private static LinkedListNode <INode> ReplaceLdgWithLdc(LinkedListNode <INode> node, ShaderConfig config, int storageIndex)
{
Operation operation = (Operation)node.Value;
Operand GetCbufOffset()
{
Operand addrLow = operation.GetSource(0);
Operand baseAddrLow = config.CreateCbuf(0, UbeBaseOffset + storageIndex * StorageDescSize);
Operand baseAddrTrunc = Local();
Operand alignMask = Const(-config.GpuAccessor.QueryHostStorageBufferOffsetAlignment());
Operation andOp = new Operation(Instruction.BitwiseAnd, baseAddrTrunc, baseAddrLow, alignMask);
node.List.AddBefore(node, andOp);
Operand byteOffset = Local();
Operand wordOffset = Local();
Operation subOp = new Operation(Instruction.Subtract, byteOffset, addrLow, baseAddrTrunc);
Operation shrOp = new Operation(Instruction.ShiftRightU32, wordOffset, byteOffset, Const(2));
node.List.AddBefore(node, subOp);
node.List.AddBefore(node, shrOp);
return(wordOffset);
}
Operand[] sources = new Operand[operation.SourcesCount];
int cbSlot = UbeFirstCbuf + storageIndex;
sources[0] = Const(cbSlot);
sources[1] = GetCbufOffset();
config.SetUsedConstantBuffer(cbSlot);
for (int index = 2; index < operation.SourcesCount; index++)
{
sources[index] = operation.GetSource(index);
}
Operation ldcOp = new Operation(Instruction.LoadConstant, operation.Dest, sources);
for (int index = 0; index < operation.SourcesCount; index++)
{
operation.SetSource(index, null);
}
LinkedListNode <INode> oldNode = node;
node = node.List.AddBefore(node, ldcOp);
node.List.Remove(oldNode);
return(node);
}
private static LinkedListNode <INode> ReplaceGlobalWithStorage(LinkedListNode <INode> node, ShaderConfig config, int storageIndex)
{
Operation operation = (Operation)node.Value;
bool isAtomic = operation.Inst.IsAtomic();
bool isStg16Or8 = operation.Inst == Instruction.StoreGlobal16 || operation.Inst == Instruction.StoreGlobal8;
bool isWrite = isAtomic || operation.Inst == Instruction.StoreGlobal || isStg16Or8;
config.SetUsedStorageBuffer(storageIndex, isWrite);
Operand GetStorageOffset()
{
Operand addrLow = operation.GetSource(0);
Operand baseAddrLow = config.CreateCbuf(0, GetStorageCbOffset(config.Stage, storageIndex));
Operand baseAddrTrunc = Local();
Operand alignMask = Const(-config.GpuAccessor.QueryHostStorageBufferOffsetAlignment());
Operation andOp = new Operation(Instruction.BitwiseAnd, baseAddrTrunc, baseAddrLow, alignMask);
node.List.AddBefore(node, andOp);
Operand byteOffset = Local();
Operation subOp = new Operation(Instruction.Subtract, byteOffset, addrLow, baseAddrTrunc);
node.List.AddBefore(node, subOp);
if (isStg16Or8)
{
return(byteOffset);
}
Operand wordOffset = Local();
Operation shrOp = new Operation(Instruction.ShiftRightU32, wordOffset, byteOffset, Const(2));
node.List.AddBefore(node, shrOp);
return(wordOffset);
}
Operand[] sources = new Operand[operation.SourcesCount];
sources[0] = Const(storageIndex);
sources[1] = GetStorageOffset();
for (int index = 2; index < operation.SourcesCount; index++)
{
sources[index] = operation.GetSource(index);
}
Operation storageOp;
if (isAtomic)
{
Instruction inst = (operation.Inst & ~Instruction.MrMask) | Instruction.MrStorage;
storageOp = new Operation(inst, operation.Dest, sources);
}
else if (operation.Inst == Instruction.LoadGlobal)
{
storageOp = new Operation(Instruction.LoadStorage, operation.Dest, sources);
}
else
{
Instruction storeInst = operation.Inst switch
{
Instruction.StoreGlobal16 => Instruction.StoreStorage16,
Instruction.StoreGlobal8 => Instruction.StoreStorage8,
_ => Instruction.StoreStorage
};
storageOp = new Operation(storeInst, null, sources);
}
for (int index = 0; index < operation.SourcesCount; index++)
{
operation.SetSource(index, null);
}
LinkedListNode <INode> oldNode = node;
node = node.List.AddBefore(node, storageOp);
node.List.Remove(oldNode);
return(node);
}
