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 Passing functions and other callable objects
269 --------------------------------------------
271 Sometimes you may want a function to be passed a callback object. In order to
272 support lambda expressions and other function objects, you should not use the
273 traditional C approach of taking a function pointer and an opaque cookie:
277 void takeCallback(bool (*Callback)(Function *, void *), void *Cookie);
279 Instead, use one of the following approaches:
284 If you don't mind putting the definition of your function into a header file,
285 make it a function template that is templated on the callable type.
289 template<typename Callable>
290 void takeCallback(Callable Callback) {
294 The ``function_ref`` class template
295 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
298 (`doxygen <http://llvm.org/doxygen/classllvm_1_1function_ref.html>`__) class
299 template represents a reference to a callable object, templated over the type
300 of the callable. This is a good choice for passing a callback to a function,
301 if you don't need to hold onto the callback after the function returns. In this
302 way, ``function_ref`` is to ``std::function`` as ``StringRef`` is to
305 ``function_ref<Ret(Param1, Param2, ...)>`` can be implicitly constructed from
306 any callable object that can be called with arguments of type ``Param1``,
307 ``Param2``, ..., and returns a value that can be converted to type ``Ret``.
312 void visitBasicBlocks(Function *F, function_ref<bool (BasicBlock*)> Callback) {
313 for (BasicBlock &BB : *F)
322 visitBasicBlocks(F, [&](BasicBlock *BB) {
328 Note that a ``function_ref`` object contains pointers to external memory, so it
329 is not generally safe to store an instance of the class (unless you know that
330 the external storage will not be freed). If you need this ability, consider
331 using ``std::function``. ``function_ref`` is small enough that it should always
336 The ``DEBUG()`` macro and ``-debug`` option
337 -------------------------------------------
339 Often when working on your pass you will put a bunch of debugging printouts and
340 other code into your pass. After you get it working, you want to remove it, but
341 you may need it again in the future (to work out new bugs that you run across).
343 Naturally, because of this, you don't want to delete the debug printouts, but
344 you don't want them to always be noisy. A standard compromise is to comment
345 them out, allowing you to enable them if you need them in the future.
347 The ``llvm/Support/Debug.h`` (`doxygen
348 <http://llvm.org/doxygen/Debug_8h-source.html>`__) file provides a macro named
349 ``DEBUG()`` that is a much nicer solution to this problem. Basically, you can
350 put arbitrary code into the argument of the ``DEBUG`` macro, and it is only
351 executed if '``opt``' (or any other tool) is run with the '``-debug``' command
356 DEBUG(errs() << "I am here!\n");
358 Then you can run your pass like this:
362 $ opt < a.bc > /dev/null -mypass
364 $ opt < a.bc > /dev/null -mypass -debug
367 Using the ``DEBUG()`` macro instead of a home-brewed solution allows you to not
368 have to create "yet another" command line option for the debug output for your
369 pass. Note that ``DEBUG()`` macros are disabled for optimized builds, so they
370 do not cause a performance impact at all (for the same reason, they should also
371 not contain side-effects!).
373 One additional nice thing about the ``DEBUG()`` macro is that you can enable or
374 disable it directly in gdb. Just use "``set DebugFlag=0``" or "``set
375 DebugFlag=1``" from the gdb if the program is running. If the program hasn't
376 been started yet, you can always just run it with ``-debug``.
380 Fine grained debug info with ``DEBUG_TYPE`` and the ``-debug-only`` option
381 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
383 Sometimes you may find yourself in a situation where enabling ``-debug`` just
384 turns on **too much** information (such as when working on the code generator).
385 If you want to enable debug information with more fine-grained control, you
386 can define the ``DEBUG_TYPE`` macro and use the ``-debug-only`` option as
392 DEBUG(errs() << "No debug type\n");
393 #define DEBUG_TYPE "foo"
394 DEBUG(errs() << "'foo' debug type\n");
396 #define DEBUG_TYPE "bar"
397 DEBUG(errs() << "'bar' debug type\n"));
399 #define DEBUG_TYPE ""
400 DEBUG(errs() << "No debug type (2)\n");
402 Then you can run your pass like this:
406 $ opt < a.bc > /dev/null -mypass
408 $ opt < a.bc > /dev/null -mypass -debug
413 $ opt < a.bc > /dev/null -mypass -debug-only=foo
415 $ opt < a.bc > /dev/null -mypass -debug-only=bar
418 Of course, in practice, you should only set ``DEBUG_TYPE`` at the top of a file,
419 to specify the debug type for the entire module (if you do this before you
420 ``#include "llvm/Support/Debug.h"``, you don't have to insert the ugly
421 ``#undef``'s). Also, you should use names more meaningful than "foo" and "bar",
422 because there is no system in place to ensure that names do not conflict. If
423 two different modules use the same string, they will all be turned on when the
424 name is specified. This allows, for example, all debug information for
425 instruction scheduling to be enabled with ``-debug-type=InstrSched``, even if
426 the source lives in multiple files.
428 The ``DEBUG_WITH_TYPE`` macro is also available for situations where you would
429 like to set ``DEBUG_TYPE``, but only for one specific ``DEBUG`` statement. It
430 takes an additional first parameter, which is the type to use. For example, the
431 preceding example could be written as:
435 DEBUG_WITH_TYPE("", errs() << "No debug type\n");
436 DEBUG_WITH_TYPE("foo", errs() << "'foo' debug type\n");
437 DEBUG_WITH_TYPE("bar", errs() << "'bar' debug type\n"));
438 DEBUG_WITH_TYPE("", errs() << "No debug type (2)\n");
442 The ``Statistic`` class & ``-stats`` option
443 -------------------------------------------
445 The ``llvm/ADT/Statistic.h`` (`doxygen
446 <http://llvm.org/doxygen/Statistic_8h-source.html>`__) file provides a class
447 named ``Statistic`` that is used as a unified way to keep track of what the LLVM
448 compiler is doing and how effective various optimizations are. It is useful to
449 see what optimizations are contributing to making a particular program run
452 Often you may run your pass on some big program, and you're interested to see
453 how many times it makes a certain transformation. Although you can do this with
454 hand inspection, or some ad-hoc method, this is a real pain and not very useful
455 for big programs. Using the ``Statistic`` class makes it very easy to keep
456 track of this information, and the calculated information is presented in a
457 uniform manner with the rest of the passes being executed.
459 There are many examples of ``Statistic`` uses, but the basics of using it are as
462 #. Define your statistic like this:
466 #define DEBUG_TYPE "mypassname" // This goes before any #includes.
467 STATISTIC(NumXForms, "The # of times I did stuff");
469 The ``STATISTIC`` macro defines a static variable, whose name is specified by
470 the first argument. The pass name is taken from the ``DEBUG_TYPE`` macro, and
471 the description is taken from the second argument. The variable defined
472 ("NumXForms" in this case) acts like an unsigned integer.
474 #. Whenever you make a transformation, bump the counter:
478 ++NumXForms; // I did stuff!
480 That's all you have to do. To get '``opt``' to print out the statistics
481 gathered, use the '``-stats``' option:
485 $ opt -stats -mypassname < program.bc > /dev/null
486 ... statistics output ...
488 When running ``opt`` on a C file from the SPEC benchmark suite, it gives a
489 report that looks like this:
493 7646 bitcodewriter - Number of normal instructions
494 725 bitcodewriter - Number of oversized instructions
495 129996 bitcodewriter - Number of bitcode bytes written
496 2817 raise - Number of insts DCEd or constprop'd
497 3213 raise - Number of cast-of-self removed
498 5046 raise - Number of expression trees converted
499 75 raise - Number of other getelementptr's formed
500 138 raise - Number of load/store peepholes
501 42 deadtypeelim - Number of unused typenames removed from symtab
502 392 funcresolve - Number of varargs functions resolved
503 27 globaldce - Number of global variables removed
504 2 adce - Number of basic blocks removed
505 134 cee - Number of branches revectored
506 49 cee - Number of setcc instruction eliminated
507 532 gcse - Number of loads removed
508 2919 gcse - Number of instructions removed
509 86 indvars - Number of canonical indvars added
510 87 indvars - Number of aux indvars removed
511 25 instcombine - Number of dead inst eliminate
512 434 instcombine - Number of insts combined
513 248 licm - Number of load insts hoisted
514 1298 licm - Number of insts hoisted to a loop pre-header
515 3 licm - Number of insts hoisted to multiple loop preds (bad, no loop pre-header)
516 75 mem2reg - Number of alloca's promoted
517 1444 cfgsimplify - Number of blocks simplified
519 Obviously, with so many optimizations, having a unified framework for this stuff
520 is very nice. Making your pass fit well into the framework makes it more
521 maintainable and useful.
525 Viewing graphs while debugging code
526 -----------------------------------
528 Several of the important data structures in LLVM are graphs: for example CFGs
529 made out of LLVM :ref:`BasicBlocks <BasicBlock>`, CFGs made out of LLVM
530 :ref:`MachineBasicBlocks <MachineBasicBlock>`, and :ref:`Instruction Selection
531 DAGs <SelectionDAG>`. In many cases, while debugging various parts of the
532 compiler, it is nice to instantly visualize these graphs.
534 LLVM provides several callbacks that are available in a debug build to do
535 exactly that. If you call the ``Function::viewCFG()`` method, for example, the
536 current LLVM tool will pop up a window containing the CFG for the function where
537 each basic block is a node in the graph, and each node contains the instructions
538 in the block. Similarly, there also exists ``Function::viewCFGOnly()`` (does
539 not include the instructions), the ``MachineFunction::viewCFG()`` and
540 ``MachineFunction::viewCFGOnly()``, and the ``SelectionDAG::viewGraph()``
541 methods. Within GDB, for example, you can usually use something like ``call
542 DAG.viewGraph()`` to pop up a window. Alternatively, you can sprinkle calls to
543 these functions in your code in places you want to debug.
545 Getting this to work requires a small amount of setup. On Unix systems
546 with X11, install the `graphviz <http://www.graphviz.org>`_ toolkit, and make
547 sure 'dot' and 'gv' are in your path. If you are running on Mac OS X, download
548 and install the Mac OS X `Graphviz program
549 <http://www.pixelglow.com/graphviz/>`_ and add
550 ``/Applications/Graphviz.app/Contents/MacOS/`` (or wherever you install it) to
551 your path. The programs need not be present when configuring, building or
552 running LLVM and can simply be installed when needed during an active debug
555 ``SelectionDAG`` has been extended to make it easier to locate *interesting*
556 nodes in large complex graphs. From gdb, if you ``call DAG.setGraphColor(node,
557 "color")``, then the next ``call DAG.viewGraph()`` would highlight the node in
558 the specified color (choices of colors can be found at `colors
559 <http://www.graphviz.org/doc/info/colors.html>`_.) More complex node attributes
560 can be provided with ``call DAG.setGraphAttrs(node, "attributes")`` (choices can
561 be found at `Graph attributes <http://www.graphviz.org/doc/info/attrs.html>`_.)
562 If you want to restart and clear all the current graph attributes, then you can
563 ``call DAG.clearGraphAttrs()``.
565 Note that graph visualization features are compiled out of Release builds to
566 reduce file size. This means that you need a Debug+Asserts or Release+Asserts
567 build to use these features.
571 Picking the Right Data Structure for a Task
572 ===========================================
574 LLVM has a plethora of data structures in the ``llvm/ADT/`` directory, and we
575 commonly use STL data structures. This section describes the trade-offs you
576 should consider when you pick one.
578 The first step is a choose your own adventure: do you want a sequential
579 container, a set-like container, or a map-like container? The most important
580 thing when choosing a container is the algorithmic properties of how you plan to
581 access the container. Based on that, you should use:
584 * a :ref:`map-like <ds_map>` container if you need efficient look-up of a
585 value based on another value. Map-like containers also support efficient
586 queries for containment (whether a key is in the map). Map-like containers
587 generally do not support efficient reverse mapping (values to keys). If you
588 need that, use two maps. Some map-like containers also support efficient
589 iteration through the keys in sorted order. Map-like containers are the most
590 expensive sort, only use them if you need one of these capabilities.
592 * a :ref:`set-like <ds_set>` container if you need to put a bunch of stuff into
593 a container that automatically eliminates duplicates. Some set-like
594 containers support efficient iteration through the elements in sorted order.
595 Set-like containers are more expensive than sequential containers.
597 * a :ref:`sequential <ds_sequential>` container provides the most efficient way
598 to add elements and keeps track of the order they are added to the collection.
599 They permit duplicates and support efficient iteration, but do not support
600 efficient look-up based on a key.
602 * a :ref:`string <ds_string>` container is a specialized sequential container or
603 reference structure that is used for character or byte arrays.
605 * a :ref:`bit <ds_bit>` container provides an efficient way to store and
606 perform set operations on sets of numeric id's, while automatically
607 eliminating duplicates. Bit containers require a maximum of 1 bit for each
608 identifier you want to store.
610 Once the proper category of container is determined, you can fine tune the
611 memory use, constant factors, and cache behaviors of access by intelligently
612 picking a member of the category. Note that constant factors and cache behavior
613 can be a big deal. If you have a vector that usually only contains a few
614 elements (but could contain many), for example, it's much better to use
615 :ref:`SmallVector <dss_smallvector>` than :ref:`vector <dss_vector>`. Doing so
616 avoids (relatively) expensive malloc/free calls, which dwarf the cost of adding
617 the elements to the container.
621 Sequential Containers (std::vector, std::list, etc)
622 ---------------------------------------------------
624 There are a variety of sequential containers available for you, based on your
625 needs. Pick the first in this section that will do what you want.
