[OpenCL] Add cl_khr_subgroup_ballot to TableGen BIFs

[svenvh]

Summary

Add the builtin functions brought by the cl_khr_subgroup_ballot
extension to -fdeclare-opencl-builtins.

Also add placeholder comments for the other Extended Subgroup
Functions from the OpenCL Extension Specification.

I have verified (semi-manually) that this adds the same ballot functions
that are also in opencl-c.h already.

[svenvh]

Add test after rebasing on dependent commit (D95616).

[Anastasia]

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

35 ↗

(On Diff #319860)

Not related to this patch - I think we should have included opencl-c-base.h by default when passing -fdeclare-opencl-builtins or do you see any issues with that?

141 ↗

(On Diff #319860)

I appreciate we haven't addressed the standard header testing holistically yet and this functionality was added in opencl-c.h without testing the functions, but would it be better to test each function at least here?

Even though the final goal should be testing all overloads, this is outside of the scope of this work that is closing the gap between opencl-c.h and this experimental function declaration mechanism.

[svenvh]

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

35 ↗

(On Diff #319860)

That's something we probably want to do in the future indeed.

141 ↗

(On Diff #319860)

but would it be better to test each function at least here?

I think that defeats the purpose of this test. This test is meant to test the basic functionality of -fdeclare-opencl-builtins. It is not a completeness test.

A completeness test shouldn't be tied to -fdeclare-opencl-builtins alone; such a test should also be run against opencl-c.h, so this would be the wrong test to turn into a completeness test in my opinion.

[Anastasia]

LGTM! Thanks!

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

141 ↗

(On Diff #319860)

Ok, this test does test functionality selectively but I wouldn't be able to tell how one selects what to test. Perhaps a comment would help for the future development...

[svenvh]

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

141 ↗

(On Diff #319860)

There are no formal selection criteria here I think. The test aims to cover the different functional aspects of -fdeclare-opencl-builtins. The new functional aspect for -fdeclare-opencl-builtins in this patch is an extension that is not in OpenCLExtensions.def, so it makes sense to add such a case to the test.

I will add a comment to the top of the test file.

[Anastasia]

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

141 ↗

(On Diff #319860)

Makes sense! Thanks!

clang/lib/Sema/OpenCLBuiltins.td

 }
 
 // Extension associated to a builtin function.
 class FunctionExtension<string _Ext> : AbstractExtension<_Ext>;
 
 // FunctionExtension definitions.
 def FuncExtNone                          : FunctionExtension<"">;
 def FuncExtKhrSubgroups                  : FunctionExtension<"cl_khr_subgroups">;
+def FuncExtKhrSubgroupBallot             : FunctionExtension<"cl_khr_subgroup_ballot">;
 def FuncExtKhrGlobalInt32BaseAtomics     : FunctionExtension<"cl_khr_global_int32_base_atomics">;
 def FuncExtKhrGlobalInt32ExtendedAtomics : FunctionExtension<"cl_khr_global_int32_extended_atomics">;
 def FuncExtKhrLocalInt32BaseAtomics      : FunctionExtension<"cl_khr_local_int32_base_atomics">;
 def FuncExtKhrLocalInt32ExtendedAtomics  : FunctionExtension<"cl_khr_local_int32_extended_atomics">;
 def FuncExtKhrInt64BaseAtomics           : FunctionExtension<"cl_khr_int64_base_atomics">;
 def FuncExtKhrInt64ExtendedAtomics       : FunctionExtension<"cl_khr_int64_extended_atomics">;
 def FuncExtKhrMipmapImage                : FunctionExtension<"cl_khr_mipmap_image">;
 def FuncExtKhrMipmapImageWrites          : FunctionExtension<"cl_khr_mipmap_image_writes">;

 // uchar abs(uchar).
 def TLAllUIntsTwice : TypeList<[UChar, UChar, UShort, UShort, UInt, UInt, ULong, ULong]>;
 
 def TLAllInts       : TypeList<[Char, UChar, Short, UShort, Int, UInt, Long, ULong]>;
 
