644 lines
30 KiB
ReStructuredText
644 lines
30 KiB
ReStructuredText
|
==============================================
|
||
|
LLVM Atomic Instructions and Concurrency Guide
|
||
|
==============================================
|
||
|
|
||
|
.. contents::
|
||
|
:local:
|
||
|
|
||
|
Introduction
|
||
|
============
|
||
|
|
||
|
LLVM supports instructions which are well-defined in the presence of threads and
|
||
|
asynchronous signals.
|
||
|
|
||
|
The atomic instructions are designed specifically to provide readable IR and
|
||
|
optimized code generation for the following:
|
||
|
|
||
|
* The C++11 ``<atomic>`` header. (`C++11 draft available here
|
||
|
<http://www.open-std.org/jtc1/sc22/wg21/>`_.) (`C11 draft available here
|
||
|
<http://www.open-std.org/jtc1/sc22/wg14/>`_.)
|
||
|
|
||
|
* Proper semantics for Java-style memory, for both ``volatile`` and regular
|
||
|
shared variables. (`Java Specification
|
||
|
<http://docs.oracle.com/javase/specs/jls/se8/html/jls-17.html>`_)
|
||
|
|
||
|
* gcc-compatible ``__sync_*`` builtins. (`Description
|
||
|
<https://gcc.gnu.org/onlinedocs/gcc/_005f_005fsync-Builtins.html>`_)
|
||
|
|
||
|
* Other scenarios with atomic semantics, including ``static`` variables with
|
||
|
non-trivial constructors in C++.
|
||
|
|
||
|
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.
|
||
|
|
||
|
.. _Optimization outside atomic:
|
||
|
|
||
|
Optimization outside atomic
|
||
|
===========================
|
||
|
|
||
|
The basic ``'load'`` and ``'store'`` allow a variety of optimizations, but can
|
||
|
lead to undefined results in a concurrent environment; see `NotAtomic`_. This
|
||
|
section specifically goes into the one optimizer restriction which applies in
|
||
|
concurrent environments, which gets a bit more of an extended description
|
||
|
because any optimization dealing with stores needs to be aware of it.
|
||
|
|
||
|
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. Take the following example:
|
||
|
|
||
|
.. code-block:: c
|
||
|
|
||
|
/* C code, for readability; run through clang -O2 -S -emit-llvm to get
|
||
|
equivalent IR */
|
||
|
int x;
|
||
|
void f(int* a) {
|
||
|
for (int i = 0; i < 100; i++) {
|
||
|
if (a[i])
|
||
|
x += 1;
|
||
|
}
|
||
|
}
|
||
|
|
||
|
The following is equivalent in non-concurrent situations:
|
||
|
|
||
|
.. code-block:: c
|
||
|
|
||
|
int x;
|
||
|
void f(int* a) {
|
||
|
int xtemp = x;
|
||
|
for (int i = 0; i < 100; i++) {
|
||
|
if (a[i])
|
||
|
xtemp += 1;
|
||
|
}
|
||
|
x = xtemp;
|
||
|
}
|
||
|
|
||
|
However, LLVM is not allowed to transform the former to the latter: it could
|
||
|
indirectly introduce undefined behavior if another thread can access ``x`` at
|
||
|
the same time. That thread would read `undef` instead of the value it was
|
||
|
expecting, which can lead to undefined behavior down the line. (This example is
|
||
|
particularly of interest because before the concurrency model was implemented,
|
||
|
LLVM would perform this transformation.)
|
||
|
|
||
|
Note that speculative loads are allowed; a load which is part of a race returns
|
||
|
``undef``, but does not have undefined behavior.
|
||
|
|
||
|
Atomic instructions
|
||
|
===================
|
||
|
|
||
|
For cases where simple loads and stores are not sufficient, LLVM provides
|
||
|
various atomic instructions. The exact guarantees provided depend on the
|
||
|
ordering; see `Atomic orderings`_.
|
||
|
|
||
|
``load atomic`` and ``store atomic`` provide the same basic functionality as
|
||
|
non-atomic loads and stores, but provide additional guarantees in situations
|
||
|
where threads and signals are involved.
