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 static cl::opt<bool> AvoidReuse(
42 "lowerbitsets-avoid-reuse",
43 cl::desc("Try to avoid reuse of byte array addresses using aliases"),
44 cl::Hidden, cl::init(true));
46 bool BitSetInfo::containsGlobalOffset(uint64_t Offset) const {
47 if (Offset < ByteOffset)
50 if ((Offset - ByteOffset) % (uint64_t(1) << AlignLog2) != 0)
53 uint64_t BitOffset = (Offset - ByteOffset) >> AlignLog2;
54 if (BitOffset >= BitSize)
57 return Bits.count(BitOffset);
60 bool BitSetInfo::containsValue(
62 const DenseMap<GlobalVariable *, uint64_t> &GlobalLayout, Value *V,
63 uint64_t COffset) const {
64 if (auto GV = dyn_cast<GlobalVariable>(V)) {
65 auto I = GlobalLayout.find(GV);
66 if (I == GlobalLayout.end())
68 return containsGlobalOffset(I->second + COffset);
71 if (auto GEP = dyn_cast<GEPOperator>(V)) {
72 APInt APOffset(DL.getPointerSizeInBits(0), 0);
73 bool Result = GEP->accumulateConstantOffset(DL, APOffset);
76 COffset += APOffset.getZExtValue();
77 return containsValue(DL, GlobalLayout, GEP->getPointerOperand(),
81 if (auto Op = dyn_cast<Operator>(V)) {
82 if (Op->getOpcode() == Instruction::BitCast)
83 return containsValue(DL, GlobalLayout, Op->getOperand(0), COffset);
85 if (Op->getOpcode() == Instruction::Select)
86 return containsValue(DL, GlobalLayout, Op->getOperand(1), COffset) &&
87 containsValue(DL, GlobalLayout, Op->getOperand(2), COffset);
93 BitSetInfo BitSetBuilder::build() {
97 // Normalize each offset against the minimum observed offset, and compute
98 // the bitwise OR of each of the offsets. The number of trailing zeros
99 // in the mask gives us the log2 of the alignment of all offsets, which
100 // allows us to compress the bitset by only storing one bit per aligned
103 for (uint64_t &Offset : Offsets) {
109 BSI.ByteOffset = Min;
113 BSI.AlignLog2 = countTrailingZeros(Mask, ZB_Undefined);
115 // Build the compressed bitset while normalizing the offsets against the
116 // computed alignment.
117 BSI.BitSize = ((Max - Min) >> BSI.AlignLog2) + 1;
118 for (uint64_t Offset : Offsets) {
119 Offset >>= BSI.AlignLog2;
120 BSI.Bits.insert(Offset);
126 void GlobalLayoutBuilder::addFragment(const std::set<uint64_t> &F) {
127 // Create a new fragment to hold the layout for F.
128 Fragments.emplace_back();
129 std::vector<uint64_t> &Fragment = Fragments.back();
130 uint64_t FragmentIndex = Fragments.size() - 1;
132 for (auto ObjIndex : F) {
133 uint64_t OldFragmentIndex = FragmentMap[ObjIndex];
134 if (OldFragmentIndex == 0) {
135 // We haven't seen this object index before, so just add it to the current
137 Fragment.push_back(ObjIndex);
139 // This index belongs to an existing fragment. Copy the elements of the
140 // old fragment into this one and clear the old fragment. We don't update
141 // the fragment map just yet, this ensures that any further references to
142 // indices from the old fragment in this fragment do not insert any more
144 std::vector<uint64_t> &OldFragment = Fragments[OldFragmentIndex];
145 Fragment.insert(Fragment.end(), OldFragment.begin(), OldFragment.end());
150 // Update the fragment map to point our object indices to this fragment.
151 for (uint64_t ObjIndex : Fragment)
152 FragmentMap[ObjIndex] = FragmentIndex;
155 void ByteArrayBuilder::allocate(const std::set<uint64_t> &Bits,
156 uint64_t BitSize, uint64_t &AllocByteOffset,
157 uint8_t &AllocMask) {
158 // Find the smallest current allocation.
