1 <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN"
2 "http://www.w3.org/TR/html4/strict.dtd">
5 <meta http-equiv="Content-Type" content="text/html; charset=utf-8">
6 <title>LLVM Alias Analysis Infrastructure</title>
7 <link rel="stylesheet" href="llvm.css" type="text/css">
12 LLVM Alias Analysis Infrastructure
16 <li><a href="#introduction">Introduction</a></li>
18 <li><a href="#overview"><tt>AliasAnalysis</tt> Class Overview</a>
20 <li><a href="#pointers">Representation of Pointers</a></li>
21 <li><a href="#alias">The <tt>alias</tt> method</a></li>
22 <li><a href="#ModRefInfo">The <tt>getModRefInfo</tt> methods</a></li>
23 <li><a href="#OtherItfs">Other useful <tt>AliasAnalysis</tt> methods</a></li>
27 <li><a href="#writingnew">Writing a new <tt>AliasAnalysis</tt> Implementation</a>
29 <li><a href="#passsubclasses">Different Pass styles</a></li>
30 <li><a href="#requiredcalls">Required initialization calls</a></li>
31 <li><a href="#interfaces">Interfaces which may be specified</a></li>
32 <li><a href="#chaining"><tt>AliasAnalysis</tt> chaining behavior</a></li>
33 <li><a href="#updating">Updating analysis results for transformations</a></li>
34 <li><a href="#implefficiency">Efficiency Issues</a></li>
35 <li><a href="#limitations">Limitations</a></li>
39 <li><a href="#using">Using alias analysis results</a>
41 <li><a href="#memdep">Using the <tt>MemoryDependenceAnalysis</tt> Pass</a></li>
42 <li><a href="#ast">Using the <tt>AliasSetTracker</tt> class</a></li>
43 <li><a href="#direct">Using the <tt>AliasAnalysis</tt> interface directly</a></li>
47 <li><a href="#exist">Existing alias analysis implementations and clients</a>
49 <li><a href="#impls">Available <tt>AliasAnalysis</tt> implementations</a></li>
50 <li><a href="#aliasanalysis-xforms">Alias analysis driven transformations</a></li>
51 <li><a href="#aliasanalysis-debug">Clients for debugging and evaluation of
52 implementations</a></li>
55 <li><a href="#memdep">Memory Dependence Analysis</a></li>
58 <div class="doc_author">
59 <p>Written by <a href="mailto:sabre@nondot.org">Chris Lattner</a></p>
62 <!-- *********************************************************************** -->
64 <a name="introduction">Introduction</a>
66 <!-- *********************************************************************** -->
70 <p>Alias Analysis (aka Pointer Analysis) is a class of techniques which attempt
71 to determine whether or not two pointers ever can point to the same object in
72 memory. There are many different algorithms for alias analysis and many
73 different ways of classifying them: flow-sensitive vs flow-insensitive,
74 context-sensitive vs context-insensitive, field-sensitive vs field-insensitive,
75 unification-based vs subset-based, etc. Traditionally, alias analyses respond
76 to a query with a <a href="#MustMayNo">Must, May, or No</a> alias response,
77 indicating that two pointers always point to the same object, might point to the
78 same object, or are known to never point to the same object.</p>
81 href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html"><tt>AliasAnalysis</tt></a>
82 class is the primary interface used by clients and implementations of alias
83 analyses in the LLVM system. This class is the common interface between clients
84 of alias analysis information and the implementations providing it, and is
85 designed to support a wide range of implementations and clients (but currently
86 all clients are assumed to be flow-insensitive). In addition to simple alias
87 analysis information, this class exposes Mod/Ref information from those
88 implementations which can provide it, allowing for powerful analyses and
89 transformations to work well together.</p>
91 <p>This document contains information necessary to successfully implement this
92 interface, use it, and to test both sides. It also explains some of the finer
93 points about what exactly results mean. If you feel that something is unclear
94 or should be added, please <a href="mailto:sabre@nondot.org">let me
99 <!-- *********************************************************************** -->
101 <a name="overview"><tt>AliasAnalysis</tt> Class Overview</a>
103 <!-- *********************************************************************** -->
108 href="http://llvm.org/doxygen/classllvm_1_1AliasAnalysis.html"><tt>AliasAnalysis</tt></a>
109 class defines the interface that the various alias analysis implementations
110 should support. This class exports two important enums: <tt>AliasResult</tt>
111 and <tt>ModRefResult</tt> which represent the result of an alias query or a
112 mod/ref query, respectively.</p>
114 <p>The <tt>AliasAnalysis</tt> interface exposes information about memory,
115 represented in several different ways. In particular, memory objects are
116 represented as a starting address and size, and function calls are represented
117 as the actual <tt>call</tt> or <tt>invoke</tt> instructions that performs the
118 call. The <tt>AliasAnalysis</tt> interface also exposes some helper methods
