stdlib: add and test strtold (DONE)

The method is rather naive - digits read as an integer, then multipled
by a power of 10 or 2. This does not always give exact results, but it's
close enough for now. A stub support for long double larger than 64 bits
is provided.

CMakeLists.txt

@@ -121,6 +121,7 @@ set(SOURCES
   src/libc/stdlib/reallocarray.c
   src/libc/stdlib/strto_int.c
   src/libc/stdlib/strtol.c
+  src/libc/stdlib/strtold.c
   src/libc/stdlib/strtoll.c
   src/libc/stdlib/strtoul.c
   src/libc/stdlib/strtoull.c

STATUS

@@ -89,7 +89,8 @@ DONE:       Function/symbol/macro is defined, builds, links, and is tested
 7.20 <stdlib.h>
 !    7.20.1.1  atof: TODO
      7.20.1.2  atoi, atol, atoll: DONE
-!    7.20.1.3  strtod, strtof, strtold: TODO
+!    7.20.1.3  strtod, strtof: TODO
+               strtold: DONE
      7.20.1.4  strtol, strtoul, strtoll, strtoull: DONE
 !    7.20.2  Pseudo-random sequence generation functions: TODO
 !    7.20.3  Memory management functions: TODO (check existing code first)

src/libc/stdlib/strtold.c

@@ -0,0 +1,193 @@
+#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
+
+/* Basically strncasecmp. */
+static int ncasecmp(char const *left, char const *right, size_t n)
+{
+	for(size_t i = 0; i < n; i++) {
+		int diff = tolower(left[i]) - tolower(right[i]);
+		if(diff) return diff;
+	}
+	return 0;
+}
+
+/*
+** 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 void parse_digits(char const * restrict *ptr0, bool *valid,
+	SIGNIFICAND_TYPE *digits, long *exponent, bool hexadecimal)
+{
+	char const *ptr = *ptr0;
+	bool dot_found = false;
+	int digits_found = 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; isdigit(*ptr) || (hexadecimal && isxdigit(*ptr))
+		|| *ptr == dot_character; i++, ptr++) {
+
+		/* Allow only one dot in the string, stop at the second one */
+		if(*ptr == dot_character && dot_found) break;
+
+		if(*ptr == dot_character) {
+			dot_found = true;
+			continue;
+		}
+
+		/* Count digits only until SIGNIFICAND_DIGITS */
+		if(digits_found < max_digits) {
+			if(hexadecimal) {
+				int v = *ptr - '0';
+				if(!isdigit(*ptr)) v = tolower(*ptr)-'a'+10;
+				*digits = (*digits << 4) + v;
+			}
+			else {
+				*digits = (*digits * 10) + (*ptr - '0');
+			}
+		}
+		else (*exponent)++;
+
+		if(dot_found) (*exponent)--;
+
+		/* But also round at the first discarded one */
+		if(digits_found == max_digits && *ptr >= '5')
+			(*digits)++;
+
+		digits_found++;
+	}
+
+	/* Require at least one digit to be present; if not, the whole string
+	   is considered invalid */
+	if(!digits_found) {
+		*valid = false;
+		return;
+	}
+
+	/* In hexadecimal, each character is worth 4 bits of exponent */
+	if(hexadecimal) (*exponent) *= 4;
+
+	/* Parse exponent */
+	if(tolower(*ptr) == exp_character) {
+		char *end;
+		long e = strtol(ptr + 1, &end, 10);
+
+		/* If an integer cannot be parsed, ignore the 'e...' part */
+		if(end != ptr + 1) {
+			ptr = end;
+			*exponent += e;
+		}
+	}
+
+	*ptr0 = ptr;
+	*valid = true;
+}
+
+long double strtold(char const * restrict ptr, char ** restrict endptr)
+{
+	/* Save the value of ptr in endptr, in case format is invalid */
+	if(endptr) *endptr = (char *)ptr;
+
+	/* Skip initial whitespace */
+	while(isspace(*ptr)) ptr++;
+
+	/* Read optional sign */
+	bool negative = false;
+	if(*ptr == '-') negative = true;
+	if(*ptr == '-' || *ptr == '+') ptr++;
+
+	/* Result variable */
+	bool valid = true;
+	long double x = 0.0;
+
+	/* NaN possibly with an argument */
+	if(!ncasecmp(ptr, "nan", 3)) {
+		if(ptr[3] == '(') {
+			x = __builtin_nanl(ptr+4);
+			while(*ptr != ')') ptr++;
+		}
+		else {
+			x = __builtin_nanl("");
+			ptr += 3;
+		}
+	}
+	/* Infinity */
+	else if(!ncasecmp(ptr, "infinity", 8)) {
+		x = __builtin_infl();
+		ptr += 8;
+	}
+	else if(!ncasecmp(ptr, "inf", 3)) {
+		x = __builtin_infl();
+		ptr += 3;
+	}
+	else {
+		SIGNIFICAND_TYPE digits = 0;
+		long exponent = 0;
+
+		if(ptr[0] == '0' && tolower(ptr[1]) == 'x') {
+			ptr += 2;
+			parse_digits(&ptr, &valid, &digits, &exponent, true);
+			x = (long double)digits * exp2(exponent);
+		}
+		else {
+			parse_digits(&ptr, &valid, &digits, &exponent, false);
+			x = (long double)digits * powl(10, exponent);
+		}
+
+		/*
+		** 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(x == HUGE_VALL) {
+			errno = ERANGE;
+		}
+	}
+
+	/* Apply sign; this method is allowed by 7.20.1.3§4.249 */
+	if(negative) x = -x;
+
+	/* Save the result pointer */
+	if(endptr && valid) *endptr = (char *)ptr;
+
+	return x;
+}