[OpenCL] Allow undefining header-only features

[svenvh]

Summary

opencl-c-base.h always defines 5 particular feature macros for
SPIR-V, making it impossible to disable those features.

To allow disabling any of those features, let the header recognize
__undef_<feature> macros. The user can then pass the
-D__undef_<feature> flag on the command line to disable a specific
feature. The __undef macro could potentially also be set from
-cl-ext=-feature, but for now only change the header and only
provide __undef macros for the 5 features that are always enabled in
opencl-c-base.h.

This is an alternative to https://reviews.llvm.org/D137652

[Anastasia]

LGTM, thanks!

Btw I wonder if in the future we could add some error or warning in case someone uses the same approach for frontend specific features, i.e.

#ifdef __undef___opencl_c_generic_address_space
#error "Feature __opencl_c_generic_address_space can only be disabled via -cl-ext flag"
#endif

[svenvh]

In D141297#4043122, @Anastasia wrote:

Btw I wonder if in the future we could add some error or warning in case someone uses the same approach for frontend specific features, i.e.

#ifdef __undef___opencl_c_generic_address_space
#error "Feature __opencl_c_generic_address_space can only be disabled via -cl-ext flag"
#endif

Interesting idea, but I'm a bit hesitant of doing so:
It increases the size of opencl-c-base.h so it will take longer to parse, which will affect every OpenCL compilation. Luckily we can avoid that cost if we keep the __undef_ mechanism an internal solution, which it will be once we let -cl-ext=-feature generate __undef_ macros for extensions that are not in OpenCLExtensions.def. So longer term users will never have to pass __undef_ macros.

[Anastasia]

In D141297#4047148, @svenvh wrote:

In D141297#4043122, @Anastasia wrote:

Btw I wonder if in the future we could add some error or warning in case someone uses the same approach for frontend specific features, i.e.

#ifdef __undef___opencl_c_generic_address_space
#error "Feature __opencl_c_generic_address_space can only be disabled via -cl-ext flag"
#endif

Interesting idea, but I'm a bit hesitant of doing so:
It increases the size of opencl-c-base.h so it will take longer to parse, which will affect every OpenCL compilation. Luckily we can avoid that cost if we keep the __undef_ mechanism an internal solution, which it will be once we let -cl-ext=-feature generate __undef_ macros for extensions that are not in OpenCLExtensions.def. So longer term users will never have to pass __undef_ macros.

Ok that would be better indeed. Btw I hope we won't have that many feature macros that it would be too long to parse those. :(

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

 // For the SPIR and SPIR-V target all features are supported.
 #if defined(__SPIR__) || defined(__SPIRV__)
 #define __opencl_c_work_group_collective_functions 1
 #define __opencl_c_atomic_order_seq_cst 1
 #define __opencl_c_atomic_scope_device 1
 #define __opencl_c_atomic_scope_all_devices 1
 #define __opencl_c_read_write_images 1
 #endif // defined(__SPIR__)
+
+// Undefine any feature macros that have been explicitly disabled using
+// an __undef_<feature> macro.
+#ifdef __undef___opencl_c_work_group_collective_functions
+#undef __opencl_c_work_group_collective_functions
+#endif
+#ifdef __undef___opencl_c_atomic_order_seq_cst
+#undef __opencl_c_atomic_order_seq_cst
+#endif
+#ifdef __undef___opencl_c_atomic_scope_device
+#undef __opencl_c_atomic_scope_device
+#endif
+#ifdef __undef___opencl_c_atomic_scope_all_devices
+#undef __opencl_c_atomic_scope_all_devices
+#endif
+#ifdef __undef___opencl_c_read_write_images
+#undef __opencl_c_read_write_images
+#endif
+
 #endif // (__OPENCL_CPP_VERSION__ == 202100 || __OPENCL_C_VERSION__ == 300)
 
 #if !defined(__opencl_c_generic_address_space)
 // Internal feature macro to provide named (global, local, private) address
 // space overloads for builtin functions that take a pointer argument.
 #define __opencl_c_named_address_space_builtins 1
 #endif // !defined(__opencl_c_generic_address_space)
 

clang/test/SemaOpenCL/features.cl

 // RUN:   | FileCheck -match-full-lines %s  --check-prefix=NO-FEATURES
 // RUN: %clang_cc1 -triple spir-unknown-unknown %s -E -dM -o - -x cl -cl-std=CL1.2 \
 // RUN:   | FileCheck -match-full-lines %s  --check-prefix=NO-FEATURES
 // RUN: %clang_cc1 -triple spir-unknown-unknown %s -E -dM -o - -x cl -cl-std=CL2.0 \
 // RUN:   | FileCheck -match-full-lines %s  --check-prefix=NO-FEATURES
 // RUN: %clang_cc1 -triple spir-unknown-unknown %s -E -dM -o - -x cl -cl-std=clc++1.0 \
 // RUN:   | FileCheck -match-full-lines %s  --check-prefix=NO-FEATURES
 
