build: aclocal.m4 -- Provide arguments to main
[nasm.git] / asm / float.c
blob187b9e5e018225709c4e34b5d96611a2a83fe555
1 /* ----------------------------------------------------------------------- *
3 * Copyright 1996-2018 The NASM Authors - All Rights Reserved
4 * See the file AUTHORS included with the NASM distribution for
5 * the specific copyright holders.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following
9 * conditions are met:
11 * * Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * * Redistributions in binary form must reproduce the above
14 * copyright notice, this list of conditions and the following
15 * disclaimer in the documentation and/or other materials provided
16 * with the distribution.
18 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND
19 * CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES,
20 * INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
21 * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
22 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
23 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
24 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
25 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
26 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
28 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
29 * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
30 * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32 * ----------------------------------------------------------------------- */
35 * float.c floating-point constant support for the Netwide Assembler
38 #include "compiler.h"
40 #include <ctype.h>
41 #include <stdio.h>
42 #include <stdlib.h>
43 #include <string.h>
45 #include "nasm.h"
46 #include "float.h"
47 #include "error.h"
50 * -----------------
51 * local variables
52 * -----------------
54 static bool daz = false; /* denormals as zero */
55 static enum float_round rc = FLOAT_RC_NEAR; /* rounding control */
58 * -----------
59 * constants
60 * -----------
63 /* "A limb is like a digit but bigger */
64 typedef uint32_t fp_limb;
65 typedef uint64_t fp_2limb;
67 #define LIMB_BITS 32
68 #define LIMB_BYTES (LIMB_BITS/8)
69 #define LIMB_TOP_BIT ((fp_limb)1 << (LIMB_BITS-1))
70 #define LIMB_MASK ((fp_limb)(~0))
71 #define LIMB_ALL_BYTES ((fp_limb)0x01010101)
72 #define LIMB_BYTE(x) ((x)*LIMB_ALL_BYTES)
74 /* 112 bits + 64 bits for accuracy + 16 bits for rounding */
75 #define MANT_LIMBS 6
77 /* 52 digits fit in 176 bits because 10^53 > 2^176 > 10^52 */
78 #define MANT_DIGITS 52
80 /* the format and the argument list depend on MANT_LIMBS */
81 #define MANT_FMT "%08x_%08x_%08x_%08x_%08x_%08x"
82 #define MANT_ARG SOME_ARG(mant, 0)
84 #define SOME_ARG(a,i) (a)[(i)+0], (a)[(i)+1], (a)[(i)+2], \
85 (a)[(i)+3], (a)[(i)+4], (a)[(i)+5]
88 * ---------------------------------------------------------------------------
89 * emit a printf()-like debug message... but only if DEBUG_FLOAT was defined
90 * ---------------------------------------------------------------------------
93 #ifdef DEBUG_FLOAT
94 #define dprintf(x) printf x
95 #else
96 #define dprintf(x) do { } while (0)
97 #endif
100 * ---------------------------------------------------------------------------
101 * multiply
102 * ---------------------------------------------------------------------------
104 static int float_multiply(fp_limb *to, fp_limb *from)
106 fp_2limb temp[MANT_LIMBS * 2];
107 int i, j;
110 * guaranteed that top bit of 'from' is set -- so we only have
111 * to worry about _one_ bit shift to the left
113 dprintf(("%s=" MANT_FMT "\n", "mul1", SOME_ARG(to, 0)));
114 dprintf(("%s=" MANT_FMT "\n", "mul2", SOME_ARG(from, 0)));
116 memset(temp, 0, sizeof temp);
118 for (i = 0; i < MANT_LIMBS; i++) {
119 for (j = 0; j < MANT_LIMBS; j++) {
120 fp_2limb n;
121 n = (fp_2limb) to[i] * (fp_2limb) from[j];
122 temp[i + j] += n >> LIMB_BITS;
123 temp[i + j + 1] += (fp_limb)n;
