[9fans] vlong double-spill botch

[9fans] vlong double-spill botch

[Fernan Bolando]

what does vlong double-spill botch mean in
r = (a * b) + (((a * u) + (b * t)) << 18);          /* low is only 35b */

a larger section of the code is shown below


t_uint64 a = ABS (s1);
t_uint64 b = ABS (s2);
t_uint64 t, u, r;

if ((a == 0) || (b == 0)) {                             /* operand = 0? */
    rs[0] = rs[1] = 0;                                  /* result 0 */
    return;
    }
if ((a & FIT32) || (b & FIT32)) {                       /* fit in 64b? */
    t = a >> 18;                                        /* no, split in half */
    a = a & RMASK;                                      /* "dp" multiply */
    u = b >> 18;
    b = b & RMASK;
    r = (a * b) + (((a * u) + (b * t)) << 18);          /* low is only 35b */
    rs[0] = ((t * u) << 1) + (r >> 35);                 /* so lsh hi 1 */
    rs[1] = r & MMASK;
    }

Re: [9fans] vlong double-spill botch

[erik quanstrom]

> what does vlong double-spill botch mean in
> r = (a * b) + (((a * u) + (b * t)) << 18);          /* low is only 35b */
>

this is a workaround that i put in the compiler;
it's not in the distribution.  the distribution
compiler happily miscompiles.

if both the left and rhs side of an expression
need to allocate more registers, then 8c often
miscompiles.  you may be able to fix this by
changing 18 to 18ull, but you might as well
do it this way:

	r = a*u + b*t;
	r <<= 18ull;
	r += a*b;

http://9fans.net/archive/2009/09/436
http://9fans.net/archive/2009/08/1044
http://9fans.net/archive/2009/08/1042

- erik

Re: [9fans] vlong double-spill botch

[Fernan Bolando]

On Mon, Jan 3, 2011 at 10:56 PM, erik quanstrom
<quanstro@labs.coraid.com> wrote:
>> what does vlong double-spill botch mean in
>> r = (a * b) + (((a * u) + (b * t)) << 18);          /* low is only 35b */
>>
>
> this is a workaround that i put in the compiler;
> it's not in the distribution.  the distribution
> compiler happily miscompiles.
>
> if both the left and rhs side of an expression
> need to allocate more registers, then 8c often
> miscompiles.  you may be able to fix this by
> changing 18 to 18ull, but you might as well
> do it this way:
>
>        r = a*u + b*t;
>        r <<= 18ull;
>        r += a*b;
>
This did not work, but this did
ra = a * b;
rb = a * u;
rc = b * t;
r = ra + ((rb + rc) << 18);