sysy-data/hidden_functional_c/asm/22_matrix_multiply.s

374 lines
5.4 KiB
ArmAsm
Raw Permalink Normal View History

2024-06-14 13:34:46 +08:00
.text
.attribute 4, 16
.attribute 5, "rv64i2p0_m2p0_a2p0_f2p0_d2p0_c2p0"
.file "22_matrix_multiply.sy"
.globl matrix_multiply
.p2align 1
.type matrix_multiply,@function
matrix_multiply:
addi sp, sp, -16
sd s0, 8(sp)
sd s1, 0(sp)
lui a0, %hi(m1)
lw a1, %lo(m1)(a0)
blez a1, .LBB0_8
lui a0, %hi(n2)
lw a2, %lo(n2)(a0)
lui a0, %hi(n1)
lw a3, %lo(n1)(a0)
sgtz a0, a2
sgtz a4, a3
and a0, a0, a4
beqz a0, .LBB0_8
li t4, 0
slli a0, a1, 32
srli a6, a0, 32
slli a0, a2, 32
srli t1, a0, 32
slli a0, a3, 32
srli t2, a0, 32
lui a0, %hi(a)
addi t5, a0, %lo(a)
lui a0, %hi(b)
addi a7, a0, %lo(b)
li t0, 400
lui a0, %hi(res)
addi t3, a0, %lo(res)
.LBB0_3:
li a1, 0
mv a4, a7
.LBB0_4:
mul a0, t4, t0
slli a2, a1, 2
add a0, a0, a2
add t6, a0, t3
lw a5, 0(t6)
mv a0, t2
mv a2, a4
mv a3, t5
.LBB0_5:
lw s0, 0(a3)
lw s1, 0(a2)
mulw s0, s1, s0
addw a5, a5, s0
addi a3, a3, 4
addi a0, a0, -1
addi a2, a2, 400
bnez a0, .LBB0_5
sw a5, 0(t6)
addi a1, a1, 1
addi a4, a4, 4
bne a1, t1, .LBB0_4
addi t4, t4, 1
addi t5, t5, 400
bne t4, a6, .LBB0_3
.LBB0_8:
ld s0, 8(sp)
ld s1, 0(sp)
addi sp, sp, 16
ret
.Lfunc_end0:
.size matrix_multiply, .Lfunc_end0-matrix_multiply
.globl main
.p2align 1
.type main,@function
main:
addi sp, sp, -64
sd ra, 56(sp)
sd s0, 48(sp)
sd s1, 40(sp)
sd s2, 32(sp)
sd s3, 24(sp)
sd s4, 16(sp)
sd s5, 8(sp)
sd s6, 0(sp)
call getint
lui s2, %hi(m1)
sw a0, %lo(m1)(s2)
call getint
lw a2, %lo(m1)(s2)
mv a1, a0
lui s5, %hi(n1)
sgtz a0, a2
sgtz a3, a1
and a0, a0, a3
sw a1, %lo(n1)(s5)
beqz a0, .LBB1_7
li s4, 0
lui a0, %hi(a)
addi s3, a0, %lo(a)
j .LBB1_3
.LBB1_2:
addi s4, s4, 1
addi s3, s3, 400
bge s4, a2, .LBB1_7
.LBB1_3:
sext.w a0, a1
blez a0, .LBB1_2
li s0, 0
mv s1, s3
.LBB1_5:
call getint
lw a1, %lo(n1)(s5)
sw a0, 0(s1)
addi s0, s0, 1
addi s1, s1, 4
blt s0, a1, .LBB1_5
lw a2, %lo(m1)(s2)
j .LBB1_2
.LBB1_7:
call getint
lui s2, %hi(m2)
sw a0, %lo(m2)(s2)
call getint
lw a2, %lo(m2)(s2)
mv s6, a0
lui s5, %hi(n2)
sgtz a0, a2
sgtz a3, s6
and a0, a0, a3
sw s6, %lo(n2)(s5)
beqz a0, .LBB1_14
li s4, 0
lui a0, %hi(b)
addi s3, a0, %lo(b)
j .LBB1_10
.LBB1_9:
addi s4, s4, 1
