Re: [PATCH 01/16] math128: Introduce various 128bit primitives

[Linus Torvalds <torvalds@linux-foundation.org>]

On Wed, Oct 24, 2012 at 2:53 PM, Juri Lelli <juri.lelli@gmail.com> wrote:
> From: Peter Zijlstra <a.p.zijlstra@chello.nl>
>
> Grow rudimentary u128 support without relying on gcc/libgcc.

I missed the part where somebody explains why and what needs this?
It's going to be very expensive indeed on some platforms, so the fact
that it is *sometimes* cheap doesn't necessarily imply it should ever
be used.

So please, explain what the pressing need is that is so worthwhile
that this is worth it. Maybe it was in a 00/16 cover letter, but not
only was that not sent out to the people who got 01, you'd still want
it in the commit message.

> +typedef union {
> +       struct {
> +#if __BYTE_ORDER__  == __ORDER_LITTLE_ENDIAN__
> +               u64 lo, hi;
> +#else
> +               u64 hi, lo;
> +#endif
> +       };
> +#ifdef __SIZEOF_INT128__ /* gcc-4.6+ */
> +       unsigned __int128 val;
> +#endif
> +} u128;

This also looks totally wrong.

If gcc has native support for __int128, then the union is pointless.
Don't do it. Just do

  #ifdef __SIZEOF_INT128__
    typedef unsigned __int128 u128;
  #else
    typedef struct { ... u64 hi/lo in the right order } u128;
  #endif

because it's possible that using the native bare type will make gcc
able to do better for various things.

Sure, it's possible that you want to use a union in low-level
architecture code that implements the actual math, BUT EVEN THEN the
above union is pure and utter garbage. On 32-bit machines, you'd want
to make it a union of 4 32-bit entities etc. So putting it like this
in a generic file looks wrong. In fact, your very own generic
mul_u64_u64() would seem to want to use the "4 32-bit words" kind of
model.

Also, the union isn't used for generic code anyway, since the generic
code has that same __SIZEOF_INT128__ test for which generic version it
should include (and I wonder if it should just be

  #ifdef __SIZEOF_INT128__
    #include <linux/native-128bit.h>
  #elif CONFIG_64BIT
    #include <linux/generic64bit-128bit.h>
  #else
    #include <linux/generic64bit-128bit.h>
  #endif

and then have separate files entirely for the "gcc handles the common
operations" vs "64-bit architecture needs two words for most things"
vs "32-bit architectures need 4 words for most things".

I dunno. But I think this is wrong.

             Linus