Add R_RISCV_ALIGN_DOWN

[palmer-dabbelt]

We're in the process of stumbling into another round of these bugs related to insufficient alignment bytes. The idea of adding another alignment reloation that rounds down has come up a few times, but I don't remember if anyone's actually written the spec. So here's one.

[jrtc27]

This needs justification within this forum (I at least have not been aware of any of the discussions that led to this proposal)

[palmer-dabbelt]

#183
riscv-collab/riscv-gnu-toolchain#445

[kito-cheng]

That remind me that we have similar issue for R_RISCV_RELAX: #116, and @jrtc27 has proposed we can point a dummy ISA symbol by the R_RISCV_RELAX, and I think...that could also use similar mechanism here? and IIRC @ptomsich has mention RISC-V not implement max alignment for the .align directive, so I guess...we could resolve together instead of add two more relocation (one for R_RISCV_ALIGN_DOWN and one for R_RISCV_ALIGN_MAX) .

c.c. @Nelson1225 @MaskRay

[MaskRay]

Thanks for tagging me. To address issues like riscv-collab/riscv-gnu-toolchain#445 , I agree that discerning RVC vs non-RVC R_RISCV_RELAX is important, and we can use @jrtc27's idea to use the symbol index part of the r_info field (say, symbol_index!=0 means non-RVC). This scheme is compatible with existing linkers.

I haven't thought much regarding R_RISCV_ALIGN. Right now I just know that the two fields in the syntax .align [abs-expr[, abs-expr[, abs-expr]]] are not representable, but it seems that people are fine with the limitation. A new relocation type doesn't seem to be needed if we just try to resolve riscv-collab/riscv-gnu-toolchain#445?