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).
133 It also shows a convention that we follow in this document. When
134 demonstrating instructions, we will follow an instruction with a comment
135 that defines the type and name of value produced.
143 LLVM programs are composed of ``Module``'s, each of which is a
144 translation unit of the input programs. Each module consists of
145 functions, global variables, and symbol table entries. Modules may be
146 combined together with the LLVM linker, which merges function (and
147 global variable) definitions, resolves forward declarations, and merges
148 symbol table entries. Here is an example of the "hello world" module:
152 ; Declare the string constant as a global constant.
153 @.str = private unnamed_addr constant [13 x i8] c"hello world\0A\00"
155 ; External declaration of the puts function
156 declare i32 @puts(i8* nocapture) nounwind
158 ; Definition of main function
159 define i32 @main() { ; i32()*
160 ; Convert [13 x i8]* to i8 *...
161 %cast210 = getelementptr [13 x i8]* @.str, i64 0, i64 0
163 ; Call puts function to write out the string to stdout.
164 call i32 @puts(i8* %cast210)
169 !1 = metadata !{i32 42}
172 This example is made up of a :ref:`global variable <globalvars>` named
173 "``.str``", an external declaration of the "``puts``" function, a
174 :ref:`function definition <functionstructure>` for "``main``" and
175 :ref:`named metadata <namedmetadatastructure>` "``foo``".
177 In general, a module is made up of a list of global values (where both
178 functions and global variables are global values). Global values are
179 represented by a pointer to a memory location (in this case, a pointer
180 to an array of char, and a pointer to a function), and have one of the
181 following :ref:`linkage types <linkage>`.
188 All Global Variables and Functions have one of the following types of
192 Global values with "``private``" linkage are only directly
193 accessible by objects in the current module. In particular, linking
194 code into a module with an private global value may cause the
195 private to be renamed as necessary to avoid collisions. Because the
196 symbol is private to the module, all references can be updated. This
197 doesn't show up in any symbol table in the object file.
199 Similar to ``private``, but the symbol is passed through the
200 assembler and evaluated by the linker. Unlike normal strong symbols,
201 they are removed by the linker from the final linked image
202 (executable or dynamic library).
203 ``linker_private_weak``
204 Similar to "``linker_private``", but the symbol is weak. Note that
205 ``linker_private_weak`` symbols are subject to coalescing by the
206 linker. The symbols are removed by the linker from the final linked
207 image (executable or dynamic library).
209 Similar to private, but the value shows as a local symbol
210 (``STB_LOCAL`` in the case of ELF) in the object file. This
211 corresponds to the notion of the '``static``' keyword in C.
212 ``available_externally``
213 Globals with "``available_externally``" linkage are never emitted
214 into the object file corresponding to the LLVM module. They exist to
215 allow inlining and other optimizations to take place given knowledge
216 of the definition of the global, which is known to be somewhere
217 outside the module. Globals with ``available_externally`` linkage
218 are allowed to be discarded at will, and are otherwise the same as
219 ``linkonce_odr``. This linkage type is only allowed on definitions,
222 Globals with "``linkonce``" linkage are merged with other globals of
223 the same name when linkage occurs. This can be used to implement
224 some forms of inline functions, templates, or other code which must
225 be generated in each translation unit that uses it, but where the
226 body may be overridden with a more definitive definition later.
227 Unreferenced ``linkonce`` globals are allowed to be discarded. Note
228 that ``linkonce`` linkage does not actually allow the optimizer to
229 inline the body of this function into callers because it doesn't
230 know if this definition of the function is the definitive definition
231 within the program or whether it will be overridden by a stronger
232 definition. To enable inlining and other optimizations, use
233 "``linkonce_odr``" linkage.
235 "``weak``" linkage has the same merging semantics as ``linkonce``
236 linkage, except that unreferenced globals with ``weak`` linkage may
237 not be discarded. This is used for globals that are declared "weak"
240 "``common``" linkage is most similar to "``weak``" linkage, but they
241 are used for tentative definitions in C, such as "``int X;``" at
242 global scope. Symbols with "``common``" linkage are merged in the
243 same way as ``weak symbols``, and they may not be deleted if
244 unreferenced. ``common`` symbols may not have an explicit section,
245 must have a zero initializer, and may not be marked
246 ':ref:`constant <globalvars>`'. Functions and aliases may not have
249 .. _linkage_appending:
252 "``appending``" linkage may only be applied to global variables of
253 pointer to array type. When two global variables with appending
254 linkage are linked together, the two global arrays are appended
255 together. This is the LLVM, typesafe, equivalent of having the
256 system linker append together "sections" with identical names when
259 The semantics of this linkage follow the ELF object file model: the
260 symbol is weak until linked, if not linked, the symbol becomes null
261 instead of being an undefined reference.
262 ``linkonce_odr``, ``weak_odr``
263 Some languages allow differing globals to be merged, such as two
264 functions with different semantics. Other languages, such as
265 ``C++``, ensure that only equivalent globals are ever merged (the
266 "one definition rule" --- "ODR"). Such languages can use the
267 ``linkonce_odr`` and ``weak_odr`` linkage types to indicate that the
268 global will only be merged with equivalent globals. These linkage
269 types are otherwise the same as their non-``odr`` versions.
271 If none of the above identifiers are used, the global is externally
272 visible, meaning that it participates in linkage and can be used to
273 resolve external symbol references.
275 The next two types of linkage are targeted for Microsoft Windows
276 platform only. They are designed to support importing (exporting)
277 symbols from (to) DLLs (Dynamic Link Libraries).
280 "``dllimport``" linkage causes the compiler to reference a function
281 or variable via a global pointer to a pointer that is set up by the
282 DLL exporting the symbol. On Microsoft Windows targets, the pointer
283 name is formed by combining ``__imp_`` and the function or variable
286 "``dllexport``" linkage causes the compiler to provide a global
287 pointer to a pointer in a DLL, so that it can be referenced with the
288 ``dllimport`` attribute. On Microsoft Windows targets, the pointer
289 name is formed by combining ``__imp_`` and the function or variable
292 For example, since the "``.LC0``" variable is defined to be internal, if
293 another module defined a "``.LC0``" variable and was linked with this
294 one, one of the two would be renamed, preventing a collision. Since
295 "``main``" and "``puts``" are external (i.e., lacking any linkage
296 declarations), they are accessible outside of the current module.
298 It is illegal for a function *declaration* to have any linkage type
299 other than ``external``, ``dllimport`` or ``extern_weak``.
306 LLVM :ref:`functions <functionstructure>`, :ref:`calls <i_call>` and
307 :ref:`invokes <i_invoke>` can all have an optional calling convention
308 specified for the call. The calling convention of any pair of dynamic
309 caller/callee must match, or the behavior of the program is undefined.
310 The following calling conventions are supported by LLVM, and more may be
313 "``ccc``" - The C calling convention
314 This calling convention (the default if no other calling convention
315 is specified) matches the target C calling conventions. This calling
316 convention supports varargs function calls and tolerates some
317 mismatch in the declared prototype and implemented declaration of
318 the function (as does normal C).
319 "``fastcc``" - The fast calling convention
320 This calling convention attempts to make calls as fast as possible
321 (e.g. by passing things in registers). This calling convention
322 allows the target to use whatever tricks it wants to produce fast
323 code for the target, without having to conform to an externally
324 specified ABI (Application Binary Interface). `Tail calls can only
325 be optimized when this, the GHC or the HiPE convention is
326 used. <CodeGenerator.html#id80>`_ This calling convention does not
327 support varargs and requires the prototype of all callees to exactly
328 match the prototype of the function definition.
329 "``coldcc``" - The cold calling convention
330 This calling convention attempts to make code in the caller as
331 efficient as possible under the assumption that the call is not
332 commonly executed. As such, these calls often preserve all registers
333 so that the call does not break any live ranges in the caller side.
334 This calling convention does not support varargs and requires the
335 prototype of all callees to exactly match the prototype of the
337 "``cc 10``" - GHC convention
338 This calling convention has been implemented specifically for use by
339 the `Glasgow Haskell Compiler (GHC) <http://www.haskell.org/ghc>`_.
340 It passes everything in registers, going to extremes to achieve this
341 by disabling callee save registers. This calling convention should
342 not be used lightly but only for specific situations such as an
343 alternative to the *register pinning* performance technique often
344 used when implementing functional programming languages. At the
345 moment only X86 supports this convention and it has the following
348 - On *X86-32* only supports up to 4 bit type parameters. No
349 floating point types are supported.
350 - On *X86-64* only supports up to 10 bit type parameters and 6
351 floating point parameters.
353 This calling convention supports `tail call
354 optimization <CodeGenerator.html#id80>`_ but requires both the
355 caller and callee are using it.
356 "``cc 11``" - The HiPE calling convention
357 This calling convention has been implemented specifically for use by
358 the `High-Performance Erlang
359 (HiPE) <http://www.it.uu.se/research/group/hipe/>`_ compiler, *the*
360 native code compiler of the `Ericsson's Open Source Erlang/OTP
361 system <http://www.erlang.org/download.shtml>`_. It uses more
362 registers for argument passing than the ordinary C calling
363 convention and defines no callee-saved registers. The calling
364 convention properly supports `tail call
365 optimization <CodeGenerator.html#id80>`_ but requires that both the
366 caller and the callee use it. It uses a *register pinning*
367 mechanism, similar to GHC's convention, for keeping frequently
368 accessed runtime components pinned to specific hardware registers.
369 At the moment only X86 supports this convention (both 32 and 64
371 "``cc <n>``" - Numbered convention
372 Any calling convention may be specified by number, allowing
373 target-specific calling conventions to be used. Target specific
374 calling conventions start at 64.
376 More calling conventions can be added/defined on an as-needed basis, to
377 support Pascal conventions or any other well-known target-independent
380 .. _visibilitystyles:
385 All Global Variables and Functions have one of the following visibility
388 "``default``" - Default style
389 On targets that use the ELF object file format, default visibility
390 means that the declaration is visible to other modules and, in
391 shared libraries, means that the declared entity may be overridden.
392 On Darwin, default visibility means that the declaration is visible
393 to other modules. Default visibility corresponds to "external
394 linkage" in the language.
395 "``hidden``" - Hidden style
396 Two declarations of an object with hidden visibility refer to the
397 same object if they are in the same shared object. Usually, hidden
398 visibility indicates that the symbol will not be placed into the
399 dynamic symbol table, so no other module (executable or shared
400 library) can reference it directly.
401 "``protected``" - Protected style
402 On ELF, protected visibility indicates that the symbol will be
403 placed in the dynamic symbol table, but that references within the
404 defining module will bind to the local symbol. That is, the symbol
405 cannot be overridden by another module.
412 LLVM IR allows you to specify name aliases for certain types. This can
413 make it easier to read the IR and make the IR more condensed
414 (particularly when recursive types are involved). An example of a name
419 %mytype = type { %mytype*, i32 }
421 You may give a name to any :ref:`type <typesystem>` except
422 ":ref:`void <t_void>`". Type name aliases may be used anywhere a type is
423 expected with the syntax "%mytype".
425 Note that type names are aliases for the structural type that they
426 indicate, and that you can therefore specify multiple names for the same
427 type. This often leads to confusing behavior when dumping out a .ll
428 file. Since LLVM IR uses structural typing, the name is not part of the
429 type. When printing out LLVM IR, the printer will pick *one name* to
430 render all types of a particular shape. This means that if you have code
431 where two different source types end up having the same LLVM type, that
432 the dumper will sometimes print the "wrong" or unexpected type. This is
433 an important design point and isn't going to change.
440 Global variables define regions of memory allocated at compilation time
443 Global variables definitions must be initialized, may have an explicit section
444 to be placed in, and may have an optional explicit alignment specified.
446 Global variables in other translation units can also be declared, in which
447 case they don't have an initializer.
449 A variable may be defined as ``thread_local``, which means that it will
450 not be shared by threads (each thread will have a separated copy of the
451 variable). Not all targets support thread-local variables. Optionally, a
452 TLS model may be specified:
455 For variables that are only used within the current shared library.
457 For variables in modules that will not be loaded dynamically.
459 For variables defined in the executable and only used within it.
461 The models correspond to the ELF TLS models; see `ELF Handling For
462 Thread-Local Storage <http://people.redhat.com/drepper/tls.pdf>`_ for
463 more information on under which circumstances the different models may
464 be used. The target may choose a different TLS model if the specified
465 model is not supported, or if a better choice of model can be made.
467 A variable may be defined as a global ``constant``, which indicates that
468 the contents of the variable will **never** be modified (enabling better
469 optimization, allowing the global data to be placed in the read-only
470 section of an executable, etc). Note that variables that need runtime
471 initialization cannot be marked ``constant`` as there is a store to the
474 LLVM explicitly allows *declarations* of global variables to be marked
475 constant, even if the final definition of the global is not. This
476 capability can be used to enable slightly better optimization of the
477 program, but requires the language definition to guarantee that
478 optimizations based on the 'constantness' are valid for the translation
479 units that do not include the definition.
481 As SSA values, global variables define pointer values that are in scope
482 (i.e. they dominate) all basic blocks in the program. Global variables
483 always define a pointer to their "content" type because they describe a
484 region of memory, and all memory objects in LLVM are accessed through
487 Global variables can be marked with ``unnamed_addr`` which indicates
488 that the address is not significant, only the content. Constants marked
489 like this can be merged with other constants if they have the same
490 initializer. Note that a constant with significant address *can* be
491 merged with a ``unnamed_addr`` constant, the result being a constant
492 whose address is significant.
494 A global variable may be declared to reside in a target-specific
495 numbered address space. For targets that support them, address spaces
496 may affect how optimizations are performed and/or what target
497 instructions are used to access the variable. The default address space
498 is zero. The address space qualifier must precede any other attributes.
500 LLVM allows an explicit section to be specified for globals. If the
501 target supports it, it will emit globals to the section specified.
503 By default, global initializers are optimized by assuming that global
504 variables defined within the module are not modified from their
505 initial values before the start of the global initializer. This is
506 true even for variables potentially accessible from outside the
507 module, including those with external linkage or appearing in
508 ``@llvm.used``. This assumption may be suppressed by marking the
509 variable with ``externally_initialized``.
511 An explicit alignment may be specified for a global, which must be a
512 power of 2. If not present, or if the alignment is set to zero, the
513 alignment of the global is set by the target to whatever it feels
514 convenient. If an explicit alignment is specified, the global is forced
515 to have exactly that alignment. Targets and optimizers are not allowed
516 to over-align the global if the global has an assigned section. In this
517 case, the extra alignment could be observable: for example, code could
518 assume that the globals are densely packed in their section and try to
519 iterate over them as an array, alignment padding would break this
522 For example, the following defines a global in a numbered address space
523 with an initializer, section, and alignment:
527 @G = addrspace(5) constant float 1.0, section "foo", align 4
529 The following example just declares a global variable
533 @G = external global i32
535 The following example defines a thread-local global with the
536 ``initialexec`` TLS model:
540 @G = thread_local(initialexec) global i32 0, align 4
542 .. _functionstructure:
547 LLVM function definitions consist of the "``define``" keyword, an
548 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
549 style <visibility>`, an optional :ref:`calling convention <callingconv>`,
550 an optional ``unnamed_addr`` attribute, a return type, an optional
551 :ref:`parameter attribute <paramattrs>` for the return type, a function
552 name, a (possibly empty) argument list (each with optional :ref:`parameter
553 attributes <paramattrs>`), optional :ref:`function attributes <fnattrs>`,
554 an optional section, an optional alignment, an optional :ref:`garbage
555 collector name <gc>`, an optional :ref:`prefix <prefixdata>`, an opening
556 curly brace, a list of basic blocks, and a closing curly brace.
558 LLVM function declarations consist of the "``declare``" keyword, an
559 optional :ref:`linkage type <linkage>`, an optional :ref:`visibility
560 style <visibility>`, an optional :ref:`calling convention <callingconv>`,
561 an optional ``unnamed_addr`` attribute, a return type, an optional
562 :ref:`parameter attribute <paramattrs>` for the return type, a function
563 name, a possibly empty list of arguments, an optional alignment, an optional
564 :ref:`garbage collector name <gc>` and an optional :ref:`prefix <prefixdata>`.
566 A function definition contains a list of basic blocks, forming the CFG (Control
567 Flow Graph) for the function. Each basic block may optionally start with a label
568 (giving the basic block a symbol table entry), contains a list of instructions,
569 and ends with a :ref:`terminator <terminators>` instruction (such as a branch or
570 function return). If an explicit label is not provided, a block is assigned an
571 implicit numbered label, using the next value from the same counter as used for
572 unnamed temporaries (:ref:`see above<identifiers>`). For example, if a function
573 entry block does not have an explicit label, it will be assigned label "%0",
574 then the first unnamed temporary in that block will be "%1", etc.
576 The first basic block in a function is special in two ways: it is
577 immediately executed on entrance to the function, and it is not allowed
578 to have predecessor basic blocks (i.e. there can not be any branches to
579 the entry block of a function). Because the block can have no
580 predecessors, it also cannot have any :ref:`PHI nodes <i_phi>`.
582 LLVM allows an explicit section to be specified for functions. If the
583 target supports it, it will emit functions to the section specified.
585 An explicit alignment may be specified for a function. If not present,
586 or if the alignment is set to zero, the alignment of the function is set
587 by the target to whatever it feels convenient. If an explicit alignment
588 is specified, the function is forced to have at least that much
589 alignment. All alignments must be a power of 2.
591 If the ``unnamed_addr`` attribute is given, the address is know to not
592 be significant and two identical functions can be merged.
596 define [linkage] [visibility]
598 <ResultType> @<FunctionName> ([argument list])
599 [fn Attrs] [section "name"] [align N]
600 [gc] [prefix Constant] { ... }
607 Aliases act as "second name" for the aliasee value (which can be either
608 function, global variable, another alias or bitcast of global value).
609 Aliases may have an optional :ref:`linkage type <linkage>`, and an optional
610 :ref:`visibility style <visibility>`.
614 @<Name> = alias [Linkage] [Visibility] <AliaseeTy> @<Aliasee>
616 The linkage must be one of ``private``, ``linker_private``,
617 ``linker_private_weak``, ``internal``, ``linkonce``, ``weak``,
618 ``linkonce_odr``, ``weak_odr``, ``external``. Note that some system linkers
619 might not correctly handle dropping a weak symbol that is aliased by a non weak
622 .. _namedmetadatastructure:
627 Named metadata is a collection of metadata. :ref:`Metadata
628 nodes <metadata>` (but not metadata strings) are the only valid
629 operands for a named metadata.
633 ; Some unnamed metadata nodes, which are referenced by the named metadata.
634 !0 = metadata !{metadata !"zero"}
635 !1 = metadata !{metadata !"one"}
636 !2 = metadata !{metadata !"two"}
638 !name = !{!0, !1, !2}
645 The return type and each parameter of a function type may have a set of
646 *parameter attributes* associated with them. Parameter attributes are
647 used to communicate additional information about the result or
648 parameters of a function. Parameter attributes are considered to be part
649 of the function, not of the function type, so functions with different
650 parameter attributes can have the same function type.
652 Parameter attributes are simple keywords that follow the type specified.
653 If multiple parameter attributes are needed, they are space separated.
658 declare i32 @printf(i8* noalias nocapture, ...)
659 declare i32 @atoi(i8 zeroext)
660 declare signext i8 @returns_signed_char()
662 Note that any attributes for the function result (``nounwind``,
663 ``readonly``) come immediately after the argument list.
665 Currently, only the following parameter attributes are defined:
668 This indicates to the code generator that the parameter or return
669 value should be zero-extended to the extent required by the target's
670 ABI (which is usually 32-bits, but is 8-bits for a i1 on x86-64) by
671 the caller (for a parameter) or the callee (for a return value).
673 This indicates to the code generator that the parameter or return
674 value should be sign-extended to the extent required by the target's
675 ABI (which is usually 32-bits) by the caller (for a parameter) or
676 the callee (for a return value).
678 This indicates that this parameter or return value should be treated
679 in a special target-dependent fashion during while emitting code for
680 a function call or return (usually, by putting it in a register as
681 opposed to memory, though some targets use it to distinguish between
682 two different kinds of registers). Use of this attribute is
685 This indicates that the pointer parameter should really be passed by
686 value to the function. The attribute implies that a hidden copy of
687 the pointee is made between the caller and the callee, so the callee
688 is unable to modify the value in the caller. This attribute is only
689 valid on LLVM pointer arguments. It is generally used to pass
690 structs and arrays by value, but is also valid on pointers to
691 scalars. The copy is considered to belong to the caller not the
692 callee (for example, ``readonly`` functions should not write to
693 ``byval`` parameters). This is not a valid attribute for return
696 The byval attribute also supports specifying an alignment with the
697 align attribute. It indicates the alignment of the stack slot to
698 form and the known alignment of the pointer specified to the call
699 site. If the alignment is not specified, then the code generator
700 makes a target-specific assumption.
703 This indicates that the pointer parameter specifies the address of a
704 structure that is the return value of the function in the source
705 program. This pointer must be guaranteed by the caller to be valid:
706 loads and stores to the structure may be assumed by the callee
707 not to trap and to be properly aligned. This may only be applied to
708 the first parameter. This is not a valid attribute for return
711 This indicates that pointer values :ref:`based <pointeraliasing>` on
712 the argument or return value do not alias pointer values which are
713 not *based* on it, ignoring certain "irrelevant" dependencies. For a
714 call to the parent function, dependencies between memory references
715 from before or after the call and from those during the call are
716 "irrelevant" to the ``noalias`` keyword for the arguments and return
717 value used in that call. The caller shares the responsibility with
718 the callee for ensuring that these requirements are met. For further
719 details, please see the discussion of the NoAlias response in `alias
720 analysis <AliasAnalysis.html#MustMayNo>`_.
722 Note that this definition of ``noalias`` is intentionally similar
723 to the definition of ``restrict`` in C99 for function arguments,
724 though it is slightly weaker.
726 For function return values, C99's ``restrict`` is not meaningful,
727 while LLVM's ``noalias`` is.
729 This indicates that the callee does not make any copies of the
730 pointer that outlive the callee itself. This is not a valid
731 attribute for return values.
736 This indicates that the pointer parameter can be excised using the
737 :ref:`trampoline intrinsics <int_trampoline>`. This is not a valid
738 attribute for return values and can only be applied to one parameter.
741 This indicates that the function always returns the argument as its return
742 value. This is an optimization hint to the code generator when generating
743 the caller, allowing tail call optimization and omission of register saves
744 and restores in some cases; it is not checked or enforced when generating
745 the callee. The parameter and the function return type must be valid
746 operands for the :ref:`bitcast instruction <i_bitcast>`. This is not a
747 valid attribute for return values and can only be applied to one parameter.
751 Garbage Collector Names
752 -----------------------
754 Each function may specify a garbage collector name, which is simply a
759 define void @f() gc "name" { ... }
761 The compiler declares the supported values of *name*. Specifying a
762 collector which will cause the compiler to alter its output in order to
763 support the named garbage collection algorithm.
770 Prefix data is data associated with a function which the code generator
771 will emit immediately before the function body. The purpose of this feature
772 is to allow frontends to associate language-specific runtime metadata with
773 specific functions and make it available through the function pointer while
774 still allowing the function pointer to be called. To access the data for a
775 given function, a program may bitcast the function pointer to a pointer to
776 the constant's type. This implies that the IR symbol points to the start
779 To maintain the semantics of ordinary function calls, the prefix data must
780 have a particular format. Specifically, it must begin with a sequence of
781 bytes which decode to a sequence of machine instructions, valid for the
782 module's target, which transfer control to the point immediately succeeding
783 the prefix data, without performing any other visible action. This allows
784 the inliner and other passes to reason about the semantics of the function
785 definition without needing to reason about the prefix data. Obviously this
786 makes the format of the prefix data highly target dependent.
788 Prefix data is laid out as if it were an initializer for a global variable
789 of the prefix data's type. No padding is automatically placed between the
790 prefix data and the function body. If padding is required, it must be part
793 A trivial example of valid prefix data for the x86 architecture is ``i8 144``,
794 which encodes the ``nop`` instruction:
798 define void @f() prefix i8 144 { ... }
800 Generally prefix data can be formed by encoding a relative branch instruction
801 which skips the metadata, as in this example of valid prefix data for the
802 x86_64 architecture, where the first two bytes encode ``jmp .+10``:
806 %0 = type <{ i8, i8, i8* }>
808 define void @f() prefix %0 <{ i8 235, i8 8, i8* @md}> { ... }
810 A function may have prefix data but no body. This has similar semantics
811 to the ``available_externally`` linkage in that the data may be used by the
812 optimizers but will not be emitted in the object file.
819 Attribute groups are groups of attributes that are referenced by objects within
820 the IR. They are important for keeping ``.ll`` files readable, because a lot of
821 functions will use the same set of attributes. In the degenerative case of a
822 ``.ll`` file that corresponds to a single ``.c`` file, the single attribute
823 group will capture the important command line flags used to build that file.
825 An attribute group is a module-level object. To use an attribute group, an
826 object references the attribute group's ID (e.g. ``#37``). An object may refer
827 to more than one attribute group. In that situation, the attributes from the
828 different groups are merged.
830 Here is an example of attribute groups for a function that should always be
831 inlined, has a stack alignment of 4, and which shouldn't use SSE instructions:
835 ; Target-independent attributes:
836 attributes #0 = { alwaysinline alignstack=4 }
838 ; Target-dependent attributes:
839 attributes #1 = { "no-sse" }
841 ; Function @f has attributes: alwaysinline, alignstack=4, and "no-sse".
842 define void @f() #0 #1 { ... }
849 Function attributes are set to communicate additional information about
850 a function. Function attributes are considered to be part of the
851 function, not of the function type, so functions with different function
852 attributes can have the same function type.
854 Function attributes are simple keywords that follow the type specified.
855 If multiple attributes are needed, they are space separated. For
860 define void @f() noinline { ... }
861 define void @f() alwaysinline { ... }
862 define void @f() alwaysinline optsize { ... }
863 define void @f() optsize { ... }
866 This attribute indicates that, when emitting the prologue and
867 epilogue, the backend should forcibly align the stack pointer.
868 Specify the desired alignment, which must be a power of two, in
871 This attribute indicates that the inliner should attempt to inline
872 this function into callers whenever possible, ignoring any active
873 inlining size threshold for this caller.
875 This indicates that the callee function at a call site should be
876 recognized as a built-in function, even though the function's declaration
877 uses the ``nobuiltin`` attribute. This is only valid at call sites for
878 direct calls to functions which are declared with the ``nobuiltin``
881 This attribute indicates that this function is rarely called. When
882 computing edge weights, basic blocks post-dominated by a cold
883 function call are also considered to be cold; and, thus, given low
886 This attribute indicates that the source code contained a hint that
887 inlining this function is desirable (such as the "inline" keyword in
888 C/C++). It is just a hint; it imposes no requirements on the
891 This attribute suggests that optimization passes and code generator
892 passes make choices that keep the code size of this function as small
893 as possible and perform optimizations that may sacrifice runtime
894 performance in order to minimize the size of the generated code.
896 This attribute disables prologue / epilogue emission for the
897 function. This can have very system-specific consequences.
899 This indicates that the callee function at a call site is not recognized as
900 a built-in function. LLVM will retain the original call and not replace it
901 with equivalent code based on the semantics of the built-in function, unless
902 the call site uses the ``builtin`` attribute. This is valid at call sites
903 and on function declarations and definitions.
905 This attribute indicates that calls to the function cannot be
906 duplicated. A call to a ``noduplicate`` function may be moved
907 within its parent function, but may not be duplicated within
910 A function containing a ``noduplicate`` call may still
911 be an inlining candidate, provided that the call is not
912 duplicated by inlining. That implies that the function has
913 internal linkage and only has one call site, so the original
914 call is dead after inlining.
916 This attributes disables implicit floating point instructions.
918 This attribute indicates that the inliner should never inline this
919 function in any situation. This attribute may not be used together
920 with the ``alwaysinline`` attribute.
922 This attribute suppresses lazy symbol binding for the function. This
923 may make calls to the function faster, at the cost of extra program
924 startup time if the function is not called during program startup.
926 This attribute indicates that the code generator should not use a
927 red zone, even if the target-specific ABI normally permits it.
929 This function attribute indicates that the function never returns
930 normally. This produces undefined behavior at runtime if the
931 function ever does dynamically return.
933 This function attribute indicates that the function never returns
934 with an unwind or exceptional control flow. If the function does
935 unwind, its runtime behavior is undefined.
937 This function attribute indicates that the function is not optimized
938 by any optimization or code generator passes with the
939 exception of interprocedural optimization passes.
940 This attribute cannot be used together with the ``alwaysinline``
941 attribute; this attribute is also incompatible
942 with the ``minsize`` attribute and the ``optsize`` attribute.
944 The inliner should never inline this function in any situation.
945 Only functions with the ``alwaysinline`` attribute are valid
946 candidates for inlining inside the body of this function.
948 This attribute suggests that optimization passes and code generator
949 passes make choices that keep the code size of this function low,
950 and otherwise do optimizations specifically to reduce code size as
951 long as they do not significantly impact runtime performance.
953 On a function, this attribute indicates that the function computes its
954 result (or decides to unwind an exception) based strictly on its arguments,
955 without dereferencing any pointer arguments or otherwise accessing
956 any mutable state (e.g. memory, control registers, etc) visible to
957 caller functions. It does not write through any pointer arguments
958 (including ``byval`` arguments) and never changes any state visible
959 to callers. This means that it cannot unwind exceptions by calling
960 the ``C++`` exception throwing methods.
962 On an argument, this attribute indicates that the function does not
963 dereference that pointer argument, even though it may read or write the
964 memory that the pointer points to if accessed through other pointers.
966 On a function, this attribute indicates that the function does not write
967 through any pointer arguments (including ``byval`` arguments) or otherwise
968 modify any state (e.g. memory, control registers, etc) visible to
969 caller functions. It may dereference pointer arguments and read
970 state that may be set in the caller. A readonly function always
971 returns the same value (or unwinds an exception identically) when
972 called with the same set of arguments and global state. It cannot
973 unwind an exception by calling the ``C++`` exception throwing
976 On an argument, this attribute indicates that the function does not write
977 through this pointer argument, even though it may write to the memory that
978 the pointer points to.
980 This attribute indicates that this function can return twice. The C
981 ``setjmp`` is an example of such a function. The compiler disables
982 some optimizations (like tail calls) in the caller of these
985 This attribute indicates that AddressSanitizer checks
986 (dynamic address safety analysis) are enabled for this function.
988 This attribute indicates that MemorySanitizer checks (dynamic detection
989 of accesses to uninitialized memory) are enabled for this function.
991 This attribute indicates that ThreadSanitizer checks
992 (dynamic thread safety analysis) are enabled for this function.
994 This attribute indicates that the function should emit a stack
995 smashing protector. It is in the form of a "canary" --- a random value
996 placed on the stack before the local variables that's checked upon
997 return from the function to see if it has been overwritten. A
998 heuristic is used to determine if a function needs stack protectors
999 or not. The heuristic used will enable protectors for functions with:
1001 - Character arrays larger than ``ssp-buffer-size`` (default 8).
1002 - Aggregates containing character arrays larger than ``ssp-buffer-size``.
1003 - Calls to alloca() with variable sizes or constant sizes greater than
1004 ``ssp-buffer-size``.
1006 If a function that has an ``ssp`` attribute is inlined into a
1007 function that doesn't have an ``ssp`` attribute, then the resulting
1008 function will have an ``ssp`` attribute.
1010 This attribute indicates that the function should *always* emit a
1011 stack smashing protector. This overrides the ``ssp`` function
1014 If a function that has an ``sspreq`` attribute is inlined into a
1015 function that doesn't have an ``sspreq`` attribute or which has an
1016 ``ssp`` or ``sspstrong`` attribute, then the resulting function will have
1017 an ``sspreq`` attribute.
1019 This attribute indicates that the function should emit a stack smashing
1020 protector. This attribute causes a strong heuristic to be used when
1021 determining if a function needs stack protectors. The strong heuristic
1022 will enable protectors for functions with:
1024 - Arrays of any size and type
1025 - Aggregates containing an array of any size and type.
1026 - Calls to alloca().
1027 - Local variables that have had their address taken.
1029 This overrides the ``ssp`` function attribute.
1031 If a function that has an ``sspstrong`` attribute is inlined into a
1032 function that doesn't have an ``sspstrong`` attribute, then the
1033 resulting function will have an ``sspstrong`` attribute.