632 The ``llvm::ArrayRef`` class is the preferred class to use in an interface that
633 accepts a sequential list of elements in memory and just reads from them. By
634 taking an ``ArrayRef``, the API can be passed a fixed size array, an
635 ``std::vector``, an ``llvm::SmallVector`` and anything else that is contiguous
643 Fixed size arrays are very simple and very fast. They are good if you know
644 exactly how many elements you have, or you have a (low) upper bound on how many
649 Heap Allocated Arrays
650 ^^^^^^^^^^^^^^^^^^^^^
652 Heap allocated arrays (``new[]`` + ``delete[]``) are also simple. They are good
653 if the number of elements is variable, if you know how many elements you will
654 need before the array is allocated, and if the array is usually large (if not,
655 consider a :ref:`SmallVector <dss_smallvector>`). The cost of a heap allocated
656 array is the cost of the new/delete (aka malloc/free). Also note that if you
657 are allocating an array of a type with a constructor, the constructor and
658 destructors will be run for every element in the array (re-sizable vectors only
659 construct those elements actually used).
661 .. _dss_tinyptrvector:
663 llvm/ADT/TinyPtrVector.h
664 ^^^^^^^^^^^^^^^^^^^^^^^^
666 ``TinyPtrVector<Type>`` is a highly specialized collection class that is
667 optimized to avoid allocation in the case when a vector has zero or one
668 elements. It has two major restrictions: 1) it can only hold values of pointer
669 type, and 2) it cannot hold a null pointer.
671 Since this container is highly specialized, it is rarely used.
675 llvm/ADT/SmallVector.h
676 ^^^^^^^^^^^^^^^^^^^^^^
678 ``SmallVector<Type, N>`` is a simple class that looks and smells just like
679 ``vector<Type>``: it supports efficient iteration, lays out elements in memory
680 order (so you can do pointer arithmetic between elements), supports efficient
681 push_back/pop_back operations, supports efficient random access to its elements,
684 The advantage of SmallVector is that it allocates space for some number of
685 elements (N) **in the object itself**. Because of this, if the SmallVector is
686 dynamically smaller than N, no malloc is performed. This can be a big win in
687 cases where the malloc/free call is far more expensive than the code that
688 fiddles around with the elements.
690 This is good for vectors that are "usually small" (e.g. the number of
691 predecessors/successors of a block is usually less than 8). On the other hand,
692 this makes the size of the SmallVector itself large, so you don't want to
693 allocate lots of them (doing so will waste a lot of space). As such,
694 SmallVectors are most useful when on the stack.
696 SmallVector also provides a nice portable and efficient replacement for
701 Prefer to use ``SmallVectorImpl<T>`` as a parameter type.
703 In APIs that don't care about the "small size" (most?), prefer to use
704 the ``SmallVectorImpl<T>`` class, which is basically just the "vector
705 header" (and methods) without the elements allocated after it. Note that
706 ``SmallVector<T, N>`` inherits from ``SmallVectorImpl<T>`` so the
707 conversion is implicit and costs nothing. E.g.
711 // BAD: Clients cannot pass e.g. SmallVector<Foo, 4>.
712 hardcodedSmallSize(SmallVector<Foo, 2> &Out);
713 // GOOD: Clients can pass any SmallVector<Foo, N>.
714 allowsAnySmallSize(SmallVectorImpl<Foo> &Out);
717 SmallVector<Foo, 8> Vec;
718 hardcodedSmallSize(Vec); // Error.
719 allowsAnySmallSize(Vec); // Works.
722 Even though it has "``Impl``" in the name, this is so widely used that
723 it really isn't "private to the implementation" anymore. A name like
724 ``SmallVectorHeader`` would be more appropriate.
731 ``std::vector`` is well loved and respected. It is useful when SmallVector
732 isn't: when the size of the vector is often large (thus the small optimization
733 will rarely be a benefit) or if you will be allocating many instances of the
734 vector itself (which would waste space for elements that aren't in the
735 container). vector is also useful when interfacing with code that expects
738 One worthwhile note about std::vector: avoid code like this:
747 Instead, write this as:
757 Doing so will save (at least) one heap allocation and free per iteration of the
765 ``std::deque`` is, in some senses, a generalized version of ``std::vector``.
766 Like ``std::vector``, it provides constant time random access and other similar
767 properties, but it also provides efficient access to the front of the list. It
768 does not guarantee continuity of elements within memory.
770 In exchange for this extra flexibility, ``std::deque`` has significantly higher
771 constant factor costs than ``std::vector``. If possible, use ``std::vector`` or
779 ``std::list`` is an extremely inefficient class that is rarely useful. It
780 performs a heap allocation for every element inserted into it, thus having an
781 extremely high constant factor, particularly for small data types.
782 ``std::list`` also only supports bidirectional iteration, not random access
785 In exchange for this high cost, std::list supports efficient access to both ends
786 of the list (like ``std::deque``, but unlike ``std::vector`` or
787 ``SmallVector``). In addition, the iterator invalidation characteristics of
788 std::list are stronger than that of a vector class: inserting or removing an
789 element into the list does not invalidate iterator or pointers to other elements
797 ``ilist<T>`` implements an 'intrusive' doubly-linked list. It is intrusive,
798 because it requires the element to store and provide access to the prev/next
799 pointers for the list.
801 ``ilist`` has the same drawbacks as ``std::list``, and additionally requires an
802 ``ilist_traits`` implementation for the element type, but it provides some novel
803 characteristics. In particular, it can efficiently store polymorphic objects,
804 the traits class is informed when an element is inserted or removed from the
805 list, and ``ilist``\ s are guaranteed to support a constant-time splice
808 These properties are exactly what we want for things like ``Instruction``\ s and
809 basic blocks, which is why these are implemented with ``ilist``\ s.
811 Related classes of interest are explained in the following subsections:
813 * :ref:`ilist_traits <dss_ilist_traits>`
815 * :ref:`iplist <dss_iplist>`
817 * :ref:`llvm/ADT/ilist_node.h <dss_ilist_node>`
819 * :ref:`Sentinels <dss_ilist_sentinel>`
821 .. _dss_packedvector:
823 llvm/ADT/PackedVector.h
824 ^^^^^^^^^^^^^^^^^^^^^^^
826 Useful for storing a vector of values using only a few number of bits for each
827 value. Apart from the standard operations of a vector-like container, it can
828 also perform an 'or' set operation.
836 FirstCondition = 0x1,
837 SecondCondition = 0x2,
842 PackedVector<State, 2> Vec1;
843 Vec1.push_back(FirstCondition);
845 PackedVector<State, 2> Vec2;
846 Vec2.push_back(SecondCondition);
849 return Vec1[0]; // returns 'Both'.
852 .. _dss_ilist_traits:
857 ``ilist_traits<T>`` is ``ilist<T>``'s customization mechanism. ``iplist<T>``
858 (and consequently ``ilist<T>``) publicly derive from this traits class.
865 ``iplist<T>`` is ``ilist<T>``'s base and as such supports a slightly narrower
866 interface. Notably, inserters from ``T&`` are absent.
868 ``ilist_traits<T>`` is a public base of this class and can be used for a wide
869 variety of customizations.
873 llvm/ADT/ilist_node.h
874 ^^^^^^^^^^^^^^^^^^^^^
876 ``ilist_node<T>`` implements a the forward and backward links that are expected
877 by the ``ilist<T>`` (and analogous containers) in the default manner.
879 ``ilist_node<T>``\ s are meant to be embedded in the node type ``T``, usually
880 ``T`` publicly derives from ``ilist_node<T>``.
882 .. _dss_ilist_sentinel:
887 ``ilist``\ s have another specialty that must be considered. To be a good
888 citizen in the C++ ecosystem, it needs to support the standard container
889 operations, such as ``begin`` and ``end`` iterators, etc. Also, the
890 ``operator--`` must work correctly on the ``end`` iterator in the case of
891 non-empty ``ilist``\ s.
893 The only sensible solution to this problem is to allocate a so-called *sentinel*
894 along with the intrusive list, which serves as the ``end`` iterator, providing
895 the back-link to the last element. However conforming to the C++ convention it
896 is illegal to ``operator++`` beyond the sentinel and it also must not be
899 These constraints allow for some implementation freedom to the ``ilist`` how to
900 allocate and store the sentinel. The corresponding policy is dictated by
901 ``ilist_traits<T>``. By default a ``T`` gets heap-allocated whenever the need
902 for a sentinel arises.
904 While the default policy is sufficient in most cases, it may break down when
905 ``T`` does not provide a default constructor. Also, in the case of many
906 instances of ``ilist``\ s, the memory overhead of the associated sentinels is
907 wasted. To alleviate the situation with numerous and voluminous
908 ``T``-sentinels, sometimes a trick is employed, leading to *ghostly sentinels*.
910 Ghostly sentinels are obtained by specially-crafted ``ilist_traits<T>`` which
911 superpose the sentinel with the ``ilist`` instance in memory. Pointer
912 arithmetic is used to obtain the sentinel, which is relative to the ``ilist``'s
913 ``this`` pointer. The ``ilist`` is augmented by an extra pointer, which serves
914 as the back-link of the sentinel. This is the only field in the ghostly
915 sentinel which can be legally accessed.
919 Other Sequential Container options
920 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
922 Other STL containers are available, such as ``std::string``.
924 There are also various STL adapter classes such as ``std::queue``,
925 ``std::priority_queue``, ``std::stack``, etc. These provide simplified access
926 to an underlying container but don't affect the cost of the container itself.
930 String-like containers
931 ----------------------
933 There are a variety of ways to pass around and use strings in C and C++, and
934 LLVM adds a few new options to choose from. Pick the first option on this list
935 that will do what you need, they are ordered according to their relative cost.
937 Note that is is generally preferred to *not* pass strings around as ``const
938 char*``'s. These have a number of problems, including the fact that they
939 cannot represent embedded nul ("\0") characters, and do not have a length
940 available efficiently. The general replacement for '``const char*``' is
943 For more information on choosing string containers for APIs, please see
944 :ref:`Passing Strings <string_apis>`.
951 The StringRef class is a simple value class that contains a pointer to a
952 character and a length, and is quite related to the :ref:`ArrayRef
953 <dss_arrayref>` class (but specialized for arrays of characters). Because
954 StringRef carries a length with it, it safely handles strings with embedded nul
955 characters in it, getting the length does not require a strlen call, and it even
956 has very convenient APIs for slicing and dicing the character range that it
959 StringRef is ideal for passing simple strings around that are known to be live,
960 either because they are C string literals, std::string, a C array, or a
961 SmallVector. Each of these cases has an efficient implicit conversion to
962 StringRef, which doesn't result in a dynamic strlen being executed.
964 StringRef has a few major limitations which make more powerful string containers
967 #. You cannot directly convert a StringRef to a 'const char*' because there is
968 no way to add a trailing nul (unlike the .c_str() method on various stronger
971 #. StringRef doesn't own or keep alive the underlying string bytes.
972 As such it can easily lead to dangling pointers, and is not suitable for
973 embedding in datastructures in most cases (instead, use an std::string or
974 something like that).
976 #. For the same reason, StringRef cannot be used as the return value of a
977 method if the method "computes" the result string. Instead, use std::string.
979 #. StringRef's do not allow you to mutate the pointed-to string bytes and it
980 doesn't allow you to insert or remove bytes from the range. For editing
981 operations like this, it interoperates with the :ref:`Twine <dss_twine>`
984 Because of its strengths and limitations, it is very common for a function to
985 take a StringRef and for a method on an object to return a StringRef that points
986 into some string that it owns.
993 The Twine class is used as an intermediary datatype for APIs that want to take a
994 string that can be constructed inline with a series of concatenations. Twine
995 works by forming recursive instances of the Twine datatype (a simple value
996 object) on the stack as temporary objects, linking them together into a tree
997 which is then linearized when the Twine is consumed. Twine is only safe to use
998 as the argument to a function, and should always be a const reference, e.g.:
1002 void foo(const Twine &T);
1006 foo(X + "." + Twine(i));
1008 This example forms a string like "blarg.42" by concatenating the values
1009 together, and does not form intermediate strings containing "blarg" or "blarg.".
1011 Because Twine is constructed with temporary objects on the stack, and because
1012 these instances are destroyed at the end of the current statement, it is an
1013 inherently dangerous API. For example, this simple variant contains undefined
1014 behavior and will probably crash:
1018 void foo(const Twine &T);
1022 const Twine &Tmp = X + "." + Twine(i);
1025 ... because the temporaries are destroyed before the call. That said, Twine's
1026 are much more efficient than intermediate std::string temporaries, and they work
1027 really well with StringRef. Just be aware of their limitations.
1029 .. _dss_smallstring:
1031 llvm/ADT/SmallString.h
1032 ^^^^^^^^^^^^^^^^^^^^^^
1034 SmallString is a subclass of :ref:`SmallVector <dss_smallvector>` that adds some
1035 convenience APIs like += that takes StringRef's. SmallString avoids allocating
1036 memory in the case when the preallocated space is enough to hold its data, and
1037 it calls back to general heap allocation when required. Since it owns its data,
1038 it is very safe to use and supports full mutation of the string.
1040 Like SmallVector's, the big downside to SmallString is their sizeof. While they
1041 are optimized for small strings, they themselves are not particularly small.
1042 This means that they work great for temporary scratch buffers on the stack, but
1043 should not generally be put into the heap: it is very rare to see a SmallString
1044 as the member of a frequently-allocated heap data structure or returned
1052 The standard C++ std::string class is a very general class that (like
1053 SmallString) owns its underlying data. sizeof(std::string) is very reasonable
1054 so it can be embedded into heap data structures and returned by-value. On the
1055 other hand, std::string is highly inefficient for inline editing (e.g.
