Google Git
Sign in
fuchsia / third_party / glslang / refs/heads/upstream/GL_EXT_buffer_reference_uvec2 / . / hlsl / hlslParseHelper.cpp
blob: 1549e3a4ce95aa58fcf27bc82efedfcaf7123421 [file] [log] [blame] [edit]
//
// Copyright (C) 2017-2018 Google, Inc.
// Copyright (C) 2017 LunarG, Inc.
//
// All rights reserved.
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions
// are met:
//
// Redistributions of source code must retain the above copyright
// notice, this list of conditions and the following disclaimer.
//
// Redistributions in binary form must reproduce the above
// copyright notice, this list of conditions and the following
// disclaimer in the documentation and/or other materials provided
// with the distribution.
//
// Neither the name of 3Dlabs Inc. Ltd. nor the names of its
// contributors may be used to endorse or promote products derived
// from this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
// COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
// LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
// CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
// LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
// ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
// POSSIBILITY OF SUCH DAMAGE.
//
#include "hlslParseHelper.h"
#include "hlslScanContext.h"
#include "hlslGrammar.h"
#include "hlslAttributes.h"
#include "../glslang/Include/Common.h"
#include "../glslang/MachineIndependent/Scan.h"
#include "../glslang/MachineIndependent/preprocessor/PpContext.h"
#include "../glslang/OSDependent/osinclude.h"
#include <algorithm>
#include <functional>
#include <cctype>
#include <array>
#include <set>
namespace glslang {
HlslParseContext::HlslParseContext(TSymbolTable& symbolTable, TIntermediate& interm, bool parsingBuiltins,
int version, EProfile profile, const SpvVersion& spvVersion, EShLanguage language,
TInfoSink& infoSink,
const TString sourceEntryPointName,
bool forwardCompatible, EShMessages messages) :
TParseContextBase(symbolTable, interm, parsingBuiltins, version, profile, spvVersion, language, infoSink,
forwardCompatible, messages, &sourceEntryPointName),
annotationNestingLevel(0),
inputPatch(nullptr),
nextInLocation(0), nextOutLocation(0),
entryPointFunction(nullptr),
entryPointFunctionBody(nullptr),
gsStreamOutput(nullptr),
clipDistanceOutput(nullptr),
cullDistanceOutput(nullptr),
clipDistanceInput(nullptr),
cullDistanceInput(nullptr)
{
globalUniformDefaults.clear();
globalUniformDefaults.layoutMatrix = ElmRowMajor;
globalUniformDefaults.layoutPacking = ElpStd140;
globalBufferDefaults.clear();
globalBufferDefaults.layoutMatrix = ElmRowMajor;
globalBufferDefaults.layoutPacking = ElpStd430;
globalInputDefaults.clear();
globalOutputDefaults.clear();
clipSemanticNSizeIn.fill(0);
cullSemanticNSizeIn.fill(0);
clipSemanticNSizeOut.fill(0);
cullSemanticNSizeOut.fill(0);
// "Shaders in the transform
// feedback capturing mode have an initial global default of
// layout(xfb_buffer = 0) out;"
if (language == EShLangVertex ||
language == EShLangTessControl ||
language == EShLangTessEvaluation ||
language == EShLangGeometry)
globalOutputDefaults.layoutXfbBuffer = 0;
if (language == EShLangGeometry)
globalOutputDefaults.layoutStream = 0;
}
HlslParseContext::~HlslParseContext()
{
}
void HlslParseContext::initializeExtensionBehavior()
{
TParseContextBase::initializeExtensionBehavior();
// HLSL allows #line by default.
extensionBehavior[E_GL_GOOGLE_cpp_style_line_directive] = EBhEnable;
}
void HlslParseContext::setLimits(const TBuiltInResource& r)
{
resources = r;
intermediate.setLimits(resources);
}
//
// Parse an array of strings using the parser in HlslRules.
//
// Returns true for successful acceptance of the shader, false if any errors.
//
bool HlslParseContext::parseShaderStrings(TPpContext& ppContext, TInputScanner& input, bool versionWillBeError)
{
currentScanner = &input;
ppContext.setInput(input, versionWillBeError);
HlslScanContext scanContext(*this, ppContext);
HlslGrammar grammar(scanContext, *this);
if (!grammar.parse()) {
// Print a message formated such that if you click on the message it will take you right to
// the line through most UIs.
const glslang::TSourceLoc& sourceLoc = input.getSourceLoc();
infoSink.info << sourceLoc.getFilenameStr() << "(" << sourceLoc.line << "): error at column " << sourceLoc.column
<< ", HLSL parsing failed.\n";
++numErrors;
return false;
}
finish();
return numErrors == 0;
}
//
// Return true if this l-value node should be converted in some manner.
// For instance: turning a load aggregate into a store in an l-value.
//
bool HlslParseContext::shouldConvertLValue(const TIntermNode* node) const
{
if (node == nullptr || node->getAsTyped() == nullptr)
return false;
const TIntermAggregate* lhsAsAggregate = node->getAsAggregate();
const TIntermBinary* lhsAsBinary = node->getAsBinaryNode();
// If it's a swizzled/indexed aggregate, look at the left node instead.
if (lhsAsBinary != nullptr &&
(lhsAsBinary->getOp() == EOpVectorSwizzle || lhsAsBinary->getOp() == EOpIndexDirect))
lhsAsAggregate = lhsAsBinary->getLeft()->getAsAggregate();
if (lhsAsAggregate != nullptr && lhsAsAggregate->getOp() == EOpImageLoad)
return true;
return false;
}
void HlslParseContext::growGlobalUniformBlock(const TSourceLoc& loc, TType& memberType, const TString& memberName,
TTypeList* newTypeList)
{
newTypeList = nullptr;
correctUniform(memberType.getQualifier());
if (memberType.isStruct()) {
auto it = ioTypeMap.find(memberType.getStruct());
if (it != ioTypeMap.end() && it->second.uniform)
newTypeList = it->second.uniform;
}
TParseContextBase::growGlobalUniformBlock(loc, memberType, memberName, newTypeList);
}
//
// Return a TLayoutFormat corresponding to the given texture type.
//
TLayoutFormat HlslParseContext::getLayoutFromTxType(const TSourceLoc& loc, const TType& txType)
{
if (txType.isStruct()) {
// TODO: implement.
error(loc, "unimplemented: structure type in image or buffer", "", "");
return ElfNone;
}
const int components = txType.getVectorSize();
const TBasicType txBasicType = txType.getBasicType();
const auto selectFormat = [this,&components](TLayoutFormat v1, TLayoutFormat v2, TLayoutFormat v4) -> TLayoutFormat {
if (intermediate.getNoStorageFormat())
return ElfNone;
return components == 1 ? v1 :
components == 2 ? v2 : v4;
};
switch (txBasicType) {
case EbtFloat: return selectFormat(ElfR32f, ElfRg32f, ElfRgba32f);
case EbtInt: return selectFormat(ElfR32i, ElfRg32i, ElfRgba32i);
case EbtUint: return selectFormat(ElfR32ui, ElfRg32ui, ElfRgba32ui);
default:
error(loc, "unknown basic type in image format", "", "");
return ElfNone;
}
}
//
// Both test and if necessary, spit out an error, to see if the node is really
// an l-value that can be operated on this way.
//
// Returns true if there was an error.
//
bool HlslParseContext::lValueErrorCheck(const TSourceLoc& loc, const char* op, TIntermTyped* node)
{
if (shouldConvertLValue(node)) {
// if we're writing to a texture, it must be an RW form.
TIntermAggregate* lhsAsAggregate = node->getAsAggregate();
TIntermTyped* object = lhsAsAggregate->getSequence()[0]->getAsTyped();
if (!object->getType().getSampler().isImage()) {
error(loc, "operator[] on a non-RW texture must be an r-value", "", "");
return true;
}
}
// We tolerate samplers as l-values, even though they are nominally
// illegal, because we expect a later optimization to eliminate them.
if (node->getType().getBasicType() == EbtSampler) {
intermediate.setNeedsLegalization();
return false;
}
// Let the base class check errors
return TParseContextBase::lValueErrorCheck(loc, op, node);
}
//
// This function handles l-value conversions and verifications. It uses, but is not synonymous
// with lValueErrorCheck. That function accepts an l-value directly, while this one must be
// given the surrounding tree - e.g, with an assignment, so we can convert the assign into a
// series of other image operations.
//
// Most things are passed through unmodified, except for error checking.
//
TIntermTyped* HlslParseContext::handleLvalue(const TSourceLoc& loc, const char* op, TIntermTyped*& node)
{
if (node == nullptr)
return nullptr;
TIntermBinary* nodeAsBinary = node->getAsBinaryNode();
TIntermUnary* nodeAsUnary = node->getAsUnaryNode();
TIntermAggregate* sequence = nullptr;
TIntermTyped* lhs = nodeAsUnary ? nodeAsUnary->getOperand() :
nodeAsBinary ? nodeAsBinary->getLeft() :
nullptr;
// Early bail out if there is no conversion to apply
if (!shouldConvertLValue(lhs)) {
if (lhs != nullptr)
if (lValueErrorCheck(loc, op, lhs))
return nullptr;
return node;
}
// *** If we get here, we're going to apply some conversion to an l-value.
// Helper to create a load.
const auto makeLoad = [&](TIntermSymbol* rhsTmp, TIntermTyped* object, TIntermTyped* coord, const TType& derefType) {
TIntermAggregate* loadOp = new TIntermAggregate(EOpImageLoad);
loadOp->setLoc(loc);
loadOp->getSequence().push_back(object);
loadOp->getSequence().push_back(intermediate.addSymbol(*coord->getAsSymbolNode()));
loadOp->setType(derefType);
sequence = intermediate.growAggregate(sequence,
intermediate.addAssign(EOpAssign, rhsTmp, loadOp, loc),
loc);
};
// Helper to create a store.
const auto makeStore = [&](TIntermTyped* object, TIntermTyped* coord, TIntermSymbol* rhsTmp) {
TIntermAggregate* storeOp = new TIntermAggregate(EOpImageStore);
storeOp->getSequence().push_back(object);
storeOp->getSequence().push_back(coord);
storeOp->getSequence().push_back(intermediate.addSymbol(*rhsTmp));
storeOp->setLoc(loc);
storeOp->setType(TType(EbtVoid));
sequence = intermediate.growAggregate(sequence, storeOp);
};
// Helper to create an assign.
const auto makeBinary = [&](TOperator op, TIntermTyped* lhs, TIntermTyped* rhs) {
sequence = intermediate.growAggregate(sequence,
intermediate.addBinaryNode(op, lhs, rhs, loc, lhs->getType()),
loc);
};
// Helper to complete sequence by adding trailing variable, so we evaluate to the right value.
const auto finishSequence = [&](TIntermSymbol* rhsTmp, const TType& derefType) -> TIntermAggregate* {
// Add a trailing use of the temp, so the sequence returns the proper value.
sequence = intermediate.growAggregate(sequence, intermediate.addSymbol(*rhsTmp));
sequence->setOperator(EOpSequence);
sequence->setLoc(loc);
sequence->setType(derefType);
return sequence;
};
// Helper to add unary op
const auto makeUnary = [&](TOperator op, TIntermSymbol* rhsTmp) {
sequence = intermediate.growAggregate(sequence,
intermediate.addUnaryNode(op, intermediate.addSymbol(*rhsTmp), loc,
rhsTmp->getType()),
loc);
};
// Return true if swizzle or index writes all components of the given variable.
const auto writesAllComponents = [&](TIntermSymbol* var, TIntermBinary* swizzle) -> bool {
if (swizzle == nullptr) // not a swizzle or index
return true;
// Track which components are being set.
std::array<bool, 4> compIsSet;
compIsSet.fill(false);
const TIntermConstantUnion* asConst = swizzle->getRight()->getAsConstantUnion();
const TIntermAggregate* asAggregate = swizzle->getRight()->getAsAggregate();
// This could be either a direct index, or a swizzle.
