1 /* float.c floating-point constant support for the Netwide Assembler
3 * The Netwide Assembler is copyright (C) 1996 Simon Tatham and
4 * Julian Hall. All rights reserved. The software is
5 * redistributable under the license given in the file "LICENSE"
6 * distributed in the NASM archive.
8 * initial version 13/ix/96 by Simon Tatham
28 static bool daz = false; /* denormals as zero */
29 static enum float_round rc = FLOAT_RC_NEAR; /* rounding control */
37 /* "A limb is like a digit but bigger */
38 typedef uint32_t fp_limb;
39 typedef uint64_t fp_2limb;
42 #define LIMB_BYTES (LIMB_BITS/8)
43 #define LIMB_TOP_BIT ((fp_limb)1 << (LIMB_BITS-1))
44 #define LIMB_MASK ((fp_limb)(~0))
45 #define LIMB_ALL_BYTES ((fp_limb)0x01010101)
46 #define LIMB_BYTE(x) ((x)*LIMB_ALL_BYTES)
48 /* 112 bits + 64 bits for accuracy + 16 bits for rounding */
51 /* 52 digits fit in 176 bits because 10^53 > 2^176 > 10^52 */
52 #define MANT_DIGITS 52
54 /* the format and the argument list depend on MANT_LIMBS */
55 #define MANT_FMT "%08x_%08x_%08x_%08x_%08x_%08x"
56 #define MANT_ARG SOME_ARG(mant, 0)
58 #define SOME_ARG(a,i) (a)[(i)+0], (a)[(i)+1], (a)[(i)+2], (a)[(i)+3], \
59 (a)[(i)+4], (a)[(i)+5]
62 * ---------------------------------------------------------------------------
63 * emit a printf()-like debug message... but only if DEBUG_FLOAT was defined
64 * ---------------------------------------------------------------------------
68 #define dprintf(x) printf x
70 #define dprintf(x) do { } while (0)
74 * ---------------------------------------------------------------------------
76 * ---------------------------------------------------------------------------
78 static int float_multiply(fp_limb *to, fp_limb *from)
80 fp_2limb temp[MANT_LIMBS * 2];
84 * guaranteed that top bit of 'from' is set -- so we only have
85 * to worry about _one_ bit shift to the left
87 dprintf(("%s=" MANT_FMT "\n", "mul1", SOME_ARG(to, 0)));
88 dprintf(("%s=" MANT_FMT "\n", "mul2", SOME_ARG(from, 0)));
90 memset(temp, 0, sizeof temp);
92 for (i = 0; i < MANT_LIMBS; i++) {
93 for (j = 0; j < MANT_LIMBS; j++) {
95 n = (fp_2limb) to[i] * (fp_2limb) from[j];
96 temp[i + j] += n >> LIMB_BITS;
97 temp[i + j + 1] += (fp_limb)n;
101 for (i = MANT_LIMBS * 2; --i;) {
102 temp[i - 1] += temp[i] >> LIMB_BITS;
103 temp[i] &= LIMB_MASK;
106 dprintf(("%s=" MANT_FMT "_" MANT_FMT "\n", "temp", SOME_ARG(temp, 0),
107 SOME_ARG(temp, MANT_LIMBS)));
109 if (temp[0] & LIMB_TOP_BIT) {
110 for (i = 0; i < MANT_LIMBS; i++) {
111 to[i] = temp[i] & LIMB_MASK;
113 dprintf(("%s=" MANT_FMT " (%i)\n", "prod", SOME_ARG(to, 0), 0));
116 for (i = 0; i < MANT_LIMBS; i++) {
117 to[i] = (temp[i] << 1) + !!(temp[i + 1] & LIMB_TOP_BIT);
119 dprintf(("%s=" MANT_FMT " (%i)\n", "prod", SOME_ARG(to, 0), -1));
125 * ---------------------------------------------------------------------------
126 * read an exponent; returns INT32_MAX on error
127 * ---------------------------------------------------------------------------
129 static int32_t read_exponent(const char *string, int32_t max)
134 if (*string == '+') {
136 } else if (*string == '-') {
141 if (*string >= '0' && *string <= '9') {
142 i = (i * 10) + (*string - '0');
145 * To ensure that underflows and overflows are
146 * handled properly we must avoid wraparounds of
147 * the signed integer value that is used to hold
148 * the exponent. Therefore we cap the exponent at
149 * +/-5000, which is slightly more/less than
150 * what's required for normal and denormal numbers
151 * in single, double, and extended precision, but
152 * sufficient to avoid signed integer wraparound.
