From b4e12ad5c9fc9d5dcb4b599ba944a1a5330acdde Mon Sep 17 00:00:00 2001
From: LinesPrower <LinesPrower@gmail.com>
Date: Sun, 27 Sep 2009 14:19:38 +0000
Subject: [PATCH] Some improvements of FP and DMA

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@4334 8ced0084-cf51-0410-be5f-012b33b47a6e
---
 .../Interpreter/Interpreter_FloatingPoint.cpp | 26 +++++++++----------
 .../Interpreter/Interpreter_LoadStore.cpp     | 11 ++++----
 .../Interpreter/Interpreter_Paired.cpp        | 12 +--------
 .../Interpreter_SystemRegisters.cpp           |  5 ++++
 4 files changed, 23 insertions(+), 31 deletions(-)

diff --git a/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_FloatingPoint.cpp b/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_FloatingPoint.cpp
index a7a44c1c6b..b9e006ab85 100644
--- a/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_FloatingPoint.cpp
+++ b/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_FloatingPoint.cpp
@@ -105,7 +105,6 @@ void fcmpu(UGeckoInstruction _inst)
 }
 
 // Apply current rounding mode
-// need to investigate this instruction.
 void fctiwx(UGeckoInstruction _inst)
 {
 	const double b = rPS0(_inst.FB);
@@ -133,7 +132,7 @@ void fctiwx(UGeckoInstruction _inst)
 			{
 				double t = b + 0.5;
 				i = (s32)t;
-				if (t - i < 0) i--;
+				if (t - i < 0 || (t - i == 0 && b > 0)) i--;
 				break;
 			}
 		case 1: // zero
@@ -161,19 +160,15 @@ void fctiwx(UGeckoInstruction _inst)
 			FPSCR.FR = fabs(di) > fabs(b);
 		}
 	}	
-	
-	//FPRF undefined
-
-	riPS0(_inst.FD) = (u64)value; // zero extend
-	if (_inst.Rc) 
+	// based on HW tests
+	// FPRF is not affected
+	riPS0(_inst.FD) = 0xfff8000000000000ull | value;
+	if (value == 0 && ( (*(u64*)&b) & DOUBLE_SIGN ))
+		riPS0(_inst.FD) |= 0x100000000ull;
+	if (_inst.Rc)
 		Helper_UpdateCR1(rPS0(_inst.FD));
 }
 
-/*
-In the float -> int direction, floating point input values larger than the largest 
-representable int result in 0x80000000 (a very negative number) rather than the 
-largest representable int on PowerPC. */
-
 // Always round toward zero
 void fctiwzx(UGeckoInstruction _inst)
 {
@@ -209,8 +204,11 @@ void fctiwzx(UGeckoInstruction _inst)
 		}
 		value = (u32)i;
 	}
-
-	riPS0(_inst.FD) = (u64)value;
+	// based on HW tests
+	// FPRF is not affected
+	riPS0(_inst.FD) = 0xfff8000000000000ull | value;
+	if (value == 0 && ( (*(u64*)&b) & DOUBLE_SIGN ))
+		riPS0(_inst.FD) |= 0x100000000ull;	
 	if (_inst.Rc)
 		Helper_UpdateCR1(rPS0(_inst.FD));
 }
diff --git a/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_LoadStore.cpp b/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_LoadStore.cpp
index 10d9cc7b00..999c5e3582 100644
--- a/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_LoadStore.cpp
+++ b/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_LoadStore.cpp
@@ -315,19 +315,18 @@ void dcbtst(UGeckoInstruction _inst)
 	_assert_msg_(POWERPC,0,"dcbtst - Not implemented");
 }
 
-// __________________________________________________________________________________________________
-// dcbz
-// TODO(ector) check docs
 void dcbz(UGeckoInstruction _inst)
 {
     // !!! after the dcbz follows a dcbf... dont clear the memory in this case !!!
     // 0x81330c2c
-    u32 NextOpcode = Memory::Read_U32(PC+4);
+    // LinesPrower: why? does it break something?
+    //              according to docs dcbz->dcbf should clear the memory immediately!
+    //              However, dcbz->dcbi won't clear anything, but that's completely senseless
+   /* u32 NextOpcode = Memory::Read_U32(PC+4);
     if (NextOpcode == 0x7C0400AC)
     {
         return;
-    }
-
+    }*/
 	// HACK but works... we think
 	Memory::Memset(Helper_Get_EA_X(_inst) & (~31), 0, 32);
 }
diff --git a/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_Paired.cpp b/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_Paired.cpp
index ce035f48ee..b172aae979 100644
--- a/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_Paired.cpp
+++ b/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_Paired.cpp
@@ -473,18 +473,8 @@ void ps_cmpo1(UGeckoInstruction _inst)
 // TODO(ector) check docs
 void dcbz_l(UGeckoInstruction _inst)
 {
-	// This is supposed to allocate a cache line in the locked cache. Not entirely sure how
-	// this is visible to the rest of the world. For now, we ignore it.
-	/*
-	addr_t ea = Helper_Get_EA(_inst);
-
-	u32 blockStart = ea & (~(CACHEBLOCKSIZE-1));
-	u32 blockEnd = blockStart + CACHEBLOCKSIZE;
-
 	//FAKE: clear memory instead of clearing the cache block
-	for (int i=blockStart; i<blockEnd; i+=4)
-	Memory::Write_U32(0,i);
-	*/
+	Memory::Memset(Helper_Get_EA_X(_inst) & (~31), 0, 32);
 }
 
 }  // namespace
diff --git a/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_SystemRegisters.cpp b/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_SystemRegisters.cpp
index e6f5a144f8..a439b28270 100644
--- a/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_SystemRegisters.cpp
+++ b/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_SystemRegisters.cpp
@@ -323,6 +323,9 @@ void mtspr(UGeckoInstruction _inst)
 			UReg_HID2 old_hid2;
 			old_hid2.Hex = oldValue;
 
+			//if (HID2.LCE && !old_hid2.LCE)
+			//	PanicAlert("Locked cache enabled!");
+
 			if (HID2.PSE == 0)
 				PanicAlert("WARNING: PSE in HID2 isnt set");
 
@@ -363,6 +366,7 @@ void mtspr(UGeckoInstruction _inst)
 			u32 dwMemAddress = DMAU.MEM_ADDR << 5;
 			u32 dwCacheAddress = DMAL.LC_ADDR << 5;
 			u32 iLength = ((DMAU.DMA_LEN_U << 2) | DMAL.DMA_LEN_L);
+			//INFO_LOG(POWERPC, "DMA: mem = %x, cache = %x, len = %u, LD = %d, PC=%x", dwMemAddress, dwCacheAddress, iLength, (int)DMAL.DMA_LD, PC);
 			if (iLength == 0) 
 				iLength = 128;
 			if (DMAL.DMA_LD)
@@ -370,6 +374,7 @@ void mtspr(UGeckoInstruction _inst)
 			else
 				Memory::DMA_LCToMemory(dwMemAddress, dwCacheAddress, iLength);
 		}
+		DMAL.DMA_T = 0;
 		break;
 
 	case SPR_DEC: