https://gcc.gnu.org/bugzilla/show_bug.cgi?id=57955
--- Comment #16 from Eric Botcazou <ebotcazou at gcc dot gnu.org> --- > I tried implementing Jakub's suggestion to have CONSTANT_ALIGNMENT return > 128 for large constructors. This doesn't fix the r201264 version of the > test case, which still generates fairly horrid code. The use of > CONSTANT_ALIGNMENT in align_variable causes the destination to receive the > 128-bit alignment, but the constructor itself remains with 32-bit alignment, > so the aligned-only lvx instruction can't be used. Isn't that because the SRA pass now undoes what the gimplifier does?
