On Thu, May 13, 2021 at 5:55 PM Bin Meng bmeng.cn@gmail.com wrote:
On Fri, May 14, 2021 at 7:56 AM Simon Glass sjg@chromium.org wrote:
On Thu, 13 May 2021 at 08:46, Heinrich Schuchardt xypron.glpk@gmx.de
wrote:
/usr/bin/ld: common/built-in.o: in function `bootdelay_process': common/autoboot.c:335: undefined reference to `_init' collect2: error: ld returned 1 exit status make: *** [Makefile:1726: u-boot] Error 1
In the ELF standard, .init sections were deprecated and replaced with .init_array sections in the early 1990s I think. It took some time for the toolchains to fully migrate from init to init_array, but by 2010 or so everyone was using init_array instead of init, with init support retained only for legacy code that hadn't been fixed to use init_array yet. Because RISC-V is a new target with no legacy code to support, we made the decision to drop support for the obsolete init section. Last I checked there are only two glibc targets that have no init section support, with the other one also being a new arch, like RISC-V. Same goes for the embedded target, so the RISC-V newlib port has no init section support also.
It sounds like you have some legacy user code that hasn't been fixed yet to use init_array instead of init. Or maybe it is a program loader that supports both? In which case it should be extended to not use init on new targets that no longer support it.
Init_array supports stack unwinding (aka C++ EH) and init doesn't, so init_array should always be preferred over init, unless you have a very old toolchain that lacks init_array support. Dropping init support from the RISC-V toolchain allows us to save some bytes of program code size, and save some cycles on program startup, which is good considering that this is a feature that we don't need anymore.
Jim