diff options
Diffstat (limited to 'modules/skjson/src/SkJSON.cpp')
-rw-r--r-- | modules/skjson/src/SkJSON.cpp | 481 |
1 files changed, 131 insertions, 350 deletions
diff --git a/modules/skjson/src/SkJSON.cpp b/modules/skjson/src/SkJSON.cpp index 8af85ba695..b513c78536 100644 --- a/modules/skjson/src/SkJSON.cpp +++ b/modules/skjson/src/SkJSON.cpp @@ -9,7 +9,6 @@ #include "SkStream.h" #include "SkString.h" -#include "SkTo.h" #include <cmath> #include <vector> @@ -18,357 +17,131 @@ namespace skjson { //#define SK_JSON_REPORT_ERRORS -namespace { - -/* - Value's impl side: - - -- fixed 64-bit size - - -- 8-byte aligned - - -- union of: - - bool - int32 - float - char[8] (short string storage) - external payload pointer - -- highest 3 bits reserved for type storage - - */ static_assert( sizeof(Value) == 8, ""); static_assert(alignof(Value) == 8, ""); static constexpr size_t kRecAlign = alignof(Value); -// The current record layout assumes LE and will take some tweaking for BE. -#if defined(SK_CPU_BENDIAN) -static_assert(false, "Big-endian builds are not supported."); -#endif - -class ValueRec : public Value { -public: - static constexpr uint64_t kTypeBits = 3, - kTypeShift = 64 - kTypeBits, - kTypeMask = ((1ULL << kTypeBits) - 1) << kTypeShift; - - enum RecType : uint64_t { - // We picked kShortString == 0 so that tag 0b000 and stored max_size-size (7-7=0) - // conveniently overlap the '\0' terminator, allowing us to store a 7 character - // C string inline. - kShortString = 0b000ULL << kTypeShift, // inline payload - kNull = 0b001ULL << kTypeShift, // no payload - kBool = 0b010ULL << kTypeShift, // inline payload - kInt = 0b011ULL << kTypeShift, // inline payload - kFloat = 0b100ULL << kTypeShift, // inline payload - kString = 0b101ULL << kTypeShift, // ptr to external storage - kArray = 0b110ULL << kTypeShift, // ptr to external storage - kObject = 0b111ULL << kTypeShift, // ptr to external storage - }; - - RecType getRecType() const { - return static_cast<RecType>(*this->cast<uint64_t>() & kTypeMask); - } - - // Access the record data as T. - // - // This is also used to access the payload for inline records. Since the record type lives in - // the high bits, sizeof(T) must be less than sizeof(Value) when accessing inline payloads. - // - // E.g. - // - // uint8_t - // ----------------------------------------------------------------------- - // | val8 | val8 | val8 | val8 | val8 | val8 | val8 | TYPE| - // ----------------------------------------------------------------------- - // - // uint32_t - // ----------------------------------------------------------------------- - // | val32 | unused | TYPE| - // ----------------------------------------------------------------------- - // - // T* (64b) - // ----------------------------------------------------------------------- - // | T* (kTypeShift bits) |TYPE| - // ----------------------------------------------------------------------- - // - template <typename T> - const T* cast() const { - static_assert(sizeof (T) <= sizeof(ValueRec), ""); - static_assert(alignof(T) <= alignof(ValueRec), ""); - return reinterpret_cast<const T*>(this); - } - - template <typename T> - T* cast() { return const_cast<T*>(const_cast<const ValueRec*>(this)->cast<T>()); } - - // Access the pointer payload. - template <typename T> - const T* ptr() const { - static_assert(sizeof(uintptr_t) == sizeof(Value) || - sizeof(uintptr_t) * 2 == sizeof(Value), ""); - - return (sizeof(uintptr_t) < sizeof(Value)) - // For 32-bit, pointers are stored unmodified. - ? *this->cast<const T*>() - // For 64-bit, we use the high bits of the pointer as type