662 lines
9.7 KiB
ArmAsm
662 lines
9.7 KiB
ArmAsm
.text
|
|
.attribute 4, 16
|
|
.attribute 5, "rv64i2p0_m2p0_a2p0_f2p0_d2p0_c2p0"
|
|
.file "conv1.sy"
|
|
.globl checkrange
|
|
.p2align 1
|
|
.type checkrange,@function
|
|
checkrange:
|
|
lui a4, 262144
|
|
addiw a1, a4, -1
|
|
addw a2, a0, a1
|
|
mv a3, a0
|
|
blt a0, a4, .LBB0_2
|
|
lui a3, 262144
|
|
.LBB0_2:
|
|
subw a2, a2, a3
|
|
lui a3, 786432
|
|
and a2, a2, a3
|
|
subw a4, a0, a2
|
|
bgtz a4, .LBB0_4
|
|
mv a4, a2
|
|
j .LBB0_5
|
|
.LBB0_4:
|
|
add a4, a4, a2
|
|
.LBB0_5:
|
|
add a1, a1, a4
|
|
sub a1, a1, a0
|
|
and a1, a1, a3
|
|
addw a0, a0, a1
|
|
subw a0, a0, a2
|
|
ret
|
|
.Lfunc_end0:
|
|
.size checkrange, .Lfunc_end0-checkrange
|
|
|
|
.globl reduce
|
|
.p2align 1
|
|
.type reduce,@function
|
|
reduce:
|
|
li a4, 4
|
|
bltu a4, a0, .LBB1_20
|
|
mv a3, a0
|
|
slli a0, a0, 2
|
|
lui a4, %hi(.LJTI1_0)
|
|
addi a4, a4, %lo(.LJTI1_0)
|
|
add a0, a0, a4
|
|
lw a0, 0(a0)
|
|
jr a0
|
|
.LBB1_2:
|
|
addw a3, a2, a1
|
|
add a1, a1, a2
|
|
lui a5, 262144
|
|
addiw a4, a5, -1
|
|
addw a0, a1, a4
|
|
mv a2, a3
|
|
blt a3, a5, .LBB1_4
|
|
lui a2, 262144
|
|
.LBB1_4:
|
|
subw a2, a0, a2
|
|
lui a0, 786432
|
|
and a5, a2, a0
|
|
subw a2, a3, a5
|
|
sub a1, a4, a1
|
|
add a1, a1, a5
|
|
blez a2, .LBB1_6
|
|
add a1, a1, a2
|
|
.LBB1_6:
|
|
and a0, a0, a1
|
|
addw a0, a0, a2
|
|
ret
|
|
.LBB1_7:
|
|
blt a2, a1, .LBB1_9
|
|
mv a1, a2
|
|
.LBB1_9:
|
|
mv a0, a1
|
|
ret
|
|
.LBB1_10:
|
|
li a0, 0
|
|
li a6, 1
|
|
li a4, 1
|
|
j .LBB1_12
|
|
.LBB1_11:
|
|
slliw a0, a0, 1
|
|
ori a0, a0, 1
|
|
srliw a3, a4, 29
|
|
slliw a4, a4, 1
|
|
bnez a3, .LBB1_23
|
|
.LBB1_12:
|
|
divw a5, a1, a4
|
|
srliw a3, a5, 31
|
|
add a3, a3, a5
|
|
andi a3, a3, -2
|
|
subw a3, a5, a3
|
|
beq a3, a6, .LBB1_11
|
|
divw a3, a2, a4
|
|
srliw a5, a3, 31
|
|
add a5, a5, a3
|
|
andi a5, a5, -2
|
|
subw a3, a3, a5
|
|
beq a3, a6, .LBB1_11
|
|
slliw a0, a0, 1
|
|
srliw a3, a4, 29
|
|
slliw a4, a4, 1
|
|
beqz a3, .LBB1_12
|
|
j .LBB1_23
|
|
.LBB1_15:
|
|
