| blob | b4aaafac508a1c3cb05bb43edcc7d975132e5260 |
3 <html>
4 <head>
7 </head>
8 <body>
11 LLVM Programmer's Manual
12 </div>
14 <ol>
17 <ul>
19 <!--
20 <li>The <tt>-time-passes</tt> option</li>
21 <li>How to use the LLVM Makefile system</li>
22 <li>How to write a regression test</li>
24 -->
25 </ul>
26 </li>
28 <ul>
32 option</a>
33 <ul>
36 </ul>
37 </li>
39 option</a></li>
40 <!--
41 <li>The <tt>InstVisitor</tt> template
42 <li>The general graph API
43 -->
45 </ul>
46 </li>
48 <ul>
50 <ul>
59 </ul></li>
61 <ul>
71 </ul></li>
73 <ul>
80 </ul></li>
82 <ul>
85 </ul></li>
86 </ul>
87 </li>
89 <ul>
91 <ul>
99 class pointer</a> </li>
101 complex example</a> </li>
103 the same way</a> </li>
105 use-def chains</a> </li>
107 successors of blocks</a></li>
108 </ul>
109 </li>
111 <ul>
118 </ul>
119 </li>
121 <!--
122 <li>Working with the Control Flow Graph
123 <ul>
124 <li>Accessing predecessors and successors of a <tt>BasicBlock</tt>
125 <li>
126 <li>
127 </ul>
128 -->
129 </ul>
130 </li>
133 <ul>
135 <ul>
140 </ul></li>
142 <li><a href="#SymbolTable">The <tt>ValueSymbolTable</tt> and <tt>TypeSymbolTable</tt> classes</a></li>
143 <li><a href="#UserLayout">The <tt>User</tt> and owned <tt>Use</tt> classes' memory layout</a></li>
144 </ul></li>
147 <ul>
151 <ul>
153 <ul>
156 <ul>
158 <ul>
161 </ul>
162 </li>
163 </ul>
164 </li>
165 </ul>
166 </li>
169 </ul>
170 </li>
171 </ul>
172 </li>
173 </ol>
181 </div>
183 <!-- *********************************************************************** -->
186 </div>
187 <!-- *********************************************************************** -->
191 <p>This document is meant to highlight some of the important classes and
192 interfaces available in the LLVM source-base. This manual is not
193 intended to explain what LLVM is, how it works, and what LLVM code looks
194 like. It assumes that you know the basics of LLVM and are interested
195 in writing transformations or otherwise analyzing or manipulating the
196 code.</p>
198 <p>This document should get you oriented so that you can find your
199 way in the continuously growing source code that makes up the LLVM
200 infrastructure. Note that this manual is not intended to serve as a
201 replacement for reading the source code, so if you think there should be
202 a method in one of these classes to do something, but it's not listed,
204 are provided to make this as easy as possible.</p>
206 <p>The first section of this document describes general information that is
207 useful to know when working in the LLVM infrastructure, and the second describes
208 the Core LLVM classes. In the future this manual will be extended with
209 information describing how to use extension libraries, such as dominator
210 information, CFG traversal routines, and useful utilities like the <tt><a
213 </div>
215 <!-- *********************************************************************** -->
218 </div>
219 <!-- *********************************************************************** -->
223 <p>This section contains general information that is useful if you are working
224 in the LLVM source-base, but that isn't specific to any particular API.</p>
226 </div>
228 <!-- ======================================================================= -->
231 </div>
235 <p>LLVM makes heavy use of the C++ Standard Template Library (STL),
236 perhaps much more than you are used to, or have seen before. Because of
237 this, you might want to do a little background reading in the
238 techniques used and capabilities of the library. There are many good
239 pages that discuss the STL, and several books on the subject that you
240 can get, so it will not be discussed in this document.</p>
244 <ol>
247 reference</a> - an excellent reference for the STL and other parts of the
248 standard C++ library.</li>
251 O'Reilly book in the making. It has a decent Standard Library
252 Reference that rivals Dinkumware's, and is unfortunately no longer free since the
253 book has been published.</li>
256 Questions</a></li>
259 Contains a useful <a
264 Page</a></li>
268 the book).</a></li>
270 </ol>
274 to write maintainable code more than where to put your curly braces.</p>
276 </div>
278 <!-- ======================================================================= -->
281 </div>
285 <ol>
287 Branch and Tag Primer</a></li>
289 static and shared libraries across platforms</a></li>
290 </ol>
292 </div>
294 <!-- *********************************************************************** -->
297 </div>
298 <!-- *********************************************************************** -->
302 <p>Here we highlight some LLVM APIs that are generally useful and good to
303 know about when writing transformations.</p>
305 </div>
307 <!-- ======================================================================= -->
311 </div>
315 <p>The LLVM source-base makes extensive use of a custom form of RTTI.
317 operator, but they don't have some drawbacks (primarily stemming from
319 have a v-table). Because they are used so often, you must know what they
320 do and how they work. All of these templates are defined in the <a
322 file (note that you very rarely have to include this file directly).</p>
324 <dl>
328 "<tt>instanceof</tt>" operator. It returns true or false depending on whether
329 a reference or pointer points to an instance of the specified class. This can
330 be very useful for constraint checking of various sorts (example below).</p>
331 </dd>
336 converts a pointer or reference from a base class to a derived class, causing
337 an assertion failure if it is not really an instance of the right type. This
338 should be used in cases where you have some information that makes you believe
343 <pre>
345 if (isa<<a href="#Constant">Constant</a>>(V) || isa<<a href="#Argument">Argument</a>>(V) || isa<<a href="#GlobalValue">GlobalValue</a>>(V))
346 return true;
350 }
351 </pre>
352 </div>
356 operator.</p>
358 </dd>
363 It checks to see if the operand is of the specified type, and if so, returns a
364 pointer to it (this operator does not work with references). If the operand is
365 not of the correct type, a null pointer is returned. Thus, this works very
369 statement like this:</p>
372 <pre>
373 if (<a href="#AllocationInst">AllocationInst</a> *AI = dyn_cast<<a href="#AllocationInst">AllocationInst</a>>(Val)) {
375 }
376 </pre>
377 </div>
381 statement, which is very convenient.</p>
386 blocks to check for lots of different variants of classes. If you find
387 yourself wanting to do this, it is much cleaner and more efficient to use the
390 </dd>
396 argument (which it then propagates). This can sometimes be useful, allowing
397 you to combine several null checks into one.</p></dd>
403 as an argument (which it then propagates). This can sometimes be useful,
404 allowing you to combine several null checks into one.</p></dd>
406 </dl>
408 <p>These five templates can be used with any classes, whether they have a
409 v-table or not. To add support for these templates, you simply need to add
411 to. Describing this is currently outside the scope of this document, but there
412 are lots of examples in the LLVM source base.</p>
414 </div>
416 <!-- ======================================================================= -->
419 </div>
423 <p>Often when working on your pass you will put a bunch of debugging printouts
424 and other code into your pass. After you get it working, you want to remove
425 it, but you may need it again in the future (to work out new bugs that you run
426 across).</p>
428 <p> Naturally, because of this, you don't want to delete the debug printouts,
429 but you don't want them to always be noisy. A standard compromise is to comment
