[OpenCL] Add generic type handling for builtin functions

[svenvh]

Summary

Generic types are an abstraction of type sets. It mimics the way
functions are defined in the OpenCL specification. For example,
floatN can abstract all the vector sizes of the float type.

This allows to

• stick more closely to the specification, which uses generic types;
• factorize definitions of functions with numerous prototypes in the tablegen file; and
• reduce the memory impact of functions with many overloads.

Patch by Pierre Gondois and Sven van Haastregt.

Continuation of https://reviews.llvm.org/D63434

[svenvh]

Main changes since D63434:

• Rename List* to Vec*.
• Rename TLnn -> TLAll, TLInt, TLFloat.
• Apply clang-format.
• Improve/update documentation.
• Factor out renaming of base types into separate commit.
• Change return type of OCL2Qual.
• Remove default case from OCL2Qual switch statement: it should be fully covering the enum.

[Anastasia]

clang/lib/Sema/OpenCLBuiltins.td

51

Are the qualifiers added elsewhere?

221

Is this float GenType?

222

Would it make sense to use the same name scheme as below?

clang/lib/Sema/SemaLookup.cpp

680

What if both return and args use GenType, does GenTypeMaxCnt return max of all? Or do they have to align?

708

I don't get this logic?

755

Btw you are not marking param as (in) or (out) here.

817

I guess this should be done conditionally for C++ mode. But perhaps we don't have to do this now. It might be worth adding a FIXME.

clang/utils/TableGen/ClangOpenCLBuiltinEmitter.cpp

70

Perhaps not something for this patch, but I am wondering if a better name for this class would be OpenCLBuiltinTrieEmitter?

94

I am a bit confused about the purpose of this function as input and output seem to be both vectors of the Record. It might make sense to explain the difference. :)

225

I find this comment confusing... may be because it belongs to the code later. I am not sure it adds any value.

281

Is this logic to avoid duplicate signatures? If so it might be worth explaining it.

377

I feel it would be good to explain the structure of the function we are trying to generate i.e something like:

• We first generate switch/case statement over all type names that populates the list of type IDs
• Then we walk over the type IDs from the list and map them to the AST type and attributes accordingly.

404

I am not sure what you are trying to say about ExtVector...

457

I am really confused as the above comment on line 432 makes examples of non scalar types too i.e. int4.

[svenvh]

clang/lib/Sema/OpenCLBuiltins.td

51

Good point, this change should be part of the next patch (D63442). I've taken it out of this patch.

221

We will be adding more definitions for integers here in followup patches.

222

No, because FGenType is for all floating point types (half / float / double), as it uses the TLFloat type list. GenTypeFloatVecAndScalar is for the float type (i.e. fp32) only.

I will update the comments of both definitions to clarify the difference.

clang/lib/Sema/SemaLookup.cpp

680

They have to be compatible, otherwise we will hit the llvm_reachable below.

708

While trying to clarify this, I realized this checking should probably be moved to the TableGen emitter, as this is checking validity of the .td input so ideally we should check that at compiler compile time. @Pierre mentioned that this might not be trivial, but I'll have a look at it.

clang/utils/TableGen/ClangOpenCLBuiltinEmitter.cpp

70

It's emitting more than a trie; also tables and structs. How about OpenCLBuiltinEmitter (which aligns nicely with the file name)?

94

Clarified comment.

225

Elaborated.

281

Added comment.

377

Added comment.

404

Rephrased.

457

I guess you got confused about the two different steps; I have rephrased the comment on line 432.

[svenvh]

Addressing review comments.

[Pierre]

Just one comment,
Thank you again for making everything clearer :)

clang/lib/Sema/OpenCLBuiltins.td

110

Maybe it would be worth adding more information on how to use GenTypes. I was thinking of something like this :

When using multiple generic types :

• The maximal cardinal of the used GenTypes must be the PGCM of all the cardinals.
• The generic types combine as if it was an array like GenType[Type][VecSize].

I.e: With GT1 = [half, <1, 2>] and GT2 = [float, int, <1, 2>], they will combine as
<half, float>, <half2, float2>, <half, int>, <half, int2>
The maximal cardinal of GT1 and GT2 is 4 (= 2 types * 2 vector sizes).
4 is the PGCM of GT1 (=2) and GT2 (=4).

[svenvh]

• Move checking of GenType compatibility from SemaLookup to TableGen emitter.
• Provide more elaborate explanation about combining GenTypes in a declaration.
• Add max/min definitions to cover "sgentype" behavior and add test for max.
• Minor scattered comment improvements.

Combining GenTypes in an incorrect way will now produce an error from TableGen, e.g.:

OpenCLBuiltins.td:1601:1: error: number of vector sizes should be equal or 1 for all gentypes in a declaration
def : Builtin<"crashme", [GenTypeFloatVecNoScalar, GenTypeFloatVecAndScalar]>
^

OpenCLBuiltins.td:1602:1: error: number of types should be equal or 1 for all gentypes in a declaration
def : Builtin<"crashme2", [IGenTypeN, FGenTypeN]>;
^

[svenvh]

clang/lib/Sema/OpenCLBuiltins.td

110

I've added some rules about combining GenTypes in the comment below based on your example.

clang/lib/Sema/SemaLookup.cpp

708

The checking has been moved into ClangOpenCLBuiltinEmitter.cpp now, see VerifySignature.

817

Done.

[Anastasia]

clang/lib/Sema/OpenCLBuiltins.td

118

What does combining mean?

222

Ok, it could though be

GenTypeAllFloats

GenTypeF

Just that naming scheme feels a bit random right now.

clang/lib/Sema/SemaLookup.cpp

680

Do we say anywhere they have to be compatible?

clang/utils/TableGen/ClangOpenCLBuiltinEmitter.cpp

70

Yeah, makes sense!

264

I feel what you are checking is whether the vector sizes and types match but not just that the number of types match...

268

Can we rename B to something meaningful or add a comment explaining it?

272

I feel there is some black magic here. :)

[svenvh]

• Update comments as per review comments.
• Rename iterator List to VecSizes in OpenCLBuiltins.td
• Format GenericType definition.

clang/lib/Sema/OpenCLBuiltins.td

118

Clarified.

222

The convention is a bit subtle perhaps, but it is not random. For manual definitions that have multiple base types, the convention is typeinfo#GenType#vecinfo. For generated definitions that have a single base type, the convention is GenType#typeinfo#vecinfo. Having different conventions prevents the manual definitions from colliding with the generated ones.

clang/lib/Sema/SemaLookup.cpp

680

Yes, in OpenCLBuiltins.td near the definition of GenericType.

clang/utils/TableGen/ClangOpenCLBuiltinEmitter.cpp

264

"actual types" here means "final" non-generic types, such as half2, int, short4, etc.

Rephrased as "actual scalar or vector types", does that help?

268

Renamed to BuiltinRec.

272

Added some comments.

