Auto merge of rust-lang#121062 - RustyYato:f32-midpoint, r=the8472

Change f32::midpoint to upcast to f64

This has been verified by kani as a correct optimization

see: rust-lang#110840 (comment)

The new implementation is branchless and only differs in which NaN values are produced (if any are produced at all), which is fine to change. Aside from NaN handling, this implementation produces bitwise identical results to the original implementation.

Question: do we need a codegen test for this? I didn't add one, since the original PR rust-lang#92048 didn't have any codegen tests.

Author: bors

library/core/src/num/f32.rs

@@ -1030,25 +1030,36 @@ impl f32 {
     /// ```
     #[unstable(feature = "num_midpoint", issue = "110840")]
     pub fn midpoint(self, other: f32) -> f32 {
-        const LO: f32 = f32::MIN_POSITIVE * 2.;
-        const HI: f32 = f32::MAX / 2.;
-
-        let (a, b) = (self, other);
-        let abs_a = a.abs_private();
-        let abs_b = b.abs_private();
-
-        if abs_a <= HI && abs_b <= HI {
-            // Overflow is impossible
-            (a + b) / 2.
-        } else if abs_a < LO {
-            // Not safe to halve a
-            a + (b / 2.)
-        } else if abs_b < LO {
-            // Not safe to halve b
-            (a / 2.) + b
-        } else {
-            // Not safe to halve a and b
-            (a / 2.) + (b / 2.)
+        cfg_if! {
+            if #[cfg(all(target_arch = "arm", target_pointer_width = "32",
+                         not(target_feature = "vfp2")))] {
+                // some 32-bit ARM architectures don't have native double-precision floats
+                // so fall back to a similar algorithm as in f64, but using f32
+                // This should only differ in the specific NaNs reported.
+
+                const LO: f32 = f32::MIN_POSITIVE * 2.;
+                const HI: f32 = f32::MAX / 2.;
+
+                let (a, b) = (self, other);
+                let abs_a = a.abs_private();
+                let abs_b = b.abs_private();
+
+                if abs_a <= HI && abs_b <= HI {
+                    // Overflow is impossible
+                    (a + b) / 2.
+                } else if abs_a < LO {
+                    // Not safe to halve a
+                    a + (b / 2.)
+                } else if abs_b < LO {
+                    // Not safe to halve b
+                    (a / 2.) + b
+                } else {
+                    // Not safe to halve a and b
+                    (a / 2.) + (b / 2.)
+                }
+            } else {
+                ((f64::from(self) + f64::from(other)) / 2.0) as f32
+            }
         }
     }
 

library/core/tests/num/mod.rs

@@ -880,6 +880,17 @@ macro_rules! test_float {
                 assert!(($nan as $fty).midpoint(1.0).is_nan());
                 assert!((1.0 as $fty).midpoint($nan).is_nan());
                 assert!(($nan as $fty).midpoint($nan).is_nan());
+
+                // test if large differences in magnitude are still correctly computed.
+                // NOTE: that because of how small x and y are, x + y can never overflow
+                // so (x + y) / 2.0 is always correct
+                for i in 64..128 {
+                    for j in 0..64u8 {
+                        let x = (2.0f32.powi(i) + f32::from(j)) / 2.0;
+                        let y = 2.0f32.powi(i).midpoint(f32::from(j));
+                        assert_eq!(x, y);
+                    }
+                }
             }
             #[test]
             fn rem_euclid() {