static uint8_t __attribute__((aligned(64))) ppu_bitreverse_lut[256] = {
	0x00, 0x80, 0x40, 0xc0, 0x20, 0xa0, 0x60, 0xe0, 0x10, 0x90, 0x50, 0xd0, 0x30, 0xb0, 0x70, 0xf0,
	0x08, 0x88, 0x48, 0xc8, 0x28, 0xa8, 0x68, 0xe8, 0x18, 0x98, 0x58, 0xd8, 0x38, 0xb8, 0x78, 0xf8,
	0x04, 0x84, 0x44, 0xc4, 0x24, 0xa4, 0x64, 0xe4, 0x14, 0x94, 0x54, 0xd4, 0x34, 0xb4, 0x74, 0xf4,
	0x0c, 0x8c, 0x4c, 0xcc, 0x2c, 0xac, 0x6c, 0xec, 0x1c, 0x9c, 0x5c, 0xdc, 0x3c, 0xbc, 0x7c, 0xfc,
	0x02, 0x82, 0x42, 0xc2, 0x22, 0xa2, 0x62, 0xe2, 0x12, 0x92, 0x52, 0xd2, 0x32, 0xb2, 0x72, 0xf2,
	0x0a, 0x8a, 0x4a, 0xca, 0x2a, 0xaa, 0x6a, 0xea, 0x1a, 0x9a, 0x5a, 0xda, 0x3a, 0xba, 0x7a, 0xfa,
	0x06, 0x86, 0x46, 0xc6, 0x26, 0xa6, 0x66, 0xe6, 0x16, 0x96, 0x56, 0xd6, 0x36, 0xb6, 0x76, 0xf6,
	0x0e, 0x8e, 0x4e, 0xce, 0x2e, 0xae, 0x6e, 0xee, 0x1e, 0x9e, 0x5e, 0xde, 0x3e, 0xbe, 0x7e, 0xfe,
	0x01, 0x81, 0x41, 0xc1, 0x21, 0xa1, 0x61, 0xe1, 0x11, 0x91, 0x51, 0xd1, 0x31, 0xb1, 0x71, 0xf1,
	0x09, 0x89, 0x49, 0xc9, 0x29, 0xa9, 0x69, 0xe9, 0x19, 0x99, 0x59, 0xd9, 0x39, 0xb9, 0x79, 0xf9,
	0x05, 0x85, 0x45, 0xc5, 0x25, 0xa5, 0x65, 0xe5, 0x15, 0x95, 0x55, 0xd5, 0x35, 0xb5, 0x75, 0xf5,
	0x0d, 0x8d, 0x4d, 0xcd, 0x2d, 0xad, 0x6d, 0xed, 0x1d, 0x9d, 0x5d, 0xdd, 0x3d, 0xbd, 0x7d, 0xfd,
	0x03, 0x83, 0x43, 0xc3, 0x23, 0xa3, 0x63, 0xe3, 0x13, 0x93, 0x53, 0xd3, 0x33, 0xb3, 0x73, 0xf3,
	0x0b, 0x8b, 0x4b, 0xcb, 0x2b, 0xab, 0x6b, 0xeb, 0x1b, 0x9b, 0x5b, 0xdb, 0x3b, 0xbb, 0x7b, 0xfb,
	0x07, 0x87, 0x47, 0xc7, 0x27, 0xa7, 0x67, 0xe7, 0x17, 0x97, 0x57, 0xd7, 0x37, 0xb7, 0x77, 0xf7,
	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, flatten, optimize("unroll-loops")))
static inline void ppu_evaluate_sprites(struct nes_state *state, uint32_t scanline) {
	struct ppu_state *restrict ppu = &state->ppu;
	uint8_t sprite_height = (ppu->reg_ctrl & PPU_CTRL_SPRITE_HEIGHT) ? 16 : 8;
	uint8_t n = 0;
	uint8_t sprite_zero_found = 0;

	uint8_t * restrict src = ppu->oam;
	uint8_t * restrict dst = ppu->secondary_oam;

	for(uint8_t i = 0; i < 64; i++, src += 4) {
		uint32_t row = scanline - src[0];
		if(row < sprite_height) {
			if(n < 8) {
				dst[0] = src[0];
				dst[1] = src[1];
				dst[2] = src[2];
				dst[3] = src[3];
				dst += 4;
				sprite_zero_found |= (i == 0);
				n++;
			} else {
				ppu->reg_status |= PPU_STATUS_SPRITE_OVERFLOW;
				break;
			}
		}
	}
	ppu->sprite_zero_in_range = sprite_zero_found;
	ppu->sprite_count = n;
}

