Implement VectorType conditional operator GNU extension.

[erichkeane]

Summary

GCC supports the conditional operator on VectorTypes that acts as a
'select' in C++ mode. This patch implements the support. Types are
converted as closely to GCC's behavior as possible, though in a few
places consistency with our existing vector type support was preferred.

Note that this implementation is different from the OpenCL version in a
number of ways, so it unfortunately required a different implementation.

First, the SEMA rules and promotion rules are significantly different.

Secondly, GCC implements COND[i] != 0 ? LHS[i] : RHS[i] (where i is in
the range 0- VectorSize, for each element). In OpenCL, the condition is
COND[i] < 0 ? LHS[i]: RHS[i].

In the process of implementing this, it was also required to make the
expression COND ? LHS : RHS type dependent if COND is type dependent,
since the type is now dependent on the condition. For example:

T ? 1 : 2;

Is not typically type dependent, since the result can be deduced from
the operands. HOWEVER, if T is a VectorType now, it could change this
to a 'select' (basically a swizzle with a non-constant mask) with the 1
and 2 being promoted to vectors themselves.

While this is a change, it is NOT a standards incompatible change. Based
on my (and D. Gregor's, at the time of writing the code) reading of the
standard, the expression is supposed to be type dependent if ANY
sub-expression is type dependent.

[erichkeane]

Sorry in advance for the reviewer choices... Blame got me you all for all touching these things in the early 2010s. Since then, Richard and I seem to be the only one to have touched these types.

[efriedma]

Can you update https://clang.llvm.org/docs/LanguageExtensions.html#langext-vectors with a description of the interaction between the different language features here?

It's unfortunate that the gcc extension is explicitly incompatible with OpenCL, but I guess we're stuck with it.

clang/include/clang/AST/Expr.h

3738

You can just refer to [temp.dep.expr] here; the type does in fact "depend" on the type of the condition. I'm a little concerned that the C++ standard might not preserve this guarantee in the future.

clang/lib/Sema/SemaExpr.cpp

8981 ↗

(On Diff #233791)

Can you separate this out into an independent patch?

clang/lib/Sema/SemaExprCXX.cpp

5826

I'm not completely sure this works the way you want it to... in particular, if the types don't match, and both operands can be promoted to int, you end up with an int vector.

[erichkeane]

In D71463#1783953, @efriedma wrote:

Can you update https://clang.llvm.org/docs/LanguageExtensions.html#langext-vectors with a description of the interaction between the different language features here?

Done, see the next patch.

It's unfortunate that the gcc extension is explicitly incompatible with OpenCL, but I guess we're stuck with it.

I think it is the other way around actually. The GCC extension predates the OpenCL one from what I can tell.

clang/include/clang/AST/Expr.h

3738

I wasn't able to find a bug report in the core issues list, and this has been known about by Doug for ~15 years now. I'm somewhat less concerned, particularly since GCC has this issue as well.

clang/lib/Sema/SemaExpr.cpp

8981 ↗

(On Diff #233791)

348f22eac83d9a3ee946e41be43fe507f04a89b6

clang/lib/Sema/SemaExprCXX.cpp

5826

I think I'm OK with that... GCC does allow some promotions here (a short and a ushort act as an int). I tried a bunch of cases and couldn't come up with a case where this didn't work in the 2 scalar versions. You wouldn't happen to have something in mind, would you?

[erichkeane]

Rebase + @eli.friedman s comments.

[erichkeane]

Ping :)

[aaron.ballman]

clang/docs/LanguageExtensions.rst

467–468

Do these columns line up in the actual source file? They don't appear to line up in the Phab viewer, but I don't know if that's an artifact of Phab or not.

clang/include/clang/Basic/DiagnosticSemaKinds.td

6882

We don't actually use enumeral in diagnostics anywhere -- I think that should be enumeration instead.

clang/lib/CodeGen/CGExprScalar.cpp

4301

auto *

Also, should these three variables have identifiers starting with an uppercase like CondV et al?

clang/lib/Sema/SemaExprCXX.cpp

5757

Do you actually need the getTypePtr() bit, or will the cast<> suffice to trigger the cast through Type*?

5761–5762

If these checks are redundant, would you prefer to assert them?

5779

const auto * (same elsewhere in the function where the type is spelled out in the initialization)

5793

Shouldn't this pass in the LHSType instead because that's what's being tested?

5799

And this one do the RHSType?

5805

Do you need to canonicalize these types before comparing them?

[erichkeane]

New patch coming as soon as my build finishes.

clang/docs/LanguageExtensions.rst

467–468

They don't, I had misread the sphinx doc it seems and was doing footnotes incorrectly. A coworker shared the live-sphinx viewer so I corrected this. See next patch.

clang/lib/CodeGen/CGExprScalar.cpp

4301

Probably, I just C&P'ed this code from the above group :/

clang/lib/Sema/SemaExprCXX.cpp

5757

TIL!

5793

Yep!

5805

No, both versions of this function canonicalize for me: https://clang.llvm.org/doxygen/ASTContext_8h_source.html#l02305

[erichkeane]

@aaron.ballman s comments done.

[erichkeane]

@eli.friedman @rjmccall Do you have further comments on these?

[Anastasia]

It looks good, I was just thinking whether it would be possible to share more common infrastructure. There is Sema::CheckVectorOperands that corresponding OpenCL methods are using internally. Do you think it is possible to share the code more?

clang/include/clang/Basic/DiagnosticSemaKinds.td

6880

Would this only apply to GNU extension? I am just wondering if this can be generalized.

clang/lib/Sema/SemaExprCXX.cpp

5797

Don't we already have a diagnostic for this used in OpenCL?

[erichkeane]

In D71463#1797436, @Anastasia wrote:

It looks good, I was just thinking whether it would be possible to share more common infrastructure. There is Sema::CheckVectorOperands that corresponding OpenCL methods are using internally. Do you think it is possible to share the code more?

I tried in my initial implementation. Unfortunately CheckVectorOperands and the OpenCL vectors have significantly different semantic rules from this, these aren't even consistent with the normal GCC vector conversion rules. The similarities/differences between all of these are frustrating, any attempt I made at unifying the logic made both really difficult to follow.

clang/include/clang/Basic/DiagnosticSemaKinds.td

6880

To be honest, I limited the scope of this a bit since the ext_vector type isn't something I have a great knowledge of. The OpenCL vector types end up having frustratingly similar/different behavior that it required significant differences in logic.

Does OpenCL prohibit void/throw expressions in their vector conditionals?

clang/lib/Sema/SemaExprCXX.cpp

5797

The one I found was less specific than just 'mismatched'. It ended up containing info about not convertible. I tried a set of 'selects' on it but it ended up getting hacked up more than I was comfortable with. Unless you're referring to a different one that I missed?

[erichkeane]

Ping?

[aaron.ballman]

LGTM!

clang/docs/LanguageExtensions.rst

.. code-block:: c++
   float4 vf = (float4)((float2)(1.0f, 2.0f), (float2)(3.0f, 4.0f));
 
 Vector Operations
 -----------------
 
 The table below shows the support for each operation by vector extension.  A
 dash indicates that an operation is not accepted according to a corresponding
 specification.
 