156 } else if (*string == '_') {
159 error(ERR_NONFATAL|ERR_PASS1,
160 "invalid character in floating-point constant %s: '%c'",
161 "exponent", *string);
171 * ---------------------------------------------------------------------------
173 * ---------------------------------------------------------------------------
175 static bool ieee_flconvert(const char *string, fp_limb *mant,
178 char digits[MANT_DIGITS];
180 fp_limb mult[MANT_LIMBS], bit;
182 int32_t tenpwr, twopwr;
184 bool started, seendot, warned;
189 started = seendot = false;
191 while (*string && *string != 'E' && *string != 'e') {
192 if (*string == '.') {
196 error(ERR_NONFATAL|ERR_PASS1,
197 "too many periods in floating-point constant");
200 } else if (*string >= '0' && *string <= '9') {
201 if (*string == '0' && !started) {
207 if (p < digits + sizeof(digits)) {
208 *p++ = *string - '0';
211 error(ERR_WARNING|ERR_WARN_FL_TOOLONG|ERR_PASS1,
212 "floating-point constant significand contains "
213 "more than %i digits", MANT_DIGITS);
221 } else if (*string == '_') {
224 error(ERR_NONFATAL|ERR_PASS1,
225 "invalid character in floating-point constant %s: '%c'",
226 "significand", *string);
235 string++; /* eat the E */
236 e = read_exponent(string, 5000);
243 * At this point, the memory interval [digits,p) contains a
244 * series of decimal digits zzzzzzz, such that our number X
245 * satisfies X = 0.zzzzzzz * 10^tenpwr.
250 dprintf(("%c", *q + '0'));
253 dprintf((" * 10^%i\n", tenpwr));
256 * Now convert [digits,p) to our internal representation.
259 for (m = mant; m < mant + MANT_LIMBS; m++) {
266 while (m < mant + MANT_LIMBS) {
268 while (p > q && !p[-1]) {
274 for (r = p; r-- > q;) {
303 * At this point, the 'mant' array contains the first frac-
304 * tional places of a base-2^16 real number which when mul-
305 * tiplied by 2^twopwr and 5^tenpwr gives X.
307 dprintf(("X = " MANT_FMT " * 2^%i * 5^%i\n", MANT_ARG, twopwr,
311 * Now multiply 'mant' by 5^tenpwr.
313 if (tenpwr < 0) { /* mult = 5^-1 = 0.2 */
314 for (m = mult; m < mult + MANT_LIMBS - 1; m++) {
315 *m = LIMB_BYTE(0xcc);
317 mult[MANT_LIMBS - 1] = LIMB_BYTE(0xcc)+1;
322 * If tenpwr was 1000...000b, then it becomes 1000...000b. See
323 * the "ANSI C" comment below for more details on that case.
325 * Because we already truncated tenpwr to +5000...-5000 inside
326 * the exponent parsing code, this shouldn't happen though.
