From b3b7ed2a2faf8af8c5ad680354e08c2a3af59b32 Mon Sep 17 00:00:00 2001
From: skidau <skidau@gmail.com>
Date: Sat, 16 Jan 2010 03:43:16 +0000
Subject: [PATCH] FPU NaN and negative zero fixes.  Verified against the real
 hardware.

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@4847 8ced0084-cf51-0410-be5f-012b33b47a6e
---
 .../PowerPC/Interpreter/Interpreter_FPUtils.h | 45 ++++++++++---------
 .../Interpreter/Interpreter_FloatingPoint.cpp | 13 +++---
 .../Interpreter/Interpreter_Paired.cpp        |  8 ++--
 3 files changed, 36 insertions(+), 30 deletions(-)

diff --git a/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_FPUtils.h b/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_FPUtils.h
index 693d317a8d..0f54bd9060 100644
--- a/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_FPUtils.h
+++ b/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_FPUtils.h
@@ -24,7 +24,7 @@
 
 using namespace MathUtil;
 
-// warining! very slow!
+// warning! very slow! This setting fixes NAN
 //#define VERY_ACCURATE_FP
 
 #define MIN_SINGLE 0xc7efffffe0000000ull
@@ -188,26 +188,31 @@ inline double NI_madd(const double a, const double b, const double c)
 
 inline double NI_msub(const double a, const double b, const double c)
 {
-#ifdef VERY_ACCURATE_FP
-	if (a != a) return a;
-	if (c != c) return c;
-	if (b != b) return b;
-	double t = a * b;
-	if (t != t)
-	{
-		SetFPException(FPSCR_VXIMZ);
-		return PPC_NAN;
-	}
-	t = t - c;
-	if (t != t)
-	{
-		SetFPException(FPSCR_VXISI);
-		return PPC_NAN;
-	}
-	return t;
-#else
+//#ifdef VERY_ACCURATE_FP
+//  This code does not produce accurate fp!  NAN's are not calculated correctly, nor negative zero.
+//	The code is kept here for reference.
+//
+//	if (a != a) return a;
+//	if (c != c) return c;
+//	if (b != b) return b;
+//	double t = a * b;
+//	if (t != t)
+//	{
+//		SetFPException(FPSCR_VXIMZ);
+//		return PPC_NAN;
+//	}
+//
+//	t = t - c;
+//	if (t != t)
+//	{
+//		SetFPException(FPSCR_VXISI);
+//		return PPC_NAN;
+//	}
+//	return t;
+//#else
+//	This code does not calculate QNAN's correctly but calculates negative zero correctly.
 	return NI_sub(NI_mul(a, b), c);
-#endif
+// #endif
 }
 
 // used by stfsXX instructions and ps_rsqrte
diff --git a/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_FloatingPoint.cpp b/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_FloatingPoint.cpp
index 3693aef7b1..2db80f8195 100644
--- a/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_FloatingPoint.cpp
+++ b/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_FloatingPoint.cpp
@@ -232,7 +232,7 @@ void fnabsx(UGeckoInstruction _inst)
 {
 	riPS0(_inst.FD) = riPS0(_inst.FB) | (1ULL << 63);
 	// This is a binary instruction. Does not alter FPSCR
-	if (_inst.Rc) Helper_UpdateCR1(rPS0(_inst.FD));
+	if (_inst.Rc) Helper_UpdateCR1(rPS0(_inst.FD));	
 }	
 
 void fnegx(UGeckoInstruction _inst)
@@ -436,27 +436,28 @@ void fmsubsx(UGeckoInstruction _inst)
 	if (_inst.Rc) Helper_UpdateCR1(rPS0(_inst.FD)); 
 }
 