+// For OpenCL C 3.0, header-only features can be disabled using macros.
+// RUN: %clang_cc1 -triple spir-unknown-unknown %s -E -dM -o - -x cl -cl-std=CL3.0 -fdeclare-opencl-builtins -finclude-default-header \
+// RUN:    -D__undef___opencl_c_work_group_collective_functions=1 \
+// RUN:    -D__undef___opencl_c_atomic_order_seq_cst=1 \
+// RUN:    -D__undef___opencl_c_atomic_scope_device=1 \
+// RUN:    -D__undef___opencl_c_atomic_scope_all_devices=1 \
+// RUN:    -D__undef___opencl_c_read_write_images=1 \
+// RUN:   | FileCheck %s --check-prefix=NO-HEADERONLY-FEATURES
+
 // Note that __opencl_c_int64 is always defined assuming
 // always compiling for FULL OpenCL profile
 
 // FEATURES: #define __opencl_c_3d_image_writes 1
 // FEATURES: #define __opencl_c_atomic_order_acq_rel 1
 // FEATURES: #define __opencl_c_atomic_order_seq_cst 1
 // FEATURES: #define __opencl_c_device_enqueue 1
 // FEATURES: #define __opencl_c_fp64 1
 // FEATURES: #define __opencl_c_generic_address_space 1
 // FEATURES: #define __opencl_c_images 1
 // FEATURES: #define __opencl_c_int64 1
 // FEATURES: #define __opencl_c_pipes 1
 // FEATURES: #define __opencl_c_program_scope_global_variables 1
 // FEATURES: #define __opencl_c_read_write_images 1
 // FEATURES: #define __opencl_c_subgroups 1
 
 // NO-FEATURES: #define __opencl_c_int64 1
-// NO-FEATURES-NOT: __opencl_c_3d_image_writes
-// NO-FEATURES-NOT: __opencl_c_atomic_order_acq_rel
-// NO-FEATURES-NOT: __opencl_c_atomic_order_seq_cst
-// NO-FEATURES-NOT: __opencl_c_device_enqueue
-// NO-FEATURES-NOT: __opencl_c_fp64
-// NO-FEATURES-NOT: __opencl_c_generic_address_space
-// NO-FEATURES-NOT: __opencl_c_images
-// NO-FEATURES-NOT: __opencl_c_pipes
-// NO-FEATURES-NOT: __opencl_c_program_scope_global_variables
-// NO-FEATURES-NOT: __opencl_c_read_write_images
-// NO-FEATURES-NOT: __opencl_c_subgroups
+// NO-FEATURES-NOT: #define __opencl_c_3d_image_writes
+// NO-FEATURES-NOT: #define __opencl_c_atomic_order_acq_rel
+// NO-FEATURES-NOT: #define __opencl_c_atomic_order_seq_cst
+// NO-FEATURES-NOT: #define __opencl_c_device_enqueue
+// NO-FEATURES-NOT: #define __opencl_c_fp64
+// NO-FEATURES-NOT: #define __opencl_c_generic_address_space
+// NO-FEATURES-NOT: #define __opencl_c_images
+// NO-FEATURES-NOT: #define __opencl_c_pipes
+// NO-FEATURES-NOT: #define __opencl_c_program_scope_global_variables
+// NO-FEATURES-NOT: #define __opencl_c_read_write_images
+// NO-FEATURES-NOT: #define __opencl_c_subgroups
+
+// NO-HEADERONLY-FEATURES-NOT: #define __opencl_c_work_group_collective_functions
+// NO-HEADERONLY-FEATURES-NOT: #define __opencl_c_atomic_order_seq_cst
+// NO-HEADERONLY-FEATURES-NOT: #define __opencl_c_atomic_scope_device
+// NO-HEADERONLY-FEATURES-NOT: #define __opencl_c_atomic_scope_all_devices
+// NO-HEADERONLY-FEATURES-NOT: #define __opencl_c_read_write_images