Re: [RFC] Mips, N32, cc, gcc, and gdb (longish)

[Daniel Jacobowitz <drow at mvista dot com>]

On Thu, Aug 01, 2002 at 06:14:32PM -0700, Michael Snyder wrote:
> OK, I think I've made this as concise as practical.   ;-)
> 
> In case the difference between Irix cc and gcc does not get 
> fixed (and I mean specifically the difference concerning 
> alignment of small struct arguments), I'd like to discuss 
> "fixing" gdb so that it will work correctly with both.
> 
> I've found a testcase for this problem in call-ar-st.exp:
> 
> (gdb) print print_long_arg_list (a, b, c, d, e, f, *struct1, 
> 	*struct2, *struct3, *struct4, *flags, *flags_combo, 
> 	*three_char, *five_char, *int_char_combo, 
> 	*d1, *d2, *d3, *f1, *f2, *f3)
> 
> This fails (or doesn't) depending on whether a small struct
> is left-justified or right-justified when it is pushed onto
> the stack (ie, when it is the 5th or higher argument).
> 
> The code in mips_push_arguments that controls this is:
> 
>       if (MIPS_STACK_ARGSIZE == 8 &&
>           (typecode == TYPE_CODE_INT ||
>            typecode == TYPE_CODE_PTR ||
>            typecode == TYPE_CODE_FLT) && len <= 4)
>         longword_offset = MIPS_STACK_ARGSIZE - len;
>       else if ((typecode == TYPE_CODE_STRUCT ||
>                 typecode == TYPE_CODE_UNION) &&
>               TYPE_LENGTH (arg_type) < MIPS_STACK_ARGSIZE)
>         longword_offset = MIPS_STACK_ARGSIZE - len;
>     }
> 
> As it is written, it "works" with gcc but fails with cc.
> 
> Kevin proposed this change to me (I don't remember if
> he has submitted it yet):
> 
> ***
>                 typecode == TYPE_CODE_UNION) &&
> !              TYPE_LENGTH (arg_type) < MIPS_STACK_ARGSIZE)
>         longword_offset = MIPS_STACK_ARGSIZE - len;
> ---
>                 typecode == TYPE_CODE_UNION) &&
> !              TYPE_LENGTH (arg_type) < MIPS_STACK_ARGSIZE  
> !              && tdep->mips_abi != MIPS_ABI_N32)
>         longword_offset = MIPS_STACK_ARGSIZE - len;
> ***
> 
> With this change, it "works" with cc, but fails with gcc.
> 
> I want to discuss the following change, which should 
> make it "work" with both cc and gcc (given the current
> behavior of both):
> 
> ***
>                 typecode == TYPE_CODE_UNION) &&
> !              TYPE_LENGTH (arg_type) < MIPS_STACK_ARGSIZE)
>         longword_offset = MIPS_STACK_ARGSIZE - len;
> ---
>                 typecode == TYPE_CODE_UNION) &&
> !              TYPE_LENGTH (arg_type) < MIPS_STACK_ARGSIZE  
> !              && (gcc_p ||tdep->mips_abi != MIPS_ABI_N32))
>         longword_offset = MIPS_STACK_ARGSIZE - len;
> ***
> 
> 
> This of course makes gdb's behavior dependent on which 
> compiler it detects.
> 
> There is one problem: there is no variable "gcc_p" within 
> the scope of mips_push_arguments, because PUSH_ARGUMENTS 
> does not pass it.  So that would need to be solved, 
> possibly by modifying the definition of PUSH_ARGUMENTS.
> 
> Comments?

Comment - you're on a slippery slope.  I know of at least one other
variation in this area; for structures of less than a word SGI CC
shifts them in register only for big endian targets, and not for little
endian (on the old versions of CC which support little endian).  There
comes a point where we just need to get GCC fixed, and I think this is
it.

-- 
Daniel Jacobowitz                           Carnegie Mellon University
MontaVista Software                         Debian GNU/Linux Developer