Re: [patch] use __builtin_ctzl in ggc-page

[Steven Bosscher <stevenb at suse dot de>]

On Monday 17 January 2005 13:38, Richard Earnshaw wrote:
> > 	* ggc-page.c (ggc_alloc_stat): Use __builtin_ctzl instead of a
> > 	loop to look for a free slot in a page entry.
>
> I would expect this makes things worse for platforms that don't have the
> builtin, since we now end up with a library call overhead on top of the
> calculation cost.

Note that there are c[lt]zm2 and ffsm2 named patterns so in theory
it should be possible to avoid a libcall for all targets.  There are
some quite efficient target specific versions of c[lt]z in libgcc2's
longlong.h and as a fall-back we could add some generic version like
the one in bitmap.c:

#if BITMAP_WORD_BITS > 32
  if (!(word & 0xffffffff))
    word >>= 32, bit_no += 32;
#endif
  if (!(word & 0xffff))
    word >>= 16, bit_no += 16;
  if (!(word & 0xff))
    word >>= 8, bit_no += 8;
  if (!(word & 0xf))
    word >>= 4, bit_no += 4;
  if (!(word & 0x3))
    word >>= 2, bit_no += 2;
  if (!(word & 0x1))
    word >>= 1, bit_no += 1;

  gcc_assert (word & 1);

It should be possible to produce a similar code sequence in RTL from
builtins.c, right?  I don't know how I'd have to do that, probably
something with optabs, but I don't know if it is possible somehow to
have an optab that generatez inline RTL for a builtin.  Hints would
be appreciated ;-)

For GCC 3.4 and later as the build compiler bitmap.c always uses the
ctz builtin, and so does ggc-page now, but apparently the only target
descriptions with a czt pattern are alpha, arm, h3800, i386, ia64,
and rs6000.  I suppose bootstrapping on other hosts would result in a
libcall for the __builtin_ctz calls.  That would, in fact, be a gcc 4
regression...

Is there some way to detect if the host needs a libcall for a builtin?

You already said in another thread of this discussion that we should
call the builtins via macros.  I suppose you're right.  I'll try to
figure out something for this.

Gr.
Steven