https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97139
--- Comment #2 from Martin Liška <marxin at gcc dot gnu.org> ---
So one can reproduce it with current master with:
-O3 -march=znver2 -g -flto=auto -flto-partition=one
where first bad debug counter value is:
-fdbg-cnt=vect_slp:357
The change is:
before:
_648 = pix.x;
_649 = _648 / 4;
pix.x = _649;
_650 = pix.y;
_651 = _650 / 4;
pix.y = _651;
_652 = pix.pos_x;
_653 = _652 / 4;
pix.pos_x = _653;
_654 = pix.pos_y;
_655 = _654 / 4;
pix.pos_y = _655;
...
after:
vect__648.11467_1941 = MEM <vector(4) int> [(int *)&pix + 8B];
_1946 = vect__648.11467_1941 < { 0, 0, 0, 0 };
vect_patt_180.11468_1944 = VEC_COND_EXPR <_1946, { 3, 3, 3, 3 }, { 0, 0, 0, 0
}>;
vect_patt_182.11469_1947 = vect__648.11467_1941 + vect_patt_180.11468_1944;
vect_patt_185.11470_1966 = vect_patt_182.11469_1947 >> 2;
_1967 = BIT_FIELD_REF <vect_patt_185.11470_1966, 32, 0>;
_1968 = BIT_FIELD_REF <vect_patt_185.11470_1966, 32, 32>;
_1950 = BIT_FIELD_REF <vect_patt_182.11469_1947, 32, 0>;
_1965 = BIT_FIELD_REF <vect_patt_182.11469_1947, 32, 32>;
_1943 = BIT_FIELD_REF <vect_patt_180.11468_1944, 32, 0>;
_1942 = BIT_FIELD_REF <vect_patt_180.11468_1944, 32, 32>;
_1945 = &pix.x + 4;
_648 = pix.x;
_649 = _648 / 4;
_650 = pix.y;
_651 = _650 / 4;
_652 = pix.pos_x;
_653 = _652 / 4;
_654 = pix.pos_y;
_655 = _654 / 4;
MEM <vector(4) int> [(int *)&pix + 8B] = vect_patt_185.11470_1966;
...
So a vector shift by 2 instead of division by 4. It looks fine (assuming the
trick for negative numbers is correct).
@Richi: can you please dig into it?
