Add alloc_size attribute to clang

[george.burgess.iv]

Summary

This patch adds the alloc_size attribute, docs, and all of the logic needed to handle it to clang. The only thing that's _really_ missing is our ability to handle non-const locals, because that's proving to be a moderately interesting challenge. Our current strategy is to lower alloc_size to LLVM, and have it handle those cases.

There's also a restructuring of TryEvaluateBuiltinObjectSize thrown in, because I felt like it was getting overly subtle. Hopefully the new version is considerably easier to reason about. :)

— Implementation notes —

• With the restructuring of TryEvaluateBuiltinObjectSize, I was able to make EM_DesignatorFold (now EM_OffsetFold) the only EvalMode we use in evaluating __builtin_object_size.

• The InvalidBase functionality of LValues was further (ab)used to get information to TryEvaluateBuiltinObjectSize. We know an LValue has been initialized by an alloc_size function if the base is invalid and the base itself is a CallExpr (or a CastExpr that contains the alloc_size CallExpr). This won’t conflict with current behavior, because (prior to this patch) all invalid bases were MemberExprs.

[george.burgess.iv]

As discussed offline:

• Added support for treating const variables as constexpr
• Added requirement for LValues with an invalid alloc_size base to have an array index as their first designator entry (along with support for having an array of unknown size as the first entry in an LValue's designator)

Additionally:

• Fixed a bug with the prior impl where we'd report too many bytes for type=3 with alloc_size
• Refactored a fair amount of duplicated logic
• Added checks to be sure that LValues with invalid bases don't get moved into APValues

...I didn't end up refactoring the LValue designator walking, because we only needed one change in one place (because invalid bases can't find their way into APValues). I'm happy to try to make a visitor-like thing and use it to refactor the ~4-5 places we walk {LValue,APValue} designators (3 of which reside in ExprConstant.cpp) if we still think that would be relatively valuable.

[george.burgess.iv]

Also, due to allowing the evaluation of const expressions as constexpr expressions, this patch needs http://reviews.llvm.org/D14877 to go in so we don't break tests. :)

[aaron.ballman]

include/clang/Basic/Attr.td

714

Should this be using HasFunctionProto instead of Function?

include/clang/Basic/DiagnosticSemaKinds.td

2144

This warning diagnostic is not used.

2145

This text is a bit misleading based on the diagnostic. The attribute doesn't apply to the integer argument. The attribute argument should refer to an integer parameter of the function declaration. I think the diagnostic should reflect that. Perhaps:

%0 attribute argument may only refer to a function parameter of integer type

and the diagnostic loc can point to the attribute argument in question. (If that's not possible, it would be good to add a %1 to identify which of the two possible attribute arguments is at fault.)

lib/AST/ExprConstant.cpp

117

Add a newline to separate the functions (here and elsewhere).

128

const auto *, here and elsewhere. Whenever you have auto, we want all the cv-qualifiers as well as */& information along with the auto.

144

Is this function going to expand to be more than just a forward to another? If not, we should remove it.

154

Would be good to have '&& "explanation of assert"' as well (here and elsewhere).

262

Missing a period.

267

Missing function documentation.

357

Missing period.

1110

Is there a reason this assert only fires in !NDEBUG mode? Is it too expensive to enable for any assert build?

lib/Sema/SemaDeclAttr.cpp

721

Please do not use auto unless the type is explicitly spelled out in the initializer (here and elsewhere).

728

Are character types and Boolean types also okay?

764

If you use HasFunctionProto in Attr.td, none of this is required. If you can't use HasFunctionProto, then there's still no need to check FD because that's handled by the common attribute code.

766

Please, no magic numbers. I don't think 4 is even the correct value, that's ExpectedParameter.

777

Please do not use auto unless the type is explicitly spelled out in the initializer.

[george.burgess.iv]

Addressed all feedback

[majnemer]

include/clang/Basic/AttrDocs.td

214

nonconst -> non-const?

[george.burgess.iv]

• Fixed wording in AttrDocs
• Realized I forgot to press 'submit' on comment responses after the last update. Will correct this in ~5 seconds :)

include/clang/Basic/Attr.td

714

Nice catch!

include/clang/Basic/AttrDocs.td

214

Danke

lib/AST/ExprConstant.cpp

1110

I was under the impression that asserts are only enabled if !NDEBUG. The code only exists to make a few assertions, so I think guarding it with "if assertions are enabled" makes that purpose more clear. If my assumption is wrong, I'm happy to remove the ifndef.

lib/Sema/SemaDeclAttr.cpp

728

Boolean makes no sense, but char is fine here. Thanks

766

Doesn't matter; dead code.

[aaron.ballman]

lib/AST/ExprConstant.cpp

1110

I was under the impression that asserts are only enabled if !NDEBUG. The code only exists to make a few assertions, so I think guarding it with "if assertions are enabled" makes that purpose more clear. If my assumption is wrong, I'm happy to remove the ifndef.

That's a good point; I forgot that we use NDEBUG as our release+asserts flag (derp). That being said, this is basically the same as:

assert((!BInvalid || isa<MemberExpr>(B.get<Expr*>()) || tryUnwrapAllocSizeCall(B.get<Expr*>())) && "Unexpected type of invalid base");

Uncertain if that is an improvement or not, but it is shorter-ish. :-)

lib/Sema/SemaDeclAttr.cpp

728

Boolean makes no sense, but char is fine here. Thanks

isIntegerType() already covers isCharType(). I think what you want is:

if (!Param->getType()->isIntegerType() || Param->getType()->isBooleanType())

[george.burgess.iv]

• Rebased
• Removed bits that were already upstreamed as part of D14877
• Added support for LLVM's allocsize attribute (under review at D14933 )

[aaron.ballman]

The attribute part looks mostly good (a few small nits), but the rest should be reviewed by @rsmith.

lib/Sema/SemaDeclAttr.cpp

718

Can be replaced with assert(Attr.isArgExpr(AttrArgNo) && "expected expression argument");

720

This whole if should be replaced by a call to checkFunctionOrMethodParameterIndex(), which will return the correct function parameter number as an output parameter.

[george.burgess.iv]

Addressed all feedback.

lib/Sema/SemaDeclAttr.cpp

718

Neat! Thanks.

[george.burgess.iv]

Now that allocsize is in LLVM (r266032), this isn't blocked. Yay!

Rebased, added a few tests, did a few tiny refactors, and fixed an overflow bug when the user tried to allocate 2**63 < N < 2**64 bytes (because CharUnits uses an int64_t as storage).

[aaron.ballman]

include/clang/Basic/Attr.td

717

I don't see any C++ tests involving templates; can you add some?

test/SemaCXX/constant-expression-cxx11.cpp

1171

This change seems out of place for a test that doesn't use alloc_size anywhere. Can you explain why this test has changed?

[george.burgess.iv]

Addressed all feedback

include/clang/Basic/Attr.td

717

Good catch

test/SemaCXX/constant-expression-cxx11.cpp

1171

Yup!

Clang's constexpr evaluator is conservative about const variables. If it kept this attitude, alloc_size evaluation can't happen in non-constexpr contexts in clang, even in cases like:

int foo() {
  void *const p = malloc(10);
  return __builtin_object_size(p, 0);
}

The change I made at ExprConstant.cpp:2789 (which Richard okayed when I talked with him offline) makes the constexpr evaluator more willing to look at const variables. With that, we can fold calls to __builtin_object_size like above in clang.

It also means that there are a few cases outside of alloc_size where the constexpr evaluator can be more aggressive. As it turns out, this is one of them (as is the change in test/CodeGenCXX/global-init.cpp, IIRC)

[aaron.ballman]

Thank you for working on this! LGTM, but please wait for @rsmith to okay as well.

test/SemaCXX/constant-expression-cxx11.cpp

1171

Thank you for the explanation!

[ahatanak]

I'm wondering what the status of this patch is since someone has asked us to add support for this attribute in clang. Are you still waiting for review?

test/CodeGenCXX/alloc-size.cpp

66 ↗

(On Diff #55478)

Is it necessary to compute __builtin_object_size in the front-end (rather than in some IR passes like instcombine) when it takes the pointer returned by a function marked alloc_size?

Also, is the IR optimization smart enough to get the exact object size in the following case?

void foo(int a, int b) {
  void *p0 = my_malloc(a);
  g0 = __builtin_object_size(p0, 0);
  void *p1 = my_calloc(a, b);
  g1 = __builtin_object_size(p1, 1);
}

void foo1() {
  foo(10, 50);
}

[george.burgess.iv]

I'm wondering what the status of this patch is since someone has asked us to add support for this attribute in clang. Are you still waiting for review?

Waiting for an LGTM from Richard, though this patch currently doesn't pass all tests. One of the changes in it causes some craziness with how we codegen blocks (in e.g. objc/OpenCL). Basically, we'll end up emitting the same block definition N times, since this makes the constexpr evaluator more accurate in some cases. I hope to fix that + get back to this in the nearish future.

(Happy to accept an LGTM now, though, since the bug isn't with this patch. ;) Naturally, I'll submit after the block-related bug is fixed.)

test/CodeGenCXX/alloc-size.cpp

66 ↗

(On Diff #55478)

Is it necessary to compute __builtin_object_size in the front-end (rather than in some IR passes like instcombine)

It's necessary to try in the frontend. If that fails, we'll lower calls to __bos to @llvm.objectsize, and let LLVM try it.

The main reason that we care for the frontend is, among other reasons, so we can use it with the enable_if attribute, like so:

void strncat(char *buf, int n, const char *from)
  __attribute__((overloadable,
                 enable_if(__builtin_object_size(buf, 0) < n, ""),
                 unavailable("'n' is larger than the target buffer!")));

void strncat(char *buf, int n, const char *from)
  __attribute__((overloadable));

int main() {
  char buf[4];
  strncat(buf, sizeof(buf)+1, "hi"); // expected-error{{'n' is larger than the target buffer!}}
}

---

Also, is the IR optimization smart enough to get the exact object size in the following case?

__builtin_object_size *always* gets lowered to a constant by either clang or LLVM. That is, if foo gets inlined into foo1, then LLVM should be able to determine accurate values for g0 and g1. If foo isn't inlined, then no: you'll get -1 for both.

If you'd like to see what we can and can't get, LLVM already has the allocsize attribute (and @llvm.objectsize intrinsic) in it. :)

[ahatanak]

test/CodeGenCXX/alloc-size.cpp

66 ↗

(On Diff #55478)

Thanks for the explanation. I see why you want to evaluate it in the front-end.

It looks like llvm optimizes function foo before it gets inlined into function foo1 and, as a result, g0 and g1 both get -1 instead of 10 and 500. I guess llvm should defer optimizing the function until after inliner is run.

[george.burgess.iv]

test/CodeGenCXX/alloc-size.cpp

66 ↗

(On Diff #55478)

Hm. Looks like how we handle allocsize in LLVM may be slightly broken, then, because it works with regular allocas.

