fxlibc/src/stdlib/strto_fp.c

#include "stdlib_p.h"
#include <stdlib.h>
#include <stdbool.h>

#include <float.h>
#include <fenv.h>
#include <math.h>

#include <string.h>
#include <errno.h>
#include <ctype.h>

/*
** In the following conversions, the significant digits are represented in an
** integer and multiplied at the last moment by a suitable power of 10 (decimal
** representation) or 2 (hexadecimal representation). An integer of a suitable
** size needs to be used; that size is the size of the long double type.
**
** TODO: vhex-x86: Using 128-bit long double is untested!
*/
#if __SIZEOF_LONG_DOUBLE__ == 8
# define SIGNIFICAND_TYPE uint64_t
# define SIGNIFICAND_DIGITS 17
#elif __SIZEOF_LONG_DOUBLE__ <= 16
# define SIGNIFICAND_TYPE unsigned __int128
# define SIGNIFICAND_DIGITS 38
#else
# error long double larger than 128 bits is not supported
#endif

/*
** Parse digits and exponent into integers, in decimal or hexadecimal notation.
**
** -> In decimal notation; we read up to 19 (64-bit) or 38 (128-bit) digits,
*     which is enough to fill the mantissa of a long double, and later multiply
**    the digits by a power of 10. The main approximation is the power of 10.
**
** -> In hexadecimal notation, we read as many bits as the mantissa of a long
**    double, then later multiply by a power of 2. There are no approximations.
*/
static bool parse_digits(struct __scanf_input *input,
	SIGNIFICAND_TYPE *digits, long *exponent, bool hexadecimal)
{
	bool dot_found = false;
	int digits_found=0, c=0;

	*digits = 0;
	*exponent = 0;

	int max_digits = hexadecimal ? LDBL_MANT_DIG / 4 : SIGNIFICAND_DIGITS;

	/* TODO: locale: use a locale-aware decimal separator */
	int dot_character = '.';
	int exp_character = (hexadecimal ? 'p' : 'e');

	for(int i = 0; true; i++) {
		c = __scanf_peek(input);
		if(!(isdigit(c) ||
		    (hexadecimal && isxdigit(c)) ||
		    (c == dot_character && !dot_found))) break;
		__scanf_in(input);

		if(c == dot_character) {
			dot_found = true;
			continue;
		}

		/* Count digits only until SIGNIFICAND_DIGITS */
		if(digits_found < max_digits) {
			if(hexadecimal) {
				int v = c - '0';
				if(!isdigit(c)) v = tolower(c) - 'a' + 10;
				*digits = (*digits << 4) + v;
			}
			else {
				*digits = (*digits * 10) + (c - '0');
			}
		}
		else (*exponent)++;

		if(dot_found) (*exponent)--;

		/* But also round at the first discarded one */
		if(digits_found == max_digits && c >= '5')
			(*digits)++;

		digits_found++;
	}

	/* Require at least one digit to be present; if not, the whole string
	   is considered invalid */
	if(!digits_found)
		return false;

	/* In hexadecimal, each character is worth 4 bits of exponent */
	if(hexadecimal) (*exponent) *= 4;

	/* Parse exponent */
	if(tolower(__scanf_peek(input)) == exp_character) {
		/* Hack: Restore the str pointer if this fails (which we
		   cannot determine with a single lookahead) so that *endptr is
		   set correctly */
		struct __scanf_input backup = *input;

		__scanf_in(input);
		long e = 0;
		if(__strto_int(input, 10, &e, NULL, false) == 0)
			*exponent += e;
		else
			*input = backup;
	}

	return true;
}

static bool expect(struct __scanf_input *input, char const *sequence)
{
	for(int i = 0; sequence[i]; i++) {
		int c = __scanf_in(input);
		if(tolower(c) != tolower(sequence[i]))
			return false;
	}
	return true;
}

int __strto_fp(struct __scanf_input *input, double *out, float *outf,
	long double *outl)
{
	/* Skip initial whitespace */
	while(isspace(__scanf_peek(input))) __scanf_in(input);

	/* Read optional sign */
	bool negative = false;
	int sign = __scanf_peek(input);
	if(sign == '-') negative = true;
	if(sign == '-' || sign == '+') __scanf_in(input);

	int errno_value = 0;
	bool valid = false;

	/* Result variable */
	if(out)  *out = 0.0;
	if(outf) *outf = 0.0f;
	if(outl) *outl = 0.0l;

	/* NaN possibly with an argument */
	if(tolower(__scanf_peek(input)) == 'n') {
		if(!expect(input, "nan"))
			return EINVAL;

		/* Get the argument for up to 32 bytes */
		char arg[32];
		int i = 0;

		if(__scanf_peek(input) == '(') {
			while(i < 31) {
				int c = __scanf_in(input);
				if(c == ')') break;
				arg[i++] = c;
			}
			arg[i] = 0;
		}

		if(out)  *out  = __builtin_nan(arg);
		if(outf) *outf = __builtin_nanf(arg);
		if(outl) *outl = __builtin_nanl(arg);
		valid = true;
	}
	else if(tolower(__scanf_peek(input)) == 'i') {
		if(!expect(input, "inf"))
			return EINVAL;
		if(tolower(__scanf_peek(input)) == 'i' &&
		   !expect(input, "inity"))
			return EINVAL;
		if(out)  *out  = __builtin_inf();
		if(outf) *outf = __builtin_inff();
		if(outl) *outl = __builtin_infl();
		valid = true;
	}
	else {
		SIGNIFICAND_TYPE digits = 0;
		long e = 0;

		/* Check for the 0x prefix. Skipping a 0 if we start with 0 but
		   not 0x isn't a problem. */
		bool hexa = false;
		if(__scanf_peek(input) == '0') {
			__scanf_in(input);
			if(tolower(__scanf_peek(input)) == 'x') {
				__scanf_in(input);
				hexa = true;
			}
			/* Count the 0 as a digit */
			else valid = true;
		}

		if(hexa) {
			valid |= parse_digits(input, &digits, &e, true);
			if(out)  *out  = (double)digits * exp2(e);
			if(outf) *outf = (float)digits * exp2f(e);
			if(outl) *outl = (long double)digits * exp2l(e);
		}
		else {
			valid |= parse_digits(input, &digits, &e, false);
			if(out)  *out  = (double)digits * pow(10, e);
			if(outf) *outf = (float)digits * powf(10, e);
			if(outl) *outl = (long double)digits * powl(10, e);
		}

		/*
		** Detect overflow, somewhat. Implementation is not required to
		** set errno on underflow, which makes things much easier for
		** us as underflow gives 0 (7.20.1.3§10).
		*/
		if((out && *out == HUGE_VAL)
			|| (outf && *outf == HUGE_VALF)
			|| (outl && *outl == HUGE_VALL)) {
			errno_value = ERANGE;
		}
	}

	/* Apply sign; this method is allowed by 7.20.1.3§4.249 */
	if(negative) {
		if(out)  *out  = -(*out);
		if(outf) *outf = -(*outf);
		if(outl) *outl = -(*outl);
	}

	return valid ? errno_value : EINVAL;
}