1 ==============================
2 LLVM Language Reference Manual
3 ==============================
12 This document is a reference manual for the LLVM assembly language. LLVM
13 is a Static Single Assignment (SSA) based representation that provides
14 type safety, low-level operations, flexibility, and the capability of
15 representing 'all' high-level languages cleanly. It is the common code
16 representation used throughout all phases of the LLVM compilation
22 The LLVM code representation is designed to be used in three different
23 forms: as an in-memory compiler IR, as an on-disk bitcode representation
24 (suitable for fast loading by a Just-In-Time compiler), and as a human
25 readable assembly language representation. This allows LLVM to provide a
26 powerful intermediate representation for efficient compiler
27 transformations and analysis, while providing a natural means to debug
28 and visualize the transformations. The three different forms of LLVM are
29 all equivalent. This document describes the human readable
30 representation and notation.
32 The LLVM representation aims to be light-weight and low-level while
33 being expressive, typed, and extensible at the same time. It aims to be
34 a "universal IR" of sorts, by being at a low enough level that
35 high-level ideas may be cleanly mapped to it (similar to how
36 microprocessors are "universal IR's", allowing many source languages to
37 be mapped to them). By providing type information, LLVM can be used as
38 the target of optimizations: for example, through pointer analysis, it
39 can be proven that a C automatic variable is never accessed outside of
40 the current function, allowing it to be promoted to a simple SSA value
41 instead of a memory location.
48 It is important to note that this document describes 'well formed' LLVM
49 assembly language. There is a difference between what the parser accepts
50 and what is considered 'well formed'. For example, the following
51 instruction is syntactically okay, but not well formed:
57 because the definition of ``%x`` does not dominate all of its uses. The
58 LLVM infrastructure provides a verification pass that may be used to
59 verify that an LLVM module is well formed. This pass is automatically
60 run by the parser after parsing input assembly and by the optimizer
61 before it outputs bitcode. The violations pointed out by the verifier
62 pass indicate bugs in transformation passes or input to the parser.
69 LLVM identifiers come in two basic types: global and local. Global
70 identifiers (functions, global variables) begin with the ``'@'``
71 character. Local identifiers (register names, types) begin with the
72 ``'%'`` character. Additionally, there are three different formats for
73 identifiers, for different purposes:
75 #. Named values are represented as a string of characters with their
76 prefix. For example, ``%foo``, ``@DivisionByZero``,
77 ``%a.really.long.identifier``. The actual regular expression used is
78 '``[%@][a-zA-Z$._][a-zA-Z$._0-9]*``'. Identifiers which require other
79 characters in their names can be surrounded with quotes. Special
80 characters may be escaped using ``"\xx"`` where ``xx`` is the ASCII
81 code for the character in hexadecimal. In this way, any character can
82 be used in a name value, even quotes themselves.
83 #. Unnamed values are represented as an unsigned numeric value with
84 their prefix. For example, ``%12``, ``@2``, ``%44``.
85 #. Constants, which are described in the section Constants_ below.
87 LLVM requires that values start with a prefix for two reasons: Compilers
88 don't need to worry about name clashes with reserved words, and the set
89 of reserved words may be expanded in the future without penalty.
90 Additionally, unnamed identifiers allow a compiler to quickly come up
91 with a temporary variable without having to avoid symbol table
94 Reserved words in LLVM are very similar to reserved words in other
95 languages. There are keywords for different opcodes ('``add``',
96 '``bitcast``', '``ret``', etc...), for primitive type names ('``void``',
97 '``i32``', etc...), and others. These reserved words cannot conflict
98 with variable names, because none of them start with a prefix character
101 Here is an example of LLVM code to multiply the integer variable
108 %result = mul i32 %X, 8
110 After strength reduction:
114 %result = shl i32 %X, 3
120 %0 = add i32 %X, %X ; yields {i32}:%0
121 %1 = add i32 %0, %0 ; yields {i32}:%1
122 %result = add i32 %1, %1
124 This last way of multiplying ``%X`` by 8 illustrates several important
125 lexical features of LLVM:
127 #. Comments are delimited with a '``;``' and go until the end of line.
128 #. Unnamed temporaries are created when the result of a computation is
129 not assigned to a named value.
130 #. Unnamed temporaries are numbered sequentially (using a per-function
131 incrementing counter, starting with 0). Note that basic blocks are
132 included in this numbering. For example, if the entry basic block is not
133 given a label name, then it will get number 0.
135 It also shows a convention that we follow in this document. When
136 demonstrating instructions, we will follow an instruction with a comment
137 that defines the type and name of value produced.
145 LLVM programs are composed of ``Module``'s, each of which is a
146 translation unit of the input programs. Each module consists of
147 functions, global variables, and symbol table entries. Modules may be
148 combined together with the LLVM linker, which merges function (and
149 global variable) definitions, resolves forward declarations, and merges
150 symbol table entries. Here is an example of the "hello world" module:
154 ; Declare the string constant as a global constant.
155 @.str = private unnamed_addr constant [13 x i8] c"hello world\0A\00"
157 ; External declaration of the puts function
158 declare i32 @puts(i8* nocapture) nounwind
160 ; Definition of main function
161 define i32 @main() { ; i32()*
162 ; Convert [13 x i8]* to i8 *...
163 %cast210 = getelementptr [13 x i8]* @.str, i64 0, i64 0
165 ; Call puts function to write out the string to stdout.
166 call i32 @puts(i8* %cast210)
171 !1 = metadata !{i32 42}
174 This example is made up of a :ref:`global variable <globalvars>` named
175 "``.str``", an external declaration of the "``puts``" function, a
176 :ref:`function definition <functionstructure>` for "``main``" and
177 :ref:`named metadata <namedmetadatastructure>` "``foo``".
179 In general, a module is made up of a list of global values (where both
180 functions and global variables are global values). Global values are
181 represented by a pointer to a memory location (in this case, a pointer
182 to an array of char, and a pointer to a function), and have one of the
183 following :ref:`linkage types <linkage>`.
190 All Global Variables and Functions have one of the following types of
194 Global values with "``private``" linkage are only directly
195 accessible by objects in the current module. In particular, linking
196 code into a module with an private global value may cause the
197 private to be renamed as necessary to avoid collisions. Because the
198 symbol is private to the module, all references can be updated. This
199 doesn't show up in any symbol table in the object file.
201 Similar to ``private``, but the symbol is passed through the
202 assembler and evaluated by the linker. Unlike normal strong symbols,
203 they are removed by the linker from the final linked image
204 (executable or dynamic library).
205 ``linker_private_weak``
206 Similar to "``linker_private``", but the symbol is weak. Note that
207 ``linker_private_weak`` symbols are subject to coalescing by the
208 linker. The symbols are removed by the linker from the final linked
209 image (executable or dynamic library).
211 Similar to private, but the value shows as a local symbol
212 (``STB_LOCAL`` in the case of ELF) in the object file. This
213 corresponds to the notion of the '``static``' keyword in C.
214 ``available_externally``
215 Globals with "``available_externally``" linkage are never emitted
216 into the object file corresponding to the LLVM module. They exist to
217 allow inlining and other optimizations to take place given knowledge
218 of the definition of the global, which is known to be somewhere
219 outside the module. Globals with ``available_externally`` linkage
220 are allowed to be discarded at will, and are otherwise the same as
221 ``linkonce_odr``. This linkage type is only allowed on definitions,
224 Globals with "``linkonce``" linkage are merged with other globals of
225 the same name when linkage occurs. This can be used to implement
226 some forms of inline functions, templates, or other code which must
227 be generated in each translation unit that uses it, but where the
228 body may be overridden with a more definitive definition later.
229 Unreferenced ``linkonce`` globals are allowed to be discarded. Note
230 that ``linkonce`` linkage does not actually allow the optimizer to
231 inline the body of this function into callers because it doesn't
232 know if this definition of the function is the definitive definition
233 within the program or whether it will be overridden by a stronger
234 definition. To enable inlining and other optimizations, use
235 "``linkonce_odr``" linkage.
237 "``weak``" linkage has the same merging semantics as ``linkonce``
238 linkage, except that unreferenced globals with ``weak`` linkage may
239 not be discarded. This is used for globals that are declared "weak"
242 "``common``" linkage is most similar to "``weak``" linkage, but they
243 are used for tentative definitions in C, such as "``int X;``" at
244 global scope. Symbols with "``common``" linkage are merged in the
245 same way as ``weak symbols``, and they may not be deleted if
246 unreferenced. ``common`` symbols may not have an explicit section,
247 must have a zero initializer, and may not be marked
248 ':ref:`constant <globalvars>`'. Functions and aliases may not have
251 .. _linkage_appending:
254 "``appending``" linkage may only be applied to global variables of
255 pointer to array type. When two global variables with appending
256 linkage are linked together, the two global arrays are appended
257 together. This is the LLVM, typesafe, equivalent of having the
258 system linker append together "sections" with identical names when
261 The semantics of this linkage follow the ELF object file model: the
262 symbol is weak until linked, if not linked, the symbol becomes null
263 instead of being an undefined reference.
264 ``linkonce_odr``, ``weak_odr``
265 Some languages allow differing globals to be merged, such as two
266 functions with different semantics. Other languages, such as
267 ``C++``, ensure that only equivalent globals are ever merged (the
268 "one definition rule" --- "ODR"). Such languages can use the
269 ``linkonce_odr`` and ``weak_odr`` linkage types to indicate that the
270 global will only be merged with equivalent globals. These linkage
271 types are otherwise the same as their non-``odr`` versions.
273 If none of the above identifiers are used, the global is externally
274 visible, meaning that it participates in linkage and can be used to
275 resolve external symbol references.
277 The next two types of linkage are targeted for Microsoft Windows
278 platform only. They are designed to support importing (exporting)
279 symbols from (to) DLLs (Dynamic Link Libraries).
282 "``dllimport``" linkage causes the compiler to reference a function
283 or variable via a global pointer to a pointer that is set up by the
284 DLL exporting the symbol. On Microsoft Windows targets, the pointer
285 name is formed by combining ``__imp_`` and the function or variable
288 "``dllexport``" linkage causes the compiler to provide a global
289 pointer to a pointer in a DLL, so that it can be referenced with the
290 ``dllimport`` attribute. On Microsoft Windows targets, the pointer
291 name is formed by combining ``__imp_`` and the function or variable
292 name. Since this linkage exists for defining a dll interface, the
293 compiler, assembler and linker know it is externally referenced and
294 must refrain from deleting the symbol.
296 It is illegal for a function *declaration* to have any linkage type
297 other than ``external``, ``dllimport`` or ``extern_weak``.
304 LLVM :ref:`functions <functionstructure>`, :ref:`calls <i_call>` and
305 :ref:`invokes <i_invoke>` can all have an optional calling convention
306 specified for the call. The calling convention of any pair of dynamic
307 caller/callee must match, or the behavior of the program is undefined.
308 The following calling conventions are supported by LLVM, and more may be
311 "``ccc``" - The C calling convention
312 This calling convention (the default if no other calling convention
313 is specified) matches the target C calling conventions. This calling
314 convention supports varargs function calls and tolerates some
315 mismatch in the declared prototype and implemented declaration of
316 the function (as does normal C).
317 "``fastcc``" - The fast calling convention
318 This calling convention attempts to make calls as fast as possible
319 (e.g. by passing things in registers). This calling convention
320 allows the target to use whatever tricks it wants to produce fast
321 code for the target, without having to conform to an externally
322 specified ABI (Application Binary Interface). `Tail calls can only
323 be optimized when this, the GHC or the HiPE convention is
324 used. <CodeGenerator.html#id80>`_ This calling convention does not
325 support varargs and requires the prototype of all callees to exactly
326 match the prototype of the function definition.
327 "``coldcc``" - The cold calling convention
328 This calling convention attempts to make code in the caller as
329 efficient as possible under the assumption that the call is not
330 commonly executed. As such, these calls often preserve all registers
331 so that the call does not break any live ranges in the caller side.
332 This calling convention does not support varargs and requires the
333 prototype of all callees to exactly match the prototype of the
335 "``cc 10``" - GHC convention
336 This calling convention has been implemented specifically for use by
337 the `Glasgow Haskell Compiler (GHC) <http://www.haskell.org/ghc>`_.
338 It passes everything in registers, going to extremes to achieve this
339 by disabling callee save registers. This calling convention should
340 not be used lightly but only for specific situations such as an
341 alternative to the *register pinning* performance technique often
342 used when implementing functional programming languages. At the
343 moment only X86 supports this convention and it has the following
346 - On *X86-32* only supports up to 4 bit type parameters. No
347 floating point types are supported.
348 - On *X86-64* only supports up to 10 bit type parameters and 6
349 floating point parameters.
351 This calling convention supports `tail call
352 optimization <CodeGenerator.html#id80>`_ but requires both the
353 caller and callee are using it.
354 "``cc 11``" - The HiPE calling convention
355 This calling convention has been implemented specifically for use by
356 the `High-Performance Erlang
357 (HiPE) <http://www.it.uu.se/research/group/hipe/>`_ compiler, *the*
358 native code compiler of the `Ericsson's Open Source Erlang/OTP
359 system <http://www.erlang.org/download.shtml>`_. It uses more
360 registers for argument passing than the ordinary C calling
361 convention and defines no callee-saved registers. The calling
362 convention properly supports `tail call
363 optimization <CodeGenerator.html#id80>`_ but requires that both the
364 caller and the callee use it. It uses a *register pinning*
365 mechanism, similar to GHC's convention, for keeping frequently
366 accessed runtime components pinned to specific hardware registers.
367 At the moment only X86 supports this convention (both 32 and 64
369 "``webkit_jscc``" - WebKit's JavaScript calling convention
370 This calling convention has been implemented for `WebKit FTL JIT
371 <https://trac.webkit.org/wiki/FTLJIT>`_. It passes arguments on the
372 stack right to left (as cdecl does), and returns a value in the
373 platform's customary return register.
374 "``anyregcc``" - Dynamic calling convention for code patching
375 This is a special convention that supports patching an arbitrary code
376 sequence in place of a call site. This convention forces the call
377 arguments into registers but allows them to be dynamcially
378 allocated. This can currently only be used with calls to
379 llvm.experimental.patchpoint because only this intrinsic records
380 the location of its arguments in a side table. See :doc:`StackMaps`.
381 "``cc <n>``" - Numbered convention
382 Any calling convention may be specified by number, allowing
383 target-specific calling conventions to be used. Target specific
384 calling conventions start at 64.
386 More calling conventions can be added/defined on an as-needed basis, to
387 support Pascal conventions or any other well-known target-independent
390 .. _visibilitystyles:
395 All Global Variables and Functions have one of the following visibility
398 "``default``" - Default style
399 On targets that use the ELF object file format, default visibility
400 means that the declaration is visible to other modules and, in
401 shared libraries, means that the declared entity may be overridden.
402 On Darwin, default visibility means that the declaration is visible
403 to other modules. Default visibility corresponds to "external
404 linkage" in the language.
405 "``hidden``" - Hidden style
406 Two declarations of an object with hidden visibility refer to the
407 same object if they are in the same shared object. Usually, hidden
408 visibility indicates that the symbol will not be placed into the
409 dynamic symbol table, so no other module (executable or shared
410 library) can reference it directly.
411 "``protected``" - Protected style
412 On ELF, protected visibility indicates that the symbol will be
413 placed in the dynamic symbol table, but that references within the
414 defining module will bind to the local symbol. That is, the symbol
415 cannot be overridden by another module.
422 LLVM IR allows you to specify name aliases for certain types. This can
423 make it easier to read the IR and make the IR more condensed
424 (particularly when recursive types are involved). An example of a name
429 %mytype = type { %mytype*, i32 }
431 You may give a name to any :ref:`type <typesystem>` except
432 ":ref:`void <t_void>`". Type name aliases may be used anywhere a type is
433 expected with the syntax "%mytype".
435 Note that type names are aliases for the structural type that they
436 indicate, and that you can therefore specify multiple names for the same
437 type. This often leads to confusing behavior when dumping out a .ll
438 file. Since LLVM IR uses structural typing, the name is not part of the
439 type. When printing out LLVM IR, the printer will pick *one name* to
440 render all types of a particular shape. This means that if you have code
441 where two different source types end up having the same LLVM type, that
442 the dumper will sometimes print the "wrong" or unexpected type. This is
443 an important design point and isn't going to change.
450 Global variables define regions of memory allocated at compilation time
453 Global variables definitions must be initialized, may have an explicit section
454 to be placed in, and may have an optional explicit alignment specified.
456 Global variables in other translation units can also be declared, in which
457 case they don't have an initializer.
459 A variable may be defined as ``thread_local``, which means that it will
460 not be shared by threads (each thread will have a separated copy of the
461 variable). Not all targets support thread-local variables. Optionally, a
462 TLS model may be specified:
465 For variables that are only used within the current shared library.
467 For variables in modules that will not be loaded dynamically.
469 For variables defined in the executable and only used within it.
471 The models correspond to the ELF TLS models; see `ELF Handling For
472 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
473 more information on under which circumstances the different models may
474 be used. The target may choose a different TLS model if the specified
475 model is not supported, or if a better choice of model can be made.
477 A variable may be defined as a global ``constant``, which indicates that
478 the contents of the variable will **never** be modified (enabling better
479 optimization, allowing the global data to be placed in the read-only
480 section of an executable, etc). Note that variables that need runtime
481 initialization cannot be marked ``constant`` as there is a store to the
484 LLVM explicitly allows *declarations* of global variables to be marked
485 constant, even if the final definition of the global is not. This
486 capability can be used to enable slightly better optimization of the
487 program, but requires the language definition to guarantee that
488 optimizations based on the 'constantness' are valid for the translation
489 units that do not include the definition.
491 As SSA values, global variables define pointer values that are in scope
492 (i.e. they dominate) all basic blocks in the program. Global variables
493 always define a pointer to their "content" type because they describe a
494 region of memory, and all memory objects in LLVM are accessed through
497 Global variables can be marked with ``unnamed_addr`` which indicates
498 that the address is not significant, only the content. Constants marked
499 like this can be merged with other constants if they have the same
500 initializer. Note that a constant with significant address *can* be
501 merged with a ``unnamed_addr`` constant, the result being a constant
502 whose address is significant.
504 A global variable may be declared to reside in a target-specific
505 numbered address space. For targets that support them, address spaces
506 may affect how optimizations are performed and/or what target
507 instructions are used to access the variable. The default address space
508 is zero. The address space qualifier must precede any other attributes.
510 LLVM allows an explicit section to be specified for globals. If the
511 target supports it, it will emit globals to the section specified.
513 By default, global initializers are optimized by assuming that global
514 variables defined within the module are not modified from their
515 initial values before the start of the global initializer. This is
516 true even for variables potentially accessible from outside the
517 module, including those with external linkage or appearing in
518 ``@llvm.used`` or dllexported variables. This assumption may be suppressed
519 by marking the variable with ``externally_initialized``.
521 An explicit alignment may be specified for a global, which must be a
522 power of 2. If not present, or if the alignment is set to zero, the
523 alignment of the global is set by the target to whatever it feels
524 convenient. If an explicit alignment is specified, the global is forced
525 to have exactly that alignment. Targets and optimizers are not allowed
526 to over-align the global if the global has an assigned section. In this
527 case, the extra alignment could be observable: for example, code could
528 assume that the globals are densely packed in their section and try to
529 iterate over them as an array, alignment padding would break this
532 For example, the following defines a global in a numbered address space
533 with an initializer, section, and alignment:
537 @G = addrspace(5) constant float 1.0, section "foo", align 4
539 The following example just declares a global variable
543 @G = external global i32
545 The following example defines a thread-local global with the
546 ``initialexec`` TLS model:
550 @G = thread_local(initialexec) global i32 0, align 4
552 .. _functionstructure:
557 LLVM function definitions consist of the "``define``" keyword, an
558 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
559 style <visibility>`, an optional :ref:`calling convention <callingconv>`,
560 an optional ``unnamed_addr`` attribute, a return type, an optional
561 :ref:`parameter attribute <paramattrs>` for the return type, a function
562 name, a (possibly empty) argument list (each with optional :ref:`parameter
563 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
564 an optional section, an optional alignment, an optional :ref:`garbage
565 collector name <gc>`, an optional :ref:`prefix <prefixdata>`, an opening
566 curly brace, a list of basic blocks, and a closing curly brace.
568 LLVM function declarations consist of the "``declare``" keyword, an
569 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
570 style <visibility>`, an optional :ref:`calling convention <callingconv>`,
571 an optional ``unnamed_addr`` attribute, a return type, an optional
572 :ref:`parameter attribute <paramattrs>` for the return type, a function
573 name, a possibly empty list of arguments, an optional alignment, an optional
574 :ref:`garbage collector name <gc>` and an optional :ref:`prefix <prefixdata>`.
576 A function definition contains a list of basic blocks, forming the CFG (Control
577 Flow Graph) for the function. Each basic block may optionally start with a label
578 (giving the basic block a symbol table entry), contains a list of instructions,
579 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
580 function return). If an explicit label is not provided, a block is assigned an
581 implicit numbered label, using the next value from the same counter as used for
582 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
583 entry block does not have an explicit label, it will be assigned label "%0",
584 then the first unnamed temporary in that block will be "%1", etc.
586 The first basic block in a function is special in two ways: it is
587 immediately executed on entrance to the function, and it is not allowed
588 to have predecessor basic blocks (i.e. there can not be any branches to
589 the entry block of a function). Because the block can have no
590 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
592 LLVM allows an explicit section to be specified for functions. If the
593 target supports it, it will emit functions to the section specified.
595 An explicit alignment may be specified for a function. If not present,
596 or if the alignment is set to zero, the alignment of the function is set
597 by the target to whatever it feels convenient. If an explicit alignment
598 is specified, the function is forced to have at least that much
599 alignment. All alignments must be a power of 2.
601 If the ``unnamed_addr`` attribute is given, the address is know to not
602 be significant and two identical functions can be merged.
606 define [linkage] [visibility]
608 <ResultType> @<FunctionName> ([argument list])
609 [fn Attrs] [section "name"] [align N]
610 [gc] [prefix Constant] { ... }
617 Aliases act as "second name" for the aliasee value (which can be either
618 function, global variable, another alias or bitcast of global value).
619 Aliases may have an optional :ref:`linkage type <linkage>`, and an optional
620 :ref:`visibility style <visibility>`.
624 @<Name> = alias [Linkage] [Visibility] <AliaseeTy> @<Aliasee>
626 The linkage must be one of ``private``, ``linker_private``,
627 ``linker_private_weak``, ``internal``, ``linkonce``, ``weak``,
628 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
629 might not correctly handle dropping a weak symbol that is aliased by a non-weak
632 .. _namedmetadatastructure:
637 Named metadata is a collection of metadata. :ref:`Metadata
638 nodes <metadata>` (but not metadata strings) are the only valid
639 operands for a named metadata.
643 ; Some unnamed metadata nodes, which are referenced by the named metadata.
644 !0 = metadata !{metadata !"zero"}
645 !1 = metadata !{metadata !"one"}
646 !2 = metadata !{metadata !"two"}
648 !name = !{!0, !1, !2}
655 The return type and each parameter of a function type may have a set of
656 *parameter attributes* associated with them. Parameter attributes are
657 used to communicate additional information about the result or
658 parameters of a function. Parameter attributes are considered to be part
659 of the function, not of the function type, so functions with different
660 parameter attributes can have the same function type.
662 Parameter attributes are simple keywords that follow the type specified.
663 If multiple parameter attributes are needed, they are space separated.
668 declare i32 @printf(i8* noalias nocapture, ...)
669 declare i32 @atoi(i8 zeroext)
670 declare signext i8 @returns_signed_char()
672 Note that any attributes for the function result (``nounwind``,
673 ``readonly``) come immediately after the argument list.
675 Currently, only the following parameter attributes are defined:
678 This indicates to the code generator that the parameter or return
679 value should be zero-extended to the extent required by the target's
680 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
681 the caller (for a parameter) or the callee (for a return value).
683 This indicates to the code generator that the parameter or return
684 value should be sign-extended to the extent required by the target's
685 ABI (which is usually 32-bits) by the caller (for a parameter) or
686 the callee (for a return value).
688 This indicates that this parameter or return value should be treated
689 in a special target-dependent fashion during while emitting code for
690 a function call or return (usually, by putting it in a register as
691 opposed to memory, though some targets use it to distinguish between
692 two different kinds of registers). Use of this attribute is
695 This indicates that the pointer parameter should really be passed by
696 value to the function. The attribute implies that a hidden copy of
697 the pointee is made between the caller and the callee, so the callee
698 is unable to modify the value in the caller. This attribute is only
699 valid on LLVM pointer arguments. It is generally used to pass
700 structs and arrays by value, but is also valid on pointers to
701 scalars. The copy is considered to belong to the caller not the
702 callee (for example, ``readonly`` functions should not write to
703 ``byval`` parameters). This is not a valid attribute for return
706 The byval attribute also supports specifying an alignment with the
707 align attribute. It indicates the alignment of the stack slot to
708 form and the known alignment of the pointer specified to the call
709 site. If the alignment is not specified, then the code generator
710 makes a target-specific assumption.
716 .. Warning:: This feature is unstable and not fully implemented.
718 The ``inalloca`` argument attribute allows the caller to get the
719 address of an outgoing argument to a ``call`` or ``invoke`` before
720 it executes. It is similar to ``byval`` in that it is used to pass
721 arguments by value, but it guarantees that the argument will not be
724 To be :ref:`well formed <wellformed>`, the caller must pass in an
725 alloca value into an ``inalloca`` parameter, and an alloca may be
726 used as an ``inalloca`` argument at most once. The attribute can
727 only be applied to parameters that would be passed in memory and not
728 registers. The ``inalloca`` attribute cannot be used in conjunction
729 with other attributes that affect argument storage, like ``inreg``,
730 ``nest``, ``sret``, or ``byval``. The ``inalloca`` stack space is
731 considered to be clobbered by any call that uses it, so any
732 ``inalloca`` parameters cannot be marked ``readonly``.
734 Allocas passed with ``inalloca`` to a call must be in the opposite
735 order of the parameter list, meaning that the rightmost argument
736 must be allocated first. If a call has inalloca arguments, no other
737 allocas can occur between the first alloca used by the call and the
738 call site, unless they are are cleared by calls to
739 :ref:`llvm.stackrestore <int_stackrestore>`. Violating these rules
740 results in undefined behavior at runtime.
742 See :doc:`InAlloca` for more information on how to use this
746 This indicates that the pointer parameter specifies the address of a
747 structure that is the return value of the function in the source
748 program. This pointer must be guaranteed by the caller to be valid:
749 loads and stores to the structure may be assumed by the callee
750 not to trap and to be properly aligned. This may only be applied to
751 the first parameter. This is not a valid attribute for return
754 This indicates that pointer values :ref:`based <pointeraliasing>` on
755 the argument or return value do not alias pointer values which are
756 not *based* on it, ignoring certain "irrelevant" dependencies. For a
757 call to the parent function, dependencies between memory references
758 from before or after the call and from those during the call are
759 "irrelevant" to the ``noalias`` keyword for the arguments and return
760 value used in that call. The caller shares the responsibility with
761 the callee for ensuring that these requirements are met. For further
762 details, please see the discussion of the NoAlias response in `alias
763 analysis <AliasAnalysis.html#MustMayNo>`_.
