LLVM  6.0.0svn
LoopUnrollRuntime.cpp
Go to the documentation of this file.
1 //===-- UnrollLoopRuntime.cpp - Runtime Loop unrolling utilities ----------===//
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 some loop unrolling utilities for loops with run-time
11 // trip counts. See LoopUnroll.cpp for unrolling loops with compile-time
12 // trip counts.
13 //
14 // The functions in this file are used to generate extra code when the
15 // run-time trip count modulo the unroll factor is not 0. When this is the
16 // case, we need to generate code to execute these 'left over' iterations.
17 //
18 // The current strategy generates an if-then-else sequence prior to the
19 // unrolled loop to execute the 'left over' iterations before or after the
20 // unrolled loop.
21 //
22 //===----------------------------------------------------------------------===//
23 
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/SmallSet.h"
28 #include "llvm/Analysis/LoopPass.h"
31 #include "llvm/IR/BasicBlock.h"
32 #include "llvm/IR/Dominators.h"
33 #include "llvm/IR/Metadata.h"
34 #include "llvm/IR/Module.h"
35 #include "llvm/Support/Debug.h"
37 #include "llvm/Transforms/Scalar.h"
42 #include <algorithm>
43 
44 using namespace llvm;
45 
46 #define DEBUG_TYPE "loop-unroll"
47 
48 STATISTIC(NumRuntimeUnrolled,
49  "Number of loops unrolled with run-time trip counts");
51  "unroll-runtime-multi-exit", cl::init(false), cl::Hidden,
52  cl::desc("Allow runtime unrolling for loops with multiple exits, when "
53  "epilog is generated"));
54 
55 /// Connect the unrolling prolog code to the original loop.
56 /// The unrolling prolog code contains code to execute the
57 /// 'extra' iterations if the run-time trip count modulo the
58 /// unroll count is non-zero.
59 ///
60 /// This function performs the following:
61 /// - Create PHI nodes at prolog end block to combine values
62 /// that exit the prolog code and jump around the prolog.
63 /// - Add a PHI operand to a PHI node at the loop exit block
64 /// for values that exit the prolog and go around the loop.
65 /// - Branch around the original loop if the trip count is less
66 /// than the unroll factor.
67 ///
68 static void ConnectProlog(Loop *L, Value *BECount, unsigned Count,
69  BasicBlock *PrologExit,
70  BasicBlock *OriginalLoopLatchExit,
71  BasicBlock *PreHeader, BasicBlock *NewPreHeader,
73  LoopInfo *LI, bool PreserveLCSSA) {
74  BasicBlock *Latch = L->getLoopLatch();
75  assert(Latch && "Loop must have a latch");
76  BasicBlock *PrologLatch = cast<BasicBlock>(VMap[Latch]);
77 
78  // Create a PHI node for each outgoing value from the original loop
79  // (which means it is an outgoing value from the prolog code too).
80  // The new PHI node is inserted in the prolog end basic block.
81  // The new PHI node value is added as an operand of a PHI node in either
82  // the loop header or the loop exit block.
83  for (BasicBlock *Succ : successors(Latch)) {
84  for (Instruction &BBI : *Succ) {
85  PHINode *PN = dyn_cast<PHINode>(&BBI);
86  // Exit when we passed all PHI nodes.
87  if (!PN)
88  break;
89  // Add a new PHI node to the prolog end block and add the
90  // appropriate incoming values.
91  PHINode *NewPN = PHINode::Create(PN->getType(), 2, PN->getName() + ".unr",
92  PrologExit->getFirstNonPHI());
93  // Adding a value to the new PHI node from the original loop preheader.
94  // This is the value that skips all the prolog code.
95  if (L->contains(PN)) {
96  NewPN->addIncoming(PN->getIncomingValueForBlock(NewPreHeader),
97  PreHeader);
98  } else {
99  NewPN->addIncoming(UndefValue::get(PN->getType()), PreHeader);
100  }
101 
102  Value *V = PN->getIncomingValueForBlock(Latch);
103  if (Instruction *I = dyn_cast<Instruction>(V)) {
104  if (L->contains(I)) {
105  V = VMap.lookup(I);
106  }
107  }
108  // Adding a value to the new PHI node from the last prolog block
109  // that was created.
110  NewPN->addIncoming(V, PrologLatch);
111 
112  // Update the existing PHI node operand with the value from the
113  // new PHI node. How this is done depends on if the existing
114  // PHI node is in the original loop block, or the exit block.
115  if (L->contains(PN)) {
116  PN->setIncomingValue(PN->getBasicBlockIndex(NewPreHeader), NewPN);
117  } else {
118  PN->addIncoming(NewPN, PrologExit);
119  }
120  }
121  }
122 
123  // Make sure that created prolog loop is in simplified form
124  SmallVector<BasicBlock *, 4> PrologExitPreds;
125  Loop *PrologLoop = LI->getLoopFor(PrologLatch);
126  if (PrologLoop) {
127  for (BasicBlock *PredBB : predecessors(PrologExit))
128  if (PrologLoop->contains(PredBB))
129  PrologExitPreds.push_back(PredBB);
130 
131  SplitBlockPredecessors(PrologExit, PrologExitPreds, ".unr-lcssa", DT, LI,
132  PreserveLCSSA);
133  }
134 
135  // Create a branch around the original loop, which is taken if there are no
136  // iterations remaining to be executed after running the prologue.
137  Instruction *InsertPt = PrologExit->getTerminator();
138  IRBuilder<> B(InsertPt);
139 
140  assert(Count != 0 && "nonsensical Count!");
141 
142  // If BECount <u (Count - 1) then (BECount + 1) % Count == (BECount + 1)
143  // This means %xtraiter is (BECount + 1) and all of the iterations of this
144  // loop were executed by the prologue. Note that if BECount <u (Count - 1)
145  // then (BECount + 1) cannot unsigned-overflow.
146  Value *BrLoopExit =
147  B.CreateICmpULT(BECount, ConstantInt::get(BECount->getType(), Count - 1));
148  // Split the exit to maintain loop canonicalization guarantees
149  SmallVector<BasicBlock *, 4> Preds(predecessors(OriginalLoopLatchExit));
150  SplitBlockPredecessors(OriginalLoopLatchExit, Preds, ".unr-lcssa", DT, LI,
151  PreserveLCSSA);
152  // Add the branch to the exit block (around the unrolled loop)
153  B.CreateCondBr(BrLoopExit, OriginalLoopLatchExit, NewPreHeader);
154  InsertPt->eraseFromParent();
155  if (DT)
156  DT->changeImmediateDominator(OriginalLoopLatchExit, PrologExit);
157 }
158 
159 /// Connect the unrolling epilog code to the original loop.
160 /// The unrolling epilog code contains code to execute the
161 /// 'extra' iterations if the run-time trip count modulo the
162 /// unroll count is non-zero.
163 ///
164 /// This function performs the following:
165 /// - Update PHI nodes at the unrolling loop exit and epilog loop exit
166 /// - Create PHI nodes at the unrolling loop exit to combine
167 /// values that exit the unrolling loop code and jump around it.
168 /// - Update PHI operands in the epilog loop by the new PHI nodes
169 /// - Branch around the epilog loop if extra iters (ModVal) is zero.
170 ///
171 static void ConnectEpilog(Loop *L, Value *ModVal, BasicBlock *NewExit,
172  BasicBlock *Exit, BasicBlock *PreHeader,
173  BasicBlock *EpilogPreHeader, BasicBlock *NewPreHeader,
174  ValueToValueMapTy &VMap, DominatorTree *DT,
175  LoopInfo *LI, bool PreserveLCSSA) {
176  BasicBlock *Latch = L->getLoopLatch();
177  assert(Latch && "Loop must have a latch");
178  BasicBlock *EpilogLatch = cast<BasicBlock>(VMap[Latch]);
179 
180  // Loop structure should be the following:
181  //
182  // PreHeader
183  // NewPreHeader
184  // Header
185  // ...
