A jump table
The same program in M68K, MIPS, RISC-V, Z80.
Four possible destinations and not one comparison between them. Choosing between four branches with
cmp and je costs up to four comparisons, and every case you add costs one more. This program costs
one memory read and one jump, and it would still cost one memory read and one jump with two hundred
cases in the table.
dq case0 is the line that makes this possible, and it is stranger than it looks. A label is a
number, the address of some code, and nothing stops you storing that number in memory the way you
would store a price or a count. Four labels, four qwords, an ordinary array that happens to be full of
addresses of instructions.
jmp [rbx + rcx*8] then reads eight bytes out of that array and puts them in rip. The brackets are
load bearing: jmp rbx without them would jump to the address of the table itself and start executing
the table, treating a list of addresses as a list of instructions.
The cmp rcx, CASES and jae above it are not a nicety. An index past the end of the table reads
eight bytes of whatever the assembler put after it and jumps to that, and "whatever was next in memory"
is not an address of anything. Set mov rcx, 9 and the check catches it and r8 comes out as -1. Now
delete the check and run it again: the program stops somewhere, with no message, and rip in the
registers panel is your only evidence about where it went.
The four labels here are ordinary ones and not local .case0 labels, and it is not a style choice. A
local label belongs to the last ordinary label above it, so a .done written inside case0 would be a
different label from a .done written inside case1, and the four jmp done lines would go to four
different places.