// // Copyright (C) 2014-2015 LunarG, Inc. // Copyright (C) 2015-2018 Google, 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. // // Helper for making SPIR-V IR. Generally, this is documented in the header // SpvBuilder.h. // #include #include #include #include #include "SpvBuilder.h" #include "hex_float.h" #ifndef _WIN32 #include #endif namespace spv { Builder::Builder(unsigned int spvVersion, unsigned int magicNumber, SpvBuildLogger* buildLogger) : spvVersion(spvVersion), source(SourceLanguageUnknown), sourceVersion(0), sourceFileStringId(NoResult), currentLine(0), currentFile(nullptr), emitOpLines(false), addressModel(AddressingModelLogical), memoryModel(MemoryModelGLSL450), builderNumber(magicNumber), buildPoint(0), uniqueId(0), entryPointFunction(0), generatingOpCodeForSpecConst(false), logger(buildLogger) { clearAccessChain(); } Builder::~Builder() { } Id Builder::import(const char* name) { Instruction* import = new Instruction(getUniqueId(), NoType, OpExtInstImport); import->addStringOperand(name); module.mapInstruction(import); imports.push_back(std::unique_ptr(import)); return import->getResultId(); } // Emit instruction for non-filename-based #line directives (ie. no filename // seen yet): emit an OpLine if we've been asked to emit OpLines and the line // number has changed since the last time, and is a valid line number. void Builder::setLine(int lineNum) { if (lineNum != 0 && lineNum != currentLine) { currentLine = lineNum; if (emitOpLines) addLine(sourceFileStringId, currentLine, 0); } } // If no filename, do non-filename-based #line emit. Else do filename-based emit. // Emit OpLine if we've been asked to emit OpLines and the line number or filename // has changed since the last time, and line number is valid. void Builder::setLine(int lineNum, const char* filename) { if (filename == nullptr) { setLine(lineNum); return; } if ((lineNum != 0 && lineNum != currentLine) || currentFile == nullptr || strncmp(filename, currentFile, strlen(currentFile) + 1) != 0) { currentLine = lineNum; currentFile = filename; if (emitOpLines) { spv::Id strId = getStringId(filename); addLine(strId, currentLine, 0); } } } void Builder::addLine(Id fileName, int lineNum, int column) { Instruction* line = new Instruction(OpLine); line->addIdOperand(fileName); line->addImmediateOperand(lineNum); line->addImmediateOperand(column); buildPoint->addInstruction(std::unique_ptr(line)); } // For creating new groupedTypes (will return old type if the requested one was already made). Id Builder::makeVoidType() { Instruction* type; if (groupedTypes[OpTypeVoid].size() == 0) { type = new Instruction(getUniqueId(), NoType, OpTypeVoid); groupedTypes[OpTypeVoid].push_back(type); constantsTypesGlobals.push_back(std::unique_ptr(type)); module.mapInstruction(type); } else type = groupedTypes[OpTypeVoid].back(); return type->getResultId(); } Id Builder::makeBoolType() { Instruction* type; if (groupedTypes[OpTypeBool].size() == 0) { type = new Instruction(getUniqueId(), NoType, OpTypeBool); groupedTypes[OpTypeBool].push_back(type); constantsTypesGlobals.push_back(std::unique_ptr(type)); module.mapInstruction(type); } else type = groupedTypes[OpTypeBool].back(); return type->getResultId(); } Id Builder::makeSamplerType() { Instruction* type; if (groupedTypes[OpTypeSampler].size() == 0) { type = new Instruction(getUniqueId(), NoType, OpTypeSampler); groupedTypes[OpTypeSampler].push_back(type); constantsTypesGlobals.push_back(std::unique_ptr(type)); module.mapInstruction(type); } else type = groupedTypes[OpTypeSampler].back(); return type->getResultId(); } Id Builder::makePointer(StorageClass storageClass, Id pointee) { // try to find it Instruction* type; for (int t = 0; t < (int)groupedTypes[OpTypePointer].size(); ++t) { type = groupedTypes[OpTypePointer][t]; if (type->getImmediateOperand(0) == (unsigned)storageClass && type->getIdOperand(1) == pointee) return type->getResultId(); } // not found, make it type = new Instruction(getUniqueId(), NoType, OpTypePointer); type->addImmediateOperand(storageClass); type->addIdOperand(pointee); groupedTypes[OpTypePointer].push_back(type); constantsTypesGlobals.push_back(std::unique_ptr(type)); module.mapInstruction(type); return type->getResultId(); } Id Builder::makeForwardPointer(StorageClass storageClass) { // Caching/uniquifying doesn't work here, because we don't know the // pointee type and there can be multiple forward pointers of the same // storage type. Somebody higher up in the stack must keep track. Instruction* type = new Instruction(getUniqueId(), NoType, OpTypeForwardPointer); type->addImmediateOperand(storageClass); constantsTypesGlobals.push_back(std::unique_ptr(type)); module.mapInstruction(type); return type->getResultId(); } Id Builder::makePointerFromForwardPointer(StorageClass storageClass, Id forwardPointerType, Id pointee) { // try to find it Instruction* type; for (int t = 0; t < (int)groupedTypes[OpTypePointer].size(); ++t) { type = groupedTypes[OpTypePointer][t]; if (type->getImmediateOperand(0) == (unsigned)storageClass && type->getIdOperand(1) == pointee) return type->getResultId(); } type = new Instruction(forwardPointerType, NoType, OpTypePointer); type->addImmediateOperand(storageClass); type->addIdOperand(pointee); groupedTypes[OpTypePointer].push_back(type); constantsTypesGlobals.push_back(std::unique_ptr(type)); module.mapInstruction(type); return type->getResultId(); } Id Builder::makeIntegerType(int width, bool hasSign) { // try to find it Instruction* type; for (int t = 0; t < (int)groupedTypes[OpTypeInt].size(); ++t) { type = groupedTypes[OpTypeInt][t]; if (type->getImmediateOperand(0) == (unsigned)width && type->getImmediateOperand(1) == (hasSign ? 1u : 0u)) return type->getResultId(); } // not found, make it type = new Instruction(getUniqueId(), NoType, OpTypeInt); type->addImmediateOperand(width); type->addImmediateOperand(hasSign ? 1 : 0); groupedTypes[OpTypeInt].push_back(type); constantsTypesGlobals.push_back(std::unique_ptr(type)); module.mapInstruction(type); // deal with capabilities switch (width) { case 8: case 16: // these are currently handled by storage-type declarations and post processing break; case 64: addCapability(CapabilityInt64); break; default: break; } return type->getResultId(); } Id Builder::makeFloatType(int width) { // try to find it Instruction* type; for (int t = 0; t < (int)groupedTypes[OpTypeFloat].size(); ++t) { type = groupedTypes[OpTypeFloat][t]; if (type->getImmediateOperand(0) == (unsigned)width) return type->getResultId(); } // not found, make it type = new Instruction(getUniqueId(), NoType, OpTypeFloat); type->addImmediateOperand(width); groupedTypes[OpTypeFloat].push_back(type); constantsTypesGlobals.push_back(std::unique_ptr(type)); module.mapInstruction(type); // deal with capabilities switch (width) { case 16: // currently handled by storage-type declarations and post processing break; case 64: addCapability(CapabilityFloat64); break; default: break; } return type->getResultId(); } // Make a struct without checking for duplication. // See makeStructResultType() for non-decorated structs // needed as the result of some instructions, which does // check for duplicates. Id Builder::makeStructType(const std::vector& members, const char* name) { // Don't look for previous one, because in the general case, // structs can be duplicated except for decorations. // not found, make it Instruction* type = new Instruction(getUniqueId(), NoType, OpTypeStruct); for (int op = 0; op < (int)members.size(); ++op) type->addIdOperand(members[op]); groupedTypes[OpTypeStruct].push_back(type); constantsTypesGlobals.push_back(std::unique_ptr(type)); module.mapInstruction(type); addName(type->getResultId(), name); return type->getResultId(); } // Make a struct for the simple results of several instructions, // checking for duplication. Id Builder::makeStructResultType(Id type0, Id type1) { // try to find it Instruction* type; for (int t = 0; t < (int)groupedTypes[OpTypeStruct].size(); ++t) { type = groupedTypes[OpTypeStruct][t]; if (type->getNumOperands() != 2) continue; if (type->getIdOperand(0) != type0 || type->getIdOperand(1) != type1) continue; return type->getResultId(); } // not found, make it std::vector members; members.push_back(type0); members.push_back(type1); return makeStructType(members, "ResType"); } Id Builder::makeVectorType(Id component, int size) { // try to find it Instruction* type; for (int t = 0; t < (int)groupedTypes[OpTypeVector].size(); ++t) { type = groupedTypes[OpTypeVector][t]; if (type->getIdOperand(0) == component && type->getImmediateOperand(1) == (unsigned)size) return type->getResultId(); } // not found, make it type = new Instruction(getUniqueId(), NoType, OpTypeVector); type->addIdOperand(component); type->addImmediateOperand(size); groupedTypes[OpTypeVector].push_back(type); constantsTypesGlobals.push_back(std::unique_ptr(type)); module.mapInstruction(type); return type->getResultId(); } Id Builder::makeMatrixType(Id component, int cols, int rows) { assert(cols <= maxMatrixSize && rows <= maxMatrixSize); Id column = makeVectorType(component, rows); // try to find it Instruction* type; for (int t = 0; t < (int)groupedTypes[OpTypeMatrix].size(); ++t) { type = groupedTypes[OpTypeMatrix][t]; if (type->getIdOperand(0) == column && type->getImmediateOperand(1) == (unsigned)cols) return type->getResultId(); } // not found, make it type = new Instruction(getUniqueId(), NoType, OpTypeMatrix); type->addIdOperand(column); type->addImmediateOperand(cols); groupedTypes[OpTypeMatrix].push_back(type); constantsTypesGlobals.push_back(std::unique_ptr(type)); module.mapInstruction(type); return type->getResultId(); } Id Builder::makeCooperativeMatrixType(Id component, Id scope, Id rows, Id cols) { // try to find it Instruction* type; for (int t = 0; t < (int)groupedTypes[OpTypeCooperativeMatrixNV].size(); ++t) { type = groupedTypes[OpTypeCooperativeMatrixNV][t]; if (type->getIdOperand(0) == component && type->getIdOperand(1) == scope && type->getIdOperand(2) == rows && type->getIdOperand(3) == cols) return type->getResultId(); } // not found, make it type = new Instruction(getUniqueId(), NoType, OpTypeCooperativeMatrixNV); type->addIdOperand(component); type->addIdOperand(scope); type->addIdOperand(rows); type->addIdOperand(cols); groupedTypes[OpTypeCooperativeMatrixNV].push_back(type); constantsTypesGlobals.push_back(std::unique_ptr(type)); module.mapInstruction(type); return type->getResultId(); } // TODO: performance: track arrays per stride // If a stride is supplied (non-zero) make an array. // If no stride (0), reuse previous array types. // 'size' is an Id of a constant or specialization constant of the array size Id Builder::makeArrayType(Id element, Id sizeId, int stride) { Instruction* type; if (stride == 0) { // try to find existing type for (int t = 0; t < (int)groupedTypes[OpTypeArray].size(); ++t) { type = groupedTypes[OpTypeArray][t]; if (type->getIdOperand(0) == element && type->getIdOperand(1) == sizeId) return