1 // Copyright 2015 The Rust Project Developers. See the COPYRIGHT
2 // file at the top-level directory of this distribution and at
3 // http://rust-lang.org/COPYRIGHT.
5 // Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
6 // http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
7 // <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
8 // option. This file may not be copied, modified, or distributed
9 // except according to those terms.
11 //! Bit fiddling on positive IEEE 754 floats. Negative numbers aren't and needn't be handled.
12 //! Normal floating point numbers have a canonical representation as (frac, exp) such that the
13 //! value is 2<sup>exp</sup> * (1 + sum(frac[N-i] / 2<sup>i</sup>)) where N is the number of bits.
14 //! Subnormals are slightly different and weird, but the same principle applies.
16 //! Here, however, we represent them as (sig, k) with f positive, such that the value is f *
17 //! 2<sup>e</sup>. Besides making the "hidden bit" explicit, this changes the exponent by the
18 //! so-called mantissa shift.
20 //! Put another way, normally floats are written as (1) but here they are written as (2):
22 //! 1. `1.101100...11 * 2^m`
23 //! 2. `1101100...11 * 2^n`
25 //! We call (1) the **fractional representation** and (2) the **integral representation**.
27 //! Many functions in this module only handle normal numbers. The dec2flt routines conservatively
28 //! take the universally-correct slow path (Algorithm M) for very small and very large numbers.
29 //! That algorithm needs only next_float() which does handle subnormals and zeros.
31 use cmp::Ordering::{Less, Equal, Greater};
32 use ops::{Mul, Div, Neg};
33 use fmt::{Debug, LowerExp};
35 use num::diy_float::Fp;
36 use num::FpCategory::{Infinite, Zero, Subnormal, Normal, Nan};
38 use num::dec2flt::num::{self, Big};
39 use num::dec2flt::table;
41 #[derive(Copy, Clone, Debug)]
48 pub fn new(sig: u64, k: i16) -> Self {
49 Unpacked { sig: sig, k: k }
53 /// A helper trait to avoid duplicating basically all the conversion code for `f32` and `f64`.
55 /// See the parent module's doc comment for why this is necessary.
57 /// Should **never ever** be implemented for other types or be used outside the dec2flt module.
58 /// Inherits from `Float` because there is some overlap, but all the reused methods are trivial.
59 pub trait RawFloat : Float + Copy + Debug + LowerExp
60 + Mul<Output=Self> + Div<Output=Self> + Neg<Output=Self>
66 /// Returns the mantissa, exponent and sign as integers.
67 fn integer_decode(self) -> (u64, i16, i8);
69 /// Get the raw binary representation of the float.
70 fn transmute(self) -> u64;
72 /// Transmute the raw binary representation into a float.
73 fn from_bits(bits: u64) -> Self;
76 fn unpack(self) -> Unpacked;
78 /// Cast from a small integer that can be represented exactly. Panic if the integer can't be
79 /// represented, the other code in this module makes sure to never let that happen.
80 fn from_int(x: u64) -> Self;
82 /// Get the value 10<sup>e</sup> from a pre-computed table.
83 /// Panics for `e >= CEIL_LOG5_OF_MAX_SIG`.
84 fn short_fast_pow10(e: usize) -> Self;
86 /// What the name says. It's easier to hard code than juggling intrinsics and
87 /// hoping LLVM constant folds it.
88 const CEIL_LOG5_OF_MAX_SIG: i16;
90 // A conservative bound on the decimal digits of inputs that can't produce overflow or zero or
91 /// subnormals. Probably the decimal exponent of the maximum normal value, hence the name.
92 const MAX_NORMAL_DIGITS: usize;
94 /// When the most significant decimal digit has a place value greater than this, the number
95 /// is certainly rounded to infinity.
96 const INF_CUTOFF: i64;
98 /// When the most significant decimal digit has a place value less than this, the number
99 /// is certainly rounded to zero.
100 const ZERO_CUTOFF: i64;
102 /// The number of bits in the exponent.
105 /// The number of bits in the singificand, *including* the hidden bit.
108 /// The number of bits in the singificand, *excluding* the hidden bit.
109 const EXPLICIT_SIG_BITS: u8;
111 /// The maximum legal exponent in fractional representation.
114 /// The minimum legal exponent in fractional representation, excluding subnormals.
117 /// `MAX_EXP` for integral representation, i.e., with the shift applied.
118 const MAX_EXP_INT: i16;
120 /// `MAX_EXP` encoded (i.e., with offset bias)
121 const MAX_ENCODED_EXP: i16;
123 /// `MIN_EXP` for integral representation, i.e., with the shift applied.
124 const MIN_EXP_INT: i16;
126 /// The maximum normalized singificand in integral representation.
129 /// The minimal normalized significand in integral representation.
133 // Mostly a workaround for #34344.
134 macro_rules! other_constants {
136 const EXPLICIT_SIG_BITS: u8 = Self::SIG_BITS - 1;
137 const MAX_EXP: i16 = (1 << (Self::EXP_BITS - 1)) - 1;
138 const MIN_EXP: i16 = -Self::MAX_EXP + 1;
139 const MAX_EXP_INT: i16 = Self::MAX_EXP - (Self::SIG_BITS as i16 - 1);
140 const MAX_ENCODED_EXP: i16 = (1 << Self::EXP_BITS) - 1;
141 const MIN_EXP_INT: i16 = Self::MIN_EXP - (Self::SIG_BITS as i16 - 1);
142 const MAX_SIG: u64 = (1 << Self::SIG_BITS) - 1;
143 const MIN_SIG: u64 = 1 << (Self::SIG_BITS - 1);
145 const INFINITY: Self = $crate::$type::INFINITY;
146 const NAN: Self = $crate::$type::NAN;
147 const ZERO: Self = 0.0;
151 impl RawFloat for f32 {
152 const SIG_BITS: u8 = 24;
153 const EXP_BITS: u8 = 8;
154 const CEIL_LOG5_OF_MAX_SIG: i16 = 11;
155 const MAX_NORMAL_DIGITS: usize = 35;
156 const INF_CUTOFF: i64 = 40;
157 const ZERO_CUTOFF: i64 = -48;
158 other_constants!(f32);
160 /// Returns the mantissa, exponent and sign as integers.
161 fn integer_decode(self) -> (u64, i16, i8) {
162 let bits: u32 = unsafe { transmute(self) };
163 let sign: i8 = if bits >> 31 == 0 { 1 } else { -1 };
164 let mut exponent: i16 = ((bits >> 23) & 0xff) as i16;
165 let mantissa = if exponent == 0 {
166 (bits & 0x7fffff) << 1
168 (bits & 0x7fffff) | 0x800000
170 // Exponent bias + mantissa shift
171 exponent -= 127 + 23;
172 (mantissa as u64, exponent, sign)
175 fn transmute(self) -> u64 {
176 let bits: u32 = unsafe { transmute(self) };
180 fn from_bits(bits: u64) -> f32 {
181 assert!(bits < u32::MAX as u64, "f32::from_bits: too many bits");
182 unsafe { transmute(bits as u32) }
185 fn unpack(self) -> Unpacked {
186 let (sig, exp, _sig) = self.integer_decode();
187 Unpacked::new(sig, exp)
190 fn from_int(x: u64) -> f32 {
191 // rkruppe is uncertain whether `as` rounds correctly on all platforms.
192 debug_assert!(x as f32 == fp_to_float(Fp { f: x, e: 0 }));
196 fn short_fast_pow10(e: usize) -> Self {
197 table::F32_SHORT_POWERS[e]
202 impl RawFloat for f64 {
203 const SIG_BITS: u8 = 53;
204 const EXP_BITS: u8 = 11;
205 const CEIL_LOG5_OF_MAX_SIG: i16 = 23;
206 const MAX_NORMAL_DIGITS: usize = 305;
207 const INF_CUTOFF: i64 = 310;
208 const ZERO_CUTOFF: i64 = -326;
209 other_constants!(f64);
211 /// Returns the mantissa, exponent and sign as integers.
212 fn integer_decode(self) -> (u64, i16, i8) {
213 let bits: u64 = unsafe { transmute(self) };
214 let sign: i8 = if bits >> 63 == 0 { 1 } else { -1 };
215 let mut exponent: i16 = ((bits >> 52) & 0x7ff) as i16;
216 let mantissa = if exponent == 0 {
217 (bits & 0xfffffffffffff) << 1
219 (bits & 0xfffffffffffff) | 0x10000000000000
221 // Exponent bias + mantissa shift
222 exponent -= 1023 + 52;
223 (mantissa, exponent, sign)
226 fn transmute(self) -> u64 {
227 let bits: u64 = unsafe { transmute(self) };
231 fn from_bits(bits: u64) -> f64 {
232 unsafe { transmute(bits) }
235 fn unpack(self) -> Unpacked {
236 let (sig, exp, _sig) = self.integer_decode();
237 Unpacked::new(sig, exp)
240 fn from_int(x: u64) -> f64 {
241 // rkruppe is uncertain whether `as` rounds correctly on all platforms.
242 debug_assert!(x as f64 == fp_to_float(Fp { f: x, e: 0 }));
246 fn short_fast_pow10(e: usize) -> Self {
247 table::F64_SHORT_POWERS[e]
251 /// Convert an Fp to the closest machine float type.
252 /// Does not handle subnormal results.
253 pub fn fp_to_float<T: RawFloat>(x: Fp) -> T {
254 let x = x.normalize();
255 // x.f is 64 bit, so x.e has a mantissa shift of 63
258 panic!("fp_to_float: exponent {} too large", e)
259 } else if e > T::MIN_EXP {
260 encode_normal(round_normal::<T>(x))
262 panic!("fp_to_float: exponent {} too small", e)
266 /// Round the 64-bit significand to T::SIG_BITS bits with half-to-even.
