LLVM  6.0.0svn
ConstantHoisting.cpp
Go to the documentation of this file.
1 //===- ConstantHoisting.cpp - Prepare code for expensive constants --------===//
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 pass identifies expensive constants to hoist and coalesces them to
11 // better prepare it for SelectionDAG-based code generation. This works around
12 // the limitations of the basic-block-at-a-time approach.
13 //
14 // First it scans all instructions for integer constants and calculates its
15 // cost. If the constant can be folded into the instruction (the cost is
16 // TCC_Free) or the cost is just a simple operation (TCC_BASIC), then we don't
17 // consider it expensive and leave it alone. This is the default behavior and
18 // the default implementation of getIntImmCost will always return TCC_Free.
19 //
20 // If the cost is more than TCC_BASIC, then the integer constant can't be folded
21 // into the instruction and it might be beneficial to hoist the constant.
22 // Similar constants are coalesced to reduce register pressure and
23 // materialization code.
24 //
25 // When a constant is hoisted, it is also hidden behind a bitcast to force it to
26 // be live-out of the basic block. Otherwise the constant would be just
27 // duplicated and each basic block would have its own copy in the SelectionDAG.
28 // The SelectionDAG recognizes such constants as opaque and doesn't perform
29 // certain transformations on them, which would create a new expensive constant.
30 //
31 // This optimization is only applied to integer constants in instructions and
32 // simple (this means not nested) constant cast expressions. For example:
33 // %0 = load i64* inttoptr (i64 big_constant to i64*)
34 //===----------------------------------------------------------------------===//
35 
37 #include "llvm/ADT/APInt.h"
38 #include "llvm/ADT/DenseMap.h"
39 #include "llvm/ADT/None.h"
40 #include "llvm/ADT/Optional.h"
41 #include "llvm/ADT/SmallPtrSet.h"
42 #include "llvm/ADT/SmallVector.h"
43 #include "llvm/ADT/Statistic.h"
46 #include "llvm/IR/BasicBlock.h"
47 #include "llvm/IR/Constants.h"
48 #include "llvm/IR/Dominators.h"
49 #include "llvm/IR/Function.h"
50 #include "llvm/IR/InstrTypes.h"
51 #include "llvm/IR/Instruction.h"
52 #include "llvm/IR/Instructions.h"
53 #include "llvm/IR/IntrinsicInst.h"
54 #include "llvm/IR/Value.h"
55 #include "llvm/Pass.h"
57 #include "llvm/Support/Casting.h"
59 #include "llvm/Support/Debug.h"
61 #include "llvm/Transforms/Scalar.h"
64 #include <algorithm>
65 #include <cassert>
66 #include <cstdint>
67 #include <iterator>
68 #include <tuple>
69 #include <utility>
70 
71 using namespace llvm;
72 using namespace consthoist;
73 
74 #define DEBUG_TYPE "consthoist"
75 
76 STATISTIC(NumConstantsHoisted, "Number of constants hoisted");
77 STATISTIC(NumConstantsRebased, "Number of constants rebased");
78 
80  "consthoist-with-block-frequency", cl::init(true), cl::Hidden,
81  cl::desc("Enable the use of the block frequency analysis to reduce the "
82  "chance to execute const materialization more frequently than "
83  "without hoisting."));
84 
85 namespace {
86 
87 /// \brief The constant hoisting pass.
88 class ConstantHoistingLegacyPass : public FunctionPass {
89 public:
90  static char ID; // Pass identification, replacement for typeid
91 
92  ConstantHoistingLegacyPass() : FunctionPass(ID) {
94  }
95 
96  bool runOnFunction(Function &Fn) override;
97 
98  StringRef getPassName() const override { return "Constant Hoisting"; }
99 
100  void getAnalysisUsage(AnalysisUsage &AU) const override {
101  AU.setPreservesCFG();
106  }
107 
108  void releaseMemory() override { Impl.releaseMemory(); }
109 
110 private:
112 };
113 
114 } // end anonymous namespace
115 
117 
118 INITIALIZE_PASS_BEGIN(ConstantHoistingLegacyPass, "consthoist",
119  "Constant Hoisting", false, false)
123 INITIALIZE_PASS_END(ConstantHoistingLegacyPass, "consthoist",
124  "Constant Hoisting", false, false)
125 
127  return new ConstantHoistingLegacyPass();
128 }
129 
130 /// \brief Perform the constant hoisting optimization for the given function.
132  if (skipFunction(Fn))
133  return false;
134 
135  DEBUG(dbgs() << "********** Begin Constant Hoisting **********\n");
136  DEBUG(dbgs() << "********** Function: " << Fn.getName() << '\n');
137 
138  bool MadeChange =
139  Impl.runImpl(Fn, getAnalysis<TargetTransformInfoWrapperPass>().getTTI(Fn),
140  getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
142  ? &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI()
143  : nullptr,
144  Fn.getEntryBlock());
145 
146  if (MadeChange) {
147  DEBUG(dbgs() << "********** Function after Constant Hoisting: "
148  << Fn.getName() << '\n');
149  DEBUG(dbgs() << Fn);
150  }
151  DEBUG(dbgs() << "********** End Constant Hoisting **********\n");
152 
153  return MadeChange;
154 }
155 
156 /// \brief Find the constant materialization insertion point.
157 Instruction *ConstantHoistingPass::findMatInsertPt(Instruction *Inst,
158  unsigned Idx) const {
159  // If the operand is a cast instruction, then we have to materialize the
160  // constant before the cast instruction.
