineBlockPlacementPass(*PassRegistry::getPassRegistry());
342   }
343 
344   bool runOnMachineFunction(MachineFunction &F) override;
345 
getAnalysisUsage(AnalysisUsage & AU) const346   void getAnalysisUsage(AnalysisUsage &AU) const override {
347     AU.addRequired<MachineBranchProbabilityInfo>();
348     AU.addRequired<MachineBlockFrequencyInfo>();
349     AU.addRequired<MachineDominatorTree>();
350     AU.addRequired<MachineLoopInfo>();
351     AU.addRequired<TargetPassConfig>();
352     MachineFunctionPass::getAnalysisUsage(AU);
353   }
354 };
355 }
356 
357 char MachineBlockPlacement::ID = 0;
358 char &llvm::MachineBlockPlacementID = MachineBlockPlacement::ID;
359 INITIALIZE_PASS_BEGIN(MachineBlockPlacement, "block-placement",
360                       "Branch Probability Basic Block Placement", false, false)
INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)361 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
362 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
363 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
364 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
365 INITIALIZE_PASS_END(MachineBlockPlacement, "block-placement",
366                     "Branch Probability Basic Block Placement", false, false)
367 
368 #ifndef NDEBUG
369 /// \brief Helper to print the name of a MBB.
370 ///
371 /// Only used by debug logging.
372 static std::string getBlockName(MachineBasicBlock *BB) {
373   std::string Result;
374   raw_string_ostream OS(Result);
375   OS << "BB#" << BB->getNumber();
376   OS << " ('" << BB->getName() << "')";
377   OS.flush();
378   return Result;
379 }
380 #endif
381 
382 /// \brief Mark a chain's successors as having one fewer preds.
383 ///
384 /// When a chain is being merged into the "placed" chain, this routine will
385 /// quickly walk the successors of each block in the chain and mark them as
386 /// having one fewer active predecessor. It also adds any successors of this
387 /// chain which reach the zero-predecessor state to the worklist passed in.
markChainSuccessors(BlockChain & Chain,MachineBasicBlock * LoopHeaderBB,const BlockFilterSet * BlockFilter)388 void MachineBlockPlacement::markChainSuccessors(
389     BlockChain &Chain, MachineBasicBlock *LoopHeaderBB,
390     const BlockFilterSet *BlockFilter) {
391   // Walk all the blocks in this chain, marking their successors as having
392   // a predecessor placed.
393   for (MachineBasicBlock *MBB : Chain) {
394     // Add any successors for which this is the only un-placed in-loop
395     // predecessor to the worklist as a viable candidate for CFG-neutral
396     // placement. No subsequent placement of this block will violate the CFG
397     // shape, so we get to use heuristics to choose a favorable placement.
398     for (MachineBasicBlock *Succ : MBB->successors()) {
399       if (BlockFilter && !BlockFilter->count(Succ))
400         continue;
401       BlockChain &SuccChain = *BlockToChain[Succ];
402       // Disregard edges within a fixed chain, or edges to the loop header.
403       if (&Chain == &SuccChain || Succ == LoopHeaderBB)
404         continue;
405 
406       // This is a cross-chain edge that is within the loop, so decrement the
407       // loop predecessor count of the destination chain.
408       if (SuccChain.UnscheduledPredecessors == 0 ||
409           --SuccChain.UnscheduledPredecessors > 0)
410         continue;
411 
412       auto *MBB = *SuccChain.begin();
413       if (MBB->isEHPad())
414         EHPadWorkList.push_back(MBB);
415       else
416         BlockWorkList.push_back(MBB);
417     }
418   }
419 }
420 
421 /// This helper function collects the set of successors of block
422 /// \p BB that are allowed to be its layout successors, and return
423 /// the total branch probability of edges from \p BB to those
424 /// blocks.
collectViableSuccessors(MachineBasicBlock * BB,BlockChain & Chain,const BlockFilterSet * BlockFilter,SmallVector<MachineBasicBlock *,4> & Successors)425 BranchProbability MachineBlockPlacement::collectViableSuccessors(
426     MachineBasicBlock *BB, BlockChain &Chain, const BlockFilterSet *BlockFilter,
427     SmallVector<MachineBasicBlock *, 4> &Successors) {
428   // Adjust edge probabilities by excluding edges pointing to blocks that is
429   // either not in BlockFilter or is already in the current chain. Consider the
430   // following CFG:
431   //
432   //     --->A
433   //     |  / \
434   //     | B   C
435   //     |  \ / \
436   //     ----D   E
437   //
438   // Assume A->C is very hot (>90%), and C->D has a 50% probability, then after
439   // A->C is chosen as a fall-through, D won't be selected as a successor of C
440   // due to CFG constraint (the probability of C->D is not greater than
441   // HotProb to break top-oorder). If we exclude E that is not in BlockFilter
442   // when calculating the  probability of C->D, D will be selected and we
443   // will get A C D B as the layout of this loop.
444   auto AdjustedSumProb = BranchProbability::getOne();
445   for (MachineBasicBlock *Succ : BB->successors()) {
446     bool SkipSucc = false;
447     if (Succ->isEHPad() || (BlockFilter && !BlockFilter->count(Succ))) {
448       SkipSucc = true;
449     } else {
450       BlockChain *SuccChain = BlockToChain[Succ];
451       if (SuccChain == &Chain) {
452         SkipSucc = true;
453       } else if (Succ != *SuccChain->begin()) {
454         DEBUG(dbgs() << "    " << getBlockName(Succ) << " -> Mid chain!\n");
455         continue;
456       }
457     }
458     if (SkipSucc)
459       AdjustedSumProb -= MBPI->getEdgeProbability(BB, Succ);
460     else
461       Successors.push_back(Succ);
462   }
463 
464   return AdjustedSumProb;
465 }
466 
467 /// The helper function returns the branch probability that is adjusted
468 /// or normalized over the new total \p AdjustedSumProb.
469 
470 static BranchProbability
getAdjustedProbability(BranchProbability OrigProb,BranchProbability AdjustedSumProb)471 getAdjustedProbability(BranchProbability OrigProb,
472                        BranchProbability AdjustedSumProb) {
473   BranchProbability SuccProb;
474   uint32_t SuccProbN = OrigProb.getNumerator();
475   uint32_t SuccProbD = AdjustedSumProb.getNumerator();
476   if (SuccProbN >= SuccProbD)
477     SuccProb = BranchProbability::getOne();
478   else
479     SuccProb = BranchProbability(SuccProbN, SuccProbD);
480 
481   return SuccProb;
482 }
483 
484 /// When the option OutlineOptionalBranches is on, this method
485 /// checks if the fallthrough candidate block \p Succ (of block
486 /// \p BB) also has other unscheduled predecessor blocks which
487 /// are also successors of \p BB (forming triagular shape CFG).
488 /// If none of such predecessors are small, it returns true.
489 /// The caller can choose to select \p Succ as the layout successors
490 /// so that \p Succ's predecessors (optional branches) can be
491 /// outlined.
492 /// FIXME: fold this with more general layout cost analysis.
shouldPredBlockBeOutlined(MachineBasicBlock * BB,MachineBasicBlock * Succ,BlockChain & Chain,const BlockFilterSet * BlockFilter,BranchProbability SuccProb,BranchProbability HotProb)493 bool MachineBlockPlacement::shouldPredBlockBeOutlined(
494     MachineBasicBlock *BB, MachineBasicBlock *Succ, BlockChain &Chain,
495     const BlockFilterSet *BlockFilter, BranchProbability SuccProb,
496     BranchProbability HotProb) {
497   if (!OutlineOptionalBranches)
498     return false;
499   // If we outline optional branches, look whether Succ is unavoidable, i.e.
500   // dominates all terminators of the MachineFunction. If it does, other
501   // successors must be optional. Don't do this for cold branches.
502   if (SuccProb > HotProb.getCompl() && UnavoidableBlocks.count(Succ) > 0) {
503     for (MachineBasicBlock *Pred : Succ->predecessors()) {
504       // Check whether there is an unplaced optional branch.
505       if (Pred == Succ || (BlockFilter && !BlockFilter->count(Pred)) ||
506           BlockToChain[Pred] == &Chain)
507         continue;
508       // Check whether the optional branch has exactly one BB.
509       if (Pred->pred_size() > 1 || *Pred->pred_begin() != BB)
510         continue;
511       // Check whether the optional branch is small.
512       if (Pred->size() < OutlineOptionalThreshold)
513         return false;
514     }
515     return true;
516   } else
517     return false;
518 }
519 
520 // When profile is not present, return the StaticLikelyProb.
