LLVM Atomic Instructions and Concurrency Guide

  1. Introduction
  2. Load and store
  3. Other atomic instructions
  4. Atomic orderings
  5. Atomics and IR optimization
  6. Atomics and Codegen

Written by Eli Friedman

Introduction

Historically, LLVM has not had very strong support for concurrency; some minimal intrinsics were provided, and volatile was used in some cases to achieve rough semantics in the presence of concurrency. However, this is changing; there are now new instructions which are well-defined in the presence of threads and asynchronous signals, and the model for existing instructions has been clarified in the IR.

The atomic instructions are designed specifically to provide readable IR and optimized code generation for the following:

Atomic and volatile in the IR are orthogonal; "volatile" is the C/C++ volatile, which ensures that every volatile load and store happens and is performed in the stated order. A couple examples: if a SequentiallyConsistent store is immediately followed by another SequentiallyConsistent store to the same address, the first store can be erased. This transformation is not allowed for a pair of volatile stores. On the other hand, a non-volatile non-atomic load can be moved across a volatile load freely, but not an Acquire load.

This document is intended to provide a guide to anyone either writing a frontend for LLVM or working on optimization passes for LLVM with a guide for how to deal with instructions with special semantics in the presence of concurrency. This is not intended to be a precise guide to the semantics; the details can get extremely complicated and unreadable, and are not usually necessary.

Load and store

The basic 'load' and 'store' allow a variety of optimizations, but can have unintuitive results in a concurrent environment. For a frontend writer, the rule is essentially that all memory accessed with basic loads and stores by multiple threads should be protected by a lock or other synchronization; otherwise, you are likely to run into undefined behavior. (Do not use volatile as a substitute for atomics; it might work on some platforms, but does not provide the necessary guarantees in general.)

From the optimizer's point of view, the rule is that if there are not any instructions with atomic ordering involved, concurrency does not matter, with one exception: if a variable might be visible to another thread or signal handler, a store cannot be inserted along a path where it might not execute otherwise. For example, suppose LICM wants to take all the loads and stores in a loop to and from a particular address and promote them to registers. LICM is not allowed to insert an unconditional store after the loop with the computed value unless a store unconditionally executes within the loop. Note that speculative loads are allowed; a load which is part of a race returns undef, but does not have undefined behavior.

For cases where simple loads and stores are not sufficient, LLVM provides atomic loads and stores with varying levels of guarantees.

Other atomic instructions

cmpxchg and atomicrmw are essentially like an atomic load followed by an atomic store (where the store is conditional for cmpxchg), but no other memory operation can happen between the load and store. Note that our cmpxchg does not have quite as many options for making cmpxchg weaker as the C++0x version.

A fence provides Acquire and/or Release ordering which is not part of another operation; it is normally used along with Monotonic memory operations. A Monotonic load followed by an Acquire fence is roughly equivalent to an Acquire load.

Frontends generating atomic instructions generally need to be aware of the target to some degree; atomic instructions are guaranteed to be lock-free, and therefore an instruction which is wider than the target natively supports can be impossible to generate.

Atomic orderings

In order to achieve a balance between performance and necessary guarantees, there are six levels of atomicity. They are listed in order of strength; each level includes all the guarantees of the previous level except for Acquire/Release.

Unordered

Unordered is the lowest level of atomicity. It essentially guarantees that races produce somewhat sane results instead of having undefined behavior. It also guarantees the operation to be lock-free, so it do not depend on the data being part of a special atomic structure or depend on a separate per-process global lock. Note that code generation will fail for unsupported atomic operations; if you need such an operation, use explicit locking.

