-j 2 exposes problem with pattern rule

[John Breitenbach]

Hi,
 I'm hoping for an explanation for some strange behavior that I see 
when attempting a parallel build.  Its
not the classic race condition, but rather something wrong with my 
pattern rule.  Hopefully this will
be an obvious problem with a nice solution.

My pattern rule reads like this:

lf/% af/% nf/% pf/%:
   do_stuff()

My intent is to invoke the same function for any make target with one of 
the 4 subdirs listed above.
There are dependencies between them.  e.g., lf/abc depends on af/abc and 
nf/aaa.  af/abc
depends on pf/ppp while nf/aaa depends on pf/p1...  Thus there are 
opportunities for parallel.
But the "notdir" part of the targets DO repeat in af and lf.

This works fine with -j 1 builds - everything builds in order as 
needed.  However, I run into a
problem with -j 2.  I think there is a problem when I have lf/abc 
dependent on af/abc using the
pattern rule above.  make builds af/abc in parallel with nf/aaa.  af/abc 
finishes, and then nf/aaa
finishes and make thinks its done!  It never attempts to build lf/abc 
... again only with multiple jobs.

Here's the last bit of make's -d output - after af/abc finishes.

Reaping winning child 0x09b975f8 PID 27863
Removing child 0x09b975f8 PID 27863 from chain.
Released token for child 0x09b975f8 (af/abc).
Live child 0x09b968e8 (nf/aaa) PID 22912
Considering target file `lf/abc'.
File `lf/abc' was considered already.
Live child 0x09b968e8 (nf/aaa) PID 22912


Thanks for reading.