160 for (unsigned I = 1; I != BitsPerByte; ++I)
161 if (BitAllocs[I] < BitAllocs[Bit])
164 AllocByteOffset = BitAllocs[Bit];
166 // Add our size to it.
167 unsigned ReqSize = AllocByteOffset + BitSize;
168 BitAllocs[Bit] = ReqSize;
169 if (Bytes.size() < ReqSize)
170 Bytes.resize(ReqSize);
173 AllocMask = 1 << Bit;
174 for (uint64_t B : Bits)
175 Bytes[AllocByteOffset + B] |= AllocMask;
180 struct ByteArrayInfo {
181 std::set<uint64_t> Bits;
183 GlobalVariable *ByteArray;
187 struct LowerBitSets : public ModulePass {
189 LowerBitSets() : ModulePass(ID) {
190 initializeLowerBitSetsPass(*PassRegistry::getPassRegistry());
195 bool LinkerSubsectionsViaSymbols;
198 IntegerType *Int32Ty;
200 IntegerType *Int64Ty;
203 // The llvm.bitsets named metadata.
204 NamedMDNode *BitSetNM;
206 // Mapping from bitset mdstrings to the call sites that test them.
207 DenseMap<MDString *, std::vector<CallInst *>> BitSetTestCallSites;
209 std::vector<ByteArrayInfo> ByteArrayInfos;
212 buildBitSet(MDString *BitSet,
213 const DenseMap<GlobalVariable *, uint64_t> &GlobalLayout);
214 ByteArrayInfo *createByteArray(BitSetInfo &BSI);
215 void allocateByteArrays();
216 Value *createBitSetTest(IRBuilder<> &B, BitSetInfo &BSI, ByteArrayInfo *&BAI,
219 lowerBitSetCall(CallInst *CI, BitSetInfo &BSI, ByteArrayInfo *&BAI,
220 GlobalVariable *CombinedGlobal,
221 const DenseMap<GlobalVariable *, uint64_t> &GlobalLayout);
222 void buildBitSetsFromGlobals(const std::vector<MDString *> &BitSets,
223 const std::vector<GlobalVariable *> &Globals);
225 bool eraseBitSetMetadata();
227 bool doInitialization(Module &M) override;
228 bool runOnModule(Module &M) override;
233 INITIALIZE_PASS_BEGIN(LowerBitSets, "lowerbitsets",
234 "Lower bitset metadata", false, false)
235 INITIALIZE_PASS_END(LowerBitSets, "lowerbitsets",
236 "Lower bitset metadata", false, false)
237 char LowerBitSets::ID = 0;
239 ModulePass *llvm::createLowerBitSetsPass() { return new LowerBitSets; }
241 bool LowerBitSets::doInitialization(Module &Mod) {
243 const DataLayout &DL = Mod.getDataLayout();
245 Triple TargetTriple(M->getTargetTriple());
246 LinkerSubsectionsViaSymbols = TargetTriple.isMacOSX();
248 Int1Ty = Type::getInt1Ty(M->getContext());
249 Int8Ty = Type::getInt8Ty(M->getContext());
250 Int32Ty = Type::getInt32Ty(M->getContext());
251 Int32PtrTy = PointerType::getUnqual(Int32Ty);
252 Int64Ty = Type::getInt64Ty(M->getContext());
253 IntPtrTy = DL.getIntPtrType(M->getContext(), 0);
255 BitSetNM = M->getNamedMetadata("llvm.bitsets");
257 BitSetTestCallSites.clear();
262 /// Build a bit set for BitSet using the object layouts in
264 BitSetInfo LowerBitSets::buildBitSet(
266 const DenseMap<GlobalVariable *, uint64_t> &GlobalLayout) {
269 // Compute the byte offset of each element of this bitset.
