1 | use core::ops::Neg; |
2 | |
3 | use crate::int::{CastFrom, CastInto, Int, MinInt}; |
4 | |
5 | use super::Float; |
6 | |
7 | /// Conversions from integers to floats. |
8 | /// |
9 | /// The algorithm is explained here: <https://blog.m-ou.se/floats/>. It roughly does the following: |
10 | /// - Calculate a base mantissa by shifting the integer into mantissa position. This gives us a |
11 | /// mantissa _with the implicit bit set_! |
12 | /// - Figure out if rounding needs to occur by classifying the bits that are to be truncated. Some |
13 | /// patterns are used to simplify this. Adjust the mantissa with the result if needed. |
14 | /// - Calculate the exponent based on the base-2 logarithm of `i` (leading zeros). Subtract one. |
15 | /// - Shift the exponent and add the mantissa to create the final representation. Subtracting one |
16 | /// from the exponent (above) accounts for the explicit bit being set in the mantissa. |
17 | /// |
18 | /// # Terminology |
19 | /// |
20 | /// - `i`: the original integer |
21 | /// - `i_m`: the integer, shifted fully left (no leading zeros) |
22 | /// - `n`: number of leading zeroes |
23 | /// - `e`: the resulting exponent. Usually 1 is subtracted to offset the mantissa implicit bit. |
24 | /// - `m_base`: the mantissa before adjusting for truncated bits. Implicit bit is usually set. |
25 | /// - `adj`: the bits that will be truncated, possibly compressed in some way. |
26 | /// - `m`: the resulting mantissa. Implicit bit is usually set. |
27 | mod int_to_float { |
28 | use super::*; |
29 | |
30 | /// Calculate the exponent from the number of leading zeros. |
31 | /// |
32 | /// Usually 1 is subtracted from this function's result, so that a mantissa with the implicit |
33 | /// bit set can be added back later. |
34 | fn exp<I: Int, F: Float<Int: CastFrom<u32>>>(n: u32) -> F::Int { |
35 | F::Int::cast_from(F::EXP_BIAS - 1 + I::BITS - n) |
36 | } |
37 | |
38 | /// Adjust a mantissa with dropped bits to perform correct rounding. |
39 | /// |
40 | /// The dropped bits should be exactly the bits that get truncated (left-aligned), but they |
41 | /// can be combined or compressed in some way that simplifies operations. |
42 | fn m_adj<F: Float>(m_base: F::Int, dropped_bits: F::Int) -> F::Int { |
43 | // Branchlessly extract a `1` if rounding up should happen, 0 otherwise |
44 | // This accounts for rounding to even. |
45 | let adj = (dropped_bits - ((dropped_bits >> (F::BITS - 1)) & !m_base)) >> (F::BITS - 1); |
46 | |
47 | // Add one when we need to round up. Break ties to even. |
48 | m_base + adj |
49 | } |
50 | |
51 | /// Shift the exponent to its position and add the mantissa. |
52 | /// |
53 | /// If the mantissa has the implicit bit set, the exponent should be one less than its actual |
54 | /// value to cancel it out. |
55 | fn repr<F: Float>(e: F::Int, m: F::Int) -> F::Int { |
56 | // + rather than | so the mantissa can overflow into the exponent |
