| blob | eaed402d8af4e9c4aff507039bcb20cec115615e |
3 <html>
4 <head>
8 </head>
9 <body>
12 LLVM Programmer's Manual
13 </div>
15 <ol>
18 <ul>
20 <!--
21 <li>The <tt>-time-passes</tt> option</li>
22 <li>How to use the LLVM Makefile system</li>
23 <li>How to write a regression test</li>
25 -->
26 </ul>
27 </li>
29 <ul>
34 <ul>
37 </ul>
38 </li>
40 option</a>
41 <ul>
44 </ul>
45 </li>
47 option</a></li>
48 <!--
49 <li>The <tt>InstVisitor</tt> template
50 <li>The general graph API
51 -->
53 </ul>
54 </li>
56 <ul>
58 <ul>
67 </ul></li>
69 <ul>
79 </ul></li>
81 <ul>
88 </ul></li>
90 <!--<ul>
91 todo
94 <ul>
97 </ul></li>
98 </ul>
99 </li>
101 <ul>
103 <ul>
111 class pointer</a> </li>
113 complex example</a> </li>
115 the same way</a> </li>
117 use-def chains</a> </li>
119 successors of blocks</a></li>
120 </ul>
121 </li>
123 <ul>
130 </ul>
131 </li>
133 <!--
134 <li>Working with the Control Flow Graph
135 <ul>
136 <li>Accessing predecessors and successors of a <tt>BasicBlock</tt>
137 <li>
138 <li>
139 </ul>
140 -->
141 </ul>
142 </li>
145 <ul>
147 </a></li>
151 </ul>
152 </li>
155 <ul>
157 <ul>
162 </ul></li>
164 <li><a href="#SymbolTable">The <tt>ValueSymbolTable</tt> and <tt>TypeSymbolTable</tt> classes</a></li>
165 <li><a href="#UserLayout">The <tt>User</tt> and owned <tt>Use</tt> classes' memory layout</a></li>
166 </ul></li>
169 <ul>
173 <ul>
175 <ul>
178 <ul>
180 <ul>
183 </ul>
184 </li>
185 </ul>
186 </li>
187 </ul>
188 </li>
191 </ul>
192 </li>
193 </ul>
194 </li>
195 </ol>
204 </div>
206 <!-- *********************************************************************** -->
209 </div>
210 <!-- *********************************************************************** -->
214 <p>This document is meant to highlight some of the important classes and
215 interfaces available in the LLVM source-base. This manual is not
216 intended to explain what LLVM is, how it works, and what LLVM code looks
217 like. It assumes that you know the basics of LLVM and are interested
218 in writing transformations or otherwise analyzing or manipulating the
219 code.</p>
221 <p>This document should get you oriented so that you can find your
222 way in the continuously growing source code that makes up the LLVM
223 infrastructure. Note that this manual is not intended to serve as a
224 replacement for reading the source code, so if you think there should be
225 a method in one of these classes to do something, but it's not listed,
227 are provided to make this as easy as possible.</p>
229 <p>The first section of this document describes general information that is
230 useful to know when working in the LLVM infrastructure, and the second describes
231 the Core LLVM classes. In the future this manual will be extended with
232 information describing how to use extension libraries, such as dominator
233 information, CFG traversal routines, and useful utilities like the <tt><a
236 </div>
238 <!-- *********************************************************************** -->
241 </div>
242 <!-- *********************************************************************** -->
246 <p>This section contains general information that is useful if you are working
247 in the LLVM source-base, but that isn't specific to any particular API.</p>
249 </div>
251 <!-- ======================================================================= -->
254 </div>
258 <p>LLVM makes heavy use of the C++ Standard Template Library (STL),
259 perhaps much more than you are used to, or have seen before. Because of
260 this, you might want to do a little background reading in the
261 techniques used and capabilities of the library. There are many good
262 pages that discuss the STL, and several books on the subject that you
263 can get, so it will not be discussed in this document.</p>
267 <ol>
270 reference</a> - an excellent reference for the STL and other parts of the
271 standard C++ library.</li>
274 O'Reilly book in the making. It has a decent Standard Library
275 Reference that rivals Dinkumware's, and is unfortunately no longer free since the
276 book has been published.</li>
279 Questions</a></li>
282 Contains a useful <a
287 Page</a></li>
291 the book).</a></li>
293 </ol>
297 to write maintainable code more than where to put your curly braces.</p>
299 </div>
301 <!-- ======================================================================= -->
304 </div>
308 <ol>
310 Branch and Tag Primer</a></li>
312 static and shared libraries across platforms</a></li>
313 </ol>
315 </div>
317 <!-- *********************************************************************** -->
320 </div>
321 <!-- *********************************************************************** -->
325 <p>Here we highlight some LLVM APIs that are generally useful and good to
326 know about when writing transformations.</p>
328 </div>
330 <!-- ======================================================================= -->
334 </div>
338 <p>The LLVM source-base makes extensive use of a custom form of RTTI.
340 operator, but they don't have some drawbacks (primarily stemming from
342 have a v-table). Because they are used so often, you must know what they
343 do and how they work. All of these templates are defined in the <a
345 file (note that you very rarely have to include this file directly).</p>
347 <dl>
351 "<tt>instanceof</tt>" operator. It returns true or false depending on whether
352 a reference or pointer points to an instance of the specified class. This can
353 be very useful for constraint checking of various sorts (example below).</p>
354 </dd>
359 converts a pointer or reference from a base class to a derived class, causing
360 an assertion failure if it is not really an instance of the right type. This
361 should be used in cases where you have some information that makes you believe
366 <pre>
368 if (isa<<a href="#Constant">Constant</a>>(V) || isa<<a href="#Argument">Argument</a>>(V) || isa<<a href="#GlobalValue">GlobalValue</a>>(V))
369 return true;
373 }
374 </pre>
375 </div>
379 operator.</p>
381 </dd>
386 It checks to see if the operand is of the specified type, and if so, returns a
387 pointer to it (this operator does not work with references). If the operand is
388 not of the correct type, a null pointer is returned. Thus, this works very
392 statement like this:</p>
395 <pre>
396 if (<a href="#AllocationInst">AllocationInst</a> *AI = dyn_cast<<a href="#AllocationInst">AllocationInst</a>>(Val)) {
398 }
399 </pre>
400 </div>
404 statement, which is very convenient.</p>
409 blocks to check for lots of different variants of classes. If you find
410 yourself wanting to do this, it is much cleaner and more efficient to use the
413 </dd>
419 argument (which it then propagates). This can sometimes be useful, allowing
420 you to combine several null checks into one.</p></dd>
426 as an argument (which it then propagates). This can sometimes be useful,
427 allowing you to combine several null checks into one.</p></dd>
429 </dl>
431 <p>These five templates can be used with any classes, whether they have a
432 v-table or not. To add support for these templates, you simply need to add
434 to. Describing this is currently outside the scope of this document, but there
435 are lots of examples in the LLVM source base.</p>
437 </div>
440 <!-- ======================================================================= -->
444 </div>
448 <p>Although LLVM generally does not do much string manipulation, we do have
449 several important APIs which take strings. Two important examples are the
450 Value class -- which has names for instructions, functions, etc. -- and the
451 StringMap class which is used extensively in LLVM and Clang.</p>
453 <p>These are generic classes, and they need to be able to accept strings which
454 may have embedded null characters. Therefore, they cannot simply take
456 clients to perform a heap allocation which is usually unnecessary. Instead,
460 </div>
462 <!-- _______________________________________________________________________ -->
465 </div>
470 (a character array and a length) and supports the common operations available
473 <p>It can be implicitly constructed using a C style null-terminated string,
478 iterator find(const StringRef &Key);
479 </div>
484 <pre>
488 </pre>
489 </div>
495 for more information.</p>
498 pointers to external memory it is not generally safe to store an instance of the
499 class (unless you know that the external storage will not be freed).</p>
501 </div>
503 <!-- _______________________________________________________________________ -->
506 </div>
511 strings. For example, a common LLVM paradigm is to name one instruction based on
512 the name of another instruction with a suffix, for example:</p>
515 <pre>
517 </pre>
518 </div>
522 which points to temporary (stack allocated) objects. Twines can be implicitly
523 constructed as the result of the plus operator applied to strings (i.e., a C
525 actual concatentation of strings until it is actually required, at which point
526 it can be efficiently rendered directly into a character array. This avoids
527 unnecessary heap allocation involved in constructing the temporary results of
528 string concatenation. See
530 for more information.</p>
533 and should almost never be stored or mentioned directly. They are intended
534 solely for use when defining a function which should be able to efficiently
535 accept concatenated strings.</p>
537 </div>
540 <!-- ======================================================================= -->
543 </div>
547 <p>Often when working on your pass you will put a bunch of debugging printouts
548 and other code into your pass. After you get it working, you want to remove
549 it, but you may need it again in the future (to work out new bugs that you run
550 across).</p>
552 <p> Naturally, because of this, you don't want to delete the debug printouts,
553 but you don't want them to always be noisy. A standard compromise is to comment
