https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78327
Andrew Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Summary|-Walloca-large-than false |Improve VRP for ranges of
|positives due to bad range |signed char
|info for signed integers in |
|[-TYPE_MAX + N, N] |
Severity|normal |enhancement
--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
The problem is related to anything described here.
The problem is VRP is not adding an assert in the true case of for n_9 usage:
<bb 2>:
n.0_1 = (unsigned char) n_9(D);
_2 = n.0_1 + 125;
_3 = (signed char) _2;
if (_3 < 0)
goto <bb 4>;
else
goto <bb 3>;
<bb 3>:
_15 = (int) n_9(D);
_17 = _15 + 128;
_19 = (long unsigned int) _17;
<bb 4>:
# prephitmp_20 = PHI <_19(3), 128(2)>
------ CUT -----
Basically when we look at the if statement of if (_3 < 0) we don't recognize it
is doing ((signed char)(((unsigned char) n_9) + 125)) < 0.
This should be an easy thing to implement in VRP.
Here is a simple testcase without warnings.
void link_failure ();
void g (signed char n)
{
if (n < -125 || 2 < n) n = 0;
if ((n + 128) > 130)
link_failure();
}
void h (signed char n)
{
if (n < -125 || 1 < n) n = 0;
if ((n + 128) > 129)
link_failure();
}
void g1 (int n)
{
if (n < -125 || 2 < n) n = 0;
if ((n + 128) > 130)
link_failure();
}
void h1 (int n)
{
if (n < -125 || 1 < n) n = 0;
if ((n + 128) > 129)
link_failure();
}
Again this has nothing to with anti-range as there is no ranges being produced
here by VRP for g.
