private static bool FindBrxTargets(ShaderConfig config, IEnumerable <Block> blocks, Func <ulong, Block> getBlock)
{
bool hasNewTarget = false;
foreach (Block block in blocks)
{
InstOp lastOp = block.GetLastOp();
bool hasNext = block.HasNext();
if (lastOp.Name == InstName.Brx && block.Successors.Count == (hasNext ? 1 : 0))
{
InstBrx opBrx = new InstBrx(lastOp.RawOpCode);
ulong baseOffset = lastOp.GetAbsoluteAddress();
// An indirect branch could go anywhere,
// try to get the possible target offsets from the constant buffer.
(int cbBaseOffset, int cbOffsetsCount) = FindBrxTargetRange(block, opBrx.SrcA);
if (cbOffsetsCount != 0)
{
hasNewTarget = true;
}
for (int i = 0; i < cbOffsetsCount; i++)
{
uint targetOffset = config.GpuAccessor.ConstantBuffer1Read(cbBaseOffset + i * 4);
Block target = getBlock(baseOffset + targetOffset);
target.Predecessors.Add(block);
block.Successors.Add(target);
}
}
}
return(hasNewTarget);
}
public static void RunPass(BasicBlock block, ShaderConfig config)
{
int sbStart = GetStorageBaseCbOffset(config.Stage);
int sbEnd = sbStart + StorageDescsSize;
for (LinkedListNode <INode> node = block.Operations.First; node != null; node = node.Next)
{
if (!(node.Value is Operation operation))
{
continue;
}
if (UsesGlobalMemory(operation.Inst))
{
Operand source = operation.GetSource(0);
if (source.AsgOp is Operation asgOperation)
{
int storageIndex = SearchForStorageBase(asgOperation, sbStart, sbEnd);
if (storageIndex >= 0)
{
node = ReplaceGlobalWithStorage(node, config, storageIndex);
}
}
}
}
}
static void Main(string[] args)
{
if (args.Length == 2)
{
GalShaderType type = GalShaderType.Vertex;
switch (args[0].ToLower())
{
case "v": type = GalShaderType.Vertex; break;
case "tc": type = GalShaderType.TessControl; break;
case "te": type = GalShaderType.TessEvaluation; break;
case "g": type = GalShaderType.Geometry; break;
case "f": type = GalShaderType.Fragment; break;
}
using (FileStream fs = new FileStream(args[1], FileMode.Open, FileAccess.Read))
{
Memory mem = new Memory(fs);
ShaderConfig config = new ShaderConfig(type, 65536);
string code = Translator.Translate(mem, 0, config).Code;
Console.WriteLine(code);
}
}
else
{
Console.WriteLine("Usage: Ryujinx.ShaderTools [v|tc|te|g|f] shader.bin");
}
}
public static GlslProgram Generate(StructuredProgramInfo info, ShaderConfig config)
{
CodeGenContext context = new CodeGenContext(info, config);
Declarations.Declare(context, info);
if (info.Functions.Count != 0)
{
for (int i = 1; i < info.Functions.Count; i++)
{
context.AppendLine($"{GetFunctionSignature(info.Functions[i])};");
}
context.AppendLine();
for (int i = 1; i < info.Functions.Count; i++)
{
PrintFunction(context, info, info.Functions[i]);
context.AppendLine();
}
}
PrintFunction(context, info, info.Functions[0], MainFunctionName);
return(new GlslProgram(
context.CBufferDescriptors.ToArray(),
context.SBufferDescriptors.ToArray(),
context.TextureDescriptors.ToArray(),
context.ImageDescriptors.ToArray(),
context.GetCode()));
}
public string GetExpression(AstOperand operand, ShaderConfig config, bool cbIndexable)
{
switch (operand.Type)
{
case OperandType.Argument:
return(GetArgumentName(operand.Value));
case OperandType.Attribute:
