[Bug tree-optimization/40770] Vectorization of complex types, vectorization of sincos missing

[kangshan0910 at hotmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=40770

--- Comment #21 from 康 珊 <kangshan0910 at hotmail dot com> ---
(In reply to Jakub Jelinek from comment #15)
> OpenMP has a way to express this,
> #pragma omp declare simd notinbranch linear(y,z)
> double sincos (double x, double *y, double *z);
> As can be seen on -O2 -fopenmp-simd:
> #pragma omp declare simd notinbranch linear(y, z)
> double foo (double x, double *y, double *z);
> 
> double a[1024], b[1024], c[1024];
> 
> void
> bar (void)
> {
>   #pragma omp simd
>   for (int i = 0; i < 1024; i++)
>     foo (a[i], &b[i], &c[i]);
> }
> it works fine with that.  But if #pragma omp simd is commented out, it is
> not, because the compiler has no assurance on what the function call
> behavior will be e.g. for data reference analysis (it could e.g. modify the
> global vars).
> Now, for sincos proper, the compiler knows what the function does if not
> -fno-builtin-sincos.
> 
> Unfortunately, I think glibc currently defines sincos in math-vector.h the
> unuseful way with just simd ("notinbranch") attribute, which effectively
> means
> that the caller is supposed to pass vectors of pointers (well integers with
> those sizes) and so needs scatter under the hood.
> So the best thing would be to change glibc first.

Thanks for your reply.(In reply to Jakub Jelinek from comment #15)
> OpenMP has a way to express this,
> #pragma omp declare simd notinbranch linear(y,z)
> double sincos (double x, double *y, double *z);
> As can be seen on -O2 -fopenmp-simd:
> #pragma omp declare simd notinbranch linear(y, z)
> double foo (double x, double *y, double *z);
> 
> double a[1024], b[1024], c[1024];
> 
> void
> bar (void)
> {
>   #pragma omp simd
>   for (int i = 0; i < 1024; i++)
>     foo (a[i], &b[i], &c[i]);
> }
> it works fine with that.  But if #pragma omp simd is commented out, it is
> not, because the compiler has no assurance on what the function call
> behavior will be e.g. for data reference analysis (it could e.g. modify the
> global vars).
> Now, for sincos proper, the compiler knows what the function does if not
> -fno-builtin-sincos.
> 
> Unfortunately, I think glibc currently defines sincos in math-vector.h the
> unuseful way with just simd ("notinbranch") attribute, which effectively
> means
> that the caller is supposed to pass vectors of pointers (well integers with
> those sizes) and so needs scatter under the hood.
> So the best thing would be to change glibc first.

Thanks for your reply. I tried your method. But it still doesn't work for
sincos. 

#pragma omp declare simd notinbranch linear(y, z)
void sincos(double x, double *y, double *z);

for (unsigned int i = 0; i < loopCount; i++)
    {
#pragma omp simd
        for (unsigned int j = i * dim; j < (i + 1) * dim; j++)
        {
            sincos(input_array[j], &result_array[j], &result_array1[j]);
        }
    }

$ gcc -m64 -g -O2 -fopenmp-simd -fPIC -Wall test_sincos.c -o test_sincos -lm
-march=skylake-avx512 -ffast-math -ftree-loop-vectorize
$ ldd test_sincos
        linux-vdso.so.1 (0x00007ffd2393f000)
        libm.so.6 => /lib64/libm.so.6 (0x00007f574f49d000)
        libc.so.6 => /lib64/libc.so.6 (0x00007f574f2a2000)
        /lib64/ld-linux-x86-64.so.2 (0x00007f574f601000)
My gcc version is "10.1.1 20200507".
And I have uploaded my test source file.