if (asConst) {
compIsSet[asConst->getConstArray()[0].getIConst()] = true;
} else if (asAggregate) {
const TIntermSequence& seq = asAggregate->getSequence();
for (int comp=0; comp<int(seq.size()); ++comp)
compIsSet[seq[comp]->getAsConstantUnion()->getConstArray()[0].getIConst()] = true;
} else {
assert(0);
}
// Return true if all components are being set by the index or swizzle
return std::all_of(compIsSet.begin(), compIsSet.begin() + var->getType().getVectorSize(),
[](bool isSet) { return isSet; } );
};
// Create swizzle matching input swizzle
const auto addSwizzle = [&](TIntermSymbol* var, TIntermBinary* swizzle) -> TIntermTyped* {
if (swizzle)
return intermediate.addBinaryNode(swizzle->getOp(), var, swizzle->getRight(), loc, swizzle->getType());
else
return var;
};
TIntermBinary* lhsAsBinary = lhs->getAsBinaryNode();
TIntermAggregate* lhsAsAggregate = lhs->getAsAggregate();
bool lhsIsSwizzle = false;
// If it's a swizzled L-value, remember the swizzle, and use the LHS.
if (lhsAsBinary != nullptr && (lhsAsBinary->getOp() == EOpVectorSwizzle || lhsAsBinary->getOp() == EOpIndexDirect)) {
lhsAsAggregate = lhsAsBinary->getLeft()->getAsAggregate();
lhsIsSwizzle = true;
}
TIntermTyped* object = lhsAsAggregate->getSequence()[0]->getAsTyped();
TIntermTyped* coord = lhsAsAggregate->getSequence()[1]->getAsTyped();
const TSampler& texSampler = object->getType().getSampler();
TType objDerefType;
getTextureReturnType(texSampler, objDerefType);
if (nodeAsBinary) {
TIntermTyped* rhs = nodeAsBinary->getRight();
const TOperator assignOp = nodeAsBinary->getOp();
bool isModifyOp = false;
switch (assignOp) {
case EOpAddAssign:
case EOpSubAssign:
case EOpMulAssign:
case EOpVectorTimesMatrixAssign:
case EOpVectorTimesScalarAssign:
case EOpMatrixTimesScalarAssign:
case EOpMatrixTimesMatrixAssign:
case EOpDivAssign:
case EOpModAssign:
case EOpAndAssign:
case EOpInclusiveOrAssign:
case EOpExclusiveOrAssign:
case EOpLeftShiftAssign:
case EOpRightShiftAssign:
isModifyOp = true;
// fall through...
case EOpAssign:
{
// Since this is an lvalue, we'll convert an image load to a sequence like this
// (to still provide the value):
// OpSequence
// OpImageStore(object, lhs, rhs)
// rhs
// But if it's not a simple symbol RHS (say, a fn call), we don't want to duplicate the RHS,
// so we'll convert instead to this:
// OpSequence
// rhsTmp = rhs
// OpImageStore(object, coord, rhsTmp)
// rhsTmp
// If this is a read-modify-write op, like +=, we issue:
// OpSequence
// coordtmp = load's param1
// rhsTmp = OpImageLoad(object, coordTmp)
// rhsTmp op= rhs
// OpImageStore(object, coordTmp, rhsTmp)
// rhsTmp
//
// If the lvalue is swizzled, we apply that when writing the temp variable, like so:
// ...
// rhsTmp.some_swizzle = ...
// For partial writes, an error is generated.
TIntermSymbol* rhsTmp = rhs->getAsSymbolNode();
TIntermTyped* coordTmp = coord;
if (rhsTmp == nullptr || isModifyOp || lhsIsSwizzle) {
rhsTmp = makeInternalVariableNode(loc, "storeTemp", objDerefType);
// Partial updates not yet supported
if (!writesAllComponents(rhsTmp, lhsAsBinary)) {
error(loc, "unimplemented: partial image updates", "", "");
}
// Assign storeTemp = rhs
if (isModifyOp) {
// We have to make a temp var for the coordinate, to avoid evaluating it twice.
coordTmp = makeInternalVariableNode(loc, "coordTemp", coord->getType());
makeBinary(EOpAssign, coordTmp, coord); // coordtmp = load[param1]
makeLoad(rhsTmp, object, coordTmp, objDerefType); // rhsTmp = OpImageLoad(object, coordTmp)
}
// rhsTmp op= rhs.
makeBinary(assignOp, addSwizzle(intermediate.addSymbol(*rhsTmp), lhsAsBinary), rhs);
}
makeStore(object, coordTmp, rhsTmp); // add a store
return finishSequence(rhsTmp, objDerefType); // return rhsTmp from sequence
}
default:
break;
}
}
if (nodeAsUnary) {
const TOperator assignOp = nodeAsUnary->getOp();
switch (assignOp) {
case EOpPreIncrement:
case EOpPreDecrement:
{
// We turn this into:
// OpSequence
// coordtmp = load's param1
// rhsTmp = OpImageLoad(object, coordTmp)
// rhsTmp op
// OpImageStore(object, coordTmp, rhsTmp)
// rhsTmp
TIntermSymbol* rhsTmp = makeInternalVariableNode(loc, "storeTemp", objDerefType);
TIntermTyped* coordTmp = makeInternalVariableNode(loc, "coordTemp", coord->getType());
makeBinary(EOpAssign, coordTmp, coord); // coordtmp = load[param1]
makeLoad(rhsTmp, object, coordTmp, objDerefType); // rhsTmp = OpImageLoad(object, coordTmp)
makeUnary(assignOp, rhsTmp); // op rhsTmp
makeStore(object, coordTmp, rhsTmp); // OpImageStore(object, coordTmp, rhsTmp)
return finishSequence(rhsTmp, objDerefType); // return rhsTmp from sequence
}
case EOpPostIncrement:
case EOpPostDecrement:
{
// We turn this into:
// OpSequence
// coordtmp = load's param1
// rhsTmp1 = OpImageLoad(object, coordTmp)
// rhsTmp2 = rhsTmp1
// rhsTmp2 op
// OpImageStore(object, coordTmp, rhsTmp2)
// rhsTmp1 (pre-op value)
TIntermSymbol* rhsTmp1 = makeInternalVariableNode(loc, "storeTempPre", objDerefType);
TIntermSymbol* rhsTmp2 = makeInternalVariableNode(loc, "storeTempPost", objDerefType);
TIntermTyped* coordTmp = makeInternalVariableNode(loc, "coordTemp", coord->getType());
makeBinary(EOpAssign, coordTmp, coord); // coordtmp = load[param1]
makeLoad(rhsTmp1, object, coordTmp, objDerefType); // rhsTmp1 = OpImageLoad(object, coordTmp)
makeBinary(EOpAssign, rhsTmp2, rhsTmp1); // rhsTmp2 = rhsTmp1
makeUnary(assignOp, rhsTmp2); // rhsTmp op
makeStore(object, coordTmp, rhsTmp2); // OpImageStore(object, coordTmp, rhsTmp2)
return finishSequence(rhsTmp1, objDerefType); // return rhsTmp from sequence
}
default:
break;
}
}
if (lhs)
if (lValueErrorCheck(loc, op, lhs))
return nullptr;
return node;
}
void HlslParseContext::handlePragma(const TSourceLoc& loc, const TVector<TString>& tokens)
{
if (pragmaCallback)
pragmaCallback(loc.line, tokens);
if (tokens.size() == 0)
return;
// These pragmas are case insensitive in HLSL, so we'll compare in lower case.
TVector<TString> lowerTokens = tokens;
for (auto it = lowerTokens.begin(); it != lowerTokens.end(); ++it)
std::transform(it->begin(), it->end(), it->begin(), ::tolower);
// Handle pack_matrix
if (tokens.size() == 4 && lowerTokens[0] == "pack_matrix" && tokens[1] == "(" && tokens[3] == ")") {
// Note that HLSL semantic order is Mrc, not Mcr like SPIR-V, so we reverse the sense.
// Row major becomes column major and vice versa.
if (lowerTokens[2] == "row_major") {
globalUniformDefaults.layoutMatrix = globalBufferDefaults.layoutMatrix = ElmColumnMajor;
} else if (lowerTokens[2] == "column_major") {
globalUniformDefaults.layoutMatrix = globalBufferDefaults.layoutMatrix = ElmRowMajor;
} else {
// unknown majorness strings are treated as (HLSL column major)==(SPIR-V row major)
warn(loc, "unknown pack_matrix pragma value", tokens[2].c_str(), "");
globalUniformDefaults.layoutMatrix = globalBufferDefaults.layoutMatrix = ElmRowMajor;
}
return;
}
// Handle once
if (lowerTokens[0] == "once") {
warn(loc, "not implemented", "#pragma once", "");
return;
}
}
//
// Look at a '.' matrix selector string and change it into components
// for a matrix. There are two types:
//
// _21 second row, first column (one based)
// _m21 third row, second column (zero based)
//
// Returns true if there is no error.
//
bool HlslParseContext::parseMatrixSwizzleSelector(const TSourceLoc& loc, const TString& fields, int cols, int rows,
TSwizzleSelectors<TMatrixSelector>& components)
{
int startPos[MaxSwizzleSelectors];
int numComps = 0;
TString compString = fields;
// Find where each component starts,
// recording the first character position after the '_'.
for (size_t c = 0; c < compString.size(); ++c) {
if (compString[c] == '_') {
if (numComps >= MaxSwizzleSelectors) {
error(loc, "matrix component swizzle has too many components", compString.c_str(), "");
return false;
}
if (c > compString.size() - 3 ||
((compString[c+1] == 'm' || compString[c+1] == 'M') && c > compString.size() - 4)) {
error(loc, "matrix component swizzle missing", compString.c_str(), "");
return false;
}
startPos[numComps++] = (int)c + 1;
}
}
// Process each component
for (int i = 0; i < numComps; ++i) {
int pos = startPos[i];
int bias = -1;
if (compString[pos] == 'm' || compString[pos] == 'M') {
bias = 0;
++pos;
}
TMatrixSelector comp;
comp.coord1 = compString[pos+0] - '0' + bias;
comp.coord2 = compString[pos+1] - '0' + bias;
if (comp.coord1 < 0 || comp.coord1 >= cols) {
error(loc, "matrix row component out of range", compString.c_str(), "");
return false;
}
if (comp.coord2 < 0 || comp.coord2 >= rows) {
error(loc, "matrix column component out of range", compString.c_str(), "");
return false;
}
components.push_back(comp);
}
return true;
}
// If the 'comps' express a column of a matrix,
// return the column. Column means the first coords all match.
//
// Otherwise, return -1.
//
int HlslParseContext::getMatrixComponentsColumn(int rows, const TSwizzleSelectors<TMatrixSelector>& selector)
{
int col = -1;
// right number of comps?
if (selector.size() != rows)
return -1;
// all comps in the same column?
// rows in order?
col = selector[0].coord1;
for (int i = 0; i < rows; ++i) {
if (col != selector[i].coord1)
return -1;
if (i != selector[i].coord2)
return -1;
}
return col;
}
//
// Handle seeing a variable identifier in the grammar.
//
TIntermTyped* HlslParseContext::handleVariable(const TSourceLoc& loc, const TString* string)
{
int thisDepth;
TSymbol* symbol = symbolTable.find(*string, thisDepth);
if (symbol && symbol->getAsVariable() && symbol->getAsVariable()->isUserType()) {
error(loc, "expected symbol, not user-defined type", string->c_str(), "");
return nullptr;
}
const TVariable* variable = nullptr;
const TAnonMember* anon = symbol ? symbol->getAsAnonMember() : nullptr;
TIntermTyped* node = nullptr;
if (anon) {
// It was a member of an anonymous container, which could be a 'this' structure.
// Create a subtree for its dereference.
if (thisDepth > 0) {
variable = getImplicitThis(thisDepth);
if (variable == nullptr)
error(loc, "cannot access member variables (static member function?)", "this", "");
}
if (variable == nullptr)
variable = anon->getAnonContainer().getAsVariable();
TIntermTyped* container = intermediate.addSymbol(*variable, loc);
TIntermTyped* constNode = intermediate.addConstantUnion(anon->getMemberNumber(), loc);
node = intermediate.addIndex(EOpIndexDirectStruct, container, constNode, loc);
node->setType(*(*variable->getType().getStruct())[anon->getMemberNumber()].type);
if (node->getType().hiddenMember())
error(loc, "member of nameless block was not redeclared", string->c_str(), "");
} else {
// Not a member of an anonymous container.
// The symbol table search was done in the lexical phase.