521 // When profile is available, we need to handle the triangle-shape CFG.
getLayoutSuccessorProbThreshold(MachineBasicBlock * BB)522 static BranchProbability getLayoutSuccessorProbThreshold(
523       MachineBasicBlock *BB) {
524   if (!BB->getParent()->getFunction()->getEntryCount())
525     return BranchProbability(StaticLikelyProb, 100);
526   if (BB->succ_size() == 2) {
527     const MachineBasicBlock *Succ1 = *BB->succ_begin();
528     const MachineBasicBlock *Succ2 = *(BB->succ_begin() + 1);
529     if (Succ1->isSuccessor(Succ2) || Succ2->isSuccessor(Succ1)) {
530       /* See case 1 below for the cost analysis. For BB->Succ to
531        * be taken with smaller cost, the following needs to hold:
532        *   Prob(BB->Succ) > 2* Prob(BB->Pred)
533        *   So the threshold T
534        *   T = 2 * (1-Prob(BB->Pred). Since T + Prob(BB->Pred) == 1,
535        * We have  T + T/2 = 1, i.e. T = 2/3. Also adding user specified
536        * branch bias, we have
537        *   T = (2/3)*(ProfileLikelyProb/50)
538        *     = (2*ProfileLikelyProb)/150)
539        */
540       return BranchProbability(2 * ProfileLikelyProb, 150);
541     }
542   }
543   return BranchProbability(ProfileLikelyProb, 100);
544 }
545 
546 /// Checks to see if the layout candidate block \p Succ has a better layout
547 /// predecessor than \c BB. If yes, returns true.
hasBetterLayoutPredecessor(MachineBasicBlock * BB,MachineBasicBlock * Succ,BlockChain & SuccChain,BranchProbability SuccProb,BranchProbability RealSuccProb,BlockChain & Chain,const BlockFilterSet * BlockFilter)548 bool MachineBlockPlacement::hasBetterLayoutPredecessor(
549     MachineBasicBlock *BB, MachineBasicBlock *Succ, BlockChain &SuccChain,
550     BranchProbability SuccProb, BranchProbability RealSuccProb,
551     BlockChain &Chain, const BlockFilterSet *BlockFilter) {
552 
553   // This is no global conflict, just return false.
554   if (SuccChain.UnscheduledPredecessors == 0)
555     return false;
556 
557   // There are two basic scenarios here:
558   // -------------------------------------
559   // Case 1: triagular shape CFG:
560   //     BB
561   //     | \
562   //     |  \
563   //     |   Pred
564   //     |   /
565   //     Succ
566   // In this case, we are evaluating whether to select edge -> Succ, e.g.
567   // set Succ as the layout successor of BB. Picking Succ as BB's
568   // successor breaks the  CFG constraints. With this layout, Pred BB
569   // is forced to be outlined, so the overall cost will be cost of the
570   // branch taken from BB to Pred, plus the cost of back taken branch
571   // from Pred to Succ, as well as the additional cost asssociated
572   // with the needed unconditional jump instruction from Pred To Succ.
573   // The cost of the topological order layout is the taken branch cost
574   // from BB to Succ, so to make BB->Succ a viable candidate, the following
575   // must hold:
576   //     2 * freq(BB->Pred) * taken_branch_cost + unconditional_jump_cost
577   //      < freq(BB->Succ) *  taken_branch_cost.
578   // Ignoring unconditional jump cost, we get
579   //    freq(BB->Succ) > 2 * freq(BB->Pred), i.e.,
580   //    prob(BB->Succ) > 2 * prob(BB->Pred)
581   //
582   // When real profile data is available, we can precisely compute the the
583   // probabililty threshold that is needed for edge BB->Succ to be considered.
584   // With out profile data, the heuristic requires the branch bias to be
585   // a lot larger to make sure the signal is very strong (e.g. 80% default).
586   // -----------------------------------------------------------------
587   // Case 2: diamond like CFG:
588   //     S
589   //    / \
590   //   |   \
591   //  BB    Pred
592   //   \    /
593   //    Succ
594   //    ..
595   // In this case, edge S->BB has already been selected, and we are evaluating
596   // candidate edge BB->Succ. Edge S->BB is selected because prob(S->BB)
597   // is no less than prob(S->Pred). When real profile data is *available*, if
598   // the condition is true, it will be always better to continue the trace with
599   // edge BB->Succ instead of laying out with topological order (i.e. laying
600   // Pred first).  The cost of S->BB->Succ is 2 * freq (S->Pred), while with
601   // the topo order, the cost is freq(S-> Pred) + Pred(S->BB) which is larger.
602   // When profile data is not available, however, we need to be more
603   // conservative. If the branch prediction is wrong, breaking the topo-order
604   // will actually yield a layout with large cost. For this reason, we need
605   // strong biaaed branch at block S with Prob(S->BB) in order to select
606   // BB->Succ. This is equialant to looking the CFG backward with backward
607   // edge: Prob(Succ->BB) needs to >= HotProb in order to be selected (without
608   // profile data).
609 
610   BranchProbability HotProb = getLayoutSuccessorProbThreshold(BB);
611 
612   // Forward checking. For case 2, SuccProb will be 1.
613   if (SuccProb < HotProb) {
614     DEBUG(dbgs() << "    " << getBlockName(Succ) << " -> " << SuccProb
615                  << " (prob) (CFG conflict)\n");
616     return true;
617   }
618 
619   // Make sure that a hot successor doesn't have a globally more
620   // important predecessor.
621   BlockFrequency CandidateEdgeFreq = MBFI->getBlockFreq(BB) * RealSuccProb;
622   bool BadCFGConflict = false;
623 
624   for (MachineBasicBlock *Pred : Succ->predecessors()) {
625     if (Pred == Succ || BlockToChain[Pred] == &SuccChain ||
626         (BlockFilter && !BlockFilter->count(Pred)) ||
627         BlockToChain[Pred] == &Chain)
628       continue;
629     // Do backward checking. For case 1, it is actually redundant check. For
630     // case 2 above, we need a backward checking to filter out edges that are
631     // not 'strongly' biased. With profile data available, the check is mostly
632     // redundant too (when threshold prob is set at 50%) unless S has more than
633     // two successors.
634     // BB  Pred
635     //  \ /
636     //  Succ
637     // We select edgee BB->Succ if
638     //      freq(BB->Succ) > freq(Succ) * HotProb
639     //      i.e. freq(BB->Succ) > freq(BB->Succ) * HotProb + freq(Pred->Succ) *
640     //      HotProb
641     //      i.e. freq((BB->Succ) * (1 - HotProb) > freq(Pred->Succ) * HotProb
642     BlockFrequency PredEdgeFreq =
643         MBFI->getBlockFreq(Pred) * MBPI->getEdgeProbability(Pred, Succ);
644     if (PredEdgeFreq * HotProb >= CandidateEdgeFreq * HotProb.getCompl()) {
645       BadCFGConflict = true;
646       break;
647     }
648   }
649 
650   if (BadCFGConflict) {
651     DEBUG(dbgs() << "    " << getBlockName(Succ) << " -> " << SuccProb
652                  << " (prob) (non-cold CFG conflict)\n");
653     return true;
654   }
655 
656   return false;
657 }
658 
659 /// \brief Select the best successor for a block.
660 ///
661 /// This looks across all successors of a particular block and attempts to
662 /// select the "best" one to be the layout successor. It only considers direct
663 /// successors which also pass the block filter. It will attempt to avoid
664 /// breaking CFG structure, but cave and break such structures in the case of
665 /// very hot successor edges.
666 ///
667 /// \returns The best successor block found, or null if none are viable.
668 MachineBasicBlock *
selectBestSuccessor(MachineBasicBlock * BB,BlockChain & Chain,const BlockFilterSet * BlockFilter)669 MachineBlockPlacement::selectBestSuccessor(MachineBasicBlock *BB,
670                                            BlockChain &Chain,
671                                            const BlockFilterSet *BlockFilter) {
672   const BranchProbability HotProb(StaticLikelyProb, 100);
673 
674   MachineBasicBlock *BestSucc = nullptr;
675   auto BestProb = BranchProbability::getZero();
676 
677   SmallVector<MachineBasicBlock *, 4> Successors;
678   auto AdjustedSumProb =
679       collectViableSuccessors(BB, Chain, BlockFilter, Successors);
680 
681   DEBUG(dbgs() << "Attempting merge from: " << getBlockName(BB) << "\n");
682   for (MachineBasicBlock *Succ : Successors) {
683     auto RealSuccProb = MBPI->getEdgeProbability(BB, Succ);
684     BranchProbability SuccProb =
685         getAdjustedProbability(RealSuccProb, AdjustedSumProb);
686 
687     // This heuristic is off by default.
688     if (shouldPredBlockBeOutlined(BB, Succ, Chain, BlockFilter, SuccProb,
689                                   HotProb))
690       return Succ;
691 
692     BlockChain &SuccChain = *BlockToChain[Succ];
693     // Skip the edge \c BB->Succ if block \c Succ has a better layout
694     // predecessor that yields lower global cost.