57 | (e << F::SIG_BITS) + m |
58 | } |
59 | |
60 | /// Shift distance from a left-aligned integer to a smaller float. |
61 | fn shift_f_lt_i<I: Int, F: Float>() -> u32 { |
62 | (I::BITS - F::BITS) + F::EXP_BITS |
63 | } |
64 | |
65 | /// Shift distance from an integer with `n` leading zeros to a smaller float. |
66 | fn shift_f_gt_i<I: Int, F: Float>(n: u32) -> u32 { |
67 | F::SIG_BITS - I::BITS + 1 + n |
68 | } |
69 | |
70 | /// Perform a signed operation as unsigned, then add the sign back. |
71 | pub fn signed<I, F, Conv>(i: I, conv: Conv) -> F |
72 | where |
73 | F: Float, |
74 | I: Int, |
75 | F::Int: CastFrom<I>, |
76 | Conv: Fn(I::UnsignedInt) -> F::Int, |
77 | { |
78 | let sign_bit = F::Int::cast_from(i >> (I::BITS - 1)) << (F::BITS - 1); |
79 | F::from_bits(conv(i.unsigned_abs()) | sign_bit) |
80 | } |
81 | |
82 | pub fn u32_to_f32_bits(i: u32) -> u32 { |
83 | if i == 0 { |
84 | return 0; |
85 | } |
86 | let n = i.leading_zeros(); |
87 | // Mantissa with implicit bit set (significant bits) |
88 | let m_base = (i << n) >> f32::EXP_BITS; |
89 | // Bits that will be dropped (insignificant bits) |
90 | let adj = (i << n) << (f32::SIG_BITS + 1); |
91 | let m = m_adj::<f32>(m_base, adj); |
92 | let e = exp::<u32, f32>(n) - 1; |
93 | repr::<f32>(e, m) |
94 | } |
95 | |
96 | pub fn u32_to_f64_bits(i: u32) -> u64 { |
97 | if i == 0 { |
98 | return 0; |
99 | } |
100 | let n = i.leading_zeros(); |
101 | // Mantissa with implicit bit set |
102 | let m = (i as u64) << shift_f_gt_i::<u32, f64>(n); |
103 | let e = exp::<u32, f64>(n) - 1; |
104 | repr::<f64>(e, m) |
105 | } |
106 | |
107 | #[cfg (f128_enabled)] |
108 | pub fn u32_to_f128_bits(i: u32) -> u128 { |
109 | if i == 0 { |
110 | return 0; |
111 | } |
112 | let n = i.leading_zeros(); |
113 | |
114 | // Shift into mantissa position that is correct for the type, but shifted into the lower |
115 | // 64 bits over so can can avoid 128-bit math. |
116 | let m = (i as u64) << (shift_f_gt_i::<u32, f128>(n) - 64); |
117 | let e = exp::<u32, f128>(n) as u64 - 1; |
118 | // High 64 bits of f128 representation. |
119 | let h = (e << (f128::SIG_BITS - 64)) + m; |
120 | |
121 | // Shift back to the high bits, the rest of the mantissa will always be 0. |
122 | (h as u128) << 64 |
123 | } |
124 | |
125 | pub fn u64_to_f32_bits(i: u64) -> u32 { |
126 | let n = i.leading_zeros(); |
127 | let i_m = i.wrapping_shl(n); |
128 | // Mantissa with implicit bit set |
129 | let m_base: u32 = (i_m >> shift_f_lt_i::<u64, f32>()) as u32; |
130 | // The entire lower half of `i` will be truncated (masked portion), plus the |
131 | // next `EXP_BITS` bits. |
132 | let adj = ((i_m >> f32::EXP_BITS) | i_m & 0xFFFF) as u32; |
133 | let m = m_adj::<f32>(m_base, adj); |
134 | let e = if i == 0 { 0 } else { exp::<u64, f32>(n) - 1 }; |
135 | repr::<f32>(e, m) |
136 | } |
137 | |
138 | pub fn u64_to_f64_bits(i: u64) -> u64 { |
139 | if i == 0 { |
140 | return 0; |
141 | } |
142 | let n = i.leading_zeros(); |
143 | // Mantissa with implicit bit set |
144 | let m_base = (i << n) >> f64::EXP_BITS; |
145 | let adj = (i << n) << (f64::SIG_BITS + 1); |
146 | let m = m_adj::<f64>(m_base, adj); |
147 | let e = exp::<u64, f64>(n) - 1; |
148 | repr::<f64>(e, m) |
149 | } |
150 | |
151 | #[cfg (f128_enabled)] |
152 | pub fn u64_to_f128_bits(i: u64) -> u128 { |
153 | if i == 0 { |
154 | return 0; |
155 | } |
156 | let n = i.leading_zeros(); |
157 | // Mantissa with implicit bit set |
158 | let m = (i as u128) << shift_f_gt_i::<u64, f128>(n); |
159 | let e = exp::<u64, f128>(n) - 1; |
160 | repr::<f128>(e, m) |
161 | } |
162 | |
163 | pub fn u128_to_f32_bits(i: u128) -> u32 { |
164 | let n = i.leading_zeros(); |
165 | let i_m = i.wrapping_shl(n); // Mantissa, shifted so the first bit is nonzero |
166 | let m_base: u32 = (i_m >> shift_f_lt_i::<u128, f32>()) as u32; |
167 | |
168 | // Within the upper `F::BITS`, everything except for the signifcand |
169 | // gets truncated |
170 | let d1: u32 = (i_m >> (u128::BITS - f32::BITS - f32::SIG_BITS - 1)).cast(); |
171 | |
172 | // The entire rest of `i_m` gets truncated. Zero the upper `F::BITS` then just |
173 | // check if it is nonzero. |
174 | let d2: u32 = (i_m << f32::BITS >> f32::BITS != 0).into(); |
175 | let adj = d1 | d2; |
176 | |
177 | // Mantissa with implicit bit set |
178 | let m = m_adj::<f32>(m_base, adj); |
179 | let e = if i == 0 { 0 } else { exp::<u128, f32>(n) - 1 }; |
180 | repr::<f32>(e, m) |
181 | } |
182 | |
183 | pub fn u128_to_f64_bits(i: u128) -> u64 { |
184 | let n = i.leading_zeros(); |
185 | let i_m = i.wrapping_shl(n); |
186 | // Mantissa with implicit bit set |
187 | let m_base: u64 = (i_m >> shift_f_lt_i::<u128, f64>()) as u64; |
188 | // The entire lower half of `i` will be truncated (masked portion), plus the |
189 | // next `EXP_BITS` bits. |
190 | let adj = ((i_m >> f64::EXP_BITS) | i_m & 0xFFFF_FFFF) as u64; |
191 | let m = m_adj::<f64>(m_base, adj); |
192 | let e = if i == 0 { 0 } else { exp::<u128, f64>(n) - 1 }; |
193 | repr::<f64>(e, m) |
194 | } |
195 | |
196 | #[cfg (f128_enabled)] |
197 | pub fn u128_to_f128_bits(i: u128) -> u128 { |
198 | if i == 0 { |
199 | return 0; |
200 | } |
201 | let n = i.leading_zeros(); |
202 | // Mantissa with implicit bit set |
203 | let m_base = (i << n) >> f128::EXP_BITS; |
204 | let adj = (i << n) << (f128::SIG_BITS + 1); |
205 | let m = m_adj::<f128>(m_base, adj); |
206 | let e = exp::<u128, f128>(n) - 1; |
207 | repr::<f128>(e, m) |
208 | } |
209 | } |
210 | |
211 | // Conversions from unsigned integers to floats. |
212 | intrinsics! { |
213 | #[arm_aeabi_alias = __aeabi_ui2f] |
214 | pub extern "C" fn __floatunsisf(i: u32) -> f32 { |
215 | f32::from_bits(int_to_float::u32_to_f32_bits(i)) |
216 | } |
217 | |
218 | #[arm_aeabi_alias = __aeabi_ui2d] |
219 | pub extern "C" fn __floatunsidf(i: u32) -> f64 { |