[Anastasia]

LGTM! Thanks!

clang/lib/Sema/OpenCLBuiltins.td

 def DefaultAS    : AddressSpace<"clang::LangAS::Default">;
 def PrivateAS    : AddressSpace<"clang::LangAS::opencl_private">;
 def GlobalAS     : AddressSpace<"clang::LangAS::opencl_global">;
 def ConstantAS   : AddressSpace<"clang::LangAS::opencl_constant">;
 def LocalAS      : AddressSpace<"clang::LangAS::opencl_local">;
 def GenericAS    : AddressSpace<"clang::LangAS::opencl_generic">;
 
 
-// Qualified Type. Allow to retrieve one ASTContext QualType.
-class QualType<string _Name> {
+// Qualified Type.  These map to ASTContext::QualType.
+class QualType<string _Name, bit _IsAbstract=0> {
   // Name of the field or function in a clang::ASTContext
   // E.g. Name="IntTy" for the int type, and "getIntPtrType()" for an intptr_t
   string Name = _Name;
+  // Some QualTypes in this file represent an abstract type for which there is
+  // no corresponding AST QualType, e.g. a GenType or an `image2d_t` type
+  // without access qualifiers.
+  bit IsAbstract = _IsAbstract;
 }
 
-// Helper class to store type access qualifiers (volatile, const, ...).
-class Qualifier<string _QualName> {
-  string QualName = _QualName;
+
+// List of integers.
+class IntList<string _Name, list<int> _List> {
+  string Name = _Name;
+  list<int> List = _List;
 }
 
 //===----------------------------------------------------------------------===//
 //                      OpenCL C classes for types
 //===----------------------------------------------------------------------===//
-// OpenCL types (int, float, ...)
+// OpenCL C basic data types (int, float, image2d_t, ...).
+// Its Child classes can represent concrete types (e.g.: VectorType) or
+// custom types (e.g.: GenType).
+// Instances of these child classes should be used in Builtin function
+// arguments.  See the definition of the "read_imagef" function as example.
 class Type<string _Name, QualType _QTName> {
-  // Name of the Type
+  // Name of the Type.
   string Name = _Name;
-  // QualType associated with this type
+  // QualType associated with this type.
   QualType QTName = _QTName;
-  // Size of the vector (if applicable)
-  int VecWidth = 0;
-  // Is pointer
+  // Size of the vector (if applicable).
+  int VecWidth = 1;
+  // Is a pointer.
   bit IsPointer = 0;
-  // List of qualifiers associated with the type (volatile, ...)
-  list<Qualifier> QualList = [];
-  // Address space
-  string AddrSpace = "clang::LangAS::Default";
   // Access qualifier. Must be one of ("RO", "WO", "RW").
   string AccessQualifier = "";
+  // Address space.
+  string AddrSpace = "clang::LangAS::Default";
 }
 
 // OpenCL vector types (e.g. int2, int3, int16, float8, ...)
 class VectorType<Type _Ty, int _VecWidth> : Type<_Ty.Name, _Ty.QTName> {
   int VecWidth = _VecWidth;
 }
 
-// OpenCL pointer types (e.g. int*, float*, ...)
+// OpenCL pointer types (e.g. int*, float*, ...).
 class PointerType<Type _Ty, AddressSpace _AS = GlobalAS> :
                                       Type<_Ty.Name, _Ty.QTName> {
   bit IsPointer = 1;
   string AddrSpace = _AS.Name;
 }
 
 // OpenCL image types (e.g. image2d_t, ...)
 class ImageType<Type _Ty, QualType _QTName, string _AccessQualifier> :
                                               Type<_Ty.Name, _QTName> {
   let AccessQualifier = _AccessQualifier;
 }
 
+// List of Types.
+class TypeList<string _Name, list<Type> _Type> {
+  string Name = _Name;
+  list<Type> List = _Type;
+}
+
+// A GenericType is an abstract type that defines a set of types as a
+// combination of Types and vector sizes.
+//
+// E.g.: If TypeListField = <int, float> and VectorList = <1, 2, 4>, then
+//       then it represents <int, int2, int4, float, float2, float4>.
+// _Ty          : Name of the GenType.
+// _TypeList    : Types this GenericType can be.
+// _VectorList  : Sizes of the vector for each type of the _TypeList. 0 being a
+//                scalar.
+class GenericType<string _Ty, TypeList _TypeList, IntList _VectorList> :
+                                            Type<_Ty, QualType<"null", 1>> {
+  // Possible element types of the generic type.
+  TypeList TypeListField = _TypeList;
+  // Possible vector sizes of the types in the TypeListField.
+  IntList VectorList = _VectorList;
+  // The VecWidth field is ignored for GenericTypes. Use VectorList instead.
+  let VecWidth = 0;
+}
+
 //===----------------------------------------------------------------------===//
 //                      OpenCL C class for builtin functions
 //===----------------------------------------------------------------------===//
 class Builtin<string _Name, list<Type> _Signature> {
   // Name of the builtin function
   string Name = _Name;
   // List of types used by the function. The first one is the return type and
   // the following are the arguments. The list must have at least one element
   // (the return type).
   list<Type> Signature = _Signature;
   // OpenCL Extension to which the function belongs (cl_khr_subgroups, ...)
   string Extension = "";
   // OpenCL Version to which the function belongs (CL10, ...)
   Version Version = CL10;
 }
 
 //===----------------------------------------------------------------------===//
-//                           Multiclass definitions
-//===----------------------------------------------------------------------===//
-// multiclass BifN: Creates Builtin class instances for OpenCL builtin
-//                  functions with N arguments.
-// _Name      : Name of the function
-// _Signature : Signature of the function (list of the Type used by the
-//              function, the first one being the return type).
-// _IsVector  : List of bit indicating if the type in the _Signature at the
-//              same index is to be a vector in the multiple overloads. The
-//              list must have at least one non-zero value.
-multiclass Bif0<string _Name, list<Type> _Signature, list<bit> _IsVector> {
-  def : Builtin<_Name, _Signature>;
-  foreach v = [2, 3, 4, 8, 16] in {
-    def : Builtin<_Name,
-                  [!if(_IsVector[0], VectorType<_Signature[0], v>, _Signature[0])]>;
-  }
-}
-multiclass Bif1<string _Name, list<Type> _Signature, list<bit> _IsVector> {
-  def : Builtin<_Name, _Signature>;
-  foreach v = [2, 3, 4, 8, 16] in {
-    def : Builtin<_Name,
-                  [!if(_IsVector[0], VectorType<_Signature[0], v>, _Signature[0]),
-                  !if(_IsVector[1], VectorType<_Signature[1], v>, _Signature[1])]>;
-  }
-}
-multiclass Bif2<string _Name, list<Type> _Signature, list<bit> _IsVector> {
-  def : Builtin<_Name, _Signature>;
-  foreach v = [2, 3, 4, 8, 16] in {
-    def : Builtin<_Name,
-                  [!if(_IsVector[0], VectorType<_Signature[0], v>, _Signature[0]),
-                  !if(_IsVector[1], VectorType<_Signature[1], v>, _Signature[1]),
-                  !if(_IsVector[2], VectorType<_Signature[2], v>, _Signature[2])]>;
-  }
-}
-multiclass Bif3<string _Name, list<Type> _Signature, list<bit> _IsVector> {
-  def : Builtin<_Name, _Signature>;
-  foreach v = [2, 3, 4, 8, 16] in {
-    def : Builtin<_Name,
-                  [!if(_IsVector[0], VectorType<_Signature[0], v>, _Signature[0]),
-                  !if(_IsVector[1], VectorType<_Signature[1], v>, _Signature[1]),
-                  !if(_IsVector[2], VectorType<_Signature[2], v>, _Signature[2]),
-                  !if(_IsVector[3], VectorType<_Signature[3], v>, _Signature[3])]>;
-  }
-}
-//===----------------------------------------------------------------------===//
 //                 Definitions of OpenCL C types
 //===----------------------------------------------------------------------===//
-// OpenCL v1.2 s6.1.1: Built-in Scalar Data Types
+
+// OpenCL v1.0/1.2/2.0 s6.1.1: Built-in Scalar Data Types.
 def Bool      : Type<"bool",      QualType<"BoolTy">>;
 def Char      : Type<"char",      QualType<"CharTy">>;
 def UChar     : Type<"uchar",     QualType<"UnsignedCharTy">>;
 def Short     : Type<"short",     QualType<"ShortTy">>;
 def UShort    : Type<"ushort",    QualType<"UnsignedShortTy">>;
 def Int       : Type<"int",       QualType<"IntTy">>;
 def UInt      : Type<"uint",      QualType<"UnsignedIntTy">>;
 def Long      : Type<"long",      QualType<"LongTy">>;
 def ULong     : Type<"ulong",     QualType<"UnsignedLongTy">>;
 def Float     : Type<"float",     QualType<"FloatTy">>;
 def Double    : Type<"double",    QualType<"DoubleTy">>;
 def Half      : Type<"half",      QualType<"HalfTy">>;
 def Size      : Type<"size_t",    QualType<"getSizeType()">>;
 def PtrDiff   : Type<"ptrdiff_t", QualType<"getPointerDiffType()">>;
 def IntPtr    : Type<"intptr_t",  QualType<"getIntPtrType()">>;
 def UIntPtr   : Type<"uintPtr_t", QualType<"getUIntPtrType()">>;
 def Void      : Type<"void_t",    QualType<"VoidTy">>;
 
