%{_target_cpu} value on i686 differes between 4.20 and 5.9*

[pmatilai]

Sometime between mid-December and mid-January my snapshot builds started failing on i386: https://copr.fedorainfracloud.org/coprs/pmatilai/rpm-snapshot/builds/ with the following:

      stderr: 
        /usr/bin/pkg-config: line 8: /usr/bin/i386-redhat-linux-gnu-pkg-config: No such file or directory

This comes from pkg-config which on Fedora calls rpm (🤪 ) macro evaluation to determine the path of the actual script to call - in this case the actual file is /usr/bin/i686-redhat-linux-gnu-pkg-config. The relevant macro is %{_target_cpu} so that's what's now evaluating differently, and the likely cause being the rpmrc STL refactoring I did in the first half of December.

Comparing between two mock shells suggests it's probably buildarch translation that's misbehaving:

sh-5.2# rpm --version;rpm --showrc|head -2
RPM version 4.20.0
ARCHITECTURE AND OS:
build arch : i686

vs

sh-5.2# rpm --version;rpm --showrc|head -3
RPM version 5.99.90
ARCHITECTURE AND OS:
build arch : i386
compatible build archs: i686 i586 i486 i386 noarch fat

Most likely this is a regression, but the rpmrc stuff being so old and strange, it's possible something was actually misbehaving earlier. Need to sort this out before 6.0 alpha though.

[pmatilai]

Bisected down to 7be4b79.

The reason: what I didn't realize is that there can be multiple key:value pairs in the source data, coming from different places (user, host, vendor etc), but only the last one counts so order matters. And the code uses insert() on the map, causing the first encountered value to be used.

Phew, for a moment I though this might be something nastier.