aarch64: Add a note saying why we use frintx rather than frintn

tgross35 · tgross35 · commit fc34c3edad29 · 2025-05-28T21:10:28.000Z
diff --git a/libm/src/math/arch/aarch64.rs b/libm/src/math/arch/aarch64.rs
@@ -30,6 +30,12 @@ pub fn fmaf(mut x: f32, y: f32, z: f32) -> f32 {
     x
 }
 
+// NB: `frintx` is technically the correct instruction for C's `rint`. However, in Rust (and LLVM
+// by default), `rint` is identical to `roundeven` (no fpenv interaction) so we use the
+// side-effect-free `frintn`.
+//
+// In general, C code that calls Rust's libm should assume that fpenv is ignored.
+
 pub fn rint(mut x: f64) -> f64 {
     // SAFETY: `frintn` is available with neon and has no side effects.
     //

Original file line number	Diff line number	Diff line change
`@@ -30,6 +30,12 @@ pub fn fmaf(mut x: f32, y: f32, z: f32) -> f32 {`
`30`	`30`	`x`
`31`	`31`	`}`
`32`	`32`
	`33`	+// NB: `frintx` is technically the correct instruction for C's `rint`. However, in Rust (and LLVM
	`34`	+// by default), `rint` is identical to `roundeven` (no fpenv interaction) so we use the
	`35`	+// side-effect-free `frintn`.
	`36`	`+//`
	`37`	`+// In general, C code that calls Rust's libm should assume that fpenv is ignored.`
	`38`	`+`
`33`	`39`	`pub fn rint(mut x: f64) -> f64 {`
`34`	`40`	// SAFETY: `frintn` is available with neon and has no side effects.
`35`	`41`	`//`