Merge pull request rust-lang#316 from tgross35/arch-module

Move arch-specific behavior and intrinsics to a separate module

Author: tgross35

Cargo.toml

@@ -14,7 +14,10 @@ exclude = ["/ci/", "/.github/workflows/"]
 rust-version = "1.63"
 
 [features]
-default = []
+default = ["arch"]
+
+# Enable architecture-specific features such as SIMD or assembly routines.
+arch = []
 
 # This tells the compiler to assume that a Nightly toolchain is being used and
 # that it should activate any useful Nightly things accordingly.

build.rs

@@ -15,6 +15,7 @@ fn main() {
     }
 
     configure_intrinsics();
+    configure_arch();
 }
 
 /// Simplify the feature logic for enabling intrinsics so code only needs to use
@@ -28,3 +29,14 @@ fn configure_intrinsics() {
         println!("cargo:rustc-cfg=intrinsics_enabled");
     }
 }
+
+/// Simplify the feature logic for enabling arch-specific features so code only needs to use
+/// `cfg(arch_enabled)`.
+fn configure_arch() {
+    println!("cargo:rustc-check-cfg=cfg(arch_enabled)");
+
+    // Enabled by default via the "arch" feature, `force-soft-floats` overrides to disable.
+    if cfg!(feature = "arch") && !cfg!(feature = "force-soft-floats") {
+        println!("cargo:rustc-cfg=arch_enabled");
+    }
+}

ci/run.sh

@@ -64,6 +64,7 @@ fi
 
 # Make sure we can build with overriding features. We test the indibidual
 # features it controls separately.
+cargo check --no-default-features
 cargo check --features "force-soft-floats"
 
 if [ "${BUILD_ONLY:-}" = "1" ]; then

crates/compiler-builtins-smoke-test/Cargo.toml

@@ -18,6 +18,7 @@ force-soft-floats = []
 
 [lints.rust]
 unexpected_cfgs = { level = "warn", check-cfg = [
+  "cfg(arch_enabled)",
   "cfg(assert_no_panic)",
   "cfg(intrinsics_enabled)",
 ] }

crates/libm-test/build.rs

@@ -156,7 +156,11 @@ mod musl_serialized_tests {
             return;
         }
 
-        let files = fs::read_dir(math_src).unwrap().map(|f| f.unwrap().path()).collect::<Vec<_>>();
+        let files = fs::read_dir(math_src)
+            .unwrap()
+            .map(|f| f.unwrap().path())
+            .filter(file_needs_test)
+            .collect::<Vec<_>>();
 
         let mut math = Vec::new();
         for file in files {
@@ -187,6 +191,19 @@ mod musl_serialized_tests {
         generate_unit_tests(&math);
     }
 
+    /// Check whether a path within `src/math` should get tests generated.
+    fn file_needs_test(path: &PathBuf) -> bool {
+        // Skip directories
+        if path.is_dir() {
+            return false;
+        }
+
+        let fname = path.file_name().unwrap().to_str().unwrap();
+
+        // Musl doesn't support `f16` or `f128`
+        !(fname.contains("f16") || fname.contains("f128"))
+    }
+
     /// A "poor man's" parser for the signature of a function
     fn parse(s: &str) -> Function {
         let s = eat(s, "pub fn ");

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;