765 Note that this definition of ``noalias`` is intentionally similar
766 to the definition of ``restrict`` in C99 for function arguments,
767 though it is slightly weaker.
769 For function return values, C99's ``restrict`` is not meaningful,
770 while LLVM's ``noalias`` is.
772 This indicates that the callee does not make any copies of the
773 pointer that outlive the callee itself. This is not a valid
774 attribute for return values.
779 This indicates that the pointer parameter can be excised using the
780 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
781 attribute for return values and can only be applied to one parameter.
784 This indicates that the function always returns the argument as its return
785 value. This is an optimization hint to the code generator when generating
786 the caller, allowing tail call optimization and omission of register saves
787 and restores in some cases; it is not checked or enforced when generating
788 the callee. The parameter and the function return type must be valid
789 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
790 valid attribute for return values and can only be applied to one parameter.
794 Garbage Collector Names
795 -----------------------
797 Each function may specify a garbage collector name, which is simply a
802 define void @f() gc "name" { ... }
804 The compiler declares the supported values of *name*. Specifying a
805 collector which will cause the compiler to alter its output in order to
806 support the named garbage collection algorithm.
813 Prefix data is data associated with a function which the code generator
814 will emit immediately before the function body. The purpose of this feature
815 is to allow frontends to associate language-specific runtime metadata with
816 specific functions and make it available through the function pointer while
817 still allowing the function pointer to be called. To access the data for a
818 given function, a program may bitcast the function pointer to a pointer to
819 the constant's type. This implies that the IR symbol points to the start
822 To maintain the semantics of ordinary function calls, the prefix data must
823 have a particular format. Specifically, it must begin with a sequence of
824 bytes which decode to a sequence of machine instructions, valid for the
825 module's target, which transfer control to the point immediately succeeding
826 the prefix data, without performing any other visible action. This allows
827 the inliner and other passes to reason about the semantics of the function
828 definition without needing to reason about the prefix data. Obviously this
829 makes the format of the prefix data highly target dependent.
831 Prefix data is laid out as if it were an initializer for a global variable
832 of the prefix data's type. No padding is automatically placed between the
833 prefix data and the function body. If padding is required, it must be part
836 A trivial example of valid prefix data for the x86 architecture is ``i8 144``,
837 which encodes the ``nop`` instruction:
841 define void @f() prefix i8 144 { ... }
843 Generally prefix data can be formed by encoding a relative branch instruction
844 which skips the metadata, as in this example of valid prefix data for the
845 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
849 %0 = type <{ i8, i8, i8* }>
851 define void @f() prefix %0 <{ i8 235, i8 8, i8* @md}> { ... }
853 A function may have prefix data but no body. This has similar semantics
854 to the ``available_externally`` linkage in that the data may be used by the
855 optimizers but will not be emitted in the object file.
862 Attribute groups are groups of attributes that are referenced by objects within
863 the IR. They are important for keeping ``.ll`` files readable, because a lot of
864 functions will use the same set of attributes. In the degenerative case of a
865 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
866 group will capture the important command line flags used to build that file.
868 An attribute group is a module-level object. To use an attribute group, an
869 object references the attribute group's ID (e.g. ``#37``). An object may refer
870 to more than one attribute group. In that situation, the attributes from the
871 different groups are merged.
873 Here is an example of attribute groups for a function that should always be
874 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
878 ; Target-independent attributes:
879 attributes #0 = { alwaysinline alignstack=4 }
881 ; Target-dependent attributes:
882 attributes #1 = { "no-sse" }
884 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
885 define void @f() #0 #1 { ... }
892 Function attributes are set to communicate additional information about
893 a function. Function attributes are considered to be part of the
894 function, not of the function type, so functions with different function
895 attributes can have the same function type.
897 Function attributes are simple keywords that follow the type specified.
898 If multiple attributes are needed, they are space separated. For
903 define void @f() noinline { ... }
904 define void @f() alwaysinline { ... }
905 define void @f() alwaysinline optsize { ... }
906 define void @f() optsize { ... }
909 This attribute indicates that, when emitting the prologue and
910 epilogue, the backend should forcibly align the stack pointer.
911 Specify the desired alignment, which must be a power of two, in
914 This attribute indicates that the inliner should attempt to inline
915 this function into callers whenever possible, ignoring any active
916 inlining size threshold for this caller.
918 This indicates that the callee function at a call site should be
919 recognized as a built-in function, even though the function's declaration
920 uses the ``nobuiltin`` attribute. This is only valid at call sites for
921 direct calls to functions which are declared with the ``nobuiltin``
924 This attribute indicates that this function is rarely called. When
925 computing edge weights, basic blocks post-dominated by a cold
926 function call are also considered to be cold; and, thus, given low
929 This attribute indicates that the source code contained a hint that
930 inlining this function is desirable (such as the "inline" keyword in
931 C/C++). It is just a hint; it imposes no requirements on the
934 This attribute suggests that optimization passes and code generator
935 passes make choices that keep the code size of this function as small
936 as possible and perform optimizations that may sacrifice runtime
937 performance in order to minimize the size of the generated code.
939 This attribute disables prologue / epilogue emission for the
940 function. This can have very system-specific consequences.
942 This indicates that the callee function at a call site is not recognized as
943 a built-in function. LLVM will retain the original call and not replace it
944 with equivalent code based on the semantics of the built-in function, unless
945 the call site uses the ``builtin`` attribute. This is valid at call sites
946 and on function declarations and definitions.
948 This attribute indicates that calls to the function cannot be
949 duplicated. A call to a ``noduplicate`` function may be moved
950 within its parent function, but may not be duplicated within
953 A function containing a ``noduplicate`` call may still
954 be an inlining candidate, provided that the call is not
955 duplicated by inlining. That implies that the function has
956 internal linkage and only has one call site, so the original
957 call is dead after inlining.
959 This attributes disables implicit floating point instructions.
961 This attribute indicates that the inliner should never inline this
962 function in any situation. This attribute may not be used together
963 with the ``alwaysinline`` attribute.
965 This attribute suppresses lazy symbol binding for the function. This
966 may make calls to the function faster, at the cost of extra program
967 startup time if the function is not called during program startup.
969 This attribute indicates that the code generator should not use a
970 red zone, even if the target-specific ABI normally permits it.
972 This function attribute indicates that the function never returns
973 normally. This produces undefined behavior at runtime if the
974 function ever does dynamically return.
976 This function attribute indicates that the function never returns
977 with an unwind or exceptional control flow. If the function does
978 unwind, its runtime behavior is undefined.
980 This function attribute indicates that the function is not optimized
981 by any optimization or code generator passes with the
982 exception of interprocedural optimization passes.
983 This attribute cannot be used together with the ``alwaysinline``
984 attribute; this attribute is also incompatible
985 with the ``minsize`` attribute and the ``optsize`` attribute.
987 This attribute requires the ``noinline`` attribute to be specified on
988 the function as well, so the function is never inlined into any caller.
989 Only functions with the ``alwaysinline`` attribute are valid
990 candidates for inlining into the body of this function.
992 This attribute suggests that optimization passes and code generator
993 passes make choices that keep the code size of this function low,
994 and otherwise do optimizations specifically to reduce code size as
995 long as they do not significantly impact runtime performance.
997 On a function, this attribute indicates that the function computes its
998 result (or decides to unwind an exception) based strictly on its arguments,
999 without dereferencing any pointer arguments or otherwise accessing
1000 any mutable state (e.g. memory, control registers, etc) visible to
1001 caller functions. It does not write through any pointer arguments
1002 (including ``byval`` arguments) and never changes any state visible
1003 to callers. This means that it cannot unwind exceptions by calling
1004 the ``C++`` exception throwing methods.
1006 On an argument, this attribute indicates that the function does not
1007 dereference that pointer argument, even though it may read or write the
1008 memory that the pointer points to if accessed through other pointers.
1010 On a function, this attribute indicates that the function does not write
1011 through any pointer arguments (including ``byval`` arguments) or otherwise
1012 modify any state (e.g. memory, control registers, etc) visible to
1013 caller functions. It may dereference pointer arguments and read
1014 state that may be set in the caller. A readonly function always
1015 returns the same value (or unwinds an exception identically) when
1016 called with the same set of arguments and global state. It cannot
1017 unwind an exception by calling the ``C++`` exception throwing
1020 On an argument, this attribute indicates that the function does not write
1021 through this pointer argument, even though it may write to the memory that
1022 the pointer points to.
1024 This attribute indicates that this function can return twice. The C
1025 ``setjmp`` is an example of such a function. The compiler disables
1026 some optimizations (like tail calls) in the caller of these
1028 ``sanitize_address``
1029 This attribute indicates that AddressSanitizer checks
1030 (dynamic address safety analysis) are enabled for this function.
1032 This attribute indicates that MemorySanitizer checks (dynamic detection
1033 of accesses to uninitialized memory) are enabled for this function.
1035 This attribute indicates that ThreadSanitizer checks
1036 (dynamic thread safety analysis) are enabled for this function.
1038 This attribute indicates that the function should emit a stack
1039 smashing protector. It is in the form of a "canary" --- a random value
1040 placed on the stack before the local variables that's checked upon
1041 return from the function to see if it has been overwritten. A
1042 heuristic is used to determine if a function needs stack protectors
1043 or not. The heuristic used will enable protectors for functions with:
1045 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1046 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1047 - Calls to alloca() with variable sizes or constant sizes greater than
1048 ``ssp-buffer-size``.
1050 If a function that has an ``ssp`` attribute is inlined into a
1051 function that doesn't have an ``ssp`` attribute, then the resulting
1052 function will have an ``ssp`` attribute.
1054 This attribute indicates that the function should *always* emit a
1055 stack smashing protector. This overrides the ``ssp`` function
1058 If a function that has an ``sspreq`` attribute is inlined into a
1059 function that doesn't have an ``sspreq`` attribute or which has an
1060 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1061 an ``sspreq`` attribute.
1063 This attribute indicates that the function should emit a stack smashing
1064 protector. This attribute causes a strong heuristic to be used when
1065 determining if a function needs stack protectors. The strong heuristic
1066 will enable protectors for functions with:
1068 - Arrays of any size and type
1069 - Aggregates containing an array of any size and type.
1070 - Calls to alloca().
1071 - Local variables that have had their address taken.
1073 This overrides the ``ssp`` function attribute.
1075 If a function that has an ``sspstrong`` attribute is inlined into a
1076 function that doesn't have an ``sspstrong`` attribute, then the
1077 resulting function will have an ``sspstrong`` attribute.
1079 This attribute indicates that the ABI being targeted requires that
1080 an unwind table entry be produce for this function even if we can
1081 show that no exceptions passes by it. This is normally the case for
1082 the ELF x86-64 abi, but it can be disabled for some compilation
1087 Module-Level Inline Assembly
1088 ----------------------------
1090 Modules may contain "module-level inline asm" blocks, which corresponds
1091 to the GCC "file scope inline asm" blocks. These blocks are internally
1092 concatenated by LLVM and treated as a single unit, but may be separated
1093 in the ``.ll`` file if desired. The syntax is very simple:
1095 .. code-block:: llvm
1097 module asm "inline asm code goes here"
1098 module asm "more can go here"
1100 The strings can contain any character by escaping non-printable
1101 characters. The escape sequence used is simply "\\xx" where "xx" is the
1102 two digit hex code for the number.
1104 The inline asm code is simply printed to the machine code .s file when
1105 assembly code is generated.
1107 .. _langref_datalayout:
1112 A module may specify a target specific data layout string that specifies
1113 how data is to be laid out in memory. The syntax for the data layout is
1116 .. code-block:: llvm
1118 target datalayout = "layout specification"
1120 The *layout specification* consists of a list of specifications
1121 separated by the minus sign character ('-'). Each specification starts
1122 with a letter and may include other information after the letter to
1123 define some aspect of the data layout. The specifications accepted are
1127 Specifies that the target lays out data in big-endian form. That is,
1128 the bits with the most significance have the lowest address
1131 Specifies that the target lays out data in little-endian form. That
1132 is, the bits with the least significance have the lowest address
1135 Specifies the natural alignment of the stack in bits. Alignment
1136 promotion of stack variables is limited to the natural stack
1137 alignment to avoid dynamic stack realignment. The stack alignment
1138 must be a multiple of 8-bits. If omitted, the natural stack
1139 alignment defaults to "unspecified", which does not prevent any
1140 alignment promotions.
1141 ``p[n]:<size>:<abi>:<pref>``
1142 This specifies the *size* of a pointer and its ``<abi>`` and
1143 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1144 bits. Specifying the ``<pref>`` alignment is optional. If omitted, the
1145 preceding ``:`` should be omitted too. The address space, ``n`` is
1146 optional, and if not specified, denotes the default address space 0.
1147 The value of ``n`` must be in the range [1,2^23).
1148 ``i<size>:<abi>:<pref>``
1149 This specifies the alignment for an integer type of a given bit
1150 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1151 ``v<size>:<abi>:<pref>``
1152 This specifies the alignment for a vector type of a given bit
1154 ``f<size>:<abi>:<pref>``
1155 This specifies the alignment for a floating point type of a given bit
1156 ``<size>``. Only values of ``<size>`` that are supported by the target
1157 will work. 32 (float) and 64 (double) are supported on all targets; 80
1158 or 128 (different flavors of long double) are also supported on some
1160 ``a<size>:<abi>:<pref>``
1161 This specifies the alignment for an aggregate type of a given bit
1164 If prerest, specifies that llvm names are mangled in the output. The
1166 * ``e``: ELF mangling: Private symbols get a ``.L`` prefix.
1167 * ``m``: Mips mangling: Private symbols get a ``$`` prefix.
1168 * ``o``: Mach-O mangling: Private symbols get ``L`` prefix. Other
1169 symbols get a ``_`` prefix.
1170 * ``c``: COFF prefix: Similar to Mach-O, but stdcall and fastcall
1171 functions also get a suffix based on the frame size.
1172 ``n<size1>:<size2>:<size3>...``
1173 This specifies a set of native integer widths for the target CPU in
1174 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1175 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1176 this set are considered to support most general arithmetic operations
1179 When constructing the data layout for a given target, LLVM starts with a
1180 default set of specifications which are then (possibly) overridden by
1181 the specifications in the ``datalayout`` keyword. The default
1182 specifications are given in this list:
1184 - ``E`` - big endian
1185 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1186 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1187 same as the default address space.
1188 - ``S0`` - natural stack alignment is unspecified
1189 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1190 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1191 - ``i16:16:16`` - i16 is 16-bit aligned
1192 - ``i32:32:32`` - i32 is 32-bit aligned
1193 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1194 alignment of 64-bits
1195 - ``f16:16:16`` - half is 16-bit aligned
1196 - ``f32:32:32`` - float is 32-bit aligned
1197 - ``f64:64:64`` - double is 64-bit aligned
1198 - ``f128:128:128`` - quad is 128-bit aligned
1199 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1200 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1201 - ``a:0:64`` - aggregates are 64-bit aligned
1203 When LLVM is determining the alignment for a given type, it uses the
1206 #. If the type sought is an exact match for one of the specifications,
1207 that specification is used.
1208 #. If no match is found, and the type sought is an integer type, then
1209 the smallest integer type that is larger than the bitwidth of the
1210 sought type is used. If none of the specifications are larger than
1211 the bitwidth then the largest integer type is used. For example,
1212 given the default specifications above, the i7 type will use the
1213 alignment of i8 (next largest) while both i65 and i256 will use the
1214 alignment of i64 (largest specified).
1215 #. If no match is found, and the type sought is a vector type, then the
1216 largest vector type that is smaller than the sought vector type will
1217 be used as a fall back. This happens because <128 x double> can be
1218 implemented in terms of 64 <2 x double>, for example.
1220 The function of the data layout string may not be what you expect.
1221 Notably, this is not a specification from the frontend of what alignment
1222 the code generator should use.
1224 Instead, if specified, the target data layout is required to match what
1225 the ultimate *code generator* expects. This string is used by the
1226 mid-level optimizers to improve code, and this only works if it matches
1227 what the ultimate code generator uses. If you would like to generate IR
1228 that does not embed this target-specific detail into the IR, then you
1229 don't have to specify the string. This will disable some optimizations
1230 that require precise layout information, but this also prevents those
1231 optimizations from introducing target specificity into the IR.
1238 A module may specify a target triple string that describes the target
1239 host. The syntax for the target triple is simply:
1241 .. code-block:: llvm
1243 target triple = "x86_64-apple-macosx10.7.0"
1245 The *target triple* string consists of a series of identifiers delimited
1246 by the minus sign character ('-'). The canonical forms are:
1250 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1251 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1253 This information is passed along to the backend so that it generates
1254 code for the proper architecture. It's possible to override this on the
1255 command line with the ``-mtriple`` command line option.
1257 .. _pointeraliasing:
1259 Pointer Aliasing Rules
1260 ----------------------
1262 Any memory access must be done through a pointer value associated with
1263 an address range of the memory access, otherwise the behavior is
1264 undefined. Pointer values are associated with address ranges according
1265 to the following rules:
1267 - A pointer value is associated with the addresses associated with any
1268 value it is *based* on.
1269 - An address of a global variable is associated with the address range
1270 of the variable's storage.
1271 - The result value of an allocation instruction is associated with the
1272 address range of the allocated storage.
1273 - A null pointer in the default address-space is associated with no
1275 - An integer constant other than zero or a pointer value returned from
1276 a function not defined within LLVM may be associated with address
1277 ranges allocated through mechanisms other than those provided by
1278 LLVM. Such ranges shall not overlap with any ranges of addresses
1279 allocated by mechanisms provided by LLVM.
1281 A pointer value is *based* on another pointer value according to the
1284 - A pointer value formed from a ``getelementptr`` operation is *based*
1285 on the first operand of the ``getelementptr``.
1286 - The result value of a ``bitcast`` is *based* on the operand of the
1288 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1289 values that contribute (directly or indirectly) to the computation of
1290 the pointer's value.
1291 - The "*based* on" relationship is transitive.
1293 Note that this definition of *"based"* is intentionally similar to the
1294 definition of *"based"* in C99, though it is slightly weaker.
1296 LLVM IR does not associate types with memory. The result type of a
1297 ``load`` merely indicates the size and alignment of the memory from
1298 which to load, as well as the interpretation of the value. The first
1299 operand type of a ``store`` similarly only indicates the size and
1300 alignment of the store.
1302 Consequently, type-based alias analysis, aka TBAA, aka
1303 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1304 :ref:`Metadata <metadata>` may be used to encode additional information
1305 which specialized optimization passes may use to implement type-based
1310 Volatile Memory Accesses
1311 ------------------------
1313 Certain memory accesses, such as :ref:`load <i_load>`'s,
1314 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1315 marked ``volatile``. The optimizers must not change the number of
1316 volatile operations or change their order of execution relative to other
1317 volatile operations. The optimizers *may* change the order of volatile
1318 operations relative to non-volatile operations. This is not Java's
1319 "volatile" and has no cross-thread synchronization behavior.
1321 IR-level volatile loads and stores cannot safely be optimized into
1322 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1323 flagged volatile. Likewise, the backend should never split or merge
1324 target-legal volatile load/store instructions.
1326 .. admonition:: Rationale
1328 Platforms may rely on volatile loads and stores of natively supported
1329 data width to be executed as single instruction. For example, in C
1330 this holds for an l-value of volatile primitive type with native
1331 hardware support, but not necessarily for aggregate types. The
1332 frontend upholds these expectations, which are intentionally
1333 unspecified in the IR. The rules above ensure that IR transformation
1334 do not violate the frontend's contract with the language.
1338 Memory Model for Concurrent Operations
1339 --------------------------------------
1341 The LLVM IR does not define any way to start parallel threads of
1342 execution or to register signal handlers. Nonetheless, there are
1343 platform-specific ways to create them, and we define LLVM IR's behavior
1344 in their presence. This model is inspired by the C++0x memory model.
1346 For a more informal introduction to this model, see the :doc:`Atomics`.
1348 We define a *happens-before* partial order as the least partial order
1351 - Is a superset of single-thread program order, and
1352 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1353 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1354 techniques, like pthread locks, thread creation, thread joining,
1355 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1356 Constraints <ordering>`).
1358 Note that program order does not introduce *happens-before* edges
1359 between a thread and signals executing inside that thread.
1361 Every (defined) read operation (load instructions, memcpy, atomic
1362 loads/read-modify-writes, etc.) R reads a series of bytes written by
1363 (defined) write operations (store instructions, atomic
1364 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1365 section, initialized globals are considered to have a write of the
1366 initializer which is atomic and happens before any other read or write
1367 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1368 may see any write to the same byte, except:
1370 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1371 write\ :sub:`2` happens before R\ :sub:`byte`, then
1372 R\ :sub:`byte` does not see write\ :sub:`1`.
1373 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1374 R\ :sub:`byte` does not see write\ :sub:`3`.
1376 Given that definition, R\ :sub:`byte` is defined as follows:
1378 - If R is volatile, the result is target-dependent. (Volatile is
1379 supposed to give guarantees which can support ``sig_atomic_t`` in
1380 C/C++, and may be used for accesses to addresses which do not behave
1381 like normal memory. It does not generally provide cross-thread
1383 - Otherwise, if there is no write to the same byte that happens before
1384 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1385 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1386 R\ :sub:`byte` returns the value written by that write.
1387 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1388 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1389 Memory Ordering Constraints <ordering>` section for additional
1390 constraints on how the choice is made.
1391 - Otherwise R\ :sub:`byte` returns ``undef``.
1393 R returns the value composed of the series of bytes it read. This
1394 implies that some bytes within the value may be ``undef`` **without**
1395 the entire value being ``undef``. Note that this only defines the
1396 semantics of the operation; it doesn't mean that targets will emit more
1397 than one instruction to read the series of bytes.
1399 Note that in cases where none of the atomic intrinsics are used, this
1400 model places only one restriction on IR transformations on top of what
1401 is required for single-threaded execution: introducing a store to a byte
1402 which might not otherwise be stored is not allowed in general.
1403 (Specifically, in the case where another thread might write to and read
1404 from an address, introducing a store can change a load that may see
1405 exactly one write into a load that may see multiple writes.)
1409 Atomic Memory Ordering Constraints
1410 ----------------------------------
1412 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1413 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1414 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1415 an ordering parameter that determines which other atomic instructions on
1416 the same address they *synchronize with*. These semantics are borrowed
1417 from Java and C++0x, but are somewhat more colloquial. If these
1418 descriptions aren't precise enough, check those specs (see spec
1419 references in the :doc:`atomics guide <Atomics>`).
1420 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1421 differently since they don't take an address. See that instruction's
1422 documentation for details.
1424 For a simpler introduction to the ordering constraints, see the
1428 The set of values that can be read is governed by the happens-before
1429 partial order. A value cannot be read unless some operation wrote
1430 it. This is intended to provide a guarantee strong enough to model
1431 Java's non-volatile shared variables. This ordering cannot be
1432 specified for read-modify-write operations; it is not strong enough
1433 to make them atomic in any interesting way.
1435 In addition to the guarantees of ``unordered``, there is a single
1436 total order for modifications by ``monotonic`` operations on each
1437 address. All modification orders must be compatible with the
1438 happens-before order. There is no guarantee that the modification
1439 orders can be combined to a global total order for the whole program
1440 (and this often will not be possible). The read in an atomic
1441 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1442 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1443 order immediately before the value it writes. If one atomic read
1444 happens before another atomic read of the same address, the later
1445 read must see the same value or a later value in the address's
1446 modification order. This disallows reordering of ``monotonic`` (or
1447 stronger) operations on the same address. If an address is written
1448 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1449 read that address repeatedly, the other threads must eventually see
1450 the write. This corresponds to the C++0x/C1x
1451 ``memory_order_relaxed``.
1453 In addition to the guarantees of ``monotonic``, a
1454 *synchronizes-with* edge may be formed with a ``release`` operation.
1455 This is intended to model C++'s ``memory_order_acquire``.
1457 In addition to the guarantees of ``monotonic``, if this operation
1458 writes a value which is subsequently read by an ``acquire``
1459 operation, it *synchronizes-with* that operation. (This isn't a
1460 complete description; see the C++0x definition of a release
1461 sequence.) This corresponds to the C++0x/C1x
1462 ``memory_order_release``.
1463 ``acq_rel`` (acquire+release)
1464 Acts as both an ``acquire`` and ``release`` operation on its
1465 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1466 ``seq_cst`` (sequentially consistent)
1467 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1468 operation which only reads, ``release`` for an operation which only
1469 writes), there is a global total order on all
1470 sequentially-consistent operations on all addresses, which is
1471 consistent with the *happens-before* partial order and with the
1472 modification orders of all the affected addresses. Each
1473 sequentially-consistent read sees the last preceding write to the
1474 same address in this global order. This corresponds to the C++0x/C1x
1475 ``memory_order_seq_cst`` and Java volatile.
1479 If an atomic operation is marked ``singlethread``, it only *synchronizes
1480 with* or participates in modification and seq\_cst total orderings with
1481 other operations running in the same thread (for example, in signal
1489 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1490 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1491 :ref:`frem <i_frem>`) have the following flags that can set to enable
1492 otherwise unsafe floating point operations
1495 No NaNs - Allow optimizations to assume the arguments and result are not
1496 NaN. Such optimizations are required to retain defined behavior over
1497 NaNs, but the value of the result is undefined.
1500 No Infs - Allow optimizations to assume the arguments and result are not
1501 +/-Inf. Such optimizations are required to retain defined behavior over
1502 +/-Inf, but the value of the result is undefined.
1505 No Signed Zeros - Allow optimizations to treat the sign of a zero
1506 argument or result as insignificant.
1509 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1510 argument rather than perform division.
1513 Fast - Allow algebraically equivalent transformations that may
1514 dramatically change results in floating point (e.g. reassociate). This
1515 flag implies all the others.
1522 The LLVM type system is one of the most important features of the
1523 intermediate representation. Being typed enables a number of
1524 optimizations to be performed on the intermediate representation
1525 directly, without having to do extra analyses on the side before the
1526 transformation. A strong type system makes it easier to read the
1527 generated code and enables novel analyses and transformations that are
1528 not feasible to perform on normal three address code representations.
1538 The void type does not represent any value and has no size.
1556 The function type can be thought of as a function signature. It consists of a
1557 return type and a list of formal parameter types. The return type of a function
1558 type is a void type or first class type --- except for :ref:`label <t_label>`
1559 and :ref:`metadata <t_metadata>` types.
