Choices for implementing the same_value cast:

• 1 A dedicated inner loop, which would double the number of casting loops: we already have 4: aligned/not aligned and strided/contiguous, so that would mean 8 loop functions
  code around the cast in the inner loop and a if flag==SAMEVALUE branch.
• 2 Adding a flags field to the context. This will impact performance
• 3 Some check after each inner loop iteration that the values are the same. But how would that work: we would still need more loops.

2: Checking only the flag on the fast path

    if (context->flags != 1234) {
        *(_TYPE2 *)dst = _CONVERT_FN(*(_TYPE1 *)src);
    }

makes this timeit test go from 640ns to 880 ns:

timeit -s "import numpy as np; a = np.arange(10_000, dtype=np.int64)" "a.astype(np.int8)"

Increasing the size to 10_000_000 and testing with/without the dummy flag makes the time go from 74.5ms to 69.5ms. The casting loops already use NPY_GCC_OPT_3.

Update: it looks like the CPU can elide the if flags out of the loop, and so the non-checking loop is as fast as not having a check at all as long as I do

uint64_t flags = context->flags;
while... {
    if (flags == NP_CHECK_OVERFLOW) {
        // do the check
    }
    *(_TYPE2 *)dst = _CONVERT_FN(*(_TYPE1 *)src);
}

Sebastian says: you cannot change the values of the enums in https://github.com/mattip/numpy/blob/005ce7630792a94a9330b9af809904bb10ff2ea6/numpy/_core/include/numpy/ndarraytypes.h#L227C9-L227C30