1056 concatenating a bunch of stuff together) and because it is provided by the
1057 standard library, its performance characteristics depend a lot of the host
1058 standard library (e.g. libc++ and MSVC provide a highly optimized string class,
1059 GCC contains a really slow implementation).
1061 The major disadvantage of std::string is that almost every operation that makes
1062 them larger can allocate memory, which is slow. As such, it is better to use
1063 SmallVector or Twine as a scratch buffer, but then use std::string to persist
1068 Set-Like Containers (std::set, SmallSet, SetVector, etc)
1069 --------------------------------------------------------
1071 Set-like containers are useful when you need to canonicalize multiple values
1072 into a single representation. There are several different choices for how to do
1073 this, providing various trade-offs.
1075 .. _dss_sortedvectorset:
1080 If you intend to insert a lot of elements, then do a lot of queries, a great
1081 approach is to use a vector (or other sequential container) with
1082 std::sort+std::unique to remove duplicates. This approach works really well if
1083 your usage pattern has these two distinct phases (insert then query), and can be
1084 coupled with a good choice of :ref:`sequential container <ds_sequential>`.
1086 This combination provides the several nice properties: the result data is
1087 contiguous in memory (good for cache locality), has few allocations, is easy to
1088 address (iterators in the final vector are just indices or pointers), and can be
1089 efficiently queried with a standard binary search (e.g.
1090 ``std::lower_bound``; if you want the whole range of elements comparing
1091 equal, use ``std::equal_range``).
1098 If you have a set-like data structure that is usually small and whose elements
1099 are reasonably small, a ``SmallSet<Type, N>`` is a good choice. This set has
1100 space for N elements in place (thus, if the set is dynamically smaller than N,
1101 no malloc traffic is required) and accesses them with a simple linear search.
1102 When the set grows beyond 'N' elements, it allocates a more expensive
1103 representation that guarantees efficient access (for most types, it falls back
1104 to std::set, but for pointers it uses something far better, :ref:`SmallPtrSet
1107 The magic of this class is that it handles small sets extremely efficiently, but
1108 gracefully handles extremely large sets without loss of efficiency. The
1109 drawback is that the interface is quite small: it supports insertion, queries
1110 and erasing, but does not support iteration.
1112 .. _dss_smallptrset:
1114 llvm/ADT/SmallPtrSet.h
1115 ^^^^^^^^^^^^^^^^^^^^^^
1117 SmallPtrSet has all the advantages of ``SmallSet`` (and a ``SmallSet`` of
1118 pointers is transparently implemented with a ``SmallPtrSet``), but also supports
1119 iterators. If more than 'N' insertions are performed, a single quadratically
1120 probed hash table is allocated and grows as needed, providing extremely
1121 efficient access (constant time insertion/deleting/queries with low constant
1122 factors) and is very stingy with malloc traffic.
1124 Note that, unlike ``std::set``, the iterators of ``SmallPtrSet`` are invalidated
1125 whenever an insertion occurs. Also, the values visited by the iterators are not
1126 visited in sorted order.
1133 DenseSet is a simple quadratically probed hash table. It excels at supporting
1134 small values: it uses a single allocation to hold all of the pairs that are
1135 currently inserted in the set. DenseSet is a great way to unique small values
1136 that are not simple pointers (use :ref:`SmallPtrSet <dss_smallptrset>` for
1137 pointers). Note that DenseSet has the same requirements for the value type that
1138 :ref:`DenseMap <dss_densemap>` has.
1142 llvm/ADT/SparseSet.h
1143 ^^^^^^^^^^^^^^^^^^^^
1145 SparseSet holds a small number of objects identified by unsigned keys of
1146 moderate size. It uses a lot of memory, but provides operations that are almost
1147 as fast as a vector. Typical keys are physical registers, virtual registers, or
1148 numbered basic blocks.
1150 SparseSet is useful for algorithms that need very fast clear/find/insert/erase
1151 and fast iteration over small sets. It is not intended for building composite
1154 .. _dss_sparsemultiset:
1156 llvm/ADT/SparseMultiSet.h
1157 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1159 SparseMultiSet adds multiset behavior to SparseSet, while retaining SparseSet's
1160 desirable attributes. Like SparseSet, it typically uses a lot of memory, but
1161 provides operations that are almost as fast as a vector. Typical keys are
1162 physical registers, virtual registers, or numbered basic blocks.
1164 SparseMultiSet is useful for algorithms that need very fast
1165 clear/find/insert/erase of the entire collection, and iteration over sets of
1166 elements sharing a key. It is often a more efficient choice than using composite
1167 data structures (e.g. vector-of-vectors, map-of-vectors). It is not intended for
1168 building composite data structures.
1172 llvm/ADT/FoldingSet.h
1173 ^^^^^^^^^^^^^^^^^^^^^
1175 FoldingSet is an aggregate class that is really good at uniquing
1176 expensive-to-create or polymorphic objects. It is a combination of a chained
1177 hash table with intrusive links (uniqued objects are required to inherit from
1178 FoldingSetNode) that uses :ref:`SmallVector <dss_smallvector>` as part of its ID
1181 Consider a case where you want to implement a "getOrCreateFoo" method for a
1182 complex object (for example, a node in the code generator). The client has a
1183 description of **what** it wants to generate (it knows the opcode and all the
1184 operands), but we don't want to 'new' a node, then try inserting it into a set
1185 only to find out it already exists, at which point we would have to delete it
1186 and return the node that already exists.
1188 To support this style of client, FoldingSet perform a query with a
1189 FoldingSetNodeID (which wraps SmallVector) that can be used to describe the
1190 element that we want to query for. The query either returns the element
1191 matching the ID or it returns an opaque ID that indicates where insertion should
1192 take place. Construction of the ID usually does not require heap traffic.
1194 Because FoldingSet uses intrusive links, it can support polymorphic objects in
1195 the set (for example, you can have SDNode instances mixed with LoadSDNodes).
1196 Because the elements are individually allocated, pointers to the elements are
1197 stable: inserting or removing elements does not invalidate any pointers to other
1205 ``std::set`` is a reasonable all-around set class, which is decent at many
1206 things but great at nothing. std::set allocates memory for each element
1207 inserted (thus it is very malloc intensive) and typically stores three pointers
1208 per element in the set (thus adding a large amount of per-element space
1209 overhead). It offers guaranteed log(n) performance, which is not particularly
1210 fast from a complexity standpoint (particularly if the elements of the set are
1211 expensive to compare, like strings), and has extremely high constant factors for
1212 lookup, insertion and removal.
1214 The advantages of std::set are that its iterators are stable (deleting or
1215 inserting an element from the set does not affect iterators or pointers to other
1216 elements) and that iteration over the set is guaranteed to be in sorted order.
1217 If the elements in the set are large, then the relative overhead of the pointers
1218 and malloc traffic is not a big deal, but if the elements of the set are small,
1219 std::set is almost never a good choice.
1223 llvm/ADT/SetVector.h
1224 ^^^^^^^^^^^^^^^^^^^^
1226 LLVM's ``SetVector<Type>`` is an adapter class that combines your choice of a
1227 set-like container along with a :ref:`Sequential Container <ds_sequential>` The
1228 important property that this provides is efficient insertion with uniquing
1229 (duplicate elements are ignored) with iteration support. It implements this by
1230 inserting elements into both a set-like container and the sequential container,
1231 using the set-like container for uniquing and the sequential container for
1234 The difference between SetVector and other sets is that the order of iteration
1235 is guaranteed to match the order of insertion into the SetVector. This property
1236 is really important for things like sets of pointers. Because pointer values
1237 are non-deterministic (e.g. vary across runs of the program on different
1238 machines), iterating over the pointers in the set will not be in a well-defined
1241 The drawback of SetVector is that it requires twice as much space as a normal
1242 set and has the sum of constant factors from the set-like container and the
1243 sequential container that it uses. Use it **only** if you need to iterate over
1244 the elements in a deterministic order. SetVector is also expensive to delete
1245 elements out of (linear time), unless you use its "pop_back" method, which is
1248 ``SetVector`` is an adapter class that defaults to using ``std::vector`` and a
1249 size 16 ``SmallSet`` for the underlying containers, so it is quite expensive.
1250 However, ``"llvm/ADT/SetVector.h"`` also provides a ``SmallSetVector`` class,
1251 which defaults to using a ``SmallVector`` and ``SmallSet`` of a specified size.
1252 If you use this, and if your sets are dynamically smaller than ``N``, you will
1253 save a lot of heap traffic.
1255 .. _dss_uniquevector:
1257 llvm/ADT/UniqueVector.h
1258 ^^^^^^^^^^^^^^^^^^^^^^^
1260 UniqueVector is similar to :ref:`SetVector <dss_setvector>` but it retains a
1261 unique ID for each element inserted into the set. It internally contains a map
1262 and a vector, and it assigns a unique ID for each value inserted into the set.
1264 UniqueVector is very expensive: its cost is the sum of the cost of maintaining
1265 both the map and vector, it has high complexity, high constant factors, and
1266 produces a lot of malloc traffic. It should be avoided.
1268 .. _dss_immutableset:
1270 llvm/ADT/ImmutableSet.h
1271 ^^^^^^^^^^^^^^^^^^^^^^^
1273 ImmutableSet is an immutable (functional) set implementation based on an AVL
1274 tree. Adding or removing elements is done through a Factory object and results
1275 in the creation of a new ImmutableSet object. If an ImmutableSet already exists
1276 with the given contents, then the existing one is returned; equality is compared
1277 with a FoldingSetNodeID. The time and space complexity of add or remove
1278 operations is logarithmic in the size of the original set.
1280 There is no method for returning an element of the set, you can only check for
1285 Other Set-Like Container Options
1286 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1288 The STL provides several other options, such as std::multiset and the various
1289 "hash_set" like containers (whether from C++ TR1 or from the SGI library). We
1290 never use hash_set and unordered_set because they are generally very expensive
1291 (each insertion requires a malloc) and very non-portable.
1293 std::multiset is useful if you're not interested in elimination of duplicates,
1294 but has all the drawbacks of std::set. A sorted vector (where you don't delete
1295 duplicate entries) or some other approach is almost always better.
1299 Map-Like Containers (std::map, DenseMap, etc)
1300 ---------------------------------------------
1302 Map-like containers are useful when you want to associate data to a key. As
1303 usual, there are a lot of different ways to do this. :)
1305 .. _dss_sortedvectormap:
1310 If your usage pattern follows a strict insert-then-query approach, you can
1311 trivially use the same approach as :ref:`sorted vectors for set-like containers
1312 <dss_sortedvectorset>`. The only difference is that your query function (which
1313 uses std::lower_bound to get efficient log(n) lookup) should only compare the
1314 key, not both the key and value. This yields the same advantages as sorted
1319 llvm/ADT/StringMap.h
1320 ^^^^^^^^^^^^^^^^^^^^
1322 Strings are commonly used as keys in maps, and they are difficult to support
1323 efficiently: they are variable length, inefficient to hash and compare when
1324 long, expensive to copy, etc. StringMap is a specialized container designed to
1325 cope with these issues. It supports mapping an arbitrary range of bytes to an
1326 arbitrary other object.
1328 The StringMap implementation uses a quadratically-probed hash table, where the
1329 buckets store a pointer to the heap allocated entries (and some other stuff).
1330 The entries in the map must be heap allocated because the strings are variable
1331 length. The string data (key) and the element object (value) are stored in the
1332 same allocation with the string data immediately after the element object.
1333 This container guarantees the "``(char*)(&Value+1)``" points to the key string
1336 The StringMap is very fast for several reasons: quadratic probing is very cache
1337 efficient for lookups, the hash value of strings in buckets is not recomputed
1338 when looking up an element, StringMap rarely has to touch the memory for
1339 unrelated objects when looking up a value (even when hash collisions happen),
1340 hash table growth does not recompute the hash values for strings already in the
1341 table, and each pair in the map is store in a single allocation (the string data
1342 is stored in the same allocation as the Value of a pair).
1344 StringMap also provides query methods that take byte ranges, so it only ever
1345 copies a string if a value is inserted into the table.
1347 StringMap iteratation order, however, is not guaranteed to be deterministic, so
1348 any uses which require that should instead use a std::map.
1352 llvm/ADT/IndexedMap.h
1353 ^^^^^^^^^^^^^^^^^^^^^
1355 IndexedMap is a specialized container for mapping small dense integers (or
1356 values that can be mapped to small dense integers) to some other type. It is
1357 internally implemented as a vector with a mapping function that maps the keys
1358 to the dense integer range.
1360 This is useful for cases like virtual registers in the LLVM code generator: they
1361 have a dense mapping that is offset by a compile-time constant (the first
1362 virtual register ID).
1369 DenseMap is a simple quadratically probed hash table. It excels at supporting
1370 small keys and values: it uses a single allocation to hold all of the pairs
1371 that are currently inserted in the map. DenseMap is a great way to map
1372 pointers to pointers, or map other small types to each other.
1374 There are several aspects of DenseMap that you should be aware of, however.
1375 The iterators in a DenseMap are invalidated whenever an insertion occurs,
1376 unlike map. Also, because DenseMap allocates space for a large number of
1377 key/value pairs (it starts with 64 by default), it will waste a lot of space if
1378 your keys or values are large. Finally, you must implement a partial
1379 specialization of DenseMapInfo for the key that you want, if it isn't already
1380 supported. This is required to tell DenseMap about two special marker values
1381 (which can never be inserted into the map) that it needs internally.
1383 DenseMap's find_as() method supports lookup operations using an alternate key
1384 type. This is useful in cases where the normal key type is expensive to
1385 construct, but cheap to compare against. The DenseMapInfo is responsible for
1386 defining the appropriate comparison and hashing methods for each alternate key
1394 ValueMap is a wrapper around a :ref:`DenseMap <dss_densemap>` mapping
1395 ``Value*``\ s (or subclasses) to another type. When a Value is deleted or
1396 RAUW'ed, ValueMap will update itself so the new version of the key is mapped to
1397 the same value, just as if the key were a WeakVH. You can configure exactly how
1398 this happens, and what else happens on these two events, by passing a ``Config``
1399 parameter to the ValueMap template.