addi s3, s3, 400
bge s4, a2, .LBB1_14
.LBB1_10:
sext.w a0, s6
blez a0, .LBB1_9
li s0, 0
mv s1, s3
.LBB1_12:
call getint
lw s6, %lo(n2)(s5)
sw a0, 0(s1)
addi s0, s0, 1
addi s1, s1, 4
blt s0, s6, .LBB1_12
lw a2, %lo(m2)(s2)
j .LBB1_9
.LBB1_14:
lui a0, %hi(m1)
lwu a7, %lo(m1)(a0)
sext.w a6, a7
blez a6, .LBB1_28
lui a0, %hi(n1)
lw a0, %lo(n1)(a0)
sext.w a2, s6
sgtz a2, a2
sgtz a3, a0
and a2, a2, a3
beqz a2, .LBB1_23
li t5, 0
slli a0, a0, 32
srli t2, a0, 32
lui a0, %hi(a)
addi t6, a0, %lo(a)
slli a0, s6, 32
srli t3, a0, 32
lui a0, %hi(b)
addi t0, a0, %lo(b)
li t1, 400
lui a0, %hi(res)
addi t4, a0, %lo(res)
.LBB1_17:
li s0, 0
mv a3, t0
.LBB1_18:
mul a0, t5, t1
slli a2, s0, 2
add a0, a0, a2
add s2, a0, t4
lw s1, 0(s2)
mv a5, t6
mv a0, a3
mv a2, t2
.LBB1_19:
lw a1, 0(a5)
lw a4, 0(a0)
mulw a1, a4, a1
addw s1, s1, a1
addi a2, a2, -1
addi a0, a0, 400
addi a5, a5, 4
bnez a2, .LBB1_19
sw s1, 0(s2)
addi s0, s0, 1
addi a3, a3, 4
bne s0, t3, .LBB1_18
addi t5, t5, 1
addi t6, t6, 400
bne t5, a7, .LBB1_17
blez a6, .LBB1_28
.LBB1_23:
li s4, 0
lui a0, %hi(res)
addi s2, a0, %lo(res)
lui s5, %hi(n2)
lui s3, %hi(m1)
sext.w a0, s6
blez a0, .LBB1_26
.LBB1_24:
li s0, 0
mv s1, s2
.LBB1_25:
lw a0, 0(s1)
call putint
li a0, 32
call putch
lw a0, %lo(n2)(s5)
addi s0, s0, 1
addi s1, s1, 4
blt s0, a0, .LBB1_25
.LBB1_26:
li a0, 10
call putch
lw a0, %lo(m1)(s3)
addi s4, s4, 1
bge s4, a0, .LBB1_28
lw s6, %lo(n2)(s5)
addi s2, s2, 400
sext.w a0, s6
bgtz a0, .LBB1_24
j .LBB1_26
.LBB1_28:
li a0, 0
ld ra, 56(sp)
ld s0, 48(sp)
ld s1, 40(sp)
ld s2, 32(sp)
ld s3, 24(sp)
ld s4, 16(sp)
ld s5, 8(sp)
ld s6, 0(sp)
addi sp, sp, 64
ret
.Lfunc_end1:
.size main, .Lfunc_end1-main
.type m1,@object
.section .sbss,"aw",@nobits
.globl m1
.p2align 2
m1:
.word 0
.size m1, 4
.type n2,@object
.globl n2
.p2align 2
n2:
.word 0
.size n2, 4
.type n1,@object
.globl n1
.p2align 2
n1:
.word 0
.size n1, 4
.type res,@object
.bss
.globl res
.p2align 2
res:
.zero 40000
.size res, 40000
.type a,@object
.globl a
.p2align 2
a:
.zero 40000
.size a, 40000
.type b,@object
.globl b
.p2align 2
b:
.zero 40000
.size b, 40000
.type m2,@object
.section .sbss,"aw",@nobits
.globl m2
.p2align 2
m2:
.word 0
.size m2, 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