type->getResultId(); } } // not found, make it type = new Instruction(getUniqueId(), NoType, OpTypeArray); type->addIdOperand(element); type->addIdOperand(sizeId); groupedTypes[OpTypeArray].push_back(type); constantsTypesGlobals.push_back(std::unique_ptr(type)); module.mapInstruction(type); return type->getResultId(); } Id Builder::makeRuntimeArray(Id element) { Instruction* type = new Instruction(getUniqueId(), NoType, OpTypeRuntimeArray); type->addIdOperand(element); constantsTypesGlobals.push_back(std::unique_ptr(type)); module.mapInstruction(type); return type->getResultId(); } Id Builder::makeFunctionType(Id returnType, const std::vector& paramTypes) { // try to find it Instruction* type; for (int t = 0; t < (int)groupedTypes[OpTypeFunction].size(); ++t) { type = groupedTypes[OpTypeFunction][t]; if (type->getIdOperand(0) != returnType || (int)paramTypes.size() != type->getNumOperands() - 1) continue; bool mismatch = false; for (int p = 0; p < (int)paramTypes.size(); ++p) { if (paramTypes[p] != type->getIdOperand(p + 1)) { mismatch = true; break; } } if (! mismatch) return type->getResultId(); } // not found, make it type = new Instruction(getUniqueId(), NoType, OpTypeFunction); type->addIdOperand(returnType); for (int p = 0; p < (int)paramTypes.size(); ++p) type->addIdOperand(paramTypes[p]); groupedTypes[OpTypeFunction].push_back(type); constantsTypesGlobals.push_back(std::unique_ptr(type)); module.mapInstruction(type); return type->getResultId(); } Id Builder::makeImageType(Id sampledType, Dim dim, bool depth, bool arrayed, bool ms, unsigned sampled, ImageFormat format) { assert(sampled == 1 || sampled == 2); // try to find it Instruction* type; for (int t = 0; t < (int)groupedTypes[OpTypeImage].size(); ++t) { type = groupedTypes[OpTypeImage][t]; if (type->getIdOperand(0) == sampledType && type->getImmediateOperand(1) == (unsigned int)dim && type->getImmediateOperand(2) == ( depth ? 1u : 0u) && type->getImmediateOperand(3) == (arrayed ? 1u : 0u) && type->getImmediateOperand(4) == ( ms ? 1u : 0u) && type->getImmediateOperand(5) == sampled && type->getImmediateOperand(6) == (unsigned int)format) return type->getResultId(); } // not found, make it type = new Instruction(getUniqueId(), NoType, OpTypeImage); type->addIdOperand(sampledType); type->addImmediateOperand( dim); type->addImmediateOperand( depth ? 1 : 0); type->addImmediateOperand(arrayed ? 1 : 0); type->addImmediateOperand( ms ? 1 : 0); type->addImmediateOperand(sampled); type->addImmediateOperand((unsigned int)format); groupedTypes[OpTypeImage].push_back(type); constantsTypesGlobals.push_back(std::unique_ptr(type)); module.mapInstruction(type); // deal with capabilities switch (dim) { case DimBuffer: if (sampled == 1) addCapability(CapabilitySampledBuffer); else addCapability(CapabilityImageBuffer); break; case Dim1D: if (sampled == 1) addCapability(CapabilitySampled1D); else addCapability(CapabilityImage1D); break; case DimCube: if (arrayed) { if (sampled == 1) addCapability(CapabilitySampledCubeArray); else addCapability(CapabilityImageCubeArray); } break; case DimRect: if (sampled == 1) addCapability(CapabilitySampledRect); else addCapability(CapabilityImageRect); break; case DimSubpassData: addCapability(CapabilityInputAttachment); break; default: break; } if (ms) { if (sampled == 2) { // Images used with subpass data are not storage // images, so don't require the capability for them. if (dim != Dim::DimSubpassData) addCapability(CapabilityStorageImageMultisample); if (arrayed) addCapability(CapabilityImageMSArray); } } return type->getResultId(); } Id Builder::makeSampledImageType(Id imageType) { // try to find it Instruction* type; for (int t = 0; t < (int)groupedTypes[OpTypeSampledImage].size(); ++t) { type = groupedTypes[OpTypeSampledImage][t]; if (type->getIdOperand(0) == imageType) return type->getResultId(); } // not found, make it type = new Instruction(getUniqueId(), NoType, OpTypeSampledImage); type->addIdOperand(imageType); groupedTypes[OpTypeSampledImage].push_back(type); constantsTypesGlobals.push_back(std::unique_ptr(type)); module.mapInstruction(type); return type->getResultId(); } #ifdef NV_EXTENSIONS Id Builder::makeAccelerationStructureNVType() { Instruction *type; if (groupedTypes[OpTypeAccelerationStructureNV].size() == 0) { type = new Instruction(getUniqueId(), NoType, OpTypeAccelerationStructureNV); groupedTypes[OpTypeAccelerationStructureNV].push_back(type); constantsTypesGlobals.push_back(std::unique_ptr(type)); module.mapInstruction(type); } else { type = groupedTypes[OpTypeAccelerationStructureNV].back(); } return type->getResultId(); } #endif Id Builder::getDerefTypeId(Id resultId) const { Id typeId = getTypeId(resultId); assert(isPointerType(typeId)); return module.getInstruction(typeId)->getIdOperand(1); } Op Builder::getMostBasicTypeClass(Id typeId) const { Instruction* instr = module.getInstruction(typeId); Op typeClass = instr->getOpCode(); switch (typeClass) { case OpTypeVector: case OpTypeMatrix: case OpTypeArray: case OpTypeRuntimeArray: return getMostBasicTypeClass(instr->getIdOperand(0)); case OpTypePointer: return getMostBasicTypeClass(instr->getIdOperand(1)); default: return typeClass; } } int Builder::getNumTypeConstituents(Id typeId) const { Instruction* instr = module.getInstruction(typeId); switch (instr->getOpCode()) { case OpTypeBool: case OpTypeInt: case OpTypeFloat: case OpTypePointer: return 1; case OpTypeVector: case OpTypeMatrix: return instr->getImmediateOperand(1); case OpTypeArray: { Id lengthId = instr->getIdOperand(1); return module.getInstruction(lengthId)->getImmediateOperand(0); } case OpTypeStruct: return instr->getNumOperands(); case OpTypeCooperativeMatrixNV: // has only one constituent when used with OpCompositeConstruct. return 1; default: assert(0); return 1; } } // Return the lowest-level type of scalar that an homogeneous composite is made out of. // Typically, this is just to find out if something is made out of ints or floats. // However, it includes returning a structure, if say, it is an array of structure. Id Builder::getScalarTypeId(Id typeId) const { Instruction* instr = module.getInstruction(typeId); Op typeClass = instr->getOpCode(); switch (typeClass) { case OpTypeVoid: case OpTypeBool: case OpTypeInt: case OpTypeFloat: case OpTypeStruct: return instr->getResultId(); case OpTypeVector: case OpTypeMatrix: case OpTypeArray: case OpTypeRuntimeArray: case OpTypePointer: return getScalarTypeId(getContainedTypeId(typeId)); default: assert(0); return NoResult; } } // Return the type of 'member' of a composite. Id Builder::getContainedTypeId(Id typeId, int member) const { Instruction* instr = module.getInstruction(typeId); Op typeClass = instr->getOpCode(); switch (typeClass) { case OpTypeVector: case OpTypeMatrix: case OpTypeArray: case OpTypeRuntimeArray: case OpTypeCooperativeMatrixNV: return instr->getIdOperand(0); case OpTypePointer: return instr->getIdOperand(1); case OpTypeStruct: return instr->getIdOperand(member); default: assert(0); return NoResult; } } // Return the immediately contained type of a given composite type. Id Builder::getContainedTypeId(Id typeId) const { return getContainedTypeId(typeId, 0); } // Returns true if 'typeId' is or contains a scalar type declared with 'typeOp' // of width 'width'. The 'width' is only consumed for int and float types. // Returns false otherwise. bool Builder::containsType(Id typeId, spv::Op typeOp, unsigned int width) const { const Instruction& instr = *module.getInstruction(typeId); Op typeClass = instr.getOpCode(); switch (typeClass) { case OpTypeInt: case OpTypeFloat: return typeClass == typeOp && instr.getImmediateOperand(0) == width; case OpTypeStruct: for (int m = 0; m < instr.getNumOperands(); ++m) { if (containsType(instr.getIdOperand(m), typeOp, width)) return true; } return false; case OpTypePointer: return false; case OpTypeVector: case OpTypeMatrix: case OpTypeArray: case OpTypeRuntimeArray: return containsType(getContainedTypeId(typeId), typeOp, width); default: return typeClass == typeOp; } } // return true if the type is a pointer to PhysicalStorageBufferEXT or an // array of such pointers. These require restrict/aliased decorations. bool Builder::containsPhysicalStorageBufferOrArray(Id typeId) const { const Instruction& instr = *module.getInstruction(typeId); Op typeClass = instr.getOpCode(); switch (typeClass) { case OpTypePointer: return getTypeStorageClass(typeId) == StorageClassPhysicalStorageBufferEXT; case OpTypeArray: return containsPhysicalStorageBufferOrArray(getContainedTypeId(typeId)); default: return false; } } // See if a scalar constant of this type has already been created, so it // can be reused rather than duplicated. (Required by the specification). Id Builder::findScalarConstant(Op typeClass, Op opcode, Id typeId, unsigned value) { Instruction* constant; for (int i = 0; i < (int)groupedConstants[typeClass].size(); ++i) { constant = groupedConstants[typeClass][i]; if (constant->getOpCode() == opcode && constant->getTypeId() == typeId && constant->getImmediateOperand(0) == value) return constant->getResultId(); } return 0; } // Version of findScalarConstant (see above) for scalars that take two operands (e.g. a 'double' or 'int64'). Id Builder::findScalarConstant(Op typeClass, Op opcode, Id typeId, unsigned v1, unsigned v2) { Instruction* constant; for (int i = 0; i < (int)groupedConstants[typeClass].size(); ++i) { constant = groupedConstants[typeClass][i]; if (constant->getOpCode() == opcode && constant->getTypeId() == typeId && constant->getImmediateOperand(0) == v1 && constant->getImmediateOperand(1) == v2) return constant->getResultId(); } return 0; } // Return true if consuming 'opcode' means consuming a constant. // "constant" here means after final transform to executable code, // the value consumed will be a constant, so includes specialization. bool Builder::isConstantOpCode(Op opcode) const { switch (opcode) { case OpUndef: case OpConstantTrue: case OpConstantFalse: case OpConstant: case OpConstantComposite: case OpConstantSampler: case OpConstantNull: case OpSpecConstantTrue: case OpSpecConstantFalse: case OpSpecConstant: case OpSpecConstantComposite: case OpSpecConstantOp: return true; default: return false; } } // Return true if consuming 'opcode' means consuming a specialization constant. bool Builder::isSpecConstantOpCode(Op opcode) const { switch (opcode) { case OpSpecConstantTrue: case OpSpecConstantFalse: case OpSpecConstant: case OpSpecConstantComposite: case OpSpecConstantOp: return true; default: return false; } } Id Builder::makeBoolConstant(bool b, bool specConstant) { Id typeId = makeBoolType(); Instruction* constant; Op opcode = specConstant ? (b ? OpSpecConstantTrue : OpSpecConstantFalse) : (b ? OpConstantTrue : OpConstantFalse); // See if we already made it. Applies only to regular constants, because specialization constants // must remain distinct for the purpose of applying a SpecId decoration. if (! specConstant) { Id existing = 0; for (int i = 0; i < (int)groupedConstants[OpTypeBool].size(); ++i) { constant = groupedConstants[OpTypeBool][i]; if (constant->getTypeId() == typeId && constant->getOpCode() == opcode) existing = constant->getResultId(); } if (existing) return existing; } // Make it Instruction* c = new Instruction(getUniqueId(), typeId, opcode); constantsTypesGlobals.push_back(std::unique_ptr(c)); groupedConstants[OpTypeBool].push_back(c); module.mapInstruction(c); return c->getResultId(); } Id Builder::makeIntConstant(Id typeId, unsigned value, bool specConstant) { Op opcode = specConstant ? OpSpecConstant : OpConstant; // See if we already made it. Applies only to regular constants, because specialization constants // must remain distinct for the purpose of applying a SpecId decoration. if (! specConstant) { Id existing = findScalarConstant(OpTypeInt, opcode, typeId, value); if (existing) return existing; } Instruction* c = new Instruction(getUniqueId(), typeId, opcode); c->addImmediateOperand(value); constantsTypesGlobals.push_back(std::unique_ptr(c)); groupedConstants[OpTypeInt].push_back(c); module.mapInstruction(c); return c->getResultId(); } Id Builder::makeInt64Constant(Id typeId, unsigned long long value, bool specConstant) { Op opcode = specConstant ? OpSpecConstant : OpConstant; unsigned op1 = value & 0xFFFFFFFF; unsigned op2 = value >> 32; // See if we already made it. Applies only to regular constants, because specialization constants // must remain distinct for the purpose of applying a SpecId decoration. if (! specConstant) { Id existing = findScalarConstant(OpTypeInt, opcode, typeId, op1, op2); if (existing) return existing; } Instruction* c = new Instruction(getUniqueId(), typeId, opcode); c->addImmediateOperand(op1); c->addImmediateOperand(op2); constantsTypesGlobals.push_back(std::unique_ptr(c)); groupedConstants[OpTypeInt].push_back(c); module.mapInstruction(c); return c->getResultId(); } Id Builder::makeFloatConstant(float f, bool specConstant) { Op opcode = specConstant ? OpSpecConstant : OpConstant; Id typeId = makeFloatType(32); union { float fl; unsigned int ui; } u; u.fl = f; unsigned value = u.ui; // See if we already made it. Applies only to regular constants, because specialization constants // must remain distinct for the purpose of applying a SpecId decoration. if (! specConstant) { Id existing = findScalarConstant(OpTypeFloat, opcode, typeId, value); if (existing) return existing; } Instruction* c = new Instruction(getUniqueId(), typeId, opcode); c->addImmediateOperand(value); constantsTypesGlobals.push_back(std::unique_ptr(c)); groupedConstants[OpTypeFloat].push_back(c); module.mapInstruction(c); return c->getResultId(); } Id Builder::makeDoubleConstant(double d, bool specConstant) { Op opcode = specConstant ? OpSpecConstant : OpConstant; Id typeId = makeFloatType(64); union { double db; unsigned long long ull; } u; u.db = d; unsigned long long value = u.ull; unsigned op1 = value & 0xFFFFFFFF; unsigned op2 = value >> 32; // See if we already made it. Applies only to regular constants, because specialization constants // must remain distinct for the purpose of applying a SpecId decoration. if (! specConstant) { Id existing = findScalarConstant(OpTypeFloat, opcode, typeId, op1, op2); if (existing) return existing; } Instruction* c = new Instruction(getUniqueId(), typeId, opcode); c->addImmediateOperand(op1); c->addImmediateOperand(op2); constantsTypesGlobals.push_back(std::unique_ptr(c)); groupedConstants[OpTypeFloat].push_back(c); module.mapInstruction(c); return c->getResultId(); } Id Builder::makeFloat16Constant(float f16, bool specConstant) { Op opcode = specConstant ? OpSpecConstant : OpConstant; Id typeId = makeFloatType(16); spvutils::HexFloat> fVal(f16); spvutils::HexFloat> f16Val(0); fVal.castTo(f16Val, spvutils::kRoundToZero); unsigned value = f16Val.value().getAsFloat().get_value(); // See if we already made it. Applies only to regular constants, because specialization constants // must remain distinct for the purpose of applying a SpecId decoration. if (!specConstant) { Id existing = findScalarConstant(OpTypeFloat, opcode, typeId, value); if (existing) return existing; } Instruction* c = new Instruction(getUniqueId(), typeId, opcode); c->addImmediateOperand(value); constantsTypesGlobals.push_back(std::unique_ptr(c)); groupedConstants[OpTypeFloat].push_back(c); module.mapInstruction(c); return c->getResultId(); } Id Builder::makeFpConstant(Id type, double d, bool specConstant) { assert(isFloatType(type)); switch (getScalarTypeWidth(type)) { case 16: return makeFloat16Constant((float)d, specConstant); case 32: return makeFloatConstant((float)d, specConstant); case 64: return makeDoubleConstant(d, specConstant); default: break; } assert(false); return NoResult; } Id Builder::findCompositeConstant(Op typeClass, Id typeId, const std::vector& comps) { Instruction* constant = 0; bool found = false; for (int i = 0; i < (int)groupedConstants[typeClass].size(); ++i) { constant = groupedConstants[typeClass][i]; if (constant->getTypeId() != typeId) continue; // same contents? bool mismatch = false; for (int op = 0; op < constant->getNumOperands(); ++op) { if (constant->getIdOperand(op) != comps[op]) { mismatch = true; break; } } if (! mismatch) { found = true; break; } } return found ? constant->getResultId() : NoResult; } Id Builder::findStructConstant(Id typeId, const std::vector& comps) { Instruction* constant = 0; bool found = false; for (int i = 0; i < (int)groupedStructConstants[typeId].size(); ++i) { constant = groupedStructConstants[typeId][i]; // same contents? bool mismatch = false; for (int op = 0; op < constant->getNumOperands(); ++op) { if (constant->getIdOperand(op) != comps[op]) { mismatch = true; break; } } if (! mismatch) { found = true; break; } } return found ? constant->getResultId() : NoResult; } // Comments in header Id Builder::makeCompositeConstant(Id typeId, const std::vector& members, bool specConstant) { Op opcode = specConstant ? OpSpecConstantComposite : OpConstantComposite; assert(typeId); Op typeClass = getTypeClass(typeId); switch (typeClass) { case OpTypeVector: case OpTypeArray: case OpTypeMatrix: case OpTypeCooperativeMatrixNV: if (! specConstant) { Id existing = findCompositeConstant(typeClass, typeId, members); if (existing) return existing; } break; case OpTypeStruct: if (! specConstant) { Id existing = findStructConstant(typeId, members); if (existing) return existing; } break; default: assert(0); return makeFloatConstant(0.0); } Instruction* c = new Instruction(getUniqueId(), typeId, opcode); for (int op = 0; op < (int)members.size(); ++op) c->addIdOperand(members[op]); constantsTypesGlobals.push_back(std::unique_ptr(c)); if (typeClass == OpTypeStruct) groupedStructConstants[typeId].push_back(c); else groupedConstants[typeClass].push_back(c); module.mapInstruction(c); return c->getResultId(); } Instruction* Builder::addEntryPoint(ExecutionModel model, Function* function, const char* name) { Instruction* entryPoint = new Instruction(OpEntryPoint); entryPoint->addImmediateOperand(model); entryPoint->addIdOperand(function->getId()); entryPoint->addStringOperand(name); entryPoints.push_back(std::unique_ptr(entryPoint)); return entryPoint; } // Currently relying on the fact that all 'value' of interest are small non-negative values. void Builder::addExecutionMode(Function* entryPoint, ExecutionMode mode, int value1, int value2, int value3) { Instruction* instr = new Instruction(OpExecutionMode); instr->addIdOperand(entryPoint->getId()); instr->addImmediateOperand(mode); if (value1 >= 0) instr->addImmediateOperand(value1); if (value2 >= 0) instr->addImmediateOperand(value2); if (value3 >= 0) instr->addImmediateOperand(value3); executionModes.push_back(std::unique_ptr(instr)); } void Builder::addName(Id id, const char* string) { Instruction* name = new Instruction(OpName); name->addIdOperand(id); name->addStringOperand(string); names.push_back(std::unique_ptr(name)); } void Builder::addMemberName(Id id, int memberNumber, const char* string) { Instruction* name = new Instruction(OpMemberName); name->addIdOperand(id); name->addImmediateOperand(memberNumber); name->addStringOperand(string); names.push_back(std::unique_ptr(name)); } void Builder::addDecoration(Id id, Decoration decoration, int num) { if (decoration == spv::DecorationMax) return; Instruction* dec = new Instruction(OpDecorate); dec->addIdOperand(id); dec->addImmediateOperand(decoration); if (num >= 0) dec->addImmediateOperand(num); decorations.push_back(std::unique_ptr(dec)); } void Builder::addDecoration(Id id, Decoration decoration, const char* s) { if (decoration == spv::DecorationMax) return; Instruction* dec = new Instruction(OpDecorateStringGOOGLE); dec->addIdOperand(id); dec->addImmediateOperand(decoration); dec->addStringOperand(s); decorations.push_back(std::unique_ptr(dec)); } void Builder::addDecorationId(Id id, Decoration decoration, Id idDecoration) { if (decoration == spv::DecorationMax) return; Instruction* dec = new Instruction(OpDecorateId); dec->addIdOperand(id); dec->addImmediateOperand(decoration); dec->addIdOperand(idDecoration); decorations.push_back(std::unique_ptr(dec)); } void Builder::addMemberDecoration(Id id, unsigned int member, Decoration decoration, int num) { if (decoration == spv::DecorationMax) return; Instruction* dec = new Instruction(OpMemberDecorate); dec->addIdOperand(id); dec->addImmediateOperand(member); dec->addImmediateOperand(decoration); if (num >= 0) dec->addImmediateOperand(num); decorations.push_back(std::unique_ptr(dec)); } void Builder::addMemberDecoration(Id id, unsigned int member, Decoration decoration, const char *s) { if (decoration == spv::DecorationMax) return; Instruction* dec = new Instruction(OpMemberDecorateStringGOOGLE); dec->addIdOperand(id); dec->addImmediateOperand(member); dec->addImmediateOperand(decoration); dec->addStringOperand(s); decorations.push_back(std::unique_ptr(dec)); } // Comments in header Function* Builder::makeEntryPoint(const char* entryPoint) { assert(! entryPointFunction); Block* entry; std::vector params; std::vector> decorations; entryPointFunction = makeFunctionEntry(NoPrecision, makeVoidType(), entryPoint, params, decorations, &entry); return entryPointFunction; } // Comments in header Function* Builder::makeFunctionEntry(Decoration precision, Id returnType, const char* name, const std::vector& paramTypes, const std::vector>& decorations, Block **entry) { // Make the function and initial instructions in it Id typeId = makeFunctionType(returnType, paramTypes); Id firstParamId = paramTypes.size() == 0 ? 