4.3.1 test failures on alphaev56-dec-osf4.0g (Tru64)

[Torbjorn Granlund]

"Daniel Richard G." <skunk at iSKUNK.ORG> writes:

  For my part, I would be inclined to think it shouldn't... I interpret
  --host=none as "generic Unix", which would imply no assembly code. Is
  there a reason why one would want to use assembly in one place, as
  opposed to all or none?
  
I would suspect people use --host=none to work around some problems with
the assembly code.  Such problems are not uncommon, since people keep
inventing new relocation limitations.

I don't see such problems with inline assembly.  Killing inline assembly
will make GMP perform dreadfully.

  >   I'd be happy to provide any further information on this, as well as
  >   test potential fixes.
  >
  > There will be no potential fixes for the FAILures before somebody with
  > access debugs it.  How about you?
  
  Debugging Alpha assembly code is a bit beyond my ken :>
  
  I tried building with CFLAGS="-O0 -g2 -DNO_ASM" and --host=none, with
  both cc(1) and GCC---but everything passes, so I can't get at the
  problem in straight C.
  
I suspect the assembler might be the culprit.  If I understood your
previous messages correctly, the crash happens in invert_limb.asm, which
has a couple of relocations.  Either the m4 used is buggy (that is
surprisingly common!) or the assembler is buggy or has new relocations
limitations.  (I tried a gcc 4.4 cross compiler tergeted for
alpha-dec-osf5, and it generates relocs identical to the ones I use in
invert_limb.asm.  Also, I am 100% sure the file used to work with OSF in
the past, with both gcc and cc.)

  If you'd like to try some "remote debugging via e-mail," I'd be happy to
  oblige.

Thanks for the offer, but I am afraid that would be poor usage of the
time I have for developing GMP.

-- 
Torbjörn