[OpenCL] Supports optional generic address space sematics in C++ for OpenCL 2021

[Topotuna]

Summary

Adds support for a feature macro __opencl_c_generic_address_space
in C++ for OpenCL 2021 enabling a respective optional core feature
from OpenCL 3.0. Testing is only performed in SemaOpenCL because
generic address space functionality is yet to be implemented in
C++ for OpenCL 2021.

This change aims to achieve compatibility between C++ for OpenCL
2021 and OpenCL 3.0.

[Anastasia]

clang/lib/Headers/opencl-c-base.h

33 ↗

(On Diff #367771)

I think we should just check here that C++ for OpenCL version is 1.0 so the same as for OpenCL C 2.0 because the same logic applies to it.

[Topotuna]

__opencl_c_generic_address_space should not have been chosen as the first feature to be addressed out of all OpenCL 3.0 optional core features. This is because __opencl_c_device_enqueue has a dependency on this feature. I will review the order in which OpenCL 3.0 optional features should be introduced to C++ for OpenCL 2021 and return to this revision when all dependencies are resolved.

[Anastasia]

In D108461#2960030, @Topotuna wrote:

__opencl_c_generic_address_space should not have been chosen as the first feature to be addressed out of all OpenCL 3.0 optional core features. This is because __opencl_c_device_enqueue has a dependency on this feature. I will review the order in which OpenCL 3.0 optional features should be introduced to C++ for OpenCL 2021 and return to this revision when all dependencies are resolved.

Ok, although it makes sense to add generic address space as multiple other features depend on this and we can unblock more work. Also I think __opencl_c_device_enqueue is being worked on by someone already but it is still not complete.

[Topotuna]

The previously suggested change was split into intermediate steps:
reviews.llvm.org/rGf9bc1b3bee55 - introduced a mapping function for C++ for OpenCL version compatibility with OpenCL and modified condition for setting OpenCLGenericAddressSpace language option.
reviews.llvm.org/rGfb321c2ea274 - enabled all OpenCL 3.0 optional core feature macros without implementing their support in C++ for OpenCL 2021.

This change only enables semantics for optional generic address space although actual functionality is not yet implemented in C++ for OpenCL 2021.

[Anastasia]

LGTM! Thanks

clang/lib/Basic/TargetInfo.cpp

413–415

Ok, I am guessing this will be changed back to one check later on?

[Topotuna]

clang/lib/Basic/TargetInfo.cpp

413–415

Yes, the change that removes this check is already put for review: D109306

clang/lib/Basic/TargetInfo.cpp

if (Opts.OpenCL) {
     // OpenCL C 2.0 or OpenCL C 3.0 with the __opencl_c_generic_address_space
     // feature
     // OpenCL C v3.0 s6.2.1 - OpenCL pipes require support of OpenCL C 2.0
     // or later and __opencl_c_pipes feature
     // FIXME: These language options are also defined in setLangDefaults()
     // for OpenCL C 2.0 but with no access to target capabilities. Target
     // should be immutable once created and thus these language options need
     // to be defined only once.
-    if (Opts.OpenCLVersion == 300) {
+    if (Opts.getOpenCLCompatibleVersion() == 300) {
       const auto &OpenCLFeaturesMap = getSupportedOpenCLOpts();
       Opts.OpenCLGenericAddressSpace = hasFeatureEnabled(
           OpenCLFeaturesMap, "__opencl_c_generic_address_space");
+      if (Opts.OpenCLVersion == 300)
         Opts.OpenCLPipes =
             hasFeatureEnabled(OpenCLFeaturesMap, "__opencl_c_pipes");
     }
   }
 
   if (Opts.DoubleSize) {
     if (Opts.DoubleSize == 32) {
       DoubleWidth = 32;
       LongDoubleWidth = 32;
       DoubleFormat = &llvm::APFloat::IEEEsingle();

