1<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN" 2 "http://www.w3.org/TR/html4/strict.dtd"> 3<html> 4<head> 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"> 8</head> 9<body> 10 11<h1> 12 LLVM Atomic Instructions and Concurrency Guide 13</h1> 14 15<ol> 16 <li><a href="#introduction">Introduction</a></li> 17 <li><a href="#outsideatomic">Optimization outside atomic</a></li> 18 <li><a href="#atomicinst">Atomic instructions</a></li> 19 <li><a href="#ordering">Atomic orderings</a></li> 20 <li><a href="#iropt">Atomics and IR optimization</a></li> 21 <li><a href="#codegen">Atomics and Codegen</a></li> 22</ol> 23 24<div class="doc_author"> 25 <p>Written by Eli Friedman</p> 26</div> 27 28<!-- *********************************************************************** --> 29<h2> 30 <a name="introduction">Introduction</a> 31</h2> 32<!-- *********************************************************************** --> 33 34<div> 35 36<p>Historically, LLVM has not had very strong support for concurrency; some 37minimal intrinsics were provided, and <code>volatile</code> was used in some 38cases to achieve rough semantics in the presence of concurrency. However, this 39is changing; there are now new instructions which are well-defined in the 40presence of threads and asynchronous signals, and the model for existing 41instructions has been clarified in the IR.</p> 42 43<p>The atomic instructions are designed specifically to provide readable IR and 44 optimized code generation for the following:</p> 45<ul> 46 <li>The new C++0x <code><atomic></code> header. 47 (<a href="http://www.open-std.org/jtc1/sc22/wg21/">C++0x draft available here</a>.) 48 (<a href="http://www.open-std.org/jtc1/sc22/wg14/">C1x draft available here</a>)</li> 49 <li>Proper semantics for Java-style memory, for both <code>volatile</code> and 50 regular shared variables. 51 (<a href="http://java.sun.com/docs/books/jls/third_edition/html/memory.html">Java Specification</a>)</li> 52 <li>gcc-compatible <code>__sync_*</code> builtins. 53 (<a href="http://gcc.gnu.org/onlinedocs/gcc/Atomic-Builtins.html">Description</a>)</li> 54 <li>Other scenarios with atomic semantics, including <code>static</code> 55 variables with non-trivial constructors in C++.</li> 56</ul> 57 58<p>Atomic and volatile in the IR are orthogonal; "volatile" is the C/C++ 59 volatile, which ensures that every volatile load and store happens and is 60 performed in the stated order. A couple examples: if a 61 SequentiallyConsistent store is immediately followed by another 62 SequentiallyConsistent store to the same address, the first store can 63 be erased. This transformation is not allowed for a pair of volatile 64 stores. On the other hand, a non-volatile non-atomic load can be moved 65 across a volatile load freely, but not an Acquire load.</p> 66 67<p>This document is intended to provide a guide to anyone either writing a 68 frontend for LLVM or working on optimization passes for LLVM with a guide 69 for how to deal with instructions with special semantics in the presence of 70 concurrency. This is not intended to be a precise guide to the semantics; 71 the details can get extremely complicated and unreadable, and are not 72 usually necessary.</p> 73 74</div> 75 76<!-- *********************************************************************** --> 77<h2> 78 <a name="outsideatomic">Optimization outside atomic</a> 79</h2> 80<!-- *********************************************************************** --> 81 82<div> 83 84<p>The basic <code>'load'</code> and <code>'store'</code> allow a variety of 85 optimizations, but can lead to undefined results in a concurrent environment; 86 see <a href="#o_nonatomic">NonAtomic</a>. This section specifically goes 87 into the one optimizer restriction which applies in concurrent environments, 88 which gets a bit more of an extended description because any optimization 89 dealing with stores needs to be aware of it.