forked from NUDT-compiler/nudt-compiler-cpp
You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
877 lines
18 KiB
877 lines
18 KiB
.data
|
|
.globl mod
|
|
.p2align 2
|
|
mod:
|
|
.word 998244353
|
|
.globl d
|
|
.p2align 2
|
|
d:
|
|
.word 0
|
|
.globl maxlen
|
|
.p2align 2
|
|
maxlen:
|
|
.word 2097152
|
|
.globl temp
|
|
.p2align 2
|
|
temp:
|
|
.zero 8388608
|
|
.globl a
|
|
.p2align 2
|
|
a:
|
|
.zero 8388608
|
|
.globl b
|
|
.p2align 2
|
|
b:
|
|
.zero 8388608
|
|
.globl c
|
|
.p2align 2
|
|
c:
|
|
.zero 8388608
|
|
|
|
.text
|
|
.globl multiply
|
|
.p2align 2
|
|
multiply:
|
|
.L.multiply.0:
|
|
stp x29, x30, [sp, #-16]!
|
|
mov x29, sp
|
|
sub sp, sp, #80
|
|
stur w0, [x29, #-4]
|
|
stur w1, [x29, #-8]
|
|
ldur w8, [x29, #-4]
|
|
stur w8, [x29, #-16]
|
|
ldur w8, [x29, #-8]
|
|
stur w8, [x29, #-20]
|
|
ldur w8, [x29, #-20]
|
|
stur w8, [x29, #-24]
|
|
ldur w8, [x29, #-24]
|
|
mov w9, #0
|
|
cmp w8, w9
|
|
b.eq .L.multiply.1
|
|
b .L.multiply.2
|
|
.L.multiply.1:
|
|
mov w0, #0
|
|
add sp, sp, #80
|
|
ldp x29, x30, [sp], #16
|
|
ret
|
|
.L.multiply.2:
|
|
ldur w8, [x29, #-20]
|
|
stur w8, [x29, #-28]
|
|
ldur w8, [x29, #-28]
|
|
mov w9, #1
|
|
cmp w8, w9
|
|
b.eq .L.multiply.3
|
|
b .L.multiply.4
|
|
.L.multiply.3:
|
|
ldur w8, [x29, #-16]
|
|
stur w8, [x29, #-32]
|
|
ldur w0, [x29, #-32]
|
|
movz w8, #1
|
|
movk w8, #15232, lsl #16
|
|
sdiv w10, w0, w8
|
|
msub w0, w10, w8, w0
|
|
add sp, sp, #80
|
|
ldp x29, x30, [sp], #16
|
|
ret
|
|
.L.multiply.4:
|
|
ldur w8, [x29, #-16]
|
|
stur w8, [x29, #-36]
|
|
ldur w8, [x29, #-20]
|
|
stur w8, [x29, #-40]
|
|
ldur w0, [x29, #-36]
|
|
ldur w1, [x29, #-40]
|
|
mov w8, #2
|
|
sdiv w1, w1, w8
|
|
bl multiply
|
|
stur w0, [x29, #-44]
|
|
ldur w8, [x29, #-44]
|
|
stur w8, [x29, #-12]
|
|
ldur w8, [x29, #-12]
|
|
stur w8, [x29, #-48]
|
|
ldur w8, [x29, #-12]
|
|
stur w8, [x29, #-52]
|
|
ldur w8, [x29, #-48]
|
|
ldur w10, [x29, #-52]
|
|
add w8, w8, w10
|
|
movz w9, #1
|
|
movk w9, #15232, lsl #16
|
|
sdiv w10, w8, w9
|
|
msub w8, w10, w9, w8
|
|
stur w8, [x29, #-12]
|
|
ldur w8, [x29, #-20]
|
|
stur w8, [x29, #-56]
|
|
ldur w8, [x29, #-56]
|
|
mov w10, #2
|
|
sdiv w10, w8, w10
|
|
msub w8, w10, w10, w8
|
|
mov w9, #1
|
|
cmp w8, w9
|
|
b.eq .L.multiply.5
|
|
b .L.multiply.6
|
|
.L.multiply.5:
|
|
ldur w8, [x29, #-12]
|
|
stur w8, [x29, #-60]
|
|
ldur w8, [x29, #-16]
|
|
stur w8, [x29, #-64]
|
|
ldur w0, [x29, #-60]
|
|
ldur w9, [x29, #-64]
|
|
add w0, w0, w9
|
|
movz w8, #1
|
|
movk w8, #15232, lsl #16
|
|
sdiv w10, w0, w8
|
|
msub w0, w10, w8, w0
|
|
add sp, sp, #80
|
|
ldp x29, x30, [sp], #16
|
|
ret
|
|
.L.multiply.6:
|
|
ldur w8, [x29, #-12]
|
|
stur w8, [x29, #-68]
|
|
ldur w0, [x29, #-68]
|
|
add sp, sp, #80
|
|
ldp x29, x30, [sp], #16
|
|
ret
|
|
.L.multiply.7:
|
|
mov w0, #0
|
|
add sp, sp, #80
|
|
ldp x29, x30, [sp], #16
|
|
ret
|
|
|
|
.text
|
|
.globl power
|
|
.p2align 2
|
|
power:
|
|
.L.power.0:
|
|
stp x29, x30, [sp, #-16]!
