Re: PATCH: Don't allocate space for SSE reg in reg save area if SSE is disabled

[Jakub Jelinek <jakub at redhat dot com>]

On Thu, Aug 28, 2008 at 10:04:49AM -0700, H.J. Lu wrote:
> On Thu, Aug 28, 2008 at 09:57:45AM -0700, H.J. Lu wrote:
> > 
> > We are allocating space for SSE reg in reg save area even if SSE is
> > disabled. -mno-sse is used on Linux kernel.  Here is patch to avoid it.
> > OK for trunk and 4.3?
> > 
> 
> Here is the updated patch with some optimization.  OK for  trunk and
> 4.3?

Does this actually work, even if e.g. there are no integer, but some
floating point va_args?
Say
double foo (int n, ...)
{
  float ret = 0.0;
  va_list ap;
  va_start (ap, n);
  for (int i = 0; i < n; i++)
    ret += va_arg (ap, double);
  va_end (ap);
  return 0;
}
I'm afraid it would just clobber wrong portion of the stack.

E.g. setup_incoming_varargs_64 does:
      emit_insn (gen_rtx_SET (VOIDmode, tmp_reg,
                              plus_constant (save_area,
                                             8 * X86_64_REGPARM_MAX + 127)));
where I'd expect you want with your patch ix86_varargs_gpr_size + 127.
Similarly, ix86_va_start needs some changes too if the area for gpr
registers is left out.  Either change:
      t = build2 (MODIFY_EXPR, type, fpr,
                  build_int_cst (type, n_fpr * 16 + 8*X86_64_REGPARM_MAX));
- if ix86_varargs_gpr_size is zero, then supposedly va_list doesn't escape
and so we don't care about ABI effects.  Or adjust ap.reg_save_area
initialization (sav) instead, if ix86_varargs_gpr_size is 0 subtract
8*X86_64_REGPARM_MAX from it.

	Jakub