1 | //===-- fp_div_impl.inc - Floating point division -----------------*- C -*-===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file implements soft-float division with the IEEE-754 default |
10 | // rounding (to nearest, ties to even). |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "fp_lib.h" |
15 | |
16 | // The __divXf3__ function implements Newton-Raphson floating point division. |
17 | // It uses 3 iterations for float32, 4 for float64 and 5 for float128, |
18 | // respectively. Due to number of significant bits being roughly doubled |
19 | // every iteration, the two modes are supported: N full-width iterations (as |
20 | // it is done for float32 by default) and (N-1) half-width iteration plus one |
21 | // final full-width iteration. It is expected that half-width integer |
22 | // operations (w.r.t rep_t size) can be performed faster for some hardware but |
23 | // they require error estimations to be computed separately due to larger |
24 | // computational errors caused by truncating intermediate results. |
25 | |
26 | // Half the bit-size of rep_t |
27 | #define HW (typeWidth / 2) |
28 | // rep_t-sized bitmask with lower half of bits set to ones |
29 | #define loMask (REP_C(-1) >> HW) |
30 | |
31 | #if NUMBER_OF_FULL_ITERATIONS < 1 |
32 | #error At least one full iteration is required |
33 | #endif |
34 | |
35 | static __inline fp_t __divXf3__(fp_t a, fp_t b) { |
36 | |
37 | const unsigned int aExponent = toRep(x: a) >> significandBits & maxExponent; |
38 | const unsigned int bExponent = toRep(x: b) >> significandBits & maxExponent; |
39 | const rep_t quotientSign = (toRep(x: a) ^ toRep(x: b)) & signBit; |
40 | |
41 | rep_t aSignificand = toRep(x: a) & significandMask; |
42 | rep_t bSignificand = toRep(x: b) & significandMask; |
43 | int scale = 0; |
44 | |
45 | // Detect if a or b is zero, denormal, infinity, or NaN. |
46 | if (aExponent - 1U >= maxExponent - 1U || |
47 | bExponent - 1U >= maxExponent - 1U) { |
48 | |
49 | const rep_t aAbs = toRep(x: a) & absMask; |
50 | const rep_t bAbs = toRep(x: b) & absMask; |
51 | |
52 | // NaN / anything = qNaN |
53 | if (aAbs > infRep) |
54 | return fromRep(x: toRep(x: a) | quietBit); |
55 | // anything / NaN = qNaN |
56 | if (bAbs > infRep) |
57 | return fromRep(x: toRep(x: b) | quietBit); |
58 | |
59 | if (aAbs == infRep) { |
60 | // infinity / infinity = NaN |
61 | if (bAbs == infRep) |
62 | return fromRep(qnanRep); |
63 | // infinity / anything else = +/- infinity |
64 | else |
65 | return fromRep(x: aAbs | quotientSign); |
66 | } |
67 | |
68 | // anything else / infinity = +/- 0 |
69 | if (bAbs == infRep) |
70 | return fromRep(x: quotientSign); |
71 | |
72 | if (!aAbs) { |
73 | // zero / zero = NaN |
74 | if (!bAbs) |
75 | return fromRep(qnanRep); |
76 | // zero / anything else = +/- zero |
77 | else |
78 | return fromRep(x: quotientSign); |
79 | } |
80 | // anything else / zero = +/- infinity |
81 | if (!bAbs) |
82 | return fromRep(infRep | quotientSign); |
83 | |
84 | // One or both of a or b is denormal. The other (if applicable) is a |
85 | // normal number. Renormalize one or both of a and b, and set scale to |
86 | // include the necessary exponent adjustment. |
87 | if (aAbs < implicitBit) |
88 | scale += normalize(significand: &aSignificand); |
89 | if (bAbs < implicitBit) |
90 | scale -= normalize(significand: &bSignificand); |