|
||
|
|
||
|
``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 on any thread between the load and store.
|
||
|
|
||
|
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, and a Monotonic store following a Release fence is roughly
|
||
|
equivalent to a Release store. SequentiallyConsistent fences behave as both
|
||
|
an Acquire and a Release fence, and offer some additional complicated
|
||
|
guarantees, see the C++11 standard for details.
|
||
|
|
||
|
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:
|
||
|
|
||
|
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. (See also `LangRef Ordering <LangRef.html#ordering>`_.)
|
||
|
|
||
|
.. _NotAtomic:
|
||
|
|
||
|
NotAtomic
|
||
|
---------
|
||
|
|
||
|
NotAtomic is the obvious, a load or store which is not atomic. (This isn't
|
||
|
really a level of atomicity, but is listed here for comparison.) This is
|
||
|
essentially a regular load or store. If there is a race on a given memory
|
||
|
location, loads from that location return undef.
|
||
|
|
||
|
Relevant standard
|
||
|
This is intended to match shared variables in C/C++, and to be used in any
|
||
|
other context where memory access is necessary, and a race is impossible. (The
|
||
|
precise definition is in `LangRef Memory Model <LangRef.html#memmodel>`_.)
|
||
|
|
||
|
Notes for frontends
|
||
|
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. If your frontend is
|
||
|
for a "safe" language like Java, use Unordered to load and store any shared
|
||
|
variable. Note that NotAtomic volatile loads and stores are not properly
|
||
|
atomic; do not try to use them as a substitute. (Per the C/C++ standards,
|
||
|
volatile does provide some limited guarantees around asynchronous signals, but
|
||
|
atomics are generally a better solution.)
|
||
|
|
||
|
Notes for optimizers
|
||
|
Introducing loads to shared variables along a codepath where they would not
|
||
|
otherwise exist is allowed; introducing stores to shared variables is not. See
|
||
|
`Optimization outside atomic`_.
|
||
|
|
||
|
Notes for code generation
|
||
|
The one interesting restriction here is that it is not allowed to write to
|
||
|
bytes outside of the bytes relevant to a store. This is mostly relevant to
|
||
|
unaligned stores: it is not allowed in general to convert an unaligned store
|
||
|
into two aligned stores of the same width as the unaligned store. Backends are
|
||
|
also expected to generate an i8 store as an i8 store, and not an instruction
|
||
|
which writes to surrounding bytes. (If you are writing a backend for an
|
||
|
architecture which cannot satisfy these restrictions and cares about
|
||
|
concurrency, please send an email to llvm-dev.)
|
||
|
|
||
|
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 does 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 without LPAE, or not naturally-aligned ``LDRD`` on LPAE 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++11/C11 ``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 are 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++11/C11 ``memory_order_acquire``. It should also be
|
||
|
used for C++11/C11 ``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. It 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++11/C11 ``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 a 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++11/C11 ``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 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++11/C11 ``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 SequentiallyConsistent stores require Release
|
||
|
barriers. Additionally, the code generator must enforce ordering between
|
||
|
SequentiallyConsistent stores followed by SequentiallyConsistent 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:
|
||
|
|
||
|
* ``isSimple()``: A load or store which is not volatile or atomic. This is
|
||
|
what, for example, memcpyopt would check for operations it might transform.
|
||
|
|
||
|
* ``isUnordered()``: A load or store which is not volatile and at most
|
||
|
Unordered. This would be checked, for example, by LICM before hoisting an
|
||
|
operation.
|
||
|
|
||
|
* ``mayReadFromMemory()``/``mayWriteToMemory()``: Existing predicate, but note
|
||
|
that they return true for any operation which is volatile or at least
|
||
|
Monotonic.
|
||
|
|
||
|
* ``isStrongerThan`` / ``isAtLeastOrStrongerThan``: These are predicates on
|
||
|
orderings. They can be useful for passes that are aware of atomics, for
|
||
|
example to do DSE across a single atomic access, but not across a
|
||
|
release-acquire pair (see MemoryDependencyAnalysis for an example of this)
|
||
|
|
||
|
* Alias analysis: Note that AA will return ModRef for anything Acquire or
|
||
|
Release, and for the address accessed by any Monotonic operation.
