binutils 2.47 gives the kernel a bogus AS_VERSION

[tamnd]

The binutils 2.47 tarball we pin builds an assembler that reports itself as GNU assembler (GNU Binutils) 2.47.20260726. The kernel's scripts/as-version.sh splits the version on dots and computes 10000 * major + 100 * minor + patch, so the date is read as the patch level and every cell built with binutils 2.47 gets CONFIG_AS_VERSION=20285426 and CONFIG_LD_VERSION=20285426.

Seen first on 7.2.8 x86_64 with gcc-13.5.0, whose bundle pairs it with binutils 2.47, in the new reports/config-differential.md (#69).

Every AS_VERSION >= and LD_VERSION >= test in the tree passes with a value that large, so nothing fails today. A test the other way round, or a range, would go wrong.

To do:

• Check whether the released binutils-2.47.tar.xz really carries a dated version or whether our build adds it (the bfd/version.h date and BFD_VERSION_STRING).
• If it is upstream, find the kernel commit that handles it, if any, and list the Current kernels that misread it.

[tamnd]

It is upstream. The signed binutils-2.47.tar.xz ships bfd/development.sh with development=true and experimental=true, where 2.45 has both false, so the release tools carry BFD_VERSION_DATE (20260726) in their version string just like a snapshot would. Our build adds nothing.

scripts/as-version.sh and scripts/ld-version.sh arrived around 5.13, so on the Current set every kernel from 5.15.221 on gets the bogus value with binutils 2.47, and 5.10.270 has no AS_VERSION at all. Next step is to grep the Current trees for AS_VERSION and LD_VERSION tests that are not a plain >= to see whether any of them changes a decision.

[tamnd]

Checked 6.12.111 and 7.2.8. The only upper bounds on these symbols are the riscv ones (AS_VERSION < 180000 and LLD_VERSION < 180000), and both are guarded by AS_IS_LLVM or LD_IS_LLD, so GNU binutils never reaches them. The one equality test is powerpc's CONFIG_LD_VERSION = 23700, which looks for 2.37 exactly and is unaffected. Everything else is >= through Kconfig or test-ge.

So on the Current set the bogus value changes no decision on our platforms. I'm leaving this open as a known shape: the persona work in rucc-kernel should not copy the dated version string, and a later kernel that adds an upper bound on AS_VERSION for GNU as would trip on 2.47.