Fortran on X86 : **2 on a complex*16 crashes

[joukj]

Hi all

I found this, which is not to be expected:

Code: Select all

rumba-jj) f90/vers
VSI Fortran X8.5-0002 (GEM 50X2N) for X86 systems
rumba-jj) ty test.f90 
program test
  complex*16 x
  x =x**2
end
  13-JUN-2023 13:33:32
rumba-jj) f90 test
rumba-jj) link test
rumba-jj) run test
%SYSTEM-F-ACCVIO, access violation, reason mask=06, virtual address=000000000000
0002, PC=FFFF830008A491FB, PS=0000001B
%TRACE-F-TRACEBACK, symbolic stack dump follows
image     module    routine               line      rel PC           abs PC     
LIBOTS                                       0 00000000800021FB FFFF830008A491FB
DPML$SHR  DPML_CPOWI.C;1                #13088 0000000080072E93 FFFF830008C51E93
TEST  test.f90;33  TEST                      3 0000000000000045 0000000080000045
                                             0 FFFF8300071FBDC6 FFFF8300071FBDC6
DCL                                          0 000000008006741B 000000007ADFF41B
%TRACE-I-LINENUMBER, Leading '#' specifies a source file record number.
%TRACE-I-END, end of TRACE stack dump
rumba-jj)

[arne_v]

That must be a bug. John Reagan!!!!

But I am curious whether it also crashes if you do:

x = 1
x = x **2

(I have no gotten my own VMS x86-64 up and running yet)

[joukj]

Yes it does. Sorry I stripped the sample code too much to get the smallest reproducer.

[arne_v]

OK.

Given your other post then it looks like the math lib on x86-64 has a few "issues".

[jreagan]

"long double" (aka REAL*16, COMPLEX*16, Pascal QUADRUPLE, etc) is currently a work in progress (the release notes should say that).

The cross-compilers used to build the underlying Math routines know only about x86 f80 (10bytes of data in a 10byte container). However, OpenVMS wants IEEE X floating (16bytes of data in a 16byte container). The newer LLVM we use for native compilers knows about all three forms:

- 10bytes of F80 in a 10byte container
- 10bytes of F80 in a 16byte container
- 16bytes of IEEE_X in a 16byte container

It will have to all wait for native builds of the Math RTL (and possibly other RTLs) to get all of the long double working)

[joukj]

I thought complex*16 is just "double precision: (real*8,real*8)

[jreagan]

Yes, my bad. I had confused the KIND= syntax.

Yes, COMPLEX*8 is a pair of REAL*4; COMPLEX*16 is a pair of REAL*8, and COMPLEX*32 is a pair of REAL*16.

COMPLEX(KIND=4) is COMPLEX*8
COMPLEX(KIND=8) is COMPLEX*16
COMPLEX(KIND=16) is COMPLEX*32

DOUBLE COMPLEX is COMPLEX(8) is COMPLEX*16

Confusing