diff --git a/clang/lib/Headers/opencl-c.h b/clang/lib/Headers/opencl-c.h
--- a/clang/lib/Headers/opencl-c.h
+++ b/clang/lib/Headers/opencl-c.h
@@ -11190,305 +11190,305 @@
  * 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);
-ushort4 __ovld vload4(size_t offset, const __constant ushort *p);
-int4 __ovld vload4(size_t offset, const __constant int *p);
-uint4 __ovld vload4(size_t offset, const __constant uint *p);
-long4 __ovld vload4(size_t offset, const __constant long *p);
-ulong4 __ovld vload4(size_t offset, const __constant ulong *p);
-float4 __ovld vload4(size_t offset, const __constant float *p);
-char8 __ovld vload8(size_t offset, const __constant char *p);
-uchar8 __ovld vload8(size_t offset, const __constant uchar *p);
-short8 __ovld vload8(size_t offset, const __constant short *p);
-ushort8 __ovld vload8(size_t offset, const __constant ushort *p);
-int8 __ovld vload8(size_t offset, const __constant int *p);
-uint8 __ovld vload8(size_t offset, const __constant uint *p);
-long8 __ovld vload8(size_t offset, const __constant long *p);
-ulong8 __ovld vload8(size_t offset, const __constant ulong *p);
-float8 __ovld vload8(size_t offset, const __constant float *p);
-char16 __ovld vload16(size_t offset, const __constant char *p);
-uchar16 __ovld vload16(size_t offset, const __constant uchar *p);
-short16 __ovld vload16(size_t offset, const __constant short *p);
-ushort16 __ovld vload16(size_t offset, const __constant ushort *p);
-int16 __ovld vload16(size_t offset, const __constant int *p);
-uint16 __ovld vload16(size_t offset, const __constant uint *p);
-long16 __ovld vload16(size_t offset, const __constant long *p);
-ulong16 __ovld vload16(size_t offset, const __constant ulong *p);
-float16 __ovld vload16(size_t offset, const __constant float *p);
+char2 __ovld __purefn vload2(size_t offset, const __constant char *p);
+uchar2 __ovld __purefn vload2(size_t offset, const __constant uchar *p);
+short2 __ovld __purefn vload2(size_t offset, const __constant short *p);
+ushort2 __ovld __purefn vload2(size_t offset, const __constant ushort *p);
+int2 __ovld __purefn vload2(size_t offset, const __constant int *p);
+uint2 __ovld __purefn vload2(size_t offset, const __constant uint *p);
+long2 __ovld __purefn vload2(size_t offset, const __constant long *p);
+ulong2 __ovld __purefn vload2(size_t offset, const __constant ulong *p);
+float2 __ovld __purefn vload2(size_t offset, const __constant float *p);
+char3 __ovld __purefn vload3(size_t offset, const __constant char *p);
+uchar3 __ovld __purefn vload3(size_t offset, const __constant uchar *p);
+short3 __ovld __purefn vload3(size_t offset, const __constant short *p);
+ushort3 __ovld __purefn vload3(size_t offset, const __constant ushort *p);
+int3 __ovld __purefn vload3(size_t offset, const __constant int *p);
+uint3 __ovld __purefn vload3(size_t offset, const __constant uint *p);
+long3 __ovld __purefn vload3(size_t offset, const __constant long *p);
+ulong3 __ovld __purefn vload3(size_t offset, const __constant ulong *p);
+float3 __ovld __purefn vload3(size_t offset, const __constant float *p);
+char4 __ovld __purefn vload4(size_t offset, const __constant char *p);
+uchar4 __ovld __purefn vload4(size_t offset, const __constant uchar *p);
+short4 __ovld __purefn vload4(size_t offset, const __constant short *p);
+ushort4 __ovld __purefn vload4(size_t offset, const __constant ushort *p);
+int4 __ovld __purefn vload4(size_t offset, const __constant int *p);
+uint4 __ovld __purefn vload4(size_t offset, const __constant uint *p);
+long4 __ovld __purefn vload4(size_t offset, const __constant long *p);
+ulong4 __ovld __purefn vload4(size_t offset, const __constant ulong *p);
+float4 __ovld __purefn vload4(size_t offset, const __constant float *p);
+char8 __ovld __purefn vload8(size_t offset, const __constant char *p);
+uchar8 __ovld __purefn vload8(size_t offset, const __constant uchar *p);
+short8 __ovld __purefn vload8(size_t offset, const __constant short *p);
+ushort8 __ovld __purefn vload8(size_t offset, const __constant ushort *p);
+int8 __ovld __purefn vload8(size_t offset, const __constant int *p);
+uint8 __ovld __purefn vload8(size_t offset, const __constant uint *p);
+long8 __ovld __purefn vload8(size_t offset, const __constant long *p);
+ulong8 __ovld __purefn vload8(size_t offset, const __constant ulong *p);
+float8 __ovld __purefn vload8(size_t offset, const __constant float *p);
+char16 __ovld __purefn vload16(size_t offset, const __constant char *p);
+uchar16 __ovld __purefn vload16(size_t offset, const __constant uchar *p);
+short16 __ovld __purefn vload16(size_t offset, const __constant short *p);
+ushort16 __ovld __purefn vload16(size_t offset, const __constant ushort *p);
+int16 __ovld __purefn vload16(size_t offset, const __constant int *p);
+uint16 __ovld __purefn vload16(size_t offset, const __constant uint *p);
+long16 __ovld __purefn vload16(size_t offset, const __constant long *p);
+ulong16 __ovld __purefn vload16(size_t offset, const __constant ulong *p);
+float16 __ovld __purefn vload16(size_t offset, const __constant float *p);
 #ifdef cl_khr_fp64
-double2 __ovld vload2(size_t offset, const __constant double *p);
-double3 __ovld vload3(size_t offset, const __constant double *p);
-double4 __ovld vload4(size_t offset, const __constant double *p);
-double8 __ovld vload8(size_t offset, const __constant double *p);
-double16 __ovld vload16(size_t offset, const __constant double *p);
+double2 __ovld __purefn vload2(size_t offset, const __constant double *p);
+double3 __ovld __purefn vload3(size_t offset, const __constant double *p);
+double4 __ovld __purefn vload4(size_t offset, const __constant double *p);
+double8 __ovld __purefn vload8(size_t offset, const __constant double *p);
+double16 __ovld __purefn vload16(size_t offset, const __constant double *p);
 #endif //cl_khr_fp64
 
 #ifdef cl_khr_fp16
-half __ovld vload(size_t offset, const __constant half *p);
-half2 __ovld vload2(size_t offset, const __constant half *p);
-half3 __ovld vload3(size_t offset, const __constant half *p);
-half4 __ovld vload4(size_t offset, const __constant half *p);
-half8 __ovld vload8(size_t offset, const __constant half *p);
-half16 __ovld vload16(size_t offset, const __constant half *p);
+half __ovld __purefn vload(size_t offset, const __constant half *p);
+half2 __ovld __purefn vload2(size_t offset, const __constant half *p);
+half3 __ovld __purefn vload3(size_t offset, const __constant half *p);
+half4 __ovld __purefn vload4(size_t offset, const __constant half *p);
+half8 __ovld __purefn vload8(size_t offset, const __constant half *p);
+half16 __ovld __purefn vload16(size_t offset, const __constant half *p);
 #endif //cl_khr_fp16
 
