Merge pull request rust-lang#152 from alexcrichton/wasm-sqrt

Optimize intrinsics on wasm32

Author: alexcrichton

.travis.yml

@@ -29,6 +29,13 @@ matrix:
     - env: TARGET=cargo-fmt
       rust: beta
 
+    - env: TARGET=wasm32-unknown-unknown
+      rust: nightly
+      install: rustup target add $TARGET
+      script:
+        - cargo build --target $TARGET
+        - cargo build --no-default-features --target $TARGET
+
 before_install: set -e
 
 install:

Cargo.toml

@@ -12,6 +12,8 @@ version = "0.1.2"
 [features]
 # only used to run our test suite
 checked = []
+default = ['stable']
+stable = []
 
 [workspace]
 members = [

src/lib.rs

@@ -11,6 +11,10 @@
 
 #![deny(warnings)]
 #![no_std]
+#![cfg_attr(
+    all(target_arch = "wasm32", not(feature = "stable")),
+    feature(core_intrinsics)
+)]
 
 mod math;
 

src/math/ceil.rs

@@ -4,6 +4,14 @@ const TOINT: f64 = 1. / f64::EPSILON;
 
 #[inline]
 pub fn ceil(x: f64) -> f64 {
+    // On wasm32 we know that LLVM's intrinsic will compile to an optimized
+    // `f64.ceil` native instruction, so we can leverage this for both code size
+    // and speed.
+    llvm_intrinsically_optimized! {
+        #[cfg(target_arch = "wasm32")] {
+            return unsafe { ::core::intrinsics::ceilf64(x) }
+        }
+    }
     let u: u64 = x.to_bits();
     let e: i64 = (u >> 52 & 0x7ff) as i64;
     let y: f64;

src/math/ceilf.rs

@@ -2,6 +2,14 @@ use core::f32;
 
 #[inline]
 pub fn ceilf(x: f32) -> f32 {
+    // On wasm32 we know that LLVM's intrinsic will compile to an optimized
+    // `f32.ceil` native instruction, so we can leverage this for both code size
+    // and speed.
+    llvm_intrinsically_optimized! {
+        #[cfg(target_arch = "wasm32")] {
+            return unsafe { ::core::intrinsics::ceilf32(x) }
+        }
+    }
     let mut ui = x.to_bits();
     let e = (((ui >> 23) & 0xff) - 0x7f) as i32;
 

src/math/fabs.rs

@@ -2,5 +2,13 @@ use core::u64;
 
 #[inline]
 pub fn fabs(x: f64) -> f64 {
+    // On wasm32 we know that LLVM's intrinsic will compile to an optimized
+    // `f64.abs` native instruction, so we can leverage this for both code size
+    // and speed.
+    llvm_intrinsically_optimized! {
+        #[cfg(target_arch = "wasm32")] {
+            return unsafe { ::core::intrinsics::fabsf64(x) }
+        }
+    }
     f64::from_bits(x.to_bits() & (u64::MAX / 2))
 }

src/math/fabsf.rs

@@ -1,4 +1,12 @@
 #[inline]
 pub fn fabsf(x: f32) -> f32 {
+    // On wasm32 we know that LLVM's intrinsic will compile to an optimized
+    // `f32.abs` native instruction, so we can leverage this for both code size
+    // and speed.
+    llvm_intrinsically_optimized! {
+        #[cfg(target_arch = "wasm32")] {
+            return unsafe { ::core::intrinsics::fabsf32(x) }
+        }
+    }
     f32::from_bits(x.to_bits() & 0x7fffffff)
 }

src/math/floor.rs

@@ -4,6 +4,14 @@ const TOINT: f64 = 1. / f64::EPSILON;
 
 #[inline]
 pub fn floor(x: f64) -> f64 {
+    // On wasm32 we know that LLVM's intrinsic will compile to an optimized
+    // `f64.floor` native instruction, so we can leverage this for both code size
+    // and speed.
+    llvm_intrinsically_optimized! {
+        #[cfg(target_arch = "wasm32")] {
+            return unsafe { ::core::intrinsics::floorf64(x) }
+        }
+    }
     let ui = x.to_bits();
     let e = ((ui >> 52) & 0x7ff) as i32;
 

src/math/floorf.rs

@@ -2,6 +2,14 @@ use core::f32;
 