186  // Latch
187  // NewExit (PN)
188  // EpilogPreHeader
189  // EpilogHeader
190  // ...
191  // EpilogLatch
192  // Exit (EpilogPN)
193 
194  // Update PHI nodes at NewExit and Exit.
195  for (Instruction &BBI : *NewExit) {
196  PHINode *PN = dyn_cast<PHINode>(&BBI);
197  // Exit when we passed all PHI nodes.
198  if (!PN)
199  break;
200  // PN should be used in another PHI located in Exit block as
201  // Exit was split by SplitBlockPredecessors into Exit and NewExit
202  // Basicaly it should look like:
203  // NewExit:
204  // PN = PHI [I, Latch]
205  // ...
206  // Exit:
207  // EpilogPN = PHI [PN, EpilogPreHeader]
208  //
209  // There is EpilogPreHeader incoming block instead of NewExit as
210  // NewExit was spilt 1 more time to get EpilogPreHeader.
211  assert(PN->hasOneUse() && "The phi should have 1 use");
212  PHINode *EpilogPN = cast<PHINode> (PN->use_begin()->getUser());
213  assert(EpilogPN->getParent() == Exit && "EpilogPN should be in Exit block");
214 
215  // Add incoming PreHeader from branch around the Loop
216  PN->addIncoming(UndefValue::get(PN->getType()), PreHeader);
217 
218  Value *V = PN->getIncomingValueForBlock(Latch);
220  if (I && L->contains(I))
221  // If value comes from an instruction in the loop add VMap value.
222  V = VMap.lookup(I);
223  // For the instruction out of the loop, constant or undefined value
224  // insert value itself.
225  EpilogPN->addIncoming(V, EpilogLatch);
226 
227  assert(EpilogPN->getBasicBlockIndex(EpilogPreHeader) >= 0 &&
228  "EpilogPN should have EpilogPreHeader incoming block");
229  // Change EpilogPreHeader incoming block to NewExit.
230  EpilogPN->setIncomingBlock(EpilogPN->getBasicBlockIndex(EpilogPreHeader),
231  NewExit);
232  // Now PHIs should look like:
233  // NewExit:
234  // PN = PHI [I, Latch], [undef, PreHeader]
235  // ...
236  // Exit:
237  // EpilogPN = PHI [PN, NewExit], [VMap[I], EpilogLatch]
238  }
239 
240  // Create PHI nodes at NewExit (from the unrolling loop Latch and PreHeader).
241  // Update corresponding PHI nodes in epilog loop.
242  for (BasicBlock *Succ : successors(Latch)) {
243  // Skip this as we already updated phis in exit blocks.
244  if (!L->contains(Succ))
245  continue;
246  for (Instruction &BBI : *Succ) {
247  PHINode *PN = dyn_cast<PHINode>(&BBI);
248  // Exit when we passed all PHI nodes.
249  if (!PN)
250  break;
251  // Add new PHI nodes to the loop exit block and update epilog
252  // PHIs with the new PHI values.
253  PHINode *NewPN = PHINode::Create(PN->getType(), 2, PN->getName() + ".unr",
254  NewExit->getFirstNonPHI());
255  // Adding a value to the new PHI node from the unrolling loop preheader.
256  NewPN->addIncoming(PN->getIncomingValueForBlock(NewPreHeader), PreHeader);
257  // Adding a value to the new PHI node from the unrolling loop latch.
258  NewPN->addIncoming(PN->getIncomingValueForBlock(Latch), Latch);
259 
260  // Update the existing PHI node operand with the value from the new PHI
261  // node. Corresponding instruction in epilog loop should be PHI.
262  PHINode *VPN = cast<PHINode>(VMap[&BBI]);
263  VPN->setIncomingValue(VPN->getBasicBlockIndex(EpilogPreHeader), NewPN);
264  }
265  }
266 
267  Instruction *InsertPt = NewExit->getTerminator();
268  IRBuilder<> B(InsertPt);
269  Value *BrLoopExit = B.CreateIsNotNull(ModVal, "lcmp.mod");
270  assert(Exit && "Loop must have a single exit block only");
271  // Split the epilogue exit to maintain loop canonicalization guarantees
273  SplitBlockPredecessors(Exit, Preds, ".epilog-lcssa", DT, LI,
274  PreserveLCSSA);
275  // Add the branch to the exit block (around the unrolling loop)
276  B.CreateCondBr(BrLoopExit, EpilogPreHeader, Exit);
277  InsertPt->eraseFromParent();
278  if (DT)
279  DT->changeImmediateDominator(Exit, NewExit);
280 
281  // Split the main loop exit to maintain canonicalization guarantees.
282  SmallVector<BasicBlock*, 4> NewExitPreds{Latch};
283  SplitBlockPredecessors(NewExit, NewExitPreds, ".loopexit", DT, LI,
284  PreserveLCSSA);
285 }
286 
287 /// Create a clone of the blocks in a loop and connect them together.
288 /// If CreateRemainderLoop is false, loop structure will not be cloned,
289 /// otherwise a new loop will be created including all cloned blocks, and the
290 /// iterator of it switches to count NewIter down to 0.
291 /// The cloned blocks should be inserted between InsertTop and InsertBot.
292 /// If loop structure is cloned InsertTop should be new preheader, InsertBot
293 /// new loop exit.
294 /// Return the new cloned loop that is created when CreateRemainderLoop is true.
295 static Loop *
296 CloneLoopBlocks(Loop *L, Value *NewIter, const bool CreateRemainderLoop,
297  const bool UseEpilogRemainder, const bool UnrollRemainder,
298  BasicBlock *InsertTop,
299  BasicBlock *InsertBot, BasicBlock *Preheader,
300  std::vector<BasicBlock *> &NewBlocks, LoopBlocksDFS &LoopBlocks,
301  ValueToValueMapTy &VMap, DominatorTree *DT, LoopInfo *LI) {
302  StringRef suffix = UseEpilogRemainder ? "epil" : "prol";
303  BasicBlock *Header = L->getHeader();
304  BasicBlock *Latch = L->getLoopLatch();
305  Function *F = Header->getParent();
306  LoopBlocksDFS::RPOIterator BlockBegin = LoopBlocks.beginRPO();
307  LoopBlocksDFS::RPOIterator BlockEnd = LoopBlocks.endRPO();
308  Loop *ParentLoop = L->getParentLoop();
309  NewLoopsMap NewLoops;
310  NewLoops[ParentLoop] = ParentLoop;
311  if (!CreateRemainderLoop)
312  NewLoops[L] = ParentLoop;
313 
314  // For each block in the original loop, create a new copy,
315  // and update the value map with the newly created values.
316  for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) {
317  BasicBlock *NewBB = CloneBasicBlock(*BB, VMap, "." + suffix, F);
318  NewBlocks.push_back(NewBB);
319 
320  // If we're unrolling the outermost loop, there's no remainder loop,
321  // and this block isn't in a nested loop, then the new block is not
322  // in any loop. Otherwise, add it to loopinfo.
323  if (CreateRemainderLoop || LI->getLoopFor(*BB) != L || ParentLoop)
324  addClonedBlockToLoopInfo(*BB, NewBB, LI, NewLoops);
325 
326  VMap[*BB] = NewBB;
327  if (Header == *BB) {
328  // For the first block, add a CFG connection to this newly
329  // created block.
