new and changed mappers

19 files changed, 338 insertions, 151 deletions

diff --git a/build.sh b/build.sh
index de8ecc5..cbe8e64 100755
--- a/build.sh
+++ b/build.sh
@@ -7,7 +7,7 @@ CC=gcc
 
 # Base configuration common to all builds
 CFLAGS="-std=gnu11 -mtune=generic "
-CFLAGS+="-mbmi -mbmi2 "
+CFLAGS+="-mbmi "
 CFLAGS+="-mfunction-return=keep -mindirect-branch=keep "
 CFLAGS+="-fwrapv -ffast-math -fno-trapping-math -fvisibility=hidden "
 CFLAGS+="-fno-stack-protector -fno-PIE -no-pie -fno-strict-aliasing -fcf-protection=none -ffunction-sections -fdata-sections "
@@ -19,13 +19,14 @@ CFLAGS+="-Wno-unused-variable -Wno-unused-const-variable -Wno-unused-function -W
 CFLAGS+="-U_FORTIFY_SOURCE -fno-pic "
 
 LDFLAGS="-Wl,--gc-sections -Wl,--as-needed "
+# LDFLAGS+="-Wl,--script=mknes.ld "
 
 # Base include paths (common for all platforms)
 INCLUDE_PATHS="-Ibase -I.."
 
 # Linux-specific includes and libraries
 LINUX_INCLUDE="-I/usr/include/pipewire-0.3 -I/usr/include/spa-0.2"
-LINUX_LIBS="-lpipewire-0.3 -lXi -lX11 -lGL -lm -ldl -pthread -lglfw -larchive"
+LINUX_LIBS="-lpipewire-0.3 -lXi -lX11 -lGL -lm -ldl -pthread -lglfw -larchive "
 
 # Windows-specific includes and libraries
 # WINDOWS_INCLUDE=""
diff --git a/mappers/mapper.c b/mappers/mapper.c
index 099b6d7..71caff1 100644
--- a/mappers/mapper.c
+++ b/mappers/mapper.c
@@ -23,7 +23,10 @@ static void mapper_default_chr_write(struct nes_state *state, uint32_t addr, uin
 static void mapper_default_tick(struct nes_state *state) { }
 
 #include "mapper_000_0.c"
+#include "mapper_001_0.c"
 #include "mapper_002_2.c"
+#include "mapper_003_0.c"
+#include "mapper_003_1.c"
 #include "mapper_003_2.c"
 #include "mapper_007_2.c"
 #include "mapper_011_0.c"
@@ -33,7 +36,10 @@ static void mapper_default_tick(struct nes_state *state) { }
 
 static void (*mapper_table[4096])(struct nes_state *state) = {
 	[MAPPER_ID( 0, 0)] = mapper_000_0_init,
+	[MAPPER_ID( 1, 0)] = mapper_001_0_init,
 	[MAPPER_ID( 2, 2)] = mapper_002_2_init,
+	[MAPPER_ID( 3, 0)] = mapper_003_0_init,
+	[MAPPER_ID( 3, 1)] = mapper_003_1_init,
 	[MAPPER_ID( 3, 2)] = mapper_003_2_init,
 	[MAPPER_ID( 7, 2)] = mapper_007_2_init,
 	[MAPPER_ID(11, 0)] = mapper_011_0_init,
@@ -52,6 +58,8 @@ static void mapper_reset(struct nes_state *state) {
 
 static void mapper_setup(struct nes_state *state) {
 	uint32_t mapper = state->ines.mapper << 4 | state->ines.submapper;
+	printf("Mapper %d_%x requested.\n", state->ines.mapper, state->ines.submapper);
+
 	mapper_reset(state);
 	if(mapper_table[mapper]) {
 		mapper_table[mapper](state);
diff --git a/mappers/mapper.h b/mappers/mapper.h
index 863201b..20ee906 100644
--- a/mappers/mapper.h
+++ b/mappers/mapper.h
@@ -1,6 +1,9 @@
 
 #include "mapper_000_0.h"
+#include "mapper_001_0.h"
 #include "mapper_002_2.h"
+#include "mapper_003_0.h"
+#include "mapper_003_1.h"
 #include "mapper_003_2.h"
 #include "mapper_007_2.h"
 #include "mapper_011_0.h"
@@ -20,7 +23,10 @@ struct mapper_functions {
 
 union mapper_data {
 	struct mapper_000_0 m000_0;
+	struct mapper_001_0 m001_0;
 	struct mapper_002_2 m002_2;
+	struct mapper_003_2 m003_0;
+	struct mapper_003_2 m003_1;
 	struct mapper_003_2 m003_2;
 	struct mapper_007_2 m007_2;
 	struct mapper_011_0 m011_0;
diff --git a/mappers/mapper_000_0.c b/mappers/mapper_000_0.c
index b79cdbf..8e46090 100644
--- a/mappers/mapper_000_0.c
+++ b/mappers/mapper_000_0.c
@@ -1,17 +1,18 @@
 
 
-__attribute__((hot))
+__attribute__((section(".mapper_000_0"), hot))
 static uint8_t mapper_000_0_prg_read(struct nes_state *state, uint32_t addr) {
 	struct mapper_000_0 *mapper = (struct mapper_000_0 *)&state->map;
 
 	return state->prg_rom[addr & mapper->mask];
 }
 
-__attribute__((hot))
+__attribute__((section(".mapper_000_0"), hot))
 static uint8_t mapper_000_0_chr_read(struct nes_state *state, uint32_t addr) {
 	return state->chr_rom[addr];
 }
 
