Rewrite InterlockedExchange128_nf with intrinsics on clang

[cbezault]

I'm not sure the alignas(16) on the __int128 is strictly necessary.
I'm also not 100% sure this is correct, though all the tests I ran on my Surface Pro X passed.

@BillyONeal @AlexGuteniev bat signals.

The codegen also seems right:
https://godbolt.org/z/1d8MnT vs

; Listing generated by Microsoft (R) Optimizing Compiler Version 19.28.29515.1

        TTL     C:\sources\test.obj
        ARM64

        AREA    |.drectve|, DRECTVE

        EXPORT  |?dest@@3RC_JC| [ DATA ]                ; dest

        AREA    |.data|, DATA
|?dest@@3RC_JC| DCQ 0x0, 0x0                            ;  = 0x0000000000000000 ; dest
        DCQ     0x1, 0x0                                ;  = 0x0000000000000001
        EXPORT  |main|

        AREA    |.pdata|, PDATA
|$pdata$main| DCD imagerel |$LN6|
        DCD     0x80004d
        ;Flags[SingleProEpi] functionLength[76] RegF[0] RegI[0] H[0] frameChainReturn[UnChained] frameSize[16]
; Function compile flags: /Odsp
; File C:\sources\test.cpp

        AREA    |.text$mn|, CODE, ARM64

|main|  PROC

; 6    : int main() {

|$LN6|
        sub         sp,sp,#0x10

; 7    :     __int64 compare[2];
; 8    :     _InterlockedCompareExchange128_nf(dest, compare[1], compare[0], compare);

        mov         x7,sp
        adrp        x8,|?dest@@3RC_JC|
        add         x15,x8,PageOffset(|?dest@@3RC_JC|)
        ldr         x14,[sp,#8]
        ldr         x13,[sp]
        ldp         x12,x11,[x7]
|$LN3@main|
        ldxp        x10,x9,[x15]
        cmp         x10,x12
        bne         |$LN4@main|
        cmp         x9,x11
        bne         |$LN4@main|
        stxp        w8,x13,x14,[x15]
        cmp         w8,#0
        bne         |$LN3@main|
|$LN4@main|
        stp         x10,x9,[x7]

; 9    :     return 0;

        mov         w0,#0
        add         sp,sp,#0x10
        ret

        ENDP  ; |main|

        END

[cbezault]

Maybe fixes #1491

[cbezault]

Okay went through all the steps of actually compiling the STL and then a test program and testing it on the Surface Pro X again. I also looked at the new codegen and it seems right.

[CaseyCarter]

FYI: I wanted to apply a small refactor to reduce duplication in the not-clang and not-ARM64 cases, which quickly turned into restoring the original code on the #else branch. I also made _Dest be const in the workaround path.

[StephanTLavavej]

Thanks for fixing and validating this! 🛠️ 😸