1 ========================
2 LLVM Programmer's Manual
3 ========================
9 This is always a work in progress.
16 This document is meant to highlight some of the important classes and interfaces
17 available in the LLVM source-base. This manual is not intended to explain what
18 LLVM is, how it works, and what LLVM code looks like. It assumes that you know
19 the basics of LLVM and are interested in writing transformations or otherwise
20 analyzing or manipulating the code.
22 This document should get you oriented so that you can find your way in the
23 continuously growing source code that makes up the LLVM infrastructure. Note
24 that this manual is not intended to serve as a replacement for reading the
25 source code, so if you think there should be a method in one of these classes to
26 do something, but it's not listed, check the source. Links to the `doxygen
27 <http://llvm.org/doxygen/>`__ sources are provided to make this as easy as
30 The first section of this document describes general information that is useful
31 to know when working in the LLVM infrastructure, and the second describes the
32 Core LLVM classes. In the future this manual will be extended with information
33 describing how to use extension libraries, such as dominator information, CFG
34 traversal routines, and useful utilities like the ``InstVisitor`` (`doxygen
35 <http://llvm.org/doxygen/InstVisitor_8h-source.html>`__) template.
42 This section contains general information that is useful if you are working in
43 the LLVM source-base, but that isn't specific to any particular API.
47 The C++ Standard Template Library
48 ---------------------------------
50 LLVM makes heavy use of the C++ Standard Template Library (STL), perhaps much
51 more than you are used to, or have seen before. Because of this, you might want
52 to do a little background reading in the techniques used and capabilities of the
53 library. There are many good pages that discuss the STL, and several books on
54 the subject that you can get, so it will not be discussed in this document.
56 Here are some useful links:
59 <http://en.cppreference.com/w/>`_ - an excellent
60 reference for the STL and other parts of the standard C++ library.
62 #. `C++ In a Nutshell <http://www.tempest-sw.com/cpp/>`_ - This is an O'Reilly
63 book in the making. It has a decent Standard Library Reference that rivals
64 Dinkumware's, and is unfortunately no longer free since the book has been
67 #. `C++ Frequently Asked Questions <http://www.parashift.com/c++-faq-lite/>`_.
69 #. `SGI's STL Programmer's Guide <http://www.sgi.com/tech/stl/>`_ - Contains a
70 useful `Introduction to the STL
71 <http://www.sgi.com/tech/stl/stl_introduction.html>`_.
73 #. `Bjarne Stroustrup's C++ Page
74 <http://www.research.att.com/%7Ebs/C++.html>`_.
76 #. `Bruce Eckel's Thinking in C++, 2nd ed. Volume 2 Revision 4.0
77 (even better, get the book)
78 <http://www.mindview.net/Books/TICPP/ThinkingInCPP2e.html>`_.
80 You are also encouraged to take a look at the :doc:`LLVM Coding Standards
81 <CodingStandards>` guide which focuses on how to write maintainable code more
82 than where to put your curly braces.
86 Other useful references
87 -----------------------
89 #. `Using static and shared libraries across platforms
90 <http://www.fortran-2000.com/ArnaudRecipes/sharedlib.html>`_
94 Important and useful LLVM APIs
95 ==============================
97 Here we highlight some LLVM APIs that are generally useful and good to know
98 about when writing transformations.
102 The ``isa<>``, ``cast<>`` and ``dyn_cast<>`` templates
103 ------------------------------------------------------
105 The LLVM source-base makes extensive use of a custom form of RTTI. These
106 templates have many similarities to the C++ ``dynamic_cast<>`` operator, but
107 they don't have some drawbacks (primarily stemming from the fact that
108 ``dynamic_cast<>`` only works on classes that have a v-table). Because they are
109 used so often, you must know what they do and how they work. All of these
110 templates are defined in the ``llvm/Support/Casting.h`` (`doxygen
111 <http://llvm.org/doxygen/Casting_8h-source.html>`__) file (note that you very
112 rarely have to include this file directly).
115 The ``isa<>`` operator works exactly like the Java "``instanceof``" operator.
116 It returns true or false depending on whether a reference or pointer points to
117 an instance of the specified class. This can be very useful for constraint
118 checking of various sorts (example below).
121 The ``cast<>`` operator is a "checked cast" operation. It converts a pointer
122 or reference from a base class to a derived class, causing an assertion
123 failure if it is not really an instance of the right type. This should be
124 used in cases where you have some information that makes you believe that
125 something is of the right type. An example of the ``isa<>`` and ``cast<>``
130 static bool isLoopInvariant(const Value *V, const Loop *L) {
131 if (isa<Constant>(V) || isa<Argument>(V) || isa<GlobalValue>(V))
134 // Otherwise, it must be an instruction...
135 return !L->contains(cast<Instruction>(V)->getParent());
138 Note that you should **not** use an ``isa<>`` test followed by a ``cast<>``,
139 for that use the ``dyn_cast<>`` operator.
142 The ``dyn_cast<>`` operator is a "checking cast" operation. It checks to see
143 if the operand is of the specified type, and if so, returns a pointer to it
144 (this operator does not work with references). If the operand is not of the
145 correct type, a null pointer is returned. Thus, this works very much like
146 the ``dynamic_cast<>`` operator in C++, and should be used in the same
147 circumstances. Typically, the ``dyn_cast<>`` operator is used in an ``if``
148 statement or some other flow control statement like this:
152 if (AllocationInst *AI = dyn_cast<AllocationInst>(Val)) {
156 This form of the ``if`` statement effectively combines together a call to
157 ``isa<>`` and a call to ``cast<>`` into one statement, which is very
160 Note that the ``dyn_cast<>`` operator, like C++'s ``dynamic_cast<>`` or Java's
161 ``instanceof`` operator, can be abused. In particular, you should not use big
162 chained ``if/then/else`` blocks to check for lots of different variants of
163 classes. If you find yourself wanting to do this, it is much cleaner and more
164 efficient to use the ``InstVisitor`` class to dispatch over the instruction
168 The ``cast_or_null<>`` operator works just like the ``cast<>`` operator,
169 except that it allows for a null pointer as an argument (which it then
170 propagates). This can sometimes be useful, allowing you to combine several
171 null checks into one.
173 ``dyn_cast_or_null<>``:
174 The ``dyn_cast_or_null<>`` operator works just like the ``dyn_cast<>``
175 operator, except that it allows for a null pointer as an argument (which it
176 then propagates). This can sometimes be useful, allowing you to combine
177 several null checks into one.
179 These five templates can be used with any classes, whether they have a v-table
180 or not. If you want to add support for these templates, see the document
181 :doc:`How to set up LLVM-style RTTI for your class hierarchy
182 <HowToSetUpLLVMStyleRTTI>`
186 Passing strings (the ``StringRef`` and ``Twine`` classes)
187 ---------------------------------------------------------
189 Although LLVM generally does not do much string manipulation, we do have several
190 important APIs which take strings. Two important examples are the Value class
191 -- which has names for instructions, functions, etc. -- and the ``StringMap``
192 class which is used extensively in LLVM and Clang.
194 These are generic classes, and they need to be able to accept strings which may
195 have embedded null characters. Therefore, they cannot simply take a ``const
196 char *``, and taking a ``const std::string&`` requires clients to perform a heap
197 allocation which is usually unnecessary. Instead, many LLVM APIs use a
198 ``StringRef`` or a ``const Twine&`` for passing strings efficiently.
202 The ``StringRef`` class
203 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
205 The ``StringRef`` data type represents a reference to a constant string (a
206 character array and a length) and supports the common operations available on
207 ``std::string``, but does not require heap allocation.
209 It can be implicitly constructed using a C style null-terminated string, an
210 ``std::string``, or explicitly with a character pointer and length. For
211 example, the ``StringRef`` find function is declared as:
215 iterator find(StringRef Key);
217 and clients can call it using any one of:
221 Map.find("foo"); // Lookup "foo"
222 Map.find(std::string("bar")); // Lookup "bar"
223 Map.find(StringRef("\0baz", 4)); // Lookup "\0baz"
225 Similarly, APIs which need to return a string may return a ``StringRef``
226 instance, which can be used directly or converted to an ``std::string`` using
227 the ``str`` member function. See ``llvm/ADT/StringRef.h`` (`doxygen
228 <http://llvm.org/doxygen/classllvm_1_1StringRef_8h-source.html>`__) for more
231 You should rarely use the ``StringRef`` class directly, because it contains
232 pointers to external memory it is not generally safe to store an instance of the
233 class (unless you know that the external storage will not be freed).
234 ``StringRef`` is small and pervasive enough in LLVM that it should always be
240 The ``Twine`` (`doxygen <http://llvm.org/doxygen/classllvm_1_1Twine.html>`__)
241 class is an efficient way for APIs to accept concatenated strings. For example,
242 a common LLVM paradigm is to name one instruction based on the name of another
243 instruction with a suffix, for example:
247 New = CmpInst::Create(..., SO->getName() + ".cmp");
249 The ``Twine`` class is effectively a lightweight `rope
250 <http://en.wikipedia.org/wiki/Rope_(computer_science)>`_ which points to
251 temporary (stack allocated) objects. Twines can be implicitly constructed as
252 the result of the plus operator applied to strings (i.e., a C strings, an
253 ``std::string``, or a ``StringRef``). The twine delays the actual concatenation
254 of strings until it is actually required, at which point it can be efficiently
255 rendered directly into a character array. This avoids unnecessary heap
256 allocation involved in constructing the temporary results of string
257 concatenation. See ``llvm/ADT/Twine.h`` (`doxygen
258 <http://llvm.org/doxygen/Twine_8h_source.html>`__) and :ref:`here <dss_twine>`
259 for more information.
261 As with a ``StringRef``, ``Twine`` objects point to external memory and should
262 almost never be stored or mentioned directly. They are intended solely for use
263 when defining a function which should be able to efficiently accept concatenated
268 The ``DEBUG()`` macro and ``-debug`` option
269 -------------------------------------------
271 Often when working on your pass you will put a bunch of debugging printouts and
272 other code into your pass. After you get it working, you want to remove it, but
273 you may need it again in the future (to work out new bugs that you run across).
275 Naturally, because of this, you don't want to delete the debug printouts, but
276 you don't want them to always be noisy. A standard compromise is to comment
277 them out, allowing you to enable them if you need them in the future.
279 The ``llvm/Support/Debug.h`` (`doxygen
280 <http://llvm.org/doxygen/Debug_8h-source.html>`__) file provides a macro named
281 ``DEBUG()`` that is a much nicer solution to this problem. Basically, you can
282 put arbitrary code into the argument of the ``DEBUG`` macro, and it is only
283 executed if '``opt``' (or any other tool) is run with the '``-debug``' command
288 DEBUG(errs() << "I am here!\n");
290 Then you can run your pass like this:
294 $ opt < a.bc > /dev/null -mypass
296 $ opt < a.bc > /dev/null -mypass -debug
299 Using the ``DEBUG()`` macro instead of a home-brewed solution allows you to not
300 have to create "yet another" command line option for the debug output for your
301 pass. Note that ``DEBUG()`` macros are disabled for optimized builds, so they
302 do not cause a performance impact at all (for the same reason, they should also
303 not contain side-effects!).
305 One additional nice thing about the ``DEBUG()`` macro is that you can enable or
306 disable it directly in gdb. Just use "``set DebugFlag=0``" or "``set
307 DebugFlag=1``" from the gdb if the program is running. If the program hasn't
308 been started yet, you can always just run it with ``-debug``.
312 Fine grained debug info with ``DEBUG_TYPE`` and the ``-debug-only`` option
313 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
315 Sometimes you may find yourself in a situation where enabling ``-debug`` just
316 turns on **too much** information (such as when working on the code generator).
317 If you want to enable debug information with more fine-grained control, you
318 define the ``DEBUG_TYPE`` macro and the ``-debug`` only option as follows:
323 DEBUG(errs() << "No debug type\n");
324 #define DEBUG_TYPE "foo"
325 DEBUG(errs() << "'foo' debug type\n");
327 #define DEBUG_TYPE "bar"
328 DEBUG(errs() << "'bar' debug type\n"));
330 #define DEBUG_TYPE ""
331 DEBUG(errs() << "No debug type (2)\n");
333 Then you can run your pass like this:
337 $ opt < a.bc > /dev/null -mypass
339 $ opt < a.bc > /dev/null -mypass -debug
344 $ opt < a.bc > /dev/null -mypass -debug-only=foo
346 $ opt < a.bc > /dev/null -mypass -debug-only=bar
349 Of course, in practice, you should only set ``DEBUG_TYPE`` at the top of a file,
350 to specify the debug type for the entire module (if you do this before you
351 ``#include "llvm/Support/Debug.h"``, you don't have to insert the ugly
352 ``#undef``'s). Also, you should use names more meaningful than "foo" and "bar",
353 because there is no system in place to ensure that names do not conflict. If
354 two different modules use the same string, they will all be turned on when the
355 name is specified. This allows, for example, all debug information for
356 instruction scheduling to be enabled with ``-debug-type=InstrSched``, even if
357 the source lives in multiple files.
359 The ``DEBUG_WITH_TYPE`` macro is also available for situations where you would
360 like to set ``DEBUG_TYPE``, but only for one specific ``DEBUG`` statement. It
361 takes an additional first parameter, which is the type to use. For example, the
362 preceding example could be written as:
366 DEBUG_WITH_TYPE("", errs() << "No debug type\n");
367 DEBUG_WITH_TYPE("foo", errs() << "'foo' debug type\n");
368 DEBUG_WITH_TYPE("bar", errs() << "'bar' debug type\n"));
369 DEBUG_WITH_TYPE("", errs() << "No debug type (2)\n");
373 The ``Statistic`` class & ``-stats`` option
374 -------------------------------------------
376 The ``llvm/ADT/Statistic.h`` (`doxygen
377 <http://llvm.org/doxygen/Statistic_8h-source.html>`__) file provides a class
378 named ``Statistic`` that is used as a unified way to keep track of what the LLVM
379 compiler is doing and how effective various optimizations are. It is useful to
380 see what optimizations are contributing to making a particular program run
383 Often you may run your pass on some big program, and you're interested to see
384 how many times it makes a certain transformation. Although you can do this with
385 hand inspection, or some ad-hoc method, this is a real pain and not very useful
386 for big programs. Using the ``Statistic`` class makes it very easy to keep
387 track of this information, and the calculated information is presented in a
388 uniform manner with the rest of the passes being executed.
390 There are many examples of ``Statistic`` uses, but the basics of using it are as
393 #. Define your statistic like this:
397 #define DEBUG_TYPE "mypassname" // This goes before any #includes.
398 STATISTIC(NumXForms, "The # of times I did stuff");
400 The ``STATISTIC`` macro defines a static variable, whose name is specified by
401 the first argument. The pass name is taken from the ``DEBUG_TYPE`` macro, and
402 the description is taken from the second argument. The variable defined
403 ("NumXForms" in this case) acts like an unsigned integer.