+__attribute__((section(".mapper_000_0")))
 static void mapper_000_0_init(struct nes_state *state) {
 	struct mapper_000_0 *mapper = (struct mapper_000_0 *)&state->map;
 	state->mapper.prg_read = mapper_000_0_prg_read;
diff --git a/mappers/mapper_001_0.c b/mappers/mapper_001_0.c
new file mode 100644
index 0000000..519738e
--- /dev/null
+++ b/mappers/mapper_001_0.c
@@ -0,0 +1,121 @@
+
+__attribute__((section(".mapper_001_0"), hot))
+static uint8_t mapper_001_0_prg_read(struct nes_state *state, uint32_t addr) {
+	struct mapper_001_0 *mapper = (struct mapper_001_0 *)&state->map;
+	if(addr >= 0x8000) {
+		if(addr < 0xc000) {
+			return mapper->prg_rom_0[addr & 0x3fff];
+		} else {
+			return mapper->prg_rom_1[addr & 0x3fff];
+		}
+	}
+	return 0;
+}
+
+__attribute__((section(".mapper_001_0"), hot))
+static void mapper_001_0_prg_write(struct nes_state *state, uint32_t addr, uint8_t value) {
+	struct mapper_001_0 *mapper = (struct mapper_001_0 *)&state->map;
+	if(addr < 0x8000) return;
+
+	if(value & 0x80) {
+		mapper->shift = 0;
+		mapper->shift_count = 0;
+		mapper->control |= 0x0c;
+		return;
+	}
+
+	mapper->shift |= (value & 1) << mapper->shift_count;
+	mapper->shift_count++;
+
+	if(mapper->shift_count == 5) {
+		uint32_t bank;
+		uint8_t reg = (addr >> 13) & 0x03;
+		switch(reg) {
+			case 0: {// Control
+				mapper->control = mapper->shift;
+			} break;
+
+			case 1: { // CHR bank 0
+				mapper->chr_bank0 = mapper->shift;
+				mapper->chr_bank_0 = state->chr_rom + (mapper->chr_bank0 * 0x1000);
+			} break;
+
+			case 2: { // CHR bank 1
+				mapper->chr_bank1 = mapper->shift;
+				mapper->chr_bank_1 = state->chr_rom + (mapper->chr_bank1 * 0x1000);
+			} break;
+
+			case 3: { // PRG bank
+				mapper->prg_bank = mapper->shift & 0x0f;
+				if(mapper->control & 0x08) {
+					// 16KB bank switching
+					if(mapper->control & 0x04) {
+						mapper->prg_rom_0 = state->prg_rom + 0x4000 * 0;
+						mapper->prg_rom_1 = state->prg_rom + 0x4000 * mapper->prg_bank;
+					} else {
+						mapper->prg_rom_0 = state->prg_rom + 0x4000 * mapper->prg_bank;
+						mapper->prg_rom_1 = state->prg_rom + 0x4000 * (state->ines.prg_size / 0x4000 - 1);
+					}
+				} else {
+					// 32KB mode
+					bank = (mapper->prg_bank & 0x0e) * 0x4000;
+					mapper->prg_rom_0 = state->prg_rom + bank;
+					mapper->prg_rom_1 = state->prg_rom + bank + 0x4000;
+				}
+			} break;
+		}
+		mapper->shift = 0;
+		mapper->shift_count = 0;
+	}
+}
+
+__attribute__((section(".mapper_001_0"), hot))
+static uint8_t mapper_001_0_chr_read(struct nes_state *state, uint32_t addr) {
+	struct mapper_001_0 *mapper = (struct mapper_001_0 *)&state->map;
+	if(mapper->control & 0x10) {
+		// 4KB mode
+		if(addr < 0x1000) {
+			return mapper->chr_bank_0[addr];
+		} else {
+			return mapper->chr_bank_1[addr - 0x1000];
+		}
+	} else {
+		// 8KB mode
+		return mapper->chr_bank_0[addr];
+	}
+}
+
+__attribute__((section(".mapper_001_0"), hot))
+static void mapper_001_0_chr_write(struct nes_state *state, uint32_t addr, uint8_t value) {
+	// CHR RAM write (if present)
+	state->chr_ram[addr] = value;
+}
+
+__attribute__((section(".mapper_001_0"), hot))
+static uint8_t mapper_001_0_ciram_read(struct nes_state *state, uint32_t addr) {
+	return state->ciram[addr & 0x3ff];
+}
+
+__attribute__((section(".mapper_001_0"), hot))
+static void mapper_001_0_ciram_write(struct nes_state *state, uint32_t addr, uint8_t value) {
+	state->ciram[addr & 0x3ff] = value;
+}
+
+__attribute__((section(".mapper_001_0")))
+static void mapper_001_0_init(struct nes_state *state) {
+	struct mapper_001_0 *mapper = (struct mapper_001_0 *)&state->map;
+
+	mapper->control = 0x0c;
+	mapper->prg_rom_0 = state->prg_rom;
+	mapper->prg_rom_1 = state->prg_rom + (state->ines.prg_size - 0x4000);
+
+	mapper->chr_bank_0 = state->chr_rom;
+	mapper->chr_bank_1 = state->chr_rom + 0x1000;
+
+	state->mapper.prg_read		= mapper_001_0_prg_read;
+	state->mapper.prg_write		= mapper_001_0_prg_write;
+	state->mapper.chr_read		= mapper_001_0_chr_read;
+	state->mapper.chr_write		= mapper_001_0_chr_write;
+	state->mapper.ciram_read	= mapper_001_0_ciram_read;
+	state->mapper.ciram_write	= mapper_001_0_ciram_write;
+}
diff --git a/mappers/mapper_001_0.h b/mappers/mapper_001_0.h
new file mode 100644
index 0000000..deb0e29
--- /dev/null
+++ b/mappers/mapper_001_0.h
@@ -0,0 +1,14 @@
+
+struct mapper_001_0 {
+	uint8_t shift;
+	uint8_t shift_count;
+	uint8_t control;
+	uint8_t prg_bank;
+	uint8_t chr_bank0;
+	uint8_t chr_bank1;
+
+	uint8_t *prg_rom_0;
+	uint8_t *prg_rom_1;
+	uint8_t *chr_bank_0;
+	uint8_t *chr_bank_1;
+} __attribute__((packed, aligned(64)));
diff --git a/mappers/mapper_002_2.c b/mappers/mapper_002_2.c
index 079fb0b..b88c95b 100644
--- a/mappers/mapper_002_2.c
+++ b/mappers/mapper_002_2.c
@@ -1,5 +1,6 @@
 
