How do I stop GCC from optimizing this byte-for-byte copy into a memcpy call?

One thing that seems to be sufficient here: instead of using -fno-builtin-memcpy use -fno-builtin for compiling the translation unit of memcpy alone!

An alternative would be to pass -fno-tree-loop-distribute-patterns; though this might be brittle as it forbids the compiler from reorganizing the loop code first and then replacing part of them with calls to mem* functions.

Or, since you cannot rely anything in the C library, perhaps using -ffreestanding could be in order.

This won't work (memcpy unconditionally calls itself), and it causes a segfault.

Redefining memcpy is undefined behavior.

How do I disable the optimization that causes the copy to be transformed into a memcpy call (preferably while still compiling with -O3)?

Don't. The best approach is fixing your code instead:

• In most cases, you should use another name.

• In the rare case you are really implementing a C library (as discussed in the comments), and you really want to reimplement memcpy, then you should be using compiler-specific options to achieve that. For GCC, see -fno-builtin* and -ffreestanding, as well as -nodefaultlibs and -nostdlib.