554 them out, allowing you to enable them if you need them in the future.</p>
558 this problem. Basically, you can put arbitrary code into the argument of the
563 <pre>
565 </pre>
566 </div>
571 <pre>
575 I am here!
576 </pre>
577 </div>
580 to not have to create "yet another" command line option for the debug output for
582 so they do not cause a performance impact at all (for the same reason, they
583 should also not contain side-effects!).</p>
586 enable or disable it directly in gdb. Just use "<tt>set DebugFlag=0</tt>" or
587 "<tt>set DebugFlag=1</tt>" from the gdb if the program is running. If the
588 program hasn't been started yet, you can always just run it with
591 </div>
593 <!-- _______________________________________________________________________ -->
597 </div>
603 generator). If you want to enable debug information with more fine-grained
605 option as follows:</p>
608 <pre>
609 #undef DEBUG_TYPE
611 #define DEBUG_TYPE "foo"
613 #undef DEBUG_TYPE
614 #define DEBUG_TYPE "bar"
616 #undef DEBUG_TYPE
617 #define DEBUG_TYPE ""
619 </pre>
620 </div>
625 <pre>
629 No debug type
630 'foo' debug type
631 'bar' debug type
632 No debug type (2)
634 'foo' debug type
636 'bar' debug type
637 </pre>
638 </div>
641 a file, to specify the debug type for the entire module (if you do this before
644 "bar", because there is no system in place to ensure that names do not
645 conflict. If two different modules use the same string, they will all be turned
646 on when the name is specified. This allows, for example, all debug information
648 even if the source lives in multiple files.</p>
652 statement. It takes an additional first parameter, which is the type to use. For
653 example, the preceeding example could be written as:</p>
657 <pre>
662 </pre>
663 </div>
665 </div>
667 <!-- ======================================================================= -->
670 option</a>
671 </div>
678 keep track of what the LLVM compiler is doing and how effective various
679 optimizations are. It is useful to see what optimizations are contributing to
680 making a particular program run faster.</p>
682 <p>Often you may run your pass on some big program, and you're interested to see
683 how many times it makes a certain transformation. Although you can do this with
684 hand inspection, or some ad-hoc method, this is a real pain and not very useful
686 keep track of this information, and the calculated information is presented in a
687 uniform manner with the rest of the passes being executed.</p>
690 it are as follows:</p>
692 <ol>
696 <pre>
697 #define <a href="#DEBUG_TYPE">DEBUG_TYPE</a> "mypassname" <i>// This goes before any #includes.</i>
698 STATISTIC(NumXForms, "The # of times I did stuff");
699 </pre>
700 </div>
703 specified by the first argument. The pass name is taken from the DEBUG_TYPE
704 macro, and the description is taken from the second argument. The variable
710 <pre>
712 </pre>
713 </div>
715 </li>
716 </ol>
722 <pre>
725 </pre>
726 </div>
729 suite, it gives a report that looks like this:</p>
732 <pre>
733 7646 bitcodewriter - Number of normal instructions
734 725 bitcodewriter - Number of oversized instructions
735 129996 bitcodewriter - Number of bitcode bytes written
736 2817 raise - Number of insts DCEd or constprop'd
737 3213 raise - Number of cast-of-self removed
738 5046 raise - Number of expression trees converted
739 75 raise - Number of other getelementptr's formed
740 138 raise - Number of load/store peepholes
741 42 deadtypeelim - Number of unused typenames removed from symtab
742 392 funcresolve - Number of varargs functions resolved
743 27 globaldce - Number of global variables removed
744 2 adce - Number of basic blocks removed
745 134 cee - Number of branches revectored
746 49 cee - Number of setcc instruction eliminated
747 532 gcse - Number of loads removed
748 2919 gcse - Number of instructions removed
749 86 indvars - Number of canonical indvars added
750 87 indvars - Number of aux indvars removed
751 25 instcombine - Number of dead inst eliminate
752 434 instcombine - Number of insts combined
753 248 licm - Number of load insts hoisted
754 1298 licm - Number of insts hoisted to a loop pre-header
755 3 licm - Number of insts hoisted to multiple loop preds (bad, no loop pre-header)
756 75 mem2reg - Number of alloca's promoted
757 1444 cfgsimplify - Number of blocks simplified
758 </pre>
759 </div>
761 <p>Obviously, with so many optimizations, having a unified framework for this
762 stuff is very nice. Making your pass fit well into the framework makes it more
763 maintainable and useful.</p>
765 </div>
767 <!-- ======================================================================= -->
770 </div>
774 <p>Several of the important data structures in LLVM are graphs: for example
778 DAGs</a>. In many cases, while debugging various parts of the compiler, it is
779 nice to instantly visualize these graphs.</p>
781 <p>LLVM provides several callbacks that are available in a debug build to do
783 the current LLVM tool will pop up a window containing the CFG for the function
784 where each basic block is a node in the graph, and each node contains the
785 instructions in the block. Similarly, there also exists
790 up a window. Alternatively, you can sprinkle calls to these functions in your
791 code in places you want to debug.</p>
793 <p>Getting this to work requires a small amount of configuration. On Unix
795 toolkit, and make sure 'dot' and 'gv' are in your path. If you are running on
796 Mac OS/X, download and install the Mac OS/X <a
799 it) to your path. Once in your system and path are set up, rerun the LLVM
800 configure script and rebuild LLVM to enable this functionality.</p>
806 specified color (choices of colors can be found at <a
808 complex node attributes can be provided with <tt>call
811 Attributes</a>.) If you want to restart and clear all the current graph
814 </div>
816 <!-- *********************************************************************** -->
819 </div>
820 <!-- *********************************************************************** -->
825 and we commonly use STL data structures. This section describes the trade-offs
826 you should consider when you pick one.</p>
828 <p>
829 The first step is a choose your own adventure: do you want a sequential
830 container, a set-like container, or a map-like container? The most important
831 thing when choosing a container is the algorithmic properties of how you plan to
832 access the container. Based on that, you should use:</p>
834 <ul>
836 of an value based on another value. Map-like containers also support
837 efficient queries for containment (whether a key is in the map). Map-like
838 containers generally do not support efficient reverse mapping (values to
839 keys). If you need that, use two maps. Some map-like containers also
840 support efficient iteration through the keys in sorted order. Map-like
841 containers are the most expensive sort, only use them if you need one of
842 these capabilities.</li>
845 stuff into a container that automatically eliminates duplicates. Some
846 set-like containers support efficient iteration through the elements in
847 sorted order. Set-like containers are more expensive than sequential
848 containers.
849 </li>
852 the most efficient way to add elements and keeps track of the order they are
853 added to the collection. They permit duplicates and support efficient
854 iteration, but do not support efficient look-up based on a key.
855 </li>
858 container or reference structure that is used for character or byte
859 arrays.</li>
862 perform set operations on sets of numeric id's, while automatically
863 eliminating duplicates. Bit containers require a maximum of 1 bit for each
864 identifier you want to store.
865 </li>
866 </ul>
868 <p>
869 Once the proper category of container is determined, you can fine tune the
870 memory use, constant factors, and cache behaviors of access by intelligently
871 picking a member of the category. Note that constant factors and cache behavior
872 can be a big deal. If you have a vector that usually only contains a few
873 elements (but could contain many), for example, it's much better to use
875 . Doing so avoids (relatively) expensive malloc/free calls, which dwarf the
876 cost of adding the elements to the container. </p>
878 </div>
880 <!-- ======================================================================= -->
883 </div>
886 There are a variety of sequential containers available for you, based on your
887 needs. Pick the first in this section that will do what you want.
888 </div>
890 <!-- _______________________________________________________________________ -->
893 </div>
896 <p>Fixed size arrays are very simple and very fast. They are good if you know
897 exactly how many elements you have, or you have a (low) upper bound on how many
898 you have.</p>
899 </div>
901 <!-- _______________________________________________________________________ -->
904 </div>
907 <p>Heap allocated arrays (new[] + delete[]) are also simple. They are good if
908 the number of elements is variable, if you know how many elements you will need
909 before the array is allocated, and if the array is usually large (if not,
911 allocated array is the cost of the new/delete (aka malloc/free). Also note that
912 if you are allocating an array of a type with a constructor, the constructor and
913 destructors will be run for every element in the array (re-sizable vectors only
914 construct those elements actually used).</p>
915 </div>
917 <!-- _______________________________________________________________________ -->
920 </div>
925 it supports efficient iteration, lays out elements in memory order (so you can
926 do pointer arithmetic between elements), supports efficient push_back/pop_back
927 operations, supports efficient random access to its elements, etc.</p>
929 <p>The advantage of SmallVector is that it allocates space for
931 the SmallVector is dynamically smaller than N, no malloc is performed. This can
932 be a big win in cases where the malloc/free call is far more expensive than the
933 code that fiddles around with the elements.</p>
936 predecessors/successors of a block is usually less than 8). On the other hand,
937 this makes the size of the SmallVector itself large, so you don't want to
938 allocate lots of them (doing so will waste a lot of space). As such,
939 SmallVectors are most useful when on the stack.</p>
941 <p>SmallVector also provides a nice portable and efficient replacement for
944 </div>
946 <!-- _______________________________________________________________________ -->
949 </div>
952 <p>
953 std::vector is well loved and respected. It is useful when SmallVector isn't:
954 when the size of the vector is often large (thus the small optimization will
955 rarely be a benefit) or if you will be allocating many instances of the vector
956 itself (which would waste space for elements that aren't in the container).
957 vector is also useful when interfacing with code that expects vectors :).
