1 //===- ConstantHoisting.cpp - Prepare code for expensive constants --------===//
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
7 //===----------------------------------------------------------------------===//
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.
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 getIntImmCostInst will always return TCC_Free.
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.
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.
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 //===----------------------------------------------------------------------===//
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/IR/BasicBlock.h"
47 #include "llvm/IR/Constants.h"
48 #include "llvm/IR/DebugInfoMetadata.h"
49 #include "llvm/IR/Dominators.h"
50 #include "llvm/IR/Function.h"
51 #include "llvm/IR/InstrTypes.h"
52 #include "llvm/IR/Instruction.h"
53 #include "llvm/IR/Instructions.h"
54 #include "llvm/IR/IntrinsicInst.h"
55 #include "llvm/IR/Value.h"
56 #include "llvm/InitializePasses.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/Local.h"
65 #include "llvm/Transforms/Utils/SizeOpts.h"
74 using namespace consthoist;
76 #define DEBUG_TYPE "consthoist"
78 STATISTIC(NumConstantsHoisted, "Number of constants hoisted");
79 STATISTIC(NumConstantsRebased, "Number of constants rebased");
81 static cl::opt<bool> ConstHoistWithBlockFrequency(
82 "consthoist-with-block-frequency", cl::init(true), cl::Hidden,
83 cl::desc("Enable the use of the block frequency analysis to reduce the "
84 "chance to execute const materialization more frequently than "
85 "without hoisting."));
87 static cl::opt<bool> ConstHoistGEP(
88 "consthoist-gep", cl::init(false), cl::Hidden,
89 cl::desc("Try hoisting constant gep expressions"));
91 static cl::opt<unsigned>
92 MinNumOfDependentToRebase("consthoist-min-num-to-rebase",
93 cl::desc("Do not rebase if number of dependent constants of a Base is less "
95 cl::init(0), cl::Hidden);
99 /// The constant hoisting pass.
100 class ConstantHoistingLegacyPass : public FunctionPass {
102 static char ID; // Pass identification, replacement for typeid
104 ConstantHoistingLegacyPass() : FunctionPass(ID) {
105 initializeConstantHoistingLegacyPassPass(*PassRegistry::getPassRegistry());
108 bool runOnFunction(Function &Fn) override;
110 StringRef getPassName() const override { return "Constant Hoisting"; }
112 void getAnalysisUsage(AnalysisUsage &AU) const override {
113 AU.setPreservesCFG();
114 if (ConstHoistWithBlockFrequency)
115 AU.addRequired<BlockFrequencyInfoWrapperPass>();
116 AU.addRequired<DominatorTreeWrapperPass>();
117 AU.addRequired<ProfileSummaryInfoWrapperPass>();
118 AU.addRequired<TargetTransformInfoWrapperPass>();
122 ConstantHoistingPass Impl;
125 } // end anonymous namespace
127 char ConstantHoistingLegacyPass::ID = 0;
129 INITIALIZE_PASS_BEGIN(ConstantHoistingLegacyPass, "consthoist",
130 "Constant Hoisting", false, false)
131 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
132 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
133 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)
134 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
135 INITIALIZE_PASS_END(ConstantHoistingLegacyPass, "consthoist",
136 "Constant Hoisting", false, false)
138 FunctionPass *llvm::createConstantHoistingPass() {
139 return new ConstantHoistingLegacyPass();
142 /// Perform the constant hoisting optimization for the given function.
143 bool ConstantHoistingLegacyPass::runOnFunction(Function &Fn) {
144 if (skipFunction(Fn))
147 LLVM_DEBUG(dbgs() << "********** Begin Constant Hoisting **********\n");
148 LLVM_DEBUG(dbgs() << "********** Function: " << Fn.getName() << '\n');
151 Impl.runImpl(Fn, getAnalysis<TargetTransformInfoWrapperPass>().getTTI(Fn),
152 getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
153 ConstHoistWithBlockFrequency
154 ? &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI()
157 &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI());
160 LLVM_DEBUG(dbgs() << "********** Function after Constant Hoisting: "
161 << Fn.getName() << '\n');
162 LLVM_DEBUG(dbgs() << Fn);
164 LLVM_DEBUG(dbgs() << "********** End Constant Hoisting **********\n");
169 /// Find the constant materialization insertion point.
170 Instruction *ConstantHoistingPass::findMatInsertPt(Instruction *Inst,
171 unsigned Idx) const {
172 // If the operand is a cast instruction, then we have to materialize the
173 // constant before the cast instruction.
175 Value *Opnd = Inst->getOperand(Idx);
176 if (auto CastInst = dyn_cast<Instruction>(Opnd))
177 if (CastInst->isCast())
181 // The simple and common case. This also includes constant expressions.
182 if (!isa<PHINode>(Inst) && !Inst->isEHPad())
185 // We can't insert directly before a phi node or an eh pad. Insert before
186 // the terminator of the incoming or dominating block.
