From 846178a588cf9aca06b0f67fafcc5ad11d1f5bc8 Mon Sep 17 00:00:00 2001
From: "Unknown W. Brackets" <checkins@unknownbrackets.org>
Date: Thu, 15 Aug 2013 01:26:16 -0700
Subject: [PATCH] Optimize thread switching a bit.

~2.5% improvement in Zettai Hero Project (while multithreading.)
---
 Core/HLE/sceKernelThread.cpp | 22 ++++++----------------
 Core/HLE/sceKernelThread.h   | 20 ++++++++++++++------
 Core/MIPS/MIPS.h             | 24 +++++++++++++++---------
 3 files changed, 35 insertions(+), 31 deletions(-)

diff --git a/Core/HLE/sceKernelThread.cpp b/Core/HLE/sceKernelThread.cpp
index 230191cae7..04d59d1488 100644
--- a/Core/HLE/sceKernelThread.cpp
+++ b/Core/HLE/sceKernelThread.cpp
@@ -1412,8 +1412,8 @@ u32 sceKernelGetThreadmanIdList(u32 type, u32 readBufPtr, u32 readBufSize, u32 i
 // Saves the current CPU context
 void __KernelSaveContext(ThreadContext *ctx, bool vfpuEnabled)
 {
-	memcpy(ctx->r, currentMIPS->r, sizeof(ctx->r));
-	memcpy(ctx->f, currentMIPS->f, sizeof(ctx->f));
+	// r and f are immediately next to each other and must be.
+	memcpy(ctx->r, currentMIPS->r, sizeof(ctx->r) + sizeof(ctx->f));
 
 	if (vfpuEnabled)
 	{
@@ -1421,19 +1421,14 @@ void __KernelSaveContext(ThreadContext *ctx, bool vfpuEnabled)
 		memcpy(ctx->vfpuCtrl, currentMIPS->vfpuCtrl, sizeof(ctx->vfpuCtrl));
 	}
 
-	ctx->pc = currentMIPS->pc;
-	ctx->hi = currentMIPS->hi;
-	ctx->lo = currentMIPS->lo;
-	ctx->fcr0 = currentMIPS->fcr0;
-	ctx->fcr31 = currentMIPS->fcr31;
-	ctx->fpcond = currentMIPS->fpcond;
+	memcpy(ctx->other, currentMIPS->other, sizeof(ctx->other));
 }
 
 // Loads a CPU context
 void __KernelLoadContext(ThreadContext *ctx, bool vfpuEnabled)
 {
-	memcpy(currentMIPS->r, ctx->r, sizeof(ctx->r));
-	memcpy(currentMIPS->f, ctx->f, sizeof(ctx->f));
+	// r and f are immediately next to each other and must be.
+	memcpy(currentMIPS->r, ctx->r, sizeof(ctx->r) + sizeof(ctx->f));
 
 	if (vfpuEnabled)
 	{
@@ -1441,12 +1436,7 @@ void __KernelLoadContext(ThreadContext *ctx, bool vfpuEnabled)
 		memcpy(currentMIPS->vfpuCtrl, ctx->vfpuCtrl, sizeof(ctx->vfpuCtrl));
 	}
 
-	currentMIPS->pc = ctx->pc;
-	currentMIPS->hi = ctx->hi;
-	currentMIPS->lo = ctx->lo;
-	currentMIPS->fcr0 = ctx->fcr0;
-	currentMIPS->fcr31 = ctx->fcr31;
-	currentMIPS->fpcond = ctx->fpcond;
+	memcpy(currentMIPS->other, ctx->other, sizeof(ctx->other));
 
 	// Reset the llBit, the other thread may have touched memory.
 	currentMIPS->llBit = 0;
diff --git a/Core/HLE/sceKernelThread.h b/Core/HLE/sceKernelThread.h
index e931f959f4..147677ce1e 100644
--- a/Core/HLE/sceKernelThread.h
+++ b/Core/HLE/sceKernelThread.h
@@ -109,19 +109,27 @@ void __KernelRegisterWaitTypeFuncs(WaitType type, WaitBeginCallbackFunc beginFun
 struct ThreadContext
 {
 	void reset();
+
+	// r must be followed by f.
 	u32 r[32];
 	float f[32];
+
 	float v[128];
 	u32 vfpuCtrl[16];
 
-	u32 pc;
+	union {
+		struct {
+			u32 pc;
 
-	u32 hi;
-	u32 lo;
+			u32 hi;
+			u32 lo;
 
-	u32 fcr0;
-	u32 fcr31;
-	u32 fpcond;
+			u32 fcr0;
+			u32 fcr31;
+			u32 fpcond;
+		};
+		u32 other[6];
+	};
 };
 
 // Internal API, used by implementations of kernel functions
diff --git a/Core/MIPS/MIPS.h b/Core/MIPS/MIPS.h
index 3503c7c73a..5ac5a7c99f 100644
--- a/Core/MIPS/MIPS.h
+++ b/Core/MIPS/MIPS.h
@@ -107,7 +107,7 @@ public:
 	void Reset();
 	void DoState(PointerWrap &p);
 
-	// MUST start with r!
+	// MUST start with r and be followed by f!
 	u32 r[32];
 	union {
 		float f[32];
@@ -122,17 +122,23 @@ public:
 	// If vfpuCtrl (prefixes) get mysterious values, check the VFPU regcache code.
 	u32 vfpuCtrl[16];
 
-	u32 pc;
+	union {
+		struct {
+			u32 pc;
+
+			u32 hi;
+			u32 lo;
+
+			u32 fcr0;
+			u32 fcr31; //fpu control register
+			u32 fpcond;  // cache the cond flag of fcr31  (& 1 << 23)
+		};
+		u32 other[6];
+	};
+
 	u32 nextPC;
 	int downcount;  // This really doesn't belong here, it belongs in CoreTiming. But you gotta do what you gotta do, this needs to be reachable in the ARM JIT.
 
-	u32 hi;
-	u32 lo;
-
-	u32 fcr0;
-	u32 fcr31; //fpu control register
-	u32 fpcond;  // cache the cond flag of fcr31  (& 1 << 23)
-
 	bool inDelaySlot;
 	int llBit;  // ll/sc