Change Linux fasttimer implementation back to RDTSC - using a reliable syscall was REALLY chewing CPU time. Sigh. I didn't realize how incredibly often this gets called. So, back to the assembly.

But be more careful with CPU clock count on linux, so the fasttimer values are much more accurate than they were the last time we were with RDTSC, in absolute terms - back in the right order of magnitude anyway.

Also change many instances of Mhz to MHz.
Also some minor comment fixes.

9 files changed, 43 insertions, 28 deletions

diff --git a/indra/llcommon/llfasttimer.h b/indra/llcommon/llfasttimer.h
index 48461df6ae..840d09d970 100644
--- a/indra/llcommon/llfasttimer.h
+++ b/indra/llcommon/llfasttimer.h
@@ -36,6 +36,10 @@
 // pull in the actual class definition
 #include "llfasttimer_class.h"
 
+//
+// Important note: These implementations must be FAST!
+//
+
 #if LL_WINDOWS
 //
 // Windows implementation of CPU clock
@@ -99,15 +103,17 @@ inline U64 LLFastTimer::getCPUClockCount64()
 #endif
 
 
-#if LL_LINUX || LL_SOLARIS
+#if (LL_LINUX || LL_SOLARIS) && !(defined(__i386__) || defined(__amd64__))
 //
-// Linux and Solaris implementation of CPU clock - all architectures.
+// Linux and Solaris implementation of CPU clock - non-x86.
+// This is accurate but SLOW!  Only use out of desperation.
 //
 // Try to use the MONOTONIC clock if available, this is a constant time counter
-// with nanosecond resolution (but not necessarily accuracy) and attempts are made
-// to synchronize this value between cores at kernel start. It should not be affected
-// by CPU frequency. If not available use the REALTIME clock, but this may be affected by
-// NTP adjustments or other user activity affecting the system time.
+// with nanosecond resolution (but not necessarily accuracy) and attempts are
+// made to synchronize this value between cores at kernel start. It should not
+// be affected by CPU frequency. If not available use the REALTIME clock, but
+// this may be affected by NTP adjustments or other user activity affecting
+// the system time.
 inline U64 LLFastTimer::getCPUClockCount64()
 {
 	struct timespec tp;
@@ -124,12 +130,12 @@ inline U32 LLFastTimer::getCPUClockCount32()
 {
 	return (U32)(LLFastTimer::getCPUClockCount64() >> 8);
 }
-#endif // (LL_LINUX || LL_SOLARIS))
+#endif // (LL_LINUX || LL_SOLARIS) && !(defined(__i386__) || defined(__amd64__))
 
 
-#if (LL_DARWIN) && (defined(__i386__) || defined(__amd64__))
+#if (LL_LINUX || LL_SOLARIS || LL_DARWIN) && (defined(__i386__) || defined(__amd64__))
 //
-// Mac x86 implementation of CPU clock
+// Mac+Linux+Solaris FAST x86 implementation of CPU clock
 inline U32 LLFastTimer::getCPUClockCount32()
 {
 	U64 x;
diff --git a/indra/llcommon/llfasttimer_class.cpp b/indra/llcommon/llfasttimer_class.cpp
index 2e5edb1f3b..f39a4e6619 100644
--- a/indra/llcommon/llfasttimer_class.cpp
+++ b/indra/llcommon/llfasttimer_class.cpp
@@ -230,17 +230,17 @@ void LLFastTimer::DeclareTimer::updateCachedPointers()
 }
 
 //static
-#if LL_LINUX || LL_SOLARIS || ( LL_DARWIN && !(defined(__i386__) || defined(__amd64__)) )
+#if (LL_DARWIN || LL_LINUX || LL_SOLARIS) && !(defined(__i386__) || defined(__amd64__))
 U64 LLFastTimer::countsPerSecond() // counts per second for the *32-bit* timer
 {
 	return sClockResolution >> 8;
 }
-#else // windows or x86-mac
+#else // windows or x86-mac or x86-linux or x86-solaris
 U64 LLFastTimer::countsPerSecond() // counts per second for the *32-bit* timer
 {
 	static U64 sCPUClockFrequency = U64(CProcessor().GetCPUFrequency(50));
 
-	// we drop the low-order byte in out timers, so report a lower frequency
+	// we drop the low-order byte in our timers, so report a lower frequency
 	return sCPUClockFrequency >> 8;
 }
 #endif
diff --git a/indra/llcommon/llprocessor.cpp b/indra/llcommon/llprocessor.cpp
index 8a4a4a8f9a..f6ab55a6b5 100644
--- a/indra/llcommon/llprocessor.cpp
+++ b/indra/llcommon/llprocessor.cpp
@@ -60,6 +60,10 @@
 #	include <windows.h>
 #endif
 