187 assert(Entry != Inst->getParent() && "PHI or landing pad in entry block!");
188 if (Idx != ~0U && isa<PHINode>(Inst))
189 return cast<PHINode>(Inst)->getIncomingBlock(Idx)->getTerminator();
191 // This must be an EH pad. Iterate over immediate dominators until we find a
192 // non-EH pad. We need to skip over catchswitch blocks, which are both EH pads
194 auto IDom = DT->getNode(Inst->getParent())->getIDom();
195 while (IDom->getBlock()->isEHPad()) {
196 assert(Entry != IDom->getBlock() && "eh pad in entry block");
197 IDom = IDom->getIDom();
200 return IDom->getBlock()->getTerminator();
203 /// Given \p BBs as input, find another set of BBs which collectively
204 /// dominates \p BBs and have the minimal sum of frequencies. Return the BB
205 /// set found in \p BBs.
206 static void findBestInsertionSet(DominatorTree &DT, BlockFrequencyInfo &BFI,
208 SetVector<BasicBlock *> &BBs) {
209 assert(!BBs.count(Entry) && "Assume Entry is not in BBs");
210 // Nodes on the current path to the root.
211 SmallPtrSet<BasicBlock *, 8> Path;
212 // Candidates includes any block 'BB' in set 'BBs' that is not strictly
213 // dominated by any other blocks in set 'BBs', and all nodes in the path
214 // in the dominator tree from Entry to 'BB'.
215 SmallPtrSet<BasicBlock *, 16> Candidates;
216 for (auto BB : BBs) {
217 // Ignore unreachable basic blocks.
218 if (!DT.isReachableFromEntry(BB))
221 // Walk up the dominator tree until Entry or another BB in BBs
222 // is reached. Insert the nodes on the way to the Path.
223 BasicBlock *Node = BB;
224 // The "Path" is a candidate path to be added into Candidates set.
225 bool isCandidate = false;
228 if (Node == Entry || Candidates.count(Node)) {
232 assert(DT.getNode(Node)->getIDom() &&
233 "Entry doens't dominate current Node");
234 Node = DT.getNode(Node)->getIDom()->getBlock();
235 } while (!BBs.count(Node));
237 // If isCandidate is false, Node is another Block in BBs dominating
238 // current 'BB'. Drop the nodes on the Path.
242 // Add nodes on the Path into Candidates.
243 Candidates.insert(Path.begin(), Path.end());
246 // Sort the nodes in Candidates in top-down order and save the nodes
249 SmallVector<BasicBlock *, 16> Orders;
250 Orders.push_back(Entry);
251 while (Idx != Orders.size()) {
252 BasicBlock *Node = Orders[Idx++];
253 for (auto ChildDomNode : DT.getNode(Node)->children()) {
254 if (Candidates.count(ChildDomNode->getBlock()))
255 Orders.push_back(ChildDomNode->getBlock());
259 // Visit Orders in bottom-up order.
260 using InsertPtsCostPair =
261 std::pair<SetVector<BasicBlock *>, BlockFrequency>;
263 // InsertPtsMap is a map from a BB to the best insertion points for the
264 // subtree of BB (subtree not including the BB itself).
265 DenseMap<BasicBlock *, InsertPtsCostPair> InsertPtsMap;
266 InsertPtsMap.reserve(Orders.size() + 1);
267 for (auto RIt = Orders.rbegin(); RIt != Orders.rend(); RIt++) {
268 BasicBlock *Node = *RIt;
269 bool NodeInBBs = BBs.count(Node);
270 auto &InsertPts = InsertPtsMap[Node].first;
271 BlockFrequency &InsertPtsFreq = InsertPtsMap[Node].second;
273 // Return the optimal insert points in BBs.
276 if (InsertPtsFreq > BFI.getBlockFreq(Node) ||
277 (InsertPtsFreq == BFI.getBlockFreq(Node) && InsertPts.size() > 1))
280 BBs.insert(InsertPts.begin(), InsertPts.end());
284 BasicBlock *Parent = DT.getNode(Node)->getIDom()->getBlock();
285 // Initially, ParentInsertPts is empty and ParentPtsFreq is 0. Every child
286 // will update its parent's ParentInsertPts and ParentPtsFreq.
287 auto &ParentInsertPts = InsertPtsMap[Parent].first;
288 BlockFrequency &ParentPtsFreq = InsertPtsMap[Parent].second;
289 // Choose to insert in Node or in subtree of Node.
290 // Don't hoist to EHPad because we may not find a proper place to insert
292 // If the total frequency of InsertPts is the same as the frequency of the
293 // target Node, and InsertPts contains more than one nodes, choose hoisting
294 // to reduce code size.
297 (InsertPtsFreq > BFI.getBlockFreq(Node) ||
298 (InsertPtsFreq == BFI.getBlockFreq(Node) && InsertPts.size() > 1)))) {
299 ParentInsertPts.insert(Node);
300 ParentPtsFreq += BFI.getBlockFreq(Node);
302 ParentInsertPts.insert(InsertPts.begin(), InsertPts.end());
303 ParentPtsFreq += InsertPtsFreq;
308 /// Find an insertion point that dominates all uses.