161  if (Idx != ~0U) {
162  Value *Opnd = Inst->getOperand(Idx);
163  if (auto CastInst = dyn_cast<Instruction>(Opnd))
164  if (CastInst->isCast())
165  return CastInst;
166  }
167 
168  // The simple and common case. This also includes constant expressions.
169  if (!isa<PHINode>(Inst) && !Inst->isEHPad())
170  return Inst;
171 
172  // We can't insert directly before a phi node or an eh pad. Insert before
173  // the terminator of the incoming or dominating block.
174  assert(Entry != Inst->getParent() && "PHI or landing pad in entry block!");
175  if (Idx != ~0U && isa<PHINode>(Inst))
176  return cast<PHINode>(Inst)->getIncomingBlock(Idx)->getTerminator();
177 
178  // This must be an EH pad. Iterate over immediate dominators until we find a
179  // non-EH pad. We need to skip over catchswitch blocks, which are both EH pads
180  // and terminators.
181  auto IDom = DT->getNode(Inst->getParent())->getIDom();
182  while (IDom->getBlock()->isEHPad()) {
183  assert(Entry != IDom->getBlock() && "eh pad in entry block");
184  IDom = IDom->getIDom();
185  }
186 
187  return IDom->getBlock()->getTerminator();
188 }
189 
190 /// \brief Given \p BBs as input, find another set of BBs which collectively
191 /// dominates \p BBs and have the minimal sum of frequencies. Return the BB
192 /// set found in \p BBs.
194  BasicBlock *Entry,
196  assert(!BBs.count(Entry) && "Assume Entry is not in BBs");
197  // Nodes on the current path to the root.
199  // Candidates includes any block 'BB' in set 'BBs' that is not strictly
200  // dominated by any other blocks in set 'BBs', and all nodes in the path
201  // in the dominator tree from Entry to 'BB'.
203  for (auto BB : BBs) {
204  Path.clear();
205  // Walk up the dominator tree until Entry or another BB in BBs
206  // is reached. Insert the nodes on the way to the Path.
207  BasicBlock *Node = BB;
208  // The "Path" is a candidate path to be added into Candidates set.
209  bool isCandidate = false;
210  do {
211  Path.insert(Node);
212  if (Node == Entry || Candidates.count(Node)) {
213  isCandidate = true;
214  break;
215  }
216  assert(DT.getNode(Node)->getIDom() &&
217  "Entry doens't dominate current Node");
218  Node = DT.getNode(Node)->getIDom()->getBlock();
219  } while (!BBs.count(Node));
220 
221  // If isCandidate is false, Node is another Block in BBs dominating
222  // current 'BB'. Drop the nodes on the Path.
223  if (!isCandidate)
224  continue;
225 
226  // Add nodes on the Path into Candidates.
227  Candidates.insert(Path.begin(), Path.end());
228  }
229 
230  // Sort the nodes in Candidates in top-down order and save the nodes
231  // in Orders.
232  unsigned Idx = 0;
234  Orders.push_back(Entry);
235  while (Idx != Orders.size()) {
236  BasicBlock *Node = Orders[Idx++];
237  for (auto ChildDomNode : DT.getNode(Node)->getChildren()) {
238  if (Candidates.count(ChildDomNode->getBlock()))
239  Orders.push_back(ChildDomNode->getBlock());
240  }
241  }
242 
243  // Visit Orders in bottom-up order.
244  using InsertPtsCostPair =
245  std::pair<SmallPtrSet<BasicBlock *, 16>, BlockFrequency>;
246 
247  // InsertPtsMap is a map from a BB to the best insertion points for the
248  // subtree of BB (subtree not including the BB itself).
250  InsertPtsMap.reserve(Orders.size() + 1);
251  for (auto RIt = Orders.rbegin(); RIt != Orders.rend(); RIt++) {
252  BasicBlock *Node = *RIt;
253  bool NodeInBBs = BBs.count(Node);
254  SmallPtrSet<BasicBlock *, 16> &InsertPts = InsertPtsMap[Node].first;
255  BlockFrequency &InsertPtsFreq = InsertPtsMap[Node].second;
256 
257  // Return the optimal insert points in BBs.
258  if (Node == Entry) {
259  BBs.clear();
260  if (InsertPtsFreq > BFI.getBlockFreq(Node) ||
261  (InsertPtsFreq == BFI.getBlockFreq(Node) && InsertPts.size() > 1))
262  BBs.insert(Entry);
263  else
264  BBs.insert(InsertPts.begin(), InsertPts.end());
265  break;
266  }
267 
268  BasicBlock *Parent = DT.getNode(Node)->getIDom()->getBlock();
269  // Initially, ParentInsertPts is empty and ParentPtsFreq is 0. Every child
270  // will update its parent's ParentInsertPts and ParentPtsFreq.
271  SmallPtrSet<BasicBlock *, 16> &ParentInsertPts = InsertPtsMap[Parent].first;
272  BlockFrequency &ParentPtsFreq = InsertPtsMap[Parent].second;
273  // Choose to insert in Node or in subtree of Node.
274  // Don't hoist to EHPad because we may not find a proper place to insert
275  // in EHPad.
276  // If the total frequency of InsertPts is the same as the frequency of the
277  // target Node, and InsertPts contains more than one nodes, choose hoisting
278  // to reduce code size.