-============================== ======= ======= ======= =======
-         Operator              OpenCL  AltiVec   GCC    NEON
-============================== ======= ======= ======= =======
-[]                               yes     yes     yes     --
-unary operators +, --            yes     yes     yes     --
-++, -- --                        yes     yes     yes     --
-+,--,*,/,%                       yes     yes     yes     --
-bitwise operators &,|,^,~        yes     yes     yes     --
->>,<<                            yes     yes     yes     --
-!, &&, ||                        yes     --      --      --
-==, !=, >, <, >=, <=             yes     yes     --      --
-=                                yes     yes     yes     yes
-:?                               yes     --      --      --
-sizeof                           yes     yes     yes     yes
-C-style cast                     yes     yes     yes     no
-reinterpret_cast                 yes     no      yes     no
-static_cast                      yes     no      yes     no
-const_cast                       no      no      no      no
-============================== ======= ======= ======= =======
+============================== ======= ======= ============= =======
+         Operator              OpenCL  AltiVec     GCC        NEON
+============================== ======= ======= ============= =======
+[]                               yes     yes       yes         --
+unary operators +, --            yes     yes       yes         --
+++, -- --                        yes     yes       yes         --
++,--,*,/,%                       yes     yes       yes         --
+bitwise operators &,|,^,~        yes     yes       yes         --
+>>,<<                            yes     yes       yes         --
+!, &&, ||                        yes     --        yes [#]_    --
+==, !=, >, <, >=, <=             yes     yes       yes         --
+=                                yes     yes       yes         yes
+:? [#]_                          yes     --        yes         --
+sizeof                           yes     yes       yes         yes
+C-style cast                     yes     yes       yes         no
+reinterpret_cast                 yes     no        yes         no
+static_cast                      yes     no        yes         no
+const_cast                       no      no        no          no
+============================== ======= ======= ============= =======
 
 See also :ref:`langext-__builtin_shufflevector`, :ref:`langext-__builtin_convertvector`.
 
+.. [#] unary operator ! is not implemented, however && and || are.
+.. [#] While OpenCL and GCC vectors both implement the comparison operator(?:) as a
+  'select', they operate somewhat differently. OpenCL selects based on signedness of
+  the condition operands, but GCC vectors use normal bool conversions (that is, != 0).
+
 Half-Precision Floating Point
 =============================
 
 Clang supports two half-precision (16-bit) floating point types: ``__fp16`` and
 ``_Float16``.  These types are supported in all language modes.
 
 ``__fp16`` is supported on every target, as it is purely a storage format; see below.
 ``_Float16`` is currently only supported on the following targets, with further

clang/include/clang/AST/Expr.h

 /// middle" extension is a BinaryConditionalOperator.
 class ConditionalOperator : public AbstractConditionalOperator {
   enum { COND, LHS, RHS, END_EXPR };
   Stmt* SubExprs[END_EXPR]; // Left/Middle/Right hand sides.
 
   friend class ASTStmtReader;
 public:
   ConditionalOperator(Expr *cond, SourceLocation QLoc, Expr *lhs,
-                      SourceLocation CLoc, Expr *rhs,
-                      QualType t, ExprValueKind VK, ExprObjectKind OK)
-    : AbstractConditionalOperator(ConditionalOperatorClass, t, VK, OK,
-           // FIXME: the type of the conditional operator doesn't
-           // depend on the type of the conditional, but the standard
-           // seems to imply that it could. File a bug!
-           (lhs->isTypeDependent() || rhs->isTypeDependent()),
+                      SourceLocation CLoc, Expr *rhs, QualType t,
+                      ExprValueKind VK, ExprObjectKind OK)
+      : AbstractConditionalOperator(
+            ConditionalOperatorClass, t, VK, OK,
+            // The type of the conditional operator depends on the type
+            // of the conditional to support the GCC vector conditional
+            // extension. Additionally, [temp.dep.expr] does specify state that
+            // this should be dependent on ALL sub expressions.
+            (cond->isTypeDependent() || lhs->isTypeDependent() ||
+             rhs->isTypeDependent()),
             (cond->isValueDependent() || lhs->isValueDependent() ||
              rhs->isValueDependent()),
             (cond->isInstantiationDependent() ||
              lhs->isInstantiationDependent() ||
              rhs->isInstantiationDependent()),
             (cond->containsUnexpandedParameterPack() ||
              lhs->containsUnexpandedParameterPack() ||
              rhs->containsUnexpandedParameterPack()),
-                                  QLoc, CLoc) {
     SubExprs[COND] = cond;
     SubExprs[LHS] = lhs;
     SubExprs[RHS] = rhs;
   }
 
   /// Build an empty conditional operator.
   explicit ConditionalOperator(EmptyShell Empty)
     : AbstractConditionalOperator(ConditionalOperatorClass, Empty) { }

clang/include/clang/Basic/DiagnosticSemaKinds.td

def err_conditional_ambiguous_ovl : Error<
   "conditional expression is ambiguous; %diff{$ and $|types}0,1 "
   "can be converted to several common types">;
 def err_conditional_vector_size : Error<
   "vector condition type %0 and result type %1 do not have the same number "
   "of elements">;
 def err_conditional_vector_element_size : Error<
   "vector condition type %0 and result type %1 do not have elements of the "
   "same size">;
+def err_conditional_vector_has_void : Error<
+  "GNU vector conditional operand cannot be %select{void|a throw expression}0">;
+def err_conditional_vector_operand_type
+    : Error<"%select{enumeration|extended vector}0 type %1 is not allowed in a "
+            "vector conditional">;
+def err_conditional_vector_mismatched_vectors
+    : Error<"vector operands to the vector conditional must be the same type "
+            "%diff{($ and $)|}0,1}">;
 
 def err_throw_incomplete : Error<
   "cannot throw object of incomplete type %0">;
 def err_throw_incomplete_ptr : Error<
   "cannot throw pointer to object of incomplete type %0">;
 def warn_throw_underaligned_obj : Warning<
   "underaligned exception object thrown">,
   InGroup<UnderalignedExceptionObject>;

clang/include/clang/Sema/Sema.h

   Expr *recreateSyntacticForm(PseudoObjectExpr *E);
 
   QualType CheckConditionalOperands( // C99 6.5.15
     ExprResult &Cond, ExprResult &LHS, ExprResult &RHS,
     ExprValueKind &VK, ExprObjectKind &OK, SourceLocation QuestionLoc);
   QualType CXXCheckConditionalOperands( // C++ 5.16
     ExprResult &cond, ExprResult &lhs, ExprResult &rhs,
     ExprValueKind &VK, ExprObjectKind &OK, SourceLocation questionLoc);
+  QualType CheckGNUVectorConditionalTypes(ExprResult &Cond, ExprResult &LHS,
+                                          ExprResult &RHS,
+                                          SourceLocation QuestionLoc);
   QualType FindCompositePointerType(SourceLocation Loc, Expr *&E1, Expr *&E2,
                                     bool ConvertArgs = true);
   QualType FindCompositePointerType(SourceLocation Loc,
                                     ExprResult &E1, ExprResult &E2,
                                     bool ConvertArgs = true) {
     Expr *E1Tmp = E1.get(), *E2Tmp = E2.get();
     QualType Composite =
         FindCompositePointerType(Loc, E1Tmp, E2Tmp, ConvertArgs);

clang/lib/AST/ExprConstant.cpp

public:
 
     bool VisitUnaryReal(const UnaryOperator *E)
       { return Visit(E->getSubExpr()); }
     bool VisitCastExpr(const CastExpr* E);
     bool VisitInitListExpr(const InitListExpr *E);
     bool VisitUnaryImag(const UnaryOperator *E);
     // FIXME: Missing: unary -, unary ~, binary add/sub/mul/div,
     //                 binary comparisons, binary and/or/xor,
+    //                 conditional operator (for GNU conditional select),
     //                 shufflevector, ExtVectorElementExpr
   };
 } // end anonymous namespace
 
 static bool EvaluateVector(const Expr* E, APValue& Result, EvalInfo &Info) {
   assert(E->isRValue() && E->getType()->isVectorType() &&"not a vector rvalue");
   return VectorExprEvaluator(Info, Result).Visit(E);
 }

clang/lib/CodeGen/CGExprScalar.cpp

if (CGF.getLangOpts().OpenCL
     llvm::Value *tmp4 = Builder.CreateAnd(LHSTmp, tmp);
     llvm::Value *tmp5 = Builder.CreateOr(tmp3, tmp4, "cond");
     if (wasCast)
       tmp5 = Builder.CreateBitCast(tmp5, RHS->getType());
 