 
+__attribute__((section(".mapper_002_2"), hot))
 static uint8_t mapper_002_2_prg_read(struct nes_state *state, uint32_t addr) {
 	struct mapper_002_2 *mapper = (struct mapper_002_2 *)&state->map;
 
@@ -12,6 +13,7 @@ static uint8_t mapper_002_2_prg_read(struct nes_state *state, uint32_t addr) {
 	return 0;
 }
 
+__attribute__((section(".mapper_002_2"), hot))
 static void mapper_002_2_prg_write(struct nes_state *state, uint32_t addr, uint8_t value) {
 	struct mapper_002_2 *mapper = (struct mapper_002_2 *)&state->map;
 
@@ -20,14 +22,17 @@ static void mapper_002_2_prg_write(struct nes_state *state, uint32_t addr, uint8
 	}
 }
 
+__attribute__((section(".mapper_002_2"), hot))
 static uint8_t mapper_002_2_chr_read(struct nes_state *state, uint32_t addr) {
 	return state->chr_ram[addr];
 }
 
+__attribute__((section(".mapper_002_2"), hot))
 static void mapper_002_2_chr_write(struct nes_state *state, uint32_t addr, uint8_t value) {
 	state->chr_ram[addr] = value;
 }
 
+__attribute__((section(".mapper_002_2")))
 static void mapper_002_2_init(struct nes_state *state) {
 	struct mapper_002_2 *mapper = (struct mapper_002_2 *)&state->map;
 
diff --git a/mappers/mapper_003_0.c b/mappers/mapper_003_0.c
new file mode 100644
index 0000000..e85d995
--- /dev/null
+++ b/mappers/mapper_003_0.c
@@ -0,0 +1,31 @@
+
+__attribute__((section(".mapper_003_0"), hot))
+static uint8_t mapper_003_0_prg_read(struct nes_state *state, uint32_t addr) {
+	return state->prg_rom[addr & 0x7fff];
+}
+
+__attribute__((section(".mapper_003_0"), hot))
+static void mapper_003_0_prg_write(struct nes_state *state, uint32_t addr, uint8_t value) {
+	struct mapper_003_0 *mapper = (struct mapper_003_0 *)&state->map;
+
+	if(addr >= 0x8000) {
+		mapper->chr_ptr = state->chr_rom + (value & 3) * 0x2000;
+	}
+}
+
+__attribute__((section(".mapper_003_0"), hot))
+static uint8_t mapper_003_0_chr_read(struct nes_state *state, uint32_t addr) {
+	struct mapper_003_0 *mapper = (struct mapper_003_0 *)&state->map;
+	return mapper->chr_ptr[addr];
+}
+
+__attribute__((section(".mapper_003_0")))
+static void mapper_003_0_init(struct nes_state *state) {
+	struct mapper_003_0 *mapper = (struct mapper_003_0 *)&state->map;
+
+	mapper->chr_ptr = state->chr_rom;
+
+	state->mapper.prg_read	= mapper_003_0_prg_read;
+	state->mapper.prg_write	= mapper_003_0_prg_write;
+	state->mapper.chr_read	= mapper_003_0_chr_read;
+}
diff --git a/mappers/mapper_003_0.h b/mappers/mapper_003_0.h
new file mode 100644
index 0000000..f03b106
--- /dev/null
+++ b/mappers/mapper_003_0.h
@@ -0,0 +1,4 @@
+
+struct mapper_003_0 {
+	uint8_t *chr_ptr;
+} __attribute__((packed, aligned(64)));
diff --git a/mappers/mapper_003_1.c b/mappers/mapper_003_1.c
new file mode 100644
index 0000000..fd7cd27
--- /dev/null
+++ b/mappers/mapper_003_1.c
@@ -0,0 +1,31 @@
+
+__attribute__((section(".mapper_003_1"), hot))
+static uint8_t mapper_003_1_prg_read(struct nes_state *state, uint32_t addr) {
+	return state->prg_rom[addr & 0x7fff];
+}
+
+__attribute__((section(".mapper_003_1"), hot))
+static void mapper_003_1_prg_write(struct nes_state *state, uint32_t addr, uint8_t value) {
+	struct mapper_003_1 *mapper = (struct mapper_003_1 *)&state->map;
+
+	if(addr >= 0x8000) {
+		mapper->chr_ptr = state->chr_rom + (value & 3) * 0x2000;
+	}
+}
+
+__attribute__((section(".mapper_003_1"), hot))
+static uint8_t mapper_003_1_chr_read(struct nes_state *state, uint32_t addr) {
+	struct mapper_003_1 *mapper = (struct mapper_003_1 *)&state->map;
+	return mapper->chr_ptr[addr];
+}
+
+__attribute__((section(".mapper_003_1")))
+static void mapper_003_1_init(struct nes_state *state) {
+	struct mapper_003_1 *mapper = (struct mapper_003_1 *)&state->map;
+
+	mapper->chr_ptr = state->chr_rom;
+
+	state->mapper.prg_read	= mapper_003_1_prg_read;
+	state->mapper.prg_write	= mapper_003_1_prg_write;
+	state->mapper.chr_read	= mapper_003_1_chr_read;
+}
diff --git a/mappers/mapper_003_1.h b/mappers/mapper_003_1.h
new file mode 100644
index 0000000..c233a1a
--- /dev/null
+++ b/mappers/mapper_003_1.h
@@ -0,0 +1,4 @@
+
+struct mapper_003_1 {
+	uint8_t *chr_ptr;
+} __attribute__((packed, aligned(64)));
diff --git a/mappers/mapper_003_2.c b/mappers/mapper_003_2.c
index e93916a..b780a8f 100644
--- a/mappers/mapper_003_2.c
+++ b/mappers/mapper_003_2.c
@@ -1,21 +1,26 @@
 