958 </p>
963 <pre>
964 for ( ... ) {
966 use V;
967 }
968 </pre>
969 </div>
974 <pre>
976 for ( ... ) {
977 use V;
978 V.clear();
979 }
980 </pre>
981 </div>
983 <p>Doing so will save (at least) one heap allocation and free per iteration of
984 the loop.</p>
986 </div>
988 <!-- _______________________________________________________________________ -->
991 </div>
994 <p>std::deque is, in some senses, a generalized version of std::vector. Like
995 std::vector, it provides constant time random access and other similar
996 properties, but it also provides efficient access to the front of the list. It
997 does not guarantee continuity of elements within memory.</p>
999 <p>In exchange for this extra flexibility, std::deque has significantly higher
1000 constant factor costs than std::vector. If possible, use std::vector or
1001 something cheaper.</p>
1002 </div>
1004 <!-- _______________________________________________________________________ -->
1007 </div>
1010 <p>std::list is an extremely inefficient class that is rarely useful.
1011 It performs a heap allocation for every element inserted into it, thus having an
1012 extremely high constant factor, particularly for small data types. std::list
1013 also only supports bidirectional iteration, not random access iteration.</p>
1015 <p>In exchange for this high cost, std::list supports efficient access to both
1016 ends of the list (like std::deque, but unlike std::vector or SmallVector). In
1017 addition, the iterator invalidation characteristics of std::list are stronger
1018 than that of a vector class: inserting or removing an element into the list does
1019 not invalidate iterator or pointers to other elements in the list.</p>
1020 </div>
1022 <!-- _______________________________________________________________________ -->
1025 </div>
1029 intrusive, because it requires the element to store and provide access to the
1030 prev/next pointers for the list.</p>
1034 provides some novel characteristics. In particular, it can efficiently store
1035 polymorphic objects, the traits class is informed when an element is inserted or
1037 constant-time splice operation.</p>
1039 <p>These properties are exactly what we want for things like
1043 Related classes of interest are explained in the following subsections:
1044 <ul>
1049 </ul>
1050 </div>
1052 <!-- _______________________________________________________________________ -->
1055 </div>
1060 publicly derive from this traits class.</p>
1061 </div>
1063 <!-- _______________________________________________________________________ -->
1066 </div>
1070 supports a slightly narrower interface. Notably, inserters from
1074 used for a wide variety of customizations.</p>
1075 </div>
1077 <!-- _______________________________________________________________________ -->
1080 </div>
1085 in the default manner.</p>
1090 </div>
1092 <!-- _______________________________________________________________________ -->
1095 </div>
1099 citizen in the C++ ecosystem, it needs to support the standard container
1104 <p>The only sensible solution to this problem is to allocate a so-called
1106 iterator, providing the back-link to the last element. However conforming to the
1108 also must not be dereferenced.</p>
1111 how to allocate and store the sentinel. The corresponding policy is dictated
1113 whenever the need for a sentinel arises.</p>
1115 <p>While the default policy is sufficient in most cases, it may break down when
1118 is wasted. To alleviate the situation with numerous and voluminous
1124 arithmetic is used to obtain the sentinel, which is relative to the
1126 extra pointer, which serves as the back-link of the sentinel. This is the only
1127 field in the ghostly sentinel which can be legally accessed.</p>
1128 </div>
1130 <!-- _______________________________________________________________________ -->
1133 </div>
1138 <p>There are also various STL adapter classes such as std::queue,
1139 std::priority_queue, std::stack, etc. These provide simplified access to an
1140 underlying container but don't affect the cost of the container itself.</p>
1142 </div>
1145 <!-- ======================================================================= -->
1148 </div>
1152 <p>Set-like containers are useful when you need to canonicalize multiple values
1153 into a single representation. There are several different choices for how to do
1154 this, providing various trade-offs.</p>
1156 </div>
1159 <!-- _______________________________________________________________________ -->
1162 </div>
1166 <p>If you intend to insert a lot of elements, then do a lot of queries, a
1167 great approach is to use a vector (or other sequential container) with
1168 std::sort+std::unique to remove duplicates. This approach works really well if
1169 your usage pattern has these two distinct phases (insert then query), and can be
1171 </p>
1173 <p>
1174 This combination provides the several nice properties: the result data is
1175 contiguous in memory (good for cache locality), has few allocations, is easy to
1176 address (iterators in the final vector are just indices or pointers), and can be
1177 efficiently queried with a standard binary or radix search.</p>
1179 </div>
1181 <!-- _______________________________________________________________________ -->
1184 </div>
1188 <p>If you have a set-like data structure that is usually small and whose elements
1190 has space for N elements in place (thus, if the set is dynamically smaller than
1191 N, no malloc traffic is required) and accesses them with a simple linear search.
1192 When the set grows beyond 'N' elements, it allocates a more expensive representation that
1193 guarantees efficient access (for most types, it falls back to std::set, but for
1194 pointers it uses something far better, <a
1197 <p>The magic of this class is that it handles small sets extremely efficiently,
1198 but gracefully handles extremely large sets without loss of efficiency. The
1199 drawback is that the interface is quite small: it supports insertion, queries
1200 and erasing, but does not support iteration.</p>
1202 </div>
1204 <!-- _______________________________________________________________________ -->
1207 </div>
1211 <p>SmallPtrSet has all the advantages of SmallSet (and a SmallSet of pointers is
1212 transparently implemented with a SmallPtrSet), but also supports iterators. If
1213 more than 'N' insertions are performed, a single quadratically
1214 probed hash table is allocated and grows as needed, providing extremely
1215 efficient access (constant time insertion/deleting/queries with low constant
1216 factors) and is very stingy with malloc traffic.</p>
1218 <p>Note that, unlike std::set, the iterators of SmallPtrSet are invalidated
1219 whenever an insertion occurs. Also, the values visited by the iterators are not
1220 visited in sorted order.</p>
1222 </div>
1224 <!-- _______________________________________________________________________ -->
1227 </div>
1231 <p>
1232 DenseSet is a simple quadratically probed hash table. It excels at supporting
1233 small values: it uses a single allocation to hold all of the pairs that
1234 are currently inserted in the set. DenseSet is a great way to unique small
1235 values that are not simple pointers (use <a
1237 the same requirements for the value type that <a
1239 </p>
1241 </div>
1243 <!-- _______________________________________________________________________ -->
1246 </div>
1250 <p>
1251 FoldingSet is an aggregate class that is really good at uniquing
1252 expensive-to-create or polymorphic objects. It is a combination of a chained
1253 hash table with intrusive links (uniqued objects are required to inherit from
1255 its ID process.</p>
1258 a complex object (for example, a node in the code generator). The client has a
1259 description of *what* it wants to generate (it knows the opcode and all the
1260 operands), but we don't want to 'new' a node, then try inserting it into a set
1261 only to find out it already exists, at which point we would have to delete it
1262 and return the node that already exists.
1263 </p>
1265 <p>To support this style of client, FoldingSet perform a query with a
1266 FoldingSetNodeID (which wraps SmallVector) that can be used to describe the
1267 element that we want to query for. The query either returns the element
1268 matching the ID or it returns an opaque ID that indicates where insertion should
1269 take place. Construction of the ID usually does not require heap traffic.</p>
1271 <p>Because FoldingSet uses intrusive links, it can support polymorphic objects
1272 in the set (for example, you can have SDNode instances mixed with LoadSDNodes).
1273 Because the elements are individually allocated, pointers to the elements are
1274 stable: inserting or removing elements does not invalidate any pointers to other
1275 elements.
1276 </p>
1278 </div>
1280 <!-- _______________________________________________________________________ -->
1283 </div>
1288 many things but great at nothing. std::set allocates memory for each element
1289 inserted (thus it is very malloc intensive) and typically stores three pointers
1290 per element in the set (thus adding a large amount of per-element space
1291 overhead). It offers guaranteed log(n) performance, which is not particularly
1292 fast from a complexity standpoint (particularly if the elements of the set are
1293 expensive to compare, like strings), and has extremely high constant factors for
1294 lookup, insertion and removal.</p>
1296 <p>The advantages of std::set are that its iterators are stable (deleting or
1297 inserting an element from the set does not affect iterators or pointers to other
1298 elements) and that iteration over the set is guaranteed to be in sorted order.
1299 If the elements in the set are large, then the relative overhead of the pointers
1300 and malloc traffic is not a big deal, but if the elements of the set are small,
1301 std::set is almost never a good choice.</p>
1303 </div>
1305 <!-- _______________________________________________________________________ -->
1308 </div>
1313 Container</a>. The important property
1314 that this provides is efficient insertion with uniquing (duplicate elements are
1315 ignored) with iteration support. It implements this by inserting elements into
1316 both a set-like container and the sequential container, using the set-like
1317 container for uniquing and the sequential container for iteration.
1318 </p>
1320 <p>The difference between SetVector and other sets is that the order of
1321 iteration is guaranteed to match the order of insertion into the SetVector.
1322 This property is really important for things like sets of pointers. Because
1323 pointer values are non-deterministic (e.g. vary across runs of the program on
1324 different machines), iterating over the pointers in the set will
1325 not be in a well-defined order.</p>
1327 <p>
1328 The drawback of SetVector is that it requires twice as much space as a normal
1329 set and has the sum of constant factors from the set-like container and the
1330 sequential container that it uses. Use it *only* if you need to iterate over
1331 the elements in a deterministic order. SetVector is also expensive to delete
1332 elements out of (linear time), unless you use it's "pop_back" method, which is
1333 faster.