271 for (MDNode *Op : BitSetNM->operands()) {
272 if (Op->getOperand(0) != BitSet || !Op->getOperand(1))
274 auto OpGlobal = cast<GlobalVariable>(
275 cast<ConstantAsMetadata>(Op->getOperand(1))->getValue());
277 cast<ConstantInt>(cast<ConstantAsMetadata>(Op->getOperand(2))
278 ->getValue())->getZExtValue();
280 Offset += GlobalLayout.find(OpGlobal)->second;
282 BSB.addOffset(Offset);
289 /// Build a test that bit BitOffset mod sizeof(Bits)*8 is set in
290 /// Bits. This pattern matches to the bt instruction on x86.
291 static Value *createMaskedBitTest(IRBuilder<> &B, Value *Bits,
293 auto BitsType = cast<IntegerType>(Bits->getType());
294 unsigned BitWidth = BitsType->getBitWidth();
296 BitOffset = B.CreateZExtOrTrunc(BitOffset, BitsType);
298 B.CreateAnd(BitOffset, ConstantInt::get(BitsType, BitWidth - 1));
299 Value *BitMask = B.CreateShl(ConstantInt::get(BitsType, 1), BitIndex);
300 Value *MaskedBits = B.CreateAnd(Bits, BitMask);
301 return B.CreateICmpNE(MaskedBits, ConstantInt::get(BitsType, 0));
304 ByteArrayInfo *LowerBitSets::createByteArray(BitSetInfo &BSI) {
305 // Create globals to stand in for byte arrays and masks. These never actually
306 // get initialized, we RAUW and erase them later in allocateByteArrays() once
307 // we know the offset and mask to use.
308 auto ByteArrayGlobal = new GlobalVariable(
309 *M, Int8Ty, /*isConstant=*/true, GlobalValue::PrivateLinkage, nullptr);
310 auto MaskGlobal = new GlobalVariable(
311 *M, Int8Ty, /*isConstant=*/true, GlobalValue::PrivateLinkage, nullptr);
313 ByteArrayInfos.emplace_back();
314 ByteArrayInfo *BAI = &ByteArrayInfos.back();
316 BAI->Bits = BSI.Bits;
317 BAI->BitSize = BSI.BitSize;
318 BAI->ByteArray = ByteArrayGlobal;
319 BAI->Mask = ConstantExpr::getPtrToInt(MaskGlobal, Int8Ty);
323 void LowerBitSets::allocateByteArrays() {
324 std::stable_sort(ByteArrayInfos.begin(), ByteArrayInfos.end(),
325 [](const ByteArrayInfo &BAI1, const ByteArrayInfo &BAI2) {
326 return BAI1.BitSize > BAI2.BitSize;
329 std::vector<uint64_t> ByteArrayOffsets(ByteArrayInfos.size());
331 ByteArrayBuilder BAB;
332 for (unsigned I = 0; I != ByteArrayInfos.size(); ++I) {
333 ByteArrayInfo *BAI = &ByteArrayInfos[I];
336 BAB.allocate(BAI->Bits, BAI->BitSize, ByteArrayOffsets[I], Mask);
338 BAI->Mask->replaceAllUsesWith(ConstantInt::get(Int8Ty, Mask));
339 cast<GlobalVariable>(BAI->Mask->getOperand(0))->eraseFromParent();
342 Constant *ByteArrayConst = ConstantDataArray::get(M->getContext(), BAB.Bytes);
344 new GlobalVariable(*M, ByteArrayConst->getType(), /*isConstant=*/true,
345 GlobalValue::PrivateLinkage, ByteArrayConst);
347 for (unsigned I = 0; I != ByteArrayInfos.size(); ++I) {
348 ByteArrayInfo *BAI = &ByteArrayInfos[I];
350 Constant *Idxs[] = {ConstantInt::get(IntPtrTy, 0),
351 ConstantInt::get(IntPtrTy, ByteArrayOffsets[I])};
352 Constant *GEP = ConstantExpr::getInBoundsGetElementPtr(
353 ByteArrayConst->getType(), ByteArray, Idxs);
355 // Create an alias instead of RAUW'ing the gep directly. On x86 this ensures
356 // that the pc-relative displacement is folded into the lea instead of the
357 // test instruction getting another displacement.
