-rw-r--r-- 13386 lib25519-20260614/crypto_multiscalar/ed25519/arm64-uma10l/ge25519_p1p1_10l_dense_to_p2_10l_dense.S raw
#include "crypto_asm_hidden.h"
// linker define ge25519_p1p1_10l_dense_to_p2_10l_dense
.p2align 4
ASM_HIDDEN _CRYPTO_SHARED_NAMESPACE(ge25519_p1p1_10l_dense_to_p2_10l_dense)
.globl _CRYPTO_SHARED_NAMESPACE(ge25519_p1p1_10l_dense_to_p2_10l_dense)
ASM_HIDDEN CRYPTO_SHARED_NAMESPACE(ge25519_p1p1_10l_dense_to_p2_10l_dense)
.globl CRYPTO_SHARED_NAMESPACE(ge25519_p1p1_10l_dense_to_p2_10l_dense)
_CRYPTO_SHARED_NAMESPACE(ge25519_p1p1_10l_dense_to_p2_10l_dense):
CRYPTO_SHARED_NAMESPACE(ge25519_p1p1_10l_dense_to_p2_10l_dense):
sub sp, sp, #112
stp x19, x20, [sp, #0]
stp x21, x22, [sp, #16]
stp x23, x24, [sp, #32]
stp x25, x26, [sp, #48]
stp x27, x28, [sp, #64]
stp x29, x30, [sp, #80]
stp x0, x1, [sp, #96]
mov w30, #19
// inputs <104,40> and <120,80>>
ldr x11, [sp, #104]
add x12, x11, #40
ld2 {v10.s, v11.s}[0], [x11], #8
ld2 {v10.s, v11.s}[1], [x12], #8
ld2 {v12.s, v13.s}[0], [x11], #8
ld2 {v12.s, v13.s}[1], [x12], #8
ld2 {v14.s, v15.s}[0], [x11], #8
ld2 {v14.s, v15.s}[1], [x12], #8
ld2 {v16.s, v17.s}[0], [x11], #8
ld2 {v16.s, v17.s}[1], [x12], #8
ld2 {v18.s, v19.s}[0], [x11], #8
ld2 {v18.s, v19.s}[1], [x12], #8
ldr x10, [sp, #104]
add x11, x10, #120
add x12, x10, #80
ld2 {v20.s, v21.s}[0], [x11], #8
ld2 {v20.s, v21.s}[1], [x12], #8
ld2 {v22.s, v23.s}[0], [x11], #8
ld2 {v22.s, v23.s}[1], [x12], #8
ld2 {v24.s, v25.s}[0], [x11], #8
ld2 {v24.s, v25.s}[1], [x12], #8
ld2 {v26.s, v27.s}[0], [x11], #8
ld2 {v26.s, v27.s}[1], [x12], #8
ld2 {v28.s, v29.s}[0], [x11], #8
ld2 {v28.s, v29.s}[1], [x12], #8
// <96,40> ← Mul(<104,40>,<120,80>)
umull v0.2d, v10.2s, v20.2s
umull v0.2d, v10.2s, v20.2s
umull v1.2d, v10.2s, v21.2s
umlal v1.2d, v11.2s, v20.2s
umull v2.2d, v10.2s, v22.2s
umlal v2.2d, v12.2s, v20.2s
umull v3.2d, v10.2s, v23.2s
umlal v3.2d, v11.2s, v22.2s
umlal v3.2d, v12.2s, v21.2s
umlal v3.2d, v13.2s, v20.2s
umull v4.2d, v10.2s, v24.2s
umlal v4.2d, v12.2s, v22.2s
umlal v4.2d, v14.2s, v20.2s
umull v5.2d, v10.2s, v25.2s
umlal v5.2d, v11.2s, v24.2s
umlal v5.2d, v12.2s, v23.2s
umlal v5.2d, v13.2s, v22.2s
umlal v5.2d, v14.2s, v21.2s
umlal v5.2d, v15.2s, v20.2s
umull v6.2d, v10.2s, v26.2s
umlal v6.2d, v12.2s, v24.2s
umlal v6.2d, v14.2s, v22.2s
umlal v6.2d, v16.2s, v20.2s
umull v7.2d, v10.2s, v27.2s
umlal v7.2d, v11.2s, v26.2s
umlal v7.2d, v12.2s, v25.2s
umlal v7.2d, v13.2s, v24.2s
umlal v7.2d, v14.2s, v23.2s
umlal v7.2d, v15.2s, v22.2s
