Change f32::midpoint to upcast to f64

This has been verified by kani as a correct optimization see: #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. The new implementation is gated on targets that have a fast 64-bit floating point implementation in hardware, and on WASM.
rust-lang · jieyouxu · Feb 14, 2024 · Jun 2, 2024 · May 31, 2024 · May 31, 2024
commit 849c5254afdeb96a740cfac989a382fde69b6067
diff --git a/library/core/src/num/f32.rs b/library/core/src/num/f32.rs
@@ -1016,25 +1016,42 @@ 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(any(
+                    target_arch = "x86_64",
+                    target_arch = "aarch64",
+                    all(any(target_arch="riscv32", target_arch= "riscv64"), target_feature="d"),
+                    all(target_arch = "arm", target_feature="vfp2"),
+                    target_arch = "wasm32",
+                    target_arch = "wasm64",
+                ))] {
+                // whitelist the faster implementation to targets that have known good 64-bit float
+                // implementations. Falling back to the branchy code on targets that don't have
+                // 64-bit hardware floats or buggy implementations.
+                // see: https://github.com/rust-lang/rust/pull/121062#issuecomment-2123408114
+                ((f64::from(self) + f64::from(other)) / 2.0) as f32
+            } else {
+                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.)
+                }
+            }
         }
     }
 

diff --git a/library/core/tests/num/mod.rs b/library/core/tests/num/mod.rs
@@ -719,7 +719,7 @@ assume_usize_width! {
 }
 
 macro_rules! test_float {
-    ($modname: ident, $fty: ty, $inf: expr, $neginf: expr, $nan: expr, $min: expr, $max: expr, $min_pos: expr) => {
+    ($modname: ident, $fty: ty, $inf: expr, $neginf: expr, $nan: expr, $min: expr, $max: expr, $min_pos: expr, $max_exp:expr) => {
         mod $modname {
             #[test]
             fn min() {
@@ -870,6 +870,27 @@ macro_rules! test_float {
                 assert!(($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
+                // in particular, `2.pow(i)` will  never be at the max exponent, so it could
+                // be safely doubled, while j is significantly smaller.
+                for i in $max_exp.saturating_sub(64)..$max_exp {
+                    for j in 0..64u8 {
+                        let large = <$fty>::from(2.0f32).powi(i);
+                        // a much smaller number, such that there is no chance of overflow to test
+                        // potential double rounding in midpoint's implementation.
+                        let small = <$fty>::from(2.0f32).powi($max_exp - 1)
+                            * <$fty>::EPSILON
+                            * <$fty>::from(j);
+
+                        let naive = (large + small) / 2.0;
+                        let midpoint = large.midpoint(small);
+
+                        assert_eq!(naive, midpoint);
+                    }
+                }
             }
             #[test]
             fn rem_euclid() {
@@ -902,7 +923,8 @@ test_float!(
     f32::NAN,
     f32::MIN,
     f32::MAX,
-    f32::MIN_POSITIVE
+    f32::MIN_POSITIVE,
+    f32::MAX_EXP
 );
 test_float!(
     f64,
@@ -912,5 +934,6 @@ test_float!(
     f64::NAN,
     f64::MIN,
     f64::MAX,
-    f64::MIN_POSITIVE
+    f64::MIN_POSITIVE,
+    f64::MAX_EXP
 );