| blob | 5e36ae14e85a8c9af332166baaf4753b0dfdbcb4 |
3 <html>
4 <head>
7 </head>
8 <body>
10 <h1>
11 LLVM Alias Analysis Infrastructure
12 </h1>
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 <!-- *********************************************************************** -->
62 <h2>
64 </h2>
65 <!-- *********************************************************************** -->
67 <div>
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 <!-- *********************************************************************** -->
99 <h2>
101 </h2>
102 <!-- *********************************************************************** -->
104 <div>
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>
121 multiple values, values which are not
123 same function.</p>
125 <!-- ======================================================================= -->
126 <h3>
128 </h3>
130 <div>
133 which are used to query whether or not two memory objects alias, whether
134 function calls can modify or read a memory object, etc. For all of these
135 queries, memory objects are represented as a pair of their starting address (a
138 <p>Representing memory objects as a starting address and a size is critically
139 important for correct Alias Analyses. For example, consider this (silly, but
140 possible) C code:</p>
143 <pre>
144 int i;
145 char C[2];
146 char A[10];
147 /* ... */
149 C[0] = A[i]; /* One byte store */
151 }
152 </pre>
153 </div>
157 locations one byte apart, and the accesses are each one byte. In this case, the
158 LICM pass can use store motion to remove the stores from the loop. In
159 constrast, the following code:</p>
162 <pre>
163 int i;
164 char C[2];
165 char A[10];
166 /* ... */
168 ((short*)C)[0] = A[i]; /* Two byte store! */
170 }
171 </pre>
172 </div>
174 <p>In this case, the two stores to C do alias each other, because the access to
176 available in the query, even the first case would have to conservatively assume
177 that the accesses alias.</p>
179 </div>
181 <!-- ======================================================================= -->
182 <h3>
184 </h3>
186 <div>
188 or not two memory objects alias each other. It takes two memory objects as
189 input and returns MustAlias, PartialAlias, MayAlias, or NoAlias as
190 appropriate.</p>
193 that either the two pointer values be defined within the same function, or at
196 <!-- _______________________________________________________________________ -->
197 <h4>
199 </h4>
201 <div>
202 <p>The NoAlias response may be used when there is never an immediate dependence
205 pointers point to non-overlapping memory ranges. Another is when the two
206 pointers are only ever used for reading memory. Another is when the memory is
207 freed and reallocated between accesses through one pointer and accesses through
208 the other -- in this case, there is a dependence, but it's mediated by the free
209 and reallocation.</p>
211 <p>As an exception to this is with the
213 dependencies are ignored.</p>
215 <p>The MayAlias response is used whenever the two pointers might refer to the
216 same object.</p>
218 <p>The PartialAlias response is used when the two memory objects are known
219 to be overlapping in some way, but do not start at the same address.</p>
221 <p>The MustAlias response may only be returned if the two memory objects are
222 guaranteed to always start at exactly the same location. A MustAlias response
223 implies that the pointers compare equal.</p>
225 </div>
227 </div>
229 <!-- ======================================================================= -->
230 <h3>
232 </h3>
234 <div>
237 execution of an instruction can read or modify a memory location. Mod/Ref
239 a location, ModRef is returned.</p>
242 method for testing dependencies between function calls. This method takes two
243 call sites (CS1 & CS2), returns NoModRef if neither call writes to memory
244 read or written by the other, Ref if CS1 reads memory written by CS2, Mod if CS1
245 writes to memory read or written by CS2, or ModRef if CS1 might read or write
246 memory written to by CS2. Note that this relation is not commutative.</p>
248 </div>
251 <!-- ======================================================================= -->
252 <h3>
254 </h3>
256 <div>
258 <p>
259 Several other tidbits of information are often collected by various alias
260 analysis implementations and can be put to good use by various clients.
261 </p>
263 <!-- _______________________________________________________________________ -->
264 <h4>
266 </h4>
268 <div>
271 analysis can prove that the pointer only points to unchanging memory locations
272 (functions, constant global variables, and the null pointer). This information
273 can be used to refine mod/ref information: it is impossible for an unchanging
274 memory location to be modified.</p>
276 </div>
278 <!-- _______________________________________________________________________ -->
279 <h4>
282 </h4>
284 <div>
286 <p>These methods are used to provide very simple mod/ref information for
288 function if the analysis can prove that the function never reads or writes to
289 memory, or if the function only reads from constant memory. Functions with this
290 property are side-effect free and only depend on their input arguments, allowing
291 them to be eliminated if they form common subexpressions or be hoisted out of
297 can prove that (at most) the function only reads from non-volatile memory.