330  InsertTop->getTerminator()->setSuccessor(0, NewBB);
331  }
332 
333  if (DT) {
334  if (Header == *BB) {
335  // The header is dominated by the preheader.
336  DT->addNewBlock(NewBB, InsertTop);
337  } else {
338  // Copy information from original loop to unrolled loop.
339  BasicBlock *IDomBB = DT->getNode(*BB)->getIDom()->getBlock();
340  DT->addNewBlock(NewBB, cast<BasicBlock>(VMap[IDomBB]));
341  }
342  }
343 
344  if (Latch == *BB) {
345  // For the last block, if CreateRemainderLoop is false, create a direct
346  // jump to InsertBot. If not, create a loop back to cloned head.
347  VMap.erase((*BB)->getTerminator());
348  BasicBlock *FirstLoopBB = cast<BasicBlock>(VMap[Header]);
349  BranchInst *LatchBR = cast<BranchInst>(NewBB->getTerminator());
350  IRBuilder<> Builder(LatchBR);
351  if (!CreateRemainderLoop) {
352  Builder.CreateBr(InsertBot);
353  } else {
354  PHINode *NewIdx = PHINode::Create(NewIter->getType(), 2,
355  suffix + ".iter",
356  FirstLoopBB->getFirstNonPHI());
357  Value *IdxSub =
358  Builder.CreateSub(NewIdx, ConstantInt::get(NewIdx->getType(), 1),
359  NewIdx->getName() + ".sub");
360  Value *IdxCmp =
361  Builder.CreateIsNotNull(IdxSub, NewIdx->getName() + ".cmp");
362  Builder.CreateCondBr(IdxCmp, FirstLoopBB, InsertBot);
363  NewIdx->addIncoming(NewIter, InsertTop);
364  NewIdx->addIncoming(IdxSub, NewBB);
365  }
366  LatchBR->eraseFromParent();
367  }
368  }
369 
370  // Change the incoming values to the ones defined in the preheader or
371  // cloned loop.
372  for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
373  PHINode *NewPHI = cast<PHINode>(VMap[&*I]);
374  if (!CreateRemainderLoop) {
375  if (UseEpilogRemainder) {
376  unsigned idx = NewPHI->getBasicBlockIndex(Preheader);
377  NewPHI->setIncomingBlock(idx, InsertTop);
378  NewPHI->removeIncomingValue(Latch, false);
379  } else {
380  VMap[&*I] = NewPHI->getIncomingValueForBlock(Preheader);
381  cast<BasicBlock>(VMap[Header])->getInstList().erase(NewPHI);
382  }
383  } else {
384  unsigned idx = NewPHI->getBasicBlockIndex(Preheader);
385  NewPHI->setIncomingBlock(idx, InsertTop);
386  BasicBlock *NewLatch = cast<BasicBlock>(VMap[Latch]);
387  idx = NewPHI->getBasicBlockIndex(Latch);
388  Value *InVal = NewPHI->getIncomingValue(idx);
389  NewPHI->setIncomingBlock(idx, NewLatch);
390  if (Value *V = VMap.lookup(InVal))
391  NewPHI->setIncomingValue(idx, V);
392  }
393  }
394  if (CreateRemainderLoop) {
395  Loop *NewLoop = NewLoops[L];
396  assert(NewLoop && "L should have been cloned");
397 
398  // Only add loop metadata if the loop is not going to be completely
399  // unrolled.
400  if (UnrollRemainder)
401  return NewLoop;
402 
403  // Add unroll disable metadata to disable future unrolling for this loop.
405  // Reserve first location for self reference to the LoopID metadata node.
406  MDs.push_back(nullptr);
407  MDNode *LoopID = NewLoop->getLoopID();
408  if (LoopID) {
409  // First remove any existing loop unrolling metadata.
410  for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
411  bool IsUnrollMetadata = false;
412  MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
413  if (MD) {
414  const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
415  IsUnrollMetadata = S && S->getString().startswith("llvm.loop.unroll.");
416  }
417  if (!IsUnrollMetadata)
418  MDs.push_back(LoopID->getOperand(i));
419  }
420  }
421 
422  LLVMContext &Context = NewLoop->getHeader()->getContext();
423  SmallVector<Metadata *, 1> DisableOperands;
424  DisableOperands.push_back(MDString::get(Context,
425  "llvm.loop.unroll.disable"));
426  MDNode *DisableNode = MDNode::get(Context, DisableOperands);
427  MDs.push_back(DisableNode);
428 
429  MDNode *NewLoopID = MDNode::get(Context, MDs);
430  // Set operand 0 to refer to the loop id itself.
431  NewLoopID->replaceOperandWith(0, NewLoopID);
432  NewLoop->setLoopID(NewLoopID);
433  return NewLoop;
434  }
435  else
436  return nullptr;
437 }
438 
439 /// Returns true if we can safely unroll a multi-exit/exiting loop. OtherExits
440 /// is populated with all the loop exit blocks other than the LatchExit block.
441 static bool
443  BasicBlock *LatchExit, bool PreserveLCSSA,
444  bool UseEpilogRemainder) {
445 
446  // We currently have some correctness constrains in unrolling a multi-exit
447  // loop. Check for these below.
448 
449  // We rely on LCSSA form being preserved when the exit blocks are transformed.
450  if (!PreserveLCSSA)
451  return false;
453  L->getUniqueExitBlocks(Exits);
454  for (auto *BB : Exits)
455  if (BB != LatchExit)
456  OtherExits.push_back(BB);
457 
458  // TODO: Support multiple exiting blocks jumping to the `LatchExit` when
459  // UnrollRuntimeMultiExit is true. This will need updating the logic in
460  // connectEpilog/connectProlog.
461  if (!LatchExit->getSinglePredecessor()) {
462  DEBUG(dbgs() << "Bailout for multi-exit handling when latch exit has >1 "
463  "predecessor.\n");
464  return false;
465  }
466  // FIXME: We bail out of multi-exit unrolling when epilog loop is generated
467  // and L is an inner loop. This is because in presence of multiple exits, the
468  // outer loop is incorrect: we do not add the EpilogPreheader and exit to the
469  // outer loop. This is automatically handled in the prolog case, so we do not
470  // have that bug in prolog generation.
471  if (UseEpilogRemainder && L->getParentLoop())
472  return false;
473 
474  // All constraints have been satisfied.
475  return true;
476 }
477 
478 /// Returns true if we can profitably unroll the multi-exit loop L. Currently,
479 /// we return true only if UnrollRuntimeMultiExit is set to true.
481  Loop *L, SmallVectorImpl<BasicBlock *> &OtherExits, BasicBlock *LatchExit,
482  bool PreserveLCSSA, bool UseEpilogRemainder) {
483 
484 #if !defined(NDEBUG)
485  SmallVector<BasicBlock *, 8> OtherExitsDummyCheck;
486  assert(canSafelyUnrollMultiExitLoop(L, OtherExitsDummyCheck, LatchExit,
487  PreserveLCSSA, UseEpilogRemainder) &&
488  "Should be safe to unroll before checking profitability!");
489 #endif
490 
491  // Priority goes to UnrollRuntimeMultiExit if it's supplied.
492  if (UnrollRuntimeMultiExit.getNumOccurrences())
493  return UnrollRuntimeMultiExit;
494 
495  // The main pain point with multi-exit loop unrolling is that once unrolled,
496  // we will not be able to merge all blocks into a straight line code.
497  // There are branches within the unrolled loop that go to the OtherExits.
498  // The second point is the increase in code size, but this is true
499  // irrespective of multiple exits.
