https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128556
Drea Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Last reconfirmed| |2026-11-08
Ever confirmed|0 |1
--- Comment #1 from Drea Pinski <pinskia at gcc dot gnu.org> ---
_1 = acc_5(D) + _2;
_7 = _1 >> 1;
is being turned into:
_7 = .AVG_FLOOR (_2, acc_5(D));
By:
```
/* Fold (x + y) >> 1 into IFN_AVG_FLOOR (x, y) and ((x + y) + 1) >> 1
into IFN_AVG_CEIL (x, y) if x and y are vectors in which the sum
cannot wrap: overflow is undefined, or each element of x and y is
known to have at least one leading zero bit, or two if the type is
signed, since the sum of two nonnegative values can then wrap to a
negative one. The IFN_AVG_CEIL fold comes first because
((x + y) + 1) >> 1 is also (x' + 1) >> 1. */
(simplify
(rshift (plus:s (plus:cs @0 @1) integer_onep) integer_onep)
(if (VECTOR_TYPE_P (type)
&& direct_internal_fn_supported_p (IFN_AVG_CEIL, type,
OPTIMIZE_FOR_BOTH))
(with { int min_clz = TYPE_UNSIGNED (type) ? 1 : 2; }
(if (TYPE_OVERFLOW_UNDEFINED (type)
|| (wi::clz (get_nonzero_bits (@0)) >= min_clz
&& wi::clz (get_nonzero_bits (@1)) >= min_clz))
(IFN_AVG_CEIL @0 @1)))))
(simplify
(rshift (plus:cs @0 @1) integer_onep)
(if (VECTOR_TYPE_P (type)
&& direct_internal_fn_supported_p (IFN_AVG_FLOOR, type,
OPTIMIZE_FOR_BOTH))
(with { int min_clz = TYPE_UNSIGNED (type) ? 1 : 2; }
(if (TYPE_OVERFLOW_UNDEFINED (type)
|| (wi::clz (get_nonzero_bits (@0)) >= min_clz
&& wi::clz (get_nonzero_bits (@1)) >= min_clz))
(IFN_AVG_FLOOR @0 @1)))))
```
Which I think is correct.
The problem is before that.
vaba_s8 is being turned into:
```
_2 = .ABD (x_3(D), y_4(D));
_1 = acc_5(D) + _2;
```
But this assumes there is no overflow for `acc_5(D) + _2` but there is in this
case.