405 #. Whenever you make a transformation, bump the counter:
409 ++NumXForms; // I did stuff!
411 That's all you have to do. To get '``opt``' to print out the statistics
412 gathered, use the '``-stats``' option:
416 $ opt -stats -mypassname < program.bc > /dev/null
417 ... statistics output ...
419 When running ``opt`` on a C file from the SPEC benchmark suite, it gives a
420 report that looks like this:
424 7646 bitcodewriter - Number of normal instructions
425 725 bitcodewriter - Number of oversized instructions
426 129996 bitcodewriter - Number of bitcode bytes written
427 2817 raise - Number of insts DCEd or constprop'd
428 3213 raise - Number of cast-of-self removed
429 5046 raise - Number of expression trees converted
430 75 raise - Number of other getelementptr's formed
431 138 raise - Number of load/store peepholes
432 42 deadtypeelim - Number of unused typenames removed from symtab
433 392 funcresolve - Number of varargs functions resolved
434 27 globaldce - Number of global variables removed
435 2 adce - Number of basic blocks removed
436 134 cee - Number of branches revectored
437 49 cee - Number of setcc instruction eliminated
438 532 gcse - Number of loads removed
439 2919 gcse - Number of instructions removed
440 86 indvars - Number of canonical indvars added
441 87 indvars - Number of aux indvars removed
442 25 instcombine - Number of dead inst eliminate
443 434 instcombine - Number of insts combined
444 248 licm - Number of load insts hoisted
445 1298 licm - Number of insts hoisted to a loop pre-header
446 3 licm - Number of insts hoisted to multiple loop preds (bad, no loop pre-header)
447 75 mem2reg - Number of alloca's promoted
448 1444 cfgsimplify - Number of blocks simplified
450 Obviously, with so many optimizations, having a unified framework for this stuff
451 is very nice. Making your pass fit well into the framework makes it more
452 maintainable and useful.
456 Viewing graphs while debugging code
457 -----------------------------------
459 Several of the important data structures in LLVM are graphs: for example CFGs
460 made out of LLVM :ref:`BasicBlocks <BasicBlock>`, CFGs made out of LLVM
461 :ref:`MachineBasicBlocks <MachineBasicBlock>`, and :ref:`Instruction Selection
462 DAGs <SelectionDAG>`. In many cases, while debugging various parts of the
463 compiler, it is nice to instantly visualize these graphs.
465 LLVM provides several callbacks that are available in a debug build to do
466 exactly that. If you call the ``Function::viewCFG()`` method, for example, the
467 current LLVM tool will pop up a window containing the CFG for the function where
468 each basic block is a node in the graph, and each node contains the instructions
469 in the block. Similarly, there also exists ``Function::viewCFGOnly()`` (does
470 not include the instructions), the ``MachineFunction::viewCFG()`` and
471 ``MachineFunction::viewCFGOnly()``, and the ``SelectionDAG::viewGraph()``
472 methods. Within GDB, for example, you can usually use something like ``call
473 DAG.viewGraph()`` to pop up a window. Alternatively, you can sprinkle calls to
474 these functions in your code in places you want to debug.
476 Getting this to work requires a small amount of configuration. On Unix systems
477 with X11, install the `graphviz <http://www.graphviz.org>`_ toolkit, and make
478 sure 'dot' and 'gv' are in your path. If you are running on Mac OS X, download
479 and install the Mac OS X `Graphviz program
480 <http://www.pixelglow.com/graphviz/>`_ and add
481 ``/Applications/Graphviz.app/Contents/MacOS/`` (or wherever you install it) to
482 your path. Once in your system and path are set up, rerun the LLVM configure
483 script and rebuild LLVM to enable this functionality.
485 ``SelectionDAG`` has been extended to make it easier to locate *interesting*
486 nodes in large complex graphs. From gdb, if you ``call DAG.setGraphColor(node,
487 "color")``, then the next ``call DAG.viewGraph()`` would highlight the node in
488 the specified color (choices of colors can be found at `colors
489 <http://www.graphviz.org/doc/info/colors.html>`_.) More complex node attributes
490 can be provided with ``call DAG.setGraphAttrs(node, "attributes")`` (choices can
491 be found at `Graph attributes <http://www.graphviz.org/doc/info/attrs.html>`_.)
492 If you want to restart and clear all the current graph attributes, then you can
493 ``call DAG.clearGraphAttrs()``.
495 Note that graph visualization features are compiled out of Release builds to
496 reduce file size. This means that you need a Debug+Asserts or Release+Asserts
497 build to use these features.
501 Picking the Right Data Structure for a Task
502 ===========================================
504 LLVM has a plethora of data structures in the ``llvm/ADT/`` directory, and we
505 commonly use STL data structures. This section describes the trade-offs you
506 should consider when you pick one.
508 The first step is a choose your own adventure: do you want a sequential
509 container, a set-like container, or a map-like container? The most important
510 thing when choosing a container is the algorithmic properties of how you plan to
511 access the container. Based on that, you should use:
514 * a :ref:`map-like <ds_map>` container if you need efficient look-up of a
515 value based on another value. Map-like containers also support efficient
516 queries for containment (whether a key is in the map). Map-like containers
517 generally do not support efficient reverse mapping (values to keys). If you
518 need that, use two maps. Some map-like containers also support efficient
519 iteration through the keys in sorted order. Map-like containers are the most
520 expensive sort, only use them if you need one of these capabilities.
522 * a :ref:`set-like <ds_set>` container if you need to put a bunch of stuff into
523 a container that automatically eliminates duplicates. Some set-like
524 containers support efficient iteration through the elements in sorted order.
525 Set-like containers are more expensive than sequential containers.
527 * a :ref:`sequential <ds_sequential>` container provides the most efficient way
528 to add elements and keeps track of the order they are added to the collection.
529 They permit duplicates and support efficient iteration, but do not support
530 efficient look-up based on a key.
532 * a :ref:`string <ds_string>` container is a specialized sequential container or
533 reference structure that is used for character or byte arrays.
535 * a :ref:`bit <ds_bit>` container provides an efficient way to store and
536 perform set operations on sets of numeric id's, while automatically
537 eliminating duplicates. Bit containers require a maximum of 1 bit for each
538 identifier you want to store.
540 Once the proper category of container is determined, you can fine tune the
541 memory use, constant factors, and cache behaviors of access by intelligently
542 picking a member of the category. Note that constant factors and cache behavior
543 can be a big deal. If you have a vector that usually only contains a few
544 elements (but could contain many), for example, it's much better to use
545 :ref:`SmallVector <dss_smallvector>` than :ref:`vector <dss_vector>`. Doing so
546 avoids (relatively) expensive malloc/free calls, which dwarf the cost of adding
547 the elements to the container.
551 Sequential Containers (std::vector, std::list, etc)
552 ---------------------------------------------------
554 There are a variety of sequential containers available for you, based on your
555 needs. Pick the first in this section that will do what you want.
562 The ``llvm::ArrayRef`` class is the preferred class to use in an interface that
563 accepts a sequential list of elements in memory and just reads from them. By
564 taking an ``ArrayRef``, the API can be passed a fixed size array, an
565 ``std::vector``, an ``llvm::SmallVector`` and anything else that is contiguous
573 Fixed size arrays are very simple and very fast. They are good if you know
574 exactly how many elements you have, or you have a (low) upper bound on how many
579 Heap Allocated Arrays
580 ^^^^^^^^^^^^^^^^^^^^^
582 Heap allocated arrays (``new[]`` + ``delete[]``) are also simple. They are good
583 if the number of elements is variable, if you know how many elements you will
584 need before the array is allocated, and if the array is usually large (if not,
585 consider a :ref:`SmallVector <dss_smallvector>`). The cost of a heap allocated
586 array is the cost of the new/delete (aka malloc/free). Also note that if you
587 are allocating an array of a type with a constructor, the constructor and
588 destructors will be run for every element in the array (re-sizable vectors only
589 construct those elements actually used).
591 .. _dss_tinyptrvector:
593 llvm/ADT/TinyPtrVector.h
594 ^^^^^^^^^^^^^^^^^^^^^^^^
596 ``TinyPtrVector<Type>`` is a highly specialized collection class that is
597 optimized to avoid allocation in the case when a vector has zero or one
598 elements. It has two major restrictions: 1) it can only hold values of pointer
599 type, and 2) it cannot hold a null pointer.
601 Since this container is highly specialized, it is rarely used.
605 llvm/ADT/SmallVector.h
606 ^^^^^^^^^^^^^^^^^^^^^^
608 ``SmallVector<Type, N>`` is a simple class that looks and smells just like
609 ``vector<Type>``: it supports efficient iteration, lays out elements in memory
610 order (so you can do pointer arithmetic between elements), supports efficient
611 push_back/pop_back operations, supports efficient random access to its elements,
614 The advantage of SmallVector is that it allocates space for some number of
615 elements (N) **in the object itself**. Because of this, if the SmallVector is
616 dynamically smaller than N, no malloc is performed. This can be a big win in
617 cases where the malloc/free call is far more expensive than the code that
618 fiddles around with the elements.
620 This is good for vectors that are "usually small" (e.g. the number of
621 predecessors/successors of a block is usually less than 8). On the other hand,
622 this makes the size of the SmallVector itself large, so you don't want to
623 allocate lots of them (doing so will waste a lot of space). As such,
624 SmallVectors are most useful when on the stack.
626 SmallVector also provides a nice portable and efficient replacement for
631 Prefer to use ``SmallVectorImpl<T>`` as a parameter type.
633 In APIs that don't care about the "small size" (most?), prefer to use
634 the ``SmallVectorImpl<T>`` class, which is basically just the "vector
635 header" (and methods) without the elements allocated after it. Note that
636 ``SmallVector<T, N>`` inherits from ``SmallVectorImpl<T>`` so the
637 conversion is implicit and costs nothing. E.g.
641 // BAD: Clients cannot pass e.g. SmallVector<Foo, 4>.
642 hardcodedSmallSize(SmallVector<Foo, 2> &Out);
643 // GOOD: Clients can pass any SmallVector<Foo, N>.
644 allowsAnySmallSize(SmallVectorImpl<Foo> &Out);
647 SmallVector<Foo, 8> Vec;
648 hardcodedSmallSize(Vec); // Error.
649 allowsAnySmallSize(Vec); // Works.
652 Even though it has "``Impl``" in the name, this is so widely used that
653 it really isn't "private to the implementation" anymore. A name like
654 ``SmallVectorHeader`` would be more appropriate.
661 ``std::vector`` is well loved and respected. It is useful when SmallVector
662 isn't: when the size of the vector is often large (thus the small optimization
663 will rarely be a benefit) or if you will be allocating many instances of the
664 vector itself (which would waste space for elements that aren't in the
665 container). vector is also useful when interfacing with code that expects
668 One worthwhile note about std::vector: avoid code like this:
677 Instead, write this as:
687 Doing so will save (at least) one heap allocation and free per iteration of the
695 ``std::deque`` is, in some senses, a generalized version of ``std::vector``.
696 Like ``std::vector``, it provides constant time random access and other similar
697 properties, but it also provides efficient access to the front of the list. It
698 does not guarantee continuity of elements within memory.
700 In exchange for this extra flexibility, ``std::deque`` has significantly higher
701 constant factor costs than ``std::vector``. If possible, use ``std::vector`` or
709 ``std::list`` is an extremely inefficient class that is rarely useful. It
710 performs a heap allocation for every element inserted into it, thus having an
711 extremely high constant factor, particularly for small data types.
712 ``std::list`` also only supports bidirectional iteration, not random access
715 In exchange for this high cost, std::list supports efficient access to both ends
716 of the list (like ``std::deque``, but unlike ``std::vector`` or
717 ``SmallVector``). In addition, the iterator invalidation characteristics of
718 std::list are stronger than that of a vector class: inserting or removing an
719 element into the list does not invalidate iterator or pointers to other elements
727 ``ilist<T>`` implements an 'intrusive' doubly-linked list. It is intrusive,
728 because it requires the element to store and provide access to the prev/next
729 pointers for the list.
731 ``ilist`` has the same drawbacks as ``std::list``, and additionally requires an
732 ``ilist_traits`` implementation for the element type, but it provides some novel
733 characteristics. In particular, it can efficiently store polymorphic objects,
734 the traits class is informed when an element is inserted or removed from the
735 list, and ``ilist``\ s are guaranteed to support a constant-time splice
738 These properties are exactly what we want for things like ``Instruction``\ s and
739 basic blocks, which is why these are implemented with ``ilist``\ s.
741 Related classes of interest are explained in the following subsections:
743 * :ref:`ilist_traits <dss_ilist_traits>`
745 * :ref:`iplist <dss_iplist>`
747 * :ref:`llvm/ADT/ilist_node.h <dss_ilist_node>`
749 * :ref:`Sentinels <dss_ilist_sentinel>`
751 .. _dss_packedvector:
753 llvm/ADT/PackedVector.h
754 ^^^^^^^^^^^^^^^^^^^^^^^
756 Useful for storing a vector of values using only a few number of bits for each
757 value. Apart from the standard operations of a vector-like container, it can
758 also perform an 'or' set operation.
766 FirstCondition = 0x1,
767 SecondCondition = 0x2,
772 PackedVector<State, 2> Vec1;
773 Vec1.push_back(FirstCondition);
775 PackedVector<State, 2> Vec2;
776 Vec2.push_back(SecondCondition);
779 return Vec1[0]; // returns 'Both'.
782 .. _dss_ilist_traits:
787 ``ilist_traits<T>`` is ``ilist<T>``'s customization mechanism. ``iplist<T>``
788 (and consequently ``ilist<T>``) publicly derive from this traits class.
795 ``iplist<T>`` is ``ilist<T>``'s base and as such supports a slightly narrower
796 interface. Notably, inserters from ``T&`` are absent.
798 ``ilist_traits<T>`` is a public base of this class and can be used for a wide
799 variety of customizations.
803 llvm/ADT/ilist_node.h
804 ^^^^^^^^^^^^^^^^^^^^^
806 ``ilist_node<T>`` implements a the forward and backward links that are expected
807 by the ``ilist<T>`` (and analogous containers) in the default manner.
809 ``ilist_node<T>``\ s are meant to be embedded in the node type ``T``, usually
810 ``T`` publicly derives from ``ilist_node<T>``.
812 .. _dss_ilist_sentinel:
817 ``ilist``\ s have another specialty that must be considered. To be a good
818 citizen in the C++ ecosystem, it needs to support the standard container
819 operations, such as ``begin`` and ``end`` iterators, etc. Also, the
820 ``operator--`` must work correctly on the ``end`` iterator in the case of
821 non-empty ``ilist``\ s.