328 } else if (tenpwr > 0) { /* mult = 5^+1 = 5.0 */
329 mult[0] = (fp_limb)5 << (LIMB_BITS-3); /* 0xA000... */
330 for (m = mult + 1; m < mult + MANT_LIMBS; m++) {
338 dprintf(("loop=" MANT_FMT " * 2^%i * 5^%i (%i)\n", MANT_ARG,
339 twopwr, tenpwr, extratwos));
341 dprintf(("mant*mult\n"));
342 twopwr += extratwos + float_multiply(mant, mult);
344 dprintf(("mult*mult\n"));
345 extratwos = extratwos * 2 + float_multiply(mult, mult);
349 * In ANSI C, the result of right-shifting a signed integer is
350 * considered implementation-specific. To ensure that the loop
351 * terminates even if tenpwr was 1000...000b to begin with, we
352 * manually clear the MSB, in case a 1 was shifted in.
354 * Because we already truncated tenpwr to +5000...-5000 inside
355 * the exponent parsing code, this shouldn't matter; neverthe-
356 * less it is the right thing to do here.
358 tenpwr &= (uint32_t) - 1 >> 1;
362 * At this point, the 'mant' array contains the first frac-
363 * tional places of a base-2^16 real number in [0.5,1) that
364 * when multiplied by 2^twopwr gives X. Or it contains zero
365 * of course. We are done.
372 * ---------------------------------------------------------------------------
373 * operations of specific bits
374 * ---------------------------------------------------------------------------
377 /* Set a bit, using *bigendian* bit numbering (0 = MSB) */
378 static void set_bit(fp_limb *mant, int bit)
380 mant[bit/LIMB_BITS] |= LIMB_TOP_BIT >> (bit & (LIMB_BITS-1));
383 /* Test a single bit */
384 static int test_bit(const fp_limb *mant, int bit)
386 return (mant[bit/LIMB_BITS] >> (~bit & (LIMB_BITS-1))) & 1;
389 /* Report if the mantissa value is all zero */
390 static bool is_zero(const fp_limb *mant)
394 for (i = 0; i < MANT_LIMBS; i++)
402 * ---------------------------------------------------------------------------
403 * round a mantissa off after i words
404 * ---------------------------------------------------------------------------
407 #define ROUND_COLLECT_BITS \
409 m = mant[i] & (2*bit-1); \
410 for (j = i+1; j < MANT_LIMBS; j++) \
414 #define ROUND_ABS_DOWN \
416 mant[i] &= ~(bit-1); \
417 for (j = i+1; j < MANT_LIMBS; j++) \
422 #define ROUND_ABS_UP \
424 mant[i] = (mant[i] & ~(bit-1)) + bit; \
425 for (j = i+1; j < MANT_LIMBS; j++) \
427 while (i > 0 && !mant[i]) \
432 static bool ieee_round(bool minus, fp_limb *mant, int bits)
436 int i = bits / LIMB_BITS;
437 int p = bits % LIMB_BITS;
438 fp_limb bit = LIMB_TOP_BIT >> p;
440 if (rc == FLOAT_RC_NEAR) {
448 if (test_bit(mant, bits-1)) {
457 } else if (rc == FLOAT_RC_ZERO ||
458 rc == (minus ? FLOAT_RC_UP : FLOAT_RC_DOWN)) {
461 /* rc == (minus ? FLOAT_RC_DOWN : FLOAT_RC_UP) */
462 /* Round toward +/- infinity */
473 /* Returns a value >= 16 if not a valid hex digit */
474 static unsigned int hexval(char c)
476 unsigned int v = (unsigned char) c;
478 if (v >= '0' && v <= '9')
481 return (v|0x20) - 'a' + 10;
484 /* Handle floating-point numbers with radix 2^bits and binary exponent */
485 static bool ieee_flconvert_bin(const char *string, int bits,
486 fp_limb *mant, int32_t *exponent)
488 static const int log2tbl[16] =
489 { -1, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3 };
490 fp_limb mult[MANT_LIMBS + 1], *mp;
493 bool seendot, seendigit;
495 int radix = 1 << bits;
499 seendot = seendigit = false;
503 memset(mult, 0, sizeof mult);
505 while ((c = *string++) != '\0') {
510 error(ERR_NONFATAL|ERR_PASS1,
511 "too many periods in floating-point constant");
514 } else if ((v = hexval(c)) < (unsigned int)radix) {
515 if (!seendigit && v) {
520 ms = (LIMB_BITS-1)-l;
522 twopwr = seendot ? twopwr-bits+l : l+1-bits;
529 if (mp > &mult[MANT_LIMBS])
530 mp = &mult[MANT_LIMBS]; /* Guard slot */
542 } else if (c == 'p' || c == 'P') {
544 e = read_exponent(string, 20000);
549 } else if (c == '_') {
552 error(ERR_NONFATAL|ERR_PASS1,
553 "floating-point constant: `%c' is invalid character", c);
559 memset(mant, 0, sizeof mult); /* Zero */
562 memcpy(mant, mult, sizeof mult);
570 * Shift a mantissa to the right by i bits.