695     if (hasBetterLayoutPredecessor(BB, Succ, SuccChain, SuccProb, RealSuccProb,
696                                    Chain, BlockFilter))
697       continue;
698 
699     DEBUG(
700         dbgs() << "    " << getBlockName(Succ) << " -> " << SuccProb
701                << " (prob)"
702                << (SuccChain.UnscheduledPredecessors != 0 ? " (CFG break)" : "")
703                << "\n");
704     if (BestSucc && BestProb >= SuccProb)
705       continue;
706     BestSucc = Succ;
707     BestProb = SuccProb;
708   }
709   return BestSucc;
710 }
711 
712 /// \brief Select the best block from a worklist.
713 ///
714 /// This looks through the provided worklist as a list of candidate basic
715 /// blocks and select the most profitable one to place. The definition of
716 /// profitable only really makes sense in the context of a loop. This returns
717 /// the most frequently visited block in the worklist, which in the case of
718 /// a loop, is the one most desirable to be physically close to the rest of the
719 /// loop body in order to improve icache behavior.
720 ///
721 /// \returns The best block found, or null if none are viable.
selectBestCandidateBlock(BlockChain & Chain,SmallVectorImpl<MachineBasicBlock * > & WorkList)722 MachineBasicBlock *MachineBlockPlacement::selectBestCandidateBlock(
723     BlockChain &Chain, SmallVectorImpl<MachineBasicBlock *> &WorkList) {
724   // Once we need to walk the worklist looking for a candidate, cleanup the
725   // worklist of already placed entries.
726   // FIXME: If this shows up on profiles, it could be folded (at the cost of
727   // some code complexity) into the loop below.
728   WorkList.erase(std::remove_if(WorkList.begin(), WorkList.end(),
729                                 [&](MachineBasicBlock *BB) {
730                                   return BlockToChain.lookup(BB) == &Chain;
731                                 }),
732                  WorkList.end());
733 
734   if (WorkList.empty())
735     return nullptr;
736 
737   bool IsEHPad = WorkList[0]->isEHPad();
738 
739   MachineBasicBlock *BestBlock = nullptr;
740   BlockFrequency BestFreq;
741   for (MachineBasicBlock *MBB : WorkList) {
742     assert(MBB->isEHPad() == IsEHPad);
743 
744     BlockChain &SuccChain = *BlockToChain[MBB];
745     if (&SuccChain == &Chain)
746       continue;
747 
748     assert(SuccChain.UnscheduledPredecessors == 0 && "Found CFG-violating block");
749 
750     BlockFrequency CandidateFreq = MBFI->getBlockFreq(MBB);
751     DEBUG(dbgs() << "    " << getBlockName(MBB) << " -> ";
752           MBFI->printBlockFreq(dbgs(), CandidateFreq) << " (freq)\n");
753 
754     // For ehpad, we layout the least probable first as to avoid jumping back
755     // from least probable landingpads to more probable ones.
756     //
757     // FIXME: Using probability is probably (!) not the best way to achieve
758     // this. We should probably have a more principled approach to layout
759     // cleanup code.
760     //
761     // The goal is to get:
762     //
763     //                 +--------------------------+
764     //                 |                          V
765     // InnerLp -> InnerCleanup    OuterLp -> OuterCleanup -> Resume
766     //
767     // Rather than:
768     //
769     //                 +-------------------------------------+
770     //                 V                                     |
771     // OuterLp -> OuterCleanup -> Resume     InnerLp -> InnerCleanup
772     if (BestBlock && (IsEHPad ^ (BestFreq >= CandidateFreq)))
773       continue;
774 
775     BestBlock = MBB;
776     BestFreq = CandidateFreq;
777   }
778 
779   return BestBlock;
780 }
781 
782 /// \brief Retrieve the first unplaced basic block.
783 ///
784 /// This routine is called when we are unable to use the CFG to walk through
785 /// all of the basic blocks and form a chain due to unnatural loops in the CFG.
786 /// We walk through the function's blocks in order, starting from the
787 /// LastUnplacedBlockIt. We update this iterator on each call to avoid
788 /// re-scanning the entire sequence on repeated calls to this routine.
getFirstUnplacedBlock(const BlockChain & PlacedChain,MachineFunction::iterator & PrevUnplacedBlockIt,const BlockFilterSet * BlockFilter)789 MachineBasicBlock *MachineBlockPlacement::getFirstUnplacedBlock(
790     const BlockChain &PlacedChain,
791     MachineFunction::iterator &PrevUnplacedBlockIt,
792     const BlockFilterSet *BlockFilter) {
793   for (MachineFunction::iterator I = PrevUnplacedBlockIt, E = F->end(); I != E;
794        ++I) {
795     if (BlockFilter && !BlockFilter->count(&*I))
796       continue;
797     if (BlockToChain[&*I] != &PlacedChain) {
798       PrevUnplacedBlockIt = I;
799       // Now select the head of the chain to which the unplaced block belongs
800       // as the block to place. This will force the entire chain to be placed,
801       // and satisfies the requirements of merging chains.
802       return *BlockToChain[&*I]->begin();
803     }
804   }
805   return nullptr;
806 }
807 
fillWorkLists(MachineBasicBlock * MBB,SmallPtrSetImpl<BlockChain * > & UpdatedPreds,const BlockFilterSet * BlockFilter=nullptr)808 void MachineBlockPlacement::fillWorkLists(
809     MachineBasicBlock *MBB,
810     SmallPtrSetImpl<BlockChain *> &UpdatedPreds,
811     const BlockFilterSet *BlockFilter = nullptr) {
812   BlockChain &Chain = *BlockToChain[MBB];
813   if (!UpdatedPreds.insert(&Chain).second)
814     return;
815 
816   assert(Chain.UnscheduledPredecessors == 0);
817   for (MachineBasicBlock *ChainBB : Chain) {
818     assert(BlockToChain[ChainBB] == &Chain);
819     for (MachineBasicBlock *Pred : ChainBB->predecessors()) {
820       if (BlockFilter && !BlockFilter->count(Pred))
821         continue;
822       if (BlockToChain[Pred] == &Chain)
823         continue;
824       ++Chain.UnscheduledPredecessors;
825     }
826   }
827 
828   if (Chain.UnscheduledPredecessors != 0)
829     return;
830 
831   MBB = *Chain.begin();
832   if (MBB->isEHPad())
833     EHPadWorkList.push_back(MBB);
834   else
835     BlockWorkList.push_back(MBB);
836 }
837 
buildChain(MachineBasicBlock * BB,BlockChain & Chain,const BlockFilterSet * BlockFilter)838 void MachineBlockPlacement::buildChain(
839     MachineBasicBlock *BB, BlockChain &Chain,
840     const BlockFilterSet *BlockFilter) {
841   assert(BB && "BB must not be null.\n");
842   assert(BlockToChain[BB] == &Chain && "BlockToChainMap mis-match.\n");
843   MachineFunction::iterator PrevUnplacedBlockIt = F->begin();
844 
845   MachineBasicBlock *LoopHeaderBB = BB;
846   markChainSuccessors(Chain, LoopHeaderBB, BlockFilter);
847   BB = *std::prev(Chain.end());
848   for (;;) {
849     assert(BB && "null block found at end of chain in loop.");
850     assert(BlockToChain[BB] == &Chain && "BlockToChainMap mis-match in loop.");
851     assert(*std::prev(Chain.end()) == BB && "BB Not found at end of chain.");
852 
853 
854     // Look for the best viable successor if there is one to place immediately
855     // after this block.
856     MachineBasicBlock *BestSucc = selectBestSuccessor(BB, Chain, BlockFilter);
857 
858     // If an immediate successor isn't available, look for the best viable
859     // block among those we've identified as not violating the loop's CFG at
860     // this point. This won't be a fallthrough, but it will increase locality.
861     if (!BestSucc)
862       BestSucc = selectBestCandidateBlock(Chain, BlockWorkList);
863     if (!BestSucc)
864       BestSucc = selectBestCandidateBlock(Chain, EHPadWorkList);
865 
866     if (!BestSucc) {
867       BestSucc = getFirstUnplacedBlock(Chain, PrevUnplacedBlockIt, BlockFilter);
868       if (!BestSucc)
869         break;
870 
871       DEBUG(dbgs() << "Unnatural loop CFG detected, forcibly merging the "
872                       "layout successor until the CFG reduces\n");
873     }
874 
875     // Place this block, updating the datastructures to reflect its placement.
876     BlockChain &SuccChain = *BlockToChain[BestSucc];
877     // Zero out UnscheduledPredecessors for the successor we're about to merge in case
878     // we selected a successor that didn't fit naturally into the CFG.
879     SuccChain.UnscheduledPredecessors = 0;
880     DEBUG(dbgs() << "Merging from " << getBlockName(BB) << " to "
881                  << getBlockName(BestSucc) << "\n");
882     markChainSuccessors(SuccChain, LoopHeaderBB, BlockFilter);
883     Chain.merge(BestSucc, &SuccChain);
884     BB = *std::prev(Chain.end());
885   }
886 
887   DEBUG(dbgs() << "Finished forming chain for header block "
888                << getBlockName(*Chain.begin()) << "\n");
889 }
890 
891 /// \brief Find the best loop top block for layout.
