Re: [gentoo-user] Re: Cannot compile texinfo: Illegal instruction -> Wrong -march and -mtune flags?

[Florian Philipp]

Am Freitag 03 August 2007 03:06 schrieb Daniel da Veiga:
> On 8/2/07, Florian Philipp <f.philipp@××××××.de> wrote:
> > Am Donnerstag 02 August 2007 23:36 schrieb Alexander Skwar:
> > > · Florian Philipp <f.philipp@××××××.de>:
> > > > You see, they are not compatible and even if some code works I
> > > > wouldn't bet multimedia apps will perform well.
> > > >
> > > > With -mtune the instruction set stays the same. It is just
> > > > "rearranged".
> > >
> > > Hm. Allright. When using just -mtune (ie. without -march), the
> > > docs at
> > > http://gcc.gnu.org/onlinedocs/gcc-4.1.2/gcc/i386-and-x86_002d64-Options
> > >.htm l
> > >
> > > say:
> > > | While picking a specific cpu-type will schedule things appropriately
> > > | for that particular chip, the compiler will not generate any code
> > > | that does not run on the i386 without the -march=cpu-type option
> > > | being used.
> > >
> > > If -mtune=athlon-xp is used, code is generated which may make
> > > use of 3dNOW!. 3dNOW! is, of course, not to be found on 386 :)
> > > If the instruction set stays the same, code generated with
> > > -mtune=athlon-xp would not be executable on 386 machines, if
> > > I understand you correctly.
> > >
> > > Hm. With -mtune, the set of available instructions (ie.
> > > stuff like 3dNOW!, I suppose?) is NOT changed from the default
> > > of i386, is it? Or what does "Tune to cpu-type everything applicable
> > > about the generated code, except for the ABI and the set of available
> > > instructions." mean - especially note the "except for [...] the set of
> > > available instructions" part.
> > >
> > > So with "-mtune=pentium-m -march=athlon-xp" I'm making the compiler
> > > generate code which is "ordered" the way it's best for pentium-m
> > > machines while allowing it to use athlon-xp instruction set? Is
> > > that what I'm doing?
> > >
> > > If so, then it seems you're right - code will run, but maybe not
> > > so well.
> > >
> > > Is that understanding correct? If so, then I really should think
> > > twice about using "-mtune=pentium-m -march=athlon-xp", shouldn't
> > > I?
> > >
> > > Curious,
> > >
> > > Alexander Skwar
> > > --
> >
> > At least that's how I understand the issue. At the moment I've got two
> > ideas to solve your problem:
> >
> > 1. set march to an inferior target (pentium-3 and pentium-3m seem okay:
> > mmx and sse) and mtune for one (or even both?) of them
> > 2. set march to one of them and disable incompatible instruction sets
> > with options like -mno-sse2 or -mno-3dnow
>
> Isn't the -march=i686 valid?
> I guess that would be the most "compatible" option for binaries that
> will run on AMD and Intel processors... Or simply use no "-march"
> setting, only "-mtune"... I did that recently to switch a whole system
> from an Athlon XP to a Intel Core Duo...
>
-march=i686 -mmmx -msse would be the better choice, then.

I'd still like to know what gcc does when you enable two mtune settings. Use 
the last one? Use both?