Move architecture-specific code to src/math/arch

Move the code and call into its new location with
`select_implementation`.

src/math/arch/i586.rs

@@ -0,0 +1,37 @@
+//! Architecture-specific support for x86-32 without SSE2
+
+use super::super::fabs;
+
+/// Use an alternative implementation on x86, because the
+/// main implementation fails with the x87 FPU used by
+/// debian i386, probably due to excess precision issues.
+/// Basic implementation taken from https://github.com/rust-lang/libm/issues/219.
+pub fn ceil(x: f64) -> f64 {
+    if fabs(x).to_bits() < 4503599627370496.0_f64.to_bits() {
+        let truncated = x as i64 as f64;
+        if truncated < x {
+            return truncated + 1.0;
+        } else {
+            return truncated;
+        }
+    } else {
+        return x;
+    }
+}
+
+/// Use an alternative implementation on x86, because the
+/// main implementation fails with the x87 FPU used by
+/// debian i386, probably due to excess precision issues.
+/// Basic implementation taken from https://github.com/rust-lang/libm/issues/219.
+pub fn floor(x: f64) -> f64 {
+    if fabs(x).to_bits() < 4503599627370496.0_f64.to_bits() {
+        let truncated = x as i64 as f64;
+        if truncated > x {
+            return truncated - 1.0;
+        } else {
+            return truncated;
+        }
+    } else {
+        return x;
+    }
+}

src/math/arch/i686.rs

@@ -0,0 +1,24 @@
+//! Architecture-specific support for x86-32 and x86-64 with SSE2
+
+#![cfg(not(feature = "force-soft-floats"))]
+
+#[cfg(target_arch = "x86")]
+use core::arch::x86::*;
+#[cfg(target_arch = "x86_64")]
+use core::arch::x86_64::*;
+
+pub fn sqrtf(x: f32) -> f32 {
+    unsafe {
+        let m = _mm_set_ss(x);
+        let m_sqrt = _mm_sqrt_ss(m);
+        _mm_cvtss_f32(m_sqrt)
+    }
+}
+
+pub fn sqrt(x: f64) -> f64 {
+    unsafe {
+        let m = _mm_set_sd(x);
+        let m_sqrt = _mm_sqrt_pd(m);
+        _mm_cvtsd_f64(m_sqrt)
+    }
+}

src/math/arch/mod.rs

@@ -7,3 +7,22 @@
 
 #[cfg(intrinsics_enabled)]
 pub mod intrinsics;
+
+// Most implementations should be defined here, to ensure they are not made available when
+// soft floats are required.
+#[cfg(arch_enabled)]
+cfg_if! {
+    if #[cfg(target_feature = "sse2")] {
+        mod i686;
+        pub use i686::{sqrt, sqrtf};
+    }
+}
+
+// There are certain architecture-specific implementations that are needed for correctness
+// even with `force-soft-float`. These are configured here.
+cfg_if! {
+    if #[cfg(all(target_arch = "x86", not(target_feature = "sse2")))] {
+        mod i586;
+        pub use i586::{ceil, floor};
+    }
+}

src/math/ceil.rs

@@ -10,28 +10,11 @@ const TOINT: f64 = 1. / f64::EPSILON;
 pub fn ceil(x: f64) -> f64 {
     select_implementation! {
         name: ceil,
+        use_arch_required: all(target_arch = "x86", not(target_feature = "sse2")),
         use_intrinsic: target_arch = "wasm32",
         args: x,
     }
 
-    #[cfg(all(target_arch = "x86", not(target_feature = "sse2")))]
-    {
-        //use an alternative implementation on x86, because the
-        //main implementation fails with the x87 FPU used by
-        //debian i386, probably due to excess precision issues.
-        //basic implementation taken from https://github.com/rust-lang/libm/issues/219
-        use super::fabs;
-        if fabs(x).to_bits() < 4503599627370496.0_f64.to_bits() {
-            let truncated = x as i64 as f64;
-            if truncated < x {
-                return truncated + 1.0;
-            } else {
-                return truncated;
-            }
-        } else {
-            return x;
-        }
-    }
     let u: u64 = x.to_bits();
     let e: i64 = (u >> 52 & 0x7ff) as i64;
     let y: f64;

src/math/floor.rs

@@ -10,28 +10,11 @@ const TOINT: f64 = 1. / f64::EPSILON;
 pub fn floor(x: f64) -> f64 {
     select_implementation! {
         name: floor,
+        use_arch_required: all(target_arch = "x86", not(target_feature = "sse2")),
         use_intrinsic: target_arch = "wasm32",
         args: x,
     }
 
-    #[cfg(all(target_arch = "x86", not(target_feature = "sse2")))]
-    {
-        //use an alternative implementation on x86, because the
-        //main implementation fails with the x87 FPU used by
-        //debian i386, probably due to excess precision issues.
-        //basic implementation taken from https://github.com/rust-lang/libm/issues/219
-        use super::fabs;
-        if fabs(x).to_bits() < 4503599627370496.0_f64.to_bits() {
-            let truncated = x as i64 as f64;
-            if truncated > x {
-                return truncated - 1.0;
-            } else {
-                return truncated;
-            }
-        } else {
-            return x;
-        }
-    }
     let ui = x.to_bits();
     let e = ((ui >> 52) & 0x7ff) as i32;