+__attribute__((section(".mapper_003_2"), hot))
 static uint8_t mapper_003_2_prg_read(struct nes_state *state, uint32_t addr) {
 	return state->prg_rom[addr & 0x7fff];
 }
 
+__attribute__((section(".mapper_003_2"), hot))
 static void mapper_003_2_prg_write(struct nes_state *state, uint32_t addr, uint8_t value) {
 	struct mapper_003_2 *mapper = (struct mapper_003_2 *)&state->map;
 
 	if(addr >= 0x8000) {
+		value &= state->prg_rom[addr & 0x7fff];
 		mapper->chr_ptr = state->chr_rom + (value & 3) * 0x2000;
 	}
 }
 
+__attribute__((section(".mapper_003_2"), hot))
 static uint8_t mapper_003_2_chr_read(struct nes_state *state, uint32_t addr) {
 	struct mapper_003_2 *mapper = (struct mapper_003_2 *)&state->map;
 	return mapper->chr_ptr[addr];
 }
 
+__attribute__((section(".mapper_003_2")))
 static void mapper_003_2_init(struct nes_state *state) {
 	struct mapper_003_2 *mapper = (struct mapper_003_2 *)&state->map;
 
diff --git a/mappers/mapper_007_2.c b/mappers/mapper_007_2.c
index ed56dd0..27125b9 100644
--- a/mappers/mapper_007_2.c
+++ b/mappers/mapper_007_2.c
@@ -1,5 +1,5 @@
 
-
+__attribute__((section(".mapper_007_2"), hot))
 static uint8_t mapper_007_2_prg_read(struct nes_state *state, uint32_t addr) {
 	struct mapper_007_2 *mapper = (struct mapper_007_2 *)&state->map;
 	if(addr >= 0x8000) {
@@ -8,6 +8,7 @@ static uint8_t mapper_007_2_prg_read(struct nes_state *state, uint32_t addr) {
 	return 0;
 }
 
+__attribute__((section(".mapper_007_2"), hot))
 static void mapper_007_2_prg_write(struct nes_state *state, uint32_t addr, uint8_t value) {
 	struct mapper_007_2 *mapper = (struct mapper_007_2 *)&state->map;
 	if(addr >= 0x8000) {
@@ -21,24 +22,29 @@ static void mapper_007_2_prg_write(struct nes_state *state, uint32_t addr, uint8
 	}
 }
 
+__attribute__((section(".mapper_007_2"), hot))
 static uint8_t mapper_007_2_chr_read(struct nes_state *state, uint32_t addr) {
 	return state->chr_ram[addr];
 }
 
+__attribute__((section(".mapper_007_2"), hot))
 static void mapper_007_2_chr_write(struct nes_state *state, uint32_t addr, uint8_t value) {
 	state->chr_ram[addr] = value;
 }
 
+__attribute__((section(".mapper_007_2"), hot))
 static uint8_t mapper_007_2_ciram_read(struct nes_state *state, uint32_t addr) {
 	struct mapper_007_2 *mapper = (struct mapper_007_2 *)&state->map;
 	return mapper->ciram[addr & 0x3ff];
 }
 
+__attribute__((section(".mapper_007_2"), hot))
 static void mapper_007_2_ciram_write(struct nes_state *state, uint32_t addr, uint8_t value) {
 	struct mapper_007_2 *mapper = (struct mapper_007_2 *)&state->map;
 	mapper->ciram[addr & 0x3ff] = value;
 }
 
+__attribute__((section(".mapper_007_2")))
 static void mapper_007_2_init(struct nes_state *state) {
 	struct mapper_007_2 *mapper = (struct mapper_007_2 *)&state->map;
 	mapper->prg_rom = state->prg_rom;
diff --git a/mappers/mapper_011_0.c b/mappers/mapper_011_0.c
index ce390a0..d6d518c 100644
--- a/mappers/mapper_011_0.c
+++ b/mappers/mapper_011_0.c
@@ -1,5 +1,5 @@
 
