Re: getting back the quote_ignores_source_dir behavior of -I-

[Daniel Villeneuve <daniel_villeneuve at sympatico dot ca>]

Zack Weinberg wrote:

> I am quite surprised to hear that behavior 2 (ignoring the source file's
> directory) does something useful for you ... it tends to break the
> standard (libc) headers.
>  

I have code that looks like this:

$PWD/foo/a.c         -> #include "bar/file.h"
$PWD/foo/bar/file.h
$PWD/bar/file.h

Our build system calls gcc from $PWD in a way that is similar to:

gcc -I- -I$PWD -I/some/other/dir -o foo/a.o -c foo/a.c

All our #include directives are relative to "stable" points ($PWD, 
/some/other/dir, ...) in a VPATH manner.

Without -I-, there is no way to get $PWD/bar/file.h from the a.c file 
(other than using an absolute path, which is not an option).  Of course, 
I could rename $PWD/foo/bar/file.h, but I expect two "modules" (foo and 
bar here) to be independent w.r.t. internal file naming conventions (and 
of course $PWD/foo/bar/file.h has this name because it relates to 
$PWD/bar/file.h ...).

I've never had problems with standard headers and I've used -I- 
(exclusively for behavior 2) on several platforms (linux, solaris, hpux, 
aix) for several years, so it might not be a problem for "users" of libc 
(by opposition to developers).

> I would prefer that you leave -I- entirely deprecated and add a new
> option, say -ino-file-directory, which has only the effect of setting
> quote_ignores_source_dir.  (As usual when I suggest option names, better
> name suggestions are welcome.)
>  
Noted.

> > With this change, one can use -I- before
> > any -I without any other unwanted side effects (such as losing the
> > result of previous -iquote options).
> >    
>
> This sounds like a bugfix which we should apply independent of your other
> changes, but it would be even better if you appended the BRACKET chain
> to the QUOTE chain if it wasn't empty; thus e.g.
>    -iquote foo -I bar -I- -I quux
> would not lose 'foo' from the QUOTE chain.
>  

That's fine for me.

> Please resubmit the bugfix and the new option as two separate patches
> (with the above suggested change, and test cases if you can hack 'em in
> the dejagnu framework - see gcc.dg/cpp/ for existing examples, and note
> the 'inc' subdirectory of that directory).

I'll look at this later this week.
Thanks for the feedback.
--
Daniel