430 them out, allowing you to enable them if you need them in the future.</p>
434 this problem. Basically, you can put arbitrary code into the argument of the
439 <pre>
441 </pre>
442 </div>
447 <pre>
451 I am here!
452 </pre>
453 </div>
456 to not have to create "yet another" command line option for the debug output for
458 so they do not cause a performance impact at all (for the same reason, they
459 should also not contain side-effects!).</p>
462 enable or disable it directly in gdb. Just use "<tt>set DebugFlag=0</tt>" or
463 "<tt>set DebugFlag=1</tt>" from the gdb if the program is running. If the
464 program hasn't been started yet, you can always just run it with
467 </div>
469 <!-- _______________________________________________________________________ -->
473 </div>
479 generator). If you want to enable debug information with more fine-grained
481 option as follows:</p>
484 <pre>
486 #undef DEBUG_TYPE
487 #define DEBUG_TYPE "foo"
489 #undef DEBUG_TYPE
490 #define DEBUG_TYPE "bar"
492 #undef DEBUG_TYPE
493 #define DEBUG_TYPE ""
495 </pre>
496 </div>
501 <pre>
505 No debug type
506 'foo' debug type
507 'bar' debug type
508 No debug type (2)
510 'foo' debug type
512 'bar' debug type
513 </pre>
514 </div>
517 a file, to specify the debug type for the entire module (if you do this before
520 "bar", because there is no system in place to ensure that names do not
521 conflict. If two different modules use the same string, they will all be turned
522 on when the name is specified. This allows, for example, all debug information
524 even if the source lives in multiple files.</p>
526 </div>
528 <!-- ======================================================================= -->
531 option</a>
532 </div>
539 keep track of what the LLVM compiler is doing and how effective various
540 optimizations are. It is useful to see what optimizations are contributing to
541 making a particular program run faster.</p>
543 <p>Often you may run your pass on some big program, and you're interested to see
544 how many times it makes a certain transformation. Although you can do this with
545 hand inspection, or some ad-hoc method, this is a real pain and not very useful
547 keep track of this information, and the calculated information is presented in a
548 uniform manner with the rest of the passes being executed.</p>
551 it are as follows:</p>
553 <ol>
557 <pre>
558 #define <a href="#DEBUG_TYPE">DEBUG_TYPE</a> "mypassname" <i>// This goes before any #includes.</i>
559 STATISTIC(NumXForms, "The # of times I did stuff");
560 </pre>
561 </div>
564 specified by the first argument. The pass name is taken from the DEBUG_TYPE
565 macro, and the description is taken from the second argument. The variable
571 <pre>
573 </pre>
574 </div>
576 </li>
577 </ol>
583 <pre>
586 </pre>
587 </div>
590 suite, it gives a report that looks like this:</p>
593 <pre>
594 7646 bitcodewriter - Number of normal instructions
595 725 bitcodewriter - Number of oversized instructions
596 129996 bitcodewriter - Number of bitcode bytes written
597 2817 raise - Number of insts DCEd or constprop'd
598 3213 raise - Number of cast-of-self removed
599 5046 raise - Number of expression trees converted
600 75 raise - Number of other getelementptr's formed
601 138 raise - Number of load/store peepholes
602 42 deadtypeelim - Number of unused typenames removed from symtab
603 392 funcresolve - Number of varargs functions resolved
604 27 globaldce - Number of global variables removed
605 2 adce - Number of basic blocks removed
606 134 cee - Number of branches revectored
607 49 cee - Number of setcc instruction eliminated
608 532 gcse - Number of loads removed
609 2919 gcse - Number of instructions removed
610 86 indvars - Number of canonical indvars added
611 87 indvars - Number of aux indvars removed
612 25 instcombine - Number of dead inst eliminate
613 434 instcombine - Number of insts combined
614 248 licm - Number of load insts hoisted
615 1298 licm - Number of insts hoisted to a loop pre-header
616 3 licm - Number of insts hoisted to multiple loop preds (bad, no loop pre-header)
617 75 mem2reg - Number of alloca's promoted
618 1444 cfgsimplify - Number of blocks simplified
619 </pre>
620 </div>
622 <p>Obviously, with so many optimizations, having a unified framework for this
623 stuff is very nice. Making your pass fit well into the framework makes it more
624 maintainable and useful.</p>
626 </div>
628 <!-- ======================================================================= -->
631 </div>
635 <p>Several of the important data structures in LLVM are graphs: for example
639 DAGs</a>. In many cases, while debugging various parts of the compiler, it is
640 nice to instantly visualize these graphs.</p>
642 <p>LLVM provides several callbacks that are available in a debug build to do
644 the current LLVM tool will pop up a window containing the CFG for the function
645 where each basic block is a node in the graph, and each node contains the
646 instructions in the block. Similarly, there also exists
651 up a window. Alternatively, you can sprinkle calls to these functions in your
652 code in places you want to debug.</p>
654 <p>Getting this to work requires a small amount of configuration. On Unix
656 toolkit, and make sure 'dot' and 'gv' are in your path. If you are running on
657 Mac OS/X, download and install the Mac OS/X <a
660 it) to your path. Once in your system and path are set up, rerun the LLVM
661 configure script and rebuild LLVM to enable this functionality.</p>
667 specified color (choices of colors can be found at <a
669 complex node attributes can be provided with <tt>call
672 Attributes</a>.) If you want to restart and clear all the current graph
675 </div>
677 <!-- *********************************************************************** -->
680 </div>
681 <!-- *********************************************************************** -->
686 and we commonly use STL data structures. This section describes the trade-offs
687 you should consider when you pick one.</p>
689 <p>
690 The first step is a choose your own adventure: do you want a sequential
691 container, a set-like container, or a map-like container? The most important
692 thing when choosing a container is the algorithmic properties of how you plan to
693 access the container. Based on that, you should use:</p>
695 <ul>
697 of an value based on another value. Map-like containers also support
698 efficient queries for containment (whether a key is in the map). Map-like
699 containers generally do not support efficient reverse mapping (values to
700 keys). If you need that, use two maps. Some map-like containers also
701 support efficient iteration through the keys in sorted order. Map-like
702 containers are the most expensive sort, only use them if you need one of
703 these capabilities.</li>
706 stuff into a container that automatically eliminates duplicates. Some
707 set-like containers support efficient iteration through the elements in
708 sorted order. Set-like containers are more expensive than sequential
709 containers.
710 </li>
713 the most efficient way to add elements and keeps track of the order they are
714 added to the collection. They permit duplicates and support efficient
715 iteration, but do not support efficient look-up based on a key.
716 </li>
719 perform set operations on sets of numeric id's, while automatically
720 eliminating duplicates. Bit containers require a maximum of 1 bit for each
721 identifier you want to store.
722 </li>
723 </ul>
725 <p>
726 Once the proper category of container is determined, you can fine tune the
727 memory use, constant factors, and cache behaviors of access by intelligently
728 picking a member of the category. Note that constant factors and cache behavior
729 can be a big deal. If you have a vector that usually only contains a few
730 elements (but could contain many), for example, it's much better to use
732 . Doing so avoids (relatively) expensive malloc/free calls, which dwarf the
733 cost of adding the elements to the container. </p>
735 </div>
737 <!-- ======================================================================= -->
740 </div>
743 There are a variety of sequential containers available for you, based on your
744 needs. Pick the first in this section that will do what you want.
745 </div>
747 <!-- _______________________________________________________________________ -->
750 </div>
753 <p>Fixed size arrays are very simple and very fast. They are good if you know
754 exactly how many elements you have, or you have a (low) upper bound on how many
755 you have.</p>
756 </div>
758 <!-- _______________________________________________________________________ -->
761 </div>
764 <p>Heap allocated arrays (new[] + delete[]) are also simple. They are good if
765 the number of elements is variable, if you know how many elements you will need
766 before the array is allocated, and if the array is usually large (if not,
768 allocated array is the cost of the new/delete (aka malloc/free). Also note that
769 if you are allocating an array of a type with a constructor, the constructor and
770 destructors will be run for every element in the array (re-sizable vectors only
771 construct those elements actually used).</p>
772 </div>
774 <!-- _______________________________________________________________________ -->
777 </div>
782 it supports efficient iteration, lays out elements in memory order (so you can
783 do pointer arithmetic between elements), supports efficient push_back/pop_back
784 operations, supports efficient random access to its elements, etc.</p>
786 <p>The advantage of SmallVector is that it allocates space for
788 the SmallVector is dynamically smaller than N, no malloc is performed. This can
789 be a big win in cases where the malloc/free call is far more expensive than the
790 code that fiddles around with the elements.</p>
793 predecessors/successors of a block is usually less than 8). On the other hand,
794 this makes the size of the SmallVector itself large, so you don't want to
795 allocate lots of them (doing so will waste a lot of space). As such,
796 SmallVectors are most useful when on the stack.</p>
798 <p>SmallVector also provides a nice portable and efficient replacement for
801 </div>
803 <!-- _______________________________________________________________________ -->
806 </div>
809 <p>
810 std::vector is well loved and respected. It is useful when SmallVector isn't:
811 when the size of the vector is often large (thus the small optimization will
812 rarely be a benefit) or if you will be allocating many instances of the vector
813 itself (which would waste space for elements that aren't in the container).
814 vector is also useful when interfacing with code that expects vectors :).
815 </p>
820 <pre>
821 for ( ... ) {
823 use V;
824 }
825 </pre>
826 </div>
831 <pre>
833 for ( ... ) {
834 use V;
835 V.clear();
836 }
837 </pre>
838 </div>
840 <p>Doing so will save (at least) one heap allocation and free per iteration of
841 the loop.</p>
843 </div>
845 <!-- _______________________________________________________________________ -->
848 </div>
851 <p>std::deque is, in some senses, a generalized version of std::vector. Like
852 std::vector, it provides constant time random access and other similar
853 properties, but it also provides efficient access to the front of the list. It
854 does not guarantee continuity of elements within memory.</p>
856 <p>In exchange for this extra flexibility, std::deque has significantly higher
857 constant factor costs than std::vector. If possible, use std::vector or
858 something cheaper.</p>
859 </div>
861 <!-- _______________________________________________________________________ -->
864 </div>
867 <p>std::list is an extremely inefficient class that is rarely useful.