-// OpenCL v1.2 s6.1.2: Built-in Vector Data Types
-foreach v = [2, 3, 4, 8, 16] in {
-  def char#v#_t    : VectorType<Char, v>;
-  def uchar#v#_t   : VectorType<UChar, v>;
-  def short#v#_t   : VectorType<Short, v>;
-  def ushort#v#_t  : VectorType<UShort, v>;
-  def "int"#v#_t   : VectorType<Int, v>;
-  def uint#v#_t    : VectorType<UInt, v>;
-  def long#v#_t    : VectorType<Long, v>;
-  def ulong#v#_t   : VectorType<ULong, v>;
-  def float#v#_t   : VectorType<Float, v>;
-  def double#v#_t  : VectorType<Double, v>;
-  def half#v#_t    : VectorType<Half, v>;
-}
+// OpenCL v1.0/1.2/2.0 s6.1.2: Built-in Vector Data Types.
+// Built-in vector data types are created by TableGen's OpenCLBuiltinEmitter.
 
 // OpenCL v1.2 s6.1.3: Other Built-in Data Types
 // These definitions with a "null" name are "abstract". They should not
 // be used in definitions of Builtin functions.
-def image2d_t         : Type<"image2d_t", QualType<"null">>;
-def image3d_t         : Type<"image3d_t", QualType<"null">>;
-def image2d_array_t   : Type<"image2d_array_t", QualType<"null">>;
-def image1d_t         : Type<"image1d_t", QualType<"null">>;
-def image1d_buffer_t  : Type<"image1d_buffer_t", QualType<"null">>;
-def image1d_array_t   : Type<"image1d_array_t", QualType<"null">>;
+def image2d_t         : Type<"image2d_t", QualType<"null", 1>>;
+def image3d_t         : Type<"image3d_t", QualType<"null", 1>>;
+def image2d_array_t   : Type<"image2d_array_t", QualType<"null", 1>>;
+def image1d_t         : Type<"image1d_t", QualType<"null", 1>>;
+def image1d_buffer_t  : Type<"image1d_buffer_t", QualType<"null", 1>>;
+def image1d_array_t   : Type<"image1d_array_t", QualType<"null", 1>>;
 // Unlike the few functions above, the following definitions can be used
 // in definitions of Builtin functions (they have a QualType with a name).
 foreach v = ["RO", "WO", "RW"] in {
   def image2d_#v#_t       : ImageType<image2d_t,
                                       QualType<"OCLImage2d"#v#"Ty">,
                                       v>;
   def image3d_#v#_t       : ImageType<image3d_t,
                                       QualType<"OCLImage3d"#v#"Ty">,
def image1d_array#v#_t  : ImageType<image1d_array_t,
                                       QualType<"OCLImage1dArray"#v#"Ty">,
                                       v>;
 }
 
 def Sampler           : Type<"Sampler", QualType<"OCLSamplerTy">>;
 def Event             : Type<"Event", QualType<"OCLEventTy">>;
 
 //===----------------------------------------------------------------------===//
+//                 Definitions of OpenCL gentype variants
+//===----------------------------------------------------------------------===//
+// The OpenCL specification often uses "gentype" in builtin function
+// declarations to indicate that a builtin function is available with various
+// argument and return types.  The types represented by "gentype" vary between
+// different parts of the specification.  The following definitions capture
+// the different type lists for gentypes in different parts of the
+// specification.
+
+// Vector width lists.
+def VecAndScalar: IntList<"VecAndScalar", [1, 2, 3, 4, 8, 16]>;
+def VecNoScalar : IntList<"VecNoScalar", [2, 3, 4, 8, 16]>;
+
+// Type lists.
+def TLFloat : TypeList<"TLFloat", [Float, Double, Half]>;
+
+// GenType definitions.
+def FGenTypeN   : GenericType<"FGenTypeN", TLFloat, VecAndScalar>;
+// Generate basic GenTypes.  Names are like: GenTypeFloatVecAndScalar.
+foreach Type = [Char, UChar, Short, UShort,
+                Int, UInt, Long, ULong,
+                Float, Double, Half] in {
+  foreach List = [VecAndScalar, VecNoScalar] in {
+    def "GenType"#Type#List :
+              GenericType<"GenType" # Type # List,
+                          TypeList<"GL" # Type.Name, [Type]>,
+                          List>;
+  }
+}
+
+
+//===----------------------------------------------------------------------===//
 //                 Definitions of OpenCL builtin functions
 //===----------------------------------------------------------------------===//
 //--------------------------------------------------------------------
 // OpenCL v1.1/1.2/2.0 s6.2.3 - Explicit conversions.
 // OpenCL v2.0 Extensions s5.1.1 and s6.1.1 - Conversions.
 
 // Generate the convert_* builtins functions.
 foreach RType = [Float, Double, Half, Char, UChar, Short,
foreach name = ["get_global_size", "get_global_id", "get_local_size",
                 "get_global_offset"] in {
   def : Builtin<name, [Size, UInt]>;
 }
 
 // OpenCL v1.2 s6.12.2: Math Functions
 foreach name = ["acos", "acosh", "acospi",
                 "asin", "asinh", "asinpi",
                 "atan", "atanh", "atanpi"] in {
-  foreach type = [Float, Double, Half] in {
-    defm : Bif1<name, [type, type], [1, 1]>;
-  }
+  def : Builtin<name, [FGenTypeN, FGenTypeN]>;
 }
 
 foreach name = ["atan2", "atan2pi"] in {
-  foreach type = [Float, Double, Half] in {
-    defm : Bif2<name, [type, type, type], [1, 1, 1]>;
-  }
+  def : Builtin<name, [FGenTypeN, FGenTypeN, FGenTypeN]>;
 }
 
 foreach name = ["fmax", "fmin"] in {
-  foreach type = [Float, Double, Half] in {
-    defm : Bif2<name, [type, type, type], [1, 1, 1]>;
-    defm : Bif2<name, [type, type, type], [1, 1, 0]>;
-  }
+  def : Builtin<name, [FGenTypeN, FGenTypeN, FGenTypeN]>;
+  def : Builtin<name, [GenTypeFloatVecNoScalar, GenTypeFloatVecNoScalar, Float]>;
+  def : Builtin<name, [GenTypeDoubleVecNoScalar, GenTypeDoubleVecNoScalar, Double]>;
+  def : Builtin<name, [GenTypeHalfVecNoScalar, GenTypeHalfVecNoScalar, Half]>;
 }
 
 // OpenCL v1.2 s6.12.14: Built-in Image Read Functions
 def read_imagef : Builtin<"read_imagef",
-                          [float4_t, image2d_RO_t, VectorType<Int, 2>]>;
+                          [VectorType<Float, 4>, image2d_RO_t, VectorType<Int, 2>]>;
 def write_imagef : Builtin<"write_imagef",
                            [Void,
                             image2d_WO_t,
                             VectorType<Int, 2>,
                             VectorType<Float, 4>]>;
 