0 : getUniqueIds((int)paramTypes.size()); Function* function = new Function(getUniqueId(), returnType, typeId, firstParamId, module); // Set up the precisions setPrecision(function->getId(), precision); for (unsigned p = 0; p < (unsigned)decorations.size(); ++p) { for (int d = 0; d < (int)decorations[p].size(); ++d) addDecoration(firstParamId + p, decorations[p][d]); } // CFG if (entry) { *entry = new Block(getUniqueId(), *function); function->addBlock(*entry); setBuildPoint(*entry); } if (name) addName(function->getId(), name); functions.push_back(std::unique_ptr(function)); return function; } // Comments in header void Builder::makeReturn(bool implicit, Id retVal) { if (retVal) { Instruction* inst = new Instruction(NoResult, NoType, OpReturnValue); inst->addIdOperand(retVal); buildPoint->addInstruction(std::unique_ptr(inst)); } else buildPoint->addInstruction(std::unique_ptr(new Instruction(NoResult, NoType, OpReturn))); if (! implicit) createAndSetNoPredecessorBlock("post-return"); } // Comments in header void Builder::leaveFunction() { Block* block = buildPoint; Function& function = buildPoint->getParent(); assert(block); // If our function did not contain a return, add a return void now. if (! block->isTerminated()) { if (function.getReturnType() == makeVoidType()) makeReturn(true); else { makeReturn(true, createUndefined(function.getReturnType())); } } } // Comments in header void Builder::makeDiscard() { buildPoint->addInstruction(std::unique_ptr(new Instruction(OpKill))); createAndSetNoPredecessorBlock("post-discard"); } // Comments in header Id Builder::createVariable(StorageClass storageClass, Id type, const char* name, Id initializer) { Id pointerType = makePointer(storageClass, type); Instruction* inst = new Instruction(getUniqueId(), pointerType, OpVariable); inst->addImmediateOperand(storageClass); if (initializer != NoResult) inst->addIdOperand(initializer); switch (storageClass) { case StorageClassFunction: // Validation rules require the declaration in the entry block buildPoint->getParent().addLocalVariable(std::unique_ptr(inst)); break; default: constantsTypesGlobals.push_back(std::unique_ptr(inst)); module.mapInstruction(inst); break; } if (name) addName(inst->getResultId(), name); return inst->getResultId(); } // Comments in header Id Builder::createUndefined(Id type) { Instruction* inst = new Instruction(getUniqueId(), type, OpUndef); buildPoint->addInstruction(std::unique_ptr(inst)); return inst->getResultId(); } // av/vis/nonprivate are unnecessary and illegal for some storage classes. spv::MemoryAccessMask Builder::sanitizeMemoryAccessForStorageClass(spv::MemoryAccessMask memoryAccess, StorageClass sc) const { switch (sc) { case spv::StorageClassUniform: case spv::StorageClassWorkgroup: case spv::StorageClassStorageBuffer: case spv::StorageClassPhysicalStorageBufferEXT: break; default: memoryAccess = spv::MemoryAccessMask(memoryAccess & ~(spv::MemoryAccessMakePointerAvailableKHRMask | spv::MemoryAccessMakePointerVisibleKHRMask | spv::MemoryAccessNonPrivatePointerKHRMask)); break; } return memoryAccess; } // Comments in header void Builder::createStore(Id rValue, Id lValue, spv::MemoryAccessMask memoryAccess, spv::Scope scope, unsigned int alignment) { Instruction* store = new Instruction(OpStore); store->addIdOperand(lValue); store->addIdOperand(rValue); memoryAccess = sanitizeMemoryAccessForStorageClass(memoryAccess, getStorageClass(lValue)); if (memoryAccess != MemoryAccessMaskNone) { store->addImmediateOperand(memoryAccess); if (memoryAccess & spv::MemoryAccessAlignedMask) { store->addImmediateOperand(alignment); } if (memoryAccess & spv::MemoryAccessMakePointerAvailableKHRMask) { store->addIdOperand(makeUintConstant(scope)); } } buildPoint->addInstruction(std::unique_ptr(store)); } // Comments in header Id Builder::createLoad(Id lValue, spv::MemoryAccessMask memoryAccess, spv::Scope scope, unsigned int alignment) { Instruction* load = new Instruction(getUniqueId(), getDerefTypeId(lValue), OpLoad); load->addIdOperand(lValue); memoryAccess = sanitizeMemoryAccessForStorageClass(memoryAccess, getStorageClass(lValue)); if (memoryAccess != MemoryAccessMaskNone) { load->addImmediateOperand(memoryAccess); if (memoryAccess & spv::MemoryAccessAlignedMask) { load->addImmediateOperand(alignment); } if (memoryAccess & spv::MemoryAccessMakePointerVisibleKHRMask) { load->addIdOperand(makeUintConstant(scope)); } } buildPoint->addInstruction(std::unique_ptr(load)); return load->getResultId(); } // Comments in header Id Builder::createAccessChain(StorageClass storageClass, Id base, const std::vector& offsets) { // Figure out the final resulting type. spv::Id typeId = getTypeId(base); assert(isPointerType(typeId) && offsets.size() > 0); typeId = getContainedTypeId(typeId); for (int i = 0; i < (int)offsets.size(); ++i) { if (isStructType(typeId)) { assert(isConstantScalar(offsets[i])); typeId = getContainedTypeId(typeId, getConstantScalar(offsets[i])); } else typeId = getContainedTypeId(typeId, offsets[i]); } typeId = makePointer(storageClass, typeId); // Make the instruction Instruction* chain = new Instruction(getUniqueId(), typeId, OpAccessChain); chain->addIdOperand(base); for (int i = 0; i < (int)offsets.size(); ++i) chain->addIdOperand(offsets[i]); buildPoint->addInstruction(std::unique_ptr(chain)); return chain->getResultId(); } Id Builder::createArrayLength(Id base, unsigned int member) { spv::Id intType = makeUintType(32); Instruction* length = new Instruction(getUniqueId(), intType, OpArrayLength); length->addIdOperand(base); length->addImmediateOperand(member); buildPoint->addInstruction(std::unique_ptr(length)); return length->getResultId(); } Id Builder::createCooperativeMatrixLength(Id type) { spv::Id intType = makeUintType(32); // Generate code for spec constants if in spec constant operation // generation mode. if (generatingOpCodeForSpecConst) { return createSpecConstantOp(OpCooperativeMatrixLengthNV, intType, std::vector(1, type), std::vector()); } Instruction* length = new Instruction(getUniqueId(), intType, OpCooperativeMatrixLengthNV); length->addIdOperand(type); buildPoint->addInstruction(std::unique_ptr(length)); return length->getResultId(); } Id Builder::createCompositeExtract(Id composite, Id typeId, unsigned index) { // Generate code for spec constants if in spec constant operation // generation mode. if (generatingOpCodeForSpecConst) { return createSpecConstantOp(OpCompositeExtract, typeId, std::vector(1, composite), std::vector(1, index)); } Instruction* extract = new Instruction(getUniqueId(), typeId, OpCompositeExtract); extract->addIdOperand(composite); extract->addImmediateOperand(index); buildPoint->addInstruction(std::unique_ptr(extract)); return extract->getResultId(); } Id Builder::createCompositeExtract(Id composite, Id typeId, const std::vector& indexes) { // Generate code for spec constants if in spec constant operation // generation mode. if (generatingOpCodeForSpecConst) { return createSpecConstantOp(OpCompositeExtract, typeId, std::vector(1, composite), indexes); } Instruction* extract = new Instruction(getUniqueId(), typeId, OpCompositeExtract); extract->addIdOperand(composite); for (int i = 0; i < (int)indexes.size(); ++i) extract->addImmediateOperand(indexes[i]); buildPoint->addInstruction(std::unique_ptr(extract)); return extract->getResultId(); } Id Builder::createCompositeInsert(Id object, Id composite, Id typeId, unsigned index) { Instruction* insert = new Instruction(getUniqueId(), typeId, OpCompositeInsert); insert->addIdOperand(object); insert->addIdOperand(composite); insert->addImmediateOperand(index); buildPoint->addInstruction(std::unique_ptr(insert)); return insert->getResultId(); } Id Builder::createCompositeInsert(Id object, Id composite, Id typeId, const std::vector& indexes) { Instruction* insert = new Instruction(getUniqueId(), typeId, OpCompositeInsert); insert->addIdOperand(object); insert->addIdOperand(composite); for (int i = 0; i < (int)indexes.size(); ++i) insert->addImmediateOperand(indexes[i]); buildPoint->addInstruction(std::unique_ptr(insert)); return insert->getResultId(); } Id Builder::createVectorExtractDynamic(Id vector, Id typeId, Id componentIndex) { Instruction* extract = new Instruction(getUniqueId(), typeId, OpVectorExtractDynamic); extract->addIdOperand(vector); extract->addIdOperand(componentIndex); buildPoint->addInstruction(std::unique_ptr(extract)); return extract->getResultId(); } Id Builder::createVectorInsertDynamic(Id vector, Id typeId, Id component, Id componentIndex) { Instruction* insert = new Instruction(getUniqueId(), typeId, OpVectorInsertDynamic); insert->addIdOperand(vector); insert->addIdOperand(component); insert->addIdOperand(componentIndex); buildPoint->addInstruction(std::unique_ptr(insert)); return insert->getResultId(); } // An opcode that has no operands, no result id, and no type void Builder::createNoResultOp(Op opCode) { Instruction* op = new Instruction(opCode); buildPoint->addInstruction(std::unique_ptr(op)); } // An opcode that has one id operand, no result id, and no type void Builder::createNoResultOp(Op opCode, Id operand) { Instruction* op = new Instruction(opCode); op->addIdOperand(operand); buildPoint->addInstruction(std::unique_ptr(op)); } // An opcode that has one or more operands, no result id, and no type void Builder::createNoResultOp(Op opCode, const std::vector& operands) { Instruction* op = new Instruction(opCode); for (auto it = operands.cbegin(); it != operands.cend(); ++it) { op->addIdOperand(*it); } buildPoint->addInstruction(std::unique_ptr(op)); } // An opcode that has multiple operands, no result id, and no type void Builder::createNoResultOp(Op opCode, const std::vector& operands) { Instruction* op = new Instruction(opCode); for (auto it = operands.cbegin(); it != operands.cend(); ++it) { if (it->isId) op->addIdOperand(it->word); else op->addImmediateOperand(it->word); } buildPoint->addInstruction(std::unique_ptr(op)); } void Builder::createControlBarrier(Scope execution, Scope memory, MemorySemanticsMask semantics) { Instruction* op = new Instruction(OpControlBarrier); op->addIdOperand(makeUintConstant(execution)); op->addIdOperand(makeUintConstant(memory)); op->addIdOperand(makeUintConstant(semantics)); buildPoint->addInstruction(std::unique_ptr(op)); } void Builder::createMemoryBarrier(unsigned executionScope, unsigned memorySemantics) { Instruction* op = new Instruction(OpMemoryBarrier); op->addIdOperand(makeUintConstant(executionScope)); op->addIdOperand(makeUintConstant(memorySemantics)); buildPoint->addInstruction(std::unique_ptr(op)); } // An opcode that has one operands, a result id, and a type Id Builder::createUnaryOp(Op opCode, Id typeId, Id operand) { // Generate code for spec constants if in spec constant operation // generation mode. if (generatingOpCodeForSpecConst) { return createSpecConstantOp(opCode, typeId, std::vector(1, operand), std::vector()); } Instruction* op = new Instruction(getUniqueId(), typeId, opCode); op->addIdOperand(operand); buildPoint->addInstruction(std::unique_ptr(op)); return op->getResultId(); } Id Builder::createBinOp(Op opCode, Id typeId, Id left, Id right) { // Generate code for spec constants if in spec constant operation // generation mode. if (generatingOpCodeForSpecConst) { std::vector operands(2); operands[0] = left; operands[1] = right; return createSpecConstantOp(opCode, typeId, operands, std::vector()); } Instruction* op = new Instruction(getUniqueId(), typeId, opCode); op->addIdOperand(left); op->addIdOperand(right); buildPoint->addInstruction(std::unique_ptr(op)); return op->getResultId(); } Id Builder::createTriOp(Op opCode, Id typeId, Id op1, Id op2, Id op3) { // Generate code for spec constants if in spec constant operation // generation mode. if (generatingOpCodeForSpecConst) { std::vector operands(3); operands[0] = op1; operands[1] = op2; operands[2] = op3; return createSpecConstantOp( opCode, typeId, operands, std::vector()); } Instruction* op = new Instruction(getUniqueId(), typeId, opCode); op->addIdOperand(op1); op->addIdOperand(op2); op->addIdOperand(op3); buildPoint->addInstruction(std::unique_ptr(op)); return op->getResultId(); } Id Builder::createOp(Op opCode, Id typeId, const std::vector& operands) { Instruction* op = new Instruction(getUniqueId(), typeId, opCode); for (auto it = operands.cbegin(); it != operands.cend(); ++it) op->addIdOperand(*it); buildPoint->addInstruction(std::unique_ptr(op)); return op->getResultId(); } Id Builder::createOp(Op opCode, Id typeId, const std::vector& operands) { Instruction* op = new Instruction(getUniqueId(), typeId, opCode); for (auto it = operands.cbegin(); it != operands.cend(); ++it) { if (it->isId) op->addIdOperand(it->word); else op->addImmediateOperand(it->word); } buildPoint->addInstruction(std::unique_ptr(op)); return op->getResultId(); } Id Builder::createSpecConstantOp(Op opCode, Id typeId, const std::vector& operands, const std::vector& literals) { Instruction* op = new Instruction(getUniqueId(), typeId, OpSpecConstantOp); op->addImmediateOperand((unsigned) opCode); for (auto it = operands.cbegin(); it != operands.cend(); ++it) op->addIdOperand(*it); for (auto it = literals.cbegin(); it != literals.cend(); ++it) op->addImmediateOperand(*it); module.mapInstruction(op); constantsTypesGlobals.push_back(std::unique_ptr(op)); return op->getResultId(); } Id Builder::createFunctionCall(spv::Function* function, const std::vector& args) { Instruction* op = new Instruction(getUniqueId(), function->getReturnType(), OpFunctionCall); op->addIdOperand(function->getId()); for (int a = 0; a < (int)args.size(); ++a) op->addIdOperand(args[a]); buildPoint->addInstruction(std::unique_ptr(op)); return op->getResultId(); } // Comments in header Id Builder::createRvalueSwizzle(Decoration precision, Id typeId, Id source, const std::vector& channels) { if (channels.size() == 1) return setPrecision(createCompositeExtract(source, typeId, channels.front()), precision); if (generatingOpCodeForSpecConst) { std::vector operands(2); operands[0] = operands[1] = source; return setPrecision(createSpecConstantOp(OpVectorShuffle, typeId, operands, channels), precision); } Instruction* swizzle = new Instruction(getUniqueId(), typeId, OpVectorShuffle); assert(isVector(source)); swizzle->addIdOperand(source); swizzle->addIdOperand(source); for (int i = 0; i < (int)channels.size(); ++i) swizzle->addImmediateOperand(channels[i]); buildPoint->addInstruction(std::unique_ptr(swizzle)); return setPrecision(swizzle->getResultId(), precision); } // Comments in header Id Builder::createLvalueSwizzle(Id typeId, Id target, Id source, const std::vector& channels) { if (channels.size() == 1 && getNumComponents(source) == 1) return createCompositeInsert(source, target, typeId, channels.front()); Instruction* swizzle = new Instruction(getUniqueId(), typeId, OpVectorShuffle); assert(isVector(target)); swizzle->addIdOperand(target); assert(getNumComponents(source) == (int)channels.size()); assert(isVector(source)); swizzle->addIdOperand(source); // Set up an identity shuffle from the base value to the result value unsigned int components[4]; int numTargetComponents = getNumComponents(target); for (int i = 0; i < numTargetComponents; ++i) components[i] = i; // Punch in the l-value swizzle for (int i = 0; i < (int)channels.size(); ++i) components[channels[i]] = numTargetComponents + i; // finish the instruction with these components selectors for (int i = 0; i < numTargetComponents; ++i) swizzle->addImmediateOperand(components[i]); buildPoint->addInstruction(std::unique_ptr(swizzle)); return swizzle->getResultId(); } // Comments in header void Builder::promoteScalar(Decoration precision, Id& left, Id& right) { int direction = getNumComponents(right) - getNumComponents(left); if (direction > 0) left = smearScalar(precision, left, makeVectorType(getTypeId(left), getNumComponents(right))); else if (direction < 0) right = smearScalar(precision, right, makeVectorType(getTypeId(right), getNumComponents(left))); return; } // Comments in header Id Builder::smearScalar(Decoration precision, Id scalar, Id vectorType) { assert(getNumComponents(scalar) == 1); assert(getTypeId(scalar) == getScalarTypeId(vectorType)); int numComponents = getNumTypeComponents(vectorType); if (numComponents == 1) return scalar; Instruction* smear = nullptr; if (generatingOpCodeForSpecConst) { auto members = std::vector(numComponents, scalar); // Sometime even in spec-constant-op mode, the temporary vector created by // promoting a scalar might not be a spec constant. This should depend on // the scalar. // e.g.: // const vec2 spec_const_result = a_spec_const_vec2 + a_front_end_const_scalar; // In such cases, the temporary vector created from a_front_end_const_scalar // is not a spec constant vector, even though the binary operation node is marked // as 'specConstant' and we are in spec-constant-op mode. auto result_id = makeCompositeConstant(vectorType, members, isSpecConstant(scalar)); smear = module.getInstruction(result_id); } else { smear = new Instruction(getUniqueId(), vectorType, OpCompositeConstruct); for (int c = 0; c < numComponents; ++c) smear->addIdOperand(scalar); buildPoint->addInstruction(std::unique_ptr(smear)); } return setPrecision(smear->getResultId(), precision); } // Comments in header Id Builder::createBuiltinCall(Id resultType, Id builtins, int entryPoint, const std::vector& args) { Instruction* inst = new Instruction(getUniqueId(), resultType, OpExtInst); inst->addIdOperand(builtins); inst->addImmediateOperand(entryPoint); for (int arg = 0; arg < (int)args.size(); ++arg) inst->addIdOperand(args[arg]); buildPoint->addInstruction(std::unique_ptr(inst)); return inst->getResultId(); } // Accept all parameters needed to create a texture instruction. // Create the correct instruction based on the inputs, and make the call. Id Builder::createTextureCall(Decoration precision, Id resultType, bool sparse, bool fetch, bool proj, bool gather, bool noImplicitLod, const TextureParameters& parameters, ImageOperandsMask signExtensionMask) { static const int maxTextureArgs = 10; Id texArgs[maxTextureArgs] = {}; // // Set up the fixed arguments // int numArgs = 0; bool explicitLod = false; texArgs[numArgs++] = parameters.sampler; texArgs[numArgs++] = parameters.coords; if (parameters.Dref != NoResult) texArgs[numArgs++] = parameters.Dref; if (parameters.component != NoResult) texArgs[numArgs++] = parameters.component; #ifdef NV_EXTENSIONS if (parameters.granularity != NoResult) texArgs[numArgs++] = parameters.granularity; if (parameters.coarse != NoResult) texArgs[numArgs++] = parameters.coarse; #endif // // Set up the optional arguments // int optArgNum = numArgs; // track which operand, if it exists, is the mask of optional arguments ++numArgs; // speculatively make room for the mask operand ImageOperandsMask mask = ImageOperandsMaskNone; // the mask operand if (parameters.bias) { mask = (ImageOperandsMask)(mask | ImageOperandsBiasMask); texArgs[numArgs++] = parameters.bias; } if (parameters.lod) { mask = (ImageOperandsMask)(mask | ImageOperandsLodMask); texArgs[numArgs++] = parameters.lod; explicitLod = true; } else if (parameters.gradX) { mask = (ImageOperandsMask)(mask | ImageOperandsGradMask); texArgs[numArgs++] = parameters.gradX; texArgs[numArgs++] = parameters.gradY; explicitLod = true; } else if (noImplicitLod && ! fetch && ! gather) { // have to explicitly use lod of 0 if not allowed to have them be implicit, and // we would otherwise be about to issue an implicit instruction mask = (ImageOperandsMask)(mask | ImageOperandsLodMask); texArgs[numArgs++] = makeFloatConstant(0.0); explicitLod = true; } if (parameters.offset) { if (isConstant(parameters.offset)) mask = (ImageOperandsMask)(mask | ImageOperandsConstOffsetMask); else { addCapability(CapabilityImageGatherExtended); mask = (ImageOperandsMask)(mask | ImageOperandsOffsetMask); } texArgs[numArgs++] = parameters.offset; } if (parameters.offsets) { addCapability(CapabilityImageGatherExtended); mask = (ImageOperandsMask)(mask | ImageOperandsConstOffsetsMask); texArgs[numArgs++] = parameters.offsets; } if (parameters.sample) { mask = (ImageOperandsMask)(mask | ImageOperandsSampleMask); texArgs[numArgs++] = parameters.sample; } if (parameters.lodClamp) { // capability if this bit is used addCapability(CapabilityMinLod); mask = (ImageOperandsMask)(mask | ImageOperandsMinLodMask); texArgs[numArgs++] = parameters.lodClamp; } if (parameters.nonprivate) { mask = mask | ImageOperandsNonPrivateTexelKHRMask; } if (parameters.volatil) { mask = mask | ImageOperandsVolatileTexelKHRMask; } mask = mask | signExtensionMask; if (mask == ImageOperandsMaskNone) --numArgs; // undo speculative reservation for the mask argument else texArgs[optArgNum] = mask; // // Set up the instruction // Op opCode = OpNop; // All paths below need to set this if (fetch) { if (sparse) opCode = OpImageSparseFetch; else opCode = OpImageFetch; #ifdef NV_EXTENSIONS } else if (parameters.granularity && parameters.coarse) { opCode = OpImageSampleFootprintNV; #endif } else if (gather) { if (parameters.Dref) if (sparse) opCode = OpImageSparseDrefGather; else opCode = OpImageDrefGather; else if (sparse) opCode = OpImageSparseGather; else opCode = OpImageGather; } else if (explicitLod) { if (parameters.Dref) { if (proj) if (sparse) opCode = OpImageSparseSampleProjDrefExplicitLod; else opCode = OpImageSampleProjDrefExplicitLod; else if (sparse) opCode = OpImageSparseSampleDrefExplicitLod; else opCode = OpImageSampleDrefExplicitLod; } else { if (proj) if (sparse) opCode = OpImageSparseSampleProjExplicitLod; else opCode = OpImageSampleProjExplicitLod; else if (sparse) opCode = OpImageSparseSampleExplicitLod; else opCode = OpImageSampleExplicitLod; } } else { if (parameters.Dref) { if (proj) if (sparse) opCode = OpImageSparseSampleProjDrefImplicitLod; else opCode = OpImageSampleProjDrefImplicitLod; else if (sparse) opCode = OpImageSparseSampleDrefImplicitLod; else opCode = OpImageSampleDrefImplicitLod; } else { if (proj) if (sparse) opCode = OpImageSparseSampleProjImplicitLod; else opCode = OpImageSampleProjImplicitLod; else if (sparse) opCode = OpImageSparseSampleImplicitLod; else opCode = OpImageSampleImplicitLod; } } // See if the result type is expecting a smeared result. // This happens when a legacy shadow*() call is made, which // gets a vec4 back instead of a float. Id smearedType = resultType; if (! isScalarType(resultType)) { switch (opCode) { case OpImageSampleDrefImplicitLod: case OpImageSampleDrefExplicitLod: case OpImageSampleProjDrefImplicitLod: case OpImageSampleProjDrefExplicitLod: resultType = getScalarTypeId(resultType); break; default: break; } } Id typeId0 = 0; Id typeId1 = 0; if (sparse) { typeId0 = resultType; typeId1 = getDerefTypeId(parameters.texelOut); resultType = makeStructResultType(typeId0, typeId1); } // Build the SPIR-V instruction Instruction* textureInst = new Instruction(getUniqueId(), resultType, opCode); for (int op = 0; op < optArgNum; ++op) textureInst->addIdOperand(texArgs[op]); if (optArgNum < numArgs) textureInst->addImmediateOperand(texArgs[optArgNum]); for (int op = optArgNum + 1; op < numArgs; ++op) textureInst->addIdOperand(texArgs[op]); setPrecision(textureInst->getResultId(), precision); buildPoint->addInstruction(std::unique_ptr(textureInst)); Id resultId = textureInst->getResultId(); if (sparse) { // set capability addCapability(CapabilitySparseResidency); // Decode the return type that was a special structure createStore(createCompositeExtract(resultId, typeId1, 1), parameters.texelOut); resultId = createCompositeExtract(resultId, typeId0, 0); setPrecision(resultId, precision); } else { // When a smear is needed, do it, as per what was computed // above when resultType was changed to a scalar type. if (resultType != smearedType) resultId = smearScalar(precision, resultId, smearedType); } return resultId; } // Comments in header Id Builder::createTextureQueryCall(Op opCode, const TextureParameters& parameters, bool isUnsignedResult) { // Figure out the result type Id resultType = 0; switch (opCode) { case OpImageQuerySize: case OpImageQuerySizeLod: { int numComponents = 0; switch (getTypeDimensionality(getImageType(parameters.sampler))) { case Dim1D: case DimBuffer: numComponents = 1; break; case Dim2D: case DimCube: case DimRect: case DimSubpassData: numComponents = 2; break; case Dim3D: numComponents = 3; break; default: assert(0); break; } if (isArrayedImageType(getImageType(parameters.sampler))) ++numComponents; Id intType = isUnsignedResult ? makeUintType(32) : makeIntType(32); if (numComponents == 1) resultType = intType; else resultType = makeVectorType(intType, numComponents); break; } case OpImageQueryLod: #ifdef AMD_EXTENSIONS resultType = makeVectorType(getScalarTypeId(getTypeId(parameters.coords)), 2); #else resultType = makeVectorType(makeFloatType(32), 2); #endif break; case OpImageQueryLevels: case OpImageQuerySamples: resultType = isUnsignedResult ? makeUintType(32) : makeIntType(32); break; default: assert(0); break; } Instruction* query = new Instruction(getUniqueId(), resultType, opCode); query->addIdOperand(parameters.sampler); if (parameters.coords) query->addIdOperand(parameters.coords); if (parameters.lod) query->addIdOperand(parameters.lod); buildPoint->addInstruction(std::unique_ptr(query)); return query->getResultId(); } // External comments in header. // Operates recursively to visit the composite's hierarchy. Id Builder::createCompositeCompare(Decoration precision, Id value1, Id value2, bool equal) { Id boolType = makeBoolType(); Id valueType = getTypeId(value1); Id resultId = NoResult; int numConstituents = getNumTypeConstituents(valueType); // Scalars and Vectors if (isScalarType(valueType) || isVectorType(valueType)) { assert(valueType == getTypeId(value2)); // These just need a single comparison, just have // to figure out what it is. Op op; switch (getMostBasicTypeClass(valueType)) { case OpTypeFloat: op = equal ? OpFOrdEqual : OpFOrdNotEqual; break; case OpTypeInt: default: op = equal ? OpIEqual : OpINotEqual; break; case OpTypeBool: op = equal ? OpLogicalEqual : OpLogicalNotEqual; precision = NoPrecision; break; } if (isScalarType(valueType)) { // scalar resultId = createBinOp(op, boolType, value1, value2); } else { // vector resultId = createBinOp(op, makeVectorType(boolType, numConstituents), value1, value2); setPrecision(resultId, precision); // reduce vector compares... resultId = createUnaryOp(equal ? OpAll : OpAny, boolType, resultId); } return setPrecision(resultId, precision); } // Only structs, arrays, and matrices should be left. // They share in common the reduction operation across their constituents. assert(isAggregateType(valueType) || isMatrixType(valueType)); // Compare each pair of constituents for (int constituent = 0; constituent < numConstituents; ++constituent) { std::vector indexes(1, constituent); Id constituentType1 = getContainedTypeId(getTypeId(value1), constituent); Id constituentType2 = getContainedTypeId(getTypeId(value2), constituent); Id constituent1 = createCompositeExtract(value1, constituentType1, indexes); Id constituent2 = createCompositeExtract(value2, constituentType2, indexes); Id subResultId = createCompositeCompare(precision, constituent1, constituent2, equal); if (constituent == 0) resultId = subResultId; else resultId = setPrecision(createBinOp(equal ? OpLogicalAnd : OpLogicalOr, boolType, resultId, subResultId), precision); } return resultId; } // OpCompositeConstruct Id Builder::createCompositeConstruct(Id typeId, const std::vector& constituents) { assert(isAggregateType(typeId) || (getNumTypeConstituents(typeId) > 1 && getNumTypeConstituents(typeId) == (int)constituents.size())); if (generatingOpCodeForSpecConst) { // Sometime, even in spec-constant-op mode, the constant composite to be // constructed may not be a specialization constant. // e.g.: // const mat2 m2 = mat2(a_spec_const, a_front_end_const, another_front_end_const, third_front_end_const); // The first column vector should be a spec constant one, as a_spec_const is a spec constant. // The second column vector should NOT be spec constant, as it does not contain any spec constants. // To handle such cases, we check the constituents of the constant vector to determine whether this // vector should be created as a spec constant. return makeCompositeConstant(typeId, constituents, std::any_of(constituents.begin(), constituents.end(), [&](spv::Id id) { return isSpecConstant(id); })); } Instruction* op = new Instruction(getUniqueId(), typeId, OpCompositeConstruct); for (int c = 0; c < (int)constituents.size(); ++c) op->addIdOperand(constituents[c]); buildPoint->addInstruction(std::unique_ptr(op)); return op->getResultId(); } // Vector or scalar constructor Id Builder::createConstructor(Decoration precision, const std::vector& sources, Id resultTypeId) { Id result = NoResult; unsigned int numTargetComponents = getNumTypeComponents(resultTypeId); unsigned int targetComponent = 0; // Special case: when calling a vector constructor with a single scalar // argument, smear the scalar if (sources.size() == 1 && isScalar(sources[0]) && numTargetComponents > 1) return smearScalar(precision, sources[0], resultTypeId); // accumulate the arguments for OpCompositeConstruct std::vector constituents; Id scalarTypeId = getScalarTypeId(resultTypeId); // lambda to store the result of visiting an argument component const auto latchResult = [&](Id comp) { if (numTargetComponents > 1) constituents.push_back(comp); else result = comp; ++targetComponent; }; // lambda to visit a vector argument's components const auto accumulateVectorConstituents = [&](Id sourceArg) { unsigned int sourceSize = getNumComponents(sourceArg); unsigned int sourcesToUse = sourceSize; if (sourcesToUse + targetComponent > numTargetComponents) sourcesToUse = numTargetComponents - targetComponent; for (unsigned int s = 0; s < sourcesToUse; ++s) { std::vector swiz; swiz.push_back(s); latchResult(createRvalueSwizzle(precision, scalarTypeId, sourceArg, swiz)); } }; // lambda to visit a matrix argument's components const auto accumulateMatrixConstituents = [&](Id sourceArg) { unsigned int sourceSize = getNumColumns(sourceArg) * getNumRows(sourceArg); unsigned int sourcesToUse = sourceSize; if (sourcesToUse + targetComponent > numTargetComponents) sourcesToUse = numTargetComponents - targetComponent; int col = 0; int row = 0; for (unsigned int s = 0; s < sourcesToUse; ++s) { if (row >= getNumRows(sourceArg)) { row = 0; col++; } std::vector indexes; indexes.push_back(col); indexes.push_back(row); latchResult(createCompositeExtract(sourceArg, scalarTypeId, indexes)); row++; } }; // Go through the source arguments, each one could have either // a single or multiple components to contribute. for (unsigned int i = 0; i < sources.size(); ++i) { if (isScalar(sources[i]) || isPointer(sources[i])) latchResult(sources[i]); else if (isVector(sources[i])) accumulateVectorConstituents(sources[i]); else if (isMatrix(sources[i])) accumulateMatrixConstituents(sources[i]); else assert(0); if (targetComponent >= numTargetComponents) break; } // If the result is a vector, make it from the gathered constituents. if (constituents.size() > 0) result = createCompositeConstruct(resultTypeId, constituents); return setPrecision(result, precision); } // Comments in header Id Builder::createMatrixConstructor(Decoration precision, const std::vector& sources, Id resultTypeId) { Id componentTypeId = getScalarTypeId(resultTypeId); int numCols = getTypeNumColumns(resultTypeId); int numRows = getTypeNumRows(resultTypeId); Instruction* instr = module.getInstruction(componentTypeId); unsigned bitCount = instr->getImmediateOperand(0); // Optimize matrix constructed from a bigger matrix if (isMatrix(sources[0]) && getNumColumns(sources[0]) >= numCols && getNumRows(sources[0]) >= numRows) { // To truncate the matrix to a smaller number of rows/columns, we need to: // 1. For each column, extract the column and truncate it to the required size using shuffle // 2. Assemble the resulting matrix from all columns Id matrix = sources[0]; Id columnTypeId = getContainedTypeId(resultTypeId); Id sourceColumnTypeId = getContainedTypeId(getTypeId(matrix)); std::vector channels; for (int row = 0; row < numRows; ++row) channels.push_back(row); std::vector matrixColumns; for (int col = 0; col < numCols; ++col) { std::vector indexes; indexes.push_back(col); Id colv = createCompositeExtract(matrix, sourceColumnTypeId, indexes); setPrecision(colv, precision); if (numRows != getNumRows(matrix)) { matrixColumns.push_back(createRvalueSwizzle(precision, columnTypeId, colv, channels)); } else { matrixColumns.push_back(colv); } } return setPrecision(createCompositeConstruct(resultTypeId, matrixColumns), precision); } // Otherwise, will use a two step process // 1. make a compile-time 2D array of values // 2. construct a matrix from that array // Step 1. // initialize the array to the identity matrix Id ids[maxMatrixSize][maxMatrixSize]; Id one = (bitCount == 64 ? makeDoubleConstant(1.0) : makeFloatConstant(1.0)); Id zero = (bitCount == 64 ? makeDoubleConstant(0.0) : makeFloatConstant(0.0)); for (int col = 0; col < 4; ++col) { for (int row = 0; row < 4; ++row) { if (col == row) ids[col][row] = one; else ids[col][row] = zero; } } // modify components as dictated by the arguments if (sources.size() == 1 && isScalar(sources[0])) { // a single