868 It performs a heap allocation for every element inserted into it, thus having an
869 extremely high constant factor, particularly for small data types. std::list
870 also only supports bidirectional iteration, not random access iteration.</p>
872 <p>In exchange for this high cost, std::list supports efficient access to both
873 ends of the list (like std::deque, but unlike std::vector or SmallVector). In
874 addition, the iterator invalidation characteristics of std::list are stronger
875 than that of a vector class: inserting or removing an element into the list does
876 not invalidate iterator or pointers to other elements in the list.</p>
877 </div>
879 <!-- _______________________________________________________________________ -->
882 </div>
886 intrusive, because it requires the element to store and provide access to the
887 prev/next pointers for the list.</p>
891 provides some novel characteristics. In particular, it can efficiently store
892 polymorphic objects, the traits class is informed when an element is inserted or
894 constant-time splice operation.</p>
896 <p>These properties are exactly what we want for things like
900 Related classes of interest are explained in the following subsections:
901 <ul>
906 </ul>
907 </div>
909 <!-- _______________________________________________________________________ -->
912 </div>
917 publicly derive from this traits class.</p>
918 </div>
920 <!-- _______________________________________________________________________ -->
923 </div>
927 supports a slightly narrower interface. Notably, inserters from
931 used for a wide variety of customizations.</p>
932 </div>
934 <!-- _______________________________________________________________________ -->
937 </div>
942 in the default manner.</p>
947 </div>
949 <!-- _______________________________________________________________________ -->
952 </div>
956 citizen in the C++ ecosystem, it needs to support the standard container
961 <p>The only sensible solution to this problem is to allocate a so-called
963 iterator, providing the back-link to the last element. However conforming to the
965 also must not be dereferenced.</p>
968 how to allocate and store the sentinel. The corresponding policy is dictated
970 whenever the need for a sentinel arises.</p>
972 <p>While the default policy is sufficient in most cases, it may break down when
975 is wasted. To alleviate the situation with numerous and voluminous
981 arithmetic is used to obtain the sentinel, which is relative to the
983 extra pointer, which serves as the back-link of the sentinel. This is the only
984 field in the ghostly sentinel which can be legally accessed.</p>
985 </div>
987 <!-- _______________________________________________________________________ -->
990 </div>
995 <p>There are also various STL adapter classes such as std::queue,
996 std::priority_queue, std::stack, etc. These provide simplified access to an
997 underlying container but don't affect the cost of the container itself.</p>
999 </div>
1002 <!-- ======================================================================= -->
1005 </div>
1009 <p>Set-like containers are useful when you need to canonicalize multiple values
1010 into a single representation. There are several different choices for how to do
1011 this, providing various trade-offs.</p>
1013 </div>
1016 <!-- _______________________________________________________________________ -->
1019 </div>
1023 <p>If you intend to insert a lot of elements, then do a lot of queries, a
1024 great approach is to use a vector (or other sequential container) with
1025 std::sort+std::unique to remove duplicates. This approach works really well if
1026 your usage pattern has these two distinct phases (insert then query), and can be
1028 </p>
1030 <p>
1031 This combination provides the several nice properties: the result data is
1032 contiguous in memory (good for cache locality), has few allocations, is easy to
1033 address (iterators in the final vector are just indices or pointers), and can be
1034 efficiently queried with a standard binary or radix search.</p>
1036 </div>
1038 <!-- _______________________________________________________________________ -->
1041 </div>
1045 <p>If you have a set-like data structure that is usually small and whose elements
1047 has space for N elements in place (thus, if the set is dynamically smaller than
1048 N, no malloc traffic is required) and accesses them with a simple linear search.
1049 When the set grows beyond 'N' elements, it allocates a more expensive representation that
1050 guarantees efficient access (for most types, it falls back to std::set, but for
1051 pointers it uses something far better, <a
1054 <p>The magic of this class is that it handles small sets extremely efficiently,
1055 but gracefully handles extremely large sets without loss of efficiency. The
1056 drawback is that the interface is quite small: it supports insertion, queries
1057 and erasing, but does not support iteration.</p>
1059 </div>
1061 <!-- _______________________________________________________________________ -->
1064 </div>
1068 <p>SmallPtrSet has all the advantages of SmallSet (and a SmallSet of pointers is
1069 transparently implemented with a SmallPtrSet), but also supports iterators. If
1070 more than 'N' insertions are performed, a single quadratically
1071 probed hash table is allocated and grows as needed, providing extremely
1072 efficient access (constant time insertion/deleting/queries with low constant
1073 factors) and is very stingy with malloc traffic.</p>
1075 <p>Note that, unlike std::set, the iterators of SmallPtrSet are invalidated
1076 whenever an insertion occurs. Also, the values visited by the iterators are not
1077 visited in sorted order.</p>
1079 </div>
1081 <!-- _______________________________________________________________________ -->
1084 </div>
1088 <p>
1089 DenseSet is a simple quadratically probed hash table. It excels at supporting
1090 small values: it uses a single allocation to hold all of the pairs that
1091 are currently inserted in the set. DenseSet is a great way to unique small
1092 values that are not simple pointers (use <a
1094 the same requirements for the value type that <a
1096 </p>
1098 </div>
1100 <!-- _______________________________________________________________________ -->
1103 </div>
1107 <p>
1108 FoldingSet is an aggregate class that is really good at uniquing
1109 expensive-to-create or polymorphic objects. It is a combination of a chained
1110 hash table with intrusive links (uniqued objects are required to inherit from
1112 its ID process.</p>
1115 a complex object (for example, a node in the code generator). The client has a
1116 description of *what* it wants to generate (it knows the opcode and all the
1117 operands), but we don't want to 'new' a node, then try inserting it into a set
1118 only to find out it already exists, at which point we would have to delete it
1119 and return the node that already exists.
1120 </p>
1122 <p>To support this style of client, FoldingSet perform a query with a
1123 FoldingSetNodeID (which wraps SmallVector) that can be used to describe the
1124 element that we want to query for. The query either returns the element
1125 matching the ID or it returns an opaque ID that indicates where insertion should
1126 take place. Construction of the ID usually does not require heap traffic.</p>
1128 <p>Because FoldingSet uses intrusive links, it can support polymorphic objects
1129 in the set (for example, you can have SDNode instances mixed with LoadSDNodes).
1130 Because the elements are individually allocated, pointers to the elements are
1131 stable: inserting or removing elements does not invalidate any pointers to other
1132 elements.
1133 </p>
1135 </div>
1137 <!-- _______________________________________________________________________ -->
1140 </div>
1145 many things but great at nothing. std::set allocates memory for each element
1146 inserted (thus it is very malloc intensive) and typically stores three pointers
1147 per element in the set (thus adding a large amount of per-element space
1148 overhead). It offers guaranteed log(n) performance, which is not particularly
1149 fast from a complexity standpoint (particularly if the elements of the set are
1150 expensive to compare, like strings), and has extremely high constant factors for
1151 lookup, insertion and removal.</p>
1153 <p>The advantages of std::set are that its iterators are stable (deleting or
1154 inserting an element from the set does not affect iterators or pointers to other
1155 elements) and that iteration over the set is guaranteed to be in sorted order.
1156 If the elements in the set are large, then the relative overhead of the pointers
1157 and malloc traffic is not a big deal, but if the elements of the set are small,
1158 std::set is almost never a good choice.</p>
1160 </div>
1162 <!-- _______________________________________________________________________ -->
1165 </div>
1170 Container</a>. The important property
1171 that this provides is efficient insertion with uniquing (duplicate elements are
1172 ignored) with iteration support. It implements this by inserting elements into
1173 both a set-like container and the sequential container, using the set-like
1174 container for uniquing and the sequential container for iteration.
1175 </p>
1177 <p>The difference between SetVector and other sets is that the order of
1178 iteration is guaranteed to match the order of insertion into the SetVector.
1179 This property is really important for things like sets of pointers. Because
1180 pointer values are non-deterministic (e.g. vary across runs of the program on
1181 different machines), iterating over the pointers in the set will
1182 not be in a well-defined order.</p>
1184 <p>
1185 The drawback of SetVector is that it requires twice as much space as a normal
1186 set and has the sum of constant factors from the set-like container and the
1187 sequential container that it uses. Use it *only* if you need to iterate over
1188 the elements in a deterministic order. SetVector is also expensive to delete
1189 elements out of (linear time), unless you use it's "pop_back" method, which is
1190 faster.
