1 ========================
2 LLVM Programmer's Manual
3 ========================
9 This is a work in progress.
11 .. sectionauthor:: Chris Lattner <sabre@nondot.org>,
12 Dinakar Dhurjati <dhurjati@cs.uiuc.edu>,
13 Gabor Greif <ggreif@gmail.com>,
14 Joel Stanley <jstanley@cs.uiuc.edu>,
15 Reid Spencer <rspencer@x10sys.com> and
16 Owen Anderson <owen@apple.com>
23 This document is meant to highlight some of the important classes and interfaces
24 available in the LLVM source-base. This manual is not intended to explain what
25 LLVM is, how it works, and what LLVM code looks like. It assumes that you know
26 the basics of LLVM and are interested in writing transformations or otherwise
27 analyzing or manipulating the code.
29 This document should get you oriented so that you can find your way in the
30 continuously growing source code that makes up the LLVM infrastructure. Note
31 that this manual is not intended to serve as a replacement for reading the
32 source code, so if you think there should be a method in one of these classes to
33 do something, but it's not listed, check the source. Links to the `doxygen
34 <http://llvm.org/doxygen/>`__ sources are provided to make this as easy as
37 The first section of this document describes general information that is useful
38 to know when working in the LLVM infrastructure, and the second describes the
39 Core LLVM classes. In the future this manual will be extended with information
40 describing how to use extension libraries, such as dominator information, CFG
41 traversal routines, and useful utilities like the ``InstVisitor`` (`doxygen
42 <http://llvm.org/doxygen/InstVisitor_8h-source.html>`__) template.
49 This section contains general information that is useful if you are working in
50 the LLVM source-base, but that isn't specific to any particular API.
54 The C++ Standard Template Library
55 ---------------------------------
57 LLVM makes heavy use of the C++ Standard Template Library (STL), perhaps much
58 more than you are used to, or have seen before. Because of this, you might want
59 to do a little background reading in the techniques used and capabilities of the
60 library. There are many good pages that discuss the STL, and several books on
61 the subject that you can get, so it will not be discussed in this document.
63 Here are some useful links:
66 <http://en.cppreference.com/w/>`_ - an excellent
67 reference for the STL and other parts of the standard C++ library.
69 #. `C++ In a Nutshell <http://www.tempest-sw.com/cpp/>`_ - This is an O'Reilly
70 book in the making. It has a decent Standard Library Reference that rivals
71 Dinkumware's, and is unfortunately no longer free since the book has been
74 #. `C++ Frequently Asked Questions <http://www.parashift.com/c++-faq-lite/>`_.
76 #. `SGI's STL Programmer's Guide <http://www.sgi.com/tech/stl/>`_ - Contains a
77 useful `Introduction to the STL
78 <http://www.sgi.com/tech/stl/stl_introduction.html>`_.
80 #. `Bjarne Stroustrup's C++ Page
81 <http://www.research.att.com/%7Ebs/C++.html>`_.
83 #. `Bruce Eckel's Thinking in C++, 2nd ed. Volume 2 Revision 4.0
84 (even better, get the book) <http://64.78.49.204/>`_.
86 You are also encouraged to take a look at the :ref:`LLVM Coding Standards
87 <coding_standards>` guide which focuses on how to write maintainable code more
88 than where to put your curly braces.
92 Other useful references
93 -----------------------
95 #. `Using static and shared libraries across platforms
96 <http://www.fortran-2000.com/ArnaudRecipes/sharedlib.html>`_
100 Important and useful LLVM APIs
101 ==============================
103 Here we highlight some LLVM APIs that are generally useful and good to know
104 about when writing transformations.
108 The ``isa<>``, ``cast<>`` and ``dyn_cast<>`` templates
109 ------------------------------------------------------
111 The LLVM source-base makes extensive use of a custom form of RTTI. These
112 templates have many similarities to the C++ ``dynamic_cast<>`` operator, but
113 they don't have some drawbacks (primarily stemming from the fact that
114 ``dynamic_cast<>`` only works on classes that have a v-table). Because they are
115 used so often, you must know what they do and how they work. All of these
116 templates are defined in the ``llvm/Support/Casting.h`` (`doxygen
117 <http://llvm.org/doxygen/Casting_8h-source.html>`__) file (note that you very
118 rarely have to include this file directly).
121 The ``isa<>`` operator works exactly like the Java "``instanceof``" operator.
122 It returns true or false depending on whether a reference or pointer points to
123 an instance of the specified class. This can be very useful for constraint
124 checking of various sorts (example below).
127 The ``cast<>`` operator is a "checked cast" operation. It converts a pointer
128 or reference from a base class to a derived class, causing an assertion
129 failure if it is not really an instance of the right type. This should be
130 used in cases where you have some information that makes you believe that
131 something is of the right type. An example of the ``isa<>`` and ``cast<>``
136 static bool isLoopInvariant(const Value *V, const Loop *L) {
137 if (isa<Constant>(V) || isa<Argument>(V) || isa<GlobalValue>(V))
140 // Otherwise, it must be an instruction...
141 return !L->contains(cast<Instruction>(V)->getParent());
144 Note that you should **not** use an ``isa<>`` test followed by a ``cast<>``,
145 for that use the ``dyn_cast<>`` operator.
148 The ``dyn_cast<>`` operator is a "checking cast" operation. It checks to see
149 if the operand is of the specified type, and if so, returns a pointer to it
150 (this operator does not work with references). If the operand is not of the
151 correct type, a null pointer is returned. Thus, this works very much like
152 the ``dynamic_cast<>`` operator in C++, and should be used in the same
153 circumstances. Typically, the ``dyn_cast<>`` operator is used in an ``if``
154 statement or some other flow control statement like this:
158 if (AllocationInst *AI = dyn_cast<AllocationInst>(Val)) {
162 This form of the ``if`` statement effectively combines together a call to
163 ``isa<>`` and a call to ``cast<>`` into one statement, which is very
166 Note that the ``dyn_cast<>`` operator, like C++'s ``dynamic_cast<>`` or Java's
167 ``instanceof`` operator, can be abused. In particular, you should not use big
168 chained ``if/then/else`` blocks to check for lots of different variants of
169 classes. If you find yourself wanting to do this, it is much cleaner and more
170 efficient to use the ``InstVisitor`` class to dispatch over the instruction
174 The ``cast_or_null<>`` operator works just like the ``cast<>`` operator,
175 except that it allows for a null pointer as an argument (which it then
176 propagates). This can sometimes be useful, allowing you to combine several
177 null checks into one.
179 ``dyn_cast_or_null<>``:
180 The ``dyn_cast_or_null<>`` operator works just like the ``dyn_cast<>``
181 operator, except that it allows for a null pointer as an argument (which it
182 then propagates). This can sometimes be useful, allowing you to combine
183 several null checks into one.
185 These five templates can be used with any classes, whether they have a v-table
186 or not. If you want to add support for these templates, see the document
187 :ref:`How to set up LLVM-style RTTI for your class hierarchy
188 <how-to-set-up-llvm-style-rtti>`
192 Passing strings (the ``StringRef`` and ``Twine`` classes)
193 ---------------------------------------------------------
195 Although LLVM generally does not do much string manipulation, we do have several
196 important APIs which take strings. Two important examples are the Value class
197 -- which has names for instructions, functions, etc. -- and the ``StringMap``
198 class which is used extensively in LLVM and Clang.
200 These are generic classes, and they need to be able to accept strings which may
201 have embedded null characters. Therefore, they cannot simply take a ``const
202 char *``, and taking a ``const std::string&`` requires clients to perform a heap
203 allocation which is usually unnecessary. Instead, many LLVM APIs use a
204 ``StringRef`` or a ``const Twine&`` for passing strings efficiently.
208 The ``StringRef`` class
209 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
211 The ``StringRef`` data type represents a reference to a constant string (a
212 character array and a length) and supports the common operations available on
213 ``std::string``, but does not require heap allocation.
215 It can be implicitly constructed using a C style null-terminated string, an
216 ``std::string``, or explicitly with a character pointer and length. For
217 example, the ``StringRef`` find function is declared as:
221 iterator find(StringRef Key);
223 and clients can call it using any one of:
227 Map.find("foo"); // Lookup "foo"
228 Map.find(std::string("bar")); // Lookup "bar"
229 Map.find(StringRef("\0baz", 4)); // Lookup "\0baz"
231 Similarly, APIs which need to return a string may return a ``StringRef``
232 instance, which can be used directly or converted to an ``std::string`` using
233 the ``str`` member function. See ``llvm/ADT/StringRef.h`` (`doxygen
234 <http://llvm.org/doxygen/classllvm_1_1StringRef_8h-source.html>`__) for more
237 You should rarely use the ``StringRef`` class directly, because it contains
238 pointers to external memory it is not generally safe to store an instance of the
239 class (unless you know that the external storage will not be freed).
240 ``StringRef`` is small and pervasive enough in LLVM that it should always be
246 The ``Twine`` (`doxygen <http://llvm.org/doxygen/classllvm_1_1Twine.html>`__)
247 class is an efficient way for APIs to accept concatenated strings. For example,
248 a common LLVM paradigm is to name one instruction based on the name of another
249 instruction with a suffix, for example:
253 New = CmpInst::Create(..., SO->getName() + ".cmp");
255 The ``Twine`` class is effectively a lightweight `rope
256 <http://en.wikipedia.org/wiki/Rope_(computer_science)>`_ which points to
257 temporary (stack allocated) objects. Twines can be implicitly constructed as
258 the result of the plus operator applied to strings (i.e., a C strings, an
259 ``std::string``, or a ``StringRef``). The twine delays the actual concatenation
260 of strings until it is actually required, at which point it can be efficiently
261 rendered directly into a character array. This avoids unnecessary heap
262 allocation involved in constructing the temporary results of string
263 concatenation. See ``llvm/ADT/Twine.h`` (`doxygen
264 <http://llvm.org/doxygen/Twine_8h_source.html>`__) and :ref:`here <dss_twine>`
265 for more information.
267 As with a ``StringRef``, ``Twine`` objects point to external memory and should
268 almost never be stored or mentioned directly. They are intended solely for use
269 when defining a function which should be able to efficiently accept concatenated
274 The ``DEBUG()`` macro and ``-debug`` option
275 -------------------------------------------
277 Often when working on your pass you will put a bunch of debugging printouts and
278 other code into your pass. After you get it working, you want to remove it, but
279 you may need it again in the future (to work out new bugs that you run across).
281 Naturally, because of this, you don't want to delete the debug printouts, but
282 you don't want them to always be noisy. A standard compromise is to comment
283 them out, allowing you to enable them if you need them in the future.
285 The ``llvm/Support/Debug.h`` (`doxygen
286 <http://llvm.org/doxygen/Debug_8h-source.html>`__) file provides a macro named
287 ``DEBUG()`` that is a much nicer solution to this problem. Basically, you can
288 put arbitrary code into the argument of the ``DEBUG`` macro, and it is only
289 executed if '``opt``' (or any other tool) is run with the '``-debug``' command
294 DEBUG(errs() << "I am here!\n");
296 Then you can run your pass like this:
300 $ opt < a.bc > /dev/null -mypass
302 $ opt < a.bc > /dev/null -mypass -debug
305 Using the ``DEBUG()`` macro instead of a home-brewed solution allows you to not
306 have to create "yet another" command line option for the debug output for your
307 pass. Note that ``DEBUG()`` macros are disabled for optimized builds, so they
308 do not cause a performance impact at all (for the same reason, they should also
309 not contain side-effects!).
311 One additional nice thing about the ``DEBUG()`` macro is that you can enable or
312 disable it directly in gdb. Just use "``set DebugFlag=0``" or "``set
313 DebugFlag=1``" from the gdb if the program is running. If the program hasn't
314 been started yet, you can always just run it with ``-debug``.
318 Fine grained debug info with ``DEBUG_TYPE`` and the ``-debug-only`` option
319 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
321 Sometimes you may find yourself in a situation where enabling ``-debug`` just
322 turns on **too much** information (such as when working on the code generator).
323 If you want to enable debug information with more fine-grained control, you
324 define the ``DEBUG_TYPE`` macro and the ``-debug`` only option as follows:
329 DEBUG(errs() << "No debug type\n");
330 #define DEBUG_TYPE "foo"
331 DEBUG(errs() << "'foo' debug type\n");
333 #define DEBUG_TYPE "bar"
334 DEBUG(errs() << "'bar' debug type\n"));
336 #define DEBUG_TYPE ""
337 DEBUG(errs() << "No debug type (2)\n");
339 Then you can run your pass like this:
343 $ opt < a.bc > /dev/null -mypass
345 $ opt < a.bc > /dev/null -mypass -debug
350 $ opt < a.bc > /dev/null -mypass -debug-only=foo
352 $ opt < a.bc > /dev/null -mypass -debug-only=bar
355 Of course, in practice, you should only set ``DEBUG_TYPE`` at the top of a file,
356 to specify the debug type for the entire module (if you do this before you
357 ``#include "llvm/Support/Debug.h"``, you don't have to insert the ugly
358 ``#undef``'s). Also, you should use names more meaningful than "foo" and "bar",
359 because there is no system in place to ensure that names do not conflict. If
360 two different modules use the same string, they will all be turned on when the
361 name is specified. This allows, for example, all debug information for
362 instruction scheduling to be enabled with ``-debug-type=InstrSched``, even if
363 the source lives in multiple files.
365 The ``DEBUG_WITH_TYPE`` macro is also available for situations where you would
366 like to set ``DEBUG_TYPE``, but only for one specific ``DEBUG`` statement. It
367 takes an additional first parameter, which is the type to use. For example, the
368 preceding example could be written as:
372 DEBUG_WITH_TYPE("", errs() << "No debug type\n");
373 DEBUG_WITH_TYPE("foo", errs() << "'foo' debug type\n");
374 DEBUG_WITH_TYPE("bar", errs() << "'bar' debug type\n"));
375 DEBUG_WITH_TYPE("", errs() << "No debug type (2)\n");
379 The ``Statistic`` class & ``-stats`` option
380 -------------------------------------------
382 The ``llvm/ADT/Statistic.h`` (`doxygen
383 <http://llvm.org/doxygen/Statistic_8h-source.html>`__) file provides a class
384 named ``Statistic`` that is used as a unified way to keep track of what the LLVM
385 compiler is doing and how effective various optimizations are. It is useful to
386 see what optimizations are contributing to making a particular program run
389 Often you may run your pass on some big program, and you're interested to see
390 how many times it makes a certain transformation. Although you can do this with
391 hand inspection, or some ad-hoc method, this is a real pain and not very useful
392 for big programs. Using the ``Statistic`` class makes it very easy to keep
393 track of this information, and the calculated information is presented in a
394 uniform manner with the rest of the passes being executed.
