[FFmpeg-devel] [FFmpeg-cvslog] r12171 - trunk/doc/optimization.txt
Thu Feb 21 21:46:38 CET 2008
On Thu, 2008-02-21 at 13:26 -0700, Loren Merritt wrote:
> On Thu, 21 Feb 2008, Uoti Urpala wrote:
> > On Thu, 2008-02-21 at 20:11 +0100, Michael Niedermayer wrote:
> >> As you can see gcc injects 2 unneeded lea instructions in the innermost loop.
> > It'd need at least one other instruction instead though. Anyway I do not
> > see that behavior with Debian's gcc-4.3, gcc always uses 3 instructions
> > for the adds (same as the handwritten asm).
> It could use (%rsi,%r8) directly in the movq. No extra instructions.
gcc could not because the constraint is "r"(s1). If you change the
constraint to "m"(*s1) then gcc does use that form.
More information about the ffmpeg-devel