#include "SABER_params.h"

.align 2
.global PQCLEAN_LIGHTSABER_AARCH64_asm_round
.global _PQCLEAN_LIGHTSABER_AARCH64_asm_round
#ifndef __clang__
.type PQCLEAN_LIGHTSABER_AARCH64_asm_round, %function
#endif
PQCLEAN_LIGHTSABER_AARCH64_asm_round:
_PQCLEAN_LIGHTSABER_AARCH64_asm_round:


    .equ srv, (SABER_EQ-SABER_EP)

    ld2 { v0.8H,  v1.8H}, [x1], #32
    ld2 { v2.8H,  v3.8H}, [x1], #32
    ld2 { v4.8H,  v5.8H}, [x1], #32
    ld2 { v6.8H,  v7.8H}, [x1], #32

    srshr  v0.8H,   v0.8H, #srv
    srshr  v2.8H,   v2.8H, #srv
    srshr  v4.8H,   v4.8H, #srv
    srshr  v6.8H,   v6.8H, #srv

    mov x7, #7
    _round_loop:

    st1 { v0.8H}, [x0], #16
    ld2 { v0.8H,  v1.8H}, [x1], #32
    st1 { v2.8H}, [x0], #16
    ld2 { v2.8H,  v3.8H}, [x1], #32
    st1 { v4.8H}, [x0], #16
    ld2 { v4.8H,  v5.8H}, [x1], #32
    st1 { v6.8H}, [x0], #16
    ld2 { v6.8H,  v7.8H}, [x1], #32

    srshr  v0.8H,   v0.8H, #srv
    srshr  v2.8H,   v2.8H, #srv
    srshr  v4.8H,   v4.8H, #srv
    srshr  v6.8H,   v6.8H, #srv

    sub x7, x7, #1
    cbnz x7, _round_loop

    st1 { v0.8H}, [x0], #16
    st1 { v2.8H}, [x0], #16
    st1 { v4.8H}, [x0], #16
    st1 { v6.8H}, [x0], #16

    br lr

.align 2
.global PQCLEAN_LIGHTSABER_AARCH64_asm_enc_add_msg
.global _PQCLEAN_LIGHTSABER_AARCH64_asm_enc_add_msg
#ifndef __clang__
.type PQCLEAN_LIGHTSABER_AARCH64_asm_enc_add_msg, %function
#endif
PQCLEAN_LIGHTSABER_AARCH64_asm_enc_add_msg:
_PQCLEAN_LIGHTSABER_AARCH64_asm_enc_add_msg:

    .equ srv, (SABER_EP-SABER_ET)
    .equ slv, (SABER_EP-1)

    dup v30.8H, w3

    ld2 { v0.8H,  v1.8H}, [x1], #32
    ld2 { v2.8H,  v3.8H}, [x1], #32
    ld2 { v4.8H,  v5.8H}, [x1], #32
    ld2 { v6.8H,  v7.8H}, [x1], #32
    ld1 { v1.8H}, [x2], #16
    ld1 { v3.8H}, [x2], #16
    ld1 { v5.8H}, [x2], #16
    ld1 { v7.8H}, [x2], #16

    add   v0.8H,   v0.8H,  v30.8H
    add   v2.8H,   v2.8H,  v30.8H
    add   v4.8H,   v4.8H,  v30.8H
    add   v6.8H,   v6.8H,  v30.8H

    shl   v1.8H,   v1.8H, #slv
    shl   v3.8H,   v3.8H, #slv
    shl   v5.8H,   v5.8H, #slv
    shl   v7.8H,   v7.8H, #slv

    sub   v0.8H,   v0.8H,   v1.8H
    sub   v2.8H,   v2.8H,   v3.8H
    sub   v4.8H,   v4.8H,   v5.8H
    sub   v6.8H,   v6.8H,   v7.8H

    sshr  v0.8H,   v0.8H, #srv
    sshr  v2.8H,   v2.8H, #srv
    sshr  v4.8H,   v4.8H, #srv
    sshr  v6.8H,   v6.8H, #srv

    mov x7, #7
    _enc_add_msg_loop:

    st1 { v0.8H}, [x0], #16
    ld2 { v0.8H,  v1.8H}, [x1], #32
    st1 { v2.8H}, [x0], #16
    ld2 { v2.8H,  v3.8H}, [x1], #32
    st1 { v4.8H}, [x0], #16
    ld2 { v4.8H,  v5.8H}, [x1], #32
    st1 { v6.8H}, [x0], #16
    ld2 { v6.8H,  v7.8H}, [x1], #32
    ld1 { v1.8H}, [x2], #16
    ld1 { v3.8H}, [x2], #16
    ld1 { v5.8H}, [x2], #16
    ld1 { v7.8H}, [x2], #16

    add   v0.8H,   v0.8H,  v30.8H
    add   v2.8H,   v2.8H,  v30.8H
    add   v4.8H,   v4.8H,  v30.8H
    add   v6.8H,   v6.8H,  v30.8H