</p> 90 91<p>From the optimizer's point of view, the rule is that if there 92 are not any instructions with atomic ordering involved, concurrency does 93 not matter, with one exception: if a variable might be visible to another 94 thread or signal handler, a store cannot be inserted along a path where it 95 might not execute otherwise. Take the following example:</p> 96 97<pre> 98/* C code, for readability; run through clang -O2 -S -emit-llvm to get 99 equivalent IR */ 100int x; 101void f(int* a) { 102 for (int i = 0; i < 100; i++) { 103 if (a[i]) 104 x += 1; 105 } 106} 107</pre> 108 109<p>The following is equivalent in non-concurrent situations:</p> 110 111<pre> 112int x; 113void f(int* a) { 114 int xtemp = x; 115 for (int i = 0; i < 100; i++) { 116 if (a[i]) 117 xtemp += 1; 118 } 119 x = xtemp; 120} 121</pre> 122 123<p>However, LLVM is not allowed to transform the former to the latter: it could 124 indirectly introduce undefined behavior if another thread can access x at 125 the same time. (This example is particularly of interest because before the 126 concurrency model was implemented, LLVM would perform this 127 transformation.)</p> 128 129<p>Note that speculative loads are allowed; a load which 130 is part of a race returns <code>undef</code>, but does not have undefined 131 behavior.</p> 132 133 134</div> 135 136<!-- *********************************************************************** --> 137<h2> 138 <a name="atomicinst">Atomic instructions</a> 139</h2> 140<!-- *********************************************************************** --> 141 142<div> 143 144<p>For cases where simple loads and stores are not sufficient, LLVM provides 145 various atomic instructions. The exact guarantees provided depend on the 146 ordering; see <a href="#ordering">Atomic orderings</a></p> 147 148<p><code>load atomic</code> and <code>store atomic</code> provide the same 149 basic functionality as non-atomic loads and stores, but provide additional 150 guarantees in situations where threads and signals are involved.</p> 151 152<p><code>cmpxchg</code> and <code>atomicrmw</code> are essentially like an 153 atomic load followed by an atomic store (where the store is conditional for 154 <code>cmpxchg</code>), but no other memory operation can happen on any thread 155 between the load and store. Note that LLVM's cmpxchg does not provide quite 156 as many options as the C++0x version.</p> 157 158<p>A <code>fence</code> provides Acquire and/or Release ordering which is not 159 part of another operation; it is normally used along with Monotonic memory 160 operations. A Monotonic load followed by an Acquire fence is roughly 161 equivalent to an Acquire load.</p> 162 163<p>Frontends generating atomic instructions generally need to be aware of the 164 target to some degree; atomic instructions are guaranteed to be lock-free, 165 and therefore an instruction which is wider than the target natively supports 166 can be impossible to generate.</p> 167 168</div> 169 170<!-- *********************************************************************** --> 171<h2> 172 <a name="ordering">Atomic orderings</a> 173</h2> 174<!-- *********************************************************************** --> 175 176<div> 177 178<p>In order to achieve a balance between performance and necessary guarantees, 179 there are six levels of atomicity. They are listed in order of strength; 180 each level includes all the guarantees of the previous level except for 181 Acquire/Release. (See also <a href="LangRef.html#ordering">LangRef</a>.)</p> 182 183<!-- ======================================================================= --> 184<h3> 185 <a name="o_notatomic">NotAtomic</a> 186</h3> 187 188<div> 189 190<p>NotAtomic is the obvious, a load or store which is not atomic. (This isn't 191 really a level of atomicity, but is listed here for comparison.) This is 192 essentially a regular load or store. If there is a race on a given memory 193 location, loads from that location return undef.</p> 194 195<dl> 196 <dt>Relevant standard</dt> 197 <dd>This is intended to match shared variables in C/C++, and to be used 198 in any other context where memory access is necessary, and 199 a race is impossible. (The precise definition is in 200 <a href="LangRef.html#memmodel">LangRef</a>.) 201 <dt>Notes for frontends</dt> 202 <dd>The rule is essentially that all memory accessed with basic loads and 203 stores by multiple threads should be protected by a lock or other 204 synchronization; otherwise, you are likely to run into undefined 205 behavior. If your frontend is for a "safe" language like Java, 206 use Unordered to load and store any shared variable. Note that NotAtomic 207 volatile loads and stores are not properly atomic; do not try to use 208 them as a substitute. (Per the C/C++ standards, volatile does provide 209 some limited guarantees around asynchronous signals, but atomics are 210 generally a better solution.) 211 <dt>Notes for optimizers</dt> 212 <dd>Introducing loads to shared variables along a codepath where they would 213 not otherwise exist is allowed; introducing stores to shared variables 214 is not. See <a href="#outsideatomic">Optimization outside 215 atomic</a>.</dd> 216 <dt>Notes for code generation</dt> 217 <dd>The one interesting restriction here is that it is not allowed to write 218 to bytes outside of the bytes relevant to a store. This is mostly 219 relevant to unaligned stores: it is not allowed in general to convert 220 an unaligned store into two aligned stores of the same width as the 221 unaligned store. Backends are also expected to generate an i8 store 222 as an i8 store, and not an instruction which writes to surrounding 223 bytes. (If you are writing a backend for an architecture which cannot 224 satisfy these restrictions and cares about concurrency, please send an 225 email to llvmdev.)</dd> 226</dl> 227 228</div> 229 230 231<!-- ======================================================================= --> 232<h3> 233 <a name="o_unordered">Unordered</a> 234</h3> 235 236<div> 237 238<p>Unordered is the lowest level of atomicity. It essentially guarantees that 239 races produce somewhat sane results instead of having undefined behavior. 240 It also guarantees the operation to be lock-free, so it do not depend on 241 the data being part of a special atomic structure or depend on a separate 242 per-process global lock. Note that code generation will fail for 243 unsupported atomic operations; if you need such an operation, use explicit 244 locking.</p> 245 246<dl> 247 <dt>Relevant standard</dt> 248 <dd>This is intended to match the Java memory model for shared 249 variables.</dd> 250 <dt>Notes for frontends</dt> 251 <dd>This cannot be used for synchronization, but is useful for Java and 252 other "safe" languages which need to guarantee that the generated 253 code never exhibits undefined behavior. Note that this guarantee 254 is cheap on common platforms for loads of a native width, but can 255 be expensive or unavailable for wider loads, like a 64-bit store 256 on ARM. (A frontend for Java or other "safe" languages would normally 257 split a 64-bit store on ARM into two 32-bit unordered stores.) 258 <dt>Notes for optimizers</dt> 259 <dd>In terms of the optimizer, this prohibits any transformation that 260 transforms a single load into multiple loads, transforms a store 261 into multiple stores, narrows a store, or stores a value which 262 would not be stored otherwise. Some examples of unsafe optimizations 263 are narrowing an assignment into a bitfield, rematerializing 264 a load, and turning loads and stores into a memcpy call. Reordering 265 unordered operations is safe, though, and optimizers should take 266 advantage of that because unordered operations are common in 267 languages that need them.</dd> 268 <dt>Notes for code generation</dt> 269 <dd>These operations are required to be atomic in the sense that if you 270 use unordered loads and unordered stores, a load cannot see a value 271 which was never stored. A normal load or store instruction is usually 272 sufficient, but note that an unordered load or store cannot 273 be split into multiple instructions (or an instruction which 274 does multiple memory operations, like <code>LDRD</code> on ARM).</dd> 275</dl> 276 277</div> 278 279<!-- ======================================================================= --> 280<h3> 281 <a name="o_monotonic">Monotonic</a> 282</h3> 283 284<div> 285 286<p>Monotonic is the weakest level of atomicity that can be used in 287 synchronization primitives, although it does not provide any general 288 synchronization. It essentially guarantees that if you take all the 289 operations affecting a specific address, a consistent ordering exists. 290 291<dl> 292 <dt>Relevant standard</dt> 293 <dd>This corresponds to the C++0x/C1x <code>memory_order_relaxed</code>; 294 see those standards for the exact definition. 295 <dt>Notes for frontends</dt> 296 <dd>If you are writing a frontend which uses this directly, use with caution. 297 The guarantees in terms of synchronization are very weak, so make 298 sure these are only used in a pattern which you know is correct. 299 Generally, these would either be used for atomic operations which 300 do not protect other memory (like an atomic counter), or along with 301 a <code>fence</code>.</dd> 302 <dt>Notes for optimizers</dt> 303 <dd>In terms of the optimizer, this can be treated as a read+write on the 304 relevant memory location (and alias analysis will take advantage of 305 that). In addition, it is legal to reorder non-atomic and Unordered 306 loads around Monotonic loads. CSE/DSE and a few other optimizations 307 are allowed, but Monotonic operations are unlikely to be used in ways 308 which would make those optimizations useful.</dd> 309 <dt>Notes for code generation</dt> 310 <dd>Code generation is essentially the same as that for unordered for loads 311 and stores. No fences are required. <code>cmpxchg</code> and 312 <code>atomicrmw</code> are required to appear as a single operation.</dd> 313</dl> 314 315</div> 316 317<!-- ======================================================================= --> 318<h3> 319 <a name="o_acquire">Acquire</a> 320</h3> 321 322<div> 323 324<p>Acquire provides a barrier of the sort necessary to acquire a lock to access 325 other memory with normal loads and stores. 326 327<dl> 328 <dt>Relevant standard</dt> 329 <dd>This corresponds to the C++0x/C1x <code>memory_order_acquire</code>. It 330 should also be used for C++0x/C1x <code>memory_order_consume</code>. 331 <dt>Notes for frontends</dt> 332 <dd>If you are writing a frontend which uses this directly, use with caution. 333 Acquire only provides a semantic guarantee when paired with a Release 334 operation.</dd> 335 <dt>Notes for optimizers</dt> 336 <dd>Optimizers not aware of atomics can treat this like a nothrow call. 337 It is also possible to move stores from before an Acquire load 338 or read-modify-write operation to after it, and move non-Acquire 339 loads from before an Acquire operation to after it.</dd> 340 <dt>Notes for code generation</dt> 341 <dd>Architectures with weak memory ordering (essentially everything relevant 342 today except x86 and SPARC) require some sort of fence to maintain 343 the Acquire semantics. The precise fences required varies widely by 344 architecture, but for a simple implementation, most architectures provide 345 a barrier which is strong enough for everything (<code>dmb</code> on ARM, 346 <code>sync</code> on PowerPC, etc.). Putting such a fence after the 347 equivalent Monotonic operation is sufficient to maintain Acquire 348 semantics for a memory operation.</dd> 349</dl> 350 351</div> 352 353<!-- ======================================================================= --> 354<h3> 355 <a name="o_acquire">Release</a> 356</h3> 357 358<div> 359 360<p>Release is similar to Acquire, but with a barrier of the sort necessary to 361 release a lock. 362 363<dl> 364 <dt>Relevant standard</dt> 365 <dd>This corresponds to the C++0x/C1x <code>memory_order_release</code>.</dd> 366 <dt>Notes for frontends</dt> 367 <dd>If you are writing a frontend which uses this directly, use with caution. 368 Release only provides a semantic guarantee when paired with a Acquire 369 operation.</dd> 370 <dt>Notes for optimizers</dt> 371 <dd>Optimizers not aware of atomics can treat this like a nothrow call. 372 It is also possible to move loads from after a Release store 373 or read-modify-write operation to before it, and move non-Release 374 stores from after an Release operation to before it.</dd> 375 <dt>Notes for code generation</dt> 376 <dd>See the section on Acquire; a fence before the relevant operation is 377 usually sufficient for Release. Note that a store-store fence is not 378 sufficient to implement Release semantics; store-store fences are 379 generally not exposed to IR because they are extremely difficult to 380 use correctly.</dd> 381</dl> 382 383</div> 384 385<!