LowerBitSets: Use byte arrays instead of bit sets to represent in-memory bit sets.

[pcc]

Summary

By loading from indexed offsets into a byte array and applying a mask, a
program can test bits from the bit set with a relatively short instruction
sequence. For example, suppose we have 15 bit sets to lay out:

A (16 bits), B (15 bits), C (14 bits), D (13 bits), E (12 bits),
F (11 bits), G (10 bits), H (9 bits), I (7 bits), J (6 bits), K (5 bits),
L (4 bits), M (3 bits), N (2 bits), O (1 bit)

These bits can be laid out in a 16-byte array like this:

  Byte Offset
0123456789ABCDEF

Bit

7 HHHHHHHHHIIIIIII
6 GGGGGGGGGGJJJJJJ
5 FFFFFFFFFFFKKKKK
4 EEEEEEEEEEEELLLL
3 DDDDDDDDDDDDDMMM
2 CCCCCCCCCCCCCCNN
1 BBBBBBBBBBBBBBBO
0 AAAAAAAAAAAAAAAA

For example, to test bit X of A, we evaluate ((bits[X] & 1) != 0), or to
test bit X of I, we evaluate ((bits[9 + X] & 0x80) != 0). This can be done
in 1-2 machine instructions on x86, or 4-6 instructions on ARM.

This uses the LPT multiprocessor scheduling algorithm to lay out the bits
efficiently.

Saves ~450KB of instructions in a recent build of Chromium.

[kcc]

Saves ~450KB of instructions in a recent build of Chromium.

Out of how many?

[kcc]

Saves ~450KB of instructions in a recent build of Chromium.

Out of how many?

[pcc]

The current (i.e. after this patch) .text overhead is around 6.2MB, so around 7%.

[kcc]

include/llvm/Transforms/IPO/LowerBitSets.h

34

Is this better than unordered_set?
(I don't know, just asking)

183

LPT == longest process time?
Maybe mention the full name here?

lib/Transforms/IPO/LowerBitSets.cpp

34–36

It would be interesting to see more details stats here,
e.g. how many bits were packed into how many bytes.

[jfb]

include/llvm/Transforms/IPO/LowerBitSets.h

174

Could you add a symbolic constant for 8? Also, why 8? Is it good for x86 but now as good for ARM?

178

memset or std::fill.

lib/Transforms/IPO/LowerBitSets.cpp

34–36

+1

How many bits are unused (didn't fit) would be interesting.

154

Update 8.

300

Capitalize RAUW.

367

Just to be sure I understand: this inlines small (<= 64 bit) bitsets. The current patch only benefits from co-locating larger bitsets, right? It would probably be good for the docs to specify this double approach to reducing bitset footprint.

[pcc]

Address reviewer comments.

include/llvm/Transforms/IPO/LowerBitSets.h

34

Maybe, not sure though. DenseSet may be even better here. But I'd rather wait until this pass isn't a small fraction of the overall compile time before trying to optimize our data structures.

174

Are you asking why I am using 8 bits per array element? Well, for one thing it gives us better packing (the more bins there are, the less evenly they will be filled). For another, the instruction sequences can be slightly shorter, both on x86 and ARM. For example, this code:

typedef unsigned long T;
typedef void (*F)(void);
T table;

unsigned char bitset[1000];
void FOO(T **obj) {
  T *vptr = *obj;
  T ptr_offset = (T)vptr - (T)table;
  T t1 = ptr_offset << (sizeof(void*)*8 - 3);
  T t2 = ptr_offset >> 3;
  T t3 = t1 | t2;
  unsigned bits = bitset[t3];
  if ((t3 >= TABLE_SIZE) || (bits & 1))
    __builtin_trap();
  F f = (F)vptr[4];
  f();
}

is one byte shorter (on x86) or one instruction shorter (on ARM) than this code:

typedef unsigned long T;
typedef void (*F)(void);
T table;

unsigned short bitset[1000];
void FOO(T **obj) {
  T *vptr = *obj;
  T ptr_offset = (T)vptr - (T)table;
  T t1 = ptr_offset << (sizeof(void*)*8 - 3);
  T t2 = ptr_offset >> 3;
  T t3 = t1 | t2;
  unsigned bits = bitset[t3];
  if ((t3 >= TABLE_SIZE) || (bits & 0x100))
    __builtin_trap();
  F f = (F)vptr[4];
  f();
}

I guess I could parameterise this but it seems like it would add more complexity than we need. Happy to document the choice of 8 bits per array element.

178

Done

183

Done

lib/Transforms/IPO/LowerBitSets.cpp

34–36

I've added some more stats here.

154

See my other comment.

300

Done

367

We already document the technique for inlining <=64 bits:

http://clang.llvm.org/docs/ControlFlowIntegrityDesign.html#short-inline-bit-vectors

and I do plan to document byte arrays on that page once this lands.

[jfb]

include/llvm/Transforms/IPO/LowerBitSets.h

174

Yes, documentation for the choice of 8 is good :)

lib/Transforms/IPO/LowerBitSets.cpp

367

sgtm

[pcc]

• Document choice of 1 byte per array element

include/llvm/Transforms/IPO/LowerBitSets.h

174

Done

[jfb]

include/llvm/Transforms/IPO/LowerBitSets.h

179

I'd still rather have to 8 pulled out as its own symbolic constant, instead of being used here and below as a magic number.