-
+__attribute__((section(".mapper_011_0"), hot))
 static uint8_t mapper_011_0_prg_read(struct nes_state *state, uint32_t addr) {
 	struct mapper_011_0 *mapper = (struct mapper_011_0 *)&state->map;
 
@@ -9,6 +9,7 @@ static uint8_t mapper_011_0_prg_read(struct nes_state *state, uint32_t addr) {
 	return 0;
 }
 
+__attribute__((section(".mapper_011_0"), hot))
 static void mapper_011_0_prg_write(struct nes_state *state, uint32_t addr, uint8_t value) {
 	struct mapper_011_0 *mapper = (struct mapper_011_0 *)&state->map;
 
@@ -18,12 +19,14 @@ static void mapper_011_0_prg_write(struct nes_state *state, uint32_t addr, uint8
 	}
 }
 
+__attribute__((section(".mapper_011_0"), hot))
 static uint8_t mapper_011_0_chr_read(struct nes_state *state, uint32_t addr) {
 	struct mapper_011_0 *mapper = (struct mapper_011_0 *)&state->map;
 
 	return mapper->chr_ptr[addr];
 }
 
+__attribute__((section(".mapper_011_0")))
 static void mapper_011_0_init(struct nes_state *state) {
 	struct mapper_011_0 *mapper = (struct mapper_011_0 *)&state->map;
 
diff --git a/mappers/mapper_066_0.c b/mappers/mapper_066_0.c
index bf31720..7a8f49c 100644
--- a/mappers/mapper_066_0.c
+++ b/mappers/mapper_066_0.c
@@ -1,4 +1,5 @@
 
+__attribute__((section(".mapper_066_0"), hot))
 static uint8_t mapper_066_0_prg_read(struct nes_state *state, uint32_t addr) {
 	struct mapper_066_0 *mapper = (struct mapper_066_0 *)&state->map;
 
@@ -8,6 +9,7 @@ static uint8_t mapper_066_0_prg_read(struct nes_state *state, uint32_t addr) {
 	return 0;
 }
 
+__attribute__((section(".mapper_066_0"), hot))
 static void mapper_066_0_prg_write(struct nes_state *state, uint32_t addr, uint8_t value) {
 	struct mapper_066_0 *mapper = (struct mapper_066_0 *)&state->map;
 
@@ -20,11 +22,13 @@ static void mapper_066_0_prg_write(struct nes_state *state, uint32_t addr, uint8
 	}
 }
 
+__attribute__((section(".mapper_066_0"), hot))
 static uint8_t mapper_066_0_chr_read(struct nes_state *state, uint32_t addr) {
 	struct mapper_066_0 *mapper = (struct mapper_066_0 *)&state->map;
 	return mapper->chr_offset[addr];
 }
 
+__attribute__((section(".mapper_066_0")))
 static void mapper_066_0_init(struct nes_state *state) {
 	struct mapper_066_0 *mapper = (struct mapper_066_0 *)&state->map;
 
diff --git a/mknes.c b/mknes.c
index b5f05d3..7220450 100644
--- a/mknes.c
+++ b/mknes.c
@@ -13,8 +13,8 @@
 #define WINDOW_WIDTH 320 * 3
 #define WINDOW_HEIGHT 240 * 3
 
-#define PRG_ROM_SIZE	(512 * 1024)
-#define CHR_ROM_SIZE	(512 * 1024)
+#define PRG_ROM_SIZE	(2 * 1024 * 1024)
+#define CHR_ROM_SIZE	(1 * 1024 * 1024)
 #define PIXELS_SIZE	(256 * 240)
 #define RAM_SIZE		0x1000		// 0x800 in reality, but for aligned alloc it must be the size of the alignment (4096)
 #define SRAM_SIZE		0x2000
@@ -104,62 +104,8 @@ static uint32_t frames;		// debug information
 #include "mappers/mapper.c"
 
 #include "callbacks.c"
-struct nes_state nstate;
-
-
-static struct nes_state *allocate_nes_state(void) {
-	struct nes_state *state = (struct nes_state*)calloc(1, sizeof(struct nes_state));
-	if(!state) return 0;
-
-	state->prg_rom	= aligned_alloc(4096, PRG_ROM_SIZE);
-	state->chr_rom	= aligned_alloc(4096, CHR_ROM_SIZE);
-	state->pixels	= aligned_alloc(4096, PIXELS_SIZE);
-	state->ram		= aligned_alloc(4096, RAM_SIZE);
-	state->sram		= aligned_alloc(4096, SRAM_SIZE);
-	state->ciram	= aligned_alloc(4096, CIRAM_SIZE);
-	state->chr_ram	= aligned_alloc(4096, CHR_RAM_SIZE);
-	memset(state->prg_rom, 0, PRG_ROM_SIZE);
-	memset(state->chr_rom, 0, CHR_ROM_SIZE);
-	memset(state->pixels, 0, PIXELS_SIZE);
-	memset(state->ram, 0, RAM_SIZE);
-	memset(state->sram, 0, SRAM_SIZE);
-	memset(state->ciram, 0, CIRAM_SIZE);
-	memset(state->chr_ram, 0, CHR_RAM_SIZE);
-	return state;
-}
-
-static void free_nes_state(struct nes_state *s) {
-	free(s->prg_rom);
-	free(s->chr_rom);
-	free(s->pixels);
-	free(s->ram);
-	free(s->sram);
-	free(s->ciram);
-	free(s->chr_ram);
-	free(s);
-}
-
-
-
-
-static void dump_nametable_text(struct nes_state *state, const char *filename) {
-	FILE *f = fopen(filename, "w");
-	if(!f) return;
-
-	for(int y = 0; y < 30; y++) {
-		for(int x = 0; x < 32; x++) {
-			uint16_t addr = y * 32 + x;
-			uint8_t tile = state->ciram[addr];	// assuming NT0 is mapped to 0x000
-			fprintf(f, "%02x ", tile);
-		}
-		fprintf(f, "\n");
-	}
-	fclose(f);
-}
-
-
-
 
