xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/ConstantHoisting.cpp (revision 6966ac055c3b7a39266fb982493330df7a097997)
1 //===- ConstantHoisting.cpp - Prepare code for expensive constants --------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This pass identifies expensive constants to hoist and coalesces them to
10 // better prepare it for SelectionDAG-based code generation. This works around
11 // the limitations of the basic-block-at-a-time approach.
12 //
13 // First it scans all instructions for integer constants and calculates its
14 // cost. If the constant can be folded into the instruction (the cost is
15 // TCC_Free) or the cost is just a simple operation (TCC_BASIC), then we don't
16 // consider it expensive and leave it alone. This is the default behavior and
17 // the default implementation of getIntImmCost will always return TCC_Free.
18 //
19 // If the cost is more than TCC_BASIC, then the integer constant can't be folded
20 // into the instruction and it might be beneficial to hoist the constant.
21 // Similar constants are coalesced to reduce register pressure and
22 // materialization code.
23 //
24 // When a constant is hoisted, it is also hidden behind a bitcast to force it to
25 // be live-out of the basic block. Otherwise the constant would be just
26 // duplicated and each basic block would have its own copy in the SelectionDAG.
27 // The SelectionDAG recognizes such constants as opaque and doesn't perform
28 // certain transformations on them, which would create a new expensive constant.
29 //
30 // This optimization is only applied to integer constants in instructions and
31 // simple (this means not nested) constant cast expressions. For example:
32 // %0 = load i64* inttoptr (i64 big_constant to i64*)
33 //===----------------------------------------------------------------------===//
34 
35 #include "llvm/Transforms/Scalar/ConstantHoisting.h"
36 #include "llvm/ADT/APInt.h"
37 #include "llvm/ADT/DenseMap.h"
38 #include "llvm/ADT/None.h"
39 #include "llvm/ADT/Optional.h"
40 #include "llvm/ADT/SmallPtrSet.h"
41 #include "llvm/ADT/SmallVector.h"
42 #include "llvm/ADT/Statistic.h"
43 #include "llvm/Analysis/BlockFrequencyInfo.h"
44 #include "llvm/Analysis/ProfileSummaryInfo.h"
45 #include "llvm/Analysis/TargetTransformInfo.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/IR/BasicBlock.h"
48 #include "llvm/IR/Constants.h"
49 #include "llvm/IR/DebugInfoMetadata.h"
50 #include "llvm/IR/Dominators.h"
51 #include "llvm/IR/Function.h"
52 #include "llvm/IR/InstrTypes.h"
53 #include "llvm/IR/Instruction.h"
54 #include "llvm/IR/Instructions.h"
55 #include "llvm/IR/IntrinsicInst.h"
56 #include "llvm/IR/Value.h"
57 #include "llvm/Pass.h"
58 #include "llvm/Support/BlockFrequency.h"
59 #include "llvm/Support/Casting.h"
60 #include "llvm/Support/CommandLine.h"
61 #include "llvm/Support/Debug.h"
62 #include "llvm/Support/raw_ostream.h"
63 #include "llvm/Transforms/Scalar.h"
64 #include "llvm/Transforms/Utils/SizeOpts.h"
65 #include <algorithm>
66 #include <cassert>
67 #include <cstdint>
68 #include <iterator>
69 #include <tuple>
70 #include <utility>
71 
72 using namespace llvm;
73 using namespace consthoist;
74 
75 #define DEBUG_TYPE "consthoist"
76 
77 STATISTIC(NumConstantsHoisted, "Number of constants hoisted");
78 STATISTIC(NumConstantsRebased, "Number of constants rebased");
79 
80 static cl::opt<bool> ConstHoistWithBlockFrequency(
81     "consthoist-with-block-frequency", cl::init(true), cl::Hidden,
82     cl::desc("Enable the use of the block frequency analysis to reduce the "
83              "chance to execute const materialization more frequently than "
84              "without hoisting."));
85 
86 static cl::opt<bool> ConstHoistGEP(
87     "consthoist-gep", cl::init(false), cl::Hidden,
88     cl::desc("Try hoisting constant gep expressions"));
89 
90 static cl::opt<unsigned>
91 MinNumOfDependentToRebase("consthoist-min-num-to-rebase",
92     cl::desc("Do not rebase if number of dependent constants of a Base is less "
93              "than this number."),
94     cl::init(0), cl::Hidden);
95 
96 namespace {
97 
98 /// The constant hoisting pass.
99 class ConstantHoistingLegacyPass : public FunctionPass {
100 public:
101   static char ID; // Pass identification, replacement for typeid
102 
103   ConstantHoistingLegacyPass() : FunctionPass(ID) {
104     initializeConstantHoistingLegacyPassPass(*PassRegistry::getPassRegistry());
105   }
106 
107   bool runOnFunction(Function &Fn) override;
108 
109   StringRef getPassName() const override { return "Constant Hoisting"; }
110 
111   void getAnalysisUsage(AnalysisUsage &AU) const override {
112     AU.setPreservesCFG();
113     if (ConstHoistWithBlockFrequency)
114       AU.addRequired<BlockFrequencyInfoWrapperPass>();
115     AU.addRequired<DominatorTreeWrapperPass>();
116     AU.addRequired<ProfileSummaryInfoWrapperPass>();
117     AU.addRequired<TargetTransformInfoWrapperPass>();
118   }
119 
120 private:
121   ConstantHoistingPass Impl;
122 };
123 
124 } // end anonymous namespace
125 
126 char ConstantHoistingLegacyPass::ID = 0;
127 
128 INITIALIZE_PASS_BEGIN(ConstantHoistingLegacyPass, "consthoist",
129                       "Constant Hoisting", false, false)
130 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
131 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
132 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
133 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
134 INITIALIZE_PASS_END(ConstantHoistingLegacyPass, "consthoist",
135                     "Constant Hoisting", false, false)
136 
137 FunctionPass *llvm::createConstantHoistingPass() {
138   return new ConstantHoistingLegacyPass();
139 }
140 
141 /// Perform the constant hoisting optimization for the given function.
142 bool ConstantHoistingLegacyPass::runOnFunction(Function &Fn) {
143   if (skipFunction(Fn))
144     return false;
145 
146   LLVM_DEBUG(dbgs() << "********** Begin Constant Hoisting **********\n");
147   LLVM_DEBUG(dbgs() << "********** Function: " << Fn.getName() << '\n');
148 
149   bool MadeChange =
150       Impl.runImpl(Fn, getAnalysis<TargetTransformInfoWrapperPass>().getTTI(Fn),
151                    getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
152                    ConstHoistWithBlockFrequency
153                        ? &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI()
154                        : nullptr,
155                    Fn.getEntryBlock(),
156                    &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI());
157 
158   if (MadeChange) {
159     LLVM_DEBUG(dbgs() << "********** Function after Constant Hoisting: "
160                       << Fn.getName() << '\n');
161     LLVM_DEBUG(dbgs() << Fn);
162   }
163   LLVM_DEBUG(dbgs() << "********** End Constant Hoisting **********\n");
164 
165   return MadeChange;
166 }
167 
168 /// Find the constant materialization insertion point.
169 Instruction *ConstantHoistingPass::findMatInsertPt(Instruction *Inst,
170                                                    unsigned Idx) const {
171   // If the operand is a cast instruction, then we have to materialize the
172   // constant before the cast instruction.
173   if (Idx != ~0U) {
174     Value *Opnd = Inst->getOperand(Idx);
175     if (auto CastInst = dyn_cast<Instruction>(Opnd))
176       if (CastInst->isCast())
177         return CastInst;
178   }
179 
180   // The simple and common case. This also includes constant expressions.
181   if (!isa<PHINode>(Inst) && !Inst->isEHPad())
182     return Inst;
183 
184   // We can't insert directly before a phi node or an eh pad. Insert before
185   // the terminator of the incoming or dominating block.
186   assert(Entry != Inst->getParent() && "PHI or landing pad in entry block!");
187   if (Idx != ~0U && isa<PHINode>(Inst))
188     return cast<PHINode>(Inst)->getIncomingBlock(Idx)->getTerminator();
189 
190   // This must be an EH pad. Iterate over immediate dominators until we find a
191   // non-EH pad. We need to skip over catchswitch blocks, which are both EH pads
192   // and terminators.
193   auto IDom = DT->getNode(Inst->getParent())->getIDom();
194   while (IDom->getBlock()->isEHPad()) {
195     assert(Entry != IDom->getBlock() && "eh pad in entry block");
196     IDom = IDom->getIDom();
197   }
198 
199   return IDom->getBlock()->getTerminator();
200 }
201 
202 /// Given \p BBs as input, find another set of BBs which collectively
203 /// dominates \p BBs and have the minimal sum of frequencies. Return the BB
204 /// set found in \p BBs.
205 static void findBestInsertionSet(DominatorTree &DT, BlockFrequencyInfo &BFI,
206                                  BasicBlock *Entry,
207                                  SmallPtrSet<BasicBlock *, 8> &BBs) {
208   assert(!BBs.count(Entry) && "Assume Entry is not in BBs");
209   // Nodes on the current path to the root.
210   SmallPtrSet<BasicBlock *, 8> Path;
211   // Candidates includes any block 'BB' in set 'BBs' that is not strictly
212   // dominated by any other blocks in set 'BBs', and all nodes in the path
213   // in the dominator tree from Entry to 'BB'.
214   SmallPtrSet<BasicBlock *, 16> Candidates;
215   for (auto BB : BBs) {
216     // Ignore unreachable basic blocks.
217     if (!DT.isReachableFromEntry(BB))
218       continue;
219     Path.clear();
220     // Walk up the dominator tree until Entry or another BB in BBs
221     // is reached. Insert the nodes on the way to the Path.
222     BasicBlock *Node = BB;
223     // The "Path" is a candidate path to be added into Candidates set.
224     bool isCandidate = false;
225     do {
226       Path.insert(Node);
227       if (Node == Entry || Candidates.count(Node)) {
228         isCandidate = true;
229         break;
230       }
231       assert(DT.getNode(Node)->getIDom() &&
232              "Entry doens't dominate current Node");
233       Node = DT.getNode(Node)->getIDom()->getBlock();
234     } while (!BBs.count(Node));
235 
236     // If isCandidate is false, Node is another Block in BBs dominating
237     // current 'BB'. Drop the nodes on the Path.
238     if (!isCandidate)
239       continue;
240 
241     // Add nodes on the Path into Candidates.
242     Candidates.insert(Path.begin(), Path.end());
243   }
244 
245   // Sort the nodes in Candidates in top-down order and save the nodes
246   // in Orders.
247   unsigned Idx = 0;
248   SmallVector<BasicBlock *, 16> Orders;
249   Orders.push_back(Entry);
250   while (Idx != Orders.size()) {
251     BasicBlock *Node = Orders[Idx++];
252     for (auto ChildDomNode : DT.getNode(Node)->getChildren()) {
253       if (Candidates.count(ChildDomNode->getBlock()))
254         Orders.push_back(ChildDomNode->getBlock());
255     }
256   }
257 
258   // Visit Orders in bottom-up order.
259   using InsertPtsCostPair =
260       std::pair<SmallPtrSet<BasicBlock *, 16>, BlockFrequency>;
261 
262   // InsertPtsMap is a map from a BB to the best insertion points for the
263   // subtree of BB (subtree not including the BB itself).