Filed PR28834. Thanks for the heads-up. :)

[hfinkel]

What's the status of this?

[george.burgess.iv]

With any luck, this'll be on the top of my to-do list next week. :)

[george.burgess.iv]

Rebased and made the __builtin_object_size code a tiny bit cleaner.

The blocks bugfix I mentioned will be up as a separate review in a few minutes. :)

[george.burgess.iv]

Ping

[george.burgess.iv]

Ping :)

[aaron.ballman]

I still think that this looks good, so if @rsmith doesn't comment over the weekend, you can go ahead and commit on Monday -- we can handle any feedback in post-commit review.

include/clang/Basic/AttrDocs.td

212

Double space after the full stop should be a single space (here and below).

[george.burgess.iv]

Thanks, everyone! :)

include/clang/Basic/Attr.td

 
 def Destructor : InheritableAttr {
   let Spellings = [GCC<"destructor">];
   let Args = [DefaultIntArgument<"Priority", 65535>];
   let Subjects = SubjectList<[Function]>;
   let Documentation = [Undocumented];
 }
 
+def AllocSize : InheritableAttr {
+  let Spellings = [GCC<"alloc_size">];
+  let Subjects = SubjectList<[HasFunctionProto], WarnDiag,
+                             "ExpectedFunctionWithProtoType">;
+  let Args = [IntArgument<"ElemSizeParam">, IntArgument<"NumElemsParam", 1>];
+  let TemplateDependent = 1;
+  let Documentation = [AllocSizeDocs];
+}
+
 def EnableIf : InheritableAttr {
   let Spellings = [GNU<"enable_if">];
   let Subjects = SubjectList<[Function]>;
   let Args = [ExprArgument<"Cond">, StringArgument<"Message">];
   let TemplateDependent = 1;
   let Documentation = [EnableIfDocs];
 }
 

include/clang/Basic/AttrDocs.td

.. code-block:: c++
   void *b() __attribute__((assume_aligned (32, 4)));
 
 Note that this attribute provides information to the compiler regarding a
 condition that the code already ensures is true. It does not cause the compiler
 to enforce the provided alignment assumption.
   }];
 }
 
+def AllocSizeDocs : Documentation {
+  let Category = DocCatFunction;
+  let Content = [{
+The ``alloc_size`` attribute can be placed on functions that return pointers in
+order to hint to the compiler how many bytes of memory will be available at the
+the returned poiner. ``alloc_size`` takes one or two arguments.
+
+- ``alloc_size(N)`` implies that argument number N equals the number of
+  available bytes at the returned pointer.
+- ``alloc_size(N, M)`` implies that the product of argument number N and
+  argument number M equals the number of available bytes at the returned
+  pointer.
+
+Argument numbers are 1-based.
+
+An example of how to use ``alloc_size``
+
+.. code-block:: c
+
+  void *my_malloc(int a) __attribute__((alloc_size(1)));
+  void *my_calloc(int a, int b) __attribute__((alloc_size(1, 2)));
+
+  int main() {
+    void *const p = my_malloc(100);
+    assert(__builtin_object_size(p, 0) == 100);
+    void *const a = my_calloc(20, 5);
+    assert(__builtin_object_size(a, 0) == 100);
+  }
+
+.. Note:: This attribute works differently in clang than it does in GCC.
+  Specifically, clang will only trace const pointers (as above); we give up on
+  non-const ones.  It's planned that LLVM will have an ``alloc_size`` attribute
+  that we can lower to in the near future. This should cover a large number of
+  nonconst cases.
+  }];
+}
+
 def EnableIfDocs : Documentation {
   let Category = DocCatFunction;
   let Content = [{
 .. Note:: Some features of this attribute are experimental. The meaning of
   multiple enable_if attributes on a single declaration is subject to change in
   a future version of clang. Also, the ABI is not standardized and the name
   mangling may change in future versions. To avoid that, use asm labels.
 

include/clang/Basic/DiagnosticSemaKinds.td

 def err_attribute_only_once_per_parameter : Error<
   "%0 attribute can only be applied once per parameter">;
 def err_attribute_uuid_malformed_guid : Error<
   "uuid attribute contains a malformed GUID">;
 def warn_attribute_pointers_only : Warning<
   "%0 attribute only applies to%select{| constant}1 pointer arguments">,
   InGroup<IgnoredAttributes>;
 def err_attribute_pointers_only : Error<warn_attribute_pointers_only.Text>;
+def err_attribute_integers_only : Error<
+  "%0 attribute argument may only refer to a function parameter of integer "
+  "type">;
 def warn_attribute_return_pointers_only : Warning<
   "%0 attribute only applies to return values that are pointers">,
   InGroup<IgnoredAttributes>;
 def warn_attribute_return_pointers_refs_only : Warning<
   "%0 attribute only applies to return values that are pointers or references">,
   InGroup<IgnoredAttributes>;
 def warn_attribute_pointer_or_reference_only : Warning<
   "%0 attribute only applies to a pointer or reference (%1 is invalid)">,

lib/AST/ExprConstant.cpp

static const CXXRecordDecl *getAsBaseClass(APValue::LValuePathEntry E) {
     return dyn_cast<CXXRecordDecl>(getAsBaseOrMember(E).getPointer());
   }
   /// Determine whether this LValue path entry for a base class names a virtual
   /// base class.
   static bool isVirtualBaseClass(APValue::LValuePathEntry E) {
     return getAsBaseOrMember(E).getInt();
   }
 
+  /// Given a CallExpr, try to get the alloc_size attribute. May return null.
+  static const AllocSizeAttr *getAllocSizeAttr(const CallExpr *CE) {
+    const auto *Callee = CE->getDirectCallee();
+    return Callee ? Callee->getAttr<AllocSizeAttr>() : nullptr;
+  }
+
+  /// Attempts to unwrap a CallExpr (with an alloc_size attribute) from an Expr.
+  /// This will look through a single cast.
+  ///
+  /// Returns null if we couldn't unwrap a function with alloc_size.
+  static const CallExpr *tryUnwrapAllocSizeCall(const Expr *E) {
+    if (!E->getType()->isPointerType())
+      return nullptr;
+
+    E = E->IgnoreParens();
+    // If we're doing a variable assignment from e.g. malloc(N), there will
+    // probably be a cast of some kind. Ignore it.
+    if (const auto *Cast = dyn_cast<CastExpr>(E))
+      E = Cast->getSubExpr()->IgnoreParens();
+
+    if (const auto *CE = dyn_cast<CallExpr>(E))
+      return getAllocSizeAttr(CE) ? CE : nullptr;
+    return nullptr;
+  }
+
+  /// Determines whether or not the given Base contains a call to a function
+  /// with the alloc_size attribute.
+  static bool isBaseAnAllocSizeCall(APValue::LValueBase Base) {
+    const auto *E = Base.dyn_cast<const Expr *>();
+    return E && E->getType()->isPointerType() && tryUnwrapAllocSizeCall(E);
+  }
+
+  /// Determines if an LValue with the given LValueBase will have an unsized
+  /// array in its designator.
   /// Find the path length and type of the most-derived subobject in the given
   /// path, and find the size of the containing array, if any.
-  static
-  unsigned findMostDerivedSubobject(ASTContext &Ctx, QualType Base,
-                                    ArrayRef<APValue::LValuePathEntry> Path,
-                                    uint64_t &ArraySize, QualType &Type,
-                                    bool &IsArray) {
+  static unsigned
+  findMostDerivedSubobject(ASTContext &Ctx, APValue::LValueBase Base,
+                           uint64_t &ArraySize, QualType &Type, bool &IsArray) {
+    // This only accepts LValueBases from APValues, and APValues don't support
+    // arrays that lack size info.
+    assert(!isBaseAnAllocSizeCall(Base) && "Unsized arrays shouldn't appear here");
     unsigned MostDerivedLength = 0;
-    Type = Base;
+    Type = getType(Base);
+
     for (unsigned I = 0, N = Path.size(); I != N; ++I) {
       if (Type->isArrayType()) {
         const ConstantArrayType *CAT =
             cast<ConstantArrayType>(Ctx.getAsArrayType(Type));
         Type = CAT->getElementType();
         ArraySize = CAT->getSize().getZExtValue();
         MostDerivedLength = I + 1;
         IsArray = true;
       } else if (Type->isAnyComplexType()) {
         const ComplexType *CT = Type->castAs<ComplexType>();
         Type = CT->getElementType();
         ArraySize = 2;
struct SubobjectDesignator {
     /// True if the subobject was named in a manner not supported by C++11. Such
     /// lvalues can still be folded, but they are not core constant expressions
     /// and we cannot perform lvalue-to-rvalue conversions on them.
     bool Invalid : 1;
 
     /// Is this a pointer one past the end of an object?
     bool IsOnePastTheEnd : 1;
 
+    /// Indicator of whether the first entry is an unsized array.
+    bool FirstEntryIsAnUnsizedArray : 1;
+
     /// Indicator of whether the most-derived object is an array element.
     bool MostDerivedIsArrayElement : 1;
 
     /// The length of the path to the most-derived object of which this is a
     /// subobject.
-    unsigned MostDerivedPathLength : 29;
+    unsigned MostDerivedPathLength : 28;
 
     /// The size of the array of which the most-derived object is an element.
     /// This will always be 0 if the most-derived object is not an array
     /// element. 0 is not an indicator of whether or not the most-derived object
     /// is an array, however, because 0-length arrays are allowed.
+    ///
+    /// If the current array is an unsized array, the value of this is
+    /// undefined.
     uint64_t MostDerivedArraySize;
 
     /// The type of the most derived object referred to by this address.
     QualType MostDerivedType;
 
     typedef APValue::LValuePathEntry PathEntry;
 
     /// The entries on the path from the glvalue to the designated subobject.
     SmallVector<PathEntry, 8> Entries;
 
     SubobjectDesignator() : Invalid(true) {}
 
     explicit SubobjectDesignator(QualType T)
         : Invalid(false), IsOnePastTheEnd(false),
-          MostDerivedIsArrayElement(false), MostDerivedPathLength(0),
-          MostDerivedArraySize(0), MostDerivedType(T) {}
+          FirstEntryIsAnUnsizedArray(false), MostDerivedIsArrayElement(false),
+          MostDerivedPathLength(0), MostDerivedArraySize(0),
+          MostDerivedType(T) {}
 