 #[inline]
 pub fn floorf(x: f32) -> f32 {
+    // On wasm32 we know that LLVM's intrinsic will compile to an optimized
+    // `f32.floor` native instruction, so we can leverage this for both code size
+    // and speed.
+    llvm_intrinsically_optimized! {
+        #[cfg(target_arch = "wasm32")] {
+            return unsafe { ::core::intrinsics::floorf32(x) }
+        }
+    }
     let mut ui = x.to_bits();
     let e = (((ui >> 23) & 0xff) - 0x7f) as i32;
 

src/math/mod.rs

@@ -58,6 +58,17 @@ macro_rules! i {
     };
 }
 
+macro_rules! llvm_intrinsically_optimized {
+    (#[cfg($($clause:tt)*)] $e:expr) => {
+        #[cfg(all(not(feature = "stable"), $($clause)*))]
+        {
+            if true { // thwart the dead code lint
+                $e
+            }
+        }
+    };
+}
+
 // Public modules
 mod acos;
 mod acosf;

src/math/sqrt.rs

@@ -82,6 +82,18 @@ const TINY: f64 = 1.0e-300;
 
 #[inline]
 pub fn sqrt(x: f64) -> f64 {
+    // On wasm32 we know that LLVM's intrinsic will compile to an optimized
+    // `f64.sqrt` native instruction, so we can leverage this for both code size
+    // and speed.
+    llvm_intrinsically_optimized! {
+        #[cfg(target_arch = "wasm32")] {
+            return if x < 0.0 {
+                f64::NAN
+            } else {
+                unsafe { ::core::intrinsics::sqrtf64(x) }
+            }
+        }
+    }
     let mut z: f64;
     let sign: u32 = 0x80000000;
     let mut ix0: i32;

src/math/sqrtf.rs

@@ -17,6 +17,18 @@ const TINY: f32 = 1.0e-30;
 
 #[inline]
 pub fn sqrtf(x: f32) -> f32 {
+    // On wasm32 we know that LLVM's intrinsic will compile to an optimized
+    // `f32.sqrt` native instruction, so we can leverage this for both code size
+    // and speed.
+    llvm_intrinsically_optimized! {
+        #[cfg(target_arch = "wasm32")] {
+            return if x < 0.0 {
+                ::core::f32::NAN
+            } else {
+                unsafe { ::core::intrinsics::sqrtf32(x) }
+            }
+        }
+    }
     let mut z: f32;
     let sign: i32 = 0x80000000u32 as i32;
     let mut ix: i32;

src/math/trunc.rs

@@ -2,6 +2,14 @@ use core::f64;
 
 #[inline]
 pub fn trunc(x: f64) -> f64 {
+    // On wasm32 we know that LLVM's intrinsic will compile to an optimized
+    // `f64.trunc` native instruction, so we can leverage this for both code size
+    // and speed.
+    llvm_intrinsically_optimized! {
+        #[cfg(target_arch = "wasm32")] {
+            return unsafe { ::core::intrinsics::truncf64(x) }
+        }
+    }
     let x1p120 = f64::from_bits(0x4770000000000000); // 0x1p120f === 2 ^ 120
 
     let mut i: u64 = x.to_bits();

src/math/truncf.rs

@@ -2,6 +2,14 @@ use core::f32;
 
 #[inline]
 pub fn truncf(x: f32) -> f32 {
+    // On wasm32 we know that LLVM's intrinsic will compile to an optimized
+    // `f32.trunc` native instruction, so we can leverage this for both code size
+    // and speed.
+    llvm_intrinsically_optimized! {
+        #[cfg(target_arch = "wasm32")] {
+            return unsafe { ::core::intrinsics::truncf32(x) }
+        }
+    }
     let x1p120 = f32::from_bits(0x7b800000); // 0x1p120f === 2 ^ 120
 
     let mut i: u32 = x.to_bits();