220 | f64::from_bits(int_to_float::u32_to_f64_bits(i)) |
221 | } |
222 | |
223 | #[arm_aeabi_alias = __aeabi_ul2f] |
224 | pub extern "C" fn __floatundisf(i: u64) -> f32 { |
225 | f32::from_bits(int_to_float::u64_to_f32_bits(i)) |
226 | } |
227 | |
228 | #[arm_aeabi_alias = __aeabi_ul2d] |
229 | pub extern "C" fn __floatundidf(i: u64) -> f64 { |
230 | f64::from_bits(int_to_float::u64_to_f64_bits(i)) |
231 | } |
232 | |
233 | #[cfg_attr(target_os = "uefi" , unadjusted_on_win64)] |
234 | pub extern "C" fn __floatuntisf(i: u128) -> f32 { |
235 | f32::from_bits(int_to_float::u128_to_f32_bits(i)) |
236 | } |
237 | |
238 | #[cfg_attr(target_os = "uefi" , unadjusted_on_win64)] |
239 | pub extern "C" fn __floatuntidf(i: u128) -> f64 { |
240 | f64::from_bits(int_to_float::u128_to_f64_bits(i)) |
241 | } |
242 | |
243 | #[ppc_alias = __floatunsikf] |
244 | #[cfg (f128_enabled)] |
245 | pub extern "C" fn __floatunsitf(i: u32) -> f128 { |
246 | f128::from_bits(int_to_float::u32_to_f128_bits(i)) |
247 | } |
248 | |
249 | #[ppc_alias = __floatundikf] |
250 | #[cfg (f128_enabled)] |
251 | pub extern "C" fn __floatunditf(i: u64) -> f128 { |
252 | f128::from_bits(int_to_float::u64_to_f128_bits(i)) |
253 | } |
254 | |
255 | #[ppc_alias = __floatuntikf] |
256 | #[cfg (f128_enabled)] |
257 | pub extern "C" fn __floatuntitf(i: u128) -> f128 { |
258 | f128::from_bits(int_to_float::u128_to_f128_bits(i)) |
259 | } |
260 | } |
261 | |
262 | // Conversions from signed integers to floats. |
263 | intrinsics! { |
264 | #[arm_aeabi_alias = __aeabi_i2f] |
265 | pub extern "C" fn __floatsisf(i: i32) -> f32 { |
266 | int_to_float::signed(i, int_to_float::u32_to_f32_bits) |
267 | } |
268 | |
269 | #[arm_aeabi_alias = __aeabi_i2d] |
270 | pub extern "C" fn __floatsidf(i: i32) -> f64 { |
271 | int_to_float::signed(i, int_to_float::u32_to_f64_bits) |
272 | } |
273 | |
274 | #[arm_aeabi_alias = __aeabi_l2f] |
275 | pub extern "C" fn __floatdisf(i: i64) -> f32 { |
276 | int_to_float::signed(i, int_to_float::u64_to_f32_bits) |
277 | } |
278 | |
279 | #[arm_aeabi_alias = __aeabi_l2d] |
280 | pub extern "C" fn __floatdidf(i: i64) -> f64 { |
281 | int_to_float::signed(i, int_to_float::u64_to_f64_bits) |
282 | } |
283 | |
284 | #[cfg_attr(target_os = "uefi" , unadjusted_on_win64)] |
285 | pub extern "C" fn __floattisf(i: i128) -> f32 { |
286 | int_to_float::signed(i, int_to_float::u128_to_f32_bits) |
287 | } |
288 | |
289 | #[cfg_attr(target_os = "uefi" , unadjusted_on_win64)] |
290 | pub extern "C" fn __floattidf(i: i128) -> f64 { |
291 | int_to_float::signed(i, int_to_float::u128_to_f64_bits) |
292 | } |
293 | |
294 | #[ppc_alias = __floatsikf] |
295 | #[cfg (f128_enabled)] |
296 | pub extern "C" fn __floatsitf(i: i32) -> f128 { |
297 | int_to_float::signed(i, int_to_float::u32_to_f128_bits) |
298 | } |
299 | |
300 | #[ppc_alias = __floatdikf] |
301 | #[cfg (f128_enabled)] |
302 | pub extern "C" fn __floatditf(i: i64) -> f128 { |