309 SetVector<Instruction *> ConstantHoistingPass::findConstantInsertionPoint(
310 const ConstantInfo &ConstInfo) const {
311 assert(!ConstInfo.RebasedConstants.empty() && "Invalid constant info entry.");
312 // Collect all basic blocks.
313 SetVector<BasicBlock *> BBs;
314 SetVector<Instruction *> InsertPts;
315 for (auto const &RCI : ConstInfo.RebasedConstants)
316 for (auto const &U : RCI.Uses)
317 BBs.insert(findMatInsertPt(U.Inst, U.OpndIdx)->getParent());
319 if (BBs.count(Entry)) {
320 InsertPts.insert(&Entry->front());
325 findBestInsertionSet(*DT, *BFI, Entry, BBs);
326 for (auto BB : BBs) {
327 BasicBlock::iterator InsertPt = BB->begin();
328 for (; isa<PHINode>(InsertPt) || InsertPt->isEHPad(); ++InsertPt)
330 InsertPts.insert(&*InsertPt);
335 while (BBs.size() >= 2) {
336 BasicBlock *BB, *BB1, *BB2;
337 BB1 = BBs.pop_back_val();
338 BB2 = BBs.pop_back_val();
339 BB = DT->findNearestCommonDominator(BB1, BB2);
341 InsertPts.insert(&Entry->front());
346 assert((BBs.size() == 1) && "Expected only one element.");
347 Instruction &FirstInst = (*BBs.begin())->front();
348 InsertPts.insert(findMatInsertPt(&FirstInst));
352 /// Record constant integer ConstInt for instruction Inst at operand
355 /// The operand at index Idx is not necessarily the constant integer itself. It
356 /// could also be a cast instruction or a constant expression that uses the
357 /// constant integer.
358 void ConstantHoistingPass::collectConstantCandidates(
359 ConstCandMapType &ConstCandMap, Instruction *Inst, unsigned Idx,
360 ConstantInt *ConstInt) {
362 // Ask the target about the cost of materializing the constant for the given
363 // instruction and operand index.
364 if (auto IntrInst = dyn_cast<IntrinsicInst>(Inst))
365 Cost = TTI->getIntImmCostIntrin(IntrInst->getIntrinsicID(), Idx,
366 ConstInt->getValue(), ConstInt->getType(),
367 TargetTransformInfo::TCK_SizeAndLatency);
369 Cost = TTI->getIntImmCostInst(Inst->getOpcode(), Idx, ConstInt->getValue(),
371 TargetTransformInfo::TCK_SizeAndLatency);
373 // Ignore cheap integer constants.
374 if (Cost > TargetTransformInfo::TCC_Basic) {
375 ConstCandMapType::iterator Itr;
377 ConstPtrUnionType Cand = ConstInt;
378 std::tie(Itr, Inserted) = ConstCandMap.insert(std::make_pair(Cand, 0));
380 ConstIntCandVec.push_back(ConstantCandidate(ConstInt));
381 Itr->second = ConstIntCandVec.size() - 1;
383 ConstIntCandVec[Itr->second].addUser(Inst, Idx, Cost);
384 LLVM_DEBUG(if (isa<ConstantInt>(Inst->getOperand(Idx))) dbgs()
385 << "Collect constant " << *ConstInt << " from " << *Inst
386 << " with cost " << Cost << '\n';
387 else dbgs() << "Collect constant " << *ConstInt
388 << " indirectly from " << *Inst << " via "
389 << *Inst->getOperand(Idx) << " with cost " << Cost
394 /// Record constant GEP expression for instruction Inst at operand index Idx.
395 void ConstantHoistingPass::collectConstantCandidates(
396 ConstCandMapType &ConstCandMap, Instruction *Inst, unsigned Idx,
397 ConstantExpr *ConstExpr) {
398 // TODO: Handle vector GEPs
399 if (ConstExpr->getType()->isVectorTy())
402 GlobalVariable *BaseGV = dyn_cast<GlobalVariable>(ConstExpr->getOperand(0));
406 // Get offset from the base GV.
407 PointerType *GVPtrTy = cast<PointerType>(BaseGV->getType());
408 IntegerType *PtrIntTy = DL->getIntPtrType(*Ctx, GVPtrTy->getAddressSpace());
409 APInt Offset(DL->getTypeSizeInBits(PtrIntTy), /*val*/0, /*isSigned*/true);
410 auto *GEPO = cast<GEPOperator>(ConstExpr);
411 if (!GEPO->accumulateConstantOffset(*DL, Offset))
414 if (!Offset.isIntN(32))
417 // A constant GEP expression that has a GlobalVariable as base pointer is
418 // usually lowered to a load from constant pool. Such operation is unlikely
419 // to be cheaper than compute it by <Base + Offset>, which can be lowered to
420 // an ADD instruction or folded into Load/Store instruction.