119 which allow you to get mod/ref information for arbitrary instructions.</p>
121 <p>All <tt>AliasAnalysis</tt> interfaces require that in queries involving
122 multiple values, values which are not
123 <a href="LangRef.html#constants">constants</a> are all defined within the
126 <!-- ======================================================================= -->
128 <a name="pointers">Representation of Pointers</a>
133 <p>Most importantly, the <tt>AliasAnalysis</tt> class provides several methods
134 which are used to query whether or not two memory objects alias, whether
135 function calls can modify or read a memory object, etc. For all of these
136 queries, memory objects are represented as a pair of their starting address (a
137 symbolic LLVM <tt>Value*</tt>) and a static size.</p>
139 <p>Representing memory objects as a starting address and a size is critically
140 important for correct Alias Analyses. For example, consider this (silly, but
141 possible) C code:</p>
143 <div class="doc_code">
149 for (i = 0; i != 10; ++i) {
150 C[0] = A[i]; /* One byte store */
151 C[1] = A[9-i]; /* One byte store */
156 <p>In this case, the <tt>basicaa</tt> pass will disambiguate the stores to
157 <tt>C[0]</tt> and <tt>C[1]</tt> because they are accesses to two distinct
158 locations one byte apart, and the accesses are each one byte. In this case, the
159 LICM pass can use store motion to remove the stores from the loop. In
160 constrast, the following code:</p>
162 <div class="doc_code">
168 for (i = 0; i != 10; ++i) {
169 ((short*)C)[0] = A[i]; /* Two byte store! */
170 C[1] = A[9-i]; /* One byte store */
175 <p>In this case, the two stores to C do alias each other, because the access to
176 the <tt>&C[0]</tt> element is a two byte access. If size information wasn't
177 available in the query, even the first case would have to conservatively assume
178 that the accesses alias.</p>
182 <!-- ======================================================================= -->
184 <a name="alias">The <tt>alias</tt> method</a>
188 <p>The <tt>alias</tt> method is the primary interface used to determine whether
189 or not two memory objects alias each other. It takes two memory objects as
190 input and returns MustAlias, PartialAlias, MayAlias, or NoAlias as
193 <p>Like all <tt>AliasAnalysis</tt> interfaces, the <tt>alias</tt> method requires
194 that either the two pointer values be defined within the same function, or at
195 least one of the values is a <a href="LangRef.html#constants">constant</a>.</p>
197 <!-- _______________________________________________________________________ -->
199 <a name="MustMayNo">Must, May, and No Alias Responses</a>
203 <p>The NoAlias response may be used when there is never an immediate dependence
204 between any memory reference <i>based</i> on one pointer and any memory
205 reference <i>based</i> the other. The most obvious example is when the two
206 pointers point to non-overlapping memory ranges. Another is when the two
207 pointers are only ever used for reading memory. Another is when the memory is
208 freed and reallocated between accesses through one pointer and accesses through
209 the other -- in this case, there is a dependence, but it's mediated by the free
210 and reallocation.</p>
212 <p>As an exception to this is with the
213 <a href="LangRef.html#noalias"><tt>noalias</tt></a> keyword; the "irrelevant"
214 dependencies are ignored.</p>
216 <p>The MayAlias response is used whenever the two pointers might refer to the
219 <p>The PartialAlias response is used when the two memory objects are known
220 to be overlapping in some way, but do not start at the same address.</p>
222 <p>The MustAlias response may only be returned if the two memory objects are
223 guaranteed to always start at exactly the same location. A MustAlias response
224 implies that the pointers compare equal.</p>
230 <!-- ======================================================================= -->
232 <a name="ModRefInfo">The <tt>getModRefInfo</tt> methods</a>
237 <p>The <tt>getModRefInfo</tt> methods return information about whether the
238 execution of an instruction can read or modify a memory location. Mod/Ref
239 information is always conservative: if an instruction <b>might</b> read or write
240 a location, ModRef is returned.</p>
242 <p>The <tt>AliasAnalysis</tt> class also provides a <tt>getModRefInfo</tt>
243 method for testing dependencies between function calls. This method takes two
244 call sites (CS1 & CS2), returns NoModRef if neither call writes to memory
245 read or written by the other, Ref if CS1 reads memory written by CS2, Mod if CS1
246 writes to memory read or written by CS2, or ModRef if CS1 might read or write
247 memory written to by CS2. Note that this relation is not commutative.</p>
252 <!-- ======================================================================= -->
254 <a name="OtherItfs">Other useful <tt>AliasAnalysis</tt> methods</a>
260 Several other tidbits of information are often collected by various alias
261 analysis implementations and can be put to good use by various clients.
264 <!-- _______________________________________________________________________ -->
266 The <tt>pointsToConstantMemory</tt> method
271 <p>The <tt>pointsToConstantMemory</tt> method returns true if and only if the
272 analysis can prove that the pointer only points to unchanging memory locations
273 (functions, constant global variables, and the null pointer). This information
274 can be used to refine mod/ref information: it is impossible for an unchanging
275 memory location to be modified.</p>
279 <!-- _______________________________________________________________________ -->
281 <a name="simplemodref">The <tt>doesNotAccessMemory</tt> and
282 <tt>onlyReadsMemory</tt> methods</a>
287 <p>These methods are used to provide very simple mod/ref information for
288 function calls. The <tt>doesNotAccessMemory</tt> method returns true for a
289 function if the analysis can prove that the function never reads or writes to
290 memory, or if the function only reads from constant memory. Functions with this
291 property are side-effect free and only depend on their input arguments, allowing
292 them to be eliminated if they form common subexpressions or be hoisted out of
293 loops. Many common functions behave this way (e.g., <tt>sin</tt> and
294 <tt>cos</tt>) but many others do not (e.g., <tt>acos</tt>, which modifies the
295 <tt>errno</tt> variable).</p>
297 <p>The <tt>onlyReadsMemory</tt> method returns true for a function if analysis
298 can prove that (at most) the function only reads from non-volatile memory.
299 Functions with this property are side-effect free, only depending on their input
300 arguments and the state of memory when they are called. This property allows
301 calls to these functions to be eliminated and moved around, as long as there is
302 no store instruction that changes the contents of memory. Note that all
303 functions that satisfy the <tt>doesNotAccessMemory</tt> method also satisfies
304 <tt>onlyReadsMemory</tt>.</p>
312 <!-- *********************************************************************** -->
314 <a name="writingnew">Writing a new <tt>AliasAnalysis</tt> Implementation</a>
316 <!-- *********************************************************************** -->
320 <p>Writing a new alias analysis implementation for LLVM is quite
321 straight-forward. There are already several implementations that you can use
322 for examples, and the following information should help fill in any details.
323 For a examples, take a look at the <a href="#impls">various alias analysis
324 implementations</a> included with LLVM.</p>
326 <!-- ======================================================================= -->
328 <a name="passsubclasses">Different Pass styles</a>
333 <p>The first step to determining what type of <a
334 href="WritingAnLLVMPass.html">LLVM pass</a> you need to use for your Alias
335 Analysis. As is the case with most other analyses and transformations, the
336 answer should be fairly obvious from what type of problem you are trying to
340 <li>If you require interprocedural analysis, it should be a
342 <li>If you are a function-local analysis, subclass <tt>FunctionPass</tt>.</li>
343 <li>If you don't need to look at the program at all, subclass
344 <tt>ImmutablePass</tt>.</li>
347 <p>In addition to the pass that you subclass, you should also inherit from the
348 <tt>AliasAnalysis</tt> interface, of course, and use the
349 <tt>RegisterAnalysisGroup</tt> template to register as an implementation of
350 <tt>AliasAnalysis</tt>.</p>
354 <!-- ======================================================================= -->
356 <a name="requiredcalls">Required initialization calls</a>
361 <p>Your subclass of <tt>AliasAnalysis</tt> is required to invoke two methods on
362 the <tt>AliasAnalysis</tt> base class: <tt>getAnalysisUsage</tt> and
363 <tt>InitializeAliasAnalysis</tt>. In particular, your implementation of
364 <tt>getAnalysisUsage</tt> should explicitly call into the
365 <tt>AliasAnalysis::getAnalysisUsage</tt> method in addition to doing any
366 declaring any pass dependencies your pass has. Thus you should have something
369 <div class="doc_code">
371 void getAnalysisUsage(AnalysisUsage &AU) const {
372 AliasAnalysis::getAnalysisUsage(AU);
373 <i>// declare your dependencies here.</i>
378 <p>Additionally, your must invoke the <tt>InitializeAliasAnalysis</tt> method
379 from your analysis run method (<tt>run</tt> for a <tt>Pass</tt>,
380 <tt>runOnFunction</tt> for a <tt>FunctionPass</tt>, or <tt>InitializePass</tt>
381 for an <tt>ImmutablePass</tt>). For example (as part of a <tt>Pass</tt>):</p>
383 <div class="doc_code">
385 bool run(Module &M) {
