Bug report: strtod drops LSB

Bug report: strtod drops LSB

[Anastasios]

[-- Attachment #1: Type: text/plain, Size: 687 bytes --]

Hello,

Consider this program, strtod.c:

    #include <stdlib.h>
    #include <stdio.h>

    int main()
    {
        printf("%lf\n", strtod("283686952306183", NULL));
    }

With current musl master from Git:

    $ musl-gcc -static strtod.c -o a.musl
    $ ./a.musl
    283686952306176.000000

By comparison, with glibc:

    $ gcc -static strtod.c -o a.glibc
    $ ./a.glibc
    283686952306183.000000

The correct binary representation of this float is

    0x42f0203040506070

but musl strtod produces

    0x42f0203040506000

i.e., it fails to set the LSB. I examined this while ruling out printf as the cause.

[-- Attachment #2: Type: text/html, Size: 1670 bytes --]

Re: Bug report: strtod drops LSB

[Szabolcs Nagy]

* Anastasios <antonbachin@yahoo.com> [2019-09-25 09:32:35 -0500]:
> Hello,
> 
> Consider this program, strtod.c:
> 
>     #include <stdlib.h>
>     #include <stdio.h>
> 
>     int main()
>     {
>         printf("%lf\n", strtod("283686952306183", NULL));
>     }
> 
> With current musl master from Git:
> 
>     $ musl-gcc -static strtod.c -o a.musl
>     $ ./a.musl
>     283686952306176.000000
> 
> By comparison, with glibc:
> 
>     $ gcc -static strtod.c -o a.glibc
>     $ ./a.glibc
>     283686952306183.000000
> 

works for me on both i486-linux-musl and aarch64-linux-musl.
i have no idea what can cause such bug.

can you try printf with %a and strtod(...)==0x1.020304050607p+48
to see if it's printf %lf that goes wrong?

> The correct binary representation of this float is
> 
>     0x42f0203040506070
> 
> but musl strtod produces
> 
>     0x42f0203040506000
> 
> i.e., it fails to set the LSB. I examined this while ruling out printf as the cause.

Re: Bug report: strtod drops LSB

[Rich Felker]

On Wed, Sep 25, 2019 at 09:32:35AM -0500, Anastasios wrote:
> Hello,
> 
> Consider this program, strtod.c:
> 
>     #include <stdlib.h>
>     #include <stdio.h>
> 
>     int main()
>     {
>         printf("%lf\n", strtod("283686952306183", NULL));
>     }
> 
> With current musl master from Git:
> 
>     $ musl-gcc -static strtod.c -o a.musl
>     $ ./a.musl
>     283686952306176.000000
> 
> By comparison, with glibc:
> 
>     $ gcc -static strtod.c -o a.glibc
>     $ ./a.glibc
>     283686952306183.000000
> 
> The correct binary representation of this float is
> 
>     0x42f0203040506070
> 
> but musl strtod produces
> 
>     0x42f0203040506000
> 
> i.e., it fails to set the LSB. I examined this while ruling out printf as the cause.

I can't reproduce this. My test program for strtod shows, for the
input "283686952306183":

d:  283686952306183 [0x1.020304050607p+48] [42f0203040506070]

I suspect you miscompiled musl, possibly by passing in CFLAGS (perhaps
from defaults in your environment?) that break floating point
semantics. We test for and refuse to build if __FAST_MATH__ is
defined, but GCC only defines it if you use -ffast-math, not if you
manually enable one or more of the individual broken options that
-ffast-math enables.

Alternatively, it's possible that you have a broken compiler version
that miscompiles floating point code.