    shl   v1.8H,   v1.8H, #slv
    shl   v3.8H,   v3.8H, #slv
    shl   v5.8H,   v5.8H, #slv
    shl   v7.8H,   v7.8H, #slv

    sub   v0.8H,   v0.8H,   v1.8H
    sub   v2.8H,   v2.8H,   v3.8H
    sub   v4.8H,   v4.8H,   v5.8H
    sub   v6.8H,   v6.8H,   v7.8H

    sshr  v0.8H,   v0.8H, #srv
    sshr  v2.8H,   v2.8H, #srv
    sshr  v4.8H,   v4.8H, #srv
    sshr  v6.8H,   v6.8H, #srv

    sub x7, x7, #1
    cbnz x7, _enc_add_msg_loop

    st1 { v0.8H}, [x0], #16
    st1 { v2.8H}, [x0], #16
    st1 { v4.8H}, [x0], #16
    st1 { v6.8H}, [x0], #16

    br lr


.align 2
.global PQCLEAN_LIGHTSABER_AARCH64_asm_dec_get_msg
.global _PQCLEAN_LIGHTSABER_AARCH64_asm_dec_get_msg
#ifndef __clang__
.type PQCLEAN_LIGHTSABER_AARCH64_asm_dec_get_msg, %function
#endif
PQCLEAN_LIGHTSABER_AARCH64_asm_dec_get_msg:
_PQCLEAN_LIGHTSABER_AARCH64_asm_dec_get_msg:

    .equ srv, (SABER_EP-1)
    .equ slv, (SABER_EP-SABER_ET)

    dup v30.8H, w3

    ld2 { v0.8H,  v1.8H}, [x1], #32
    ld2 { v2.8H,  v3.8H}, [x1], #32
    ld2 { v4.8H,  v5.8H}, [x1], #32
    ld2 { v6.8H,  v7.8H}, [x1], #32
    ld1 { v1.8H}, [x2], #16
    ld1 { v3.8H}, [x2], #16
    ld1 { v5.8H}, [x2], #16
    ld1 { v7.8H}, [x2], #16

    add   v0.8H,   v0.8H,  v30.8H
    add   v2.8H,   v2.8H,  v30.8H
    add   v4.8H,   v4.8H,  v30.8H
    add   v6.8H,   v6.8H,  v30.8H

    shl   v1.8H,   v1.8H, #slv
    shl   v3.8H,   v3.8H, #slv
    shl   v5.8H,   v5.8H, #slv
    shl   v7.8H,   v7.8H, #slv

    sub   v0.8H,   v0.8H,   v1.8H
    sub   v2.8H,   v2.8H,   v3.8H
    sub   v4.8H,   v4.8H,   v5.8H
    sub   v6.8H,   v6.8H,   v7.8H

    sshr  v0.8H,   v0.8H, #srv
    sshr  v2.8H,   v2.8H, #srv
    sshr  v4.8H,   v4.8H, #srv
    sshr  v6.8H,   v6.8H, #srv

    mov x7, #7
    _dec_get_msg_loop:

    st1 { v0.8H}, [x0], #16
    ld2 { v0.8H,  v1.8H}, [x1], #32
    st1 { v2.8H}, [x0], #16
    ld2 { v2.8H,  v3.8H}, [x1], #32
    st1 { v4.8H}, [x0], #16
    ld2 { v4.8H,  v5.8H}, [x1], #32
    st1 { v6.8H}, [x0], #16
    ld2 { v6.8H,  v7.8H}, [x1], #32
    ld1 { v1.8H}, [x2], #16
    ld1 { v3.8H}, [x2], #16
    ld1 { v5.8H}, [x2], #16
    ld1 { v7.8H}, [x2], #16

    add   v0.8H,   v0.8H,  v30.8H
    add   v2.8H,   v2.8H,  v30.8H
    add   v4.8H,   v4.8H,  v30.8H
    add   v6.8H,   v6.8H,  v30.8H

    shl   v1.8H,   v1.8H, #slv
    shl   v3.8H,   v3.8H, #slv
    shl   v5.8H,   v5.8H, #slv
    shl   v7.8H,   v7.8H, #slv

    sub   v0.8H,   v0.8H,   v1.8H
    sub   v2.8H,   v2.8H,   v3.8H
    sub   v4.8H,   v4.8H,   v5.8H
    sub   v6.8H,   v6.8H,   v7.8H

    sshr  v0.8H,   v0.8H, #srv
    sshr  v2.8H,   v2.8H, #srv
    sshr  v4.8H,   v4.8H, #srv
    sshr  v6.8H,   v6.8H, #srv

    sub x7, x7, #1
    cbnz x7, _dec_get_msg_loop

    st1 { v0.8H}, [x0], #16
    st1 { v2.8H}, [x0], #16
    st1 { v4.8H}, [x0], #16
    st1 { v6.8H}, [x0], #16

    br lr