On Tue, Mar 18, 2014 at 10:42:01 +0000, Daniel Berrange wrote:
On Tue, Mar 18, 2014 at 11:36:51AM +0100, Jiri Denemark wrote:
> On Mon, Mar 17, 2014 at 09:33:11 -0600, Eric Blake wrote:
> > On 03/14/2014 12:55 PM, Jiri Denemark wrote:
> > > On Fri, Mar 14, 2014 at 07:54:58 -0600, Eric Blake wrote:
> > >> On 03/14/2014 04:43 AM, Jiri Denemark wrote:
> > >>> Ancient automake (such as from RHEL5) does not provide abs_srcdir
and
> > >>> abs_builddir variables which are used by a recent commit of mine
> > >>> (e562e82).
> > >>>
> > >>> Signed-off-by: Jiri Denemark <jdenemar(a)redhat.com>
> > >>> ---
> >
> > >>> +# old automake does not provide abs_{src,build}dir variables
> > >>> +abs_builddir = $(shell pwd)
> > >>> +abs_srcdir = $(shell cd $(srcdir) && pwd)
> >
> > Hmm, just noticed another thing - with NEWER automake, we are now less
> > efficient (calling out to $(shell) overwrites the value that is already
> > provided for free without a subprocess by newer automake). Does it work
> > if you use:
> >
> > abs_builddir ?= $(shell pwd)
>
> Automake does a fair amount of magic here so abs_builddir = $(shel ...)
> does not actually overwrite the old value because there is no old value
> :-) Automake just does not put it's own abs_builddir definition in that
> case. If I switch to abs_builddir ?= $(shell ...), automake adds it's
> own definition so it seems it could work except that it doesn't. With
> "=", automake adds the definition above the other variable definition
> and namely above libvirt_cpu_la_DEPENDENCIES which uses it:
>
> abs_builddir = $(shell ...)
> ...
> libvirt_cpu_la_DEPENDENCIES = $(abs_builddir)/...
> ...
> rules
>
> When I switch to "?=", automake apparently does not recognize it as a
> variable definition and puts it between the block with variable
> definitions and rules. Thus the result is:
>
> abs_builddir = /some/build/path
> ...
> libvirt_cpu_la_DEPENDENCIES = $(abs_builddir)/...
> ...
> abs_builddir ?= $(shell ...)
> ...
> rules
>
> which of course does not work with old automake which does not add the
> first abs_builddir definition there and thus libvirt_cpu_la_DEPENDENCIES
> sees $(abs_builddir) empty.
>
> Anyway, making two extra shell commands once per build does not seem
> like anything we should really care about so I'd just stick with
>
> abs_builddir = $(shell ...)
Why don't we just avoid the whole issue by removing use of abs_srcdir
and abs_builddir. Can this rule:
$(abs_builddir)/cpu/cpu_map.xml:
$(AM_V_GEN)ln -s $(abs_srcdir)/cpu/cpu_map.xml $@
be just changed to
cpu/cpu_map.xml:
$(AM_V_GEN)ln -s $(srcdir)/cpu/cpu_map.xml $@
That's what I tried first but it does not work at all. I don't
understand why but make thinks cpu/cpu_map.xml target is uptodate even
though the file does not exist in builddir. And $(srcdir) doesn't work
for relative VPATH. For example, if VPATH is .., then
abs_srcdir = /some/path/src
abs_builddir = /some/path/build/src
srcdir = ../../src
and the /some/path/build/src/cpu/cpu_map.xml link will be
../../src/cpu/cpu_map.xml and thus will point to itself. And we can't
just blindly do
ln -s ../$(srcdir)/cpu/cpu_map.xml $@
because this would not work for absolute VPATH, when srcdir is
/some/path/src.
Yeah, it's a mess.
Jirka