[pcc]

• Use a constant for the number of bits per byte

include/llvm/Transforms/IPO/LowerBitSets.h

179

done

[jfb]

lgtm

include/llvm/Transforms/IPO/LowerBitSets.h

 
 namespace llvm {
 
 class DataLayout;
 class GlobalVariable;
 class Value;
 
 struct BitSetInfo {
-  // The actual bitset.
-  std::vector<uint8_t> Bits;
+  // The indices of the set bits in the bitset.
+  std::set<uint64_t> Bits;
 
   // The byte offset into the combined global represented by the bitset.
   uint64_t ByteOffset;
 
   // The size of the bitset in bits.
   uint64_t BitSize;
 
   // Log2 alignment of the bit set relative to the combined global.
   // For example, a log2 alignment of 3 means that bits in the bitset
   // represent addresses 8 bytes apart.
   unsigned AlignLog2;
 
   bool isSingleOffset() const {
-    return Bits.size() == 1 && Bits[0] == 1;
+    return Bits.size() == 1;
   }
 
   bool isAllOnes() const {
-    for (unsigned I = 0; I != Bits.size() - 1; ++I)
-      if (Bits[I] != 0xFF)
-        return false;
-
-    if (BitSize % 8 == 0)
-      return Bits[Bits.size() - 1] == 0xFF;
-
-    return Bits[Bits.size() - 1] == (1 << (BitSize % 8)) - 1;
+    return Bits.size() == BitSize;
   }
 
   bool containsGlobalOffset(uint64_t Offset) const;
 
   bool containsValue(const DataLayout *DL,
                      const DenseMap<GlobalVariable *, uint64_t> &GlobalLayout,
                      Value *V, uint64_t COffset = 0) const;
-
 };
 
 struct BitSetBuilder {
   SmallVector<uint64_t, 16> Offsets;
   uint64_t Min, Max;
 
   BitSetBuilder() : Min(std::numeric_limits<uint64_t>::max()), Max(0) {}
 
GlobalLayoutBuilder(uint64_t NumObjects)
       : Fragments(1), FragmentMap(NumObjects) {}
 
   /// Add F to the layout while trying to keep its indices contiguous.
   /// If a previously seen fragment uses any of F's indices, that
   /// fragment will be laid out inside F.
   void addFragment(const std::set<uint64_t> &F);
 };
 
+/// This class is used to build a byte array containing overlapping bit sets. By
+/// loading from indexed offsets into the byte array and applying a mask, a
+/// program can test bits from the bit set with a relatively short instruction
+/// sequence. For example, suppose we have 15 bit sets to lay out:
+///
+/// A (16 bits), B (15 bits), C (14 bits), D (13 bits), E (12 bits),
+/// F (11 bits), G (10 bits), H (9 bits), I (7 bits), J (6 bits), K (5 bits),
+/// L (4 bits), M (3 bits), N (2 bits), O (1 bit)
+///
+/// These bits can be laid out in a 16-byte array like this:
+///
+///       Byte Offset
+///     0123456789ABCDEF
+/// Bit
+///   7 HHHHHHHHHIIIIIII
+///   6 GGGGGGGGGGJJJJJJ
+///   5 FFFFFFFFFFFKKKKK
+///   4 EEEEEEEEEEEELLLL
+///   3 DDDDDDDDDDDDDMMM
+///   2 CCCCCCCCCCCCCCNN
+///   1 BBBBBBBBBBBBBBBO
+///   0 AAAAAAAAAAAAAAAA
+///
+/// For example, to test bit X of A, we evaluate ((bits[X] & 1) != 0), or to
+/// test bit X of I, we evaluate ((bits[9 + X] & 0x80) != 0). This can be done
+/// in 1-2 machine instructions on x86, or 4-6 instructions on ARM.
+struct ByteArrayBuilder {
+  /// The byte array built so far.
+  std::vector<uint8_t> Bytes;
+
+  /// The number of bytes allocated so far for each of the bits.
+  uint64_t BitAllocs[8];
+
+  ByteArrayBuilder() {
+    memset(BitAllocs, 0, sizeof(BitAllocs));
+  }
+
+  /// Allocate BitSize bits in the byte array where Bits contains the bits to
+  /// set. AllocByteOffset is set to the offset within the byte array and
+  /// AllocMask is set to the bitmask for those bits. This uses the LPT (Longest
+  /// Processing Time) multiprocessor scheduling algorithm to lay out the bits
+  /// efficiently; the pass allocates bit sets in decreasing size order.
+  void allocate(const std::set<uint64_t> &Bits, uint64_t BitSize,
+                uint64_t &AllocByteOffset, uint8_t &AllocMask);
+};
+
 } // namespace llvm
 
 #endif

lib/Transforms/IPO/LowerBitSets.cpp

 #include "llvm/IR/Operator.h"
 #include "llvm/Pass.h"
 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
 
 using namespace llvm;
 
 #define DEBUG_TYPE "lowerbitsets"
 
-STATISTIC(NumBitSetsCreated, "Number of bitsets created");
+STATISTIC(ByteArraySizeBits, "Byte array size in bits");
+STATISTIC(ByteArraySizeBytes, "Byte array size in bytes");
+STATISTIC(NumByteArraysCreated, "Number of byte arrays created");
 STATISTIC(NumBitSetCallsLowered, "Number of bitset calls lowered");
 STATISTIC(NumBitSetDisjointSets, "Number of disjoint sets of bitsets");
 
 bool BitSetInfo::containsGlobalOffset(uint64_t Offset) const {
   if (Offset < ByteOffset)
     return false;
 
   if ((Offset - ByteOffset) % (uint64_t(1) << AlignLog2) != 0)
     return false;
 
   uint64_t BitOffset = (Offset - ByteOffset) >> AlignLog2;
   if (BitOffset >= BitSize)
     return false;
 
-  return (Bits[BitOffset / 8] >> (BitOffset % 8)) & 1;
+  return Bits.count(BitOffset);
 }
 
 bool BitSetInfo::containsValue(
     const DataLayout *DL,
     const DenseMap<GlobalVariable *, uint64_t> &GlobalLayout, Value *V,
     uint64_t COffset) const {
   if (auto GV = dyn_cast<GlobalVariable>(V)) {
     auto I = GlobalLayout.find(GV);
for (uint64_t &Offset : Offsets) {
     Offset -= Min;
     Mask |= Offset;
   }
 
   BitSetInfo BSI;
   BSI.ByteOffset = Min;
 
   BSI.AlignLog2 = 0;
-  // FIXME: Can probably do something smarter if all offsets are 0.
   if (Mask != 0)
     BSI.AlignLog2 = countTrailingZeros(Mask, ZB_Undefined);
 