 #if defined(__opencl_c_generic_address_space)
-char2 __ovld vload2(size_t offset, const char *p);
-uchar2 __ovld vload2(size_t offset, const uchar *p);
-short2 __ovld vload2(size_t offset, const short *p);
-ushort2 __ovld vload2(size_t offset, const ushort *p);
-int2 __ovld vload2(size_t offset, const int *p);
-uint2 __ovld vload2(size_t offset, const uint *p);
-long2 __ovld vload2(size_t offset, const long *p);
-ulong2 __ovld vload2(size_t offset, const ulong *p);
-float2 __ovld vload2(size_t offset, const float *p);
-char3 __ovld vload3(size_t offset, const char *p);
-uchar3 __ovld vload3(size_t offset, const uchar *p);
-short3 __ovld vload3(size_t offset, const short *p);
-ushort3 __ovld vload3(size_t offset, const ushort *p);
-int3 __ovld vload3(size_t offset, const int *p);
-uint3 __ovld vload3(size_t offset, const uint *p);
-long3 __ovld vload3(size_t offset, const long *p);
-ulong3 __ovld vload3(size_t offset, const ulong *p);
-float3 __ovld vload3(size_t offset, const float *p);
-char4 __ovld vload4(size_t offset, const char *p);
-uchar4 __ovld vload4(size_t offset, const uchar *p);
-short4 __ovld vload4(size_t offset, const short *p);
-ushort4 __ovld vload4(size_t offset, const ushort *p);
-int4 __ovld vload4(size_t offset, const int *p);
-uint4 __ovld vload4(size_t offset, const uint *p);
-long4 __ovld vload4(size_t offset, const long *p);
-ulong4 __ovld vload4(size_t offset, const ulong *p);
-float4 __ovld vload4(size_t offset, const float *p);
-char8 __ovld vload8(size_t offset, const char *p);
-uchar8 __ovld vload8(size_t offset, const uchar *p);
-short8 __ovld vload8(size_t offset, const short *p);
-ushort8 __ovld vload8(size_t offset, const ushort *p);
-int8 __ovld vload8(size_t offset, const int *p);
-uint8 __ovld vload8(size_t offset, const uint *p);
-long8 __ovld vload8(size_t offset, const long *p);
-ulong8 __ovld vload8(size_t offset, const ulong *p);
-float8 __ovld vload8(size_t offset, const float *p);
-char16 __ovld vload16(size_t offset, const char *p);
-uchar16 __ovld vload16(size_t offset, const uchar *p);
-short16 __ovld vload16(size_t offset, const short *p);
-ushort16 __ovld vload16(size_t offset, const ushort *p);
-int16 __ovld vload16(size_t offset, const int *p);
-uint16 __ovld vload16(size_t offset, const uint *p);
-long16 __ovld vload16(size_t offset, const long *p);
-ulong16 __ovld vload16(size_t offset, const ulong *p);
-float16 __ovld vload16(size_t offset, const float *p);
+char2 __ovld __purefn vload2(size_t offset, const char *p);
+uchar2 __ovld __purefn vload2(size_t offset, const uchar *p);
+short2 __ovld __purefn vload2(size_t offset, const short *p);
+ushort2 __ovld __purefn vload2(size_t offset, const ushort *p);
+int2 __ovld __purefn vload2(size_t offset, const int *p);
+uint2 __ovld __purefn vload2(size_t offset, const uint *p);
+long2 __ovld __purefn vload2(size_t offset, const long *p);
+ulong2 __ovld __purefn vload2(size_t offset, const ulong *p);
+float2 __ovld __purefn vload2(size_t offset, const float *p);
+char3 __ovld __purefn vload3(size_t offset, const char *p);
+uchar3 __ovld __purefn vload3(size_t offset, const uchar *p);
+short3 __ovld __purefn vload3(size_t offset, const short *p);
+ushort3 __ovld __purefn vload3(size_t offset, const ushort *p);
+int3 __ovld __purefn vload3(size_t offset, const int *p);
+uint3 __ovld __purefn vload3(size_t offset, const uint *p);
+long3 __ovld __purefn vload3(size_t offset, const long *p);
+ulong3 __ovld __purefn vload3(size_t offset, const ulong *p);
+float3 __ovld __purefn vload3(size_t offset, const float *p);
+char4 __ovld __purefn vload4(size_t offset, const char *p);
+uchar4 __ovld __purefn vload4(size_t offset, const uchar *p);
+short4 __ovld __purefn vload4(size_t offset, const short *p);
+ushort4 __ovld __purefn vload4(size_t offset, const ushort *p);
+int4 __ovld __purefn vload4(size_t offset, const int *p);
+uint4 __ovld __purefn vload4(size_t offset, const uint *p);
+long4 __ovld __purefn vload4(size_t offset, const long *p);
+ulong4 __ovld __purefn vload4(size_t offset, const ulong *p);
+float4 __ovld __purefn vload4(size_t offset, const float *p);
+char8 __ovld __purefn vload8(size_t offset, const char *p);
+uchar8 __ovld __purefn vload8(size_t offset, const uchar *p);
+short8 __ovld __purefn vload8(size_t offset, const short *p);
+ushort8 __ovld __purefn vload8(size_t offset, const ushort *p);
+int8 __ovld __purefn vload8(size_t offset, const int *p);
+uint8 __ovld __purefn vload8(size_t offset, const uint *p);
+long8 __ovld __purefn vload8(size_t offset, const long *p);
+ulong8 __ovld __purefn vload8(size_t offset, const ulong *p);
+float8 __ovld __purefn vload8(size_t offset, const float *p);
+char16 __ovld __purefn vload16(size_t offset, const char *p);
+uchar16 __ovld __purefn vload16(size_t offset, const uchar *p);
+short16 __ovld __purefn vload16(size_t offset, const short *p);
+ushort16 __ovld __purefn vload16(size_t offset, const ushort *p);
+int16 __ovld __purefn vload16(size_t offset, const int *p);
+uint16 __ovld __purefn vload16(size_t offset, const uint *p);
+long16 __ovld __purefn vload16(size_t offset, const long *p);
+ulong16 __ovld __purefn vload16(size_t offset, const ulong *p);
+float16 __ovld __purefn vload16(size_t offset, const float *p);
 
 #ifdef cl_khr_fp64
-double2 __ovld vload2(size_t offset, const double *p);
-double3 __ovld vload3(size_t offset, const double *p);
-double4 __ovld vload4(size_t offset, const double *p);
-double8 __ovld vload8(size_t offset, const double *p);
-double16 __ovld vload16(size_t offset, const double *p);
+double2 __ovld __purefn vload2(size_t offset, const double *p);
+double3 __ovld __purefn vload3(size_t offset, const double *p);
+double4 __ovld __purefn vload4(size_t offset, const double *p);
+double8 __ovld __purefn vload8(size_t offset, const double *p);
+double16 __ovld __purefn vload16(size_t offset, const double *p);
 #endif //cl_khr_fp64
 