umlal v7.2d, v16.2s, v21.2s
umlal v7.2d, v17.2s, v20.2s
umull v8.2d, v10.2s, v28.2s
umlal v8.2d, v12.2s, v26.2s
umlal v8.2d, v14.2s, v24.2s
umlal v8.2d, v16.2s, v22.2s
umlal v8.2d, v18.2s, v20.2s
umull v9.2d, v10.2s, v29.2s
umlal v9.2d, v11.2s, v28.2s
umlal v9.2d, v12.2s, v27.2s
umlal v9.2d, v13.2s, v26.2s
umlal v9.2d, v14.2s, v25.2s
umlal v9.2d, v15.2s, v24.2s
umlal v9.2d, v16.2s, v23.2s
umlal v9.2d, v17.2s, v22.2s
umlal v9.2d, v18.2s, v21.2s
umlal v9.2d, v19.2s, v20.2s
dup v31.2s, w30
mul v22.2s, v22.2s, v31.2s
mul v24.2s, v24.2s, v31.2s
mul v26.2s, v26.2s, v31.2s
mul v28.2s, v28.2s, v31.2s
umlal v0.2d, v12.2s, v28.2s
umlal v0.2d, v14.2s, v26.2s
umlal v0.2d, v16.2s, v24.2s
umlal v0.2d, v18.2s, v22.2s
umlal v1.2d, v13.2s, v28.2s
umlal v1.2d, v15.2s, v26.2s
umlal v1.2d, v17.2s, v24.2s
umlal v1.2d, v19.2s, v22.2s
umlal v2.2d, v14.2s, v28.2s
umlal v2.2d, v16.2s, v26.2s
umlal v2.2d, v18.2s, v24.2s
umlal v3.2d, v15.2s, v28.2s
umlal v3.2d, v17.2s, v26.2s
umlal v3.2d, v19.2s, v24.2s
umlal v4.2d, v16.2s, v28.2s
umlal v4.2d, v18.2s, v26.2s
umlal v5.2d, v17.2s, v28.2s
umlal v5.2d, v19.2s, v26.2s
umlal v6.2d, v18.2s, v28.2s
umlal v7.2d, v19.2s, v28.2s
shl v11.2s, v11.2s, #1
shl v13.2s, v13.2s, #1
shl v15.2s, v15.2s, #1
shl v17.2s, v17.2s, #1
shl v19.2s, v19.2s, #1
umlal v2.2d, v11.2s, v21.2s
umlal v4.2d, v11.2s, v23.2s
umlal v4.2d, v13.2s, v21.2s
umlal v6.2d, v11.2s, v25.2s
umlal v6.2d, v13.2s, v23.2s
umlal v6.2d, v15.2s, v21.2s
umlal v8.2d, v11.2s, v27.2s
umlal v8.2d, v13.2s, v25.2s
umlal v8.2d, v15.2s, v23.2s
umlal v8.2d, v17.2s, v21.2s
mul v21.2s, v21.2s, v31.2s
mul v23.2s, v23.2s, v31.2s
mul v25.2s, v25.2s, v31.2s
mul v27.2s, v27.2s, v31.2s
mul v29.2s, v29.2s, v31.2s
umlal v0.2d, v11.2s, v29.2s
umlal v0.2d, v13.2s, v27.2s
umlal v0.2d, v15.2s, v25.2s
umlal v0.2d, v17.2s, v23.2s
umlal v0.2d, v19.2s, v21.2s
umlal v1.2d, v12.2s, v29.2s
umlal v1.2d, v14.2s, v27.2s
umlal v1.2d, v16.2s, v25.2s
umlal v1.2d, v18.2s, v23.2s
umlal v2.2d, v13.2s, v29.2s
umlal v2.2d, v15.2s, v27.2s
umlal v2.2d, v17.2s, v25.2s
umlal v2.2d, v19.2s, v23.2s
umlal v3.2d, v14.2s, v29.2s
umlal v3.2d, v16.2s, v27.2s
umlal v3.2d, v18.2s, v25.2s
umlal v4.2d, v15.2s, v29.2s
umlal v4.2d, v17.2s, v27.2s
umlal v4.2d, v19.2s, v25.2s
umlal v5.2d, v16.2s, v29.2s
umlal v5.2d, v18.2s, v27.2s
umlal v6.2d, v17.2s, v29.2s
umlal v6.2d, v19.2s, v27.2s
umlal v7.2d, v18.2s, v29.2s
umlal v8.2d, v19.2s, v29.2s
mov w29, #0x03ffffff
dup v30.2d, x29
ushr v25.2d, v30.2d, #1
