before ppu rearrangement and fix of scanline 261

7 files changed, 75 insertions, 36 deletions

diff --git a/build.sh b/build.sh
index 611ca19..8858da4 100755
--- a/build.sh
+++ b/build.sh
@@ -9,12 +9,13 @@ CFLAGS+="-mbmi2 "
 CFLAGS+="-mfunction-return=keep "
 CFLAGS+="-mindirect-branch=keep "
 CFLAGS+="-fwrapv -ffast-math -fno-trapping-math -fwhole-program -fvisibility=hidden "
-CFLAGS+="-fno-stack-protector -fno-PIE -no-pie -fno-strict-aliasing -ffunction-sections -fdata-sections "
+CFLAGS+="-fno-stack-protector -fno-PIE -no-pie -fno-strict-aliasing -fcf-protection=none -ffunction-sections -fdata-sections "
 # CFLAGS+="-fno-exceptions -fno-rtti -fno-use-cxa-atexit "
 CFLAGS+="-fno-non-call-exceptions -fno-unwind-tables -fno-asynchronous-unwind-tables "
 CFLAGS+="-Wall -Wextra "
 CFLAGS+="-Wno-unused-parameter -Wno-sign-compare -Wno-trigraphs -Wno-maybe-uninitialized "
 CFLAGS+="-Wno-unused-variable -Wno-unused-const-variable -Wno-unused-function -Wno-write-strings -Wno-missing-field-initializers "
+CFLAGS+="-U_FORTIFY_SOURCE -fno-pic "
 
 LDFLAGS="-Wl,--gc-sections -Wl,--as-needed "
 
diff --git a/cpu_opcodes.c b/cpu_opcodes.c
index b879015..7e174be 100644
--- a/cpu_opcodes.c
+++ b/cpu_opcodes.c
@@ -1,7 +1,7 @@
 
 
 // ADC
-
+__attribute__((always_inline, hot))
 static inline void adc(struct cpu_state * restrict cpu, uint8_t value) {
 #ifdef ENABLE_DECIMAL_MODE
 	if(cpu->d) {
@@ -407,7 +407,7 @@ static void opcode_brk(struct nes_state * restrict state) {
 
 
 // BRANCHES
-
+__attribute__((always_inline, hot))
 static inline int page_crossed(uint16_t a, uint16_t b) {
 	return (a & 0xff00) != (b & 0xff00);
 }
@@ -583,6 +583,7 @@ static void opcode_sei(struct nes_state * restrict state) {
 
 
 // CMP
+__attribute__((always_inline, hot))
 static inline void cmp(struct cpu_state * restrict cpu, uint8_t value) {
 	uint8_t result = cpu->a - value;
 	cpu->c = (cpu->a >= value);
@@ -695,7 +696,7 @@ static void opcode_cmp_absx(struct nes_state * restrict state) {
 
 
 // CPX
-
+__attribute__((always_inline, hot))
 static inline void cpx(struct cpu_state * restrict cpu, uint8_t value) {
 	uint8_t result = cpu->x - value;
 	cpu->c = (cpu->x >= value);
@@ -731,7 +732,7 @@ static void opcode_cpx_abs(struct nes_state * restrict state) {
 
 
 // CPY
-
+__attribute__((always_inline, hot))
 static inline void cpy(struct cpu_state * restrict cpu, uint8_t value) {
 	uint8_t result = cpu->y - value;
 	cpu->c = (cpu->y >= value);
@@ -1752,7 +1753,7 @@ static void opcode_ror_absx(struct nes_state * restrict state) {
 
 
 // SBC
-
+__attribute__((always_inline, hot))
 static inline void sbc(struct cpu_state * restrict cpu, uint8_t value) {
 #ifdef ENABLE_DECIMAL_MODE
 	if(cpu->d) {
@@ -2123,6 +2124,9 @@ static void opcode_sty_zpx(struct nes_state * restrict state) {
 