   // Build the compressed bitset while normalizing the offsets against the
   // computed alignment.
   BSI.BitSize = ((Max - Min) >> BSI.AlignLog2) + 1;
-  uint64_t ByteSize = (BSI.BitSize + 7) / 8;
-  BSI.Bits.resize(ByteSize);
   for (uint64_t Offset : Offsets) {
     Offset >>= BSI.AlignLog2;
-    BSI.Bits[Offset / 8] |= 1 << (Offset % 8);
+    BSI.Bits.insert(Offset);
   }
 
   return BSI;
 }
 
 void GlobalLayoutBuilder::addFragment(const std::set<uint64_t> &F) {
   // Create a new fragment to hold the layout for F.
   Fragments.emplace_back();
for (auto ObjIndex : F) {
     }
   }
 
   // Update the fragment map to point our object indices to this fragment.
   for (uint64_t ObjIndex : Fragment)
     FragmentMap[ObjIndex] = FragmentIndex;
 }
 
+void ByteArrayBuilder::allocate(const std::set<uint64_t> &Bits,
+                                uint64_t BitSize, uint64_t &AllocByteOffset,
+                                uint8_t &AllocMask) {
+  // Find the smallest current allocation.
+  unsigned Bit = 0;
+  for (unsigned I = 1; I != 8; ++I)
+    if (BitAllocs[I] < BitAllocs[Bit])
+      Bit = I;
+
+  AllocByteOffset = BitAllocs[Bit];
+
+  // Add our size to it.
+  unsigned ReqSize = AllocByteOffset + BitSize;
+  BitAllocs[Bit] = ReqSize;
+  if (Bytes.size() < ReqSize)
+    Bytes.resize(ReqSize);
+
+  // Set our bits.
+  AllocMask = 1 << Bit;
+  for (uint64_t B : Bits)
+    Bytes[AllocByteOffset + B] |= AllocMask;
+}
+
 namespace {
 
+struct ByteArrayInfo {
+  std::set<uint64_t> Bits;
+  uint64_t BitSize;
+  GlobalVariable *ByteArray;
+  Constant *Mask;
+};
+
 struct LowerBitSets : public ModulePass {
   static char ID;
   LowerBitSets() : ModulePass(ID) {
     initializeLowerBitSetsPass(*PassRegistry::getPassRegistry());
   }
 
+  Module *M;
+
   const DataLayout *DL;
   IntegerType *Int1Ty;
   IntegerType *Int8Ty;
   IntegerType *Int32Ty;
   Type *Int32PtrTy;
   IntegerType *Int64Ty;
   Type *IntPtrTy;
 
   // The llvm.bitsets named metadata.
   NamedMDNode *BitSetNM;
 
   // Mapping from bitset mdstrings to the call sites that test them.
   DenseMap<MDString *, std::vector<CallInst *>> BitSetTestCallSites;
 
+  std::vector<ByteArrayInfo> ByteArrayInfos;
+
   BitSetInfo
   buildBitSet(MDString *BitSet,
               const DenseMap<GlobalVariable *, uint64_t> &GlobalLayout);
-  Value *createBitSetTest(IRBuilder<> &B, const BitSetInfo &BSI,
-                          GlobalVariable *BitSetGlobal, Value *BitOffset);
+  ByteArrayInfo *createByteArray(BitSetInfo &BSI);
+  void allocateByteArrays();
+  Value *createBitSetTest(IRBuilder<> &B, BitSetInfo &BSI, ByteArrayInfo *&BAI,
+                          Value *BitOffset);
   Value *
-  lowerBitSetCall(CallInst *CI, const BitSetInfo &BSI,
-                  GlobalVariable *BitSetGlobal, GlobalVariable *CombinedGlobal,
+  lowerBitSetCall(CallInst *CI, BitSetInfo &BSI, ByteArrayInfo *&BAI,
+                  GlobalVariable *CombinedGlobal,
                   const DenseMap<GlobalVariable *, uint64_t> &GlobalLayout);
-  void buildBitSetsFromGlobals(Module &M,
-                               const std::vector<MDString *> &BitSets,
+  void buildBitSetsFromGlobals(const std::vector<MDString *> &BitSets,
                                const std::vector<GlobalVariable *> &Globals);
-  bool buildBitSets(Module &M);
-  bool eraseBitSetMetadata(Module &M);
+  bool buildBitSets();
+  bool eraseBitSetMetadata();
 
   bool doInitialization(Module &M) override;
   bool runOnModule(Module &M) override;
 };
 
 } // namespace
 
 INITIALIZE_PASS_BEGIN(LowerBitSets, "lowerbitsets",
                 "Lower bitset metadata", false, false)
 INITIALIZE_PASS_END(LowerBitSets, "lowerbitsets",
                 "Lower bitset metadata", false, false)
 char LowerBitSets::ID = 0;
 
 ModulePass *llvm::createLowerBitSetsPass() { return new LowerBitSets; }
 
-bool LowerBitSets::doInitialization(Module &M) {
-  DL = M.getDataLayout();
+bool LowerBitSets::doInitialization(Module &Mod) {
+  M = &Mod;
+
+  DL = M->getDataLayout();
   if (!DL)
     report_fatal_error("Data layout required");
 
-  Int1Ty = Type::getInt1Ty(M.getContext());
-  Int8Ty = Type::getInt8Ty(M.getContext());
-  Int32Ty = Type::getInt32Ty(M.getContext());
+  Int1Ty = Type::getInt1Ty(M->getContext());
+  Int8Ty = Type::getInt8Ty(M->getContext());
+  Int32Ty = Type::getInt32Ty(M->getContext());
   Int32PtrTy = PointerType::getUnqual(Int32Ty);
-  Int64Ty = Type::getInt64Ty(M.getContext());
-  IntPtrTy = DL->getIntPtrType(M.getContext(), 0);
+  Int64Ty = Type::getInt64Ty(M->getContext());
+  IntPtrTy = DL->getIntPtrType(M->getContext(), 0);
 
-  BitSetNM = M.getNamedMetadata("llvm.bitsets");
+  BitSetNM = M->getNamedMetadata("llvm.bitsets");
 