|
|
mov x29, sp
|
|
sub sp, sp, #80
|
|
stur w0, [x29, #-4]
|
|
stur w1, [x29, #-8]
|
|
ldur w8, [x29, #-4]
|
|
stur w8, [x29, #-16]
|
|
ldur w8, [x29, #-8]
|
|
stur w8, [x29, #-20]
|
|
ldur w8, [x29, #-20]
|
|
stur w8, [x29, #-24]
|
|
ldur w8, [x29, #-24]
|
|
mov w9, #0
|
|
cmp w8, w9
|
|
b.eq .L.power.1
|
|
b .L.power.2
|
|
.L.power.1:
|
|
mov w0, #1
|
|
add sp, sp, #80
|
|
ldp x29, x30, [sp], #16
|
|
ret
|
|
.L.power.2:
|
|
ldur w8, [x29, #-16]
|
|
stur w8, [x29, #-28]
|
|
ldur w8, [x29, #-20]
|
|
stur w8, [x29, #-32]
|
|
ldur w0, [x29, #-28]
|
|
ldur w1, [x29, #-32]
|
|
mov w8, #2
|
|
sdiv w1, w1, w8
|
|
bl power
|
|
stur w0, [x29, #-36]
|
|
ldur w8, [x29, #-36]
|
|
stur w8, [x29, #-12]
|
|
ldur w8, [x29, #-12]
|
|
stur w8, [x29, #-40]
|
|
ldur w8, [x29, #-12]
|
|
stur w8, [x29, #-44]
|
|
ldur w0, [x29, #-40]
|
|
ldur w1, [x29, #-44]
|
|
bl multiply
|
|
stur w0, [x29, #-48]
|
|
ldur w8, [x29, #-48]
|
|
stur w8, [x29, #-12]
|
|
ldur w8, [x29, #-20]
|
|
stur w8, [x29, #-52]
|
|
ldur w8, [x29, #-52]
|
|
mov w10, #2
|
|
sdiv w10, w8, w10
|
|
msub w8, w10, w10, w8
|
|
mov w9, #1
|
|
cmp w8, w9
|
|
b.eq .L.power.3
|
|
b .L.power.4
|
|
.L.power.3:
|
|
ldur w8, [x29, #-12]
|
|
stur w8, [x29, #-56]
|
|
ldur w8, [x29, #-16]
|
|
stur w8, [x29, #-60]
|
|
ldur w0, [x29, #-56]
|
|
ldur w1, [x29, #-60]
|
|
bl multiply
|
|
stur w0, [x29, #-64]
|
|
ldur w0, [x29, #-64]
|
|
add sp, sp, #80
|
|
ldp x29, x30, [sp], #16
|
|
ret
|
|
.L.power.4:
|
|
ldur w8, [x29, #-12]
|
|
stur w8, [x29, #-68]
|
|
ldur w0, [x29, #-68]
|
|
add sp, sp, #80
|
|
ldp x29, x30, [sp], #16
|
|
ret
|
|
.L.power.5:
|
|
mov w0, #0
|
|
add sp, sp, #80
|
|
ldp x29, x30, [sp], #16
|
|
ret
|
|
|
|
.text
|
|
.globl memmove
|
|
.p2align 2
|
|
memmove:
|
|
.L.memmove.0:
|
|
stp x29, x30, [sp, #-16]!