421 int Cost = TTI->getIntImmCostInst(Instruction::Add, 1, Offset, PtrIntTy,
422 TargetTransformInfo::TCK_SizeAndLatency);
423 ConstCandVecType &ExprCandVec = ConstGEPCandMap[BaseGV];
424 ConstCandMapType::iterator Itr;
426 ConstPtrUnionType Cand = ConstExpr;
427 std::tie(Itr, Inserted) = ConstCandMap.insert(std::make_pair(Cand, 0));
429 ExprCandVec.push_back(ConstantCandidate(
430 ConstantInt::get(Type::getInt32Ty(*Ctx), Offset.getLimitedValue()),
432 Itr->second = ExprCandVec.size() - 1;
434 ExprCandVec[Itr->second].addUser(Inst, Idx, Cost);
437 /// Check the operand for instruction Inst at index Idx.
438 void ConstantHoistingPass::collectConstantCandidates(
439 ConstCandMapType &ConstCandMap, Instruction *Inst, unsigned Idx) {
440 Value *Opnd = Inst->getOperand(Idx);
442 // Visit constant integers.
443 if (auto ConstInt = dyn_cast<ConstantInt>(Opnd)) {
444 collectConstantCandidates(ConstCandMap, Inst, Idx, ConstInt);
448 // Visit cast instructions that have constant integers.
449 if (auto CastInst = dyn_cast<Instruction>(Opnd)) {
450 // Only visit cast instructions, which have been skipped. All other
451 // instructions should have already been visited.
452 if (!CastInst->isCast())
455 if (auto *ConstInt = dyn_cast<ConstantInt>(CastInst->getOperand(0))) {
456 // Pretend the constant is directly used by the instruction and ignore
457 // the cast instruction.
458 collectConstantCandidates(ConstCandMap, Inst, Idx, ConstInt);
463 // Visit constant expressions that have constant integers.
464 if (auto ConstExpr = dyn_cast<ConstantExpr>(Opnd)) {
465 // Handle constant gep expressions.
466 if (ConstHoistGEP && ConstExpr->isGEPWithNoNotionalOverIndexing())
467 collectConstantCandidates(ConstCandMap, Inst, Idx, ConstExpr);
469 // Only visit constant cast expressions.
470 if (!ConstExpr->isCast())
473 if (auto ConstInt = dyn_cast<ConstantInt>(ConstExpr->getOperand(0))) {
474 // Pretend the constant is directly used by the instruction and ignore
475 // the constant expression.
476 collectConstantCandidates(ConstCandMap, Inst, Idx, ConstInt);
482 /// Scan the instruction for expensive integer constants and record them
483 /// in the constant candidate vector.
484 void ConstantHoistingPass::collectConstantCandidates(
485 ConstCandMapType &ConstCandMap, Instruction *Inst) {
486 // Skip all cast instructions. They are visited indirectly later on.
490 // Scan all operands.
491 for (unsigned Idx = 0, E = Inst->getNumOperands(); Idx != E; ++Idx) {
492 // The cost of materializing the constants (defined in
493 // `TargetTransformInfo::getIntImmCostInst`) for instructions which only
494 // take constant variables is lower than `TargetTransformInfo::TCC_Basic`.
495 // So it's safe for us to collect constant candidates from all
497 if (canReplaceOperandWithVariable(Inst, Idx)) {
498 collectConstantCandidates(ConstCandMap, Inst, Idx);
500 } // end of for all operands
503 /// Collect all integer constants in the function that cannot be folded
504 /// into an instruction itself.
505 void ConstantHoistingPass::collectConstantCandidates(Function &Fn) {
506 ConstCandMapType ConstCandMap;
507 for (BasicBlock &BB : Fn) {
508 // Ignore unreachable basic blocks.
509 if (!DT->isReachableFromEntry(&BB))
511 for (Instruction &Inst : BB)
512 collectConstantCandidates(ConstCandMap, &Inst);
516 // This helper function is necessary to deal with values that have different
517 // bit widths (APInt Operator- does not like that). If the value cannot be
518 // represented in uint64 we return an "empty" APInt. This is then interpreted
519 // as the value is not in range.
520 static Optional<APInt> calculateOffsetDiff(const APInt &V1, const APInt &V2) {
521 Optional<APInt> Res = None;
522 unsigned BW = V1.getBitWidth() > V2.getBitWidth() ?
