/* * Copyright 2015-2016 ARM Limited * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. * You may obtain a copy of the License at * * http://www.apache.org/licenses/LICENSE-2.0 * * Unless required by applicable law or agreed to in writing, software * distributed under the License is distributed on an "AS IS" BASIS, * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. * See the License for the specific language governing permissions and * limitations under the License. */ #include "spirv_cross.hpp" #include "GLSL.std.450.h" #include #include #include using namespace std; using namespace spv; using namespace spirv_cross; #define log(...) fprintf(stderr, __VA_ARGS__) Instruction::Instruction(const vector &spirv, uint32_t &index) { op = spirv[index] & 0xffff; count = (spirv[index] >> 16) & 0xffff; offset = index + 1; length = count - 1; index += count; if (index > spirv.size()) throw CompilerError("SPIR-V instruction goes out of bounds."); } Compiler::Compiler(vector ir) : spirv(move(ir)) { parse(); } string Compiler::compile() { return ""; } bool Compiler::variable_storage_is_aliased(const SPIRVariable &v) { auto &type = get(v.basetype); bool ssbo = (meta[type.self].decoration.decoration_flags & (1ull << DecorationBufferBlock)) != 0; bool image = type.basetype == SPIRType::Image; bool counter = type.basetype == SPIRType::AtomicCounter; return ssbo || image || counter; } bool Compiler::block_is_pure(const SPIRBlock &block) { for (auto &i : block.ops) { auto ops = stream(i); auto op = static_cast(i.op); switch (op) { case OpFunctionCall: { uint32_t func = ops[2]; if (!function_is_pure(get(func))) return false; break; } case OpStore: { auto &type = expression_type(ops[0]); if (type.storage != StorageClassFunction) return false; break; } case OpImageWrite: return false; // Atomics are impure. case OpAtomicLoad: case OpAtomicStore: case OpAtomicExchange: case OpAtomicCompareExchange: case OpAtomicIIncrement: case OpAtomicIDecrement: case OpAtomicIAdd: case OpAtomicISub: case OpAtomicSMin: case OpAtomicUMin: case OpAtomicSMax: case OpAtomicUMax: case OpAtomicAnd: case OpAtomicOr: case OpAtomicXor: return false; // Geometry shader builtins modify global state. case OpEndPrimitive: case OpEmitStreamVertex: case OpEndStreamPrimitive: case OpEmitVertex: return false; // Barriers disallow any reordering, so we should treat blocks with barrier as writing. case OpControlBarrier: case OpMemoryBarrier: return false; // OpExtInst is potentially impure depending on extension, but GLSL builtins are at least pure. default: break; } } return true; } string Compiler::to_name(uint32_t id, bool allow_alias) { if (allow_alias && ids.at(id).get_type() == TypeType) { // If this type is a simple alias, emit the // name of the original type instead. // We don't want to override the meta alias // as that can be overridden by the reflection APIs after parse. auto &type = get(id); if (type.type_alias) return to_name(type.type_alias); } if (meta[id].decoration.alias.empty()) return join("_", id); else return meta.at(id).decoration.alias; } bool Compiler::function_is_pure(const SPIRFunction &func) { for (auto block : func.blocks) { if (!block_is_pure(get(block))) { //fprintf(stderr, "Function %s is impure!\n", to_name(func.self).c_str()); return false; } } //fprintf(stderr, "Function %s is pure!\n", to_name(func.self).c_str()); return true; } void Compiler::register_global_read_dependencies(const SPIRBlock &block, uint32_t id) { for (auto &i : block.ops) { auto ops = stream(i); auto op = static_cast(i.op); switch (op) { case OpFunctionCall: { uint32_t func = ops[2]; register_global_read_dependencies(get(func), id); break; } case OpLoad: case OpImageRead: { // If we're in a storage class which does not get invalidated, adding dependencies here is no big deal. auto *var = maybe_get_backing_variable(ops[2]); if (var && var->storage != StorageClassFunction) { auto &type = get(var->basetype); // InputTargets are immutable. if (type.basetype != SPIRType::Image && type.image.dim != DimSubpassData) var->dependees.push_back(id); } break; } default: break; } } } void Compiler::register_global_read_dependencies(const SPIRFunction &func, uint32_t id) { for (auto block : func.blocks) register_global_read_dependencies(get(block), id); } SPIRVariable *Compiler::maybe_get_backing_variable(uint32_t chain) { auto *var = maybe_get(chain); if (!var) { auto *cexpr = maybe_get(chain); if (cexpr) var = maybe_get(cexpr->loaded_from); } return var; } void Compiler::register_read(uint32_t expr, uint32_t chain, bool forwarded) { auto &e = get(expr); auto *var = maybe_get_backing_variable(chain); if (var) { e.loaded_from = var->self; // If the backing variable is immutable, we do not need to depend on the variable. if (forwarded && !is_immutable(var->self)) var->dependees.push_back(e.self); // If we load from a parameter, make sure we create "inout" if we also write to the parameter. // The default is "in" however, so we never invalidate our compilation by reading. if (var && var->parameter) var->parameter->read_count++; } } void Compiler::register_write(uint32_t chain) { auto *var = maybe_get(chain); if (!var) { // If we're storing through an access chain, invalidate the backing variable instead. auto *expr = maybe_get(chain); if (expr && expr->loaded_from) var = maybe_get(expr->loaded_from); } if (var) { // If our variable is in a storage class which can alias with other buffers, // invalidate