remove_spaces.S

.intel_syntax noprefix
.global remove_spaces_asm16

.section .mydata, "a"
    maps:
        .incbin "map.bin"
    shifts = maps + 0x10000*16
    spaces16:
        .fill 16, 1, 0x20

.text
    remove_spaces_asm16:
        // char* remove_spaces_asm16(char *in, char *out, unsigned int count);
        // rdi, rsi, rdx

        vmovdqu xmm2, [spaces16]

        1:
            vmovdqu     xmm0, [rdi]
            vpcmpeqb    xmm1, xmm2, xmm0
            vpmovmskb   rax, xmm1
            movzx       ecx, byte ptr [shifts + rax]
            shl         eax, 4
            vpshufb     xmm0, xmm0, [maps + rax]
            vmovdqu     [rsi], xmm0
            add         rdi, 16
            add         rsi, rcx
            sub         rdx, 16
            jne 1b

        mov rax, rsi

        ret