523 V1.getBitWidth() : V2.getBitWidth();
524 uint64_t LimVal1 = V1.getLimitedValue();
525 uint64_t LimVal2 = V2.getLimitedValue();
527 if (LimVal1 == ~0ULL || LimVal2 == ~0ULL)
530 uint64_t Diff = LimVal1 - LimVal2;
531 return APInt(BW, Diff, true);
534 // From a list of constants, one needs to picked as the base and the other
535 // constants will be transformed into an offset from that base constant. The
536 // question is which we can pick best? For example, consider these constants
537 // and their number of uses:
539 // Constants| 2 | 4 | 12 | 42 |
540 // NumUses | 3 | 2 | 8 | 7 |
542 // Selecting constant 12 because it has the most uses will generate negative
543 // offsets for constants 2 and 4 (i.e. -10 and -8 respectively). If negative
544 // offsets lead to less optimal code generation, then there might be better
545 // solutions. Suppose immediates in the range of 0..35 are most optimally
546 // supported by the architecture, then selecting constant 2 is most optimal
547 // because this will generate offsets: 0, 2, 10, 40. Offsets 0, 2 and 10 are in
548 // range 0..35, and thus 3 + 2 + 8 = 13 uses are in range. Selecting 12 would
549 // have only 8 uses in range, so choosing 2 as a base is more optimal. Thus, in
550 // selecting the base constant the range of the offsets is a very important
551 // factor too that we take into account here. This algorithm calculates a total
552 // costs for selecting a constant as the base and substract the costs if
553 // immediates are out of range. It has quadratic complexity, so we call this
554 // function only when we're optimising for size and there are less than 100
555 // constants, we fall back to the straightforward algorithm otherwise
556 // which does not do all the offset calculations.
558 ConstantHoistingPass::maximizeConstantsInRange(ConstCandVecType::iterator S,
559 ConstCandVecType::iterator E,
560 ConstCandVecType::iterator &MaxCostItr) {
561 unsigned NumUses = 0;
563 bool OptForSize = Entry->getParent()->hasOptSize() ||
564 llvm::shouldOptimizeForSize(Entry->getParent(), PSI, BFI,
565 PGSOQueryType::IRPass);
566 if (!OptForSize || std::distance(S,E) > 100) {
567 for (auto ConstCand = S; ConstCand != E; ++ConstCand) {
568 NumUses += ConstCand->Uses.size();
569 if (ConstCand->CumulativeCost > MaxCostItr->CumulativeCost)
570 MaxCostItr = ConstCand;
575 LLVM_DEBUG(dbgs() << "== Maximize constants in range ==\n");
577 for (auto ConstCand = S; ConstCand != E; ++ConstCand) {
578 auto Value = ConstCand->ConstInt->getValue();
579 Type *Ty = ConstCand->ConstInt->getType();
581 NumUses += ConstCand->Uses.size();
582 LLVM_DEBUG(dbgs() << "= Constant: " << ConstCand->ConstInt->getValue()
585 for (auto User : ConstCand->Uses) {
586 unsigned Opcode = User.Inst->getOpcode();
587 unsigned OpndIdx = User.OpndIdx;
588 Cost += TTI->getIntImmCostInst(Opcode, OpndIdx, Value, Ty,
589 TargetTransformInfo::TCK_SizeAndLatency);
590 LLVM_DEBUG(dbgs() << "Cost: " << Cost << "\n");
592 for (auto C2 = S; C2 != E; ++C2) {
593 Optional<APInt> Diff = calculateOffsetDiff(
594 C2->ConstInt->getValue(),
595 ConstCand->ConstInt->getValue());
598 TTI->getIntImmCodeSizeCost(Opcode, OpndIdx, Diff.getValue(), Ty);
600 LLVM_DEBUG(dbgs() << "Offset " << Diff.getValue() << " "
601 << "has penalty: " << ImmCosts << "\n"
602 << "Adjusted cost: " << Cost << "\n");
606 LLVM_DEBUG(dbgs() << "Cumulative cost: " << Cost << "\n");
607 if (Cost > MaxCost) {
609 MaxCostItr = ConstCand;
610 LLVM_DEBUG(dbgs() << "New candidate: " << MaxCostItr->ConstInt->getValue()
617 /// Find the base constant within the given range and rebase all other
618 /// constants with respect to the base constant.
619 void ConstantHoistingPass::findAndMakeBaseConstant(
620 ConstCandVecType::iterator S, ConstCandVecType::iterator E,
621 SmallVectorImpl<consthoist::ConstantInfo> &ConstInfoVec) {
623 unsigned NumUses = maximizeConstantsInRange(S, E, MaxCostItr);
625 // Don't hoist constants that have only one use.
629 ConstantInt *ConstInt = MaxCostItr->ConstInt;
630 ConstantExpr *ConstExpr = MaxCostItr->ConstExpr;
631 ConstantInfo ConstInfo;
632 ConstInfo.BaseInt = ConstInt;
633 ConstInfo.BaseExpr = ConstExpr;
634 Type *Ty = ConstInt->getType();
636 // Rebase the constants with respect to the base constant.
637 for (auto ConstCand = S; ConstCand != E; ++ConstCand) {
638 APInt Diff = ConstCand->ConstInt->getValue() - ConstInt->getValue();
639 Constant *Offset = Diff == 0 ? nullptr : ConstantInt::get(Ty, Diff);
641 ConstCand->ConstExpr ? ConstCand->ConstExpr->getType() : nullptr;
642 ConstInfo.RebasedConstants.push_back(
643 RebasedConstantInfo(std::move(ConstCand->Uses), Offset, ConstTy));
645 ConstInfoVec.push_back(std::move(ConstInfo));
648 /// Finds and combines constant candidates that can be easily
649 /// rematerialized with an add from a common base constant.
