Prepare profiling macros

1 files changed, 21 insertions, 10 deletions

diff --git a/src/gba/gba.h b/src/gba/gba.h
index b02d745..27a6a9a 100644
--- a/src/gba/gba.h
+++ b/src/gba/gba.h
@@ -98,18 +98,18 @@ typedef u16 Color;
 typedef Color Palette[16];
 
 // Inline function to calculate the 15 bit color value.
-#define RGB15(R,G,B) (u16)(((B) << 10) | ((G) << 5) | (R));
+#define RGB15(R,G,B) (u16)(((B) << 10) | ((G) << 5) | (R))
 
 // Some nice default colors.
-#define COLOR_BLACK  RGB15(0,  0, 0)
-#define COLOR_RED    RGB15(31, 0,10)
-#define COLOR_GREEN  RGB15(31, 0,10)
-#define COLOR_YELLOW RGB15(31, 0,10)
-#define COLOR_BLUE   RGB15(2, 17,31)
-#define COLOR_PURPLE RGB15(2, 17,31)
-#define COLOR_CYAN   RGB15(0, 27,30)
-#define COLOR_GREY   RGB15(16,17,19)
-#define COLOR_WHITE  RGB15(28,28,28)
+#define COLOR_BLACK  RGB15( 0,  0,  0)
+#define COLOR_RED    RGB15(31,  0, 10)
+#define COLOR_GREEN  RGB15( 0, 31, 18)
+#define COLOR_YELLOW RGB15(31, 31,  0)
+#define COLOR_BLUE   RGB15( 2, 17, 31)
+#define COLOR_PURPLE RGB15(15,  7, 31)
+#define COLOR_CYAN   RGB15( 0, 27, 30)
+#define COLOR_GREY   RGB15(16, 17, 19)
+#define COLOR_WHITE  RGB15(28, 28, 28)
 
 //
 // Tile memory access.
@@ -419,6 +419,8 @@ inline
 void
 dma_copy(void *dst, const void *src, u32 size, int channel) {
     dma_transfer_copy(dst, src, size / 4, channel, DMA_CHUNK_32 | DMA_ENABLE);
+    // Stall for 2 cycles in case we call this function more than once.
+    asm("nop"); asm("nop");
 }
 
 // Fill the dst location with the word set at src.
@@ -426,6 +428,8 @@ inline
 void
 dma_fill(void *dst, vu32 src, u32 size, int channel) {
     dma_transfer_fill(dst, src, size / 4, channel, DMA_CHUNK_32 | DMA_ENABLE);
+    // Stall for 2 cycles in case we call this function more than once.
+    asm("nop"); asm("nop");
 }
 
 //
@@ -661,6 +665,7 @@ wait_vsync(void) {
 #define LEN(ARR) (sizeof(ARR) / sizeof((ARR)[0]))
 
 // Fixed-point arithmetic for (i.P) numbers.
+#define FP_NUM(A,P)        ((A) << (P))
 #define FP_MUL(A,B,P)      (((A) * (B)) >> (P))
 #define FP_DIV(A,B,P)      (((A) << (P)) / (B))
 #define FP_LERP(Y0,Y1,X,P) ((Y0) + FP_MUL((X), ((Y1) - (Y0)), P))
@@ -683,4 +688,10 @@ memcpy32(u32 *dst, const u32 *src, u32 size) {
     }
 }
 
+//
+// Compiler hints.
+//
+
+#define UNROLL_LOOPS __attribute__((optimize("unroll-loops")))
+
 #endif // GBA_H