Section alignment requirements and inserted padding

[Mansindam]

I wasted a few hours finding out that sh_align is handled differently by ld then mold (and gold)
ld ignores it, mold respects it and will pad data with zeros

The thing is, programs cannot rely on either behavior (eg. rip-rela addressing into a .rodata), since if using the wrong linker it generates the worst kind of obscure, delayed failures.

Perhaps mold should emit a warning rather than silently insert padding between elements of a section

[Mansindam]

I've also discovered that an alignment of 1 for a .rodata is capable of truncating/overwriting symbols explicitly sized 16

[rui314]

Can you give me an example to demonstrate the issue? As far as I know, GNU ld respects the section alignment.

[Mansindam]

clang -fuse-ld=mold main.c test.o && ./a.out ; clang main.c test.o && ./a.out
KO
OK

#include <unistd.h>
#include <stdint.h>
#include <stdalign.h>
uint64_t ffilex(char const []);
int main() {
  uint64_t r = ffilex("showU64B ptr x = VWrite ptr (SubB (ReplicateB 48) (MovmB x)) %c\n");
  write(1 , r == 10235907949189046785ul ? "OK\n" : "KO\n", 3);
  return 0;
}```

Raw relocatable
test.txt

[rui314]

I have no idea what your ffilex function is supposed to do. Please share me an example that I can actually try myself. I need to reproduce the issue to debug.

[Mansindam]

On this input it returns a fixed number IFF the .rodata is setup correctly,
more importantly, why do you even want to run it? If you do just link it and run it..?
readelf shows the setup, you can also load it in gdb and inspect the .rodata memory layout, in particular symbols d0 and d1