     return tmp5;
   }
 
+  if (condExpr->getType()->isVectorType()) {
+    CGF.incrementProfileCounter(E);
+
+    llvm::Value *CondV = CGF.EmitScalarExpr(condExpr);
+    llvm::Value *LHS = Visit(lhsExpr);
+    llvm::Value *RHS = Visit(rhsExpr);
+
+    llvm::Type *CondType = ConvertType(condExpr->getType());
+    auto *VecTy = cast<llvm::VectorType>(CondType);
+    llvm::Value *ZeroVec = llvm::Constant::getNullValue(VecTy);
+
+    CondV = Builder.CreateICmpNE(CondV, ZeroVec, "vector_cond");
+    return Builder.CreateSelect(CondV, LHS, RHS, "vector_select");
+  }
+
   // If this is a really simple expression (like x ? 4 : 5), emit this as a
   // select instead of as control flow.  We can only do this if it is cheap and
   // safe to evaluate the LHS and RHS unconditionally.
   if (isCheapEnoughToEvaluateUnconditionally(lhsExpr, CGF) &&
       isCheapEnoughToEvaluateUnconditionally(rhsExpr, CGF)) {
     llvm::Value *CondV = CGF.EvaluateExprAsBool(condExpr);
     llvm::Value *StepV = Builder.CreateZExtOrBitCast(CondV, CGF.Int64Ty);
 

clang/lib/Sema/SemaExprCXX.cpp

static bool ConvertForConditional(Sema &Self, ExprResult &E, QualType T) {
   ExprResult Result = InitSeq.Perform(Self, Entity, Kind, Arg);
   if (Result.isInvalid())
     return true;
 
   E = Result;
   return false;
 }
 
+// Check the condition operand of ?: to see if it is valid for the GCC
+// extension.
+static bool isValidVectorForConditionalCondition(ASTContext &Ctx,
+                                                 QualType CondTy) {
+  if (!CondTy->isVectorType() || CondTy->isExtVectorType())
+    return false;
+  const QualType EltTy =
+      cast<VectorType>(CondTy.getCanonicalType())->getElementType();
+
+  assert(!EltTy->isBooleanType() && !EltTy->isEnumeralType() &&
+         "Vectors cant be boolean or enum types");
+  return EltTy->isIntegralType(Ctx);
+}
+
+QualType Sema::CheckGNUVectorConditionalTypes(ExprResult &Cond, ExprResult &LHS,
+                                              ExprResult &RHS,
+                                              SourceLocation QuestionLoc) {
+  LHS = DefaultFunctionArrayLvalueConversion(LHS.get());
+  RHS = DefaultFunctionArrayLvalueConversion(RHS.get());
+
+  QualType CondType = Cond.get()->getType();
+  const auto *CondVT = CondType->getAs<VectorType>();
+  QualType CondElementTy = CondVT->getElementType();
+  unsigned CondElementCount = CondVT->getNumElements();
+  QualType LHSType = LHS.get()->getType();
+  const auto *LHSVT = LHSType->getAs<VectorType>();
+  QualType RHSType = RHS.get()->getType();
+  const auto *RHSVT = RHSType->getAs<VectorType>();
+
+  QualType ResultType;
+
+  // FIXME: In the future we should define what the Extvector conditional
+  // operator looks like.
+  if (LHSVT && isa<ExtVectorType>(LHSVT)) {
+    Diag(QuestionLoc, diag::err_conditional_vector_operand_type)
+        << /*isExtVector*/ true << LHSType;
+    return {};
+  }
+
+  if (RHSVT && isa<ExtVectorType>(RHSVT)) {
+    Diag(QuestionLoc, diag::err_conditional_vector_operand_type)
+        << /*isExtVector*/ true << RHSType;
+    return {};
+  }
+
+  if (LHSVT && RHSVT) {
+    // If both are vector types, they must be the same type.
+    if (!Context.hasSameType(LHSType, RHSType)) {
+      Diag(QuestionLoc, diag::err_conditional_vector_mismatched_vectors)
+          << LHSType << RHSType;
+      return {};
+    }
+    ResultType = LHSType;
+  } else if (LHSVT || RHSVT) {
+    ResultType = CheckVectorOperands(
+        LHS, RHS, QuestionLoc, /*isCompAssign*/ false, /*AllowBothBool*/ true,
+        /*AllowBoolConversions*/ false);
+    if (ResultType.isNull())
+      return {};
+  } else {
+    // Both are scalar.
+    QualType ResultElementTy;
+    LHSType = LHSType.getCanonicalType().getUnqualifiedType();
+    RHSType = RHSType.getCanonicalType().getUnqualifiedType();
+
+    if (Context.hasSameType(LHSType, RHSType))
+      ResultElementTy = LHSType;
+    else
+      ResultElementTy = UsualArithmeticConversions(LHS, RHS);
+
+    if (ResultElementTy->isEnumeralType()) {
+      Diag(QuestionLoc, diag::err_conditional_vector_operand_type)
+          << /*isExtVector*/ false << ResultElementTy;
+      return {};
+    }
+    ResultType = Context.getVectorType(
+        ResultElementTy, CondType->getAs<VectorType>()->getNumElements(),
+        VectorType::GenericVector);
+
+    LHS = ImpCastExprToType(LHS.get(), ResultType, CK_VectorSplat);
+    RHS = ImpCastExprToType(RHS.get(), ResultType, CK_VectorSplat);
+  }
+
+  assert(!ResultType.isNull() && ResultType->isVectorType() &&
+         "Result should have been a vector type");
+  QualType ResultElementTy = ResultType->getAs<VectorType>()->getElementType();
+  unsigned ResultElementCount =
+      ResultType->getAs<VectorType>()->getNumElements();
+
+  if (ResultElementCount != CondElementCount) {
+    Diag(QuestionLoc, diag::err_conditional_vector_size) << CondType
+                                                         << ResultType;
+    return {};
+  }
+
+  if (Context.getTypeSize(ResultElementTy) !=
+      Context.getTypeSize(CondElementTy)) {
+    Diag(QuestionLoc, diag::err_conditional_vector_element_size) << CondType
+                                                                 << ResultType;
+    return {};
+  }
+
+  return ResultType;
+}
+
 /// Check the operands of ?: under C++ semantics.
 ///
 /// See C++ [expr.cond]. Note that LHS is never null, even for the GNU x ?: y
 /// extension. In this case, LHS == Cond. (But they're not aliases.)
+///
+/// This function also implements GCC's vector extension for conditionals.
+///  GCC's vector extension permits the use of a?b:c where the type of
+///  a is that of a integer vector with the same number of elements and
+///  size as the vectors of b and c. If one of either b or c is a scalar
+///  it is implicitly converted to match the type of the vector.
+///  Otherwise the expression is ill-formed. If both b and c are scalars,
+///  then b and c are checked and converted to the type of a if possible.
+///  Unlike the OpenCL ?: operator, the expression is evaluated as
+///  (a[0] != 0 ? b[0] : c[0], .. , a[n] != 0 ? b[n] : c[n]).
 QualType Sema::CXXCheckConditionalOperands(ExprResult &Cond, ExprResult &LHS,
                                            ExprResult &RHS, ExprValueKind &VK,
                                            ExprObjectKind &OK,
                                            SourceLocation QuestionLoc) {
-  // FIXME: Handle C99's complex types, vector types, block pointers and Obj-C++
-  // interface pointers.
+  // FIXME: Handle C99's complex types, block pointers and Obj-C++ interface
+  // pointers.
+
+  // Assume r-value.
+  VK = VK_RValue;
+  OK = OK_Ordinary;
+  bool IsVectorConditional =
+      isValidVectorForConditionalCondition(Context, Cond.get()->getType());
 
