| blob | e2525ae84e0e5e766f170fa63d3d8a9b379290c6 |
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>
120 <!--
121 <li>Working with the Control Flow Graph
122 <ul>
123 <li>Accessing predecessors and successors of a <tt>BasicBlock</tt>
124 <li>
125 <li>
126 </ul>
127 -->
128 </ul>
129 </li>
132 <ul>
134 <ul>
139 </ul></li>
141 <li><a href="#SymbolTable">The <tt>ValueSymbolTable</tt> and <tt>TypeSymbolTable</tt> classes</a></li>
142 <li><a href="#UserLayout">The <tt>User</tt> and owned <tt>Use</tt> classes' memory layout</a></li>
143 </ul></li>
146 <ul>
150 <ul>
152 <ul>
155 <ul>
157 <ul>
160 </ul>
161 </li>
162 </ul>
163 </li>
164 </ul>
165 </li>
168 </ul>
169 </li>
170 </ul>
171 </li>
172 </ol>
180 </div>
182 <!-- *********************************************************************** -->
185 </div>
186 <!-- *********************************************************************** -->
190 <p>This document is meant to highlight some of the important classes and
191 interfaces available in the LLVM source-base. This manual is not
192 intended to explain what LLVM is, how it works, and what LLVM code looks
193 like. It assumes that you know the basics of LLVM and are interested
194 in writing transformations or otherwise analyzing or manipulating the
195 code.</p>
197 <p>This document should get you oriented so that you can find your
198 way in the continuously growing source code that makes up the LLVM
199 infrastructure. Note that this manual is not intended to serve as a
200 replacement for reading the source code, so if you think there should be
201 a method in one of these classes to do something, but it's not listed,
203 are provided to make this as easy as possible.</p>
205 <p>The first section of this document describes general information that is
206 useful to know when working in the LLVM infrastructure, and the second describes
207 the Core LLVM classes. In the future this manual will be extended with
208 information describing how to use extension libraries, such as dominator
209 information, CFG traversal routines, and useful utilities like the <tt><a
212 </div>
214 <!-- *********************************************************************** -->
217 </div>
218 <!-- *********************************************************************** -->
222 <p>This section contains general information that is useful if you are working
223 in the LLVM source-base, but that isn't specific to any particular API.</p>
225 </div>
227 <!-- ======================================================================= -->
230 </div>
234 <p>LLVM makes heavy use of the C++ Standard Template Library (STL),
235 perhaps much more than you are used to, or have seen before. Because of
236 this, you might want to do a little background reading in the
237 techniques used and capabilities of the library. There are many good
238 pages that discuss the STL, and several books on the subject that you
239 can get, so it will not be discussed in this document.</p>
243 <ol>
246 reference</a> - an excellent reference for the STL and other parts of the
247 standard C++ library.</li>
250 O'Reilly book in the making. It has a decent Standard Library
251 Reference that rivals Dinkumware's, and is unfortunately no longer free since the
252 book has been published.</li>
255 Questions</a></li>
258 Contains a useful <a
263 Page</a></li>
267 the book).</a></li>
269 </ol>
273 to write maintainable code more than where to put your curly braces.</p>
275 </div>
277 <!-- ======================================================================= -->
280 </div>
284 <ol>
286 Branch and Tag Primer</a></li>
288 static and shared libraries across platforms</a></li>
289 </ol>
291 </div>
293 <!-- *********************************************************************** -->
296 </div>
297 <!-- *********************************************************************** -->
301 <p>Here we highlight some LLVM APIs that are generally useful and good to
302 know about when writing transformations.</p>
304 </div>
306 <!-- ======================================================================= -->
310 </div>
314 <p>The LLVM source-base makes extensive use of a custom form of RTTI.
316 operator, but they don't have some drawbacks (primarily stemming from
318 have a v-table). Because they are used so often, you must know what they
319 do and how they work. All of these templates are defined in the <a
321 file (note that you very rarely have to include this file directly).</p>
323 <dl>
327 "<tt>instanceof</tt>" operator. It returns true or false depending on whether
328 a reference or pointer points to an instance of the specified class. This can
329 be very useful for constraint checking of various sorts (example below).</p>
330 </dd>
335 converts a pointer or reference from a base class to a derived class, causing
336 an assertion failure if it is not really an instance of the right type. This
337 should be used in cases where you have some information that makes you believe
342 <pre>
344 if (isa<<a href="#Constant">Constant</a>>(V) || isa<<a href="#Argument">Argument</a>>(V) || isa<<a href="#GlobalValue">GlobalValue</a>>(V))
345 return true;
349 }
350 </pre>
351 </div>
355 operator.</p>
357 </dd>
362 It checks to see if the operand is of the specified type, and if so, returns a
363 pointer to it (this operator does not work with references). If the operand is
364 not of the correct type, a null pointer is returned. Thus, this works very
368 statement like this:</p>
371 <pre>
372 if (<a href="#AllocationInst">AllocationInst</a> *AI = dyn_cast<<a href="#AllocationInst">AllocationInst</a>>(Val)) {
374 }
375 </pre>
376 </div>
380 statement, which is very convenient.</p>
385 blocks to check for lots of different variants of classes. If you find
386 yourself wanting to do this, it is much cleaner and more efficient to use the
389 </dd>
395 argument (which it then propagates). This can sometimes be useful, allowing
396 you to combine several null checks into one.</p></dd>
402 as an argument (which it then propagates). This can sometimes be useful,
403 allowing you to combine several null checks into one.</p></dd>
405 </dl>
407 <p>These five templates can be used with any classes, whether they have a
408 v-table or not. To add support for these templates, you simply need to add
410 to. Describing this is currently outside the scope of this document, but there
411 are lots of examples in the LLVM source base.</p>
413 </div>
415 <!-- ======================================================================= -->
418 </div>
422 <p>Often when working on your pass you will put a bunch of debugging printouts
423 and other code into your pass. After you get it working, you want to remove
424 it, but you may need it again in the future (to work out new bugs that you run
425 across).</p>
427 <p> Naturally, because of this, you don't want to delete the debug printouts,
428 but you don't want them to always be noisy. A standard compromise is to comment
429 them out, allowing you to enable them if you need them in the future.</p>
433 this problem. Basically, you can put arbitrary code into the argument of the
438 <pre>
440 </pre>
441 </div>
446 <pre>
450 I am here!
451 </pre>
452 </div>
455 to not have to create "yet another" command line option for the debug output for
457 so they do not cause a performance impact at all (for the same reason, they
458 should also not contain side-effects!).</p>
461 enable or disable it directly in gdb. Just use "<tt>set DebugFlag=0</tt>" or
462 "<tt>set DebugFlag=1</tt>" from the gdb if the program is running. If the
463 program hasn't been started yet, you can always just run it with
466 </div>
468 <!-- _______________________________________________________________________ -->
472 </div>
478 generator). If you want to enable debug information with more fine-grained
480 option as follows:</p>
483 <pre>
485 #undef DEBUG_TYPE
486 #define DEBUG_TYPE "foo"
488 #undef DEBUG_TYPE
489 #define DEBUG_TYPE "bar"
491 #undef DEBUG_TYPE
492 #define DEBUG_TYPE ""
494 </pre>
495 </div>
500 <pre>
504 No debug type
505 'foo' debug type
506 'bar' debug type
507 No debug type (2)
509 'foo' debug type
511 'bar' debug type
512 </pre>
513 </div>
516 a file, to specify the debug type for the entire module (if you do this before
519 "bar", because there is no system in place to ensure that names do not
520 conflict. If two different modules use the same string, they will all be turned
521 on when the name is specified. This allows, for example, all debug information
523 even if the source lives in multiple files.</p>
525 </div>
527 <!-- ======================================================================= -->
530 option</a>
531 </div>
538 keep track of what the LLVM compiler is doing and how effective various
539 optimizations are. It is useful to see what optimizations are contributing to
540 making a particular program run faster.</p>
542 <p>Often you may run your pass on some big program, and you're interested to see
543 how many times it makes a certain transformation. Although you can do this with
544 hand inspection, or some ad-hoc method, this is a real pain and not very useful
546 keep track of this information, and the calculated information is presented in a
547 uniform manner with the rest of the passes being executed.</p>
550 it are as follows:</p>
552 <ol>
556 <pre>
557 #define <a href="#DEBUG_TYPE">DEBUG_TYPE</a> "mypassname" <i>// This goes before any #includes.</i>
558 STATISTIC(NumXForms, "The # of times I did stuff");
559 </pre>
560 </div>
563 specified by the first argument. The pass name is taken from the DEBUG_TYPE
564 macro, and the description is taken from the second argument. The variable
570 <pre>
572 </pre>
573 </div>
575 </li>
576 </ol>
582 <pre>
585 </pre>
586 </div>
589 suite, it gives a report that looks like this:</p>
592 <pre>
593 7646 bitcodewriter - Number of normal instructions
594 725 bitcodewriter - Number of oversized instructions
595 129996 bitcodewriter - Number of bitcode bytes written
596 2817 raise - Number of insts DCEd or constprop'd
597 3213 raise - Number of cast-of-self removed
598 5046 raise - Number of expression trees converted
599 75 raise - Number of other getelementptr's formed
600 138 raise - Number of load/store peepholes
601 42 deadtypeelim - Number of unused typenames removed from symtab
602 392 funcresolve - Number of varargs functions resolved
603 27 globaldce - Number of global variables removed
604 2 adce - Number of basic blocks removed
605 134 cee - Number of branches revectored
606 49 cee - Number of setcc instruction eliminated
607 532 gcse - Number of loads removed
608 2919 gcse - Number of instructions removed
609 86 indvars - Number of canonical indvars added
610 87 indvars - Number of aux indvars removed
611 25 instcombine - Number of dead inst eliminate
612 434 instcombine - Number of insts combined
613 248 licm - Number of load insts hoisted
614 1298 licm - Number of insts hoisted to a loop pre-header
615 3 licm - Number of insts hoisted to multiple loop preds (bad, no loop pre-header)
616 75 mem2reg - Number of alloca's promoted
617 1444 cfgsimplify - Number of blocks simplified
618 </pre>
619 </div>
621 <p>Obviously, with so many optimizations, having a unified framework for this
622 stuff is very nice. Making your pass fit well into the framework makes it more
623 maintainable and useful.</p>
625 </div>
627 <!-- ======================================================================= -->
630 </div>
634 <p>Several of the important data structures in LLVM are graphs: for example
638 DAGs</a>. In many cases, while debugging various parts of the compiler, it is
639 nice to instantly visualize these graphs.</p>
641 <p>LLVM provides several callbacks that are available in a debug build to do
643 the current LLVM tool will pop up a window containing the CFG for the function
644 where each basic block is a node in the graph, and each node contains the
645 instructions in the block. Similarly, there also exists
650 up a window. Alternatively, you can sprinkle calls to these functions in your
651 code in places you want to debug.</p>
653 <p>Getting this to work requires a small amount of configuration. On Unix
655 toolkit, and make sure 'dot' and 'gv' are in your path. If you are running on
656 Mac OS/X, download and install the Mac OS/X <a
659 it) to your path. Once in your system and path are set up, rerun the LLVM
660 configure script and rebuild LLVM to enable this functionality.</p>
666 specified color (choices of colors can be found at <a
668 complex node attributes can be provided with <tt>call
671 Attributes</a>.) If you want to restart and clear all the current graph
674 </div>
676 <!-- *********************************************************************** -->
679 </div>
680 <!-- *********************************************************************** -->
685 and we commonly use STL data structures. This section describes the trade-offs
686 you should consider when you pick one.</p>
688 <p>
689 The first step is a choose your own adventure: do you want a sequential
690 container, a set-like container, or a map-like container? The most important
691 thing when choosing a container is the algorithmic properties of how you plan to
692 access the container. Based on that, you should use:</p>
694 <ul>
696 of an value based on another value. Map-like containers also support
697 efficient queries for containment (whether a key is in the map). Map-like
698 containers generally do not support efficient reverse mapping (values to
699 keys). If you need that, use two maps. Some map-like containers also
700 support efficient iteration through the keys in sorted order. Map-like
701 containers are the most expensive sort, only use them if you need one of
702 these capabilities.</li>
705 stuff into a container that automatically eliminates duplicates. Some
706 set-like containers support efficient iteration through the elements in
707 sorted order. Set-like containers are more expensive than sequential
708 containers.
709 </li>
712 the most efficient way to add elements and keeps track of the order they are
713 added to the collection. They permit duplicates and support efficient
714 iteration, but do not support efficient look-up based on a key.
715 </li>
718 perform set operations on sets of numeric id's, while automatically
719 eliminating duplicates. Bit containers require a maximum of 1 bit for each
720 identifier you want to store.
721 </li>
722 </ul>
724 <p>
725 Once the proper category of container is determined, you can fine tune the
726 memory use, constant factors, and cache behaviors of access by intelligently
727 picking a member of the category. Note that constant factors and cache behavior
728 can be a big deal. If you have a vector that usually only contains a few
729 elements (but could contain many), for example, it's much better to use
731 . Doing so avoids (relatively) expensive malloc/free calls, which dwarf the
732 cost of adding the elements to the container. </p>
734 </div>
736 <!-- ======================================================================= -->
739 </div>
742 There are a variety of sequential containers available for you, based on your
743 needs. Pick the first in this section that will do what you want.
744 </div>
746 <!-- _______________________________________________________________________ -->
749 </div>
752 <p>Fixed size arrays are very simple and very fast. They are good if you know
753 exactly how many elements you have, or you have a (low) upper bound on how many
754 you have.</p>
755 </div>
757 <!-- _______________________________________________________________________ -->
760 </div>
763 <p>Heap allocated arrays (new[] + delete[]) are also simple. They are good if
764 the number of elements is variable, if you know how many elements you will need
765 before the array is allocated, and if the array is usually large (if not,
767 allocated array is the cost of the new/delete (aka malloc/free). Also note that
768 if you are allocating an array of a type with a constructor, the constructor and
769 destructors will be run for every element in the array (re-sizable vectors only
770 construct those elements actually used).</p>
771 </div>
773 <!-- _______________________________________________________________________ -->
776 </div>
781 it supports efficient iteration, lays out elements in memory order (so you can
782 do pointer arithmetic between elements), supports efficient push_back/pop_back
783 operations, supports efficient random access to its elements, etc.</p>
785 <p>The advantage of SmallVector is that it allocates space for
787 the SmallVector is dynamically smaller than N, no malloc is performed. This can
788 be a big win in cases where the malloc/free call is far more expensive than the
789 code that fiddles around with the elements.</p>
792 predecessors/successors of a block is usually less than 8). On the other hand,
793 this makes the size of the SmallVector itself large, so you don't want to
794 allocate lots of them (doing so will waste a lot of space). As such,
795 SmallVectors are most useful when on the stack.</p>
797 <p>SmallVector also provides a nice portable and efficient replacement for
800 </div>
802 <!-- _______________________________________________________________________ -->
805 </div>
808 <p>
809 std::vector is well loved and respected. It is useful when SmallVector isn't:
810 when the size of the vector is often large (thus the small optimization will
811 rarely be a benefit) or if you will be allocating many instances of the vector
812 itself (which would waste space for elements that aren't in the container).
813 vector is also useful when interfacing with code that expects vectors :).
814 </p>
819 <pre>
820 for ( ... ) {
822 use V;
823 }
824 </pre>
825 </div>
830 <pre>
832 for ( ... ) {
833 use V;
834 V.clear();
835 }
836 </pre>
837 </div>
839 <p>Doing so will save (at least) one heap allocation and free per iteration of
840 the loop.</p>
842 </div>
844 <!-- _______________________________________________________________________ -->
847 </div>
850 <p>std::deque is, in some senses, a generalized version of std::vector. Like
851 std::vector, it provides constant time random access and other similar
852 properties, but it also provides efficient access to the front of the list. It
853 does not guarantee continuity of elements within memory.</p>
855 <p>In exchange for this extra flexibility, std::deque has significantly higher
856 constant factor costs than std::vector. If possible, use std::vector or
857 something cheaper.</p>
858 </div>
860 <!-- _______________________________________________________________________ -->
863 </div>
866 <p>std::list is an extremely inefficient class that is rarely useful.