1191 </p>
1193 <p>SetVector is an adapter class that defaults to using std::vector and std::set
1194 for the underlying containers, so it is quite expensive. However,
1196 defaults to using a SmallVector and SmallSet of a specified size. If you use
1197 this, and if your sets are dynamically smaller than N, you will save a lot of
1198 heap traffic.</p>
1200 </div>
1202 <!-- _______________________________________________________________________ -->
1205 </div>
1209 <p>
1211 retains a unique ID for each element inserted into the set. It internally
1212 contains a map and a vector, and it assigns a unique ID for each value inserted
1213 into the set.</p>
1215 <p>UniqueVector is very expensive: its cost is the sum of the cost of
1216 maintaining both the map and vector, it has high complexity, high constant
1217 factors, and produces a lot of malloc traffic. It should be avoided.</p>
1219 </div>
1222 <!-- _______________________________________________________________________ -->
1225 </div>
1229 <p>
1230 The STL provides several other options, such as std::multiset and the various
1231 "hash_set" like containers (whether from C++ TR1 or from the SGI library). We
1232 never use hash_set and unordered_set because they are generally very expensive
1233 (each insertion requires a malloc) and very non-portable.
1234 </p>
1236 <p>std::multiset is useful if you're not interested in elimination of
1237 duplicates, but has all the drawbacks of std::set. A sorted vector (where you
1238 don't delete duplicate entries) or some other approach is almost always
1239 better.</p>
1241 </div>
1243 <!-- ======================================================================= -->
1246 </div>
1249 Map-like containers are useful when you want to associate data to a key. As
1250 usual, there are a lot of different ways to do this. :)
1251 </div>
1253 <!-- _______________________________________________________________________ -->
1256 </div>
1260 <p>
1261 If your usage pattern follows a strict insert-then-query approach, you can
1263 for set-like containers</a>. The only difference is that your query function
1264 (which uses std::lower_bound to get efficient log(n) lookup) should only compare
1265 the key, not both the key and value. This yields the same advantages as sorted
1266 vectors for sets.
1267 </p>
1268 </div>
1270 <!-- _______________________________________________________________________ -->
1273 </div>
1277 <p>
1278 Strings are commonly used as keys in maps, and they are difficult to support
1279 efficiently: they are variable length, inefficient to hash and compare when
1280 long, expensive to copy, etc. StringMap is a specialized container designed to
1281 cope with these issues. It supports mapping an arbitrary range of bytes to an
1282 arbitrary other object.</p>
1284 <p>The StringMap implementation uses a quadratically-probed hash table, where
1285 the buckets store a pointer to the heap allocated entries (and some other
1286 stuff). The entries in the map must be heap allocated because the strings are
1287 variable length. The string data (key) and the element object (value) are
1288 stored in the same allocation with the string data immediately after the element
1289 object. This container guarantees the "<tt>(char*)(&Value+1)</tt>" points
1290 to the key string for a value.</p>
1292 <p>The StringMap is very fast for several reasons: quadratic probing is very
1293 cache efficient for lookups, the hash value of strings in buckets is not
1294 recomputed when lookup up an element, StringMap rarely has to touch the
1295 memory for unrelated objects when looking up a value (even when hash collisions
1296 happen), hash table growth does not recompute the hash values for strings
1297 already in the table, and each pair in the map is store in a single allocation
1298 (the string data is stored in the same allocation as the Value of a pair).</p>
1300 <p>StringMap also provides query methods that take byte ranges, so it only ever
1301 copies a string if a value is inserted into the table.</p>
1302 </div>
1304 <!-- _______________________________________________________________________ -->
1307 </div>
1310 <p>
1311 IndexedMap is a specialized container for mapping small dense integers (or
1312 values that can be mapped to small dense integers) to some other type. It is
1313 internally implemented as a vector with a mapping function that maps the keys to
1314 the dense integer range.
1315 </p>
1317 <p>
1318 This is useful for cases like virtual registers in the LLVM code generator: they
1319 have a dense mapping that is offset by a compile-time constant (the first
1320 virtual register ID).</p>
1322 </div>
1324 <!-- _______________________________________________________________________ -->
1327 </div>
1331 <p>
1332 DenseMap is a simple quadratically probed hash table. It excels at supporting
1333 small keys and values: it uses a single allocation to hold all of the pairs that
1334 are currently inserted in the map. DenseMap is a great way to map pointers to
1335 pointers, or map other small types to each other.
1336 </p>
1338 <p>
1339 There are several aspects of DenseMap that you should be aware of, however. The
1340 iterators in a densemap are invalidated whenever an insertion occurs, unlike
1341 map. Also, because DenseMap allocates space for a large number of key/value
1342 pairs (it starts with 64 by default), it will waste a lot of space if your keys
1343 or values are large. Finally, you must implement a partial specialization of
1344 DenseMapInfo for the key that you want, if it isn't already supported. This
1345 is required to tell DenseMap about two special marker values (which can never be
1346 inserted into the map) that it needs internally.</p>
1348 </div>
1350 <!-- _______________________________________________________________________ -->
1353 </div>
1357 <p>
1359 a single allocation per pair inserted into the map, it offers log(n) lookup with
1360 an extremely large constant factor, imposes a space penalty of 3 pointers per
1361 pair in the map, etc.</p>
1363 <p>std::map is most useful when your keys or values are very large, if you need
1364 to iterate over the collection in sorted order, or if you need stable iterators
1365 into the map (i.e. they don't get invalidated if an insertion or deletion of
1366 another element takes place).</p>
1368 </div>
1370 <!-- _______________________________________________________________________ -->
1373 </div>
1377 <p>
1378 The STL provides several other options, such as std::multimap and the various
1379 "hash_map" like containers (whether from C++ TR1 or from the SGI library). We
1380 never use hash_set and unordered_set because they are generally very expensive
1381 (each insertion requires a malloc) and very non-portable.</p>
1383 <p>std::multimap is useful if you want to map a key to multiple values, but has
1384 all the drawbacks of std::map. A sorted vector or some other approach is almost
1385 always better.</p>
1387 </div>
1389 <!-- ======================================================================= -->
1392 </div>
1395 <p>Unlike the other containers, there are only two bit storage containers, and
1396 choosing when to use each is relatively straightforward.</p>
1398 <p>One additional option is
1400 implementation in many common compilers (e.g. commonly available versions of
1401 GCC) is extremely inefficient and 2) the C++ standards committee is likely to
1402 deprecate this container and/or change it significantly somehow. In any case,
1403 please don't use it.</p>
1404 </div>
1406 <!-- _______________________________________________________________________ -->
1409 </div>
1412 <p> The BitVector container provides a fixed size set of bits for manipulation.
1413 It supports individual bit setting/testing, as well as set operations. The set
1414 operations take time O(size of bitvector), but operations are performed one word
1415 at a time, instead of one bit at a time. This makes the BitVector very fast for
1416 set operations compared to other containers. Use the BitVector when you expect
1417 the number of set bits to be high (IE a dense set).
1418 </p>
1419 </div>
1421 <!-- _______________________________________________________________________ -->
1424 </div>
1427 <p> The SparseBitVector container is much like BitVector, with one major
1428 difference: Only the bits that are set, are stored. This makes the
1429 SparseBitVector much more space efficient than BitVector when the set is sparse,
1430 as well as making set operations O(number of set bits) instead of O(size of
1431 universe). The downside to the SparseBitVector is that setting and testing of random bits is O(N), and on large SparseBitVectors, this can be slower than BitVector. In our implementation, setting or testing bits in sorted order
1432 (either forwards or reverse) is O(1) worst case. Testing and setting bits within 128 bits (depends on size) of the current bit is also O(1). As a general statement, testing/setting bits in a SparseBitVector is O(distance away from last set bit).