823 The only sensible solution to this problem is to allocate a so-called *sentinel*
824 along with the intrusive list, which serves as the ``end`` iterator, providing
825 the back-link to the last element. However conforming to the C++ convention it
826 is illegal to ``operator++`` beyond the sentinel and it also must not be
829 These constraints allow for some implementation freedom to the ``ilist`` how to
830 allocate and store the sentinel. The corresponding policy is dictated by
831 ``ilist_traits<T>``. By default a ``T`` gets heap-allocated whenever the need
832 for a sentinel arises.
834 While the default policy is sufficient in most cases, it may break down when
835 ``T`` does not provide a default constructor. Also, in the case of many
836 instances of ``ilist``\ s, the memory overhead of the associated sentinels is
837 wasted. To alleviate the situation with numerous and voluminous
838 ``T``-sentinels, sometimes a trick is employed, leading to *ghostly sentinels*.
840 Ghostly sentinels are obtained by specially-crafted ``ilist_traits<T>`` which
841 superpose the sentinel with the ``ilist`` instance in memory. Pointer
842 arithmetic is used to obtain the sentinel, which is relative to the ``ilist``'s
843 ``this`` pointer. The ``ilist`` is augmented by an extra pointer, which serves
844 as the back-link of the sentinel. This is the only field in the ghostly
845 sentinel which can be legally accessed.
849 Other Sequential Container options
850 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
852 Other STL containers are available, such as ``std::string``.
854 There are also various STL adapter classes such as ``std::queue``,
855 ``std::priority_queue``, ``std::stack``, etc. These provide simplified access
856 to an underlying container but don't affect the cost of the container itself.
860 String-like containers
861 ----------------------
863 There are a variety of ways to pass around and use strings in C and C++, and
864 LLVM adds a few new options to choose from. Pick the first option on this list
865 that will do what you need, they are ordered according to their relative cost.
867 Note that is is generally preferred to *not* pass strings around as ``const
868 char*``'s. These have a number of problems, including the fact that they
869 cannot represent embedded nul ("\0") characters, and do not have a length
870 available efficiently. The general replacement for '``const char*``' is
873 For more information on choosing string containers for APIs, please see
874 :ref:`Passing Strings <string_apis>`.
881 The StringRef class is a simple value class that contains a pointer to a
882 character and a length, and is quite related to the :ref:`ArrayRef
883 <dss_arrayref>` class (but specialized for arrays of characters). Because
884 StringRef carries a length with it, it safely handles strings with embedded nul
885 characters in it, getting the length does not require a strlen call, and it even
886 has very convenient APIs for slicing and dicing the character range that it
889 StringRef is ideal for passing simple strings around that are known to be live,
890 either because they are C string literals, std::string, a C array, or a
891 SmallVector. Each of these cases has an efficient implicit conversion to
892 StringRef, which doesn't result in a dynamic strlen being executed.
894 StringRef has a few major limitations which make more powerful string containers
897 #. You cannot directly convert a StringRef to a 'const char*' because there is
898 no way to add a trailing nul (unlike the .c_str() method on various stronger
901 #. StringRef doesn't own or keep alive the underlying string bytes.
902 As such it can easily lead to dangling pointers, and is not suitable for
903 embedding in datastructures in most cases (instead, use an std::string or
904 something like that).
906 #. For the same reason, StringRef cannot be used as the return value of a
907 method if the method "computes" the result string. Instead, use std::string.
909 #. StringRef's do not allow you to mutate the pointed-to string bytes and it
910 doesn't allow you to insert or remove bytes from the range. For editing
911 operations like this, it interoperates with the :ref:`Twine <dss_twine>`
914 Because of its strengths and limitations, it is very common for a function to
915 take a StringRef and for a method on an object to return a StringRef that points
916 into some string that it owns.
923 The Twine class is used as an intermediary datatype for APIs that want to take a
924 string that can be constructed inline with a series of concatenations. Twine
925 works by forming recursive instances of the Twine datatype (a simple value
926 object) on the stack as temporary objects, linking them together into a tree
927 which is then linearized when the Twine is consumed. Twine is only safe to use
928 as the argument to a function, and should always be a const reference, e.g.:
932 void foo(const Twine &T);
936 foo(X + "." + Twine(i));
938 This example forms a string like "blarg.42" by concatenating the values
939 together, and does not form intermediate strings containing "blarg" or "blarg.".
941 Because Twine is constructed with temporary objects on the stack, and because
942 these instances are destroyed at the end of the current statement, it is an
943 inherently dangerous API. For example, this simple variant contains undefined
944 behavior and will probably crash:
948 void foo(const Twine &T);
952 const Twine &Tmp = X + "." + Twine(i);
955 ... because the temporaries are destroyed before the call. That said, Twine's
956 are much more efficient than intermediate std::string temporaries, and they work
957 really well with StringRef. Just be aware of their limitations.
961 llvm/ADT/SmallString.h
962 ^^^^^^^^^^^^^^^^^^^^^^
964 SmallString is a subclass of :ref:`SmallVector <dss_smallvector>` that adds some
965 convenience APIs like += that takes StringRef's. SmallString avoids allocating
966 memory in the case when the preallocated space is enough to hold its data, and
967 it calls back to general heap allocation when required. Since it owns its data,
968 it is very safe to use and supports full mutation of the string.
970 Like SmallVector's, the big downside to SmallString is their sizeof. While they
971 are optimized for small strings, they themselves are not particularly small.
972 This means that they work great for temporary scratch buffers on the stack, but
973 should not generally be put into the heap: it is very rare to see a SmallString
974 as the member of a frequently-allocated heap data structure or returned
982 The standard C++ std::string class is a very general class that (like
983 SmallString) owns its underlying data. sizeof(std::string) is very reasonable
984 so it can be embedded into heap data structures and returned by-value. On the
985 other hand, std::string is highly inefficient for inline editing (e.g.
986 concatenating a bunch of stuff together) and because it is provided by the
987 standard library, its performance characteristics depend a lot of the host
988 standard library (e.g. libc++ and MSVC provide a highly optimized string class,
989 GCC contains a really slow implementation).
991 The major disadvantage of std::string is that almost every operation that makes
992 them larger can allocate memory, which is slow. As such, it is better to use
993 SmallVector or Twine as a scratch buffer, but then use std::string to persist
998 Set-Like Containers (std::set, SmallSet, SetVector, etc)
999 --------------------------------------------------------
1001 Set-like containers are useful when you need to canonicalize multiple values
1002 into a single representation. There are several different choices for how to do
1003 this, providing various trade-offs.
1005 .. _dss_sortedvectorset:
1010 If you intend to insert a lot of elements, then do a lot of queries, a great
1011 approach is to use a vector (or other sequential container) with
1012 std::sort+std::unique to remove duplicates. This approach works really well if
1013 your usage pattern has these two distinct phases (insert then query), and can be
1014 coupled with a good choice of :ref:`sequential container <ds_sequential>`.
1016 This combination provides the several nice properties: the result data is
1017 contiguous in memory (good for cache locality), has few allocations, is easy to
1018 address (iterators in the final vector are just indices or pointers), and can be
1019 efficiently queried with a standard binary search (e.g.
1020 ``std::lower_bound``; if you want the whole range of elements comparing
1021 equal, use ``std::equal_range``).
1028 If you have a set-like data structure that is usually small and whose elements
1029 are reasonably small, a ``SmallSet<Type, N>`` is a good choice. This set has
1030 space for N elements in place (thus, if the set is dynamically smaller than N,
1031 no malloc traffic is required) and accesses them with a simple linear search.
1032 When the set grows beyond 'N' elements, it allocates a more expensive
1033 representation that guarantees efficient access (for most types, it falls back
1034 to std::set, but for pointers it uses something far better, :ref:`SmallPtrSet
1037 The magic of this class is that it handles small sets extremely efficiently, but
1038 gracefully handles extremely large sets without loss of efficiency. The
1039 drawback is that the interface is quite small: it supports insertion, queries
1040 and erasing, but does not support iteration.
1042 .. _dss_smallptrset:
1044 llvm/ADT/SmallPtrSet.h
1045 ^^^^^^^^^^^^^^^^^^^^^^
1047 SmallPtrSet has all the advantages of ``SmallSet`` (and a ``SmallSet`` of
1048 pointers is transparently implemented with a ``SmallPtrSet``), but also supports
1049 iterators. If more than 'N' insertions are performed, a single quadratically
1050 probed hash table is allocated and grows as needed, providing extremely
1051 efficient access (constant time insertion/deleting/queries with low constant
1052 factors) and is very stingy with malloc traffic.
1054 Note that, unlike ``std::set``, the iterators of ``SmallPtrSet`` are invalidated
1055 whenever an insertion occurs. Also, the values visited by the iterators are not
1056 visited in sorted order.
1063 DenseSet is a simple quadratically probed hash table. It excels at supporting
1064 small values: it uses a single allocation to hold all of the pairs that are
1065 currently inserted in the set. DenseSet is a great way to unique small values
1066 that are not simple pointers (use :ref:`SmallPtrSet <dss_smallptrset>` for
1067 pointers). Note that DenseSet has the same requirements for the value type that
1068 :ref:`DenseMap <dss_densemap>` has.
1072 llvm/ADT/SparseSet.h
1073 ^^^^^^^^^^^^^^^^^^^^
1075 SparseSet holds a small number of objects identified by unsigned keys of
1076 moderate size. It uses a lot of memory, but provides operations that are almost
1077 as fast as a vector. Typical keys are physical registers, virtual registers, or
1078 numbered basic blocks.
1080 SparseSet is useful for algorithms that need very fast clear/find/insert/erase
1081 and fast iteration over small sets. It is not intended for building composite
1084 .. _dss_sparsemultiset:
1086 llvm/ADT/SparseMultiSet.h
1087 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1089 SparseMultiSet adds multiset behavior to SparseSet, while retaining SparseSet's
1090 desirable attributes. Like SparseSet, it typically uses a lot of memory, but
1091 provides operations that are almost as fast as a vector. Typical keys are
1092 physical registers, virtual registers, or numbered basic blocks.
1094 SparseMultiSet is useful for algorithms that need very fast
1095 clear/find/insert/erase of the entire collection, and iteration over sets of
1096 elements sharing a key. It is often a more efficient choice than using composite
1097 data structures (e.g. vector-of-vectors, map-of-vectors). It is not intended for
1098 building composite data structures.
1102 llvm/ADT/FoldingSet.h
1103 ^^^^^^^^^^^^^^^^^^^^^
1105 FoldingSet is an aggregate class that is really good at uniquing
1106 expensive-to-create or polymorphic objects. It is a combination of a chained
1107 hash table with intrusive links (uniqued objects are required to inherit from
1108 FoldingSetNode) that uses :ref:`SmallVector <dss_smallvector>` as part of its ID
1111 Consider a case where you want to implement a "getOrCreateFoo" method for a
1112 complex object (for example, a node in the code generator). The client has a
1113 description of **what** it wants to generate (it knows the opcode and all the
1114 operands), but we don't want to 'new' a node, then try inserting it into a set
1115 only to find out it already exists, at which point we would have to delete it
1116 and return the node that already exists.
1118 To support this style of client, FoldingSet perform a query with a
1119 FoldingSetNodeID (which wraps SmallVector) that can be used to describe the
1120 element that we want to query for. The query either returns the element
1121 matching the ID or it returns an opaque ID that indicates where insertion should
1122 take place. Construction of the ID usually does not require heap traffic.
1124 Because FoldingSet uses intrusive links, it can support polymorphic objects in
1125 the set (for example, you can have SDNode instances mixed with LoadSDNodes).
1126 Because the elements are individually allocated, pointers to the elements are
1127 stable: inserting or removing elements does not invalidate any pointers to other
1135 ``std::set`` is a reasonable all-around set class, which is decent at many
1136 things but great at nothing. std::set allocates memory for each element
1137 inserted (thus it is very malloc intensive) and typically stores three pointers
1138 per element in the set (thus adding a large amount of per-element space
1139 overhead). It offers guaranteed log(n) performance, which is not particularly
1140 fast from a complexity standpoint (particularly if the elements of the set are
1141 expensive to compare, like strings), and has extremely high constant factors for
1142 lookup, insertion and removal.
1144 The advantages of std::set are that its iterators are stable (deleting or
1145 inserting an element from the set does not affect iterators or pointers to other
1146 elements) and that iteration over the set is guaranteed to be in sorted order.
1147 If the elements in the set are large, then the relative overhead of the pointers
1148 and malloc traffic is not a big deal, but if the elements of the set are small,
1149 std::set is almost never a good choice.
1153 llvm/ADT/SetVector.h
1154 ^^^^^^^^^^^^^^^^^^^^
1156 LLVM's ``SetVector<Type>`` is an adapter class that combines your choice of a
1157 set-like container along with a :ref:`Sequential Container <ds_sequential>` The
1158 important property that this provides is efficient insertion with uniquing
1159 (duplicate elements are ignored) with iteration support. It implements this by
1160 inserting elements into both a set-like container and the sequential container,
1161 using the set-like container for uniquing and the sequential container for
1164 The difference between SetVector and other sets is that the order of iteration
1165 is guaranteed to match the order of insertion into the SetVector. This property
1166 is really important for things like sets of pointers. Because pointer values
1167 are non-deterministic (e.g. vary across runs of the program on different
1168 machines), iterating over the pointers in the set will not be in a well-defined
1171 The drawback of SetVector is that it requires twice as much space as a normal
1172 set and has the sum of constant factors from the set-like container and the
1173 sequential container that it uses. Use it **only** if you need to iterate over
1174 the elements in a deterministic order. SetVector is also expensive to delete
1175 elements out of (linear time), unless you use its "pop_back" method, which is
1178 ``SetVector`` is an adapter class that defaults to using ``std::vector`` and a
1179 size 16 ``SmallSet`` for the underlying containers, so it is quite expensive.
1180 However, ``"llvm/ADT/SetVector.h"`` also provides a ``SmallSetVector`` class,
1181 which defaults to using a ``SmallVector`` and ``SmallSet`` of a specified size.
1182 If you use this, and if your sets are dynamically smaller than ``N``, you will
1183 save a lot of heap traffic.
1185 .. _dss_uniquevector:
1187 llvm/ADT/UniqueVector.h
1188 ^^^^^^^^^^^^^^^^^^^^^^^
1190 UniqueVector is similar to :ref:`SetVector <dss_setvector>` but it retains a
1191 unique ID for each element inserted into the set. It internally contains a map
1192 and a vector, and it assigns a unique ID for each value inserted into the set.
1194 UniqueVector is very expensive: its cost is the sum of the cost of maintaining
1195 both the map and vector, it has high complexity, high constant factors, and
1196 produces a lot of malloc traffic. It should be avoided.
