xref: /plan9/sys/src/ape/lib/ap/stdio/strtod.c (revision 43aadf5e5598b9159a98f64e309c3ae860328a56)
1 #include "fconv.h"
2 
3 /* strtod for IEEE-, VAX-, and IBM-arithmetic machines (dmg).
4  *
5  * This strtod returns a nearest machine number to the input decimal
6  * string (or sets errno to ERANGE).  With IEEE arithmetic, ties are
7  * broken by the IEEE round-even rule.  Otherwise ties are broken by
8  * biased rounding (add half and chop).
9  *
10  * Inspired loosely by William D. Clinger's paper "How to Read Floating
11  * Point Numbers Accurately" [Proc. ACM SIGPLAN '90, pp. 92-101].
12  *
13  * Modifications:
14  *
15  *	1. We only require IEEE, IBM, or VAX double-precision
16  *		arithmetic (not IEEE double-extended).
17  *	2. We get by with floating-point arithmetic in a case that
18  *		Clinger missed -- when we're computing d * 10^n
19  *		for a small integer d and the integer n is not too
20  *		much larger than 22 (the maximum integer k for which
21  *		we can represent 10^k exactly), we may be able to
22  *		compute (d*10^k) * 10^(e-k) with just one roundoff.
23  *	3. Rather than a bit-at-a-time adjustment of the binary
24  *		result in the hard case, we use floating-point
25  *		arithmetic to determine the adjustment to within
26  *		one bit; only in really hard cases do we need to
27  *		compute a second residual.
28  *	4. Because of 3., we don't need a large table of powers of 10
29  *		for ten-to-e (just some small tables, e.g. of 10^k
30  *		for 0 <= k <= 22).
31  */
32 
33 #ifdef RND_PRODQUOT
34 #define rounded_product(a,b) a = rnd_prod(a, b)
35 #define rounded_quotient(a,b) a = rnd_quot(a, b)
36 extern double rnd_prod(double, double), rnd_quot(double, double);
37 #else
38 #define rounded_product(a,b) a *= b
39 #define rounded_quotient(a,b) a /= b
40 #endif
41 
42  static double
ulp(double xarg)43 ulp(double xarg)
44 {
45 	register long L;
46 	Dul a;
47 	Dul x;
48 
49 	x.d = xarg;
50 	L = (word0(x) & Exp_mask) - (P-1)*Exp_msk1;
51 #ifndef Sudden_Underflow
52 	if (L > 0) {
53 #endif
54 #ifdef IBM
55 		L |= Exp_msk1 >> 4;
56 #endif
57 		word0(a) = L;
58 		word1(a) = 0;
59 #ifndef Sudden_Underflow
60 		}
61 	else {
62 		L = -L >> Exp_shift;
63 		if (L < Exp_shift) {
64 			word0(a) = 0x80000 >> L;
65 			word1(a) = 0;
66 			}
67 		else {
68 			word0(a) = 0;
69 			L -= Exp_shift;
70 			word1(a) = L >= 31 ? 1 : 1 << 31 - L;
71 			}
72 		}
73 #endif
74 	return a.d;
75 	}
76 
77  static Bigint *
s2b(CONST char * s,int nd0,int nd,unsigned long y9)78 s2b(CONST char *s, int nd0, int nd, unsigned long y9)
79 {
80 	Bigint *b;
81 	int i, k;
82 	long x, y;
83 
84 	x = (nd + 8) / 9;
85 	for(k = 0, y = 1; x > y; y <<= 1, k++) ;
86 #ifdef Pack_32
87 	b = Balloc(k);
88 	b->x[0] = y9;
89 	b->wds = 1;
90 #else
91 	b = Balloc(k+1);
92 	b->x[0] = y9 & 0xffff;
93 	b->wds = (b->x[1] = y9 >> 16) ? 2 : 1;
94 #endif
95 
96 	i = 9;
97 	if (9 < nd0) {
98 		s += 9;
