[Draft] Arancini Calling Convention

[martin-fink]

This adds an arancini calling convention in LLVM that passes and returns all 16 GPRS + PC + FS + GS in host registers.
This should, in theory, decrease the amount of loads/stores from and to the cpu state.

Untested so far.

Example for a small function with just three params/return values on AArch64.

define arancini { i64, i64, i64 } @func2(i64 %0, i64 %1, i64 %2) {
  %4 = add i64 %0, 1
  %5 = add i64 %1, 1
  %6 = add i64 %2, 1

  %ret_val = insertvalue { i64, i64, i64 } undef, i64 %4, 0
  %ret_val2 = insertvalue { i64, i64, i64 } %ret_val, i64 %5, 1
  %ret_val3 = insertvalue { i64, i64, i64 } %ret_val2, i64 %6, 2

  ret { i64, i64, i64 } %ret_val3
}

lowers to the following:

add	x0, x0, #1
add	x1, x1, #1
add	x2, x2, #1
ret

There's some stuff in here that should not be merged. Also, the LLVM code for this currently lives at https://github.com/martin-fink/llvm-project/tree/arancini-calling-conv, we might want to move it here if this makes sense.

[martin-fink]

I cannot convert this to a draft since I'm lacking some permissions.

[martin-fink]

For the microbenchmark of virt fn calls, this speeds those up a bit:

baseline	arancini-cc	re-added switch	internal static lookup fn
3,041.24 ms	810.88 ms	795.16 ms	709.98 ms

[martin-fink]

Function signatures now look like this:

define arancini { i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64, i64 } @func(ptr noalias nocapture noundef nonnull align 16 dereferenceable(2232) %0, i64 %1, i64 %2, i64 %3, i64 %4, i64 %5, i64 %6, i64 %7, i64 %8, i64 %9, i64 %10, i64 %11, i64 %12, i64 %13, i64 %14, i64 %15, i64 %16, i64 %17)

We pass the CPU state + PC + 16 * GPR + FS + GS and return the same except for the CPU state, as the caller already has the CPU state. To ensure that we then don't need to shuffle stuff around, we skip one return register in the calling convention.