return(GetAttributeName(operand, config));
case OperandType.Constant:
return(NumberFormatter.FormatInt(operand.Value));
case OperandType.ConstantBuffer:
return(GetConstantBufferName(operand.CbufSlot, operand.CbufOffset, config.Stage, cbIndexable));
case OperandType.LocalVariable:
return(_locals[operand]);
case OperandType.Undefined:
return(DefaultNames.UndefinedName);
}
throw new ArgumentException($"Invalid operand type \"{operand.Type}\".");
}
public static string Generate(StructuredProgramInfo info, ShaderConfig config)
{
CodeGenContext context = new CodeGenContext(info, config);
Declarations.Declare(context, info);
if (info.Functions.Count != 0)
{
for (int i = 1; i < info.Functions.Count; i++)
{
context.AppendLine($"{GetFunctionSignature(info.Functions[i])};");
}
context.AppendLine();
for (int i = 1; i < info.Functions.Count; i++)
{
PrintFunction(context, info, info.Functions[i]);
context.AppendLine();
}
}
PrintFunction(context, info, info.Functions[0], MainFunctionName);
return(context.GetCode());
}
private static void SetUserAttributeUses(ShaderConfig config, IOpCodeAttribute opAttr)
{
if (opAttr.Indexed)
{
if (opAttr.Emitter == InstEmit.Ast)
{
config.SetAllOutputUserAttributes();
}
else
{
config.SetAllInputUserAttributes();
}
}
else
{
for (int elemIndex = 0; elemIndex < opAttr.Count; elemIndex++)
{
int attr = opAttr.AttributeOffset + elemIndex * 4;
if (attr >= AttributeConsts.UserAttributeBase && attr < AttributeConsts.UserAttributeEnd)
{
int index = (attr - AttributeConsts.UserAttributeBase) / 16;
if (opAttr.Emitter == InstEmit.Ast)
{
config.SetOutputUserAttribute(index);
}
else
{
config.SetInputUserAttribute(index);
}
}
}
}
}
public CodeGenContext(StructuredProgramInfo info, ShaderConfig config)
{
_info = info;
Config = config;
OperandManager = new OperandManager();
_sb = new StringBuilder();
}
private static void FillBlock(ShaderConfig config, Block block, ulong limitAddress, ulong startAddress)
{
IGpuAccessor gpuAccessor = config.GpuAccessor;
ulong address = block.Address;
int bufferOffset = 0;
ReadOnlySpan <ulong> buffer = ReadOnlySpan <ulong> .Empty;
InstOp op = default;
do
{
if (address + 7 >= limitAddress)
{
break;
}
// Ignore scheduling instructions, which are written every 32 bytes.
if ((address & 0x1f) == 0)
{
address += 8;
bufferOffset++;
continue;
}
if (bufferOffset >= buffer.Length)
{
buffer = gpuAccessor.GetCode(startAddress + address, 8);
bufferOffset = 0;
}
ulong opCode = buffer[bufferOffset++];
op = InstTable.GetOp(address, opCode);
if (op.Props.HasFlag(InstProps.TexB))
{
config.SetUsedFeature(FeatureFlags.Bindless);
}
if (op.Name == InstName.Ald || op.Name == InstName.Ast || op.Name == InstName.Ipa)
{
SetUserAttributeUses(config, op.Name, opCode);
}
else if (op.Name == InstName.Ssy || op.Name == InstName.Pbk)
{
block.AddPushOp(op);
}
block.OpCodes.Add(op);
address += 8;
}while (!op.Props.HasFlag(InstProps.Bra));
block.EndAddress = address;
}
public static void RunPass(BasicBlock block, ShaderConfig config)
{
// We can turn a bindless into regular access by recognizing the pattern
// produced by the compiler for separate texture and sampler.
// We check for the following conditions:
// - The handle is the result of a bitwise OR logical operation.