1334 </p>
1336 <p>SetVector is an adapter class that defaults to using std::vector and std::set
1337 for the underlying containers, so it is quite expensive. However,
1339 defaults to using a SmallVector and SmallSet of a specified size. If you use
1340 this, and if your sets are dynamically smaller than N, you will save a lot of
1341 heap traffic.</p>
1343 </div>
1345 <!-- _______________________________________________________________________ -->
1348 </div>
1352 <p>
1354 retains a unique ID for each element inserted into the set. It internally
1355 contains a map and a vector, and it assigns a unique ID for each value inserted
1356 into the set.</p>
1358 <p>UniqueVector is very expensive: its cost is the sum of the cost of
1359 maintaining both the map and vector, it has high complexity, high constant
1360 factors, and produces a lot of malloc traffic. It should be avoided.</p>
1362 </div>
1365 <!-- _______________________________________________________________________ -->
1368 </div>
1372 <p>
1373 The STL provides several other options, such as std::multiset and the various
1374 "hash_set" like containers (whether from C++ TR1 or from the SGI library). We
1375 never use hash_set and unordered_set because they are generally very expensive
1376 (each insertion requires a malloc) and very non-portable.
1377 </p>
1379 <p>std::multiset is useful if you're not interested in elimination of
1380 duplicates, but has all the drawbacks of std::set. A sorted vector (where you
1381 don't delete duplicate entries) or some other approach is almost always
1382 better.</p>
1384 </div>
1386 <!-- ======================================================================= -->
1389 </div>
1392 Map-like containers are useful when you want to associate data to a key. As
1393 usual, there are a lot of different ways to do this. :)
1394 </div>
1396 <!-- _______________________________________________________________________ -->
1399 </div>
1403 <p>
1404 If your usage pattern follows a strict insert-then-query approach, you can
1406 for set-like containers</a>. The only difference is that your query function
1407 (which uses std::lower_bound to get efficient log(n) lookup) should only compare
1408 the key, not both the key and value. This yields the same advantages as sorted
1409 vectors for sets.
1410 </p>
1411 </div>
1413 <!-- _______________________________________________________________________ -->
1416 </div>
1420 <p>
1421 Strings are commonly used as keys in maps, and they are difficult to support
1422 efficiently: they are variable length, inefficient to hash and compare when
1423 long, expensive to copy, etc. StringMap is a specialized container designed to
1424 cope with these issues. It supports mapping an arbitrary range of bytes to an
1425 arbitrary other object.</p>
1427 <p>The StringMap implementation uses a quadratically-probed hash table, where
1428 the buckets store a pointer to the heap allocated entries (and some other
1429 stuff). The entries in the map must be heap allocated because the strings are
1430 variable length. The string data (key) and the element object (value) are
1431 stored in the same allocation with the string data immediately after the element
1432 object. This container guarantees the "<tt>(char*)(&Value+1)</tt>" points
1433 to the key string for a value.</p>
1435 <p>The StringMap is very fast for several reasons: quadratic probing is very
1436 cache efficient for lookups, the hash value of strings in buckets is not
1437 recomputed when lookup up an element, StringMap rarely has to touch the
1438 memory for unrelated objects when looking up a value (even when hash collisions
1439 happen), hash table growth does not recompute the hash values for strings
1440 already in the table, and each pair in the map is store in a single allocation
1441 (the string data is stored in the same allocation as the Value of a pair).</p>
1443 <p>StringMap also provides query methods that take byte ranges, so it only ever
1444 copies a string if a value is inserted into the table.</p>
1445 </div>
1447 <!-- _______________________________________________________________________ -->
1450 </div>
1453 <p>
1454 IndexedMap is a specialized container for mapping small dense integers (or
1455 values that can be mapped to small dense integers) to some other type. It is
1456 internally implemented as a vector with a mapping function that maps the keys to
1457 the dense integer range.
1458 </p>
1460 <p>
1461 This is useful for cases like virtual registers in the LLVM code generator: they
1462 have a dense mapping that is offset by a compile-time constant (the first
1463 virtual register ID).</p>
1465 </div>
1467 <!-- _______________________________________________________________________ -->
1470 </div>
1474 <p>
1475 DenseMap is a simple quadratically probed hash table. It excels at supporting
1476 small keys and values: it uses a single allocation to hold all of the pairs that
1477 are currently inserted in the map. DenseMap is a great way to map pointers to
1478 pointers, or map other small types to each other.
1479 </p>
1481 <p>
1482 There are several aspects of DenseMap that you should be aware of, however. The
1483 iterators in a densemap are invalidated whenever an insertion occurs, unlike
1484 map. Also, because DenseMap allocates space for a large number of key/value
1485 pairs (it starts with 64 by default), it will waste a lot of space if your keys
1486 or values are large. Finally, you must implement a partial specialization of
1487 DenseMapInfo for the key that you want, if it isn't already supported. This
1488 is required to tell DenseMap about two special marker values (which can never be
1489 inserted into the map) that it needs internally.</p>
1491 </div>
1493 <!-- _______________________________________________________________________ -->
1496 </div>
1500 <p>
1502 a single allocation per pair inserted into the map, it offers log(n) lookup with
1503 an extremely large constant factor, imposes a space penalty of 3 pointers per
1504 pair in the map, etc.</p>
1506 <p>std::map is most useful when your keys or values are very large, if you need
1507 to iterate over the collection in sorted order, or if you need stable iterators
1508 into the map (i.e. they don't get invalidated if an insertion or deletion of
1509 another element takes place).</p>
1511 </div>
1513 <!-- _______________________________________________________________________ -->
1516 </div>
1520 <p>
1521 The STL provides several other options, such as std::multimap and the various
1522 "hash_map" like containers (whether from C++ TR1 or from the SGI library). We
1523 never use hash_set and unordered_set because they are generally very expensive
1524 (each insertion requires a malloc) and very non-portable.</p>
1526 <p>std::multimap is useful if you want to map a key to multiple values, but has
1527 all the drawbacks of std::map. A sorted vector or some other approach is almost
1528 always better.</p>
1530 </div>
1532 <!-- ======================================================================= -->
1535 </div>
1539 <p>
1540 TODO: const char* vs stringref vs smallstring vs std::string. Describe twine,
1541 xref to #string_apis.
1542 </p>
1544 </div>
1546 <!-- ======================================================================= -->
1549 </div>
1552 <p>Unlike the other containers, there are only two bit storage containers, and
1553 choosing when to use each is relatively straightforward.</p>
1555 <p>One additional option is
1557 implementation in many common compilers (e.g. commonly available versions of
1558 GCC) is extremely inefficient and 2) the C++ standards committee is likely to
1559 deprecate this container and/or change it significantly somehow. In any case,
1560 please don't use it.</p>
1561 </div>
1563 <!-- _______________________________________________________________________ -->
1566 </div>
1569 <p> The BitVector container provides a fixed size set of bits for manipulation.
1570 It supports individual bit setting/testing, as well as set operations. The set
1571 operations take time O(size of bitvector), but operations are performed one word
1572 at a time, instead of one bit at a time. This makes the BitVector very fast for
1573 set operations compared to other containers. Use the BitVector when you expect
1574 the number of set bits to be high (IE a dense set).
1575 </p>
1576 </div>
1578 <!-- _______________________________________________________________________ -->
1581 </div>
1584 <p> The SparseBitVector container is much like BitVector, with one major
1585 difference: Only the bits that are set, are stored. This makes the
1586 SparseBitVector much more space efficient than BitVector when the set is sparse,
1587 as well as making set operations O(number of set bits) instead of O(size of
1588 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
1589 (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).
1590 </p>
1591 </div>
1593 <!-- *********************************************************************** -->
1596 </div>
1597 <!-- *********************************************************************** -->
1601 <p>This section describes how to perform some very simple transformations of
1602 LLVM code. This is meant to give examples of common idioms used, showing the
1604 you should also read about the main classes that you will be working with. The
1606 and descriptions of the main classes that you should know about.</p>
1608 </div>
1610 <!-- NOTE: this section should be heavy on example code -->
1611 <!-- ======================================================================= -->
1614 </div>
1618 <p>The LLVM compiler infrastructure have many different data structures that may
1619 be traversed. Following the example of the C++ standard template library, the
1620 techniques used to traverse these various data structures are all basically the
1623 function returns an iterator pointing to one past the last valid element of the
1625 between the two operations.</p>
1627 <p>Because the pattern for iteration is common across many different aspects of
1628 the program representation, the standard template library algorithms may be used
1629 on them, and it is easier to remember how to iterate. First we show a few common
1630 examples of the data structures that need to be traversed. Other data
1631 structures are traversed in very similar ways.</p>
1633 </div>
1635 <!-- _______________________________________________________________________ -->
1640 </div>
1645 transform in some way; in particular, you'd like to manipulate its
1652 <pre>
1659 </pre>
1660 </div>
1662 <p>Note that i can be used as if it were a pointer for the purposes of
1664 because the indirection operator is overloaded for the iterator
1668 </div>
1670 <!-- _______________________________________________________________________ -->
1675 </div>
1680 easy to iterate over the individual instructions that make up
1685 <pre>
1691 </pre>
1692 </div>
1694 <p>However, this isn't really the best way to print out the contents of a
1696 anything you'll care about, you could have just invoked the print routine on the
1699 </div>
1701 <!-- _______________________________________________________________________ -->
1706 </div>
1715 small example that shows how to dump all instructions in a function to the standard error stream:<p>
1718 <pre>
1722 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
1724 </pre>
1725 </div>
1728 work list with its initial contents. For example, if you wanted to
1730 F, all you would need to do is something like:</p>
1733 <pre>
1737 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
1738 worklist.insert(&*I);
1739 </pre>
1740 </div>
1745 </div>
1747 <!-- _______________________________________________________________________ -->
1750 vice-versa)</a>
1751 </div>
1755 <p>Sometimes, it'll be useful to grab a reference (or pointer) to a class
1756 instance when all you've got at hand is an iterator. Well, extracting
1757 a reference or a pointer from an iterator is very straight-forward.