303 | int_to_float::signed(i, int_to_float::u64_to_f128_bits) |
304 | } |
305 | |
306 | #[ppc_alias = __floattikf] |
307 | #[cfg (f128_enabled)] |
308 | pub extern "C" fn __floattitf(i: i128) -> f128 { |
309 | int_to_float::signed(i, int_to_float::u128_to_f128_bits) |
310 | } |
311 | } |
312 | |
313 | /// Generic float to unsigned int conversions. |
314 | fn float_to_unsigned_int<F, U>(f: F) -> U |
315 | where |
316 | F: Float, |
317 | U: Int<UnsignedInt = U>, |
318 | F::Int: CastInto<U>, |
319 | F::Int: CastFrom<u32>, |
320 | F::Int: CastInto<U::UnsignedInt>, |
321 | u32: CastFrom<F::Int>, |
322 | { |
323 | float_to_int_inner::<F, U, _, _>(fbits:f.to_bits(), |i: U| i, || U::MAX) |
324 | } |
325 | |
326 | /// Generic float to signed int conversions. |
327 | fn float_to_signed_int<F, I>(f: F) -> I |
328 | where |
329 | F: Float, |
330 | I: Int + Neg<Output = I>, |
331 | I::UnsignedInt: Int, |
332 | F::Int: CastInto<I::UnsignedInt>, |
333 | F::Int: CastFrom<u32>, |
334 | u32: CastFrom<F::Int>, |
335 | { |
336 | float_to_int_inner::<F, I, _, _>( |
337 | fbits:f.to_bits() & !F::SIGN_MASK, |
338 | |i: I| if f.is_sign_negative() { -i } else { i }, |
339 | || if f.is_sign_negative() { I::MIN } else { I::MAX }, |
340 | ) |
341 | } |
342 | |
343 | /// Float to int conversions, generic for both signed and unsigned. |
344 | /// |
345 | /// Parameters: |
346 | /// - `fbits`: `abg(f)` bitcasted to an integer. |
347 | /// - `map_inbounds`: apply this transformation to integers that are within range (add the sign back). |
348 | /// - `out_of_bounds`: return value when out of range for `I`. |
349 | fn float_to_int_inner<F, I, FnFoo, FnOob>( |
350 | fbits: F::Int, |
351 | map_inbounds: FnFoo, |
352 | out_of_bounds: FnOob, |
353 | ) -> I |
354 | where |
355 | F: Float, |
356 | I: Int, |
357 | FnFoo: FnOnce(I) -> I, |
358 | FnOob: FnOnce() -> I, |
359 | I::UnsignedInt: Int, |
360 | F::Int: CastInto<I::UnsignedInt>, |
361 | F::Int: CastFrom<u32>, |
362 | u32: CastFrom<F::Int>, |
363 | { |
364 | let int_max_exp = F::EXP_BIAS + I::MAX.ilog2() + 1; |
365 | let foobar = F::EXP_BIAS + I::UnsignedInt::BITS - 1; |
366 | |
367 | if fbits < F::ONE.to_bits() { |
368 | // < 0 gets rounded to 0 |
369 | I::ZERO |
370 | } else if fbits < F::Int::cast_from(int_max_exp) << F::SIG_BITS { |
371 | // >= 1, < integer max |
372 | let m_base = if I::UnsignedInt::BITS >= F::Int::BITS { |
373 | I::UnsignedInt::cast_from(fbits) << (I::BITS - F::SIG_BITS - 1) |
374 | } else { |
375 | I::UnsignedInt::cast_from(fbits >> (F::SIG_BITS - I::BITS + 1)) |
376 | }; |
377 | |
378 | // Set the implicit 1-bit. |
379 | let m: I::UnsignedInt = (I::UnsignedInt::ONE << (I::BITS - 1)) | m_base; |
380 | |
381 | // Shift based on the exponent and bias. |
382 | let s: u32 = (foobar) - u32::cast_from(fbits >> F::SIG_BITS); |
383 | |
384 | let unsigned = m >> s; |
385 | map_inbounds(I::from_unsigned(unsigned)) |
386 | } else if fbits <= F::EXP_MASK { |