li a0, 0
|
|
li a6, 1
|
|
li a4, 1
|
|
j .LBB1_17
|
|
.LBB1_16:
|
|
slliw a0, a0, 1
|
|
srliw a3, a4, 29
|
|
slliw a4, a4, 1
|
|
bnez a3, .LBB1_23
|
|
.LBB1_17:
|
|
divw a5, a1, a4
|
|
srliw a3, a5, 31
|
|
add a3, a3, a5
|
|
andi a3, a3, -2
|
|
subw a3, a5, a3
|
|
bne a3, a6, .LBB1_16
|
|
divw a3, a2, a4
|
|
srliw a5, a3, 31
|
|
add a5, a5, a3
|
|
andi a5, a5, -2
|
|
subw a3, a3, a5
|
|
bne a3, a6, .LBB1_16
|
|
slliw a0, a0, 1
|
|
ori a0, a0, 1
|
|
srliw a3, a4, 29
|
|
slliw a4, a4, 1
|
|
beqz a3, .LBB1_17
|
|
j .LBB1_23
|
|
.LBB1_20:
|
|
li a0, 0
|
|
ret
|
|
.LBB1_21:
|
|
li a0, 0
|
|
.LBB1_22:
|
|
divw a4, a1, a3
|
|
srliw a5, a4, 31
|
|
add a5, a5, a4
|
|
andi a5, a5, -2
|
|
subw a6, a4, a5
|
|
divw a5, a2, a3
|
|
srliw a4, a5, 31
|
|
add a4, a4, a5
|
|
andi a4, a4, -2
|
|
subw a4, a5, a4
|
|
xor a4, a6, a4
|
|
snez a4, a4
|
|
slliw a0, a0, 1
|
|
srliw a5, a3, 29
|
|
slliw a3, a3, 1
|
|
or a0, a0, a4
|
|
beqz a5, .LBB1_22
|
|
.LBB1_23:
|
|
ret
|
|
.Lfunc_end1:
|
|
.size reduce, .Lfunc_end1-reduce
|
|
.section .rodata,"a",@progbits
|
|
.p2align 2
|
|
.LJTI1_0:
|
|
.word .LBB1_2
|
|
.word .LBB1_21
|
|
.word .LBB1_7
|
|
.word .LBB1_10
|
|
.word .LBB1_15
|
|
|
|
.text
|
|
.globl getvalue
|
|
.p2align 1
|
|
.type getvalue,@function
|
|
getvalue:
|
|
or a5, a4, a3
|
|
slti a5, a5, 0
|
|
slt a1, a3, a1
|
|
xori a1, a1, 1
|
|
or a1, a1, a5
|
|
slt a5, a4, a2
|
|
xori a5, a5, 1
|
|
or a5, a5, a1
|
|
li a1, 0
|
|
bnez a5, .LBB2_2
|
|
mulw a1, a3, a2
|
|
addw a1, a1, a4
|
|
slli a1, a1, 2
|
|
add a0, a0, a1
|
|
lw a1, 0(a0)
|
|
.LBB2_2:
|
|
mv a0, a1
|
|
ret
|
|
.Lfunc_end2:
|
|
.size getvalue, .Lfunc_end2-getvalue
|
|
|
|
.globl convn
|
|
.p2align 1
|
|
.type convn,@function
|
|
convn:
|
|
addi sp, sp, -160
|
|
sd ra, 152(sp)
|
|
sd s0, 144(sp)
|
|
sd s1, 136(sp)
|
|
sd s2, 128(sp)
|
|
sd s3, 120(sp)
|
|
sd s4, 112(sp)
|
|
sd s5, 104(sp)
|
|
sd s6, 96(sp)
|
|
sd s7, 88(sp)
|
|
sd s8, 80(sp)
|
|
sd s9, 72(sp)
|
|
sd s10, 64(sp)
|
|
sd s11, 56(sp)
|
|
sd a2, 40(sp)
|
|
