https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122749
Andrew Pinski <pinskia at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Last reconfirmed| |2025-11-20
Ever confirmed|0 |1
Summary|[16 Regression] useless |[16 Regression] signed to
|type conversions inserted |unsigned type conversions
|during vectorization |inserted during
|blocking MLA |vectorization blocking MLA
Component|target |tree-optimization
--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
```
vect_x_19.27_96 = .COND_FMA (loop_mask_87, vect__21.23_89,
vect_vec_iv_.24_91, vect_x_26.26_95, vect_x_26.26_95);
```
vs now:
```
vect__20.26_113 = vect__21.24_109 * vect_vec_iv_.25_111;
vect_x_19.29_116 = VIEW_CONVERT_EXPR<vector([4,4]) unsigned
int>(vect__20.26_113);
vect_x_19.29_117 = VIEW_CONVERT_EXPR<vector([4,4]) unsigned
int>(vect_x_26.27_115);
vect_x_19.29_118 = .COND_ADD (loop_mask_107, vect_x_19.29_116,
vect_x_19.29_117, vect_x_19.29_117);
vect_x_19.28_119 = VIEW_CONVERT_EXPR<vector([4,4]) int>(vect_x_19.29_118);
```
I suspect the code in tree-ssa-math-opts.cc needs to add the support for
optional nop_converts and convert the multiply into unsigned case too. They are
not useless with respect to overflow.
And there is no non-predicated FMA:
```
Trying 39 -> 42:
39: r178:VNx4SI=r174:VNx4SI*r119:VNx4SI
REG_DEAD r174:VNx4SI
42: r127:VNx4SI=r178:VNx4SI+r127:VNx4SI
REG_DEAD r178:VNx4SI
Failed to match this instruction:
(set (reg:VNx4SI 127 [ vect_x_12.16 ])
(plus:VNx4SI (mult:VNx4SI (reg:VNx4SI 174 [ vect__4.12_45 ])
(reg:VNx4SI 119 [ vect_vec_iv_.13 ]))
(reg:VNx4SI 127 [ vect_x_12.16 ])))
```
