[compiler-rt] [Compiler-rt] Add AArch64 routines for __arm_agnostic("sme_za_state") (PR #120059)

[Zequan Wu via llvm-commits]

ZequanWu wrote:

> > Is that expected? Are we holding something wrong? Or is that too little information to say something useful?
> 
> I don't think it's expected (looks like the chrome build is clang trunk).
> 
> It looks like the `+sme2` support detection is not working:
> 
> ```c++
> // Fails to build on clang trunk (+sme2 not set)
> void foo(void)  __arm_streaming_compatible {
>   asm(".arch armv9-a+sme2");
>   asm("smstart");
>   asm("ldr zt0, [sp]");
> }
> // (side note: It seems smstart does not require +sme to be assembled)
> ```
> 
> See: https://godbolt.org/z/osdfGd6dn
> 
> It looks like either of these would work:
> 
> ```c++
> // Builds successfully on clang trunk.
> __attribute__((target("arch=armv9-a+sme2")))
> void foo(void)  __arm_streaming_compatible {
>   asm("smstart");
>   asm("ldr zt0, [sp]");
> }
> 
> // OR:
> 
> void foo2(void)  __arm_streaming_compatible {
>   asm(".arch armv9-a+sme2\n"
>       "smstart\n"
>       "ldr zt0, [sp]\n");
> }
> ```
> 
> See: https://godbolt.org/z/aqTTvKcE5

Is it a regression introduced by this change? If so, please revert it if it takes a while to fix it.

https://github.com/llvm/llvm-project/pull/120059