127 for (i = MANT_LIMBS * 2; --i;) {
128 temp[i - 1] += temp[i] >> LIMB_BITS;
129 temp[i] &= LIMB_MASK;
132 dprintf(("%s=" MANT_FMT "_" MANT_FMT "\n", "temp", SOME_ARG(temp, 0),
133 SOME_ARG(temp, MANT_LIMBS)));
135 if (temp[0] & LIMB_TOP_BIT) {
136 for (i = 0; i < MANT_LIMBS; i++) {
137 to[i] = temp[i] & LIMB_MASK;
139 dprintf(("%s=" MANT_FMT " (%i)\n", "prod", SOME_ARG(to, 0), 0));
140 return 0;
141 } else {
142 for (i = 0; i < MANT_LIMBS; i++) {
143 to[i] = (temp[i] << 1) + !!(temp[i + 1] & LIMB_TOP_BIT);
145 dprintf(("%s=" MANT_FMT " (%i)\n", "prod", SOME_ARG(to, 0), -1));
146 return -1;
151 * ---------------------------------------------------------------------------
152 * read an exponent; returns INT32_MAX on error
153 * ---------------------------------------------------------------------------
155 static int32_t read_exponent(const char *string, int32_t max)
157 int32_t i = 0;
158 bool neg = false;
160 if (*string == '+') {
161 string++;
162 } else if (*string == '-') {
163 neg = true;
164 string++;
166 while (*string) {
167 if (*string >= '0' && *string <= '9') {
168 i = (i * 10) + (*string - '0');
171 * To ensure that underflows and overflows are
172 * handled properly we must avoid wraparounds of
173 * the signed integer value that is used to hold
174 * the exponent. Therefore we cap the exponent at
175 * +/-5000, which is slightly more/less than
176 * what's required for normal and denormal numbers
177 * in single, double, and extended precision, but
178 * sufficient to avoid signed integer wraparound.
180 if (i > max)
181 i = max;
182 } else if (*string == '_') {
183 /* do nothing */
184 } else {
185 nasm_nonfatal("invalid character in floating-point constant %s: '%c'",
186 "exponent", *string);
187 return INT32_MAX;
189 string++;
192 return neg ? -i : i;
196 * ---------------------------------------------------------------------------
197 * convert
198 * ---------------------------------------------------------------------------
200 static bool ieee_flconvert(const char *string, fp_limb *mant,
201 int32_t * exponent)
203 char digits[MANT_DIGITS];
204 char *p, *q, *r;
205 fp_limb mult[MANT_LIMBS], bit;
206 fp_limb *m;
207 int32_t tenpwr, twopwr;
208 int32_t extratwos;
209 bool started, seendot, warned;
211 warned = false;
212 p = digits;
213 tenpwr = 0;
214 started = seendot = false;
216 while (*string && *string != 'E' && *string != 'e') {
217 if (*string == '.') {
218 if (!seendot) {
219 seendot = true;
220 } else {
221 nasm_nonfatal("too many periods in floating-point constant");
222 return false;
224 } else if (*string >= '0' && *string <= '9') {
225 if (*string == '0' && !started) {
226 if (seendot) {
227 tenpwr--;
229 } else {
230 started = true;
231 if (p < digits + sizeof(digits)) {
232 *p++ = *string - '0';
233 } else {
234 if (!warned) {
236 *!float-toolong [on] too many digits in floating-point number
237 *! warns about too many digits in floating-point numbers.
239 nasm_warn(WARN_FLOAT_TOOLONG|ERR_PASS2,
240 "floating-point constant significand contains "
241 "more than %i digits", MANT_DIGITS);
242 warned = true;
245 if (!seendot) {
246 tenpwr++;
249 } else if (*string == '_') {
250 /* do nothing */
251 } else {
252 nasm_nonfatalf(ERR_PASS2,
253 "invalid character in floating-point constant %s: '%c'",
254 "significand", *string);
255 return false;
257 string++;
260 if (*string) {
261 int32_t e;
263 string++; /* eat the E */
264 e = read_exponent(string, 5000);
265 if (e == INT32_MAX)
266 return false;
267 tenpwr += e;
271 * At this point, the memory interval [digits,p) contains a
272 * series of decimal digits zzzzzzz, such that our number X
273 * satisfies X = 0.zzzzzzz * 10^tenpwr.
275 q = digits;
276 dprintf(("X = 0."));
277 while (q < p) {
278 dprintf(("%c", *q + '0'));
279 q++;
281 dprintf((" * 10^%i\n", tenpwr));
284 * Now convert [digits,p) to our internal representation.