     SubobjectDesignator(ASTContext &Ctx, const APValue &V)
         : Invalid(!V.isLValue() || !V.hasLValuePath()), IsOnePastTheEnd(false),
-          MostDerivedIsArrayElement(false), MostDerivedPathLength(0),
-          MostDerivedArraySize(0) {
+          FirstEntryIsAnUnsizedArray(false), MostDerivedIsArrayElement(false),
+          MostDerivedPathLength(0), MostDerivedArraySize(0) {
+      assert(V.isLValue() && "Non-LValue used to make an LValue designator?");
       if (!Invalid) {
         IsOnePastTheEnd = V.isLValueOnePastTheEnd();
         ArrayRef<PathEntry> VEntries = V.getLValuePath();
         Entries.insert(Entries.end(), VEntries.begin(), VEntries.end());
         if (V.getLValueBase()) {
           bool IsArray = false;
-          MostDerivedPathLength =
-              findMostDerivedSubobject(Ctx, getType(V.getLValueBase()),
-                                       V.getLValuePath(), MostDerivedArraySize,
-                                       MostDerivedType, IsArray);
+          MostDerivedPathLength = findMostDerivedSubobject(
+              Ctx, V.getLValueBase(), V.getLValuePath(), MostDerivedArraySize,
           MostDerivedIsArrayElement = IsArray;
         }
       }
     }
 
     void setInvalid() {
       Invalid = true;
       Entries.clear();
     }
 
+    /// Determine whether the most derived subobject is an array without a
+    /// known bound.
+    bool isMostDerivedAnUnsizedArray() const {
+      return FirstEntryIsAnUnsizedArray && Entries.size() == 1;
+    }
+
+    /// Determine what the most derived array's size is. Results in an assertion
+    /// failure if the most derived array lacks a size.
+    uint64_t getMostDerivedArraySize() const {
+      assert(!isMostDerivedAnUnsizedArray() && "Unsized array has no size");
+      return MostDerivedArraySize;
+    }
+
     /// Determine whether this is a one-past-the-end pointer.
     bool isOnePastTheEnd() const {
       assert(!Invalid);
       if (IsOnePastTheEnd)
         return true;
-      if (MostDerivedIsArrayElement &&
+      if (!isMostDerivedAnUnsizedArray() && MostDerivedIsArrayElement &&
           Entries[MostDerivedPathLength - 1].ArrayIndex == MostDerivedArraySize)
         return true;
       return false;
     }
 
     /// Check that this refers to a valid subobject.
     bool isValidSubobject() const {
       if (Invalid)
void addArrayUnchecked(const ConstantArrayType *CAT) {
       Entries.push_back(Entry);
 
       // This is a most-derived object.
       MostDerivedType = CAT->getElementType();
       MostDerivedIsArrayElement = true;
       MostDerivedArraySize = CAT->getSize().getZExtValue();
       MostDerivedPathLength = Entries.size();
     }
+    /// Update this designator to refer to the first element within the array of
+    /// elements of type T. This is an array of unknown size.
+    void addUnsizedArrayUnchecked(QualType ElemTy) {
+      PathEntry Entry;
+      Entry.ArrayIndex = 0;
+      Entries.push_back(Entry);
+
+      MostDerivedType = ElemTy;
+      MostDerivedIsArrayElement = true;
+      // The value in MostDerivedArraySize is undefined in this case. So, set it
+      // to an arbitrary value that's likely to loudly break things if it's
+      // used.
+      MostDerivedArraySize = std::numeric_limits<uint64_t>::max() / 2;
+      MostDerivedPathLength = Entries.size();
+    }
     /// Update this designator to refer to the given base or member of this
     /// object.
     void addDeclUnchecked(const Decl *D, bool Virtual = false) {
       PathEntry Entry;
       APValue::BaseOrMemberType Value(D, Virtual);
       Entry.BaseOrMember = Value.getOpaqueValue();
       Entries.push_back(Entry);
 
void addComplexUnchecked(QualType EltTy, bool Imag) {
       MostDerivedIsArrayElement = true;
       MostDerivedArraySize = 2;
       MostDerivedPathLength = Entries.size();
     }
     void diagnosePointerArithmetic(EvalInfo &Info, const Expr *E, uint64_t N);
     /// Add N to the address of this subobject.
     void adjustIndex(EvalInfo &Info, const Expr *E, uint64_t N) {
       if (Invalid) return;
+      if (isMostDerivedAnUnsizedArray()) {
+        // Can't verify -- trust that the user is doing the right thing (or if
+        // not, trust that the caller will catch the bad behavior).
+        Entries.back().ArrayIndex += N;
+        return;
+      }
       if (MostDerivedPathLength == Entries.size() &&
           MostDerivedIsArrayElement) {
         Entries.back().ArrayIndex += N;
-        if (Entries.back().ArrayIndex > MostDerivedArraySize) {
+        if (Entries.back().ArrayIndex > getMostDerivedArraySize()) {
           diagnosePointerArithmetic(Info, E, Entries.back().ArrayIndex);
           setInvalid();
         }
         return;
       }
       // [expr.add]p4: For the purposes of these operators, a pointer to a
       // nonarray object behaves the same as a pointer to the first element of
       // an array of length one with the type of the object as its element type.
enum EvaluationMode {
       /// construct that we can't evaluate yet (because we don't yet know the
       /// value of something) but stop if we hit something that could never be
       /// a constant expression. Some expressions can be retried in the
       /// optimizer if we don't constant fold them here, but in an unevaluated
       /// context we try to fold them immediately since the optimizer never
       /// gets a chance to look at it.
       EM_PotentialConstantExpressionUnevaluated,
 
-      /// Evaluate as a constant expression. Continue evaluating if we find a
-      /// MemberExpr with a base that can't be evaluated.
-      EM_DesignatorFold,
+      /// Evaluate as a constant expression. Continue evaluating if either:
+      /// - We find a MemberExpr with a base that can't be evaluated.
+      /// - We find a variable initialized with a call to a function that has
+      ///   the alloc_size attribute on it.
+      /// In either case, the LValue returned shall have an invalid base; in the
+      /// former, the base will be the invalid MemberExpr, in the latter, the
+      /// base will be either the alloc_size CallExpr or a CastExpr wrapping
+      /// said CallExpr.
+      EM_OffsetFold,
     } EvalMode;
 
     /// Are we checking whether the expression is a potential constant
     /// expression?
     bool checkingPotentialConstantExpression() const {
       return EvalMode == EM_PotentialConstantExpression ||
              EvalMode == EM_PotentialConstantExpressionUnevaluated;
     }
OptionalDiagnostic Diag(SourceLocation Loc, diag::kind DiagId
           case EM_EvaluateForOverflow:
             if (!HasFoldFailureDiagnostic)
               break;
             // We've already failed to fold something. Keep that diagnostic.
           case EM_ConstantExpression:
           case EM_PotentialConstantExpression:
           case EM_ConstantExpressionUnevaluated:
           case EM_PotentialConstantExpressionUnevaluated:
-          case EM_DesignatorFold:
+          case EM_OffsetFold:
             HasActiveDiagnostic = false;
             return OptionalDiagnostic();
           }
         }
 
         unsigned CallStackNotes = CallStackDepth - 1;
         unsigned Limit = Ctx.getDiagnostics().getConstexprBacktraceLimit();
         if (Limit)
bool keepEvaluatingAfterSideEffect() {
       case EM_PotentialConstantExpressionUnevaluated:
       case EM_EvaluateForOverflow:
       case EM_IgnoreSideEffects:
         return true;
 
       case EM_ConstantExpression:
       case EM_ConstantExpressionUnevaluated:
       case EM_ConstantFold:
-      case EM_DesignatorFold:
+      case EM_OffsetFold:
         return false;
       }
       llvm_unreachable("Missed EvalMode case");
     }
 
     /// Note that we have had a side-effect, and determine whether we should
     /// keep evaluating.
     bool noteSideEffect() {
       EvalStatus.HasSideEffects = true;
       return keepEvaluatingAfterSideEffect();
     }
 
     /// Should we continue evaluation after encountering undefined behavior?
     bool keepEvaluatingAfterUndefinedBehavior() {
       switch (EvalMode) {
       case EM_EvaluateForOverflow:
       case EM_IgnoreSideEffects:
       case EM_ConstantFold:
-      case EM_DesignatorFold:
+      case EM_OffsetFold:
         return true;
 
       case EM_PotentialConstantExpression:
       case EM_PotentialConstantExpressionUnevaluated:
       case EM_ConstantExpression:
       case EM_ConstantExpressionUnevaluated:
         return false;
       }
bool keepEvaluatingAfterFailure() {
       case EM_PotentialConstantExpressionUnevaluated:
       case EM_EvaluateForOverflow:
         return true;
 
       case EM_ConstantExpression:
       case EM_ConstantExpressionUnevaluated:
       case EM_ConstantFold:
       case EM_IgnoreSideEffects:
-      case EM_DesignatorFold:
+      case EM_OffsetFold:
         return false;
       }
       llvm_unreachable("Missed EvalMode case");
     }
 
-    bool allowInvalidBaseExpr() const {
-      return EvalMode == EM_DesignatorFold;
-    }
+    bool allowInvalidBaseExpr() const { return EvalMode == EM_OffsetFold; }
   };
 
   /// Object used to treat all foldable expressions as constant expressions.
   struct FoldConstant {
     EvalInfo &Info;
     bool Enabled;
     bool HadNoPriorDiags;
     EvalInfo::EvaluationMode OldMode;
struct FoldConstant {
     }
   };
 
   /// RAII object used to treat the current evaluation as the correct pointer
   /// offset fold for the current EvalMode
   struct FoldOffsetRAII {
     EvalInfo &Info;
     EvalInfo::EvaluationMode OldMode;
-    explicit FoldOffsetRAII(EvalInfo &Info, bool Subobject)
+    explicit FoldOffsetRAII(EvalInfo &Info)
         : Info(Info), OldMode(Info.EvalMode) {
       if (!Info.checkingPotentialConstantExpression())
-        Info.EvalMode = Subobject ? EvalInfo::EM_DesignatorFold
-                                  : EvalInfo::EM_ConstantFold;
+        Info.EvalMode = EvalInfo::EM_OffsetFold;
     }
 
     ~FoldOffsetRAII() { Info.EvalMode = OldMode; }
   };
 
   /// RAII object used to suppress diagnostics and side-effects from a
   /// speculative evaluation.
   class SpeculativeEvaluationRAII {
if (isOnePastTheEnd()) {
     setInvalid();
     return false;
   }
   return true;
 }
 
 void SubobjectDesignator::diagnosePointerArithmetic(EvalInfo &Info,
                                                     const Expr *E, uint64_t N) {
+  // If we're complaining, we must statically know the size of the most derived
+  // array.
   if (MostDerivedPathLength == Entries.size() && MostDerivedIsArrayElement)
     Info.CCEDiag(E, diag::note_constexpr_array_index)
       << static_cast<int>(N) << /*array*/ 0
-      << static_cast<unsigned>(MostDerivedArraySize);
+      << static_cast<unsigned>(getMostDerivedArraySize());
   else
     Info.CCEDiag(E, diag::note_constexpr_array_index)
       << static_cast<int>(N) << /*non-array*/ 1;
   setInvalid();
 }
 