867 It performs a heap allocation for every element inserted into it, thus having an
868 extremely high constant factor, particularly for small data types. std::list
869 also only supports bidirectional iteration, not random access iteration.</p>
871 <p>In exchange for this high cost, std::list supports efficient access to both
872 ends of the list (like std::deque, but unlike std::vector or SmallVector). In
873 addition, the iterator invalidation characteristics of std::list are stronger
874 than that of a vector class: inserting or removing an element into the list does
875 not invalidate iterator or pointers to other elements in the list.</p>
876 </div>
878 <!-- _______________________________________________________________________ -->
881 </div>
885 intrusive, because it requires the element to store and provide access to the
886 prev/next pointers for the list.</p>
890 provides some novel characteristics. In particular, it can efficiently store
891 polymorphic objects, the traits class is informed when an element is inserted or
893 constant-time splice operation.</p>
895 <p>These properties are exactly what we want for things like
899 Related classes of interest are explained in the following subsections:
900 <ul>
905 </ul>
906 </div>
908 <!-- _______________________________________________________________________ -->
911 </div>
916 publicly derive from this traits class.</p>
917 </div>
919 <!-- _______________________________________________________________________ -->
922 </div>
926 supports a slightly narrower interface. Notably, inserters from
930 used for a wide variety of customizations.</p>
931 </div>
933 <!-- _______________________________________________________________________ -->
936 </div>
941 in the default manner.</p>
946 </div>
948 <!-- _______________________________________________________________________ -->
951 </div>
955 citizen in the C++ ecosystem, it needs to support the standard container
960 <p>The only sensible solution to this problem is to allocate a so-called
962 iterator, providing the back-link to the last element. However conforming to the
964 also must not be dereferenced.</p>
967 how to allocate and store the sentinel. The corresponding policy is dictated
969 whenever the need for a sentinel arises.</p>
971 <p>While the default policy is sufficient in most cases, it may break down when
974 is wasted. To alleviate the situation with numerous and voluminous
980 arithmetic is used to obtain the sentinel, which is relative to the
982 extra pointer, which serves as the back-link of the sentinel. This is the only
983 field in the ghostly sentinel which can be legally accessed.</p>
984 </div>
986 <!-- _______________________________________________________________________ -->
989 </div>
994 <p>There are also various STL adapter classes such as std::queue,
995 std::priority_queue, std::stack, etc. These provide simplified access to an
996 underlying container but don't affect the cost of the container itself.</p>
998 </div>
1001 <!-- ======================================================================= -->
1004 </div>
1008 <p>Set-like containers are useful when you need to canonicalize multiple values
1009 into a single representation. There are several different choices for how to do
1010 this, providing various trade-offs.</p>
1012 </div>
1015 <!-- _______________________________________________________________________ -->
1018 </div>
1022 <p>If you intend to insert a lot of elements, then do a lot of queries, a
1023 great approach is to use a vector (or other sequential container) with
1024 std::sort+std::unique to remove duplicates. This approach works really well if
1025 your usage pattern has these two distinct phases (insert then query), and can be
1027 </p>
1029 <p>
1030 This combination provides the several nice properties: the result data is
1031 contiguous in memory (good for cache locality), has few allocations, is easy to
1032 address (iterators in the final vector are just indices or pointers), and can be
1033 efficiently queried with a standard binary or radix search.</p>
1035 </div>
1037 <!-- _______________________________________________________________________ -->
1040 </div>
1044 <p>If you have a set-like data structure that is usually small and whose elements
1046 has space for N elements in place (thus, if the set is dynamically smaller than
1047 N, no malloc traffic is required) and accesses them with a simple linear search.
1048 When the set grows beyond 'N' elements, it allocates a more expensive representation that
1049 guarantees efficient access (for most types, it falls back to std::set, but for
1050 pointers it uses something far better, <a
1053 <p>The magic of this class is that it handles small sets extremely efficiently,
1054 but gracefully handles extremely large sets without loss of efficiency. The
1055 drawback is that the interface is quite small: it supports insertion, queries
1056 and erasing, but does not support iteration.</p>
1058 </div>
1060 <!-- _______________________________________________________________________ -->
1063 </div>
1067 <p>SmallPtrSet has all the advantages of SmallSet (and a SmallSet of pointers is
1068 transparently implemented with a SmallPtrSet), but also supports iterators. If
1069 more than 'N' insertions are performed, a single quadratically
1070 probed hash table is allocated and grows as needed, providing extremely
1071 efficient access (constant time insertion/deleting/queries with low constant
1072 factors) and is very stingy with malloc traffic.</p>
1074 <p>Note that, unlike std::set, the iterators of SmallPtrSet are invalidated
1075 whenever an insertion occurs. Also, the values visited by the iterators are not
1076 visited in sorted order.</p>
1078 </div>
1080 <!-- _______________________________________________________________________ -->
1083 </div>
1087 <p>
1088 DenseSet is a simple quadratically probed hash table. It excels at supporting
1089 small values: it uses a single allocation to hold all of the pairs that
1090 are currently inserted in the set. DenseSet is a great way to unique small
1091 values that are not simple pointers (use <a
1093 the same requirements for the value type that <a
1095 </p>
1097 </div>
1099 <!-- _______________________________________________________________________ -->
1102 </div>
1106 <p>
1107 FoldingSet is an aggregate class that is really good at uniquing
1108 expensive-to-create or polymorphic objects. It is a combination of a chained
1109 hash table with intrusive links (uniqued objects are required to inherit from
1111 its ID process.</p>
1114 a complex object (for example, a node in the code generator). The client has a
1115 description of *what* it wants to generate (it knows the opcode and all the
1116 operands), but we don't want to 'new' a node, then try inserting it into a set
1117 only to find out it already exists, at which point we would have to delete it
1118 and return the node that already exists.
1119 </p>
1121 <p>To support this style of client, FoldingSet perform a query with a
1122 FoldingSetNodeID (which wraps SmallVector) that can be used to describe the
1123 element that we want to query for. The query either returns the element
1124 matching the ID or it returns an opaque ID that indicates where insertion should
1125 take place. Construction of the ID usually does not require heap traffic.</p>
1127 <p>Because FoldingSet uses intrusive links, it can support polymorphic objects
1128 in the set (for example, you can have SDNode instances mixed with LoadSDNodes).
1129 Because the elements are individually allocated, pointers to the elements are
1130 stable: inserting or removing elements does not invalidate any pointers to other
1131 elements.
1132 </p>
1134 </div>
1136 <!-- _______________________________________________________________________ -->
1139 </div>
1144 many things but great at nothing. std::set allocates memory for each element
1145 inserted (thus it is very malloc intensive) and typically stores three pointers
1146 per element in the set (thus adding a large amount of per-element space
1147 overhead). It offers guaranteed log(n) performance, which is not particularly
1148 fast from a complexity standpoint (particularly if the elements of the set are
1149 expensive to compare, like strings), and has extremely high constant factors for
1150 lookup, insertion and removal.</p>
1152 <p>The advantages of std::set are that its iterators are stable (deleting or
1153 inserting an element from the set does not affect iterators or pointers to other
1154 elements) and that iteration over the set is guaranteed to be in sorted order.
1155 If the elements in the set are large, then the relative overhead of the pointers
1156 and malloc traffic is not a big deal, but if the elements of the set are small,
1157 std::set is almost never a good choice.</p>
1159 </div>
1161 <!-- _______________________________________________________________________ -->
1164 </div>
1169 Container</a>. The important property
1170 that this provides is efficient insertion with uniquing (duplicate elements are
1171 ignored) with iteration support. It implements this by inserting elements into
1172 both a set-like container and the sequential container, using the set-like
1173 container for uniquing and the sequential container for iteration.
1174 </p>
1176 <p>The difference between SetVector and other sets is that the order of
1177 iteration is guaranteed to match the order of insertion into the SetVector.
1178 This property is really important for things like sets of pointers. Because
1179 pointer values are non-deterministic (e.g. vary across runs of the program on
1180 different machines), iterating over the pointers in the set will
1181 not be in a well-defined order.</p>
1183 <p>
1184 The drawback of SetVector is that it requires twice as much space as a normal
1185 set and has the sum of constant factors from the set-like container and the
1186 sequential container that it uses. Use it *only* if you need to iterate over
1187 the elements in a deterministic order. SetVector is also expensive to delete
1188 elements out of (linear time), unless you use it's "pop_back" method, which is
1189 faster.