386 InitializeAliasAnalysis(this);
387 <i>// Perform analysis here...</i>
395 <!-- ======================================================================= -->
397 <a name="interfaces">Interfaces which may be specified</a>
403 href="/doxygen/classllvm_1_1AliasAnalysis.html"><tt>AliasAnalysis</tt></a>
404 virtual methods default to providing <a href="#chaining">chaining</a> to another
405 alias analysis implementation, which ends up returning conservatively correct
406 information (returning "May" Alias and "Mod/Ref" for alias and mod/ref queries
407 respectively). Depending on the capabilities of the analysis you are
408 implementing, you just override the interfaces you can improve.</p>
414 <!-- ======================================================================= -->
416 <a name="chaining"><tt>AliasAnalysis</tt> chaining behavior</a>
421 <p>With only two special exceptions (the <tt><a
422 href="#basic-aa">basicaa</a></tt> and <a href="#no-aa"><tt>no-aa</tt></a>
423 passes) every alias analysis pass chains to another alias analysis
424 implementation (for example, the user can specify "<tt>-basicaa -ds-aa
425 -licm</tt>" to get the maximum benefit from both alias
426 analyses). The alias analysis class automatically takes care of most of this
427 for methods that you don't override. For methods that you do override, in code
428 paths that return a conservative MayAlias or Mod/Ref result, simply return
429 whatever the superclass computes. For example:</p>
431 <div class="doc_code">
433 AliasAnalysis::AliasResult alias(const Value *V1, unsigned V1Size,
434 const Value *V2, unsigned V2Size) {
439 <i>// Couldn't determine a must or no-alias result.</i>
440 return AliasAnalysis::alias(V1, V1Size, V2, V2Size);
445 <p>In addition to analysis queries, you must make sure to unconditionally pass
446 LLVM <a href="#updating">update notification</a> methods to the superclass as
447 well if you override them, which allows all alias analyses in a change to be
453 <!-- ======================================================================= -->
455 <a name="updating">Updating analysis results for transformations</a>
460 Alias analysis information is initially computed for a static snapshot of the
461 program, but clients will use this information to make transformations to the
462 code. All but the most trivial forms of alias analysis will need to have their
463 analysis results updated to reflect the changes made by these transformations.
467 The <tt>AliasAnalysis</tt> interface exposes four methods which are used to
468 communicate program changes from the clients to the analysis implementations.
469 Various alias analysis implementations should use these methods to ensure that
470 their internal data structures are kept up-to-date as the program changes (for
471 example, when an instruction is deleted), and clients of alias analysis must be
472 sure to call these interfaces appropriately.
475 <!-- _______________________________________________________________________ -->
476 <h4>The <tt>deleteValue</tt> method</h4>
479 The <tt>deleteValue</tt> method is called by transformations when they remove an
480 instruction or any other value from the program (including values that do not
481 use pointers). Typically alias analyses keep data structures that have entries
482 for each value in the program. When this method is called, they should remove
483 any entries for the specified value, if they exist.
486 <!-- _______________________________________________________________________ -->
487 <h4>The <tt>copyValue</tt> method</h4>
490 The <tt>copyValue</tt> method is used when a new value is introduced into the
491 program. There is no way to introduce a value into the program that did not
492 exist before (this doesn't make sense for a safe compiler transformation), so
493 this is the only way to introduce a new value. This method indicates that the
494 new value has exactly the same properties as the value being copied.
497 <!-- _______________________________________________________________________ -->
498 <h4>The <tt>replaceWithNewValue</tt> method</h4>
501 This method is a simple helper method that is provided to make clients easier to
502 use. It is implemented by copying the old analysis information to the new
503 value, then deleting the old value. This method cannot be overridden by alias
504 analysis implementations.
507 <!-- _______________________________________________________________________ -->
508 <h4>The <tt>addEscapingUse</tt> method</h4>
511 <p>The <tt>addEscapingUse</tt> method is used when the uses of a pointer
512 value have changed in ways that may invalidate precomputed analysis information.
