https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120955
--- Comment #32 from Georg-Johann Lay <gjl at gcc dot gnu.org> --- (In reply to fiesh from comment #31) > Also, can anyone provide hints what optimization step to look for? As it seems, the tree oprimizers are not involved for the test case: With -O2 -fdump-tree-all you'll find code with "= 97" from .007t.gimple through .271t.optimized, whereas with -Os -fdump-tree-all code like grep "LC0" is present from .gimple through .optimized. This indicates that the frontend plays a role. Maybe acr.cc::avr_use_by_pieces_infrastructure_p() is sub-optimal and and has an effect.