1190 </p>
1192 <p>SetVector is an adapter class that defaults to using std::vector and std::set
1193 for the underlying containers, so it is quite expensive. However,
1195 defaults to using a SmallVector and SmallSet of a specified size. If you use
1196 this, and if your sets are dynamically smaller than N, you will save a lot of
1197 heap traffic.</p>
1199 </div>
1201 <!-- _______________________________________________________________________ -->
1204 </div>
1208 <p>
1210 retains a unique ID for each element inserted into the set. It internally
1211 contains a map and a vector, and it assigns a unique ID for each value inserted
1212 into the set.</p>
1214 <p>UniqueVector is very expensive: its cost is the sum of the cost of
1215 maintaining both the map and vector, it has high complexity, high constant
1216 factors, and produces a lot of malloc traffic. It should be avoided.</p>
1218 </div>
1221 <!-- _______________________________________________________________________ -->
1224 </div>
1228 <p>
1229 The STL provides several other options, such as std::multiset and the various
1230 "hash_set" like containers (whether from C++ TR1 or from the SGI library). We
1231 never use hash_set and unordered_set because they are generally very expensive
1232 (each insertion requires a malloc) and very non-portable.
1233 </p>
1235 <p>std::multiset is useful if you're not interested in elimination of
1236 duplicates, but has all the drawbacks of std::set. A sorted vector (where you
1237 don't delete duplicate entries) or some other approach is almost always
1238 better.</p>
1240 </div>
1242 <!-- ======================================================================= -->
1245 </div>
1248 Map-like containers are useful when you want to associate data to a key. As
1249 usual, there are a lot of different ways to do this. :)
1250 </div>
1252 <!-- _______________________________________________________________________ -->
1255 </div>
1259 <p>
1260 If your usage pattern follows a strict insert-then-query approach, you can
1262 for set-like containers</a>. The only difference is that your query function
1263 (which uses std::lower_bound to get efficient log(n) lookup) should only compare
1264 the key, not both the key and value. This yields the same advantages as sorted
1265 vectors for sets.
1266 </p>
1267 </div>
1269 <!-- _______________________________________________________________________ -->
1272 </div>
1276 <p>
1277 Strings are commonly used as keys in maps, and they are difficult to support
1278 efficiently: they are variable length, inefficient to hash and compare when
1279 long, expensive to copy, etc. StringMap is a specialized container designed to
1280 cope with these issues. It supports mapping an arbitrary range of bytes to an
1281 arbitrary other object.</p>
1283 <p>The StringMap implementation uses a quadratically-probed hash table, where
1284 the buckets store a pointer to the heap allocated entries (and some other
1285 stuff). The entries in the map must be heap allocated because the strings are
1286 variable length. The string data (key) and the element object (value) are
1287 stored in the same allocation with the string data immediately after the element
1288 object. This container guarantees the "<tt>(char*)(&Value+1)</tt>" points
1289 to the key string for a value.</p>
1291 <p>The StringMap is very fast for several reasons: quadratic probing is very
1292 cache efficient for lookups, the hash value of strings in buckets is not
1293 recomputed when lookup up an element, StringMap rarely has to touch the
1294 memory for unrelated objects when looking up a value (even when hash collisions
1295 happen), hash table growth does not recompute the hash values for strings
1296 already in the table, and each pair in the map is store in a single allocation
1297 (the string data is stored in the same allocation as the Value of a pair).</p>
1299 <p>StringMap also provides query methods that take byte ranges, so it only ever
1300 copies a string if a value is inserted into the table.</p>
1301 </div>
1303 <!-- _______________________________________________________________________ -->
1306 </div>
1309 <p>
1310 IndexedMap is a specialized container for mapping small dense integers (or
1311 values that can be mapped to small dense integers) to some other type. It is
1312 internally implemented as a vector with a mapping function that maps the keys to
1313 the dense integer range.
1314 </p>
1316 <p>
1317 This is useful for cases like virtual registers in the LLVM code generator: they
1318 have a dense mapping that is offset by a compile-time constant (the first
1319 virtual register ID).</p>
1321 </div>
1323 <!-- _______________________________________________________________________ -->
1326 </div>
1330 <p>
1331 DenseMap is a simple quadratically probed hash table. It excels at supporting
1332 small keys and values: it uses a single allocation to hold all of the pairs that
1333 are currently inserted in the map. DenseMap is a great way to map pointers to
1334 pointers, or map other small types to each other.
1335 </p>
1337 <p>
1338 There are several aspects of DenseMap that you should be aware of, however. The
1339 iterators in a densemap are invalidated whenever an insertion occurs, unlike
1340 map. Also, because DenseMap allocates space for a large number of key/value
1341 pairs (it starts with 64 by default), it will waste a lot of space if your keys
1342 or values are large. Finally, you must implement a partial specialization of
1343 DenseMapInfo for the key that you want, if it isn't already supported. This
1344 is required to tell DenseMap about two special marker values (which can never be
1345 inserted into the map) that it needs internally.</p>
1347 </div>
1349 <!-- _______________________________________________________________________ -->
1352 </div>
1356 <p>
1358 a single allocation per pair inserted into the map, it offers log(n) lookup with
1359 an extremely large constant factor, imposes a space penalty of 3 pointers per
1360 pair in the map, etc.</p>
1362 <p>std::map is most useful when your keys or values are very large, if you need
1363 to iterate over the collection in sorted order, or if you need stable iterators
1364 into the map (i.e. they don't get invalidated if an insertion or deletion of
1365 another element takes place).</p>
1367 </div>
1369 <!-- _______________________________________________________________________ -->
1372 </div>
1376 <p>
1377 The STL provides several other options, such as std::multimap and the various
1378 "hash_map" like containers (whether from C++ TR1 or from the SGI library). We
1379 never use hash_set and unordered_set because they are generally very expensive
1380 (each insertion requires a malloc) and very non-portable.</p>
1382 <p>std::multimap is useful if you want to map a key to multiple values, but has
1383 all the drawbacks of std::map. A sorted vector or some other approach is almost
1384 always better.</p>
1386 </div>
1388 <!-- ======================================================================= -->
1391 </div>
1394 <p>Unlike the other containers, there are only two bit storage containers, and
1395 choosing when to use each is relatively straightforward.</p>
1397 <p>One additional option is
1399 implementation in many common compilers (e.g. commonly available versions of
1400 GCC) is extremely inefficient and 2) the C++ standards committee is likely to
1401 deprecate this container and/or change it significantly somehow. In any case,
1402 please don't use it.</p>
1403 </div>
1405 <!-- _______________________________________________________________________ -->
1408 </div>
1411 <p> The BitVector container provides a fixed size set of bits for manipulation.
1412 It supports individual bit setting/testing, as well as set operations. The set
1413 operations take time O(size of bitvector), but operations are performed one word
1414 at a time, instead of one bit at a time. This makes the BitVector very fast for
1415 set operations compared to other containers. Use the BitVector when you expect
1416 the number of set bits to be high (IE a dense set).
1417 </p>
1418 </div>
1420 <!-- _______________________________________________________________________ -->
1423 </div>
1426 <p> The SparseBitVector container is much like BitVector, with one major
1427 difference: Only the bits that are set, are stored. This makes the
1428 SparseBitVector much more space efficient than BitVector when the set is sparse,
1429 as well as making set operations O(number of set bits) instead of O(size of
1430 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
1431 (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).
1432 </p>
1433 </div>
1435 <!-- *********************************************************************** -->
1438 </div>
1439 <!-- *********************************************************************** -->
1443 <p>This section describes how to perform some very simple transformations of
1444 LLVM code. This is meant to give examples of common idioms used, showing the
1446 you should also read about the main classes that you will be working with. The
1448 and descriptions of the main classes that you should know about.</p>
1450 </div>
1452 <!-- NOTE: this section should be heavy on example code -->
1453 <!-- ======================================================================= -->
1456 </div>
1460 <p>The LLVM compiler infrastructure have many different data structures that may
1461 be traversed. Following the example of the C++ standard template library, the
1462 techniques used to traverse these various data structures are all basically the
1465 function returns an iterator pointing to one past the last valid element of the
1467 between the two operations.</p>
1469 <p>Because the pattern for iteration is common across many different aspects of
1470 the program representation, the standard template library algorithms may be used
1471 on them, and it is easier to remember how to iterate. First we show a few common
1472 examples of the data structures that need to be traversed. Other data
1473 structures are traversed in very similar ways.</p>
1475 </div>
1477 <!-- _______________________________________________________________________ -->
1482 </div>
1487 transform in some way; in particular, you'd like to manipulate its
1494 <pre>
1501 </pre>
1502 </div>
1504 <p>Note that i can be used as if it were a pointer for the purposes of
1506 because the indirection operator is overloaded for the iterator
1510 </div>
1512 <!-- _______________________________________________________________________ -->
1517 </div>
1522 easy to iterate over the individual instructions that make up
1527 <pre>
1533 </pre>
1534 </div>
1536 <p>However, this isn't really the best way to print out the contents of a
1538 anything you'll care about, you could have just invoked the print routine on the
1541 </div>
1543 <!-- _______________________________________________________________________ -->
1548 </div>
1557 small example that shows how to dump all instructions in a function to the standard error stream:<p>
1560 <pre>
1564 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
1566 </pre>
1567 </div>
1570 work list with its initial contents. For example, if you wanted to
1572 F, all you would need to do is something like:</p>
1575 <pre>
1579 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
1580 worklist.insert(&*I);
1581 </pre>
1582 </div>
1587 </div>
1589 <!-- _______________________________________________________________________ -->
1592 vice-versa)</a>
1593 </div>
1597 <p>Sometimes, it'll be useful to grab a reference (or pointer) to a class
1598 instance when all you've got at hand is an iterator. Well, extracting
1599 a reference or a pointer from an iterator is very straight-forward.