264   DenseMap<BasicBlock *, InsertPtsCostPair> InsertPtsMap;
265   InsertPtsMap.reserve(Orders.size() + 1);
266   for (auto RIt = Orders.rbegin(); RIt != Orders.rend(); RIt++) {
267     BasicBlock *Node = *RIt;
268     bool NodeInBBs = BBs.count(Node);
269     SmallPtrSet<BasicBlock *, 16> &InsertPts = InsertPtsMap[Node].first;
270     BlockFrequency &InsertPtsFreq = InsertPtsMap[Node].second;
271 
272     // Return the optimal insert points in BBs.
273     if (Node == Entry) {
274       BBs.clear();
275       if (InsertPtsFreq > BFI.getBlockFreq(Node) ||
276           (InsertPtsFreq == BFI.getBlockFreq(Node) && InsertPts.size() > 1))
277         BBs.insert(Entry);
278       else
279         BBs.insert(InsertPts.begin(), InsertPts.end());
280       break;
281     }
282 
283     BasicBlock *Parent = DT.getNode(Node)->getIDom()->getBlock();
284     // Initially, ParentInsertPts is empty and ParentPtsFreq is 0. Every child
285     // will update its parent's ParentInsertPts and ParentPtsFreq.
286     SmallPtrSet<BasicBlock *, 16> &ParentInsertPts = InsertPtsMap[Parent].first;
287     BlockFrequency &ParentPtsFreq = InsertPtsMap[Parent].second;
288     // Choose to insert in Node or in subtree of Node.
289     // Don't hoist to EHPad because we may not find a proper place to insert
290     // in EHPad.
291     // If the total frequency of InsertPts is the same as the frequency of the
292     // target Node, and InsertPts contains more than one nodes, choose hoisting
293     // to reduce code size.
294     if (NodeInBBs ||
295         (!Node->isEHPad() &&
296          (InsertPtsFreq > BFI.getBlockFreq(Node) ||
297           (InsertPtsFreq == BFI.getBlockFreq(Node) && InsertPts.size() > 1)))) {
298       ParentInsertPts.insert(Node);
299       ParentPtsFreq += BFI.getBlockFreq(Node);
300     } else {
301       ParentInsertPts.insert(InsertPts.begin(), InsertPts.end());
302       ParentPtsFreq += InsertPtsFreq;
303     }
304   }
305 }
306 
307 /// Find an insertion point that dominates all uses.
308 SmallPtrSet<Instruction *, 8> ConstantHoistingPass::findConstantInsertionPoint(
309     const ConstantInfo &ConstInfo) const {
310   assert(!ConstInfo.RebasedConstants.empty() && "Invalid constant info entry.");
311   // Collect all basic blocks.
312   SmallPtrSet<BasicBlock *, 8> BBs;
313   SmallPtrSet<Instruction *, 8> InsertPts;
314   for (auto const &RCI : ConstInfo.RebasedConstants)
315     for (auto const &U : RCI.Uses)
316       BBs.insert(findMatInsertPt(U.Inst, U.OpndIdx)->getParent());
317 
318   if (BBs.count(Entry)) {
319     InsertPts.insert(&Entry->front());
320     return InsertPts;
321   }
322 
323   if (BFI) {
324     findBestInsertionSet(*DT, *BFI, Entry, BBs);
325     for (auto BB : BBs) {
326       BasicBlock::iterator InsertPt = BB->begin();
327       for (; isa<PHINode>(InsertPt) || InsertPt->isEHPad(); ++InsertPt)
328         ;
329       InsertPts.insert(&*InsertPt);
330     }
331     return InsertPts;
332   }
333 
334   while (BBs.size() >= 2) {
335     BasicBlock *BB, *BB1, *BB2;
336     BB1 = *BBs.begin();
337     BB2 = *std::next(BBs.begin());
338     BB = DT->findNearestCommonDominator(BB1, BB2);
339     if (BB == Entry) {
340       InsertPts.insert(&Entry->front());
341       return InsertPts;
342     }
343     BBs.erase(BB1);
344     BBs.erase(BB2);
345     BBs.insert(BB);
346   }
347   assert((BBs.size() == 1) && "Expected only one element.");
348   Instruction &FirstInst = (*BBs.begin())->front();
349   InsertPts.insert(findMatInsertPt(&FirstInst));
350   return InsertPts;
351 }
352 
353 /// Record constant integer ConstInt for instruction Inst at operand
354 /// index Idx.
355 ///
356 /// The operand at index Idx is not necessarily the constant integer itself. It
357 /// could also be a cast instruction or a constant expression that uses the
358 /// constant integer.
359 void ConstantHoistingPass::collectConstantCandidates(
360     ConstCandMapType &ConstCandMap, Instruction *Inst, unsigned Idx,
361     ConstantInt *ConstInt) {
362   unsigned Cost;
363   // Ask the target about the cost of materializing the constant for the given
364   // instruction and operand index.
365   if (auto IntrInst = dyn_cast<IntrinsicInst>(Inst))
366     Cost = TTI->getIntImmCost(IntrInst->getIntrinsicID(), Idx,
367                               ConstInt->getValue(), ConstInt->getType());
368   else
369     Cost = TTI->getIntImmCost(Inst->getOpcode(), Idx, ConstInt->getValue(),
370                               ConstInt->getType());
371 
372   // Ignore cheap integer constants.
373   if (Cost > TargetTransformInfo::TCC_Basic) {
374     ConstCandMapType::iterator Itr;
375     bool Inserted;
376     ConstPtrUnionType Cand = ConstInt;
377     std::tie(Itr, Inserted) = ConstCandMap.insert(std::make_pair(Cand, 0));
378     if (Inserted) {
379       ConstIntCandVec.push_back(ConstantCandidate(ConstInt));
380       Itr->second = ConstIntCandVec.size() - 1;
381     }
382     ConstIntCandVec[Itr->second].addUser(Inst, Idx, Cost);
383     LLVM_DEBUG(if (isa<ConstantInt>(Inst->getOperand(Idx))) dbgs()
384                    << "Collect constant " << *ConstInt << " from " << *Inst
385                    << " with cost " << Cost << '\n';
386                else dbgs() << "Collect constant " << *ConstInt
387                            << " indirectly from " << *Inst << " via "
388                            << *Inst->getOperand(Idx) << " with cost " << Cost
389                            << '\n';);
390   }
391 }
392 
393 /// Record constant GEP expression for instruction Inst at operand index Idx.