 CallStackFrame::CallStackFrame(EvalInfo &Info, SourceLocation CallLoc,
                                const FunctionDecl *Callee, const LValue *This,
struct LValue {
     const CharUnits &getLValueOffset() const { return Offset; }
     unsigned getLValueCallIndex() const { return CallIndex; }
     SubobjectDesignator &getLValueDesignator() { return Designator; }
     const SubobjectDesignator &getLValueDesignator() const { return Designator;}
 
     void moveInto(APValue &V) const {
       if (Designator.Invalid)
         V = APValue(Base, Offset, APValue::NoLValuePath(), CallIndex);
-      else
+      else {
+        assert(!InvalidBase && "APValues can't handle invalid LValue bases");
+        assert(!Designator.FirstEntryIsAnUnsizedArray &&
+               "Unsized array with a valid base?");
         V = APValue(Base, Offset, Designator.Entries,
                     Designator.IsOnePastTheEnd, CallIndex);
       }
+    }
     void setFrom(ASTContext &Ctx, const APValue &V) {
-      assert(V.isLValue());
+      assert(V.isLValue() && "Setting LValue from a non-LValue?");
       Base = V.getLValueBase();
       Offset = V.getLValueOffset();
       InvalidBase = false;
       CallIndex = V.getLValueCallIndex();
       Designator = SubobjectDesignator(Ctx, V);
     }
 
     void set(APValue::LValueBase B, unsigned I = 0, bool BInvalid = false) {
+#ifndef NDEBUG
+      // We only allow a few types of invalid bases. Enforce that here.
+      if (BInvalid) {
+        const auto *E = B.get<const Expr *>();
+        assert((isa<MemberExpr>(E) || tryUnwrapAllocSizeCall(E)) &&
+               "Unexpected type of invalid base");
+      }
+#endif
+
       Base = B;
       Offset = CharUnits::Zero();
       InvalidBase = BInvalid;
       CallIndex = I;
       Designator = SubobjectDesignator(getType(B));
     }
 
     void setInvalid(APValue::LValueBase B, unsigned I = 0) {
bool checkSubobject(EvalInfo &Info, const Expr *E, CheckSubobjectKind CSK) {
              Designator.checkSubobject(Info, E, CSK);
     }
 
     void addDecl(EvalInfo &Info, const Expr *E,
                  const Decl *D, bool Virtual = false) {
       if (checkSubobject(Info, E, isa<FieldDecl>(D) ? CSK_Field : CSK_Base))
         Designator.addDeclUnchecked(D, Virtual);
     }
+    void addUnsizedArray(EvalInfo &Info, QualType ElemTy) {
+      assert(Designator.Entries.empty() && getType(Base)->isPointerType());
+      assert(isBaseAnAllocSizeCall(Base) &&
+             "Only alloc_size bases can have unsized arrays");
+      Designator.FirstEntryIsAnUnsizedArray = true;
+      Designator.addUnsizedArrayUnchecked(ElemTy);
+    }
     void addArray(EvalInfo &Info, const Expr *E, const ConstantArrayType *CAT) {
       if (checkSubobject(Info, E, CSK_ArrayToPointer))
         Designator.addArrayUnchecked(CAT);
     }
     void addComplex(EvalInfo &Info, const Expr *E, QualType EltTy, bool Imag) {
       if (checkSubobject(Info, E, Imag ? CSK_Imag : CSK_Real))
         Designator.addComplexUnchecked(EltTy, Imag);
     }
if (!Frame) {
           VD == Info.EvaluatingDecl.dyn_cast<const ValueDecl *>()) {
         // OK, we can read and modify an object if we're in the process of
         // evaluating its initializer, because its lifetime began in this
         // evaluation.
       } else if (AK != AK_Read) {
         // All the remaining cases only permit reading.
         Info.Diag(E, diag::note_constexpr_modify_global);
         return CompleteObject();
-      } else if (VD->isConstexpr()) {
+      } else if (VD->isConstexpr() || BaseType.isConstQualified()) {
         // OK, we can read this variable.
       } else if (BaseType->isIntegralOrEnumerationType()) {
         if (!BaseType.isConstQualified()) {
           if (Info.getLangOpts().CPlusPlus) {
             Info.Diag(E, diag::note_constexpr_ltor_non_const_int, 1) << VD;
             Info.Note(VD->getLocation(), diag::note_declared_at);
           } else {
             Info.Diag(E);
bool LValueExprEvaluator::VisitBinAssign(const BinaryOperator *E) {
   return handleAssignment(this->Info, E, Result, E->getLHS()->getType(),
                           NewVal);
 }
 
 //===----------------------------------------------------------------------===//
 // Pointer Evaluation
 //===----------------------------------------------------------------------===//
 
+/// \brief Attempts to compute the number of bytes available at the pointer
+/// returned by a function with the alloc_size attribute. Returns true if we
+/// were successful.
+///
+/// This expects the given CallExpr to be a call to a function with an
+/// alloc_size attribute.
+static bool getBytesReturnedByAllocSizeCall(const ASTContext &Ctx,
+                                            const CallExpr *Call,
+                                            CharUnits &Result) {
+  const auto *AllocSize = getAllocSizeAttr(Call);
+
+  // alloc_size args are base-1, 0 means not present.
+  assert(AllocSize && AllocSize->getElemSizeParam() != 0);
+  unsigned SizeArgNo = AllocSize->getElemSizeParam() - 1;
+  unsigned BitsInSizeT = Ctx.getTypeSize(Ctx.getSizeType());
+  if (Call->getNumArgs() <= SizeArgNo)
+    return false;
+
+  auto EvaluateAsSizeT = [&](const Expr *E, APSInt &Into) {
+    if (!E->EvaluateAsInt(Into, Ctx, Expr::SE_AllowSideEffects))
+      return false;
+    if (Into.isNegative() || !Into.isIntN(BitsInSizeT))
+      return false;
+    Into = Into.zextOrSelf(BitsInSizeT);
+    return true;
+  };
+
+  APSInt SizeOfElem;
+  if (!EvaluateAsSizeT(Call->getArg(SizeArgNo), SizeOfElem))
+    return false;
+
+  if (!AllocSize->getNumElemsParam()) {
+    Result = CharUnits::fromQuantity(SizeOfElem.getZExtValue());
+    return true;
+  }
+
+  APSInt NumberOfElems;
+  // Argument numbers start at 1
+  unsigned NumArgNo = AllocSize->getNumElemsParam() - 1;
+  if (!EvaluateAsSizeT(Call->getArg(NumArgNo), NumberOfElems))
+    return false;
+
+  bool Overflow;
+  llvm::APInt BytesAvailable = SizeOfElem.umul_ov(NumberOfElems, Overflow);
+  if (Overflow)
+    return false;
+
+  Result = CharUnits::fromQuantity(BytesAvailable.getZExtValue());
+  return true;
+}
+
+/// \brief Convenience function. LVal's base must be a call to an alloc_size
+/// function.
+static bool getBytesReturnedByAllocSizeCall(const ASTContext &Ctx,
+                                            const LValue &LVal,
+                                            CharUnits &Result) {
+  assert(isBaseAnAllocSizeCall(LVal.getLValueBase()) &&
+         "Can't get the size of a non alloc_size function");
+  const auto *Base = LVal.getLValueBase().get<const Expr *>();
+  const CallExpr *CE = tryUnwrapAllocSizeCall(Base);
+  return getBytesReturnedByAllocSizeCall(Ctx, CE, Result);
+}
+
+/// \brief Attempts to evaluate the given LValueBase as the result of a call to
+/// a function with the alloc_size attribute. If it was possible to evaluate the
+/// LValue as such, this function will return true, make Result's Base point to
+/// said function call, and mark Result's Base as invalid.
+static bool evaluateLValueAsAllocSize(EvalInfo &Info, APValue::LValueBase Base,
+                                      LValue &Result) {
+  if (!Info.allowInvalidBaseExpr() || Base.isNull())
+    return false;
+
+  // Because we do no form of static analysis, we only support const variables.
+  //
+  // Additionally, we can't support parameters, nor can we support static
+  // variables (in the latter case, use-before-assign isn't UB; in the former,
+  // we have no clue what they'll be assigned to).
+  const auto *VD =
+      dyn_cast_or_null<VarDecl>(Base.dyn_cast<const ValueDecl *>());
+  if (!VD || !VD->isLocalVarDecl() || !VD->getType().isConstQualified())
+    return false;
+
+  const Expr *Init = VD->getAnyInitializer();
+  if (!Init)
+    return false;
+
+  const Expr *E = Init->IgnoreParens();
+  if (!tryUnwrapAllocSizeCall(E))
+    return false;
+
+  // Store E instead of E unwrapped so that the type of the LValue's base is
+  // what the user wanted.
+  Result.setInvalid(E);
+
+  QualType Pointee = E->getType()->castAs<PointerType>()->getPointeeType();
+  Result.addUnsizedArray(Info, Pointee);
+  return true;
+}
+
 namespace {
 class PointerExprEvaluator
   : public ExprEvaluatorBase<PointerExprEvaluator> {
   LValue &Result;
 
   bool Success(const Expr *E) {
     Result.set(E);
     return true;
if (const ConstantArrayType *CAT
           = Info.Ctx.getAsConstantArrayType(SubExpr->getType()))
       Result.addArray(Info, E, CAT);
     else
       Result.Designator.setInvalid();
     return true;
 
   case CK_FunctionToPointerDecay:
     return EvaluateLValue(SubExpr, Result, Info);
+
+  case CK_LValueToRValue: {
+    LValue LVal;
+    if (!EvaluateLValue(E->getSubExpr(), LVal, Info))
+      return false;
+
+    APValue RVal;
+    // Note, we use the subexpression's type in order to retain cv-qualifiers.
+    if (!handleLValueToRValueConversion(Info, E, E->getSubExpr()->getType(),
+                                        LVal, RVal))
+      return evaluateLValueAsAllocSize(Info, LVal.Base, Result);
+    return Success(RVal, E);
+  }
   }
 
   return ExprEvaluatorBaseTy::VisitCastExpr(E);
 }
 
 static CharUnits GetAlignOfType(EvalInfo &Info, QualType T) {
   // C++ [expr.alignof]p3:
   //     When alignof is applied to a reference type, the result is the
if (OffsetResult.Offset.RoundUpToAlignment(Align) != OffsetResult.Offset) {
           << Offset << (unsigned) getExtValue(Alignment);
 
       return false;
     }
 
     return true;
   }
   default:
-    return ExprEvaluatorBaseTy::VisitCallExpr(E);
+    if (ExprEvaluatorBaseTy::VisitCallExpr(E))
+      return true;
+    if (Info.allowInvalidBaseExpr() && getAllocSizeAttr(E)) {
+      Result.setInvalid(E);
+      QualType PointeeTy =
+          E->getType()->castAs<PointerType>()->getPointeeType();
+      Result.addUnsizedArray(Info, PointeeTy);
+      return true;
+    }
+    return false;
   }
 }
 
