[IR] Increase max alignment to 4GB

Currently the max alignment representable is 1GB, see D108661.
Setting the align of an object to 4GB is desirable in some cases to make sure the lower 32 bits are clear which can be used for some optimizations, e.g. https://crbug.com/1016945.

This uses an extra bit in instructions that carry an alignment. We can store 15 bits of "free" information, and with this change some instructions (e.g. AtomicCmpXchgInst) use 14 bits.
We can increase the max alignment representable above 4GB (up to 2^62) since we're only using 33 of the 64 values, but I've just limited it to 4GB for now.

The one place we have to update the bitcode format is for the alloca instruction. It stores its alignment into 5 bits of a 32 bit bitfield. I've added another field which is 8 bits and should be future proof for a while. For backward compatibility, we check if the old field has a value and use that, otherwise use the new field.

Updating clang's max allowed alignment will come in a future patch.

Reviewed By: hans

Differential Revision: https://reviews.llvm.org/D110451

clang/lib/CodeGen/CGBlocks.cpp

@@ -2695,8 +2695,8 @@ const BlockByrefInfo &CodeGenFunction::getBlockByrefInfo(const VarDecl *D) {
     size = varOffset;
 
   // Conversely, we might have to prevent LLVM from inserting padding.
-  } else if (CGM.getDataLayout().getABITypeAlignment(varTy)
-               > varAlign.getQuantity()) {
+  } else if (CGM.getDataLayout().getABITypeAlignment(varTy) >
+             uint64_t(varAlign.getQuantity())) {
     packed = true;
   }
   types.push_back(varTy);

clang/lib/CodeGen/CGCall.cpp

@@ -5021,12 +5021,12 @@ RValue CodeGenFunction::EmitCall(const CGFunctionInfo &CallInfo,
         auto scalarAlign = CGM.getDataLayout().getPrefTypeAlignment(scalarType);
 
         // Materialize to a temporary.
-        addr = CreateTempAlloca(
-            RV.getScalarVal()->getType(),
-            CharUnits::fromQuantity(std::max(
-                (unsigned)layout->getAlignment().value(), scalarAlign)),
-            "tmp",
-            /*ArraySize=*/nullptr, &AllocaAddr);
+        addr =
+            CreateTempAlloca(RV.getScalarVal()->getType(),
+                             CharUnits::fromQuantity(std::max(
+                                 layout->getAlignment().value(), scalarAlign)),
+                             "tmp",
+                             /*ArraySize=*/nullptr, &AllocaAddr);
         tempSize = EmitLifetimeStart(scalarSize, AllocaAddr.getPointer());
 
         Builder.CreateStore(RV.getScalarVal(), addr);

clang/lib/CodeGen/CGDecl.cpp

@@ -1142,7 +1142,7 @@ Address CodeGenModule::createUnnamedGlobalFrom(const VarDecl &D,
     GV->setAlignment(Align.getAsAlign());
     GV->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
     CacheEntry = GV;
-  } else if (CacheEntry->getAlignment() < Align.getQuantity()) {
+  } else if (CacheEntry->getAlignment() < uint64_t(Align.getQuantity())) {
     CacheEntry->setAlignment(Align.getAsAlign());
   }
 

clang/lib/CodeGen/CodeGenModule.cpp

@@ -5386,7 +5386,7 @@ CodeGenModule::GetAddrOfConstantStringFromLiteral(const StringLiteral *S,
   if (!LangOpts.WritableStrings) {
     Entry = &ConstantStringMap[C];
     if (auto GV = *Entry) {
-      if (Alignment.getQuantity() > GV->getAlignment())
+      if (uint64_t(Alignment.getQuantity()) > GV->getAlignment())
         GV->setAlignment(Alignment.getAsAlign());
       return ConstantAddress(castStringLiteralToDefaultAddressSpace(*this, GV),
                              Alignment);
@@ -5449,7 +5449,7 @@ ConstantAddress CodeGenModule::GetAddrOfConstantCString(
   if (!LangOpts.WritableStrings) {
     Entry = &ConstantStringMap[C];
     if (auto GV = *Entry) {
-      if (Alignment.getQuantity() > GV->getAlignment())
+      if (uint64_t(Alignment.getQuantity()) > GV->getAlignment())
         GV->setAlignment(Alignment.getAsAlign());
       return ConstantAddress(castStringLiteralToDefaultAddressSpace(*this, GV),
                              Alignment);