1401 .. _dss_intervalmap:
1403 llvm/ADT/IntervalMap.h
1404 ^^^^^^^^^^^^^^^^^^^^^^
1406 IntervalMap is a compact map for small keys and values. It maps key intervals
1407 instead of single keys, and it will automatically coalesce adjacent intervals.
1408 When then map only contains a few intervals, they are stored in the map object
1409 itself to avoid allocations.
1411 The IntervalMap iterators are quite big, so they should not be passed around as
1412 STL iterators. The heavyweight iterators allow a smaller data structure.
1419 std::map has similar characteristics to :ref:`std::set <dss_set>`: it uses a
1420 single allocation per pair inserted into the map, it offers log(n) lookup with
1421 an extremely large constant factor, imposes a space penalty of 3 pointers per
1422 pair in the map, etc.
1424 std::map is most useful when your keys or values are very large, if you need to
1425 iterate over the collection in sorted order, or if you need stable iterators
1426 into the map (i.e. they don't get invalidated if an insertion or deletion of
1427 another element takes place).
1431 llvm/ADT/MapVector.h
1432 ^^^^^^^^^^^^^^^^^^^^
1434 ``MapVector<KeyT,ValueT>`` provides a subset of the DenseMap interface. The
1435 main difference is that the iteration order is guaranteed to be the insertion
1436 order, making it an easy (but somewhat expensive) solution for non-deterministic
1437 iteration over maps of pointers.
1439 It is implemented by mapping from key to an index in a vector of key,value
1440 pairs. This provides fast lookup and iteration, but has two main drawbacks:
1441 the key is stored twice and removing elements takes linear time. If it is
1442 necessary to remove elements, it's best to remove them in bulk using
1445 .. _dss_inteqclasses:
1447 llvm/ADT/IntEqClasses.h
1448 ^^^^^^^^^^^^^^^^^^^^^^^
1450 IntEqClasses provides a compact representation of equivalence classes of small
1451 integers. Initially, each integer in the range 0..n-1 has its own equivalence
1452 class. Classes can be joined by passing two class representatives to the
1453 join(a, b) method. Two integers are in the same class when findLeader() returns
1454 the same representative.
1456 Once all equivalence classes are formed, the map can be compressed so each
1457 integer 0..n-1 maps to an equivalence class number in the range 0..m-1, where m
1458 is the total number of equivalence classes. The map must be uncompressed before
1459 it can be edited again.
1461 .. _dss_immutablemap:
1463 llvm/ADT/ImmutableMap.h
1464 ^^^^^^^^^^^^^^^^^^^^^^^
1466 ImmutableMap is an immutable (functional) map implementation based on an AVL
1467 tree. Adding or removing elements is done through a Factory object and results
1468 in the creation of a new ImmutableMap object. If an ImmutableMap already exists
1469 with the given key set, then the existing one is returned; equality is compared
1470 with a FoldingSetNodeID. The time and space complexity of add or remove
1471 operations is logarithmic in the size of the original map.
1475 Other Map-Like Container Options
1476 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1478 The STL provides several other options, such as std::multimap and the various
1479 "hash_map" like containers (whether from C++ TR1 or from the SGI library). We
1480 never use hash_set and unordered_set because they are generally very expensive
1481 (each insertion requires a malloc) and very non-portable.
1483 std::multimap is useful if you want to map a key to multiple values, but has all
1484 the drawbacks of std::map. A sorted vector or some other approach is almost
1489 Bit storage containers (BitVector, SparseBitVector)
1490 ---------------------------------------------------
1492 Unlike the other containers, there are only two bit storage containers, and
1493 choosing when to use each is relatively straightforward.
1495 One additional option is ``std::vector<bool>``: we discourage its use for two
1496 reasons 1) the implementation in many common compilers (e.g. commonly
1497 available versions of GCC) is extremely inefficient and 2) the C++ standards
1498 committee is likely to deprecate this container and/or change it significantly
1499 somehow. In any case, please don't use it.
1506 The BitVector container provides a dynamic size set of bits for manipulation.
1507 It supports individual bit setting/testing, as well as set operations. The set
1508 operations take time O(size of bitvector), but operations are performed one word
1509 at a time, instead of one bit at a time. This makes the BitVector very fast for
1510 set operations compared to other containers. Use the BitVector when you expect
1511 the number of set bits to be high (i.e. a dense set).
1513 .. _dss_smallbitvector:
1518 The SmallBitVector container provides the same interface as BitVector, but it is
1519 optimized for the case where only a small number of bits, less than 25 or so,
1520 are needed. It also transparently supports larger bit counts, but slightly less
1521 efficiently than a plain BitVector, so SmallBitVector should only be used when
1522 larger counts are rare.
1524 At this time, SmallBitVector does not support set operations (and, or, xor), and
1525 its operator[] does not provide an assignable lvalue.
1527 .. _dss_sparsebitvector:
1532 The SparseBitVector container is much like BitVector, with one major difference:
1533 Only the bits that are set, are stored. This makes the SparseBitVector much
1534 more space efficient than BitVector when the set is sparse, as well as making
1535 set operations O(number of set bits) instead of O(size of universe). The
1536 downside to the SparseBitVector is that setting and testing of random bits is
1537 O(N), and on large SparseBitVectors, this can be slower than BitVector. In our
1538 implementation, setting or testing bits in sorted order (either forwards or
1539 reverse) is O(1) worst case. Testing and setting bits within 128 bits (depends
1540 on size) of the current bit is also O(1). As a general statement,
1541 testing/setting bits in a SparseBitVector is O(distance away from last set bit).
1545 Helpful Hints for Common Operations
1546 ===================================
1548 This section describes how to perform some very simple transformations of LLVM
1549 code. This is meant to give examples of common idioms used, showing the
1550 practical side of LLVM transformations.
1552 Because this is a "how-to" section, you should also read about the main classes
1553 that you will be working with. The :ref:`Core LLVM Class Hierarchy Reference
1554 <coreclasses>` contains details and descriptions of the main classes that you
1559 Basic Inspection and Traversal Routines
1560 ---------------------------------------
1562 The LLVM compiler infrastructure have many different data structures that may be
1563 traversed. Following the example of the C++ standard template library, the
1564 techniques used to traverse these various data structures are all basically the
1565 same. For a enumerable sequence of values, the ``XXXbegin()`` function (or
1566 method) returns an iterator to the start of the sequence, the ``XXXend()``
1567 function returns an iterator pointing to one past the last valid element of the
1568 sequence, and there is some ``XXXiterator`` data type that is common between the
1571 Because the pattern for iteration is common across many different aspects of the
1572 program representation, the standard template library algorithms may be used on
1573 them, and it is easier to remember how to iterate. First we show a few common
1574 examples of the data structures that need to be traversed. Other data
1575 structures are traversed in very similar ways.
1577 .. _iterate_function:
1579 Iterating over the ``BasicBlock`` in a ``Function``
1580 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1582 It's quite common to have a ``Function`` instance that you'd like to transform
1583 in some way; in particular, you'd like to manipulate its ``BasicBlock``\ s. To
1584 facilitate this, you'll need to iterate over all of the ``BasicBlock``\ s that
1585 constitute the ``Function``. The following is an example that prints the name
1586 of a ``BasicBlock`` and the number of ``Instruction``\ s it contains:
1590 // func is a pointer to a Function instance
1591 for (Function::iterator i = func->begin(), e = func->end(); i != e; ++i)
1592 // Print out the name of the basic block if it has one, and then the
1593 // number of instructions that it contains
1594 errs() << "Basic block (name=" << i->getName() << ") has "
1595 << i->size() << " instructions.\n";
1597 Note that i can be used as if it were a pointer for the purposes of invoking
1598 member functions of the ``Instruction`` class. This is because the indirection
1599 operator is overloaded for the iterator classes. In the above code, the
1600 expression ``i->size()`` is exactly equivalent to ``(*i).size()`` just like
1603 .. _iterate_basicblock:
1605 Iterating over the ``Instruction`` in a ``BasicBlock``
1606 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1608 Just like when dealing with ``BasicBlock``\ s in ``Function``\ s, it's easy to
1609 iterate over the individual instructions that make up ``BasicBlock``\ s. Here's
1610 a code snippet that prints out each instruction in a ``BasicBlock``:
1614 // blk is a pointer to a BasicBlock instance
1615 for (BasicBlock::iterator i = blk->begin(), e = blk->end(); i != e; ++i)
1616 // The next statement works since operator<<(ostream&,...)
1617 // is overloaded for Instruction&
1618 errs() << *i << "\n";
1621 However, this isn't really the best way to print out the contents of a
1622 ``BasicBlock``! Since the ostream operators are overloaded for virtually
1623 anything you'll care about, you could have just invoked the print routine on the
1624 basic block itself: ``errs() << *blk << "\n";``.
1626 .. _iterate_insiter:
1628 Iterating over the ``Instruction`` in a ``Function``
1629 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1631 If you're finding that you commonly iterate over a ``Function``'s
1632 ``BasicBlock``\ s and then that ``BasicBlock``'s ``Instruction``\ s,
1633 ``InstIterator`` should be used instead. You'll need to include
1634 ``llvm/IR/InstIterator.h`` (`doxygen
1635 <http://llvm.org/doxygen/InstIterator_8h.html>`__) and then instantiate
1636 ``InstIterator``\ s explicitly in your code. Here's a small example that shows
1637 how to dump all instructions in a function to the standard error stream:
1641 #include "llvm/IR/InstIterator.h"
1643 // F is a pointer to a Function instance
1644 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
1645 errs() << *I << "\n";
1647 Easy, isn't it? You can also use ``InstIterator``\ s to fill a work list with
1648 its initial contents. For example, if you wanted to initialize a work list to
1649 contain all instructions in a ``Function`` F, all you would need to do is
1654 std::set<Instruction*> worklist;
1655 // or better yet, SmallPtrSet<Instruction*, 64> worklist;
1657 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I)
1658 worklist.insert(&*I);
1660 The STL set ``worklist`` would now contain all instructions in the ``Function``
1663 .. _iterate_convert:
1665 Turning an iterator into a class pointer (and vice-versa)
1666 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1668 Sometimes, it'll be useful to grab a reference (or pointer) to a class instance
1669 when all you've got at hand is an iterator. Well, extracting a reference or a
1670 pointer from an iterator is very straight-forward. Assuming that ``i`` is a
1671 ``BasicBlock::iterator`` and ``j`` is a ``BasicBlock::const_iterator``:
1675 Instruction& inst = *i; // Grab reference to instruction reference
1676 Instruction* pinst = &*i; // Grab pointer to instruction reference
1677 const Instruction& inst = *j;
1679 However, the iterators you'll be working with in the LLVM framework are special:
1680 they will automatically convert to a ptr-to-instance type whenever they need to.
1681 Instead of derferencing the iterator and then taking the address of the result,
1682 you can simply assign the iterator to the proper pointer type and you get the
1683 dereference and address-of operation as a result of the assignment (behind the
1684 scenes, this is a result of overloading casting mechanisms). Thus the last line
1685 of the last example,
1689 Instruction *pinst = &*i;
1691 is semantically equivalent to
1695 Instruction *pinst = i;
1697 It's also possible to turn a class pointer into the corresponding iterator, and
1698 this is a constant time operation (very efficient). The following code snippet
1699 illustrates use of the conversion constructors provided by LLVM iterators. By
1700 using these, you can explicitly grab the iterator of something without actually
1701 obtaining it via iteration over some structure:
1705 void printNextInstruction(Instruction* inst) {
1706 BasicBlock::iterator it(inst);
1707 ++it; // After this line, it refers to the instruction after *inst
1708 if (it != inst->getParent()->end()) errs() << *it << "\n";
1711 Unfortunately, these implicit conversions come at a cost; they prevent these
1712 iterators from conforming to standard iterator conventions, and thus from being
1713 usable with standard algorithms and containers. For example, they prevent the
1714 following code, where ``B`` is a ``BasicBlock``, from compiling:
1718 llvm::SmallVector<llvm::Instruction *, 16>(B->begin(), B->end());
1720 Because of this, these implicit conversions may be removed some day, and
1721 ``operator*`` changed to return a pointer instead of a reference.
1723 .. _iterate_complex:
1725 Finding call sites: a slightly more complex example
1726 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1728 Say that you're writing a FunctionPass and would like to count all the locations
1729 in the entire module (that is, across every ``Function``) where a certain
1730 function (i.e., some ``Function *``) is already in scope. As you'll learn
1731 later, you may want to use an ``InstVisitor`` to accomplish this in a much more
1732 straight-forward manner, but this example will allow us to explore how you'd do
1733 it if you didn't have ``InstVisitor`` around. In pseudo-code, this is what we
1736 .. code-block:: none
1738 initialize callCounter to zero
1739 for each Function f in the Module
1740 for each BasicBlock b in f
1741 for each Instruction i in b
1742 if (i is a CallInst and calls the given function)
1743 increment callCounter
1745 And the actual code is (remember, because we're writing a ``FunctionPass``, our
1746 ``FunctionPass``-derived class simply has to override the ``runOnFunction``
1751 Function* targetFunc = ...;
1753 class OurFunctionPass : public FunctionPass {
1755 OurFunctionPass(): callCounter(0) { }
1757 virtual runOnFunction(Function& F) {
1758 for (Function::iterator b = F.begin(), be = F.end(); b != be; ++b) {
1759 for (BasicBlock::iterator i = b->begin(), ie = b->end(); i != ie; ++i) {
1760 if (CallInst* callInst = dyn_cast<CallInst>(&*i)) {
1761 // We know we've encountered a call instruction, so we
1762 // need to determine if it's a call to the
1763 // function pointed to by m_func or not.