|
|
mov x29, sp
|
|
sub sp, sp, #80
|
|
stur x0, [x29, #-8]
|
|
stur w1, [x29, #-12]
|
|
stur x2, [x29, #-24]
|
|
stur w3, [x29, #-28]
|
|
ldur w8, [x29, #-12]
|
|
stur w8, [x29, #-36]
|
|
ldur w8, [x29, #-28]
|
|
stur w8, [x29, #-40]
|
|
mov w8, #0
|
|
stur w8, [x29, #-32]
|
|
b .L.memmove.1
|
|
.L.memmove.1:
|
|
ldur w8, [x29, #-32]
|
|
stur w8, [x29, #-44]
|
|
ldur w8, [x29, #-40]
|
|
stur w8, [x29, #-48]
|
|
ldur w8, [x29, #-44]
|
|
ldur w9, [x29, #-48]
|
|
cmp w8, w9
|
|
b.lt .L.memmove.2
|
|
b .L.memmove.3
|
|
.L.memmove.2:
|
|
ldur w8, [x29, #-36]
|
|
stur w8, [x29, #-52]
|
|
ldur w8, [x29, #-32]
|
|
stur w8, [x29, #-56]
|
|
ldur w8, [x29, #-32]
|
|
stur w8, [x29, #-60]
|
|
ldur x15, [x29, #-24]
|
|
ldur w14, [x29, #-60]
|
|
sxtw x14, w14
|
|
lsl x14, x14, #2
|
|
add x15, x15, x14
|
|
ldr w8, [x15]
|
|
stur w8, [x29, #-64]
|
|
ldur w8, [x29, #-64]
|
|
ldur x15, [x29, #-8]
|
|
ldur w14, [x29, #-52]
|
|
ldur w9, [x29, #-56]
|
|
add w14, w14, w9
|
|
sxtw x14, w14
|
|
lsl x14, x14, #2
|
|
add x15, x15, x14
|
|
str w8, [x15]
|
|
ldur w8, [x29, #-32]
|
|
stur w8, [x29, #-68]
|
|
ldur w8, [x29, #-68]
|
|
mov w9, #1
|
|
add w8, w8, w9
|
|
stur w8, [x29, #-32]
|
|
b .L.memmove.1
|
|
.L.memmove.3:
|
|
ldur w8, [x29, #-32]
|
|
stur w8, [x29, #-72]
|
|
ldur w0, [x29, #-72]
|
|
add sp, sp, #80
|
|
ldp x29, x30, [sp], #16
|
|
ret
|
|
|
|
.text
|
|
.globl fft
|
|
.p2align 2
|
|
fft:
|
|
.L.fft.0:
|
|
stp x29, x30, [sp, #-16]!
|
|
mov x29, sp
|
|
sub sp, sp, #272
|
|
stur x0, [x29, #-8]
|
|
stur w1, [x29, #-12]
|
|
stur w2, [x29, #-16]
|
|
stur w3, [x29, #-20]
|
|
ldur w8, [x29, #-12]
|
|
stur w8, [x29, #-40]
|
|
ldur w8, [x29, #-16]
|
|
stur w8, [x29, #-44]
|
|
ldur w8, [x29, #-20]
|
|
stur w8, [x29, #-48]
|
|
ldur w8, [x29, #-44]
|
|
stur w8, [x29, #-52]
|
|
ldur w8, [x29, #-52]
|
|
mov w9, #1
|
|
cmp w8, w9
|
|
b.eq .L.fft.1
|
|
b .L.fft.2
|
|
.L.fft.1:
|
|
mov w0, #1
|
|
add sp, sp, #272
|
|
ldp x29, x30, [sp], #16
|
|
ret
|
|
.L.fft.2:
|
|
mov w8, #0
|
|
stur w8, [x29, #-24]
|
|
b .L.fft.3
|
|
.L.fft.3:
|
|
ldur w8, [x29, #-24]
|
|
stur w8, [x29, #-56]
|
|
ldur w8, [x29, #-44]
|
|
stur w8, [x29, #-60]
|
|
ldur w8, [x29, #-56]
|
|
ldur w9, [x29, #-60]
|
|
cmp w8, w9
|
|
b.lt .L.fft.4
|
|
b .L.fft.5
