| blob | 8e6a7fd4d7c606fba2ed5320227b0dc51e942b14 |
3 <html>
4 <head>
7 </head>
8 <body>
11 LLVM Alias Analysis Infrastructure
12 </div>
14 <ol>
18 <ul>
23 </ul>
24 </li>
27 <ul>
35 </ul>
36 </li>
39 <ul>
43 </ul>
44 </li>
47 <ul>
51 implementations</a></li>
52 </ul>
53 </li>
55 </ol>
59 </div>
61 <!-- *********************************************************************** -->
64 </div>
65 <!-- *********************************************************************** -->
69 <p>Alias Analysis (aka Pointer Analysis) is a class of techniques which attempt
70 to determine whether or not two pointers ever can point to the same object in
71 memory. There are many different algorithms for alias analysis and many
72 different ways of classifying them: flow-sensitive vs flow-insensitive,
73 context-sensitive vs context-insensitive, field-sensitive vs field-insensitive,
74 unification-based vs subset-based, etc. Traditionally, alias analyses respond
76 indicating that two pointers always point to the same object, might point to the
77 same object, or are known to never point to the same object.</p>
81 class is the primary interface used by clients and implementations of alias
82 analyses in the LLVM system. This class is the common interface between clients
83 of alias analysis information and the implementations providing it, and is
84 designed to support a wide range of implementations and clients (but currently
85 all clients are assumed to be flow-insensitive). In addition to simple alias
86 analysis information, this class exposes Mod/Ref information from those
87 implementations which can provide it, allowing for powerful analyses and
88 transformations to work well together.</p>
90 <p>This document contains information necessary to successfully implement this
91 interface, use it, and to test both sides. It also explains some of the finer
92 points about what exactly results mean. If you feel that something is unclear
96 </div>
98 <!-- *********************************************************************** -->
101 </div>
102 <!-- *********************************************************************** -->
108 class defines the interface that the various alias analysis implementations
111 mod/ref query, respectively.</p>
114 represented in several different ways. In particular, memory objects are
115 represented as a starting address and size, and function calls are represented
118 which allow you to get mod/ref information for arbitrary instructions.</p>
120 </div>
122 <!-- ======================================================================= -->
125 </div>
130 which are used to query whether or not two memory objects alias, whether
131 function calls can modify or read a memory object, etc. For all of these
132 queries, memory objects are represented as a pair of their starting address (a
135 <p>Representing memory objects as a starting address and a size is critically
136 important for correct Alias Analyses. For example, consider this (silly, but
137 possible) C code:</p>
140 <pre>
141 int i;
142 char C[2];
143 char A[10];
144 /* ... */
146 C[0] = A[i]; /* One byte store */
148 }
149 </pre>
150 </div>
154 locations one byte apart, and the accesses are each one byte. In this case, the
155 LICM pass can use store motion to remove the stores from the loop. In
156 constrast, the following code:</p>
159 <pre>
160 int i;
161 char C[2];
162 char A[10];
163 /* ... */
165 ((short*)C)[0] = A[i]; /* Two byte store! */
167 }
168 </pre>
169 </div>
171 <p>In this case, the two stores to C do alias each other, because the access to
173 available in the query, even the first case would have to conservatively assume
174 that the accesses alias.</p>
176 </div>
178 <!-- ======================================================================= -->
181 </div>
185 not two memory objects alias each other. It takes two memory objects as input
186 and returns MustAlias, MayAlias, or NoAlias as appropriate.
