string_of_float returns -NaN for most numbers formatted by printf

[timbertson]

I found this issue where yojson objects with timestamps (floats) were being serialized OK, but when deserialized they would came back as -NaN.

To test, I ran the following code:

    let float_test f : unit =
      let s = string_of_float f in
      let s_printf = Printf.sprintf "%f" f in
      Printf.printf "float=%f -> string_of_float=%s -> parsed=%f\n" f s (float_of_string s);
      Printf.printf "float=%f -> printf-string  =%s -> parsed=%f\n" f s_printf (float_of_string s_printf)
    in

    float_test 2.4;
    float_test 1493895831.871153;
    let attempts = ref 500 in
    while !attempts > 0 do
      decr attempts;
      let f = Random.float 999999999999999.0 in
      float_test f;
      float_test f
    done;

I got the following output in mirage-ukvm, with ocaml-freestanding 0.2.1:

float=2.400000 -> string_of_float=2.4 -> parsed=2.400000
float=2.400000 -> printf-string  =2.400000 -> parsed=2.400000
float=1493895831.871153 -> string_of_float=1493895831.87 -> parsed=1493895831.870000
float=1493895831.871153 -> printf-string  =1493895831.871153 -> parsed=-nan
float=481705019586434.312500 -> string_of_float=4.81705019586e+14 -> parsed=481705019586000.000000
float=481705019586434.312500 -> printf-string  =481705019586434.312500 -> parsed=-nan
float=481705019586434.312500 -> string_of_float=4.81705019586e+14 -> parsed=481705019586000.000000
float=481705019586434.312500 -> printf-string  =481705019586434.312500 -> parsed=-nan
float=554380010410619.812500 -> string_of_float=5.54380010411e+14 -> parsed=554380010411000.000000
float=554380010410619.812500 -> printf-string  =554380010410619.812500 -> parsed=-nan
float=554380010410619.812500 -> string_of_float=5.54380010411e+14 -> parsed=554380010411000.000000
float=554380010410619.812500 -> printf-string  =554380010410619.812500 -> parsed=-nan

i.e. when using string_of_float the number correctly roundtrips to a string and back, but using printf's float formatting it consistently gets parsed as -NaN.
2.4 was a random simple case I tried, and funnily enough it works both ways - perhaps the bug only affects numbers with a large number of decimal places?

I remember float_of_string used to not work at all in xen because strtod was missing(mirage/mirage-platform#118), so I'm guessing this is probably a bug in ocaml-freestanding's strtod?

[mato]

That's possible; the strtod implementation in ocaml-freestanding comes from the standalone dtoa.c by David M. Gay, I've not really tested it much. If you could provide a test case in C that reproduces the problem I'll look into it.

(Aside: I wanted to use the musl implementation of strtod but that depends on dragging in more of stdio, so using the standalone dtoa.c was the more expedient solution at the time. Feel free to take this path instead...)

[hannesm]

FWIW the dtoa.c in mirage-xen and here are identical.. (modulo -Wmaybe-uninitialized)

[timbertson]

To clarify: I never got this far on xen since it lacked strtod last time I tried, so this bug could definitely be there too.

I whipped up a quick C version with just one problematic number. Results were... baffling:

#include <stdio.h>
#include "dtoa.c"

#define buflen 20

int main() {
	double f = 1493895831.871153;
	char buf[buflen];
	snprintf(buf, buflen, "%F", f);

	double f_parsed = strtod(buf, NULL);
	printf("double = %F; str = %s; double_of_str = %F\n", f, buf, f_parsed);
}

Compiling the above with gcc -O2 -std=c99 -Wall -Wno-parentheses -Werror float_test.c gives:

double = 1493895831.871153; str = 1493895831.871153; double_of_str = 1493895831.871153

But without -O2:

double = 1493895831.871153; str = 1493895831.871153; double_of_str = -NAN

And if I change the snprintf line to the seemingly-identical:

	snprintf(buf, buflen, "%s", "1493895831.871153");

... then with or without -O2, I get:

double = 1493895831.871153; str = 1493895831.871153; double_of_str = -NAN

This seems consistent on my machine, but I'd be interested if others get the same results.

[mato]

Interesting. I get exactly the same results as you on Debian 9 (gcc 6.3.0) but on Debian 8 (gcc 4.9.2) I get double_of_str = -NAN regardless of which optimization setting I use.

[mato]

Fixed in #28.