1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
2 "http://www.w3.org/TR/html4/strict.dtd">
5 <title>LLVM Atomic Instructions and Concurrency Guide</title>
6 <meta http-equiv="Content-Type" content="text/html; charset=utf-8">
7 <link rel="stylesheet" href="llvm.css" type="text/css">
12 LLVM Atomic Instructions and Concurrency Guide
16 <li><a href="#introduction">Introduction</a></li>
17 <li><a href="#loadstore">Load and store</a></li>
18 <li><a href="#ordering">Atomic orderings</a></li>
19 <li><a href="#otherinst">Other atomic instructions</a></li>
20 <li><a href="#iropt">Atomics and IR optimization</a></li>
21 <li><a href="#codegen">Atomics and Codegen</a></li>
24 <div class="doc_author">
25 <p>Written by Eli Friedman</p>
28 <!-- *********************************************************************** -->
30 <a name="introduction">Introduction</a>
32 <!-- *********************************************************************** -->
36 <p>Historically, LLVM has not had very strong support for concurrency; some
37 minimal intrinsics were provided, and <code>volatile</code> was used in some
38 cases to achieve rough semantics in the presence of concurrency. However, this
39 is changing; there are now new instructions which are well-defined in the
40 presence of threads and asynchronous signals, and the model for existing
41 instructions has been clarified in the IR.</p>
43 <p>The atomic instructions are designed specifically to provide readable IR and
44 optimized code generation for the following:</p>
46 <li>The new C++0x <code><atomic></code> header.</li>
47 <li>Proper semantics for Java-style memory, for both <code>volatile</code> and
48 regular shared variables.</li>
49 <li>gcc-compatible <code>__sync_*</code> builtins.</li>
50 <li>Other scenarios with atomic semantics, including <code>static</code>
51 variables with non-trivial constructors in C++.</li>
54 <p>This document is intended to provide a guide to anyone either writing a
55 frontend for LLVM or working on optimization passes for LLVM with a guide
56 for how to deal with instructions with special semantics in the presence of
57 concurrency. This is not intended to be a precise guide to the semantics;
58 the details can get extremely complicated and unreadable, and are not
59 usually necessary.</p>
63 <!-- *********************************************************************** -->
65 <a name="loadstore">Load and store</a>
67 <!-- *********************************************************************** -->
71 <p>The basic <code>'load'</code> and <code>'store'</code> allow a variety of
72 optimizations, but can have unintuitive results in a concurrent environment.
73 For a frontend writer, the rule is essentially that all memory accessed
74 with basic loads and stores by multiple threads should be protected by a
75 lock or other synchronization; otherwise, you are likely to run into
76 undefined behavior. (Do not use volatile as a substitute for atomics; it
77 might work on some platforms, but does not provide the necessary guarantees
80 <p>From the optimizer's point of view, the rule is that if there
81 are not any instructions with atomic ordering involved, concurrency does not
82 matter, with one exception: if a variable might be visible to another
83 thread or signal handler, a store cannot be inserted along a path where it
84 might not execute otherwise. Note that speculative loads are allowed;
85 a load which is part of a race returns <code>undef</code>, but is not
86 undefined behavior.</p>
88 <p>For cases where simple loads and stores are not sufficient, LLVM provides
89 atomic loads and stores with varying levels of guarantees.</p>
93 <!-- *********************************************************************** -->
95 <a name="ordering">Atomic orderings</a>
97 <!-- *********************************************************************** -->
101 <p>In order to achieve a balance between performance and necessary guarantees,
102 there are six levels of atomicity. They are listed in order of strength;
103 each level includes all the guarantees of the previous level except for
106 <p>Unordered is the lowest level of atomicity. It essentially guarantees that
107 races produce somewhat sane results instead of having undefined behavior.