   BitSetTestCallSites.clear();
 
   return false;
 }
 
 /// Build a bit set for BitSet using the object layouts in
 /// GlobalLayout.
static Value *createMaskedBitTest(IRBuilder<> &B, Value *Bits,
   BitOffset = B.CreateZExtOrTrunc(BitOffset, BitsType);
   Value *BitIndex =
       B.CreateAnd(BitOffset, ConstantInt::get(BitsType, BitWidth - 1));
   Value *BitMask = B.CreateShl(ConstantInt::get(BitsType, 1), BitIndex);
   Value *MaskedBits = B.CreateAnd(Bits, BitMask);
   return B.CreateICmpNE(MaskedBits, ConstantInt::get(BitsType, 0));
 }
 
+ByteArrayInfo *LowerBitSets::createByteArray(BitSetInfo &BSI) {
+  // Create globals to stand in for byte arrays and masks. These never actually
+  // get initialized, we RAUW and erase them later in allocateByteArrays() once
+  // we know the offset and mask to use.
+  auto ByteArrayGlobal = new GlobalVariable(
+      *M, Int8Ty, /*isConstant=*/true, GlobalValue::PrivateLinkage, nullptr);
+  auto MaskGlobal = new GlobalVariable(
+      *M, Int8Ty, /*isConstant=*/true, GlobalValue::PrivateLinkage, nullptr);
+
+  ByteArrayInfos.emplace_back();
+  ByteArrayInfo *BAI = &ByteArrayInfos.back();
+
+  BAI->Bits = BSI.Bits;
+  BAI->BitSize = BSI.BitSize;
+  BAI->ByteArray = ByteArrayGlobal;
+  BAI->Mask = ConstantExpr::getPtrToInt(MaskGlobal, Int8Ty);
+  return BAI;
+}
+
+void LowerBitSets::allocateByteArrays() {
+  std::stable_sort(ByteArrayInfos.begin(), ByteArrayInfos.end(),
+                   [](const ByteArrayInfo &BAI1, const ByteArrayInfo &BAI2) {
+                     return BAI1.BitSize > BAI2.BitSize;
+                   });
+
+  std::vector<uint64_t> ByteArrayOffsets(ByteArrayInfos.size());
+
+  ByteArrayBuilder BAB;
+  for (unsigned I = 0; I != ByteArrayInfos.size(); ++I) {
+    ByteArrayInfo *BAI = &ByteArrayInfos[I];
+
+    uint8_t Mask;
+    BAB.allocate(BAI->Bits, BAI->BitSize, ByteArrayOffsets[I], Mask);
+
+    BAI->Mask->replaceAllUsesWith(ConstantInt::get(Int8Ty, Mask));
+    cast<GlobalVariable>(BAI->Mask->getOperand(0))->eraseFromParent();
+  }
+
+  Constant *ByteArrayConst = ConstantDataArray::get(M->getContext(), BAB.Bytes);
+  auto ByteArray =
+      new GlobalVariable(*M, ByteArrayConst->getType(), /*isConstant=*/true,
+                         GlobalValue::PrivateLinkage, ByteArrayConst);
+
+  for (unsigned I = 0; I != ByteArrayInfos.size(); ++I) {
+    ByteArrayInfo *BAI = &ByteArrayInfos[I];
+
+    Constant *Idxs[] = {ConstantInt::get(IntPtrTy, 0),
+                        ConstantInt::get(IntPtrTy, ByteArrayOffsets[I])};
+    Constant *GEP = ConstantExpr::getInBoundsGetElementPtr(ByteArray, Idxs);
+
+    // Create an alias instead of RAUW'ing the gep directly. On x86 this ensures
+    // that the pc-relative displacement is folded into the lea instead of the
+    // test instruction getting another displacement.
+    GlobalAlias *Alias = GlobalAlias::create(
+        Int8Ty, 0, GlobalValue::PrivateLinkage, "bits", GEP, M);
+    BAI->ByteArray->replaceAllUsesWith(Alias);
+    BAI->ByteArray->eraseFromParent();
+  }
+
+  ByteArraySizeBits = BAB.BitAllocs[0] + BAB.BitAllocs[1] + BAB.BitAllocs[2] +
+                      BAB.BitAllocs[3] + BAB.BitAllocs[4] + BAB.BitAllocs[5] +
+                      BAB.BitAllocs[6] + BAB.BitAllocs[7];
+  ByteArraySizeBytes = BAB.Bytes.size();
+}
+
 /// Build a test that bit BitOffset is set in BSI, where
 /// BitSetGlobal is a global containing the bits in BSI.
-Value *LowerBitSets::createBitSetTest(IRBuilder<> &B, const BitSetInfo &BSI,
-                                      GlobalVariable *BitSetGlobal,
-                                      Value *BitOffset) {
-  if (BSI.Bits.size() <= 8) {
+Value *LowerBitSets::createBitSetTest(IRBuilder<> &B, BitSetInfo &BSI,
+                                      ByteArrayInfo *&BAI, Value *BitOffset) {
+  if (BSI.BitSize <= 64) {
     // If the bit set is sufficiently small, we can avoid a load by bit testing
     // a constant.
     IntegerType *BitsTy;
-    if (BSI.Bits.size() <= 4)
+    if (BSI.BitSize <= 32)
       BitsTy = Int32Ty;
     else
       BitsTy = Int64Ty;
 
     uint64_t Bits = 0;
-    for (auto I = BSI.Bits.rbegin(), E = BSI.Bits.rend(); I != E; ++I) {
-      Bits <<= 8;
-      Bits |= *I;
-    }
+    for (auto Bit : BSI.Bits)
+      Bits |= uint64_t(1) << Bit;
     Constant *BitsConst = ConstantInt::get(BitsTy, Bits);
     return createMaskedBitTest(B, BitsConst, BitOffset);
   } else {
-    // TODO: We might want to use the memory variant of the bt instruction
-    // with the previously computed bit offset at -Os. This instruction does
-    // exactly what we want but has been benchmarked as being slower than open
-    // coding the load+bt.
-    Value *BitSetGlobalOffset =
-        B.CreateLShr(BitOffset, ConstantInt::get(IntPtrTy, 5));
-    Value *BitSetEntryAddr = B.CreateGEP(
-        ConstantExpr::getBitCast(BitSetGlobal, Int32PtrTy), BitSetGlobalOffset);
-    Value *BitSetEntry = B.CreateLoad(BitSetEntryAddr);
+    if (!BAI) {
+      ++NumByteArraysCreated;
+      BAI = createByteArray(BSI);
+    }
+
+    Value *ByteAddr = B.CreateGEP(BAI->ByteArray, BitOffset);
+    Value *Byte = B.CreateLoad(ByteAddr);
 