 //===----------------------------------------------------------------------===//
 // Member Pointer Evaluation
 //===----------------------------------------------------------------------===//
 
 namespace {
if (CastKind != CK_NoOp && CastKind != CK_BitCast &&
     return NoParens;
 
   auto *SubExpr = Cast->getSubExpr();
   if (!SubExpr->getType()->hasPointerRepresentation() || !SubExpr->isRValue())
     return NoParens;
   return ignorePointerCastsAndParens(SubExpr);
 }
 
+/// Checks if FD is the last FieldDecl (with respect to layout) in its
+/// container.
+bool isLastFieldDecl(const ASTContext &Ctx, const FieldDecl *FD) {
+  if (FD->getParent()->isUnion())
+    return true;
+  const ASTRecordLayout &Layout = Ctx.getASTRecordLayout(FD->getParent());
+  return FD->getFieldIndex() + 1 == Layout.getFieldCount();
+}
+
 /// Checks to see if the given LValue's Designator is at the end of the LValue's
 /// record layout. e.g.
 ///   struct { struct { int a, b; } fst, snd; } obj;
 ///   obj.fst   // no
 ///   obj.snd   // yes
 ///   obj.fst.a // no
 ///   obj.fst.b // no
 ///   obj.snd.a // no
if (auto *FD = dyn_cast<FieldDecl>(ME->getMemberDecl())) {
         return false;
     } else if (auto *IFD = dyn_cast<IndirectFieldDecl>(ME->getMemberDecl())) {
       for (auto *FD : IFD->chain())
         if (!IsLastFieldDecl(cast<FieldDecl>(FD)))
           return false;
     }
   }
 
+  unsigned I = 0;
   QualType BaseType = getType(Base);
-  for (int I = 0, E = LVal.Designator.Entries.size(); I != E; ++I) {
+  if (LVal.Designator.FirstEntryIsAnUnsizedArray) {
+    assert(isBaseAnAllocSizeCall(Base) &&
+           "Unsized array in non-alloc_size call?");
+    // If this is an alloc_size base, we simply ignore the array index
+    ++I;
+    BaseType = BaseType->castAs<PointerType>()->getPointeeType();
+  }
+
+  for (unsigned E = LVal.Designator.Entries.size(); I != E; ++I) {
+    const auto &Entry = LVal.Designator.Entries[I];
     if (BaseType->isArrayType()) {
       // Because __builtin_object_size treats arrays as objects, we can ignore
       // the index iff this is the last array in the Designator.
       if (I + 1 == E)
         return true;
-      auto *CAT = cast<ConstantArrayType>(Ctx.getAsArrayType(BaseType));
-      uint64_t Index = LVal.Designator.Entries[I].ArrayIndex;
+      const auto *CAT = cast<ConstantArrayType>(Ctx.getAsArrayType(BaseType));
+      uint64_t Index = Entry.ArrayIndex;
       if (Index + 1 != CAT->getSize())
         return false;
       BaseType = CAT->getElementType();
     } else if (BaseType->isAnyComplexType()) {
-      auto *CT = BaseType->castAs<ComplexType>();
-      uint64_t Index = LVal.Designator.Entries[I].ArrayIndex;
+      const auto *CT = BaseType->castAs<ComplexType>();
+      uint64_t Index = Entry.ArrayIndex;
       if (Index != 1)
         return false;
       BaseType = CT->getElementType();
-    } else if (auto *FD = getAsField(LVal.Designator.Entries[I])) {
+    } else if (const FieldDecl *FD = getAsField(Entry)) {
       if (!IsLastFieldDecl(FD))
         return false;
       BaseType = FD->getType();
     } else {
-      assert(getAsBaseClass(LVal.Designator.Entries[I]) != nullptr &&
+      assert(getAsBaseClass(Entry) != nullptr &&
              "Expecting cast to a base class");
       return false;
     }
   }
   return true;
 }
 
-/// Tests to see if the LValue has a designator (that isn't necessarily valid).
+/// Tests to see if the LValue has a user-specified designator (that isn't
+/// necessarily valid). Note that this always returns 'true' if the LValue has
+/// an unsized array as its first designator entry, because there's currently no
+/// way to tell if the user typed *foo or foo[0].
 static bool refersToCompleteObject(const LValue &LVal) {
   if (LVal.Designator.Invalid || !LVal.Designator.Entries.empty())
     return false;
 
   if (!LVal.InvalidBase)
     return true;
 
-  auto *E = LVal.Base.dyn_cast<const Expr *>();
-  (void)E;
-  assert(E != nullptr && isa<MemberExpr>(E));
+  // If `E` is a MemberExpr, then the first part of the designator is hiding in
+  // the LValueBase.
+  const auto *E = LVal.Base.dyn_cast<const Expr *>();
+  return !E || !isa<MemberExpr>(E);
+}
+
+/// Attempts to detect a user writing into a piece of memory that's impossible
+/// to figure out the size of by just using types.
+static bool isUserWritingOffTheEnd(EvalInfo &Info, const LValue &LVal) {
+  const auto &Designator = LVal.Designator;
+  // Notes:
+  // - Users can only write off of the end when we have an invalid base. Invalid
+  //   bases imply we don't know where the memory came from.
+  // - This idiom is used most often with arrays of size 0 or 1. Giving up on
+  //   anything more than that can hurt accuracy a lot.
+  return LVal.InvalidBase &&
+         Designator.Entries.size() == Designator.MostDerivedPathLength &&
+         Designator.MostDerivedIsArrayElement &&
+         Designator.getMostDerivedArraySize() < 2 &&
+         isDesignatorAtObjectEnd(Info.Ctx, LVal);
+}
+
+/// Helper for TryEvaluateBuiltinObjectSize -- Given an LValue, this will
+/// determine how many bytes exist from the beginning of the object to either
+/// the end of the current subobject, or the end of the object itself, depending
+/// on what the LValue looks like + the value of Type.
+///
+/// If this returns false, the value of Result is undefined.
+static bool determineEndOffset(EvalInfo &Info, SourceLocation ExprLoc,
+                               unsigned Type, const LValue &LVal,
+                               CharUnits &Result) {
+  // In the case where we're not dealing with a subobject in the expression, we
+  // can ignore the subobject bit.
+  //
+  // Additionally, as a special case, we can deal with LValues with a single
+  // array index in their designator as a complete object. (This simplifies
+  // subobject handling logic)
+  bool EvaluateAsCompleteObject = (Type & 1) == 0 ||
+                                  refersToCompleteObject(LVal) ||
+                                  LVal.Designator.isMostDerivedAnUnsizedArray();
+
+  // We can't give a correct lower bound for Type=3 if the designator is invalid
+  // and we're meant to be evaluating it. Give up.
+  if (Type == 3 && LVal.Designator.Invalid && !EvaluateAsCompleteObject)
+    return false;
+
+  // We want to evaluate the size of the entire object. This is a valid fallback
+  // for when Type=1 and the designator is invalid, because we're asked for an
+  // upper-bound.
+  if (LVal.Designator.Invalid || EvaluateAsCompleteObject) {
+    if (isBaseAnAllocSizeCall(LVal.getLValueBase()) &&
+        getBytesReturnedByAllocSizeCall(Info.Ctx, LVal, Result))
+      return true;
+
+    if (LVal.InvalidBase)
+      return false;
+
+    QualType BaseTy = getObjectType(LVal.getLValueBase());
+    if (BaseTy.isNull())
       return false;
+
+    return HandleSizeof(Info, ExprLoc, BaseTy, Result);
   }
 
-/// Tries to evaluate the __builtin_object_size for @p E. If successful, returns
-/// true and stores the result in @p Size.
-///
-/// If @p WasError is non-null, this will report whether the failure to evaluate
-/// is to be treated as an Error in IntExprEvaluator.
-static bool tryEvaluateBuiltinObjectSize(const Expr *E, unsigned Type,
-                                         EvalInfo &Info, uint64_t &Size,
-                                         bool *WasError = nullptr) {
-  if (WasError != nullptr)
-    *WasError = false;
+  // We want to evaluate the size of a subobject.
+  const auto &Designator = LVal.Designator;
+
+  // The following is a moderately common idiom in C:
+  //
+  // struct Foo { int a; char c[1]; };
+  // struct Foo *F = (struct Foo *)malloc(sizeof(struct Foo) + strlen(Bar));
+  // strcpy(&F->c[0], Bar);
+  if (isUserWritingOffTheEnd(Info, LVal)) {
+    // If we can resolve this to an alloc_size call, we can hand that back,
+    // because we know for certain how many bytes there are to write to.
+    if (isBaseAnAllocSizeCall(LVal.getLValueBase()) &&
+        getBytesReturnedByAllocSizeCall(Info.Ctx, LVal, Result))
+      return true;
 
-  auto Error = [&](const Expr *E) {
-    if (WasError != nullptr)
-      *WasError = true;
+    // If we cannot determine the size of the initial allocation, then we need
+    // to give up for Type=1; we can't accurately determine an upper-bound.
+    // OTOH, we can continue for Type=3, because the size of the most derived
+    // type (or array) is still guaranteed to be there (read: we explicitly
+    // don't support cases of "you've allocated < sizeof(Type) bytes").
+    if (Type == 1)
+      return false;
+  }
+
+  CharUnits BytesPerElem;
+  if (!HandleSizeof(Info, ExprLoc, Designator.MostDerivedType, BytesPerElem))
     return false;
-  };
 
-  auto Success = [&](uint64_t S, const Expr *E) {
-    Size = S;
+  // According to the GCC documentation, we want the size of the subobject
+  // denoted by the pointer. But that's not quite right -- what we actually
+  // want is the size of the immediately-enclosing array, if there is one.
+  int64_t ElemsRemaining = 1;
+  if (Designator.MostDerivedIsArrayElement &&
+      Designator.Entries.size() == Designator.MostDerivedPathLength) {
+    ElemsRemaining = Designator.getMostDerivedArraySize() -
+                     Designator.Entries.back().ArrayIndex;
+  } else if (Designator.isOnePastTheEnd())
+    ElemsRemaining = 0;
+
+  if (ElemsRemaining <= 0) {
+    // We've gone passed the end -- pretend there were zero bytes.
+    Result = CharUnits::Zero();
     return true;
-  };
+  }
+
+  Result = LVal.getLValueOffset() + BytesPerElem * ElemsRemaining;
+  return true;
+}
 