358 if (LinkerSubsectionsViaSymbols) {
359 BAI->ByteArray->replaceAllUsesWith(GEP);
361 GlobalAlias *Alias = GlobalAlias::create(
362 Int8Ty, 0, GlobalValue::PrivateLinkage, "bits", GEP, M);
363 BAI->ByteArray->replaceAllUsesWith(Alias);
365 BAI->ByteArray->eraseFromParent();
368 ByteArraySizeBits = BAB.BitAllocs[0] + BAB.BitAllocs[1] + BAB.BitAllocs[2] +
369 BAB.BitAllocs[3] + BAB.BitAllocs[4] + BAB.BitAllocs[5] +
370 BAB.BitAllocs[6] + BAB.BitAllocs[7];
371 ByteArraySizeBytes = BAB.Bytes.size();
374 /// Build a test that bit BitOffset is set in BSI, where
375 /// BitSetGlobal is a global containing the bits in BSI.
376 Value *LowerBitSets::createBitSetTest(IRBuilder<> &B, BitSetInfo &BSI,
377 ByteArrayInfo *&BAI, Value *BitOffset) {
378 if (BSI.BitSize <= 64) {
379 // If the bit set is sufficiently small, we can avoid a load by bit testing
382 if (BSI.BitSize <= 32)
388 for (auto Bit : BSI.Bits)
389 Bits |= uint64_t(1) << Bit;
390 Constant *BitsConst = ConstantInt::get(BitsTy, Bits);
391 return createMaskedBitTest(B, BitsConst, BitOffset);
394 ++NumByteArraysCreated;
395 BAI = createByteArray(BSI);
398 Constant *ByteArray = BAI->ByteArray;
399 if (!LinkerSubsectionsViaSymbols && AvoidReuse) {
400 // Each use of the byte array uses a different alias. This makes the
401 // backend less likely to reuse previously computed byte array addresses,
402 // improving the security of the CFI mechanism based on this pass.
403 ByteArray = GlobalAlias::create(
404 BAI->ByteArray->getType()->getElementType(), 0,
405 GlobalValue::PrivateLinkage, "bits_use", ByteArray, M);
408 Value *ByteAddr = B.CreateGEP(ByteArray, BitOffset);
409 Value *Byte = B.CreateLoad(ByteAddr);
411 Value *ByteAndMask = B.CreateAnd(Byte, BAI->Mask);
412 return B.CreateICmpNE(ByteAndMask, ConstantInt::get(Int8Ty, 0));
416 /// Lower a llvm.bitset.test call to its implementation. Returns the value to
417 /// replace the call with.
418 Value *LowerBitSets::lowerBitSetCall(
419 CallInst *CI, BitSetInfo &BSI, ByteArrayInfo *&BAI,
420 GlobalVariable *CombinedGlobal,
421 const DenseMap<GlobalVariable *, uint64_t> &GlobalLayout) {
422 Value *Ptr = CI->getArgOperand(0);
423 const DataLayout &DL = M->getDataLayout();
425 if (BSI.containsValue(DL, GlobalLayout, Ptr))
426 return ConstantInt::getTrue(CombinedGlobal->getParent()->getContext());
428 Constant *GlobalAsInt = ConstantExpr::getPtrToInt(CombinedGlobal, IntPtrTy);
429 Constant *OffsetedGlobalAsInt = ConstantExpr::getAdd(
430 GlobalAsInt, ConstantInt::get(IntPtrTy, BSI.ByteOffset));
432 BasicBlock *InitialBB = CI->getParent();
436 Value *PtrAsInt = B.CreatePtrToInt(Ptr, IntPtrTy);
438 if (BSI.isSingleOffset())
439 return B.CreateICmpEQ(PtrAsInt, OffsetedGlobalAsInt);
441 Value *PtrOffset = B.CreateSub(PtrAsInt, OffsetedGlobalAsInt);
444 if (BSI.AlignLog2 == 0) {
445 BitOffset = PtrOffset;
447 // We need to check that the offset both falls within our range and is
448 // suitably aligned. We can check both properties at the same time by
449 // performing a right rotate by log2(alignment) followed by an integer
450 // comparison against the bitset size. The rotate will move the lower
451 // order bits that need to be zero into the higher order bits of the
452 // result, causing the comparison to fail if they are nonzero. The rotate
453 // also conveniently gives us a bit offset to use during the load from
456 B.CreateLShr(PtrOffset, ConstantInt::get(IntPtrTy, BSI.AlignLog2));
457 Value *OffsetSHL = B.CreateShl(
459 ConstantInt::get(IntPtrTy, DL.getPointerSizeInBits(0) - BSI.AlignLog2));
460 BitOffset = B.CreateOr(OffsetSHR, OffsetSHL);
463 Constant *BitSizeConst = ConstantInt::get(IntPtrTy, BSI.BitSize);
464 Value *OffsetInRange = B.CreateICmpULT(BitOffset, BitSizeConst);
466 // If the bit set is all ones, testing against it is unnecessary.