|
||
|
|
||
|
To support optimizing around atomic operations, make sure you are using the
|
||
|
right predicates; everything should work if that is done. If your pass should
|
||
|
optimize some atomic operations (Unordered operations in particular), make sure
|
||
|
it doesn't replace an atomic load or store with a non-atomic operation.
|
||
|
|
||
|
Some examples of how optimizations interact with various kinds of atomic
|
||
|
operations:
|
||
|
|
||
|
* ``memcpyopt``: An atomic operation cannot be optimized into part of a
|
||
|
memcpy/memset, including unordered loads/stores. It can pull operations
|
||
|
across some atomic operations.
|
||
|
|
||
|
* LICM: Unordered loads/stores can be moved out of a loop. It just treats
|
||
|
monotonic operations like a read+write to a memory location, and anything
|
||
|
stricter than that like a nothrow call.
|
||
|
|
||
|
* DSE: Unordered stores can be DSE'ed like normal stores. Monotonic stores can
|
||
|
be DSE'ed in some cases, but it's tricky to reason about, and not especially
|
||
|
important. It is possible in some case for DSE to operate across a stronger
|
||
|
atomic operation, but it is fairly tricky. DSE delegates this reasoning to
|
||
|
MemoryDependencyAnalysis (which is also used by other passes like GVN).
|
||
|
|
||
|
* Folding a load: Any atomic load from a constant global can be constant-folded,
|
||
|
because it cannot be observed. Similar reasoning allows sroa with
|
||
|
atomic loads and stores.
|
||
|
|
||
|
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 can be emitted by the AtomicExpand Codegen pass if
|
||
|
``setInsertFencesForAtomic()`` was used.
|
||
|
|
||
|
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.
|
||
|
|
||
|
One very important property of the atomic operations is that if your backend
|
||
|
supports any inline lock-free atomic operations of a given size, you should
|
||
|
support *ALL* operations of that size in a lock-free manner.
|
||
|
|
||
|
When the target implements atomic ``cmpxchg`` or LL/SC instructions (as most do)
|
||
|
this is trivial: all the other operations can be implemented on top of those
|
||
|
primitives. However, on many older CPUs (e.g. ARMv5, SparcV8, Intel 80386) there
|
||
|
are atomic load and store instructions, but no ``cmpxchg`` or LL/SC. As it is
|
||
|
invalid to implement ``atomic load`` using the native instruction, but
|
||
|
``cmpxchg`` using a library call to a function that uses a mutex, ``atomic
|
||
|
load`` must *also* expand to a library call on such architectures, so that it
|
||
|
can remain atomic with regards to a simultaneous ``cmpxchg``, by using the same
|
||
|
mutex.
|
||
|
|
||
|
AtomicExpandPass can help with that: it will expand all atomic operations to the
|
||
|
proper ``__atomic_*`` libcalls for any size above the maximum set by
|
||
|
``setMaxAtomicSizeInBitsSupported`` (which defaults to 0).
|
||
|
|
||
|
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 (before v8), 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.).
|
||
|
|
||
|
It is often easiest for backends to use AtomicExpandPass to lower some of the
|
||
|
atomic constructs. Here are some lowerings it can do:
|
||
|
|
||
|
* cmpxchg -> loop with load-linked/store-conditional
|
||
|
by overriding ``shouldExpandAtomicCmpXchgInIR()``, ``emitLoadLinked()``,
|
||
|
``emitStoreConditional()``
|
||
|
* large loads/stores -> ll-sc/cmpxchg
|
||
|
by overriding ``shouldExpandAtomicStoreInIR()``/``shouldExpandAtomicLoadInIR()``
|
||
|
* strong atomic accesses -> monotonic accesses + fences by overriding
|
||
|
``shouldInsertFencesForAtomic()``, ``emitLeadingFence()``, and
|
||
|
``emitTrailingFence()``
|
||
|
* atomic rmw -> loop with cmpxchg or load-linked/store-conditional
|
||
|
by overriding ``expandAtomicRMWInIR()``
|
||
|
* expansion to __atomic_* libcalls for unsupported sizes.