   // C++11 [expr.cond]p1
   //   The first expression is contextually converted to bool.
-  //
-  // FIXME; GCC's vector extension permits the use of a?b:c where the type of
-  //        a is that of a integer vector with the same number of elements and
-  //        size as the vectors of b and c. If one of either b or c is a scalar
-  //        it is implicitly converted to match the type of the vector.
-  //        Otherwise the expression is ill-formed. If both b and c are scalars,
-  //        then b and c are checked and converted to the type of a if possible.
-  //        Unlike the OpenCL ?: operator, the expression is evaluated as
-  //        (a[0] != 0 ? b[0] : c[0], .. , a[n] != 0 ? b[n] : c[n]).
   if (!Cond.get()->isTypeDependent()) {
-    ExprResult CondRes = CheckCXXBooleanCondition(Cond.get());
+    ExprResult CondRes = IsVectorConditional
+                             ? DefaultFunctionArrayLvalueConversion(Cond.get())
+                             : CheckCXXBooleanCondition(Cond.get());
     if (CondRes.isInvalid())
       return QualType();
     Cond = CondRes;
+  } else {
+    // To implement C++, the first expression typically doesn't alter the result
+    // type of the conditional, however the GCC compatible vector extension
+    // changes the result type to be that of the conditional. Since we cannot
+    // know if this is a vector extension here, delay the conversion of the
+    // LHS/RHS below until later.
+    return Context.DependentTy;
   }
 
-  // Assume r-value.
-  VK = VK_RValue;
-  OK = OK_Ordinary;
 
   // Either of the arguments dependent?
   if (LHS.get()->isTypeDependent() || RHS.get()->isTypeDependent())
     return Context.DependentTy;
 
   // C++11 [expr.cond]p2
   //   If either the second or the third operand has type (cv) void, ...
   QualType LTy = LHS.get()->getType();
   QualType RTy = RHS.get()->getType();
   bool LVoid = LTy->isVoidType();
   bool RVoid = RTy->isVoidType();
   if (LVoid || RVoid) {
     //   ... one of the following shall hold:
     //   -- The second or the third operand (but not both) is a (possibly
     //      parenthesized) throw-expression; the result is of the type
     //      and value category of the other.
     bool LThrow = isa<CXXThrowExpr>(LHS.get()->IgnoreParenImpCasts());
     bool RThrow = isa<CXXThrowExpr>(RHS.get()->IgnoreParenImpCasts());
+
+    // Void expressions aren't legal in the vector-conditional expressions.
+    if (IsVectorConditional) {
+      SourceRange DiagLoc =
+          LVoid ? LHS.get()->getSourceRange() : RHS.get()->getSourceRange();
+      bool IsThrow = LVoid ? LThrow : RThrow;
+      Diag(DiagLoc.getBegin(), diag::err_conditional_vector_has_void)
+          << DiagLoc << IsThrow;
+      return QualType();
+    }
+
     if (LThrow != RThrow) {
       Expr *NonThrow = LThrow ? RHS.get() : LHS.get();
       VK = NonThrow->getValueKind();
       // DR (no number yet): the result is a bit-field if the
       // non-throw-expression operand is a bit-field.
       OK = NonThrow->getObjectKind();
       return NonThrow->getType();
     }
 
     //   -- Both the second and third operands have type void; the result is of
     //      type void and is a prvalue.
     if (LVoid && RVoid)
       return Context.VoidTy;
 
     // Neither holds, error.
     Diag(QuestionLoc, diag::err_conditional_void_nonvoid)
       << (LVoid ? RTy : LTy) << (LVoid ? 0 : 1)
       << LHS.get()->getSourceRange() << RHS.get()->getSourceRange();
     return QualType();
   }
 
   // Neither is void.
+  if (IsVectorConditional)
+    return CheckGNUVectorConditionalTypes(Cond, LHS, RHS, QuestionLoc);
 
   // C++11 [expr.cond]p3
   //   Otherwise, if the second and third operand have different types, and
   //   either has (cv) class type [...] an attempt is made to convert each of
   //   those operands to the type of the other.
   if (!Context.hasSameType(LTy, RTy) &&
       (LTy->isRecordType() || RTy->isRecordType())) {
     // These return true if a single direction is already ambiguous.

