[OpenCL] Stop opencl-c-base.h leaking extension enabling

[svenvh]

Summary

opencl-c.h disables all extensions at its end, but opencl-c-base.h
does not, and that causes any inclusion of only opencl-c-base.h to
leave some extensions (such as cl_khr_fp16) enabled. This affects the
-fdeclare-opencl-builtins option for example.

This violates the OpenCL Extension Specification which specifies that
"The initial state of the compiler is as if the directive #pragma
OPENCL EXTENSION all : disable was issued".

Fix by disabling all extensions at the end of opencl-c-base.h and
enable extensions inside opencl.h which relied on opencl-c-base.h
enabling the cl_khr_fp16/64 extensions.

[Anastasia]

LGTM! Thanks!

[mantognini]

LGMT as it reduces the divergence in compilation flows. I let @Anastasia, or someone else more familiar with the codebase, give the final approval though.

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

 #define CLK_AVC_IME_RESULT_SINGLE_REFERENCE_STREAMOUT_INITIALIZE_INTEL 0x0
 #define CLK_AVC_IME_RESULT_SINGLE_REFERENCE_STREAMIN_INITIALIZE_INTEL 0x0
 #define CLK_AVC_IME_RESULT_DUAL_REFERENCE_STREAMOUT_INITIALIZE_INTEL 0x0
 #define CLK_AVC_IME_RESULT_DUAL_REFERENCE_STREAMIN_INITIALIZE_INTEL 0x0
 
 #pragma OPENCL EXTENSION cl_intel_device_side_avc_motion_estimation : end
 #endif // cl_intel_device_side_avc_motion_estimation
 
+// Disable any extensions we may have enabled previously.
+#pragma OPENCL EXTENSION all : disable
+
 #endif //_OPENCL_BASE_H_

clang/lib/Headers/opencl-c.h

 float16 __ovld __cnfn convert_float16_rtz(float16);
 float16 __ovld __cnfn convert_float16_rtp(float16);
 float16 __ovld __cnfn convert_float16_rtn(float16);
 float16 __ovld __cnfn convert_float16(float16);
 
 // Conversions with double data type parameters or return value.
 
 #ifdef cl_khr_fp64
+#pragma OPENCL EXTENSION cl_khr_fp64 : enable
 char __ovld __cnfn convert_char(double);
 char __ovld __cnfn convert_char_rte(double);
 char __ovld __cnfn convert_char_rtn(double);
 char __ovld __cnfn convert_char_rtp(double);
 char __ovld __cnfn convert_char_rtz(double);
 char __ovld __cnfn convert_char_sat(double);
 char __ovld __cnfn convert_char_sat_rte(double);
 char __ovld __cnfn convert_char_sat_rtn(double);

 double16 __ovld __cnfn convert_double16_rtz(short16);
 double16 __ovld __cnfn convert_double16_rtz(uchar16);
 double16 __ovld __cnfn convert_double16_rtz(uint16);
 double16 __ovld __cnfn convert_double16_rtz(ulong16);
 double16 __ovld __cnfn convert_double16_rtz(ushort16);
 #endif //cl_khr_fp64
 
 #ifdef cl_khr_fp16
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
 // Convert half types to non-double types.
 uchar __ovld __cnfn convert_uchar(half);
 uchar __ovld __cnfn convert_uchar_rte(half);
 uchar __ovld __cnfn convert_uchar_rtp(half);
 uchar __ovld __cnfn convert_uchar_rtn(half);
 uchar __ovld __cnfn convert_uchar_rtz(half);
 uchar __ovld __cnfn convert_uchar_sat(half);
 uchar __ovld __cnfn convert_uchar_sat_rte(half);

  *
  * The address computed as (p + (offset * n)) must be
  * 8-bit aligned if gentype is char, uchar;
  * 16-bit aligned if gentype is short, ushort, half;
  * 32-bit aligned if gentype is int, uint, float;
  * 64-bit aligned if gentype is long, ulong, double.
  */
 
 char2 __ovld vload2(size_t offset, const __constant char *p);
 uchar2 __ovld vload2(size_t offset, const __constant uchar *p);
 short2 __ovld vload2(size_t offset, const __constant short *p);
 ushort2 __ovld vload2(size_t offset, const __constant ushort *p);
 int2 __ovld vload2(size_t offset, const __constant int *p);
 uint2 __ovld vload2(size_t offset, const __constant uint *p);
 long2 __ovld vload2(size_t offset, const __constant long *p);
 ulong2 __ovld vload2(size_t offset, const __constant ulong *p);
 float2 __ovld vload2(size_t offset, const __constant float *p);
 char3 __ovld vload3(size_t offset, const __constant char *p);
 uchar3 __ovld vload3(size_t offset, const __constant uchar *p);
 short3 __ovld vload3(size_t offset, const __constant short *p);
 ushort3 __ovld vload3(size_t offset, const __constant ushort *p);
 int3 __ovld vload3(size_t offset, const __constant int *p);
 uint3 __ovld vload3(size_t offset, const __constant uint *p);
 long3 __ovld vload3(size_t offset, const __constant long *p);
 ulong3 __ovld vload3(size_t offset, const __constant ulong *p);
 float3 __ovld vload3(size_t offset, const __constant float *p);
 char4 __ovld vload4(size_t offset, const __constant char *p);
 uchar4 __ovld vload4(size_t offset, const __constant uchar *p);
 short4 __ovld vload4(size_t offset, const __constant short *p);

clang/test/SemaOpenCL/half.cl

 // RUN: %clang_cc1 %s -verify -pedantic -fsyntax-only -Wno-unused-value -triple spir-unknown-unknown
+// RUN: %clang_cc1 %s -verify -pedantic -fsyntax-only -Wno-unused-value -triple spir-unknown-unknown -fdeclare-opencl-builtins -finclude-default-header
 
-#pragma OPENCL EXTENSION cl_khr_fp16 : disable
 constant float f = 1.0h; // expected-error{{half precision constant requires cl_khr_fp16}}
 
 half half_disabled(half *p, // expected-error{{declaring function return value of type 'half' is not allowed}}
                    half h)  // expected-error{{declaring function parameter of type '__private half' is not allowed}}
 {
   half a[2]; // expected-error{{declaring variable of type '__private half [2]' is not allowed}}
   half b;    // expected-error{{declaring variable of type '__private half' is not allowed}}
   *p; // expected-error{{loading directly from pointer to type '__private half' requires cl_khr_fp16. Use vector data load builtin functions instead}}