1 | //! Converting decimal strings into IEEE 754 binary floating point numbers. |
2 | //! |
3 | //! # Problem statement |
4 | //! |
5 | //! We are given a decimal string such as `12.34e56`. This string consists of integral (`12`), |
6 | //! fractional (`34`), and exponent (`56`) parts. All parts are optional and interpreted as zero |
7 | //! when missing. |
8 | //! |
9 | //! We seek the IEEE 754 floating point number that is closest to the exact value of the decimal |
10 | //! string. It is well-known that many decimal strings do not have terminating representations in |
11 | //! base two, so we round to 0.5 units in the last place (in other words, as well as possible). |
12 | //! Ties, decimal values exactly half-way between two consecutive floats, are resolved with the |
13 | //! half-to-even strategy, also known as banker's rounding. |
14 | //! |
15 | //! Needless to say, this is quite hard, both in terms of implementation complexity and in terms |
16 | //! of CPU cycles taken. |
17 | //! |
18 | //! # Implementation |
19 | //! |
20 | //! First, we ignore signs. Or rather, we remove it at the very beginning of the conversion |
21 | //! process and re-apply it at the very end. This is correct in all edge cases since IEEE |
22 | //! floats are symmetric around zero, negating one simply flips the first bit. |
23 | //! |
24 | //! Then we remove the decimal point by adjusting the exponent: Conceptually, `12.34e56` turns |
25 | //! into `1234e54`, which we describe with a positive integer `f = 1234` and an integer `e = 54`. |
26 | //! The `(f, e)` representation is used by almost all code past the parsing stage. |
27 | //! |
28 | //! We then try a long chain of progressively more general and expensive special cases using |
29 | //! machine-sized integers and small, fixed-sized floating point numbers (first `f32`/`f64`, then |
30 | //! a type with 64 bit significand). The extended-precision algorithm |
31 | //! uses the Eisel-Lemire algorithm, which uses a 128-bit (or 192-bit) |
32 | //! representation that can accurately and quickly compute the vast majority |
33 | //! of floats. When all these fail, we bite the bullet and resort to using |
34 | //! a large-decimal representation, shifting the digits into range, calculating |
35 | //! the upper significant bits and exactly round to the nearest representation. |
36 | //! |
37 | //! Another aspect that needs attention is the ``RawFloat`` trait by which almost all functions |
38 | //! are parametrized. One might think that it's enough to parse to `f64` and cast the result to |
39 | //! `f32`. Unfortunately this is not the world we live in, and this has nothing to do with using |
40 | //! base two or half-to-even rounding. |
41 | //! |
42 | //! Consider for example two types `d2` and `d4` representing a decimal type with two decimal |
43 | //! digits and four decimal digits each and take "0.01499" as input. Let's use half-up rounding. |
44 | //! Going directly to two decimal digits gives `0.01`, but if we round to four digits first, |
45 | //! we get `0.0150`, which is then rounded up to `0.02`. The same principle applies to other |
46 | //! operations as well, if you want 0.5 ULP accuracy you need to do *everything* in full precision |
47 | //! and round *exactly once, at the end*, by considering all truncated bits at once. |
48 | //! |
49 | //! Primarily, this module and its children implement the algorithms described in: |
50 | //! "Number Parsing at a Gigabyte per Second", available online: |
51 | //! <https://arxiv.org/abs/2101.11408>. |
52 | //! |
53 | //! # Other |
54 | //! |
55 | //! The conversion should *never* panic. There are assertions and explicit panics in the code, |
56 | //! but they should never be triggered and only serve as internal sanity checks. Any panics should |
57 | //! be considered a bug. |
58 | //! |
59 | //! There are unit tests but they are woefully inadequate at ensuring correctness, they only cover |
