1 //===-- LowerBitSets.cpp - Bitset lowering pass ---------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass lowers bitset metadata and calls to the llvm.bitset.test intrinsic.
11 // See http://llvm.org/docs/LangRef.html#bitsets for more information.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/Transforms/IPO/LowerBitSets.h"
16 #include "llvm/Transforms/IPO.h"
17 #include "llvm/ADT/EquivalenceClasses.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/ADT/Triple.h"
20 #include "llvm/IR/Constant.h"
21 #include "llvm/IR/Constants.h"
22 #include "llvm/IR/GlobalVariable.h"
23 #include "llvm/IR/IRBuilder.h"
24 #include "llvm/IR/Instructions.h"
25 #include "llvm/IR/Intrinsics.h"
26 #include "llvm/IR/Module.h"
27 #include "llvm/IR/Operator.h"
28 #include "llvm/Pass.h"
29 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
33 #define DEBUG_TYPE "lowerbitsets"
35 STATISTIC(ByteArraySizeBits, "Byte array size in bits");
36 STATISTIC(ByteArraySizeBytes, "Byte array size in bytes");
37 STATISTIC(NumByteArraysCreated, "Number of byte arrays created");
38 STATISTIC(NumBitSetCallsLowered, "Number of bitset calls lowered");
39 STATISTIC(NumBitSetDisjointSets, "Number of disjoint sets of bitsets");
41 bool BitSetInfo::containsGlobalOffset(uint64_t Offset) const {
42 if (Offset < ByteOffset)
45 if ((Offset - ByteOffset) % (uint64_t(1) << AlignLog2) != 0)
48 uint64_t BitOffset = (Offset - ByteOffset) >> AlignLog2;
49 if (BitOffset >= BitSize)
52 return Bits.count(BitOffset);
55 bool BitSetInfo::containsValue(
57 const DenseMap<GlobalVariable *, uint64_t> &GlobalLayout, Value *V,
58 uint64_t COffset) const {
59 if (auto GV = dyn_cast<GlobalVariable>(V)) {
60 auto I = GlobalLayout.find(GV);
61 if (I == GlobalLayout.end())
63 return containsGlobalOffset(I->second + COffset);
66 if (auto GEP = dyn_cast<GEPOperator>(V)) {
67 APInt APOffset(DL.getPointerSizeInBits(0), 0);
68 bool Result = GEP->accumulateConstantOffset(DL, APOffset);
71 COffset += APOffset.getZExtValue();
72 return containsValue(DL, GlobalLayout, GEP->getPointerOperand(),
76 if (auto Op = dyn_cast<Operator>(V)) {
77 if (Op->getOpcode() == Instruction::BitCast)
78 return containsValue(DL, GlobalLayout, Op->getOperand(0), COffset);
80 if (Op->getOpcode() == Instruction::Select)
81 return containsValue(DL, GlobalLayout, Op->getOperand(1), COffset) &&
82 containsValue(DL, GlobalLayout, Op->getOperand(2), COffset);
88 BitSetInfo BitSetBuilder::build() {
92 // Normalize each offset against the minimum observed offset, and compute
93 // the bitwise OR of each of the offsets. The number of trailing zeros
94 // in the mask gives us the log2 of the alignment of all offsets, which
95 // allows us to compress the bitset by only storing one bit per aligned
98 for (uint64_t &Offset : Offsets) {
104 BSI.ByteOffset = Min;
108 BSI.AlignLog2 = countTrailingZeros(Mask, ZB_Undefined);
110 // Build the compressed bitset while normalizing the offsets against the
111 // computed alignment.
112 BSI.BitSize = ((Max - Min) >> BSI.AlignLog2) + 1;
113 for (uint64_t Offset : Offsets) {
114 Offset >>= BSI.AlignLog2;
115 BSI.Bits.insert(Offset);
121 void GlobalLayoutBuilder::addFragment(const std::set<uint64_t> &F) {
122 // Create a new fragment to hold the layout for F.