396 There are many examples of ``Statistic`` uses, but the basics of using it are as
399 #. Define your statistic like this:
403 #define DEBUG_TYPE "mypassname" // This goes before any #includes.
404 STATISTIC(NumXForms, "The # of times I did stuff");
406 The ``STATISTIC`` macro defines a static variable, whose name is specified by
407 the first argument. The pass name is taken from the ``DEBUG_TYPE`` macro, and
408 the description is taken from the second argument. The variable defined
409 ("NumXForms" in this case) acts like an unsigned integer.
411 #. Whenever you make a transformation, bump the counter:
415 ++NumXForms; // I did stuff!
417 That's all you have to do. To get '``opt``' to print out the statistics
418 gathered, use the '``-stats``' option:
422 $ opt -stats -mypassname < program.bc > /dev/null
423 ... statistics output ...
425 When running ``opt`` on a C file from the SPEC benchmark suite, it gives a
426 report that looks like this:
430 7646 bitcodewriter - Number of normal instructions
431 725 bitcodewriter - Number of oversized instructions
432 129996 bitcodewriter - Number of bitcode bytes written
433 2817 raise - Number of insts DCEd or constprop'd
434 3213 raise - Number of cast-of-self removed
435 5046 raise - Number of expression trees converted
436 75 raise - Number of other getelementptr's formed
437 138 raise - Number of load/store peepholes
438 42 deadtypeelim - Number of unused typenames removed from symtab
439 392 funcresolve - Number of varargs functions resolved
440 27 globaldce - Number of global variables removed
441 2 adce - Number of basic blocks removed
442 134 cee - Number of branches revectored
443 49 cee - Number of setcc instruction eliminated
444 532 gcse - Number of loads removed
445 2919 gcse - Number of instructions removed
446 86 indvars - Number of canonical indvars added
447 87 indvars - Number of aux indvars removed
448 25 instcombine - Number of dead inst eliminate
449 434 instcombine - Number of insts combined
450 248 licm - Number of load insts hoisted
451 1298 licm - Number of insts hoisted to a loop pre-header
452 3 licm - Number of insts hoisted to multiple loop preds (bad, no loop pre-header)
453 75 mem2reg - Number of alloca's promoted
454 1444 cfgsimplify - Number of blocks simplified
456 Obviously, with so many optimizations, having a unified framework for this stuff
457 is very nice. Making your pass fit well into the framework makes it more
458 maintainable and useful.
462 Viewing graphs while debugging code
463 -----------------------------------
465 Several of the important data structures in LLVM are graphs: for example CFGs
466 made out of LLVM :ref:`BasicBlocks <BasicBlock>`, CFGs made out of LLVM
467 :ref:`MachineBasicBlocks <MachineBasicBlock>`, and :ref:`Instruction Selection
468 DAGs <SelectionDAG>`. In many cases, while debugging various parts of the
469 compiler, it is nice to instantly visualize these graphs.
471 LLVM provides several callbacks that are available in a debug build to do
472 exactly that. If you call the ``Function::viewCFG()`` method, for example, the
473 current LLVM tool will pop up a window containing the CFG for the function where
474 each basic block is a node in the graph, and each node contains the instructions
475 in the block. Similarly, there also exists ``Function::viewCFGOnly()`` (does
476 not include the instructions), the ``MachineFunction::viewCFG()`` and
477 ``MachineFunction::viewCFGOnly()``, and the ``SelectionDAG::viewGraph()``
478 methods. Within GDB, for example, you can usually use something like ``call
479 DAG.viewGraph()`` to pop up a window. Alternatively, you can sprinkle calls to
480 these functions in your code in places you want to debug.
482 Getting this to work requires a small amount of configuration. On Unix systems
483 with X11, install the `graphviz <http://www.graphviz.org>`_ toolkit, and make
484 sure 'dot' and 'gv' are in your path. If you are running on Mac OS/X, download
485 and install the Mac OS/X `Graphviz program
486 <http://www.pixelglow.com/graphviz/>`_ and add
487 ``/Applications/Graphviz.app/Contents/MacOS/`` (or wherever you install it) to
488 your path. Once in your system and path are set up, rerun the LLVM configure
489 script and rebuild LLVM to enable this functionality.
491 ``SelectionDAG`` has been extended to make it easier to locate *interesting*
492 nodes in large complex graphs. From gdb, if you ``call DAG.setGraphColor(node,
493 "color")``, then the next ``call DAG.viewGraph()`` would highlight the node in
494 the specified color (choices of colors can be found at `colors
495 <http://www.graphviz.org/doc/info/colors.html>`_.) More complex node attributes
496 can be provided with ``call DAG.setGraphAttrs(node, "attributes")`` (choices can
497 be found at `Graph attributes <http://www.graphviz.org/doc/info/attrs.html>`_.)
498 If you want to restart and clear all the current graph attributes, then you can
499 ``call DAG.clearGraphAttrs()``.
501 Note that graph visualization features are compiled out of Release builds to
502 reduce file size. This means that you need a Debug+Asserts or Release+Asserts
503 build to use these features.
507 Picking the Right Data Structure for a Task
508 ===========================================
510 LLVM has a plethora of data structures in the ``llvm/ADT/`` directory, and we
511 commonly use STL data structures. This section describes the trade-offs you
512 should consider when you pick one.
514 The first step is a choose your own adventure: do you want a sequential
515 container, a set-like container, or a map-like container? The most important
516 thing when choosing a container is the algorithmic properties of how you plan to
517 access the container. Based on that, you should use:
520 * a :ref:`map-like <ds_map>` container if you need efficient look-up of a
521 value based on another value. Map-like containers also support efficient
522 queries for containment (whether a key is in the map). Map-like containers
523 generally do not support efficient reverse mapping (values to keys). If you
524 need that, use two maps. Some map-like containers also support efficient
525 iteration through the keys in sorted order. Map-like containers are the most
526 expensive sort, only use them if you need one of these capabilities.
528 * a :ref:`set-like <ds_set>` container if you need to put a bunch of stuff into
529 a container that automatically eliminates duplicates. Some set-like
530 containers support efficient iteration through the elements in sorted order.
531 Set-like containers are more expensive than sequential containers.
533 * a :ref:`sequential <ds_sequential>` container provides the most efficient way
534 to add elements and keeps track of the order they are added to the collection.
535 They permit duplicates and support efficient iteration, but do not support
536 efficient look-up based on a key.
538 * a :ref:`string <ds_string>` container is a specialized sequential container or
539 reference structure that is used for character or byte arrays.
541 * a :ref:`bit <ds_bit>` container provides an efficient way to store and
542 perform set operations on sets of numeric id's, while automatically
543 eliminating duplicates. Bit containers require a maximum of 1 bit for each
544 identifier you want to store.
546 Once the proper category of container is determined, you can fine tune the
547 memory use, constant factors, and cache behaviors of access by intelligently
548 picking a member of the category. Note that constant factors and cache behavior
549 can be a big deal. If you have a vector that usually only contains a few
550 elements (but could contain many), for example, it's much better to use
551 :ref:`SmallVector <dss_smallvector>` than :ref:`vector <dss_vector>`. Doing so
552 avoids (relatively) expensive malloc/free calls, which dwarf the cost of adding
553 the elements to the container.
557 Sequential Containers (std::vector, std::list, etc)
558 ---------------------------------------------------
560 There are a variety of sequential containers available for you, based on your
561 needs. Pick the first in this section that will do what you want.
568 The ``llvm::ArrayRef`` class is the preferred class to use in an interface that
569 accepts a sequential list of elements in memory and just reads from them. By
570 taking an ``ArrayRef``, the API can be passed a fixed size array, an
571 ``std::vector``, an ``llvm::SmallVector`` and anything else that is contiguous
579 Fixed size arrays are very simple and very fast. They are good if you know
580 exactly how many elements you have, or you have a (low) upper bound on how many
585 Heap Allocated Arrays
586 ^^^^^^^^^^^^^^^^^^^^^
588 Heap allocated arrays (``new[]`` + ``delete[]``) are also simple. They are good
589 if the number of elements is variable, if you know how many elements you will
590 need before the array is allocated, and if the array is usually large (if not,
591 consider a :ref:`SmallVector <dss_smallvector>`). The cost of a heap allocated
592 array is the cost of the new/delete (aka malloc/free). Also note that if you
593 are allocating an array of a type with a constructor, the constructor and
594 destructors will be run for every element in the array (re-sizable vectors only
595 construct those elements actually used).
597 .. _dss_tinyptrvector:
599 llvm/ADT/TinyPtrVector.h
600 ^^^^^^^^^^^^^^^^^^^^^^^^
602 ``TinyPtrVector<Type>`` is a highly specialized collection class that is
603 optimized to avoid allocation in the case when a vector has zero or one
604 elements. It has two major restrictions: 1) it can only hold values of pointer
605 type, and 2) it cannot hold a null pointer.
607 Since this container is highly specialized, it is rarely used.
611 llvm/ADT/SmallVector.h
612 ^^^^^^^^^^^^^^^^^^^^^^
614 ``SmallVector<Type, N>`` is a simple class that looks and smells just like
615 ``vector<Type>``: it supports efficient iteration, lays out elements in memory
616 order (so you can do pointer arithmetic between elements), supports efficient
617 push_back/pop_back operations, supports efficient random access to its elements,
620 The advantage of SmallVector is that it allocates space for some number of
621 elements (N) **in the object itself**. Because of this, if the SmallVector is
622 dynamically smaller than N, no malloc is performed. This can be a big win in
623 cases where the malloc/free call is far more expensive than the code that
624 fiddles around with the elements.
626 This is good for vectors that are "usually small" (e.g. the number of
627 predecessors/successors of a block is usually less than 8). On the other hand,
628 this makes the size of the SmallVector itself large, so you don't want to
629 allocate lots of them (doing so will waste a lot of space). As such,
630 SmallVectors are most useful when on the stack.
632 SmallVector also provides a nice portable and efficient replacement for
640 ``std::vector`` is well loved and respected. It is useful when SmallVector
641 isn't: when the size of the vector is often large (thus the small optimization
642 will rarely be a benefit) or if you will be allocating many instances of the
643 vector itself (which would waste space for elements that aren't in the
644 container). vector is also useful when interfacing with code that expects
647 One worthwhile note about std::vector: avoid code like this:
656 Instead, write this as:
666 Doing so will save (at least) one heap allocation and free per iteration of the
674 ``std::deque`` is, in some senses, a generalized version of ``std::vector``.
675 Like ``std::vector``, it provides constant time random access and other similar
676 properties, but it also provides efficient access to the front of the list. It
677 does not guarantee continuity of elements within memory.
679 In exchange for this extra flexibility, ``std::deque`` has significantly higher
680 constant factor costs than ``std::vector``. If possible, use ``std::vector`` or
688 ``std::list`` is an extremely inefficient class that is rarely useful. It
689 performs a heap allocation for every element inserted into it, thus having an
690 extremely high constant factor, particularly for small data types.
691 ``std::list`` also only supports bidirectional iteration, not random access
694 In exchange for this high cost, std::list supports efficient access to both ends
695 of the list (like ``std::deque``, but unlike ``std::vector`` or
696 ``SmallVector``). In addition, the iterator invalidation characteristics of
697 std::list are stronger than that of a vector class: inserting or removing an
698 element into the list does not invalidate iterator or pointers to other elements
706 ``ilist<T>`` implements an 'intrusive' doubly-linked list. It is intrusive,
707 because it requires the element to store and provide access to the prev/next
708 pointers for the list.
710 ``ilist`` has the same drawbacks as ``std::list``, and additionally requires an
711 ``ilist_traits`` implementation for the element type, but it provides some novel
712 characteristics. In particular, it can efficiently store polymorphic objects,
713 the traits class is informed when an element is inserted or removed from the
714 list, and ``ilist``\ s are guaranteed to support a constant-time splice
717 These properties are exactly what we want for things like ``Instruction``\ s and
718 basic blocks, which is why these are implemented with ``ilist``\ s.
720 Related classes of interest are explained in the following subsections:
722 * :ref:`ilist_traits <dss_ilist_traits>`
724 * :ref:`iplist <dss_iplist>`
726 * :ref:`llvm/ADT/ilist_node.h <dss_ilist_node>`
728 * :ref:`Sentinels <dss_ilist_sentinel>`
730 .. _dss_packedvector:
732 llvm/ADT/PackedVector.h
733 ^^^^^^^^^^^^^^^^^^^^^^^
735 Useful for storing a vector of values using only a few number of bits for each
736 value. Apart from the standard operations of a vector-like container, it can
737 also perform an 'or' set operation.
745 FirstCondition = 0x1,
746 SecondCondition = 0x2,
751 PackedVector<State, 2> Vec1;
752 Vec1.push_back(FirstCondition);
754 PackedVector<State, 2> Vec2;
755 Vec2.push_back(SecondCondition);
758 return Vec1[0]; // returns 'Both'.
761 .. _dss_ilist_traits:
766 ``ilist_traits<T>`` is ``ilist<T>``'s customization mechanism. ``iplist<T>``
767 (and consequently ``ilist<T>``) publicly derive from this traits class.
774 ``iplist<T>`` is ``ilist<T>``'s base and as such supports a slightly narrower
775 interface. Notably, inserters from ``T&`` are absent.
777 ``ilist_traits<T>`` is a public base of this class and can be used for a wide
778 variety of customizations.
782 llvm/ADT/ilist_node.h
783 ^^^^^^^^^^^^^^^^^^^^^
785 ``ilist_node<T>`` implements a the forward and backward links that are expected
786 by the ``ilist<T>`` (and analogous containers) in the default manner.
788 ``ilist_node<T>``\ s are meant to be embedded in the node type ``T``, usually
789 ``T`` publicly derives from ``ilist_node<T>``.
791 .. _dss_ilist_sentinel:
796 ``ilist``\ s have another specialty that must be considered. To be a good
797 citizen in the C++ ecosystem, it needs to support the standard container
798 operations, such as ``begin`` and ``end`` iterators, etc. Also, the
799 ``operator--`` must work correctly on the ``end`` iterator in the case of
800 non-empty ``ilist``\ s.
802 The only sensible solution to this problem is to allocate a so-called *sentinel*
803 along with the intrusive list, which serves as the ``end`` iterator, providing
804 the back-link to the last element. However conforming to the C++ convention it
805 is illegal to ``operator++`` beyond the sentinel and it also must not be
808 These constraints allow for some implementation freedom to the ``ilist`` how to
809 allocate and store the sentinel. The corresponding policy is dictated by
810 ``ilist_traits<T>``. By default a ``T`` gets heap-allocated whenever the need
811 for a sentinel arises.