108 This is intended to match the Java memory model for shared variables. It
109 cannot be used for synchronization, but is useful for Java and other
110 "safe" languages which need to guarantee that the generated code never
111 exhibits undefined behavior. Note that this guarantee is cheap on common
112 platforms for loads of a native width, but can be expensive or unavailable
113 for wider loads, like a 64-bit load on ARM. (A frontend for a "safe"
114 language would normally split a 64-bit load on ARM into two 32-bit
115 unordered loads.) In terms of the optimizer, this prohibits any
116 transformation that transforms a single load into multiple loads,
117 transforms a store into multiple stores, narrows a store, or stores a
118 value which would not be stored otherwise. Some examples of unsafe
119 optimizations are narrowing an assignment into a bitfield, rematerializing
120 a load, and turning loads and stores into a memcpy call. Reordering
121 unordered operations is safe, though, and optimizers should take
122 advantage of that because unordered operations are common in
123 languages that need them.</p>
125 <p>Monotonic is the weakest level of atomicity that can be used in
126 synchronization primitives, although it does not provide any general
127 synchronization. It essentially guarantees that if you take all the
128 operations affecting a specific address, a consistent ordering exists.
129 This corresponds to the C++0x/C1x <code>memory_order_relaxed</code>; see
130 those standards for the exact definition. If you are writing a frontend, do
131 not use the low-level synchronization primitives unless you are compiling
132 a language which requires it or are sure a given pattern is correct. In
133 terms of the optimizer, this can be treated as a read+write on the relevant
134 memory location (and alias analysis will take advantage of that). In
135 addition, it is legal to reorder non-atomic and Unordered loads around
136 Monotonic loads. CSE/DSE and a few other optimizations are allowed, but
137 Monotonic operations are unlikely to be used in ways which would make
138 those optimizations useful.</p>
140 <p>Acquire provides a barrier of the sort necessary to acquire a lock to access
141 other memory with normal loads and stores. This corresponds to the
142 C++0x/C1x <code>memory_order_acquire</code>. It should also be used for
143 C++0x/C1x <code>memory_order_consume</code>. This is a low-level
144 synchronization primitive. In general, optimizers should treat this like
147 <p>Release is similar to Acquire, but with a barrier of the sort necessary to
148 release a lock. This corresponds to the C++0x/C1x
149 <code>memory_order_release</code>. In general, optimizers should treat this
150 like a nothrow call.</p>
152 <p>AcquireRelease (<code>acq_rel</code> in IR) provides both an Acquire and a Release barrier.
153 This corresponds to the C++0x/C1x <code>memory_order_acq_rel</code>. In general,
154 optimizers should treat this like a nothrow call.</p>
156 <p>SequentiallyConsistent (<code>seq_cst</code> in IR) provides Acquire and/or
157 Release semantics, and in addition guarantees a total ordering exists with
158 all other SequentiallyConsistent operations. This corresponds to the
159 C++0x/C1x <code>memory_order_seq_cst</code>, and Java volatile. The intent
160 of this ordering level is to provide a programming model which is relatively
161 easy to understand. In general, optimizers should treat this like a
166 <!-- *********************************************************************** -->
168 <a name="otherinst">Other atomic instructions</a>
170 <!-- *********************************************************************** -->
174 <p><code>cmpxchg</code> and <code>atomicrmw</code> are essentially like an
175 atomic load followed by an atomic store (where the store is conditional for
176 <code>cmpxchg</code>), but no other memory operation can happen between
177 the load and store. Note that our cmpxchg does not have quite as many
178 options for making cmpxchg weaker as the C++0x version.</p>
180 <p>A <code>fence</code> provides Acquire and/or Release ordering which is not
181 part of another operation; it is normally used along with Monotonic memory
182 operations. A Monotonic load followed by an Acquire fence is roughly
183 equivalent to an Acquire load.</p>
185 <p>Frontends generating atomic instructions generally need to be aware of the
186 target to some degree; atomic instructions are guaranteed to be lock-free,
187 and therefore an instruction which is wider than the target natively supports
188 can be impossible to generate.</p>
192 <!-- *********************************************************************** -->
194 <a name="iropt">Atomics and IR optimization</a>
196 <!-- *********************************************************************** -->
200 <p>Predicates for optimizer writers to query:
202 <li>isSimple(): A load or store which is not volatile or atomic. This is
203 what, for example, memcpyopt would check for operations it might
205 <li>isUnordered(): A load or store which is not volatile and at most
206 Unordered. This would be checked, for example, by LICM before hoisting
208 <li>mayReadFromMemory()/mayWriteToMemory(): Existing predicate, but note
209 that they return true for any operation which is volatile or at least
211 <li>Alias analysis: Note that AA will return ModRef for anything Acquire or
212 Release, and for the address accessed by any Monotonic operation.
215 <p>There are essentially two components to supporting atomic operations. The
216 first is making sure to query isSimple() or isUnordered() instead
217 of isVolatile() before transforming an operation. The other piece is
218 making sure that a transform does not end up replacing, for example, an
219 Unordered operation with a non-atomic operation. Most of the other
220 necessary checks automatically fall out from existing predicates and
221 alias analysis queries.</p>
223 <p>Some examples of how optimizations interact with various kinds of atomic
226 <li>memcpyopt: An atomic operation cannot be optimized into part of a
227 memcpy/memset, including unordered loads/stores. It can pull operations
228 across some atomic operations.
229 <li>LICM: Unordered loads/stores can be moved out of a loop. It just treats
230 monotonic operations like a read+write to a memory location, and anything
231 stricter than that like a nothrow call.
232 <li>DSE: Unordered stores can be DSE'ed like normal stores. Monotonic stores
233 can be DSE'ed in some cases, but it's tricky to reason about, and not
234 especially important.
235 <li>Folding a load: Any atomic load from a constant global can be
236 constant-folded, because it cannot be observed. Similar reasoning allows
237 scalarrepl with atomic loads and stores.
242 <!-- *********************************************************************** -->
244 <a name="codegen">Atomics and Codegen</a>
246 <!-- *********************************************************************** -->
250 <p>Atomic operations are represented in the SelectionDAG with
251 <code>ATOMIC_*</code> opcodes. On architectures which use barrier
252 instructions for all atomic ordering (like ARM), appropriate fences are
253 split out as the DAG is built.</p>
255 <p>The MachineMemOperand for all atomic operations is currently marked as
256 volatile; this is not correct in the IR sense of volatile, but CodeGen
257 handles anything marked volatile very conservatively. This should get
258 fixed at some point.</p>
260 <p>The implementation of atomics on LL/SC architectures (like ARM) is currently
261 a bit of a mess; there is a lot of copy-pasted code across targets, and
262 the representation is relatively unsuited to optimization (it would be nice
263 to be able to optimize loops involving cmpxchg etc.).</p>
265 <p>On x86, all atomic loads generate a <code>MOV</code>.
266 SequentiallyConsistent stores generate an <code>XCHG</code>, other stores
267 generate a <code>MOV</code>. SequentiallyConsistent fences generate an
268 <code>MFENCE</code>, other fences do not cause any code to be generated.
269 cmpxchg uses the <code>LOCK CMPXCHG</code> instruction.
270 <code>atomicrmw xchg</code> uses <code>XCHG</code>,
271 <code>atomicrmw add</code> and <code>atomicrmw sub</code> use
272 <code>XADD</code>, and all other <code>atomicrmw</code> operations generate
273 a loop with <code>LOCK CMPXCHG</code>. Depending on the users of the
274 result, some <code>atomicrmw</code> operations can be translated into
275 operations like <code>LOCK AND</code>, but that does not work in
278 <p>On ARM, MIPS, and many other RISC architectures, Acquire, Release, and
279 SequentiallyConsistent semantics require barrier instructions
280 for every such operation. Loads and stores generate normal instructions.
281 <code>atomicrmw</code> and <code>cmpxchg</code> generate LL/SC loops.</p>
285 <!-- *********************************************************************** -->
289 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
290 src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a>
291 <a href="http://validator.w3.org/check/referer"><img
292 src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a>
294 <a href="http://llvm.org/">LLVM Compiler Infrastructure</a><br>
295 Last modified: $Date: 2011-08-09 02:07:00 -0700 (Tue, 09 Aug 2011) $