91 | } |
92 | |
93 | // Set the implicit significand bit. If we fell through from the |
94 | // denormal path it was already set by normalize( ), but setting it twice |
95 | // won't hurt anything. |
96 | aSignificand |= implicitBit; |
97 | bSignificand |= implicitBit; |
98 | |
99 | int writtenExponent = (aExponent - bExponent + scale) + exponentBias; |
100 | |
101 | const rep_t b_UQ1 = bSignificand << (typeWidth - significandBits - 1); |
102 | |
103 | // Align the significand of b as a UQ1.(n-1) fixed-point number in the range |
104 | // [1.0, 2.0) and get a UQ0.n approximate reciprocal using a small minimax |
105 | // polynomial approximation: x0 = 3/4 + 1/sqrt(2) - b/2. |
106 | // The max error for this approximation is achieved at endpoints, so |
107 | // abs(x0(b) - 1/b) <= abs(x0(1) - 1/1) = 3/4 - 1/sqrt(2) = 0.04289..., |
108 | // which is about 4.5 bits. |
109 | // The initial approximation is between x0(1.0) = 0.9571... and x0(2.0) = 0.4571... |
110 | |
111 | // Then, refine the reciprocal estimate using a quadratically converging |
112 | // Newton-Raphson iteration: |
113 | // x_{n+1} = x_n * (2 - x_n * b) |
114 | // |
115 | // Let b be the original divisor considered "in infinite precision" and |
116 | // obtained from IEEE754 representation of function argument (with the |
117 | // implicit bit set). Corresponds to rep_t-sized b_UQ1 represented in |
118 | // UQ1.(W-1). |
119 | // |
120 | // Let b_hw be an infinitely precise number obtained from the highest (HW-1) |
121 | // bits of divisor significand (with the implicit bit set). Corresponds to |
122 | // half_rep_t-sized b_UQ1_hw represented in UQ1.(HW-1) that is a **truncated** |
123 | // version of b_UQ1. |
124 | // |
125 | // Let e_n := x_n - 1/b_hw |
126 | // E_n := x_n - 1/b |
127 | // abs(E_n) <= abs(e_n) + (1/b_hw - 1/b) |
128 | // = abs(e_n) + (b - b_hw) / (b*b_hw) |
129 | // <= abs(e_n) + 2 * 2^-HW |
130 | |
131 | // rep_t-sized iterations may be slower than the corresponding half-width |
132 | // variant depending on the handware and whether single/double/quad precision |
133 | // is selected. |
134 | // NB: Using half-width iterations increases computation errors due to |
135 | // rounding, so error estimations have to be computed taking the selected |
136 | // mode into account! |
137 | #if NUMBER_OF_HALF_ITERATIONS > 0 |
138 | // Starting with (n-1) half-width iterations |
139 | const half_rep_t b_UQ1_hw = bSignificand >> (significandBits + 1 - HW); |
140 | |
141 | // C is (3/4 + 1/sqrt(2)) - 1 truncated to W0 fractional bits as UQ0.HW |
142 | // with W0 being either 16 or 32 and W0 <= HW. |
143 | // That is, C is the aforementioned 3/4 + 1/sqrt(2) constant (from which |
144 | // b/2 is subtracted to obtain x0) wrapped to [0, 1) range. |
145 | #if defined(SINGLE_PRECISION) |
146 | // Use 16-bit initial estimation in case we are using half-width iterations |
147 | // for float32 division. This is expected to be useful for some 16-bit |
148 | // targets. Not used by default as it requires performing more work during |
149 | // rounding and would hardly help on regular 32- or 64-bit targets. |
150 | const half_rep_t C_hw = HALF_REP_C(0x7504); |
151 | #else |
152 | // HW is at least 32. Shifting into the highest bits if needed. |
153 | const half_rep_t C_hw = HALF_REP_C(0x7504F333) << (HW - 32); |
154 | #endif |
155 | |