394 void ConstantHoistingPass::collectConstantCandidates(
395     ConstCandMapType &ConstCandMap, Instruction *Inst, unsigned Idx,
396     ConstantExpr *ConstExpr) {
397   // TODO: Handle vector GEPs
398   if (ConstExpr->getType()->isVectorTy())
399     return;
400 
401   GlobalVariable *BaseGV = dyn_cast<GlobalVariable>(ConstExpr->getOperand(0));
402   if (!BaseGV)
403     return;
404 
405   // Get offset from the base GV.
406   PointerType *GVPtrTy = dyn_cast<PointerType>(BaseGV->getType());
407   IntegerType *PtrIntTy = DL->getIntPtrType(*Ctx, GVPtrTy->getAddressSpace());
408   APInt Offset(DL->getTypeSizeInBits(PtrIntTy), /*val*/0, /*isSigned*/true);
409   auto *GEPO = cast<GEPOperator>(ConstExpr);
410   if (!GEPO->accumulateConstantOffset(*DL, Offset))
411     return;
412 
413   if (!Offset.isIntN(32))
414     return;
415 
416   // A constant GEP expression that has a GlobalVariable as base pointer is
417   // usually lowered to a load from constant pool. Such operation is unlikely
418   // to be cheaper than compute it by <Base + Offset>, which can be lowered to
419   // an ADD instruction or folded into Load/Store instruction.
420   int Cost = TTI->getIntImmCost(Instruction::Add, 1, Offset, PtrIntTy);
421   ConstCandVecType &ExprCandVec = ConstGEPCandMap[BaseGV];
422   ConstCandMapType::iterator Itr;
423   bool Inserted;
424   ConstPtrUnionType Cand = ConstExpr;
425   std::tie(Itr, Inserted) = ConstCandMap.insert(std::make_pair(Cand, 0));
426   if (Inserted) {
427     ExprCandVec.push_back(ConstantCandidate(
428         ConstantInt::get(Type::getInt32Ty(*Ctx), Offset.getLimitedValue()),
429         ConstExpr));
430     Itr->second = ExprCandVec.size() - 1;
431   }
432   ExprCandVec[Itr->second].addUser(Inst, Idx, Cost);
433 }
434 
435 /// Check the operand for instruction Inst at index Idx.
436 void ConstantHoistingPass::collectConstantCandidates(
437     ConstCandMapType &ConstCandMap, Instruction *Inst, unsigned Idx) {
438   Value *Opnd = Inst->getOperand(Idx);
439 
440   // Visit constant integers.
441   if (auto ConstInt = dyn_cast<ConstantInt>(Opnd)) {
442     collectConstantCandidates(ConstCandMap, Inst, Idx, ConstInt);
443     return;
444   }
445 
446   // Visit cast instructions that have constant integers.
447   if (auto CastInst = dyn_cast<Instruction>(Opnd)) {
448     // Only visit cast instructions, which have been skipped. All other
449     // instructions should have already been visited.
450     if (!CastInst->isCast())
451       return;
452 
453     if (auto *ConstInt = dyn_cast<ConstantInt>(CastInst->getOperand(0))) {
454       // Pretend the constant is directly used by the instruction and ignore
455       // the cast instruction.
456       collectConstantCandidates(ConstCandMap, Inst, Idx, ConstInt);
457       return;
458     }
459   }
460 
461   // Visit constant expressions that have constant integers.
462   if (auto ConstExpr = dyn_cast<ConstantExpr>(Opnd)) {
463     // Handle constant gep expressions.
464     if (ConstHoistGEP && ConstExpr->isGEPWithNoNotionalOverIndexing())
465       collectConstantCandidates(ConstCandMap, Inst, Idx, ConstExpr);
466 
467     // Only visit constant cast expressions.
468     if (!ConstExpr->isCast())
469       return;
470 
471     if (auto ConstInt = dyn_cast<ConstantInt>(ConstExpr->getOperand(0))) {
472       // Pretend the constant is directly used by the instruction and ignore
473       // the constant expression.
474       collectConstantCandidates(ConstCandMap, Inst, Idx, ConstInt);
475       return;
476     }
477   }
478 }
479 
480 /// Scan the instruction for expensive integer constants and record them
481 /// in the constant candidate vector.
482 void ConstantHoistingPass::collectConstantCandidates(
483     ConstCandMapType &ConstCandMap, Instruction *Inst) {
484   // Skip all cast instructions. They are visited indirectly later on.
485   if (Inst->isCast())
486     return;
487 
488   // Scan all operands.
489   for (unsigned Idx = 0, E = Inst->getNumOperands(); Idx != E; ++Idx) {
490     // The cost of materializing the constants (defined in
491     // `TargetTransformInfo::getIntImmCost`) for instructions which only take
492     // constant variables is lower than `TargetTransformInfo::TCC_Basic`. So
493     // it's safe for us to collect constant candidates from all IntrinsicInsts.
494     if (canReplaceOperandWithVariable(Inst, Idx) || isa<IntrinsicInst>(Inst)) {
495       collectConstantCandidates(ConstCandMap, Inst, Idx);
496     }
497   } // end of for all operands
498 }
499 
500 /// Collect all integer constants in the function that cannot be folded
501 /// into an instruction itself.