1035 This attribute indicates that the ABI being targeted requires that
1036 an unwind table entry be produce for this function even if we can
1037 show that no exceptions passes by it. This is normally the case for
1038 the ELF x86-64 abi, but it can be disabled for some compilation
1043 Module-Level Inline Assembly
1044 ----------------------------
1046 Modules may contain "module-level inline asm" blocks, which corresponds
1047 to the GCC "file scope inline asm" blocks. These blocks are internally
1048 concatenated by LLVM and treated as a single unit, but may be separated
1049 in the ``.ll`` file if desired. The syntax is very simple:
1051 .. code-block:: llvm
1053 module asm "inline asm code goes here"
1054 module asm "more can go here"
1056 The strings can contain any character by escaping non-printable
1057 characters. The escape sequence used is simply "\\xx" where "xx" is the
1058 two digit hex code for the number.
1060 The inline asm code is simply printed to the machine code .s file when
1061 assembly code is generated.
1063 .. _langref_datalayout:
1068 A module may specify a target specific data layout string that specifies
1069 how data is to be laid out in memory. The syntax for the data layout is
1072 .. code-block:: llvm
1074 target datalayout = "layout specification"
1076 The *layout specification* consists of a list of specifications
1077 separated by the minus sign character ('-'). Each specification starts
1078 with a letter and may include other information after the letter to
1079 define some aspect of the data layout. The specifications accepted are
1083 Specifies that the target lays out data in big-endian form. That is,
1084 the bits with the most significance have the lowest address
1087 Specifies that the target lays out data in little-endian form. That
1088 is, the bits with the least significance have the lowest address
1091 Specifies the natural alignment of the stack in bits. Alignment
1092 promotion of stack variables is limited to the natural stack
1093 alignment to avoid dynamic stack realignment. The stack alignment
1094 must be a multiple of 8-bits. If omitted, the natural stack
1095 alignment defaults to "unspecified", which does not prevent any
1096 alignment promotions.
1097 ``p[n]:<size>:<abi>:<pref>``
1098 This specifies the *size* of a pointer and its ``<abi>`` and
1099 ``<pref>``\erred alignments for address space ``n``. All sizes are in
1100 bits. Specifying the ``<pref>`` alignment is optional. If omitted, the
1101 preceding ``:`` should be omitted too. The address space, ``n`` is
1102 optional, and if not specified, denotes the default address space 0.
1103 The value of ``n`` must be in the range [1,2^23).
1104 ``i<size>:<abi>:<pref>``
1105 This specifies the alignment for an integer type of a given bit
1106 ``<size>``. The value of ``<size>`` must be in the range [1,2^23).
1107 ``v<size>:<abi>:<pref>``
1108 This specifies the alignment for a vector type of a given bit
1110 ``f<size>:<abi>:<pref>``
1111 This specifies the alignment for a floating point type of a given bit
1112 ``<size>``. Only values of ``<size>`` that are supported by the target
1113 will work. 32 (float) and 64 (double) are supported on all targets; 80
1114 or 128 (different flavors of long double) are also supported on some
1116 ``a<size>:<abi>:<pref>``
1117 This specifies the alignment for an aggregate type of a given bit
1119 ``s<size>:<abi>:<pref>``
1120 This specifies the alignment for a stack object of a given bit
1122 ``n<size1>:<size2>:<size3>...``
1123 This specifies a set of native integer widths for the target CPU in
1124 bits. For example, it might contain ``n32`` for 32-bit PowerPC,
1125 ``n32:64`` for PowerPC 64, or ``n8:16:32:64`` for X86-64. Elements of
1126 this set are considered to support most general arithmetic operations
1129 When constructing the data layout for a given target, LLVM starts with a
1130 default set of specifications which are then (possibly) overridden by
1131 the specifications in the ``datalayout`` keyword. The default
1132 specifications are given in this list:
1134 - ``E`` - big endian
1135 - ``p:64:64:64`` - 64-bit pointers with 64-bit alignment.
1136 - ``p[n]:64:64:64`` - Other address spaces are assumed to be the
1137 same as the default address space.
1138 - ``S0`` - natural stack alignment is unspecified
1139 - ``i1:8:8`` - i1 is 8-bit (byte) aligned
1140 - ``i8:8:8`` - i8 is 8-bit (byte) aligned
1141 - ``i16:16:16`` - i16 is 16-bit aligned
1142 - ``i32:32:32`` - i32 is 32-bit aligned
1143 - ``i64:32:64`` - i64 has ABI alignment of 32-bits but preferred
1144 alignment of 64-bits
1145 - ``f16:16:16`` - half is 16-bit aligned
1146 - ``f32:32:32`` - float is 32-bit aligned
1147 - ``f64:64:64`` - double is 64-bit aligned
1148 - ``f128:128:128`` - quad is 128-bit aligned
1149 - ``v64:64:64`` - 64-bit vector is 64-bit aligned
1150 - ``v128:128:128`` - 128-bit vector is 128-bit aligned
1151 - ``a0:0:64`` - aggregates are 64-bit aligned
1153 When LLVM is determining the alignment for a given type, it uses the
1156 #. If the type sought is an exact match for one of the specifications,
1157 that specification is used.
1158 #. If no match is found, and the type sought is an integer type, then
1159 the smallest integer type that is larger than the bitwidth of the
1160 sought type is used. If none of the specifications are larger than
1161 the bitwidth then the largest integer type is used. For example,
1162 given the default specifications above, the i7 type will use the
1163 alignment of i8 (next largest) while both i65 and i256 will use the
1164 alignment of i64 (largest specified).
1165 #. If no match is found, and the type sought is a vector type, then the
1166 largest vector type that is smaller than the sought vector type will
1167 be used as a fall back. This happens because <128 x double> can be
1168 implemented in terms of 64 <2 x double>, for example.
1170 The function of the data layout string may not be what you expect.
1171 Notably, this is not a specification from the frontend of what alignment
1172 the code generator should use.
1174 Instead, if specified, the target data layout is required to match what
1175 the ultimate *code generator* expects. This string is used by the
1176 mid-level optimizers to improve code, and this only works if it matches
1177 what the ultimate code generator uses. If you would like to generate IR
1178 that does not embed this target-specific detail into the IR, then you
1179 don't have to specify the string. This will disable some optimizations
1180 that require precise layout information, but this also prevents those
1181 optimizations from introducing target specificity into the IR.
1188 A module may specify a target triple string that describes the target
1189 host. The syntax for the target triple is simply:
1191 .. code-block:: llvm
1193 target triple = "x86_64-apple-macosx10.7.0"
1195 The *target triple* string consists of a series of identifiers delimited
1196 by the minus sign character ('-'). The canonical forms are:
1200 ARCHITECTURE-VENDOR-OPERATING_SYSTEM
1201 ARCHITECTURE-VENDOR-OPERATING_SYSTEM-ENVIRONMENT
1203 This information is passed along to the backend so that it generates
1204 code for the proper architecture. It's possible to override this on the
1205 command line with the ``-mtriple`` command line option.
1207 .. _pointeraliasing:
1209 Pointer Aliasing Rules
1210 ----------------------
1212 Any memory access must be done through a pointer value associated with
1213 an address range of the memory access, otherwise the behavior is
1214 undefined. Pointer values are associated with address ranges according
1215 to the following rules:
1217 - A pointer value is associated with the addresses associated with any
1218 value it is *based* on.
1219 - An address of a global variable is associated with the address range
1220 of the variable's storage.
1221 - The result value of an allocation instruction is associated with the
1222 address range of the allocated storage.
1223 - A null pointer in the default address-space is associated with no
1225 - An integer constant other than zero or a pointer value returned from
1226 a function not defined within LLVM may be associated with address
1227 ranges allocated through mechanisms other than those provided by
1228 LLVM. Such ranges shall not overlap with any ranges of addresses
1229 allocated by mechanisms provided by LLVM.
1231 A pointer value is *based* on another pointer value according to the
1234 - A pointer value formed from a ``getelementptr`` operation is *based*
1235 on the first operand of the ``getelementptr``.
1236 - The result value of a ``bitcast`` is *based* on the operand of the
1238 - A pointer value formed by an ``inttoptr`` is *based* on all pointer
1239 values that contribute (directly or indirectly) to the computation of
1240 the pointer's value.
1241 - The "*based* on" relationship is transitive.
1243 Note that this definition of *"based"* is intentionally similar to the
1244 definition of *"based"* in C99, though it is slightly weaker.
1246 LLVM IR does not associate types with memory. The result type of a
1247 ``load`` merely indicates the size and alignment of the memory from
1248 which to load, as well as the interpretation of the value. The first
1249 operand type of a ``store`` similarly only indicates the size and
1250 alignment of the store.
1252 Consequently, type-based alias analysis, aka TBAA, aka
1253 ``-fstrict-aliasing``, is not applicable to general unadorned LLVM IR.
1254 :ref:`Metadata <metadata>` may be used to encode additional information
1255 which specialized optimization passes may use to implement type-based
1260 Volatile Memory Accesses
1261 ------------------------
1263 Certain memory accesses, such as :ref:`load <i_load>`'s,
1264 :ref:`store <i_store>`'s, and :ref:`llvm.memcpy <int_memcpy>`'s may be
1265 marked ``volatile``. The optimizers must not change the number of
1266 volatile operations or change their order of execution relative to other
1267 volatile operations. The optimizers *may* change the order of volatile
1268 operations relative to non-volatile operations. This is not Java's
1269 "volatile" and has no cross-thread synchronization behavior.
1271 IR-level volatile loads and stores cannot safely be optimized into
1272 llvm.memcpy or llvm.memmove intrinsics even when those intrinsics are
1273 flagged volatile. Likewise, the backend should never split or merge
1274 target-legal volatile load/store instructions.
1276 .. admonition:: Rationale
1278 Platforms may rely on volatile loads and stores of natively supported
1279 data width to be executed as single instruction. For example, in C
1280 this holds for an l-value of volatile primitive type with native
1281 hardware support, but not necessarily for aggregate types. The
1282 frontend upholds these expectations, which are intentionally
1283 unspecified in the IR. The rules above ensure that IR transformation
1284 do not violate the frontend's contract with the language.
1288 Memory Model for Concurrent Operations
1289 --------------------------------------
1291 The LLVM IR does not define any way to start parallel threads of
1292 execution or to register signal handlers. Nonetheless, there are
1293 platform-specific ways to create them, and we define LLVM IR's behavior
1294 in their presence. This model is inspired by the C++0x memory model.
1296 For a more informal introduction to this model, see the :doc:`Atomics`.
1298 We define a *happens-before* partial order as the least partial order
1301 - Is a superset of single-thread program order, and
1302 - When a *synchronizes-with* ``b``, includes an edge from ``a`` to
1303 ``b``. *Synchronizes-with* pairs are introduced by platform-specific
1304 techniques, like pthread locks, thread creation, thread joining,
1305 etc., and by atomic instructions. (See also :ref:`Atomic Memory Ordering
1306 Constraints <ordering>`).
1308 Note that program order does not introduce *happens-before* edges
1309 between a thread and signals executing inside that thread.
1311 Every (defined) read operation (load instructions, memcpy, atomic
1312 loads/read-modify-writes, etc.) R reads a series of bytes written by
1313 (defined) write operations (store instructions, atomic
1314 stores/read-modify-writes, memcpy, etc.). For the purposes of this
1315 section, initialized globals are considered to have a write of the
1316 initializer which is atomic and happens before any other read or write
1317 of the memory in question. For each byte of a read R, R\ :sub:`byte`
1318 may see any write to the same byte, except:
1320 - If write\ :sub:`1` happens before write\ :sub:`2`, and
1321 write\ :sub:`2` happens before R\ :sub:`byte`, then
1322 R\ :sub:`byte` does not see write\ :sub:`1`.
1323 - If R\ :sub:`byte` happens before write\ :sub:`3`, then
1324 R\ :sub:`byte` does not see write\ :sub:`3`.
1326 Given that definition, R\ :sub:`byte` is defined as follows:
1328 - If R is volatile, the result is target-dependent. (Volatile is
1329 supposed to give guarantees which can support ``sig_atomic_t`` in
1330 C/C++, and may be used for accesses to addresses which do not behave
1331 like normal memory. It does not generally provide cross-thread
1333 - Otherwise, if there is no write to the same byte that happens before
1334 R\ :sub:`byte`, R\ :sub:`byte` returns ``undef`` for that byte.
1335 - Otherwise, if R\ :sub:`byte` may see exactly one write,
1336 R\ :sub:`byte` returns the value written by that write.
1337 - Otherwise, if R is atomic, and all the writes R\ :sub:`byte` may
1338 see are atomic, it chooses one of the values written. See the :ref:`Atomic
1339 Memory Ordering Constraints <ordering>` section for additional
1340 constraints on how the choice is made.
1341 - Otherwise R\ :sub:`byte` returns ``undef``.
1343 R returns the value composed of the series of bytes it read. This
1344 implies that some bytes within the value may be ``undef`` **without**
1345 the entire value being ``undef``. Note that this only defines the
1346 semantics of the operation; it doesn't mean that targets will emit more
1347 than one instruction to read the series of bytes.
1349 Note that in cases where none of the atomic intrinsics are used, this
1350 model places only one restriction on IR transformations on top of what
1351 is required for single-threaded execution: introducing a store to a byte
1352 which might not otherwise be stored is not allowed in general.
1353 (Specifically, in the case where another thread might write to and read
1354 from an address, introducing a store can change a load that may see
1355 exactly one write into a load that may see multiple writes.)
1359 Atomic Memory Ordering Constraints
1360 ----------------------------------
1362 Atomic instructions (:ref:`cmpxchg <i_cmpxchg>`,
1363 :ref:`atomicrmw <i_atomicrmw>`, :ref:`fence <i_fence>`,
1364 :ref:`atomic load <i_load>`, and :ref:`atomic store <i_store>`) take
1365 an ordering parameter that determines which other atomic instructions on
1366 the same address they *synchronize with*. These semantics are borrowed
1367 from Java and C++0x, but are somewhat more colloquial. If these
1368 descriptions aren't precise enough, check those specs (see spec
1369 references in the :doc:`atomics guide <Atomics>`).
1370 :ref:`fence <i_fence>` instructions treat these orderings somewhat
1371 differently since they don't take an address. See that instruction's
1372 documentation for details.
1374 For a simpler introduction to the ordering constraints, see the
1378 The set of values that can be read is governed by the happens-before
1379 partial order. A value cannot be read unless some operation wrote
1380 it. This is intended to provide a guarantee strong enough to model
1381 Java's non-volatile shared variables. This ordering cannot be
1382 specified for read-modify-write operations; it is not strong enough
1383 to make them atomic in any interesting way.
1385 In addition to the guarantees of ``unordered``, there is a single
1386 total order for modifications by ``monotonic`` operations on each
1387 address. All modification orders must be compatible with the
1388 happens-before order. There is no guarantee that the modification
1389 orders can be combined to a global total order for the whole program
1390 (and this often will not be possible). The read in an atomic
1391 read-modify-write operation (:ref:`cmpxchg <i_cmpxchg>` and
1392 :ref:`atomicrmw <i_atomicrmw>`) reads the value in the modification
1393 order immediately before the value it writes. If one atomic read
1394 happens before another atomic read of the same address, the later
1395 read must see the same value or a later value in the address's
1396 modification order. This disallows reordering of ``monotonic`` (or
1397 stronger) operations on the same address. If an address is written
1398 ``monotonic``-ally by one thread, and other threads ``monotonic``-ally
1399 read that address repeatedly, the other threads must eventually see
1400 the write. This corresponds to the C++0x/C1x
1401 ``memory_order_relaxed``.
1403 In addition to the guarantees of ``monotonic``, a
1404 *synchronizes-with* edge may be formed with a ``release`` operation.
1405 This is intended to model C++'s ``memory_order_acquire``.
1407 In addition to the guarantees of ``monotonic``, if this operation
1408 writes a value which is subsequently read by an ``acquire``
1409 operation, it *synchronizes-with* that operation. (This isn't a
1410 complete description; see the C++0x definition of a release
1411 sequence.) This corresponds to the C++0x/C1x
1412 ``memory_order_release``.
1413 ``acq_rel`` (acquire+release)
1414 Acts as both an ``acquire`` and ``release`` operation on its
1415 address. This corresponds to the C++0x/C1x ``memory_order_acq_rel``.
1416 ``seq_cst`` (sequentially consistent)
1417 In addition to the guarantees of ``acq_rel`` (``acquire`` for an
1418 operation which only reads, ``release`` for an operation which only
1419 writes), there is a global total order on all
1420 sequentially-consistent operations on all addresses, which is
1421 consistent with the *happens-before* partial order and with the
1422 modification orders of all the affected addresses. Each
1423 sequentially-consistent read sees the last preceding write to the
1424 same address in this global order. This corresponds to the C++0x/C1x
1425 ``memory_order_seq_cst`` and Java volatile.
1429 If an atomic operation is marked ``singlethread``, it only *synchronizes
1430 with* or participates in modification and seq\_cst total orderings with
1431 other operations running in the same thread (for example, in signal
1439 LLVM IR floating-point binary ops (:ref:`fadd <i_fadd>`,
1440 :ref:`fsub <i_fsub>`, :ref:`fmul <i_fmul>`, :ref:`fdiv <i_fdiv>`,
1441 :ref:`frem <i_frem>`) have the following flags that can set to enable
1442 otherwise unsafe floating point operations
1445 No NaNs - Allow optimizations to assume the arguments and result are not
1446 NaN. Such optimizations are required to retain defined behavior over
1447 NaNs, but the value of the result is undefined.
1450 No Infs - Allow optimizations to assume the arguments and result are not
1451 +/-Inf. Such optimizations are required to retain defined behavior over
1452 +/-Inf, but the value of the result is undefined.
1455 No Signed Zeros - Allow optimizations to treat the sign of a zero
1456 argument or result as insignificant.
1459 Allow Reciprocal - Allow optimizations to use the reciprocal of an
1460 argument rather than perform division.
1463 Fast - Allow algebraically equivalent transformations that may
1464 dramatically change results in floating point (e.g. reassociate). This
1465 flag implies all the others.
1472 The LLVM type system is one of the most important features of the
1473 intermediate representation. Being typed enables a number of
1474 optimizations to be performed on the intermediate representation
1475 directly, without having to do extra analyses on the side before the
1476 transformation. A strong type system makes it easier to read the
1477 generated code and enables novel analyses and transformations that are
1478 not feasible to perform on normal three address code representations.
1480 .. _typeclassifications:
1482 Type Classifications
1483 --------------------
1485 The types fall into a few useful classifications:
1494 * - :ref:`integer <t_integer>`
1495 - ``i1``, ``i2``, ``i3``, ... ``i8``, ... ``i16``, ... ``i32``, ...
1498 * - :ref:`floating point <t_floating>`
1499 - ``half``, ``float``, ``double``, ``x86_fp80``, ``fp128``,
1507 - :ref:`integer <t_integer>`, :ref:`floating point <t_floating>`,
1508 :ref:`pointer <t_pointer>`, :ref:`vector <t_vector>`,
1509 :ref:`structure <t_struct>`, :ref:`array <t_array>`,
1510 :ref:`label <t_label>`, :ref:`metadata <t_metadata>`.
1512 * - :ref:`primitive <t_primitive>`
1513 - :ref:`label <t_label>`,
1514 :ref:`void <t_void>`,
1515 :ref:`integer <t_integer>`,
1516 :ref:`floating point <t_floating>`,
1517 :ref:`x86mmx <t_x86mmx>`,
1518 :ref:`metadata <t_metadata>`.
1520 * - :ref:`derived <t_derived>`
1521 - :ref:`array <t_array>`,
1522 :ref:`function <t_function>`,
1523 :ref:`pointer <t_pointer>`,
1524 :ref:`structure <t_struct>`,
1525 :ref:`vector <t_vector>`,
1526 :ref:`opaque <t_opaque>`.
1528 The :ref:`first class <t_firstclass>` types are perhaps the most important.
1529 Values of these types are the only ones which can be produced by
1537 The primitive types are the fundamental building blocks of the LLVM
1548 The integer type is a very simple type that simply specifies an
1549 arbitrary bit width for the integer type desired. Any bit width from 1
1550 bit to 2\ :sup:`23`\ -1 (about 8 million) can be specified.
1559 The number of bits the integer will occupy is specified by the ``N``
1565 +----------------+------------------------------------------------+
1566 | ``i1`` | a single-bit integer. |
1567 +----------------+------------------------------------------------+
1568 | ``i32`` | a 32-bit integer. |
1569 +----------------+------------------------------------------------+
1570 | ``i1942652`` | a really big integer of over 1 million bits. |
1571 +----------------+------------------------------------------------+
1575 Floating Point Types
1576 ^^^^^^^^^^^^^^^^^^^^
1585 - 16-bit floating point value
1588 - 32-bit floating point value
1591 - 64-bit floating point value
1594 - 128-bit floating point value (112-bit mantissa)
1597 - 80-bit floating point value (X87)
1600 - 128-bit floating point value (two 64-bits)
1610 The x86mmx type represents a value held in an MMX register on an x86
1611 machine. The operations allowed on it are quite limited: parameters and
1612 return values, load and store, and bitcast. User-specified MMX
1613 instructions are represented as intrinsic or asm calls with arguments
1614 and/or results of this type. There are no arrays, vectors or constants
1632 The void type does not represent any value and has no size.
1649 The label type represents code labels.
1666 The metadata type represents embedded metadata. No derived types may be
1667 created from metadata except for :ref:`function <t_function>` arguments.
1681 The real power in LLVM comes from the derived types in the system. This
1682 is what allows a programmer to represent arrays, functions, pointers,
1683 and other useful types. Each of these types contain one or more element
1684 types which may be a primitive type, or another derived type. For
1685 example, it is possible to have a two dimensional array, using an array
1686 as the element type of another array.
1693 Aggregate Types are a subset of derived types that can contain multiple
1694 member types. :ref:`Arrays <t_array>` and :ref:`structs <t_struct>` are
1695 aggregate types. :ref:`Vectors <t_vector>` are not considered to be
1706 The array type is a very simple derived type that arranges elements
1707 sequentially in memory. The array type requires a size (number of
1708 elements) and an underlying data type.
1715 [<# elements> x <elementtype>]
1717 The number of elements is a constant integer value; ``elementtype`` may
1718 be any type with a size.
1723 +------------------+--------------------------------------+
1724 | ``[40 x i32]`` | Array of 40 32-bit integer values. |
1725 +------------------+--------------------------------------+
1726 | ``[41 x i32]`` | Array of 41 32-bit integer values. |
1727 +------------------+--------------------------------------+
1728 | ``[4 x i8]`` | Array of 4 8-bit integer values. |
1729 +------------------+--------------------------------------+
1731 Here are some examples of multidimensional arrays:
1733 +-----------------------------+----------------------------------------------------------+
1734 | ``[3 x [4 x i32]]`` | 3x4 array of 32-bit integer values. |
1735 +-----------------------------+----------------------------------------------------------+
1736 | ``[12 x [10 x float]]`` | 12x10 array of single precision floating point values. |
1737 +-----------------------------+----------------------------------------------------------+
1738 | ``[2 x [3 x [4 x i16]]]`` | 2x3x4 array of 16-bit integer values. |
1739 +-----------------------------+----------------------------------------------------------+
1741 There is no restriction on indexing beyond the end of the array implied
1742 by a static type (though there are restrictions on indexing beyond the
1743 bounds of an allocated object in some cases). This means that
1744 single-dimension 'variable sized array' addressing can be implemented in
1745 LLVM with a zero length array type. An implementation of 'pascal style
1746 arrays' in LLVM could use the type "``{ i32, [0 x float]}``", for
1757 The function type can be thought of as a function signature. It consists of a
1758 return type and a list of formal parameter types. The return type of a function
1759 type is a void type or first class type --- except for :ref:`label <t_label>`
1760 and :ref:`metadata <t_metadata>` types.
1767 <returntype> (<parameter list>)
1769 ...where '``<parameter list>``' is a comma-separated list of type
1770 specifiers. Optionally, the parameter list may include a type ``...``, which
1771 indicates that the function takes a variable number of arguments. Variable
1772 argument functions can access their arguments with the :ref:`variable argument
1773 handling intrinsic <int_varargs>` functions. '``<returntype>``' is any type
1774 except :ref:`label <t_label>` and :ref:`metadata <t_metadata>`.
1779 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1780 | ``i32 (i32)`` | function taking an ``i32``, returning an ``i32`` |
1781 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1782 | ``float (i16, i32 *) *`` | :ref:`Pointer <t_pointer>` to a function that takes an ``i16`` and a :ref:`pointer <t_pointer>` to ``i32``, returning ``float``. |
1783 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1784 | ``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. |
1785 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1786 | ``{i32, i32} (i32)`` | A function taking an ``i32``, returning a :ref:`structure <t_struct>` containing two ``i32`` values |
1787 +---------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1797 The structure type is used to represent a collection of data members
1798 together in memory. The elements of a structure may be any type that has
1801 Structures in memory are accessed using '``load``' and '``store``' by
1802 getting a pointer to a field with the '``getelementptr``' instruction.
1803 Structures in registers are accessed using the '``extractvalue``' and
1804 '``insertvalue``' instructions.
1806 Structures may optionally be "packed" structures, which indicate that
1807 the alignment of the struct is one byte, and that there is no padding
1808 between the elements. In non-packed structs, padding between field types
1809 is inserted as defined by the DataLayout string in the module, which is
1810 required to match what the underlying code generator expects.
1812 Structures can either be "literal" or "identified". A literal structure
1813 is defined inline with other types (e.g. ``{i32, i32}*``) whereas
1814 identified types are always defined at the top level with a name.
1815 Literal types are uniqued by their contents and can never be recursive
1816 or opaque since there is no way to write one. Identified types can be
1817 recursive, can be opaqued, and are never uniqued.
1824 %T1 = type { <type list> } ; Identified normal struct type
1825 %T2 = type <{ <type list> }> ; Identified packed struct type
1830 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1831 | ``{ i32, i32, i32 }`` | A triple of three ``i32`` values |
1832 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1833 | ``{ 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``. |
1834 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1835 | ``<{ i8, i32 }>`` | A packed struct known to be 5 bytes in size. |
1836 +------------------------------+---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
1840 Opaque Structure Types
1841 ^^^^^^^^^^^^^^^^^^^^^^
1846 Opaque structure types are used to represent named structure types that
1847 do not have a body specified. This corresponds (for example) to the C
1848 notion of a forward declared structure.
1861 +--------------+-------------------+
1862 | ``opaque`` | An opaque type. |
1863 +--------------+-------------------+
1873 The pointer type is used to specify memory locations. Pointers are
1874 commonly used to reference objects in memory.
1876 Pointer types may have an optional address space attribute defining the
1877 numbered address space where the pointed-to object resides. The default
1878 address space is number zero. The semantics of non-zero address spaces
1879 are target-specific.
1881 Note that LLVM does not permit pointers to void (``void*``) nor does it
1882 permit pointers to labels (``label*``). Use ``i8*`` instead.
1894 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1895 | ``[4 x i32]*`` | A :ref:`pointer <t_pointer>` to :ref:`array <t_array>` of four ``i32`` values. |
1896 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1897 | ``i32 (i32*) *`` | A :ref:`pointer <t_pointer>` to a :ref:`function <t_function>` that takes an ``i32*``, returning an ``i32``. |
1898 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1899 | ``i32 addrspace(5)*`` | A :ref:`pointer <t_pointer>` to an ``i32`` value that resides in address space #5. |
1900 +-------------------------+--------------------------------------------------------------------------------------------------------------+
1910 A vector type is a simple derived type that represents a vector of
1911 elements. Vector types are used when multiple primitive data are
1912 operated in parallel using a single instruction (SIMD). A vector type
1913 requires a size (number of elements) and an underlying primitive data
1914 type. Vector types are considered :ref:`first class <t_firstclass>`.
1921 < <# elements> x <elementtype> >
1923 The number of elements is a constant integer value larger than 0;
1924 elementtype may be any integer or floating point type, or a pointer to
1925 these types. Vectors of size zero are not allowed.
1930 +-------------------+--------------------------------------------------+
1931 | ``<4 x i32>`` | Vector of 4 32-bit integer values. |
1932 +-------------------+--------------------------------------------------+
1933 | ``<8 x float>`` | Vector of 8 32-bit floating-point values. |
1934 +-------------------+--------------------------------------------------+
1935 | ``<2 x i64>`` | Vector of 2 64-bit integer values. |
1936 +-------------------+--------------------------------------------------+
1937 | ``<4 x i64*>`` | Vector of 4 pointers to 64-bit integer values. |
1938 +-------------------+--------------------------------------------------+
1943 LLVM has several different basic types of constants. This section
1944 describes them all and their syntax.
1949 **Boolean constants**
1950 The two strings '``true``' and '``false``' are both valid constants
1952 **Integer constants**
1953 Standard integers (such as '4') are constants of the
1954 :ref:`integer <t_integer>` type. Negative numbers may be used with
1956 **Floating point constants**
1957 Floating point constants use standard decimal notation (e.g.
1958 123.421), exponential notation (e.g. 1.23421e+2), or a more precise
1959 hexadecimal notation (see below). The assembler requires the exact
1960 decimal value of a floating-point constant. For example, the
1961 assembler accepts 1.25 but rejects 1.3 because 1.3 is a repeating
1962 decimal in binary. Floating point constants must have a :ref:`floating
1963 point <t_floating>` type.
1964 **Null pointer constants**
1965 The identifier '``null``' is recognized as a null pointer constant
1966 and must be of :ref:`pointer type <t_pointer>`.
1968 The one non-intuitive notation for constants is the hexadecimal form of
1969 floating point constants. For example, the form
1970 '``double 0x432ff973cafa8000``' is equivalent to (but harder to read
1971 than) '``double 4.5e+15``'. The only time hexadecimal floating point
1972 constants are required (and the only time that they are generated by the
1973 disassembler) is when a floating point constant must be emitted but it
1974 cannot be represented as a decimal floating point number in a reasonable
1975 number of digits. For example, NaN's, infinities, and other special
1976 values are represented in their IEEE hexadecimal format so that assembly
1977 and disassembly do not cause any bits to change in the constants.
1979 When using the hexadecimal form, constants of types half, float, and
1980 double are represented using the 16-digit form shown above (which
1981 matches the IEEE754 representation for double); half and float values
1982 must, however, be exactly representable as IEEE 754 half and single
1983 precision, respectively. Hexadecimal format is always used for long
1984 double, and there are three forms of long double. The 80-bit format used
1985 by x86 is represented as ``0xK`` followed by 20 hexadecimal digits. The
1986 128-bit format used by PowerPC (two adjacent doubles) is represented by
1987 ``0xM`` followed by 32 hexadecimal digits. The IEEE 128-bit format is
1988 represented by ``0xL`` followed by 32 hexadecimal digits. Long doubles
1989 will only work if they match the long double format on your target.
1990 The IEEE 16-bit format (half precision) is represented by ``0xH``
1991 followed by 4 hexadecimal digits. All hexadecimal formats are big-endian
1992 (sign bit at the left).
1994 There are no constants of type x86mmx.
1996 .. _complexconstants:
2001 Complex constants are a (potentially recursive) combination of simple
2002 constants and smaller complex constants.
2004 **Structure constants**
2005 Structure constants are represented with notation similar to
2006 structure type definitions (a comma separated list of elements,
2007 surrounded by braces (``{}``)). For example:
2008 "``{ i32 4, float 17.0, i32* @G }``", where "``@G``" is declared as
2009 "``@G = external global i32``". Structure constants must have
2010 :ref:`structure type <t_struct>`, and the number and types of elements
2011 must match those specified by the type.
2013 Array constants are represented with notation similar to array type
2014 definitions (a comma separated list of elements, surrounded by
2015 square brackets (``[]``)). For example:
2016 "``[ i32 42, i32 11, i32 74 ]``". Array constants must have
2017 :ref:`array type <t_array>`, and the number and types of elements must
2018 match those specified by the type.
2019 **Vector constants**
2020 Vector constants are represented with notation similar to vector
2021 type definitions (a comma separated list of elements, surrounded by
2022 less-than/greater-than's (``<>``)). For example:
2023 "``< i32 42, i32 11, i32 74, i32 100 >``". Vector constants
2024 must have :ref:`vector type <t_vector>`, and the number and types of
2025 elements must match those specified by the type.
2026 **Zero initialization**
2027 The string '``zeroinitializer``' can be used to zero initialize a
2028 value to zero of *any* type, including scalar and
2029 :ref:`aggregate <t_aggregate>` types. This is often used to avoid
2030 having to print large zero initializers (e.g. for large arrays) and
2031 is always exactly equivalent to using explicit zero initializers.
2033 A metadata node is a structure-like constant with :ref:`metadata
2034 type <t_metadata>`. For example:
2035 "``metadata !{ i32 0, metadata !"test" }``". Unlike other
2036 constants that are meant to be interpreted as part of the
2037 instruction stream, metadata is a place to attach additional
2038 information such as debug info.