clang/test/CodeGenCXX/vector-conditional.cpp

+// RUN: %clang_cc1 %s -triple x86_64-linux-gnu -Wno-unused -std=c++11 -emit-llvm -o - | FileCheck %s
+
+using FourShorts = short __attribute__((__vector_size__(8)));
+using TwoInts = int __attribute__((__vector_size__(8)));
+using TwoUInts = unsigned __attribute__((__vector_size__(8)));
+using FourInts = int __attribute__((__vector_size__(16)));
+using FourUInts = unsigned __attribute__((__vector_size__(16)));
+using TwoLongLong = long long __attribute__((__vector_size__(16)));
+using FourLongLong = long long __attribute__((__vector_size__(32)));
+using TwoFloats = float __attribute__((__vector_size__(8)));
+using FourFloats = float __attribute__((__vector_size__(16)));
+using TwoDoubles = double __attribute__((__vector_size__(16)));
+using FourDoubles = double __attribute__((__vector_size__(32)));
+
+FourShorts four_shorts;
+TwoInts two_ints;
+TwoUInts two_uints;
+FourInts four_ints;
+FourUInts four_uints;
+TwoLongLong two_ll;
+FourLongLong four_ll;
+TwoFloats two_floats;
+FourFloats four_floats;
+TwoDoubles two_doubles;
+FourDoubles four_doubles;
+
+short some_short;
+unsigned short some_ushort;
+int some_int;
+float some_float;
+unsigned int some_uint;
+long long some_ll;
+unsigned long long some_ull;
+double some_double;
+
+// CHECK: TwoVectorOps
+void TwoVectorOps() {
+  two_ints ? two_ints : two_ints;
+  // CHECK: %[[COND:.+]] = load <2 x i32>
+  // CHECK: %[[LHS:.+]] = load <2 x i32>
+  // CHECK: %[[RHS:.+]] = load <2 x i32>
+  // CHECK: %[[NEZERO:.+]] = icmp ne <2 x i32> %[[COND]], zeroinitializer
+  // CHECK: %[[SELECT:.+]] = select <2 x i1> %[[NEZERO]], <2 x i32> %[[LHS]], <2 x i32> %[[RHS]]
+
+  two_ints ? two_floats : two_floats;
+  // CHECK: %[[COND:.+]] = load <2 x i32>
+  // CHECK: %[[LHS:.+]] = load <2 x float>
+  // CHECK: %[[RHS:.+]] = load <2 x float>
+  // CHECK: %[[NEZERO:.+]] = icmp ne <2 x i32> %[[COND]], zeroinitializer
+  // CHECK: %[[SELECT:.+]] = select <2 x i1> %[[NEZERO]], <2 x float> %[[LHS]], <2 x float> %[[RHS]]
+
+  two_ll ? two_doubles : two_doubles;
+  // CHECK: %[[COND:.+]] = load <2 x i64>
+  // CHECK: %[[LHS:.+]] = load <2 x double>
+  // CHECK: %[[RHS:.+]] = load <2 x double>
+  // CHECK: %[[NEZERO:.+]] = icmp ne <2 x i64> %[[COND]], zeroinitializer
+  // CHECK: %[[SELECT:.+]] = select <2 x i1> %[[NEZERO]], <2 x double> %[[LHS]], <2 x double> %[[RHS]]
+}
+
+// CHECK: TwoScalarOps
+void TwoScalarOps() {
+  four_shorts ? some_short : some_short;
+  // CHECK: %[[COND:.+]] = load <4 x i16>
+  // CHECK: %[[LHS:.+]] = load i16
+  // CHECK: %[[LHS_SPLAT_INSERT:.+]] = insertelement <4 x i16> undef, i16 %[[LHS]], i32 0
+  // CHECK: %[[LHS_SPLAT:.+]] = shufflevector <4 x i16> %[[LHS_SPLAT_INSERT]], <4 x i16> undef, <4 x i32> zeroinitializer
+  // CHECK: %[[RHS:.+]] = load i16
+  // CHECK: %[[RHS_SPLAT_INSERT:.+]] = insertelement <4 x i16> undef, i16 %[[RHS]], i32 0
+  // CHECK: %[[RHS_SPLAT:.+]] = shufflevector <4 x i16> %[[RHS_SPLAT_INSERT]], <4 x i16> undef, <4 x i32> zeroinitializer
+  // CHECK: %[[NEZERO:.+]] = icmp ne <4 x i16> %[[COND]], zeroinitializer
+  // CHECK: %[[SELECT:.+]] = select <4 x i1> %[[NEZERO]], <4 x i16> %[[LHS_SPLAT]], <4 x i16> %[[RHS_SPLAT]]
+
+  four_shorts ? some_ushort : some_ushort;
+  // CHECK: %[[COND:.+]] = load <4 x i16>
+  // CHECK: %[[LHS:.+]] = load i16
+  // CHECK: %[[LHS_SPLAT_INSERT:.+]] = insertelement <4 x i16> undef, i16 %[[LHS]], i32 0
+  // CHECK: %[[LHS_SPLAT:.+]] = shufflevector <4 x i16> %[[LHS_SPLAT_INSERT]], <4 x i16> undef, <4 x i32> zeroinitializer
+  // CHECK: %[[RHS:.+]] = load i16
+  // CHECK: %[[RHS_SPLAT_INSERT:.+]] = insertelement <4 x i16> undef, i16 %[[RHS]], i32 0
+  // CHECK: %[[RHS_SPLAT:.+]] = shufflevector <4 x i16> %[[RHS_SPLAT_INSERT]], <4 x i16> undef, <4 x i32> zeroinitializer
+  // CHECK: %[[NEZERO:.+]] = icmp ne <4 x i16> %[[COND]], zeroinitializer
+  // CHECK: %[[SELECT:.+]] = select <4 x i1> %[[NEZERO]], <4 x i16> %[[LHS_SPLAT]], <4 x i16> %[[RHS_SPLAT]]
+
+  four_ints ? some_ushort : some_short;
+  // CHECK: %[[COND:.+]] = load <4 x i32>
+  // CHECK: %[[LHS:.+]] = load i16
+  // CHECK: %[[LHS_ZEXT:.+]] = zext i16 %[[LHS]] to i32
+  // CHECK: %[[LHS_SPLAT_INSERT:.+]] = insertelement <4 x i32> undef, i32 %[[LHS_ZEXT]], i32 0
+  // CHECK: %[[LHS_SPLAT:.+]] = shufflevector <4 x i32> %[[LHS_SPLAT_INSERT]], <4 x i32> undef, <4 x i32> zeroinitializer
+  // CHECK: %[[RHS:.+]] = load i16
+  // CHECK: %[[RHS_SEXT:.+]] = sext i16 %[[RHS]] to i32
+  // CHECK: %[[RHS_SPLAT_INSERT:.+]] = insertelement <4 x i32> undef, i32 %[[RHS_SEXT]], i32 0
+  // CHECK: %[[RHS_SPLAT:.+]] = shufflevector <4 x i32> %[[RHS_SPLAT_INSERT]], <4 x i32> undef, <4 x i32> zeroinitializer
+  // CHECK: %[[NEZERO:.+]] = icmp ne <4 x i32> %[[COND]], zeroinitializer
+  // CHECK: %[[SELECT:.+]] = select <4 x i1> %[[NEZERO]], <4 x i32> %[[LHS_SPLAT]], <4 x i32> %[[RHS_SPLAT]]
+
+  four_ints ? some_int : some_float;
+  // CHECK: %[[COND:.+]] = load <4 x i32>
+  // CHECK: %[[LHS:.+]] = load i32
+  // CHECK: %[[LHS_CONV:.+]] = sitofp i32 %[[LHS]] to float
+  // CHECK: %[[LHS_SPLAT_INSERT:.+]] = insertelement <4 x float> undef, float %[[LHS_CONV]], i32 0
+  // CHECK: %[[LHS_SPLAT:.+]] = shufflevector <4 x float> %[[LHS_SPLAT_INSERT]], <4 x float> undef, <4 x i32> zeroinitializer
+  // CHECK: %[[RHS:.+]] = load float
+  // CHECK: %[[RHS_SPLAT_INSERT:.+]] = insertelement <4 x float> undef, float %[[RHS]], i32 0
+  // CHECK: %[[RHS_SPLAT:.+]] = shufflevector <4 x float> %[[RHS_SPLAT_INSERT]], <4 x float> undef, <4 x i32> zeroinitializer
+  // CHECK: %[[NEZERO:.+]] = icmp ne <4 x i32> %[[COND]], zeroinitializer
+  // CHECK: %[[SELECT:.+]] = select <4 x i1> %[[NEZERO]], <4 x float> %[[LHS_SPLAT]], <4 x float> %[[RHS_SPLAT]]
+
+  four_ll ? some_double : some_ll;
+  // CHECK: %[[COND:.+]] = load <4 x i64>
+  // CHECK: %[[LHS:.+]] = load double
+  // CHECK: %[[LHS_SPLAT_INSERT:.+]] = insertelement <4 x double> undef, double %[[LHS]], i32 0
+  // CHECK: %[[LHS_SPLAT:.+]] = shufflevector <4 x double> %[[LHS_SPLAT_INSERT]], <4 x double> undef, <4 x i32> zeroinitializer
+  // CHECK: %[[RHS:.+]] = load i64
+  // CHECK: %[[RHS_CONV:.+]] = sitofp i64 %[[RHS]] to double
+  // CHECK: %[[RHS_SPLAT_INSERT:.+]] = insertelement <4 x double> undef, double %[[RHS_CONV]], i32 0
+  // CHECK: %[[RHS_SPLAT:.+]] = shufflevector <4 x double> %[[RHS_SPLAT_INSERT]], <4 x double> undef, <4 x i32> zeroinitializer
+  // CHECK: %[[NEZERO:.+]] = icmp ne <4 x i64> %[[COND]], zeroinitializer
+  // CHECK: %[[SELECT:.+]] = select <4 x i1> %[[NEZERO]], <4 x double> %[[LHS_SPLAT]], <4 x double> %[[RHS_SPLAT]]
