AMD backend failure:instruction offset out of range

After moving from ftn95 version 8.61 to 8.62 I observed an AMD backend failure if compiling the following source named backup_failure.for:

        SUBROUTINE GETINW (TXART,IW,RW,TW) 
        !+ OK for ftn95 8.61 , error for ftn95 8.62: AMD backend failure:Internal error - instruction offset out of range
        INCLUDE 'datlizp.for'
        INCLUDE 'p0918in0.for'
        INCLUDE 'p0918in1.for'
        CHARACTER
     *  TXART*(*),TW*(*)
        INTEGER*4 IW
        REAL*4 RW
        REAL*8 DW
C       DW=1.0 !+ ok
        DW=ZW(LKP) !+ bad

        RETURN
        END

The exact error message when using ftn95 8.62 is

*** AMD backend failure:Internal error - instruction offset out of range

the command for building is

ftn95 backup_failure.for /CFPP /DEFINE SALFORD64 1 /ALT_KINDS /FIXED_FORMAT /64

The compile succeeds and creates the object file expected if using ftn95 8.61 for compiling.

Moreover if I substitute

      DW=ZW(LKP)

by

      DW=1.0

in file backup_failure.for, then the compile succeeds for both version ftn95 8.61 and 8.62.

Unfortunately I cannot yet reproduce the error without using the 3 include files above the sizes of which are not so small.

I will try to reduce the size of the include files to create a 'real' reproducer.

This problem is occuring for many files which have successfully been compiled when using ftn95 8.61.

Regards Dietmar

The following link will probably provide a fix.

https://www.dropbox.com/s/wjxroa49dz8cxxx/ftn95_8_62_1.zip?dl=0

Paul,

thanks for the link containing version 8.62.1 of ftn95.

It solved the problem addressed above. So I do not have to send you my reproducer (I have been able to strip the include files down to an acceptable size).

Regards Dietmar

There is little value in updating to this latest patch unless you encounter the same failure. It only fixes the one issue.