Name the softfp world so makepkg can build in it

[frangarcj]

No package can be built in the softfp world today. The first attempt was this morning — 32e026c in vitasdk-autobuild put softfp in the try-build matrix — and both packages tried since died the same way before compiling anything:

vita-makepkg: line 381: unset: `provides_vita-softfp': not a valid identifier
==> ERROR: An unknown error has occurred. Exiting...

Why

merge_arch_attrs() in makepkg builds shell variable names out of CARCH:

for attr in "${supported_attrs[@]}"; do
    eval "$attr+=(\"\${${attr}_$CARCH[@]}\")"
done
unset -v "${supported_attrs[@]/%/_$CARCH}"

A hyphen makes that name unusable, and bash does not say so — it reads ${provides_vita-softfp[@]} as ${provides_vita-...}: the variable provides_vita, the - default-value operator, and the word softfp[@].

$ CARCH=vita-softfp; provides=(); eval "provides+=(\"\${provides_$CARCH[@]}\")"
$ printf 'provides=[%s]\n' "${provides[@]}"
provides=[softfp[@]]
$ unset -v "provides_$CARCH"
bash: unset: `provides_vita-softfp': not a valid identifier

So the literal string softfp[@] lands in all seven arrays — provides, conflicts, depends, replaces, optdepends, makedepends, checkdepends — and then the unset fails and makepkg aborts.

The abort is the lucky part. Without it every softfp package would have shipped with a fabricated depends=softfp[@].

pacman assumes architectures are identifiers and never says so: x86_64, aarch64, armv7h, i686 — nobody had reached for a hyphen.

What changes

The world is vita_softfp, and Profiles.cmake now refuses any profile name that would break the same way. The tag, repository and channel names are separate fields and keep their hyphens.

The published softfp core has CARCH="vita-softfp" stamped in bin/makepkg.conf, so it has to be rebuilt — it is a nightly, and nothing has been built against it.

🤖 Generated with Claude Code