502 void ConstantHoistingPass::collectConstantCandidates(Function &Fn) {
503   ConstCandMapType ConstCandMap;
504   for (BasicBlock &BB : Fn)
505     for (Instruction &Inst : BB)
506       collectConstantCandidates(ConstCandMap, &Inst);
507 }
508 
509 // This helper function is necessary to deal with values that have different
510 // bit widths (APInt Operator- does not like that). If the value cannot be
511 // represented in uint64 we return an "empty" APInt. This is then interpreted
512 // as the value is not in range.
513 static Optional<APInt> calculateOffsetDiff(const APInt &V1, const APInt &V2) {
514   Optional<APInt> Res = None;
515   unsigned BW = V1.getBitWidth() > V2.getBitWidth() ?
516                 V1.getBitWidth() : V2.getBitWidth();
517   uint64_t LimVal1 = V1.getLimitedValue();
518   uint64_t LimVal2 = V2.getLimitedValue();
519 
520   if (LimVal1 == ~0ULL || LimVal2 == ~0ULL)
521     return Res;
522 
523   uint64_t Diff = LimVal1 - LimVal2;
524   return APInt(BW, Diff, true);
525 }
526 
527 // From a list of constants, one needs to picked as the base and the other
528 // constants will be transformed into an offset from that base constant. The
529 // question is which we can pick best? For example, consider these constants
530 // and their number of uses:
531 //
532 //  Constants| 2 | 4 | 12 | 42 |
533 //  NumUses  | 3 | 2 |  8 |  7 |
534 //
535 // Selecting constant 12 because it has the most uses will generate negative
536 // offsets for constants 2 and 4 (i.e. -10 and -8 respectively). If negative
537 // offsets lead to less optimal code generation, then there might be better
538 // solutions. Suppose immediates in the range of 0..35 are most optimally
539 // supported by the architecture, then selecting constant 2 is most optimal
540 // because this will generate offsets: 0, 2, 10, 40. Offsets 0, 2 and 10 are in
541 // range 0..35, and thus 3 + 2 + 8 = 13 uses are in range. Selecting 12 would
542 // have only 8 uses in range, so choosing 2 as a base is more optimal. Thus, in
543 // selecting the base constant the range of the offsets is a very important
544 // factor too that we take into account here. This algorithm calculates a total
545 // costs for selecting a constant as the base and substract the costs if
546 // immediates are out of range. It has quadratic complexity, so we call this
547 // function only when we're optimising for size and there are less than 100
548 // constants, we fall back to the straightforward algorithm otherwise
549 // which does not do all the offset calculations.
550 unsigned
551 ConstantHoistingPass::maximizeConstantsInRange(ConstCandVecType::iterator S,
552                                            ConstCandVecType::iterator E,
553                                            ConstCandVecType::iterator &MaxCostItr) {
554   unsigned NumUses = 0;
555 
556   bool OptForSize = Entry->getParent()->hasOptSize() ||
557                     llvm::shouldOptimizeForSize(Entry->getParent(), PSI, BFI);
558   if (!OptForSize || std::distance(S,E) > 100) {
559     for (auto ConstCand = S; ConstCand != E; ++ConstCand) {
560       NumUses += ConstCand->Uses.size();
561       if (ConstCand->CumulativeCost > MaxCostItr->CumulativeCost)
562         MaxCostItr = ConstCand;
563     }
564     return NumUses;
565   }
566 
567   LLVM_DEBUG(dbgs() << "== Maximize constants in range ==\n");
568   int MaxCost = -1;
569   for (auto ConstCand = S; ConstCand != E; ++ConstCand) {
570     auto Value = ConstCand->ConstInt->getValue();
571     Type *Ty = ConstCand->ConstInt->getType();
572     int Cost = 0;
573     NumUses += ConstCand->Uses.size();
574     LLVM_DEBUG(dbgs() << "= Constant: " << ConstCand->ConstInt->getValue()
575                       << "\n");
576 
577     for (auto User : ConstCand->Uses) {
578       unsigned Opcode = User.Inst->getOpcode();
579       unsigned OpndIdx = User.OpndIdx;
580       Cost += TTI->getIntImmCost(Opcode, OpndIdx, Value, Ty);
581       LLVM_DEBUG(dbgs() << "Cost: " << Cost << "\n");
582 
583       for (auto C2 = S; C2 != E; ++C2) {
584         Optional<APInt> Diff = calculateOffsetDiff(
585                                    C2->ConstInt->getValue(),
586                                    ConstCand->ConstInt->getValue());
587         if (Diff) {
588           const int ImmCosts =
589             TTI->getIntImmCodeSizeCost(Opcode, OpndIdx, Diff.getValue(), Ty);
590           Cost -= ImmCosts;
591           LLVM_DEBUG(dbgs() << "Offset " << Diff.getValue() << " "
592                             << "has penalty: " << ImmCosts << "\n"
593                             << "Adjusted cost: " << Cost << "\n");
594         }
595       }
596     }
597     LLVM_DEBUG(dbgs() << "Cumulative cost: " << Cost << "\n");
598     if (Cost > MaxCost) {
599       MaxCost = Cost;
600       MaxCostItr = ConstCand;
601       LLVM_DEBUG(dbgs() << "New candidate: " << MaxCostItr->ConstInt->getValue()
602                         << "\n");
603     }
604   }
605   return NumUses;
606 }
607 
608 /// Find the base constant within the given range and rebase all other
609 /// constants with respect to the base constant.
610 void ConstantHoistingPass::findAndMakeBaseConstant(
611     ConstCandVecType::iterator S, ConstCandVecType::iterator E,
612     SmallVectorImpl<consthoist::ConstantInfo> &ConstInfoVec) {
613   auto MaxCostItr = S;
614   unsigned NumUses = maximizeConstantsInRange(S, E, MaxCostItr);
615 
616   // Don't hoist constants that have only one use.