 #ifdef cl_khr_fp16
-half __ovld vload(size_t offset, const half *p);
-half2 __ovld vload2(size_t offset, const half *p);
-half3 __ovld vload3(size_t offset, const half *p);
-half4 __ovld vload4(size_t offset, const half *p);
-half8 __ovld vload8(size_t offset, const half *p);
-half16 __ovld vload16(size_t offset, const half *p);
+half __ovld __purefn vload(size_t offset, const half *p);
+half2 __ovld __purefn vload2(size_t offset, const half *p);
+half3 __ovld __purefn vload3(size_t offset, const half *p);
+half4 __ovld __purefn vload4(size_t offset, const half *p);
+half8 __ovld __purefn vload8(size_t offset, const half *p);
+half16 __ovld __purefn vload16(size_t offset, const half *p);
 #endif //cl_khr_fp16
 #else
-char2 __ovld vload2(size_t offset, const __global char *p);
-uchar2 __ovld vload2(size_t offset, const __global uchar *p);
-short2 __ovld vload2(size_t offset, const __global short *p);
-ushort2 __ovld vload2(size_t offset, const __global ushort *p);
-int2 __ovld vload2(size_t offset, const __global int *p);
-uint2 __ovld vload2(size_t offset, const __global uint *p);
-long2 __ovld vload2(size_t offset, const __global long *p);
-ulong2 __ovld vload2(size_t offset, const __global ulong *p);
-float2 __ovld vload2(size_t offset, const __global float *p);
-char3 __ovld vload3(size_t offset, const __global char *p);
-uchar3 __ovld vload3(size_t offset, const __global uchar *p);
-short3 __ovld vload3(size_t offset, const __global short *p);
-ushort3 __ovld vload3(size_t offset, const __global ushort *p);
-int3 __ovld vload3(size_t offset, const __global int *p);
-uint3 __ovld vload3(size_t offset, const __global uint *p);
-long3 __ovld vload3(size_t offset, const __global long *p);
-ulong3 __ovld vload3(size_t offset, const __global ulong *p);
-float3 __ovld vload3(size_t offset, const __global float *p);
-char4 __ovld vload4(size_t offset, const __global char *p);
-uchar4 __ovld vload4(size_t offset, const __global uchar *p);
-short4 __ovld vload4(size_t offset, const __global short *p);
-ushort4 __ovld vload4(size_t offset, const __global ushort *p);
-int4 __ovld vload4(size_t offset, const __global int *p);
-uint4 __ovld vload4(size_t offset, const __global uint *p);
-long4 __ovld vload4(size_t offset, const __global long *p);
-ulong4 __ovld vload4(size_t offset, const __global ulong *p);
-float4 __ovld vload4(size_t offset, const __global float *p);
-char8 __ovld vload8(size_t offset, const __global char *p);
-uchar8 __ovld vload8(size_t offset, const __global uchar *p);
-short8 __ovld vload8(size_t offset, const __global short *p);
-ushort8 __ovld vload8(size_t offset, const __global ushort *p);
-int8 __ovld vload8(size_t offset, const __global int *p);
-uint8 __ovld vload8(size_t offset, const __global uint *p);
-long8 __ovld vload8(size_t offset, const __global long *p);
-ulong8 __ovld vload8(size_t offset, const __global ulong *p);
-float8 __ovld vload8(size_t offset, const __global float *p);
-char16 __ovld vload16(size_t offset, const __global char *p);
-uchar16 __ovld vload16(size_t offset, const __global uchar *p);
-short16 __ovld vload16(size_t offset, const __global short *p);
-ushort16 __ovld vload16(size_t offset, const __global ushort *p);
-int16 __ovld vload16(size_t offset, const __global int *p);
-uint16 __ovld vload16(size_t offset, const __global uint *p);
-long16 __ovld vload16(size_t offset, const __global long *p);
-ulong16 __ovld vload16(size_t offset, const __global ulong *p);
-float16 __ovld vload16(size_t offset, const __global float *p);
-char2 __ovld vload2(size_t offset, const __local char *p);
-uchar2 __ovld vload2(size_t offset, const __local uchar *p);
-short2 __ovld vload2(size_t offset, const __local short *p);
-ushort2 __ovld vload2(size_t offset, const __local ushort *p);
-int2 __ovld vload2(size_t offset, const __local int *p);
-uint2 __ovld vload2(size_t offset, const __local uint *p);
-long2 __ovld vload2(size_t offset, const __local long *p);
-ulong2 __ovld vload2(size_t offset, const __local ulong *p);
-float2 __ovld vload2(size_t offset, const __local float *p);
-char3 __ovld vload3(size_t offset, const __local char *p);
-uchar3 __ovld vload3(size_t offset, const __local uchar *p);
-short3 __ovld vload3(size_t offset, const __local short *p);
-ushort3 __ovld vload3(size_t offset, const __local ushort *p);
-int3 __ovld vload3(size_t offset, const __local int *p);
-uint3 __ovld vload3(size_t offset, const __local uint *p);
-long3 __ovld vload3(size_t offset, const __local long *p);
-ulong3 __ovld vload3(size_t offset, const __local ulong *p);
-float3 __ovld vload3(size_t offset, const __local float *p);
-char4 __ovld vload4(size_t offset, const __local char *p);
-uchar4 __ovld vload4(size_t offset, const __local uchar *p);
-short4 __ovld vload4(size_t offset, const __local short *p);
-ushort4 __ovld vload4(size_t offset, const __local ushort *p);
-int4 __ovld vload4(size_t offset, const __local int *p);
-uint4 __ovld vload4(size_t offset, const __local uint *p);
-long4 __ovld vload4(size_t offset, const __local long *p);
-ulong4 __ovld vload4(size_t offset, const __local ulong *p);
-float4 __ovld vload4(size_t offset, const __local float *p);
-char8 __ovld vload8(size_t offset, const __local char *p);
-uchar8 __ovld vload8(size_t offset, const __local uchar *p);
-short8 __ovld vload8(size_t offset, const __local short *p);
-ushort8 __ovld vload8(size_t offset, const __local ushort *p);
-int8 __ovld vload8(size_t offset, const __local int *p);
-uint8 __ovld vload8(size_t offset, const __local uint *p);
-long8 __ovld vload8(size_t offset, const __local long *p);
-ulong8 __ovld vload8(size_t offset, const __local ulong *p);
-float8 __ovld vload8(size_t offset, const __local float *p);
-char16 __ovld vload16(size_t offset, const __local char *p);
-uchar16 __ovld vload16(size_t offset, const __local uchar *p);
-short16 __ovld vload16(size_t offset, const __local short *p);
-ushort16 __ovld vload16(size_t offset, const __local ushort *p);
-int16 __ovld vload16(size_t offset, const __local int *p);
-uint16 __ovld vload16(size_t offset, const __local uint *p);
-long16 __ovld vload16(size_t offset, const __local long *p);
-ulong16 __ovld vload16(size_t offset, const __local ulong *p);
-float16 __ovld vload16(size_t offset, const __local float *p);
-char2 __ovld vload2(size_t offset, const __private char *p);
-uchar2 __ovld vload2(size_t offset, const __private uchar *p);
-short2 __ovld vload2(size_t offset, const __private short *p);
-ushort2 __ovld vload2(size_t offset, const __private ushort *p);
-int2 __ovld vload2(size_t offset, const __private int *p);
-uint2 __ovld vload2(size_t offset, const __private uint *p);
-long2 __ovld vload2(size_t offset, const __private long *p);
-ulong2 __ovld vload2(size_t offset, const __private ulong *p);
-float2 __ovld vload2(size_t offset, const __private float *p);
-char3 __ovld vload3(size_t offset, const __private char *p);
-uchar3 __ovld vload3(size_t offset, const __private uchar *p);
-short3 __ovld vload3(size_t offset, const __private short *p);
-ushort3 __ovld vload3(size_t offset, const __private ushort *p);
-int3 __ovld vload3(size_t offset, const __private int *p);
-uint3 __ovld vload3(size_t offset, const __private uint *p);
-long3 __ovld vload3(size_t offset, const __private long *p);
-ulong3 __ovld vload3(size_t offset, const __private ulong *p);
-float3 __ovld vload3(size_t offset, const __private float *p);
-char4 __ovld vload4(size_t offset, const __private char *p);
-uchar4 __ovld vload4(size_t offset, const __private uchar *p);
-short4 __ovld vload4(size_t offset, const __private short *p);
-ushort4 __ovld vload4(size_t offset, const __private ushort *p);
-int4 __ovld vload4(size_t offset, const __private int *p);
-uint4 __ovld vload4(size_t offset, const __private uint *p);
-long4 __ovld vload4(size_t offset, const __private long *p);
-ulong4 __ovld vload4(size_t offset, const __private ulong *p);
-float4 __ovld vload4(size_t offset, const __private float *p);
-char8 __ovld vload8(size_t offset, const __private char *p);
-uchar8 __ovld vload8(size_t offset, const __private uchar *p);
-short8 __ovld vload8(size_t offset, const __private short *p);
-ushort8 __ovld vload8(size_t offset, const __private ushort *p);
-int8 __ovld vload8(size_t offset, const __private int *p);
-uint8 __ovld vload8(size_t offset, const __private uint *p);
-long8 __ovld vload8(size_t offset, const __private long *p);
-ulong8 __ovld vload8(size_t offset, const __private ulong *p);
-float8 __ovld vload8(size_t offset, const __private float *p);
-char16 __ovld vload16(size_t offset, const __private char *p);
-uchar16 __ovld vload16(size_t offset, const __private uchar *p);
-short16 __ovld vload16(size_t offset, const __private short *p);
-ushort16 __ovld vload16(size_t offset, const __private ushort *p);
-int16 __ovld vload16(size_t offset, const __private int *p);
-uint16 __ovld vload16(size_t offset, const __private uint *p);
-long16 __ovld vload16(size_t offset, const __private long *p);
-ulong16 __ovld vload16(size_t offset, const __private ulong *p);
-float16 __ovld vload16(size_t offset, const __private float *p);
+char2 __ovld __purefn vload2(size_t offset, const __global char *p);
+uchar2 __ovld __purefn vload2(size_t offset, const __global uchar *p);
+short2 __ovld __purefn vload2(size_t offset, const __global short *p);