|
|
.L.fft.4:
|
|
ldur w8, [x29, #-24]
|
|
stur w8, [x29, #-64]
|
|
ldur w8, [x29, #-64]
|
|
mov w9, #2
|
|
sdiv w10, w8, w9
|
|
msub w8, w10, w9, w8
|
|
mov w9, #0
|
|
cmp w8, w9
|
|
b.eq .L.fft.6
|
|
b .L.fft.7
|
|
.L.fft.5:
|
|
ldur w8, [x29, #-40]
|
|
stur w8, [x29, #-68]
|
|
ldur w8, [x29, #-44]
|
|
stur w8, [x29, #-72]
|
|
ldur x0, [x29, #-8]
|
|
ldur w1, [x29, #-68]
|
|
adrp x2, temp
|
|
add x2, x2, :lo12:temp
|
|
ldur w3, [x29, #-72]
|
|
bl memmove
|
|
stur w0, [x29, #-76]
|
|
ldur w8, [x29, #-40]
|
|
stur w8, [x29, #-80]
|
|
ldur w8, [x29, #-44]
|
|
stur w8, [x29, #-84]
|
|
ldur w8, [x29, #-48]
|
|
stur w8, [x29, #-88]
|
|
ldur w8, [x29, #-48]
|
|
stur w8, [x29, #-92]
|
|
ldur w0, [x29, #-88]
|
|
ldur w1, [x29, #-92]
|
|
bl multiply
|
|
stur w0, [x29, #-96]
|
|
ldur x0, [x29, #-8]
|
|
ldur w1, [x29, #-80]
|
|
ldur w2, [x29, #-84]
|
|
mov w8, #2
|
|
sdiv w2, w2, w8
|
|
ldur w3, [x29, #-96]
|
|
bl fft
|
|
stur w0, [x29, #-100]
|
|
ldur w8, [x29, #-40]
|
|
stur w8, [x29, #-104]
|
|
ldur w8, [x29, #-44]
|
|
stur w8, [x29, #-108]
|
|
ldur w8, [x29, #-44]
|
|
stur w8, [x29, #-112]
|
|
ldur w8, [x29, #-48]
|
|
stur w8, [x29, #-116]
|
|
ldur w8, [x29, #-48]
|
|
stur w8, [x29, #-120]
|
|
ldur w0, [x29, #-116]
|
|
ldur w1, [x29, #-120]
|
|
bl multiply
|
|
stur w0, [x29, #-124]
|
|
ldur x0, [x29, #-8]
|
|
ldur w1, [x29, #-104]
|
|
ldur w8, [x29, #-108]
|
|
mov w9, #2
|
|
sdiv w8, w8, w9
|
|
add w1, w1, w8
|
|
ldur w2, [x29, #-112]
|
|
mov w8, #2
|
|
sdiv w2, w2, w8
|
|
ldur w3, [x29, #-124]
|
|
bl fft
|
|
stur w0, [x29, #-128]
|
|
mov w8, #0
|
|
stur w8, [x29, #-24]
|
|
mov w8, #1
|
|
stur w8, [x29, #-28]
|
|
b .L.fft.9
|
|
.L.fft.6:
|
|
ldur w8, [x29, #-24]
|
|
stur w8, [x29, #-132]
|
|
ldur w8, [x29, #-24]
|
|
stur w8, [x29, #-136]
|
|
ldur w8, [x29, #-40]
|
|
stur w8, [x29, #-140]
|
|
ldur x15, [x29, #-8]
|
|
ldur w14, [x29, #-136]
|
|
ldur w8, [x29, #-140]
|
|
add w14, w14, w8
|
|
sxtw x14, w14
|
|
lsl x14, x14, #2
|
|
add x15, x15, x14
|
|
ldr w8, [x15]
|
|
stur w8, [x29, #-144]
|
|
ldur w8, [x29, #-144]
|
|
adrp x15, temp
|
|
add x15, x15, :lo12:temp
|
|
ldur w14, [x29, #-132]
|
|
mov w9, #2
|
|
sdiv w14, w14, w9
|
|
sxtw x14, w14
|
|
lsl x14, x14, #2
|
|
add x15, x15, x14
|
|
str w8, [x15]