1764 if (callInst->getCalledFunction() == targetFunc)
1772 unsigned callCounter;
1775 .. _calls_and_invokes:
1777 Treating calls and invokes the same way
1778 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1780 You may have noticed that the previous example was a bit oversimplified in that
1781 it did not deal with call sites generated by 'invoke' instructions. In this,
1782 and in other situations, you may find that you want to treat ``CallInst``\ s and
1783 ``InvokeInst``\ s the same way, even though their most-specific common base
1784 class is ``Instruction``, which includes lots of less closely-related things.
1785 For these cases, LLVM provides a handy wrapper class called ``CallSite``
1786 (`doxygen <http://llvm.org/doxygen/classllvm_1_1CallSite.html>`__) It is
1787 essentially a wrapper around an ``Instruction`` pointer, with some methods that
1788 provide functionality common to ``CallInst``\ s and ``InvokeInst``\ s.
1790 This class has "value semantics": it should be passed by value, not by reference
1791 and it should not be dynamically allocated or deallocated using ``operator new``
1792 or ``operator delete``. It is efficiently copyable, assignable and
1793 constructable, with costs equivalents to that of a bare pointer. If you look at
1794 its definition, it has only a single pointer member.
1798 Iterating over def-use & use-def chains
1799 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1801 Frequently, we might have an instance of the ``Value`` class (`doxygen
1802 <http://llvm.org/doxygen/classllvm_1_1Value.html>`__) and we want to determine
1803 which ``User`` s use the ``Value``. The list of all ``User``\ s of a particular
1804 ``Value`` is called a *def-use* chain. For example, let's say we have a
1805 ``Function*`` named ``F`` to a particular function ``foo``. Finding all of the
1806 instructions that *use* ``foo`` is as simple as iterating over the *def-use*
1813 for (User *U : GV->users()) {
1814 if (Instruction *Inst = dyn_cast<Instruction>(U)) {
1815 errs() << "F is used in instruction:\n";
1816 errs() << *Inst << "\n";
1819 Alternatively, it's common to have an instance of the ``User`` Class (`doxygen
1820 <http://llvm.org/doxygen/classllvm_1_1User.html>`__) and need to know what
1821 ``Value``\ s are used by it. The list of all ``Value``\ s used by a ``User`` is
1822 known as a *use-def* chain. Instances of class ``Instruction`` are common
1823 ``User`` s, so we might want to iterate over all of the values that a particular
1824 instruction uses (that is, the operands of the particular ``Instruction``):
1828 Instruction *pi = ...;
1830 for (Use &U : pi->operands()) {
1835 Declaring objects as ``const`` is an important tool of enforcing mutation free
1836 algorithms (such as analyses, etc.). For this purpose above iterators come in
1837 constant flavors as ``Value::const_use_iterator`` and
1838 ``Value::const_op_iterator``. They automatically arise when calling
1839 ``use/op_begin()`` on ``const Value*``\ s or ``const User*``\ s respectively.
1840 Upon dereferencing, they return ``const Use*``\ s. Otherwise the above patterns
1845 Iterating over predecessors & successors of blocks
1846 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1848 Iterating over the predecessors and successors of a block is quite easy with the
1849 routines defined in ``"llvm/Support/CFG.h"``. Just use code like this to
1850 iterate over all predecessors of BB:
1854 #include "llvm/Support/CFG.h"
1855 BasicBlock *BB = ...;
1857 for (BasicBlock *Pred : predecessors(BB)) {
1861 Similarly, to iterate over successors use ``successors``.
1865 Making simple changes
1866 ---------------------
1868 There are some primitive transformation operations present in the LLVM
1869 infrastructure that are worth knowing about. When performing transformations,
1870 it's fairly common to manipulate the contents of basic blocks. This section
1871 describes some of the common methods for doing so and gives example code.
1873 .. _schanges_creating:
1875 Creating and inserting new ``Instruction``\ s
1876 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
1878 *Instantiating Instructions*
1880 Creation of ``Instruction``\ s is straight-forward: simply call the constructor
1881 for the kind of instruction to instantiate and provide the necessary parameters.
1882 For example, an ``AllocaInst`` only *requires* a (const-ptr-to) ``Type``. Thus:
1886 AllocaInst* ai = new AllocaInst(Type::Int32Ty);
1888 will create an ``AllocaInst`` instance that represents the allocation of one
1889 integer in the current stack frame, at run time. Each ``Instruction`` subclass
1890 is likely to have varying default parameters which change the semantics of the
1891 instruction, so refer to the `doxygen documentation for the subclass of
1892 Instruction <http://llvm.org/doxygen/classllvm_1_1Instruction.html>`_ that
1893 you're interested in instantiating.
1897 It is very useful to name the values of instructions when you're able to, as
1898 this facilitates the debugging of your transformations. If you end up looking
1899 at generated LLVM machine code, you definitely want to have logical names
1900 associated with the results of instructions! By supplying a value for the
1901 ``Name`` (default) parameter of the ``Instruction`` constructor, you associate a
1902 logical name with the result of the instruction's execution at run time. For
1903 example, say that I'm writing a transformation that dynamically allocates space
1904 for an integer on the stack, and that integer is going to be used as some kind
1905 of index by some other code. To accomplish this, I place an ``AllocaInst`` at
1906 the first point in the first ``BasicBlock`` of some ``Function``, and I'm
1907 intending to use it within the same ``Function``. I might do:
1911 AllocaInst* pa = new AllocaInst(Type::Int32Ty, 0, "indexLoc");
1913 where ``indexLoc`` is now the logical name of the instruction's execution value,
1914 which is a pointer to an integer on the run time stack.
1916 *Inserting instructions*
1918 There are essentially three ways to insert an ``Instruction`` into an existing
1919 sequence of instructions that form a ``BasicBlock``:
1921 * Insertion into an explicit instruction list
1923 Given a ``BasicBlock* pb``, an ``Instruction* pi`` within that ``BasicBlock``,
1924 and a newly-created instruction we wish to insert before ``*pi``, we do the
1929 BasicBlock *pb = ...;
1930 Instruction *pi = ...;
1931 Instruction *newInst = new Instruction(...);
1933 pb->getInstList().insert(pi, newInst); // Inserts newInst before pi in pb
1935 Appending to the end of a ``BasicBlock`` is so common that the ``Instruction``
1936 class and ``Instruction``-derived classes provide constructors which take a
1937 pointer to a ``BasicBlock`` to be appended to. For example code that looked
1942 BasicBlock *pb = ...;
1943 Instruction *newInst = new Instruction(...);
1945 pb->getInstList().push_back(newInst); // Appends newInst to pb
1951 BasicBlock *pb = ...;
1952 Instruction *newInst = new Instruction(..., pb);
1954 which is much cleaner, especially if you are creating long instruction
1957 * Insertion into an implicit instruction list
1959 ``Instruction`` instances that are already in ``BasicBlock``\ s are implicitly
1960 associated with an existing instruction list: the instruction list of the
1961 enclosing basic block. Thus, we could have accomplished the same thing as the
1962 above code without being given a ``BasicBlock`` by doing:
1966 Instruction *pi = ...;
1967 Instruction *newInst = new Instruction(...);
1969 pi->getParent()->getInstList().insert(pi, newInst);
1971 In fact, this sequence of steps occurs so frequently that the ``Instruction``
1972 class and ``Instruction``-derived classes provide constructors which take (as
1973 a default parameter) a pointer to an ``Instruction`` which the newly-created
1974 ``Instruction`` should precede. That is, ``Instruction`` constructors are
1975 capable of inserting the newly-created instance into the ``BasicBlock`` of a
1976 provided instruction, immediately before that instruction. Using an
1977 ``Instruction`` constructor with a ``insertBefore`` (default) parameter, the
1982 Instruction* pi = ...;
1983 Instruction* newInst = new Instruction(..., pi);
1985 which is much cleaner, especially if you're creating a lot of instructions and
1986 adding them to ``BasicBlock``\ s.
1988 * Insertion using an instance of ``IRBuilder``
1990 Inserting several ``Instruction``\ s can be quite laborious using the previous
1991 methods. The ``IRBuilder`` is a convenience class that can be used to add
1992 several instructions to the end of a ``BasicBlock`` or before a particular
1993 ``Instruction``. It also supports constant folding and renaming named
1994 registers (see ``IRBuilder``'s template arguments).
1996 The example below demonstrates a very simple use of the ``IRBuilder`` where
1997 three instructions are inserted before the instruction ``pi``. The first two
1998 instructions are Call instructions and third instruction multiplies the return
1999 value of the two calls.
2003 Instruction *pi = ...;
2004 IRBuilder<> Builder(pi);
2005 CallInst* callOne = Builder.CreateCall(...);
2006 CallInst* callTwo = Builder.CreateCall(...);
2007 Value* result = Builder.CreateMul(callOne, callTwo);
2009 The example below is similar to the above example except that the created
2010 ``IRBuilder`` inserts instructions at the end of the ``BasicBlock`` ``pb``.
2014 BasicBlock *pb = ...;
2015 IRBuilder<> Builder(pb);
2016 CallInst* callOne = Builder.CreateCall(...);
2017 CallInst* callTwo = Builder.CreateCall(...);
2018 Value* result = Builder.CreateMul(callOne, callTwo);
2020 See :doc:`tutorial/LangImpl3` for a practical use of the ``IRBuilder``.
2023 .. _schanges_deleting:
2025 Deleting Instructions
2026 ^^^^^^^^^^^^^^^^^^^^^
2028 Deleting an instruction from an existing sequence of instructions that form a
2029 BasicBlock_ is very straight-forward: just call the instruction's
2030 ``eraseFromParent()`` method. For example:
2034 Instruction *I = .. ;
2035 I->eraseFromParent();
2037 This unlinks the instruction from its containing basic block and deletes it. If
2038 you'd just like to unlink the instruction from its containing basic block but
2039 not delete it, you can use the ``removeFromParent()`` method.
2041 .. _schanges_replacing:
2043 Replacing an Instruction with another Value
2044 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2046 Replacing individual instructions
2047 """""""""""""""""""""""""""""""""
2049 Including "`llvm/Transforms/Utils/BasicBlockUtils.h
2050 <http://llvm.org/doxygen/BasicBlockUtils_8h-source.html>`_" permits use of two
2051 very useful replace functions: ``ReplaceInstWithValue`` and
2052 ``ReplaceInstWithInst``.
2054 .. _schanges_deleting_sub:
2056 Deleting Instructions
2057 """""""""""""""""""""
2059 * ``ReplaceInstWithValue``
2061 This function replaces all uses of a given instruction with a value, and then
2062 removes the original instruction. The following example illustrates the
2063 replacement of the result of a particular ``AllocaInst`` that allocates memory
2064 for a single integer with a null pointer to an integer.
2068 AllocaInst* instToReplace = ...;
2069 BasicBlock::iterator ii(instToReplace);
2071 ReplaceInstWithValue(instToReplace->getParent()->getInstList(), ii,
2072 Constant::getNullValue(PointerType::getUnqual(Type::Int32Ty)));
2074 * ``ReplaceInstWithInst``
2076 This function replaces a particular instruction with another instruction,
2077 inserting the new instruction into the basic block at the location where the
2078 old instruction was, and replacing any uses of the old instruction with the
2079 new instruction. The following example illustrates the replacement of one
2080 ``AllocaInst`` with another.
2084 AllocaInst* instToReplace = ...;
2085 BasicBlock::iterator ii(instToReplace);
2087 ReplaceInstWithInst(instToReplace->getParent()->getInstList(), ii,
2088 new AllocaInst(Type::Int32Ty, 0, "ptrToReplacedInt"));
2091 Replacing multiple uses of Users and Values
2092 """""""""""""""""""""""""""""""""""""""""""
2094 You can use ``Value::replaceAllUsesWith`` and ``User::replaceUsesOfWith`` to
2095 change more than one use at a time. See the doxygen documentation for the
2096 `Value Class <http://llvm.org/doxygen/classllvm_1_1Value.html>`_ and `User Class
2097 <http://llvm.org/doxygen/classllvm_1_1User.html>`_, respectively, for more
2100 .. _schanges_deletingGV:
2102 Deleting GlobalVariables
2103 ^^^^^^^^^^^^^^^^^^^^^^^^
2105 Deleting a global variable from a module is just as easy as deleting an
2106 Instruction. First, you must have a pointer to the global variable that you
2107 wish to delete. You use this pointer to erase it from its parent, the module.
2112 GlobalVariable *GV = .. ;
2114 GV->eraseFromParent();
2122 In generating IR, you may need some complex types. If you know these types
2123 statically, you can use ``TypeBuilder<...>::get()``, defined in
2124 ``llvm/Support/TypeBuilder.h``, to retrieve them. ``TypeBuilder`` has two forms
2125 depending on whether you're building types for cross-compilation or native
2126 library use. ``TypeBuilder<T, true>`` requires that ``T`` be independent of the
2127 host environment, meaning that it's built out of types from the ``llvm::types``
2128 (`doxygen <http://llvm.org/doxygen/namespacellvm_1_1types.html>`__) namespace
2129 and pointers, functions, arrays, etc. built of those. ``TypeBuilder<T, false>``
2130 additionally allows native C types whose size may depend on the host compiler.
2135 FunctionType *ft = TypeBuilder<types::i<8>(types::i<32>*), true>::get();
2137 is easier to read and write than the equivalent
2141 std::vector<const Type*> params;
2142 params.push_back(PointerType::getUnqual(Type::Int32Ty));
2143 FunctionType *ft = FunctionType::get(Type::Int8Ty, params, false);
2145 See the `class comment
2146 <http://llvm.org/doxygen/TypeBuilder_8h-source.html#l00001>`_ for more details.