 __attribute__((noinline))
 static void init_opcode_lut(void) {
+	for(uint32_t i = 0; i < 256; ++i) {
+		opcode_lut[i] = opcode_nop;		// make sure erroneous opcodes just do "NOP", this might change in the future!
+	}
 	opcode_lut[0x00] = opcode_brk;
 	opcode_lut[0x01] = opcode_ora_indx;
 	opcode_lut[0x05] = opcode_ora_zp;
diff --git a/memory.c b/memory.c
index 54962a5..e934136 100644
--- a/memory.c
+++ b/memory.c
@@ -5,6 +5,7 @@ __attribute__((hot))
 static uint8_t memory_read(struct nes_state *restrict state, uint32_t offset) {
 	state->cycles++;
 	ppu_tick(state);
+	// apu_tick(state);
 
 	if(LIKELY(offset < 0x2000)) {
 		return state->ram[offset & 0x07ff];
@@ -18,6 +19,10 @@ static uint8_t memory_read(struct nes_state *restrict state, uint32_t offset) {
 		state->input_bit[index]++;
 		return value | 0x40;	// Bit 6 open bus high, bit 7 low
 
+	// } else if(offset == 4015) {
+	// 	static uint32_t apuread = 0;
+	// 	// printf("%.5d apu\n", apuread++);
+
 	} else if(LIKELY(offset >= 0x6000)) {
 		return state->mapper.prg_read(state, offset);
 
@@ -29,6 +34,7 @@ __attribute__((hot))
 static void memory_write(struct nes_state *restrict state, uint32_t offset, uint8_t value) {
 	state->cycles++;
 	ppu_tick(state);
+	// apu_tick(state);
 
 	if(LIKELY(offset < 0x2000)) {
 		state->ram[offset & 0x07ff] = value;
@@ -36,21 +42,34 @@ static void memory_write(struct nes_state *restrict state, uint32_t offset, uint
 	} else if(offset < 0x4000) {
 		ppu_write(state, offset, value);
 
-	} else if(offset == 0x4014) {
-		ppu_dma_4014(state, value);
-
-	} else if(offset == 0x4016) {
-		uint8_t prev = state->input_strobe;
-		state->input_strobe = value & 1;
-
-		if(prev == 1 && (value & 1) == 0) {
-			// Latch current inputs
-			state->input_latch[0] = state->input[0];
-			state->input_latch[1] = state->input[1];
-			state->input_bit[0] = 0;
-			state->input_bit[1] = 0;
+	} else if(offset < 0x4018) {
+		// APU + joypad + DMA
+		if(offset == 0x4014) {
+			ppu_dma_4014(state, value);
+
+		// } else if(offset == 0x4015) {	// DMC Enable (APU)
+		// 	state->apu.dmc_dma_enabled = (value & 0x10) ? 1 : 0;
+
+		} else if(offset == 0x4016) {
+			uint8_t prev = state->input_strobe;
+			state->input_strobe = value & 1;
+
+			if(prev == 1 && (value & 1) == 0) {
+				// Latch current inputs
+				state->input_latch[0] = state->input[0];
+				state->input_latch[1] = state->input[1];
+				state->input_bit[0] = 0;
+				state->input_bit[1] = 0;
+			}
+
+		// } else if(offset == 0x4017) {	// Frame Counter (APU)
+		// 	state->apu.mode = (value >> 7) & 1;
+		// 	state->apu.irq_inhibit = (value >> 6) & 1;
+		// 	state->apu.frame_cycle = 0;
+		// 	if(state->apu.irq_inhibit) {
+		// 		state->apu.irq_pending = 0;
+		// 	}
 		}
-
 	} else if(offset >= 0x6000) {
 		state->mapper.prg_write(state, offset, value);
 	}
@@ -73,6 +92,7 @@ __attribute__((hot))
 static uint8_t memory_read_dummy(struct nes_state *restrict state, uint32_t offset) {
 	state->cycles++;
 	ppu_tick(state);
+	// apu_tick(state);
 
 	if(UNLIKELY(offset >= 0x2000 && offset < 0x4000)) {
 		return ppu_read(state, offset);
diff --git a/mknes.c b/mknes.c
index 6ef4e02..a365ddb 100644
--- a/mknes.c
+++ b/mknes.c
@@ -90,7 +90,7 @@ static GLFWwindow *window;
 #include "render.c"
 
 static uint32_t frames;		// debug information
-// #include "smb_tas.h"	// REMOVE ME
+#include "smb_tas.h"	// REMOVE ME
 
 // NES core
 #include "mapper.h"
@@ -224,6 +224,19 @@ int main(int argc, char **argv) {
 	uint32_t hi = nstate->mapper.prg_read(nstate, 0xfffd);
 	nstate->cpu.pc = (hi << 8) | lo;
 