156 | // b >= 1, thus an upper bound for 3/4 + 1/sqrt(2) - b/2 is about 0.9572, |
157 | // so x0 fits to UQ0.HW without wrapping. |
158 | half_rep_t x_UQ0_hw = C_hw - (b_UQ1_hw /* exact b_hw/2 as UQ0.HW */); |
159 | // An e_0 error is comprised of errors due to |
160 | // * x0 being an inherently imprecise first approximation of 1/b_hw |
161 | // * C_hw being some (irrational) number **truncated** to W0 bits |
162 | // Please note that e_0 is calculated against the infinitely precise |
163 | // reciprocal of b_hw (that is, **truncated** version of b). |
164 | // |
165 | // e_0 <= 3/4 - 1/sqrt(2) + 2^-W0 |
166 | |
167 | // By construction, 1 <= b < 2 |
168 | // f(x) = x * (2 - b*x) = 2*x - b*x^2 |
169 | // f'(x) = 2 * (1 - b*x) |
170 | // |
171 | // On the [0, 1] interval, f(0) = 0, |
172 | // then it increses until f(1/b) = 1 / b, maximum on (0, 1), |
173 | // then it decreses to f(1) = 2 - b |
174 | // |
175 | // Let g(x) = x - f(x) = b*x^2 - x. |
176 | // On (0, 1/b), g(x) < 0 <=> f(x) > x |
177 | // On (1/b, 1], g(x) > 0 <=> f(x) < x |
178 | // |
179 | // For half-width iterations, b_hw is used instead of b. |
180 | REPEAT_N_TIMES(NUMBER_OF_HALF_ITERATIONS, { |
181 | // corr_UQ1_hw can be **larger** than 2 - b_hw*x by at most 1*Ulp |
182 | // of corr_UQ1_hw. |
183 | // "0.0 - (...)" is equivalent to "2.0 - (...)" in UQ1.(HW-1). |
184 | // On the other hand, corr_UQ1_hw should not overflow from 2.0 to 0.0 provided |
185 | // no overflow occurred earlier: ((rep_t)x_UQ0_hw * b_UQ1_hw >> HW) is |
186 | // expected to be strictly positive because b_UQ1_hw has its highest bit set |
187 | // and x_UQ0_hw should be rather large (it converges to 1/2 < 1/b_hw <= 1). |
188 | half_rep_t corr_UQ1_hw = 0 - ((rep_t)x_UQ0_hw * b_UQ1_hw >> HW); |
189 | |
190 | // Now, we should multiply UQ0.HW and UQ1.(HW-1) numbers, naturally |
191 | // obtaining an UQ1.(HW-1) number and proving its highest bit could be |
192 | // considered to be 0 to be able to represent it in UQ0.HW. |
193 | // From the above analysis of f(x), if corr_UQ1_hw would be represented |
194 | // without any intermediate loss of precision (that is, in twice_rep_t) |
195 | // x_UQ0_hw could be at most [1.]000... if b_hw is exactly 1.0 and strictly |
196 | // less otherwise. On the other hand, to obtain [1.]000..., one have to pass |
197 | // 1/b_hw == 1.0 to f(x), so this cannot occur at all without overflow (due |
198 | // to 1.0 being not representable as UQ0.HW). |
199 | // The fact corr_UQ1_hw was virtually round up (due to result of |
200 | // multiplication being **first** truncated, then negated - to improve |
201 | // error estimations) can increase x_UQ0_hw by up to 2*Ulp of x_UQ0_hw. |
202 | x_UQ0_hw = (rep_t)x_UQ0_hw * corr_UQ1_hw >> (HW - 1); |
203 | // Now, either no overflow occurred or x_UQ0_hw is 0 or 1 in its half_rep_t |
204 | // representation. In the latter case, x_UQ0_hw will be either 0 or 1 after |
205 | // any number of iterations, so just subtract 2 from the reciprocal |
206 | // approximation after last iteration. |
207 | |
208 | // In infinite precision, with 0 <= eps1, eps2 <= U = 2^-HW: |
209 | // corr_UQ1_hw = 2 - (1/b_hw + e_n) * b_hw + 2*eps1 |
210 | // = 1 - e_n * b_hw + 2*eps1 |
211 | // x_UQ0_hw = (1/b_hw + e_n) * (1 - e_n*b_hw + 2*eps1) - eps2 |
212 | // = 1/b_hw - e_n + 2*eps1/b_hw + e_n - e_n^2*b_hw + 2*e_n*eps1 - eps2 |