// - Both sources of the OR operation comes from CB2 (used by NVN to hold texture handles).
for (LinkedListNode <INode> node = block.Operations.First; node != null; node = node.Next)
{
if (!(node.Value is TextureOperation texOp))
{
continue;
}
if ((texOp.Flags & TextureFlags.Bindless) == 0)
{
continue;
}
if (texOp.Inst == Instruction.TextureSample)
{
if (!(texOp.GetSource(0).AsgOp is Operation handleCombineOp))
{
continue;
}
if (handleCombineOp.Inst != Instruction.BitwiseOr)
{
continue;
}
Operand src0 = handleCombineOp.GetSource(0);
Operand src1 = handleCombineOp.GetSource(1);
if (src0.Type != OperandType.ConstantBuffer || src0.GetCbufSlot() != NvnTextureBufferSlot ||
src1.Type != OperandType.ConstantBuffer || src1.GetCbufSlot() != NvnTextureBufferSlot)
{
continue;
}
texOp.SetHandle(src0.GetCbufOffset() | (src1.GetCbufOffset() << 16));
}
else if (texOp.Inst == Instruction.ImageLoad || texOp.Inst == Instruction.ImageStore)
{
Operand src0 = texOp.GetSource(0);
if (src0.Type == OperandType.ConstantBuffer && src0.GetCbufSlot() == NvnTextureBufferSlot)
{
texOp.SetHandle(src0.GetCbufOffset());
texOp.Format = config.GetTextureFormat(texOp.Handle);
}
}
}
}
public CodeGenContext(ShaderConfig config)
{
Config = config;
CBufferDescriptors = new List <CBufferDescriptor>();
TextureDescriptors = new List <TextureDescriptor>();
OperandManager = new OperandManager();
_sb = new StringBuilder();
}
private static void SetHandle(ShaderConfig config, TextureOperation texOp, int cbufOffset, int cbufSlot, bool rewriteSamplerType)
{
texOp.SetHandle(cbufOffset, cbufSlot);
if (rewriteSamplerType)
{
texOp.Type = config.GpuAccessor.QuerySamplerType(cbufOffset, cbufSlot);
}
config.SetUsedTexture(texOp.Inst, texOp.Type, texOp.Format, texOp.Flags, cbufSlot, cbufOffset);
}
private static int FixedFuncToUserAttribute(ShaderConfig config, int attr, bool isOutput)
{
if (attr >= AttributeConsts.FrontColorDiffuseR && attr < AttributeConsts.ClipDistance0)
{
attr = FixedFuncToUserAttribute(config, attr, AttributeConsts.FrontColorDiffuseR, 0, isOutput);
}
else if (attr >= AttributeConsts.TexCoordBase && attr < AttributeConsts.TexCoordEnd)
{
attr = FixedFuncToUserAttribute(config, attr, AttributeConsts.TexCoordBase, 4, isOutput);
}
return(attr);
}
public static GlslProgram Generate(StructuredProgramInfo info, ShaderConfig config)
{
CodeGenContext context = new CodeGenContext(config);
Declarations.Declare(context, info);
PrintMainBlock(context, info);
return(new GlslProgram(
context.CBufferDescriptors.ToArray(),
context.TextureDescriptors.ToArray(),
context.GetCode()));
}
public string GetExpression(AstOperand operand, ShaderConfig config)
{
return(operand.Type switch
{
OperandType.Argument => GetArgumentName(operand.Value),
OperandType.Attribute => GetAttributeName(operand.Value, config, perPatch: false),
OperandType.AttributePerPatch => GetAttributeName(operand.Value, config, perPatch: true),
OperandType.Constant => NumberFormatter.FormatInt(operand.Value),
OperandType.ConstantBuffer => GetConstantBufferName(operand, config),
OperandType.LocalVariable => _locals[operand],
OperandType.Undefined => DefaultNames.UndefinedName,
_ => throw new ArgumentException($"Invalid operand type \"{operand.Type}\".")
});
public CodeGenContext(ShaderConfig config, bool cbIndexable)
{
Config = config;
CbIndexable = cbIndexable;
CBufferDescriptors = new List <BufferDescriptor>();
SBufferDescriptors = new List <BufferDescriptor>();
TextureDescriptors = new List <TextureDescriptor>();
ImageDescriptors = new List <TextureDescriptor>();
OperandManager = new OperandManager();
_sb = new StringBuilder();
}
public CodeGenContext(StructuredProgramInfo info, ShaderConfig config)
{
_info = info;
Config = config;
CBufferDescriptors = new List <BufferDescriptor>();
SBufferDescriptors = new List <BufferDescriptor>();
TextureDescriptors = new List <TextureDescriptor>();
ImageDescriptors = new List <TextureDescriptor>();
OperandManager = new OperandManager();
_sb = new StringBuilder();
}
public static void RunPass(BasicBlock[] blocks, ShaderConfig config)
{
RunOptimizationPasses(blocks);
// Those passes are looking for specific patterns and only needs to run once.