scalar; resets the diagonals for (int col = 0; col < 4; ++col) ids[col][col] = sources[0]; } else if (isMatrix(sources[0])) { // constructing from another matrix; copy over the parts that exist in both the argument and constructee Id matrix = sources[0]; int minCols = std::min(numCols, getNumColumns(matrix)); int minRows = std::min(numRows, getNumRows(matrix)); for (int col = 0; col < minCols; ++col) { std::vector indexes; indexes.push_back(col); for (int row = 0; row < minRows; ++row) { indexes.push_back(row); ids[col][row] = createCompositeExtract(matrix, componentTypeId, indexes); indexes.pop_back(); setPrecision(ids[col][row], precision); } } } else { // fill in the matrix in column-major order with whatever argument components are available int row = 0; int col = 0; for (int arg = 0; arg < (int)sources.size(); ++arg) { Id argComp = sources[arg]; for (int comp = 0; comp < getNumComponents(sources[arg]); ++comp) { if (getNumComponents(sources[arg]) > 1) { argComp = createCompositeExtract(sources[arg], componentTypeId, comp); setPrecision(argComp, precision); } ids[col][row++] = argComp; if (row == numRows) { row = 0; col++; } } } } // Step 2: Construct a matrix from that array. // First make the column vectors, then make the matrix. // make the column vectors Id columnTypeId = getContainedTypeId(resultTypeId); std::vector matrixColumns; for (int col = 0; col < numCols; ++col) { std::vector vectorComponents; for (int row = 0; row < numRows; ++row) vectorComponents.push_back(ids[col][row]); Id column = createCompositeConstruct(columnTypeId, vectorComponents); setPrecision(column, precision); matrixColumns.push_back(column); } // make the matrix return setPrecision(createCompositeConstruct(resultTypeId, matrixColumns), precision); } // Comments in header Builder::If::If(Id cond, unsigned int ctrl, Builder& gb) : builder(gb), condition(cond), control(ctrl), elseBlock(0) { function = &builder.getBuildPoint()->getParent(); // make the blocks, but only put the then-block into the function, // the else-block and merge-block will be added later, in order, after // earlier code is emitted thenBlock = new Block(builder.getUniqueId(), *function); mergeBlock = new Block(builder.getUniqueId(), *function); // Save the current block, so that we can add in the flow control split when // makeEndIf is called. headerBlock = builder.getBuildPoint(); function->addBlock(thenBlock); builder.setBuildPoint(thenBlock); } // Comments in header void Builder::If::makeBeginElse() { // Close out the "then" by having it jump to the mergeBlock builder.createBranch(mergeBlock); // Make the first else block and add it to the function elseBlock = new Block(builder.getUniqueId(), *function); function->addBlock(elseBlock); // Start building the else block builder.setBuildPoint(elseBlock); } // Comments in header void Builder::If::makeEndIf() { // jump to the merge block builder.createBranch(mergeBlock); // Go back to the headerBlock and make the flow control split builder.setBuildPoint(headerBlock); builder.createSelectionMerge(mergeBlock, control); if (elseBlock) builder.createConditionalBranch(condition, thenBlock, elseBlock); else builder.createConditionalBranch(condition, thenBlock, mergeBlock); // add the merge block to the function function->addBlock(mergeBlock); builder.setBuildPoint(mergeBlock); } // Comments in header void Builder::makeSwitch(Id selector, unsigned int control, int numSegments, const std::vector& caseValues, const std::vector& valueIndexToSegment, int defaultSegment, std::vector& segmentBlocks) { Function& function = buildPoint->getParent(); // make all the blocks for (int s = 0; s < numSegments; ++s) segmentBlocks.push_back(new Block(getUniqueId(), function)); Block* mergeBlock = new Block(getUniqueId(), function); // make and insert the switch's selection-merge instruction createSelectionMerge(mergeBlock, control); // make the switch instruction Instruction* switchInst = new Instruction(NoResult, NoType, OpSwitch); switchInst->addIdOperand(selector); auto defaultOrMerge = (defaultSegment >= 0) ? segmentBlocks[defaultSegment] : mergeBlock; switchInst->addIdOperand(defaultOrMerge->getId()); defaultOrMerge->addPredecessor(buildPoint); for (int i = 0; i < (int)caseValues.size(); ++i) { switchInst->addImmediateOperand(caseValues[i]); switchInst->addIdOperand(segmentBlocks[valueIndexToSegment[i]]->getId()); segmentBlocks[valueIndexToSegment[i]]->addPredecessor(buildPoint); } buildPoint->addInstruction(std::unique_ptr(switchInst)); // push the merge block switchMerges.push(mergeBlock); } // Comments in header void Builder::addSwitchBreak() { // branch to the top of the merge block stack createBranch(switchMerges.top()); createAndSetNoPredecessorBlock("post-switch-break"); } // Comments in header void Builder::nextSwitchSegment(std::vector& segmentBlock, int nextSegment) { int lastSegment = nextSegment - 1; if (lastSegment >= 0) { // Close out previous segment by jumping, if necessary, to next segment if (! buildPoint->isTerminated()) createBranch(segmentBlock[nextSegment]); } Block* block = segmentBlock[nextSegment]; block->getParent().addBlock(block); setBuildPoint(block); } // Comments in header void Builder::endSwitch(std::vector& /*segmentBlock*/) { // Close out previous segment by jumping, if necessary, to next segment if (! buildPoint->isTerminated()) addSwitchBreak(); switchMerges.top()->getParent().addBlock(switchMerges.top()); setBuildPoint(switchMerges.top()); switchMerges.pop(); } Block& Builder::makeNewBlock() { Function& function = buildPoint->getParent(); auto block = new Block(getUniqueId(), function); function.addBlock(block); return *block; } Builder::LoopBlocks& Builder::makeNewLoop() { // This verbosity is needed to simultaneously get the same behavior // everywhere (id's in the same order), have a syntax that works // across lots of versions of C++, have no warnings from pedantic // compilation modes, and leave the rest of the code alone. Block& head = makeNewBlock(); Block& body = makeNewBlock(); Block& merge = makeNewBlock(); Block& continue_target = makeNewBlock(); LoopBlocks blocks(head, body, merge, continue_target); loops.push(blocks); return loops.top(); } void Builder::createLoopContinue() { createBranch(&loops.top().continue_target); // Set up a block for dead code. createAndSetNoPredecessorBlock("post-loop-continue"); } void Builder::createLoopExit() { createBranch(&loops.top().merge); // Set up a block for dead code. createAndSetNoPredecessorBlock("post-loop-break"); } void Builder::closeLoop() { loops.pop(); } void Builder::clearAccessChain() { accessChain.base = NoResult; accessChain.indexChain.clear(); accessChain.instr = NoResult; accessChain.swizzle.clear(); accessChain.component = NoResult; accessChain.preSwizzleBaseType = NoType; accessChain.isRValue = false; accessChain.coherentFlags.clear(); accessChain.alignment = 0; } // Comments in header void Builder::accessChainPushSwizzle(std::vector& swizzle, Id preSwizzleBaseType, AccessChain::CoherentFlags coherentFlags, unsigned int alignment) { accessChain.coherentFlags |= coherentFlags; accessChain.alignment |= alignment; // swizzles can be stacked in GLSL, but simplified to a single // one here; the base type doesn't change if (accessChain.preSwizzleBaseType == NoType) accessChain.preSwizzleBaseType = preSwizzleBaseType; // if needed, propagate the swizzle for the current access chain if (accessChain.swizzle.size() > 0) { std::vector oldSwizzle = accessChain.swizzle; accessChain.swizzle.resize(0); for (unsigned int i = 0; i < swizzle.size(); ++i) { assert(swizzle[i] < oldSwizzle.size()); accessChain.swizzle.push_back(oldSwizzle[swizzle[i]]); } } else accessChain.swizzle = swizzle; // determine if we need to track this swizzle anymore simplifyAccessChainSwizzle(); } // Comments in header void Builder::accessChainStore(Id rvalue, spv::MemoryAccessMask memoryAccess, spv::Scope scope, unsigned int alignment) { assert(accessChain.isRValue == false); transferAccessChainSwizzle(true); Id base = collapseAccessChain(); Id source = rvalue; // dynamic component should be gone assert(accessChain.component == NoResult); // If swizzle still exists, it is out-of-order or not full, we must load the target vector, // extract and insert elements to perform writeMask and/or swizzle. if (accessChain.swizzle.size() > 0) { Id tempBaseId = createLoad(base); source = createLvalueSwizzle(getTypeId(tempBaseId), tempBaseId, source, accessChain.swizzle); } // take LSB of alignment alignment = alignment & ~(alignment & (alignment-1)); if (getStorageClass(base) == StorageClassPhysicalStorageBufferEXT) { memoryAccess = (spv::MemoryAccessMask)(memoryAccess | spv::MemoryAccessAlignedMask); } createStore(source, base, memoryAccess, scope, alignment); } // Comments in header Id Builder::accessChainLoad(Decoration precision, Decoration nonUniform, Id resultType, spv::MemoryAccessMask memoryAccess, spv::Scope scope, unsigned int alignment) { Id id; if (accessChain.isRValue) { // transfer access chain, but try to stay in registers transferAccessChainSwizzle(false); if (accessChain.indexChain.size() > 0) { Id swizzleBase = accessChain.preSwizzleBaseType != NoType ? accessChain.preSwizzleBaseType : resultType; // if all the accesses are constants, we can use OpCompositeExtract std::vector indexes; bool constant = true; for (int i = 0; i < (int)accessChain.indexChain.size(); ++i) { if (isConstantScalar(accessChain.indexChain[i])) indexes.push_back(getConstantScalar(accessChain.indexChain[i])); else { constant = false; break; } } if (constant) { id = createCompositeExtract(accessChain.base, swizzleBase, indexes); } else { Id lValue = NoResult; if (spvVersion >= Spv_1_4) { // make a new function variable for this r-value, using an initializer, // and mark it as NonWritable so that downstream it can be detected as a lookup // table lValue = createVariable(StorageClassFunction, getTypeId(accessChain.base), "indexable", accessChain.base); addDecoration(lValue, DecorationNonWritable); } else { lValue = createVariable(StorageClassFunction, getTypeId(accessChain.base), "indexable"); // store into it createStore(accessChain.base, lValue); } // move base to the new variable accessChain.base = lValue; accessChain.isRValue = false; // load through the access chain id = createLoad(collapseAccessChain()); } setPrecision(id, precision); } else id = accessChain.base; // no precision, it was set when this was defined } else { transferAccessChainSwizzle(true); // take LSB of alignment alignment = alignment & ~(alignment & (alignment-1)); if (getStorageClass(accessChain.base) == StorageClassPhysicalStorageBufferEXT) { memoryAccess = (spv::MemoryAccessMask)(memoryAccess | spv::MemoryAccessAlignedMask); } // load through the access chain id = createLoad(collapseAccessChain(), memoryAccess, scope, alignment); setPrecision(id, precision); addDecoration(id, nonUniform); } // Done, unless there are swizzles to do if (accessChain.swizzle.size() == 0 && accessChain.component == NoResult) return id; // Do