| blob | 97a813344d894dea99fa9489ee2a02136ee82f96 |
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>
34 </ul>
35 </li>
38 <ul>
42 </ul>
43 </li>
46 <ul>
50 implementations</a></li>
51 </ul>
52 </li>
54 </ol>
58 </div>
60 <!-- *********************************************************************** -->
63 </div>
64 <!-- *********************************************************************** -->
68 <p>Alias Analysis (aka Pointer Analysis) is a class of techniques which attempt
69 to determine whether or not two pointers ever can point to the same object in
70 memory. There are many different algorithms for alias analysis and many
71 different ways of classifying them: flow-sensitive vs flow-insensitive,
72 context-sensitive vs context-insensitive, field-sensitive vs field-insensitive,
73 unification-based vs subset-based, etc. Traditionally, alias analyses respond
75 indicating that two pointers always point to the same object, might point to the
76 same object, or are known to never point to the same object.</p>
80 class is the primary interface used by clients and implementations of alias
81 analyses in the LLVM system. This class is the common interface between clients
82 of alias analysis information and the implementations providing it, and is
83 designed to support a wide range of implementations and clients (but currently
84 all clients are assumed to be flow-insensitive). In addition to simple alias
85 analysis information, this class exposes Mod/Ref information from those
86 implementations which can provide it, allowing for powerful analyses and
87 transformations to work well together.</p>
89 <p>This document contains information necessary to successfully implement this
90 interface, use it, and to test both sides. It also explains some of the finer
91 points about what exactly results mean. If you feel that something is unclear
95 </div>
97 <!-- *********************************************************************** -->
100 </div>
101 <!-- *********************************************************************** -->
107 class defines the interface that the various alias analysis implementations
110 mod/ref query, respectively.</p>
113 represented in several different ways. In particular, memory objects are
114 represented as a starting address and size, and function calls are represented
117 which allow you to get mod/ref information for arbitrary instructions.</p>
119 </div>
121 <!-- ======================================================================= -->
124 </div>
129 which are used to query whether or not two memory objects alias, whether
130 function calls can modify or read a memory object, etc. For all of these
131 queries, memory objects are represented as a pair of their starting address (a
134 <p>Representing memory objects as a starting address and a size is critically
135 important for correct Alias Analyses. For example, consider this (silly, but
136 possible) C code:</p>
139 <pre>
140 int i;
141 char C[2];
142 char A[10];
143 /* ... */
145 C[0] = A[i]; /* One byte store */
147 }
148 </pre>
149 </div>
153 locations one byte apart, and the accesses are each one byte. In this case, the
154 LICM pass can use store motion to remove the stores from the loop. In
155 constrast, the following code:</p>
158 <pre>
159 int i;
160 char C[2];
161 char A[10];
162 /* ... */
164 ((short*)C)[0] = A[i]; /* Two byte store! */
166 }
167 </pre>
168 </div>
170 <p>In this case, the two stores to C do alias each other, because the access to
172 available in the query, even the first case would have to conservatively assume
173 that the accesses alias.</p>
175 </div>
177 <!-- ======================================================================= -->
180 </div>
184 not two memory objects alias each other. It takes two memory objects as input
185 and returns MustAlias, MayAlias, or NoAlias as appropriate.
