[Ffmpeg-cvslog] r6729 - trunk/configure

[Måns Rullgård]

Guillaume POIRIER said:
> Hi,
>
> On 10/18/06, M?ns Rullg?rd <mru at inprovide.com> wrote:
>>
>> gpoirier said:
>> > Author: gpoirier
>> > Date: Wed Oct 18 14:33:42 2006
>> > New Revision: 6729
>> >
>> > Modified:
>> >    trunk/configure
>> >
>> > Log:
>> > rename --tune to --cpu and make --cpu pass the apropriate -march=xx option to
>> the
>> > compiler.
>> > Note that previous implementation (--tune) wasn't consistent with regards to
>> > setting -march/-mcpu/-mtune whereas current --cpu now is
>>
>> This is wrong.  GCC is inconsistent between architectures about the names of
>> those
>> flags.  This change is likely to break builds on all but x86 machines.
>
> Let me see. As far as I can see, x86 is the only arch for which -march
> is documented as setting -mpcu or -mtune too.
>
> On other arches, it's not documented as such (but maybe the behaviour
> is the same, who knows?)
>
> M?ns, if you haven't already, please read the patch and tell me if it
> does what it's supposed to:
>   --cpu sets the base intruction emitting scheme (-march);
>   --tune (the one that is to come when I write it, not the old --tune)
> sets the scheduling heuristic and other things that just make it run
> better on the selected target. This is supposed to be controlled by
> -mcpu or -mtune depending on GCC version (prior 3.4 uses -mcpu, the
> newer uses -mtune).
>
> So as far as I can see, unless you give me further details in what is
> wrong with the way --cpu is implemented, my implementation seems
> correct to me...

What you say is true for x86, but not most other arches.  On some -mcpu
does more or less what -march does on x86, and -march doesn't exist (error).
Some other arches have both -mcpu and -march, but with different meanings
than x86.  The same goes for -mtune.  Some arches have it, others don't, and
where it does exist, it means different things.  The GCC info manual has
all the details.

-- 
M?ns Rullg?rd
mru at inprovide.com