617   if (NumUses <= 1)
618     return;
619 
620   ConstantInt *ConstInt = MaxCostItr->ConstInt;
621   ConstantExpr *ConstExpr = MaxCostItr->ConstExpr;
622   ConstantInfo ConstInfo;
623   ConstInfo.BaseInt = ConstInt;
624   ConstInfo.BaseExpr = ConstExpr;
625   Type *Ty = ConstInt->getType();
626 
627   // Rebase the constants with respect to the base constant.
628   for (auto ConstCand = S; ConstCand != E; ++ConstCand) {
629     APInt Diff = ConstCand->ConstInt->getValue() - ConstInt->getValue();
630     Constant *Offset = Diff == 0 ? nullptr : ConstantInt::get(Ty, Diff);
631     Type *ConstTy =
632         ConstCand->ConstExpr ? ConstCand->ConstExpr->getType() : nullptr;
633     ConstInfo.RebasedConstants.push_back(
634       RebasedConstantInfo(std::move(ConstCand->Uses), Offset, ConstTy));
635   }
636   ConstInfoVec.push_back(std::move(ConstInfo));
637 }
638 
639 /// Finds and combines constant candidates that can be easily
640 /// rematerialized with an add from a common base constant.
641 void ConstantHoistingPass::findBaseConstants(GlobalVariable *BaseGV) {
642   // If BaseGV is nullptr, find base among candidate constant integers;
643   // Otherwise find base among constant GEPs that share the same BaseGV.
644   ConstCandVecType &ConstCandVec = BaseGV ?
645       ConstGEPCandMap[BaseGV] : ConstIntCandVec;
646   ConstInfoVecType &ConstInfoVec = BaseGV ?
647       ConstGEPInfoMap[BaseGV] : ConstIntInfoVec;
648 
649   // Sort the constants by value and type. This invalidates the mapping!
650   llvm::stable_sort(ConstCandVec, [](const ConstantCandidate &LHS,
651                                      const ConstantCandidate &RHS) {
652     if (LHS.ConstInt->getType() != RHS.ConstInt->getType())
653       return LHS.ConstInt->getType()->getBitWidth() <
654              RHS.ConstInt->getType()->getBitWidth();
655     return LHS.ConstInt->getValue().ult(RHS.ConstInt->getValue());
656   });
657 
658   // Simple linear scan through the sorted constant candidate vector for viable
659   // merge candidates.
660   auto MinValItr = ConstCandVec.begin();
661   for (auto CC = std::next(ConstCandVec.begin()), E = ConstCandVec.end();
662        CC != E; ++CC) {
663     if (MinValItr->ConstInt->getType() == CC->ConstInt->getType()) {
664       Type *MemUseValTy = nullptr;
665       for (auto &U : CC->Uses) {
666         auto *UI = U.Inst;
667         if (LoadInst *LI = dyn_cast<LoadInst>(UI)) {
668           MemUseValTy = LI->getType();
669           break;
670         } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
671           // Make sure the constant is used as pointer operand of the StoreInst.
672           if (SI->getPointerOperand() == SI->getOperand(U.OpndIdx)) {
673             MemUseValTy = SI->getValueOperand()->getType();
674             break;
675           }
676         }
677       }
678 
679       // Check if the constant is in range of an add with immediate.
680       APInt Diff = CC->ConstInt->getValue() - MinValItr->ConstInt->getValue();
681       if ((Diff.getBitWidth() <= 64) &&
682           TTI->isLegalAddImmediate(Diff.getSExtValue()) &&
683           // Check if Diff can be used as offset in addressing mode of the user
684           // memory instruction.
685           (!MemUseValTy || TTI->isLegalAddressingMode(MemUseValTy,
686            /*BaseGV*/nullptr, /*BaseOffset*/Diff.getSExtValue(),
687            /*HasBaseReg*/true, /*Scale*/0)))
688         continue;
689     }
690     // We either have now a different constant type or the constant is not in
691     // range of an add with immediate anymore.
692     findAndMakeBaseConstant(MinValItr, CC, ConstInfoVec);
693     // Start a new base constant search.
694     MinValItr = CC;
695   }
696   // Finalize the last base constant search.
697   findAndMakeBaseConstant(MinValItr, ConstCandVec.end(), ConstInfoVec);
698 }
699 
700 /// Updates the operand at Idx in instruction Inst with the result of
701 ///        instruction Mat. If the instruction is a PHI node then special
702 ///        handling for duplicate values form the same incoming basic block is
703 ///        required.
704 /// \return The update will always succeed, but the return value indicated if
705 ///         Mat was used for the update or not.
706 static bool updateOperand(Instruction *Inst, unsigned Idx, Instruction *Mat) {
707   if (auto PHI = dyn_cast<PHINode>(Inst)) {
708     // Check if any previous operand of the PHI node has the same incoming basic
709     // block. This is a very odd case that happens when the incoming basic block
710     // has a switch statement. In this case use the same value as the previous
711     // operand(s), otherwise we will fail verification due to different values.
712     // The values are actually the same, but the variable names are different
713     // and the verifier doesn't like that.
714     BasicBlock *IncomingBB = PHI->getIncomingBlock(Idx);
715     for (unsigned i = 0; i < Idx; ++i) {
716       if (PHI->getIncomingBlock(i) == IncomingBB) {
717         Value *IncomingVal = PHI->getIncomingValue(i);
718         Inst->setOperand(Idx, IncomingVal);
719         return false;
720       }
721     }
722   }
723 
724   Inst->setOperand(Idx, Mat);
725   return true;
726 }
727 
728 /// Emit materialization code for all rebased constants and update their
729 /// users.
730 void ConstantHoistingPass::emitBaseConstants(Instruction *Base,
731                                              Constant *Offset,
732                                              Type *Ty,
733                                              const ConstantUser &ConstUser) {
734   Instruction *Mat = Base;
735 
736   // The same offset can be dereferenced to different types in nested struct.