298 Functions with this property are side-effect free, only depending on their input
299 arguments and the state of memory when they are called. This property allows
300 calls to these functions to be eliminated and moved around, as long as there is
301 no store instruction that changes the contents of memory. Note that all
305 </div>
307 </div>
309 </div>
311 <!-- *********************************************************************** -->
312 <h2>
314 </h2>
315 <!-- *********************************************************************** -->
317 <div>
319 <p>Writing a new alias analysis implementation for LLVM is quite
320 straight-forward. There are already several implementations that you can use
321 for examples, and the following information should help fill in any details.
325 <!-- ======================================================================= -->
326 <h3>
328 </h3>
330 <div>
334 Analysis. As is the case with most other analyses and transformations, the
335 answer should be fairly obvious from what type of problem you are trying to
336 solve:</p>
338 <ol>
339 <li>If you require interprocedural analysis, it should be a
342 <li>If you don't need to look at the program at all, subclass
344 </ol>
346 <p>In addition to the pass that you subclass, you should also inherit from the
351 </div>
353 <!-- ======================================================================= -->
354 <h3>
356 </h3>
358 <div>
365 declaring any pass dependencies your pass has. Thus you should have something
366 like this:</p>
369 <pre>
370 void getAnalysisUsage(AnalysisUsage &AU) const {
371 AliasAnalysis::getAnalysisUsage(AU);
373 }
374 </pre>
375 </div>
383 <pre>
384 bool run(Module &M) {
385 InitializeAliasAnalysis(this);
387 return false;
388 }
389 </pre>
390 </div>
392 </div>
394 <!-- ======================================================================= -->
395 <h3>
397 </h3>
399 <div>
404 alias analysis implementation, which ends up returning conservatively correct
406 respectively). Depending on the capabilities of the analysis you are
407 implementing, you just override the interfaces you can improve.</p>
409 </div>
413 <!-- ======================================================================= -->
414 <h3>
416 </h3>
418 <div>
422 passes) every alias analysis pass chains to another alias analysis
423 implementation (for example, the user can specify "<tt>-basicaa -ds-aa
424 -licm</tt>" to get the maximum benefit from both alias
425 analyses). The alias analysis class automatically takes care of most of this
426 for methods that you don't override. For methods that you do override, in code
427 paths that return a conservative MayAlias or Mod/Ref result, simply return
428 whatever the superclass computes. For example:</p>
431 <pre>
432 AliasAnalysis::AliasResult alias(const Value *V1, unsigned V1Size,
433 const Value *V2, unsigned V2Size) {
434 if (...)
435 return NoAlias;
436 ...
439 return AliasAnalysis::alias(V1, V1Size, V2, V2Size);
440 }
441 </pre>
442 </div>
444 <p>In addition to analysis queries, you must make sure to unconditionally pass
446 well if you override them, which allows all alias analyses in a change to be
447 updated.</p>
449 </div>
452 <!-- ======================================================================= -->
453 <h3>
455 </h3>
457 <div>
458 <p>
459 Alias analysis information is initially computed for a static snapshot of the
460 program, but clients will use this information to make transformations to the
461 code. All but the most trivial forms of alias analysis will need to have their
462 analysis results updated to reflect the changes made by these transformations.
463 </p>
465 <p>
467 communicate program changes from the clients to the analysis implementations.
468 Various alias analysis implementations should use these methods to ensure that
469 their internal data structures are kept up-to-date as the program changes (for
470 example, when an instruction is deleted), and clients of alias analysis must be
471 sure to call these interfaces appropriately.
472 </p>
474 <!-- _______________________________________________________________________ -->
477 <div>
479 instruction or any other value from the program (including values that do not
480 use pointers). Typically alias analyses keep data structures that have entries
481 for each value in the program. When this method is called, they should remove
482 any entries for the specified value, if they exist.
483 </div>
485 <!-- _______________________________________________________________________ -->
488 <div>
490 program. There is no way to introduce a value into the program that did not
491 exist before (this doesn't make sense for a safe compiler transformation), so
492 this is the only way to introduce a new value. This method indicates that the
493 new value has exactly the same properties as the value being copied.