-- ======================================================================= --> 386<h3> 387 <a name="o_acqrel">AcquireRelease</a> 388</h3> 389 390<div> 391 392<p>AcquireRelease (<code>acq_rel</code> in IR) provides both an Acquire and a 393 Release barrier (for fences and operations which both read and write memory). 394 395<dl> 396 <dt>Relevant standard</dt> 397 <dd>This corresponds to the C++0x/C1x <code>memory_order_acq_rel</code>. 398 <dt>Notes for frontends</dt> 399 <dd>If you are writing a frontend which uses this directly, use with caution. 400 Acquire only provides a semantic guarantee when paired with a Release 401 operation, and vice versa.</dd> 402 <dt>Notes for optimizers</dt> 403 <dd>In general, optimizers should treat this like a nothrow call; the 404 the possible optimizations are usually not interesting.</dd> 405 <dt>Notes for code generation</dt> 406 <dd>This operation has Acquire and Release semantics; see the sections on 407 Acquire and Release.</dd> 408</dl> 409 410</div> 411 412<!-- ======================================================================= --> 413<h3> 414 <a name="o_seqcst">SequentiallyConsistent</a> 415</h3> 416 417<div> 418 419<p>SequentiallyConsistent (<code>seq_cst</code> in IR) provides 420 Acquire semantics for loads and Release semantics for 421 stores. Additionally, it guarantees that a total ordering exists 422 between all SequentiallyConsistent operations. 423 424<dl> 425 <dt>Relevant standard</dt> 426 <dd>This corresponds to the C++0x/C1x <code>memory_order_seq_cst</code>, 427 Java volatile, and the gcc-compatible <code>__sync_*</code> builtins 428 which do not specify otherwise. 429 <dt>Notes for frontends</dt> 430 <dd>If a frontend is exposing atomic operations, these are much easier to 431 reason about for the programmer than other kinds of operations, and using 432 them is generally a practical performance tradeoff.</dd> 433 <dt>Notes for optimizers</dt> 434 <dd>Optimizers not aware of atomics can treat this like a nothrow call. 435 For SequentiallyConsistent loads and stores, the same reorderings are 436 allowed as for Acquire loads and Release stores, except that 437 SequentiallyConsistent operations may not be reordered.</dd> 438 <dt>Notes for code generation</dt> 439 <dd>SequentiallyConsistent loads minimally require the same barriers 440 as Acquire operations and SequentiallyConsistent stores require 441 Release barriers. Additionally, the code generator must enforce 442 ordering between SequentiallyConsistent stores followed by 443 SequentiallyConsistent loads. This is usually done by emitting 444 either a full fence before the loads or a full fence after the 445 stores; which is preferred varies by architecture.</dd> 446</dl> 447 448</div> 449 450</div> 451 452<!-- *********************************************************************** --> 453<h2> 454 <a name="iropt">Atomics and IR optimization</a> 455</h2> 456<!-- *********************************************************************** --> 457 458<div> 459 460<p>Predicates for optimizer writers to query: 461<ul> 462 <li>isSimple(): A load or store which is not volatile or atomic. This is 463 what, for example, memcpyopt would check for operations it might 464 transform.</li> 465 <li>isUnordered(): A load or store which is not volatile and at most 466 Unordered. This would be checked, for example, by LICM before hoisting 467 an operation.</li> 468 <li>mayReadFromMemory()/mayWriteToMemory(): Existing predicate, but note 469 that they return true for any operation which is volatile or at least 470 Monotonic.</li> 471 <li>Alias analysis: Note that AA will return ModRef for anything Acquire or 472 Release, and for the address accessed by any Monotonic operation.</li> 473</ul> 474 475<p>To support optimizing around atomic operations, make sure you are using 476 the right predicates; everything should work if that is done. If your 477 pass should optimize some atomic operations (Unordered operations in 478 particular), make sure it doesn't replace an atomic load or store with 479 a non-atomic operation.</p> 480 481<p>Some examples of how optimizations interact with various kinds of atomic 482 operations: 483<ul> 484 <li>memcpyopt: An atomic operation cannot be optimized into part of a 485 memcpy/memset, including unordered loads/stores. It can pull operations 486 across some atomic operations. 