513 Implementations may either use this callback to provide conservative responses
514 for points whose uses have change since analysis time, or may recompute some
515 or all of their internal state to continue providing accurate responses.</p>
517 <p>In general, any new use of a pointer value is considered an escaping use,
518 and must be reported through this callback, <em>except</em> for the
522 <li>A <tt>bitcast</tt> or <tt>getelementptr</tt> of the pointer</li>
523 <li>A <tt>store</tt> through the pointer (but not a <tt>store</tt>
524 <em>of</em> the pointer)</li>
525 <li>A <tt>load</tt> through the pointer</li>
531 <!-- ======================================================================= -->
533 <a name="implefficiency">Efficiency Issues</a>
538 <p>From the LLVM perspective, the only thing you need to do to provide an
539 efficient alias analysis is to make sure that alias analysis <b>queries</b> are
540 serviced quickly. The actual calculation of the alias analysis results (the
541 "run" method) is only performed once, but many (perhaps duplicate) queries may
542 be performed. Because of this, try to move as much computation to the run
543 method as possible (within reason).</p>
547 <!-- ======================================================================= -->
549 <a name="limitations">Limitations</a>
554 <p>The AliasAnalysis infrastructure has several limitations which make
555 writing a new <tt>AliasAnalysis</tt> implementation difficult.</p>
557 <p>There is no way to override the default alias analysis. It would
558 be very useful to be able to do something like "opt -my-aa -O2" and
559 have it use -my-aa for all passes which need AliasAnalysis, but there
560 is currently no support for that, short of changing the source code
561 and recompiling. Similarly, there is also no way of setting a chain
562 of analyses as the default.</p>
564 <p>There is no way for transform passes to declare that they preserve
565 <tt>AliasAnalysis</tt> implementations. The <tt>AliasAnalysis</tt>
566 interface includes <tt>deleteValue</tt> and <tt>copyValue</tt> methods
567 which are intended to allow a pass to keep an AliasAnalysis consistent,
568 however there's no way for a pass to declare in its
569 <tt>getAnalysisUsage</tt> that it does so. Some passes attempt to use
570 <tt>AU.addPreserved<AliasAnalysis></tt>, however this doesn't
571 actually have any effect.</p>
573 <p><tt>AliasAnalysisCounter</tt> (<tt>-count-aa</tt>) and <tt>AliasDebugger</tt>
574 (<tt>-debug-aa</tt>) are implemented as <tt>ModulePass</tt> classes, so if your
575 alias analysis uses <tt>FunctionPass</tt>, it won't be able to use
576 these utilities. If you try to use them, the pass manager will
577 silently route alias analysis queries directly to
578 <tt>BasicAliasAnalysis</tt> instead.</p>
580 <p>Similarly, the <tt>opt -p</tt> option introduces <tt>ModulePass</tt>
581 passes between each pass, which prevents the use of <tt>FunctionPass</tt>
582 alias analysis passes.</p>
584 <p>The <tt>AliasAnalysis</tt> API does have functions for notifying
585 implementations when values are deleted or copied, however these
586 aren't sufficient. There are many other ways that LLVM IR can be
587 modified which could be relevant to <tt>AliasAnalysis</tt>
588 implementations which can not be expressed.</p>
590 <p>The <tt>AliasAnalysisDebugger</tt> utility seems to suggest that
591 <tt>AliasAnalysis</tt> implementations can expect that they will be
592 informed of any relevant <tt>Value</tt> before it appears in an
593 alias query. However, popular clients such as <tt>GVN</tt> don't
594 support this, and are known to trigger errors when run with the
595 <tt>AliasAnalysisDebugger</tt>.</p>
597 <p>Due to several of the above limitations, the most obvious use for
598 the <tt>AliasAnalysisCounter</tt> utility, collecting stats on all
599 alias queries in a compilation, doesn't work, even if the
600 <tt>AliasAnalysis</tt> implementations don't use <tt>FunctionPass</tt>.
601 There's no way to set a default, much less a default sequence,
602 and there's no way to preserve it.</p>
604 <p>The <tt>AliasSetTracker</tt> class (which is used by <tt>LICM</tt>
605 makes a non-deterministic number of alias queries. This can cause stats
606 collected by <tt>AliasAnalysisCounter</tt> to have fluctuations among
607 identical runs, for example. Another consequence is that debugging
608 techniques involving pausing execution after a predetermined number
609 of queries can be unreliable.</p>
611 <p>Many alias queries can be reformulated in terms of other alias
612 queries. When multiple <tt>AliasAnalysis</tt> queries are chained together,
613 it would make sense to start those queries from the beginning of the chain,
614 with care taken to avoid infinite looping, however currently an
615 implementation which wants to do this can only start such queries
622 <!-- *********************************************************************** -->
624 <a name="using">Using alias analysis results</a>
626 <!-- *********************************************************************** -->
630 <p>There are several different ways to use alias analysis results. In order of
631 preference, these are...</p>
633 <!-- ======================================================================= -->
635 <a name="memdep">Using the <tt>MemoryDependenceAnalysis</tt> Pass</a>
640 <p>The <tt>memdep</tt> pass uses alias analysis to provide high-level dependence
641 information about memory-using instructions. This will tell you which store
642 feeds into a load, for example. It uses caching and other techniques to be
643 efficient, and is used by Dead Store Elimination, GVN, and memcpy optimizations.
