#include "crypto_asm_hidden.h" // linker define ge25519_multi_scalarmult_process /* Assembly for multi scalar multiplication */ .p2align 4 ASM_HIDDEN _CRYPTO_SHARED_NAMESPACE(ge25519_multi_scalarmult_process) .globl _CRYPTO_SHARED_NAMESPACE(ge25519_multi_scalarmult_process) ASM_HIDDEN CRYPTO_SHARED_NAMESPACE(ge25519_multi_scalarmult_process) .globl CRYPTO_SHARED_NAMESPACE(ge25519_multi_scalarmult_process) _CRYPTO_SHARED_NAMESPACE(ge25519_multi_scalarmult_process): CRYPTO_SHARED_NAMESPACE(ge25519_multi_scalarmult_process): sub sp, sp, #816 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] movz x21, #0xffda movk x21, #0x07ff, lsl 16 movk x21, #0xfffe, lsl 32 movk x21, #0x03ff, lsl 48 movz x22, #0xfffe movk x22, #0x07ff, lsl 16 movk x22, #0xfffe, lsl 32 movk x22, #0x03ff, lsl 48 stp x21, x22, [sp, #192] movz x21, #0xf159 movk x21, #0x02b2, lsl 16 movk x21, #0xe509, lsl 32 movk x21, #0x01a6, lsl 48 movz x22, #0xdd7a movk x22, #0x022a, lsl 16 movk x22, #0x141d, lsl 32 movk x22, #0x00d4, lsl 48 movz x23, #0x8052 movk x23, #0x0003, lsl 16 movk x23, #0xd130, lsl 32 movk x23, #0x00f3, lsl 48 movz x24, #0x7977 movk x24, #0x0340, lsl 16 movk x24, #0xe331, lsl 32 movk x24, #0x019c, lsl 48 movz x25, #0x6dff movk x25, #0x01c5, lsl 16 movk x25, #0x1b67, lsl 32 movk x25, #0x0090, lsl 48 stp x21, x22, [sp, #208] stp x23, x24, [sp, #224] str x25, [sp, #240] mov x27, #1 stp xzr, xzr, [sp, #248] stp xzr, xzr, [sp, #264] str xzr, [sp, #280] stp x27, xzr, [sp, #288] stp xzr, xzr, [sp, #304] str xzr, [sp, #320] stp x27, xzr, [sp, #328] stp xzr, xzr, [sp, #344] str xzr, [sp, #360] stp xzr, xzr, [sp, #368] stp xzr, xzr, [sp, #384] str xzr, [sp, #400] stp x0, x1, [sp, #96] stp x2, x3, [sp, #112] mov x25, xzr mov x26, #160 mul x26, x26, x4 mov x30, #255 add x29, x1, x30 stp x25, x26, [sp, #160] stp x29, x30, [sp, #176] mov w30, #19 .L1: ldr x25, [sp, #160] cmp x25, xzr beq .L2 /* dbl p1p1 */ // square add x29, sp, #328 ldp w10, w11, [x29, #0] ldp w12, w13, [x29, #8] ldp w14, w15, [x29, #16] ldp w16, w17, [x29, #24] ldp w18, w19, [x29, #32] mul w20, w16, w30 mul w21, w18, w30 add w25, w30, w30 mul w22, w15, w25 mul w23, w17, w25 mul w24, w19, w25 umull x0, w10, w10 add w25, w10, w10 umull x1, w25, w11 umull x2, w25, w12 umull x3, w25, w13 umull x4, w25, w14 umull x5, w25, w15 umull x6, w25, w16 umull x7, w25, w17 umull x8, w25, w18 umaddl x4, w12, w12, x4 umaddl x8, w14, w14, x8 add w10, w15, w15 umaddl x1, w20, w10, x1 umaddl x2, w20, w16, x2 add w9, w17, w17 umaddl x3, w21, w10, x3 umaddl x5, w21, w9, x5 umaddl x6, w21, w18, x6 umaddl x0, w22, w15, x0 umaddl x1, w23, w14, x1 umaddl x2, w23, w10, x2 umaddl x3, w23, w16, x3 umaddl x4, w23, w17, x4 umaddl x1, w24, w12, x1 umaddl x3, w24, w14, x3 umaddl x4, w24, w10, x4 umaddl x5, w24, w16, x5 umaddl x6, w24, w9, x6 umaddl x7, w24, w18, x7 umaddl x8, w24, w19, x8 add w26, w11, w11 umaddl x0, w26, w24, x0 umaddl x2, w26, w11, x2 umaddl x3, w26, w12, x3 umaddl x5, w26, w14, x5 umaddl x6, w26, w10, x6 umaddl x7, w26, w16, x7 umaddl x8, w26, w9, x8 umull x9, w25, w19 umaddl x9, w26, w18, x9 add w27, w12, w12 umaddl x0, w27, w21, x0 umaddl x5, w27, w13, x5 umaddl x8, w27, w16, x8 umaddl x6, w27, w14, x6 umaddl x7, w27, w15, x7 umaddl x9, w27, w17, x9 add w28, w13, w13 umaddl x4, w26, w28, x4 umaddl x0, w28, w23, x0 umaddl x1, w28, w21, x1 umaddl x2, w28, w24, x2 umaddl x8, w28, w10, x8 umaddl x6, w28, w13, x6 umaddl x7, w28, w14, x7 umaddl x9, w28, w16, x9 add w29, w14, w14 umaddl x0, w29, w20, x0 umaddl x2, w29, w21, x2 umaddl x9, w29, w15, x9 add w18, w16, w16 umaddl x4, w18, w21, x4 // double and then reduce add x0, x0, x0 add x1, x1, x1 add x2, x2, x2 add x3, x3, x3 add x4, x4, x4 add x5, x5, x5 add x6, x6, x6 add x7, x7, x7 add x8, x8, x8 add x9, x9, x9 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 add x29, sp, #768 stp x0, x2, [x29, #0] stp x4, x6, [x29, #16] str x8, [x29, #32] // square add x29, sp, #288 ldp w10, w11, [x29, #0] ldp w12, w13, [x29, #8] ldp