From b45b91bf07289caa31f0fe1e68b5235c0f063ee6 Mon Sep 17 00:00:00 2001
From: "Jake.Stine" <Jake.Stine@96395faa-99c1-11dd-bbfe-3dabce05a288>
Date: Sun, 20 Dec 2009 01:28:23 +0000
Subject: [PATCH] More minor improvements to cpuSpeed calculations, since my
 experimental fix apparently works. :)

git-svn-id: http://pcsx2.googlecode.com/svn/trunk@2364 96395faa-99c1-11dd-bbfe-3dabce05a288
---
 common/src/x86emitter/WinCpuDetect.cpp |  9 ++++++++-
 common/src/x86emitter/cpudetect.cpp    | 18 ++++++++++++++----
 2 files changed, 22 insertions(+), 5 deletions(-)

diff --git a/common/src/x86emitter/WinCpuDetect.cpp b/common/src/x86emitter/WinCpuDetect.cpp
index af87a20a69..06888dd982 100644
--- a/common/src/x86emitter/WinCpuDetect.cpp
+++ b/common/src/x86emitter/WinCpuDetect.cpp
@@ -88,8 +88,15 @@ SingleCoreAffinity::SingleCoreAffinity()
 		);
 	}
 	
-	// Force Windows to timeslice (hoping this fixes some affinity issues)
 	Sleep( 2 );
+
+	// Sleep Explained: I arbitrarily pick Core 0 to lock to for running the CPU test.  This
+	// means that the current thread will need to be switched to Core 0 if it's currently
+	// scheduled on a difference cpu/core.  However, Windows does not necessarily perform
+	// that scheduling immediately upon the call to SetThreadAffinityMask (seems dependent
+	// on version: XP does, Win7 does not).  So by issuing a Sleep here we give Win7 time
+	// to issue a timeslice and move our thread to Core 0.  Without this, it tends to move
+	// the thread during the cpuSpeed test instead, causing totally wacky results.
 };
 
 SingleCoreAffinity::~SingleCoreAffinity() throw()
diff --git a/common/src/x86emitter/cpudetect.cpp b/common/src/x86emitter/cpudetect.cpp
index 3eb850297c..efbf17f424 100644
--- a/common/src/x86emitter/cpudetect.cpp
+++ b/common/src/x86emitter/cpudetect.cpp
@@ -34,7 +34,7 @@ static const char* bool_to_char( bool testcond )
 static s64 CPUSpeedHz( u64 time )
 {
 	u64 timeStart, timeStop;
-	s64 startTick, endTick;
+	s64 startCycle, endCycle;
 
 	if( ! x86caps.hasTimeStampCounter )
 		return 0;
@@ -49,18 +49,28 @@ static s64 CPUSpeedHz( u64 time )
 	do
 	{
 		timeStop = GetCPUTicks();
-		startTick = __rdtsc();
+		startCycle = __rdtsc();
 	} while( ( timeStop - timeStart ) == 0 );
 
 	timeStart = timeStop;
 	do
 	{
 		timeStop = GetCPUTicks();
-		endTick = __rdtsc();
+		endCycle = __rdtsc();
 	}
 	while( ( timeStop - timeStart ) < time );
 
-	return (s64)( endTick - startTick );
+	s64 cycleCount = endCycle - startCycle;
+	s64 timeCount = timeStop - timeStart;
+	s64 overrun = timeCount - time;
+	if( !overrun ) return cycleCount;
+
+	// interference could cause us to overshoot the target time, compensate:
+	
+	double cyclesPerTick = (double)cycleCount / (double)timeCount;
+	double newCycleCount = (double)cycleCount - (cyclesPerTick * overrun);
+
+	return (s64)newCycleCount;
 }
 
 // Recompiled code buffer for SSE and MXCSR feature testing.