123 Fragments.emplace_back();
124 std::vector<uint64_t> &Fragment = Fragments.back();
125 uint64_t FragmentIndex = Fragments.size() - 1;
127 for (auto ObjIndex : F) {
128 uint64_t OldFragmentIndex = FragmentMap[ObjIndex];
129 if (OldFragmentIndex == 0) {
130 // We haven't seen this object index before, so just add it to the current
132 Fragment.push_back(ObjIndex);
134 // This index belongs to an existing fragment. Copy the elements of the
135 // old fragment into this one and clear the old fragment. We don't update
136 // the fragment map just yet, this ensures that any further references to
137 // indices from the old fragment in this fragment do not insert any more
139 std::vector<uint64_t> &OldFragment = Fragments[OldFragmentIndex];
140 Fragment.insert(Fragment.end(), OldFragment.begin(), OldFragment.end());
145 // Update the fragment map to point our object indices to this fragment.
146 for (uint64_t ObjIndex : Fragment)
147 FragmentMap[ObjIndex] = FragmentIndex;
150 void ByteArrayBuilder::allocate(const std::set<uint64_t> &Bits,
151 uint64_t BitSize, uint64_t &AllocByteOffset,
152 uint8_t &AllocMask) {
153 // Find the smallest current allocation.
155 for (unsigned I = 1; I != BitsPerByte; ++I)
156 if (BitAllocs[I] < BitAllocs[Bit])
159 AllocByteOffset = BitAllocs[Bit];
161 // Add our size to it.
162 unsigned ReqSize = AllocByteOffset + BitSize;
163 BitAllocs[Bit] = ReqSize;
164 if (Bytes.size() < ReqSize)
165 Bytes.resize(ReqSize);
168 AllocMask = 1 << Bit;
169 for (uint64_t B : Bits)
170 Bytes[AllocByteOffset + B] |= AllocMask;
175 struct ByteArrayInfo {
176 std::set<uint64_t> Bits;
178 GlobalVariable *ByteArray;
182 struct LowerBitSets : public ModulePass {
184 LowerBitSets() : ModulePass(ID) {
185 initializeLowerBitSetsPass(*PassRegistry::getPassRegistry());
190 bool LinkerSubsectionsViaSymbols;
193 IntegerType *Int32Ty;
195 IntegerType *Int64Ty;
198 // The llvm.bitsets named metadata.
199 NamedMDNode *BitSetNM;
201 // Mapping from bitset mdstrings to the call sites that test them.
202 DenseMap<MDString *, std::vector<CallInst *>> BitSetTestCallSites;
204 std::vector<ByteArrayInfo> ByteArrayInfos;
207 buildBitSet(MDString *BitSet,
208 const DenseMap<GlobalVariable *, uint64_t> &GlobalLayout);
209 ByteArrayInfo *createByteArray(BitSetInfo &BSI);
210 void allocateByteArrays();
211 Value *createBitSetTest(IRBuilder<> &B, BitSetInfo &BSI, ByteArrayInfo *&BAI,
214 lowerBitSetCall(CallInst *CI, BitSetInfo &BSI, ByteArrayInfo *&BAI,
215 GlobalVariable *CombinedGlobal,
216 const DenseMap<GlobalVariable *, uint64_t> &GlobalLayout);
217 void buildBitSetsFromGlobals(const std::vector<MDString *> &BitSets,
218 const std::vector<GlobalVariable *> &Globals);
220 bool eraseBitSetMetadata();
222 bool doInitialization(Module &M) override;
223 bool runOnModule(Module &M) override;
228 INITIALIZE_PASS_BEGIN(LowerBitSets, "lowerbitsets",
229 "Lower bitset metadata", false, false)
230 INITIALIZE_PASS_END(LowerBitSets, "lowerbitsets",
231 "Lower bitset metadata", false, false)
232 char LowerBitSets::ID = 0;
234 ModulePass *llvm::createLowerBitSetsPass() { return new LowerBitSets; }
236 bool LowerBitSets::doInitialization(Module &Mod) {
238 const DataLayout &DL = Mod.getDataLayout();
240 Triple TargetTriple(M->getTargetTriple());
241 LinkerSubsectionsViaSymbols = TargetTriple.isMacOSX();
243 Int1Ty = Type::getInt1Ty(M->getContext());
244 Int8Ty = Type::getInt8Ty(M->getContext());
245 Int32Ty = Type::getInt32Ty(M->getContext());
246 Int32PtrTy = PointerType::getUnqual(Int32Ty);
247 Int64Ty = Type::getInt64Ty(M->getContext());
248 IntPtrTy = DL.getIntPtrType(M->getContext(), 0);
250 BitSetNM = M->getNamedMetadata("llvm.bitsets");
252 BitSetTestCallSites.clear();
257 /// Build a bit set for BitSet using the object layouts in
259 BitSetInfo LowerBitSets::buildBitSet(
261 const DenseMap<GlobalVariable *, uint64_t> &GlobalLayout) {
264 // Compute the byte offset of each element of this bitset.
