Re: AGHI instruction sets wrong condition code

[Ilya Leoshkevich]

On Wed, 2023-02-01 at 13:37 +0100, Thomas Huth wrote:
> 
>   Hi Richard,
> 
> while looking at an issue that Michael reported in IRC today 
> (https://bugs.debian.org/987110) I noticed that the AGHI instructions
> behaves differently in TCG compared to a KVM guest:
> 
> With the KVM guest:
> 
> (gdb) p/x $r14
> $1 = 0x3ffffffaa48
> (gdb) p/x $pswm
> $2 = 0x705000180000000
> (gdb) x/1i $pc
> => 0x3fff7f99d6e <dl_main+22>:  aghi    %r14,-32
> (gdb) stepi
> 0x000003fff7f99d72      1357    in rtld.c
> (gdb) p/x $r14
> $3 = 0x3ffffffaa28
> (gdb) p/x $pswm
> $4 = 0x705200180000000
> 
> With the TCG guest:
> 
> (gdb) p/x $r14
> $1 = 0x3ffffffaa48
> (gdb) p/x $pswm
> $2 = 0x705000180000000
> (gdb) x/1i $pc
> => 0x3fff7f99d6e <dl_main+22>:  aghi    %r14,-32
> (gdb) stepi
> 0x000003fff7f99d72      1357    in rtld.c
> (gdb) p/x $r14
> $3 = 0x3ffffffaa28
> (gdb) p/x $pswm
> $4 = 0x705000180000000
> 
> As you can see, the PSW mask is different after the AGHI instruction,
> real 
> hardware sets CC to 2, while TCG keeps it a 0.
> 
> I stared at the translate.c and cc_helper.c code for a while, but I
> can't 
> quite figure it out ... if you've got some spare minutes, could you
> maybe 
> have a look at this, too?
> 
>   Thanks,
>    Thomas
> 
> 
> PS: I guess it's not the bug causing the issue reported by Michael,
> since 
> there is no instruction looking at the CC right after that AGHI

Interestingly enough, the following works fine with linux-user:

#include <assert.h>
#include <stdlib.h>

int main(void) {
    long val = 0x3ffffffaa48;
    int cc;

    asm("aghi %[val],-32\n"
        "ipm %[cc]" : [val] "+r" (val), [cc] "=r" (cc) : : "cc");
    cc >>= 28;
    assert(val == 0x3ffffffaa28);
    assert(cc == 2);

    return EXIT_SUCCESS;
}

I wonder if this could be a PER/singlestep issue?