add and use lzss decompressor, baku variant

kernel.asm

@@ -100,15 +100,15 @@ Drive64Done:
     nop
     nop
 
-    // try out an interrupt:
-    //sw      r0, 0(r0)
-    mfc0    t1, CP0_Status
-    ori     t1, 2
-    mtc0    t1, CP0_Status
-    la      t0, WipeRegisters
-    mtc0    t0, CP0_EPC
-    j       InterruptHandler
-    nop
+//  // try out an interrupt:
+//  //sw      r0, 0(r0)
+//  mfc0    t1, CP0_Status
+//  ori     t1, 2
+//  mtc0    t1, CP0_Status
+//  la      t0, WipeRegisters
+//  mtc0    t0, CP0_EPC
+//  j       InterruptHandler
+//  nop
 
 WipeRegisters:
     // load up most registers with a dummy value for debugging

lz.asm

@@ -3,7 +3,7 @@
 
 LZInit:
     la a0, LZ+4 // A0 = Source Address
-    lui a1, 0xA010 // A1 = Destination Address (DRAM Start Offset)
+    lui a1, 0x8010 // A1 = Destination Address (DRAM Start Offset)
 
     lbu t0, -1(a0) // T0 = HI Data Length Byte
     sll t0, 8

lzss.baku.asm

@@ -0,0 +1,134 @@
+LzDecomp:
+    // a0: pointer to compressed data (must be RDRAM, cart is unsupported)
+    // a1: compressed size
+    // a2: output pointer
+    // a3: maximum uncompressed size
+    // v0: error code (0 is OK)
+
+    // t0: current pointer for reading
+    // t1: end pointer (exclusive)
+    // t2: current pointer for writing
+    // t3: end pointer (exclusive)
+    // t4: current code, shifted as necessary
+    // t5: 1 means raw, 0 means copy
+    // t6: either raw byte or part of copy command
+    // t7: current pseudo-position
+    // t8: match length
+    // t9: match position
+
+    beqz    a1, LzExit // nothing to decompress? nothing to do!
+    nop
+
+    blez    a3, LzNoSpace
+    nop
+
+    or      t0, a0, r0
+    addu    t1, a0, a1
+    or      t2, a2, r0
+    addu    t3, a2, a3
+
+    lli     t7, 0x3BE
+
+LzNextCode:
+    // READ
+    lbu     t4, 0(t0)
+    addiu   t0, 1
+
+    ori     t4, t4, 0x100 // add end marker
+
+    andi    t5, t4, 1
+LzReadCode:
+    beq     t2, t3, LzReiterate
+    nop
+
+    // READ
+    beq     t0, t1, LzBadEnd
+    nop
+    lbu     t6, 0(t0)
+    addiu   t0, 1
+
+    beqz    t5, LzCopy
+    nop
+
+LzRaw:
+    // WRITE
+    beq     t2, t3, LzNoSpace
+    addiu   t7, 1
+    andi    t7, t7, 0x3FF
+    sb      t6, 0(t2)
+    addiu   t2, 1
+
+    // SHIFT
+    srl     t4, 1
+    xori    at, t4, 1
+    bnez    at, LzReadCode // branch if t4 != 1
+    andi    t5, t4, 1
+
+    b       LzReiterate
+    nop
+
+LzCopy:
+    // READ
+    beq     t0, t1, LzBadEnd
+    nop
+    lbu     t8, 0(t0)
+    addiu   t0, 1
+
+    // extract match position
+    sll     t9, t8, 2
+    andi    t9, 0x0300
+    or      t9, t6
+
+    // extract and prepare match length
+    andi    t8, 0x003F
+    addiu   t8, 3
+
+    // prepare absolute match position
+    subu    t9, t7, t9
+    subiu   t9, 1
+    andi    t9, 0x03FF
+    addiu   t9, 1
+    subu    t9, t2, t9
+
+    // pre-emptively move psuedo-position ahead (don't need it for the loop)
+    addu    t7, t8
+    andi    t7, t7, 0x3FF
+
+LzCopyLoop:
+    // repurposing t6: byte being copied
+    slt     at, t9, a2
+    bnez    at,+ // don't copy bytes before the output pointer
+    lli     t6, 0 // use nulls instead
+    lbu     t6, 0(t9)
++
+
+    // WRITE
+    beq     t2, t3, LzNoSpace
+    subiu   t8, 1
+    sb      t6, 0(t2)
+    addiu   t2, 1
+
+    bnez    t8, LzCopyLoop
+    addiu   t9, 1
+
+    // SHIFT
+    srl     t4, 1
+    xori    at, t4, 1
+    bnez    at, LzReadCode // branch if t4 != 1
+    andi    t5, t4, 1
+
+LzReiterate:
+    bne     t0, t1, LzNextCode
+    nop
+
+LzExit:
+    jr      ra
+    lli     v0, 0
+
+LzNoSpace:
+    jr      ra
+    lli     v0, 1
+
+LzBadEnd: // aka Unexpected End Of File
+    jr      ra
+    lli     v0, 2

