Ryujinx/ChocolArm64/Translation/ILEmitterCtx.cs
gdkchan e21ebbf666 Misc. CPU optimizations (#575)
* Add optimizations related to caller/callee saved registers, thread synchronization and disable tier 0

* Refactoring

* Add a config entry to enable or disable the reg load/store opt.

* Remove unnecessary register state stores for calls when the callee is know

* Rename IoType to VarType

* Enable tier 0 while fixing some perf issues related to tier 0

* Small tweak -- Compile before adding to the cache, to avoid lags

* Add required config entry
2019-02-28 13:03:31 +11:00

782 lines
23 KiB
C#

using ChocolArm64.Decoders;
using ChocolArm64.Instructions;
using ChocolArm64.Memory;
using ChocolArm64.State;
using System;
using System.Collections.Generic;
using System.Reflection;
using System.Reflection.Emit;
namespace ChocolArm64.Translation
{
class ILEmitterCtx
{
public MemoryManager Memory { get; }
private TranslatorCache _cache;
private TranslatorQueue _queue;
private Dictionary<long, ILLabel> _labels;
private long _subPosition;
private int _opcIndex;
private Block _currBlock;
public Block CurrBlock => _currBlock;
public OpCode64 CurrOp => _currBlock?.OpCodes[_opcIndex];
public TranslationTier Tier { get; }
public Aarch32Mode Mode { get; } = Aarch32Mode.User; //TODO
public bool HasIndirectJump { get; set; }
public bool HasSlowCall { get; set; }
private Dictionary<Block, ILBlock> _visitedBlocks;
private Queue<Block> _branchTargets;
private List<ILBlock> _ilBlocks;
private ILBlock _ilBlock;
private OpCode64 _optOpLastCompare;
private OpCode64 _optOpLastFlagSet;
//This is the index of the temporary register, used to store temporary
//values needed by some functions, since IL doesn't have a swap instruction.
//You can use any value here as long it doesn't conflict with the indices
//for the other registers. Any value >= 64 or < 0 will do.
private const int ReservedLocalsCount = 64;
private const int RorTmpIndex = ReservedLocalsCount + 0;
private const int CmpOptTmp1Index = ReservedLocalsCount + 1;
private const int CmpOptTmp2Index = ReservedLocalsCount + 2;
private const int IntGpTmp1Index = ReservedLocalsCount + 3;
private const int IntGpTmp2Index = ReservedLocalsCount + 4;
private const int UserIntTempStart = ReservedLocalsCount + 5;
//Vectors are part of another "set" of locals.
private const int VecGpTmp1Index = ReservedLocalsCount + 0;
private const int VecGpTmp2Index = ReservedLocalsCount + 1;
private const int UserVecTempStart = ReservedLocalsCount + 2;
private static int _userIntTempCount;
private static int _userVecTempCount;
public ILEmitterCtx(
MemoryManager memory,
TranslatorCache cache,
TranslatorQueue queue,
TranslationTier tier,
Block graph)
{
Memory = memory ?? throw new ArgumentNullException(nameof(memory));
_cache = cache ?? throw new ArgumentNullException(nameof(cache));
_queue = queue ?? throw new ArgumentNullException(nameof(queue));
_currBlock = graph ?? throw new ArgumentNullException(nameof(graph));
Tier = tier;
_labels = new Dictionary<long, ILLabel>();
_visitedBlocks = new Dictionary<Block, ILBlock>();
_visitedBlocks.Add(graph, new ILBlock());
_branchTargets = new Queue<Block>();
_ilBlocks = new List<ILBlock>();
_subPosition = graph.Position;
ResetBlockState();
if (AdvanceOpCode())
{
EmitSynchronization();
_ilBlock.Add(new ILOpCodeLoadState(_ilBlock, isSubEntry: true));