1433 </p>
1434 </div>
1436 <!-- *********************************************************************** -->
1439 </div>
1440 <!-- *********************************************************************** -->
1444 <p>This section describes how to perform some very simple transformations of
1445 LLVM code. This is meant to give examples of common idioms used, showing the
1447 you should also read about the main classes that you will be working with. The
1449 and descriptions of the main classes that you should know about.</p>
1451 </div>
1453 <!-- NOTE: this section should be heavy on example code -->
1454 <!-- ======================================================================= -->
1457 </div>
1461 <p>The LLVM compiler infrastructure have many different data structures that may
1462 be traversed. Following the example of the C++ standard template library, the
1463 techniques used to traverse these various data structures are all basically the
1466 function returns an iterator pointing to one past the last valid element of the
1468 between the two operations.</p>
1470 <p>Because the pattern for iteration is common across many different aspects of
1471 the program representation, the standard template library algorithms may be used
1472 on them, and it is easier to remember how to iterate. First we show a few common
1473 examples of the data structures that need to be traversed. Other data
1474 structures are traversed in very similar ways.</p>
1476 </div>
1478 <!-- _______________________________________________________________________ -->
1483 </div>
1488 transform in some way; in particular, you'd like to manipulate its
1495 <pre>
1502 </pre>
1503 </div>
1505 <p>Note that i can be used as if it were a pointer for the purposes of
1507 because the indirection operator is overloaded for the iterator
1511 </div>
1513 <!-- _______________________________________________________________________ -->
1518 </div>
1523 easy to iterate over the individual instructions that make up
1528 <pre>
1534 </pre>
1535 </div>
1537 <p>However, this isn't really the best way to print out the contents of a
1539 anything you'll care about, you could have just invoked the print routine on the
1542 </div>
1544 <!-- _______________________________________________________________________ -->
1549 </div>
1558 small example that shows how to dump all instructions in a function to the standard error stream:<p>
1561 <pre>
1565 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
1567 </pre>
1568 </div>
1571 work list with its initial contents. For example, if you wanted to
1573 F, all you would need to do is something like:</p>
1576 <pre>
1580 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
1581 worklist.insert(&*I);
1582 </pre>
1583 </div>
1588 </div>
1590 <!-- _______________________________________________________________________ -->
1593 vice-versa)</a>
1594 </div>
1598 <p>Sometimes, it'll be useful to grab a reference (or pointer) to a class
1599 instance when all you've got at hand is an iterator. Well, extracting
1600 a reference or a pointer from an iterator is very straight-forward.
1605 <pre>
1608 const Instruction& inst = *j;
1609 </pre>
1610 </div>
1612 <p>However, the iterators you'll be working with in the LLVM framework are
1613 special: they will automatically convert to a ptr-to-instance type whenever they
1614 need to. Instead of dereferencing the iterator and then taking the address of
1615 the result, you can simply assign the iterator to the proper pointer type and
1616 you get the dereference and address-of operation as a result of the assignment
1617 (behind the scenes, this is a result of overloading casting mechanisms). Thus
1618 the last line of the last example,</p>
1621 <pre>
1622 Instruction *pinst = &*i;
1623 </pre>
1624 </div>
1629 <pre>
1630 Instruction *pinst = i;
1631 </pre>
1632 </div>
1634 <p>It's also possible to turn a class pointer into the corresponding iterator,
1635 and this is a constant time operation (very efficient). The following code
1636 snippet illustrates use of the conversion constructors provided by LLVM
1637 iterators. By using these, you can explicitly grab the iterator of something
1638 without actually obtaining it via iteration over some structure:</p>
1641 <pre>
1642 void printNextInstruction(Instruction* inst) {
1643 BasicBlock::iterator it(inst);
1646 }
1647 </pre>
1648 </div>
1650 </div>
1652 <!--_______________________________________________________________________-->
1655 example</a>
1656 </div>
1660 <p>Say that you're writing a FunctionPass and would like to count all the
1664 much more straight-forward manner, but this example will allow us to explore how
1666 is what we want to do:</p>
1669 <pre>
1670 initialize callCounter to zero
1671 for each Function f in the Module
1672 for each BasicBlock b in f
1673 for each Instruction i in b
1674 if (i is a CallInst and calls the given function)
1675 increment callCounter
1676 </pre>
1677 </div>
1679 <p>And the actual code is (remember, because we're writing a
1684 <pre>
1685 Function* targetFunc = ...;
1687 class OurFunctionPass : public FunctionPass {
1688 public:
1689 OurFunctionPass(): callCounter(0) { }
1691 virtual runOnFunction(Function& F) {
1692 for (Function::iterator b = F.begin(), be = F.end(); b != be; ++b) {
1699 if (callInst->getCalledFunction() == targetFunc)
1700 ++callCounter;
1701 }
1702 }
1703 }
1704 }
1706 private:
1707 unsigned callCounter;
1708 };
1709 </pre>
1710 </div>
1712 </div>
1714 <!--_______________________________________________________________________-->
1717 </div>
1721 <p>You may have noticed that the previous example was a bit oversimplified in
1722 that it did not deal with call sites generated by 'invoke' instructions. In
1723 this, and in other situations, you may find that you want to treat
1726 less closely-related things. For these cases, LLVM provides a handy wrapper
1727 class called <a
1734 reference and it should not be dynamically allocated or deallocated using
1736 assignable and constructable, with costs equivalents to that of a bare pointer.
1737 If you look at its definition, it has only a single pointer member.</p>
1739 </div>
1741 <!--_______________________________________________________________________-->
1744 </div>
1758 <pre>
1759 Function *F = ...;
1765 }
1766 </pre>
1767 </div>
1774 all of the values that a particular instruction uses (that is, the operands of
1778 <pre>
1779 Instruction *pi = ...;
1782 Value *v = *i;
1784 }
1785 </pre>
1786 </div>
1788 <!--
1789 def-use chains ("finding all users of"): Value::use_begin/use_end
1790 use-def chains ("finding all values used"): User::op_begin/op_end [op=operand]
1791 -->
1793 </div>
1795 <!--_______________________________________________________________________-->
1798 successors of blocks</a>
1799 </div>
1803 <p>Iterating over the predecessors and successors of a block is quite easy
1805 this to iterate over all predecessors of BB:</p>
1808 <pre>
1809 #include "llvm/Support/CFG.h"
1810 BasicBlock *BB = ...;
1812 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1813 BasicBlock *Pred = *PI;
1815 }
1816 </pre>
1817 </div>
1819 <p>Similarly, to iterate over successors use
1820 succ_iterator/succ_begin/succ_end.</p>
1822 </div>
1825 <!-- ======================================================================= -->
1828 </div>
1832 <p>There are some primitive transformation operations present in the LLVM
1833 infrastructure that are worth knowing about. When performing
1834 transformations, it's fairly common to manipulate the contents of basic
1835 blocks. This section describes some of the common methods for doing so
1836 and gives example code.</p>
1838 </div>
1840 <!--_______________________________________________________________________-->
1844 </div>
1851 constructor for the kind of instruction to instantiate and provide the necessary
1856 <pre>
1857 AllocaInst* ai = new AllocaInst(Type::Int32Ty);
1858 </pre>
1859 </div>
1863 subclass is likely to have varying default parameters which change the semantics
1864 of the instruction, so refer to the <a
1870 <p>It is very useful to name the values of instructions when you're able to, as
1871 this facilitates the debugging of your transformations. If you end up looking
1872 at generated LLVM machine code, you definitely want to have logical names
1873 associated with the results of instructions! By supplying a value for the
1875 associate a logical name with the result of the instruction's execution at
1876 run time. For example, say that I'm writing a transformation that dynamically
1877 allocates space for an integer on the stack, and that integer is going to be
1878 used as some kind of index by some other code. To accomplish this, I place an
1884 <pre>
1886 </pre>
1887 </div>
1890 execution value, which is a pointer to an integer on the run time stack.</p>
1897 <ul>
1898 <li>Insertion into an explicit instruction list
1905 <pre>
1906 BasicBlock *pb = ...;
1907 Instruction *pi = ...;
1908 Instruction *newInst = new Instruction(...);
1911 </pre>
1912 </div>
1916 classes provide constructors which take a pointer to a
1918 looked like: </p>
1921 <pre>
1922 BasicBlock *pb = ...;
1923 Instruction *newInst = new Instruction(...);
1926 </pre>
1927 </div>
1932 <pre>
1933 BasicBlock *pb = ...;
1934 Instruction *newInst = new Instruction(..., pb);
1935 </pre>
1936 </div>
1938 <p>which is much cleaner, especially if you are creating
1939 long instruction streams.</p></li>
1941 <li>Insertion into an implicit instruction list
1944 are implicitly associated with an existing instruction list: the instruction
1945 list of the enclosing basic block. Thus, we could have accomplished the same
1947 </p>
1950 <pre>
1951 Instruction *pi = ...;
1952 Instruction *newInst = new Instruction(...);
1955 </pre>
1956 </div>
1958 <p>In fact, this sequence of steps occurs so frequently that the
1960 constructors which take (as a default parameter) a pointer to an
1964 provided instruction, immediately before that instruction. Using an
1966 parameter, the above code becomes:</p>
1969 <pre>
1970 Instruction* pi = ...;
1971 Instruction* newInst = new Instruction(..., pi);
1972 </pre>
1973 </div>
1975 <p>which is much cleaner, especially if you're creating a lot of
1977 </ul>
1979 </div>
1981 <!--_______________________________________________________________________-->
1984 </div>
1988 <p>Deleting an instruction from an existing sequence of instructions that form a
1990 you must have a pointer to the instruction that you wish to delete. Second, you
1991 need to obtain the pointer to that instruction's basic block. You use the
1992 pointer to the basic block to get its list of instructions and then use the
1993 erase function to remove your instruction. For example:</p>
1996 <pre>
1998 I->eraseFromParent();
1999 </pre>
2000 </div>
2002 </div>
2004 <!--_______________________________________________________________________-->
2008 </div>
2014 <p>Including "<a href="/doxygen/BasicBlockUtils_8h-source.html">llvm/Transforms/Utils/BasicBlockUtils.h</a>"
2020 <ul>
2023 <p>This function replaces all uses of a given instruction with a value,
2024 and then removes the original instruction. The following example
2025 illustrates the replacement of the result of a particular
2027 pointer to an integer.</p>
2030 <pre>
2031 AllocaInst* instToReplace = ...;
2032 BasicBlock::iterator ii(instToReplace);
2035 Constant::getNullValue(PointerType::getUnqual(Type::Int32Ty)));
2036 </pre></div></li>
2040 <p>This function replaces a particular instruction with another
2041 instruction, inserting the new instruction into the basic block at the
2042 location where the old instruction was, and replacing any uses of the old
2043 instruction with the new instruction. The following example illustrates
2047 <pre>
2048 AllocaInst* instToReplace = ...;
2049 BasicBlock::iterator ii(instToReplace);
2053 </pre></div></li>
2054 </ul>
2062 information.</p>
2064 <!-- Value::replaceAllUsesWith User::replaceUsesOfWith Point out:
2065 include/llvm/Transforms/Utils/ especially BasicBlockUtils.h with:
2066 ReplaceInstWithValue, ReplaceInstWithInst -->
2068 </div>
2070 <!--_______________________________________________________________________-->
2073 </div>
2077 <p>Deleting a global variable from a module is just as easy as deleting an
2078 Instruction. First, you must have a pointer to the global variable that you wish
2079 to delete. You use this pointer to erase it from its parent, the module.
