Re: Patch for Makefile race against current cvs

Tom Lane writes:

> What bothers me is that you seem to be saying that *any* construct
> involving multiple outputs from one rule is unsafe in a parallel make.
> That strikes me as a huge restriction, and one that would surely be
> mentioned prominently in the gmake manual if it were real.  But I can't
> find anything that says that.

I think we have been victims of a misunderstanding of how multiple-target
rules work.  The gmake manual says:

"A rule with multiple targets is equivalent to writing many rules, each
with one target, and all identical aside from that."

So when we write this:

| $(srcdir)/preproc.c $(srcdir)/preproc.h: preproc.y
|         $(YACC) -d $(YFLAGS) $<
|         mv y.tab.c $(srcdir)/preproc.c
|         mv y.tab.h $(srcdir)/preproc.h

make reads it as this:

| $(srcdir)/preproc.c: preproc.y
|         $(YACC) -d $(YFLAGS) $<
|         mv y.tab.c $(srcdir)/preproc.c
|         mv y.tab.h $(srcdir)/preproc.h
|
| $(srcdir)/preproc.h: preproc.y
|         $(YACC) -d $(YFLAGS) $<
|         mv y.tab.c $(srcdir)/preproc.c
|         mv y.tab.h $(srcdir)/preproc.h

So it's completely conceivable that both rules get executed in parallel.
(In general, you cannot optimize this away, because the actual commands
may differ depending on the value of $@.)

This basically means that you cannot correctly write a target that
generates more than one file.  The only workaround I see is to cheat about
the dependencies, as Klaus' patch does:

y.tab.c: preproc.y
        $(YACC) -d $(YFLAGS) $<

$(srcdir)/preproc.c: y.tab.c
        mv y.tab.c $(srcdir)/preproc.c

$(srcdir)/preproc.h: y.tab.c
        mv y.tab.h $(srcdir)/preproc.h

It's the latter rule where we're technically cheating and which I was
initially regarding as part of the "details I don't like", but now I think
it's the best (only?) way to go.

--
Peter Eisentraut   peter_e@gmx.net