// See if it was a variable.
variable = symbol ? symbol->getAsVariable() : nullptr;
if (variable) {
if ((variable->getType().getBasicType() == EbtBlock ||
variable->getType().getBasicType() == EbtStruct) && variable->getType().getStruct() == nullptr) {
error(loc, "cannot be used (maybe an instance name is needed)", string->c_str(), "");
variable = nullptr;
}
} else {
if (symbol)
error(loc, "variable name expected", string->c_str(), "");
}
// Recovery, if it wasn't found or was not a variable.
if (variable == nullptr) {
error(loc, "unknown variable", string->c_str(), "");
variable = new TVariable(string, TType(EbtVoid));
}
if (variable->getType().getQualifier().isFrontEndConstant())
node = intermediate.addConstantUnion(variable->getConstArray(), variable->getType(), loc);
else
node = intermediate.addSymbol(*variable, loc);
}
if (variable->getType().getQualifier().isIo())
intermediate.addIoAccessed(*string);
return node;
}
//
// Handle operator[] on any objects it applies to. Currently:
// Textures
// Buffers
//
TIntermTyped* HlslParseContext::handleBracketOperator(const TSourceLoc& loc, TIntermTyped* base, TIntermTyped* index)
{
// handle r-value operator[] on textures and images. l-values will be processed later.
if (base->getType().getBasicType() == EbtSampler && !base->isArray()) {
const TSampler& sampler = base->getType().getSampler();
if (sampler.isImage() || sampler.isTexture()) {
if (! mipsOperatorMipArg.empty() && mipsOperatorMipArg.back().mipLevel == nullptr) {
// The first operator[] to a .mips[] sequence is the mip level. We'll remember it.
mipsOperatorMipArg.back().mipLevel = index;
return base; // next [] index is to the same base.
} else {
TIntermAggregate* load = new TIntermAggregate(sampler.isImage() ? EOpImageLoad : EOpTextureFetch);
TType sampReturnType;
getTextureReturnType(sampler, sampReturnType);
load->setType(sampReturnType);
load->setLoc(loc);
load->getSequence().push_back(base);
load->getSequence().push_back(index);
// Textures need a MIP. If we saw one go by, use it. Otherwise, use zero.
if (sampler.isTexture()) {
if (! mipsOperatorMipArg.empty()) {
load->getSequence().push_back(mipsOperatorMipArg.back().mipLevel);
mipsOperatorMipArg.pop_back();
} else {
load->getSequence().push_back(intermediate.addConstantUnion(0, loc, true));
}
}
return load;
}
}
}
// Handle operator[] on structured buffers: this indexes into the array element of the buffer.
// indexStructBufferContent returns nullptr if it isn't a structuredbuffer (SSBO).
TIntermTyped* sbArray = indexStructBufferContent(loc, base);
if (sbArray != nullptr) {
if (sbArray == nullptr)
return nullptr;
// Now we'll apply the [] index to that array
const TOperator idxOp = (index->getQualifier().storage == EvqConst) ? EOpIndexDirect : EOpIndexIndirect;
TIntermTyped* element = intermediate.addIndex(idxOp, sbArray, index, loc);
const TType derefType(sbArray->getType(), 0);
element->setType(derefType);
return element;
}
return nullptr;
}
//
// Cast index value to a uint if it isn't already (for operator[], load indexes, etc)
TIntermTyped* HlslParseContext::makeIntegerIndex(TIntermTyped* index)
{
const TBasicType indexBasicType = index->getType().getBasicType();
const int vecSize = index->getType().getVectorSize();
// We can use int types directly as the index
if (indexBasicType == EbtInt || indexBasicType == EbtUint ||
indexBasicType == EbtInt64 || indexBasicType == EbtUint64)
return index;
// Cast index to unsigned integer if it isn't one.
return intermediate.addConversion(EOpConstructUint, TType(EbtUint, EvqTemporary, vecSize), index);
}
//
// Handle seeing a base[index] dereference in the grammar.
//
TIntermTyped* HlslParseContext::handleBracketDereference(const TSourceLoc& loc, TIntermTyped* base, TIntermTyped* index)
{
index = makeIntegerIndex(index);
if (index == nullptr) {
error(loc, " unknown index type ", "", "");
return nullptr;
}
TIntermTyped* result = handleBracketOperator(loc, base, index);
if (result != nullptr)
return result; // it was handled as an operator[]
bool flattened = false;
int indexValue = 0;
if (index->getQualifier().isFrontEndConstant())
indexValue = index->getAsConstantUnion()->getConstArray()[0].getIConst();
variableCheck(base);
if (! base->isArray() && ! base->isMatrix() && ! base->isVector()) {
if (base->getAsSymbolNode())
error(loc, " left of '[' is not of type array, matrix, or vector ",
base->getAsSymbolNode()->getName().c_str(), "");
else
error(loc, " left of '[' is not of type array, matrix, or vector ", "expression", "");
} else if (base->getType().getQualifier().isFrontEndConstant() &&
index->getQualifier().isFrontEndConstant()) {
// both base and index are front-end constants
checkIndex(loc, base->getType(), indexValue);
return intermediate.foldDereference(base, indexValue, loc);
} else {
// at least one of base and index is variable...
if (index->getQualifier().isFrontEndConstant())
checkIndex(loc, base->getType(), indexValue);
if (base->getType().isScalarOrVec1())
result = base;
else if (base->getAsSymbolNode() && wasFlattened(base)) {
if (index->getQualifier().storage != EvqConst)
error(loc, "Invalid variable index to flattened array", base->getAsSymbolNode()->getName().c_str(), "");
result = flattenAccess(base, indexValue);
flattened = (result != base);
} else {
if (index->getQualifier().isFrontEndConstant()) {
if (base->getType().isUnsizedArray())
base->getWritableType().updateImplicitArraySize(indexValue + 1);
else
checkIndex(loc, base->getType(), indexValue);
result = intermediate.addIndex(EOpIndexDirect, base, index, loc);
} else
result = intermediate.addIndex(EOpIndexIndirect, base, index, loc);
}
}
if (result == nullptr) {
// Insert dummy error-recovery result
result = intermediate.addConstantUnion(0.0, EbtFloat, loc);
} else {
// If the array reference was flattened, it has the correct type. E.g, if it was
// a uniform array, it was flattened INTO a set of scalar uniforms, not scalar temps.
// In that case, we preserve the qualifiers.
if (!flattened) {
// Insert valid dereferenced result
TType newType(base->getType(), 0); // dereferenced type
if (base->getType().getQualifier().storage == EvqConst && index->getQualifier().storage == EvqConst)
newType.getQualifier().storage = EvqConst;
else
newType.getQualifier().storage = EvqTemporary;
result->setType(newType);
}
}
return result;
}
// Handle seeing a binary node with a math operation.
TIntermTyped* HlslParseContext::handleBinaryMath(const TSourceLoc& loc, const char* str, TOperator op,
TIntermTyped* left, TIntermTyped* right)
{
TIntermTyped* result = intermediate.addBinaryMath(op, left, right, loc);
if (result == nullptr)
binaryOpError(loc, str, left->getCompleteString(), right->getCompleteString());
return result;
}
// Handle seeing a unary node with a math operation.
TIntermTyped* HlslParseContext::handleUnaryMath(const TSourceLoc& loc, const char* str, TOperator op,
TIntermTyped* childNode)
{
TIntermTyped* result = intermediate.addUnaryMath(op, childNode, loc);
if (result)
return result;
else
unaryOpError(loc, str, childNode->getCompleteString());
return childNode;
}
//
// Return true if the name is a struct buffer method
//
bool HlslParseContext::isStructBufferMethod(const TString& name) const
{
return
name == "GetDimensions" ||
name == "Load" ||
name == "Load2" ||
name == "Load3" ||
name == "Load4" ||
name == "Store" ||
name == "Store2" ||
name == "Store3" ||
name == "Store4" ||
name == "InterlockedAdd" ||
name == "InterlockedAnd" ||
name == "InterlockedCompareExchange" ||
name == "InterlockedCompareStore" ||
name == "InterlockedExchange" ||
name == "InterlockedMax" ||
name == "InterlockedMin" ||
name == "InterlockedOr" ||
name == "InterlockedXor" ||
name == "IncrementCounter" ||
name == "DecrementCounter" ||
name == "Append" ||
name == "Consume";
}
//
// Handle seeing a base.field dereference in the grammar, where 'field' is a
// swizzle or member variable.
//
TIntermTyped* HlslParseContext::handleDotDereference(const TSourceLoc& loc, TIntermTyped* base, const TString& field)
{
variableCheck(base);
if (base->isArray()) {
error(loc, "cannot apply to an array:", ".", field.c_str());
return base;
}
TIntermTyped* result = base;
if (base->getType().getBasicType() == EbtSampler) {
// Handle .mips[mipid][pos] operation on textures
const TSampler& sampler = base->getType().getSampler();
if (sampler.isTexture() && field == "mips") {
// Push a null to signify that we expect a mip level under operator[] next.
mipsOperatorMipArg.push_back(tMipsOperatorData(loc, nullptr));
// Keep 'result' pointing to 'base', since we expect an operator[] to go by next.
} else {
if (field == "mips")
error(loc, "unexpected texture type for .mips[][] operator:",
base->getType().getCompleteString().c_str(), "");
else
error(loc, "unexpected operator on texture type:", field.c_str(),
base->getType().getCompleteString().c_str());
}
} else if (base->isVector() || base->isScalar()) {
TSwizzleSelectors<TVectorSelector> selectors;
parseSwizzleSelector(loc, field, base->getVectorSize(), selectors);
if (base->isScalar()) {
if (selectors.size() == 1)
return result;
else {
TType type(base->getBasicType(), EvqTemporary, selectors.size());
return addConstructor(loc, base, type);
}
}
if (base->getVectorSize() == 1) {
TType scalarType(base->getBasicType(), EvqTemporary, 1);
if (selectors.size() == 1)
return addConstructor(loc, base, scalarType);
else {
TType vectorType(base->getBasicType(), EvqTemporary, selectors.size());
return addConstructor(loc, addConstructor(loc, base, scalarType), vectorType);
}
}
if (base->getType().getQualifier().isFrontEndConstant())
result = intermediate.foldSwizzle(base, selectors, loc);
else {
if (selectors.size() == 1) {
TIntermTyped* index = intermediate.addConstantUnion(selectors[0], loc);
result = intermediate.addIndex(EOpIndexDirect, base, index, loc);
result->setType(TType(base->getBasicType(), EvqTemporary));
} else {
TIntermTyped* index = intermediate.addSwizzle(selectors, loc);
result = intermediate.addIndex(EOpVectorSwizzle, base, index, loc);
result->setType(TType(base->getBasicType(), EvqTemporary, base->getType().getQualifier().precision,
selectors.size()));
}
}
} else if (base->isMatrix()) {
TSwizzleSelectors<TMatrixSelector> selectors;
if (! parseMatrixSwizzleSelector(loc, field, base->getMatrixCols(), base->getMatrixRows(), selectors))
return result;
if (selectors.size() == 1) {
// Representable by m[c][r]
if (base->getType().getQualifier().isFrontEndConstant()) {
result = intermediate.foldDereference(base, selectors[0].coord1, loc);
result = intermediate.foldDereference(result, selectors[0].coord2, loc);
} else {
result = intermediate.addIndex(EOpIndexDirect, base,
intermediate.addConstantUnion(selectors[0].coord1, loc),
loc);
TType dereferencedCol(base->getType(), 0);
result->setType(dereferencedCol);
result = intermediate.addIndex(EOpIndexDirect, result,
intermediate.addConstantUnion(selectors[0].coord2, loc),
loc);
TType dereferenced(dereferencedCol, 0);
result->setType(dereferenced);
}
} else {
int column = getMatrixComponentsColumn(base->getMatrixRows(), selectors);
if (column >= 0) {
// Representable by m[c]
if (base->getType().getQualifier().isFrontEndConstant())
result = intermediate.foldDereference(base, column, loc);
else {
result = intermediate.addIndex(EOpIndexDirect, base, intermediate.addConstantUnion(column, loc),
loc);
TType dereferenced(base->getType(), 0);
result->setType(dereferenced);
}
} else {
// general case, not a column, not a single component
TIntermTyped* index = intermediate.addSwizzle(selectors, loc);
result = intermediate.addIndex(EOpMatrixSwizzle, base, index, loc);
result->setType(TType(base->getBasicType(), EvqTemporary, base->getType().getQualifier().precision,
selectors.size()));
}
}
} else if (base->getBasicType() == EbtStruct || base->getBasicType() == EbtBlock) {
const TTypeList* fields = base->getType().getStruct();
bool fieldFound = false;
int member;
for (member = 0; member < (int)fields->size(); ++member) {
if ((*fields)[member].type->getFieldName() == field) {
fieldFound = true;
break;
}
}
if (fieldFound) {
if (base->getAsSymbolNode() && wasFlattened(base)) {
result = flattenAccess(base, member);
} else {
if (base->getType().getQualifier().storage == EvqConst)
result = intermediate.foldDereference(base, member, loc);
else {
TIntermTyped* index = intermediate.addConstantUnion(member, loc);
result = intermediate.addIndex(EOpIndexDirectStruct, base, index, loc);
result->setType(*(*fields)[member].type);
}
}
} else
error(loc, "no such field in structure", field.c_str(), "");
} else
error(loc, "does not apply to this type:", field.c_str(), base->getType().getCompleteString().c_str());
return result;
}
//
// Return true if the field should be treated as a built-in method.