813 While the default policy is sufficient in most cases, it may break down when
814 ``T`` does not provide a default constructor. Also, in the case of many
815 instances of ``ilist``\ s, the memory overhead of the associated sentinels is
816 wasted. To alleviate the situation with numerous and voluminous
817 ``T``-sentinels, sometimes a trick is employed, leading to *ghostly sentinels*.
819 Ghostly sentinels are obtained by specially-crafted ``ilist_traits<T>`` which
820 superpose the sentinel with the ``ilist`` instance in memory. Pointer
821 arithmetic is used to obtain the sentinel, which is relative to the ``ilist``'s
822 ``this`` pointer. The ``ilist`` is augmented by an extra pointer, which serves
823 as the back-link of the sentinel. This is the only field in the ghostly
824 sentinel which can be legally accessed.
828 Other Sequential Container options
829 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
831 Other STL containers are available, such as ``std::string``.
833 There are also various STL adapter classes such as ``std::queue``,
834 ``std::priority_queue``, ``std::stack``, etc. These provide simplified access
835 to an underlying container but don't affect the cost of the container itself.
839 String-like containers
840 ----------------------
842 There are a variety of ways to pass around and use strings in C and C++, and
843 LLVM adds a few new options to choose from. Pick the first option on this list
844 that will do what you need, they are ordered according to their relative cost.
846 Note that is is generally preferred to *not* pass strings around as ``const
847 char*``'s. These have a number of problems, including the fact that they
848 cannot represent embedded nul ("\0") characters, and do not have a length
849 available efficiently. The general replacement for '``const char*``' is
852 For more information on choosing string containers for APIs, please see
853 :ref:`Passing Strings <string_apis>`.
860 The StringRef class is a simple value class that contains a pointer to a
861 character and a length, and is quite related to the :ref:`ArrayRef
862 <dss_arrayref>` class (but specialized for arrays of characters). Because
863 StringRef carries a length with it, it safely handles strings with embedded nul
864 characters in it, getting the length does not require a strlen call, and it even
865 has very convenient APIs for slicing and dicing the character range that it
868 StringRef is ideal for passing simple strings around that are known to be live,
869 either because they are C string literals, std::string, a C array, or a
870 SmallVector. Each of these cases has an efficient implicit conversion to
871 StringRef, which doesn't result in a dynamic strlen being executed.
873 StringRef has a few major limitations which make more powerful string containers
876 #. You cannot directly convert a StringRef to a 'const char*' because there is
877 no way to add a trailing nul (unlike the .c_str() method on various stronger
880 #. StringRef doesn't own or keep alive the underlying string bytes.
881 As such it can easily lead to dangling pointers, and is not suitable for
882 embedding in datastructures in most cases (instead, use an std::string or
883 something like that).
885 #. For the same reason, StringRef cannot be used as the return value of a
886 method if the method "computes" the result string. Instead, use std::string.
888 #. StringRef's do not allow you to mutate the pointed-to string bytes and it
889 doesn't allow you to insert or remove bytes from the range. For editing
890 operations like this, it interoperates with the :ref:`Twine <dss_twine>`
893 Because of its strengths and limitations, it is very common for a function to
894 take a StringRef and for a method on an object to return a StringRef that points
895 into some string that it owns.
902 The Twine class is used as an intermediary datatype for APIs that want to take a
903 string that can be constructed inline with a series of concatenations. Twine
904 works by forming recursive instances of the Twine datatype (a simple value
905 object) on the stack as temporary objects, linking them together into a tree
906 which is then linearized when the Twine is consumed. Twine is only safe to use
907 as the argument to a function, and should always be a const reference, e.g.:
911 void foo(const Twine &T);
915 foo(X + "." + Twine(i));
917 This example forms a string like "blarg.42" by concatenating the values
918 together, and does not form intermediate strings containing "blarg" or "blarg.".
920 Because Twine is constructed with temporary objects on the stack, and because
921 these instances are destroyed at the end of the current statement, it is an
922 inherently dangerous API. For example, this simple variant contains undefined
923 behavior and will probably crash:
927 void foo(const Twine &T);
931 const Twine &Tmp = X + "." + Twine(i);
934 ... because the temporaries are destroyed before the call. That said, Twine's
935 are much more efficient than intermediate std::string temporaries, and they work
936 really well with StringRef. Just be aware of their limitations.
940 llvm/ADT/SmallString.h
941 ^^^^^^^^^^^^^^^^^^^^^^
943 SmallString is a subclass of :ref:`SmallVector <dss_smallvector>` that adds some
944 convenience APIs like += that takes StringRef's. SmallString avoids allocating
945 memory in the case when the preallocated space is enough to hold its data, and
946 it calls back to general heap allocation when required. Since it owns its data,
947 it is very safe to use and supports full mutation of the string.
949 Like SmallVector's, the big downside to SmallString is their sizeof. While they
950 are optimized for small strings, they themselves are not particularly small.
951 This means that they work great for temporary scratch buffers on the stack, but
952 should not generally be put into the heap: it is very rare to see a SmallString
953 as the member of a frequently-allocated heap data structure or returned
961 The standard C++ std::string class is a very general class that (like
962 SmallString) owns its underlying data. sizeof(std::string) is very reasonable
963 so it can be embedded into heap data structures and returned by-value. On the
964 other hand, std::string is highly inefficient for inline editing (e.g.
965 concatenating a bunch of stuff together) and because it is provided by the
966 standard library, its performance characteristics depend a lot of the host
967 standard library (e.g. libc++ and MSVC provide a highly optimized string class,
968 GCC contains a really slow implementation).
970 The major disadvantage of std::string is that almost every operation that makes
971 them larger can allocate memory, which is slow. As such, it is better to use
972 SmallVector or Twine as a scratch buffer, but then use std::string to persist
977 Set-Like Containers (std::set, SmallSet, SetVector, etc)
978 --------------------------------------------------------
980 Set-like containers are useful when you need to canonicalize multiple values
981 into a single representation. There are several different choices for how to do
982 this, providing various trade-offs.
984 .. _dss_sortedvectorset:
989 If you intend to insert a lot of elements, then do a lot of queries, a great
990 approach is to use a vector (or other sequential container) with
991 std::sort+std::unique to remove duplicates. This approach works really well if
992 your usage pattern has these two distinct phases (insert then query), and can be
993 coupled with a good choice of :ref:`sequential container <ds_sequential>`.
995 This combination provides the several nice properties: the result data is
996 contiguous in memory (good for cache locality), has few allocations, is easy to
997 address (iterators in the final vector are just indices or pointers), and can be
998 efficiently queried with a standard binary or radix search.
1005 If you have a set-like data structure that is usually small and whose elements
1006 are reasonably small, a ``SmallSet<Type, N>`` is a good choice. This set has
1007 space for N elements in place (thus, if the set is dynamically smaller than N,
1008 no malloc traffic is required) and accesses them with a simple linear search.
1009 When the set grows beyond 'N' elements, it allocates a more expensive
1010 representation that guarantees efficient access (for most types, it falls back
1011 to std::set, but for pointers it uses something far better, :ref:`SmallPtrSet
1014 The magic of this class is that it handles small sets extremely efficiently, but
1015 gracefully handles extremely large sets without loss of efficiency. The
1016 drawback is that the interface is quite small: it supports insertion, queries
1017 and erasing, but does not support iteration.
1019 .. _dss_smallptrset:
1021 llvm/ADT/SmallPtrSet.h
1022 ^^^^^^^^^^^^^^^^^^^^^^
1024 SmallPtrSet has all the advantages of ``SmallSet`` (and a ``SmallSet`` of
1025 pointers is transparently implemented with a ``SmallPtrSet``), but also supports
1026 iterators. If more than 'N' insertions are performed, a single quadratically
1027 probed hash table is allocated and grows as needed, providing extremely
1028 efficient access (constant time insertion/deleting/queries with low constant
1029 factors) and is very stingy with malloc traffic.
1031 Note that, unlike ``std::set``, the iterators of ``SmallPtrSet`` are invalidated
1032 whenever an insertion occurs. Also, the values visited by the iterators are not
1033 visited in sorted order.
1040 DenseSet is a simple quadratically probed hash table. It excels at supporting
1041 small values: it uses a single allocation to hold all of the pairs that are
1042 currently inserted in the set. DenseSet is a great way to unique small values
1043 that are not simple pointers (use :ref:`SmallPtrSet <dss_smallptrset>` for
1044 pointers). Note that DenseSet has the same requirements for the value type that
1045 :ref:`DenseMap <dss_densemap>` has.
1049 llvm/ADT/SparseSet.h
1050 ^^^^^^^^^^^^^^^^^^^^
1052 SparseSet holds a small number of objects identified by unsigned keys of
1053 moderate size. It uses a lot of memory, but provides operations that are almost
1054 as fast as a vector. Typical keys are physical registers, virtual registers, or
1055 numbered basic blocks.
1057 SparseSet is useful for algorithms that need very fast clear/find/insert/erase
1058 and fast iteration over small sets. It is not intended for building composite
1063 llvm/ADT/FoldingSet.h
1064 ^^^^^^^^^^^^^^^^^^^^^
1066 FoldingSet is an aggregate class that is really good at uniquing
1067 expensive-to-create or polymorphic objects. It is a combination of a chained
1068 hash table with intrusive links (uniqued objects are required to inherit from
1069 FoldingSetNode) that uses :ref:`SmallVector <dss_smallvector>` as part of its ID
1072 Consider a case where you want to implement a "getOrCreateFoo" method for a
1073 complex object (for example, a node in the code generator). The client has a
1074 description of **what** it wants to generate (it knows the opcode and all the
1075 operands), but we don't want to 'new' a node, then try inserting it into a set
1076 only to find out it already exists, at which point we would have to delete it
1077 and return the node that already exists.
1079 To support this style of client, FoldingSet perform a query with a
1080 FoldingSetNodeID (which wraps SmallVector) that can be used to describe the
1081 element that we want to query for. The query either returns the element
1082 matching the ID or it returns an opaque ID that indicates where insertion should
1083 take place. Construction of the ID usually does not require heap traffic.
1085 Because FoldingSet uses intrusive links, it can support polymorphic objects in
1086 the set (for example, you can have SDNode instances mixed with LoadSDNodes).
1087 Because the elements are individually allocated, pointers to the elements are
1088 stable: inserting or removing elements does not invalidate any pointers to other
1096 ``std::set`` is a reasonable all-around set class, which is decent at many
1097 things but great at nothing. std::set allocates memory for each element
1098 inserted (thus it is very malloc intensive) and typically stores three pointers
1099 per element in the set (thus adding a large amount of per-element space
1100 overhead). It offers guaranteed log(n) performance, which is not particularly
1101 fast from a complexity standpoint (particularly if the elements of the set are
1102 expensive to compare, like strings), and has extremely high constant factors for
1103 lookup, insertion and removal.
1105 The advantages of std::set are that its iterators are stable (deleting or
1106 inserting an element from the set does not affect iterators or pointers to other
1107 elements) and that iteration over the set is guaranteed to be in sorted order.
1108 If the elements in the set are large, then the relative overhead of the pointers
1109 and malloc traffic is not a big deal, but if the elements of the set are small,
1110 std::set is almost never a good choice.
1114 llvm/ADT/SetVector.h
1115 ^^^^^^^^^^^^^^^^^^^^
1117 LLVM's ``SetVector<Type>`` is an adapter class that combines your choice of a
1118 set-like container along with a :ref:`Sequential Container <ds_sequential>` The
1119 important property that this provides is efficient insertion with uniquing
1120 (duplicate elements are ignored) with iteration support. It implements this by
1121 inserting elements into both a set-like container and the sequential container,
1122 using the set-like container for uniquing and the sequential container for
1125 The difference between SetVector and other sets is that the order of iteration
1126 is guaranteed to match the order of insertion into the SetVector. This property
1127 is really important for things like sets of pointers. Because pointer values
1128 are non-deterministic (e.g. vary across runs of the program on different
1129 machines), iterating over the pointers in the set will not be in a well-defined
1132 The drawback of SetVector is that it requires twice as much space as a normal
1133 set and has the sum of constant factors from the set-like container and the
1134 sequential container that it uses. Use it **only** if you need to iterate over
1135 the elements in a deterministic order. SetVector is also expensive to delete
1136 elements out of (linear time), unless you use it's "pop_back" method, which is
1139 ``SetVector`` is an adapter class that defaults to using ``std::vector`` and a
1140 size 16 ``SmallSet`` for the underlying containers, so it is quite expensive.
1141 However, ``"llvm/ADT/SetVector.h"`` also provides a ``SmallSetVector`` class,
1142 which defaults to using a ``SmallVector`` and ``SmallSet`` of a specified size.
1143 If you use this, and if your sets are dynamically smaller than ``N``, you will
1144 save a lot of heap traffic.
1146 .. _dss_uniquevector:
1148 llvm/ADT/UniqueVector.h
1149 ^^^^^^^^^^^^^^^^^^^^^^^
1151 UniqueVector is similar to :ref:`SetVector <dss_setvector>` but it retains a
1152 unique ID for each element inserted into the set. It internally contains a map
1153 and a vector, and it assigns a unique ID for each value inserted into the set.
1155 UniqueVector is very expensive: its cost is the sum of the cost of maintaining
1156 both the map and vector, it has high complexity, high constant factors, and
1157 produces a lot of malloc traffic. It should be avoided.
1159 .. _dss_immutableset:
1161 llvm/ADT/ImmutableSet.h
1162 ^^^^^^^^^^^^^^^^^^^^^^^
1164 ImmutableSet is an immutable (functional) set implementation based on an AVL
1165 tree. Adding or removing elements is done through a Factory object and results
1166 in the creation of a new ImmutableSet object. If an ImmutableSet already exists
1167 with the given contents, then the existing one is returned; equality is compared
1168 with a FoldingSetNodeID. The time and space complexity of add or remove
1169 operations is logarithmic in the size of the original set.
1171 There is no method for returning an element of the set, you can only check for
1176 Other Set-Like Container Options
1177 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1179 The STL provides several other options, such as std::multiset and the various
1180 "hash_set" like containers (whether from C++ TR1 or from the SGI library). We
1181 never use hash_set and unordered_set because they are generally very expensive
1182 (each insertion requires a malloc) and very non-portable.
1184 std::multiset is useful if you're not interested in elimination of duplicates,
1185 but has all the drawbacks of std::set. A sorted vector (where you don't delete
1186 duplicate entries) or some other approach is almost always better.
