Re: [Boost-build] architecture/instruction-set

From: Vladimir Prus (ghost_at_[hidden])
Date: 2007-12-01 14:55:03

On Saturday 01 December 2007 21:04:27 Felipe Magno de Almeida wrote:
> Hi Volodya,
> 
> On Dec 1, 2007 3:45 PM, Vladimir Prus <ghost_at_[hidden]> wrote:
> >
> > I was doing some documentation, and it seems to me that
> > architecture/instruction-set are confusingly named, and implemented.
> 
> I don't use boost.build for cross-compiling, so my opinion might be a
> little biased.
> 
> > For start, most compilers distinguish between generating code for
> > a given processor (which won't run on different processors, in general),
> > and tuning code for a given processor. gcc.jam uses <instruction-set>
> > to set -march, which is the hard processor selection. msvc.jam uses
> > <instruction-set> for /flavor -- which is "soft" tuning selection.
> > Should we have another <tune> option?
> 
> I think tune is a nice name, though maybe a little too short.
> How about <cpu-tune>?
> 
> > Also, now 'architecture' is general family (x86, ia64, sparc), and
> > 'instruction-set' is specific processor there. It does not seem to be
> > that 'architecture' is right name here
> 
> I find it a good name. Though sometimes architecture might be
> interpreted with respect to which OS it will run too. Not only the
> CPU.

Well, looking at: http://www.arm.com/products/CPUs/architecture.html

        ... In 2000, the ARMv5TEJ architecture added ...
        ...The ARMv6 architecture, announced in 2001, features...
        ...The ARMv7 architecture lies below the CortexTM family...

So "architecture" is actually used in some rather specific sense -- not
in the highly generic sense used now by Boost.Build.

> > -- after all, how the processor
> > is implemented internally is surely changed radically, several time,
> > during lifetime of x86.
> 
> I don't think it matters how a processor is implemented or
> reimplemented to boost.build naming scheme.

Well, if the established meaning of 'architecture' is different
from what we use, it's confusing.

> 
> > Would a better naming be:
> >
> >         - cpu-family (for general family, keeping backward
> >         compatibility of some sort)
> >         - cpu (for exact cpu to target)
> >
> > Or maybe we should just use 'cpu' with subfeatures for exact cpu?
> > Like
> >
> >         cpu=x86-nocona
> >
> > ?
> 
> How about:
> 
> <architecture>x86
> <cpu-tune>pentium4
> <cpu>pentium3
> ?
> 
> Where <cpu> is the minimum required to run.
> Does it mirror well to other architectures?

Well, using arm as example, I think:

        <architecture>arm <cpu>armv5t

is probably redundant a bit.

- Volodya

-- 
Vladimir Prus
http://vladimir_prus.blogspot.com
Boost.Build V2: http://boost.org/boost-build2