storage. - : reinterpret_cast<T*>(*this->cast<uintptr_t>() & ~kTypeMask); - } - - // Type-bound recs only store their type. - static ValueRec MakeTypeBound(RecType t) { - ValueRec v; - *v.cast<uint64_t>() = t; - SkASSERT(v.getRecType() == t); - return v; - } - - // Primitive recs store a type and inline primitive payload. - template <typename T> - static ValueRec MakePrimitive(RecType t, T src) { - ValueRec v = MakeTypeBound(t); - *v.cast<T>() = src; - SkASSERT(v.getRecType() == t); - return v; - } - - // Pointer recs store a type (in the upper kTypeBits bits) and a pointer. - template <typename T> - static ValueRec MakePtr(RecType t, const T* p) { - SkASSERT((t & kTypeMask) == t); - if (sizeof(uintptr_t) == sizeof(Value)) { - // For 64-bit, we rely on the pointer hi bits being unused. - SkASSERT(!(reinterpret_cast<uintptr_t>(p) & kTypeMask)); - } - - ValueRec v = MakeTypeBound(t); - *v.cast<uintptr_t>() |= reinterpret_cast<uintptr_t>(p); - - SkASSERT(v.getRecType() == t); - SkASSERT(v.ptr<T>() == p); - - return v; - } - - // Vector recs point to externally allocated slabs with the following layout: - // - // [size_t n] [REC_0] ... [REC_n-1] [optional extra trailing storage] - // - // Long strings use extra_alloc_size == 1 to store the \0 terminator. - template <typename T, size_t extra_alloc_size = 0> - static ValueRec MakeVector(RecType t, const T* src, size_t size, SkArenaAlloc& alloc) { - // For zero-size arrays, we just store a nullptr. - size_t* size_ptr = nullptr; - - if (size) { - // The Ts are already in memory, so their size should be safeish. - const auto total_size = sizeof(size_t) + sizeof(T) * size + extra_alloc_size; - size_ptr = reinterpret_cast<size_t*>(alloc.makeBytesAlignedTo(total_size, kRecAlign)); - auto* data_ptr = reinterpret_cast<T*>(size_ptr + 1); - *size_ptr = size; - memcpy(data_ptr, src, sizeof(T) * size); - } - - return MakePtr(t, size_ptr); - } - - size_t vectorSize(RecType t) const { - if (this->is<NullValue>()) return 0; - SkASSERT(this->getRecType() == t); - - const auto* size_ptr = this->ptr<const size_t>(); - return size_ptr ? *size_ptr : 0; - } - - template <typename T> - const T* vectorBegin(RecType t) const { - if (this->is<NullValue>()) return nullptr; - SkASSERT(this->getRecType() == t); - - const auto* size_ptr = this->ptr<const size_t>(); - return size_ptr ? reinterpret_cast<const T*>(size_ptr + 1) : nullptr; - } - - template <typename T> - const T* vectorEnd(RecType t) const { - if (this->is<NullValue>()) return nullptr; - SkASSERT(this->getRecType() == t); - - const auto* size_ptr = this->ptr<const size_t>(); - return size_ptr ? reinterpret_cast<const T*>(size_ptr + 1) + *size_ptr : nullptr; - } - - // Strings have two flavors: - // - // -- short strings (len <= 7) -> these are stored inline, in the record - // (one byte reserved for null terminator/type): - // - // [str] [\0]|[max_len - actual_len] - // - // Storing [max_len - actual_len] allows the 'len' field to double-up as a - // null terminator when size == max_len (this works 'cause kShortString == 0). - // - // -- long strings (len > 7) -> these are externally allocated vectors (VectorRec<char>). - // - // The string data plus a null-char terminator are copied over. - static constexpr size_t kMaxInlineStringSize = sizeof(Value) - 1; - - static ValueRec MakeString(const char* src, size_t size, SkArenaAlloc& alloc) { - ValueRec v; - - if (size > kMaxInlineStringSize) { - v = MakeVector<char, 1>(kString, src, size, alloc); - const_cast<char *>(v.vectorBegin<char>(ValueRec::kString))[size] = '\0'; - } else { - v = MakeTypeBound(kShortString); - auto* payload = v.cast<char>(); - memcpy(payload, src, size); - payload[size] = '\0'; - - const auto len_tag = SkTo<char>(kMaxInlineStringSize - size); - // This technically overwrites the type hi bits, but is safe because - // 1) kShortString == 0 - // 2) 0 <= len_tag <= 7 - static_assert(kShortString == 0, "please don't break this"); - payload[kMaxInlineStringSize] = len_tag; - SkASSERT(v.getRecType() == kShortString); - } - return v; - } - - size_t stringSize() const { - if (this->getRecType() == ValueRec::kShortString) { - const auto* payload = this->cast<char>(); - return kMaxInlineStringSize - SkToSizeT(payload[kMaxInlineStringSize]); - } - - return this->vectorSize(ValueRec::kString); - } - - const char* stringBegin() const { - if (this->getRecType() == ValueRec::kShortString) { - return this->cast<char>(); - } - - return this->vectorBegin<char>(ValueRec::kString); - } +void Value::init_tagged(Tag t) { + memset(fData8, 0, sizeof(fData8)); + fData8[Value::kTagOffset] = SkTo<uint8_t>(t); + SkASSERT(this->getTag() == t); +} - const char* stringEnd() const { - if (this->getRecType() == ValueRec::kShortString) { - const auto* payload = this->cast<char>(); - return payload + kMaxInlineStringSize - SkToSizeT(payload[kMaxInlineStringSize]); - } +// Pointer values store a type (in the upper kTagBits bits) and a pointer. +void Value::init_tagged_pointer(Tag t, void* p) { + *this->cast<uintptr_t>() = reinterpret_cast<uintptr_t>(p); - return this->vectorEnd<char>(ValueRec::kString); + if (sizeof(Value) == sizeof(uintptr_t)) { + // For 64-bit, we rely on the pointer upper bits being unused/zero. + SkASSERT(!(fData8[kTagOffset] & kTagMask)); + fData8[kTagOffset] |= SkTo<uint8_t>(t); + } else { + // For 32-bit, we need to zero-initialize the upper 32 bits + SkASSERT(sizeof(Value) == sizeof(uintptr_t) * 2); + this->cast<uintptr_t>()[kTagOffset >> 2] = 0; + fData8[kTagOffset] = SkTo<uint8_t>(t); } -}; -} // namespace - - -// Boring public Value glue. - -const Value& Value::Null() { - static const Value g_null = ValueRec::MakeTypeBound(ValueRec::kNull); - return g_null; + SkASSERT(this->getTag() == t); + SkASSERT(this->ptr<void>() == p); } -const Member& Member::Null() { - static const Member g_null = { Value::Null().as<StringValue>(), Value::Null() }; - return g_null; +NullValue::NullValue() { + this->init_tagged(Tag::kNull); + SkASSERT(this->getTag() == Tag::kNull); } -Value::Type Value::getType() const { - static constexpr Value::Type kTypeMap[] = { - Value::Type::kString, // kShortString - Value::Type::kNull, // kNull - Value::Type::kBool, // kBool - Value::Type::kNumber, // kInt - Value::Type::kNumber, // kFloat - Value::Type::kString, // kString - Value::Type::kArray, // kArray - Value::Type::kObject, // kObject - }; - - const auto& rec = *reinterpret_cast<const ValueRec*>(this); - const auto type_index = static_cast<size_t>(rec.getRecType() >> ValueRec::kTypeShift); - SkASSERT(type_index < SK_ARRAY_COUNT(kTypeMap)); - - return kTypeMap[type_index]; +BoolValue::BoolValue(bool b) { + this->init_tagged(Tag::kBool); + *this->cast<bool>() = b; + SkASSERT(this->getTag() == Tag::kBool); } -template <> -bool PrimitiveValue<bool, Value::Type::kBool>::operator*() const { - const auto& rec = *reinterpret_cast<const ValueRec*>(this); - - if (rec.is<NullValue>()) return false; - - SkASSERT(rec.getRecType() == ValueRec::kBool); - - return *rec.cast<bool>(); +NumberValue::NumberValue(int32_t i) { + this->init_tagged(Tag::kInt); + *this->cast<int32_t>() = i; + SkASSERT(this->getTag() == Tag::kInt); } -template <> -double PrimitiveValue<double, Value::Type::kNumber>::operator*() const { - const auto& rec = *reinterpret_cast<const ValueRec*>(this); - - if (rec.is<NullValue>()) return 0; - - SkASSERT(rec.getRecType() == ValueRec::kInt || - rec.getRecType() == ValueRec::kFloat); - - return rec.getRecType() == ValueRec::kInt - ? static_cast<double>(*rec.cast<int32_t>()) - : static_cast<double>(*rec.cast<float>()); +NumberValue::NumberValue(float f) { + this->init_tagged(Tag::kFloat); + *this->cast<float>() = f; + SkASSERT(this->getTag() == Tag::kFloat); } -template <> -size_t VectorValue<Value, Value::Type::kArray>::size() const { - return reinterpret_cast<const ValueRec*>(this)->vectorSize(ValueRec::kArray); +// Vector recs point to externally allocated slabs with the following layout: +// +// [size_t n] [REC_0] ... [REC_n-1] [optional extra trailing storage] +// +// Long strings use extra_alloc_size == 1 to store the \0 terminator. +// +template <typename T, size_t extra_alloc_size = 0> +static void* MakeVector(const void* src, size_t size, SkArenaAlloc& alloc) { + // The Ts are already in memory, so their size should be safe. + const auto total_size = sizeof(size_t) + size * sizeof(T) + extra_alloc_size; + auto* size_ptr = reinterpret_cast<size_t*>(alloc.makeBytesAlignedTo(total_size, kRecAlign)); + auto* data_ptr = reinterpret_cast<void*>(size_ptr + 1); + *size_ptr = size; + memcpy(data_ptr, src, size * sizeof(T)); + + return size_ptr; } -template <> -const Value* VectorValue<Value, Value::Type::kArray>::begin() const { - return reinterpret_cast<const ValueRec*>(this)->vectorBegin<Value>(ValueRec::kArray); +ArrayValue::ArrayValue(const Value* src, size_t size, SkArenaAlloc& alloc) { + this->init_tagged_pointer(Tag::kArray, MakeVector<Value>(src, size, alloc)); + SkASSERT(this->getTag() == Tag::kArray); } -template <> -const Value* VectorValue<Value, Value::Type::kArray>::end() const { - return reinterpret_cast<const ValueRec*>(this)->vectorEnd<Value>(ValueRec::kArray); -} +// Strings have two flavors: +// +// -- short strings (len <= 7) -> these are stored inline, in the record +// (one byte reserved for null terminator/type): +// +// [str] [\0]|[max_len - actual_len] +// +// Storing [max_len - actual_len] allows the 'len' field to double-up as a +// null terminator when size == max_len (this works 'cause kShortString == 0). +// +// -- long strings (len > 7) -> these are externally allocated vectors (VectorRec<char>). +// +// The string data plus a null-char terminator are copied over. +// +StringValue::StringValue(const char* src, size_t size, SkArenaAlloc& alloc) { + if (size > kMaxInlineStringSize) { + this->init_tagged_pointer(Tag::kString, MakeVector<char, 1>(src, size, alloc)); -template <> -size_t VectorValue<Member, Value::Type::kObject>::size() const { - return reinterpret_cast<const ValueRec*>(this)->vectorSize(ValueRec::kObject); -} + auto* data = this->cast<VectorValue<char, Value::Type::kString>>()->begin(); + const_cast<char*>(data)[size] = '\0'; + SkASSERT(this->getTag() == Tag::kString); + return; + } -template <> -const Member* VectorValue<Member, Value::Type::kObject>::begin() const { - return reinterpret_cast<const ValueRec*>(this)->vectorBegin<Member>(ValueRec::kObject); -} + this->init_tagged(Tag::kShortString); -template <> -const Member* VectorValue<Member, Value::Type::kObject>::end() const { - return reinterpret_cast<const ValueRec*>(this)->vectorEnd<Member>(ValueRec::kObject); -} + auto* payload = this->cast<char>(); + memcpy(payload, src, size); + payload[size] = '\0'; -template <> -size_t VectorValue<char, Value::Type::kString>::size() const { - return reinterpret_cast<const ValueRec*>(this)->stringSize(); -} + const auto len_tag = SkTo<char>(kMaxInlineStringSize - size); + // This