clang/test/SemaOpenCL/address-spaces-conversions-cl2.0.cl

 // RUN: %clang_cc1 %s -ffake-address-space-map -verify -pedantic -fsyntax-only -DCONSTANT -cl-std=CL2.0
 // RUN: %clang_cc1 %s -ffake-address-space-map -verify -pedantic -fsyntax-only -DGLOBAL -cl-std=CL2.0
 // RUN: %clang_cc1 %s -ffake-address-space-map -verify -pedantic -fsyntax-only -DGENERIC -cl-std=CL2.0
-// RUN: %clang_cc1 %s -ffake-address-space-map -verify -pedantic -fsyntax-only -DCONSTANT -cl-std=clc++
-// RUN: %clang_cc1 %s -ffake-address-space-map -verify -pedantic -fsyntax-only -DGLOBAL -cl-std=clc++
-// RUN: %clang_cc1 %s -ffake-address-space-map -verify -pedantic -fsyntax-only -DGENERIC -cl-std=clc++
+// RUN: %clang_cc1 %s -ffake-address-space-map -verify -pedantic -fsyntax-only -DCONSTANT -cl-std=clc++1.0
+// RUN: %clang_cc1 %s -ffake-address-space-map -verify -pedantic -fsyntax-only -DGLOBAL -cl-std=clc++1.0
+// RUN: %clang_cc1 %s -ffake-address-space-map -verify -pedantic -fsyntax-only -DGENERIC -cl-std=clc++1.0
 // RUN: %clang_cc1 %s -ffake-address-space-map -verify -pedantic -fsyntax-only -DCONSTANT -cl-std=CL3.0 -cl-ext=+__opencl_c_generic_address_space
 // RUN: %clang_cc1 %s -ffake-address-space-map -verify -pedantic -fsyntax-only -DGLOBAL -cl-std=CL3.0 -cl-ext=+__opencl_c_generic_address_space
 // RUN: %clang_cc1 %s -ffake-address-space-map -verify -pedantic -fsyntax-only -DGENERIC -cl-std=CL3.0 -cl-ext=+__opencl_c_generic_address_space
+// RUN: %clang_cc1 %s -ffake-address-space-map -verify -pedantic -fsyntax-only -DCONSTANT -cl-std=clc++2021 -cl-ext=+__opencl_c_generic_address_space
+// RUN: %clang_cc1 %s -ffake-address-space-map -verify -pedantic -fsyntax-only -DGLOBAL -cl-std=clc++2021 -cl-ext=+__opencl_c_generic_address_space
+// RUN: %clang_cc1 %s -ffake-address-space-map -verify -pedantic -fsyntax-only -DGENERIC -cl-std=clc++2021 -cl-ext=+__opencl_c_generic_address_space
 
 /* OpenCLC v2.0 adds a set of restrictions for conversions between pointers to
 *  different address spaces, mainly described in Sections 6.5.5 and 6.5.6.
 *
 *  It adds notion of overlapping address spaces. The main differention is that
 *  an unnamed address space is added, called '__generic'. Pointers to the
 *  generic address space can be interchangabley used with pointers to any
 *  other address space except for __constant address space (Section 6.5.5).

clang/test/SemaOpenCL/address-spaces.cl

 // RUN: %clang_cc1 %s -verify -pedantic -fsyntax-only
 // RUN: %clang_cc1 %s -cl-std=CL2.0 -verify -pedantic -fsyntax-only
 // RUN: %clang_cc1 %s -cl-std=CL3.0 -cl-ext=+__opencl_c_generic_address_space -verify -pedantic -fsyntax-only
-// RUN: %clang_cc1 %s -cl-std=clc++ -verify -pedantic -fsyntax-only
+// RUN: %clang_cc1 %s -cl-std=clc++1.0 -verify -pedantic -fsyntax-only
+// RUN: %clang_cc1 %s -cl-std=clc++2021 -cl-ext=+__opencl_c_generic_address_space -verify -pedantic -fsyntax-only
 
 __constant int ci = 1;
 
 // __constant ints are allowed in constant expressions.
 enum use_ci_in_enum { enumerator = ci };
 typedef int use_ci_in_array_bound[ci];
 
 // general constant folding of array bounds is not permitted