494 </div>
496 <!-- _______________________________________________________________________ -->
499 <div>
500 This method is a simple helper method that is provided to make clients easier to
501 use. It is implemented by copying the old analysis information to the new
502 value, then deleting the old value. This method cannot be overridden by alias
503 analysis implementations.
504 </div>
506 <!-- _______________________________________________________________________ -->
509 <div>
511 value have changed in ways that may invalidate precomputed analysis information.
512 Implementations may either use this callback to provide conservative responses
513 for points whose uses have change since analysis time, or may recompute some
514 or all of their internal state to continue providing accurate responses.</p>
516 <p>In general, any new use of a pointer value is considered an escaping use,
518 uses below:</p>
520 <ul>
525 </ul>
526 </div>
528 </div>
530 <!-- ======================================================================= -->
531 <h3>
533 </h3>
535 <div>
537 <p>From the LLVM perspective, the only thing you need to do to provide an
539 serviced quickly. The actual calculation of the alias analysis results (the
540 "run" method) is only performed once, but many (perhaps duplicate) queries may
541 be performed. Because of this, try to move as much computation to the run
542 method as possible (within reason).</p>
544 </div>
546 <!-- ======================================================================= -->
547 <h3>
549 </h3>
551 <div>
553 <p>The AliasAnalysis infrastructure has several limitations which make
556 <p>There is no way to override the default alias analysis. It would
557 be very useful to be able to do something like "opt -my-aa -O2" and
558 have it use -my-aa for all passes which need AliasAnalysis, but there
559 is currently no support for that, short of changing the source code
560 and recompiling. Similarly, there is also no way of setting a chain
561 of analyses as the default.</p>
563 <p>There is no way for transform passes to declare that they preserve
566 which are intended to allow a pass to keep an AliasAnalysis consistent,
567 however there's no way for a pass to declare in its
570 actually have any effect.</p>
575 these utilities. If you try to use them, the pass manager will
576 silently route alias analysis queries directly to
581 alias analysis passes.</p>
584 implementations when values are deleted or copied, however these
585 aren't sufficient. There are many other ways that LLVM IR can be
587 implementations which can not be expressed.</p>
593 support this, and are known to trigger errors when run with the
596 <p>Due to several of the above limitations, the most obvious use for
598 alias queries in a compilation, doesn't work, even if the
600 There's no way to set a default, much less a default sequence,
601 and there's no way to preserve it.</p>
604 makes a non-deterministic number of alias queries. This can cause stats
606 identical runs, for example. Another consequence is that debugging
607 techniques involving pausing execution after a predetermined number
608 of queries can be unreliable.</p>
610 <p>Many alias queries can be reformulated in terms of other alias
612 it would make sense to start those queries from the beginning of the chain,
613 with care taken to avoid infinite looping, however currently an
614 implementation which wants to do this can only start such queries
615 from itself.</p>
617 </div>
619 </div>
621 <!-- *********************************************************************** -->
622 <h2>
624 </h2>
625 <!-- *********************************************************************** -->
627 <div>
629 <p>There are several different ways to use alias analysis results. In order of
630 preference, these are...</p>
632 <!-- ======================================================================= -->
633 <h3>
635 </h3>
637 <div>
640 information about memory-using instructions. This will tell you which store
641 feeds into a load, for example. It uses caching and other techniques to be
642 efficient, and is used by Dead Store Elimination, GVN, and memcpy optimizations.
643 </p>
645 </div>
647 <!-- ======================================================================= -->
648 <h3>
650 </h3>
652 <div>
655 in some scope, rather than information about pairwise aliasing. The <tt><a
657 is used to efficiently build these Alias Sets from the pairwise alias analysis
661 to add information about various potentially aliasing instructions in the scope
662 you are interested in. Once all of the alias sets are completed, your pass
663 should simply iterate through the constructed alias sets, using the
667 to be disjoint, calculate mod/ref information and volatility for the set, and
668 keep track of whether or not all of the pointers in the set are Must aliases.