186 </div>
188 <!-- _______________________________________________________________________ -->
191 </div>
194 <p>The NoAlias response is used when the two pointers refer to distinct objects,
195 regardless of whether the pointers compare equal. For example, freed pointers
196 don't alias any pointers that were allocated afterwards. As a degenerate case,
197 pointers returned by malloc(0) have no bytes for an object, and are considered
198 NoAlias even when malloc returns the same pointer. The same rule applies to
199 NULL pointers.</p>
201 <p>The MayAlias response is used whenever the two pointers might refer to the
202 same object. If the two memory objects overlap, but do not start at the same
203 location, return MayAlias.</p>
205 <p>The MustAlias response may only be returned if the two memory objects are
206 guaranteed to always start at exactly the same location. A MustAlias response
207 implies that the pointers compare equal.</p>
209 </div>
211 <!-- ======================================================================= -->
214 </div>
219 execution of an instruction can read or modify a memory location. Mod/Ref
221 a location, ModRef is returned.</p>
224 method for testing dependencies between function calls. This method takes two
225 call sites (CS1 & CS2), returns NoModRef if the two calls refer to disjoint
226 memory locations, Ref if CS1 reads memory written by CS2, Mod if CS1 writes to
227 memory read or written by CS2, or ModRef if CS1 might read or write memory
228 accessed by CS2. Note that this relation is not commutative. Clients that use
230 method, which indicates whether or not an analysis can provide mod/ref
231 information for function call pairs (most can not). If this predicate is false,
233 method many times.</p>
235 </div>
238 <!-- ======================================================================= -->
241 </div>
245 <p>
246 Several other tidbits of information are often collected by various alias
247 analysis implementations and can be put to good use by various clients.
248 </p>
250 </div>
252 <!-- _______________________________________________________________________ -->
255 </div>
260 always must alias a pointer. This information can be provided in some cases for
261 important objects like the null pointer and global values. Knowing that a
262 pointer always points to a particular function allows indirect calls to be
263 turned into direct calls, for example.</p>
265 </div>
267 <!-- _______________________________________________________________________ -->
270 </div>
275 analysis can prove that the pointer only points to unchanging memory locations
276 (functions, constant global variables, and the null pointer). This information
277 can be used to refine mod/ref information: it is impossible for an unchanging
278 memory location to be modified.</p>
280 </div>
282 <!-- _______________________________________________________________________ -->
286 </div>
290 <p>These methods are used to provide very simple mod/ref information for
292 function if the analysis can prove that the function never reads or writes to
293 memory, or if the function only reads from constant memory. Functions with this
294 property are side-effect free and only depend on their input arguments, allowing
295 them to be eliminated if they form common subexpressions or be hoisted out of
301 can prove that (at most) the function only reads from non-volatile memory.
302 Functions with this property are side-effect free, only depending on their input
303 arguments and the state of memory when they are called. This property allows
304 calls to these functions to be eliminated and moved around, as long as there is
305 no store instruction that changes the contents of memory. Note that all
309 </div>
311 <!-- *********************************************************************** -->
314 </div>
315 <!-- *********************************************************************** -->
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 </div>
327 <!-- ======================================================================= -->
330 </div>
336 Analysis. As is the case with most other analyses and transformations, the
337 answer should be fairly obvious from what type of problem you are trying to
338 solve:</p>
340 <ol>
341 <li>If you require interprocedural analysis, it should be a
344 <li>If you don't need to look at the program at all, subclass
346 </ol>
348 <p>In addition to the pass that you subclass, you should also inherit from the
353 </div>
355 <!-- ======================================================================= -->
358 </div>
367 declaring any pass dependencies your pass has. Thus you should have something
368 like this:</p>
371 <pre>
372 void getAnalysisUsage(AnalysisUsage &AU) const {
373 AliasAnalysis::getAnalysisUsage(AU);
375 }
376 </pre>
377 </div>
385 <pre>
386 bool run(Module &M) {
387 InitializeAliasAnalysis(this);
389 return false;
390 }
391 </pre>
392 </div>
394 </div>
396 <!-- ======================================================================= -->
399 </div>
406 alias analysis implementation, which ends up returning conservatively correct
408 respectively). Depending on the capabilities of the analysis you are
409 implementing, you just override the interfaces you can improve.</p>
411 </div>
415 <!-- ======================================================================= -->
418 </div>
424 passes) every alias analysis pass chains to another alias analysis
425 implementation (for example, the user can specify "<tt>-basicaa -ds-aa
426 -anders-aa -licm</tt>" to get the maximum benefit from the three alias
427 analyses). The alias analysis class automatically takes care of most of this
428 for methods that you don't override. For methods that you do override, in code
429 paths that return a conservative MayAlias or Mod/Ref result, simply return
430 whatever the superclass computes. For example:</p>
433 <pre>
434 AliasAnalysis::AliasResult alias(const Value *V1, unsigned V1Size,
435 const Value *V2, unsigned V2Size) {
436 if (...)