2153 This section describes the interaction of the LLVM APIs with multithreading,
2154 both on the part of client applications, and in the JIT, in the hosted
2157 Note that LLVM's support for multithreading is still relatively young. Up
2158 through version 2.5, the execution of threaded hosted applications was
2159 supported, but not threaded client access to the APIs. While this use case is
2160 now supported, clients *must* adhere to the guidelines specified below to ensure
2161 proper operation in multithreaded mode.
2163 Note that, on Unix-like platforms, LLVM requires the presence of GCC's atomic
2164 intrinsics in order to support threaded operation. If you need a
2165 multhreading-capable LLVM on a platform without a suitably modern system
2166 compiler, consider compiling LLVM and LLVM-GCC in single-threaded mode, and
2167 using the resultant compiler to build a copy of LLVM with multithreading
2172 Ending Execution with ``llvm_shutdown()``
2173 -----------------------------------------
2175 When you are done using the LLVM APIs, you should call ``llvm_shutdown()`` to
2176 deallocate memory used for internal structures.
2180 Lazy Initialization with ``ManagedStatic``
2181 ------------------------------------------
2183 ``ManagedStatic`` is a utility class in LLVM used to implement static
2184 initialization of static resources, such as the global type tables. In a
2185 single-threaded environment, it implements a simple lazy initialization scheme.
2186 When LLVM is compiled with support for multi-threading, however, it uses
2187 double-checked locking to implement thread-safe lazy initialization.
2191 Achieving Isolation with ``LLVMContext``
2192 ----------------------------------------
2194 ``LLVMContext`` is an opaque class in the LLVM API which clients can use to
2195 operate multiple, isolated instances of LLVM concurrently within the same
2196 address space. For instance, in a hypothetical compile-server, the compilation
2197 of an individual translation unit is conceptually independent from all the
2198 others, and it would be desirable to be able to compile incoming translation
2199 units concurrently on independent server threads. Fortunately, ``LLVMContext``
2200 exists to enable just this kind of scenario!
2202 Conceptually, ``LLVMContext`` provides isolation. Every LLVM entity
2203 (``Module``\ s, ``Value``\ s, ``Type``\ s, ``Constant``\ s, etc.) in LLVM's
2204 in-memory IR belongs to an ``LLVMContext``. Entities in different contexts
2205 *cannot* interact with each other: ``Module``\ s in different contexts cannot be
2206 linked together, ``Function``\ s cannot be added to ``Module``\ s in different
2207 contexts, etc. What this means is that is is safe to compile on multiple
2208 threads simultaneously, as long as no two threads operate on entities within the
2211 In practice, very few places in the API require the explicit specification of a
2212 ``LLVMContext``, other than the ``Type`` creation/lookup APIs. Because every
2213 ``Type`` carries a reference to its owning context, most other entities can
2214 determine what context they belong to by looking at their own ``Type``. If you
2215 are adding new entities to LLVM IR, please try to maintain this interface
2218 For clients that do *not* require the benefits of isolation, LLVM provides a
2219 convenience API ``getGlobalContext()``. This returns a global, lazily
2220 initialized ``LLVMContext`` that may be used in situations where isolation is
2228 LLVM's "eager" JIT compiler is safe to use in threaded programs. Multiple
2229 threads can call ``ExecutionEngine::getPointerToFunction()`` or
2230 ``ExecutionEngine::runFunction()`` concurrently, and multiple threads can run
2231 code output by the JIT concurrently. The user must still ensure that only one
2232 thread accesses IR in a given ``LLVMContext`` while another thread might be
2233 modifying it. One way to do that is to always hold the JIT lock while accessing
2234 IR outside the JIT (the JIT *modifies* the IR by adding ``CallbackVH``\ s).
2235 Another way is to only call ``getPointerToFunction()`` from the
2236 ``LLVMContext``'s thread.
2238 When the JIT is configured to compile lazily (using
2239 ``ExecutionEngine::DisableLazyCompilation(false)``), there is currently a `race
2240 condition <http://llvm.org/bugs/show_bug.cgi?id=5184>`_ in updating call sites
2241 after a function is lazily-jitted. It's still possible to use the lazy JIT in a
2242 threaded program if you ensure that only one thread at a time can call any
2243 particular lazy stub and that the JIT lock guards any IR access, but we suggest
2244 using only the eager JIT in threaded programs.
2251 This section describes some of the advanced or obscure API's that most clients
2252 do not need to be aware of. These API's tend manage the inner workings of the
2253 LLVM system, and only need to be accessed in unusual circumstances.
2257 The ``ValueSymbolTable`` class
2258 ------------------------------
2260 The ``ValueSymbolTable`` (`doxygen
2261 <http://llvm.org/doxygen/classllvm_1_1ValueSymbolTable.html>`__) class provides
2262 a symbol table that the :ref:`Function <c_Function>` and Module_ classes use for
2263 naming value definitions. The symbol table can provide a name for any Value_.
2265 Note that the ``SymbolTable`` class should not be directly accessed by most
2266 clients. It should only be used when iteration over the symbol table names
2267 themselves are required, which is very special purpose. Note that not all LLVM
2268 Value_\ s have names, and those without names (i.e. they have an empty name) do
2269 not exist in the symbol table.
2271 Symbol tables support iteration over the values in the symbol table with
2272 ``begin/end/iterator`` and supports querying to see if a specific name is in the
2273 symbol table (with ``lookup``). The ``ValueSymbolTable`` class exposes no
2274 public mutator methods, instead, simply call ``setName`` on a value, which will
2275 autoinsert it into the appropriate symbol table.
2279 The ``User`` and owned ``Use`` classes' memory layout
2280 -----------------------------------------------------
2282 The ``User`` (`doxygen <http://llvm.org/doxygen/classllvm_1_1User.html>`__)
2283 class provides a basis for expressing the ownership of ``User`` towards other
2284 `Value instance <http://llvm.org/doxygen/classllvm_1_1Value.html>`_\ s. The
2285 ``Use`` (`doxygen <http://llvm.org/doxygen/classllvm_1_1Use.html>`__) helper
2286 class is employed to do the bookkeeping and to facilitate *O(1)* addition and
2291 Interaction and relationship between ``User`` and ``Use`` objects
2292 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2294 A subclass of ``User`` can choose between incorporating its ``Use`` objects or
2295 refer to them out-of-line by means of a pointer. A mixed variant (some ``Use``
2296 s inline others hung off) is impractical and breaks the invariant that the
2297 ``Use`` objects belonging to the same ``User`` form a contiguous array.
2299 We have 2 different layouts in the ``User`` (sub)classes:
2303 The ``Use`` object(s) are inside (resp. at fixed offset) of the ``User``
2304 object and there are a fixed number of them.
2308 The ``Use`` object(s) are referenced by a pointer to an array from the
2309 ``User`` object and there may be a variable number of them.
2311 As of v2.4 each layout still possesses a direct pointer to the start of the
2312 array of ``Use``\ s. Though not mandatory for layout a), we stick to this
2313 redundancy for the sake of simplicity. The ``User`` object also stores the
2314 number of ``Use`` objects it has. (Theoretically this information can also be
2315 calculated given the scheme presented below.)
2317 Special forms of allocation operators (``operator new``) enforce the following
2320 * Layout a) is modelled by prepending the ``User`` object by the ``Use[]``
2323 .. code-block:: none
2325 ...---.---.---.---.-------...
2326 | P | P | P | P | User
2327 '''---'---'---'---'-------'''
2329 * Layout b) is modelled by pointing at the ``Use[]`` array.
2331 .. code-block:: none
2342 *(In the above figures* '``P``' *stands for the* ``Use**`` *that is stored in
2343 each* ``Use`` *object in the member* ``Use::Prev`` *)*
2347 The waymarking algorithm
2348 ^^^^^^^^^^^^^^^^^^^^^^^^
2350 Since the ``Use`` objects are deprived of the direct (back)pointer to their
2351 ``User`` objects, there must be a fast and exact method to recover it. This is
2352 accomplished by the following scheme:
2354 A bit-encoding in the 2 LSBits (least significant bits) of the ``Use::Prev``
2355 allows to find the start of the ``User`` object:
2357 * ``00`` --- binary digit 0
2359 * ``01`` --- binary digit 1
2361 * ``10`` --- stop and calculate (``s``)
2363 * ``11`` --- full stop (``S``)
2365 Given a ``Use*``, all we have to do is to walk till we get a stop and we either
2366 have a ``User`` immediately behind or we have to walk to the next stop picking
2367 up digits and calculating the offset:
2369 .. code-block:: none
2371 .---.---.---.---.---.---.---.---.---.---.---.---.---.---.---.---.----------------
2372 | 1 | s | 1 | 0 | 1 | 0 | s | 1 | 1 | 0 | s | 1 | 1 | s | 1 | S | User (or User*)
2373 '---'---'---'---'---'---'---'---'---'---'---'---'---'---'---'---'----------------
2374 |+15 |+10 |+6 |+3 |+1
2377 | | | ______________________>
2378 | | ______________________________________>
2379 | __________________________________________________________>
2381 Only the significant number of bits need to be stored between the stops, so that
2382 the *worst case is 20 memory accesses* when there are 1000 ``Use`` objects
2383 associated with a ``User``.
2387 Reference implementation
2388 ^^^^^^^^^^^^^^^^^^^^^^^^
2390 The following literate Haskell fragment demonstrates the concept:
2392 .. code-block:: haskell
2394 > import Test.QuickCheck
2396 > digits :: Int -> [Char] -> [Char]
2397 > digits 0 acc = '0' : acc
2398 > digits 1 acc = '1' : acc
2399 > digits n acc = digits (n `div` 2) $ digits (n `mod` 2) acc
2401 > dist :: Int -> [Char] -> [Char]
2404 > dist 1 acc = let r = dist 0 acc in 's' : digits (length r) r
2405 > dist n acc = dist (n - 1) $ dist 1 acc
2407 > takeLast n ss = reverse $ take n $ reverse ss
2409 > test = takeLast 40 $ dist 20 []
2412 Printing <test> gives: ``"1s100000s11010s10100s1111s1010s110s11s1S"``
2414 The reverse algorithm computes the length of the string just by examining a
2417 .. code-block:: haskell
2419 > pref :: [Char] -> Int
2421 > pref ('s':'1':rest) = decode 2 1 rest
2422 > pref (_:rest) = 1 + pref rest
2424 > decode walk acc ('0':rest) = decode (walk + 1) (acc * 2) rest
2425 > decode walk acc ('1':rest) = decode (walk + 1) (acc * 2 + 1) rest
2426 > decode walk acc _ = walk + acc
2429 Now, as expected, printing <pref test> gives ``40``.
2431 We can *quickCheck* this with following property:
2433 .. code-block:: haskell
2435 > testcase = dist 2000 []
2436 > testcaseLength = length testcase
2438 > identityProp n = n > 0 && n <= testcaseLength ==> length arr == pref arr
2439 > where arr = takeLast n testcase
2442 As expected <quickCheck identityProp> gives:
2446 *Main> quickCheck identityProp
2447 OK, passed 100 tests.
2449 Let's be a bit more exhaustive:
2451 .. code-block:: haskell
2454 > deepCheck p = check (defaultConfig { configMaxTest = 500 }) p
2457 And here is the result of <deepCheck identityProp>:
2461 *Main> deepCheck identityProp
2462 OK, passed 500 tests.
2466 Tagging considerations
2467 ^^^^^^^^^^^^^^^^^^^^^^
2469 To maintain the invariant that the 2 LSBits of each ``Use**`` in ``Use`` never
2470 change after being set up, setters of ``Use::Prev`` must re-tag the new
2471 ``Use**`` on every modification. Accordingly getters must strip the tag bits.
2473 For layout b) instead of the ``User`` we find a pointer (``User*`` with LSBit
2474 set). Following this pointer brings us to the ``User``. A portable trick
2475 ensures that the first bytes of ``User`` (if interpreted as a pointer) never has
2476 the LSBit set. (Portability is relying on the fact that all known compilers
2477 place the ``vptr`` in the first word of the instances.)
2481 The Core LLVM Class Hierarchy Reference
2482 =======================================
2484 ``#include "llvm/IR/Type.h"``
2486 header source: `Type.h <http://llvm.org/doxygen/Type_8h-source.html>`_
2488 doxygen info: `Type Clases <http://llvm.org/doxygen/classllvm_1_1Type.html>`_
2490 The Core LLVM classes are the primary means of representing the program being
2491 inspected or transformed. The core LLVM classes are defined in header files in
2492 the ``include/llvm/`` directory, and implemented in the ``lib/VMCore``
2497 The Type class and Derived Types
2498 --------------------------------
2500 ``Type`` is a superclass of all type classes. Every ``Value`` has a ``Type``.
2501 ``Type`` cannot be instantiated directly but only through its subclasses.
2502 Certain primitive types (``VoidType``, ``LabelType``, ``FloatType`` and
2503 ``DoubleType``) have hidden subclasses. They are hidden because they offer no
2504 useful functionality beyond what the ``Type`` class offers except to distinguish
2505 themselves from other subclasses of ``Type``.
2507 All other types are subclasses of ``DerivedType``. Types can be named, but this
2508 is not a requirement. There exists exactly one instance of a given shape at any
2509 one time. This allows type equality to be performed with address equality of
2510 the Type Instance. That is, given two ``Type*`` values, the types are identical
2511 if the pointers are identical.
2515 Important Public Methods
2516 ^^^^^^^^^^^^^^^^^^^^^^^^
2518 * ``bool isIntegerTy() const``: Returns true for any integer type.