99 		do b = multadd(b, 10, *s++ - '0');
100 			while(++i < nd0);
101 		s++;
102 		}
103 	else
104 		s += 10;
105 	for(; i < nd; i++)
106 		b = multadd(b, 10, *s++ - '0');
107 	return b;
108 	}
109 
110  static double
b2d(Bigint * a,int * e)111 b2d(Bigint *a, int *e)
112 {
113 	unsigned long *xa, *xa0, w, y, z;
114 	int k;
115 	Dul d;
116 #ifdef VAX
117 	unsigned long d0, d1;
118 #else
119 #define d0 word0(d)
120 #define d1 word1(d)
121 #endif
122 
123 	xa0 = a->x;
124 	xa = xa0 + a->wds;
125 	y = *--xa;
126 #ifdef DEBUG
127 	if (!y) Bug("zero y in b2d");
128 #endif
129 	k = hi0bits(y);
130 	*e = 32 - k;
131 #ifdef Pack_32
132 	if (k < Ebits) {
133 		d0 = Exp_1 | y >> Ebits - k;
134 		w = xa > xa0 ? *--xa : 0;
135 		d1 = y << (32-Ebits) + k | w >> Ebits - k;
136 		goto ret_d;
137 		}
138 	z = xa > xa0 ? *--xa : 0;
139 	if (k -= Ebits) {
140 		d0 = Exp_1 | y << k | z >> 32 - k;
141 		y = xa > xa0 ? *--xa : 0;
142 		d1 = z << k | y >> 32 - k;
143 		}
144 	else {
145 		d0 = Exp_1 | y;
146 		d1 = z;
147 		}
148 #else
149 	if (k < Ebits + 16) {
150 		z = xa > xa0 ? *--xa : 0;
151 		d0 = Exp_1 | y << k - Ebits | z >> Ebits + 16 - k;
152 		w = xa > xa0 ? *--xa : 0;
153 		y = xa > xa0 ? *--xa : 0;
154 		d1 = z << k + 16 - Ebits | w << k - Ebits | y >> 16 + Ebits - k;
155 		goto ret_d;
156 		}
157 	z = xa > xa0 ? *--xa : 0;
158 	w = xa > xa0 ? *--xa : 0;
159 	k -= Ebits + 16;
160 	d0 = Exp_1 | y << k + 16 | z << k | w >> 16 - k;
161 	y = xa > xa0 ? *--xa : 0;
162 	d1 = w << k + 16 | y << k;
163 #endif
164  ret_d:
165 #ifdef VAX
166 	word0(d) = d0 >> 16 | d0 << 16;
167 	word1(d) = d1 >> 16 | d1 << 16;
168 #else
169 #undef d0
170 #undef d1
171 #endif
172 	return d.d;
173 	}
174 
175  static double
ratio(Bigint * a,Bigint * b)176 ratio(Bigint *a, Bigint *b)
177 {
178 	Dul da, db;
179 	int k, ka, kb;
180 
181 	da.d = b2d(a, &ka);
182 	db.d = b2d(b, &kb);
183 #ifdef Pack_32
184 	k = ka - kb + 32*(a->wds - b->wds);
185 #else
186 	k = ka - kb + 16*(a->wds - b->wds);
187 #endif
188 #ifdef IBM
189 	if (k > 0) {
190 		word0(da) += (k >> 2)*Exp_msk1;
191 		if (k &= 3)
192 			da *= 1 << k;
193 		}
194 	else {
195 		k = -k;
196 		word0(db) += (k >> 2)*Exp_msk1;
197 		if (k &= 3)
198 			db *= 1 << k;
199 		}
200 #else
201 	if (k > 0)
202 		word0(da) += k*Exp_msk1;
203 	else {
204 		k = -k;
205 		word0(db) += k*Exp_msk1;
206 		}
207 #endif
208 	return da.d / db.d;
209 	}
210 
211  double
strtod(CONST char * s00,char ** se)212 strtod(CONST char *s00, char **se)
213 {
214 	int bb2, bb5, bbe, bd2, bd5, bbbits, bs2, c, dsign,
215 		 e, e1, esign, i, j, k, nd, nd0, nf, nz, nz0, sign;
216 	CONST char *s, *s0, *s1;
217 	double aadj, aadj1, adj;
218 	Dul rv, rv0;
219 	long L;
220 	unsigned long y, z;
221 	Bigint *bb, *bb1, *bd, *bd0, *bs, *delta;
222 	sign = nz0 = nz = 0;
223 	rv.d = 0.;
224 	for(s = s00;;s++) switch(*s) {
225 		case '-':
226 			sign = 1;
227 			/* no break */
228 		case '+':
229 			if (*++s)
