Ryujinx/ARMeilleure/Translation/Translator.cs
LDj3SNuD 68f6b79fd3
Add a simple Pools Limiter. (#1830)
* Added support for offline invalidation, via PPTC, of low cq translations replaced by high cq translations; both on a single run and between runs.

Added invalidation of .cache files in the event of reuse on a different user operating system.

Added .info and .cache files invalidation in case of a failed stream decompression.

Nits.

* InternalVersion = 1712;

* Nits.

* Address comment.

* Get rid of BinaryFormatter.

Nits.

* Move Ptc.LoadTranslations().

Nits.

* Nits.

* Fixed corner cases (in case backup copies have to be used). Added save logs.

* Not core fixes.

* Complement to the previous commit. Added load logs. Removed BinaryFormatter leftovers.

* Add LoadTranslations log.

* Nits.

* Removed the search and management of LowCq overlapping functions.

* Final increment of .info and .cache flags.

* Nit.

* Free up memory allocated by Pools during any PPTC translations at boot time.

* Nit due to rebase.

* Add a simple Pools Limiter.

* Nits.

* Fix missing JumpTable.RegisterFunction() due to rebase.

Clear MemoryStreams as soon as possible, when they are no longer needed.

* Code cleaning.

* Nit for retrigger Checks.

* Update Ptc.cs

* Contextual refactoring of Translator. Ignore resetting of pools for DirectCallStubs.

