Re: [RFA] mips 32/64 register/stack fix

[Daniel Jacobowitz <drow at mvista dot com>]

On Thu, Apr 29, 2004 at 09:40:23AM -0400, Andrew Cagney wrote:
> >On Wed, Apr 28, 2004 at 05:45:22PM -0400, Andrew Cagney wrote:
> >
> >>>>>Here's your tdep info:
> >>>>>{elf_flags = 1610625025, mips_abi = MIPS_ABI_EABI32,
> >>
> >>>
> >>>I was looking to see if anything like that was present -> it is clearly 
> >>>identified as a 32-bit ABI binary.
> >>>
> >>>I think the bug is in mips_register_type, the tail end should probably 
> >>>be changed to read something like:
> >>>
> >>>  else if (regnum >= NUM_REGS
> >>>	   && gdbarch_tdep (gdbarch)->mips64_transfers_32bit_regs_p)
> >>>    /* The target, while using a 64-bit raw register buffer, is only
> >>>       transfering 32-bits of each integer register.  Reflect this in
> >>>       the cooked/pseudo register value.  */
> >>>    return builtin_type_int32;
> >>>  else if (regnum > NUM_REGS
> 
> >= NUM_REGS
> 
> >>>	   && mips_saved_regsize (gdbarch) == 4)
> >>>    /* A 32-bit ABI such as o32 possibly running on a 64-bit ISA.  */
> >>>    return builtin_type_int32;
> >>>  else if (mips_regsize (gdbarch) == 8)
> >>>    /* 64-bit ISA.  */
> >>>    return builtin_type_int64;
> >>>  else
> >>>    /* 32-bit ISA.  */
> >>>    return builtin_type_int32;
> >
> >
> >Would you mind clarifying the reason for having mips_regsize, which is
> >used in a number of other places, return a regsize based on the ISA?
> 
> Hasn't this all come up before?

Yes, but obviously it hasn't been resolved :)  I've dropped it a couple
of times - got distracted.

> It's the register size of the underlying ISA, on top of which the ABI is 
> implemented.  I think I'll rename mips_regsize and mips_saved_regsize to 
> clarify this.

Please.

> >This change to mips_register_type conveys the fact that we've only got
> >32 bits of data.  But we'll choose to print (in info registers) a
> >64-bit wide field for each GPR if the binary is tagged E_MIPS_ARCH_64 |
> >E_MIPS_ABI_EABI32, and a 32-bit field if it's tagged E_MIPS_ARCH_2 |
> >E_MIPS_ABI_EABI32.  Conceptually, I think we're interested in some
> >combination of the available register size (-> a property of the target)
> >and the size of registers known to the inferior program (-> unclear mix
> >of its ABI and ISA).
> 
> It does?  With the above it should always print a 32-bit GPR reflecting 
> the 32-bit ABI.  If you think it should display something else, look at 
> user-regs which lets you implement registers using the frame register 
> values (something psuedo registers can't handle).

It makes decisions based on mips_regsize.  See print_gp_register_row.
So does mips32_heuristic_proc_desc; which has probably been converted
incorrectly at some point in the past, since it takes "8 - mips_regsize
(current_gdbarch)" and wants 4 for N32.  I think that whole block of
code is bogus.

Hmm, I was just using print_gp_register_row as an example - but it
turns out to be the only example.  Only the two functions above and
mips_register_type reference mips_regsize now - and a bunch of NetBSD
specific code which is almost certainly wrong.  With the functions
renamed, I guess it'll be much easier to spot the errors.

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer