[C++ PATCH] Don't clear DECL_BUILT_IN_CLASS/DECL_FUNCTION_CODE when redeclaring a builtin if types match

[Jakub Jelinek <jakub at redhat dot com>]

Hi!

The C FE always keeps DECL_BUILT_IN_CLASS when types match, even when
newdecl defines a function.
      if (DECL_BUILT_IN (olddecl))
        {
          /* If redeclaring a builtin function, it stays built in.
             But it gets tagged as having been declared.  */
          DECL_BUILT_IN_CLASS (newdecl) = DECL_BUILT_IN_CLASS (olddecl);
          DECL_FUNCTION_CODE (newdecl) = DECL_FUNCTION_CODE (olddecl);
          C_DECL_DECLARED_BUILTIN (newdecl) = 1;
          if (new_is_prototype)
            C_DECL_BUILTIN_PROTOTYPE (newdecl) = 0;
          else
            C_DECL_BUILTIN_PROTOTYPE (newdecl)
              = C_DECL_BUILTIN_PROTOTYPE (olddecl);
        }
But the C++ FE doesn't do this:
      if (new_defines_function)
	...
      else if (types_match)
	{
          /* If redeclaring a builtin function, and not a definition,
             it stays built in.  */
          if (DECL_BUILT_IN (olddecl))
            {
              DECL_BUILT_IN_CLASS (newdecl) = DECL_BUILT_IN_CLASS (olddecl);
              DECL_FUNCTION_CODE (newdecl) = DECL_FUNCTION_CODE (olddecl);
              /* If we're keeping the built-in definition, keep the rtl,
                 regardless of declaration matches.  */
              COPY_DECL_RTL (olddecl, newdecl);
            }
	...

This causes quite undesirable behavior e.g. on glibc headers on ppc with
-mlong-double-64 (as shown on the testcase).  Attached are two different
patches, one does what the C FE does and copies DECL_BUILT_IN_CLASS
etc. whenever types_match (bootstrapped/regtested on x86_64-linux).
IMHO when you define memcpy in current CU (or some other builtins), there
is no reason why any calls to that builtin from within other functions in
the CU shouldn't be suddenly optimized as builtins.

The second alternative patch only does that for gnu_inline wrappers
- especially if the function is just extern inline
__attribute__((gnu_inline)) without always_inline attribute, it might
not be inlined in some places.  Still they should be treated as builtin
calls.  I very much prefer the first patch though.

Ok for trunk?

	Jakub

Attachment: E2
Description: Text document

Attachment: E
Description: Text document