From 385f77fb480f7bdebbc72e0b24e528085e0dadf2 Mon Sep 17 00:00:00 2001 From: dark-samus Date: Mon, 23 Mar 2020 06:35:43 -0400 Subject: [PATCH 1/2] match the final non-matching in boot! --- src/boot/yaz0.c | 31 +++++++++++++------------------ 1 file changed, 13 insertions(+), 18 deletions(-) diff --git a/src/boot/yaz0.c b/src/boot/yaz0.c index 5954e28aaa..379d4ca467 100644 --- a/src/boot/yaz0.c +++ b/src/boot/yaz0.c @@ -51,10 +51,6 @@ void* Yaz0_NextDMA(void* curSrcPos) { return dst; } -#ifdef NON_MATCHING -// has an extra check at the start of the small compressed loop -// along with an extra move instruction -// maybe some regalloc differences besides these issues void Yaz0_DecompressImpl(Yaz0Header* hdr, u8* dst) { u32 bitIdx; u8* src; @@ -63,46 +59,45 @@ void Yaz0_DecompressImpl(Yaz0Header* hdr, u8* dst) { u32 nibble; u8* backPtr; s32 chunkSize; - u32 curDecBit; u32 off; bitIdx = 0; - src = hdr->data; + src = (u8*)hdr->data; dstEnd = dst + hdr->decSize; do { if (bitIdx == 0) { - if ((sYaz0MaxPtr < (u32)src) && (sYaz0CurSize != 0)) + if ((sYaz0MaxPtr < (u32)src) && (sYaz0CurSize != 0)) { src = Yaz0_NextDMA(src); + } chunkHeader = *src++; bitIdx = 8; } - curDecBit = chunkHeader & (1 << 7); - chunkHeader <<= 1; - - if (curDecBit) { // uncompressed - *dst++ = *src++; + if (chunkHeader & (1 << 7)) { // uncompressed + *dst = *src; + dst++; + src++; } else { // compressed + off = ((*src & 0xF) << 8 | *(src + 1)); nibble = *src >> 4; - backPtr = dst - (*(src + 1) | (*src & 0xF) << 8); + backPtr = dst - off; src += 2; chunkSize = (nibble == 0) // N = chunkSize; B = back offset ? *src++ + 0x12 // 3 bytes 0B BB NN : nibble + 2; // 2 bytes NB BB - for (off = chunkSize; off > 0; off--) + do { *dst++ = *(backPtr++ - 1); + chunkSize--; + } while (chunkSize != 0); } - + chunkHeader <<= 1; bitIdx--; } while (dst != dstEnd); } -#else -#pragma GLOBAL_ASM("asm/non_matchings/boot/yaz0/Yaz0_DecompressImpl.s") -#endif void Yaz0_Decompress(u32 romStart, void* dst, u32 size) { sYaz0CurRomStart = romStart; From 7db0ed672a1c4517d64a559b7a8644be81731d2c Mon Sep 17 00:00:00 2001 From: dark-samus Date: Mon, 23 Mar 2020 06:54:52 -0400 Subject: [PATCH 2/2] remove asm for Yaz0_DecompressImpl --- .../boot/yaz0/Yaz0_DecompressImpl.s | 99 ------------------- 1 file changed, 99 deletions(-) delete mode 100644 asm/non_matchings/boot/yaz0/Yaz0_DecompressImpl.s diff --git a/asm/non_matchings/boot/yaz0/Yaz0_DecompressImpl.s