|
||
|
* part-word atomicrmw/cmpxchg -> target-specific intrinsic by overriding
|
||
|
``shouldExpandAtomicRMWInIR``, ``emitMaskedAtomicRMWIntrinsic``,
|
||
|
``shouldExpandAtomicCmpXchgInIR``, and ``emitMaskedAtomicCmpXchgIntrinsic``.
|
||
|
|
||
|
For an example of these look at the ARM (first five lowerings) or RISC-V (last
|
||
|
lowering) backend.
|
||
|
|
||
|
AtomicExpandPass supports two strategies for lowering atomicrmw/cmpxchg to
|
||
|
load-linked/store-conditional (LL/SC) loops. The first expands the LL/SC loop
|
||
|
in IR, calling target lowering hooks to emit intrinsics for the LL and SC
|
||
|
operations. However, many architectures have strict requirements for LL/SC
|
||
|
loops to ensure forward progress, such as restrictions on the number and type
|
||
|
of instructions in the loop. It isn't possible to enforce these restrictions
|
||
|
when the loop is expanded in LLVM IR, and so affected targets may prefer to
|
||
|
expand to LL/SC loops at a very late stage (i.e. after register allocation).
|
||
|
AtomicExpandPass can help support lowering of part-word atomicrmw or cmpxchg
|
||
|
using this strategy by producing IR for any shifting and masking that can be
|
||
|
performed outside of the LL/SC loop.
|
||
|
|
||
|
Libcalls: __atomic_*
|
||
|
====================
|
||
|
|
||
|
There are two kinds of atomic library calls that are generated by LLVM. Please
|
||
|
note that both sets of library functions somewhat confusingly share the names of
|
||
|
builtin functions defined by clang. Despite this, the library functions are
|
||
|
not directly related to the builtins: it is *not* the case that ``__atomic_*``
|
||
|
builtins lower to ``__atomic_*`` library calls and ``__sync_*`` builtins lower
|
||
|
to ``__sync_*`` library calls.
|
||
|
|
||
|
The first set of library functions are named ``__atomic_*``. This set has been
|
||
|
"standardized" by GCC, and is described below. (See also `GCC's documentation
|
||
|
<https://gcc.gnu.org/wiki/Atomic/GCCMM/LIbrary>`_)
|
||
|
|
||
|
LLVM's AtomicExpandPass will translate atomic operations on data sizes above
|
||
|
``MaxAtomicSizeInBitsSupported`` into calls to these functions.
|
||
|
|
||
|
There are four generic functions, which can be called with data of any size or
|
||
|
alignment::
|
||
|
|
||
|
void __atomic_load(size_t size, void *ptr, void *ret, int ordering)
|
||
|
void __atomic_store(size_t size, void *ptr, void *val, int ordering)
|
||
|
void __atomic_exchange(size_t size, void *ptr, void *val, void *ret, int ordering)
|
||
|
bool __atomic_compare_exchange(size_t size, void *ptr, void *expected, void *desired, int success_order, int failure_order)
|
||
|
|
||
|
There are also size-specialized versions of the above functions, which can only
|
||
|
be used with *naturally-aligned* pointers of the appropriate size. In the
|
||
|
signatures below, "N" is one of 1, 2, 4, 8, and 16, and "iN" is the appropriate
|
||
|
integer type of that size; if no such integer type exists, the specialization
|
||
|
cannot be used::
|
||
|
|
||
|
iN __atomic_load_N(iN *ptr, iN val, int ordering)
|
||
|
void __atomic_store_N(iN *ptr, iN val, int ordering)
|
||
|
iN __atomic_exchange_N(iN *ptr, iN val, int ordering)
|
||
|
bool __atomic_compare_exchange_N(iN *ptr, iN *expected, iN desired, int success_order, int failure_order)
|
||
|
|
||
|
Finally there are some read-modify-write functions, which are only available in
|
||
|
the size-specific variants (any other sizes use a ``__atomic_compare_exchange``
|
||
|
loop)::
|
||
|
|
||
|
iN __atomic_fetch_add_N(iN *ptr, iN val, int ordering)
|
||
|
iN __atomic_fetch_sub_N(iN *ptr, iN val, int ordering)
|
||
|
iN __atomic_fetch_and_N(iN *ptr, iN val, int ordering)
|
||
|
iN __atomic_fetch_or_N(iN *ptr, iN val, int ordering)
|
||
|
iN __atomic_fetch_xor_N(iN *ptr, iN val, int ordering)
|
||
|
iN __atomic_fetch_nand_N(iN *ptr, iN val, int ordering)
|
||
|
|
||
|
This set of library functions have some interesting implementation requirements
|
||
|
to take note of:
|
||
|
|
||
|
- They support all sizes and alignments -- including those which cannot be
|
||
|
implemented natively on any existing hardware. Therefore, they will certainly
|
||
|
use mutexes in for some sizes/alignments.
