optimized away an if(dot == 256) in the hot path ~5% improvement

3 files changed, 59 insertions, 44 deletions

diff --git a/mknes.c b/mknes.c
index 7417bed..51e737e 100644
--- a/mknes.c
+++ b/mknes.c
@@ -227,8 +227,8 @@ int main(int argc, char **argv) {
 
 			// while(!glfwWindowShouldClose(window)) {
 			for(uint32_t i = 0; i < 0x5000; ++ i) {
-				// timer_wait(timer);
-				// glfwPollEvents();
+			// 	timer_wait(timer);
+			// 	glfwPollEvents();
 
 // //
 				while(!nstate->ppu.frame_ready) {
diff --git a/mknes.h b/mknes.h
index a296fe4..10ae358 100644
--- a/mknes.h
+++ b/mknes.h
@@ -62,7 +62,7 @@ struct ppu_state {
 
 	uint8_t sprite_zero_hit_possible;
 	uint8_t sprite_count;
-} __attribute__((packed, aligned(64)));
+} __attribute__((aligned(64)));
 
 struct cpu_state {
 	uint32_t pc;	// Program Counter
@@ -80,7 +80,7 @@ struct cpu_state {
 	uint8_t c;		// Carry Flag
 // --
 	uint8_t die;	// KIL instruction found!
-} __attribute__((packed, aligned(64)));
+} __attribute__((aligned(64)));
 
 
 struct ines_state {
diff --git a/ppu.c b/ppu.c
index da360e0..e9a6894 100644
--- a/ppu.c
+++ b/ppu.c
@@ -280,29 +280,46 @@ static void ppu_tick(struct nes_state *state) {
 			switch(scanline) {
 				case 0 ... 239: {
 					switch(dot) {
-						case 1:
-						    __attribute__((fallthrough));
-
-						case 2 ... 256:		// fallthrough: this is 1->256
-							ppu_render_pixel(state);
-
-							if(UNLIKELY(dot == 256)) {
-								if((ppu->vram_addr & 0x7000) != 0x7000) {
-									ppu->vram_addr += 0x1000;
+						case 256: {
+
+							if((ppu->vram_addr & 0x7000) != 0x7000) {
+								ppu->vram_addr += 0x1000;
+							} else {
+								ppu->vram_addr &= ~0x7000;
+								uint32_t y = (ppu->vram_addr & 0x03e0) >> 5;
+								if(y == 29) {
+									y = 0;
+									ppu->vram_addr ^= 0x0800;
+								} else if(y == 31) {
+									y = 0;
 								} else {
-									ppu->vram_addr &= ~0x7000;
-									uint32_t y = (ppu->vram_addr & 0x03e0) >> 5;
-									if(y == 29) {
-										y = 0;
-										ppu->vram_addr ^= 0x0800;
-									} else if(y == 31) {
-										y = 0;
-									} else {
-										y++;
-									}
-									ppu->vram_addr = (ppu->vram_addr & ~0x03e0) | (y << 5);
+									y++;
 								}
+								ppu->vram_addr = (ppu->vram_addr & ~0x03e0) | (y << 5);
 							}
+							goto render_and_shift;
+						} break;
+
+						case 1 ... 255:		// fallthrough: this is 1->256
+render_and_shift:		ppu_render_pixel(state);
+
+							// if(UNLIKELY(dot == 256)) {
+							// 	if((ppu->vram_addr & 0x7000) != 0x7000) {
+							// 		ppu->vram_addr += 0x1000;
+							// 	} else {
+							// 		ppu->vram_addr &= ~0x7000;
+							// 		uint32_t y = (ppu->vram_addr & 0x03e0) >> 5;
+							// 		if(y == 29) {
+							// 			y = 0;
+							// 			ppu->vram_addr ^= 0x0800;
+							// 		} else if(y == 31) {
+							// 			y = 0;
+							// 		} else {
+							// 			y++;
+							// 		}
+							// 		ppu->vram_addr = (ppu->vram_addr & ~0x03e0) | (y << 5);
+							// 	}
+							// }
 							__attribute__((fallthrough));
 
 						case 321 ... 336: {	// fallthrough: the code below has to run  1->256 + 321->336
@@ -386,30 +403,28 @@ static void ppu_tick(struct nes_state *state) {
 
 				case 261: {
 					switch(dot) {
-						case 1:
-							// Clear vblank, sprite 0 hit, etc.
-							__attribute__((fallthrough));
-
-						case 2 ... 256:
-							if(UNLIKELY(dot == 256)) {
-								if((ppu->vram_addr & 0x7000) != 0x7000) {
-									ppu->vram_addr += 0x1000;
+						case 256: {
+
+							if((ppu->vram_addr & 0x7000) != 0x7000) {
+								ppu->vram_addr += 0x1000;
+							} else {
+								ppu->vram_addr &= ~0x7000;
+								uint32_t y = (ppu->vram_addr & 0x03e0) >> 5;
+								if(y == 29) {
+									y = 0;
+									ppu->vram_addr ^= 0x0800;
+								} else if(y == 31) {
+									y = 0;
 								} else {
-									ppu->vram_addr &= ~0x7000;
-									uint32_t y = (ppu->vram_addr & 0x03e0) >> 5;
-									if(y == 29) {
-										y = 0;
-										ppu->vram_addr ^= 0x0800;
-									} else if(y == 31) {
-										y = 0;
-									} else {
-										y++;
-									}
-									ppu->vram_addr = (ppu->vram_addr & ~0x03e0) | (y << 5);
+									y++;
 								}
+								ppu->vram_addr = (ppu->vram_addr & ~0x03e0) | (y << 5);
 							}
-
+						}
 							__attribute__((fallthrough));
+
+						case 1 ... 255:		__attribute__((fallthrough));
+
 						case 321 ... 336: {		// Rendering and tile fetch
 							if(ppu->reg_mask & 0x10) {
 								for(uint32_t i = 0; i < ppu->sprite_count; i++) {