///////////////////////////////////////////////////////////////////////////////
//
//  Microsoft Research Singularity
//
//  Copyright (c) Microsoft Corporation.  All rights reserved.
//
//  File:   PerfSnap.sg
//
//  Note:   Performance measurer for Singularity Benchmark
//

using System;
using Microsoft.Singularity.V1.Services;

namespace Microsoft.Singularity.Applications
{
    public struct PerfSnap
    {
        // Global options for all instances
        static bool xmlOutput = false;
        static bool atRing3   = false;

        private bool disabled;
        private long begCycleCount;
        private long endCycleCount;
        private long begSwitchCount;
        private long endSwitchCount;
        private long begInterruptCount;
        private long endInterruptCount;
        private long begKernelGcCount;
        private long endKernelGcCount;
        private long begProcessGcCount;
        private long endProcessGcCount;
        private long iterations;
        private ulong begAllocatedCount;
        private ulong begAllocatedBytes;
        private ulong begFreedCount;
        private ulong begFreedBytes;
        private ulong endAllocatedCount;
        private ulong endAllocatedBytes;
        private ulong endFreedCount;
        private ulong endFreedBytes;
        private ulong begStackGets;
        private ulong begStackRets;
        private ulong endStackGets;
        private ulong endStackRets;


        public long Cycles      { get { return endCycleCount - begCycleCount; } }
        public long Interrupts  { get { return endInterruptCount - begInterruptCount; } }
        public long Switches    { get { return endSwitchCount - begSwitchCount; } }
        public long KernelGCs   { get { return endKernelGcCount - begKernelGcCount; } }
        public long ProcessGCs  { get { return endProcessGcCount - begProcessGcCount; } }
        public ulong AllocatedCount { get { return endAllocatedCount-begAllocatedCount; } }
        public ulong AllocatedBytes { get { return endAllocatedBytes-begAllocatedBytes; } }
        public ulong FreedCount { get { return endFreedCount - begFreedCount; } }
        public ulong FreedBytes { get { return endFreedBytes - begFreedBytes; } }
        public ulong StackGets { get { return endStackGets - begStackGets; } }
        public ulong StackRets { get { return endStackRets - begStackRets; } }

        public void Start()
        {
            if (!atRing3) {
                disabled = Processor.DisableInterrupts();
            }

            int collectorCount;
            long collectorMillis;
            long collectorBytes;
            GC.PerformanceCounters(out collectorCount,
                                   out collectorMillis,
                                   out collectorBytes);

            ulong stackGets;
            ulong stackRets;
            StackService.GetUsageStatistics(out stackGets,
                                            out stackRets);
            begStackGets = stackGets;
            begStackRets = stackRets;

            ulong allocatedCount;
            ulong allocatedBytes;
            ulong freedCount;
            ulong freedBytes;
            PageTableService.GetUsageStatistics(out allocatedCount,
                                                out allocatedBytes,
                                                out freedCount,
                                                out freedBytes);
            begAllocatedCount = allocatedCount;
            begAllocatedBytes = allocatedBytes;
            begFreedCount = freedCount;
            begFreedBytes = freedBytes;

            begInterruptCount = ProcessService.GetKernelInterruptCount();
            begSwitchCount = ProcessService.GetContextSwitchCount();
            begKernelGcCount = ProcessService.GetKernelGcCount();
            begProcessGcCount = collectorCount;

#if x64_PERF
            // Set up for perf counting
            if (!atRing3) {
                // Reset the performance counters to what we're interested in.
                Reset(0, PerfEvtSel.COUNT | PerfEvtSel.CyclesNotHalted);
                Reset(1, PerfEvtSel.COUNT | PerfEvtSel.RetiredInstructions);
                Reset(2, PerfEvtSel.COUNT | PerfEvtSel.RetiredBranchInstructions);
                Reset(3, PerfEvtSel.COUNT | PerfEvtSel.RequestsToL2Cache | 0x400);
            } else {
                // We're not allowed to reset the perf counters, so take note
                // of their current values; we will subtract from this later.
                x64_i0 = Processor.ReadPmc(0);
                x64_i1 = Processor.ReadPmc(1);
                x64_i2 = Processor.ReadPmc(2);
                x64_i3 = Processor.ReadPmc(3);
            }
#endif

            begCycleCount = unchecked((long)Processor.CycleCount);
        }

        public void Finish(long iterations)
        {
            endCycleCount = unchecked((long)Processor.CycleCount);