2040 Global Variable and Function Addresses
2041 --------------------------------------
2043 The addresses of :ref:`global variables <globalvars>` and
2044 :ref:`functions <functionstructure>` are always implicitly valid
2045 (link-time) constants. These constants are explicitly referenced when
2046 the :ref:`identifier for the global <identifiers>` is used and always have
2047 :ref:`pointer <t_pointer>` type. For example, the following is a legal LLVM
2050 .. code-block:: llvm
2054 @Z = global [2 x i32*] [ i32* @X, i32* @Y ]
2061 The string '``undef``' can be used anywhere a constant is expected, and
2062 indicates that the user of the value may receive an unspecified
2063 bit-pattern. Undefined values may be of any type (other than '``label``'
2064 or '``void``') and be used anywhere a constant is permitted.
2066 Undefined values are useful because they indicate to the compiler that
2067 the program is well defined no matter what value is used. This gives the
2068 compiler more freedom to optimize. Here are some examples of
2069 (potentially surprising) transformations that are valid (in pseudo IR):
2071 .. code-block:: llvm
2081 This is safe because all of the output bits are affected by the undef
2082 bits. Any output bit can have a zero or one depending on the input bits.
2084 .. code-block:: llvm
2095 These logical operations have bits that are not always affected by the
2096 input. For example, if ``%X`` has a zero bit, then the output of the
2097 '``and``' operation will always be a zero for that bit, no matter what
2098 the corresponding bit from the '``undef``' is. As such, it is unsafe to
2099 optimize or assume that the result of the '``and``' is '``undef``'.
2100 However, it is safe to assume that all bits of the '``undef``' could be
2101 0, and optimize the '``and``' to 0. Likewise, it is safe to assume that
2102 all the bits of the '``undef``' operand to the '``or``' could be set,
2103 allowing the '``or``' to be folded to -1.
2105 .. code-block:: llvm
2107 %A = select undef, %X, %Y
2108 %B = select undef, 42, %Y
2109 %C = select %X, %Y, undef
2119 This set of examples shows that undefined '``select``' (and conditional
2120 branch) conditions can go *either way*, but they have to come from one
2121 of the two operands. In the ``%A`` example, if ``%X`` and ``%Y`` were
2122 both known to have a clear low bit, then ``%A`` would have to have a
2123 cleared low bit. However, in the ``%C`` example, the optimizer is
2124 allowed to assume that the '``undef``' operand could be the same as
2125 ``%Y``, allowing the whole '``select``' to be eliminated.
2127 .. code-block:: llvm
2129 %A = xor undef, undef
2146 This example points out that two '``undef``' operands are not
2147 necessarily the same. This can be surprising to people (and also matches
2148 C semantics) where they assume that "``X^X``" is always zero, even if
2149 ``X`` is undefined. This isn't true for a number of reasons, but the
2150 short answer is that an '``undef``' "variable" can arbitrarily change
2151 its value over its "live range". This is true because the variable
2152 doesn't actually *have a live range*. Instead, the value is logically
2153 read from arbitrary registers that happen to be around when needed, so
2154 the value is not necessarily consistent over time. In fact, ``%A`` and
2155 ``%C`` need to have the same semantics or the core LLVM "replace all
2156 uses with" concept would not hold.
2158 .. code-block:: llvm
2166 These examples show the crucial difference between an *undefined value*
2167 and *undefined behavior*. An undefined value (like '``undef``') is
2168 allowed to have an arbitrary bit-pattern. This means that the ``%A``
2169 operation can be constant folded to '``undef``', because the '``undef``'
2170 could be an SNaN, and ``fdiv`` is not (currently) defined on SNaN's.
2171 However, in the second example, we can make a more aggressive
2172 assumption: because the ``undef`` is allowed to be an arbitrary value,
2173 we are allowed to assume that it could be zero. Since a divide by zero
2174 has *undefined behavior*, we are allowed to assume that the operation
2175 does not execute at all. This allows us to delete the divide and all
2176 code after it. Because the undefined operation "can't happen", the
2177 optimizer can assume that it occurs in dead code.
2179 .. code-block:: llvm
2181 a: store undef -> %X
2182 b: store %X -> undef
2187 These examples reiterate the ``fdiv`` example: a store *of* an undefined
2188 value can be assumed to not have any effect; we can assume that the
2189 value is overwritten with bits that happen to match what was already
2190 there. However, a store *to* an undefined location could clobber
2191 arbitrary memory, therefore, it has undefined behavior.
2198 Poison values are similar to :ref:`undef values <undefvalues>`, however
2199 they also represent the fact that an instruction or constant expression
2200 which cannot evoke side effects has nevertheless detected a condition
2201 which results in undefined behavior.
2203 There is currently no way of representing a poison value in the IR; they
2204 only exist when produced by operations such as :ref:`add <i_add>` with
2207 Poison value behavior is defined in terms of value *dependence*:
2209 - Values other than :ref:`phi <i_phi>` nodes depend on their operands.
2210 - :ref:`Phi <i_phi>` nodes depend on the operand corresponding to
2211 their dynamic predecessor basic block.
2212 - Function arguments depend on the corresponding actual argument values
2213 in the dynamic callers of their functions.
2214 - :ref:`Call <i_call>` instructions depend on the :ref:`ret <i_ret>`
2215 instructions that dynamically transfer control back to them.
2216 - :ref:`Invoke <i_invoke>` instructions depend on the
2217 :ref:`ret <i_ret>`, :ref:`resume <i_resume>`, or exception-throwing
2218 call instructions that dynamically transfer control back to them.
2219 - Non-volatile loads and stores depend on the most recent stores to all
2220 of the referenced memory addresses, following the order in the IR
2221 (including loads and stores implied by intrinsics such as
2222 :ref:`@llvm.memcpy <int_memcpy>`.)
2223 - An instruction with externally visible side effects depends on the
2224 most recent preceding instruction with externally visible side
2225 effects, following the order in the IR. (This includes :ref:`volatile
2226 operations <volatile>`.)
2227 - An instruction *control-depends* on a :ref:`terminator
2228 instruction <terminators>` if the terminator instruction has
2229 multiple successors and the instruction is always executed when
2230 control transfers to one of the successors, and may not be executed
2231 when control is transferred to another.
2232 - Additionally, an instruction also *control-depends* on a terminator
2233 instruction if the set of instructions it otherwise depends on would
2234 be different if the terminator had transferred control to a different
2236 - Dependence is transitive.
2238 Poison Values have the same behavior as :ref:`undef values <undefvalues>`,
2239 with the additional affect that any instruction which has a *dependence*
2240 on a poison value has undefined behavior.
2242 Here are some examples:
2244 .. code-block:: llvm
2247 %poison = sub nuw i32 0, 1 ; Results in a poison value.
2248 %still_poison = and i32 %poison, 0 ; 0, but also poison.
2249 %poison_yet_again = getelementptr i32* @h, i32 %still_poison
2250 store i32 0, i32* %poison_yet_again ; memory at @h[0] is poisoned
2252 store i32 %poison, i32* @g ; Poison value stored to memory.
2253 %poison2 = load i32* @g ; Poison value loaded back from memory.
2255 store volatile i32 %poison, i32* @g ; External observation; undefined behavior.
2257 %narrowaddr = bitcast i32* @g to i16*
2258 %wideaddr = bitcast i32* @g to i64*
2259 %poison3 = load i16* %narrowaddr ; Returns a poison value.
2260 %poison4 = load i64* %wideaddr ; Returns a poison value.
2262 %cmp = icmp slt i32 %poison, 0 ; Returns a poison value.
2263 br i1 %cmp, label %true, label %end ; Branch to either destination.
2266 store volatile i32 0, i32* @g ; This is control-dependent on %cmp, so
2267 ; it has undefined behavior.
2271 %p = phi i32 [ 0, %entry ], [ 1, %true ]
2272 ; Both edges into this PHI are
2273 ; control-dependent on %cmp, so this
2274 ; always results in a poison value.
2276 store volatile i32 0, i32* @g ; This would depend on the store in %true
2277 ; if %cmp is true, or the store in %entry
2278 ; otherwise, so this is undefined behavior.
2280 br i1 %cmp, label %second_true, label %second_end
2281 ; The same branch again, but this time the
2282 ; true block doesn't have side effects.
2289 store volatile i32 0, i32* @g ; This time, the instruction always depends
2290 ; on the store in %end. Also, it is
2291 ; control-equivalent to %end, so this is
2292 ; well-defined (ignoring earlier undefined
2293 ; behavior in this example).
2297 Addresses of Basic Blocks
2298 -------------------------
2300 ``blockaddress(@function, %block)``
2302 The '``blockaddress``' constant computes the address of the specified
2303 basic block in the specified function, and always has an ``i8*`` type.
2304 Taking the address of the entry block is illegal.
2306 This value only has defined behavior when used as an operand to the
2307 ':ref:`indirectbr <i_indirectbr>`' instruction, or for comparisons
2308 against null. Pointer equality tests between labels addresses results in
2309 undefined behavior --- though, again, comparison against null is ok, and
2310 no label is equal to the null pointer. This may be passed around as an
2311 opaque pointer sized value as long as the bits are not inspected. This
2312 allows ``ptrtoint`` and arithmetic to be performed on these values so
2313 long as the original value is reconstituted before the ``indirectbr``
2316 Finally, some targets may provide defined semantics when using the value
2317 as the operand to an inline assembly, but that is target specific.
2321 Constant Expressions
2322 --------------------
2324 Constant expressions are used to allow expressions involving other
2325 constants to be used as constants. Constant expressions may be of any
2326 :ref:`first class <t_firstclass>` type and may involve any LLVM operation
2327 that does not have side effects (e.g. load and call are not supported).
2328 The following is the syntax for constant expressions:
2330 ``trunc (CST to TYPE)``
2331 Truncate a constant to another type. The bit size of CST must be
2332 larger than the bit size of TYPE. Both types must be integers.
2333 ``zext (CST to TYPE)``
2334 Zero extend a constant to another type. The bit size of CST must be
2335 smaller than the bit size of TYPE. Both types must be integers.
2336 ``sext (CST to TYPE)``
2337 Sign extend a constant to another type. The bit size of CST must be
2338 smaller than the bit size of TYPE. Both types must be integers.
2339 ``fptrunc (CST to TYPE)``
2340 Truncate a floating point constant to another floating point type.
2341 The size of CST must be larger than the size of TYPE. Both types
2342 must be floating point.
2343 ``fpext (CST to TYPE)``
2344 Floating point extend a constant to another type. The size of CST
2345 must be smaller or equal to the size of TYPE. Both types must be
2347 ``fptoui (CST to TYPE)``
2348 Convert a floating point constant to the corresponding unsigned
2349 integer constant. TYPE must be a scalar or vector integer type. CST
2350 must be of scalar or vector floating point type. Both CST and TYPE
2351 must be scalars, or vectors of the same number of elements. If the
2352 value won't fit in the integer type, the results are undefined.
2353 ``fptosi (CST to TYPE)``
2354 Convert a floating point constant to the corresponding signed
2355 integer constant. TYPE must be a scalar or vector integer type. CST
2356 must be of scalar or vector floating point type. Both CST and TYPE
2357 must be scalars, or vectors of the same number of elements. If the
2358 value won't fit in the integer type, the results are undefined.
2359 ``uitofp (CST to TYPE)``
2360 Convert an unsigned integer constant to the corresponding floating
2361 point constant. TYPE must be a scalar or vector floating point type.
2362 CST must be of scalar or vector integer type. Both CST and TYPE must
2363 be scalars, or vectors of the same number of elements. If the value
2364 won't fit in the floating point type, the results are undefined.
2365 ``sitofp (CST to TYPE)``
2366 Convert a signed integer constant to the corresponding floating
2367 point constant. TYPE must be a scalar or vector floating point type.
2368 CST must be of scalar or vector integer type. Both CST and TYPE must
2369 be scalars, or vectors of the same number of elements. If the value
2370 won't fit in the floating point type, the results are undefined.
2371 ``ptrtoint (CST to TYPE)``
2372 Convert a pointer typed constant to the corresponding integer
2373 constant. ``TYPE`` must be an integer type. ``CST`` must be of
2374 pointer type. The ``CST`` value is zero extended, truncated, or
2375 unchanged to make it fit in ``TYPE``.
2376 ``inttoptr (CST to TYPE)``
2377 Convert an integer constant to a pointer constant. TYPE must be a
2378 pointer type. CST must be of integer type. The CST value is zero
2379 extended, truncated, or unchanged to make it fit in a pointer size.
2380 This one is *really* dangerous!
2381 ``bitcast (CST to TYPE)``
2382 Convert a constant, CST, to another TYPE. The constraints of the
2383 operands are the same as those for the :ref:`bitcast
2384 instruction <i_bitcast>`.
2385 ``addrspacecast (CST to TYPE)``
2386 Convert a constant pointer or constant vector of pointer, CST, to another
2387 TYPE in a different address space. The constraints of the operands are the
2388 same as those for the :ref:`addrspacecast instruction <i_addrspacecast>`.
2389 ``getelementptr (CSTPTR, IDX0, IDX1, ...)``, ``getelementptr inbounds (CSTPTR, IDX0, IDX1, ...)``
2390 Perform the :ref:`getelementptr operation <i_getelementptr>` on
2391 constants. As with the :ref:`getelementptr <i_getelementptr>`
2392 instruction, the index list may have zero or more indexes, which are
2393 required to make sense for the type of "CSTPTR".
2394 ``select (COND, VAL1, VAL2)``
2395 Perform the :ref:`select operation <i_select>` on constants.
2396 ``icmp COND (VAL1, VAL2)``
2397 Performs the :ref:`icmp operation <i_icmp>` on constants.
2398 ``fcmp COND (VAL1, VAL2)``
2399 Performs the :ref:`fcmp operation <i_fcmp>` on constants.
2400 ``extractelement (VAL, IDX)``
2401 Perform the :ref:`extractelement operation <i_extractelement>` on
2403 ``insertelement (VAL, ELT, IDX)``
2404 Perform the :ref:`insertelement operation <i_insertelement>` on
2406 ``shufflevector (VEC1, VEC2, IDXMASK)``
2407 Perform the :ref:`shufflevector operation <i_shufflevector>` on
2409 ``extractvalue (VAL, IDX0, IDX1, ...)``
2410 Perform the :ref:`extractvalue operation <i_extractvalue>` on
2411 constants. The index list is interpreted in a similar manner as
2412 indices in a ':ref:`getelementptr <i_getelementptr>`' operation. At
2413 least one index value must be specified.
2414 ``insertvalue (VAL, ELT, IDX0, IDX1, ...)``
2415 Perform the :ref:`insertvalue operation <i_insertvalue>` on constants.
2416 The index list is interpreted in a similar manner as indices in a
2417 ':ref:`getelementptr <i_getelementptr>`' operation. At least one index
2418 value must be specified.
2419 ``OPCODE (LHS, RHS)``
2420 Perform the specified operation of the LHS and RHS constants. OPCODE
2421 may be any of the :ref:`binary <binaryops>` or :ref:`bitwise
2422 binary <bitwiseops>` operations. The constraints on operands are
2423 the same as those for the corresponding instruction (e.g. no bitwise
2424 operations on floating point values are allowed).
2431 Inline Assembler Expressions
2432 ----------------------------
2434 LLVM supports inline assembler expressions (as opposed to :ref:`Module-Level
2435 Inline Assembly <moduleasm>`) through the use of a special value. This
2436 value represents the inline assembler as a string (containing the
2437 instructions to emit), a list of operand constraints (stored as a
2438 string), a flag that indicates whether or not the inline asm expression
2439 has side effects, and a flag indicating whether the function containing
2440 the asm needs to align its stack conservatively. An example inline
2441 assembler expression is:
2443 .. code-block:: llvm
2445 i32 (i32) asm "bswap $0", "=r,r"
2447 Inline assembler expressions may **only** be used as the callee operand
2448 of a :ref:`call <i_call>` or an :ref:`invoke <i_invoke>` instruction.
2449 Thus, typically we have:
2451 .. code-block:: llvm
2453 %X = call i32 asm "bswap $0", "=r,r"(i32 %Y)
2455 Inline asms with side effects not visible in the constraint list must be
2456 marked as having side effects. This is done through the use of the
2457 '``sideeffect``' keyword, like so:
2459 .. code-block:: llvm
2461 call void asm sideeffect "eieio", ""()
2463 In some cases inline asms will contain code that will not work unless
2464 the stack is aligned in some way, such as calls or SSE instructions on
2465 x86, yet will not contain code that does that alignment within the asm.
2466 The compiler should make conservative assumptions about what the asm
2467 might contain and should generate its usual stack alignment code in the
2468 prologue if the '``alignstack``' keyword is present:
2470 .. code-block:: llvm
2472 call void asm alignstack "eieio", ""()
2474 Inline asms also support using non-standard assembly dialects. The
2475 assumed dialect is ATT. When the '``inteldialect``' keyword is present,
2476 the inline asm is using the Intel dialect. Currently, ATT and Intel are
2477 the only supported dialects. An example is:
2479 .. code-block:: llvm
2481 call void asm inteldialect "eieio", ""()
2483 If multiple keywords appear the '``sideeffect``' keyword must come
2484 first, the '``alignstack``' keyword second and the '``inteldialect``'
2490 The call instructions that wrap inline asm nodes may have a
2491 "``!srcloc``" MDNode attached to it that contains a list of constant
2492 integers. If present, the code generator will use the integer as the
2493 location cookie value when report errors through the ``LLVMContext``
2494 error reporting mechanisms. This allows a front-end to correlate backend
2495 errors that occur with inline asm back to the source code that produced
2498 .. code-block:: llvm
2500 call void asm sideeffect "something bad", ""(), !srcloc !42
2502 !42 = !{ i32 1234567 }
2504 It is up to the front-end to make sense of the magic numbers it places
2505 in the IR. If the MDNode contains multiple constants, the code generator
2506 will use the one that corresponds to the line of the asm that the error
2511 Metadata Nodes and Metadata Strings
2512 -----------------------------------
2514 LLVM IR allows metadata to be attached to instructions in the program
2515 that can convey extra information about the code to the optimizers and
2516 code generator. One example application of metadata is source-level
2517 debug information. There are two metadata primitives: strings and nodes.
2518 All metadata has the ``metadata`` type and is identified in syntax by a
2519 preceding exclamation point ('``!``').
2521 A metadata string is a string surrounded by double quotes. It can
2522 contain any character by escaping non-printable characters with
2523 "``\xx``" where "``xx``" is the two digit hex code. For example:
2526 Metadata nodes are represented with notation similar to structure
2527 constants (a comma separated list of elements, surrounded by braces and
2528 preceded by an exclamation point). Metadata nodes can have any values as
2529 their operand. For example:
2531 .. code-block:: llvm
2533 !{ metadata !"test\00", i32 10}
2535 A :ref:`named metadata <namedmetadatastructure>` is a collection of
2536 metadata nodes, which can be looked up in the module symbol table. For
2539 .. code-block:: llvm
2541 !foo = metadata !{!4, !3}
2543 Metadata can be used as function arguments. Here ``llvm.dbg.value``
2544 function is using two metadata arguments:
2546 .. code-block:: llvm
2548 call void @llvm.dbg.value(metadata !24, i64 0, metadata !25)
2550 Metadata can be attached with an instruction. Here metadata ``!21`` is
2551 attached to the ``add`` instruction using the ``!dbg`` identifier:
2553 .. code-block:: llvm
2555 %indvar.next = add i64 %indvar, 1, !dbg !21
2557 More information about specific metadata nodes recognized by the
2558 optimizers and code generator is found below.
2563 In LLVM IR, memory does not have types, so LLVM's own type system is not
2564 suitable for doing TBAA. Instead, metadata is added to the IR to
2565 describe a type system of a higher level language. This can be used to
2566 implement typical C/C++ TBAA, but it can also be used to implement
2567 custom alias analysis behavior for other languages.
2569 The current metadata format is very simple. TBAA metadata nodes have up
2570 to three fields, e.g.:
2572 .. code-block:: llvm
2574 !0 = metadata !{ metadata !"an example type tree" }
2575 !1 = metadata !{ metadata !"int", metadata !0 }
2576 !2 = metadata !{ metadata !"float", metadata !0 }
2577 !3 = metadata !{ metadata !"const float", metadata !2, i64 1 }
2579 The first field is an identity field. It can be any value, usually a
2580 metadata string, which uniquely identifies the type. The most important
2581 name in the tree is the name of the root node. Two trees with different
2582 root node names are entirely disjoint, even if they have leaves with
2585 The second field identifies the type's parent node in the tree, or is
2586 null or omitted for a root node. A type is considered to alias all of
2587 its descendants and all of its ancestors in the tree. Also, a type is
2588 considered to alias all types in other trees, so that bitcode produced
2589 from multiple front-ends is handled conservatively.
2591 If the third field is present, it's an integer which if equal to 1
2592 indicates that the type is "constant" (meaning
2593 ``pointsToConstantMemory`` should return true; see `other useful
2594 AliasAnalysis methods <AliasAnalysis.html#OtherItfs>`_).
2596 '``tbaa.struct``' Metadata
2597 ^^^^^^^^^^^^^^^^^^^^^^^^^^
2599 The :ref:`llvm.memcpy <int_memcpy>` is often used to implement
2600 aggregate assignment operations in C and similar languages, however it
2601 is defined to copy a contiguous region of memory, which is more than
2602 strictly necessary for aggregate types which contain holes due to
2603 padding. Also, it doesn't contain any TBAA information about the fields
2606 ``!tbaa.struct`` metadata can describe which memory subregions in a
2607 memcpy are padding and what the TBAA tags of the struct are.
2609 The current metadata format is very simple. ``!tbaa.struct`` metadata
2610 nodes are a list of operands which are in conceptual groups of three.
2611 For each group of three, the first operand gives the byte offset of a
2612 field in bytes, the second gives its size in bytes, and the third gives
2615 .. code-block:: llvm
2617 !4 = metadata !{ i64 0, i64 4, metadata !1, i64 8, i64 4, metadata !2 }
2619 This describes a struct with two fields. The first is at offset 0 bytes
2620 with size 4 bytes, and has tbaa tag !1. The second is at offset 8 bytes
2621 and has size 4 bytes and has tbaa tag !2.
2623 Note that the fields need not be contiguous. In this example, there is a
2624 4 byte gap between the two fields. This gap represents padding which
2625 does not carry useful data and need not be preserved.
2627 '``fpmath``' Metadata
2628 ^^^^^^^^^^^^^^^^^^^^^
2630 ``fpmath`` metadata may be attached to any instruction of floating point
2631 type. It can be used to express the maximum acceptable error in the
2632 result of that instruction, in ULPs, thus potentially allowing the
2633 compiler to use a more efficient but less accurate method of computing
2634 it. ULP is defined as follows:
2636 If ``x`` is a real number that lies between two finite consecutive
2637 floating-point numbers ``a`` and ``b``, without being equal to one
2638 of them, then ``ulp(x) = |b - a|``, otherwise ``ulp(x)`` is the
2639 distance between the two non-equal finite floating-point numbers
2640 nearest ``x``. Moreover, ``ulp(NaN)`` is ``NaN``.
2642 The metadata node shall consist of a single positive floating point
2643 number representing the maximum relative error, for example:
2645 .. code-block:: llvm
2647 !0 = metadata !{ float 2.5 } ; maximum acceptable inaccuracy is 2.5 ULPs
2649 '``range``' Metadata
2650 ^^^^^^^^^^^^^^^^^^^^
2652 ``range`` metadata may be attached only to loads of integer types. It
2653 expresses the possible ranges the loaded value is in. The ranges are
2654 represented with a flattened list of integers. The loaded value is known
2655 to be in the union of the ranges defined by each consecutive pair. Each
2656 pair has the following properties:
2658 - The type must match the type loaded by the instruction.
2659 - The pair ``a,b`` represents the range ``[a,b)``.
2660 - Both ``a`` and ``b`` are constants.
2661 - The range is allowed to wrap.
2662 - The range should not represent the full or empty set. That is,
2665 In addition, the pairs must be in signed order of the lower bound and
2666 they must be non-contiguous.
2670 .. code-block:: llvm
2672 %a = load i8* %x, align 1, !range !0 ; Can only be 0 or 1
2673 %b = load i8* %y, align 1, !range !1 ; Can only be 255 (-1), 0 or 1
2674 %c = load i8* %z, align 1, !range !2 ; Can only be 0, 1, 3, 4 or 5
2675 %d = load i8* %z, align 1, !range !3 ; Can only be -2, -1, 3, 4 or 5
2677 !0 = metadata !{ i8 0, i8 2 }
2678 !1 = metadata !{ i8 255, i8 2 }
2679 !2 = metadata !{ i8 0, i8 2, i8 3, i8 6 }
2680 !3 = metadata !{ i8 -2, i8 0, i8 3, i8 6 }
2685 It is sometimes useful to attach information to loop constructs. Currently,
2686 loop metadata is implemented as metadata attached to the branch instruction
2687 in the loop latch block. This type of metadata refer to a metadata node that is
2688 guaranteed to be separate for each loop. The loop identifier metadata is
2689 specified with the name ``llvm.loop``.
2691 The loop identifier metadata is implemented using a metadata that refers to
2692 itself to avoid merging it with any other identifier metadata, e.g.,
2693 during module linkage or function inlining. That is, each loop should refer
2694 to their own identification metadata even if they reside in separate functions.
2695 The following example contains loop identifier metadata for two separate loop
2698 .. code-block:: llvm
2700 !0 = metadata !{ metadata !0 }
2701 !1 = metadata !{ metadata !1 }
2703 The loop identifier metadata can be used to specify additional per-loop
2704 metadata. Any operands after the first operand can be treated as user-defined
2705 metadata. For example the ``llvm.vectorizer.unroll`` metadata is understood
2706 by the loop vectorizer to indicate how many times to unroll the loop:
2708 .. code-block:: llvm
2710 br i1 %exitcond, label %._crit_edge, label %.lr.ph, !llvm.loop !0
2712 !0 = metadata !{ metadata !0, metadata !1 }
2713 !1 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 2 }
2718 Metadata types used to annotate memory accesses with information helpful
2719 for optimizations are prefixed with ``llvm.mem``.
2721 '``llvm.mem.parallel_loop_access``' Metadata
2722 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2724 For a loop to be parallel, in addition to using
2725 the ``llvm.loop`` metadata to mark the loop latch branch instruction,
2726 also all of the memory accessing instructions in the loop body need to be
2727 marked with the ``llvm.mem.parallel_loop_access`` metadata. If there
2728 is at least one memory accessing instruction not marked with the metadata,
2729 the loop must be considered a sequential loop. This causes parallel loops to be
2730 converted to sequential loops due to optimization passes that are unaware of
2731 the parallel semantics and that insert new memory instructions to the loop
2734 Example of a loop that is considered parallel due to its correct use of
2735 both ``llvm.loop`` and ``llvm.mem.parallel_loop_access``
2736 metadata types that refer to the same loop identifier metadata.
2738 .. code-block:: llvm
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 %for.end, label %for.body, !llvm.loop !0
2750 !0 = metadata !{ metadata !0 }
2752 It is also possible to have nested parallel loops. In that case the
2753 memory accesses refer to a list of loop identifier metadata nodes instead of
2754 the loop identifier metadata node directly:
2756 .. code-block:: llvm
2763 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2765 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2767 br i1 %exitcond, label %inner.for.end, label %inner.for.body, !llvm.loop !1
2771 %0 = load i32* %arrayidx, align 4, !llvm.mem.parallel_loop_access !0
2773 store i32 %0, i32* %arrayidx4, align 4, !llvm.mem.parallel_loop_access !0
2775 br i1 %exitcond, label %outer.for.end, label %outer.for.body, !llvm.loop !2
2777 outer.for.end: ; preds = %for.body
2779 !0 = metadata !{ metadata !1, metadata !2 } ; a list of loop identifiers
2780 !1 = metadata !{ metadata !1 } ; an identifier for the inner loop
2781 !2 = metadata !{ metadata !2 } ; an identifier for the outer loop
2783 '``llvm.vectorizer``'
2784 ^^^^^^^^^^^^^^^^^^^^^
2786 Metadata prefixed with ``llvm.vectorizer`` is used to control per-loop
2787 vectorization parameters such as vectorization factor and unroll factor.
2789 ``llvm.vectorizer`` metadata should be used in conjunction with ``llvm.loop``
2790 loop identification metadata.
2792 '``llvm.vectorizer.unroll``' Metadata
2793 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2795 This metadata instructs the loop vectorizer to unroll the specified
2796 loop exactly ``N`` times.
2798 The first operand is the string ``llvm.vectorizer.unroll`` and the second
2799 operand is an integer specifying the unroll factor. For example:
2801 .. code-block:: llvm
2803 !0 = metadata !{ metadata !"llvm.vectorizer.unroll", i32 4 }
2805 Note that setting ``llvm.vectorizer.unroll`` to 1 disables unrolling of the
2808 If ``llvm.vectorizer.unroll`` is set to 0 then the amount of unrolling will be
2809 determined automatically.
2811 '``llvm.vectorizer.width``' Metadata
2812 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2814 This metadata sets the target width of the vectorizer to ``N``. Without
2815 this metadata, the vectorizer will choose a width automatically.
2816 Regardless of this metadata, the vectorizer will only vectorize loops if
2817 it believes it is valid to do so.
2819 The first operand is the string ``llvm.vectorizer.width`` and the second
2820 operand is an integer specifying the width. For example:
2822 .. code-block:: llvm
2824 !0 = metadata !{ metadata !"llvm.vectorizer.width", i32 4 }
2826 Note that setting ``llvm.vectorizer.width`` to 1 disables vectorization of the
2829 If ``llvm.vectorizer.width`` is set to 0 then the width will be determined
2832 Module Flags Metadata
2833 =====================
2835 Information about the module as a whole is difficult to convey to LLVM's
2836 subsystems. The LLVM IR isn't sufficient to transmit this information.
2837 The ``llvm.module.flags`` named metadata exists in order to facilitate
2838 this. These flags are in the form of key / value pairs --- much like a
2839 dictionary --- making it easy for any subsystem who cares about a flag to
2842 The ``llvm.module.flags`` metadata contains a list of metadata triplets.
2843 Each triplet has the following form:
2845 - The first element is a *behavior* flag, which specifies the behavior
2846 when two (or more) modules are merged together, and it encounters two
2847 (or more) metadata with the same ID. The supported behaviors are
2849 - The second element is a metadata string that is a unique ID for the
2850 metadata. Each module may only have one flag entry for each unique ID (not
2851 including entries with the **Require** behavior).
2852 - The third element is the value of the flag.
2854 When two (or more) modules are merged together, the resulting
2855 ``llvm.module.flags`` metadata is the union of the modules' flags. That is, for
2856 each unique metadata ID string, there will be exactly one entry in the merged
2857 modules ``llvm.module.flags`` metadata table, and the value for that entry will
2858 be determined by the merge behavior flag, as described below. The only exception
2859 is that entries with the *Require* behavior are always preserved.
2861 The following behaviors are supported:
2872 Emits an error if two values disagree, otherwise the resulting value
2873 is that of the operands.
2877 Emits a warning if two values disagree. The result value will be the
2878 operand for the flag from the first module being linked.
2882 Adds a requirement that another module flag be present and have a
2883 specified value after linking is performed. The value must be a
2884 metadata pair, where the first element of the pair is the ID of the
2885 module flag to be restricted, and the second element of the pair is
2886 the value the module flag should be restricted to. This behavior can
2887 be used to restrict the allowable results (via triggering of an
2888 error) of linking IDs with the **Override** behavior.
2892 Uses the specified value, regardless of the behavior or value of the
2893 other module. If both modules specify **Override**, but the values
2894 differ, an error will be emitted.
2898 Appends the two values, which are required to be metadata nodes.
2902 Appends the two values, which are required to be metadata
2903 nodes. However, duplicate entries in the second list are dropped
2904 during the append operation.
2906 It is an error for a particular unique flag ID to have multiple behaviors,
2907 except in the case of **Require** (which adds restrictions on another metadata
2908 value) or **Override**.
2910 An example of module flags:
2912 .. code-block:: llvm
2914 !0 = metadata !{ i32 1, metadata !"foo", i32 1 }
2915 !1 = metadata !{ i32 4, metadata !"bar", i32 37 }
2916 !2 = metadata !{ i32 2, metadata !"qux", i32 42 }
2917 !3 = metadata !{ i32 3, metadata !"qux",
2919 metadata !"foo", i32 1
2922 !llvm.module.flags = !{ !0, !1, !2, !3 }
2924 - Metadata ``!0`` has the ID ``!"foo"`` and the value '1'. The behavior
2925 if two or more ``!"foo"`` flags are seen is to emit an error if their
2926 values are not equal.