892 ///
893 /// Look for a block which is strictly better than the loop header for laying
894 /// out at the top of the loop. This looks for one and only one pattern:
895 /// a latch block with no conditional exit. This block will cause a conditional
896 /// jump around it or will be the bottom of the loop if we lay it out in place,
897 /// but if it it doesn't end up at the bottom of the loop for any reason,
898 /// rotation alone won't fix it. Because such a block will always result in an
899 /// unconditional jump (for the backedge) rotating it in front of the loop
900 /// header is always profitable.
901 MachineBasicBlock *
findBestLoopTop(MachineLoop & L,const BlockFilterSet & LoopBlockSet)902 MachineBlockPlacement::findBestLoopTop(MachineLoop &L,
903                                        const BlockFilterSet &LoopBlockSet) {
904   // Check that the header hasn't been fused with a preheader block due to
905   // crazy branches. If it has, we need to start with the header at the top to
906   // prevent pulling the preheader into the loop body.
907   BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
908   if (!LoopBlockSet.count(*HeaderChain.begin()))
909     return L.getHeader();
910 
911   DEBUG(dbgs() << "Finding best loop top for: " << getBlockName(L.getHeader())
912                << "\n");
913 
914   BlockFrequency BestPredFreq;
915   MachineBasicBlock *BestPred = nullptr;
916   for (MachineBasicBlock *Pred : L.getHeader()->predecessors()) {
917     if (!LoopBlockSet.count(Pred))
918       continue;
919     DEBUG(dbgs() << "    header pred: " << getBlockName(Pred) << ", "
920                  << Pred->succ_size() << " successors, ";
921           MBFI->printBlockFreq(dbgs(), Pred) << " freq\n");
922     if (Pred->succ_size() > 1)
923       continue;
924 
925     BlockFrequency PredFreq = MBFI->getBlockFreq(Pred);
926     if (!BestPred || PredFreq > BestPredFreq ||
927         (!(PredFreq < BestPredFreq) &&
928          Pred->isLayoutSuccessor(L.getHeader()))) {
929       BestPred = Pred;
930       BestPredFreq = PredFreq;
931     }
932   }
933 
934   // If no direct predecessor is fine, just use the loop header.
935   if (!BestPred) {
936     DEBUG(dbgs() << "    final top unchanged\n");
937     return L.getHeader();
938   }
939 
940   // Walk backwards through any straight line of predecessors.
941   while (BestPred->pred_size() == 1 &&
942          (*BestPred->pred_begin())->succ_size() == 1 &&
943          *BestPred->pred_begin() != L.getHeader())
944     BestPred = *BestPred->pred_begin();
945 
946   DEBUG(dbgs() << "    final top: " << getBlockName(BestPred) << "\n");
947   return BestPred;
948 }
949 
950 /// \brief Find the best loop exiting block for layout.
951 ///
952 /// This routine implements the logic to analyze the loop looking for the best
953 /// block to layout at the top of the loop. Typically this is done to maximize
954 /// fallthrough opportunities.
955 MachineBasicBlock *
findBestLoopExit(MachineLoop & L,const BlockFilterSet & LoopBlockSet)956 MachineBlockPlacement::findBestLoopExit(MachineLoop &L,
957                                         const BlockFilterSet &LoopBlockSet) {
958   // We don't want to layout the loop linearly in all cases. If the loop header
959   // is just a normal basic block in the loop, we want to look for what block
960   // within the loop is the best one to layout at the top. However, if the loop
961   // header has be pre-merged into a chain due to predecessors not having
962   // analyzable branches, *and* the predecessor it is merged with is *not* part
963   // of the loop, rotating the header into the middle of the loop will create
964   // a non-contiguous range of blocks which is Very Bad. So start with the
965   // header and only rotate if safe.
966   BlockChain &HeaderChain = *BlockToChain[L.getHeader()];
967   if (!LoopBlockSet.count(*HeaderChain.begin()))
968     return nullptr;
969 
970   BlockFrequency BestExitEdgeFreq;
971   unsigned BestExitLoopDepth = 0;
972   MachineBasicBlock *ExitingBB = nullptr;
973   // If there are exits to outer loops, loop rotation can severely limit
974   // fallthrough opportunites unless it selects such an exit. Keep a set of
975   // blocks where rotating to exit with that block will reach an outer loop.
976   SmallPtrSet<MachineBasicBlock *, 4> BlocksExitingToOuterLoop;
977 
978   DEBUG(dbgs() << "Finding best loop exit for: " << getBlockName(L.getHeader())
979                << "\n");
980   for (MachineBasicBlock *MBB : L.getBlocks()) {
981     BlockChain &Chain = *BlockToChain[MBB];
982     // Ensure that this block is at the end of a chain; otherwise it could be
983     // mid-way through an inner loop or a successor of an unanalyzable branch.
984     if (MBB != *std::prev(Chain.end()))
985       continue;
986 
987     // Now walk the successors. We need to establish whether this has a viable
988     // exiting successor and whether it has a viable non-exiting successor.
989     // We store the old exiting state and restore it if a viable looping
990     // successor isn't found.
991     MachineBasicBlock *OldExitingBB = ExitingBB;
992     BlockFrequency OldBestExitEdgeFreq = BestExitEdgeFreq;
993     bool HasLoopingSucc = false;
994     for (MachineBasicBlock *Succ : MBB->successors()) {
995       if (Succ->isEHPad())
996         continue;
997       if (Succ == MBB)
998         continue;
999       BlockChain &SuccChain = *BlockToChain[Succ];
1000       // Don't split chains, either this chain or the successor's chain.
1001       if (&Chain == &SuccChain) {
1002         DEBUG(dbgs() << "    exiting: " << getBlockName(MBB) << " -> "
1003                      << getBlockName(Succ) << " (chain conflict)\n");
1004         continue;
1005       }
1006 
1007       auto SuccProb = MBPI->getEdgeProbability(MBB, Succ);
1008       if (LoopBlockSet.count(Succ)) {
1009         DEBUG(dbgs() << "    looping: " << getBlockName(MBB) << " -> "
1010                      << getBlockName(Succ) << " (" << SuccProb << ")\n");
1011         HasLoopingSucc = true;
1012         continue;
1013       }
1014 
1015       unsigned SuccLoopDepth = 0;
1016       if (MachineLoop *ExitLoop = MLI->getLoopFor(Succ)) {
1017         SuccLoopDepth = ExitLoop->getLoopDepth();
1018         if (ExitLoop->contains(&L))
1019           BlocksExitingToOuterLoop.insert(MBB);
1020       }
1021 
1022       BlockFrequency ExitEdgeFreq = MBFI->getBlockFreq(MBB) * SuccProb;
1023       DEBUG(dbgs() << "    exiting: " << getBlockName(MBB) << " -> "
1024                    << getBlockName(Succ) << " [L:" << SuccLoopDepth << "] (";
1025             MBFI->printBlockFreq(dbgs(), ExitEdgeFreq) << ")\n");
1026       // Note that we bias this toward an existing layout successor to retain
1027       // incoming order in the absence of better information. The exit must have
1028       // a frequency higher than the current exit before we consider breaking
1029       // the layout.
1030       BranchProbability Bias(100 - ExitBlockBias, 100);
1031       if (!ExitingBB || SuccLoopDepth > BestExitLoopDepth ||
1032           ExitEdgeFreq > BestExitEdgeFreq ||
1033           (MBB->isLayoutSuccessor(Succ) &&
1034            !(ExitEdgeFreq < BestExitEdgeFreq * Bias))) {
1035         BestExitEdgeFreq = ExitEdgeFreq;
1036         ExitingBB = MBB;
1037       }
1038     }
1039 
1040     if (!HasLoopingSucc) {
1041       // Restore the old exiting state, no viable looping successor was found.
1042       ExitingBB = OldExitingBB;
1043       BestExitEdgeFreq = OldBestExitEdgeFreq;
1044     }
1045   }
1046   // Without a candidate exiting block or with only a single block in the
1047   // loop, just use the loop header to layout the loop.
1048   if (!ExitingBB || L.getNumBlocks() == 1)
1049     return nullptr;
1050 
1051   // Also, if we have exit blocks which lead to outer loops but didn't select
1052   // one of them as the exiting block we are rotating toward, disable loop
1053   // rotation altogether.
1054   if (!BlocksExitingToOuterLoop.empty() &&
1055       !BlocksExitingToOuterLoop.count(ExitingBB))
1056     return nullptr;
1057 
1058   DEBUG(dbgs() << "  Best exiting block: " << getBlockName(ExitingBB) << "\n");
1059   return ExitingBB;
1060 }
1061 
1062 /// \brief Attempt to rotate an exiting block to the bottom of the loop.
