RyuKen/ARMeilleure/Instructions/InstEmitSimdCmp32.cs
FICTURE7 22b2cb39af
Reduce JIT GC allocations (#2515)
* Turn `MemoryOperand` into a struct

* Remove `IntrinsicOperation`

* Remove `PhiNode`

* Remove `Node`

* Turn `Operand` into a struct

* Turn `Operation` into a struct

* Clean up pool management methods

* Add `Arena` allocator

* Move `OperationHelper` to `Operation.Factory`

* Move `OperandHelper` to `Operand.Factory`

* Optimize `Operation` a bit

* Fix `Arena` initialization

* Rename `NativeList<T>` to `ArenaList<T>`

* Reduce `Operand` size from 88 to 56 bytes

* Reduce `Operation` size from 56 to 40 bytes

* Add optimistic interning of Register & Constant operands

* Optimize `RegisterUsage` pass a bit

* Optimize `RemoveUnusedNodes` pass a bit

Iterating in reverse-order allows killing dependency chains in a single
pass.

* Fix PPTC symbols

* Optimize `BasicBlock` a bit

Reduce allocations from `_successor` & `DominanceFrontiers`

* Fix `Operation` resize

* Make `Arena` expandable

Change the arena allocator to be expandable by allocating in pages, with
some of them being pooled. Currently 32 pages are pooled. An LRU removal
mechanism should probably be added to it.

Apparently MHR can allocate bitmaps large enough to exceed the 16MB
limit for the type.

* Move `Arena` & `ArenaList` to `Common`

* Remove `ThreadStaticPool` & co

* Add `PhiOperation`

* Reduce `Operand` size from 56 from 48 bytes

* Add linear-probing to `Operand` intern table

* Optimize `HybridAllocator` a bit

* Add `Allocators` class

* Tune `ArenaAllocator` sizes

* Add page removal mechanism to `ArenaAllocator`

Remove pages which have not been used for more than 5s after each reset.

I am on fence if this would be better using a Gen2 callback object like
the one in System.Buffers.ArrayPool<T>, to trim the pool. Because right
now if a large translation happens, the pages will be freed only after a
reset. This reset may not happen for a while because no new translation
is hit, but the arena base sizes are rather small.

* Fix `OOM` when allocating larger than page size in `ArenaAllocator`

Tweak resizing mechanism for Operand.Uses and Assignemnts.

* Optimize `Optimizer` a bit

* Optimize `Operand.Add<T>/Remove<T>` a bit

* Clean up `PreAllocator`

* Fix phi insertion order

Reduce codegen diffs.

* Fix code alignment

* Use new heuristics for degree of parallelism

* Suppress warnings

* Address gdkchan's feedback

Renamed `GetValue()` to `GetValueUnsafe()` to make it more clear that
`Operand.Value` should usually not be modified directly.

* Add fast path to `ArenaAllocator`

* Assembly for `ArenaAllocator.Allocate(ulong)`:

  .L0:
    mov rax, [rcx+0x18]
    lea r8, [rax+rdx]
    cmp r8, [rcx+0x10]
    ja short .L2
  .L1:
    mov rdx, [rcx+8]
    add rax, [rdx+8]
    mov [rcx+0x18], r8
    ret
  .L2:
    jmp ArenaAllocator.AllocateSlow(UInt64)

  A few variable/field had to be changed to ulong so that RyuJIT avoids
  emitting zero-extends.

* Implement a new heuristic to free pooled pages.

  If an arena is used often, it is more likely that its pages will be
  needed, so the pages are kept for longer (e.g: during PPTC rebuild or
  burst sof compilations). If is not used often, then it is more likely
  that its pages will not be needed (e.g: after PPTC rebuild or bursts
  of compilations).