2928 - Metadata ``!1`` has the ID ``!"bar"`` and the value '37'. The
2929 behavior if two or more ``!"bar"`` flags are seen is to use the value
2932 - Metadata ``!2`` has the ID ``!"qux"`` and the value '42'. The
2933 behavior if two or more ``!"qux"`` flags are seen is to emit a
2934 warning if their values are not equal.
2936 - Metadata ``!3`` has the ID ``!"qux"`` and the value:
2940 metadata !{ metadata !"foo", i32 1 }
2942 The behavior is to emit an error if the ``llvm.module.flags`` does not
2943 contain a flag with the ID ``!"foo"`` that has the value '1' after linking is
2946 Objective-C Garbage Collection Module Flags Metadata
2947 ----------------------------------------------------
2949 On the Mach-O platform, Objective-C stores metadata about garbage
2950 collection in a special section called "image info". The metadata
2951 consists of a version number and a bitmask specifying what types of
2952 garbage collection are supported (if any) by the file. If two or more
2953 modules are linked together their garbage collection metadata needs to
2954 be merged rather than appended together.
2956 The Objective-C garbage collection module flags metadata consists of the
2957 following key-value pairs:
2966 * - ``Objective-C Version``
2967 - **[Required]** --- The Objective-C ABI version. Valid values are 1 and 2.
2969 * - ``Objective-C Image Info Version``
2970 - **[Required]** --- The version of the image info section. Currently
2973 * - ``Objective-C Image Info Section``
2974 - **[Required]** --- The section to place the metadata. Valid values are
2975 ``"__OBJC, __image_info, regular"`` for Objective-C ABI version 1, and
2976 ``"__DATA,__objc_imageinfo, regular, no_dead_strip"`` for
2977 Objective-C ABI version 2.
2979 * - ``Objective-C Garbage Collection``
2980 - **[Required]** --- Specifies whether garbage collection is supported or
2981 not. Valid values are 0, for no garbage collection, and 2, for garbage
2982 collection supported.
2984 * - ``Objective-C GC Only``
2985 - **[Optional]** --- Specifies that only garbage collection is supported.
2986 If present, its value must be 6. This flag requires that the
2987 ``Objective-C Garbage Collection`` flag have the value 2.
2989 Some important flag interactions:
2991 - If a module with ``Objective-C Garbage Collection`` set to 0 is
2992 merged with a module with ``Objective-C Garbage Collection`` set to
2993 2, then the resulting module has the
2994 ``Objective-C Garbage Collection`` flag set to 0.
2995 - A module with ``Objective-C Garbage Collection`` set to 0 cannot be
2996 merged with a module with ``Objective-C GC Only`` set to 6.
2998 Automatic Linker Flags Module Flags Metadata
2999 --------------------------------------------
3001 Some targets support embedding flags to the linker inside individual object
3002 files. Typically this is used in conjunction with language extensions which
3003 allow source files to explicitly declare the libraries they depend on, and have
3004 these automatically be transmitted to the linker via object files.
3006 These flags are encoded in the IR using metadata in the module flags section,
3007 using the ``Linker Options`` key. The merge behavior for this flag is required
3008 to be ``AppendUnique``, and the value for the key is expected to be a metadata
3009 node which should be a list of other metadata nodes, each of which should be a
3010 list of metadata strings defining linker options.
3012 For example, the following metadata section specifies two separate sets of
3013 linker options, presumably to link against ``libz`` and the ``Cocoa``
3016 !0 = metadata !{ i32 6, metadata !"Linker Options",
3018 metadata !{ metadata !"-lz" },
3019 metadata !{ metadata !"-framework", metadata !"Cocoa" } } }
3020 !llvm.module.flags = !{ !0 }
3022 The metadata encoding as lists of lists of options, as opposed to a collapsed
3023 list of options, is chosen so that the IR encoding can use multiple option
3024 strings to specify e.g., a single library, while still having that specifier be
3025 preserved as an atomic element that can be recognized by a target specific
3026 assembly writer or object file emitter.
3028 Each individual option is required to be either a valid option for the target's
3029 linker, or an option that is reserved by the target specific assembly writer or
3030 object file emitter. No other aspect of these options is defined by the IR.
3032 .. _intrinsicglobalvariables:
3034 Intrinsic Global Variables
3035 ==========================
3037 LLVM has a number of "magic" global variables that contain data that
3038 affect code generation or other IR semantics. These are documented here.
3039 All globals of this sort should have a section specified as
3040 "``llvm.metadata``". This section and all globals that start with
3041 "``llvm.``" are reserved for use by LLVM.
3045 The '``llvm.used``' Global Variable
3046 -----------------------------------
3048 The ``@llvm.used`` global is an array which has
3049 :ref:`appending linkage <linkage_appending>`. This array contains a list of
3050 pointers to named global variables, functions and aliases which may optionally
3051 have a pointer cast formed of bitcast or getelementptr. For example, a legal
3054 .. code-block:: llvm
3059 @llvm.used = appending global [2 x i8*] [
3061 i8* bitcast (i32* @Y to i8*)
3062 ], section "llvm.metadata"
3064 If a symbol appears in the ``@llvm.used`` list, then the compiler, assembler,
3065 and linker are required to treat the symbol as if there is a reference to the
3066 symbol that it cannot see (which is why they have to be named). For example, if
3067 a variable has internal linkage and no references other than that from the
3068 ``@llvm.used`` list, it cannot be deleted. This is commonly used to represent
3069 references from inline asms and other things the compiler cannot "see", and
3070 corresponds to "``attribute((used))``" in GNU C.
3072 On some targets, the code generator must emit a directive to the
3073 assembler or object file to prevent the assembler and linker from
3074 molesting the symbol.
3076 .. _gv_llvmcompilerused:
3078 The '``llvm.compiler.used``' Global Variable
3079 --------------------------------------------
3081 The ``@llvm.compiler.used`` directive is the same as the ``@llvm.used``
3082 directive, except that it only prevents the compiler from touching the
3083 symbol. On targets that support it, this allows an intelligent linker to
3084 optimize references to the symbol without being impeded as it would be
3087 This is a rare construct that should only be used in rare circumstances,
3088 and should not be exposed to source languages.
3090 .. _gv_llvmglobalctors:
3092 The '``llvm.global_ctors``' Global Variable
3093 -------------------------------------------
3095 .. code-block:: llvm
3097 %0 = type { i32, void ()* }
3098 @llvm.global_ctors = appending global [1 x %0] [%0 { i32 65535, void ()* @ctor }]
3100 The ``@llvm.global_ctors`` array contains a list of constructor
3101 functions and associated priorities. The functions referenced by this
3102 array will be called in ascending order of priority (i.e. lowest first)
3103 when the module is loaded. The order of functions with the same priority
3106 .. _llvmglobaldtors:
3108 The '``llvm.global_dtors``' Global Variable
3109 -------------------------------------------
3111 .. code-block:: llvm
3113 %0 = type { i32, void ()* }
3114 @llvm.global_dtors = appending global [1 x %0] [%0 { i32 65535, void ()* @dtor }]
3116 The ``@llvm.global_dtors`` array contains a list of destructor functions
3117 and associated priorities. The functions referenced by this array will
3118 be called in descending order of priority (i.e. highest first) when the
3119 module is loaded. The order of functions with the same priority is not
3122 Instruction Reference
3123 =====================
3125 The LLVM instruction set consists of several different classifications
3126 of instructions: :ref:`terminator instructions <terminators>`, :ref:`binary
3127 instructions <binaryops>`, :ref:`bitwise binary
3128 instructions <bitwiseops>`, :ref:`memory instructions <memoryops>`, and
3129 :ref:`other instructions <otherops>`.
3133 Terminator Instructions
3134 -----------------------
3136 As mentioned :ref:`previously <functionstructure>`, every basic block in a
3137 program ends with a "Terminator" instruction, which indicates which
3138 block should be executed after the current block is finished. These
3139 terminator instructions typically yield a '``void``' value: they produce
3140 control flow, not values (the one exception being the
3141 ':ref:`invoke <i_invoke>`' instruction).
3143 The terminator instructions are: ':ref:`ret <i_ret>`',
3144 ':ref:`br <i_br>`', ':ref:`switch <i_switch>`',
3145 ':ref:`indirectbr <i_indirectbr>`', ':ref:`invoke <i_invoke>`',
3146 ':ref:`resume <i_resume>`', and ':ref:`unreachable <i_unreachable>`'.
3150 '``ret``' Instruction
3151 ^^^^^^^^^^^^^^^^^^^^^
3158 ret <type> <value> ; Return a value from a non-void function
3159 ret void ; Return from void function
3164 The '``ret``' instruction is used to return control flow (and optionally
3165 a value) from a function back to the caller.
3167 There are two forms of the '``ret``' instruction: one that returns a
3168 value and then causes control flow, and one that just causes control
3174 The '``ret``' instruction optionally accepts a single argument, the
3175 return value. The type of the return value must be a ':ref:`first
3176 class <t_firstclass>`' type.
3178 A function is not :ref:`well formed <wellformed>` if it it has a non-void
3179 return type and contains a '``ret``' instruction with no return value or
3180 a return value with a type that does not match its type, or if it has a
3181 void return type and contains a '``ret``' instruction with a return
3187 When the '``ret``' instruction is executed, control flow returns back to
3188 the calling function's context. If the caller is a
3189 ":ref:`call <i_call>`" instruction, execution continues at the
3190 instruction after the call. If the caller was an
3191 ":ref:`invoke <i_invoke>`" instruction, execution continues at the
3192 beginning of the "normal" destination block. If the instruction returns
3193 a value, that value shall set the call or invoke instruction's return
3199 .. code-block:: llvm
3201 ret i32 5 ; Return an integer value of 5
3202 ret void ; Return from a void function
3203 ret { i32, i8 } { i32 4, i8 2 } ; Return a struct of values 4 and 2
3207 '``br``' Instruction
3208 ^^^^^^^^^^^^^^^^^^^^
3215 br i1 <cond>, label <iftrue>, label <iffalse>
3216 br label <dest> ; Unconditional branch
3221 The '``br``' instruction is used to cause control flow to transfer to a
3222 different basic block in the current function. There are two forms of
3223 this instruction, corresponding to a conditional branch and an
3224 unconditional branch.
3229 The conditional branch form of the '``br``' instruction takes a single
3230 '``i1``' value and two '``label``' values. The unconditional form of the
3231 '``br``' instruction takes a single '``label``' value as a target.
3236 Upon execution of a conditional '``br``' instruction, the '``i1``'
3237 argument is evaluated. If the value is ``true``, control flows to the
3238 '``iftrue``' ``label`` argument. If "cond" is ``false``, control flows
3239 to the '``iffalse``' ``label`` argument.
3244 .. code-block:: llvm
3247 %cond = icmp eq i32 %a, %b
3248 br i1 %cond, label %IfEqual, label %IfUnequal
3256 '``switch``' Instruction
3257 ^^^^^^^^^^^^^^^^^^^^^^^^
3264 switch <intty> <value>, label <defaultdest> [ <intty> <val>, label <dest> ... ]
3269 The '``switch``' instruction is used to transfer control flow to one of
3270 several different places. It is a generalization of the '``br``'
3271 instruction, allowing a branch to occur to one of many possible
3277 The '``switch``' instruction uses three parameters: an integer
3278 comparison value '``value``', a default '``label``' destination, and an
3279 array of pairs of comparison value constants and '``label``'s. The table
3280 is not allowed to contain duplicate constant entries.
3285 The ``switch`` instruction specifies a table of values and destinations.
3286 When the '``switch``' instruction is executed, this table is searched
3287 for the given value. If the value is found, control flow is transferred
3288 to the corresponding destination; otherwise, control flow is transferred
3289 to the default destination.
3294 Depending on properties of the target machine and the particular
3295 ``switch`` instruction, this instruction may be code generated in
3296 different ways. For example, it could be generated as a series of
3297 chained conditional branches or with a lookup table.
3302 .. code-block:: llvm
3304 ; Emulate a conditional br instruction
3305 %Val = zext i1 %value to i32
3306 switch i32 %Val, label %truedest [ i32 0, label %falsedest ]
3308 ; Emulate an unconditional br instruction
3309 switch i32 0, label %dest [ ]
3311 ; Implement a jump table:
3312 switch i32 %val, label %otherwise [ i32 0, label %onzero
3314 i32 2, label %ontwo ]
3318 '``indirectbr``' Instruction
3319 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3326 indirectbr <somety>* <address>, [ label <dest1>, label <dest2>, ... ]
3331 The '``indirectbr``' instruction implements an indirect branch to a
3332 label within the current function, whose address is specified by
3333 "``address``". Address must be derived from a
3334 :ref:`blockaddress <blockaddress>` constant.
3339 The '``address``' argument is the address of the label to jump to. The
3340 rest of the arguments indicate the full set of possible destinations
3341 that the address may point to. Blocks are allowed to occur multiple
3342 times in the destination list, though this isn't particularly useful.
3344 This destination list is required so that dataflow analysis has an
3345 accurate understanding of the CFG.
3350 Control transfers to the block specified in the address argument. All
3351 possible destination blocks must be listed in the label list, otherwise
3352 this instruction has undefined behavior. This implies that jumps to
3353 labels defined in other functions have undefined behavior as well.
3358 This is typically implemented with a jump through a register.
3363 .. code-block:: llvm
3365 indirectbr i8* %Addr, [ label %bb1, label %bb2, label %bb3 ]
3369 '``invoke``' Instruction
3370 ^^^^^^^^^^^^^^^^^^^^^^^^
3377 <result> = invoke [cconv] [ret attrs] <ptr to function ty> <function ptr val>(<function args>) [fn attrs]
3378 to label <normal label> unwind label <exception label>
3383 The '``invoke``' instruction causes control to transfer to a specified
3384 function, with the possibility of control flow transfer to either the
3385 '``normal``' label or the '``exception``' label. If the callee function
3386 returns with the "``ret``" instruction, control flow will return to the
3387 "normal" label. If the callee (or any indirect callees) returns via the
3388 ":ref:`resume <i_resume>`" instruction or other exception handling
3389 mechanism, control is interrupted and continued at the dynamically
3390 nearest "exception" label.
3392 The '``exception``' label is a `landing
3393 pad <ExceptionHandling.html#overview>`_ for the exception. As such,
3394 '``exception``' label is required to have the
3395 ":ref:`landingpad <i_landingpad>`" instruction, which contains the
3396 information about the behavior of the program after unwinding happens,
3397 as its first non-PHI instruction. The restrictions on the
3398 "``landingpad``" instruction's tightly couples it to the "``invoke``"
3399 instruction, so that the important information contained within the
3400 "``landingpad``" instruction can't be lost through normal code motion.
3405 This instruction requires several arguments:
3407 #. The optional "cconv" marker indicates which :ref:`calling
3408 convention <callingconv>` the call should use. If none is
3409 specified, the call defaults to using C calling conventions.
3410 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
3411 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
3413 #. '``ptr to function ty``': shall be the signature of the pointer to
3414 function value being invoked. In most cases, this is a direct
3415 function invocation, but indirect ``invoke``'s are just as possible,
3416 branching off an arbitrary pointer to function value.
3417 #. '``function ptr val``': An LLVM value containing a pointer to a
3418 function to be invoked.
3419 #. '``function args``': argument list whose types match the function
3420 signature argument types and parameter attributes. All arguments must
3421 be of :ref:`first class <t_firstclass>` type. If the function signature
3422 indicates the function accepts a variable number of arguments, the
3423 extra arguments can be specified.
3424 #. '``normal label``': the label reached when the called function
3425 executes a '``ret``' instruction.
3426 #. '``exception label``': the label reached when a callee returns via
3427 the :ref:`resume <i_resume>` instruction or other exception handling
3429 #. The optional :ref:`function attributes <fnattrs>` list. Only
3430 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
3431 attributes are valid here.
3436 This instruction is designed to operate as a standard '``call``'
3437 instruction in most regards. The primary difference is that it
3438 establishes an association with a label, which is used by the runtime
3439 library to unwind the stack.
3441 This instruction is used in languages with destructors to ensure that
3442 proper cleanup is performed in the case of either a ``longjmp`` or a
3443 thrown exception. Additionally, this is important for implementation of
3444 '``catch``' clauses in high-level languages that support them.
3446 For the purposes of the SSA form, the definition of the value returned
3447 by the '``invoke``' instruction is deemed to occur on the edge from the
3448 current block to the "normal" label. If the callee unwinds then no
3449 return value is available.
3454 .. code-block:: llvm
3456 %retval = invoke i32 @Test(i32 15) to label %Continue
3457 unwind label %TestCleanup ; {i32}:retval set
3458 %retval = invoke coldcc i32 %Testfnptr(i32 15) to label %Continue
3459 unwind label %TestCleanup ; {i32}:retval set
3463 '``resume``' Instruction
3464 ^^^^^^^^^^^^^^^^^^^^^^^^
3471 resume <type> <value>
3476 The '``resume``' instruction is a terminator instruction that has no
3482 The '``resume``' instruction requires one argument, which must have the
3483 same type as the result of any '``landingpad``' instruction in the same
3489 The '``resume``' instruction resumes propagation of an existing
3490 (in-flight) exception whose unwinding was interrupted with a
3491 :ref:`landingpad <i_landingpad>` instruction.
3496 .. code-block:: llvm
3498 resume { i8*, i32 } %exn
3502 '``unreachable``' Instruction
3503 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
3515 The '``unreachable``' instruction has no defined semantics. This
3516 instruction is used to inform the optimizer that a particular portion of
3517 the code is not reachable. This can be used to indicate that the code
3518 after a no-return function cannot be reached, and other facts.
3523 The '``unreachable``' instruction has no defined semantics.
3530 Binary operators are used to do most of the computation in a program.
3531 They require two operands of the same type, execute an operation on
3532 them, and produce a single value. The operands might represent multiple
3533 data, as is the case with the :ref:`vector <t_vector>` data type. The
3534 result value has the same type as its operands.
3536 There are several different binary operators:
3540 '``add``' Instruction
3541 ^^^^^^^^^^^^^^^^^^^^^
3548 <result> = add <ty> <op1>, <op2> ; yields {ty}:result
3549 <result> = add nuw <ty> <op1>, <op2> ; yields {ty}:result
3550 <result> = add nsw <ty> <op1>, <op2> ; yields {ty}:result
3551 <result> = add nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3556 The '``add``' instruction returns the sum of its two operands.
3561 The two arguments to the '``add``' instruction must be
3562 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3563 arguments must have identical types.
3568 The value produced is the integer sum of the two operands.
3570 If the sum has unsigned overflow, the result returned is the
3571 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3574 Because LLVM integers use a two's complement representation, this
3575 instruction is appropriate for both signed and unsigned integers.
3577 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3578 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3579 result value of the ``add`` is a :ref:`poison value <poisonvalues>` if
3580 unsigned and/or signed overflow, respectively, occurs.
3585 .. code-block:: llvm
3587 <result> = add i32 4, %var ; yields {i32}:result = 4 + %var
3591 '``fadd``' Instruction
3592 ^^^^^^^^^^^^^^^^^^^^^^
3599 <result> = fadd [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3604 The '``fadd``' instruction returns the sum of its two operands.
3609 The two arguments to the '``fadd``' instruction must be :ref:`floating
3610 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3611 Both arguments must have identical types.
3616 The value produced is the floating point sum of the two operands. This
3617 instruction can also take any number of :ref:`fast-math flags <fastmath>`,
3618 which are optimization hints to enable otherwise unsafe floating point
3624 .. code-block:: llvm
3626 <result> = fadd float 4.0, %var ; yields {float}:result = 4.0 + %var
3628 '``sub``' Instruction
3629 ^^^^^^^^^^^^^^^^^^^^^
3636 <result> = sub <ty> <op1>, <op2> ; yields {ty}:result
3637 <result> = sub nuw <ty> <op1>, <op2> ; yields {ty}:result
3638 <result> = sub nsw <ty> <op1>, <op2> ; yields {ty}:result
3639 <result> = sub nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3644 The '``sub``' instruction returns the difference of its two operands.
3646 Note that the '``sub``' instruction is used to represent the '``neg``'
3647 instruction present in most other intermediate representations.
3652 The two arguments to the '``sub``' instruction must be
3653 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3654 arguments must have identical types.
3659 The value produced is the integer difference of the two operands.
3661 If the difference has unsigned overflow, the result returned is the
3662 mathematical result modulo 2\ :sup:`n`\ , where n is the bit width of
3665 Because LLVM integers use a two's complement representation, this
3666 instruction is appropriate for both signed and unsigned integers.
3668 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3669 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3670 result value of the ``sub`` is a :ref:`poison value <poisonvalues>` if
3671 unsigned and/or signed overflow, respectively, occurs.
3676 .. code-block:: llvm
3678 <result> = sub i32 4, %var ; yields {i32}:result = 4 - %var
3679 <result> = sub i32 0, %val ; yields {i32}:result = -%var
3683 '``fsub``' Instruction
3684 ^^^^^^^^^^^^^^^^^^^^^^
3691 <result> = fsub [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3696 The '``fsub``' instruction returns the difference of its two operands.
3698 Note that the '``fsub``' instruction is used to represent the '``fneg``'
3699 instruction present in most other intermediate representations.
3704 The two arguments to the '``fsub``' instruction must be :ref:`floating
3705 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3706 Both arguments must have identical types.
3711 The value produced is the floating point difference of the two operands.
3712 This instruction can also take any number of :ref:`fast-math
3713 flags <fastmath>`, which are optimization hints to enable otherwise
3714 unsafe floating point optimizations:
3719 .. code-block:: llvm
3721 <result> = fsub float 4.0, %var ; yields {float}:result = 4.0 - %var
3722 <result> = fsub float -0.0, %val ; yields {float}:result = -%var
3724 '``mul``' Instruction
3725 ^^^^^^^^^^^^^^^^^^^^^
3732 <result> = mul <ty> <op1>, <op2> ; yields {ty}:result
3733 <result> = mul nuw <ty> <op1>, <op2> ; yields {ty}:result
3734 <result> = mul nsw <ty> <op1>, <op2> ; yields {ty}:result
3735 <result> = mul nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
3740 The '``mul``' instruction returns the product of its two operands.
3745 The two arguments to the '``mul``' instruction must be
3746 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3747 arguments must have identical types.
3752 The value produced is the integer product of the two operands.
3754 If the result of the multiplication has unsigned overflow, the result
3755 returned is the mathematical result modulo 2\ :sup:`n`\ , where n is the
3756 bit width of the result.
3758 Because LLVM integers use a two's complement representation, and the
3759 result is the same width as the operands, this instruction returns the
3760 correct result for both signed and unsigned integers. If a full product
3761 (e.g. ``i32`` * ``i32`` -> ``i64``) is needed, the operands should be
3762 sign-extended or zero-extended as appropriate to the width of the full
3765 ``nuw`` and ``nsw`` stand for "No Unsigned Wrap" and "No Signed Wrap",
3766 respectively. If the ``nuw`` and/or ``nsw`` keywords are present, the
3767 result value of the ``mul`` is a :ref:`poison value <poisonvalues>` if
3768 unsigned and/or signed overflow, respectively, occurs.
3773 .. code-block:: llvm
3775 <result> = mul i32 4, %var ; yields {i32}:result = 4 * %var
3779 '``fmul``' Instruction
3780 ^^^^^^^^^^^^^^^^^^^^^^
3787 <result> = fmul [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3792 The '``fmul``' instruction returns the product of its two operands.
3797 The two arguments to the '``fmul``' instruction must be :ref:`floating
3798 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3799 Both arguments must have identical types.
3804 The value produced is the floating point product of the two operands.
3805 This instruction can also take any number of :ref:`fast-math
3806 flags <fastmath>`, which are optimization hints to enable otherwise
3807 unsafe floating point optimizations:
3812 .. code-block:: llvm
3814 <result> = fmul float 4.0, %var ; yields {float}:result = 4.0 * %var
3816 '``udiv``' Instruction
3817 ^^^^^^^^^^^^^^^^^^^^^^
3824 <result> = udiv <ty> <op1>, <op2> ; yields {ty}:result
3825 <result> = udiv exact <ty> <op1>, <op2> ; yields {ty}:result
3830 The '``udiv``' instruction returns the quotient of its two operands.
3835 The two arguments to the '``udiv``' instruction must be
3836 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3837 arguments must have identical types.
3842 The value produced is the unsigned integer quotient of the two operands.
3844 Note that unsigned integer division and signed integer division are
3845 distinct operations; for signed integer division, use '``sdiv``'.
3847 Division by zero leads to undefined behavior.
3849 If the ``exact`` keyword is present, the result value of the ``udiv`` is
3850 a :ref:`poison value <poisonvalues>` if %op1 is not a multiple of %op2 (as
3851 such, "((a udiv exact b) mul b) == a").
3856 .. code-block:: llvm
3858 <result> = udiv i32 4, %var ; yields {i32}:result = 4 / %var
3860 '``sdiv``' Instruction
3861 ^^^^^^^^^^^^^^^^^^^^^^
3868 <result> = sdiv <ty> <op1>, <op2> ; yields {ty}:result
3869 <result> = sdiv exact <ty> <op1>, <op2> ; yields {ty}:result
3874 The '``sdiv``' instruction returns the quotient of its two operands.
3879 The two arguments to the '``sdiv``' instruction must be
3880 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3881 arguments must have identical types.
3886 The value produced is the signed integer quotient of the two operands
3887 rounded towards zero.
3889 Note that signed integer division and unsigned integer division are
3890 distinct operations; for unsigned integer division, use '``udiv``'.
3892 Division by zero leads to undefined behavior. Overflow also leads to
3893 undefined behavior; this is a rare case, but can occur, for example, by
3894 doing a 32-bit division of -2147483648 by -1.
3896 If the ``exact`` keyword is present, the result value of the ``sdiv`` is
3897 a :ref:`poison value <poisonvalues>` if the result would be rounded.
3902 .. code-block:: llvm
3904 <result> = sdiv i32 4, %var ; yields {i32}:result = 4 / %var
3908 '``fdiv``' Instruction
3909 ^^^^^^^^^^^^^^^^^^^^^^
3916 <result> = fdiv [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
3921 The '``fdiv``' instruction returns the quotient of its two operands.
3926 The two arguments to the '``fdiv``' instruction must be :ref:`floating
3927 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
3928 Both arguments must have identical types.
3933 The value produced is the floating point quotient of the two operands.
3934 This instruction can also take any number of :ref:`fast-math
3935 flags <fastmath>`, which are optimization hints to enable otherwise
3936 unsafe floating point optimizations:
3941 .. code-block:: llvm
3943 <result> = fdiv float 4.0, %var ; yields {float}:result = 4.0 / %var
3945 '``urem``' Instruction
3946 ^^^^^^^^^^^^^^^^^^^^^^
3953 <result> = urem <ty> <op1>, <op2> ; yields {ty}:result
3958 The '``urem``' instruction returns the remainder from the unsigned
3959 division of its two arguments.
3964 The two arguments to the '``urem``' instruction must be
3965 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
3966 arguments must have identical types.
3971 This instruction returns the unsigned integer *remainder* of a division.
3972 This instruction always performs an unsigned division to get the
3975 Note that unsigned integer remainder and signed integer remainder are
3976 distinct operations; for signed integer remainder, use '``srem``'.
3978 Taking the remainder of a division by zero leads to undefined behavior.
3983 .. code-block:: llvm
3985 <result> = urem i32 4, %var ; yields {i32}:result = 4 % %var
3987 '``srem``' Instruction
3988 ^^^^^^^^^^^^^^^^^^^^^^
3995 <result> = srem <ty> <op1>, <op2> ; yields {ty}:result
4000 The '``srem``' instruction returns the remainder from the signed
4001 division of its two operands. This instruction can also take
4002 :ref:`vector <t_vector>` versions of the values in which case the elements
4008 The two arguments to the '``srem``' instruction must be
4009 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4010 arguments must have identical types.
4015 This instruction returns the *remainder* of a division (where the result
4016 is either zero or has the same sign as the dividend, ``op1``), not the
4017 *modulo* operator (where the result is either zero or has the same sign
4018 as the divisor, ``op2``) of a value. For more information about the
4019 difference, see `The Math
4020 Forum <http://mathforum.org/dr.math/problems/anne.4.28.99.html>`_. For a
4021 table of how this is implemented in various languages, please see
4023 operation <http://en.wikipedia.org/wiki/Modulo_operation>`_.
4025 Note that signed integer remainder and unsigned integer remainder are
4026 distinct operations; for unsigned integer remainder, use '``urem``'.
4028 Taking the remainder of a division by zero leads to undefined behavior.
4029 Overflow also leads to undefined behavior; this is a rare case, but can
4030 occur, for example, by taking the remainder of a 32-bit division of
4031 -2147483648 by -1. (The remainder doesn't actually overflow, but this
4032 rule lets srem be implemented using instructions that return both the
4033 result of the division and the remainder.)
4038 .. code-block:: llvm
4040 <result> = srem i32 4, %var ; yields {i32}:result = 4 % %var
4044 '``frem``' Instruction
4045 ^^^^^^^^^^^^^^^^^^^^^^
4052 <result> = frem [fast-math flags]* <ty> <op1>, <op2> ; yields {ty}:result
4057 The '``frem``' instruction returns the remainder from the division of
4063 The two arguments to the '``frem``' instruction must be :ref:`floating
4064 point <t_floating>` or :ref:`vector <t_vector>` of floating point values.
4065 Both arguments must have identical types.
4070 This instruction returns the *remainder* of a division. The remainder
4071 has the same sign as the dividend. This instruction can also take any
4072 number of :ref:`fast-math flags <fastmath>`, which are optimization hints
4073 to enable otherwise unsafe floating point optimizations:
4078 .. code-block:: llvm
4080 <result> = frem float 4.0, %var ; yields {float}:result = 4.0 % %var
4084 Bitwise Binary Operations
4085 -------------------------
4087 Bitwise binary operators are used to do various forms of bit-twiddling
4088 in a program. They are generally very efficient instructions and can
4089 commonly be strength reduced from other instructions. They require two
4090 operands of the same type, execute an operation on them, and produce a
4091 single value. The resulting value is the same type as its operands.
4093 '``shl``' Instruction
4094 ^^^^^^^^^^^^^^^^^^^^^
4101 <result> = shl <ty> <op1>, <op2> ; yields {ty}:result
4102 <result> = shl nuw <ty> <op1>, <op2> ; yields {ty}:result
4103 <result> = shl nsw <ty> <op1>, <op2> ; yields {ty}:result
4104 <result> = shl nuw nsw <ty> <op1>, <op2> ; yields {ty}:result
4109 The '``shl``' instruction returns the first operand shifted to the left
4110 a specified number of bits.
4115 Both arguments to the '``shl``' instruction must be the same
4116 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4117 '``op2``' is treated as an unsigned value.
4122 The value produced is ``op1`` \* 2\ :sup:`op2` mod 2\ :sup:`n`,
4123 where ``n`` is the width of the result. If ``op2`` is (statically or
4124 dynamically) negative or equal to or larger than the number of bits in
4125 ``op1``, the result is undefined. If the arguments are vectors, each
4126 vector element of ``op1`` is shifted by the corresponding shift amount
4129 If the ``nuw`` keyword is present, then the shift produces a :ref:`poison
4130 value <poisonvalues>` if it shifts out any non-zero bits. If the
4131 ``nsw`` keyword is present, then the shift produces a :ref:`poison
4132 value <poisonvalues>` if it shifts out any bits that disagree with the
4133 resultant sign bit. As such, NUW/NSW have the same semantics as they
4134 would if the shift were expressed as a mul instruction with the same
4135 nsw/nuw bits in (mul %op1, (shl 1, %op2)).
4140 .. code-block:: llvm
4142 <result> = shl i32 4, %var ; yields {i32}: 4 << %var
4143 <result> = shl i32 4, 2 ; yields {i32}: 16
4144 <result> = shl i32 1, 10 ; yields {i32}: 1024
4145 <result> = shl i32 1, 32 ; undefined
4146 <result> = shl <2 x i32> < i32 1, i32 1>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 2, i32 4>
4148 '``lshr``' Instruction
4149 ^^^^^^^^^^^^^^^^^^^^^^
4156 <result> = lshr <ty> <op1>, <op2> ; yields {ty}:result
4157 <result> = lshr exact <ty> <op1>, <op2> ; yields {ty}:result
4162 The '``lshr``' instruction (logical shift right) returns the first
4163 operand shifted to the right a specified number of bits with zero fill.
4168 Both arguments to the '``lshr``' instruction must be the same
4169 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4170 '``op2``' is treated as an unsigned value.
4175 This instruction always performs a logical shift right operation. The
4176 most significant bits of the result will be filled with zero bits after
4177 the shift. If ``op2`` is (statically or dynamically) equal to or larger
4178 than the number of bits in ``op1``, the result is undefined. If the
4179 arguments are vectors, each vector element of ``op1`` is shifted by the
4180 corresponding shift amount in ``op2``.
