[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <f9012138-c4c3-4c5a-9dbd-717167cf1c85@lenardszolnoki.com>
Date: Wed, 9 Apr 2025 14:20:04 +0100
From: Lénárd Szolnoki <cpp@...ardszolnoki.com>
To: musl@...ts.openwall.com
Cc: Szabolcs Nagy <nsz@...t70.net>
Subject: Re: strtod doesn't set errno to ERANGE for slightly
out-of-range hex float
On 09/04/2025 13:55, Szabolcs Nagy wrote:
> * Lénárd Szolnoki <cpp@...ardszolnoki.com> [2025-04-09 11:38:01 +0100]:
>> On 08/04/2025 23:25, Szabolcs Nagy wrote:
>>> * Lénárd Szolnoki <cpp@...ardszolnoki.com> [2025-03-24 10:55:36 +0000]:
>>>> Hi,
>>>>
>>>> #include <stdlib.h>
>>>> #include <assert.h>
>>>> #include <errno.h>
>>>>
>>>> int main() {
>>>> const char* const input = "0x1.fffffffffffff8p+1023";
>>>> double ret = strtod(input, NULL);
>>>> assert(errno == ERANGE);
>>>> }
>>>>
>>>> The input is slightly out of range, DBL_MAX is 1.fffffffffffffp+1023 (one
>>>> fewer binary 1 in the mantissa).
>>>>
>>>> The assert fails with musl libc, passes with glibc.
> ...
>>> the overflow threshold is not calculated precisely
>>> for the hexfloat case. __floatscan returns an in
>>> range long double value that is rounded to inf by
>>> strtod, but there is no overflow check there.
>>
>> An other test case is with a different behaviour is
>> "0x1.fffffffffffff00000000000000000000000000000000000000000000001p+1023"
>>
>> This is even more slightly larger than DBL_MAX, therefore it should still
>> return inf and set errno to ERANGE.
>>
>
> in nearest rounding mode i'd expect DBL_MAX.
It appears that you are right. I expected DBL_MAX to be the threshold for all rounding
modes, but I double checked the IEEE standard and it seems that for nearest rounding mode
the threshold is exactly my first example.
>
> in upward rounding mode musl seems to return inf.
>
>> For this one strtod returns DBL_MAX. What happens is that this rounds
>> towards the value of DBL_MAX within long double, and the conversion from
>> long double to double is exact, so it does not overflow. Sort of a double
>> rounding problem, but at the edge of the range of double.
>
> there is only a single rounding, and i think that's fine.
>
>
>>>
>>> i guess hexfloat overflow check should be fixed.
>>>
>>>>
>>>> musl version: 1.2.5 (as shipped in Debian sid)
>>>>
>>>> Regards,
>>>> Lénárd Szolnoki
Powered by blists - more mailing lists
Confused about mailing lists and their use?
Read about mailing lists on Wikipedia
and check out these
guidelines on proper formatting of your messages.