1565 <returntype> (<parameter list>)
1567 ...where '``<parameter list>``' is a comma-separated list of type
1568 specifiers. Optionally, the parameter list may include a type ``...``, which
1569 indicates that the function takes a variable number of arguments. Variable
1570 argument functions can access their arguments with the :ref:`variable argument
1571 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
1572 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
1576 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1577 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1578 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1579 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1580 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1581 | ``i32 (i8*, ...)`` | A vararg function that takes at least one :ref:`pointer <t_pointer>` to ``i8`` (char in C), which returns an integer. This is the signature for ``printf`` in LLVM. |
1582 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1583 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1584 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1591 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1592 Values of these types are the only ones which can be produced by
1600 These are the types that are valid in registers from CodeGen's perspective.
1609 The integer type is a very simple type that simply specifies an
1610 arbitrary bit width for the integer type desired. Any bit width from 1
1611 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1619 The number of bits the integer will occupy is specified by the ``N``
1625 +----------------+------------------------------------------------+
1626 | ``i1`` | a single-bit integer. |
1627 +----------------+------------------------------------------------+
1628 | ``i32`` | a 32-bit integer. |
1629 +----------------+------------------------------------------------+
1630 | ``i1942652`` | a really big integer of over 1 million bits. |
1631 +----------------+------------------------------------------------+
1635 Floating Point Types
1636 """"""""""""""""""""
1645 - 16-bit floating point value
1648 - 32-bit floating point value
1651 - 64-bit floating point value
1654 - 128-bit floating point value (112-bit mantissa)
1657 - 80-bit floating point value (X87)
1660 - 128-bit floating point value (two 64-bits)
1669 The x86mmx type represents a value held in an MMX register on an x86
1670 machine. The operations allowed on it are quite limited: parameters and
1671 return values, load and store, and bitcast. User-specified MMX
1672 instructions are represented as intrinsic or asm calls with arguments
1673 and/or results of this type. There are no arrays, vectors or constants
1690 The pointer type is used to specify memory locations. Pointers are
1691 commonly used to reference objects in memory.
1693 Pointer types may have an optional address space attribute defining the
1694 numbered address space where the pointed-to object resides. The default
1695 address space is number zero. The semantics of non-zero address spaces
1696 are target-specific.
1698 Note that LLVM does not permit pointers to void (``void*``) nor does it
1699 permit pointers to labels (``label*``). Use ``i8*`` instead.
1709 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1710 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1711 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1712 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1713 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1714 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1715 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1724 A vector type is a simple derived type that represents a vector of
1725 elements. Vector types are used when multiple primitive data are
1726 operated in parallel using a single instruction (SIMD). A vector type
1727 requires a size (number of elements) and an underlying primitive data
1728 type. Vector types are considered :ref:`first class <t_firstclass>`.
1734 < <# elements> x <elementtype> >
1736 The number of elements is a constant integer value larger than 0;
1737 elementtype may be any integer or floating point type, or a pointer to
1738 these types. Vectors of size zero are not allowed.
1742 +-------------------+--------------------------------------------------+
1743 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
1744 +-------------------+--------------------------------------------------+
1745 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
1746 +-------------------+--------------------------------------------------+
1747 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
1748 +-------------------+--------------------------------------------------+
1749 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
1750 +-------------------+--------------------------------------------------+
1759 The label type represents code labels.
1774 The metadata type represents embedded metadata. No derived types may be
1775 created from metadata except for :ref:`function <t_function>` arguments.
1788 Aggregate Types are a subset of derived types that can contain multiple
1789 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
1790 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
1800 The array type is a very simple derived type that arranges elements
1801 sequentially in memory. The array type requires a size (number of
1802 elements) and an underlying data type.
1808 [<# elements> x <elementtype>]
1810 The number of elements is a constant integer value; ``elementtype`` may
1811 be any type with a size.
1815 +------------------+--------------------------------------+
1816 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
1817 +------------------+--------------------------------------+
1818 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
1819 +------------------+--------------------------------------+
1820 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
1821 +------------------+--------------------------------------+
1823 Here are some examples of multidimensional arrays:
1825 +-----------------------------+----------------------------------------------------------+
1826 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
1827 +-----------------------------+----------------------------------------------------------+
1828 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
1829 +-----------------------------+----------------------------------------------------------+
1830 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
1831 +-----------------------------+----------------------------------------------------------+
1833 There is no restriction on indexing beyond the end of the array implied
1834 by a static type (though there are restrictions on indexing beyond the
1835 bounds of an allocated object in some cases). This means that
1836 single-dimension 'variable sized array' addressing can be implemented in
1837 LLVM with a zero length array type. An implementation of 'pascal style
1838 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
1848 The structure type is used to represent a collection of data members
1849 together in memory. The elements of a structure may be any type that has
1852 Structures in memory are accessed using '``load``' and '``store``' by
1853 getting a pointer to a field with the '``getelementptr``' instruction.
1854 Structures in registers are accessed using the '``extractvalue``' and
1855 '``insertvalue``' instructions.
1857 Structures may optionally be "packed" structures, which indicate that
1858 the alignment of the struct is one byte, and that there is no padding
1859 between the elements. In non-packed structs, padding between field types
1860 is inserted as defined by the DataLayout string in the module, which is
1861 required to match what the underlying code generator expects.
1863 Structures can either be "literal" or "identified". A literal structure
1864 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
1865 identified types are always defined at the top level with a name.
1866 Literal types are uniqued by their contents and can never be recursive
1867 or opaque since there is no way to write one. Identified types can be
1868 recursive, can be opaqued, and are never uniqued.
1874 %T1 = type { <type list> } ; Identified normal struct type
1875 %T2 = type <{ <type list> }> ; Identified packed struct type
1879 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1880 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
1881 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1882 | ``{ float, i32 (i32) * }`` | A pair, where the first element is a ``float`` and the second element is a :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32``, returning an ``i32``. |
1883 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1884 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
1885 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1889 Opaque Structure Types
1890 """"""""""""""""""""""
1894 Opaque structure types are used to represent named structure types that
1895 do not have a body specified. This corresponds (for example) to the C
1896 notion of a forward declared structure.
1907 +--------------+-------------------+
1908 | ``opaque`` | An opaque type. |
1909 +--------------+-------------------+
1914 LLVM has several different basic types of constants. This section
1915 describes them all and their syntax.
1920 **Boolean constants**
1921 The two strings '``true``' and '``false``' are both valid constants
1923 **Integer constants**
1924 Standard integers (such as '4') are constants of the
1925 :ref:`integer <t_integer>` type. Negative numbers may be used with
1927 **Floating point constants**
1928 Floating point constants use standard decimal notation (e.g.
1929 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
1930 hexadecimal notation (see below). The assembler requires the exact
1931 decimal value of a floating-point constant. For example, the
1932 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
1933 decimal in binary. Floating point constants must have a :ref:`floating
1934 point <t_floating>` type.
1935 **Null pointer constants**
1936 The identifier '``null``' is recognized as a null pointer constant
1937 and must be of :ref:`pointer type <t_pointer>`.
1939 The one non-intuitive notation for constants is the hexadecimal form of
1940 floating point constants. For example, the form
1941 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
1942 than) '``double 4.5e+15``'. The only time hexadecimal floating point
1943 constants are required (and the only time that they are generated by the
1944 disassembler) is when a floating point constant must be emitted but it
1945 cannot be represented as a decimal floating point number in a reasonable
1946 number of digits. For example, NaN's, infinities, and other special
1947 values are represented in their IEEE hexadecimal format so that assembly
1948 and disassembly do not cause any bits to change in the constants.
1950 When using the hexadecimal form, constants of types half, float, and
1951 double are represented using the 16-digit form shown above (which
1952 matches the IEEE754 representation for double); half and float values
1953 must, however, be exactly representable as IEEE 754 half and single
1954 precision, respectively. Hexadecimal format is always used for long
1955 double, and there are three forms of long double. The 80-bit format used
1956 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
1957 128-bit format used by PowerPC (two adjacent doubles) is represented by
1958 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
1959 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
1960 will only work if they match the long double format on your target.
1961 The IEEE 16-bit format (half precision) is represented by ``0xH``
1962 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
1963 (sign bit at the left).
1965 There are no constants of type x86mmx.
1967 .. _complexconstants:
1972 Complex constants are a (potentially recursive) combination of simple
1973 constants and smaller complex constants.
1975 **Structure constants**
1976 Structure constants are represented with notation similar to
1977 structure type definitions (a comma separated list of elements,
1978 surrounded by braces (``{}``)). For example:
1979 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
1980 "``@G = external global i32``". Structure constants must have
1981 :ref:`structure type <t_struct>`, and the number and types of elements
1982 must match those specified by the type.
1984 Array constants are represented with notation similar to array type
1985 definitions (a comma separated list of elements, surrounded by
1986 square brackets (``[]``)). For example:
1987 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
1988 :ref:`array type <t_array>`, and the number and types of elements must
1989 match those specified by the type.
1990 **Vector constants**
1991 Vector constants are represented with notation similar to vector
1992 type definitions (a comma separated list of elements, surrounded by
1993 less-than/greater-than's (``<>``)). For example:
1994 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
1995 must have :ref:`vector type <t_vector>`, and the number and types of
1996 elements must match those specified by the type.
1997 **Zero initialization**
1998 The string '``zeroinitializer``' can be used to zero initialize a
1999 value to zero of *any* type, including scalar and
2000 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2001 having to print large zero initializers (e.g. for large arrays) and
2002 is always exactly equivalent to using explicit zero initializers.
2004 A metadata node is a structure-like constant with :ref:`metadata
2005 type <t_metadata>`. For example:
2006 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
2007 constants that are meant to be interpreted as part of the
2008 instruction stream, metadata is a place to attach additional
2009 information such as debug info.
2011 Global Variable and Function Addresses
2012 --------------------------------------
2014 The addresses of :ref:`global variables <globalvars>` and
2015 :ref:`functions <functionstructure>` are always implicitly valid
2016 (link-time) constants. These constants are explicitly referenced when
2017 the :ref:`identifier for the global <identifiers>` is used and always have
2018 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2021 .. code-block:: llvm
2025 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2032 The string '``undef``' can be used anywhere a constant is expected, and
2033 indicates that the user of the value may receive an unspecified
2034 bit-pattern. Undefined values may be of any type (other than '``label``'
2035 or '``void``') and be used anywhere a constant is permitted.
2037 Undefined values are useful because they indicate to the compiler that
2038 the program is well defined no matter what value is used. This gives the
2039 compiler more freedom to optimize. Here are some examples of
2040 (potentially surprising) transformations that are valid (in pseudo IR):
2042 .. code-block:: llvm
2052 This is safe because all of the output bits are affected by the undef
2053 bits. Any output bit can have a zero or one depending on the input bits.
2055 .. code-block:: llvm
2066 These logical operations have bits that are not always affected by the
2067 input. For example, if ``%X`` has a zero bit, then the output of the
2068 '``and``' operation will always be a zero for that bit, no matter what
2069 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2070 optimize or assume that the result of the '``and``' is '``undef``'.
2071 However, it is safe to assume that all bits of the '``undef``' could be
2072 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2073 all the bits of the '``undef``' operand to the '``or``' could be set,
2074 allowing the '``or``' to be folded to -1.
2076 .. code-block:: llvm
2078 %A = select undef, %X, %Y
2079 %B = select undef, 42, %Y
2080 %C = select %X, %Y, undef
2090 This set of examples shows that undefined '``select``' (and conditional
2091 branch) conditions can go *either way*, but they have to come from one
2092 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2093 both known to have a clear low bit, then ``%A`` would have to have a
2094 cleared low bit. However, in the ``%C`` example, the optimizer is
2095 allowed to assume that the '``undef``' operand could be the same as
2096 ``%Y``, allowing the whole '``select``' to be eliminated.
2098 .. code-block:: llvm
2100 %A = xor undef, undef
2117 This example points out that two '``undef``' operands are not
2118 necessarily the same. This can be surprising to people (and also matches
2119 C semantics) where they assume that "``X^X``" is always zero, even if
2120 ``X`` is undefined. This isn't true for a number of reasons, but the
2121 short answer is that an '``undef``' "variable" can arbitrarily change
2122 its value over its "live range". This is true because the variable
2123 doesn't actually *have a live range*. Instead, the value is logically
2124 read from arbitrary registers that happen to be around when needed, so
2125 the value is not necessarily consistent over time. In fact, ``%A`` and
2126 ``%C`` need to have the same semantics or the core LLVM "replace all
2127 uses with" concept would not hold.
2129 .. code-block:: llvm
2137 These examples show the crucial difference between an *undefined value*
2138 and *undefined behavior*. An undefined value (like '``undef``') is
2139 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2140 operation can be constant folded to '``undef``', because the '``undef``'
2141 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2142 However, in the second example, we can make a more aggressive
2143 assumption: because the ``undef`` is allowed to be an arbitrary value,
2144 we are allowed to assume that it could be zero. Since a divide by zero
2145 has *undefined behavior*, we are allowed to assume that the operation
2146 does not execute at all. This allows us to delete the divide and all
2147 code after it. Because the undefined operation "can't happen", the
2148 optimizer can assume that it occurs in dead code.
2150 .. code-block:: llvm
2152 a: store undef -> %X
2153 b: store %X -> undef
2158 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2159 value can be assumed to not have any effect; we can assume that the
2160 value is overwritten with bits that happen to match what was already
2161 there. However, a store *to* an undefined location could clobber
2162 arbitrary memory, therefore, it has undefined behavior.
2169 Poison values are similar to :ref:`undef values <undefvalues>`, however
2170 they also represent the fact that an instruction or constant expression
2171 which cannot evoke side effects has nevertheless detected a condition
2172 which results in undefined behavior.
2174 There is currently no way of representing a poison value in the IR; they
2175 only exist when produced by operations such as :ref:`add <i_add>` with
2178 Poison value behavior is defined in terms of value *dependence*:
2180 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2181 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2182 their dynamic predecessor basic block.
2183 - Function arguments depend on the corresponding actual argument values
2184 in the dynamic callers of their functions.
2185 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2186 instructions that dynamically transfer control back to them.
2187 - :ref:`Invoke <i_invoke>` instructions depend on the
2188 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2189 call instructions that dynamically transfer control back to them.
2190 - Non-volatile loads and stores depend on the most recent stores to all
2191 of the referenced memory addresses, following the order in the IR
2192 (including loads and stores implied by intrinsics such as
2193 :ref:`@llvm.memcpy <int_memcpy>`.)
2194 - An instruction with externally visible side effects depends on the
2195 most recent preceding instruction with externally visible side
2196 effects, following the order in the IR. (This includes :ref:`volatile
2197 operations <volatile>`.)
2198 - An instruction *control-depends* on a :ref:`terminator
2199 instruction <terminators>` if the terminator instruction has
2200 multiple successors and the instruction is always executed when
2201 control transfers to one of the successors, and may not be executed
2202 when control is transferred to another.
2203 - Additionally, an instruction also *control-depends* on a terminator
2204 instruction if the set of instructions it otherwise depends on would
2205 be different if the terminator had transferred control to a different
2207 - Dependence is transitive.
2209 Poison Values have the same behavior as :ref:`undef values <undefvalues>`,
2210 with the additional affect that any instruction which has a *dependence*
2211 on a poison value has undefined behavior.
2213 Here are some examples:
2215 .. code-block:: llvm
2218 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2219 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2220 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2221 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2223 store i32 %poison, i32* @g ; Poison value stored to memory.
2224 %poison2 = load i32* @g ; Poison value loaded back from memory.
2226 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2228 %narrowaddr = bitcast i32* @g to i16*
2229 %wideaddr = bitcast i32* @g to i64*
2230 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2231 %poison4 = load i64* %wideaddr ; Returns a poison value.
2233 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2234 br i1 %cmp, label %true, label %end ; Branch to either destination.
2237 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2238 ; it has undefined behavior.
2242 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2243 ; Both edges into this PHI are
2244 ; control-dependent on %cmp, so this
2245 ; always results in a poison value.
2247 store volatile i32 0, i32* @g ; This would depend on the store in %true
2248 ; if %cmp is true, or the store in %entry
2249 ; otherwise, so this is undefined behavior.
2251 br i1 %cmp, label %second_true, label %second_end
2252 ; The same branch again, but this time the
2253 ; true block doesn't have side effects.
2260 store volatile i32 0, i32* @g ; This time, the instruction always depends
2261 ; on the store in %end. Also, it is
2262 ; control-equivalent to %end, so this is
2263 ; well-defined (ignoring earlier undefined
2264 ; behavior in this example).
2268 Addresses of Basic Blocks
2269 -------------------------
2271 ``blockaddress(@function, %block)``
2273 The '``blockaddress``' constant computes the address of the specified
2274 basic block in the specified function, and always has an ``i8*`` type.
2275 Taking the address of the entry block is illegal.
2277 This value only has defined behavior when used as an operand to the
2278 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2279 against null. Pointer equality tests between labels addresses results in
2280 undefined behavior --- though, again, comparison against null is ok, and
2281 no label is equal to the null pointer. This may be passed around as an
2282 opaque pointer sized value as long as the bits are not inspected. This
2283 allows ``ptrtoint`` and arithmetic to be performed on these values so
2284 long as the original value is reconstituted before the ``indirectbr``
2287 Finally, some targets may provide defined semantics when using the value
2288 as the operand to an inline assembly, but that is target specific.
2292 Constant Expressions
2293 --------------------
2295 Constant expressions are used to allow expressions involving other
2296 constants to be used as constants. Constant expressions may be of any
2297 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2298 that does not have side effects (e.g. load and call are not supported).
2299 The following is the syntax for constant expressions:
2301 ``trunc (CST to TYPE)``
2302 Truncate a constant to another type. The bit size of CST must be
2303 larger than the bit size of TYPE. Both types must be integers.
2304 ``zext (CST to TYPE)``
2305 Zero extend a constant to another type. The bit size of CST must be
2306 smaller than the bit size of TYPE. Both types must be integers.
2307 ``sext (CST to TYPE)``
2308 Sign extend a constant to another type. The bit size of CST must be
2309 smaller than the bit size of TYPE. Both types must be integers.
2310 ``fptrunc (CST to TYPE)``
2311 Truncate a floating point constant to another floating point type.
2312 The size of CST must be larger than the size of TYPE. Both types
2313 must be floating point.
2314 ``fpext (CST to TYPE)``
2315 Floating point extend a constant to another type. The size of CST
2316 must be smaller or equal to the size of TYPE. Both types must be
2318 ``fptoui (CST to TYPE)``
2319 Convert a floating point constant to the corresponding unsigned
2320 integer constant. TYPE must be a scalar or vector integer type. CST
2321 must be of scalar or vector floating point type. Both CST and TYPE
2322 must be scalars, or vectors of the same number of elements. If the
2323 value won't fit in the integer type, the results are undefined.
2324 ``fptosi (CST to TYPE)``
2325 Convert a floating point constant to the corresponding signed
2326 integer constant. TYPE must be a scalar or vector integer type. CST
2327 must be of scalar or vector floating point type. Both CST and TYPE
2328 must be scalars, or vectors of the same number of elements. If the
2329 value won't fit in the integer type, the results are undefined.
2330 ``uitofp (CST to TYPE)``
2331 Convert an unsigned integer constant to the corresponding floating
2332 point constant. TYPE must be a scalar or vector floating point type.
2333 CST must be of scalar or vector integer type. Both CST and TYPE must
2334 be scalars, or vectors of the same number of elements. If the value
2335 won't fit in the floating point type, the results are undefined.
2336 ``sitofp (CST to TYPE)``
2337 Convert a signed integer constant to the corresponding floating
2338 point constant. TYPE must be a scalar or vector floating point type.
2339 CST must be of scalar or vector integer type. Both CST and TYPE must
2340 be scalars, or vectors of the same number of elements. If the value
2341 won't fit in the floating point type, the results are undefined.
2342 ``ptrtoint (CST to TYPE)``
2343 Convert a pointer typed constant to the corresponding integer
2344 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2345 pointer type. The ``CST`` value is zero extended, truncated, or
2346 unchanged to make it fit in ``TYPE``.
2347 ``inttoptr (CST to TYPE)``
2348 Convert an integer constant to a pointer constant. TYPE must be a
2349 pointer type. CST must be of integer type. The CST value is zero
2350 extended, truncated, or unchanged to make it fit in a pointer size.
2351 This one is *really* dangerous!
2352 ``bitcast (CST to TYPE)``
2353 Convert a constant, CST, to another TYPE. The constraints of the
2354 operands are the same as those for the :ref:`bitcast
2355 instruction <i_bitcast>`.
2356 ``addrspacecast (CST to TYPE)``
2357 Convert a constant pointer or constant vector of pointer, CST, to another
2358 TYPE in a different address space. The constraints of the operands are the
2359 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2360 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2361 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2362 constants. As with the :ref:`getelementptr <i_getelementptr>`
2363 instruction, the index list may have zero or more indexes, which are
2364 required to make sense for the type of "CSTPTR".
2365 ``select (COND, VAL1, VAL2)``
2366 Perform the :ref:`select operation <i_select>` on constants.
2367 ``icmp COND (VAL1, VAL2)``
2368 Performs the :ref:`icmp operation <i_icmp>` on constants.
2369 ``fcmp COND (VAL1, VAL2)``
2370 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2371 ``extractelement (VAL, IDX)``
2372 Perform the :ref:`extractelement operation <i_extractelement>` on
2374 ``insertelement (VAL, ELT, IDX)``
2375 Perform the :ref:`insertelement operation <i_insertelement>` on
2377 ``shufflevector (VEC1, VEC2, IDXMASK)``
2378 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2380 ``extractvalue (VAL, IDX0, IDX1, ...)``
2381 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2382 constants. The index list is interpreted in a similar manner as
2383 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2384 least one index value must be specified.
2385 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2386 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2387 The index list is interpreted in a similar manner as indices in a
2388 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2389 value must be specified.
2390 ``OPCODE (LHS, RHS)``
2391 Perform the specified operation of the LHS and RHS constants. OPCODE
2392 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2393 binary <bitwiseops>` operations. The constraints on operands are
2394 the same as those for the corresponding instruction (e.g. no bitwise
2395 operations on floating point values are allowed).
2402 Inline Assembler Expressions
2403 ----------------------------
2405 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2406 Inline Assembly <moduleasm>`) through the use of a special value. This
2407 value represents the inline assembler as a string (containing the
2408 instructions to emit), a list of operand constraints (stored as a
2409 string), a flag that indicates whether or not the inline asm expression
2410 has side effects, and a flag indicating whether the function containing
2411 the asm needs to align its stack conservatively. An example inline
2412 assembler expression is:
2414 .. code-block:: llvm
2416 i32 (i32) asm "bswap $0", "=r,r"
2418 Inline assembler expressions may **only** be used as the callee operand
2419 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2420 Thus, typically we have:
2422 .. code-block:: llvm
2424 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2426 Inline asms with side effects not visible in the constraint list must be
2427 marked as having side effects. This is done through the use of the
2428 '``sideeffect``' keyword, like so:
2430 .. code-block:: llvm
2432 call void asm sideeffect "eieio", ""()
2434 In some cases inline asms will contain code that will not work unless
2435 the stack is aligned in some way, such as calls or SSE instructions on
2436 x86, yet will not contain code that does that alignment within the asm.
2437 The compiler should make conservative assumptions about what the asm
2438 might contain and should generate its usual stack alignment code in the
2439 prologue if the '``alignstack``' keyword is present:
2441 .. code-block:: llvm
2443 call void asm alignstack "eieio", ""()
2445 Inline asms also support using non-standard assembly dialects. The
2446 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2447 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2448 the only supported dialects. An example is:
2450 .. code-block:: llvm
2452 call void asm inteldialect "eieio", ""()
2454 If multiple keywords appear the '``sideeffect``' keyword must come
2455 first, the '``alignstack``' keyword second and the '``inteldialect``'
2461 The call instructions that wrap inline asm nodes may have a
2462 "``!srcloc``" MDNode attached to it that contains a list of constant
2463 integers. If present, the code generator will use the integer as the
2464 location cookie value when report errors through the ``LLVMContext``
2465 error reporting mechanisms. This allows a front-end to correlate backend
2466 errors that occur with inline asm back to the source code that produced
2469 .. code-block:: llvm
2471 call void asm sideeffect "something bad", ""(), !srcloc !42
2473 !42 = !{ i32 1234567 }
2475 It is up to the front-end to make sense of the magic numbers it places
2476 in the IR. If the MDNode contains multiple constants, the code generator
2477 will use the one that corresponds to the line of the asm that the error
2482 Metadata Nodes and Metadata Strings
2483 -----------------------------------
2485 LLVM IR allows metadata to be attached to instructions in the program
2486 that can convey extra information about the code to the optimizers and
2487 code generator. One example application of metadata is source-level
2488 debug information. There are two metadata primitives: strings and nodes.
2489 All metadata has the ``metadata`` type and is identified in syntax by a
2490 preceding exclamation point ('``!``').
2492 A metadata string is a string surrounded by double quotes. It can
2493 contain any character by escaping non-printable characters with
2494 "``\xx``" where "``xx``" is the two digit hex code. For example:
2497 Metadata nodes are represented with notation similar to structure
2498 constants (a comma separated list of elements, surrounded by braces and
2499 preceded by an exclamation point). Metadata nodes can have any values as
2500 their operand. For example:
2502 .. code-block:: llvm
2504 !{ metadata !"test\00", i32 10}
2506 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2507 metadata nodes, which can be looked up in the module symbol table. For
2510 .. code-block:: llvm
2512 !foo = metadata !{!4, !3}
2514 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2515 function is using two metadata arguments:
2517 .. code-block:: llvm
2519 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2521 Metadata can be attached with an instruction. Here metadata ``!21`` is
2522 attached to the ``add`` instruction using the ``!dbg`` identifier:
2524 .. code-block:: llvm
2526 %indvar.next = add i64 %indvar, 1, !dbg !21
2528 More information about specific metadata nodes recognized by the
2529 optimizers and code generator is found below.
2534 In LLVM IR, memory does not have types, so LLVM's own type system is not
2535 suitable for doing TBAA. Instead, metadata is added to the IR to
2536 describe a type system of a higher level language. This can be used to
2537 implement typical C/C++ TBAA, but it can also be used to implement
2538 custom alias analysis behavior for other languages.
2540 The current metadata format is very simple. TBAA metadata nodes have up
2541 to three fields, e.g.:
2543 .. code-block:: llvm
2545 !0 = metadata !{ metadata !"an example type tree" }
2546 !1 = metadata !{ metadata !"int", metadata !0 }
2547 !2 = metadata !{ metadata !"float", metadata !0 }
2548 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2550 The first field is an identity field. It can be any value, usually a
2551 metadata string, which uniquely identifies the type. The most important
2552 name in the tree is the name of the root node. Two trees with different
2553 root node names are entirely disjoint, even if they have leaves with
2556 The second field identifies the type's parent node in the tree, or is
2557 null or omitted for a root node. A type is considered to alias all of
2558 its descendants and all of its ancestors in the tree. Also, a type is
2559 considered to alias all types in other trees, so that bitcode produced
2560 from multiple front-ends is handled conservatively.