279  if (NodeInBBs ||
280  (!Node->isEHPad() &&
281  (InsertPtsFreq > BFI.getBlockFreq(Node) ||
282  (InsertPtsFreq == BFI.getBlockFreq(Node) && InsertPts.size() > 1)))) {
283  ParentInsertPts.insert(Node);
284  ParentPtsFreq += BFI.getBlockFreq(Node);
285  } else {
286  ParentInsertPts.insert(InsertPts.begin(), InsertPts.end());
287  ParentPtsFreq += InsertPtsFreq;
288  }
289  }
290 }
291 
292 /// \brief Find an insertion point that dominates all uses.
293 SmallPtrSet<Instruction *, 8> ConstantHoistingPass::findConstantInsertionPoint(
294  const ConstantInfo &ConstInfo) const {
295  assert(!ConstInfo.RebasedConstants.empty() && "Invalid constant info entry.");
296  // Collect all basic blocks.
299  for (auto const &RCI : ConstInfo.RebasedConstants)
300  for (auto const &U : RCI.Uses)
301  BBs.insert(findMatInsertPt(U.Inst, U.OpndIdx)->getParent());
302 
303  if (BBs.count(Entry)) {
304  InsertPts.insert(&Entry->front());
305  return InsertPts;
306  }
307 
308  if (BFI) {
309  findBestInsertionSet(*DT, *BFI, Entry, BBs);
310  for (auto BB : BBs) {
311  BasicBlock::iterator InsertPt = BB->begin();
312  for (; isa<PHINode>(InsertPt) || InsertPt->isEHPad(); ++InsertPt)
313  ;
314  InsertPts.insert(&*InsertPt);
315  }
316  return InsertPts;
317  }
318 
319  while (BBs.size() >= 2) {
320  BasicBlock *BB, *BB1, *BB2;
321  BB1 = *BBs.begin();
322  BB2 = *std::next(BBs.begin());
323  BB = DT->findNearestCommonDominator(BB1, BB2);
324  if (BB == Entry) {
325  InsertPts.insert(&Entry->front());
326  return InsertPts;
327  }
328  BBs.erase(BB1);
329  BBs.erase(BB2);
330  BBs.insert(BB);
331  }
332  assert((BBs.size() == 1) && "Expected only one element.");
333  Instruction &FirstInst = (*BBs.begin())->front();
334  InsertPts.insert(findMatInsertPt(&FirstInst));
335  return InsertPts;
336 }
337 
338 /// \brief Record constant integer ConstInt for instruction Inst at operand
339 /// index Idx.
340 ///
341 /// The operand at index Idx is not necessarily the constant integer itself. It
342 /// could also be a cast instruction or a constant expression that uses the
343 // constant integer.
344 void ConstantHoistingPass::collectConstantCandidates(
345  ConstCandMapType &ConstCandMap, Instruction *Inst, unsigned Idx,
346  ConstantInt *ConstInt) {
347  unsigned Cost;
348  // Ask the target about the cost of materializing the constant for the given
349  // instruction and operand index.
350  if (auto IntrInst = dyn_cast<IntrinsicInst>(Inst))
351  Cost = TTI->getIntImmCost(IntrInst->getIntrinsicID(), Idx,
352  ConstInt->getValue(), ConstInt->getType());
353  else
354  Cost = TTI->getIntImmCost(Inst->getOpcode(), Idx, ConstInt->getValue(),
355  ConstInt->getType());
356 
357  // Ignore cheap integer constants.
358  if (Cost > TargetTransformInfo::TCC_Basic) {
359  ConstCandMapType::iterator Itr;
360  bool Inserted;
361  std::tie(Itr, Inserted) = ConstCandMap.insert(std::make_pair(ConstInt, 0));
362  if (Inserted) {
363  ConstCandVec.push_back(ConstantCandidate(ConstInt));
364  Itr->second = ConstCandVec.size() - 1;
365  }
366  ConstCandVec[Itr->second].addUser(Inst, Idx, Cost);
367  DEBUG(if (isa<ConstantInt>(Inst->getOperand(Idx)))
368  dbgs() << "Collect constant " << *ConstInt << " from " << *Inst
369  << " with cost " << Cost << '\n';
370  else
371  dbgs() << "Collect constant " << *ConstInt << " indirectly from "
372  << *Inst << " via " << *Inst->getOperand(Idx) << " with cost "
373  << Cost << '\n';
374  );
375  }
376 }
377 
378 /// \brief Check the operand for instruction Inst at index Idx.
379 void ConstantHoistingPass::collectConstantCandidates(
380  ConstCandMapType &ConstCandMap, Instruction *Inst, unsigned Idx) {
381  Value *Opnd = Inst->getOperand(Idx);
382 
383  // Visit constant integers.
384  if (auto ConstInt = dyn_cast<ConstantInt>(Opnd)) {
385  collectConstantCandidates(ConstCandMap, Inst, Idx, ConstInt);
386  return;
387  }
388 
389  // Visit cast instructions that have constant integers.
390  if (auto CastInst = dyn_cast<Instruction>(Opnd)) {
391  // Only visit cast instructions, which have been skipped. All other
392  // instructions should have already been visited.
393  if (!CastInst->isCast())
394  return;
395 
396  if (auto *ConstInt = dyn_cast<ConstantInt>(CastInst->getOperand(0))) {
397  // Pretend the constant is directly used by the instruction and ignore
398  // the cast instruction.
399  collectConstantCandidates(ConstCandMap, Inst, Idx, ConstInt);
400  return;
401  }
402  }
403 
404  // Visit constant expressions that have constant integers.
