Re: request for configure clarifications on darwin10

[Bob Friesenhahn]

On Wed, 16 Sep 2009, Jack Howarth wrote:
>   The question I have is, in such a case where the detected triplet from
> config.guess reports a radically different architecture (i386-apple-darwin10)
> than the actual default code execution and compiler code generation
> (x86_64-apple-darwin10), might that not be considered a form of
> cross-compilation requiring the user pass at least the --target triplet to
> configure? Currently, many developers on darwin10 seem to be under the
> impression that it is okay to leave configure believing the target is
> i386-apple-darwin10 while the actual code generation is x86_64-apple-darwin10.
> My view is that this is a misuse of configure since it is basically working
> on incorrect assumptions about the architecture.

Autoconf won't think that the build is a cross-build unless 
config.guess output does not match the target output.  By default, 
config.guess output is is used for all triplets so architecture 
details don't matter.

I think it is correct for the architecture type reported by 
config.guess to match the current execution architecture (which may 
not match the compiler).

Fixing config.guess to produce a more accurate architecture will not 
make configure run.  To make configure run by default would require 
enhancements in Autoconf (or a new macro included in all projects) 
which tries to add necessary compiler options (e.g. -m32) so that 
output of the compiler will actually run.  This sort of change would 
take years to implement and propagate.  It is much easier to provide a 
macports build wrapper which sets compilation flags correctly.

It is good to keep in mind that while Apple's compiler could be 
considered the "system" compiler, it is an optional installation only 
usually installed by developers.  It is easily replaced.  Over the 
years config.guess has been used by far more than Autoconf and for 
more than software compilation.  Ideally it should report the same 
results even if the optional compiler is not currently installed.

> config.guess misreporting the architecture in this manner. The best 
> approach would seem to be decoupling the architecture from the 
> kernel type and instead relying on the actual code generation. I 
> believe this is in fact done currently for solaris in 
> config.guess...

The Solaris case is interesting but it is rarely used for anything. 
Last time I tried, GCC was not prepared to build with 64-bit default 
output on Solaris.  64-bit default output can be achieved using a 
small wrapper script.  As implemented by Autoconf, results are not 
very accurate since if CC is set by config.site it is (apparently) too 
late since config.guess has already been executed.  That explains why 
all of my 64-bit Solaris builds still say "i386-pc-solaris2.10".

Bob
--
Bob Friesenhahn
address@hidden, http://www.simplesystems.org/users/bfriesen/
GraphicsMagick Maintainer,    http://www.GraphicsMagick.org/