Support for tanhf on aarch64 with a vectorized SVE implementation

[hiraditya]

Cloning openxla/xla#24201 to bring it up with ARM folks.

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: AdityaK (hiraditya)

Cloning https://github.com/openxla/xla/issues/24201 to bring it up with ARM folks.

[efriedma-quic]

If you've configured an appropriate vector library, the LLVM loop vectorizer supports vectorizing tanhf using <vscale x 4 x float> libcalls, which would be 8 elements on a Neoverse-V1.

Not sure what else you're asking for.

[hiraditya]

can you share a repro or existing testcase somewhere? for example this doesn't vectorize: #145763

#include <math.h>
#include <stdio.h>
void foo(float*);

int main(void)
{
    const int N = 1024;
    float arr[N], trr[N];
    for (int i = 0; i < N; ++i) {
        arr[i] = rand() % 100;
    }

    for (int i = 0; i < N; ++i) {
        trr[i] = tanhf(arr[i]);
    }
    foo(trr);
}

[efriedma-quic]

https://godbolt.org/z/GYqxKdnsv

[hiraditya]

Thanks! i'll communicate it further. closing this.