dranke/upkr
1
0
Fork 0
mirror of https://github.com/exoticorn/upkr.git synced 2026-01-20 11:36:42 +01:00
Code Issues Projects Releases Wiki Activity

some more optimizations to the rv unpacker

Browse Source
This commit is contained in:
Dennis Ranke
2022-10-03 08:47:12 +02:00
parent 32cd8e5b6c
commit a46eb0e7f5
1 changed files with 28 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files

60
asm_unpackers/unpack_riscv.S
View File

@@ -1,13 +1,11 @@
.section .text .section .text
#define FRAME_SIZE (256+32*4+4)
// x8 prob array ptr
// x9 prev was literal // x9 prev was literal
// x10 out ptr // x10 out ptr
// x11 in ptr // x11 in ptr
// x12 offset // x12 offset
// x13 state // x13 state
// x14 context index
.global upkr_unpack .global upkr_unpack
.type upkr_unpack, %function .type upkr_unpack, %function
@@ -15,11 +13,11 @@ upkr_unpack:
mv t4, ra mv t4, ra
mv x17, x8 mv x17, x8
mv t6, x9 mv t6, x9
li x13, FRAME_SIZE li x9, 256 + 128
li x9, 128 mv x13, x9
1: 1:
addi sp, sp, -1 sub x8, sp, x13
sb x9, 0(sp) sb x9, 0(x8)
addi x13, x13, -1 addi x13, x13, -1
bnez x13, 1b bnez x13, 1b
@@ -35,7 +33,7 @@ upkr_unpack:
.Lfinished_offset: .Lfinished_offset:
addi x14, x14, 64 addi x14, x14, 64
jal t3, upkr_decode_number jal upkr_decode_number
1: 1:
add x14, x10, t0 add x14, x10, t0
lbu x14, (x14) lbu x14, (x14)
@@ -58,11 +56,10 @@ upkr_unpack:
.Lread_offset_inc_x14: .Lread_offset_inc_x14:
addi x14, x14, 1 addi x14, x14, 1
.Lread_offset: .Lread_offset:
jal t3, upkr_decode_number jal upkr_decode_number
addi t0, x9, 1 addi t0, x9, 1
bnez t0, .Lfinished_offset bnez t0, .Lfinished_offset
.Ldone: .Ldone:
addi sp, sp, FRAME_SIZE
mv x8, x17 mv x8, x17
mv x9, t6 mv x9, t6
jr t4 jr t4
@@ -70,20 +67,21 @@ upkr_unpack:
// x14 context index // x14 context index
// return: x9 negtive decoded number // return: x9 negtive decoded number
upkr_decode_number: upkr_decode_number:
mv t3, ra
mv t5, x14 mv t5, x14
li x9, 0 li x9, 0
li x8, -1 li t1, -1
1: 1:
jal upkr_decode_bit jal upkr_decode_bit
beqz x15, 1f beqz x15, 1f
jal upkr_decode_bit jal upkr_decode_bit
beqz x15, 2f beqz x15, 2f
add x9, x9, x8 add x9, x9, t1
2: 2:
slli x8, x8, 1 add t1, t1, t1
j 1b j 1b
1: 1:
add x9, x9, x8 add x9, x9, t1
mv x14, t5 mv x14, t5
jr t3 jr t3
@@ -104,39 +102,37 @@ upkr_decode_bit:
srli x15, x13, 12 srli x15, x13, 12
beqz x15, upkr_load_byte beqz x15, upkr_load_byte
mv t1, x14 addi x14, x14, 1
mv t2, x10
add x14, x14, sp sub sp, sp, x14
lbu x12, 0(x14) lbu x12, 0(sp)
andi x10, x13, 255 andi x8, x13, 255
sltu x15, x10, x12 sltu x15, x8, x12
srli x13, x13, 8 srli x13, x13, 8
beqz x15, .Lelse beqz x15, .Lelse
mul x13, x13, x12 mul x13, x13, x12
add x13, x13, x10 add x13, x13, x8
li x10, 256 + 8 li x8, 256 + 8
sub x10, x10, x12 sub x8, x8, x12
srli x10, x10, 4 srli x8, x8, 4
add x12, x12, x10 add x12, x12, x8
j .Lendif j .Lendif
.Lelse: .Lelse:
li x16, 256 li x16, 256
sub x16, x16, x12 sub x16, x16, x12
mul x13, x13, x16 mul x13, x13, x16
add x13, x13, x10 add x13, x13, x8
sub x13, x13, x12 sub x13, x13, x12
addi x10, x12, 8 addi x8, x12, 8
srli x10, x10, 4 srli x8, x8, 4
sub x12, x12, x10 sub x12, x12, x8
.Lendif: .Lendif:
sb x12, 0(x14) sb x12, 0(sp)
add sp, sp, x14
addi x14, t1, 1
mv x10, t2
ret ret