Hans de Goede wrote:
The question is can we make things generic and still set off reasonably optimal code for a wide range of arm systems. I'm not asking for the last 5-10%, but we should be able to get atleast upto 90% with regards to code-size, but also speed of a custom build toolchain for a specific target if our generic libc becomes much much larger then a special one, and cannot be modularized then I'm afraid that having a generic toolchain isn't much good as lots of arm usage is embedded and size often matters there.
Any arm-linux-gnu{,eabi} toolchain is going to support a pretty wide range of arm systems, but not be optimal for any of them. That's not too big a deal if your tools provide a wide range of multilibs and you are willing to set the right optimization flags. It may be worth having a wide range of libc versions (Like ia32 Fedora being "i386", but having some i686 packages where it counts).
You can also get arm-linux-gnueabi tools from here:
ftp://ftp.ges.redhat.com/private/releng/arm-linux-beta
Like Lennert's tools, these are EABI, but completely self contained in a single source rpm. Making ABI tools is a one line change in the spec file.