|
|
b .L.fft.8
|
|
.L.fft.7:
|
|
ldur w8, [x29, #-44]
|
|
stur w8, [x29, #-148]
|
|
ldur w8, [x29, #-24]
|
|
stur w8, [x29, #-152]
|
|
ldur w8, [x29, #-24]
|
|
stur w8, [x29, #-156]
|
|
ldur w8, [x29, #-40]
|
|
stur w8, [x29, #-160]
|
|
ldur x15, [x29, #-8]
|
|
ldur w14, [x29, #-156]
|
|
ldur w8, [x29, #-160]
|
|
add w14, w14, w8
|
|
sxtw x14, w14
|
|
lsl x14, x14, #2
|
|
add x15, x15, x14
|
|
ldr w8, [x15]
|
|
stur w8, [x29, #-164]
|
|
ldur w8, [x29, #-164]
|
|
adrp x15, temp
|
|
add x15, x15, :lo12:temp
|
|
ldur w14, [x29, #-148]
|
|
mov w10, #2
|
|
sdiv w14, w14, w10
|
|
ldur w9, [x29, #-152]
|
|
mov w10, #2
|
|
sdiv w9, w9, w10
|
|
add w14, w14, w9
|
|
sxtw x14, w14
|
|
lsl x14, x14, #2
|
|
add x15, x15, x14
|
|
str w8, [x15]
|
|
b .L.fft.8
|
|
.L.fft.8:
|
|
ldur w8, [x29, #-24]
|
|
stur w8, [x29, #-168]
|
|
ldur w8, [x29, #-168]
|
|
mov w9, #1
|
|
add w8, w8, w9
|
|
stur w8, [x29, #-24]
|
|
b .L.fft.3
|
|
.L.fft.9:
|
|
ldur w8, [x29, #-24]
|
|
stur w8, [x29, #-172]
|
|
ldur w8, [x29, #-44]
|
|
stur w8, [x29, #-176]
|
|
ldur w8, [x29, #-172]
|
|
ldur w9, [x29, #-176]
|
|
mov w10, #2
|
|
sdiv w9, w9, w10
|
|
cmp w8, w9
|
|
b.lt .L.fft.10
|
|
b .L.fft.11
|
|
.L.fft.10:
|
|
ldur w8, [x29, #-40]
|
|
stur w8, [x29, #-180]
|
|
ldur w8, [x29, #-24]
|
|
stur w8, [x29, #-184]
|
|
ldur x15, [x29, #-8]
|
|
ldur w14, [x29, #-180]
|
|
ldur w8, [x29, #-184]
|
|
add w14, w14, w8
|
|
sxtw x14, w14
|
|
lsl x14, x14, #2
|
|
add x15, x15, x14
|
|
ldr w8, [x15]
|
|
stur w8, [x29, #-188]
|
|
ldur w8, [x29, #-188]
|
|
stur w8, [x29, #-32]
|
|
ldur w8, [x29, #-40]
|
|
stur w8, [x29, #-192]
|
|
ldur w8, [x29, #-24]
|
|
stur w8, [x29, #-196]
|
|
ldur w8, [x29, #-44]
|
|
stur w8, [x29, #-200]
|
|
ldur x15, [x29, #-8]
|
|
ldur w14, [x29, #-192]
|
|
ldur w9, [x29, #-196]
|
|
add w14, w14, w9
|
|
ldur w8, [x29, #-200]
|
|
mov w9, #2
|
|
sdiv w8, w8, w9
|
|
add w14, w14, w8
|
|
sxtw x14, w14
|
|
lsl x14, x14, #2
|
|
add x15, x15, x14
|
|
ldr w8, [x15]
|
|
stur w8, [x29, #-204]
|
|
ldur w8, [x29, #-204]
|
|
stur w8, [x29, #-36]
|
|
ldur w8, [x29, #-40]
|
|
stur w8, [x29, #-208]
|
|
ldur w8, [x29, #-24]
|
|
stur w8, [x29, #-212]
|
|
ldur w8, [x29, #-32]
|
|
stur w8, [x29, #-216]
|
|
ldur w8, [x29, #-28]
|
|
stur w8, [x29, #-220]