+ushort2 __ovld __purefn vload2(size_t offset, const __global ushort *p);
+int2 __ovld __purefn vload2(size_t offset, const __global int *p);
+uint2 __ovld __purefn vload2(size_t offset, const __global uint *p);
+long2 __ovld __purefn vload2(size_t offset, const __global long *p);
+ulong2 __ovld __purefn vload2(size_t offset, const __global ulong *p);
+float2 __ovld __purefn vload2(size_t offset, const __global float *p);
+char3 __ovld __purefn vload3(size_t offset, const __global char *p);
+uchar3 __ovld __purefn vload3(size_t offset, const __global uchar *p);
+short3 __ovld __purefn vload3(size_t offset, const __global short *p);
+ushort3 __ovld __purefn vload3(size_t offset, const __global ushort *p);
+int3 __ovld __purefn vload3(size_t offset, const __global int *p);
+uint3 __ovld __purefn vload3(size_t offset, const __global uint *p);
+long3 __ovld __purefn vload3(size_t offset, const __global long *p);
+ulong3 __ovld __purefn vload3(size_t offset, const __global ulong *p);
+float3 __ovld __purefn vload3(size_t offset, const __global float *p);
+char4 __ovld __purefn vload4(size_t offset, const __global char *p);
+uchar4 __ovld __purefn vload4(size_t offset, const __global uchar *p);
+short4 __ovld __purefn vload4(size_t offset, const __global short *p);
+ushort4 __ovld __purefn vload4(size_t offset, const __global ushort *p);
+int4 __ovld __purefn vload4(size_t offset, const __global int *p);
+uint4 __ovld __purefn vload4(size_t offset, const __global uint *p);
+long4 __ovld __purefn vload4(size_t offset, const __global long *p);
+ulong4 __ovld __purefn vload4(size_t offset, const __global ulong *p);
+float4 __ovld __purefn vload4(size_t offset, const __global float *p);
+char8 __ovld __purefn vload8(size_t offset, const __global char *p);
+uchar8 __ovld __purefn vload8(size_t offset, const __global uchar *p);
+short8 __ovld __purefn vload8(size_t offset, const __global short *p);
+ushort8 __ovld __purefn vload8(size_t offset, const __global ushort *p);
+int8 __ovld __purefn vload8(size_t offset, const __global int *p);
+uint8 __ovld __purefn vload8(size_t offset, const __global uint *p);
+long8 __ovld __purefn vload8(size_t offset, const __global long *p);
+ulong8 __ovld __purefn vload8(size_t offset, const __global ulong *p);
+float8 __ovld __purefn vload8(size_t offset, const __global float *p);
+char16 __ovld __purefn vload16(size_t offset, const __global char *p);
+uchar16 __ovld __purefn vload16(size_t offset, const __global uchar *p);
+short16 __ovld __purefn vload16(size_t offset, const __global short *p);
+ushort16 __ovld __purefn vload16(size_t offset, const __global ushort *p);
+int16 __ovld __purefn vload16(size_t offset, const __global int *p);
+uint16 __ovld __purefn vload16(size_t offset, const __global uint *p);
+long16 __ovld __purefn vload16(size_t offset, const __global long *p);
+ulong16 __ovld __purefn vload16(size_t offset, const __global ulong *p);
+float16 __ovld __purefn vload16(size_t offset, const __global float *p);
+char2 __ovld __purefn vload2(size_t offset, const __local char *p);
+uchar2 __ovld __purefn vload2(size_t offset, const __local uchar *p);
+short2 __ovld __purefn vload2(size_t offset, const __local short *p);
+ushort2 __ovld __purefn vload2(size_t offset, const __local ushort *p);
+int2 __ovld __purefn vload2(size_t offset, const __local int *p);
+uint2 __ovld __purefn vload2(size_t offset, const __local uint *p);
+long2 __ovld __purefn vload2(size_t offset, const __local long *p);
+ulong2 __ovld __purefn vload2(size_t offset, const __local ulong *p);
+float2 __ovld __purefn vload2(size_t offset, const __local float *p);
+char3 __ovld __purefn vload3(size_t offset, const __local char *p);
+uchar3 __ovld __purefn vload3(size_t offset, const __local uchar *p);
+short3 __ovld __purefn vload3(size_t offset, const __local short *p);
+ushort3 __ovld __purefn vload3(size_t offset, const __local ushort *p);
+int3 __ovld __purefn vload3(size_t offset, const __local int *p);
+uint3 __ovld __purefn vload3(size_t offset, const __local uint *p);
+long3 __ovld __purefn vload3(size_t offset, const __local long *p);
+ulong3 __ovld __purefn vload3(size_t offset, const __local ulong *p);
+float3 __ovld __purefn vload3(size_t offset, const __local float *p);
+char4 __ovld __purefn vload4(size_t offset, const __local char *p);
+uchar4 __ovld __purefn vload4(size_t offset, const __local uchar *p);
+short4 __ovld __purefn vload4(size_t offset, const __local short *p);
+ushort4 __ovld __purefn vload4(size_t offset, const __local ushort *p);
+int4 __ovld __purefn vload4(size_t offset, const __local int *p);
+uint4 __ovld __purefn vload4(size_t offset, const __local uint *p);
+long4 __ovld __purefn vload4(size_t offset, const __local long *p);
+ulong4 __ovld __purefn vload4(size_t offset, const __local ulong *p);
+float4 __ovld __purefn vload4(size_t offset, const __local float *p);
+char8 __ovld __purefn vload8(size_t offset, const __local char *p);
+uchar8 __ovld __purefn vload8(size_t offset, const __local uchar *p);
+short8 __ovld __purefn vload8(size_t offset, const __local short *p);
+ushort8 __ovld __purefn vload8(size_t offset, const __local ushort *p);
+int8 __ovld __purefn vload8(size_t offset, const __local int *p);
+uint8 __ovld __purefn vload8(size_t offset, const __local uint *p);
+long8 __ovld __purefn vload8(size_t offset, const __local long *p);
+ulong8 __ovld __purefn vload8(size_t offset, const __local ulong *p);
+float8 __ovld __purefn vload8(size_t offset, const __local float *p);
+char16 __ovld __purefn vload16(size_t offset, const __local char *p);
+uchar16 __ovld __purefn vload16(size_t offset, const __local uchar *p);
+short16 __ovld __purefn vload16(size_t offset, const __local short *p);
+ushort16 __ovld __purefn vload16(size_t offset, const __local ushort *p);
+int16 __ovld __purefn vload16(size_t offset, const __local int *p);
+uint16 __ovld __purefn vload16(size_t offset, const __local uint *p);
+long16 __ovld __purefn vload16(size_t offset, const __local long *p);
+ulong16 __ovld __purefn vload16(size_t offset, const __local ulong *p);
+float16 __ovld __purefn vload16(size_t offset, const __local float *p);
+char2 __ovld __purefn vload2(size_t offset, const __private char *p);
+uchar2 __ovld __purefn vload2(size_t offset, const __private uchar *p);
+short2 __ovld __purefn vload2(size_t offset, const __private short *p);
+ushort2 __ovld __purefn vload2(size_t offset, const __private ushort *p);
+int2 __ovld __purefn vload2(size_t offset, const __private int *p);
+uint2 __ovld __purefn vload2(size_t offset, const __private uint *p);
+long2 __ovld __purefn vload2(size_t offset, const __private long *p);
+ulong2 __ovld __purefn vload2(size_t offset, const __private ulong *p);
+float2 __ovld __purefn vload2(size_t offset, const __private float *p);
+char3 __ovld __purefn vload3(size_t offset, const __private char *p);
+uchar3 __ovld __purefn vload3(size_t offset, const __private uchar *p);
+short3 __ovld __purefn vload3(size_t offset, const __private short *p);
+ushort3 __ovld __purefn vload3(size_t offset, const __private ushort *p);
+int3 __ovld __purefn vload3(size_t offset, const __private int *p);
+uint3 __ovld __purefn vload3(size_t offset, const __private uint *p);
+long3 __ovld __purefn vload3(size_t offset, const __private long *p);
+ulong3 __ovld __purefn vload3(size_t offset, const __private ulong *p);
+float3 __ovld __purefn vload3(size_t offset, const __private float *p);
+char4 __ovld __purefn vload4(size_t offset, const __private char *p);
+uchar4 __ovld __purefn vload4(size_t offset, const __private uchar *p);
+short4 __ovld __purefn vload4(size_t offset, const __private short *p);
+ushort4 __ovld __purefn vload4(size_t offset, const __private ushort *p);
+int4 __ovld __purefn vload4(size_t offset, const __private int *p);
+uint4 __ovld __purefn vload4(size_t offset, const __private uint *p);
+long4 __ovld __purefn vload4(size_t offset, const __private long *p);
+ulong4 __ovld __purefn vload4(size_t offset, const __private ulong *p);
+float4 __ovld __purefn vload4(size_t offset, const __private float *p);
+char8 __ovld __purefn vload8(size_t offset, const __private char *p);
+uchar8 __ovld __purefn vload8(size_t offset, const __private uchar *p);
+short8 __ovld __purefn vload8(size_t offset, const __private short *p);
+ushort8 __ovld __purefn vload8(size_t offset, const __private ushort *p);
+int8 __ovld __purefn vload8(size_t offset, const __private int *p);
+uint8 __ovld __purefn vload8(size_t offset, const __private uint *p);
+long8 __ovld __purefn vload8(size_t offset, const __private long *p);
+ulong8 __ovld __purefn vload8(size_t offset, const __private ulong *p);
+float8 __ovld __purefn vload8(size_t offset, const __private float *p);
+char16 __ovld __purefn vload16(size_t offset, const __private char *p);
+uchar16 __ovld __purefn vload16(size_t offset, const __private uchar *p);
+short16 __ovld __purefn vload16(size_t offset, const __private short *p);
+ushort16 __ovld __purefn vload16(size_t offset, const __private ushort *p);
+int16 __ovld __purefn vload16(size_t offset, const __private int *p);
+uint16 __ovld __purefn vload16(size_t offset, const __private uint *p);
+long16 __ovld __purefn vload16(size_t offset, const __private long *p);
+ulong16 __ovld __purefn vload16(size_t offset, const __private ulong *p);
+float16 __ovld __purefn vload16(size_t offset, const __private float *p);
 
