Fix sret for AArch64

[bjorn3]

AArch64 requires the struct return address argument to be stored in the x8 register. This register is never used for regular arguments.

According to https://github.com/Gankra/abi-checker cg_clif with this PR matches the abi of LLVM for big structs.

Marking as draft as the test fails with:

Register allocation error for vcode
VCode {
  Entry block: 0
Block 0:
    (original IR block: block0)
    (instruction range: 0 .. 6)
  Inst 0: mov %v128, x8
  Inst 1: mov %v129, %v128
  Inst 2: mov %v129, %v128
  Inst 3: mov x8, %v129
  Inst 4: mov x0, %v130
  Inst 5: ret
}

Error: EntryLivein
CLIF for error:
function %f17(i64 sret) -> i64 fast {
block0(v0: i64):
    return v0
}

[afonso360]

I did some investigation into this, and i think this is being broken by the ensure_struct_return_ptr_is_returned legalization pass.

If I understand this correctly it inserts a second return arg, which is causing some trouble in the regalloc (still don't fully understand why though).

However, by adding that return arg I think it is breaking the AArch64 ABI anyway, so we should probably restrict that to x86. (probably by moving it to the x86 backend as suggested in #4618 (comment)).

[bjorn3]

I think I found the problem. (i64 sret) -> i64 is legalized to (i64 sret) -> i64 sret, i64, but the return instruction doesn't account for this. This is a pre-existing problem with all backends. It isn't required to work either. In other words I wrote a bad test on which Cranelift failed correctly, but with a confusing error. I will fix the test. If at a later point sret legalization is removed in favor of direct integration with the backends, that should fix this issue.

function %f17(i64 sret) -> i64 {
block0(v0: i64):
    return v0
}

[bjorn3]

Apart from i128, all abi-checker tests now pass with this PR.