266 for (MDNode *Op : BitSetNM->operands()) {
267 if (Op->getOperand(0) != BitSet || !Op->getOperand(1))
269 auto OpGlobal = cast<GlobalVariable>(
270 cast<ConstantAsMetadata>(Op->getOperand(1))->getValue());
272 cast<ConstantInt>(cast<ConstantAsMetadata>(Op->getOperand(2))
273 ->getValue())->getZExtValue();
275 Offset += GlobalLayout.find(OpGlobal)->second;
277 BSB.addOffset(Offset);
284 /// Build a test that bit BitOffset mod sizeof(Bits)*8 is set in
285 /// Bits. This pattern matches to the bt instruction on x86.
286 static Value *createMaskedBitTest(IRBuilder<> &B, Value *Bits,
288 auto BitsType = cast<IntegerType>(Bits->getType());
289 unsigned BitWidth = BitsType->getBitWidth();
291 BitOffset = B.CreateZExtOrTrunc(BitOffset, BitsType);
293 B.CreateAnd(BitOffset, ConstantInt::get(BitsType, BitWidth - 1));
294 Value *BitMask = B.CreateShl(ConstantInt::get(BitsType, 1), BitIndex);
295 Value *MaskedBits = B.CreateAnd(Bits, BitMask);
296 return B.CreateICmpNE(MaskedBits, ConstantInt::get(BitsType, 0));
299 ByteArrayInfo *LowerBitSets::createByteArray(BitSetInfo &BSI) {
300 // Create globals to stand in for byte arrays and masks. These never actually
301 // get initialized, we RAUW and erase them later in allocateByteArrays() once
302 // we know the offset and mask to use.
303 auto ByteArrayGlobal = new GlobalVariable(
304 *M, Int8Ty, /*isConstant=*/true, GlobalValue::PrivateLinkage, nullptr);
305 auto MaskGlobal = new GlobalVariable(
306 *M, Int8Ty, /*isConstant=*/true, GlobalValue::PrivateLinkage, nullptr);
308 ByteArrayInfos.emplace_back();
309 ByteArrayInfo *BAI = &ByteArrayInfos.back();
311 BAI->Bits = BSI.Bits;
312 BAI->BitSize = BSI.BitSize;
313 BAI->ByteArray = ByteArrayGlobal;
314 BAI->Mask = ConstantExpr::getPtrToInt(MaskGlobal, Int8Ty);
318 void LowerBitSets::allocateByteArrays() {
319 std::stable_sort(ByteArrayInfos.begin(), ByteArrayInfos.end(),
320 [](const ByteArrayInfo &BAI1, const ByteArrayInfo &BAI2) {
321 return BAI1.BitSize > BAI2.BitSize;
324 std::vector<uint64_t> ByteArrayOffsets(ByteArrayInfos.size());
326 ByteArrayBuilder BAB;
327 for (unsigned I = 0; I != ByteArrayInfos.size(); ++I) {
328 ByteArrayInfo *BAI = &ByteArrayInfos[I];
331 BAB.allocate(BAI->Bits, BAI->BitSize, ByteArrayOffsets[I], Mask);
333 BAI->Mask->replaceAllUsesWith(ConstantInt::get(Int8Ty, Mask));
334 cast<GlobalVariable>(BAI->Mask->getOperand(0))->eraseFromParent();
337 Constant *ByteArrayConst = ConstantDataArray::get(M->getContext(), BAB.Bytes);
339 new GlobalVariable(*M, ByteArrayConst->getType(), /*isConstant=*/true,
340 GlobalValue::PrivateLinkage, ByteArrayConst);
342 for (unsigned I = 0; I != ByteArrayInfos.size(); ++I) {
343 ByteArrayInfo *BAI = &ByteArrayInfos[I];
345 Constant *Idxs[] = {ConstantInt::get(IntPtrTy, 0),
346 ConstantInt::get(IntPtrTy, ByteArrayOffsets[I])};
347 Constant *GEP = ConstantExpr::getInBoundsGetElementPtr(ByteArray, Idxs);
349 // Create an alias instead of RAUW'ing the gep directly. On x86 this ensures
350 // that the pc-relative displacement is folded into the lea instead of the
351 // test instruction getting another displacement.