-    return createMaskedBitTest(B, BitSetEntry, BitOffset);
+    Value *ByteAndMask = B.CreateAnd(Byte, BAI->Mask);
+    return B.CreateICmpNE(ByteAndMask, ConstantInt::get(Int8Ty, 0));
   }
 }
 
 /// Lower a llvm.bitset.test call to its implementation. Returns the value to
 /// replace the call with.
 Value *LowerBitSets::lowerBitSetCall(
-    CallInst *CI, const BitSetInfo &BSI, GlobalVariable *BitSetGlobal,
+    CallInst *CI, BitSetInfo &BSI, ByteArrayInfo *&BAI,
     GlobalVariable *CombinedGlobal,
     const DenseMap<GlobalVariable *, uint64_t> &GlobalLayout) {
   Value *Ptr = CI->getArgOperand(0);
 
   if (BSI.containsValue(DL, GlobalLayout, Ptr))
-    return ConstantInt::getTrue(BitSetGlobal->getParent()->getContext());
+    return ConstantInt::getTrue(CombinedGlobal->getParent()->getContext());
 
   Constant *GlobalAsInt = ConstantExpr::getPtrToInt(CombinedGlobal, IntPtrTy);
   Constant *OffsetedGlobalAsInt = ConstantExpr::getAdd(
       GlobalAsInt, ConstantInt::get(IntPtrTy, BSI.ByteOffset));
 
   BasicBlock *InitialBB = CI->getParent();
 
   IRBuilder<> B(CI);
Value *LowerBitSets::lowerBitSetCall(
   if (BSI.isAllOnes())
     return OffsetInRange;
 
   TerminatorInst *Term = SplitBlockAndInsertIfThen(OffsetInRange, CI, false);
   IRBuilder<> ThenB(Term);
 
   // Now that we know that the offset is in range and aligned, load the
   // appropriate bit from the bitset.
-  Value *Bit = createBitSetTest(ThenB, BSI, BitSetGlobal, BitOffset);
+  Value *Bit = createBitSetTest(ThenB, BSI, BAI, BitOffset);
 
   // The value we want is 0 if we came directly from the initial block
   // (having failed the range or alignment checks), or the loaded bit if
   // we came from the block in which we loaded it.
   B.SetInsertPoint(CI);
   PHINode *P = B.CreatePHI(Int1Ty, 2);
   P->addIncoming(ConstantInt::get(Int1Ty, 0), InitialBB);
   P->addIncoming(Bit, ThenB.GetInsertBlock());
   return P;
 }
 
 /// Given a disjoint set of bitsets and globals, layout the globals, build the
 /// bit sets and lower the llvm.bitset.test calls.
 void LowerBitSets::buildBitSetsFromGlobals(
-    Module &M,
     const std::vector<MDString *> &BitSets,
     const std::vector<GlobalVariable *> &Globals) {
   // Build a new global with the combined contents of the referenced globals.
   std::vector<Constant *> GlobalInits;
   for (GlobalVariable *G : Globals) {
     GlobalInits.push_back(G->getInitializer());
     uint64_t InitSize = DL->getTypeAllocSize(G->getInitializer()->getType());
 
     // Compute the amount of padding required to align the next element to the
     // next power of 2.
     uint64_t Padding = NextPowerOf2(InitSize - 1) - InitSize;
 
     // Cap at 128 was found experimentally to have a good data/instruction
     // overhead tradeoff.
     if (Padding > 128)
       Padding = RoundUpToAlignment(InitSize, 128) - InitSize;
 
     GlobalInits.push_back(
         ConstantAggregateZero::get(ArrayType::get(Int8Ty, Padding)));
   }
   if (!GlobalInits.empty())
     GlobalInits.pop_back();
-  Constant *NewInit = ConstantStruct::getAnon(M.getContext(), GlobalInits);
+  Constant *NewInit = ConstantStruct::getAnon(M->getContext(), GlobalInits);
   auto CombinedGlobal =
-      new GlobalVariable(M, NewInit->getType(), /*isConstant=*/true,
+      new GlobalVariable(*M, NewInit->getType(), /*isConstant=*/true,
                          GlobalValue::PrivateLinkage, NewInit);
 
   const StructLayout *CombinedGlobalLayout =
       DL->getStructLayout(cast<StructType>(NewInit->getType()));
 
   // Compute the offsets of the original globals within the new global.
   DenseMap<GlobalVariable *, uint64_t> GlobalLayout;
   for (unsigned I = 0; I != Globals.size(); ++I)
     // Multiply by 2 to account for padding elements.
     GlobalLayout[Globals[I]] = CombinedGlobalLayout->getElementOffset(I * 2);
 
   // For each bitset in this disjoint set...
   for (MDString *BS : BitSets) {
     // Build the bitset.
     BitSetInfo BSI = buildBitSet(BS, GlobalLayout);
 
-    // Create a global in which to store it.
-    ++NumBitSetsCreated;
-    Constant *BitsConst = ConstantDataArray::get(M.getContext(), BSI.Bits);
-    auto BitSetGlobal = new GlobalVariable(
-        M, BitsConst->getType(), /*isConstant=*/true,
-        GlobalValue::PrivateLinkage, BitsConst, BS->getString() + ".bits");
+    ByteArrayInfo *BAI = 0;
 
     // Lower each call to llvm.bitset.test for this bitset.
     for (CallInst *CI : BitSetTestCallSites[BS]) {
       ++NumBitSetCallsLowered;
-      Value *Lowered =
-          lowerBitSetCall(CI, BSI, BitSetGlobal, CombinedGlobal, GlobalLayout);
+      Value *Lowered = lowerBitSetCall(CI, BSI, BAI, CombinedGlobal, GlobalLayout);
       CI->replaceAllUsesWith(Lowered);
       CI->eraseFromParent();
     }
   }
 
