minimal_rt: fix intrinsics to correctly return original pointer

[chris-oo]

The implementation for intrinsics in minimal_rt was incorrect. For example, the inline asm block for memcpy correctly describes that rdi will be clobbered by rep movsb, but memcpy's API requires that the return value is the original destination pointer, not the incremented pointer. Fix this by removing the mut on dest and indicating to the compiler to discard the clobbered value.

It seems that in most cases, the compiler ignored the return value of memcpy, except when #2031 was being added. Certain compilation options or function implementations would cause the compiler to use the inlined returned value of memcpy directly from the call to realloc, which would then cause a later an assertion failure due to the pointer not being what was expected and overlapping with the new destination.

Fix all the intrinsics in minimal_rt to correctly return the unmodified dest or ptr as they should.

[Copilot]

Pull Request Overview

This PR fixes a bug in the memcpy implementation in minimal_rt where the function was returning the modified destination pointer instead of the original destination pointer as required by the C API standard.

• Removes the mut keyword from the dest parameter to prevent accidental modification
• Changes the inline assembly output to discard the clobbered rdi register value instead of capturing it

[smalis-msft]

Check memset, ARM impls, and any other similar functions too?

[smalis-msft]

I am amazed we got this far before this became an issue.

[chris-oo]

I know right? Seems like pure dumb luck if anything. But I am wondering if in a full optimized build with fat LTO, if the compiler actually knows that dest is mutated and can't be relied on, and must spill the original value if it wants it. That's at least the final assembly code I saw for the working binaries, on release builds.