352 if (LinkerSubsectionsViaSymbols) {
353 BAI->ByteArray->replaceAllUsesWith(GEP);
355 GlobalAlias *Alias = GlobalAlias::create(
356 Int8Ty, 0, GlobalValue::PrivateLinkage, "bits", GEP, M);
357 BAI->ByteArray->replaceAllUsesWith(Alias);
359 BAI->ByteArray->eraseFromParent();
362 ByteArraySizeBits = BAB.BitAllocs[0] + BAB.BitAllocs[1] + BAB.BitAllocs[2] +
363 BAB.BitAllocs[3] + BAB.BitAllocs[4] + BAB.BitAllocs[5] +
364 BAB.BitAllocs[6] + BAB.BitAllocs[7];
365 ByteArraySizeBytes = BAB.Bytes.size();
368 /// Build a test that bit BitOffset is set in BSI, where
369 /// BitSetGlobal is a global containing the bits in BSI.
370 Value *LowerBitSets::createBitSetTest(IRBuilder<> &B, BitSetInfo &BSI,
371 ByteArrayInfo *&BAI, Value *BitOffset) {
372 if (BSI.BitSize <= 64) {
373 // If the bit set is sufficiently small, we can avoid a load by bit testing
376 if (BSI.BitSize <= 32)
382 for (auto Bit : BSI.Bits)
383 Bits |= uint64_t(1) << Bit;
384 Constant *BitsConst = ConstantInt::get(BitsTy, Bits);
385 return createMaskedBitTest(B, BitsConst, BitOffset);
388 ++NumByteArraysCreated;
389 BAI = createByteArray(BSI);
392 Value *ByteAddr = B.CreateGEP(BAI->ByteArray, BitOffset);
393 Value *Byte = B.CreateLoad(ByteAddr);
395 Value *ByteAndMask = B.CreateAnd(Byte, BAI->Mask);
396 return B.CreateICmpNE(ByteAndMask, ConstantInt::get(Int8Ty, 0));
400 /// Lower a llvm.bitset.test call to its implementation. Returns the value to
401 /// replace the call with.
