Improvements to the `RenderLine()` function (#920)

* Improvements to the `RenderLine()` function
- Simplify by using indices instead of incrementing pointers
- Improve performance in the case where mask != -1 by only processing the bits that are set

Source/render.cpp

@@ -133,54 +133,66 @@ static DWORD LeftFoliageMask[TILE_HEIGHT] = {
 	0xFFFFFFF0, 0xFFFFFFFC,
 };
 
-inline static void RenderLine(BYTE **dst, BYTE **src, int n, BYTE *tbl, DWORD mask)
-{
+inline static int count_leading_zeros(DWORD mask) {
+	// Note: This assumes that the argument is not zero,
+	// which means there is at least one bit set.
+#if defined(__GNUC__) || defined(__clang__)
+	return __builtin_clz(mask);
+#else
 	int i;
+	for (i = 0; (mask & 0x80000000) == 0; i++, mask <<= 1);
+	return i;
+#endif
+}
 
+template <typename F>
+void foreach_set_bit(DWORD mask, const F& f) {
+	int i = 0;
+	while (mask != 0) {
+		int z = count_leading_zeros(mask);
+		i += z, mask <<= z;
+		for (; mask & 0x80000000; i++, mask <<= 1)
+			f(i);
+	}
+}
+
+inline static void RenderLine(BYTE **dst, BYTE **src, int n, BYTE *tbl, DWORD mask)
+{
 #ifdef NO_OVERDRAW
 	if (*dst < gpBufStart || *dst > gpBufEnd) {
-		*src += n;
-		*dst += n;
-		return;
+		goto skip;
 	}
 #endif
 
 	if (mask == 0xFFFFFFFF) {
 		if (light_table_index == lightmax) {
 			memset(*dst, 0, n);
-			(*src) += n;
-			(*dst) += n;
 		} else if (light_table_index == 0) {
 			memcpy(*dst, *src, n);
-			(*src) += n;
-			(*dst) += n;
 		} else {
-			for (i = 0; i < n; i++, (*src)++, (*dst)++) {
-				(*dst)[0] = tbl[(*src)[0]];
+			for (int i = 0; i < n; i++) {
+				(*dst)[i] = tbl[(*src)[i]];
 			}
 		}
 	} else {
+		// The number of iterations is anyway limited by the size of the mask.
+		// So we can limit it by ANDing the mask with another mask that only keeps
+		// iterations that are lower than n. We can now avoid testing if i < n
+		// at every loop iteration.
+		mask &= ((((DWORD)1) << n) - 1) << ((sizeof(DWORD) * CHAR_BIT) - n);
+
 		if (light_table_index == lightmax) {
-			(*src) += n;
-			for (i = 0; i < n; i++, (*dst)++, mask <<= 1) {
-				if (mask & 0x80000000) {
-					(*dst)[0] = 0;
-				}
-			}
+			foreach_set_bit(mask, [=] (int i) { (*dst)[i] = 0; });
 		} else if (light_table_index == 0) {
-			for (i = 0; i < n; i++, (*src)++, (*dst)++, mask <<= 1) {
-				if (mask & 0x80000000) {
-					(*dst)[0] = (*src)[0];
-				}
-			}
+			foreach_set_bit(mask, [=] (int i) { (*dst)[i] = (*src)[i]; });
 		} else {
-			for (i = 0; i < n; i++, (*src)++, (*dst)++, mask <<= 1) {
-				if (mask & 0x80000000) {
-					(*dst)[0] = tbl[(*src)[0]];
-				}
-			}
+			foreach_set_bit(mask, [=] (int i) { (*dst)[i] = tbl[(*src)[i]]; });
 		}
 	}
+
+skip:
+	(*src) += n;
+	(*dst) += n;
 }
 
 #if defined(__clang__) || defined(__GNUC__)