286 bit = LIMB_TOP_BIT;
287 for (m = mant; m < mant + MANT_LIMBS; m++) {
288 *m = 0;
290 m = mant;
291 q = digits;
292 started = false;
293 twopwr = 0;
294 while (m < mant + MANT_LIMBS) {
295 fp_limb carry = 0;
296 while (p > q && !p[-1]) {
297 p--;
299 if (p <= q) {
300 break;
302 for (r = p; r-- > q;) {
303 int32_t i;
304 i = 2 * *r + carry;
305 if (i >= 10) {
306 carry = 1;
307 i -= 10;
308 } else {
309 carry = 0;
311 *r = i;
313 if (carry) {
314 *m |= bit;
315 started = true;
317 if (started) {
318 if (bit == 1) {
319 bit = LIMB_TOP_BIT;
320 m++;
321 } else {
322 bit >>= 1;
324 } else {
325 twopwr--;
328 twopwr += tenpwr;
331 * At this point, the 'mant' array contains the first frac-
332 * tional places of a base-2^16 real number which when mul-
333 * tiplied by 2^twopwr and 5^tenpwr gives X.
335 dprintf(("X = " MANT_FMT " * 2^%i * 5^%i\n", MANT_ARG, twopwr,
336 tenpwr));
339 * Now multiply 'mant' by 5^tenpwr.
341 if (tenpwr < 0) { /* mult = 5^-1 = 0.2 */
342 for (m = mult; m < mult + MANT_LIMBS - 1; m++) {
343 *m = LIMB_BYTE(0xcc);
345 mult[MANT_LIMBS - 1] = LIMB_BYTE(0xcc)+1;
346 extratwos = -2;
347 tenpwr = -tenpwr;
350 * If tenpwr was 1000...000b, then it becomes 1000...000b. See
351 * the "ANSI C" comment below for more details on that case.
353 * Because we already truncated tenpwr to +5000...-5000 inside
354 * the exponent parsing code, this shouldn't happen though.
356 } else if (tenpwr > 0) { /* mult = 5^+1 = 5.0 */
357 mult[0] = (fp_limb)5 << (LIMB_BITS-3); /* 0xA000... */
358 for (m = mult + 1; m < mult + MANT_LIMBS; m++) {
359 *m = 0;
361 extratwos = 3;
362 } else {
363 extratwos = 0;
365 while (tenpwr) {
366 dprintf(("loop=" MANT_FMT " * 2^%i * 5^%i (%i)\n", MANT_ARG,
367 twopwr, tenpwr, extratwos));
368 if (tenpwr & 1) {
369 dprintf(("mant*mult\n"));
370 twopwr += extratwos + float_multiply(mant, mult);
372 dprintf(("mult*mult\n"));
373 extratwos = extratwos * 2 + float_multiply(mult, mult);
374 tenpwr >>= 1;
377 * In ANSI C, the result of right-shifting a signed integer is
378 * considered implementation-specific. To ensure that the loop
379 * terminates even if tenpwr was 1000...000b to begin with, we
380 * manually clear the MSB, in case a 1 was shifted in.
382 * Because we already truncated tenpwr to +5000...-5000 inside
383 * the exponent parsing code, this shouldn't matter; neverthe-
384 * less it is the right thing to do here.
386 tenpwr &= (uint32_t) - 1 >> 1;
390 * At this point, the 'mant' array contains the first frac-
391 * tional places of a base-2^16 real number in [0.5,1) that
392 * when multiplied by 2^twopwr gives X. Or it contains zero
393 * of course. We are done.