 #ifdef cl_khr_fp64
-double2 __ovld vload2(size_t offset, const __global double *p);
-double3 __ovld vload3(size_t offset, const __global double *p);
-double4 __ovld vload4(size_t offset, const __global double *p);
-double8 __ovld vload8(size_t offset, const __global double *p);
-double16 __ovld vload16(size_t offset, const __global double *p);
-double2 __ovld vload2(size_t offset, const __local double *p);
-double3 __ovld vload3(size_t offset, const __local double *p);
-double4 __ovld vload4(size_t offset, const __local double *p);
-double8 __ovld vload8(size_t offset, const __local double *p);
-double16 __ovld vload16(size_t offset, const __local double *p);
-double2 __ovld vload2(size_t offset, const __private double *p);
-double3 __ovld vload3(size_t offset, const __private double *p);
-double4 __ovld vload4(size_t offset, const __private double *p);
-double8 __ovld vload8(size_t offset, const __private double *p);
-double16 __ovld vload16(size_t offset, const __private double *p);
+double2 __ovld __purefn vload2(size_t offset, const __global double *p);
+double3 __ovld __purefn vload3(size_t offset, const __global double *p);
+double4 __ovld __purefn vload4(size_t offset, const __global double *p);
+double8 __ovld __purefn vload8(size_t offset, const __global double *p);
+double16 __ovld __purefn vload16(size_t offset, const __global double *p);
+double2 __ovld __purefn vload2(size_t offset, const __local double *p);
+double3 __ovld __purefn vload3(size_t offset, const __local double *p);
+double4 __ovld __purefn vload4(size_t offset, const __local double *p);
+double8 __ovld __purefn vload8(size_t offset, const __local double *p);
+double16 __ovld __purefn vload16(size_t offset, const __local double *p);
+double2 __ovld __purefn vload2(size_t offset, const __private double *p);
+double3 __ovld __purefn vload3(size_t offset, const __private double *p);
+double4 __ovld __purefn vload4(size_t offset, const __private double *p);
+double8 __ovld __purefn vload8(size_t offset, const __private double *p);
+double16 __ovld __purefn vload16(size_t offset, const __private double *p);
 #endif //cl_khr_fp64
 