187 </div>
189 <!-- _______________________________________________________________________ -->
192 </div>
195 <p>The NoAlias response may be used when there is never an immediate dependence
198 pointers point to non-overlapping memory ranges. Another is when the two
199 pointers are only ever used for reading memory. Another is when the memory is
200 freed and reallocated between accesses through one pointer and accesses through
201 the other -- in this case, there is a dependence, but it's mediated by the free
202 and reallocation.</p>
204 <p>As an exception to this is with the
207 the "irrelevant" dependencies, provided their clients do not need to be aware
208 of these dependencies for correctness.</p>
210 <p>The MayAlias response is used whenever the two pointers might refer to the
211 same object. If the two memory objects overlap, but do not start at the same
212 location, return MayAlias.</p>
214 <p>The MustAlias response may only be returned if the two memory objects are
215 guaranteed to always start at exactly the same location. A MustAlias response
216 implies that the pointers compare equal.</p>
218 </div>
220 <!-- ======================================================================= -->
223 </div>
228 execution of an instruction can read or modify a memory location. Mod/Ref
230 a location, ModRef is returned.</p>
233 method for testing dependencies between function calls. This method takes two
234 call sites (CS1 & CS2), returns NoModRef if the two calls refer to disjoint
235 memory locations, Ref if CS1 reads memory written by CS2, Mod if CS1 writes to
236 memory read or written by CS2, or ModRef if CS1 might read or write memory
237 accessed by CS2. Note that this relation is not commutative.</p>
239 </div>
242 <!-- ======================================================================= -->
245 </div>
249 <p>
250 Several other tidbits of information are often collected by various alias
251 analysis implementations and can be put to good use by various clients.
252 </p>
254 </div>
256 <!-- _______________________________________________________________________ -->
259 </div>
264 analysis can prove that the pointer only points to unchanging memory locations
265 (functions, constant global variables, and the null pointer). This information
266 can be used to refine mod/ref information: it is impossible for an unchanging
267 memory location to be modified.</p>
269 </div>
271 <!-- _______________________________________________________________________ -->
275 </div>
279 <p>These methods are used to provide very simple mod/ref information for
281 function if the analysis can prove that the function never reads or writes to
282 memory, or if the function only reads from constant memory. Functions with this
283 property are side-effect free and only depend on their input arguments, allowing
284 them to be eliminated if they form common subexpressions or be hoisted out of
290 can prove that (at most) the function only reads from non-volatile memory.
291 Functions with this property are side-effect free, only depending on their input
292 arguments and the state of memory when they are called. This property allows
293 calls to these functions to be eliminated and moved around, as long as there is
294 no store instruction that changes the contents of memory. Note that all
298 </div>
300 <!-- *********************************************************************** -->
303 </div>
304 <!-- *********************************************************************** -->
308 <p>Writing a new alias analysis implementation for LLVM is quite
309 straight-forward. There are already several implementations that you can use
310 for examples, and the following information should help fill in any details.
314 </div>
316 <!-- ======================================================================= -->
319 </div>
325 Analysis. As is the case with most other analyses and transformations, the
326 answer should be fairly obvious from what type of problem you are trying to
327 solve:</p>
329 <ol>
330 <li>If you require interprocedural analysis, it should be a
333 <li>If you don't need to look at the program at all, subclass
335 </ol>
337 <p>In addition to the pass that you subclass, you should also inherit from the
342 </div>
344 <!-- ======================================================================= -->
347 </div>
356 declaring any pass dependencies your pass has. Thus you should have something
357 like this:</p>
360 <pre>
361 void getAnalysisUsage(AnalysisUsage &AU) const {
362 AliasAnalysis::getAnalysisUsage(AU);
364 }
365 </pre>
366 </div>
374 <pre>
375 bool run(Module &M) {
376 InitializeAliasAnalysis(this);
378 return false;
379 }
380 </pre>
381 </div>
383 </div>
385 <!-- ======================================================================= -->
388 </div>
395 alias analysis implementation, which ends up returning conservatively correct
397 respectively). Depending on the capabilities of the analysis you are
398 implementing, you just override the interfaces you can improve.</p>
400 </div>
404 <!-- ======================================================================= -->
407 </div>
413 passes) every alias analysis pass chains to another alias analysis
414 implementation (for example, the user can specify "<tt>-basicaa -ds-aa
415 -licm</tt>" to get the maximum benefit from both alias
416 analyses). The alias analysis class automatically takes care of most of this
417 for methods that you don't override. For methods that you do override, in code
418 paths that return a conservative MayAlias or Mod/Ref result, simply return
419 whatever the superclass computes. For example:</p>
422 <pre>
423 AliasAnalysis::AliasResult alias(const Value *V1, unsigned V1Size,
424 const Value *V2, unsigned V2Size) {
425 if (...)