for (int blkIndex = 0; blkIndex < blocks.Length; blkIndex++)
{
GlobalToStorage.RunPass(blocks[blkIndex], config);
BindlessToIndexed.RunPass(blocks[blkIndex]);
BindlessElimination.RunPass(blocks[blkIndex], config);
}
// Run optimizations one last time to remove any code that is now optimizable after above passes.
RunOptimizationPasses(blocks);
}
public static void RunPass(BasicBlock block, ShaderConfig config)
{
int sbStart = GetStorageBaseCbOffset(config.Stage);
int sbEnd = sbStart + StorageDescsSize;
for (LinkedListNode <INode> node = block.Operations.First; node != null; node = node.Next)
{
if (!(node.Value is Operation operation))
{
continue;
}
if (UsesGlobalMemory(operation.Inst))
{
Operand source = operation.GetSource(0);
if (source.AsgOp is Operation asgOperation)
{
int storageIndex = SearchForStorageBase(asgOperation, sbStart, sbEnd);
if (storageIndex >= 0)
{
// Storage buffers are implemented using global memory access.
// If we know from where the base address of the access is loaded,
// we can guess which storage buffer it is accessing.
// We can then replace the global memory access with a storage
// buffer access.
node = ReplaceGlobalWithStorage(node, config, storageIndex);
}
else if (config.Stage == ShaderStage.Compute && operation.Inst == Instruction.LoadGlobal)
{
// Here we effectively try to replace a LDG instruction with LDC.
// The hardware only supports a limited amount of constant buffers
// so NVN "emulates" more constant buffers using global memory access.
// Here we try to replace the global access back to a constant buffer
// load.
storageIndex = SearchForStorageBase(asgOperation, UbeBaseOffset, UbeBaseOffset + UbeDescsSize);
if (storageIndex >= 0)
{
node = ReplaceLdgWithLdc(node, config, storageIndex);
}
}
}
}
}
}
public string GetExpression(AstOperand operand, ShaderConfig config)
{
return(operand.Type switch
{
OperandType.Argument => GetArgumentName(operand.Value),
OperandType.Attribute => GetAttributeName(operand, config),
OperandType.Constant => NumberFormatter.FormatInt(operand.Value),
OperandType.ConstantBuffer => GetConstantBufferName(
operand.CbufSlot,
operand.CbufOffset,
config.Stage,
config.UsedFeatures.HasFlag(FeatureFlags.CbIndexing)),
OperandType.LocalVariable => _locals[operand],
OperandType.Undefined => DefaultNames.UndefinedName,
_ => throw new ArgumentException($"Invalid operand type \"{operand.Type}\".")
