[AArch64] Why doesn't clang generate stp/ldp for 256-bit value ?

[Zhenhang1213]

In arm document: it shows that if use keywords like volatile with a 128-bit structure type, it is possible to treat two 64-bit values as a 128-bit value. I have confirmed this works in Clang/GCC. However, when I extend this to a 256-bit data structure, LLVM cannot handle it while GCC can. Is it possible to extend LLVM's behavior?

demo:
https://godbolt.org/z/Eoj454o6T

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Austin (Zhenhang1213)

In arm [document](https://developer.arm.com/documentation/ka004805/latest/): it shows that if use keywords like volatile with a 128-bit structure type, it is possible to treat two 64-bit values as a 128-bit value. I have confirmed this works in Clang/GCC. However, when I extend this to a 256-bit data structure, LLVM cannot handle it while GCC can. Is it possible to extend LLVM's behavior?

demo:
https://godbolt.org/z/Eoj454o6T

[davemgreen]

There is code to handle the volatile load/store to LDP case (https://godbolt.org/z/Yzbb1EoKM), but not from a memcpy, which will be lowered to "legal" types (i64) and not folded into LDP (https://godbolt.org/z/n4eWejhMr). Getting LLVM to do the same through a memcpy sounds sensible to me if you can figure out a way to do it. It goes via getOptimalMemOpType and gets split up in DAG creation.