2520 * ``bool isFloatingPointTy()``: Return true if this is one of the five
2521 floating point types.
2523 * ``bool isSized()``: Return true if the type has known size. Things
2524 that don't have a size are abstract types, labels and void.
2528 Important Derived Types
2529 ^^^^^^^^^^^^^^^^^^^^^^^
2532 Subclass of DerivedType that represents integer types of any bit width. Any
2533 bit width between ``IntegerType::MIN_INT_BITS`` (1) and
2534 ``IntegerType::MAX_INT_BITS`` (~8 million) can be represented.
2536 * ``static const IntegerType* get(unsigned NumBits)``: get an integer
2537 type of a specific bit width.
2539 * ``unsigned getBitWidth() const``: Get the bit width of an integer type.
2542 This is subclassed by ArrayType, PointerType and VectorType.
2544 * ``const Type * getElementType() const``: Returns the type of each
2545 of the elements in the sequential type.
2548 This is a subclass of SequentialType and defines the interface for array
2551 * ``unsigned getNumElements() const``: Returns the number of elements
2555 Subclass of SequentialType for pointer types.
2558 Subclass of SequentialType for vector types. A vector type is similar to an
2559 ArrayType but is distinguished because it is a first class type whereas
2560 ArrayType is not. Vector types are used for vector operations and are usually
2561 small vectors of of an integer or floating point type.
2564 Subclass of DerivedTypes for struct types.
2569 Subclass of DerivedTypes for function types.
2571 * ``bool isVarArg() const``: Returns true if it's a vararg function.
2573 * ``const Type * getReturnType() const``: Returns the return type of the
2576 * ``const Type * getParamType (unsigned i)``: Returns the type of the ith
2579 * ``const unsigned getNumParams() const``: Returns the number of formal
2584 The ``Module`` class
2585 --------------------
2587 ``#include "llvm/IR/Module.h"``
2589 header source: `Module.h <http://llvm.org/doxygen/Module_8h-source.html>`_
2591 doxygen info: `Module Class <http://llvm.org/doxygen/classllvm_1_1Module.html>`_
2593 The ``Module`` class represents the top level structure present in LLVM
2594 programs. An LLVM module is effectively either a translation unit of the
2595 original program or a combination of several translation units merged by the
2596 linker. The ``Module`` class keeps track of a list of :ref:`Function
2597 <c_Function>`\ s, a list of GlobalVariable_\ s, and a SymbolTable_.
2598 Additionally, it contains a few helpful member functions that try to make common
2603 Important Public Members of the ``Module`` class
2604 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2606 * ``Module::Module(std::string name = "")``
2608 Constructing a Module_ is easy. You can optionally provide a name for it
2609 (probably based on the name of the translation unit).
2611 * | ``Module::iterator`` - Typedef for function list iterator
2612 | ``Module::const_iterator`` - Typedef for const_iterator.
2613 | ``begin()``, ``end()``, ``size()``, ``empty()``
2615 These are forwarding methods that make it easy to access the contents of a
2616 ``Module`` object's :ref:`Function <c_Function>` list.
2618 * ``Module::FunctionListType &getFunctionList()``
2620 Returns the list of :ref:`Function <c_Function>`\ s. This is necessary to use
2621 when you need to update the list or perform a complex action that doesn't have
2622 a forwarding method.
2626 * | ``Module::global_iterator`` - Typedef for global variable list iterator
2627 | ``Module::const_global_iterator`` - Typedef for const_iterator.
2628 | ``global_begin()``, ``global_end()``, ``global_size()``, ``global_empty()``
2630 These are forwarding methods that make it easy to access the contents of a
2631 ``Module`` object's GlobalVariable_ list.
2633 * ``Module::GlobalListType &getGlobalList()``
2635 Returns the list of GlobalVariable_\ s. This is necessary to use when you
2636 need to update the list or perform a complex action that doesn't have a
2641 * ``SymbolTable *getSymbolTable()``
2643 Return a reference to the SymbolTable_ for this ``Module``.
2647 * ``Function *getFunction(StringRef Name) const``
2649 Look up the specified function in the ``Module`` SymbolTable_. If it does not
2650 exist, return ``null``.
2652 * ``Function *getOrInsertFunction(const std::string &Name, const FunctionType
2655 Look up the specified function in the ``Module`` SymbolTable_. If it does not
2656 exist, add an external declaration for the function and return it.
2658 * ``std::string getTypeName(const Type *Ty)``
2660 If there is at least one entry in the SymbolTable_ for the specified Type_,
2661 return it. Otherwise return the empty string.
2663 * ``bool addTypeName(const std::string &Name, const Type *Ty)``
2665 Insert an entry in the SymbolTable_ mapping ``Name`` to ``Ty``. If there is
2666 already an entry for this name, true is returned and the SymbolTable_ is not
2674 ``#include "llvm/IR/Value.h"``
2676 header source: `Value.h <http://llvm.org/doxygen/Value_8h-source.html>`_
2678 doxygen info: `Value Class <http://llvm.org/doxygen/classllvm_1_1Value.html>`_
2680 The ``Value`` class is the most important class in the LLVM Source base. It
2681 represents a typed value that may be used (among other things) as an operand to
2682 an instruction. There are many different types of ``Value``\ s, such as
2683 Constant_\ s, Argument_\ s. Even Instruction_\ s and :ref:`Function
2684 <c_Function>`\ s are ``Value``\ s.
2686 A particular ``Value`` may be used many times in the LLVM representation for a
2687 program. For example, an incoming argument to a function (represented with an
2688 instance of the Argument_ class) is "used" by every instruction in the function
2689 that references the argument. To keep track of this relationship, the ``Value``
2690 class keeps a list of all of the ``User``\ s that is using it (the User_ class
2691 is a base class for all nodes in the LLVM graph that can refer to ``Value``\ s).
2692 This use list is how LLVM represents def-use information in the program, and is
2693 accessible through the ``use_*`` methods, shown below.
2695 Because LLVM is a typed representation, every LLVM ``Value`` is typed, and this
2696 Type_ is available through the ``getType()`` method. In addition, all LLVM
2697 values can be named. The "name" of the ``Value`` is a symbolic string printed
2700 .. code-block:: llvm
2706 The name of this instruction is "foo". **NOTE** that the name of any value may
2707 be missing (an empty string), so names should **ONLY** be used for debugging
2708 (making the source code easier to read, debugging printouts), they should not be
2709 used to keep track of values or map between them. For this purpose, use a
2710 ``std::map`` of pointers to the ``Value`` itself instead.
2712 One important aspect of LLVM is that there is no distinction between an SSA
2713 variable and the operation that produces it. Because of this, any reference to
2714 the value produced by an instruction (or the value available as an incoming
2715 argument, for example) is represented as a direct pointer to the instance of the
2716 class that represents this value. Although this may take some getting used to,
2717 it simplifies the representation and makes it easier to manipulate.
2721 Important Public Members of the ``Value`` class
2722 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2724 * | ``Value::use_iterator`` - Typedef for iterator over the use-list
2725 | ``Value::const_use_iterator`` - Typedef for const_iterator over the
2727 | ``unsigned use_size()`` - Returns the number of users of the value.
2728 | ``bool use_empty()`` - Returns true if there are no users.
2729 | ``use_iterator use_begin()`` - Get an iterator to the start of the
2731 | ``use_iterator use_end()`` - Get an iterator to the end of the use-list.
2732 | ``User *use_back()`` - Returns the last element in the list.
2734 These methods are the interface to access the def-use information in LLVM.
2735 As with all other iterators in LLVM, the naming conventions follow the
2736 conventions defined by the STL_.
2738 * ``Type *getType() const``
2739 This method returns the Type of the Value.
2741 * | ``bool hasName() const``
2742 | ``std::string getName() const``
2743 | ``void setName(const std::string &Name)``
2745 This family of methods is used to access and assign a name to a ``Value``, be
2746 aware of the :ref:`precaution above <nameWarning>`.
2748 * ``void replaceAllUsesWith(Value *V)``
2750 This method traverses the use list of a ``Value`` changing all User_\ s of the
2751 current value to refer to "``V``" instead. For example, if you detect that an
2752 instruction always produces a constant value (for example through constant
2753 folding), you can replace all uses of the instruction with the constant like
2758 Inst->replaceAllUsesWith(ConstVal);
2765 ``#include "llvm/IR/User.h"``
2767 header source: `User.h <http://llvm.org/doxygen/User_8h-source.html>`_
2769 doxygen info: `User Class <http://llvm.org/doxygen/classllvm_1_1User.html>`_
2773 The ``User`` class is the common base class of all LLVM nodes that may refer to
2774 ``Value``\ s. It exposes a list of "Operands" that are all of the ``Value``\ s
2775 that the User is referring to. The ``User`` class itself is a subclass of
2778 The operands of a ``User`` point directly to the LLVM ``Value`` that it refers
2779 to. Because LLVM uses Static Single Assignment (SSA) form, there can only be
2780 one definition referred to, allowing this direct connection. This connection
2781 provides the use-def information in LLVM.
2785 Important Public Members of the ``User`` class
2786 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2788 The ``User`` class exposes the operand list in two ways: through an index access
2789 interface and through an iterator based interface.
2791 * | ``Value *getOperand(unsigned i)``
2792 | ``unsigned getNumOperands()``
2794 These two methods expose the operands of the ``User`` in a convenient form for
2797 * | ``User::op_iterator`` - Typedef for iterator over the operand list
2798 | ``op_iterator op_begin()`` - Get an iterator to the start of the operand
2800 | ``op_iterator op_end()`` - Get an iterator to the end of the operand list.
2802 Together, these methods make up the iterator based interface to the operands
2808 The ``Instruction`` class
2809 -------------------------
2811 ``#include "llvm/IR/Instruction.h"``
2813 header source: `Instruction.h
2814 <http://llvm.org/doxygen/Instruction_8h-source.html>`_
2816 doxygen info: `Instruction Class
2817 <http://llvm.org/doxygen/classllvm_1_1Instruction.html>`_
2819 Superclasses: User_, Value_
2821 The ``Instruction`` class is the common base class for all LLVM instructions.
2822 It provides only a few methods, but is a very commonly used class. The primary
2823 data tracked by the ``Instruction`` class itself is the opcode (instruction
2824 type) and the parent BasicBlock_ the ``Instruction`` is embedded into. To
2825 represent a specific type of instruction, one of many subclasses of
2826 ``Instruction`` are used.
2828 Because the ``Instruction`` class subclasses the User_ class, its operands can
2829 be accessed in the same way as for other ``User``\ s (with the
2830 ``getOperand()``/``getNumOperands()`` and ``op_begin()``/``op_end()`` methods).
2831 An important file for the ``Instruction`` class is the ``llvm/Instruction.def``
2832 file. This file contains some meta-data about the various different types of
2833 instructions in LLVM. It describes the enum values that are used as opcodes
2834 (for example ``Instruction::Add`` and ``Instruction::ICmp``), as well as the
2835 concrete sub-classes of ``Instruction`` that implement the instruction (for
2836 example BinaryOperator_ and CmpInst_). Unfortunately, the use of macros in this
2837 file confuses doxygen, so these enum values don't show up correctly in the
2838 `doxygen output <http://llvm.org/doxygen/classllvm_1_1Instruction.html>`_.
2842 Important Subclasses of the ``Instruction`` class
2843 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2847 * ``BinaryOperator``
2849 This subclasses represents all two operand instructions whose operands must be
2850 the same type, except for the comparison instructions.
2855 This subclass is the parent of the 12 casting instructions. It provides
2856 common operations on cast instructions.
2862 This subclass respresents the two comparison instructions,
2863 `ICmpInst <LangRef.html#i_icmp>`_ (integer opreands), and
2864 `FCmpInst <LangRef.html#i_fcmp>`_ (floating point operands).
2868 * ``TerminatorInst``
2870 This subclass is the parent of all terminator instructions (those which can
2875 Important Public Members of the ``Instruction`` class
2876 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2878 * ``BasicBlock *getParent()``
2880 Returns the BasicBlock_ that this
2881 ``Instruction`` is embedded into.
2883 * ``bool mayWriteToMemory()``
2885 Returns true if the instruction writes to memory, i.e. it is a ``call``,
2886 ``free``, ``invoke``, or ``store``.
2888 * ``unsigned getOpcode()``
2890 Returns the opcode for the ``Instruction``.
2892 * ``Instruction *clone() const``
2894 Returns another instance of the specified instruction, identical in all ways
2895 to the original except that the instruction has no parent (i.e. it's not
2896 embedded into a BasicBlock_), and it has no name.
2900 The ``Constant`` class and subclasses
2901 -------------------------------------
2903 Constant represents a base class for different types of constants. It is
2904 subclassed by ConstantInt, ConstantArray, etc. for representing the various
2905 types of Constants. GlobalValue_ is also a subclass, which represents the
2906 address of a global variable or function.
2910 Important Subclasses of Constant
2911 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2913 * ConstantInt : This subclass of Constant represents an integer constant of
2916 * ``const APInt& getValue() const``: Returns the underlying
2917 value of this constant, an APInt value.
2919 * ``int64_t getSExtValue() const``: Converts the underlying APInt value to an
2920 int64_t via sign extension. If the value (not the bit width) of the APInt
2921 is too large to fit in an int64_t, an assertion will result. For this
2922 reason, use of this method is discouraged.
2924 * ``uint64_t getZExtValue() const``: Converts the underlying APInt value
2925 to a uint64_t via zero extension. IF the value (not the bit width) of the
2926 APInt is too large to fit in a uint64_t, an assertion will result. For this
2927 reason, use of this method is discouraged.