   // Build aliases pointing to offsets into the combined global for each
   // global from which we built the combined global, and replace references
   // to the original globals with references to the aliases.
   for (unsigned I = 0; I != Globals.size(); ++I) {
     // Multiply by 2 to account for padding elements.
     Constant *CombinedGlobalIdxs[] = {ConstantInt::get(Int32Ty, 0),
                                       ConstantInt::get(Int32Ty, I * 2)};
     Constant *CombinedGlobalElemPtr =
         ConstantExpr::getGetElementPtr(CombinedGlobal, CombinedGlobalIdxs);
     GlobalAlias *GAlias = GlobalAlias::create(
         Globals[I]->getType()->getElementType(),
         Globals[I]->getType()->getAddressSpace(), Globals[I]->getLinkage(),
-        "", CombinedGlobalElemPtr, &M);
+        "", CombinedGlobalElemPtr, M);
     GAlias->takeName(Globals[I]);
     Globals[I]->replaceAllUsesWith(GAlias);
     Globals[I]->eraseFromParent();
   }
 }
 
 /// Lower all bit sets in this module.
-bool LowerBitSets::buildBitSets(Module &M) {
+bool LowerBitSets::buildBitSets() {
   Function *BitSetTestFunc =
-      M.getFunction(Intrinsic::getName(Intrinsic::bitset_test));
+      M->getFunction(Intrinsic::getName(Intrinsic::bitset_test));
   if (!BitSetTestFunc)
     return false;
 
   // Equivalence class set containing bitsets and the globals they reference.
   // This is used to partition the set of bitsets in the module into disjoint
   // sets.
   typedef EquivalenceClasses<PointerUnion<GlobalVariable *, MDString *>>
       GlobalClassesTy;
for (auto &&F : GLB.Fragments)
         *OGI++ = Globals[Offset];
 
     // Order bitsets by name for determinism.
     std::sort(BitSets.begin(), BitSets.end(), [](MDString *S1, MDString *S2) {
       return S1->getString() < S2->getString();
     });
 
     // Build the bitsets from this disjoint set.
-    buildBitSetsFromGlobals(M, BitSets, OrderedGlobals);
+    buildBitSetsFromGlobals(BitSets, OrderedGlobals);
   }
 
+  allocateByteArrays();
+
   return true;
 }
 
-bool LowerBitSets::eraseBitSetMetadata(Module &M) {
+bool LowerBitSets::eraseBitSetMetadata() {
   if (!BitSetNM)
     return false;
 
-  M.eraseNamedMetadata(BitSetNM);
+  M->eraseNamedMetadata(BitSetNM);
   return true;
 }
 
 bool LowerBitSets::runOnModule(Module &M) {
-  bool Changed = buildBitSets(M);
-  Changed |= eraseBitSetMetadata(M);
+  bool Changed = buildBitSets();
+  Changed |= eraseBitSetMetadata();
   return Changed;
 }

test/Transforms/LowerBitSets/simple.ll

 ; RUN: opt -S -lowerbitsets < %s | FileCheck %s
 ; RUN: opt -S -O3 < %s | FileCheck -check-prefix=CHECK-NODISCARD %s
 
 target datalayout = "e-p:32:32"
 
 ; CHECK: [[G:@[^ ]*]] = private constant { i32, [0 x i8], [63 x i32], [4 x i8], i32, [0 x i8], [2 x i32] } { i32 1, [0 x i8] zeroinitializer, [63 x i32] zeroinitializer, [4 x i8] zeroinitializer, i32 3, [0 x i8] zeroinitializer, [2 x i32] [i32 4, i32 5] }
 @a = constant i32 1
 @b = constant [63 x i32] zeroinitializer
 @c = constant i32 3
 @d = constant [2 x i32] [i32 4, i32 5]
 
+; CHECK: [[BA:@[^ ]*]] = private constant [68 x i8] c"\03\01\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\00\02\00\01"
+
 ; Offset 0, 4 byte alignment
-; CHECK: @bitset1.bits = private constant [9 x i8] c"\03\00\00\00\00\00\00\00\08"
 !0 = !{!"bitset1", i32* @a, i32 0}
 ; CHECK-NODISCARD-DAG: !{!"bitset1", i32* @a, i32 0}
 !1 = !{!"bitset1", [63 x i32]* @b, i32 0}
 ; CHECK-NODISCARD-DAG: !{!"bitset1", [63 x i32]* @b, i32 0}
 !2 = !{!"bitset1", [2 x i32]* @d, i32 4}
 ; CHECK-NODISCARD-DAG: !{!"bitset1", [2 x i32]* @d, i32 4}
 
 ; Offset 4, 256 byte alignment
-; CHECK: @bitset2.bits = private constant [1 x i8] c"\03"
 !3 = !{!"bitset2", [63 x i32]* @b, i32 0}
 ; CHECK-NODISCARD-DAG: !{!"bitset2", [63 x i32]* @b, i32 0}
 !4 = !{!"bitset2", i32* @c, i32 0}
 ; CHECK-NODISCARD-DAG: !{!"bitset2", i32* @c, i32 0}
 
 ; Offset 0, 4 byte alignment
-; CHECK: @bitset3.bits = private constant [9 x i8] c"\01\00\00\00\00\00\00\00\02"
 !5 = !{!"bitset3", i32* @a, i32 0}
 ; CHECK-NODISCARD-DAG: !{!"bitset3", i32* @a, i32 0}
 !6 = !{!"bitset3", i32* @c, i32 0}
 ; CHECK-NODISCARD-DAG: !{!"bitset3", i32* @c, i32 0}
 
 ; Entries whose second operand is null (the result of a global being DCE'd)
 ; should be ignored.
 !7 = !{!"bitset2", null, i32 0}
 
