Re: [Qemu-devel] [RFC] alpha qemu arithmetic exceptions

[Al Viro <viro@ZenIV.linux.org.uk>]

On Wed, Jul 02, 2014 at 05:05:08AM +0100, Al Viro wrote:
> OK, DS10 resurrected and so far seems to be stable (I'll know by tomorrow;
> there's a possibility that chipset heatsink is dodgy, but so far it seems
> to be doing OK).  That gives us a EV6 box.
> 
> Which glibc version it is?  I don't see such failures with your
> axp/axp-next (head at 6b38f4e7f); could you post the details on your
> reproducer?  I've tried to guess the likely version by glibc.git, but
> I don't see nearbyint tests with such argument in any version there,
> so I couldn't find it that way...

FWIW,

; cat >a.c <<'EOF'
#include <stdio.h>
#include <fenv.h>

volatile long x;
void __attribute__((noinline)) f(double v)
{
        x = v;
}

main()
{
        unsigned long tmp, ret;
        static char *names[] = {"IOV", "INE", "UNF", "OVF", "DZE", "INV"};
        int i;

        feclearexcept(FE_ALL_EXCEPT);
        f(4.5);

        __asm__ __volatile__ (
                "stt $f0,%0\n\t"
                "trapb\n\t"
                "mf_fpcr $f0\n\t"
                "trapb\n\t"
                "stt $f0,%1\n\t"
                "ldt $f0,%0"
                : "=m"(tmp), "=m"(ret));
        for (i = 0; i < 6; i++)
                printf(" %s", (ret >> (57-i)) & 1 ? names[i] : "   ");
        printf(" %x ", fetestexcept(FE_ALL_EXCEPT));
        printf("FE_INEXACT = %x\n", FE_INEXACT);
}
EOF
; gcc -lm a.c
; ./a.out
     INE                 200000 FE_INEXACT = 200000
; uname -a
Linux wynton 2.6.22-rc7 #1 Thu Aug 30 02:03:17 EDT 2007 alpha GNU/Linux

That's on freshly resurrected DS10, just brought to the last debian/alpha
(i.e. lenny).  Kernel had been locally built back before the box has died.

On miles (3.3.6+, AS200) result is different:
     INE                 0 FE_INEXACT = 200000

On qemu (with debian kernel from lenny - 2.6.26) it's the same as on DS10:
     INE                 200000 FE_INEXACT = 200000

It _might_ be the difference between 3.3 and 2.6.20-somethine, but I doubt
that.  It's definitely not a matter of difference in libc versions - AS200
box has 2.13-38, but static binary built on DS10 (with its 2.7-18) copied on
AS200 behaves there as locally built one (i.e. 0 from fetestexcept(), as
opposed to FE_INEXACT the same static binary produces on DS10 and under
qemu).

AFAICS, it leaves two possibilities - EV45 (AS200) vs. EV6 (DS10) and EV67
(qemu) _or_ some change in the kernel.  I'll build 3.x kernel for DS10 and
post the results; shouldn't take long...