+#if LL_LINUX
+#include "llsys.h"
+#endif // LL_LINUX
+
 #if !LL_DARWIN && !LL_SOLARIS
 
 #ifdef PROCESSOR_FREQUENCY_MEASURE_AVAILABLE
@@ -116,6 +120,11 @@ CProcessor::CProcessor()
 ////////////////////////////////////////////////////////////////////////////
 F64 CProcessor::GetCPUFrequency(unsigned int uiMeasureMSecs)
 {
+#if LL_LINUX
+	// use the shinier LLCPUInfo interface
+	return 1000000.0F * gSysCPU.getMHz();
+#endif
+
 #ifndef PROCESSOR_FREQUENCY_MEASURE_AVAILABLE
 	return 0;
 #else
@@ -781,7 +790,7 @@ bool CProcessor::AnalyzeAMDProcessor()
 		case 5:			// Family = 5:  K5 / K6 processor family
 			switch (CPUInfo.uiModel)
 			{
-				case 0:			// Model = 0:  K5 SSA 5 (Pentium Rating *ggg* 75, 90 and 100 Mhz)
+				case 0:			// Model = 0:  K5 SSA 5 (Pentium Rating *ggg* 75, 90 and 100 MHz)
 					strcpy(CPUInfo.strModel, "AMD K5 SSA5 (PR75, PR90, PR100)");		/* Flawfinder: ignore */
 					strncat(strCPUName, "AMD K5 SSA5 (PR75, PR90, PR100)", sizeof(strCPUName) - strlen(strCPUName) -1);		/* Flawfinder: ignore */
 					break;
diff --git a/indra/llcommon/llsys.cpp b/indra/llcommon/llsys.cpp
index 0272c55db2..52b1b63209 100644
--- a/indra/llcommon/llsys.cpp
+++ b/indra/llcommon/llsys.cpp
@@ -519,15 +519,15 @@ LLCPUInfo::LLCPUInfo()
 	mHasSSE = info->_Ext.SSE_StreamingSIMD_Extensions;
 	mHasSSE2 = info->_Ext.SSE2_StreamingSIMD2_Extensions;
 	mHasAltivec = info->_Ext.Altivec_Extensions;
-	mCPUMhz = (S32)(proc.GetCPUFrequency(50)/1000000.0);
+	mCPUMHz = (F64)(proc.GetCPUFrequency(50)/1000000.0F);
 	mFamily.assign( info->strFamily );
 	mCPUString = "Unknown";
 
 #if LL_WINDOWS || LL_DARWIN || LL_SOLARIS
 	out << proc.strCPUName;
-	if (200 < mCPUMhz && mCPUMhz < 10000)           // *NOTE: cpu speed is often way wrong, do a sanity check
+	if (200 < mCPUMHz && mCPUMHz < 10000)           // *NOTE: cpu speed is often way wrong, do a sanity check
 	{
-		out << " (" << mCPUMhz << " MHz)";
+		out << " (" << mCPUMHz << " MHz)";
 	}
 	mCPUString = out.str();
 	
@@ -572,7 +572,7 @@ LLCPUInfo::LLCPUInfo()
 	if (LLStringUtil::convertToF64(cpuinfo["cpu mhz"], mhz)
 	    && 200.0 < mhz && mhz < 10000.0)
 	{
-		mCPUMhz = (S32)llrint(mhz);
+		mCPUMHz = (F64)(mhz);
 	}
 	if (!cpuinfo["model name"].empty())
 		mCPUString = cpuinfo["model name"];
@@ -595,9 +595,9 @@ bool LLCPUInfo::hasSSE2() const
 	return mHasSSE2;
 }
 
-S32 LLCPUInfo::getMhz() const
+F64 LLCPUInfo::getMHz() const
 {
-	return mCPUMhz;
+	return mCPUMHz;
 }
 
 std::string LLCPUInfo::getCPUString() const
@@ -644,7 +644,7 @@ void LLCPUInfo::stream(std::ostream& s) const
 	s << "->mHasSSE:     " << (U32)mHasSSE << std::endl;
 	s << "->mHasSSE2:    " << (U32)mHasSSE2 << std::endl;
 	s << "->mHasAltivec: " << (U32)mHasAltivec << std::endl;
-	s << "->mCPUMhz:     " << mCPUMhz << std::endl;
+	s << "->mCPUMHz:     " << mCPUMHz << std::endl;
 	s << "->mCPUString:  " << mCPUString << std::endl;
 }
 
diff --git a/indra/llcommon/llsys.h b/indra/llcommon/llsys.h
index c2c45bec9a..0b34951149 100644
--- a/indra/llcommon/llsys.h
+++ b/indra/llcommon/llsys.h
@@ -81,7 +81,7 @@ public:
 	bool hasAltivec() const;
 	bool hasSSE() const;
 	bool hasSSE2() const;
-	S32	 getMhz() const;
+	F64 getMHz() const;
 