1190 Map-Like Containers (std::map, DenseMap, etc)
1191 ---------------------------------------------
1193 Map-like containers are useful when you want to associate data to a key. As
1194 usual, there are a lot of different ways to do this. :)
1196 .. _dss_sortedvectormap:
1201 If your usage pattern follows a strict insert-then-query approach, you can
1202 trivially use the same approach as :ref:`sorted vectors for set-like containers
1203 <dss_sortedvectorset>`. The only difference is that your query function (which
1204 uses std::lower_bound to get efficient log(n) lookup) should only compare the
1205 key, not both the key and value. This yields the same advantages as sorted
1210 llvm/ADT/StringMap.h
1211 ^^^^^^^^^^^^^^^^^^^^
1213 Strings are commonly used as keys in maps, and they are difficult to support
1214 efficiently: they are variable length, inefficient to hash and compare when
1215 long, expensive to copy, etc. StringMap is a specialized container designed to
1216 cope with these issues. It supports mapping an arbitrary range of bytes to an
1217 arbitrary other object.
1219 The StringMap implementation uses a quadratically-probed hash table, where the
1220 buckets store a pointer to the heap allocated entries (and some other stuff).
1221 The entries in the map must be heap allocated because the strings are variable
1222 length. The string data (key) and the element object (value) are stored in the
1223 same allocation with the string data immediately after the element object.
1224 This container guarantees the "``(char*)(&Value+1)``" points to the key string
1227 The StringMap is very fast for several reasons: quadratic probing is very cache
1228 efficient for lookups, the hash value of strings in buckets is not recomputed
1229 when looking up an element, StringMap rarely has to touch the memory for
1230 unrelated objects when looking up a value (even when hash collisions happen),
1231 hash table growth does not recompute the hash values for strings already in the
1232 table, and each pair in the map is store in a single allocation (the string data
1233 is stored in the same allocation as the Value of a pair).
1235 StringMap also provides query methods that take byte ranges, so it only ever
1236 copies a string if a value is inserted into the table.
1238 StringMap iteratation order, however, is not guaranteed to be deterministic, so
1239 any uses which require that should instead use a std::map.
1243 llvm/ADT/IndexedMap.h
1244 ^^^^^^^^^^^^^^^^^^^^^
1246 IndexedMap is a specialized container for mapping small dense integers (or
1247 values that can be mapped to small dense integers) to some other type. It is
1248 internally implemented as a vector with a mapping function that maps the keys
1249 to the dense integer range.
1251 This is useful for cases like virtual registers in the LLVM code generator: they
1252 have a dense mapping that is offset by a compile-time constant (the first
1253 virtual register ID).
1260 DenseMap is a simple quadratically probed hash table. It excels at supporting
1261 small keys and values: it uses a single allocation to hold all of the pairs
1262 that are currently inserted in the map. DenseMap is a great way to map
1263 pointers to pointers, or map other small types to each other.
1265 There are several aspects of DenseMap that you should be aware of, however.
1266 The iterators in a DenseMap are invalidated whenever an insertion occurs,
1267 unlike map. Also, because DenseMap allocates space for a large number of
1268 key/value pairs (it starts with 64 by default), it will waste a lot of space if
1269 your keys or values are large. Finally, you must implement a partial
1270 specialization of DenseMapInfo for the key that you want, if it isn't already
1271 supported. This is required to tell DenseMap about two special marker values
1272 (which can never be inserted into the map) that it needs internally.
1274 DenseMap's find_as() method supports lookup operations using an alternate key
1275 type. This is useful in cases where the normal key type is expensive to
1276 construct, but cheap to compare against. The DenseMapInfo is responsible for
1277 defining the appropriate comparison and hashing methods for each alternate key
1285 ValueMap is a wrapper around a :ref:`DenseMap <dss_densemap>` mapping
1286 ``Value*``\ s (or subclasses) to another type. When a Value is deleted or
1287 RAUW'ed, ValueMap will update itself so the new version of the key is mapped to
1288 the same value, just as if the key were a WeakVH. You can configure exactly how
1289 this happens, and what else happens on these two events, by passing a ``Config``
1290 parameter to the ValueMap template.
1292 .. _dss_intervalmap:
1294 llvm/ADT/IntervalMap.h
1295 ^^^^^^^^^^^^^^^^^^^^^^
1297 IntervalMap is a compact map for small keys and values. It maps key intervals
1298 instead of single keys, and it will automatically coalesce adjacent intervals.
1299 When then map only contains a few intervals, they are stored in the map object
1300 itself to avoid allocations.
1302 The IntervalMap iterators are quite big, so they should not be passed around as
1303 STL iterators. The heavyweight iterators allow a smaller data structure.
1310 std::map has similar characteristics to :ref:`std::set <dss_set>`: it uses a
1311 single allocation per pair inserted into the map, it offers log(n) lookup with
1312 an extremely large constant factor, imposes a space penalty of 3 pointers per
1313 pair in the map, etc.
1315 std::map is most useful when your keys or values are very large, if you need to
1316 iterate over the collection in sorted order, or if you need stable iterators
1317 into the map (i.e. they don't get invalidated if an insertion or deletion of
1318 another element takes place).
1322 llvm/ADT/MapVector.h
1323 ^^^^^^^^^^^^^^^^^^^^
1325 ``MapVector<KeyT,ValueT>`` provides a subset of the DenseMap interface. The
1326 main difference is that the iteration order is guaranteed to be the insertion
1327 order, making it an easy (but somewhat expensive) solution for non-deterministic
1328 iteration over maps of pointers.
1330 It is implemented by mapping from key to an index in a vector of key,value
1331 pairs. This provides fast lookup and iteration, but has two main drawbacks: The
1332 key is stored twice and it doesn't support removing elements.
1334 .. _dss_inteqclasses:
1336 llvm/ADT/IntEqClasses.h
1337 ^^^^^^^^^^^^^^^^^^^^^^^
1339 IntEqClasses provides a compact representation of equivalence classes of small
1340 integers. Initially, each integer in the range 0..n-1 has its own equivalence
1341 class. Classes can be joined by passing two class representatives to the
1342 join(a, b) method. Two integers are in the same class when findLeader() returns
1343 the same representative.
1345 Once all equivalence classes are formed, the map can be compressed so each
1346 integer 0..n-1 maps to an equivalence class number in the range 0..m-1, where m
1347 is the total number of equivalence classes. The map must be uncompressed before
1348 it can be edited again.
1350 .. _dss_immutablemap:
1352 llvm/ADT/ImmutableMap.h
1353 ^^^^^^^^^^^^^^^^^^^^^^^
1355 ImmutableMap is an immutable (functional) map implementation based on an AVL
1356 tree. Adding or removing elements is done through a Factory object and results
1357 in the creation of a new ImmutableMap object. If an ImmutableMap already exists
1358 with the given key set, then the existing one is returned; equality is compared
1359 with a FoldingSetNodeID. The time and space complexity of add or remove
1360 operations is logarithmic in the size of the original map.
1364 Other Map-Like Container Options
1365 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1367 The STL provides several other options, such as std::multimap and the various
1368 "hash_map" like containers (whether from C++ TR1 or from the SGI library). We
1369 never use hash_set and unordered_set because they are generally very expensive
1370 (each insertion requires a malloc) and very non-portable.
1372 std::multimap is useful if you want to map a key to multiple values, but has all
1373 the drawbacks of std::map. A sorted vector or some other approach is almost
1378 Bit storage containers (BitVector, SparseBitVector)
1379 ---------------------------------------------------
1381 Unlike the other containers, there are only two bit storage containers, and
1382 choosing when to use each is relatively straightforward.
1384 One additional option is ``std::vector<bool>``: we discourage its use for two
1385 reasons 1) the implementation in many common compilers (e.g. commonly
1386 available versions of GCC) is extremely inefficient and 2) the C++ standards
1387 committee is likely to deprecate this container and/or change it significantly
1388 somehow. In any case, please don't use it.
1395 The BitVector container provides a dynamic size set of bits for manipulation.
1396 It supports individual bit setting/testing, as well as set operations. The set
1397 operations take time O(size of bitvector), but operations are performed one word
1398 at a time, instead of one bit at a time. This makes the BitVector very fast for
1399 set operations compared to other containers. Use the BitVector when you expect
1400 the number of set bits to be high (i.e. a dense set).
1402 .. _dss_smallbitvector:
1407 The SmallBitVector container provides the same interface as BitVector, but it is
1408 optimized for the case where only a small number of bits, less than 25 or so,
1409 are needed. It also transparently supports larger bit counts, but slightly less
1410 efficiently than a plain BitVector, so SmallBitVector should only be used when
1411 larger counts are rare.
1413 At this time, SmallBitVector does not support set operations (and, or, xor), and
1414 its operator[] does not provide an assignable lvalue.
1416 .. _dss_sparsebitvector:
1421 The SparseBitVector container is much like BitVector, with one major difference:
1422 Only the bits that are set, are stored. This makes the SparseBitVector much
1423 more space efficient than BitVector when the set is sparse, as well as making
1424 set operations O(number of set bits) instead of O(size of universe). The
1425 downside to the SparseBitVector is that setting and testing of random bits is
1426 O(N), and on large SparseBitVectors, this can be slower than BitVector. In our
1427 implementation, setting or testing bits in sorted order (either forwards or
1428 reverse) is O(1) worst case. Testing and setting bits within 128 bits (depends
1429 on size) of the current bit is also O(1). As a general statement,
1430 testing/setting bits in a SparseBitVector is O(distance away from last set bit).
1434 Helpful Hints for Common Operations
1435 ===================================
1437 This section describes how to perform some very simple transformations of LLVM
1438 code. This is meant to give examples of common idioms used, showing the
1439 practical side of LLVM transformations.
1441 Because this is a "how-to" section, you should also read about the main classes
1442 that you will be working with. The :ref:`Core LLVM Class Hierarchy Reference
1443 <coreclasses>` contains details and descriptions of the main classes that you
1448 Basic Inspection and Traversal Routines
1449 ---------------------------------------
1451 The LLVM compiler infrastructure have many different data structures that may be
1452 traversed. Following the example of the C++ standard template library, the
1453 techniques used to traverse these various data structures are all basically the
1454 same. For a enumerable sequence of values, the ``XXXbegin()`` function (or
1455 method) returns an iterator to the start of the sequence, the ``XXXend()``
1456 function returns an iterator pointing to one past the last valid element of the
1457 sequence, and there is some ``XXXiterator`` data type that is common between the
1460 Because the pattern for iteration is common across many different aspects of the
1461 program representation, the standard template library algorithms may be used on
1462 them, and it is easier to remember how to iterate. First we show a few common
1463 examples of the data structures that need to be traversed. Other data
1464 structures are traversed in very similar ways.
1466 .. _iterate_function:
1468 Iterating over the ``BasicBlock`` in a ``Function``
1469 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1471 It's quite common to have a ``Function`` instance that you'd like to transform
1472 in some way; in particular, you'd like to manipulate its ``BasicBlock``\ s. To
1473 facilitate this, you'll need to iterate over all of the ``BasicBlock``\ s that
1474 constitute the ``Function``. The following is an example that prints the name
1475 of a ``BasicBlock`` and the number of ``Instruction``\ s it contains:
1479 // func is a pointer to a Function instance
1480 for (Function::iterator i = func->begin(), e = func->end(); i != e; ++i)
1481 // Print out the name of the basic block if it has one, and then the
1482 // number of instructions that it contains
1483 errs() << "Basic block (name=" << i->getName() << ") has "
1484 << i->size() << " instructions.\n";
1486 Note that i can be used as if it were a pointer for the purposes of invoking
1487 member functions of the ``Instruction`` class. This is because the indirection
1488 operator is overloaded for the iterator classes. In the above code, the
1489 expression ``i->size()`` is exactly equivalent to ``(*i).size()`` just like
1492 .. _iterate_basicblock:
1494 Iterating over the ``Instruction`` in a ``BasicBlock``
1495 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1497 Just like when dealing with ``BasicBlock``\ s in ``Function``\ s, it's easy to
1498 iterate over the individual instructions that make up ``BasicBlock``\ s. Here's
1499 a code snippet that prints out each instruction in a ``BasicBlock``:
1503 // blk is a pointer to a BasicBlock instance
1504 for (BasicBlock::iterator i = blk->begin(), e = blk->end(); i != e; ++i)
1505 // The next statement works since operator<<(ostream&,...)
1506 // is overloaded for Instruction&
1507 errs() << *i << "\n";
1510 However, this isn't really the best way to print out the contents of a
1511 ``BasicBlock``! Since the ostream operators are overloaded for virtually
1512 anything you'll care about, you could have just invoked the print routine on the
1513 basic block itself: ``errs() << *blk << "\n";``.
1515 .. _iterate_insiter:
1517 Iterating over the ``Instruction`` in a ``Function``
1518 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1520 If you're finding that you commonly iterate over a ``Function``'s
1521 ``BasicBlock``\ s and then that ``BasicBlock``'s ``Instruction``\ s,
1522 ``InstIterator`` should be used instead. You'll need to include
1523 ``llvm/Support/InstIterator.h`` (`doxygen
1524 <http://llvm.org/doxygen/InstIterator_8h-source.html>`__) and then instantiate
1525 ``InstIterator``\ s explicitly in your code. Here's a small example that shows
1526 how to dump all instructions in a function to the standard error stream:
1530 #include "llvm/Support/InstIterator.h"
1532 // F is a pointer to a Function instance
1533 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
1534 errs() << *I << "\n";
1536 Easy, isn't it? You can also use ``InstIterator``\ s to fill a work list with
1537 its initial contents. For example, if you wanted to initialize a work list to
1538 contain all instructions in a ``Function`` F, all you would need to do is
1543 std::set<Instruction*> worklist;
1544 // or better yet, SmallPtrSet<Instruction*, 64> worklist;
1546 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
1547 worklist.insert(&*I);
1549 The STL set ``worklist`` would now contain all instructions in the ``Function``
1552 .. _iterate_convert:
1554 Turning an iterator into a class pointer (and vice-versa)
1555 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1557 Sometimes, it'll be useful to grab a reference (or pointer) to a class instance
1558 when all you've got at hand is an iterator. Well, extracting a reference or a
1559 pointer from an iterator is very straight-forward. Assuming that ``i`` is a
1560 ``BasicBlock::iterator`` and ``j`` is a ``BasicBlock::const_iterator``:
1564 Instruction& inst = *i; // Grab reference to instruction reference
1565 Instruction* pinst = &*i; // Grab pointer to instruction reference
1566 const Instruction& inst = *j;
1568 However, the iterators you'll be working with in the LLVM framework are special:
1569 they will automatically convert to a ptr-to-instance type whenever they need to.
1570 Instead of derferencing the iterator and then taking the address of the result,
1571 you can simply assign the iterator to the proper pointer type and you get the
1572 dereference and address-of operation as a result of the assignment (behind the
1573 scenes, this is a result of overloading casting mechanisms). Thus the last line
1574 of the last example,
1578 Instruction *pinst = &*i;
1580 is semantically equivalent to
1584 Instruction *pinst = i;
1586 It's also possible to turn a class pointer into the corresponding iterator, and
1587 this is a constant time operation (very efficient). The following code snippet
1588 illustrates use of the conversion constructors provided by LLVM iterators. By
1589 using these, you can explicitly grab the iterator of something without actually
1590 obtaining it via iteration over some structure:
1594 void printNextInstruction(Instruction* inst) {
1595 BasicBlock::iterator it(inst);
1596 ++it; // After this line, it refers to the instruction after *inst
1597 if (it != inst->getParent()->end()) errs() << *it << "\n";
1600 Unfortunately, these implicit conversions come at a cost; they prevent these
1601 iterators from conforming to standard iterator conventions, and thus from being
1602 usable with standard algorithms and containers. For example, they prevent the
1603 following code, where ``B`` is a ``BasicBlock``, from compiling:
1607 llvm::SmallVector<llvm::Instruction *, 16>(B->begin(), B->end());
1609 Because of this, these implicit conversions may be removed some day, and
1610 ``operator*`` changed to return a pointer instead of a reference.
1612 .. _iterate_complex:
1614 Finding call sites: a slightly more complex example
1615 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1617 Say that you're writing a FunctionPass and would like to count all the locations
1618 in the entire module (that is, across every ``Function``) where a certain
1619 function (i.e., some ``Function *``) is already in scope. As you'll learn
1620 later, you may want to use an ``InstVisitor`` to accomplish this in a much more
1621 straight-forward manner, but this example will allow us to explore how you'd do
1622 it if you didn't have ``InstVisitor`` around. In pseudo-code, this is what we
1625 .. code-block:: none
1627 initialize callCounter to zero
1628 for each Function f in the Module
1629 for each BasicBlock b in f
1630 for each Instruction i in b
1631 if (i is a CallInst and calls the given function)
1632 increment callCounter
1634 And the actual code is (remember, because we're writing a ``FunctionPass``, our
1635 ``FunctionPass``-derived class simply has to override the ``runOnFunction``
1640 Function* targetFunc = ...;
1642 class OurFunctionPass : public FunctionPass {
1644 OurFunctionPass(): callCounter(0) { }
1646 virtual runOnFunction(Function& F) {
1647 for (Function::iterator b = F.begin(), be = F.end(); b != be; ++b) {
1648 for (BasicBlock::iterator i = b->begin(), ie = b->end(); i != ie; ++i) {
1649 if (CallInst* callInst = dyn_cast<CallInst>(&*i)) {
1650 // We know we've encountered a call instruction, so we
1651 // need to determine if it's a call to the
1652 // function pointed to by m_func or not.
1653 if (callInst->getCalledFunction() == targetFunc)
1661 unsigned callCounter;
1664 .. _calls_and_invokes:
1666 Treating calls and invokes the same way
1667 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1669 You may have noticed that the previous example was a bit oversimplified in that
1670 it did not deal with call sites generated by 'invoke' instructions. In this,
1671 and in other situations, you may find that you want to treat ``CallInst``\ s and
1672 ``InvokeInst``\ s the same way, even though their most-specific common base
1673 class is ``Instruction``, which includes lots of less closely-related things.
1674 For these cases, LLVM provides a handy wrapper class called ``CallSite``
1675 (`doxygen <http://llvm.org/doxygen/classllvm_1_1CallSite.html>`__) It is
1676 essentially a wrapper around an ``Instruction`` pointer, with some methods that
1677 provide functionality common to ``CallInst``\ s and ``InvokeInst``\ s.
1679 This class has "value semantics": it should be passed by value, not by reference
1680 and it should not be dynamically allocated or deallocated using ``operator new``
1681 or ``operator delete``. It is efficiently copyable, assignable and
1682 constructable, with costs equivalents to that of a bare pointer. If you look at
1683 its definition, it has only a single pointer member.
1687 Iterating over def-use & use-def chains
1688 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1690 Frequently, we might have an instance of the ``Value`` class (`doxygen
1691 <http://llvm.org/doxygen/classllvm_1_1Value.html>`__) and we want to determine
1692 which ``User`` s use the ``Value``. The list of all ``User``\ s of a particular
1693 ``Value`` is called a *def-use* chain. For example, let's say we have a
1694 ``Function*`` named ``F`` to a particular function ``foo``. Finding all of the
1695 instructions that *use* ``foo`` is as simple as iterating over the *def-use*
1702 for (Value::use_iterator i = F->use_begin(), e = F->use_end(); i != e; ++i)
1703 if (Instruction *Inst = dyn_cast<Instruction>(*i)) {
1704 errs() << "F is used in instruction:\n";
1705 errs() << *Inst << "\n";
1708 Note that dereferencing a ``Value::use_iterator`` is not a very cheap operation.
1709 Instead of performing ``*i`` above several times, consider doing it only once in
1710 the loop body and reusing its result.
1712 Alternatively, it's common to have an instance of the ``User`` Class (`doxygen
1713 <http://llvm.org/doxygen/classllvm_1_1User.html>`__) and need to know what
1714 ``Value``\ s are used by it. The list of all ``Value``\ s used by a ``User`` is
1715 known as a *use-def* chain. Instances of class ``Instruction`` are common
1716 ``User`` s, so we might want to iterate over all of the values that a particular
1717 instruction uses (that is, the operands of the particular ``Instruction``):
1721 Instruction *pi = ...;
1723 for (User::op_iterator i = pi->op_begin(), e = pi->op_end(); i != e; ++i) {
1728 Declaring objects as ``const`` is an important tool of enforcing mutation free
1729 algorithms (such as analyses, etc.). For this purpose above iterators come in
1730 constant flavors as ``Value::const_use_iterator`` and
1731 ``Value::const_op_iterator``. They automatically arise when calling
1732 ``use/op_begin()`` on ``const Value*``\ s or ``const User*``\ s respectively.
1733 Upon dereferencing, they return ``const Use*``\ s. Otherwise the above patterns
1738 Iterating over predecessors & successors of blocks
1739 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1741 Iterating over the predecessors and successors of a block is quite easy with the
1742 routines defined in ``"llvm/Support/CFG.h"``. Just use code like this to
1743 iterate over all predecessors of BB:
1747 #include "llvm/Support/CFG.h"
1748 BasicBlock *BB = ...;
1750 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1751 BasicBlock *Pred = *PI;
1755 Similarly, to iterate over successors use ``succ_iterator/succ_begin/succ_end``.
1759 Making simple changes
1760 ---------------------
1762 There are some primitive transformation operations present in the LLVM
1763 infrastructure that are worth knowing about. When performing transformations,
1764 it's fairly common to manipulate the contents of basic blocks. This section
1765 describes some of the common methods for doing so and gives example code.
1767 .. _schanges_creating:
1769 Creating and inserting new ``Instruction``\ s
1770 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1772 *Instantiating Instructions*
1774 Creation of ``Instruction``\ s is straight-forward: simply call the constructor
1775 for the kind of instruction to instantiate and provide the necessary parameters.
1776 For example, an ``AllocaInst`` only *requires* a (const-ptr-to) ``Type``. Thus:
1780 AllocaInst* ai = new AllocaInst(Type::Int32Ty);
1782 will create an ``AllocaInst`` instance that represents the allocation of one
1783 integer in the current stack frame, at run time. Each ``Instruction`` subclass
1784 is likely to have varying default parameters which change the semantics of the
1785 instruction, so refer to the `doxygen documentation for the subclass of
1786 Instruction <http://llvm.org/doxygen/classllvm_1_1Instruction.html>`_ that
1787 you're interested in instantiating.
1791 It is very useful to name the values of instructions when you're able to, as
1792 this facilitates the debugging of your transformations. If you end up looking
1793 at generated LLVM machine code, you definitely want to have logical names
1794 associated with the results of instructions! By supplying a value for the
1795 ``Name`` (default) parameter of the ``Instruction`` constructor, you associate a
1796 logical name with the result of the instruction's execution at run time. For
1797 example, say that I'm writing a transformation that dynamically allocates space
1798 for an integer on the stack, and that integer is going to be used as some kind
1799 of index by some other code. To accomplish this, I place an ``AllocaInst`` at
1800 the first point in the first ``BasicBlock`` of some ``Function``, and I'm
1801 intending to use it within the same ``Function``. I might do:
1805 AllocaInst* pa = new AllocaInst(Type::Int32Ty, 0, "indexLoc");
1807 where ``indexLoc`` is now the logical name of the instruction's execution value,
1808 which is a pointer to an integer on the run time stack.
1810 *Inserting instructions*
1812 There are essentially two ways to insert an ``Instruction`` into an existing
1813 sequence of instructions that form a ``BasicBlock``:
1815 * Insertion into an explicit instruction list
1817 Given a ``BasicBlock* pb``, an ``Instruction* pi`` within that ``BasicBlock``,
1818 and a newly-created instruction we wish to insert before ``*pi``, we do the
1823 BasicBlock *pb = ...;
1824 Instruction *pi = ...;
1825 Instruction *newInst = new Instruction(...);
1827 pb->getInstList().insert(pi, newInst); // Inserts newInst before pi in pb
1829 Appending to the end of a ``BasicBlock`` is so common that the ``Instruction``
1830 class and ``Instruction``-derived classes provide constructors which take a
1831 pointer to a ``BasicBlock`` to be appended to. For example code that looked
1836 BasicBlock *pb = ...;
1837 Instruction *newInst = new Instruction(...);
1839 pb->getInstList().push_back(newInst); // Appends newInst to pb
1845 BasicBlock *pb = ...;
1846 Instruction *newInst = new Instruction(..., pb);
1848 which is much cleaner, especially if you are creating long instruction
1851 * Insertion into an implicit instruction list
1853 ``Instruction`` instances that are already in ``BasicBlock``\ s are implicitly
1854 associated with an existing instruction list: the instruction list of the
1855 enclosing basic block. Thus, we could have accomplished the same thing as the
1856 above code without being given a ``BasicBlock`` by doing:
1860 Instruction *pi = ...;
1861 Instruction *newInst = new Instruction(...);
1863 pi->getParent()->getInstList().insert(pi, newInst);
1865 In fact, this sequence of steps occurs so frequently that the ``Instruction``
1866 class and ``Instruction``-derived classes provide constructors which take (as
1867 a default parameter) a pointer to an ``Instruction`` which the newly-created
1868 ``Instruction`` should precede. That is, ``Instruction`` constructors are
1869 capable of inserting the newly-created instance into the ``BasicBlock`` of a
1870 provided instruction, immediately before that instruction. Using an
1871 ``Instruction`` constructor with a ``insertBefore`` (default) parameter, the
1876 Instruction* pi = ...;
1877 Instruction* newInst = new Instruction(..., pi);
1879 which is much cleaner, especially if you're creating a lot of instructions and
1880 adding them to ``BasicBlock``\ s.
1882 .. _schanges_deleting:
1884 Deleting Instructions
1885 ^^^^^^^^^^^^^^^^^^^^^
1887 Deleting an instruction from an existing sequence of instructions that form a
1888 BasicBlock_ is very straight-forward: just call the instruction's
1889 ``eraseFromParent()`` method. For example:
1893 Instruction *I = .. ;
1894 I->eraseFromParent();
1896 This unlinks the instruction from its containing basic block and deletes it. If
1897 you'd just like to unlink the instruction from its containing basic block but
1898 not delete it, you can use the ``removeFromParent()`` method.
1900 .. _schanges_replacing:
1902 Replacing an Instruction with another Value
1903 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1905 Replacing individual instructions
1906 """""""""""""""""""""""""""""""""
1908 Including "`llvm/Transforms/Utils/BasicBlockUtils.h
1909 <http://llvm.org/doxygen/BasicBlockUtils_8h-source.html>`_" permits use of two
1910 very useful replace functions: ``ReplaceInstWithValue`` and
1911 ``ReplaceInstWithInst``.
1913 .. _schanges_deleting_sub:
1915 Deleting Instructions
1916 """""""""""""""""""""
1918 * ``ReplaceInstWithValue``
1920 This function replaces all uses of a given instruction with a value, and then
1921 removes the original instruction. The following example illustrates the
1922 replacement of the result of a particular ``AllocaInst`` that allocates memory
1923 for a single integer with a null pointer to an integer.
1927 AllocaInst* instToReplace = ...;
1928 BasicBlock::iterator ii(instToReplace);
1930 ReplaceInstWithValue(instToReplace->getParent()->getInstList(), ii,
1931 Constant::getNullValue(PointerType::getUnqual(Type::Int32Ty)));
1933 * ``ReplaceInstWithInst``
1935 This function replaces a particular instruction with another instruction,
1936 inserting the new instruction into the basic block at the location where the
1937 old instruction was, and replacing any uses of the old instruction with the
1938 new instruction. The following example illustrates the replacement of one
1939 ``AllocaInst`` with another.
1943 AllocaInst* instToReplace = ...;
1944 BasicBlock::iterator ii(instToReplace);
1946 ReplaceInstWithInst(instToReplace->getParent()->getInstList(), ii,
1947 new AllocaInst(Type::Int32Ty, 0, "ptrToReplacedInt"));
1950 Replacing multiple uses of Users and Values
1951 """""""""""""""""""""""""""""""""""""""""""
1953 You can use ``Value::replaceAllUsesWith`` and ``User::replaceUsesOfWith`` to
1954 change more than one use at a time. See the doxygen documentation for the
1955 `Value Class <http://llvm.org/doxygen/classllvm_1_1Value.html>`_ and `User Class
1956 <http://llvm.org/doxygen/classllvm_1_1User.html>`_, respectively, for more
1959 .. _schanges_deletingGV:
1961 Deleting GlobalVariables
1962 ^^^^^^^^^^^^^^^^^^^^^^^^
1964 Deleting a global variable from a module is just as easy as deleting an
1965 Instruction. First, you must have a pointer to the global variable that you
1966 wish to delete. You use this pointer to erase it from its parent, the module.
1971 GlobalVariable *GV = .. ;
1973 GV->eraseFromParent();
1981 In generating IR, you may need some complex types. If you know these types
1982 statically, you can use ``TypeBuilder<...>::get()``, defined in
1983 ``llvm/Support/TypeBuilder.h``, to retrieve them. ``TypeBuilder`` has two forms
1984 depending on whether you're building types for cross-compilation or native
1985 library use. ``TypeBuilder<T, true>`` requires that ``T`` be independent of the
1986 host environment, meaning that it's built out of types from the ``llvm::types``
1987 (`doxygen <http://llvm.org/doxygen/namespacellvm_1_1types.html>`__) namespace
1988 and pointers, functions, arrays, etc. built of those. ``TypeBuilder<T, false>``
1989 additionally allows native C types whose size may depend on the host compiler.
