Re: [Qemu-devel] signed division overflow workaround

[Peter Maydell]

On 12 August 2011 10:44, Max Filippov <address@hidden> wrote:
> Hello.
>
>> +            case 13: /*QUOSi*/
>> +                tcg_gen_div_i32(cpu_R[RRR_R], cpu_R[RRR_S], cpu_R[RRR_T]);
>> +                break;
>
> I'm currently developing test suite for xtensa port and found that
> with this implementation of QUOS (signed 32-bit division) guest that
> divide 0x80000000 by -1 crashes qemu with 'floating point exception'
> on x86_64 host.

Yes. See tcg/README, which says:
# Undefined behavior if division by zero or overflow.

This is basically because the behaviour varies rather from architecture
to architecture so (a) it would be a pain to implement consistent
handling in the tcg backends and (b) all the front-ends would probably
still need special case handling to give the results the guest arch
requires. (What is xtensa's behaviour for div by zero and the overflow
case ?)

> I guess that this is a known issue, at least target-arm has a special
> case to handle this. Is there a better way to handle this issue, or
> special-casing is the only option?

I wouldn't do what target-arm does: it goes out to a helper function,
probably because that code predates the addition of brcond to TCG.
PPC and MIPS are probably better candidates to crib from, as they
both do the special-case checks inline.

-- PMM