srliw s0, a5, 31
|
|
li a6, 1
|
|
addw s0, s0, a5
|
|
mv a5, a4
|
|
blt a6, a4, .LBB3_2
|
|
li a5, 1
|
|
.LBB3_2:
|
|
lui t0, 262144
|
|
sraiw a2, s0, 1
|
|
sd a2, 48(sp)
|
|
mv s1, a3
|
|
blt a6, a3, .LBB3_4
|
|
li s1, 1
|
|
.LBB3_4:
|
|
li s0, 0
|
|
addiw t3, t0, -1
|
|
ld a2, 48(sp)
|
|
neg a2, a2
|
|
slli s1, s1, 32
|
|
srli s1, s1, 32
|
|
sd s1, 8(sp)
|
|
slli a5, a5, 32
|
|
srli t2, a5, 32
|
|
li t4, 4
|
|
lui a5, %hi(.LJTI3_0)
|
|
addi t6, a5, %lo(.LJTI3_0)
|
|
lui t5, 786432
|
|
li ra, 1
|
|
sd a2, 16(sp)
|
|
mv a5, a2
|
|
j .LBB3_6
|
|
.LBB3_5:
|
|
ld s0, 32(sp)
|
|
addi s0, s0, 1
|
|
ld a5, 24(sp)
|
|
addiw a5, a5, 1
|
|
ld a2, 8(sp)
|
|
beq s0, a2, .LBB3_37
|
|
.LBB3_6:
|
|
li s6, 0
|
|
sd a5, 24(sp)
|
|
sext.w s4, a5
|
|
mul s3, s0, a4
|
|
ld a2, 48(sp)
|
|
sd s0, 32(sp)
|
|
add s7, s0, a2
|
|
ld s5, 16(sp)
|
|
j .LBB3_8
|
|
.LBB3_7:
|
|
add a2, s6, s3
|
|
slli a2, a2, 2
|
|
ld a5, 40(sp)
|
|
add a2, a2, a5
|
|
sw a6, 0(a2)
|
|
addi s6, s6, 1
|
|
addiw s5, s5, 1
|
|
beq s6, t2, .LBB3_5
|
|
.LBB3_8:
|
|
li a6, 0
|
|
sext.w s2, s5
|
|
ld a2, 48(sp)
|
|
add s10, s6, a2
|
|
mv s11, s4
|
|
j .LBB3_10
|
|
.LBB3_9:
|
|
addi s11, s11, 1
|
|
bge s11, s7, .LBB3_7
|
|
.LBB3_10:
|
|
mul s9, s11, a4
|
|
mv s8, s2
|
|
j .LBB3_13
|
|
.LBB3_11:
|
|
mv a6, t1
|
|
.LBB3_12:
|
|
addi s8, s8, 1
|
|
bge s8, s10, .LBB3_9
|
|
.LBB3_13:
|
|
slt a5, s11, a3
|
|
xori a5, a5, 1
|
|
or s0, s8, s11
|
|
srliw s0, s0, 31
|
|
or a5, a5, s0
|
|
slt s0, s8, a4
|
|
xori s0, s0, 1
|
|
or s0, s0, a5
|
|
li a7, 0
|
|
mv t1, a6
|
|
bnez s0, .LBB3_15
|
|
add s1, s8, s9
|
|
slli s1, s1, 2
|
|
add s1, s1, a1
|
|
lw a7, 0(s1)
|
|
.LBB3_15:
|
|
bltu t4, a0, .LBB3_22
|
|
slli s1, a0, 2
|
|
add s1, s1, t6
|
|
lw s0, 0(s1)
|
|
jr s0
|
|
.LBB3_17:
|
|
addw s1, a7, t1
|
|
lui t0, 262144
|
|
mv a6, s1
|
|
blt s1, t0, .LBB3_19
|
|
lui a6, 262144
|
|
.LBB3_19:
|
|
add s0, a7, t1
|
|
addw a5, s0, t3
|
|
subw a5, a5, a6
|
|
and a5, a5, t5
|
|
subw s1, s1, a5
|
|
sub s0, t3, s0
|
|
add a5, a5, s0
|
|
blez s1, .LBB3_21