395 *exponent = twopwr;
396 return true;
400 * ---------------------------------------------------------------------------
401 * operations of specific bits
402 * ---------------------------------------------------------------------------
405 /* Set a bit, using *bigendian* bit numbering (0 = MSB) */
406 static void set_bit(fp_limb *mant, int bit)
408 mant[bit/LIMB_BITS] |= LIMB_TOP_BIT >> (bit & (LIMB_BITS-1));
411 /* Test a single bit */
412 static int test_bit(const fp_limb *mant, int bit)
414 return (mant[bit/LIMB_BITS] >> (~bit & (LIMB_BITS-1))) & 1;
417 /* Report if the mantissa value is all zero */
418 static bool is_zero(const fp_limb *mant)
420 int i;
422 for (i = 0; i < MANT_LIMBS; i++)
423 if (mant[i])
424 return false;
426 return true;
430 * ---------------------------------------------------------------------------
431 * round a mantissa off after i words
432 * ---------------------------------------------------------------------------
435 #define ROUND_COLLECT_BITS \
436 do { \
437 m = mant[i] & (2*bit-1); \
438 for (j = i+1; j < MANT_LIMBS; j++) \
439 m = m | mant[j]; \
440 } while (0)
442 #define ROUND_ABS_DOWN \
443 do { \
444 mant[i] &= ~(bit-1); \
445 for (j = i+1; j < MANT_LIMBS; j++) \
446 mant[j] = 0; \
447 return false; \
448 } while (0)
450 #define ROUND_ABS_UP \
451 do { \
452 mant[i] = (mant[i] & ~(bit-1)) + bit; \
453 for (j = i+1; j < MANT_LIMBS; j++) \
454 mant[j] = 0; \
455 while (i > 0 && !mant[i]) \
456 ++mant[--i]; \
457 return !mant[0]; \
458 } while (0)
460 static bool ieee_round(bool minus, fp_limb *mant, int bits)
462 fp_limb m = 0;
463 int32_t j;
464 int i = bits / LIMB_BITS;
465 int p = bits % LIMB_BITS;
466 fp_limb bit = LIMB_TOP_BIT >> p;
468 if (rc == FLOAT_RC_NEAR) {
469 if (mant[i] & bit) {
470 mant[i] &= ~bit;
471 ROUND_COLLECT_BITS;
472 mant[i] |= bit;
473 if (m) {
474 ROUND_ABS_UP;
475 } else {
476 if (test_bit(mant, bits-1)) {
477 ROUND_ABS_UP;
478 } else {
479 ROUND_ABS_DOWN;
482 } else {
483 ROUND_ABS_DOWN;
485 } else if (rc == FLOAT_RC_ZERO ||
486 rc == (minus ? FLOAT_RC_UP : FLOAT_RC_DOWN)) {
487 ROUND_ABS_DOWN;
488 } else {
489 /* rc == (minus ? FLOAT_RC_DOWN : FLOAT_RC_UP) */
490 /* Round toward +/- infinity */
491 ROUND_COLLECT_BITS;
492 if (m) {
493 ROUND_ABS_UP;
494 } else {
495 ROUND_ABS_DOWN;
498 return false;
501 /* Returns a value >= 16 if not a valid hex digit */
502 static unsigned int hexval(char c)
504 unsigned int v = (unsigned char) c;
506 if (v >= '0' && v <= '9')
507 return v - '0';
508 else
509 return (v|0x20) - 'a' + 10;
512 /* Handle floating-point numbers with radix 2^bits and binary exponent */
513 static bool ieee_flconvert_bin(const char *string, int bits,
514 fp_limb *mant, int32_t *exponent)
516 static const int log2tbl[16] =
517 { -1, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3 };
518 fp_limb mult[MANT_LIMBS + 1], *mp;
519 int ms;
520 int32_t twopwr;
521 bool seendot, seendigit;
522 unsigned char c;
523 const int radix = 1 << bits;
524 fp_limb v;
526 twopwr = 0;
527 seendot = seendigit = false;
528 ms = 0;
529 mp = NULL;
531 memset(mult, 0, sizeof mult);
533 while ((c = *string++) != '\0') {
534 if (c == '.') {
535 if (!seendot)
536 seendot = true;
537 else {
538 nasm_nonfatal("too many periods in floating-point constant");
539 return false;
541 } else if ((v = hexval(c)) < (unsigned int)radix) {
542 if (!seendigit && v) {
543 int l = log2tbl[v];
545 seendigit = true;
546 mp = mult;
547 ms = (LIMB_BITS-1)-l;
549 twopwr += l+1-bits;
552 if (seendigit) {
553 if (ms < 0) {
554 /* Cast to fp_2limb as ms == -LIMB_BITS is possible. */
555 *mp |= (fp_2limb)v >> -ms;
556 mp++;
557 if (mp > &mult[MANT_LIMBS])
558 mp = &mult[MANT_LIMBS]; /* Guard slot */
559 ms += LIMB_BITS;
561 *mp |= v << ms;
562 ms -= bits;
564 if (!seendot)
565 twopwr += bits;
566 } else {
567 if (seendot)
568 twopwr -= bits;
570 } else if (c == 'p' || c == 'P') {
571 int32_t e;
572 e = read_exponent(string, 20000);
573 if (e == INT32_MAX)
574 return false;
575 twopwr += e;
576 break;
577 } else if (c == '_') {
578 /* ignore */
579 } else {
580 nasm_nonfatal("floating-point constant: `%c' is invalid character", c);
581 return false;
585 if (!seendigit) {
586 memset(mant, 0, MANT_LIMBS*sizeof(fp_limb)); /* Zero */
587 *exponent = 0;
588 } else {
589 memcpy(mant, mult, MANT_LIMBS*sizeof(fp_limb));
590 *exponent = twopwr;
593 return true;
597 * Shift a mantissa to the right by i bits.