1994 FunctionType *ft = TypeBuilder<types::i<8>(types::i<32>*), true>::get();
1996 is easier to read and write than the equivalent
2000 std::vector<const Type*> params;
2001 params.push_back(PointerType::getUnqual(Type::Int32Ty));
2002 FunctionType *ft = FunctionType::get(Type::Int8Ty, params, false);
2004 See the `class comment
2005 <http://llvm.org/doxygen/TypeBuilder_8h-source.html#l00001>`_ for more details.
2012 This section describes the interaction of the LLVM APIs with multithreading,
2013 both on the part of client applications, and in the JIT, in the hosted
2016 Note that LLVM's support for multithreading is still relatively young. Up
2017 through version 2.5, the execution of threaded hosted applications was
2018 supported, but not threaded client access to the APIs. While this use case is
2019 now supported, clients *must* adhere to the guidelines specified below to ensure
2020 proper operation in multithreaded mode.
2022 Note that, on Unix-like platforms, LLVM requires the presence of GCC's atomic
2023 intrinsics in order to support threaded operation. If you need a
2024 multhreading-capable LLVM on a platform without a suitably modern system
2025 compiler, consider compiling LLVM and LLVM-GCC in single-threaded mode, and
2026 using the resultant compiler to build a copy of LLVM with multithreading
2029 .. _startmultithreaded:
2031 Entering and Exiting Multithreaded Mode
2032 ---------------------------------------
2034 In order to properly protect its internal data structures while avoiding
2035 excessive locking overhead in the single-threaded case, the LLVM must intialize
2036 certain data structures necessary to provide guards around its internals. To do
2037 so, the client program must invoke ``llvm_start_multithreaded()`` before making
2038 any concurrent LLVM API calls. To subsequently tear down these structures, use
2039 the ``llvm_stop_multithreaded()`` call. You can also use the
2040 ``llvm_is_multithreaded()`` call to check the status of multithreaded mode.
2042 Note that both of these calls must be made *in isolation*. That is to say that
2043 no other LLVM API calls may be executing at any time during the execution of
2044 ``llvm_start_multithreaded()`` or ``llvm_stop_multithreaded``. It's is the
2045 client's responsibility to enforce this isolation.
2047 The return value of ``llvm_start_multithreaded()`` indicates the success or
2048 failure of the initialization. Failure typically indicates that your copy of
2049 LLVM was built without multithreading support, typically because GCC atomic
2050 intrinsics were not found in your system compiler. In this case, the LLVM API
2051 will not be safe for concurrent calls. However, it *will* be safe for hosting
2052 threaded applications in the JIT, though :ref:`care must be taken
2053 <jitthreading>` to ensure that side exits and the like do not accidentally
2054 result in concurrent LLVM API calls.
2058 Ending Execution with ``llvm_shutdown()``
2059 -----------------------------------------
2061 When you are done using the LLVM APIs, you should call ``llvm_shutdown()`` to
2062 deallocate memory used for internal structures. This will also invoke
2063 ``llvm_stop_multithreaded()`` if LLVM is operating in multithreaded mode. As
2064 such, ``llvm_shutdown()`` requires the same isolation guarantees as
2065 ``llvm_stop_multithreaded()``.
2067 Note that, if you use scope-based shutdown, you can use the
2068 ``llvm_shutdown_obj`` class, which calls ``llvm_shutdown()`` in its destructor.
2072 Lazy Initialization with ``ManagedStatic``
2073 ------------------------------------------
2075 ``ManagedStatic`` is a utility class in LLVM used to implement static
2076 initialization of static resources, such as the global type tables. Before the
2077 invocation of ``llvm_shutdown()``, it implements a simple lazy initialization
2078 scheme. Once ``llvm_start_multithreaded()`` returns, however, it uses
2079 double-checked locking to implement thread-safe lazy initialization.
2081 Note that, because no other threads are allowed to issue LLVM API calls before
2082 ``llvm_start_multithreaded()`` returns, it is possible to have
2083 ``ManagedStatic``\ s of ``llvm::sys::Mutex``\ s.
2085 The ``llvm_acquire_global_lock()`` and ``llvm_release_global_lock`` APIs provide
2086 access to the global lock used to implement the double-checked locking for lazy
2087 initialization. These should only be used internally to LLVM, and only if you
2088 know what you're doing!
2092 Achieving Isolation with ``LLVMContext``
2093 ----------------------------------------
2095 ``LLVMContext`` is an opaque class in the LLVM API which clients can use to
2096 operate multiple, isolated instances of LLVM concurrently within the same
2097 address space. For instance, in a hypothetical compile-server, the compilation
2098 of an individual translation unit is conceptually independent from all the
2099 others, and it would be desirable to be able to compile incoming translation
2100 units concurrently on independent server threads. Fortunately, ``LLVMContext``
2101 exists to enable just this kind of scenario!
2103 Conceptually, ``LLVMContext`` provides isolation. Every LLVM entity
2104 (``Module``\ s, ``Value``\ s, ``Type``\ s, ``Constant``\ s, etc.) in LLVM's
2105 in-memory IR belongs to an ``LLVMContext``. Entities in different contexts
2106 *cannot* interact with each other: ``Module``\ s in different contexts cannot be
2107 linked together, ``Function``\ s cannot be added to ``Module``\ s in different
2108 contexts, etc. What this means is that is is safe to compile on multiple
2109 threads simultaneously, as long as no two threads operate on entities within the
2112 In practice, very few places in the API require the explicit specification of a
2113 ``LLVMContext``, other than the ``Type`` creation/lookup APIs. Because every
2114 ``Type`` carries a reference to its owning context, most other entities can
2115 determine what context they belong to by looking at their own ``Type``. If you
2116 are adding new entities to LLVM IR, please try to maintain this interface
2119 For clients that do *not* require the benefits of isolation, LLVM provides a
2120 convenience API ``getGlobalContext()``. This returns a global, lazily
2121 initialized ``LLVMContext`` that may be used in situations where isolation is
2129 LLVM's "eager" JIT compiler is safe to use in threaded programs. Multiple
2130 threads can call ``ExecutionEngine::getPointerToFunction()`` or
2131 ``ExecutionEngine::runFunction()`` concurrently, and multiple threads can run
2132 code output by the JIT concurrently. The user must still ensure that only one
2133 thread accesses IR in a given ``LLVMContext`` while another thread might be
2134 modifying it. One way to do that is to always hold the JIT lock while accessing
2135 IR outside the JIT (the JIT *modifies* the IR by adding ``CallbackVH``\ s).
2136 Another way is to only call ``getPointerToFunction()`` from the
2137 ``LLVMContext``'s thread.
2139 When the JIT is configured to compile lazily (using
2140 ``ExecutionEngine::DisableLazyCompilation(false)``), there is currently a `race
2141 condition <http://llvm.org/bugs/show_bug.cgi?id=5184>`_ in updating call sites
2142 after a function is lazily-jitted. It's still possible to use the lazy JIT in a
2143 threaded program if you ensure that only one thread at a time can call any
2144 particular lazy stub and that the JIT lock guards any IR access, but we suggest
2145 using only the eager JIT in threaded programs.
2152 This section describes some of the advanced or obscure API's that most clients
2153 do not need to be aware of. These API's tend manage the inner workings of the
2154 LLVM system, and only need to be accessed in unusual circumstances.
2158 The ``ValueSymbolTable`` class
2159 ------------------------------
2161 The ``ValueSymbolTable`` (`doxygen
2162 <http://llvm.org/doxygen/classllvm_1_1ValueSymbolTable.html>`__) class provides
2163 a symbol table that the :ref:`Function <c_Function>` and Module_ classes use for
2164 naming value definitions. The symbol table can provide a name for any Value_.
2166 Note that the ``SymbolTable`` class should not be directly accessed by most
2167 clients. It should only be used when iteration over the symbol table names
2168 themselves are required, which is very special purpose. Note that not all LLVM
2169 Value_\ s have names, and those without names (i.e. they have an empty name) do
2170 not exist in the symbol table.
2172 Symbol tables support iteration over the values in the symbol table with
2173 ``begin/end/iterator`` and supports querying to see if a specific name is in the
2174 symbol table (with ``lookup``). The ``ValueSymbolTable`` class exposes no
2175 public mutator methods, instead, simply call ``setName`` on a value, which will
2176 autoinsert it into the appropriate symbol table.
2180 The ``User`` and owned ``Use`` classes' memory layout
2181 -----------------------------------------------------
2183 The ``User`` (`doxygen <http://llvm.org/doxygen/classllvm_1_1User.html>`__)
2184 class provides a basis for expressing the ownership of ``User`` towards other
2185 `Value instance <http://llvm.org/doxygen/classllvm_1_1Value.html>`_\ s. The
2186 ``Use`` (`doxygen <http://llvm.org/doxygen/classllvm_1_1Use.html>`__) helper
2187 class is employed to do the bookkeeping and to facilitate *O(1)* addition and
2192 Interaction and relationship between ``User`` and ``Use`` objects
2193 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2195 A subclass of ``User`` can choose between incorporating its ``Use`` objects or
2196 refer to them out-of-line by means of a pointer. A mixed variant (some ``Use``
2197 s inline others hung off) is impractical and breaks the invariant that the
2198 ``Use`` objects belonging to the same ``User`` form a contiguous array.
2200 We have 2 different layouts in the ``User`` (sub)classes:
2204 The ``Use`` object(s) are inside (resp. at fixed offset) of the ``User``
2205 object and there are a fixed number of them.
2209 The ``Use`` object(s) are referenced by a pointer to an array from the
2210 ``User`` object and there may be a variable number of them.
2212 As of v2.4 each layout still possesses a direct pointer to the start of the
2213 array of ``Use``\ s. Though not mandatory for layout a), we stick to this
2214 redundancy for the sake of simplicity. The ``User`` object also stores the
2215 number of ``Use`` objects it has. (Theoretically this information can also be
2216 calculated given the scheme presented below.)
2218 Special forms of allocation operators (``operator new``) enforce the following
2221 * Layout a) is modelled by prepending the ``User`` object by the ``Use[]``
2224 .. code-block:: none
2226 ...---.---.---.---.-------...
2227 | P | P | P | P | User
2228 '''---'---'---'---'-------'''
2230 * Layout b) is modelled by pointing at the ``Use[]`` array.
2232 .. code-block:: none
2243 *(In the above figures* '``P``' *stands for the* ``Use**`` *that is stored in
2244 each* ``Use`` *object in the member* ``Use::Prev`` *)*
2248 The waymarking algorithm
2249 ^^^^^^^^^^^^^^^^^^^^^^^^
2251 Since the ``Use`` objects are deprived of the direct (back)pointer to their
2252 ``User`` objects, there must be a fast and exact method to recover it. This is
2253 accomplished by the following scheme:
2255 A bit-encoding in the 2 LSBits (least significant bits) of the ``Use::Prev``
2256 allows to find the start of the ``User`` object:
2258 * ``00`` –> binary digit 0
2260 * ``01`` –> binary digit 1
2262 * ``10`` –> stop and calculate (``s``)
2264 * ``11`` –> full stop (``S``)
2266 Given a ``Use*``, all we have to do is to walk till we get a stop and we either
2267 have a ``User`` immediately behind or we have to walk to the next stop picking
2268 up digits and calculating the offset:
2270 .. code-block:: none
2272 .---.---.---.---.---.---.---.---.---.---.---.---.---.---.---.---.----------------
2273 | 1 | s | 1 | 0 | 1 | 0 | s | 1 | 1 | 0 | s | 1 | 1 | s | 1 | S | User (or User*)
2274 '---'---'---'---'---'---'---'---'---'---'---'---'---'---'---'---'----------------
2275 |+15 |+10 |+6 |+3 |+1
2278 | | | ______________________>
2279 | | ______________________________________>
2280 | __________________________________________________________>
2282 Only the significant number of bits need to be stored between the stops, so that
2283 the *worst case is 20 memory accesses* when there are 1000 ``Use`` objects
2284 associated with a ``User``.
2288 Reference implementation
2289 ^^^^^^^^^^^^^^^^^^^^^^^^
2291 The following literate Haskell fragment demonstrates the concept:
2293 .. code-block:: haskell
2295 > import Test.QuickCheck
2297 > digits :: Int -> [Char] -> [Char]
2298 > digits 0 acc = '0' : acc
2299 > digits 1 acc = '1' : acc
2300 > digits n acc = digits (n `div` 2) $ digits (n `mod` 2) acc
2302 > dist :: Int -> [Char] -> [Char]
2305 > dist 1 acc = let r = dist 0 acc in 's' : digits (length r) r
2306 > dist n acc = dist (n - 1) $ dist 1 acc
2308 > takeLast n ss = reverse $ take n $ reverse ss
2310 > test = takeLast 40 $ dist 20 []
2313 Printing <test> gives: ``"1s100000s11010s10100s1111s1010s110s11s1S"``
2315 The reverse algorithm computes the length of the string just by examining a
2318 .. code-block:: haskell
2320 > pref :: [Char] -> Int
2322 > pref ('s':'1':rest) = decode 2 1 rest
2323 > pref (_:rest) = 1 + pref rest
2325 > decode walk acc ('0':rest) = decode (walk + 1) (acc * 2) rest
2326 > decode walk acc ('1':rest) = decode (walk + 1) (acc * 2 + 1) rest
2327 > decode walk acc _ = walk + acc
2330 Now, as expected, printing <pref test> gives ``40``.
2332 We can *quickCheck* this with following property:
2334 .. code-block:: haskell
2336 > testcase = dist 2000 []
2337 > testcaseLength = length testcase
2339 > identityProp n = n > 0 && n <= testcaseLength ==> length arr == pref arr
2340 > where arr = takeLast n testcase
2343 As expected <quickCheck identityProp> gives:
2347 *Main> quickCheck identityProp
2348 OK, passed 100 tests.
2350 Let's be a bit more exhaustive:
2352 .. code-block:: haskell
2355 > deepCheck p = check (defaultConfig { configMaxTest = 500 }) p
2358 And here is the result of <deepCheck identityProp>:
2362 *Main> deepCheck identityProp
2363 OK, passed 500 tests.
2367 Tagging considerations
2368 ^^^^^^^^^^^^^^^^^^^^^^
2370 To maintain the invariant that the 2 LSBits of each ``Use**`` in ``Use`` never
2371 change after being set up, setters of ``Use::Prev`` must re-tag the new
2372 ``Use**`` on every modification. Accordingly getters must strip the tag bits.
