Big endian baremetal toolchain for HS38 generates invalid code

[kolerov]

This bug appears in snps-arceb-multilib-elf32 configuration for crosstool-ng. It's presented if you use both arc-2022.09 and arc-2023.03 branches for newlib, gcc and binutils.

Consider this example:

#include <stdio.h>

int main()
{
    puts("hello");
    return 0;
}

Then build and run it using nSIM:

arceb-elf32-gcc -mcpu=hs38 -specs=nsim.specs main.c -o main.elf
nsimdrv -tcf $NSIM_HOME/etc/tcf/templates/hs38_full.tcf -on nsim_isa_big_endian -on nsim_emt main.elf

Then nsimdrv shows nothing. I've found that such behavior may be observed only for -mcpu targets with -mmpy-option=9 (with support of all MPY instructions): -mcpu=hs38 and -mcpu=hs38_linux.

The example works well with -specs=nano.specs. Bug doesn't appear if the toolchain is built with -O0 flag. I suspect that the toolchain generates invalid optimized code for Newlib.