all variables which depend on aliased variables. if (variable_storage_is_aliased(*var)) flush_all_aliased_variables(); else if (var) flush_dependees(*var); // We tried to write to a parameter which is not marked with out qualifier, force a recompile. if (var->parameter && var->parameter->write_count == 0) { var->parameter->write_count++; force_recompile = true; } } } void Compiler::flush_dependees(SPIRVariable &var) { for (auto expr : var.dependees) { invalid_expressions.insert(expr); get(expr).invalidated_by.push_back(var.self); } var.dependees.clear(); } void Compiler::flush_all_aliased_variables() { for (auto aliased : aliased_variables) flush_dependees(get(aliased)); } void Compiler::flush_all_atomic_capable_variables() { for (auto global : global_variables) flush_dependees(get(global)); flush_all_aliased_variables(); } void Compiler::flush_all_active_variables() { // Invalidate all temporaries we read from variables in this block since they were forwarded. // Invalidate all temporaries we read from globals. for (auto &v : current_function->local_variables) flush_dependees(get(v)); for (auto &arg : current_function->arguments) flush_dependees(get(arg.id)); for (auto global : global_variables) flush_dependees(get(global)); flush_all_aliased_variables(); } const SPIRType &Compiler::expression_type(uint32_t id) const { switch (ids[id].get_type()) { case TypeVariable: return get(get(id).basetype); case TypeExpression: return get(get(id).expression_type); case TypeConstant: return get(get(id).constant_type); case TypeUndef: return get(get(id).basetype); default: throw CompilerError("Cannot resolve expression type."); } } bool Compiler::expression_is_lvalue(uint32_t id) const { auto &type = expression_type(id); switch (type.basetype) { case SPIRType::SampledImage: case SPIRType::Image: case SPIRType::Sampler: return false; default: return true; } } bool Compiler::is_immutable(uint32_t id) const { if (ids[id].get_type() == TypeVariable) { auto &var = get(id); // Anything we load from the UniformConstant address space is guaranteed to be immutable. bool pointer_to_const = var.storage == StorageClassUniformConstant; return pointer_to_const || var.phi_variable || var.forwardable || !expression_is_lvalue(id); } else if (ids[id].get_type() == TypeExpression) return get(id).immutable; else if (ids[id].get_type() == TypeConstant || ids[id].get_type() == TypeUndef) return true; else return false; } bool Compiler::is_builtin_variable(const SPIRVariable &var) const { if (var.compat_builtin || meta[var.self].decoration.builtin) return true; // We can have builtin structs as well. If one member of a struct is builtin, the struct must also be builtin. for (auto &m : meta[get(var.basetype).self].members) if (m.builtin) return true; return false; } bool Compiler::is_member_builtin(const SPIRType &type, uint32_t index, BuiltIn *builtin) const { auto &memb = meta[type.self].members; if (index < memb.size() && memb[index].builtin) { if (builtin) *builtin = memb[index].builtin_type; return true; } return false; } bool Compiler::is_scalar(const SPIRType &type) const { return type.vecsize == 1 && type.columns == 1; } bool Compiler::is_vector(const SPIRType &type) const { return type.vecsize > 1 && type.columns == 1; } bool Compiler::is_matrix(const SPIRType &type) const { return type.vecsize > 1 && type.columns > 1; } ShaderResources Compiler::get_shader_resources() const { ShaderResources res; for (auto &id : ids) { if (id.get_type() != TypeVariable) continue; auto &var = id.get(); auto &type = get(var.basetype); if (!type.pointer || is_builtin_variable(var)) continue; // Input if (var.storage == StorageClassInput) { if (meta[type.self].decoration.decoration_flags & (1ull << DecorationBlock)) res.stage_inputs.push_back({ var.self, type.self, meta[type.self].decoration.alias }); else res.stage_inputs.push_back({ var.self, type.self, meta[var.self].decoration.alias }); } // Subpass inputs else if (var.storage == StorageClassUniformConstant && type.image.dim == DimSubpassData) { res.subpass_inputs.push_back({ var.self, type.self, meta[var.self].decoration.alias }); } // Outputs else if (var.storage == StorageClassOutput) { if (meta[type.self].decoration.decoration_flags & (1ull << DecorationBlock)) res.stage_outputs.push_back({ var.self, type.self, meta[type.self].decoration.alias }); else res.stage_outputs.push_back({ var.self, type.self, meta[var.self].decoration.alias }); } // UBOs else if (type.storage == StorageClassUniform && (meta[type.self].decoration.decoration_flags & (1ull << DecorationBlock))) { res.uniform_buffers.push_back({ var.self, type.self, meta[type.self].decoration.alias }); } // SSBOs else if (type.storage == StorageClassUniform && (meta[type.self].decoration.decoration_flags & (1ull << DecorationBufferBlock))) { res.storage_buffers.push_back({ var.self, type.self, meta[type.self].decoration.alias }); } // Push constant blocks else if (type.storage == StorageClassPushConstant) { // There can only be one push constant block, but keep the vector in case this restriction is lifted // in the future. res.push_constant_buffers.push_back({ var.self, type.self, meta[var.self].decoration.alias }); } // Images else if (type.storage == StorageClassUniformConstant && type.basetype == SPIRType::Image) { res.storage_images.push_back({ var.self, type.self, meta[var.self].decoration.alias }); } // Textures else if (type.storage == StorageClassUniformConstant && type.basetype == SPIRType::SampledImage) { res.sampled_images.push_back({ var.self, type.self, meta[var.self].decoration.alias }); } // Atomic counters else if (type.storage == StorageClassAtomicCounter) { res.atomic_counters.push_back({ var.self, type.self, meta[var.self].decoration.alias }); } } return res; } static inline uint32_t swap_endian(uint32_t v) { return ((v >> 24) & 0x000000ffu) | ((v >> 8) & 0x0000ff00u) | ((v << 8) & 0x00ff0000u) | ((v << 24) & 0xff000000u); } static string extract_string(const vector &spirv, uint32_t offset) { string ret; for (uint32_t i = offset; i < spirv.size(); i++) { uint32_t w = spirv[i]; for (uint32_t j = 0; j < 4; j++, w >>= 8) { char c = w & 0xff; if (c == '\0') return ret; ret += c; } } throw CompilerError("String was not terminated before EOF"); } static bool is_valid_spirv_version(uint32_t version) { switch (version) { // Allow v99 since it tends to just work. case 99: case 0x10000: // SPIR-V 1.0 case 0x10100: // SPIR-V 1.1 return true; default: return false; } } void Compiler::parse() { auto len = spirv.size(); if (len < 5) throw CompilerError("SPIRV file too small."); auto s = spirv.data(); // Endian-swap if we need to. if (s[0] == swap_endian(MagicNumber)) transform(begin(spirv), end(spirv), begin(spirv), [](uint32_t c) { return swap_endian(c); }); if (s[0] != MagicNumber || !is_valid_spirv_version(s[1])) throw CompilerError("Invalid SPIRV format."); uint32_t bound = s[3]; ids.resize(bound); meta.resize(bound); uint32_t offset = 5; while (offset < len) inst.emplace_back(spirv, offset); for (auto &i : inst) parse(i); if (current_function) throw CompilerError("Function was not terminated."); if (current_block) throw CompilerError("Block was not terminated."); } void Compiler::flatten_interface_block(uint32_t id) { auto &var = get(id); auto &type = get(var.basetype); auto flags = meta.at(type.self).decoration.decoration_flags; if (!type.array.empty()) throw CompilerError("Type is array of UBOs."); if (type.basetype != SPIRType::Struct) throw CompilerError("Type is not a struct."); if ((flags & (1ull << DecorationBlock)) == 0) throw CompilerError("Type is not a block."); if (type.member_types.empty()) throw CompilerError("Member list of struct is empty."); uint32_t t = type.member_types[0]; for (auto &m : type.member_types) if (t != m) throw CompilerError("Types in block differ."); auto &mtype = get(t); if (!mtype.array.empty()) throw CompilerError("Member type cannot be arrays."); if (mtype.basetype == SPIRType::Struct) throw CompilerError("Member type cannot be struct."); // Inherit variable name from interface block name. meta.at(var.self).decoration.alias = meta.at(type.self).decoration.alias; auto storage = var.storage; if (storage == StorageClassUniform) storage = StorageClassUniformConstant; // Change type definition in-place into an array instead. // Access chains will still work as-is. uint32_t array_size = uint32_t(type.member_types.size()); type = mtype; type.array.push_back(array_size); type.pointer = true; type.storage = storage; var.storage = storage; } void Compiler::update_name_cache(unordered_set &cache, string &name) { if (name.empty()) return; if (cache.find(name) == end(cache)) { cache.insert(name); return; } uint32_t counter = 0; auto tmpname = name; // If there is a collision (very rare), // keep tacking on extra identifier until it's unique. do { counter++; name = tmpname + "_" + convert_to_string(counter); } while (cache.find(name) != end(cache)); cache.insert(name); } void Compiler::set_name(uint32_t id, const std::string &name) { auto &str = meta.at(id).decoration.alias; str.clear(); if (name.empty()) return; // Reserved for temporaries. if (name[0] == '_') return; // Functions in glslangValidator are mangled with name( stuff. // Normally, we would never see '(' in any legal indentifiers, so just strip them out. str = name.substr(0, name.find('(')); for (uint32_t i = 0; i < str.size(); i++) { auto &c = str[i]; // _ variables are reserved by the internal implementation, // otherwise, make sure the name is a valid identifier. if (i == 0 || (str[0] == '_' && i == 1)) c = isalpha(c) ? c : '_'; else c = isalnum(c) ? c : '_'; } } const SPIRType &Compiler::get_type(uint32_t id) const { return get(id); } void Compiler::set_member_decoration(uint32_t id, uint32_t index, Decoration decoration, uint32_t argument) { meta.at(id).members.resize(max(meta[id].members.size(), size_t(index) + 1)); auto &dec = meta.at(id).members[index]; dec.decoration_flags |= 1ull << decoration; switch (decoration) { case DecorationBuiltIn: dec.builtin = true; dec.builtin_type = static_cast(argument); break; case DecorationLocation: dec.location = argument; break; case DecorationOffset: dec.offset = argument; break; default: break; } } void Compiler::set_member_name(uint32_t id, uint32_t index, const std::string &name) { meta.at(id).members.resize(max(meta[id].members.size(), size_t(index) + 1)); meta.at(id).members[index].alias = name; } const std::string &Compiler::get_member_name(uint32_t id, uint32_t index) const { auto &m = meta.at(id); if (index >= m.members.size()) { static string empty; return empty; } return m.members[index].alias; } uint32_t Compiler::get_member_decoration(uint32_t id, uint32_t index, Decoration decoration) const { auto &dec = meta.at(id).members.at(index); if (!