+/// \brief Tries to evaluate the __builtin_object_size for @p E. If successful,
+/// returns true and stores the result in @p Size.
+///
+/// If @p WasError is non-null, this will report whether the failure to evaluate
+/// is to be treated as an Error in IntExprEvaluator.
+static bool tryEvaluateBuiltinObjectSize(const Expr *E, unsigned Type,
+                                         EvalInfo &Info, uint64_t &Size) {
   // Determine the denoted object.
-  LValue Base;
+  LValue LVal;
   {
     // The operand of __builtin_object_size is never evaluated for side-effects.
     // If there are any, but we can determine the pointed-to object anyway, then
     // ignore the side-effects.
     SpeculativeEvaluationRAII SpeculativeEval(Info);
-    FoldOffsetRAII Fold(Info, Type & 1);
+    FoldOffsetRAII Fold(Info);
 
     if (E->isGLValue()) {
       // It's possible for us to be given GLValues if we're called via
       // Expr::tryEvaluateObjectSize.
       APValue RVal;
       if (!EvaluateAsRValue(Info, E, RVal))
         return false;
-      Base.setFrom(Info.Ctx, RVal);
-    } else if (!EvaluatePointer(ignorePointerCastsAndParens(E), Base, Info))
+      LVal.setFrom(Info.Ctx, RVal);
+    } else if (!EvaluatePointer(ignorePointerCastsAndParens(E), LVal, Info))
       return false;
   }
 
-  CharUnits BaseOffset = Base.getLValueOffset();
   // If we point to before the start of the object, there are no accessible
   // bytes.
-  if (BaseOffset.isNegative())
-    return Success(0, E);
-
-  // In the case where we're not dealing with a subobject, we discard the
-  // subobject bit.
-  bool SubobjectOnly = (Type & 1) != 0 && !refersToCompleteObject(Base);
-
-  // If Type & 1 is 0, we need to be able to statically guarantee that the bytes
-  // exist. If we can't verify the base, then we can't do that.
-  //
-  // As a special case, we produce a valid object size for an unknown object
-  // with a known designator if Type & 1 is 1. For instance:
-  //
-  //   extern struct X { char buff[32]; int a, b, c; } *p;
-  //   int a = __builtin_object_size(p->buff + 4, 3); // returns 28
-  //   int b = __builtin_object_size(p->buff + 4, 2); // returns 0, not 40
-  //
-  // This matches GCC's behavior.
-  if (Base.InvalidBase && !SubobjectOnly)
-    return Error(E);
-
-  // If we're not examining only the subobject, then we reset to a complete
-  // object designator
-  //
-  // If Type is 1 and we've lost track of the subobject, just find the complete
-  // object instead. (If Type is 3, that's not correct behavior and we should
-  // return 0 instead.)
-  LValue End = Base;
-  if (!SubobjectOnly || (End.Designator.Invalid && Type == 1)) {
-    QualType T = getObjectType(End.getLValueBase());
-    if (T.isNull())
-      End.Designator.setInvalid();
-    else {
-      End.Designator = SubobjectDesignator(T);
-      End.Offset = CharUnits::Zero();
-    }
-  }
-
-  // If it is not possible to determine which objects ptr points to at compile
-  // time, __builtin_object_size should return (size_t) -1 for type 0 or 1
-  // and (size_t) 0 for type 2 or 3.
-  if (End.Designator.Invalid)
-    return false;
-
-  // According to the GCC documentation, we want the size of the subobject
-  // denoted by the pointer. But that's not quite right -- what we actually
-  // want is the size of the immediately-enclosing array, if there is one.
-  int64_t AmountToAdd = 1;
-  if (End.Designator.MostDerivedIsArrayElement &&
-      End.Designator.Entries.size() == End.Designator.MostDerivedPathLength) {
-    // We got a pointer to an array. Step to its end.
-    AmountToAdd = End.Designator.MostDerivedArraySize -
-                  End.Designator.Entries.back().ArrayIndex;
-  } else if (End.Designator.isOnePastTheEnd()) {
-    // We're already pointing at the end of the object.
-    AmountToAdd = 0;
+  if (LVal.getLValueOffset().isNegative()) {
+    Size = 0;
+    return true;
   }
 
-  QualType PointeeType = End.Designator.MostDerivedType;
-  assert(!PointeeType.isNull());
-  if (PointeeType->isIncompleteType() || PointeeType->isFunctionType())
-    return Error(E);
-
-  if (!HandleLValueArrayAdjustment(Info, E, End, End.Designator.MostDerivedType,
-                                   AmountToAdd))
+  CharUnits EndOffset;
+  if (!determineEndOffset(Info, E->getExprLoc(), Type, LVal, EndOffset))
     return false;
 
-  auto EndOffset = End.getLValueOffset();
-
-  // The following is a moderately common idiom in C:
-  //
-  // struct Foo { int a; char c[1]; };
-  // struct Foo *F = (struct Foo *)malloc(sizeof(struct Foo) + strlen(Bar));
-  // strcpy(&F->c[0], Bar);
-  //
-  // So, if we see that we're examining a 1-length (or 0-length) array at the
-  // end of a struct with an unknown base, we give up instead of breaking code
-  // that behaves this way. Note that we only do this when Type=1, because
-  // Type=3 is a lower bound, so answering conservatively is fine.
-  if (End.InvalidBase && SubobjectOnly && Type == 1 &&
-      End.Designator.Entries.size() == End.Designator.MostDerivedPathLength &&
-      End.Designator.MostDerivedIsArrayElement &&
-      End.Designator.MostDerivedArraySize < 2 &&
-      isDesignatorAtObjectEnd(Info.Ctx, End))
-    return false;
-
-  if (BaseOffset > EndOffset)
-    return Success(0, E);
-
-  return Success((EndOffset - BaseOffset).getQuantity(), E);
+  CharUnits BytesRemaining = EndOffset - LVal.getLValueOffset();
+  Size = std::max(0L, BytesRemaining.getQuantity());
+  return true;
 }
 
 bool IntExprEvaluator::TryEvaluateBuiltinObjectSize(const CallExpr *E,
                                                     unsigned Type) {
   uint64_t Size;
-  bool WasError;
-  if (::tryEvaluateBuiltinObjectSize(E->getArg(0), Type, Info, Size, &WasError))
+  if (::tryEvaluateBuiltinObjectSize(E->getArg(0), Type, Info, Size))
     return Success(Size, E);
-  if (WasError)
-    return Error(E);
   return false;
 }
 
 bool IntExprEvaluator::VisitCallExpr(const CallExpr *E) {
   switch (unsigned BuiltinOp = E->getBuiltinCallee()) {
   default:
     return ExprEvaluatorBaseTy::VisitCallExpr(E);
 
case Builtin::BI__builtin_object_size: {
     // Expression had no side effects, but we couldn't statically determine the
     // size of the referenced object.
     switch (Info.EvalMode) {
     case EvalInfo::EM_ConstantExpression:
     case EvalInfo::EM_PotentialConstantExpression:
     case EvalInfo::EM_ConstantFold:
     case EvalInfo::EM_EvaluateForOverflow:
     case EvalInfo::EM_IgnoreSideEffects:
-    case EvalInfo::EM_DesignatorFold:
+    case EvalInfo::EM_OffsetFold:
       // Leave it to IR generation.
       return Error(E);
     case EvalInfo::EM_ConstantExpressionUnevaluated:
     case EvalInfo::EM_PotentialConstantExpressionUnevaluated:
       // Reduce it to a constant now.
       return Success((Type & 2) ? 0 : -1, E);
     }
   }

lib/Sema/SemaDeclAttr.cpp

static void handleAssertExclusiveLockAttr(Sema &S, Decl *D,
   unsigned Size = Args.size();
   Expr **StartArg = Size == 0 ? nullptr : &Args[0];
   D->addAttr(::new (S.Context)
              AssertExclusiveLockAttr(Attr.getRange(), S.Context,
                                      StartArg, Size,
                                      Attr.getAttributeSpellingListIndex()));
 }
 
+/// \brief Checks to be sure that the given parameter number is inbounds, and is
+/// an some integral type. Will emit appropriate diagnostics if this returns
+/// false.
+///
+/// FuncParamNo is expected to be from the user, so is base-1. AttrArgNo is used
+/// to actually retrieve the argument, so it's base-0.
+static bool checkParamIsIntegerType(Sema &S, const FunctionDecl *FD,
+                                    const AttributeList &Attr,
+                                    unsigned FuncParamNo, unsigned AttrArgNo) {
+  assert(Attr.getArg(AttrArgNo).is<Expr *>());
+  // FuncParamNo is base-1
+  if (FuncParamNo < 1 || FuncParamNo > FD->getNumParams()) {
+    SourceLocation SrcLoc = Attr.getArgAsExpr(AttrArgNo)->getLocStart();
+    unsigned UserArgNo = AttrArgNo + 1;
+    S.Diag(SrcLoc, diag::err_attribute_argument_out_of_range)
+        << Attr.getName() << UserArgNo << FD->getNumParams();
+    return false;
+  }
+
+  const ParmVarDecl *Param = FD->getParamDecl(FuncParamNo - 1);
+  if (!Param->getType()->isIntegerType() && !Param->getType()->isCharType()) {
+    SourceLocation SrcLoc = Attr.getArgAsExpr(AttrArgNo)->getLocStart();
+    S.Diag(SrcLoc, diag::err_attribute_integers_only)
+        << Attr.getName() << Param->getSourceRange();
+    return false;
+  }
+  return true;
+}
+
+/// \brief Wrapper around checkUInt32Argument, with an extra check to be sure
+/// that the result will fit into a regular (signed) int. All args have the same
+/// purpose as they do in checkUInt32Argument.
+static bool checkPositiveIntArgument(Sema &S, const AttributeList &Attr,
+                                     const Expr *Expr, int &Val,
+                                     unsigned Idx = UINT_MAX) {
+  uint32_t UVal;
+  if (!checkUInt32Argument(S, Attr, Expr, UVal, Idx))
+    return false;
+
+  if (UVal > std::numeric_limits<int>::max()) {
+    llvm::APSInt I(32); // for toString
+    I = UVal;
+    S.Diag(Expr->getExprLoc(), diag::err_ice_too_large)
+        << I.toString(10, false) << 32 << /* Unsigned */ 0;
+    return false;
+  }
+
+  Val = (int)UVal;
+  return true;
+}
+
+static void handleAllocSizeAttr(Sema &S, Decl *D, const AttributeList &Attr) {
+  if (!checkAttributeAtLeastNumArgs(S, Attr, 1) ||
+      !checkAttributeAtMostNumArgs(S, Attr, 2))
+    return;
+
+  const auto *FD = cast<FunctionDecl>(D);
+  if (!FD->getReturnType()->isPointerType()) {
+    S.Diag(Attr.getLoc(), diag::warn_attribute_return_pointers_only)
+        << Attr.getName();
+    return;
+  }
+
+  const Expr *SizeExpr = Attr.getArgAsExpr(0);
+  int SizeArgNo;
+  // Paramater indices are 1-based, hence Index=1
+  if (!checkPositiveIntArgument(S, Attr, SizeExpr, SizeArgNo, /*Index=*/1))
+    return;
+
+  if (!checkParamIsIntegerType(S, FD, Attr, SizeArgNo, /*AttrArgNo=*/0))
+    return;
+
+  // Args are 1-based, so 0 implies that the arg was not present
+  int NumberArgNo = 0;
+  if (Attr.getNumArgs() == 2) {
+    const Expr *NumberExpr = Attr.getArgAsExpr(1);
+    // Paramater indices are 1-based, hence Index=2
+    if (!checkPositiveIntArgument(S, Attr, NumberExpr, NumberArgNo,
+                                  /*Index=*/2))
+      return;
+    if (!checkParamIsIntegerType(S, FD, Attr, NumberArgNo, /*AttrArgNo=*/1))
+      return;
+  }
+
+  D->addAttr(::new (S.Context) AllocSizeAttr(
+      Attr.getRange(), S.Context, SizeArgNo, NumberArgNo,
+      Attr.getAttributeSpellingListIndex()));
+}
 