650 void ConstantHoistingPass::findBaseConstants(GlobalVariable *BaseGV) {
651 // If BaseGV is nullptr, find base among candidate constant integers;
652 // Otherwise find base among constant GEPs that share the same BaseGV.
653 ConstCandVecType &ConstCandVec = BaseGV ?
654 ConstGEPCandMap[BaseGV] : ConstIntCandVec;
655 ConstInfoVecType &ConstInfoVec = BaseGV ?
656 ConstGEPInfoMap[BaseGV] : ConstIntInfoVec;
658 // Sort the constants by value and type. This invalidates the mapping!
659 llvm::stable_sort(ConstCandVec, [](const ConstantCandidate &LHS,
660 const ConstantCandidate &RHS) {
661 if (LHS.ConstInt->getType() != RHS.ConstInt->getType())
662 return LHS.ConstInt->getType()->getBitWidth() <
663 RHS.ConstInt->getType()->getBitWidth();
664 return LHS.ConstInt->getValue().ult(RHS.ConstInt->getValue());
667 // Simple linear scan through the sorted constant candidate vector for viable
669 auto MinValItr = ConstCandVec.begin();
670 for (auto CC = std::next(ConstCandVec.begin()), E = ConstCandVec.end();
672 if (MinValItr->ConstInt->getType() == CC->ConstInt->getType()) {
673 Type *MemUseValTy = nullptr;
674 for (auto &U : CC->Uses) {
676 if (LoadInst *LI = dyn_cast<LoadInst>(UI)) {
677 MemUseValTy = LI->getType();
679 } else if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
680 // Make sure the constant is used as pointer operand of the StoreInst.
681 if (SI->getPointerOperand() == SI->getOperand(U.OpndIdx)) {
682 MemUseValTy = SI->getValueOperand()->getType();
688 // Check if the constant is in range of an add with immediate.
689 APInt Diff = CC->ConstInt->getValue() - MinValItr->ConstInt->getValue();
690 if ((Diff.getBitWidth() <= 64) &&
691 TTI->isLegalAddImmediate(Diff.getSExtValue()) &&
692 // Check if Diff can be used as offset in addressing mode of the user
693 // memory instruction.
694 (!MemUseValTy || TTI->isLegalAddressingMode(MemUseValTy,
695 /*BaseGV*/nullptr, /*BaseOffset*/Diff.getSExtValue(),
696 /*HasBaseReg*/true, /*Scale*/0)))
699 // We either have now a different constant type or the constant is not in
700 // range of an add with immediate anymore.
701 findAndMakeBaseConstant(MinValItr, CC, ConstInfoVec);
702 // Start a new base constant search.
705 // Finalize the last base constant search.
706 findAndMakeBaseConstant(MinValItr, ConstCandVec.end(), ConstInfoVec);
709 /// Updates the operand at Idx in instruction Inst with the result of
710 /// instruction Mat. If the instruction is a PHI node then special
711 /// handling for duplicate values form the same incoming basic block is
713 /// \return The update will always succeed, but the return value indicated if
714 /// Mat was used for the update or not.
715 static bool updateOperand(Instruction *Inst, unsigned Idx, Instruction *Mat) {
716 if (auto PHI = dyn_cast<PHINode>(Inst)) {
717 // Check if any previous operand of the PHI node has the same incoming basic
718 // block. This is a very odd case that happens when the incoming basic block
719 // has a switch statement. In this case use the same value as the previous
720 // operand(s), otherwise we will fail verification due to different values.
721 // The values are actually the same, but the variable names are different
722 // and the verifier doesn't like that.
723 BasicBlock *IncomingBB = PHI->getIncomingBlock(Idx);
724 for (unsigned i = 0; i < Idx; ++i) {
725 if (PHI->getIncomingBlock(i) == IncomingBB) {
726 Value *IncomingVal = PHI->getIncomingValue(i);
727 Inst->setOperand(Idx, IncomingVal);
733 Inst->setOperand(Idx, Mat);
737 /// Emit materialization code for all rebased constants and update their
739 void ConstantHoistingPass::emitBaseConstants(Instruction *Base,
742 const ConstantUser &ConstUser) {
743 Instruction *Mat = Base;
745 // The same offset can be dereferenced to different types in nested struct.
746 if (!Offset && Ty && Ty != Base->getType())
747 Offset = ConstantInt::get(Type::getInt32Ty(*Ctx), 0);
750 Instruction *InsertionPt = findMatInsertPt(ConstUser.Inst,
753 // Constant being rebased is a ConstantExpr.
