Re: Multiple file compilation patch

[Jeff Sturm <jsturm at one-point dot com>]

On Sun, 18 Mar 2001, Bryce McKinlay wrote:
> Could we please revert to the old behaviour by default (on the branch,
> at least)?. It seems to me that this mode would be better as an
> option, at least until it works more reliably.

I second that, unless the fix is obvious and quick (it's not obvious at
all to me).

I see three problems with the combined inputs patch right now that don't
appear on the surface to be related:

- multiple class files with -c and without -o are being combined (as I
already mentioned)

- if the command line contains non-source files (i.e. libgcjgc.so) they
are being passed as file lists to jc1 when combining inputs (this is the
cause the "can't open .ELF.." message)

Example: gcj G19990301_01.class foo.class bar.o -o G19990301_01

- command lines like "gcj A.class B.class -o ab" are combining inputs,
which would probably be OK if it weren't exposing bugs in the backend (see
bytecode verification errors in the testsuite).  This is a behavior change
in the driver that is not easy to work around.

I'm strongly in favor of having the combined inputs enhancement in 3.0,
_provided that it's use is optional_.  I was under the impression that it
would only apply to command lines with -c, -o, and multiple sources, which
was nonsensical prior to this patch anyway.

Maybe there could be a flag like -fcombine-inputs that is off by default.

Jeff