405  if (auto ConstExpr = dyn_cast<ConstantExpr>(Opnd)) {
406  // Only visit constant cast expressions.
407  if (!ConstExpr->isCast())
408  return;
409 
410  if (auto ConstInt = dyn_cast<ConstantInt>(ConstExpr->getOperand(0))) {
411  // Pretend the constant is directly used by the instruction and ignore
412  // the constant expression.
413  collectConstantCandidates(ConstCandMap, Inst, Idx, ConstInt);
414  return;
415  }
416  }
417 }
418 
419 /// \brief Scan the instruction for expensive integer constants and record them
420 /// in the constant candidate vector.
421 void ConstantHoistingPass::collectConstantCandidates(
422  ConstCandMapType &ConstCandMap, Instruction *Inst) {
423  // Skip all cast instructions. They are visited indirectly later on.
424  if (Inst->isCast())
425  return;
426 
427  // Scan all operands.
428  for (unsigned Idx = 0, E = Inst->getNumOperands(); Idx != E; ++Idx) {
429  // The cost of materializing the constants (defined in
430  // `TargetTransformInfo::getIntImmCost`) for instructions which only take
431  // constant variables is lower than `TargetTransformInfo::TCC_Basic`. So
432  // it's safe for us to collect constant candidates from all IntrinsicInsts.
433  if (canReplaceOperandWithVariable(Inst, Idx) || isa<IntrinsicInst>(Inst)) {
434  collectConstantCandidates(ConstCandMap, Inst, Idx);
435  }
436  } // end of for all operands
437 }
438 
439 /// \brief Collect all integer constants in the function that cannot be folded
440 /// into an instruction itself.
441 void ConstantHoistingPass::collectConstantCandidates(Function &Fn) {
442  ConstCandMapType ConstCandMap;
443  for (BasicBlock &BB : Fn)
444  for (Instruction &Inst : BB)
445  collectConstantCandidates(ConstCandMap, &Inst);
446 }
447 
448 // This helper function is necessary to deal with values that have different
449 // bit widths (APInt Operator- does not like that). If the value cannot be
450 // represented in uint64 we return an "empty" APInt. This is then interpreted
451 // as the value is not in range.
452 static Optional<APInt> calculateOffsetDiff(const APInt &V1, const APInt &V2) {
453  Optional<APInt> Res = None;
454  unsigned BW = V1.getBitWidth() > V2.getBitWidth() ?
455  V1.getBitWidth() : V2.getBitWidth();
456  uint64_t LimVal1 = V1.getLimitedValue();
457  uint64_t LimVal2 = V2.getLimitedValue();
458 
459  if (LimVal1 == ~0ULL || LimVal2 == ~0ULL)
460  return Res;
461 
462  uint64_t Diff = LimVal1 - LimVal2;
463  return APInt(BW, Diff, true);
464 }
465 
466 // From a list of constants, one needs to picked as the base and the other
467 // constants will be transformed into an offset from that base constant. The
468 // question is which we can pick best? For example, consider these constants
469 // and their number of uses:
470 //
471 // Constants| 2 | 4 | 12 | 42 |
472 // NumUses | 3 | 2 | 8 | 7 |
473 //
474 // Selecting constant 12 because it has the most uses will generate negative
475 // offsets for constants 2 and 4 (i.e. -10 and -8 respectively). If negative
476 // offsets lead to less optimal code generation, then there might be better
477 // solutions. Suppose immediates in the range of 0..35 are most optimally
478 // supported by the architecture, then selecting constant 2 is most optimal
479 // because this will generate offsets: 0, 2, 10, 40. Offsets 0, 2 and 10 are in
480 // range 0..35, and thus 3 + 2 + 8 = 13 uses are in range. Selecting 12 would
481 // have only 8 uses in range, so choosing 2 as a base is more optimal. Thus, in
482 // selecting the base constant the range of the offsets is a very important
483 // factor too that we take into account here. This algorithm calculates a total
484 // costs for selecting a constant as the base and substract the costs if
485 // immediates are out of range. It has quadratic complexity, so we call this
486 // function only when we're optimising for size and there are less than 100
487 // constants, we fall back to the straightforward algorithm otherwise
488 // which does not do all the offset calculations.
489 unsigned
490 ConstantHoistingPass::maximizeConstantsInRange(ConstCandVecType::iterator S,
491  ConstCandVecType::iterator E,
492  ConstCandVecType::iterator &MaxCostItr) {
493  unsigned NumUses = 0;
494 
495  if(!Entry->getParent()->optForSize() || std::distance(S,E) > 100) {
496  for (auto ConstCand = S; ConstCand != E; ++ConstCand) {
497  NumUses += ConstCand->Uses.size();
498  if (ConstCand->CumulativeCost > MaxCostItr->CumulativeCost)
499  MaxCostItr = ConstCand;
500  }
501  return NumUses;
502  }
503 
504  DEBUG(dbgs() << "== Maximize constants in range ==\n");
505  int MaxCost = -1;
506  for (auto ConstCand = S; ConstCand != E; ++ConstCand) {
507  auto Value = ConstCand->ConstInt->getValue();
508  Type *Ty = ConstCand->ConstInt->getType();
509  int Cost = 0;
510  NumUses += ConstCand->Uses.size();
511  DEBUG(dbgs() << "= Constant: " << ConstCand->ConstInt->getValue() << "\n");
512 
513  for (auto User : ConstCand->Uses) {
514  unsigned Opcode = User.Inst->getOpcode();
515  unsigned OpndIdx = User.OpndIdx;
516  Cost += TTI->getIntImmCost(Opcode, OpndIdx, Value, Ty);
517  DEBUG(dbgs() << "Cost: " << Cost << "\n");
518 
519  for (auto C2 = S; C2 != E; ++C2) {
521  C2->ConstInt->getValue(),
522  ConstCand->ConstInt->getValue());
523  if (Diff) {
524  const int ImmCosts =
525  TTI->getIntImmCodeSizeCost(Opcode, OpndIdx, Diff.getValue(), Ty);
526  Cost -= ImmCosts;
527  DEBUG(dbgs() << "Offset " << Diff.getValue() << " "
528  << "has penalty: " << ImmCosts << "\n"
529  << "Adjusted cost: " << Cost << "\n");
530  }
531  }
532  }
533  DEBUG(dbgs() << "Cumulative cost: " << Cost << "\n");
534  if (Cost > MaxCost) {
535  MaxCost = Cost;
536  MaxCostItr = ConstCand;
537  DEBUG(dbgs() << "New candidate: " << MaxCostItr->ConstInt->getValue()
538  << "\n");
539  }
540  }
541  return NumUses;
542 }
543 
544 /// \brief Find the base constant within the given range and rebase all other
545 /// constants with respect to the base constant.
