Rearranged the ppu_state to be even more cacheline aware, gained another 1% performance

1 files changed, 10 insertions, 9 deletions

diff --git a/mknes.h b/mknes.h
index b6739d3..944f762 100644
--- a/mknes.h
+++ b/mknes.h
@@ -56,7 +56,6 @@ struct ppu_state {
 	uint8_t open_bus;						// 29
 
 	uint8_t sprite_count;				// 30 - Number of sprites in secondary OAM (0-8)
-	uint8_t palette[32];					// 31
 
 	// NOTE(peter): CACHELINE 2
 	uint8_t secondary_oam[32] __attribute__((aligned(64)));
@@ -68,16 +67,18 @@ struct ppu_state {
 		uint8_t position;
 		uint8_t priority;
 		uint8_t palette;
-	} __attribute__((packed)) sprites[8];
+	} __attribute__((packed)) sprites[8];	// 32 -> +40
 
-	uint8_t input[2];						// 40 - Controller 1 & 2
-	uint8_t input_latch[2];				// 42 - Latched inputs after strobe
-	uint8_t input_bit[2];				// 44 - Current bit position being shifted out
-	uint8_t input_strobe;				// 46 - Control bit (0 or 1)
-	uint8_t frame_ready;					// 47
-	uint8_t sprite_zero_in_range;		// 48 - Boolean: is sprite 0 in range (will always be slot 0 if true)
+	uint8_t input_strobe;				// 76 - Control bit (0 or 1)
+	uint8_t input[2];						// 73 - Controller 1 & 2
+	uint8_t input_latch[2];				// 74 - Latched inputs after strobe
+	uint8_t input_bit[2];				// 75 - Current bit position being shifted out
+	uint8_t frame_ready;					// 77
+	uint8_t sprite_zero_in_range;		// 78 - Boolean: is sprite 0 in range (will always be slot 0 if true)
 
-	// NOTE(peter): CACHELINE 4
+	uint8_t palette[32];					// 79
+
+	// NOTE(peter): CACHELINE 5
 	uint8_t oam[256] __attribute__((aligned(64)));
 } __attribute__((packed));