1604 <pre>
1607 const Instruction& inst = *j;
1608 </pre>
1609 </div>
1611 <p>However, the iterators you'll be working with in the LLVM framework are
1612 special: they will automatically convert to a ptr-to-instance type whenever they
1613 need to. Instead of dereferencing the iterator and then taking the address of
1614 the result, you can simply assign the iterator to the proper pointer type and
1615 you get the dereference and address-of operation as a result of the assignment
1616 (behind the scenes, this is a result of overloading casting mechanisms). Thus
1617 the last line of the last example,</p>
1620 <pre>
1621 Instruction *pinst = &*i;
1622 </pre>
1623 </div>
1628 <pre>
1629 Instruction *pinst = i;
1630 </pre>
1631 </div>
1633 <p>It's also possible to turn a class pointer into the corresponding iterator,
1634 and this is a constant time operation (very efficient). The following code
1635 snippet illustrates use of the conversion constructors provided by LLVM
1636 iterators. By using these, you can explicitly grab the iterator of something
1637 without actually obtaining it via iteration over some structure:</p>
1640 <pre>
1641 void printNextInstruction(Instruction* inst) {
1642 BasicBlock::iterator it(inst);
1645 }
1646 </pre>
1647 </div>
1649 </div>
1651 <!--_______________________________________________________________________-->
1654 example</a>
1655 </div>
1659 <p>Say that you're writing a FunctionPass and would like to count all the
1663 much more straight-forward manner, but this example will allow us to explore how
1665 is what we want to do:</p>
1668 <pre>
1669 initialize callCounter to zero
1670 for each Function f in the Module
1671 for each BasicBlock b in f
1672 for each Instruction i in b
1673 if (i is a CallInst and calls the given function)
1674 increment callCounter
1675 </pre>
1676 </div>
1678 <p>And the actual code is (remember, because we're writing a
1683 <pre>
1684 Function* targetFunc = ...;
1686 class OurFunctionPass : public FunctionPass {
1687 public:
1688 OurFunctionPass(): callCounter(0) { }
1690 virtual runOnFunction(Function& F) {
1691 for (Function::iterator b = F.begin(), be = F.end(); b != be; ++b) {
1698 if (callInst->getCalledFunction() == targetFunc)
1699 ++callCounter;
1700 }
1701 }
1702 }
1703 }
1705 private:
1706 unsigned callCounter;
1707 };
1708 </pre>
1709 </div>
1711 </div>
1713 <!--_______________________________________________________________________-->
1716 </div>
1720 <p>You may have noticed that the previous example was a bit oversimplified in
1721 that it did not deal with call sites generated by 'invoke' instructions. In
1722 this, and in other situations, you may find that you want to treat
1725 less closely-related things. For these cases, LLVM provides a handy wrapper
1726 class called <a
1733 reference and it should not be dynamically allocated or deallocated using
1735 assignable and constructable, with costs equivalents to that of a bare pointer.
1736 If you look at its definition, it has only a single pointer member.</p>
1738 </div>
1740 <!--_______________________________________________________________________-->
1743 </div>
1757 <pre>
1758 Function *F = ...;
1764 }
1765 </pre>
1766 </div>
1773 all of the values that a particular instruction uses (that is, the operands of
1777 <pre>
1778 Instruction *pi = ...;
1781 Value *v = *i;
1783 }
1784 </pre>
1785 </div>
1787 <!--
1788 def-use chains ("finding all users of"): Value::use_begin/use_end
1789 use-def chains ("finding all values used"): User::op_begin/op_end [op=operand]
1790 -->
1792 </div>
1794 <!--_______________________________________________________________________-->
1797 successors of blocks</a>
1798 </div>
1802 <p>Iterating over the predecessors and successors of a block is quite easy
1804 this to iterate over all predecessors of BB:</p>
1807 <pre>
1808 #include "llvm/Support/CFG.h"
1809 BasicBlock *BB = ...;
1811 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1812 BasicBlock *Pred = *PI;
1814 }
1815 </pre>
1816 </div>
1818 <p>Similarly, to iterate over successors use
1819 succ_iterator/succ_begin/succ_end.</p>
1821 </div>
1824 <!-- ======================================================================= -->
1827 </div>
1831 <p>There are some primitive transformation operations present in the LLVM
1832 infrastructure that are worth knowing about. When performing
1833 transformations, it's fairly common to manipulate the contents of basic
1834 blocks. This section describes some of the common methods for doing so
1835 and gives example code.</p>
1837 </div>
1839 <!--_______________________________________________________________________-->
1843 </div>
1850 constructor for the kind of instruction to instantiate and provide the necessary
1855 <pre>
1856 AllocaInst* ai = new AllocaInst(Type::Int32Ty);
1857 </pre>
1858 </div>
1862 subclass is likely to have varying default parameters which change the semantics
1863 of the instruction, so refer to the <a
1869 <p>It is very useful to name the values of instructions when you're able to, as
1870 this facilitates the debugging of your transformations. If you end up looking
1871 at generated LLVM machine code, you definitely want to have logical names
1872 associated with the results of instructions! By supplying a value for the
1874 associate a logical name with the result of the instruction's execution at
1875 run time. For example, say that I'm writing a transformation that dynamically
1876 allocates space for an integer on the stack, and that integer is going to be
1877 used as some kind of index by some other code. To accomplish this, I place an
1883 <pre>
1885 </pre>
1886 </div>
1889 execution value, which is a pointer to an integer on the run time stack.</p>
1896 <ul>
1897 <li>Insertion into an explicit instruction list
1904 <pre>
1905 BasicBlock *pb = ...;
1906 Instruction *pi = ...;
1907 Instruction *newInst = new Instruction(...);
1910 </pre>
1911 </div>
1915 classes provide constructors which take a pointer to a
1917 looked like: </p>
1920 <pre>
1921 BasicBlock *pb = ...;
1922 Instruction *newInst = new Instruction(...);
1925 </pre>
1926 </div>
1931 <pre>
1932 BasicBlock *pb = ...;
1933 Instruction *newInst = new Instruction(..., pb);
1934 </pre>
1935 </div>
1937 <p>which is much cleaner, especially if you are creating
1938 long instruction streams.</p></li>
1940 <li>Insertion into an implicit instruction list
1943 are implicitly associated with an existing instruction list: the instruction
1944 list of the enclosing basic block. Thus, we could have accomplished the same
1946 </p>
1949 <pre>
1950 Instruction *pi = ...;
1951 Instruction *newInst = new Instruction(...);
1954 </pre>
1955 </div>
1957 <p>In fact, this sequence of steps occurs so frequently that the
1959 constructors which take (as a default parameter) a pointer to an
1963 provided instruction, immediately before that instruction. Using an
1965 parameter, the above code becomes:</p>
1968 <pre>
1969 Instruction* pi = ...;
1970 Instruction* newInst = new Instruction(..., pi);
1971 </pre>
1972 </div>
1974 <p>which is much cleaner, especially if you're creating a lot of
1976 </ul>
1978 </div>
1980 <!--_______________________________________________________________________-->
1983 </div>
1987 <p>Deleting an instruction from an existing sequence of instructions that form a
1989 you must have a pointer to the instruction that you wish to delete. Second, you
1990 need to obtain the pointer to that instruction's basic block. You use the
1991 pointer to the basic block to get its list of instructions and then use the
1992 erase function to remove your instruction. For example:</p>
1995 <pre>
1997 I->eraseFromParent();
1998 </pre>
1999 </div>
2001 </div>
2003 <!--_______________________________________________________________________-->
2007 </div>
2013 <p>Including "<a href="/doxygen/BasicBlockUtils_8h-source.html">llvm/Transforms/Utils/BasicBlockUtils.h</a>"
2019 <ul>
2022 <p>This function replaces all uses of a given instruction with a value,
2023 and then removes the original instruction. The following example
2024 illustrates the replacement of the result of a particular
2026 pointer to an integer.</p>
2029 <pre>
2030 AllocaInst* instToReplace = ...;
2031 BasicBlock::iterator ii(instToReplace);
2034 Constant::getNullValue(PointerType::getUnqual(Type::Int32Ty)));
2035 </pre></div></li>
2039 <p>This function replaces a particular instruction with another
2040 instruction, inserting the new instruction into the basic block at the
2041 location where the old instruction was, and replacing any uses of the old
2042 instruction with the new instruction. The following example illustrates
2046 <pre>
2047 AllocaInst* instToReplace = ...;
2048 BasicBlock::iterator ii(instToReplace);
2052 </pre></div></li>
2053 </ul>
2061 information.</p>
2063 <!-- Value::replaceAllUsesWith User::replaceUsesOfWith Point out:
2064 include/llvm/Transforms/Utils/ especially BasicBlockUtils.h with:
2065 ReplaceInstWithValue, ReplaceInstWithInst -->
2067 </div>
2069 <!--_______________________________________________________________________-->
2072 </div>
2076 <p>Deleting a global variable from a module is just as easy as deleting an
2077 Instruction. First, you must have a pointer to the global variable that you wish
2078 to delete. You use this pointer to erase it from its parent, the module.
