[llvm] [AVR] Support stack realignment (PR #187129)

[Ben Shi via llvm-commits]

benshi001 wrote:

I think current implementation is much better than the previous one, except two issues:

1. the above two failures in `make check-llvm-codegen -j`.
2. some wrong assembly is generated, for example

```c++
// clang a.c -O2 -Wall -S --target=avr -mmcu=atmega328
oid g(void *, void *);

void f() {
        int a = 22;
        _Alignas(8) int d = 55;
        g(&a, &d);
}
```

The following assembly is generated

```
        push    r28
        push    r29
        in      r28, 61
        in      r29, 62
        sbiw    r28, 11
        in      r0, 63
        cli
        out     62, r29
        out     63, r0
        out     61, r28
        movw    r24, r28
        adiw    r24, 10
        andi    r24, 248
        ldi     r18, 22
        ldi     r19, 0
        std     Y+2, r19
        std     Y+1, r18
        ldi     r18, 55
        ldi     r19, 0
        movw    r22, r24
        movw    r30, r22
        std     Z+1, r19
        st      Z, r18
        movw    r24, r28
        adiw    r24, 1
        call    g
        adiw    r28, 11
        in      r0, 63
        cli
        out     62, r29
        out     63, r0
        out     61, r28
        pop     r29
        pop     r28
        ret
```

`22` is stored to (Y+2):(Y+1) and `55` is stored to an location aligned to 8-byte. It seems the order is reverse.

https://github.com/llvm/llvm-project/pull/187129