546 void ConstantHoistingPass::findAndMakeBaseConstant(
547  ConstCandVecType::iterator S, ConstCandVecType::iterator E) {
548  auto MaxCostItr = S;
549  unsigned NumUses = maximizeConstantsInRange(S, E, MaxCostItr);
550 
551  // Don't hoist constants that have only one use.
552  if (NumUses <= 1)
553  return;
554 
555  ConstantInfo ConstInfo;
556  ConstInfo.BaseConstant = MaxCostItr->ConstInt;
557  Type *Ty = ConstInfo.BaseConstant->getType();
558 
559  // Rebase the constants with respect to the base constant.
560  for (auto ConstCand = S; ConstCand != E; ++ConstCand) {
561  APInt Diff = ConstCand->ConstInt->getValue() -
562  ConstInfo.BaseConstant->getValue();
563  Constant *Offset = Diff == 0 ? nullptr : ConstantInt::get(Ty, Diff);
564  ConstInfo.RebasedConstants.push_back(
565  RebasedConstantInfo(std::move(ConstCand->Uses), Offset));
566  }
567  ConstantVec.push_back(std::move(ConstInfo));
568 }
569 
570 /// \brief Finds and combines constant candidates that can be easily
571 /// rematerialized with an add from a common base constant.
572 void ConstantHoistingPass::findBaseConstants() {
573  // Sort the constants by value and type. This invalidates the mapping!
574  std::sort(ConstCandVec.begin(), ConstCandVec.end(),
575  [](const ConstantCandidate &LHS, const ConstantCandidate &RHS) {
576  if (LHS.ConstInt->getType() != RHS.ConstInt->getType())
577  return LHS.ConstInt->getType()->getBitWidth() <
578  RHS.ConstInt->getType()->getBitWidth();
579  return LHS.ConstInt->getValue().ult(RHS.ConstInt->getValue());
580  });
581 
582  // Simple linear scan through the sorted constant candidate vector for viable
583  // merge candidates.
584  auto MinValItr = ConstCandVec.begin();
585  for (auto CC = std::next(ConstCandVec.begin()), E = ConstCandVec.end();
586  CC != E; ++CC) {
587  if (MinValItr->ConstInt->getType() == CC->ConstInt->getType()) {
588  // Check if the constant is in range of an add with immediate.
589  APInt Diff = CC->ConstInt->getValue() - MinValItr->ConstInt->getValue();
590  if ((Diff.getBitWidth() <= 64) &&
591  TTI->isLegalAddImmediate(Diff.getSExtValue()))
592  continue;
593  }
594  // We either have now a different constant type or the constant is not in
595  // range of an add with immediate anymore.
596  findAndMakeBaseConstant(MinValItr, CC);
597  // Start a new base constant search.
598  MinValItr = CC;
599  }
600  // Finalize the last base constant search.
601  findAndMakeBaseConstant(MinValItr, ConstCandVec.end());
602 }
603 
604 /// \brief Updates the operand at Idx in instruction Inst with the result of
605 /// instruction Mat. If the instruction is a PHI node then special
606 /// handling for duplicate values form the same incoming basic block is
607 /// required.
608 /// \return The update will always succeed, but the return value indicated if
609 /// Mat was used for the update or not.
610 static bool updateOperand(Instruction *Inst, unsigned Idx, Instruction *Mat) {
611  if (auto PHI = dyn_cast<PHINode>(Inst)) {
612  // Check if any previous operand of the PHI node has the same incoming basic
613  // block. This is a very odd case that happens when the incoming basic block
614  // has a switch statement. In this case use the same value as the previous
615  // operand(s), otherwise we will fail verification due to different values.
616  // The values are actually the same, but the variable names are different
617  // and the verifier doesn't like that.
618  BasicBlock *IncomingBB = PHI->getIncomingBlock(Idx);
619  for (unsigned i = 0; i < Idx; ++i) {
620  if (PHI->getIncomingBlock(i) == IncomingBB) {
621  Value *IncomingVal = PHI->getIncomingValue(i);
622  Inst->setOperand(Idx, IncomingVal);
623  return false;
624  }
625  }
626  }
627 
628  Inst->setOperand(Idx, Mat);
629  return true;
630 }
631 
632 /// \brief Emit materialization code for all rebased constants and update their
633 /// users.