 
 // OpenCL v2.0 s9.17.3: Additions to section 6.13.1: Work-Item Functions
 let Version = CL20 in {
   let Extension = "cl_khr_subgroups" in {
     def get_sub_group_size : Builtin<"get_sub_group_size", [UInt]>;
     def get_max_sub_group_size : Builtin<"get_max_sub_group_size", [UInt]>;
     def get_num_sub_groups : Builtin<"get_num_sub_groups", [UInt]>;
   }
 }

clang/lib/Sema/SemaLookup.cpp

 
 LLVM_DUMP_METHOD void LookupResult::dump() {
   llvm::errs() << "lookup results for " << getLookupName().getAsString()
                << ":\n";
   for (NamedDecl *D : *this)
     D->dump();
 }
 
-/// When trying to resolve a function name, if the isOpenCLBuiltin function
-/// defined in "OpenCLBuiltins.inc" returns a non-null <Index, Len>, then the
-/// identifier is referencing an OpenCL builtin function. Thus, all its
-/// prototypes are added to the LookUpResult.
-///
-/// \param S The Sema instance
-/// \param LR  The LookupResult instance
-/// \param II  The identifier being resolved
-/// \param Index  The list of prototypes starts at Index in OpenCLBuiltins[]
-/// \param Len  The list of prototypes has Len elements
-static void InsertOCLBuiltinDeclarations(Sema &S, LookupResult &LR,
-                                         IdentifierInfo *II, unsigned Index,
-                                         unsigned Len) {
+/// Helper function to get the QualType instances of the return type and
+/// arguments for OpenCL library functions
+/// \param Context (in) The Context instance.
+/// \param OpenCLBuiltin (in) The signature currently handled.
+/// \param GenTypeMaxCnt (out) Maximum number of types contained in a generic
+///        type used as return type or as argument.
+///        Only meaningful for generic types, otherwise equals 1.
+/// \param RetTypes (out) List of the possible return types.
+/// \param ArgTypes (out) List of the possible types for the arguments.
+static void GetQualTypesForOpenCLBuiltin(
+    ASTContext &Context, const OpenCLBuiltinStruct &OpenCLBuiltin,
+    unsigned &GenTypeMaxCnt, std::vector<QualType> &RetTypes,
+    SmallVector<std::vector<QualType>, 5> &ArgTypes) {
+  // Get the QualType instances of the return types.
+  unsigned Sig = SignatureTable[OpenCLBuiltin.SigTableIndex];
+  OCL2Qual(Context, TypeTable[Sig], RetTypes);
+  GenTypeMaxCnt = RetTypes.size();
+
+  // Get the QualType instances of the arguments.
+  // First type is the return type, skip it.
+  for (unsigned Index = 1; Index < OpenCLBuiltin.NumTypes; Index++) {
+    std::vector<QualType> Ty;
+    OCL2Qual(Context,
+        TypeTable[SignatureTable[OpenCLBuiltin.SigTableIndex + Index]], Ty);
+    ArgTypes.push_back(Ty);
+    GenTypeMaxCnt = (Ty.size() > GenTypeMaxCnt) ? Ty.size() : GenTypeMaxCnt;
+  }
 
-  for (unsigned i = 0; i < Len; ++i) {
-    const OpenCLBuiltinDecl &Decl = OpenCLBuiltins[Index - 1 + i];
-    ASTContext &Context = S.Context;
+  // Check that that lengths of the gentypes are compatibles.
+  if ((GenTypeMaxCnt % RetTypes.size()) != 0) {
+    goto error_handler;
+  }
+  for (unsigned Index = 0; Index < ArgTypes.size(); Index++) {
+    unsigned TypeCntAtIndex = ArgTypes[Index].size();
+    if ((GenTypeMaxCnt % TypeCntAtIndex) != 0) {
+      goto error_handler;
+    }
+  }
 
-    // Ignore this BIF if the version is incorrect.
-    if (Context.getLangOpts().OpenCLVersion < Decl.Version)
-      continue;
+  return;
 
-    FunctionProtoType::ExtProtoInfo PI;
-    PI.Variadic = false;
+error_handler:
+  llvm_unreachable("Lengths of genTypes used are incompatible in the "
+                   "definition of the builtin function");
+}
+
+/// Helper function to create a list of the possible signatures for the
+/// OpenCL library functions.
+/// \param Context (in) The Context instance.
+/// \param GenTypeMaxCnt (in) Maximum number of types contained in a generic
+///        type used as return type or as argument.
+///        Only meaningful for generic types, otherwise equals 1.
+/// \param FunctionList (out) List of the function declarations.
+/// \param RetTypes (in) List of the possible return types.
+/// \param ArgTypes (in) List of the possible types for the arguments.
+static void
+GetOpenCLBuiltinFctOverloads(ASTContext &Context, unsigned GenTypeMaxCnt,
+                             std::vector<QualType> &FunctionList,
+                             std::vector<QualType> &RetTypes,
+                             SmallVector<std::vector<QualType>, 5> &ArgTypes) {
 
-    // Defined in "OpenCLBuiltins.inc"
-    QualType RT = OCL2Qual(Context, OpenCLSignature[Decl.ArgTableIndex]);
+  for (unsigned IGenType = 0; IGenType < GenTypeMaxCnt; IGenType++) {
+    SmallVector<QualType, 5> ArgList;
 
-    SmallVector<QualType, 5> ArgTypes;
-    for (unsigned I = 1; I < Decl.NumArgs; I++) {
-      QualType Ty = OCL2Qual(Context, OpenCLSignature[Decl.ArgTableIndex + I]);
-      ArgTypes.push_back(Ty);
+    for (unsigned I = 0; I < ArgTypes.size(); I++) {
+      ArgList.push_back(ArgTypes[I][IGenType % ArgTypes[I].size()]);
     }
 
-    QualType R = Context.getFunctionType(RT, ArgTypes, PI);
-    SourceLocation Loc = LR.getNameLoc();
+    FunctionList.push_back(Context.getFunctionType(
+        RetTypes[(RetTypes.size() != 1) ? IGenType : 0], ArgList, PI));
+  }
+}
 
-    // TODO: This part is taken from Sema::LazilyCreateBuiltin,
-    // maybe refactor it.
+/// When trying to resolve a function name, if isOpenCLBuiltin() returns a
+/// non-null <Index, Len> pair, then the name is referencing an OpenCL
+/// builtin function.  Add all candidate signatures to the LookUpResult.
+///
+/// \param S The Sema instance.
+/// \param LR The LookupResult instance.
+/// \param II The identifier being resolved.
+/// \param FctIndex Starting index in the BuiltinTable.
+/// \param Len The list of signatures has Len elements.
+static void InsertOCLBuiltinDeclarationsFromTable(Sema &S, LookupResult &LR,
+                                                  IdentifierInfo *II,
+                                                  const unsigned FctIndex,
+                                                  const unsigned Len) {
+  // The builtin function declaration uses generic types (gentype).
+  bool HasGenType = false;
+
+  // Maximum number of types contained in a generic type used as return type or
+  // as argument.  Only meaningful for generic types, otherwise equals 1.
+  unsigned GenTypeMaxCnt;
+
+  for (unsigned SignatureIndex = 0; SignatureIndex < Len; SignatureIndex++) {
+    const OpenCLBuiltinStruct &OpenCLBuiltin =
+        BuiltinTable[FctIndex + SignatureIndex];
+    ASTContext &Context = S.Context;
+
+    std::vector<QualType> RetTypes;
+    SmallVector<std::vector<QualType>, 5> ArgTypes;
+
+    GetQualTypesForOpenCLBuiltin(Context, OpenCLBuiltin, GenTypeMaxCnt,
+                                 RetTypes, ArgTypes);
+
+    if (GenTypeMaxCnt > 1) {
+      HasGenType = true;
+    }
+
+    // Create function overload for each type combination.
+    std::vector<QualType> FunctionList;
+    GetOpenCLBuiltinFctOverloads(Context, GenTypeMaxCnt, FunctionList, RetTypes,
+                                 ArgTypes);
+
+    SourceLocation Loc = LR.getNameLoc();
     DeclContext *Parent = Context.getTranslationUnitDecl();
-    FunctionDecl *New = FunctionDecl::Create(Context, Parent, Loc, Loc, II, R,
-                                             /*TInfo=*/nullptr, SC_Extern,
-                                             false, R->isFunctionProtoType());
-    New->setImplicit();
+    FunctionDecl *NewOpenCLBuiltin;
+
+    for (unsigned Index = 0; Index < GenTypeMaxCnt; Index++) {
+      NewOpenCLBuiltin = FunctionDecl::Create(
+          Context, Parent, Loc, Loc, II, FunctionList[Index],
+          /*TInfo=*/nullptr, SC_Extern, false,
+          FunctionList[Index]->isFunctionProtoType());
+      NewOpenCLBuiltin->setImplicit();
 