213 | // = 1/b_hw + 2*eps1/b_hw - e_n^2*b_hw + 2*e_n*eps1 - eps2 |
214 | // e_{n+1} = -e_n^2*b_hw + 2*eps1/b_hw + 2*e_n*eps1 - eps2 |
215 | // = 2*e_n*eps1 - (e_n^2*b_hw + eps2) + 2*eps1/b_hw |
216 | // \------ >0 -------/ \-- >0 ---/ |
217 | // abs(e_{n+1}) <= 2*abs(e_n)*U + max(2*e_n^2 + U, 2 * U) |
218 | }) |
219 | // For initial half-width iterations, U = 2^-HW |
220 | // Let abs(e_n) <= u_n * U, |
221 | // then abs(e_{n+1}) <= 2 * u_n * U^2 + max(2 * u_n^2 * U^2 + U, 2 * U) |
222 | // u_{n+1} <= 2 * u_n * U + max(2 * u_n^2 * U + 1, 2) |
223 | |
224 | // Account for possible overflow (see above). For an overflow to occur for the |
225 | // first time, for "ideal" corr_UQ1_hw (that is, without intermediate |
226 | // truncation), the result of x_UQ0_hw * corr_UQ1_hw should be either maximum |
227 | // value representable in UQ0.HW or less by 1. This means that 1/b_hw have to |
228 | // be not below that value (see g(x) above), so it is safe to decrement just |
229 | // once after the final iteration. On the other hand, an effective value of |
230 | // divisor changes after this point (from b_hw to b), so adjust here. |
231 | x_UQ0_hw -= 1U; |
232 | rep_t x_UQ0 = (rep_t)x_UQ0_hw << HW; |
233 | x_UQ0 -= 1U; |
234 | |
235 | #else |
236 | // C is (3/4 + 1/sqrt(2)) - 1 truncated to 32 fractional bits as UQ0.n |
237 | const rep_t C = REP_C(0x7504F333) << (typeWidth - 32); |
238 | rep_t x_UQ0 = C - b_UQ1; |
239 | // E_0 <= 3/4 - 1/sqrt(2) + 2 * 2^-32 |
240 | #endif |
241 | |
242 | // Error estimations for full-precision iterations are calculated just |
243 | // as above, but with U := 2^-W and taking extra decrementing into account. |
244 | // We need at least one such iteration. |
245 | |
246 | #ifdef USE_NATIVE_FULL_ITERATIONS |
247 | REPEAT_N_TIMES(NUMBER_OF_FULL_ITERATIONS, { |
248 | rep_t corr_UQ1 = 0 - ((twice_rep_t)x_UQ0 * b_UQ1 >> typeWidth); |
249 | x_UQ0 = (twice_rep_t)x_UQ0 * corr_UQ1 >> (typeWidth - 1); |
250 | }) |
251 | #else |
252 | #if NUMBER_OF_FULL_ITERATIONS != 1 |
253 | #error Only a single emulated full iteration is supported |
254 | #endif |
255 | #if !(NUMBER_OF_HALF_ITERATIONS > 0) |
256 | // Cannot normally reach here: only one full-width iteration is requested and |
257 | // the total number of iterations should be at least 3 even for float32. |
258 | #error Check NUMBER_OF_HALF_ITERATIONS, NUMBER_OF_FULL_ITERATIONS and USE_NATIVE_FULL_ITERATIONS. |
259 | #endif |
260 | // Simulating operations on a twice_rep_t to perform a single final full-width |
261 | // iteration. Using ad-hoc multiplication implementations to take advantage |
262 | // of particular structure of operands. |
263 | rep_t blo = b_UQ1 & loMask; |
264 | // x_UQ0 = x_UQ0_hw * 2^HW - 1 |
265 | // x_UQ0 * b_UQ1 = (x_UQ0_hw * 2^HW) * (b_UQ1_hw * 2^HW + blo) - b_UQ1 |
266 | // |
267 | // <--- higher half ---><--- lower half ---> |
268 | // [x_UQ0_hw * b_UQ1_hw] |
269 | // + [ x_UQ0_hw * blo ] |
270 | // - [ b_UQ1 ] |
271 | // = [ result ][.... discarded ...] |
272 | rep_t corr_UQ1 = 0U - ( (rep_t)x_UQ0_hw * b_UQ1_hw |
273 | + ((rep_t)x_UQ0_hw * blo >> HW) |
274 | - REP_C(1)); // account for *possible* carry |
275 | rep_t lo_corr = corr_UQ1 & loMask; |
276 | rep_t hi_corr = corr_UQ1 >> HW; |
277 | // x_UQ0 * corr_UQ1 = (x_UQ0_hw * 2^HW) * (hi_corr * 2^HW + lo_corr) - corr_UQ1 |
278 | x_UQ0 = ((rep_t)x_UQ0_hw * hi_corr << 1) |