}
}
public static int GetIntTempIndex()
{
return UserIntTempStart + _userIntTempCount++;
}
public static int GetVecTempIndex()
{
return UserVecTempStart + _userVecTempCount++;
}
public ILBlock[] GetILBlocks()
{
EmitAllOpCodes();
return _ilBlocks.ToArray();
}
private void EmitAllOpCodes()
{
do
{
EmitOpCode();
}
while (AdvanceOpCode());
}
private void EmitOpCode()
{
if (_currBlock == null)
{
return;
}
int opcIndex = _opcIndex;
if (opcIndex == 0)
{
MarkLabel(GetLabel(_currBlock.Position));
}
bool isLastOp = opcIndex == CurrBlock.OpCodes.Count - 1;
if (isLastOp && CurrBlock.Branch != null &&
(ulong)CurrBlock.Branch.Position <= (ulong)CurrBlock.Position)
{
EmitSynchronization();
}
//On AARCH32 mode, (almost) all instruction can be conditionally
//executed, and the required condition is encoded on the opcode.
//We handle that here, skipping the instruction if the condition
//is not met. We can just ignore it when the condition is "Always",
//because in this case the instruction is always going to be executed.
//Condition "Never" is also ignored because this is a special encoding
//used by some unconditional instructions.
ILLabel lblSkip = null;
if (CurrOp is OpCode32 op && op.Cond < Condition.Al)
{
lblSkip = new ILLabel();
EmitCondBranch(lblSkip, GetInverseCond(op.Cond));
}
CurrOp.Emitter(this);
if (lblSkip != null)
{
MarkLabel(lblSkip);
//If this is the last op on the block, and there's no "next" block
//after this one, then we have to return right now, with the address
//of the next instruction to be executed (in the case that the condition
//is false, and the branch was not taken, as all basic blocks should end with
//some kind of branch).
if (isLastOp && CurrBlock.Next == null)
{
EmitStoreState();
EmitLdc_I8(CurrOp.Position + CurrOp.OpCodeSizeInBytes);
Emit(OpCodes.Ret);
}
}
_ilBlock.Add(new ILBarrier());
}
private static Condition GetInverseCond(Condition cond)
{
//Bit 0 of all conditions is basically a negation bit, so
//inverting this bit has the effect of inverting the condition.
return (Condition)((int)cond ^ 1);
}
private void EmitSynchronization()
{
EmitLdarg(TranslatedSub.StateArgIdx);
EmitLdc_I4(_currBlock.OpCodes.Count);
EmitPrivateCall(typeof(CpuThreadState), nameof(CpuThreadState.Synchronize));
EmitLdc_I4(0);
ILLabel lblContinue = new ILLabel();
Emit(OpCodes.Bne_Un_S, lblContinue);
EmitLdc_I8(0);
Emit(OpCodes.Ret);
MarkLabel(lblContinue);
}
private bool AdvanceOpCode()
{
if (_currBlock == null)
{
return false;
}
while (++_opcIndex >= _currBlock.OpCodes.Count)
{
if (!AdvanceBlock())
{
return false;
}
ResetBlockState();
}
return true;
}
private bool AdvanceBlock()
{
if (_currBlock.Branch != null)
{
if (_visitedBlocks.TryAdd(_currBlock.Branch, _ilBlock.Branch))
{
_branchTargets.Enqueue(_currBlock.Branch);
}
}
if (_currBlock.Next != null)
{
if (_visitedBlocks.TryAdd(_currBlock.Next, _ilBlock.Next))
{
_currBlock = _currBlock.Next;
return true;
}
else
{
Emit(OpCodes.Br, GetLabel(_currBlock.Next.Position));
}
}
return _branchTargets.TryDequeue(out _currBlock);
}
private void ResetBlockState()
{
_ilBlock = _visitedBlocks[_currBlock];
_ilBlocks.Add(_ilBlock);
_ilBlock.Next = GetOrCreateILBlock(_currBlock.Next);
_ilBlock.Branch = GetOrCreateILBlock(_currBlock.Branch);
_opcIndex = -1;
_optOpLastFlagSet = null;
_optOpLastCompare = null;
}
private ILBlock GetOrCreateILBlock(Block block)