 !llvm.bitsets = !{ !0, !1, !2, !3, !4, !5, !6, !7 }
 
 ; CHECK: @a = alias getelementptr inbounds ({ i32, [0 x i8], [63 x i32], [4 x i8], i32, [0 x i8], [2 x i32] }* [[G]], i32 0, i32 0)
 ; CHECK: @b = alias getelementptr inbounds ({ i32, [0 x i8], [63 x i32], [4 x i8], i32, [0 x i8], [2 x i32] }* [[G]], i32 0, i32 2)
 ; CHECK: @c = alias getelementptr inbounds ({ i32, [0 x i8], [63 x i32], [4 x i8], i32, [0 x i8], [2 x i32] }* [[G]], i32 0, i32 4)
 ; CHECK: @d = alias getelementptr inbounds ({ i32, [0 x i8], [63 x i32], [4 x i8], i32, [0 x i8], [2 x i32] }* [[G]], i32 0, i32 6)
 
+; CHECK: @bits = private alias getelementptr inbounds ([68 x i8]* [[BA]], i32 0, i32 0)
+; CHECK: @bits1 = private alias getelementptr inbounds ([68 x i8]* [[BA]], i32 0, i32 0)
+
 declare i1 @llvm.bitset.test(i8* %ptr, metadata %bitset) nounwind readnone
 
 ; CHECK: @foo(i32* [[A0:%[^ ]*]])
 define i1 @foo(i32* %p) {
   ; CHECK-NOT: llvm.bitset.test
 
   ; CHECK: [[R0:%[^ ]*]] = bitcast i32* [[A0]] to i8*
   %pi8 = bitcast i32* %p to i8*
   ; CHECK: [[R1:%[^ ]*]] = ptrtoint i8* [[R0]] to i32
   ; CHECK: [[R2:%[^ ]*]] = sub i32 [[R1]], ptrtoint ({ i32, [0 x i8], [63 x i32], [4 x i8], i32, [0 x i8], [2 x i32] }* [[G]] to i32)
   ; CHECK: [[R3:%[^ ]*]] = lshr i32 [[R2]], 2
   ; CHECK: [[R4:%[^ ]*]] = shl i32 [[R2]], 30
   ; CHECK: [[R5:%[^ ]*]] = or i32 [[R3]], [[R4]]
   ; CHECK: [[R6:%[^ ]*]] = icmp ult i32 [[R5]], 68
   ; CHECK: br i1 [[R6]]
 
-  ; CHECK: [[R8:%[^ ]*]] = lshr i32 [[R5]], 5
-  ; CHECK: [[R9:%[^ ]*]] = getelementptr i32, i32* bitcast ([9 x i8]* @bitset1.bits to i32*), i32 [[R8]]
-  ; CHECK: [[R10:%[^ ]*]] = load i32, i32* [[R9]]
-  ; CHECK: [[R11:%[^ ]*]] = and i32 [[R5]], 31
-  ; CHECK: [[R12:%[^ ]*]] = shl i32 1, [[R11]]
-  ; CHECK: [[R13:%[^ ]*]] = and i32 [[R10]], [[R12]]
-  ; CHECK: [[R14:%[^ ]*]] = icmp ne i32 [[R13]], 0
+  ; CHECK: [[R8:%[^ ]*]] = getelementptr i8, i8* @bits, i32 [[R5]]
+  ; CHECK: [[R9:%[^ ]*]] = load i8, i8* [[R8]]
+  ; CHECK: [[R10:%[^ ]*]] = and i8 [[R9]], 1
+  ; CHECK: [[R11:%[^ ]*]] = icmp ne i8 [[R10]], 0
 
-  ; CHECK: [[R16:%[^ ]*]] = phi i1 [ false, {{%[^ ]*}} ], [ [[R14]], {{%[^ ]*}} ]
+  ; CHECK: [[R16:%[^ ]*]] = phi i1 [ false, {{%[^ ]*}} ], [ [[R11]], {{%[^ ]*}} ]
   %x = call i1 @llvm.bitset.test(i8* %pi8, metadata !"bitset1")
 
   ; CHECK-NOT: llvm.bitset.test
   %y = call i1 @llvm.bitset.test(i8* %pi8, metadata !"bitset1")
 
   ; CHECK: ret i1 [[R16]]
   ret i1 %x
 }
define i1 @baz(i32* %p) {
   ; CHECK: [[T1:%[^ ]*]] = ptrtoint i8* [[T0]] to i32
   ; CHECK: [[T2:%[^ ]*]] = sub i32 [[T1]], ptrtoint ({ i32, [0 x i8], [63 x i32], [4 x i8], i32, [0 x i8], [2 x i32] }* [[G]] to i32)
   ; CHECK: [[T3:%[^ ]*]] = lshr i32 [[T2]], 2
   ; CHECK: [[T4:%[^ ]*]] = shl i32 [[T2]], 30
   ; CHECK: [[T5:%[^ ]*]] = or i32 [[T3]], [[T4]]
   ; CHECK: [[T6:%[^ ]*]] = icmp ult i32 [[T5]], 66
   ; CHECK: br i1 [[T6]]
 
-  ; CHECK: [[T8:%[^ ]*]] = lshr i32 [[T5]], 5
-  ; CHECK: [[T9:%[^ ]*]] = getelementptr i32, i32* bitcast ([9 x i8]* @bitset3.bits to i32*), i32 [[T8]]
-  ; CHECK: [[T10:%[^ ]*]] = load i32, i32* [[T9]]
-  ; CHECK: [[T11:%[^ ]*]] = and i32 [[T5]], 31
-  ; CHECK: [[T12:%[^ ]*]] = shl i32 1, [[T11]]
-  ; CHECK: [[T13:%[^ ]*]] = and i32 [[T10]], [[T12]]
-  ; CHECK: [[T14:%[^ ]*]] = icmp ne i32 [[T13]], 0
+  ; CHECK: [[T8:%[^ ]*]] = getelementptr i8, i8* @bits1, i32 [[T5]]
+  ; CHECK: [[T9:%[^ ]*]] = load i8, i8* [[T8]]
+  ; CHECK: [[T10:%[^ ]*]] = and i8 [[T9]], 2
+  ; CHECK: [[T11:%[^ ]*]] = icmp ne i8 [[T10]], 0
 