       // Create Decl objects for each parameter, adding them to the
       // FunctionDecl.
-    if (const FunctionProtoType *FT = dyn_cast<FunctionProtoType>(R)) {
-      SmallVector<ParmVarDecl *, 16> Params;
-      for (unsigned i = 0, e = FT->getNumParams(); i != e; ++i) {
-        ParmVarDecl *Parm =
-            ParmVarDecl::Create(Context, New, SourceLocation(),
-                                SourceLocation(), nullptr, FT->getParamType(i),
-                                /*TInfo=*/nullptr, SC_None, nullptr);
-        Parm->setScopeInfo(0, i);
-        Params.push_back(Parm);
+      if (const FunctionProtoType *FP =
+              dyn_cast<FunctionProtoType>(FunctionList[Index])) {
+        SmallVector<ParmVarDecl *, 16> ParmList;
+        for (unsigned IParm = 0, e = FP->getNumParams(); IParm != e; ++IParm) {
+          ParmVarDecl *Parm = ParmVarDecl::Create(
+              Context, NewOpenCLBuiltin, SourceLocation(), SourceLocation(),
+              nullptr, FP->getParamType(IParm),
+          Parm->setScopeInfo(0, IParm);
+          ParmList.push_back(Parm);
         }
-      New->setParams(Params);
+        NewOpenCLBuiltin->setParams(ParmList);
+      }
+      NewOpenCLBuiltin->addAttr(OverloadableAttr::CreateImplicit(Context));
+      LR.addDecl(NewOpenCLBuiltin);
     }
-
-    New->addAttr(OverloadableAttr::CreateImplicit(Context));
-
-    if (strlen(Decl.Extension))
-      S.setOpenCLExtensionForDecl(New, Decl.Extension);
-
-    LR.addDecl(New);
   }
 
   // If we added overloads, need to resolve the lookup result.
-  if (Len > 1)
+  if (Len > 1 || HasGenType)
     LR.resolveKind();
 }
 
 /// Lookup a builtin function, when name lookup would otherwise
 /// fail.
 static bool LookupBuiltin(Sema &S, LookupResult &R) {
   Sema::LookupNameKind NameKind = R.getLookupKind();
 
if (II) {
           return true;
         }
       }
 
       // Check if this is an OpenCL Builtin, and if so, insert its overloads.
       if (S.getLangOpts().OpenCL && S.getLangOpts().DeclareOpenCLBuiltins) {
         auto Index = isOpenCLBuiltin(II->getName());
         if (Index.first) {
-          InsertOCLBuiltinDeclarations(S, R, II, Index.first, Index.second);
+          InsertOCLBuiltinDeclarationsFromTable(S, R, II, Index.first - 1,
+                                                Index.second);
           return true;
         }
       }
 
       // If this is a builtin on this (or all) targets, create the decl.
       if (unsigned BuiltinID = II->getBuiltinID()) {
         // In C++ and OpenCL (spec v1.2 s6.9.f), we don't have any predefined
         // library functions like 'malloc'. Instead, we'll just error.

clang/test/SemaOpenCL/fdeclare-opencl-builtins.cl

 // RUN: %clang_cc1 %s -triple spir -verify -pedantic -fsyntax-only -cl-std=CL2.0 -fdeclare-opencl-builtins -DNO_HEADER
 // RUN: %clang_cc1 %s -triple spir -verify -pedantic -fsyntax-only -cl-std=CL2.0 -fdeclare-opencl-builtins -finclude-default-header
+// expected-no-diagnostics
 
 // Test the -fdeclare-opencl-builtins option.
 
 // Provide typedefs when invoking clang without -finclude-default-header.
 #ifdef NO_HEADER
+typedef char char2 __attribute__((ext_vector_type(2)));
 typedef float float4 __attribute__((ext_vector_type(4)));
-typedef int int4 __attribute__((ext_vector_type(4)));
 typedef int int2 __attribute__((ext_vector_type(2)));
+typedef int int4 __attribute__((ext_vector_type(4)));
+typedef long long2 __attribute__((ext_vector_type(2)));
 typedef unsigned int uint;
 typedef __SIZE_TYPE__ size_t;
 #endif
 
-kernel void basic_conversion(global float4 *buf, global int4 *res) {
-  res[0] = convert_int4(buf[0]);
-}
+kernel void basic_conversion() {
+  double d;
+  float f;
+  char2 c2;
+  long2 l2;
+  float4 f4;
+  int4 i4;
 
-kernel void basic_readonly_image_type(__read_only image2d_t img, int2 coord, global float4 *out) {
-  out[0] = read_imagef(img, coord);
+  f = convert_float(d);
+  d = convert_double_sat_rtp(f);
+  l2 = convert_long2_rtz(c2);
+  i4 = convert_int4_sat(f4);
 }
 
 kernel void basic_subgroup(global uint *out) {
   out[0] = get_sub_group_size();
-// expected-error@-1{{use of declaration 'get_sub_group_size' requires cl_khr_subgroups extension to be enabled}}
-#pragma OPENCL EXTENSION cl_khr_subgroups : enable
-  out[1] = get_sub_group_size();
 }

clang/utils/TableGen/ClangOpenCLBuiltinEmitter.cpp

 //
 // This tablegen backend emits code for checking whether a function is an
 // OpenCL builtin function. If so, all overloads of this function are
 // added to the LookupResult. The generated include file is used by
 // SemaLookup.cpp
 //
 // For a successful lookup of e.g. the "cos" builtin, isOpenCLBuiltin("cos")
 // returns a pair <Index, Len>.
-// OpenCLBuiltins[Index] to OpenCLBuiltins[Index + Len] contains the pairs
+// BuiltinTable[Index] to BuiltinTable[Index + Len] contains the pairs
 // <SigIndex, SigLen> of the overloads of "cos".
-// OpenCLSignature[SigIndex] to OpenCLSignature[SigIndex + SigLen] contains
-// one of the signatures of "cos". The OpenCLSignature entry can be
-// referenced by other functions, i.e. "sin", since multiple OpenCL builtins
-// share the same signature.
+// SignatureTable[SigIndex] to SignatureTable[SigIndex + SigLen] contains
+// one of the signatures of "cos". The SignatureTable entry can be
+// referenced by other functions, e.g. "sin", to exploit the fact that
+// many OpenCL builtins share the same signature.
+//
+// The file generated by this TableGen emitter contains the following:
+//
+//  * Structs and enums to represent types and function signatures.
+//
+//  * OpenCLTypeStruct TypeTable[]
+//    Type information for return types and arguments.
+//
+//  * unsigned SignatureTable[]
+//    A list of types representing function signatures.  Each entry is an index
+//    into the above TypeTable.  Multiple entries following each other form a
+//    signature, where the first entry is the return type and subsequent
+//    entries are the argument types.
+//
+//  * OpenCLBuiltinStruct BuiltinTable[]
+//    Each entry represents one overload of an OpenCL builtin function and
+//    consists of an index into the SignatureTable and the number of arguments.
+//
+//  * std::pair<unsigned, unsigned> isOpenCLBuiltin(llvm::StringRef Name)
+//    Find out whether a string matches an existing OpenCL builtin function
+//    name and return an index into BuiltinTable and the number of overloads.
+//
+//  * void OCL2Qual(ASTContext&, OpenCLTypeStruct, std::vector<QualType>&)
+//    Convert an OpenCLTypeStruct type to a list of QualType instances.
+//    One OpenCLTypeStruct can represent multiple types, primarily when using
+//    GenTypes.
+//
 //===----------------------------------------------------------------------===//
 