1762 <pre>
1765 const Instruction& inst = *j;
1766 </pre>
1767 </div>
1769 <p>However, the iterators you'll be working with in the LLVM framework are
1770 special: they will automatically convert to a ptr-to-instance type whenever they
1771 need to. Instead of dereferencing the iterator and then taking the address of
1772 the result, you can simply assign the iterator to the proper pointer type and
1773 you get the dereference and address-of operation as a result of the assignment
1774 (behind the scenes, this is a result of overloading casting mechanisms). Thus
1775 the last line of the last example,</p>
1778 <pre>
1779 Instruction *pinst = &*i;
1780 </pre>
1781 </div>
1786 <pre>
1787 Instruction *pinst = i;
1788 </pre>
1789 </div>
1791 <p>It's also possible to turn a class pointer into the corresponding iterator,
1792 and this is a constant time operation (very efficient). The following code
1793 snippet illustrates use of the conversion constructors provided by LLVM
1794 iterators. By using these, you can explicitly grab the iterator of something
1795 without actually obtaining it via iteration over some structure:</p>
1798 <pre>
1799 void printNextInstruction(Instruction* inst) {
1800 BasicBlock::iterator it(inst);
1803 }
1804 </pre>
1805 </div>
1807 </div>
1809 <!--_______________________________________________________________________-->
1812 example</a>
1813 </div>
1817 <p>Say that you're writing a FunctionPass and would like to count all the
1821 much more straight-forward manner, but this example will allow us to explore how
1823 is what we want to do:</p>
1826 <pre>
1827 initialize callCounter to zero
1828 for each Function f in the Module
1829 for each BasicBlock b in f
1830 for each Instruction i in b
1831 if (i is a CallInst and calls the given function)
1832 increment callCounter
1833 </pre>
1834 </div>
1836 <p>And the actual code is (remember, because we're writing a
1841 <pre>
1842 Function* targetFunc = ...;
1844 class OurFunctionPass : public FunctionPass {
1845 public:
1846 OurFunctionPass(): callCounter(0) { }
1848 virtual runOnFunction(Function& F) {
1849 for (Function::iterator b = F.begin(), be = F.end(); b != be; ++b) {
1856 if (callInst->getCalledFunction() == targetFunc)
1857 ++callCounter;
1858 }
1859 }
1860 }
1861 }
1863 private:
1864 unsigned callCounter;
1865 };
1866 </pre>
1867 </div>
1869 </div>
1871 <!--_______________________________________________________________________-->
1874 </div>
1878 <p>You may have noticed that the previous example was a bit oversimplified in
1879 that it did not deal with call sites generated by 'invoke' instructions. In
1880 this, and in other situations, you may find that you want to treat
1883 less closely-related things. For these cases, LLVM provides a handy wrapper
1884 class called <a
1891 reference and it should not be dynamically allocated or deallocated using
1893 assignable and constructable, with costs equivalents to that of a bare pointer.
1894 If you look at its definition, it has only a single pointer member.</p>
1896 </div>
1898 <!--_______________________________________________________________________-->
1901 </div>
1915 <pre>
1916 Function *F = ...;
1922 }
1923 </pre>
1924 </div>
1931 all of the values that a particular instruction uses (that is, the operands of
1935 <pre>
1936 Instruction *pi = ...;
1939 Value *v = *i;
1941 }
1942 </pre>
1943 </div>
1945 <!--
1946 def-use chains ("finding all users of"): Value::use_begin/use_end
1947 use-def chains ("finding all values used"): User::op_begin/op_end [op=operand]
1948 -->
1950 </div>
1952 <!--_______________________________________________________________________-->
1955 successors of blocks</a>
1956 </div>
1960 <p>Iterating over the predecessors and successors of a block is quite easy
1962 this to iterate over all predecessors of BB:</p>
1965 <pre>
1966 #include "llvm/Support/CFG.h"
1967 BasicBlock *BB = ...;
1969 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1970 BasicBlock *Pred = *PI;
1972 }
1973 </pre>
1974 </div>
1976 <p>Similarly, to iterate over successors use
1977 succ_iterator/succ_begin/succ_end.</p>
1979 </div>
1982 <!-- ======================================================================= -->
1985 </div>
1989 <p>There are some primitive transformation operations present in the LLVM
1990 infrastructure that are worth knowing about. When performing
1991 transformations, it's fairly common to manipulate the contents of basic
1992 blocks. This section describes some of the common methods for doing so
1993 and gives example code.</p>
1995 </div>
1997 <!--_______________________________________________________________________-->
2001 </div>
2008 constructor for the kind of instruction to instantiate and provide the necessary
2013 <pre>
2014 AllocaInst* ai = new AllocaInst(Type::Int32Ty);
2015 </pre>
2016 </div>
2020 subclass is likely to have varying default parameters which change the semantics
2021 of the instruction, so refer to the <a
2027 <p>It is very useful to name the values of instructions when you're able to, as
2028 this facilitates the debugging of your transformations. If you end up looking
2029 at generated LLVM machine code, you definitely want to have logical names
2030 associated with the results of instructions! By supplying a value for the
2032 associate a logical name with the result of the instruction's execution at
2033 run time. For example, say that I'm writing a transformation that dynamically
2034 allocates space for an integer on the stack, and that integer is going to be
2035 used as some kind of index by some other code. To accomplish this, I place an
2041 <pre>
2043 </pre>
2044 </div>
2047 execution value, which is a pointer to an integer on the run time stack.</p>
2054 <ul>
2055 <li>Insertion into an explicit instruction list
2062 <pre>
2063 BasicBlock *pb = ...;
2064 Instruction *pi = ...;
2065 Instruction *newInst = new Instruction(...);
2068 </pre>
2069 </div>
2073 classes provide constructors which take a pointer to a
2075 looked like: </p>
2078 <pre>
2079 BasicBlock *pb = ...;
2080 Instruction *newInst = new Instruction(...);
2083 </pre>
2084 </div>
2089 <pre>
2090 BasicBlock *pb = ...;
2091 Instruction *newInst = new Instruction(..., pb);
2092 </pre>
2093 </div>
2095 <p>which is much cleaner, especially if you are creating
2096 long instruction streams.</p></li>
2098 <li>Insertion into an implicit instruction list
2101 are implicitly associated with an existing instruction list: the instruction
2102 list of the enclosing basic block. Thus, we could have accomplished the same
2104 </p>
2107 <pre>
2108 Instruction *pi = ...;
2109 Instruction *newInst = new Instruction(...);
2112 </pre>
2113 </div>
2115 <p>In fact, this sequence of steps occurs so frequently that the
2117 constructors which take (as a default parameter) a pointer to an
2121 provided instruction, immediately before that instruction. Using an
2123 parameter, the above code becomes:</p>
2126 <pre>
2127 Instruction* pi = ...;
2128 Instruction* newInst = new Instruction(..., pi);
2129 </pre>
2130 </div>
2132 <p>which is much cleaner, especially if you're creating a lot of
2134 </ul>
2136 </div>
2138 <!--_______________________________________________________________________-->
2141 </div>
2145 <p>Deleting an instruction from an existing sequence of instructions that form a
2147 you must have a pointer to the instruction that you wish to delete. Second, you
2148 need to obtain the pointer to that instruction's basic block. You use the
2149 pointer to the basic block to get its list of instructions and then use the
2150 erase function to remove your instruction. For example:</p>
2153 <pre>
2155 I->eraseFromParent();
2156 </pre>
2157 </div>
2159 </div>
2161 <!--_______________________________________________________________________-->
2165 </div>
2171 <p>Including "<a href="/doxygen/BasicBlockUtils_8h-source.html">llvm/Transforms/Utils/BasicBlockUtils.h</a>"
2177 <ul>
2180 <p>This function replaces all uses of a given instruction with a value,
2181 and then removes the original instruction. The following example
2182 illustrates the replacement of the result of a particular
2184 pointer to an integer.</p>
2187 <pre>
2188 AllocaInst* instToReplace = ...;
2189 BasicBlock::iterator ii(instToReplace);
2192 Constant::getNullValue(PointerType::getUnqual(Type::Int32Ty)));
2193 </pre></div></li>
2197 <p>This function replaces a particular instruction with another
2198 instruction, inserting the new instruction into the basic block at the
2199 location where the old instruction was, and replacing any uses of the old
2200 instruction with the new instruction. The following example illustrates
2204 <pre>
2205 AllocaInst* instToReplace = ...;
2206 BasicBlock::iterator ii(instToReplace);
2210 </pre></div></li>
2211 </ul>
2219 information.</p>
2221 <!-- Value::replaceAllUsesWith User::replaceUsesOfWith Point out:
2222 include/llvm/Transforms/Utils/ especially BasicBlockUtils.h with:
2223 ReplaceInstWithValue, ReplaceInstWithInst -->
2225 </div>
2227 <!--_______________________________________________________________________-->
2230 </div>
2234 <p>Deleting a global variable from a module is just as easy as deleting an
2235 Instruction. First, you must have a pointer to the global variable that you wish
2236 to delete. You use this pointer to erase it from its parent, the module.