230 				goto break2;
231 			/* no break */
232 		case 0:
233 			s = s00;
234 			goto ret;
235 		case '\t':
236 		case '\n':
237 		case '\v':
238 		case '\f':
239 		case '\r':
240 		case ' ':
241 			continue;
242 		default:
243 			goto break2;
244 		}
245  break2:
246 	if (*s == '0') {
247 		nz0 = 1;
248 		while(*++s == '0') ;
249 		if (!*s)
250 			goto ret;
251 		}
252 	s0 = s;
253 	y = z = 0;
254 	for(nd = nf = 0; (c = *s) >= '0' && c <= '9'; nd++, s++)
255 		if (nd < 9)
256 			y = 10*y + c - '0';
257 		else if (nd < 16)
258 			z = 10*z + c - '0';
259 	nd0 = nd;
260 	if (c == '.') {
261 		c = *++s;
262 		if (!nd) {
263 			for(; c == '0'; c = *++s)
264 				nz++;
265 			if (c > '0' && c <= '9') {
266 				s0 = s;
267 				nf += nz;
268 				nz = 0;
269 				goto have_dig;
270 				}
271 			goto dig_done;
272 			}
273 		for(; c >= '0' && c <= '9'; c = *++s) {
274  have_dig:
275 			nz++;
276 			if (c -= '0') {
277 				nf += nz;
278 				for(i = 1; i < nz; i++)
279 					if (nd++ < 9)
280 						y *= 10;
281 					else if (nd <= DBL_DIG + 1)
282 						z *= 10;
283 				if (nd++ < 9)
284 					y = 10*y + c;
285 				else if (nd <= DBL_DIG + 1)
286 					z = 10*z + c;
287 				nz = 0;
288 				}
289 			}
290 		}
291  dig_done:
292 	e = 0;
293 	if (c == 'e' || c == 'E') {
294 		if (!nd && !nz && !nz0) {
295 			s = s00;
296 			goto ret;
297 			}
298 		s00 = s;
299 		esign = 0;
300 		switch(c = *++s) {
301 			case '-':
302 				esign = 1;
303 			case '+':
304 				c = *++s;
305 			}
306 		if (c >= '0' && c <= '9') {
307 			while(c == '0')
308 				c = *++s;
309 			if (c > '0' && c <= '9') {
310 				e = c - '0';
311 				s1 = s;
312 				while((c = *++s) >= '0' && c <= '9')
313 					e = 10*e + c - '0';
314 				if (s - s1 > 8)
315 					/* Avoid confusion from exponents
316 					 * so large that e might overflow.
317 					 */
318 					e = 9999999;
319 				if (esign)
320 					e = -e;
321 				}
322 			else
323 				e = 0;
324 			}
325 		else
326 			s = s00;
327 		}
328 	if (!nd) {
329 		if (!nz && !nz0)
330 			s = s00;
331 		goto ret;
332 		}
333 	e1 = e -= nf;
334 
335 	/* Now we have nd0 digits, starting at s0, followed by a
336 	 * decimal point, followed by nd-nd0 digits.  The number we're
337 	 * after is the integer represented by those digits times
338 	 * 10**e */
339 
340 	if (!nd0)
341 		nd0 = nd;
342 	k = nd < DBL_DIG + 1 ? nd : DBL_DIG + 1;
343 	rv.d = y;
344 	if (k > 9)
345 		rv.d = tens[k - 9] * rv.d + z;
346 	bd0 = 0;
347 	if (nd <= DBL_DIG
348 #ifndef RND_PRODQUOT
349 		&& FLT_ROUNDS == 1
350 #endif
351 			) {
352 		if (!e)
353 			goto ret;
354 		if (e > 0) {
355 			if (e <= Ten_pmax) {
356 #ifdef VAX
357 				goto vax_ovfl_check;
358 #else
359 				/* rv = */ rounded_product(rv.d, tens[e]);
360 				goto ret;
361 #endif
362 				}
363 			i = DBL_DIG - nd;
364 			if (e <= Ten_pmax + i) {
365 				/* A fancier test would sometimes let us do
366 				 * this for larger i values.