(dec.decoration_flags & (1ull << decoration))) return 0; switch (decoration) { case DecorationBuiltIn: return dec.builtin_type; case DecorationLocation: return dec.location; case DecorationOffset: return dec.offset; default: return 0; } } uint64_t Compiler::get_member_decoration_mask(uint32_t id, uint32_t index) const { auto &m = meta.at(id); if (index >= m.members.size()) return 0; return m.members[index].decoration_flags; } void Compiler::unset_member_decoration(uint32_t id, uint32_t index, Decoration decoration) { auto &m = meta.at(id); if (index >= m.members.size()) return; auto &dec = m.members[index]; dec.decoration_flags &= ~(1ull << decoration); switch (decoration) { case DecorationBuiltIn: dec.builtin = false; break; case DecorationLocation: dec.location = 0; break; case DecorationOffset: dec.offset = 0; break; default: break; } } void Compiler::set_decoration(uint32_t id, Decoration decoration, uint32_t argument) { auto &dec = meta.at(id).decoration; dec.decoration_flags |= 1ull << decoration; switch (decoration) { case DecorationBuiltIn: dec.builtin = true; dec.builtin_type = static_cast(argument); break; case DecorationLocation: dec.location = argument; break; case DecorationOffset: dec.offset = argument; break; case DecorationArrayStride: dec.array_stride = argument; break; case DecorationBinding: dec.binding = argument; break; case DecorationDescriptorSet: dec.set = argument; break; case DecorationInputAttachmentIndex: dec.input_attachment = argument; break; default: break; } } StorageClass Compiler::get_storage_class(uint32_t id) const { return get(id).storage; } const std::string &Compiler::get_name(uint32_t id) const { return meta.at(id).decoration.alias; } uint64_t Compiler::get_decoration_mask(uint32_t id) const { auto &dec = meta.at(id).decoration; return dec.decoration_flags; } uint32_t Compiler::get_decoration(uint32_t id, Decoration decoration) const { auto &dec = meta.at(id).decoration; if (!(dec.decoration_flags & (1ull << decoration))) return 0; switch (decoration) { case DecorationBuiltIn: return dec.builtin_type; case DecorationLocation: return dec.location; case DecorationOffset: return dec.offset; case DecorationBinding: return dec.binding; case DecorationDescriptorSet: return dec.set; case DecorationInputAttachmentIndex: return dec.input_attachment; default: return 0; } } void Compiler::unset_decoration(uint32_t id, Decoration decoration) { auto &dec = meta.at(id).decoration; dec.decoration_flags &= ~(1ull << decoration); switch (decoration) { case DecorationBuiltIn: dec.builtin = false; break; case DecorationLocation: dec.location = 0; break; case DecorationOffset: dec.offset = 0; break; case DecorationBinding: dec.binding = 0; break; case DecorationDescriptorSet: dec.set = 0; break; default: break; } } void Compiler::parse(const Instruction &instruction) { auto ops = stream(instruction); auto op = static_cast(instruction.op); uint32_t length = instruction.length; switch (op) { case OpMemoryModel: case OpSourceExtension: case OpNop: break; case OpSource: { auto lang = static_cast(ops[0]); switch (lang) { case SourceLanguageESSL: source.es = true; source.version = ops[1]; source.known = true; break; case SourceLanguageGLSL: source.es = false; source.version = ops[1]; source.known = true; break; default: source.known = false; break; } break; } case OpUndef: { uint32_t result_type = ops[0]; uint32_t id = ops[1]; set(id, result_type); break; } case OpCapability: { uint32_t cap = ops[0]; if (cap == CapabilityKernel) throw CompilerError("Kernel capability not supported."); break; } case OpExtInstImport: { uint32_t id = ops[0]; auto ext = extract_string(spirv, instruction.offset + 1); if (ext == "GLSL.std.450") set(id, SPIRExtension::GLSL); else throw CompilerError("Only GLSL.std.450 extension interface supported."); break; } case OpEntryPoint: { if (execution.entry_point) throw CompilerError("More than one entry point not supported."); execution.model = static_cast(ops[0]); execution.entry_point = ops[1]; break; } case OpExecutionMode: { uint32_t entry = ops[0]; if (entry != execution.entry_point) throw CompilerError("Cannot set execution mode to non-existing entry point."); auto mode = static_cast(ops[1]); execution.flags |= 1ull << mode; switch (mode) { case ExecutionModeInvocations: execution.invocations = ops[2]; break; case ExecutionModeLocalSize: execution.workgroup_size.x = ops[2]; execution.workgroup_size.y = ops[3]; execution.workgroup_size.z = ops[4]; break; case ExecutionModeOutputVertices: execution.output_vertices = ops[2]; break; default: break; } break; } case OpName: { uint32_t id = ops[0]; set_name(id, extract_string(spirv, instruction.offset + 1)); break; } case OpMemberName: { uint32_t id = ops[0]; uint32_t member = ops[1]; set_member_name(id, member, extract_string(spirv, instruction.offset + 2)); break; } case OpDecorate: { uint32_t id = ops[0]; auto decoration = static_cast(ops[1]); if (length >= 3) set_decoration(id, decoration, ops[2]); else set_decoration(id, decoration); break; } case OpMemberDecorate: { uint32_t id = ops[0]; uint32_t member = ops[1]; auto decoration = static_cast(ops[2]); if (length >= 4) set_member_decoration(id, member, decoration, ops[3]); else set_member_decoration(id, member, decoration); break; } // Build up basic types. case OpTypeVoid: { uint32_t id = ops[0]; auto &type = set(id); type.basetype = SPIRType::Void; break; } case OpTypeBool: { uint32_t id = ops[0]; auto &type = set(id); type.basetype = SPIRType::Bool; type.width = 1; break; } case OpTypeFloat: { uint32_t id = ops[0]; uint32_t width = ops[1]; auto &type = set(id); type.basetype = SPIRType::Float; type.width = width; break; } case OpTypeInt: { uint32_t id = ops[0]; uint32_t width = ops[1]; auto &type = set(id); type.basetype = ops[2] ? SPIRType::Int : SPIRType::UInt; type.width = width; break; } // Build composite types by "inheriting". // NOTE: The self member is also copied! For pointers and array modifiers this is a good thing // since we can refer to decorations on pointee classes which is needed for UBO/SSBO, I/O blocks in geometry/tess etc. case OpTypeVector: { uint32_t id = ops[0]; uint32_t vecsize = ops[2]; auto &base = get(ops[1]); auto &vecbase = set(id); vecbase = base; vecbase.vecsize = vecsize; vecbase.self = id; break; } case OpTypeMatrix: { uint32_t id = ops[0]; uint32_t colcount = ops[2]; auto &base = get(ops[1]); auto &matrixbase = set(id); matrixbase = base; matrixbase.columns = colcount; matrixbase.self = id; break; } case OpTypeArray: { uint32_t id = ops[0]; auto &base = get(ops[1]); auto &arraybase = set(id); arraybase = base; arraybase.array.push_back(get(ops[2]).scalar()); // Do NOT set arraybase.self! break; } case OpTypeRuntimeArray: { uint32_t id = ops[0]; auto &base = get(ops[1]); auto &arraybase = set(id); arraybase = base; arraybase.array.push_back(0); // Do NOT set arraybase.self! break; } case OpTypeImage: { uint32_t id = ops[0]; auto &type = set(id); type.basetype = SPIRType::Image; type.image.type = ops[1]; type.image.dim = static_cast(ops[2]); type.image.depth = ops[3] != 0; type.image.arrayed = ops[4] != 0; type.image.ms = ops[5] != 0; type.image.sampled = ops[6]; type.image.format = static_cast(ops[7]); break; } case OpTypeSampledImage: { uint32_t id = ops[0]; uint32_t imagetype = ops[1]; auto &type = set(id); type = get(imagetype); type.basetype = SPIRType::SampledImage; type.self = id; break; } // Not really used. case OpTypeSampler: { uint32_t id = ops[0]; auto &type = set(id); type.basetype = SPIRType::Sampler; break; } case OpTypePointer: { uint32_t id = ops[0]; auto &base = get(ops[2]); auto &ptrbase = set(id); ptrbase = base; if (ptrbase.pointer) throw CompilerError("Cannot make pointer-to-pointer type."); ptrbase.pointer = true; ptrbase.storage = static_cast(ops[1]); if (ptrbase.storage == StorageClassAtomicCounter) ptrbase.basetype = SPIRType::AtomicCounter; // Do NOT set ptrbase.self! break; } case OpTypeStruct: { uint32_t id = ops[0]; auto &type = set(id); type.basetype = SPIRType::Struct; for (uint32_t i = 1; i < length; i++) type.member_types.push_back(ops[i]); // Check if we have seen this struct type before, with just different // decorations. for (auto &other : global_struct_cache) { if (types_are_logically_equivalent(type, get(other))) { type.type_alias = other; break; } } if (type.type_alias == 0) global_struct_cache.push_back(id); break; } case OpTypeFunction: { uint32_t id = ops[0]; uint32_t ret = ops[1]; auto &func = set(id, ret); for (uint32_t i = 2; i < length; i++) func.parameter_types.push_back(ops[i]); break; } // Variable declaration // All variables are essentially pointers with a storage qualifier. case OpVariable: { uint32_t type = ops[0]; uint32_t id = ops[1]; auto storage = static_cast(ops[2]); uint32_t initializer = length == 4 ? ops[3] : 0; if (storage == StorageClassFunction) { if (!current_function) throw CompilerError("No function currently in scope"); current_function->add_local_variable(id); } else if (storage == StorageClassPrivate || storage == StorageClassWorkgroup || storage == StorageClassOutput) { global_variables.push_back(id); } auto &var = set(id, type, storage, initializer); if (variable_storage_is_aliased(var)) aliased_variables.push_back(var.self); // glslangValidator does not emit required qualifiers here. // Solve this by making the image access as restricted as possible // and loosen up if we need to. auto &vartype = expression_type(id); if (vartype.basetype == SPIRType::Image) { auto &flags = meta.at(id).decoration.decoration_flags; flags |= 1ull << DecorationNonWritable; flags |= 1ull << DecorationNonReadable; } break; } // OpPhi // OpPhi is a fairly magical opcode. // It selects temporary variables based on which parent block we *came from*. // In high-level languages we can "de-SSA" by creating a function local, and flush out temporaries to this function-local // variable to emulate SSA Phi. case OpPhi: { if (!current_function) throw CompilerError("No function currently in scope"); if (!current_block) throw CompilerError("No block currently in scope"); uint32_t result_type = ops[0]; uint32_t id = ops[1]; // Instead of a temporary, create a new function-wide temporary with this ID instead. auto &var = set(id, result_type, spv::StorageClassFunction); var.phi_variable = true; current_function->add_local_variable(id); for (uint32_t i = 2; i + 2 <= length; i += 2) current_block->phi_variables.push_back({ ops[i], ops[i + 1], id }); break; } // Constants case OpSpecConstant: case OpConstant: { uint32_t id = ops[1]; set(id, ops[0], ops[2]).specialization = op == OpSpecConstant; break; } case OpSpecConstantFalse: case OpConstantFalse: { uint32_t id = ops[1]; set(id, ops[0], 0).specialization = op == OpSpecConstantFalse; break; } case OpSpecConstantTrue: case OpConstantTrue: { uint32_t id = ops[1]; set(id, ops[0], 1).specialization = op == OpSpecConstantTrue; break; } case OpSpecConstantComposite: case OpConstantComposite: { uint32_t id = ops[1]; uint32_t type = ops[0]; auto &ctype = get(type); SPIRConstant *constant = nullptr; // We can have constants which are structs and arrays. // In this case, our SPIRConstant will be a list of other SPIRConstant ids which we // can refer to. if (ctype.basetype == SPIRType::Struct || !ctype.array.empty()) { constant = &set(id, type, ops + 2, length - 2); constant->specialization = op == OpSpecConstantComposite; break; } bool matrix = ctype.columns > 1; if (matrix) { switch (length - 2) { case 1: constant = &set(id, type, get(ops[2]).vector()); break; case 2: constant = &set(id, type, get(ops[2]).vector(), get(ops[3]).vector()); break; case 3: constant = &set(id, type, get(ops[2]).vector(), get(ops[3]).vector(), get(ops[4]).vector()); break; case 4: constant = &set(id, type, get(ops[2]).vector(), get(ops[3]).vector(), get(ops[4]).vector(), get(ops[5]).vector()); break; default: throw CompilerError("OpConstantComposite only supports 1, 2, 3 and 4 columns."); } } else { switch (length - 2) { case 1: constant = &set(id, type, get(ops[2]).scalar()); break; case 2: constant = &set(id, type, get(ops[2]).scalar(), get(ops[3]).scalar()); break; case 3: constant = &set(id, type, get(ops[2]).scalar(), get(ops[3]).scalar(), get(ops[4]).scalar()); break; case 4: constant = &set(id, type, get(ops[2]).scalar(), get(ops[3]).scalar(), get(ops[4]).scalar(), get(ops[5]).scalar()); break; default: throw CompilerError("OpConstantComposite only supports 1, 2, 3 and 4 components."); } } constant->specialization = op == OpSpecConstantComposite; break; } // Functions case OpFunction: { uint32_t res = ops[0]; uint32_t id = ops[1]; // Control uint32_t type = ops[3]; if (current_function) throw CompilerError("Must end a function before starting a new one!"); current_function = &set(id, res, type); break; } case OpFunctionParameter: { uint32_t type = ops[0]; uint32_t id = ops[1]; if (!current_function) throw CompilerError("Must be in a function!"); current_function->add_parameter(type, id); set(id, type, StorageClassFunction); break; } case OpFunctionEnd: { current_function = nullptr; break; } // Blocks case OpLabel: { // OpLabel always starts a block. if (!current_function) throw CompilerError("Blocks cannot exist outside functions!"); uint32_t id = ops[0]; current_function->blocks.push_back(id); if (!current_function->entry_block) current_function->entry_block = id; if (current_block) throw CompilerError("Cannot start a block before ending the current block."); current_block = &set(id); break; } // Branch instructions end blocks. case OpBranch: { if (!current_block) throw CompilerError("Trying to end a non-existing block."); uint32_t target = ops[0]; current_block->terminator = SPIRBlock::Direct; current_block->next_block = target; current_block = nullptr; break; } case OpBranchConditional: { if (!current_block) throw CompilerError("Trying to end a non-existing block."); current_block->condition = ops[0]; current_block->true_block = ops[1]; current_block->false_block = ops[2]; current_block->terminator = SPIRBlock::Select; current_block = nullptr; break; } case OpSwitch: { if (!current_block) throw CompilerError("Trying to end a non-existing block."); if (current_block->merge == SPIRBlock::MergeNone) throw CompilerError("Switch statement is not structured"); current_block->terminator = SPIRBlock::MultiSelect; current_block->condition = ops[0]; current_block->default_block = ops[1]; for (uint32_t i = 2; i + 2 <= length; i += 2) current_block->cases.push_back({ ops[i], ops[i + 1] }); // If we jump to next block, make it break instead since we're inside a switch case block at that point. multiselect_merge_targets.insert(current_block->next_block); current_block = nullptr; break; } case OpKill: { if (!current_block) throw CompilerError("Trying to end a non-existing block."); current_block->terminator = SPIRBlock::Kill; current_block = nullptr; break; } case OpReturn: { if (!current_block) throw CompilerError("Trying to end a non-existing block."); current_block->terminator = SPIRBlock::Return; current_block = nullptr; break; } case OpReturnValue: { if (!current_block) throw CompilerError("Trying to end a non-existing block."); current_block->terminator = SPIRBlock::Return; current_block->return_value = ops[0]; current_block = nullptr; break; } case OpUnreachable: { if (!current_block) throw CompilerError("Trying to end a non-existing block."); current_block->terminator = SPIRBlock::Unreachable; current_block = nullptr; break; } case OpSelectionMerge: { if (!current_block) throw CompilerError("Trying to modify a non-existing block."); current_block->next_block = ops[0]; current_block->merge = SPIRBlock::MergeSelection; selection_merge_targets.insert(current_block->next_block); break; } case OpLoopMerge: { if (!current_block) throw CompilerError("Trying to modify a non-existing block."); current_block->merge_block = ops[0]; current_block->continue_block = ops[1]; current_block->merge = SPIRBlock::MergeLoop; loop_blocks.insert(current_block->self); loop_merge_targets.insert(current_block->merge_block); // Don't add loop headers to continue blocks, // which would make it impossible branch into the loop header since // they are treated as continues. if (current_block->continue_block != current_block->self) continue_blocks.insert(current_block->continue_block); break; } // Actual opcodes. default: { if (!current_block) throw CompilerError("Currently no block to insert opcode."); current_block->ops.push_back(instruction); break; } } } bool