2374 For layout b) instead of the ``User`` we find a pointer (``User*`` with LSBit
2375 set). Following this pointer brings us to the ``User``. A portable trick
2376 ensures that the first bytes of ``User`` (if interpreted as a pointer) never has
2377 the LSBit set. (Portability is relying on the fact that all known compilers
2378 place the ``vptr`` in the first word of the instances.)
2382 The Core LLVM Class Hierarchy Reference
2383 =======================================
2385 ``#include "llvm/Type.h"``
2387 header source: `Type.h <http://llvm.org/doxygen/Type_8h-source.html>`_
2389 doxygen info: `Type Clases <http://llvm.org/doxygen/classllvm_1_1Type.html>`_
2391 The Core LLVM classes are the primary means of representing the program being
2392 inspected or transformed. The core LLVM classes are defined in header files in
2393 the ``include/llvm/`` directory, and implemented in the ``lib/VMCore``
2398 The Type class and Derived Types
2399 --------------------------------
2401 ``Type`` is a superclass of all type classes. Every ``Value`` has a ``Type``.
2402 ``Type`` cannot be instantiated directly but only through its subclasses.
2403 Certain primitive types (``VoidType``, ``LabelType``, ``FloatType`` and
2404 ``DoubleType``) have hidden subclasses. They are hidden because they offer no
2405 useful functionality beyond what the ``Type`` class offers except to distinguish
2406 themselves from other subclasses of ``Type``.
2408 All other types are subclasses of ``DerivedType``. Types can be named, but this
2409 is not a requirement. There exists exactly one instance of a given shape at any
2410 one time. This allows type equality to be performed with address equality of
2411 the Type Instance. That is, given two ``Type*`` values, the types are identical
2412 if the pointers are identical.
2416 Important Public Methods
2417 ^^^^^^^^^^^^^^^^^^^^^^^^
2419 * ``bool isIntegerTy() const``: Returns true for any integer type.
2421 * ``bool isFloatingPointTy()``: Return true if this is one of the five
2422 floating point types.
2424 * ``bool isSized()``: Return true if the type has known size. Things
2425 that don't have a size are abstract types, labels and void.
2429 Important Derived Types
2430 ^^^^^^^^^^^^^^^^^^^^^^^
2433 Subclass of DerivedType that represents integer types of any bit width. Any
2434 bit width between ``IntegerType::MIN_INT_BITS`` (1) and
2435 ``IntegerType::MAX_INT_BITS`` (~8 million) can be represented.
2437 * ``static const IntegerType* get(unsigned NumBits)``: get an integer
2438 type of a specific bit width.
2440 * ``unsigned getBitWidth() const``: Get the bit width of an integer type.
2443 This is subclassed by ArrayType, PointerType and VectorType.
2445 * ``const Type * getElementType() const``: Returns the type of each
2446 of the elements in the sequential type.
2449 This is a subclass of SequentialType and defines the interface for array
2452 * ``unsigned getNumElements() const``: Returns the number of elements
2456 Subclass of SequentialType for pointer types.
2459 Subclass of SequentialType for vector types. A vector type is similar to an
2460 ArrayType but is distinguished because it is a first class type whereas
2461 ArrayType is not. Vector types are used for vector operations and are usually
2462 small vectors of of an integer or floating point type.
2465 Subclass of DerivedTypes for struct types.
2470 Subclass of DerivedTypes for function types.
2472 * ``bool isVarArg() const``: Returns true if it's a vararg function.
2474 * ``const Type * getReturnType() const``: Returns the return type of the
2477 * ``const Type * getParamType (unsigned i)``: Returns the type of the ith
2480 * ``const unsigned getNumParams() const``: Returns the number of formal
2485 The ``Module`` class
2486 --------------------
2488 ``#include "llvm/Module.h"``
2490 header source: `Module.h <http://llvm.org/doxygen/Module_8h-source.html>`_
2492 doxygen info: `Module Class <http://llvm.org/doxygen/classllvm_1_1Module.html>`_
2494 The ``Module`` class represents the top level structure present in LLVM
2495 programs. An LLVM module is effectively either a translation unit of the
2496 original program or a combination of several translation units merged by the
2497 linker. The ``Module`` class keeps track of a list of :ref:`Function
2498 <c_Function>`\ s, a list of GlobalVariable_\ s, and a SymbolTable_.
2499 Additionally, it contains a few helpful member functions that try to make common
2504 Important Public Members of the ``Module`` class
2505 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2507 * ``Module::Module(std::string name = "")``
2509 Constructing a Module_ is easy. You can optionally provide a name for it
2510 (probably based on the name of the translation unit).
2512 * | ``Module::iterator`` - Typedef for function list iterator
2513 | ``Module::const_iterator`` - Typedef for const_iterator.
2514 | ``begin()``, ``end()``, ``size()``, ``empty()``
2516 These are forwarding methods that make it easy to access the contents of a
2517 ``Module`` object's :ref:`Function <c_Function>` list.
2519 * ``Module::FunctionListType &getFunctionList()``
2521 Returns the list of :ref:`Function <c_Function>`\ s. This is necessary to use
2522 when you need to update the list or perform a complex action that doesn't have
2523 a forwarding method.
2527 * | ``Module::global_iterator`` - Typedef for global variable list iterator
2528 | ``Module::const_global_iterator`` - Typedef for const_iterator.
2529 | ``global_begin()``, ``global_end()``, ``global_size()``, ``global_empty()``
2531 These are forwarding methods that make it easy to access the contents of a
2532 ``Module`` object's GlobalVariable_ list.
2534 * ``Module::GlobalListType &getGlobalList()``
2536 Returns the list of GlobalVariable_\ s. This is necessary to use when you
2537 need to update the list or perform a complex action that doesn't have a
2542 * ``SymbolTable *getSymbolTable()``
2544 Return a reference to the SymbolTable_ for this ``Module``.
2548 * ``Function *getFunction(StringRef Name) const``
2550 Look up the specified function in the ``Module`` SymbolTable_. If it does not
2551 exist, return ``null``.
2553 * ``Function *getOrInsertFunction(const std::string &Name, const FunctionType
2556 Look up the specified function in the ``Module`` SymbolTable_. If it does not
2557 exist, add an external declaration for the function and return it.
2559 * ``std::string getTypeName(const Type *Ty)``
2561 If there is at least one entry in the SymbolTable_ for the specified Type_,
2562 return it. Otherwise return the empty string.
2564 * ``bool addTypeName(const std::string &Name, const Type *Ty)``
2566 Insert an entry in the SymbolTable_ mapping ``Name`` to ``Ty``. If there is
2567 already an entry for this name, true is returned and the SymbolTable_ is not
2575 ``#include "llvm/Value.h"``
2577 header source: `Value.h <http://llvm.org/doxygen/Value_8h-source.html>`_
2579 doxygen info: `Value Class <http://llvm.org/doxygen/classllvm_1_1Value.html>`_
2581 The ``Value`` class is the most important class in the LLVM Source base. It
2582 represents a typed value that may be used (among other things) as an operand to
2583 an instruction. There are many different types of ``Value``\ s, such as
2584 Constant_\ s, Argument_\ s. Even Instruction_\ s and :ref:`Function
2585 <c_Function>`\ s are ``Value``\ s.
2587 A particular ``Value`` may be used many times in the LLVM representation for a
2588 program. For example, an incoming argument to a function (represented with an
2589 instance of the Argument_ class) is "used" by every instruction in the function
2590 that references the argument. To keep track of this relationship, the ``Value``
2591 class keeps a list of all of the ``User``\ s that is using it (the User_ class
2592 is a base class for all nodes in the LLVM graph that can refer to ``Value``\ s).
2593 This use list is how LLVM represents def-use information in the program, and is
2594 accessible through the ``use_*`` methods, shown below.
2596 Because LLVM is a typed representation, every LLVM ``Value`` is typed, and this
2597 Type_ is available through the ``getType()`` method. In addition, all LLVM
2598 values can be named. The "name" of the ``Value`` is a symbolic string printed
2601 .. code-block:: llvm
2607 The name of this instruction is "foo". **NOTE** that the name of any value may
2608 be missing (an empty string), so names should **ONLY** be used for debugging
2609 (making the source code easier to read, debugging printouts), they should not be
2610 used to keep track of values or map between them. For this purpose, use a
2611 ``std::map`` of pointers to the ``Value`` itself instead.
2613 One important aspect of LLVM is that there is no distinction between an SSA
2614 variable and the operation that produces it. Because of this, any reference to
2615 the value produced by an instruction (or the value available as an incoming
2616 argument, for example) is represented as a direct pointer to the instance of the
2617 class that represents this value. Although this may take some getting used to,
2618 it simplifies the representation and makes it easier to manipulate.
2622 Important Public Members of the ``Value`` class
2623 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2625 * | ``Value::use_iterator`` - Typedef for iterator over the use-list
2626 | ``Value::const_use_iterator`` - Typedef for const_iterator over the
2628 | ``unsigned use_size()`` - Returns the number of users of the value.
2629 | ``bool use_empty()`` - Returns true if there are no users.
2630 | ``use_iterator use_begin()`` - Get an iterator to the start of the
2632 | ``use_iterator use_end()`` - Get an iterator to the end of the use-list.
2633 | ``User *use_back()`` - Returns the last element in the list.
2635 These methods are the interface to access the def-use information in LLVM.
2636 As with all other iterators in LLVM, the naming conventions follow the
2637 conventions defined by the STL_.
2639 * ``Type *getType() const``
2640 This method returns the Type of the Value.
2642 * | ``bool hasName() const``
2643 | ``std::string getName() const``
2644 | ``void setName(const std::string &Name)``
2646 This family of methods is used to access and assign a name to a ``Value``, be
2647 aware of the :ref:`precaution above <nameWarning>`.
2649 * ``void replaceAllUsesWith(Value *V)``
2651 This method traverses the use list of a ``Value`` changing all User_\ s of the
2652 current value to refer to "``V``" instead. For example, if you detect that an
2653 instruction always produces a constant value (for example through constant
2654 folding), you can replace all uses of the instruction with the constant like
2659 Inst->replaceAllUsesWith(ConstVal);
2666 ``#include "llvm/User.h"``
2668 header source: `User.h <http://llvm.org/doxygen/User_8h-source.html>`_
2670 doxygen info: `User Class <http://llvm.org/doxygen/classllvm_1_1User.html>`_
2674 The ``User`` class is the common base class of all LLVM nodes that may refer to
2675 ``Value``\ s. It exposes a list of "Operands" that are all of the ``Value``\ s
2676 that the User is referring to. The ``User`` class itself is a subclass of
2679 The operands of a ``User`` point directly to the LLVM ``Value`` that it refers
2680 to. Because LLVM uses Static Single Assignment (SSA) form, there can only be
2681 one definition referred to, allowing this direct connection. This connection
2682 provides the use-def information in LLVM.
2686 Important Public Members of the ``User`` class
2687 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2689 The ``User`` class exposes the operand list in two ways: through an index access
2690 interface and through an iterator based interface.
2692 * | ``Value *getOperand(unsigned i)``
2693 | ``unsigned getNumOperands()``
2695 These two methods expose the operands of the ``User`` in a convenient form for
2698 * | ``User::op_iterator`` - Typedef for iterator over the operand list
2699 | ``op_iterator op_begin()`` - Get an iterator to the start of the operand
2701 | ``op_iterator op_end()`` - Get an iterator to the end of the operand list.
2703 Together, these methods make up the iterator based interface to the operands
2709 The ``Instruction`` class
2710 -------------------------
2712 ``#include "llvm/Instruction.h"``
2714 header source: `Instruction.h
2715 <http://llvm.org/doxygen/Instruction_8h-source.html>`_
2717 doxygen info: `Instruction Class
2718 <http://llvm.org/doxygen/classllvm_1_1Instruction.html>`_
2720 Superclasses: User_, Value_
2722 The ``Instruction`` class is the common base class for all LLVM instructions.
2723 It provides only a few methods, but is a very commonly used class. The primary
2724 data tracked by the ``Instruction`` class itself is the opcode (instruction
2725 type) and the parent BasicBlock_ the ``Instruction`` is embedded into. To
2726 represent a specific type of instruction, one of many subclasses of
2727 ``Instruction`` are used.
2729 Because the ``Instruction`` class subclasses the User_ class, its operands can
2730 be accessed in the same way as for other ``User``\ s (with the
2731 ``getOperand()``/``getNumOperands()`` and ``op_begin()``/``op_end()`` methods).
2732 An important file for the ``Instruction`` class is the ``llvm/Instruction.def``
2733 file. This file contains some meta-data about the various different types of
2734 instructions in LLVM. It describes the enum values that are used as opcodes
2735 (for example ``Instruction::Add`` and ``Instruction::ICmp``), as well as the
2736 concrete sub-classes of ``Instruction`` that implement the instruction (for
2737 example BinaryOperator_ and CmpInst_). Unfortunately, the use of macros in this
2738 file confuses doxygen, so these enum values don't show up correctly in the
2739 `doxygen output <http://llvm.org/doxygen/classllvm_1_1Instruction.html>`_.
2743 Important Subclasses of the ``Instruction`` class
2744 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2748 * ``BinaryOperator``
2750 This subclasses represents all two operand instructions whose operands must be
2751 the same type, except for the comparison instructions.
2756 This subclass is the parent of the 12 casting instructions. It provides
2757 common operations on cast instructions.
2763 This subclass respresents the two comparison instructions,
2764 `ICmpInst <LangRef.html#i_icmp>`_ (integer opreands), and
2765 `FCmpInst <LangRef.html#i_fcmp>`_ (floating point operands).
2769 * ``TerminatorInst``
2771 This subclass is the parent of all terminator instructions (those which can
2776 Important Public Members of the ``Instruction`` class
2777 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2779 * ``BasicBlock *getParent()``
2781 Returns the BasicBlock_ that this
2782 ``Instruction`` is embedded into.
2784 * ``bool mayWriteToMemory()``
2786 Returns true if the instruction writes to memory, i.e. it is a ``call``,
2787 ``free``, ``invoke``, or ``store``.
2789 * ``unsigned getOpcode()``
2791 Returns the opcode for the ``Instruction``.
2793 * ``Instruction *clone() const``
2795 Returns another instance of the specified instruction, identical in all ways
2796 to the original except that the instruction has no parent (i.e. it's not
2797 embedded into a BasicBlock_), and it has no name.
2801 The ``Constant`` class and subclasses
2802 -------------------------------------
2804 Constant represents a base class for different types of constants. It is
2805 subclassed by ConstantInt, ConstantArray, etc. for representing the various
2806 types of Constants. GlobalValue_ is also a subclass, which represents the
2807 address of a global variable or function.