Relevant standard
This is intended to match the Java memory model for shared variables.
Notes for frontends
This cannot be used for synchronization, but is useful for Java and other "safe" languages which need to guarantee that the generated code never exhibits undefined behavior. Note that this guarantee is cheap on common platforms for loads of a native width, but can be expensive or unavailable for wider loads, like a 64-bit store on ARM. (A frontend for Java or other "safe" languages would normally split a 64-bit store on ARM into two 32-bit unordered stores.)
Notes for optimizers
In terms of the optimizer, this prohibits any transformation that transforms a single load into multiple loads, transforms a store into multiple stores, narrows a store, or stores a value which would not be stored otherwise. Some examples of unsafe optimizations are narrowing an assignment into a bitfield, rematerializing a load, and turning loads and stores into a memcpy call. Reordering unordered operations is safe, though, and optimizers should take advantage of that because unordered operations are common in languages that need them.
Notes for code generation
These operations are required to be atomic in the sense that if you use unordered loads and unordered stores, a load cannot see a value which was never stored. A normal load or store instruction is usually sufficient, but note that an unordered load or store cannot be split into multiple instructions (or an instruction which does multiple memory operations, like LDRD on ARM).

Monotonic

Monotonic is the weakest level of atomicity that can be used in synchronization primitives, although it does not provide any general synchronization. It essentially guarantees that if you take all the operations affecting a specific address, a consistent ordering exists.

Relevant standard
This corresponds to the C++0x/C1x memory_order_relaxed; see those standards for the exact definition.
Notes for frontends
If you are writing a frontend which uses this directly, use with caution. The guarantees in terms of synchronization are very weak, so make sure these are only used in a pattern which you know is correct. Generally, these would either be used for atomic operations which do not protect other memory (like an atomic counter), or along with a fence.
Notes for optimizers
In terms of the optimizer, this can be treated as a read+write on the relevant memory location (and alias analysis will take advantage of that). In addition, it is legal to reorder non-atomic and Unordered loads around Monotonic loads. CSE/DSE and a few other optimizations are allowed, but Monotonic operations are unlikely to be used in ways which would make those optimizations useful.
Notes for code generation
Code generation is essentially the same as that for unordered for loads and stores. No fences is required. cmpxchg and atomicrmw are required to appear as a single operation.

Acquire

Acquire provides a barrier of the sort necessary to acquire a lock to access other memory with normal loads and stores.

Relevant standard
This corresponds to the C++0x/C1x memory_order_acquire. It should also be used for C++0x/C1x memory_order_consume.
Notes for frontends
If you are writing a frontend which uses this directly, use with caution. Acquire only provides a semantic guarantee when paired with a Release operation.
Notes for optimizers
Optimizers not aware of atomics can treat this like a nothrow call. Tt is also possible to move stores from before an Acquire load or read-modify-write operation to after it, and move non-Acquire loads from before an Acquire operation to after it.
Notes for code generation
Architectures with weak memory ordering (essentially everything relevant today except x86 and SPARC) require some sort of fence to maintain the Acquire semantics. The precise fences required varies widely by architecture, but for a simple implementation, most architectures provide a barrier which is strong enough for everything (dmb on ARM, sync on PowerPC, etc.). Putting such a fence after the equivalent Monotonic operation is sufficient to maintain Acquire semantics for a memory operation.

Release

Release is similar to Acquire, but with a barrier of the sort necessary to release a lock.

Relevant standard
This corresponds to the C++0x/C1x memory_order_release.
Notes for frontends
If you are writing a frontend which uses this directly, use with caution. Release only provides a semantic guarantee when paired with a Acquire operation.
Notes for optimizers
Optimizers not aware of atomics can treat this like a nothrow call. It is also possible to move loads from after a Release store or read-modify-write operation to before it, and move non-Release stores from after an Release operation to before it.
Notes for code generation
See the section on Acquire; a fence before the relevant operation is usually sufficient for Release. Note that a store-store fence is not sufficient to implement Release semantics; store-store fences are generally not exposed to IR because they are extremely difficult to use correctly.

AcquireRelease

AcquireRelease (acq_rel in IR) provides both an Acquire and a Release barrier (for fences and operations which both read and write memory).

Relevant standard
This corresponds to the C++0x/C1x memory_order_acq_rel.
Notes for frontends
If you are writing a frontend which uses this directly, use with caution. Acquire only provides a semantic guarantee when paired with a Release operation, and vice versa.
Notes for optimizers
In general, optimizers should treat this like a nothrow call; the the possible optimizations are usually not interesting.
Notes for code generation
This operation has Acquire and Release semantics; see the sections on Acquire and Release.

