exp10: Add cr_exp10 from CORE-MATH

This was our function with the worst precision. Replace our current
implementation with the CORE-MATH correctly rounded version.

Source: https://gitlab.inria.fr/core-math/core-math/-/blob/93d9f3bab7561cfb62f746f7e70c0888bb5c9a00/src/binary64/exp10/exp10.c

Author: tgross35

libm-test/src/f8_impl.rs

@@ -90,6 +90,10 @@ impl Float for f8 {
         libm::generic::copysign(self, other)
     }
 
+    fn roundeven(self) -> Self {
+        libm::generic::rint_status(self).val
+    }
+
     fn fma(self, _y: Self, _z: Self) -> Self {
         unimplemented!()
     }

libm-test/src/precision.rs

@@ -92,7 +92,8 @@ pub fn default_ulp(ctx: &CheckCtx) -> Option<u32> {
         Bn::Erf => 1,
         Bn::Erfc => 4,
         Bn::Exp => 1,
-        Bn::Exp10 => 6,
+        Bn::Exp10 if ctx.fn_ident == Id::Exp10f => 6,
+        Bn::Exp10 => 0,
         Bn::Exp2 => 1,
         Bn::Expm1 => 1,
         Bn::Hypot => 1,

libm/src/math/approx/exp10f64.rs

@@ -9,12 +9,6 @@ const P10: &[f64] = &[
 /// Calculates 10 raised to the power of `x` (f64).
 #[cfg_attr(assert_no_panic, no_panic::no_panic)]
 pub fn exp10f64(x: f64) -> f64 {
-    select_implementation! {
-        name: x87_exp10,
-        use_arch_required: x86_no_sse2,
-        args: x,
-    }
-
     let (mut y, n) = modf(x);
     let u: u64 = n.to_bits();
     /* fabs(n) < 16 without raising invalid on nan */