Back at 2618fps

2 files changed, 30 insertions, 27 deletions

diff --git a/mknes.h b/mknes.h
index 10ae358..f038fa6 100644
--- a/mknes.h
+++ b/mknes.h
@@ -21,12 +21,16 @@
 #define MIRROR_FOURSCREEN 2
 
 struct ppu_state {
+	uint16_t bg_shift_pattern_low;
+	uint16_t bg_shift_pattern_high;
+	uint16_t bg_shift_attrib_low;
+	uint16_t bg_shift_attrib_high;
+
 	uint32_t scanline;
 	uint32_t dot;
-	uint32_t bg_shift_pattern_low;
-	uint32_t bg_shift_pattern_high;
-	uint32_t bg_shift_attrib_low;
-	uint32_t bg_shift_attrib_high;
+	uint32_t vram_addr;
+	uint32_t temp_addr;
+	uint32_t fine_x;
 	uint8_t bg_next_tile_id;
 	uint8_t bg_next_tile_attrib;
 	uint8_t bg_next_tile_lsb;
@@ -35,26 +39,13 @@ struct ppu_state {
 	uint8_t oam_data;
 	uint8_t even_frame;
 
-	uint8_t oam[256];
-	uint8_t secondary_oam[32];
-	uint8_t palette[0x20];
-	uint8_t sprite_indexes[8];
-	uint32_t sprite_patterns[8];
-	uint8_t sprite_positions[8];
-	uint8_t sprite_priorities[8];
-	uint8_t sprite_shift_lo[8];
-	uint8_t sprite_shift_hi[8];
-
 	uint8_t reg_ctrl;
 	uint8_t reg_mask;
 	uint8_t reg_status;
 
-	uint32_t vram_addr;
-	uint32_t temp_addr;
-	uint32_t fine_x;
+	uint8_t write_latch;
 
 	uint8_t vram_read_buffer;
-	uint8_t write_latch;
 
 	uint8_t open_bus;
 
@@ -62,7 +53,17 @@ struct ppu_state {
 
 	uint8_t sprite_zero_hit_possible;
 	uint8_t sprite_count;
-} __attribute__((aligned(64)));
+
+	uint8_t sprite_indexes[8];
+	uint8_t sprite_positions[8];
+	uint8_t sprite_priorities[8];
+	uint8_t sprite_shift_lo[8];
+	uint8_t sprite_shift_hi[8];
+	uint8_t palette[32];
+	uint8_t secondary_oam[32];
+	uint8_t oam[256];
+
+} __attribute__((packed, aligned(64)));
 
 struct cpu_state {
 	uint32_t pc;	// Program Counter
@@ -80,7 +81,7 @@ struct cpu_state {
 	uint8_t c;		// Carry Flag
 // --
 	uint8_t die;	// KIL instruction found!
-} __attribute__((aligned(64)));
+};// __attribute__((aligned(64)));
 
 
 struct ines_state {
@@ -93,18 +94,18 @@ struct ines_state {
 
 struct nes_state {
 	size_t cycles;
-	struct ines_state ines;
 	struct cpu_state cpu;
 	uint8_t irq_pending;
 	uint8_t nmi_pending;
 	uint8_t input[2];				// Controller 1 & 2
 	uint8_t input_latch[2];		// Latched inputs after strobe
-	uint8_t input_strobe;		// Control bit (0 or 1)
 	uint8_t input_bit[2];		// Current bit position being shifted out
+	uint8_t input_strobe;		// Control bit (0 or 1)
 	struct ppu_state ppu;
 
 	struct mapper_entry mapper;
 	union mapper_data map;
+	struct ines_state ines;
 
 	uint8_t *pixels;
 	uint8_t *ram;
diff --git a/ppu.c b/ppu.c
index cc223b7..cef45fa 100644
--- a/ppu.c
+++ b/ppu.c
@@ -11,14 +11,14 @@ static uint8_t __attribute__((aligned(64))) ppu_bitreverse_lut[256] = {
 #undef R6
 
 static void ppu_reset(struct nes_state *state) {
-	struct ppu_state *ppu = &state->ppu;
+	struct ppu_state *restrict ppu = &state->ppu;
 	memset(ppu, 0, sizeof(struct ppu_state));
 }
 
 
 __attribute__((always_inline, hot))
 static inline void ppu_evaluate_sprites(struct nes_state *state) {
-	struct ppu_state *ppu = &state->ppu;
+	struct ppu_state *restrict ppu = &state->ppu;
 	uint8_t sprite_height = (ppu->reg_ctrl & 0x20) ? 16 : 8;
 	uint8_t n = 0;
 
@@ -52,7 +52,7 @@ static inline void ppu_evaluate_sprites(struct nes_state *state) {
 
 __attribute__((always_inline, hot))
 static inline void ppu_fetch_sprite_patterns(struct nes_state *state) {
-	struct ppu_state *ppu = &state->ppu;
+	struct ppu_state *restrict ppu = &state->ppu;
 	uint32_t addr;
 	uint32_t bank;
 
@@ -87,6 +87,7 @@ static inline void ppu_fetch_sprite_patterns(struct nes_state *state) {
 
 		ppu->sprite_shift_lo[i] = lsb;
 		ppu->sprite_shift_hi[i] = msb;
+
 		ppu->sprite_positions[i] = x;
 		ppu->sprite_priorities[i] = attr & 0x20;
 		s += 4;
@@ -95,7 +96,7 @@ static inline void ppu_fetch_sprite_patterns(struct nes_state *state) {
 
 __attribute__((always_inline, hot))
 static inline void ppu_render_pixel(struct nes_state *state) {
-	struct ppu_state *ppu = &state->ppu;
+	struct ppu_state *restrict ppu = &state->ppu;
 
 	uint32_t x = ppu->dot - 1;
 	uint32_t y = ppu->scanline;
@@ -160,7 +161,8 @@ static inline void ppu_render_pixel(struct nes_state *state) {
 
 __attribute__((hot, flatten))
 static void ppu_tick(struct nes_state *state) {
-	struct ppu_state *ppu = &state->ppu;
+	struct ppu_state *restrict ppu = &state->ppu;
+
 
 	uint32_t dot = ppu->dot;
 	uint32_t scanline = ppu->scanline;