Rework some multi-output target dependency handling.

This reworks my last commit in r301285 to more closely match what was in
r241298 (but reverted in r294878).

This is addressing "missing .meta file" rebuilds but also ensuring that
files are always generated when needed in each case.

Note that this is not a complete rework of the problem areas identified
in r301285 as most are "good enough" right now as the new pattern
is too verbose. It's only worth making this current change where headers
may be generated in the INCS list; where missing .meta file rebuilds are
spotted.

--- Technical details follow ---

Several attempts to deal with this problem of multi-output targets, with and
without META MODE, were explained in r241298, r294878, and r301285.

The general problem is with multi-output targets such as:
        foo.c foo.h:
                touch foo.c foo.h
        foo.c foo.h:
                touch foo.c
                touch foo.h
        foo.c foo.h: foo.in
                ./generator ${.ALLSRC}

This pattern is problematic in jobs mode as both files end up being
built concurrently and leads to races. With META MODE it is worse
as both targets end up rebuilding if they lack a .meta file. So the
generator is force built twice even though it is only needed once.
There are also problems in that 'make foo.h' may be ran before 'make foo.c';
The order of make generating the targets is not guaranteed.

An older attempted workaround to this (discussed in r294878) was:
        foo.h: foo.c
        foo.c: foo.in
                ./generator ${.ALLSRC}
This appears fine except that if foo.h is missing and foo.c exists then
foo.h will never be regenerated. This pattern is close to the solution
in this commit though:

        foo.h: foo.c .NOMETA
        .if !exists(foo.h)
        foo.c: .PHONY .META
        .endif
        foo.c: foo.in
                ./generator ${.ALLSRC}

There's 2 differences here:
1. foo.h will never expect to have a .meta file since the foo.c target
   will generate both and own the .meta file.
2. If foo.h does not exist then it needs to force foo.c to be rebuilt
   with .PHONY. That normally disables META MODE though so .META is
   given to tell bmake we do really expect a .meta file.

This pattern cannot work with implicit suffix rules since the .c and .h files
may be generated at different times (buildincludes vs depend/all).

Sponsored by:	Dell EMC
MFC after:	2 weeks

1 files changed, 8 insertions, 3 deletions

diff --git a/share/mk/bsd.snmpmod.mk b/share/mk/bsd.snmpmod.mk
index c45937cc82a7..6ecbbcf09bf7 100644
--- a/share/mk/bsd.snmpmod.mk
+++ b/share/mk/bsd.snmpmod.mk
@@ -12,9 +12,14 @@ GENSNMPTREEFLAGS+=	-I${SHAREDIR}/snmpdefs
 ${MOD}_oid.h: ${MOD}_tree.def ${EXTRAMIBDEFS} ${EXTRAMIBSYMS}
 	cat ${.ALLSRC} | gensnmptree ${GENSNMPTREEFLAGS} -e ${XSYM} > ${.TARGET}
 
-.ORDER: ${MOD}_tree.c ${MOD}_tree.h
-${MOD}_tree.h: .NOMETA
-${MOD}_tree.c ${MOD}_tree.h: ${MOD}_tree.def ${EXTRAMIBDEFS}
+# Multi-output targets both expect a .meta file and will fight over it. Only
+# allow it on the .c file instead.
+${MOD}_tree.h: ${MOD}_tree.c .NOMETA
+# Force rebuild the .c file if any of its other outputs are missing.
+.if !exists(${MOD}_tree.h)
+${MOD}_tree.c: .PHONY .META
+.endif
+${MOD}_tree.c: ${MOD}_tree.def ${EXTRAMIBDEFS}
 	cat ${.ALLSRC} | gensnmptree -f ${GENSNMPTREEFLAGS} -p ${MOD}_
 
 .if defined(DEFS)