500 
501  // Note: Both the heuristics below are coarse grained. We are essentially
502  // enabling unrolling of loops that have a single side exit other than the
503  // normal LatchExit (i.e. exiting into a deoptimize block).
504  // The heuristics considered are:
505  // 1. low number of branches in the unrolled version.
506  // 2. high predictability of these extra branches.
507  // We avoid unrolling loops that have more than two exiting blocks. This
508  // limits the total number of branches in the unrolled loop to be atmost
509  // the unroll factor (since one of the exiting blocks is the latch block).
510  SmallVector<BasicBlock*, 4> ExitingBlocks;
511  L->getExitingBlocks(ExitingBlocks);
512  if (ExitingBlocks.size() > 2)
513  return false;
514 
515  // The second heuristic is that L has one exit other than the latchexit and
516  // that exit is a deoptimize block. We know that deoptimize blocks are rarely
517  // taken, which also implies the branch leading to the deoptimize block is
518  // highly predictable.
519  return (OtherExits.size() == 1 &&
520  OtherExits[0]->getTerminatingDeoptimizeCall());
521  // TODO: These can be fine-tuned further to consider code size or deopt states
522  // that are captured by the deoptimize exit block.
523  // Also, we can extend this to support more cases, if we actually
524  // know of kinds of multiexit loops that would benefit from unrolling.
525 }
526 
527 /// Insert code in the prolog/epilog code when unrolling a loop with a
528 /// run-time trip-count.
529 ///
530 /// This method assumes that the loop unroll factor is total number
531 /// of loop bodies in the loop after unrolling. (Some folks refer
532 /// to the unroll factor as the number of *extra* copies added).
533 /// We assume also that the loop unroll factor is a power-of-two. So, after
534 /// unrolling the loop, the number of loop bodies executed is 2,
535 /// 4, 8, etc. Note - LLVM converts the if-then-sequence to a switch
536 /// instruction in SimplifyCFG.cpp. Then, the backend decides how code for
537 /// the switch instruction is generated.
538 ///
539 /// ***Prolog case***
540 /// extraiters = tripcount % loopfactor
541 /// if (extraiters == 0) jump Loop:
542 /// else jump Prol:
543 /// Prol: LoopBody;
544 /// extraiters -= 1 // Omitted if unroll factor is 2.
545 /// if (extraiters != 0) jump Prol: // Omitted if unroll factor is 2.
546 /// if (tripcount < loopfactor) jump End:
547 /// Loop:
548 /// ...
549 /// End:
550 ///
551 /// ***Epilog case***
552 /// extraiters = tripcount % loopfactor
553 /// if (tripcount < loopfactor) jump LoopExit:
554 /// unroll_iters = tripcount - extraiters
555 /// Loop: LoopBody; (executes unroll_iter times);
556 /// unroll_iter -= 1
557 /// if (unroll_iter != 0) jump Loop:
558 /// LoopExit:
559 /// if (extraiters == 0) jump EpilExit:
560 /// Epil: LoopBody; (executes extraiters times)
561 /// extraiters -= 1 // Omitted if unroll factor is 2.
562 /// if (extraiters != 0) jump Epil: // Omitted if unroll factor is 2.
563 /// EpilExit:
564 
565 bool llvm::UnrollRuntimeLoopRemainder(Loop *L, unsigned Count,
566  bool AllowExpensiveTripCount,
567  bool UseEpilogRemainder,
568  bool UnrollRemainder,
569  LoopInfo *LI, ScalarEvolution *SE,
572  bool PreserveLCSSA) {
573  DEBUG(dbgs() << "Trying runtime unrolling on Loop: \n");
574  DEBUG(L->dump());
575  DEBUG(UseEpilogRemainder ? dbgs() << "Using epilog remainder.\n" :
576  dbgs() << "Using prolog remainder.\n");
577 
578  // Make sure the loop is in canonical form.
579  if (!L->isLoopSimplifyForm()) {
580  DEBUG(dbgs() << "Not in simplify form!\n");
581  return false;
582  }
583 
584  // Guaranteed by LoopSimplifyForm.
585  BasicBlock *Latch = L->getLoopLatch();
586  BasicBlock *Header = L->getHeader();
587 
588  BranchInst *LatchBR = cast<BranchInst>(Latch->getTerminator());
589  unsigned ExitIndex = LatchBR->getSuccessor(0) == Header ? 1 : 0;
590  BasicBlock *LatchExit = LatchBR->getSuccessor(ExitIndex);
591  // Cloning the loop basic blocks (`CloneLoopBlocks`) requires that one of the
592  // targets of the Latch be an exit block out of the loop. This needs
593  // to be guaranteed by the callers of UnrollRuntimeLoopRemainder.
594  assert(!L->contains(LatchExit) &&
595  "one of the loop latch successors should be the exit block!");
596  // These are exit blocks other than the target of the latch exiting block.
597  SmallVector<BasicBlock *, 4> OtherExits;
598  bool isMultiExitUnrollingEnabled =
599  canSafelyUnrollMultiExitLoop(L, OtherExits, LatchExit, PreserveLCSSA,
600  UseEpilogRemainder) &&
601  canProfitablyUnrollMultiExitLoop(L, OtherExits, LatchExit, PreserveLCSSA,
602  UseEpilogRemainder);
603  // Support only single exit and exiting block unless multi-exit loop unrolling is enabled.
604  if (!isMultiExitUnrollingEnabled &&
605  (!L->getExitingBlock() || OtherExits.size())) {
606  DEBUG(
607  dbgs()
608  << "Multiple exit/exiting blocks in loop and multi-exit unrolling not "
609  "enabled!\n");
610  return false;
611  }
612  // Use Scalar Evolution to compute the trip count. This allows more loops to
613  // be unrolled than relying on induction var simplification.
614  if (!SE)
615  return false;
616 
617  // Only unroll loops with a computable trip count, and the trip count needs
618  // to be an int value (allowing a pointer type is a TODO item).
619  // We calculate the backedge count by using getExitCount on the Latch block,
620  // which is proven to be the only exiting block in this loop. This is same as
621  // calculating getBackedgeTakenCount on the loop (which computes SCEV for all
622  // exiting blocks).
623  const SCEV *BECountSC = SE->getExitCount(L, Latch);
624  if (isa<SCEVCouldNotCompute>(BECountSC) ||
625  !BECountSC->getType()->isIntegerTy()) {
626  DEBUG(dbgs() << "Could not compute exit block SCEV\n");
627  return false;
628  }
629 
630  unsigned BEWidth = cast<IntegerType>(BECountSC->getType())->getBitWidth();
631 
632  // Add 1 since the backedge count doesn't include the first loop iteration.
633  const SCEV *TripCountSC =
634  SE->getAddExpr(BECountSC, SE->getConstant(BECountSC->getType(), 1));
635  if (isa<SCEVCouldNotCompute>(TripCountSC)) {
636  DEBUG(dbgs() << "Could not compute trip count SCEV.\n");
637  return false;
638  }
639 
640  BasicBlock *PreHeader = L->getLoopPreheader();
641  BranchInst *PreHeaderBR = cast<BranchInst>(PreHeader->getTerminator());
642  const DataLayout &DL = Header->getModule()->getDataLayout();
643  SCEVExpander Expander(*SE, DL, "loop-unroll");
644  if (!AllowExpensiveTripCount &&
645  Expander.isHighCostExpansion(TripCountSC, L, PreHeaderBR)) {
646  DEBUG(dbgs() << "High cost for expanding trip count scev!\n");
647  return false;
648  }
649 
650  // This constraint lets us deal with an overflowing trip count easily; see the
651  // comment on ModVal below.