599 static void ieee_shr(fp_limb *mant, int i)
601 fp_limb n, m;
602 int j = 0;
603 int sr, sl, offs;
605 sr = i % LIMB_BITS; sl = LIMB_BITS-sr;
606 offs = i/LIMB_BITS;
608 if (sr == 0) {
609 if (offs)
610 for (j = MANT_LIMBS-1; j >= offs; j--)
611 mant[j] = mant[j-offs];
612 } else if (MANT_LIMBS-1-offs < 0) {
613 j = MANT_LIMBS-1;
614 } else {
615 n = mant[MANT_LIMBS-1-offs] >> sr;
616 for (j = MANT_LIMBS-1; j > offs; j--) {
617 m = mant[j-offs-1];
618 mant[j] = (m << sl) | n;
619 n = m >> sr;
621 mant[j--] = n;
623 while (j >= 0)
624 mant[j--] = 0;
627 /* Produce standard IEEE formats, with implicit or explicit integer
628 bit; this makes the following assumptions:
630 - the sign bit is the MSB, followed by the exponent,
631 followed by the integer bit if present.
632 - the sign bit plus exponent fit in 16 bits.
633 - the exponent bias is 2^(n-1)-1 for an n-bit exponent */
635 struct ieee_format {
636 int bytes;
637 int mantissa; /* Fractional bits in the mantissa */
638 int explicit; /* Explicit integer */
639 int exponent; /* Bits in the exponent */
643 * The 16- and 128-bit formats are expected to be in IEEE 754r.
644 * AMD SSE5 uses the 16-bit format.
646 * The 32- and 64-bit formats are the original IEEE 754 formats.
648 * The 80-bit format is x87-specific, but widely used.
650 * The 8-bit format appears to be the consensus 8-bit floating-point
651 * format. It is apparently used in graphics applications.
653 static const struct ieee_format ieee_8 = { 1, 3, 0, 4 };
654 static const struct ieee_format ieee_16 = { 2, 10, 0, 5 };
655 static const struct ieee_format ieee_32 = { 4, 23, 0, 8 };
656 static const struct ieee_format ieee_64 = { 8, 52, 0, 11 };
657 static const struct ieee_format ieee_80 = { 10, 63, 1, 15 };
658 static const struct ieee_format ieee_128 = { 16, 112, 0, 15 };
660 /* Types of values we can generate */
661 enum floats {
662 FL_ZERO,
663 FL_DENORMAL,
664 FL_NORMAL,
665 FL_INFINITY,
666 FL_QNAN,
667 FL_SNAN
670 static int to_packed_bcd(const char *str, const char *p,
671 int s, uint8_t *result,
672 const struct ieee_format *fmt)
674 int n = 0;
675 char c;
676 int tv = -1;
678 if (fmt != &ieee_80) {
679 nasm_nonfatal("packed BCD requires an 80-bit format");
680 return 0;
683 while (p >= str) {
684 c = *p--;
685 if (c >= '0' && c <= '9') {
686 if (tv < 0) {
687 if (n == 9)
688 nasm_warn(WARN_OTHER|ERR_PASS2, "packed BCD truncated to 18 digits");
689 tv = c-'0';
690 } else {
691 if (n < 9)
692 *result++ = tv + ((c-'0') << 4);
693 n++;
694 tv = -1;
696 } else if (c == '_') {
697 /* do nothing */
698 } else {
699 nasm_nonfatal("invalid character `%c' in packed BCD constant", c);
700 return 0;
703 if (tv >= 0) {
704 if (n < 9)
705 *result++ = tv;
706 n++;
708 while (n < 9) {
709 *result++ = 0;
710 n++;
712 *result = (s < 0) ? 0x80 : 0;
714 return 1; /* success */
717 static int to_float(const char *str, int s, uint8_t *result,
718 const struct ieee_format *fmt)
720 fp_limb mant[MANT_LIMBS];
721 int32_t exponent = 0;
