Re: [Qemu-devel] [PATCH 14/17] s390x: Implement opcode helpers

[Alexander Graf <agraf@suse.de>]

Am 24.03.2011 um 18:41 schrieb Richard Henderson <rth@twiddle.net>:

> On 03/24/2011 10:29 AM, Peter Maydell wrote:
>> On 24 March 2011 15:58, Alexander Graf <agraf@suse.de> wrote:
>> 
>> This is more random comments in passing than a thorough review; sorry.
>> 
>>> +#if HOST_LONG_BITS == 64 && defined(__GNUC__)
>>> +        /* assuming 64-bit hosts have __uint128_t */
>>> +        __uint128_t dividend = (((__uint128_t)env->regs[r1]) << 64) |
>>> +                               (env->regs[r1+1]);
>>> +        __uint128_t quotient = dividend / divisor;
>>> +        env->regs[r1+1] = quotient;
>>> +        __uint128_t remainder = dividend % divisor;
>>> +        env->regs[r1] = remainder;
>>> +#else
>>> +        /* 32-bit hosts would need special wrapper functionality - just abort if
>>> +           we encounter such a case; it's very unlikely anyways. */
>>> +        cpu_abort(env, "128 -> 64/64 division not implemented\n");
>>> +#endif
>> 
>> ...I'm still using a 32 bit system :-)
> 
> A couple of options:
> 
> (1) Steal code from gcc's __[u]divdi3 for implementing double-word division via
>    single-word division.  In this case, your "single-word" will be long long.
> 
> (2) Implement a simple bit reduction loop.  This is probably easiest.
> 
> (3) Reuse some of the softfloat code that manipulates 128bit quantities.  This
>    is probably the best option, particularly if the availability of __uint128
>    is taught to softfloat so that it doesn't always open-code stuff that the
>    compiler could take care of.

In all applications I've run so far this abort has _never_ fired. I'm not sure gcc even emits it.

So IMHO this abort doesn't hurt. Once we get a bug report of a user hitting it, we can think of ways to implement the missing bits. For now, it's not worse than a not implemented opcode (of which we still have quite a number).


Alex

>