* Nit for retrigger Checks.
2021-01-12 19:04:02 +01:00

452 lines
15 KiB
C#

using ARMeilleure.Decoders;
using ARMeilleure.Diagnostics;
using ARMeilleure.Instructions;
using ARMeilleure.IntermediateRepresentation;
using ARMeilleure.Memory;
using ARMeilleure.State;
using ARMeilleure.Translation.Cache;
using ARMeilleure.Translation.PTC;
using System;
using System.Collections.Concurrent;
using System.Collections.Generic;
using System.Diagnostics;
using System.Linq;
using System.Threading;
using static ARMeilleure.IntermediateRepresentation.OperandHelper;
using static ARMeilleure.IntermediateRepresentation.OperationHelper;
namespace ARMeilleure.Translation
{
public class Translator
{
private readonly IJitMemoryAllocator _allocator;
private readonly IMemoryManager _memory;
private readonly ConcurrentDictionary<ulong, TranslatedFunction> _funcs;
private readonly ConcurrentQueue<KeyValuePair<ulong, IntPtr>> _oldFuncs;
private readonly ConcurrentStack<RejitRequest> _backgroundStack;
private readonly AutoResetEvent _backgroundTranslatorEvent;
private readonly ReaderWriterLock _backgroundTranslatorLock;
private JumpTable _jumpTable;
internal JumpTable JumpTable => _jumpTable;
private volatile int _threadCount;
// FIXME: Remove this once the init logic of the emulator will be redone.
public static ManualResetEvent IsReadyForTranslation = new ManualResetEvent(false);
public Translator(IJitMemoryAllocator allocator, IMemoryManager memory)
{
_allocator = allocator;
_memory = memory;
_funcs = new ConcurrentDictionary<ulong, TranslatedFunction>();
_oldFuncs = new ConcurrentQueue<KeyValuePair<ulong, IntPtr>>();
_backgroundStack = new ConcurrentStack<RejitRequest>();
_backgroundTranslatorEvent = new AutoResetEvent(false);
_backgroundTranslatorLock = new ReaderWriterLock();
JitCache.Initialize(allocator);
DirectCallStubs.InitializeStubs();
}
private void TranslateStackedSubs()
{
while (_threadCount != 0)
{
_backgroundTranslatorLock.AcquireReaderLock(Timeout.Infinite);
if (_backgroundStack.TryPop(out RejitRequest request))
{
TranslatedFunction func = Translate(_memory, _jumpTable, request.Address, request.Mode, highCq: true);
_funcs.AddOrUpdate(request.Address, func, (key, oldFunc) =>
{
EnqueueForDeletion(key, oldFunc);
return func;
});
_jumpTable.RegisterFunction(request.Address, func);
if (PtcProfiler.Enabled)
{
PtcProfiler.UpdateEntry(request.Address, request.Mode, highCq: true);
}
_backgroundTranslatorLock.ReleaseReaderLock();
}
else
{
_backgroundTranslatorLock.ReleaseReaderLock();
_backgroundTranslatorEvent.WaitOne();
}
}
_backgroundTranslatorEvent.Set(); // Wake up any other background translator threads, to encourage them to exit.
}
public void Execute(State.ExecutionContext context, ulong address)
{
if (Interlocked.Increment(ref _threadCount) == 1)
{
IsReadyForTranslation.WaitOne();
Debug.Assert(_jumpTable == null);
_jumpTable = new JumpTable(_allocator);
if (Ptc.State == PtcState.Enabled)
{
Ptc.LoadTranslations(_funcs, _memory, _jumpTable);
Ptc.MakeAndSaveTranslations(_funcs, _memory, _jumpTable);
}
PtcProfiler.Start();
Ptc.Disable();
// Simple heuristic, should be user configurable in future. (1 for 4 core/ht or less, 2 for 6 core+ht etc).
// All threads are normal priority except from the last, which just fills as much of the last core as the os lets it with a low priority.
// If we only have one rejit thread, it should be normal priority as highCq code is performance critical.
// TODO: Use physical cores rather than logical. This only really makes sense for processors with hyperthreading. Requires OS specific code.
int unboundedThreadCount = Math.Max(1, (Environment.ProcessorCount - 6) / 3);
int threadCount = Math.Min(4, unboundedThreadCount);
for (int i = 0; i < threadCount; i++)
{
bool last = i != 0 && i == unboundedThreadCount - 1;
Thread backgroundTranslatorThread = new Thread(TranslateStackedSubs)
{
Name = "CPU.BackgroundTranslatorThread." + i,
Priority = last ? ThreadPriority.Lowest : ThreadPriority.Normal
};
backgroundTranslatorThread.Start();
}
}
Statistics.InitializeTimer();
NativeInterface.RegisterThread(context, _memory, this);
do
{
address = ExecuteSingle(context, address);
}
while (context.Running && address != 0);
NativeInterface.UnregisterThread();
if (Interlocked.Decrement(ref _threadCount) == 0)
{
_backgroundTranslatorEvent.Set();
ClearJitCache();
ResetPools();
_jumpTable.Dispose();
_jumpTable = null;
}
}
public ulong ExecuteSingle(State.ExecutionContext context, ulong address)
{
TranslatedFunction func = GetOrTranslate(address, context.ExecutionMode);
Statistics.StartTimer();
ulong nextAddr = func.Execute(context);
Statistics.StopTimer(address);
return nextAddr;
}
internal TranslatedFunction GetOrTranslate(ulong address, ExecutionMode mode, bool hintRejit = false)
{
if (!_funcs.TryGetValue(address, out TranslatedFunction func))
{
func = Translate(_memory, _jumpTable, address, mode, highCq: false);
TranslatedFunction getFunc = _funcs.GetOrAdd(address, func);
if (getFunc != func)
{
JitCache.Unmap(func.FuncPtr);
func = getFunc;
}
if (PtcProfiler.Enabled)
{
PtcProfiler.AddEntry(address, mode, highCq: false);
}
}
if (hintRejit && func.ShouldRejit())
{
_backgroundStack.Push(new RejitRequest(address, mode));
_backgroundTranslatorEvent.Set();
}
return func;
}
internal static TranslatedFunction Translate(IMemoryManager memory, JumpTable jumpTable, ulong address, ExecutionMode mode, bool highCq)
{
ArmEmitterContext context = new ArmEmitterContext(memory, jumpTable, address, highCq, Aarch32Mode.User);
