Dennis Wassel wrote: > What strikes me about this is the fact that ATLAS' configure (all by > itself, no options given) thinks that using -m64 on a 32-bit system > might be a good idea (not to actually link something executable, > probably just some low-level assembler-ish stuff, no clue), since I'm > using a 64-bit compiler - am I up to something here, or do you think > it's just One Of Those Things? It seems common for numeric computation libraries to have configure checks to detect the specific processor variant so that they will be built with the best optimization, cf. gmp/mpfr. The problem is that they tend to do this without regard for the flavor of operating system. For example I seem to recall a recent PR in which gmp was configured on e.g. sparc-sun-solaris2.10 but it detected that the machine was really sparc64-* and built binaries of the wrong flavor, leading to an error in the gcc build when it tried to link the stage 1 xgcc. I think in all these cases you can simply spell out the --host= explicitly to override the autodetection logic. Brian