1063 ///
1064 /// Once we have built a chain, try to rotate it to line up the hot exit block
1065 /// with fallthrough out of the loop if doing so doesn't introduce unnecessary
1066 /// branches. For example, if the loop has fallthrough into its header and out
1067 /// of its bottom already, don't rotate it.
rotateLoop(BlockChain & LoopChain,MachineBasicBlock * ExitingBB,const BlockFilterSet & LoopBlockSet)1068 void MachineBlockPlacement::rotateLoop(BlockChain &LoopChain,
1069                                        MachineBasicBlock *ExitingBB,
1070                                        const BlockFilterSet &LoopBlockSet) {
1071   if (!ExitingBB)
1072     return;
1073 
1074   MachineBasicBlock *Top = *LoopChain.begin();
1075   bool ViableTopFallthrough = false;
1076   for (MachineBasicBlock *Pred : Top->predecessors()) {
1077     BlockChain *PredChain = BlockToChain[Pred];
1078     if (!LoopBlockSet.count(Pred) &&
1079         (!PredChain || Pred == *std::prev(PredChain->end()))) {
1080       ViableTopFallthrough = true;
1081       break;
1082     }
1083   }
1084 
1085   // If the header has viable fallthrough, check whether the current loop
1086   // bottom is a viable exiting block. If so, bail out as rotating will
1087   // introduce an unnecessary branch.
1088   if (ViableTopFallthrough) {
1089     MachineBasicBlock *Bottom = *std::prev(LoopChain.end());
1090     for (MachineBasicBlock *Succ : Bottom->successors()) {
1091       BlockChain *SuccChain = BlockToChain[Succ];
1092       if (!LoopBlockSet.count(Succ) &&
1093           (!SuccChain || Succ == *SuccChain->begin()))
1094         return;
1095     }
1096   }
1097 
1098   BlockChain::iterator ExitIt =
1099       std::find(LoopChain.begin(), LoopChain.end(), ExitingBB);
1100   if (ExitIt == LoopChain.end())
1101     return;
1102 
1103   std::rotate(LoopChain.begin(), std::next(ExitIt), LoopChain.end());
1104 }
1105 
1106 /// \brief Attempt to rotate a loop based on profile data to reduce branch cost.
1107 ///
1108 /// With profile data, we can determine the cost in terms of missed fall through
1109 /// opportunities when rotating a loop chain and select the best rotation.
1110 /// Basically, there are three kinds of cost to consider for each rotation:
1111 ///    1. The possibly missed fall through edge (if it exists) from BB out of
1112 ///    the loop to the loop header.
1113 ///    2. The possibly missed fall through edges (if they exist) from the loop
1114 ///    exits to BB out of the loop.
1115 ///    3. The missed fall through edge (if it exists) from the last BB to the
1116 ///    first BB in the loop chain.
1117 ///  Therefore, the cost for a given rotation is the sum of costs listed above.
1118 ///  We select the best rotation with the smallest cost.
rotateLoopWithProfile(BlockChain & LoopChain,MachineLoop & L,const BlockFilterSet & LoopBlockSet)1119 void MachineBlockPlacement::rotateLoopWithProfile(
1120     BlockChain &LoopChain, MachineLoop &L, const BlockFilterSet &LoopBlockSet) {
1121   auto HeaderBB = L.getHeader();
1122   auto HeaderIter = std::find(LoopChain.begin(), LoopChain.end(), HeaderBB);
1123   auto RotationPos = LoopChain.end();
1124 
1125   BlockFrequency SmallestRotationCost = BlockFrequency::getMaxFrequency();
1126 
1127   // A utility lambda that scales up a block frequency by dividing it by a
1128   // branch probability which is the reciprocal of the scale.
1129   auto ScaleBlockFrequency = [](BlockFrequency Freq,
1130                                 unsigned Scale) -> BlockFrequency {
1131     if (Scale == 0)
1132       return 0;
1133     // Use operator / between BlockFrequency and BranchProbability to implement
1134     // saturating multiplication.
1135     return Freq / BranchProbability(1, Scale);
1136   };
1137 
1138   // Compute the cost of the missed fall-through edge to the loop header if the
1139   // chain head is not the loop header. As we only consider natural loops with
1140   // single header, this computation can be done only once.
1141   BlockFrequency HeaderFallThroughCost(0);
1142   for (auto *Pred : HeaderBB->predecessors()) {
1143     BlockChain *PredChain = BlockToChain[Pred];
1144     if (!LoopBlockSet.count(Pred) &&
1145         (!PredChain || Pred == *std::prev(PredChain->end()))) {
1146       auto EdgeFreq =
1147           MBFI->getBlockFreq(Pred) * MBPI->getEdgeProbability(Pred, HeaderBB);
1148       auto FallThruCost = ScaleBlockFrequency(EdgeFreq, MisfetchCost);
1149       // If the predecessor has only an unconditional jump to the header, we
1150       // need to consider the cost of this jump.
1151       if (Pred->succ_size() == 1)
1152         FallThruCost += ScaleBlockFrequency(EdgeFreq, JumpInstCost);
1153       HeaderFallThroughCost = std::max(HeaderFallThroughCost, FallThruCost);
1154     }
1155   }
1156 
1157   // Here we collect all exit blocks in the loop, and for each exit we find out
1158   // its hottest exit edge. For each loop rotation, we define the loop exit cost
1159   // as the sum of frequencies of exit edges we collect here, excluding the exit
1160   // edge from the tail of the loop chain.
1161   SmallVector<std::pair<MachineBasicBlock *, BlockFrequency>, 4> ExitsWithFreq;
1162   for (auto BB : LoopChain) {
1163     auto LargestExitEdgeProb = BranchProbability::getZero();
1164     for (auto *Succ : BB->successors()) {
1165       BlockChain *SuccChain = BlockToChain[Succ];
1166       if (!LoopBlockSet.count(Succ) &&
1167           (!SuccChain || Succ == *SuccChain->begin())) {
1168         auto SuccProb = MBPI->getEdgeProbability(BB, Succ);
1169         LargestExitEdgeProb = std::max(LargestExitEdgeProb, SuccProb);
1170       }
1171     }
1172     if (LargestExitEdgeProb > BranchProbability::getZero()) {
1173       auto ExitFreq = MBFI->getBlockFreq(BB) * LargestExitEdgeProb;
1174       ExitsWithFreq.emplace_back(BB, ExitFreq);
1175     }
1176   }
1177 
1178   // In this loop we iterate every block in the loop chain and calculate the
1179   // cost assuming the block is the head of the loop chain. When the loop ends,
1180   // we should have found the best candidate as the loop chain's head.
1181   for (auto Iter = LoopChain.begin(), TailIter = std::prev(LoopChain.end()),
1182             EndIter = LoopChain.end();
1183        Iter != EndIter; Iter++, TailIter++) {
1184     // TailIter is used to track the tail of the loop chain if the block we are
1185     // checking (pointed by Iter) is the head of the chain.
1186     if (TailIter == LoopChain.end())
1187       TailIter = LoopChain.begin();
1188 
1189     auto TailBB = *TailIter;
1190 
1191     // Calculate the cost by putting this BB to the top.
1192     BlockFrequency Cost = 0;
1193 
1194     // If the current BB is the loop header, we need to take into account the
1195     // cost of the missed fall through edge from outside of the loop to the
1196     // header.
1197     if (Iter != HeaderIter)
1198       Cost += HeaderFallThroughCost;
1199 
1200     // Collect the loop exit cost by summing up frequencies of all exit edges
1201     // except the one from the chain tail.
1202     for (auto &ExitWithFreq : ExitsWithFreq)
1203       if (TailBB != ExitWithFreq.first)
1204         Cost += ExitWithFreq.second;
1205 
1206     // The cost of breaking the once fall-through edge from the tail to the top
1207     // of the loop chain. Here we need to consider three cases:
1208     // 1. If the tail node has only one successor, then we will get an
1209     //    additional jmp instruction. So the cost here is (MisfetchCost +
1210     //    JumpInstCost) * tail node frequency.
1211     // 2. If the tail node has two successors, then we may still get an
1212     //    additional jmp instruction if the layout successor after the loop
1213     //    chain is not its CFG successor. Note that the more frequently executed
1214     //    jmp instruction will be put ahead of the other one. Assume the
1215     //    frequency of those two branches are x and y, where x is the frequency
1216     //    of the edge to the chain head, then the cost will be
1217     //    (x * MisfetechCost + min(x, y) * JumpInstCost) * tail node frequency.
1218     // 3. If the tail node has more than two successors (this rarely happens),
1219     //    we won't consider any additional cost.