|
||
|
|
||
|
- As a consequence, they cannot be shipped in a statically linked
|
||
|
compiler-support library, as they have state which must be shared amongst all
|
||
|
DSOs loaded in the program. They must be provided in a shared library used by
|
||
|
all objects.
|
||
|
|
||
|
- The set of atomic sizes supported lock-free must be a superset of the sizes
|
||
|
any compiler can emit. That is: if a new compiler introduces support for
|
||
|
inline-lock-free atomics of size N, the ``__atomic_*`` functions must also have a
|
||
|
lock-free implementation for size N. This is a requirement so that code
|
||
|
produced by an old compiler (which will have called the ``__atomic_*`` function)
|
||
|
interoperates with code produced by the new compiler (which will use native
|
||
|
the atomic instruction).
|
||
|
|
||
|
Note that it's possible to write an entirely target-independent implementation
|
||
|
of these library functions by using the compiler atomic builtins themselves to
|
||
|
implement the operations on naturally-aligned pointers of supported sizes, and a
|
||
|
generic mutex implementation otherwise.
|
||
|
|
||
|
Libcalls: __sync_*
|
||
|
==================
|
||
|
|
||
|
Some targets or OS/target combinations can support lock-free atomics, but for
|
||
|
various reasons, it is not practical to emit the instructions inline.
|
||
|
|
||
|
There's two typical examples of this.
|
||
|
|
||
|
Some CPUs support multiple instruction sets which can be switched back and forth
|
||
|
on function-call boundaries. For example, MIPS supports the MIPS16 ISA, which
|
||
|
has a smaller instruction encoding than the usual MIPS32 ISA. ARM, similarly,
|
||
|
has the Thumb ISA. In MIPS16 and earlier versions of Thumb, the atomic
|
||
|
instructions are not encodable. However, those instructions are available via a
|
||
|
function call to a function with the longer encoding.
|
||
|
|
||
|
Additionally, a few OS/target pairs provide kernel-supported lock-free
|
||
|
atomics. ARM/Linux is an example of this: the kernel `provides
|
||
|
<https://www.kernel.org/doc/Documentation/arm/kernel_user_helpers.txt>`_ a
|
||
|
function which on older CPUs contains a "magically-restartable" atomic sequence
|
||
|
(which looks atomic so long as there's only one CPU), and contains actual atomic
|
||
|
instructions on newer multicore models. This sort of functionality can typically
|
||
|
be provided on any architecture, if all CPUs which are missing atomic
|
||
|
compare-and-swap support are uniprocessor (no SMP). This is almost always the
|
||
|
case. The only common architecture without that property is SPARC -- SPARCV8 SMP
|
||
|
systems were common, yet it doesn't support any sort of compare-and-swap
|
||
|
operation.
|
||
|
|
||
|
In either of these cases, the Target in LLVM can claim support for atomics of an
|
||
|
appropriate size, and then implement some subset of the operations via libcalls
|
||
|
to a ``__sync_*`` function. Such functions *must* not use locks in their
|
||
|
implementation, because unlike the ``__atomic_*`` routines used by
|
||
|
AtomicExpandPass, these may be mixed-and-matched with native instructions by the
|
||
|
target lowering.
