tests

1 files changed, 39 insertions, 36 deletions

diff --git a/ppu.c b/ppu.c
index e0fce44..3b6200e 100644
--- a/ppu.c
+++ b/ppu.c
@@ -60,20 +60,23 @@ static inline void ppu_evaluate_sprites(struct nes_state *state) {
 __attribute__((hot))
 static inline void ppu_fetch_sprite_patterns(struct nes_state *state) {
 	struct ppu_state *restrict ppu = &state->ppu;
-	uint32_t addr;
-	uint32_t bank;
-	uint8_t lsb;
-	uint8_t msb;
 
 	uint8_t * restrict s = ppu->secondary_oam;
 	uint8_t height = (ppu->reg_ctrl & 0x20) ? 16 : 8;
+	uint8_t ctrl = ppu->reg_ctrl;
+	uint8_t scanline = ppu->scanline;
+	uint32_t sprite_pattern_table_base = (ctrl & 0x08) << 9;
 
 	for(uint8_t i = 0; i < ppu->sprite_count; i++) {
-		uint8_t y = s[0], tile = s[1], attr = s[2], x = s[3];
-		uint8_t row = ppu->scanline - y;
+		uint8_t y = s[0];
+		uint8_t tile = s[1];
+		uint8_t attr = s[2];
+		uint8_t x = s[3];
 
+		uint8_t row = scanline - y;
 		row = (attr & 0x80) ? height - 1 - row : row;
 
+		uint32_t bank, addr;
 		if(height == 16) {
 			bank = (tile & 1) << 12;
 			tile &= 0xfe;
@@ -81,29 +84,29 @@ static inline void ppu_fetch_sprite_patterns(struct nes_state *state) {
 				tile++;
 				row -= 8;
 			}
-		} else {
-			bank = (ppu->reg_ctrl & 0x08) << 9;
-		}
-		addr = bank + tile * 16 + row;
+			addr = bank + tile * 16 + row;
 
-		if(attr & 0x40) {
-			lsb = ppu_bitreverse_lut[state->mapper_function.chr_read(state, addr)];
-			msb = ppu_bitreverse_lut[state->mapper_function.chr_read(state, addr + 8)];
 		} else {
-			lsb = state->mapper_function.chr_read(state, addr);
-			msb = state->mapper_function.chr_read(state, addr + 8);
+			addr = sprite_pattern_table_base + tile * 16 + row;
 		}
 
+		uint8_t val_lo = state->mapper_function.chr_read(state, addr);
+		uint8_t val_hi = state->mapper_function.chr_read(state, addr + 8);
+
+		uint8_t rev = -(!!(attr & 0x40));
+		uint8_t lsb = (rev & ppu_bitreverse_lut[val_lo]) | (~rev & val_lo);
+		uint8_t msb = (rev & ppu_bitreverse_lut[val_hi]) | (~rev & val_hi);
+
 		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;
 	}
 }
 
-__attribute__((always_inline, hot, flatten))
+__attribute__((always_inline, hot))
 static inline void ppu_render_pixel(struct nes_state *state) {
 	struct ppu_state *restrict ppu = &state->ppu;
 
@@ -112,8 +115,6 @@ static inline void ppu_render_pixel(struct nes_state *state) {
 
 	uint16_t bit = 0x8000 >> ppu->fine_x;
 
-	uint8_t bg_pixel = 0;
-	uint8_t bg_palette = 0;
 	uint8_t sp_pixel = 0;
 	uint8_t sp_palette = 0;
 	uint8_t sp_prio = 0;
@@ -133,23 +134,25 @@ static inline void ppu_render_pixel(struct nes_state *state) {
 	uint8_t a0 = !!(ppu->bg_shift_attrib_low & bit);
 	uint8_t a1 = !!(ppu->bg_shift_attrib_high & bit);
 
-	bg_pixel   = ((p1 << 1) | p0) & bg_mask;
-	bg_palette = ((a1 << 1) | a0) & bg_mask;
+	uint8_t bg_pixel   = ((p1 << 1) | p0) & bg_mask;
+	uint8_t bg_palette = ((a1 << 1) | a0) & bg_mask;
 
 	// Sprite
-	for(uint8_t i = 0; i < ppu->sprite_count; i++) {
-		if(ppu->sprite_positions[i]) continue;
+	if(sp_mask) {
+		for(uint8_t i = 0; i < ppu->sprite_count; i++) {
+			if(ppu->sprite_positions[i]) continue;
 
-		uint8_t lo = ppu->sprite_shift_lo[i];
-		uint8_t hi = ppu->sprite_shift_hi[i];
-		sp_pixel = (((hi & 0x80) >> 6) | ((lo & 0x80) >> 7)) & sp_mask;
+			uint8_t lo = ppu->sprite_shift_lo[i];
+			uint8_t hi = ppu->sprite_shift_hi[i];
+			sp_pixel = (((hi & 0x80) >> 6) | ((lo & 0x80) >> 7));
 
-		if(!sp_pixel) continue;
+			if(!sp_pixel) continue;
 
-		sp_palette = ppu->secondary_oam[i * 4 + 2] & 3;
-		sp_prio = ppu->sprite_priorities[i];
-		sp_zero = (ppu->sprite_indexes[i] == 0);
-		break;
+			sp_palette	= ppu->secondary_oam[i * 4 + 2] & 3;
+			sp_prio		= ppu->sprite_priorities[i];
+			sp_zero		= (ppu->sprite_indexes[i] == 0);
+			break;
+		}
 	}
 
 	// Final pixel composition
@@ -170,8 +173,8 @@ static inline void ppu_render_pixel(struct nes_state *state) {
 	state->pixels[y * 256 + x] = ppu->palette[palette_index];		// NOTE(peter): Add color_emphasis bits (expand palette to 8x).
 }
 
-__attribute__((hot, flatten))
-__attribute__((optimize("no-jump-tables")))
+
+__attribute__((hot, optimize("no-jump-tables")))
 static inline void ppu_tick(struct nes_state *state) {
 	struct ppu_state *restrict ppu = &state->ppu;
 
@@ -208,10 +211,10 @@ static inline void ppu_tick(struct nes_state *state) {
 
 						case 1 ... 255:		// fallthrough: this is 1->256
 							ppu_render_pixel(state);
-							__attribute__((fallthrough));
+							__attribute__((fallthrough));	// fallthrough: the code below has to run  1->256 + 321->336
+
+						case 321 ... 336: {	// Rendering and tile fetch;
 
-						case 321 ... 336: {	// fallthrough: the code below has to run  1->256 + 321->336
-							// Rendering and tile fetch goes here
 							if(ppu->reg_mask & 0x10) {
 								for(uint32_t i = 0; i < ppu->sprite_count; i++) {
 									if(ppu->sprite_positions[i] > 0) {