__attribute__((hot))
static inline void ppu_fetch_sprite_patterns(struct nes_state * restrict state, uint32_t scanline) {
	struct ppu_state *restrict ppu = &state->ppu;

	uint8_t * restrict sec_oam = ppu->secondary_oam;
	uint8_t ctrl = ppu->reg_ctrl;
	uint8_t sprite_height = (ctrl & PPU_CTRL_SPRITE_HEIGHT) ? 16 : 8;

	uint32_t sprite_pattern_table_base = (ctrl & PPU_CTRL_SPRITE_TILE) << 9;

	for(uint8_t i = 0; i < ppu->sprite_count; i++, sec_oam += 4) {
		uint8_t y = sec_oam[0];
		uint8_t tile = sec_oam[1];
		uint8_t attr = sec_oam[2];
		uint8_t x = sec_oam[3];

		uint32_t row = scanline - y;
		row = (attr & SPRITE_ATTR_FLIP_VERTICAL) ? sprite_height - 1 - row : row;

		uint32_t bank;
		uint32_t addr;
		if(sprite_height == 16) {
			// For 8x16 sprites:
			// - Bank comes from tile bit 0 (bits 1-7 are the tile index)
			// - Row 0-7 uses base tile, row 8-15 uses base tile + 1
			// - Row offset wraps to 0-7 within each 8-pixel half
			//
			// Original logic:
			//   bank = (tile & 1) << 12;
			//   tile &= 0xfe;
			//   if(row >= 8) { tile++; row -= 8; }
			//   addr = bank + tile * 16 + row;
			addr = ((tile & 1) << 12) + ((tile & 0xfe) + (row >> 3)) * 16 + (row & 7);

		} else {
			addr = sprite_pattern_table_base + tile * 16 + row;

		}

		uint8_t val_lo = state->mapper_function.chr_read(state, addr);
		uint8_t val_hi = state->mapper_function.chr_read(state, addr + 8);

		uint8_t rev = -(!!(attr & SPRITE_ATTR_FLIP_HORIZONTAL));
		uint8_t lsb = (rev & ppu_bitreverse_lut[val_lo]) | (~rev & val_lo);
		uint8_t msb = (rev & ppu_bitreverse_lut[val_hi]) | (~rev & val_hi);

		ppu->sprites[i].shift_lo = lsb;
		ppu->sprites[i].shift_hi = msb;
		ppu->sprites[i].position = x;
		ppu->sprites[i].priority = attr & SPRITE_ATTR_PRIORITY;
		ppu->sprites[i].palette = attr & SPRITE_ATTR_PALETTE_MASK;
	}
}

__attribute__((always_inline, hot, optimize("no-jump-tables", "no-unroll-loops")))
static inline void ppu_render_pixel(struct nes_state * restrict state, uint32_t x, uint32_t y, uint8_t mask_reg) {
	struct ppu_state *restrict ppu = &state->ppu;

	uint16_t bit = 0x8000 >> ppu->fine_x;

	uint8_t sp_pixel = 0;
	uint8_t sp_palette = 0;
	uint8_t sp_prio = 0;
	uint8_t sp_zero = 0;

	// uint8_t mask_reg = ppu->reg_mask;  // Single load
	uint8_t show_bg = mask_reg & PPU_MASK_SHOW_BG;
	uint8_t show_sprites = mask_reg & PPU_MASK_SHOW_SPRITES;
	uint8_t left_bg = mask_reg & 0x02;
	uint8_t left_sp = mask_reg & 0x04;

	uint8_t bg_mask = (show_bg && (left_bg || x & ~7)) ? 0xff : 0x00;
	uint8_t sp_mask = (show_sprites && (left_sp || x & ~7));