{
if (block == null)
{
return null;
}
if (_visitedBlocks.TryGetValue(block, out ILBlock ilBlock))
{
return ilBlock;
}
return new ILBlock();
}
public void TranslateAhead(long position, ExecutionMode mode = ExecutionMode.Aarch64)
{
if (_cache.TryGetSubroutine(position, out TranslatedSub sub) && sub.Tier != TranslationTier.Tier0)
{
return;
}
_queue.Enqueue(position, mode, TranslationTier.Tier1, isComplete: true);
}
public bool TryOptEmitSubroutineCall()
{
//Calls should always have a next block, unless
//we're translating a single basic block.
if (_currBlock.Next == null)
{
return false;
}
if (!(CurrOp is IOpCodeBImm op))
{
return false;
}
if (!_cache.TryGetSubroutine(op.Imm, out TranslatedSub sub))
{
return false;
}
//It's not worth to call a Tier0 method, because
//it contains slow code, rather than the entire function.
if (sub.Tier == TranslationTier.Tier0)
{
return false;
}
EmitStoreState(sub);
for (int index = 0; index < TranslatedSub.FixedArgTypes.Length; index++)
{
EmitLdarg(index);
}
EmitCall(sub.Method);
return true;
}
public void TryOptMarkCondWithoutCmp()
{
_optOpLastCompare = CurrOp;
InstEmitAluHelper.EmitAluLoadOpers(this);
Stloc(CmpOptTmp2Index, VarType.Int);
Stloc(CmpOptTmp1Index, VarType.Int);
}
private Dictionary<Condition, OpCode> _branchOps = new Dictionary<Condition, OpCode>()
{
{ Condition.Eq, OpCodes.Beq },
{ Condition.Ne, OpCodes.Bne_Un },
{ Condition.GeUn, OpCodes.Bge_Un },
{ Condition.LtUn, OpCodes.Blt_Un },
{ Condition.GtUn, OpCodes.Bgt_Un },
{ Condition.LeUn, OpCodes.Ble_Un },
{ Condition.Ge, OpCodes.Bge },
{ Condition.Lt, OpCodes.Blt },
{ Condition.Gt, OpCodes.Bgt },
{ Condition.Le, OpCodes.Ble }
};
public void EmitCondBranch(ILLabel target, Condition cond)
{
if (_optOpLastCompare != null &&
_optOpLastCompare == _optOpLastFlagSet && _branchOps.ContainsKey(cond))
{
if (_optOpLastCompare.Emitter == InstEmit.Subs)
{
Ldloc(CmpOptTmp1Index, VarType.Int, _optOpLastCompare.RegisterSize);
Ldloc(CmpOptTmp2Index, VarType.Int, _optOpLastCompare.RegisterSize);
Emit(_branchOps[cond], target);
return;
}
else if (_optOpLastCompare.Emitter == InstEmit.Adds && cond != Condition.GeUn
&& cond != Condition.LtUn
&& cond != Condition.GtUn
&& cond != Condition.LeUn)
{
//There are several limitations that needs to be taken into account for CMN comparisons:
//* The unsigned comparisons are not valid, as they depend on the
//carry flag value, and they will have different values for addition and
//subtraction. For addition, it's carry, and for subtraction, it's borrow.
//So, we need to make sure we're not doing a unsigned compare for the CMN case.
//* We can only do the optimization for the immediate variants,
//because when the second operand value is exactly INT_MIN, we can't
//negate the value as theres no positive counterpart.
//Such invalid values can't be encoded on the immediate encodings.
if (_optOpLastCompare is IOpCodeAluImm64 op)
{
Ldloc(CmpOptTmp1Index, VarType.Int, _optOpLastCompare.RegisterSize);
if (_optOpLastCompare.RegisterSize == RegisterSize.Int32)
{
EmitLdc_I4((int)-op.Imm);
}
else
{
EmitLdc_I8(-op.Imm);
}
Emit(_branchOps[cond], target);
return;
}
}
}
OpCode ilOp;
int intCond = (int)cond;
if (intCond < 14)
{
int condTrue = intCond >> 1;
switch (condTrue)
{
case 0: EmitLdflg((int)PState.ZBit); break;
case 1: EmitLdflg((int)PState.CBit); break;