 	// Family is "AMD Duron" or "Intel Pentium Pro"
 	const std::string& getFamily() const { return mFamily; }
@@ -90,7 +90,7 @@ private:
 	bool mHasSSE;
 	bool mHasSSE2;
 	bool mHasAltivec;
-	S32 mCPUMhz;
+	F64 mCPUMHz;
 	std::string mFamily;
 	std::string mCPUString;
 };
diff --git a/indra/llcommon/lltimer.cpp b/indra/llcommon/lltimer.cpp
index 25b768079b..6111db2bfa 100644
--- a/indra/llcommon/lltimer.cpp
+++ b/indra/llcommon/lltimer.cpp
@@ -209,7 +209,7 @@ F64 calc_clock_frequency(U32 uiMeasureMSecs)
 // Both Linux and Mac use gettimeofday for accurate time
 F64 calc_clock_frequency(unsigned int uiMeasureMSecs)
 {
-	return 1000000.0; // microseconds, so 1 Mhz.
+	return 1000000.0; // microseconds, so 1 MHz.
 }
 
 U64 get_clock_count()
diff --git a/indra/newview/llappviewer.cpp b/indra/newview/llappviewer.cpp
index 78163adf47..319d9b02a9 100644
--- a/indra/newview/llappviewer.cpp
+++ b/indra/newview/llappviewer.cpp
@@ -862,7 +862,7 @@ bool LLAppViewer::init()
 			minSpecs += "\n";
 			unsupported = true;
 		}
-		if(gSysCPU.getMhz() < minCPU)
+		if(gSysCPU.getMHz() < minCPU)
 		{
 			minSpecs += LLNotifications::instance().getGlobalString("UnsupportedCPU");
 			minSpecs += "\n";
@@ -2524,7 +2524,7 @@ void LLAppViewer::writeSystemInfo()
 
 	gDebugInfo["CPUInfo"]["CPUString"] = gSysCPU.getCPUString();
 	gDebugInfo["CPUInfo"]["CPUFamily"] = gSysCPU.getFamily();
-	gDebugInfo["CPUInfo"]["CPUMhz"] = gSysCPU.getMhz();
+	gDebugInfo["CPUInfo"]["CPUMhz"] = (S32)gSysCPU.getMHz();
 	gDebugInfo["CPUInfo"]["CPUAltivec"] = gSysCPU.hasAltivec();
 	gDebugInfo["CPUInfo"]["CPUSSE"] = gSysCPU.hasSSE();
 	gDebugInfo["CPUInfo"]["CPUSSE2"] = gSysCPU.hasSSE2();
diff --git a/indra/newview/llfeaturemanager.cpp b/indra/newview/llfeaturemanager.cpp
index fbb90c69f3..50b08f782a 100644
--- a/indra/newview/llfeaturemanager.cpp
+++ b/indra/newview/llfeaturemanager.cpp
@@ -621,9 +621,9 @@ void LLFeatureManager::applyBaseMasks()
 	
 #if LL_SOLARIS && defined(__sparc) 	//  even low MHz SPARCs are fast
 #error The 800 is hinky. Would something like a LL_MIN_MHZ make more sense here?
-	if (gSysCPU.getMhz() < 800)
+	if (gSysCPU.getMHz() < 800)
 #else
-	if (gSysCPU.getMhz() < 1100)
+	if (gSysCPU.getMHz() < 1100)
 #endif
 	{
 		maskFeatures("CPUSlow");
diff --git a/indra/newview/llviewerjoystick.cpp b/indra/newview/llviewerjoystick.cpp
index b758f6c701..240a539f2e 100644
--- a/indra/newview/llviewerjoystick.cpp
+++ b/indra/newview/llviewerjoystick.cpp
@@ -163,7 +163,7 @@ LLViewerJoystick::LLViewerJoystick()
 	memset(mBtn, 0, sizeof(mBtn));
 
 	// factor in bandwidth? bandwidth = gViewerStats->mKBitStat
-	mPerfScale = 4000.f / gSysCPU.getMhz();
+	mPerfScale = 4000.f / gSysCPU.getMHz(); // hmm.  why?
 }
 
 // -----------------------------------------------------------------------------