remaining swizzling // Do the basic swizzle if (accessChain.swizzle.size() > 0) { Id swizzledType = getScalarTypeId(getTypeId(id)); if (accessChain.swizzle.size() > 1) swizzledType = makeVectorType(swizzledType, (int)accessChain.swizzle.size()); id = createRvalueSwizzle(precision, swizzledType, id, accessChain.swizzle); } // Do the dynamic component if (accessChain.component != NoResult) id = setPrecision(createVectorExtractDynamic(id, resultType, accessChain.component), precision); addDecoration(id, nonUniform); return id; } Id Builder::accessChainGetLValue() { assert(accessChain.isRValue == false); transferAccessChainSwizzle(true); Id lvalue = collapseAccessChain(); // If swizzle exists, it is out-of-order or not full, we must load the target vector, // extract and insert elements to perform writeMask and/or swizzle. This does not // go with getting a direct l-value pointer. assert(accessChain.swizzle.size() == 0); assert(accessChain.component == NoResult); return lvalue; } // comment in header Id Builder::accessChainGetInferredType() { // anything to operate on? if (accessChain.base == NoResult) return NoType; Id type = getTypeId(accessChain.base); // do initial dereference if (! accessChain.isRValue) type = getContainedTypeId(type); // dereference each index for (auto it = accessChain.indexChain.cbegin(); it != accessChain.indexChain.cend(); ++it) { if (isStructType(type)) type = getContainedTypeId(type, getConstantScalar(*it)); else type = getContainedTypeId(type); } // dereference swizzle if (accessChain.swizzle.size() == 1) type = getContainedTypeId(type); else if (accessChain.swizzle.size() > 1) type = makeVectorType(getContainedTypeId(type), (int)accessChain.swizzle.size()); // dereference component selection if (accessChain.component) type = getContainedTypeId(type); return type; } void Builder::dump(std::vector& out) const { // Header, before first instructions: out.push_back(MagicNumber); out.push_back(spvVersion); out.push_back(builderNumber); out.push_back(uniqueId + 1); out.push_back(0); // Capabilities for (auto it = capabilities.cbegin(); it != capabilities.cend(); ++it) { Instruction capInst(0, 0, OpCapability); capInst.addImmediateOperand(*it); capInst.dump(out); } for (auto it = extensions.cbegin(); it != extensions.cend(); ++it) { Instruction extInst(0, 0, OpExtension); extInst.addStringOperand(it->c_str()); extInst.dump(out); } dumpInstructions(out, imports); Instruction memInst(0, 0, OpMemoryModel); memInst.addImmediateOperand(addressModel); memInst.addImmediateOperand(memoryModel); memInst.dump(out); // Instructions saved up while building: dumpInstructions(out, entryPoints); dumpInstructions(out, executionModes); // Debug instructions dumpInstructions(out, strings); dumpSourceInstructions(out); for (int e = 0; e < (int)sourceExtensions.size(); ++e) { Instruction sourceExtInst(0, 0, OpSourceExtension); sourceExtInst.addStringOperand(sourceExtensions[e]); sourceExtInst.dump(out); } dumpInstructions(out, names); dumpModuleProcesses(out); // Annotation instructions dumpInstructions(out, decorations); dumpInstructions(out, constantsTypesGlobals); dumpInstructions(out, externals); // The functions module.dump(out); } // // Protected methods. // // Turn the described access chain in 'accessChain' into an instruction(s) // computing its address. This *cannot* include complex swizzles, which must // be handled after this is called. // // Can generate code. Id Builder::collapseAccessChain() { assert(accessChain.isRValue == false); // did we already emit an access chain for this? if (accessChain.instr != NoResult) return accessChain.instr; // If we have a dynamic component, we can still transfer // that into a final operand to the access chain. We need to remap the // dynamic component through the swizzle to get a new dynamic component to // update. // // This was not done in transferAccessChainSwizzle() because it might // generate code. remapDynamicSwizzle(); if (accessChain.component != NoResult) { // transfer the dynamic component to the access chain accessChain.indexChain.push_back(accessChain.component); accessChain.component = NoResult; } // note that non-trivial swizzling is left pending // do we have an access chain? if (accessChain.indexChain.size() == 0) return accessChain.base; // emit the access chain StorageClass storageClass = (StorageClass)module.getStorageClass(getTypeId(accessChain.base)); accessChain.instr = createAccessChain(storageClass, accessChain.base, accessChain.indexChain); return accessChain.instr; } // For a dynamic component selection of a swizzle. // // Turn the swizzle and dynamic component into just a dynamic component. // // Generates code. void Builder::remapDynamicSwizzle() { // do we have a swizzle to remap a dynamic component through? if (accessChain.component != NoResult && accessChain.swizzle.size() > 1) { // build a vector of the swizzle for the component to map into std::vector components; for (int c = 0; c < (int)accessChain.swizzle.size(); ++c) components.push_back(makeUintConstant(accessChain.swizzle[c])); Id mapType = makeVectorType(makeUintType(32), (int)accessChain.swizzle.size()); Id map = makeCompositeConstant(mapType, components); // use it accessChain.component = createVectorExtractDynamic(map, makeUintType(32), accessChain.component); accessChain.swizzle.clear(); } } // clear out swizzle if it is redundant, that is reselecting the same components // that would be present without the swizzle. void Builder::simplifyAccessChainSwizzle() { // If the swizzle has fewer components than the vector, it is subsetting, and must stay // to preserve that fact. if (getNumTypeComponents(accessChain.preSwizzleBaseType) > (int)accessChain.swizzle.size()) return; // if components are out of order, it is a swizzle for (unsigned int i = 0; i < accessChain.swizzle.size(); ++i) { if (i != accessChain.swizzle[i]) return; } // otherwise, there is no need to track this swizzle accessChain.swizzle.clear(); if (accessChain.component == NoResult) accessChain.preSwizzleBaseType = NoType; } // To the extent any swizzling can become part of the chain // of accesses instead of a post operation, make it so. // If 'dynamic' is true, include transferring the dynamic component, // otherwise, leave it pending. // // Does not generate code. just updates the access chain. void Builder::transferAccessChainSwizzle(bool dynamic) { // non existent? if (accessChain.swizzle.size() == 0 && accessChain.component == NoResult) return; // too complex? // (this requires either a swizzle, or generating code for a dynamic component) if (accessChain.swizzle.size() > 1) return; // single component, either in the swizzle and/or dynamic component if (accessChain.swizzle.size() == 1) { assert(accessChain.component == NoResult); // handle static component selection accessChain.indexChain.push_back(makeUintConstant(accessChain.swizzle.front())); accessChain.swizzle.clear(); accessChain.preSwizzleBaseType = NoType; } else if (dynamic && accessChain.component != NoResult) { assert(accessChain.swizzle.size() == 0); // handle dynamic component accessChain.indexChain.push_back(accessChain.component); accessChain.preSwizzleBaseType = NoType; accessChain.component = NoResult; } } // Utility method for creating a new block and setting the insert point to // be in it. This is useful for flow-control operations that need a "dummy" // block proceeding them (e.g. instructions after a discard, etc). void Builder::createAndSetNoPredecessorBlock(const char* /*name*/) { Block* block = new Block(getUniqueId(), buildPoint->getParent()); block->setUnreachable(); buildPoint->getParent().addBlock(block); setBuildPoint(block); // if (name) // addName(block->getId(), name); } // Comments in header void Builder::createBranch(Block* block) { Instruction* branch = new Instruction(OpBranch); branch->addIdOperand(block->getId()); buildPoint->addInstruction(std::unique_ptr(branch)); block->addPredecessor(buildPoint); } void Builder::createSelectionMerge(Block* mergeBlock, unsigned int control) { Instruction* merge = new Instruction(OpSelectionMerge); merge->addIdOperand(mergeBlock->getId()); merge->addImmediateOperand(control); buildPoint->addInstruction(std::unique_ptr(merge)); } void Builder::createLoopMerge(Block* mergeBlock, Block* continueBlock, unsigned int control, const std::vector& operands) { Instruction* merge = new Instruction(OpLoopMerge); merge->addIdOperand(mergeBlock->getId()); merge->addIdOperand(continueBlock->getId()); merge->addImmediateOperand(control); for (int op = 0; op < (int)operands.size(); ++op) merge->addImmediateOperand(operands[op]); buildPoint->addInstruction(std::unique_ptr(merge)); } void Builder::createConditionalBranch(Id condition, Block* thenBlock, Block* elseBlock) { Instruction* branch = new Instruction(OpBranchConditional); branch->addIdOperand(condition); branch->addIdOperand(thenBlock->getId()); branch->addIdOperand(elseBlock->getId()); buildPoint->addInstruction(std::unique_ptr(branch)); thenBlock->addPredecessor(buildPoint); elseBlock->addPredecessor(buildPoint); } // OpSource // [OpSourceContinued] // ... void Builder::dumpSourceInstructions(const spv::Id fileId, const std::string& text, std::vector& out) const { const int maxWordCount = 0xFFFF; const int opSourceWordCount = 4; const int nonNullBytesPerInstruction = 4 * (maxWordCount - opSourceWordCount) - 1; if (source != SourceLanguageUnknown) { // OpSource Language Version File Source Instruction sourceInst(NoResult, NoType, OpSource); sourceInst.addImmediateOperand(source); sourceInst.addImmediateOperand(sourceVersion); // File operand if (fileId != NoResult) { sourceInst.addIdOperand(fileId); // Source operand if (text.size() > 0) { int nextByte = 0; std::string subString; while ((int)text.size() - nextByte > 0) { subString = text.substr(nextByte, nonNullBytesPerInstruction); if (nextByte == 0) { // OpSource sourceInst.addStringOperand(subString.c_str()); sourceInst.dump(out); } else { // OpSourcContinued Instruction sourceContinuedInst(OpSourceContinued); sourceContinuedInst.addStringOperand(subString.c_str()); sourceContinuedInst.dump(out); } nextByte += nonNullBytesPerInstruction; } } else sourceInst.dump(out); } else sourceInst.dump(out); } } // Dump an OpSource[Continued] sequence for the source and every include file void Builder::dumpSourceInstructions(std::vector& out) const { dumpSourceInstructions(sourceFileStringId, sourceText, out); for (auto iItr = includeFiles.begin(); iItr != includeFiles.end(); ++iItr) dumpSourceInstructions(iItr->first, *iItr->second, out); } void Builder::dumpInstructions(std::vector& out, const std::vector >& instructions) const { for (int i = 0; i < (int)instructions.size(); ++i) { instructions[i]->dump(out); } } void Builder::dumpModuleProcesses(std::vector& out) const { for (int i = 0; i < (int)moduleProcesses.size(); ++i) { Instruction moduleProcessed(OpModuleProcessed); moduleProcessed.addStringOperand(moduleProcesses[i]); moduleProcessed.dump(out); } } }; // end spv namespace