402 Value *LowerBitSets::lowerBitSetCall(
403 CallInst *CI, BitSetInfo &BSI, ByteArrayInfo *&BAI,
404 GlobalVariable *CombinedGlobal,
405 const DenseMap<GlobalVariable *, uint64_t> &GlobalLayout) {
406 Value *Ptr = CI->getArgOperand(0);
407 const DataLayout &DL = M->getDataLayout();
409 if (BSI.containsValue(DL, GlobalLayout, Ptr))
410 return ConstantInt::getTrue(CombinedGlobal->getParent()->getContext());
412 Constant *GlobalAsInt = ConstantExpr::getPtrToInt(CombinedGlobal, IntPtrTy);
413 Constant *OffsetedGlobalAsInt = ConstantExpr::getAdd(
414 GlobalAsInt, ConstantInt::get(IntPtrTy, BSI.ByteOffset));
416 BasicBlock *InitialBB = CI->getParent();
420 Value *PtrAsInt = B.CreatePtrToInt(Ptr, IntPtrTy);
422 if (BSI.isSingleOffset())
423 return B.CreateICmpEQ(PtrAsInt, OffsetedGlobalAsInt);
425 Value *PtrOffset = B.CreateSub(PtrAsInt, OffsetedGlobalAsInt);
428 if (BSI.AlignLog2 == 0) {
429 BitOffset = PtrOffset;
431 // We need to check that the offset both falls within our range and is
432 // suitably aligned. We can check both properties at the same time by
433 // performing a right rotate by log2(alignment) followed by an integer
434 // comparison against the bitset size. The rotate will move the lower
435 // order bits that need to be zero into the higher order bits of the
436 // result, causing the comparison to fail if they are nonzero. The rotate
437 // also conveniently gives us a bit offset to use during the load from
440 B.CreateLShr(PtrOffset, ConstantInt::get(IntPtrTy, BSI.AlignLog2));
441 Value *OffsetSHL = B.CreateShl(
443 ConstantInt::get(IntPtrTy, DL.getPointerSizeInBits(0) - BSI.AlignLog2));
444 BitOffset = B.CreateOr(OffsetSHR, OffsetSHL);
447 Constant *BitSizeConst = ConstantInt::get(IntPtrTy, BSI.BitSize);
448 Value *OffsetInRange = B.CreateICmpULT(BitOffset, BitSizeConst);
450 // If the bit set is all ones, testing against it is unnecessary.
452 return OffsetInRange;
454 TerminatorInst *Term = SplitBlockAndInsertIfThen(OffsetInRange, CI, false);
455 IRBuilder<> ThenB(Term);
457 // Now that we know that the offset is in range and aligned, load the
458 // appropriate bit from the bitset.
459 Value *Bit = createBitSetTest(ThenB, BSI, BAI, BitOffset);
461 // The value we want is 0 if we came directly from the initial block
462 // (having failed the range or alignment checks), or the loaded bit if
463 // we came from the block in which we loaded it.
464 B.SetInsertPoint(CI);
465 PHINode *P = B.CreatePHI(Int1Ty, 2);
466 P->addIncoming(ConstantInt::get(Int1Ty, 0), InitialBB);
467 P->addIncoming(Bit, ThenB.GetInsertBlock());
471 /// Given a disjoint set of bitsets and globals, layout the globals, build the
472 /// bit sets and lower the llvm.bitset.test calls.
473 void LowerBitSets::buildBitSetsFromGlobals(
474 const std::vector<MDString *> &BitSets,
475 const std::vector<GlobalVariable *> &Globals) {
476 // Build a new global with the combined contents of the referenced globals.
477 std::vector<Constant *> GlobalInits;
478 const DataLayout &DL = M->getDataLayout();
479 for (GlobalVariable *G : Globals) {
480 GlobalInits.push_back(G->getInitializer());
481 uint64_t InitSize = DL.getTypeAllocSize(G->getInitializer()->getType());
483 // Compute the amount of padding required to align the next element to the
485 uint64_t Padding = NextPowerOf2(InitSize - 1) - InitSize;
487 // Cap at 128 was found experimentally to have a good data/instruction
488 // overhead tradeoff.
490 Padding = RoundUpToAlignment(InitSize, 128) - InitSize;
492 GlobalInits.push_back(
493 ConstantAggregateZero::get(ArrayType::get(Int8Ty, Padding)));
495 if (!GlobalInits.empty())
496 GlobalInits.pop_back();
497 Constant *NewInit = ConstantStruct::getAnon(M->getContext(), GlobalInits);
498 auto CombinedGlobal =
499 new GlobalVariable(*M, NewInit->getType(), /*isConstant=*/true,
500 GlobalValue::PrivateLinkage, NewInit);
502 const StructLayout *CombinedGlobalLayout =
503 DL.getStructLayout(cast<StructType>(NewInit->getType()));
505 // Compute the offsets of the original globals within the new global.
506 DenseMap<GlobalVariable *, uint64_t> GlobalLayout;
507 for (unsigned I = 0; I != Globals.size(); ++I)
508 // Multiply by 2 to account for padding elements.
