https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121759
Bug ID: 121759
Summary: Windows/MinGW: formatted reads lose decimal digits
after binary stream read
Product: gcc
Version: 15.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: fortran
Assignee: unassigned at gcc dot gnu.org
Reporter: francois.depicciotto at gmail dot com
Target Milestone: ---
Created attachment 62267
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=62267&action=edit
Code to reproduce
There is a reproducible but non-deterministic bug in libgfortran under Windows
(MinGW/MinGW-w64):
after performing a binary read (or write) from a file, subsequent formatted
reads (FORM='FORMATTED') from text files sometimes lose the decimal part of
floating-point numbers.
This issue does not occur under Linux with the same gfortran version.
How to reproduce
****************
Create two files in the working directory:
toto.txt containing exactly:
123.456
toto.bin containing at least 8 zero bytes (dummy binary file)
Program (minimal reproducer):
program toto
call readtxt
call readtxt
call readbin
call readtxt
call readtxt
end program
subroutine readtxt
real :: t
integer :: kf, IPB
t = 0.0
open(newunit=kf, file="toto.txt", form="FORMATTED",
& status="OLD", action="READ", iostat=IPB)
if (IPB == 0) then
read(kf, *) t
print *, "t=", t
endif
close(kf)
end subroutine
subroutine readbin
integer :: io, IPB
real*8 :: x
open(newunit=io, file="toto.bin", access="STREAM", &
form="UNFORMATTED", status="OLD", iostat=IPB)
read(io) x
close(io)
end subroutine
Another code to reproduce is attached (that creates the txt itself, and with
slightly more tests).
Observed behavior
*****************
The first two calls to readtxt always read and display the correct value
(123.456).
After the binary read (readbin), the next calls to readtxt *sometimes* still
read correctly, but randomly lose the decimal part and print 123.000000.
*This happens non-deterministically: multiple executions of the same binary
produce different outcomes ; sometimes 123.456, sometimes 123.000.*
Example (Windows 11, gfortran 15.2.0, MinGW-w64 13.0.0):
t= 123.456001
t= 123.456001
t= 123.000000
t= 123.000000
On Linux with the same compiler version, the issue does not reproduce: the
value is always read correctly.
Expected behavior
*****************
Formatted reads from a text file should always return the correct
floating-point value, regardless of whether a binary read or wirte was
performed earlier.
Notes
*****
Adding DECIMAL='POINT' to the OPEN statement does not fix the issue.
Using explicit formats (e.g. read(kf,'(F15.8)') t) also does not fix it.
Performing only the OPEN/CLOSE of the binary file without an actual READ avoids
triggering the bug.
Reported by multiple users in Fortran-lang:
https://fortran-lang.discourse.group/t/problem-after-reading-binary-file-with-windows-gfortran/7407
