feraiseexcept does not raise traps in Cygwin

feraiseexcept does not raise traps in Cygwin

[Bruno Haible]

Hi,

Seen on
  - Cygwin 3.4.6 or 2.9.0 on x86_64.
  - Cygwin 2.9.0 on i386.

According to ISO C 23 § 7.6.4.3
  "The feraiseexcept function attempts to raise the supported floating-point
   exceptions represented by its argument. 266)
   Footnote 266) The effect is intended to be similar to that of floating-point
   exceptions raised by arithmetic operations. Hence, implementation extensions
   associated with raising a floating-point exception (for example, enabled
   traps or IEC 60559 alternate exception handling) should be honored."

This does not work. How to reproduce:
=================================== foo.c ===================================
#define _GNU_SOURCE 1
#include <fenv.h>
#include <assert.h>

int
main ()
{
  /* Clear FE_INVALID exceptions from past operations.  */
  if (feclearexcept (FE_INVALID))
    return 1;

  /* An FE_INVALID exception shall trigger a SIGFPE signal, which by default
     terminates the program.  */
  if (feenableexcept (FE_INVALID) == -1)
    return 2;

  if (feraiseexcept (FE_INVALID))
    return 3;

  return 0;
}
=============================================================================

For x86_64:

$ x86_64-pc-cygwin-gcc -Wall foo.c
$ ./a.exe; echo $?

For i386:

$ i686-pc-cygwin-gcc -Wall foo.c
$ ./a.exe; echo $?

Expected result (like seen e.g. on Linux/glibc):
Floating-point exception (core dumped)
136

Actual result:
0

The workaround for x86_64 is to redefine feeraiseexcept in the same way
as glibc does. This modified test program includes the workaround:

=================================== foo.c ===================================
#define _GNU_SOURCE 1
#include <fenv.h>
#include <assert.h>

/* The floating-point environment of the 387 unit.  */
typedef struct
  {
    /* 7 32-bit words:  */
    unsigned short __control_word;      /* fctrl register */
    unsigned short __reserved1;
    unsigned short __status_word;       /* fstat register */
    unsigned short __reserved2;
    unsigned int more[5];
  }
x86_387_fenv_t;

int
feraiseexcept (int exceptions)
{
  exceptions &= FE_ALL_EXCEPT;
  if ((exceptions & ~(FE_INVALID | FE_DIVBYZERO)) == 0 && 0)
    {
      /* First: invalid exception.  */
      if (exceptions & FE_INVALID)
        {
          double volatile a;
          double volatile b;
          a = 0; b = a / a;
          (void) b;
        }
      /* Next: division by zero.  */
      if (exceptions & FE_DIVBYZERO)
        {
          double volatile a, b;
          double volatile c;
          a = 1; b = 0; c = a / b;
          (void) c;
        }
    }
  else
    {
      /* The general case.  */

      /* Set the bits in the 387 unit.  */
      x86_387_fenv_t env;
      unsigned short orig_status_word;
      __asm__ __volatile__ ("fnstenv %0" : "=m" (*&env));
      orig_status_word = env.__status_word;
      env.__status_word |= exceptions;
      if (env.__status_word != orig_status_word)
        {
          __asm__ __volatile__ ("fldenv %0" : : "m" (*&env));
          /* A trap (if enabled) is triggered only at the next floating-point
             instruction.  Force it to occur here.  */
          __asm__ __volatile__ ("fwait");
        }
    }
  return 0;
}

int
main ()
{
  /* Clear FE_INVALID exceptions from past operations.  */
  if (feclearexcept (FE_INVALID))
    return 1;

  /* An FE_INVALID exception shall trigger a SIGFPE signal, which by default
     terminates the program.  */
  if (feenableexcept (FE_INVALID) == -1)
    return 2;

  if (feraiseexcept (FE_INVALID))
    return 3;

  return 0;
}
=============================================================================

(This workaround *should* also work on i386, but it doesn't. I don't know
why.)

Bruno

Re: feraiseexcept does not raise traps in newlib on x86/_64 [was Cygwin]

[Brian Inglis]

On 2023-10-25 03:21, Bruno Haible via Cygwin wrote:
> Hi,
> 
> Seen on
>    - Cygwin 3.4.6 or 2.9.0 on x86_64.
>    - Cygwin 2.9.0 on i386.
> 
> According to ISO C 23 § 7.6.4.3
>    "The feraiseexcept function attempts to raise the supported floating-point
>     exceptions represented by its argument. 266)
>     Footnote 266) The effect is intended to be similar to that of floating-point
>     exceptions raised by arithmetic operations. Hence, implementation extensions
>     associated with raising a floating-point exception (for example, enabled
>     traps or IEC 60559 alternate exception handling) should be honored."
> 
> This does not work. How to reproduce:
> =================================== foo.c ===================================
> #define _GNU_SOURCE 1
> #include <fenv.h>
> #include <assert.h>
> 
> int
> main ()
> {
>    /* Clear FE_INVALID exceptions from past operations.  */
>    if (feclearexcept (FE_INVALID))
>      return 1;
> 
>    /* An FE_INVALID exception shall trigger a SIGFPE signal, which by default
>       terminates the program.  */
>    if (feenableexcept (FE_INVALID) == -1)
>      return 2;
> 
>    if (feraiseexcept (FE_INVALID))
>      return 3;
> 
>    return 0;
> }
> =============================================================================
> 
> For x86_64:
> 
> $ x86_64-pc-cygwin-gcc -Wall foo.c
> $ ./a.exe; echo $?
> 
> For i386:
> 
> $ i686-pc-cygwin-gcc -Wall foo.c
> $ ./a.exe; echo $?
> 
> Expected result (like seen e.g. on Linux/glibc):
> Floating-point exception (core dumped)
> 136
> 
> Actual result:
> 0
> 
> The workaround for x86_64 is to redefine feeraiseexcept in the same way
> as glibc does. This modified test program includes the workaround:
> 
> =================================== foo.c ===================================
> #define _GNU_SOURCE 1
> #include <fenv.h>
> #include <assert.h>