2079 For example:</p>
2082 <pre>
2085 GV->eraseFromParent();
2086 </pre>
2087 </div>
2089 </div>
2091 <!-- *********************************************************************** -->
2094 </div>
2095 <!-- *********************************************************************** -->
2098 <p>
2099 This section describes some of the advanced or obscure API's that most clients
2100 do not need to be aware of. These API's tend manage the inner workings of the
2101 LLVM system, and only need to be accessed in unusual circumstances.
2102 </p>
2103 </div>
2105 <!-- ======================================================================= -->
2108 </div>
2112 <p>
2113 The LLVM type system has a very simple goal: allow clients to compare types for
2114 structural equality with a simple pointer comparison (aka a shallow compare).
2115 This goal makes clients much simpler and faster, and is used throughout the LLVM
2116 system.
2117 </p>
2119 <p>
2120 Unfortunately achieving this goal is not a simple matter. In particular,
2121 recursive types and late resolution of opaque types makes the situation very
2122 difficult to handle. Fortunately, for the most part, our implementation makes
2123 most clients able to be completely unaware of the nasty internal details. The
2124 primary case where clients are exposed to the inner workings of it are when
2125 building a recursive type. In addition to this case, the LLVM bitcode reader,
2126 assembly parser, and linker also have to be aware of the inner workings of this
2127 system.
2128 </p>
2130 <p>
2131 For our purposes below, we need three concepts. First, an "Opaque Type" is
2134 opaque type as part of its type graph (for example "<tt>{ opaque, i32 }</tt>").
2135 Third, a concrete type is a type that is not an abstract type (e.g. "<tt>{ i32,
2136 float }</tt>").
2137 </p>
2139 </div>
2141 <!-- ______________________________________________________________________ -->
2144 </div>
2148 <p>
2149 Because the most common question is "how do I build a recursive type with LLVM",
2150 we answer it now and explain it as we go. Here we include enough to cause this
2151 to be emitted to an output .ll file:
2152 </p>
2155 <pre>
2156 %mylist = type { %mylist*, i32 }
2157 </pre>
2158 </div>
2160 <p>
2161 To build this, use the following LLVM APIs:
2162 </p>
2165 <pre>
2169 Elts.push_back(PointerType::getUnqual(StructTy));
2170 Elts.push_back(Type::Int32Ty);
2171 StructType *NewSTy = StructType::get(Elts);
2175 cast<OpaqueType>(StructTy.get())-><a href="#refineAbstractTypeTo">refineAbstractTypeTo</a>(NewSTy);
2177 // <i>NewSTy is potentially invalidated, but StructTy (a <a href="#PATypeHolder">PATypeHolder</a>) is</i>
2183 </pre>
2184 </div>
2186 <p>
2187 This code shows the basic approach used to build recursive types: build a
2188 non-recursive type using 'opaque', then use type unification to close the cycle.
2189 The type unification step is performed by the <tt><a
2191 described next. After that, we describe the <a
2193 </p>
2195 </div>
2197 <!-- ______________________________________________________________________ -->
2200 </div>
2203 <p>
2205 While this method is actually a member of the DerivedType class, it is most
2206 often used on OpaqueType instances. Type unification is actually a recursive
2207 process. After unification, types can become structurally isomorphic to
2208 existing types, and all duplicates are deleted (to preserve pointer equality).
2209 </p>
2211 <p>
2212 In the example above, the OpaqueType object is definitely deleted.
2213 Additionally, if there is an "{ \2*, i32}" type already created in the system,
2215 a type is deleted, any "Type*" pointers in the program are invalidated. As
2218 types can never move or be deleted). To deal with this, the <a
2220 reference to a possibly refined type, and the <a
2222 complex datastructures.
2223 </p>
2225 </div>
2227 <!-- ______________________________________________________________________ -->
2230 </div>
2233 <p>
2234 PATypeHolder is a form of a "smart pointer" for Type objects. When VMCore
2235 happily goes about nuking types that become isomorphic to existing types, it
2236 automatically updates all PATypeHolder objects to point to the new type. In the
2237 example above, this allows the code to maintain a pointer to the resultant
2238 resolved recursive type, even though the Type*'s are potentially invalidated.
2239 </p>
2241 <p>
2242 PATypeHolder is an extremely light-weight object that uses a lazy union-find
2243 implementation to update pointers. For example the pointer from a Value to its
2244 Type is maintained by PATypeHolder objects.
2245 </p>
2247 </div>
2249 <!-- ______________________________________________________________________ -->
2252 </div>
2256 <p>
2257 Some data structures need more to perform more complex updates when types get
2258 resolved. To support this, a class can derive from the AbstractTypeUser class.
2259 This class
2260 allows it to get callbacks when certain types are resolved. To register to get
2261 callbacks for a particular type, the DerivedType::{add/remove}AbstractTypeUser
2262 methods can be called on a type. Note that these methods only work for <i>
2263 abstract</i> types. Concrete types (those that do not include any opaque
2264 objects) can never be refined.
2265 </p>
2266 </div>
2269 <!-- ======================================================================= -->
2273 </div>
2283 names for types.</p>
2286 by most clients. It should only be used when iteration over the symbol table
2287 names themselves are required, which is very special purpose. Note that not
2288 all LLVM
2290 an empty name) do not exist in the symbol table.
2291 </p>
2293 <p>These symbol tables support iteration over the values/types in the symbol
2298 appropriate symbol table. For types, use the Module::addTypeName method to
2299 insert entries into the symbol table.</p>
2301 </div>
2305 <!-- ======================================================================= -->
2308 </div>
2316 addition and removal.</p>
2318 <!-- ______________________________________________________________________ -->
2320 <a name="Use2User">Interaction and relationship between <tt>User</tt> and <tt>Use</tt> objects</a>
2321 </div>
2324 <p>
2326 or refer to them out-of-line by means of a pointer. A mixed variant
2329 </p>
2330 </div>
2332 <p>
2334 <ul>
2335 <li><p>Layout a)
2337 object and there are a fixed number of them.</p>
2339 <li><p>Layout b)
2342 number of them.</p>
2343 </ul>
2344 <p>
2345 As of v2.4 each layout still possesses a direct pointer to the
2347 we stick to this redundancy for the sake of simplicity.
2349 has. (Theoretically this information can also be calculated
2350 given the scheme presented below.)</p>
2351 <p>
2353 enforce the following memory layouts:</p>
2355 <ul>
2356 <li><p>Layout a) is modelled by prepending the <tt>User</tt> object by the <tt>Use[]</tt> array.</p>
2358 <pre>
2359 ...---.---.---.---.-------...
2360 | P | P | P | P | User
2361 '''---'---'---'---'-------'''
2362 </pre>
2365 <pre>
2366 .-------...
2367 | User
2368 '-------'''
2369 |
2370 v
2371 .---.---.---.---...