387 | // >= max (incl. inf) |
388 | out_of_bounds() |
389 | } else { |
390 | I::ZERO |
391 | } |
392 | } |
393 | |
394 | // Conversions from floats to unsigned integers. |
395 | intrinsics! { |
396 | #[arm_aeabi_alias = __aeabi_f2uiz] |
397 | pub extern "C" fn __fixunssfsi(f: f32) -> u32 { |
398 | float_to_unsigned_int(f) |
399 | } |
400 | |
401 | #[arm_aeabi_alias = __aeabi_f2ulz] |
402 | pub extern "C" fn __fixunssfdi(f: f32) -> u64 { |
403 | float_to_unsigned_int(f) |
404 | } |
405 | |
406 | pub extern "C" fn __fixunssfti(f: f32) -> u128 { |
407 | float_to_unsigned_int(f) |
408 | } |
409 | |
410 | #[arm_aeabi_alias = __aeabi_d2uiz] |
411 | pub extern "C" fn __fixunsdfsi(f: f64) -> u32 { |
412 | float_to_unsigned_int(f) |
413 | } |
414 | |
415 | #[arm_aeabi_alias = __aeabi_d2ulz] |
416 | pub extern "C" fn __fixunsdfdi(f: f64) -> u64 { |
417 | float_to_unsigned_int(f) |
418 | } |
419 | |
420 | pub extern "C" fn __fixunsdfti(f: f64) -> u128 { |
421 | float_to_unsigned_int(f) |
422 | } |
423 | |
424 | #[ppc_alias = __fixunskfsi] |
425 | #[cfg (f128_enabled)] |
426 | pub extern "C" fn __fixunstfsi(f: f128) -> u32 { |
427 | float_to_unsigned_int(f) |
428 | } |
429 | |
430 | #[ppc_alias = __fixunskfdi] |
431 | #[cfg (f128_enabled)] |
432 | pub extern "C" fn __fixunstfdi(f: f128) -> u64 { |
433 | float_to_unsigned_int(f) |
434 | } |
435 | |
436 | #[ppc_alias = __fixunskfti] |
437 | #[cfg (f128_enabled)] |
438 | pub extern "C" fn __fixunstfti(f: f128) -> u128 { |
439 | float_to_unsigned_int(f) |
440 | } |
441 | } |
442 | |
443 | // Conversions from floats to signed integers. |
444 | intrinsics! { |
445 | #[arm_aeabi_alias = __aeabi_f2iz] |
446 | pub extern "C" fn __fixsfsi(f: f32) -> i32 { |
447 | float_to_signed_int(f) |
448 | } |
449 | |
450 | #[arm_aeabi_alias = __aeabi_f2lz] |
451 | pub extern "C" fn __fixsfdi(f: f32) -> i64 { |
452 | float_to_signed_int(f) |
453 | } |
454 | |
455 | pub extern "C" fn __fixsfti(f: f32) -> i128 { |
456 | float_to_signed_int(f) |
457 | } |
458 | |
459 | #[arm_aeabi_alias = __aeabi_d2iz] |
460 | pub extern "C" fn __fixdfsi(f: f64) -> i32 { |
461 | float_to_signed_int(f) |
462 | } |
463 | |
464 | #[arm_aeabi_alias = __aeabi_d2lz] |
465 | pub extern "C" fn __fixdfdi(f: f64) -> i64 { |
466 | float_to_signed_int(f) |
467 | } |
468 | |
469 | pub extern "C" fn __fixdfti(f: f64) -> i128 { |
470 | float_to_signed_int(f) |
471 | } |
472 | |
473 | #[ppc_alias = __fixkfsi] |
474 | #[cfg (f128_enabled)] |
475 | pub extern "C" fn __fixtfsi(f: f128) -> i32 { |
476 | float_to_signed_int(f) |
477 | } |
478 | |
479 | #[ppc_alias = __fixkfdi] |
480 | #[cfg (f128_enabled)] |
481 | pub extern "C" fn __fixtfdi(f: f128) -> i64 { |
482 | float_to_signed_int(f) |
483 | } |
484 | |
485 | #[ppc_alias = __fixkfti] |
486 | #[cfg (f128_enabled)] |
487 | pub extern "C" fn __fixtfti(f: f128) -> i128 { |
488 | float_to_signed_int(f) |
489 | } |
490 | } |
491 | |