 #include "llvm/ADT/MapVector.h"
 #include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/SmallString.h"
 #include "llvm/ADT/StringExtras.h"
 #include "llvm/ADT/StringRef.h"
 #include "llvm/ADT/StringSet.h"
 #include "llvm/Support/ErrorHandling.h"
 #include "llvm/Support/raw_ostream.h"
 #include "llvm/TableGen/Error.h"
 #include "llvm/TableGen/Record.h"
 #include "llvm/TableGen/StringMatcher.h"
 #include "llvm/TableGen/TableGenBackend.h"
 #include <set>
 
 using namespace llvm;
 
 namespace {
 class BuiltinNameEmitter {
 public:
   BuiltinNameEmitter(RecordKeeper &Records, raw_ostream &OS)
       : Records(Records), OS(OS) {}
 
   // Entrypoint to generate the functions and structures for checking
   // whether a function is an OpenCL builtin function.
   void Emit();
 
 private:
   // Contains OpenCL builtin functions and related information, stored as
   // Record instances. They are coming from the associated TableGen file.
   RecordKeeper &Records;
 
   // The output file.
   raw_ostream &OS;
 
-  // Emit the enums and structs.
+  // Helper function for BuiltinNameEmitter::EmitDeclarations.  Generate enum
+  // definitions in the Output string parameter, and save their Record instance
+  // in the List parameter.
+  // \param Types (in) List containing the Types to extract.
+  // \param TypesSeen (out) List containing the Types already extracted.
+  // \param Output (out) String containing the enums to emit in the output file.
+  // \param List (out) List to fill with the extracted types.
+  void ExtractEnumTypes(std::vector<Record *> &Types,
+                        StringMap<bool> &TypesSeen, std::string &Output,
+                        std::vector<const Record *> &List);
+
+  // Emit the enum or struct used in the generated file.
+  // Populate the TypeList at the same time.
   void EmitDeclarations();
 
-  // Parse the Records generated by TableGen and populate OverloadInfo and
-  // SignatureSet.
+  // Parse the Records generated by TableGen to populate the SignaturesList,
+  // FctOverloadMap and TypeMap.
   void GetOverloads();
 
-  // Emit the OpenCLSignature table. This table contains all possible
-  // signatures, and is a struct OpenCLType. A signature is composed of a
-  // return type (mandatory), followed by zero or more argument types.
+  // Emit the TypeTable. This table contains all the possible types. A type can
+  // have several attributes (e.g. vector size, whether it is a pointer type,
+  // constness, ...).
+  // For example
+  //   {OCLT_Double, 2},
+  // is describing the type "Double", being a vector of 2 elements. See the
+  // struct OpenCLTypeStruct for more information.
+  void EmitTypeTable();
+
+  // Emit the SignatureTable. This table contains all the possible signatures.
+  // A signature is stored as a list of indexes of the TypeTable.
+  // The first index references the return type (mandatory), and the followings
+  // reference its arguments.
   // E.g.:
-  // // 12
-  // { OCLT_uchar, 4, clang::LangAS::Default, false },
-  // { OCLT_float, 4, clang::LangAS::Default, false },
-  // This means that index 12 represents a signature
-  //   - returning a uchar vector of 4 elements, and
-  //   - taking as first argument a float vector of 4 elements.
+  // 15, 2, 15 can represent a function with the signature:
+  // int func(float, int)
+  // The "int" type being at the index 15 in the TypeTable.
   void EmitSignatureTable();
 
-  // Emit the OpenCLBuiltins table. This table contains all overloads of
+  // Emit the BuiltinTable table. This table contains all the overloads of
   // each function, and is a struct OpenCLBuiltinDecl.
   // E.g.:
-  // // acos
-  //   { 2, 0, "", 100 },
-  // This means that the signature of this acos overload is defined in OpenCL
-  // version 1.0 (100) and does not belong to any extension ("").  It has a
-  // 1 argument (+1 for the return type), stored at index 0 in the
-  // OpenCLSignature table.
+  // // convert_float2_rtn
+  //   { 58, 2 },
+  // This means that the signature of this convert_float2_rtn overload has
+  // 1 argument (+1 for the return type), stored at index 58 in
+  // the SignatureTable.
   void EmitBuiltinTable();
 
   // Emit a StringMatcher function to check whether a function name is an
   // OpenCL builtin function name.
   void EmitStringMatcher();
 
   // Emit a function returning the clang QualType instance associated with
   // the TableGen Record Type.
   void EmitQualTypeFinder();
 
   // Contains a list of the available signatures, without the name of the
   // function. Each pair consists of a signature and a cumulative index.
   // E.g.:  <<float, float>, 0>,
   //        <<float, int, int, 2>>,
   //        <<float>, 5>,
   //        ...
   //        <<double, double>, 35>.
-  std::vector<std::pair<std::vector<Record *>, unsigned>> SignatureSet;
+  std::vector<std::pair<std::vector<Record *>, unsigned>> SignaturesList;
 
   // Map the name of a builtin function to its prototypes (instances of the
   // TableGen "Builtin" class).
   // Each prototype is registered as a pair of:
   //   <pointer to the "Builtin" instance,
-  //    cumulative index of the associated signature in the SignatureSet>
+  //    cumulative index of the associated signature in the SignaturesList>
   // E.g.:  The function cos: (float cos(float), double cos(double), ...)
   //        <"cos", <<ptrToPrototype0, 5>,
-  //                <ptrToPrototype1, 35>>
-  //                <ptrToPrototype2, 79>>
+  //                 <ptrToPrototype1, 35>,
+  //                 <ptrToPrototype2, 79>>
   // ptrToPrototype1 has the following signature: <double, double>
   MapVector<StringRef, std::vector<std::pair<const Record *, unsigned>>>
-      OverloadInfo;
+      FctOverloadMap;
+
+  // Contains the map of OpenCL types to their index in the TypeTable.
+  MapVector<const Record *, unsigned> TypeMap;
+
+  // List of OpenCL type names in the same order as in enum OpenCLTypeID.
+  // This list does not contain generic types.
+  std::vector<const Record *> TypeList;
+
+  // Same as TypeList, but for generic types only.
+  std::vector<const Record *> GenTypeList;
 };
 } // namespace
 
 void BuiltinNameEmitter::Emit() {
   emitSourceFileHeader("OpenCL Builtin handling", OS);
 
   OS << "#include \"llvm/ADT/StringRef.h\"\n";
   OS << "using namespace clang;\n\n";
 
+  // Emit enums and structs.
   EmitDeclarations();
 
   GetOverloads();
 
+  // Emit tables.
+  EmitTypeTable();
   EmitSignatureTable();
-
   EmitBuiltinTable();
 
   EmitStringMatcher();
 