2080 For example:</p>
2083 <pre>
2086 GV->eraseFromParent();
2087 </pre>
2088 </div>
2090 </div>
2092 <!-- ======================================================================= -->
2095 </div>
2099 <p>In generating IR, you may need some complex types. If you know these types
2102 has two forms depending on whether you're building types for cross-compilation
2105 out of types from
2107 namespace and pointers, functions, arrays, etc. built of
2109 whose size may depend on the host compiler. For example,</p>
2112 <pre>
2114 </pre>
2115 </div>
2120 <pre>
2122 params.push_back(PointerType::getUnqual(Type::Int32Ty));
2123 FunctionType *ft = FunctionType::get(Type::Int8Ty, params, false);
2124 </pre>
2125 </div>
2130 </div>
2132 <!-- *********************************************************************** -->
2135 </div>
2136 <!-- *********************************************************************** -->
2139 <p>
2140 This section describes some of the advanced or obscure API's that most clients
2141 do not need to be aware of. These API's tend manage the inner workings of the
2142 LLVM system, and only need to be accessed in unusual circumstances.
2143 </p>
2144 </div>
2146 <!-- ======================================================================= -->
2149 </div>
2153 <p>
2154 The LLVM type system has a very simple goal: allow clients to compare types for
2155 structural equality with a simple pointer comparison (aka a shallow compare).
2156 This goal makes clients much simpler and faster, and is used throughout the LLVM
2157 system.
2158 </p>
2160 <p>
2161 Unfortunately achieving this goal is not a simple matter. In particular,
2162 recursive types and late resolution of opaque types makes the situation very
2163 difficult to handle. Fortunately, for the most part, our implementation makes
2164 most clients able to be completely unaware of the nasty internal details. The
2165 primary case where clients are exposed to the inner workings of it are when
2166 building a recursive type. In addition to this case, the LLVM bitcode reader,
2167 assembly parser, and linker also have to be aware of the inner workings of this
2168 system.
2169 </p>
2171 <p>
2172 For our purposes below, we need three concepts. First, an "Opaque Type" is
2175 opaque type as part of its type graph (for example "<tt>{ opaque, i32 }</tt>").
2176 Third, a concrete type is a type that is not an abstract type (e.g. "<tt>{ i32,
2177 float }</tt>").
2178 </p>
2180 </div>
2182 <!-- ______________________________________________________________________ -->
2185 </div>
2189 <p>
2190 Because the most common question is "how do I build a recursive type with LLVM",
2191 we answer it now and explain it as we go. Here we include enough to cause this
2192 to be emitted to an output .ll file:
2193 </p>
2196 <pre>
2197 %mylist = type { %mylist*, i32 }
2198 </pre>
2199 </div>
2201 <p>
2202 To build this, use the following LLVM APIs:
2203 </p>
2206 <pre>
2210 Elts.push_back(PointerType::getUnqual(StructTy));
2211 Elts.push_back(Type::Int32Ty);
2212 StructType *NewSTy = StructType::get(Elts);
2216 cast<OpaqueType>(StructTy.get())-><a href="#refineAbstractTypeTo">refineAbstractTypeTo</a>(NewSTy);
2218 // <i>NewSTy is potentially invalidated, but StructTy (a <a href="#PATypeHolder">PATypeHolder</a>) is</i>
2224 </pre>
2225 </div>
2227 <p>
2228 This code shows the basic approach used to build recursive types: build a
2229 non-recursive type using 'opaque', then use type unification to close the cycle.
2230 The type unification step is performed by the <tt><a
2232 described next. After that, we describe the <a
2234 </p>
2236 </div>
2238 <!-- ______________________________________________________________________ -->
2241 </div>
2244 <p>
2246 While this method is actually a member of the DerivedType class, it is most
2247 often used on OpaqueType instances. Type unification is actually a recursive
2248 process. After unification, types can become structurally isomorphic to
2249 existing types, and all duplicates are deleted (to preserve pointer equality).
2250 </p>
2252 <p>
2253 In the example above, the OpaqueType object is definitely deleted.
2254 Additionally, if there is an "{ \2*, i32}" type already created in the system,
2256 a type is deleted, any "Type*" pointers in the program are invalidated. As
2259 types can never move or be deleted). To deal with this, the <a
2261 reference to a possibly refined type, and the <a
2263 complex datastructures.
2264 </p>
2266 </div>
2268 <!-- ______________________________________________________________________ -->
2271 </div>
2274 <p>
2275 PATypeHolder is a form of a "smart pointer" for Type objects. When VMCore
2276 happily goes about nuking types that become isomorphic to existing types, it
2277 automatically updates all PATypeHolder objects to point to the new type. In the
2278 example above, this allows the code to maintain a pointer to the resultant
2279 resolved recursive type, even though the Type*'s are potentially invalidated.
2280 </p>
2282 <p>
2283 PATypeHolder is an extremely light-weight object that uses a lazy union-find
2284 implementation to update pointers. For example the pointer from a Value to its
2285 Type is maintained by PATypeHolder objects.
2286 </p>
2288 </div>
2290 <!-- ______________________________________________________________________ -->
2293 </div>
2297 <p>
2298 Some data structures need more to perform more complex updates when types get
2299 resolved. To support this, a class can derive from the AbstractTypeUser class.
2300 This class
2301 allows it to get callbacks when certain types are resolved. To register to get
2302 callbacks for a particular type, the DerivedType::{add/remove}AbstractTypeUser
2303 methods can be called on a type. Note that these methods only work for <i>
2304 abstract</i> types. Concrete types (those that do not include any opaque
2305 objects) can never be refined.
2306 </p>
2307 </div>
2310 <!-- ======================================================================= -->
2314 </div>
2324 names for types.</p>
2327 by most clients. It should only be used when iteration over the symbol table
2328 names themselves are required, which is very special purpose. Note that not
2329 all LLVM
2331 an empty name) do not exist in the symbol table.
2332 </p>
2334 <p>These symbol tables support iteration over the values/types in the symbol
2339 appropriate symbol table. For types, use the Module::addTypeName method to
2340 insert entries into the symbol table.</p>
2342 </div>
2346 <!-- ======================================================================= -->
2349 </div>
2357 addition and removal.</p>
2359 <!-- ______________________________________________________________________ -->
2361 <a name="Use2User">Interaction and relationship between <tt>User</tt> and <tt>Use</tt> objects</a>
2362 </div>
2365 <p>
2367 or refer to them out-of-line by means of a pointer. A mixed variant
2370 </p>
2371 </div>
2373 <p>
2375 <ul>
2376 <li><p>Layout a)
2378 object and there are a fixed number of them.</p>
2380 <li><p>Layout b)
2383 number of them.</p>
2384 </ul>
2385 <p>
2386 As of v2.4 each layout still possesses a direct pointer to the
2388 we stick to this redundancy for the sake of simplicity.
