[bananas] Serious problem with LINUX AIPS Binaries

Patrick P. Murphy pmurphy at cv.nrao.edu
Wed Dec 23 13:43:07 EST 1998


Summary:

  The AIPS group at NRAO has uncovered a serious problem with the LINUX
  binaries that have been pre-built and made available via anonymous ftp
  and on CDROM for the 15OCT98 version of AIPS.  New binaries and
  libraries are now available on our ftp server, effective 9:30am US/EST
  (14:30 UT) today, December 23, 1998.  The old binaries were made
  unavailable at 9:00am.

Description:

  The problem stems from our use of the EGCS version 1.1 g77 compiler.  In
  a couple of key subroutines that perform byte swapping (ZRLR32 and
  ZR32RL), the use of INTEGER*2 entities appears to cause random, non-
  deterministic changes in the data.  The data written to tape or disk by
  FITTP are corrupted in such a way that neither FITTP, FITLD or PRTTP
  detect any error. The data cannot be repaired.

Are you affected:

  The problem has been found only in the EGCS 1.1 and 1.1.1 compilers; it
  appears to be absent if the binaries are built with EGCS 1.0.2 or 1.0.3.
  You are affected if:
     - You built AIPS 15OCT98 with the EGCS version 1.1 or 1.1.1 of g77 on
       LINUX. 
     - You installed LINUX binaries from any 15OCT98 CDROM from NRAO with
       a sequence number between 1 and 40;
     - You retrieved the 15OCT98 LINUX binaries from NRAO's ftp site prior
       to 9:30am (US/EST; 14:30 UT) on December 23, 1998.
  The Linux AIPS release of 15OCT98 generated by NRAO after September 1998
  is the only known version with this problem.

  The version of an egcs compiler can be determined by the following:
	g77 -v | grep egcs | grep release | head -1
  This will clearly show what version you have.  If there is no output,
  you are not using an EGCS compiler.

  Any 15OCT98 CDROM with a (handwritten) sequence number between 1 and 40
  should be either destroyed (contact us for a free replacement; see
  below) or only used for source and/or Solaris binaries.

What we are doing:

  We have already rebuilt the LINUX binaries and libraries for 15OCT98
  with the 1.0.2 version of EGCS g77/gcc, and these are now available on
  our ftp server.  We are cutting a re-issue of the CDROM; sequence
  numbers on this release will start with #100.  

Remedies:

  If you are affected, you have several options:
     - Install an older version of the EGCS compilers (1.0.2 or 1.0.3),
       delete all the 15OCT98 AIPS libraries in $LIBR/*/SUBLIB, and redo 
       INSTEP2 and INSTEP4; or
     - Request a new CDROM (free) from aipsmail at nrao.edu (we will try to
       contact you anyway, but if we miss you, contact us); or
     - Download the new binaries and libraries from our ftp site.

  Anyone who already received a 15OCT98 CDROM from NRAO should receive a
  new one within the next month or so.  

  The new binaries are in /aips/15OCT98/LINUX/LOAD/*.EXE.gz and the
  libraries are in /aips/15OCT98/LINUX/LIBR/*/SUBLIB on our ftp server
  ftp.cv.nrao.edu.  If you get a new CDROM and you are running in "copied
  to disk" mode, simply re-run the CDSETUP script to revert to the mode of
  running from CDROM, then back again.

"Midnight Job" impact:

  All sites on the midnight job using the Linux egcs compiler version 1.1
  or higher will have had this problem since they upgraded their egcs
  compiler.  Any sites on the 15APR99 midnight job (including all NRAO
  sites) do not need to do anything; the problem is in the process of
  being corrected for this version (the routines will be modified so as to
  not need INTEGER*2).

Our apologies for not catching this problem sooner.  May your Christmas be
more bug-free than our recent experiences!

				- Tony Beasley
				  Eric Greisen
				  Pat Murphy




More information about the Bananas mailing list