2562 If the third field is present, it's an integer which if equal to 1
2563 indicates that the type is "constant" (meaning
2564 ``pointsToConstantMemory`` should return true; see `other useful
2565 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2567 '``tbaa.struct``' Metadata
2568 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2570 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2571 aggregate assignment operations in C and similar languages, however it
2572 is defined to copy a contiguous region of memory, which is more than
2573 strictly necessary for aggregate types which contain holes due to
2574 padding. Also, it doesn't contain any TBAA information about the fields
2577 ``!tbaa.struct`` metadata can describe which memory subregions in a
2578 memcpy are padding and what the TBAA tags of the struct are.
2580 The current metadata format is very simple. ``!tbaa.struct`` metadata
2581 nodes are a list of operands which are in conceptual groups of three.
2582 For each group of three, the first operand gives the byte offset of a
2583 field in bytes, the second gives its size in bytes, and the third gives
2586 .. code-block:: llvm
2588 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2590 This describes a struct with two fields. The first is at offset 0 bytes
2591 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2592 and has size 4 bytes and has tbaa tag !2.
2594 Note that the fields need not be contiguous. In this example, there is a
2595 4 byte gap between the two fields. This gap represents padding which
2596 does not carry useful data and need not be preserved.
2598 '``fpmath``' Metadata
2599 ^^^^^^^^^^^^^^^^^^^^^
2601 ``fpmath`` metadata may be attached to any instruction of floating point
2602 type. It can be used to express the maximum acceptable error in the
2603 result of that instruction, in ULPs, thus potentially allowing the
2604 compiler to use a more efficient but less accurate method of computing
2605 it. ULP is defined as follows:
2607 If ``x`` is a real number that lies between two finite consecutive
2608 floating-point numbers ``a`` and ``b``, without being equal to one
2609 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2610 distance between the two non-equal finite floating-point numbers
2611 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2613 The metadata node shall consist of a single positive floating point
2614 number representing the maximum relative error, for example:
2616 .. code-block:: llvm
2618 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2620 '``range``' Metadata
2621 ^^^^^^^^^^^^^^^^^^^^
2623 ``range`` metadata may be attached only to loads of integer types. It
2624 expresses the possible ranges the loaded value is in. The ranges are
2625 represented with a flattened list of integers. The loaded value is known
2626 to be in the union of the ranges defined by each consecutive pair. Each
2627 pair has the following properties:
2629 - The type must match the type loaded by the instruction.
2630 - The pair ``a,b`` represents the range ``[a,b)``.
2631 - Both ``a`` and ``b`` are constants.
2632 - The range is allowed to wrap.
2633 - The range should not represent the full or empty set. That is,
2636 In addition, the pairs must be in signed order of the lower bound and
2637 they must be non-contiguous.
2641 .. code-block:: llvm
2643 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2644 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2645 %c = load i8* %z, align 1, !range !2 ; Can only be 0, 1, 3, 4 or 5
2646 %d = load i8* %z, align 1, !range !3 ; Can only be -2, -1, 3, 4 or 5
2648 !0 = metadata !{ i8 0, i8 2 }
2649 !1 = metadata !{ i8 255, i8 2 }
2650 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2651 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2656 It is sometimes useful to attach information to loop constructs. Currently,
2657 loop metadata is implemented as metadata attached to the branch instruction
2658 in the loop latch block. This type of metadata refer to a metadata node that is
2659 guaranteed to be separate for each loop. The loop identifier metadata is
2660 specified with the name ``llvm.loop``.
2662 The loop identifier metadata is implemented using a metadata that refers to
2663 itself to avoid merging it with any other identifier metadata, e.g.,
2664 during module linkage or function inlining. That is, each loop should refer
2665 to their own identification metadata even if they reside in separate functions.
2666 The following example contains loop identifier metadata for two separate loop
2669 .. code-block:: llvm
2671 !0 = metadata !{ metadata !0 }
2672 !1 = metadata !{ metadata !1 }
2674 The loop identifier metadata can be used to specify additional per-loop
2675 metadata. Any operands after the first operand can be treated as user-defined
2676 metadata. For example the ``llvm.vectorizer.unroll`` metadata is understood
2677 by the loop vectorizer to indicate how many times to unroll the loop:
2679 .. code-block:: llvm
2681 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
2683 !0 = metadata !{ metadata !0, metadata !1 }
2684 !1 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 2 }
2689 Metadata types used to annotate memory accesses with information helpful
2690 for optimizations are prefixed with ``llvm.mem``.
2692 '``llvm.mem.parallel_loop_access``' Metadata
2693 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2695 For a loop to be parallel, in addition to using
2696 the ``llvm.loop`` metadata to mark the loop latch branch instruction,
2697 also all of the memory accessing instructions in the loop body need to be
2698 marked with the ``llvm.mem.parallel_loop_access`` metadata. If there
2699 is at least one memory accessing instruction not marked with the metadata,
2700 the loop must be considered a sequential loop. This causes parallel loops to be
2701 converted to sequential loops due to optimization passes that are unaware of
2702 the parallel semantics and that insert new memory instructions to the loop
2705 Example of a loop that is considered parallel due to its correct use of
2706 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
2707 metadata types that refer to the same loop identifier metadata.
2709 .. code-block:: llvm
2713 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2715 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2717 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
2721 !0 = metadata !{ metadata !0 }
2723 It is also possible to have nested parallel loops. In that case the
2724 memory accesses refer to a list of loop identifier metadata nodes instead of
2725 the loop identifier metadata node directly:
2727 .. code-block:: llvm
2734 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2736 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2738 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
2742 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2744 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2746 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
2748 outer.for.end: ; preds = %for.body
2750 !0 = metadata !{ metadata !1, metadata !2 } ; a list of loop identifiers
2751 !1 = metadata !{ metadata !1 } ; an identifier for the inner loop
2752 !2 = metadata !{ metadata !2 } ; an identifier for the outer loop
2754 '``llvm.vectorizer``'
2755 ^^^^^^^^^^^^^^^^^^^^^
2757 Metadata prefixed with ``llvm.vectorizer`` is used to control per-loop
2758 vectorization parameters such as vectorization factor and unroll factor.
2760 ``llvm.vectorizer`` metadata should be used in conjunction with ``llvm.loop``
2761 loop identification metadata.
2763 '``llvm.vectorizer.unroll``' Metadata
2764 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2766 This metadata instructs the loop vectorizer to unroll the specified
2767 loop exactly ``N`` times.
2769 The first operand is the string ``llvm.vectorizer.unroll`` and the second
2770 operand is an integer specifying the unroll factor. For example:
2772 .. code-block:: llvm
2774 !0 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 4 }
2776 Note that setting ``llvm.vectorizer.unroll`` to 1 disables unrolling of the
2779 If ``llvm.vectorizer.unroll`` is set to 0 then the amount of unrolling will be
2780 determined automatically.
2782 '``llvm.vectorizer.width``' Metadata
2783 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2785 This metadata sets the target width of the vectorizer to ``N``. Without
2786 this metadata, the vectorizer will choose a width automatically.
2787 Regardless of this metadata, the vectorizer will only vectorize loops if
2788 it believes it is valid to do so.
2790 The first operand is the string ``llvm.vectorizer.width`` and the second
2791 operand is an integer specifying the width. For example:
2793 .. code-block:: llvm
2795 !0 = metadata !{ metadata !"llvm.vectorizer.width", i32 4 }
2797 Note that setting ``llvm.vectorizer.width`` to 1 disables vectorization of the
2800 If ``llvm.vectorizer.width`` is set to 0 then the width will be determined
2803 Module Flags Metadata
2804 =====================
2806 Information about the module as a whole is difficult to convey to LLVM's
2807 subsystems. The LLVM IR isn't sufficient to transmit this information.
2808 The ``llvm.module.flags`` named metadata exists in order to facilitate
2809 this. These flags are in the form of key / value pairs --- much like a
2810 dictionary --- making it easy for any subsystem who cares about a flag to
2813 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
2814 Each triplet has the following form:
2816 - The first element is a *behavior* flag, which specifies the behavior
2817 when two (or more) modules are merged together, and it encounters two
2818 (or more) metadata with the same ID. The supported behaviors are
2820 - The second element is a metadata string that is a unique ID for the
2821 metadata. Each module may only have one flag entry for each unique ID (not
2822 including entries with the **Require** behavior).
2823 - The third element is the value of the flag.
2825 When two (or more) modules are merged together, the resulting
2826 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
2827 each unique metadata ID string, there will be exactly one entry in the merged
2828 modules ``llvm.module.flags`` metadata table, and the value for that entry will
2829 be determined by the merge behavior flag, as described below. The only exception
2830 is that entries with the *Require* behavior are always preserved.
2832 The following behaviors are supported:
2843 Emits an error if two values disagree, otherwise the resulting value
2844 is that of the operands.
2848 Emits a warning if two values disagree. The result value will be the
2849 operand for the flag from the first module being linked.
2853 Adds a requirement that another module flag be present and have a
2854 specified value after linking is performed. The value must be a
2855 metadata pair, where the first element of the pair is the ID of the
2856 module flag to be restricted, and the second element of the pair is
2857 the value the module flag should be restricted to. This behavior can
2858 be used to restrict the allowable results (via triggering of an
2859 error) of linking IDs with the **Override** behavior.
2863 Uses the specified value, regardless of the behavior or value of the
2864 other module. If both modules specify **Override**, but the values
2865 differ, an error will be emitted.
2869 Appends the two values, which are required to be metadata nodes.
2873 Appends the two values, which are required to be metadata
2874 nodes. However, duplicate entries in the second list are dropped
2875 during the append operation.
2877 It is an error for a particular unique flag ID to have multiple behaviors,
2878 except in the case of **Require** (which adds restrictions on another metadata
2879 value) or **Override**.
2881 An example of module flags:
2883 .. code-block:: llvm
2885 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
2886 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
2887 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
2888 !3 = metadata !{ i32 3, metadata !"qux",
2890 metadata !"foo", i32 1
2893 !llvm.module.flags = !{ !0, !1, !2, !3 }
2895 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
2896 if two or more ``!"foo"`` flags are seen is to emit an error if their
2897 values are not equal.
2899 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
2900 behavior if two or more ``!"bar"`` flags are seen is to use the value
2903 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
2904 behavior if two or more ``!"qux"`` flags are seen is to emit a
2905 warning if their values are not equal.
2907 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
2911 metadata !{ metadata !"foo", i32 1 }
2913 The behavior is to emit an error if the ``llvm.module.flags`` does not
2914 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
2917 Objective-C Garbage Collection Module Flags Metadata
2918 ----------------------------------------------------
2920 On the Mach-O platform, Objective-C stores metadata about garbage
2921 collection in a special section called "image info". The metadata
2922 consists of a version number and a bitmask specifying what types of
2923 garbage collection are supported (if any) by the file. If two or more
2924 modules are linked together their garbage collection metadata needs to
2925 be merged rather than appended together.
2927 The Objective-C garbage collection module flags metadata consists of the
2928 following key-value pairs:
2937 * - ``Objective-C Version``
2938 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
2940 * - ``Objective-C Image Info Version``
2941 - **[Required]** --- The version of the image info section. Currently
2944 * - ``Objective-C Image Info Section``
2945 - **[Required]** --- The section to place the metadata. Valid values are
2946 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
2947 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
2948 Objective-C ABI version 2.
2950 * - ``Objective-C Garbage Collection``
2951 - **[Required]** --- Specifies whether garbage collection is supported or
2952 not. Valid values are 0, for no garbage collection, and 2, for garbage
2953 collection supported.
2955 * - ``Objective-C GC Only``
2956 - **[Optional]** --- Specifies that only garbage collection is supported.
2957 If present, its value must be 6. This flag requires that the
2958 ``Objective-C Garbage Collection`` flag have the value 2.
2960 Some important flag interactions:
2962 - If a module with ``Objective-C Garbage Collection`` set to 0 is
2963 merged with a module with ``Objective-C Garbage Collection`` set to
2964 2, then the resulting module has the
2965 ``Objective-C Garbage Collection`` flag set to 0.
2966 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
2967 merged with a module with ``Objective-C GC Only`` set to 6.
2969 Automatic Linker Flags Module Flags Metadata
2970 --------------------------------------------
2972 Some targets support embedding flags to the linker inside individual object
2973 files. Typically this is used in conjunction with language extensions which
2974 allow source files to explicitly declare the libraries they depend on, and have
2975 these automatically be transmitted to the linker via object files.
2977 These flags are encoded in the IR using metadata in the module flags section,
2978 using the ``Linker Options`` key. The merge behavior for this flag is required
2979 to be ``AppendUnique``, and the value for the key is expected to be a metadata
2980 node which should be a list of other metadata nodes, each of which should be a
2981 list of metadata strings defining linker options.
2983 For example, the following metadata section specifies two separate sets of
2984 linker options, presumably to link against ``libz`` and the ``Cocoa``
2987 !0 = metadata !{ i32 6, metadata !"Linker Options",
2989 metadata !{ metadata !"-lz" },
2990 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
2991 !llvm.module.flags = !{ !0 }
2993 The metadata encoding as lists of lists of options, as opposed to a collapsed
2994 list of options, is chosen so that the IR encoding can use multiple option
2995 strings to specify e.g., a single library, while still having that specifier be
2996 preserved as an atomic element that can be recognized by a target specific
2997 assembly writer or object file emitter.
2999 Each individual option is required to be either a valid option for the target's
3000 linker, or an option that is reserved by the target specific assembly writer or
3001 object file emitter. No other aspect of these options is defined by the IR.
3003 .. _intrinsicglobalvariables:
3005 Intrinsic Global Variables
3006 ==========================
3008 LLVM has a number of "magic" global variables that contain data that
3009 affect code generation or other IR semantics. These are documented here.
3010 All globals of this sort should have a section specified as
3011 "``llvm.metadata``". This section and all globals that start with
3012 "``llvm.``" are reserved for use by LLVM.
3016 The '``llvm.used``' Global Variable
3017 -----------------------------------
3019 The ``@llvm.used`` global is an array which has
3020 :ref:`appending linkage <linkage_appending>`. This array contains a list of
3021 pointers to named global variables, functions and aliases which may optionally
3022 have a pointer cast formed of bitcast or getelementptr. For example, a legal
3025 .. code-block:: llvm
3030 @llvm.used = appending global [2 x i8*] [
3032 i8* bitcast (i32* @Y to i8*)
3033 ], section "llvm.metadata"
3035 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
3036 and linker are required to treat the symbol as if there is a reference to the
3037 symbol that it cannot see (which is why they have to be named). For example, if
3038 a variable has internal linkage and no references other than that from the
3039 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
3040 references from inline asms and other things the compiler cannot "see", and
3041 corresponds to "``attribute((used))``" in GNU C.
3043 On some targets, the code generator must emit a directive to the
3044 assembler or object file to prevent the assembler and linker from
3045 molesting the symbol.
3047 .. _gv_llvmcompilerused:
3049 The '``llvm.compiler.used``' Global Variable
3050 --------------------------------------------
3052 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
3053 directive, except that it only prevents the compiler from touching the
3054 symbol. On targets that support it, this allows an intelligent linker to
3055 optimize references to the symbol without being impeded as it would be
3058 This is a rare construct that should only be used in rare circumstances,
3059 and should not be exposed to source languages.
3061 .. _gv_llvmglobalctors:
3063 The '``llvm.global_ctors``' Global Variable
3064 -------------------------------------------
3066 .. code-block:: llvm
3068 %0 = type { i32, void ()* }
3069 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor }]
3071 The ``@llvm.global_ctors`` array contains a list of constructor
3072 functions and associated priorities. The functions referenced by this
3073 array will be called in ascending order of priority (i.e. lowest first)
3074 when the module is loaded. The order of functions with the same priority
3077 .. _llvmglobaldtors:
3079 The '``llvm.global_dtors``' Global Variable
3080 -------------------------------------------
3082 .. code-block:: llvm
3084 %0 = type { i32, void ()* }
3085 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor }]
3087 The ``@llvm.global_dtors`` array contains a list of destructor functions
3088 and associated priorities. The functions referenced by this array will
3089 be called in descending order of priority (i.e. highest first) when the
3090 module is loaded. The order of functions with the same priority is not
3093 Instruction Reference
3094 =====================
3096 The LLVM instruction set consists of several different classifications
3097 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
3098 instructions <binaryops>`, :ref:`bitwise binary
3099 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
3100 :ref:`other instructions <otherops>`.
3104 Terminator Instructions
3105 -----------------------
3107 As mentioned :ref:`previously <functionstructure>`, every basic block in a
3108 program ends with a "Terminator" instruction, which indicates which
3109 block should be executed after the current block is finished. These
3110 terminator instructions typically yield a '``void``' value: they produce
3111 control flow, not values (the one exception being the
3112 ':ref:`invoke <i_invoke>`' instruction).
3114 The terminator instructions are: ':ref:`ret <i_ret>`',
3115 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
3116 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
3117 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
3121 '``ret``' Instruction
3122 ^^^^^^^^^^^^^^^^^^^^^
3129 ret <type> <value> ; Return a value from a non-void function
3130 ret void ; Return from void function
3135 The '``ret``' instruction is used to return control flow (and optionally
3136 a value) from a function back to the caller.
3138 There are two forms of the '``ret``' instruction: one that returns a
3139 value and then causes control flow, and one that just causes control
3145 The '``ret``' instruction optionally accepts a single argument, the
3146 return value. The type of the return value must be a ':ref:`first
3147 class <t_firstclass>`' type.
3149 A function is not :ref:`well formed <wellformed>` if it it has a non-void
3150 return type and contains a '``ret``' instruction with no return value or
3151 a return value with a type that does not match its type, or if it has a
3152 void return type and contains a '``ret``' instruction with a return
3158 When the '``ret``' instruction is executed, control flow returns back to
3159 the calling function's context. If the caller is a
3160 ":ref:`call <i_call>`" instruction, execution continues at the
3161 instruction after the call. If the caller was an
3162 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3163 beginning of the "normal" destination block. If the instruction returns
3164 a value, that value shall set the call or invoke instruction's return
3170 .. code-block:: llvm
3172 ret i32 5 ; Return an integer value of 5
3173 ret void ; Return from a void function
3174 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3178 '``br``' Instruction
3179 ^^^^^^^^^^^^^^^^^^^^
3186 br i1 <cond>, label <iftrue>, label <iffalse>
3187 br label <dest> ; Unconditional branch
3192 The '``br``' instruction is used to cause control flow to transfer to a
3193 different basic block in the current function. There are two forms of
3194 this instruction, corresponding to a conditional branch and an
3195 unconditional branch.
3200 The conditional branch form of the '``br``' instruction takes a single
3201 '``i1``' value and two '``label``' values. The unconditional form of the
3202 '``br``' instruction takes a single '``label``' value as a target.
3207 Upon execution of a conditional '``br``' instruction, the '``i1``'
3208 argument is evaluated. If the value is ``true``, control flows to the
3209 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3210 to the '``iffalse``' ``label`` argument.
3215 .. code-block:: llvm
3218 %cond = icmp eq i32 %a, %b
3219 br i1 %cond, label %IfEqual, label %IfUnequal
3227 '``switch``' Instruction
3228 ^^^^^^^^^^^^^^^^^^^^^^^^
3235 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3240 The '``switch``' instruction is used to transfer control flow to one of
3241 several different places. It is a generalization of the '``br``'
3242 instruction, allowing a branch to occur to one of many possible
3248 The '``switch``' instruction uses three parameters: an integer
3249 comparison value '``value``', a default '``label``' destination, and an
3250 array of pairs of comparison value constants and '``label``'s. The table
3251 is not allowed to contain duplicate constant entries.
3256 The ``switch`` instruction specifies a table of values and destinations.
3257 When the '``switch``' instruction is executed, this table is searched
3258 for the given value. If the value is found, control flow is transferred
3259 to the corresponding destination; otherwise, control flow is transferred
3260 to the default destination.
3265 Depending on properties of the target machine and the particular
3266 ``switch`` instruction, this instruction may be code generated in
3267 different ways. For example, it could be generated as a series of
3268 chained conditional branches or with a lookup table.
3273 .. code-block:: llvm
3275 ; Emulate a conditional br instruction
3276 %Val = zext i1 %value to i32
3277 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3279 ; Emulate an unconditional br instruction
3280 switch i32 0, label %dest [ ]
3282 ; Implement a jump table:
3283 switch i32 %val, label %otherwise [ i32 0, label %onzero
3285 i32 2, label %ontwo ]
3289 '``indirectbr``' Instruction
3290 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3297 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3302 The '``indirectbr``' instruction implements an indirect branch to a
3303 label within the current function, whose address is specified by
3304 "``address``". Address must be derived from a
3305 :ref:`blockaddress <blockaddress>` constant.
3310 The '``address``' argument is the address of the label to jump to. The
3311 rest of the arguments indicate the full set of possible destinations
3312 that the address may point to. Blocks are allowed to occur multiple
3313 times in the destination list, though this isn't particularly useful.
3315 This destination list is required so that dataflow analysis has an
3316 accurate understanding of the CFG.
3321 Control transfers to the block specified in the address argument. All
3322 possible destination blocks must be listed in the label list, otherwise
3323 this instruction has undefined behavior. This implies that jumps to
3324 labels defined in other functions have undefined behavior as well.
3329 This is typically implemented with a jump through a register.
3334 .. code-block:: llvm
3336 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3340 '``invoke``' Instruction
3341 ^^^^^^^^^^^^^^^^^^^^^^^^
3348 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3349 to label <normal label> unwind label <exception label>
3354 The '``invoke``' instruction causes control to transfer to a specified
3355 function, with the possibility of control flow transfer to either the
3356 '``normal``' label or the '``exception``' label. If the callee function
3357 returns with the "``ret``" instruction, control flow will return to the
3358 "normal" label. If the callee (or any indirect callees) returns via the
3359 ":ref:`resume <i_resume>`" instruction or other exception handling
3360 mechanism, control is interrupted and continued at the dynamically
3361 nearest "exception" label.
3363 The '``exception``' label is a `landing
3364 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3365 '``exception``' label is required to have the
3366 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3367 information about the behavior of the program after unwinding happens,
3368 as its first non-PHI instruction. The restrictions on the
3369 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3370 instruction, so that the important information contained within the
3371 "``landingpad``" instruction can't be lost through normal code motion.
3376 This instruction requires several arguments:
3378 #. The optional "cconv" marker indicates which :ref:`calling
3379 convention <callingconv>` the call should use. If none is
3380 specified, the call defaults to using C calling conventions.
3381 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3382 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3384 #. '``ptr to function ty``': shall be the signature of the pointer to
3385 function value being invoked. In most cases, this is a direct
3386 function invocation, but indirect ``invoke``'s are just as possible,
3387 branching off an arbitrary pointer to function value.
3388 #. '``function ptr val``': An LLVM value containing a pointer to a
3389 function to be invoked.
3390 #. '``function args``': argument list whose types match the function
3391 signature argument types and parameter attributes. All arguments must
3392 be of :ref:`first class <t_firstclass>` type. If the function signature
3393 indicates the function accepts a variable number of arguments, the
3394 extra arguments can be specified.
3395 #. '``normal label``': the label reached when the called function
3396 executes a '``ret``' instruction.
3397 #. '``exception label``': the label reached when a callee returns via
3398 the :ref:`resume <i_resume>` instruction or other exception handling
3400 #. The optional :ref:`function attributes <fnattrs>` list. Only
3401 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3402 attributes are valid here.
3407 This instruction is designed to operate as a standard '``call``'
3408 instruction in most regards. The primary difference is that it
3409 establishes an association with a label, which is used by the runtime
3410 library to unwind the stack.
3412 This instruction is used in languages with destructors to ensure that
3413 proper cleanup is performed in the case of either a ``longjmp`` or a
3414 thrown exception. Additionally, this is important for implementation of
3415 '``catch``' clauses in high-level languages that support them.
3417 For the purposes of the SSA form, the definition of the value returned
3418 by the '``invoke``' instruction is deemed to occur on the edge from the
3419 current block to the "normal" label. If the callee unwinds then no
3420 return value is available.
3425 .. code-block:: llvm
3427 %retval = invoke i32 @Test(i32 15) to label %Continue
3428 unwind label %TestCleanup ; {i32}:retval set
3429 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3430 unwind label %TestCleanup ; {i32}:retval set
3434 '``resume``' Instruction
3435 ^^^^^^^^^^^^^^^^^^^^^^^^
3442 resume <type> <value>
3447 The '``resume``' instruction is a terminator instruction that has no
3453 The '``resume``' instruction requires one argument, which must have the
3454 same type as the result of any '``landingpad``' instruction in the same
3460 The '``resume``' instruction resumes propagation of an existing
3461 (in-flight) exception whose unwinding was interrupted with a
3462 :ref:`landingpad <i_landingpad>` instruction.
3467 .. code-block:: llvm
3469 resume { i8*, i32 } %exn
3473 '``unreachable``' Instruction
3474 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3486 The '``unreachable``' instruction has no defined semantics. This
3487 instruction is used to inform the optimizer that a particular portion of
3488 the code is not reachable. This can be used to indicate that the code
3489 after a no-return function cannot be reached, and other facts.
3494 The '``unreachable``' instruction has no defined semantics.
3501 Binary operators are used to do most of the computation in a program.
3502 They require two operands of the same type, execute an operation on
3503 them, and produce a single value. The operands might represent multiple
3504 data, as is the case with the :ref:`vector <t_vector>` data type. The
3505 result value has the same type as its operands.
3507 There are several different binary operators:
3511 '``add``' Instruction
3512 ^^^^^^^^^^^^^^^^^^^^^
3519 <result> = add <ty> <op1>, <op2> ; yields {ty}:result
3520 <result> = add nuw <ty> <op1>, <op2> ; yields {ty}:result
3521 <result> = add nsw <ty> <op1>, <op2> ; yields {ty}:result
3522 <result> = add nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3527 The '``add``' instruction returns the sum of its two operands.
3532 The two arguments to the '``add``' instruction must be
3533 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3534 arguments must have identical types.
3539 The value produced is the integer sum of the two operands.
3541 If the sum has unsigned overflow, the result returned is the
3542 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3545 Because LLVM integers use a two's complement representation, this
3546 instruction is appropriate for both signed and unsigned integers.
3548 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3549 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3550 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
3551 unsigned and/or signed overflow, respectively, occurs.
3556 .. code-block:: llvm
3558 <result> = add i32 4, %var ; yields {i32}:result = 4 + %var
3562 '``fadd``' Instruction
3563 ^^^^^^^^^^^^^^^^^^^^^^
3570 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3575 The '``fadd``' instruction returns the sum of its two operands.
3580 The two arguments to the '``fadd``' instruction must be :ref:`floating
3581 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3582 Both arguments must have identical types.
3587 The value produced is the floating point sum of the two operands. This
3588 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
3589 which are optimization hints to enable otherwise unsafe floating point
3595 .. code-block:: llvm
3597 <result> = fadd float 4.0, %var ; yields {float}:result = 4.0 + %var
3599 '``sub``' Instruction
3600 ^^^^^^^^^^^^^^^^^^^^^
3607 <result> = sub <ty> <op1>, <op2> ; yields {ty}:result
3608 <result> = sub nuw <ty> <op1>, <op2> ; yields {ty}:result
3609 <result> = sub nsw <ty> <op1>, <op2> ; yields {ty}:result
3610 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3615 The '``sub``' instruction returns the difference of its two operands.