   #ifdef REDEFFE

> /* The floating-point environment of the 387 unit.  */
> typedef struct
>    {
>      /* 7 32-bit words:  */
>      unsigned short __control_word;      /* fctrl register */
>      unsigned short __reserved1;
>      unsigned short __status_word;       /* fstat register */
>      unsigned short __reserved2;
>      unsigned int more[5];
>    }
> x86_387_fenv_t;
> 
> int
> feraiseexcept (int exceptions)
> {
>    exceptions &= FE_ALL_EXCEPT;
>    if ((exceptions & ~(FE_INVALID | FE_DIVBYZERO)) == 0 && 0)
>      {
>        /* First: invalid exception.  */
>        if (exceptions & FE_INVALID)
>          {
>            double volatile a;
>            double volatile b;
>            a = 0; b = a / a;
>            (void) b;
>          }
>        /* Next: division by zero.  */
>        if (exceptions & FE_DIVBYZERO)
>          {
>            double volatile a, b;
>            double volatile c;
>            a = 1; b = 0; c = a / b;
>            (void) c;
>          }
>      }
>    else
>      {
>        /* The general case.  */
> 
>        /* Set the bits in the 387 unit.  */
>        x86_387_fenv_t env;
>        unsigned short orig_status_word;
>        __asm__ __volatile__ ("fnstenv %0" : "=m" (*&env));
>        orig_status_word = env.__status_word;
>        env.__status_word |= exceptions;
>        if (env.__status_word != orig_status_word)
>          {
>            __asm__ __volatile__ ("fldenv %0" : : "m" (*&env));
>            /* A trap (if enabled) is triggered only at the next floating-point
>               instruction.  Force it to occur here.  */
>            __asm__ __volatile__ ("fwait");
>          }
>      }
>    return 0;
> }

   #endif /* REDEFFE */

> int
> main ()
> {
>    /* Clear FE_INVALID exceptions from past operations.  */
>    if (feclearexcept (FE_INVALID))
>      return 1;
> 
>    /* An FE_INVALID exception shall trigger a SIGFPE signal, which by default
>       terminates the program.  */
>    if (feenableexcept (FE_INVALID) == -1)
>      return 2;
> 
>    if (feraiseexcept (FE_INVALID))
>      return 3;
> 
>    return 0;
> }
> =============================================================================
> 
> (This workaround *should* also work on i386, but it doesn't. I don't know
> why.)

Thanks for the report Bruno,

Confirmed on Cygwin current stable with #ifdef added for ease of rebuild:

$ uname -srvmo
CYGWIN_NT-10.0-19045 3.4.9-1.x86_64 2023-09-06 11:19 UTC x86_64 Cygwin
$ gcc -UREDEFFE -o feraiseexcept{,.c} && ./feraiseexcept; echo $?
0
$ gcc -DREDEFFE -o feraiseexcept{,.c} && ./feraiseexcept; echo $?
Floating point exception (core dumped)
136

Cygwin no longer supports x86/i686 32 bit.

Anyone have a Cygwin 3.5.0 test release environment available, just in case it 
might be different in some way, and could confirm?

The code is in newlib, so redirecting there:

https://sourceware.org/cgit/newlib-cygwin/tree/newlib/libm/machine/shared_x86/fenv.c#n253

That only has similar common code as in glibc, not the specific exceptions.

Please, could anyone test on a non-Cygwin newlib x86/_64 build?

The BSDs store, add exceptions, then load first the X87 then SSE MXCSR environments:

https://cgit.freebsd.org/src/tree/lib/msun/amd64/fenv.c#n32

https://github.com/NetBSD/src/blob/trunk/lib/libm/arch/x86_64/fenv.c#L195

https://github.com/openbsd/src/blob/master/lib/libm/arch/amd64/fenv.c#L116

which is handled in newlib fenv.c functions by fgetenv/fsetenv *except* in 
feraiseexcept!

-- 
Take care. Thanks, Brian Inglis              Calgary, Alberta, Canada

La perfection est atteinte                   Perfection is achieved
non pas lorsqu'il n'y a plus rien à ajouter  not when there is no more to add
mais lorsqu'il n'y a plus rien à retirer     but when there is no more to cut
                                 -- Antoine de Saint-Exupéry

Re: feraiseexcept does not raise traps in Cygwin

[Corinna Vinschen]

Hi Bruno,

On Oct 25 11:21, Bruno Haible via Cygwin wrote:
> Hi,
> 
> Seen on
>   - Cygwin 3.4.6 or 2.9.0 on x86_64.
>   - Cygwin 2.9.0 on i386.
> 
> According to ISO C 23 § 7.6.4.3
>   "The feraiseexcept function attempts to raise the supported floating-point
>    exceptions represented by its argument. 266)
>    Footnote 266) The effect is intended to be similar to that of floating-point
>    exceptions raised by arithmetic operations. Hence, implementation extensions
>    associated with raising a floating-point exception (for example, enabled
>    traps or IEC 60559 alternate exception handling) should be honored."
> 
> This does not work. How to reproduce:
> [...]

Thanks for the report and the simple testcase.  I checked in a patch.
It will be in the test build cygwin-3.5.0-0.450.g2ef627c46eb2 in an
hour or two, if you want to give it a try.


Thanks again,
Corinna