Bug 16383

image_types_wic.bbclass gates syslinux-native on the build host rather
than on the target, so an x86 image built on a non-x86 host is given the
target bootloader without the installer wic needs in order to use it.

This is the other half of bug 13276:
https://bugzilla.yoctoproject.org/show_bug.cgi?id=13276

That bug reported builds failing on non-x86 hosts, and it was correctly
fixed for the case reported. Filing separately rather than reopening,
because the reported symptom really was resolved; what follows is a
different case that the fix left behind.

Every table below assumes an image that uses wic, and shows whether that
image ends up depending on syslinux-native.

1. As bug 13276 was reported. WKS_FILE_DEPENDS_DEFAULT listed
   syslinux-native unconditionally, and the syslinux recipe set
   COMPATIBLE_HOST to x86 for every variant, native included:

       build host | target  | syslinux-native | result
       -----------+---------+-----------------+---------------------
       x86        | x86     | yes             | works
       x86        | non-x86 | yes             | works, but wasteful
       non-x86    | x86     | yes             | BUILD FAILS
       non-x86    | non-x86 | yes             | BUILD FAILS  <- reported

2. After the fix that closed bug 13276, which gated the dependency on
   BUILD_ARCH:

   https://git.openembedded.org/openembedded-core/commit/?id=7e2ee2b59319

       build host | target  | syslinux-native | result
       -----------+---------+-----------------+---------------------
       x86        | x86     | yes             | works
       x86        | non-x86 | yes             | works, but wasteful
       non-x86    | x86     | no              | installer missing
       non-x86    | non-x86 | no              | works  <- reported case fixed

   The reported failure is gone. But the dependency was dropped from
   both non-x86-host rows, including the one that legitimately needs
   syslinux-native, so an x86 image built on a non-x86 host now gets the
   target bootloader without the installer wic has to run to use it.

   Richard made exactly this point in comment 1 on that bug: "The
   dependency would still seem inaccurate and should probably only be
   added if the target machine is x86?" The patch that landed gated on
   the build host instead.

3. After the syslinux recipe was reworked so that only its target code
   is x86-specific:

   https://git.openembedded.org/openembedded-core/commit/?id=7273e131bfc7

       build host | target  | syslinux-native | result
       -----------+---------+-----------------+----------------------------
       x86        | x86     | yes             | works
       x86        | non-x86 | yes             | works, but wasteful
       non-x86    | x86     | no              | installer missing, and now
                  |         |                 | needlessly so
       non-x86    | non-x86 | no              | works

   Nothing moved, and that is the point: the constraint that justified
   gating on the build host disappeared in 2020. That commit message
   even names the case it was enabling, "arm64 build host, qemux86
   target", which is exactly the row that is still broken.

4. Gating on the target instead, the way wic-tools has always done it:

       build host | target  | syslinux-native | result
       -----------+---------+-----------------+---------------------
       x86        | x86     | yes             | works
       x86        | non-x86 | no              | waste removed
       non-x86    | x86     | yes             | works
       non-x86    | non-x86 | no              | works  <- still fixed

The bottom-right row is the one bug 13276 reported, and it never regains
the dependency: a target gate excludes a non-x86 target whatever the
build host is. The build-host gate was never load-bearing for the
reported case.

The only row that changes behaviour is non-x86 host with an x86 target,
which starts building syslinux-native where it previously skipped it. I
do not have a non-x86 build host to confirm that, so a test from someone
who does would be very welcome. Three things suggest it is safe:
7273e131bfc7 was written specifically to enable it, its author was at
Arm at the time, and wic-tools already gates syslinux-native on the
target, so "bitbake wic-tools" on an arm64 host with an x86 target
exercises the same path today. If it does turn out to fail, that would
be a syslinux recipe issue -- it would need a COMPATIBLE_HOST for
native, contradicting the intent of 7273e131bfc7 -- rather than a wic
one.

A patch is included in a wic series about to be posted to
openembedded-core.