634 void ConstantHoistingPass::emitBaseConstants(Instruction *Base,
635  Constant *Offset,
636  const ConstantUser &ConstUser) {
637  Instruction *Mat = Base;
638  if (Offset) {
639  Instruction *InsertionPt = findMatInsertPt(ConstUser.Inst,
640  ConstUser.OpndIdx);
641  Mat = BinaryOperator::Create(Instruction::Add, Base, Offset,
642  "const_mat", InsertionPt);
643 
644  DEBUG(dbgs() << "Materialize constant (" << *Base->getOperand(0)
645  << " + " << *Offset << ") in BB "
646  << Mat->getParent()->getName() << '\n' << *Mat << '\n');
647  Mat->setDebugLoc(ConstUser.Inst->getDebugLoc());
648  }
649  Value *Opnd = ConstUser.Inst->getOperand(ConstUser.OpndIdx);
650 
651  // Visit constant integer.
652  if (isa<ConstantInt>(Opnd)) {
653  DEBUG(dbgs() << "Update: " << *ConstUser.Inst << '\n');
654  if (!updateOperand(ConstUser.Inst, ConstUser.OpndIdx, Mat) && Offset)
655  Mat->eraseFromParent();
656  DEBUG(dbgs() << "To : " << *ConstUser.Inst << '\n');
657  return;
658  }
659 
660  // Visit cast instruction.
661  if (auto CastInst = dyn_cast<Instruction>(Opnd)) {
662  assert(CastInst->isCast() && "Expected an cast instruction!");
663  // Check if we already have visited this cast instruction before to avoid
664  // unnecessary cloning.
665  Instruction *&ClonedCastInst = ClonedCastMap[CastInst];
666  if (!ClonedCastInst) {
667  ClonedCastInst = CastInst->clone();
668  ClonedCastInst->setOperand(0, Mat);
669  ClonedCastInst->insertAfter(CastInst);
670  // Use the same debug location as the original cast instruction.
671  ClonedCastInst->setDebugLoc(CastInst->getDebugLoc());
672  DEBUG(dbgs() << "Clone instruction: " << *CastInst << '\n'
673  << "To : " << *ClonedCastInst << '\n');
674  }
675 
676  DEBUG(dbgs() << "Update: " << *ConstUser.Inst << '\n');
677  updateOperand(ConstUser.Inst, ConstUser.OpndIdx, ClonedCastInst);
678  DEBUG(dbgs() << "To : " << *ConstUser.Inst << '\n');
679  return;
680  }
681 
682  // Visit constant expression.
683  if (auto ConstExpr = dyn_cast<ConstantExpr>(Opnd)) {
684  Instruction *ConstExprInst = ConstExpr->getAsInstruction();
685  ConstExprInst->setOperand(0, Mat);
686  ConstExprInst->insertBefore(findMatInsertPt(ConstUser.Inst,
687  ConstUser.OpndIdx));
688 
689  // Use the same debug location as the instruction we are about to update.
690  ConstExprInst->setDebugLoc(ConstUser.Inst->getDebugLoc());
691 
692  DEBUG(dbgs() << "Create instruction: " << *ConstExprInst << '\n'
693  << "From : " << *ConstExpr << '\n');
694  DEBUG(dbgs() << "Update: " << *ConstUser.Inst << '\n');
695  if (!updateOperand(ConstUser.Inst, ConstUser.OpndIdx, ConstExprInst)) {
696  ConstExprInst->eraseFromParent();
697  if (Offset)
698  Mat->eraseFromParent();
699  }
700  DEBUG(dbgs() << "To : " << *ConstUser.Inst << '\n');
701  return;
702  }
703 }
704 
705 /// \brief Hoist and hide the base constant behind a bitcast and emit
706 /// materialization code for derived constants.
707 bool ConstantHoistingPass::emitBaseConstants() {
708  bool MadeChange = false;
709  for (auto const &ConstInfo : ConstantVec) {
710  // Hoist and hide the base constant behind a bitcast.
711  SmallPtrSet<Instruction *, 8> IPSet = findConstantInsertionPoint(ConstInfo);
712  assert(!IPSet.empty() && "IPSet is empty");
713 
714  unsigned UsesNum = 0;
715  unsigned ReBasesNum = 0;
716  for (Instruction *IP : IPSet) {
717  IntegerType *Ty = ConstInfo.BaseConstant->getType();
718  Instruction *Base =
719  new BitCastInst(ConstInfo.BaseConstant, Ty, "const", IP);
720 
721  Base->setDebugLoc(IP->getDebugLoc());
722 
723  DEBUG(dbgs() << "Hoist constant (" << *ConstInfo.BaseConstant
724  << ") to BB " << IP->getParent()->getName() << '\n'
725  << *Base << '\n');
726 
727  // Emit materialization code for all rebased constants.
728  unsigned Uses = 0;
729  for (auto const &RCI : ConstInfo.RebasedConstants) {
730  for (auto const &U : RCI.Uses) {
731  Uses++;
732  BasicBlock *OrigMatInsertBB =
733  findMatInsertPt(U.Inst, U.OpndIdx)->getParent();
734  // If Base constant is to be inserted in multiple places,
735  // generate rebase for U using the Base dominating U.