737   if (!Offset && Ty && Ty != Base->getType())
738     Offset = ConstantInt::get(Type::getInt32Ty(*Ctx), 0);
739 
740   if (Offset) {
741     Instruction *InsertionPt = findMatInsertPt(ConstUser.Inst,
742                                                ConstUser.OpndIdx);
743     if (Ty) {
744       // Constant being rebased is a ConstantExpr.
745       PointerType *Int8PtrTy = Type::getInt8PtrTy(*Ctx,
746           cast<PointerType>(Ty)->getAddressSpace());
747       Base = new BitCastInst(Base, Int8PtrTy, "base_bitcast", InsertionPt);
748       Mat = GetElementPtrInst::Create(Int8PtrTy->getElementType(), Base,
749           Offset, "mat_gep", InsertionPt);
750       Mat = new BitCastInst(Mat, Ty, "mat_bitcast", InsertionPt);
751     } else
752       // Constant being rebased is a ConstantInt.
753       Mat = BinaryOperator::Create(Instruction::Add, Base, Offset,
754                                  "const_mat", InsertionPt);
755 
756     LLVM_DEBUG(dbgs() << "Materialize constant (" << *Base->getOperand(0)
757                       << " + " << *Offset << ") in BB "
758                       << Mat->getParent()->getName() << '\n'
759                       << *Mat << '\n');
760     Mat->setDebugLoc(ConstUser.Inst->getDebugLoc());
761   }
762   Value *Opnd = ConstUser.Inst->getOperand(ConstUser.OpndIdx);
763 
764   // Visit constant integer.
765   if (isa<ConstantInt>(Opnd)) {
766     LLVM_DEBUG(dbgs() << "Update: " << *ConstUser.Inst << '\n');
767     if (!updateOperand(ConstUser.Inst, ConstUser.OpndIdx, Mat) && Offset)
768       Mat->eraseFromParent();
769     LLVM_DEBUG(dbgs() << "To    : " << *ConstUser.Inst << '\n');
770     return;
771   }
772 
773   // Visit cast instruction.
774   if (auto CastInst = dyn_cast<Instruction>(Opnd)) {
775     assert(CastInst->isCast() && "Expected an cast instruction!");
776     // Check if we already have visited this cast instruction before to avoid
777     // unnecessary cloning.
778     Instruction *&ClonedCastInst = ClonedCastMap[CastInst];
779     if (!ClonedCastInst) {
780       ClonedCastInst = CastInst->clone();
781       ClonedCastInst->setOperand(0, Mat);
782       ClonedCastInst->insertAfter(CastInst);
783       // Use the same debug location as the original cast instruction.
784       ClonedCastInst->setDebugLoc(CastInst->getDebugLoc());
785       LLVM_DEBUG(dbgs() << "Clone instruction: " << *CastInst << '\n'
786                         << "To               : " << *ClonedCastInst << '\n');
787     }
788 
789     LLVM_DEBUG(dbgs() << "Update: " << *ConstUser.Inst << '\n');
790     updateOperand(ConstUser.Inst, ConstUser.OpndIdx, ClonedCastInst);
791     LLVM_DEBUG(dbgs() << "To    : " << *ConstUser.Inst << '\n');
792     return;
793   }
794 
795   // Visit constant expression.
796   if (auto ConstExpr = dyn_cast<ConstantExpr>(Opnd)) {
797     if (ConstExpr->isGEPWithNoNotionalOverIndexing()) {
798       // Operand is a ConstantGEP, replace it.
799       updateOperand(ConstUser.Inst, ConstUser.OpndIdx, Mat);
800       return;
801     }
802 
803     // Aside from constant GEPs, only constant cast expressions are collected.
804     assert(ConstExpr->isCast() && "ConstExpr should be a cast");
805     Instruction *ConstExprInst = ConstExpr->getAsInstruction();
806     ConstExprInst->setOperand(0, Mat);
807     ConstExprInst->insertBefore(findMatInsertPt(ConstUser.Inst,
808                                                 ConstUser.OpndIdx));
809 
810     // Use the same debug location as the instruction we are about to update.
811     ConstExprInst->setDebugLoc(ConstUser.Inst->getDebugLoc());
812 
813     LLVM_DEBUG(dbgs() << "Create instruction: " << *ConstExprInst << '\n'
814                       << "From              : " << *ConstExpr << '\n');
815     LLVM_DEBUG(dbgs() << "Update: " << *ConstUser.Inst << '\n');
816     if (!updateOperand(ConstUser.Inst, ConstUser.OpndIdx, ConstExprInst)) {
817       ConstExprInst->eraseFromParent();
818       if (Offset)
819         Mat->eraseFromParent();
820     }
821     LLVM_DEBUG(dbgs() << "To    : " << *ConstUser.Inst << '\n');
822     return;
823   }
824 }
825 
826 /// Hoist and hide the base constant behind a bitcast and emit
827 /// materialization code for derived constants.
828 bool ConstantHoistingPass::emitBaseConstants(GlobalVariable *BaseGV) {
829   bool MadeChange = false;
830   SmallVectorImpl<consthoist::ConstantInfo> &ConstInfoVec =
831       BaseGV ? ConstGEPInfoMap[BaseGV] : ConstIntInfoVec;
832   for (auto const &ConstInfo : ConstInfoVec) {
833     SmallPtrSet<Instruction *, 8> IPSet = findConstantInsertionPoint(ConstInfo);
834     // We can have an empty set if the function contains unreachable blocks.
835     if (IPSet.empty())
836       continue;
837 
838     unsigned UsesNum = 0;
839     unsigned ReBasesNum = 0;
840     unsigned NotRebasedNum = 0;
841     for (Instruction *IP : IPSet) {
842       // First, collect constants depending on this IP of the base.