648 <!-- ======================================================================= -->
650 <a name="ast">Using the <tt>AliasSetTracker</tt> class</a>
655 <p>Many transformations need information about alias <b>sets</b> that are active
656 in some scope, rather than information about pairwise aliasing. The <tt><a
657 href="/doxygen/classllvm_1_1AliasSetTracker.html">AliasSetTracker</a></tt> class
658 is used to efficiently build these Alias Sets from the pairwise alias analysis
659 information provided by the <tt>AliasAnalysis</tt> interface.</p>
661 <p>First you initialize the AliasSetTracker by using the "<tt>add</tt>" methods
662 to add information about various potentially aliasing instructions in the scope
663 you are interested in. Once all of the alias sets are completed, your pass
664 should simply iterate through the constructed alias sets, using the
665 <tt>AliasSetTracker</tt> <tt>begin()</tt>/<tt>end()</tt> methods.</p>
667 <p>The <tt>AliasSet</tt>s formed by the <tt>AliasSetTracker</tt> are guaranteed
668 to be disjoint, calculate mod/ref information and volatility for the set, and
669 keep track of whether or not all of the pointers in the set are Must aliases.
670 The AliasSetTracker also makes sure that sets are properly folded due to call
671 instructions, and can provide a list of pointers in each set.</p>
673 <p>As an example user of this, the <a href="/doxygen/structLICM.html">Loop
674 Invariant Code Motion</a> pass uses <tt>AliasSetTracker</tt>s to calculate alias
675 sets for each loop nest. If an <tt>AliasSet</tt> in a loop is not modified,
676 then all load instructions from that set may be hoisted out of the loop. If any
677 alias sets are stored to <b>and</b> are must alias sets, then the stores may be
678 sunk to outside of the loop, promoting the memory location to a register for the
679 duration of the loop nest. Both of these transformations only apply if the
680 pointer argument is loop-invariant.</p>
682 <!-- _______________________________________________________________________ -->
684 The AliasSetTracker implementation
689 <p>The AliasSetTracker class is implemented to be as efficient as possible. It
690 uses the union-find algorithm to efficiently merge AliasSets when a pointer is
691 inserted into the AliasSetTracker that aliases multiple sets. The primary data
692 structure is a hash table mapping pointers to the AliasSet they are in.</p>
694 <p>The AliasSetTracker class must maintain a list of all of the LLVM Value*'s
695 that are in each AliasSet. Since the hash table already has entries for each
696 LLVM Value* of interest, the AliasesSets thread the linked list through these
697 hash-table nodes to avoid having to allocate memory unnecessarily, and to make
698 merging alias sets extremely efficient (the linked list merge is constant time).
701 <p>You shouldn't need to understand these details if you are just a client of
702 the AliasSetTracker, but if you look at the code, hopefully this brief
703 description will help make sense of why things are designed the way they
710 <!-- ======================================================================= -->
712 <a name="direct">Using the <tt>AliasAnalysis</tt> interface directly</a>
717 <p>If neither of these utility class are what your pass needs, you should use
718 the interfaces exposed by the <tt>AliasAnalysis</tt> class directly. Try to use
719 the higher-level methods when possible (e.g., use mod/ref information instead of
720 the <a href="#alias"><tt>alias</tt></a> method directly if possible) to get the
721 best precision and efficiency.</p>
727 <!-- *********************************************************************** -->
729 <a name="exist">Existing alias analysis implementations and clients</a>
731 <!-- *********************************************************************** -->
735 <p>If you're going to be working with the LLVM alias analysis infrastructure,
736 you should know what clients and implementations of alias analysis are
737 available. In particular, if you are implementing an alias analysis, you should
738 be aware of the <a href="#aliasanalysis-debug">the clients</a> that are useful
739 for monitoring and evaluating different implementations.</p>
741 <!-- ======================================================================= -->
743 <a name="impls">Available <tt>AliasAnalysis</tt> implementations</a>
748 <p>This section lists the various implementations of the <tt>AliasAnalysis</tt>
749 interface. With the exception of the <a href="#no-aa"><tt>-no-aa</tt></a>
750 implementation, all of these <a href="#chaining">chain</a> to other alias
751 analysis implementations.</p>