w14, w15, [x29, #16] ldp w16, w17, [x29, #24] ldp w18, w19, [x29, #32] mul w20, w16, w30 mul w21, w18, w30 add w25, w30, w30 mul w22, w15, w25 mul w23, w17, w25 mul w24, w19, w25 umull x0, w10, w10 add w25, w10, w10 umull x1, w25, w11 umull x2, w25, w12 umull x3, w25, w13 umull x4, w25, w14 umull x5, w25, w15 umull x6, w25, w16 umull x7, w25, w17 umull x8, w25, w18 umaddl x4, w12, w12, x4 umaddl x8, w14, w14, x8 add w10, w15, w15 umaddl x1, w20, w10, x1 umaddl x2, w20, w16, x2 add w9, w17, w17 umaddl x3, w21, w10, x3 umaddl x5, w21, w9, x5 umaddl x6, w21, w18, x6 umaddl x0, w22, w15, x0 umaddl x1, w23, w14, x1 umaddl x2, w23, w10, x2 umaddl x3, w23, w16, x3 umaddl x4, w23, w17, x4 umaddl x1, w24, w12, x1 umaddl x3, w24, w14, x3 umaddl x4, w24, w10, x4 umaddl x5, w24, w16, x5 umaddl x6, w24, w9, x6 umaddl x7, w24, w18, x7 umaddl x8, w24, w19, x8 add w26, w11, w11 umaddl x0, w26, w24, x0 umaddl x2, w26, w11, x2 umaddl x3, w26, w12, x3 umaddl x5, w26, w14, x5 umaddl x6, w26, w10, x6 umaddl x7, w26, w16, x7 umaddl x8, w26, w9, x8 umull x9, w25, w19 umaddl x9, w26, w18, x9 add w27, w12, w12 umaddl x0, w27, w21, x0 umaddl x5, w27, w13, x5 umaddl x8, w27, w16, x8 umaddl x6, w27, w14, x6 umaddl x7, w27, w15, x7 umaddl x9, w27, w17, x9 add w28, w13, w13 umaddl x4, w26, w28, x4 umaddl x0, w28, w23, x0 umaddl x1, w28, w21, x1 umaddl x2, w28, w24, x2 umaddl x8, w28, w10, x8 umaddl x6, w28, w13, x6 umaddl x7, w28, w14, x7 umaddl x9, w28, w16, x9 add w29, w14, w14 umaddl x0, w29, w20, x0 umaddl x2, w29, w21, x2 umaddl x9, w29, w15, x9 add w18, w16, w16 umaddl x4, w18, w21, x4 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 add x29, sp, #728 stp x0, x2, [x29, #0] stp x4, x6, [x29, #16] str x8, [x29, #32] // square add x29, sp, #248 ldp w10, w11, [x29, #0] ldp w12, w13, [x29, #8] ldp w14, w15, [x29, #16] ldp w16, w17, [x29, #24] ldp w18, w19, [x29, #32] mul w20, w16, w30 mul w21, w18, w30 add w25, w30, w30 mul w22, w15, w25 mul w23, w17, w25 mul w24, w19, w25 umull x0, w10, w10 add w25, w10, w10 umull x1, w25, w11 umull x2, w25, w12 umull x3, w25, w13 umull x4, w25, w14 umull x5, w25, w15 umull x6, w25, w16 umull x7, w25, w17 umull x8, w25, w18 umaddl x4, w12, w12, x4 umaddl x8, w14, w14, x8 add w10, w15, w15 umaddl x1, w20, w10, x1 umaddl x2, w20, w16, x2 add w9, w17, w17 umaddl x3, w21, w10, x3 umaddl x5, w21, w9, x5 umaddl x6, w21, w18, x6 umaddl x0, w22, w15, x0 umaddl x1, w23, w14, x1 umaddl x2, w23, w10, x2 umaddl x3, w23, w16, x3 umaddl x4, w23, w17, x4 umaddl x1, w24, w12, x1 umaddl x3, w24, w14, x3 umaddl x4, w24, w10, x4 umaddl x5, w24, w16, x5 umaddl x6, w24, w9, x6 umaddl x7, w24, w18, x7 umaddl x8, w24, w19, x8 add w26, w11, w11 umaddl x0, w26, w24, x0 umaddl x2, w26, w11, x2 umaddl x3, w26, w12, x3 umaddl x5, w26, w14, x5 umaddl x6, w26, w10, x6 umaddl x7, w26, w16, x7 umaddl x8, w26, w9, x8 umull x9, w25, w19 umaddl x9, w26, w18, x9 add w27, w12, w12 umaddl x0, w27, w21, x0 umaddl x5, w27, w13, x5 umaddl x8, w27, w16, x8 umaddl x6, w27, w14, x6 umaddl x7, w27, w15, x7 umaddl x9, w27, w17, x9 add w28, w13, w13 umaddl x4, w26, w28, x4 umaddl x0, w28, w23, x0 umaddl x1, w28, w21, x1 umaddl x2, w28, w24, x2 umaddl x8, w28, w10, x8 umaddl x6, w28, w13, x6 umaddl x7, w28, w14, x7 umaddl x9, w28, w16, x9 add w29, w14, w14 umaddl x0, w29, w20, x0 umaddl x2, w29, w21, x2 umaddl x9, w29, w15, x9 add w18, w16, w16 umaddl x4, w18, w21, x4 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 // neg ldp x21, x22, [sp, #192] sub x11, x21, x0 sub x12, x22, x2 sub x18, x22, x4 sub x19, x22, x6 sub x20, x22, x8 // add add x29, sp, #608 ldp x13, x14, [x29, #120] ldp x15, x16, [x29, #136] ldr x17, [x29, #152] add x0, x11, x13 add x2, x12, x14 add x4, x18, x15 add x6, x19, x16 add x8, x20, x17 stp x0, x2, [sp, #448] stp x4, x6, [sp, #464] str x8, [sp, #480] // sub ldp x13, x14, [x29, #160] ldp x15, x16, [x29, #176] ldr x17, [x29, #192] sub x0, x0, x13 sub x2, x2, x14 sub x4, x4, x15 sub x6, x6, x16 sub