3617 Note that the '``sub``' instruction is used to represent the '``neg``'
3618 instruction present in most other intermediate representations.
3623 The two arguments to the '``sub``' instruction must be
3624 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3625 arguments must have identical types.
3630 The value produced is the integer difference of the two operands.
3632 If the difference has unsigned overflow, the result returned is the
3633 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3636 Because LLVM integers use a two's complement representation, this
3637 instruction is appropriate for both signed and unsigned integers.
3639 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3640 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3641 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
3642 unsigned and/or signed overflow, respectively, occurs.
3647 .. code-block:: llvm
3649 <result> = sub i32 4, %var ; yields {i32}:result = 4 - %var
3650 <result> = sub i32 0, %val ; yields {i32}:result = -%var
3654 '``fsub``' Instruction
3655 ^^^^^^^^^^^^^^^^^^^^^^
3662 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3667 The '``fsub``' instruction returns the difference of its two operands.
3669 Note that the '``fsub``' instruction is used to represent the '``fneg``'
3670 instruction present in most other intermediate representations.
3675 The two arguments to the '``fsub``' instruction must be :ref:`floating
3676 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3677 Both arguments must have identical types.
3682 The value produced is the floating point difference of the two operands.
3683 This instruction can also take any number of :ref:`fast-math
3684 flags <fastmath>`, which are optimization hints to enable otherwise
3685 unsafe floating point optimizations:
3690 .. code-block:: llvm
3692 <result> = fsub float 4.0, %var ; yields {float}:result = 4.0 - %var
3693 <result> = fsub float -0.0, %val ; yields {float}:result = -%var
3695 '``mul``' Instruction
3696 ^^^^^^^^^^^^^^^^^^^^^
3703 <result> = mul <ty> <op1>, <op2> ; yields {ty}:result
3704 <result> = mul nuw <ty> <op1>, <op2> ; yields {ty}:result
3705 <result> = mul nsw <ty> <op1>, <op2> ; yields {ty}:result
3706 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3711 The '``mul``' instruction returns the product of its two operands.
3716 The two arguments to the '``mul``' instruction must be
3717 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3718 arguments must have identical types.
3723 The value produced is the integer product of the two operands.
3725 If the result of the multiplication has unsigned overflow, the result
3726 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
3727 bit width of the result.
3729 Because LLVM integers use a two's complement representation, and the
3730 result is the same width as the operands, this instruction returns the
3731 correct result for both signed and unsigned integers. If a full product
3732 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
3733 sign-extended or zero-extended as appropriate to the width of the full
3736 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3737 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3738 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
3739 unsigned and/or signed overflow, respectively, occurs.
3744 .. code-block:: llvm
3746 <result> = mul i32 4, %var ; yields {i32}:result = 4 * %var
3750 '``fmul``' Instruction
3751 ^^^^^^^^^^^^^^^^^^^^^^
3758 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3763 The '``fmul``' instruction returns the product of its two operands.
3768 The two arguments to the '``fmul``' instruction must be :ref:`floating
3769 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3770 Both arguments must have identical types.
3775 The value produced is the floating point product of the two operands.
3776 This instruction can also take any number of :ref:`fast-math
3777 flags <fastmath>`, which are optimization hints to enable otherwise
3778 unsafe floating point optimizations:
3783 .. code-block:: llvm
3785 <result> = fmul float 4.0, %var ; yields {float}:result = 4.0 * %var
3787 '``udiv``' Instruction
3788 ^^^^^^^^^^^^^^^^^^^^^^
3795 <result> = udiv <ty> <op1>, <op2> ; yields {ty}:result
3796 <result> = udiv exact <ty> <op1>, <op2> ; yields {ty}:result
3801 The '``udiv``' instruction returns the quotient of its two operands.
3806 The two arguments to the '``udiv``' instruction must be
3807 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3808 arguments must have identical types.
3813 The value produced is the unsigned integer quotient of the two operands.
3815 Note that unsigned integer division and signed integer division are
3816 distinct operations; for signed integer division, use '``sdiv``'.
3818 Division by zero leads to undefined behavior.
3820 If the ``exact`` keyword is present, the result value of the ``udiv`` is
3821 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
3822 such, "((a udiv exact b) mul b) == a").
3827 .. code-block:: llvm
3829 <result> = udiv i32 4, %var ; yields {i32}:result = 4 / %var
3831 '``sdiv``' Instruction
3832 ^^^^^^^^^^^^^^^^^^^^^^
3839 <result> = sdiv <ty> <op1>, <op2> ; yields {ty}:result
3840 <result> = sdiv exact <ty> <op1>, <op2> ; yields {ty}:result
3845 The '``sdiv``' instruction returns the quotient of its two operands.
3850 The two arguments to the '``sdiv``' instruction must be
3851 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3852 arguments must have identical types.
3857 The value produced is the signed integer quotient of the two operands
3858 rounded towards zero.
3860 Note that signed integer division and unsigned integer division are
3861 distinct operations; for unsigned integer division, use '``udiv``'.
3863 Division by zero leads to undefined behavior. Overflow also leads to
3864 undefined behavior; this is a rare case, but can occur, for example, by
3865 doing a 32-bit division of -2147483648 by -1.
3867 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
3868 a :ref:`poison value <poisonvalues>` if the result would be rounded.
3873 .. code-block:: llvm
3875 <result> = sdiv i32 4, %var ; yields {i32}:result = 4 / %var
3879 '``fdiv``' Instruction
3880 ^^^^^^^^^^^^^^^^^^^^^^
3887 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3892 The '``fdiv``' instruction returns the quotient of its two operands.
3897 The two arguments to the '``fdiv``' instruction must be :ref:`floating
3898 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3899 Both arguments must have identical types.
3904 The value produced is the floating point quotient of the two operands.
3905 This instruction can also take any number of :ref:`fast-math
3906 flags <fastmath>`, which are optimization hints to enable otherwise
3907 unsafe floating point optimizations:
3912 .. code-block:: llvm
3914 <result> = fdiv float 4.0, %var ; yields {float}:result = 4.0 / %var
3916 '``urem``' Instruction
3917 ^^^^^^^^^^^^^^^^^^^^^^
3924 <result> = urem <ty> <op1>, <op2> ; yields {ty}:result
3929 The '``urem``' instruction returns the remainder from the unsigned
3930 division of its two arguments.
3935 The two arguments to the '``urem``' instruction must be
3936 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3937 arguments must have identical types.
3942 This instruction returns the unsigned integer *remainder* of a division.
3943 This instruction always performs an unsigned division to get the
3946 Note that unsigned integer remainder and signed integer remainder are
3947 distinct operations; for signed integer remainder, use '``srem``'.
3949 Taking the remainder of a division by zero leads to undefined behavior.
3954 .. code-block:: llvm
3956 <result> = urem i32 4, %var ; yields {i32}:result = 4 % %var
3958 '``srem``' Instruction
3959 ^^^^^^^^^^^^^^^^^^^^^^
3966 <result> = srem <ty> <op1>, <op2> ; yields {ty}:result
3971 The '``srem``' instruction returns the remainder from the signed
3972 division of its two operands. This instruction can also take
3973 :ref:`vector <t_vector>` versions of the values in which case the elements
3979 The two arguments to the '``srem``' instruction must be
3980 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3981 arguments must have identical types.
3986 This instruction returns the *remainder* of a division (where the result
3987 is either zero or has the same sign as the dividend, ``op1``), not the
3988 *modulo* operator (where the result is either zero or has the same sign
3989 as the divisor, ``op2``) of a value. For more information about the
3990 difference, see `The Math
3991 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
3992 table of how this is implemented in various languages, please see
3994 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
3996 Note that signed integer remainder and unsigned integer remainder are
3997 distinct operations; for unsigned integer remainder, use '``urem``'.
3999 Taking the remainder of a division by zero leads to undefined behavior.
4000 Overflow also leads to undefined behavior; this is a rare case, but can
4001 occur, for example, by taking the remainder of a 32-bit division of
4002 -2147483648 by -1. (The remainder doesn't actually overflow, but this
4003 rule lets srem be implemented using instructions that return both the
4004 result of the division and the remainder.)
4009 .. code-block:: llvm
4011 <result> = srem i32 4, %var ; yields {i32}:result = 4 % %var
4015 '``frem``' Instruction
4016 ^^^^^^^^^^^^^^^^^^^^^^
4023 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
4028 The '``frem``' instruction returns the remainder from the division of
4034 The two arguments to the '``frem``' instruction must be :ref:`floating
4035 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4036 Both arguments must have identical types.
4041 This instruction returns the *remainder* of a division. The remainder
4042 has the same sign as the dividend. This instruction can also take any
4043 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
4044 to enable otherwise unsafe floating point optimizations:
4049 .. code-block:: llvm
4051 <result> = frem float 4.0, %var ; yields {float}:result = 4.0 % %var
4055 Bitwise Binary Operations
4056 -------------------------
4058 Bitwise binary operators are used to do various forms of bit-twiddling
4059 in a program. They are generally very efficient instructions and can
4060 commonly be strength reduced from other instructions. They require two
4061 operands of the same type, execute an operation on them, and produce a
4062 single value. The resulting value is the same type as its operands.
4064 '``shl``' Instruction
4065 ^^^^^^^^^^^^^^^^^^^^^
4072 <result> = shl <ty> <op1>, <op2> ; yields {ty}:result
4073 <result> = shl nuw <ty> <op1>, <op2> ; yields {ty}:result
4074 <result> = shl nsw <ty> <op1>, <op2> ; yields {ty}:result
4075 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
4080 The '``shl``' instruction returns the first operand shifted to the left
4081 a specified number of bits.
4086 Both arguments to the '``shl``' instruction must be the same
4087 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4088 '``op2``' is treated as an unsigned value.
4093 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
4094 where ``n`` is the width of the result. If ``op2`` is (statically or
4095 dynamically) negative or equal to or larger than the number of bits in
4096 ``op1``, the result is undefined. If the arguments are vectors, each
4097 vector element of ``op1`` is shifted by the corresponding shift amount
4100 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
4101 value <poisonvalues>` if it shifts out any non-zero bits. If the
4102 ``nsw`` keyword is present, then the shift produces a :ref:`poison
4103 value <poisonvalues>` if it shifts out any bits that disagree with the
4104 resultant sign bit. As such, NUW/NSW have the same semantics as they
4105 would if the shift were expressed as a mul instruction with the same
4106 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
4111 .. code-block:: llvm
4113 <result> = shl i32 4, %var ; yields {i32}: 4 << %var
4114 <result> = shl i32 4, 2 ; yields {i32}: 16
4115 <result> = shl i32 1, 10 ; yields {i32}: 1024
4116 <result> = shl i32 1, 32 ; undefined
4117 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
4119 '``lshr``' Instruction
4120 ^^^^^^^^^^^^^^^^^^^^^^
4127 <result> = lshr <ty> <op1>, <op2> ; yields {ty}:result
4128 <result> = lshr exact <ty> <op1>, <op2> ; yields {ty}:result
4133 The '``lshr``' instruction (logical shift right) returns the first
4134 operand shifted to the right a specified number of bits with zero fill.
4139 Both arguments to the '``lshr``' instruction must be the same
4140 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4141 '``op2``' is treated as an unsigned value.
4146 This instruction always performs a logical shift right operation. The
4147 most significant bits of the result will be filled with zero bits after
4148 the shift. If ``op2`` is (statically or dynamically) equal to or larger
4149 than the number of bits in ``op1``, the result is undefined. If the
4150 arguments are vectors, each vector element of ``op1`` is shifted by the
4151 corresponding shift amount in ``op2``.
4153 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4154 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4160 .. code-block:: llvm
4162 <result> = lshr i32 4, 1 ; yields {i32}:result = 2
4163 <result> = lshr i32 4, 2 ; yields {i32}:result = 1
4164 <result> = lshr i8 4, 3 ; yields {i8}:result = 0
4165 <result> = lshr i8 -2, 1 ; yields {i8}:result = 0x7F
4166 <result> = lshr i32 1, 32 ; undefined
4167 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4169 '``ashr``' Instruction
4170 ^^^^^^^^^^^^^^^^^^^^^^
4177 <result> = ashr <ty> <op1>, <op2> ; yields {ty}:result
4178 <result> = ashr exact <ty> <op1>, <op2> ; yields {ty}:result
4183 The '``ashr``' instruction (arithmetic shift right) returns the first
4184 operand shifted to the right a specified number of bits with sign
4190 Both arguments to the '``ashr``' instruction must be the same
4191 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4192 '``op2``' is treated as an unsigned value.
4197 This instruction always performs an arithmetic shift right operation,
4198 The most significant bits of the result will be filled with the sign bit
4199 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4200 than the number of bits in ``op1``, the result is undefined. If the
4201 arguments are vectors, each vector element of ``op1`` is shifted by the
4202 corresponding shift amount in ``op2``.
4204 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4205 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4211 .. code-block:: llvm
4213 <result> = ashr i32 4, 1 ; yields {i32}:result = 2
4214 <result> = ashr i32 4, 2 ; yields {i32}:result = 1
4215 <result> = ashr i8 4, 3 ; yields {i8}:result = 0
4216 <result> = ashr i8 -2, 1 ; yields {i8}:result = -1
4217 <result> = ashr i32 1, 32 ; undefined
4218 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4220 '``and``' Instruction
4221 ^^^^^^^^^^^^^^^^^^^^^
4228 <result> = and <ty> <op1>, <op2> ; yields {ty}:result
4233 The '``and``' instruction returns the bitwise logical and of its two
4239 The two arguments to the '``and``' instruction must be
4240 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4241 arguments must have identical types.
4246 The truth table used for the '``and``' instruction is:
4263 .. code-block:: llvm
4265 <result> = and i32 4, %var ; yields {i32}:result = 4 & %var
4266 <result> = and i32 15, 40 ; yields {i32}:result = 8
4267 <result> = and i32 4, 8 ; yields {i32}:result = 0
4269 '``or``' Instruction
4270 ^^^^^^^^^^^^^^^^^^^^
4277 <result> = or <ty> <op1>, <op2> ; yields {ty}:result
4282 The '``or``' instruction returns the bitwise logical inclusive or of its
4288 The two arguments to the '``or``' instruction must be
4289 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4290 arguments must have identical types.
4295 The truth table used for the '``or``' instruction is:
4314 <result> = or i32 4, %var ; yields {i32}:result = 4 | %var
4315 <result> = or i32 15, 40 ; yields {i32}:result = 47
4316 <result> = or i32 4, 8 ; yields {i32}:result = 12
4318 '``xor``' Instruction
4319 ^^^^^^^^^^^^^^^^^^^^^
4326 <result> = xor <ty> <op1>, <op2> ; yields {ty}:result
4331 The '``xor``' instruction returns the bitwise logical exclusive or of
4332 its two operands. The ``xor`` is used to implement the "one's
4333 complement" operation, which is the "~" operator in C.
4338 The two arguments to the '``xor``' instruction must be
4339 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4340 arguments must have identical types.
4345 The truth table used for the '``xor``' instruction is:
4362 .. code-block:: llvm
4364 <result> = xor i32 4, %var ; yields {i32}:result = 4 ^ %var
4365 <result> = xor i32 15, 40 ; yields {i32}:result = 39
4366 <result> = xor i32 4, 8 ; yields {i32}:result = 12
4367 <result> = xor i32 %V, -1 ; yields {i32}:result = ~%V
4372 LLVM supports several instructions to represent vector operations in a
4373 target-independent manner. These instructions cover the element-access
4374 and vector-specific operations needed to process vectors effectively.
4375 While LLVM does directly support these vector operations, many
4376 sophisticated algorithms will want to use target-specific intrinsics to
4377 take full advantage of a specific target.
4379 .. _i_extractelement:
4381 '``extractelement``' Instruction
4382 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4389 <result> = extractelement <n x <ty>> <val>, i32 <idx> ; yields <ty>
4394 The '``extractelement``' instruction extracts a single scalar element
4395 from a vector at a specified index.
4400 The first operand of an '``extractelement``' instruction is a value of
4401 :ref:`vector <t_vector>` type. The second operand is an index indicating
4402 the position from which to extract the element. The index may be a
4408 The result is a scalar of the same type as the element type of ``val``.
4409 Its value is the value at position ``idx`` of ``val``. If ``idx``
4410 exceeds the length of ``val``, the results are undefined.
4415 .. code-block:: llvm
4417 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4419 .. _i_insertelement:
4421 '``insertelement``' Instruction
4422 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4429 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> ; yields <n x <ty>>
4434 The '``insertelement``' instruction inserts a scalar element into a
4435 vector at a specified index.
4440 The first operand of an '``insertelement``' instruction is a value of
4441 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4442 type must equal the element type of the first operand. The third operand
4443 is an index indicating the position at which to insert the value. The
4444 index may be a variable.
4449 The result is a vector of the same type as ``val``. Its element values
4450 are those of ``val`` except at position ``idx``, where it gets the value
4451 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4457 .. code-block:: llvm
4459 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4461 .. _i_shufflevector:
4463 '``shufflevector``' Instruction
4464 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4471 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4476 The '``shufflevector``' instruction constructs a permutation of elements
4477 from two input vectors, returning a vector with the same element type as
4478 the input and length that is the same as the shuffle mask.
4483 The first two operands of a '``shufflevector``' instruction are vectors
4484 with the same type. The third argument is a shuffle mask whose element
4485 type is always 'i32'. The result of the instruction is a vector whose
4486 length is the same as the shuffle mask and whose element type is the
4487 same as the element type of the first two operands.
4489 The shuffle mask operand is required to be a constant vector with either
4490 constant integer or undef values.
4495 The elements of the two input vectors are numbered from left to right
4496 across both of the vectors. The shuffle mask operand specifies, for each
4497 element of the result vector, which element of the two input vectors the
4498 result element gets. The element selector may be undef (meaning "don't
4499 care") and the second operand may be undef if performing a shuffle from
4505 .. code-block:: llvm
4507 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4508 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4509 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4510 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4511 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4512 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4513 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4514 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4516 Aggregate Operations
4517 --------------------
4519 LLVM supports several instructions for working with
4520 :ref:`aggregate <t_aggregate>` values.
4524 '``extractvalue``' Instruction
4525 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4532 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4537 The '``extractvalue``' instruction extracts the value of a member field
4538 from an :ref:`aggregate <t_aggregate>` value.
4543 The first operand of an '``extractvalue``' instruction is a value of
4544 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4545 constant indices to specify which value to extract in a similar manner
4546 as indices in a '``getelementptr``' instruction.
4548 The major differences to ``getelementptr`` indexing are:
4550 - Since the value being indexed is not a pointer, the first index is
4551 omitted and assumed to be zero.
4552 - At least one index must be specified.
4553 - Not only struct indices but also array indices must be in bounds.
4558 The result is the value at the position in the aggregate specified by
4564 .. code-block:: llvm
4566 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
4570 '``insertvalue``' Instruction
4571 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4578 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
4583 The '``insertvalue``' instruction inserts a value into a member field in
4584 an :ref:`aggregate <t_aggregate>` value.
4589 The first operand of an '``insertvalue``' instruction is a value of
4590 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
4591 a first-class value to insert. The following operands are constant
4592 indices indicating the position at which to insert the value in a
4593 similar manner as indices in a '``extractvalue``' instruction. The value
4594 to insert must have the same type as the value identified by the
4600 The result is an aggregate of the same type as ``val``. Its value is
4601 that of ``val`` except that the value at the position specified by the
4602 indices is that of ``elt``.
4607 .. code-block:: llvm
4609 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
4610 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
4611 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
4615 Memory Access and Addressing Operations
4616 ---------------------------------------
4618 A key design point of an SSA-based representation is how it represents
4619 memory. In LLVM, no memory locations are in SSA form, which makes things
4620 very simple. This section describes how to read, write, and allocate
4625 '``alloca``' Instruction
4626 ^^^^^^^^^^^^^^^^^^^^^^^^
4633 <result> = alloca <type>[, <ty> <NumElements>][, align <alignment>] ; yields {type*}:result
4638 The '``alloca``' instruction allocates memory on the stack frame of the
4639 currently executing function, to be automatically released when this
4640 function returns to its caller. The object is always allocated in the
4641 generic address space (address space zero).
4646 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
4647 bytes of memory on the runtime stack, returning a pointer of the
4648 appropriate type to the program. If "NumElements" is specified, it is
4649 the number of elements allocated, otherwise "NumElements" is defaulted
4650 to be one. If a constant alignment is specified, the value result of the
4651 allocation is guaranteed to be aligned to at least that boundary. If not
4652 specified, or if zero, the target can choose to align the allocation on
4653 any convenient boundary compatible with the type.
4655 '``type``' may be any sized type.
4660 Memory is allocated; a pointer is returned. The operation is undefined
4661 if there is insufficient stack space for the allocation. '``alloca``'d
4662 memory is automatically released when the function returns. The
4663 '``alloca``' instruction is commonly used to represent automatic
4664 variables that must have an address available. When the function returns
4665 (either with the ``ret`` or ``resume`` instructions), the memory is
4666 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4667 The order in which memory is allocated (ie., which way the stack grows)
4673 .. code-block:: llvm
4675 %ptr = alloca i32 ; yields {i32*}:ptr
4676 %ptr = alloca i32, i32 4 ; yields {i32*}:ptr
4677 %ptr = alloca i32, i32 4, align 1024 ; yields {i32*}:ptr
4678 %ptr = alloca i32, align 1024 ; yields {i32*}:ptr
4682 '``load``' Instruction
4683 ^^^^^^^^^^^^^^^^^^^^^^
4690 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4691 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4692 !<index> = !{ i32 1 }
4697 The '``load``' instruction is used to read from memory.
4702 The argument to the ``load`` instruction specifies the memory address
4703 from which to load. The pointer must point to a :ref:`first
4704 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
4705 then the optimizer is not allowed to modify the number or order of
4706 execution of this ``load`` with other :ref:`volatile
4707 operations <volatile>`.
4709 If the ``load`` is marked as ``atomic``, it takes an extra
4710 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4711 ``release`` and ``acq_rel`` orderings are not valid on ``load``
4712 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4713 when they may see multiple atomic stores. The type of the pointee must
4714 be an integer type whose bit width is a power of two greater than or
4715 equal to eight and less than or equal to a target-specific size limit.
4716 ``align`` must be explicitly specified on atomic loads, and the load has
4717 undefined behavior if the alignment is not set to a value which is at
4718 least the size in bytes of the pointee. ``!nontemporal`` does not have
4719 any defined semantics for atomic loads.
4721 The optional constant ``align`` argument specifies the alignment of the
4722 operation (that is, the alignment of the memory address). A value of 0
4723 or an omitted ``align`` argument means that the operation has the ABI
4724 alignment for the target. It is the responsibility of the code emitter
4725 to ensure that the alignment information is correct. Overestimating the
4726 alignment results in undefined behavior. Underestimating the alignment
4727 may produce less efficient code. An alignment of 1 is always safe.
4729 The optional ``!nontemporal`` metadata must reference a single
4730 metadata name ``<index>`` corresponding to a metadata node with one
4731 ``i32`` entry of value 1. The existence of the ``!nontemporal``
4732 metadata on the instruction tells the optimizer and code generator
4733 that this load is not expected to be reused in the cache. The code
4734 generator may select special instructions to save cache bandwidth, such
4735 as the ``MOVNT`` instruction on x86.
4737 The optional ``!invariant.load`` metadata must reference a single
4738 metadata name ``<index>`` corresponding to a metadata node with no
4739 entries. The existence of the ``!invariant.load`` metadata on the
4740 instruction tells the optimizer and code generator that this load
4741 address points to memory which does not change value during program
4742 execution. The optimizer may then move this load around, for example, by
4743 hoisting it out of loops using loop invariant code motion.
4748 The location of memory pointed to is loaded. If the value being loaded
4749 is of scalar type then the number of bytes read does not exceed the
4750 minimum number of bytes needed to hold all bits of the type. For
4751 example, loading an ``i24`` reads at most three bytes. When loading a
4752 value of a type like ``i20`` with a size that is not an integral number
4753 of bytes, the result is undefined if the value was not originally
4754 written using a store of the same type.
4759 .. code-block:: llvm
4761 %ptr = alloca i32 ; yields {i32*}:ptr
4762 store i32 3, i32* %ptr ; yields {void}
4763 %val = load i32* %ptr ; yields {i32}:val = i32 3
4767 '``store``' Instruction
4768 ^^^^^^^^^^^^^^^^^^^^^^^
4775 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields {void}
4776 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields {void}
4781 The '``store``' instruction is used to write to memory.
4786 There are two arguments to the ``store`` instruction: a value to store
4787 and an address at which to store it. The type of the ``<pointer>``
4788 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
4789 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
4790 then the optimizer is not allowed to modify the number or order of
4791 execution of this ``store`` with other :ref:`volatile
4792 operations <volatile>`.
4794 If the ``store`` is marked as ``atomic``, it takes an extra
4795 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4796 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
4797 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4798 when they may see multiple atomic stores. The type of the pointee must
4799 be an integer type whose bit width is a power of two greater than or
4800 equal to eight and less than or equal to a target-specific size limit.
4801 ``align`` must be explicitly specified on atomic stores, and the store
4802 has undefined behavior if the alignment is not set to a value which is
4803 at least the size in bytes of the pointee. ``!nontemporal`` does not
4804 have any defined semantics for atomic stores.
4806 The optional constant ``align`` argument specifies the alignment of the
4807 operation (that is, the alignment of the memory address). A value of 0
4808 or an omitted ``align`` argument means that the operation has the ABI
4809 alignment for the target. It is the responsibility of the code emitter
4810 to ensure that the alignment information is correct. Overestimating the
4811 alignment results in undefined behavior. Underestimating the
4812 alignment may produce less efficient code. An alignment of 1 is always
4815 The optional ``!nontemporal`` metadata must reference a single metadata
4816 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
4817 value 1. The existence of the ``!nontemporal`` metadata on the instruction
4818 tells the optimizer and code generator that this load is not expected to
4819 be reused in the cache. The code generator may select special
4820 instructions to save cache bandwidth, such as the MOVNT instruction on
4826 The contents of memory are updated to contain ``<value>`` at the
4827 location specified by the ``<pointer>`` operand. If ``<value>`` is
4828 of scalar type then the number of bytes written does not exceed the
4829 minimum number of bytes needed to hold all bits of the type. For
4830 example, storing an ``i24`` writes at most three bytes. When writing a
4831 value of a type like ``i20`` with a size that is not an integral number
4832 of bytes, it is unspecified what happens to the extra bits that do not
4833 belong to the type, but they will typically be overwritten.
