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

[Torbjorn Granlund]

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

  On Tue, 2010 Sep  7 20:33+0200, Torbjorn Granlund wrote:
  >
  > Please use gdb to see which exact instruction fails.
  
  I thought I'd provided that information already. Should I attach the
  preprocessed .s file?
  
Sorry, I didn't see that before.  It is he first LEA that crashes.

Let's blindly test if moving the float constant into data space helps:

diff -r c173a930ec87 mpn/alpha/default.m4
--- a/mpn/alpha/default.m4	Tue May 25 10:07:38 2010 +0200
+++ b/mpn/alpha/default.m4	Wed Sep 08 23:45:33 2010 +0200
@@ -35,6 +35,7 @@
 dnl  Usage: FLOAT64(label,value)
 define(`FLOAT64',
 m4_assert_numargs(2)
+	DATA
 `	.align	3
 $1:	.t_floating $2')
 

  This is a static-only build, as the later GCC build will be built with a
  static libgcc. (I'm also using --with-pic, but the same failures result
  without that.)
  
I am not sure --with-pic has any effect.

  Interestingly, a shared-only build passes all tests.

Curious.  If one build type causes trouble, it is usually the shared
build!
  
  I just see an invocation of plain "as" in there. If the problem is
  down to the linker, however, this would be the only instance of it
  that I'm aware of.
  
WHich as version is this?  What does "which as" print?
What does "as --version" or somesuch option print?

  Surely this is something simple enough! Would you want to see assembly
  from one of the generic .c files?
  
If the above patch fails, I'll send some C snippets for compilation...

-- 
Torbjörn