__attribute__((hot))
static inline uint8_t memory_read(struct nes_state *state, uint32_t offset) {
	state->cpu.cycles++;
	ppu_tick(state);
	apu_tick(state);

	if(offset <= 0x1fff) {
		return state->ram[offset & 0x07ff];
	}

	if(offset >= 0x8000) {
		return state->mapper_function.prg_rom_read(state, offset);
	}

	if(offset >= 0x2000 && offset <= 0x3fff) {
		return ppu_read(state, offset);
	}

	if(offset == 0x4015) {
		return apu_read4015(state);
	}

	if(offset == 0x4016 || offset == 0x4017) {
		uint32_t index = offset & 1;
		uint8_t value = (state->ppu.input_latch[index] >> state->ppu.input_bit[index]) & 1;
		state->ppu.input_bit[index]++;
		return value | 0x40; // Bit 6 open bus high, bit 7 low
	}

	if(offset >= 0x6000 && offset <= 0x7fff) {
		return state->mapper_function.prg_ram_read(state, offset);
	}

	return 0;
}


__attribute__((hot))
static inline uint8_t memory_read_dummy(struct nes_state *state, uint32_t offset) {
	state->cpu.cycles++;
	ppu_tick(state);
	apu_tick(state);

	if(offset >= 0x2000 && offset < 0x4000) {
		return ppu_read(state, offset);
	}
	return 0;
}


__attribute__((hot, optimize("no-jump-tables")))
static inline void memory_write(struct nes_state *state, uint32_t offset, uint8_t value) {
	state->cpu.cycles++;
	ppu_tick(state);
	apu_tick(state);

	if(offset <= 0x1fff) {
		state->ram[offset & 0x07ff] = value;
	}

	if(offset >= 0x2000 && offset <= 0x3fff) {
		ppu_write(state, offset, value);
	}

	if(offset >= 0x4000 && offset <= 0x4017) {
		if(offset == 0x4014) {
			ppu_dma_4014(state, value);
		}

		if(offset == 0x4016) {
			uint8_t s = value & 1;
			uint8_t prev = state->ppu.input_strobe;
			state->ppu.input_strobe = s;

			if(prev == 1 && s == 0) {
					state->ppu.input_latch[0] = state->ppu.input[0];
					state->ppu.input_latch[1] = state->ppu.input[1];
					state->ppu.input_bit[0] = 0;
					state->ppu.input_bit[1] = 0;
			}
		}

		if(offset != 0x4014 && offset != 0x4016) {
			apu_write(state, offset, value);
		}
	}

	if(offset >= 0x6000 && offset <= 0x7fff) {
		state->mapper_function.prg_ram_write(state, offset, value);
	}

	if(offset >= 0x8000) {
		state->mapper_function.prg_rom_write(state, offset, value);
	}

}

__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];
}