|
|
ldur w8, [x29, #-36]
|
|
stur w8, [x29, #-224]
|
|
ldur w0, [x29, #-220]
|
|
ldur w1, [x29, #-224]
|
|
bl multiply
|
|
stur w0, [x29, #-228]
|
|
ldur w8, [x29, #-216]
|
|
ldur w10, [x29, #-228]
|
|
add w8, w8, w10
|
|
movz w9, #1
|
|
movk w9, #15232, lsl #16
|
|
sdiv w10, w8, w9
|
|
msub w8, w10, w9, w8
|
|
ldur x15, [x29, #-8]
|
|
ldur w14, [x29, #-208]
|
|
ldur w9, [x29, #-212]
|
|
add w14, w14, w9
|
|
sxtw x14, w14
|
|
lsl x14, x14, #2
|
|
add x15, x15, x14
|
|
str w8, [x15]
|
|
ldur w8, [x29, #-40]
|
|
stur w8, [x29, #-232]
|
|
ldur w8, [x29, #-24]
|
|
stur w8, [x29, #-236]
|
|
ldur w8, [x29, #-44]
|
|
stur w8, [x29, #-240]
|
|
ldur w8, [x29, #-32]
|
|
stur w8, [x29, #-244]
|
|
ldur w8, [x29, #-28]
|
|
stur w8, [x29, #-248]
|
|
ldur w8, [x29, #-36]
|
|
stur w8, [x29, #-252]
|
|
ldur w0, [x29, #-248]
|
|
ldur w1, [x29, #-252]
|
|
bl multiply
|
|
stur w0, [x29, #-256]
|
|
ldur w8, [x29, #-244]
|
|
ldur w11, [x29, #-256]
|
|
sub w8, w8, w11
|
|
movz w10, #1
|
|
movk w10, #15232, lsl #16
|
|
add w8, w8, w10
|
|
movz w9, #1
|
|
movk w9, #15232, lsl #16
|
|
sdiv w10, w8, w9
|
|
msub w8, w10, w9, w8
|
|
ldur x15, [x29, #-8]
|
|
ldur w14, [x29, #-232]
|
|
ldur w10, [x29, #-236]
|
|
add w14, w14, w10
|
|
ldur w9, [x29, #-240]
|
|
mov w10, #2
|
|
sdiv w9, w9, w10
|
|
add w14, w14, w9
|
|
sxtw x14, w14
|
|
lsl x14, x14, #2
|
|
add x15, x15, x14
|
|
str w8, [x15]
|
|
ldur w8, [x29, #-28]
|
|
mov x13, x29
|
|
sub x13, x13, #260
|
|
str w8, [x13]
|
|
ldur w8, [x29, #-48]
|
|
mov x13, x29
|
|
sub x13, x13, #264
|
|
str w8, [x13]
|
|
mov x13, x29
|
|
sub x13, x13, #260
|
|
ldr w0, [x13]
|
|
mov x13, x29
|
|
sub x13, x13, #264
|
|
ldr w1, [x13]
|
|
bl multiply
|
|
mov x13, x29
|
|
sub x13, x13, #268
|
|
str w0, [x13]
|
|
mov x13, x29
|
|
sub x13, x13, #268
|
|
ldr w8, [x13]
|
|
stur w8, [x29, #-28]
|
|
ldur w8, [x29, #-24]
|
|
mov x13, x29
|
|
sub x13, x13, #272
|
|
str w8, [x13]
|
|
mov x13, x29
|
|
sub x13, x13, #272
|
|
ldr w8, [x13]
|
|
mov w9, #1
|
|
add w8, w8, w9
|
|
stur w8, [x29, #-24]
|
|
b .L.fft.9
|
|
.L.fft.11:
|
|
mov w0, #0
|
|
add sp, sp, #272
|
|
ldp x29, x30, [sp], #16
|
|
ret
|
|
|
|
.text
|
|
.globl main
|
|
.p2align 2
|
|
main:
|
|
.L.main.0:
|
|
stp x29, x30, [sp, #-16]!
