Transactional memory
Transactional memory is a concurrency synchronization mechanism that combines groups of statements in transactions, that are
- atomic (either all statements occur, or nothing occurs)
- isolated (statements in a transaction may not observe half-written writes made by another transaction, even if they execute in parallel)
Typical implementations use hardware transactional memory where supported and to the limits that it is available (e.g. until the changeset is saturated) and fall back to software transactional memory, usually implemented with optimistic concurrency: if another transaction updated some of the variables used by a transaction, it is silently retried. For that reason, retriable transactions ("atomic blocks") can only call transaction-safe functions.
Note that accessing a variable in a transaction and out of a transaction without other external synchronization is a data race.
If feature testing is supported, the features described here are indicated by the macro constant __cpp_transactional_memory with a value equal or greater 201505.
Contents |
Synchronized blocks
synchronized compound-statement
Executes the compound statement as if under a global lock: all outermost synchronized blocks in the program execute in a single total order. The end of each synchronized block synchronizes with the beginning of the next synchronized block in that order. Synchronized blocks that are nested within other synchronized blocks have no special semantics.
Synchronized blocks are not transactions (unlike the atomic blocks below) and may call transaction-unsafe functions.
#include <iostream> #include <vector> #include <thread> int f() { static int i = 0; synchronized { // begin synchronized block std::cout << i << " -> "; ++i; // each call to f() obtains a unique value of i std::cout << i << '\n'; return i; // end synchronized block } } int main() { std::vector<std::thread> v(10); for(auto& t: v) t = std::thread([]{ for(int n = 0; n < 10; ++n) f(); }); for(auto& t: v) t.join(); }
Output:
0 -> 1 1 -> 2 2 -> 3 ... 99 -> 100
Leaving a synchronized block by any means (reaching the end, executing goto, break, continue, or return, or throwing an exception) exits the block and synchronizes-with the next block in the single total order if the exited block was an outer block.
Entering a synchronized block by goto or switch is not allowed.
Although synchronized blocks execute as-if under a global lock, the implementations are expected to examine the code within each block and use optimistic concurrency (backed up by hardware transactional memory where available) for transaction-safe code and minimal locking for non-transaction safe code. When a synchronized block makes a call to a non-inlined function, the compiler may have to hold a lock around the entire call unless the attribute [[optimize_for_synchronized]] (see below) is used.
Atomic blocks
| This section is incomplete |
atomic_noexcept compound-statement
atomic_cancel compound-statement
atomic_commit compound-statement
The exceptions used for transaction cancellation in atomic_cancel blocks are std::bad_alloc, std::bad_array_length, std::bad_array_new_length, std::bad_cast, std::bad_typeid, std::bad_exception, std::exception and all standard library exceptions derived from it, and the new exception std::tx_exception<T>
The compound-statement in an atomic block is not allowed to execute any expression or statement or call any function that isn't transaction_safe (this is a compile time error)
// each call to f() retrieves a unique value of i, even when done in parallel int f() { static int i = 0; atomic_noexcept { // begin transaction // printf("before %d\n", i); // error: cannot call a non transaction-safe function ++i; return i; // commit transaction } }
Leaving an atomic block by any means other than exception (reaching the end, goto, break, continue, return) commits the transaction.
Transaction-safe functions
| This section is incomplete |
A function can be explicitly declared to be transaction-safe by using the keyword transaction_safe in its declaration.
| This section is incomplete |
In a lambda declaration, it appears either immediately after the capture list, or immediately after the (keyword mutable (if one is used).
| This section is incomplete |
extern volatile int * p = 0; struct S { virtual ~S(); }; int f() transaction_safe { int x = 0; // ok: not volatile p = &x; // ok: the pointer is not volatile int i = *p; // error: read through volatile glvalue S s; // error: invocation of unsafe destructor }
int f(int x) { // implicitly transaction-safe if (x <= 0) return 0; return x + f(x-1); }
If a function that is not transaction-safe is called through a reference or pointer to a non-transaction-safe function, the behavior is undefined.
Pointers to transaction-safe functions and pointers to transaction-safe member functions are implicitly convertible to pointers to functions and pointers to member functions respectively. It is unspecified if the resulting pointer compares equal to the original.
Transaction-safe virtual functions
| This section is incomplete |
If the final overrider of a transaction_safe_dynamic function is not declared transaction_safe, calling it in an atomic block is undefined behavior.
Attributes
The attribute [[optimize_for_synchronized]] may be applied to a declarator in a function declaration and must appear on the first declaration of the function.
If a function is declared [[optimize_for_synchronized]] in one translation unit and the same function is declared without [[optimize_for_synchronized]] in another translation unit, the program is ill-formed; no diagnostic required.
It indicates that a the function definition should be optimized for invocation from a synchronized statement. In particular, it avoids serializing synchronized blocks that make a call to a function that is transaction-safe for the majority of calls, but not for all calls (e.g. hash table insertion that may have to rehash, allocator that may have to request a new block, a simple function that may rarely log)
std::atomic<bool> rehash{false}; // maintenance thread runs this loop void maintenance_thread(void*) { while (!shutdown) { synchronized { if (rehash) { hash.rehash(); rehash = false; } } } } // worker threads execute hundreds of thousands of calls to this function // every second. Calls to insert_key() from synchronized blocks in other // translation units will cause those blocks to serialize, unless insert_key() // is marked [[optimize_for_synchronized]] [[optimize_for_synchronized]] void insert_key(char* key, char* value) { bool concern = hash.insert(key, value); if (concern) rehash = true; } // GCC assembly without the attribute: the entire function is serialized insert_key(char*, char*): subq $8, %rsp movq %rsi, %rdx movq %rdi, %rsi movl $hash, %edi call Hash::insert(char*, char*) testb %al, %al je .L20 movb $1, rehash(%rip) mfence .L20: addq $8, %rsp ret // GCC assembly with the attribute: transaction clone for insert_key(char*, char*): subq $8, %rsp movq %rsi, %rdx movq %rdi, %rsi movl $hash, %edi call transaction clone for Hash::insert(char*, char*) testb %al, %al je .L27 xorl %edi, %edi call _ITM_changeTransactionMode // Note: this is the serialization point movb $1, rehash(%rip) mfence .L27: addq $8, %rsp ret
| This section is incomplete Reason: check assembly with trunk, also show caller-side changes |
Notes
| This section is incomplete Reason: experience notes from Wyatt paper/talk |
Compiler support
This technical specification is supported by GCC as of version 6.1 (requires -fgnu-tm to enable). An older variant of this specification was supported in GCC as of 4.7.