754 PointerType *Int8PtrTy = Type::getInt8PtrTy(*Ctx,
755 cast<PointerType>(Ty)->getAddressSpace());
756 Base = new BitCastInst(Base, Int8PtrTy, "base_bitcast", InsertionPt);
757 Mat = GetElementPtrInst::Create(Int8PtrTy->getElementType(), Base,
758 Offset, "mat_gep", InsertionPt);
759 Mat = new BitCastInst(Mat, Ty, "mat_bitcast", InsertionPt);
761 // Constant being rebased is a ConstantInt.
762 Mat = BinaryOperator::Create(Instruction::Add, Base, Offset,
763 "const_mat", InsertionPt);
765 LLVM_DEBUG(dbgs() << "Materialize constant (" << *Base->getOperand(0)
766 << " + " << *Offset << ") in BB "
767 << Mat->getParent()->getName() << '\n'
769 Mat->setDebugLoc(ConstUser.Inst->getDebugLoc());
771 Value *Opnd = ConstUser.Inst->getOperand(ConstUser.OpndIdx);
773 // Visit constant integer.
774 if (isa<ConstantInt>(Opnd)) {
775 LLVM_DEBUG(dbgs() << "Update: " << *ConstUser.Inst << '\n');
776 if (!updateOperand(ConstUser.Inst, ConstUser.OpndIdx, Mat) && Offset)
777 Mat->eraseFromParent();
778 LLVM_DEBUG(dbgs() << "To : " << *ConstUser.Inst << '\n');
782 // Visit cast instruction.
783 if (auto CastInst = dyn_cast<Instruction>(Opnd)) {
784 assert(CastInst->isCast() && "Expected an cast instruction!");
785 // Check if we already have visited this cast instruction before to avoid
786 // unnecessary cloning.
787 Instruction *&ClonedCastInst = ClonedCastMap[CastInst];
788 if (!ClonedCastInst) {
789 ClonedCastInst = CastInst->clone();
790 ClonedCastInst->setOperand(0, Mat);
791 ClonedCastInst->insertAfter(CastInst);
792 // Use the same debug location as the original cast instruction.
793 ClonedCastInst->setDebugLoc(CastInst->getDebugLoc());
794 LLVM_DEBUG(dbgs() << "Clone instruction: " << *CastInst << '\n'
795 << "To : " << *ClonedCastInst << '\n');
798 LLVM_DEBUG(dbgs() << "Update: " << *ConstUser.Inst << '\n');
799 updateOperand(ConstUser.Inst, ConstUser.OpndIdx, ClonedCastInst);
800 LLVM_DEBUG(dbgs() << "To : " << *ConstUser.Inst << '\n');
804 // Visit constant expression.
805 if (auto ConstExpr = dyn_cast<ConstantExpr>(Opnd)) {
806 if (ConstExpr->isGEPWithNoNotionalOverIndexing()) {
807 // Operand is a ConstantGEP, replace it.
808 updateOperand(ConstUser.Inst, ConstUser.OpndIdx, Mat);
812 // Aside from constant GEPs, only constant cast expressions are collected.
813 assert(ConstExpr->isCast() && "ConstExpr should be a cast");
814 Instruction *ConstExprInst = ConstExpr->getAsInstruction();
815 ConstExprInst->setOperand(0, Mat);
816 ConstExprInst->insertBefore(findMatInsertPt(ConstUser.Inst,
819 // Use the same debug location as the instruction we are about to update.
820 ConstExprInst->setDebugLoc(ConstUser.Inst->getDebugLoc());
822 LLVM_DEBUG(dbgs() << "Create instruction: " << *ConstExprInst << '\n'
823 << "From : " << *ConstExpr << '\n');
824 LLVM_DEBUG(dbgs() << "Update: " << *ConstUser.Inst << '\n');
825 if (!updateOperand(ConstUser.Inst, ConstUser.OpndIdx, ConstExprInst)) {
826 ConstExprInst->eraseFromParent();
828 Mat->eraseFromParent();
830 LLVM_DEBUG(dbgs() << "To : " << *ConstUser.Inst << '\n');
835 /// Hoist and hide the base constant behind a bitcast and emit
836 /// materialization code for derived constants.
837 bool ConstantHoistingPass::emitBaseConstants(GlobalVariable *BaseGV) {
838 bool MadeChange = false;
839 SmallVectorImpl<consthoist::ConstantInfo> &ConstInfoVec =
840 BaseGV ? ConstGEPInfoMap[BaseGV] : ConstIntInfoVec;
841 for (auto const &ConstInfo : ConstInfoVec) {
842 SetVector<Instruction *> IPSet = findConstantInsertionPoint(ConstInfo);
843 // We can have an empty set if the function contains unreachable blocks.
847 unsigned UsesNum = 0;
848 unsigned ReBasesNum = 0;
849 unsigned NotRebasedNum = 0;
850 for (Instruction *IP : IPSet) {
851 // First, collect constants depending on this IP of the base.