652  if (Log2_32(Count) > BEWidth) {
653  DEBUG(dbgs()
654  << "Count failed constraint on overflow trip count calculation.\n");
655  return false;
656  }
657 
658  // Loop structure is the following:
659  //
660  // PreHeader
661  // Header
662  // ...
663  // Latch
664  // LatchExit
665 
666  BasicBlock *NewPreHeader;
667  BasicBlock *NewExit = nullptr;
668  BasicBlock *PrologExit = nullptr;
669  BasicBlock *EpilogPreHeader = nullptr;
670  BasicBlock *PrologPreHeader = nullptr;
671 
672  if (UseEpilogRemainder) {
673  // If epilog remainder
674  // Split PreHeader to insert a branch around loop for unrolling.
675  NewPreHeader = SplitBlock(PreHeader, PreHeader->getTerminator(), DT, LI);
676  NewPreHeader->setName(PreHeader->getName() + ".new");
677  // Split LatchExit to create phi nodes from branch above.
678  SmallVector<BasicBlock*, 4> Preds(predecessors(LatchExit));
679  NewExit = SplitBlockPredecessors(LatchExit, Preds, ".unr-lcssa",
680  DT, LI, PreserveLCSSA);
681  // Split NewExit to insert epilog remainder loop.
682  EpilogPreHeader = SplitBlock(NewExit, NewExit->getTerminator(), DT, LI);
683  EpilogPreHeader->setName(Header->getName() + ".epil.preheader");
684  } else {
685  // If prolog remainder
686  // Split the original preheader twice to insert prolog remainder loop
687  PrologPreHeader = SplitEdge(PreHeader, Header, DT, LI);
688  PrologPreHeader->setName(Header->getName() + ".prol.preheader");
689  PrologExit = SplitBlock(PrologPreHeader, PrologPreHeader->getTerminator(),
690  DT, LI);
691  PrologExit->setName(Header->getName() + ".prol.loopexit");
692  // Split PrologExit to get NewPreHeader.
693  NewPreHeader = SplitBlock(PrologExit, PrologExit->getTerminator(), DT, LI);
694  NewPreHeader->setName(PreHeader->getName() + ".new");
695  }
696  // Loop structure should be the following:
697  // Epilog Prolog
698  //
699  // PreHeader PreHeader
700  // *NewPreHeader *PrologPreHeader
701  // Header *PrologExit
702  // ... *NewPreHeader
703  // Latch Header
704  // *NewExit ...
705  // *EpilogPreHeader Latch
706  // LatchExit LatchExit
707 
708  // Calculate conditions for branch around loop for unrolling
709  // in epilog case and around prolog remainder loop in prolog case.
710  // Compute the number of extra iterations required, which is:
711  // extra iterations = run-time trip count % loop unroll factor
712  PreHeaderBR = cast<BranchInst>(PreHeader->getTerminator());
713  Value *TripCount = Expander.expandCodeFor(TripCountSC, TripCountSC->getType(),
714  PreHeaderBR);
715  Value *BECount = Expander.expandCodeFor(BECountSC, BECountSC->getType(),
716  PreHeaderBR);
717  IRBuilder<> B(PreHeaderBR);
718  Value *ModVal;
719  // Calculate ModVal = (BECount + 1) % Count.
720  // Note that TripCount is BECount + 1.
721  if (isPowerOf2_32(Count)) {
722  // When Count is power of 2 we don't BECount for epilog case, however we'll
723  // need it for a branch around unrolling loop for prolog case.
724  ModVal = B.CreateAnd(TripCount, Count - 1, "xtraiter");
725  // 1. There are no iterations to be run in the prolog/epilog loop.
726  // OR
727  // 2. The addition computing TripCount overflowed.
728  //
729  // If (2) is true, we know that TripCount really is (1 << BEWidth) and so
730  // the number of iterations that remain to be run in the original loop is a
731  // multiple Count == (1 << Log2(Count)) because Log2(Count) <= BEWidth (we
732  // explicitly check this above).
733  } else {
734  // As (BECount + 1) can potentially unsigned overflow we count
735  // (BECount % Count) + 1 which is overflow safe as BECount % Count < Count.
736  Value *ModValTmp = B.CreateURem(BECount,
737  ConstantInt::get(BECount->getType(),
738  Count));
739  Value *ModValAdd = B.CreateAdd(ModValTmp,
740  ConstantInt::get(ModValTmp->getType(), 1));
741  // At that point (BECount % Count) + 1 could be equal to Count.
742  // To handle this case we need to take mod by Count one more time.
743  ModVal = B.CreateURem(ModValAdd,
744  ConstantInt::get(BECount->getType(), Count),
745  "xtraiter");
746  }
747  Value *BranchVal =
748  UseEpilogRemainder ? B.CreateICmpULT(BECount,
749  ConstantInt::get(BECount->getType(),
750  Count - 1)) :
751  B.CreateIsNotNull(ModVal, "lcmp.mod");
752  BasicBlock *RemainderLoop = UseEpilogRemainder ? NewExit : PrologPreHeader;
753  BasicBlock *UnrollingLoop = UseEpilogRemainder ? NewPreHeader : PrologExit;
754  // Branch to either remainder (extra iterations) loop or unrolling loop.
755  B.CreateCondBr(BranchVal, RemainderLoop, UnrollingLoop);
756  PreHeaderBR->eraseFromParent();
757  if (DT) {
758  if (UseEpilogRemainder)
759  DT->changeImmediateDominator(NewExit, PreHeader);
760  else
761  DT->changeImmediateDominator(PrologExit, PreHeader);
762  }
763  Function *F = Header->getParent();
764  // Get an ordered list of blocks in the loop to help with the ordering of the
765  // cloned blocks in the prolog/epilog code
766  LoopBlocksDFS LoopBlocks(L);
767  LoopBlocks.perform(LI);
768 
769  //
770  // For each extra loop iteration, create a copy of the loop's basic blocks
771  // and generate a condition that branches to the copy depending on the
772  // number of 'left over' iterations.
773  //
774  std::vector<BasicBlock *> NewBlocks;
775  ValueToValueMapTy VMap;
776 
777  // For unroll factor 2 remainder loop will have 1 iterations.
778  // Do not create 1 iteration loop.
779  bool CreateRemainderLoop = (Count != 2);
780 
781  // Clone all the basic blocks in the loop. If Count is 2, we don't clone
782  // the loop, otherwise we create a cloned loop to execute the extra
783  // iterations. This function adds the appropriate CFG connections.
784  BasicBlock *InsertBot = UseEpilogRemainder ? LatchExit : PrologExit;
785  BasicBlock *InsertTop = UseEpilogRemainder ? EpilogPreHeader : PrologPreHeader;
786  Loop *remainderLoop = CloneLoopBlocks(
787  L, ModVal, CreateRemainderLoop, UseEpilogRemainder, UnrollRemainder,
788  InsertTop, InsertBot,
789  NewPreHeader, NewBlocks, LoopBlocks, VMap, DT, LI);
790 
791  // Insert the cloned blocks into the function.
792  F->getBasicBlockList().splice(InsertBot->getIterator(),
793  F->getBasicBlockList(),
794  NewBlocks[0]->getIterator(),
795  F->end());
796 
797  // Now the loop blocks are cloned and the other exiting blocks from the
798  // remainder are connected to the original Loop's exit blocks. The remaining
799  // work is to update the phi nodes in the original loop, and take in the
800  // values from the cloned region. Also update the dominator info for
801  // OtherExits and their immediate successors, since we have new edges into
802  // OtherExits.