367 				 */
368 				e -= i;
369 				rv.d *= tens[i];
370 #ifdef VAX
371 				/* VAX exponent range is so narrow we must
372 				 * worry about overflow here...
373 				 */
374  vax_ovfl_check:
375 				word0(rv) -= P*Exp_msk1;
376 				/* rv = */ rounded_product(rv.d, tens[e]);
377 				if ((word0(rv) & Exp_mask)
378 				 > Exp_msk1*(DBL_MAX_EXP+Bias-1-P))
379 					goto ovfl;
380 				word0(rv) += P*Exp_msk1;
381 #else
382 				/* rv = */ rounded_product(rv.d, tens[e]);
383 #endif
384 				goto ret;
385 				}
386 			}
387 		else if (e >= -Ten_pmax) {
388 			/* rv = */ rounded_quotient(rv.d, tens[-e]);
389 			goto ret;
390 			}
391 		}
392 	e1 += nd - k;
393 
394 	/* Get starting approximation = rv * 10**e1 */
395 
396 	if (e1 > 0) {
397 		if (nd0 + e1 - 1 > DBL_MAX_10_EXP)
398 			goto ovfl;
399 		if (i = e1 & 15)
400 			rv.d *= tens[i];
401 		if (e1 &= ~15) {
402 			if (e1 > DBL_MAX_10_EXP) {
403  ovfl:
404 				errno = ERANGE;
405 				rv.d = HUGE_VAL;
406 				if (bd0)
407 					goto retfree;
408 				goto ret;
409 				}
410 			if (e1 >>= 4) {
411 				for(j = 0; e1 > 1; j++, e1 >>= 1)
412 					if (e1 & 1)
413 						rv.d *= bigtens[j];
414 			/* The last multiplication could overflow. */
415 				word0(rv) -= P*Exp_msk1;
416 				rv.d *= bigtens[j];
417 				if ((z = word0(rv) & Exp_mask)
418 				 > Exp_msk1*(DBL_MAX_EXP+Bias-P))
419 					goto ovfl;
420 				if (z > Exp_msk1*(DBL_MAX_EXP+Bias-1-P)) {
421 					/* set to largest number */
422 					/* (Can't trust DBL_MAX) */
423 					word0(rv) = Big0;
424 					word1(rv) = Big1;
425 					}
426 				else
427 					word0(rv) += P*Exp_msk1;
428 				}
429 
430 			}
431 		}
432 	else if (e1 < 0) {
433 		e1 = -e1;
434 		if (i = e1 & 15)
435 			rv.d /= tens[i];
436 		if (e1 &= ~15) {
437 			e1 >>= 4;
438 			if (e1 >= 1 << n_bigtens)
439 				goto undfl;
440 			for(j = 0; e1 > 1; j++, e1 >>= 1)
441 				if (e1 & 1)
442 					rv.d *= tinytens[j];
443 			/* The last multiplication could underflow. */
444 			rv0.d = rv.d;
445 			rv.d *= tinytens[j];
446 			if (rv.d == 0) {
447 				rv.d = 2.*rv0.d;
448 				rv.d *= tinytens[j];
449 				if (rv.d == 0) {
450  undfl:
451 					rv.d = 0.;
452 					errno = ERANGE;
453 					if (bd0)
454 						goto retfree;
455 					goto ret;
456 					}
457 				word0(rv) = Tiny0;
458 				word1(rv) = Tiny1;
459 				/* The refinement below will clean
460 				 * this approximation up.
