Re: failure of cross-compilation detection on BlueGene/L

[Steven G. Johnson]

Steffen Dettmer wrote:
> Does this mean you have no way to compile any program running on
> the front-end GNU/Linux?

No, if you ./configure without a --host argument it uses the gcc for the 
front-end (build system) and compiles for the BlueGene front end like 
one would expect.

> You mean, you want to detect BlueGene/L by concluding that from
> the fact that fopen does not work?

No, that's not what I want to do.

If the user specifies --host, autoconf checks whether it is 
cross-compiling by checking whether executables work.  This check gives 
a false negative on BlueGene.  I want to correct the test so that it 
gives the correct result.

> Not that I have a clue about, but for me this sounds really
> strange.  I think, obviously it cannot work, because there are
> platforms that do not even have fopen() or the compiled binaries
> (or its libc) may apply additional checking and may refuse to
> open files (e.g. extra permission checking, or execution in a
> read-only folder or without write cababilities). I think it could
> be bad if on such a platform cross-compiling would be detected.

The configure script cannot run in a read-only folder -- it creates 
files all over the place.  Furthermore, autoconf already assumes that 
fopen works when not cross-compiling (this is assumed in 
AC_CHECK_SIZEOF, for example).  Since I'm not proposing to test for any 
capability that we're not relying on already, it should be safe.

All of this was discussed in my previous email; you probably need to 
re-read it more carefully.

Steven