267 /// Does not handle exponent overflow.
268 pub fn round_normal<T: RawFloat>(x: Fp) -> Unpacked {
269 let excess = 64 - T::SIG_BITS as i16;
270 let half: u64 = 1 << (excess - 1);
271 let (q, rem) = (x.f >> excess, x.f & ((1 << excess) - 1));
272 assert_eq!(q << excess | rem, x.f);
273 // Adjust mantissa shift
274 let k = x.e + excess;
277 } else if rem == half && (q % 2) == 0 {
279 } else if q == T::MAX_SIG {
280 Unpacked::new(T::MIN_SIG, k + 1)
282 Unpacked::new(q + 1, k)
286 /// Inverse of `RawFloat::unpack()` for normalized numbers.
287 /// Panics if the significand or exponent are not valid for normalized numbers.
288 pub fn encode_normal<T: RawFloat>(x: Unpacked) -> T {
289 debug_assert!(T::MIN_SIG <= x.sig && x.sig <= T::MAX_SIG,
290 "encode_normal: significand not normalized");
291 // Remove the hidden bit
292 let sig_enc = x.sig & !(1 << T::EXPLICIT_SIG_BITS);
293 // Adjust the exponent for exponent bias and mantissa shift
294 let k_enc = x.k + T::MAX_EXP + T::EXPLICIT_SIG_BITS as i16;
295 debug_assert!(k_enc != 0 && k_enc < T::MAX_ENCODED_EXP,
296 "encode_normal: exponent out of range");
297 // Leave sign bit at 0 ("+"), our numbers are all positive
298 let bits = (k_enc as u64) << T::EXPLICIT_SIG_BITS | sig_enc;
302 /// Construct a subnormal. A mantissa of 0 is allowed and constructs zero.
303 pub fn encode_subnormal<T: RawFloat>(significand: u64) -> T {
304 assert!(significand < T::MIN_SIG, "encode_subnormal: not actually subnormal");
305 // Encoded exponent is 0, the sign bit is 0, so we just have to reinterpret the bits.
306 T::from_bits(significand)
309 /// Approximate a bignum with an Fp. Rounds within 0.5 ULP with half-to-even.
310 pub fn big_to_fp(f: &Big) -> Fp {
311 let end = f.bit_length();
312 assert!(end != 0, "big_to_fp: unexpectedly, input is zero");
313 let start = end.saturating_sub(64);
314 let leading = num::get_bits(f, start, end);
315 // We cut off all bits prior to the index `start`, i.e., we effectively right-shift by
316 // an amount of `start`, so this is also the exponent we need.
317 let e = start as i16;
318 let rounded_down = Fp { f: leading, e: e }.normalize();
319 // Round (half-to-even) depending on the truncated bits.
320 match num::compare_with_half_ulp(f, start) {
321 Less => rounded_down,
322 Equal if leading % 2 == 0 => rounded_down,
323 Equal | Greater => match leading.checked_add(1) {
324 Some(f) => Fp { f: f, e: e }.normalize(),
325 None => Fp { f: 1 << 63, e: e + 1 },
330 /// Find the largest floating point number strictly smaller than the argument.
331 /// Does not handle subnormals, zero, or exponent underflow.
332 pub fn prev_float<T: RawFloat>(x: T) -> T {
334 Infinite => panic!("prev_float: argument is infinite"),
335 Nan => panic!("prev_float: argument is NaN"),
336 Subnormal => panic!("prev_float: argument is subnormal"),
337 Zero => panic!("prev_float: argument is zero"),
339 let Unpacked { sig, k } = x.unpack();
340 if sig == T::MIN_SIG {
341 encode_normal(Unpacked::new(T::MAX_SIG, k - 1))
343 encode_normal(Unpacked::new(sig - 1, k))
349 // Find the smallest floating point number strictly larger than the argument.
350 // This operation is saturating, i.e. next_float(inf) == inf.
351 // Unlike most code in this module, this function does handle zero, subnormals, and infinities.
352 // However, like all other code here, it does not deal with NaN and negative numbers.
353 pub fn next_float<T: RawFloat>(x: T) -> T {
355 Nan => panic!("next_float: argument is NaN"),
356 Infinite => T::INFINITY,
357 // This seems too good to be true, but it works.
358 // 0.0 is encoded as the all-zero word. Subnormals are 0x000m...m where m is the mantissa.
359 // In particular, the smallest subnormal is 0x0...01 and the largest is 0x000F...F.
360 // The smallest normal number is 0x0010...0, so this corner case works as well.
361 // If the increment overflows the mantissa, the carry bit increments the exponent as we
362 // want, and the mantissa bits become zero. Because of the hidden bit convention, this
363 // too is exactly what we want!
364 // Finally, f64::MAX + 1 = 7eff...f + 1 = 7ff0...0 = f64::INFINITY.
365 Zero | Subnormal | Normal => {
366 let bits: u64 = x.transmute();
367 T::from_bits(bits + 1)