	// Background
	uint8_t p0 = !!(ppu->bg_shift_pattern_low & bit);
	uint8_t p1 = !!(ppu->bg_shift_pattern_high & bit);
	uint8_t a0 = !!(ppu->bg_shift_attrib_low & bit);
	uint8_t a1 = !!(ppu->bg_shift_attrib_high & bit);

	uint8_t bg_pixel   = ((p1 << 1) | p0) & bg_mask;
	uint8_t bg_palette = ((a1 << 1) | a0) & bg_mask;

	// Sprites - evaluate in forward order (0 has highest priority)
	if(sp_mask) {
		uint8_t found_sprite = 0xff;
		for(uint8_t i = 0; i < ppu->sprite_count; i++) {
			if(!ppu->sprites[i].position) {
				sp_pixel = (((ppu->sprites[i].shift_hi & 0x80) >> 6) | ((ppu->sprites[i].shift_lo & 0x80) >> 7));
				if(sp_pixel) {
					found_sprite = i;
					goto sprite_found;
				}
			}
		}
		goto no_sprite;

sprite_found:
		sp_prio = ppu->sprites[found_sprite].priority;
		sp_palette = ppu->sprites[found_sprite].palette;
		sp_zero = ppu->sprite_zero_in_range & !(found_sprite);
	}
no_sprite:

	// Final pixel composition
	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);
	uint8_t palette_index = 0;

	if(selector == 1) {
		palette_index = 0x10 | sp_index;
	} else if(selector == 2) {
		palette_index = bg_index;
	} else if(selector == 3) {
		palette_index = (sp_prio) ? bg_index : (0x10 | sp_index);
		if (sp_zero && x <= 254) ppu->reg_status |= PPU_STATUS_SPRITE_ZERO_HIT;
	}

	state->pixels[y * 256 + x] = ppu->palette[palette_index];		// NOTE(peter): Add color_emphasis bits (expand palette to 8x).
}


__attribute__((noinline, hot, optimize("no-jump-tables", "unroll-loops")))
static void ppu_tick(struct nes_state *state) {
	struct ppu_state *restrict ppu = &state->ppu;

	uint32_t dot = ppu->dot;
	uint32_t scanline = ppu->scanline;
	uint8_t reg_mask = ppu->reg_mask;
	uint8_t rendering = (reg_mask & (PPU_MASK_SHOW_SPRITES | PPU_MASK_SHOW_BG));

	for(uint8_t ppu_loops = 0; ppu_loops < 3; ++ppu_loops) {

		if(rendering) {
			if(scanline <= 239) {
				if(dot >= 1 && dot < 256) {	// NOTE(peter): dot 256 will be done in the next else in the chain
					ppu_render_pixel(state, dot - 1, scanline, reg_mask);
					goto shift_and_fetch;

				} else if(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);
					}
					ppu_render_pixel(state, dot - 1, scanline, reg_mask);
					goto shift_and_fetch;

				} else if(dot == 257) {
					ppu->vram_addr = (ppu->vram_addr & ~0x041f) | (ppu->temp_addr & 0x041f);
					ppu_evaluate_sprites(state, scanline);

				} else if(dot >= 321 && dot <= 336) {
					goto shift_and_fetch;

				} else if(dot == 340) {
					ppu_fetch_sprite_patterns(state, scanline);
				}
			}

			if(scanline == 261) {

				if(dot >= 1 && dot < 256) {	// NOTE(peter): dot 256 will be done in the next else in the chain
					goto shift_and_fetch;

				} else if(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);
					}
					goto shift_and_fetch;

				} else if(dot == 257) {
					ppu->vram_addr = (ppu->vram_addr & ~0x041f) | (ppu->temp_addr & 0x041f);

				} else if(dot >= 280 && dot <= 304) {
					ppu->vram_addr = (ppu->vram_addr & ~0x7be0) | (ppu->temp_addr & 0x7be0);