437 return NoAlias;
438 ...
441 return AliasAnalysis::alias(V1, V1Size, V2, V2Size);
442 }
443 </pre>
444 </div>
446 <p>In addition to analysis queries, you must make sure to unconditionally pass
448 well if you override them, which allows all alias analyses in a change to be
449 updated.</p>
451 </div>
454 <!-- ======================================================================= -->
457 </div>
460 <p>
461 Alias analysis information is initially computed for a static snapshot of the
462 program, but clients will use this information to make transformations to the
463 code. All but the most trivial forms of alias analysis will need to have their
464 analysis results updated to reflect the changes made by these transformations.
465 </p>
467 <p>
469 communicate program changes from the clients to the analysis implementations.
470 Various alias analysis implementations should use these methods to ensure that
471 their internal data structures are kept up-to-date as the program changes (for
472 example, when an instruction is deleted), and clients of alias analysis must be
473 sure to call these interfaces appropriately.
474 </p>
475 </div>
477 <!-- _______________________________________________________________________ -->
482 instruction or any other value from the program (including values that do not
483 use pointers). Typically alias analyses keep data structures that have entries
484 for each value in the program. When this method is called, they should remove
485 any entries for the specified value, if they exist.
486 </div>
488 <!-- _______________________________________________________________________ -->
493 program. There is no way to introduce a value into the program that did not
494 exist before (this doesn't make sense for a safe compiler transformation), so
495 this is the only way to introduce a new value. This method indicates that the
496 new value has exactly the same properties as the value being copied.
497 </div>
499 <!-- _______________________________________________________________________ -->
503 This method is a simple helper method that is provided to make clients easier to
504 use. It is implemented by copying the old analysis information to the new
505 value, then deleting the old value. This method cannot be overridden by alias
506 analysis implementations.
507 </div>
509 <!-- ======================================================================= -->
512 </div>
516 <p>From the LLVM perspective, the only thing you need to do to provide an
518 serviced quickly. The actual calculation of the alias analysis results (the
519 "run" method) is only performed once, but many (perhaps duplicate) queries may
520 be performed. Because of this, try to move as much computation to the run
521 method as possible (within reason).</p>
523 </div>
525 <!-- *********************************************************************** -->
528 </div>
529 <!-- *********************************************************************** -->
533 <p>There are several different ways to use alias analysis results. In order of
534 preference, these are...</p>
536 </div>
538 <!-- ======================================================================= -->
541 </div>
547 or transformation can be modeled in a form that uses value numbering
548 information, you don't have to do anything special to handle load instructions:
551 </div>
553 <!-- ======================================================================= -->
556 </div>
561 in some scope, rather than information about pairwise aliasing. The <tt><a
563 is used to efficiently build these Alias Sets from the pairwise alias analysis
567 to add information about various potentially aliasing instructions in the scope
568 you are interested in. Once all of the alias sets are completed, your pass
569 should simply iterate through the constructed alias sets, using the
573 to be disjoint, calculate mod/ref information and volatility for the set, and
574 keep track of whether or not all of the pointers in the set are Must aliases.
575 The AliasSetTracker also makes sure that sets are properly folded due to call
576 instructions, and can provide a list of pointers in each set.</p>
581 then all load instructions from that set may be hoisted out of the loop. If any
583 sunk to outside of the loop, promoting the memory location to a register for the
584 duration of the loop nest. Both of these transformations only apply if the
585 pointer argument is loop-invariant.</p>
587 </div>
589 <!-- _______________________________________________________________________ -->
591 The AliasSetTracker implementation
592 </div>
596 <p>The AliasSetTracker class is implemented to be as efficient as possible. It
597 uses the union-find algorithm to efficiently merge AliasSets when a pointer is
598 inserted into the AliasSetTracker that aliases multiple sets. The primary data
599 structure is a hash table mapping pointers to the AliasSet they are in.</p>
601 <p>The AliasSetTracker class must maintain a list of all of the LLVM Value*'s
602 that are in each AliasSet. Since the hash table already has entries for each
603 LLVM Value* of interest, the AliasesSets thread the linked list through these
604 hash-table nodes to avoid having to allocate memory unnecessarily, and to make
605 merging alias sets extremely efficient (the linked list merge is constant time).