803  SmallSet<BasicBlock*, 8> ImmediateSuccessorsOfExitBlocks;
804  for (auto *BB : OtherExits) {
805  for (auto &II : *BB) {
806 
807  // Given we preserve LCSSA form, we know that the values used outside the
808  // loop will be used through these phi nodes at the exit blocks that are
809  // transformed below.
810  if (!isa<PHINode>(II))
811  break;
812  PHINode *Phi = cast<PHINode>(&II);
813  unsigned oldNumOperands = Phi->getNumIncomingValues();
814  // Add the incoming values from the remainder code to the end of the phi
815  // node.
816  for (unsigned i =0; i < oldNumOperands; i++){
817  Value *newVal = VMap.lookup(Phi->getIncomingValue(i));
818  // newVal can be a constant or derived from values outside the loop, and
819  // hence need not have a VMap value. Also, since lookup already generated
820  // a default "null" VMap entry for this value, we need to populate that
821  // VMap entry correctly, with the mapped entry being itself.
822  if (!newVal) {
823  newVal = Phi->getIncomingValue(i);
824  VMap[Phi->getIncomingValue(i)] = Phi->getIncomingValue(i);
825  }
826  Phi->addIncoming(newVal,
827  cast<BasicBlock>(VMap[Phi->getIncomingBlock(i)]));
828  }
829  }
830 #if defined(EXPENSIVE_CHECKS) && !defined(NDEBUG)
831  for (BasicBlock *SuccBB : successors(BB)) {
832  assert(!(any_of(OtherExits,
833  [SuccBB](BasicBlock *EB) { return EB == SuccBB; }) ||
834  SuccBB == LatchExit) &&
835  "Breaks the definition of dedicated exits!");
836  }
837 #endif
838  // Update the dominator info because the immediate dominator is no longer the
839  // header of the original Loop. BB has edges both from L and remainder code.
840  // Since the preheader determines which loop is run (L or directly jump to
841  // the remainder code), we set the immediate dominator as the preheader.
842  if (DT) {
843  DT->changeImmediateDominator(BB, PreHeader);
844  // Also update the IDom for immediate successors of BB. If the current
845  // IDom is the header, update the IDom to be the preheader because that is
846  // the nearest common dominator of all predecessors of SuccBB. We need to
847  // check for IDom being the header because successors of exit blocks can
848  // have edges from outside the loop, and we should not incorrectly update
849  // the IDom in that case.
850  for (BasicBlock *SuccBB: successors(BB))
851  if (ImmediateSuccessorsOfExitBlocks.insert(SuccBB).second) {
852  if (DT->getNode(SuccBB)->getIDom()->getBlock() == Header) {
853  assert(!SuccBB->getSinglePredecessor() &&
854  "BB should be the IDom then!");
855  DT->changeImmediateDominator(SuccBB, PreHeader);
856  }
857  }
858  }
859  }
860 
861  // Loop structure should be the following:
862  // Epilog Prolog
863  //
864  // PreHeader PreHeader
865  // NewPreHeader PrologPreHeader
866  // Header PrologHeader
867  // ... ...
868  // Latch PrologLatch
869  // NewExit PrologExit
870  // EpilogPreHeader NewPreHeader
871  // EpilogHeader Header
872  // ... ...
873  // EpilogLatch Latch
874  // LatchExit LatchExit
875 
876  // Rewrite the cloned instruction operands to use the values created when the
877  // clone is created.
878  for (BasicBlock *BB : NewBlocks) {
879  for (Instruction &I : *BB) {
880  RemapInstruction(&I, VMap,
882  }
883  }
884 
885  if (UseEpilogRemainder) {
886  // Connect the epilog code to the original loop and update the
887  // PHI functions.
888  ConnectEpilog(L, ModVal, NewExit, LatchExit, PreHeader,
889  EpilogPreHeader, NewPreHeader, VMap, DT, LI,
890  PreserveLCSSA);
891 
892  // Update counter in loop for unrolling.
893  // I should be multiply of Count.
894  IRBuilder<> B2(NewPreHeader->getTerminator());
895  Value *TestVal = B2.CreateSub(TripCount, ModVal, "unroll_iter");
896  BranchInst *LatchBR = cast<BranchInst>(Latch->getTerminator());
897  B2.SetInsertPoint(LatchBR);
898  PHINode *NewIdx = PHINode::Create(TestVal->getType(), 2, "niter",
899  Header->getFirstNonPHI());
900  Value *IdxSub =
901  B2.CreateSub(NewIdx, ConstantInt::get(NewIdx->getType(), 1),
902  NewIdx->getName() + ".nsub");
903  Value *IdxCmp;
904  if (LatchBR->getSuccessor(0) == Header)
905  IdxCmp = B2.CreateIsNotNull(IdxSub, NewIdx->getName() + ".ncmp");
906  else
907  IdxCmp = B2.CreateIsNull(IdxSub, NewIdx->getName() + ".ncmp");
908  NewIdx->addIncoming(TestVal, NewPreHeader);
909  NewIdx->addIncoming(IdxSub, Latch);
910  LatchBR->setCondition(IdxCmp);
911  } else {
912  // Connect the prolog code to the original loop and update the
913  // PHI functions.
914  ConnectProlog(L, BECount, Count, PrologExit, LatchExit, PreHeader,
915  NewPreHeader, VMap, DT, LI, PreserveLCSSA);
916  }
917 
918  // If this loop is nested, then the loop unroller changes the code in the
919  // parent loop, so the Scalar Evolution pass needs to be run again.
920  if (Loop *ParentLoop = L->getParentLoop())
921  SE->forgetLoop(ParentLoop);
922 
923  // Canonicalize to LoopSimplifyForm both original and remainder loops. We
924  // cannot rely on the LoopUnrollPass to do this because it only does
925  // canonicalization for parent/subloops and not the sibling loops.
926  if (OtherExits.size() > 0) {
927  // Generate dedicated exit blocks for the original loop, to preserve
928  // LoopSimplifyForm.
929  formDedicatedExitBlocks(L, DT, LI, PreserveLCSSA);
930  // Generate dedicated exit blocks for the remainder loop if one exists, to
931  // preserve LoopSimplifyForm.
932  if (remainderLoop)
933  formDedicatedExitBlocks(remainderLoop, DT, LI, PreserveLCSSA);
934  }
935 
936  if (remainderLoop && UnrollRemainder) {
937  DEBUG(dbgs() << "Unrolling remainder loop\n");
938  UnrollLoop(remainderLoop, /*Count*/Count - 1, /*TripCount*/Count - 1,
939  /*Force*/false, /*AllowRuntime*/false,
940  /*AllowExpensiveTripCount*/false, /*PreserveCondBr*/true,
941  /*PreserveOnlyFirst*/false, /*TripMultiple*/1,
942  /*PeelCount*/0, /*UnrollRemainder*/false, LI, SE, DT, AC, ORE,
943  PreserveLCSSA);
944  }
945 
946  NumRuntimeUnrolled++;
947  return true;
948 }
static unsigned getBitWidth(Type *Ty, const DataLayout &DL)
Returns the bitwidth of the given scalar or pointer type.
DomTreeNodeBase< NodeT > * getNode(NodeT *BB) const
getNode - return the (Post)DominatorTree node for the specified basic block.
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks &#39;this&#39; from the containing basic block and deletes it.
Definition: Instruction.cpp:69
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:109
BranchInst * CreateCondBr(Value *Cond, BasicBlock *True, BasicBlock *False, MDNode *BranchWeights=nullptr, MDNode *Unpredictable=nullptr)
Create a conditional &#39;br Cond, TrueDest, FalseDest&#39; instruction.