case 2: EmitLdflg((int)PState.NBit); break;
case 3: EmitLdflg((int)PState.VBit); break;
case 4:
EmitLdflg((int)PState.CBit);
EmitLdflg((int)PState.ZBit);
Emit(OpCodes.Not);
Emit(OpCodes.And);
break;
case 5:
case 6:
EmitLdflg((int)PState.NBit);
EmitLdflg((int)PState.VBit);
Emit(OpCodes.Ceq);
if (condTrue == 6)
{
EmitLdflg((int)PState.ZBit);
Emit(OpCodes.Not);
Emit(OpCodes.And);
}
break;
}
ilOp = (intCond & 1) != 0
? OpCodes.Brfalse
: OpCodes.Brtrue;
}
else
{
ilOp = OpCodes.Br;
}
Emit(ilOp, target);
}
public void EmitCast(IntType intType)
{
switch (intType)
{
case IntType.UInt8: Emit(OpCodes.Conv_U1); break;
case IntType.UInt16: Emit(OpCodes.Conv_U2); break;
case IntType.UInt32: Emit(OpCodes.Conv_U4); break;
case IntType.UInt64: Emit(OpCodes.Conv_U8); break;
case IntType.Int8: Emit(OpCodes.Conv_I1); break;
case IntType.Int16: Emit(OpCodes.Conv_I2); break;
case IntType.Int32: Emit(OpCodes.Conv_I4); break;
case IntType.Int64: Emit(OpCodes.Conv_I8); break;
}
bool sz64 = CurrOp.RegisterSize != RegisterSize.Int32;
if (sz64 == (intType == IntType.UInt64 ||
intType == IntType.Int64))
{
return;
}
if (sz64)
{
Emit(intType >= IntType.Int8
? OpCodes.Conv_I8
: OpCodes.Conv_U8);
}
else
{
Emit(OpCodes.Conv_U4);
}
}
public void EmitLsl(int amount) => EmitILShift(amount, OpCodes.Shl);
public void EmitLsr(int amount) => EmitILShift(amount, OpCodes.Shr_Un);
public void EmitAsr(int amount) => EmitILShift(amount, OpCodes.Shr);
private void EmitILShift(int amount, OpCode ilOp)
{
if (amount > 0)
{
EmitLdc_I4(amount);
Emit(ilOp);
}
}
public void EmitRor(int amount)
{
if (amount > 0)
{
Stloc(RorTmpIndex, VarType.Int);
Ldloc(RorTmpIndex, VarType.Int);
EmitLdc_I4(amount);
Emit(OpCodes.Shr_Un);
Ldloc(RorTmpIndex, VarType.Int);
EmitLdc_I4(CurrOp.GetBitsCount() - amount);
Emit(OpCodes.Shl);
Emit(OpCodes.Or);
}
}
public ILLabel GetLabel(long position)
{
if (!_labels.TryGetValue(position, out ILLabel output))
{
output = new ILLabel();
_labels.Add(position, output);
}
return output;
}
public void MarkLabel(ILLabel label)
{
_ilBlock.Add(label);
}
public void Emit(OpCode ilOp)
{
_ilBlock.Add(new ILOpCode(ilOp));
}
public void Emit(OpCode ilOp, ILLabel label)
{
_ilBlock.Add(new ILOpCodeBranch(ilOp, label));
}
public void EmitFieldLoad(FieldInfo info)
{
_ilBlock.Add(new ILOpCodeLoadField(info));
}
public void EmitPrint(string text)
{
_ilBlock.Add(new ILOpCodeLog(text));
}
public void EmitLdarg(int index)
{
_ilBlock.Add(new ILOpCodeLoad(index, VarType.Arg));
}
public void EmitLdintzr(int index)
{
if (index != RegisterAlias.Zr)
{
EmitLdint(index);
}
else
{
EmitLdc_I(0);
}
}
public void EmitStintzr(int index)
{
if (index != RegisterAlias.Zr)
{
EmitStint(index);
}
else
{
Emit(OpCodes.Pop);
}
}
public void EmitLoadState()
{
if (_ilBlock.Next == null)
{
throw new InvalidOperationException("Can't load state for next block, because there's no next block.");
}
_ilBlock.Add(new ILOpCodeLoadState(_ilBlock.Next));
}
public void EmitStoreState()
{
_ilBlock.Add(new ILOpCodeStoreState(_ilBlock));
}
private void EmitStoreState(TranslatedSub callSub)
{
_ilBlock.Add(new ILOpCodeStoreState(_ilBlock, callSub));
}
public void EmitLdtmp() => EmitLdint(IntGpTmp1Index);
public void EmitSttmp() => EmitStint(IntGpTmp1Index);
public void EmitLdtmp2() => EmitLdint(IntGpTmp2Index);
public void EmitSttmp2() => EmitStint(IntGpTmp2Index);