lzss.baku.unsafe.asm

@@ -0,0 +1,123 @@
+LzDecomp:
+    // a0: pointer to compressed data (must be RDRAM, cart is unsupported)
+    // a1: compressed size
+    // a2: output pointer
+    // a3: maximum uncompressed size
+    // v0: error code (0 is OK)
+
+    // t0: current pointer for reading
+    // t1: end pointer (exclusive)
+    // t2: current pointer for writing
+    // t3: end pointer (exclusive)
+    // t4: current code, shifted as necessary
+    // t5: 1 means raw, 0 means copy
+    // t6: either raw byte or part of copy command
+    // t7: current pseudo-position
+    // t8: match length
+    // t9: match position
+
+    beqz    a1, LzExit // nothing to decompress? nothing to do!
+    lli     v0, 0
+
+    blez    a3, LzExit
+    lli     v0, 1
+
+    or      t0, a0, r0
+    addu    t1, a0, a1
+    or      t2, a2, r0
+    addu    t3, a2, a3
+
+    lli     t7, 0x3BE
+
+LzNextCode:
+    // READ
+    lbu     t4, 0(t0)
+    addiu   t0, 1
+
+    ori     t4, t4, 0x100 // add end marker
+
+    andi    t5, t4, 1
+LzReadCode:
+    beq     t2, t3, LzExit
+    lli     v0, 0
+
+    // READ
+    lbu     t6, 0(t0)
+    addiu   t0, 1
+
+    beqz    t5, LzCopy
+    nop
+
+LzRaw:
+    // WRITE
+    addiu   t7, 1
+    andi    t7, t7, 0x3FF
+    sb      t6, 0(t2)
+    addiu   t2, 1
+
+    // SHIFT
+    srl     t4, 1
+    xori    at, t4, 1
+    bnez    at, LzReadCode // branch if t4 != 1
+    andi    t5, t4, 1
+
+    // reiterate
+    bne     t0, t1, LzNextCode
+    nop
+
+    b       LzExit
+    lli     v0, 0
+
+LzCopy:
+    // READ
+    lbu     t8, 0(t0)
+    addiu   t0, 1
+
+    // extract match position
+    sll     t9, t8, 2
+    andi    t9, 0x0300
+    or      t9, t6
+
+    // extract and prepare match length
+    andi    t8, 0x003F
+    addiu   t8, 3
+
+    // prepare absolute match position
+    subu    t9, t7, t9
+    subiu   t9, 1
+    andi    t9, 0x03FF
+    addiu   t9, 1
+    subu    t9, t2, t9
+
+    // pre-emptively move psuedo-position ahead (don't need it for the loop)
+    addu    t7, t8
+    andi    t7, t7, 0x3FF
+
+LzCopyLoop:
+    // repurposing t6: byte being copied
+    lbu     t6, 0(t9)
+
+    // WRITE
+    subiu   t8, 1
+    sb      t6, 0(t2)
+    addiu   t2, 1
+
+    bnez    t8, LzCopyLoop
+    addiu   t9, 1
+
+    // SHIFT
+    srl     t4, 1
+    xori    at, t4, 1
+    bnez    at, LzReadCode // branch if t4 != 1
+    andi    t5, t4, 1
+
+    // reiterate
+    bne     t0, t1, LzNextCode
+    nop
+
+    b       LzExit
+    lli     v0, 0
+
+LzExit:
+    jr      ra
+    nop

main.asm

@@ -27,17 +27,29 @@ Main:
     lui     t0, K_BASE
 
     lui     s0, BLAH_BASE
-
-    mfc0    t0, CP0_Count
     mfc0    t1, CP0_Status+0
-    sw      t0, BLAH_COUNTS+0(s0)
     sw      t1, 8(s0)
 
-// decompress our picture
-include "lz.asm"
+    nop; nop; nop; nop
+    mfc0    t0, CP0_Count
+    sw      t0, BLAH_COUNTS+0(s0)
+
+    // decompress our picture
+    la      a0, LZ_BAKU + 4
+    lw      a3, -4(a0) // load uncompressed size from the file itself
+    li      a1, LZ_BAKU.size - 4
+    li      a2, VIDEO_BUFFER | 0x80000000
+    jal     LzDecomp
+    nop
+    // TODO: flush cache on video buffer
 
     mfc0    t0, CP0_Count
+    nop; nop; nop; nop
+
+    lw      t1, BLAH_COUNTS+0(s0)
     sw      t0, BLAH_COUNTS+8(s0)
+    subu    t1, t0, t1
+    sw      t1, BLAH_COUNTS+0xC(s0)
 
     lui     a0, BLAH_BASE
     lli     a1, 0x20
@@ -150,8 +162,10 @@ PushRSPTask:
     jr      ra
     nop
 
+include "lzss.baku.unsafe.asm"
+
 align(16); insert F3DZEX_BOOT, "bin/F3DZEX2.boot.bin"
 align(16); insert F3DZEX_DMEM, "bin/F3DZEX2.data.bin"
 align(16); insert F3DZEX_IMEM, "bin/F3DZEX2.bin"
 align(16); insert FONT, "res/dwarf.1bpp"
-align(16); insert LZ, "res/Image.lz"
+align(16); insert LZ_BAKU, "res/Image.baku.lzss"