2811 Important Subclasses of Constant
2812 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2814 * ConstantInt : This subclass of Constant represents an integer constant of
2817 * ``const APInt& getValue() const``: Returns the underlying
2818 value of this constant, an APInt value.
2820 * ``int64_t getSExtValue() const``: Converts the underlying APInt value to an
2821 int64_t via sign extension. If the value (not the bit width) of the APInt
2822 is too large to fit in an int64_t, an assertion will result. For this
2823 reason, use of this method is discouraged.
2825 * ``uint64_t getZExtValue() const``: Converts the underlying APInt value
2826 to a uint64_t via zero extension. IF the value (not the bit width) of the
2827 APInt is too large to fit in a uint64_t, an assertion will result. For this
2828 reason, use of this method is discouraged.
2830 * ``static ConstantInt* get(const APInt& Val)``: Returns the ConstantInt
2831 object that represents the value provided by ``Val``. The type is implied
2832 as the IntegerType that corresponds to the bit width of ``Val``.
2834 * ``static ConstantInt* get(const Type *Ty, uint64_t Val)``: Returns the
2835 ConstantInt object that represents the value provided by ``Val`` for integer
2838 * ConstantFP : This class represents a floating point constant.
2840 * ``double getValue() const``: Returns the underlying value of this constant.
2842 * ConstantArray : This represents a constant array.
2844 * ``const std::vector<Use> &getValues() const``: Returns a vector of
2845 component constants that makeup this array.
2847 * ConstantStruct : This represents a constant struct.
2849 * ``const std::vector<Use> &getValues() const``: Returns a vector of
2850 component constants that makeup this array.
2852 * GlobalValue : This represents either a global variable or a function. In
2853 either case, the value is a constant fixed address (after linking).
2857 The ``GlobalValue`` class
2858 -------------------------
2860 ``#include "llvm/GlobalValue.h"``
2862 header source: `GlobalValue.h
2863 <http://llvm.org/doxygen/GlobalValue_8h-source.html>`_
2865 doxygen info: `GlobalValue Class
2866 <http://llvm.org/doxygen/classllvm_1_1GlobalValue.html>`_
2868 Superclasses: Constant_, User_, Value_
2870 Global values ( GlobalVariable_\ s or :ref:`Function <c_Function>`\ s) are the
2871 only LLVM values that are visible in the bodies of all :ref:`Function
2872 <c_Function>`\ s. Because they are visible at global scope, they are also
2873 subject to linking with other globals defined in different translation units.
2874 To control the linking process, ``GlobalValue``\ s know their linkage rules.
2875 Specifically, ``GlobalValue``\ s know whether they have internal or external
2876 linkage, as defined by the ``LinkageTypes`` enumeration.
2878 If a ``GlobalValue`` has internal linkage (equivalent to being ``static`` in C),
2879 it is not visible to code outside the current translation unit, and does not
2880 participate in linking. If it has external linkage, it is visible to external
2881 code, and does participate in linking. In addition to linkage information,
2882 ``GlobalValue``\ s keep track of which Module_ they are currently part of.
2884 Because ``GlobalValue``\ s are memory objects, they are always referred to by
2885 their **address**. As such, the Type_ of a global is always a pointer to its
2886 contents. It is important to remember this when using the ``GetElementPtrInst``
2887 instruction because this pointer must be dereferenced first. For example, if
2888 you have a ``GlobalVariable`` (a subclass of ``GlobalValue)`` that is an array
2889 of 24 ints, type ``[24 x i32]``, then the ``GlobalVariable`` is a pointer to
2890 that array. Although the address of the first element of this array and the
2891 value of the ``GlobalVariable`` are the same, they have different types. The
2892 ``GlobalVariable``'s type is ``[24 x i32]``. The first element's type is
2893 ``i32.`` Because of this, accessing a global value requires you to dereference
2894 the pointer with ``GetElementPtrInst`` first, then its elements can be accessed.
2895 This is explained in the `LLVM Language Reference Manual
2896 <LangRef.html#globalvars>`_.
2900 Important Public Members of the ``GlobalValue`` class
2901 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2903 * | ``bool hasInternalLinkage() const``
2904 | ``bool hasExternalLinkage() const``
2905 | ``void setInternalLinkage(bool HasInternalLinkage)``
2907 These methods manipulate the linkage characteristics of the ``GlobalValue``.
2909 * ``Module *getParent()``
2911 This returns the Module_ that the
2912 GlobalValue is currently embedded into.
2916 The ``Function`` class
2917 ----------------------
2919 ``#include "llvm/Function.h"``
2921 header source: `Function.h <http://llvm.org/doxygen/Function_8h-source.html>`_
2923 doxygen info: `Function Class
2924 <http://llvm.org/doxygen/classllvm_1_1Function.html>`_
2926 Superclasses: GlobalValue_, Constant_, User_, Value_
2928 The ``Function`` class represents a single procedure in LLVM. It is actually
2929 one of the more complex classes in the LLVM hierarchy because it must keep track
2930 of a large amount of data. The ``Function`` class keeps track of a list of
2931 BasicBlock_\ s, a list of formal Argument_\ s, and a SymbolTable_.
2933 The list of BasicBlock_\ s is the most commonly used part of ``Function``
2934 objects. The list imposes an implicit ordering of the blocks in the function,
2935 which indicate how the code will be laid out by the backend. Additionally, the
2936 first BasicBlock_ is the implicit entry node for the ``Function``. It is not
2937 legal in LLVM to explicitly branch to this initial block. There are no implicit
2938 exit nodes, and in fact there may be multiple exit nodes from a single
2939 ``Function``. If the BasicBlock_ list is empty, this indicates that the
2940 ``Function`` is actually a function declaration: the actual body of the function
2941 hasn't been linked in yet.
2943 In addition to a list of BasicBlock_\ s, the ``Function`` class also keeps track
2944 of the list of formal Argument_\ s that the function receives. This container
2945 manages the lifetime of the Argument_ nodes, just like the BasicBlock_ list does
2946 for the BasicBlock_\ s.
2948 The SymbolTable_ is a very rarely used LLVM feature that is only used when you
2949 have to look up a value by name. Aside from that, the SymbolTable_ is used
2950 internally to make sure that there are not conflicts between the names of
2951 Instruction_\ s, BasicBlock_\ s, or Argument_\ s in the function body.
2953 Note that ``Function`` is a GlobalValue_ and therefore also a Constant_. The
2954 value of the function is its address (after linking) which is guaranteed to be
2959 Important Public Members of the ``Function``
2960 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2962 * ``Function(const FunctionType *Ty, LinkageTypes Linkage,
2963 const std::string &N = "", Module* Parent = 0)``
2965 Constructor used when you need to create new ``Function``\ s to add the
2966 program. The constructor must specify the type of the function to create and
2967 what type of linkage the function should have. The FunctionType_ argument
2968 specifies the formal arguments and return value for the function. The same
2969 FunctionType_ value can be used to create multiple functions. The ``Parent``
2970 argument specifies the Module in which the function is defined. If this
2971 argument is provided, the function will automatically be inserted into that
2972 module's list of functions.
2974 * ``bool isDeclaration()``
2976 Return whether or not the ``Function`` has a body defined. If the function is
2977 "external", it does not have a body, and thus must be resolved by linking with
2978 a function defined in a different translation unit.
2980 * | ``Function::iterator`` - Typedef for basic block list iterator
2981 | ``Function::const_iterator`` - Typedef for const_iterator.
2982 | ``begin()``, ``end()``, ``size()``, ``empty()``
2984 These are forwarding methods that make it easy to access the contents of a
2985 ``Function`` object's BasicBlock_ list.
2987 * ``Function::BasicBlockListType &getBasicBlockList()``
2989 Returns the list of BasicBlock_\ s. This is necessary to use when you need to
2990 update the list or perform a complex action that doesn't have a forwarding
2993 * | ``Function::arg_iterator`` - Typedef for the argument list iterator
2994 | ``Function::const_arg_iterator`` - Typedef for const_iterator.
2995 | ``arg_begin()``, ``arg_end()``, ``arg_size()``, ``arg_empty()``
2997 These are forwarding methods that make it easy to access the contents of a
2998 ``Function`` object's Argument_ list.
3000 * ``Function::ArgumentListType &getArgumentList()``
3002 Returns the list of Argument_. This is necessary to use when you need to
3003 update the list or perform a complex action that doesn't have a forwarding
3006 * ``BasicBlock &getEntryBlock()``
3008 Returns the entry ``BasicBlock`` for the function. Because the entry block
3009 for the function is always the first block, this returns the first block of
3012 * | ``Type *getReturnType()``
3013 | ``FunctionType *getFunctionType()``
3015 This traverses the Type_ of the ``Function`` and returns the return type of
3016 the function, or the FunctionType_ of the actual function.
3018 * ``SymbolTable *getSymbolTable()``
3020 Return a pointer to the SymbolTable_ for this ``Function``.
3024 The ``GlobalVariable`` class
3025 ----------------------------
3027 ``#include "llvm/GlobalVariable.h"``
3029 header source: `GlobalVariable.h
3030 <http://llvm.org/doxygen/GlobalVariable_8h-source.html>`_
3032 doxygen info: `GlobalVariable Class
3033 <http://llvm.org/doxygen/classllvm_1_1GlobalVariable.html>`_
3035 Superclasses: GlobalValue_, Constant_, User_, Value_
3037 Global variables are represented with the (surprise surprise) ``GlobalVariable``
3038 class. Like functions, ``GlobalVariable``\ s are also subclasses of
3039 GlobalValue_, and as such are always referenced by their address (global values
3040 must live in memory, so their "name" refers to their constant address). See
3041 GlobalValue_ for more on this. Global variables may have an initial value
3042 (which must be a Constant_), and if they have an initializer, they may be marked
3043 as "constant" themselves (indicating that their contents never change at
3046 .. _m_GlobalVariable:
3048 Important Public Members of the ``GlobalVariable`` class
3049 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3051 * ``GlobalVariable(const Type *Ty, bool isConstant, LinkageTypes &Linkage,
3052 Constant *Initializer = 0, const std::string &Name = "", Module* Parent = 0)``
3054 Create a new global variable of the specified type. If ``isConstant`` is true
3055 then the global variable will be marked as unchanging for the program. The
3056 Linkage parameter specifies the type of linkage (internal, external, weak,
3057 linkonce, appending) for the variable. If the linkage is InternalLinkage,
3058 WeakAnyLinkage, WeakODRLinkage, LinkOnceAnyLinkage or LinkOnceODRLinkage, then
3059 the resultant global variable will have internal linkage. AppendingLinkage
3060 concatenates together all instances (in different translation units) of the
3061 variable into a single variable but is only applicable to arrays. See the
3062 `LLVM Language Reference <LangRef.html#modulestructure>`_ for further details
3063 on linkage types. Optionally an initializer, a name, and the module to put
3064 the variable into may be specified for the global variable as well.
3066 * ``bool isConstant() const``
3068 Returns true if this is a global variable that is known not to be modified at
3071 * ``bool hasInitializer()``
3073 Returns true if this ``GlobalVariable`` has an intializer.
3075 * ``Constant *getInitializer()``
3077 Returns the initial value for a ``GlobalVariable``. It is not legal to call
3078 this method if there is no initializer.
3082 The ``BasicBlock`` class
3083 ------------------------
3085 ``#include "llvm/BasicBlock.h"``
3087 header source: `BasicBlock.h
3088 <http://llvm.org/doxygen/BasicBlock_8h-source.html>`_
3090 doxygen info: `BasicBlock Class
3091 <http://llvm.org/doxygen/classllvm_1_1BasicBlock.html>`_
3095 This class represents a single entry single exit section of the code, commonly
3096 known as a basic block by the compiler community. The ``BasicBlock`` class
3097 maintains a list of Instruction_\ s, which form the body of the block. Matching
3098 the language definition, the last element of this list of instructions is always
3099 a terminator instruction (a subclass of the TerminatorInst_ class).
3101 In addition to tracking the list of instructions that make up the block, the
3102 ``BasicBlock`` class also keeps track of the :ref:`Function <c_Function>` that
3103 it is embedded into.
3105 Note that ``BasicBlock``\ s themselves are Value_\ s, because they are
3106 referenced by instructions like branches and can go in the switch tables.
3107 ``BasicBlock``\ s have type ``label``.
3111 Important Public Members of the ``BasicBlock`` class
3112 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3114 * ``BasicBlock(const std::string &Name = "", Function *Parent = 0)``
3116 The ``BasicBlock`` constructor is used to create new basic blocks for
3117 insertion into a function. The constructor optionally takes a name for the
3118 new block, and a :ref:`Function <c_Function>` to insert it into. If the
3119 ``Parent`` parameter is specified, the new ``BasicBlock`` is automatically
3120 inserted at the end of the specified :ref:`Function <c_Function>`, if not
3121 specified, the BasicBlock must be manually inserted into the :ref:`Function
3124 * | ``BasicBlock::iterator`` - Typedef for instruction list iterator
3125 | ``BasicBlock::const_iterator`` - Typedef for const_iterator.
3126 | ``begin()``, ``end()``, ``front()``, ``back()``,
3127 ``size()``, ``empty()``
3128 STL-style functions for accessing the instruction list.
3130 These methods and typedefs are forwarding functions that have the same
3131 semantics as the standard library methods of the same names. These methods
3132 expose the underlying instruction list of a basic block in a way that is easy
3133 to manipulate. To get the full complement of container operations (including
3134 operations to update the list), you must use the ``getInstList()`` method.
3136 * ``BasicBlock::InstListType &getInstList()``
3138 This method is used to get access to the underlying container that actually
3139 holds the Instructions. This method must be used when there isn't a
3140 forwarding function in the ``BasicBlock`` class for the operation that you
3141 would like to perform. Because there are no forwarding functions for
3142 "updating" operations, you need to use this if you want to update the contents
3143 of a ``BasicBlock``.
3145 * ``Function *getParent()``
3147 Returns a pointer to :ref:`Function <c_Function>` the block is embedded into,
3148 or a null pointer if it is homeless.
3150 * ``TerminatorInst *getTerminator()``
3152 Returns a pointer to the terminator instruction that appears at the end of the
3153 ``BasicBlock``. If there is no terminator instruction, or if the last
3154 instruction in the block is not a terminator, then a null pointer is returned.
3158 The ``Argument`` class
3159 ----------------------
3161 This subclass of Value defines the interface for incoming formal arguments to a
3162 function. A Function maintains a list of its formal arguments. An argument has
3163 a pointer to the parent Function.