572 static void ieee_shr(fp_limb *mant, int i)
578 sr = i % LIMB_BITS; sl = LIMB_BITS-sr;
583 for (j = MANT_LIMBS-1; j >= offs; j--)
584 mant[j] = mant[j-offs];
586 n = mant[MANT_LIMBS-1-offs] >> sr;
587 for (j = MANT_LIMBS-1; j > offs; j--) {
589 mant[j] = (m << sl) | n;
598 /* Produce standard IEEE formats, with implicit or explicit integer
599 bit; this makes the following assumptions:
601 - the sign bit is the MSB, followed by the exponent,
602 followed by the integer bit if present.
603 - the sign bit plus exponent fit in 16 bits.
604 - the exponent bias is 2^(n-1)-1 for an n-bit exponent */
608 int mantissa; /* Fractional bits in the mantissa */
609 int explicit; /* Explicit integer */
610 int exponent; /* Bits in the exponent */
614 * The 16- and 128-bit formats are expected to be in IEEE 754r.
615 * AMD SSE5 uses the 16-bit format.
617 * The 32- and 64-bit formats are the original IEEE 754 formats.
619 * The 80-bit format is x87-specific, but widely used.
621 * The 8-bit format appears to be the consensus 8-bit floating-point
622 * format. It is apparently used in graphics applications.
624 static const struct ieee_format ieee_8 = { 1, 3, 0, 4 };
625 static const struct ieee_format ieee_16 = { 2, 10, 0, 5 };
626 static const struct ieee_format ieee_32 = { 4, 23, 0, 8 };
627 static const struct ieee_format ieee_64 = { 8, 52, 0, 11 };
628 static const struct ieee_format ieee_80 = { 10, 63, 1, 15 };
629 static const struct ieee_format ieee_128 = { 16, 112, 0, 15 };
631 /* Types of values we can generate */
641 static int to_float(const char *str, int s, uint8_t * result,
642 const struct ieee_format *fmt)
644 fp_limb mant[MANT_LIMBS];
645 int32_t exponent = 0;
646 int32_t expmax = 1 << (fmt->exponent - 1);
647 fp_limb one_mask = LIMB_TOP_BIT >>
648 ((fmt->exponent+fmt->explicit) % LIMB_BITS);
649 int one_pos = (fmt->exponent+fmt->explicit)/LIMB_BITS;
655 int bits = fmt->bytes * 8;
661 case 'n': /* __nan__ */
663 case 'q': /* __qnan__ */
667 case 's': /* __snan__ */
671 case 'i': /* __infinity__ */
676 error(ERR_NONFATAL|ERR_PASS1,
677 "internal error: unknown FP constant token `%s'\n", str);
686 ok = ieee_flconvert_bin(str+2, 4, mant, &exponent);
690 ok = ieee_flconvert_bin(str+2, 3, mant, &exponent);
694 ok = ieee_flconvert_bin(str+2, 1, mant, &exponent);
698 ok = ieee_flconvert(str+2, mant, &exponent);
701 /* Leading zero was just a zero? */
702 ok = ieee_flconvert(str, mant, &exponent);
705 } else if (str[0] == '$') {
706 ok = ieee_flconvert_bin(str+1, 4, mant, &exponent);
708 ok = ieee_flconvert(str, mant, &exponent);
713 } else if (mant[0] & LIMB_TOP_BIT) {
718 if (exponent >= 2 - expmax && exponent <= expmax) {
