Merge pull request #427 from tgross35/generic-trunc

Add `truncf16` and `truncf128`

Author: tgross35

libm/crates/compiler-builtins-smoke-test/src/lib.rs

@@ -157,6 +157,8 @@ no_mangle! {
     tgammaf(x: f32) -> f32;
     trunc(x: f64) -> f64;
     truncf(x: f32) -> f32;
+    truncf128(x: f128) -> f128;
+    truncf16(x: f16) -> f16;
     y0(x: f64) -> f64;
     y0f(x: f32) -> f32;
     y1(x: f64) -> f64;

libm/crates/libm-macros/src/shared.rs

@@ -9,7 +9,7 @@ const ALL_OPERATIONS_NESTED: &[(FloatTy, Signature, Option<Signature>, &[&str])]
         FloatTy::F16,
         Signature { args: &[Ty::F16], returns: &[Ty::F16] },
         None,
-        &["fabsf16"],
+        &["fabsf16", "truncf16"],
     ),
     (
         // `fn(f32) -> f32`
@@ -40,7 +40,7 @@ const ALL_OPERATIONS_NESTED: &[(FloatTy, Signature, Option<Signature>, &[&str])]
         FloatTy::F128,
         Signature { args: &[Ty::F128], returns: &[Ty::F128] },
         None,
-        &["fabsf128"],
+        &["fabsf128", "truncf128"],
     ),
     (
         // `(f16, f16) -> f16`

libm/crates/libm-test/benches/random.rs

@@ -117,7 +117,7 @@ libm_macros::for_each_function! {
         exp10 | exp10f | exp2 | exp2f => (true, Some(musl_math_sys::MACRO_FN_NAME)),
 
         // Musl does not provide `f16` and `f128` functions
-        copysignf16 | copysignf128 | fabsf16 | fabsf128 => (false, None),
+        copysignf16 | copysignf128 | fabsf16 | fabsf128 | truncf16 | truncf128  => (false, None),
 
         // By default we never skip (false) and always have a musl function available
         _ => (false, Some(musl_math_sys::MACRO_FN_NAME))

libm/crates/libm-test/src/domain.rs

@@ -199,3 +199,13 @@ impl HasDomain<f16> for crate::op::fabsf16::Routine {
 impl HasDomain<f128> for crate::op::fabsf128::Routine {
     const DOMAIN: Domain<f128> = Domain::<f128>::UNBOUNDED;
 }
+
+#[cfg(f16_enabled)]
+impl HasDomain<f16> for crate::op::truncf16::Routine {
+    const DOMAIN: Domain<f16> = Domain::<f16>::UNBOUNDED;
+}
+
+#[cfg(f128_enabled)]
+impl HasDomain<f128> for crate::op::truncf128::Routine {
+    const DOMAIN: Domain<f128> = Domain::<f128>::UNBOUNDED;
+}

libm/crates/libm-test/src/mpfloat.rs

@@ -141,6 +141,7 @@ libm_macros::for_each_function! {
         lgamma_r, lgammaf_r, modf, modff, nextafter, nextafterf, pow,powf,
         remquo, remquof, scalbn, scalbnf, sincos, sincosf, yn, ynf,
         copysignf16, copysignf128, fabsf16, fabsf128,
+        truncf16, truncf128,
     ],
     fn_extra: match MACRO_FN_NAME {
         // Remap function names that are different between mpfr and libm
@@ -202,11 +203,13 @@ impl_no_round! {
 #[cfg(f16_enabled)]
 impl_no_round! {
     fabsf16 => abs_mut;
+    truncf16 => trunc_mut;
 }
 
 #[cfg(f128_enabled)]
 impl_no_round! {
     fabsf128 => abs_mut;
+    truncf128 => trunc_mut;
 }
 
 /// Some functions are difficult to do in a generic way. Implement them here.