1198 .. _dss_immutableset:
1200 llvm/ADT/ImmutableSet.h
1201 ^^^^^^^^^^^^^^^^^^^^^^^
1203 ImmutableSet is an immutable (functional) set implementation based on an AVL
1204 tree. Adding or removing elements is done through a Factory object and results
1205 in the creation of a new ImmutableSet object. If an ImmutableSet already exists
1206 with the given contents, then the existing one is returned; equality is compared
1207 with a FoldingSetNodeID. The time and space complexity of add or remove
1208 operations is logarithmic in the size of the original set.
1210 There is no method for returning an element of the set, you can only check for
1215 Other Set-Like Container Options
1216 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1218 The STL provides several other options, such as std::multiset and the various
1219 "hash_set" like containers (whether from C++ TR1 or from the SGI library). We
1220 never use hash_set and unordered_set because they are generally very expensive
1221 (each insertion requires a malloc) and very non-portable.
1223 std::multiset is useful if you're not interested in elimination of duplicates,
1224 but has all the drawbacks of std::set. A sorted vector (where you don't delete
1225 duplicate entries) or some other approach is almost always better.
1229 Map-Like Containers (std::map, DenseMap, etc)
1230 ---------------------------------------------
1232 Map-like containers are useful when you want to associate data to a key. As
1233 usual, there are a lot of different ways to do this. :)
1235 .. _dss_sortedvectormap:
1240 If your usage pattern follows a strict insert-then-query approach, you can
1241 trivially use the same approach as :ref:`sorted vectors for set-like containers
1242 <dss_sortedvectorset>`. The only difference is that your query function (which
1243 uses std::lower_bound to get efficient log(n) lookup) should only compare the
1244 key, not both the key and value. This yields the same advantages as sorted
1249 llvm/ADT/StringMap.h
1250 ^^^^^^^^^^^^^^^^^^^^
1252 Strings are commonly used as keys in maps, and they are difficult to support
1253 efficiently: they are variable length, inefficient to hash and compare when
1254 long, expensive to copy, etc. StringMap is a specialized container designed to
1255 cope with these issues. It supports mapping an arbitrary range of bytes to an
1256 arbitrary other object.
1258 The StringMap implementation uses a quadratically-probed hash table, where the
1259 buckets store a pointer to the heap allocated entries (and some other stuff).
1260 The entries in the map must be heap allocated because the strings are variable
1261 length. The string data (key) and the element object (value) are stored in the
1262 same allocation with the string data immediately after the element object.
1263 This container guarantees the "``(char*)(&Value+1)``" points to the key string
1266 The StringMap is very fast for several reasons: quadratic probing is very cache
1267 efficient for lookups, the hash value of strings in buckets is not recomputed
1268 when looking up an element, StringMap rarely has to touch the memory for
1269 unrelated objects when looking up a value (even when hash collisions happen),
1270 hash table growth does not recompute the hash values for strings already in the
1271 table, and each pair in the map is store in a single allocation (the string data
1272 is stored in the same allocation as the Value of a pair).
1274 StringMap also provides query methods that take byte ranges, so it only ever
1275 copies a string if a value is inserted into the table.
1277 StringMap iteratation order, however, is not guaranteed to be deterministic, so
1278 any uses which require that should instead use a std::map.
1282 llvm/ADT/IndexedMap.h
1283 ^^^^^^^^^^^^^^^^^^^^^
1285 IndexedMap is a specialized container for mapping small dense integers (or
1286 values that can be mapped to small dense integers) to some other type. It is
1287 internally implemented as a vector with a mapping function that maps the keys
1288 to the dense integer range.
1290 This is useful for cases like virtual registers in the LLVM code generator: they
1291 have a dense mapping that is offset by a compile-time constant (the first
1292 virtual register ID).
1299 DenseMap is a simple quadratically probed hash table. It excels at supporting
1300 small keys and values: it uses a single allocation to hold all of the pairs
1301 that are currently inserted in the map. DenseMap is a great way to map
1302 pointers to pointers, or map other small types to each other.
1304 There are several aspects of DenseMap that you should be aware of, however.
1305 The iterators in a DenseMap are invalidated whenever an insertion occurs,
1306 unlike map. Also, because DenseMap allocates space for a large number of
1307 key/value pairs (it starts with 64 by default), it will waste a lot of space if
1308 your keys or values are large. Finally, you must implement a partial
1309 specialization of DenseMapInfo for the key that you want, if it isn't already
1310 supported. This is required to tell DenseMap about two special marker values
1311 (which can never be inserted into the map) that it needs internally.
1313 DenseMap's find_as() method supports lookup operations using an alternate key
1314 type. This is useful in cases where the normal key type is expensive to
1315 construct, but cheap to compare against. The DenseMapInfo is responsible for
1316 defining the appropriate comparison and hashing methods for each alternate key
1324 ValueMap is a wrapper around a :ref:`DenseMap <dss_densemap>` mapping
1325 ``Value*``\ s (or subclasses) to another type. When a Value is deleted or
1326 RAUW'ed, ValueMap will update itself so the new version of the key is mapped to
1327 the same value, just as if the key were a WeakVH. You can configure exactly how
1328 this happens, and what else happens on these two events, by passing a ``Config``
1329 parameter to the ValueMap template.
1331 .. _dss_intervalmap:
1333 llvm/ADT/IntervalMap.h
1334 ^^^^^^^^^^^^^^^^^^^^^^
1336 IntervalMap is a compact map for small keys and values. It maps key intervals
1337 instead of single keys, and it will automatically coalesce adjacent intervals.
1338 When then map only contains a few intervals, they are stored in the map object
1339 itself to avoid allocations.
1341 The IntervalMap iterators are quite big, so they should not be passed around as
1342 STL iterators. The heavyweight iterators allow a smaller data structure.
1349 std::map has similar characteristics to :ref:`std::set <dss_set>`: it uses a
1350 single allocation per pair inserted into the map, it offers log(n) lookup with
1351 an extremely large constant factor, imposes a space penalty of 3 pointers per
1352 pair in the map, etc.
1354 std::map is most useful when your keys or values are very large, if you need to
1355 iterate over the collection in sorted order, or if you need stable iterators
1356 into the map (i.e. they don't get invalidated if an insertion or deletion of
1357 another element takes place).
1361 llvm/ADT/MapVector.h
1362 ^^^^^^^^^^^^^^^^^^^^
1364 ``MapVector<KeyT,ValueT>`` provides a subset of the DenseMap interface. The
1365 main difference is that the iteration order is guaranteed to be the insertion
1366 order, making it an easy (but somewhat expensive) solution for non-deterministic
1367 iteration over maps of pointers.
1369 It is implemented by mapping from key to an index in a vector of key,value
1370 pairs. This provides fast lookup and iteration, but has two main drawbacks: The
1371 key is stored twice and it doesn't support removing elements.
1373 .. _dss_inteqclasses:
1375 llvm/ADT/IntEqClasses.h
1376 ^^^^^^^^^^^^^^^^^^^^^^^
1378 IntEqClasses provides a compact representation of equivalence classes of small
1379 integers. Initially, each integer in the range 0..n-1 has its own equivalence
1380 class. Classes can be joined by passing two class representatives to the
1381 join(a, b) method. Two integers are in the same class when findLeader() returns
1382 the same representative.
1384 Once all equivalence classes are formed, the map can be compressed so each
1385 integer 0..n-1 maps to an equivalence class number in the range 0..m-1, where m
1386 is the total number of equivalence classes. The map must be uncompressed before
1387 it can be edited again.
1389 .. _dss_immutablemap:
1391 llvm/ADT/ImmutableMap.h
1392 ^^^^^^^^^^^^^^^^^^^^^^^
1394 ImmutableMap is an immutable (functional) map implementation based on an AVL
1395 tree. Adding or removing elements is done through a Factory object and results
1396 in the creation of a new ImmutableMap object. If an ImmutableMap already exists
1397 with the given key set, then the existing one is returned; equality is compared
1398 with a FoldingSetNodeID. The time and space complexity of add or remove
1399 operations is logarithmic in the size of the original map.
1403 Other Map-Like Container Options
1404 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1406 The STL provides several other options, such as std::multimap and the various
1407 "hash_map" like containers (whether from C++ TR1 or from the SGI library). We
1408 never use hash_set and unordered_set because they are generally very expensive
1409 (each insertion requires a malloc) and very non-portable.
1411 std::multimap is useful if you want to map a key to multiple values, but has all
1412 the drawbacks of std::map. A sorted vector or some other approach is almost
1417 Bit storage containers (BitVector, SparseBitVector)
1418 ---------------------------------------------------
1420 Unlike the other containers, there are only two bit storage containers, and
1421 choosing when to use each is relatively straightforward.
1423 One additional option is ``std::vector<bool>``: we discourage its use for two
1424 reasons 1) the implementation in many common compilers (e.g. commonly
1425 available versions of GCC) is extremely inefficient and 2) the C++ standards
1426 committee is likely to deprecate this container and/or change it significantly
1427 somehow. In any case, please don't use it.
1434 The BitVector container provides a dynamic size set of bits for manipulation.
1435 It supports individual bit setting/testing, as well as set operations. The set
1436 operations take time O(size of bitvector), but operations are performed one word
1437 at a time, instead of one bit at a time. This makes the BitVector very fast for
1438 set operations compared to other containers. Use the BitVector when you expect
1439 the number of set bits to be high (i.e. a dense set).
1441 .. _dss_smallbitvector:
1446 The SmallBitVector container provides the same interface as BitVector, but it is
1447 optimized for the case where only a small number of bits, less than 25 or so,
1448 are needed. It also transparently supports larger bit counts, but slightly less
1449 efficiently than a plain BitVector, so SmallBitVector should only be used when
1450 larger counts are rare.
1452 At this time, SmallBitVector does not support set operations (and, or, xor), and
1453 its operator[] does not provide an assignable lvalue.
1455 .. _dss_sparsebitvector:
1460 The SparseBitVector container is much like BitVector, with one major difference:
1461 Only the bits that are set, are stored. This makes the SparseBitVector much
1462 more space efficient than BitVector when the set is sparse, as well as making
1463 set operations O(number of set bits) instead of O(size of universe). The
1464 downside to the SparseBitVector is that setting and testing of random bits is
1465 O(N), and on large SparseBitVectors, this can be slower than BitVector. In our
1466 implementation, setting or testing bits in sorted order (either forwards or
1467 reverse) is O(1) worst case. Testing and setting bits within 128 bits (depends
1468 on size) of the current bit is also O(1). As a general statement,
1469 testing/setting bits in a SparseBitVector is O(distance away from last set bit).
1473 Helpful Hints for Common Operations
1474 ===================================
1476 This section describes how to perform some very simple transformations of LLVM
1477 code. This is meant to give examples of common idioms used, showing the
1478 practical side of LLVM transformations.
1480 Because this is a "how-to" section, you should also read about the main classes
1481 that you will be working with. The :ref:`Core LLVM Class Hierarchy Reference
1482 <coreclasses>` contains details and descriptions of the main classes that you
1487 Basic Inspection and Traversal Routines
1488 ---------------------------------------
1490 The LLVM compiler infrastructure have many different data structures that may be
1491 traversed. Following the example of the C++ standard template library, the
1492 techniques used to traverse these various data structures are all basically the
1493 same. For a enumerable sequence of values, the ``XXXbegin()`` function (or
1494 method) returns an iterator to the start of the sequence, the ``XXXend()``
1495 function returns an iterator pointing to one past the last valid element of the
1496 sequence, and there is some ``XXXiterator`` data type that is common between the
1499 Because the pattern for iteration is common across many different aspects of the
1500 program representation, the standard template library algorithms may be used on
1501 them, and it is easier to remember how to iterate. First we show a few common
1502 examples of the data structures that need to be traversed. Other data
1503 structures are traversed in very similar ways.
1505 .. _iterate_function:
1507 Iterating over the ``BasicBlock`` in a ``Function``
1508 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1510 It's quite common to have a ``Function`` instance that you'd like to transform
1511 in some way; in particular, you'd like to manipulate its ``BasicBlock``\ s. To
1512 facilitate this, you'll need to iterate over all of the ``BasicBlock``\ s that
1513 constitute the ``Function``. The following is an example that prints the name
1514 of a ``BasicBlock`` and the number of ``Instruction``\ s it contains:
1518 // func is a pointer to a Function instance
1519 for (Function::iterator i = func->begin(), e = func->end(); i != e; ++i)
1520 // Print out the name of the basic block if it has one, and then the
1521 // number of instructions that it contains
1522 errs() << "Basic block (name=" << i->getName() << ") has "
1523 << i->size() << " instructions.\n";
1525 Note that i can be used as if it were a pointer for the purposes of invoking
1526 member functions of the ``Instruction`` class. This is because the indirection
1527 operator is overloaded for the iterator classes. In the above code, the
1528 expression ``i->size()`` is exactly equivalent to ``(*i).size()`` just like
1531 .. _iterate_basicblock:
1533 Iterating over the ``Instruction`` in a ``BasicBlock``
1534 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1536 Just like when dealing with ``BasicBlock``\ s in ``Function``\ s, it's easy to
1537 iterate over the individual instructions that make up ``BasicBlock``\ s. Here's
1538 a code snippet that prints out each instruction in a ``BasicBlock``:
1542 // blk is a pointer to a BasicBlock instance
1543 for (BasicBlock::iterator i = blk->begin(), e = blk->end(); i != e; ++i)
1544 // The next statement works since operator<<(ostream&,...)
1545 // is overloaded for Instruction&
1546 errs() << *i << "\n";
1549 However, this isn't really the best way to print out the contents of a
1550 ``BasicBlock``! Since the ostream operators are overloaded for virtually
1551 anything you'll care about, you could have just invoked the print routine on the
1552 basic block itself: ``errs() << *blk << "\n";``.
1554 .. _iterate_insiter:
1556 Iterating over the ``Instruction`` in a ``Function``
1557 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1559 If you're finding that you commonly iterate over a ``Function``'s
1560 ``BasicBlock``\ s and then that ``BasicBlock``'s ``Instruction``\ s,
1561 ``InstIterator`` should be used instead. You'll need to include
1562 ``llvm/Support/InstIterator.h`` (`doxygen
1563 <http://llvm.org/doxygen/InstIterator_8h-source.html>`__) and then instantiate
1564 ``InstIterator``\ s explicitly in your code. Here's a small example that shows
1565 how to dump all instructions in a function to the standard error stream:
1569 #include "llvm/Support/InstIterator.h"
1571 // F is a pointer to a Function instance
1572 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
1573 errs() << *I << "\n";
1575 Easy, isn't it? You can also use ``InstIterator``\ s to fill a work list with
1576 its initial contents. For example, if you wanted to initialize a work list to
1577 contain all instructions in a ``Function`` F, all you would need to do is
1582 std::set<Instruction*> worklist;
1583 // or better yet, SmallPtrSet<Instruction*, 64> worklist;
1585 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
1586 worklist.insert(&*I);
1588 The STL set ``worklist`` would now contain all instructions in the ``Function``
1591 .. _iterate_convert:
1593 Turning an iterator into a class pointer (and vice-versa)
1594 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1596 Sometimes, it'll be useful to grab a reference (or pointer) to a class instance
1597 when all you've got at hand is an iterator. Well, extracting a reference or a
1598 pointer from an iterator is very straight-forward. Assuming that ``i`` is a
1599 ``BasicBlock::iterator`` and ``j`` is a ``BasicBlock::const_iterator``:
1603 Instruction& inst = *i; // Grab reference to instruction reference
1604 Instruction* pinst = &*i; // Grab pointer to instruction reference
1605 const Instruction& inst = *j;
1607 However, the iterators you'll be working with in the LLVM framework are special:
1608 they will automatically convert to a ptr-to-instance type whenever they need to.