736  if (IPSet.size() == 1 ||
737  DT->dominates(Base->getParent(), OrigMatInsertBB)) {
738  emitBaseConstants(Base, RCI.Offset, U);
739  ReBasesNum++;
740  }
741 
742  Base->setDebugLoc(DILocation::getMergedLocation(Base->getDebugLoc(), U.Inst->getDebugLoc()));
743  }
744  }
745  UsesNum = Uses;
746 
747  // Use the same debug location as the last user of the constant.
748  assert(!Base->use_empty() && "The use list is empty!?");
749  assert(isa<Instruction>(Base->user_back()) &&
750  "All uses should be instructions.");
751  }
752  (void)UsesNum;
753  (void)ReBasesNum;
754  // Expect all uses are rebased after rebase is done.
755  assert(UsesNum == ReBasesNum && "Not all uses are rebased");
756 
757  NumConstantsHoisted++;
758 
759  // Base constant is also included in ConstInfo.RebasedConstants, so
760  // deduct 1 from ConstInfo.RebasedConstants.size().
761  NumConstantsRebased = ConstInfo.RebasedConstants.size() - 1;
762 
763  MadeChange = true;
764  }
765  return MadeChange;
766 }
767 
768 /// \brief Check all cast instructions we made a copy of and remove them if they
769 /// have no more users.
770 void ConstantHoistingPass::deleteDeadCastInst() const {
771  for (auto const &I : ClonedCastMap)
772  if (I.first->use_empty())
773  I.first->eraseFromParent();
774 }
775 
776 /// \brief Optimize expensive integer constants in the given function.
779  BasicBlock &Entry) {
780  this->TTI = &TTI;
781  this->DT = &DT;
782  this->BFI = BFI;
783  this->Entry = &Entry;
784  // Collect all constant candidates.
785  collectConstantCandidates(Fn);
786 
787  // There are no constant candidates to worry about.
788  if (ConstCandVec.empty())
789  return false;
790 
791  // Combine constants that can be easily materialized with an add from a common
792  // base constant.
793  findBaseConstants();
794 
795  // There are no constants to emit.
796  if (ConstantVec.empty())
797  return false;
798 
799  // Finally hoist the base constant and emit materialization code for dependent
800  // constants.
801  bool MadeChange = emitBaseConstants();
802 
803  // Cleanup dead instructions.
804  deleteDeadCastInst();
805 
806  return MadeChange;
807 }
808 
811  auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
812  auto &TTI = AM.getResult<TargetIRAnalysis>(F);
815  : nullptr;
816  if (!runImpl(F, TTI, DT, BFI, F.getEntryBlock()))
817  return PreservedAnalyses::all();
818 
820  PA.preserveSet<CFGAnalyses>();
821  return PA;
822 }
const NoneType None
Definition: None.h:24
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
IntegerType * getType() const
getType - Specialize the getType() method to always return an IntegerType, which reduces the amount o...
Definition: Constants.h:172
INITIALIZE_PASS_BEGIN(ConstantHoistingLegacyPass, "consthoist", "Constant Hoisting", false, false) INITIALIZE_PASS_END(ConstantHoistingLegacyPass
static void findBestInsertionSet(DominatorTree &DT, BlockFrequencyInfo &BFI, BasicBlock *Entry, SmallPtrSet< BasicBlock *, 8 > &BBs)
Given BBs as input, find another set of BBs which collectively dominates BBs and have the minimal sum...
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
A base constant and all its rebased constants.
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:687
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
void initializeConstantHoistingLegacyPassPass(PassRegistry &)
LLVM_ATTRIBUTE_ALWAYS_INLINE size_type size() const
Definition: SmallVector.h:136
virtual void releaseMemory()
releaseMemory() - This member can be implemented by a pass if it wants to be able to release its memo...
Definition: Pass.cpp:95
Keeps track of a constant candidate and its uses.
Analysis pass providing the TargetTransformInfo.
static bool updateOperand(Instruction *Inst, unsigned Idx, Instruction *Mat)
Updates the operand at Idx in instruction Inst with the result of instruction Mat.
STATISTIC(NumFunctions, "Total number of functions")
Analysis pass which computes a DominatorTree.
Definition: Dominators.h:238
F(f)
This represents a constant that has been rebased with respect to a base constant. ...
unsigned getBitWidth() const
Return the number of bits in the APInt.
Definition: APInt.h:1488
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:252
AnalysisUsage & addRequired()
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:51
This is the base class for all instructions that perform data casts.
Definition: InstrTypes.h:560
Constant Hoisting
static bool runImpl(CallGraphSCC &SCC, AARGetterT AARGetter)
Legacy analysis pass which computes BlockFrequencyInfo.
This file implements a class to represent arbitrary precision integral constant values and operations...
Instruction * clone() const
Create a copy of &#39;this&#39; instruction that is identical in all ways except the following: ...
int64_t getSExtValue() const
Get sign extended value.
Definition: APInt.h:1554
const T & getValue() const LLVM_LVALUE_FUNCTION
Definition: Optional.h:127
This class represents a no-op cast from one type to another.
const std::vector< DomTreeNodeBase * > & getChildren() const
const APInt & getValue() const
Return the constant as an APInt value reference.
Definition: Constants.h:138
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:125
unsigned getBitWidth() const
Get the number of bits in this IntegerType.
Definition: DerivedTypes.h:66
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:140
Value * getOperand(unsigned i) const
Definition: User.h:154
const BasicBlock & getEntryBlock() const
Definition: Function.h:572
NodeT * getBlock() const
BlockFrequencyInfo pass uses BlockFrequencyInfoImpl implementation to estimate IR basic block frequen...
static bool runOnFunction(Function &F, bool PostInlining)
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:406
Wrapper pass for TargetTransformInfo.