public void EmitLdvectmp() => EmitLdvec(VecGpTmp1Index);
public void EmitStvectmp() => EmitStvec(VecGpTmp1Index);
public void EmitLdvectmp2() => EmitLdvec(VecGpTmp2Index);
public void EmitStvectmp2() => EmitStvec(VecGpTmp2Index);
public void EmitLdint(int index) => Ldloc(index, VarType.Int);
public void EmitStint(int index) => Stloc(index, VarType.Int);
public void EmitLdvec(int index) => Ldloc(index, VarType.Vector);
public void EmitStvec(int index) => Stloc(index, VarType.Vector);
public void EmitLdflg(int index) => Ldloc(index, VarType.Flag);
public void EmitStflg(int index)
{
//Set this only if any of the NZCV flag bits were modified.
//This is used to ensure that, when emiting a direct IL branch
//instruction for compare + branch sequences, we're not expecting
//to use comparison values from an old instruction, when in fact
//the flags were already overwritten by another instruction further along.
if (index >= (int)PState.VBit)
{
_optOpLastFlagSet = CurrOp;
}
Stloc(index, VarType.Flag);
}
private void Ldloc(int index, VarType varType)
{
_ilBlock.Add(new ILOpCodeLoad(index, varType, CurrOp.RegisterSize));
}
private void Ldloc(int index, VarType varType, RegisterSize registerSize)
{
_ilBlock.Add(new ILOpCodeLoad(index, varType, registerSize));
}
private void Stloc(int index, VarType varType)
{
_ilBlock.Add(new ILOpCodeStore(index, varType, CurrOp.RegisterSize));
}
public void EmitCallPropGet(Type objType, string propName)
{
EmitCall(objType, $"get_{propName}");
}
public void EmitCallPropSet(Type objType, string propName)
{
EmitCall(objType, $"set_{propName}");
}
public void EmitCall(Type objType, string mthdName)
{
if (objType == null)
{
throw new ArgumentNullException(nameof(objType));
}
if (mthdName == null)
{
throw new ArgumentNullException(nameof(mthdName));
}
EmitCall(objType.GetMethod(mthdName));
}
public void EmitCallPrivatePropGet(Type objType, string propName)
{
EmitPrivateCall(objType, $"get_{propName}");
}
public void EmitCallPrivatePropSet(Type objType, string propName)
{
EmitPrivateCall(objType, $"set_{propName}");
}
public void EmitPrivateCall(Type objType, string mthdName)
{
if (objType == null)
{
throw new ArgumentNullException(nameof(objType));
}
if (mthdName == null)
{
throw new ArgumentNullException(nameof(mthdName));
}
EmitCall(objType.GetMethod(mthdName, BindingFlags.Instance | BindingFlags.NonPublic));
}
public void EmitCall(MethodInfo mthdInfo, bool isVirtual = false)
{
_ilBlock.Add(new ILOpCodeCall(mthdInfo ?? throw new ArgumentNullException(nameof(mthdInfo)), isVirtual));
}
public void EmitLdc_I(long value)
{
if (CurrOp.RegisterSize == RegisterSize.Int32)
{
EmitLdc_I4((int)value);
}
else
{
EmitLdc_I8(value);
}
}
public void EmitLdc_I4(int value)
{
_ilBlock.Add(new ILOpCodeConst(value));
}
public void EmitLdc_I8(long value)
{
_ilBlock.Add(new ILOpCodeConst(value));
}
public void EmitLdc_R4(float value)
{
_ilBlock.Add(new ILOpCodeConst(value));
}
public void EmitLdc_R8(double value)
{
_ilBlock.Add(new ILOpCodeConst(value));
}
public void EmitZnFlagCheck()
{
EmitZnCheck(OpCodes.Ceq, (int)PState.ZBit);
EmitZnCheck(OpCodes.Clt, (int)PState.NBit);
}
private void EmitZnCheck(OpCode ilCmpOp, int flag)
{
Emit(OpCodes.Dup);
Emit(OpCodes.Ldc_I4_0);
if (CurrOp.RegisterSize != RegisterSize.Int32)
{
Emit(OpCodes.Conv_I8);
}
Emit(ilCmpOp);
EmitStflg(flag);
}
}
}