Re: [PATCH] stdarg optimization (4.1 projects, stage 1.2)

[Eric Botcazou <ebotcazou at libertysurf dot fr>]

Hi Jakub, Richard,

I briefly looked at this optimization for SPARC and I now wonder if it worth 
the cycles there:

#include <stdarg.h>

int foo(int i, ...)
{
  va_list ap;
  int j;

  va_start(ap, i);
  j = va_arg(ap, int);
  va_end(ap);
  return j;
}

at -O0:
foo:
        save    %sp, -120, %sp
        st      %i1, [%fp+72]
        st      %i2, [%fp+76]
        st      %i3, [%fp+80]
        st      %i4, [%fp+84]
        st      %i5, [%fp+88]
        st      %i0, [%fp+68]
        add     %fp, 72, %g1
        st      %g1, [%fp-24]
        ld      [%fp-24], %i5
        mov     %i5, %g1
        ld      [%g1], %g1
        st      %g1, [%fp-20]
        add     %i5, 4, %g1
        st      %g1, [%fp-24]
        ld      [%fp-20], %g1
        mov     %g1, %i0
        restore
        jmp     %o7+8
         nop

at -O1:
foo:
        add     %sp, -120, %sp
        mov     %o1, %o0
        jmp     %o7+8
         sub    %sp, -120, %sp

And of course the 6 registers are not spilled to the variable part of the 
frame, but to the fixed part.  Same for the 64-bit ABI.

So the best solution could be to define a target hook, defaulting to false and 
set to true on SPARC, so as to always return false in gate_optimize_stdarg.

What do you think?  Thanks in advance.


Btw, Jakub, TARGET_STDARG_OPTIMIZE_HOOK is not documented AFAICS.

-- 
Eric Botcazou