Auto merge of #127013 - tgross35:f16-format-parse, r=Mark-Simulacrum

Add `f16` formatting and parsing

Use the same algorithms as for `f32` and `f64` to implement `f16` parsing and printing.

try-job: x86_64-gnu-aux

Author: bors

library/core/src/fmt/float.rs

@@ -20,6 +20,8 @@ macro_rules! impl_general_format {
     }
 }
 
+#[cfg(target_has_reliable_f16)]
+impl_general_format! { f16 }
 impl_general_format! { f32 f64 }
 
 // Don't inline this so callers don't use the stack space this function
@@ -231,6 +233,13 @@ macro_rules! floating {
 
 floating! { f32 f64 }
 
+#[cfg(target_has_reliable_f16)]
+floating! { f16 }
+
+// FIXME(f16_f128): A fallback is used when the backend+target does not support f16 well, in order
+// to avoid ICEs.
+
+#[cfg(not(target_has_reliable_f16))]
 #[stable(feature = "rust1", since = "1.0.0")]
 impl Debug for f16 {
     #[inline]
@@ -239,6 +248,33 @@ impl Debug for f16 {
     }
 }
 
+#[cfg(not(target_has_reliable_f16))]
+#[stable(feature = "rust1", since = "1.0.0")]
+impl Display for f16 {
+    #[inline]
+    fn fmt(&self, fmt: &mut Formatter<'_>) -> Result {
+        Debug::fmt(self, fmt)
+    }
+}
+
+#[cfg(not(target_has_reliable_f16))]
+#[stable(feature = "rust1", since = "1.0.0")]
+impl LowerExp for f16 {
+    #[inline]
+    fn fmt(&self, fmt: &mut Formatter<'_>) -> Result {
+        Debug::fmt(self, fmt)
+    }
+}
+
+#[cfg(not(target_has_reliable_f16))]
+#[stable(feature = "rust1", since = "1.0.0")]
+impl UpperExp for f16 {
+    #[inline]
+    fn fmt(&self, fmt: &mut Formatter<'_>) -> Result {
+        Debug::fmt(self, fmt)
+    }
+}
+
 #[stable(feature = "rust1", since = "1.0.0")]
 impl Debug for f128 {
     #[inline]

library/core/src/lib.rs

@@ -101,6 +101,7 @@
 #![feature(bstr)]
 #![feature(bstr_internals)]
 #![feature(cfg_match)]
+#![feature(cfg_target_has_reliable_f16_f128)]
 #![feature(const_carrying_mul_add)]
 #![feature(const_eval_select)]
 #![feature(core_intrinsics)]

library/core/src/num/dec2flt/float.rs

@@ -45,7 +45,7 @@ macro_rules! int {
     }
 }
 
-int!(u32, u64);
+int!(u16, u32, u64);
 
 /// A helper trait to avoid duplicating basically all the conversion code for IEEE floats.
 ///
@@ -189,9 +189,14 @@ pub trait RawFloat:
 
     /// Returns the mantissa, exponent and sign as integers.
     ///
-    /// That is, this returns `(m, p, s)` such that `s * m * 2^p` represents the original float.
-    /// For 0, the exponent will be `-(EXP_BIAS + SIG_BITS`, which is the
-    /// minimum subnormal power.
+    /// This returns `(m, p, s)` such that `s * m * 2^p` represents the original float. For 0, the
+    /// exponent will be `-(EXP_BIAS + SIG_BITS)`, which is the minimum subnormal power. For
+    /// infinity or NaN, the exponent will be `EXP_SAT - EXP_BIAS - SIG_BITS`.
+    ///
+    /// If subnormal, the mantissa will be shifted one bit to the left. Otherwise, it is returned
+    /// with the explicit bit set but otherwise unshifted
+    ///
+    /// `s` is only ever +/-1.
     fn integer_decode(self) -> (u64, i16, i8) {
         let bits = self.to_bits();
         let sign: i8 = if bits >> (Self::BITS - 1) == Self::Int::ZERO { 1 } else { -1 };
@@ -213,6 +218,50 @@ const fn pow2_to_pow10(a: i64) -> i64 {
     res as i64
 }
 