|
|
add a5, a5, s1
|
|
.LBB3_21:
|
|
and a5, a5, t5
|
|
addw a6, a5, s1
|
|
j .LBB3_12
|
|
.LBB3_22:
|
|
li a6, 0
|
|
j .LBB3_12
|
|
.LBB3_23:
|
|
li a6, 0
|
|
mv s0, a0
|
|
.LBB3_24:
|
|
divw a5, t1, s0
|
|
srliw a2, a5, 31
|
|
add a2, a2, a5
|
|
andi a2, a2, -2
|
|
subw a2, a5, a2
|
|
divw a5, a7, s0
|
|
srliw s1, a5, 31
|
|
add s1, s1, a5
|
|
andi s1, s1, -2
|
|
subw a5, a5, s1
|
|
xor a2, a2, a5
|
|
snez a2, a2
|
|
slliw a5, a6, 1
|
|
srliw s1, s0, 29
|
|
slliw s0, s0, 1
|
|
or a6, a5, a2
|
|
beqz s1, .LBB3_24
|
|
j .LBB3_12
|
|
.LBB3_25:
|
|
blt a7, t1, .LBB3_11
|
|
mv t1, a7
|
|
j .LBB3_11
|
|
.LBB3_27:
|
|
li a6, 0
|
|
li t0, 1
|
|
j .LBB3_29
|
|
.LBB3_28:
|
|
slliw a2, a6, 1
|
|
ori a6, a2, 1
|
|
srliw a2, t0, 29
|
|
slliw t0, t0, 1
|
|
bnez a2, .LBB3_12
|
|
.LBB3_29:
|
|
divw a2, t1, t0
|
|
srliw a5, a2, 31
|
|
add a5, a5, a2
|
|
andi a5, a5, -2
|
|
subw a2, a2, a5
|
|
beq a2, ra, .LBB3_28
|
|
divw a2, a7, t0
|
|
srliw a5, a2, 31
|
|
add a5, a5, a2
|
|
andi a5, a5, -2
|
|
subw a2, a2, a5
|
|
beq a2, ra, .LBB3_28
|
|
slliw a6, a6, 1
|
|
srliw a2, t0, 29
|
|
slliw t0, t0, 1
|
|
beqz a2, .LBB3_29
|
|
j .LBB3_12
|
|
.LBB3_32:
|
|
li a6, 0
|
|
li s0, 1
|
|
j .LBB3_34
|
|
.LBB3_33:
|
|
slliw a6, a6, 1
|
|
srliw a2, s0, 29
|
|
slliw s0, s0, 1
|
|
bnez a2, .LBB3_12
|
|
.LBB3_34:
|
|
divw a2, t1, s0
|
|
srliw a5, a2, 31
|
|
add a5, a5, a2
|
|
andi a5, a5, -2
|
|
subw a2, a2, a5
|
|
bne a2, ra, .LBB3_33
|
|
divw a2, a7, s0
|
|
srliw a5, a2, 31
|
|
add a5, a5, a2
|
|
andi a5, a5, -2
|
|
subw a2, a2, a5
|
|
bne a2, ra, .LBB3_33
|
|
slliw a2, a6, 1
|
|
ori a6, a2, 1
|
|
srliw a2, s0, 29
|
|
slliw s0, s0, 1
|
|
beqz a2, .LBB3_34
|
|
j .LBB3_12
|
|
.LBB3_37:
|
|
li a0, 0
|
|
ld ra, 152(sp)
|
|
ld s0, 144(sp)
|
|
ld s1, 136(sp)
|
|
ld s2, 128(sp)
|
|
ld s3, 120(sp)
|
|
ld s4, 112(sp)
|
|
ld s5, 104(sp)
|
|
ld s6, 96(sp)
|
|
ld s7, 88(sp)
|
|
ld s8, 80(sp)
|
|
ld s9, 72(sp)
|
|
ld s10, 64(sp)
|
|
ld s11, 56(sp)
|
|
addi sp, sp, 160
|
|
ret
|
|
.Lfunc_end3:
|
|
.size convn, .Lfunc_end3-convn
