I found memcpySmallAllowReadWriteOverflow15Impl will call built-in memcpy because of clang loop-idiom optimization. It is as expected, or an issue?

[taiyang-li]

https://godbolt.org/z/YxM3rc7Eh

[alexey-milovidov]

Not expected. We already noticed this behavior in GCC a long time ago, but not in Clang:
https://presentations.clickhouse.com/zero_cost_conf/memcpy/#11

[nickitat]

it is known
llvm/llvm-project#56467

[alexey-milovidov]

@taiyang-li, let's use [[clang::no_builtin("memcpy")]] to better express the intent.

[taiyang-li]

@taiyang-li, let's use [[clang::no_builtin("memcpy")]] to better express the intent.

i tried before, but the function won't be inlined, which is slow. if i add always inline attribute, the compiler will compain with an error.

[alexey-milovidov]

Ok, then your approach will suffice.