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:153
void setDebugLoc(DebugLoc Loc)
Set the debug location information for this instruction.
Definition: Instruction.h:281
void insertBefore(Instruction *InsertPos)
Insert an unlinked instruction into a basic block immediately before the specified instruction...
Definition: Instruction.cpp:75
LLVM Basic Block Representation.
Definition: BasicBlock.h:59
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:46
bool ult(const APInt &RHS) const
Unsigned less than comparison.
Definition: APInt.h:1164
DomTreeNodeBase * getIDom() const
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
This is an important base class in LLVM.
Definition: Constant.h:42
LLVM_NODISCARD bool empty() const
Definition: SmallPtrSet.h:92
bool canReplaceOperandWithVariable(const Instruction *I, unsigned OpIdx)
Given an instruction, is it legal to set operand OpIdx to a non-constant value?
Definition: Local.cpp:2230
This file contains the declarations for the subclasses of Constant, which represent the different fla...
std::pair< iterator, bool > insert(PtrType Ptr)
Inserts Ptr if and only if there is no element in the container equal to Ptr.
Definition: SmallPtrSet.h:371
Represent the analysis usage information of a pass.
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:285
void reserve(size_type NumEntries)
Grow the densemap so that it can contain at least NumEntries items before resizing again...
Definition: DenseMap.h:100
size_type count(ConstPtrType Ptr) const
count - Return 1 if the specified pointer is in the set, 0 otherwise.
Definition: SmallPtrSet.h:382
Class to represent integer types.
Definition: DerivedTypes.h:40
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:159
bool isCast() const
Definition: Instruction.h:131
static cl::opt< bool > ConstHoistWithBlockFrequency("consthoist-with-block-frequency", cl::init(true), cl::Hidden, cl::desc("Enable the use of the block frequency analysis to reduce the " "chance to execute const materialization more frequently than " "without hoisting."))
INITIALIZE_PASS_END(RegBankSelect, DEBUG_TYPE, "Assign register bank of generic virtual registers", false, false) RegBankSelect
Keeps track of the user of a constant and the operand index where the constant is used...
size_type size() const
Definition: SmallPtrSet.h:93
static const DILocation * getMergedLocation(const DILocation *LocA, const DILocation *LocB, const Instruction *ForInst=nullptr)
When two instructions are combined into a single instruction we also need to combine the original loc...
Analysis pass which computes BlockFrequencyInfo.
Iterator for intrusive lists based on ilist_node.
unsigned getNumOperands() const
Definition: User.h:176
SmallPtrSet - This class implements a set which is optimized for holding SmallSize or less elements...
Definition: SmallPtrSet.h:418
This is the shared class of boolean and integer constants.
Definition: Constants.h:84
This pass provides access to the codegen interfaces that are needed for IR-level transformations.
bool erase(PtrType Ptr)
erase - If the set contains the specified pointer, remove it and return true, otherwise return false...
Definition: SmallPtrSet.h:378
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:864
Instruction * user_back()
Specialize the methods defined in Value, as we know that an instruction can only be used by other ins...
Definition: Instruction.h:63
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
void setPreservesCFG()
This function should be called by the pass, iff they do not:
Definition: Pass.cpp:285
static Optional< APInt > calculateOffsetDiff(const APInt &V1, const APInt &V2)
BlockFrequency getBlockFreq(const BasicBlock *BB) const
getblockFreq - Return block frequency.
void setOperand(unsigned i, Value *Val)
Definition: User.h:159
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:132
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)
Class for arbitrary precision integers.
Definition: APInt.h:69
static BinaryOperator * Create(BinaryOps Op, Value *S1, Value *S2, const Twine &Name=Twine(), Instruction *InsertBefore=nullptr)
Construct a binary instruction, given the opcode and the two operands.
Represents analyses that only rely on functions&#39; control flow.
Definition: PassManager.h:114
const DebugLoc & getDebugLoc() const
Return the debug location for this node as a DebugLoc.
Definition: Instruction.h:284
iterator begin() const
Definition: SmallPtrSet.h:397
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:61
uint64_t getLimitedValue(uint64_t Limit=UINT64_MAX) const
If this value is smaller than the specified limit, return it, otherwise return the limit value...
Definition: APInt.h:475
void preserveSet()
Mark an analysis set as preserved.
Definition: PassManager.h:189
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:220
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
iterator end() const
Definition: SmallPtrSet.h:402
bool runImpl(Function &F, TargetTransformInfo &TTI, DominatorTree &DT, BlockFrequencyInfo *BFI, BasicBlock &Entry)
Optimize expensive integer constants in the given function.
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
The cost of a typical &#39;add&#39; instruction.
bool isEHPad() const
Return true if this basic block is an exception handling block.
Definition: BasicBlock.h:383
LLVM Value Representation.
Definition: Value.h:73
#define DEBUG(X)
Definition: Debug.h:118
bool isEHPad() const
Return true if the instruction is a variety of EH-block.
Definition: Instruction.h:538
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:49
A container for analyses that lazily runs them and caches their results.
RebasedConstantListType RebasedConstants
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:267
void sort(Policy policy, RandomAccessIterator Start, RandomAccessIterator End, const Comparator &Comp=Comparator())
Definition: Parallel.h:199
This pass exposes codegen information to IR-level passes.
bool use_empty() const
Definition: Value.h:328
const BasicBlock * getParent() const
Definition: Instruction.h:66
FunctionPass * createConstantHoistingPass()