On Fri, Sep 04, 2015 at 06:04:15PM -0500, Segher Boessenkool wrote: > On Fri, Sep 04, 2015 at 04:16:40PM -0400, Rich Felker wrote: > > One thing I've noticed that's odd is that gcc -mfdpic -fPIC produces > > different (less efficient) code from just gcc -mfdpic, which seems > > wrong, but agrees with sh.c which has a number of checks for flag_pic > > not matched with a TARGET_FDPIC check. > > Generic code tests flag_pic in important places as well. > > > I'm thinking all of these > > should either be flag_pic||TARGET_PIC or flag_pic&&!TARGET_FDPIC, > > depending on whether the code applies to all PIC or is specific to the > > non-FDPIC PIC model where r12 is call-saved. Does this sound correct? > > I think we need spurious -fPIC to work (although it could be handled > > with spec magic) and not pessimize code, since most library builds > > will use -fPIC. > > If you never want -fPIC (or -fpic) if fdpic is enabled, you can disable > it (in sh_option_override)?
It turns out that with !flag_pic, gcc is generating broken code and/or ICE, and this happens even after changing all the remaining flag_pic tests in sh.c to flag_pic||TARGET_FDPIC. There are a few more in sh.md I didn't try changing but they did not look relevant; the ICE came via expand_binop in prepare_move_operands. Before I look at it further, is there any reason to expect !flag_pic in the generic code to break things when the target-specific code has PIC-like constraints? For now I just made sh_option_override force flag_pic when TARGET_FDPIC is set. Note that flag_pic by itself is equivalent to -fPIE; -fPIC also sets flag_shlib which affects other things like TLS model and binds_locally interpretation. This seems like a viable solution (and I got rid of the suboptimal codegen by fixing the condition in sh_function_ok_for_sibcall so that flag_pic doesn't preclude sibcalls when TARGET_FDPIC is set) but I'd still like to figure out why gcc is breaking without flag_pic... Rich