+struct nes_state nstate;
 
 int main(int argc, char **argv) {
 #ifdef _WIN32
@@ -167,70 +113,72 @@ int main(int argc, char **argv) {
 #endif
 
 	state.toggle_crt_emulation = 1;
-	struct nes_state *nstate = allocate_nes_state();
 
 	setbuf(stdout, 0);
 	init_opcode_lut();
 	init_opcode_ud_lut();
 	// protect_opcode_lut();
-	ppu_reset(nstate);
-	// ines2_load(nstate, "data/0000/10-Yard Fight (USA, Europe).nes");
-	// ines2_load(nstate, "data/0000/Balloon Fight (USA).nes");
-	// ines2_load(nstate, "data/0000/Excitebike (Japan, USA).nes");
-	// ines2_load(nstate, "data/0000/Ice Climber (USA, Europe, Korea).nes");
-	// ines2_load(nstate, "data/0000/Kung Fu (Japan, USA).nes");
-	ines2_load(nstate, "data/0000/Super Mario Bros. (World) (HVC-SM).nes");
-	// ines2_load(nstate, "data/Super Mario Bros. (W) (V1.0) [!].nes");
-	// ines2_load(nstate, "data/Super Mario Bros. (JU) [!].nes");
-	// ines2_load(nstate, "data/0000/Urban Champion (World).nes");
-	// ines2_load(nstate, "data/0000/Wrecking Crew (World).nes");
-	// ines2_load(nstate, "data/0000/scanline.nes");
-	// ines2_load(nstate, "data/0000/Sayoonara!.NES");
-	// ines2_load(nstate, "data/0000/raster_demos/RasterChromaLuma.NES");
-	// ines2_load(nstate, "data/0000/raster_demos/RasterTest1.NES");
-	// ines2_load(nstate, "data/0000/raster_demos/RasterTest2.NES");
-	// ines2_load(nstate, "data/0000/raster_demos/RasterTest3.NES");
-	// ines2_load(nstate, "data/0000/raster_demos/RasterTest3a.NES");
-	// ines2_load(nstate, "data/0000/raster_demos/RasterTest3b.NES");
-	// ines2_load(nstate, "data/0000/raster_demos/RasterTest3c.NES");
-	// ines2_load(nstate, "data/0000/raster_demos/RasterTest3d.NES");
-	// ines2_load(nstate, "data/0000/raster_demos/RasterTest3e.NES");
-	// ines2_load(nstate, "data/0000/NEStress.NES");
-	// ines2_load(nstate, "data/0000/Super Mario Bros. (World) (HVC-SM).zip");
-	// ines2_load(nstate, "data/0042/Super Mario Bros. + Duck Hunt (USA).zip");
-	// ines2_load(nstate, "data/0000/Xevious - The Avenger (USA).zip");
-	// ines2_load(nstate, "data/tv.nes");
-
-	// ines2_load(nstate, "data/Life Force (USA).zip");	// 2002
-
-	// ines2_load(nstate, "data/0003/Flipull - An Exciting Cube Game (Japan) (En).zip");
-	// ines2_load(nstate, "data/0003/Friday the 13th (USA).zip");
-	// ines2_load(nstate, "data/0003/Ghostbusters (Japan).zip");
-	ines2_load(nstate, "data/0003/Gradius (USA).zip");
-	// ines2_load(nstate, "data/0007/Battletoads (USA).zip");
-	// ines2_load(nstate, "data/0007/Beetlejuice (USA).zip");
-	// ines2_load(nstate, "data/0007/Cabal (USA).zip");
-
-	// ines2_load(nstate, "data/000b/Baby Boomer (USA) (Unl).zip");
-	// ines2_load(nstate, "data/000b/Captain Comic - The Adventure (USA) (Unl).zip");
-	// ines2_load(nstate, "data/000b/King Neptune's Adventure (USA) (Unl).zip");
-
-	// ines2_load(nstate, "data/2002/Attack Animal Gakuen (Japan).zip");
-	// ines2_load(nstate, "data/2002/Ballblazer (Japan).zip");
-	// ines2_load(nstate, "data/2002/Best of the Best - Championship Karate (USA).zip");
-
-	mapper_setup(nstate);
-	uint32_t lo = nstate->mapper.prg_read(nstate, 0xfffc);
-	uint32_t hi = nstate->mapper.prg_read(nstate, 0xfffd);
-	nstate->cpu.pc = (hi << 8) | lo;
-
-#if 0
+	ppu_reset(&nstate);
+	// ines2_load(&nstate, "data/0000/10-Yard Fight (USA, Europe).nes");
+	// ines2_load(&nstate, "data/0000/Balloon Fight (USA).nes");
+	// ines2_load(&nstate, "data/0000/Excitebike (Japan, USA).nes");
+	// ines2_load(&nstate, "data/0000/Ice Climber (USA, Europe, Korea).nes");
+	// ines2_load(&nstate, "data/0000/Kung Fu (Japan, USA).nes");
+	ines2_load(&nstate, "data/0000/Super Mario Bros. (World) (HVC-SM).nes");
+	// ines2_load(&nstate, "data/Super Mario Bros. (W) (V1.0) [!].nes");
+	// ines2_load(&nstate, "data/Super Mario Bros. (JU) [!].nes");
+	// ines2_load(&nstate, "data/0000/Urban Champion (World).nes");
+	// ines2_load(&nstate, "data/0000/Wrecking Crew (World).nes");
+	// ines2_load(&nstate, "data/0000/scanline.nes");
+	// ines2_load(&nstate, "data/0000/Sayoonara!.NES");
+	// ines2_load(&nstate, "data/0000/raster_demos/RasterChromaLuma.NES");
+	// ines2_load(&nstate, "data/0000/raster_demos/RasterTest1.NES");
+	// ines2_load(&nstate, "data/0000/raster_demos/RasterTest2.NES");
+	// ines2_load(&nstate, "data/0000/raster_demos/RasterTest3.NES");
+	// ines2_load(&nstate, "data/0000/raster_demos/RasterTest3a.NES");
+	// ines2_load(&nstate, "data/0000/raster_demos/RasterTest3b.NES");
+	// ines2_load(&nstate, "data/0000/raster_demos/RasterTest3c.NES");
+	// ines2_load(&nstate, "data/0000/raster_demos/RasterTest3d.NES");
+	// ines2_load(&nstate, "data/0000/raster_demos/RasterTest3e.NES");
+	// ines2_load(&nstate, "data/0000/NEStress.NES");
+	// ines2_load(&nstate, "data/0000/Super Mario Bros. (World) (HVC-SM).zip");
+	// ines2_load(&nstate, "data/0042/Super Mario Bros. + Duck Hunt (USA).zip");
+	// ines2_load(&nstate, "data/0000/Xevious - The Avenger (USA).zip");
+	// ines2_load(&nstate, "data/tv.nes");
+
+	ines2_load(&nstate, "data/Life Force (USA).zip");	// 2002
+
+	// ines2_load(&nstate, "data/0003/Flipull - An Exciting Cube Game (Japan) (En).zip");
+	// ines2_load(&nstate, "data/0003/Friday the 13th (USA).zip");
+	// ines2_load(&nstate, "data/0003/Ghostbusters (Japan).zip");
+	// ines2_load(&nstate, "data/0003/Gradius (USA).zip");
+	// ines2_load(&nstate, "data/0007/Battletoads (USA).zip");
+	// ines2_load(&nstate, "data/0007/Beetlejuice (USA).zip");
+	// ines2_load(&nstate, "data/0007/Cabal (USA).zip");
+
+	// ines2_load(&nstate, "data/000b/Baby Boomer (USA) (Unl).zip");
+	// ines2_load(&nstate, "data/000b/Captain Comic - The Adventure (USA) (Unl).zip");
+	// ines2_load(&nstate, "data/000b/King Neptune's Adventure (USA) (Unl).zip");
+
+	// ines2_load(&nstate, "data/2002/Attack Animal Gakuen (Japan).zip");
+	// ines2_load(&nstate, "data/2002/Ballblazer (Japan).zip");
+	// ines2_load(&nstate, "data/2002/Best of the Best - Championship Karate (USA).zip");
+
+	// ines2_load(&nstate, "data/Blaster Master (USA).zip");
+
+
+	mapper_setup(&nstate);
+	uint32_t lo = nstate.mapper.prg_read(&nstate, 0xfffc);
+	uint32_t hi = nstate.mapper.prg_read(&nstate, 0xfffd);
+	nstate.cpu.pc = (hi << 8) | lo;
+
+#if 1
 	for(uint32_t i = 0; i < 0x5000; ++ i) {
-		while(!nstate->ppu.frame_ready) {
+		while(!nstate.ppu.frame_ready) {
 			// PROFILE_NAMED("nes emulator");
-			cpu_tick(nstate);
+			cpu_tick(&nstate);
 		}
-		nstate->ppu.frame_ready = 0;
+		nstate.ppu.frame_ready = 0;
 		frames++;
 	}
 	return 0;
@@ -250,8 +198,8 @@ int main(int argc, char **argv) {
 
