Some people (even here) mentioned about the cheat to vectorize your code: translate it to C, feed to GCC with desired `-march=...` / `__attribute__((__target__("...")))` (you still need intrinsics for good results, but at least they are there and work...), copy the assembly back, and dispatch these clumps with `uses cpu` and function pointers. But `x86(-64)` assembler warns about negative offsets in certain places where they are quite normal. ```pascal {$asmmode intel} begin asm lea rsp, [rsp-32] // (1) lea rax, @table[rip] movss xmm0, dword ptr [rax-8] // (2) lea rsp, [rsp+32] jmp @next db 0, 1, 2, 3, 4, 5, 6, 7 @table: db 8, 9, 10 @next: end ['rax', 'xmm0']; end. ``` (1) emits `Warning: Use of [RSP-offset], access may cause a crash or value may be lost.`, good for any other instruction, or for any other destination but `rsp` itself. (2) emits `Warning: Check "[movups mem,xmmreg0]: offset of memory operand is negative "-8 byte"`. Idk, GCC sometimes, at a whim, moves pointer increments to the beginning of the loop body and then accesses these pointers with such offsets.