279 | + ((rep_t)x_UQ0_hw * lo_corr >> (HW - 1)) |
280 | - REP_C(2); // 1 to account for the highest bit of corr_UQ1 can be 1 |
281 | // 1 to account for possible carry |
282 | // Just like the case of half-width iterations but with possibility |
283 | // of overflowing by one extra Ulp of x_UQ0. |
284 | x_UQ0 -= 1U; |
285 | // ... and then traditional fixup by 2 should work |
286 | |
287 | // On error estimation: |
288 | // abs(E_{N-1}) <= (u_{N-1} + 2 /* due to conversion e_n -> E_n */) * 2^-HW |
289 | // + (2^-HW + 2^-W)) |
290 | // abs(E_{N-1}) <= (u_{N-1} + 3.01) * 2^-HW |
291 | |
292 | // Then like for the half-width iterations: |
293 | // With 0 <= eps1, eps2 < 2^-W |
294 | // E_N = 4 * E_{N-1} * eps1 - (E_{N-1}^2 * b + 4 * eps2) + 4 * eps1 / b |
295 | // abs(E_N) <= 2^-W * [ 4 * abs(E_{N-1}) + max(2 * abs(E_{N-1})^2 * 2^W + 4, 8)) ] |
296 | // abs(E_N) <= 2^-W * [ 4 * (u_{N-1} + 3.01) * 2^-HW + max(4 + 2 * (u_{N-1} + 3.01)^2, 8) ] |
297 | #endif |
298 | |
299 | // Finally, account for possible overflow, as explained above. |
300 | x_UQ0 -= 2U; |
301 | |
302 | // u_n for different precisions (with N-1 half-width iterations): |
303 | // W0 is the precision of C |
304 | // u_0 = (3/4 - 1/sqrt(2) + 2^-W0) * 2^HW |
305 | |
306 | // Estimated with bc: |
307 | // define half1(un) { return 2.0 * (un + un^2) / 2.0^hw + 1.0; } |
308 | // define half2(un) { return 2.0 * un / 2.0^hw + 2.0; } |
309 | // define full1(un) { return 4.0 * (un + 3.01) / 2.0^hw + 2.0 * (un + 3.01)^2 + 4.0; } |
310 | // define full2(un) { return 4.0 * (un + 3.01) / 2.0^hw + 8.0; } |
311 | |
312 | // | f32 (0 + 3) | f32 (2 + 1) | f64 (3 + 1) | f128 (4 + 1) |
313 | // u_0 | < 184224974 | < 2812.1 | < 184224974 | < 791240234244348797 |
314 | // u_1 | < 15804007 | < 242.7 | < 15804007 | < 67877681371350440 |
315 | // u_2 | < 116308 | < 2.81 | < 116308 | < 499533100252317 |
316 | // u_3 | < 7.31 | | < 7.31 | < 27054456580 |
317 | // u_4 | | | | < 80.4 |
318 | // Final (U_N) | same as u_3 | < 72 | < 218 | < 13920 |
319 | |
320 | // Add 2 to U_N due to final decrement. |
321 | |
322 | #if defined(SINGLE_PRECISION) && NUMBER_OF_HALF_ITERATIONS == 2 && NUMBER_OF_FULL_ITERATIONS == 1 |
323 | #define RECIPROCAL_PRECISION REP_C(74) |
324 | #elif defined(SINGLE_PRECISION) && NUMBER_OF_HALF_ITERATIONS == 0 && NUMBER_OF_FULL_ITERATIONS == 3 |
325 | #define RECIPROCAL_PRECISION REP_C(10) |
326 | #elif defined(DOUBLE_PRECISION) && NUMBER_OF_HALF_ITERATIONS == 3 && NUMBER_OF_FULL_ITERATIONS == 1 |
327 | #define RECIPROCAL_PRECISION REP_C(220) |
328 | #elif defined(QUAD_PRECISION) && NUMBER_OF_HALF_ITERATIONS == 4 && NUMBER_OF_FULL_ITERATIONS == 1 |
329 | #define RECIPROCAL_PRECISION REP_C(13922) |
330 | #else |
331 | #error Invalid number of iterations |
332 | #endif |
333 | |
334 | // Suppose 1/b - P * 2^-W < x < 1/b + P * 2^-W |
335 | x_UQ0 -= RECIPROCAL_PRECISION; |
336 | // Now 1/b - (2*P) * 2^-W < x < 1/b |
337 | // FIXME Is x_UQ0 still >= 0.5? |
338 | |
339 | rep_t quotient_UQ1, dummy; |
340 | wideMultiply(a: x_UQ0, b: aSignificand << 1, hi: "ient_UQ1, lo: &dummy); |
341 | // Now, a/b - 4*P * 2^-W < q < a/b for q=<quotient_UQ1:dummy> in UQ1.(SB+1+W). |
342 | |
343 | // quotient_UQ1 is in [0.5, 2.0) as UQ1.(SB+1), |
344 | // adjust it to be in [1.0, 2.0) as UQ1.SB. |
345 | rep_t residualLo; |
346 | if (quotient_UQ1 < (implicitBit << 1)) { |