|
|
.section .rodata,"a",@progbits
|
|
.p2align 2
|
|
.LJTI3_0:
|
|
.word .LBB3_17
|
|
.word .LBB3_23
|
|
.word .LBB3_25
|
|
.word .LBB3_27
|
|
.word .LBB3_32
|
|
|
|
.text
|
|
.globl memmove
|
|
.p2align 1
|
|
.type memmove,@function
|
|
memmove:
|
|
blez a2, .LBB4_3
|
|
slli a2, a2, 32
|
|
srli a2, a2, 32
|
|
.LBB4_2:
|
|
lw a3, 0(a1)
|
|
sw a3, 0(a0)
|
|
addi a0, a0, 4
|
|
addi a2, a2, -1
|
|
addi a1, a1, 4
|
|
bnez a2, .LBB4_2
|
|
.LBB4_3:
|
|
ret
|
|
.Lfunc_end4:
|
|
.size memmove, .Lfunc_end4-memmove
|
|
|
|
.globl main
|
|
.p2align 1
|
|
.type main,@function
|
|
main:
|
|
addi sp, sp, -96
|
|
sd ra, 88(sp)
|
|
sd s0, 80(sp)
|
|
sd s1, 72(sp)
|
|
sd s2, 64(sp)
|
|
sd s3, 56(sp)
|
|
sd s4, 48(sp)
|
|
sd s5, 40(sp)
|
|
sd s6, 32(sp)
|
|
sd s7, 24(sp)
|
|
sd s8, 16(sp)
|
|
sd s9, 8(sp)
|
|
call getint
|
|
mv s3, a0
|
|
call getint
|
|
mv s5, a0
|
|
call getint
|
|
mv s6, a0
|
|
lui a0, %hi(a)
|
|
addi s2, a0, %lo(a)
|
|
mv a0, s2
|
|
call getarray
|
|
lui a0, %hi(kernelid)
|
|
addi a0, a0, %lo(kernelid)
|
|
call getarray
|
|
mv s1, a0
|
|
li a0, 110
|
|
call _sysy_starttime
|
|
mulw s4, s6, s5
|
|
blez s1, .LBB5_6
|
|
blez s4, .LBB5_4
|
|
slli a0, s4, 32
|
|
srli s7, a0, 30
|
|
lui a0, %hi(kernelid)
|
|
addi s0, a0, %lo(kernelid)
|
|
slli a0, s1, 32
|
|
srli s1, a0, 32
|
|
lui a0, %hi(a)
|
|
addi s8, a0, %lo(a)
|
|
lui a0, %hi(b)
|
|
addi s9, a0, %lo(b)
|
|
.LBB5_3:
|
|
lw a0, 0(s0)
|
|
mv a1, s8
|
|
mv a2, s9
|
|
mv a3, s5
|
|
mv a4, s6
|
|
mv a5, s3
|
|
call convn
|
|
mv a0, s8
|
|
mv a1, s9
|
|
mv a2, s7
|
|
call memcpy@plt
|
|
addi s1, s1, -1
|
|
addi s0, s0, 4
|
|
bnez s1, .LBB5_3
|
|
j .LBB5_6
|
|
.LBB5_4:
|
|
lui a0, %hi(kernelid)
|
|
addi s0, a0, %lo(kernelid)
|
|
slli a0, s1, 32
|
|
srli s1, a0, 32
|
|
lui a0, %hi(a)
|
|
addi s7, a0, %lo(a)
|
|
lui a0, %hi(b)
|
|
addi s8, a0, %lo(b)
|
|
.LBB5_5:
|
|
lw a0, 0(s0)
|
|
mv a1, s7
|
|
mv a2, s8
|
|
mv a3, s5
|
|
mv a4, s6
|
|
mv a5, s3
|
|
call convn
|
|
addi s1, s1, -1
|
|
addi s0, s0, 4
|
|
