[FFmpeg-devel] [PATCH] x86/intmath: allow the source operand for icc/icl ff_ctzll to be a memory location

[Matt Oliver]

On 25 October 2015 at 12:22, Ganesh Ajjanagadde <gajjanag at mit.edu> wrote:

> On Sat, Oct 24, 2015 at 7:03 PM, James Almer <jamrial at gmail.com> wrote:
> > On 10/24/2015 7:48 PM, Ganesh Ajjanagadde wrote:
> >> On Sat, Oct 24, 2015 at 6:08 PM, James Almer <jamrial at gmail.com> wrote:
> >>> This gives the compiler some flexibility
> >>>
> >>> Signed-off-by: James Almer <jamrial at gmail.com>
> >>> ---
> >>>  libavutil/x86/intmath.h | 2 +-
> >>>  1 file changed, 1 insertion(+), 1 deletion(-)
> >>>
> >>> diff --git a/libavutil/x86/intmath.h b/libavutil/x86/intmath.h
> >>> index 7881e3c..10d3e7f 100644
> >>> --- a/libavutil/x86/intmath.h
> >>> +++ b/libavutil/x86/intmath.h
> >>> @@ -36,7 +36,7 @@ static av_always_inline av_const int
> ff_ctzll_x86(long long v)
> >>>  {
> >>>  #   if ARCH_X86_64
> >>>      uint64_t c;
> >>> -    __asm__("bsfq %1,%0" : "=r" (c) : "r" (v));
> >>> +    __asm__ ("bsfq %1,%0" : "=r" (c) : "rm" (v));
> >>>      return c;
> >>>  #   else
> >>>      return ((uint32_t)v == 0) ? _bit_scan_forward((uint32_t)(v >>
> 32)) + 32 : _bit_scan_forward((uint32_t)v);
> >>> --
> >>> 2.6.2
> >>
> >> This question may be silly, but can you clarify when this asm code is
> >> used instead of __builtin_ctzll? For instance, I think on GNU/Linux,
> >> x86-64, sufficiently modern GCC (even with asm enabled), we should
> >> always use the builtin: the compiler knows best what to do with its
> >> builtin.
> >
> > This is ICC/ICL, not GCC.
>
> Ah, now I recall that _bit_scan_forward64 was not always available on
> ICC. Anyway, this is just a heads up to whoever uses ICC/ICL and the
> like: you may want to find out to which versions this built in
> applies.


bit_scan_forward64 isnt available on ICL at all, hence the use of asm.

Patch works with ICL fine, LGTM.