x8, x8, x17 stp x0, x2, [x29, #0] stp x4, x6, [x29, #16] str x8, [x29, #32] // sub ldp x13, x14, [x29, #120] ldp x15, x16, [x29, #136] ldr x17, [x29, #152] movz x21, #0xffb4 movk x21, #0x0fff, lsl 16 movk x21, #0xfffc, lsl 32 movk x21, #0x07ff, lsl 48 movz x23, #0xfffc movk x23, #0x0fff, lsl 16 movk x23, #0xfffc, lsl 32 movk x23, #0x07ff, lsl 48 add x11, x11, x21 add x12, x12, x23 add x18, x18, x23 add x19, x19, x23 add x20, x20, x23 sub x0, x11, x13 sub x2, x12, x14 sub x4, x18, x15 sub x6, x19, x16 sub x8, x20, x17 lsr x1, x0, #32 mov w0, w0 lsr x3, x2, #32 mov w2, w2 lsr x5, x4, #32 mov w4, w4 lsr x7, x6, #32 mov w6, w6 lsr x9, x8, #32 mov w8, w8 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 add x29, sp, #528 stp x0, x2, [x29, #0] stp x4, x6, [x29, #16] str x8, [x29, #32] // add ldp x10, x12, [sp, #248] ldp x14, x16, [sp, #264] ldr x18, [sp, #280] ldp x20, x22, [sp, #288] ldp x24, x26, [sp, #304] ldr x28, [sp, #320] add x10, x10, x20 add x12, x12, x22 add x14, x14, x24 add x16, x16, x26 add x18, x18, x28 // square lsr x11, x10, #32 lsr x13, x12, #32 lsr x15, x14, #32 lsr x17, x16, #32 lsr x19, x18, #32 mul w20, w16, w30 mul w21, w18, w30 add w25, w30, w30 mul w22, w15, w25 mul w23, w17, w25 mul w24, w19, w25 umull x0, w10, w10 add w25, w10, w10 umull x1, w25, w11 umull x2, w25, w12 umull x3, w25, w13 umull x4, w25, w14 umull x5, w25, w15 umull x6, w25, w16 umull x7, w25, w17 umull x8, w25, w18 umaddl x4, w12, w12, x4 umaddl x8, w14, w14, x8 add w10, w15, w15 umaddl x1, w20, w10, x1 umaddl x2, w20, w16, x2 add w9, w17, w17 umaddl x3, w21, w10, x3 umaddl x5, w21, w9, x5 umaddl x6, w21, w18, x6 umaddl x0, w22, w15, x0 umaddl x1, w23, w14, x1 umaddl x2, w23, w10, x2 umaddl x3, w23, w16, x3 umaddl x4, w23, w17, x4 umaddl x1, w24, w12, x1 umaddl x3, w24, w14, x3 umaddl x4, w24, w10, x4 umaddl x5, w24, w16, x5 umaddl x6, w24, w9, x6 umaddl x7, w24, w18, x7 umaddl x8, w24, w19, x8 add w26, w11, w11 umaddl x0, w26, w24, x0 umaddl x2, w26, w11, x2 umaddl x3, w26, w12, x3 umaddl x5, w26, w14, x5 umaddl x6, w26, w10, x6 umaddl x7, w26, w16, x7 umaddl x8, w26, w9, x8 umull x9, w25, w19 umaddl x9, w26, w18, x9 add w27, w12, w12 umaddl x0, w27, w21, x0 umaddl x5, w27, w13, x5 umaddl x8, w27, w16, x8 umaddl x6, w27, w14, x6 umaddl x7, w27, w15, x7 umaddl x9, w27, w17, x9 add w28, w13, w13 umaddl x4, w26, w28, x4 umaddl x0, w28, w23, x0 umaddl x1, w28, w21, x1 umaddl x2, w28, w24, x2 umaddl x8, w28, w10, x8 umaddl x6, w28, w13, x6 umaddl x7, w28, w14, x7 umaddl x9, w28, w16, x9 add w29, w14, w14 umaddl x0, w29, w20, x0 umaddl x2, w29, w21, x2 umaddl x9, w29, w15, x9 add w18, w16, w16 umaddl x4, w18, w21, x4 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 // add add x29, sp, #408 ldp x13, x14, [x29, #120] ldp x15, x16, [x29, #136] ldr x17, [x29, #152] add x10, x0, x13 add x12, x2, x14 add x14, x4, x15 add x16, x6, x16 add x18, x8, x17 stp x10, x12, [x29, #0] stp x14, x16, [x29, #16] str x18, [x29, #32] /* p1p1 to p3 */ // mul ldp w20, w21, [x29, #200] ldp w22, w23, [x29, #208] ldp w24, w25, [x29, #216] ldp w26, w27, [x29, #224] ldp w28, w29, [x29, #232] lsr x11, x10, #32 lsr x13, x12, #32 lsr x15, x14, #32 lsr x17, x16, #32 lsr x19, x18, #32 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 stp x0, x2, [sp, #248] stp x4, x6, [sp, #264] str x8, [sp, #280] // mul add x29, sp, #448 ldp w10, w11, [x29, #0] ldp w12, w13, [x29, #8] ldp w14, w15, [x29, #16] ldp w16, w17, [x29, #24] ldp w18, w19, [x29, #32] ldp w20, w21, [x29, #80] ldp w22, w23, [x29, #88] ldp w24, w25, [x29, #96] ldp w26, w27, [x29, #104] ldp w28, w29, [x29, #112] 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 stp x0, x2, [sp, #288] stp x4, x6, [sp, #304] str x8, [sp, #320] // mul add x29, sp, #448 ldp w10, w11, [x29, #0] ldp w12, w13, [x29, #8] ldp w14, w15, [x29, #16] ldp w16, w17, [x29, #24] ldp w18, w19, [x29, #32] ldp w20, w21, [x29, #160] ldp w22, w23, [x29, #168] ldp w24, w25, [x29, #176] ldp w26, w27, [x29, #184] ldp