bnez s1, .LBB5_5
|
|
.LBB5_6:
|
|
li a0, 117
|
|
call _sysy_stoptime
|
|
mv a0, s4
|
|
mv a1, s2
|
|
call putarray
|
|
li a0, 0
|
|
ld ra, 88(sp)
|
|
ld s0, 80(sp)
|
|
ld s1, 72(sp)
|
|
ld s2, 64(sp)
|
|
ld s3, 56(sp)
|
|
ld s4, 48(sp)
|
|
ld s5, 40(sp)
|
|
ld s6, 32(sp)
|
|
ld s7, 24(sp)
|
|
ld s8, 16(sp)
|
|
ld s9, 8(sp)
|
|
addi sp, sp, 96
|
|
ret
|
|
.Lfunc_end5:
|
|
.size main, .Lfunc_end5-main
|
|
|
|
.type a,@object
|
|
.bss
|
|
.globl a
|
|
.p2align 2
|
|
a:
|
|
.zero 40000000
|
|
.size a, 40000000
|
|
|
|
.type kernelid,@object
|
|
.globl kernelid
|
|
.p2align 2
|
|
kernelid:
|
|
.zero 40000
|
|
.size kernelid, 40000
|
|
|
|
.type b,@object
|
|
.globl b
|
|
.p2align 2
|
|
b:
|
|
.zero 40000000
|
|
.size b, 40000000
|
|
|
|
.type _sysy_start,@object
|
|
.globl _sysy_start
|
|
.p2align 3
|
|
_sysy_start:
|
|
.zero 16
|
|
.size _sysy_start, 16
|
|
|
|
.type _sysy_end,@object
|
|
.globl _sysy_end
|
|
.p2align 3
|
|
_sysy_end:
|
|
.zero 16
|
|
.size _sysy_end, 16
|
|
|
|
.type _sysy_l1,@object
|
|
.globl _sysy_l1
|
|
.p2align 2
|
|
_sysy_l1:
|
|
.zero 4096
|
|
.size _sysy_l1, 4096
|
|
|
|
.type _sysy_l2,@object
|
|
.globl _sysy_l2
|
|
.p2align 2
|
|
_sysy_l2:
|
|
.zero 4096
|
|
.size _sysy_l2, 4096
|
|
|
|
.type _sysy_h,@object
|
|
.globl _sysy_h
|
|
.p2align 2
|
|
_sysy_h:
|
|
.zero 4096
|
|
.size _sysy_h, 4096
|
|
|
|
.type _sysy_m,@object
|
|
.globl _sysy_m
|
|
.p2align 2
|
|
_sysy_m:
|
|
.zero 4096
|
|
.size _sysy_m, 4096
|
|
|
|
.type _sysy_s,@object
|
|
.globl _sysy_s
|
|
.p2align 2
|
|
_sysy_s:
|
|
.zero 4096
|
|
.size _sysy_s, 4096
|
|
|
|
.type _sysy_us,@object
|
|
.globl _sysy_us
|
|
.p2align 2
|
|
_sysy_us:
|
|
.zero 4096
|
|
.size _sysy_us, 4096
|
|
|
|
.type _sysy_idx,@object
|
|
.section .sbss,"aw",@nobits
|
|
.globl _sysy_idx
|
|
.p2align 2
|
|
_sysy_idx:
|
|
.word 0
|
|
.size _sysy_idx, 4
|
|
|
|
.ident "Debian clang version 14.0.6"
|
|
.section ".note.GNU-stack","",@progbits
|
|
.addrsig
|
|
.addrsig_sym a
|
|
.addrsig_sym kernelid
|
|
.addrsig_sym b
|