853 using RebasedUse = std::tuple<Constant *, Type *, ConstantUser>;
854 SmallVector<RebasedUse, 4> ToBeRebased;
855 for (auto const &RCI : ConstInfo.RebasedConstants) {
856 for (auto const &U : RCI.Uses) {
858 BasicBlock *OrigMatInsertBB =
859 findMatInsertPt(U.Inst, U.OpndIdx)->getParent();
860 // If Base constant is to be inserted in multiple places,
861 // generate rebase for U using the Base dominating U.
862 if (IPSet.size() == 1 ||
863 DT->dominates(IP->getParent(), OrigMatInsertBB))
864 ToBeRebased.push_back(RebasedUse(RCI.Offset, RCI.Ty, U));
869 // If only few constants depend on this IP of base, skip rebasing,
870 // assuming the base and the rebased have the same materialization cost.
871 if (ToBeRebased.size() < MinNumOfDependentToRebase) {
872 NotRebasedNum += ToBeRebased.size();
876 // Emit an instance of the base at this IP.
877 Instruction *Base = nullptr;
878 // Hoist and hide the base constant behind a bitcast.
879 if (ConstInfo.BaseExpr) {
880 assert(BaseGV && "A base constant expression must have an base GV");
881 Type *Ty = ConstInfo.BaseExpr->getType();
882 Base = new BitCastInst(ConstInfo.BaseExpr, Ty, "const", IP);
884 IntegerType *Ty = ConstInfo.BaseInt->getType();
885 Base = new BitCastInst(ConstInfo.BaseInt, Ty, "const", IP);
888 Base->setDebugLoc(IP->getDebugLoc());
890 LLVM_DEBUG(dbgs() << "Hoist constant (" << *ConstInfo.BaseInt
891 << ") to BB " << IP->getParent()->getName() << '\n'
894 // Emit materialization code for rebased constants depending on this IP.
895 for (auto const &R : ToBeRebased) {
896 Constant *Off = std::get<0>(R);
897 Type *Ty = std::get<1>(R);
898 ConstantUser U = std::get<2>(R);
899 emitBaseConstants(Base, Off, Ty, U);
901 // Use the same debug location as the last user of the constant.
902 Base->setDebugLoc(DILocation::getMergedLocation(
903 Base->getDebugLoc(), U.Inst->getDebugLoc()));
905 assert(!Base->use_empty() && "The use list is empty!?");
906 assert(isa<Instruction>(Base->user_back()) &&
907 "All uses should be instructions.");
912 // Expect all uses are rebased after rebase is done.
913 assert(UsesNum == (ReBasesNum + NotRebasedNum) &&
914 "Not all uses are rebased");
916 NumConstantsHoisted++;
918 // Base constant is also included in ConstInfo.RebasedConstants, so
919 // deduct 1 from ConstInfo.RebasedConstants.size().
920 NumConstantsRebased += ConstInfo.RebasedConstants.size() - 1;
927 /// Check all cast instructions we made a copy of and remove them if they
928 /// have no more users.
929 void ConstantHoistingPass::deleteDeadCastInst() const {
930 for (auto const &I : ClonedCastMap)
931 if (I.first->use_empty())
932 I.first->eraseFromParent();
935 /// Optimize expensive integer constants in the given function.
936 bool ConstantHoistingPass::runImpl(Function &Fn, TargetTransformInfo &TTI,
937 DominatorTree &DT, BlockFrequencyInfo *BFI,
938 BasicBlock &Entry, ProfileSummaryInfo *PSI) {
942 this->DL = &Fn.getParent()->getDataLayout();
943 this->Ctx = &Fn.getContext();
944 this->Entry = &Entry;
946 // Collect all constant candidates.
947 collectConstantCandidates(Fn);
949 // Combine constants that can be easily materialized with an add from a common
951 if (!ConstIntCandVec.empty())
952 findBaseConstants(nullptr);
953 for (auto &MapEntry : ConstGEPCandMap)
954 if (!MapEntry.second.empty())
955 findBaseConstants(MapEntry.first);
957 // Finally hoist the base constant and emit materialization code for dependent
959 bool MadeChange = false;
960 if (!ConstIntInfoVec.empty())
961 MadeChange = emitBaseConstants(nullptr);
962 for (auto MapEntry : ConstGEPInfoMap)
963 if (!MapEntry.second.empty())
964 MadeChange |= emitBaseConstants(MapEntry.first);
967 // Cleanup dead instructions.
968 deleteDeadCastInst();
975 PreservedAnalyses ConstantHoistingPass::run(Function &F,
976 FunctionAnalysisManager &AM) {
977 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
978 auto &TTI = AM.getResult<TargetIRAnalysis>(F);
979 auto BFI = ConstHoistWithBlockFrequency
980 ? &AM.getResult<BlockFrequencyAnalysis>(F)
982 auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F);
983 auto *PSI = MAMProxy.getCachedResult<ProfileSummaryAnalysis>(*F.getParent());
984 if (!runImpl(F, TTI, DT, BFI, F.getEntryBlock(), PSI))
985 return PreservedAnalyses::all();
987 PreservedAnalyses PA;
988 PA.preserveSet<CFGAnalyses>();