[cfe-dev] scalar operations to vectors and implicit conversions

Petar Jovanovic via cfe-dev cfe-dev at lists.llvm.org
Tue May 24 08:27:28 PDT 2016


Hi All,

Clang currently objects when scalar operations are applied to vectors (vector_size), while GCC does not.

Here is an example for X86:

#include <stdio.h>
#include <xmmintrin.h>

int main(int argc, char* argv[])
{
__m128 vec_a = _mm_set_ps(0, 1, 2, 3);
__m128 vec_b = _mm_set_ps(1, 2, 3, 4);
__m128 vec_r;

vec_a = vec_a + 5;
vec_r = _mm_add_ps(vec_a, vec_b);

printf("result = { %2.1f, %2.1f, %2.1f, %2.1f }\n", vec_r[0], vec_r[1], vec_r[2], vec_r[3]);

return 0;
}

Building this example with Clang will raise a build error:

$ clang -msse x86-simd.c -o x86-simd.exe
x86-simd.c:10:15: error: cannot convert between vector values of different size ('__m128' (vector of 4 'float' values) and 'int')
vec_a = vec_a + 5;
        ~~~~~ ^ ~
1 error generated.


Would anyone object to taking a vector x scalar operation to mean that the scalar should be replicated to the same width as the vector, same way GCC does? AFAICS, this is already the case with OpenCl's ext_vector_type.

This probably could be fixed in Sema::CheckVectorOperands() along with the changes in the test examples (test/Sema/zvector.c and possibly other). How should "-fno-lax-vector-conversions" interfere in those case?

Please share your opinion. Thanks.

Regards,
Petar
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20160524/edb1616a/attachment.html>


More information about the cfe-dev mailing list