-
 void fnmaddx(UGeckoInstruction _inst)
 {
-	rPS0(_inst.FD) = ForceDouble(0.0-NI_madd(rPS0(_inst.FA), rPS0(_inst.FC), rPS0(_inst.FB)));
+	rPS0(_inst.FD) = ForceDouble(-NI_madd(rPS0(_inst.FA), rPS0(_inst.FC), rPS0(_inst.FB)));
  	UpdateFPRF(rPS0(_inst.FD));
 	if (_inst.Rc) Helper_UpdateCR1(rPS0(_inst.FD));
 }
 void fnmaddsx(UGeckoInstruction _inst)
 {
 	rPS0(_inst.FD) = rPS1(_inst.FD) = 
-		ForceSingle(0.0-NI_madd(rPS0(_inst.FA), rPS0(_inst.FC), rPS0(_inst.FB)));
+		ForceSingle(-NI_madd(rPS0(_inst.FA), rPS0(_inst.FC), rPS0(_inst.FB)));
  	UpdateFPRF(rPS0(_inst.FD));
 	if (_inst.Rc) Helper_UpdateCR1(rPS0(_inst.FD)); 
 }
 
 void fnmsubx(UGeckoInstruction _inst)
 {
-	rPS0(_inst.FD) = ForceDouble(0.0-NI_msub(rPS0(_inst.FA), rPS0(_inst.FC), rPS0(_inst.FB)));
+	rPS0(_inst.FD) = ForceDouble(-NI_msub(rPS0(_inst.FA), rPS0(_inst.FC), rPS0(_inst.FB)));
  	UpdateFPRF(rPS0(_inst.FD));
 	if (_inst.Rc) Helper_UpdateCR1(rPS0(_inst.FD));
 }
+
+// fnmsubsx does not handle QNAN properly - see NI_msub
 void fnmsubsx(UGeckoInstruction _inst)
 {
 	rPS0(_inst.FD) = rPS1(_inst.FD) =
@@ -465,13 +466,13 @@ void fnmsubsx(UGeckoInstruction _inst)
 	if (_inst.Rc) Helper_UpdateCR1(rPS0(_inst.FD)); 
 }
 
-
 void fsubx(UGeckoInstruction _inst)
 {
 	rPS0(_inst.FD) = ForceDouble(NI_sub(rPS0(_inst.FA), rPS0(_inst.FB)));
  	UpdateFPRF(rPS0(_inst.FD));
 	if (_inst.Rc) Helper_UpdateCR1(rPS0(_inst.FD));
 }
+
 void fsubsx(UGeckoInstruction _inst)
 {
 	rPS0(_inst.FD) = rPS1(_inst.FD) = ForceSingle(NI_sub(rPS0(_inst.FA), rPS0(_inst.FB)));
diff --git a/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_Paired.cpp b/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_Paired.cpp
index b172aae979..7222d390ac 100644
--- a/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_Paired.cpp
+++ b/Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_Paired.cpp
@@ -296,16 +296,16 @@ void ps_madd(UGeckoInstruction _inst)
 
 void ps_nmsub(UGeckoInstruction _inst)
 {
-	rPS0(_inst.FD) = ForceSingle( 0.0-NI_msub( rPS0(_inst.FA), rPS0(_inst.FC), rPS0(_inst.FB) ) );
-	rPS1(_inst.FD) = ForceSingle( 0.0-NI_msub( rPS1(_inst.FA), rPS1(_inst.FC), rPS1(_inst.FB) ) );
+	rPS0(_inst.FD) = ForceSingle( -NI_msub( rPS0(_inst.FA), rPS0(_inst.FC), rPS0(_inst.FB) ) );
+	rPS1(_inst.FD) = ForceSingle( -NI_msub( rPS1(_inst.FA), rPS1(_inst.FC), rPS1(_inst.FB) ) );
 	UpdateFPRF(rPS0(_inst.FD));
 	if (_inst.Rc) Helper_UpdateCR1(rPS0(_inst.FD));
 }
 
 void ps_nmadd(UGeckoInstruction _inst)
 {
-	rPS0(_inst.FD) = ForceSingle( 0.0-NI_madd( rPS0(_inst.FA), rPS0(_inst.FC), rPS0(_inst.FB) ) );
-	rPS1(_inst.FD) = ForceSingle( 0.0-NI_madd( rPS1(_inst.FA), rPS1(_inst.FC), rPS1(_inst.FB) ) );
+	rPS0(_inst.FD) = ForceSingle( -NI_madd( rPS0(_inst.FA), rPS0(_inst.FC), rPS0(_inst.FB) ) );
+	rPS1(_inst.FD) = ForceSingle( -NI_madd( rPS1(_inst.FA), rPS1(_inst.FC), rPS1(_inst.FB) ) );
 	UpdateFPRF(rPS0(_inst.FD));
 	if (_inst.Rc) Helper_UpdateCR1(rPS0(_inst.FD));
 }