// Return false otherwise.
//
bool HlslParseContext::isBuiltInMethod(const TSourceLoc&, TIntermTyped* base, const TString& field)
{
if (base == nullptr)
return false;
variableCheck(base);
if (base->getType().getBasicType() == EbtSampler) {
return true;
} else if (isStructBufferType(base->getType()) && isStructBufferMethod(field)) {
return true;
} else if (field == "Append" ||
field == "RestartStrip") {
// We cannot check the type here: it may be sanitized if we're not compiling a geometry shader, but
// the code is around in the shader source.
return true;
} else
return false;
}
// Independently establish a built-in that is a member of a structure.
// 'arraySizes' are what's desired for the independent built-in, whatever
// the higher-level source/expression of them was.
void HlslParseContext::splitBuiltIn(const TString& baseName, const TType& memberType, const TArraySizes* arraySizes,
const TQualifier& outerQualifier)
{
// Because of arrays of structs, we might be asked more than once,
// but the arraySizes passed in should have captured the whole thing
// the first time.
// However, clip/cull rely on multiple updates.
if (!isClipOrCullDistance(memberType))
if (splitBuiltIns.find(tInterstageIoData(memberType.getQualifier().builtIn, outerQualifier.storage)) !=
splitBuiltIns.end())
return;
TVariable* ioVar = makeInternalVariable(baseName + "." + memberType.getFieldName(), memberType);
if (arraySizes != nullptr && !memberType.isArray())
ioVar->getWritableType().copyArraySizes(*arraySizes);
splitBuiltIns[tInterstageIoData(memberType.getQualifier().builtIn, outerQualifier.storage)] = ioVar;
if (!isClipOrCullDistance(ioVar->getType()))
trackLinkage(*ioVar);
// Merge qualifier from the user structure
mergeQualifiers(ioVar->getWritableType().getQualifier(), outerQualifier);
// Fix the builtin type if needed (e.g, some types require fixed array sizes, no matter how the
// shader declared them). This is done after mergeQualifiers(), in case fixBuiltInIoType looks
// at the qualifier to determine e.g, in or out qualifications.
fixBuiltInIoType(ioVar->getWritableType());
// But, not location, we're losing that
ioVar->getWritableType().getQualifier().layoutLocation = TQualifier::layoutLocationEnd;
}
// Split a type into
// 1. a struct of non-I/O members
// 2. a collection of independent I/O variables
void HlslParseContext::split(const TVariable& variable)
{
// Create a new variable:
const TType& clonedType = *variable.getType().clone();
const TType& splitType = split(clonedType, variable.getName(), clonedType.getQualifier());
splitNonIoVars[variable.getUniqueId()] = makeInternalVariable(variable.getName(), splitType);
}
// Recursive implementation of split().
// Returns reference to the modified type.
const TType& HlslParseContext::split(const TType& type, const TString& name, const TQualifier& outerQualifier)
{
if (type.isStruct()) {
TTypeList* userStructure = type.getWritableStruct();
for (auto ioType = userStructure->begin(); ioType != userStructure->end(); ) {
if (ioType->type->isBuiltIn()) {
// move out the built-in
splitBuiltIn(name, *ioType->type, type.getArraySizes(), outerQualifier);
ioType = userStructure->erase(ioType);
} else {
split(*ioType->type, name + "." + ioType->type->getFieldName(), outerQualifier);
++ioType;
}
}
}
return type;
}
// Is this an aggregate that should be flattened?
// Can be applied to intermediate levels of type in a hierarchy.
// Some things like flattening uniform arrays are only about the top level
// of the aggregate, triggered on 'topLevel'.
bool HlslParseContext::shouldFlatten(const TType& type, TStorageQualifier qualifier, bool topLevel) const
{
switch (qualifier) {
case EvqVaryingIn:
case EvqVaryingOut:
return type.isStruct() || type.isArray();
case EvqUniform:
return (type.isArray() && intermediate.getFlattenUniformArrays() && topLevel) ||
(type.isStruct() && type.containsOpaque());
default:
return false;
};
}
// Top level variable flattening: construct data
void HlslParseContext::flatten(const TVariable& variable, bool linkage)
{
const TType& type = variable.getType();
// If it's a standalone built-in, there is nothing to flatten
if (type.isBuiltIn() && !type.isStruct())
return;
auto entry = flattenMap.insert(std::make_pair(variable.getUniqueId(),
TFlattenData(type.getQualifier().layoutBinding,
type.getQualifier().layoutLocation)));
// the item is a map pair, so first->second is the TFlattenData itself.
flatten(variable, type, entry.first->second, variable.getName(), linkage, type.getQualifier(), nullptr);
}
// Recursively flatten the given variable at the provided type, building the flattenData as we go.
//
// This is mutually recursive with flattenStruct and flattenArray.
// We are going to flatten an arbitrarily nested composite structure into a linear sequence of
// members, and later on, we want to turn a path through the tree structure into a final
// location in this linear sequence.
//
// If the tree was N-ary, that can be directly calculated. However, we are dealing with
// arbitrary numbers - perhaps a struct of 7 members containing an array of 3. Thus, we must
// build a data structure to allow the sequence of bracket and dot operators on arrays and
// structs to arrive at the proper member.
//
// To avoid storing a tree with pointers, we are going to flatten the tree into a vector of integers.
// The leaves are the indexes into the flattened member array.
// Each level will have the next location for the Nth item stored sequentially, so for instance:
//
// struct { float2 a[2]; int b; float4 c[3] };
//
// This will produce the following flattened tree:
// Pos: 0 1 2 3 4 5 6 7 8 9 10 11 12 13
// (3, 7, 8, 5, 6, 0, 1, 2, 11, 12, 13, 3, 4, 5}
//
// Given a reference to mystruct.c[1], the access chain is (2,1), so we traverse:
// (0+2) = 8 --> (8+1) = 12 --> 12 = 4
//
// so the 4th flattened member in traversal order is ours.
//
int HlslParseContext::flatten(const TVariable& variable, const TType& type,
TFlattenData& flattenData, TString name, bool linkage,
const TQualifier& outerQualifier,
const TArraySizes* builtInArraySizes)
{
// If something is an arrayed struct, the array flattener will recursively call flatten()
// to then flatten the struct, so this is an "if else": we don't do both.
if (type.isArray())
return flattenArray(variable, type, flattenData, name, linkage, outerQualifier);
else if (type.isStruct())
return flattenStruct(variable, type, flattenData, name, linkage, outerQualifier, builtInArraySizes);
else {
assert(0); // should never happen
return -1;
}
}
// Add a single flattened member to the flattened data being tracked for the composite
// Returns true for the final flattening level.
int HlslParseContext::addFlattenedMember(const TVariable& variable, const TType& type, TFlattenData& flattenData,
const TString& memberName, bool linkage,
const TQualifier& outerQualifier,
const TArraySizes* builtInArraySizes)
{
if (!shouldFlatten(type, outerQualifier.storage, false)) {
// This is as far as we flatten. Insert the variable.
TVariable* memberVariable = makeInternalVariable(memberName, type);
mergeQualifiers(memberVariable->getWritableType().getQualifier(), variable.getType().getQualifier());
if (flattenData.nextBinding != TQualifier::layoutBindingEnd)
memberVariable->getWritableType().getQualifier().layoutBinding = flattenData.nextBinding++;
if (memberVariable->getType().isBuiltIn()) {
// inherited locations are nonsensical for built-ins (TODO: what if semantic had a number)
memberVariable->getWritableType().getQualifier().layoutLocation = TQualifier::layoutLocationEnd;
} else {
// inherited locations must be auto bumped, not replicated
if (flattenData.nextLocation != TQualifier::layoutLocationEnd) {
memberVariable->getWritableType().getQualifier().layoutLocation = flattenData.nextLocation;
flattenData.nextLocation += intermediate.computeTypeLocationSize(memberVariable->getType(), language);
nextOutLocation = std::max(nextOutLocation, flattenData.nextLocation);
}
}
flattenData.offsets.push_back(static_cast<int>(flattenData.members.size()));
flattenData.members.push_back(memberVariable);
if (linkage)
trackLinkage(*memberVariable);
return static_cast<int>(flattenData.offsets.size()) - 1; // location of the member reference
} else {
// Further recursion required
return flatten(variable, type, flattenData, memberName, linkage, outerQualifier, builtInArraySizes);
}
}
// Figure out the mapping between an aggregate's top members and an
// equivalent set of individual variables.
//
// Assumes shouldFlatten() or equivalent was called first.
int HlslParseContext::flattenStruct(const TVariable& variable, const TType& type,
TFlattenData& flattenData, TString name, bool linkage,
const TQualifier& outerQualifier,
const TArraySizes* builtInArraySizes)
{
assert(type.isStruct());
auto members = *type.getStruct();
// Reserve space for this tree level.
int start = static_cast<int>(flattenData.offsets.size());
int pos = start;
flattenData.offsets.resize(int(pos + members.size()), -1);
for (int member = 0; member < (int)members.size(); ++member) {
TType& dereferencedType = *members[member].type;
if (dereferencedType.isBuiltIn())
splitBuiltIn(variable.getName(), dereferencedType, builtInArraySizes, outerQualifier);
else {
const int mpos = addFlattenedMember(variable, dereferencedType, flattenData,
name + "." + dereferencedType.getFieldName(),
linkage, outerQualifier,
builtInArraySizes == nullptr && dereferencedType.isArray()
? dereferencedType.getArraySizes()
: builtInArraySizes);
flattenData.offsets[pos++] = mpos;
}
}
return start;
}
// Figure out mapping between an array's members and an
// equivalent set of individual variables.
//
// Assumes shouldFlatten() or equivalent was called first.
int HlslParseContext::flattenArray(const TVariable& variable, const TType& type,
TFlattenData& flattenData, TString name, bool linkage,
const TQualifier& outerQualifier)
{
assert(type.isSizedArray());
const int size = type.getOuterArraySize();
const TType dereferencedType(type, 0);
if (name.empty())
name = variable.getName();
// Reserve space for this tree level.
int start = static_cast<int>(flattenData.offsets.size());
int pos = start;
flattenData.offsets.resize(int(pos + size), -1);
for (int element=0; element < size; ++element) {
char elementNumBuf[20]; // sufficient for MAXINT
snprintf(elementNumBuf, sizeof(elementNumBuf)-1, "[%d]", element);
const int mpos = addFlattenedMember(variable, dereferencedType, flattenData,
name + elementNumBuf, linkage, outerQualifier,
type.getArraySizes());
flattenData.offsets[pos++] = mpos;
}
return start;
}
// Return true if we have flattened this node.
bool HlslParseContext::wasFlattened(const TIntermTyped* node) const
{
return node != nullptr && node->getAsSymbolNode() != nullptr &&
wasFlattened(node->getAsSymbolNode()->getId());
}
// Return true if we have split this structure
bool HlslParseContext::wasSplit(const TIntermTyped* node) const
{
return node != nullptr && node->getAsSymbolNode() != nullptr &&
wasSplit(node->getAsSymbolNode()->getId());
}
// Turn an access into an aggregate that was flattened to instead be
// an access to the individual variable the member was flattened to.
// Assumes wasFlattened() or equivalent was called first.
TIntermTyped* HlslParseContext::flattenAccess(TIntermTyped* base, int member)
{
const TType dereferencedType(base->getType(), member); // dereferenced type
const TIntermSymbol& symbolNode = *base->getAsSymbolNode();
TIntermTyped* flattened = flattenAccess(symbolNode.getId(), member, base->getQualifier().storage,
dereferencedType, symbolNode.getFlattenSubset());
return flattened ? flattened : base;
}
TIntermTyped* HlslParseContext::flattenAccess(int uniqueId, int member, TStorageQualifier outerStorage,
const TType& dereferencedType, int subset)
{
const auto flattenData = flattenMap.find(uniqueId);
if (flattenData == flattenMap.end())
return nullptr;
// Calculate new cumulative offset from the packed tree
int newSubset = flattenData->second.offsets[subset >= 0 ? subset + member : member];
TIntermSymbol* subsetSymbol;
if (!shouldFlatten(dereferencedType, outerStorage, false)) {
// Finished flattening: create symbol for variable
member = flattenData->second.offsets[newSubset];
const TVariable* memberVariable = flattenData->second.members[member];
subsetSymbol = intermediate.addSymbol(*memberVariable);
subsetSymbol->setFlattenSubset(-1);
} else {
// If this is not the final flattening, accumulate the position and return
// an object of the partially dereferenced type.
subsetSymbol = new TIntermSymbol(uniqueId, "flattenShadow", dereferencedType);
subsetSymbol->setFlattenSubset(newSubset);
}
return subsetSymbol;
}
// For finding where the first leaf is in a subtree of a multi-level aggregate
// that is just getting a subset assigned. Follows the same logic as flattenAccess,
// but logically going down the "left-most" tree branch each step of the way.
//
// Returns the offset into the first leaf of the subset.
int HlslParseContext::findSubtreeOffset(const TIntermNode& node) const
{
const TIntermSymbol* sym = node.getAsSymbolNode();
if (sym == nullptr)
return 0;
if (!sym->isArray() && !sym->isStruct())
return 0;
int subset = sym->getFlattenSubset();
if (subset == -1)
return 0;
// Getting this far means a partial aggregate is identified by the flatten subset.
// Find the first leaf of the subset.
const auto flattenData = flattenMap.find(sym->getId());
if (flattenData == flattenMap.end())
return 0;
return findSubtreeOffset(sym->getType(), subset, flattenData->second.offsets);
do {
subset = flattenData->second.offsets[subset];
} while (true);
}
// Recursively do the desent
int HlslParseContext::findSubtreeOffset(const TType& type, int subset, const TVector<int>& offsets) const
{
if (!type.isArray() && !type.isStruct())
return offsets[subset];
TType derefType(type, 0);
return findSubtreeOffset(derefType, offsets[subset], offsets);
};
// Find and return the split IO TVariable for id, or nullptr if none.
TVariable* HlslParseContext::getSplitNonIoVar(int id) const
{
const auto splitNonIoVar = splitNonIoVars.find(id);
if (splitNonIoVar == splitNonIoVars.end())
return nullptr;
return splitNonIoVar->second;
}
// Pass through to base class after remembering built-in mappings.
void HlslParseContext::trackLinkage(TSymbol& symbol)
{
TBuiltInVariable biType = symbol.getType().getQualifier().builtIn;
if (biType != EbvNone)
builtInTessLinkageSymbols[biType] = symbol.clone();
TParseContextBase::trackLinkage(symbol);
}
// Returns true if the built-in is a clip or cull distance variable.
bool HlslParseContext::isClipOrCullDistance(TBuiltInVariable builtIn)
{
return builtIn == EbvClipDistance || builtIn == EbvCullDistance;
}
// Some types require fixed array sizes in SPIR-V, but can be scalars or
// arrays of sizes SPIR-V doesn't allow. For example, tessellation factors.
// This creates the right size. A conversion is performed when the internal
// type is copied to or from the external type. This corrects the externally
// facing input or output type to abide downstream semantics.
void HlslParseContext::fixBuiltInIoType(TType& type)
{
int requiredArraySize = 0;
int requiredVectorSize = 0;
switch (type.getQualifier().builtIn) {
case EbvTessLevelOuter: requiredArraySize = 4; break;
case EbvTessLevelInner: requiredArraySize = 2; break;
case EbvSampleMask:
{
// Promote scalar to array of size 1. Leave existing arrays alone.
if (!type.isArray())
requiredArraySize = 1;
break;
}
case EbvWorkGroupId: requiredVectorSize = 3; break;
case EbvGlobalInvocationId: requiredVectorSize = 3; break;
case EbvLocalInvocationId: requiredVectorSize = 3; break;
case EbvTessCoord: requiredVectorSize = 3; break;
default:
if (isClipOrCullDistance(type)) {
const int loc = type.getQualifier().layoutLocation;
if (type.getQualifier().builtIn == EbvClipDistance) {
if (type.getQualifier().storage == EvqVaryingIn)
clipSemanticNSizeIn[loc] = type.getVectorSize();
else
clipSemanticNSizeOut[loc] = type.getVectorSize();
} else {
if (type.getQualifier().storage == EvqVaryingIn)
cullSemanticNSizeIn[loc] = type.getVectorSize();
else
cullSemanticNSizeOut[loc] = type.getVectorSize();
}
}
return;
}
// Alter or set vector size as needed.
if (requiredVectorSize > 0) {
TType newType(type.getBasicType(), type.getQualifier().storage, requiredVectorSize);
newType.getQualifier() = type.getQualifier();
type.shallowCopy(newType);
}
// Alter or set array size as needed.
if (requiredArraySize > 0) {
if (!type.isArray() || type.getOuterArraySize() != requiredArraySize) {
TArraySizes* arraySizes = new TArraySizes;
arraySizes->addInnerSize(requiredArraySize);
type.transferArraySizes(arraySizes);
}
}
}
// Variables that correspond to the user-interface in and out of a stage
// (not the built-in interface) are
// - assigned locations
// - registered as a linkage node (part of the stage's external interface).
// Assumes it is called in the order in which locations should be assigned.
void HlslParseContext::assignToInterface(TVariable& variable)
{
const auto assignLocation = [&](TVariable& variable) {
TType& type = variable.getWritableType();
if (!type.isStruct() || type.getStruct()->size() > 0) {
TQualifier& qualifier = type.getQualifier();
if (qualifier.storage == EvqVaryingIn || qualifier.storage == EvqVaryingOut) {
if (qualifier.builtIn == EbvNone && !qualifier.hasLocation()) {
// Strip off the outer array dimension for those having an extra one.
int size;
if (type.isArray() && qualifier.isArrayedIo(language)) {
TType elementType(type, 0);
size = intermediate.computeTypeLocationSize(elementType, language);
} else
size = intermediate.computeTypeLocationSize(type, language);
if (qualifier.storage == EvqVaryingIn) {
variable.getWritableType().getQualifier().layoutLocation = nextInLocation;
nextInLocation += size;
} else {
variable.getWritableType().getQualifier().layoutLocation = nextOutLocation;
nextOutLocation += size;
}
}
trackLinkage(variable);
}
}
};
if (wasFlattened(variable.getUniqueId())) {
auto& memberList = flattenMap[variable.getUniqueId()].members;
for (auto member = memberList.begin(); member != memberList.end(); ++member)
assignLocation(**member);
} else if (wasSplit(variable.getUniqueId())) {
TVariable* splitIoVar = getSplitNonIoVar(variable.getUniqueId());
assignLocation(*splitIoVar);
} else {
assignLocation(variable);
}
}
//
// Handle seeing a function declarator in the grammar. This is the precursor
// to recognizing a function prototype or function definition.
//
void HlslParseContext::handleFunctionDeclarator(const TSourceLoc& loc, TFunction& function, bool prototype)
{
//
// Multiple declarations of the same function name are allowed.
//
// If this is a definition, the definition production code will check for redefinitions
// (we don't know at this point if it's a definition or not).
//
bool builtIn;
TSymbol* symbol = symbolTable.find(function.getMangledName(), &builtIn);
const TFunction* prevDec = symbol ? symbol->getAsFunction() : 0;
if (prototype) {
// All built-in functions are defined, even though they don't have a body.
// Count their prototype as a definition instead.
if (symbolTable.atBuiltInLevel())
function.setDefined();
else {
if (prevDec && ! builtIn)
symbol->getAsFunction()->setPrototyped(); // need a writable one, but like having prevDec as a const
function.setPrototyped();
}
}
// This insert won't actually insert it if it's a duplicate signature, but it will still check for
// other forms of name collisions.
if (! symbolTable.insert(function))
error(loc, "function name is redeclaration of existing name", function.getName().c_str(), "");
}
// For struct buffers with counters, we must pass the counter buffer as hidden parameter.
// This adds the hidden parameter to the parameter list in 'paramNodes' if needed.
// Otherwise, it's a no-op
void HlslParseContext::addStructBufferHiddenCounterParam(const TSourceLoc& loc, TParameter& param,
TIntermAggregate*& paramNodes)
{
if (! hasStructBuffCounter(*param.type))
return;
const TString counterBlockName(intermediate.addCounterBufferName(*param.name));
TType counterType;
counterBufferType(loc, counterType);
TVariable *variable = makeInternalVariable(counterBlockName, counterType);
if (! symbolTable.insert(*variable))
error(loc, "redefinition", variable->getName().c_str(), "");
paramNodes = intermediate.growAggregate(paramNodes,
intermediate.addSymbol(*variable, loc),
loc);
}
//
// Handle seeing the function prototype in front of a function definition in the grammar.
// The body is handled after this function returns.
//
// Returns an aggregate of parameter-symbol nodes.
//
TIntermAggregate* HlslParseContext::handleFunctionDefinition(const TSourceLoc& loc, TFunction& function,
const TAttributes& attributes,
TIntermNode*& entryPointTree)
{
currentCaller = function.getMangledName();
TSymbol* symbol = symbolTable.find(function.getMangledName());
TFunction* prevDec = symbol ? symbol->getAsFunction() : nullptr;
if (prevDec == nullptr)
error(loc, "can't find function", function.getName().c_str(), "");
// Note: 'prevDec' could be 'function' if this is the first time we've seen function
// as it would have just been put in the symbol table. Otherwise, we're looking up
// an earlier occurrence.
if (prevDec && prevDec->isDefined()) {
// Then this function already has a body.
error(loc, "function already has a body", function.getName().c_str(), "");
}
if (prevDec && ! prevDec->isDefined()) {
prevDec->setDefined();
// Remember the return type for later checking for RETURN statements.
currentFunctionType = &(prevDec->getType());
} else
currentFunctionType = new TType(EbtVoid);
functionReturnsValue = false;
// Entry points need different I/O and other handling, transform it so the
// rest of this function doesn't care.
entryPointTree = transformEntryPoint(loc, function, attributes);
//
// New symbol table scope for body of function plus its arguments
//
pushScope();
//
// Insert parameters into the symbol table.
// If the parameter has no name, it's not an error, just don't insert it
// (could be used for unused args).
//
// Also, accumulate the list of parameters into the AST, so lower level code
// knows where to find parameters.
//
TIntermAggregate* paramNodes = new TIntermAggregate;
for (int i = 0; i < function.getParamCount(); i++) {
TParameter& param = function[i];
if (param.name != nullptr) {
TVariable *variable = new TVariable(param.name, *param.type);
if (i == 0 && function.hasImplicitThis()) {
// Anonymous 'this' members are already in a symbol-table level,
// and we need to know what function parameter to map them to.
symbolTable.makeInternalVariable(*variable);
pushImplicitThis(variable);
}
// Insert the parameters with name in the symbol table.
if (! symbolTable.insert(*variable))
error(loc, "redefinition", variable->getName().c_str(), "");
// Add parameters to the AST list.
if (shouldFlatten(variable->getType(), variable->getType().getQualifier().storage, true)) {
// Expand the AST parameter nodes (but not the name mangling or symbol table view)
// for structures that need to be flattened.
flatten(*variable, false);
const TTypeList* structure = variable->getType().getStruct();
for (int mem = 0; mem < (int)structure->size(); ++mem) {
paramNodes = intermediate.growAggregate(paramNodes,
flattenAccess(variable->getUniqueId(), mem,
variable->getType().getQualifier().storage,
*(*structure)[mem].type),
loc);
}
} else {
// Add the parameter to the AST
paramNodes = intermediate.growAggregate(paramNodes,
intermediate.addSymbol(*variable, loc),
loc);
}
// Add hidden AST parameter for struct buffer counters, if needed.
addStructBufferHiddenCounterParam(loc, param, paramNodes);
} else
paramNodes = intermediate.growAggregate(paramNodes, intermediate.addSymbol(*param.type, loc), loc);
}
if (function.hasIllegalImplicitThis())
pushImplicitThis(nullptr);
intermediate.setAggregateOperator(paramNodes, EOpParameters, TType(EbtVoid), loc);
loopNestingLevel = 0;
controlFlowNestingLevel = 0;
postEntryPointReturn = false;
return paramNodes;
}
// Handle all [attrib] attribute for the shader entry point
void HlslParseContext::handleEntryPointAttributes(const TSourceLoc& loc, const TAttributes& attributes)
{
for (auto it = attributes.begin(); it != attributes.end(); ++it) {
switch (it->name) {
case EatNumThreads:
{
const TIntermSequence& sequence = it->args->getSequence();
for (int lid = 0; lid < int(sequence.size()); ++lid)
intermediate.setLocalSize(lid, sequence[lid]->getAsConstantUnion()->getConstArray()[0].getIConst());
break;
}
case EatMaxVertexCount:
{
int maxVertexCount;
if (! it->getInt(maxVertexCount)) {
error(loc, "invalid maxvertexcount", "", "");
} else {
if (! intermediate.setVertices(maxVertexCount))
error(loc, "cannot change previously set maxvertexcount attribute", "", "");
}
break;
}
case EatPatchConstantFunc:
{
TString pcfName;
if (! it->getString(pcfName, 0, false)) {
error(loc, "invalid patch constant function", "", "");
} else {
patchConstantFunctionName = pcfName;
}
break;
}
case EatDomain:
{
// Handle [domain("...")]
TString domainStr;
if (! it->getString(domainStr)) {
error(loc, "invalid domain", "", "");
} else {
TLayoutGeometry domain = ElgNone;
if (domainStr == "tri") {
domain = ElgTriangles;
} else if (domainStr == "quad") {
domain = ElgQuads;
} else if (domainStr == "isoline") {
domain = ElgIsolines;
} else {
error(loc, "unsupported domain type", domainStr.c_str(), "");
}
if (language == EShLangTessEvaluation) {
if (! intermediate.setInputPrimitive(domain))
error(loc, "cannot change previously set domain", TQualifier::getGeometryString(domain), "");
} else {
if (! intermediate.setOutputPrimitive(domain))
error(loc, "cannot change previously set domain", TQualifier::getGeometryString(domain), "");
}
}
break;
}
case EatOutputTopology:
{
// Handle [outputtopology("...")]
TString topologyStr;
if (! it->getString(topologyStr)) {
error(loc, "invalid outputtopology", "", "");
} else {
TVertexOrder vertexOrder = EvoNone;
TLayoutGeometry primitive = ElgNone;
if (topologyStr == "point") {
intermediate.setPointMode();
} else if (topologyStr == "line") {
primitive = ElgIsolines;
} else if (topologyStr == "triangle_cw") {
vertexOrder = EvoCw;
primitive = ElgTriangles;
} else if (topologyStr == "triangle_ccw") {
vertexOrder = EvoCcw;
primitive = ElgTriangles;
} else {
error(loc, "unsupported outputtopology type", topologyStr.c_str(), "");
}
if (vertexOrder != EvoNone) {
if (! intermediate.setVertexOrder(vertexOrder)) {
error(loc, "cannot change previously set outputtopology",
TQualifier::getVertexOrderString(vertexOrder), "");
}
}
if (primitive != ElgNone)
intermediate.setOutputPrimitive(primitive);
}
break;
}
case EatPartitioning:
{
// Handle [partitioning("...")]
TString partitionStr;
if (! it->getString(partitionStr)) {
error(loc, "invalid partitioning", "", "");
} else {
TVertexSpacing partitioning = EvsNone;
if (partitionStr == "integer") {
partitioning = EvsEqual;
} else if (partitionStr == "fractional_even") {
partitioning = EvsFractionalEven;
} else if (partitionStr == "fractional_odd") {
partitioning = EvsFractionalOdd;
//} else if (partition == "pow2") { // TODO: currently nothing to map this to.
} else {
error(loc, "unsupported partitioning type", partitionStr.c_str(), "");
}
if (! intermediate.setVertexSpacing(partitioning))
error(loc, "cannot change previously set partitioning",
TQualifier::getVertexSpacingString(partitioning), "");
}
break;
}
case EatOutputControlPoints:
{
// Handle [outputcontrolpoints("...")]
int ctrlPoints;
if (! it->getInt(ctrlPoints)) {
error(loc, "invalid outputcontrolpoints", "", "");
} else {
if (! intermediate.setVertices(ctrlPoints)) {
error(loc, "cannot change previously set outputcontrolpoints attribute", "", "");
}
}
break;
}
case EatEarlyDepthStencil:
intermediate.setEarlyFragmentTests();
break;
case EatBuiltIn:
case EatLocation:
// tolerate these because of dual use of entrypoint and type attributes
break;
default:
warn(loc, "attribute does not apply to entry point", "", "");
break;
}
}
}
// Update the given type with any type-like attribute information in the
// attributes.
void HlslParseContext::transferTypeAttributes(const TSourceLoc& loc, const TAttributes& attributes, TType& type,
bool allowEntry)
{
if (attributes.size() == 0)
return;
int value;
TString builtInString;
for (auto it = attributes.begin(); it != attributes.end(); ++it) {
switch (it->name) {
case EatLocation:
// location
if (it->getInt(value))
type.getQualifier().layoutLocation = value;
else
error(loc, "needs a literal integer", "location", "");
break;
case EatBinding:
// binding
if (it->getInt(value)) {
type.getQualifier().layoutBinding = value;
type.getQualifier().layoutSet = 0;
} else
error(loc, "needs a literal integer", "binding", "");
// set
if (it->getInt(value, 1))
type.getQualifier().layoutSet = value;
break;
case EatGlobalBinding:
// global cbuffer binding
if (it->getInt(value))
globalUniformBinding = value;
else
error(loc, "needs a literal integer", "global binding", "");
// global cbuffer set
if (it->getInt(value, 1))
globalUniformSet = value;
break;
case EatInputAttachment:
// input attachment
if (it->getInt(value))
type.getQualifier().layoutAttachment = value;
else
error(loc, "needs a literal integer", "input attachment", "");
break;
case EatBuiltIn:
// PointSize built-in
if (it->getString(builtInString, 0, false)) {
if (builtInString == "PointSize")
type.getQualifier().builtIn = EbvPointSize;
}
break;
case EatPushConstant:
// push_constant
type.getQualifier().layoutPushConstant = true;
break;
case EatConstantId:
// specialization constant
if (it->getInt(value)) {
TSourceLoc loc;
loc.init();
setSpecConstantId(loc, type.getQualifier(), value);
}
break;
default:
if (! allowEntry)
warn(loc, "attribute does not apply to a type", "", "");
break;
}
}
}
//
// Do all special handling for the entry point, including wrapping
// the shader's entry point with the official entry point that will call it.
//
// The following:
//
// retType shaderEntryPoint(args...) // shader declared entry point
// { body }
//
// Becomes
//
// out retType ret;
// in iargs<that are input>...;
// out oargs<that are output> ...;
//
// void shaderEntryPoint() // synthesized, but official, entry point
// {
// args<that are input> = iargs...;
// ret = @shaderEntryPoint(args...);
// oargs = args<that are output>...;
// }
// retType @shaderEntryPoint(args...)
// { body }
//
// The symbol table will still map the original entry point name to the
// the modified function and its new name:
//
// symbol table: shaderEntryPoint -> @shaderEntryPoint
//
// Returns nullptr if no entry-point tree was built, otherwise, returns
// a subtree that creates the entry point.
//
TIntermNode* HlslParseContext::transformEntryPoint(const TSourceLoc& loc, TFunction& userFunction,
const TAttributes& attributes)
{
// Return true if this is a tessellation patch constant function input to a domain shader.
const auto isDsPcfInput = [this](const TType& type) {
return language == EShLangTessEvaluation &&
type.contains([](const TType* t) {
return t->getQualifier().builtIn == EbvTessLevelOuter ||
t->getQualifier().builtIn == EbvTessLevelInner;
});
};
// if we aren't in the entry point, fix the IO as such and exit
if (userFunction.getName().compare(intermediate.getEntryPointName().c_str()) != 0) {
remapNonEntryPointIO(userFunction);
return nullptr;
}
entryPointFunction = &userFunction; // needed in finish()
// Handle entry point attributes
handleEntryPointAttributes(loc, attributes);
// entry point logic...
// Move parameters and return value to shader in/out
TVariable* entryPointOutput; // gets created in remapEntryPointIO
TVector<TVariable*> inputs;
TVector<TVariable*> outputs;
remapEntryPointIO(userFunction, entryPointOutput, inputs, outputs);
// Further this return/in/out transform by flattening, splitting, and assigning locations
const auto makeVariableInOut = [&](TVariable& variable) {
if (variable.getType().isStruct()) {
if (variable.getType().getQualifier().isArrayedIo(language)) {
if (variable.getType().containsBuiltIn())
split(variable);
} else if (shouldFlatten(variable.getType(), EvqVaryingIn /* not assigned yet, but close enough */, true))
flatten(variable, false /* don't track linkage here, it will be tracked in assignToInterface() */);
}
// TODO: flatten arrays too
// TODO: flatten everything in I/O
// TODO: replace all split with flatten, make all paths can create flattened I/O, then split code can be removed
// For clip and cull distance, multiple output variables potentially get merged
// into one in assignClipCullDistance. That code in assignClipCullDistance
// handles the interface logic, so we avoid it here in that case.
if (!isClipOrCullDistance(variable.getType()))
assignToInterface(variable);
};
if (entryPointOutput != nullptr)
makeVariableInOut(*entryPointOutput);
for (auto it = inputs.begin(); it != inputs.end(); ++it)
if (!isDsPcfInput((*it)->getType())) // wait until the end for PCF input (see comment below)
makeVariableInOut(*(*it));
for (auto it = outputs.begin(); it != outputs.end(); ++it)
makeVariableInOut(*(*it));
// In the domain shader, PCF input must be at the end of the linkage. That's because in the
// hull shader there is no ordering: the output comes from the separate PCF, which does not
// participate in the argument list. That is always put at the end of the HS linkage, so the
// input side of the DS must match. The argument may be in any position in the DS argument list
// however, so this ensures the linkage is built in the correct order regardless of argument order.
if (language == EShLangTessEvaluation) {
for (auto it = inputs.begin(); it != inputs.end(); ++it)
if (isDsPcfInput((*it)->getType()))
makeVariableInOut(*(*it));
}
// Synthesize the call
pushScope(); // matches the one in handleFunctionBody()
// new signature
TType voidType(EbtVoid);
TFunction synthEntryPoint(&userFunction.getName(), voidType);
TIntermAggregate* synthParams = new TIntermAggregate();
intermediate.setAggregateOperator(synthParams, EOpParameters, voidType, loc);
intermediate.setEntryPointMangledName(synthEntryPoint.getMangledName().c_str());
intermediate.incrementEntryPointCount();
TFunction callee(&userFunction.getName(), voidType); // call based on old name, which is still in the symbol table
// change original name
userFunction.addPrefix("@"); // change the name in the function, but not in the symbol table
// Copy inputs (shader-in -> calling arg), while building up the call node
TVector<TVariable*> argVars;
TIntermAggregate* synthBody = new TIntermAggregate();
auto inputIt = inputs.begin();
TIntermTyped* callingArgs = nullptr;
for (int i = 0; i < userFunction.getParamCount(); i++) {
TParameter& param = userFunction[i];
argVars.push_back(makeInternalVariable(*param.name, *param.type));
argVars.back()->getWritableType().getQualifier().makeTemporary();
// Track the input patch, which is the only non-builtin supported by hull shader PCF.
if (param.getDeclaredBuiltIn() == EbvInputPatch)
inputPatch = argVars.back();
TIntermSymbol* arg = intermediate.addSymbol(*argVars.back());
handleFunctionArgument(&callee, callingArgs, arg);
if (param.type->getQualifier().isParamInput()) {
intermediate.growAggregate(synthBody, handleAssign(loc, EOpAssign, arg,
intermediate.addSymbol(**inputIt)));
inputIt++;
}
}
// Call
currentCaller = synthEntryPoint.getMangledName();
TIntermTyped* callReturn = handleFunctionCall(loc, &callee, callingArgs);
currentCaller = userFunction.getMangledName();
// Return value
if (entryPointOutput) {
TIntermTyped* returnAssign;
// For hull shaders, the wrapped entry point return value is written to
// an array element as indexed by invocation ID, which we might have to make up.
// This is required to match SPIR-V semantics.
if (language == EShLangTessControl) {
TIntermSymbol* invocationIdSym = findTessLinkageSymbol(EbvInvocationId);
// If there is no user declared invocation ID, we must make one.
if (invocationIdSym == nullptr) {
TType invocationIdType(EbtUint, EvqIn, 1);
TString* invocationIdName = NewPoolTString("InvocationId");
invocationIdType.getQualifier().builtIn = EbvInvocationId;
TVariable* variable = makeInternalVariable(*invocationIdName, invocationIdType);
globalQualifierFix(loc, variable->getWritableType().getQualifier());
trackLinkage(*variable);
invocationIdSym = intermediate.addSymbol(*variable);
}
TIntermTyped* element = intermediate.addIndex(EOpIndexIndirect, intermediate.addSymbol(*entryPointOutput),
invocationIdSym, loc);
// Set the type of the array element being dereferenced
const TType derefElementType(entryPointOutput->getType(), 0);
element->setType(derefElementType);
returnAssign = handleAssign(loc, EOpAssign, element, callReturn);
} else {
returnAssign = handleAssign(loc, EOpAssign, intermediate.addSymbol(*entryPointOutput), callReturn);
}
intermediate.growAggregate(synthBody, returnAssign);
} else
intermediate.growAggregate(synthBody, callReturn);
// Output copies
auto outputIt = outputs.begin();
for (int i = 0; i < userFunction.getParamCount(); i++) {
TParameter& param = userFunction[i];
// GS outputs are via emit, so we do not copy them here.
if (param.type->getQualifier().isParamOutput()) {
if (param.getDeclaredBuiltIn() == EbvGsOutputStream) {
// GS output stream does not assign outputs here: it's the Append() method
// which writes to the output, probably multiple times separated by Emit.
// We merely remember the output to use, here.
gsStreamOutput = *outputIt;
} else {
intermediate.growAggregate(synthBody, handleAssign(loc, EOpAssign,
intermediate.addSymbol(**outputIt),
intermediate.addSymbol(*argVars[i])));
}
outputIt++;
}
}
// Put the pieces together to form a full function subtree
// for the synthesized entry point.
synthBody->setOperator(EOpSequence);
TIntermNode* synthFunctionDef = synthParams;
handleFunctionBody(loc, synthEntryPoint, synthBody, synthFunctionDef);
entryPointFunctionBody = synthBody;
return synthFunctionDef;
}
void HlslParseContext::handleFunctionBody(const TSourceLoc& loc, TFunction& function, TIntermNode* functionBody,
TIntermNode*& node)
{
node = intermediate.growAggregate(node, functionBody);
intermediate.setAggregateOperator(node, EOpFunction, function.getType(), loc);
node->getAsAggregate()->setName(function.getMangledName().c_str());
popScope();
if (function.hasImplicitThis())
popImplicitThis();
if (function.getType().getBasicType() != EbtVoid && ! functionReturnsValue)
error(loc, "function does not return a value:", "", function.getName().c_str());
}
// AST I/O is done through shader globals declared in the 'in' or 'out'
// storage class. An HLSL entry point has a return value, input parameters
// and output parameters. These need to get remapped to the AST I/O.
void HlslParseContext::remapEntryPointIO(TFunction& function, TVariable*& returnValue,
TVector<TVariable*>& inputs, TVector<TVariable*>& outputs)
{
// We might have in input structure type with no decorations that caused it
// to look like an input type, yet it has (e.g.) interpolation types that
// must be modified that turn it into an input type.
// Hence, a missing ioTypeMap for 'input' might need to be synthesized.
const auto synthesizeEditedInput = [this](TType& type) {
// True if a type needs to be 'flat'
const auto needsFlat = [](const TType& type) {
return type.containsBasicType(EbtInt) ||
type.containsBasicType(EbtUint) ||
type.containsBasicType(EbtInt64) ||
type.containsBasicType(EbtUint64) ||
type.containsBasicType(EbtBool) ||
type.containsBasicType(EbtDouble);
};
if (language == EShLangFragment && needsFlat(type)) {
if (type.isStruct()) {
TTypeList* finalList = nullptr;
auto it = ioTypeMap.find(type.getStruct());
if (it == ioTypeMap.end() || it->second.input == nullptr) {
// Getting here means we have no input struct, but we need one.
auto list = new TTypeList;
for (auto member = type.getStruct()->begin(); member != type.getStruct()->end(); ++member) {
TType* newType = new TType;
newType->shallowCopy(*member->type);
TTypeLoc typeLoc = { newType, member->loc };
list->push_back(typeLoc);
}
// install the new input type
if (it == ioTypeMap.end()) {
tIoKinds newLists = { list, nullptr, nullptr };
ioTypeMap[type.getStruct()] = newLists;
} else
it->second.input = list;
finalList = list;
} else
finalList = it->second.input;
// edit for 'flat'
for (auto member = finalList->begin(); member != finalList->end(); ++member) {
if (needsFlat(*member->type)) {
member->type->getQualifier().clearInterpolation();
member->type->getQualifier().flat = true;
}
}
} else {
type.getQualifier().clearInterpolation();
type.getQualifier().flat = true;
}
}
};
// Do the actual work to make a type be a shader input or output variable,
// and clear the original to be non-IO (for use as a normal function parameter/return).
const auto makeIoVariable = [this](const char* name, TType& type, TStorageQualifier storage) -> TVariable* {
TVariable* ioVariable = makeInternalVariable(name, type);
clearUniformInputOutput(type.getQualifier());
if (type.isStruct()) {
auto newLists = ioTypeMap.find(ioVariable->getType().getStruct());
if (newLists != ioTypeMap.end()) {
if (storage == EvqVaryingIn && newLists->second.input)
ioVariable->getWritableType().setStruct(newLists->second.input);
else if (storage == EvqVaryingOut && newLists->second.output)
ioVariable->getWritableType().setStruct(newLists->second.output);
}
}
if (storage == EvqVaryingIn) {
correctInput(ioVariable->getWritableType().getQualifier());
if (language == EShLangTessEvaluation)
if (!ioVariable->getType().isArray())
ioVariable->getWritableType().getQualifier().patch = true;
} else {
correctOutput(ioVariable->getWritableType().getQualifier());
}
ioVariable->getWritableType().getQualifier().storage = storage;
fixBuiltInIoType(ioVariable->getWritableType());
return ioVariable;
};
// return value is actually a shader-scoped output (out)
if (function.getType().getBasicType() == EbtVoid) {
returnValue = nullptr;
} else {
if (language == EShLangTessControl) {
// tessellation evaluation in HLSL writes a per-ctrl-pt value, but it needs to be an
// array in SPIR-V semantics. We'll write to it indexed by invocation ID.
returnValue = makeIoVariable("@entryPointOutput", function.getWritableType(), EvqVaryingOut);
TType outputType;
outputType.shallowCopy(function.getType());
// vertices has necessarily already been set when handling entry point attributes.
TArraySizes* arraySizes = new TArraySizes;
arraySizes->addInnerSize(intermediate.getVertices());
outputType.transferArraySizes(arraySizes);
clearUniformInputOutput(function.getWritableType().getQualifier());
returnValue = makeIoVariable("@entryPointOutput", outputType, EvqVaryingOut);
} else {
returnValue = makeIoVariable("@entryPointOutput", function.getWritableType(), EvqVaryingOut);
}
}
// parameters are actually shader-scoped inputs and outputs (in or out)
for (int i = 0; i < function.getParamCount(); i++) {
TType& paramType = *function[i].type;
if (paramType.getQualifier().isParamInput()) {
synthesizeEditedInput(paramType);
TVariable* argAsGlobal = makeIoVariable(function[i].name->c_str(), paramType, EvqVaryingIn);
inputs.push_back(argAsGlobal);
}
if (paramType.getQualifier().isParamOutput()) {
TVariable* argAsGlobal = makeIoVariable(function[i].name->c_str(), paramType, EvqVaryingOut);
outputs.push_back(argAsGlobal);
}
}
}
// An HLSL function that looks like an entry point, but is not,
// declares entry point IO built-ins, but these have to be undone.
void HlslParseContext::remapNonEntryPointIO(TFunction& function)
{
// return value
if (function.getType().getBasicType() != EbtVoid)
clearUniformInputOutput(function.getWritableType().getQualifier());
// parameters.
// References to structuredbuffer types are left unmodified
for (int i = 0; i < function.getParamCount(); i++)
if (!isReference(*function[i].type))
clearUniformInputOutput(function[i].type->getQualifier());
}
// Handle function returns, including type conversions to the function return type
// if necessary.
TIntermNode* HlslParseContext::handleReturnValue(const TSourceLoc& loc, TIntermTyped* value)
{
functionReturnsValue = true;
if (currentFunctionType->getBasicType() == EbtVoid) {
error(loc, "void function cannot return a value", "return", "");
return intermediate.addBranch(EOpReturn, loc);
} else if (*currentFunctionType != value->getType()) {
value = intermediate.addConversion(EOpReturn, *currentFunctionType, value);
if (value && *currentFunctionType != value->getType())
value = intermediate.addUniShapeConversion(EOpReturn, *currentFunctionType, value);
if (value == nullptr || *currentFunctionType != value->getType()) {
error(loc, "type does not match, or is not convertible to, the function's return type", "return", "");
return value;
}
}
return intermediate.addBranch(EOpReturn, value, loc);
}
void HlslParseContext::handleFunctionArgument(TFunction* function,
TIntermTyped*& arguments, TIntermTyped* newArg)
{
TParameter param = { 0, new TType, nullptr };
param.type->shallowCopy(newArg->getType());
function->addParameter(param);
if (arguments)
arguments = intermediate.growAggregate(arguments, newArg);
else
arguments = newArg;
}
// Position may require special handling: we can optionally invert Y.
// See: https://github.com/KhronosGroup/glslang/issues/1173
// https://github.com/KhronosGroup/glslang/issues/494
TIntermTyped* HlslParseContext::assignPosition(const TSourceLoc& loc, TOperator op,
TIntermTyped* left, TIntermTyped* right)
{
// If we are not asked for Y inversion, use a plain old assign.
if (!intermediate.getInvertY())
return intermediate.addAssign(op, left, right, loc);
// If we get here, we should invert Y.
TIntermAggregate* assignList = nullptr;
// If this is a complex rvalue, we don't want to dereference it many times. Create a temporary.
TVariable* rhsTempVar = nullptr;
rhsTempVar = makeInternalVariable("@position", right->getType());
rhsTempVar->getWritableType().getQualifier().makeTemporary();
{
TIntermTyped* rhsTempSym = intermediate.addSymbol(*rhsTempVar, loc);
assignList = intermediate.growAggregate(assignList,
intermediate.addAssign(EOpAssign, rhsTempSym, right, loc), loc);
}
// pos.y = -pos.y
{
const int Y = 1;
TIntermTyped* tempSymL = intermediate.addSymbol(*rhsTempVar, loc);
TIntermTyped* tempSymR = intermediate.addSymbol(*rhsTempVar, loc);
TIntermTyped* index = intermediate.addConstantUnion(Y, loc);
TIntermTyped* lhsElement = intermediate.addIndex(EOpIndexDirect, tempSymL, index, loc);
TIntermTyped* rhsElement = intermediate.addIndex(EOpIndexDirect, tempSymR, index, loc);
const TType derefType(right->getType(), 0);
lhsElement->setType(derefType);
rhsElement->setType(derefType);
TIntermTyped* yNeg = intermediate.addUnaryMath(EOpNegative, rhsElement, loc);
assignList = intermediate.growAggregate(assignList, intermediate.addAssign(EOpAssign, lhsElement, yNeg, loc));
}
// Assign the rhs temp (now with Y inversion) to the final output
{
TIntermTyped* rhsTempSym = intermediate.addSymbol(*rhsTempVar, loc);
assignList = intermediate.growAggregate(assignList, intermediate.addAssign(op, left, rhsTempSym, loc));
}
assert(assignList != nullptr);
assignList->setOperator(EOpSequence);
return assignList;
}
// Clip and cull distance require special handling due to a semantic mismatch. In HLSL,
// these can be float scalar, float vector, or arrays of float scalar or float vector.
// In SPIR-V, they are arrays of scalar floats in all cases. We must copy individual components
// (e.g, both x and y components of a float2) out into the destination float array.
//
// The values are assigned to sequential members of the output array. The inner dimension
// is vector components. The outer dimension is array elements.