+#if 0
+	for(uint32_t i = 0; i < 0x5000; ++ i) {
+		while(!nstate->ppu.frame_ready) {
+			// PROFILE_NAMED("nes emulator");
+			cpu_tick(nstate);
+		}
+		nstate->ppu.frame_ready = 0;
+		frames++;
+	}
+	return 0;
+#endif
+
+
 	struct timer_handle *timer = timer_new(FRAME_INTERVAL_NS);
 	if(!timer) {
 		fprintf(stderr, "Failed to create timer\n");
@@ -263,16 +276,6 @@ int main(int argc, char **argv) {
 
 			timer_start(timer);
 
-#if 1
-			for(uint32_t i = 0; i < 0x5000; ++ i) {
-				while(!nstate->ppu.frame_ready) {
-					// PROFILE_NAMED("nes emulator");
-					cpu_tick(nstate);
-				}
-				nstate->ppu.frame_ready = 0;
-				frames++;
-			}
-#else
 			while(!glfwWindowShouldClose(window)) {
 				timer_wait(timer);
 				glfwPollEvents();
@@ -305,7 +308,6 @@ int main(int argc, char **argv) {
 				render_frame();
 				glfwSwapBuffers(window);
 			}
-#endif
 printf("total frames: %6.6d  total cycles: %ld\n", frames, nstate->cycles);
 			glfwDestroyWindow(window);
 		} else {
diff --git a/mknes.h b/mknes.h
index f8e4be7..fb1228b 100644
--- a/mknes.h
+++ b/mknes.h
@@ -52,7 +52,6 @@ struct ppu_state {
 
 	uint8_t sprite_zero_hit_possible;
 	uint8_t sprite_count;
-
 	uint8_t palette[32];
 
 	// NOTE(peter): CACHELINE 2 start here!
@@ -68,6 +67,15 @@ struct ppu_state {
 	uint8_t frame_ready;
 } __attribute__((packed, aligned(64)));
 
+struct apu_state {
+	uint32_t frame_cycle;
+	uint8_t mode;
+	uint8_t irq_inhibit;
+	uint8_t irq_pending;
+	uint8_t dmc_dma_enabled;
+	uint32_t dmc_sample_timer;
+};
+
 struct cpu_state {
 	uint32_t pc;	// Program Counter
 	uint8_t sp;		// Stack Pointer
@@ -104,6 +112,7 @@ struct nes_state {
 	uint8_t input_latch[2];		// Latched inputs after strobe
 	uint8_t input_bit[2];		// Current bit position being shifted out
 	uint8_t input_strobe;		// Control bit (0 or 1)
+	// struct apu_state apu;
 	struct ppu_state ppu;
 
 	struct mapper_entry mapper;
diff --git a/ppu.c b/ppu.c
index b34115c..cd549d1 100644
--- a/ppu.c
+++ b/ppu.c
@@ -397,8 +397,8 @@ static void ppu_tick(struct nes_state *state) {
 		}
 
 		if(UNLIKELY(scanline == 241) && dot == 1) {
-			// static int32_t tas_frame = 0;
-			// state->input[0] = tas_input[tas_frame++];
+			static int32_t tas_frame = 0;
+			state->input[0] = tas_input[tas_frame++];
 
 			ppu->reg_status |= 0x80;
 			if(ppu->reg_ctrl & 0x80) {
@@ -432,4 +432,3 @@ static void ppu_tick(struct nes_state *state) {
 		ppu->scanline = scanline;
 	}
 }
-
diff --git a/ppu_registers.c b/ppu_registers.c
index e2e7940..1f738a1 100644
--- a/ppu_registers.c
+++ b/ppu_registers.c
@@ -122,6 +122,7 @@ static inline void ppu_dma_4014(struct nes_state *state, uint8_t page) {
 	for(uint8_t i = 0; i < idle_cycles; i++) {
 		state->cycles++;
 		ppu_tick(state);
+		// apu_tick(state);
 	}
 
 	for(uint32_t i = 0; i < 256; i++) {
@@ -129,10 +130,13 @@ static inline void ppu_dma_4014(struct nes_state *state, uint8_t page) {
 
 		state->cycles++;
 		ppu_tick(state);
+		// apu_tick(state);
 		uint8_t value = memory_read_dma(state, addr);
 
 		state->cycles++;
 		ppu_tick(state);
+		// apu_tick(state);
+
 		ppu_write(state, 4, value);
 	}
 }