722 const int32_t expmax = 1 << (fmt->exponent - 1);
723 fp_limb one_mask = LIMB_TOP_BIT >>
724 ((fmt->exponent+fmt->explicit) % LIMB_BITS);
725 const int one_pos = (fmt->exponent+fmt->explicit)/LIMB_BITS;
726 int i;
727 int shift;
728 enum floats type;
729 bool ok;
730 const bool minus = s < 0;
731 const int bits = fmt->bytes * 8;
732 const char *strend;
734 nasm_assert(str[0]);
736 strend = strchr(str, '\0');
737 if (strend[-1] == 'P' || strend[-1] == 'p')
738 return to_packed_bcd(str, strend-2, s, result, fmt);
740 if (str[0] == '_') {
741 /* Special tokens */
743 switch (str[2]) {
744 case 'n': /* __nan__ */
745 case 'N':
746 case 'q': /* __qnan__ */
747 case 'Q':
748 type = FL_QNAN;
749 break;
750 case 's': /* __snan__ */
751 case 'S':
752 type = FL_SNAN;
753 break;
754 case 'i': /* __infinity__ */
755 case 'I':
756 type = FL_INFINITY;
757 break;
758 default:
759 nasm_nonfatal("internal error: unknown FP constant token `%s'", str);
760 type = FL_QNAN;
761 break;
763 } else {
764 if (str[0] == '0') {
765 switch (str[1]) {
766 case 'x': case 'X':
767 case 'h': case 'H':
768 ok = ieee_flconvert_bin(str+2, 4, mant, &exponent);
769 break;
770 case 'o': case 'O':
771 case 'q': case 'Q':
772 ok = ieee_flconvert_bin(str+2, 3, mant, &exponent);
773 break;
774 case 'b': case 'B':
775 case 'y': case 'Y':
776 ok = ieee_flconvert_bin(str+2, 1, mant, &exponent);
777 break;
778 case 'd': case 'D':
779 case 't': case 'T':
780 ok = ieee_flconvert(str+2, mant, &exponent);
781 break;
782 case 'p': case 'P':
783 return to_packed_bcd(str+2, strend-1, s, result, fmt);
784 default:
785 /* Leading zero was just a zero? */
786 ok = ieee_flconvert(str, mant, &exponent);
787 break;
789 } else if (str[0] == '$') {
790 ok = ieee_flconvert_bin(str+1, 4, mant, &exponent);
791 } else {
792 ok = ieee_flconvert(str, mant, &exponent);
795 if (!ok) {
796 type = FL_QNAN;
797 } else if (mant[0] & LIMB_TOP_BIT) {
799 * Non-zero.
801 exponent--;
802 if (exponent >= 2 - expmax && exponent <= expmax) {
803 type = FL_NORMAL;
804 } else if (exponent > 0) {
805 nasm_warn(WARN_FLOAT_OVERFLOW|ERR_PASS2,
806 "overflow in floating-point constant");
807 type = FL_INFINITY;
808 } else {
809 /* underflow or denormal; the denormal code handles
810 actual underflow. */
811 type = FL_DENORMAL;
813 } else {
814 /* Zero */
815 type = FL_ZERO;
819 switch (type) {
820 case FL_ZERO:
821 zero:
822 memset(mant, 0, sizeof mant);
823 break;
825 case FL_DENORMAL:
827 shift = -(exponent + expmax - 2 - fmt->exponent)
828 + fmt->explicit;
829 ieee_shr(mant, shift);
830 ieee_round(minus, mant, bits);
831 if (mant[one_pos] & one_mask) {
832 /* One's position is set, we rounded up into normal range */
833 exponent = 1;
834 if (!fmt->explicit)
835 mant[one_pos] &= ~one_mask; /* remove explicit one */
836 mant[0] |= exponent << (LIMB_BITS-1 - fmt->exponent);
837 } else {
838 if (daz || is_zero(mant)) {
840 *!float-underflow [off] floating point underflow
841 *! warns about floating point underflow (a nonzero
842 *! constant rounded to zero.)