w28, w29, [x29, #192] 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 stp x0, x2, [sp, #328] stp x4, x6, [sp, #344] str x8, [sp, #360] // mul add x29, sp, #408 ldp w10, w11, [x29, #0] ldp w12, w13, [x29, #8] ldp w14, w15, [x29, #16] ldp w16, w17, [x29, #24] ldp w18, w19, [x29, #32] ldp w20, w21, [x29, #120] ldp w22, w23, [x29, #128] ldp w24, w25, [x29, #136] ldp w26, w27, [x29, #144] ldp w28, w29, [x29, #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 stp x0, x2, [sp, #368] stp x4, x6, [sp, #384] str x8, [sp, #400] .L2: str xzr, [sp, #128] ldr x1, [sp, #176] str x1, [sp, #136] .L3: ldrsb w14, [x1, #0] cmp w14, wzr bne .L4 add x1, x1, #256 str x1, [sp, #136] ldr x3, [sp, #128] add x3, x3, #1 str x3, [sp, #128] ldr x4, [sp, #120] cmp x3, x4 blt .L3 ldp x28, x29, [sp, #176] sub x28, x28, #1 sub x29, x29, #1 stp x28, x29, [sp, #176] cmp x29, xzr bge .L1 b .L8 .L4: mov x25, #1 str x25, [sp, #160] str w14, [sp, #144] ldr x0, [sp, #112] ldr x9, [sp, #128] ldr x26, [sp, #168] mul x8, x26, x9 add x0, x0, x8 str x0, [sp, #152] ldrsb w14, [sp, #144] cmp w14, wzr blt .L5 lsr w14, w14, #1 mov x9, #160 mul x14, x14, x9 add x0, x0, x14 str x0, [sp, #152] /* add p1p1 */ // sub ldp x3, x4, [sp, #248] ldp x5, x6, [sp, #264] ldr x7, [sp, #280] ldp x13, x14, [sp, #288] ldp x15, x16, [sp, #304] ldr x17, [sp, #320] ldp x21, x22, [sp, #192] add x8, x13, x21 add x9, x14, x22 add x10, x15, x22 add x11, x16, x22 add x12, x17, x22 sub x8, x8, x3 sub x9, x9, x4 sub x10, x10, x5 sub x11, x11, x6 sub x12, x12, x7 stp x8, x9, [sp, #408] stp x10, x11, [sp, #424] str x12, [sp, #440] // add add x3, x3, x13 add x4, x4, x14 add x5, x5, x15 add x6, x6, x16 add x7, x7, x17 add x29, sp, #488 stp x3, x4, [x29, #0] stp x5, x6, [x29, #16] str x7, [x29, #32] // add ldr x0, [sp, #152] ldp x1, x3, [x0, #0] ldp x5, x7, [x0, #16] ldr x9, [x0, #32] ldp x10, x12, [x0, #40] ldp x14, x16, [x0, #56] ldr x18, [x0, #72] add x0, x10, x1 add x2, x12, x3 add x4, x14, x5 add x6, x16, x7 add x8, x18, x9 stp x0, x2, [x29, #160] stp x4, x6, [x29, #176] str x8, [x29, #192] // sub ldp x21, x22, [sp, #192] add x10, x10, x21 add x12, x12, x22 add x14, x14, x22 add x16, x16, x22 add x18, x18, x22 sub x10, x10, x1 sub x12, x12, x3 sub x14, x14, x5 sub x16, x16, x7 sub x18, x18, x9 // mul add x29, sp, #408 ldp w20, w21, [x29, #0] ldp w22, w23, [x29, #8] ldp w24, w25, [x29, #16] ldp w26, w27, [x29, #24] ldp w28, w29, [x29, #32] lsr x11, x10, #32 lsr x13, x12, #32 lsr x15, x14, #32 lsr x17, x16, #32 lsr x19, x18, #32 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 stp x0, x2, [sp, #408] stp x4, x6, [sp, #424] str x8, [sp, #440] // mul add x29, sp, #488 ldp w10, w11, [x29, #0] ldp w12, w13, [x29, #8] ldp w14, w15, [x29, #16] ldp w16, w17, [x29, #24] ldp w18, w19, [x29, #32] ldp w20, w21, [x29, #160] ldp w22, w23, [x29, #168] ldp w24, w25, [x29, #176] ldp w26, w27, [x29, #184] ldp w28, w29, [x29, #192] 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 // add ldp x13, x14, [sp, #408] ldp x15, x16, [sp, #424] ldr x17, [sp, #440] add x1, x0, x13 add x3, x2, x14 add x5, x4, x15 add x7, x6, x16 add x9, x8, x17 add x29, sp, #368 stp x1, x3, [x29, #280] stp x5, x7, [x29, #296] str x9, [x29, #312] // sub ldp x21, x22, [sp, #192] add x0, x0, x21 add x2, x2, x22 add x4, x4, x22 add x6, x6, x22 add x8, x8, x22 sub x0, x0, x13 sub x2, x2, x14 sub x4, x4, x15 sub x6, x6, x16 sub x8, x8, x17 stp x0, x2, [x29, #120] stp x4, x6, [x29, #136] str x8, [x29, #152] // mul ldr x0, [sp, #152] ldp w10, w11, [x0, #120] ldp w12, w13, [x0, #128] ldp w14, w15, [x0, #136] ldp w16, w17, [x0, #144] ldp w18, w19, [x0, #152] ldp w20, w21, [x29, #0] ldp w22, w23, [x29, #8] ldp w24, w25, [x29, #16] ldp w26, w27, [x29, #24] ldp w28, w29, [x29, #32] 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 x17, x7, #0x1ffffff add x3, x3, x2, lsr #26 and x12, x2, #0x3ffffff add x9, x9, x8, lsr #26 and x18, x8, #0x3ffffff add x4, x4, x3, lsr #25 and x13, x3, #0x1ffffff bic