Ryujinx/ARMeilleure/CodeGen/X86/X86Optimizer.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

255 lines
8.5 KiB
C#

using ARMeilleure.CodeGen.Optimizations;
using ARMeilleure.IntermediateRepresentation;
using ARMeilleure.Translation;
using System.Collections.Generic;
using static ARMeilleure.IntermediateRepresentation.Operand.Factory;
using static ARMeilleure.IntermediateRepresentation.Operation.Factory;
namespace ARMeilleure.CodeGen.X86
{
static class X86Optimizer
{
public static void RunPass(ControlFlowGraph cfg)
{
var constants = new Dictionary<ulong, Operand>();
Operand GetConstantCopy(BasicBlock block, Operation operation, Operand source)
{
if (!constants.TryGetValue(source.Value, out var constant))
{
constant = Local(source.Type);
Operation copyOp = Operation(Instruction.Copy, constant, source);
block.Operations.AddBefore(operation, copyOp);
constants.Add(source.Value, constant);
}
return constant;
}
for (BasicBlock block = cfg.Blocks.First; block != null; block = block.ListNext)
{
constants.Clear();
Operation nextNode;
for (Operation node = block.Operations.First; node != default; node = nextNode)
{
nextNode = node.ListNext;
// Insert copies for constants that can't fit on a 32-bits immediate.
// Doing this early unblocks a few optimizations.
if (node.Instruction == Instruction.Add)
{
Operand src1 = node.GetSource(0);
Operand src2 = node.GetSource(1);
if (src1.Kind == OperandKind.Constant && (src1.Relocatable || CodeGenCommon.IsLongConst(src1)))
{
node.SetSource(0, GetConstantCopy(block, node, src1));
}
if (src2.Kind == OperandKind.Constant && (src2.Relocatable || CodeGenCommon.IsLongConst(src2)))
{
node.SetSource(1, GetConstantCopy(block, node, src2));
}
}
// Try to fold something like:
// shl rbx, 2
// add rax, rbx
// add rax, 0xcafe
// mov rax, [rax]
// Into:
// mov rax, [rax+rbx*4+0xcafe]
if (IsMemoryLoadOrStore(node.Instruction))
{
OperandType type;
if (node.Destination != default)
{
type = node.Destination.Type;
}
else
{
type = node.GetSource(1).Type;
}
Operand memOp = GetMemoryOperandOrNull(node.GetSource(0), type);
if (memOp != default)
{
node.SetSource(0, memOp);
}
}
}
}
Optimizer.RemoveUnusedNodes(cfg);
}
private static Operand GetMemoryOperandOrNull(Operand addr, OperandType type)
{
Operand baseOp = addr;
// First we check if the address is the result of a local X with 32-bits immediate
// addition. If that is the case, then the baseOp is X, and the memory operand immediate
// becomes the addition immediate. Otherwise baseOp keeps being the address.
int imm = GetConstOp(ref baseOp);
// Now we check if the baseOp is the result of a local Y with a local Z addition.
// If that is the case, we now set baseOp to Y and indexOp to Z. We further check
// if Z is the result of a left shift of local W by a value >= 0 and <= 3, if that
// is the case, we set indexOp to W and adjust the scale value of the memory operand
// to match that of the left shift.
// There is one missed case, which is the address being a shift result, but this is
// probably not worth optimizing as it should never happen.
(Operand indexOp, Multiplier scale) = GetIndexOp(ref baseOp);
// If baseOp is still equal to address, then there's nothing that can be optimized.
if (baseOp == addr)
{
return default;
}
if (imm == 0 && scale == Multiplier.x1 && indexOp != default)
{
imm = GetConstOp(ref indexOp);
}
return MemoryOp(type, baseOp, indexOp, scale, imm);
}
private static int GetConstOp(ref Operand baseOp)
{
Operation operation = GetAsgOpWithInst(baseOp, Instruction.Add);
if (operation == default)
{
return 0;
}
Operand src1 = operation.GetSource(0);
Operand src2 = operation.GetSource(1);
Operand constOp;
Operand otherOp;
if (src1.Kind == OperandKind.Constant && src2.Kind == OperandKind.LocalVariable)
{
constOp = src1;
otherOp = src2;
}
else if (src1.Kind == OperandKind.LocalVariable && src2.Kind == OperandKind.Constant)
{
constOp = src2;
otherOp = src1;
}
else
{
return 0;
}
// If we have addition by 64-bits constant, then we can't optimize it further,
// as we can't encode a 64-bits immediate on the memory operand.
if (CodeGenCommon.IsLongConst(constOp))
{
return 0;
}
baseOp = otherOp;
return constOp.AsInt32();
}
private static (Operand, Multiplier) GetIndexOp(ref Operand baseOp)
{
Operand indexOp = default;
Multiplier scale = Multiplier.x1;
Operation addOp = GetAsgOpWithInst(baseOp, Instruction.Add);
if (addOp == default)
{
return (indexOp, scale);
}
Operand src1 = addOp.GetSource(0);
Operand src2 = addOp.GetSource(1);
if (src1.Kind != OperandKind.LocalVariable || src2.Kind != OperandKind.LocalVariable)
{
return (indexOp, scale);
}
baseOp = src1;
indexOp = src2;
Operation shlOp = GetAsgOpWithInst(src1, Instruction.ShiftLeft);
bool indexOnSrc2 = false;
if (shlOp == default)
{
shlOp = GetAsgOpWithInst(src2, Instruction.ShiftLeft);
indexOnSrc2 = true;
}
if (shlOp != default)
{
Operand shSrc = shlOp.GetSource(0);
Operand shift = shlOp.GetSource(1);
if (shSrc.Kind == OperandKind.LocalVariable && shift.Kind == OperandKind.Constant && shift.Value <= 3)
{
scale = shift.Value switch
{
1 => Multiplier.x2,
2 => Multiplier.x4,
3 => Multiplier.x8,
_ => Multiplier.x1
};
baseOp = indexOnSrc2 ? src1 : src2;
indexOp = shSrc;
}
}
return (indexOp, scale);
}
private static Operation GetAsgOpWithInst(Operand op, Instruction inst)
{
// If we have multiple assignments, folding is not safe
// as the value may be different depending on the
// control flow path.
if (op.AssignmentsCount != 1)
{
return default;
}
Operation asgOp = op.Assignments[0];
if (asgOp.Instruction != inst)
{
return default;
}
return asgOp;
}
private static bool IsMemoryLoadOrStore(Instruction inst)
{
return inst == Instruction.Load ||
inst == Instruction.Load16 ||
inst == Instruction.Load8 ||
inst == Instruction.Store ||
inst == Instruction.Store16 ||
inst == Instruction.Store8;
}
}
}