sysy-data/performance_c/asm/median2.s

390 lines
5.7 KiB
ArmAsm

.text
.attribute 4, 16
.attribute 5, "rv64i2p0_m2p0_a2p0_f2p0_d2p0_c2p0"
.file "median2.sy"
.globl meanless_calculation
.p2align 1
.type meanless_calculation,@function
meanless_calculation:
slti a2, a0, 1
slti a3, a1, 1
or a3, a3, a2
li a2, 0
bnez a3, .LBB0_4
addiw a1, a1, -1
addiw a2, a0, -1
bltu a1, a2, .LBB0_3
mv a1, a2
.LBB0_3:
addiw a2, a0, 2
mulw a2, a1, a2
addw a0, a0, a2
addiw a2, a1, -1
slli a1, a1, 32
slli a2, a2, 32
mulhu a1, a1, a2
srli a1, a1, 1
addw a0, a0, a1
addiw a2, a0, 1
.LBB0_4:
mv a0, a2
ret
.Lfunc_end0:
.size meanless_calculation, .Lfunc_end0-meanless_calculation
.globl swap
.p2align 1
.type swap,@function
swap:
mv a6, a0
li a0, 0
slli a4, a2, 2
add a4, a4, a6
lw a5, 0(a4)
slli a3, a1, 2
add a3, a3, a6
lw a6, 0(a3)
sw a5, 0(a3)
slti a3, a1, 1
slti a5, a2, 1
or a3, a3, a5
sw a6, 0(a4)
bnez a3, .LBB1_4
addiw a0, a2, -1
addiw a2, a1, -1
bltu a0, a2, .LBB1_3
mv a0, a2
.LBB1_3:
addiw a2, a1, 2
mulw a2, a0, a2
addiw a3, a0, -1
slli a0, a0, 32
slli a3, a3, 32
mulhu a0, a3, a0
srli a0, a0, 1
addw a1, a1, a2
addw a0, a0, a1
addiw a0, a0, 1
.LBB1_4:
ret
.Lfunc_end1:
.size swap, .Lfunc_end1-swap
.globl median
.p2align 1
.type median,@function
median:
addi sp, sp, -16
sd s0, 8(sp)
sd s1, 0(sp)
addi a6, a0, -4
addi a7, a0, 4
.LBB2_1:
sext.w a4, a1
slli a4, a4, 2
add t0, a0, a4
mv t1, a2
.LBB2_2:
lw t3, 0(t0)
li t4, 0
mv a2, t1
addiw t1, t1, 1
mv t2, a1
sext.w a4, t1
sext.w s1, t2
bge s1, a4, .LBB2_7
.LBB2_3:
sext.w t5, t2
addw a5, t1, t4
subw t1, a5, t2
slli a5, a4, 2
add a5, a5, a6
.LBB2_4:
lw s0, 0(a5)
blt s0, t3, .LBB2_11
addi a4, a4, -1
addiw t4, t4, 1
addi a5, a5, -4
blt t5, a4, .LBB2_4
mv t4, t1
mv t1, t2
.LBB2_7:
sext.w t5, t1
bge s1, t5, .LBB2_12
.LBB2_8:
slli a4, s1, 2
add a4, a4, a7
mv t6, t4
.LBB2_9:
lw a5, 0(a4)
bge a5, t3, .LBB2_13
addi s1, s1, 1
addiw t6, t6, -1
addi a4, a4, 4
bne t5, s1, .LBB2_9
j .LBB2_15
.LBB2_11:
addi t1, a4, -1
sext.w t5, t1
blt s1, t5, .LBB2_8
.LBB2_12:
mv t6, t4
sext.w a4, t2
bne a4, t5, .LBB2_14
j .LBB2_16
.LBB2_13:
addi t2, s1, 1
sext.w a4, t2
beq a4, t5, .LBB2_16
.LBB2_14:
slli a4, a4, 2
add a4, a4, a0
slli a5, t5, 2
add a5, a5, a0
lw s1, 0(a5)
lw s0, 0(a4)
sw s1, 0(a4)
sw s0, 0(a5)
mv t4, t6
sext.w a4, t1
sext.w s1, t2
blt s1, a4, .LBB2_3
j .LBB2_7
.LBB2_15:
addw a4, t4, t2
subw t6, a4, t1
.LBB2_16:
sw t3, 0(t0)
slli a4, t5, 2
add a4, a4, a0
lw a5, 0(a4)
sw a5, 0(t0)
sw t3, 0(a4)
blt a3, t5, .LBB2_2
addiw a1, t1, 1
blt t5, a3, .LBB2_1
mv a0, t6
ld s0, 8(sp)
ld s1, 0(sp)
addi sp, sp, 16
ret
.Lfunc_end2:
.size median, .Lfunc_end2-median
.globl main
.p2align 1
.type main,@function
main:
addi sp, sp, -32
sd ra, 24(sp)
sd s0, 16(sp)
sd s1, 8(sp)
lui a0, %hi(a)
addi s0, a0, %lo(a)
mv a0, s0
call getarray
lui s1, %hi(n)
sw a0, %lo(n)(s1)
li a0, 60
call _sysy_starttime
lwu a0, %lo(n)(s1)
li a7, 0
srli a1, a0, 31
addw a1, a1, a0
addiw t0, a0, -1
sraiw a6, a1, 1
j .LBB3_2
.LBB3_1:
sext.w a0, a5
addiw a7, a5, 1
bge a0, a6, .LBB3_18
.LBB3_2:
slli a0, a7, 2
add t1, a0, s0
mv a5, t0
j .LBB3_5
.LBB3_3:
mv a1, a2
.LBB3_4:
sw a4, 0(t1)
slli a0, a1, 2
add a0, a0, s0
lw a1, 0(a0)
sw a1, 0(t1)
sw a4, 0(a0)
bge a6, a2, .LBB3_1
.LBB3_5:
lw a4, 0(t1)
mv t0, a5
addiw a5, a5, 1
mv a1, a7
sext.w a2, a5
sext.w a0, a1
bge a0, a2, .LBB3_10
.LBB3_6:
sext.w a5, a1
slli a3, a2, 2
add a3, a3, s0
addi a3, a3, -4
.LBB3_7:
lw s1, 0(a3)
blt s1, a4, .LBB3_12
addi a2, a2, -1
addi a3, a3, -4
blt a5, a2, .LBB3_7
mv a5, a1
.LBB3_10:
sext.w a2, a5
blt a0, a2, .LBB3_13
.LBB3_11:
sext.w a0, a1
bne a0, a2, .LBB3_17
j .LBB3_3
.LBB3_12:
addi a5, a2, -1
sext.w a2, a5
bge a0, a2, .LBB3_11
.LBB3_13:
sext.w a1, a5
slli a3, a0, 2
add a3, a3, s0
addi a3, a3, 4
.LBB3_14:
lw s1, 0(a3)
bge s1, a4, .LBB3_16
addi a0, a0, 1
addi a3, a3, 4
bne a1, a0, .LBB3_14
j .LBB3_4
.LBB3_16:
addi a1, a0, 1
sext.w a0, a1
beq a0, a2, .LBB3_3
.LBB3_17:
slli a0, a0, 2
add a0, a0, s0
slli a2, a2, 2
add a2, a2, s0
lw a3, 0(a2)
lw s1, 0(a0)
sw a3, 0(a0)
sw s1, 0(a2)
sext.w a2, a5
sext.w a0, a1
blt a0, a2, .LBB3_6
j .LBB3_10
.LBB3_18:
li a0, 62
call _sysy_stoptime
lui s1, %hi(n)
lw a0, %lo(n)(s1)
lui a1, %hi(a)
addi s0, a1, %lo(a)
mv a1, s0
call putarray
lwu a0, %lo(n)(s1)
srli a1, a0, 31
addw a0, a0, a1
slli a0, a0, 1
andi a0, a0, -4
add a0, a0, s0
lw a0, 0(a0)
sraiw a1, a0, 31
srliw a1, a1, 24
add a1, a1, a0
andi a1, a1, -256
subw a0, a0, a1
ld ra, 24(sp)
ld s0, 16(sp)
ld s1, 8(sp)
addi sp, sp, 32
ret
.Lfunc_end3:
.size main, .Lfunc_end3-main
.type a,@object
.bss
.globl a
.p2align 2
a:
.zero 40000000
.size a, 40000000
.type n,@object
.section .sbss,"aw",@nobits
.globl n
.p2align 2
n:
.word 0
.size n, 4
.type _sysy_start,@object
.bss
.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