468 return OffsetInRange;
470 TerminatorInst *Term = SplitBlockAndInsertIfThen(OffsetInRange, CI, false);
471 IRBuilder<> ThenB(Term);
473 // Now that we know that the offset is in range and aligned, load the
474 // appropriate bit from the bitset.
475 Value *Bit = createBitSetTest(ThenB, BSI, BAI, BitOffset);
477 // The value we want is 0 if we came directly from the initial block
478 // (having failed the range or alignment checks), or the loaded bit if
479 // we came from the block in which we loaded it.
480 B.SetInsertPoint(CI);
481 PHINode *P = B.CreatePHI(Int1Ty, 2);
482 P->addIncoming(ConstantInt::get(Int1Ty, 0), InitialBB);
483 P->addIncoming(Bit, ThenB.GetInsertBlock());
487 /// Given a disjoint set of bitsets and globals, layout the globals, build the
488 /// bit sets and lower the llvm.bitset.test calls.
489 void LowerBitSets::buildBitSetsFromGlobals(
490 const std::vector<MDString *> &BitSets,
491 const std::vector<GlobalVariable *> &Globals) {
492 // Build a new global with the combined contents of the referenced globals.
493 std::vector<Constant *> GlobalInits;
494 const DataLayout &DL = M->getDataLayout();
495 for (GlobalVariable *G : Globals) {
496 GlobalInits.push_back(G->getInitializer());
497 uint64_t InitSize = DL.getTypeAllocSize(G->getInitializer()->getType());
499 // Compute the amount of padding required to align the next element to the
501 uint64_t Padding = NextPowerOf2(InitSize - 1) - InitSize;
503 // Cap at 128 was found experimentally to have a good data/instruction
504 // overhead tradeoff.
506 Padding = RoundUpToAlignment(InitSize, 128) - InitSize;
508 GlobalInits.push_back(
509 ConstantAggregateZero::get(ArrayType::get(Int8Ty, Padding)));
511 if (!GlobalInits.empty())
512 GlobalInits.pop_back();
513 Constant *NewInit = ConstantStruct::getAnon(M->getContext(), GlobalInits);
514 auto CombinedGlobal =
515 new GlobalVariable(*M, NewInit->getType(), /*isConstant=*/true,
516 GlobalValue::PrivateLinkage, NewInit);
518 const StructLayout *CombinedGlobalLayout =
519 DL.getStructLayout(cast<StructType>(NewInit->getType()));
521 // Compute the offsets of the original globals within the new global.
522 DenseMap<GlobalVariable *, uint64_t> GlobalLayout;
523 for (unsigned I = 0; I != Globals.size(); ++I)
524 // Multiply by 2 to account for padding elements.
525 GlobalLayout[Globals[I]] = CombinedGlobalLayout->getElementOffset(I * 2);
527 // For each bitset in this disjoint set...
