On Thu, 2018-07-12 at 07:17 +0100, Richard Sandiford wrote: > > So it only calls targetm.hard_regno_call_part_clobbered if such a > call is known to exist somewhere between the two references to > regno (although we don't have the calls themselves to hand). > > Thanks, > Richard
Having the specific calls is the problem here because the normal ABI and the SIMD ABI are going to have different values for caller_save and part_clobbered. But I think I have found a better way to address this. Looking at 'need_for_call_save_p', when 'flag_ipa_ra' is set, we look at 'actual_call_used_reg_set' to see if we need to save a register. But even if a particular register doesn't show up as used, if 'hard_regno_call_part_clobbered' is set for that register we are going to return true and say we need a save/restore of the register. On Aarch64, if I 'really' didn't use the register I shouldn't need to save/restore it. If I used it but it is callee saved then on 'normal' functions I may need to save/restore it (to protect the upper bits) but on SIMD functions I do not need to save/restore it at all because the callee will save/restore the entire register and not just the lower 64 bits. I think that the patch I want to create is: when 'flag_ipa_ra' is true, remove the check of 'hard_regno_call_part_clobbered' from 'need_for_call_save_p'. Instead, check 'hard_regno_call_part_clobbered' in 'process_bb_lives' (where we know exactly what function is being called) and use that to set 'actual_call_used_reg_set'. In process_bb_live we know exactly what function we are calling and can check to see if it is a 'normal' function or a SIMD function. Steve Ellcey sell...@cavium.com