2237 For example:</p>
2240 <pre>
2243 GV->eraseFromParent();
2244 </pre>
2245 </div>
2247 </div>
2249 <!-- ======================================================================= -->
2252 </div>
2256 <p>In generating IR, you may need some complex types. If you know these types
2259 has two forms depending on whether you're building types for cross-compilation
2262 out of types from
2264 namespace and pointers, functions, arrays, etc. built of
2266 whose size may depend on the host compiler. For example,</p>
2269 <pre>
2271 </pre>
2272 </div>
2277 <pre>
2279 params.push_back(PointerType::getUnqual(Type::Int32Ty));
2280 FunctionType *ft = FunctionType::get(Type::Int8Ty, params, false);
2281 </pre>
2282 </div>
2287 </div>
2289 <!-- *********************************************************************** -->
2292 </div>
2293 <!-- *********************************************************************** -->
2296 <p>
2297 This section describes the interaction of the LLVM APIs with multithreading,
2298 both on the part of client applications, and in the JIT, in the hosted
2299 application.
2300 </p>
2302 <p>
2303 Note that LLVM's support for multithreading is still relatively young. Up
2304 through version 2.5, the execution of threaded hosted applications was
2305 supported, but not threaded client access to the APIs. While this use case is
2307 ensure proper operation in multithreaded mode.
2308 </p>
2310 <p>
2311 Note that, on Unix-like platforms, LLVM requires the presence of GCC's atomic
2312 intrinsics in order to support threaded operation. If you need a
2313 multhreading-capable LLVM on a platform without a suitably modern system
2314 compiler, consider compiling LLVM and LLVM-GCC in single-threaded mode, and
2315 using the resultant compiler to build a copy of LLVM with multithreading
2316 support.
2317 </p>
2318 </div>
2320 <!-- ======================================================================= -->
2323 </div>
2327 <p>
2328 In order to properly protect its internal data structures while avoiding
2329 excessive locking overhead in the single-threaded case, the LLVM must intialize
2330 certain data structures necessary to provide guards around its internals. To do
2332 making any concurrent LLVM API calls. To subsequently tear down these
2335 mode.
2336 </p>
2338 <p>
2340 say that no other LLVM API calls may be executing at any time during the
2342 </tt>. It's is the client's responsibility to enforce this isolation.
2343 </p>
2345 <p>
2347 failure of the initialization. Failure typically indicates that your copy of
2348 LLVM was built without multithreading support, typically because GCC atomic
2349 intrinsics were not found in your system compiler. In this case, the LLVM API
2351 hosting threaded applications in the JIT, though care must be taken to ensure
2352 that side exits and the like do not accidentally result in concurrent LLVM API
2353 calls.
2354 </p>
2355 </div>
2357 <!-- ======================================================================= -->
2360 </div>
2363 <p>
2365 to deallocate memory used for internal structures. This will also invoke
2369 </p>
2371 <p>
2372 Note that, if you use scope-based shutdown, you can use the
2374 destructor.
2375 </div>
2377 <!-- ======================================================================= -->
2380 </div>
2383 <p>
2385 initialization of static resources, such as the global type tables. Before the
2388 however, it uses double-checked locking to implement thread-safe lazy
2389 initialization.
2390 </p>
2392 <p>
2393 Note that, because no other threads are allowed to issue LLVM API calls before
2396 </p>
2398 <p>
2400 APIs provide access to the global lock used to implement the double-checked
2401 locking for lazy initialization. These should only be used internally to LLVM,
2402 and only if you know what you're doing!
2403 </p>
2404 </div>
2406 <!-- ======================================================================= -->
2409 </div>
2412 <p>
2414 to operate multiple, isolated instances of LLVM concurrently within the same
2415 address space. For instance, in a hypothetical compile-server, the compilation
2416 of an individual translation unit is conceptually independent from all the
2417 others, and it would be desirable to be able to compile incoming translation
2418 units concurrently on independent server threads. Fortunately,
2420 </p>
2422 <p>
2429 safe to compile on multiple threads simultaneously, as long as no two threads
2430 operate on entities within the same context.
2431 </p>
2433 <p>
2434 In practice, very few places in the API require the explicit specification of a
2437 other entities can determine what context they belong to by looking at their
2439 maintain this interface design.
2440 </p>
2442 <p>
2446 isolation is not a concern.
2447 </p>
2448 </div>
2450 <!-- *********************************************************************** -->
2453 </div>
2454 <!-- *********************************************************************** -->
2457 <p>
2458 This section describes some of the advanced or obscure API's that most clients
2459 do not need to be aware of. These API's tend manage the inner workings of the
2460 LLVM system, and only need to be accessed in unusual circumstances.
2461 </p>
2462 </div>
2464 <!-- ======================================================================= -->
2467 </div>
2471 <p>
2472 The LLVM type system has a very simple goal: allow clients to compare types for
2473 structural equality with a simple pointer comparison (aka a shallow compare).
2474 This goal makes clients much simpler and faster, and is used throughout the LLVM
2475 system.
2476 </p>
2478 <p>
2479 Unfortunately achieving this goal is not a simple matter. In particular,
2480 recursive types and late resolution of opaque types makes the situation very
2481 difficult to handle. Fortunately, for the most part, our implementation makes
2482 most clients able to be completely unaware of the nasty internal details. The
2483 primary case where clients are exposed to the inner workings of it are when
2484 building a recursive type. In addition to this case, the LLVM bitcode reader,
2485 assembly parser, and linker also have to be aware of the inner workings of this
2486 system.
2487 </p>
2489 <p>
2490 For our purposes below, we need three concepts. First, an "Opaque Type" is
2493 opaque type as part of its type graph (for example "<tt>{ opaque, i32 }</tt>").
2494 Third, a concrete type is a type that is not an abstract type (e.g. "<tt>{ i32,
2495 float }</tt>").
2496 </p>
2498 </div>
2500 <!-- ______________________________________________________________________ -->
2503 </div>
2507 <p>
2508 Because the most common question is "how do I build a recursive type with LLVM",
2509 we answer it now and explain it as we go. Here we include enough to cause this
2510 to be emitted to an output .ll file:
2511 </p>
2514 <pre>
2515 %mylist = type { %mylist*, i32 }
2516 </pre>
2517 </div>
2519 <p>
2520 To build this, use the following LLVM APIs:
2521 </p>
2524 <pre>
2528 Elts.push_back(PointerType::getUnqual(StructTy));
2529 Elts.push_back(Type::Int32Ty);
2530 StructType *NewSTy = StructType::get(Elts);
2534 cast<OpaqueType>(StructTy.get())-><a href="#refineAbstractTypeTo">refineAbstractTypeTo</a>(NewSTy);
2536 // <i>NewSTy is potentially invalidated, but StructTy (a <a href="#PATypeHolder">PATypeHolder</a>) is</i>
2542 </pre>
2543 </div>
2545 <p>
2546 This code shows the basic approach used to build recursive types: build a
2547 non-recursive type using 'opaque', then use type unification to close the cycle.
2548 The type unification step is performed by the <tt><a
2550 described next. After that, we describe the <a
2552 </p>
2554 </div>
2556 <!-- ______________________________________________________________________ -->
2559 </div>
2562 <p>
2564 While this method is actually a member of the DerivedType class, it is most
2565 often used on OpaqueType instances. Type unification is actually a recursive
2566 process. After unification, types can become structurally isomorphic to
2567 existing types, and all duplicates are deleted (to preserve pointer equality).
2568 </p>
2570 <p>
2571 In the example above, the OpaqueType object is definitely deleted.
2572 Additionally, if there is an "{ \2*, i32}" type already created in the system,
2574 a type is deleted, any "Type*" pointers in the program are invalidated. As
2577 types can never move or be deleted). To deal with this, the <a
2579 reference to a possibly refined type, and the <a
2581 complex datastructures.
2582 </p>
2584 </div>
2586 <!-- ______________________________________________________________________ -->
2589 </div>
2592 <p>
2593 PATypeHolder is a form of a "smart pointer" for Type objects. When VMCore
2594 happily goes about nuking types that become isomorphic to existing types, it
2595 automatically updates all PATypeHolder objects to point to the new type. In the
2596 example above, this allows the code to maintain a pointer to the resultant
2597 resolved recursive type, even though the Type*'s are potentially invalidated.
2598 </p>
2600 <p>
2601 PATypeHolder is an extremely light-weight object that uses a lazy union-find
2602 implementation to update pointers. For example the pointer from a Value to its
2603 Type is maintained by PATypeHolder objects.
2604 </p>
2606 </div>
2608 <!-- ______________________________________________________________________ -->
2611 </div>
2615 <p>
2616 Some data structures need more to perform more complex updates when types get
2617 resolved. To support this, a class can derive from the AbstractTypeUser class.
2618 This class
2619 allows it to get callbacks when certain types are resolved. To register to get
2620 callbacks for a particular type, the DerivedType::{add/remove}AbstractTypeUser
2621 methods can be called on a type. Note that these methods only work for <i>
2622 abstract</i> types. Concrete types (those that do not include any opaque
2623 objects) can never be refined.