llvm/include/llvm/Bitcode/BitcodeCommon.h

@@ -19,10 +19,14 @@
 namespace llvm {
 
 struct AllocaPackedValues {
-  using Align = Bitfield::Element<unsigned, 0, 5>;
-  using UsedWithInAlloca = Bitfield::Element<bool, Align::NextBit, 1>;
+  // We increased the number of bits needed to represent alignment to be more
+  // than 5, but to preserve backward compatibility we store the upper bits
+  // separately.
+  using AlignLower = Bitfield::Element<unsigned, 0, 5>;
+  using UsedWithInAlloca = Bitfield::Element<bool, AlignLower::NextBit, 1>;
   using ExplicitType = Bitfield::Element<bool, UsedWithInAlloca::NextBit, 1>;
   using SwiftError = Bitfield::Element<bool, ExplicitType::NextBit, 1>;
+  using AlignUpper = Bitfield::Element<unsigned, SwiftError::NextBit, 3>;
 };
 
 } // namespace llvm

llvm/include/llvm/IR/DataLayout.h

@@ -518,7 +518,7 @@ class DataLayout {
 
   /// Returns the minimum ABI-required alignment for the specified type.
   /// FIXME: Deprecate this function once migration to Align is over.
-  unsigned getABITypeAlignment(Type *Ty) const;
+  uint64_t getABITypeAlignment(Type *Ty) const;
 
   /// Returns the minimum ABI-required alignment for the specified type.
   Align getABITypeAlign(Type *Ty) const;
@@ -541,7 +541,7 @@ class DataLayout {
   ///
   /// This is always at least as good as the ABI alignment.
   /// FIXME: Deprecate this function once migration to Align is over.
-  unsigned getPrefTypeAlignment(Type *Ty) const;
+  uint64_t getPrefTypeAlignment(Type *Ty) const;
 
   /// Returns the preferred stack/global alignment for the specified
   /// type.

llvm/include/llvm/IR/GlobalObject.h

@@ -51,7 +51,7 @@ class GlobalObject : public GlobalValue {
 
   Comdat *ObjComdat;
   enum {
-    LastAlignmentBit = 4,
+    LastAlignmentBit = 5,
     HasSectionHashEntryBit,
 
     GlobalObjectBits,
@@ -68,7 +68,7 @@ class GlobalObject : public GlobalValue {
   GlobalObject(const GlobalObject &) = delete;
 
   /// FIXME: Remove this function once transition to Align is over.
-  unsigned getAlignment() const {
+  uint64_t getAlignment() const {
     MaybeAlign Align = getAlign();
     return Align ? Align->value() : 0;
   }

llvm/include/llvm/IR/Instruction.h

@@ -59,11 +59,11 @@ class Instruction : public User,
   // Template alias so that all Instruction storing alignment use the same
   // definiton.
   // Valid alignments are powers of two from 2^0 to 2^MaxAlignmentExponent =
-  // 2^30. We store them as Log2(Alignment), so we need 5 bits to encode the 31
+  // 2^32. We store them as Log2(Alignment), so we need 6 bits to encode the 33
   // possible values.
   template <unsigned Offset>
   using AlignmentBitfieldElementT =
-      typename Bitfield::Element<unsigned, Offset, 5,
+      typename Bitfield::Element<unsigned, Offset, 6,
                                  Value::MaxAlignmentExponent>;
 
   template <unsigned Offset>

llvm/include/llvm/IR/Instructions.h

@@ -126,7 +126,7 @@ class AllocaInst : public UnaryInstruction {
   }
 
   // FIXME: Remove this one transition to Align is over.
-  unsigned getAlignment() const { return getAlign().value(); }
+  uint64_t getAlignment() const { return getAlign().value(); }
 
   /// Return true if this alloca is in the entry block of the function and is a
   /// constant size. If so, the code generator will fold it into the
@@ -217,7 +217,7 @@ class LoadInst : public UnaryInstruction {
   /// Return the alignment of the access that is being performed.
   /// FIXME: Remove this function once transition to Align is over.
   /// Use getAlign() instead.
-  unsigned getAlignment() const { return getAlign().value(); }
+  uint64_t getAlignment() const { return getAlign().value(); }
 