 // GenType definitions for multiple base types (e.g. all floating point types,
 // or all integer types).
 // All types
+def AGenType1              : GenericType<"AGenType1", TLAll, Vec1>;
 def AGenTypeN              : GenericType<"AGenTypeN", TLAll, VecAndScalar>;
 def AGenTypeNNoScalar      : GenericType<"AGenTypeNNoScalar", TLAll, VecNoScalar>;
 // All integer
 def AIGenType1             : GenericType<"AIGenType1", TLAllInts, Vec1>;
 def AIGenTypeN             : GenericType<"AIGenTypeN", TLAllInts, VecAndScalar>;
 def AIGenTypeNNoScalar     : GenericType<"AIGenTypeNNoScalar", TLAllInts, VecNoScalar>;
 // All integer to unsigned
 def AI2UGenTypeN           : GenericType<"AI2UGenTypeN", TLAllUIntsTwice, VecAndScalar>;
let Extension = FuncExtKhrSubgroups in {
   foreach name = ["sub_group_reduce_", "sub_group_scan_exclusive_",
                   "sub_group_scan_inclusive_"] in {
     foreach op = ["add", "min", "max"] in {
       def : Builtin<name # op, [IntLongFloatGenType1, IntLongFloatGenType1], Attr.Convergent>;
     }
   }
 }
 
+// OpenCL Extension v3.0 s38 - Extended Subgroup Functions
+
+// Section 38.4.1 - cl_khr_subgroup_extended_types
+// TODO
+
+// Section 38.5.1 - cl_khr_subgroup_non_uniform_vote
+// TODO
+
+// Section 38.6.1 - cl_khr_subgroup_ballot
+let Extension = FuncExtKhrSubgroupBallot in {
+  def : Builtin<"sub_group_non_uniform_broadcast", [AGenTypeN, AGenTypeN, UInt]>;
+  def : Builtin<"sub_group_broadcast_first", [AGenType1, AGenType1]>;
+  def : Builtin<"sub_group_ballot", [VectorType<UInt, 4>, Int]>;
+  def : Builtin<"sub_group_inverse_ballot", [Int, VectorType<UInt, 4>], Attr.Const>;
+  def : Builtin<"sub_group_ballot_bit_extract", [Int, VectorType<UInt, 4>, UInt], Attr.Const>;
+  def : Builtin<"sub_group_ballot_bit_count", [UInt, VectorType<UInt, 4>], Attr.Const>;
+  def : Builtin<"sub_group_ballot_inclusive_scan", [UInt, VectorType<UInt, 4>]>;
+  def : Builtin<"sub_group_ballot_exclusive_scan", [UInt, VectorType<UInt, 4>]>;
+  def : Builtin<"sub_group_ballot_find_lsb", [UInt, VectorType<UInt, 4>]>;
+  def : Builtin<"sub_group_ballot_find_msb", [UInt, VectorType<UInt, 4>]>;
+
+  foreach op = ["eq", "ge", "gt", "le", "lt"] in {
+    def : Builtin<"get_sub_group_" # op # "_mask", [VectorType<UInt, 4>], Attr.Const>;
+  }
+}
+
+// Section 38.7.1 - cl_khr_subgroup_non_uniform_arithmetic
+// TODO
+
+// Section 38.8.1 - cl_khr_subgroup_shuffle
+// TODO
+
+// Section 38.9.1 - cl_khr_subgroup_shuffle_relative
+// TODO
+
+// Section 38.10.1 - cl_khr_subgroup_clustered_reduce
+// TODO
+
 //--------------------------------------------------------------------
 // Arm extensions.
 let Extension = ArmIntegerDotProductInt8 in {
   foreach name = ["arm_dot"] in {
     def : Builtin<name, [UInt, VectorType<UChar, 4>, VectorType<UChar, 4>]>;
     def : Builtin<name, [Int, VectorType<Char, 4>, VectorType<Char, 4>]>;
   }
 }