[Bug tree-optimization/72785] [7 Regression] kernel build error since r236831

["dhowells at redhat dot com" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72785

--- Comment #15 from dhowells at redhat dot com <dhowells at redhat dot com> ---
(In reply to Jakub Jelinek from comment #14)
> (In reply to dhowells@redhat.com from comment #13)
> ...
> Ugh, no.  Why not just x && (x & -x) == x ? __builtin_ctz (x) : -1
> (or ctzl or ctzll depending on what the type is)?

Comparing the kernel's ilog2() to your suggestion:

        int kernel_ilog2(unsigned long x)
        {
                return ilog2(x);
        }

        int jakub_ilog2(unsigned long x)
        {
                return x && (x & -x) == x ? __builtin_ctz (x) : -1;
        }

compiled with -Os for x86_64 gives:

0000000000000000 <kernel_ilog2>:
   0:   83 c8 ff                or     $0xffffffff,%eax
   3:   48 0f bd c7             bsr    %rdi,%rax
   7:   c3                      retq   

0000000000000008 <jakub_ilog2>:
   8:   83 c8 ff                or     $0xffffffff,%eax
   b:   48 85 ff                test   %rdi,%rdi
   e:   74 17                   je     27 <jakub_ilog2+0x1f>
  10:   48 89 fa                mov    %rdi,%rdx
  13:   0f bc c7                bsf    %edi,%eax
  16:   48 f7 da                neg    %rdx
  19:   48 21 fa                and    %rdi,%rdx
  1c:   48 39 d7                cmp    %rdx,%rdi
  1f:   ba ff ff ff ff          mov    $0xffffffff,%edx
  24:   0f 45 c2                cmovne %edx,%eax
  27:   c3                      retq   

Note that in the kernel variant, the initial OR instruction is superfluous, but
is required by fls() and fls64() which x86_64 is using behind the scenes.  Not
all arches, however, use fls() and fls64() to implement ilog2().