606 </p>
608 <p>You shouldn't need to understand these details if you are just a client of
609 the AliasSetTracker, but if you look at the code, hopefully this brief
610 description will help make sense of why things are designed the way they
611 are.</p>
613 </div>
615 <!-- ======================================================================= -->
618 </div>
622 <p>If neither of these utility class are what your pass needs, you should use
624 the higher-level methods when possible (e.g., use mod/ref information instead of
626 best precision and efficiency.</p>
628 </div>
630 <!-- *********************************************************************** -->
633 </div>
634 <!-- *********************************************************************** -->
638 <p>If you're going to be working with the LLVM alias analysis infrastructure,
639 you should know what clients and implementations of alias analysis are
640 available. In particular, if you are implementing an alias analysis, you should
642 for monitoring and evaluating different implementations.</p>
644 </div>
646 <!-- ======================================================================= -->
649 </div>
658 </div>
660 <!-- _______________________________________________________________________ -->
663 </div>
668 never returns any useful information. This pass can be useful if you think that
669 alias analysis is doing something wrong and are trying to narrow down a
670 problem.</p>
672 </div>
674 <!-- _______________________________________________________________________ -->
677 </div>
684 <ul>
685 <li>Distinct globals, stack allocations, and heap allocations can never
686 alias.</li>
687 <li>Globals, stack allocations, and heap allocations never alias the null
688 pointer.</li>
693 <li>Pointers that obviously point to constant globals
695 <li>Function calls can not modify or references stack allocations if they never
696 escape from the function that allocates them (a common case for automatic
697 arrays).</li>
698 </ul>
700 </div>
702 <!-- _______________________________________________________________________ -->
705 </div>
709 <p>This pass implements a simple context-sensitive mod/ref and alias analysis
710 for internal global variables that don't "have their address taken". If a
711 global does not have its address taken, the pass knows that no pointers alias
712 the global. This pass also keeps track of functions that it knows never access
713 memory or never read memory. This allows certain optimizations (e.g. GCSE) to
714 eliminate call instructions entirely.
715 </p>
717 <p>The real power of this pass is that it provides context-sensitive mod/ref
718 information for call instructions. This allows the optimizer to know that
719 calls to a function do not clobber or read the value of the global, allowing
720 loads and stores to be eliminated.</p>
722 <p>Note that this pass is somewhat limited in its scope (only support
723 non-address taken globals), but is very quick analysis.</p>
724 </div>
726 <!-- _______________________________________________________________________ -->
729 </div>
734 for interprocedural alias analysis. This algorithm is a subset-based,
735 flow-insensitive, context-insensitive, and field-insensitive alias analysis that
736 is widely believed to be fairly precise. Unfortunately, this algorithm is also
739 substitution") to improve its efficiency, so it can be quite slow in common
740 cases.
741 </p>
743 </div>
745 <!-- _______________________________________________________________________ -->
748 </div>
753 "Steensgaard's algorithm" for interprocedural alias analysis. Steensgaard's
754 algorithm is a unification-based, flow-insensitive, context-insensitive, and
755 field-insensitive alias analysis that is also very scalable (effectively linear
756 time).</p>
759 field-<b>sensitive</b>" version of Steensgaard's algorithm using the Data
760 Structure Analysis framework. This gives it substantially more precision than
761 the standard algorithm while maintaining excellent analysis scalability.</p>
764 module, it is not part of the LLVM core.</p>
766 </div>
768 <!-- _______________________________________________________________________ -->
771 </div>
776 algorithm. Data Structure Analysis is a modular unification-based,
779 O(n*log(n)).</p>
781 <p>This algorithm is capable of responding to a full variety of alias analysis
782 queries, and can provide context-sensitive mod/ref information as well. The
783 only major facility not implemented so far is support for must-alias
784 information.</p>
787 module, it is not part of the LLVM core.</p>
789 </div>
792 <!-- ======================================================================= -->
795 </div>
798 LLVM includes several alias-analysis driven transformations which can be used
799 with any of the implementations above.