libm/crates/libm-test/tests/compare_built_musl.rs

@@ -48,7 +48,7 @@ where
 libm_macros::for_each_function! {
     callback: musl_rand_tests,
     // Musl does not support `f16` and `f128` on all platforms.
-    skip: [copysignf16, copysignf128, fabsf16, fabsf128],
+    skip: [copysignf16, copysignf128, fabsf16, fabsf128, truncf16, truncf128],
     attributes: [
         #[cfg_attr(x86_no_sse, ignore)] // FIXME(correctness): wrong result on i586
         [exp10, exp10f, exp2, exp2f, rint]
@@ -146,5 +146,7 @@ libm_macros::for_each_function! {
         // Not provided by musl
         fabsf16,
         fabsf128,
+        truncf16,
+        truncf128,
     ],
 }

libm/crates/util/src/main.rs

@@ -84,7 +84,7 @@ fn do_eval(basis: &str, op: &str, inputs: &[&str]) {
         emit_types: [CFn, RustFn, RustArgs],
         extra: (basis, op, inputs),
         fn_extra: match MACRO_FN_NAME {
-            copysignf16 | copysignf128 | fabsf16 | fabsf128 => None,
+            copysignf16 | copysignf128 | fabsf16 | fabsf128 | truncf16 | truncf128  => None,
             _ => Some(musl_math_sys::MACRO_FN_NAME)
         }
     }

libm/etc/function-definitions.json

@@ -743,17 +743,33 @@
         "sources": [
             "src/libm_helper.rs",
             "src/math/arch/wasm32.rs",
+            "src/math/generic/trunc.rs",
             "src/math/trunc.rs"
         ],
         "type": "f64"
     },
     "truncf": {
         "sources": [
             "src/math/arch/wasm32.rs",
+            "src/math/generic/trunc.rs",
             "src/math/truncf.rs"
         ],
         "type": "f32"
     },
+    "truncf128": {
+        "sources": [
+            "src/math/generic/trunc.rs",
+            "src/math/truncf128.rs"
+        ],
+        "type": "f128"
+    },
+    "truncf16": {
+        "sources": [
+            "src/math/generic/trunc.rs",
+            "src/math/truncf16.rs"
+        ],
+        "type": "f16"
+    },
     "y0": {
         "sources": [
             "src/libm_helper.rs",

libm/etc/function-list.txt

@@ -111,6 +111,8 @@ tgamma
 tgammaf
 trunc
 truncf
+truncf128
+truncf16
 y0
 y0f
 y1

libm/src/math/generic/mod.rs

@@ -1,5 +1,7 @@
 mod copysign;
 mod fabs;
+mod trunc;
 
 pub use copysign::copysign;
 pub use fabs::fabs;
+pub use trunc::trunc;

libm/src/math/generic/trunc.rs

@@ -0,0 +1,57 @@
+/* SPDX-License-Identifier: MIT
+ * origin: musl src/math/trunc.c */
+
+use super::super::{Float, Int, IntTy, MinInt};
+
+pub fn trunc<F: Float>(x: F) -> F {
+    let mut xi: F::Int = x.to_bits();
+    let e: i32 = x.exp_unbiased();
+
+    // C1: The represented value has no fractional part, so no truncation is needed
+    if e >= F::SIG_BITS as i32 {
+        return x;
+    }
+
+    let mask = if e < 0 {
+        // C2: If the exponent is negative, the result will be zero so we mask out everything
+        // except the sign.
+        F::SIGN_MASK
+    } else {
+        // C3: Otherwise, we mask out the last `e` bits of the significand.
+        !(F::SIG_MASK >> e.unsigned())
+    };
+
+    // C4: If the to-be-masked-out portion is already zero, we have an exact result
+    if (xi & !mask) == IntTy::<F>::ZERO {
+        return x;
+    }
+
+    // C5: Otherwise the result is inexact and we will truncate. Raise `FE_INEXACT`, mask the
+    // result, and return.
+    force_eval!(x + F::MAX);
+    xi &= mask;
+    F::from_bits(xi)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn sanity_check() {
+        assert_biteq!(trunc(1.1f32), 1.0);
+        assert_biteq!(trunc(1.1f64), 1.0);
+
+        // C1
+        assert_biteq!(trunc(hf32!("0x1p23")), hf32!("0x1p23"));
+        assert_biteq!(trunc(hf64!("0x1p52")), hf64!("0x1p52"));
+        assert_biteq!(trunc(hf32!("-0x1p23")), hf32!("-0x1p23"));
+        assert_biteq!(trunc(hf64!("-0x1p52")), hf64!("-0x1p52"));
+
+        // C2
+        assert_biteq!(trunc(hf32!("0x1p-1")), 0.0);
+        assert_biteq!(trunc(hf64!("0x1p-1")), 0.0);
+        assert_biteq!(trunc(hf32!("-0x1p-1")), -0.0);
+        assert_biteq!(trunc(hf64!("-0x1p-1")), -0.0);
+    }
+}

libm/src/math/mod.rs

@@ -121,7 +121,7 @@ use self::rem_pio2::rem_pio2;
 use self::rem_pio2_large::rem_pio2_large;
 use self::rem_pio2f::rem_pio2f;
 #[allow(unused_imports)]
-use self::support::{CastFrom, CastInto, DInt, Float, HInt, Int, MinInt};
+use self::support::{CastFrom, CastInto, DInt, Float, HInt, Int, IntTy, MinInt};
 
