Re: FPE in localtime.cc
- Date: Tue, 10 Jul 2018 09:35:11 -0400
- From: cyg Simple <cygsimple@xxxxxxxxx>
- Subject: Re: FPE in localtime.cc
On 7/9/2018 3:07 PM, Corinna Vinschen wrote:
> On Jul 9 11:16, Brian Inglis wrote:
>> On 2018-07-09 10:49, Corinna Vinschen wrote:
>>> On Jul 9 15:47, Lavrentiev, Anton (NIH/NLM/NCBI) [C] wrote:
>>>> the following sample coredumps with FPE at localhost.cc:1962 with the latest snapshot (6/29/2018):
>>> You can simplify your testcase by not calling any time functions:
>>> #define _GNU_SOURCE
>>> #include <fenv.h>
>>> #include <stdio.h>
>>> #include <stdlib.h>
>>> #define SECSPERDAY 86400
>>> int main(int argc, char **argv)
>>> long tdays = argc > 1 ? strtol (argv, NULL, 10) : 189;
>>> long seconds = tdays * SECSPERDAY + 0.5;
>>> printf ("%ld\n", seconds);
>>> This generates a SIGFPE on Linux as well.
>>> The line computing seconds is the same line as used by the localtime
>>> function. Cygwin shares the entire localtime code with the various
>>> BSDs, so I guess they would have the same problem.
>> What is that line meant to do? Am I missing something?
>> It should be the equivalent of (tdays*SECSPERDAY*2 + 1)/2!
>> It converts an integer value to double, adds 1/2, and throws it away on
>> conversion back, unless the intermediate has insufficient mantissa bits, in
>> which case, it loses the low bits?
> You may want to ask the original author why he used FP arithmetic in
> this place. Maybe it's a way to avoid integer overflow. I'm reluctant
> to change this given that this code is still used in BSD as well.
I don't see a fetestexcept() being used. Setting all the bits is
definitely going to raise an exception.
>>> Bottom line is, don't bulk enable FP exceptions, but only if you really
>>> need it for certain parts of your code. Don't expect library functions
>>> to be SIGFPE clean under all circumstances.
>> Maybe selectively enable specific FPEs to check for where needed.
>> Or be careful what you wish for, as you just might get a lot more than you
>> bargained for ;^>
> That's what I meant.
Problem reports: http://cygwin.com/problems.html
Unsubscribe info: http://cygwin.com/ml/#unsubscribe-simple