2624 </p>
2625 </div>
2628 <!-- ======================================================================= -->
2632 </div>
2642 names for types.</p>
2645 by most clients. It should only be used when iteration over the symbol table
2646 names themselves are required, which is very special purpose. Note that not
2647 all LLVM
2649 an empty name) do not exist in the symbol table.
2650 </p>
2652 <p>These symbol tables support iteration over the values/types in the symbol
2657 appropriate symbol table. For types, use the Module::addTypeName method to
2658 insert entries into the symbol table.</p>
2660 </div>
2664 <!-- ======================================================================= -->
2667 </div>
2675 addition and removal.</p>
2677 <!-- ______________________________________________________________________ -->
2679 <a name="Use2User">Interaction and relationship between <tt>User</tt> and <tt>Use</tt> objects</a>
2680 </div>
2683 <p>
2685 or refer to them out-of-line by means of a pointer. A mixed variant
2688 </p>
2689 </div>
2691 <p>
2693 <ul>
2694 <li><p>Layout a)
2696 object and there are a fixed number of them.</p>
2698 <li><p>Layout b)
2701 number of them.</p>
2702 </ul>
2703 <p>
2704 As of v2.4 each layout still possesses a direct pointer to the
2706 we stick to this redundancy for the sake of simplicity.
2708 has. (Theoretically this information can also be calculated
2709 given the scheme presented below.)</p>
2710 <p>
2712 enforce the following memory layouts:</p>
2714 <ul>
2715 <li><p>Layout a) is modelled by prepending the <tt>User</tt> object by the <tt>Use[]</tt> array.</p>
2717 <pre>
2718 ...---.---.---.---.-------...
2719 | P | P | P | P | User
2720 '''---'---'---'---'-------'''
2721 </pre>
2724 <pre>
2725 .-------...
2726 | User
2727 '-------'''
2728 |
2729 v
2730 .---.---.---.---...
2731 | P | P | P | P |
2732 '---'---'---'---'''
2733 </pre>
2734 </ul>
2738 <!-- ______________________________________________________________________ -->
2741 </div>
2744 <p>
2747 recover it. This is accomplished by the following scheme:</p>
2748 </div>
2750 A bit-encoding in the 2 LSBits (least significant bits) of the <tt>Use::Prev</tt> allows to find the
2752 <ul>
2757 </ul>
2758 <p>
2761 we have to walk to the next stop picking up digits
2762 and calculating the offset:</p>
2763 <pre>
2764 .---.---.---.---.---.---.---.---.---.---.---.---.---.---.---.---.----------------
2766 '---'---'---'---'---'---'---'---'---'---'---'---'---'---'---'---'----------------
2768 | | | | |__>
2769 | | | |__________>
2770 | | |______________________>
2771 | |______________________________________>
2772 |__________________________________________________________>
2773 </pre>
2774 <p>
2775 Only the significant number of bits need to be stored between the
2779 <!-- ______________________________________________________________________ -->
2782 </div>
2785 <p>
2786 The following literate Haskell fragment demonstrates the concept:</p>
2787 </div>
2790 <pre>
2791 > import Test.QuickCheck
2792 >
2797 >
2803 >
2804 > takeLast n ss = reverse $ take n $ reverse ss
2805 >
2807 >
2808 </pre>
2809 </div>
2810 <p>
2812 <p>
2813 The reverse algorithm computes the length of the string just by examining
2814 a certain prefix:</p>
2817 <pre>
2822 >
2825 > decode walk acc _ = walk + acc
2826 >
2827 </pre>
2828 </div>
2829 <p>
2831 <p>
2835 <pre>
2837 > testcaseLength = length testcase
2838 >
2840 > where arr = takeLast n testcase
2841 >
2842 </pre>
2843 </div>
2844 <p>
2847 <pre>
2848 *Main> quickCheck identityProp
2849 OK, passed 100 tests.
2850 </pre>
2851 <p>
2852 Let's be a bit more exhaustive:</p>
2855 <pre>
2856 >
2858 >
2859 </pre>
2860 </div>
2861 <p>
2864 <pre>
2865 *Main> deepCheck identityProp
2866 OK, passed 500 tests.
2867 </pre>
2869 <!-- ______________________________________________________________________ -->
2872 </div>
2874 <p>
2878 tag bits.</p>
2879 <p>
2883 the LSBit set. (Portability is relying on the fact that all known compilers place the
2886 </div>
2888 <!-- *********************************************************************** -->
2891 </div>
2892 <!-- *********************************************************************** -->
2898 <p>The Core LLVM classes are the primary means of representing the program
2899 being inspected or transformed. The core LLVM classes are defined in
2903 </div>
2905 <!-- ======================================================================= -->
2908 </div>
2916 subclasses. They are hidden because they offer no useful functionality beyond
2920 named, but this is not a requirement. There exists exactly
2921 one instance of a given shape at any one time. This allows type equality to
2922 be performed with address equality of the Type Instance. That is, given two
2924 </p>
2925 </div>
2927 <!-- _______________________________________________________________________ -->
2930 </div>
2934 <ul>
2938 floating point types.</li>
2941 an OpaqueType anywhere in its definition).</li>
2944 that don't have a size are abstract types, labels and void.</li>
2946 </ul>
2947 </div>
2949 <!-- _______________________________________________________________________ -->
2952 </div>
2954 <dl>
2956 <dd>Subclass of DerivedType that represents integer types of any bit width.
2959 <ul>
2961 type of a specific bit width.</li>
2963 type.</li>
2964 </ul>
2965 </dd>
2967 <dd>This is subclassed by ArrayType and PointerType
2968 <ul>
2970 of the elements in the sequential type. </li>
2971 </ul>
2972 </dd>
2974 <dd>This is a subclass of SequentialType and defines the interface for array
2975 types.
2976 <ul>
2978 elements in the array. </li>
2979 </ul>
2980 </dd>
2984 <dd>Subclass of SequentialType for vector types. A
2985 vector type is similar to an ArrayType but is distinguished because it is
2986 a first class type wherease ArrayType is not. Vector types are used for
2987 vector operations and are usually small vectors of of an integer or floating
2988 point type.</dd>
2992 <dd>Subclass of DerivedTypes for function types.
2993 <ul>
2995 function</li>
2997 return type of the function.</li>
2999 the type of the ith parameter.</li>
3001 number of formal parameters.</li>
3002 </ul>
3003 </dd>
3005 <dd>Sublcass of DerivedType for abstract types. This class
3006 defines no content and is used as a placeholder for some other type. Note
3007 that OpaqueType is used (temporarily) during type resolution for forward
3008 references of types. Once the referenced type is resolved, the OpaqueType
3009 is replaced with the actual type. OpaqueType can also be used for data
3010 abstraction. At link time opaque types can be resolved to actual types
3011 of the same name.</dd>
3012 </dl>
3013 </div>
3017 <!-- ======================================================================= -->
3020 </div>
3029 programs. An LLVM module is effectively either a translation unit of the
3030 original program or a combination of several translation units merged by the
3035 helpful member functions that try to make common operations easy.</p>
3037 </div>
3039 <!-- _______________________________________________________________________ -->
3042 </div>
3046 <ul>
3048 </ul>
3051 provide a name for it (probably based on the name of the translation unit).</p>
3053 <ul>
3060 <p>These are forwarding methods that make it easy to access the contents of
3062 list.</p></li>
3067 necessary to use when you need to update the list or perform a complex
3068 action that doesn't have a forwarding method.</p>
3071 </ul>
3073 <hr>
3075 <ul>
3083 <p> These are forwarding methods that make it easy to access the contents of
3091 use when you need to update the list or perform a complex action that
3092 doesn't have a forwarding method.</p>
3095 </ul>
3097 <hr>
3099 <ul>
3106 </ul>
3108 <hr>
3110 <ul>
3123 external declaration for the function and return it.</p></li>
3130 string.</p></li>
3137 name, true is returned and the <a
3139 </ul>
3141 </div>
3144 <!-- ======================================================================= -->
3147 </div>
3152 <br>
3156 base. It represents a typed value that may be used (among other things) as an
3164 for a program. For example, an incoming argument to a function (represented
3166 every instruction in the function that references the argument. To keep track
3172 methods, shown below.</p>
3176 method. In addition, all LLVM values can be named. The "name" of the
3180 <pre>
3182 </pre>
3183 </div>
3186 that the name of any value may be missing (an empty string), so names should
3188 debugging printouts), they should not be used to keep track of values or map
3192 <p>One important aspect of LLVM is that there is no distinction between an SSA
3193 variable and the operation that produces it. Because of this, any reference to
3194 the value produced by an instruction (or the value available as an incoming
3195 argument, for example) is represented as a direct pointer to the instance of
3196 the class that
3197 represents this value. Although this may take some getting used to, it
3198 simplifies the representation and makes it easier to manipulate.</p>
3200 </div>
3202 <!-- _______________________________________________________________________ -->
3205 </div>
3209 <ul>
3211 use-list<br>
3213 the use-list<br>
3215 value.<br>
3218 the use-list.<br>
3220 use-list.<br>
3222 element in the list.
3223 <p> These methods are the interface to access the def-use
3224 information in LLVM. As with all other iterators in LLVM, the naming
3226 </li>
3229 </li>
3235 </li>
3240 "<tt>V</tt>" instead. For example, if you detect that an instruction always
3241 produces a constant value (for example through constant folding), you can
3242 replace all uses of the instruction with the constant like this:</p>
3245 <pre>
3246 Inst->replaceAllUsesWith(ConstVal);
3247 </pre>
3248 </div>
3250 </ul>
3252 </div>
3254 <!-- ======================================================================= -->
3257 </div>
3261 <p>
3274 Single Assignment (SSA) form, there can only be one definition referred to,
3275 allowing this direct connection. This connection provides the use-def
3276 information in LLVM.</p>
3278 </div>
3280 <!-- _______________________________________________________________________ -->
3283 </div>
3288 an index access interface and through an iterator based interface.</p>
3290 <ul>
3294 convenient form for direct access.</p></li>
3297 list<br>
3299 the operand list.<br>
3301 operand list.
3302 <p> Together, these methods make up the iterator based interface to
3304 </ul>
3306 </div>
3308 <!-- ======================================================================= -->
3311 </div>
3322 instructions. It provides only a few methods, but is a very commonly used
3324 opcode (instruction type) and the parent <a
3326 into. To represent a specific type of instruction, one of many subclasses of
3335 file contains some meta-data about the various different types of instructions
3336 in LLVM. It describes the enum values that are used as opcodes (for example
3341 this file confuses doxygen, so these enum values don't show up correctly in the
3344 </div>
3346 <!-- _______________________________________________________________________ -->
3349 class</a>
3350 </div>
3352 <ul>
3354 <p>This subclasses represents all two operand instructions whose operands
3355 must be the same type, except for the comparison instructions.</p></li>
3358 common operations on cast instructions.</p>
3360 <p>This subclass respresents the two comparison instructions,
3364 <p>This subclass is the parent of all terminator instructions (those which
3365 can terminate a block).</p>
3366 </ul>
3367 </div>
3369 <!-- _______________________________________________________________________ -->
3372 class</a>
3373 </div>
3377 <ul>
3382 <p>Returns true if the instruction writes to memory, i.e. it is a
3387 <p>Returns another instance of the specified instruction, identical
3388 in all ways to the original except that the instruction has no parent
3390 and it has no name</p></li>
3391 </ul>
3393 </div>
3395 <!-- ======================================================================= -->
3398 </div>
3402 <p>Constant represents a base class for different types of constants. It
3403 is subclassed by ConstantInt, ConstantArray, etc. for representing
3405 a subclass, which represents the address of a global variable or function.
3406 </p>
3408 </div>
3410 <!-- _______________________________________________________________________ -->
3413 <ul>
3414 <li>ConstantInt : This subclass of Constant represents an integer constant of
3415 any width.
3416 <ul>
3418 value of this constant, an APInt value.</li>
3420 value to an int64_t via sign extension. If the value (not the bit width)
3421 of the APInt is too large to fit in an int64_t, an assertion will result.
3422 For this reason, use of this method is discouraged.</li>
3424 value to a uint64_t via zero extension. IF the value (not the bit width)
3425 of the APInt is too large to fit in a uint64_t, an assertion will result.
3426 For this reason, use of this method is discouraged.</li>
3429 The type is implied as the IntegerType that corresponds to the bit width
3432 Returns the ConstantInt object that represents the value provided by
3434 </ul>
3435 </li>
3436 <li>ConstantFP : This class represents a floating point constant.
3437 <ul>
3439 this constant. </li>
3440 </ul>
3441 </li>
3442 <li>ConstantArray : This represents a constant array.
3443 <ul>
3445 a vector of component constants that makeup this array. </li>
3446 </ul>
3447 </li>
3448 <li>ConstantStruct : This represents a constant struct.
3449 <ul>
3451 a vector of component constants that makeup this array. </li>
3452 </ul>
3453 </li>
3454 <li>GlobalValue : This represents either a global variable or a function. In
3455 either case, the value is a constant fixed address (after linking).
3456 </li>
3457 </ul>
3458 </div>
3461 <!-- ======================================================================= -->
3464 </div>
3471 Class</a><br>
3478 Because they are visible at global scope, they are also subject to linking with
3479 other globals defined in different translation units. To control the linking
3486 unit, and does not participate in linking. If it has external linkage, it is
3487 visible to external code, and does participate in linking. In addition to
3493 global is always a pointer to its contents. It is important to remember this
3498 the address of the first element of this array and the value of the
3506 </div>
3508 <!-- _______________________________________________________________________ -->
3511 class</a>
3512 </div>
3516 <ul>
3521 <p> </p>
3522 </li>
3525 GlobalValue is currently embedded into.</p></li>
3526 </ul>
3528 </div>
3530 <!-- ======================================================================= -->
3533 </div>
3546 actually one of the more complex classes in the LLVM heirarchy because it must
3554 ordering of the blocks in the function, which indicate how the code will be
3555 layed out by the backend. Additionally, the first <a
3558 block. There are no implicit exit nodes, and in fact there may be multiple exit
3562 function hasn't been linked in yet.</p>
3572 LLVM feature that is only used when you have to look up a value by name. Aside
3574 internally to make sure that there are not conflicts between the names of <a
3581 is its address (after linking) which is guaranteed to be constant.</p>
3582 </div>
3584 <!-- _______________________________________________________________________ -->
3587 class</a>
3588 </div>
3592 <ul>
3597 the the program. The constructor must specify the type of the function to
3598 create and what type of linkage the function should have. The <a
3600 specifies the formal arguments and return value for the function. The same
3603 in which the function is defined. If this argument is provided, the function
3604 will automatically be inserted into that module's list of
3605 functions.</p></li>
3610 function is "external", it does not have a body, and thus must be resolved
3611 by linking with a function defined in a different translation unit.</p></li>
3619 <p>These are forwarding methods that make it easy to access the contents of
3621 list.</p></li>
3626 is necessary to use when you need to update the list or perform a complex
3627 action that doesn't have a forwarding method.</p></li>
3630 iterator<br>
3636 <p>These are forwarding methods that make it easy to access the contents of
3638 list.</p></li>
3643 necessary to use when you need to update the list or perform a complex
3644 action that doesn't have a forwarding method.</p></li>
3649 function. Because the entry block for the function is always the first
3658 function.</p></li>
3664 </ul>
3666 </div>
3668 <!-- ======================================================================= -->
3671 </div>
3677 <br>
3679 Class</a><br>
3685 <p>Global variables are represented with the (suprise suprise)
3688 always referenced by their address (global values must live in memory, so their
3689 "name" refers to their constant address). See
3691 variables may have an initial value (which must be a
3693 they may be marked as "constant" themselves (indicating that their contents
3694 never change at runtime).</p>
3695 </div>
3697 <!-- _______________________________________________________________________ -->
3701 </div>
3705 <ul>
3710 <p>Create a new global variable of the specified type. If
3712 unchanging for the program. The Linkage parameter specifies the type of
3713 linkage (internal, external, weak, linkonce, appending) for the variable.
3714 If the linkage is InternalLinkage, WeakAnyLinkage, WeakODRLinkage,
3715 LinkOnceAnyLinkage or LinkOnceODRLinkage, then the resultant
3716 global variable will have internal linkage. AppendingLinkage concatenates
3717 together all instances (in different translation units) of the variable
3718 into a single variable but is only applicable to arrays. See
3720 further details on linkage types. Optionally an initializer, a name, and the
3721 module to put the variable into may be specified for the global variable as
3722 well.</p></li>
3726 <p>Returns true if this is a global variable that is known not to
3727 be modified at runtime.</p></li>
3736 to call this method if there is no initializer.</p></li>
3737 </ul>
3739 </div>
3742 <!-- ======================================================================= -->
3745 </div>
3752 Class</a><br>
3755 <p>This class represents a single entry multiple exit section of the code,
3756 commonly known as a basic block by the compiler community. The
3759 Matching the language definition, the last element of this list of instructions
3760 is always a terminator instruction (a subclass of the <a
3763 <p>In addition to tracking the list of instructions that make up the block, the
3772 </div>
3774 <!-- _______________________________________________________________________ -->
3777 class</a>
3778 </div>
3781 <ul>
3787 insertion into a function. The constructor optionally takes a name for the new
3790 automatically inserted at the end of the specified <a
3798 STL-style functions for accessing the instruction list.
3800 <p>These methods and typedefs are forwarding functions that have the same
3801 semantics as the standard library methods of the same names. These methods
3802 expose the underlying instruction list of a basic block in a way that is easy to
3803 manipulate. To get the full complement of container operations (including
3805 method.</p></li>
3809 <p>This method is used to get access to the underlying container that actually
3810 holds the Instructions. This method must be used when there isn't a forwarding
3812 to perform. Because there are no forwarding functions for "updating"
3813 operations, you need to use this if you want to update the contents of a
3819 embedded into, or a null pointer if it is homeless.</p></li>
3823 <p> Returns a pointer to the terminator instruction that appears at the end of
3825 instruction in the block is not a terminator, then a null pointer is
3826 returned.</p></li>
3828 </ul>
3830 </div>
3833 <!-- ======================================================================= -->
3836 </div>
3840 <p>This subclass of Value defines the interface for incoming formal
3841 arguments to a function. A Function maintains a list of its formal
3842 arguments. An argument has a pointer to the parent Function.</p>
3844 </div>
3846 <!-- *********************************************************************** -->
3847 <hr>
3848 <address>
3857 Last modified: $Date$
3858 </address>
3860 </body>
3861 </html>