487 <li>LICM: Unordered loads/stores can be moved out of a loop. It just treats 488 monotonic operations like a read+write to a memory location, and anything 489 stricter than that like a nothrow call. 490 <li>DSE: Unordered stores can be DSE'ed like normal stores. Monotonic stores 491 can be DSE'ed in some cases, but it's tricky to reason about, and not 492 especially important. 493 <li>Folding a load: Any atomic load from a constant global can be 494 constant-folded, because it cannot be observed. Similar reasoning allows 495 scalarrepl with atomic loads and stores. 496</ul> 497 498</div> 499 500<!-- *********************************************************************** --> 501<h2> 502 <a name="codegen">Atomics and Codegen</a> 503</h2> 504<!-- *********************************************************************** --> 505 506<div> 507 508<p>Atomic operations are represented in the SelectionDAG with 509 <code>ATOMIC_*</code> opcodes. On architectures which use barrier 510 instructions for all atomic ordering (like ARM), appropriate fences are 511 split out as the DAG is built.</p> 512 513<p>The MachineMemOperand for all atomic operations is currently marked as 514 volatile; this is not correct in the IR sense of volatile, but CodeGen 515 handles anything marked volatile very conservatively. This should get 516 fixed at some point.</p> 517 518<p>Common architectures have some way of representing at least a pointer-sized 519 lock-free <code>cmpxchg</code>; such an operation can be used to implement 520 all the other atomic operations which can be represented in IR up to that 521 size. Backends are expected to implement all those operations, but not 522 operations which cannot be implemented in a lock-free manner. It is 523 expected that backends will give an error when given an operation which 524 cannot be implemented. (The LLVM code generator is not very helpful here 525 at the moment, but hopefully that will change.)</p> 526 527<p>The implementation of atomics on LL/SC architectures (like ARM) is currently 528 a bit of a mess; there is a lot of copy-pasted code across targets, and 529 the representation is relatively unsuited to optimization (it would be nice 530 to be able to optimize loops involving cmpxchg etc.).</p> 531 532<p>On x86, all atomic loads generate a <code>MOV</code>. 533 SequentiallyConsistent stores generate an <code>XCHG</code>, other stores 534 generate a <code>MOV</code>. SequentiallyConsistent fences generate an 535 <code>MFENCE</code>, other fences do not cause any code to be generated. 536 cmpxchg uses the <code>LOCK CMPXCHG</code> instruction. 537 <code>atomicrmw xchg</code> uses <code>XCHG</code>, 538 <code>atomicrmw add</code> and <code>atomicrmw sub</code> use 539 <code>XADD</code>, and all other <code>atomicrmw</code> operations generate 540 a loop with <code>LOCK CMPXCHG</code>. Depending on the users of the 541 result, some <code>atomicrmw</code> operations can be translated into 542 operations like <code>LOCK AND</code>, but that does not work in 543 general.</p> 544 545<p>On ARM, MIPS, and many other RISC architectures, Acquire, Release, and 546 SequentiallyConsistent semantics require barrier instructions 547 for every such operation. Loads and stores generate normal instructions. 548 <code>cmpxchg</code> and <code>atomicrmw</code> can be represented using 549 a loop with LL/SC-style instructions which take some sort of exclusive 550 lock on a cache line (<code>LDREX</code> and <code>STREX</code> on 551 ARM, etc.). At the moment, the IR does not provide any way to represent a 552 weak <code>cmpxchg</code> which would not require a loop.</p> 553</div> 554 555<!-- *********************************************************************** --> 556 557<hr> 558<address> 559 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img 560 src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a> 561 <a href="http://validator.w3.org/check/referer"><img 562 src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a> 563 564 <a href="http://llvm.org/">LLVM Compiler Infrastructure</a><br> 565 Last modified: $Date: 2011-08-09 02:07:00 -0700 (Tue, 09 Aug 2011) $ 566</address> 567 568</body> 569</html> 570