509 GlobalLayout[Globals[I]] = CombinedGlobalLayout->getElementOffset(I * 2);
511 // For each bitset in this disjoint set...
512 for (MDString *BS : BitSets) {
514 BitSetInfo BSI = buildBitSet(BS, GlobalLayout);
516 ByteArrayInfo *BAI = 0;
518 // Lower each call to llvm.bitset.test for this bitset.
519 for (CallInst *CI : BitSetTestCallSites[BS]) {
520 ++NumBitSetCallsLowered;
521 Value *Lowered = lowerBitSetCall(CI, BSI, BAI, CombinedGlobal, GlobalLayout);
522 CI->replaceAllUsesWith(Lowered);
523 CI->eraseFromParent();
527 // Build aliases pointing to offsets into the combined global for each
528 // global from which we built the combined global, and replace references
529 // to the original globals with references to the aliases.
530 for (unsigned I = 0; I != Globals.size(); ++I) {
531 // Multiply by 2 to account for padding elements.
532 Constant *CombinedGlobalIdxs[] = {ConstantInt::get(Int32Ty, 0),
533 ConstantInt::get(Int32Ty, I * 2)};
534 Constant *CombinedGlobalElemPtr =
535 ConstantExpr::getGetElementPtr(CombinedGlobal, CombinedGlobalIdxs);
536 if (LinkerSubsectionsViaSymbols) {
537 Globals[I]->replaceAllUsesWith(CombinedGlobalElemPtr);
539 GlobalAlias *GAlias = GlobalAlias::create(
540 Globals[I]->getType()->getElementType(),
541 Globals[I]->getType()->getAddressSpace(), Globals[I]->getLinkage(),
542 "", CombinedGlobalElemPtr, M);
543 GAlias->takeName(Globals[I]);
544 Globals[I]->replaceAllUsesWith(GAlias);
546 Globals[I]->eraseFromParent();
550 /// Lower all bit sets in this module.
551 bool LowerBitSets::buildBitSets() {
552 Function *BitSetTestFunc =
553 M->getFunction(Intrinsic::getName(Intrinsic::bitset_test));
557 // Equivalence class set containing bitsets and the globals they reference.
558 // This is used to partition the set of bitsets in the module into disjoint
560 typedef EquivalenceClasses<PointerUnion<GlobalVariable *, MDString *>>
562 GlobalClassesTy GlobalClasses;
564 for (const Use &U : BitSetTestFunc->uses()) {
565 auto CI = cast<CallInst>(U.getUser());
567 auto BitSetMDVal = dyn_cast<MetadataAsValue>(CI->getArgOperand(1));
568 if (!BitSetMDVal || !isa<MDString>(BitSetMDVal->getMetadata()))
570 "Second argument of llvm.bitset.test must be metadata string");
571 auto BitSet = cast<MDString>(BitSetMDVal->getMetadata());
573 // Add the call site to the list of call sites for this bit set. We also use
574 // BitSetTestCallSites to keep track of whether we have seen this bit set
575 // before. If we have, we don't need to re-add the referenced globals to the
576 // equivalence class.
577 std::pair<DenseMap<MDString *, std::vector<CallInst *>>::iterator,
579 BitSetTestCallSites.insert(
580 std::make_pair(BitSet, std::vector<CallInst *>()));
581 Ins.first->second.push_back(CI);
585 // Add the bitset to the equivalence class.
586 GlobalClassesTy::iterator GCI = GlobalClasses.insert(BitSet);
587 GlobalClassesTy::member_iterator CurSet = GlobalClasses.findLeader(GCI);
592 // Verify the bitset metadata and add the referenced globals to the bitset's
593 // equivalence class.