753 <!-- _______________________________________________________________________ -->
755 <a name="no-aa">The <tt>-no-aa</tt> pass</a>
760 <p>The <tt>-no-aa</tt> pass is just like what it sounds: an alias analysis that
761 never returns any useful information. This pass can be useful if you think that
762 alias analysis is doing something wrong and are trying to narrow down a
767 <!-- _______________________________________________________________________ -->
769 <a name="basic-aa">The <tt>-basicaa</tt> pass</a>
774 <p>The <tt>-basicaa</tt> pass is an aggressive local analysis that "knows"
775 many important facts:</p>
778 <li>Distinct globals, stack allocations, and heap allocations can never
780 <li>Globals, stack allocations, and heap allocations never alias the null
782 <li>Different fields of a structure do not alias.</li>
783 <li>Indexes into arrays with statically differing subscripts cannot alias.</li>
784 <li>Many common standard C library functions <a
785 href="#simplemodref">never access memory or only read memory</a>.</li>
786 <li>Pointers that obviously point to constant globals
787 "<tt>pointToConstantMemory</tt>".</li>
788 <li>Function calls can not modify or references stack allocations if they never
789 escape from the function that allocates them (a common case for automatic
795 <!-- _______________________________________________________________________ -->
797 <a name="globalsmodref">The <tt>-globalsmodref-aa</tt> pass</a>
802 <p>This pass implements a simple context-sensitive mod/ref and alias analysis
803 for internal global variables that don't "have their address taken". If a
804 global does not have its address taken, the pass knows that no pointers alias
805 the global. This pass also keeps track of functions that it knows never access
806 memory or never read memory. This allows certain optimizations (e.g. GVN) to
807 eliminate call instructions entirely.
810 <p>The real power of this pass is that it provides context-sensitive mod/ref
811 information for call instructions. This allows the optimizer to know that
812 calls to a function do not clobber or read the value of the global, allowing
813 loads and stores to be eliminated.</p>
815 <p>Note that this pass is somewhat limited in its scope (only support
816 non-address taken globals), but is very quick analysis.</p>
819 <!-- _______________________________________________________________________ -->
821 <a name="steens-aa">The <tt>-steens-aa</tt> pass</a>
826 <p>The <tt>-steens-aa</tt> pass implements a variation on the well-known
827 "Steensgaard's algorithm" for interprocedural alias analysis. Steensgaard's
828 algorithm is a unification-based, flow-insensitive, context-insensitive, and
829 field-insensitive alias analysis that is also very scalable (effectively linear
832 <p>The LLVM <tt>-steens-aa</tt> pass implements a "speculatively
833 field-<b>sensitive</b>" version of Steensgaard's algorithm using the Data
834 Structure Analysis framework. This gives it substantially more precision than
835 the standard algorithm while maintaining excellent analysis scalability.</p>
837 <p>Note that <tt>-steens-aa</tt> is available in the optional "poolalloc"
838 module, it is not part of the LLVM core.</p>
842 <!-- _______________________________________________________________________ -->
844 <a name="ds-aa">The <tt>-ds-aa</tt> pass</a>
849 <p>The <tt>-ds-aa</tt> pass implements the full Data Structure Analysis
850 algorithm. Data Structure Analysis is a modular unification-based,
851 flow-insensitive, context-<b>sensitive</b>, and speculatively
852 field-<b>sensitive</b> alias analysis that is also quite scalable, usually at
855 <p>This algorithm is capable of responding to a full variety of alias analysis
856 queries, and can provide context-sensitive mod/ref information as well. The
857 only major facility not implemented so far is support for must-alias
860 <p>Note that <tt>-ds-aa</tt> is available in the optional "poolalloc"
861 module, it is not part of the LLVM core.</p>
865 <!-- _______________________________________________________________________ -->
867 <a name="scev-aa">The <tt>-scev-aa</tt> pass</a>
872 <p>The <tt>-scev-aa</tt> pass implements AliasAnalysis queries by
873 translating them into ScalarEvolution queries. This gives it a
874 more complete understanding of <tt>getelementptr</tt> instructions
875 and loop induction variables than other alias analyses have.</p>
881 <!-- ======================================================================= -->
883 <a name="aliasanalysis-xforms">Alias analysis driven transformations</a>
887 LLVM includes several alias-analysis driven transformations which can be used
888 with any of the implementations above.