x10, x9, #0x1ffffff add x0, x0, x10, lsr #25 add x0, x0, x10, lsr #24 add x0, x0, x10, lsr #21 and x19, x9, #0x1ffffff add x5, x5, x4, lsr #26 and x14, x4, #0x3ffffff add x11, x1, x0, lsr #26 and x10, x0, #0x3ffffff add x16, x6, x5, lsr #25 and x15, x5, #0x1ffffff // mul ldp w20, w21, [sp, #208] ldp w22, w23, [sp, #216] ldp w24, w25, [sp, #224] ldp w26, w27, [sp, #232] ldp w28, w29, [sp, #240] 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 stp x0, x2, [sp, #408] stp x4, x6, [sp, #424] str x8, [sp, #440] // mul ldr x0, [sp, #152] ldp w10, w11, [x0, #80] ldp w12, w13, [x0, #88] ldp w14, w15, [x0, #96] ldp w16, w17, [x0, #104] ldp w18, w19, [x0, #112] add x29, sp, #328 ldp w20, w21, [x29, #0] ldp w22, w23, [x29, #8] ldp w24, w25, [x29, #16] ldp w26, w27, [x29, #24] ldp w28, w29, [x29, #32] 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 // double and then reduce add x0, x0, x0 add x1, x1, x1 add x2, x2, x2 add x3, x3, x3 add x4, x4, x4 add x5, x5, x5 add x6, x6, x6 add x7, x7, x7 add x8, x8, x8 add x9, x9, x9 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 // add ldp x13, x14, [sp, #408] ldp x15, x16, [sp, #424] ldr x17, [sp, #440] add x1, x0, x13 add x3, x2, x14 add x5, x4, x15 add x7, x6, x16 add x9, x8, x17 add x29, sp, #488 stp x1, x3, [x29, #80] stp x5, x7, [x29, #96] str x9, [x29, #112] // sub ldp x21, x22, [sp, #192] add x0, x0, x21 add x2, x2, x22 add x4, x4, x22 add x6, x6, x22 add x8, x8, x22 sub x20, x0, x13 sub x22, x2, x14 sub x24, x4, x15 sub x26, x6, x16 sub x28, x8, x17 stp x20, x22, [x29, #200] stp x24, x26, [x29, #216] str x28, [x29, #232] /* p1p1 to p3 */ // mul ldp w10, w11, [x29, #0] ldp w12, w13, [x29, #8] ldp w14, w15, [x29, #16] ldp w16, w17, [x29, #24] ldp w18, w19, [x29, #32] lsr x21, x20, #32 lsr x23, x22, #32 lsr x25, x24, #32 lsr x27, x26, #32 lsr x29, x28, #32 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 stp x0, x2, [sp, #248] stp x4, x6, [sp, #264] str x8, [sp, #280] // mul add x29, sp, #568 ldp w10, w11, [x29, #0] ldp w12, w13, [x29, #8] ldp w14, w15, [x29, #16] ldp w16, w17, [x29, #24] ldp w18, w19, [x29, #32] ldp w20, w21, [x29, #80] ldp w22, w23, [x29, #88] ldp w24, w25, [x29, #96] ldp w26, w27, [x29, #104] ldp w28, w29, [x29, #112] 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 stp x0, x2, [sp, #288] stp x4, x6, [sp, #304] str x8, [sp, #320] // mul add x29, sp, #568 ldp w10, w11, [x29, #0] ldp w12, w13, [x29, #8] ldp w14, w15, [x29, #16] ldp w16, w17, [x29, #24] ldp w18, w19, [x29, #32] ldp w20, w21, [x29, #120] ldp w22, w23, [x29, #128] ldp w24, w25, [x29, #136] ldp w26, w27, [x29, #144] ldp w28, w29, [x29, #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 stp x0, x2, [sp, #328] stp x4, x6, [sp, #344] str x8, [sp, #360] // mul add x29, sp, #488 ldp w10, w11, [x29, #0] ldp w12, w13, [x29, #8] ldp w14, w15, [x29, #16] ldp w16, w17, [x29, #24] ldp w18, w19, [x29, #32] ldp w20, w21, [x29, #160] ldp w22, w23, [x29, #168] ldp w24, w25, [x29, #176] ldp w26, w27, [x29, #184] ldp w28, w29, [x29, #192] 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 stp x0, x2, [sp, #368] stp x4, x6, [sp, #384] str x8, [sp, #400] b .L6 .L5: mov w15, wzr sub w15, w15, w14 lsr w15, w15, #1 mov x9, #160 mul x15, x15, x9 add x0, x0, x15 str x0, [sp, #152] /* sub p1p1 */ // neg ldp x7, x8, [x0, #0] ldp x9, x10, [x0, #16] ldr x11, [x0, #32] ldp x21, x22, [sp, #192] sub x7, x21, x7 sub x8, x22, x8 sub x9, x22, x9 sub x10, x22, x10 sub x11, x22, x11 add x29, sp, #728 stp x7, x8, [x29, #0] stp x9, x10, [x29, #16] str x11, [x29, #32] // neg ldp x7, x8, [x0, #120] ldp x9, x10, [x0, #136] ldr x11, [x0, #152] sub x7, x21, x7 sub x8, x22, x8 sub x9, x22, x9 sub x10, x22, x10 sub x11, x22, x11 stp x7, x8, [x29, #40] stp x9, x10, [x29, #56] str x11, [x29, #72] // sub ldp x3, x4, [sp, #248] ldp x5, x6, [sp, #264] ldr x7, [sp, #280] ldp x13, x14, [sp, #288] ldp x15, x16, [sp, #304] ldr x17, [sp, #320] add x8, x13, x21 add x9, x14, x22 add x10, x15, x22 add