Compiler::block_is_loop_candidate(const SPIRBlock &block, SPIRBlock::Method method) const { // Tried and failed. if (block.disable_block_optimization || block.complex_continue) return false; if (method == SPIRBlock::MergeToSelectForLoop) { // Try to detect common for loop pattern // which the code backend can use to create cleaner code. // for(;;) { if (cond) { some_body; } else { break; } } // is the pattern we're looking for. bool ret = block.terminator == SPIRBlock::Select && block.merge == SPIRBlock::MergeLoop && block.true_block != block.merge_block && block.true_block != block.self && block.false_block == block.merge_block; // If we have OpPhi which depends on branches which came from our own block, // we need to flush phi variables in else block instead of a trivial break, // so we cannot assume this is a for loop candidate. if (ret) { for (auto &phi : block.phi_variables) if (phi.parent == block.self) return false; auto *merge = maybe_get(block.merge_block); if (merge) for (auto &phi : merge->phi_variables) if (phi.parent == block.self) return false; } return ret; } else if (method == SPIRBlock::MergeToDirectForLoop) { // Empty loop header that just sets up merge target // and branches to loop body. bool ret = block.terminator == SPIRBlock::Direct && block.merge == SPIRBlock::MergeLoop && block.ops.empty(); if (!ret) return false; auto &child = get(block.next_block); ret = child.terminator == SPIRBlock::Select && child.merge == SPIRBlock::MergeNone && child.false_block == block.merge_block && child.true_block != block.merge_block && child.true_block != block.self; // If we have OpPhi which depends on branches which came from our own block, // we need to flush phi variables in else block instead of a trivial break, // so we cannot assume this is a for loop candidate. if (ret) { for (auto &phi : block.phi_variables) if (phi.parent == block.self || phi.parent == child.self) return false; for (auto &phi : child.phi_variables) if (phi.parent == block.self) return false; auto *merge = maybe_get(block.merge_block); if (merge) for (auto &phi : merge->phi_variables) if (phi.parent == block.self || phi.parent == child.false_block) return false; } return ret; } else return false; } bool Compiler::block_is_outside_flow_control_from_block(const SPIRBlock &from, const SPIRBlock &to) { auto *start = &from; if (start->self == to.self) return true; // Break cycles. if (is_continue(start->self)) return false; // If our select block doesn't merge, we must break or continue in these blocks, // so if continues occur branchless within these blocks, consider them branchless as well. // This is typically used for loop control. if (start->terminator == SPIRBlock::Select && start->merge == SPIRBlock::MergeNone && (block_is_outside_flow_control_from_block(get(start->true_block), to) || block_is_outside_flow_control_from_block(get(start->false_block), to))) { return true; } else if (start->merge_block && block_is_outside_flow_control_from_block(get(start->merge_block), to)) { return true; } else if (start->next_block && block_is_outside_flow_control_from_block(get(start->next_block), to)) { return true; } else return false; } bool Compiler::execution_is_noop(const SPIRBlock &from, const SPIRBlock &to) const { if (!execution_is_branchless(from, to)) return false; auto *start = &from; for (;;) { if (start->self == to.self) return true; if (!start->ops.empty()) return false; start = &get(start->next_block); } } bool Compiler::execution_is_branchless(const SPIRBlock &from, const SPIRBlock &to) const { auto *start = &from; for (;;) { if (start->self == to.self) return true; if (start->terminator == SPIRBlock::Direct && start->merge == SPIRBlock::MergeNone) start = &get(start->next_block); else return false; } } SPIRBlock::ContinueBlockType Compiler::continue_block_type(const SPIRBlock &block) const { // The block was deemed too complex during code emit, pick conservative fallback paths. if (block.complex_continue) return SPIRBlock::ComplexLoop; // In older glslang output continue block can be equal to the loop header. // In this case, execution is clearly branchless, so just assume a while loop header here. if (block.merge == SPIRBlock::MergeLoop) return SPIRBlock::WhileLoop; auto &dominator = get(block.loop_dominator); if (execution_is_noop(block, dominator)) return SPIRBlock::WhileLoop; else if (execution_is_branchless(block, dominator)) return SPIRBlock::ForLoop; else { if (block.merge == SPIRBlock::MergeNone && block.terminator == SPIRBlock::Select && block.true_block == dominator.self && block.false_block == dominator.merge_block) { return SPIRBlock::DoWhileLoop; } else return SPIRBlock::ComplexLoop; } } bool Compiler::traverse_all_reachable_opcodes(const SPIRBlock &block, OpcodeHandler &handler) const { // Ideally, perhaps traverse the CFG instead of all blocks in order to eliminate dead blocks, // but this shouldn't be a problem in practice unless the SPIR-V is doing insane things like recursing // inside dead blocks ... for (auto &i : block.ops) { auto ops = stream(i); auto op = static_cast(i.op); if (!handler.handle(op, ops, i.length)) return false; uint32_t func = ops[2]; if (op == OpFunctionCall && !traverse_all_reachable_opcodes(get(func), handler)) return false; } return true; } bool Compiler::traverse_all_reachable_opcodes(const SPIRFunction &func, OpcodeHandler &handler) const { for (auto block : func.blocks) if (!traverse_all_reachable_opcodes(get(block), handler)) return false; return true; } uint32_t Compiler::type_struct_member_offset(const SPIRType &type, uint32_t index) const { // Decoration must be set in valid SPIR-V, otherwise throw. auto &dec = meta[type.self].members.at(index); if (dec.decoration_flags & (1ull << DecorationOffset)) return dec.offset; else throw CompilerError("Struct member does not have Offset set."); } uint32_t Compiler::type_struct_member_array_stride(const SPIRType &type, uint32_t index) const { // Decoration must be set in valid SPIR-V, otherwise throw. // ArrayStride is part of the array type not OpMemberDecorate. auto &dec = meta[type.member_types[index]].decoration; if (dec.decoration_flags & (1ull << DecorationArrayStride)) return dec.array_stride; else throw CompilerError("Struct member does not have ArrayStride set."); } size_t Compiler::get_declared_struct_size(const SPIRType &type) const { uint32_t last = uint32_t(type.member_types.size() - 1); size_t offset = type_struct_member_offset(type, last); size_t size = get_declared_struct_member_size(type, last); return offset + size; } size_t Compiler::get_declared_struct_member_size(const SPIRType &struct_type, uint32_t index) const { auto flags = get_member_decoration_mask(struct_type.self, index); auto &type = get(struct_type.member_types[index]); if (type.basetype != SPIRType::Struct) { switch (type.basetype) { case SPIRType::Unknown: case SPIRType::Void: case SPIRType::Bool: // Bools are purely logical, and cannot be used for externally visible types. case SPIRType::AtomicCounter: case SPIRType::Image: case SPIRType::SampledImage: case SPIRType::Sampler: throw CompilerError("Querying size for object with opaque size.\n"); default: break; } size_t component_size = type.width / 8; unsigned vecsize = type.vecsize; unsigned columns = type.columns; if (type.array.empty()) { // Vectors. if (columns == 1) return vecsize * component_size; else { // Per SPIR-V spec, matrices must be tightly packed and aligned up for vec3 accesses. if ((flags & (1ull << DecorationRowMajor)) && columns == 3) columns = 4; else if ((flags & (1ull << DecorationColMajor)) && vecsize == 3) vecsize = 4; return vecsize * columns * component_size; } } else { // For arrays, we can use ArrayStride to get an easy check. return type_struct_member_array_stride(struct_type, index) * type.array.back(); } } else { // Recurse. uint32_t last = uint32_t(struct_type.member_types.size() - 1); uint32_t offset = type_struct_member_offset(struct_type, last); size_t size = get_declared_struct_size(get(struct_type.member_types.back())); return offset + size; } } bool Compiler::BufferAccessHandler::handle(Op opcode, const uint32_t *args, uint32_t length) { if (opcode != OpAccessChain && opcode != OpInBoundsAccessChain) return true; // Invalid SPIR-V. if (length < 4) return false; if (args[2] != id) return true; // Don't bother traversing the entire access chain tree yet. // If we access a struct member, assume we access the entire member. uint32_t index = compiler.get(args[3]).scalar(); // Seen this index already. if (seen.find(index) != end(seen)) return true; seen.insert(index); auto &type = compiler.expression_type(id); uint32_t offset = compiler.type_struct_member_offset(type, index); size_t range; // If we have another member in the struct, deduce the range by looking at the next member. // This is okay since structs in SPIR-V can have padding, but Offset decoration must be // monotonically increasing. // Of course, this doesn't take into account if the SPIR-V for some reason decided to add // very large amounts of padding, but that's not really a big deal. if (index + 1 < type.member_types.size()) { range = compiler.type_struct_member_offset(type, index + 1) - offset; } else { // No padding, so just deduce it from the size of the member directly. range = compiler.get_declared_struct_member_size(type, index); } ranges.push_back({ index, offset, range }); return true; } std::vector Compiler::get_active_buffer_ranges(uint32_t id) const { std::vector ranges; BufferAccessHandler handler(*this, ranges, id); traverse_all_reachable_opcodes(get(execution.entry_point), handler); return ranges; } // Increase the number of IDs by the specified incremental amount. // Returns the value of the first ID available for use in the expanded bound. uint32_t Compiler::increase_bound_by(uint32_t incr_amount) { uint32_t curr_bound = (uint32_t)ids.size(); uint32_t new_bound = curr_bound + incr_amount; ids.resize(new_bound); meta.resize(new_bound); return curr_bound; } bool Compiler::types_are_logically_equivalent(const SPIRType &a, const SPIRType &b) const { if (a.basetype != b.basetype) return false; if (a.width != b.width) return false; if (a.vecsize != b.vecsize) return false; if (a.columns != b.columns) return false; if (a.array.size() != b.array.size()) return false; unsigned array_count = a.array.size(); if (array_count && memcmp(a.array.data(), b.array.data(), array_count * sizeof(uint32_t)) != 0) return false; if (a.basetype == SPIRType::Image || a.basetype == SPIRType::SampledImage) { if (memcmp(&a.image, &b.image, sizeof(SPIRType::Image)) != 0) return false; } if (a.member_types.size() != b.member_types.size()) return false; unsigned member_types = a.member_types.size(); for (unsigned i = 0; i < member_types; i++) { if (!types_are_logically_equivalent(get(a.member_types[i]), get(b.member_types[i]))) return false; } return true; }