Re: [RFA] MIPS DWARF2 CFI support

[Andrew Cagney <cagney@gnu.org>]

> On Mon, Jun 07, 2004 at 08:49:08PM -0700, Martin M. Hunt wrote:
> 
>>> This is really kevinb's work, with some minor tweaks.  I've tested this
>>> on over 50 multilib combinations and it looks goood.
>>> 
>>> 2004-05-26  Kevin Buettner: 
>>> 
>>> 	* dwarf2-frame.c (execute_cfa_program): Fix typo in which the
>>> 	alignment was being added to the offset instead of multiplied.
>>> 
>>> 	* dwarf2-frame.c (struct comp_unit): Add new field ``signed_addr_p''.
>>> 	(encoding_for_size): Add new parameter ``signed_addr_p''.  Adjust
>>> 	all callers.  Add code for handling signed encodings.
>>> 	(dwarf2_build_frame_info): Initialize ``unit.signed_addr_p''.
>>> 
>>> 	* dwarf2-frame.c (dwarf2_build_frame_info): Set unit.addr_size.
>>> 
>>> 	* config/mips/tm-mips.h (SP_REGNUM): Delete define.
>>> 	* mips-tdep.h (MIPS_SP_REGNUM): Define.
>>> 	* mips-tdep.c (mips_gdbarch_init): Set SP_REGNUM via call
>>> 	to set_gdbarch_sp_regnum().  Use cooked register number.
>>> 	(SP_REGNUM): Replace all occurrences with MIPS_SP_REGNUM.
>>> 
>>> 	* mips-tdep.c (mips_gdbarch_init): Set PC_REGNUM (via
>>> 	call to set_gdbarch_pc_regnum()) to cooked register number.
>>> 
>>> 	* mips-tdep.c (dwarf2-frame.h): Include.
>>> 	(mips_gdbarch_init): Enable DWARF2 CFI support.
> 
> 
> I've got three objections:
>   - This is a whole lot of changes.  Please submit them independently,
>     with some sort of attempt at explanation.

(Yes please, the SP_REGNUM looked to be an easy one to split out and the 
bulk of the change).

>>> -  unit.addr_size = objfile->obfd->arch_info->bits_per_address / 8;
>>> +  unit.addr_size = TYPE_LENGTH (builtin_type_void_data_ptr);
>>> +  unit.signed_addr_p = bfd_get_sign_extend_vma (unit.abfd);
> 
> 
>   - This is wrong; please don't introduce dependencies on GDB's type
>     system in dwarf2-frame.c.  I imagine what's there isn't right
>     either because arch_info is completely unreliable on MIPS.  There
>     is already a dwarf64_p where this is used; is that right?
> 
>     Actually, this sounds like a problem in the dwarf3 draft that
>     someone (maybe Jim?) brought up on the dwarf2 list a few weeks ago.
>     Jim, do you remember?
> 
> And somewhat theoretical:
>   - I'd really rather not enable dwarf2 unwinding for MIPS until
>     someone makes a go at fixing the MIPS prologue unwinder.  Either
>     that or rip it out; because after this patch it will get almost
>     zero coverage in the testsuite.  This is an ongoing problem.
>     I know that the existing prologue unwinder is broken at the
>     beginning of functions and in a number of other situations.

I don't think delaying CFI until the prologue unwinder is working is 
reasonable.  However, what would help is the addition of a new more 
specialized stub-unwinder (see hppa and s390) (perhaphs later we can 
return to cleaning up the existing prologue unwinder).  Patches for the 
former (the stub unwinder) more than welcome.

Andrew