#if X64_PERF
            x64_p0 = Processor.ReadPmc(0);
            x64_p1 = Processor.ReadPmc(1);
            x64_p2 = Processor.ReadPmc(2);
            x64_p3 = Processor.ReadPmc(3);
#endif

            endInterruptCount = ProcessService.GetKernelInterruptCount();
            endSwitchCount = ProcessService.GetContextSwitchCount();
            endKernelGcCount = ProcessService.GetKernelGcCount();

            int collectorCount;
            long collectorMillis;
            long collectorBytes;
            GC.PerformanceCounters(out collectorCount,
                                   out collectorMillis,
                                   out collectorBytes);
            endProcessGcCount = collectorCount;

            ulong allocatedCount;
            ulong allocatedBytes;
            ulong freedCount;
            ulong freedBytes;
            PageTableService.GetUsageStatistics(out allocatedCount,
                                                out allocatedBytes,
                                                out freedCount,
                                                out freedBytes);
            endAllocatedCount = allocatedCount;
            endAllocatedBytes = allocatedBytes;
            endFreedCount = freedCount;
            endFreedBytes = freedBytes;

            ulong stackGets;
            ulong stackRets;
            StackService.GetUsageStatistics(out stackGets,
                                            out stackRets);
            endStackGets = stackGets;
            endStackRets = stackRets;

            if (!atRing3) {
                Processor.RestoreInterrupts(disabled);
            }

            this.iterations = iterations;
        }

        public void Display(string name)
        {
            if (xmlOutput) {
                DisplayXml(name);
            } else {
                DisplayText(name);
            }
        }

        private static void WriteXmlValue(string name, long value)
        {
            DualWriteLine(
                String.Format(" <Measurement> <name> {0} </name> <value> {1:d} </value> </Measurement>", name, value)
                );
        }

        private static void WriteXmlValue(string name, ulong value)
        {
            DualWriteLine(
                String.Format(" <Measurement> <name> {0} </name> <value> {1:d} </value> </Measurement>", name, value)
                );
        }

        private void DisplayXml(string name)
        {
            DualWriteLine(
                String.Format("<Microbenchmark Name=\"{0}\">", name)
                );
            WriteXmlValue("CyclesPerIteration", Cycles / iterations);
            WriteXmlValue("Cycles", Cycles);
            WriteXmlValue("Iterations", iterations);
            WriteXmlValue("Switches", Switches);
            WriteXmlValue("Interrupts", Interrupts);
            WriteXmlValue("KernelGCs", KernelGCs);
            WriteXmlValue("ProcessGCs", ProcessGCs);
            WriteXmlValue("AllocationCount", AllocatedCount);
            WriteXmlValue("AllocationBytes", AllocatedBytes);
            WriteXmlValue("FreedCount", FreedCount);
            WriteXmlValue("FreedBytes", FreedBytes);
            WriteXmlValue("StackGetCount", StackGets);
            WriteXmlValue("StackReturnCount", StackRets);
            DualWriteLine("</Microbenchmark>");
        }

        private void DisplayText(string name)
        {
            DualWriteLine(
                String.Format("{0,-16} {1,6:d} x{2,8:d} ={3,12:d} " +
                              "[swi={4,6:d} int={5,3:d} gc={6:d}/{7:d}]",
                              name,
                              iterations,
                              Cycles / iterations,
                              Cycles,
                              Switches,
                              Interrupts,
                              KernelGCs,
                              ProcessGCs)
                );

            if (AllocatedCount > 1 || FreedCount > 1 || StackGets > 1) {
                DualWriteLine(
                    string.Format(
                        "                       " +
                        "[alloc={0,4:d}/{1,8:x} free={2,4:d}/{3,8:x} " +
                        "stack={4,4:d}/{5,4:d}]",
                        AllocatedCount,
                        AllocatedBytes,
                        FreedCount,
                        FreedBytes,
                        StackGets,
                        StackRets)
                    );
            }

#if x64_PERF
            if (!atRing3) {
                // Read off the current MSR values and turn them
                // into nice labels
                ulong e0 = Processor.ReadMsr(0xc0010000);
                ulong e1 = Processor.ReadMsr(0xc0010001);
                ulong e2 = Processor.ReadMsr(0xc0010002);
                ulong e3 = Processor.ReadMsr(0xc0010003);

                DualWriteLine(
                    String.Format("evt: {0,16} {1,16} {2,16} {3,16}",
                                  EvtSelToString(e0),
                                  EvtSelToString(e1),
                                  EvtSelToString(e2),
                                  EvtSelToString(e3)));
            } else {
                // Subtract from the initial perf-counter values to
                // get the delta we want
                x64_p0 -= x64_i0;
                x64_p1 -= x64_i1;
                x64_p2 -= x64_i2;
                x64_p3 -= x64_i3;
            }

            DualWriteLine(
                String.Format("pmc: {0:d16} {1:d16} {2:d16} {3:d16} {4:d16}\n\n",
                              Cycles, x64_p0, x64_p1, x64_p2, x64_p3));
#endif
        }

        private static void DualWriteLine(string message)
        {
            Console.WriteLine(message);
            DebugStub.WriteLine(message);
        }

        public static void SetOptions(bool ring3, bool doXmlOutput)
        {
            PerfSnap.atRing3   = ring3;
            PerfSnap.xmlOutput = doXmlOutput;
        }
    }
}