https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118154
--- Comment #5 from Robin Dapp <rdapp at gcc dot gnu.org> ---
Confirmed, funnily only happens with a QEMU VLEN=128 and not with VLEN >= 256.
-fwrapv and -fno-strict-aliasing are not necessary for me.
Another "funny" thing:
vect__5.15_44 = .COND_LEN_MAX ({ -1, ... }, _62, { 17, ... }, _62, 6, 0);
vect__5.15_89 = .COND_LEN_MAX ({ -1, ... }, vect__5.15_44, { 17, ... },
vect__5.15_44, 6, 0);
vect__5.15_56 = .COND_LEN_MAX ({ -1, ... }, vect__5.15_89, { 17, ... },
vect__5.15_89, 6, 0);
vect__5.15_25 = .COND_LEN_MAX ({ -1, ... }, vect__5.15_56, { 17, ... },
vect__5.15_56, 6, 0);
vect__5.15_22 = .COND_LEN_MAX ({ -1, ... }, vect__5.15_25, { 17, ... },
vect__5.15_25, 6, 0);
vect__5.15_72 = .COND_LEN_MAX ({ -1, ... }, vect__5.15_22, { 17, ... },
vect__5.15_22, 6, 0);
vect__5.15_115 = .COND_LEN_MAX ({ -1, ... }, vect__5.15_72, { 17, ... },
vect__5.15_72, 6, 0);
vect__5.15_120 = .COND_LEN_MAX ({ -1, ... }, vect__5.15_115, { 17, ... },
vect__5.15_115, 6, 0);
vect__5.15_125 = .COND_LEN_MAX ({ -1, ... }, vect__5.15_120, { 17, ... },
vect__5.15_120, 6, 0);
vect__5.15_130 = .COND_LEN_MAX ({ -1, ... }, vect__5.15_125, { 17, ... },
vect__5.15_125, 6, 0);
vect__5.15_135 = .COND_LEN_MAX ({ -1, ... }, vect__5.15_130, { 17, ... },
vect__5.15_130, 6, 0);
vect__5.15_140 = .COND_LEN_MAX ({ -1, ... }, vect__5.15_135, { 17, ... },
vect__5.15_135, 6, 0);
vect__5.15_145 = .COND_LEN_MAX ({ -1, ... }, vect__5.15_140, { 17, ... },
vect__5.15_140, 6, 0);
vect__5.15_150 = .COND_LEN_MAX ({ -1, ... }, vect__5.15_145, { 17, ... },
vect__5.15_145, 6, 0);
vect__5.15_155 = .COND_LEN_MAX ({ -1, ... }, vect__5.15_150, { 17, ... },
vect__5.15_150, 6, 0);
vect__5.15_160 = .COND_LEN_MAX ({ -1, ... }, vect__5.15_155, { 17, ... },
vect__5.15_155, 6, 0);
vect__5.15_66 = .COND_LEN_MAX ({ -1, ... }, vect__5.15_160, { 17, ... },
vect__5.15_160, 6, 0);
_69 = .REDUC_MAX (vect__5.15_66);
That's not the reason for the bug but can't be what we want either :) Will
need to have a look at that separately.
As for the main issue: Very likely strided stores and a backend problem. If I
disable the pattern in our backend the test passes. Going to have a deeper
look.