2390 has. (Theoretically this information can also be calculated
2391 given the scheme presented below.)</p>
2392 <p>
2394 enforce the following memory layouts:</p>
2396 <ul>
2397 <li><p>Layout a) is modelled by prepending the <tt>User</tt> object by the <tt>Use[]</tt> array.</p>
2399 <pre>
2400 ...---.---.---.---.-------...
2401 | P | P | P | P | User
2402 '''---'---'---'---'-------'''
2403 </pre>
2406 <pre>
2407 .-------...
2408 | User
2409 '-------'''
2410 |
2411 v
2412 .---.---.---.---...
2413 | P | P | P | P |
2414 '---'---'---'---'''
2415 </pre>
2416 </ul>
2420 <!-- ______________________________________________________________________ -->
2423 </div>
2426 <p>
2429 recover it. This is accomplished by the following scheme:</p>
2430 </div>
2432 A bit-encoding in the 2 LSBits (least significant bits) of the <tt>Use::Prev</tt> allows to find the
2434 <ul>
2439 </ul>
2440 <p>
2443 we have to walk to the next stop picking up digits
2444 and calculating the offset:</p>
2445 <pre>
2446 .---.---.---.---.---.---.---.---.---.---.---.---.---.---.---.---.----------------
2448 '---'---'---'---'---'---'---'---'---'---'---'---'---'---'---'---'----------------
2450 | | | | |__>
2451 | | | |__________>
2452 | | |______________________>
2453 | |______________________________________>
2454 |__________________________________________________________>
2455 </pre>
2456 <p>
2457 Only the significant number of bits need to be stored between the
2461 <!-- ______________________________________________________________________ -->
2464 </div>
2467 <p>
2468 The following literate Haskell fragment demonstrates the concept:</p>
2469 </div>
2472 <pre>
2473 > import Test.QuickCheck
2474 >
2479 >
2485 >
2486 > takeLast n ss = reverse $ take n $ reverse ss
2487 >
2489 >
2490 </pre>
2491 </div>
2492 <p>
2494 <p>
2495 The reverse algorithm computes the length of the string just by examining
2496 a certain prefix:</p>
2499 <pre>
2504 >
2507 > decode walk acc _ = walk + acc
2508 >
2509 </pre>
2510 </div>
2511 <p>
2513 <p>
2517 <pre>
2519 > testcaseLength = length testcase
2520 >
2522 > where arr = takeLast n testcase
2523 >
2524 </pre>
2525 </div>
2526 <p>
2529 <pre>
2530 *Main> quickCheck identityProp
2531 OK, passed 100 tests.
2532 </pre>
2533 <p>
2534 Let's be a bit more exhaustive:</p>
2537 <pre>
2538 >
2540 >
2541 </pre>
2542 </div>
2543 <p>
2546 <pre>
2547 *Main> deepCheck identityProp
2548 OK, passed 500 tests.
2549 </pre>
2551 <!-- ______________________________________________________________________ -->
2554 </div>
2556 <p>
2560 tag bits.</p>
2561 <p>
2565 the LSBit set. (Portability is relying on the fact that all known compilers place the
2568 </div>
2570 <!-- *********************************************************************** -->
2573 </div>
2574 <!-- *********************************************************************** -->
2580 <p>The Core LLVM classes are the primary means of representing the program
2581 being inspected or transformed. The core LLVM classes are defined in
2585 </div>
2587 <!-- ======================================================================= -->
2590 </div>
2598 subclasses. They are hidden because they offer no useful functionality beyond
2602 named, but this is not a requirement. There exists exactly
2603 one instance of a given shape at any one time. This allows type equality to
2604 be performed with address equality of the Type Instance. That is, given two
2606 </p>
2607 </div>
2609 <!-- _______________________________________________________________________ -->
2612 </div>
2616 <ul>
2620 floating point types.</li>
2623 an OpaqueType anywhere in its definition).</li>
2626 that don't have a size are abstract types, labels and void.</li>
2628 </ul>
2629 </div>
2631 <!-- _______________________________________________________________________ -->
2634 </div>
2636 <dl>
2638 <dd>Subclass of DerivedType that represents integer types of any bit width.
2641 <ul>
2643 type of a specific bit width.</li>
2645 type.</li>
2646 </ul>
2647 </dd>
2649 <dd>This is subclassed by ArrayType and PointerType
2650 <ul>
2652 of the elements in the sequential type. </li>
2653 </ul>
2654 </dd>
2656 <dd>This is a subclass of SequentialType and defines the interface for array
2657 types.
2658 <ul>
2660 elements in the array. </li>
2661 </ul>
2662 </dd>
2666 <dd>Subclass of SequentialType for vector types. A
2667 vector type is similar to an ArrayType but is distinguished because it is
2668 a first class type wherease ArrayType is not. Vector types are used for
2669 vector operations and are usually small vectors of of an integer or floating
2670 point type.</dd>
2674 <dd>Subclass of DerivedTypes for function types.
2675 <ul>
2677 function</li>
2679 return type of the function.</li>
2681 the type of the ith parameter.</li>
2683 number of formal parameters.</li>
2684 </ul>
2685 </dd>
2687 <dd>Sublcass of DerivedType for abstract types. This class
2688 defines no content and is used as a placeholder for some other type. Note
2689 that OpaqueType is used (temporarily) during type resolution for forward
2690 references of types. Once the referenced type is resolved, the OpaqueType
2691 is replaced with the actual type. OpaqueType can also be used for data
2692 abstraction. At link time opaque types can be resolved to actual types
2693 of the same name.</dd>
2694 </dl>
2695 </div>
2699 <!-- ======================================================================= -->
2702 </div>
2711 programs. An LLVM module is effectively either a translation unit of the
2712 original program or a combination of several translation units merged by the
2717 helpful member functions that try to make common operations easy.</p>
2719 </div>
2721 <!-- _______________________________________________________________________ -->
2724 </div>
2728 <ul>
2730 </ul>
2733 provide a name for it (probably based on the name of the translation unit).</p>
2735 <ul>
2742 <p>These are forwarding methods that make it easy to access the contents of
2744 list.</p></li>
2749 necessary to use when you need to update the list or perform a complex
2750 action that doesn't have a forwarding method.</p>
2753 </ul>
2755 <hr>
2757 <ul>
2765 <p> These are forwarding methods that make it easy to access the contents of
2773 use when you need to update the list or perform a complex action that
2774 doesn't have a forwarding method.</p>
2777 </ul>
2779 <hr>
2781 <ul>
2788 </ul>
2790 <hr>
2792 <ul>
2805 external declaration for the function and return it.</p></li>
2812 string.</p></li>
2819 name, true is returned and the <a
2821 </ul>
2823 </div>
2826 <!-- ======================================================================= -->
2829 </div>
2834 <br>
2838 base. It represents a typed value that may be used (among other things) as an
2846 for a program. For example, an incoming argument to a function (represented
2848 every instruction in the function that references the argument. To keep track
2854 methods, shown below.</p>
2858 method. In addition, all LLVM values can be named. The "name" of the
2862 <pre>
2864 </pre>
2865 </div>
2868 that the name of any value may be missing (an empty string), so names should
2870 debugging printouts), they should not be used to keep track of values or map
2874 <p>One important aspect of LLVM is that there is no distinction between an SSA
2875 variable and the operation that produces it. Because of this, any reference to
2876 the value produced by an instruction (or the value available as an incoming
2877 argument, for example) is represented as a direct pointer to the instance of
2878 the class that
2879 represents this value. Although this may take some getting used to, it
2880 simplifies the representation and makes it easier to manipulate.</p>
2882 </div>
2884 <!-- _______________________________________________________________________ -->
2887 </div>
2891 <ul>
2893 use-list<br>
2895 the use-list<br>
2897 value.<br>
2900 the use-list.<br>
2902 use-list.<br>
2904 element in the list.
2905 <p> These methods are the interface to access the def-use
2906 information in LLVM. As with all other iterators in LLVM, the naming
2908 </li>
2911 </li>
2917 </li>
2922 "<tt>V</tt>" instead. For example, if you detect that an instruction always
2923 produces a constant value (for example through constant folding), you can
2924 replace all uses of the instruction with the constant like this:</p>
2927 <pre>
2928 Inst->replaceAllUsesWith(ConstVal);
2929 </pre>
2930 </div>
2932 </ul>
2934 </div>
2936 <!-- ======================================================================= -->
2939 </div>
2943 <p>
2956 Single Assignment (SSA) form, there can only be one definition referred to,
2957 allowing this direct connection. This connection provides the use-def
2958 information in LLVM.</p>
2960 </div>
2962 <!-- _______________________________________________________________________ -->
2965 </div>
2970 an index access interface and through an iterator based interface.</p>
2972 <ul>
2976 convenient form for direct access.</p></li>
2979 list<br>
2981 the operand list.<br>
2983 operand list.
