Re: glibc 2.1.97

[Christian Iseli <chris at ludwig-alpha dot unil dot ch>]

jakub@redhat.com said:
> printf_fphex is only miscompiled with -mcpu=ev4/-mcpu=ev5 AFAIK. I'll
> try to look into it soon (unless Richard beats me to do that). 

I'm pretty sure I had it miscompiled with -mcpu=ev56 too.

> Please don't. Errors like 156 ulps should not be in libm-test-ulps, it
> is a bug to compute it with such errors, not something that should be
> tolerated.

Ok.  I had no idea what the upper limit should be, and (probably wrongly) 
guessed it would vary depending on the kind of function evaluated.  Last time 
I asked on the list, I was told that the errors in math test on alpha were 
expected, and I (again, probably wrongly) assumed they were "normal" and could 
be dealt with by adjusting the expected results...

> Please make sure you have Richard's sqrt fixes in the
> compiler (or compile e_sqrt.c with -O1). 

Did the fix get included in GCC CVS yet ?

> (with yesterday's binutils patches and alpha dl-machine.h patch posted
> today). 

Bleeding edge huh ? ;-)  You mean HJ released another binutils, or do you use 
the binutils CVS from sourceware ?

Cheers,
					Christian