1220     if (TailBB->isSuccessor(*Iter)) {
1221       auto TailBBFreq = MBFI->getBlockFreq(TailBB);
1222       if (TailBB->succ_size() == 1)
1223         Cost += ScaleBlockFrequency(TailBBFreq.getFrequency(),
1224                                     MisfetchCost + JumpInstCost);
1225       else if (TailBB->succ_size() == 2) {
1226         auto TailToHeadProb = MBPI->getEdgeProbability(TailBB, *Iter);
1227         auto TailToHeadFreq = TailBBFreq * TailToHeadProb;
1228         auto ColderEdgeFreq = TailToHeadProb > BranchProbability(1, 2)
1229                                   ? TailBBFreq * TailToHeadProb.getCompl()
1230                                   : TailToHeadFreq;
1231         Cost += ScaleBlockFrequency(TailToHeadFreq, MisfetchCost) +
1232                 ScaleBlockFrequency(ColderEdgeFreq, JumpInstCost);
1233       }
1234     }
1235 
1236     DEBUG(dbgs() << "The cost of loop rotation by making " << getBlockName(*Iter)
1237                  << " to the top: " << Cost.getFrequency() << "\n");
1238 
1239     if (Cost < SmallestRotationCost) {
1240       SmallestRotationCost = Cost;
1241       RotationPos = Iter;
1242     }
1243   }
1244 
1245   if (RotationPos != LoopChain.end()) {
1246     DEBUG(dbgs() << "Rotate loop by making " << getBlockName(*RotationPos)
1247                  << " to the top\n");
1248     std::rotate(LoopChain.begin(), RotationPos, LoopChain.end());
1249   }
1250 }
1251 
1252 /// \brief Collect blocks in the given loop that are to be placed.
1253 ///
1254 /// When profile data is available, exclude cold blocks from the returned set;
1255 /// otherwise, collect all blocks in the loop.
1256 MachineBlockPlacement::BlockFilterSet
collectLoopBlockSet(MachineLoop & L)1257 MachineBlockPlacement::collectLoopBlockSet(MachineLoop &L) {
1258   BlockFilterSet LoopBlockSet;
1259 
1260   // Filter cold blocks off from LoopBlockSet when profile data is available.
1261   // Collect the sum of frequencies of incoming edges to the loop header from
1262   // outside. If we treat the loop as a super block, this is the frequency of
1263   // the loop. Then for each block in the loop, we calculate the ratio between
1264   // its frequency and the frequency of the loop block. When it is too small,
1265   // don't add it to the loop chain. If there are outer loops, then this block
1266   // will be merged into the first outer loop chain for which this block is not
1267   // cold anymore. This needs precise profile data and we only do this when
1268   // profile data is available.
1269   if (F->getFunction()->getEntryCount()) {
1270     BlockFrequency LoopFreq(0);
1271     for (auto LoopPred : L.getHeader()->predecessors())
1272       if (!L.contains(LoopPred))
1273         LoopFreq += MBFI->getBlockFreq(LoopPred) *
1274                     MBPI->getEdgeProbability(LoopPred, L.getHeader());
1275 
1276     for (MachineBasicBlock *LoopBB : L.getBlocks()) {
1277       auto Freq = MBFI->getBlockFreq(LoopBB).getFrequency();
1278       if (Freq == 0 || LoopFreq.getFrequency() / Freq > LoopToColdBlockRatio)
1279         continue;
1280       LoopBlockSet.insert(LoopBB);
1281     }
1282   } else
1283     LoopBlockSet.insert(L.block_begin(), L.block_end());
1284 
1285   return LoopBlockSet;
1286 }
1287 
1288 /// \brief Forms basic block chains from the natural loop structures.
1289 ///
1290 /// These chains are designed to preserve the existing *structure* of the code
1291 /// as much as possible. We can then stitch the chains together in a way which
1292 /// both preserves the topological structure and minimizes taken conditional
1293 /// branches.
buildLoopChains(MachineLoop & L)1294 void MachineBlockPlacement::buildLoopChains(MachineLoop &L) {
1295   // First recurse through any nested loops, building chains for those inner
1296   // loops.
1297   for (MachineLoop *InnerLoop : L)
1298     buildLoopChains(*InnerLoop);
1299 
1300   assert(BlockWorkList.empty());
1301   assert(EHPadWorkList.empty());
1302   BlockFilterSet LoopBlockSet = collectLoopBlockSet(L);
1303 
1304   // Check if we have profile data for this function. If yes, we will rotate
1305   // this loop by modeling costs more precisely which requires the profile data
1306   // for better layout.
1307   bool RotateLoopWithProfile =
1308       ForcePreciseRotationCost ||
1309       (PreciseRotationCost && F->getFunction()->getEntryCount());
1310 
1311   // First check to see if there is an obviously preferable top block for the
1312   // loop. This will default to the header, but may end up as one of the
1313   // predecessors to the header if there is one which will result in strictly
1314   // fewer branches in the loop body.
1315   // When we use profile data to rotate the loop, this is unnecessary.
1316   MachineBasicBlock *LoopTop =
1317       RotateLoopWithProfile ? L.getHeader() : findBestLoopTop(L, LoopBlockSet);
1318 
1319   // If we selected just the header for the loop top, look for a potentially
1320   // profitable exit block in the event that rotating the loop can eliminate
1321   // branches by placing an exit edge at the bottom.
1322   MachineBasicBlock *ExitingBB = nullptr;
1323   if (!RotateLoopWithProfile && LoopTop == L.getHeader())
1324     ExitingBB = findBestLoopExit(L, LoopBlockSet);
1325 
1326   BlockChain &LoopChain = *BlockToChain[LoopTop];
1327 
1328   // FIXME: This is a really lame way of walking the chains in the loop: we
1329   // walk the blocks, and use a set to prevent visiting a particular chain
1330   // twice.
1331   SmallPtrSet<BlockChain *, 4> UpdatedPreds;
1332   assert(LoopChain.UnscheduledPredecessors == 0);
1333   UpdatedPreds.insert(&LoopChain);
1334 
1335   for (MachineBasicBlock *LoopBB : LoopBlockSet)
1336     fillWorkLists(LoopBB, UpdatedPreds, &LoopBlockSet);
1337 
1338   buildChain(LoopTop, LoopChain, &LoopBlockSet);
1339 
1340   if (RotateLoopWithProfile)
1341     rotateLoopWithProfile(LoopChain, L, LoopBlockSet);
1342   else
1343     rotateLoop(LoopChain, ExitingBB, LoopBlockSet);
1344 
1345   DEBUG({
1346     // Crash at the end so we get all of the debugging output first.
1347     bool BadLoop = false;
1348     if (LoopChain.UnscheduledPredecessors) {
1349       BadLoop = true;
1350       dbgs() << "Loop chain contains a block without its preds placed!\n"
1351              << "  Loop header:  " << getBlockName(*L.block_begin()) << "\n"
1352              << "  Chain header: " << getBlockName(*LoopChain.begin()) << "\n";
1353     }
1354     for (MachineBasicBlock *ChainBB : LoopChain) {
1355       dbgs() << "          ... " << getBlockName(ChainBB) << "\n";
1356       if (!LoopBlockSet.erase(ChainBB)) {
1357         // We don't mark the loop as bad here because there are real situations
1358         // where this can occur. For example, with an unanalyzable fallthrough
1359         // from a loop block to a non-loop block or vice versa.
1360         dbgs() << "Loop chain contains a block not contained by the loop!\n"
1361                << "  Loop header:  " << getBlockName(*L.block_begin()) << "\n"
1362                << "  Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
1363                << "  Bad block:    " << getBlockName(ChainBB) << "\n";
1364       }
1365     }
1366 
1367     if (!LoopBlockSet.empty()) {
1368       BadLoop = true;
1369       for (MachineBasicBlock *LoopBB : LoopBlockSet)
1370         dbgs() << "Loop contains blocks never placed into a chain!\n"
1371                << "  Loop header:  " << getBlockName(*L.block_begin()) << "\n"
1372                << "  Chain header: " << getBlockName(*LoopChain.begin()) << "\n"
1373                << "  Bad block:    " << getBlockName(LoopBB) << "\n";
1374     }
1375     assert(!BadLoop && "Detected problems with the placement of this loop.");
1376   });
1377 
1378   BlockWorkList.clear();
1379   EHPadWorkList.clear();
1380 }
1381 
1382 /// When OutlineOpitonalBranches is on, this method colects BBs that
1383 /// dominates all terminator blocks of the function \p F.
collectMustExecuteBBs()1384 void MachineBlockPlacement::collectMustExecuteBBs() {
1385   if (OutlineOptionalBranches) {
1386     // Find the nearest common dominator of all of F's terminators.
1387     MachineBasicBlock *Terminator = nullptr;
1388     for (MachineBasicBlock &MBB : *F) {
1389       if (MBB.succ_size() == 0) {
1390         if (Terminator == nullptr)
1391           Terminator = &MBB;
1392         else
1393           Terminator = MDT->findNearestCommonDominator(Terminator, &MBB);
1394       }
1395     }
1396 
1397     // MBBs dominating this common dominator are unavoidable.