594 for (MDNode *Op : BitSetNM->operands()) {
595 if (Op->getNumOperands() != 3)
597 "All operands of llvm.bitsets metadata must have 3 elements");
599 if (Op->getOperand(0) != BitSet || !Op->getOperand(1))
602 auto OpConstMD = dyn_cast<ConstantAsMetadata>(Op->getOperand(1));
604 report_fatal_error("Bit set element must be a constant");
605 auto OpGlobal = dyn_cast<GlobalVariable>(OpConstMD->getValue());
607 report_fatal_error("Bit set element must refer to global");
609 auto OffsetConstMD = dyn_cast<ConstantAsMetadata>(Op->getOperand(2));
611 report_fatal_error("Bit set element offset must be a constant");
612 auto OffsetInt = dyn_cast<ConstantInt>(OffsetConstMD->getValue());
615 "Bit set element offset must be an integer constant");
617 CurSet = GlobalClasses.unionSets(
618 CurSet, GlobalClasses.findLeader(GlobalClasses.insert(OpGlobal)));
622 if (GlobalClasses.empty())
625 // For each disjoint set we found...
626 for (GlobalClassesTy::iterator I = GlobalClasses.begin(),
627 E = GlobalClasses.end();
629 if (!I->isLeader()) continue;
631 ++NumBitSetDisjointSets;
633 // Build the list of bitsets and referenced globals in this disjoint set.
634 std::vector<MDString *> BitSets;
635 std::vector<GlobalVariable *> Globals;
636 llvm::DenseMap<MDString *, uint64_t> BitSetIndices;
637 llvm::DenseMap<GlobalVariable *, uint64_t> GlobalIndices;
638 for (GlobalClassesTy::member_iterator MI = GlobalClasses.member_begin(I);
639 MI != GlobalClasses.member_end(); ++MI) {
640 if ((*MI).is<MDString *>()) {
641 BitSetIndices[MI->get<MDString *>()] = BitSets.size();
642 BitSets.push_back(MI->get<MDString *>());
644 GlobalIndices[MI->get<GlobalVariable *>()] = Globals.size();
645 Globals.push_back(MI->get<GlobalVariable *>());
649 // For each bitset, build a set of indices that refer to globals referenced
651 std::vector<std::set<uint64_t>> BitSetMembers(BitSets.size());
653 for (MDNode *Op : BitSetNM->operands()) {
654 // Op = { bitset name, global, offset }
655 if (!Op->getOperand(1))
657 auto I = BitSetIndices.find(cast<MDString>(Op->getOperand(0)));
658 if (I == BitSetIndices.end())
661 auto OpGlobal = cast<GlobalVariable>(
662 cast<ConstantAsMetadata>(Op->getOperand(1))->getValue());
663 BitSetMembers[I->second].insert(GlobalIndices[OpGlobal]);
667 // Order the sets of indices by size. The GlobalLayoutBuilder works best
668 // when given small index sets first.
670 BitSetMembers.begin(), BitSetMembers.end(),
671 [](const std::set<uint64_t> &O1, const std::set<uint64_t> &O2) {
672 return O1.size() < O2.size();
675 // Create a GlobalLayoutBuilder and provide it with index sets as layout
676 // fragments. The GlobalLayoutBuilder tries to lay out members of fragments
677 // as close together as possible.
678 GlobalLayoutBuilder GLB(Globals.size());
679 for (auto &&MemSet : BitSetMembers)
680 GLB.addFragment(MemSet);
682 // Build a vector of globals with the computed layout.
683 std::vector<GlobalVariable *> OrderedGlobals(Globals.size());
684 auto OGI = OrderedGlobals.begin();
685 for (auto &&F : GLB.Fragments)
686 for (auto &&Offset : F)
687 *OGI++ = Globals[Offset];
689 // Order bitsets by name for determinism.
690 std::sort(BitSets.begin(), BitSets.end(), [](MDString *S1, MDString *S2) {
691 return S1->getString() < S2->getString();
694 // Build the bitsets from this disjoint set.
695 buildBitSetsFromGlobals(BitSets, OrderedGlobals);
698 allocateByteArrays();
703 bool LowerBitSets::eraseBitSetMetadata() {
707 M->eraseNamedMetadata(BitSetNM);
711 bool LowerBitSets::runOnModule(Module &M) {
712 bool Changed = buildBitSets();
713 Changed |= eraseBitSetMetadata();