+
+  four_ints ? some_int : some_short;
+  // CHECK: %[[COND:.+]] = load <4 x i32>
+  // CHECK: %[[LHS:.+]] = load i32
+  // CHECK: %[[LHS_SPLAT_INSERT:.+]] = insertelement <4 x i32> undef, i32 %[[LHS]], i32 0
+  // CHECK: %[[LHS_SPLAT:.+]] = shufflevector <4 x i32> %[[LHS_SPLAT_INSERT]], <4 x i32> undef, <4 x i32> zeroinitializer
+  // CHECK: %[[RHS:.+]] = load i16
+  // CHECK: %[[RHS_SEXT:.+]] = sext i16 %[[RHS]] to i32
+  // CHECK: %[[RHS_SPLAT_INSERT:.+]] = insertelement <4 x i32> undef, i32 %[[RHS_SEXT]], i32 0
+  // CHECK: %[[RHS_SPLAT:.+]] = shufflevector <4 x i32> %[[RHS_SPLAT_INSERT]], <4 x i32> undef, <4 x i32> zeroinitializer
+  // CHECK: %[[NEZERO:.+]] = icmp ne <4 x i32> %[[COND]], zeroinitializer
+  // CHECK: %[[SELECT:.+]] = select <4 x i1> %[[NEZERO]], <4 x i32> %[[LHS_SPLAT]], <4 x i32> %[[RHS_SPLAT]]
+}
+
+// CHECK: OneScalarOp
+void OneScalarOp() {
+  four_ints ? four_ints : some_int;
+  // CHECK: %[[COND:.+]] = load <4 x i32>
+  // CHECK: %[[LHS:.+]] = load <4 x i32>
+  // CHECK: %[[RHS:.+]] = load i32
+  // CHECK: %[[RHS_SPLAT_INSERT:.+]] = insertelement <4 x i32> undef, i32 %[[RHS]], i32 0
+  // CHECK: %[[RHS_SPLAT:.+]] = shufflevector <4 x i32> %[[RHS_SPLAT_INSERT]], <4 x i32> undef, <4 x i32> zeroinitializer
+  // CHECK: %[[NEZERO:.+]] = icmp ne <4 x i32> %[[COND]], zeroinitializer
+  // CHECK: %[[SELECT:.+]] = select <4 x i1> %[[NEZERO]], <4 x i32> %[[LHS]], <4 x i32> %[[RHS_SPLAT]]
+
+  four_ints ? four_ints : 5;
+  // CHECK: %[[COND:.+]] = load <4 x i32>
+  // CHECK: %[[LHS:.+]] = load <4 x i32>
+  // CHECK: %[[NEZERO:.+]] = icmp ne <4 x i32> %[[COND]], zeroinitializer
+  // CHECK: %[[SELECT:.+]] = select <4 x i1> %[[NEZERO]], <4 x i32> %[[LHS]], <4 x i32> <i32 5, i32 5, i32 5, i32 5>
+
+  four_ints ?: some_float;
+  // CHECK: %[[COND:.+]] = load <4 x i32>
+  // CHECK: %[[RHS:.+]] = load float
+  // CHECK: %[[RHS_CONV:.+]] = fptosi float %[[RHS]] to i32
+  // CHECK: %[[RHS_SPLAT_INSERT:.+]] = insertelement <4 x i32> undef, i32 %[[RHS_CONV]], i32 0
+  // CHECK: %[[RHS_SPLAT:.+]] = shufflevector <4 x i32> %[[RHS_SPLAT_INSERT]], <4 x i32> undef, <4 x i32> zeroinitializer
+  // CHECK: %[[NEZERO:.+]] = icmp ne <4 x i32> %[[COND]], zeroinitializer
+  // CHECK: %[[SELECT:.+]] = select <4 x i1> %[[NEZERO]], <4 x i32> %[[COND]], <4 x i32> %[[RHS_SPLAT]]
+
+  four_ints ? four_ints : 5.0f;
+  // CHECK: %[[COND:.+]] = load <4 x i32>
+  // CHECK: %[[LHS:.+]] = load <4 x i32>
+  // CHECK: %[[NEZERO:.+]] = icmp ne <4 x i32> %[[COND]], zeroinitializer
+  // CHECK: %[[SELECT:.+]] = select <4 x i1> %[[NEZERO]], <4 x i32> %[[LHS]], <4 x i32> <i32 5, i32 5, i32 5, i32 5>
+
+  four_ints ? some_float : four_ints;
+  // CHECK: %[[COND:.+]] = load <4 x i32>
+  // CHECK: %[[LHS:.+]] = load float
+  // CHECK: %[[LHS_CONV:.+]] = fptosi float %[[LHS]] to i32
+  // CHECK: %[[LHS_SPLAT_INSERT:.+]] = insertelement <4 x i32> undef, i32 %[[LHS_CONV]], i32 0
+  // CHECK: %[[LHS_SPLAT:.+]] = shufflevector <4 x i32> %[[LHS_SPLAT_INSERT]], <4 x i32> undef, <4 x i32> zeroinitializer
+  // CHECK: %[[RHS:.+]] = load <4 x i32>
+  // CHECK: %[[NEZERO:.+]] = icmp ne <4 x i32> %[[COND]], zeroinitializer
+  // CHECK: %[[SELECT:.+]] = select <4 x i1> %[[NEZERO]], <4 x i32> %[[LHS_SPLAT]], <4 x i32> %[[RHS]]
+
+  four_ints ? four_floats : some_float;
+  // CHECK: %[[COND:.+]] = load <4 x i32>
+  // CHECK: %[[LHS:.+]] = load <4 x float>
+  // CHECK: %[[RHS:.+]] = load float
+  // CHECK: %[[RHS_SPLAT_INSERT:.+]] = insertelement <4 x float> undef, float %[[RHS]], i32 0
+  // CHECK: %[[RHS_SPLAT:.+]] = shufflevector <4 x float> %[[RHS_SPLAT_INSERT]], <4 x float> undef, <4 x i32> zeroinitializer
+  // CHECK: %[[NEZERO:.+]] = icmp ne <4 x i32> %[[COND]], zeroinitializer
+  // CHECK: %[[SELECT:.+]] = select <4 x i1> %[[NEZERO]], <4 x float> %[[LHS]], <4 x float> %[[RHS_SPLAT]]
+
+  four_ll ? four_doubles : 6.0;
+  // CHECK: %[[COND:.+]] = load <4 x i64>
+  // CHECK: %[[LHS:.+]] = load <4 x double>
+  // CHECK: %[[NEZERO:.+]] = icmp ne <4 x i64> %[[COND]], zeroinitializer
+  // CHECK: %[[SELECT:.+]] = select <4 x i1> %[[NEZERO]], <4 x double> %[[LHS]], <4 x double> <double 6.{{.+}}, double 6.{{.+}}, double 6.{{.+}}>
+
+  four_ll ? four_ll : 6.0;
+  // CHECK: %[[COND:.+]] = load <4 x i64>
+  // CHECK: %[[LHS:.+]] = load <4 x i64>
+  // CHECK: %[[NEZERO:.+]] = icmp ne <4 x i64> %[[COND]], zeroinitializer
+  // CHECK: %[[SELECT:.+]] = select <4 x i1> %[[NEZERO]], <4 x i64> %[[LHS]], <4 x i64> <i64 6, i64 6, i64 6, i64 6>
+
+  four_ll ? four_ll : 6;
+  // CHECK: %[[COND:.+]] = load <4 x i64>
+  // CHECK: %[[LHS:.+]] = load <4 x i64>
+  // CHECK: %[[NEZERO:.+]] = icmp ne <4 x i64> %[[COND]], zeroinitializer
+  // CHECK: %[[SELECT:.+]] = select <4 x i1> %[[NEZERO]], <4 x i64> %[[LHS]], <4 x i64> <i64 6, i64 6, i64 6, i64 6>
+
+  four_ll ? four_ll : some_int;
+  // CHECK: %[[COND:.+]] = load <4 x i64>
+  // CHECK: %[[LHS:.+]] = load <4 x i64>
+  // CHECK: %[[RHS:.+]] = load i32
+  // CHECK: %[[RHS_CONV:.+]] = sext i32 %[[RHS]] to i64
+  // CHECK: %[[RHS_SPLAT_INSERT:.+]] = insertelement <4 x i64> undef, i64 %[[RHS_CONV]], i32 0
+  // CHECK: %[[RHS_SPLAT:.+]] = shufflevector <4 x i64> %[[RHS_SPLAT_INSERT]], <4 x i64> undef, <4 x i32> zeroinitializer
+  // CHECK: %[[NEZERO:.+]] = icmp ne <4 x i64> %[[COND]], zeroinitializer
+  // CHECK: %[[SELECT:.+]] = select <4 x i1> %[[NEZERO]], <4 x i64> %[[LHS]], <4 x i64> %[[RHS_SPLAT]]
+
+  four_ll ? four_ll : some_ll;
+  // CHECK: %[[COND:.+]] = load <4 x i64>
+  // CHECK: %[[LHS:.+]] = load <4 x i64>
+  // CHECK: %[[RHS:.+]] = load i64
+  // CHECK: %[[RHS_SPLAT_INSERT:.+]] = insertelement <4 x i64> undef, i64 %[[RHS]], i32 0
+  // CHECK: %[[RHS_SPLAT:.+]] = shufflevector <4 x i64> %[[RHS_SPLAT_INSERT]], <4 x i64> undef, <4 x i32> zeroinitializer
+  // CHECK: %[[NEZERO:.+]] = icmp ne <4 x i64> %[[COND]], zeroinitializer
+  // CHECK: %[[SELECT:.+]] = select <4 x i1> %[[NEZERO]], <4 x i64> %[[LHS]], <4 x i64> %[[RHS_SPLAT]]
+
+  four_ll ? four_ll : some_double;
+  // CHECK: %[[COND:.+]] = load <4 x i64>
+  // CHECK: %[[LHS:.+]] = load <4 x i64>
+  // CHECK: %[[RHS:.+]] = load double
+  // CHECK: %[[RHS_CONV:.+]] = fptosi double %[[RHS]] to i64
+  // CHECK: %[[RHS_SPLAT_INSERT:.+]] = insertelement <4 x i64> undef, i64 %[[RHS_CONV]], i32 0
+  // CHECK: %[[RHS_SPLAT:.+]] = shufflevector <4 x i64> %[[RHS_SPLAT_INSERT]], <4 x i64> undef, <4 x i32> zeroinitializer
+  // CHECK: %[[NEZERO:.+]] = icmp ne <4 x i64> %[[COND]], zeroinitializer
+  // CHECK: %[[SELECT:.+]] = select <4 x i1> %[[NEZERO]], <4 x i64> %[[LHS]], <4 x i64> %[[RHS_SPLAT]]
+}