* Address riperiperi's feedback

* Use `EqualityComparer<T>` in `IntrusiveList<T>`

Avoids a potential GC hole in `Equals(T, T)`.
2021-08-17 15:08:34 -03:00

413 lines
14 KiB
C#

using ARMeilleure.Decoders;
using ARMeilleure.IntermediateRepresentation;
using ARMeilleure.State;
using ARMeilleure.Translation;
using System;
using System.Reflection;
using static ARMeilleure.Instructions.InstEmitHelper;
using static ARMeilleure.Instructions.InstEmitSimdHelper;
using static ARMeilleure.Instructions.InstEmitSimdHelper32;
using static ARMeilleure.IntermediateRepresentation.Operand.Factory;
namespace ARMeilleure.Instructions
{
using Func2I = Func<Operand, Operand, Operand>;
static partial class InstEmit32
{
public static void Vceq_V(ArmEmitterContext context)
{
if (Optimizations.FastFP && Optimizations.UseSse2)
{
EmitSse2OrAvxCmpOpF32(context, CmpCondition.Equal, false);
}
else
{
EmitCmpOpF32(context, nameof(SoftFloat32.FPCompareEQFpscr), false);
}
}
public static void Vceq_I(ArmEmitterContext context)
{
EmitCmpOpI32(context, context.ICompareEqual, context.ICompareEqual, false, false);
}
public static void Vceq_Z(ArmEmitterContext context)
{
OpCode32Simd op = (OpCode32Simd)context.CurrOp;
if (op.F)
{
if (Optimizations.FastFP && Optimizations.UseSse2)
{
EmitSse2OrAvxCmpOpF32(context, CmpCondition.Equal, true);
}
else
{
EmitCmpOpF32(context, nameof(SoftFloat32.FPCompareEQFpscr), true);
}
}
else
{
EmitCmpOpI32(context, context.ICompareEqual, context.ICompareEqual, true, false);
}
}
public static void Vcge_V(ArmEmitterContext context)
{
if (Optimizations.FastFP && Optimizations.UseAvx)
{
EmitSse2OrAvxCmpOpF32(context, CmpCondition.GreaterThanOrEqual, false);
}
else
{
EmitCmpOpF32(context, nameof(SoftFloat32.FPCompareGEFpscr), false);
}
}
public static void Vcge_I(ArmEmitterContext context)
{
OpCode32SimdReg op = (OpCode32SimdReg)context.CurrOp;
EmitCmpOpI32(context, context.ICompareGreaterOrEqual, context.ICompareGreaterOrEqualUI, false, !op.U);
}
public static void Vcge_Z(ArmEmitterContext context)
{
OpCode32Simd op = (OpCode32Simd)context.CurrOp;
if (op.F)
{
if (Optimizations.FastFP && Optimizations.UseAvx)
{
EmitSse2OrAvxCmpOpF32(context, CmpCondition.GreaterThanOrEqual, true);
}
else
{
EmitCmpOpF32(context, nameof(SoftFloat32.FPCompareGEFpscr), true);
}
}
else
{
EmitCmpOpI32(context, context.ICompareGreaterOrEqual, context.ICompareGreaterOrEqualUI, true, true);
}
}
public static void Vcgt_V(ArmEmitterContext context)
{
if (Optimizations.FastFP && Optimizations.UseAvx)
{
EmitSse2OrAvxCmpOpF32(context, CmpCondition.GreaterThan, false);
}
else
{
EmitCmpOpF32(context, nameof(SoftFloat32.FPCompareGTFpscr), false);
}
}
public static void Vcgt_I(ArmEmitterContext context)
{
OpCode32SimdReg op = (OpCode32SimdReg)context.CurrOp;
EmitCmpOpI32(context, context.ICompareGreater, context.ICompareGreaterUI, false, !op.U);
}
public static void Vcgt_Z(ArmEmitterContext context)
{
OpCode32Simd op = (OpCode32Simd)context.CurrOp;
if (op.F)
{
if (Optimizations.FastFP && Optimizations.UseAvx)
{
EmitSse2OrAvxCmpOpF32(context, CmpCondition.GreaterThan, true);
}
else
{
EmitCmpOpF32(context, nameof(SoftFloat32.FPCompareGTFpscr), true);
}
}
else
{