   /// Return the alignment of the access that is being performed.
   Align getAlign() const {
@@ -348,7 +348,7 @@ class StoreInst : public Instruction {
   /// Return the alignment of the access that is being performed
   /// FIXME: Remove this function once transition to Align is over.
   /// Use getAlign() instead.
-  unsigned getAlignment() const { return getAlign().value(); }
+  uint64_t getAlignment() const { return getAlign().value(); }
 
   Align getAlign() const {
     return Align(1ULL << (getSubclassData<AlignmentField>()));

llvm/include/llvm/IR/Value.h

@@ -785,8 +785,8 @@ class Value {
   ///
   /// This is the greatest alignment value supported by load, store, and alloca
   /// instructions, and global values.
-  static constexpr unsigned MaxAlignmentExponent = 30;
-  static constexpr unsigned MaximumAlignment = 1u << MaxAlignmentExponent;
+  static constexpr unsigned MaxAlignmentExponent = 32;
+  static constexpr uint64_t MaximumAlignment = 1ULL << MaxAlignmentExponent;
 
   /// Mutate the type of this Value to be of the specified type.
   ///

llvm/include/llvm/Transforms/IPO/Attributor.h

@@ -3413,18 +3413,18 @@ struct AADereferenceable
 };
 
 using AAAlignmentStateType =
-    IncIntegerState<uint32_t, Value::MaximumAlignment, 1>;
+    IncIntegerState<uint64_t, Value::MaximumAlignment, 1>;
 /// An abstract interface for all align attributes.
 struct AAAlign : public IRAttribute<
                      Attribute::Alignment,
                      StateWrapper<AAAlignmentStateType, AbstractAttribute>> {
   AAAlign(const IRPosition &IRP, Attributor &A) : IRAttribute(IRP) {}
 
   /// Return assumed alignment.
-  unsigned getAssumedAlign() const { return getAssumed(); }
+  uint64_t getAssumedAlign() const { return getAssumed(); }
 
   /// Return known alignment.
-  unsigned getKnownAlign() const { return getKnown(); }
+  uint64_t getKnownAlign() const { return getKnown(); }
 
   /// See AbstractAttribute::getName()
   const std::string getName() const override { return "AAAlign"; }

llvm/lib/AsmParser/LLParser.cpp

@@ -1927,7 +1927,7 @@ bool LLParser::parseOptionalAlignment(MaybeAlign &Alignment, bool AllowParens) {
   if (!EatIfPresent(lltok::kw_align))
     return false;
   LocTy AlignLoc = Lex.getLoc();
-  uint32_t Value = 0;
+  uint64_t Value = 0;
 
   LocTy ParenLoc = Lex.getLoc();
   bool HaveParens = false;
@@ -1936,13 +1936,13 @@ bool LLParser::parseOptionalAlignment(MaybeAlign &Alignment, bool AllowParens) {
       HaveParens = true;
   }
 
-  if (parseUInt32(Value))
+  if (parseUInt64(Value))
     return true;
 
   if (HaveParens && !EatIfPresent(lltok::rparen))
     return error(ParenLoc, "expected ')'");
 
-  if (!isPowerOf2_32(Value))
+  if (!isPowerOf2_64(Value))
     return error(AlignLoc, "alignment is not a power of two");
   if (Value > Value::MaximumAlignment)
     return error(AlignLoc, "huge alignments are not supported yet");

llvm/lib/Bitcode/Reader/BitcodeReader.cpp

@@ -4930,8 +4930,10 @@ Error BitcodeReader::parseFunctionBody(Function *F) {
       Type *OpTy = getTypeByID(Record[1]);
       Value *Size = getFnValueByID(Record[2], OpTy);
       MaybeAlign Align;
-      if (Error Err =
-              parseAlignmentValue(Bitfield::get<APV::Align>(Rec), Align)) {
+      uint64_t AlignExp =
+          Bitfield::get<APV::AlignLower>(Rec) |
+          (Bitfield::get<APV::AlignUpper>(Rec) << APV::AlignLower::Bits);
+      if (Error Err = parseAlignmentValue(AlignExp, Align)) {
         return Err;
       }
       if (!Ty || !Size)