4182 If the ``exact`` keyword is present, the result value of the ``lshr`` is
4183 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4189 .. code-block:: llvm
4191 <result> = lshr i32 4, 1 ; yields {i32}:result = 2
4192 <result> = lshr i32 4, 2 ; yields {i32}:result = 1
4193 <result> = lshr i8 4, 3 ; yields {i8}:result = 0
4194 <result> = lshr i8 -2, 1 ; yields {i8}:result = 0x7F
4195 <result> = lshr i32 1, 32 ; undefined
4196 <result> = lshr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 2> ; yields: result=<2 x i32> < i32 0x7FFFFFFF, i32 1>
4198 '``ashr``' Instruction
4199 ^^^^^^^^^^^^^^^^^^^^^^
4206 <result> = ashr <ty> <op1>, <op2> ; yields {ty}:result
4207 <result> = ashr exact <ty> <op1>, <op2> ; yields {ty}:result
4212 The '``ashr``' instruction (arithmetic shift right) returns the first
4213 operand shifted to the right a specified number of bits with sign
4219 Both arguments to the '``ashr``' instruction must be the same
4220 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer type.
4221 '``op2``' is treated as an unsigned value.
4226 This instruction always performs an arithmetic shift right operation,
4227 The most significant bits of the result will be filled with the sign bit
4228 of ``op1``. If ``op2`` is (statically or dynamically) equal to or larger
4229 than the number of bits in ``op1``, the result is undefined. If the
4230 arguments are vectors, each vector element of ``op1`` is shifted by the
4231 corresponding shift amount in ``op2``.
4233 If the ``exact`` keyword is present, the result value of the ``ashr`` is
4234 a :ref:`poison value <poisonvalues>` if any of the bits shifted out are
4240 .. code-block:: llvm
4242 <result> = ashr i32 4, 1 ; yields {i32}:result = 2
4243 <result> = ashr i32 4, 2 ; yields {i32}:result = 1
4244 <result> = ashr i8 4, 3 ; yields {i8}:result = 0
4245 <result> = ashr i8 -2, 1 ; yields {i8}:result = -1
4246 <result> = ashr i32 1, 32 ; undefined
4247 <result> = ashr <2 x i32> < i32 -2, i32 4>, < i32 1, i32 3> ; yields: result=<2 x i32> < i32 -1, i32 0>
4249 '``and``' Instruction
4250 ^^^^^^^^^^^^^^^^^^^^^
4257 <result> = and <ty> <op1>, <op2> ; yields {ty}:result
4262 The '``and``' instruction returns the bitwise logical and of its two
4268 The two arguments to the '``and``' instruction must be
4269 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4270 arguments must have identical types.
4275 The truth table used for the '``and``' instruction is:
4292 .. code-block:: llvm
4294 <result> = and i32 4, %var ; yields {i32}:result = 4 & %var
4295 <result> = and i32 15, 40 ; yields {i32}:result = 8
4296 <result> = and i32 4, 8 ; yields {i32}:result = 0
4298 '``or``' Instruction
4299 ^^^^^^^^^^^^^^^^^^^^
4306 <result> = or <ty> <op1>, <op2> ; yields {ty}:result
4311 The '``or``' instruction returns the bitwise logical inclusive or of its
4317 The two arguments to the '``or``' instruction must be
4318 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4319 arguments must have identical types.
4324 The truth table used for the '``or``' instruction is:
4343 <result> = or i32 4, %var ; yields {i32}:result = 4 | %var
4344 <result> = or i32 15, 40 ; yields {i32}:result = 47
4345 <result> = or i32 4, 8 ; yields {i32}:result = 12
4347 '``xor``' Instruction
4348 ^^^^^^^^^^^^^^^^^^^^^
4355 <result> = xor <ty> <op1>, <op2> ; yields {ty}:result
4360 The '``xor``' instruction returns the bitwise logical exclusive or of
4361 its two operands. The ``xor`` is used to implement the "one's
4362 complement" operation, which is the "~" operator in C.
4367 The two arguments to the '``xor``' instruction must be
4368 :ref:`integer <t_integer>` or :ref:`vector <t_vector>` of integer values. Both
4369 arguments must have identical types.
4374 The truth table used for the '``xor``' instruction is:
4391 .. code-block:: llvm
4393 <result> = xor i32 4, %var ; yields {i32}:result = 4 ^ %var
4394 <result> = xor i32 15, 40 ; yields {i32}:result = 39
4395 <result> = xor i32 4, 8 ; yields {i32}:result = 12
4396 <result> = xor i32 %V, -1 ; yields {i32}:result = ~%V
4401 LLVM supports several instructions to represent vector operations in a
4402 target-independent manner. These instructions cover the element-access
4403 and vector-specific operations needed to process vectors effectively.
4404 While LLVM does directly support these vector operations, many
4405 sophisticated algorithms will want to use target-specific intrinsics to
4406 take full advantage of a specific target.
4408 .. _i_extractelement:
4410 '``extractelement``' Instruction
4411 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4418 <result> = extractelement <n x <ty>> <val>, i32 <idx> ; yields <ty>
4423 The '``extractelement``' instruction extracts a single scalar element
4424 from a vector at a specified index.
4429 The first operand of an '``extractelement``' instruction is a value of
4430 :ref:`vector <t_vector>` type. The second operand is an index indicating
4431 the position from which to extract the element. The index may be a
4437 The result is a scalar of the same type as the element type of ``val``.
4438 Its value is the value at position ``idx`` of ``val``. If ``idx``
4439 exceeds the length of ``val``, the results are undefined.
4444 .. code-block:: llvm
4446 <result> = extractelement <4 x i32> %vec, i32 0 ; yields i32
4448 .. _i_insertelement:
4450 '``insertelement``' Instruction
4451 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4458 <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> ; yields <n x <ty>>
4463 The '``insertelement``' instruction inserts a scalar element into a
4464 vector at a specified index.
4469 The first operand of an '``insertelement``' instruction is a value of
4470 :ref:`vector <t_vector>` type. The second operand is a scalar value whose
4471 type must equal the element type of the first operand. The third operand
4472 is an index indicating the position at which to insert the value. The
4473 index may be a variable.
4478 The result is a vector of the same type as ``val``. Its element values
4479 are those of ``val`` except at position ``idx``, where it gets the value
4480 ``elt``. If ``idx`` exceeds the length of ``val``, the results are
4486 .. code-block:: llvm
4488 <result> = insertelement <4 x i32> %vec, i32 1, i32 0 ; yields <4 x i32>
4490 .. _i_shufflevector:
4492 '``shufflevector``' Instruction
4493 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4500 <result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask> ; yields <m x <ty>>
4505 The '``shufflevector``' instruction constructs a permutation of elements
4506 from two input vectors, returning a vector with the same element type as
4507 the input and length that is the same as the shuffle mask.
4512 The first two operands of a '``shufflevector``' instruction are vectors
4513 with the same type. The third argument is a shuffle mask whose element
4514 type is always 'i32'. The result of the instruction is a vector whose
4515 length is the same as the shuffle mask and whose element type is the
4516 same as the element type of the first two operands.
4518 The shuffle mask operand is required to be a constant vector with either
4519 constant integer or undef values.
4524 The elements of the two input vectors are numbered from left to right
4525 across both of the vectors. The shuffle mask operand specifies, for each
4526 element of the result vector, which element of the two input vectors the
4527 result element gets. The element selector may be undef (meaning "don't
4528 care") and the second operand may be undef if performing a shuffle from
4534 .. code-block:: llvm
4536 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4537 <4 x i32> <i32 0, i32 4, i32 1, i32 5> ; yields <4 x i32>
4538 <result> = shufflevector <4 x i32> %v1, <4 x i32> undef,
4539 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32> - Identity shuffle.
4540 <result> = shufflevector <8 x i32> %v1, <8 x i32> undef,
4541 <4 x i32> <i32 0, i32 1, i32 2, i32 3> ; yields <4 x i32>
4542 <result> = shufflevector <4 x i32> %v1, <4 x i32> %v2,
4543 <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7 > ; yields <8 x i32>
4545 Aggregate Operations
4546 --------------------
4548 LLVM supports several instructions for working with
4549 :ref:`aggregate <t_aggregate>` values.
4553 '``extractvalue``' Instruction
4554 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4561 <result> = extractvalue <aggregate type> <val>, <idx>{, <idx>}*
4566 The '``extractvalue``' instruction extracts the value of a member field
4567 from an :ref:`aggregate <t_aggregate>` value.
4572 The first operand of an '``extractvalue``' instruction is a value of
4573 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The operands are
4574 constant indices to specify which value to extract in a similar manner
4575 as indices in a '``getelementptr``' instruction.
4577 The major differences to ``getelementptr`` indexing are:
4579 - Since the value being indexed is not a pointer, the first index is
4580 omitted and assumed to be zero.
4581 - At least one index must be specified.
4582 - Not only struct indices but also array indices must be in bounds.
4587 The result is the value at the position in the aggregate specified by
4593 .. code-block:: llvm
4595 <result> = extractvalue {i32, float} %agg, 0 ; yields i32
4599 '``insertvalue``' Instruction
4600 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
4607 <result> = insertvalue <aggregate type> <val>, <ty> <elt>, <idx>{, <idx>}* ; yields <aggregate type>
4612 The '``insertvalue``' instruction inserts a value into a member field in
4613 an :ref:`aggregate <t_aggregate>` value.
4618 The first operand of an '``insertvalue``' instruction is a value of
4619 :ref:`struct <t_struct>` or :ref:`array <t_array>` type. The second operand is
4620 a first-class value to insert. The following operands are constant
4621 indices indicating the position at which to insert the value in a
4622 similar manner as indices in a '``extractvalue``' instruction. The value
4623 to insert must have the same type as the value identified by the
4629 The result is an aggregate of the same type as ``val``. Its value is
4630 that of ``val`` except that the value at the position specified by the
4631 indices is that of ``elt``.
4636 .. code-block:: llvm
4638 %agg1 = insertvalue {i32, float} undef, i32 1, 0 ; yields {i32 1, float undef}
4639 %agg2 = insertvalue {i32, float} %agg1, float %val, 1 ; yields {i32 1, float %val}
4640 %agg3 = insertvalue {i32, {float}} %agg1, float %val, 1, 0 ; yields {i32 1, float %val}
4644 Memory Access and Addressing Operations
4645 ---------------------------------------
4647 A key design point of an SSA-based representation is how it represents
4648 memory. In LLVM, no memory locations are in SSA form, which makes things
4649 very simple. This section describes how to read, write, and allocate
4654 '``alloca``' Instruction
4655 ^^^^^^^^^^^^^^^^^^^^^^^^
4662 <result> = alloca <type>[, <ty> <NumElements>][, align <alignment>] ; yields {type*}:result
4667 The '``alloca``' instruction allocates memory on the stack frame of the
4668 currently executing function, to be automatically released when this
4669 function returns to its caller. The object is always allocated in the
4670 generic address space (address space zero).
4675 The '``alloca``' instruction allocates ``sizeof(<type>)*NumElements``
4676 bytes of memory on the runtime stack, returning a pointer of the
4677 appropriate type to the program. If "NumElements" is specified, it is
4678 the number of elements allocated, otherwise "NumElements" is defaulted
4679 to be one. If a constant alignment is specified, the value result of the
4680 allocation is guaranteed to be aligned to at least that boundary. If not
4681 specified, or if zero, the target can choose to align the allocation on
4682 any convenient boundary compatible with the type.
4684 '``type``' may be any sized type.
4689 Memory is allocated; a pointer is returned. The operation is undefined
4690 if there is insufficient stack space for the allocation. '``alloca``'d
4691 memory is automatically released when the function returns. The
4692 '``alloca``' instruction is commonly used to represent automatic
4693 variables that must have an address available. When the function returns
4694 (either with the ``ret`` or ``resume`` instructions), the memory is
4695 reclaimed. Allocating zero bytes is legal, but the result is undefined.
4696 The order in which memory is allocated (ie., which way the stack grows)
4702 .. code-block:: llvm
4704 %ptr = alloca i32 ; yields {i32*}:ptr
4705 %ptr = alloca i32, i32 4 ; yields {i32*}:ptr
4706 %ptr = alloca i32, i32 4, align 1024 ; yields {i32*}:ptr
4707 %ptr = alloca i32, align 1024 ; yields {i32*}:ptr
4711 '``load``' Instruction
4712 ^^^^^^^^^^^^^^^^^^^^^^
4719 <result> = load [volatile] <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>][, !invariant.load !<index>]
4720 <result> = load atomic [volatile] <ty>* <pointer> [singlethread] <ordering>, align <alignment>
4721 !<index> = !{ i32 1 }
4726 The '``load``' instruction is used to read from memory.
4731 The argument to the ``load`` instruction specifies the memory address
4732 from which to load. The pointer must point to a :ref:`first
4733 class <t_firstclass>` type. If the ``load`` is marked as ``volatile``,
4734 then the optimizer is not allowed to modify the number or order of
4735 execution of this ``load`` with other :ref:`volatile
4736 operations <volatile>`.
4738 If the ``load`` is marked as ``atomic``, it takes an extra
4739 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4740 ``release`` and ``acq_rel`` orderings are not valid on ``load``
4741 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4742 when they may see multiple atomic stores. The type of the pointee must
4743 be an integer type whose bit width is a power of two greater than or
4744 equal to eight and less than or equal to a target-specific size limit.
4745 ``align`` must be explicitly specified on atomic loads, and the load has
4746 undefined behavior if the alignment is not set to a value which is at
4747 least the size in bytes of the pointee. ``!nontemporal`` does not have
4748 any defined semantics for atomic loads.
4750 The optional constant ``align`` argument specifies the alignment of the
4751 operation (that is, the alignment of the memory address). A value of 0
4752 or an omitted ``align`` argument means that the operation has the ABI
4753 alignment for the target. It is the responsibility of the code emitter
4754 to ensure that the alignment information is correct. Overestimating the
4755 alignment results in undefined behavior. Underestimating the alignment
4756 may produce less efficient code. An alignment of 1 is always safe.
4758 The optional ``!nontemporal`` metadata must reference a single
4759 metadata name ``<index>`` corresponding to a metadata node with one
4760 ``i32`` entry of value 1. The existence of the ``!nontemporal``
4761 metadata on the instruction tells the optimizer and code generator
4762 that this load is not expected to be reused in the cache. The code
4763 generator may select special instructions to save cache bandwidth, such
4764 as the ``MOVNT`` instruction on x86.
4766 The optional ``!invariant.load`` metadata must reference a single
4767 metadata name ``<index>`` corresponding to a metadata node with no
4768 entries. The existence of the ``!invariant.load`` metadata on the
4769 instruction tells the optimizer and code generator that this load
4770 address points to memory which does not change value during program
4771 execution. The optimizer may then move this load around, for example, by
4772 hoisting it out of loops using loop invariant code motion.
4777 The location of memory pointed to is loaded. If the value being loaded
4778 is of scalar type then the number of bytes read does not exceed the
4779 minimum number of bytes needed to hold all bits of the type. For
4780 example, loading an ``i24`` reads at most three bytes. When loading a
4781 value of a type like ``i20`` with a size that is not an integral number
4782 of bytes, the result is undefined if the value was not originally
4783 written using a store of the same type.
4788 .. code-block:: llvm
4790 %ptr = alloca i32 ; yields {i32*}:ptr
4791 store i32 3, i32* %ptr ; yields {void}
4792 %val = load i32* %ptr ; yields {i32}:val = i32 3
4796 '``store``' Instruction
4797 ^^^^^^^^^^^^^^^^^^^^^^^
4804 store [volatile] <ty> <value>, <ty>* <pointer>[, align <alignment>][, !nontemporal !<index>] ; yields {void}
4805 store atomic [volatile] <ty> <value>, <ty>* <pointer> [singlethread] <ordering>, align <alignment> ; yields {void}
4810 The '``store``' instruction is used to write to memory.
4815 There are two arguments to the ``store`` instruction: a value to store
4816 and an address at which to store it. The type of the ``<pointer>``
4817 operand must be a pointer to the :ref:`first class <t_firstclass>` type of
4818 the ``<value>`` operand. If the ``store`` is marked as ``volatile``,
4819 then the optimizer is not allowed to modify the number or order of
4820 execution of this ``store`` with other :ref:`volatile
4821 operations <volatile>`.
4823 If the ``store`` is marked as ``atomic``, it takes an extra
4824 :ref:`ordering <ordering>` and optional ``singlethread`` argument. The
4825 ``acquire`` and ``acq_rel`` orderings aren't valid on ``store``
4826 instructions. Atomic loads produce :ref:`defined <memmodel>` results
4827 when they may see multiple atomic stores. The type of the pointee must
4828 be an integer type whose bit width is a power of two greater than or
4829 equal to eight and less than or equal to a target-specific size limit.
4830 ``align`` must be explicitly specified on atomic stores, and the store
4831 has undefined behavior if the alignment is not set to a value which is
4832 at least the size in bytes of the pointee. ``!nontemporal`` does not
4833 have any defined semantics for atomic stores.
4835 The optional constant ``align`` argument specifies the alignment of the
4836 operation (that is, the alignment of the memory address). A value of 0
4837 or an omitted ``align`` argument means that the operation has the ABI
4838 alignment for the target. It is the responsibility of the code emitter
4839 to ensure that the alignment information is correct. Overestimating the
4840 alignment results in undefined behavior. Underestimating the
4841 alignment may produce less efficient code. An alignment of 1 is always
4844 The optional ``!nontemporal`` metadata must reference a single metadata
4845 name ``<index>`` corresponding to a metadata node with one ``i32`` entry of
4846 value 1. The existence of the ``!nontemporal`` metadata on the instruction
4847 tells the optimizer and code generator that this load is not expected to
4848 be reused in the cache. The code generator may select special
4849 instructions to save cache bandwidth, such as the MOVNT instruction on
4855 The contents of memory are updated to contain ``<value>`` at the
4856 location specified by the ``<pointer>`` operand. If ``<value>`` is
4857 of scalar type then the number of bytes written does not exceed the
4858 minimum number of bytes needed to hold all bits of the type. For
4859 example, storing an ``i24`` writes at most three bytes. When writing a
4860 value of a type like ``i20`` with a size that is not an integral number
4861 of bytes, it is unspecified what happens to the extra bits that do not
4862 belong to the type, but they will typically be overwritten.
4867 .. code-block:: llvm
4869 %ptr = alloca i32 ; yields {i32*}:ptr
4870 store i32 3, i32* %ptr ; yields {void}
4871 %val = load i32* %ptr ; yields {i32}:val = i32 3
4875 '``fence``' Instruction
4876 ^^^^^^^^^^^^^^^^^^^^^^^
4883 fence [singlethread] <ordering> ; yields {void}
4888 The '``fence``' instruction is used to introduce happens-before edges
4894 '``fence``' instructions take an :ref:`ordering <ordering>` argument which
4895 defines what *synchronizes-with* edges they add. They can only be given
4896 ``acquire``, ``release``, ``acq_rel``, and ``seq_cst`` orderings.
4901 A fence A which has (at least) ``release`` ordering semantics
4902 *synchronizes with* a fence B with (at least) ``acquire`` ordering
4903 semantics if and only if there exist atomic operations X and Y, both
4904 operating on some atomic object M, such that A is sequenced before X, X
4905 modifies M (either directly or through some side effect of a sequence
4906 headed by X), Y is sequenced before B, and Y observes M. This provides a
4907 *happens-before* dependency between A and B. Rather than an explicit
4908 ``fence``, one (but not both) of the atomic operations X or Y might
4909 provide a ``release`` or ``acquire`` (resp.) ordering constraint and
4910 still *synchronize-with* the explicit ``fence`` and establish the
4911 *happens-before* edge.
4913 A ``fence`` which has ``seq_cst`` ordering, in addition to having both
4914 ``acquire`` and ``release`` semantics specified above, participates in
4915 the global program order of other ``seq_cst`` operations and/or fences.
4917 The optional ":ref:`singlethread <singlethread>`" argument specifies
4918 that the fence only synchronizes with other fences in the same thread.
4919 (This is useful for interacting with signal handlers.)
4924 .. code-block:: llvm
4926 fence acquire ; yields {void}
4927 fence singlethread seq_cst ; yields {void}
4931 '``cmpxchg``' Instruction
4932 ^^^^^^^^^^^^^^^^^^^^^^^^^
4939 cmpxchg [volatile] <ty>* <pointer>, <ty> <cmp>, <ty> <new> [singlethread] <ordering> ; yields {ty}
4944 The '``cmpxchg``' instruction is used to atomically modify memory. It
4945 loads a value in memory and compares it to a given value. If they are
4946 equal, it stores a new value into the memory.
4951 There are three arguments to the '``cmpxchg``' instruction: an address
4952 to operate on, a value to compare to the value currently be at that
4953 address, and a new value to place at that address if the compared values
4954 are equal. The type of '<cmp>' must be an integer type whose bit width
4955 is a power of two greater than or equal to eight and less than or equal
4956 to a target-specific size limit. '<cmp>' and '<new>' must have the same
4957 type, and the type of '<pointer>' must be a pointer to that type. If the
4958 ``cmpxchg`` is marked as ``volatile``, then the optimizer is not allowed
4959 to modify the number or order of execution of this ``cmpxchg`` with
4960 other :ref:`volatile operations <volatile>`.
4962 The :ref:`ordering <ordering>` argument specifies how this ``cmpxchg``
4963 synchronizes with other atomic operations.
4965 The optional "``singlethread``" argument declares that the ``cmpxchg``
4966 is only atomic with respect to code (usually signal handlers) running in
4967 the same thread as the ``cmpxchg``. Otherwise the cmpxchg is atomic with
4968 respect to all other code in the system.
4970 The pointer passed into cmpxchg must have alignment greater than or
4971 equal to the size in memory of the operand.
4976 The contents of memory at the location specified by the '``<pointer>``'
4977 operand is read and compared to '``<cmp>``'; if the read value is the
4978 equal, '``<new>``' is written. The original value at the location is
4981 A successful ``cmpxchg`` is a read-modify-write instruction for the purpose
4982 of identifying release sequences. A failed ``cmpxchg`` is equivalent to an
4983 atomic load with an ordering parameter determined by dropping any
4984 ``release`` part of the ``cmpxchg``'s ordering.
4989 .. code-block:: llvm
4992 %orig = atomic load i32* %ptr unordered ; yields {i32}
4996 %cmp = phi i32 [ %orig, %entry ], [%old, %loop]
4997 %squared = mul i32 %cmp, %cmp
4998 %old = cmpxchg i32* %ptr, i32 %cmp, i32 %squared ; yields {i32}
4999 %success = icmp eq i32 %cmp, %old
5000 br i1 %success, label %done, label %loop
5007 '``atomicrmw``' Instruction
5008 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
5015 atomicrmw [volatile] <operation> <ty>* <pointer>, <ty> <value> [singlethread] <ordering> ; yields {ty}
5020 The '``atomicrmw``' instruction is used to atomically modify memory.
5025 There are three arguments to the '``atomicrmw``' instruction: an
5026 operation to apply, an address whose value to modify, an argument to the
5027 operation. The operation must be one of the following keywords:
5041 The type of '<value>' must be an integer type whose bit width is a power
5042 of two greater than or equal to eight and less than or equal to a
5043 target-specific size limit. The type of the '``<pointer>``' operand must
5044 be a pointer to that type. If the ``atomicrmw`` is marked as
5045 ``volatile``, then the optimizer is not allowed to modify the number or
5046 order of execution of this ``atomicrmw`` with other :ref:`volatile
5047 operations <volatile>`.
5052 The contents of memory at the location specified by the '``<pointer>``'
5053 operand are atomically read, modified, and written back. The original
5054 value at the location is returned. The modification is specified by the
5057 - xchg: ``*ptr = val``
5058 - add: ``*ptr = *ptr + val``
5059 - sub: ``*ptr = *ptr - val``
5060 - and: ``*ptr = *ptr & val``
5061 - nand: ``*ptr = ~(*ptr & val)``
5062 - or: ``*ptr = *ptr | val``
5063 - xor: ``*ptr = *ptr ^ val``
5064 - max: ``*ptr = *ptr > val ? *ptr : val`` (using a signed comparison)
5065 - min: ``*ptr = *ptr < val ? *ptr : val`` (using a signed comparison)
5066 - umax: ``*ptr = *ptr > val ? *ptr : val`` (using an unsigned
5068 - umin: ``*ptr = *ptr < val ? *ptr : val`` (using an unsigned
5074 .. code-block:: llvm
5076 %old = atomicrmw add i32* %ptr, i32 1 acquire ; yields {i32}
5078 .. _i_getelementptr:
5080 '``getelementptr``' Instruction
5081 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5088 <result> = getelementptr <pty>* <ptrval>{, <ty> <idx>}*
5089 <result> = getelementptr inbounds <pty>* <ptrval>{, <ty> <idx>}*
5090 <result> = getelementptr <ptr vector> ptrval, <vector index type> idx
5095 The '``getelementptr``' instruction is used to get the address of a
5096 subelement of an :ref:`aggregate <t_aggregate>` data structure. It performs
5097 address calculation only and does not access memory.
5102 The first argument is always a pointer or a vector of pointers, and
5103 forms the basis of the calculation. The remaining arguments are indices
5104 that indicate which of the elements of the aggregate object are indexed.
5105 The interpretation of each index is dependent on the type being indexed
5106 into. The first index always indexes the pointer value given as the
5107 first argument, the second index indexes a value of the type pointed to
5108 (not necessarily the value directly pointed to, since the first index
5109 can be non-zero), etc. The first type indexed into must be a pointer
5110 value, subsequent types can be arrays, vectors, and structs. Note that
5111 subsequent types being indexed into can never be pointers, since that
5112 would require loading the pointer before continuing calculation.
5114 The type of each index argument depends on the type it is indexing into.
5115 When indexing into a (optionally packed) structure, only ``i32`` integer
5116 **constants** are allowed (when using a vector of indices they must all
5117 be the **same** ``i32`` integer constant). When indexing into an array,
5118 pointer or vector, integers of any width are allowed, and they are not
5119 required to be constant. These integers are treated as signed values
5122 For example, let's consider a C code fragment and how it gets compiled
5138 int *foo(struct ST *s) {
5139 return &s[1].Z.B[5][13];
5142 The LLVM code generated by Clang is:
5144 .. code-block:: llvm
5146 %struct.RT = type { i8, [10 x [20 x i32]], i8 }
5147 %struct.ST = type { i32, double, %struct.RT }
5149 define i32* @foo(%struct.ST* %s) nounwind uwtable readnone optsize ssp {
5151 %arrayidx = getelementptr inbounds %struct.ST* %s, i64 1, i32 2, i32 1, i64 5, i64 13
5158 In the example above, the first index is indexing into the
5159 '``%struct.ST*``' type, which is a pointer, yielding a '``%struct.ST``'
5160 = '``{ i32, double, %struct.RT }``' type, a structure. The second index
5161 indexes into the third element of the structure, yielding a
5162 '``%struct.RT``' = '``{ i8 , [10 x [20 x i32]], i8 }``' type, another
5163 structure. The third index indexes into the second element of the
5164 structure, yielding a '``[10 x [20 x i32]]``' type, an array. The two
5165 dimensions of the array are subscripted into, yielding an '``i32``'
5166 type. The '``getelementptr``' instruction returns a pointer to this
5167 element, thus computing a value of '``i32*``' type.
5169 Note that it is perfectly legal to index partially through a structure,
5170 returning a pointer to an inner element. Because of this, the LLVM code
5171 for the given testcase is equivalent to:
5173 .. code-block:: llvm
5175 define i32* @foo(%struct.ST* %s) {
5176 %t1 = getelementptr %struct.ST* %s, i32 1 ; yields %struct.ST*:%t1
5177 %t2 = getelementptr %struct.ST* %t1, i32 0, i32 2 ; yields %struct.RT*:%t2
5178 %t3 = getelementptr %struct.RT* %t2, i32 0, i32 1 ; yields [10 x [20 x i32]]*:%t3
5179 %t4 = getelementptr [10 x [20 x i32]]* %t3, i32 0, i32 5 ; yields [20 x i32]*:%t4
5180 %t5 = getelementptr [20 x i32]* %t4, i32 0, i32 13 ; yields i32*:%t5
5184 If the ``inbounds`` keyword is present, the result value of the
5185 ``getelementptr`` is a :ref:`poison value <poisonvalues>` if the base
5186 pointer is not an *in bounds* address of an allocated object, or if any
5187 of the addresses that would be formed by successive addition of the
5188 offsets implied by the indices to the base address with infinitely
5189 precise signed arithmetic are not an *in bounds* address of that
5190 allocated object. The *in bounds* addresses for an allocated object are
5191 all the addresses that point into the object, plus the address one byte
5192 past the end. In cases where the base is a vector of pointers the
5193 ``inbounds`` keyword applies to each of the computations element-wise.
5195 If the ``inbounds`` keyword is not present, the offsets are added to the
5196 base address with silently-wrapping two's complement arithmetic. If the
5197 offsets have a different width from the pointer, they are sign-extended
5198 or truncated to the width of the pointer. The result value of the
5199 ``getelementptr`` may be outside the object pointed to by the base
5200 pointer. The result value may not necessarily be used to access memory
5201 though, even if it happens to point into allocated storage. See the
5202 :ref:`Pointer Aliasing Rules <pointeraliasing>` section for more
5205 The getelementptr instruction is often confusing. For some more insight
5206 into how it works, see :doc:`the getelementptr FAQ <GetElementPtr>`.
5211 .. code-block:: llvm
5213 ; yields [12 x i8]*:aptr
5214 %aptr = getelementptr {i32, [12 x i8]}* %saptr, i64 0, i32 1
5216 %vptr = getelementptr {i32, <2 x i8>}* %svptr, i64 0, i32 1, i32 1
5218 %eptr = getelementptr [12 x i8]* %aptr, i64 0, i32 1
5220 %iptr = getelementptr [10 x i32]* @arr, i16 0, i16 0
5222 In cases where the pointer argument is a vector of pointers, each index
5223 must be a vector with the same number of elements. For example:
5225 .. code-block:: llvm
5227 %A = getelementptr <4 x i8*> %ptrs, <4 x i64> %offsets,
5229 Conversion Operations
5230 ---------------------
5232 The instructions in this category are the conversion instructions
5233 (casting) which all take a single operand and a type. They perform
5234 various bit conversions on the operand.
5236 '``trunc .. to``' Instruction
5237 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5244 <result> = trunc <ty> <value> to <ty2> ; yields ty2
5249 The '``trunc``' instruction truncates its operand to the type ``ty2``.
5254 The '``trunc``' instruction takes a value to trunc, and a type to trunc
5255 it to. Both types must be of :ref:`integer <t_integer>` types, or vectors
5256 of the same number of integers. The bit size of the ``value`` must be
5257 larger than the bit size of the destination type, ``ty2``. Equal sized
5258 types are not allowed.
5263 The '``trunc``' instruction truncates the high order bits in ``value``
5264 and converts the remaining bits to ``ty2``. Since the source size must
5265 be larger than the destination size, ``trunc`` cannot be a *no-op cast*.
5266 It will always truncate bits.
5271 .. code-block:: llvm
5273 %X = trunc i32 257 to i8 ; yields i8:1
5274 %Y = trunc i32 123 to i1 ; yields i1:true
5275 %Z = trunc i32 122 to i1 ; yields i1:false
5276 %W = trunc <2 x i16> <i16 8, i16 7> to <2 x i8> ; yields <i8 8, i8 7>
5278 '``zext .. to``' Instruction
5279 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5286 <result> = zext <ty> <value> to <ty2> ; yields ty2
5291 The '``zext``' instruction zero extends its operand to type ``ty2``.
5296 The '``zext``' instruction takes a value to cast, and a type to cast it
5297 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5298 the same number of integers. The bit size of the ``value`` must be
5299 smaller than the bit size of the destination type, ``ty2``.
5304 The ``zext`` fills the high order bits of the ``value`` with zero bits
5305 until it reaches the size of the destination type, ``ty2``.
5307 When zero extending from i1, the result will always be either 0 or 1.
5312 .. code-block:: llvm
5314 %X = zext i32 257 to i64 ; yields i64:257
5315 %Y = zext i1 true to i32 ; yields i32:1
5316 %Z = zext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5318 '``sext .. to``' Instruction
5319 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5326 <result> = sext <ty> <value> to <ty2> ; yields ty2
5331 The '``sext``' sign extends ``value`` to the type ``ty2``.
5336 The '``sext``' instruction takes a value to cast, and a type to cast it
5337 to. Both types must be of :ref:`integer <t_integer>` types, or vectors of
5338 the same number of integers. The bit size of the ``value`` must be
5339 smaller than the bit size of the destination type, ``ty2``.