|
|
mov x29, sp
|
|
sub sp, sp, #176
|
|
adrp x0, a
|
|
add x0, x0, :lo12:a
|
|
bl getarray
|
|
stur w0, [x29, #-16]
|
|
ldur w8, [x29, #-16]
|
|
stur w8, [x29, #-4]
|
|
adrp x0, b
|
|
add x0, x0, :lo12:b
|
|
bl getarray
|
|
stur w0, [x29, #-20]
|
|
ldur w8, [x29, #-20]
|
|
stur w8, [x29, #-8]
|
|
bl starttime
|
|
mov w8, #1
|
|
adrp x13, d
|
|
str w8, [x13, #:lo12:d]
|
|
b .L.main.1
|
|
.L.main.1:
|
|
adrp x13, d
|
|
ldr w8, [x13, #:lo12:d]
|
|
stur w8, [x29, #-24]
|
|
ldur w8, [x29, #-4]
|
|
stur w8, [x29, #-28]
|
|
ldur w8, [x29, #-8]
|
|
stur w8, [x29, #-32]
|
|
ldur w8, [x29, #-24]
|
|
ldur w9, [x29, #-28]
|
|
ldur w11, [x29, #-32]
|
|
add w9, w9, w11
|
|
mov w10, #1
|
|
sub w9, w9, w10
|
|
cmp w8, w9
|
|
b.lt .L.main.2
|
|
b .L.main.3
|
|
.L.main.2:
|
|
adrp x13, d
|
|
ldr w8, [x13, #:lo12:d]
|
|
stur w8, [x29, #-36]
|
|
ldur w8, [x29, #-36]
|
|
mov w9, #2
|
|
lsl w8, w8, #1
|
|
adrp x13, d
|
|
str w8, [x13, #:lo12:d]
|
|
b .L.main.1
|
|
.L.main.3:
|
|
adrp x13, d
|
|
ldr w8, [x13, #:lo12:d]
|
|
stur w8, [x29, #-40]
|
|
adrp x13, d
|
|
ldr w8, [x13, #:lo12:d]
|
|
stur w8, [x29, #-44]
|
|
mov w0, #3
|
|
movz w1, #0
|
|
movk w1, #15232, lsl #16
|
|
ldur w8, [x29, #-44]
|
|
sdiv w1, w1, w8
|
|
bl power
|
|
stur w0, [x29, #-48]
|
|
adrp x0, a
|
|
add x0, x0, :lo12:a
|
|
mov w1, #0
|
|
ldur w2, [x29, #-40]
|
|
ldur w3, [x29, #-48]
|
|
bl fft
|
|
stur w0, [x29, #-52]
|
|
adrp x13, d
|
|
ldr w8, [x13, #:lo12:d]
|
|
stur w8, [x29, #-56]
|
|
adrp x13, d
|
|
ldr w8, [x13, #:lo12:d]
|
|
stur w8, [x29, #-60]
|
|
mov w0, #3
|
|
movz w1, #0
|
|
movk w1, #15232, lsl #16
|
|
ldur w8, [x29, #-60]
|
|
sdiv w1, w1, w8
|
|
bl power
|
|
stur w0, [x29, #-64]
|
|
adrp x0, b
|
|
add x0, x0, :lo12:b
|
|
mov w1, #0
|
|
ldur w2, [x29, #-56]
|
|
ldur w3, [x29, #-64]
|
|
bl fft
|
|
stur w0, [x29, #-68]
|
|
mov w8, #0
|
|
stur w8, [x29, #-12]
|
|
b .L.main.4
|
|
.L.main.4:
|
|
ldur w8, [x29, #-12]
|
|
stur w8, [x29, #-72]
|
|
adrp x13, d
|
|
ldr w8, [x13, #:lo12:d]
|
|
stur w8, [x29, #-76]
|
|
ldur w8, [x29, #-72]
|
|
ldur w9, [x29, #-76]
|
|
cmp w8, w9
|
|
b.lt .L.main.5
|
|
b .L.main.6
|
|
.L.main.5:
|
|
ldur w8, [x29, #-12]
|
|
stur w8, [x29, #-80]
|
|
ldur w8, [x29, #-12]
|
|
stur w8, [x29, #-84]
|
|
adrp x15, a
|
|
add x15, x15, :lo12:a
|
|
ldur w14, [x29, #-84]
|
|
sxtw x14, w14
|
|
lsl x14, x14, #2
|
|
add x15, x15, x14
|
|
ldr w8, [x15]
|
|
stur w8, [x29, #-88]
|
|
ldur w8, [x29, #-12]
|
|
stur w8, [x29, #-92]
|
|
adrp x15, b
|
|
add x15, x15, :lo12:b
|
|
ldur w14, [x29, #-92]
|
|
sxtw x14, w14
|
|
lsl x14, x14, #2