461 				 */
462 				}
463 			}
464 		}
465 
466 	/* Now the hard part -- adjusting rv to the correct value.*/
467 
468 	/* Put digits into bd: true value = bd * 10^e */
469 
470 	bd0 = s2b(s0, nd0, nd, y);
471 
472 	for(;;) {
473 		bd = Balloc(bd0->k);
474 		Bcopy(bd, bd0);
475 		bb = d2b(rv.d, &bbe, &bbbits);	/* rv = bb * 2^bbe */
476 		bs = i2b(1);
477 
478 		if (e >= 0) {
479 			bb2 = bb5 = 0;
480 			bd2 = bd5 = e;
481 			}
482 		else {
483 			bb2 = bb5 = -e;
484 			bd2 = bd5 = 0;
485 			}
486 		if (bbe >= 0)
487 			bb2 += bbe;
488 		else
489 			bd2 -= bbe;
490 		bs2 = bb2;
491 #ifdef Sudden_Underflow
492 #ifdef IBM
493 		j = 1 + 4*P - 3 - bbbits + ((bbe + bbbits - 1) & 3);
494 #else
495 		j = P + 1 - bbbits;
496 #endif
497 #else
498 		i = bbe + bbbits - 1;	/* logb(rv) */
499 		if (i < Emin)	/* denormal */
500 			j = bbe + (P-Emin);
501 		else
502 			j = P + 1 - bbbits;
503 #endif
504 		bb2 += j;
505 		bd2 += j;
506 		i = bb2 < bd2 ? bb2 : bd2;
507 		if (i > bs2)
508 			i = bs2;
509 		if (i > 0) {
510 			bb2 -= i;
511 			bd2 -= i;
512 			bs2 -= i;
513 			}
514 		if (bb5 > 0) {
515 			bs = pow5mult(bs, bb5);
516 			bb1 = mult(bs, bb);
517 			Bfree(bb);
518 			bb = bb1;
519 			}
520 		if (bb2 > 0)
521 			bb = lshift(bb, bb2);
522 		if (bd5 > 0)
523 			bd = pow5mult(bd, bd5);
524 		if (bd2 > 0)
525 			bd = lshift(bd, bd2);
526 		if (bs2 > 0)
527 			bs = lshift(bs, bs2);
528 		delta = diff(bb, bd);
529 		dsign = delta->sign;
530 		delta->sign = 0;
531 		i = cmp(delta, bs);
532 		if (i < 0) {
533 			/* Error is less than half an ulp -- check for
534 			 * special case of mantissa a power of two.
535 			 */
536 			if (dsign || word1(rv) || word0(rv) & Bndry_mask)
537 				break;
538 			delta = lshift(delta,Log2P);
539 			if (cmp(delta, bs) > 0)
540 				goto drop_down;
541 			break;
542 			}
543 		if (i == 0) {
544 			/* exactly half-way between */
545 			if (dsign) {
546 				if ((word0(rv) & Bndry_mask1) == Bndry_mask1
547 				 &&  word1(rv) == 0xffffffff) {
548 					/*boundary case -- increment exponent*/
549 					word0(rv) = (word0(rv) & Exp_mask)
550 						+ Exp_msk1
551 #ifdef IBM
552 						| Exp_msk1 >> 4
553 #endif
554 						;
555 					word1(rv) = 0;
556 					break;
557 					}
558 				}
559 			else if (!(word0(rv) & Bndry_mask) && !word1(rv)) {
560  drop_down:
561 				/* boundary case -- decrement exponent */
562 #ifdef Sudden_Underflow
563 				L = word0(rv) & Exp_mask;
564 #ifdef IBM
565 				if (L <  Exp_msk1)
566 #else
567 				if (L <= Exp_msk1)
568 #endif
569 					goto undfl;
570 				L -= Exp_msk1;
571 #else
572 				L = (word0(rv) & Exp_mask) - Exp_msk1;
573 #endif
574 				word0(rv) = L | Bndry_mask1;
575 				word1(rv) = 0xffffffff;
576 #ifdef IBM
577 				goto cont;
578 #else
579 				break;
580 #endif
581 				}
582 #ifndef ROUND_BIASED
583 			if (!(word1(rv) & LSB))
584 				break;
585 #endif
586 			if (dsign)
587 				rv.d += ulp(rv.d);
588 #ifndef ROUND_BIASED
589 			else {
590 				rv.d -= ulp(rv.d);
591 #ifndef Sudden_Underflow
592 				if (rv.d == 0)
593 					goto undfl;
594 #endif
595 				}
596 #endif
597 			break;
598 			}
599 		if ((aadj = ratio(delta, bs)) <= 2.) {
600 			if (dsign)
601 				aadj = aadj1 = 1.;
