Re: dwarf2-frame.c question for maintainers

["Martin M. Hunt" <hunt at redhat dot com>]

On Thu, 2004-07-15 at 10:58, Andrew Cagney wrote:
> > On Wed, 2004-07-14 at 12:28, Andrew Cagney wrote:
> > 
> >>>>>> >>> How come extract_typed_address, in read_reg, doesn't sign extend?
> >>>
> >>>> > 
> >>>> > 
> >>>> > I should have explained that. It does.  However extract_typed_address is
> >>>> > incorrect because it makes the invalid assumption that sizeof(address)
> >>>> > == sizeof(register).  So that has to go and be replaced with something
> >>>> > like
> >>>> > extract_signed_integer (buf, register_size (current_gdbarch, regnum));
> >>
> >>> 
> >>> You mean the builtin_type_void_data_ptr parameter to 
> >>> extract_typed_address?  Ah.
> >>> 
> >>> I see builtin_type_void_data_ptr dates back to 1.1 (Mark?).  It could 
> >>> instead use the register's type?
> > 
> > 
> > extract_typed_address calls extract_[un]signed_integer with size =
> > TYPE_LENGTH of builtin_type_void_data_ptr.
> > 
> > Here's exactly what I am seeing.  Maybe you can tell me if read_reg is
> > the problem.
> > 
> > For example big-endian Mips, with o64 or (eabi and mlong32):
> > (registers are 64 bits and pointers are 32 bits)
> > 
> > read_reg calls frame_unwind_register (next_frame, regnum, buf)
> > after that, buf has something like ffffffff801fffb8
> > 
> > Now if you do extract_typed_address(), it knows addresses are 4 bytes
> > and returns 0xffffffff sign extended to 0xfffffffffffffff
> 
> Right, as it stands, that call is just wrong.
> 
> > If instead, you call extract_[un]signed_integer((buf, register_size
> > (current_gdbarch, regnum)), it returns 0xffffffff801fffb8
> > 
> > The real problem here is the the size.  AFAICT, sign-extension here is
> > unimportant; I get the same test results calling 
> > extract_unsigned_integer in read_reg() for mips, because, as you can
> > see, nothing needs extending, just the whole register needs read.
> > However, I can't prove that is always the case because I am not familiar
> > enough with the code.
> 
> Consider o32.  Both the ABI and ISA are 32-bits, but GDB's CORE_ADDR may 
> be 64-bits.  Even if it doesn't appear to make a difference, the MIPS 
> needs to always sign extend addresses/registers - that's the dogma :-)

Right.  And so back to the original question.  What is the best way to
have read_reg detect if it should sign-extend?  We agreed that passing a
pointer to the CU was out, as was using a global.  Do I need to add
something to gdbarch?  

-- 
Martin M. Hunt <hunt@redhat.com>
Red Hat Inc.