-  ; CHECK: [[T16:%[^ ]*]] = phi i1 [ false, {{%[^ ]*}} ], [ [[T14]], {{%[^ ]*}} ]
+  ; CHECK: [[T16:%[^ ]*]] = phi i1 [ false, {{%[^ ]*}} ], [ [[T11]], {{%[^ ]*}} ]
   %x = call i1 @llvm.bitset.test(i8* %pi8, metadata !"bitset3")
   ; CHECK: ret i1 [[T16]]
   ret i1 %x
 }
 
 ; CHECK-NOT: !llvm.bitsets

unittests/Transforms/IPO/LowerBitSets.cpp

 #include "llvm/Transforms/IPO/LowerBitSets.h"
 #include "gtest/gtest.h"
 
 using namespace llvm;
 
 TEST(LowerBitSets, BitSetBuilder) {
   struct {
     std::vector<uint64_t> Offsets;
-    std::vector<uint8_t> Bits;
+    std::set<uint64_t> Bits;
     uint64_t ByteOffset;
     uint64_t BitSize;
     unsigned AlignLog2;
     bool IsSingleOffset;
     bool IsAllOnes;
   } BSBTests[] = {
-      {{}, {0}, 0, 1, 0, false, false},
-      {{0}, {1}, 0, 1, 0, true, true},
-      {{4}, {1}, 4, 1, 0, true, true},
-      {{37}, {1}, 37, 1, 0, true, true},
-      {{0, 1}, {3}, 0, 2, 0, false, true},
-      {{0, 4}, {3}, 0, 2, 2, false, true},
-      {{0, uint64_t(1) << 33}, {3}, 0, 2, 33, false, true},
-      {{3, 7}, {3}, 3, 2, 2, false, true},
-      {{0, 1, 7}, {131}, 0, 8, 0, false, false},
-      {{0, 2, 14}, {131}, 0, 8, 1, false, false},
-      {{0, 1, 8}, {3, 1}, 0, 9, 0, false, false},
-      {{0, 2, 16}, {3, 1}, 0, 9, 1, false, false},
-      {{0, 1, 2, 3, 4, 5, 6, 7}, {255}, 0, 8, 0, false, true},
-      {{0, 1, 2, 3, 4, 5, 6, 7, 8}, {255, 1}, 0, 9, 0, false, true},
+      {{}, {}, 0, 1, 0, false, false},
+      {{0}, {0}, 0, 1, 0, true, true},
+      {{4}, {0}, 4, 1, 0, true, true},
+      {{37}, {0}, 37, 1, 0, true, true},
+      {{0, 1}, {0, 1}, 0, 2, 0, false, true},
+      {{0, 4}, {0, 1}, 0, 2, 2, false, true},
+      {{0, uint64_t(1) << 33}, {0, 1}, 0, 2, 33, false, true},
+      {{3, 7}, {0, 1}, 3, 2, 2, false, true},
+      {{0, 1, 7}, {0, 1, 7}, 0, 8, 0, false, false},
+      {{0, 2, 14}, {0, 1, 7}, 0, 8, 1, false, false},
+      {{0, 1, 8}, {0, 1, 8}, 0, 9, 0, false, false},
+      {{0, 2, 16}, {0, 1, 8}, 0, 9, 1, false, false},
+      {{0, 1, 2, 3, 4, 5, 6, 7},
+       {0, 1, 2, 3, 4, 5, 6, 7},
+       0,
+       8,
+       0,
+       false,
+       true},
+      {{0, 1, 2, 3, 4, 5, 6, 7, 8},
+       {0, 1, 2, 3, 4, 5, 6, 7, 8},
+       0,
+       9,
+       0,
+       false,
+       true},
   };
 
   for (auto &&T : BSBTests) {
     BitSetBuilder BSB;
     for (auto Offset : T.Offsets)
       BSB.addOffset(Offset);
 
     BitSetInfo BSI = BSB.build();
for (auto &&T : GLBTests) {
 
     std::vector<uint64_t> ComputedLayout;
     for (auto &&F : GLB.Fragments)
       ComputedLayout.insert(ComputedLayout.end(), F.begin(), F.end());
 
     EXPECT_EQ(T.WantLayout, ComputedLayout);
   }
 }
+
+TEST(LowerBitSets, ByteArrayBuilder) {
+  struct BABAlloc {
+    std::set<uint64_t> Bits;
+    uint64_t BitSize;
+    uint64_t WantByteOffset;
+    uint8_t WantMask;
+  };
+
+  struct {
+    std::vector<BABAlloc> Allocs;
+    std::vector<uint8_t> WantBytes;
+  } BABTests[] = {
+      {{{{0}, 1, 0, 1}, {{0}, 1, 0, 2}}, {3}},
+      {{{{0}, 16, 0, 1},
+        {{1}, 15, 0, 2},
+        {{2}, 14, 0, 4},
+        {{3}, 13, 0, 8},
+        {{4}, 12, 0, 0x10},
+        {{5}, 11, 0, 0x20},
+        {{6}, 10, 0, 0x40},
+        {{7}, 9, 0, 0x80},
+        {{0}, 7, 9, 0x80},
+        {{0}, 6, 10, 0x40},
+        {{0}, 5, 11, 0x20},
+        {{0}, 4, 12, 0x10},
+        {{0}, 3, 13, 8},
+        {{0}, 2, 14, 4},
+        {{0}, 1, 15, 2}},
+       {1, 2, 4, 8, 0x10, 0x20, 0x40, 0x80, 0, 0x80, 0x40, 0x20, 0x10, 8, 4,
+        2}},
+  };
+
+  for (auto &&T : BABTests) {
+    ByteArrayBuilder BABuilder;
+
+    for (auto &&A : T.Allocs) {
+      uint64_t GotByteOffset;
+      uint8_t GotMask;
+
+      BABuilder.allocate(A.Bits, A.BitSize, GotByteOffset, GotMask);
+      EXPECT_EQ(A.WantByteOffset, GotByteOffset);
+      EXPECT_EQ(A.WantMask, GotMask);
+    }
+
+    EXPECT_EQ(T.WantBytes, BABuilder.Bytes);
+  }
+}