Re: [Qemu-devel] [PATCH 1/3] Add support for 128-bit arithmeticRe: [PATCH 1/3] Add support for 128-bit arithmetic

[Andreas Färber]

Am 10.02.2012 10:51, schrieb Jay Foad:
> On 30 Oct 2011, Avi Kivity wrote:
>> The memory API supports 64-bit buses (e.g. PCI).  A size on such a bus cannot
>> be represented with a 64-bit data type, if both 0 and the entire address
>> space size are to be represented.  Futhermore, any address arithemetic may
>> overflow and return unexpected results.
>>
>> Introduce a 128-bit signed integer type for use in such cases.  Addition,
>> subtraction, and comparison are the only operations supported.
>>
>> Signed-off-by: Avi Kivity <address@hidden>
>> ---
>>  int128.h |  116 
>> ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
>>  1 files changed, 116 insertions(+), 0 deletions(-)
>>  create mode 100644 int128.h
>>
>> diff --git a/int128.h b/int128.h
>> new file mode 100644
>> index 0000000..b3864b6
>> --- /dev/null
>> +++ b/int128.h
>> @@ -0,0 +1,116 @@
>> +#ifndef INT128_H
>> +#define INT128_H
>> +
>> +typedef struct Int128 Int128;
>> +
>> +struct Int128 {
>> +    uint64_t lo;
>> +    int64_t hi;
>> +};
> 
>> +static inline Int128 int128_add(Int128 a, Int128 b)
>> +{
>> +    Int128 r = { a.lo + b.lo, a.hi + b.hi };
>> +    r.hi += (r.lo < a.lo) || (r.lo < b.lo);
>> +    return r;
>> +}
> 
> This is a bit redundant. You only need either:
> 
>    r.hi += r.lo < a.lo;
> 
> or:
> 
>    r.hi += r.lo < b.lo;
> 
> because the way that two's complement addition works means that r.lo
> will always be less than both a.lo and b.lo, or
> greater-than-or-equal-to both of them.
> 
>> +static inline bool int128_ge(Int128 a, Int128 b)
>> +{
>> +    return int128_nonneg(int128_sub(a, b));
>> +}
> 
> This is wrong if you get signed overflow in int128_sub(a, b).
> 
>> Regardless, the need for careful coding means subtle bugs,
> 
> Indeed :-)

Are these just theoretical issues or does anything in particular break
for you? These functions were introduced to tackle int64_t overflow
issues in the Memory API, not as an arbitrary API AFAIU.

If nothing's broken in practice it would be best if you could just send
a patch with your proposed fix rather than describing it in words. :)

Andreas

-- 
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany
GF: Jeff Hawn, Jennifer Guild, Felix Imendörffer; HRB 16746 AG Nürnberg