tools/nxflat: Add an Apache-licensed NXFLAT converter

[casaroli]

Summary

ldnxflat is the last part of the NXFLAT toolchain that NuttX cannot carry, so a board that builds NXFLAT modules has to fetch a GPL tool from buildroot. This adds an Apache-2.0 replacement, and a driver that builds every test module through the toolchain.

mknxflat came in with #19600 and was relicensed with its author's agreement, because Gregory Nutt owned all of it. ldnxflat cannot follow: it descends from elf2flt and obj-res.c through eleven sets of copyright holders reaching back to 1996, several of them companies that no longer exist. It is GPL by that descent rather than by its libbfd dependency, so no set of permissions can be assembled and a rewrite is the only route.

NXFLAT is not an ARM format. binfmt/Kconfig puts no architecture dependency on it and the loader only ever adds a base to a 32-bit word, so an architecture here supplies a table entry, an entry-point convention and a relocation handler, and an object for an unknown machine is refused by name rather than converted wrongly.

Impact

Nothing changes for a configuration that does not set CONFIG_NXFLAT. For one that does, LDNXFLAT names the tool in the tree instead of looking for one on PATH, and mknxflat's output is unchanged: the thunk it generates for each of the eleven test modules is byte identical to the one the merged version generates.

Two defects of the old converter do not survive the rewrite. A GOT entry naming a .bss object now resolves to the object, where the old tool dropped the section's address and pointed it at the start of D-Space — the GOT itself — so a module reaching a static through the GOT read and wrote its own GOT. And the alignment gap before .bss no longer goes missing from h_bssend.

A module can also carry a string, which is a compiler flag rather than the converter, and not a new finding: lm3s6965-ek has had -mno-pic-data-is-text-relative in its own scripts/Make.defs since 2021 (issue #3737) and the CMake build applies it to every PIC configuration. The first commit moves it where it belonged; the other sixteen NXFLAT configurations were without it.

Testing

Nine of the seventeen configurations that set CONFIG_NXFLAT are on the CI blacklist, and they are exactly the ones that build a module — the converter was not there to build them with. eagle100:nxflat and lm3s6965-ek:qemu-nxflat come off it here, and both were built with the toolchain CI uses. The other seven stay off for reasons of their own: the thttpd configurations want CONFIG_BOARDCTL_ROMDISK, which their defconfigs do not set, and some of those boards define their own CPICFLAGS without filtering --fixed-r9, so the compiler refuses r9 as the PIC register.

lm3s6965-ek:qemu-nxflat also runs. Under qemu-system-arm -M lm3s6965evb its ROMFS modules execute to End-of-Test.. Exit-ing, struct reporting pf = 0x13335 — odd, so the Thumb bit survives — and then In dummyfunc() -- PASS. Built with the out-of-tree tool instead, the same image panics in the third test with PC: 00000000 and the remaining five never run: longjmp's jmp_buf is a static, its GOT entry lands on the GOT, and setjmp overwrites it.

tools/nxflat/testsuite.sh builds and converts every module of apps/examples/nxflat/tests and reports the relocations each one carried. It needs no board: a module is built for cortex-m3 against a configured tree's headers.

tools/checkpatch.sh -c -u -m -g passes.

Provenance

The container is defined by include/nxflat.h and by what binfmt/libnxflat does with it, the segment layout by gnu-nxflat-gotoff.ld, and the relocation arithmetic is that of libs/libc/machine/arm/armv7-m/arch_elf.c. Those files and the new one are Apache-2.0.

No part of the out-of-tree tool is used. During development it was run on the same inputs to compare output against, which is how the conventions the container does not state — where the GOT sits, the order of its entries, the order of the relocation records — were matched. Six modules come out byte identical to it, hello++3 converts where it refuses R_ARM_TARGET1, and four differ where it is wrong. Two of its results are deliberately not matched, and the file says which.

apache/nuttx-apps#3803 builds the test modules this exercises. It is not needed for this one to be correct.

[github-actions]

MemBrowse Memory Report

No memory changes detected for:

• arduino-mega2560
• esp32-devkitc
• hifive1-revb
• mirtoo
• qemu-armv8a
• qemu-intel64
• rx65n-rsk2mb
• s698pm-dkit
• stm32-nucleo-f103rb