Rich

Re: Bug report: strtod drops LSB

[Anastasios]

[-- Attachment #1: Type: text/plain, Size: 1737 bytes --]

Having looked into it more, I found this is due to being on WSL, and it is a duplicate of https://github.com/microsoft/WSL/issues/830 [https://github.com/microsoft/WSL/issues/830], which, I see, you are already well aware of. Please excuse the noise.
On 9/25/2019 10:51:53 AM, Rich Felker <dalias@libc.org> wrote:
On Wed, Sep 25, 2019 at 09:32:35AM -0500, Anastasios wrote:
> Hello,
>
> Consider this program, strtod.c:
>
>     #include
>     #include
>
>     int main()
>     {
>         printf("%lf\n", strtod("283686952306183", NULL));
>     }
>
> With current musl master from Git:
>
>     $ musl-gcc -static strtod.c -o a.musl
>     $ ./a.musl
>     283686952306176.000000
>
> By comparison, with glibc:
>
>     $ gcc -static strtod.c -o a.glibc
>     $ ./a.glibc
>     283686952306183.000000
>
> The correct binary representation of this float is
>
>     0x42f0203040506070
>
> but musl strtod produces
>
>     0x42f0203040506000
>
> i.e., it fails to set the LSB. I examined this while ruling out printf as the cause.

I can't reproduce this. My test program for strtod shows, for the
input "283686952306183":

d: 283686952306183 [0x1.020304050607p+48] [42f0203040506070]

I suspect you miscompiled musl, possibly by passing in CFLAGS (perhaps
from defaults in your environment?) that break floating point
semantics. We test for and refuse to build if __FAST_MATH__ is
defined, but GCC only defines it if you use -ffast-math, not if you
manually enable one or more of the individual broken options that
-ffast-math enables.

Alternatively, it's possible that you have a broken compiler version
that miscompiles floating point code.

Rich

[-- Attachment #2: Type: text/html, Size: 2679 bytes --]

Re: Bug report: strtod drops LSB

[Anastasios]

[-- Attachment #1: Type: text/plain, Size: 2207 bytes --]

For the record, this is the program I used to confirm that it's a duplicate:

    #include <stdlib.h>
    #include <stdio.h>

    void main() {
        unsigned short control_word = 0x37f;
        asm ("fldcw %0" : : "m" (control_word));
        printf("%lf\n", strtod("283686952306183", NULL));
    }


Perhaps it can help if anyone on WSL reports this again.
On 9/25/2019 12:56:35 PM, Anastasios <antonbachin@yahoo.com> wrote:
Having looked into it more, I found this is due to being on WSL, and it is a duplicate of https://github.com/microsoft/WSL/issues/830 [https://github.com/microsoft/WSL/issues/830], which, I see, you are already well aware of. Please excuse the noise.
On 9/25/2019 10:51:53 AM, Rich Felker <dalias@libc.org> wrote:
On Wed, Sep 25, 2019 at 09:32:35AM -0500, Anastasios wrote:
> Hello,
>
> Consider this program, strtod.c:
>
>     #include
>     #include
>
>     int main()
>     {
>         printf("%lf\n", strtod("283686952306183", NULL));
>     }
>
> With current musl master from Git:
>
>     $ musl-gcc -static strtod.c -o a.musl
>     $ ./a.musl
>     283686952306176.000000
>
> By comparison, with glibc:
>
>     $ gcc -static strtod.c -o a.glibc
>     $ ./a.glibc
>     283686952306183.000000
>
> The correct binary representation of this float is
>
>     0x42f0203040506070
>
> but musl strtod produces
>
>     0x42f0203040506000
>
> i.e., it fails to set the LSB. I examined this while ruling out printf as the cause.

I can't reproduce this. My test program for strtod shows, for the
input "283686952306183":

d: 283686952306183 [0x1.020304050607p+48] [42f0203040506070]

I suspect you miscompiled musl, possibly by passing in CFLAGS (perhaps
from defaults in your environment?) that break floating point
semantics. We test for and refuse to build if __FAST_MATH__ is
defined, but GCC only defines it if you use -ffast-math, not if you
manually enable one or more of the individual broken options that
-ffast-math enables.

Alternatively, it's possible that you have a broken compiler version
that miscompiles floating point code.

Rich

[-- Attachment #2: Type: text/html, Size: 4835 bytes --]