1398     UnavoidableBlocks.clear();
1399     for (MachineBasicBlock &MBB : *F) {
1400       if (MDT->dominates(&MBB, Terminator)) {
1401         UnavoidableBlocks.insert(&MBB);
1402       }
1403     }
1404   }
1405 }
1406 
buildCFGChains()1407 void MachineBlockPlacement::buildCFGChains() {
1408   // Ensure that every BB in the function has an associated chain to simplify
1409   // the assumptions of the remaining algorithm.
1410   SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
1411   for (MachineFunction::iterator FI = F->begin(), FE = F->end(); FI != FE;
1412        ++FI) {
1413     MachineBasicBlock *BB = &*FI;
1414     BlockChain *Chain =
1415         new (ChainAllocator.Allocate()) BlockChain(BlockToChain, BB);
1416     // Also, merge any blocks which we cannot reason about and must preserve
1417     // the exact fallthrough behavior for.
1418     for (;;) {
1419       Cond.clear();
1420       MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
1421       if (!TII->analyzeBranch(*BB, TBB, FBB, Cond) || !FI->canFallThrough())
1422         break;
1423 
1424       MachineFunction::iterator NextFI = std::next(FI);
1425       MachineBasicBlock *NextBB = &*NextFI;
1426       // Ensure that the layout successor is a viable block, as we know that
1427       // fallthrough is a possibility.
1428       assert(NextFI != FE && "Can't fallthrough past the last block.");
1429       DEBUG(dbgs() << "Pre-merging due to unanalyzable fallthrough: "
1430                    << getBlockName(BB) << " -> " << getBlockName(NextBB)
1431                    << "\n");
1432       Chain->merge(NextBB, nullptr);
1433       FI = NextFI;
1434       BB = NextBB;
1435     }
1436   }
1437 
1438   // Turned on with OutlineOptionalBranches option
1439   collectMustExecuteBBs();
1440 
1441   // Build any loop-based chains.
1442   for (MachineLoop *L : *MLI)
1443     buildLoopChains(*L);
1444 
1445   assert(BlockWorkList.empty());
1446   assert(EHPadWorkList.empty());
1447 
1448   SmallPtrSet<BlockChain *, 4> UpdatedPreds;
1449   for (MachineBasicBlock &MBB : *F)
1450     fillWorkLists(&MBB, UpdatedPreds);
1451 
1452   BlockChain &FunctionChain = *BlockToChain[&F->front()];
1453   buildChain(&F->front(), FunctionChain);
1454 
1455 #ifndef NDEBUG
1456   typedef SmallPtrSet<MachineBasicBlock *, 16> FunctionBlockSetType;
1457 #endif
1458   DEBUG({
1459     // Crash at the end so we get all of the debugging output first.
1460     bool BadFunc = false;
1461     FunctionBlockSetType FunctionBlockSet;
1462     for (MachineBasicBlock &MBB : *F)
1463       FunctionBlockSet.insert(&MBB);
1464 
1465     for (MachineBasicBlock *ChainBB : FunctionChain)
1466       if (!FunctionBlockSet.erase(ChainBB)) {
1467         BadFunc = true;
1468         dbgs() << "Function chain contains a block not in the function!\n"
1469                << "  Bad block:    " << getBlockName(ChainBB) << "\n";
1470       }
1471 
1472     if (!FunctionBlockSet.empty()) {
1473       BadFunc = true;
1474       for (MachineBasicBlock *RemainingBB : FunctionBlockSet)
1475         dbgs() << "Function contains blocks never placed into a chain!\n"
1476                << "  Bad block:    " << getBlockName(RemainingBB) << "\n";
1477     }
1478     assert(!BadFunc && "Detected problems with the block placement.");
1479   });
1480 
1481   // Splice the blocks into place.
1482   MachineFunction::iterator InsertPos = F->begin();
1483   DEBUG(dbgs() << "[MBP] Function: "<< F->getName() << "\n");
1484   for (MachineBasicBlock *ChainBB : FunctionChain) {
1485     DEBUG(dbgs() << (ChainBB == *FunctionChain.begin() ? "Placing chain "
1486                                                        : "          ... ")
1487                  << getBlockName(ChainBB) << "\n");
1488     if (InsertPos != MachineFunction::iterator(ChainBB))
1489       F->splice(InsertPos, ChainBB);
1490     else
1491       ++InsertPos;
1492 
1493     // Update the terminator of the previous block.
1494     if (ChainBB == *FunctionChain.begin())
1495       continue;
1496     MachineBasicBlock *PrevBB = &*std::prev(MachineFunction::iterator(ChainBB));
1497 
1498     // FIXME: It would be awesome of updateTerminator would just return rather
1499     // than assert when the branch cannot be analyzed in order to remove this
1500     // boiler plate.
1501     Cond.clear();
1502     MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
1503 
1504     // The "PrevBB" is not yet updated to reflect current code layout, so,
1505     //   o. it may fall-through to a block without explict "goto" instruction
1506     //      before layout, and no longer fall-through it after layout; or
1507     //   o. just opposite.
1508     //
1509     // analyzeBranch() may return erroneous value for FBB when these two
1510     // situations take place. For the first scenario FBB is mistakenly set NULL;
1511     // for the 2nd scenario, the FBB, which is expected to be NULL, is
1512     // mistakenly pointing to "*BI".
1513     // Thus, if the future change needs to use FBB before the layout is set, it
1514     // has to correct FBB first by using the code similar to the following:
1515     //
1516     // if (!Cond.empty() && (!FBB || FBB == ChainBB)) {
1517     //   PrevBB->updateTerminator();
1518     //   Cond.clear();
1519     //   TBB = FBB = nullptr;
1520     //   if (TII->analyzeBranch(*PrevBB, TBB, FBB, Cond)) {
1521     //     // FIXME: This should never take place.
1522     //     TBB = FBB = nullptr;
1523     //   }
1524     // }
1525     if (!TII->analyzeBranch(*PrevBB, TBB, FBB, Cond))
1526       PrevBB->updateTerminator();
1527   }
1528 
1529   // Fixup the last block.
1530   Cond.clear();
1531   MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
1532   if (!TII->analyzeBranch(F->back(), TBB, FBB, Cond))
1533     F->back().updateTerminator();
1534 
1535   BlockWorkList.clear();
1536   EHPadWorkList.clear();
1537 }
1538 
optimizeBranches()1539 void MachineBlockPlacement::optimizeBranches() {
1540   BlockChain &FunctionChain = *BlockToChain[&F->front()];
1541   SmallVector<MachineOperand, 4> Cond; // For AnalyzeBranch.
1542 
1543   // Now that all the basic blocks in the chain have the proper layout,
1544   // make a final call to AnalyzeBranch with AllowModify set.
1545   // Indeed, the target may be able to optimize the branches in a way we
1546   // cannot because all branches may not be analyzable.
1547   // E.g., the target may be able to remove an unconditional branch to
1548   // a fallthrough when it occurs after predicated terminators.
1549   for (MachineBasicBlock *ChainBB : FunctionChain) {
1550     Cond.clear();
1551     MachineBasicBlock *TBB = nullptr, *FBB = nullptr; // For AnalyzeBranch.
1552     if (!TII->analyzeBranch(*ChainBB, TBB, FBB, Cond, /*AllowModify*/ true)) {
1553       // If PrevBB has a two-way branch, try to re-order the branches
1554       // such that we branch to the successor with higher probability first.
1555       if (TBB && !Cond.empty() && FBB &&
1556           MBPI->getEdgeProbability(ChainBB, FBB) >
1557               MBPI->getEdgeProbability(ChainBB, TBB) &&
1558           !TII->ReverseBranchCondition(Cond)) {
1559         DEBUG(dbgs() << "Reverse order of the two branches: "
1560                      << getBlockName(ChainBB) << "\n");
1561         DEBUG(dbgs() << "    Edge probability: "
1562                      << MBPI->getEdgeProbability(ChainBB, FBB) << " vs "
1563                      << MBPI->getEdgeProbability(ChainBB, TBB) << "\n");
1564         DebugLoc dl; // FIXME: this is nowhere
1565         TII->RemoveBranch(*ChainBB);
1566         TII->InsertBranch(*ChainBB, FBB, TBB, Cond, dl);
1567         ChainBB->updateTerminator();
1568       }
1569     }
1570   }
1571 }
1572 
alignBlocks()1573 void MachineBlockPlacement::alignBlocks() {
1574   // Walk through the backedges of the function now that we have fully laid out
1575   // the basic blocks and align the destination of each backedge. We don't rely
1576   // exclusively on the loop info here so that we can align backedges in
1577   // unnatural CFGs and backedges that were introduced purely because of the
1578   // loop rotations done during this layout pass.
