VxWorks specs vs fixincl, machname.h and machine_name_test

[Earl Chew <earl_chew at agilent dot com>]

I'm building a vxWorks PPC cross using gcc 3.4.2 and I'm trying
to understand how the machname fix is intended to work.

I've provided the compiler with a set of stock vxWorks headers.

Selecting --target=powerpc-wrs-vxworks and
--with-headers=/gnu/tornado/target/h causes config/rs6000/vxworks.h
to be used, and it contains lines like:

#define CPP_SPEC \
"-DCPU_FAMILY=PPC -D__ppc -D__EABI__  \
 %{t403: -DCPU=PPC403 -D_SOFT_FLOAT ; \
   t405: -DCPU=PPC405 -D_SOFT_FLOAT ; \
   t440: -DCPU=PPC440 -D_SOFT_FLOAT ; \
   t603: -DCPU=PPC603               ; \
   t604: -DCPU=PPC604               ; \
   t860: -DCPU=PPC860 -D_SOFT_FLOAT ; \
       : -DCPU=PPC604}  \
 %{!msoft-float:-D__hardfp}	   \
 %{fpic|fpie: -D__PIC__=1 -D__pic__=1 ; \
   fPIC|fPIE: -D__PIC__=2 -D__pic__=2 } \
 %(cpp_cpu)"

So far so good.

I get to end of the compiler build, then past "Fixing headers", and
onto the build of libgcc.

Here's the problem:

o Fixing headers applies the machine_name_test fix and converts
  occurrences of CPU to __CPU__.

o The specs used by xgcc still say CPU=xyz, so the compile of libgcc
  terminates in confusion as the headers have been rewritten to
  look for __CPU__.

It looks like fixincl has modified to the headers to a form that
is simply unworkable with the specs.

Can anyone enlighten as to what's supposed to happen? Is fixincl
doing the wrong thing, or is the specs file wrong?

Earl