technically overwrites the tag, but is safe because + // 1) kShortString == 0 + // 2) 0 <= len_tag <= 7 + static_assert(static_cast<uint8_t>(Tag::kShortString) == 0, "please don't break this"); + payload[kMaxInlineStringSize] = len_tag; -template <> -const char* VectorValue<char, Value::Type::kString>::begin() const { - return reinterpret_cast<const ValueRec*>(this)->stringBegin(); + SkASSERT(this->getTag() == Tag::kShortString); } -template <> -const char* VectorValue<char, Value::Type::kString>::end() const { - return reinterpret_cast<const ValueRec*>(this)->stringEnd(); +ObjectValue::ObjectValue(const Member* src, size_t size, SkArenaAlloc& alloc) { + this->init_tagged_pointer(Tag::kObject, MakeVector<Member>(src, size, alloc)); + SkASSERT(this->getTag() == Tag::kObject); } + +// Boring public Value glue. + const Value& ObjectValue::operator[](const char* key) const { // Reverse search for duplicates resolution (policy: return last). const auto* begin = this->begin(); @@ -381,7 +154,8 @@ const Value& ObjectValue::operator[](const char* key) const { } } - return Value::Null(); + static const Value g_null = NullValue(); + return g_null; } namespace { @@ -457,7 +231,7 @@ public: fScopeStack.reserve(kScopeStackReserve); } - const Value& parse(const char* p) { + const Value parse(const char* p) { p = skip_ws(p); switch (*p) { @@ -466,7 +240,7 @@ public: case '[': goto match_array; default: - return this->error(Value::Null(), p, "invalid top-level value"); + return this->error(NullValue(), p, "invalid top-level value"); } match_object: @@ -480,15 +254,15 @@ public: // goto match_object_key; match_object_key: p = skip_ws(p); - if (*p != '"') return this->error(Value::Null(), p, "expected object key"); + if (*p != '"') return this->error(NullValue(), p, "expected object key"); p = this->matchString(p, [this](const char* key, size_t size) { this->pushObjectKey(key, size); }); - if (!p) return Value::Null(); + if (!p) return NullValue(); p = skip_ws(p); - if (*p != ':') return this->error(Value::Null(), p, "expected ':' separator"); + if (*p != ':') return this->error(NullValue(), p, "expected ':' separator"); ++p; @@ -498,7 +272,7 @@ public: switch (*p) { case '\0': - return this->error(Value::Null(), p, "unexpected input end"); + return this->error(NullValue(), p, "unexpected input end"); case '"': p = this->matchString(p, [this](const char* str, size_t size) { this->pushString(str, size); @@ -522,7 +296,7 @@ public: break; } - if (!p) return Value::Null(); + if (!p) return NullValue(); // goto match_post_value; match_post_value: @@ -542,7 +316,7 @@ public: case '}': goto pop_object; default: - return this->error(Value::Null(), p - 1, "unexpected value-trailing token"); + return this->error(NullValue(), p - 1, "unexpected value-trailing token"); } // unreachable @@ -552,7 +326,7 @@ public: SkASSERT(*p == '}'); if (fScopeStack.back() < 0) { - return this->error(Value::Null(), p, "unexpected object terminator"); + return this->error(NullValue(), p, "unexpected object terminator"); } this->popObjectScope(); @@ -565,13 +339,11 @@ public: if (fScopeStack.empty()) { SkASSERT(fValueStack.size() == 1); - auto* root = fAlloc.make<Value>(); - *root = fValueStack.front(); // Stop condition: parsed the top level element and there is no trailing garbage. return *skip_ws(p) == '\0' - ? *root - : this->error(Value::Null(), p, "trailing root garbage"); + ? fValueStack.front() + : this->error(NullValue(), p, "trailing root garbage"); } goto match_post_value; @@ -589,7 +361,7 @@ public: SkASSERT(*p == ']'); if (fScopeStack.back() >= 0) { - return this->error(Value::Null(), p, "unexpected array