});
public static byte[] Generate(StructuredProgramInfo info, ShaderConfig config)
{
CodeGenContext context = new CodeGenContext(config);
context.AddCapability(Capability.GroupNonUniformBallot);
context.AddCapability(Capability.ImageBuffer);
context.AddCapability(Capability.SampledBuffer);
context.AddCapability(Capability.SubgroupBallotKHR);
context.AddCapability(Capability.SubgroupVoteKHR);
context.AddExtension("SPV_KHR_shader_ballot");
context.AddExtension("SPV_KHR_subgroup_vote");
Declarations.DeclareAll(context, info);
if ((info.HelperFunctionsMask & NeedsInvocationIdMask) != 0)
{
Declarations.DeclareInvocationId(context);
}
for (int funcIndex = 0; funcIndex < info.Functions.Count; funcIndex++)
{
var function = info.Functions[funcIndex];
var retType = context.GetType(function.ReturnType.Convert());
var funcArgs = new SpvInstruction[function.InArguments.Length + function.OutArguments.Length];
for (int argIndex = 0; argIndex < funcArgs.Length; argIndex++)
{
var argType = context.GetType(function.GetArgumentType(argIndex).Convert());
var argPointerType = context.TypePointer(StorageClass.Function, argType);
funcArgs[argIndex] = argPointerType;
}
var funcType = context.TypeFunction(retType, false, funcArgs);
var spvFunc = context.Function(retType, FunctionControlMask.MaskNone, funcType);
context.DeclareFunction(funcIndex, function, spvFunc);
}
for (int funcIndex = 0; funcIndex < info.Functions.Count; funcIndex++)
{
Generate(context, info, funcIndex);
}
return(context.Generate());
}
private static int FixedFuncToUserAttribute(ShaderConfig config, int attr, int baseAttr, int baseIndex, bool isOutput)
{
int index = (attr - baseAttr) >> 4;
int userAttrIndex = config.GetFreeUserAttribute(isOutput, index);
if ((uint)userAttrIndex < Constants.MaxAttributes)
{
userAttrIndex += baseIndex;
attr = AttributeConsts.UserAttributeBase + userAttrIndex * 16 + (attr & 0xf);
if (isOutput)
{
config.SetOutputUserAttributeFixedFunc(userAttrIndex);
}
else
{
config.SetInputUserAttributeFixedFunc(userAttrIndex);
}
}
return(attr);
}
private OglShaderStage ShaderStageFactory(
IGalMemory memory,
long position,
long positionB,
bool isDualVp,
GalShaderType type)
{
ShaderConfig config = new ShaderConfig(type, OglLimit.MaxUboSize);
ShaderProgram program;
if (isDualVp)
{
ShaderDumper.Dump(memory, position, type, "a");
ShaderDumper.Dump(memory, positionB, type, "b");
program = Translator.Translate(memory, (ulong)position, (ulong)positionB, config);
}
else
{
ShaderDumper.Dump(memory, position, type);
program = Translator.Translate(memory, (ulong)position, config);
}
string code = program.Code;
if (ShaderDumper.IsDumpEnabled())
{
int shaderDumpIndex = ShaderDumper.DumpIndex;
code = "//Shader " + shaderDumpIndex + Environment.NewLine + code;
}
return(new OglShaderStage(type, code, program.Info.CBuffers, program.Info.Textures));
}
public static void RunPass(BasicBlock[] blocks, ShaderConfig config)
{
for (int blkIndex = 0; blkIndex < blocks.Length; blkIndex++)
{
GlobalToStorage.RunPass(blocks[blkIndex], config);
}
bool modified;
do
{
modified = false;
for (int blkIndex = 0; blkIndex < blocks.Length; blkIndex++)
{
BasicBlock block = blocks[blkIndex];
LinkedListNode <INode> node = block.Operations.First;
while (node != null)
{
LinkedListNode <INode> nextNode = node.Next;
bool isUnused = IsUnused(node.Value);
if (!(node.Value is Operation operation) || isUnused)
{
if (isUnused)
{
RemoveNode(block, node);
modified = true;
}
node = nextNode;
continue;
}
ConstantFolding.RunPass(operation);
Simplification.RunPass(operation);
if (DestIsLocalVar(operation))
{
if (operation.Inst == Instruction.Copy)
{
PropagateCopy(operation);
RemoveNode(block, node);
modified = true;
}
else if ((operation.Inst == Instruction.PackHalf2x16 && PropagatePack(operation)) ||
(operation.Inst == Instruction.ShuffleXor && MatchDdxOrDdy(operation)))
{
if (operation.Dest.UseOps.Count == 0)
{
RemoveNode(block, node);
}
modified = true;
}
}
node = nextNode;
}
if (BranchElimination.RunPass(block))
{
RemoveNode(block, block.Operations.Last);
modified = true;
}
}
}while (modified);
for (int blkIndex = 0; blkIndex < blocks.Length; blkIndex++)
{
BindlessToIndexed.RunPass(blocks[blkIndex]);
BindlessElimination.RunPass(blocks[blkIndex]);
// Try to eliminate any operations that are now unused.
