There is one other issue with the VPATH solution which I just realized
while trying to make changes to the original patch: the order-only
dependencies don't work with VPATH because the directories already exist
the source directory, so we'd need to use the solution with .dirstamps.

On Thu, Nov 12, 2015 at 6:50 AM Szabolcs Nagy <nsz@port70.net> wrote:

> * Petr Hosek <phosek@chromium.org> [2015-11-12 00:28:44 +0000]:
> > On Wed, Nov 11, 2015 at 3:09 PM Rich Felker <dalias@libc.org> wrote:
> > > On Wed, Nov 11, 2015 at 10:02:50PM +0000, Petr Hosek wrote:
> > > >  MEMOPS_SRCS = src/string/memcpy.c src/string/memmove.c
> > > src/string/memcmp.c src/string/memset.c
> > > >  $(MEMOPS_SRCS:%.c=%.o) $(MEMOPS_SRCS:%.c=%.lo): CFLAGS_ALL +=
> > > $(CFLAGS_MEMOPS)
> > > > @@ -119,34 +144,35 @@ $(CRT_LIBS:lib/%=crt/%): CFLAGS_ALL += -DCRT
> > > >  # force the corresponding object file to be rebuilt, even if the
> > > implicit
> > > >  # rule below goes indirectly through a .sub file.
> > > >  define mkasmdep
> > > > -$(dir $(patsubst %/,%,$(dir $(1))))$(notdir $(1:.s=.o)): $(1)
> > > > +$(dir $(patsubst $(srcdir)/%/,%,$(dir $(1))))$(notdir $(1:.s=.o)):
> $(1)
> > > > +$(dir $(patsubst $(srcdir)/%/,%,$(dir $(1))))$(notdir $(1:.s=.lo)):
> $(1)
> > > >  endef
> > > > -$(foreach s,$(wildcard src/*/$(ARCH)*/*.s),$(eval $(call
> > > mkasmdep,$(s))))
> > > > +$(foreach s,$(wildcard $(srcdir)/src/*/$(ARCH)*/*.s),$(eval $(call
> > > mkasmdep,$(s))))
> > >
> > > Was this missing the dep rules for .lo files?
> > >
> >
> > I'm not sure how this worked before, I was getting build errors without
> the
> > explicit dep rule for .lo files.
> >
> >
> > > > -%.o: $(ARCH)$(ASMSUBARCH)/%.sub
> > > > +$(OBJS_SUB): %.o:
> > > >       $(CC) $(CFLAGS_ALL_STATIC) -c -o $@ $(dir $<)$(shell cat $<)
> > > >
> > > > -%.o: $(ARCH)/%.s
> > > > +$(OBJS_S): %.o:
> > > >       $(AS_CMD) $(CFLAGS_ALL_STATIC)
> > > >
> > > > -%.o: %.c $(GENH) $(IMPH)
> > > > +$(OBJS_C): %.o: $(srcdir)/%.c $(GENH) $(IMPH)
> > > >       $(CC) $(CFLAGS_ALL_STATIC) -c -o $@ $<
> > > >
> > > > -%.lo: $(ARCH)$(ASMSUBARCH)/%.sub
> > > > +$(LOBJS_SUB): %.lo:
> > > >       $(CC) $(CFLAGS_ALL_SHARED) -c -o $@ $(dir $<)$(shell cat $<)
> > > >
> > > > -%.lo: $(ARCH)/%.s
> > > > +$(LOBJS_S): %.lo:
> > > >       $(AS_CMD) $(CFLAGS_ALL_SHARED)
> > > >
> > > > -%.lo: %.c $(GENH) $(IMPH)
> > > > +$(LOBJS_C): %.lo: $(srcdir)/%.c $(GENH) $(IMPH)
> > > >       $(CC) $(CFLAGS_ALL_SHARED) -c -o $@ $<
> > >
> > > I don't understand these rules with two :'s. I assume it's some trick
> > > I don't yet know. But in the case of the %.s ones, the new rules have
> > > no %.s in them... this looks wrong, no?
> > >
> >
> > This is bit of hack which uses the static rules. I think we might be able
> > to get rid of those if we remove the *.sub system as mentioned earlier,
> but
> > this is the only way I got it to work without having a per file rule in
> the
> > current setup.
>
> i'm fine with VPATH + a make time check that in/out is not mixed
>
> this second approach is ok if there is some solution for the
> %.o: %.s rules (i guess AS_CMD only works without %.s because
> mkasmdeps adds the deps).
>
> the .sub approach is not flexible and i don't see an easy way to
> encode all configury bits in the directory tree
> (e.g. if we ever want to allow musl to be compiled for sse4,
> avx, or other more finegrained arch extensions.. maybe .S
> with #ifdefs for subarch is more flexible)
>