EmitCmpOpI32(context, context.ICompareGreater, context.ICompareGreaterUI, true, true);
}
}
public static void Vcle_Z(ArmEmitterContext context)
{
OpCode32Simd op = (OpCode32Simd)context.CurrOp;
if (op.F)
{
if (Optimizations.FastFP && Optimizations.UseSse2)
{
EmitSse2OrAvxCmpOpF32(context, CmpCondition.LessThanOrEqual, true);
}
else
{
EmitCmpOpF32(context, nameof(SoftFloat32.FPCompareLEFpscr), true);
}
}
else
{
EmitCmpOpI32(context, context.ICompareLessOrEqual, context.ICompareLessOrEqualUI, true, true);
}
}
public static void Vclt_Z(ArmEmitterContext context)
{
OpCode32Simd op = (OpCode32Simd)context.CurrOp;
if (op.F)
{
if (Optimizations.FastFP && Optimizations.UseSse2)
{
EmitSse2OrAvxCmpOpF32(context, CmpCondition.LessThan, true);
}
else
{
EmitCmpOpF32(context, nameof(SoftFloat32.FPCompareLTFpscr), true);
}
}
else
{
EmitCmpOpI32(context, context.ICompareLess, context.ICompareLessUI, true, true);
}
}
private static void EmitCmpOpF32(ArmEmitterContext context, string name, bool zero)
{
Operand one = Const(1);
if (zero)
{
EmitVectorUnaryOpF32(context, (m) =>
{
OperandType type = m.Type;
if (type == OperandType.FP64)
{
return context.Call(typeof(SoftFloat64).GetMethod(name), m, ConstF(0.0d), one);
}
else
{
return context.Call(typeof(SoftFloat32).GetMethod(name), m, ConstF(0.0f), one);
}
});
}
else
{
EmitVectorBinaryOpF32(context, (n, m) =>
{
OperandType type = n.Type;
if (type == OperandType.FP64)
{
return context.Call(typeof(SoftFloat64).GetMethod(name), n, m, one);
}
else
{
return context.Call(typeof(SoftFloat32).GetMethod(name), n, m, one);
}
});
}
}
private static Operand ZerosOrOnes(ArmEmitterContext context, Operand fromBool, OperandType baseType)
{
var ones = (baseType == OperandType.I64) ? Const(-1L) : Const(-1);
return context.ConditionalSelect(fromBool, ones, Const(baseType, 0L));
}
private static void EmitCmpOpI32(
ArmEmitterContext context,
Func2I signedOp,
Func2I unsignedOp,
bool zero,
bool signed)
{
if (zero)
{
if (signed)
{
EmitVectorUnaryOpSx32(context, (m) =>
{
OperandType type = m.Type;
Operand zeroV = (type == OperandType.I64) ? Const(0L) : Const(0);
return ZerosOrOnes(context, signedOp(m, zeroV), type);
});
}
else
{
EmitVectorUnaryOpZx32(context, (m) =>
{
OperandType type = m.Type;
Operand zeroV = (type == OperandType.I64) ? Const(0L) : Const(0);
return ZerosOrOnes(context, unsignedOp(m, zeroV), type);
});
}
}
else
{
if (signed)
{
EmitVectorBinaryOpSx32(context, (n, m) => ZerosOrOnes(context, signedOp(n, m), n.Type));
}
else
{
EmitVectorBinaryOpZx32(context, (n, m) => ZerosOrOnes(context, unsignedOp(n, m), n.Type));
}
}
}
public static void Vcmp(ArmEmitterContext context)
{
EmitVcmpOrVcmpe(context, false);
}
public static void Vcmpe(ArmEmitterContext context)
{
EmitVcmpOrVcmpe(context, true);
}
private static void EmitVcmpOrVcmpe(ArmEmitterContext context, bool signalNaNs)
{
OpCode32SimdS op = (OpCode32SimdS)context.CurrOp;
bool cmpWithZero = (op.Opc & 2) != 0;
int sizeF = op.Size & 1;
if (Optimizations.FastFP && (signalNaNs ? Optimizations.UseAvx : Optimizations.UseSse2))
{
CmpCondition cmpOrdered = signalNaNs ? CmpCondition.OrderedS : CmpCondition.OrderedQ;
bool doubleSize = sizeF != 0;
int shift = doubleSize ? 1 : 2;