LinkedListNode <INode> node = blocks[blkIndex].Operations.First;
while (node != null)
{
LinkedListNode <INode> nextNode = node.Next;
if (IsUnused(node.Value))
{
RemoveNode(blocks[blkIndex], node);
}
node = nextNode;
}
}
}
public static string GetAttributeName(AstOperand attr, ShaderConfig config, bool isOutAttr = false, string indexExpr = "0")
{
int value = attr.Value;
char swzMask = GetSwizzleMask((value >> 2) & 3);
if (value >= AttributeConsts.UserAttributeBase && value < AttributeConsts.UserAttributeEnd)
{
value -= AttributeConsts.UserAttributeBase;
string prefix = isOutAttr
? DefaultNames.OAttributePrefix
: DefaultNames.IAttributePrefix;
if ((config.Flags & TranslationFlags.Feedback) != 0)
{
string name = $"{prefix}{(value >> 4)}_{swzMask}";
if (config.Stage == ShaderStage.Geometry && !isOutAttr)
{
name += $"[{indexExpr}]";
}
return(name);
}
else
{
string name = $"{prefix}{(value >> 4)}";
if (config.Stage == ShaderStage.Geometry && !isOutAttr)
{
name += $"[{indexExpr}]";
}
return(name + '.' + swzMask);
}
}
else
{
if (value >= AttributeConsts.FragmentOutputColorBase && value < AttributeConsts.FragmentOutputColorEnd)
{
value -= AttributeConsts.FragmentOutputColorBase;
return($"{DefaultNames.OAttributePrefix}{(value >> 4)}.{swzMask}");
}
else if (_builtInAttributes.TryGetValue(value & ~3, out BuiltInAttribute builtInAttr))
{
// TODO: There must be a better way to handle this...
if (config.Stage == ShaderStage.Fragment)
{
switch (value & ~3)
{
case AttributeConsts.PositionX: return("(gl_FragCoord.x / fp_renderScale[0])");
case AttributeConsts.PositionY: return("(gl_FragCoord.y / fp_renderScale[0])");
case AttributeConsts.PositionZ: return("gl_FragCoord.z");
case AttributeConsts.PositionW: return("gl_FragCoord.w");
}
}
string name = builtInAttr.Name;
if (config.Stage == ShaderStage.Geometry && (value & AttributeConsts.SpecialMask) == 0 && !isOutAttr)
{
name = $"gl_in[{indexExpr}].{name}";
}
return(name);
}
}
// TODO: Warn about unknown built-in attribute.
return(isOutAttr ? "// bad_attr0x" + value.ToString("X") : "0.0");
}
public static string GetOutAttributeName(AstOperand attr, ShaderConfig config)
{
return(GetAttributeName(attr, config, isOutAttr: true));
}
private static void SetUserAttributeUses(ShaderConfig config, InstName name, ulong opCode)
{
int offset;
int count = 1;
bool isStore = false;
bool indexed = false;
bool perPatch = false;
if (name == InstName.Ast)
{
InstAst opAst = new InstAst(opCode);
count = (int)opAst.AlSize + 1;
offset = opAst.Imm11;
indexed = opAst.Phys;
perPatch = opAst.P;
isStore = true;
}
else if (name == InstName.Ald)
{
InstAld opAld = new InstAld(opCode);
count = (int)opAld.AlSize + 1;
offset = opAld.Imm11;
indexed = opAld.Phys;
perPatch = opAld.P;
isStore = opAld.O;
}
else /* if (name == InstName.Ipa) */
{
InstIpa opIpa = new InstIpa(opCode);
offset = opIpa.Imm10;
indexed = opIpa.Idx;
}
if (indexed)
{
if (isStore)
{
config.SetAllOutputUserAttributes();
}
else
{
config.SetAllInputUserAttributes();
}
}
else
{
for (int elemIndex = 0; elemIndex < count; elemIndex++)
{
int attr = offset + elemIndex * 4;
if (attr >= AttributeConsts.UserAttributeBase && attr < AttributeConsts.UserAttributeEnd)
{
int index = (attr - AttributeConsts.UserAttributeBase) / 16;
if (isStore)
{
config.SetOutputUserAttribute(index, perPatch);
}
else
{
config.SetInputUserAttribute(index, perPatch);
}
}
if (!isStore &&
((attr >= AttributeConsts.FrontColorDiffuseR && attr < AttributeConsts.ClipDistance0) ||