5344 The '``sext``' instruction performs a sign extension by copying the sign
5345 bit (highest order bit) of the ``value`` until it reaches the bit size
5346 of the type ``ty2``.
5348 When sign extending from i1, the extension always results in -1 or 0.
5353 .. code-block:: llvm
5355 %X = sext i8 -1 to i16 ; yields i16 :65535
5356 %Y = sext i1 true to i32 ; yields i32:-1
5357 %Z = sext <2 x i16> <i16 8, i16 7> to <2 x i32> ; yields <i32 8, i32 7>
5359 '``fptrunc .. to``' Instruction
5360 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5367 <result> = fptrunc <ty> <value> to <ty2> ; yields ty2
5372 The '``fptrunc``' instruction truncates ``value`` to type ``ty2``.
5377 The '``fptrunc``' instruction takes a :ref:`floating point <t_floating>`
5378 value to cast and a :ref:`floating point <t_floating>` type to cast it to.
5379 The size of ``value`` must be larger than the size of ``ty2``. This
5380 implies that ``fptrunc`` cannot be used to make a *no-op cast*.
5385 The '``fptrunc``' instruction truncates a ``value`` from a larger
5386 :ref:`floating point <t_floating>` type to a smaller :ref:`floating
5387 point <t_floating>` type. If the value cannot fit within the
5388 destination type, ``ty2``, then the results are undefined.
5393 .. code-block:: llvm
5395 %X = fptrunc double 123.0 to float ; yields float:123.0
5396 %Y = fptrunc double 1.0E+300 to float ; yields undefined
5398 '``fpext .. to``' Instruction
5399 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5406 <result> = fpext <ty> <value> to <ty2> ; yields ty2
5411 The '``fpext``' extends a floating point ``value`` to a larger floating
5417 The '``fpext``' instruction takes a :ref:`floating point <t_floating>`
5418 ``value`` to cast, and a :ref:`floating point <t_floating>` type to cast it
5419 to. The source type must be smaller than the destination type.
5424 The '``fpext``' instruction extends the ``value`` from a smaller
5425 :ref:`floating point <t_floating>` type to a larger :ref:`floating
5426 point <t_floating>` type. The ``fpext`` cannot be used to make a
5427 *no-op cast* because it always changes bits. Use ``bitcast`` to make a
5428 *no-op cast* for a floating point cast.
5433 .. code-block:: llvm
5435 %X = fpext float 3.125 to double ; yields double:3.125000e+00
5436 %Y = fpext double %X to fp128 ; yields fp128:0xL00000000000000004000900000000000
5438 '``fptoui .. to``' Instruction
5439 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5446 <result> = fptoui <ty> <value> to <ty2> ; yields ty2
5451 The '``fptoui``' converts a floating point ``value`` to its unsigned
5452 integer equivalent of type ``ty2``.
5457 The '``fptoui``' instruction takes a value to cast, which must be a
5458 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5459 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5460 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5461 type with the same number of elements as ``ty``
5466 The '``fptoui``' instruction converts its :ref:`floating
5467 point <t_floating>` operand into the nearest (rounding towards zero)
5468 unsigned integer value. If the value cannot fit in ``ty2``, the results
5474 .. code-block:: llvm
5476 %X = fptoui double 123.0 to i32 ; yields i32:123
5477 %Y = fptoui float 1.0E+300 to i1 ; yields undefined:1
5478 %Z = fptoui float 1.04E+17 to i8 ; yields undefined:1
5480 '``fptosi .. to``' Instruction
5481 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5488 <result> = fptosi <ty> <value> to <ty2> ; yields ty2
5493 The '``fptosi``' instruction converts :ref:`floating point <t_floating>`
5494 ``value`` to type ``ty2``.
5499 The '``fptosi``' instruction takes a value to cast, which must be a
5500 scalar or vector :ref:`floating point <t_floating>` value, and a type to
5501 cast it to ``ty2``, which must be an :ref:`integer <t_integer>` type. If
5502 ``ty`` is a vector floating point type, ``ty2`` must be a vector integer
5503 type with the same number of elements as ``ty``
5508 The '``fptosi``' instruction converts its :ref:`floating
5509 point <t_floating>` operand into the nearest (rounding towards zero)
5510 signed integer value. If the value cannot fit in ``ty2``, the results
5516 .. code-block:: llvm
5518 %X = fptosi double -123.0 to i32 ; yields i32:-123
5519 %Y = fptosi float 1.0E-247 to i1 ; yields undefined:1
5520 %Z = fptosi float 1.04E+17 to i8 ; yields undefined:1
5522 '``uitofp .. to``' Instruction
5523 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5530 <result> = uitofp <ty> <value> to <ty2> ; yields ty2
5535 The '``uitofp``' instruction regards ``value`` as an unsigned integer
5536 and converts that value to the ``ty2`` type.
5541 The '``uitofp``' instruction takes a value to cast, which must be a
5542 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5543 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5544 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5545 type with the same number of elements as ``ty``
5550 The '``uitofp``' instruction interprets its operand as an unsigned
5551 integer quantity and converts it to the corresponding floating point
5552 value. If the value cannot fit in the floating point value, the results
5558 .. code-block:: llvm
5560 %X = uitofp i32 257 to float ; yields float:257.0
5561 %Y = uitofp i8 -1 to double ; yields double:255.0
5563 '``sitofp .. to``' Instruction
5564 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5571 <result> = sitofp <ty> <value> to <ty2> ; yields ty2
5576 The '``sitofp``' instruction regards ``value`` as a signed integer and
5577 converts that value to the ``ty2`` type.
5582 The '``sitofp``' instruction takes a value to cast, which must be a
5583 scalar or vector :ref:`integer <t_integer>` value, and a type to cast it to
5584 ``ty2``, which must be an :ref:`floating point <t_floating>` type. If
5585 ``ty`` is a vector integer type, ``ty2`` must be a vector floating point
5586 type with the same number of elements as ``ty``
5591 The '``sitofp``' instruction interprets its operand as a signed integer
5592 quantity and converts it to the corresponding floating point value. If
5593 the value cannot fit in the floating point value, the results are
5599 .. code-block:: llvm
5601 %X = sitofp i32 257 to float ; yields float:257.0
5602 %Y = sitofp i8 -1 to double ; yields double:-1.0
5606 '``ptrtoint .. to``' Instruction
5607 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5614 <result> = ptrtoint <ty> <value> to <ty2> ; yields ty2
5619 The '``ptrtoint``' instruction converts the pointer or a vector of
5620 pointers ``value`` to the integer (or vector of integers) type ``ty2``.
5625 The '``ptrtoint``' instruction takes a ``value`` to cast, which must be
5626 a a value of type :ref:`pointer <t_pointer>` or a vector of pointers, and a
5627 type to cast it to ``ty2``, which must be an :ref:`integer <t_integer>` or
5628 a vector of integers type.
5633 The '``ptrtoint``' instruction converts ``value`` to integer type
5634 ``ty2`` by interpreting the pointer value as an integer and either
5635 truncating or zero extending that value to the size of the integer type.
5636 If ``value`` is smaller than ``ty2`` then a zero extension is done. If
5637 ``value`` is larger than ``ty2`` then a truncation is done. If they are
5638 the same size, then nothing is done (*no-op cast*) other than a type
5644 .. code-block:: llvm
5646 %X = ptrtoint i32* %P to i8 ; yields truncation on 32-bit architecture
5647 %Y = ptrtoint i32* %P to i64 ; yields zero extension on 32-bit architecture
5648 %Z = ptrtoint <4 x i32*> %P to <4 x i64>; yields vector zero extension for a vector of addresses on 32-bit architecture
5652 '``inttoptr .. to``' Instruction
5653 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5660 <result> = inttoptr <ty> <value> to <ty2> ; yields ty2
5665 The '``inttoptr``' instruction converts an integer ``value`` to a
5666 pointer type, ``ty2``.
5671 The '``inttoptr``' instruction takes an :ref:`integer <t_integer>` value to
5672 cast, and a type to cast it to, which must be a :ref:`pointer <t_pointer>`
5678 The '``inttoptr``' instruction converts ``value`` to type ``ty2`` by
5679 applying either a zero extension or a truncation depending on the size
5680 of the integer ``value``. If ``value`` is larger than the size of a
5681 pointer then a truncation is done. If ``value`` is smaller than the size
5682 of a pointer then a zero extension is done. If they are the same size,
5683 nothing is done (*no-op cast*).
5688 .. code-block:: llvm
5690 %X = inttoptr i32 255 to i32* ; yields zero extension on 64-bit architecture
5691 %Y = inttoptr i32 255 to i32* ; yields no-op on 32-bit architecture
5692 %Z = inttoptr i64 0 to i32* ; yields truncation on 32-bit architecture
5693 %Z = inttoptr <4 x i32> %G to <4 x i8*>; yields truncation of vector G to four pointers
5697 '``bitcast .. to``' Instruction
5698 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5705 <result> = bitcast <ty> <value> to <ty2> ; yields ty2
5710 The '``bitcast``' instruction converts ``value`` to type ``ty2`` without
5716 The '``bitcast``' instruction takes a value to cast, which must be a
5717 non-aggregate first class value, and a type to cast it to, which must
5718 also be a non-aggregate :ref:`first class <t_firstclass>` type. The
5719 bit sizes of ``value`` and the destination type, ``ty2``, must be
5720 identical. If the source type is a pointer, the destination type must
5721 also be a pointer of the same size. This instruction supports bitwise
5722 conversion of vectors to integers and to vectors of other types (as
5723 long as they have the same size).
5728 The '``bitcast``' instruction converts ``value`` to type ``ty2``. It
5729 is always a *no-op cast* because no bits change with this
5730 conversion. The conversion is done as if the ``value`` had been stored
5731 to memory and read back as type ``ty2``. Pointer (or vector of
5732 pointers) types may only be converted to other pointer (or vector of
5733 pointers) types with the same address space through this instruction.
5734 To convert pointers to other types, use the :ref:`inttoptr <i_inttoptr>`
5735 or :ref:`ptrtoint <i_ptrtoint>` instructions first.
5740 .. code-block:: llvm
5742 %X = bitcast i8 255 to i8 ; yields i8 :-1
5743 %Y = bitcast i32* %x to sint* ; yields sint*:%x
5744 %Z = bitcast <2 x int> %V to i64; ; yields i64: %V
5745 %Z = bitcast <2 x i32*> %V to <2 x i64*> ; yields <2 x i64*>
5747 .. _i_addrspacecast:
5749 '``addrspacecast .. to``' Instruction
5750 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
5757 <result> = addrspacecast <pty> <ptrval> to <pty2> ; yields pty2
5762 The '``addrspacecast``' instruction converts ``ptrval`` from ``pty`` in
5763 address space ``n`` to type ``pty2`` in address space ``m``.
5768 The '``addrspacecast``' instruction takes a pointer or vector of pointer value
5769 to cast and a pointer type to cast it to, which must have a different
5775 The '``addrspacecast``' instruction converts the pointer value
5776 ``ptrval`` to type ``pty2``. It can be a *no-op cast* or a complex
5777 value modification, depending on the target and the address space
5778 pair. Pointer conversions within the same address space must be
5779 performed with the ``bitcast`` instruction. Note that if the address space
5780 conversion is legal then both result and operand refer to the same memory
5786 .. code-block:: llvm
5788 %X = addrspacecast i32* %x to i32 addrspace(1)* ; yields i32 addrspace(1)*:%x
5789 %Y = addrspacecast i32 addrspace(1)* %y to i64 addrspace(2)* ; yields i64 addrspace(2)*:%y
5790 %Z = addrspacecast <4 x i32*> %z to <4 x float addrspace(3)*> ; yields <4 x float addrspace(3)*>:%z
5797 The instructions in this category are the "miscellaneous" instructions,
5798 which defy better classification.
5802 '``icmp``' Instruction
5803 ^^^^^^^^^^^^^^^^^^^^^^
5810 <result> = icmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5815 The '``icmp``' instruction returns a boolean value or a vector of
5816 boolean values based on comparison of its two integer, integer vector,
5817 pointer, or pointer vector operands.
5822 The '``icmp``' instruction takes three operands. The first operand is
5823 the condition code indicating the kind of comparison to perform. It is
5824 not a value, just a keyword. The possible condition code are:
5827 #. ``ne``: not equal
5828 #. ``ugt``: unsigned greater than
5829 #. ``uge``: unsigned greater or equal
5830 #. ``ult``: unsigned less than
5831 #. ``ule``: unsigned less or equal
5832 #. ``sgt``: signed greater than
5833 #. ``sge``: signed greater or equal
5834 #. ``slt``: signed less than
5835 #. ``sle``: signed less or equal
5837 The remaining two arguments must be :ref:`integer <t_integer>` or
5838 :ref:`pointer <t_pointer>` or integer :ref:`vector <t_vector>` typed. They
5839 must also be identical types.
5844 The '``icmp``' compares ``op1`` and ``op2`` according to the condition
5845 code given as ``cond``. The comparison performed always yields either an
5846 :ref:`i1 <t_integer>` or vector of ``i1`` result, as follows:
5848 #. ``eq``: yields ``true`` if the operands are equal, ``false``
5849 otherwise. No sign interpretation is necessary or performed.
5850 #. ``ne``: yields ``true`` if the operands are unequal, ``false``
5851 otherwise. No sign interpretation is necessary or performed.
5852 #. ``ugt``: interprets the operands as unsigned values and yields
5853 ``true`` if ``op1`` is greater than ``op2``.
5854 #. ``uge``: interprets the operands as unsigned values and yields
5855 ``true`` if ``op1`` is greater than or equal to ``op2``.
5856 #. ``ult``: interprets the operands as unsigned values and yields
5857 ``true`` if ``op1`` is less than ``op2``.
5858 #. ``ule``: interprets the operands as unsigned values and yields
5859 ``true`` if ``op1`` is less than or equal to ``op2``.
5860 #. ``sgt``: interprets the operands as signed values and yields ``true``
5861 if ``op1`` is greater than ``op2``.
5862 #. ``sge``: interprets the operands as signed values and yields ``true``
5863 if ``op1`` is greater than or equal to ``op2``.
5864 #. ``slt``: interprets the operands as signed values and yields ``true``
5865 if ``op1`` is less than ``op2``.
5866 #. ``sle``: interprets the operands as signed values and yields ``true``
5867 if ``op1`` is less than or equal to ``op2``.
5869 If the operands are :ref:`pointer <t_pointer>` typed, the pointer values
5870 are compared as if they were integers.
5872 If the operands are integer vectors, then they are compared element by
5873 element. The result is an ``i1`` vector with the same number of elements
5874 as the values being compared. Otherwise, the result is an ``i1``.
5879 .. code-block:: llvm
5881 <result> = icmp eq i32 4, 5 ; yields: result=false
5882 <result> = icmp ne float* %X, %X ; yields: result=false
5883 <result> = icmp ult i16 4, 5 ; yields: result=true
5884 <result> = icmp sgt i16 4, 5 ; yields: result=false
5885 <result> = icmp ule i16 -4, 5 ; yields: result=false
5886 <result> = icmp sge i16 4, 5 ; yields: result=false
5888 Note that the code generator does not yet support vector types with the
5889 ``icmp`` instruction.
5893 '``fcmp``' Instruction
5894 ^^^^^^^^^^^^^^^^^^^^^^
5901 <result> = fcmp <cond> <ty> <op1>, <op2> ; yields {i1} or {<N x i1>}:result
5906 The '``fcmp``' instruction returns a boolean value or vector of boolean
5907 values based on comparison of its operands.
5909 If the operands are floating point scalars, then the result type is a
5910 boolean (:ref:`i1 <t_integer>`).
5912 If the operands are floating point vectors, then the result type is a
5913 vector of boolean with the same number of elements as the operands being
5919 The '``fcmp``' instruction takes three operands. The first operand is
5920 the condition code indicating the kind of comparison to perform. It is
5921 not a value, just a keyword. The possible condition code are:
5923 #. ``false``: no comparison, always returns false
5924 #. ``oeq``: ordered and equal
5925 #. ``ogt``: ordered and greater than
5926 #. ``oge``: ordered and greater than or equal
5927 #. ``olt``: ordered and less than
5928 #. ``ole``: ordered and less than or equal
5929 #. ``one``: ordered and not equal
5930 #. ``ord``: ordered (no nans)
5931 #. ``ueq``: unordered or equal
5932 #. ``ugt``: unordered or greater than
5933 #. ``uge``: unordered or greater than or equal
5934 #. ``ult``: unordered or less than
5935 #. ``ule``: unordered or less than or equal
5936 #. ``une``: unordered or not equal
5937 #. ``uno``: unordered (either nans)
5938 #. ``true``: no comparison, always returns true
5940 *Ordered* means that neither operand is a QNAN while *unordered* means
5941 that either operand may be a QNAN.
5943 Each of ``val1`` and ``val2`` arguments must be either a :ref:`floating
5944 point <t_floating>` type or a :ref:`vector <t_vector>` of floating point
5945 type. They must have identical types.
5950 The '``fcmp``' instruction compares ``op1`` and ``op2`` according to the
5951 condition code given as ``cond``. If the operands are vectors, then the
5952 vectors are compared element by element. Each comparison performed
5953 always yields an :ref:`i1 <t_integer>` result, as follows:
5955 #. ``false``: always yields ``false``, regardless of operands.
5956 #. ``oeq``: yields ``true`` if both operands are not a QNAN and ``op1``
5957 is equal to ``op2``.
5958 #. ``ogt``: yields ``true`` if both operands are not a QNAN and ``op1``
5959 is greater than ``op2``.
5960 #. ``oge``: yields ``true`` if both operands are not a QNAN and ``op1``
5961 is greater than or equal to ``op2``.
5962 #. ``olt``: yields ``true`` if both operands are not a QNAN and ``op1``
5963 is less than ``op2``.
5964 #. ``ole``: yields ``true`` if both operands are not a QNAN and ``op1``
5965 is less than or equal to ``op2``.
5966 #. ``one``: yields ``true`` if both operands are not a QNAN and ``op1``
5967 is not equal to ``op2``.
5968 #. ``ord``: yields ``true`` if both operands are not a QNAN.
5969 #. ``ueq``: yields ``true`` if either operand is a QNAN or ``op1`` is
5971 #. ``ugt``: yields ``true`` if either operand is a QNAN or ``op1`` is
5972 greater than ``op2``.
5973 #. ``uge``: yields ``true`` if either operand is a QNAN or ``op1`` is
5974 greater than or equal to ``op2``.
5975 #. ``ult``: yields ``true`` if either operand is a QNAN or ``op1`` is
5977 #. ``ule``: yields ``true`` if either operand is a QNAN or ``op1`` is
5978 less than or equal to ``op2``.
5979 #. ``une``: yields ``true`` if either operand is a QNAN or ``op1`` is
5980 not equal to ``op2``.
5981 #. ``uno``: yields ``true`` if either operand is a QNAN.
5982 #. ``true``: always yields ``true``, regardless of operands.
5987 .. code-block:: llvm
5989 <result> = fcmp oeq float 4.0, 5.0 ; yields: result=false
5990 <result> = fcmp one float 4.0, 5.0 ; yields: result=true
5991 <result> = fcmp olt float 4.0, 5.0 ; yields: result=true
5992 <result> = fcmp ueq double 1.0, 2.0 ; yields: result=false
5994 Note that the code generator does not yet support vector types with the
5995 ``fcmp`` instruction.
5999 '``phi``' Instruction
6000 ^^^^^^^^^^^^^^^^^^^^^
6007 <result> = phi <ty> [ <val0>, <label0>], ...
6012 The '``phi``' instruction is used to implement the φ node in the SSA
6013 graph representing the function.
6018 The type of the incoming values is specified with the first type field.
6019 After this, the '``phi``' instruction takes a list of pairs as
6020 arguments, with one pair for each predecessor basic block of the current
6021 block. Only values of :ref:`first class <t_firstclass>` type may be used as
6022 the value arguments to the PHI node. Only labels may be used as the
6025 There must be no non-phi instructions between the start of a basic block
6026 and the PHI instructions: i.e. PHI instructions must be first in a basic
6029 For the purposes of the SSA form, the use of each incoming value is
6030 deemed to occur on the edge from the corresponding predecessor block to
6031 the current block (but after any definition of an '``invoke``'
6032 instruction's return value on the same edge).
6037 At runtime, the '``phi``' instruction logically takes on the value
6038 specified by the pair corresponding to the predecessor basic block that
6039 executed just prior to the current block.
6044 .. code-block:: llvm
6046 Loop: ; Infinite loop that counts from 0 on up...
6047 %indvar = phi i32 [ 0, %LoopHeader ], [ %nextindvar, %Loop ]
6048 %nextindvar = add i32 %indvar, 1
6053 '``select``' Instruction
6054 ^^^^^^^^^^^^^^^^^^^^^^^^
6061 <result> = select selty <cond>, <ty> <val1>, <ty> <val2> ; yields ty
6063 selty is either i1 or {<N x i1>}
6068 The '``select``' instruction is used to choose one value based on a
6069 condition, without branching.
6074 The '``select``' instruction requires an 'i1' value or a vector of 'i1'
6075 values indicating the condition, and two values of the same :ref:`first
6076 class <t_firstclass>` type. If the val1/val2 are vectors and the
6077 condition is a scalar, then entire vectors are selected, not individual
6083 If the condition is an i1 and it evaluates to 1, the instruction returns
6084 the first value argument; otherwise, it returns the second value
6087 If the condition is a vector of i1, then the value arguments must be
6088 vectors of the same size, and the selection is done element by element.
6093 .. code-block:: llvm
6095 %X = select i1 true, i8 17, i8 42 ; yields i8:17
6099 '``call``' Instruction
6100 ^^^^^^^^^^^^^^^^^^^^^^
6107 <result> = [tail] call [cconv] [ret attrs] <ty> [<fnty>*] <fnptrval>(<function args>) [fn attrs]
6112 The '``call``' instruction represents a simple function call.
6117 This instruction requires several arguments:
6119 #. The optional "tail" marker indicates that the callee function does
6120 not access any allocas or varargs in the caller. Note that calls may
6121 be marked "tail" even if they do not occur before a
6122 :ref:`ret <i_ret>` instruction. If the "tail" marker is present, the
6123 function call is eligible for tail call optimization, but `might not
6124 in fact be optimized into a jump <CodeGenerator.html#tailcallopt>`_.
6125 The code generator may optimize calls marked "tail" with either 1)
6126 automatic `sibling call
6127 optimization <CodeGenerator.html#sibcallopt>`_ when the caller and
6128 callee have matching signatures, or 2) forced tail call optimization
6129 when the following extra requirements are met:
6131 - Caller and callee both have the calling convention ``fastcc``.
6132 - The call is in tail position (ret immediately follows call and ret
6133 uses value of call or is void).
6134 - Option ``-tailcallopt`` is enabled, or
6135 ``llvm::GuaranteedTailCallOpt`` is ``true``.
6136 - `Platform specific constraints are
6137 met. <CodeGenerator.html#tailcallopt>`_
6139 #. The optional "cconv" marker indicates which :ref:`calling
6140 convention <callingconv>` the call should use. If none is
6141 specified, the call defaults to using C calling conventions. The
6142 calling convention of the call must match the calling convention of
6143 the target function, or else the behavior is undefined.
6144 #. The optional :ref:`Parameter Attributes <paramattrs>` list for return
6145 values. Only '``zeroext``', '``signext``', and '``inreg``' attributes
6147 #. '``ty``': the type of the call instruction itself which is also the
6148 type of the return value. Functions that return no value are marked
6150 #. '``fnty``': shall be the signature of the pointer to function value
6151 being invoked. The argument types must match the types implied by
6152 this signature. This type can be omitted if the function is not
6153 varargs and if the function type does not return a pointer to a
6155 #. '``fnptrval``': An LLVM value containing a pointer to a function to
6156 be invoked. In most cases, this is a direct function invocation, but
6157 indirect ``call``'s are just as possible, calling an arbitrary pointer
6159 #. '``function args``': argument list whose types match the function
6160 signature argument types and parameter attributes. All arguments must
6161 be of :ref:`first class <t_firstclass>` type. If the function signature
6162 indicates the function accepts a variable number of arguments, the
6163 extra arguments can be specified.
6164 #. The optional :ref:`function attributes <fnattrs>` list. Only
6165 '``noreturn``', '``nounwind``', '``readonly``' and '``readnone``'
6166 attributes are valid here.
6171 The '``call``' instruction is used to cause control flow to transfer to
6172 a specified function, with its incoming arguments bound to the specified
6173 values. Upon a '``ret``' instruction in the called function, control
6174 flow continues with the instruction after the function call, and the
6175 return value of the function is bound to the result argument.
6180 .. code-block:: llvm
6182 %retval = call i32 @test(i32 %argc)
6183 call i32 (i8*, ...)* @printf(i8* %msg, i32 12, i8 42) ; yields i32
6184 %X = tail call i32 @foo() ; yields i32
6185 %Y = tail call fastcc i32 @foo() ; yields i32
6186 call void %foo(i8 97 signext)
6188 %struct.A = type { i32, i8 }
6189 %r = call %struct.A @foo() ; yields { 32, i8 }
6190 %gr = extractvalue %struct.A %r, 0 ; yields i32
6191 %gr1 = extractvalue %struct.A %r, 1 ; yields i8
6192 %Z = call void @foo() noreturn ; indicates that %foo never returns normally
6193 %ZZ = call zeroext i32 @bar() ; Return value is %zero extended
6195 llvm treats calls to some functions with names and arguments that match
6196 the standard C99 library as being the C99 library functions, and may
6197 perform optimizations or generate code for them under that assumption.
6198 This is something we'd like to change in the future to provide better
6199 support for freestanding environments and non-C-based languages.
6203 '``va_arg``' Instruction
6204 ^^^^^^^^^^^^^^^^^^^^^^^^
6211 <resultval> = va_arg <va_list*> <arglist>, <argty>
6216 The '``va_arg``' instruction is used to access arguments passed through
6217 the "variable argument" area of a function call. It is used to implement
6218 the ``va_arg`` macro in C.
6223 This instruction takes a ``va_list*`` value and the type of the
6224 argument. It returns a value of the specified argument type and
6225 increments the ``va_list`` to point to the next argument. The actual
6226 type of ``va_list`` is target specific.
6231 The '``va_arg``' instruction loads an argument of the specified type
6232 from the specified ``va_list`` and causes the ``va_list`` to point to
6233 the next argument. For more information, see the variable argument
6234 handling :ref:`Intrinsic Functions <int_varargs>`.
6236 It is legal for this instruction to be called in a function which does
6237 not take a variable number of arguments, for example, the ``vfprintf``
6240 ``va_arg`` is an LLVM instruction instead of an :ref:`intrinsic
6241 function <intrinsics>` because it takes a type as an argument.
6246 See the :ref:`variable argument processing <int_varargs>` section.
6248 Note that the code generator does not yet fully support va\_arg on many
6249 targets. Also, it does not currently support va\_arg with aggregate
6250 types on any target.
6254 '``landingpad``' Instruction
6255 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6262 <resultval> = landingpad <resultty> personality <type> <pers_fn> <clause>+
6263 <resultval> = landingpad <resultty> personality <type> <pers_fn> cleanup <clause>*
6265 <clause> := catch <type> <value>
6266 <clause> := filter <array constant type> <array constant>
6271 The '``landingpad``' instruction is used by `LLVM's exception handling
6272 system <ExceptionHandling.html#overview>`_ to specify that a basic block
6273 is a landing pad --- one where the exception lands, and corresponds to the
6274 code found in the ``catch`` portion of a ``try``/``catch`` sequence. It
6275 defines values supplied by the personality function (``pers_fn``) upon
6276 re-entry to the function. The ``resultval`` has the type ``resultty``.
6281 This instruction takes a ``pers_fn`` value. This is the personality
6282 function associated with the unwinding mechanism. The optional
6283 ``cleanup`` flag indicates that the landing pad block is a cleanup.
6285 A ``clause`` begins with the clause type --- ``catch`` or ``filter`` --- and
6286 contains the global variable representing the "type" that may be caught
6287 or filtered respectively. Unlike the ``catch`` clause, the ``filter``
6288 clause takes an array constant as its argument. Use
6289 "``[0 x i8**] undef``" for a filter which cannot throw. The
6290 '``landingpad``' instruction must contain *at least* one ``clause`` or
6291 the ``cleanup`` flag.
6296 The '``landingpad``' instruction defines the values which are set by the
6297 personality function (``pers_fn``) upon re-entry to the function, and
6298 therefore the "result type" of the ``landingpad`` instruction. As with
6299 calling conventions, how the personality function results are
6300 represented in LLVM IR is target specific.
6302 The clauses are applied in order from top to bottom. If two
6303 ``landingpad`` instructions are merged together through inlining, the
6304 clauses from the calling function are appended to the list of clauses.
6305 When the call stack is being unwound due to an exception being thrown,
6306 the exception is compared against each ``clause`` in turn. If it doesn't
6307 match any of the clauses, and the ``cleanup`` flag is not set, then
6308 unwinding continues further up the call stack.
6310 The ``landingpad`` instruction has several restrictions:
6312 - A landing pad block is a basic block which is the unwind destination
6313 of an '``invoke``' instruction.
6314 - A landing pad block must have a '``landingpad``' instruction as its
6315 first non-PHI instruction.
6316 - There can be only one '``landingpad``' instruction within the landing
6318 - A basic block that is not a landing pad block may not include a
6319 '``landingpad``' instruction.
6320 - All '``landingpad``' instructions in a function must have the same
6321 personality function.
6326 .. code-block:: llvm
6328 ;; A landing pad which can catch an integer.
6329 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6331 ;; A landing pad that is a cleanup.
6332 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6334 ;; A landing pad which can catch an integer and can only throw a double.
6335 %res = landingpad { i8*, i32 } personality i32 (...)* @__gxx_personality_v0
6337 filter [1 x i8**] [@_ZTId]
6344 LLVM supports the notion of an "intrinsic function". These functions
6345 have well known names and semantics and are required to follow certain
6346 restrictions. Overall, these intrinsics represent an extension mechanism
6347 for the LLVM language that does not require changing all of the
6348 transformations in LLVM when adding to the language (or the bitcode
6349 reader/writer, the parser, etc...).
6351 Intrinsic function names must all start with an "``llvm.``" prefix. This
6352 prefix is reserved in LLVM for intrinsic names; thus, function names may
6353 not begin with this prefix. Intrinsic functions must always be external
6354 functions: you cannot define the body of intrinsic functions. Intrinsic
6355 functions may only be used in call or invoke instructions: it is illegal
6356 to take the address of an intrinsic function. Additionally, because
6357 intrinsic functions are part of the LLVM language, it is required if any
6358 are added that they be documented here.
6360 Some intrinsic functions can be overloaded, i.e., the intrinsic
6361 represents a family of functions that perform the same operation but on
6362 different data types. Because LLVM can represent over 8 million
6363 different integer types, overloading is used commonly to allow an
6364 intrinsic function to operate on any integer type. One or more of the
6365 argument types or the result type can be overloaded to accept any
6366 integer type. Argument types may also be defined as exactly matching a
6367 previous argument's type or the result type. This allows an intrinsic
6368 function which accepts multiple arguments, but needs all of them to be
6369 of the same type, to only be overloaded with respect to a single
6370 argument or the result.
6372 Overloaded intrinsics will have the names of its overloaded argument
6373 types encoded into its function name, each preceded by a period. Only
6374 those types which are overloaded result in a name suffix. Arguments
6375 whose type is matched against another type do not. For example, the
6376 ``llvm.ctpop`` function can take an integer of any width and returns an
6377 integer of exactly the same integer width. This leads to a family of
6378 functions such as ``i8 @llvm.ctpop.i8(i8 %val)`` and
6379 ``i29 @llvm.ctpop.i29(i29 %val)``. Only one type, the return type, is
6380 overloaded, and only one type suffix is required. Because the argument's
6381 type is matched against the return type, it does not require its own
6384 To learn how to add an intrinsic function, please see the `Extending
6385 LLVM Guide <ExtendingLLVM.html>`_.
6389 Variable Argument Handling Intrinsics
6390 -------------------------------------
6392 Variable argument support is defined in LLVM with the
6393 :ref:`va_arg <i_va_arg>` instruction and these three intrinsic
6394 functions. These functions are related to the similarly named macros
6395 defined in the ``<stdarg.h>`` header file.
6397 All of these functions operate on arguments that use a target-specific
6398 value type "``va_list``". The LLVM assembly language reference manual
6399 does not define what this type is, so all transformations should be
6400 prepared to handle these functions regardless of the type used.
6402 This example shows how the :ref:`va_arg <i_va_arg>` instruction and the
6403 variable argument handling intrinsic functions are used.