Logger.StartPass(PassName.Decoding);
Block[] blocks = Decoder.Decode(memory, address, mode, highCq, singleBlock: false);
Logger.EndPass(PassName.Decoding);
PreparePool(highCq);
Logger.StartPass(PassName.Translation);
EmitSynchronization(context);
if (blocks[0].Address != address)
{
context.Branch(context.GetLabel(address));
}
ControlFlowGraph cfg = EmitAndGetCFG(context, blocks, out Range funcRange);
ulong funcSize = funcRange.End - funcRange.Start;
Logger.EndPass(PassName.Translation);
Logger.StartPass(PassName.RegisterUsage);
RegisterUsage.RunPass(cfg, mode);
Logger.EndPass(PassName.RegisterUsage);
OperandType[] argTypes = new OperandType[] { OperandType.I64 };
CompilerOptions options = highCq ? CompilerOptions.HighCq : CompilerOptions.None;
GuestFunction func;
if (Ptc.State == PtcState.Disabled)
{
func = Compiler.Compile<GuestFunction>(cfg, argTypes, OperandType.I64, options);
ReturnPool(highCq);
}
else using (PtcInfo ptcInfo = new PtcInfo())
{
func = Compiler.Compile<GuestFunction>(cfg, argTypes, OperandType.I64, options, ptcInfo);
ReturnPool(highCq);
Ptc.WriteInfoCodeRelocUnwindInfo(address, funcSize, highCq, ptcInfo);
}
return new TranslatedFunction(func, funcSize, highCq);
}
internal static void PreparePool(bool highCq)
{
PrepareOperandPool(highCq);
PrepareOperationPool(highCq);
}
internal static void ReturnPool(bool highCq)
{
ReturnOperandPool(highCq);
ReturnOperationPool(highCq);
}
internal static void ResetPools()
{
ResetOperandPools();
ResetOperationPools();
}
private struct Range
{
public ulong Start { get; }
public ulong End { get; }
public Range(ulong start, ulong end)
{
Start = start;
End = end;
}
}
private static ControlFlowGraph EmitAndGetCFG(ArmEmitterContext context, Block[] blocks, out Range range)
{
ulong rangeStart = ulong.MaxValue;
ulong rangeEnd = 0;
for (int blkIndex = 0; blkIndex < blocks.Length; blkIndex++)
{
Block block = blocks[blkIndex];
if (!block.Exit)
{
if (rangeStart > block.Address)
{
rangeStart = block.Address;
}
if (rangeEnd < block.EndAddress)
{
rangeEnd = block.EndAddress;
}
}
context.CurrBlock = block;
context.MarkLabel(context.GetLabel(block.Address));
if (block.Exit)
{
InstEmitFlowHelper.EmitTailContinue(context, Const(block.Address), block.TailCall);
}
else
{
for (int opcIndex = 0; opcIndex < block.OpCodes.Count; opcIndex++)
{
OpCode opCode = block.OpCodes[opcIndex];
context.CurrOp = opCode;
bool isLastOp = opcIndex == block.OpCodes.Count - 1;
if (isLastOp && block.Branch != null && !block.Branch.Exit && block.Branch.Address <= block.Address)
{
EmitSynchronization(context);
}
Operand lblPredicateSkip = null;
if (opCode is OpCode32 op && op.Cond < Condition.Al)
{
lblPredicateSkip = Label();
InstEmitFlowHelper.EmitCondBranch(context, lblPredicateSkip, op.Cond.Invert());
}
if (opCode.Instruction.Emitter != null)
{
opCode.Instruction.Emitter(context);
}
else
{
throw new InvalidOperationException($"Invalid instruction \"{opCode.Instruction.Name}\".");
}
if (lblPredicateSkip != null)
{
context.MarkLabel(lblPredicateSkip);
}
}
}
}
range = new Range(rangeStart, rangeEnd);
return context.GetControlFlowGraph();
}
internal static void EmitSynchronization(EmitterContext context)
{
long countOffs = NativeContext.GetCounterOffset();
Operand countAddr = context.Add(context.LoadArgument(OperandType.I64, 0), Const(countOffs));
Operand count = context.Load(OperandType.I32, countAddr);
Operand lblNonZero = Label();
Operand lblExit = Label();
context.BranchIfTrue(lblNonZero, count, BasicBlockFrequency.Cold);
Operand running = context.Call(typeof(NativeInterface).GetMethod(nameof(NativeInterface.CheckSynchronization)));
context.BranchIfTrue(lblExit, running, BasicBlockFrequency.Cold);
context.Return(Const(0L));
context.MarkLabel(lblNonZero);
count = context.Subtract(count, Const(1));
context.Store(countAddr, count);
context.MarkLabel(lblExit);
}
public void InvalidateJitCacheRegion(ulong address, ulong size)
{
static bool OverlapsWith(ulong funcAddress, ulong funcSize, ulong address, ulong size)
{
return funcAddress < address + size && address < funcAddress + funcSize;
}
// Make a copy of all overlapping functions, as we can't otherwise
// remove elements from the collection we are iterating.
// Doing that before clearing the rejit queue is fine, even
// if a function is translated after this, it would only replace
// a existing function, as rejit is only triggered on functions
// that were already executed before.
var toDelete = _funcs.Where(x => OverlapsWith(x.Key, x.Value.GuestSize, address, size)).ToArray();
if (toDelete.Length != 0)
{
// If rejit is running, stop it as it may be trying to rejit the functions we are
// supposed to remove.
ClearRejitQueue();
}
foreach (var kv in toDelete)
{
if (_funcs.TryRemove(kv.Key, out TranslatedFunction func))
{
EnqueueForDeletion(kv.Key, func);
}
}
}
private void EnqueueForDeletion(ulong guestAddress, TranslatedFunction func)
{
_oldFuncs.Enqueue(new KeyValuePair<ulong, IntPtr>(guestAddress, func.FuncPtr));
}
private void ClearJitCache()
{
// Ensure no attempt will be made to compile new functions due to rejit.
ClearRejitQueue();
foreach (var kv in _funcs)
{
JitCache.Unmap(kv.Value.FuncPtr);
}
_funcs.Clear();
while (_oldFuncs.TryDequeue(out var kv))
{
JitCache.Unmap(kv.Value);
}
}
private void ClearRejitQueue()
{
_backgroundTranslatorLock.AcquireWriterLock(Timeout.Infinite);
_backgroundStack.Clear();
_backgroundTranslatorLock.ReleaseWriterLock();
}
}
}