fedora for ARM

[Lennert Buytenhek]

On Thu, Jun 07, 2007 at 11:26:47AM -0600, Brendan Conoboy wrote:

> > Are you suggesting to use the target triple as the rpm arch name?
> >
> > Isn't gcc-4.1.2-12.armv5l-softvfp-linux-gnueabi.rpm kind of long?
> 
> Yes, it's much too long, but it is a starting point should Fedora
> deem armv5l-softvfp-linux-gnueabi an important variant :-)

The current package set _is_ basically that, we just call it
'armv5tel' since the rest doesn't add much.  Especially the
linux-gnu part.  :-)


> > (Note that our package repo uses eabi, where vfp byte/word order
> > and softfloat are the default, so we don't encode those bits of
> > info explicitly.  The packages use 'armv5tel' as the rpm arch name.)
> 
> I like armv5l and armv5b arch names, but this seems like a
> relatively minor point.

Well, armv5l implies that the CPU doesn't support the "T" or "E"
extensions.  This is an important distinction, since even though a
package compiled with -march=armv5te will probably not automagically
contain thumb insns, it _is_ allowed for gcc to emit ldrd (load
double word) and strd (guess..) instructions when the "E" extension
is specified to be present.

I know of at least one theoretically armv5te compliant ARM CPU where
ldrd/strd are buggy, and where the workaround is to make the kernel
report 'armv5t' to userspace instead of 'armv5te', and to build
userland with -march=armv5t instead of -march=armv5te.

I.e. you can't compile a package with -march=armv5te and then call
it 'armv5l', because people will then try running it on armv5l and
armv5tl machines, which might be less than successful.