6405 .. code-block:: llvm
6407 define i32 @test(i32 %X, ...) {
6408 ; Initialize variable argument processing
6410 %ap2 = bitcast i8** %ap to i8*
6411 call void @llvm.va_start(i8* %ap2)
6413 ; Read a single integer argument
6414 %tmp = va_arg i8** %ap, i32
6416 ; Demonstrate usage of llvm.va_copy and llvm.va_end
6418 %aq2 = bitcast i8** %aq to i8*
6419 call void @llvm.va_copy(i8* %aq2, i8* %ap2)
6420 call void @llvm.va_end(i8* %aq2)
6422 ; Stop processing of arguments.
6423 call void @llvm.va_end(i8* %ap2)
6427 declare void @llvm.va_start(i8*)
6428 declare void @llvm.va_copy(i8*, i8*)
6429 declare void @llvm.va_end(i8*)
6433 '``llvm.va_start``' Intrinsic
6434 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6441 declare void @llvm.va_start(i8* <arglist>)
6446 The '``llvm.va_start``' intrinsic initializes ``*<arglist>`` for
6447 subsequent use by ``va_arg``.
6452 The argument is a pointer to a ``va_list`` element to initialize.
6457 The '``llvm.va_start``' intrinsic works just like the ``va_start`` macro
6458 available in C. In a target-dependent way, it initializes the
6459 ``va_list`` element to which the argument points, so that the next call
6460 to ``va_arg`` will produce the first variable argument passed to the
6461 function. Unlike the C ``va_start`` macro, this intrinsic does not need
6462 to know the last argument of the function as the compiler can figure
6465 '``llvm.va_end``' Intrinsic
6466 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6473 declare void @llvm.va_end(i8* <arglist>)
6478 The '``llvm.va_end``' intrinsic destroys ``*<arglist>``, which has been
6479 initialized previously with ``llvm.va_start`` or ``llvm.va_copy``.
6484 The argument is a pointer to a ``va_list`` to destroy.
6489 The '``llvm.va_end``' intrinsic works just like the ``va_end`` macro
6490 available in C. In a target-dependent way, it destroys the ``va_list``
6491 element to which the argument points. Calls to
6492 :ref:`llvm.va_start <int_va_start>` and
6493 :ref:`llvm.va_copy <int_va_copy>` must be matched exactly with calls to
6498 '``llvm.va_copy``' Intrinsic
6499 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6506 declare void @llvm.va_copy(i8* <destarglist>, i8* <srcarglist>)
6511 The '``llvm.va_copy``' intrinsic copies the current argument position
6512 from the source argument list to the destination argument list.
6517 The first argument is a pointer to a ``va_list`` element to initialize.
6518 The second argument is a pointer to a ``va_list`` element to copy from.
6523 The '``llvm.va_copy``' intrinsic works just like the ``va_copy`` macro
6524 available in C. In a target-dependent way, it copies the source
6525 ``va_list`` element into the destination ``va_list`` element. This
6526 intrinsic is necessary because the `` llvm.va_start`` intrinsic may be
6527 arbitrarily complex and require, for example, memory allocation.
6529 Accurate Garbage Collection Intrinsics
6530 --------------------------------------
6532 LLVM support for `Accurate Garbage Collection <GarbageCollection.html>`_
6533 (GC) requires the implementation and generation of these intrinsics.
6534 These intrinsics allow identification of :ref:`GC roots on the
6535 stack <int_gcroot>`, as well as garbage collector implementations that
6536 require :ref:`read <int_gcread>` and :ref:`write <int_gcwrite>` barriers.
6537 Front-ends for type-safe garbage collected languages should generate
6538 these intrinsics to make use of the LLVM garbage collectors. For more
6539 details, see `Accurate Garbage Collection with
6540 LLVM <GarbageCollection.html>`_.
6542 The garbage collection intrinsics only operate on objects in the generic
6543 address space (address space zero).
6547 '``llvm.gcroot``' Intrinsic
6548 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6555 declare void @llvm.gcroot(i8** %ptrloc, i8* %metadata)
6560 The '``llvm.gcroot``' intrinsic declares the existence of a GC root to
6561 the code generator, and allows some metadata to be associated with it.
6566 The first argument specifies the address of a stack object that contains
6567 the root pointer. The second pointer (which must be either a constant or
6568 a global value address) contains the meta-data to be associated with the
6574 At runtime, a call to this intrinsic stores a null pointer into the
6575 "ptrloc" location. At compile-time, the code generator generates
6576 information to allow the runtime to find the pointer at GC safe points.
6577 The '``llvm.gcroot``' intrinsic may only be used in a function which
6578 :ref:`specifies a GC algorithm <gc>`.
6582 '``llvm.gcread``' Intrinsic
6583 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6590 declare i8* @llvm.gcread(i8* %ObjPtr, i8** %Ptr)
6595 The '``llvm.gcread``' intrinsic identifies reads of references from heap
6596 locations, allowing garbage collector implementations that require read
6602 The second argument is the address to read from, which should be an
6603 address allocated from the garbage collector. The first object is a
6604 pointer to the start of the referenced object, if needed by the language
6605 runtime (otherwise null).
6610 The '``llvm.gcread``' intrinsic has the same semantics as a load
6611 instruction, but may be replaced with substantially more complex code by
6612 the garbage collector runtime, as needed. The '``llvm.gcread``'
6613 intrinsic may only be used in a function which :ref:`specifies a GC
6618 '``llvm.gcwrite``' Intrinsic
6619 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6626 declare void @llvm.gcwrite(i8* %P1, i8* %Obj, i8** %P2)
6631 The '``llvm.gcwrite``' intrinsic identifies writes of references to heap
6632 locations, allowing garbage collector implementations that require write
6633 barriers (such as generational or reference counting collectors).
6638 The first argument is the reference to store, the second is the start of
6639 the object to store it to, and the third is the address of the field of
6640 Obj to store to. If the runtime does not require a pointer to the
6641 object, Obj may be null.
6646 The '``llvm.gcwrite``' intrinsic has the same semantics as a store
6647 instruction, but may be replaced with substantially more complex code by
6648 the garbage collector runtime, as needed. The '``llvm.gcwrite``'
6649 intrinsic may only be used in a function which :ref:`specifies a GC
6652 Code Generator Intrinsics
6653 -------------------------
6655 These intrinsics are provided by LLVM to expose special features that
6656 may only be implemented with code generator support.
6658 '``llvm.returnaddress``' Intrinsic
6659 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6666 declare i8 *@llvm.returnaddress(i32 <level>)
6671 The '``llvm.returnaddress``' intrinsic attempts to compute a
6672 target-specific value indicating the return address of the current
6673 function or one of its callers.
6678 The argument to this intrinsic indicates which function to return the
6679 address for. Zero indicates the calling function, one indicates its
6680 caller, etc. The argument is **required** to be a constant integer
6686 The '``llvm.returnaddress``' intrinsic either returns a pointer
6687 indicating the return address of the specified call frame, or zero if it
6688 cannot be identified. The value returned by this intrinsic is likely to
6689 be incorrect or 0 for arguments other than zero, so it should only be
6690 used for debugging purposes.
6692 Note that calling this intrinsic does not prevent function inlining or
6693 other aggressive transformations, so the value returned may not be that
6694 of the obvious source-language caller.
6696 '``llvm.frameaddress``' Intrinsic
6697 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6704 declare i8* @llvm.frameaddress(i32 <level>)
6709 The '``llvm.frameaddress``' intrinsic attempts to return the
6710 target-specific frame pointer value for the specified stack frame.
6715 The argument to this intrinsic indicates which function to return the
6716 frame pointer for. Zero indicates the calling function, one indicates
6717 its caller, etc. The argument is **required** to be a constant integer
6723 The '``llvm.frameaddress``' intrinsic either returns a pointer
6724 indicating the frame address of the specified call frame, or zero if it
6725 cannot be identified. The value returned by this intrinsic is likely to
6726 be incorrect or 0 for arguments other than zero, so it should only be
6727 used for debugging purposes.
6729 Note that calling this intrinsic does not prevent function inlining or
6730 other aggressive transformations, so the value returned may not be that
6731 of the obvious source-language caller.
6735 '``llvm.stacksave``' Intrinsic
6736 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6743 declare i8* @llvm.stacksave()
6748 The '``llvm.stacksave``' intrinsic is used to remember the current state
6749 of the function stack, for use with
6750 :ref:`llvm.stackrestore <int_stackrestore>`. This is useful for
6751 implementing language features like scoped automatic variable sized
6757 This intrinsic returns a opaque pointer value that can be passed to
6758 :ref:`llvm.stackrestore <int_stackrestore>`. When an
6759 ``llvm.stackrestore`` intrinsic is executed with a value saved from
6760 ``llvm.stacksave``, it effectively restores the state of the stack to
6761 the state it was in when the ``llvm.stacksave`` intrinsic executed. In
6762 practice, this pops any :ref:`alloca <i_alloca>` blocks from the stack that
6763 were allocated after the ``llvm.stacksave`` was executed.
6765 .. _int_stackrestore:
6767 '``llvm.stackrestore``' Intrinsic
6768 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6775 declare void @llvm.stackrestore(i8* %ptr)
6780 The '``llvm.stackrestore``' intrinsic is used to restore the state of
6781 the function stack to the state it was in when the corresponding
6782 :ref:`llvm.stacksave <int_stacksave>` intrinsic executed. This is
6783 useful for implementing language features like scoped automatic variable
6784 sized arrays in C99.
6789 See the description for :ref:`llvm.stacksave <int_stacksave>`.
6791 '``llvm.prefetch``' Intrinsic
6792 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6799 declare void @llvm.prefetch(i8* <address>, i32 <rw>, i32 <locality>, i32 <cache type>)
6804 The '``llvm.prefetch``' intrinsic is a hint to the code generator to
6805 insert a prefetch instruction if supported; otherwise, it is a noop.
6806 Prefetches have no effect on the behavior of the program but can change
6807 its performance characteristics.
6812 ``address`` is the address to be prefetched, ``rw`` is the specifier
6813 determining if the fetch should be for a read (0) or write (1), and
6814 ``locality`` is a temporal locality specifier ranging from (0) - no
6815 locality, to (3) - extremely local keep in cache. The ``cache type``
6816 specifies whether the prefetch is performed on the data (1) or
6817 instruction (0) cache. The ``rw``, ``locality`` and ``cache type``
6818 arguments must be constant integers.
6823 This intrinsic does not modify the behavior of the program. In
6824 particular, prefetches cannot trap and do not produce a value. On
6825 targets that support this intrinsic, the prefetch can provide hints to
6826 the processor cache for better performance.
6828 '``llvm.pcmarker``' Intrinsic
6829 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6836 declare void @llvm.pcmarker(i32 <id>)
6841 The '``llvm.pcmarker``' intrinsic is a method to export a Program
6842 Counter (PC) in a region of code to simulators and other tools. The
6843 method is target specific, but it is expected that the marker will use
6844 exported symbols to transmit the PC of the marker. The marker makes no
6845 guarantees that it will remain with any specific instruction after
6846 optimizations. It is possible that the presence of a marker will inhibit
6847 optimizations. The intended use is to be inserted after optimizations to
6848 allow correlations of simulation runs.
6853 ``id`` is a numerical id identifying the marker.
6858 This intrinsic does not modify the behavior of the program. Backends
6859 that do not support this intrinsic may ignore it.
6861 '``llvm.readcyclecounter``' Intrinsic
6862 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6869 declare i64 @llvm.readcyclecounter()
6874 The '``llvm.readcyclecounter``' intrinsic provides access to the cycle
6875 counter register (or similar low latency, high accuracy clocks) on those
6876 targets that support it. On X86, it should map to RDTSC. On Alpha, it
6877 should map to RPCC. As the backing counters overflow quickly (on the
6878 order of 9 seconds on alpha), this should only be used for small
6884 When directly supported, reading the cycle counter should not modify any
6885 memory. Implementations are allowed to either return a application
6886 specific value or a system wide value. On backends without support, this
6887 is lowered to a constant 0.
6889 Note that runtime support may be conditional on the privilege-level code is
6890 running at and the host platform.
6892 Standard C Library Intrinsics
6893 -----------------------------
6895 LLVM provides intrinsics for a few important standard C library
6896 functions. These intrinsics allow source-language front-ends to pass
6897 information about the alignment of the pointer arguments to the code
6898 generator, providing opportunity for more efficient code generation.
6902 '``llvm.memcpy``' Intrinsic
6903 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
6908 This is an overloaded intrinsic. You can use ``llvm.memcpy`` on any
6909 integer bit width and for different address spaces. Not all targets
6910 support all bit widths however.
6914 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6915 i32 <len>, i32 <align>, i1 <isvolatile>)
6916 declare void @llvm.memcpy.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6917 i64 <len>, i32 <align>, i1 <isvolatile>)
6922 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6923 source location to the destination location.
6925 Note that, unlike the standard libc function, the ``llvm.memcpy.*``
6926 intrinsics do not return a value, takes extra alignment/isvolatile
6927 arguments and the pointers can be in specified address spaces.
6932 The first argument is a pointer to the destination, the second is a
6933 pointer to the source. The third argument is an integer argument
6934 specifying the number of bytes to copy, the fourth argument is the
6935 alignment of the source and destination locations, and the fifth is a
6936 boolean indicating a volatile access.
6938 If the call to this intrinsic has an alignment value that is not 0 or 1,
6939 then the caller guarantees that both the source and destination pointers
6940 are aligned to that boundary.
6942 If the ``isvolatile`` parameter is ``true``, the ``llvm.memcpy`` call is
6943 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
6944 very cleanly specified and it is unwise to depend on it.
6949 The '``llvm.memcpy.*``' intrinsics copy a block of memory from the
6950 source location to the destination location, which are not allowed to
6951 overlap. It copies "len" bytes of memory over. If the argument is known
6952 to be aligned to some boundary, this can be specified as the fourth
6953 argument, otherwise it should be set to 0 or 1 (both meaning no alignment).
6955 '``llvm.memmove``' Intrinsic
6956 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
6961 This is an overloaded intrinsic. You can use llvm.memmove on any integer
6962 bit width and for different address space. Not all targets support all
6967 declare void @llvm.memmove.p0i8.p0i8.i32(i8* <dest>, i8* <src>,
6968 i32 <len>, i32 <align>, i1 <isvolatile>)
6969 declare void @llvm.memmove.p0i8.p0i8.i64(i8* <dest>, i8* <src>,
6970 i64 <len>, i32 <align>, i1 <isvolatile>)
6975 The '``llvm.memmove.*``' intrinsics move a block of memory from the
6976 source location to the destination location. It is similar to the
6977 '``llvm.memcpy``' intrinsic but allows the two memory locations to
6980 Note that, unlike the standard libc function, the ``llvm.memmove.*``
6981 intrinsics do not return a value, takes extra alignment/isvolatile
6982 arguments and the pointers can be in specified address spaces.
6987 The first argument is a pointer to the destination, the second is a
6988 pointer to the source. The third argument is an integer argument
6989 specifying the number of bytes to copy, the fourth argument is the
6990 alignment of the source and destination locations, and the fifth is a
6991 boolean indicating a volatile access.
6993 If the call to this intrinsic has an alignment value that is not 0 or 1,
6994 then the caller guarantees that the source and destination pointers are
6995 aligned to that boundary.
6997 If the ``isvolatile`` parameter is ``true``, the ``llvm.memmove`` call
6998 is a :ref:`volatile operation <volatile>`. The detailed access behavior is
6999 not very cleanly specified and it is unwise to depend on it.
7004 The '``llvm.memmove.*``' intrinsics copy a block of memory from the
7005 source location to the destination location, which may overlap. It
7006 copies "len" bytes of memory over. If the argument is known to be
7007 aligned to some boundary, this can be specified as the fourth argument,
7008 otherwise it should be set to 0 or 1 (both meaning no alignment).
7010 '``llvm.memset.*``' Intrinsics
7011 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7016 This is an overloaded intrinsic. You can use llvm.memset on any integer
7017 bit width and for different address spaces. However, not all targets
7018 support all bit widths.
7022 declare void @llvm.memset.p0i8.i32(i8* <dest>, i8 <val>,
7023 i32 <len>, i32 <align>, i1 <isvolatile>)
7024 declare void @llvm.memset.p0i8.i64(i8* <dest>, i8 <val>,
7025 i64 <len>, i32 <align>, i1 <isvolatile>)
7030 The '``llvm.memset.*``' intrinsics fill a block of memory with a
7031 particular byte value.
7033 Note that, unlike the standard libc function, the ``llvm.memset``
7034 intrinsic does not return a value and takes extra alignment/volatile
7035 arguments. Also, the destination can be in an arbitrary address space.
7040 The first argument is a pointer to the destination to fill, the second
7041 is the byte value with which to fill it, the third argument is an
7042 integer argument specifying the number of bytes to fill, and the fourth
7043 argument is the known alignment of the destination location.
7045 If the call to this intrinsic has an alignment value that is not 0 or 1,
7046 then the caller guarantees that the destination pointer is aligned to
7049 If the ``isvolatile`` parameter is ``true``, the ``llvm.memset`` call is
7050 a :ref:`volatile operation <volatile>`. The detailed access behavior is not
7051 very cleanly specified and it is unwise to depend on it.
7056 The '``llvm.memset.*``' intrinsics fill "len" bytes of memory starting
7057 at the destination location. If the argument is known to be aligned to
7058 some boundary, this can be specified as the fourth argument, otherwise
7059 it should be set to 0 or 1 (both meaning no alignment).
7061 '``llvm.sqrt.*``' Intrinsic
7062 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7067 This is an overloaded intrinsic. You can use ``llvm.sqrt`` on any
7068 floating point or vector of floating point type. Not all targets support
7073 declare float @llvm.sqrt.f32(float %Val)
7074 declare double @llvm.sqrt.f64(double %Val)
7075 declare x86_fp80 @llvm.sqrt.f80(x86_fp80 %Val)
7076 declare fp128 @llvm.sqrt.f128(fp128 %Val)
7077 declare ppc_fp128 @llvm.sqrt.ppcf128(ppc_fp128 %Val)
7082 The '``llvm.sqrt``' intrinsics return the sqrt of the specified operand,
7083 returning the same value as the libm '``sqrt``' functions would. Unlike
7084 ``sqrt`` in libm, however, ``llvm.sqrt`` has undefined behavior for
7085 negative numbers other than -0.0 (which allows for better optimization,
7086 because there is no need to worry about errno being set).
7087 ``llvm.sqrt(-0.0)`` is defined to return -0.0 like IEEE sqrt.
7092 The argument and return value are floating point numbers of the same
7098 This function returns the sqrt of the specified operand if it is a
7099 nonnegative floating point number.
7101 '``llvm.powi.*``' Intrinsic
7102 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7107 This is an overloaded intrinsic. You can use ``llvm.powi`` on any
7108 floating point or vector of floating point type. Not all targets support
7113 declare float @llvm.powi.f32(float %Val, i32 %power)
7114 declare double @llvm.powi.f64(double %Val, i32 %power)
7115 declare x86_fp80 @llvm.powi.f80(x86_fp80 %Val, i32 %power)
7116 declare fp128 @llvm.powi.f128(fp128 %Val, i32 %power)
7117 declare ppc_fp128 @llvm.powi.ppcf128(ppc_fp128 %Val, i32 %power)
7122 The '``llvm.powi.*``' intrinsics return the first operand raised to the
7123 specified (positive or negative) power. The order of evaluation of
7124 multiplications is not defined. When a vector of floating point type is
7125 used, the second argument remains a scalar integer value.
7130 The second argument is an integer power, and the first is a value to
7131 raise to that power.
7136 This function returns the first value raised to the second power with an
7137 unspecified sequence of rounding operations.
7139 '``llvm.sin.*``' Intrinsic
7140 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7145 This is an overloaded intrinsic. You can use ``llvm.sin`` on any
7146 floating point or vector of floating point type. Not all targets support
7151 declare float @llvm.sin.f32(float %Val)
7152 declare double @llvm.sin.f64(double %Val)
7153 declare x86_fp80 @llvm.sin.f80(x86_fp80 %Val)
7154 declare fp128 @llvm.sin.f128(fp128 %Val)
7155 declare ppc_fp128 @llvm.sin.ppcf128(ppc_fp128 %Val)
7160 The '``llvm.sin.*``' intrinsics return the sine of the operand.
7165 The argument and return value are floating point numbers of the same
7171 This function returns the sine of the specified operand, returning the
7172 same values as the libm ``sin`` functions would, and handles error
7173 conditions in the same way.
7175 '``llvm.cos.*``' Intrinsic
7176 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7181 This is an overloaded intrinsic. You can use ``llvm.cos`` on any
7182 floating point or vector of floating point type. Not all targets support
7187 declare float @llvm.cos.f32(float %Val)
7188 declare double @llvm.cos.f64(double %Val)
7189 declare x86_fp80 @llvm.cos.f80(x86_fp80 %Val)
7190 declare fp128 @llvm.cos.f128(fp128 %Val)
7191 declare ppc_fp128 @llvm.cos.ppcf128(ppc_fp128 %Val)
7196 The '``llvm.cos.*``' intrinsics return the cosine of the operand.
7201 The argument and return value are floating point numbers of the same
7207 This function returns the cosine of the specified operand, returning the
7208 same values as the libm ``cos`` functions would, and handles error
7209 conditions in the same way.
7211 '``llvm.pow.*``' Intrinsic
7212 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7217 This is an overloaded intrinsic. You can use ``llvm.pow`` on any
7218 floating point or vector of floating point type. Not all targets support
7223 declare float @llvm.pow.f32(float %Val, float %Power)
7224 declare double @llvm.pow.f64(double %Val, double %Power)
7225 declare x86_fp80 @llvm.pow.f80(x86_fp80 %Val, x86_fp80 %Power)
7226 declare fp128 @llvm.pow.f128(fp128 %Val, fp128 %Power)
7227 declare ppc_fp128 @llvm.pow.ppcf128(ppc_fp128 %Val, ppc_fp128 Power)
7232 The '``llvm.pow.*``' intrinsics return the first operand raised to the
7233 specified (positive or negative) power.
7238 The second argument is a floating point power, and the first is a value
7239 to raise to that power.
7244 This function returns the first value raised to the second power,
7245 returning the same values as the libm ``pow`` functions would, and
7246 handles error conditions in the same way.
7248 '``llvm.exp.*``' Intrinsic
7249 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7254 This is an overloaded intrinsic. You can use ``llvm.exp`` on any
7255 floating point or vector of floating point type. Not all targets support
7260 declare float @llvm.exp.f32(float %Val)
7261 declare double @llvm.exp.f64(double %Val)
7262 declare x86_fp80 @llvm.exp.f80(x86_fp80 %Val)
7263 declare fp128 @llvm.exp.f128(fp128 %Val)
7264 declare ppc_fp128 @llvm.exp.ppcf128(ppc_fp128 %Val)
7269 The '``llvm.exp.*``' intrinsics perform the exp function.
7274 The argument and return value are floating point numbers of the same
7280 This function returns the same values as the libm ``exp`` functions
7281 would, and handles error conditions in the same way.
7283 '``llvm.exp2.*``' Intrinsic
7284 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7289 This is an overloaded intrinsic. You can use ``llvm.exp2`` on any
7290 floating point or vector of floating point type. Not all targets support
7295 declare float @llvm.exp2.f32(float %Val)
7296 declare double @llvm.exp2.f64(double %Val)
7297 declare x86_fp80 @llvm.exp2.f80(x86_fp80 %Val)
7298 declare fp128 @llvm.exp2.f128(fp128 %Val)
7299 declare ppc_fp128 @llvm.exp2.ppcf128(ppc_fp128 %Val)
7304 The '``llvm.exp2.*``' intrinsics perform the exp2 function.
7309 The argument and return value are floating point numbers of the same
7315 This function returns the same values as the libm ``exp2`` functions
7316 would, and handles error conditions in the same way.
7318 '``llvm.log.*``' Intrinsic
7319 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7324 This is an overloaded intrinsic. You can use ``llvm.log`` on any
7325 floating point or vector of floating point type. Not all targets support
7330 declare float @llvm.log.f32(float %Val)
7331 declare double @llvm.log.f64(double %Val)
7332 declare x86_fp80 @llvm.log.f80(x86_fp80 %Val)
7333 declare fp128 @llvm.log.f128(fp128 %Val)
7334 declare ppc_fp128 @llvm.log.ppcf128(ppc_fp128 %Val)
7339 The '``llvm.log.*``' intrinsics perform the log function.
7344 The argument and return value are floating point numbers of the same
7350 This function returns the same values as the libm ``log`` functions
7351 would, and handles error conditions in the same way.
7353 '``llvm.log10.*``' Intrinsic
7354 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7359 This is an overloaded intrinsic. You can use ``llvm.log10`` on any
7360 floating point or vector of floating point type. Not all targets support
7365 declare float @llvm.log10.f32(float %Val)
7366 declare double @llvm.log10.f64(double %Val)
7367 declare x86_fp80 @llvm.log10.f80(x86_fp80 %Val)
7368 declare fp128 @llvm.log10.f128(fp128 %Val)
7369 declare ppc_fp128 @llvm.log10.ppcf128(ppc_fp128 %Val)
7374 The '``llvm.log10.*``' intrinsics perform the log10 function.
7379 The argument and return value are floating point numbers of the same
7385 This function returns the same values as the libm ``log10`` functions
7386 would, and handles error conditions in the same way.
7388 '``llvm.log2.*``' Intrinsic
7389 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7394 This is an overloaded intrinsic. You can use ``llvm.log2`` on any
7395 floating point or vector of floating point type. Not all targets support
7400 declare float @llvm.log2.f32(float %Val)
7401 declare double @llvm.log2.f64(double %Val)
7402 declare x86_fp80 @llvm.log2.f80(x86_fp80 %Val)
7403 declare fp128 @llvm.log2.f128(fp128 %Val)
7404 declare ppc_fp128 @llvm.log2.ppcf128(ppc_fp128 %Val)
7409 The '``llvm.log2.*``' intrinsics perform the log2 function.
7414 The argument and return value are floating point numbers of the same
7420 This function returns the same values as the libm ``log2`` functions
7421 would, and handles error conditions in the same way.
7423 '``llvm.fma.*``' Intrinsic
7424 ^^^^^^^^^^^^^^^^^^^^^^^^^^
7429 This is an overloaded intrinsic. You can use ``llvm.fma`` on any
7430 floating point or vector of floating point type. Not all targets support
7435 declare float @llvm.fma.f32(float %a, float %b, float %c)
7436 declare double @llvm.fma.f64(double %a, double %b, double %c)
7437 declare x86_fp80 @llvm.fma.f80(x86_fp80 %a, x86_fp80 %b, x86_fp80 %c)
7438 declare fp128 @llvm.fma.f128(fp128 %a, fp128 %b, fp128 %c)
7439 declare ppc_fp128 @llvm.fma.ppcf128(ppc_fp128 %a, ppc_fp128 %b, ppc_fp128 %c)
7444 The '``llvm.fma.*``' intrinsics perform the fused multiply-add
7450 The argument and return value are floating point numbers of the same
7456 This function returns the same values as the libm ``fma`` functions
7459 '``llvm.fabs.*``' Intrinsic
7460 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7465 This is an overloaded intrinsic. You can use ``llvm.fabs`` on any
7466 floating point or vector of floating point type. Not all targets support
7471 declare float @llvm.fabs.f32(float %Val)
7472 declare double @llvm.fabs.f64(double %Val)
7473 declare x86_fp80 @llvm.fabs.f80(x86_fp80 %Val)
7474 declare fp128 @llvm.fabs.f128(fp128 %Val)
7475 declare ppc_fp128 @llvm.fabs.ppcf128(ppc_fp128 %Val)
7480 The '``llvm.fabs.*``' intrinsics return the absolute value of the
7486 The argument and return value are floating point numbers of the same
7492 This function returns the same values as the libm ``fabs`` functions
7493 would, and handles error conditions in the same way.
7495 '``llvm.copysign.*``' Intrinsic
7496 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7501 This is an overloaded intrinsic. You can use ``llvm.copysign`` on any
7502 floating point or vector of floating point type. Not all targets support
7507 declare float @llvm.copysign.f32(float %Mag, float %Sgn)
7508 declare double @llvm.copysign.f64(double %Mag, double %Sgn)
7509 declare x86_fp80 @llvm.copysign.f80(x86_fp80 %Mag, x86_fp80 %Sgn)
7510 declare fp128 @llvm.copysign.f128(fp128 %Mag, fp128 %Sgn)
7511 declare ppc_fp128 @llvm.copysign.ppcf128(ppc_fp128 %Mag, ppc_fp128 %Sgn)
7516 The '``llvm.copysign.*``' intrinsics return a value with the magnitude of the
7517 first operand and the sign of the second operand.
7522 The arguments and return value are floating point numbers of the same
7528 This function returns the same values as the libm ``copysign``
7529 functions would, and handles error conditions in the same way.
7531 '``llvm.floor.*``' Intrinsic
7532 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7537 This is an overloaded intrinsic. You can use ``llvm.floor`` on any
7538 floating point or vector of floating point type. Not all targets support
7543 declare float @llvm.floor.f32(float %Val)
7544 declare double @llvm.floor.f64(double %Val)
7545 declare x86_fp80 @llvm.floor.f80(x86_fp80 %Val)
7546 declare fp128 @llvm.floor.f128(fp128 %Val)
7547 declare ppc_fp128 @llvm.floor.ppcf128(ppc_fp128 %Val)
7552 The '``llvm.floor.*``' intrinsics return the floor of the operand.
7557 The argument and return value are floating point numbers of the same
7563 This function returns the same values as the libm ``floor`` functions
7564 would, and handles error conditions in the same way.
7566 '``llvm.ceil.*``' Intrinsic
7567 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7572 This is an overloaded intrinsic. You can use ``llvm.ceil`` on any
7573 floating point or vector of floating point type. Not all targets support
7578 declare float @llvm.ceil.f32(float %Val)
7579 declare double @llvm.ceil.f64(double %Val)
7580 declare x86_fp80 @llvm.ceil.f80(x86_fp80 %Val)
7581 declare fp128 @llvm.ceil.f128(fp128 %Val)
7582 declare ppc_fp128 @llvm.ceil.ppcf128(ppc_fp128 %Val)
7587 The '``llvm.ceil.*``' intrinsics return the ceiling of the operand.
7592 The argument and return value are floating point numbers of the same
7598 This function returns the same values as the libm ``ceil`` functions
7599 would, and handles error conditions in the same way.
7601 '``llvm.trunc.*``' Intrinsic
7602 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7607 This is an overloaded intrinsic. You can use ``llvm.trunc`` on any
7608 floating point or vector of floating point type. Not all targets support
7613 declare float @llvm.trunc.f32(float %Val)
7614 declare double @llvm.trunc.f64(double %Val)
7615 declare x86_fp80 @llvm.trunc.f80(x86_fp80 %Val)
7616 declare fp128 @llvm.trunc.f128(fp128 %Val)
7617 declare ppc_fp128 @llvm.trunc.ppcf128(ppc_fp128 %Val)
7622 The '``llvm.trunc.*``' intrinsics returns the operand rounded to the
7623 nearest integer not larger in magnitude than the operand.
7628 The argument and return value are floating point numbers of the same
7634 This function returns the same values as the libm ``trunc`` functions
7635 would, and handles error conditions in the same way.
7637 '``llvm.rint.*``' Intrinsic
7638 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7643 This is an overloaded intrinsic. You can use ``llvm.rint`` on any
7644 floating point or vector of floating point type. Not all targets support
7649 declare float @llvm.rint.f32(float %Val)
7650 declare double @llvm.rint.f64(double %Val)
7651 declare x86_fp80 @llvm.rint.f80(x86_fp80 %Val)
7652 declare fp128 @llvm.rint.f128(fp128 %Val)
7653 declare ppc_fp128 @llvm.rint.ppcf128(ppc_fp128 %Val)
7658 The '``llvm.rint.*``' intrinsics returns the operand rounded to the
7659 nearest integer. It may raise an inexact floating-point exception if the
7660 operand isn't an integer.
7665 The argument and return value are floating point numbers of the same
7671 This function returns the same values as the libm ``rint`` functions
7672 would, and handles error conditions in the same way.
7674 '``llvm.nearbyint.*``' Intrinsic
7675 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7680 This is an overloaded intrinsic. You can use ``llvm.nearbyint`` on any
7681 floating point or vector of floating point type. Not all targets support
7686 declare float @llvm.nearbyint.f32(float %Val)
7687 declare double @llvm.nearbyint.f64(double %Val)
7688 declare x86_fp80 @llvm.nearbyint.f80(x86_fp80 %Val)
7689 declare fp128 @llvm.nearbyint.f128(fp128 %Val)
7690 declare ppc_fp128 @llvm.nearbyint.ppcf128(ppc_fp128 %Val)
7695 The '``llvm.nearbyint.*``' intrinsics returns the operand rounded to the
7701 The argument and return value are floating point numbers of the same
7707 This function returns the same values as the libm ``nearbyint``
7708 functions would, and handles error conditions in the same way.
