Blitter: Use 32 bit memset where available

Fix Blitter_40bppAnim::SetRect

src/blitter/32bpp_anim.cpp

@@ -212,12 +212,10 @@ inline void Blitter_32bppAnim::Draw(const Blitter::BlitterParams *bp, ZoomLevel
 
 
 				case BM_BLACK_REMAP:
-					do {
-						*dst++ = Colour(0, 0, 0);
-						*anim++ = 0;
-						src_px++;
-						src_n++;
-					} while (--n != 0);
+					memset_colour(dst, _black_colour, n);
+					memset(anim, 0, n * sizeof(*anim));
+					src_px += n;
+					src_n += n;
 					break;
 
 				case BM_TRANSPARENT:

src/blitter/32bpp_optimized.cpp

@@ -11,6 +11,7 @@
 #include "../zoom_func.h"
 #include "../settings_type.h"
 #include "32bpp_optimized.hpp"
+#include "common.hpp"
 
 #include "../safeguards.h"
 
@@ -181,12 +182,9 @@ inline void Blitter_32bppOptimized::Draw(const Blitter::BlitterParams *bp, ZoomL
 					break;
 
 				case BM_BLACK_REMAP:
-					do {
-						*dst = Colour(0, 0, 0);
-						dst++;
-						src_px++;
-						src_n++;
-					} while (--n != 0);
+					memset_colour(dst, _black_colour, n);
+					src_px += n;
+					src_n += n;
 					break;
 
 				case BM_TRANSPARENT:

src/blitter/40bpp_anim.cpp

@@ -22,9 +22,6 @@
 /** Instantiation of the 40bpp with animation blitter factory. */
 static FBlitter_40bppAnim iFBlitter_40bppAnim;
 
-/** Cached black value. */
-static const Colour _black_colour(0, 0, 0);
-
 
 void Blitter_40bppAnim::SetPixel(void *video, int x, int y, uint8 colour)
 {
@@ -58,7 +55,7 @@ void Blitter_40bppAnim::SetRect(void *video, int x, int y, const uint8 *colours,
 		Colour *dst = (Colour *)video + x + y * _screen.pitch;
 		uint8 *dstanim = ((uint32 *)dst - (uint32 *)_screen.dst_ptr) + VideoDriver::GetInstance()->GetAnimBuffer();
 		do {
-			memset((uint32 *)dst, _black_colour.data, width * sizeof(uint32));
+			memset_colour(dst, _black_colour, width);
 			memcpy(dstanim, colours, width * sizeof(uint8));
 			dst += _screen.pitch;
 			dstanim += _screen.pitch;
@@ -96,15 +93,9 @@ void Blitter_40bppAnim::DrawRect(void *video, int width, int height, uint8 colou
 	uint8 *anim_line = ((uint32 *)video - (uint32 *)_screen.dst_ptr) + VideoDriver::GetInstance()->GetAnimBuffer();
 
 	do {
-		Colour *dst = (Colour *)video;
-		uint8 *anim = anim_line;
+		memset_colour((Colour *)video, _black_colour, width);
+		memset(anim_line, colour, width);
 
-		for (int i = width; i > 0; i--) {
-			*dst = _black_colour;
-			*anim = colour;
-			dst++;
-			anim++;
-		}
 		video = (uint32 *)video + _screen.pitch;
 		anim_line += _screen.pitch;
 	} while (--height);

src/blitter/common.hpp

@@ -14,6 +14,14 @@
 #include "../core/math_func.hpp"
 
 #include <utility>
+#include <wchar.h>
+
+#ifdef __APPLE__
+#include <string.h>
+#endif
+
+/** Cached black value. */
+static const Colour _black_colour(0, 0, 0);
 
 template <typename SetPixelT>
 void Blitter::DrawLineGeneric(int x1, int y1, int x2, int y2, int screen_width, int screen_height, int width, int dash, SetPixelT set_pixel)
@@ -192,4 +200,25 @@ void Blitter::DrawLineGeneric(int x1, int y1, int x2, int y2, int screen_width,
 	}
 }
 
+inline void memset_uint32(uint32 *s, uint32 c, size_t n)
+{
+#ifdef __APPLE__
+	memset_pattern4(static_cast<void *>(s), static_cast<const void *>(&c), n * sizeof(uint32));
+#else
+	if constexpr (sizeof(wchar_t) == sizeof(uint32)) {
+		wmemset((wchar_t *)s, (wchar_t)c, n);
+	} else {
+		for (; n > 0; n--) {
+			*s = c;
+			s++;
+		}
+	}
+#endif
+}
+
+inline void memset_colour(Colour *s, Colour c, size_t n)
+{
+	memset_uint32((uint32 *)s, c.data, n);
+}
+
 #endif /* BLITTER_COMMON_HPP */