Refactored ASM code of some hooks using 4.x as ref

Code edits to DisplayKarmaChanges.
Implemented the buf_to_buf_ replacement function from 4.x.

sfall/FalloutEngine.cpp

@@ -1057,6 +1057,72 @@ void __fastcall TransCscale(long i_width, long i_height, long s_width, long s_he
 	}
 }
 
+// buf_to_buf_ function with pure MMX implementation
+void __cdecl BufToBuf(void* src, long width, long height, long src_width, void* dst, long dst_width) {
+	if (height <= 0 || width <= 0) return;
+
+	size_t blockCount = width / 64; // 64 bytes
+	size_t remainder = width % 64;
+	size_t remainderD = remainder / 4;
+	size_t remainderB = remainder % 4;
+	size_t s_pitch = src_width - width;
+	size_t d_pitch = dst_width - width;
+
+	__asm {
+		mov  ebx, s_pitch;
+		mov  edx, d_pitch;
+		mov  esi, src;
+		mov  edi, dst;
+		mov  eax, height;
+	startLoop:
+		mov  ecx, blockCount;
+		test ecx, ecx;
+		jz   copySmall;
+	copyBlock: // copies block of 64 bytes
+		movq mm0, [esi];      // movups xmm0, [esi]; // SSE implementation
+		movq mm1, [esi + 8];
+		movq mm2, [esi + 16]; // movups xmm1, [esi + 16];
+		movq mm3, [esi + 24];
+		movq mm4, [esi + 32]; // movups xmm2, [esi + 32];
+		movq mm5, [esi + 40];
+		movq mm6, [esi + 48]; // movups xmm3, [esi + 48];
+		movq mm7, [esi + 56];
+		movq [edi], mm0;      // movups [edi], xmm0;
+		movq [edi + 8], mm1;
+		movq [edi + 16], mm2; // movups [edi + 16], xmm1;
+		movq [edi + 24], mm3;
+		movq [edi + 32], mm4; // movups xmm2, [esi + 32];
+		movq [edi + 40], mm5;
+		movq [edi + 48], mm6; // movups xmm3, [esi + 48];
+		movq [edi + 56], mm7;
+		add  esi, 64;
+		lea  edi, [edi + 64];
+		dec  ecx; // blockCount
+		jnz  copyBlock;
+		// copies the remaining bytes
+		mov  ecx, remainderD;
+		rep  movsd;
+		mov  ecx, remainderB;
+		rep  movsb;
+		add  esi, ebx; // s_pitch
+		add  edi, edx; // d_pitch
+		dec  eax;      // height
+		jnz  startLoop;
+		emms;
+		jmp  end;
+	copySmall: // copies the small size data
+		mov  ecx, remainderD;
+		rep  movsd;
+		mov  ecx, remainderB;
+		rep  movsb;
+		add  esi, ebx; // s_pitch
+		add  edi, edx; // d_pitch
+		dec  eax;      // height
+		jnz  copySmall;
+end:
+	}
+}
+
 long __fastcall GetGameConfigString(const char* outValue, const char* section, const char* param) {
 	__asm {
 		mov  ebx, param;

sfall/FalloutEngine.h

@@ -1167,6 +1167,9 @@ void __fastcall DisplayInWindow(long w_here, long width, long height, void* data
 
 void __fastcall TransCscale(long i_width, long i_height, long s_width, long s_height, long xy_shift, long w_width, void* data);
 
+// buf_to_buf_ function with pure MMX implementation
+void __cdecl BufToBuf(void* src, long width, long height, long src_width, void* dst, long dst_width);
+
 long __fastcall GetGameConfigString(const char* outValue, const char* section, const char* param);
 
 /* stdcall */

sfall/Graphics.cpp

@@ -781,10 +781,11 @@ public:
 		int pitch = dRect.Pitch;
 
 		if (GPUBlt) {
-			char* pBits = (char*)dRect.pBits;
-			for (DWORD y = 0; y < mveDesc.dwHeight; y++) {
-				CopyMemory(&pBits[y * pitch], &lockTarget[y * width], width);
-			}
+			BufToBuf(lockTarget, width, mveDesc.dwHeight, width, dRect.pBits, pitch);
+			//char* pBits = (char*)dRect.pBits;
+			//for (DWORD y = 0; y < mveDesc.dwHeight; y++) {
+			//	CopyMemory(&pBits[y * pitch], &lockTarget[y * width], width);
+			//}
 		} else {
 			pitch /= 4;
 			for (DWORD y = 0; y < mveDesc.dwHeight; y++) {
@@ -1011,7 +1012,7 @@ public:
 	HRESULT __stdcall SetEntries(DWORD a, DWORD b, DWORD c, LPPALETTEENTRY destPal) { // used to set palette for splash screen, fades, subtitles
 		if (!windowInit || c == 0 || b + c > 256) return DDERR_INVALIDPARAMS;
 
-		CopyMemory(&palette[b], destPal, c * 4);
+		__movsd(&palette[b], (unsigned long*)destPal, c);
 
 		if (GPUBlt && gpuPalette) {
 			D3DLOCKED_RECT rect;
@@ -1308,6 +1309,9 @@ void GraphicsInit() {
 		fadeMulti = ((double)fadeMulti) / 100.0;
 		dlogr(" Done", DL_INIT);
 	}
+
+	// Replace the srcCopy_ function with a pure MMX implementation
+	MakeJump(0x4D36D4, BufToBuf); // buf_to_buf_
 }
 
 void GraphicsExit() {

sfall/Karma.cpp

@@ -61,30 +61,26 @@ skip:
 }
 
 static void __stdcall SetKarma(int value) {
-	int old;
-	__asm {
-		xor eax, eax;
-		call game_get_global_var_;
-		mov old, eax;
-	}
-	old = value - old;
 	char buf[128];
-	if (old == 0) return;
-	if (old > 0) {
-		sprintf_s(buf, karmaGainMsg, old);
+	if (value > 0) {
+		sprintf_s(buf, karmaGainMsg, value);
 	} else {
-		sprintf_s(buf, karmaLossMsg, -old);
+		sprintf_s(buf, karmaLossMsg, -value);
 	}
 	DisplayPrint(buf);
 }
 
 static void __declspec(naked) SetGlobalVarWrapper() {
 	__asm {
-		test eax, eax; // GVar number
+		test eax, eax; // Gvar number
 		jnz  end;
 		pushadc;
+		call game_get_global_var_;
+		sub  edx, eax; // value -= old
+		jz   skip;
 		push edx;
 		call SetKarma;
+skip:
 		popadc;
 end:
 		jmp  game_set_global_var_;