1609 Instead of derferencing the iterator and then taking the address of the result,
1610 you can simply assign the iterator to the proper pointer type and you get the
1611 dereference and address-of operation as a result of the assignment (behind the
1612 scenes, this is a result of overloading casting mechanisms). Thus the last line
1613 of the last example,
1617 Instruction *pinst = &*i;
1619 is semantically equivalent to
1623 Instruction *pinst = i;
1625 It's also possible to turn a class pointer into the corresponding iterator, and
1626 this is a constant time operation (very efficient). The following code snippet
1627 illustrates use of the conversion constructors provided by LLVM iterators. By
1628 using these, you can explicitly grab the iterator of something without actually
1629 obtaining it via iteration over some structure:
1633 void printNextInstruction(Instruction* inst) {
1634 BasicBlock::iterator it(inst);
1635 ++it; // After this line, it refers to the instruction after *inst
1636 if (it != inst->getParent()->end()) errs() << *it << "\n";
1639 Unfortunately, these implicit conversions come at a cost; they prevent these
1640 iterators from conforming to standard iterator conventions, and thus from being
1641 usable with standard algorithms and containers. For example, they prevent the
1642 following code, where ``B`` is a ``BasicBlock``, from compiling:
1646 llvm::SmallVector<llvm::Instruction *, 16>(B->begin(), B->end());
1648 Because of this, these implicit conversions may be removed some day, and
1649 ``operator*`` changed to return a pointer instead of a reference.
1651 .. _iterate_complex:
1653 Finding call sites: a slightly more complex example
1654 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1656 Say that you're writing a FunctionPass and would like to count all the locations
1657 in the entire module (that is, across every ``Function``) where a certain
1658 function (i.e., some ``Function *``) is already in scope. As you'll learn
1659 later, you may want to use an ``InstVisitor`` to accomplish this in a much more
1660 straight-forward manner, but this example will allow us to explore how you'd do
1661 it if you didn't have ``InstVisitor`` around. In pseudo-code, this is what we
1664 .. code-block:: none
1666 initialize callCounter to zero
1667 for each Function f in the Module
1668 for each BasicBlock b in f
1669 for each Instruction i in b
1670 if (i is a CallInst and calls the given function)
1671 increment callCounter
1673 And the actual code is (remember, because we're writing a ``FunctionPass``, our
1674 ``FunctionPass``-derived class simply has to override the ``runOnFunction``
1679 Function* targetFunc = ...;
1681 class OurFunctionPass : public FunctionPass {
1683 OurFunctionPass(): callCounter(0) { }
1685 virtual runOnFunction(Function& F) {
1686 for (Function::iterator b = F.begin(), be = F.end(); b != be; ++b) {
1687 for (BasicBlock::iterator i = b->begin(), ie = b->end(); i != ie; ++i) {
1688 if (CallInst* callInst = dyn_cast<CallInst>(&*i)) {
1689 // We know we've encountered a call instruction, so we
1690 // need to determine if it's a call to the
1691 // function pointed to by m_func or not.
1692 if (callInst->getCalledFunction() == targetFunc)
1700 unsigned callCounter;
1703 .. _calls_and_invokes:
1705 Treating calls and invokes the same way
1706 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1708 You may have noticed that the previous example was a bit oversimplified in that
1709 it did not deal with call sites generated by 'invoke' instructions. In this,
1710 and in other situations, you may find that you want to treat ``CallInst``\ s and
1711 ``InvokeInst``\ s the same way, even though their most-specific common base
1712 class is ``Instruction``, which includes lots of less closely-related things.
1713 For these cases, LLVM provides a handy wrapper class called ``CallSite``
1714 (`doxygen <http://llvm.org/doxygen/classllvm_1_1CallSite.html>`__) It is
1715 essentially a wrapper around an ``Instruction`` pointer, with some methods that
1716 provide functionality common to ``CallInst``\ s and ``InvokeInst``\ s.
1718 This class has "value semantics": it should be passed by value, not by reference
1719 and it should not be dynamically allocated or deallocated using ``operator new``
1720 or ``operator delete``. It is efficiently copyable, assignable and
1721 constructable, with costs equivalents to that of a bare pointer. If you look at
1722 its definition, it has only a single pointer member.
1726 Iterating over def-use & use-def chains
1727 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1729 Frequently, we might have an instance of the ``Value`` class (`doxygen
1730 <http://llvm.org/doxygen/classllvm_1_1Value.html>`__) and we want to determine
1731 which ``User`` s use the ``Value``. The list of all ``User``\ s of a particular
1732 ``Value`` is called a *def-use* chain. For example, let's say we have a
1733 ``Function*`` named ``F`` to a particular function ``foo``. Finding all of the
1734 instructions that *use* ``foo`` is as simple as iterating over the *def-use*
1741 for (Value::use_iterator i = F->use_begin(), e = F->use_end(); i != e; ++i)
1742 if (Instruction *Inst = dyn_cast<Instruction>(*i)) {
1743 errs() << "F is used in instruction:\n";
1744 errs() << *Inst << "\n";
1747 Note that dereferencing a ``Value::use_iterator`` is not a very cheap operation.
1748 Instead of performing ``*i`` above several times, consider doing it only once in
1749 the loop body and reusing its result.
1751 Alternatively, it's common to have an instance of the ``User`` Class (`doxygen
1752 <http://llvm.org/doxygen/classllvm_1_1User.html>`__) and need to know what
1753 ``Value``\ s are used by it. The list of all ``Value``\ s used by a ``User`` is
1754 known as a *use-def* chain. Instances of class ``Instruction`` are common
1755 ``User`` s, so we might want to iterate over all of the values that a particular
1756 instruction uses (that is, the operands of the particular ``Instruction``):
1760 Instruction *pi = ...;
1762 for (User::op_iterator i = pi->op_begin(), e = pi->op_end(); i != e; ++i) {
1767 Declaring objects as ``const`` is an important tool of enforcing mutation free
1768 algorithms (such as analyses, etc.). For this purpose above iterators come in
1769 constant flavors as ``Value::const_use_iterator`` and
1770 ``Value::const_op_iterator``. They automatically arise when calling
1771 ``use/op_begin()`` on ``const Value*``\ s or ``const User*``\ s respectively.
1772 Upon dereferencing, they return ``const Use*``\ s. Otherwise the above patterns
1777 Iterating over predecessors & successors of blocks
1778 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1780 Iterating over the predecessors and successors of a block is quite easy with the
1781 routines defined in ``"llvm/Support/CFG.h"``. Just use code like this to
1782 iterate over all predecessors of BB:
1786 #include "llvm/Support/CFG.h"
1787 BasicBlock *BB = ...;
1789 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
1790 BasicBlock *Pred = *PI;
1794 Similarly, to iterate over successors use ``succ_iterator/succ_begin/succ_end``.
1798 Making simple changes
1799 ---------------------
1801 There are some primitive transformation operations present in the LLVM
1802 infrastructure that are worth knowing about. When performing transformations,
1803 it's fairly common to manipulate the contents of basic blocks. This section
1804 describes some of the common methods for doing so and gives example code.
1806 .. _schanges_creating:
1808 Creating and inserting new ``Instruction``\ s
1809 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1811 *Instantiating Instructions*
1813 Creation of ``Instruction``\ s is straight-forward: simply call the constructor
1814 for the kind of instruction to instantiate and provide the necessary parameters.
1815 For example, an ``AllocaInst`` only *requires* a (const-ptr-to) ``Type``. Thus:
1819 AllocaInst* ai = new AllocaInst(Type::Int32Ty);
1821 will create an ``AllocaInst`` instance that represents the allocation of one
1822 integer in the current stack frame, at run time. Each ``Instruction`` subclass
1823 is likely to have varying default parameters which change the semantics of the
1824 instruction, so refer to the `doxygen documentation for the subclass of
1825 Instruction <http://llvm.org/doxygen/classllvm_1_1Instruction.html>`_ that
1826 you're interested in instantiating.
1830 It is very useful to name the values of instructions when you're able to, as
1831 this facilitates the debugging of your transformations. If you end up looking
1832 at generated LLVM machine code, you definitely want to have logical names
1833 associated with the results of instructions! By supplying a value for the
1834 ``Name`` (default) parameter of the ``Instruction`` constructor, you associate a
1835 logical name with the result of the instruction's execution at run time. For
1836 example, say that I'm writing a transformation that dynamically allocates space
1837 for an integer on the stack, and that integer is going to be used as some kind
1838 of index by some other code. To accomplish this, I place an ``AllocaInst`` at
1839 the first point in the first ``BasicBlock`` of some ``Function``, and I'm
1840 intending to use it within the same ``Function``. I might do:
1844 AllocaInst* pa = new AllocaInst(Type::Int32Ty, 0, "indexLoc");
1846 where ``indexLoc`` is now the logical name of the instruction's execution value,
1847 which is a pointer to an integer on the run time stack.
1849 *Inserting instructions*
1851 There are essentially two ways to insert an ``Instruction`` into an existing
1852 sequence of instructions that form a ``BasicBlock``:
1854 * Insertion into an explicit instruction list
1856 Given a ``BasicBlock* pb``, an ``Instruction* pi`` within that ``BasicBlock``,
1857 and a newly-created instruction we wish to insert before ``*pi``, we do the
1862 BasicBlock *pb = ...;
1863 Instruction *pi = ...;
1864 Instruction *newInst = new Instruction(...);
1866 pb->getInstList().insert(pi, newInst); // Inserts newInst before pi in pb
1868 Appending to the end of a ``BasicBlock`` is so common that the ``Instruction``
1869 class and ``Instruction``-derived classes provide constructors which take a
1870 pointer to a ``BasicBlock`` to be appended to. For example code that looked
1875 BasicBlock *pb = ...;
1876 Instruction *newInst = new Instruction(...);
1878 pb->getInstList().push_back(newInst); // Appends newInst to pb
1884 BasicBlock *pb = ...;
1885 Instruction *newInst = new Instruction(..., pb);
1887 which is much cleaner, especially if you are creating long instruction
1890 * Insertion into an implicit instruction list
1892 ``Instruction`` instances that are already in ``BasicBlock``\ s are implicitly
1893 associated with an existing instruction list: the instruction list of the
1894 enclosing basic block. Thus, we could have accomplished the same thing as the
1895 above code without being given a ``BasicBlock`` by doing:
1899 Instruction *pi = ...;
1900 Instruction *newInst = new Instruction(...);
1902 pi->getParent()->getInstList().insert(pi, newInst);
1904 In fact, this sequence of steps occurs so frequently that the ``Instruction``
1905 class and ``Instruction``-derived classes provide constructors which take (as
1906 a default parameter) a pointer to an ``Instruction`` which the newly-created
1907 ``Instruction`` should precede. That is, ``Instruction`` constructors are
1908 capable of inserting the newly-created instance into the ``BasicBlock`` of a
1909 provided instruction, immediately before that instruction. Using an
1910 ``Instruction`` constructor with a ``insertBefore`` (default) parameter, the
1915 Instruction* pi = ...;
1916 Instruction* newInst = new Instruction(..., pi);
1918 which is much cleaner, especially if you're creating a lot of instructions and
1919 adding them to ``BasicBlock``\ s.
1921 .. _schanges_deleting:
1923 Deleting Instructions
1924 ^^^^^^^^^^^^^^^^^^^^^
1926 Deleting an instruction from an existing sequence of instructions that form a
1927 BasicBlock_ is very straight-forward: just call the instruction's
1928 ``eraseFromParent()`` method. For example:
1932 Instruction *I = .. ;
1933 I->eraseFromParent();
1935 This unlinks the instruction from its containing basic block and deletes it. If
1936 you'd just like to unlink the instruction from its containing basic block but
1937 not delete it, you can use the ``removeFromParent()`` method.
1939 .. _schanges_replacing:
1941 Replacing an Instruction with another Value
1942 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1944 Replacing individual instructions
1945 """""""""""""""""""""""""""""""""
1947 Including "`llvm/Transforms/Utils/BasicBlockUtils.h
1948 <http://llvm.org/doxygen/BasicBlockUtils_8h-source.html>`_" permits use of two
1949 very useful replace functions: ``ReplaceInstWithValue`` and
1950 ``ReplaceInstWithInst``.
1952 .. _schanges_deleting_sub:
1954 Deleting Instructions
1955 """""""""""""""""""""
1957 * ``ReplaceInstWithValue``
1959 This function replaces all uses of a given instruction with a value, and then
1960 removes the original instruction. The following example illustrates the
1961 replacement of the result of a particular ``AllocaInst`` that allocates memory
1962 for a single integer with a null pointer to an integer.
1966 AllocaInst* instToReplace = ...;
1967 BasicBlock::iterator ii(instToReplace);
1969 ReplaceInstWithValue(instToReplace->getParent()->getInstList(), ii,
1970 Constant::getNullValue(PointerType::getUnqual(Type::Int32Ty)));
1972 * ``ReplaceInstWithInst``
1974 This function replaces a particular instruction with another instruction,
1975 inserting the new instruction into the basic block at the location where the
1976 old instruction was, and replacing any uses of the old instruction with the
1977 new instruction. The following example illustrates the replacement of one
1978 ``AllocaInst`` with another.
1982 AllocaInst* instToReplace = ...;
1983 BasicBlock::iterator ii(instToReplace);
1985 ReplaceInstWithInst(instToReplace->getParent()->getInstList(), ii,
1986 new AllocaInst(Type::Int32Ty, 0, "ptrToReplacedInt"));
1989 Replacing multiple uses of Users and Values
1990 """""""""""""""""""""""""""""""""""""""""""
1992 You can use ``Value::replaceAllUsesWith`` and ``User::replaceUsesOfWith`` to
1993 change more than one use at a time. See the doxygen documentation for the
1994 `Value Class <http://llvm.org/doxygen/classllvm_1_1Value.html>`_ and `User Class
1995 <http://llvm.org/doxygen/classllvm_1_1User.html>`_, respectively, for more
1998 .. _schanges_deletingGV:
2000 Deleting GlobalVariables
2001 ^^^^^^^^^^^^^^^^^^^^^^^^
2003 Deleting a global variable from a module is just as easy as deleting an
2004 Instruction. First, you must have a pointer to the global variable that you
2005 wish to delete. You use this pointer to erase it from its parent, the module.