669 The AliasSetTracker also makes sure that sets are properly folded due to call
670 instructions, and can provide a list of pointers in each set.</p>
675 then all load instructions from that set may be hoisted out of the loop. If any
677 sunk to outside of the loop, promoting the memory location to a register for the
678 duration of the loop nest. Both of these transformations only apply if the
679 pointer argument is loop-invariant.</p>
681 <!-- _______________________________________________________________________ -->
682 <h4>
683 The AliasSetTracker implementation
684 </h4>
686 <div>
688 <p>The AliasSetTracker class is implemented to be as efficient as possible. It
689 uses the union-find algorithm to efficiently merge AliasSets when a pointer is
690 inserted into the AliasSetTracker that aliases multiple sets. The primary data
691 structure is a hash table mapping pointers to the AliasSet they are in.</p>
693 <p>The AliasSetTracker class must maintain a list of all of the LLVM Value*'s
694 that are in each AliasSet. Since the hash table already has entries for each
695 LLVM Value* of interest, the AliasesSets thread the linked list through these
696 hash-table nodes to avoid having to allocate memory unnecessarily, and to make
697 merging alias sets extremely efficient (the linked list merge is constant time).
698 </p>
700 <p>You shouldn't need to understand these details if you are just a client of
701 the AliasSetTracker, but if you look at the code, hopefully this brief
702 description will help make sense of why things are designed the way they
703 are.</p>
705 </div>
707 </div>
709 <!-- ======================================================================= -->
710 <h3>
712 </h3>
714 <div>
716 <p>If neither of these utility class are what your pass needs, you should use
718 the higher-level methods when possible (e.g., use mod/ref information instead of
720 best precision and efficiency.</p>
722 </div>
724 </div>
726 <!-- *********************************************************************** -->
727 <h2>
729 </h2>
730 <!-- *********************************************************************** -->
732 <div>
734 <p>If you're going to be working with the LLVM alias analysis infrastructure,
735 you should know what clients and implementations of alias analysis are
736 available. In particular, if you are implementing an alias analysis, you should
738 for monitoring and evaluating different implementations.</p>
740 <!-- ======================================================================= -->
741 <h3>
743 </h3>
745 <div>
750 analysis implementations.</p>
752 <!-- _______________________________________________________________________ -->
753 <h4>
755 </h4>
757 <div>
760 never returns any useful information. This pass can be useful if you think that
761 alias analysis is doing something wrong and are trying to narrow down a
762 problem.</p>
764 </div>
766 <!-- _______________________________________________________________________ -->
767 <h4>
769 </h4>
771 <div>
774 many important facts:</p>
776 <ul>
777 <li>Distinct globals, stack allocations, and heap allocations can never
778 alias.</li>
779 <li>Globals, stack allocations, and heap allocations never alias the null
780 pointer.</li>
785 <li>Pointers that obviously point to constant globals
787 <li>Function calls can not modify or references stack allocations if they never
788 escape from the function that allocates them (a common case for automatic
789 arrays).</li>
790 </ul>
792 </div>
794 <!-- _______________________________________________________________________ -->
795 <h4>
797 </h4>
799 <div>
801 <p>This pass implements a simple context-sensitive mod/ref and alias analysis
802 for internal global variables that don't "have their address taken". If a
803 global does not have its address taken, the pass knows that no pointers alias
804 the global. This pass also keeps track of functions that it knows never access
805 memory or never read memory. This allows certain optimizations (e.g. GVN) to
806 eliminate call instructions entirely.
807 </p>
809 <p>The real power of this pass is that it provides context-sensitive mod/ref
810 information for call instructions. This allows the optimizer to know that
811 calls to a function do not clobber or read the value of the global, allowing
812 loads and stores to be eliminated.</p>
814 <p>Note that this pass is somewhat limited in its scope (only support
815 non-address taken globals), but is very quick analysis.</p>
816 </div>
818 <!-- _______________________________________________________________________ -->
819 <h4>
821 </h4>
823 <div>
826 "Steensgaard's algorithm" for interprocedural alias analysis. Steensgaard's
827 algorithm is a unification-based, flow-insensitive, context-insensitive, and
828 field-insensitive alias analysis that is also very scalable (effectively linear
829 time).</p>
832 field-<b>sensitive</b>" version of Steensgaard's algorithm using the Data
833 Structure Analysis framework. This gives it substantially more precision than
834 the standard algorithm while maintaining excellent analysis scalability.</p>
837 module, it is not part of the LLVM core.</p>
839 </div>
841 <!-- _______________________________________________________________________ -->
842 <h4>
844 </h4>
846 <div>
849 algorithm. Data Structure Analysis is a modular unification-based,
852 O(n*log(n)).</p>
854 <p>This algorithm is capable of responding to a full variety of alias analysis
855 queries, and can provide context-sensitive mod/ref information as well. The
856 only major facility not implemented so far is support for must-alias
857 information.</p>
860 module, it is not part of the LLVM core.</p>
862 </div>
864 <!-- _______________________________________________________________________ -->
865 <h4>
867 </h4>
869 <div>
872 translating them into ScalarEvolution queries. This gives it a
874 and loop induction variables than other alias analyses have.</p>
876 </div>
878 </div>
880 <!-- ======================================================================= -->
881 <h3>
883 </h3>
885 <div>
886 LLVM includes several alias-analysis driven transformations which can be used
887 with any of the implementations above.