Definition: IRBuilder.h:775
static bool canSafelyUnrollMultiExitLoop(Loop *L, SmallVectorImpl< BasicBlock *> &OtherExits, BasicBlock *LatchExit, bool PreserveLCSSA, bool UseEpilogRemainder)
Returns true if we can safely unroll a multi-exit/exiting loop.
void addIncoming(Value *V, BasicBlock *BB)
Add an incoming value to the end of the PHI list.
BlockT * getLoopLatch() const
If there is a single latch block for this loop, return it.
Definition: LoopInfoImpl.h:157
LLVMContext & Context
Value * CreateIsNotNull(Value *Arg, const Twine &Name="")
Return an i1 value testing if Arg is not null.
Definition: IRBuilder.h:1784
const SCEV * getConstant(ConstantInt *V)
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
BasicBlock * SplitBlock(BasicBlock *Old, Instruction *SplitPt, DominatorTree *DT=nullptr, LoopInfo *LI=nullptr)
Split the specified block at the specified instruction - everything before SplitPt stays in Old and e...
Value * CreateICmpULT(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1558
LLVM_ATTRIBUTE_ALWAYS_INLINE size_type size() const
Definition: SmallVector.h:136
void replaceOperandWith(unsigned I, Metadata *New)
Replace a specific operand.
Definition: Metadata.cpp:851
iterator end()
Definition: Function.h:590
static MDString * get(LLVMContext &Context, StringRef Str)
Definition: Metadata.cpp:446
The main scalar evolution driver.
This file contains the declarations for metadata subclasses.
BlockT * getLoopPreheader() const
If there is a preheader for this loop, return it.
Definition: LoopInfoImpl.h:106
bool isHighCostExpansion(const SCEV *Expr, Loop *L, const Instruction *At=nullptr)
Return true for expressions that may incur non-trivial cost to evaluate at runtime.
A cache of .assume calls within a function.
BasicBlock * getSuccessor(unsigned i) const
STATISTIC(NumFunctions, "Total number of functions")
Metadata node.
Definition: Metadata.h:862
F(f)
const MDOperand & getOperand(unsigned I) const
Definition: Metadata.h:1067
LLVMContext & getContext() const
Get the context in which this basic block lives.
Definition: BasicBlock.cpp:33
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:252
const Module * getModule() const
Return the module owning the function this basic block belongs to, or nullptr it the function does no...
Definition: BasicBlock.cpp:116
Value * removeIncomingValue(unsigned Idx, bool DeletePHIIfEmpty=true)
Remove an incoming value.
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:361
int getBasicBlockIndex(const BasicBlock *BB) const
Return the first index of the specified basic block in the value list for this PHI.
LoopT * getLoopFor(const BlockT *BB) const
Return the inner most loop that BB lives in.
Definition: LoopInfo.h:631
static Loop * CloneLoopBlocks(Loop *L, Value *NewIter, const bool CreateRemainderLoop, const bool UseEpilogRemainder, const bool UnrollRemainder, BasicBlock *InsertTop, BasicBlock *InsertBot, BasicBlock *Preheader, std::vector< BasicBlock *> &NewBlocks, LoopBlocksDFS &LoopBlocks, ValueToValueMapTy &VMap, DominatorTree *DT, LoopInfo *LI)
Create a clone of the blocks in a loop and connect them together.
bool isIntegerTy() const
True if this is an instance of IntegerType.
Definition: Type.h:197
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:664
Value * CreateAdd(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:889
void setName(const Twine &Name)
Change the name of the value.
Definition: Value.cpp:284
RPOIterator endRPO() const
Definition: LoopIterator.h:141
BlockT * getHeader() const
Definition: LoopInfo.h:107
LLVM_NODISCARD LLVM_ATTRIBUTE_ALWAYS_INLINE bool startswith(StringRef Prefix) const
Check if this string starts with the given Prefix.
Definition: StringRef.h:267
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:245
void setLoopID(MDNode *LoopID) const
Set the llvm.loop loop id metadata for this loop.
Definition: LoopInfo.cpp:249
ValueT lookup(const KeyT &Val) const
lookup - Return the entry for the specified key, or a default constructed value if no such entry exis...
Definition: ValueMap.h:167
void perform(LoopInfo *LI)
Traverse the loop blocks and store the DFS result.
Definition: LoopInfo.cpp:764
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:140
If this flag is set, the remapper knows that only local values within a function (such as an instruct...
Definition: ValueMapper.h:65
void getUniqueExitBlocks(SmallVectorImpl< BasicBlock *> &ExitBlocks) const
Return all unique successor blocks of this loop.
Definition: LoopInfo.cpp:360
static void ConnectProlog(Loop *L, Value *BECount, unsigned Count, BasicBlock *PrologExit, BasicBlock *OriginalLoopLatchExit, BasicBlock *PreHeader, BasicBlock *NewPreHeader, ValueToValueMapTy &VMap, DominatorTree *DT, LoopInfo *LI, bool PreserveLCSSA)
Connect the unrolling prolog code to the original loop.
StringRef getString() const
Definition: Metadata.cpp:456
NodeT * getBlock() const
static MDTuple * get(LLVMContext &Context, ArrayRef< Metadata *> MDs)
Definition: Metadata.h:1164
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:406
LoopUnrollStatus UnrollLoop(Loop *L, unsigned Count, unsigned TripCount, bool Force, bool AllowRuntime, bool AllowExpensiveTripCount, bool PreserveCondBr, bool PreserveOnlyFirst, unsigned TripMultiple, unsigned PeelCount, bool UnrollRemainder, LoopInfo *LI, ScalarEvolution *SE, DominatorTree *DT, AssumptionCache *AC, OptimizationRemarkEmitter *ORE, bool PreserveLCSSA)
Unroll the given loop by Count.
Definition: LoopUnroll.cpp:294
static cl::opt< bool > UnrollRuntimeMultiExit("unroll-runtime-multi-exit", cl::init(false), cl::Hidden, cl::desc("Allow runtime unrolling for loops with multiple exits, when " "epilog is generated"))
const Instruction * getFirstNonPHI() const
Returns a pointer to the first instruction in this block that is not a PHINode instruction.
Definition: BasicBlock.cpp:171
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
void setSuccessor(unsigned idx, BasicBlock *B)
Update the specified successor to point at the provided block.
void dump() const
Definition: LoopInfo.cpp:411
const BasicBlock * getSinglePredecessor() const
Return the predecessor of this block if it has a single predecessor block.
Definition: BasicBlock.cpp:217
constexpr bool isPowerOf2_32(uint32_t Value)
Return true if the argument is a power of two > 0.
Definition: MathExtras.h:421
LLVM Basic Block Representation.
Definition: BasicBlock.h:59
This is an important class for using LLVM in a threaded context.
Definition: LLVMContext.h:69
Conditional or Unconditional Branch instruction.
void changeImmediateDominator(DomTreeNodeBase< NodeT > *N, DomTreeNodeBase< NodeT > *NewIDom)
changeImmediateDominator - This method is used to update the dominator tree information when a node&#39;s...
DomTreeNodeBase * getIDom() const
Value * getIncomingValueForBlock(const BasicBlock *BB) const
SmallSet - This maintains a set of unique values, optimizing for the case when the set is small (less...
