https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90913
--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
The ICE is because .MASK_LOAD/.MASK_STORE calls are supported just for vectors.
The way it is done is that during ifcvt those internal calls are added to the
.LOOP_VECTORIZED guarded loops, and either we successfully vectorize those,
then the .MASK_LOAD/.MASK_STORE ifn calls are replaced with their vectorized
versions and all the checking is performed, or we should DCE that loop.
The r272239 change which I haven't really understood breaks this on this
testcase, we end up with
_116 = .LOOP_VECTORIZED (1, 4);
if (_116 != 0)
goto <bb 25>; [100.00%]
else
goto <bb 26>; [100.00%]
that guards these scalar .MASK_STORE calls being replaced with
_116 = 0;
if (_190 != 0)
goto <bb 25>; [100.00%]
else
goto <bb 26>; [100.00%]
that is just wrong, we can't and shouldn't reuse the for vectorization only
loop if we don't vectorize it. Both because of these .MASK* calls, but also
because e.g. there are COND_EXPRs all around in that which often isn't
beneficial in scalar loops, etc.