llvm/lib/Bitcode/Writer/BitcodeWriter.cpp

@@ -3055,7 +3055,11 @@ void ModuleBitcodeWriter::writeInstruction(const Instruction &I,
     Vals.push_back(VE.getValueID(I.getOperand(0))); // size.
     using APV = AllocaPackedValues;
     unsigned Record = 0;
-    Bitfield::set<APV::Align>(Record, getEncodedAlign(AI.getAlign()));
+    unsigned EncodedAlign = getEncodedAlign(AI.getAlign());
+    Bitfield::set<APV::AlignLower>(
+        Record, EncodedAlign & ((1 << APV::AlignLower::Bits) - 1));
+    Bitfield::set<APV::AlignUpper>(Record,
+                                   EncodedAlign >> APV::AlignLower::Bits);
     Bitfield::set<APV::UsedWithInAlloca>(Record, AI.isUsedWithInAlloca());
     Bitfield::set<APV::ExplicitType>(Record, true);
     Bitfield::set<APV::SwiftError>(Record, AI.isSwiftError());

llvm/lib/CodeGen/MIRParser/MIParser.cpp

@@ -498,7 +498,7 @@ class MIParser {
                                          MachineOperand &Dest,
                                          Optional<unsigned> &TiedDefIdx);
   bool parseOffset(int64_t &Offset);
-  bool parseAlignment(unsigned &Alignment);
+  bool parseAlignment(uint64_t &Alignment);
   bool parseAddrspace(unsigned &Addrspace);
   bool parseSectionID(Optional<MBBSectionID> &SID);
   bool parseOperandsOffset(MachineOperand &Op);
@@ -676,7 +676,7 @@ bool MIParser::parseBasicBlockDefinition(
   bool IsLandingPad = false;
   bool IsEHFuncletEntry = false;
   Optional<MBBSectionID> SectionID;
-  unsigned Alignment = 0;
+  uint64_t Alignment = 0;
   BasicBlock *BB = nullptr;
   if (consumeIfPresent(MIToken::lparen)) {
     do {
@@ -2898,16 +2898,16 @@ bool MIParser::parseOffset(int64_t &Offset) {
   return false;
 }
 
-bool MIParser::parseAlignment(unsigned &Alignment) {
+bool MIParser::parseAlignment(uint64_t &Alignment) {
   assert(Token.is(MIToken::kw_align) || Token.is(MIToken::kw_basealign));
   lex();
   if (Token.isNot(MIToken::IntegerLiteral) || Token.integerValue().isSigned())
     return error("expected an integer literal after 'align'");
-  if (getUnsigned(Alignment))
+  if (getUint64(Alignment))
     return true;
   lex();
 
-  if (!isPowerOf2_32(Alignment))
+  if (!isPowerOf2_64(Alignment))
     return error("expected a power-of-2 literal after 'align'");
 
   return false;
@@ -3261,7 +3261,7 @@ bool MIParser::parseMachineMemoryOperand(MachineMemOperand *&Dest) {
     if (parseMachinePointerInfo(Ptr))
       return true;
   }
-  unsigned BaseAlignment =
+  uint64_t BaseAlignment =
       (Size != MemoryLocation::UnknownSize ? PowerOf2Ceil(Size) : 1);
   AAMDNodes AAInfo;
   MDNode *Range = nullptr;

llvm/lib/CodeGen/SafeStack.cpp

@@ -147,7 +147,7 @@ class SafeStack {
   ///
   /// 16 seems like a reasonable upper bound on the alignment of objects that we
   /// might expect to appear on the stack on most common targets.
-  enum { StackAlignment = 16 };
+  static constexpr uint64_t StackAlignment = 16;
 
   /// Return the value of the stack canary.
   Value *getStackGuard(IRBuilder<> &IRB, Function &F);
@@ -544,8 +544,7 @@ Value *SafeStack::moveStaticAllocasToUnsafeStack(
       Size = 1; // Don't create zero-sized stack objects.
 
     // Ensure the object is properly aligned.
-    unsigned Align =
-        std::max((unsigned)DL.getPrefTypeAlignment(Ty), AI->getAlignment());
+    uint64_t Align = std::max(DL.getPrefTypeAlignment(Ty), AI->getAlignment());
 
     SSL.addObject(AI, Size, Align,
                   ClColoring ? SSC.getLiveRange(AI) : NoColoringRange);
@@ -676,9 +675,9 @@ void SafeStack::moveDynamicAllocasToUnsafeStack(
     SP = IRB.CreateSub(SP, Size);
 