528 for (MDString *BS : BitSets) {
530 BitSetInfo BSI = buildBitSet(BS, GlobalLayout);
532 ByteArrayInfo *BAI = 0;
534 // Lower each call to llvm.bitset.test for this bitset.
535 for (CallInst *CI : BitSetTestCallSites[BS]) {
536 ++NumBitSetCallsLowered;
537 Value *Lowered = lowerBitSetCall(CI, BSI, BAI, CombinedGlobal, GlobalLayout);
538 CI->replaceAllUsesWith(Lowered);
539 CI->eraseFromParent();
543 // Build aliases pointing to offsets into the combined global for each
544 // global from which we built the combined global, and replace references
545 // to the original globals with references to the aliases.
546 for (unsigned I = 0; I != Globals.size(); ++I) {
547 // Multiply by 2 to account for padding elements.
548 Constant *CombinedGlobalIdxs[] = {ConstantInt::get(Int32Ty, 0),
549 ConstantInt::get(Int32Ty, I * 2)};
550 Constant *CombinedGlobalElemPtr = ConstantExpr::getGetElementPtr(
551 NewInit->getType(), CombinedGlobal, CombinedGlobalIdxs);
552 if (LinkerSubsectionsViaSymbols) {
553 Globals[I]->replaceAllUsesWith(CombinedGlobalElemPtr);
555 GlobalAlias *GAlias = GlobalAlias::create(
556 Globals[I]->getType()->getElementType(),
557 Globals[I]->getType()->getAddressSpace(), Globals[I]->getLinkage(),
558 "", CombinedGlobalElemPtr, M);
559 GAlias->takeName(Globals[I]);
560 Globals[I]->replaceAllUsesWith(GAlias);
562 Globals[I]->eraseFromParent();
566 /// Lower all bit sets in this module.
567 bool LowerBitSets::buildBitSets() {
568 Function *BitSetTestFunc =
569 M->getFunction(Intrinsic::getName(Intrinsic::bitset_test));
573 // Equivalence class set containing bitsets and the globals they reference.
574 // This is used to partition the set of bitsets in the module into disjoint
576 typedef EquivalenceClasses<PointerUnion<GlobalVariable *, MDString *>>
578 GlobalClassesTy GlobalClasses;
580 for (const Use &U : BitSetTestFunc->uses()) {
581 auto CI = cast<CallInst>(U.getUser());
583 auto BitSetMDVal = dyn_cast<MetadataAsValue>(CI->getArgOperand(1));
584 if (!BitSetMDVal || !isa<MDString>(BitSetMDVal->getMetadata()))
586 "Second argument of llvm.bitset.test must be metadata string");
587 auto BitSet = cast<MDString>(BitSetMDVal->getMetadata());
589 // Add the call site to the list of call sites for this bit set. We also use
590 // BitSetTestCallSites to keep track of whether we have seen this bit set
591 // before. If we have, we don't need to re-add the referenced globals to the
592 // equivalence class.
593 std::pair<DenseMap<MDString *, std::vector<CallInst *>>::iterator,
595 BitSetTestCallSites.insert(
596 std::make_pair(BitSet, std::vector<CallInst *>()));
597 Ins.first->second.push_back(CI);
601 // Add the bitset to the equivalence class.
602 GlobalClassesTy::iterator GCI = GlobalClasses.insert(BitSet);
603 GlobalClassesTy::member_iterator CurSet = GlobalClasses.findLeader(GCI);
608 // Verify the bitset metadata and add the referenced globals to the bitset's
609 // equivalence class.
