On Fri, Aug 05, 2016 at 11:35:11AM +0930, Alan Modra wrote: > diff --git a/gcc/config/rs6000/rs6000.c b/gcc/config/rs6000/rs6000.c > index 779ba1f..c59d07a 100644 > --- a/gcc/config/rs6000/rs6000.c > +++ b/gcc/config/rs6000/rs6000.c > @@ -7717,8 +7717,34 @@ mem_operand_gpr (rtx op, machine_mode mode) > if (TARGET_POWERPC64 && (offset & 3) != 0) > return false; > > - if (mode_supports_vsx_dform_quad (mode) > - && !quad_address_offset_p (offset)) > + extra = GET_MODE_SIZE (mode) - UNITS_PER_WORD; > + if (extra < 0) > + extra = 0; > + > + if (GET_CODE (addr) == LO_SUM) > + /* For lo_sum addresses, we must allow any offset except one that > + causes a wrap, so test only the low 16 bits. */ > + offset = ((offset & 0xffff) ^ 0x8000) - 0x8000;
You could use sext_hwi instead? offset = sext_hwi (offset, 16); There are many other explicit manipulations in the code already of course, just FYI. This seems like an improvement; okay for trunk, thanks. I don't think we're there yet though, there is more wrong / suboptimal than just the PR :-/ Segher
