https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83385
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
Nothing obvious (well, the .optimized result is obviously broken). ISL
transforms
[scheduler] original ast:
{
for (int c0 = 0; c0 <= 1; c0 += 1) {
S_9(c0);
for (int c1 = 0; c1 <= 1; c1 += 1) {
S_3(c0, c1);
S_5(c0, c1);
for (int c2 = 0; c2 <= 1; c2 += 1)
S_6(c0, c1, c2);
S_20(c0, c1);
}
S_8(c0);
S_21(c0);
}
S_10();
for (int c1 = 0; c1 <= 1; c1 += 1)
S_4(0, c1);
}
[scheduler] AST generated by isl:
{
S_9(0);
S_3(0, 0);
S_5(0, 0);
S_20(0, 0);
for (int c0 = 1; c0 <= 2; c0 += 1) {
if (c0 == 2)
S_9(1);
S_3(c0 - 1, -c0 + 2);
S_5(c0 - 1, -c0 + 2);
S_20(c0 - 1, -c0 + 2);
if (c0 == 1) {
S_8(0);
S_21(0);
}
}
S_3(1, 1);
S_5(1, 1);
S_20(1, 1);
S_8(1);
S_21(1);
for (int c0 = 0; c0 <= 1; c0 += 1)
S_4(0, c0);
S_10();
for (int c0 = 0; c0 <= 1; c0 += 1)
for (int c1 = 0; c1 <= 3; c1 += 1)
S_6(c1 / 2, c1 % 2, c0);
}
what is odd is the last loop in the original AST - that belongs to the
if (aj != 0)
--n1;
stmt and hints at a bogus initial schedule generation (eventually again
confusing the order we visit BBs in...). Indeed - scop->pbbs has that
block ordered last :/
testing fix.
