[very low priority, upstream?] Relax away addi/cincoffset-0 resulting from highly aligned addresses?

[nwf]

Consider the construction of an address from a symbolic value, something like

1:
  lui %0, %hi(__export_mem_uart)
  addi %0, %0, %lo(1b)

If the value of __export_mem_uart has all of its 12 least significant bits zero, then the lui has done all the necessary work and the addi is going to be a no-op. Presently, AFAICT, the linker relaxation will leave the no-op in place, as they consider the zero-ness of %hi, but not of %lo.

While this is not likely to happen "by chance" given a random symbol, almost all of our MMIO addresses (__export_mem_) are sufficiently aligned for such a relaxation to fire.

Similar considerations apply for auipcc- and auicgp-based constructions (see also #153).

[davidchisnall]

I believe these will be relaxed if the compiler generates them, but our inline assembly doesn't emit the relax relocations.

[nwf]

FWIW, they don't seem to be relaxed if they come from the assembler either, even with -mrelax. Perhaps I am simply failing to use the right directive(s)?

[davidchisnall]

Assembly needs explicit directives to add the relocations that permit relaxation.