Operand m = GetVecA32(op.Vm >> shift);
Operand n = GetVecA32(op.Vd >> shift);
n = EmitSwapScalar(context, n, op.Vd, doubleSize);
m = cmpWithZero ? context.VectorZero() : EmitSwapScalar(context, m, op.Vm, doubleSize);
Operand lblNaN = Label();
Operand lblEnd = Label();
if (!doubleSize)
{
Operand ordMask = context.AddIntrinsic(Intrinsic.X86Cmpss, n, m, Const((int)cmpOrdered));
Operand isOrdered = context.AddIntrinsicInt(Intrinsic.X86Cvtsi2si, ordMask);
context.BranchIfFalse(lblNaN, isOrdered);
Operand cf = context.AddIntrinsicInt(Intrinsic.X86Comissge, n, m);
Operand zf = context.AddIntrinsicInt(Intrinsic.X86Comisseq, n, m);
Operand nf = context.AddIntrinsicInt(Intrinsic.X86Comisslt, n, m);
SetFpFlag(context, FPState.VFlag, Const(0));
SetFpFlag(context, FPState.CFlag, cf);
SetFpFlag(context, FPState.ZFlag, zf);
SetFpFlag(context, FPState.NFlag, nf);
}
else
{
Operand ordMask = context.AddIntrinsic(Intrinsic.X86Cmpsd, n, m, Const((int)cmpOrdered));
Operand isOrdered = context.AddIntrinsicLong(Intrinsic.X86Cvtsi2si, ordMask);
context.BranchIfFalse(lblNaN, isOrdered);
Operand cf = context.AddIntrinsicInt(Intrinsic.X86Comisdge, n, m);
Operand zf = context.AddIntrinsicInt(Intrinsic.X86Comisdeq, n, m);
Operand nf = context.AddIntrinsicInt(Intrinsic.X86Comisdlt, n, m);
SetFpFlag(context, FPState.VFlag, Const(0));
SetFpFlag(context, FPState.CFlag, cf);
SetFpFlag(context, FPState.ZFlag, zf);
SetFpFlag(context, FPState.NFlag, nf);
}
context.Branch(lblEnd);
context.MarkLabel(lblNaN);
SetFpFlag(context, FPState.VFlag, Const(1));
SetFpFlag(context, FPState.CFlag, Const(1));
SetFpFlag(context, FPState.ZFlag, Const(0));
SetFpFlag(context, FPState.NFlag, Const(0));
context.MarkLabel(lblEnd);
}
else
{
OperandType type = sizeF != 0 ? OperandType.FP64 : OperandType.FP32;
Operand ne = ExtractScalar(context, type, op.Vd);
Operand me;
if (cmpWithZero)
{
me = sizeF == 0 ? ConstF(0f) : ConstF(0d);
}
else
{
me = ExtractScalar(context, type, op.Vm);
}
MethodInfo info = sizeF != 0
? typeof(SoftFloat64).GetMethod(nameof(SoftFloat64.FPCompare))
: typeof(SoftFloat32).GetMethod(nameof(SoftFloat32.FPCompare));
Operand nzcv = context.Call(info, ne, me, Const(signalNaNs));
EmitSetFpscrNzcv(context, nzcv);
}
}
private static void EmitSetFpscrNzcv(ArmEmitterContext context, Operand nzcv)
{
Operand Extract(Operand value, int bit)
{
if (bit != 0)
{
value = context.ShiftRightUI(value, Const(bit));
}
value = context.BitwiseAnd(value, Const(1));
return value;
}
SetFpFlag(context, FPState.VFlag, Extract(nzcv, 0));
SetFpFlag(context, FPState.CFlag, Extract(nzcv, 1));
SetFpFlag(context, FPState.ZFlag, Extract(nzcv, 2));
SetFpFlag(context, FPState.NFlag, Extract(nzcv, 3));
}
private static void EmitSse2OrAvxCmpOpF32(ArmEmitterContext context, CmpCondition cond, bool zero)
{
OpCode32Simd op = (OpCode32Simd)context.CurrOp;
int sizeF = op.Size & 1;
Intrinsic inst = (sizeF == 0) ? Intrinsic.X86Cmpps : Intrinsic.X86Cmppd;
if (zero)
{
EmitVectorUnaryOpSimd32(context, (m) =>
{
return context.AddIntrinsic(inst, m, context.VectorZero(), Const((int)cond));
});
}
else
{
EmitVectorBinaryOpSimd32(context, (n, m) =>
{
return context.AddIntrinsic(inst, n, m, Const((int)cond));
});
}
}
}
}