843       unsigned Uses = 0;
844       using RebasedUse = std::tuple<Constant *, Type *, ConstantUser>;
845       SmallVector<RebasedUse, 4> ToBeRebased;
846       for (auto const &RCI : ConstInfo.RebasedConstants) {
847         for (auto const &U : RCI.Uses) {
848           Uses++;
849           BasicBlock *OrigMatInsertBB =
850               findMatInsertPt(U.Inst, U.OpndIdx)->getParent();
851           // If Base constant is to be inserted in multiple places,
852           // generate rebase for U using the Base dominating U.
853           if (IPSet.size() == 1 ||
854               DT->dominates(IP->getParent(), OrigMatInsertBB))
855             ToBeRebased.push_back(RebasedUse(RCI.Offset, RCI.Ty, U));
856         }
857       }
858       UsesNum = Uses;
859 
860       // If only few constants depend on this IP of base, skip rebasing,
861       // assuming the base and the rebased have the same materialization cost.
862       if (ToBeRebased.size() < MinNumOfDependentToRebase) {
863         NotRebasedNum += ToBeRebased.size();
864         continue;
865       }
866 
867       // Emit an instance of the base at this IP.
868       Instruction *Base = nullptr;
869       // Hoist and hide the base constant behind a bitcast.
870       if (ConstInfo.BaseExpr) {
871         assert(BaseGV && "A base constant expression must have an base GV");
872         Type *Ty = ConstInfo.BaseExpr->getType();
873         Base = new BitCastInst(ConstInfo.BaseExpr, Ty, "const", IP);
874       } else {
875         IntegerType *Ty = ConstInfo.BaseInt->getType();
876         Base = new BitCastInst(ConstInfo.BaseInt, Ty, "const", IP);
877       }
878 
879       Base->setDebugLoc(IP->getDebugLoc());
880 
881       LLVM_DEBUG(dbgs() << "Hoist constant (" << *ConstInfo.BaseInt
882                         << ") to BB " << IP->getParent()->getName() << '\n'
883                         << *Base << '\n');
884 
885       // Emit materialization code for rebased constants depending on this IP.
886       for (auto const &R : ToBeRebased) {
887         Constant *Off = std::get<0>(R);
888         Type *Ty = std::get<1>(R);
889         ConstantUser U = std::get<2>(R);
890         emitBaseConstants(Base, Off, Ty, U);
891         ReBasesNum++;
892         // Use the same debug location as the last user of the constant.
893         Base->setDebugLoc(DILocation::getMergedLocation(
894             Base->getDebugLoc(), U.Inst->getDebugLoc()));
895       }
896       assert(!Base->use_empty() && "The use list is empty!?");
897       assert(isa<Instruction>(Base->user_back()) &&
898              "All uses should be instructions.");
899     }
900     (void)UsesNum;
901     (void)ReBasesNum;
902     (void)NotRebasedNum;
903     // Expect all uses are rebased after rebase is done.
904     assert(UsesNum == (ReBasesNum + NotRebasedNum) &&
905            "Not all uses are rebased");
906 
907     NumConstantsHoisted++;
908 
909     // Base constant is also included in ConstInfo.RebasedConstants, so
910     // deduct 1 from ConstInfo.RebasedConstants.size().
911     NumConstantsRebased += ConstInfo.RebasedConstants.size() - 1;
912 
913     MadeChange = true;
914   }
915   return MadeChange;
916 }
917 
918 /// Check all cast instructions we made a copy of and remove them if they
919 /// have no more users.
920 void ConstantHoistingPass::deleteDeadCastInst() const {
921   for (auto const &I : ClonedCastMap)
922     if (I.first->use_empty())
923       I.first->eraseFromParent();
924 }
925 
926 /// Optimize expensive integer constants in the given function.
927 bool ConstantHoistingPass::runImpl(Function &Fn, TargetTransformInfo &TTI,
928                                    DominatorTree &DT, BlockFrequencyInfo *BFI,
929                                    BasicBlock &Entry, ProfileSummaryInfo *PSI) {
930   this->TTI = &TTI;
931   this->DT = &DT;
932   this->BFI = BFI;
933   this->DL = &Fn.getParent()->getDataLayout();
934   this->Ctx = &Fn.getContext();
935   this->Entry = &Entry;
936   this->PSI = PSI;
937   // Collect all constant candidates.
938   collectConstantCandidates(Fn);
939 
940   // Combine constants that can be easily materialized with an add from a common
941   // base constant.
942   if (!ConstIntCandVec.empty())
943     findBaseConstants(nullptr);
944   for (auto &MapEntry : ConstGEPCandMap)
945     if (!MapEntry.second.empty())
946       findBaseConstants(MapEntry.first);
947 
948   // Finally hoist the base constant and emit materialization code for dependent
949   // constants.
950   bool MadeChange = false;
951   if (!ConstIntInfoVec.empty())
952     MadeChange = emitBaseConstants(nullptr);
953   for (auto MapEntry : ConstGEPInfoMap)
954     if (!MapEntry.second.empty())
955       MadeChange |= emitBaseConstants(MapEntry.first);
956 
957 
958   // Cleanup dead instructions.
959   deleteDeadCastInst();
960 
961   cleanup();
962 
963   return MadeChange;
964 }
965 
966 PreservedAnalyses ConstantHoistingPass::run(Function &F,
967                                             FunctionAnalysisManager &AM) {
968   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
969   auto &TTI = AM.getResult<TargetIRAnalysis>(F);
970   auto BFI = ConstHoistWithBlockFrequency
971                  ? &AM.getResult<BlockFrequencyAnalysis>(F)
972                  : nullptr;
973   auto &MAM = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F).getManager();
974   auto *PSI = MAM.getCachedResult<ProfileSummaryAnalysis>(*F.getParent());
975   if (!runImpl(F, TTI, DT, BFI, F.getEntryBlock(), PSI))
976     return PreservedAnalyses::all();
977 
978   PreservedAnalyses PA;
979   PA.preserveSet<CFGAnalyses>();
980   return PA;
981 }
982