clang/test/Sema/vector-gcc-compat.cpp

void logicTest(void) {
   v2i64 v2i64_a = (v2i64){0, 1}; // expected-warning {{compound literals are a C99-specific feature}}
   v2i64 v2i64_b = (v2i64){2, 1}; // expected-warning {{compound literals are a C99-specific feature}}
   v2i64 v2i64_c = (v2i64){3, 1}; // expected-warning {{compound literals are a C99-specific feature}}
   v2i64 v2i64_r;
 
   v2i64_r = !v2i64_a;  // expected-error {{invalid argument type 'v2i64' (vector of 2 'long long' values) to unary expression}}
   v2i64_r = ~v2i64_a;
 
-  v2i64_r = v2i64_a ? v2i64_b : v2i64_c; // expected-error {{value of type 'v2i64' (vector of 2 'long long' values) is not contextually convertible to 'bool'}}
+  v2i64_r = v2i64_a ? v2i64_b : v2i64_c;
 
   v2i64_r = v2i64_a & 1;
   v2i64_r = v2i64_a | 1;
   v2i64_r = v2i64_a ^ 1;
 
   v2i64_r = 1 & v2i64_a;
   v2i64_r = 1 | v2i64_a;
   v2i64_r = 1 ^ v2i64_a;

clang/test/SemaCXX/vector-conditional.cpp

+// RUN: %clang_cc1 -triple x86_64-linux-pc -fsyntax-only -verify -fexceptions -fcxx-exceptions %s -std=c++17
+// Note that this test depends on the size of long-long to be different from
+// int, so it specifies a triple.
+
+using FourShorts = short __attribute__((__vector_size__(8)));
+using TwoInts = int __attribute__((__vector_size__(8)));
+using TwoUInts = unsigned __attribute__((__vector_size__(8)));
+using FourInts = int __attribute__((__vector_size__(16)));
+using FourUInts = unsigned __attribute__((__vector_size__(16)));
+using TwoLongLong = long long __attribute__((__vector_size__(16)));
+using FourLongLong = long long __attribute__((__vector_size__(32)));
+using TwoFloats = float __attribute__((__vector_size__(8)));
+using FourFloats = float __attribute__((__vector_size__(16)));
+using TwoDoubles = double __attribute__((__vector_size__(16)));
+using FourDoubles = double __attribute__((__vector_size__(32)));
+
+FourShorts four_shorts;
+TwoInts two_ints;
+TwoUInts two_uints;
+FourInts four_ints;
+FourUInts four_uints;
+TwoLongLong two_ll;
+FourLongLong four_ll;
+TwoFloats two_floats;
+FourFloats four_floats;
+TwoDoubles two_doubles;
+FourDoubles four_doubles;
+
+enum E {};
+enum class SE {};
+E e;
+SE se;
+
+// Check the rules of the condition of the conditional operator.
+void Condition() {
+  // Only int types are allowed here, the rest should fail to convert to bool.
+  (void)(four_floats ? 1 : 1); // expected-error {{is not contextually convertible to 'bool'}}}
+  (void)(two_doubles ? 1 : 1); // expected-error {{is not contextually convertible to 'bool'}}}
+}
+
+// Check the rules of the LHS/RHS of the conditional operator.
+void Operands() {
+  (void)(four_ints ? four_ints : throw 1); // expected-error {{GNU vector conditional operand cannot be a throw expression}}
+  (void)(four_ints ? throw 1 : four_ints); // expected-error {{GNU vector conditional operand cannot be a throw expression}}
+  (void)(four_ints ?: throw 1);            // expected-error {{GNU vector conditional operand cannot be a throw expression}}
+  (void)(four_ints ? (void)1 : four_ints); // expected-error {{GNU vector conditional operand cannot be void}}
+  (void)(four_ints ?: (void)1);            // expected-error {{GNU vector conditional operand cannot be void}}
+
+  // Vector types must be the same element size as the condition.
+  (void)(four_ints ? two_ll : two_ll);             // expected-error {{vector condition type 'FourInts' (vector of 4 'int' values) and result type 'TwoLongLong' (vector of 2 'long long' values) do not have the same number of elements}}
+  (void)(four_ints ? four_ll : four_ll);           // expected-error {{vector condition type 'FourInts' (vector of 4 'int' values) and result type 'FourLongLong' (vector of 4 'long long' values) do not have elements of the same size}}
+  (void)(four_ints ? two_doubles : two_doubles);   // expected-error {{vector condition type 'FourInts' (vector of 4 'int' values) and result type 'TwoDoubles' (vector of 2 'double' values) do not have the same number of elements}}
+  (void)(four_ints ? four_doubles : four_doubles); // expected-error {{vector condition type 'FourInts' (vector of 4 'int' values) and result type 'FourDoubles' (vector of 4 'double' values) do not have elements of the same size}}
+  (void)(four_ints ?: two_ints);                   // expected-error {{vector operands to the vector conditional must be the same type ('FourInts' (vector of 4 'int' values) and 'TwoInts' (vector of 2 'int' values)}}
+  (void)(four_ints ?: four_doubles);               // expected-error {{vector operands to the vector conditional must be the same type ('FourInts' (vector of 4 'int' values) and 'FourDoubles' (vector of 4 'double' values)}}
+
+  // Scalars are promoted, but must be the same element size.
+  (void)(four_ints ? 3.0f : 3.0); // expected-error {{vector condition type 'FourInts' (vector of 4 'int' values) and result type '__attribute__((__vector_size__(4 * sizeof(double)))) double' (vector of 4 'double' values) do not have elements of the same size}}
+  (void)(four_ints ? 5ll : 5);    // expected-error {{vector condition type 'FourInts' (vector of 4 'int' values) and result type '__attribute__((__vector_size__(4 * sizeof(long long)))) long long' (vector of 4 'long long' values) do not have elements of the same size}}
+  (void)(four_ints ?: 3.0);       // expected-error {{cannot convert between scalar type 'double' and vector type 'FourInts' (vector of 4 'int' values) as implicit conversion would cause truncation}}
+  (void)(four_ints ?: 5ll);       // We allow this despite GCc not allowing this since we support integral->vector-integral conversions despite integer rank.
+
+  // This one would be allowed in GCC, but we don't allow vectors of enum. Also,
+  // the error message isn't perfect, since it is only going to be a problem
+  // when both sides are an enum, otherwise it'll be promoted to whatever type
+  // the other side causes.
+  (void)(four_ints ? e : e);                          // expected-error {{enumeral type 'E' is not allowed in a vector conditional}}
+  (void)(four_ints ? se : se);                        // expected-error {{enumeral type 'SE' is not allowed in a vector conditional}}
+  (void)(four_shorts ? (short)5 : (unsigned short)5); // expected-error {{vector condition type 'FourShorts' (vector of 4 'short' values) and result type '__attribute__((__vector_size__(4 * sizeof(int)))) int' (vector of 4 'int' values) do not have elements of the same size}}