7710 '``llvm.round.*``' Intrinsic
7711 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7716 This is an overloaded intrinsic. You can use ``llvm.round`` on any
7717 floating point or vector of floating point type. Not all targets support
7722 declare float @llvm.round.f32(float %Val)
7723 declare double @llvm.round.f64(double %Val)
7724 declare x86_fp80 @llvm.round.f80(x86_fp80 %Val)
7725 declare fp128 @llvm.round.f128(fp128 %Val)
7726 declare ppc_fp128 @llvm.round.ppcf128(ppc_fp128 %Val)
7731 The '``llvm.round.*``' intrinsics returns the operand rounded to the
7737 The argument and return value are floating point numbers of the same
7743 This function returns the same values as the libm ``round``
7744 functions would, and handles error conditions in the same way.
7746 Bit Manipulation Intrinsics
7747 ---------------------------
7749 LLVM provides intrinsics for a few important bit manipulation
7750 operations. These allow efficient code generation for some algorithms.
7752 '``llvm.bswap.*``' Intrinsics
7753 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7758 This is an overloaded intrinsic function. You can use bswap on any
7759 integer type that is an even number of bytes (i.e. BitWidth % 16 == 0).
7763 declare i16 @llvm.bswap.i16(i16 <id>)
7764 declare i32 @llvm.bswap.i32(i32 <id>)
7765 declare i64 @llvm.bswap.i64(i64 <id>)
7770 The '``llvm.bswap``' family of intrinsics is used to byte swap integer
7771 values with an even number of bytes (positive multiple of 16 bits).
7772 These are useful for performing operations on data that is not in the
7773 target's native byte order.
7778 The ``llvm.bswap.i16`` intrinsic returns an i16 value that has the high
7779 and low byte of the input i16 swapped. Similarly, the ``llvm.bswap.i32``
7780 intrinsic returns an i32 value that has the four bytes of the input i32
7781 swapped, so that if the input bytes are numbered 0, 1, 2, 3 then the
7782 returned i32 will have its bytes in 3, 2, 1, 0 order. The
7783 ``llvm.bswap.i48``, ``llvm.bswap.i64`` and other intrinsics extend this
7784 concept to additional even-byte lengths (6 bytes, 8 bytes and more,
7787 '``llvm.ctpop.*``' Intrinsic
7788 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7793 This is an overloaded intrinsic. You can use llvm.ctpop on any integer
7794 bit width, or on any vector with integer elements. Not all targets
7795 support all bit widths or vector types, however.
7799 declare i8 @llvm.ctpop.i8(i8 <src>)
7800 declare i16 @llvm.ctpop.i16(i16 <src>)
7801 declare i32 @llvm.ctpop.i32(i32 <src>)
7802 declare i64 @llvm.ctpop.i64(i64 <src>)
7803 declare i256 @llvm.ctpop.i256(i256 <src>)
7804 declare <2 x i32> @llvm.ctpop.v2i32(<2 x i32> <src>)
7809 The '``llvm.ctpop``' family of intrinsics counts the number of bits set
7815 The only argument is the value to be counted. The argument may be of any
7816 integer type, or a vector with integer elements. The return type must
7817 match the argument type.
7822 The '``llvm.ctpop``' intrinsic counts the 1's in a variable, or within
7823 each element of a vector.
7825 '``llvm.ctlz.*``' Intrinsic
7826 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7831 This is an overloaded intrinsic. You can use ``llvm.ctlz`` on any
7832 integer bit width, or any vector whose elements are integers. Not all
7833 targets support all bit widths or vector types, however.
7837 declare i8 @llvm.ctlz.i8 (i8 <src>, i1 <is_zero_undef>)
7838 declare i16 @llvm.ctlz.i16 (i16 <src>, i1 <is_zero_undef>)
7839 declare i32 @llvm.ctlz.i32 (i32 <src>, i1 <is_zero_undef>)
7840 declare i64 @llvm.ctlz.i64 (i64 <src>, i1 <is_zero_undef>)
7841 declare i256 @llvm.ctlz.i256(i256 <src>, i1 <is_zero_undef>)
7842 declase <2 x i32> @llvm.ctlz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7847 The '``llvm.ctlz``' family of intrinsic functions counts the number of
7848 leading zeros in a variable.
7853 The first argument is the value to be counted. This argument may be of
7854 any integer type, or a vectory with integer element type. The return
7855 type must match the first argument type.
7857 The second argument must be a constant and is a flag to indicate whether
7858 the intrinsic should ensure that a zero as the first argument produces a
7859 defined result. Historically some architectures did not provide a
7860 defined result for zero values as efficiently, and many algorithms are
7861 now predicated on avoiding zero-value inputs.
7866 The '``llvm.ctlz``' intrinsic counts the leading (most significant)
7867 zeros in a variable, or within each element of the vector. If
7868 ``src == 0`` then the result is the size in bits of the type of ``src``
7869 if ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7870 ``llvm.ctlz(i32 2) = 30``.
7872 '``llvm.cttz.*``' Intrinsic
7873 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
7878 This is an overloaded intrinsic. You can use ``llvm.cttz`` on any
7879 integer bit width, or any vector of integer elements. Not all targets
7880 support all bit widths or vector types, however.
7884 declare i8 @llvm.cttz.i8 (i8 <src>, i1 <is_zero_undef>)
7885 declare i16 @llvm.cttz.i16 (i16 <src>, i1 <is_zero_undef>)
7886 declare i32 @llvm.cttz.i32 (i32 <src>, i1 <is_zero_undef>)
7887 declare i64 @llvm.cttz.i64 (i64 <src>, i1 <is_zero_undef>)
7888 declare i256 @llvm.cttz.i256(i256 <src>, i1 <is_zero_undef>)
7889 declase <2 x i32> @llvm.cttz.v2i32(<2 x i32> <src>, i1 <is_zero_undef>)
7894 The '``llvm.cttz``' family of intrinsic functions counts the number of
7900 The first argument is the value to be counted. This argument may be of
7901 any integer type, or a vectory with integer element type. The return
7902 type must match the first argument type.
7904 The second argument must be a constant and is a flag to indicate whether
7905 the intrinsic should ensure that a zero as the first argument produces a
7906 defined result. Historically some architectures did not provide a
7907 defined result for zero values as efficiently, and many algorithms are
7908 now predicated on avoiding zero-value inputs.
7913 The '``llvm.cttz``' intrinsic counts the trailing (least significant)
7914 zeros in a variable, or within each element of a vector. If ``src == 0``
7915 then the result is the size in bits of the type of ``src`` if
7916 ``is_zero_undef == 0`` and ``undef`` otherwise. For example,
7917 ``llvm.cttz(2) = 1``.
7919 Arithmetic with Overflow Intrinsics
7920 -----------------------------------
7922 LLVM provides intrinsics for some arithmetic with overflow operations.
7924 '``llvm.sadd.with.overflow.*``' Intrinsics
7925 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7930 This is an overloaded intrinsic. You can use ``llvm.sadd.with.overflow``
7931 on any integer bit width.
7935 declare {i16, i1} @llvm.sadd.with.overflow.i16(i16 %a, i16 %b)
7936 declare {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7937 declare {i64, i1} @llvm.sadd.with.overflow.i64(i64 %a, i64 %b)
7942 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7943 a signed addition of the two arguments, and indicate whether an overflow
7944 occurred during the signed summation.
7949 The arguments (%a and %b) and the first element of the result structure
7950 may be of integer types of any bit width, but they must have the same
7951 bit width. The second element of the result structure must be of type
7952 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
7958 The '``llvm.sadd.with.overflow``' family of intrinsic functions perform
7959 a signed addition of the two variables. They return a structure --- the
7960 first element of which is the signed summation, and the second element
7961 of which is a bit specifying if the signed summation resulted in an
7967 .. code-block:: llvm
7969 %res = call {i32, i1} @llvm.sadd.with.overflow.i32(i32 %a, i32 %b)
7970 %sum = extractvalue {i32, i1} %res, 0
7971 %obit = extractvalue {i32, i1} %res, 1
7972 br i1 %obit, label %overflow, label %normal
7974 '``llvm.uadd.with.overflow.*``' Intrinsics
7975 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
7980 This is an overloaded intrinsic. You can use ``llvm.uadd.with.overflow``
7981 on any integer bit width.
7985 declare {i16, i1} @llvm.uadd.with.overflow.i16(i16 %a, i16 %b)
7986 declare {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
7987 declare {i64, i1} @llvm.uadd.with.overflow.i64(i64 %a, i64 %b)
7992 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
7993 an unsigned addition of the two arguments, and indicate whether a carry
7994 occurred during the unsigned summation.
7999 The arguments (%a and %b) and the first element of the result structure
8000 may be of integer types of any bit width, but they must have the same
8001 bit width. The second element of the result structure must be of type
8002 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8008 The '``llvm.uadd.with.overflow``' family of intrinsic functions perform
8009 an unsigned addition of the two arguments. They return a structure --- the
8010 first element of which is the sum, and the second element of which is a
8011 bit specifying if the unsigned summation resulted in a carry.
8016 .. code-block:: llvm
8018 %res = call {i32, i1} @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
8019 %sum = extractvalue {i32, i1} %res, 0
8020 %obit = extractvalue {i32, i1} %res, 1
8021 br i1 %obit, label %carry, label %normal
8023 '``llvm.ssub.with.overflow.*``' Intrinsics
8024 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8029 This is an overloaded intrinsic. You can use ``llvm.ssub.with.overflow``
8030 on any integer bit width.
8034 declare {i16, i1} @llvm.ssub.with.overflow.i16(i16 %a, i16 %b)
8035 declare {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8036 declare {i64, i1} @llvm.ssub.with.overflow.i64(i64 %a, i64 %b)
8041 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8042 a signed subtraction of the two arguments, and indicate whether an
8043 overflow occurred during the signed subtraction.
8048 The arguments (%a and %b) and the first element of the result structure
8049 may be of integer types of any bit width, but they must have the same
8050 bit width. The second element of the result structure must be of type
8051 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8057 The '``llvm.ssub.with.overflow``' family of intrinsic functions perform
8058 a signed subtraction of the two arguments. They return a structure --- the
8059 first element of which is the subtraction, and the second element of
8060 which is a bit specifying if the signed subtraction resulted in an
8066 .. code-block:: llvm
8068 %res = call {i32, i1} @llvm.ssub.with.overflow.i32(i32 %a, i32 %b)
8069 %sum = extractvalue {i32, i1} %res, 0
8070 %obit = extractvalue {i32, i1} %res, 1
8071 br i1 %obit, label %overflow, label %normal
8073 '``llvm.usub.with.overflow.*``' Intrinsics
8074 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8079 This is an overloaded intrinsic. You can use ``llvm.usub.with.overflow``
8080 on any integer bit width.
8084 declare {i16, i1} @llvm.usub.with.overflow.i16(i16 %a, i16 %b)
8085 declare {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8086 declare {i64, i1} @llvm.usub.with.overflow.i64(i64 %a, i64 %b)
8091 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8092 an unsigned subtraction of the two arguments, and indicate whether an
8093 overflow occurred during the unsigned subtraction.
8098 The arguments (%a and %b) and the first element of the result structure
8099 may be of integer types of any bit width, but they must have the same
8100 bit width. The second element of the result structure must be of type
8101 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8107 The '``llvm.usub.with.overflow``' family of intrinsic functions perform
8108 an unsigned subtraction of the two arguments. They return a structure ---
8109 the first element of which is the subtraction, and the second element of
8110 which is a bit specifying if the unsigned subtraction resulted in an
8116 .. code-block:: llvm
8118 %res = call {i32, i1} @llvm.usub.with.overflow.i32(i32 %a, i32 %b)
8119 %sum = extractvalue {i32, i1} %res, 0
8120 %obit = extractvalue {i32, i1} %res, 1
8121 br i1 %obit, label %overflow, label %normal
8123 '``llvm.smul.with.overflow.*``' Intrinsics
8124 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8129 This is an overloaded intrinsic. You can use ``llvm.smul.with.overflow``
8130 on any integer bit width.
8134 declare {i16, i1} @llvm.smul.with.overflow.i16(i16 %a, i16 %b)
8135 declare {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8136 declare {i64, i1} @llvm.smul.with.overflow.i64(i64 %a, i64 %b)
8141 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8142 a signed multiplication of the two arguments, and indicate whether an
8143 overflow occurred during the signed multiplication.
8148 The arguments (%a and %b) and the first element of the result structure
8149 may be of integer types of any bit width, but they must have the same
8150 bit width. The second element of the result structure must be of type
8151 ``i1``. ``%a`` and ``%b`` are the two values that will undergo signed
8157 The '``llvm.smul.with.overflow``' family of intrinsic functions perform
8158 a signed multiplication of the two arguments. They return a structure ---
8159 the first element of which is the multiplication, and the second element
8160 of which is a bit specifying if the signed multiplication resulted in an
8166 .. code-block:: llvm
8168 %res = call {i32, i1} @llvm.smul.with.overflow.i32(i32 %a, i32 %b)
8169 %sum = extractvalue {i32, i1} %res, 0
8170 %obit = extractvalue {i32, i1} %res, 1
8171 br i1 %obit, label %overflow, label %normal
8173 '``llvm.umul.with.overflow.*``' Intrinsics
8174 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8179 This is an overloaded intrinsic. You can use ``llvm.umul.with.overflow``
8180 on any integer bit width.
8184 declare {i16, i1} @llvm.umul.with.overflow.i16(i16 %a, i16 %b)
8185 declare {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8186 declare {i64, i1} @llvm.umul.with.overflow.i64(i64 %a, i64 %b)
8191 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8192 a unsigned multiplication of the two arguments, and indicate whether an
8193 overflow occurred during the unsigned multiplication.
8198 The arguments (%a and %b) and the first element of the result structure
8199 may be of integer types of any bit width, but they must have the same
8200 bit width. The second element of the result structure must be of type
8201 ``i1``. ``%a`` and ``%b`` are the two values that will undergo unsigned
8207 The '``llvm.umul.with.overflow``' family of intrinsic functions perform
8208 an unsigned multiplication of the two arguments. They return a structure ---
8209 the first element of which is the multiplication, and the second
8210 element of which is a bit specifying if the unsigned multiplication
8211 resulted in an overflow.
8216 .. code-block:: llvm
8218 %res = call {i32, i1} @llvm.umul.with.overflow.i32(i32 %a, i32 %b)
8219 %sum = extractvalue {i32, i1} %res, 0
8220 %obit = extractvalue {i32, i1} %res, 1
8221 br i1 %obit, label %overflow, label %normal
8223 Specialised Arithmetic Intrinsics
8224 ---------------------------------
8226 '``llvm.fmuladd.*``' Intrinsic
8227 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8234 declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
8235 declare double @llvm.fmuladd.f64(double %a, double %b, double %c)
8240 The '``llvm.fmuladd.*``' intrinsic functions represent multiply-add
8241 expressions that can be fused if the code generator determines that (a) the
8242 target instruction set has support for a fused operation, and (b) that the
8243 fused operation is more efficient than the equivalent, separate pair of mul
8244 and add instructions.
8249 The '``llvm.fmuladd.*``' intrinsics each take three arguments: two
8250 multiplicands, a and b, and an addend c.
8259 %0 = call float @llvm.fmuladd.f32(%a, %b, %c)
8261 is equivalent to the expression a \* b + c, except that rounding will
8262 not be performed between the multiplication and addition steps if the
8263 code generator fuses the operations. Fusion is not guaranteed, even if
8264 the target platform supports it. If a fused multiply-add is required the
8265 corresponding llvm.fma.\* intrinsic function should be used instead.
8270 .. code-block:: llvm
8272 %r2 = call float @llvm.fmuladd.f32(float %a, float %b, float %c) ; yields {float}:r2 = (a * b) + c
8274 Half Precision Floating Point Intrinsics
8275 ----------------------------------------
8277 For most target platforms, half precision floating point is a
8278 storage-only format. This means that it is a dense encoding (in memory)
8279 but does not support computation in the format.
8281 This means that code must first load the half-precision floating point
8282 value as an i16, then convert it to float with
8283 :ref:`llvm.convert.from.fp16 <int_convert_from_fp16>`. Computation can
8284 then be performed on the float value (including extending to double
8285 etc). To store the value back to memory, it is first converted to float
8286 if needed, then converted to i16 with
8287 :ref:`llvm.convert.to.fp16 <int_convert_to_fp16>`, then storing as an
8290 .. _int_convert_to_fp16:
8292 '``llvm.convert.to.fp16``' Intrinsic
8293 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8300 declare i16 @llvm.convert.to.fp16(f32 %a)
8305 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8306 from single precision floating point format to half precision floating
8312 The intrinsic function contains single argument - the value to be
8318 The '``llvm.convert.to.fp16``' intrinsic function performs a conversion
8319 from single precision floating point format to half precision floating
8320 point format. The return value is an ``i16`` which contains the
8326 .. code-block:: llvm
8328 %res = call i16 @llvm.convert.to.fp16(f32 %a)
8329 store i16 %res, i16* @x, align 2
8331 .. _int_convert_from_fp16:
8333 '``llvm.convert.from.fp16``' Intrinsic
8334 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8341 declare f32 @llvm.convert.from.fp16(i16 %a)
8346 The '``llvm.convert.from.fp16``' intrinsic function performs a
8347 conversion from half precision floating point format to single precision
8348 floating point format.
8353 The intrinsic function contains single argument - the value to be
8359 The '``llvm.convert.from.fp16``' intrinsic function performs a
8360 conversion from half single precision floating point format to single
8361 precision floating point format. The input half-float value is
8362 represented by an ``i16`` value.
8367 .. code-block:: llvm
8369 %a = load i16* @x, align 2
8370 %res = call f32 @llvm.convert.from.fp16(i16 %a)
8375 The LLVM debugger intrinsics (which all start with ``llvm.dbg.``
8376 prefix), are described in the `LLVM Source Level
8377 Debugging <SourceLevelDebugging.html#format_common_intrinsics>`_
8380 Exception Handling Intrinsics
8381 -----------------------------
8383 The LLVM exception handling intrinsics (which all start with
8384 ``llvm.eh.`` prefix), are described in the `LLVM Exception
8385 Handling <ExceptionHandling.html#format_common_intrinsics>`_ document.
8389 Trampoline Intrinsics
8390 ---------------------
8392 These intrinsics make it possible to excise one parameter, marked with
8393 the :ref:`nest <nest>` attribute, from a function. The result is a
8394 callable function pointer lacking the nest parameter - the caller does
8395 not need to provide a value for it. Instead, the value to use is stored
8396 in advance in a "trampoline", a block of memory usually allocated on the
8397 stack, which also contains code to splice the nest value into the
8398 argument list. This is used to implement the GCC nested function address
8401 For example, if the function is ``i32 f(i8* nest %c, i32 %x, i32 %y)``
8402 then the resulting function pointer has signature ``i32 (i32, i32)*``.
8403 It can be created as follows:
8405 .. code-block:: llvm
8407 %tramp = alloca [10 x i8], align 4 ; size and alignment only correct for X86
8408 %tramp1 = getelementptr [10 x i8]* %tramp, i32 0, i32 0
8409 call i8* @llvm.init.trampoline(i8* %tramp1, i8* bitcast (i32 (i8*, i32, i32)* @f to i8*), i8* %nval)
8410 %p = call i8* @llvm.adjust.trampoline(i8* %tramp1)
8411 %fp = bitcast i8* %p to i32 (i32, i32)*
8413 The call ``%val = call i32 %fp(i32 %x, i32 %y)`` is then equivalent to
8414 ``%val = call i32 %f(i8* %nval, i32 %x, i32 %y)``.
8418 '``llvm.init.trampoline``' Intrinsic
8419 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8426 declare void @llvm.init.trampoline(i8* <tramp>, i8* <func>, i8* <nval>)
8431 This fills the memory pointed to by ``tramp`` with executable code,
8432 turning it into a trampoline.
8437 The ``llvm.init.trampoline`` intrinsic takes three arguments, all
8438 pointers. The ``tramp`` argument must point to a sufficiently large and
8439 sufficiently aligned block of memory; this memory is written to by the
8440 intrinsic. Note that the size and the alignment are target-specific -
8441 LLVM currently provides no portable way of determining them, so a
8442 front-end that generates this intrinsic needs to have some
8443 target-specific knowledge. The ``func`` argument must hold a function
8444 bitcast to an ``i8*``.
8449 The block of memory pointed to by ``tramp`` is filled with target
8450 dependent code, turning it into a function. Then ``tramp`` needs to be
8451 passed to :ref:`llvm.adjust.trampoline <int_at>` to get a pointer which can
8452 be :ref:`bitcast (to a new function) and called <int_trampoline>`. The new
8453 function's signature is the same as that of ``func`` with any arguments
8454 marked with the ``nest`` attribute removed. At most one such ``nest``
8455 argument is allowed, and it must be of pointer type. Calling the new
8456 function is equivalent to calling ``func`` with the same argument list,
8457 but with ``nval`` used for the missing ``nest`` argument. If, after
8458 calling ``llvm.init.trampoline``, the memory pointed to by ``tramp`` is
8459 modified, then the effect of any later call to the returned function
8460 pointer is undefined.
8464 '``llvm.adjust.trampoline``' Intrinsic
8465 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8472 declare i8* @llvm.adjust.trampoline(i8* <tramp>)
8477 This performs any required machine-specific adjustment to the address of
8478 a trampoline (passed as ``tramp``).
8483 ``tramp`` must point to a block of memory which already has trampoline
8484 code filled in by a previous call to
8485 :ref:`llvm.init.trampoline <int_it>`.
8490 On some architectures the address of the code to be executed needs to be
8491 different to the address where the trampoline is actually stored. This
8492 intrinsic returns the executable address corresponding to ``tramp``
8493 after performing the required machine specific adjustments. The pointer
8494 returned can then be :ref:`bitcast and executed <int_trampoline>`.
8499 This class of intrinsics exists to information about the lifetime of
8500 memory objects and ranges where variables are immutable.
8502 '``llvm.lifetime.start``' Intrinsic
8503 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8510 declare void @llvm.lifetime.start(i64 <size>, i8* nocapture <ptr>)
8515 The '``llvm.lifetime.start``' intrinsic specifies the start of a memory
8521 The first argument is a constant integer representing the size of the
8522 object, or -1 if it is variable sized. The second argument is a pointer
8528 This intrinsic indicates that before this point in the code, the value
8529 of the memory pointed to by ``ptr`` is dead. This means that it is known
8530 to never be used and has an undefined value. A load from the pointer
8531 that precedes this intrinsic can be replaced with ``'undef'``.
8533 '``llvm.lifetime.end``' Intrinsic
8534 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8541 declare void @llvm.lifetime.end(i64 <size>, i8* nocapture <ptr>)
8546 The '``llvm.lifetime.end``' intrinsic specifies the end of a memory
8552 The first argument is a constant integer representing the size of the
8553 object, or -1 if it is variable sized. The second argument is a pointer
8559 This intrinsic indicates that after this point in the code, the value of
8560 the memory pointed to by ``ptr`` is dead. This means that it is known to
8561 never be used and has an undefined value. Any stores into the memory
8562 object following this intrinsic may be removed as dead.
8564 '``llvm.invariant.start``' Intrinsic
8565 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8572 declare {}* @llvm.invariant.start(i64 <size>, i8* nocapture <ptr>)
8577 The '``llvm.invariant.start``' intrinsic specifies that the contents of
8578 a memory object will not change.
8583 The first argument is a constant integer representing the size of the
8584 object, or -1 if it is variable sized. The second argument is a pointer
8590 This intrinsic indicates that until an ``llvm.invariant.end`` that uses
8591 the return value, the referenced memory location is constant and
8594 '``llvm.invariant.end``' Intrinsic
8595 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8602 declare void @llvm.invariant.end({}* <start>, i64 <size>, i8* nocapture <ptr>)
8607 The '``llvm.invariant.end``' intrinsic specifies that the contents of a
8608 memory object are mutable.
8613 The first argument is the matching ``llvm.invariant.start`` intrinsic.
8614 The second argument is a constant integer representing the size of the
8615 object, or -1 if it is variable sized and the third argument is a
8616 pointer to the object.
8621 This intrinsic indicates that the memory is mutable again.
8626 This class of intrinsics is designed to be generic and has no specific
8629 '``llvm.var.annotation``' Intrinsic
8630 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8637 declare void @llvm.var.annotation(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8642 The '``llvm.var.annotation``' intrinsic.
8647 The first argument is a pointer to a value, the second is a pointer to a
8648 global string, the third is a pointer to a global string which is the
8649 source file name, and the last argument is the line number.
8654 This intrinsic allows annotation of local variables with arbitrary
8655 strings. This can be useful for special purpose optimizations that want
8656 to look for these annotations. These have no other defined use; they are
8657 ignored by code generation and optimization.
8659 '``llvm.ptr.annotation.*``' Intrinsic
8660 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8665 This is an overloaded intrinsic. You can use '``llvm.ptr.annotation``' on a
8666 pointer to an integer of any width. *NOTE* you must specify an address space for
8667 the pointer. The identifier for the default address space is the integer
8672 declare i8* @llvm.ptr.annotation.p<address space>i8(i8* <val>, i8* <str>, i8* <str>, i32 <int>)
8673 declare i16* @llvm.ptr.annotation.p<address space>i16(i16* <val>, i8* <str>, i8* <str>, i32 <int>)
8674 declare i32* @llvm.ptr.annotation.p<address space>i32(i32* <val>, i8* <str>, i8* <str>, i32 <int>)
8675 declare i64* @llvm.ptr.annotation.p<address space>i64(i64* <val>, i8* <str>, i8* <str>, i32 <int>)
8676 declare i256* @llvm.ptr.annotation.p<address space>i256(i256* <val>, i8* <str>, i8* <str>, i32 <int>)
8681 The '``llvm.ptr.annotation``' intrinsic.
8686 The first argument is a pointer to an integer value of arbitrary bitwidth
8687 (result of some expression), the second is a pointer to a global string, the
8688 third is a pointer to a global string which is the source file name, and the
8689 last argument is the line number. It returns the value of the first argument.
8694 This intrinsic allows annotation of a pointer to an integer with arbitrary
8695 strings. This can be useful for special purpose optimizations that want to look
8696 for these annotations. These have no other defined use; they are ignored by code
8697 generation and optimization.
8699 '``llvm.annotation.*``' Intrinsic
8700 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8705 This is an overloaded intrinsic. You can use '``llvm.annotation``' on
8706 any integer bit width.
8710 declare i8 @llvm.annotation.i8(i8 <val>, i8* <str>, i8* <str>, i32 <int>)
8711 declare i16 @llvm.annotation.i16(i16 <val>, i8* <str>, i8* <str>, i32 <int>)
8712 declare i32 @llvm.annotation.i32(i32 <val>, i8* <str>, i8* <str>, i32 <int>)
8713 declare i64 @llvm.annotation.i64(i64 <val>, i8* <str>, i8* <str>, i32 <int>)
8714 declare i256 @llvm.annotation.i256(i256 <val>, i8* <str>, i8* <str>, i32 <int>)
8719 The '``llvm.annotation``' intrinsic.
8724 The first argument is an integer value (result of some expression), the
8725 second is a pointer to a global string, the third is a pointer to a
8726 global string which is the source file name, and the last argument is
8727 the line number. It returns the value of the first argument.
8732 This intrinsic allows annotations to be put on arbitrary expressions
8733 with arbitrary strings. This can be useful for special purpose
8734 optimizations that want to look for these annotations. These have no
8735 other defined use; they are ignored by code generation and optimization.
8737 '``llvm.trap``' Intrinsic
8738 ^^^^^^^^^^^^^^^^^^^^^^^^^
8745 declare void @llvm.trap() noreturn nounwind
8750 The '``llvm.trap``' intrinsic.
8760 This intrinsic is lowered to the target dependent trap instruction. If
8761 the target does not have a trap instruction, this intrinsic will be
8762 lowered to a call of the ``abort()`` function.
8764 '``llvm.debugtrap``' Intrinsic
8765 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8772 declare void @llvm.debugtrap() nounwind
8777 The '``llvm.debugtrap``' intrinsic.
8787 This intrinsic is lowered to code which is intended to cause an
8788 execution trap with the intention of requesting the attention of a
8791 '``llvm.stackprotector``' Intrinsic
8792 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8799 declare void @llvm.stackprotector(i8* <guard>, i8** <slot>)
8804 The ``llvm.stackprotector`` intrinsic takes the ``guard`` and stores it
8805 onto the stack at ``slot``. The stack slot is adjusted to ensure that it
8806 is placed on the stack before local variables.
8811 The ``llvm.stackprotector`` intrinsic requires two pointer arguments.
8812 The first argument is the value loaded from the stack guard
8813 ``@__stack_chk_guard``. The second variable is an ``alloca`` that has
8814 enough space to hold the value of the guard.
8819 This intrinsic causes the prologue/epilogue inserter to force the position of
8820 the ``AllocaInst`` stack slot to be before local variables on the stack. This is
8821 to ensure that if a local variable on the stack is overwritten, it will destroy
8822 the value of the guard. When the function exits, the guard on the stack is
8823 checked against the original guard by ``llvm.stackprotectorcheck``. If they are
8824 different, then ``llvm.stackprotectorcheck`` causes the program to abort by
8825 calling the ``__stack_chk_fail()`` function.
8827 '``llvm.stackprotectorcheck``' Intrinsic
8828 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8835 declare void @llvm.stackprotectorcheck(i8** <guard>)
8840 The ``llvm.stackprotectorcheck`` intrinsic compares ``guard`` against an already
8841 created stack protector and if they are not equal calls the
8842 ``__stack_chk_fail()`` function.
8847 The ``llvm.stackprotectorcheck`` intrinsic requires one pointer argument, the
8848 the variable ``@__stack_chk_guard``.
8853 This intrinsic is provided to perform the stack protector check by comparing
8854 ``guard`` with the stack slot created by ``llvm.stackprotector`` and if the
8855 values do not match call the ``__stack_chk_fail()`` function.
8857 The reason to provide this as an IR level intrinsic instead of implementing it
8858 via other IR operations is that in order to perform this operation at the IR
8859 level without an intrinsic, one would need to create additional basic blocks to
8860 handle the success/failure cases. This makes it difficult to stop the stack
8861 protector check from disrupting sibling tail calls in Codegen. With this
8862 intrinsic, we are able to generate the stack protector basic blocks late in
8863 codegen after the tail call decision has occurred.
8865 '``llvm.objectsize``' Intrinsic
8866 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8873 declare i32 @llvm.objectsize.i32(i8* <object>, i1 <min>)
8874 declare i64 @llvm.objectsize.i64(i8* <object>, i1 <min>)
8879 The ``llvm.objectsize`` intrinsic is designed to provide information to
8880 the optimizers to determine at compile time whether a) an operation
8881 (like memcpy) will overflow a buffer that corresponds to an object, or
8882 b) that a runtime check for overflow isn't necessary. An object in this
8883 context means an allocation of a specific class, structure, array, or
8889 The ``llvm.objectsize`` intrinsic takes two arguments. The first
8890 argument is a pointer to or into the ``object``. The second argument is
8891 a boolean and determines whether ``llvm.objectsize`` returns 0 (if true)
8892 or -1 (if false) when the object size is unknown. The second argument
8893 only accepts constants.
8898 The ``llvm.objectsize`` intrinsic is lowered to a constant representing
8899 the size of the object concerned. If the size cannot be determined at
8900 compile time, ``llvm.objectsize`` returns ``i32/i64 -1 or 0`` (depending
8901 on the ``min`` argument).
8903 '``llvm.expect``' Intrinsic
8904 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
8911 declare i32 @llvm.expect.i32(i32 <val>, i32 <expected_val>)
8912 declare i64 @llvm.expect.i64(i64 <val>, i64 <expected_val>)
8917 The ``llvm.expect`` intrinsic provides information about expected (the
8918 most probable) value of ``val``, which can be used by optimizers.
8923 The ``llvm.expect`` intrinsic takes two arguments. The first argument is
8924 a value. The second argument is an expected value, this needs to be a
8925 constant value, variables are not allowed.
8930 This intrinsic is lowered to the ``val``.
8932 '``llvm.donothing``' Intrinsic
8933 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
8940 declare void @llvm.donothing() nounwind readnone
8945 The ``llvm.donothing`` intrinsic doesn't perform any operation. It's the
8946 only intrinsic that can be called with an invoke instruction.
8956 This intrinsic does nothing, and it's removed by optimizers and ignored