/3rd_party/llvm/docs/Atomics.html

https://code.google.com/p/softart/ · HTML · 569 lines · 462 code · 87 blank · 20 comment · 0 complexity · 524b94a8b8c8277ecce225571f00359e MD5 · raw file

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