Definition: SmallSet.h:36
const SCEV * getAddExpr(SmallVectorImpl< const SCEV *> &Ops, SCEV::NoWrapFlags Flags=SCEV::FlagAnyWrap, unsigned Depth=0)
Get a canonical add expression, or something simpler if possible.
size_t size() const
Definition: BasicBlock.h:262
static bool canProfitablyUnrollMultiExitLoop(Loop *L, SmallVectorImpl< BasicBlock *> &OtherExits, BasicBlock *LatchExit, bool PreserveLCSSA, bool UseEpilogRemainder)
Returns true if we can profitably unroll the multi-exit loop L.
void splice(iterator where, iplist_impl &L2)
Definition: ilist.h:342
bool any_of(R &&Range, UnaryPredicate P)
Provide wrappers to std::any_of which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:823
Value * expandCodeFor(const SCEV *SH, Type *Ty, Instruction *I)
Insert code to directly compute the specified SCEV expression into the program.
std::vector< BasicBlock * >::const_reverse_iterator RPOIterator
Definition: LoopIterator.h:102
BasicBlock * SplitBlockPredecessors(BasicBlock *BB, ArrayRef< BasicBlock *> Preds, const char *Suffix, DominatorTree *DT=nullptr, LoopInfo *LI=nullptr, bool PreserveLCSSA=false)
This method introduces at least one new basic block into the function and moves some of the predecess...
bool UnrollRuntimeLoopRemainder(Loop *L, unsigned Count, bool AllowExpensiveTripCount, bool UseEpilogRemainder, bool UnrollRemainder, LoopInfo *LI, ScalarEvolution *SE, DominatorTree *DT, AssumptionCache *AC, OptimizationRemarkEmitter *ORE, bool PreserveLCSSA)
Insert code in the prolog/epilog code when unrolling a loop with a run-time trip-count.
self_iterator getIterator()
Definition: ilist_node.h:82
std::pair< NoneType, bool > insert(const T &V)
insert - Insert an element into the set if it isn&#39;t already there.
Definition: SmallSet.h:81
void getExitingBlocks(SmallVectorImpl< BlockT *> &ExitingBlocks) const
Return all blocks inside the loop that have successors outside of the loop.
Definition: LoopInfoImpl.h:35
static UndefValue * get(Type *T)
Static factory methods - Return an &#39;undef&#39; object of the specified type.
Definition: Constants.cpp:1320
Value * getIncomingValue(unsigned i) const
Return incoming value number x.
static void ConnectEpilog(Loop *L, Value *ModVal, BasicBlock *NewExit, BasicBlock *Exit, BasicBlock *PreHeader, BasicBlock *EpilogPreHeader, BasicBlock *NewPreHeader, ValueToValueMapTy &VMap, DominatorTree *DT, LoopInfo *LI, bool PreserveLCSSA)
Connect the unrolling epilog code to the original loop.
bool contains(const LoopT *L) const
Return true if the specified loop is contained within in this loop.
Definition: LoopInfo.h:117
Iterator for intrusive lists based on ilist_node.
Type * getType() const
Return the LLVM type of this SCEV expression.
void setIncomingBlock(unsigned i, BasicBlock *BB)
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:864
Module.h This file contains the declarations for the Module class.
static Constant * get(Type *Ty, uint64_t V, bool isSigned=false)
If Ty is a vector type, return a Constant with a splat of the given value.
Definition: Constants.cpp:560
static PHINode * Create(Type *Ty, unsigned NumReservedValues, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
Constructors - NumReservedValues is a hint for the number of incoming edges that this phi node will h...
pred_range predecessors(BasicBlock *BB)
Definition: CFG.h:110
unsigned getNumIncomingValues() const
Return the number of incoming edges.
Value * CreateURem(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:987
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:132
unsigned Log2_32(uint32_t Value)
Return the floor log base 2 of the specified value, -1 if the value is zero.
Definition: MathExtras.h:531
const Loop * addClonedBlockToLoopInfo(BasicBlock *OriginalBB, BasicBlock *ClonedBB, LoopInfo *LI, NewLoopsMap &NewLoops)
Adds ClonedBB to LoopInfo, creates a new loop for ClonedBB if necessary and adds a mapping from the o...
Definition: LoopUnroll.cpp:190
Store the result of a depth first search within basic blocks contained by a single loop...
Definition: LoopIterator.h:98
BasicBlock * CloneBasicBlock(const BasicBlock *BB, ValueToValueMapTy &VMap, const Twine &NameSuffix="", Function *F=nullptr, ClonedCodeInfo *CodeInfo=nullptr, DebugInfoFinder *DIFinder=nullptr)
CloneBasicBlock - Return a copy of the specified basic block, but without embedding the block into a ...
void RemapInstruction(Instruction *I, ValueToValueMapTy &VM, RemapFlags Flags=RF_None, ValueMapTypeRemapper *TypeMapper=nullptr, ValueMaterializer *Materializer=nullptr)
Convert the instruction operands from referencing the current values into those specified by VM...
Definition: ValueMapper.h:243
This class uses information about analyze scalars to rewrite expressions in canonical form...
If this flag is set, the remapper ignores missing function-local entries (Argument, Instruction, BasicBlock) that are not in the value map.
Definition: ValueMapper.h:83
LoopT * getParentLoop() const
Definition: LoopInfo.h:108
use_iterator use_begin()
Definition: Value.h:334
bool isLoopSimplifyForm() const
Return true if the Loop is in the form that the LoopSimplify form transforms loops to...
Definition: LoopInfo.cpp:191
MDNode * getLoopID() const
Return the llvm.loop loop id metadata node for this loop if it is present.
Definition: LoopInfo.cpp:213
void forgetLoop(const Loop *L)
This method should be called by the client when it has changed a loop in a way that may effect Scalar...
This class represents an analyzed expression in the program.
Represents a single loop in the control flow graph.
Definition: LoopInfo.h:404
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:218
BasicBlock * getIncomingBlock(unsigned i) const
Return incoming basic block number i.
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:108
#define I(x, y, z)
Definition: MD5.cpp:58
DomTreeNodeBase< NodeT > * addNewBlock(NodeT *BB, NodeT *DomBB)
Add a new node to the dominator tree information.
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:323
const BasicBlockListType & getBasicBlockList() const
Get the underlying elements of the Function...
Definition: Function.h:565
Value * CreateAnd(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1063
void setCondition(Value *V)
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
LLVM Value Representation.
Definition: Value.h:73
succ_range successors(BasicBlock *BB)
Definition: CFG.h:143
RPOIterator beginRPO() const
Reverse iterate over the cached postorder blocks.
Definition: LoopIterator.h:137
bool formDedicatedExitBlocks(Loop *L, DominatorTree *DT, LoopInfo *LI, bool PreserveLCSSA)
Ensure that all exit blocks of the loop are dedicated exits.
Definition: LoopUtils.cpp:941
#define DEBUG(X)
Definition: Debug.h:118
BasicBlock * SplitEdge(BasicBlock *From, BasicBlock *To, DominatorTree *DT=nullptr, LoopInfo *LI=nullptr)
Split the edge connecting specified block.
bool hasOneUse() const
Return true if there is exactly one user of this value.
Definition: Value.h:408
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:49
A single uniqued string.
Definition: Metadata.h:602
const SCEV * getExitCount(const Loop *L, BasicBlock *ExitingBlock)
Get the expression for the number of loop iterations for which this loop is guaranteed not to exit vi...
const TerminatorInst * getTerminator() const LLVM_READONLY
Returns the terminator instruction if the block is well formed or null if the block is not well forme...
Definition: BasicBlock.cpp:120
void setIncomingValue(unsigned i, Value *V)
unsigned getNumOperands() const
Return number of MDNode operands.
Definition: Metadata.h:1073
BlockT * getExitingBlock() const
If getExitingBlocks would return exactly one block, return that block.
Definition: LoopInfoImpl.h:50
The optimization diagnostic interface.
bool erase(const KeyT &Val)
Definition: ValueMap.h:193