x11, x16, x22 add x12, x17, x22 sub x8, x8, x3 sub x9, x9, x4 sub x10, x10, x5 sub x11, x11, x6 sub x12, x12, x7 stp x8, x9, [sp, #408] stp x10, x11, [sp, #424] str x12, [sp, #440] // add add x3, x3, x13 add x4, x4, x14 add x5, x5, x15 add x6, x6, x16 add x7, x7, x17 add x29, sp, #488 stp x3, x4, [x29, #0] stp x5, x6, [x29, #16] str x7, [x29, #32] // add ldp x1, x3, [x29, #240] ldp x5, x7, [x29, #256] ldr x9, [x29, #272] ldr x0, [sp, #152] ldp x10, x12, [x0, #40] ldp x14, x16, [x0, #56] ldr x18, [x0, #72] add x11, x10, x1 add x13, x12, x3 add x15, x14, x5 add x17, x16, x7 add x19, x18, x9 stp x11, x13, [x29, #160] stp x15, x17, [x29, #176] str x19, [x29, #192] // sub ldp x21, x22, [sp, #192] add x10, x10, x21 add x12, x12, x22 add x14, x14, x22 add x16, x16, x22 add x18, x18, x22 sub x10, x10, x1 sub x12, x12, x3 sub x14, x14, x5 sub x16, x16, x7 sub x18, x18, x9 // mul add x29, sp, #408 ldp w20, w21, [x29, #0] ldp w22, w23, [x29, #8] ldp w24, w25, [x29, #16] ldp w26, w27, [x29, #24] ldp w28, w29, [x29, #32] lsr x11, x10, #32 lsr x13, x12, #32 lsr x15, x14, #32 lsr x17, x16, #32 lsr x19, x18, #32 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 stp x0, x2, [sp, #408] stp x4, x6, [sp, #424] str x8, [sp, #440] // mul add x29, sp, #488 ldp w10, w11, [x29, #0] ldp w12, w13, [x29, #8] ldp w14, w15, [x29, #16] ldp w16, w17, [x29, #24] ldp w18, w19, [x29, #32] ldp w20, w21, [x29, #160] ldp w22, w23, [x29, #168] ldp w24, w25, [x29, #176] ldp w26, w27, [x29, #184] ldp w28, w29, [x29, #192] 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 // add ldp x13, x14, [sp, #408] ldp x15, x16, [sp, #424] ldr x17, [sp, #440] add x1, x0, x13 add x3, x2, x14 add x5, x4, x15 add x7, x6, x16 add x9, x8, x17 add x29, sp, #368 stp x1, x3, [x29, #280] stp x5, x7, [x29, #296] str x9, [x29, #312] // sub ldp x21, x22, [sp, #192] add x0, x0, x21 add x2, x2, x22 add x4, x4, x22 add x6, x6, x22 add x8, x8, x22 sub x0, x0, x13 sub x2, x2, x14 sub x4, x4, x15 sub x6, x6, x16 sub x8, x8, x17 stp x0, x2, [x29, #120] stp x4, x6, [x29, #136] str x8, [x29, #152] // mul ldp w10, w11, [x29, #0] ldp w12, w13, [x29, #8] ldp w14, w15, [x29, #16] ldp w16, w17, [x29, #24] ldp w18, w19, [x29, #32] add x29, sp, #768 ldp w20, w21, [x29, #0] ldp w22, w23, [x29, #8] ldp w24, w25, [x29, #16] ldp w26, w27, [x29, #24] ldp w28, w29, [x29, #32] 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 x17, x7, #0x1ffffff add x3, x3, x2, lsr #26 and x12, x2, #0x3ffffff add x9, x9, x8, lsr #26 and x18, x8, #0x3ffffff add x4, x4, x3, lsr #25 and x13, x3, #0x1ffffff bic x10, x9, #0x1ffffff add x0, x0, x10, lsr #25 add x0, x0, x10, lsr #24 add x0, x0, x10, lsr #21 and x19, x9, #0x1ffffff add x5, x5, x4, lsr #26 and x14, x4, #0x3ffffff add x11, x1, x0, lsr #26 and x10, x0, #0x3ffffff add x16, x6, x5, lsr #25 and x15, x5, #0x1ffffff // mul ldp w20, w21, [sp, #208] ldp w22, w23, [sp, #216] ldp w24, w25, [sp, #224] ldp w26, w27, [sp, #232] ldp w28, w29, [sp, #240] 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 stp x0, x2, [sp, #408] stp x4, x6, [sp, #424] str x8, [sp, #440] // mul ldr x0, [sp, #152] ldp w10, w11, [x0, #80] ldp w12, w13, [x0, #88] ldp w14, w15, [x0, #96] ldp w16, w17, [x0, #104] ldp w18, w19, [x0, #112] add x29, sp, #328 ldp w20, w21, [x29, #0] ldp w22, w23, [x29, #8] ldp w24, w25, [x29, #16] ldp w26, w27, [x29, #24] ldp w28, w29, [x29, #32] 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 // double and then reduce add x0, x0, x0 add x1, x1, x1 add x2, x2, x2 add x3, x3, x3 add x4, x4, x4 add x5, x5, x5 add x6, x6, x6 add x7, x7, x7 add x8, x8, x8 add x9, x9, x9 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 // add ldp x13, x14, [sp, #408] ldp x15, x16, [sp, #424] ldr x17, [sp, #440] add x1, x0, x13 add x3, x2, x14 add x5, x4, x15 add x7, x6, x16 add x9, x8, x17 add x29, sp, #488 stp x1, x3, [x29, #80] stp x5, x7, [x29, #96] str x9, [x29, #112] // sub ldp x21, x22, [sp, #192] add x0, x0, x21 add x2, x2, x22 add