 		window = glfwCreateWindow(WINDOW_WIDTH, WINDOW_HEIGHT, "NES Emulator", 0, 0);
 		if(window) {
-			glfwSetWindowUserPointer(window, (void*)nstate);
-			glfwSetWindowAspectRatio(window, 320, 240);		// Need to set a 4:3 resolution for it to look correct!
+			glfwSetWindowUserPointer(window, (void*)&nstate);
+			glfwSetWindowAspectRatio(window, 320, 240);		// Need to set a 4:3 resolution for things to look correct!
 			glfwSetWindowSizeLimits(window, WINDOW_WIDTH, WINDOW_HEIGHT, GLFW_DONT_CARE, GLFW_DONT_CARE);
 
 			glfwMakeContextCurrent(window);
@@ -278,22 +226,15 @@ int main(int argc, char **argv) {
 				timer_wait(timer);
 				glfwPollEvents();
 
-// printf("Frame: %d\n", frames);
-
-
-// static int32_t tas_frame = 0;
-// 				nstate->input[0] = tas_input[tas_frame++];
-				while(!nstate->ppu.frame_ready) {
+				while(!nstate.ppu.frame_ready) {
 					// PROFILE_NAMED("nes emulator");
-					cpu_tick(nstate);
+					cpu_tick(&nstate);
 				}
-				nstate->ppu.frame_ready = 0;
+				nstate.ppu.frame_ready = 0;
 				frames++;
 