+#[cfg(target_has_reliable_f16)]
+impl RawFloat for f16 {
+    type Int = u16;
+
+    const INFINITY: Self = Self::INFINITY;
+    const NEG_INFINITY: Self = Self::NEG_INFINITY;
+    const NAN: Self = Self::NAN;
+    const NEG_NAN: Self = -Self::NAN;
+
+    const BITS: u32 = 16;
+    const SIG_TOTAL_BITS: u32 = Self::MANTISSA_DIGITS;
+    const EXP_MASK: Self::Int = Self::EXP_MASK;
+    const SIG_MASK: Self::Int = Self::MAN_MASK;
+
+    const MIN_EXPONENT_ROUND_TO_EVEN: i32 = -22;
+    const MAX_EXPONENT_ROUND_TO_EVEN: i32 = 5;
+    const SMALLEST_POWER_OF_TEN: i32 = -27;
+
+    #[inline]
+    fn from_u64(v: u64) -> Self {
+        debug_assert!(v <= Self::MAX_MANTISSA_FAST_PATH);
+        v as _
+    }
+
+    #[inline]
+    fn from_u64_bits(v: u64) -> Self {
+        Self::from_bits((v & 0xFFFF) as u16)
+    }
+
+    fn pow10_fast_path(exponent: usize) -> Self {
+        #[allow(clippy::use_self)]
+        const TABLE: [f16; 8] = [1e0, 1e1, 1e2, 1e3, 1e4, 0.0, 0.0, 0.];
+        TABLE[exponent & 7]
+    }
+
+    fn to_bits(self) -> Self::Int {
+        self.to_bits()
+    }
+
+    fn classify(self) -> FpCategory {
+        self.classify()
+    }
+}
+
 impl RawFloat for f32 {
     type Int = u32;
 

library/core/src/num/dec2flt/mod.rs

@@ -171,9 +171,25 @@ macro_rules! from_str_float_impl {
         }
     };
 }
+
+#[cfg(target_has_reliable_f16)]
+from_str_float_impl!(f16);
 from_str_float_impl!(f32);
 from_str_float_impl!(f64);
 
+// FIXME(f16_f128): A fallback is used when the backend+target does not support f16 well, in order
+// to avoid ICEs.
+
+#[cfg(not(target_has_reliable_f16))]
+impl FromStr for f16 {
+    type Err = ParseFloatError;
+
+    #[inline]
+    fn from_str(_src: &str) -> Result<Self, ParseFloatError> {
+        unimplemented!("requires target_has_reliable_f16")
+    }
+}
+
 /// An error which can be returned when parsing a float.
 ///
 /// This error is used as the error type for the [`FromStr`] implementation

library/core/src/num/flt2dec/decoder.rs

@@ -45,6 +45,13 @@ pub trait DecodableFloat: RawFloat + Copy {
     fn min_pos_norm_value() -> Self;
 }
 
+#[cfg(target_has_reliable_f16)]
+impl DecodableFloat for f16 {
+    fn min_pos_norm_value() -> Self {
+        f16::MIN_POSITIVE
+    }
+}
+
 impl DecodableFloat for f32 {
     fn min_pos_norm_value() -> Self {
         f32::MIN_POSITIVE

library/coretests/tests/num/dec2flt/decimal.rs

@@ -7,6 +7,20 @@ const FPATHS_F32: &[FPath<f32>] =
 const FPATHS_F64: &[FPath<f64>] =
     &[((0, 0, false, false), Some(0.0)), ((0, 0, false, false), Some(0.0))];
 