x4, x4, x22 add x6, x6, x22 add x8, x8, x22 sub x20, x0, x13 sub x22, x2, x14 sub x24, x4, x15 sub x26, x6, x16 sub x28, x8, x17 stp x20, x22, [x29, #200] stp x24, x26, [x29, #216] str x28, [x29, #232] /* p1p1 to p3 */ // mul ldp w10, w11, [x29, #0] ldp w12, w13, [x29, #8] ldp w14, w15, [x29, #16] ldp w16, w17, [x29, #24] ldp w18, w19, [x29, #32] lsr x21, x20, #32 lsr x23, x22, #32 lsr x25, x24, #32 lsr x27, x26, #32 lsr x29, x28, #32 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 stp x0, x2, [sp, #248] stp x4, x6, [sp, #264] str x8, [sp, #280] // mul add x29, sp, #568 ldp w10, w11, [x29, #0] ldp w12, w13, [x29, #8] ldp w14, w15, [x29, #16] ldp w16, w17, [x29, #24] ldp w18, w19, [x29, #32] ldp w20, w21, [x29, #80] ldp w22, w23, [x29, #88] ldp w24, w25, [x29, #96] ldp w26, w27, [x29, #104] ldp w28, w29, [x29, #112] 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 stp x0, x2, [sp, #288] stp x4, x6, [sp, #304] str x8, [sp, #320] // mul add x29, sp, #568 ldp w10, w11, [x29, #0] ldp w12, w13, [x29, #8] ldp w14, w15, [x29, #16] ldp w16, w17, [x29, #24] ldp w18, w19, [x29, #32] ldp w20, w21, [x29, #120] ldp w22, w23, [x29, #128] ldp w24, w25, [x29, #136] ldp w26, w27, [x29, #144] ldp w28, w29, [x29, #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 stp x0, x2, [sp, #328] stp x4, x6, [sp, #344] str x8, [sp, #360] // mul add x29, sp, #488 ldp w10, w11, [x29, #0] ldp w12, w13, [x29, #8] ldp w14, w15, [x29, #16] ldp w16, w17, [x29, #24] ldp w18, w19, [x29, #32] ldp w20, w21, [x29, #160] ldp w22, w23, [x29, #168] ldp w24, w25, [x29, #176] ldp w26, w27, [x29, #184] ldp w28, w29, [x29, #192] 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 stp x0, x2, [sp, #368] stp x4, x6, [sp, #384] str x8, [sp, #400] .L6: ldr x1, [sp, #136] add x1, x1, #256 str x1, [sp, #136] ldr x3, [sp, #128] add x3, x3, #1 str x3, [sp, #128] ldr x5, [sp, #120] cmp x3, x5 blt .L3 .L7: ldp x28, x29, [sp, #176] sub x28, x28, #1 sub x29, x29, #1 stp x28, x29, [sp, #176] cmp x29, xzr bge .L1 .L8: /* add p1p1 */ // sub ldr x0, [sp, #96] ldp x3, x4, [x0, #0] ldp x5, x6, [x0, #16] ldr x7, [x0, #32] ldp x13, x14, [x0, #40] ldp x15, x16, [x0, #56] ldr x17, [x0, #72] ldp x21, x22, [sp, #192] add x8, x13, x21 add x9, x14, x22 add x10, x15, x22 add x11, x16, x22 add x12, x17, x22 sub x8, x8, x3 sub x9, x9, x4 sub x10, x10, x5 sub x11, x11, x6 sub x12, x12, x7 stp x8, x9, [sp, #408] stp x10, x11, [sp, #424] str x12, [sp, #440] // add add x3, x3, x13 add x4, x4, x14 add x5, x5, x15 add x6, x6, x16 add x7, x7, x17 add x29, sp, #488 stp x3, x4, [x29, #0] stp x5, x6, [x29, #16] str x7, [x29, #32] // add ldp x1, x3, [sp, #248] ldp x5, x7, [sp, #264] ldr x9, [sp, #280] ldp x10, x12, [sp, #288] ldp x14, x16, [sp, #304] ldr x18, [sp, #320] add x11, x10, x1 add x13, x12, x3 add x15, x14, x5 add x17, x16, x7 add x19, x18, x9 stp x11, x13, [x29, #160] stp x15, x17, [x29, #176] str x19, [x29, #192] // sub ldp x21, x22, [sp, #192] add x10, x10, x21 add x12, x12, x22 add x14, x14, x22 add x16, x16, x22 add x18, x18, x22 sub x10, x10, x1 sub x12, x12, x3 sub x14, x14, x5 sub x16, x16, x7 sub x18, x18, x9 // mul add x29, sp, #408 ldp w20, w21, [x29, #0] ldp w22, w23, [x29, #8] ldp w24, w25, [x29, #16] ldp w26, w27, [x29, #24] ldp w28, w29, [x29, #32] lsr x11, x10, #32 lsr x13, x12, #32 lsr x15, x14, #32 lsr x17, x16, #32 lsr x19, x18, #32 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 stp x0, x2, [sp, #408] stp x4, x6, [sp, #424] str x8, [sp, #440] // mul add x29, sp, #488 ldp w10, w11, [x29, #0] ldp w12, w13, [x29, #8] ldp w14, w15, [x29, #16] ldp w16, w17, [x29, #24] ldp w18, w19, [x29, #32] ldp w20, w21, [x29, #160] ldp w22, w23, [x29, #168] ldp w24, w25, [x29, #176] ldp w26, w27, [x29, #184] ldp w28, w29, [x29, #192] 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 // add ldp x13, x14, [sp, #408] ldp x15, x16, [sp, #424] ldr x17, [sp, #440] add x1, x0, x13 add x3, x2, x14 add x5, x4, x15 add x7, x6, x16 add x9, x8, x17 add