 #ifdef cl_khr_fp16
-half __ovld vload(size_t offset, const __global half *p);
-half2 __ovld vload2(size_t offset, const __global half *p);
-half3 __ovld vload3(size_t offset, const __global half *p);
-half4 __ovld vload4(size_t offset, const __global half *p);
-half8 __ovld vload8(size_t offset, const __global half *p);
-half16 __ovld vload16(size_t offset, const __global half *p);
-half __ovld vload(size_t offset, const __local half *p);
-half2 __ovld vload2(size_t offset, const __local half *p);
-half3 __ovld vload3(size_t offset, const __local half *p);
-half4 __ovld vload4(size_t offset, const __local half *p);
-half8 __ovld vload8(size_t offset, const __local half *p);
-half16 __ovld vload16(size_t offset, const __local half *p);
-half __ovld vload(size_t offset, const __private half *p);
-half2 __ovld vload2(size_t offset, const __private half *p);
-half3 __ovld vload3(size_t offset, const __private half *p);
-half4 __ovld vload4(size_t offset, const __private half *p);
-half8 __ovld vload8(size_t offset, const __private half *p);
-half16 __ovld vload16(size_t offset, const __private half *p);
+half __ovld __purefn vload(size_t offset, const __global half *p);
+half2 __ovld __purefn vload2(size_t offset, const __global half *p);
+half3 __ovld __purefn vload3(size_t offset, const __global half *p);
+half4 __ovld __purefn vload4(size_t offset, const __global half *p);
+half8 __ovld __purefn vload8(size_t offset, const __global half *p);
+half16 __ovld __purefn vload16(size_t offset, const __global half *p);
+half __ovld __purefn vload(size_t offset, const __local half *p);
+half2 __ovld __purefn vload2(size_t offset, const __local half *p);
+half3 __ovld __purefn vload3(size_t offset, const __local half *p);
+half4 __ovld __purefn vload4(size_t offset, const __local half *p);
+half8 __ovld __purefn vload8(size_t offset, const __local half *p);
+half16 __ovld __purefn vload16(size_t offset, const __local half *p);
+half __ovld __purefn vload(size_t offset, const __private half *p);
+half2 __ovld __purefn vload2(size_t offset, const __private half *p);
+half3 __ovld __purefn vload3(size_t offset, const __private half *p);
+half4 __ovld __purefn vload4(size_t offset, const __private half *p);
+half8 __ovld __purefn vload8(size_t offset, const __private half *p);
+half16 __ovld __purefn vload16(size_t offset, const __private half *p);
 #endif //cl_khr_fp16
 #endif //defined(__opencl_c_generic_address_space)
 