TIntermAggregate* HlslParseContext::assignClipCullDistance(const TSourceLoc& loc, TOperator op, int semanticId,
TIntermTyped* left, TIntermTyped* right)
{
switch (language) {
case EShLangFragment:
case EShLangVertex:
case EShLangGeometry:
break;
default:
error(loc, "unimplemented: clip/cull not currently implemented for this stage", "", "");
return nullptr;
}
TVariable** clipCullVar = nullptr;
// Figure out if we are assigning to, or from, clip or cull distance.
const bool isOutput = isClipOrCullDistance(left->getType());
// This is the rvalue or lvalue holding the clip or cull distance.
TIntermTyped* clipCullNode = isOutput ? left : right;
// This is the value going into or out of the clip or cull distance.
TIntermTyped* internalNode = isOutput ? right : left;
const TBuiltInVariable builtInType = clipCullNode->getQualifier().builtIn;
decltype(clipSemanticNSizeIn)* semanticNSize = nullptr;
// Refer to either the clip or the cull distance, depending on semantic.
switch (builtInType) {
case EbvClipDistance:
clipCullVar = isOutput ? &clipDistanceOutput : &clipDistanceInput;
semanticNSize = isOutput ? &clipSemanticNSizeOut : &clipSemanticNSizeIn;
break;
case EbvCullDistance:
clipCullVar = isOutput ? &cullDistanceOutput : &cullDistanceInput;
semanticNSize = isOutput ? &cullSemanticNSizeOut : &cullSemanticNSizeIn;
break;
// called invalidly: we expected a clip or a cull distance.
// static compile time problem: should not happen.
default: assert(0); return nullptr;
}
// This is the offset in the destination array of a given semantic's data
std::array<int, maxClipCullRegs> semanticOffset;
// Calculate offset of variable of semantic N in destination array
int arrayLoc = 0;
int vecItems = 0;
for (int x = 0; x < maxClipCullRegs; ++x) {
// See if we overflowed the vec4 packing
if ((vecItems + (*semanticNSize)[x]) > 4) {
arrayLoc = (arrayLoc + 3) & (~0x3); // round up to next multiple of 4
vecItems = 0;
}
semanticOffset[x] = arrayLoc;
vecItems += (*semanticNSize)[x];
arrayLoc += (*semanticNSize)[x];
}
// It can have up to 2 array dimensions (in the case of geometry shader inputs)
const TArraySizes* const internalArraySizes = internalNode->getType().getArraySizes();
const int internalArrayDims = internalNode->getType().isArray() ? internalArraySizes->getNumDims() : 0;
// vector sizes:
const int internalVectorSize = internalNode->getType().getVectorSize();
// array sizes, or 1 if it's not an array:
const int internalInnerArraySize = (internalArrayDims > 0 ? internalArraySizes->getDimSize(internalArrayDims-1) : 1);
const int internalOuterArraySize = (internalArrayDims > 1 ? internalArraySizes->getDimSize(0) : 1);
// The created type may be an array of arrays, e.g, for geometry shader inputs.
const bool isImplicitlyArrayed = (language == EShLangGeometry && !isOutput);
// If we haven't created the output already, create it now.
if (*clipCullVar == nullptr) {
// ClipDistance and CullDistance are handled specially in the entry point input/output copy
// algorithm, because they may need to be unpacked from components of vectors (or a scalar)
// into a float array, or vice versa. Here, we make the array the right size and type,
// which depends on the incoming data, which has several potential dimensions:
// * Semantic ID
// * vector size
// * array size
// Of those, semantic ID and array size cannot appear simultaneously.
//
// Also to note: for implicitly arrayed forms (e.g, geometry shader inputs), we need to create two
// array dimensions. The shader's declaration may have one or two array dimensions. One is always
// the geometry's dimension.
const bool useInnerSize = internalArrayDims > 1 || !isImplicitlyArrayed;
const int requiredInnerArraySize = arrayLoc * (useInnerSize ? internalInnerArraySize : 1);
const int requiredOuterArraySize = (internalArrayDims > 0) ? internalArraySizes->getDimSize(0) : 1;
TType clipCullType(EbtFloat, clipCullNode->getType().getQualifier().storage, 1);
clipCullType.getQualifier() = clipCullNode->getType().getQualifier();
// Create required array dimension
TArraySizes* arraySizes = new TArraySizes;
if (isImplicitlyArrayed)
arraySizes->addInnerSize(requiredOuterArraySize);
arraySizes->addInnerSize(requiredInnerArraySize);
clipCullType.transferArraySizes(arraySizes);
// Obtain symbol name: we'll use that for the symbol we introduce.
TIntermSymbol* sym = clipCullNode->getAsSymbolNode();
assert(sym != nullptr);
// We are moving the semantic ID from the layout location, so it is no longer needed or
// desired there.
clipCullType.getQualifier().layoutLocation = TQualifier::layoutLocationEnd;
// Create variable and track its linkage
*clipCullVar = makeInternalVariable(sym->getName().c_str(), clipCullType);
trackLinkage(**clipCullVar);
}
// Create symbol for the clip or cull variable.
TIntermSymbol* clipCullSym = intermediate.addSymbol(**clipCullVar);
// vector sizes:
const int clipCullVectorSize = clipCullSym->getType().getVectorSize();
// array sizes, or 1 if it's not an array:
const TArraySizes* const clipCullArraySizes = clipCullSym->getType().getArraySizes();
const int clipCullOuterArraySize = isImplicitlyArrayed ? clipCullArraySizes->getDimSize(0) : 1;
const int clipCullInnerArraySize = clipCullArraySizes->getDimSize(isImplicitlyArrayed ? 1 : 0);
// clipCullSym has got to be an array of scalar floats, per SPIR-V semantics.
// fixBuiltInIoType() should have handled that upstream.
assert(clipCullSym->getType().isArray());
assert(clipCullSym->getType().getVectorSize() == 1);
assert(clipCullSym->getType().getBasicType() == EbtFloat);
// We may be creating multiple sub-assignments. This is an aggregate to hold them.
// TODO: it would be possible to be clever sometimes and avoid the sequence node if not needed.
TIntermAggregate* assignList = nullptr;
// Holds individual component assignments as we make them.
TIntermTyped* clipCullAssign = nullptr;
// If the types are homomorphic, use a simple assign. No need to mess about with
// individual components.
if (clipCullSym->getType().isArray() == internalNode->getType().isArray() &&
clipCullInnerArraySize == internalInnerArraySize &&
clipCullOuterArraySize == internalOuterArraySize &&
clipCullVectorSize == internalVectorSize) {
if (isOutput)
clipCullAssign = intermediate.addAssign(op, clipCullSym, internalNode, loc);
else
clipCullAssign = intermediate.addAssign(op, internalNode, clipCullSym, loc);
assignList = intermediate.growAggregate(assignList, clipCullAssign);
assignList->setOperator(EOpSequence);
return assignList;
}
// We are going to copy each component of the internal (per array element if indicated) to sequential
// array elements of the clipCullSym. This tracks the lhs element we're writing to as we go along.
// We may be starting in the middle - e.g, for a non-zero semantic ID calculated above.
int clipCullInnerArrayPos = semanticOffset[semanticId];
int clipCullOuterArrayPos = 0;
// Lambda to add an index to a node, set the type of the result, and return the new node.
const auto addIndex = [this, &loc](TIntermTyped* node, int pos) -> TIntermTyped* {
const TType derefType(node->getType(), 0);
node = intermediate.addIndex(EOpIndexDirect, node, intermediate.addConstantUnion(pos, loc), loc);
node->setType(derefType);
return node;
};
// Loop through every component of every element of the internal, and copy to or from the matching external.
for (int internalOuterArrayPos = 0; internalOuterArrayPos < internalOuterArraySize; ++internalOuterArrayPos) {
for (int internalInnerArrayPos = 0; internalInnerArrayPos < internalInnerArraySize; ++internalInnerArrayPos) {
for (int internalComponent = 0; internalComponent < internalVectorSize; ++internalComponent) {
// clip/cull array member to read from / write to:
TIntermTyped* clipCullMember = clipCullSym;
// If implicitly arrayed, there is an outer array dimension involved
if (isImplicitlyArrayed)
clipCullMember = addIndex(clipCullMember, clipCullOuterArrayPos);
// Index into proper array position for clip cull member
clipCullMember = addIndex(clipCullMember, clipCullInnerArrayPos++);
// if needed, start over with next outer array slice.
if (isImplicitlyArrayed && clipCullInnerArrayPos >= clipCullInnerArraySize) {
clipCullInnerArrayPos = semanticOffset[semanticId];
++clipCullOuterArrayPos;
}
// internal member to read from / write to:
TIntermTyped* internalMember = internalNode;
// If internal node has outer array dimension, index appropriately.
if (internalArrayDims > 1)
internalMember = addIndex(internalMember, internalOuterArrayPos);
// If internal node has inner array dimension, index appropriately.
if (internalArrayDims > 0)
internalMember = addIndex(internalMember, internalInnerArrayPos);
// If internal node is a vector, extract the component of interest.
if (internalNode->getType().isVector())
internalMember = addIndex(internalMember, internalComponent);
// Create an assignment: output from internal to clip cull, or input from clip cull to internal.
if (isOutput)
clipCullAssign = intermediate.addAssign(op, clipCullMember, internalMember, loc);
else
clipCullAssign = intermediate.addAssign(op, internalMember, clipCullMember, loc);
// Track assignment in the sequence.
assignList = intermediate.growAggregate(assignList, clipCullAssign);
}
}
}
assert(assignList != nullptr);
assignList->setOperator(EOpSequence);
return assignList;
}
// Some simple source assignments need to be flattened to a sequence
// of AST assignments. Catch these and flatten, otherwise, pass through
// to intermediate.addAssign().
//
// Also, assignment to matrix swizzles requires multiple component assignments,
// intercept those as well.
TIntermTyped* HlslParseContext::handleAssign(const TSourceLoc& loc, TOperator op, TIntermTyped* left,
TIntermTyped* right)
{
if (left == nullptr || right == nullptr)
return nullptr;
// writing to opaques will require fixing transforms
if (left->getType().containsOpaque())
intermediate.setNeedsLegalization();
if (left->getAsOperator() && left->getAsOperator()->getOp() == EOpMatrixSwizzle)
return handleAssignToMatrixSwizzle(loc, op, left, right);
// Return true if the given node is an index operation into a split variable.
const auto indexesSplit = [this](const TIntermTyped* node) -> bool {
const TIntermBinary* binaryNode = node->getAsBinaryNode();
if (binaryNode == nullptr)
return false;
return (binaryNode->getOp() == EOpIndexDirect || binaryNode->getOp() == EOpIndexIndirect) &&
wasSplit(binaryNode->getLeft());
};
// Return true if this stage assigns clip position with potentially inverted Y
const auto assignsClipPos = [this](const TIntermTyped* node) -> bool {
return node->getType().getQualifier().builtIn == EbvPosition &&
(language == EShLangVertex || language == EShLangGeometry || language == EShLangTessEvaluation);
};
const bool isSplitLeft = wasSplit(left) || indexesSplit(left);
const bool isSplitRight = wasSplit(right) || indexesSplit(right);
const bool isFlattenLeft = wasFlattened(left);
const bool isFlattenRight = wasFlattened(right);
// OK to do a single assign if neither side is split or flattened. Otherwise,
// fall through to a member-wise copy.
if (!isFlattenLeft && !isFlattenRight && !isSplitLeft && !isSplitRight) {
// Clip and cull distance requires more processing. See comment above assignClipCullDistance.
if (isClipOrCullDistance(left->getType()) || isClipOrCullDistance(right->getType())) {
const bool isOutput = isClipOrCullDistance(left->getType());
const int semanticId = (isOutput ? left : right)->getType().getQualifier().layoutLocation;
return assignClipCullDistance(loc, op, semanticId, left, right);
} else if (assignsClipPos(left)) {
// Position can require special handling: see comment above assignPosition
return assignPosition(loc, op, left, right);
} else if (left->getQualifier().builtIn == EbvSampleMask) {
// Certain builtins are required to be arrayed outputs in SPIR-V, but may internally be scalars
// in the shader. Copy the scalar RHS into the LHS array element zero, if that happens.
if (left->isArray() && !right->isArray()) {
const TType derefType(left->getType(), 0);
left = intermediate.addIndex(EOpIndexDirect, left, intermediate.addConstantUnion(0, loc), loc);
left->setType(derefType);
// Fall through to add assign.
}
}
return intermediate.addAssign