426 return NoAlias;
427 ...
430 return AliasAnalysis::alias(V1, V1Size, V2, V2Size);
431 }
432 </pre>
433 </div>
435 <p>In addition to analysis queries, you must make sure to unconditionally pass
437 well if you override them, which allows all alias analyses in a change to be
438 updated.</p>
440 </div>
443 <!-- ======================================================================= -->
446 </div>
449 <p>
450 Alias analysis information is initially computed for a static snapshot of the
451 program, but clients will use this information to make transformations to the
452 code. All but the most trivial forms of alias analysis will need to have their
453 analysis results updated to reflect the changes made by these transformations.
454 </p>
456 <p>
458 communicate program changes from the clients to the analysis implementations.
459 Various alias analysis implementations should use these methods to ensure that
460 their internal data structures are kept up-to-date as the program changes (for
461 example, when an instruction is deleted), and clients of alias analysis must be
462 sure to call these interfaces appropriately.
463 </p>
464 </div>
466 <!-- _______________________________________________________________________ -->
471 instruction or any other value from the program (including values that do not
472 use pointers). Typically alias analyses keep data structures that have entries
473 for each value in the program. When this method is called, they should remove
474 any entries for the specified value, if they exist.
475 </div>
477 <!-- _______________________________________________________________________ -->
482 program. There is no way to introduce a value into the program that did not
483 exist before (this doesn't make sense for a safe compiler transformation), so
484 this is the only way to introduce a new value. This method indicates that the
485 new value has exactly the same properties as the value being copied.
486 </div>
488 <!-- _______________________________________________________________________ -->
492 This method is a simple helper method that is provided to make clients easier to
493 use. It is implemented by copying the old analysis information to the new
494 value, then deleting the old value. This method cannot be overridden by alias
495 analysis implementations.
496 </div>
498 <!-- ======================================================================= -->
501 </div>
505 <p>From the LLVM perspective, the only thing you need to do to provide an
507 serviced quickly. The actual calculation of the alias analysis results (the
508 "run" method) is only performed once, but many (perhaps duplicate) queries may
509 be performed. Because of this, try to move as much computation to the run
510 method as possible (within reason).</p>
512 </div>
514 <!-- ======================================================================= -->
517 </div>
521 <p>PassManager support for alternative AliasAnalysis implementation
522 has some issues.</p>
524 <p>There is no way to override the default alias analysis. It would
525 be very useful to be able to do something like "opt -my-aa -O2" and
526 have it use -my-aa for all passes which need AliasAnalysis, but there
527 is currently no support for that, short of changing the source code
528 and recompiling. Similarly, there is also no way of setting a chain
529 of analyses as the default.</p>
531 <p>There is no way for transform passes to declare that they preserve
534 which are intended to allow a pass to keep an AliasAnalysis consistent,
535 however there's no way for a pass to declare in its
538 actually have any effect.</tt>
543 these utilities. If you try to use them, the pass manager will
544 silently route alias analysis queries directly to
549 alias analysis passes.</p>
551 </div>
553 <!-- *********************************************************************** -->
556 </div>
557 <!-- *********************************************************************** -->
561 <p>There are several different ways to use alias analysis results. In order of
562 preference, these are...</p>
564 </div>
566 <!-- ======================================================================= -->
569 </div>
574 information about memory-using instructions. This will tell you which store
575 feeds into a load, for example. It uses caching and other techniques to be
576 efficient, and is used by Dead Store Elimination, GVN, and memcpy optimizations.
577 </p>
579 </div>
581 <!-- ======================================================================= -->
584 </div>
589 in some scope, rather than information about pairwise aliasing. The <tt><a
591 is used to efficiently build these Alias Sets from the pairwise alias analysis
595 to add information about various potentially aliasing instructions in the scope
596 you are interested in. Once all of the alias sets are completed, your pass
597 should simply iterate through the constructed alias sets, using the
601 to be disjoint, calculate mod/ref information and volatility for the set, and
602 keep track of whether or not all of the pointers in the set are Must aliases.