+// FIXME(f16_f128): enable on all targets once possible.
+#[test]
+#[cfg(target_has_reliable_f16)]
+fn check_fast_path_f16() {
+    const FPATHS_F16: &[FPath<f16>] =
+        &[((0, 0, false, false), Some(0.0)), ((0, 0, false, false), Some(0.0))];
+    for ((exponent, mantissa, negative, many_digits), expected) in FPATHS_F16.iter().copied() {
+        let dec = Decimal { exponent, mantissa, negative, many_digits };
+        let actual = dec.try_fast_path::<f16>();
+
+        assert_eq!(actual, expected);
+    }
+}
+
 #[test]
 fn check_fast_path_f32() {
     for ((exponent, mantissa, negative, many_digits), expected) in FPATHS_F32.iter().copied() {

library/coretests/tests/num/dec2flt/float.rs

@@ -1,5 +1,24 @@
 use core::num::dec2flt::float::RawFloat;
 
+// FIXME(f16_f128): enable on all targets once possible.
+#[test]
+#[cfg(target_has_reliable_f16)]
+fn test_f16_integer_decode() {
+    assert_eq!(3.14159265359f16.integer_decode(), (1608, -9, 1));
+    assert_eq!((-8573.5918555f16).integer_decode(), (1072, 3, -1));
+    #[cfg(not(miri))] // miri doesn't have powf16
+    assert_eq!(2f16.powf(14.0).integer_decode(), (1 << 10, 4, 1));
+    assert_eq!(0f16.integer_decode(), (0, -25, 1));
+    assert_eq!((-0f16).integer_decode(), (0, -25, -1));
+    assert_eq!(f16::INFINITY.integer_decode(), (1 << 10, 6, 1));
+    assert_eq!(f16::NEG_INFINITY.integer_decode(), (1 << 10, 6, -1));
+
+    // Ignore the "sign" (quiet / signalling flag) of NAN.
+    // It can vary between runtime operations and LLVM folding.
+    let (nan_m, nan_p, _nan_s) = f16::NAN.integer_decode();
+    assert_eq!((nan_m, nan_p), (1536, 6));
+}
+
 #[test]
 fn test_f32_integer_decode() {
     assert_eq!(3.14159265359f32.integer_decode(), (13176795, -22, 1));
@@ -34,6 +53,27 @@ fn test_f64_integer_decode() {
 
 /* Sanity checks of computed magic numbers */
 
+// FIXME(f16_f128): enable on all targets once possible.
+#[test]
+#[cfg(target_has_reliable_f16)]
+fn test_f16_consts() {
+    assert_eq!(<f16 as RawFloat>::INFINITY, f16::INFINITY);
+    assert_eq!(<f16 as RawFloat>::NEG_INFINITY, -f16::INFINITY);
+    assert_eq!(<f16 as RawFloat>::NAN.to_bits(), f16::NAN.to_bits());
+    assert_eq!(<f16 as RawFloat>::NEG_NAN.to_bits(), (-f16::NAN).to_bits());
+    assert_eq!(<f16 as RawFloat>::SIG_BITS, 10);
+    assert_eq!(<f16 as RawFloat>::MIN_EXPONENT_ROUND_TO_EVEN, -22);
+    assert_eq!(<f16 as RawFloat>::MAX_EXPONENT_ROUND_TO_EVEN, 5);
+    assert_eq!(<f16 as RawFloat>::MIN_EXPONENT_FAST_PATH, -4);
+    assert_eq!(<f16 as RawFloat>::MAX_EXPONENT_FAST_PATH, 4);
+    assert_eq!(<f16 as RawFloat>::MAX_EXPONENT_DISGUISED_FAST_PATH, 7);
+    assert_eq!(<f16 as RawFloat>::EXP_MIN, -14);
+    assert_eq!(<f16 as RawFloat>::EXP_SAT, 0x1f);
+    assert_eq!(<f16 as RawFloat>::SMALLEST_POWER_OF_TEN, -27);
+    assert_eq!(<f16 as RawFloat>::LARGEST_POWER_OF_TEN, 4);
+    assert_eq!(<f16 as RawFloat>::MAX_MANTISSA_FAST_PATH, 2048);
+}
+
 #[test]
 fn test_f32_consts() {
     assert_eq!(<f32 as RawFloat>::INFINITY, f32::INFINITY);