60 | //! a small percentage of possible errors. Far more extensive tests are located in the directory |
61 | //! `src/etc/test-float-parse` as a Python script. |
62 | //! |
63 | //! A note on integer overflow: Many parts of this file perform arithmetic with the decimal |
64 | //! exponent `e`. Primarily, we shift the decimal point around: Before the first decimal digit, |
65 | //! after the last decimal digit, and so on. This could overflow if done carelessly. We rely on |
66 | //! the parsing submodule to only hand out sufficiently small exponents, where "sufficient" means |
67 | //! "such that the exponent +/- the number of decimal digits fits into a 64 bit integer". |
68 | //! Larger exponents are accepted, but we don't do arithmetic with them, they are immediately |
69 | //! turned into {positive,negative} {zero,infinity}. |
70 | |
71 | #![doc (hidden)] |
72 | #![unstable ( |
73 | feature = "dec2flt" , |
74 | reason = "internal routines only exposed for testing" , |
75 | issue = "none" |
76 | )] |
77 | |
78 | use crate::error::Error; |
79 | use crate::fmt; |
80 | use crate::str::FromStr; |
81 | |
82 | use self::common::BiasedFp; |
83 | use self::float::RawFloat; |
84 | use self::lemire::compute_float; |
85 | use self::parse::{parse_inf_nan, parse_number}; |
86 | use self::slow::parse_long_mantissa; |
87 | |
88 | mod common; |
89 | mod decimal; |
90 | mod fpu; |
91 | mod slow; |
92 | mod table; |
93 | // float is used in flt2dec, and all are used in unit tests. |
94 | pub mod float; |
95 | pub mod lemire; |
96 | pub mod number; |
97 | pub mod parse; |
98 | |
99 | macro_rules! from_str_float_impl { |
100 | ($t:ty) => { |
101 | #[stable(feature = "rust1" , since = "1.0.0" )] |
102 | impl FromStr for $t { |
103 | type Err = ParseFloatError; |
104 | |
105 | /// Converts a string in base 10 to a float. |
106 | /// Accepts an optional decimal exponent. |
107 | /// |
108 | /// This function accepts strings such as |
109 | /// |
110 | /// * '3.14' |
111 | /// * '-3.14' |
112 | /// * '2.5E10', or equivalently, '2.5e10' |
113 | /// * '2.5E-10' |
114 | /// * '5.' |
115 | /// * '.5', or, equivalently, '0.5' |
116 | /// * 'inf', '-inf', '+infinity', 'NaN' |
117 | /// |
118 | /// Note that alphabetical characters are not case-sensitive. |
119 | /// |
120 | /// Leading and trailing whitespace represent an error. |
121 | /// |
122 | /// # Grammar |
123 | /// |
124 | /// All strings that adhere to the following [EBNF] grammar when |
125 | /// lowercased will result in an [`Ok`] being returned: |
126 | /// |
127 | /// ```txt |
128 | /// Float ::= Sign? ( 'inf' | 'infinity' | 'nan' | Number ) |
129 | /// Number ::= ( Digit+ | |
130 | /// Digit+ '.' Digit* | |
131 | /// Digit* '.' Digit+ ) Exp? |
132 | /// Exp ::= 'e' Sign? Digit+ |
133 | /// Sign ::= [+-] |
134 | /// Digit ::= [0-9] |
135 | /// ``` |
136 | /// |
137 | /// [EBNF]: https://www.w3.org/TR/REC-xml/#sec-notation |
138 | /// |
139 | /// # Arguments |
140 | /// |
141 | /// * src - A string |
142 | /// |
143 | /// # Return value |
144 | /// |
145 | /// `Err(ParseFloatError)` if the string did not represent a valid |
146 | /// number. Otherwise, `Ok(n)` where `n` is the closest |
147 | /// representable floating-point number to the number represented |
148 | /// by `src` (following the same rules for rounding as for the |
149 | /// results of primitive operations). |
150 | // We add the `#[inline(never)]` attribute, since its content will |
151 | // be filled with that of `dec2flt`, which has #[inline(always)]. |
152 | // Since `dec2flt` is generic, a normal inline attribute on this function |
153 | // with `dec2flt` having no attributes results in heavily repeated |
154 | // generation of `dec2flt`, despite the fact only a maximum of 2 |
155 | // possible instances can ever exist. Adding #[inline(never)] avoids this. |
156 | #[inline(never)] |
157 | fn from_str(src: &str) -> Result<Self, ParseFloatError> { |
158 | dec2flt(src) |
159 | } |
160 | } |
