Re: DWARF location descriptor and multi-register frame pointer

[Richard Henderson <rth at redhat dot com>]

On 11/20/2012 02:22 AM, Senthil Kumar Selvaraj wrote:
> Hi all,
> 
>  For the AVR target, the FP register spans two registers, yet the DWARF
>  location information (DW_AT_location) for a local variable refers to only one 
>  of them.
> 
>  Looking at the code, I found that based_loc_descr in dwar2out.c, which handles
>  base+offset based location descriptions, gets the register number from
>  the rtx representing the reg and directly proceeds to create DW_OP_breg0+regno,
>  without considering the mode. This is unlike reg_loc_descriptor, which
>  checks hard_regno_nregs[regno][mode] and acts accordingly.
> 
>  How are frame pointer registers that span more than one hard register
>  handled? 

I would have to say "poorly", because it simply doesn't happen for any
other target.

I'm not really sure what advantage you'd have for using a multiple_reg_loc
to represent the frame pointer in the DW_AT_location for the frame base.
Especially since if you consider Dwarf3 and later should be indicating
the frame base via the CFA (DW_OP_call_frame_cfa).

>  A related question - why does based_loc_descr generate DW_OP_fbreg 
>  only if eliminate_regs eliminates the frame/arg pointer register? I
>  traced through the execution for my native (x86_64) compiler, and found
>  that the elimination is actually from the soft frame pointer to the hard frame
>  pointer. What if HARD_FRAME_POINTER_REGNUM is the same as
>  FRAME_POINTER_REGNUM? Shouldn't DW_OP_fbreg be generated for those
>  targets too?

Probably.  It would take some digging to figure out why fbreg might not
be in use unless elimination succeeded.


r~