usra v6.2d, v5.2d, #25
and v5.16b, v5.16b, v25.16b
usra v1.2d, v0.2d, #26
and v0.16b, v0.16b, v30.16b
usra v7.2d, v6.2d, #26
and v6.16b, v6.16b, v30.16b
usra v2.2d, v1.2d, #25
and v1.16b, v1.16b, v25.16b
usra v8.2d, v7.2d, #25
and v7.16b, v7.16b, v25.16b
usra v3.2d, v2.2d, #26
and v2.16b, v2.16b, v30.16b
usra v9.2d, v8.2d, #26
and v8.16b, v8.16b, v30.16b
usra v4.2d, v3.2d, #25
and v3.16b, v3.16b, v25.16b
bic v10.16b, v9.16b, v25.16b
usra v0.2d, v10.2d, #25
usra v0.2d, v10.2d, #24
usra v0.2d, v10.2d, #21
and v9.16b, v9.16b, v25.16b
usra v5.2d, v4.2d, #26
and v4.16b, v4.16b, v30.16b
usra v1.2d, v0.2d, #26
and v0.16b, v0.16b, v30.16b
usra v6.2d, v5.2d, #25
and v5.16b, v5.16b, v25.16b
ldr x11, [sp, #96]
add x12, x11, #40
st2 {v0.s, v1.s}[0], [x11], #8
st2 {v0.s, v1.s}[2], [x12], #8
st2 {v2.s, v3.s}[0], [x11], #8
st2 {v2.s, v3.s}[2], [x12], #8
st2 {v4.s, v5.s}[0], [x11], #8
st2 {v4.s, v5.s}[2], [x12], #8
st2 {v6.s, v7.s}[0], [x11], #8
st2 {v6.s, v7.s}[2], [x12], #8
st2 {v8.s, v9.s}[0], [x11], #8
st2 {v8.s, v9.s}[2], [x12], #8
// mul
ldr x1, [sp, #104]
ldp w10, w11, [x1, #40]
ldp w12, w13, [x1, #48]
ldp w14, w15, [x1, #56]
ldp w16, w17, [x1, #64]
ldp w18, w19, [x1, #72]
ldp w20, w21, [x1, #120]
ldp w22, w23, [x1, #128]
ldp w24, w25, [x1, #136]
ldp w26, w27, [x1, #144]
ldp w28, w29, [x1, #152]
umull x0, w10, w20
umull x1, w10, w21
umull x2, w10, w22
umull x3, w10, w23
umull x4, w10, w24
umull x5, w10, w25
umull x6, w10, w26
umull x7, w10, w27
umull x8, w10, w28
umull x9, w10, w29
umaddl x1, w11, w20, x1
umaddl x3, w11, w22, x3
umaddl x5, w11, w24, x5
umaddl x7, w11, w26, x7
umaddl x9, w11, w28, x9
umaddl x2, w12, w20, x2
umaddl x3, w12, w21, x3
umaddl x4, w12, w22, x4
umaddl x5, w12, w23, x5
umaddl x6, w12, w24, x6
umaddl x7, w12, w25, x7
umaddl x8, w12, w26, x8
umaddl x9, w12, w27, x9
umaddl x3, w13, w20, x3
umaddl x5, w13, w22, x5
umaddl x7, w13, w24, x7
umaddl x9, w13, w26, x9
umaddl x4, w14, w20, x4
umaddl x5, w14, w21, x5
umaddl x6, w14, w22, x6
umaddl x7, w14, w23, x7
umaddl x8, w14, w24, x8
umaddl x9, w14, w25, x9
umaddl x5, w15, w20, x5
umaddl x7, w15, w22, x7
umaddl x9, w15, w24, x9
umaddl x6, w16, w20, x6
umaddl x7, w16, w21, x7
umaddl x8, w16, w22, x8
umaddl x9, w16, w23, x9
umaddl x7, w17, w20, x7
umaddl x9, w17, w22, x9
umaddl x8, w18, w20, x8
umaddl x9, w18, w21, x9
umaddl x9, w19, w20, x9
mul w22, w22, w30
mul w24, w24, w30
mul w26, w26, w30
mul w28, w28, w30
umaddl x0, w12, w28, x0
umaddl x1, w13, w28, x1
umaddl x0, w14, w26, x0