|
||
|
|
||
|
Further, these routines do not need to be shared, as they are stateless. So,
|
||
|
there is no issue with having multiple copies included in one binary. Thus,
|
||
|
typically these routines are implemented by the statically-linked compiler
|
||
|
runtime support library.
|
||
|
|
||
|
LLVM will emit a call to an appropriate ``__sync_*`` routine if the target
|
||
|
ISelLowering code has set the corresponding ``ATOMIC_CMPXCHG``, ``ATOMIC_SWAP``,
|
||
|
or ``ATOMIC_LOAD_*`` operation to "Expand", and if it has opted-into the
|
||
|
availability of those library functions via a call to ``initSyncLibcalls()``.
|
||
|
|
||
|
The full set of functions that may be called by LLVM is (for ``N`` being 1, 2,
|
||
|
4, 8, or 16)::
|
||
|
|
||
|
iN __sync_val_compare_and_swap_N(iN *ptr, iN expected, iN desired)
|
||
|
iN __sync_lock_test_and_set_N(iN *ptr, iN val)
|
||
|
iN __sync_fetch_and_add_N(iN *ptr, iN val)
|
||
|
iN __sync_fetch_and_sub_N(iN *ptr, iN val)
|
||
|
iN __sync_fetch_and_and_N(iN *ptr, iN val)
|
||
|
iN __sync_fetch_and_or_N(iN *ptr, iN val)
|
||
|
iN __sync_fetch_and_xor_N(iN *ptr, iN val)
|
||
|
iN __sync_fetch_and_nand_N(iN *ptr, iN val)
|
||
|
iN __sync_fetch_and_max_N(iN *ptr, iN val)
|
||
|
iN __sync_fetch_and_umax_N(iN *ptr, iN val)
|
||
|
iN __sync_fetch_and_min_N(iN *ptr, iN val)
|
||
|
iN __sync_fetch_and_umin_N(iN *ptr, iN val)
|
||
|
|
||
|
This list doesn't include any function for atomic load or store; all known
|
||
|
architectures support atomic loads and stores directly (possibly by emitting a
|
||
|
fence on either side of a normal load or store.)
|
||
|
|
||
|
There's also, somewhat separately, the possibility to lower ``ATOMIC_FENCE`` to
|
||
|
``__sync_synchronize()``. This may happen or not happen independent of all the
|
||
|
above, controlled purely by ``setOperationAction(ISD::ATOMIC_FENCE, ...)``.
|
||
|
|
||
|
On AArch64, a variant of the __sync_* routines is used which contain the memory
|
||
|
order as part of the function name. These routines may determine at runtime
|
||
|
whether the single-instruction atomic operations which were introduced as part
|
||
|
of AArch64 Large System Extensions "LSE" instruction set are available, or if
|
||
|
it needs to fall back to an LL/SC loop. The following helper functions are
|
||
|
implemented in both ``compiler-rt`` and ``libgcc`` libraries
|
||
|
(``N`` is one of 1, 2, 4, 8, and ``M`` is one of 1, 2, 4, 8 and 16, and
|
||
|
``ORDER`` is one of 'relax', 'acq', 'rel', 'acq_rel')::
|
||
|
|
||
|
iM __aarch64_casM_ORDER(iM expected, iM desired, iM *ptr)
|
||
|
iN __aarch64_swpN_ORDER(iN val, iN *ptr)
|
||
|
iN __aarch64_ldaddN_ORDER(iN val, iN *ptr)
|
||
|
iN __aarch64_ldclrN_ORDER(iN val, iN *ptr)
|
||
|
iN __aarch64_ldeorN_ORDER(iN val, iN *ptr)
|
||
|
iN __aarch64_ldsetN_ORDER(iN val, iN *ptr)
|
||
|
|
||
|
Please note, if LSE instruction set is specified for AArch64 target then
|
||
|
out-of-line atomics calls are not generated and single-instruction atomic
|
||
|
operations are used in place.
|