1579   if (F->getFunction()->optForSize())
1580     return;
1581   BlockChain &FunctionChain = *BlockToChain[&F->front()];
1582   if (FunctionChain.begin() == FunctionChain.end())
1583     return; // Empty chain.
1584 
1585   const BranchProbability ColdProb(1, 5); // 20%
1586   BlockFrequency EntryFreq = MBFI->getBlockFreq(&F->front());
1587   BlockFrequency WeightedEntryFreq = EntryFreq * ColdProb;
1588   for (MachineBasicBlock *ChainBB : FunctionChain) {
1589     if (ChainBB == *FunctionChain.begin())
1590       continue;
1591 
1592     // Don't align non-looping basic blocks. These are unlikely to execute
1593     // enough times to matter in practice. Note that we'll still handle
1594     // unnatural CFGs inside of a natural outer loop (the common case) and
1595     // rotated loops.
1596     MachineLoop *L = MLI->getLoopFor(ChainBB);
1597     if (!L)
1598       continue;
1599 
1600     unsigned Align = TLI->getPrefLoopAlignment(L);
1601     if (!Align)
1602       continue; // Don't care about loop alignment.
1603 
1604     // If the block is cold relative to the function entry don't waste space
1605     // aligning it.
1606     BlockFrequency Freq = MBFI->getBlockFreq(ChainBB);
1607     if (Freq < WeightedEntryFreq)
1608       continue;
1609 
1610     // If the block is cold relative to its loop header, don't align it
1611     // regardless of what edges into the block exist.
1612     MachineBasicBlock *LoopHeader = L->getHeader();
1613     BlockFrequency LoopHeaderFreq = MBFI->getBlockFreq(LoopHeader);
1614     if (Freq < (LoopHeaderFreq * ColdProb))
1615       continue;
1616 
1617     // Check for the existence of a non-layout predecessor which would benefit
1618     // from aligning this block.
1619     MachineBasicBlock *LayoutPred =
1620         &*std::prev(MachineFunction::iterator(ChainBB));
1621 
1622     // Force alignment if all the predecessors are jumps. We already checked
1623     // that the block isn't cold above.
1624     if (!LayoutPred->isSuccessor(ChainBB)) {
1625       ChainBB->setAlignment(Align);
1626       continue;
1627     }
1628 
1629     // Align this block if the layout predecessor's edge into this block is
1630     // cold relative to the block. When this is true, other predecessors make up
1631     // all of the hot entries into the block and thus alignment is likely to be
1632     // important.
1633     BranchProbability LayoutProb =
1634         MBPI->getEdgeProbability(LayoutPred, ChainBB);
1635     BlockFrequency LayoutEdgeFreq = MBFI->getBlockFreq(LayoutPred) * LayoutProb;
1636     if (LayoutEdgeFreq <= (Freq * ColdProb))
1637       ChainBB->setAlignment(Align);
1638   }
1639 }
1640 
runOnMachineFunction(MachineFunction & MF)1641 bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &MF) {
1642   if (skipFunction(*MF.getFunction()))
1643     return false;
1644 
1645   // Check for single-block functions and skip them.
1646   if (std::next(MF.begin()) == MF.end())
1647     return false;
1648 
1649   F = &MF;
1650   MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
1651   MBFI = llvm::make_unique<BranchFolder::MBFIWrapper>(
1652       getAnalysis<MachineBlockFrequencyInfo>());
1653   MLI = &getAnalysis<MachineLoopInfo>();
1654   TII = MF.getSubtarget().getInstrInfo();
1655   TLI = MF.getSubtarget().getTargetLowering();
1656   MDT = &getAnalysis<MachineDominatorTree>();
1657   assert(BlockToChain.empty());
1658 
1659   buildCFGChains();
1660 
1661   // Changing the layout can create new tail merging opportunities.
1662   TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
1663   // TailMerge can create jump into if branches that make CFG irreducible for
1664   // HW that requires structurized CFG.
1665   bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
1666                          PassConfig->getEnableTailMerge() &&
1667                          BranchFoldPlacement;
1668   // No tail merging opportunities if the block number is less than four.
1669   if (MF.size() > 3 && EnableTailMerge) {
1670     BranchFolder BF(/*EnableTailMerge=*/true, /*CommonHoist=*/false, *MBFI,
1671                     *MBPI);
1672 
1673     if (BF.OptimizeFunction(MF, TII, MF.getSubtarget().getRegisterInfo(),
1674                             getAnalysisIfAvailable<MachineModuleInfo>(), MLI,
1675                             /*AfterBlockPlacement=*/true)) {
1676       // Redo the layout if tail merging creates/removes/moves blocks.
1677       BlockToChain.clear();
1678       ChainAllocator.DestroyAll();
1679       buildCFGChains();
1680     }
1681   }
1682 
1683   optimizeBranches();
1684   alignBlocks();
1685 
1686   BlockToChain.clear();
1687   ChainAllocator.DestroyAll();
1688 
1689   if (AlignAllBlock)
1690     // Align all of the blocks in the function to a specific alignment.
1691     for (MachineBasicBlock &MBB : MF)
1692       MBB.setAlignment(AlignAllBlock);
1693   else if (AlignAllNonFallThruBlocks) {
1694     // Align all of the blocks that have no fall-through predecessors to a
1695     // specific alignment.
1696     for (auto MBI = std::next(MF.begin()), MBE = MF.end(); MBI != MBE; ++MBI) {
1697       auto LayoutPred = std::prev(MBI);
1698       if (!LayoutPred->isSuccessor(&*MBI))
1699         MBI->setAlignment(AlignAllNonFallThruBlocks);
1700     }
1701   }
1702 
1703   // We always return true as we have no way to track whether the final order
1704   // differs from the original order.
1705   return true;
1706 }
1707 
1708 namespace {
1709 /// \brief A pass to compute block placement statistics.
1710 ///
1711 /// A separate pass to compute interesting statistics for evaluating block
1712 /// placement. This is separate from the actual placement pass so that they can
1713 /// be computed in the absence of any placement transformations or when using
1714 /// alternative placement strategies.
1715 class MachineBlockPlacementStats : public MachineFunctionPass {
1716   /// \brief A handle to the branch probability pass.
1717   const MachineBranchProbabilityInfo *MBPI;
1718 
1719   /// \brief A handle to the function-wide block frequency pass.
1720   const MachineBlockFrequencyInfo *MBFI;
1721 
1722 public:
1723   static char ID; // Pass identification, replacement for typeid
MachineBlockPlacementStats()1724   MachineBlockPlacementStats() : MachineFunctionPass(ID) {
1725     initializeMachineBlockPlacementStatsPass(*PassRegistry::getPassRegistry());
1726   }
1727 
1728   bool runOnMachineFunction(MachineFunction &F) override;
1729 
getAnalysisUsage(AnalysisUsage & AU) const1730   void getAnalysisUsage(AnalysisUsage &AU) const override {
1731     AU.addRequired<MachineBranchProbabilityInfo>();
1732     AU.addRequired<MachineBlockFrequencyInfo>();
1733     AU.setPreservesAll();
1734     MachineFunctionPass::getAnalysisUsage(AU);
1735   }
1736 };
1737 }
1738 
1739 char MachineBlockPlacementStats::ID = 0;
1740 char &llvm::MachineBlockPlacementStatsID = MachineBlockPlacementStats::ID;
1741 INITIALIZE_PASS_BEGIN(MachineBlockPlacementStats, "block-placement-stats",
1742                       "Basic Block Placement Stats", false, false)
INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)1743 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
1744 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
1745 INITIALIZE_PASS_END(MachineBlockPlacementStats, "block-placement-stats",
1746                     "Basic Block Placement Stats", false, false)
1747 
1748 bool MachineBlockPlacementStats::runOnMachineFunction(MachineFunction &F) {
1749   // Check for single-block functions and skip them.
1750   if (std::next(F.begin()) == F.end())
1751     return false;
1752 
1753   MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
1754   MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
1755 
1756   for (MachineBasicBlock &MBB : F) {
1757     BlockFrequency BlockFreq = MBFI->getBlockFreq(&MBB);
1758     Statistic &NumBranches =
1759         (MBB.succ_size() > 1) ? NumCondBranches : NumUncondBranches;
1760     Statistic &BranchTakenFreq =
1761         (MBB.succ_size() > 1) ? CondBranchTakenFreq : UncondBranchTakenFreq;
1762     for (MachineBasicBlock *Succ : MBB.successors()) {
1763       // Skip if this successor is a fallthrough.
1764       if (MBB.isLayoutSuccessor(Succ))
1765         continue;
1766 
1767       BlockFrequency EdgeFreq =
1768           BlockFreq * MBPI->getEdgeProbability(&MBB, Succ);
1769       ++NumBranches;
1770       BranchTakenFreq += EdgeFreq.getFrequency();
1771     }
1772   }
1773 
1774   return false;
1775 }
1776