[FFmpeg-devel] [RFC]Move is-align-a-power-of-two test up

[Reimar Döffinger]

On Sat, Apr 03, 2010 at 05:42:43PM +0100, M?ns Rullg?rd wrote:
> Carl Eugen Hoyos <cehoyos at rainbow.studorg.tuwien.ac.at> writes:
> > FFmpeg compilation with icc's interprocedural optimization between
> > files (-ipo) improves performance by 10% (h264) to 20% (theora).
> > (Only tested with --enable-shared.)
> > Unfortunately, it currently does not work out-of-the-box, because
> > compilation with -ipo does not fail if an invalid value for .align is
> > given ("linking" fails):
> >
> > $ cat align.c
> > void foo(void){ __asm__ volatile(".align 3"); }
> > $ icc align.c -c
> > /tmp/iccgyMZFzas_.s: Assembler messages:
> > /tmp/iccgyMZFzas_.s:13: Error: alignment not a power of 2
> > $ icc align.c -c -ipo
> > (succeeds and builds align.o)
> > $ icc align.o
> > ipo: remark #11001: performing single-file optimizations
> > ipo: remark #11005: generating object file /tmp/ipo_iccWkeAyV.o
> > /tmp/iccTgD2V5as_.s: Assembler messages:
> > /tmp/iccTgD2V5as_.s:13: Error: alignment not a power of 2
> > icc: error #10014: problem during multi-file optimization compilation
> > (code 1)
> 
> Do you have any idea why this is so?

My guess: In this mode, icc only generates some intermediate code but no
actual object code and does the real optimization, converting to asm
and then to object code at the linking stage.