Support swizzled framebuffer downloads.

Used in God Eater 2 when showing the load save screen.

Core/HLE/ReplaceTables.cpp

@@ -142,6 +142,42 @@ static int Replace_memcpy16() {
 	return 10 + bytes / 4;  // approximation
 }
 
+static int Replace_memcpy_swizzled() {
+	u32 destPtr = PARAM(0);
+	u32 srcPtr = PARAM(1);
+	u32 pitch = PARAM(2);
+	u32 h = PARAM(4);
+	if (Memory::IsVRAMAddress(srcPtr)) {
+		// Cheat a bit to force a download of the framebuffer.
+		// VRAM + 0x00400000 is simply a VRAM mirror.
+		gpu->UpdateMemory(srcPtr ^ 0x00400000, srcPtr, pitch * h);
+	}
+	u8 *dstp = Memory::GetPointerUnchecked(destPtr);
+	const u8 *srcp = Memory::GetPointerUnchecked(srcPtr);
+
+	const u8 *ysrcp = srcp;
+	for (u32 y = 0; y < h; y += 8) {
+		const u8 *xsrcp = ysrcp;
+		for (u32 x = 0; x < pitch; x += 16) {
+			const u8 *src = xsrcp;
+			for (int n = 0; n < 8; ++n) {
+				memcpy(dstp, src, 16);
+				src += pitch;
+				dstp += 16;
+			}
+			xsrcp += 16;
+		}
+		ysrcp += 8 * pitch;
+	}
+
+	RETURN(0);
+#ifndef MOBILE_DEVICE
+	CBreakPoints::ExecMemCheck(srcPtr, false, pitch * h, currentMIPS->pc);
+	CBreakPoints::ExecMemCheck(destPtr, true, pitch * h, currentMIPS->pc);
+#endif
+	return 10 + (pitch * h) / 4;  // approximation
+}
+
 static int Replace_memmove() {
 	u32 destPtr = PARAM(0);
 	u32 srcPtr = PARAM(1);
@@ -420,6 +456,7 @@ static const ReplacementTableEntry entries[] = {
 	{ "ceilf", &Replace_ceilf, 0, 0},
 	{ "memcpy", &Replace_memcpy, 0, 0},
 	{ "memcpy16", &Replace_memcpy16, 0, 0},
+	{ "memcpy_swizzled", &Replace_memcpy_swizzled, 0, 0},
 	{ "memmove", &Replace_memmove, 0, 0},
 	{ "memset", &Replace_memset, 0, 0},
 	{ "strlen", &Replace_strlen, 0, 0},

Core/MIPS/MIPSAnalyst.cpp

@@ -220,6 +220,7 @@ static const HardHashTableEntry hardcodedHashes[] = {
 	{ 0x6301fa5149bd973a, 120, "wcscat", },
 	{ 0x658b07240a690dbd, 36, "strlen", },
 	{ 0x66122f0ab50b2ef9, 296, "dl_write_dither_matrix_5", },
+	{ 0x66f7f1beccbc104a, 256, "memcpy_swizzled", }, // God Eater 2
 	{ 0x679e647e34ecf7f1, 132, "roundf", },
 	{ 0x67afe74d9ec72f52, 4380, "_strtod_r", },
 	{ 0x68b22c2aa4b8b915, 400, "sqrt", },
@@ -738,6 +739,7 @@ skip:
 		bool looking = false;
 		bool end = false;
 		bool isStraightLeaf = true;
+		bool decreasedSp = false;
 
 		u32 addr;
 		u32 addrNextSym = 0;
@@ -761,6 +763,8 @@ skip:
 				furthestBranch = 0;
 				looking = false;
 				end = false;
+				isStraightLeaf = false;
+				decreasedSp = false;
 				continue;
 			}
 
@@ -771,6 +775,7 @@ skip:
 				if (target > furthestBranch) {
 					furthestBranch = target;
 				}
+			// j X
 			} else if ((op & 0xFC000000) == 0x08000000) {
 				u32 sureTarget = GetJumpTarget(addr);
 				// Check for a tail call.  Might not even have a jr ra.
@@ -782,6 +787,18 @@ skip:
 						end = true;
 					}
 				} else if (sureTarget != INVALIDTARGET && sureTarget > addr && sureTarget > furthestBranch) {
+					static const u32 MAX_JUMP_FORWARD = 128;
+					// If it's a nearby forward jump, and not a stackless leaf, assume not a tail call.
+					if (sureTarget <= addr + MAX_JUMP_FORWARD && decreasedSp) {
+						// But let's check the delay slot.
+						MIPSOpcode op = Memory::Read_Instruction(addr + 4);
+						// addiu sp, sp, +X
+						if ((op & 0xFFFF8000) != 0x27BD0000) {
+							furthestBranch = sureTarget;
+							continue;
+						}
+					}
+
 					// A jump later.  Probably tail, but let's check if it jumps back.
 					u32 knownEnd = furthestBranch == 0 ? addr : furthestBranch;
 					u32 jumpback = ScanAheadForJumpback(sureTarget, currentFunction.start, knownEnd);
@@ -806,6 +823,14 @@ skip:
 					end = true;
 				}
 			}
+			// addiu sp, sp, -X
+			if ((op & 0xFFFF8000) == 0x27BD8000) {
+				decreasedSp = true;
+			}
+			// addiu sp, sp, +X
+			if ((op & 0xFFFF8000) == 0x27BD0000) {
+				decreasedSp = false;
+			}
 
 			if (looking) {
 				if (addr >= furthestBranch) {
@@ -838,6 +863,7 @@ skip:
 				looking = false;
 				end = false;
 				isStraightLeaf = true;
+				decreasedSp = false;
 				currentFunction.start = addr+4;
 			}
 		}

assets/knownfuncs.ini

@@ -140,6 +140,7 @@
 6301fa5149bd973a:120 = wcscat
 658b07240a690dbd:36 = strlen
 66122f0ab50b2ef9:296 = dl_write_dither_matrix_5
+66f7f1beccbc104a:256 = memcpy_swizzled
 679e647e34ecf7f1:132 = roundf
 67afe74d9ec72f52:4380 = _strtod_r
 68b22c2aa4b8b915:400 = sqrt