[casaroli]

Summary

With CONFIG_LTO_FULL (GCC), the Toolchain.defs of arm, arm64 and risc-v set LD to the gcc driver, and a loadable module (and a program of a kernel build) is linked with it, because MODULELD defaults to LD. The partial link (-r) of LTO objects keeps only LTO data, which the loader cannot use: exec failed: 22, and nm says plugin needed to handle lto object.

Where these Toolchain.defs set LD to gcc for LTO, they now also set LDMODULELTOFLAGS and LDELFLTOFLAGS: -flinker-output=nolto-rel and the compile options of the module. LDMODULEFLAGS and LDELFFLAGS use them next to -r. The link then compiles the objects to machine code with the options of the module (-mthumb, -mlong-calls, ...). Without these options the arm-none-eabi assembler stopped with invalid constant (3e8) after fixup.

crt0.o is now compiled without -flto. It must be machine code in any case, and as LTO data it made LTO warn that crt0 and the program declare main() differently, which -Werror makes an error.

The other commits fix the other links that gcc makes with LTO:

• risc-v: LDELFFLAGS has --oformat elf32-littleriscv (or elf64), an option of the linker, and the gcc driver (LTO) and clang (the clang toolchain sets LD to clang) stop with unrecognized command-line option '--oformat'. The branches that set LD to a driver now set LDXLINKER to -Xlinker, and LDELFFLAGS has $(LDXLINKER) --oformat=.... With ld the option is the same as before.
• arm64: Toolchain.defs adds -nostdlib to LDFLAGS, and the Makefile gives LDFLAGS to ld through -Xlinker, so gcc still added its start files. qemu-armv8a:knsh then stopped on the .eh_frame of crtbegin.o. -nostdlib now goes to gcc too, as on arm.
• A kernel build: the export (mkexport.sh -u) had no LDSTARTGROUP and LDENDGROUP, so apps linked each program with --start-group, which gcc rejects. The branches that set LD to the driver now set them to -Wl,--start-group and -Wl,--end-group, and mkexport.sh writes them for every export.
• risc-v without CONFIG_BINFMT_ELF_RELOCATABLE (rv-virt:knsh64): a program is linked to an executable with no compile options, so gcc added the crt0.o of its C library (multiple definition of '_start') and crashed in LTRANS without the ABI options. LDELFEXEFLAGS now gives the ELF options to this link.
• arm with CONFIG_PIC (not FDPIC) or CONFIG_BINFMT_ELF_EXECUTABLE: a program is linked to an executable too. The PIC link gave it options of ld (unrecognized command-line option '--unresolved-symbols=ignore-in-object-files'), and gcc added its start files. These links now give the options of ld with $(LDXLINKER) and add $(LDELFEXEFLAGS) (the ELF options and -nostdlib).

This replaces the Make.defs part of apache/nuttx-apps#3884, as asked there. That PR keeps only the -Wl,--defsym= form of the program attributes for the gcc driver; a module links only with both.

Impact

• CONFIG_LTO_FULL with GCC on arm, arm64 and risc-v: the modules (flat build) and the programs (kernel build) link to machine code and load.
• risc-v with clang: LDELFFLAGS no longer stops the driver.
• Not covered here, with CONFIG_LTO_FULL: a protected build fails on master already in the link of the user image (the board kernel/Makefile gives --undefined= and --start-group to $(LD)). The qemu-armv7a:knsh image (armv7-a, also armv7-r and armv8-r) fails on master already: dispatch_syscall() reads g_stublookup only in assembly, which LTO does not see (undefined reference to 'g_stublookup').
• Other architectures do not set LD to a compiler driver: no change.
• Without LTO, with clang (ld.lld) or with FDPIC (MODULELD is ld): no change. crt0.o is machine code also with LTO.
• No documentation change.

Testing

qemu-armv8a:nsh (flat, Arm GNU 14.2, QEMU HVF) with CONFIG_LTO_FULL, CONFIG_ELF, CONFIG_EXECFUNCS_SYSTEM_SYMTAB and CONFIG_EXAMPLES_HELLO=m, apps with apache/nuttx-apps#3884, hello loaded from a 9p share:

	master	this PR
hello link	-Wl,--defsym=... only	-flinker-output=nolto-rel and the module options (with -Werror)
nm hello	plugin needed to handle lto object	18 defined symbols, no LTO data
run	nxposix_spawn_exec: ERROR: exec failed: 22	Hello, World!!

The same test with QEMU TCG and hello on a semihosting hostfs, apps with apache/nuttx-apps#3884:

configuration	build	hello
rv-virt:nsh64, CONFIG_LTO_FULL (master: unrecognized command-line option '--oformat')	flat	Hello, World!!
rv-virt:nsh64, no LTO (ld)	flat	Hello, World!!
mps2-an500:nsh, CONFIG_LTO_FULL, Arm GNU 15.3	flat	Hello, World!!
rv-virt:knsh64, no LTO	kernel	ostest Exiting with status 0
rv-virt:knsh64, CONFIG_LTO_FULL (master: unrecognized command-line option '--start-group')	kernel	ostest Exiting with status 0
qemu-armv8a:knsh, CONFIG_LTO_FULL, HVF (master: .eh_frame of crtbegin.o)	kernel	ostest Exiting with status 0
qemu-armv8a:nsh, CONFIG_LTO_FULL, HVF	flat	ostest Exiting with status 0
qemu-armv8a:knsh and :nsh, no LTO	kernel, flat	ostest Exiting with status 0
mps3-an547:picostest (CONFIG_PIC), CONFIG_LTO_FULL, romfs on /pic (master: unrecognized command-line option '--unresolved-symbols=ignore-in-object-files')	flat	/pic/hello Hello, World!!, /pic/ostest Exiting with status 0
mps3-an547:picostest, no LTO	flat	the same
qemu-armv7a:knsh (CONFIG_BINFMT_ELF_EXECUTABLE), no LTO	kernel	ostest Exiting with status 0

LDELFFLAGS and LDMODULEFLAGS as make expands them, before and after this change: the same for qemu-armv8a:nsh, qemu-armv8a:knsh, mps2-an500:nsh, rv-virt:nsh and rv-virt:knsh (risc-v: --oformat=elf32-littleriscv in place of --oformat elf32-littleriscv), and for mps3-an547:picostest and qemu-armv7a:knsh; with CONFIG_LTO_FULL (qemu-armv8a:nsh, mps2-an500:nsh, rv-virt:nsh) they have -flinker-output=nolto-rel and the module options.

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

[github-actions]

MemBrowse Memory Report · Explore what changed →

No memory changes detected for:

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

---

_{Catch memory regressions in CI with MemBrowse. Free for open source.}