> I see that's not the problem. AFAICT from the code that deals with > builtin_apply and builtin_return it looks like untyped_call and > untyped_return should be doing what you expect for the code to save the > values (a quick glance at sparc.md didn't show any problems, but...). Or > is something else clobbering later?
Again, the problem is that apply_result_mode is now wrong for values returned in multi hard registers: suppose that a function returns 'long long', so DImode. The canonical return register on SPARC is (reg/i:DI 8 %o0) in that case. Now the reg_raw_mode of %o0 is SImode, not DImode, on SPARC 32-bit so apply_result_mode[8] = SImode. And, given FUNCTION_VALUE_REGNO_P, apply_result_mode[9] = VOIDmode and size == 4. The return value is, therefore, only partially reloaded. Same for _Complex int, double and so on. Sure, untyped_call and untyped_return can be hacked to work around this (a la ARM), but this would need to be done in every back-end. So I think you should find another approach to fix your MIPS-specific problem, possibly by hacking MIPS' untyped_call and untyped_return. -- Eric Botcazou