2929 * ``static ConstantInt* get(const APInt& Val)``: Returns the ConstantInt
2930 object that represents the value provided by ``Val``. The type is implied
2931 as the IntegerType that corresponds to the bit width of ``Val``.
2933 * ``static ConstantInt* get(const Type *Ty, uint64_t Val)``: Returns the
2934 ConstantInt object that represents the value provided by ``Val`` for integer
2937 * ConstantFP : This class represents a floating point constant.
2939 * ``double getValue() const``: Returns the underlying value of this constant.
2941 * ConstantArray : This represents a constant array.
2943 * ``const std::vector<Use> &getValues() const``: Returns a vector of
2944 component constants that makeup this array.
2946 * ConstantStruct : This represents a constant struct.
2948 * ``const std::vector<Use> &getValues() const``: Returns a vector of
2949 component constants that makeup this array.
2951 * GlobalValue : This represents either a global variable or a function. In
2952 either case, the value is a constant fixed address (after linking).
2956 The ``GlobalValue`` class
2957 -------------------------
2959 ``#include "llvm/IR/GlobalValue.h"``
2961 header source: `GlobalValue.h
2962 <http://llvm.org/doxygen/GlobalValue_8h-source.html>`_
2964 doxygen info: `GlobalValue Class
2965 <http://llvm.org/doxygen/classllvm_1_1GlobalValue.html>`_
2967 Superclasses: Constant_, User_, Value_
2969 Global values ( GlobalVariable_\ s or :ref:`Function <c_Function>`\ s) are the
2970 only LLVM values that are visible in the bodies of all :ref:`Function
2971 <c_Function>`\ s. Because they are visible at global scope, they are also
2972 subject to linking with other globals defined in different translation units.
2973 To control the linking process, ``GlobalValue``\ s know their linkage rules.
2974 Specifically, ``GlobalValue``\ s know whether they have internal or external
2975 linkage, as defined by the ``LinkageTypes`` enumeration.
2977 If a ``GlobalValue`` has internal linkage (equivalent to being ``static`` in C),
2978 it is not visible to code outside the current translation unit, and does not
2979 participate in linking. If it has external linkage, it is visible to external
2980 code, and does participate in linking. In addition to linkage information,
2981 ``GlobalValue``\ s keep track of which Module_ they are currently part of.
2983 Because ``GlobalValue``\ s are memory objects, they are always referred to by
2984 their **address**. As such, the Type_ of a global is always a pointer to its
2985 contents. It is important to remember this when using the ``GetElementPtrInst``
2986 instruction because this pointer must be dereferenced first. For example, if
2987 you have a ``GlobalVariable`` (a subclass of ``GlobalValue)`` that is an array
2988 of 24 ints, type ``[24 x i32]``, then the ``GlobalVariable`` is a pointer to
2989 that array. Although the address of the first element of this array and the
2990 value of the ``GlobalVariable`` are the same, they have different types. The
2991 ``GlobalVariable``'s type is ``[24 x i32]``. The first element's type is
2992 ``i32.`` Because of this, accessing a global value requires you to dereference
2993 the pointer with ``GetElementPtrInst`` first, then its elements can be accessed.
2994 This is explained in the `LLVM Language Reference Manual
2995 <LangRef.html#globalvars>`_.
2999 Important Public Members of the ``GlobalValue`` class
3000 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3002 * | ``bool hasInternalLinkage() const``
3003 | ``bool hasExternalLinkage() const``
3004 | ``void setInternalLinkage(bool HasInternalLinkage)``
3006 These methods manipulate the linkage characteristics of the ``GlobalValue``.
3008 * ``Module *getParent()``
3010 This returns the Module_ that the
3011 GlobalValue is currently embedded into.
3015 The ``Function`` class
3016 ----------------------
3018 ``#include "llvm/IR/Function.h"``
3020 header source: `Function.h <http://llvm.org/doxygen/Function_8h-source.html>`_
3022 doxygen info: `Function Class
3023 <http://llvm.org/doxygen/classllvm_1_1Function.html>`_
3025 Superclasses: GlobalValue_, Constant_, User_, Value_
3027 The ``Function`` class represents a single procedure in LLVM. It is actually
3028 one of the more complex classes in the LLVM hierarchy because it must keep track
3029 of a large amount of data. The ``Function`` class keeps track of a list of
3030 BasicBlock_\ s, a list of formal Argument_\ s, and a SymbolTable_.
3032 The list of BasicBlock_\ s is the most commonly used part of ``Function``
3033 objects. The list imposes an implicit ordering of the blocks in the function,
3034 which indicate how the code will be laid out by the backend. Additionally, the
3035 first BasicBlock_ is the implicit entry node for the ``Function``. It is not
3036 legal in LLVM to explicitly branch to this initial block. There are no implicit
3037 exit nodes, and in fact there may be multiple exit nodes from a single
3038 ``Function``. If the BasicBlock_ list is empty, this indicates that the
3039 ``Function`` is actually a function declaration: the actual body of the function
3040 hasn't been linked in yet.
3042 In addition to a list of BasicBlock_\ s, the ``Function`` class also keeps track
3043 of the list of formal Argument_\ s that the function receives. This container
3044 manages the lifetime of the Argument_ nodes, just like the BasicBlock_ list does
3045 for the BasicBlock_\ s.
3047 The SymbolTable_ is a very rarely used LLVM feature that is only used when you
3048 have to look up a value by name. Aside from that, the SymbolTable_ is used
3049 internally to make sure that there are not conflicts between the names of
3050 Instruction_\ s, BasicBlock_\ s, or Argument_\ s in the function body.
3052 Note that ``Function`` is a GlobalValue_ and therefore also a Constant_. The
3053 value of the function is its address (after linking) which is guaranteed to be
3058 Important Public Members of the ``Function``
3059 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3061 * ``Function(const FunctionType *Ty, LinkageTypes Linkage,
3062 const std::string &N = "", Module* Parent = 0)``
3064 Constructor used when you need to create new ``Function``\ s to add the
3065 program. The constructor must specify the type of the function to create and
3066 what type of linkage the function should have. The FunctionType_ argument
3067 specifies the formal arguments and return value for the function. The same
3068 FunctionType_ value can be used to create multiple functions. The ``Parent``
3069 argument specifies the Module in which the function is defined. If this
3070 argument is provided, the function will automatically be inserted into that
3071 module's list of functions.
3073 * ``bool isDeclaration()``
3075 Return whether or not the ``Function`` has a body defined. If the function is
3076 "external", it does not have a body, and thus must be resolved by linking with
3077 a function defined in a different translation unit.
3079 * | ``Function::iterator`` - Typedef for basic block list iterator
3080 | ``Function::const_iterator`` - Typedef for const_iterator.
3081 | ``begin()``, ``end()``, ``size()``, ``empty()``
3083 These are forwarding methods that make it easy to access the contents of a
3084 ``Function`` object's BasicBlock_ list.
3086 * ``Function::BasicBlockListType &getBasicBlockList()``
3088 Returns the list of BasicBlock_\ s. This is necessary to use when you need to
3089 update the list or perform a complex action that doesn't have a forwarding
3092 * | ``Function::arg_iterator`` - Typedef for the argument list iterator
3093 | ``Function::const_arg_iterator`` - Typedef for const_iterator.
3094 | ``arg_begin()``, ``arg_end()``, ``arg_size()``, ``arg_empty()``
3096 These are forwarding methods that make it easy to access the contents of a
3097 ``Function`` object's Argument_ list.
3099 * ``Function::ArgumentListType &getArgumentList()``
3101 Returns the list of Argument_. This is necessary to use when you need to
3102 update the list or perform a complex action that doesn't have a forwarding
3105 * ``BasicBlock &getEntryBlock()``
3107 Returns the entry ``BasicBlock`` for the function. Because the entry block
3108 for the function is always the first block, this returns the first block of
3111 * | ``Type *getReturnType()``
3112 | ``FunctionType *getFunctionType()``
3114 This traverses the Type_ of the ``Function`` and returns the return type of
3115 the function, or the FunctionType_ of the actual function.
3117 * ``SymbolTable *getSymbolTable()``
3119 Return a pointer to the SymbolTable_ for this ``Function``.
3123 The ``GlobalVariable`` class
3124 ----------------------------
3126 ``#include "llvm/IR/GlobalVariable.h"``
3128 header source: `GlobalVariable.h
3129 <http://llvm.org/doxygen/GlobalVariable_8h-source.html>`_
3131 doxygen info: `GlobalVariable Class
3132 <http://llvm.org/doxygen/classllvm_1_1GlobalVariable.html>`_
3134 Superclasses: GlobalValue_, Constant_, User_, Value_
3136 Global variables are represented with the (surprise surprise) ``GlobalVariable``
3137 class. Like functions, ``GlobalVariable``\ s are also subclasses of
3138 GlobalValue_, and as such are always referenced by their address (global values
3139 must live in memory, so their "name" refers to their constant address). See
3140 GlobalValue_ for more on this. Global variables may have an initial value
3141 (which must be a Constant_), and if they have an initializer, they may be marked
3142 as "constant" themselves (indicating that their contents never change at
3145 .. _m_GlobalVariable:
3147 Important Public Members of the ``GlobalVariable`` class
3148 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3150 * ``GlobalVariable(const Type *Ty, bool isConstant, LinkageTypes &Linkage,
3151 Constant *Initializer = 0, const std::string &Name = "", Module* Parent = 0)``
3153 Create a new global variable of the specified type. If ``isConstant`` is true
3154 then the global variable will be marked as unchanging for the program. The
3155 Linkage parameter specifies the type of linkage (internal, external, weak,
3156 linkonce, appending) for the variable. If the linkage is InternalLinkage,
3157 WeakAnyLinkage, WeakODRLinkage, LinkOnceAnyLinkage or LinkOnceODRLinkage, then
3158 the resultant global variable will have internal linkage. AppendingLinkage
3159 concatenates together all instances (in different translation units) of the
3160 variable into a single variable but is only applicable to arrays. See the
3161 `LLVM Language Reference <LangRef.html#modulestructure>`_ for further details
3162 on linkage types. Optionally an initializer, a name, and the module to put
3163 the variable into may be specified for the global variable as well.
3165 * ``bool isConstant() const``
3167 Returns true if this is a global variable that is known not to be modified at
3170 * ``bool hasInitializer()``
3172 Returns true if this ``GlobalVariable`` has an intializer.
3174 * ``Constant *getInitializer()``
3176 Returns the initial value for a ``GlobalVariable``. It is not legal to call
3177 this method if there is no initializer.
3181 The ``BasicBlock`` class
3182 ------------------------
3184 ``#include "llvm/IR/BasicBlock.h"``
3186 header source: `BasicBlock.h
3187 <http://llvm.org/doxygen/BasicBlock_8h-source.html>`_
3189 doxygen info: `BasicBlock Class
3190 <http://llvm.org/doxygen/classllvm_1_1BasicBlock.html>`_
3194 This class represents a single entry single exit section of the code, commonly
3195 known as a basic block by the compiler community. The ``BasicBlock`` class
3196 maintains a list of Instruction_\ s, which form the body of the block. Matching
3197 the language definition, the last element of this list of instructions is always
3198 a terminator instruction (a subclass of the TerminatorInst_ class).
3200 In addition to tracking the list of instructions that make up the block, the
3201 ``BasicBlock`` class also keeps track of the :ref:`Function <c_Function>` that
3202 it is embedded into.
3204 Note that ``BasicBlock``\ s themselves are Value_\ s, because they are
3205 referenced by instructions like branches and can go in the switch tables.
3206 ``BasicBlock``\ s have type ``label``.
3210 Important Public Members of the ``BasicBlock`` class
3211 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3213 * ``BasicBlock(const std::string &Name = "", Function *Parent = 0)``
3215 The ``BasicBlock`` constructor is used to create new basic blocks for
3216 insertion into a function. The constructor optionally takes a name for the
3217 new block, and a :ref:`Function <c_Function>` to insert it into. If the
3218 ``Parent`` parameter is specified, the new ``BasicBlock`` is automatically
3219 inserted at the end of the specified :ref:`Function <c_Function>`, if not
3220 specified, the BasicBlock must be manually inserted into the :ref:`Function
3223 * | ``BasicBlock::iterator`` - Typedef for instruction list iterator
3224 | ``BasicBlock::const_iterator`` - Typedef for const_iterator.
3225 | ``begin()``, ``end()``, ``front()``, ``back()``,
3226 ``size()``, ``empty()``
3227 STL-style functions for accessing the instruction list.
3229 These methods and typedefs are forwarding functions that have the same
3230 semantics as the standard library methods of the same names. These methods
3231 expose the underlying instruction list of a basic block in a way that is easy
3232 to manipulate. To get the full complement of container operations (including
3233 operations to update the list), you must use the ``getInstList()`` method.
3235 * ``BasicBlock::InstListType &getInstList()``
3237 This method is used to get access to the underlying container that actually
3238 holds the Instructions. This method must be used when there isn't a
3239 forwarding function in the ``BasicBlock`` class for the operation that you
3240 would like to perform. Because there are no forwarding functions for
3241 "updating" operations, you need to use this if you want to update the contents
3242 of a ``BasicBlock``.
3244 * ``Function *getParent()``
3246 Returns a pointer to :ref:`Function <c_Function>` the block is embedded into,
3247 or a null pointer if it is homeless.
3249 * ``TerminatorInst *getTerminator()``
3251 Returns a pointer to the terminator instruction that appears at the end of the
3252 ``BasicBlock``. If there is no terminator instruction, or if the last
3253 instruction in the block is not a terminator, then a null pointer is returned.
3257 The ``Argument`` class
3258 ----------------------
3260 This subclass of Value defines the interface for incoming formal arguments to a
3261 function. A Function maintains a list of its formal arguments. An argument has
3262 a pointer to the parent Function.