+
+  // They must also be convertible.
+  (void)(four_ints ? 3.0f : 5u);
+  (void)(four_ints ? 3.0f : 5);
+  unsigned us = 5u;
+  int sint = 5;
+  short shrt = 5;
+  unsigned short uss = 5u;
+  // The following 2 error in GCC for truncation errors, but it seems
+  // unimportant and inconsistent to enforce that rule.
+  (void)(four_ints ? 3.0f : us);
+  (void)(four_ints ? 3.0f : sint);
+
+  // Test promotion:
+  (void)(four_shorts ? uss : shrt);  // expected-error {{vector condition type 'FourShorts' (vector of 4 'short' values) and result type '__attribute__((__vector_size__(4 * sizeof(int)))) int' (vector of 4 'int' values) do not have elements of the same size}}
+  (void)(four_shorts ? shrt : shrt); // should be fine.
+  (void)(four_ints ? uss : shrt);    // should be fine, since they get promoted to int.
+  (void)(four_ints ? shrt : shrt);   //expected-error {{vector condition type 'FourInts' (vector of 4 'int' values) and result type '__attribute__((__vector_size__(4 * sizeof(short)))) short' (vector of 4 'short' values) do not have elements of the same size}}
+
+  // Vectors must be the same type as eachother.
+  (void)(four_ints ? four_uints : four_floats); // expected-error {{vector operands to the vector conditional must be the same type ('FourUInts' (vector of 4 'unsigned int' values) and 'FourFloats' (vector of 4 'float' values))}}
+  (void)(four_ints ? four_uints : four_ints);   // expected-error {{vector operands to the vector conditional must be the same type ('FourUInts' (vector of 4 'unsigned int' values) and 'FourInts' (vector of 4 'int' values))}}
+  (void)(four_ints ? four_ints : four_uints);   // expected-error {{vector operands to the vector conditional must be the same type ('FourInts' (vector of 4 'int' values) and 'FourUInts' (vector of 4 'unsigned int' values))}}
+
+  // GCC rejects these, but our lax vector conversions don't seem to have a problem with them. Allow conversion of the float to an int as an extension.
+  (void)(four_ints ? four_uints : 3.0f);
+  (void)(four_ints ? four_ints : 3.0f);
+
+  // When there is a vector and a scalar, conversions must be legal.
+  (void)(four_ints ? four_floats : 3); // should work, ints can convert to floats.
+  (void)(four_ints ? four_uints : e);  // should work, non-scoped enum can convert to uint.
+  (void)(four_ints ? four_uints : se); // expected-error {{cannot convert between vector and non-scalar values ('FourUInts' (vector of 4 'unsigned int' values) and 'SE'}}
+  // GCC permits this, but our conversion rules reject this for truncation.
+  (void)(two_ints ? two_ints : us);        // expected-error {{cannot convert between scalar type 'unsigned int' and vector type 'TwoInts'}}
+  (void)(four_shorts ? four_shorts : uss); // expected-error {{cannot convert between scalar type 'unsigned short' and vector type 'FourShorts'}}
+  (void)(four_ints ? four_floats : us);    // expected-error {{cannot convert between scalar type 'unsigned int' and vector type 'FourFloats'}}
+  (void)(four_ints ? four_floats : sint);  // expected-error {{cannot convert between scalar type 'int' and vector type 'FourFloats'}}
+}
+
+template <typename T1, typename T2>
+struct is_same {
+  static constexpr bool value = false;
+};
+template <typename T>
+struct is_same<T, T> {
+  static constexpr bool value = true;
+};
+template <typename T1, typename T2>
+constexpr bool is_same_v = is_same<T1, T2>::value;
+template <typename T>
+T &&declval();
+
+// Check the result types when given two vector types.
+void ResultTypes() {
+  // Vectors must be the same, but result is the type of the LHS/RHS.
+  static_assert(is_same_v<TwoInts, decltype(declval<TwoInts>() ? declval<TwoInts>() : declval<TwoInts>())>);
+  static_assert(is_same_v<TwoFloats, decltype(declval<TwoInts>() ? declval<TwoFloats>() : declval<TwoFloats>())>);
+
+  // When both are scalars, converts to vectors of common type.
+  static_assert(is_same_v<TwoUInts, decltype(declval<TwoInts>() ? declval<int>() : declval<unsigned int>())>);
+
+  // Constant is allowed since it doesn't truncate, and should promote to float.
+  static_assert(is_same_v<TwoFloats, decltype(declval<TwoInts>() ? declval<float>() : 5u)>);
+  static_assert(is_same_v<TwoFloats, decltype(declval<TwoInts>() ? 5 : declval<float>())>);
+
+  // when only 1 is a scalar, it should convert to a compatible type.
+  static_assert(is_same_v<TwoFloats, decltype(declval<TwoInts>() ? declval<TwoFloats>() : declval<float>())>);
+  static_assert(is_same_v<TwoInts, decltype(declval<TwoInts>() ? declval<TwoInts>() : declval<int>())>);
+  static_assert(is_same_v<TwoFloats, decltype(declval<TwoInts>() ? declval<TwoFloats>() : 5)>);
+
+  // For the Binary conditional operator, the result type is either the vector on the RHS (that fits the rules on size/count), or the scalar extended to the correct count.
+  static_assert(is_same_v<TwoInts, decltype(declval<TwoInts>() ?: declval<TwoInts>())>);
+  static_assert(is_same_v<TwoInts, decltype(declval<TwoInts>() ?: declval<int>())>);
+}
+
+template <typename Cond>
+void dependent_cond(Cond C) {
+  (void)(C ? 1 : 2);
+}
+
+template <typename Operand>
+void dependent_operand(Operand C) {
+  (void)(two_ints ? 1 : C);
+  (void)(two_ints ? C : 1);
+  (void)(two_ints ? C : C);
+}
+
+template <typename Cond, typename LHS, typename RHS>
+void all_dependent(Cond C, LHS L, RHS R) {
+  (void)(C ? L : R);
+}
+
+// Check dependent cases.
+void Templates() {
+  dependent_cond(two_ints);
+  dependent_operand(two_floats);
+  // expected-error@159 {{vector operands to the vector conditional must be the same type ('__attribute__((__vector_size__(4 * sizeof(unsigned int)))) unsigned int' (vector of 4 'unsigned int' values) and '__attribute__((__vector_size__(4 * sizeof(double)))) double' (vector of 4 'double' values))}}}
+  all_dependent(four_ints, four_uints, four_doubles); // expected-note {{in instantiation of}}
+
+  // expected-error@159 {{vector operands to the vector conditional must be the same type ('__attribute__((__vector_size__(4 * sizeof(unsigned int)))) unsigned int' (vector of 4 'unsigned int' values) and '__attribute__((__vector_size__(2 * sizeof(unsigned int)))) unsigned int' (vector of 2 'unsigned int' values))}}}
+  all_dependent(four_ints, four_uints, two_uints); // expected-note {{in instantiation of}}
+  all_dependent(four_ints, four_uints, four_uints);
+}