|
|
add x15, x15, x14
|
|
ldr w8, [x15]
|
|
stur w8, [x29, #-96]
|
|
ldur w0, [x29, #-88]
|
|
ldur w1, [x29, #-96]
|
|
bl multiply
|
|
stur w0, [x29, #-100]
|
|
ldur w8, [x29, #-100]
|
|
adrp x15, a
|
|
add x15, x15, :lo12:a
|
|
ldur w14, [x29, #-80]
|
|
sxtw x14, w14
|
|
lsl x14, x14, #2
|
|
add x15, x15, x14
|
|
str w8, [x15]
|
|
ldur w8, [x29, #-12]
|
|
stur w8, [x29, #-104]
|
|
ldur w8, [x29, #-104]
|
|
mov w9, #1
|
|
add w8, w8, w9
|
|
stur w8, [x29, #-12]
|
|
b .L.main.4
|
|
.L.main.6:
|
|
adrp x13, d
|
|
ldr w8, [x13, #:lo12:d]
|
|
stur w8, [x29, #-108]
|
|
adrp x13, d
|
|
ldr w8, [x13, #:lo12:d]
|
|
stur w8, [x29, #-112]
|
|
mov w0, #3
|
|
movz w1, #0
|
|
movk w1, #15232, lsl #16
|
|
movz w8, #0
|
|
movk w8, #15232, lsl #16
|
|
ldur w9, [x29, #-112]
|
|
sdiv w8, w8, w9
|
|
sub w1, w1, w8
|
|
bl power
|
|
stur w0, [x29, #-116]
|
|
adrp x0, a
|
|
add x0, x0, :lo12:a
|
|
mov w1, #0
|
|
ldur w2, [x29, #-108]
|
|
ldur w3, [x29, #-116]
|
|
bl fft
|
|
stur w0, [x29, #-120]
|
|
mov w8, #0
|
|
stur w8, [x29, #-12]
|
|
b .L.main.7
|
|
.L.main.7:
|
|
ldur w8, [x29, #-12]
|
|
stur w8, [x29, #-124]
|
|
adrp x13, d
|
|
ldr w8, [x13, #:lo12:d]
|
|
stur w8, [x29, #-128]
|
|
ldur w8, [x29, #-124]
|
|
ldur w9, [x29, #-128]
|
|
cmp w8, w9
|
|
b.lt .L.main.8
|
|
b .L.main.9
|
|
.L.main.8:
|
|
ldur w8, [x29, #-12]
|
|
stur w8, [x29, #-132]
|
|
ldur w8, [x29, #-12]
|
|
stur w8, [x29, #-136]
|
|
adrp x15, a
|
|
add x15, x15, :lo12:a
|
|
ldur w14, [x29, #-136]
|
|
sxtw x14, w14
|
|
lsl x14, x14, #2
|
|
add x15, x15, x14
|
|
ldr w8, [x15]
|
|
stur w8, [x29, #-140]
|
|
adrp x13, d
|
|
ldr w8, [x13, #:lo12:d]
|
|
stur w8, [x29, #-144]
|
|
ldur w0, [x29, #-144]
|
|
movz w1, #65535
|
|
movk w1, #15231, lsl #16
|
|
bl power
|
|
stur w0, [x29, #-148]
|
|
ldur w0, [x29, #-140]
|
|
ldur w1, [x29, #-148]
|
|
bl multiply
|
|
stur w0, [x29, #-152]
|
|
ldur w8, [x29, #-152]
|
|
adrp x15, a
|
|
add x15, x15, :lo12:a
|
|
ldur w14, [x29, #-132]
|
|
sxtw x14, w14
|
|
lsl x14, x14, #2
|
|
add x15, x15, x14
|
|
str w8, [x15]
|
|
ldur w8, [x29, #-12]
|
|
stur w8, [x29, #-156]
|
|
ldur w8, [x29, #-156]
|
|
mov w9, #1
|
|
add w8, w8, w9
|
|
stur w8, [x29, #-12]
|
|
b .L.main.7
|
|
.L.main.9:
|
|
bl stoptime
|
|
ldur w8, [x29, #-4]
|
|
stur w8, [x29, #-160]
|
|
ldur w8, [x29, #-8]
|
|
stur w8, [x29, #-164]
|
|
ldur w0, [x29, #-160]
|
|
ldur w9, [x29, #-164]
|
|
add w0, w0, w9
|
|
mov w8, #1
|
|
sub w0, w0, w8
|
|
adrp x1, a
|
|
add x1, x1, :lo12:a
|
|
bl putarray
|
|
mov w0, #0
|
|
add sp, sp, #176
|
|
ldp x29, x30, [sp], #16
|
|
ret
|