 static bool checkTryLockFunAttrCommon(Sema &S, Decl *D,
                                       const AttributeList &Attr,
                                       SmallVectorImpl<Expr *> &Args) {
   if (!checkAttributeAtLeastNumArgs(S, Attr, 1))
     return false;
 
   if (!isIntOrBool(Attr.getArgAsExpr(0))) {
case AttributeList::AT_Alias:
     handleAliasAttr(S, D, Attr);
     break;
   case AttributeList::AT_Aligned:
     handleAlignedAttr(S, D, Attr);
     break;
   case AttributeList::AT_AlignValue:
     handleAlignValueAttr(S, D, Attr);
     break;
+  case AttributeList::AT_AllocSize:
+    handleAllocSizeAttr(S, D, Attr);
+    break;
   case AttributeList::AT_AlwaysInline:
     handleAlwaysInlineAttr(S, D, Attr);
     break;
   case AttributeList::AT_AnalyzerNoReturn:
     handleAnalyzerNoReturnAttr(S, D, Attr);
     break;
   case AttributeList::AT_TLSModel:
     handleTLSModelAttr(S, D, Attr);

test/CodeGen/alloc-size.c

+// RUN: %clang_cc1 -triple x86_64-apple-darwin -emit-llvm %s -o - 2>&1 | FileCheck %s
+
+#define NULL ((void *)0)
+
+int gi;
+
+typedef unsigned long size_t;
+
+void *my_malloc(size_t) __attribute__((alloc_size(1)));
+void *my_calloc(size_t, size_t)
+    __attribute__((alloc_size(1, 2)));
+
+// CHECK-LABEL: @test1
+void test1() {
+  void *const vp = my_malloc(100);
+  // CHECK: store i32 100
+  gi = __builtin_object_size(vp, 0);
+  // CHECK: store i32 100
+  gi = __builtin_object_size(vp, 1);
+  // CHECK: store i32 100
+  gi = __builtin_object_size(vp, 2);
+  // CHECK: store i32 100
+  gi = __builtin_object_size(vp, 3);
+
+  void *const arr = my_calloc(100, 5);
+  // CHECK: store i32 500
+  gi = __builtin_object_size(arr, 0);
+  // CHECK: store i32 500
+  gi = __builtin_object_size(arr, 1);
+  // CHECK: store i32 500
+  gi = __builtin_object_size(arr, 2);
+  // CHECK: store i32 500
+  gi = __builtin_object_size(arr, 3);
+
+  // CHECK: store i32 100
+  gi = __builtin_object_size(my_malloc(100), 0);
+  // CHECK: store i32 100
+  gi = __builtin_object_size(my_malloc(100), 1);
+  // CHECK: store i32 100
+  gi = __builtin_object_size(my_malloc(100), 2);
+  // CHECK: store i32 100
+  gi = __builtin_object_size(my_malloc(100), 3);
+
+  // CHECK: store i32 500
+  gi = __builtin_object_size(my_calloc(100, 5), 0);
+  // CHECK: store i32 500
+  gi = __builtin_object_size(my_calloc(100, 5), 1);
+  // CHECK: store i32 500
+  gi = __builtin_object_size(my_calloc(100, 5), 2);
+  // CHECK: store i32 500
+  gi = __builtin_object_size(my_calloc(100, 5), 3);
+
+  void *const zeroPtr = my_malloc(0);
+  // CHECK: store i32 0
+  gi = __builtin_object_size(zeroPtr, 0);
+  // CHECK: store i32 0
+  gi = __builtin_object_size(my_malloc(0), 0);
+
+  void *const zeroArr1 = my_calloc(0, 1);
+  void *const zeroArr2 = my_calloc(1, 0);
+  // CHECK: store i32 0
+  gi = __builtin_object_size(zeroArr1, 0);
+  // CHECK: store i32 0
+  gi = __builtin_object_size(zeroArr2, 0);
+  // CHECK: store i32 0
+  gi = __builtin_object_size(my_calloc(1, 0), 0);
+  // CHECK: store i32 0
+  gi = __builtin_object_size(my_calloc(0, 1), 0);
+}
+
+// CHECK-LABEL: @test2
+void test2() {
+  void *const vp = my_malloc(gi);
+  // CHECK: @llvm.objectsize
+  gi = __builtin_object_size(vp, 0);
+
+  void *const arr1 = my_calloc(gi, 1);
+  // CHECK: @llvm.objectsize
+  gi = __builtin_object_size(arr1, 0);
+
+  void *const arr2 = my_calloc(1, gi);
+  // CHECK: @llvm.objectsize
+  gi = __builtin_object_size(arr2, 0);
+}
+
+// CHECK-LABEL: @test3
+void test3() {
+  char *const buf = (char *)my_calloc(100, 5);
+  // CHECK: store i32 500
+  gi = __builtin_object_size(buf, 0);
+  // CHECK: store i32 500
+  gi = __builtin_object_size(buf, 1);
+  // CHECK: store i32 500
+  gi = __builtin_object_size(buf, 2);
+  // CHECK: store i32 500
+  gi = __builtin_object_size(buf, 3);
+}
+
+struct Data {
+  int a;
+  int t[10];
+  char pad[3];
+  char end[1];
+};
+
+// CHECK-LABEL: @test5
+void test5() {
+  struct Data *const data = my_malloc(sizeof(*data));
+  // CHECK: store i32 48
+  gi = __builtin_object_size(data, 0);
+  // CHECK: store i32 48
+  gi = __builtin_object_size(data, 1);
+  // CHECK: store i32 48
+  gi = __builtin_object_size(data, 2);
+  // CHECK: store i32 48
+  gi = __builtin_object_size(data, 3);
+
+  // CHECK: store i32 40
+  gi = __builtin_object_size(&data->t[1], 0);
+  // CHECK: store i32 36
+  gi = __builtin_object_size(&data->t[1], 1);
+  // CHECK: store i32 40
+  gi = __builtin_object_size(&data->t[1], 2);
+  // CHECK: store i32 36
+  gi = __builtin_object_size(&data->t[1], 3);
+
+  struct Data *const arr = my_calloc(sizeof(*data), 2);
+  // CHECK: store i32 96
+  gi = __builtin_object_size(arr, 0);
+  // CHECK: store i32 96
+  gi = __builtin_object_size(arr, 1);
+  // CHECK: store i32 96
+  gi = __builtin_object_size(arr, 2);
+  // CHECK: store i32 96
+  gi = __builtin_object_size(arr, 3);
+
+  // CHECK: store i32 88
+  gi = __builtin_object_size(&arr->t[1], 0);
+  // CHECK: store i32 36
+  gi = __builtin_object_size(&arr->t[1], 1);
+  // CHECK: store i32 88
+  gi = __builtin_object_size(&arr->t[1], 2);
+  // CHECK: store i32 36
+  gi = __builtin_object_size(&arr->t[1], 3);
+}
+
+// CHECK-LABEL: @test6
+void test6() {
+  // Things that would normally trigger conservative estimates don't need to do
+  // so when we know the source of the allocation.
+  struct Data *const data = my_malloc(sizeof(*data) + 10);
+  // CHECK: store i32 11
+  gi = __builtin_object_size(data->end, 0);
+  // CHECK: store i32 11
+  gi = __builtin_object_size(data->end, 1);
+  // CHECK: store i32 11
+  gi = __builtin_object_size(data->end, 2);
+  // CHECK: store i32 11
+  gi = __builtin_object_size(data->end, 3);
+
+  struct Data *const arr = my_calloc(sizeof(*arr) + 5, 3);
+  // AFAICT, GCC treats malloc and calloc identically. So, we should do the
+  // same.
+  //
+  // Additionally, GCC ignores the initial array index when determining whether
+  // we're writing off the end of an alloc_size base. e.g.
+  //   arr[0].end
+  //   arr[1].end
+  //   arr[2].end
+  // ...Are all considered "writing off the end", because there's no way to tell
+  // with high accuracy if the user meant "allocate a single N-byte `Data`",
+  // or "allocate M smaller `Data`s with extra padding".
+
+  // CHECK: store i32 112
+  gi = __builtin_object_size(arr->end, 0);
+  // CHECK: store i32 112
+  gi = __builtin_object_size(arr->end, 1);
+  // CHECK: store i32 112
+  gi = __builtin_object_size(arr->end, 2);
+  // CHECK: store i32 112
+  gi = __builtin_object_size(arr->end, 3);
+
+  // CHECK: store i32 112
+  gi = __builtin_object_size(arr[0].end, 0);
+  // CHECK: store i32 112
+  gi = __builtin_object_size(arr[0].end, 1);
+  // CHECK: store i32 112
+  gi = __builtin_object_size(arr[0].end, 2);
+  // CHECK: store i32 112
+  gi = __builtin_object_size(arr[0].end, 3);
+
+  // CHECK: store i32 64
+  gi = __builtin_object_size(arr[1].end, 0);
+  // CHECK: store i32 64
+  gi = __builtin_object_size(arr[1].end, 1);
+  // CHECK: store i32 64
+  gi = __builtin_object_size(arr[1].end, 2);
+  // CHECK: store i32 64
+  gi = __builtin_object_size(arr[1].end, 3);
+
+  // CHECK: store i32 16
+  gi = __builtin_object_size(arr[2].end, 0);
+  // CHECK: store i32 16
+  gi = __builtin_object_size(arr[2].end, 1);
+  // CHECK: store i32 16
+  gi = __builtin_object_size(arr[2].end, 2);
+  // CHECK: store i32 16
+  gi = __builtin_object_size(arr[2].end, 3);
+}
+
+// CHECK-LABEL: @test7
+void test7() {
+  struct Data *const data = my_malloc(sizeof(*data) + 5);
+  // CHECK: store i32 9
+  gi = __builtin_object_size(data->pad, 0);
+  // CHECK: store i32 3
+  gi = __builtin_object_size(data->pad, 1);
+  // CHECK: store i32 9
+  gi = __builtin_object_size(data->pad, 2);
+  // CHECK: store i32 3
+  gi = __builtin_object_size(data->pad, 3);
+}
+
+// CHECK-LABEL: @test8
+void test8() {
+  // Non-const pointers aren't currently supported.
+  void *buf = my_calloc(100, 5);
+  // CHECK: @llvm.objectsize.i64.p0i8(i8* %{{.*}}, i1 false)
+  gi = __builtin_object_size(buf, 0);
+  // CHECK: @llvm.objectsize
+  gi = __builtin_object_size(buf, 1);
+  // CHECK: @llvm.objectsize
+  gi = __builtin_object_size(buf, 2);
+  // CHECK: store i32 0
+  gi = __builtin_object_size(buf, 3);
+}
+
+// CHECK-LABEL: @test9
+void test9() {
+  // Check to be sure that we unwrap things correctly.
+  short *const buf0 = (my_malloc(100));
+  short *const buf1 = (short*)(my_malloc(100));
+  short *const buf2 = ((short*)(my_malloc(100)));
+
+  // CHECK: store i32 100
+  gi = __builtin_object_size(buf0, 0);
+  // CHECK: store i32 100
+  gi = __builtin_object_size(buf1, 0);
+  // CHECK: store i32 100
+  gi = __builtin_object_size(buf2, 0);
+}
+
+// CHECK-LABEL: @test10
+void test10() {
+  // Yay overflow
+  short *const buf = my_calloc((size_t)-1 / 2 + 1, 2);
+  // CHECK: @llvm.objectsize
+  gi = __builtin_object_size(buf, 0);
+  // CHECK: @llvm.objectsize
+  gi = __builtin_object_size(buf, 1);
+  // CHECK: @llvm.objectsize
+  gi = __builtin_object_size(buf, 2);
+  // CHECK: store i32 0
+  gi = __builtin_object_size(buf, 3);
+}
+
+void *my_tiny_malloc(char) __attribute__((alloc_size(1)));
+void *my_tiny_calloc(char, char) __attribute__((alloc_size(1, 2)));
+
+// CHECK-LABEL: @test11
+void test11() {
+  void *const vp = my_tiny_malloc(100);
+  // CHECK: store i32 100
+  gi = __builtin_object_size(vp, 0);
+  // CHECK: store i32 100
+  gi = __builtin_object_size(vp, 1);
+  // CHECK: store i32 100
+  gi = __builtin_object_size(vp, 2);
+  // CHECK: store i32 100
+  gi = __builtin_object_size(vp, 3);
+
+  // N.B. This causes char overflow, but not size_t overflow, so it should be
+  // supported.
+  void *const arr = my_tiny_calloc(100, 5);
+  // CHECK: store i32 500
+  gi = __builtin_object_size(arr, 0);
+  // CHECK: store i32 500
+  gi = __builtin_object_size(arr, 1);
+  // CHECK: store i32 500
+  gi = __builtin_object_size(arr, 2);
+  // CHECK: store i32 500
+  gi = __builtin_object_size(arr, 3);
+
+  // CHECK: store i32 100
+  gi = __builtin_object_size(my_malloc(100), 0);
+  // CHECK: store i32 100
+  gi = __builtin_object_size(my_malloc(100), 1);
+  // CHECK: store i32 100
+  gi = __builtin_object_size(my_malloc(100), 2);
+  // CHECK: store i32 100
+  gi = __builtin_object_size(my_malloc(100), 3);
+
+  // CHECK: store i32 500
+  gi = __builtin_object_size(my_calloc(100, 5), 0);
+  // CHECK: store i32 500
+  gi = __builtin_object_size(my_calloc(100, 5), 1);
+  // CHECK: store i32 500
+  gi = __builtin_object_size(my_calloc(100, 5), 2);
+  // CHECK: store i32 500
+  gi = __builtin_object_size(my_calloc(100, 5), 3);
+}
+
+void *my_signed_malloc(long) __attribute__((alloc_size(1)));
+void *my_signed_calloc(long, long) __attribute__((alloc_size(1, 2)));
+
+// CHECK-LABEL: @test12
+void test12() {
+  // CHECK: store i32 100
+  gi = __builtin_object_size(my_signed_malloc(100), 0);
+  // CHECK: store i32 500
+  gi = __builtin_object_size(my_signed_calloc(100, 5), 0);
+
+  void *const vp = my_signed_malloc(-2);
+  // CHECK: @llvm.objectsize
+  gi = __builtin_object_size(vp, 0);
+  // N.B. These get lowered to -1 because the function calls may have
+  // side-effects, and we can't determine the objectsize.
+  // CHECK: store i32 -1
+  gi = __builtin_object_size(my_signed_malloc(-2), 0);
+
+  void *const arr1 = my_signed_calloc(-2, 1);
+  void *const arr2 = my_signed_calloc(1, -2);
+  // CHECK: @llvm.objectsize
+  gi = __builtin_object_size(arr1, 0);
+  // CHECK: @llvm.objectsize
+  gi = __builtin_object_size(arr2, 0);
+  // CHECK: store i32 -1
+  gi = __builtin_object_size(my_signed_calloc(1, -2), 0);
+  // CHECK: store i32 -1
+  gi = __builtin_object_size(my_signed_calloc(-2, 1), 0);
+}