b/asm/non_matchings/boot/yaz0/Yaz0_DecompressImpl.s deleted file mode 100644 index 4219a519cc..0000000000 --- a/asm/non_matchings/boot/yaz0/Yaz0_DecompressImpl.s +++ /dev/null @@ -1,99 +0,0 @@ -glabel Yaz0_DecompressImpl -/* 0028B0 80001CB0 27BDFFC0 */ addiu $sp, $sp, -0x40 -/* 0028B4 80001CB4 AFBF002C */ sw $ra, 0x2c($sp) -/* 0028B8 80001CB8 AFB40028 */ sw $s4, 0x28($sp) -/* 0028BC 80001CBC AFB30024 */ sw $s3, 0x24($sp) -/* 0028C0 80001CC0 AFB20020 */ sw $s2, 0x20($sp) -/* 0028C4 80001CC4 AFB1001C */ sw $s1, 0x1c($sp) -/* 0028C8 80001CC8 AFB00018 */ sw $s0, 0x18($sp) -/* 0028CC 80001CCC 8C8E0004 */ lw $t6, 4($a0) -/* 0028D0 80001CD0 3C128001 */ lui $s2, %hi(sYaz0MaxPtr) # $s2, 0x8001 -/* 0028D4 80001CD4 3C138001 */ lui $s3, %hi(sYaz0CurSize) # $s3, 0x8001 -/* 0028D8 80001CD8 00A08025 */ move $s0, $a1 -/* 0028DC 80001CDC 00003025 */ move $a2, $zero -/* 0028E0 80001CE0 24910010 */ addiu $s1, $a0, 0x10 -/* 0028E4 80001CE4 267344E8 */ addiu $s3, %lo(sYaz0CurSize) # addiu $s3, $s3, 0x44e8 -/* 0028E8 80001CE8 265244EC */ addiu $s2, %lo(sYaz0MaxPtr) # addiu $s2, $s2, 0x44ec -/* 0028EC 80001CEC 8FA70030 */ lw $a3, 0x30($sp) -/* 0028F0 80001CF0 01C5A021 */ addu $s4, $t6, $a1 -.L80001CF4: -/* 0028F4 80001CF4 54C0000F */ bnezl $a2, .L80001D34 -/* 0028F8 80001CF8 30F90080 */ andi $t9, $a3, 0x80 -/* 0028FC 80001CFC 8E4F0000 */ lw $t7, ($s2) -/* 002900 80001D00 01F1082B */ sltu $at, $t7, $s1 -/* 002904 80001D04 50200008 */ beql $at, $zero, .L80001D28 -/* 002908 80001D08 92270000 */ lbu $a3, ($s1) -/* 00290C 80001D0C 8E780000 */ lw $t8, ($s3) -/* 002910 80001D10 53000005 */ beql $t8, $zero, .L80001D28 -/* 002914 80001D14 92270000 */ lbu $a3, ($s1) -/* 002918 80001D18 0C0006F2 */ jal Yaz0_NextDMA -/* 00291C 80001D1C 02202025 */ move $a0, $s1 -/* 002920 80001D20 00408825 */ move $s1, $v0 -/* 002924 80001D24 92270000 */ lbu $a3, ($s1) -.L80001D28: -/* 002928 80001D28 26310001 */ addiu $s1, $s1, 1 -/* 00292C 80001D2C 24060008 */ li $a2, 8 -/* 002930 80001D30 30F90080 */ andi $t9, $a3, 0x80 -.L80001D34: -/* 002934 80001D34 13200006 */ beqz $t9, .L80001D50 -/* 002938 80001D38 00073840 */ sll $a3, $a3, 1 -/* 00293C 80001D3C 92280000 */ lbu $t0, ($s1) -/* 002940 80001D40 26100001 */ addiu $s0, $s0, 1 -/* 002944 80001D44 26310001 */ addiu $s1, $s1, 1 -/* 002948 80001D48 10000027 */ b .L80001DE8 -/* 00294C 80001D4C A208FFFF */ sb $t0, -1($s0) -.L80001D50: -/* 002950 80001D50 92230000 */ lbu $v1, ($s1) -/* 002954 80001D54 92290001 */ lbu $t1, 1($s1) -/* 002958 80001D58 26310002 */ addiu $s1, $s1, 2 -/* 00295C 80001D5C 306A000F */ andi $t2, $v1, 0xf -/* 002960 80001D60 000A5A00 */ sll $t3, $t2, 8 -/* 002964 80001D64 00032903 */ sra $a1, $v1, 4 -/* 002968 80001D68 012B2025 */ or $a0, $t1, $t3 -/* 00296C 80001D6C 14A00005 */ bnez $a1, .L80001D84 -/* 002970 80001D70 02041023 */ subu $v0, $s0, $a0 -/* 002974 80001D74 92230000 */ lbu $v1, ($s1) -/* 002978 80001D78 26310001 */ addiu $s1, $s1, 1 -/* 00297C 80001D7C 10000002 */ b .L80001D88 -/* 002980 80001D80 24630012 */ addiu $v1, $v1, 0x12 -.L80001D84: -/* 002984 80001D84 24A30002 */ addiu $v1, $a1, 2 -.L80001D88: -/* 002988 80001D88 30650003 */ andi $a1, $v1, 3 -/* 00298C 80001D8C 00052823 */ negu $a1, $a1 -/* 002990 80001D90 10A00009 */ beqz $a1, .L80001DB8 -/* 002994 80001D94 00A32021 */ addu $a0, $a1, $v1 -.L80001D98: -/* 002998 80001D98 904CFFFF */ lbu $t4, -1($v0) -/* 00299C 80001D9C 2463FFFF */ addiu $v1, $v1, -1 -/* 0029A0 80001DA0 26100001 */ addiu $s0, $s0, 1 -/* 0029A4 80001DA4 24420001 */ addiu $v0, $v0, 1 -/* 0029A8 80001DA8 1483FFFB */ bne $a0, $v1, .L80001D98 -/* 0029AC 80001DAC A20CFFFF */ sb $t4, -1($s0) -/* 0029B0 80001DB0 1060000D */ beqz $v1, .L80001DE8 -/* 0029B4 80001DB4 00000000 */ nop -.L80001DB8: -/* 0029B8 80001DB8 904DFFFF */ lbu $t5, -1($v0) -/* 0029BC 80001DBC 2463FFFC */ addiu $v1, $v1, -4 -/* 0029C0 80001DC0 26100004 */ addiu $s0, $s0, 4 -/* 0029C4 80001DC4 A20DFFFC */ sb $t5, -4($s0) -/* 0029C8 80001DC8 904E0000 */ lbu $t6, ($v0) -/* 0029CC 80001DCC 24420004 */ addiu $v0, $v0, 4 -/* 0029D0 80001DD0 A20EFFFD */ sb $t6, -3($s0) -/* 0029D4 80001DD4 904FFFFD */ lbu $t7, -3($v0) -/* 0029D8 80001DD8 A20FFFFE */ sb $t7, -2($s0) -/* 0029DC 80001DDC 9058FFFE */ lbu $t8, -2($v0) -/* 0029E0 80001DE0 1460FFF5 */ bnez $v1, .L80001DB8 -/* 0029E4 80001DE4 A218FFFF */ sb $t8, -1($s0) -.L80001DE8: -/* 0029E8 80001DE8 1614FFC2 */ bne $s0, $s4, .L80001CF4 -/* 0029EC 80001DEC 24C6FFFF */ addiu $a2, $a2, -1 -/* 0029F0 80001DF0 8FBF002C */ lw $ra, 0x2c($sp) -/* 0029F4 80001DF4 AFA70030 */ sw $a3, 0x30($sp) -/* 0029F8 80001DF8 8FB00018 */ lw $s0, 0x18($sp) -/* 0029FC 80001DFC 8FB1001C */ lw $s1, 0x1c($sp) -/* 002A00 80001E00 8FB20020 */ lw $s2, 0x20($sp) -/* 002A04 80001E04 8FB30024 */ lw $s3, 0x24($sp) -/* 002A08 80001E08 8FB40028 */ lw $s4, 0x28($sp) -/* 002A0C 80001E0C 03E00008 */ jr $ra -/* 002A10 80001E10 27BD0040 */ addiu $sp, $sp, 0x40 \ No newline at end of file