Hi, My target has HI and PSI hardware registers. HI hwregs can hold PSI values, and PSI registers (ex A0 reg) can hold HI values. My target is big endian and my problem occurs when I split a movpsi when gcc computes sub registers of PSI hw register. For example when using simplify_gen_subreg: - (subreg:HI (reg:PSI A0) 2) which is the least significant part returns NULL_RTX - (subreg:HI (reg:PSI A0) 0) which is the most significant part returns (reg:HI A0). This means gcc consider that the HI part of a PSI reg is located at the MS part of the register which is wrong on my target. Hence, the cprop_hardreg pass may replace some registers by wrong ones when PSI registers are split. Is it a wishable behavior and is there a way to make such a split work for PSI moves, or should I manage such a split at assembly level, preventing some further optimizations ? Thanks, Aurélien
