>Yes.  I believe that 'make depend' should be a *completely optional* stage
>which figures out 'indirect dependencies'.  For example, if 'foo.o' depends
>on 'foo.c' which includes 'foo.h', 'make depend' should create the file
>that says that 'foo.o' depends on 'foo.h' too.  This is logically separate
>from the issue of building things which are needed to build other things,
>and I think the latter should be done automatically even by a plain 'make'.

Build the source itself --- .c or .h files?  yes.

Re-build the program generators which build the .c or .h files?
No. That way lies madness: how far down do you stop?

If the change is in something like lint (e.g., adding the -d flag)
there's really no good way to encode the dependency on a specific
version, and nowhere to put a lint built against your old libc (which
you need in order to build the new lint libraries to finish building a
new libc).

Same goes for the rest of the "toolchain" -- not just
gcc/as/ld/ranlib/nm/lorder/tsort/size, but make, lex, yacc, config,
rpcgen, and the rest of the usual suspects, too.

Heck, I dont know if we even have a complete list of suspects.