https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89821
kargl at gcc dot gnu.org changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |kargl at gcc dot gnu.org
--- Comment #2 from kargl at gcc dot gnu.org ---
(In reply to Richard Biener from comment #1)
> The invalid exception is raised here:
>
> Program received signal SIGFPE, Arithmetic exception.
> 0x0000000000400cde in demo_nan () at t.f90:17
> 17 if ( (r32<=0.0_real32) .or. (r32>=0.0_real32) )then
>
> 0x0000000000400cd2 <+416>: movss -0x134(%rbp),%xmm1
> 0x0000000000400cda <+424>: pxor %xmm0,%xmm0
> => 0x0000000000400cde <+428>: comiss %xmm1,%xmm0
>
> where %xmm1 is NaN.
>
> (gdb) p $xmm1
> $2 = ( v4_float = (nan(0x400000), 0, 0, 0),
>
> so not sure what you are expecting? Is Fortran supposed to use the
> C equivalent of isgreaterequal (aka comparisons that do not raise
> exceptions?)
The Fortran standard says nothing about the option -ffpe-trap=invalid.
If one uses an option requesting a trap, then one might anticipate
that it cause a SIGFPE when an NaN is used in an expression.
This reduced testcase fails with 8-branch and trunk when compiled
with -ffpe-trap=invalid.
program demo_nan
use,intrinsic :: iso_fortran_env, only: real32
implicit none
character(len=3),save :: STRING='NaN'
real(kind=real32) :: r32
character(len=256) :: message
integer :: ios
read(STRING,*)r32
! an option to terminate a program when a NaN is encountered
! (if X is NaN the comparison with 0. is always false.)
if ( (r32<=0.0_real32) .or. (r32>=0.0_real32) )then
write(*,*)'did not produce a nan_real32'
else
write(*,*)'found a nan_real32. Handle it.'
endif
! list directed format
write(*,*,iomsg=message,iostat=ios)r32
if(ios.ne.0)write(*,*)trim(message)
! hexadecimal format to show different kinds
write(*,'(z0)',iomsg=message,iostat=ios)r32
end program demo_nan
The above compiles with both 6-branch and an old 7-branch
(need to update 7 to test).
There are two ways in which I can avoid the SIGFPE and
still use an NaN. First, one can use the intrinsic
function isnan() to change the conditional to
if (.not.isnan(r32))then
write(*,*)'did not produce a nan_real32'
else
write(*,*)'found a nan_real32. Handle it.'
endif
The second, and likely preferred method, is to use the IEEE_ARITHMETIC
module that is provided by the Fortran Standard.
