https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100492
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Known to fail| |10.3.1, 11.1.1, 12.0
Known to work| |9.3.1
--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> ---
We manage to do one more CSE of 'a' which in turn triggers loop distribution
which, when I disable it, fixes the testcase. The disambiguation in question
only triggers in FRE{1,3,5}, DOM2, PRE and CUNROLL. That points to a
latent issue in loop distribution.
The following modified testcase also is miscompiled on the GCC 10 and 11
branch,
-O2 -ftree-loop-distribution is enough to trigger that.
int printf (const char *, ...);
char a, c;
int b, d, *e = &d, g;
static char f;
int main() {
int h = 0;
int a_ = a;
for (; a_ < 1; a = ++a_) {
int *i[5], **j = &i[4], ***k[3][2] = {{&j}}, ****l = &k[2][1], *****m = &l;
char *n = &c;
f = *e = g = 0;
for (; g < 2; g++) {
for (b = 0; b < 3; b++)
h = (h && (*n = 0)) == 0;
if (g)
break;
}
}
printf("%d\n", f);
return 0;
}
Specifically we distribute f = 0 from the rest.
What is odd is that we have a loop with multiple latches at the start of loop
distribution.
in .optimized we have
<bb 9> [local count: 955630225]:
goto <bb 9>; [100.00%]
which results from the tail partition which ends up "canonicalizing" not
included GIMPLE_CONDs in a way that forms the endless loop.