720 } else if (exponent > 0) {
722 error(ERR_WARNING|ERR_WARN_FL_OVERFLOW|ERR_PASS1,
723 "overflow in floating-point constant");
726 /* underflow or denormal; the denormal code handles
739 memset(mant, 0, sizeof mant);
744 shift = -(exponent + expmax - 2 - fmt->exponent)
746 ieee_shr(mant, shift);
747 ieee_round(minus, mant, bits);
748 if (mant[one_pos] & one_mask) {
749 /* One's position is set, we rounded up into normal range */
752 mant[one_pos] &= ~one_mask; /* remove explicit one */
753 mant[0] |= exponent << (LIMB_BITS-1 - fmt->exponent);
755 if (daz || is_zero(mant)) {
756 /* Flush denormals to zero */
757 error(ERR_WARNING|ERR_WARN_FL_UNDERFLOW|ERR_PASS1,
758 "underflow in floating-point constant");
761 error(ERR_WARNING|ERR_WARN_FL_DENORM|ERR_PASS1,
762 "denormal floating-point constant");
769 exponent += expmax - 1;
770 ieee_shr(mant, fmt->exponent+fmt->explicit);
771 ieee_round(minus, mant, bits);
772 /* did we scale up by one? */
773 if (test_bit(mant, fmt->exponent+fmt->explicit-1)) {
776 if (exponent >= (expmax << 1)-1) {
777 error(ERR_WARNING|ERR_WARN_FL_OVERFLOW|ERR_PASS1,
778 "overflow in floating-point constant");
785 mant[one_pos] &= ~one_mask; /* remove explicit one */
786 mant[0] |= exponent << (LIMB_BITS-1 - fmt->exponent);
793 memset(mant, 0, sizeof mant);
794 mant[0] = (((fp_limb)1 << fmt->exponent)-1)
795 << (LIMB_BITS-1 - fmt->exponent);
797 mant[one_pos] |= one_mask;
799 set_bit(mant, fmt->exponent+fmt->explicit+1);
800 else if (type == FL_SNAN)
801 set_bit(mant, fmt->exponent+fmt->explicit+fmt->mantissa);
805 mant[0] |= minus ? LIMB_TOP_BIT : 0;
807 for (i = fmt->bytes - 1; i >= 0; i--)
808 *result++ = mant[i/LIMB_BYTES] >> (((LIMB_BYTES-1)-(i%LIMB_BYTES))*8);
810 return 1; /* success */
813 int float_const(const char *number, int sign, uint8_t * result,
814 int bytes, efunc err)
820 return to_float(number, sign, result, &ieee_8);
822 return to_float(number, sign, result, &ieee_16);
824 return to_float(number, sign, result, &ieee_32);
826 return to_float(number, sign, result, &ieee_64);
828 return to_float(number, sign, result, &ieee_80);
830 return to_float(number, sign, result, &ieee_128);
832 error(ERR_PANIC, "strange value %d passed to float_const", bytes);
837 /* Set floating-point options */
838 int float_option(const char *option)
840 if (!nasm_stricmp(option, "daz")) {
843 } else if (!nasm_stricmp(option, "nodaz")) {
846 } else if (!nasm_stricmp(option, "near")) {
849 } else if (!nasm_stricmp(option, "down")) {
852 } else if (!nasm_stricmp(option, "up")) {
855 } else if (!nasm_stricmp(option, "zero")) {
858 } else if (!nasm_stricmp(option, "default")) {
863 return -1; /* Unknown option */