4838 .. code-block:: llvm
4840 %ptr = alloca i32 ; yields {i32*}:ptr
4841 store i32 3, i32* %ptr ; yields {void}
4842 %val = load i32* %ptr ; yields {i32}:val = i32 3
4846 '``fence``' Instruction
4847 ^^^^^^^^^^^^^^^^^^^^^^^
4854 fence [singlethread] <ordering> ; yields {void}
4859 The '``fence``' instruction is used to introduce happens-before edges
4865 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
4866 defines what *synchronizes-with* edges they add. They can only be given
4867 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
4872 A fence A which has (at least) ``release`` ordering semantics
4873 *synchronizes with* a fence B with (at least) ``acquire`` ordering
4874 semantics if and only if there exist atomic operations X and Y, both
4875 operating on some atomic object M, such that A is sequenced before X, X
4876 modifies M (either directly or through some side effect of a sequence
4877 headed by X), Y is sequenced before B, and Y observes M. This provides a
4878 *happens-before* dependency between A and B. Rather than an explicit
4879 ``fence``, one (but not both) of the atomic operations X or Y might
4880 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
4881 still *synchronize-with* the explicit ``fence`` and establish the
4882 *happens-before* edge.
4884 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
4885 ``acquire`` and ``release`` semantics specified above, participates in
4886 the global program order of other ``seq_cst`` operations and/or fences.
4888 The optional ":ref:`singlethread <singlethread>`" argument specifies
4889 that the fence only synchronizes with other fences in the same thread.
4890 (This is useful for interacting with signal handlers.)
4895 .. code-block:: llvm
4897 fence acquire ; yields {void}
4898 fence singlethread seq_cst ; yields {void}
4902 '``cmpxchg``' Instruction
4903 ^^^^^^^^^^^^^^^^^^^^^^^^^
4910 cmpxchg [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <ordering> ; yields {ty}
4915 The '``cmpxchg``' instruction is used to atomically modify memory. It
4916 loads a value in memory and compares it to a given value. If they are
4917 equal, it stores a new value into the memory.
4922 There are three arguments to the '``cmpxchg``' instruction: an address
4923 to operate on, a value to compare to the value currently be at that
4924 address, and a new value to place at that address if the compared values
4925 are equal. The type of '<cmp>' must be an integer type whose bit width
4926 is a power of two greater than or equal to eight and less than or equal
4927 to a target-specific size limit. '<cmp>' and '<new>' must have the same
4928 type, and the type of '<pointer>' must be a pointer to that type. If the
4929 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
4930 to modify the number or order of execution of this ``cmpxchg`` with
4931 other :ref:`volatile operations <volatile>`.
4933 The :ref:`ordering <ordering>` argument specifies how this ``cmpxchg``
4934 synchronizes with other atomic operations.
4936 The optional "``singlethread``" argument declares that the ``cmpxchg``
4937 is only atomic with respect to code (usually signal handlers) running in
4938 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
4939 respect to all other code in the system.
4941 The pointer passed into cmpxchg must have alignment greater than or
4942 equal to the size in memory of the operand.
4947 The contents of memory at the location specified by the '``<pointer>``'
4948 operand is read and compared to '``<cmp>``'; if the read value is the
4949 equal, '``<new>``' is written. The original value at the location is
4952 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose
4953 of identifying release sequences. A failed ``cmpxchg`` is equivalent to an
4954 atomic load with an ordering parameter determined by dropping any
4955 ``release`` part of the ``cmpxchg``'s ordering.
4960 .. code-block:: llvm
4963 %orig = atomic load i32* %ptr unordered ; yields {i32}
4967 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
4968 %squared = mul i32 %cmp, %cmp
4969 %old = cmpxchg i32* %ptr, i32 %cmp, i32 %squared ; yields {i32}
4970 %success = icmp eq i32 %cmp, %old
4971 br i1 %success, label %done, label %loop
4978 '``atomicrmw``' Instruction
4979 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
4986 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields {ty}
4991 The '``atomicrmw``' instruction is used to atomically modify memory.
4996 There are three arguments to the '``atomicrmw``' instruction: an
4997 operation to apply, an address whose value to modify, an argument to the
4998 operation. The operation must be one of the following keywords:
5012 The type of '<value>' must be an integer type whose bit width is a power
5013 of two greater than or equal to eight and less than or equal to a
5014 target-specific size limit. The type of the '``<pointer>``' operand must
5015 be a pointer to that type. If the ``atomicrmw`` is marked as
5016 ``volatile``, then the optimizer is not allowed to modify the number or
5017 order of execution of this ``atomicrmw`` with other :ref:`volatile
5018 operations <volatile>`.
5023 The contents of memory at the location specified by the '``<pointer>``'
5024 operand are atomically read, modified, and written back. The original
5025 value at the location is returned. The modification is specified by the
5028 - xchg: ``*ptr = val``
5029 - add: ``*ptr = *ptr + val``
5030 - sub: ``*ptr = *ptr - val``
5031 - and: ``*ptr = *ptr & val``
5032 - nand: ``*ptr = ~(*ptr & val)``
5033 - or: ``*ptr = *ptr | val``
5034 - xor: ``*ptr = *ptr ^ val``
5035 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
5036 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
5037 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
5039 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
5045 .. code-block:: llvm
5047 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields {i32}
5049 .. _i_getelementptr:
5051 '``getelementptr``' Instruction
5052 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5059 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
5060 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
5061 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
5066 The '``getelementptr``' instruction is used to get the address of a
5067 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
5068 address calculation only and does not access memory.
5073 The first argument is always a pointer or a vector of pointers, and
5074 forms the basis of the calculation. The remaining arguments are indices
5075 that indicate which of the elements of the aggregate object are indexed.
5076 The interpretation of each index is dependent on the type being indexed
5077 into. The first index always indexes the pointer value given as the
5078 first argument, the second index indexes a value of the type pointed to
5079 (not necessarily the value directly pointed to, since the first index
5080 can be non-zero), etc. The first type indexed into must be a pointer
5081 value, subsequent types can be arrays, vectors, and structs. Note that
5082 subsequent types being indexed into can never be pointers, since that
5083 would require loading the pointer before continuing calculation.
5085 The type of each index argument depends on the type it is indexing into.
5086 When indexing into a (optionally packed) structure, only ``i32`` integer
5087 **constants** are allowed (when using a vector of indices they must all
5088 be the **same** ``i32`` integer constant). When indexing into an array,
5089 pointer or vector, integers of any width are allowed, and they are not
5090 required to be constant. These integers are treated as signed values
5093 For example, let's consider a C code fragment and how it gets compiled
5109 int *foo(struct ST *s) {
5110 return &s[1].Z.B[5][13];
5113 The LLVM code generated by Clang is:
5115 .. code-block:: llvm
5117 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
5118 %struct.ST = type { i32, double, %struct.RT }
5120 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
5122 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
5129 In the example above, the first index is indexing into the
5130 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
5131 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
5132 indexes into the third element of the structure, yielding a
5133 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
5134 structure. The third index indexes into the second element of the
5135 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
5136 dimensions of the array are subscripted into, yielding an '``i32``'
5137 type. The '``getelementptr``' instruction returns a pointer to this
5138 element, thus computing a value of '``i32*``' type.
5140 Note that it is perfectly legal to index partially through a structure,
5141 returning a pointer to an inner element. Because of this, the LLVM code
5142 for the given testcase is equivalent to:
5144 .. code-block:: llvm
5146 define i32* @foo(%struct.ST* %s) {
5147 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
5148 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
5149 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
5150 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
5151 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5155 If the ``inbounds`` keyword is present, the result value of the
5156 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5157 pointer is not an *in bounds* address of an allocated object, or if any
5158 of the addresses that would be formed by successive addition of the
5159 offsets implied by the indices to the base address with infinitely
5160 precise signed arithmetic are not an *in bounds* address of that
5161 allocated object. The *in bounds* addresses for an allocated object are
5162 all the addresses that point into the object, plus the address one byte
5163 past the end. In cases where the base is a vector of pointers the
5164 ``inbounds`` keyword applies to each of the computations element-wise.
5166 If the ``inbounds`` keyword is not present, the offsets are added to the
5167 base address with silently-wrapping two's complement arithmetic. If the
5168 offsets have a different width from the pointer, they are sign-extended
5169 or truncated to the width of the pointer. The result value of the
5170 ``getelementptr`` may be outside the object pointed to by the base
5171 pointer. The result value may not necessarily be used to access memory
5172 though, even if it happens to point into allocated storage. See the
5173 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5176 The getelementptr instruction is often confusing. For some more insight
5177 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5182 .. code-block:: llvm
5184 ; yields [12 x i8]*:aptr
5185 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5187 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5189 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5191 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5193 In cases where the pointer argument is a vector of pointers, each index
5194 must be a vector with the same number of elements. For example:
5196 .. code-block:: llvm
5198 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5200 Conversion Operations
5201 ---------------------
5203 The instructions in this category are the conversion instructions
5204 (casting) which all take a single operand and a type. They perform
5205 various bit conversions on the operand.
5207 '``trunc .. to``' Instruction
5208 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5215 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5220 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5225 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5226 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5227 of the same number of integers. The bit size of the ``value`` must be
5228 larger than the bit size of the destination type, ``ty2``. Equal sized
5229 types are not allowed.
5234 The '``trunc``' instruction truncates the high order bits in ``value``
5235 and converts the remaining bits to ``ty2``. Since the source size must
5236 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5237 It will always truncate bits.
5242 .. code-block:: llvm
5244 %X = trunc i32 257 to i8 ; yields i8:1
5245 %Y = trunc i32 123 to i1 ; yields i1:true
5246 %Z = trunc i32 122 to i1 ; yields i1:false
5247 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5249 '``zext .. to``' Instruction
5250 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5257 <result> = zext <ty> <value> to <ty2> ; yields ty2
5262 The '``zext``' instruction zero extends its operand to type ``ty2``.
5267 The '``zext``' instruction takes a value to cast, and a type to cast it
5268 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5269 the same number of integers. The bit size of the ``value`` must be
5270 smaller than the bit size of the destination type, ``ty2``.
5275 The ``zext`` fills the high order bits of the ``value`` with zero bits
5276 until it reaches the size of the destination type, ``ty2``.
5278 When zero extending from i1, the result will always be either 0 or 1.
5283 .. code-block:: llvm
5285 %X = zext i32 257 to i64 ; yields i64:257
5286 %Y = zext i1 true to i32 ; yields i32:1
5287 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5289 '``sext .. to``' Instruction
5290 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5297 <result> = sext <ty> <value> to <ty2> ; yields ty2
5302 The '``sext``' sign extends ``value`` to the type ``ty2``.
5307 The '``sext``' instruction takes a value to cast, and a type to cast it
5308 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5309 the same number of integers. The bit size of the ``value`` must be
5310 smaller than the bit size of the destination type, ``ty2``.
5315 The '``sext``' instruction performs a sign extension by copying the sign
5316 bit (highest order bit) of the ``value`` until it reaches the bit size
5317 of the type ``ty2``.
5319 When sign extending from i1, the extension always results in -1 or 0.
5324 .. code-block:: llvm
5326 %X = sext i8 -1 to i16 ; yields i16 :65535
5327 %Y = sext i1 true to i32 ; yields i32:-1
5328 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5330 '``fptrunc .. to``' Instruction
5331 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5338 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5343 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5348 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5349 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5350 The size of ``value`` must be larger than the size of ``ty2``. This
5351 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5356 The '``fptrunc``' instruction truncates a ``value`` from a larger
5357 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5358 point <t_floating>` type. If the value cannot fit within the
5359 destination type, ``ty2``, then the results are undefined.
5364 .. code-block:: llvm
5366 %X = fptrunc double 123.0 to float ; yields float:123.0
5367 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5369 '``fpext .. to``' Instruction
5370 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5377 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5382 The '``fpext``' extends a floating point ``value`` to a larger floating
5388 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5389 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5390 to. The source type must be smaller than the destination type.
5395 The '``fpext``' instruction extends the ``value`` from a smaller
5396 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5397 point <t_floating>` type. The ``fpext`` cannot be used to make a
5398 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5399 *no-op cast* for a floating point cast.
5404 .. code-block:: llvm
5406 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5407 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5409 '``fptoui .. to``' Instruction
5410 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5417 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5422 The '``fptoui``' converts a floating point ``value`` to its unsigned
5423 integer equivalent of type ``ty2``.
5428 The '``fptoui``' instruction takes a value to cast, which must be a
5429 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5430 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5431 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5432 type with the same number of elements as ``ty``
5437 The '``fptoui``' instruction converts its :ref:`floating
5438 point <t_floating>` operand into the nearest (rounding towards zero)
5439 unsigned integer value. If the value cannot fit in ``ty2``, the results
5445 .. code-block:: llvm
5447 %X = fptoui double 123.0 to i32 ; yields i32:123
5448 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5449 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5451 '``fptosi .. to``' Instruction
5452 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5459 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5464 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5465 ``value`` to type ``ty2``.
5470 The '``fptosi``' instruction takes a value to cast, which must be a
5471 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5472 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5473 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5474 type with the same number of elements as ``ty``
5479 The '``fptosi``' instruction converts its :ref:`floating
5480 point <t_floating>` operand into the nearest (rounding towards zero)
5481 signed integer value. If the value cannot fit in ``ty2``, the results
5487 .. code-block:: llvm
5489 %X = fptosi double -123.0 to i32 ; yields i32:-123
5490 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5491 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5493 '``uitofp .. to``' Instruction
5494 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5501 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5506 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5507 and converts that value to the ``ty2`` type.
5512 The '``uitofp``' instruction takes a value to cast, which must be a
5513 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5514 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5515 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5516 type with the same number of elements as ``ty``
5521 The '``uitofp``' instruction interprets its operand as an unsigned
5522 integer quantity and converts it to the corresponding floating point
5523 value. If the value cannot fit in the floating point value, the results
5529 .. code-block:: llvm
5531 %X = uitofp i32 257 to float ; yields float:257.0
5532 %Y = uitofp i8 -1 to double ; yields double:255.0
5534 '``sitofp .. to``' Instruction
5535 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5542 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
5547 The '``sitofp``' instruction regards ``value`` as a signed integer and
5548 converts that value to the ``ty2`` type.
5553 The '``sitofp``' instruction takes a value to cast, which must be a
5554 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5555 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5556 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5557 type with the same number of elements as ``ty``
5562 The '``sitofp``' instruction interprets its operand as a signed integer
5563 quantity and converts it to the corresponding floating point value. If
5564 the value cannot fit in the floating point value, the results are
5570 .. code-block:: llvm
5572 %X = sitofp i32 257 to float ; yields float:257.0
5573 %Y = sitofp i8 -1 to double ; yields double:-1.0
5577 '``ptrtoint .. to``' Instruction
5578 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5585 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
5590 The '``ptrtoint``' instruction converts the pointer or a vector of
5591 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
5596 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
5597 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
5598 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
5599 a vector of integers type.
5604 The '``ptrtoint``' instruction converts ``value`` to integer type
5605 ``ty2`` by interpreting the pointer value as an integer and either
5606 truncating or zero extending that value to the size of the integer type.
5607 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
5608 ``value`` is larger than ``ty2`` then a truncation is done. If they are
5609 the same size, then nothing is done (*no-op cast*) other than a type
5615 .. code-block:: llvm
5617 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
5618 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
5619 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
5623 '``inttoptr .. to``' Instruction
5624 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5631 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
5636 The '``inttoptr``' instruction converts an integer ``value`` to a
5637 pointer type, ``ty2``.
5642 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
5643 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
5649 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
5650 applying either a zero extension or a truncation depending on the size
5651 of the integer ``value``. If ``value`` is larger than the size of a
5652 pointer then a truncation is done. If ``value`` is smaller than the size
5653 of a pointer then a zero extension is done. If they are the same size,
5654 nothing is done (*no-op cast*).
5659 .. code-block:: llvm
5661 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
5662 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
5663 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
5664 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
5668 '``bitcast .. to``' Instruction
5669 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5676 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
5681 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
5687 The '``bitcast``' instruction takes a value to cast, which must be a
5688 non-aggregate first class value, and a type to cast it to, which must
5689 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
5690 bit sizes of ``value`` and the destination type, ``ty2``, must be
5691 identical. If the source type is a pointer, the destination type must
5692 also be a pointer of the same size. This instruction supports bitwise
5693 conversion of vectors to integers and to vectors of other types (as
5694 long as they have the same size).
5699 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
5700 is always a *no-op cast* because no bits change with this
5701 conversion. The conversion is done as if the ``value`` had been stored
5702 to memory and read back as type ``ty2``. Pointer (or vector of
5703 pointers) types may only be converted to other pointer (or vector of
5704 pointers) types with the same address space through this instruction.
5705 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
5706 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
5711 .. code-block:: llvm
5713 %X = bitcast i8 255 to i8 ; yields i8 :-1
5714 %Y = bitcast i32* %x to sint* ; yields sint*:%x
5715 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
5716 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
5718 .. _i_addrspacecast:
5720 '``addrspacecast .. to``' Instruction
5721 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5728 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
5733 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
5734 address space ``n`` to type ``pty2`` in address space ``m``.
5739 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
5740 to cast and a pointer type to cast it to, which must have a different
5746 The '``addrspacecast``' instruction converts the pointer value
5747 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
5748 value modification, depending on the target and the address space
5749 pair. Pointer conversions within the same address space must be
5750 performed with the ``bitcast`` instruction. Note that if the address space
5751 conversion is legal then both result and operand refer to the same memory
5757 .. code-block:: llvm
5759 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
5760 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
5761 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
5768 The instructions in this category are the "miscellaneous" instructions,
5769 which defy better classification.
5773 '``icmp``' Instruction
5774 ^^^^^^^^^^^^^^^^^^^^^^
5781 <result> = icmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5786 The '``icmp``' instruction returns a boolean value or a vector of
5787 boolean values based on comparison of its two integer, integer vector,
5788 pointer, or pointer vector operands.
5793 The '``icmp``' instruction takes three operands. The first operand is
5794 the condition code indicating the kind of comparison to perform. It is
5795 not a value, just a keyword. The possible condition code are:
5798 #. ``ne``: not equal
5799 #. ``ugt``: unsigned greater than
5800 #. ``uge``: unsigned greater or equal
5801 #. ``ult``: unsigned less than
5802 #. ``ule``: unsigned less or equal
5803 #. ``sgt``: signed greater than
5804 #. ``sge``: signed greater or equal
5805 #. ``slt``: signed less than
5806 #. ``sle``: signed less or equal
5808 The remaining two arguments must be :ref:`integer <t_integer>` or
5809 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
5810 must also be identical types.
5815 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
5816 code given as ``cond``. The comparison performed always yields either an
5817 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
5819 #. ``eq``: yields ``true`` if the operands are equal, ``false``
5820 otherwise. No sign interpretation is necessary or performed.
5821 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
5822 otherwise. No sign interpretation is necessary or performed.
5823 #. ``ugt``: interprets the operands as unsigned values and yields
5824 ``true`` if ``op1`` is greater than ``op2``.
5825 #. ``uge``: interprets the operands as unsigned values and yields
5826 ``true`` if ``op1`` is greater than or equal to ``op2``.
5827 #. ``ult``: interprets the operands as unsigned values and yields
5828 ``true`` if ``op1`` is less than ``op2``.
5829 #. ``ule``: interprets the operands as unsigned values and yields
5830 ``true`` if ``op1`` is less than or equal to ``op2``.
5831 #. ``sgt``: interprets the operands as signed values and yields ``true``
5832 if ``op1`` is greater than ``op2``.
5833 #. ``sge``: interprets the operands as signed values and yields ``true``
5834 if ``op1`` is greater than or equal to ``op2``.
5835 #. ``slt``: interprets the operands as signed values and yields ``true``
5836 if ``op1`` is less than ``op2``.
5837 #. ``sle``: interprets the operands as signed values and yields ``true``
5838 if ``op1`` is less than or equal to ``op2``.
5840 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
5841 are compared as if they were integers.
5843 If the operands are integer vectors, then they are compared element by
5844 element. The result is an ``i1`` vector with the same number of elements
5845 as the values being compared. Otherwise, the result is an ``i1``.
5850 .. code-block:: llvm
5852 <result> = icmp eq i32 4, 5 ; yields: result=false
5853 <result> = icmp ne float* %X, %X ; yields: result=false
5854 <result> = icmp ult i16 4, 5 ; yields: result=true
5855 <result> = icmp sgt i16 4, 5 ; yields: result=false
5856 <result> = icmp ule i16 -4, 5 ; yields: result=false
5857 <result> = icmp sge i16 4, 5 ; yields: result=false
5859 Note that the code generator does not yet support vector types with the
5860 ``icmp`` instruction.
5864 '``fcmp``' Instruction
5865 ^^^^^^^^^^^^^^^^^^^^^^
5872 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5877 The '``fcmp``' instruction returns a boolean value or vector of boolean
5878 values based on comparison of its operands.
5880 If the operands are floating point scalars, then the result type is a
5881 boolean (:ref:`i1 <t_integer>`).
5883 If the operands are floating point vectors, then the result type is a
5884 vector of boolean with the same number of elements as the operands being
5890 The '``fcmp``' instruction takes three operands. The first operand is
5891 the condition code indicating the kind of comparison to perform. It is
5892 not a value, just a keyword. The possible condition code are:
5894 #. ``false``: no comparison, always returns false
5895 #. ``oeq``: ordered and equal
5896 #. ``ogt``: ordered and greater than
5897 #. ``oge``: ordered and greater than or equal
5898 #. ``olt``: ordered and less than
5899 #. ``ole``: ordered and less than or equal
5900 #. ``one``: ordered and not equal
5901 #. ``ord``: ordered (no nans)
5902 #. ``ueq``: unordered or equal
5903 #. ``ugt``: unordered or greater than
5904 #. ``uge``: unordered or greater than or equal
5905 #. ``ult``: unordered or less than
5906 #. ``ule``: unordered or less than or equal
5907 #. ``une``: unordered or not equal
5908 #. ``uno``: unordered (either nans)
5909 #. ``true``: no comparison, always returns true
5911 *Ordered* means that neither operand is a QNAN while *unordered* means
5912 that either operand may be a QNAN.
5914 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
5915 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
5916 type. They must have identical types.
5921 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
5922 condition code given as ``cond``. If the operands are vectors, then the
5923 vectors are compared element by element. Each comparison performed
5924 always yields an :ref:`i1 <t_integer>` result, as follows:
5926 #. ``false``: always yields ``false``, regardless of operands.
5927 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
5928 is equal to ``op2``.
5929 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
5930 is greater than ``op2``.
5931 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
5932 is greater than or equal to ``op2``.
5933 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
5934 is less than ``op2``.
5935 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
5936 is less than or equal to ``op2``.
5937 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
5938 is not equal to ``op2``.
5939 #. ``ord``: yields ``true`` if both operands are not a QNAN.
5940 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
5942 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
5943 greater than ``op2``.
5944 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
5945 greater than or equal to ``op2``.
5946 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
5948 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
5949 less than or equal to ``op2``.
5950 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
5951 not equal to ``op2``.
5952 #. ``uno``: yields ``true`` if either operand is a QNAN.
5953 #. ``true``: always yields ``true``, regardless of operands.
5958 .. code-block:: llvm
5960 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
5961 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
5962 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
5963 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
5965 Note that the code generator does not yet support vector types with the
5966 ``fcmp`` instruction.
5970 '``phi``' Instruction
5971 ^^^^^^^^^^^^^^^^^^^^^
5978 <result> = phi <ty> [ <val0>, <label0>], ...
5983 The '``phi``' instruction is used to implement the φ node in the SSA
5984 graph representing the function.
5989 The type of the incoming values is specified with the first type field.
5990 After this, the '``phi``' instruction takes a list of pairs as
5991 arguments, with one pair for each predecessor basic block of the current
5992 block. Only values of :ref:`first class <t_firstclass>` type may be used as
5993 the value arguments to the PHI node. Only labels may be used as the
5996 There must be no non-phi instructions between the start of a basic block
5997 and the PHI instructions: i.e. PHI instructions must be first in a basic
6000 For the purposes of the SSA form, the use of each incoming value is
6001 deemed to occur on the edge from the corresponding predecessor block to
6002 the current block (but after any definition of an '``invoke``'
6003 instruction's return value on the same edge).
6008 At runtime, the '``phi``' instruction logically takes on the value
6009 specified by the pair corresponding to the predecessor basic block that
6010 executed just prior to the current block.
6015 .. code-block:: llvm
6017 Loop: ; Infinite loop that counts from 0 on up...
6018 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
6019 %nextindvar = add i32 %indvar, 1
6024 '``select``' Instruction
6025 ^^^^^^^^^^^^^^^^^^^^^^^^
6032 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
6034 selty is either i1 or {<N x i1>}
6039 The '``select``' instruction is used to choose one value based on a
6040 condition, without branching.
6045 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
6046 values indicating the condition, and two values of the same :ref:`first
6047 class <t_firstclass>` type. If the val1/val2 are vectors and the
6048 condition is a scalar, then entire vectors are selected, not individual
6054 If the condition is an i1 and it evaluates to 1, the instruction returns
6055 the first value argument; otherwise, it returns the second value
6058 If the condition is a vector of i1, then the value arguments must be
6059 vectors of the same size, and the selection is done element by element.
6064 .. code-block:: llvm
6066 %X = select i1 true, i8 17, i8 42 ; yields i8:17
6070 '``call``' Instruction
6071 ^^^^^^^^^^^^^^^^^^^^^^
6078 <result> = [tail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
6083 The '``call``' instruction represents a simple function call.
6088 This instruction requires several arguments:
6090 #. The optional "tail" marker indicates that the callee function does
6091 not access any allocas or varargs in the caller. Note that calls may
6092 be marked "tail" even if they do not occur before a
6093 :ref:`ret <i_ret>` instruction. If the "tail" marker is present, the
6094 function call is eligible for tail call optimization, but `might not
6095 in fact be optimized into a jump <CodeGenerator.html#tailcallopt>`_.
6096 The code generator may optimize calls marked "tail" with either 1)
6097 automatic `sibling call
6098 optimization <CodeGenerator.html#sibcallopt>`_ when the caller and
6099 callee have matching signatures, or 2) forced tail call optimization
6100 when the following extra requirements are met:
6102 - Caller and callee both have the calling convention ``fastcc``.
6103 - The call is in tail position (ret immediately follows call and ret
6104 uses value of call or is void).
6105 - Option ``-tailcallopt`` is enabled, or
6106 ``llvm::GuaranteedTailCallOpt`` is ``true``.
6107 - `Platform specific constraints are
6108 met. <CodeGenerator.html#tailcallopt>`_
6110 #. The optional "cconv" marker indicates which :ref:`calling
6111 convention <callingconv>` the call should use. If none is
6112 specified, the call defaults to using C calling conventions. The
6113 calling convention of the call must match the calling convention of
6114 the target function, or else the behavior is undefined.