161 | }; |
162 | } |
163 | from_str_float_impl!(f32); |
164 | from_str_float_impl!(f64); |
165 | |
166 | /// An error which can be returned when parsing a float. |
167 | /// |
168 | /// This error is used as the error type for the [`FromStr`] implementation |
169 | /// for [`f32`] and [`f64`]. |
170 | /// |
171 | /// # Example |
172 | /// |
173 | /// ``` |
174 | /// use std::str::FromStr; |
175 | /// |
176 | /// if let Err(e) = f64::from_str("a.12" ) { |
177 | /// println!("Failed conversion to f64: {e}" ); |
178 | /// } |
179 | /// ``` |
180 | #[derive (Debug, Clone, PartialEq, Eq)] |
181 | #[stable (feature = "rust1" , since = "1.0.0" )] |
182 | pub struct ParseFloatError { |
183 | kind: FloatErrorKind, |
184 | } |
185 | |
186 | #[derive (Debug, Clone, PartialEq, Eq)] |
187 | enum FloatErrorKind { |
188 | Empty, |
189 | Invalid, |
190 | } |
191 | |
192 | #[stable (feature = "rust1" , since = "1.0.0" )] |
193 | impl Error for ParseFloatError { |
194 | #[allow (deprecated)] |
195 | fn description(&self) -> &str { |
196 | match self.kind { |
197 | FloatErrorKind::Empty => "cannot parse float from empty string" , |
198 | FloatErrorKind::Invalid => "invalid float literal" , |
199 | } |
200 | } |
201 | } |
202 | |
203 | #[stable (feature = "rust1" , since = "1.0.0" )] |
204 | impl fmt::Display for ParseFloatError { |
205 | fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { |
206 | #[allow (deprecated)] |
207 | self.description().fmt(f) |
208 | } |
209 | } |
210 | |
211 | #[inline ] |
212 | pub(super) fn pfe_empty() -> ParseFloatError { |
213 | ParseFloatError { kind: FloatErrorKind::Empty } |
214 | } |
215 | |
216 | // Used in unit tests, keep public. |
217 | // This is much better than making FloatErrorKind and ParseFloatError::kind public. |
218 | #[inline ] |
219 | pub fn pfe_invalid() -> ParseFloatError { |
220 | ParseFloatError { kind: FloatErrorKind::Invalid } |
221 | } |
222 | |
223 | /// Converts a `BiasedFp` to the closest machine float type. |
224 | fn biased_fp_to_float<T: RawFloat>(x: BiasedFp) -> T { |
225 | let mut word: u64 = x.f; |
226 | word |= (x.e as u64) << T::MANTISSA_EXPLICIT_BITS; |
227 | T::from_u64_bits(word) |
228 | } |
229 | |
230 | /// Converts a decimal string into a floating point number. |
231 | #[inline (always)] // Will be inlined into a function with `#[inline(never)]`, see above |
232 | pub fn dec2flt<F: RawFloat>(s: &str) -> Result<F, ParseFloatError> { |
233 | let mut s = s.as_bytes(); |
234 | let c = if let Some(&c) = s.first() { |
235 | c |
236 | } else { |
237 | return Err(pfe_empty()); |
238 | }; |
239 | let negative = c == b'-' ; |
240 | if c == b'-' || c == b'+' { |
241 | s = &s[1..]; |
242 | } |
243 | if s.is_empty() { |
244 | return Err(pfe_invalid()); |
245 | } |
246 | |
247 | let mut num = match parse_number(s) { |
248 | Some(r) => r, |
249 | None if let Some(value) = parse_inf_nan(s, negative) => return Ok(value), |
250 | None => return Err(pfe_invalid()), |
251 | }; |
252 | num.negative = negative; |
253 | if let Some(value) = num.try_fast_path::<F>() { |
254 | return Ok(value); |
255 | } |
256 | |
257 | // If significant digits were truncated, then we can have rounding error |
258 | // only if `mantissa + 1` produces a different result. We also avoid |
259 | // redundantly using the Eisel-Lemire algorithm if it was unable to |
260 | // correctly round on the first pass. |
261 | let mut fp = compute_float::<F>(num.exponent, num.mantissa); |
262 | if num.many_digits && fp.e >= 0 && fp != compute_float::<F>(num.exponent, num.mantissa + 1) { |
263 | fp.e = -1; |
264 | } |
265 | // Unable to correctly round the float using the Eisel-Lemire algorithm. |
266 | // Fallback to a slower, but always correct algorithm. |
267 | if fp.e < 0 { |
268 | fp = parse_long_mantissa::<F>(s); |
269 | } |
270 | |
271 | let mut float = biased_fp_to_float::<F>(fp); |
272 | if num.negative { |
273 | float = -float; |
274 | } |
275 | Ok(float) |
276 | } |
277 | |