   EmitQualTypeFinder();
 }
 
-void BuiltinNameEmitter::EmitDeclarations() {
-  OS << "enum OpenCLTypeID {\n";
-  std::vector<Record *> Types = Records.getAllDerivedDefinitions("Type");
-  StringMap<bool> TypesSeen;
+void BuiltinNameEmitter::ExtractEnumTypes(std::vector<Record *> &Types,
+                                          StringMap<bool> &TypesSeen,
+                                          std::string &Output,
+                                          std::vector<const Record *> &List) {
+  raw_string_ostream SS(Output);
+
   for (const auto *T : Types) {
-    if (TypesSeen.find(T->getValueAsString("Name")) == TypesSeen.end())
-      OS << "  OCLT_" + T->getValueAsString("Name") << ",\n";
+    if (TypesSeen.find(T->getValueAsString("Name")) == TypesSeen.end()) {
+      SS << "  OCLT_" + T->getValueAsString("Name") << ",\n";
+      // Save the type names in the same order as their enum value. Note that
+      // the Record can be a VectorType or something else, only the name is
+      // important.
+      List.push_back(T);
       TypesSeen.insert(std::make_pair(T->getValueAsString("Name"), true));
     }
+  }
+  SS.flush();
+}
+
+void BuiltinNameEmitter::EmitDeclarations() {
+  // Enum of scalar type names (float, int, ...) and generic type sets.
+  OS << "enum OpenCLTypeID {\n";
+
+  StringMap<bool> TypesSeen;
+  std::string GenTypeEnums;
+  std::string TypeEnums;
+
+  // Generic types are defined at the end of the enum.
+  std::vector<Record *> GenTypes =
+      Records.getAllDerivedDefinitions("GenericType");
+  ExtractEnumTypes(GenTypes, TypesSeen, GenTypeEnums, GenTypeList);
+
+  std::vector<Record *> Types = Records.getAllDerivedDefinitions("Type");
+  ExtractEnumTypes(Types, TypesSeen, TypeEnums, TypeList);
+
+  OS << TypeEnums;
+  OS << GenTypeEnums;
   OS << "};\n";
 
+  // Structure definitions.
   OS << R"(
 
-// Type used in a prototype of an OpenCL builtin function.
-struct OpenCLType {
-  // A type (e.g.: float, int, ...)
-  OpenCLTypeID ID;
-  // Size of vector (if applicable)
-  unsigned VectorWidth;
-  // Address space of the pointer (if applicable)
-  LangAS AS;
-  // Whether the type is a pointer
-  bool isPointer;
+// Represents a return type or argument type.
+struct OpenCLTypeStruct {
+  // A type (e.g. float, int, ...)
+  const OpenCLTypeID ID;
+  // Vector size (if applicable; 0 for scalars and generic types).
+  const unsigned VectorWidth;
 };
 
 // One overload of an OpenCL builtin function.
-struct OpenCLBuiltinDecl {
-  // Number of arguments for the signature
-  unsigned NumArgs;
-  // Index in the OpenCLSignature table to get the required types
-  unsigned ArgTableIndex;
-  // Extension to which it belongs (e.g. cl_khr_subgroups)
-  const char *Extension;
-  // Version in which it was introduced (e.g. CL20)
-  unsigned Version;
+struct OpenCLBuiltinStruct {
+  // Index of the signature in the OpenCLTypeStruct table.
+  const unsigned SigTableIndex;
+  // Entries between index SigTableIndex and (SigTableIndex + NumTypes - 1) in
+  // the SignatureTable represent the complete signature.  The first type at
+  // index SigTableIndex is the return type.
+  const unsigned NumTypes;
 };
 
 )";
 }
 
 void BuiltinNameEmitter::GetOverloads() {
+  // Populate the TypeMap.
+  std::vector<Record *> Types = Records.getAllDerivedDefinitions("Type");
+  unsigned I = 0;
+  for (const auto &T : Types) {
+    TypeMap.insert(std::make_pair(T, I++));
+  }
+
+  // Populate the SignaturesList and the FctOverloadMap.
   unsigned CumulativeSignIndex = 0;
   std::vector<Record *> Builtins = Records.getAllDerivedDefinitions("Builtin");
   for (const auto *B : Builtins) {
     StringRef BName = B->getValueAsString("Name");
-    if (OverloadInfo.find(BName) == OverloadInfo.end()) {
-      OverloadInfo.insert(std::make_pair(
+    if (FctOverloadMap.find(BName) == FctOverloadMap.end()) {
+      FctOverloadMap.insert(std::make_pair(
           BName, std::vector<std::pair<const Record *, unsigned>>{}));
     }
 
     auto Signature = B->getValueAsListOfDefs("Signature");
     auto it =
-        std::find_if(SignatureSet.begin(), SignatureSet.end(),
+        std::find_if(SignaturesList.begin(), SignaturesList.end(),
                      [&](const std::pair<std::vector<Record *>, unsigned> &a) {
                        return a.first == Signature;
                      });
     unsigned SignIndex;
-    if (it == SignatureSet.end()) {
-      SignatureSet.push_back(std::make_pair(Signature, CumulativeSignIndex));
+    if (it == SignaturesList.end()) {
+      SignaturesList.push_back(std::make_pair(Signature, CumulativeSignIndex));
       SignIndex = CumulativeSignIndex;
       CumulativeSignIndex += Signature.size();
     } else {
       SignIndex = it->second;
     }
-    OverloadInfo[BName].push_back(std::make_pair(B, SignIndex));
-  }
+    FctOverloadMap[BName].push_back(std::make_pair(B, SignIndex));
+  }
+
+void BuiltinNameEmitter::EmitTypeTable() {
+  OS << "static const OpenCLTypeStruct TypeTable[] = {\n";
+  for (const auto &T : TypeMap) {
+    OS << "  // " << T.second << "\n";
+    OS << "  { OCLT_" << T.first->getValueAsString("Name") << ", "
+       << T.first->getValueAsInt("VecWidth") << "},\n";
+  }
+  OS << "};\n\n";
 }
 
 void BuiltinNameEmitter::EmitSignatureTable() {
-  OS << "static const OpenCLType OpenCLSignature[] = {\n";
-  for (auto &P : SignatureSet) {
+  // Store a type (e.g. int, float, int2, ...). The type is stored as an index
+  // of a struct OpenCLType table. Multiple entries following each other form a
+  // signature.
+  OS << "static const unsigned SignatureTable[] = {\n";
+  for (const auto &P : SignaturesList) {
     OS << "// " << P.second << "\n";
-    for (Record *R : P.first) {
-      OS << "{ OCLT_" << R->getValueAsString("Name") << ", "
-         << R->getValueAsInt("VecWidth") << ", "
-         << R->getValueAsString("AddrSpace") << ", "
-         << R->getValueAsBit("IsPointer") << "},";
-      OS << "\n";
+    for (const Record *R : P.first) {
+      OS << TypeMap.find(R)->second << ", ";
     }
+    OS << "\n";
   }
   OS << "};\n\n";
 }
 
 void BuiltinNameEmitter::EmitBuiltinTable() {
-  OS << "static const OpenCLBuiltinDecl OpenCLBuiltins[] = {\n";
-  for (auto &i : OverloadInfo) {
-    StringRef Name = i.first;
-    OS << "// " << Name << "\n";
-    for (auto &Overload : i.second) {
-      OS << "  { " << Overload.first->getValueAsListOfDefs("Signature").size()
-         << ", " << Overload.second << ", " << '"'
-         << Overload.first->getValueAsString("Extension") << "\", "
-         << Overload.first->getValueAsDef("Version")->getValueAsInt("Version")
+  unsigned Index = 0;
+
+  OS << "static const OpenCLBuiltinStruct BuiltinTable[] = {\n";
+  for (const auto &FOM : FctOverloadMap) {
+
+    OS << "// " << Index << " " << FOM.first << "\n";
+
+    for (const auto &Overload : FOM.second) {
+      OS << "  { "
+         << Overload.second << ", "
+         << Overload.first->getValueAsListOfDefs("Signature").size()
          << " },\n";
+         Index++;
     }
   }
   OS << "};\n\n";
 }
 
 void BuiltinNameEmitter::EmitStringMatcher() {
   std::vector<StringMatcher::StringPair> ValidBuiltins;
   unsigned CumulativeIndex = 1;
-  for (auto &i : OverloadInfo) {
+  for (auto &i : FctOverloadMap) {
     auto &Ov = i.second;
     std::string RetStmt;
     raw_string_ostream SS(RetStmt);
     SS << "return std::make_pair(" << CumulativeIndex << ", " << Ov.size()
        << ");";
     SS.flush();
     CumulativeIndex += Ov.size();
 