2372 | P | P | P | P |
2373 '---'---'---'---'''
2374 </pre>
2375 </ul>
2379 <!-- ______________________________________________________________________ -->
2382 </div>
2385 <p>
2388 recover it. This is accomplished by the following scheme:</p>
2389 </div>
2391 A bit-encoding in the 2 LSBits (least significant bits) of the <tt>Use::Prev</tt> allows to find the
2393 <ul>
2398 </ul>
2399 <p>
2402 we have to walk to the next stop picking up digits
2403 and calculating the offset:</p>
2404 <pre>
2405 .---.---.---.---.---.---.---.---.---.---.---.---.---.---.---.---.----------------
2407 '---'---'---'---'---'---'---'---'---'---'---'---'---'---'---'---'----------------
2409 | | | | |__>
2410 | | | |__________>
2411 | | |______________________>
2412 | |______________________________________>
2413 |__________________________________________________________>
2414 </pre>
2415 <p>
2416 Only the significant number of bits need to be stored between the
2420 <!-- ______________________________________________________________________ -->
2423 </div>
2426 <p>
2427 The following literate Haskell fragment demonstrates the concept:</p>
2428 </div>
2431 <pre>
2432 > import Test.QuickCheck
2433 >
2438 >
2444 >
2445 > takeLast n ss = reverse $ take n $ reverse ss
2446 >
2448 >
2449 </pre>
2450 </div>
2451 <p>
2453 <p>
2454 The reverse algorithm computes the length of the string just by examining
2455 a certain prefix:</p>
2458 <pre>
2463 >
2466 > decode walk acc _ = walk + acc
2467 >
2468 </pre>
2469 </div>
2470 <p>
2472 <p>
2476 <pre>
2478 > testcaseLength = length testcase
2479 >
2481 > where arr = takeLast n testcase
2482 >
2483 </pre>
2484 </div>
2485 <p>
2488 <pre>
2489 *Main> quickCheck identityProp
2490 OK, passed 100 tests.
2491 </pre>
2492 <p>
2493 Let's be a bit more exhaustive:</p>
2496 <pre>
2497 >
2499 >
2500 </pre>
2501 </div>
2502 <p>
2505 <pre>
2506 *Main> deepCheck identityProp
2507 OK, passed 500 tests.
2508 </pre>
2510 <!-- ______________________________________________________________________ -->
2513 </div>
2515 <p>
2519 tag bits.</p>
2520 <p>
2524 the LSBit set. (Portability is relying on the fact that all known compilers place the
2527 </div>
2529 <!-- *********************************************************************** -->
2532 </div>
2533 <!-- *********************************************************************** -->
2539 <p>The Core LLVM classes are the primary means of representing the program
2540 being inspected or transformed. The core LLVM classes are defined in
2544 </div>
2546 <!-- ======================================================================= -->
2549 </div>
2557 subclasses. They are hidden because they offer no useful functionality beyond
2561 named, but this is not a requirement. There exists exactly
2562 one instance of a given shape at any one time. This allows type equality to
2563 be performed with address equality of the Type Instance. That is, given two
2565 </p>
2566 </div>
2568 <!-- _______________________________________________________________________ -->
2571 </div>
2575 <ul>
2579 floating point types.</li>
2582 an OpaqueType anywhere in its definition).</li>
2585 that don't have a size are abstract types, labels and void.</li>
2587 </ul>
2588 </div>
2590 <!-- _______________________________________________________________________ -->
2593 </div>
2595 <dl>
2597 <dd>Subclass of DerivedType that represents integer types of any bit width.
2600 <ul>
2602 type of a specific bit width.</li>
2604 type.</li>
2605 </ul>
2606 </dd>
2608 <dd>This is subclassed by ArrayType and PointerType
2609 <ul>
2611 of the elements in the sequential type. </li>
2612 </ul>
2613 </dd>
2615 <dd>This is a subclass of SequentialType and defines the interface for array
2616 types.
2617 <ul>
2619 elements in the array. </li>
2620 </ul>
2621 </dd>
2625 <dd>Subclass of SequentialType for vector types. A
2626 vector type is similar to an ArrayType but is distinguished because it is
2627 a first class type wherease ArrayType is not. Vector types are used for
2628 vector operations and are usually small vectors of of an integer or floating
2629 point type.</dd>
2633 <dd>Subclass of DerivedTypes for function types.
2634 <ul>
2636 function</li>
2638 return type of the function.</li>
2640 the type of the ith parameter.</li>
2642 number of formal parameters.</li>
2643 </ul>
2644 </dd>
2646 <dd>Sublcass of DerivedType for abstract types. This class
2647 defines no content and is used as a placeholder for some other type. Note
2648 that OpaqueType is used (temporarily) during type resolution for forward
2649 references of types. Once the referenced type is resolved, the OpaqueType
2650 is replaced with the actual type. OpaqueType can also be used for data
2651 abstraction. At link time opaque types can be resolved to actual types
2652 of the same name.</dd>
2653 </dl>
2654 </div>
2658 <!-- ======================================================================= -->
2661 </div>
2670 programs. An LLVM module is effectively either a translation unit of the
2671 original program or a combination of several translation units merged by the
2676 helpful member functions that try to make common operations easy.</p>
2678 </div>
2680 <!-- _______________________________________________________________________ -->
2683 </div>
2687 <ul>
2689 </ul>
2692 provide a name for it (probably based on the name of the translation unit).</p>
2694 <ul>
2701 <p>These are forwarding methods that make it easy to access the contents of
2703 list.</p></li>
2708 necessary to use when you need to update the list or perform a complex
2709 action that doesn't have a forwarding method.</p>
2712 </ul>
2714 <hr>
2716 <ul>
2724 <p> These are forwarding methods that make it easy to access the contents of
2732 use when you need to update the list or perform a complex action that
2733 doesn't have a forwarding method.</p>
2736 </ul>
2738 <hr>
2740 <ul>
2747 </ul>
2749 <hr>
2751 <ul>
2764 external declaration for the function and return it.</p></li>
2771 string.</p></li>
2778 name, true is returned and the <a
2780 </ul>
2782 </div>
2785 <!-- ======================================================================= -->
2788 </div>
2793 <br>
2797 base. It represents a typed value that may be used (among other things) as an
2805 for a program. For example, an incoming argument to a function (represented
2807 every instruction in the function that references the argument. To keep track
2813 methods, shown below.</p>
2817 method. In addition, all LLVM values can be named. The "name" of the
2821 <pre>
2823 </pre>
2824 </div>
2827 that the name of any value may be missing (an empty string), so names should
2829 debugging printouts), they should not be used to keep track of values or map
2833 <p>One important aspect of LLVM is that there is no distinction between an SSA
2834 variable and the operation that produces it. Because of this, any reference to
2835 the value produced by an instruction (or the value available as an incoming
2836 argument, for example) is represented as a direct pointer to the instance of
2837 the class that
2838 represents this value. Although this may take some getting used to, it
2839 simplifies the representation and makes it easier to manipulate.</p>
2841 </div>
2843 <!-- _______________________________________________________________________ -->
2846 </div>
2850 <ul>
2852 use-list<br>
2854 the use-list<br>
2856 value.<br>
2859 the use-list.<br>
2861 use-list.<br>
2863 element in the list.
2864 <p> These methods are the interface to access the def-use
2865 information in LLVM. As with all other iterators in LLVM, the naming
2867 </li>
2870 </li>
2876 </li>
2881 "<tt>V</tt>" instead. For example, if you detect that an instruction always
2882 produces a constant value (for example through constant folding), you can
2883 replace all uses of the instruction with the constant like this:</p>
2886 <pre>
2887 Inst->replaceAllUsesWith(ConstVal);
2888 </pre>
2889 </div>
2891 </ul>
2893 </div>
2895 <!-- ======================================================================= -->
2898 </div>
2902 <p>
2915 Single Assignment (SSA) form, there can only be one definition referred to,
2916 allowing this direct connection. This connection provides the use-def
2917 information in LLVM.</p>
2919 </div>
2921 <!-- _______________________________________________________________________ -->
2924 </div>
2929 an index access interface and through an iterator based interface.</p>
2931 <ul>
2935 convenient form for direct access.</p></li>
2938 list<br>
2940 the operand list.<br>
2942 operand list.