@@ -11736,13 +11736,13 @@
  * The read address computed as (p + offset)
  * must be 16-bit aligned.
  */
-float __ovld vload_half(size_t offset, const __constant half *p);
+float __ovld __purefn vload_half(size_t offset, const __constant half *p);
 #if defined(__opencl_c_generic_address_space)
-float __ovld vload_half(size_t offset, const half *p);
+float __ovld __purefn vload_half(size_t offset, const half *p);
 #else
-float __ovld vload_half(size_t offset, const __global half *p);
-float __ovld vload_half(size_t offset, const __local half *p);
-float __ovld vload_half(size_t offset, const __private half *p);
+float __ovld __purefn vload_half(size_t offset, const __global half *p);
+float __ovld __purefn vload_half(size_t offset, const __local half *p);
+float __ovld __purefn vload_half(size_t offset, const __private half *p);
 #endif //defined(__opencl_c_generic_address_space)
 
 /**
@@ -11753,33 +11753,33 @@
  * value is returned. The read address computed
  * as (p + (offset * n)) must be 16-bit aligned.
  */
-float2 __ovld vload_half2(size_t offset, const __constant half *p);
-float3 __ovld vload_half3(size_t offset, const __constant half *p);
-float4 __ovld vload_half4(size_t offset, const __constant half *p);
-float8 __ovld vload_half8(size_t offset, const __constant half *p);
-float16 __ovld vload_half16(size_t offset, const __constant half *p);
+float2 __ovld __purefn vload_half2(size_t offset, const __constant half *p);
+float3 __ovld __purefn vload_half3(size_t offset, const __constant half *p);
+float4 __ovld __purefn vload_half4(size_t offset, const __constant half *p);
+float8 __ovld __purefn vload_half8(size_t offset, const __constant half *p);
+float16 __ovld __purefn vload_half16(size_t offset, const __constant half *p);
 #if defined(__opencl_c_generic_address_space)
-float2 __ovld vload_half2(size_t offset, const half *p);
-float3 __ovld vload_half3(size_t offset, const half *p);
-float4 __ovld vload_half4(size_t offset, const half *p);
-float8 __ovld vload_half8(size_t offset, const half *p);
-float16 __ovld vload_half16(size_t offset, const half *p);
+float2 __ovld __purefn vload_half2(size_t offset, const half *p);
+float3 __ovld __purefn vload_half3(size_t offset, const half *p);
+float4 __ovld __purefn vload_half4(size_t offset, const half *p);
+float8 __ovld __purefn vload_half8(size_t offset, const half *p);
+float16 __ovld __purefn vload_half16(size_t offset, const half *p);
 #else
-float2 __ovld vload_half2(size_t offset, const __global half *p);
-float3 __ovld vload_half3(size_t offset, const __global half *p);
-float4 __ovld vload_half4(size_t offset, const __global half *p);
-float8 __ovld vload_half8(size_t offset, const __global half *p);
-float16 __ovld vload_half16(size_t offset, const __global half *p);
-float2 __ovld vload_half2(size_t offset, const __local half *p);
-float3 __ovld vload_half3(size_t offset, const __local half *p);
-float4 __ovld vload_half4(size_t offset, const __local half *p);
-float8 __ovld vload_half8(size_t offset, const __local half *p);
-float16 __ovld vload_half16(size_t offset, const __local half *p);
-float2 __ovld vload_half2(size_t offset, const __private half *p);
-float3 __ovld vload_half3(size_t offset, const __private half *p);
-float4 __ovld vload_half4(size_t offset, const __private half *p);
-float8 __ovld vload_half8(size_t offset, const __private half *p);
-float16 __ovld vload_half16(size_t offset, const __private half *p);
+float2 __ovld __purefn vload_half2(size_t offset, const __global half *p);
+float3 __ovld __purefn vload_half3(size_t offset, const __global half *p);
+float4 __ovld __purefn vload_half4(size_t offset, const __global half *p);
+float8 __ovld __purefn vload_half8(size_t offset, const __global half *p);
+float16 __ovld __purefn vload_half16(size_t offset, const __global half *p);
+float2 __ovld __purefn vload_half2(size_t offset, const __local half *p);
+float3 __ovld __purefn vload_half3(size_t offset, const __local half *p);
+float4 __ovld __purefn vload_half4(size_t offset, const __local half *p);
+float8 __ovld __purefn vload_half8(size_t offset, const __local half *p);
+float16 __ovld __purefn vload_half16(size_t offset, const __local half *p);
+float2 __ovld __purefn vload_half2(size_t offset, const __private half *p);
+float3 __ovld __purefn vload_half3(size_t offset, const __private half *p);
+float4 __ovld __purefn vload_half4(size_t offset, const __private half *p);
+float8 __ovld __purefn vload_half8(size_t offset, const __private half *p);
+float16 __ovld __purefn vload_half16(size_t offset, const __private half *p);
 #endif //defined(__opencl_c_generic_address_space)
 
 /**
@@ -12073,33 +12073,33 @@
  * The address computed as (p + (offset * 4))
  * must be aligned to sizeof (half) * 4 bytes.
  */
-float2 __ovld vloada_half2(size_t offset, const __constant half *p);
-float3 __ovld vloada_half3(size_t offset, const __constant half *p);
-float4 __ovld vloada_half4(size_t offset, const __constant half *p);
-float8 __ovld vloada_half8(size_t offset, const __constant half *p);
-float16 __ovld vloada_half16(size_t offset, const __constant half *p);
+float2 __ovld __purefn vloada_half2(size_t offset, const __constant half *p);
+float3 __ovld __purefn vloada_half3(size_t offset, const __constant half *p);
+float4 __ovld __purefn vloada_half4(size_t offset, const __constant half *p);
+float8 __ovld __purefn vloada_half8(size_t offset, const __constant half *p);
+float16 __ovld __purefn vloada_half16(size_t offset, const __constant half *p);
 #if defined(__opencl_c_generic_address_space)
-float2 __ovld vloada_half2(size_t offset, const half *p);
-float3 __ovld vloada_half3(size_t offset, const half *p);
-float4 __ovld vloada_half4(size_t offset, const half *p);
-float8 __ovld vloada_half8(size_t offset, const half *p);
-float16 __ovld vloada_half16(size_t offset, const half *p);
+float2 __ovld __purefn vloada_half2(size_t offset, const half *p);
+float3 __ovld __purefn vloada_half3(size_t offset, const half *p);
+float4 __ovld __purefn vloada_half4(size_t offset, const half *p);
+float8 __ovld __purefn vloada_half8(size_t offset, const half *p);
+float16 __ovld __purefn vloada_half16(size_t offset, const half *p);
 #else
-float2 __ovld vloada_half2(size_t offset, const __global half *p);
-float3 __ovld vloada_half3(size_t offset, const __global half *p);
-float4 __ovld vloada_half4(size_t offset, const __global half *p);
-float8 __ovld vloada_half8(size_t offset, const __global half *p);
-float16 __ovld vloada_half16(size_t offset, const __global half *p);
-float2 __ovld vloada_half2(size_t offset, const __local half *p);
-float3 __ovld vloada_half3(size_t offset, const __local half *p);
-float4 __ovld vloada_half4(size_t offset, const __local half *p);
-float8 __ovld vloada_half8(size_t offset, const __local half *p);
-float16 __ovld vloada_half16(size_t offset, const __local half *p);
-float2 __ovld vloada_half2(size_t offset, const __private half *p);
-float3 __ovld vloada_half3(size_t offset, const __private half *p);
-float4 __ovld vloada_half4(size_t offset, const __private half *p);
-float8 __ovld vloada_half8(size_t offset, const __private half *p);
-float16 __ovld vloada_half16(size_t offset, const __private half *p);
+float2 __ovld __purefn vloada_half2(size_t offset, const __global half *p);
+float3 __ovld __purefn vloada_half3(size_t offset, const __global half *p);
+float4 __ovld __purefn vloada_half4(size_t offset, const __global half *p);
+float8 __ovld __purefn vloada_half8(size_t offset, const __global half *p);
+float16 __ovld __purefn vloada_half16(size_t offset, const __global half *p);
+float2 __ovld __purefn vloada_half2(size_t offset, const __local half *p);
+float3 __ovld __purefn vloada_half3(size_t offset, const __local half *p);
+float4 __ovld __purefn vloada_half4(size_t offset, const __local half *p);
+float8 __ovld __purefn vloada_half8(size_t offset, const __local half *p);
+float16 __ovld __purefn vloada_half16(size_t offset, const __local half *p);
+float2 __ovld __purefn vloada_half2(size_t offset, const __private half *p);
+float3 __ovld __purefn vloada_half3(size_t offset, const __private half *p);
+float4 __ovld __purefn vloada_half4(size_t offset, const __private half *p);
+float8 __ovld __purefn vloada_half8(size_t offset, const __private half *p);
+float16 __ovld __purefn vloada_half16(size_t offset, const __private half *p);
 #endif //defined(__opencl_c_generic_address_space)
 
