Calls to sin() don't get converted to sinf, leading to inefficient use of vector variants with -fveclib=ArmPL

[willlovett-arm]

https://godbolt.org/z/fE16nGPxP

#include <math.h>

void foo(float *p)
{
for (int x=0; x<999; ++x) {
    p[x] = sin(p[x]);
  }
}

generates calls to armpl_vsinq_f64 instead of armpl_vsinq_f32.

If we replace that code with sinf it works fine:

#include <math.h>

void foo(float *p)
{
for (int x=0; x<999; ++x) {
    p[x] = sinf(p[x]);
  }
}

I'm reasonably sure it's legitimate to implicitly switch to sinf(), and this would lead to double the vector throughput.

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Will Lovett (willlovett-arm)

https://godbolt.org/z/fE16nGPxP

#include <math.h>

void foo(float *p)
{
for (int x=0; x<999; ++x) {
    p[x] = sin(p[x]);
  }
}

generates calls to armpl_vsinq_f64 instead of armpl_vsinq_f32.

If we replace that code with sinf it works fine:

#include <math.h>

void foo(float *p)
{
for (int x=0; x<999; ++x) {
    p[x] = sinf(p[x]);
  }
}

I'm reasonably sure it's legitimate to implicitly switch to sinf(), and this would lead to double the vector throughput.

[davemgreen]

We do this if the function is a @sin, but not for @llvm.sin intrinsics.
https://godbolt.org/z/MTKefx5Yr

[fhahn]

Should be fixed by #139082

[willlovett-arm]

Amazing turnaround! Thanks @guy-david , all 🙏