Re: PATCH: PR rtl-optimization/55093: [4.8 Regression] [x32] -maddress-mode=long failed

[Richard Sandiford <rdsandiford at googlemail dot com>]

"H.J. Lu" <hjl.tools@gmail.com> writes:
>>> LRA has
>>>
>>>       if (REG_P (reg) && (ep = get_elimination (reg)) != NULL)
>>>         {
>>>           rtx to_rtx = replace_p ? ep->to_rtx : ep->from_rtx;
>>>
>>>           if (! replace_p)
>>>             {
>>>               offset += (ep->offset - ep->previous_offset);
>>>               offset = trunc_int_for_mode (offset, GET_MODE (plus_cst_src));
>>>             }
>>>
>>>           if (GET_CODE (XEXP (plus_cst_src, 0)) == SUBREG)
>>>             to_rtx = gen_lowpart (GET_MODE (XEXP (plus_cst_src, 0)), to_rtx);
>>>
>>> Reload has
>>>
>>>             rtx to_rtx = ep->to_rtx;
>>>             offset += ep->offset;
>>>             offset = trunc_int_for_mode (offset, GET_MODE (plus_cst_src));
>>>
>>>             if (GET_CODE (XEXP (plus_cst_src, 0)) == SUBREG)
>>>               to_rtx = gen_lowpart (GET_MODE (XEXP (plus_cst_src, 0)),
>>>                                     to_rtx);
>>>
>>> (gdb) call debug_rtx (ep->to_rtx)
>>> (reg/f:DI 7 sp)
>>> (gdb) call debug_rtx (ep->from_rtx)
>>> (reg/f:DI 16 argp)
>>> (gdb)
>>>
>>> gen_lowpart returns (reg/f:DI 7 sp) for reload and (reg:SI 16 argp)
>>> for LRA.   They are caused by
>>>
>>>   if (FRAME_POINTER_REGNUM != ARG_POINTER_REGNUM
>>>       /* We should convert arg register in LRA after the elimination
>>>          if it is possible.  */
>>>       && xregno == ARG_POINTER_REGNUM
>>>       && ! lra_in_progress)
>>>     return -1;
>>>
>>> It doesn't work in this case.
>>>
>>
>> This testcase shows that LRA can't convert arg register after
>> the elimination.
>>
> Here is a patch to remove ra_in_progress check for
> ARG_POINTER_REGNUM.  Tested on Linux.x86-64.
> OK to install?

Thanks HJ.  This looks good to me.  As well as your testcase, I think
it would be dangerous to reduce this kind of subreg during non-final
elimination in cases where the argument pointer occupies more than one
hard register (like avr IIRC).  We could end up with something like
ARG_POINTER_REGNUM+1, which wouldn't show up as an elimination register
during the rest of LRA.

It's important that we do get rid of the subreg during the final
elimination stage, but I think alter_subreg already handles that case.

Since this code is outside the LRA files: patch is OK if Vlad agrees.

Richard