[S] Change in coreboot[main]: commonlib: Add assembly optimization for ipchksum() on arm64

[Julius Werner (Code Review)]

Attention is currently required from: Paul Menzel.

Julius Werner has posted comments on this change. ( https://review.coreboot.org/c/coreboot/+/80254?usp=email )

Change subject: commonlib: Add assembly optimization for ipchksum() on arm64
......................................................................

Patch Set 3:

(2 comments)

Commit Message:

https://review.coreboot.org/c/coreboot/+/80254/comment/541599d8_0ad72bee :
PS2, Line 10: to take advantage of larger load
            : sizes and the add-with-carry instruction
> Excuse my ignorance, but why can’t the compilers do that themselves?

It's very hard to write C code for this in the first place because the C language has no means to describe the carry that gets shifted out of the end of a data type. The best I can come up with is to check it afterwards like this
```
for (size_t i = 0; i < size; i++) {
        uint64_t new_sum = result + *data++;
        if (new_sum < result)
                new_sum++;
        result = new_sum;
}
```
and interestingly the compiler does make something clever out of that on both x86
```
  10:   48 03 04 d7             add    (%rdi,%rdx,8),%rax
  14:   48 83 d0 00             adc    $0x0,%rax
  18:   48 83 c2 01             add    $0x1,%rdx
  1c:   48 39 d6                cmp    %rdx,%rsi
  1f:   75 ef                   jne    10 <sum+0x10> (File Offset: 0x50)
```
and Arm
```
   8:   f8637804        ldr     x4, [x0, x3, lsl #3]
   c:   91000463        add     x3, x3, #0x1
  10:   ab040042        adds    x2, x2, x4
  14:   9a823442        cinc    x2, x2, cs      // cs = hs, nlast
  18:   eb03003f        cmp     x1, x3
  1c:   54ffff61        b.ne    8 <sum+0x8> (File Offset: 0x48)  // b.any
```
but it's not really as clever as my code. In general I found that for very tight loops when every instruction can make a significant performance difference, GCC still doesn't really get as good as handcrafted assembly. (Also, when writing code that so explicitly tries to bait a specific instruction, one could argue that just writing the assembly directly is actually more readable.)

https://review.coreboot.org/c/coreboot/+/80254/comment/f6d7ec59_1c24940a :
PS2, Line 12: by more than 20x
> Thank you for documenting this. […]

They're not really worth mentioning (a couple of microseconds) because the coreboot tables that get checksummed by this are pretty small (at least on the platform I tested). I did this mostly for fun tbh, not for serious gain. I think some other platforms are also using that checksum for other things so it might help them a bit more (we used to use it for MRC cache where the impact was actually significant, but now that was replaced with xxhash already).