602 			else if (word1(rv) || word0(rv) & Bndry_mask) {
603 #ifndef Sudden_Underflow
604 				if (word1(rv) == Tiny1 && !word0(rv))
605 					goto undfl;
606 #endif
607 				aadj = 1.;
608 				aadj1 = -1.;
609 				}
610 			else {
611 				/* special case -- power of FLT_RADIX to be */
612 				/* rounded down... */
613 
614 				if (aadj < 2./FLT_RADIX)
615 					aadj = 1./FLT_RADIX;
616 				else
617 					aadj *= 0.5;
618 				aadj1 = -aadj;
619 				}
620 			}
621 		else {
622 			aadj *= 0.5;
623 			aadj1 = dsign ? aadj : -aadj;
624 #ifdef Check_FLT_ROUNDS
625 			switch(FLT_ROUNDS) {
626 				case 2: /* towards +infinity */
627 					aadj1 -= 0.5;
628 					break;
629 				case 0: /* towards 0 */
630 				case 3: /* towards -infinity */
631 					aadj1 += 0.5;
632 				}
633 #else
634 			if (FLT_ROUNDS == 0)
635 				aadj1 += 0.5;
636 #endif
637 			}
638 		y = word0(rv) & Exp_mask;
639 
640 		/* Check for overflow */
641 
642 		if (y == Exp_msk1*(DBL_MAX_EXP+Bias-1)) {
643 			rv0.d = rv.d;
644 			word0(rv) -= P*Exp_msk1;
645 			adj = aadj1 * ulp(rv.d);
646 			rv.d += adj;
647 			if ((word0(rv) & Exp_mask) >=
648 					Exp_msk1*(DBL_MAX_EXP+Bias-P)) {
649 				if (word0(rv0) == Big0 && word1(rv0) == Big1)
650 					goto ovfl;
651 				word0(rv) = Big0;
652 				word1(rv) = Big1;
653 				goto cont;
654 				}
655 			else
656 				word0(rv) += P*Exp_msk1;
657 			}
658 		else {
659 #ifdef Sudden_Underflow
660 			if ((word0(rv) & Exp_mask) <= P*Exp_msk1) {
661 				rv0.d = rv.d;
662 				word0(rv) += P*Exp_msk1;
663 				adj = aadj1 * ulp(rv.d);
664 				rv.d += adj;
665 #ifdef IBM
666 				if ((word0(rv) & Exp_mask) <  P*Exp_msk1)
667 #else
668 				if ((word0(rv) & Exp_mask) <= P*Exp_msk1)
669 #endif
670 					{
671 					if (word0(rv0) == Tiny0
672 					 && word1(rv0) == Tiny1)
673 						goto undfl;
674 					word0(rv) = Tiny0;
675 					word1(rv) = Tiny1;
676 					goto cont;
677 					}
678 				else
679 					word0(rv) -= P*Exp_msk1;
680 				}
681 			else {
682 				adj = aadj1 * ulp(rv.d);
683 				rv.d += adj;
684 				}
685 #else
686 			/* Compute adj so that the IEEE rounding rules will
687 			 * correctly round rv + adj in some half-way cases.
688 			 * If rv * ulp(rv) is denormalized (i.e.,
689 			 * y <= (P-1)*Exp_msk1), we must adjust aadj to avoid
690 			 * trouble from bits lost to denormalization;
691 			 * example: 1.2e-307 .
692 			 */
693 			if (y <= (P-1)*Exp_msk1 && aadj >= 1.) {
694 				aadj1 = (double)(int)(aadj + 0.5);
695 				if (!dsign)
696 					aadj1 = -aadj1;
697 				}
698 			adj = aadj1 * ulp(rv.d);
699 			rv.d += adj;
700 #endif
701 			}
702 		z = word0(rv) & Exp_mask;
703 		if (y == z) {
704 			/* Can we stop now? */
705 			L = aadj;
706 			aadj -= L;
707 			/* The tolerances below are conservative. */
708 			if (dsign || word1(rv) || word0(rv) & Bndry_mask) {
709 				if (aadj < .4999999 || aadj > .5000001)
710 					break;
711 				}
712 			else if (aadj < .4999999/FLT_RADIX)
713 				break;
714 			}
715  cont:
716 		Bfree(bb);
717 		Bfree(bd);
718 		Bfree(bs);
719 		Bfree(delta);
720 		}
721  retfree:
722 	Bfree(bb);
723 	Bfree(bd);
724 	Bfree(bs);
725 	Bfree(bd0);
726 	Bfree(delta);
727  ret:
728 	if (se)
729 		*se = (char *)s;
730 	return sign ? -rv.d : rv.d;
731 	}
732