     // Align the SP value to satisfy the AllocaInst, type and stack alignments.
-    unsigned Align = std::max(
-        std::max((unsigned)DL.getPrefTypeAlignment(Ty), AI->getAlignment()),
-        (unsigned)StackAlignment);
+    uint64_t Align =
+        std::max(std::max(DL.getPrefTypeAlignment(Ty), AI->getAlignment()),
+                 StackAlignment);
 
     assert(isPowerOf2_32(Align));
     Value *NewTop = IRB.CreateIntToPtr(

llvm/lib/IR/DataLayout.cpp

@@ -819,7 +819,7 @@ Align DataLayout::getAlignment(Type *Ty, bool abi_or_pref) const {
 }
 
 /// TODO: Remove this function once the transition to Align is over.
-unsigned DataLayout::getABITypeAlignment(Type *Ty) const {
+uint64_t DataLayout::getABITypeAlignment(Type *Ty) const {
   return getABITypeAlign(Ty).value();
 }
 
@@ -828,7 +828,7 @@ Align DataLayout::getABITypeAlign(Type *Ty) const {
 }
 
 /// TODO: Remove this function once the transition to Align is over.
-unsigned DataLayout::getPrefTypeAlignment(Type *Ty) const {
+uint64_t DataLayout::getPrefTypeAlignment(Type *Ty) const {
   return getPrefTypeAlign(Ty).value();
 }
 

llvm/lib/Transforms/Instrumentation/AddressSanitizer.cpp

@@ -176,7 +176,7 @@ const char kAMDGPUAddressPrivateName[] = "llvm.amdgcn.is.private";
 // Accesses sizes are powers of two: 1, 2, 4, 8, 16.
 static const size_t kNumberOfAccessSizes = 5;
 
-static const unsigned kAllocaRzSize = 32;
+static const uint64_t kAllocaRzSize = 32;
 
 // ASanAccessInfo implementation constants.
 constexpr size_t kCompileKernelShift = 0;
@@ -3578,7 +3578,7 @@ void FunctionStackPoisoner::poisonAlloca(Value *V, uint64_t Size,
 void FunctionStackPoisoner::handleDynamicAllocaCall(AllocaInst *AI) {
   IRBuilder<> IRB(AI);
 
-  const unsigned Alignment = std::max(kAllocaRzSize, AI->getAlignment());
+  const uint64_t Alignment = std::max(kAllocaRzSize, AI->getAlignment());
   const uint64_t AllocaRedzoneMask = kAllocaRzSize - 1;
 
   Value *Zero = Constant::getNullValue(IntptrTy);

llvm/lib/Transforms/Instrumentation/HWAddressSanitizer.cpp

@@ -356,7 +356,7 @@ class HWAddressSanitizer {
     bool WithFrameRecord;
 
     void init(Triple &TargetTriple, bool InstrumentWithCalls);
-    unsigned getObjectAlignment() const { return 1U << Scale; }
+    uint64_t getObjectAlignment() const { return 1ULL << Scale; }
   };
 
   ShadowMapping Mapping;

llvm/test/Assembler/align-inst-alloca.ll

@@ -1,6 +1,6 @@
 ; RUN: not llvm-as %s -o /dev/null 2>/dev/null
 
 define void @foo() {
-  %p = alloca i1, align 2147483648
+  %p = alloca i1, align 8589934592
   ret void
 }

llvm/test/Assembler/align-inst-load.ll

@@ -1,6 +1,6 @@
 ; RUN: not llvm-as %s -o /dev/null 2>/dev/null
 
 define void @foo(i1* %p) {
-  load i1, i1* %p, align 2147483648
+  load i1, i1* %p, align 8589934592
   ret void
 }

llvm/test/Assembler/align-inst-store.ll

@@ -1,6 +1,6 @@
 ; RUN: not llvm-as %s -o /dev/null 2>/dev/null
 
 define void @foo(i1* %p) {
-  store i1 false, i1* %p, align 2147483648
+  store i1 false, i1* %p, align 8589934592
   ret void
 }

llvm/test/Assembler/align-inst.ll

@@ -1,11 +1,11 @@
 ; RUN: llvm-as %s -o /dev/null
 ; RUN: verify-uselistorder %s
 
-@A = global i1 0, align 1073741824
+@A = global i1 0, align 4294967296
 
 define void @foo() {
-  %p = alloca i1, align 1073741824
-  load i1, i1* %p, align 1073741824
-  store i1 false, i1* %p, align 1073741824
+  %p = alloca i1, align 4294967296
+  load i1, i1* %p, align 4294967296
+  store i1 false, i1* %p, align 4294967296
   ret void
 }