Re: Medium model support in GCC

[Richard Henderson <rth at redhat dot com>]

On Fri, Jul 01, 2005 at 08:10:16PM +0200, Jan Hubicka wrote:
> + /* Variables which are this size or smaller are put in the data/bss
> +    or ldata/lbss sections.  */
> + 
> + static int ix86_section_threshold = 65536;

You'll need to make this controlled by some sort of compiler switch.

> +   if ((ix86_cmodel != CM_MEDIUM && ix86_cmodel != CM_MEDIUM_PIC)
> +       || !ix86_in_large_data_p (decl))
> +     default_elf_select_section (decl, reloc, align);
> +   else
...
> +       if (sname)
> +         named_section (decl, sname, reloc);
> +       else
> +         default_elf_select_section (decl, reloc, align);

I think this would be better as

  if ((ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_MEDIUM_PIC)
      && ix86_in_large_data_p (decl))
    {
      ...
      if (sname)
	{
	  named_section (decl, sname, reloc);
	  return;
	}
    }
  default_elf_select_section (decl, reloc, align);

> + x86_64_elf_unique_section (tree decl, int reloc)

Likewise.

> +     fprintf (file, ".largecomm ");

We usually use tabs.

> + 	  case UNSPEC_GOTOFF:
> + 	    /* For 64bit target we have to deal with GOTOFF as 64bit values,
> + 	       for 32bit targets these are always part of addressing mode
> + 	       there is no point to allow dealing with them separately -
> + 	       doing so just increase register pressure.  */
> + 	    if (!TARGET_64BIT)
> + 	      return false;
> + 	    return local_symbolic_operand (XVECEXP (x, 0, 0), Pmode);

I don't like this.

Primarily becuase we should never be creating GOTOFF relocations to
non-local symbols, and thus there should be nothing to check.  The
code here should be "return true", because the constant should, by
definition, ALWAYS be legitimate.

>         if (GET_CODE (disp) == SYMBOL_REF
> ! 	  && (ix86_cmodel == CM_SMALL_PIC
> ! 	      || (ix86_cmodel == CM_MEDIUM_PIC
> ! 		   && !SYMBOL_REF_FAR_ADDR_P (disp)))
>   	  && SYMBOL_REF_LOCAL_P (disp))

Couldn't you just always test SYMBOL_REF_FAR_ADDR_P and ignore ix86_cmodel?

> +   else if (TARGET_64BIT && local_symbolic_operand (addr, Pmode)
> + 	   && ix86_cmodel != CM_SMALL_PIC)

Do easy tests first.  I.e. the function call test last.

> + /* Flag to mark data that is in the large address area (addressable
> +    via "addl", that is, within a 2GByte offset of 0.  */
> + #define SYMBOL_FLAG_FAR_ADDR		(SYMBOL_FLAG_MACH_DEP << 0)

Surely *not* addressable via addl.

> !  	    (and (ne (symbol_ref "flag_pic && ix86_cmodel == CM_SMALL_PIC")
> ! 	    	     (const_int 0))
> !   		 (match_operand:DI 1 "symbolic_operand" ""))

This shouldn't be necessary.  We'll have legitimized the move to
something else.

> + (define_insn "set_got_rex64"
> +   [(set (match_operand:DI 0 "register_operand" "=r")
> + 	(unspec:DI [(const_int 0)] UNSPEC_SET_GOT))
> +    (clobber (reg:CC 17))]
> +   "TARGET_64BIT"
> +   "lea{q}\t_GLOBAL_OFFSET_TABLE_(%%rip), %0"

This doesn't clobber reg:CC.

> Index: testsuite/gcc.dg/x86-64-medium-1.c

The new tests should go in gcc.target.

They also need some check vs lp64, otherwise you're going to fail the
32-bit multilib.

> + /* { dg-options "-fpic -mcmodel=medium" { target { { x86_64-*-* } && ilp32 } } } */

Huh?  Why ilp32?



r~