2010 GlobalVariable *GV = .. ;
2012 GV->eraseFromParent();
2020 In generating IR, you may need some complex types. If you know these types
2021 statically, you can use ``TypeBuilder<...>::get()``, defined in
2022 ``llvm/Support/TypeBuilder.h``, to retrieve them. ``TypeBuilder`` has two forms
2023 depending on whether you're building types for cross-compilation or native
2024 library use. ``TypeBuilder<T, true>`` requires that ``T`` be independent of the
2025 host environment, meaning that it's built out of types from the ``llvm::types``
2026 (`doxygen <http://llvm.org/doxygen/namespacellvm_1_1types.html>`__) namespace
2027 and pointers, functions, arrays, etc. built of those. ``TypeBuilder<T, false>``
2028 additionally allows native C types whose size may depend on the host compiler.
2033 FunctionType *ft = TypeBuilder<types::i<8>(types::i<32>*), true>::get();
2035 is easier to read and write than the equivalent
2039 std::vector<const Type*> params;
2040 params.push_back(PointerType::getUnqual(Type::Int32Ty));
2041 FunctionType *ft = FunctionType::get(Type::Int8Ty, params, false);
2043 See the `class comment
2044 <http://llvm.org/doxygen/TypeBuilder_8h-source.html#l00001>`_ for more details.
2051 This section describes the interaction of the LLVM APIs with multithreading,
2052 both on the part of client applications, and in the JIT, in the hosted
2055 Note that LLVM's support for multithreading is still relatively young. Up
2056 through version 2.5, the execution of threaded hosted applications was
2057 supported, but not threaded client access to the APIs. While this use case is
2058 now supported, clients *must* adhere to the guidelines specified below to ensure
2059 proper operation in multithreaded mode.
2061 Note that, on Unix-like platforms, LLVM requires the presence of GCC's atomic
2062 intrinsics in order to support threaded operation. If you need a
2063 multhreading-capable LLVM on a platform without a suitably modern system
2064 compiler, consider compiling LLVM and LLVM-GCC in single-threaded mode, and
2065 using the resultant compiler to build a copy of LLVM with multithreading
2068 .. _startmultithreaded:
2070 Entering and Exiting Multithreaded Mode
2071 ---------------------------------------
2073 In order to properly protect its internal data structures while avoiding
2074 excessive locking overhead in the single-threaded case, the LLVM must intialize
2075 certain data structures necessary to provide guards around its internals. To do
2076 so, the client program must invoke ``llvm_start_multithreaded()`` before making
2077 any concurrent LLVM API calls. To subsequently tear down these structures, use
2078 the ``llvm_stop_multithreaded()`` call. You can also use the
2079 ``llvm_is_multithreaded()`` call to check the status of multithreaded mode.
2081 Note that both of these calls must be made *in isolation*. That is to say that
2082 no other LLVM API calls may be executing at any time during the execution of
2083 ``llvm_start_multithreaded()`` or ``llvm_stop_multithreaded``. It is the
2084 client's responsibility to enforce this isolation.
2086 The return value of ``llvm_start_multithreaded()`` indicates the success or
2087 failure of the initialization. Failure typically indicates that your copy of
2088 LLVM was built without multithreading support, typically because GCC atomic
2089 intrinsics were not found in your system compiler. In this case, the LLVM API
2090 will not be safe for concurrent calls. However, it *will* be safe for hosting
2091 threaded applications in the JIT, though :ref:`care must be taken
2092 <jitthreading>` to ensure that side exits and the like do not accidentally
2093 result in concurrent LLVM API calls.
2097 Ending Execution with ``llvm_shutdown()``
2098 -----------------------------------------
2100 When you are done using the LLVM APIs, you should call ``llvm_shutdown()`` to
2101 deallocate memory used for internal structures. This will also invoke
2102 ``llvm_stop_multithreaded()`` if LLVM is operating in multithreaded mode. As
2103 such, ``llvm_shutdown()`` requires the same isolation guarantees as
2104 ``llvm_stop_multithreaded()``.
2106 Note that, if you use scope-based shutdown, you can use the
2107 ``llvm_shutdown_obj`` class, which calls ``llvm_shutdown()`` in its destructor.
2111 Lazy Initialization with ``ManagedStatic``
2112 ------------------------------------------
2114 ``ManagedStatic`` is a utility class in LLVM used to implement static
2115 initialization of static resources, such as the global type tables. Before the
2116 invocation of ``llvm_shutdown()``, it implements a simple lazy initialization
2117 scheme. Once ``llvm_start_multithreaded()`` returns, however, it uses
2118 double-checked locking to implement thread-safe lazy initialization.
2120 Note that, because no other threads are allowed to issue LLVM API calls before
2121 ``llvm_start_multithreaded()`` returns, it is possible to have
2122 ``ManagedStatic``\ s of ``llvm::sys::Mutex``\ s.
2124 The ``llvm_acquire_global_lock()`` and ``llvm_release_global_lock`` APIs provide
2125 access to the global lock used to implement the double-checked locking for lazy
2126 initialization. These should only be used internally to LLVM, and only if you
2127 know what you're doing!
2131 Achieving Isolation with ``LLVMContext``
2132 ----------------------------------------
2134 ``LLVMContext`` is an opaque class in the LLVM API which clients can use to
2135 operate multiple, isolated instances of LLVM concurrently within the same
2136 address space. For instance, in a hypothetical compile-server, the compilation
2137 of an individual translation unit is conceptually independent from all the
2138 others, and it would be desirable to be able to compile incoming translation
2139 units concurrently on independent server threads. Fortunately, ``LLVMContext``
2140 exists to enable just this kind of scenario!
2142 Conceptually, ``LLVMContext`` provides isolation. Every LLVM entity
2143 (``Module``\ s, ``Value``\ s, ``Type``\ s, ``Constant``\ s, etc.) in LLVM's
2144 in-memory IR belongs to an ``LLVMContext``. Entities in different contexts
2145 *cannot* interact with each other: ``Module``\ s in different contexts cannot be
2146 linked together, ``Function``\ s cannot be added to ``Module``\ s in different
2147 contexts, etc. What this means is that is is safe to compile on multiple
2148 threads simultaneously, as long as no two threads operate on entities within the
2151 In practice, very few places in the API require the explicit specification of a
2152 ``LLVMContext``, other than the ``Type`` creation/lookup APIs. Because every
2153 ``Type`` carries a reference to its owning context, most other entities can
2154 determine what context they belong to by looking at their own ``Type``. If you
2155 are adding new entities to LLVM IR, please try to maintain this interface
2158 For clients that do *not* require the benefits of isolation, LLVM provides a
2159 convenience API ``getGlobalContext()``. This returns a global, lazily
2160 initialized ``LLVMContext`` that may be used in situations where isolation is
2168 LLVM's "eager" JIT compiler is safe to use in threaded programs. Multiple
2169 threads can call ``ExecutionEngine::getPointerToFunction()`` or
2170 ``ExecutionEngine::runFunction()`` concurrently, and multiple threads can run
2171 code output by the JIT concurrently. The user must still ensure that only one
2172 thread accesses IR in a given ``LLVMContext`` while another thread might be
2173 modifying it. One way to do that is to always hold the JIT lock while accessing
2174 IR outside the JIT (the JIT *modifies* the IR by adding ``CallbackVH``\ s).
2175 Another way is to only call ``getPointerToFunction()`` from the
2176 ``LLVMContext``'s thread.
2178 When the JIT is configured to compile lazily (using
2179 ``ExecutionEngine::DisableLazyCompilation(false)``), there is currently a `race
2180 condition <http://llvm.org/bugs/show_bug.cgi?id=5184>`_ in updating call sites
2181 after a function is lazily-jitted. It's still possible to use the lazy JIT in a
2182 threaded program if you ensure that only one thread at a time can call any
2183 particular lazy stub and that the JIT lock guards any IR access, but we suggest
2184 using only the eager JIT in threaded programs.
2191 This section describes some of the advanced or obscure API's that most clients
2192 do not need to be aware of. These API's tend manage the inner workings of the
2193 LLVM system, and only need to be accessed in unusual circumstances.
2197 The ``ValueSymbolTable`` class
2198 ------------------------------
2200 The ``ValueSymbolTable`` (`doxygen
2201 <http://llvm.org/doxygen/classllvm_1_1ValueSymbolTable.html>`__) class provides
2202 a symbol table that the :ref:`Function <c_Function>` and Module_ classes use for
2203 naming value definitions. The symbol table can provide a name for any Value_.
2205 Note that the ``SymbolTable`` class should not be directly accessed by most
2206 clients. It should only be used when iteration over the symbol table names
2207 themselves are required, which is very special purpose. Note that not all LLVM
2208 Value_\ s have names, and those without names (i.e. they have an empty name) do
2209 not exist in the symbol table.
2211 Symbol tables support iteration over the values in the symbol table with
2212 ``begin/end/iterator`` and supports querying to see if a specific name is in the
2213 symbol table (with ``lookup``). The ``ValueSymbolTable`` class exposes no
2214 public mutator methods, instead, simply call ``setName`` on a value, which will
2215 autoinsert it into the appropriate symbol table.
2219 The ``User`` and owned ``Use`` classes' memory layout
2220 -----------------------------------------------------
2222 The ``User`` (`doxygen <http://llvm.org/doxygen/classllvm_1_1User.html>`__)
2223 class provides a basis for expressing the ownership of ``User`` towards other
2224 `Value instance <http://llvm.org/doxygen/classllvm_1_1Value.html>`_\ s. The
2225 ``Use`` (`doxygen <http://llvm.org/doxygen/classllvm_1_1Use.html>`__) helper
2226 class is employed to do the bookkeeping and to facilitate *O(1)* addition and
2231 Interaction and relationship between ``User`` and ``Use`` objects
2232 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2234 A subclass of ``User`` can choose between incorporating its ``Use`` objects or
2235 refer to them out-of-line by means of a pointer. A mixed variant (some ``Use``
2236 s inline others hung off) is impractical and breaks the invariant that the
2237 ``Use`` objects belonging to the same ``User`` form a contiguous array.
2239 We have 2 different layouts in the ``User`` (sub)classes:
2243 The ``Use`` object(s) are inside (resp. at fixed offset) of the ``User``
2244 object and there are a fixed number of them.
2248 The ``Use`` object(s) are referenced by a pointer to an array from the
2249 ``User`` object and there may be a variable number of them.
2251 As of v2.4 each layout still possesses a direct pointer to the start of the
2252 array of ``Use``\ s. Though not mandatory for layout a), we stick to this
2253 redundancy for the sake of simplicity. The ``User`` object also stores the
2254 number of ``Use`` objects it has. (Theoretically this information can also be
2255 calculated given the scheme presented below.)
2257 Special forms of allocation operators (``operator new``) enforce the following
2260 * Layout a) is modelled by prepending the ``User`` object by the ``Use[]``
2263 .. code-block:: none
2265 ...---.---.---.---.-------...
2266 | P | P | P | P | User
2267 '''---'---'---'---'-------'''
2269 * Layout b) is modelled by pointing at the ``Use[]`` array.
2271 .. code-block:: none
2282 *(In the above figures* '``P``' *stands for the* ``Use**`` *that is stored in
2283 each* ``Use`` *object in the member* ``Use::Prev`` *)*
2287 The waymarking algorithm
2288 ^^^^^^^^^^^^^^^^^^^^^^^^
2290 Since the ``Use`` objects are deprived of the direct (back)pointer to their
2291 ``User`` objects, there must be a fast and exact method to recover it. This is
2292 accomplished by the following scheme:
2294 A bit-encoding in the 2 LSBits (least significant bits) of the ``Use::Prev``
2295 allows to find the start of the ``User`` object:
2297 * ``00`` --- binary digit 0
2299 * ``01`` --- binary digit 1
2301 * ``10`` --- stop and calculate (``s``)
2303 * ``11`` --- full stop (``S``)
2305 Given a ``Use*``, all we have to do is to walk till we get a stop and we either
2306 have a ``User`` immediately behind or we have to walk to the next stop picking
2307 up digits and calculating the offset:
2309 .. code-block:: none
2311 .---.---.---.---.---.---.---.---.---.---.---.---.---.---.---.---.----------------
2312 | 1 | s | 1 | 0 | 1 | 0 | s | 1 | 1 | 0 | s | 1 | 1 | s | 1 | S | User (or User*)
2313 '---'---'---'---'---'---'---'---'---'---'---'---'---'---'---'---'----------------
2314 |+15 |+10 |+6 |+3 |+1
2317 | | | ______________________>
2318 | | ______________________________________>
2319 | __________________________________________________________>
2321 Only the significant number of bits need to be stored between the stops, so that
2322 the *worst case is 20 memory accesses* when there are 1000 ``Use`` objects
2323 associated with a ``User``.
2327 Reference implementation
2328 ^^^^^^^^^^^^^^^^^^^^^^^^
2330 The following literate Haskell fragment demonstrates the concept:
2332 .. code-block:: haskell
2334 > import Test.QuickCheck
2336 > digits :: Int -> [Char] -> [Char]
2337 > digits 0 acc = '0' : acc
2338 > digits 1 acc = '1' : acc
2339 > digits n acc = digits (n `div` 2) $ digits (n `mod` 2) acc
2341 > dist :: Int -> [Char] -> [Char]
2344 > dist 1 acc = let r = dist 0 acc in 's' : digits (length r) r
2345 > dist n acc = dist (n - 1) $ dist 1 acc
2347 > takeLast n ss = reverse $ take n $ reverse ss
2349 > test = takeLast 40 $ dist 20 []
2352 Printing <test> gives: ``"1s100000s11010s10100s1111s1010s110s11s1S"``
2354 The reverse algorithm computes the length of the string just by examining a
2357 .. code-block:: haskell
2359 > pref :: [Char] -> Int
2361 > pref ('s':'1':rest) = decode 2 1 rest
2362 > pref (_:rest) = 1 + pref rest
2364 > decode walk acc ('0':rest) = decode (walk + 1) (acc * 2) rest
2365 > decode walk acc ('1':rest) = decode (walk + 1) (acc * 2 + 1) rest
2366 > decode walk acc _ = walk + acc
2369 Now, as expected, printing <pref test> gives ``40``.
2371 We can *quickCheck* this with following property:
2373 .. code-block:: haskell
2375 > testcase = dist 2000 []
2376 > testcaseLength = length testcase
2378 > identityProp n = n > 0 && n <= testcaseLength ==> length arr == pref arr
2379 > where arr = takeLast n testcase
2382 As expected <quickCheck identityProp> gives:
2386 *Main> quickCheck identityProp
2387 OK, passed 100 tests.
2389 Let's be a bit more exhaustive:
2391 .. code-block:: haskell
2394 > deepCheck p = check (defaultConfig { configMaxTest = 500 }) p
2397 And here is the result of <deepCheck identityProp>:
2401 *Main> deepCheck identityProp
2402 OK, passed 500 tests.
2406 Tagging considerations
2407 ^^^^^^^^^^^^^^^^^^^^^^
2409 To maintain the invariant that the 2 LSBits of each ``Use**`` in ``Use`` never
2410 change after being set up, setters of ``Use::Prev`` must re-tag the new
2411 ``Use**`` on every modification. Accordingly getters must strip the tag bits.
2413 For layout b) instead of the ``User`` we find a pointer (``User*`` with LSBit
2414 set). Following this pointer brings us to the ``User``. A portable trick
2415 ensures that the first bytes of ``User`` (if interpreted as a pointer) never has
2416 the LSBit set. (Portability is relying on the fact that all known compilers
2417 place the ``vptr`` in the first word of the instances.)
2421 The Core LLVM Class Hierarchy Reference
2422 =======================================
2424 ``#include "llvm/IR/Type.h"``
2426 header source: `Type.h <http://llvm.org/doxygen/Type_8h-source.html>`_
2428 doxygen info: `Type Clases <http://llvm.org/doxygen/classllvm_1_1Type.html>`_
2430 The Core LLVM classes are the primary means of representing the program being
2431 inspected or transformed. The core LLVM classes are defined in header files in
2432 the ``include/llvm/`` directory, and implemented in the ``lib/VMCore``
2437 The Type class and Derived Types
2438 --------------------------------
2440 ``Type`` is a superclass of all type classes. Every ``Value`` has a ``Type``.
2441 ``Type`` cannot be instantiated directly but only through its subclasses.
2442 Certain primitive types (``VoidType``, ``LabelType``, ``FloatType`` and
2443 ``DoubleType``) have hidden subclasses. They are hidden because they offer no
2444 useful functionality beyond what the ``Type`` class offers except to distinguish
2445 themselves from other subclasses of ``Type``.
2447 All other types are subclasses of ``DerivedType``. Types can be named, but this
2448 is not a requirement. There exists exactly one instance of a given shape at any
2449 one time. This allows type equality to be performed with address equality of
2450 the Type Instance. That is, given two ``Type*`` values, the types are identical
2451 if the pointers are identical.
2455 Important Public Methods
2456 ^^^^^^^^^^^^^^^^^^^^^^^^
2458 * ``bool isIntegerTy() const``: Returns true for any integer type.
2460 * ``bool isFloatingPointTy()``: Return true if this is one of the five
2461 floating point types.
2463 * ``bool isSized()``: Return true if the type has known size. Things
2464 that don't have a size are abstract types, labels and void.
2468 Important Derived Types
2469 ^^^^^^^^^^^^^^^^^^^^^^^
2472 Subclass of DerivedType that represents integer types of any bit width. Any
2473 bit width between ``IntegerType::MIN_INT_BITS`` (1) and
2474 ``IntegerType::MAX_INT_BITS`` (~8 million) can be represented.
2476 * ``static const IntegerType* get(unsigned NumBits)``: get an integer
2477 type of a specific bit width.
2479 * ``unsigned getBitWidth() const``: Get the bit width of an integer type.
2482 This is subclassed by ArrayType, PointerType and VectorType.
2484 * ``const Type * getElementType() const``: Returns the type of each
2485 of the elements in the sequential type.
2488 This is a subclass of SequentialType and defines the interface for array
2491 * ``unsigned getNumElements() const``: Returns the number of elements
2495 Subclass of SequentialType for pointer types.
2498 Subclass of SequentialType for vector types. A vector type is similar to an
2499 ArrayType but is distinguished because it is a first class type whereas
2500 ArrayType is not. Vector types are used for vector operations and are usually
2501 small vectors of of an integer or floating point type.
2504 Subclass of DerivedTypes for struct types.
2509 Subclass of DerivedTypes for function types.
2511 * ``bool isVarArg() const``: Returns true if it's a vararg function.
2513 * ``const Type * getReturnType() const``: Returns the return type of the
2516 * ``const Type * getParamType (unsigned i)``: Returns the type of the ith
2519 * ``const unsigned getNumParams() const``: Returns the number of formal
2524 The ``Module`` class
2525 --------------------
2527 ``#include "llvm/IR/Module.h"``
2529 header source: `Module.h <http://llvm.org/doxygen/Module_8h-source.html>`_
2531 doxygen info: `Module Class <http://llvm.org/doxygen/classllvm_1_1Module.html>`_
2533 The ``Module`` class represents the top level structure present in LLVM
2534 programs. An LLVM module is effectively either a translation unit of the
2535 original program or a combination of several translation units merged by the
2536 linker. The ``Module`` class keeps track of a list of :ref:`Function
2537 <c_Function>`\ s, a list of GlobalVariable_\ s, and a SymbolTable_.
2538 Additionally, it contains a few helpful member functions that try to make common
2543 Important Public Members of the ``Module`` class
2544 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2546 * ``Module::Module(std::string name = "")``
2548 Constructing a Module_ is easy. You can optionally provide a name for it
2549 (probably based on the name of the translation unit).
2551 * | ``Module::iterator`` - Typedef for function list iterator
2552 | ``Module::const_iterator`` - Typedef for const_iterator.
2553 | ``begin()``, ``end()``, ``size()``, ``empty()``
2555 These are forwarding methods that make it easy to access the contents of a
2556 ``Module`` object's :ref:`Function <c_Function>` list.
2558 * ``Module::FunctionListType &getFunctionList()``
2560 Returns the list of :ref:`Function <c_Function>`\ s. This is necessary to use
2561 when you need to update the list or perform a complex action that doesn't have
2562 a forwarding method.
2566 * | ``Module::global_iterator`` - Typedef for global variable list iterator
2567 | ``Module::const_global_iterator`` - Typedef for const_iterator.
2568 | ``global_begin()``, ``global_end()``, ``global_size()``, ``global_empty()``
2570 These are forwarding methods that make it easy to access the contents of a
2571 ``Module`` object's GlobalVariable_ list.
2573 * ``Module::GlobalListType &getGlobalList()``
2575 Returns the list of GlobalVariable_\ s. This is necessary to use when you
2576 need to update the list or perform a complex action that doesn't have a
2581 * ``SymbolTable *getSymbolTable()``
2583 Return a reference to the SymbolTable_ for this ``Module``.
2587 * ``Function *getFunction(StringRef Name) const``
2589 Look up the specified function in the ``Module`` SymbolTable_. If it does not
2590 exist, return ``null``.
2592 * ``Function *getOrInsertFunction(const std::string &Name, const FunctionType
2595 Look up the specified function in the ``Module`` SymbolTable_. If it does not
2596 exist, add an external declaration for the function and return it.
2598 * ``std::string getTypeName(const Type *Ty)``
2600 If there is at least one entry in the SymbolTable_ for the specified Type_,
2601 return it. Otherwise return the empty string.
2603 * ``bool addTypeName(const std::string &Name, const Type *Ty)``
2605 Insert an entry in the SymbolTable_ mapping ``Name`` to ``Ty``. If there is
2606 already an entry for this name, true is returned and the SymbolTable_ is not
2614 ``#include "llvm/IR/Value.h"``
2616 header source: `Value.h <http://llvm.org/doxygen/Value_8h-source.html>`_
2618 doxygen info: `Value Class <http://llvm.org/doxygen/classllvm_1_1Value.html>`_
2620 The ``Value`` class is the most important class in the LLVM Source base. It
2621 represents a typed value that may be used (among other things) as an operand to
2622 an instruction. There are many different types of ``Value``\ s, such as
2623 Constant_\ s, Argument_\ s. Even Instruction_\ s and :ref:`Function
2624 <c_Function>`\ s are ``Value``\ s.
2626 A particular ``Value`` may be used many times in the LLVM representation for a
2627 program. For example, an incoming argument to a function (represented with an
2628 instance of the Argument_ class) is "used" by every instruction in the function
2629 that references the argument. To keep track of this relationship, the ``Value``
2630 class keeps a list of all of the ``User``\ s that is using it (the User_ class
2631 is a base class for all nodes in the LLVM graph that can refer to ``Value``\ s).
2632 This use list is how LLVM represents def-use information in the program, and is
2633 accessible through the ``use_*`` methods, shown below.
2635 Because LLVM is a typed representation, every LLVM ``Value`` is typed, and this
2636 Type_ is available through the ``getType()`` method. In addition, all LLVM
2637 values can be named. The "name" of the ``Value`` is a symbolic string printed
2640 .. code-block:: llvm
2646 The name of this instruction is "foo". **NOTE** that the name of any value may
2647 be missing (an empty string), so names should **ONLY** be used for debugging
2648 (making the source code easier to read, debugging printouts), they should not be
2649 used to keep track of values or map between them. For this purpose, use a
2650 ``std::map`` of pointers to the ``Value`` itself instead.
2652 One important aspect of LLVM is that there is no distinction between an SSA
2653 variable and the operation that produces it. Because of this, any reference to
2654 the value produced by an instruction (or the value available as an incoming
2655 argument, for example) is represented as a direct pointer to the instance of the
2656 class that represents this value. Although this may take some getting used to,
2657 it simplifies the representation and makes it easier to manipulate.
2661 Important Public Members of the ``Value`` class
2662 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2664 * | ``Value::use_iterator`` - Typedef for iterator over the use-list
2665 | ``Value::const_use_iterator`` - Typedef for const_iterator over the
2667 | ``unsigned use_size()`` - Returns the number of users of the value.
2668 | ``bool use_empty()`` - Returns true if there are no users.
2669 | ``use_iterator use_begin()`` - Get an iterator to the start of the
2671 | ``use_iterator use_end()`` - Get an iterator to the end of the use-list.
2672 | ``User *use_back()`` - Returns the last element in the list.
2674 These methods are the interface to access the def-use information in LLVM.
2675 As with all other iterators in LLVM, the naming conventions follow the
2676 conventions defined by the STL_.
2678 * ``Type *getType() const``
2679 This method returns the Type of the Value.
2681 * | ``bool hasName() const``
2682 | ``std::string getName() const``
2683 | ``void setName(const std::string &Name)``
2685 This family of methods is used to access and assign a name to a ``Value``, be
2686 aware of the :ref:`precaution above <nameWarning>`.
2688 * ``void replaceAllUsesWith(Value *V)``
2690 This method traverses the use list of a ``Value`` changing all User_\ s of the
2691 current value to refer to "``V``" instead. For example, if you detect that an
2692 instruction always produces a constant value (for example through constant
2693 folding), you can replace all uses of the instruction with the constant like
2698 Inst->replaceAllUsesWith(ConstVal);
2705 ``#include "llvm/IR/User.h"``
2707 header source: `User.h <http://llvm.org/doxygen/User_8h-source.html>`_
2709 doxygen info: `User Class <http://llvm.org/doxygen/classllvm_1_1User.html>`_
2713 The ``User`` class is the common base class of all LLVM nodes that may refer to
2714 ``Value``\ s. It exposes a list of "Operands" that are all of the ``Value``\ s
2715 that the User is referring to. The ``User`` class itself is a subclass of
2718 The operands of a ``User`` point directly to the LLVM ``Value`` that it refers
2719 to. Because LLVM uses Static Single Assignment (SSA) form, there can only be
2720 one definition referred to, allowing this direct connection. This connection
2721 provides the use-def information in LLVM.
2725 Important Public Members of the ``User`` class
2726 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2728 The ``User`` class exposes the operand list in two ways: through an index access
2729 interface and through an iterator based interface.
2731 * | ``Value *getOperand(unsigned i)``
2732 | ``unsigned getNumOperands()``
2734 These two methods expose the operands of the ``User`` in a convenient form for
2737 * | ``User::op_iterator`` - Typedef for iterator over the operand list
2738 | ``op_iterator op_begin()`` - Get an iterator to the start of the operand
2740 | ``op_iterator op_end()`` - Get an iterator to the end of the operand list.
2742 Together, these methods make up the iterator based interface to the operands
2748 The ``Instruction`` class
2749 -------------------------
2751 ``#include "llvm/IR/Instruction.h"``
2753 header source: `Instruction.h
2754 <http://llvm.org/doxygen/Instruction_8h-source.html>`_
2756 doxygen info: `Instruction Class
2757 <http://llvm.org/doxygen/classllvm_1_1Instruction.html>`_
2759 Superclasses: User_, Value_
2761 The ``Instruction`` class is the common base class for all LLVM instructions.
2762 It provides only a few methods, but is a very commonly used class. The primary
2763 data tracked by the ``Instruction`` class itself is the opcode (instruction
2764 type) and the parent BasicBlock_ the ``Instruction`` is embedded into. To
2765 represent a specific type of instruction, one of many subclasses of
2766 ``Instruction`` are used.
2768 Because the ``Instruction`` class subclasses the User_ class, its operands can
2769 be accessed in the same way as for other ``User``\ s (with the
2770 ``getOperand()``/``getNumOperands()`` and ``op_begin()``/``op_end()`` methods).
2771 An important file for the ``Instruction`` class is the ``llvm/Instruction.def``
2772 file. This file contains some meta-data about the various different types of
2773 instructions in LLVM. It describes the enum values that are used as opcodes
2774 (for example ``Instruction::Add`` and ``Instruction::ICmp``), as well as the
2775 concrete sub-classes of ``Instruction`` that implement the instruction (for
2776 example BinaryOperator_ and CmpInst_). Unfortunately, the use of macros in this
2777 file confuses doxygen, so these enum values don't show up correctly in the
2778 `doxygen output <http://llvm.org/doxygen/classllvm_1_1Instruction.html>`_.
2782 Important Subclasses of the ``Instruction`` class
2783 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2787 * ``BinaryOperator``
2789 This subclasses represents all two operand instructions whose operands must be
2790 the same type, except for the comparison instructions.
2795 This subclass is the parent of the 12 casting instructions. It provides
2796 common operations on cast instructions.
2802 This subclass respresents the two comparison instructions,
2803 `ICmpInst <LangRef.html#i_icmp>`_ (integer opreands), and
2804 `FCmpInst <LangRef.html#i_fcmp>`_ (floating point operands).
2808 * ``TerminatorInst``
2810 This subclass is the parent of all terminator instructions (those which can
2815 Important Public Members of the ``Instruction`` class
2816 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2818 * ``BasicBlock *getParent()``
2820 Returns the BasicBlock_ that this
2821 ``Instruction`` is embedded into.
2823 * ``bool mayWriteToMemory()``
2825 Returns true if the instruction writes to memory, i.e. it is a ``call``,
2826 ``free``, ``invoke``, or ``store``.
2828 * ``unsigned getOpcode()``
2830 Returns the opcode for the ``Instruction``.
2832 * ``Instruction *clone() const``
2834 Returns another instance of the specified instruction, identical in all ways
2835 to the original except that the instruction has no parent (i.e. it's not
2836 embedded into a BasicBlock_), and it has no name.
2840 The ``Constant`` class and subclasses
2841 -------------------------------------
2843 Constant represents a base class for different types of constants. It is
2844 subclassed by ConstantInt, ConstantArray, etc. for representing the various
2845 types of Constants. GlobalValue_ is also a subclass, which represents the
2846 address of a global variable or function.
2850 Important Subclasses of Constant
2851 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2853 * ConstantInt : This subclass of Constant represents an integer constant of
2856 * ``const APInt& getValue() const``: Returns the underlying
2857 value of this constant, an APInt value.
2859 * ``int64_t getSExtValue() const``: Converts the underlying APInt value to an
2860 int64_t via sign extension. If the value (not the bit width) of the APInt
2861 is too large to fit in an int64_t, an assertion will result. For this
2862 reason, use of this method is discouraged.
2864 * ``uint64_t getZExtValue() const``: Converts the underlying APInt value
2865 to a uint64_t via zero extension. IF the value (not the bit width) of the
2866 APInt is too large to fit in a uint64_t, an assertion will result. For this
2867 reason, use of this method is discouraged.
2869 * ``static ConstantInt* get(const APInt& Val)``: Returns the ConstantInt
2870 object that represents the value provided by ``Val``. The type is implied
2871 as the IntegerType that corresponds to the bit width of ``Val``.
2873 * ``static ConstantInt* get(const Type *Ty, uint64_t Val)``: Returns the
2874 ConstantInt object that represents the value provided by ``Val`` for integer
2877 * ConstantFP : This class represents a floating point constant.
2879 * ``double getValue() const``: Returns the underlying value of this constant.
2881 * ConstantArray : This represents a constant array.
2883 * ``const std::vector<Use> &getValues() const``: Returns a vector of
2884 component constants that makeup this array.
2886 * ConstantStruct : This represents a constant struct.
2888 * ``const std::vector<Use> &getValues() const``: Returns a vector of
2889 component constants that makeup this array.
2891 * GlobalValue : This represents either a global variable or a function. In
2892 either case, the value is a constant fixed address (after linking).
2896 The ``GlobalValue`` class
2897 -------------------------
2899 ``#include "llvm/IR/GlobalValue.h"``
2901 header source: `GlobalValue.h
2902 <http://llvm.org/doxygen/GlobalValue_8h-source.html>`_
2904 doxygen info: `GlobalValue Class
2905 <http://llvm.org/doxygen/classllvm_1_1GlobalValue.html>`_
2907 Superclasses: Constant_, User_, Value_
2909 Global values ( GlobalVariable_\ s or :ref:`Function <c_Function>`\ s) are the
2910 only LLVM values that are visible in the bodies of all :ref:`Function
2911 <c_Function>`\ s. Because they are visible at global scope, they are also
2912 subject to linking with other globals defined in different translation units.
2913 To control the linking process, ``GlobalValue``\ s know their linkage rules.
2914 Specifically, ``GlobalValue``\ s know whether they have internal or external
2915 linkage, as defined by the ``LinkageTypes`` enumeration.
2917 If a ``GlobalValue`` has internal linkage (equivalent to being ``static`` in C),
2918 it is not visible to code outside the current translation unit, and does not
2919 participate in linking. If it has external linkage, it is visible to external
2920 code, and does participate in linking. In addition to linkage information,
2921 ``GlobalValue``\ s keep track of which Module_ they are currently part of.
2923 Because ``GlobalValue``\ s are memory objects, they are always referred to by
2924 their **address**. As such, the Type_ of a global is always a pointer to its
2925 contents. It is important to remember this when using the ``GetElementPtrInst``
2926 instruction because this pointer must be dereferenced first. For example, if
2927 you have a ``GlobalVariable`` (a subclass of ``GlobalValue)`` that is an array
2928 of 24 ints, type ``[24 x i32]``, then the ``GlobalVariable`` is a pointer to
2929 that array. Although the address of the first element of this array and the
2930 value of the ``GlobalVariable`` are the same, they have different types. The
2931 ``GlobalVariable``'s type is ``[24 x i32]``. The first element's type is
2932 ``i32.`` Because of this, accessing a global value requires you to dereference
2933 the pointer with ``GetElementPtrInst`` first, then its elements can be accessed.
2934 This is explained in the `LLVM Language Reference Manual
2935 <LangRef.html#globalvars>`_.
2939 Important Public Members of the ``GlobalValue`` class
2940 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2942 * | ``bool hasInternalLinkage() const``
2943 | ``bool hasExternalLinkage() const``
2944 | ``void setInternalLinkage(bool HasInternalLinkage)``
2946 These methods manipulate the linkage characteristics of the ``GlobalValue``.
2948 * ``Module *getParent()``
2950 This returns the Module_ that the
2951 GlobalValue is currently embedded into.
2955 The ``Function`` class
2956 ----------------------
2958 ``#include "llvm/IR/Function.h"``
2960 header source: `Function.h <http://llvm.org/doxygen/Function_8h-source.html>`_
2962 doxygen info: `Function Class
2963 <http://llvm.org/doxygen/classllvm_1_1Function.html>`_
2965 Superclasses: GlobalValue_, Constant_, User_, Value_
2967 The ``Function`` class represents a single procedure in LLVM. It is actually
2968 one of the more complex classes in the LLVM hierarchy because it must keep track
2969 of a large amount of data. The ``Function`` class keeps track of a list of
2970 BasicBlock_\ s, a list of formal Argument_\ s, and a SymbolTable_.
2972 The list of BasicBlock_\ s is the most commonly used part of ``Function``
2973 objects. The list imposes an implicit ordering of the blocks in the function,
2974 which indicate how the code will be laid out by the backend. Additionally, the
2975 first BasicBlock_ is the implicit entry node for the ``Function``. It is not
2976 legal in LLVM to explicitly branch to this initial block. There are no implicit
2977 exit nodes, and in fact there may be multiple exit nodes from a single
2978 ``Function``. If the BasicBlock_ list is empty, this indicates that the
2979 ``Function`` is actually a function declaration: the actual body of the function
2980 hasn't been linked in yet.
2982 In addition to a list of BasicBlock_\ s, the ``Function`` class also keeps track
2983 of the list of formal Argument_\ s that the function receives. This container
2984 manages the lifetime of the Argument_ nodes, just like the BasicBlock_ list does
2985 for the BasicBlock_\ s.
2987 The SymbolTable_ is a very rarely used LLVM feature that is only used when you
2988 have to look up a value by name. Aside from that, the SymbolTable_ is used
2989 internally to make sure that there are not conflicts between the names of
2990 Instruction_\ s, BasicBlock_\ s, or Argument_\ s in the function body.
2992 Note that ``Function`` is a GlobalValue_ and therefore also a Constant_. The
2993 value of the function is its address (after linking) which is guaranteed to be
2998 Important Public Members of the ``Function``
2999 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3001 * ``Function(const FunctionType *Ty, LinkageTypes Linkage,
3002 const std::string &N = "", Module* Parent = 0)``
3004 Constructor used when you need to create new ``Function``\ s to add the
3005 program. The constructor must specify the type of the function to create and
3006 what type of linkage the function should have. The FunctionType_ argument
3007 specifies the formal arguments and return value for the function. The same
3008 FunctionType_ value can be used to create multiple functions. The ``Parent``
3009 argument specifies the Module in which the function is defined. If this
3010 argument is provided, the function will automatically be inserted into that
3011 module's list of functions.
3013 * ``bool isDeclaration()``
3015 Return whether or not the ``Function`` has a body defined. If the function is
3016 "external", it does not have a body, and thus must be resolved by linking with
3017 a function defined in a different translation unit.
3019 * | ``Function::iterator`` - Typedef for basic block list iterator
3020 | ``Function::const_iterator`` - Typedef for const_iterator.
3021 | ``begin()``, ``end()``, ``size()``, ``empty()``
3023 These are forwarding methods that make it easy to access the contents of a
3024 ``Function`` object's BasicBlock_ list.
3026 * ``Function::BasicBlockListType &getBasicBlockList()``
3028 Returns the list of BasicBlock_\ s. This is necessary to use when you need to
3029 update the list or perform a complex action that doesn't have a forwarding
3032 * | ``Function::arg_iterator`` - Typedef for the argument list iterator
3033 | ``Function::const_arg_iterator`` - Typedef for const_iterator.
3034 | ``arg_begin()``, ``arg_end()``, ``arg_size()``, ``arg_empty()``
3036 These are forwarding methods that make it easy to access the contents of a
3037 ``Function`` object's Argument_ list.
3039 * ``Function::ArgumentListType &getArgumentList()``
3041 Returns the list of Argument_. This is necessary to use when you need to
3042 update the list or perform a complex action that doesn't have a forwarding
3045 * ``BasicBlock &getEntryBlock()``
3047 Returns the entry ``BasicBlock`` for the function. Because the entry block
3048 for the function is always the first block, this returns the first block of
3051 * | ``Type *getReturnType()``
3052 | ``FunctionType *getFunctionType()``
3054 This traverses the Type_ of the ``Function`` and returns the return type of
3055 the function, or the FunctionType_ of the actual function.
3057 * ``SymbolTable *getSymbolTable()``
3059 Return a pointer to the SymbolTable_ for this ``Function``.
3063 The ``GlobalVariable`` class
3064 ----------------------------
3066 ``#include "llvm/IR/GlobalVariable.h"``
3068 header source: `GlobalVariable.h
3069 <http://llvm.org/doxygen/GlobalVariable_8h-source.html>`_
3071 doxygen info: `GlobalVariable Class
3072 <http://llvm.org/doxygen/classllvm_1_1GlobalVariable.html>`_
3074 Superclasses: GlobalValue_, Constant_, User_, Value_
3076 Global variables are represented with the (surprise surprise) ``GlobalVariable``
3077 class. Like functions, ``GlobalVariable``\ s are also subclasses of
3078 GlobalValue_, and as such are always referenced by their address (global values
3079 must live in memory, so their "name" refers to their constant address). See
3080 GlobalValue_ for more on this. Global variables may have an initial value
3081 (which must be a Constant_), and if they have an initializer, they may be marked
3082 as "constant" themselves (indicating that their contents never change at
3085 .. _m_GlobalVariable:
3087 Important Public Members of the ``GlobalVariable`` class
3088 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3090 * ``GlobalVariable(const Type *Ty, bool isConstant, LinkageTypes &Linkage,
3091 Constant *Initializer = 0, const std::string &Name = "", Module* Parent = 0)``
3093 Create a new global variable of the specified type. If ``isConstant`` is true
3094 then the global variable will be marked as unchanging for the program. The
3095 Linkage parameter specifies the type of linkage (internal, external, weak,
3096 linkonce, appending) for the variable. If the linkage is InternalLinkage,
3097 WeakAnyLinkage, WeakODRLinkage, LinkOnceAnyLinkage or LinkOnceODRLinkage, then
3098 the resultant global variable will have internal linkage. AppendingLinkage
3099 concatenates together all instances (in different translation units) of the
3100 variable into a single variable but is only applicable to arrays. See the
3101 `LLVM Language Reference <LangRef.html#modulestructure>`_ for further details
3102 on linkage types. Optionally an initializer, a name, and the module to put
3103 the variable into may be specified for the global variable as well.
3105 * ``bool isConstant() const``
3107 Returns true if this is a global variable that is known not to be modified at
3110 * ``bool hasInitializer()``
3112 Returns true if this ``GlobalVariable`` has an intializer.
3114 * ``Constant *getInitializer()``
3116 Returns the initial value for a ``GlobalVariable``. It is not legal to call
3117 this method if there is no initializer.
3121 The ``BasicBlock`` class
3122 ------------------------
3124 ``#include "llvm/IR/BasicBlock.h"``
3126 header source: `BasicBlock.h
3127 <http://llvm.org/doxygen/BasicBlock_8h-source.html>`_
3129 doxygen info: `BasicBlock Class
3130 <http://llvm.org/doxygen/classllvm_1_1BasicBlock.html>`_
3134 This class represents a single entry single exit section of the code, commonly
3135 known as a basic block by the compiler community. The ``BasicBlock`` class
3136 maintains a list of Instruction_\ s, which form the body of the block. Matching
3137 the language definition, the last element of this list of instructions is always
3138 a terminator instruction (a subclass of the TerminatorInst_ class).
3140 In addition to tracking the list of instructions that make up the block, the
3141 ``BasicBlock`` class also keeps track of the :ref:`Function <c_Function>` that
3142 it is embedded into.
3144 Note that ``BasicBlock``\ s themselves are Value_\ s, because they are
3145 referenced by instructions like branches and can go in the switch tables.
3146 ``BasicBlock``\ s have type ``label``.
3150 Important Public Members of the ``BasicBlock`` class
3151 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3153 * ``BasicBlock(const std::string &Name = "", Function *Parent = 0)``
3155 The ``BasicBlock`` constructor is used to create new basic blocks for
3156 insertion into a function. The constructor optionally takes a name for the
3157 new block, and a :ref:`Function <c_Function>` to insert it into. If the
3158 ``Parent`` parameter is specified, the new ``BasicBlock`` is automatically
3159 inserted at the end of the specified :ref:`Function <c_Function>`, if not
3160 specified, the BasicBlock must be manually inserted into the :ref:`Function
3163 * | ``BasicBlock::iterator`` - Typedef for instruction list iterator
3164 | ``BasicBlock::const_iterator`` - Typedef for const_iterator.
3165 | ``begin()``, ``end()``, ``front()``, ``back()``,
3166 ``size()``, ``empty()``
3167 STL-style functions for accessing the instruction list.
3169 These methods and typedefs are forwarding functions that have the same
3170 semantics as the standard library methods of the same names. These methods
3171 expose the underlying instruction list of a basic block in a way that is easy
3172 to manipulate. To get the full complement of container operations (including
3173 operations to update the list), you must use the ``getInstList()`` method.
3175 * ``BasicBlock::InstListType &getInstList()``
3177 This method is used to get access to the underlying container that actually
3178 holds the Instructions. This method must be used when there isn't a
3179 forwarding function in the ``BasicBlock`` class for the operation that you
3180 would like to perform. Because there are no forwarding functions for
3181 "updating" operations, you need to use this if you want to update the contents
3182 of a ``BasicBlock``.
3184 * ``Function *getParent()``
3186 Returns a pointer to :ref:`Function <c_Function>` the block is embedded into,
3187 or a null pointer if it is homeless.
3189 * ``TerminatorInst *getTerminator()``
3191 Returns a pointer to the terminator instruction that appears at the end of the
3192 ``BasicBlock``. If there is no terminator instruction, or if the last
3193 instruction in the block is not a terminator, then a null pointer is returned.
3197 The ``Argument`` class
3198 ----------------------
3200 This subclass of Value defines the interface for incoming formal arguments to a
3201 function. A Function maintains a list of its formal arguments. An argument has
3202 a pointer to the parent Function.