 /**
diff --git a/clang/lib/Sema/OpenCLBuiltins.td b/clang/lib/Sema/OpenCLBuiltins.td
--- a/clang/lib/Sema/OpenCLBuiltins.td
+++ b/clang/lib/Sema/OpenCLBuiltins.td
@@ -806,17 +806,17 @@
   foreach AS = addrspaces in {
     foreach VSize = [2, 3, 4, 8, 16] in {
       foreach name = ["vload" # VSize] in {
-        def : Builtin<name, [VectorType<Char, VSize>, Size, PointerType<ConstType<Char>, AS>]>;
-        def : Builtin<name, [VectorType<UChar, VSize>, Size, PointerType<ConstType<UChar>, AS>]>;
-        def : Builtin<name, [VectorType<Short, VSize>, Size, PointerType<ConstType<Short>, AS>]>;
-        def : Builtin<name, [VectorType<UShort, VSize>, Size, PointerType<ConstType<UShort>, AS>]>;
-        def : Builtin<name, [VectorType<Int, VSize>, Size, PointerType<ConstType<Int>, AS>]>;
-        def : Builtin<name, [VectorType<UInt, VSize>, Size, PointerType<ConstType<UInt>, AS>]>;
-        def : Builtin<name, [VectorType<Long, VSize>, Size, PointerType<ConstType<Long>, AS>]>;
-        def : Builtin<name, [VectorType<ULong, VSize>, Size, PointerType<ConstType<ULong>, AS>]>;
-        def : Builtin<name, [VectorType<Float, VSize>, Size, PointerType<ConstType<Float>, AS>]>;
-        def : Builtin<name, [VectorType<Double, VSize>, Size, PointerType<ConstType<Double>, AS>]>;
-        def : Builtin<name, [VectorType<Half, VSize>, Size, PointerType<ConstType<Half>, AS>]>;
+        def : Builtin<name, [VectorType<Char, VSize>, Size, PointerType<ConstType<Char>, AS>], Attr.Pure>;
+        def : Builtin<name, [VectorType<UChar, VSize>, Size, PointerType<ConstType<UChar>, AS>], Attr.Pure>;
+        def : Builtin<name, [VectorType<Short, VSize>, Size, PointerType<ConstType<Short>, AS>], Attr.Pure>;
+        def : Builtin<name, [VectorType<UShort, VSize>, Size, PointerType<ConstType<UShort>, AS>], Attr.Pure>;
+        def : Builtin<name, [VectorType<Int, VSize>, Size, PointerType<ConstType<Int>, AS>], Attr.Pure>;
+        def : Builtin<name, [VectorType<UInt, VSize>, Size, PointerType<ConstType<UInt>, AS>], Attr.Pure>;
+        def : Builtin<name, [VectorType<Long, VSize>, Size, PointerType<ConstType<Long>, AS>], Attr.Pure>;
+        def : Builtin<name, [VectorType<ULong, VSize>, Size, PointerType<ConstType<ULong>, AS>], Attr.Pure>;
+        def : Builtin<name, [VectorType<Float, VSize>, Size, PointerType<ConstType<Float>, AS>], Attr.Pure>;
+        def : Builtin<name, [VectorType<Double, VSize>, Size, PointerType<ConstType<Double>, AS>], Attr.Pure>;
+        def : Builtin<name, [VectorType<Half, VSize>, Size, PointerType<ConstType<Half>, AS>], Attr.Pure>;
       }
       if defStores then {
         foreach name = ["vstore" # VSize] in {
@@ -848,10 +848,10 @@
 
 multiclass VloadVstoreHalf<list<AddressSpace> addrspaces, bit defStores> {
   foreach AS = addrspaces in {
-    def : Builtin<"vload_half", [Float, Size, PointerType<ConstType<Half>, AS>]>;
+    def : Builtin<"vload_half", [Float, Size, PointerType<ConstType<Half>, AS>], Attr.Pure>;
     foreach VSize = [2, 3, 4, 8, 16] in {
       foreach name = ["vload_half" # VSize, "vloada_half" # VSize] in {
-        def : Builtin<name, [VectorType<Float, VSize>, Size, PointerType<ConstType<Half>, AS>]>;
+        def : Builtin<name, [VectorType<Float, VSize>, Size, PointerType<ConstType<Half>, AS>], Attr.Pure>;
       }
     }
     if defStores then {
@@ -877,7 +877,7 @@
 let MinVersion = CL20 in {
   defm : VloadVstoreHalf<[GenericAS], 1>;
 }
-// vload with constant address space is available regardless of version.
+// vload_half and vloada_half with constant address space are available regardless of version.
 defm : VloadVstoreHalf<[ConstantAS], 0>;
 
 // OpenCL v3.0 s6.15.8 - Synchronization Functions.
diff --git a/clang/test/SemaOpenCL/fdeclare-opencl-builtins.cl b/clang/test/SemaOpenCL/fdeclare-opencl-builtins.cl
--- a/clang/test/SemaOpenCL/fdeclare-opencl-builtins.cl
+++ b/clang/test/SemaOpenCL/fdeclare-opencl-builtins.cl
@@ -46,12 +46,19 @@
 typedef char char2 __attribute__((ext_vector_type(2)));
 typedef char char4 __attribute__((ext_vector_type(4)));
 typedef uchar uchar4 __attribute__((ext_vector_type(4)));
+typedef uchar uchar16 __attribute__((ext_vector_type(16)));
 typedef float float4 __attribute__((ext_vector_type(4)));
+typedef float float16 __attribute__((ext_vector_type(16)));
 typedef half half4 __attribute__((ext_vector_type(4)));
 typedef int int2 __attribute__((ext_vector_type(2)));
 typedef int int4 __attribute__((ext_vector_type(4)));
+typedef uint uint2 __attribute__((ext_vector_type(2)));
 typedef uint uint4 __attribute__((ext_vector_type(4)));
 typedef long long2 __attribute__((ext_vector_type(2)));
+typedef long long8 __attribute__((ext_vector_type(8)));
+typedef ulong ulong4 __attribute__((ext_vector_type(4)));
+typedef short short16 __attribute__((ext_vector_type(16)));
+typedef ushort ushort3 __attribute__((ext_vector_type(3)));
 
 typedef int clk_profiling_info;
 #define CLK_PROFILING_COMMAND_EXEC_TIME 0x1
@@ -284,18 +291,27 @@
   global void *global_p;
   private void *private_p;
   size_t s;
+  ulong4 ul4;
+  short16 s16;
+#if __OPENCL_C_VERSION__ >= CL_VERSION_2_0
+  ushort3 us3;
+  uchar16 uc16;
+#endif
+  long8 l8;
+  uint2 ui2;
+  float16 f16;
 
-  vload4(s, (const __constant ulong *) constant_p);
-  vload16(s, (const __constant short *) constant_p);
+  ul4 = vload4(s, (const __constant ulong *) constant_p);
+  s16 = vload16(s, (const __constant short *) constant_p);
 
 #if __OPENCL_C_VERSION__ >= CL_VERSION_2_0
-  vload3(s, (const __generic ushort *) generic_p);
-  vload16(s, (const __generic uchar *) generic_p);
+  us3 = vload3(s, (const __generic ushort *) generic_p);
+  uc16 = vload16(s, (const __generic uchar *) generic_p);
 #endif
 
-  vload8(s, (const __global long *) global_p);
-  vload2(s, (const __local uint *) local_p);
-  vload16(s, (const __private float *) private_p);
+  l8 = vload8(s, (const __global long *) global_p);
+  ui2 = vload2(s, (const __local uint *) local_p);
+  f16 = vload16(s, (const __private float *) private_p);
 }
 
 kernel void basic_work_item() {