force_const_mem patch for unspecs

[HPENNER at de dot ibm dot com]

We would like to propose a change of some details of the implementation of
decode_rtx_const. This changes are essential for the integration of s390
backend into the gcc-3.1.

Problem:
     The s390 backend forces (const (unspec (symbol_ref))) and (const (plus
(unspec (symbol_ref)) const_int)) into the
constant pool. The unspec index is used to distinguish between the
extenstions of the symbols (@plt, @got, ...).
decode_rtx_const sets value->un.addr.base to rtx expression (unspec
(symbol_ref)). Since each expression
is unique, the hash-value of such expression will be different, leading to
duplicate entries in the constant pool.

Proposed solution:
     decode_rtx_const checks, if the expression is a simple umspec
expression. The address of the string is then used for
value->un.addr.base, the unspec index is coded into value->kind. This will
inhibit dublicate entries in the constant pool
and distinguish between same symbol_refs and different unspec indexes
(a@got and a@PLT could be both in the same constant pool)


Index: varasm.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/varasm.c,v
retrieving revision 1.176
diff -r1.176 varasm.c
2227c2227
< enum kind { RTX_DOUBLE, RTX_INT };
---
> enum kind { RTX_DOUBLE, RTX_INT, RTX_UNSPEC };
3433c3433,3445
<   if (value->kind == RTX_INT && value->un.addr.base != 0)
---
>   if (value->kind == RTX_INT && value->un.addr.base != 0
>       && GET_CODE (value->un.addr.base) == UNSPEC)
>     {
>       /* For a simple UNSPEC, compare index and,
>     if available the string's address of SYMBOL_REF */
>       if (XVECLEN (value->un.addr.base, 0) == 1)
>    {
>      value->kind = RTX_UNSPEC + XINT (value->un.addr.base, 1);
>      value->un.addr.base = XVECEXP (value->un.addr.base, 0, 0);
>    }
>     }
>
>   if (value->kind != RTX_DOUBLE && value->un.addr.base != 0)
3435a3448
>
3445c3458
<
---
>
3463c3476,3477
<   if (val0.un.addr.base == val1.un.addr.base)
---
>   if (val0.kind != RTX_DOUBLE && val0.kind == val1.kind
>       && val0.un.addr.base == val1.un.addr.base)

Mit freundlichem Gruß / Best regards,

Hartmut Penner

Internet Mail Address : hpenner@de.ibm.com

Tel: (49)-7031-16-4364