reached 2944fps again

7 files changed, 18 insertions, 27 deletions

diff --git a/.gitignore b/.gitignore
index 989490d..cd96a51 100644
--- a/.gitignore
+++ b/.gitignore
@@ -20,4 +20,4 @@ record.txt
 _Bench.sh
 toolchain
 accuracycoin.asm
-
+CLAUDE.md
diff --git a/Bench.sh b/Bench.sh
index 768d5b6..f91495a 100755
--- a/Bench.sh
+++ b/Bench.sh
@@ -2,7 +2,7 @@
 
 ./build.sh clean
 ./build.sh profile
-./mknes -n 1 &> /dev/null
+./mknes -n 1 -f 8192 &> /dev/null
 ./build.sh profile_release
 
 # Run full benchmark
diff --git a/mknes.c b/mknes.c
index e3e9ffd..b82fc55 100644
--- a/mknes.c
+++ b/mknes.c
@@ -271,7 +271,7 @@ int main(int argc, char **argv) {
 	// 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/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");
@@ -293,7 +293,7 @@ int main(int argc, char **argv) {
 	// ines2_load(nstate, "data/0000/Xevious - The Avenger (USA).zip");
 	// ines2_load(nstate, "data/tv.nes");
 
-	// ines2_load(nstate, "data/Life Force (USA).nes");	// 2002
+	ines2_load(nstate, "data/Life Force (USA).nes");	// 2002
 
 	// ines2_load(nstate, "data/0003/Flipull - An Exciting Cube Game (Japan) (En).zip");
 	// ines2_load(nstate, "data/0003/Friday the 13th (USA).zip");
diff --git a/mknes_apu.c b/mknes_apu.c
index 027b1e5..f1c27a8 100644
--- a/mknes_apu.c
+++ b/mknes_apu.c
@@ -1,8 +1,7 @@
 
 
-static inline void ppu_tick(struct nes_state *state);
+static void ppu_tick(struct nes_state *state);
 static uint8_t memory_read(struct nes_state *state, uint32_t offset);
-static uint8_t memory_read_dma(struct nes_state *state, uint32_t offset);
 
 // DMC frequency table (NTSC)
 static const uint16_t dmc_rate_table[16] = {
diff --git a/mknes_memory.c b/mknes_memory.c
index 49546b2..57f527e 100644
--- a/mknes_memory.c
+++ b/mknes_memory.c
@@ -1,11 +1,12 @@
 
 __attribute__((hot))
 static inline uint8_t memory_read(struct nes_state *state, uint32_t offset) {
+	if(offset >= 0x10000) __builtin_unreachable();
+
 	state->cpu.cycles++;
 	apu_tick(state);
 	ppu_tick(state);
 
-	if(offset >= 0x10000) __builtin_unreachable();
 
 	if(offset <= 0x1fff) {
 		return state->ram[offset & 0x07ff];
@@ -51,13 +52,14 @@ static inline uint8_t memory_read_dummy(struct nes_state *state, uint32_t offset
 	return result;
 }
 
-__attribute__((hot, optimize("no-jump-tables")))
+__attribute__((hot))
 static inline void memory_write(struct nes_state *state, uint32_t offset, uint8_t value) {
+	if(offset >= 0x10000) __builtin_unreachable();
+
 	state->cpu.cycles++;
 	apu_tick(state);
 	ppu_tick(state);
 
-	if(offset >= 0x10000) __builtin_unreachable();
 
 	if(offset <= 0x1fff) {
 		state->ram[offset & 0x07ff] = value;
@@ -109,9 +111,3 @@ static inline void memory_write(struct nes_state *state, uint32_t offset, uint8_
 		return;
 	}
 }
-
-__attribute__((hot, flatten))
-static inline uint8_t memory_read_dma(struct nes_state *state, uint32_t offset) {
-	// NOTE(peter): DO NOT tick CPU/PPU/APU — caller handles timing
-	return state->ram[offset & 0x07ff];
-}
diff --git a/mknes_ppu.c b/mknes_ppu.c
index 7c2b4ac..6e8e3ae 100644
--- a/mknes_ppu.c
+++ b/mknes_ppu.c
@@ -164,21 +164,22 @@ no_sprite:
 	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);
-
-	// NOTE(peter): It's actually faster to preset case 3 version of palette_index than to start from zero
-	uint8_t palette_index = (sp_prio) ? bg_index : 0x10 | sp_index;
+	uint8_t palette_index = 0;
 
 	switch(selector) {
-		case 0: { palette_index = 0;												} 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) ? PPU_STATUS_SPRITE_ZERO_HIT : 0;	} break;		// NOTE(peter): Sprite zero hit!
+		case 3: {
+			palette_index = (sp_prio) ? bg_index : 0x10 | sp_index;
+			ppu->reg_status |= (sp_zero && x < 255) ? PPU_STATUS_SPRITE_ZERO_HIT : 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).
 }
 
-__attribute__((hot, optimize("no-jump-tables", "unroll-loops")))
+__attribute__((noinline, hot, optimize("no-jump-tables", "unroll-loops")))
 static void ppu_tick(struct nes_state *state) {
 	struct ppu_state *restrict ppu = &state->ppu;
 
@@ -387,9 +388,7 @@ stupid2:			if(reg_mask & PPU_MASK_SHOW_SPRITES) {
 				if(dot >= 280 && dot <= 304) {
 					ppu->vram_addr = (ppu->vram_addr & ~0x7be0) | (ppu->temp_addr & 0x7be0);
 				}
-
 			}
-
 		}
 
 		if(dot == 1) {
diff --git a/mknes_ppu_registers.c b/mknes_ppu_registers.c
index b66f7e5..536bf2a 100644
--- a/mknes_ppu_registers.c
+++ b/mknes_ppu_registers.c
@@ -126,9 +126,6 @@ static inline uint8_t ppu_read(struct nes_state *state, uint32_t offset) {
 	return result;
 }
 
-
-static uint8_t memory_read_dma(struct nes_state *state, uint32_t offset);
-
 __attribute__((hot))
 static inline void ppu_dma_4014(struct nes_state *state, uint8_t page) {
 	uint32_t base = page << 8;
@@ -147,7 +144,7 @@ static inline void ppu_dma_4014(struct nes_state *state, uint8_t page) {
 		state->cpu.cycles++;
 		ppu_tick(state);
 		apu_tick(state);
-		uint8_t value = memory_read_dma(state, addr);
+		uint8_t value = state->ram[addr & 0x07ff];	// NOTE(peter): was; memory_read_dma(state, addr);
 
 		state->cpu.cycles++;
 		ppu_tick(state);