2943 <p> Together, these methods make up the iterator based interface to
2945 </ul>
2947 </div>
2949 <!-- ======================================================================= -->
2952 </div>
2963 instructions. It provides only a few methods, but is a very commonly used
2965 opcode (instruction type) and the parent <a
2967 into. To represent a specific type of instruction, one of many subclasses of
2976 file contains some meta-data about the various different types of instructions
2977 in LLVM. It describes the enum values that are used as opcodes (for example
2982 this file confuses doxygen, so these enum values don't show up correctly in the
2985 </div>
2987 <!-- _______________________________________________________________________ -->
2990 class</a>
2991 </div>
2993 <ul>
2995 <p>This subclasses represents all two operand instructions whose operands
2996 must be the same type, except for the comparison instructions.</p></li>
2999 common operations on cast instructions.</p>
3001 <p>This subclass respresents the two comparison instructions,
3005 <p>This subclass is the parent of all terminator instructions (those which
3006 can terminate a block).</p>
3007 </ul>
3008 </div>
3010 <!-- _______________________________________________________________________ -->
3013 class</a>
3014 </div>
3018 <ul>
3023 <p>Returns true if the instruction writes to memory, i.e. it is a
3028 <p>Returns another instance of the specified instruction, identical
3029 in all ways to the original except that the instruction has no parent
3031 and it has no name</p></li>
3032 </ul>
3034 </div>
3036 <!-- ======================================================================= -->
3039 </div>
3043 <p>Constant represents a base class for different types of constants. It
3044 is subclassed by ConstantInt, ConstantArray, etc. for representing
3046 a subclass, which represents the address of a global variable or function.
3047 </p>
3049 </div>
3051 <!-- _______________________________________________________________________ -->
3054 <ul>
3055 <li>ConstantInt : This subclass of Constant represents an integer constant of
3056 any width.
3057 <ul>
3059 value of this constant, an APInt value.</li>
3061 value to an int64_t via sign extension. If the value (not the bit width)
3062 of the APInt is too large to fit in an int64_t, an assertion will result.
3063 For this reason, use of this method is discouraged.</li>
3065 value to a uint64_t via zero extension. IF the value (not the bit width)
3066 of the APInt is too large to fit in a uint64_t, an assertion will result.
3067 For this reason, use of this method is discouraged.</li>
3070 The type is implied as the IntegerType that corresponds to the bit width
3073 Returns the ConstantInt object that represents the value provided by
3075 </ul>
3076 </li>
3077 <li>ConstantFP : This class represents a floating point constant.
3078 <ul>
3080 this constant. </li>
3081 </ul>
3082 </li>
3083 <li>ConstantArray : This represents a constant array.
3084 <ul>
3086 a vector of component constants that makeup this array. </li>
3087 </ul>
3088 </li>
3089 <li>ConstantStruct : This represents a constant struct.
3090 <ul>
3092 a vector of component constants that makeup this array. </li>
3093 </ul>
3094 </li>
3095 <li>GlobalValue : This represents either a global variable or a function. In
3096 either case, the value is a constant fixed address (after linking).
3097 </li>
3098 </ul>
3099 </div>
3102 <!-- ======================================================================= -->
3105 </div>
3112 Class</a><br>
3119 Because they are visible at global scope, they are also subject to linking with
3120 other globals defined in different translation units. To control the linking
3127 unit, and does not participate in linking. If it has external linkage, it is
3128 visible to external code, and does participate in linking. In addition to
3134 global is always a pointer to its contents. It is important to remember this
3139 the address of the first element of this array and the value of the
3147 </div>
3149 <!-- _______________________________________________________________________ -->
3152 class</a>
3153 </div>
3157 <ul>
3162 <p> </p>
3163 </li>
3166 GlobalValue is currently embedded into.</p></li>
3167 </ul>
3169 </div>
3171 <!-- ======================================================================= -->
3174 </div>
3187 actually one of the more complex classes in the LLVM heirarchy because it must
3195 ordering of the blocks in the function, which indicate how the code will be
3196 layed out by the backend. Additionally, the first <a
3199 block. There are no implicit exit nodes, and in fact there may be multiple exit
3203 function hasn't been linked in yet.</p>
3213 LLVM feature that is only used when you have to look up a value by name. Aside
3215 internally to make sure that there are not conflicts between the names of <a
3222 is its address (after linking) which is guaranteed to be constant.</p>
3223 </div>
3225 <!-- _______________________________________________________________________ -->
3228 class</a>
3229 </div>
3233 <ul>
3238 the the program. The constructor must specify the type of the function to
3239 create and what type of linkage the function should have. The <a
3241 specifies the formal arguments and return value for the function. The same
3244 in which the function is defined. If this argument is provided, the function
3245 will automatically be inserted into that module's list of
3246 functions.</p></li>
3251 function is "external", it does not have a body, and thus must be resolved
3252 by linking with a function defined in a different translation unit.</p></li>
3260 <p>These are forwarding methods that make it easy to access the contents of
3262 list.</p></li>
3267 is necessary to use when you need to update the list or perform a complex
3268 action that doesn't have a forwarding method.</p></li>
3271 iterator<br>
3277 <p>These are forwarding methods that make it easy to access the contents of
3279 list.</p></li>
3284 necessary to use when you need to update the list or perform a complex
3285 action that doesn't have a forwarding method.</p></li>
3290 function. Because the entry block for the function is always the first
3299 function.</p></li>
3305 </ul>
3307 </div>
3309 <!-- ======================================================================= -->
3312 </div>
3318 <br>
3320 Class</a><br>
3326 <p>Global variables are represented with the (suprise suprise)
3329 always referenced by their address (global values must live in memory, so their
3330 "name" refers to their constant address). See
3332 variables may have an initial value (which must be a
3334 they may be marked as "constant" themselves (indicating that their contents
3335 never change at runtime).</p>
3336 </div>
3338 <!-- _______________________________________________________________________ -->
3342 </div>
3346 <ul>
3351 <p>Create a new global variable of the specified type. If
3353 unchanging for the program. The Linkage parameter specifies the type of
3354 linkage (internal, external, weak, linkonce, appending) for the variable.
3355 If the linkage is InternalLinkage, WeakAnyLinkage, WeakODRLinkage,
3356 LinkOnceAnyLinkage or LinkOnceODRLinkage, then the resultant
3357 global variable will have internal linkage. AppendingLinkage concatenates
3358 together all instances (in different translation units) of the variable
3359 into a single variable but is only applicable to arrays. See
3361 further details on linkage types. Optionally an initializer, a name, and the
3362 module to put the variable into may be specified for the global variable as
3363 well.</p></li>
3367 <p>Returns true if this is a global variable that is known not to
3368 be modified at runtime.</p></li>
3377 to call this method if there is no initializer.</p></li>
3378 </ul>
3380 </div>
3383 <!-- ======================================================================= -->
3386 </div>
3393 Class</a><br>
3396 <p>This class represents a single entry multiple exit section of the code,
3397 commonly known as a basic block by the compiler community. The
3400 Matching the language definition, the last element of this list of instructions
3401 is always a terminator instruction (a subclass of the <a
3404 <p>In addition to tracking the list of instructions that make up the block, the
3413 </div>
3415 <!-- _______________________________________________________________________ -->
3418 class</a>
3419 </div>
3422 <ul>
3428 insertion into a function. The constructor optionally takes a name for the new
3431 automatically inserted at the end of the specified <a
3439 STL-style functions for accessing the instruction list.
3441 <p>These methods and typedefs are forwarding functions that have the same
3442 semantics as the standard library methods of the same names. These methods
3443 expose the underlying instruction list of a basic block in a way that is easy to
3444 manipulate. To get the full complement of container operations (including
3446 method.</p></li>
3450 <p>This method is used to get access to the underlying container that actually
3451 holds the Instructions. This method must be used when there isn't a forwarding
3453 to perform. Because there are no forwarding functions for "updating"
3454 operations, you need to use this if you want to update the contents of a
3460 embedded into, or a null pointer if it is homeless.</p></li>
3464 <p> Returns a pointer to the terminator instruction that appears at the end of
3466 instruction in the block is not a terminator, then a null pointer is
3467 returned.</p></li>
3469 </ul>
3471 </div>
3474 <!-- ======================================================================= -->
3477 </div>
3481 <p>This subclass of Value defines the interface for incoming formal
3482 arguments to a function. A Function maintains a list of its formal
3483 arguments. An argument has a pointer to the parent Function.</p>
3485 </div>
3487 <!-- *********************************************************************** -->
3488 <hr>
3489 <address>
3498 Last modified: $Date$
3499 </address>
3501 </body>
3502 </html>