844 nasm_warn(WARN_FLOAT_UNDERFLOW|ERR_PASS2,
845 "underflow in floating-point constant");
846 goto zero;
847 } else {
849 *!float-denorm [off] floating point denormal
850 *! warns about denormal floating point constants.
852 nasm_warn(WARN_FLOAT_DENORM|ERR_PASS2,
853 "denormal floating-point constant");
856 break;
859 case FL_NORMAL:
860 exponent += expmax - 1;
861 ieee_shr(mant, fmt->exponent+fmt->explicit);
862 ieee_round(minus, mant, bits);
863 /* did we scale up by one? */
864 if (test_bit(mant, fmt->exponent+fmt->explicit-1)) {
865 ieee_shr(mant, 1);
866 exponent++;
867 if (exponent >= (expmax << 1)-1) {
869 *!float-overflow [on] floating point overflow
870 *! warns about floating point underflow.
872 nasm_warn(WARN_FLOAT_OVERFLOW|ERR_PASS2,
873 "overflow in floating-point constant");
874 type = FL_INFINITY;
875 goto overflow;
879 if (!fmt->explicit)
880 mant[one_pos] &= ~one_mask; /* remove explicit one */
881 mant[0] |= exponent << (LIMB_BITS-1 - fmt->exponent);
882 break;
884 case FL_INFINITY:
885 case FL_QNAN:
886 case FL_SNAN:
887 overflow:
888 memset(mant, 0, sizeof mant);
889 mant[0] = (((fp_limb)1 << fmt->exponent)-1)
890 << (LIMB_BITS-1 - fmt->exponent);
891 if (fmt->explicit)
892 mant[one_pos] |= one_mask;
893 if (type == FL_QNAN)
894 set_bit(mant, fmt->exponent+fmt->explicit+1);
895 else if (type == FL_SNAN)
896 set_bit(mant, fmt->exponent+fmt->explicit+fmt->mantissa);
897 break;
900 mant[0] |= minus ? LIMB_TOP_BIT : 0;
902 for (i = fmt->bytes - 1; i >= 0; i--)
903 *result++ = mant[i/LIMB_BYTES] >> (((LIMB_BYTES-1)-(i%LIMB_BYTES))*8);
905 return 1; /* success */
908 int float_const(const char *number, int sign, uint8_t *result, int bytes)
910 switch (bytes) {
911 case 1:
912 return to_float(number, sign, result, &ieee_8);
913 case 2:
914 return to_float(number, sign, result, &ieee_16);
915 case 4:
916 return to_float(number, sign, result, &ieee_32);
917 case 8:
918 return to_float(number, sign, result, &ieee_64);
919 case 10:
920 return to_float(number, sign, result, &ieee_80);
921 case 16:
922 return to_float(number, sign, result, &ieee_128);
923 default:
924 nasm_panic("strange value %d passed to float_const", bytes);
925 return 0;
929 /* Set floating-point options */
930 int float_option(const char *option)
932 if (!nasm_stricmp(option, "daz")) {
933 daz = true;
934 return 0;
935 } else if (!nasm_stricmp(option, "nodaz")) {
936 daz = false;
937 return 0;
938 } else if (!nasm_stricmp(option, "near")) {
939 rc = FLOAT_RC_NEAR;
940 return 0;
941 } else if (!nasm_stricmp(option, "down")) {
942 rc = FLOAT_RC_DOWN;
943 return 0;
944 } else if (!nasm_stricmp(option, "up")) {
945 rc = FLOAT_RC_UP;
946 return 0;
947 } else if (!nasm_stricmp(option, "zero")) {
948 rc = FLOAT_RC_ZERO;
949 return 0;
950 } else if (!nasm_stricmp(option, "default")) {
951 rc = FLOAT_RC_NEAR;
952 daz = false;
953 return 0;
954 } else {
955 return -1; /* Unknown option */