889 <!-- _______________________________________________________________________ -->
890 <h4>
892 </h4>
894 <div>
898 not have side-effects and are not used.</p>
900 </div>
903 <!-- _______________________________________________________________________ -->
904 <h4>
906 </h4>
908 <div>
912 different transformations:</p>
914 <ul>
915 <li>It uses mod/ref information to hoist or sink load instructions out of loops
916 if there are no instructions in the loop that modifies the memory loaded.</li>
918 <li>It uses mod/ref information to hoist function calls out of loops that do not
919 write to memory and are loop-invariant.</li>
921 <li>If uses alias information to promote memory objects that are loaded and
922 stored to in loops to live in a register instead. It can do this if there are
923 no may aliases to the loaded/stored memory location.</li>
924 </ul>
926 </div>
928 <!-- _______________________________________________________________________ -->
929 <h4>
931 </h4>
933 <div>
934 <p>
936 by-value instead. In particular, if pointer arguments are only loaded from it
937 passes in the value loaded instead of the address to the function. This pass
938 uses alias information to make sure that the value loaded from the argument
939 pointer is not modified between the entry of the function and any load of the
940 pointer.</p>
941 </div>
943 <!-- _______________________________________________________________________ -->
944 <h4>
946 passes</a>
947 </h4>
949 <div>
951 <p>These passes use AliasAnalysis information to reason about loads and stores.
952 </p>
954 </div>
956 </div>
958 <!-- ======================================================================= -->
959 <h3>
961 implementations</a>
962 </h3>
964 <div>
966 <p>These passes are useful for evaluating the various alias analysis
967 implementations. You can use them with commands like '<tt>opt -ds-aa
970 <!-- _______________________________________________________________________ -->
971 <h4>
973 </h4>
975 <div>
983 <pre>
984 % opt -ds-aa -print-alias-sets -disable-output
985 </pre>
986 </div>
988 </div>
991 <!-- _______________________________________________________________________ -->
992 <h4>
994 </h4>
996 <div>
999 pass is making and what responses are returned by the alias analysis. As an
1000 example,</p>
1003 <pre>
1004 % opt -basicaa -count-aa -ds-aa -count-aa -licm
1005 </pre>
1006 </div>
1008 <p>will print out how many queries (and what responses are returned) by the
1011 when debugging a transformation or an alias analysis implementation.</p>
1013 </div>
1015 <!-- _______________________________________________________________________ -->
1016 <h4>
1018 </h4>
1020 <div>
1023 function and asks an alias analysis whether or not the pointers alias. This
1024 gives an indication of the precision of the alias analysis. Statistics are
1025 printed indicating the percent of no/may/must aliases found (a more precise
1026 algorithm will have a lower number of may aliases).</p>
1028 </div>
1030 </div>
1032 </div>
1034 <!-- *********************************************************************** -->
1035 <h2>
1037 </h2>
1038 <!-- *********************************************************************** -->
1040 <div>
1042 <p>If you're just looking to be a client of alias analysis information, consider
1043 using the Memory Dependence Analysis interface instead. MemDep is a lazy,
1044 caching layer on top of alias analysis that is able to answer the question of
1045 what preceding memory operations a given instruction depends on, either at an
1046 intra- or inter-block level. Because of its laziness and caching
1047 policy, using MemDep can be a significant performance win over accessing alias
1048 analysis directly.</p>
1050 </div>
1052 <!-- *********************************************************************** -->
1054 <hr>
1055 <address>
1063 Last modified: $Date$
1064 </address>
1066 </body>
1067 </html>