support hardware sqrt for f32 on arm32 targets

etc/function-definitions.json

@@ -985,6 +985,7 @@
     "sqrt": {
         "sources": [
             "libm/src/math/arch/aarch64/sqrt.rs",
+            "libm/src/math/arch/arm/sqrt.rs",
             "libm/src/math/arch/loongarch/sqrt.rs",
             "libm/src/math/arch/wasm32/sqrt.rs",
             "libm/src/math/arch/x86/sqrt.rs",
@@ -996,6 +997,7 @@
     "sqrtf": {
         "sources": [
             "libm/src/math/arch/aarch64/sqrt.rs",
+            "libm/src/math/arch/arm/sqrt.rs",
             "libm/src/math/arch/loongarch/sqrt.rs",
             "libm/src/math/arch/wasm32/sqrt.rs",
             "libm/src/math/arch/x86/sqrt.rs",

libm/configure.rs

@@ -19,6 +19,7 @@ pub struct Config {
     pub target_triple: String,
     pub target_triple_split: Vec<String>,
     pub target_arch: String,
+    pub target_abi: Option<String>,
     pub target_env: String,
     pub target_families: Vec<String>,
     pub target_os: String,
@@ -63,6 +64,8 @@ impl Config {
             opt_level: env::var("OPT_LEVEL").unwrap(),
             cargo_features,
             target_arch: env::var("CARGO_CFG_TARGET_ARCH").unwrap(),
+            // FIXME(msrv): defined since 1.78.
+            target_abi: env::var("CARGO_CFG_TARGET_ABI").ok(),
             target_env: env::var("CARGO_CFG_TARGET_ENV").unwrap(),
             target_families,
             target_os: env::var("CARGO_CFG_TARGET_OS").unwrap(),
@@ -114,6 +117,9 @@ pub fn emit(cfg: &Config) {
     // We have to cfg our code accordingly.
     let thumb_1 = split[0] == "thumbv6m" || split[0] == "thumbv8m.base";
 
+    // FIXME(msrv): `cfg(target_abi)` is usable since 1.78.
+    let target_abi_eabihf = cfg.target_abi.as_deref() == Some("eabihf");
+
     // Shorthand to detect i586 targets
     let x86_no_sse2 = cfg.target_arch == "x86" && !cfg.target_features.iter().any(|f| f == "sse2");
 
@@ -123,7 +129,10 @@ pub fn emit(cfg: &Config) {
     // Arch shorthand config is used in most crates.
     set_cfg("thumb", thumb);
     set_cfg("thumb_1", thumb_1);
+    set_cfg("target_abi_eabihf", target_abi_eabihf);
     set_cfg("x86_no_sse2", x86_no_sse2);
+    // FIXME(arm_target_feature): the exact target features are not yet stable.
+    println!("cargo:rustc-check-cfg=cfg(target_feature, values(\"vfp2sp\", \"vfp2\"))");
 
     match cfg.library {
         Library::CompilerBuiltins => {

libm/src/math/arch/arm/mod.rs

@@ -0,0 +1,8 @@
+//! Architecture-specific support for Arm32/Thumb with hard float ABI.
+
+mod sqrt;
+
+#[cfg(target_feature = "vfp2")]
+pub use sqrt::sqrt;
+#[cfg(target_feature = "vfp2sp")]
+pub use sqrt::sqrtf;

libm/src/math/arch/arm/sqrt.rs

@@ -0,0 +1,29 @@
+#[cfg(target_feature = "vfp2sp")]
+pub fn sqrtf(mut x: f32) -> f32 {
+    // SAFETY: vsqrt.f32 is available with `vfp2sp`.
+    unsafe {
+        core::arch::asm!(
+            "vsqrt.f32 {x}, {x}",
+            x = inout(sreg) x,
+            options(nostack, nomem, pure),
+        );
+    }
+    x
+}
+
+#[cfg(target_feature = "vfp2")]
+pub fn sqrt(mut x: f64) -> f64 {
+    // SAFETY: vsqrt.f64 is available with `vfp2`.
+    // Can't just use plain old `dreg` because `vfp2` only provides a base
+    // level of FPU support with a non-specified amount of fp64 registers.
+    // It's fine to use `dreg_low16` here, but use `dreg_low8` instead to
+    // keep the requirements as minimal as possible.
+    unsafe {
+        core::arch::asm!(
+            "vsqrt.f64 {x}, {x}",
+            x = inout(dreg_low8) x,
+            options(nostack, nomem, pure),
+        );
+    }
+    x
+}

libm/src/math/arch/mod.rs

@@ -41,6 +41,14 @@ cfg_select_nofmt! {
             sqrtf16,
         };
     }
+    all(target_arch = "arm", target_abi_eabihf) => {
+        mod arm;
+
+        #[cfg(target_feature = "vfp2sp")]
+        pub use arm::sqrtf;
+        #[cfg(target_feature = "vfp2")]
+        pub use arm::sqrt;
+    }
     any(target_arch = "loongarch32", target_arch = "loongarch64") => {
         mod loongarch;
 

libm/src/math/sqrt.rs

@@ -21,6 +21,7 @@ pub fn sqrtf(x: f32) -> f32 {
         name: sqrtf,
         use_arch: any(
             all(target_arch = "aarch64", target_feature = "neon"),
+            all(target_arch = "arm", target_abi_eabihf, target_feature = "vfp2sp"),
             all(target_arch = "wasm32", intrinsics_enabled),
             all(any(target_arch = "loongarch32", target_arch = "loongarch64"), target_feature = "f"),
             target_feature = "sse2"
@@ -38,6 +39,7 @@ pub fn sqrt(x: f64) -> f64 {
         name: sqrt,
         use_arch: any(
             all(target_arch = "aarch64", target_feature = "neon"),
+            all(target_arch = "arm", target_abi_eabihf, target_feature = "vfp2"),
             all(target_arch = "wasm32", intrinsics_enabled),
             all(any(target_arch = "loongarch32", target_arch = "loongarch64"), target_feature = "d"),
             target_feature = "sse2"