added mapper_tick() functionality, regressed 200fps, and optimized it back up to ~1940fps

1 files changed, 15 insertions, 14 deletions

diff --git a/ppu.c b/ppu.c
index 3096573..8280265 100644
--- a/ppu.c
+++ b/ppu.c
@@ -19,14 +19,12 @@ static uint8_t __attribute__((aligned(64))) ppu_bitreverse_lut[256] = {
 	0x0f, 0x8f, 0x4f, 0xcf, 0x2f, 0xaf, 0x6f, 0xef, 0x1f, 0x9f, 0x5f, 0xdf, 0x3f, 0xbf, 0x7f, 0xff
 };
 
-
 static void ppu_reset(struct nes_state *state) {
 	struct ppu_state *restrict ppu = &state->ppu;
 	memset(ppu, 0, sizeof(struct ppu_state));
 }
 
-
-// __attribute__((hot))
+__attribute__((hot))
 static inline void ppu_evaluate_sprites(struct nes_state *state) {
 	struct ppu_state *restrict ppu = &state->ppu;
 	uint8_t sprite_height = (ppu->reg_ctrl & 0x20) ? 16 : 8;
@@ -59,7 +57,7 @@ static inline void ppu_evaluate_sprites(struct nes_state *state) {
 	ppu->sprite_count = n;
 }
 
-// __attribute__((hot))
+__attribute__((hot))
 static inline void ppu_fetch_sprite_patterns(struct nes_state *state) {
 	struct ppu_state *restrict ppu = &state->ppu;
 	uint32_t addr;
@@ -105,7 +103,7 @@ static inline void ppu_fetch_sprite_patterns(struct nes_state *state) {
 	}
 }
 
-__attribute__((hot))
+__attribute__((hot, flatten))
 static inline void ppu_render_pixel(struct nes_state *state) {
 	struct ppu_state *restrict ppu = &state->ppu;
 
@@ -182,7 +180,7 @@ static inline void ppu_tick(struct nes_state *state) {
 
 	for(uint32_t ppu_loops = 0; ppu_loops < 3; ++ppu_loops) {
 
-		if(LIKELY(rendering)) {
+		if(rendering) {
 			switch(scanline) {
 				case 0 ... 239: {
 					switch(dot) {
@@ -358,12 +356,12 @@ static inline void ppu_tick(struct nes_state *state) {
 								} break;
 
 								case 0: {
-									ppu->bg_shift_pattern_low	= (ppu->bg_shift_pattern_low & 0xff00) | ppu->bg_next_tile_lsb;
-									ppu->bg_shift_pattern_high	= (ppu->bg_shift_pattern_high & 0xff00) | ppu->bg_next_tile_msb;
+									ppu->bg_shift_pattern_low = (ppu->bg_shift_pattern_low & 0xff00) | ppu->bg_next_tile_lsb;
+									ppu->bg_shift_pattern_high = (ppu->bg_shift_pattern_high & 0xff00) | ppu->bg_next_tile_msb;
 
 									uint8_t a = ppu->bg_next_tile_attrib;
-									ppu->bg_shift_attrib_low	= (ppu->bg_shift_attrib_low & 0xff00) | ((a & 1) ? 0xff : 0x00);
-									ppu->bg_shift_attrib_high	= (ppu->bg_shift_attrib_high & 0xff00) | ((a & 2) ? 0xff : 0x00);
+									ppu->bg_shift_attrib_low = (ppu->bg_shift_attrib_low & 0xff00) | ((a & 1) ? 0xff : 0x00);
+									ppu->bg_shift_attrib_high = (ppu->bg_shift_attrib_high & 0xff00) | ((a & 2) ? 0xff : 0x00);
 
 									if((ppu->vram_addr & 0x001f) == 31) {
 										ppu->vram_addr &= ~0x001f;
@@ -394,7 +392,7 @@ static inline void ppu_tick(struct nes_state *state) {
 			}
 		}
 
-		if(UNLIKELY(scanline == 241) && dot == 1) {
+		if(scanline == 241 && dot == 1) {
 			// static int32_t tas_frame = 0;
 			// state->input[0] = tas_input[tas_frame++];
 
@@ -402,7 +400,7 @@ static inline void ppu_tick(struct nes_state *state) {
 			state->cpu.nmi_pending = (ppu->reg_ctrl & 0x80);	// NOTE(peter): Set NMI if enabled.
 		}
 
-		if(UNLIKELY(scanline == 261) && dot == 1) {
+		if(scanline == 261 && dot == 1) {
 			ppu->reg_status &= ~0x80;
 			ppu->reg_status &= ~0x40;
 		}
@@ -412,7 +410,7 @@ static inline void ppu_tick(struct nes_state *state) {
 			dot = 0;
 			scanline++;
 
-			if(UNLIKELY(scanline == 261 && !ppu->even_frame && (ppu->reg_mask & 0x18))) {
+			if(scanline == 261 && !ppu->even_frame && (ppu->reg_mask & 0x18)) {
 				dot = 1;
 			}
 
@@ -426,6 +424,9 @@ static inline void ppu_tick(struct nes_state *state) {
 
 		ppu->dot = dot;
 		ppu->scanline = scanline;
-		if(UNLIKELY(state->mapper_function.tick)) state->mapper_function.tick(state);
+
+		if(state->mapper_function.tick) {
+			state->mapper_function.tick(state);
+		}
 	}
 }