terminator"); + return this->error(NullValue(), p, "unexpected array terminator"); } this->popArrayScope(); @@ -597,7 +369,7 @@ public: goto pop_common; SkASSERT(false); - return Value::Null(); + return NullValue(); } const SkString& getError() const { @@ -614,11 +386,12 @@ private: SkString fError; - template <typename T> - void popScopeAsVec(ValueRec::RecType type, size_t scope_start) { + template <typename VectorT> + void popScopeAsVec(size_t scope_start) { SkASSERT(scope_start > 0); SkASSERT(scope_start <= fValueStack.size()); + using T = typename VectorT::ValueT; static_assert( sizeof(T) >= sizeof(Value), ""); static_assert( sizeof(T) % sizeof(Value) == 0, ""); static_assert(alignof(T) == alignof(Value), ""); @@ -630,7 +403,7 @@ private: const auto* begin = reinterpret_cast<const T*>(fValueStack.data() + scope_start); // Instantiate the placeholder value added in onPush{Object/Array}. - fValueStack[scope_start - 1] = ValueRec::MakeVector<T>(type, begin, count, fAlloc); + fValueStack[scope_start - 1] = VectorT(begin, count, fAlloc); // Drop the current scope. fScopeStack.pop_back(); @@ -648,7 +421,7 @@ private: void popObjectScope() { const auto scope_start = fScopeStack.back(); SkASSERT(scope_start > 0); - this->popScopeAsVec<Member>(ValueRec::kObject, SkTo<size_t>(scope_start)); + this->popScopeAsVec<ObjectValue>(SkTo<size_t>(scope_start)); SkDEBUGCODE( const auto& obj = fValueStack.back().as<ObjectValue>(); @@ -670,7 +443,7 @@ private: void popArrayScope() { const auto scope_start = -fScopeStack.back(); SkASSERT(scope_start > 0); - this->popScopeAsVec<Value>(ValueRec::kArray, SkTo<size_t>(scope_start)); + this->popScopeAsVec<ArrayValue>(SkTo<size_t>(scope_start)); SkDEBUGCODE( const auto& arr = fValueStack.back().as<ArrayValue>(); @@ -686,31 +459,31 @@ private: } void pushTrue() { - fValueStack.push_back(ValueRec::MakePrimitive<bool>(ValueRec::kBool, true)); + fValueStack.push_back(BoolValue(true)); } void pushFalse() { - fValueStack.push_back(ValueRec::MakePrimitive<bool>(ValueRec::kBool, false)); + fValueStack.push_back(BoolValue(false)); } void pushNull() { - fValueStack.push_back(ValueRec::MakeTypeBound(ValueRec::kNull)); + fValueStack.push_back(NullValue()); } void pushString(const char* s, size_t size) { - fValueStack.push_back(ValueRec::MakeString(s, size, fAlloc)); + fValueStack.push_back(StringValue(s, size, fAlloc)); } void pushInt32(int32_t i) { - fValueStack.push_back(ValueRec::MakePrimitive<int32_t>(ValueRec::kInt, i)); + fValueStack.push_back(NumberValue(i)); } void pushFloat(float f) { - fValueStack.push_back(ValueRec::MakePrimitive<float>(ValueRec::kFloat, f)); + fValueStack.push_back(NumberValue(f)); } template <typename T> - const T& error(const T& ret_val, const char* p, const char* msg) { + T error(T&& ret_val, const char* p, const char* msg) { #if defined(SK_JSON_REPORT_ERRORS) static constexpr size_t kMaxContext = 128; fError = SkStringPrintf("%s: >", msg); @@ -930,17 +703,25 @@ void Write(const Value& v, SkWStream* stream) { } // namespace +SkString Value::toString() const { + SkDynamicMemoryWStream wstream; + Write(*this, &wstream); + const auto data = wstream.detachAsData(); + // TODO: is there a better way to pass data around without copying? + return SkString(static_cast<const char*>(data->data()), data->size()); +} + static constexpr size_t kMinChunkSize = 4096; -DOM::DOM(const char* cstr) +DOM::DOM(const char* c_str) : fAlloc(kMinChunkSize) { DOMParser parser(fAlloc); - fRoot = &parser.parse(cstr); + fRoot = parser.parse(c_str); } void DOM::write(SkWStream* stream) const { - Write(*fRoot, stream); + Write(fRoot, stream); } } // namespace skjson |