890 <!-- _______________________________________________________________________ -->
892 <a name="adce">The <tt>-adce</tt> pass</a>
897 <p>The <tt>-adce</tt> pass, which implements Aggressive Dead Code Elimination
898 uses the <tt>AliasAnalysis</tt> interface to delete calls to functions that do
899 not have side-effects and are not used.</p>
904 <!-- _______________________________________________________________________ -->
906 <a name="licm">The <tt>-licm</tt> pass</a>
911 <p>The <tt>-licm</tt> pass implements various Loop Invariant Code Motion related
912 transformations. It uses the <tt>AliasAnalysis</tt> interface for several
913 different transformations:</p>
916 <li>It uses mod/ref information to hoist or sink load instructions out of loops
917 if there are no instructions in the loop that modifies the memory loaded.</li>
919 <li>It uses mod/ref information to hoist function calls out of loops that do not
920 write to memory and are loop-invariant.</li>
922 <li>If uses alias information to promote memory objects that are loaded and
923 stored to in loops to live in a register instead. It can do this if there are
924 no may aliases to the loaded/stored memory location.</li>
929 <!-- _______________________________________________________________________ -->
931 <a name="argpromotion">The <tt>-argpromotion</tt> pass</a>
936 The <tt>-argpromotion</tt> pass promotes by-reference arguments to be passed in
937 by-value instead. In particular, if pointer arguments are only loaded from it
938 passes in the value loaded instead of the address to the function. This pass
939 uses alias information to make sure that the value loaded from the argument
940 pointer is not modified between the entry of the function and any load of the
944 <!-- _______________________________________________________________________ -->
946 <a name="gvn">The <tt>-gvn</tt>, <tt>-memcpyopt</tt>, and <tt>-dse</tt>
952 <p>These passes use AliasAnalysis information to reason about loads and stores.
959 <!-- ======================================================================= -->
961 <a name="aliasanalysis-debug">Clients for debugging and evaluation of
967 <p>These passes are useful for evaluating the various alias analysis
968 implementations. You can use them with commands like '<tt>opt -ds-aa
969 -aa-eval foo.bc -disable-output -stats</tt>'.</p>
971 <!-- _______________________________________________________________________ -->
973 <a name="print-alias-sets">The <tt>-print-alias-sets</tt> pass</a>
978 <p>The <tt>-print-alias-sets</tt> pass is exposed as part of the
979 <tt>opt</tt> tool to print out the Alias Sets formed by the <a
980 href="#ast"><tt>AliasSetTracker</tt></a> class. This is useful if you're using
981 the <tt>AliasSetTracker</tt> class. To use it, use something like:</p>
983 <div class="doc_code">
985 % opt -ds-aa -print-alias-sets -disable-output
992 <!-- _______________________________________________________________________ -->
994 <a name="count-aa">The <tt>-count-aa</tt> pass</a>
999 <p>The <tt>-count-aa</tt> pass is useful to see how many queries a particular
1000 pass is making and what responses are returned by the alias analysis. As an
1003 <div class="doc_code">
1005 % opt -basicaa -count-aa -ds-aa -count-aa -licm
1009 <p>will print out how many queries (and what responses are returned) by the
1010 <tt>-licm</tt> pass (of the <tt>-ds-aa</tt> pass) and how many queries are made
1011 of the <tt>-basicaa</tt> pass by the <tt>-ds-aa</tt> pass. This can be useful
1012 when debugging a transformation or an alias analysis implementation.</p>
1016 <!-- _______________________________________________________________________ -->
1018 <a name="aa-eval">The <tt>-aa-eval</tt> pass</a>
1023 <p>The <tt>-aa-eval</tt> pass simply iterates through all pairs of pointers in a
1024 function and asks an alias analysis whether or not the pointers alias. This
1025 gives an indication of the precision of the alias analysis. Statistics are
1026 printed indicating the percent of no/may/must aliases found (a more precise
1027 algorithm will have a lower number of may aliases).</p>
1035 <!-- *********************************************************************** -->
1037 <a name="memdep">Memory Dependence Analysis</a>
1039 <!-- *********************************************************************** -->
1043 <p>If you're just looking to be a client of alias analysis information, consider
1044 using the Memory Dependence Analysis interface instead. MemDep is a lazy,
1045 caching layer on top of alias analysis that is able to answer the question of
1046 what preceding memory operations a given instruction depends on, either at an
1047 intra- or inter-block level. Because of its laziness and caching
1048 policy, using MemDep can be a significant performance win over accessing alias
1049 analysis directly.</p>
1053 <!-- *********************************************************************** -->
1057 <a href="http://jigsaw.w3.org/css-validator/check/referer"><img
1058 src="http://jigsaw.w3.org/css-validator/images/vcss-blue" alt="Valid CSS"></a>
1059 <a href="http://validator.w3.org/check/referer"><img
1060 src="http://www.w3.org/Icons/valid-html401-blue" alt="Valid HTML 4.01"></a>
1062 <a href="mailto:sabre@nondot.org">Chris Lattner</a><br>
1063 <a href="http://llvm.org/">LLVM Compiler Infrastructure</a><br>
1064 Last modified: $Date$