2984 <p> Together, these methods make up the iterator based interface to
2986 </ul>
2988 </div>
2990 <!-- ======================================================================= -->
2993 </div>
3004 instructions. It provides only a few methods, but is a very commonly used
3006 opcode (instruction type) and the parent <a
3008 into. To represent a specific type of instruction, one of many subclasses of
3017 file contains some meta-data about the various different types of instructions
3018 in LLVM. It describes the enum values that are used as opcodes (for example
3023 this file confuses doxygen, so these enum values don't show up correctly in the
3026 </div>
3028 <!-- _______________________________________________________________________ -->
3031 class</a>
3032 </div>
3034 <ul>
3036 <p>This subclasses represents all two operand instructions whose operands
3037 must be the same type, except for the comparison instructions.</p></li>
3040 common operations on cast instructions.</p>
3042 <p>This subclass respresents the two comparison instructions,
3046 <p>This subclass is the parent of all terminator instructions (those which
3047 can terminate a block).</p>
3048 </ul>
3049 </div>
3051 <!-- _______________________________________________________________________ -->
3054 class</a>
3055 </div>
3059 <ul>
3064 <p>Returns true if the instruction writes to memory, i.e. it is a
3069 <p>Returns another instance of the specified instruction, identical
3070 in all ways to the original except that the instruction has no parent
3072 and it has no name</p></li>
3073 </ul>
3075 </div>
3077 <!-- ======================================================================= -->
3080 </div>
3084 <p>Constant represents a base class for different types of constants. It
3085 is subclassed by ConstantInt, ConstantArray, etc. for representing
3087 a subclass, which represents the address of a global variable or function.
3088 </p>
3090 </div>
3092 <!-- _______________________________________________________________________ -->
3095 <ul>
3096 <li>ConstantInt : This subclass of Constant represents an integer constant of
3097 any width.
3098 <ul>
3100 value of this constant, an APInt value.</li>
3102 value to an int64_t via sign extension. If the value (not the bit width)
3103 of the APInt is too large to fit in an int64_t, an assertion will result.
3104 For this reason, use of this method is discouraged.</li>
3106 value to a uint64_t via zero extension. IF the value (not the bit width)
3107 of the APInt is too large to fit in a uint64_t, an assertion will result.
3108 For this reason, use of this method is discouraged.</li>
3111 The type is implied as the IntegerType that corresponds to the bit width
3114 Returns the ConstantInt object that represents the value provided by
3116 </ul>
3117 </li>
3118 <li>ConstantFP : This class represents a floating point constant.
3119 <ul>
3121 this constant. </li>
3122 </ul>
3123 </li>
3124 <li>ConstantArray : This represents a constant array.
3125 <ul>
3127 a vector of component constants that makeup this array. </li>
3128 </ul>
3129 </li>
3130 <li>ConstantStruct : This represents a constant struct.
3131 <ul>
3133 a vector of component constants that makeup this array. </li>
3134 </ul>
3135 </li>
3136 <li>GlobalValue : This represents either a global variable or a function. In
3137 either case, the value is a constant fixed address (after linking).
3138 </li>
3139 </ul>
3140 </div>
3143 <!-- ======================================================================= -->
3146 </div>
3153 Class</a><br>
3160 Because they are visible at global scope, they are also subject to linking with
3161 other globals defined in different translation units. To control the linking
3168 unit, and does not participate in linking. If it has external linkage, it is
3169 visible to external code, and does participate in linking. In addition to
3175 global is always a pointer to its contents. It is important to remember this
3180 the address of the first element of this array and the value of the
3188 </div>
3190 <!-- _______________________________________________________________________ -->
3193 class</a>
3194 </div>
3198 <ul>
3203 <p> </p>
3204 </li>
3207 GlobalValue is currently embedded into.</p></li>
3208 </ul>
3210 </div>
3212 <!-- ======================================================================= -->
3215 </div>
3228 actually one of the more complex classes in the LLVM heirarchy because it must
3236 ordering of the blocks in the function, which indicate how the code will be
3237 layed out by the backend. Additionally, the first <a
3240 block. There are no implicit exit nodes, and in fact there may be multiple exit
3244 function hasn't been linked in yet.</p>
3254 LLVM feature that is only used when you have to look up a value by name. Aside
3256 internally to make sure that there are not conflicts between the names of <a
3263 is its address (after linking) which is guaranteed to be constant.</p>
3264 </div>
3266 <!-- _______________________________________________________________________ -->
3269 class</a>
3270 </div>
3274 <ul>
3279 the the program. The constructor must specify the type of the function to
3280 create and what type of linkage the function should have. The <a
3282 specifies the formal arguments and return value for the function. The same
3285 in which the function is defined. If this argument is provided, the function
3286 will automatically be inserted into that module's list of
3287 functions.</p></li>
3292 function is "external", it does not have a body, and thus must be resolved
3293 by linking with a function defined in a different translation unit.</p></li>
3301 <p>These are forwarding methods that make it easy to access the contents of
3303 list.</p></li>
3308 is necessary to use when you need to update the list or perform a complex
3309 action that doesn't have a forwarding method.</p></li>
3312 iterator<br>
3318 <p>These are forwarding methods that make it easy to access the contents of
3320 list.</p></li>
3325 necessary to use when you need to update the list or perform a complex
3326 action that doesn't have a forwarding method.</p></li>
3331 function. Because the entry block for the function is always the first
3340 function.</p></li>
3346 </ul>
3348 </div>
3350 <!-- ======================================================================= -->
3353 </div>
3359 <br>
3361 Class</a><br>
3367 <p>Global variables are represented with the (suprise suprise)
3370 always referenced by their address (global values must live in memory, so their
3371 "name" refers to their constant address). See
3373 variables may have an initial value (which must be a
3375 they may be marked as "constant" themselves (indicating that their contents
3376 never change at runtime).</p>
3377 </div>
3379 <!-- _______________________________________________________________________ -->
3383 </div>
3387 <ul>
3392 <p>Create a new global variable of the specified type. If
3394 unchanging for the program. The Linkage parameter specifies the type of
3395 linkage (internal, external, weak, linkonce, appending) for the variable.
3396 If the linkage is InternalLinkage, WeakAnyLinkage, WeakODRLinkage,
3397 LinkOnceAnyLinkage or LinkOnceODRLinkage, then the resultant
3398 global variable will have internal linkage. AppendingLinkage concatenates
3399 together all instances (in different translation units) of the variable
3400 into a single variable but is only applicable to arrays. See
3402 further details on linkage types. Optionally an initializer, a name, and the
3403 module to put the variable into may be specified for the global variable as
3404 well.</p></li>
3408 <p>Returns true if this is a global variable that is known not to
3409 be modified at runtime.</p></li>
3418 to call this method if there is no initializer.</p></li>
3419 </ul>
3421 </div>
3424 <!-- ======================================================================= -->
3427 </div>
3434 Class</a><br>
3437 <p>This class represents a single entry multiple exit section of the code,
3438 commonly known as a basic block by the compiler community. The
3441 Matching the language definition, the last element of this list of instructions
3442 is always a terminator instruction (a subclass of the <a
3445 <p>In addition to tracking the list of instructions that make up the block, the
3454 </div>
3456 <!-- _______________________________________________________________________ -->
3459 class</a>
3460 </div>
3463 <ul>
3469 insertion into a function. The constructor optionally takes a name for the new
3472 automatically inserted at the end of the specified <a
3480 STL-style functions for accessing the instruction list.
3482 <p>These methods and typedefs are forwarding functions that have the same
3483 semantics as the standard library methods of the same names. These methods
3484 expose the underlying instruction list of a basic block in a way that is easy to
3485 manipulate. To get the full complement of container operations (including
3487 method.</p></li>
3491 <p>This method is used to get access to the underlying container that actually
3492 holds the Instructions. This method must be used when there isn't a forwarding
3494 to perform. Because there are no forwarding functions for "updating"
3495 operations, you need to use this if you want to update the contents of a
3501 embedded into, or a null pointer if it is homeless.</p></li>
3505 <p> Returns a pointer to the terminator instruction that appears at the end of
3507 instruction in the block is not a terminator, then a null pointer is
3508 returned.</p></li>
3510 </ul>
3512 </div>
3515 <!-- ======================================================================= -->
3518 </div>
3522 <p>This subclass of Value defines the interface for incoming formal
3523 arguments to a function. A Function maintains a list of its formal
3524 arguments. An argument has a pointer to the parent Function.</p>
3526 </div>
3528 <!-- *********************************************************************** -->
3529 <hr>
3530 <address>
3539 Last modified: $Date$
3540 </address>
3542 </body>
3543 </html>