				} else if(dot >= 321 && dot <= 336) {
shift_and_fetch:
					if(reg_mask & PPU_MASK_SHOW_SPRITES) {
						for(uint32_t i = 0; i < ppu->sprite_count; i++) {
							if(ppu->sprites[i].position > 0) {
								ppu->sprites[i].position--;
							} else {
								ppu->sprites[i].shift_lo <<= 1;
								ppu->sprites[i].shift_hi <<= 1;
							}
						}
					}

					ppu->bg_shift_pattern_low <<= 1;
					ppu->bg_shift_pattern_high <<= 1;
					ppu->bg_shift_attrib_low <<= 1;
					ppu->bg_shift_attrib_high <<= 1;

					switch(dot % 8) {
						case 1: {
							uint32_t nt_addr = 0x2000 | (ppu->vram_addr & 0x0fff);
							ppu->bg_next_tile_id = state->mapper_function.ciram_read(state, nt_addr);
						} break;

						case 3: {
							uint32_t attr_addr = 0x23c0 | (ppu->vram_addr & 0x0c00) | ((ppu->vram_addr >> 4) & 0x38) | ((ppu->vram_addr >> 2) & 0x07);
							uint8_t attr = state->mapper_function.ciram_read(state, attr_addr & 0x0fff);
							uint8_t shift = ((ppu->vram_addr >> 4) & 4) | (ppu->vram_addr & 2);
							ppu->bg_next_tile_attrib = (attr >> shift) & 3;
						} break;

						case 5: {
							uint32_t base = (ppu->reg_ctrl & PPU_CTRL_BG_TILE_SELECT) << 8;
							uint32_t tile = ppu->bg_next_tile_id;
							uint32_t fine_y = (ppu->vram_addr >> 12) & 7;
							uint32_t addr_lsb = (base + tile * 16 + fine_y) & 0x1fff;
							ppu->bg_next_tile_lsb = state->mapper_function.chr_read(state, addr_lsb);
						} break;

						case 7: {
							uint32_t base = (ppu->reg_ctrl & PPU_CTRL_BG_TILE_SELECT) << 8;
							uint32_t tile = ppu->bg_next_tile_id;
							uint32_t fine_y = (ppu->vram_addr >> 12) & 7;
							uint32_t addr_msb = (base + tile * 16 + fine_y + 8) & 0x1fff;
							ppu->bg_next_tile_msb =  state->mapper_function.chr_read(state, addr_msb);
						} break;

						case 0: {
							ppu->bg_shift_pattern_low = (ppu->bg_shift_pattern_low) | ppu->bg_next_tile_lsb;
							ppu->bg_shift_pattern_high = (ppu->bg_shift_pattern_high) | ppu->bg_next_tile_msb;

							uint8_t a = ppu->bg_next_tile_attrib;
							ppu->bg_shift_attrib_low = (ppu->bg_shift_attrib_low) | ((a & 1) ? 0xff : 0x00);
							ppu->bg_shift_attrib_high = (ppu->bg_shift_attrib_high) | ((a & 2) ? 0xff : 0x00);

							if((ppu->vram_addr & 0x001f) == 31) {
								ppu->vram_addr &= ~0x001f;
								ppu->vram_addr ^= 0x0400;
							} else {
								ppu->vram_addr++;
							}
						} break;
					}
				}
			}
		}

		if(dot == 1) {
			switch(scanline) {
				case 241: {
					ppu->reg_status |= PPU_STATUS_VBLANK;
					state->cpu.nmi_pending = (ppu->reg_ctrl & PPU_CTRL_NMI);
				} break;

				case 261: {
					ppu->reg_status &= ~PPU_STATUS_VBLANK;
					ppu->reg_status &= ~PPU_STATUS_SPRITE_ZERO_HIT;
					ppu->reg_status &= ~PPU_STATUS_SPRITE_OVERFLOW;
				} break;
			}
		}

		dot++;
		if(dot > 340) {
			dot = 0;
			scanline++;

			if(scanline == 261 && !ppu->even_frame && rendering) {
				dot = 1;
			}

			if(scanline > 261) {
				scanline = 0;
				ppu->frame_ready = 1;
				ppu->even_frame = !ppu->even_frame;
			}

		}

		if(state->mapper_function.tick) {
			state->mapper_function.tick(state);	// TODO(peter): This signature has to be changed to supply  dot and scanline!
		}
	}
	ppu->dot = dot;
	ppu->scanline = scanline;
}