     ValidBuiltins.push_back(StringMatcher::StringPair(i.first, RetStmt));
   }
 
   OS << R"(
-// Return 0 if name is not a recognized OpenCL builtin, or an index
-// into a table of declarations if it is an OpenCL builtin.
-static std::pair<unsigned, unsigned> isOpenCLBuiltin(llvm::StringRef name) {
+// Find out whether a string matches an existing OpenCL builtin function name.
+// Returns: A pair <0, 0> if no name matches.
+//          A pair <Index, Len> indexing the BuiltinTable if the name is
+//          matching an OpenCL builtin function.
+static std::pair<unsigned, unsigned> isOpenCLBuiltin(llvm::StringRef Name) {
 
 )";
 
   StringMatcher("name", ValidBuiltins, OS).Emit(0, true);
 
   OS << "  return std::make_pair(0, 0);\n";
-  OS << "}\n";
+  OS << "} // isOpenCLBuiltin\n";
 }
 
 void BuiltinNameEmitter::EmitQualTypeFinder() {
   OS << R"(
 
-static QualType OCL2Qual(ASTContext &Context, OpenCLType Ty) {
-  QualType RT = Context.VoidTy;
-  switch (Ty.ID) {
+// Convert an OpenCLTypeStruct type to a list of QualTypes.
+// Generic types represent multiple types and vector sizes, thus a vector
+// is returned.
+static void OCL2Qual(ASTContext &Context, const OpenCLTypeStruct &Ty,
+                     std::vector<QualType> &QT) {
+  // Number of scalar types in the GenType.
+  unsigned GenTypeNumTypes;
+  // Pointer to the list of vector sizes for the GenType.
+  llvm::SmallVector<unsigned, 6> *GenVectorSizes;
 )";
 
+  // Generate list of vector sizes for each generic type.
+  for (const auto *VectList : Records.getAllDerivedDefinitions("IntList")) {
+    OS << "  llvm::SmallVector<unsigned, 6> List"
+       << VectList->getValueAsString("Name") << "{";
+    for (const auto V : VectList->getValueAsListOfInts("List")) {
+      OS << V << ", ";
+    }
+    OS << "};\n";
+  }
+
+  // Start of switch statement over all types.
+  OS << "\n  switch (Ty.ID) {\n";
+
+  // Switch cases for generic types.
+  for (const auto *GenType : Records.getAllDerivedDefinitions("GenericType")) {
+    OS << "    case OCLT_" << GenType->getValueAsString("Name") << ":\n";
+
+    // Build the Cartesian product of (vector sizes) x (types).  Only insert
+    // the plain scalar types for now; the ExtVector type will be added after
+    // the switch statement such that it is only added for the case matching
+    // the input to the OCL2Qual function.
+    for (unsigned I = 0; I < GenType->getValueAsDef("VectorList")
+                             ->getValueAsListOfInts("List")
+                             .size();
+         I++) {
+      for (const auto *T : GenType->getValueAsDef("TypeListField")
+                               ->getValueAsListOfDefs("List")) {
+        OS << "      QT.push_back(Context."
+           << T->getValueAsDef("QTName")->getValueAsString("Name")
+           << ");\n";
+      }
+    }
+    // GenTypeNumTypes is the number of types in the GenType
+    // (e.g. float/double/half).
+    OS << "      GenTypeNumTypes = "
+       << GenType->getValueAsDef("TypeListField")
+              ->getValueAsListOfDefs("List")
+              .size()
+       << ";\n";
+    // GenVectorSizes is the list of vector sizes for this GenType.
+    // QT contains GenTypeNumTypes * #GenVectorSizes elements.
+    OS << "      GenVectorSizes = &List"
+       << GenType->getValueAsDef("VectorList")->getValueAsString("Name")
+       << ";\n";
+    OS << "      break;\n";
+  }
+
+  // Switch cases for non generic, non image types (int, int4, float, ...).
   std::vector<Record *> Types = Records.getAllDerivedDefinitions("Type");
   StringMap<bool> TypesSeen;
 
   for (const auto *T : Types) {
     // Check we have not seen this Type
     if (TypesSeen.find(T->getValueAsString("Name")) != TypesSeen.end())
       continue;
     TypesSeen.insert(std::make_pair(T->getValueAsString("Name"), true));
 
     // Check the Type does not have an "abstract" QualType
     auto QT = T->getValueAsDef("QTName");
-    if (QT->getValueAsString("Name") == "null")
+    if (QT->getValueAsBit("IsAbstract") == 1)
       continue;
-
-    OS << "  case OCLT_" << T->getValueAsString("Name") << ":\n";
-    OS << "    RT = Context." << QT->getValueAsString("Name") << ";\n";
-    OS << "    break;\n";
+    // Emit the cases for non generic, non image types.
+    OS << "    case OCLT_" << T->getValueAsString("Name") << ":\n";
+    OS << "      QT.push_back(Context." << QT->getValueAsString("Name")
+       << ");\n";
+    OS << "      break;\n";
   }
-  OS << "  }\n";
 
-  // Special cases
-  OS << R"(
-  if (Ty.VectorWidth > 0)
-    RT = Context.getExtVectorType(RT, Ty.VectorWidth);
+  // End of switch statement.
+  OS << "  } // end of switch (Ty.ID)\n\n";
 
-  if (Ty.isPointer) {
-    RT = Context.getAddrSpaceQualType(RT, Ty.AS);
-    RT = Context.getPointerType(RT);
+  // Add ExtVector types if this was a generic type, as the switch statement
+  // above only populated the list with scalar types.  This completes the
+  // construction of the Cartesian product of (vector sizes) x (types).
+  OS << "  // Construct the different vector types for each generic type.\n";
+  OS << "  if (Ty.ID >= " << TypeList.size() << ") {";
+  OS << R"(
+    for (unsigned I = 0; I < QT.size(); I++) {
+      // For scalars, size is 1.
+      if ((*GenVectorSizes)[I / GenTypeNumTypes] != 1) {
+        QT[I] = Context.getExtVectorType(QT[I],
+                          (*GenVectorSizes)[I / GenTypeNumTypes]);
+      }
     }
+  }
+)";
 
-  return RT;
+  // Assign the right attributes to the types (e.g. vector size).
+  OS << R"(
+  // Set vector size for non-generic vector types.
+  if (Ty.VectorWidth > 1) {
+    for (unsigned Index = 0; Index < QT.size(); Index++) {
+      QT[Index] = Context.getExtVectorType(QT[Index], Ty.VectorWidth);
+    }
   }
 )";
+
+  // End of the "OCL2Qual" function.
+  OS << "\n} // OCL2Qual\n";
 }
 
 namespace clang {
 
 void EmitClangOpenCLBuiltins(RecordKeeper &Records, raw_ostream &OS) {
   BuiltinNameEmitter NameChecker(Records, OS);
   NameChecker.Emit();
 }
 
 } // end namespace clang