6115 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6116 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6118 #. '``ty``': the type of the call instruction itself which is also the
6119 type of the return value. Functions that return no value are marked
6121 #. '``fnty``': shall be the signature of the pointer to function value
6122 being invoked. The argument types must match the types implied by
6123 this signature. This type can be omitted if the function is not
6124 varargs and if the function type does not return a pointer to a
6126 #. '``fnptrval``': An LLVM value containing a pointer to a function to
6127 be invoked. In most cases, this is a direct function invocation, but
6128 indirect ``call``'s are just as possible, calling an arbitrary pointer
6130 #. '``function args``': argument list whose types match the function
6131 signature argument types and parameter attributes. All arguments must
6132 be of :ref:`first class <t_firstclass>` type. If the function signature
6133 indicates the function accepts a variable number of arguments, the
6134 extra arguments can be specified.
6135 #. The optional :ref:`function attributes <fnattrs>` list. Only
6136 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
6137 attributes are valid here.
6142 The '``call``' instruction is used to cause control flow to transfer to
6143 a specified function, with its incoming arguments bound to the specified
6144 values. Upon a '``ret``' instruction in the called function, control
6145 flow continues with the instruction after the function call, and the
6146 return value of the function is bound to the result argument.
6151 .. code-block:: llvm
6153 %retval = call i32 @test(i32 %argc)
6154 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
6155 %X = tail call i32 @foo() ; yields i32
6156 %Y = tail call fastcc i32 @foo() ; yields i32
6157 call void %foo(i8 97 signext)
6159 %struct.A = type { i32, i8 }
6160 %r = call %struct.A @foo() ; yields { 32, i8 }
6161 %gr = extractvalue %struct.A %r, 0 ; yields i32
6162 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
6163 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
6164 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
6166 llvm treats calls to some functions with names and arguments that match
6167 the standard C99 library as being the C99 library functions, and may
6168 perform optimizations or generate code for them under that assumption.
6169 This is something we'd like to change in the future to provide better
6170 support for freestanding environments and non-C-based languages.
6174 '``va_arg``' Instruction
6175 ^^^^^^^^^^^^^^^^^^^^^^^^
6182 <resultval> = va_arg <va_list*> <arglist>, <argty>
6187 The '``va_arg``' instruction is used to access arguments passed through
6188 the "variable argument" area of a function call. It is used to implement
6189 the ``va_arg`` macro in C.
6194 This instruction takes a ``va_list*`` value and the type of the
6195 argument. It returns a value of the specified argument type and
6196 increments the ``va_list`` to point to the next argument. The actual
6197 type of ``va_list`` is target specific.
6202 The '``va_arg``' instruction loads an argument of the specified type
6203 from the specified ``va_list`` and causes the ``va_list`` to point to
6204 the next argument. For more information, see the variable argument
6205 handling :ref:`Intrinsic Functions <int_varargs>`.
6207 It is legal for this instruction to be called in a function which does
6208 not take a variable number of arguments, for example, the ``vfprintf``
6211 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6212 function <intrinsics>` because it takes a type as an argument.
6217 See the :ref:`variable argument processing <int_varargs>` section.
6219 Note that the code generator does not yet fully support va\_arg on many
6220 targets. Also, it does not currently support va\_arg with aggregate
6221 types on any target.
6225 '``landingpad``' Instruction
6226 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6233 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6234 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6236 <clause> := catch <type> <value>
6237 <clause> := filter <array constant type> <array constant>
6242 The '``landingpad``' instruction is used by `LLVM's exception handling
6243 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6244 is a landing pad --- one where the exception lands, and corresponds to the
6245 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6246 defines values supplied by the personality function (``pers_fn``) upon
6247 re-entry to the function. The ``resultval`` has the type ``resultty``.
6252 This instruction takes a ``pers_fn`` value. This is the personality
6253 function associated with the unwinding mechanism. The optional
6254 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6256 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6257 contains the global variable representing the "type" that may be caught
6258 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6259 clause takes an array constant as its argument. Use
6260 "``[0 x i8**] undef``" for a filter which cannot throw. The
6261 '``landingpad``' instruction must contain *at least* one ``clause`` or
6262 the ``cleanup`` flag.
6267 The '``landingpad``' instruction defines the values which are set by the
6268 personality function (``pers_fn``) upon re-entry to the function, and
6269 therefore the "result type" of the ``landingpad`` instruction. As with
6270 calling conventions, how the personality function results are
6271 represented in LLVM IR is target specific.
6273 The clauses are applied in order from top to bottom. If two
6274 ``landingpad`` instructions are merged together through inlining, the
6275 clauses from the calling function are appended to the list of clauses.
6276 When the call stack is being unwound due to an exception being thrown,
6277 the exception is compared against each ``clause`` in turn. If it doesn't
6278 match any of the clauses, and the ``cleanup`` flag is not set, then
6279 unwinding continues further up the call stack.
6281 The ``landingpad`` instruction has several restrictions:
6283 - A landing pad block is a basic block which is the unwind destination
6284 of an '``invoke``' instruction.
6285 - A landing pad block must have a '``landingpad``' instruction as its
6286 first non-PHI instruction.
6287 - There can be only one '``landingpad``' instruction within the landing
6289 - A basic block that is not a landing pad block may not include a
6290 '``landingpad``' instruction.
6291 - All '``landingpad``' instructions in a function must have the same
6292 personality function.
6297 .. code-block:: llvm
6299 ;; A landing pad which can catch an integer.
6300 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6302 ;; A landing pad that is a cleanup.
6303 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6305 ;; A landing pad which can catch an integer and can only throw a double.
6306 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6308 filter [1 x i8**] [@_ZTId]
6315 LLVM supports the notion of an "intrinsic function". These functions
6316 have well known names and semantics and are required to follow certain
6317 restrictions. Overall, these intrinsics represent an extension mechanism
6318 for the LLVM language that does not require changing all of the
6319 transformations in LLVM when adding to the language (or the bitcode
6320 reader/writer, the parser, etc...).
6322 Intrinsic function names must all start with an "``llvm.``" prefix. This
6323 prefix is reserved in LLVM for intrinsic names; thus, function names may
6324 not begin with this prefix. Intrinsic functions must always be external
6325 functions: you cannot define the body of intrinsic functions. Intrinsic
6326 functions may only be used in call or invoke instructions: it is illegal
6327 to take the address of an intrinsic function. Additionally, because
6328 intrinsic functions are part of the LLVM language, it is required if any
6329 are added that they be documented here.
6331 Some intrinsic functions can be overloaded, i.e., the intrinsic
6332 represents a family of functions that perform the same operation but on
6333 different data types. Because LLVM can represent over 8 million
6334 different integer types, overloading is used commonly to allow an
6335 intrinsic function to operate on any integer type. One or more of the
6336 argument types or the result type can be overloaded to accept any
6337 integer type. Argument types may also be defined as exactly matching a
6338 previous argument's type or the result type. This allows an intrinsic
6339 function which accepts multiple arguments, but needs all of them to be
6340 of the same type, to only be overloaded with respect to a single
6341 argument or the result.
6343 Overloaded intrinsics will have the names of its overloaded argument
6344 types encoded into its function name, each preceded by a period. Only
6345 those types which are overloaded result in a name suffix. Arguments
6346 whose type is matched against another type do not. For example, the
6347 ``llvm.ctpop`` function can take an integer of any width and returns an
6348 integer of exactly the same integer width. This leads to a family of
6349 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6350 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6351 overloaded, and only one type suffix is required. Because the argument's
6352 type is matched against the return type, it does not require its own
6355 To learn how to add an intrinsic function, please see the `Extending
6356 LLVM Guide <ExtendingLLVM.html>`_.
6360 Variable Argument Handling Intrinsics
6361 -------------------------------------
6363 Variable argument support is defined in LLVM with the
6364 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6365 functions. These functions are related to the similarly named macros
6366 defined in the ``<stdarg.h>`` header file.
6368 All of these functions operate on arguments that use a target-specific
6369 value type "``va_list``". The LLVM assembly language reference manual
6370 does not define what this type is, so all transformations should be
6371 prepared to handle these functions regardless of the type used.
6373 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6374 variable argument handling intrinsic functions are used.
6376 .. code-block:: llvm
6378 define i32 @test(i32 %X, ...) {
6379 ; Initialize variable argument processing
6381 %ap2 = bitcast i8** %ap to i8*
6382 call void @llvm.va_start(i8* %ap2)
6384 ; Read a single integer argument
6385 %tmp = va_arg i8** %ap, i32
6387 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6389 %aq2 = bitcast i8** %aq to i8*
6390 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6391 call void @llvm.va_end(i8* %aq2)
6393 ; Stop processing of arguments.
6394 call void @llvm.va_end(i8* %ap2)
6398 declare void @llvm.va_start(i8*)
6399 declare void @llvm.va_copy(i8*, i8*)
6400 declare void @llvm.va_end(i8*)
6404 '``llvm.va_start``' Intrinsic
6405 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6412 declare void @llvm.va_start(i8* <arglist>)
6417 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6418 subsequent use by ``va_arg``.
6423 The argument is a pointer to a ``va_list`` element to initialize.
6428 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6429 available in C. In a target-dependent way, it initializes the
6430 ``va_list`` element to which the argument points, so that the next call
6431 to ``va_arg`` will produce the first variable argument passed to the
6432 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6433 to know the last argument of the function as the compiler can figure
6436 '``llvm.va_end``' Intrinsic
6437 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6444 declare void @llvm.va_end(i8* <arglist>)
6449 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6450 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6455 The argument is a pointer to a ``va_list`` to destroy.
6460 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6461 available in C. In a target-dependent way, it destroys the ``va_list``
6462 element to which the argument points. Calls to
6463 :ref:`llvm.va_start <int_va_start>` and
6464 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6469 '``llvm.va_copy``' Intrinsic
6470 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6477 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6482 The '``llvm.va_copy``' intrinsic copies the current argument position
6483 from the source argument list to the destination argument list.
6488 The first argument is a pointer to a ``va_list`` element to initialize.
6489 The second argument is a pointer to a ``va_list`` element to copy from.
6494 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6495 available in C. In a target-dependent way, it copies the source
6496 ``va_list`` element into the destination ``va_list`` element. This
6497 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6498 arbitrarily complex and require, for example, memory allocation.
6500 Accurate Garbage Collection Intrinsics
6501 --------------------------------------
6503 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6504 (GC) requires the implementation and generation of these intrinsics.
6505 These intrinsics allow identification of :ref:`GC roots on the
6506 stack <int_gcroot>`, as well as garbage collector implementations that
6507 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6508 Front-ends for type-safe garbage collected languages should generate
6509 these intrinsics to make use of the LLVM garbage collectors. For more
6510 details, see `Accurate Garbage Collection with
6511 LLVM <GarbageCollection.html>`_.
6513 The garbage collection intrinsics only operate on objects in the generic
6514 address space (address space zero).
6518 '``llvm.gcroot``' Intrinsic
6519 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6526 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6531 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
6532 the code generator, and allows some metadata to be associated with it.
6537 The first argument specifies the address of a stack object that contains
6538 the root pointer. The second pointer (which must be either a constant or
6539 a global value address) contains the meta-data to be associated with the
6545 At runtime, a call to this intrinsic stores a null pointer into the
6546 "ptrloc" location. At compile-time, the code generator generates
6547 information to allow the runtime to find the pointer at GC safe points.
6548 The '``llvm.gcroot``' intrinsic may only be used in a function which
6549 :ref:`specifies a GC algorithm <gc>`.
6553 '``llvm.gcread``' Intrinsic
6554 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6561 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6566 The '``llvm.gcread``' intrinsic identifies reads of references from heap
6567 locations, allowing garbage collector implementations that require read
6573 The second argument is the address to read from, which should be an
6574 address allocated from the garbage collector. The first object is a
6575 pointer to the start of the referenced object, if needed by the language
6576 runtime (otherwise null).
6581 The '``llvm.gcread``' intrinsic has the same semantics as a load
6582 instruction, but may be replaced with substantially more complex code by
6583 the garbage collector runtime, as needed. The '``llvm.gcread``'
6584 intrinsic may only be used in a function which :ref:`specifies a GC
6589 '``llvm.gcwrite``' Intrinsic
6590 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6597 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6602 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
6603 locations, allowing garbage collector implementations that require write
6604 barriers (such as generational or reference counting collectors).
6609 The first argument is the reference to store, the second is the start of
6610 the object to store it to, and the third is the address of the field of
6611 Obj to store to. If the runtime does not require a pointer to the
6612 object, Obj may be null.
6617 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
6618 instruction, but may be replaced with substantially more complex code by
6619 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
6620 intrinsic may only be used in a function which :ref:`specifies a GC
6623 Code Generator Intrinsics
6624 -------------------------
6626 These intrinsics are provided by LLVM to expose special features that
6627 may only be implemented with code generator support.
6629 '``llvm.returnaddress``' Intrinsic
6630 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6637 declare i8 *@llvm.returnaddress(i32 <level>)
6642 The '``llvm.returnaddress``' intrinsic attempts to compute a
6643 target-specific value indicating the return address of the current
6644 function or one of its callers.
6649 The argument to this intrinsic indicates which function to return the
6650 address for. Zero indicates the calling function, one indicates its
6651 caller, etc. The argument is **required** to be a constant integer
6657 The '``llvm.returnaddress``' intrinsic either returns a pointer
6658 indicating the return address of the specified call frame, or zero if it
6659 cannot be identified. The value returned by this intrinsic is likely to
6660 be incorrect or 0 for arguments other than zero, so it should only be
6661 used for debugging purposes.
6663 Note that calling this intrinsic does not prevent function inlining or
6664 other aggressive transformations, so the value returned may not be that
6665 of the obvious source-language caller.
6667 '``llvm.frameaddress``' Intrinsic
6668 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6675 declare i8* @llvm.frameaddress(i32 <level>)
6680 The '``llvm.frameaddress``' intrinsic attempts to return the
6681 target-specific frame pointer value for the specified stack frame.
6686 The argument to this intrinsic indicates which function to return the
6687 frame pointer for. Zero indicates the calling function, one indicates
6688 its caller, etc. The argument is **required** to be a constant integer
6694 The '``llvm.frameaddress``' intrinsic either returns a pointer
6695 indicating the frame address of the specified call frame, or zero if it
6696 cannot be identified. The value returned by this intrinsic is likely to
6697 be incorrect or 0 for arguments other than zero, so it should only be
6698 used for debugging purposes.
6700 Note that calling this intrinsic does not prevent function inlining or
6701 other aggressive transformations, so the value returned may not be that
6702 of the obvious source-language caller.
6706 '``llvm.stacksave``' Intrinsic
6707 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6714 declare i8* @llvm.stacksave()
6719 The '``llvm.stacksave``' intrinsic is used to remember the current state
6720 of the function stack, for use with
6721 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
6722 implementing language features like scoped automatic variable sized
6728 This intrinsic returns a opaque pointer value that can be passed to
6729 :ref:`llvm.stackrestore <int_stackrestore>`. When an
6730 ``llvm.stackrestore`` intrinsic is executed with a value saved from
6731 ``llvm.stacksave``, it effectively restores the state of the stack to
6732 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
6733 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
6734 were allocated after the ``llvm.stacksave`` was executed.
6736 .. _int_stackrestore:
6738 '``llvm.stackrestore``' Intrinsic
6739 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6746 declare void @llvm.stackrestore(i8* %ptr)
6751 The '``llvm.stackrestore``' intrinsic is used to restore the state of
6752 the function stack to the state it was in when the corresponding
6753 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
6754 useful for implementing language features like scoped automatic variable
6755 sized arrays in C99.
6760 See the description for :ref:`llvm.stacksave <int_stacksave>`.
6762 '``llvm.prefetch``' Intrinsic
6763 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6770 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
6775 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
6776 insert a prefetch instruction if supported; otherwise, it is a noop.
6777 Prefetches have no effect on the behavior of the program but can change
6778 its performance characteristics.
6783 ``address`` is the address to be prefetched, ``rw`` is the specifier
6784 determining if the fetch should be for a read (0) or write (1), and
6785 ``locality`` is a temporal locality specifier ranging from (0) - no
6786 locality, to (3) - extremely local keep in cache. The ``cache type``
6787 specifies whether the prefetch is performed on the data (1) or
6788 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
6789 arguments must be constant integers.
6794 This intrinsic does not modify the behavior of the program. In
6795 particular, prefetches cannot trap and do not produce a value. On
6796 targets that support this intrinsic, the prefetch can provide hints to
6797 the processor cache for better performance.
6799 '``llvm.pcmarker``' Intrinsic
6800 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6807 declare void @llvm.pcmarker(i32 <id>)
6812 The '``llvm.pcmarker``' intrinsic is a method to export a Program
6813 Counter (PC) in a region of code to simulators and other tools. The
6814 method is target specific, but it is expected that the marker will use
6815 exported symbols to transmit the PC of the marker. The marker makes no
6816 guarantees that it will remain with any specific instruction after
6817 optimizations. It is possible that the presence of a marker will inhibit
6818 optimizations. The intended use is to be inserted after optimizations to
6819 allow correlations of simulation runs.
6824 ``id`` is a numerical id identifying the marker.
6829 This intrinsic does not modify the behavior of the program. Backends
6830 that do not support this intrinsic may ignore it.
6832 '``llvm.readcyclecounter``' Intrinsic
6833 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6840 declare i64 @llvm.readcyclecounter()
6845 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
6846 counter register (or similar low latency, high accuracy clocks) on those
6847 targets that support it. On X86, it should map to RDTSC. On Alpha, it
6848 should map to RPCC. As the backing counters overflow quickly (on the
6849 order of 9 seconds on alpha), this should only be used for small
6855 When directly supported, reading the cycle counter should not modify any
6856 memory. Implementations are allowed to either return a application
6857 specific value or a system wide value. On backends without support, this
6858 is lowered to a constant 0.
6860 Note that runtime support may be conditional on the privilege-level code is
6861 running at and the host platform.
6863 Standard C Library Intrinsics
6864 -----------------------------
6866 LLVM provides intrinsics for a few important standard C library
6867 functions. These intrinsics allow source-language front-ends to pass
6868 information about the alignment of the pointer arguments to the code
6869 generator, providing opportunity for more efficient code generation.
6873 '``llvm.memcpy``' Intrinsic
6874 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6879 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
6880 integer bit width and for different address spaces. Not all targets
6881 support all bit widths however.
6885 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6886 i32 <len>, i32 <align>, i1 <isvolatile>)
6887 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6888 i64 <len>, i32 <align>, i1 <isvolatile>)
6893 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6894 source location to the destination location.
6896 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
6897 intrinsics do not return a value, takes extra alignment/isvolatile
6898 arguments and the pointers can be in specified address spaces.
6903 The first argument is a pointer to the destination, the second is a
6904 pointer to the source. The third argument is an integer argument
6905 specifying the number of bytes to copy, the fourth argument is the
6906 alignment of the source and destination locations, and the fifth is a
6907 boolean indicating a volatile access.
6909 If the call to this intrinsic has an alignment value that is not 0 or 1,
6910 then the caller guarantees that both the source and destination pointers
6911 are aligned to that boundary.
6913 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
6914 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6915 very cleanly specified and it is unwise to depend on it.
6920 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6921 source location to the destination location, which are not allowed to
6922 overlap. It copies "len" bytes of memory over. If the argument is known
6923 to be aligned to some boundary, this can be specified as the fourth
6924 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
6926 '``llvm.memmove``' Intrinsic
6927 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6932 This is an overloaded intrinsic. You can use llvm.memmove on any integer
6933 bit width and for different address space. Not all targets support all
6938 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6939 i32 <len>, i32 <align>, i1 <isvolatile>)
6940 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6941 i64 <len>, i32 <align>, i1 <isvolatile>)
6946 The '``llvm.memmove.*``' intrinsics move a block of memory from the
6947 source location to the destination location. It is similar to the
6948 '``llvm.memcpy``' intrinsic but allows the two memory locations to
6951 Note that, unlike the standard libc function, the ``llvm.memmove.*``
6952 intrinsics do not return a value, takes extra alignment/isvolatile
6953 arguments and the pointers can be in specified address spaces.
6958 The first argument is a pointer to the destination, the second is a
6959 pointer to the source. The third argument is an integer argument
6960 specifying the number of bytes to copy, the fourth argument is the
6961 alignment of the source and destination locations, and the fifth is a
6962 boolean indicating a volatile access.
6964 If the call to this intrinsic has an alignment value that is not 0 or 1,
6965 then the caller guarantees that the source and destination pointers are
6966 aligned to that boundary.
6968 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
6969 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
6970 not very cleanly specified and it is unwise to depend on it.
6975 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
6976 source location to the destination location, which may overlap. It
6977 copies "len" bytes of memory over. If the argument is known to be
6978 aligned to some boundary, this can be specified as the fourth argument,
6979 otherwise it should be set to 0 or 1 (both meaning no alignment).
6981 '``llvm.memset.*``' Intrinsics
6982 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6987 This is an overloaded intrinsic. You can use llvm.memset on any integer
6988 bit width and for different address spaces. However, not all targets
6989 support all bit widths.
6993 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
6994 i32 <len>, i32 <align>, i1 <isvolatile>)
6995 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
6996 i64 <len>, i32 <align>, i1 <isvolatile>)
7001 The '``llvm.memset.*``' intrinsics fill a block of memory with a
7002 particular byte value.
7004 Note that, unlike the standard libc function, the ``llvm.memset``
7005 intrinsic does not return a value and takes extra alignment/volatile
7006 arguments. Also, the destination can be in an arbitrary address space.
7011 The first argument is a pointer to the destination to fill, the second
7012 is the byte value with which to fill it, the third argument is an
7013 integer argument specifying the number of bytes to fill, and the fourth
7014 argument is the known alignment of the destination location.
7016 If the call to this intrinsic has an alignment value that is not 0 or 1,
7017 then the caller guarantees that the destination pointer is aligned to
7020 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
7021 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7022 very cleanly specified and it is unwise to depend on it.
7027 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
7028 at the destination location. If the argument is known to be aligned to
7029 some boundary, this can be specified as the fourth argument, otherwise
7030 it should be set to 0 or 1 (both meaning no alignment).
7032 '``llvm.sqrt.*``' Intrinsic
7033 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7038 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
7039 floating point or vector of floating point type. Not all targets support
7044 declare float @llvm.sqrt.f32(float %Val)
7045 declare double @llvm.sqrt.f64(double %Val)
7046 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
7047 declare fp128 @llvm.sqrt.f128(fp128 %Val)
7048 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
7053 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
7054 returning the same value as the libm '``sqrt``' functions would. Unlike
7055 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
7056 negative numbers other than -0.0 (which allows for better optimization,
7057 because there is no need to worry about errno being set).
7058 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
7063 The argument and return value are floating point numbers of the same
7069 This function returns the sqrt of the specified operand if it is a
7070 nonnegative floating point number.
7072 '``llvm.powi.*``' Intrinsic
7073 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7078 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
7079 floating point or vector of floating point type. Not all targets support
7084 declare float @llvm.powi.f32(float %Val, i32 %power)
7085 declare double @llvm.powi.f64(double %Val, i32 %power)
7086 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
7087 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
7088 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
7093 The '``llvm.powi.*``' intrinsics return the first operand raised to the
7094 specified (positive or negative) power. The order of evaluation of
7095 multiplications is not defined. When a vector of floating point type is
7096 used, the second argument remains a scalar integer value.
7101 The second argument is an integer power, and the first is a value to
7102 raise to that power.
7107 This function returns the first value raised to the second power with an
7108 unspecified sequence of rounding operations.
7110 '``llvm.sin.*``' Intrinsic
7111 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7116 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
7117 floating point or vector of floating point type. Not all targets support
7122 declare float @llvm.sin.f32(float %Val)
7123 declare double @llvm.sin.f64(double %Val)
7124 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
7125 declare fp128 @llvm.sin.f128(fp128 %Val)
7126 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
7131 The '``llvm.sin.*``' intrinsics return the sine of the operand.
7136 The argument and return value are floating point numbers of the same
7142 This function returns the sine of the specified operand, returning the
7143 same values as the libm ``sin`` functions would, and handles error
7144 conditions in the same way.
7146 '``llvm.cos.*``' Intrinsic
7147 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7152 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
7153 floating point or vector of floating point type. Not all targets support
7158 declare float @llvm.cos.f32(float %Val)
7159 declare double @llvm.cos.f64(double %Val)
7160 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
7161 declare fp128 @llvm.cos.f128(fp128 %Val)
7162 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
7167 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
7172 The argument and return value are floating point numbers of the same
7178 This function returns the cosine of the specified operand, returning the
7179 same values as the libm ``cos`` functions would, and handles error
7180 conditions in the same way.
7182 '``llvm.pow.*``' Intrinsic
7183 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7188 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
7189 floating point or vector of floating point type. Not all targets support
7194 declare float @llvm.pow.f32(float %Val, float %Power)
7195 declare double @llvm.pow.f64(double %Val, double %Power)
7196 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7197 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7198 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7203 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7204 specified (positive or negative) power.
7209 The second argument is a floating point power, and the first is a value
7210 to raise to that power.
7215 This function returns the first value raised to the second power,
7216 returning the same values as the libm ``pow`` functions would, and
7217 handles error conditions in the same way.
7219 '``llvm.exp.*``' Intrinsic
7220 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7225 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
7226 floating point or vector of floating point type. Not all targets support
7231 declare float @llvm.exp.f32(float %Val)
7232 declare double @llvm.exp.f64(double %Val)
7233 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7234 declare fp128 @llvm.exp.f128(fp128 %Val)
7235 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7240 The '``llvm.exp.*``' intrinsics perform the exp function.
7245 The argument and return value are floating point numbers of the same
7251 This function returns the same values as the libm ``exp`` functions
7252 would, and handles error conditions in the same way.
7254 '``llvm.exp2.*``' Intrinsic
7255 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7260 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
7261 floating point or vector of floating point type. Not all targets support
7266 declare float @llvm.exp2.f32(float %Val)
7267 declare double @llvm.exp2.f64(double %Val)
7268 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
7269 declare fp128 @llvm.exp2.f128(fp128 %Val)
7270 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
7275 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
7280 The argument and return value are floating point numbers of the same
7286 This function returns the same values as the libm ``exp2`` functions
7287 would, and handles error conditions in the same way.
7289 '``llvm.log.*``' Intrinsic
7290 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7295 This is an overloaded intrinsic. You can use ``llvm.log`` on any
7296 floating point or vector of floating point type. Not all targets support
7301 declare float @llvm.log.f32(float %Val)
7302 declare double @llvm.log.f64(double %Val)
7303 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7304 declare fp128 @llvm.log.f128(fp128 %Val)
7305 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7310 The '``llvm.log.*``' intrinsics perform the log function.
7315 The argument and return value are floating point numbers of the same
7321 This function returns the same values as the libm ``log`` functions
7322 would, and handles error conditions in the same way.
7324 '``llvm.log10.*``' Intrinsic
7325 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7330 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
7331 floating point or vector of floating point type. Not all targets support
7336 declare float @llvm.log10.f32(float %Val)
7337 declare double @llvm.log10.f64(double %Val)
7338 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
7339 declare fp128 @llvm.log10.f128(fp128 %Val)
7340 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
7345 The '``llvm.log10.*``' intrinsics perform the log10 function.
7350 The argument and return value are floating point numbers of the same
7356 This function returns the same values as the libm ``log10`` functions
7357 would, and handles error conditions in the same way.