SequentiallyConsistent

SequentiallyConsistent (seq_cst in IR) provides Acquire semantics for loads and Release semantics for stores. Additionally, it guarantees that a total ordering exists between all SequentiallyConsistent operations.

Relevant standard
This corresponds to the C++0x/C1x memory_order_seq_cst, Java volatile, and the gcc-compatible __sync_* builtins which do not specify otherwise.
Notes for frontends
If a frontend is exposing atomic operations, these are much easier to reason about for the programmer than other kinds of operations, and using them is generally a practical performance tradeoff.
Notes for optimizers
Optimizers not aware of atomics can treat this like a nothrow call. For SequentiallyConsistent loads and stores, the same reorderings are allowed as for Acquire loads and Release stores, except that SequentiallyConsistent operations may not be reordered.
Notes for code generation
SequentiallyConsistent loads minimally require the same barriers as Acquire operations and SequeuentiallyConsistent stores require Release barriers. Additionally, the code generator must enforce ordering between SequeuentiallyConsistent stores followed by SequeuentiallyConsistent loads. This is usually done by emitting either a full fence before the loads or a full fence after the stores; which is preferred varies by architecture.

Atomics and IR optimization

Predicates for optimizer writers to query:

There are essentially two components to supporting atomic operations. The first is making sure to query isSimple() or isUnordered() instead of isVolatile() before transforming an operation. The other piece is making sure that a transform does not end up replacing, for example, an Unordered operation with a non-atomic operation. Most of the other necessary checks automatically fall out from existing predicates and alias analysis queries.

Some examples of how optimizations interact with various kinds of atomic operations:

Atomics and Codegen

Atomic operations are represented in the SelectionDAG with ATOMIC_* opcodes. On architectures which use barrier instructions for all atomic ordering (like ARM), appropriate fences are split out as the DAG is built.

The MachineMemOperand for all atomic operations is currently marked as volatile; this is not correct in the IR sense of volatile, but CodeGen handles anything marked volatile very conservatively. This should get fixed at some point.

Common architectures have some way of representing at least a pointer-sized lock-free cmpxchg; such an operation can be used to implement all the other atomic operations which can be represented in IR up to that size. Backends are expected to implement all those operations, but not operations which cannot be implemented in a lock-free manner. It is expected that backends will give an error when given an operation which cannot be implemented. (The LLVM code generator is not very helpful here at the moment, but hopefully that will change.)

The implementation of atomics on LL/SC architectures (like ARM) is currently a bit of a mess; there is a lot of copy-pasted code across targets, and the representation is relatively unsuited to optimization (it would be nice to be able to optimize loops involving cmpxchg etc.).

On x86, all atomic loads generate a MOV. SequentiallyConsistent stores generate an XCHG, other stores generate a MOV. SequentiallyConsistent fences generate an MFENCE, other fences do not cause any code to be generated. cmpxchg uses the LOCK CMPXCHG instruction. atomicrmw xchg uses XCHG, atomicrmw add and atomicrmw sub use XADD, and all other atomicrmw operations generate a loop with LOCK CMPXCHG. Depending on the users of the result, some atomicrmw operations can be translated into operations like LOCK AND, but that does not work in general.

On ARM, MIPS, and many other RISC architectures, Acquire, Release, and SequentiallyConsistent semantics require barrier instructions for every such operation. Loads and stores generate normal instructions. cmpxchg and atomicrmw can be represented using a loop with LL/SC-style instructions which take some sort of exclusive lock on a cache line (LDREX and STREX on ARM, etc.). At the moment, the IR does not provide any way to represent a weak cmpxchg which would not require a loop.


Valid CSS Valid HTML 4.01 LLVM Compiler Infrastructure
Last modified: $Date: 2011-08-09 02:07:00 -0700 (Tue, 09 Aug 2011) $