610 for (MDNode *Op : BitSetNM->operands()) {
611 if (Op->getNumOperands() != 3)
613 "All operands of llvm.bitsets metadata must have 3 elements");
615 if (Op->getOperand(0) != BitSet || !Op->getOperand(1))
618 auto OpConstMD = dyn_cast<ConstantAsMetadata>(Op->getOperand(1));
620 report_fatal_error("Bit set element must be a constant");
621 auto OpGlobal = dyn_cast<GlobalVariable>(OpConstMD->getValue());
623 report_fatal_error("Bit set element must refer to global");
625 auto OffsetConstMD = dyn_cast<ConstantAsMetadata>(Op->getOperand(2));
627 report_fatal_error("Bit set element offset must be a constant");
628 auto OffsetInt = dyn_cast<ConstantInt>(OffsetConstMD->getValue());
631 "Bit set element offset must be an integer constant");
633 CurSet = GlobalClasses.unionSets(
634 CurSet, GlobalClasses.findLeader(GlobalClasses.insert(OpGlobal)));
638 if (GlobalClasses.empty())
641 // For each disjoint set we found...
642 for (GlobalClassesTy::iterator I = GlobalClasses.begin(),
643 E = GlobalClasses.end();
645 if (!I->isLeader()) continue;
647 ++NumBitSetDisjointSets;
649 // Build the list of bitsets and referenced globals in this disjoint set.
650 std::vector<MDString *> BitSets;
651 std::vector<GlobalVariable *> Globals;
652 llvm::DenseMap<MDString *, uint64_t> BitSetIndices;
653 llvm::DenseMap<GlobalVariable *, uint64_t> GlobalIndices;
654 for (GlobalClassesTy::member_iterator MI = GlobalClasses.member_begin(I);
655 MI != GlobalClasses.member_end(); ++MI) {
656 if ((*MI).is<MDString *>()) {
657 BitSetIndices[MI->get<MDString *>()] = BitSets.size();
658 BitSets.push_back(MI->get<MDString *>());
660 GlobalIndices[MI->get<GlobalVariable *>()] = Globals.size();
661 Globals.push_back(MI->get<GlobalVariable *>());
665 // For each bitset, build a set of indices that refer to globals referenced
667 std::vector<std::set<uint64_t>> BitSetMembers(BitSets.size());
669 for (MDNode *Op : BitSetNM->operands()) {
670 // Op = { bitset name, global, offset }
671 if (!Op->getOperand(1))
673 auto I = BitSetIndices.find(cast<MDString>(Op->getOperand(0)));
674 if (I == BitSetIndices.end())
677 auto OpGlobal = cast<GlobalVariable>(
678 cast<ConstantAsMetadata>(Op->getOperand(1))->getValue());
679 BitSetMembers[I->second].insert(GlobalIndices[OpGlobal]);
683 // Order the sets of indices by size. The GlobalLayoutBuilder works best
684 // when given small index sets first.
686 BitSetMembers.begin(), BitSetMembers.end(),
687 [](const std::set<uint64_t> &O1, const std::set<uint64_t> &O2) {
688 return O1.size() < O2.size();
691 // Create a GlobalLayoutBuilder and provide it with index sets as layout
692 // fragments. The GlobalLayoutBuilder tries to lay out members of fragments
693 // as close together as possible.
694 GlobalLayoutBuilder GLB(Globals.size());
695 for (auto &&MemSet : BitSetMembers)
696 GLB.addFragment(MemSet);
698 // Build a vector of globals with the computed layout.
699 std::vector<GlobalVariable *> OrderedGlobals(Globals.size());
700 auto OGI = OrderedGlobals.begin();
701 for (auto &&F : GLB.Fragments)
702 for (auto &&Offset : F)
703 *OGI++ = Globals[Offset];
705 // Order bitsets by name for determinism.
706 std::sort(BitSets.begin(), BitSets.end(), [](MDString *S1, MDString *S2) {
707 return S1->getString() < S2->getString();
710 // Build the bitsets from this disjoint set.
711 buildBitSetsFromGlobals(BitSets, OrderedGlobals);
714 allocateByteArrays();
719 bool LowerBitSets::eraseBitSetMetadata() {
723 M->eraseNamedMetadata(BitSetNM);
727 bool LowerBitSets::runOnModule(Module &M) {
728 bool Changed = buildBitSets();
729 Changed |= eraseBitSetMetadata();