603 The AliasSetTracker also makes sure that sets are properly folded due to call
604 instructions, and can provide a list of pointers in each set.</p>
609 then all load instructions from that set may be hoisted out of the loop. If any
611 sunk to outside of the loop, promoting the memory location to a register for the
612 duration of the loop nest. Both of these transformations only apply if the
613 pointer argument is loop-invariant.</p>
615 </div>
617 <!-- _______________________________________________________________________ -->
619 The AliasSetTracker implementation
620 </div>
624 <p>The AliasSetTracker class is implemented to be as efficient as possible. It
625 uses the union-find algorithm to efficiently merge AliasSets when a pointer is
626 inserted into the AliasSetTracker that aliases multiple sets. The primary data
627 structure is a hash table mapping pointers to the AliasSet they are in.</p>
629 <p>The AliasSetTracker class must maintain a list of all of the LLVM Value*'s
630 that are in each AliasSet. Since the hash table already has entries for each
631 LLVM Value* of interest, the AliasesSets thread the linked list through these
632 hash-table nodes to avoid having to allocate memory unnecessarily, and to make
633 merging alias sets extremely efficient (the linked list merge is constant time).
634 </p>
636 <p>You shouldn't need to understand these details if you are just a client of
637 the AliasSetTracker, but if you look at the code, hopefully this brief
638 description will help make sense of why things are designed the way they
639 are.</p>
641 </div>
643 <!-- ======================================================================= -->
646 </div>
650 <p>If neither of these utility class are what your pass needs, you should use
652 the higher-level methods when possible (e.g., use mod/ref information instead of
654 best precision and efficiency.</p>
656 </div>
658 <!-- *********************************************************************** -->
661 </div>
662 <!-- *********************************************************************** -->
666 <p>If you're going to be working with the LLVM alias analysis infrastructure,
667 you should know what clients and implementations of alias analysis are
668 available. In particular, if you are implementing an alias analysis, you should
670 for monitoring and evaluating different implementations.</p>
672 </div>
674 <!-- ======================================================================= -->
677 </div>
686 </div>
688 <!-- _______________________________________________________________________ -->
691 </div>
696 never returns any useful information. This pass can be useful if you think that
697 alias analysis is doing something wrong and are trying to narrow down a
698 problem.</p>
700 </div>
702 <!-- _______________________________________________________________________ -->
705 </div>
712 <ul>
713 <li>Distinct globals, stack allocations, and heap allocations can never
714 alias.</li>
715 <li>Globals, stack allocations, and heap allocations never alias the null
716 pointer.</li>
721 <li>Pointers that obviously point to constant globals
723 <li>Function calls can not modify or references stack allocations if they never
724 escape from the function that allocates them (a common case for automatic
725 arrays).</li>
726 </ul>
728 </div>
730 <!-- _______________________________________________________________________ -->
733 </div>
737 <p>This pass implements a simple context-sensitive mod/ref and alias analysis
738 for internal global variables that don't "have their address taken". If a
739 global does not have its address taken, the pass knows that no pointers alias
740 the global. This pass also keeps track of functions that it knows never access
741 memory or never read memory. This allows certain optimizations (e.g. GVN) to
742 eliminate call instructions entirely.