347 | // Highest bit is 0, so just reinterpret quotient_UQ1 as UQ1.SB, |
348 | // effectively doubling its value as well as its error estimation. |
349 | residualLo = (aSignificand << (significandBits + 1)) - quotient_UQ1 * bSignificand; |
350 | writtenExponent -= 1; |
351 | aSignificand <<= 1; |
352 | } else { |
353 | // Highest bit is 1 (the UQ1.(SB+1) value is in [1, 2)), convert it |
354 | // to UQ1.SB by right shifting by 1. Least significant bit is omitted. |
355 | quotient_UQ1 >>= 1; |
356 | residualLo = (aSignificand << significandBits) - quotient_UQ1 * bSignificand; |
357 | } |
358 | // NB: residualLo is calculated above for the normal result case. |
359 | // It is re-computed on denormal path that is expected to be not so |
360 | // performance-sensitive. |
361 | |
362 | // Now, q cannot be greater than a/b and can differ by at most 8*P * 2^-W + 2^-SB |
363 | // Each NextAfter() increments the floating point value by at least 2^-SB |
364 | // (more, if exponent was incremented). |
365 | // Different cases (<---> is of 2^-SB length, * = a/b that is shown as a midpoint): |
366 | // q |
367 | // | | * | | | | | |
368 | // <---> 2^t |
369 | // | | | | | * | | |
370 | // q |
371 | // To require at most one NextAfter(), an error should be less than 1.5 * 2^-SB. |
372 | // (8*P) * 2^-W + 2^-SB < 1.5 * 2^-SB |
373 | // (8*P) * 2^-W < 0.5 * 2^-SB |
374 | // P < 2^(W-4-SB) |
375 | // Generally, for at most R NextAfter() to be enough, |
376 | // P < (2*R - 1) * 2^(W-4-SB) |
377 | // For f32 (0+3): 10 < 32 (OK) |
378 | // For f32 (2+1): 32 < 74 < 32 * 3, so two NextAfter() are required |
379 | // For f64: 220 < 256 (OK) |
380 | // For f128: 4096 * 3 < 13922 < 4096 * 5 (three NextAfter() are required) |
381 | |
382 | // If we have overflowed the exponent, return infinity |
383 | if (writtenExponent >= maxExponent) |
384 | return fromRep(infRep | quotientSign); |
385 | |
386 | // Now, quotient_UQ1_SB <= the correctly-rounded result |
387 | // and may need taking NextAfter() up to 3 times (see error estimates above) |
388 | // r = a - b * q |
389 | rep_t absResult; |
390 | if (writtenExponent > 0) { |
391 | // Clear the implicit bit |
392 | absResult = quotient_UQ1 & significandMask; |
393 | // Insert the exponent |
394 | absResult |= (rep_t)writtenExponent << significandBits; |
395 | residualLo <<= 1; |
396 | } else { |
397 | // Prevent shift amount from being negative |
398 | if (significandBits + writtenExponent < 0) |
399 | return fromRep(x: quotientSign); |
400 | |
401 | absResult = quotient_UQ1 >> (-writtenExponent + 1); |
402 | |
403 | // multiplied by two to prevent shift amount to be negative |
404 | residualLo = (aSignificand << (significandBits + writtenExponent)) - (absResult * bSignificand << 1); |
405 | } |
406 | |
407 | // Round |
408 | residualLo += absResult & 1; // tie to even |
409 | // The above line conditionally turns the below LT comparison into LTE |
410 | absResult += residualLo > bSignificand; |
411 | #if defined(QUAD_PRECISION) || (defined(SINGLE_PRECISION) && NUMBER_OF_HALF_ITERATIONS > 0) |
412 | // Do not round Infinity to NaN |
413 | absResult += absResult < infRep && residualLo > (2 + 1) * bSignificand; |
414 | #endif |
415 | #if defined(QUAD_PRECISION) |
416 | absResult += absResult < infRep && residualLo > (4 + 1) * bSignificand; |
417 | #endif |
418 | return fromRep(x: absResult | quotientSign); |
419 | } |
420 | |