x29, sp, #368 stp x1, x3, [x29, #280] stp x5, x7, [x29, #296] str x9, [x29, #312] // sub ldp x21, x22, [sp, #192] add x0, x0, x21 add x2, x2, x22 add x4, x4, x22 add x6, x6, x22 add x8, x8, x22 sub x0, x0, x13 sub x2, x2, x14 sub x4, x4, x15 sub x6, x6, x16 sub x8, x8, x17 stp x0, x2, [x29, #120] stp x4, x6, [x29, #136] str x8, [x29, #152] // mul ldr x0, [sp, #96] ldp w10, w11, [x0, #120] ldp w12, w13, [x0, #128] ldp w14, w15, [x0, #136] ldp w16, w17, [x0, #144] ldp w18, w19, [x0, #152] ldp w20, w21, [x29, #0] ldp w22, w23, [x29, #8] ldp w24, w25, [x29, #16] ldp w26, w27, [x29, #24] ldp w28, w29, [x29, #32] 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 x17, x7, #0x1ffffff add x3, x3, x2, lsr #26 and x12, x2, #0x3ffffff add x9, x9, x8, lsr #26 and x18, x8, #0x3ffffff add x4, x4, x3, lsr #25 and x13, x3, #0x1ffffff bic x10, x9, #0x1ffffff add x0, x0, x10, lsr #25 add x0, x0, x10, lsr #24 add x0, x0, x10, lsr #21 and x19, x9, #0x1ffffff add x5, x5, x4, lsr #26 and x14, x4, #0x3ffffff add x11, x1, x0, lsr #26 and x10, x0, #0x3ffffff add x16, x6, x5, lsr #25 and x15, x5, #0x1ffffff // mul ldp w20, w21, [sp, #208] ldp w22, w23, [sp, #216] ldp w24, w25, [sp, #224] ldp w26, w27, [sp, #232] ldp w28, w29, [sp, #240] 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 stp x0, x2, [sp, #408] stp x4, x6, [sp, #424] str x8, [sp, #440] // mul ldr x0, [sp, #96] ldp w10, w11, [x0, #80] ldp w12, w13, [x0, #88] ldp w14, w15, [x0, #96] ldp w16, w17, [x0, #104] ldp w18, w19, [x0, #112] add x29, sp, #328 ldp w20, w21, [x29, #0] ldp w22, w23, [x29, #8] ldp w24, w25, [x29, #16] ldp w26, w27, [x29, #24] ldp w28, w29, [x29, #32] 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 // double and then reduce add x0, x0, x0 add x1, x1, x1 add x2, x2, x2 add x3, x3, x3 add x4, x4, x4 add x5, x5, x5 add x6, x6, x6 add x7, x7, x7 add x8, x8, x8 add x9, x9, x9 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 // add ldp x13, x14, [sp, #408] ldp x15, x16, [sp, #424] ldr x17, [sp, #440] add x1, x0, x13 add x3, x2, x14 add x5, x4, x15 add x7, x6, x16 add x9, x8, x17 add x29, sp, #488 stp x1, x3, [x29, #80] stp x5, x7, [x29, #96] str x9, [x29, #112] // sub ldp x21, x22, [sp, #192] add x0, x0, x21 add x2, x2, x22 add x4, x4, x22 add x6, x6, x22 add x8, x8, x22 sub x20, x0, x13 sub x22, x2, x14 sub x24, x4, x15 sub x26, x6, x16 sub x28, x8, x17 stp x20, x22, [x29, #200] stp x24, x26, [x29, #216] str x28, [x29, #232] /* p1p1 to p3 */ // mul ldp w10, w11, [x29, #0] ldp w12, w13, [x29, #8] ldp w14, w15, [x29, #16] ldp w16, w17, [x29, #24] ldp w18, w19, [x29, #32] lsr x21, x20, #32 lsr x23, x22, #32 lsr x25, x24, #32 lsr x27, x26, #32 lsr x29, x28, #32 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, #0] stp x4, x6, [x10, #16] str x8, [x10, #32] // mul add x29, sp, #568 ldp w10, w11, [x29, #0] ldp w12, w13, [x29, #8] ldp w14, w15, [x29, #16] ldp w16, w17, [x29, #24] ldp w18, w19, [x29, #32] ldp w20, w21, [x29, #80] ldp w22, w23, [x29, #88] ldp w24, w25, [x29, #96] ldp w26, w27, [x29, #104] ldp w28, w29, [x29, #112] 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, #40] stp x4, x6, [x10, #56] str x8, [x10, #72] // mul add x29, sp, #568 ldp w10, w11, [x29, #0] ldp w12, w13, [x29, #8] ldp w14, w15, [x29, #16] ldp w16, w17, [x29, #24] ldp w18, w19, [x29, #32] ldp w20, w21, [x29, #120] ldp w22, w23, [x29, #128] ldp w24, w25, [x29, #136] ldp w26, w27, [x29, #144] ldp w28, w29, [x29, #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] // mul add x29, sp, #488 ldp w10, w11, [x29, #0] ldp w12, w13, [x29, #8] ldp w14, w15, [x29, #16] ldp w16, w17, [x29, #24] ldp w18, w19, [x29, #32] ldp w20, w21, [x29, #160] ldp w22, w23, [x29, #168] ldp w24, w25, [x29, #176] ldp w26, w27, [x29, #184] ldp w28, w29, [x29, #192] 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, #120] stp x4, x6, [x10, #136] str x8, [x10, #152] 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, #816 ret .section .note.GNU-stack,"",@progbits