On 12/08/2016 05:55 AM, Martin Liška wrote:
With the patch applied, one can distinguish between PROB_VERY_UNLIKELY and
real zero probability:
I tried to see if formatting the expression
e->probability * 100.0 / REG_BR_PROB_BASE
with "%.2f" is guaranteed to output non-zero when e->probability
is non-zero. If I got the values right then the smallest non-zero
e->probability can be as low as 1, and REG_BR_PROB_BASE is 10000.
That evaluates to 0.0099999999999999985 which Glibc formats as
0.01 because (AFAIK) its printf rounds to nearest.
If this is guaranteed then I think it's fine. Otherwise, if there
is a chance that the printed result could be 0.00% for a non-zero
probability it might be worth to detect it and make sure it's at
least 0.01% to avoid the same confusion I had with the 0.00%.
FWIW, if GCC uses integers rather than floats internally to make
decisions (I don't know) then also printing integers would give
the most accurate results. E.g., something like
printf ("%2u.%u%%", e->probability / 100, e->probability % 100)
Thanks
Martin
f ()
{
int _1;
<bb 2> [100.00%]:
_1 = __builtin_sprintf (&d2, "%i", 12);
if (_1 != 2)
goto <bb 3>; [0.04%]
else
goto <bb 4>; [99.96%]
<bb 3> [0.04%]:
__builtin_abort ();
<bb 4> [99.96%]:
return;
}
Patch can bootstrap on ppc64le-redhat-linux and survives regression tests.
Ready to be installed?
Martin