 // Public modules
 mod acos;
@@ -343,19 +343,23 @@ cfg_if! {
     if #[cfg(f16_enabled)] {
         mod copysignf16;
         mod fabsf16;
+        mod truncf16;
 
         pub use self::copysignf16::copysignf16;
         pub use self::fabsf16::fabsf16;
+        pub use self::truncf16::truncf16;
     }
 }
 
 cfg_if! {
     if #[cfg(f128_enabled)] {
         mod copysignf128;
         mod fabsf128;
+        mod truncf128;
 
         pub use self::copysignf128::copysignf128;
         pub use self::fabsf128::fabsf128;
+        pub use self::truncf128::truncf128;
     }
 }
 

libm/src/math/trunc.rs

@@ -1,5 +1,3 @@
-use core::f64;
-
 /// Rounds the number toward 0 to the closest integral value (f64).
 ///
 /// This effectively removes the decimal part of the number, leaving the integral part.
@@ -11,31 +9,5 @@ pub fn trunc(x: f64) -> f64 {
         args: x,
     }
 
-    let x1p120 = f64::from_bits(0x4770000000000000); // 0x1p120f === 2 ^ 120
-
-    let mut i: u64 = x.to_bits();
-    let mut e: i64 = ((i >> 52) & 0x7ff) as i64 - 0x3ff + 12;
-    let m: u64;
-
-    if e >= 52 + 12 {
-        return x;
-    }
-    if e < 12 {
-        e = 1;
-    }
-    m = -1i64 as u64 >> e;
-    if (i & m) == 0 {
-        return x;
-    }
-    force_eval!(x + x1p120);
-    i &= !m;
-    f64::from_bits(i)
-}
-
-#[cfg(test)]
-mod tests {
-    #[test]
-    fn sanity_check() {
-        assert_eq!(super::trunc(1.1), 1.0);
-    }
+    super::generic::trunc(x)
 }

libm/src/math/truncf.rs

@@ -1,5 +1,3 @@
-use core::f32;
-
 /// Rounds the number toward 0 to the closest integral value (f32).
 ///
 /// This effectively removes the decimal part of the number, leaving the integral part.
@@ -11,25 +9,7 @@ pub fn truncf(x: f32) -> f32 {
         args: x,
     }
 
-    let x1p120 = f32::from_bits(0x7b800000); // 0x1p120f === 2 ^ 120
-
-    let mut i: u32 = x.to_bits();
-    let mut e: i32 = ((i >> 23) & 0xff) as i32 - 0x7f + 9;
-    let m: u32;
-
-    if e >= 23 + 9 {
-        return x;
-    }
-    if e < 9 {
-        e = 1;
-    }
-    m = -1i32 as u32 >> e;
-    if (i & m) == 0 {
-        return x;
-    }
-    force_eval!(x + x1p120);
-    i &= !m;
-    f32::from_bits(i)
+    super::generic::trunc(x)
 }
 
 // PowerPC tests are failing on LLVM 13: https://github.com/rust-lang/rust/issues/88520

libm/src/math/truncf128.rs

@@ -0,0 +1,7 @@
+/// Rounds the number toward 0 to the closest integral value (f128).
+///
+/// This effectively removes the decimal part of the number, leaving the integral part.
+#[cfg_attr(all(test, assert_no_panic), no_panic::no_panic)]
+pub fn truncf128(x: f128) -> f128 {
+    super::generic::trunc(x)
+}

libm/src/math/truncf16.rs

@@ -0,0 +1,7 @@
+/// Rounds the number toward 0 to the closest integral value (f16).
+///
+/// This effectively removes the decimal part of the number, leaving the integral part.
+#[cfg_attr(all(test, assert_no_panic), no_panic::no_panic)]
+pub fn truncf16(x: f16) -> f16 {
+    super::generic::trunc(x)
+}