https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126059
Hongyu Wang <hongyuw at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |hongyuw at gcc dot gnu.org
--- Comment #2 from Hongyu Wang <hongyuw at gcc dot gnu.org> ---
For i128 arith we use earlyclobber &r to NDD alternatives to avoid
reg-allocation overlapping for paired regs.
We can add peephole to fold them back to NDD for add/add_cst/sub_cst_i128, but
unless we see speedup we don't have reason to do it. The legacy sequence using
mov+add is shorter than the NDD counterparts, and the llvm-mca shows no better
cycle.
See https://godbolt.org/z/5csjG9YM7
So suppose we don't have enough reason to do such peephole.