(attr >= AttributeConsts.TexCoordBase && attr < AttributeConsts.TexCoordEnd)))
{
config.SetUsedFeature(FeatureFlags.FixedFuncAttr);
}
}
}
}
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 = Cbuf(0, UbeBaseOffset + storageIndex * StorageDescSize);
Operand baseAddrTrunc = Local();
Operand alignMask = Const(-config.QueryInfo(QueryInfoName.StorageBufferOffsetAlignment));
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];
sources[0] = Const(UbeFirstCbuf + storageIndex);
sources[1] = GetCbufOffset();
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;
Operand GetStorageOffset()
{
Operand addrLow = operation.GetSource(0);
Operand baseAddrLow = Cbuf(0, GetStorageCbOffset(config.Stage, storageIndex));
Operand baseAddrTrunc = Local();
Operand alignMask = Const(-config.QueryInfo(QueryInfoName.StorageBufferOffsetAlignment));
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];
sources[0] = Const(storageIndex);
sources[1] = GetStorageOffset();
for (int index = 2; index < operation.SourcesCount; index++)
{
sources[index] = operation.GetSource(index);
}
Operation storageOp;
if (operation.Inst.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
{
storageOp = new Operation(Instruction.StoreStorage, 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);
}
public static void RunPass(BasicBlock block, ShaderConfig config)
{
// We can turn a bindless into regular access by recognizing the pattern
// produced by the compiler for separate texture and sampler.
// We check for the following conditions:
// - The handle is a constant buffer value.
// - The handle is the result of a bitwise OR logical operation.
// - Both sources of the OR operation comes from a constant buffer.
for (LinkedListNode <INode> node = block.Operations.First; node != null; node = node.Next)
{
if (!(node.Value is TextureOperation texOp))
{
continue;
}
if ((texOp.Flags & TextureFlags.Bindless) == 0)
{
continue;
}
if (texOp.Inst == Instruction.Lod ||
texOp.Inst == Instruction.TextureSample ||
texOp.Inst == Instruction.TextureSize)
{
Operand bindlessHandle = Utils.FindLastOperation(texOp.GetSource(0), block);
bool rewriteSamplerType = texOp.Inst == Instruction.TextureSize;
if (bindlessHandle.Type == OperandType.ConstantBuffer)
{
SetHandle(config, texOp, bindlessHandle.GetCbufOffset(), bindlessHandle.GetCbufSlot(), rewriteSamplerType);
continue;
}
if (!(bindlessHandle.AsgOp is Operation handleCombineOp))
{
continue;
}
if (handleCombineOp.Inst != Instruction.BitwiseOr)
{
continue;
}
Operand src0 = Utils.FindLastOperation(handleCombineOp.GetSource(0), block);
Operand src1 = Utils.FindLastOperation(handleCombineOp.GetSource(1), block);
if (src0.Type != OperandType.ConstantBuffer || src1.Type != OperandType.ConstantBuffer)
{
continue;
}
SetHandle(
config,
texOp,
src0.GetCbufOffset() | ((src1.GetCbufOffset() + 1) << 16),
src0.GetCbufSlot() | ((src1.GetCbufSlot() + 1) << 16),
rewriteSamplerType);
}
else if (texOp.Inst == Instruction.ImageLoad ||
texOp.Inst == Instruction.ImageStore ||
texOp.Inst == Instruction.ImageAtomic)
{
Operand src0 = Utils.FindLastOperation(texOp.GetSource(0), block);
if (src0.Type == OperandType.ConstantBuffer)
{
int cbufOffset = src0.GetCbufOffset();
int cbufSlot = src0.GetCbufSlot();
if (texOp.Inst == Instruction.ImageAtomic)
{
texOp.Format = config.GetTextureFormatAtomic(cbufOffset, cbufSlot);
}
else
{
texOp.Format = config.GetTextureFormat(cbufOffset, cbufSlot);
}
SetHandle(config, texOp, cbufOffset, cbufSlot, false);
}
}
}
}