-// dump_nametable_text(nstate, "_foofbomb.txt");
-
 				uint32_t * restrict dst = display_buffer;	//buffer;
-				uint8_t * restrict src = nstate->pixels;
+				uint8_t * restrict src = nstate.pixels;
 				for(uint32_t y = 0; y < 240; ++y) {
 					for(uint32_t x = 0; x < 256; ++x) {
 						uint8_t val = *src++;
@@ -306,8 +247,10 @@ int main(int argc, char **argv) {
 				render_frame();
 				glfwSwapBuffers(window);
 			}
-printf("total frames: %6.6d  total cycles: %ld\n", frames, nstate->cpu.cycles);
+
+			printf("total frames: %6.6d  total cycles: %ld\n", frames, nstate.cpu.cycles);
 			glfwDestroyWindow(window);
+
 		} else {
 			fprintf(stderr, "Failed to create window\n");
 		}
@@ -317,7 +260,7 @@ printf("total frames: %6.6d  total cycles: %ld\n", frames, nstate->cpu.cycles);
 		fprintf(stderr, "Failed to initialize GLFW\n");
 	}
 	timer_destroy(timer);
-	free_nes_state(nstate);
+	// free_nes_state(&nstate);
 #ifdef _WIN32
 	timeEndPeriod(1);
 #endif
diff --git a/mknes.h b/mknes.h
index e1dc8f2..fba1406 100644
--- a/mknes.h
+++ b/mknes.h
@@ -117,15 +117,14 @@ struct nes_state {
 	struct cpu_state cpu;
 	// struct apu_state apu;
 
-	uint8_t *ram;
-	uint8_t *ciram;
-	uint8_t *prg_rom;
-	uint8_t *chr_rom;
-	uint8_t *chr_ram;
-	uint8_t *pixels;
-	uint8_t *sram;
-} __attribute__((packed, aligned(64)));
-
+	uint8_t ram[RAM_SIZE]			__attribute__((aligned(4096)));
+	uint8_t ciram[CIRAM_SIZE]		__attribute__((aligned(4096)));
+	uint8_t prg_rom[PRG_ROM_SIZE]	__attribute__((aligned(4096)));
+	uint8_t chr_rom[CHR_ROM_SIZE]	__attribute__((aligned(4096)));
+	uint8_t chr_ram[CHR_RAM_SIZE]	__attribute__((aligned(4096)));
+	uint8_t pixels[PIXELS_SIZE]	__attribute__((aligned(4096)));
+	uint8_t sram[SRAM_SIZE]			__attribute__((aligned(4096)));
+} __attribute__((packed, aligned(4096)));
 
 __attribute__((aligned(4096))) static uint32_t nes_palette[65] = {
 	0x585858ff, 0x00237cff, 0x0d1099ff, 0x300092ff, 0x4f006cff, 0x600035ff, 0x5c0500ff, 0x461800ff,
diff --git a/opengl.c b/opengl.c
index 255eac3..953a494 100644
--- a/opengl.c
+++ b/opengl.c
@@ -5,6 +5,7 @@
 #include "vertex_shader.h"
 
 /* [=]===^=[ compile_shader ]==============================================================^===[=] */
+__attribute__((cold, noinline, section(".init_section")))
 static GLuint compile_shader(GLenum shader_type, const char *shader_source) {
 	GLuint shader = glCreateShader(shader_type);
 	glShaderSource(shader, 1, &shader_source, 0);
@@ -21,6 +22,7 @@ static GLuint compile_shader(GLenum shader_type, const char *shader_source) {
 }
 
 /* [=]===^=[ opengl_setup ]================================================================^===[=] */
+__attribute__((cold, noinline, section(".init_section")))
 static void opengl_setup(void) {
 	gl_loader();
 
diff --git a/ppu.c b/ppu.c
index fe54be5..6f5a4eb 100644
--- a/ppu.c
+++ b/ppu.c
@@ -155,19 +155,18 @@ static inline void ppu_render_pixel(struct nes_state *state) {
 	}
 
 	// Final pixel composition
-	uint8_t palette_index = 0;
 	uint8_t bg_index = (bg_palette << 2) + bg_pixel;
 	uint8_t sp_index = (sp_palette << 2) + sp_pixel;
 	uint8_t selector = (bg_pixel ? 2 : 0) | (sp_pixel ? 1 : 0);
 
+	// uint8_t palette_index;
+	uint8_t palette_index = (sp_prio) ? bg_index : 0x10 | sp_index;
+
 	switch(selector) {
-		case 0: { palette_index = 0;						} break;
-		case 1: { palette_index = 0x10 | sp_index;	} break;
-		case 2: { palette_index = bg_index;				} break;
-		case 3: {
-			ppu->reg_status |= (sp_zero && x < 255) ? 0x40 : 0;		// NOTE(peter): Sprite zero hit!
-			palette_index = (sp_prio) ? bg_index : 0x10 | sp_index;
-		} break;
+		case 0: { palette_index = 0;												} break;
+		case 1: { palette_index = 0x10 | sp_index;							} break;
+		case 2: { palette_index = bg_index;										} break;
+		case 3: { ppu->reg_status |= (sp_zero && x < 255) ? 0x40 : 0;	} break;		// NOTE(peter): Sprite zero hit!
 	}
 
 	state->pixels[y * 256 + x] = ppu->palette[palette_index];		// NOTE(peter): Add color_emphasis bits (expand palette to 8x).