umaddl x2, w14, w28, x2
umaddl x1, w15, w26, x1
umaddl x3, w15, w28, x3
umaddl x0, w16, w24, x0
umaddl x2, w16, w26, x2
umaddl x4, w16, w28, x4
umaddl x1, w17, w24, x1
umaddl x3, w17, w26, x3
umaddl x5, w17, w28, x5
umaddl x0, w18, w22, x0
umaddl x2, w18, w24, x2
umaddl x6, w18, w28, x6
umaddl x4, w18, w26, x4
umaddl x1, w19, w22, x1
umaddl x3, w19, w24, x3
umaddl x5, w19, w26, x5
umaddl x7, w19, w28, x7
add w11, w11, w11
umaddl x2, w11, w21, x2
umaddl x4, w11, w23, x4
umaddl x6, w11, w25, x6
umaddl x8, w11, w27, x8
add w13, w13, w13
umaddl x4, w13, w21, x4
umaddl x6, w13, w23, x6
umaddl x8, w13, w25, x8
add w15, w15, w15
umaddl x6, w15, w21, x6
umaddl x8, w15, w23, x8
add w17, w17, w17
umaddl x8, w17, w21, x8
mul w21, w21, w30
mul w23, w23, w30
mul w25, w25, w30
mul w27, w27, w30
mul w29, w29, w30
umaddl x0, w11, w29, x0
umaddl x1, w12, w29, x1
umaddl x0, w13, w27, x0
umaddl x2, w13, w29, x2
umaddl x1, w14, w27, x1
umaddl x3, w14, w29, x3
umaddl x0, w15, w25, x0
umaddl x2, w15, w27, x2
umaddl x4, w15, w29, x4
umaddl x1, w16, w25, x1
umaddl x3, w16, w27, x3
umaddl x5, w16, w29, x5
umaddl x0, w17, w23, x0
umaddl x2, w17, w25, x2
umaddl x4, w17, w27, x4
umaddl x6, w17, w29, x6
umaddl x1, w18, w23, x1
umaddl x3, w18, w25, x3
umaddl x5, w18, w27, x5
umaddl x7, w18, w29, x7
add w19, w19, w19
umaddl x0, w19, w21, x0
umaddl x2, w19, w23, x2
umaddl x4, w19, w25, x4
umaddl x6, w19, w27, x6
umaddl x8, w19, w29, x8
add x6, x6, x5, lsr #25
and x5, x5, #0x1ffffff
add x1, x1, x0, lsr #26
and x0, x0, #0x3ffffff
add x7, x7, x6, lsr #26
and x6, x6, #0x3ffffff
add x2, x2, x1, lsr #25
and x1, x1, #0x1ffffff
add x8, x8, x7, lsr #25
and x7, x7, #0x1ffffff
add x3, x3, x2, lsr #26
and x2, x2, #0x3ffffff
add x9, x9, x8, lsr #26
and x8, x8, #0x3ffffff
add x4, x4, x3, lsr #25
and x3, x3, #0x1ffffff
bfi x2, x3, #32, #25
bic x10, x9, #0x1ffffff
add x0, x0, x10, lsr #25
add x0, x0, x10, lsr #24
add x0, x0, x10, lsr #21
and x9, x9, #0x1ffffff
bfi x8, x9, #32, #25
add x5, x5, x4, lsr #26
and x4, x4, #0x3ffffff
add x1, x1, x0, lsr #26
and x0, x0, #0x3ffffff
bfi x0, x1, #32, #26
add x6, x6, x5, lsr #25
bfi x6, x7, #32, #25
and x5, x5, #0x1ffffff
bfi x4, x5, #32, #25
ldr x10, [sp, #96]
stp x0, x2, [x10, #80]
stp x4, x6, [x10, #96]
str x8, [x10, #112]
ldp x29, x30, [sp, #80]
ldp x27, x28, [sp, #64]
ldp x25, x26, [sp, #48]
ldp x23, x24, [sp, #32]
ldp x21, x22, [sp, #16]
ldp x19, x20, [sp, #0]
add sp, sp, #112
ret
.section .note.GNU-stack,"",@progbits