800 </div>
802 <!-- _______________________________________________________________________ -->
805 </div>
811 not have side-effects and are not used.</p>
813 </div>
816 <!-- _______________________________________________________________________ -->
819 </div>
825 different transformations:</p>
827 <ul>
828 <li>It uses mod/ref information to hoist or sink load instructions out of loops
829 if there are no instructions in the loop that modifies the memory loaded.</li>
831 <li>It uses mod/ref information to hoist function calls out of loops that do not
832 write to memory and are loop-invariant.</li>
834 <li>If uses alias information to promote memory objects that are loaded and
835 stored to in loops to live in a register instead. It can do this if there are
836 no may aliases to the loaded/stored memory location.</li>
837 </ul>
839 </div>
841 <!-- _______________________________________________________________________ -->
844 </div>
847 <p>
849 by-value instead. In particular, if pointer arguments are only loaded from it
850 passes in the value loaded instead of the address to the function. This pass
851 uses alias information to make sure that the value loaded from the argument
852 pointer is not modified between the entry of the function and any load of the
853 pointer.</p>
854 </div>
856 <!-- _______________________________________________________________________ -->
859 </div>
864 number</a>" loads and pointers values, which is used by the GCSE pass to
866 and must-alias information. This combination of passes can make the following
867 transformations:</p>
869 <ul>
871 <li>Load instructions that follow a store to the same location are replaced with
873 <li>Pointers values (e.g. formal arguments) that must-alias simpler expressions
874 (e.g. global variables or the null pointer) are replaced. Note that this
875 implements transformations like "virtual method resolution", turning indirect
876 calls into direct calls.</li>
877 </ul>
879 </div>
881 <!-- ======================================================================= -->
884 implementations</a>
885 </div>
889 <p>These passes are useful for evaluating the various alias analysis
890 implementations. You can use them with commands like '<tt>opt -anders-aa -ds-aa
893 </div>
895 <!-- _______________________________________________________________________ -->
898 </div>
908 <pre>
909 % opt -ds-aa -print-alias-sets -disable-output
910 </pre>
911 </div>
913 </div>
916 <!-- _______________________________________________________________________ -->
919 </div>
924 pass is making and what responses are returned by the alias analysis. As an
925 example,</p>
928 <pre>
929 % opt -basicaa -count-aa -ds-aa -count-aa -licm
930 </pre>
931 </div>
933 <p>will print out how many queries (and what responses are returned) by the
936 when debugging a transformation or an alias analysis implementation.</p>
938 </div>
940 <!-- _______________________________________________________________________ -->
943 </div>
948 function and asks an alias analysis whether or not the pointers alias. This
949 gives an indication of the precision of the alias analysis. Statistics are
950 printed indicating the percent of no/may/must aliases found (a more precise
951 algorithm will have a lower number of may aliases).</p>
953 </div>
955 <!-- *********************************************************************** -->
958 </div>
959 <!-- *********************************************************************** -->
963 <p>If you're just looking to be a client of alias analysis information, consider
964 using the Memory Dependence Analysis interface instead. MemDep is a lazy,
965 caching layer on top of alias analysis that is able to answer the question of
966 what preceding memory operations a given instruction depends on, either at an
967 intra- or inter-block level. Because of its laziness and caching
968 policy, using MemDep can be a significant performance win over accessing alias
969 analysis directly.</p>
971 </div>
973 <!-- *********************************************************************** -->
975 <hr>
976 <address>
984 Last modified: $Date$
985 </address>
987 </body>
988 </html>