test/CodeGenCXX/global-init.cpp

 
 struct C { void *field; };
 
 struct D { ~D(); };
 
 // CHECK: @__dso_handle = external global i8
 // CHECK: @c = global %struct.C zeroinitializer, align 8
 
-// It's okay if we ever implement the IR-generation optimization to remove this.
-// CHECK: @_ZN5test3L3varE = internal constant i8* getelementptr inbounds ([7 x i8], [7 x i8]* 
-
 // PR6205: The casts should not require global initializers
 // CHECK: @_ZN6PR59741cE = external global %"struct.PR5974::C"
 // CHECK: @_ZN6PR59741aE = global %"struct.PR5974::A"* getelementptr inbounds (%"struct.PR5974::C", %"struct.PR5974::C"* @_ZN6PR59741cE, i32 0, i32 0)
 // CHECK: @_ZN6PR59741bE = global %"struct.PR5974::B"* bitcast (i8* getelementptr (i8, i8* bitcast (%"struct.PR5974::C"* @_ZN6PR59741cE to i8*), i64 4) to %"struct.PR5974::B"*), align 8
 
 // CHECK: call void @_ZN1AC1Ev(%struct.A* @a)
 // CHECK: call i32 @__cxa_atexit(void (i8*)* bitcast (void (%struct.A*)* @_ZN1AD1Ev to void (i8*)*), i8* getelementptr inbounds (%struct.A, %struct.A* @a, i32 0, i32 0), i8* @__dso_handle)
 A a;

test/Sema/alloc-size.c

+// RUN: %clang_cc1 %s -verify
+
+void *fail1(int a) __attribute__((alloc_size)); //expected-error{{'alloc_size' attribute takes at least 1 argument}}
+void *fail2(int a) __attribute__((alloc_size())); //expected-error{{'alloc_size' attribute takes at least 1 argument}}
+
+void *fail3(int a) __attribute__((alloc_size(0))); //expected-error{{'alloc_size' attribute parameter 1 is out of bounds: can only be 1, since there is one parameter}}
+void *fail4(int a) __attribute__((alloc_size(2))); //expected-error{{'alloc_size' attribute parameter 1 is out of bounds: can only be 1, since there is one parameter}}
+
+void *fail5(int a, int b) __attribute__((alloc_size(0, 1))); //expected-error{{'alloc_size' attribute parameter 1 is out of bounds: must be between 1 and 2}}
+void *fail6(int a, int b) __attribute__((alloc_size(3, 1))); //expected-error{{'alloc_size' attribute parameter 1 is out of bounds: must be between 1 and 2}}
+
+void *fail7(int a, int b) __attribute__((alloc_size(1, 0))); //expected-error{{'alloc_size' attribute parameter 2 is out of bounds: must be between 1 and 2}}
+void *fail8(int a, int b) __attribute__((alloc_size(1, 3))); //expected-error{{'alloc_size' attribute parameter 2 is out of bounds: must be between 1 and 2}}
+
+int fail9(int a) __attribute__((alloc_size(1))); //expected-warning{{'alloc_size' attribute only applies to return values that are pointers}}
+
+int fail10 __attribute__((alloc_size(1))); //expected-warning{{'alloc_size' attribute only applies to non-K&R-style functions}}
+
+void *fail11(void *a) __attribute__((alloc_size(1))); //expected-error{{'alloc_size' attribute argument may only refer to a function parameter of integer type}}
+
+void *fail12(int a) __attribute__((alloc_size("abc"))); //expected-error{{'alloc_size' attribute requires parameter 1 to be an integer constant}}
+void *fail12(int a) __attribute__((alloc_size(1, "abc"))); //expected-error{{'alloc_size' attribute requires parameter 2 to be an integer constant}}
+void *fail13(int a) __attribute__((alloc_size(1U<<31))); //expected-error{{integer constant expression evaluates to value 2147483648 that cannot be represented in a 32-bit signed integer type}}

test/SemaCXX/constant-expression-cxx11.cpp

 int n3 = 37; // expected-note {{declared here}}
 
 constexpr int m1 = n1; // expected-error {{constant expression}} expected-note {{read of volatile-qualified type 'const volatile int'}}
 constexpr int m2 = n2; // expected-error {{constant expression}} expected-note {{read of volatile-qualified type 'const volatile int'}}
 constexpr int m1b = const_cast<const int&>(n1); // expected-error {{constant expression}} expected-note {{read of volatile object 'n1'}}
 constexpr int m2b = const_cast<const int&>(n2); // expected-error {{constant expression}} expected-note {{read of volatile object 'n2'}}
 
 struct T { int n; };
-const T t = { 42 }; // expected-note {{declared here}}
+const T t = { 42 };
 
 constexpr int f(volatile int &&r) {
   return r; // expected-note {{read of volatile-qualified type 'volatile int'}}
 }
 constexpr int g(volatile int &&r) {
   return const_cast<int&>(r); // expected-note {{read of volatile temporary is not allowed in a constant expression}}
 }
 struct S {
   int j : f(0); // expected-error {{constant expression}} expected-note {{in call to 'f(0)'}}
   int k : g(0); // expected-error {{constant expression}} expected-note {{temporary created here}} expected-note {{in call to 'g(0)'}}
   int l : n3; // expected-error {{constant expression}} expected-note {{read of non-const variable}}
-  int m : t.n; // expected-error {{constant expression}} expected-note {{read of non-constexpr variable}}
+  int m : t.n; // expected-warning{{width of bit-field 'm' (42 bits)}}
 };
 
 }
 
 namespace ExternConstexpr {
   extern constexpr int n = 0;
   extern constexpr int m; // expected-error {{constexpr variable declaration must be a definition}}
   void f() {