RE: CPU specific binaries
All this comes out of thin air, but here goes:
* I think that currently the goal is to support the highest
optimization that does not hurt (noticeably) the low end
machines. I think this includes Pentium optimizations,
but naturally none of the instruction set extensions.
Optimization for PII or higher is probably not ok.
* Any optimization should not introduce compile time errors.
Some may result in execution errors. Most of these are
illegal instructions. Some may be legal sequences of
instructions that just give a different result. Intel
has tried hard to avoid these, however, so it may be that
there are no such things.
* There should be a set of optimizations that is always safe
to apply in any x86 architecture machine. The larger we
get this set and the better we ensure that all Debian
packages use it the better our base line is.
* The next thing would be to define a standard way to enable
some set of processor model specific optimizations. This
is really also a prerequisite for parallel binary model-
optimized distributions. It has to use auotobuild to work.
t.aa
Reply to: