It's not entirely folklore. We used to do this sort of thing all the time in the AppleII and 64KB PC days. Heck I used to have a switch construct in PC assembler that pushed the the segment and instruction pointer to stack, added a few bytes, then pushed a destination onto the stack and did an IRET. The routines it jumped to did an IRET to come back. I saved like 24 bytes by doing it "wrong" and basically broke any disassembler out there. Also, it was faster... =)

Evil self-modifying code was the only way you could get anything to work in non-glacial speed on the Apple. My friend Rick wrote an 8 direction generalized 8x8 pixel block move that self modified itself in a tight loop. It could smoothly move about 50 cells of 8x8 tiles faster than the screen refresh and was about 130ish bytes. The original way had taken over a 1k and he wanted more room for tiles. And yes, we explored modifying the code a bit to use it as graphic tiles first. =)

Writing in a memory budget was terrific for helping you think outside the box...

--
$you = new YOU;
honk() if $you->love(perl)