743 </p>
745 <p>The real power of this pass is that it provides context-sensitive mod/ref
746 information for call instructions. This allows the optimizer to know that
747 calls to a function do not clobber or read the value of the global, allowing
748 loads and stores to be eliminated.</p>
750 <p>Note that this pass is somewhat limited in its scope (only support
751 non-address taken globals), but is very quick analysis.</p>
752 </div>
754 <!-- _______________________________________________________________________ -->
757 </div>
762 "Steensgaard's algorithm" for interprocedural alias analysis. Steensgaard's
763 algorithm is a unification-based, flow-insensitive, context-insensitive, and
764 field-insensitive alias analysis that is also very scalable (effectively linear
765 time).</p>
768 field-<b>sensitive</b>" version of Steensgaard's algorithm using the Data
769 Structure Analysis framework. This gives it substantially more precision than
770 the standard algorithm while maintaining excellent analysis scalability.</p>
773 module, it is not part of the LLVM core.</p>
775 </div>
777 <!-- _______________________________________________________________________ -->
780 </div>
785 algorithm. Data Structure Analysis is a modular unification-based,
788 O(n*log(n)).</p>
790 <p>This algorithm is capable of responding to a full variety of alias analysis
791 queries, and can provide context-sensitive mod/ref information as well. The
792 only major facility not implemented so far is support for must-alias
793 information.</p>
796 module, it is not part of the LLVM core.</p>
798 </div>
800 <!-- _______________________________________________________________________ -->
803 </div>
808 translating them into ScalarEvolution queries. This gives it a
810 and loop induction variables than other alias analyses have.</p>
812 </div>
814 <!-- ======================================================================= -->
817 </div>
820 LLVM includes several alias-analysis driven transformations which can be used
821 with any of the implementations above.
822 </div>
824 <!-- _______________________________________________________________________ -->
827 </div>
833 not have side-effects and are not used.</p>
835 </div>
838 <!-- _______________________________________________________________________ -->
841 </div>
847 different transformations:</p>
849 <ul>
850 <li>It uses mod/ref information to hoist or sink load instructions out of loops
851 if there are no instructions in the loop that modifies the memory loaded.</li>
853 <li>It uses mod/ref information to hoist function calls out of loops that do not
854 write to memory and are loop-invariant.</li>
856 <li>If uses alias information to promote memory objects that are loaded and
857 stored to in loops to live in a register instead. It can do this if there are
858 no may aliases to the loaded/stored memory location.</li>
859 </ul>
861 </div>
863 <!-- _______________________________________________________________________ -->
866 </div>
869 <p>
871 by-value instead. In particular, if pointer arguments are only loaded from it
872 passes in the value loaded instead of the address to the function. This pass
873 uses alias information to make sure that the value loaded from the argument
874 pointer is not modified between the entry of the function and any load of the
875 pointer.</p>
876 </div>
878 <!-- _______________________________________________________________________ -->
881 passes</a>
882 </div>
886 <p>These passes use AliasAnalysis information to reason about loads and stores.
887 </p>
889 </div>
891 <!-- ======================================================================= -->
894 implementations</a>
895 </div>
899 <p>These passes are useful for evaluating the various alias analysis
900 implementations. You can use them with commands like '<tt>opt -ds-aa
903 </div>
905 <!-- _______________________________________________________________________ -->
908 </div>
918 <pre>
919 % opt -ds-aa -print-alias-sets -disable-output
920 </pre>
921 </div>
923 </div>
926 <!-- _______________________________________________________________________ -->
929 </div>
934 pass is making and what responses are returned by the alias analysis. As an
935 example,</p>
938 <pre>
939 % opt -basicaa -count-aa -ds-aa -count-aa -licm
940 </pre>
941 </div>
943 <p>will print out how many queries (and what responses are returned) by the
946 when debugging a transformation or an alias analysis implementation.</p>
948 </div>
950 <!-- _______________________________________________________________________ -->
953 </div>
958 function and asks an alias analysis whether or not the pointers alias. This
959 gives an indication of the precision of the alias analysis. Statistics are
960 printed indicating the percent of no/may/must aliases found (a more precise
961 algorithm will have a lower number of may aliases).</p>
963 </div>
965 <!-- *********************************************************************** -->
968 </div>
969 <!-- *********************************************************************** -->
973 <p>If you're just looking to be a client of alias analysis information, consider
974 using the Memory Dependence Analysis interface instead. MemDep is a lazy,
975 caching layer on top of alias analysis that is able to answer the question of
976 what preceding memory operations a given instruction depends on, either at an
977 intra- or inter-block level. Because of its laziness and caching
978 policy, using MemDep can be a significant performance win over accessing alias
979 analysis directly.</p>
981 </div>
983 <!-- *********************************************************************** -->
985 <hr>
986 <address>
994 Last modified: $Date$
995 </address>
997 </body>
998 </html>