7359 '``llvm.log2.*``' Intrinsic
7360 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7365 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
7366 floating point or vector of floating point type. Not all targets support
7371 declare float @llvm.log2.f32(float %Val)
7372 declare double @llvm.log2.f64(double %Val)
7373 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7374 declare fp128 @llvm.log2.f128(fp128 %Val)
7375 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7380 The '``llvm.log2.*``' intrinsics perform the log2 function.
7385 The argument and return value are floating point numbers of the same
7391 This function returns the same values as the libm ``log2`` functions
7392 would, and handles error conditions in the same way.
7394 '``llvm.fma.*``' Intrinsic
7395 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7400 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7401 floating point or vector of floating point type. Not all targets support
7406 declare float @llvm.fma.f32(float %a, float %b, float %c)
7407 declare double @llvm.fma.f64(double %a, double %b, double %c)
7408 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7409 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7410 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7415 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7421 The argument and return value are floating point numbers of the same
7427 This function returns the same values as the libm ``fma`` functions
7430 '``llvm.fabs.*``' Intrinsic
7431 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7436 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7437 floating point or vector of floating point type. Not all targets support
7442 declare float @llvm.fabs.f32(float %Val)
7443 declare double @llvm.fabs.f64(double %Val)
7444 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7445 declare fp128 @llvm.fabs.f128(fp128 %Val)
7446 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7451 The '``llvm.fabs.*``' intrinsics return the absolute value of the
7457 The argument and return value are floating point numbers of the same
7463 This function returns the same values as the libm ``fabs`` functions
7464 would, and handles error conditions in the same way.
7466 '``llvm.copysign.*``' Intrinsic
7467 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7472 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
7473 floating point or vector of floating point type. Not all targets support
7478 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
7479 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
7480 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
7481 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
7482 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
7487 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
7488 first operand and the sign of the second operand.
7493 The arguments and return value are floating point numbers of the same
7499 This function returns the same values as the libm ``copysign``
7500 functions would, and handles error conditions in the same way.
7502 '``llvm.floor.*``' Intrinsic
7503 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7508 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
7509 floating point or vector of floating point type. Not all targets support
7514 declare float @llvm.floor.f32(float %Val)
7515 declare double @llvm.floor.f64(double %Val)
7516 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
7517 declare fp128 @llvm.floor.f128(fp128 %Val)
7518 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
7523 The '``llvm.floor.*``' intrinsics return the floor of the operand.
7528 The argument and return value are floating point numbers of the same
7534 This function returns the same values as the libm ``floor`` functions
7535 would, and handles error conditions in the same way.
7537 '``llvm.ceil.*``' Intrinsic
7538 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7543 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
7544 floating point or vector of floating point type. Not all targets support
7549 declare float @llvm.ceil.f32(float %Val)
7550 declare double @llvm.ceil.f64(double %Val)
7551 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
7552 declare fp128 @llvm.ceil.f128(fp128 %Val)
7553 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
7558 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
7563 The argument and return value are floating point numbers of the same
7569 This function returns the same values as the libm ``ceil`` functions
7570 would, and handles error conditions in the same way.
7572 '``llvm.trunc.*``' Intrinsic
7573 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7578 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
7579 floating point or vector of floating point type. Not all targets support
7584 declare float @llvm.trunc.f32(float %Val)
7585 declare double @llvm.trunc.f64(double %Val)
7586 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
7587 declare fp128 @llvm.trunc.f128(fp128 %Val)
7588 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
7593 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
7594 nearest integer not larger in magnitude than the operand.
7599 The argument and return value are floating point numbers of the same
7605 This function returns the same values as the libm ``trunc`` functions
7606 would, and handles error conditions in the same way.
7608 '``llvm.rint.*``' Intrinsic
7609 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7614 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
7615 floating point or vector of floating point type. Not all targets support
7620 declare float @llvm.rint.f32(float %Val)
7621 declare double @llvm.rint.f64(double %Val)
7622 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
7623 declare fp128 @llvm.rint.f128(fp128 %Val)
7624 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
7629 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
7630 nearest integer. It may raise an inexact floating-point exception if the
7631 operand isn't an integer.
7636 The argument and return value are floating point numbers of the same
7642 This function returns the same values as the libm ``rint`` functions
7643 would, and handles error conditions in the same way.
7645 '``llvm.nearbyint.*``' Intrinsic
7646 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7651 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
7652 floating point or vector of floating point type. Not all targets support
7657 declare float @llvm.nearbyint.f32(float %Val)
7658 declare double @llvm.nearbyint.f64(double %Val)
7659 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
7660 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
7661 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
7666 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
7672 The argument and return value are floating point numbers of the same
7678 This function returns the same values as the libm ``nearbyint``
7679 functions would, and handles error conditions in the same way.
7681 '``llvm.round.*``' Intrinsic
7682 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7687 This is an overloaded intrinsic. You can use ``llvm.round`` on any
7688 floating point or vector of floating point type. Not all targets support
7693 declare float @llvm.round.f32(float %Val)
7694 declare double @llvm.round.f64(double %Val)
7695 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
7696 declare fp128 @llvm.round.f128(fp128 %Val)
7697 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
7702 The '``llvm.round.*``' intrinsics returns the operand rounded to the
7708 The argument and return value are floating point numbers of the same
7714 This function returns the same values as the libm ``round``
7715 functions would, and handles error conditions in the same way.
7717 Bit Manipulation Intrinsics
7718 ---------------------------
7720 LLVM provides intrinsics for a few important bit manipulation
7721 operations. These allow efficient code generation for some algorithms.
7723 '``llvm.bswap.*``' Intrinsics
7724 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7729 This is an overloaded intrinsic function. You can use bswap on any
7730 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
7734 declare i16 @llvm.bswap.i16(i16 <id>)
7735 declare i32 @llvm.bswap.i32(i32 <id>)
7736 declare i64 @llvm.bswap.i64(i64 <id>)
7741 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
7742 values with an even number of bytes (positive multiple of 16 bits).
7743 These are useful for performing operations on data that is not in the
7744 target's native byte order.
7749 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
7750 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
7751 intrinsic returns an i32 value that has the four bytes of the input i32
7752 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
7753 returned i32 will have its bytes in 3, 2, 1, 0 order. The
7754 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
7755 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
7758 '``llvm.ctpop.*``' Intrinsic
7759 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7764 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
7765 bit width, or on any vector with integer elements. Not all targets
7766 support all bit widths or vector types, however.
7770 declare i8 @llvm.ctpop.i8(i8 <src>)
7771 declare i16 @llvm.ctpop.i16(i16 <src>)
7772 declare i32 @llvm.ctpop.i32(i32 <src>)
7773 declare i64 @llvm.ctpop.i64(i64 <src>)
7774 declare i256 @llvm.ctpop.i256(i256 <src>)
7775 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
7780 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
7786 The only argument is the value to be counted. The argument may be of any
7787 integer type, or a vector with integer elements. The return type must
7788 match the argument type.
7793 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
7794 each element of a vector.
7796 '``llvm.ctlz.*``' Intrinsic
7797 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7802 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
7803 integer bit width, or any vector whose elements are integers. Not all
7804 targets support all bit widths or vector types, however.
7808 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
7809 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
7810 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
7811 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
7812 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
7813 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7818 The '``llvm.ctlz``' family of intrinsic functions counts the number of
7819 leading zeros in a variable.
7824 The first argument is the value to be counted. This argument may be of
7825 any integer type, or a vectory with integer element type. The return
7826 type must match the first argument type.
7828 The second argument must be a constant and is a flag to indicate whether
7829 the intrinsic should ensure that a zero as the first argument produces a
7830 defined result. Historically some architectures did not provide a
7831 defined result for zero values as efficiently, and many algorithms are
7832 now predicated on avoiding zero-value inputs.
7837 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
7838 zeros in a variable, or within each element of the vector. If
7839 ``src == 0`` then the result is the size in bits of the type of ``src``
7840 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7841 ``llvm.ctlz(i32 2) = 30``.
7843 '``llvm.cttz.*``' Intrinsic
7844 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7849 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
7850 integer bit width, or any vector of integer elements. Not all targets
7851 support all bit widths or vector types, however.
7855 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
7856 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
7857 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
7858 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
7859 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
7860 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7865 The '``llvm.cttz``' family of intrinsic functions counts the number of
7871 The first argument is the value to be counted. This argument may be of
7872 any integer type, or a vectory with integer element type. The return
7873 type must match the first argument type.
7875 The second argument must be a constant and is a flag to indicate whether
7876 the intrinsic should ensure that a zero as the first argument produces a
7877 defined result. Historically some architectures did not provide a
7878 defined result for zero values as efficiently, and many algorithms are
7879 now predicated on avoiding zero-value inputs.
7884 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
7885 zeros in a variable, or within each element of a vector. If ``src == 0``
7886 then the result is the size in bits of the type of ``src`` if
7887 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7888 ``llvm.cttz(2) = 1``.
7890 Arithmetic with Overflow Intrinsics
7891 -----------------------------------
7893 LLVM provides intrinsics for some arithmetic with overflow operations.
7895 '``llvm.sadd.with.overflow.*``' Intrinsics
7896 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7901 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
7902 on any integer bit width.
7906 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
7907 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7908 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
7913 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7914 a signed addition of the two arguments, and indicate whether an overflow
7915 occurred during the signed summation.
7920 The arguments (%a and %b) and the first element of the result structure
7921 may be of integer types of any bit width, but they must have the same
7922 bit width. The second element of the result structure must be of type
7923 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7929 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7930 a signed addition of the two variables. They return a structure --- the
7931 first element of which is the signed summation, and the second element
7932 of which is a bit specifying if the signed summation resulted in an
7938 .. code-block:: llvm
7940 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7941 %sum = extractvalue {i32, i1} %res, 0
7942 %obit = extractvalue {i32, i1} %res, 1
7943 br i1 %obit, label %overflow, label %normal
7945 '``llvm.uadd.with.overflow.*``' Intrinsics
7946 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7951 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
7952 on any integer bit width.
7956 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
7957 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7958 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
7963 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7964 an unsigned addition of the two arguments, and indicate whether a carry
7965 occurred during the unsigned summation.
7970 The arguments (%a and %b) and the first element of the result structure
7971 may be of integer types of any bit width, but they must have the same
7972 bit width. The second element of the result structure must be of type
7973 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
7979 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7980 an unsigned addition of the two arguments. They return a structure --- the
7981 first element of which is the sum, and the second element of which is a
7982 bit specifying if the unsigned summation resulted in a carry.
7987 .. code-block:: llvm
7989 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7990 %sum = extractvalue {i32, i1} %res, 0
7991 %obit = extractvalue {i32, i1} %res, 1
7992 br i1 %obit, label %carry, label %normal
7994 '``llvm.ssub.with.overflow.*``' Intrinsics
7995 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8000 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
8001 on any integer bit width.
8005 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
8006 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8007 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
8012 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8013 a signed subtraction of the two arguments, and indicate whether an
8014 overflow occurred during the signed subtraction.
8019 The arguments (%a and %b) and the first element of the result structure
8020 may be of integer types of any bit width, but they must have the same
8021 bit width. The second element of the result structure must be of type
8022 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8028 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8029 a signed subtraction of the two arguments. They return a structure --- the
8030 first element of which is the subtraction, and the second element of
8031 which is a bit specifying if the signed subtraction resulted in an
8037 .. code-block:: llvm
8039 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8040 %sum = extractvalue {i32, i1} %res, 0
8041 %obit = extractvalue {i32, i1} %res, 1
8042 br i1 %obit, label %overflow, label %normal
8044 '``llvm.usub.with.overflow.*``' Intrinsics
8045 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8050 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
8051 on any integer bit width.
8055 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
8056 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8057 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
8062 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8063 an unsigned subtraction of the two arguments, and indicate whether an
8064 overflow occurred during the unsigned subtraction.
8069 The arguments (%a and %b) and the first element of the result structure
8070 may be of integer types of any bit width, but they must have the same
8071 bit width. The second element of the result structure must be of type
8072 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8078 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8079 an unsigned subtraction of the two arguments. They return a structure ---
8080 the first element of which is the subtraction, and the second element of
8081 which is a bit specifying if the unsigned subtraction resulted in an
8087 .. code-block:: llvm
8089 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8090 %sum = extractvalue {i32, i1} %res, 0
8091 %obit = extractvalue {i32, i1} %res, 1
8092 br i1 %obit, label %overflow, label %normal
8094 '``llvm.smul.with.overflow.*``' Intrinsics
8095 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8100 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
8101 on any integer bit width.
8105 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
8106 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8107 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
8112 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8113 a signed multiplication of the two arguments, and indicate whether an
8114 overflow occurred during the signed multiplication.
8119 The arguments (%a and %b) and the first element of the result structure
8120 may be of integer types of any bit width, but they must have the same
8121 bit width. The second element of the result structure must be of type
8122 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8128 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8129 a signed multiplication of the two arguments. They return a structure ---
8130 the first element of which is the multiplication, and the second element
8131 of which is a bit specifying if the signed multiplication resulted in an
8137 .. code-block:: llvm
8139 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8140 %sum = extractvalue {i32, i1} %res, 0
8141 %obit = extractvalue {i32, i1} %res, 1
8142 br i1 %obit, label %overflow, label %normal
8144 '``llvm.umul.with.overflow.*``' Intrinsics
8145 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8150 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
8151 on any integer bit width.
8155 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
8156 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8157 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
8162 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8163 a unsigned multiplication of the two arguments, and indicate whether an
8164 overflow occurred during the unsigned multiplication.
8169 The arguments (%a and %b) and the first element of the result structure
8170 may be of integer types of any bit width, but they must have the same
8171 bit width. The second element of the result structure must be of type
8172 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8178 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8179 an unsigned multiplication of the two arguments. They return a structure ---
8180 the first element of which is the multiplication, and the second
8181 element of which is a bit specifying if the unsigned multiplication
8182 resulted in an overflow.
8187 .. code-block:: llvm
8189 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8190 %sum = extractvalue {i32, i1} %res, 0
8191 %obit = extractvalue {i32, i1} %res, 1
8192 br i1 %obit, label %overflow, label %normal
8194 Specialised Arithmetic Intrinsics
8195 ---------------------------------
8197 '``llvm.fmuladd.*``' Intrinsic
8198 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8205 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
8206 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
8211 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
8212 expressions that can be fused if the code generator determines that (a) the
8213 target instruction set has support for a fused operation, and (b) that the
8214 fused operation is more efficient than the equivalent, separate pair of mul
8215 and add instructions.
8220 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
8221 multiplicands, a and b, and an addend c.
8230 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
8232 is equivalent to the expression a \* b + c, except that rounding will
8233 not be performed between the multiplication and addition steps if the
8234 code generator fuses the operations. Fusion is not guaranteed, even if
8235 the target platform supports it. If a fused multiply-add is required the
8236 corresponding llvm.fma.\* intrinsic function should be used instead.
8241 .. code-block:: llvm
8243 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields {float}:r2 = (a * b) + c
8245 Half Precision Floating Point Intrinsics
8246 ----------------------------------------
8248 For most target platforms, half precision floating point is a
8249 storage-only format. This means that it is a dense encoding (in memory)
8250 but does not support computation in the format.
8252 This means that code must first load the half-precision floating point
8253 value as an i16, then convert it to float with
8254 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
8255 then be performed on the float value (including extending to double
8256 etc). To store the value back to memory, it is first converted to float
8257 if needed, then converted to i16 with
8258 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
8261 .. _int_convert_to_fp16:
8263 '``llvm.convert.to.fp16``' Intrinsic
8264 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8271 declare i16 @llvm.convert.to.fp16(f32 %a)
8276 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8277 from single precision floating point format to half precision floating
8283 The intrinsic function contains single argument - the value to be
8289 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8290 from single precision floating point format to half precision floating
8291 point format. The return value is an ``i16`` which contains the
8297 .. code-block:: llvm
8299 %res = call i16 @llvm.convert.to.fp16(f32 %a)
8300 store i16 %res, i16* @x, align 2
8302 .. _int_convert_from_fp16:
8304 '``llvm.convert.from.fp16``' Intrinsic
8305 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8312 declare f32 @llvm.convert.from.fp16(i16 %a)
8317 The '``llvm.convert.from.fp16``' intrinsic function performs a
8318 conversion from half precision floating point format to single precision
8319 floating point format.
8324 The intrinsic function contains single argument - the value to be
8330 The '``llvm.convert.from.fp16``' intrinsic function performs a
8331 conversion from half single precision floating point format to single
8332 precision floating point format. The input half-float value is
8333 represented by an ``i16`` value.
8338 .. code-block:: llvm
8340 %a = load i16* @x, align 2
8341 %res = call f32 @llvm.convert.from.fp16(i16 %a)
8346 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
8347 prefix), are described in the `LLVM Source Level
8348 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
8351 Exception Handling Intrinsics
8352 -----------------------------
8354 The LLVM exception handling intrinsics (which all start with
8355 ``llvm.eh.`` prefix), are described in the `LLVM Exception
8356 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
8360 Trampoline Intrinsics
8361 ---------------------
8363 These intrinsics make it possible to excise one parameter, marked with
8364 the :ref:`nest <nest>` attribute, from a function. The result is a
8365 callable function pointer lacking the nest parameter - the caller does
8366 not need to provide a value for it. Instead, the value to use is stored
8367 in advance in a "trampoline", a block of memory usually allocated on the
8368 stack, which also contains code to splice the nest value into the
8369 argument list. This is used to implement the GCC nested function address
8372 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
8373 then the resulting function pointer has signature ``i32 (i32, i32)*``.
8374 It can be created as follows:
8376 .. code-block:: llvm
8378 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
8379 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
8380 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
8381 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
8382 %fp = bitcast i8* %p to i32 (i32, i32)*
8384 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
8385 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
8389 '``llvm.init.trampoline``' Intrinsic
8390 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8397 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
8402 This fills the memory pointed to by ``tramp`` with executable code,
8403 turning it into a trampoline.
8408 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
8409 pointers. The ``tramp`` argument must point to a sufficiently large and
8410 sufficiently aligned block of memory; this memory is written to by the
8411 intrinsic. Note that the size and the alignment are target-specific -
8412 LLVM currently provides no portable way of determining them, so a
8413 front-end that generates this intrinsic needs to have some
8414 target-specific knowledge. The ``func`` argument must hold a function
8415 bitcast to an ``i8*``.
8420 The block of memory pointed to by ``tramp`` is filled with target
8421 dependent code, turning it into a function. Then ``tramp`` needs to be
8422 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
8423 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
8424 function's signature is the same as that of ``func`` with any arguments
8425 marked with the ``nest`` attribute removed. At most one such ``nest``
8426 argument is allowed, and it must be of pointer type. Calling the new
8427 function is equivalent to calling ``func`` with the same argument list,
8428 but with ``nval`` used for the missing ``nest`` argument. If, after
8429 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
8430 modified, then the effect of any later call to the returned function
8431 pointer is undefined.
8435 '``llvm.adjust.trampoline``' Intrinsic
8436 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8443 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
8448 This performs any required machine-specific adjustment to the address of
8449 a trampoline (passed as ``tramp``).
8454 ``tramp`` must point to a block of memory which already has trampoline
8455 code filled in by a previous call to
8456 :ref:`llvm.init.trampoline <int_it>`.
8461 On some architectures the address of the code to be executed needs to be
8462 different to the address where the trampoline is actually stored. This
8463 intrinsic returns the executable address corresponding to ``tramp``
8464 after performing the required machine specific adjustments. The pointer
8465 returned can then be :ref:`bitcast and executed <int_trampoline>`.
8470 This class of intrinsics exists to information about the lifetime of
8471 memory objects and ranges where variables are immutable.
8475 '``llvm.lifetime.start``' Intrinsic
8476 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8483 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8488 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
8494 The first argument is a constant integer representing the size of the
8495 object, or -1 if it is variable sized. The second argument is a pointer
8501 This intrinsic indicates that before this point in the code, the value
8502 of the memory pointed to by ``ptr`` is dead. This means that it is known
8503 to never be used and has an undefined value. A load from the pointer
8504 that precedes this intrinsic can be replaced with ``'undef'``.
8508 '``llvm.lifetime.end``' Intrinsic
8509 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8516 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8521 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
8527 The first argument is a constant integer representing the size of the
8528 object, or -1 if it is variable sized. The second argument is a pointer
8534 This intrinsic indicates that after this point in the code, the value of
8535 the memory pointed to by ``ptr`` is dead. This means that it is known to
8536 never be used and has an undefined value. Any stores into the memory
8537 object following this intrinsic may be removed as dead.
8539 '``llvm.invariant.start``' Intrinsic
8540 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8547 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8552 The '``llvm.invariant.start``' intrinsic specifies that the contents of
8553 a memory object will not change.
8558 The first argument is a constant integer representing the size of the
8559 object, or -1 if it is variable sized. The second argument is a pointer
8565 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
8566 the return value, the referenced memory location is constant and
8569 '``llvm.invariant.end``' Intrinsic
8570 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8577 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8582 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
8583 memory object are mutable.
8588 The first argument is the matching ``llvm.invariant.start`` intrinsic.
8589 The second argument is a constant integer representing the size of the
8590 object, or -1 if it is variable sized and the third argument is a
8591 pointer to the object.
8596 This intrinsic indicates that the memory is mutable again.
8601 This class of intrinsics is designed to be generic and has no specific
8604 '``llvm.var.annotation``' Intrinsic
8605 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8612 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8617 The '``llvm.var.annotation``' intrinsic.
8622 The first argument is a pointer to a value, the second is a pointer to a
8623 global string, the third is a pointer to a global string which is the
8624 source file name, and the last argument is the line number.
8629 This intrinsic allows annotation of local variables with arbitrary
8630 strings. This can be useful for special purpose optimizations that want
8631 to look for these annotations. These have no other defined use; they are
8632 ignored by code generation and optimization.
8634 '``llvm.ptr.annotation.*``' Intrinsic
8635 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8640 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
8641 pointer to an integer of any width. *NOTE* you must specify an address space for
8642 the pointer. The identifier for the default address space is the integer
8647 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8648 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
8649 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
8650 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
8651 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
8656 The '``llvm.ptr.annotation``' intrinsic.
8661 The first argument is a pointer to an integer value of arbitrary bitwidth
8662 (result of some expression), the second is a pointer to a global string, the
8663 third is a pointer to a global string which is the source file name, and the
8664 last argument is the line number. It returns the value of the first argument.
8669 This intrinsic allows annotation of a pointer to an integer with arbitrary
8670 strings. This can be useful for special purpose optimizations that want to look
8671 for these annotations. These have no other defined use; they are ignored by code
8672 generation and optimization.
8674 '``llvm.annotation.*``' Intrinsic
8675 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8680 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
8681 any integer bit width.
8685 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
8686 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
8687 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
8688 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
8689 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
8694 The '``llvm.annotation``' intrinsic.
8699 The first argument is an integer value (result of some expression), the
8700 second is a pointer to a global string, the third is a pointer to a
8701 global string which is the source file name, and the last argument is
8702 the line number. It returns the value of the first argument.
8707 This intrinsic allows annotations to be put on arbitrary expressions
8708 with arbitrary strings. This can be useful for special purpose
8709 optimizations that want to look for these annotations. These have no
8710 other defined use; they are ignored by code generation and optimization.
8712 '``llvm.trap``' Intrinsic
8713 ^^^^^^^^^^^^^^^^^^^^^^^^^
8720 declare void @llvm.trap() noreturn nounwind
8725 The '``llvm.trap``' intrinsic.
8735 This intrinsic is lowered to the target dependent trap instruction. If
8736 the target does not have a trap instruction, this intrinsic will be
8737 lowered to a call of the ``abort()`` function.
8739 '``llvm.debugtrap``' Intrinsic
8740 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8747 declare void @llvm.debugtrap() nounwind
8752 The '``llvm.debugtrap``' intrinsic.
8762 This intrinsic is lowered to code which is intended to cause an
8763 execution trap with the intention of requesting the attention of a
8766 '``llvm.stackprotector``' Intrinsic
8767 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8774 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
8779 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
8780 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
8781 is placed on the stack before local variables.
8786 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
8787 The first argument is the value loaded from the stack guard
8788 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
8789 enough space to hold the value of the guard.
8794 This intrinsic causes the prologue/epilogue inserter to force the position of
8795 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
8796 to ensure that if a local variable on the stack is overwritten, it will destroy
8797 the value of the guard. When the function exits, the guard on the stack is
8798 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
8799 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
8800 calling the ``__stack_chk_fail()`` function.
8802 '``llvm.stackprotectorcheck``' Intrinsic
8803 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8810 declare void @llvm.stackprotectorcheck(i8** <guard>)
8815 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
8816 created stack protector and if they are not equal calls the
8817 ``__stack_chk_fail()`` function.
8822 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
8823 the variable ``@__stack_chk_guard``.
8828 This intrinsic is provided to perform the stack protector check by comparing
8829 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
8830 values do not match call the ``__stack_chk_fail()`` function.
8832 The reason to provide this as an IR level intrinsic instead of implementing it
8833 via other IR operations is that in order to perform this operation at the IR
8834 level without an intrinsic, one would need to create additional basic blocks to
8835 handle the success/failure cases. This makes it difficult to stop the stack
8836 protector check from disrupting sibling tail calls in Codegen. With this
8837 intrinsic, we are able to generate the stack protector basic blocks late in
8838 codegen after the tail call decision has occurred.
8840 '``llvm.objectsize``' Intrinsic
8841 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8848 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
8849 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
8854 The ``llvm.objectsize`` intrinsic is designed to provide information to
8855 the optimizers to determine at compile time whether a) an operation
8856 (like memcpy) will overflow a buffer that corresponds to an object, or
8857 b) that a runtime check for overflow isn't necessary. An object in this
8858 context means an allocation of a specific class, structure, array, or
8864 The ``llvm.objectsize`` intrinsic takes two arguments. The first
8865 argument is a pointer to or into the ``object``. The second argument is
8866 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
8867 or -1 (if false) when the object size is unknown. The second argument
8868 only accepts constants.
8873 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
8874 the size of the object concerned. If the size cannot be determined at
8875 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
8876 on the ``min`` argument).
8878 '``llvm.expect``' Intrinsic
8879 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8886 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
8887 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
8892 The ``llvm.expect`` intrinsic provides information about expected (the
8893 most probable) value of ``val``, which can be used by optimizers.
8898 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
8899 a value. The second argument is an expected value, this needs to be a
8900 constant value, variables are not allowed.
8905 This intrinsic is lowered to the ``val``.
8907 '``llvm.donothing``' Intrinsic
8908 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8915 declare void @llvm.donothing() nounwind readnone
8920 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
8921 only intrinsic that can be called with an invoke instruction.
8931 This intrinsic does nothing, and it's removed by optimizers and ignored
8934 Stack Map Intrinsics
8935 --------------------
8937 LLVM provides experimental intrinsics to support runtime patching
8938 mechanisms commonly desired in dynamic language JITs. These intrinsics
8939 are described in :doc:`StackMaps`.